diff --git a/include/sys/zthr.h b/include/sys/zthr.h index ae8c57e9eea2..19be89eeebe5 100644 --- a/include/sys/zthr.h +++ b/include/sys/zthr.h @@ -1,42 +1,43 @@ /* * CDDL HEADER START * * This file and its contents are supplied under the terms of the * Common Development and Distribution License ("CDDL"), version 1.0. * You may only use this file in accordance with the terms of version * 1.0 of the CDDL. * * A full copy of the text of the CDDL should have accompanied this * source. A copy of the CDDL is also available via the Internet at * http://www.illumos.org/license/CDDL. * * CDDL HEADER END */ /* * Copyright (c) 2017, 2018 by Delphix. All rights reserved. */ #ifndef _SYS_ZTHR_H #define _SYS_ZTHR_H typedef struct zthr zthr_t; typedef void (zthr_func_t)(void *, zthr_t *); typedef boolean_t (zthr_checkfunc_t)(void *, zthr_t *); extern zthr_t *zthr_create(const char *zthr_name, - zthr_checkfunc_t checkfunc, zthr_func_t *func, void *arg); + zthr_checkfunc_t checkfunc, zthr_func_t *func, void *arg, + pri_t pri); extern zthr_t *zthr_create_timer(const char *zthr_name, zthr_checkfunc_t *checkfunc, zthr_func_t *func, void *arg, - hrtime_t nano_wait); + hrtime_t nano_wait, pri_t pri); extern void zthr_destroy(zthr_t *t); extern void zthr_wakeup(zthr_t *t); extern void zthr_cancel(zthr_t *t); extern void zthr_resume(zthr_t *t); extern void zthr_wait_cycle_done(zthr_t *t); extern boolean_t zthr_iscancelled(zthr_t *t); extern boolean_t zthr_has_waiters(zthr_t *t); #endif /* _SYS_ZTHR_H */ diff --git a/module/zfs/arc.c b/module/zfs/arc.c index 81cb4dabc038..0a72d8708e83 100644 --- a/module/zfs/arc.c +++ b/module/zfs/arc.c @@ -1,11078 +1,11078 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2018, Joyent, Inc. * Copyright (c) 2011, 2020, Delphix. All rights reserved. * Copyright (c) 2014, Saso Kiselkov. All rights reserved. * Copyright (c) 2017, Nexenta Systems, Inc. All rights reserved. * Copyright (c) 2019, loli10K . All rights reserved. * Copyright (c) 2020, George Amanakis. All rights reserved. * Copyright (c) 2019, Klara Inc. * Copyright (c) 2019, Allan Jude * Copyright (c) 2020, The FreeBSD Foundation [1] * * [1] Portions of this software were developed by Allan Jude * under sponsorship from the FreeBSD Foundation. */ /* * DVA-based Adjustable Replacement Cache * * While much of the theory of operation used here is * based on the self-tuning, low overhead replacement cache * presented by Megiddo and Modha at FAST 2003, there are some * significant differences: * * 1. The Megiddo and Modha model assumes any page is evictable. * Pages in its cache cannot be "locked" into memory. This makes * the eviction algorithm simple: evict the last page in the list. * This also make the performance characteristics easy to reason * about. Our cache is not so simple. At any given moment, some * subset of the blocks in the cache are un-evictable because we * have handed out a reference to them. Blocks are only evictable * when there are no external references active. This makes * eviction far more problematic: we choose to evict the evictable * blocks that are the "lowest" in the list. * * There are times when it is not possible to evict the requested * space. In these circumstances we are unable to adjust the cache * size. To prevent the cache growing unbounded at these times we * implement a "cache throttle" that slows the flow of new data * into the cache until we can make space available. * * 2. The Megiddo and Modha model assumes a fixed cache size. * Pages are evicted when the cache is full and there is a cache * miss. Our model has a variable sized cache. It grows with * high use, but also tries to react to memory pressure from the * operating system: decreasing its size when system memory is * tight. * * 3. The Megiddo and Modha model assumes a fixed page size. All * elements of the cache are therefore exactly the same size. So * when adjusting the cache size following a cache miss, its simply * a matter of choosing a single page to evict. In our model, we * have variable sized cache blocks (ranging from 512 bytes to * 128K bytes). We therefore choose a set of blocks to evict to make * space for a cache miss that approximates as closely as possible * the space used by the new block. * * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" * by N. Megiddo & D. Modha, FAST 2003 */ /* * The locking model: * * A new reference to a cache buffer can be obtained in two * ways: 1) via a hash table lookup using the DVA as a key, * or 2) via one of the ARC lists. The arc_read() interface * uses method 1, while the internal ARC algorithms for * adjusting the cache use method 2. We therefore provide two * types of locks: 1) the hash table lock array, and 2) the * ARC list locks. * * Buffers do not have their own mutexes, rather they rely on the * hash table mutexes for the bulk of their protection (i.e. most * fields in the arc_buf_hdr_t are protected by these mutexes). * * buf_hash_find() returns the appropriate mutex (held) when it * locates the requested buffer in the hash table. It returns * NULL for the mutex if the buffer was not in the table. * * buf_hash_remove() expects the appropriate hash mutex to be * already held before it is invoked. * * Each ARC state also has a mutex which is used to protect the * buffer list associated with the state. When attempting to * obtain a hash table lock while holding an ARC list lock you * must use: mutex_tryenter() to avoid deadlock. Also note that * the active state mutex must be held before the ghost state mutex. * * It as also possible to register a callback which is run when the * arc_meta_limit is reached and no buffers can be safely evicted. In * this case the arc user should drop a reference on some arc buffers so * they can be reclaimed and the arc_meta_limit honored. For example, * when using the ZPL each dentry holds a references on a znode. These * dentries must be pruned before the arc buffer holding the znode can * be safely evicted. * * Note that the majority of the performance stats are manipulated * with atomic operations. * * The L2ARC uses the l2ad_mtx on each vdev for the following: * * - L2ARC buflist creation * - L2ARC buflist eviction * - L2ARC write completion, which walks L2ARC buflists * - ARC header destruction, as it removes from L2ARC buflists * - ARC header release, as it removes from L2ARC buflists */ /* * ARC operation: * * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. * This structure can point either to a block that is still in the cache or to * one that is only accessible in an L2 ARC device, or it can provide * information about a block that was recently evicted. If a block is * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough * information to retrieve it from the L2ARC device. This information is * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block * that is in this state cannot access the data directly. * * Blocks that are actively being referenced or have not been evicted * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within * the arc_buf_hdr_t that will point to the data block in memory. A block can * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd). * * The L1ARC's data pointer may or may not be uncompressed. The ARC has the * ability to store the physical data (b_pabd) associated with the DVA of the * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block, * it will match its on-disk compression characteristics. This behavior can be * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the * compressed ARC functionality is disabled, the b_pabd will point to an * uncompressed version of the on-disk data. * * Data in the L1ARC is not accessed by consumers of the ARC directly. Each * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC * consumer. The ARC will provide references to this data and will keep it * cached until it is no longer in use. The ARC caches only the L1ARC's physical * data block and will evict any arc_buf_t that is no longer referenced. The * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the * "overhead_size" kstat. * * Depending on the consumer, an arc_buf_t can be requested in uncompressed or * compressed form. The typical case is that consumers will want uncompressed * data, and when that happens a new data buffer is allocated where the data is * decompressed for them to use. Currently the only consumer who wants * compressed arc_buf_t's is "zfs send", when it streams data exactly as it * exists on disk. When this happens, the arc_buf_t's data buffer is shared * with the arc_buf_hdr_t. * * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The * first one is owned by a compressed send consumer (and therefore references * the same compressed data buffer as the arc_buf_hdr_t) and the second could be * used by any other consumer (and has its own uncompressed copy of the data * buffer). * * arc_buf_hdr_t * +-----------+ * | fields | * | common to | * | L1- and | * | L2ARC | * +-----------+ * | l2arc_buf_hdr_t * | | * +-----------+ * | l1arc_buf_hdr_t * | | arc_buf_t * | b_buf +------------>+-----------+ arc_buf_t * | b_pabd +-+ |b_next +---->+-----------+ * +-----------+ | |-----------| |b_next +-->NULL * | |b_comp = T | +-----------+ * | |b_data +-+ |b_comp = F | * | +-----------+ | |b_data +-+ * +->+------+ | +-----------+ | * compressed | | | | * data | |<--------------+ | uncompressed * +------+ compressed, | data * shared +-->+------+ * data | | * | | * +------+ * * When a consumer reads a block, the ARC must first look to see if the * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new * arc_buf_t and either copies uncompressed data into a new data buffer from an * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the * hdr is compressed and the desired compression characteristics of the * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be * the last buffer in the hdr's b_buf list, however a shared compressed buf can * be anywhere in the hdr's list. * * The diagram below shows an example of an uncompressed ARC hdr that is * sharing its data with an arc_buf_t (note that the shared uncompressed buf is * the last element in the buf list): * * arc_buf_hdr_t * +-----------+ * | | * | | * | | * +-----------+ * l2arc_buf_hdr_t| | * | | * +-----------+ * l1arc_buf_hdr_t| | * | | arc_buf_t (shared) * | b_buf +------------>+---------+ arc_buf_t * | | |b_next +---->+---------+ * | b_pabd +-+ |---------| |b_next +-->NULL * +-----------+ | | | +---------+ * | |b_data +-+ | | * | +---------+ | |b_data +-+ * +->+------+ | +---------+ | * | | | | * uncompressed | | | | * data +------+ | | * ^ +->+------+ | * | uncompressed | | | * | data | | | * | +------+ | * +---------------------------------+ * * Writing to the ARC requires that the ARC first discard the hdr's b_pabd * since the physical block is about to be rewritten. The new data contents * will be contained in the arc_buf_t. As the I/O pipeline performs the write, * it may compress the data before writing it to disk. The ARC will be called * with the transformed data and will bcopy the transformed on-disk block into * a newly allocated b_pabd. Writes are always done into buffers which have * either been loaned (and hence are new and don't have other readers) or * buffers which have been released (and hence have their own hdr, if there * were originally other readers of the buf's original hdr). This ensures that * the ARC only needs to update a single buf and its hdr after a write occurs. * * When the L2ARC is in use, it will also take advantage of the b_pabd. The * L2ARC will always write the contents of b_pabd to the L2ARC. This means * that when compressed ARC is enabled that the L2ARC blocks are identical * to the on-disk block in the main data pool. This provides a significant * advantage since the ARC can leverage the bp's checksum when reading from the * L2ARC to determine if the contents are valid. However, if the compressed * ARC is disabled, then the L2ARC's block must be transformed to look * like the physical block in the main data pool before comparing the * checksum and determining its validity. * * The L1ARC has a slightly different system for storing encrypted data. * Raw (encrypted + possibly compressed) data has a few subtle differences from * data that is just compressed. The biggest difference is that it is not * possible to decrypt encrypted data (or vice-versa) if the keys aren't loaded. * The other difference is that encryption cannot be treated as a suggestion. * If a caller would prefer compressed data, but they actually wind up with * uncompressed data the worst thing that could happen is there might be a * performance hit. If the caller requests encrypted data, however, we must be * sure they actually get it or else secret information could be leaked. Raw * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore, * may have both an encrypted version and a decrypted version of its data at * once. When a caller needs a raw arc_buf_t, it is allocated and the data is * copied out of this header. To avoid complications with b_pabd, raw buffers * cannot be shared. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifndef _KERNEL /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ boolean_t arc_watch = B_FALSE; #endif /* * This thread's job is to keep enough free memory in the system, by * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves * arc_available_memory(). */ static zthr_t *arc_reap_zthr; /* * This thread's job is to keep arc_size under arc_c, by calling * arc_evict(), which improves arc_is_overflowing(). */ static zthr_t *arc_evict_zthr; static kmutex_t arc_evict_lock; static boolean_t arc_evict_needed = B_FALSE; /* * Count of bytes evicted since boot. */ static uint64_t arc_evict_count; /* * List of arc_evict_waiter_t's, representing threads waiting for the * arc_evict_count to reach specific values. */ static list_t arc_evict_waiters; /* * When arc_is_overflowing(), arc_get_data_impl() waits for this percent of * the requested amount of data to be evicted. For example, by default for * every 2KB that's evicted, 1KB of it may be "reused" by a new allocation. * Since this is above 100%, it ensures that progress is made towards getting * arc_size under arc_c. Since this is finite, it ensures that allocations * can still happen, even during the potentially long time that arc_size is * more than arc_c. */ int zfs_arc_eviction_pct = 200; /* * The number of headers to evict in arc_evict_state_impl() before * dropping the sublist lock and evicting from another sublist. A lower * value means we're more likely to evict the "correct" header (i.e. the * oldest header in the arc state), but comes with higher overhead * (i.e. more invocations of arc_evict_state_impl()). */ int zfs_arc_evict_batch_limit = 10; /* number of seconds before growing cache again */ int arc_grow_retry = 5; /* * Minimum time between calls to arc_kmem_reap_soon(). */ int arc_kmem_cache_reap_retry_ms = 1000; /* shift of arc_c for calculating overflow limit in arc_get_data_impl */ int zfs_arc_overflow_shift = 8; /* shift of arc_c for calculating both min and max arc_p */ int arc_p_min_shift = 4; /* log2(fraction of arc to reclaim) */ int arc_shrink_shift = 7; /* percent of pagecache to reclaim arc to */ #ifdef _KERNEL uint_t zfs_arc_pc_percent = 0; #endif /* * log2(fraction of ARC which must be free to allow growing). * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, * when reading a new block into the ARC, we will evict an equal-sized block * from the ARC. * * This must be less than arc_shrink_shift, so that when we shrink the ARC, * we will still not allow it to grow. */ int arc_no_grow_shift = 5; /* * minimum lifespan of a prefetch block in clock ticks * (initialized in arc_init()) */ static int arc_min_prefetch_ms; static int arc_min_prescient_prefetch_ms; /* * If this percent of memory is free, don't throttle. */ int arc_lotsfree_percent = 10; /* * The arc has filled available memory and has now warmed up. */ boolean_t arc_warm; /* * These tunables are for performance analysis. */ unsigned long zfs_arc_max = 0; unsigned long zfs_arc_min = 0; unsigned long zfs_arc_meta_limit = 0; unsigned long zfs_arc_meta_min = 0; unsigned long zfs_arc_dnode_limit = 0; unsigned long zfs_arc_dnode_reduce_percent = 10; int zfs_arc_grow_retry = 0; int zfs_arc_shrink_shift = 0; int zfs_arc_p_min_shift = 0; int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ /* * ARC dirty data constraints for arc_tempreserve_space() throttle. */ unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */ unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */ unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */ /* * Enable or disable compressed arc buffers. */ int zfs_compressed_arc_enabled = B_TRUE; /* * ARC will evict meta buffers that exceed arc_meta_limit. This * tunable make arc_meta_limit adjustable for different workloads. */ unsigned long zfs_arc_meta_limit_percent = 75; /* * Percentage that can be consumed by dnodes of ARC meta buffers. */ unsigned long zfs_arc_dnode_limit_percent = 10; /* * These tunables are Linux specific */ unsigned long zfs_arc_sys_free = 0; int zfs_arc_min_prefetch_ms = 0; int zfs_arc_min_prescient_prefetch_ms = 0; int zfs_arc_p_dampener_disable = 1; int zfs_arc_meta_prune = 10000; int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED; int zfs_arc_meta_adjust_restarts = 4096; int zfs_arc_lotsfree_percent = 10; /* The 6 states: */ arc_state_t ARC_anon; arc_state_t ARC_mru; arc_state_t ARC_mru_ghost; arc_state_t ARC_mfu; arc_state_t ARC_mfu_ghost; arc_state_t ARC_l2c_only; arc_stats_t arc_stats = { { "hits", KSTAT_DATA_UINT64 }, { "misses", KSTAT_DATA_UINT64 }, { "demand_data_hits", KSTAT_DATA_UINT64 }, { "demand_data_misses", KSTAT_DATA_UINT64 }, { "demand_metadata_hits", KSTAT_DATA_UINT64 }, { "demand_metadata_misses", KSTAT_DATA_UINT64 }, { "prefetch_data_hits", KSTAT_DATA_UINT64 }, { "prefetch_data_misses", KSTAT_DATA_UINT64 }, { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, { "mru_hits", KSTAT_DATA_UINT64 }, { "mru_ghost_hits", KSTAT_DATA_UINT64 }, { "mfu_hits", KSTAT_DATA_UINT64 }, { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, { "deleted", KSTAT_DATA_UINT64 }, { "mutex_miss", KSTAT_DATA_UINT64 }, { "access_skip", KSTAT_DATA_UINT64 }, { "evict_skip", KSTAT_DATA_UINT64 }, { "evict_not_enough", KSTAT_DATA_UINT64 }, { "evict_l2_cached", KSTAT_DATA_UINT64 }, { "evict_l2_eligible", KSTAT_DATA_UINT64 }, { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 }, { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 }, { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, { "evict_l2_skip", KSTAT_DATA_UINT64 }, { "hash_elements", KSTAT_DATA_UINT64 }, { "hash_elements_max", KSTAT_DATA_UINT64 }, { "hash_collisions", KSTAT_DATA_UINT64 }, { "hash_chains", KSTAT_DATA_UINT64 }, { "hash_chain_max", KSTAT_DATA_UINT64 }, { "p", KSTAT_DATA_UINT64 }, { "c", KSTAT_DATA_UINT64 }, { "c_min", KSTAT_DATA_UINT64 }, { "c_max", KSTAT_DATA_UINT64 }, { "size", KSTAT_DATA_UINT64 }, { "compressed_size", KSTAT_DATA_UINT64 }, { "uncompressed_size", KSTAT_DATA_UINT64 }, { "overhead_size", KSTAT_DATA_UINT64 }, { "hdr_size", KSTAT_DATA_UINT64 }, { "data_size", KSTAT_DATA_UINT64 }, { "metadata_size", KSTAT_DATA_UINT64 }, { "dbuf_size", KSTAT_DATA_UINT64 }, { "dnode_size", KSTAT_DATA_UINT64 }, { "bonus_size", KSTAT_DATA_UINT64 }, #if defined(COMPAT_FREEBSD11) { "other_size", KSTAT_DATA_UINT64 }, #endif { "anon_size", KSTAT_DATA_UINT64 }, { "anon_evictable_data", KSTAT_DATA_UINT64 }, { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, { "mru_size", KSTAT_DATA_UINT64 }, { "mru_evictable_data", KSTAT_DATA_UINT64 }, { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, { "mru_ghost_size", KSTAT_DATA_UINT64 }, { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, { "mfu_size", KSTAT_DATA_UINT64 }, { "mfu_evictable_data", KSTAT_DATA_UINT64 }, { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, { "mfu_ghost_size", KSTAT_DATA_UINT64 }, { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, { "l2_hits", KSTAT_DATA_UINT64 }, { "l2_misses", KSTAT_DATA_UINT64 }, { "l2_prefetch_asize", KSTAT_DATA_UINT64 }, { "l2_mru_asize", KSTAT_DATA_UINT64 }, { "l2_mfu_asize", KSTAT_DATA_UINT64 }, { "l2_bufc_data_asize", KSTAT_DATA_UINT64 }, { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 }, { "l2_feeds", KSTAT_DATA_UINT64 }, { "l2_rw_clash", KSTAT_DATA_UINT64 }, { "l2_read_bytes", KSTAT_DATA_UINT64 }, { "l2_write_bytes", KSTAT_DATA_UINT64 }, { "l2_writes_sent", KSTAT_DATA_UINT64 }, { "l2_writes_done", KSTAT_DATA_UINT64 }, { "l2_writes_error", KSTAT_DATA_UINT64 }, { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, { "l2_evict_reading", KSTAT_DATA_UINT64 }, { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, { "l2_free_on_write", KSTAT_DATA_UINT64 }, { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, { "l2_cksum_bad", KSTAT_DATA_UINT64 }, { "l2_io_error", KSTAT_DATA_UINT64 }, { "l2_size", KSTAT_DATA_UINT64 }, { "l2_asize", KSTAT_DATA_UINT64 }, { "l2_hdr_size", KSTAT_DATA_UINT64 }, { "l2_log_blk_writes", KSTAT_DATA_UINT64 }, { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 }, { "l2_log_blk_asize", KSTAT_DATA_UINT64 }, { "l2_log_blk_count", KSTAT_DATA_UINT64 }, { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 }, { "l2_rebuild_success", KSTAT_DATA_UINT64 }, { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 }, { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 }, { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 }, { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 }, { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 }, { "l2_rebuild_size", KSTAT_DATA_UINT64 }, { "l2_rebuild_asize", KSTAT_DATA_UINT64 }, { "l2_rebuild_bufs", KSTAT_DATA_UINT64 }, { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 }, { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 }, { "memory_throttle_count", KSTAT_DATA_UINT64 }, { "memory_direct_count", KSTAT_DATA_UINT64 }, { "memory_indirect_count", KSTAT_DATA_UINT64 }, { "memory_all_bytes", KSTAT_DATA_UINT64 }, { "memory_free_bytes", KSTAT_DATA_UINT64 }, { "memory_available_bytes", KSTAT_DATA_INT64 }, { "arc_no_grow", KSTAT_DATA_UINT64 }, { "arc_tempreserve", KSTAT_DATA_UINT64 }, { "arc_loaned_bytes", KSTAT_DATA_UINT64 }, { "arc_prune", KSTAT_DATA_UINT64 }, { "arc_meta_used", KSTAT_DATA_UINT64 }, { "arc_meta_limit", KSTAT_DATA_UINT64 }, { "arc_dnode_limit", KSTAT_DATA_UINT64 }, { "arc_meta_max", KSTAT_DATA_UINT64 }, { "arc_meta_min", KSTAT_DATA_UINT64 }, { "async_upgrade_sync", KSTAT_DATA_UINT64 }, { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 }, { "arc_need_free", KSTAT_DATA_UINT64 }, { "arc_sys_free", KSTAT_DATA_UINT64 }, { "arc_raw_size", KSTAT_DATA_UINT64 }, { "cached_only_in_progress", KSTAT_DATA_UINT64 }, { "abd_chunk_waste_size", KSTAT_DATA_UINT64 }, }; arc_sums_t arc_sums; #define ARCSTAT_MAX(stat, val) { \ uint64_t m; \ while ((val) > (m = arc_stats.stat.value.ui64) && \ (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ continue; \ } /* * We define a macro to allow ARC hits/misses to be easily broken down by * two separate conditions, giving a total of four different subtypes for * each of hits and misses (so eight statistics total). */ #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ if (cond1) { \ if (cond2) { \ ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ } else { \ ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ } \ } else { \ if (cond2) { \ ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ } else { \ ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ } \ } /* * This macro allows us to use kstats as floating averages. Each time we * update this kstat, we first factor it and the update value by * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall * average. This macro assumes that integer loads and stores are atomic, but * is not safe for multiple writers updating the kstat in parallel (only the * last writer's update will remain). */ #define ARCSTAT_F_AVG_FACTOR 3 #define ARCSTAT_F_AVG(stat, value) \ do { \ uint64_t x = ARCSTAT(stat); \ x = x - x / ARCSTAT_F_AVG_FACTOR + \ (value) / ARCSTAT_F_AVG_FACTOR; \ ARCSTAT(stat) = x; \ _NOTE(CONSTCOND) \ } while (0) kstat_t *arc_ksp; /* * There are several ARC variables that are critical to export as kstats -- * but we don't want to have to grovel around in the kstat whenever we wish to * manipulate them. For these variables, we therefore define them to be in * terms of the statistic variable. This assures that we are not introducing * the possibility of inconsistency by having shadow copies of the variables, * while still allowing the code to be readable. */ #define arc_tempreserve ARCSTAT(arcstat_tempreserve) #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes) #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ /* max size for dnodes */ #define arc_dnode_size_limit ARCSTAT(arcstat_dnode_limit) #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */ hrtime_t arc_growtime; list_t arc_prune_list; kmutex_t arc_prune_mtx; taskq_t *arc_prune_taskq; #define GHOST_STATE(state) \ ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ (state) == arc_l2c_only) #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) #define HDR_PRESCIENT_PREFETCH(hdr) \ ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) #define HDR_COMPRESSION_ENABLED(hdr) \ ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) #define HDR_L2_READING(hdr) \ (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED) #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH) #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) #define HDR_ISTYPE_METADATA(hdr) \ ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) #define HDR_HAS_RABD(hdr) \ (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \ (hdr)->b_crypt_hdr.b_rabd != NULL) #define HDR_ENCRYPTED(hdr) \ (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot)) #define HDR_AUTHENTICATED(hdr) \ (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot)) /* For storing compression mode in b_flags */ #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED) /* * Other sizes */ #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr)) #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) /* * Hash table routines */ #define BUF_LOCKS 2048 typedef struct buf_hash_table { uint64_t ht_mask; arc_buf_hdr_t **ht_table; kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned; } buf_hash_table_t; static buf_hash_table_t buf_hash_table; #define BUF_HASH_INDEX(spa, dva, birth) \ (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) #define HDR_LOCK(hdr) \ (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) uint64_t zfs_crc64_table[256]; /* * Level 2 ARC */ #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ #define L2ARC_HEADROOM 2 /* num of writes */ /* * If we discover during ARC scan any buffers to be compressed, we boost * our headroom for the next scanning cycle by this percentage multiple. */ #define L2ARC_HEADROOM_BOOST 200 #define L2ARC_FEED_SECS 1 /* caching interval secs */ #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ /* * We can feed L2ARC from two states of ARC buffers, mru and mfu, * and each of the state has two types: data and metadata. */ #define L2ARC_FEED_TYPES 4 /* L2ARC Performance Tunables */ unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */ unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */ unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */ unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */ int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ int l2arc_feed_again = B_TRUE; /* turbo warmup */ int l2arc_norw = B_FALSE; /* no reads during writes */ int l2arc_meta_percent = 33; /* limit on headers size */ /* * L2ARC Internals */ static list_t L2ARC_dev_list; /* device list */ static list_t *l2arc_dev_list; /* device list pointer */ static kmutex_t l2arc_dev_mtx; /* device list mutex */ static l2arc_dev_t *l2arc_dev_last; /* last device used */ static list_t L2ARC_free_on_write; /* free after write buf list */ static list_t *l2arc_free_on_write; /* free after write list ptr */ static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ static uint64_t l2arc_ndev; /* number of devices */ typedef struct l2arc_read_callback { arc_buf_hdr_t *l2rcb_hdr; /* read header */ blkptr_t l2rcb_bp; /* original blkptr */ zbookmark_phys_t l2rcb_zb; /* original bookmark */ int l2rcb_flags; /* original flags */ abd_t *l2rcb_abd; /* temporary buffer */ } l2arc_read_callback_t; typedef struct l2arc_data_free { /* protected by l2arc_free_on_write_mtx */ abd_t *l2df_abd; size_t l2df_size; arc_buf_contents_t l2df_type; list_node_t l2df_list_node; } l2arc_data_free_t; typedef enum arc_fill_flags { ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */ ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */ ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */ ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */ ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */ } arc_fill_flags_t; typedef enum arc_ovf_level { ARC_OVF_NONE, /* ARC within target size. */ ARC_OVF_SOME, /* ARC is slightly overflowed. */ ARC_OVF_SEVERE /* ARC is severely overflowed. */ } arc_ovf_level_t; static kmutex_t l2arc_feed_thr_lock; static kcondvar_t l2arc_feed_thr_cv; static uint8_t l2arc_thread_exit; static kmutex_t l2arc_rebuild_thr_lock; static kcondvar_t l2arc_rebuild_thr_cv; enum arc_hdr_alloc_flags { ARC_HDR_ALLOC_RDATA = 0x1, ARC_HDR_DO_ADAPT = 0x2, }; static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, boolean_t); static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, boolean_t); static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t); static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int); static void arc_access(arc_buf_hdr_t *, kmutex_t *); static void arc_buf_watch(arc_buf_t *); static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); static uint32_t arc_bufc_to_flags(arc_buf_contents_t); static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); static void l2arc_read_done(zio_t *); static void l2arc_do_free_on_write(void); static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr, boolean_t state_only); #define l2arc_hdr_arcstats_increment(hdr) \ l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE) #define l2arc_hdr_arcstats_decrement(hdr) \ l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE) #define l2arc_hdr_arcstats_increment_state(hdr) \ l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE) #define l2arc_hdr_arcstats_decrement_state(hdr) \ l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE) /* * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU * metadata and data are cached from ARC into L2ARC. */ int l2arc_mfuonly = 0; /* * L2ARC TRIM * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of * the current write size (l2arc_write_max) we should TRIM if we * have filled the device. It is defined as a percentage of the * write size. If set to 100 we trim twice the space required to * accommodate upcoming writes. A minimum of 64MB will be trimmed. * It also enables TRIM of the whole L2ARC device upon creation or * addition to an existing pool or if the header of the device is * invalid upon importing a pool or onlining a cache device. The * default is 0, which disables TRIM on L2ARC altogether as it can * put significant stress on the underlying storage devices. This * will vary depending of how well the specific device handles * these commands. */ unsigned long l2arc_trim_ahead = 0; /* * Performance tuning of L2ARC persistence: * * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding * an L2ARC device (either at pool import or later) will attempt * to rebuild L2ARC buffer contents. * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls * whether log blocks are written to the L2ARC device. If the L2ARC * device is less than 1GB, the amount of data l2arc_evict() * evicts is significant compared to the amount of restored L2ARC * data. In this case do not write log blocks in L2ARC in order * not to waste space. */ int l2arc_rebuild_enabled = B_TRUE; unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024; /* L2ARC persistence rebuild control routines. */ void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen); static void l2arc_dev_rebuild_thread(void *arg); static int l2arc_rebuild(l2arc_dev_t *dev); /* L2ARC persistence read I/O routines. */ static int l2arc_dev_hdr_read(l2arc_dev_t *dev); static int l2arc_log_blk_read(l2arc_dev_t *dev, const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp, l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb, zio_t *this_io, zio_t **next_io); static zio_t *l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb); static void l2arc_log_blk_fetch_abort(zio_t *zio); /* L2ARC persistence block restoration routines. */ static void l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb, uint64_t lb_asize); static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev); /* L2ARC persistence write I/O routines. */ static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb); /* L2ARC persistence auxiliary routines. */ boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp); static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *ab); boolean_t l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check); static void l2arc_blk_fetch_done(zio_t *zio); static inline uint64_t l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev); /* * We use Cityhash for this. It's fast, and has good hash properties without * requiring any large static buffers. */ static uint64_t buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) { return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); } #define HDR_EMPTY(hdr) \ ((hdr)->b_dva.dva_word[0] == 0 && \ (hdr)->b_dva.dva_word[1] == 0) #define HDR_EMPTY_OR_LOCKED(hdr) \ (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr))) #define HDR_EQUAL(spa, dva, birth, hdr) \ ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) static void buf_discard_identity(arc_buf_hdr_t *hdr) { hdr->b_dva.dva_word[0] = 0; hdr->b_dva.dva_word[1] = 0; hdr->b_birth = 0; } static arc_buf_hdr_t * buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) { const dva_t *dva = BP_IDENTITY(bp); uint64_t birth = BP_PHYSICAL_BIRTH(bp); uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); kmutex_t *hash_lock = BUF_HASH_LOCK(idx); arc_buf_hdr_t *hdr; mutex_enter(hash_lock); for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; hdr = hdr->b_hash_next) { if (HDR_EQUAL(spa, dva, birth, hdr)) { *lockp = hash_lock; return (hdr); } } mutex_exit(hash_lock); *lockp = NULL; return (NULL); } /* * Insert an entry into the hash table. If there is already an element * equal to elem in the hash table, then the already existing element * will be returned and the new element will not be inserted. * Otherwise returns NULL. * If lockp == NULL, the caller is assumed to already hold the hash lock. */ static arc_buf_hdr_t * buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) { uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); kmutex_t *hash_lock = BUF_HASH_LOCK(idx); arc_buf_hdr_t *fhdr; uint32_t i; ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); ASSERT(hdr->b_birth != 0); ASSERT(!HDR_IN_HASH_TABLE(hdr)); if (lockp != NULL) { *lockp = hash_lock; mutex_enter(hash_lock); } else { ASSERT(MUTEX_HELD(hash_lock)); } for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; fhdr = fhdr->b_hash_next, i++) { if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) return (fhdr); } hdr->b_hash_next = buf_hash_table.ht_table[idx]; buf_hash_table.ht_table[idx] = hdr; arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); /* collect some hash table performance data */ if (i > 0) { ARCSTAT_BUMP(arcstat_hash_collisions); if (i == 1) ARCSTAT_BUMP(arcstat_hash_chains); ARCSTAT_MAX(arcstat_hash_chain_max, i); } uint64_t he = atomic_inc_64_nv( &arc_stats.arcstat_hash_elements.value.ui64); ARCSTAT_MAX(arcstat_hash_elements_max, he); return (NULL); } static void buf_hash_remove(arc_buf_hdr_t *hdr) { arc_buf_hdr_t *fhdr, **hdrp; uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); ASSERT(HDR_IN_HASH_TABLE(hdr)); hdrp = &buf_hash_table.ht_table[idx]; while ((fhdr = *hdrp) != hdr) { ASSERT3P(fhdr, !=, NULL); hdrp = &fhdr->b_hash_next; } *hdrp = hdr->b_hash_next; hdr->b_hash_next = NULL; arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); /* collect some hash table performance data */ atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64); if (buf_hash_table.ht_table[idx] && buf_hash_table.ht_table[idx]->b_hash_next == NULL) ARCSTAT_BUMPDOWN(arcstat_hash_chains); } /* * Global data structures and functions for the buf kmem cache. */ static kmem_cache_t *hdr_full_cache; static kmem_cache_t *hdr_full_crypt_cache; static kmem_cache_t *hdr_l2only_cache; static kmem_cache_t *buf_cache; static void buf_fini(void) { int i; #if defined(_KERNEL) /* * Large allocations which do not require contiguous pages * should be using vmem_free() in the linux kernel\ */ vmem_free(buf_hash_table.ht_table, (buf_hash_table.ht_mask + 1) * sizeof (void *)); #else kmem_free(buf_hash_table.ht_table, (buf_hash_table.ht_mask + 1) * sizeof (void *)); #endif for (i = 0; i < BUF_LOCKS; i++) mutex_destroy(BUF_HASH_LOCK(i)); kmem_cache_destroy(hdr_full_cache); kmem_cache_destroy(hdr_full_crypt_cache); kmem_cache_destroy(hdr_l2only_cache); kmem_cache_destroy(buf_cache); } /* * Constructor callback - called when the cache is empty * and a new buf is requested. */ /* ARGSUSED */ static int hdr_full_cons(void *vbuf, void *unused, int kmflag) { arc_buf_hdr_t *hdr = vbuf; bzero(hdr, HDR_FULL_SIZE); hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); zfs_refcount_create(&hdr->b_l1hdr.b_refcnt); mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); list_link_init(&hdr->b_l1hdr.b_arc_node); list_link_init(&hdr->b_l2hdr.b_l2node); multilist_link_init(&hdr->b_l1hdr.b_arc_node); arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); return (0); } /* ARGSUSED */ static int hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag) { arc_buf_hdr_t *hdr = vbuf; hdr_full_cons(vbuf, unused, kmflag); bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr)); arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS); return (0); } /* ARGSUSED */ static int hdr_l2only_cons(void *vbuf, void *unused, int kmflag) { arc_buf_hdr_t *hdr = vbuf; bzero(hdr, HDR_L2ONLY_SIZE); arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); return (0); } /* ARGSUSED */ static int buf_cons(void *vbuf, void *unused, int kmflag) { arc_buf_t *buf = vbuf; bzero(buf, sizeof (arc_buf_t)); mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); return (0); } /* * Destructor callback - called when a cached buf is * no longer required. */ /* ARGSUSED */ static void hdr_full_dest(void *vbuf, void *unused) { arc_buf_hdr_t *hdr = vbuf; ASSERT(HDR_EMPTY(hdr)); cv_destroy(&hdr->b_l1hdr.b_cv); zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt); mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); } /* ARGSUSED */ static void hdr_full_crypt_dest(void *vbuf, void *unused) { arc_buf_hdr_t *hdr = vbuf; hdr_full_dest(vbuf, unused); arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS); } /* ARGSUSED */ static void hdr_l2only_dest(void *vbuf, void *unused) { arc_buf_hdr_t *hdr __maybe_unused = vbuf; ASSERT(HDR_EMPTY(hdr)); arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); } /* ARGSUSED */ static void buf_dest(void *vbuf, void *unused) { arc_buf_t *buf = vbuf; mutex_destroy(&buf->b_evict_lock); arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); } static void buf_init(void) { uint64_t *ct = NULL; uint64_t hsize = 1ULL << 12; int i, j; /* * The hash table is big enough to fill all of physical memory * with an average block size of zfs_arc_average_blocksize (default 8K). * By default, the table will take up * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). */ while (hsize * zfs_arc_average_blocksize < arc_all_memory()) hsize <<= 1; retry: buf_hash_table.ht_mask = hsize - 1; #if defined(_KERNEL) /* * Large allocations which do not require contiguous pages * should be using vmem_alloc() in the linux kernel */ buf_hash_table.ht_table = vmem_zalloc(hsize * sizeof (void*), KM_SLEEP); #else buf_hash_table.ht_table = kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); #endif if (buf_hash_table.ht_table == NULL) { ASSERT(hsize > (1ULL << 8)); hsize >>= 1; goto retry; } hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0); hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt", HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest, NULL, NULL, NULL, 0); hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL, NULL, NULL, 0); buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); for (i = 0; i < 256; i++) for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); for (i = 0; i < BUF_LOCKS; i++) mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL); } #define ARC_MINTIME (hz>>4) /* 62 ms */ /* * This is the size that the buf occupies in memory. If the buf is compressed, * it will correspond to the compressed size. You should use this method of * getting the buf size unless you explicitly need the logical size. */ uint64_t arc_buf_size(arc_buf_t *buf) { return (ARC_BUF_COMPRESSED(buf) ? HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); } uint64_t arc_buf_lsize(arc_buf_t *buf) { return (HDR_GET_LSIZE(buf->b_hdr)); } /* * This function will return B_TRUE if the buffer is encrypted in memory. * This buffer can be decrypted by calling arc_untransform(). */ boolean_t arc_is_encrypted(arc_buf_t *buf) { return (ARC_BUF_ENCRYPTED(buf) != 0); } /* * Returns B_TRUE if the buffer represents data that has not had its MAC * verified yet. */ boolean_t arc_is_unauthenticated(arc_buf_t *buf) { return (HDR_NOAUTH(buf->b_hdr) != 0); } void arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt, uint8_t *iv, uint8_t *mac) { arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT(HDR_PROTECTED(hdr)); bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN); bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN); bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN); *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ? ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER; } /* * Indicates how this buffer is compressed in memory. If it is not compressed * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with * arc_untransform() as long as it is also unencrypted. */ enum zio_compress arc_get_compression(arc_buf_t *buf) { return (ARC_BUF_COMPRESSED(buf) ? HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); } /* * Return the compression algorithm used to store this data in the ARC. If ARC * compression is enabled or this is an encrypted block, this will be the same * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF. */ static inline enum zio_compress arc_hdr_get_compress(arc_buf_hdr_t *hdr) { return (HDR_COMPRESSION_ENABLED(hdr) ? HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF); } uint8_t arc_get_complevel(arc_buf_t *buf) { return (buf->b_hdr->b_complevel); } static inline boolean_t arc_buf_is_shared(arc_buf_t *buf) { boolean_t shared = (buf->b_data != NULL && buf->b_hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); IMPLY(shared, ARC_BUF_SHARED(buf)); IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); /* * It would be nice to assert arc_can_share() too, but the "hdr isn't * already being shared" requirement prevents us from doing that. */ return (shared); } /* * Free the checksum associated with this header. If there is no checksum, this * is a no-op. */ static inline void arc_cksum_free(arc_buf_hdr_t *hdr) { ASSERT(HDR_HAS_L1HDR(hdr)); mutex_enter(&hdr->b_l1hdr.b_freeze_lock); if (hdr->b_l1hdr.b_freeze_cksum != NULL) { kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); hdr->b_l1hdr.b_freeze_cksum = NULL; } mutex_exit(&hdr->b_l1hdr.b_freeze_lock); } /* * Return true iff at least one of the bufs on hdr is not compressed. * Encrypted buffers count as compressed. */ static boolean_t arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) { ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr)); for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { if (!ARC_BUF_COMPRESSED(b)) { return (B_TRUE); } } return (B_FALSE); } /* * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data * matches the checksum that is stored in the hdr. If there is no checksum, * or if the buf is compressed, this is a no-op. */ static void arc_cksum_verify(arc_buf_t *buf) { arc_buf_hdr_t *hdr = buf->b_hdr; zio_cksum_t zc; if (!(zfs_flags & ZFS_DEBUG_MODIFY)) return; if (ARC_BUF_COMPRESSED(buf)) return; ASSERT(HDR_HAS_L1HDR(hdr)); mutex_enter(&hdr->b_l1hdr.b_freeze_lock); if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { mutex_exit(&hdr->b_l1hdr.b_freeze_lock); return; } fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) panic("buffer modified while frozen!"); mutex_exit(&hdr->b_l1hdr.b_freeze_lock); } /* * This function makes the assumption that data stored in the L2ARC * will be transformed exactly as it is in the main pool. Because of * this we can verify the checksum against the reading process's bp. */ static boolean_t arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) { ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); /* * Block pointers always store the checksum for the logical data. * If the block pointer has the gang bit set, then the checksum * it represents is for the reconstituted data and not for an * individual gang member. The zio pipeline, however, must be able to * determine the checksum of each of the gang constituents so it * treats the checksum comparison differently than what we need * for l2arc blocks. This prevents us from using the * zio_checksum_error() interface directly. Instead we must call the * zio_checksum_error_impl() so that we can ensure the checksum is * generated using the correct checksum algorithm and accounts for the * logical I/O size and not just a gang fragment. */ return (zio_checksum_error_impl(zio->io_spa, zio->io_bp, BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, zio->io_offset, NULL) == 0); } /* * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a * checksum and attaches it to the buf's hdr so that we can ensure that the buf * isn't modified later on. If buf is compressed or there is already a checksum * on the hdr, this is a no-op (we only checksum uncompressed bufs). */ static void arc_cksum_compute(arc_buf_t *buf) { arc_buf_hdr_t *hdr = buf->b_hdr; if (!(zfs_flags & ZFS_DEBUG_MODIFY)) return; ASSERT(HDR_HAS_L1HDR(hdr)); mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) { mutex_exit(&hdr->b_l1hdr.b_freeze_lock); return; } ASSERT(!ARC_BUF_ENCRYPTED(buf)); ASSERT(!ARC_BUF_COMPRESSED(buf)); hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, hdr->b_l1hdr.b_freeze_cksum); mutex_exit(&hdr->b_l1hdr.b_freeze_lock); arc_buf_watch(buf); } #ifndef _KERNEL void arc_buf_sigsegv(int sig, siginfo_t *si, void *unused) { panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr); } #endif /* ARGSUSED */ static void arc_buf_unwatch(arc_buf_t *buf) { #ifndef _KERNEL if (arc_watch) { ASSERT0(mprotect(buf->b_data, arc_buf_size(buf), PROT_READ | PROT_WRITE)); } #endif } /* ARGSUSED */ static void arc_buf_watch(arc_buf_t *buf) { #ifndef _KERNEL if (arc_watch) ASSERT0(mprotect(buf->b_data, arc_buf_size(buf), PROT_READ)); #endif } static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *hdr) { arc_buf_contents_t type; if (HDR_ISTYPE_METADATA(hdr)) { type = ARC_BUFC_METADATA; } else { type = ARC_BUFC_DATA; } VERIFY3U(hdr->b_type, ==, type); return (type); } boolean_t arc_is_metadata(arc_buf_t *buf) { return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); } static uint32_t arc_bufc_to_flags(arc_buf_contents_t type) { switch (type) { case ARC_BUFC_DATA: /* metadata field is 0 if buffer contains normal data */ return (0); case ARC_BUFC_METADATA: return (ARC_FLAG_BUFC_METADATA); default: break; } panic("undefined ARC buffer type!"); return ((uint32_t)-1); } void arc_buf_thaw(arc_buf_t *buf) { arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); ASSERT(!HDR_IO_IN_PROGRESS(hdr)); arc_cksum_verify(buf); /* * Compressed buffers do not manipulate the b_freeze_cksum. */ if (ARC_BUF_COMPRESSED(buf)) return; ASSERT(HDR_HAS_L1HDR(hdr)); arc_cksum_free(hdr); arc_buf_unwatch(buf); } void arc_buf_freeze(arc_buf_t *buf) { if (!(zfs_flags & ZFS_DEBUG_MODIFY)) return; if (ARC_BUF_COMPRESSED(buf)) return; ASSERT(HDR_HAS_L1HDR(buf->b_hdr)); arc_cksum_compute(buf); } /* * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, * the following functions should be used to ensure that the flags are * updated in a thread-safe way. When manipulating the flags either * the hash_lock must be held or the hdr must be undiscoverable. This * ensures that we're not racing with any other threads when updating * the flags. */ static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) { ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); hdr->b_flags |= flags; } static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) { ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); hdr->b_flags &= ~flags; } /* * Setting the compression bits in the arc_buf_hdr_t's b_flags is * done in a special way since we have to clear and set bits * at the same time. Consumers that wish to set the compression bits * must use this function to ensure that the flags are updated in * thread-safe manner. */ static void arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) { ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); /* * Holes and embedded blocks will always have a psize = 0 so * we ignore the compression of the blkptr and set the * want to uncompress them. Mark them as uncompressed. */ if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); } else { arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); ASSERT(HDR_COMPRESSION_ENABLED(hdr)); } HDR_SET_COMPRESS(hdr, cmp); ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); } /* * Looks for another buf on the same hdr which has the data decompressed, copies * from it, and returns true. If no such buf exists, returns false. */ static boolean_t arc_buf_try_copy_decompressed_data(arc_buf_t *buf) { arc_buf_hdr_t *hdr = buf->b_hdr; boolean_t copied = B_FALSE; ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT3P(buf->b_data, !=, NULL); ASSERT(!ARC_BUF_COMPRESSED(buf)); for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; from = from->b_next) { /* can't use our own data buffer */ if (from == buf) { continue; } if (!ARC_BUF_COMPRESSED(from)) { bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); copied = B_TRUE; break; } } /* * There were no decompressed bufs, so there should not be a * checksum on the hdr either. */ if (zfs_flags & ZFS_DEBUG_MODIFY) EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); return (copied); } /* * Allocates an ARC buf header that's in an evicted & L2-cached state. * This is used during l2arc reconstruction to make empty ARC buffers * which circumvent the regular disk->arc->l2arc path and instead come * into being in the reverse order, i.e. l2arc->arc. */ static arc_buf_hdr_t * arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev, dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth, enum zio_compress compress, uint8_t complevel, boolean_t protected, boolean_t prefetch, arc_state_type_t arcs_state) { arc_buf_hdr_t *hdr; ASSERT(size != 0); hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP); hdr->b_birth = birth; hdr->b_type = type; hdr->b_flags = 0; arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR); HDR_SET_LSIZE(hdr, size); HDR_SET_PSIZE(hdr, psize); arc_hdr_set_compress(hdr, compress); hdr->b_complevel = complevel; if (protected) arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED); if (prefetch) arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa); hdr->b_dva = dva; hdr->b_l2hdr.b_dev = dev; hdr->b_l2hdr.b_daddr = daddr; hdr->b_l2hdr.b_arcs_state = arcs_state; return (hdr); } /* * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. */ static uint64_t arc_hdr_size(arc_buf_hdr_t *hdr) { uint64_t size; if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF && HDR_GET_PSIZE(hdr) > 0) { size = HDR_GET_PSIZE(hdr); } else { ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); size = HDR_GET_LSIZE(hdr); } return (size); } static int arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj) { int ret; uint64_t csize; uint64_t lsize = HDR_GET_LSIZE(hdr); uint64_t psize = HDR_GET_PSIZE(hdr); void *tmpbuf = NULL; abd_t *abd = hdr->b_l1hdr.b_pabd; ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); ASSERT(HDR_AUTHENTICATED(hdr)); ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); /* * The MAC is calculated on the compressed data that is stored on disk. * However, if compressed arc is disabled we will only have the * decompressed data available to us now. Compress it into a temporary * abd so we can verify the MAC. The performance overhead of this will * be relatively low, since most objects in an encrypted objset will * be encrypted (instead of authenticated) anyway. */ if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) { tmpbuf = zio_buf_alloc(lsize); abd = abd_get_from_buf(tmpbuf, lsize); abd_take_ownership_of_buf(abd, B_TRUE); csize = zio_compress_data(HDR_GET_COMPRESS(hdr), hdr->b_l1hdr.b_pabd, tmpbuf, lsize, hdr->b_complevel); ASSERT3U(csize, <=, psize); abd_zero_off(abd, csize, psize - csize); } /* * Authentication is best effort. We authenticate whenever the key is * available. If we succeed we clear ARC_FLAG_NOAUTH. */ if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) { ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); ASSERT3U(lsize, ==, psize); ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd, psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); } else { ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize, hdr->b_crypt_hdr.b_mac); } if (ret == 0) arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH); else if (ret != ENOENT) goto error; if (tmpbuf != NULL) abd_free(abd); return (0); error: if (tmpbuf != NULL) abd_free(abd); return (ret); } /* * This function will take a header that only has raw encrypted data in * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in * b_l1hdr.b_pabd. If designated in the header flags, this function will * also decompress the data. */ static int arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb) { int ret; abd_t *cabd = NULL; void *tmp = NULL; boolean_t no_crypt = B_FALSE; boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); ASSERT(HDR_ENCRYPTED(hdr)); arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT); ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot, B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv, hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd, &no_crypt); if (ret != 0) goto error; if (no_crypt) { abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd, HDR_GET_PSIZE(hdr)); } /* * If this header has disabled arc compression but the b_pabd is * compressed after decrypting it, we need to decompress the newly * decrypted data. */ if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) { /* * We want to make sure that we are correctly honoring the * zfs_abd_scatter_enabled setting, so we allocate an abd here * and then loan a buffer from it, rather than allocating a * linear buffer and wrapping it in an abd later. */ cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, B_TRUE); tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr)); ret = zio_decompress_data(HDR_GET_COMPRESS(hdr), hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr), &hdr->b_complevel); if (ret != 0) { abd_return_buf(cabd, tmp, arc_hdr_size(hdr)); goto error; } abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, arc_hdr_size(hdr), hdr); hdr->b_l1hdr.b_pabd = cabd; } return (0); error: arc_hdr_free_abd(hdr, B_FALSE); if (cabd != NULL) arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr); return (ret); } /* * This function is called during arc_buf_fill() to prepare the header's * abd plaintext pointer for use. This involves authenticated protected * data and decrypting encrypted data into the plaintext abd. */ static int arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa, const zbookmark_phys_t *zb, boolean_t noauth) { int ret; ASSERT(HDR_PROTECTED(hdr)); if (hash_lock != NULL) mutex_enter(hash_lock); if (HDR_NOAUTH(hdr) && !noauth) { /* * The caller requested authenticated data but our data has * not been authenticated yet. Verify the MAC now if we can. */ ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset); if (ret != 0) goto error; } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) { /* * If we only have the encrypted version of the data, but the * unencrypted version was requested we take this opportunity * to store the decrypted version in the header for future use. */ ret = arc_hdr_decrypt(hdr, spa, zb); if (ret != 0) goto error; } ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); if (hash_lock != NULL) mutex_exit(hash_lock); return (0); error: if (hash_lock != NULL) mutex_exit(hash_lock); return (ret); } /* * This function is used by the dbuf code to decrypt bonus buffers in place. * The dbuf code itself doesn't have any locking for decrypting a shared dnode * block, so we use the hash lock here to protect against concurrent calls to * arc_buf_fill(). */ static void arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock) { arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT(HDR_ENCRYPTED(hdr)); ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE); ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data, arc_buf_size(buf)); buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; hdr->b_crypt_hdr.b_ebufcnt -= 1; } /* * Given a buf that has a data buffer attached to it, this function will * efficiently fill the buf with data of the specified compression setting from * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr * are already sharing a data buf, no copy is performed. * * If the buf is marked as compressed but uncompressed data was requested, this * will allocate a new data buffer for the buf, remove that flag, and fill the * buf with uncompressed data. You can't request a compressed buf on a hdr with * uncompressed data, and (since we haven't added support for it yet) if you * want compressed data your buf must already be marked as compressed and have * the correct-sized data buffer. */ static int arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb, arc_fill_flags_t flags) { int error = 0; arc_buf_hdr_t *hdr = buf->b_hdr; boolean_t hdr_compressed = (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0; boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0; dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr); ASSERT3P(buf->b_data, !=, NULL); IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf)); IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); IMPLY(encrypted, HDR_ENCRYPTED(hdr)); IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf)); IMPLY(encrypted, ARC_BUF_COMPRESSED(buf)); IMPLY(encrypted, !ARC_BUF_SHARED(buf)); /* * If the caller wanted encrypted data we just need to copy it from * b_rabd and potentially byteswap it. We won't be able to do any * further transforms on it. */ if (encrypted) { ASSERT(HDR_HAS_RABD(hdr)); abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd, HDR_GET_PSIZE(hdr)); goto byteswap; } /* * Adjust encrypted and authenticated headers to accommodate * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are * allowed to fail decryption due to keys not being loaded * without being marked as an IO error. */ if (HDR_PROTECTED(hdr)) { error = arc_fill_hdr_crypt(hdr, hash_lock, spa, zb, !!(flags & ARC_FILL_NOAUTH)); if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) { return (error); } else if (error != 0) { if (hash_lock != NULL) mutex_enter(hash_lock); arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); if (hash_lock != NULL) mutex_exit(hash_lock); return (error); } } /* * There is a special case here for dnode blocks which are * decrypting their bonus buffers. These blocks may request to * be decrypted in-place. This is necessary because there may * be many dnodes pointing into this buffer and there is * currently no method to synchronize replacing the backing * b_data buffer and updating all of the pointers. Here we use * the hash lock to ensure there are no races. If the need * arises for other types to be decrypted in-place, they must * add handling here as well. */ if ((flags & ARC_FILL_IN_PLACE) != 0) { ASSERT(!hdr_compressed); ASSERT(!compressed); ASSERT(!encrypted); if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) { ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE); if (hash_lock != NULL) mutex_enter(hash_lock); arc_buf_untransform_in_place(buf, hash_lock); if (hash_lock != NULL) mutex_exit(hash_lock); /* Compute the hdr's checksum if necessary */ arc_cksum_compute(buf); } return (0); } if (hdr_compressed == compressed) { if (!arc_buf_is_shared(buf)) { abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, arc_buf_size(buf)); } } else { ASSERT(hdr_compressed); ASSERT(!compressed); ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); /* * If the buf is sharing its data with the hdr, unlink it and * allocate a new data buffer for the buf. */ if (arc_buf_is_shared(buf)) { ASSERT(ARC_BUF_COMPRESSED(buf)); /* We need to give the buf its own b_data */ buf->b_flags &= ~ARC_BUF_FLAG_SHARED; buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); /* Previously overhead was 0; just add new overhead */ ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); } else if (ARC_BUF_COMPRESSED(buf)) { /* We need to reallocate the buf's b_data */ arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), buf); buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); /* We increased the size of b_data; update overhead */ ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); } /* * Regardless of the buf's previous compression settings, it * should not be compressed at the end of this function. */ buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; /* * Try copying the data from another buf which already has a * decompressed version. If that's not possible, it's time to * bite the bullet and decompress the data from the hdr. */ if (arc_buf_try_copy_decompressed_data(buf)) { /* Skip byteswapping and checksumming (already done) */ return (0); } else { error = zio_decompress_data(HDR_GET_COMPRESS(hdr), hdr->b_l1hdr.b_pabd, buf->b_data, HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr), &hdr->b_complevel); /* * Absent hardware errors or software bugs, this should * be impossible, but log it anyway so we can debug it. */ if (error != 0) { zfs_dbgmsg( "hdr %px, compress %d, psize %d, lsize %d", hdr, arc_hdr_get_compress(hdr), HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); if (hash_lock != NULL) mutex_enter(hash_lock); arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); if (hash_lock != NULL) mutex_exit(hash_lock); return (SET_ERROR(EIO)); } } } byteswap: /* Byteswap the buf's data if necessary */ if (bswap != DMU_BSWAP_NUMFUNCS) { ASSERT(!HDR_SHARED_DATA(hdr)); ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); } /* Compute the hdr's checksum if necessary */ arc_cksum_compute(buf); return (0); } /* * If this function is being called to decrypt an encrypted buffer or verify an * authenticated one, the key must be loaded and a mapping must be made * available in the keystore via spa_keystore_create_mapping() or one of its * callers. */ int arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb, boolean_t in_place) { int ret; arc_fill_flags_t flags = 0; if (in_place) flags |= ARC_FILL_IN_PLACE; ret = arc_buf_fill(buf, spa, zb, flags); if (ret == ECKSUM) { /* * Convert authentication and decryption errors to EIO * (and generate an ereport) before leaving the ARC. */ ret = SET_ERROR(EIO); spa_log_error(spa, zb); (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, spa, NULL, zb, NULL, 0); } return (ret); } /* * Increment the amount of evictable space in the arc_state_t's refcount. * We account for the space used by the hdr and the arc buf individually * so that we can add and remove them from the refcount individually. */ static void arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) { arc_buf_contents_t type = arc_buf_type(hdr); ASSERT(HDR_HAS_L1HDR(hdr)); if (GHOST_STATE(state)) { ASSERT0(hdr->b_l1hdr.b_bufcnt); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); (void) zfs_refcount_add_many(&state->arcs_esize[type], HDR_GET_LSIZE(hdr), hdr); return; } if (hdr->b_l1hdr.b_pabd != NULL) { (void) zfs_refcount_add_many(&state->arcs_esize[type], arc_hdr_size(hdr), hdr); } if (HDR_HAS_RABD(hdr)) { (void) zfs_refcount_add_many(&state->arcs_esize[type], HDR_GET_PSIZE(hdr), hdr); } for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { if (arc_buf_is_shared(buf)) continue; (void) zfs_refcount_add_many(&state->arcs_esize[type], arc_buf_size(buf), buf); } } /* * Decrement the amount of evictable space in the arc_state_t's refcount. * We account for the space used by the hdr and the arc buf individually * so that we can add and remove them from the refcount individually. */ static void arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) { arc_buf_contents_t type = arc_buf_type(hdr); ASSERT(HDR_HAS_L1HDR(hdr)); if (GHOST_STATE(state)) { ASSERT0(hdr->b_l1hdr.b_bufcnt); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); (void) zfs_refcount_remove_many(&state->arcs_esize[type], HDR_GET_LSIZE(hdr), hdr); return; } if (hdr->b_l1hdr.b_pabd != NULL) { (void) zfs_refcount_remove_many(&state->arcs_esize[type], arc_hdr_size(hdr), hdr); } if (HDR_HAS_RABD(hdr)) { (void) zfs_refcount_remove_many(&state->arcs_esize[type], HDR_GET_PSIZE(hdr), hdr); } for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { if (arc_buf_is_shared(buf)) continue; (void) zfs_refcount_remove_many(&state->arcs_esize[type], arc_buf_size(buf), buf); } } /* * Add a reference to this hdr indicating that someone is actively * referencing that memory. When the refcount transitions from 0 to 1, * we remove it from the respective arc_state_t list to indicate that * it is not evictable. */ static void add_reference(arc_buf_hdr_t *hdr, void *tag) { arc_state_t *state; ASSERT(HDR_HAS_L1HDR(hdr)); if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) { ASSERT(hdr->b_l1hdr.b_state == arc_anon); ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); } state = hdr->b_l1hdr.b_state; if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && (state != arc_anon)) { /* We don't use the L2-only state list. */ if (state != arc_l2c_only) { multilist_remove(&state->arcs_list[arc_buf_type(hdr)], hdr); arc_evictable_space_decrement(hdr, state); } /* remove the prefetch flag if we get a reference */ if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_decrement_state(hdr); arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_increment_state(hdr); } } /* * Remove a reference from this hdr. When the reference transitions from * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's * list making it eligible for eviction. */ static int remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) { int cnt; arc_state_t *state = hdr->b_l1hdr.b_state; ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); ASSERT(!GHOST_STATE(state)); /* * arc_l2c_only counts as a ghost state so we don't need to explicitly * check to prevent usage of the arc_l2c_only list. */ if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && (state != arc_anon)) { multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr); ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); arc_evictable_space_increment(hdr, state); } return (cnt); } /* * Returns detailed information about a specific arc buffer. When the * state_index argument is set the function will calculate the arc header * list position for its arc state. Since this requires a linear traversal * callers are strongly encourage not to do this. However, it can be helpful * for targeted analysis so the functionality is provided. */ void arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index) { arc_buf_hdr_t *hdr = ab->b_hdr; l1arc_buf_hdr_t *l1hdr = NULL; l2arc_buf_hdr_t *l2hdr = NULL; arc_state_t *state = NULL; memset(abi, 0, sizeof (arc_buf_info_t)); if (hdr == NULL) return; abi->abi_flags = hdr->b_flags; if (HDR_HAS_L1HDR(hdr)) { l1hdr = &hdr->b_l1hdr; state = l1hdr->b_state; } if (HDR_HAS_L2HDR(hdr)) l2hdr = &hdr->b_l2hdr; if (l1hdr) { abi->abi_bufcnt = l1hdr->b_bufcnt; abi->abi_access = l1hdr->b_arc_access; abi->abi_mru_hits = l1hdr->b_mru_hits; abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits; abi->abi_mfu_hits = l1hdr->b_mfu_hits; abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits; abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt); } if (l2hdr) { abi->abi_l2arc_dattr = l2hdr->b_daddr; abi->abi_l2arc_hits = l2hdr->b_hits; } abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON; abi->abi_state_contents = arc_buf_type(hdr); abi->abi_size = arc_hdr_size(hdr); } /* * Move the supplied buffer to the indicated state. The hash lock * for the buffer must be held by the caller. */ static void arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, kmutex_t *hash_lock) { arc_state_t *old_state; int64_t refcnt; uint32_t bufcnt; boolean_t update_old, update_new; arc_buf_contents_t buftype = arc_buf_type(hdr); /* * We almost always have an L1 hdr here, since we call arc_hdr_realloc() * in arc_read() when bringing a buffer out of the L2ARC. However, the * L1 hdr doesn't always exist when we change state to arc_anon before * destroying a header, in which case reallocating to add the L1 hdr is * pointless. */ if (HDR_HAS_L1HDR(hdr)) { old_state = hdr->b_l1hdr.b_state; refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt); bufcnt = hdr->b_l1hdr.b_bufcnt; update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); } else { old_state = arc_l2c_only; refcnt = 0; bufcnt = 0; update_old = B_FALSE; } update_new = update_old; ASSERT(MUTEX_HELD(hash_lock)); ASSERT3P(new_state, !=, old_state); ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); ASSERT(old_state != arc_anon || bufcnt <= 1); /* * If this buffer is evictable, transfer it from the * old state list to the new state list. */ if (refcnt == 0) { if (old_state != arc_anon && old_state != arc_l2c_only) { ASSERT(HDR_HAS_L1HDR(hdr)); multilist_remove(&old_state->arcs_list[buftype], hdr); if (GHOST_STATE(old_state)) { ASSERT0(bufcnt); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); update_old = B_TRUE; } arc_evictable_space_decrement(hdr, old_state); } if (new_state != arc_anon && new_state != arc_l2c_only) { /* * An L1 header always exists here, since if we're * moving to some L1-cached state (i.e. not l2c_only or * anonymous), we realloc the header to add an L1hdr * beforehand. */ ASSERT(HDR_HAS_L1HDR(hdr)); multilist_insert(&new_state->arcs_list[buftype], hdr); if (GHOST_STATE(new_state)) { ASSERT0(bufcnt); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); update_new = B_TRUE; } arc_evictable_space_increment(hdr, new_state); } } ASSERT(!HDR_EMPTY(hdr)); if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) buf_hash_remove(hdr); /* adjust state sizes (ignore arc_l2c_only) */ if (update_new && new_state != arc_l2c_only) { ASSERT(HDR_HAS_L1HDR(hdr)); if (GHOST_STATE(new_state)) { ASSERT0(bufcnt); /* * When moving a header to a ghost state, we first * remove all arc buffers. Thus, we'll have a * bufcnt of zero, and no arc buffer to use for * the reference. As a result, we use the arc * header pointer for the reference. */ (void) zfs_refcount_add_many(&new_state->arcs_size, HDR_GET_LSIZE(hdr), hdr); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); } else { uint32_t buffers = 0; /* * Each individual buffer holds a unique reference, * thus we must remove each of these references one * at a time. */ for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { ASSERT3U(bufcnt, !=, 0); buffers++; /* * When the arc_buf_t is sharing the data * block with the hdr, the owner of the * reference belongs to the hdr. Only * add to the refcount if the arc_buf_t is * not shared. */ if (arc_buf_is_shared(buf)) continue; (void) zfs_refcount_add_many( &new_state->arcs_size, arc_buf_size(buf), buf); } ASSERT3U(bufcnt, ==, buffers); if (hdr->b_l1hdr.b_pabd != NULL) { (void) zfs_refcount_add_many( &new_state->arcs_size, arc_hdr_size(hdr), hdr); } if (HDR_HAS_RABD(hdr)) { (void) zfs_refcount_add_many( &new_state->arcs_size, HDR_GET_PSIZE(hdr), hdr); } } } if (update_old && old_state != arc_l2c_only) { ASSERT(HDR_HAS_L1HDR(hdr)); if (GHOST_STATE(old_state)) { ASSERT0(bufcnt); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); /* * When moving a header off of a ghost state, * the header will not contain any arc buffers. * We use the arc header pointer for the reference * which is exactly what we did when we put the * header on the ghost state. */ (void) zfs_refcount_remove_many(&old_state->arcs_size, HDR_GET_LSIZE(hdr), hdr); } else { uint32_t buffers = 0; /* * Each individual buffer holds a unique reference, * thus we must remove each of these references one * at a time. */ for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { ASSERT3U(bufcnt, !=, 0); buffers++; /* * When the arc_buf_t is sharing the data * block with the hdr, the owner of the * reference belongs to the hdr. Only * add to the refcount if the arc_buf_t is * not shared. */ if (arc_buf_is_shared(buf)) continue; (void) zfs_refcount_remove_many( &old_state->arcs_size, arc_buf_size(buf), buf); } ASSERT3U(bufcnt, ==, buffers); ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); if (hdr->b_l1hdr.b_pabd != NULL) { (void) zfs_refcount_remove_many( &old_state->arcs_size, arc_hdr_size(hdr), hdr); } if (HDR_HAS_RABD(hdr)) { (void) zfs_refcount_remove_many( &old_state->arcs_size, HDR_GET_PSIZE(hdr), hdr); } } } if (HDR_HAS_L1HDR(hdr)) { hdr->b_l1hdr.b_state = new_state; if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) { l2arc_hdr_arcstats_decrement_state(hdr); hdr->b_l2hdr.b_arcs_state = new_state->arcs_state; l2arc_hdr_arcstats_increment_state(hdr); } } /* * L2 headers should never be on the L2 state list since they don't * have L1 headers allocated. */ ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); } void arc_space_consume(uint64_t space, arc_space_type_t type) { ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); switch (type) { default: break; case ARC_SPACE_DATA: ARCSTAT_INCR(arcstat_data_size, space); break; case ARC_SPACE_META: ARCSTAT_INCR(arcstat_metadata_size, space); break; case ARC_SPACE_BONUS: ARCSTAT_INCR(arcstat_bonus_size, space); break; case ARC_SPACE_DNODE: aggsum_add(&arc_sums.arcstat_dnode_size, space); break; case ARC_SPACE_DBUF: ARCSTAT_INCR(arcstat_dbuf_size, space); break; case ARC_SPACE_HDRS: ARCSTAT_INCR(arcstat_hdr_size, space); break; case ARC_SPACE_L2HDRS: aggsum_add(&arc_sums.arcstat_l2_hdr_size, space); break; case ARC_SPACE_ABD_CHUNK_WASTE: /* * Note: this includes space wasted by all scatter ABD's, not * just those allocated by the ARC. But the vast majority of * scatter ABD's come from the ARC, because other users are * very short-lived. */ ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space); break; } if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE) aggsum_add(&arc_sums.arcstat_meta_used, space); aggsum_add(&arc_sums.arcstat_size, space); } void arc_space_return(uint64_t space, arc_space_type_t type) { ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); switch (type) { default: break; case ARC_SPACE_DATA: ARCSTAT_INCR(arcstat_data_size, -space); break; case ARC_SPACE_META: ARCSTAT_INCR(arcstat_metadata_size, -space); break; case ARC_SPACE_BONUS: ARCSTAT_INCR(arcstat_bonus_size, -space); break; case ARC_SPACE_DNODE: aggsum_add(&arc_sums.arcstat_dnode_size, -space); break; case ARC_SPACE_DBUF: ARCSTAT_INCR(arcstat_dbuf_size, -space); break; case ARC_SPACE_HDRS: ARCSTAT_INCR(arcstat_hdr_size, -space); break; case ARC_SPACE_L2HDRS: aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space); break; case ARC_SPACE_ABD_CHUNK_WASTE: ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space); break; } if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE) { ASSERT(aggsum_compare(&arc_sums.arcstat_meta_used, space) >= 0); ARCSTAT_MAX(arcstat_meta_max, aggsum_upper_bound(&arc_sums.arcstat_meta_used)); aggsum_add(&arc_sums.arcstat_meta_used, -space); } ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0); aggsum_add(&arc_sums.arcstat_size, -space); } /* * Given a hdr and a buf, returns whether that buf can share its b_data buffer * with the hdr's b_pabd. */ static boolean_t arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) { /* * The criteria for sharing a hdr's data are: * 1. the buffer is not encrypted * 2. the hdr's compression matches the buf's compression * 3. the hdr doesn't need to be byteswapped * 4. the hdr isn't already being shared * 5. the buf is either compressed or it is the last buf in the hdr list * * Criterion #5 maintains the invariant that shared uncompressed * bufs must be the final buf in the hdr's b_buf list. Reading this, you * might ask, "if a compressed buf is allocated first, won't that be the * last thing in the list?", but in that case it's impossible to create * a shared uncompressed buf anyway (because the hdr must be compressed * to have the compressed buf). You might also think that #3 is * sufficient to make this guarantee, however it's possible * (specifically in the rare L2ARC write race mentioned in * arc_buf_alloc_impl()) there will be an existing uncompressed buf that * is shareable, but wasn't at the time of its allocation. Rather than * allow a new shared uncompressed buf to be created and then shuffle * the list around to make it the last element, this simply disallows * sharing if the new buf isn't the first to be added. */ ASSERT3P(buf->b_hdr, ==, hdr); boolean_t hdr_compressed = arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF; boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; return (!ARC_BUF_ENCRYPTED(buf) && buf_compressed == hdr_compressed && hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && !HDR_SHARED_DATA(hdr) && (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); } /* * Allocate a buf for this hdr. If you care about the data that's in the hdr, * or if you want a compressed buffer, pass those flags in. Returns 0 if the * copy was made successfully, or an error code otherwise. */ static int arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb, void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth, boolean_t fill, arc_buf_t **ret) { arc_buf_t *buf; arc_fill_flags_t flags = ARC_FILL_LOCKED; ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); VERIFY(hdr->b_type == ARC_BUFC_DATA || hdr->b_type == ARC_BUFC_METADATA); ASSERT3P(ret, !=, NULL); ASSERT3P(*ret, ==, NULL); IMPLY(encrypted, compressed); hdr->b_l1hdr.b_mru_hits = 0; hdr->b_l1hdr.b_mru_ghost_hits = 0; hdr->b_l1hdr.b_mfu_hits = 0; hdr->b_l1hdr.b_mfu_ghost_hits = 0; hdr->b_l1hdr.b_l2_hits = 0; buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); buf->b_hdr = hdr; buf->b_data = NULL; buf->b_next = hdr->b_l1hdr.b_buf; buf->b_flags = 0; add_reference(hdr, tag); /* * We're about to change the hdr's b_flags. We must either * hold the hash_lock or be undiscoverable. */ ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); /* * Only honor requests for compressed bufs if the hdr is actually * compressed. This must be overridden if the buffer is encrypted since * encrypted buffers cannot be decompressed. */ if (encrypted) { buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED; flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED; } else if (compressed && arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) { buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; flags |= ARC_FILL_COMPRESSED; } if (noauth) { ASSERT0(encrypted); flags |= ARC_FILL_NOAUTH; } /* * If the hdr's data can be shared then we share the data buffer and * set the appropriate bit in the hdr's b_flags to indicate the hdr is * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new * buffer to store the buf's data. * * There are two additional restrictions here because we're sharing * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be * actively involved in an L2ARC write, because if this buf is used by * an arc_write() then the hdr's data buffer will be released when the * write completes, even though the L2ARC write might still be using it. * Second, the hdr's ABD must be linear so that the buf's user doesn't * need to be ABD-aware. It must be allocated via * zio_[data_]buf_alloc(), not as a page, because we need to be able * to abd_release_ownership_of_buf(), which isn't allowed on "linear * page" buffers because the ABD code needs to handle freeing them * specially. */ boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd) && !abd_is_linear_page(hdr->b_l1hdr.b_pabd); /* Set up b_data and sharing */ if (can_share) { buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); buf->b_flags |= ARC_BUF_FLAG_SHARED; arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); } else { buf->b_data = arc_get_data_buf(hdr, arc_buf_size(buf), buf); ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); } VERIFY3P(buf->b_data, !=, NULL); hdr->b_l1hdr.b_buf = buf; hdr->b_l1hdr.b_bufcnt += 1; if (encrypted) hdr->b_crypt_hdr.b_ebufcnt += 1; /* * If the user wants the data from the hdr, we need to either copy or * decompress the data. */ if (fill) { ASSERT3P(zb, !=, NULL); return (arc_buf_fill(buf, spa, zb, flags)); } return (0); } static char *arc_onloan_tag = "onloan"; static inline void arc_loaned_bytes_update(int64_t delta) { atomic_add_64(&arc_loaned_bytes, delta); /* assert that it did not wrap around */ ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); } /* * Loan out an anonymous arc buffer. Loaned buffers are not counted as in * flight data by arc_tempreserve_space() until they are "returned". Loaned * buffers must be returned to the arc before they can be used by the DMU or * freed. */ arc_buf_t * arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) { arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); arc_loaned_bytes_update(arc_buf_size(buf)); return (buf); } arc_buf_t * arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, enum zio_compress compression_type, uint8_t complevel) { arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, psize, lsize, compression_type, complevel); arc_loaned_bytes_update(arc_buf_size(buf)); return (buf); } arc_buf_t * arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder, const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_object_type_t ot, uint64_t psize, uint64_t lsize, enum zio_compress compression_type, uint8_t complevel) { arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj, byteorder, salt, iv, mac, ot, psize, lsize, compression_type, complevel); atomic_add_64(&arc_loaned_bytes, psize); return (buf); } /* * Return a loaned arc buffer to the arc. */ void arc_return_buf(arc_buf_t *buf, void *tag) { arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT3P(buf->b_data, !=, NULL); ASSERT(HDR_HAS_L1HDR(hdr)); (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag); (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); arc_loaned_bytes_update(-arc_buf_size(buf)); } /* Detach an arc_buf from a dbuf (tag) */ void arc_loan_inuse_buf(arc_buf_t *buf, void *tag) { arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT3P(buf->b_data, !=, NULL); ASSERT(HDR_HAS_L1HDR(hdr)); (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); arc_loaned_bytes_update(arc_buf_size(buf)); } static void l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) { l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); df->l2df_abd = abd; df->l2df_size = size; df->l2df_type = type; mutex_enter(&l2arc_free_on_write_mtx); list_insert_head(l2arc_free_on_write, df); mutex_exit(&l2arc_free_on_write_mtx); } static void arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata) { arc_state_t *state = hdr->b_l1hdr.b_state; arc_buf_contents_t type = arc_buf_type(hdr); uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr); /* protected by hash lock, if in the hash table */ if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); ASSERT(state != arc_anon && state != arc_l2c_only); (void) zfs_refcount_remove_many(&state->arcs_esize[type], size, hdr); } (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr); if (type == ARC_BUFC_METADATA) { arc_space_return(size, ARC_SPACE_META); } else { ASSERT(type == ARC_BUFC_DATA); arc_space_return(size, ARC_SPACE_DATA); } if (free_rdata) { l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type); } else { l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); } } /* * Share the arc_buf_t's data with the hdr. Whenever we are sharing the * data buffer, we transfer the refcount ownership to the hdr and update * the appropriate kstats. */ static void arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) { ASSERT(arc_can_share(hdr, buf)); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!ARC_BUF_ENCRYPTED(buf)); ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); /* * Start sharing the data buffer. We transfer the * refcount ownership to the hdr since it always owns * the refcount whenever an arc_buf_t is shared. */ zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size, arc_hdr_size(hdr), buf, hdr); hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, HDR_ISTYPE_METADATA(hdr)); arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); buf->b_flags |= ARC_BUF_FLAG_SHARED; /* * Since we've transferred ownership to the hdr we need * to increment its compressed and uncompressed kstats and * decrement the overhead size. */ ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); } static void arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) { ASSERT(arc_buf_is_shared(buf)); ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); /* * We are no longer sharing this buffer so we need * to transfer its ownership to the rightful owner. */ zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size, arc_hdr_size(hdr), hdr, buf); arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); abd_free(hdr->b_l1hdr.b_pabd); hdr->b_l1hdr.b_pabd = NULL; buf->b_flags &= ~ARC_BUF_FLAG_SHARED; /* * Since the buffer is no longer shared between * the arc buf and the hdr, count it as overhead. */ ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); } /* * Remove an arc_buf_t from the hdr's buf list and return the last * arc_buf_t on the list. If no buffers remain on the list then return * NULL. */ static arc_buf_t * arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) { ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; arc_buf_t *lastbuf = NULL; /* * Remove the buf from the hdr list and locate the last * remaining buffer on the list. */ while (*bufp != NULL) { if (*bufp == buf) *bufp = buf->b_next; /* * If we've removed a buffer in the middle of * the list then update the lastbuf and update * bufp. */ if (*bufp != NULL) { lastbuf = *bufp; bufp = &(*bufp)->b_next; } } buf->b_next = NULL; ASSERT3P(lastbuf, !=, buf); IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); return (lastbuf); } /* * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's * list and free it. */ static void arc_buf_destroy_impl(arc_buf_t *buf) { arc_buf_hdr_t *hdr = buf->b_hdr; /* * Free up the data associated with the buf but only if we're not * sharing this with the hdr. If we are sharing it with the hdr, the * hdr is responsible for doing the free. */ if (buf->b_data != NULL) { /* * We're about to change the hdr's b_flags. We must either * hold the hash_lock or be undiscoverable. */ ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); arc_cksum_verify(buf); arc_buf_unwatch(buf); if (arc_buf_is_shared(buf)) { arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); } else { uint64_t size = arc_buf_size(buf); arc_free_data_buf(hdr, buf->b_data, size, buf); ARCSTAT_INCR(arcstat_overhead_size, -size); } buf->b_data = NULL; ASSERT(hdr->b_l1hdr.b_bufcnt > 0); hdr->b_l1hdr.b_bufcnt -= 1; if (ARC_BUF_ENCRYPTED(buf)) { hdr->b_crypt_hdr.b_ebufcnt -= 1; /* * If we have no more encrypted buffers and we've * already gotten a copy of the decrypted data we can * free b_rabd to save some space. */ if (hdr->b_crypt_hdr.b_ebufcnt == 0 && HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL && !HDR_IO_IN_PROGRESS(hdr)) { arc_hdr_free_abd(hdr, B_TRUE); } } } arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { /* * If the current arc_buf_t is sharing its data buffer with the * hdr, then reassign the hdr's b_pabd to share it with the new * buffer at the end of the list. The shared buffer is always * the last one on the hdr's buffer list. * * There is an equivalent case for compressed bufs, but since * they aren't guaranteed to be the last buf in the list and * that is an exceedingly rare case, we just allow that space be * wasted temporarily. We must also be careful not to share * encrypted buffers, since they cannot be shared. */ if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) { /* Only one buf can be shared at once */ VERIFY(!arc_buf_is_shared(lastbuf)); /* hdr is uncompressed so can't have compressed buf */ VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); arc_hdr_free_abd(hdr, B_FALSE); /* * We must setup a new shared block between the * last buffer and the hdr. The data would have * been allocated by the arc buf so we need to transfer * ownership to the hdr since it's now being shared. */ arc_share_buf(hdr, lastbuf); } } else if (HDR_SHARED_DATA(hdr)) { /* * Uncompressed shared buffers are always at the end * of the list. Compressed buffers don't have the * same requirements. This makes it hard to * simply assert that the lastbuf is shared so * we rely on the hdr's compression flags to determine * if we have a compressed, shared buffer. */ ASSERT3P(lastbuf, !=, NULL); ASSERT(arc_buf_is_shared(lastbuf) || arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); } /* * Free the checksum if we're removing the last uncompressed buf from * this hdr. */ if (!arc_hdr_has_uncompressed_buf(hdr)) { arc_cksum_free(hdr); } /* clean up the buf */ buf->b_hdr = NULL; kmem_cache_free(buf_cache, buf); } static void arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags) { uint64_t size; boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0); boolean_t do_adapt = ((alloc_flags & ARC_HDR_DO_ADAPT) != 0); ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata); IMPLY(alloc_rdata, HDR_PROTECTED(hdr)); if (alloc_rdata) { size = HDR_GET_PSIZE(hdr); ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL); hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr, do_adapt); ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL); ARCSTAT_INCR(arcstat_raw_size, size); } else { size = arc_hdr_size(hdr); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr, do_adapt); ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); } ARCSTAT_INCR(arcstat_compressed_size, size); ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); } static void arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata) { uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr); ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); IMPLY(free_rdata, HDR_HAS_RABD(hdr)); /* * If the hdr is currently being written to the l2arc then * we defer freeing the data by adding it to the l2arc_free_on_write * list. The l2arc will free the data once it's finished * writing it to the l2arc device. */ if (HDR_L2_WRITING(hdr)) { arc_hdr_free_on_write(hdr, free_rdata); ARCSTAT_BUMP(arcstat_l2_free_on_write); } else if (free_rdata) { arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr); } else { arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr); } if (free_rdata) { hdr->b_crypt_hdr.b_rabd = NULL; ARCSTAT_INCR(arcstat_raw_size, -size); } else { hdr->b_l1hdr.b_pabd = NULL; } if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr)) hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; ARCSTAT_INCR(arcstat_compressed_size, -size); ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); } static arc_buf_hdr_t * arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, boolean_t protected, enum zio_compress compression_type, uint8_t complevel, arc_buf_contents_t type, boolean_t alloc_rdata) { arc_buf_hdr_t *hdr; int flags = ARC_HDR_DO_ADAPT; VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); if (protected) { hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE); } else { hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); } flags |= alloc_rdata ? ARC_HDR_ALLOC_RDATA : 0; ASSERT(HDR_EMPTY(hdr)); ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); HDR_SET_PSIZE(hdr, psize); HDR_SET_LSIZE(hdr, lsize); hdr->b_spa = spa; hdr->b_type = type; hdr->b_flags = 0; arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); arc_hdr_set_compress(hdr, compression_type); hdr->b_complevel = complevel; if (protected) arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED); hdr->b_l1hdr.b_state = arc_anon; hdr->b_l1hdr.b_arc_access = 0; hdr->b_l1hdr.b_bufcnt = 0; hdr->b_l1hdr.b_buf = NULL; /* * Allocate the hdr's buffer. This will contain either * the compressed or uncompressed data depending on the block * it references and compressed arc enablement. */ arc_hdr_alloc_abd(hdr, flags); ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); return (hdr); } /* * Transition between the two allocation states for the arc_buf_hdr struct. * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller * version is used when a cache buffer is only in the L2ARC in order to reduce * memory usage. */ static arc_buf_hdr_t * arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) { ASSERT(HDR_HAS_L2HDR(hdr)); arc_buf_hdr_t *nhdr; l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || (old == hdr_l2only_cache && new == hdr_full_cache)); /* * if the caller wanted a new full header and the header is to be * encrypted we will actually allocate the header from the full crypt * cache instead. The same applies to freeing from the old cache. */ if (HDR_PROTECTED(hdr) && new == hdr_full_cache) new = hdr_full_crypt_cache; if (HDR_PROTECTED(hdr) && old == hdr_full_cache) old = hdr_full_crypt_cache; nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); buf_hash_remove(hdr); bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); if (new == hdr_full_cache || new == hdr_full_crypt_cache) { arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); /* * arc_access and arc_change_state need to be aware that a * header has just come out of L2ARC, so we set its state to * l2c_only even though it's about to change. */ nhdr->b_l1hdr.b_state = arc_l2c_only; /* Verify previous threads set to NULL before freeing */ ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); } else { ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); ASSERT0(hdr->b_l1hdr.b_bufcnt); ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); /* * If we've reached here, We must have been called from * arc_evict_hdr(), as such we should have already been * removed from any ghost list we were previously on * (which protects us from racing with arc_evict_state), * thus no locking is needed during this check. */ ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); /* * A buffer must not be moved into the arc_l2c_only * state if it's not finished being written out to the * l2arc device. Otherwise, the b_l1hdr.b_pabd field * might try to be accessed, even though it was removed. */ VERIFY(!HDR_L2_WRITING(hdr)); VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); } /* * The header has been reallocated so we need to re-insert it into any * lists it was on. */ (void) buf_hash_insert(nhdr, NULL); ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); mutex_enter(&dev->l2ad_mtx); /* * We must place the realloc'ed header back into the list at * the same spot. Otherwise, if it's placed earlier in the list, * l2arc_write_buffers() could find it during the function's * write phase, and try to write it out to the l2arc. */ list_insert_after(&dev->l2ad_buflist, hdr, nhdr); list_remove(&dev->l2ad_buflist, hdr); mutex_exit(&dev->l2ad_mtx); /* * Since we're using the pointer address as the tag when * incrementing and decrementing the l2ad_alloc refcount, we * must remove the old pointer (that we're about to destroy) and * add the new pointer to the refcount. Otherwise we'd remove * the wrong pointer address when calling arc_hdr_destroy() later. */ (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); buf_discard_identity(hdr); kmem_cache_free(old, hdr); return (nhdr); } /* * This function allows an L1 header to be reallocated as a crypt * header and vice versa. If we are going to a crypt header, the * new fields will be zeroed out. */ static arc_buf_hdr_t * arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt) { arc_buf_hdr_t *nhdr; arc_buf_t *buf; kmem_cache_t *ncache, *ocache; unsigned nsize, osize; /* * This function requires that hdr is in the arc_anon state. * Therefore it won't have any L2ARC data for us to worry * about copying. */ ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT(!HDR_HAS_L2HDR(hdr)); ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt); ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node)); ASSERT3P(hdr->b_hash_next, ==, NULL); if (need_crypt) { ncache = hdr_full_crypt_cache; nsize = sizeof (hdr->b_crypt_hdr); ocache = hdr_full_cache; osize = HDR_FULL_SIZE; } else { ncache = hdr_full_cache; nsize = HDR_FULL_SIZE; ocache = hdr_full_crypt_cache; osize = sizeof (hdr->b_crypt_hdr); } nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE); /* * Copy all members that aren't locks or condvars to the new header. * No lists are pointing to us (as we asserted above), so we don't * need to worry about the list nodes. */ nhdr->b_dva = hdr->b_dva; nhdr->b_birth = hdr->b_birth; nhdr->b_type = hdr->b_type; nhdr->b_flags = hdr->b_flags; nhdr->b_psize = hdr->b_psize; nhdr->b_lsize = hdr->b_lsize; nhdr->b_spa = hdr->b_spa; nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum; nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt; nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap; nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state; nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access; nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits; nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits; nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits; nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits; nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits; nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb; nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd; /* * This zfs_refcount_add() exists only to ensure that the individual * arc buffers always point to a header that is referenced, avoiding * a small race condition that could trigger ASSERTs. */ (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG); nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf; for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { mutex_enter(&buf->b_evict_lock); buf->b_hdr = nhdr; mutex_exit(&buf->b_evict_lock); } zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt); (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG); ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); if (need_crypt) { arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED); } else { arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED); } /* unset all members of the original hdr */ bzero(&hdr->b_dva, sizeof (dva_t)); hdr->b_birth = 0; hdr->b_type = ARC_BUFC_INVALID; hdr->b_flags = 0; hdr->b_psize = 0; hdr->b_lsize = 0; hdr->b_spa = 0; hdr->b_l1hdr.b_freeze_cksum = NULL; hdr->b_l1hdr.b_buf = NULL; hdr->b_l1hdr.b_bufcnt = 0; hdr->b_l1hdr.b_byteswap = 0; hdr->b_l1hdr.b_state = NULL; hdr->b_l1hdr.b_arc_access = 0; hdr->b_l1hdr.b_mru_hits = 0; hdr->b_l1hdr.b_mru_ghost_hits = 0; hdr->b_l1hdr.b_mfu_hits = 0; hdr->b_l1hdr.b_mfu_ghost_hits = 0; hdr->b_l1hdr.b_l2_hits = 0; hdr->b_l1hdr.b_acb = NULL; hdr->b_l1hdr.b_pabd = NULL; if (ocache == hdr_full_crypt_cache) { ASSERT(!HDR_HAS_RABD(hdr)); hdr->b_crypt_hdr.b_ot = DMU_OT_NONE; hdr->b_crypt_hdr.b_ebufcnt = 0; hdr->b_crypt_hdr.b_dsobj = 0; bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); } buf_discard_identity(hdr); kmem_cache_free(ocache, hdr); return (nhdr); } /* * This function is used by the send / receive code to convert a newly * allocated arc_buf_t to one that is suitable for a raw encrypted write. It * is also used to allow the root objset block to be updated without altering * its embedded MACs. Both block types will always be uncompressed so we do not * have to worry about compression type or psize. */ void arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder, dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv, const uint8_t *mac) { arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET); ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED); if (!HDR_PROTECTED(hdr)) hdr = arc_hdr_realloc_crypt(hdr, B_TRUE); hdr->b_crypt_hdr.b_dsobj = dsobj; hdr->b_crypt_hdr.b_ot = ot; hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ? DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot); if (!arc_hdr_has_uncompressed_buf(hdr)) arc_cksum_free(hdr); if (salt != NULL) bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); if (iv != NULL) bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); if (mac != NULL) bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); } /* * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. * The buf is returned thawed since we expect the consumer to modify it. */ arc_buf_t * arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) { arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, B_FALSE, ZIO_COMPRESS_OFF, 0, type, B_FALSE); arc_buf_t *buf = NULL; VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE, B_FALSE, B_FALSE, &buf)); arc_buf_thaw(buf); return (buf); } /* * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this * for bufs containing metadata. */ arc_buf_t * arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, enum zio_compress compression_type, uint8_t complevel) { ASSERT3U(lsize, >, 0); ASSERT3U(lsize, >=, psize); ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF); ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS); arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_FALSE, compression_type, complevel, ARC_BUFC_DATA, B_FALSE); arc_buf_t *buf = NULL; VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_TRUE, B_FALSE, B_FALSE, &buf)); arc_buf_thaw(buf); ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); if (!arc_buf_is_shared(buf)) { /* * To ensure that the hdr has the correct data in it if we call * arc_untransform() on this buf before it's been written to * disk, it's easiest if we just set up sharing between the * buf and the hdr. */ arc_hdr_free_abd(hdr, B_FALSE); arc_share_buf(hdr, buf); } return (buf); } arc_buf_t * arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder, const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_object_type_t ot, uint64_t psize, uint64_t lsize, enum zio_compress compression_type, uint8_t complevel) { arc_buf_hdr_t *hdr; arc_buf_t *buf; arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ? ARC_BUFC_METADATA : ARC_BUFC_DATA; ASSERT3U(lsize, >, 0); ASSERT3U(lsize, >=, psize); ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF); ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS); hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE, compression_type, complevel, type, B_TRUE); hdr->b_crypt_hdr.b_dsobj = dsobj; hdr->b_crypt_hdr.b_ot = ot; hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ? DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot); bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); /* * This buffer will be considered encrypted even if the ot is not an * encrypted type. It will become authenticated instead in * arc_write_ready(). */ buf = NULL; VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE, B_FALSE, B_FALSE, &buf)); arc_buf_thaw(buf); ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); return (buf); } static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr, boolean_t state_only) { l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; l2arc_dev_t *dev = l2hdr->b_dev; uint64_t lsize = HDR_GET_LSIZE(hdr); uint64_t psize = HDR_GET_PSIZE(hdr); uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize); arc_buf_contents_t type = hdr->b_type; int64_t lsize_s; int64_t psize_s; int64_t asize_s; if (incr) { lsize_s = lsize; psize_s = psize; asize_s = asize; } else { lsize_s = -lsize; psize_s = -psize; asize_s = -asize; } /* If the buffer is a prefetch, count it as such. */ if (HDR_PREFETCH(hdr)) { ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s); } else { /* * We use the value stored in the L2 header upon initial * caching in L2ARC. This value will be updated in case * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC * metadata (log entry) cannot currently be updated. Having * the ARC state in the L2 header solves the problem of a * possibly absent L1 header (apparent in buffers restored * from persistent L2ARC). */ switch (hdr->b_l2hdr.b_arcs_state) { case ARC_STATE_MRU_GHOST: case ARC_STATE_MRU: ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s); break; case ARC_STATE_MFU_GHOST: case ARC_STATE_MFU: ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s); break; default: break; } } if (state_only) return; ARCSTAT_INCR(arcstat_l2_psize, psize_s); ARCSTAT_INCR(arcstat_l2_lsize, lsize_s); switch (type) { case ARC_BUFC_DATA: ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s); break; case ARC_BUFC_METADATA: ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s); break; default: break; } } static void arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) { l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; l2arc_dev_t *dev = l2hdr->b_dev; uint64_t psize = HDR_GET_PSIZE(hdr); uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize); ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); ASSERT(HDR_HAS_L2HDR(hdr)); list_remove(&dev->l2ad_buflist, hdr); l2arc_hdr_arcstats_decrement(hdr); vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); } static void arc_hdr_destroy(arc_buf_hdr_t *hdr) { if (HDR_HAS_L1HDR(hdr)) { ASSERT(hdr->b_l1hdr.b_buf == NULL || hdr->b_l1hdr.b_bufcnt > 0); ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); } ASSERT(!HDR_IO_IN_PROGRESS(hdr)); ASSERT(!HDR_IN_HASH_TABLE(hdr)); if (HDR_HAS_L2HDR(hdr)) { l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); if (!buflist_held) mutex_enter(&dev->l2ad_mtx); /* * Even though we checked this conditional above, we * need to check this again now that we have the * l2ad_mtx. This is because we could be racing with * another thread calling l2arc_evict() which might have * destroyed this header's L2 portion as we were waiting * to acquire the l2ad_mtx. If that happens, we don't * want to re-destroy the header's L2 portion. */ if (HDR_HAS_L2HDR(hdr)) arc_hdr_l2hdr_destroy(hdr); if (!buflist_held) mutex_exit(&dev->l2ad_mtx); } /* * The header's identify can only be safely discarded once it is no * longer discoverable. This requires removing it from the hash table * and the l2arc header list. After this point the hash lock can not * be used to protect the header. */ if (!HDR_EMPTY(hdr)) buf_discard_identity(hdr); if (HDR_HAS_L1HDR(hdr)) { arc_cksum_free(hdr); while (hdr->b_l1hdr.b_buf != NULL) arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); if (hdr->b_l1hdr.b_pabd != NULL) arc_hdr_free_abd(hdr, B_FALSE); if (HDR_HAS_RABD(hdr)) arc_hdr_free_abd(hdr, B_TRUE); } ASSERT3P(hdr->b_hash_next, ==, NULL); if (HDR_HAS_L1HDR(hdr)) { ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); if (!HDR_PROTECTED(hdr)) { kmem_cache_free(hdr_full_cache, hdr); } else { kmem_cache_free(hdr_full_crypt_cache, hdr); } } else { kmem_cache_free(hdr_l2only_cache, hdr); } } void arc_buf_destroy(arc_buf_t *buf, void* tag) { arc_buf_hdr_t *hdr = buf->b_hdr; if (hdr->b_l1hdr.b_state == arc_anon) { ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); ASSERT(!HDR_IO_IN_PROGRESS(hdr)); VERIFY0(remove_reference(hdr, NULL, tag)); arc_hdr_destroy(hdr); return; } kmutex_t *hash_lock = HDR_LOCK(hdr); mutex_enter(hash_lock); ASSERT3P(hdr, ==, buf->b_hdr); ASSERT(hdr->b_l1hdr.b_bufcnt > 0); ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); ASSERT3P(buf->b_data, !=, NULL); (void) remove_reference(hdr, hash_lock, tag); arc_buf_destroy_impl(buf); mutex_exit(hash_lock); } /* * Evict the arc_buf_hdr that is provided as a parameter. The resultant * state of the header is dependent on its state prior to entering this * function. The following transitions are possible: * * - arc_mru -> arc_mru_ghost * - arc_mfu -> arc_mfu_ghost * - arc_mru_ghost -> arc_l2c_only * - arc_mru_ghost -> deleted * - arc_mfu_ghost -> arc_l2c_only * - arc_mfu_ghost -> deleted * * Return total size of evicted data buffers for eviction progress tracking. * When evicting from ghost states return logical buffer size to make eviction * progress at the same (or at least comparable) rate as from non-ghost states. * * Return *real_evicted for actual ARC size reduction to wake up threads * waiting for it. For non-ghost states it includes size of evicted data * buffers (the headers are not freed there). For ghost states it includes * only the evicted headers size. */ static int64_t arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, uint64_t *real_evicted) { arc_state_t *evicted_state, *state; int64_t bytes_evicted = 0; int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ? arc_min_prescient_prefetch_ms : arc_min_prefetch_ms; ASSERT(MUTEX_HELD(hash_lock)); ASSERT(HDR_HAS_L1HDR(hdr)); *real_evicted = 0; state = hdr->b_l1hdr.b_state; if (GHOST_STATE(state)) { ASSERT(!HDR_IO_IN_PROGRESS(hdr)); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); /* * l2arc_write_buffers() relies on a header's L1 portion * (i.e. its b_pabd field) during it's write phase. * Thus, we cannot push a header onto the arc_l2c_only * state (removing its L1 piece) until the header is * done being written to the l2arc. */ if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { ARCSTAT_BUMP(arcstat_evict_l2_skip); return (bytes_evicted); } ARCSTAT_BUMP(arcstat_deleted); bytes_evicted += HDR_GET_LSIZE(hdr); DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); if (HDR_HAS_L2HDR(hdr)) { ASSERT(hdr->b_l1hdr.b_pabd == NULL); ASSERT(!HDR_HAS_RABD(hdr)); /* * This buffer is cached on the 2nd Level ARC; * don't destroy the header. */ arc_change_state(arc_l2c_only, hdr, hash_lock); /* * dropping from L1+L2 cached to L2-only, * realloc to remove the L1 header. */ hdr = arc_hdr_realloc(hdr, hdr_full_cache, hdr_l2only_cache); *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE; } else { arc_change_state(arc_anon, hdr, hash_lock); arc_hdr_destroy(hdr); *real_evicted += HDR_FULL_SIZE; } return (bytes_evicted); } ASSERT(state == arc_mru || state == arc_mfu); evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; /* prefetch buffers have a minimum lifespan */ if (HDR_IO_IN_PROGRESS(hdr) || ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < MSEC_TO_TICK(min_lifetime))) { ARCSTAT_BUMP(arcstat_evict_skip); return (bytes_evicted); } ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); while (hdr->b_l1hdr.b_buf) { arc_buf_t *buf = hdr->b_l1hdr.b_buf; if (!mutex_tryenter(&buf->b_evict_lock)) { ARCSTAT_BUMP(arcstat_mutex_miss); break; } if (buf->b_data != NULL) { bytes_evicted += HDR_GET_LSIZE(hdr); *real_evicted += HDR_GET_LSIZE(hdr); } mutex_exit(&buf->b_evict_lock); arc_buf_destroy_impl(buf); } if (HDR_HAS_L2HDR(hdr)) { ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); } else { if (l2arc_write_eligible(hdr->b_spa, hdr)) { ARCSTAT_INCR(arcstat_evict_l2_eligible, HDR_GET_LSIZE(hdr)); switch (state->arcs_state) { case ARC_STATE_MRU: ARCSTAT_INCR( arcstat_evict_l2_eligible_mru, HDR_GET_LSIZE(hdr)); break; case ARC_STATE_MFU: ARCSTAT_INCR( arcstat_evict_l2_eligible_mfu, HDR_GET_LSIZE(hdr)); break; default: break; } } else { ARCSTAT_INCR(arcstat_evict_l2_ineligible, HDR_GET_LSIZE(hdr)); } } if (hdr->b_l1hdr.b_bufcnt == 0) { arc_cksum_free(hdr); bytes_evicted += arc_hdr_size(hdr); *real_evicted += arc_hdr_size(hdr); /* * If this hdr is being evicted and has a compressed * buffer then we discard it here before we change states. * This ensures that the accounting is updated correctly * in arc_free_data_impl(). */ if (hdr->b_l1hdr.b_pabd != NULL) arc_hdr_free_abd(hdr, B_FALSE); if (HDR_HAS_RABD(hdr)) arc_hdr_free_abd(hdr, B_TRUE); arc_change_state(evicted_state, hdr, hash_lock); ASSERT(HDR_IN_HASH_TABLE(hdr)); arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); } return (bytes_evicted); } static void arc_set_need_free(void) { ASSERT(MUTEX_HELD(&arc_evict_lock)); int64_t remaining = arc_free_memory() - arc_sys_free / 2; arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters); if (aw == NULL) { arc_need_free = MAX(-remaining, 0); } else { arc_need_free = MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count)); } } static uint64_t arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, uint64_t spa, uint64_t bytes) { multilist_sublist_t *mls; uint64_t bytes_evicted = 0, real_evicted = 0; arc_buf_hdr_t *hdr; kmutex_t *hash_lock; int evict_count = zfs_arc_evict_batch_limit; ASSERT3P(marker, !=, NULL); mls = multilist_sublist_lock(ml, idx); for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL); hdr = multilist_sublist_prev(mls, marker)) { if ((evict_count <= 0) || (bytes_evicted >= bytes)) break; /* * To keep our iteration location, move the marker * forward. Since we're not holding hdr's hash lock, we * must be very careful and not remove 'hdr' from the * sublist. Otherwise, other consumers might mistake the * 'hdr' as not being on a sublist when they call the * multilist_link_active() function (they all rely on * the hash lock protecting concurrent insertions and * removals). multilist_sublist_move_forward() was * specifically implemented to ensure this is the case * (only 'marker' will be removed and re-inserted). */ multilist_sublist_move_forward(mls, marker); /* * The only case where the b_spa field should ever be * zero, is the marker headers inserted by * arc_evict_state(). It's possible for multiple threads * to be calling arc_evict_state() concurrently (e.g. * dsl_pool_close() and zio_inject_fault()), so we must * skip any markers we see from these other threads. */ if (hdr->b_spa == 0) continue; /* we're only interested in evicting buffers of a certain spa */ if (spa != 0 && hdr->b_spa != spa) { ARCSTAT_BUMP(arcstat_evict_skip); continue; } hash_lock = HDR_LOCK(hdr); /* * We aren't calling this function from any code path * that would already be holding a hash lock, so we're * asserting on this assumption to be defensive in case * this ever changes. Without this check, it would be * possible to incorrectly increment arcstat_mutex_miss * below (e.g. if the code changed such that we called * this function with a hash lock held). */ ASSERT(!MUTEX_HELD(hash_lock)); if (mutex_tryenter(hash_lock)) { uint64_t revicted; uint64_t evicted = arc_evict_hdr(hdr, hash_lock, &revicted); mutex_exit(hash_lock); bytes_evicted += evicted; real_evicted += revicted; /* * If evicted is zero, arc_evict_hdr() must have * decided to skip this header, don't increment * evict_count in this case. */ if (evicted != 0) evict_count--; } else { ARCSTAT_BUMP(arcstat_mutex_miss); } } multilist_sublist_unlock(mls); /* * Increment the count of evicted bytes, and wake up any threads that * are waiting for the count to reach this value. Since the list is * ordered by ascending aew_count, we pop off the beginning of the * list until we reach the end, or a waiter that's past the current * "count". Doing this outside the loop reduces the number of times * we need to acquire the global arc_evict_lock. * * Only wake when there's sufficient free memory in the system * (specifically, arc_sys_free/2, which by default is a bit more than * 1/64th of RAM). See the comments in arc_wait_for_eviction(). */ mutex_enter(&arc_evict_lock); arc_evict_count += real_evicted; if (arc_free_memory() > arc_sys_free / 2) { arc_evict_waiter_t *aw; while ((aw = list_head(&arc_evict_waiters)) != NULL && aw->aew_count <= arc_evict_count) { list_remove(&arc_evict_waiters, aw); cv_broadcast(&aw->aew_cv); } } arc_set_need_free(); mutex_exit(&arc_evict_lock); /* * If the ARC size is reduced from arc_c_max to arc_c_min (especially * if the average cached block is small), eviction can be on-CPU for * many seconds. To ensure that other threads that may be bound to * this CPU are able to make progress, make a voluntary preemption * call here. */ cond_resched(); return (bytes_evicted); } /* * Evict buffers from the given arc state, until we've removed the * specified number of bytes. Move the removed buffers to the * appropriate evict state. * * This function makes a "best effort". It skips over any buffers * it can't get a hash_lock on, and so, may not catch all candidates. * It may also return without evicting as much space as requested. * * If bytes is specified using the special value ARC_EVICT_ALL, this * will evict all available (i.e. unlocked and evictable) buffers from * the given arc state; which is used by arc_flush(). */ static uint64_t arc_evict_state(arc_state_t *state, uint64_t spa, uint64_t bytes, arc_buf_contents_t type) { uint64_t total_evicted = 0; multilist_t *ml = &state->arcs_list[type]; int num_sublists; arc_buf_hdr_t **markers; num_sublists = multilist_get_num_sublists(ml); /* * If we've tried to evict from each sublist, made some * progress, but still have not hit the target number of bytes * to evict, we want to keep trying. The markers allow us to * pick up where we left off for each individual sublist, rather * than starting from the tail each time. */ markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); for (int i = 0; i < num_sublists; i++) { multilist_sublist_t *mls; markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); /* * A b_spa of 0 is used to indicate that this header is * a marker. This fact is used in arc_evict_type() and * arc_evict_state_impl(). */ markers[i]->b_spa = 0; mls = multilist_sublist_lock(ml, i); multilist_sublist_insert_tail(mls, markers[i]); multilist_sublist_unlock(mls); } /* * While we haven't hit our target number of bytes to evict, or * we're evicting all available buffers. */ while (total_evicted < bytes) { int sublist_idx = multilist_get_random_index(ml); uint64_t scan_evicted = 0; /* * Try to reduce pinned dnodes with a floor of arc_dnode_limit. * Request that 10% of the LRUs be scanned by the superblock * shrinker. */ if (type == ARC_BUFC_DATA && aggsum_compare( &arc_sums.arcstat_dnode_size, arc_dnode_size_limit) > 0) { arc_prune_async((aggsum_upper_bound( &arc_sums.arcstat_dnode_size) - arc_dnode_size_limit) / sizeof (dnode_t) / zfs_arc_dnode_reduce_percent); } /* * Start eviction using a randomly selected sublist, * this is to try and evenly balance eviction across all * sublists. Always starting at the same sublist * (e.g. index 0) would cause evictions to favor certain * sublists over others. */ for (int i = 0; i < num_sublists; i++) { uint64_t bytes_remaining; uint64_t bytes_evicted; if (total_evicted < bytes) bytes_remaining = bytes - total_evicted; else break; bytes_evicted = arc_evict_state_impl(ml, sublist_idx, markers[sublist_idx], spa, bytes_remaining); scan_evicted += bytes_evicted; total_evicted += bytes_evicted; /* we've reached the end, wrap to the beginning */ if (++sublist_idx >= num_sublists) sublist_idx = 0; } /* * If we didn't evict anything during this scan, we have * no reason to believe we'll evict more during another * scan, so break the loop. */ if (scan_evicted == 0) { /* This isn't possible, let's make that obvious */ ASSERT3S(bytes, !=, 0); /* * When bytes is ARC_EVICT_ALL, the only way to * break the loop is when scan_evicted is zero. * In that case, we actually have evicted enough, * so we don't want to increment the kstat. */ if (bytes != ARC_EVICT_ALL) { ASSERT3S(total_evicted, <, bytes); ARCSTAT_BUMP(arcstat_evict_not_enough); } break; } } for (int i = 0; i < num_sublists; i++) { multilist_sublist_t *mls = multilist_sublist_lock(ml, i); multilist_sublist_remove(mls, markers[i]); multilist_sublist_unlock(mls); kmem_cache_free(hdr_full_cache, markers[i]); } kmem_free(markers, sizeof (*markers) * num_sublists); return (total_evicted); } /* * Flush all "evictable" data of the given type from the arc state * specified. This will not evict any "active" buffers (i.e. referenced). * * When 'retry' is set to B_FALSE, the function will make a single pass * over the state and evict any buffers that it can. Since it doesn't * continually retry the eviction, it might end up leaving some buffers * in the ARC due to lock misses. * * When 'retry' is set to B_TRUE, the function will continually retry the * eviction until *all* evictable buffers have been removed from the * state. As a result, if concurrent insertions into the state are * allowed (e.g. if the ARC isn't shutting down), this function might * wind up in an infinite loop, continually trying to evict buffers. */ static uint64_t arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, boolean_t retry) { uint64_t evicted = 0; while (zfs_refcount_count(&state->arcs_esize[type]) != 0) { evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); if (!retry) break; } return (evicted); } /* * Evict the specified number of bytes from the state specified, * restricting eviction to the spa and type given. This function * prevents us from trying to evict more from a state's list than * is "evictable", and to skip evicting altogether when passed a * negative value for "bytes". In contrast, arc_evict_state() will * evict everything it can, when passed a negative value for "bytes". */ static uint64_t arc_evict_impl(arc_state_t *state, uint64_t spa, int64_t bytes, arc_buf_contents_t type) { uint64_t delta; if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) { delta = MIN(zfs_refcount_count(&state->arcs_esize[type]), bytes); return (arc_evict_state(state, spa, delta, type)); } return (0); } /* * The goal of this function is to evict enough meta data buffers from the * ARC in order to enforce the arc_meta_limit. Achieving this is slightly * more complicated than it appears because it is common for data buffers * to have holds on meta data buffers. In addition, dnode meta data buffers * will be held by the dnodes in the block preventing them from being freed. * This means we can't simply traverse the ARC and expect to always find * enough unheld meta data buffer to release. * * Therefore, this function has been updated to make alternating passes * over the ARC releasing data buffers and then newly unheld meta data * buffers. This ensures forward progress is maintained and meta_used * will decrease. Normally this is sufficient, but if required the ARC * will call the registered prune callbacks causing dentry and inodes to * be dropped from the VFS cache. This will make dnode meta data buffers * available for reclaim. */ static uint64_t arc_evict_meta_balanced(uint64_t meta_used) { int64_t delta, prune = 0, adjustmnt; uint64_t total_evicted = 0; arc_buf_contents_t type = ARC_BUFC_DATA; int restarts = MAX(zfs_arc_meta_adjust_restarts, 0); restart: /* * This slightly differs than the way we evict from the mru in * arc_evict because we don't have a "target" value (i.e. no * "meta" arc_p). As a result, I think we can completely * cannibalize the metadata in the MRU before we evict the * metadata from the MFU. I think we probably need to implement a * "metadata arc_p" value to do this properly. */ adjustmnt = meta_used - arc_meta_limit; if (adjustmnt > 0 && zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) { delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]), adjustmnt); total_evicted += arc_evict_impl(arc_mru, 0, delta, type); adjustmnt -= delta; } /* * We can't afford to recalculate adjustmnt here. If we do, * new metadata buffers can sneak into the MRU or ANON lists, * thus penalize the MFU metadata. Although the fudge factor is * small, it has been empirically shown to be significant for * certain workloads (e.g. creating many empty directories). As * such, we use the original calculation for adjustmnt, and * simply decrement the amount of data evicted from the MRU. */ if (adjustmnt > 0 && zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) { delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]), adjustmnt); total_evicted += arc_evict_impl(arc_mfu, 0, delta, type); } adjustmnt = meta_used - arc_meta_limit; if (adjustmnt > 0 && zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) { delta = MIN(adjustmnt, zfs_refcount_count(&arc_mru_ghost->arcs_esize[type])); total_evicted += arc_evict_impl(arc_mru_ghost, 0, delta, type); adjustmnt -= delta; } if (adjustmnt > 0 && zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) { delta = MIN(adjustmnt, zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type])); total_evicted += arc_evict_impl(arc_mfu_ghost, 0, delta, type); } /* * If after attempting to make the requested adjustment to the ARC * the meta limit is still being exceeded then request that the * higher layers drop some cached objects which have holds on ARC * meta buffers. Requests to the upper layers will be made with * increasingly large scan sizes until the ARC is below the limit. */ if (meta_used > arc_meta_limit) { if (type == ARC_BUFC_DATA) { type = ARC_BUFC_METADATA; } else { type = ARC_BUFC_DATA; if (zfs_arc_meta_prune) { prune += zfs_arc_meta_prune; arc_prune_async(prune); } } if (restarts > 0) { restarts--; goto restart; } } return (total_evicted); } /* * Evict metadata buffers from the cache, such that arcstat_meta_used is * capped by the arc_meta_limit tunable. */ static uint64_t arc_evict_meta_only(uint64_t meta_used) { uint64_t total_evicted = 0; int64_t target; /* * If we're over the meta limit, we want to evict enough * metadata to get back under the meta limit. We don't want to * evict so much that we drop the MRU below arc_p, though. If * we're over the meta limit more than we're over arc_p, we * evict some from the MRU here, and some from the MFU below. */ target = MIN((int64_t)(meta_used - arc_meta_limit), (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + zfs_refcount_count(&arc_mru->arcs_size) - arc_p)); total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA); /* * Similar to the above, we want to evict enough bytes to get us * below the meta limit, but not so much as to drop us below the * space allotted to the MFU (which is defined as arc_c - arc_p). */ target = MIN((int64_t)(meta_used - arc_meta_limit), (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); return (total_evicted); } static uint64_t arc_evict_meta(uint64_t meta_used) { if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY) return (arc_evict_meta_only(meta_used)); else return (arc_evict_meta_balanced(meta_used)); } /* * Return the type of the oldest buffer in the given arc state * * This function will select a random sublist of type ARC_BUFC_DATA and * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist * is compared, and the type which contains the "older" buffer will be * returned. */ static arc_buf_contents_t arc_evict_type(arc_state_t *state) { multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; int data_idx = multilist_get_random_index(data_ml); int meta_idx = multilist_get_random_index(meta_ml); multilist_sublist_t *data_mls; multilist_sublist_t *meta_mls; arc_buf_contents_t type; arc_buf_hdr_t *data_hdr; arc_buf_hdr_t *meta_hdr; /* * We keep the sublist lock until we're finished, to prevent * the headers from being destroyed via arc_evict_state(). */ data_mls = multilist_sublist_lock(data_ml, data_idx); meta_mls = multilist_sublist_lock(meta_ml, meta_idx); /* * These two loops are to ensure we skip any markers that * might be at the tail of the lists due to arc_evict_state(). */ for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { if (data_hdr->b_spa != 0) break; } for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { if (meta_hdr->b_spa != 0) break; } if (data_hdr == NULL && meta_hdr == NULL) { type = ARC_BUFC_DATA; } else if (data_hdr == NULL) { ASSERT3P(meta_hdr, !=, NULL); type = ARC_BUFC_METADATA; } else if (meta_hdr == NULL) { ASSERT3P(data_hdr, !=, NULL); type = ARC_BUFC_DATA; } else { ASSERT3P(data_hdr, !=, NULL); ASSERT3P(meta_hdr, !=, NULL); /* The headers can't be on the sublist without an L1 header */ ASSERT(HDR_HAS_L1HDR(data_hdr)); ASSERT(HDR_HAS_L1HDR(meta_hdr)); if (data_hdr->b_l1hdr.b_arc_access < meta_hdr->b_l1hdr.b_arc_access) { type = ARC_BUFC_DATA; } else { type = ARC_BUFC_METADATA; } } multilist_sublist_unlock(meta_mls); multilist_sublist_unlock(data_mls); return (type); } /* * Evict buffers from the cache, such that arcstat_size is capped by arc_c. */ static uint64_t arc_evict(void) { uint64_t total_evicted = 0; uint64_t bytes; int64_t target; uint64_t asize = aggsum_value(&arc_sums.arcstat_size); uint64_t ameta = aggsum_value(&arc_sums.arcstat_meta_used); /* * If we're over arc_meta_limit, we want to correct that before * potentially evicting data buffers below. */ total_evicted += arc_evict_meta(ameta); /* * Adjust MRU size * * If we're over the target cache size, we want to evict enough * from the list to get back to our target size. We don't want * to evict too much from the MRU, such that it drops below * arc_p. So, if we're over our target cache size more than * the MRU is over arc_p, we'll evict enough to get back to * arc_p here, and then evict more from the MFU below. */ target = MIN((int64_t)(asize - arc_c), (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); /* * If we're below arc_meta_min, always prefer to evict data. * Otherwise, try to satisfy the requested number of bytes to * evict from the type which contains older buffers; in an * effort to keep newer buffers in the cache regardless of their * type. If we cannot satisfy the number of bytes from this * type, spill over into the next type. */ if (arc_evict_type(arc_mru) == ARC_BUFC_METADATA && ameta > arc_meta_min) { bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA); total_evicted += bytes; /* * If we couldn't evict our target number of bytes from * metadata, we try to get the rest from data. */ target -= bytes; total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA); } else { bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA); total_evicted += bytes; /* * If we couldn't evict our target number of bytes from * data, we try to get the rest from metadata. */ target -= bytes; total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA); } /* * Re-sum ARC stats after the first round of evictions. */ asize = aggsum_value(&arc_sums.arcstat_size); ameta = aggsum_value(&arc_sums.arcstat_meta_used); /* * Adjust MFU size * * Now that we've tried to evict enough from the MRU to get its * size back to arc_p, if we're still above the target cache * size, we evict the rest from the MFU. */ target = asize - arc_c; if (arc_evict_type(arc_mfu) == ARC_BUFC_METADATA && ameta > arc_meta_min) { bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); total_evicted += bytes; /* * If we couldn't evict our target number of bytes from * metadata, we try to get the rest from data. */ target -= bytes; total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA); } else { bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA); total_evicted += bytes; /* * If we couldn't evict our target number of bytes from * data, we try to get the rest from data. */ target -= bytes; total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); } /* * Adjust ghost lists * * In addition to the above, the ARC also defines target values * for the ghost lists. The sum of the mru list and mru ghost * list should never exceed the target size of the cache, and * the sum of the mru list, mfu list, mru ghost list, and mfu * ghost list should never exceed twice the target size of the * cache. The following logic enforces these limits on the ghost * caches, and evicts from them as needed. */ target = zfs_refcount_count(&arc_mru->arcs_size) + zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c; bytes = arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); total_evicted += bytes; target -= bytes; total_evicted += arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); /* * We assume the sum of the mru list and mfu list is less than * or equal to arc_c (we enforced this above), which means we * can use the simpler of the two equations below: * * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c * mru ghost + mfu ghost <= arc_c */ target = zfs_refcount_count(&arc_mru_ghost->arcs_size) + zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; bytes = arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); total_evicted += bytes; target -= bytes; total_evicted += arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); return (total_evicted); } void arc_flush(spa_t *spa, boolean_t retry) { uint64_t guid = 0; /* * If retry is B_TRUE, a spa must not be specified since we have * no good way to determine if all of a spa's buffers have been * evicted from an arc state. */ ASSERT(!retry || spa == 0); if (spa != NULL) guid = spa_load_guid(spa); (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); } void arc_reduce_target_size(int64_t to_free) { uint64_t asize = aggsum_value(&arc_sums.arcstat_size); /* * All callers want the ARC to actually evict (at least) this much * memory. Therefore we reduce from the lower of the current size and * the target size. This way, even if arc_c is much higher than * arc_size (as can be the case after many calls to arc_freed(), we will * immediately have arc_c < arc_size and therefore the arc_evict_zthr * will evict. */ uint64_t c = MIN(arc_c, asize); if (c > to_free && c - to_free > arc_c_min) { arc_c = c - to_free; atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); if (arc_p > arc_c) arc_p = (arc_c >> 1); ASSERT(arc_c >= arc_c_min); ASSERT((int64_t)arc_p >= 0); } else { arc_c = arc_c_min; } if (asize > arc_c) { /* See comment in arc_evict_cb_check() on why lock+flag */ mutex_enter(&arc_evict_lock); arc_evict_needed = B_TRUE; mutex_exit(&arc_evict_lock); zthr_wakeup(arc_evict_zthr); } } /* * Determine if the system is under memory pressure and is asking * to reclaim memory. A return value of B_TRUE indicates that the system * is under memory pressure and that the arc should adjust accordingly. */ boolean_t arc_reclaim_needed(void) { return (arc_available_memory() < 0); } void arc_kmem_reap_soon(void) { size_t i; kmem_cache_t *prev_cache = NULL; kmem_cache_t *prev_data_cache = NULL; extern kmem_cache_t *zio_buf_cache[]; extern kmem_cache_t *zio_data_buf_cache[]; #ifdef _KERNEL if ((aggsum_compare(&arc_sums.arcstat_meta_used, arc_meta_limit) >= 0) && zfs_arc_meta_prune) { /* * We are exceeding our meta-data cache limit. * Prune some entries to release holds on meta-data. */ arc_prune_async(zfs_arc_meta_prune); } #if defined(_ILP32) /* * Reclaim unused memory from all kmem caches. */ kmem_reap(); #endif #endif for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { #if defined(_ILP32) /* reach upper limit of cache size on 32-bit */ if (zio_buf_cache[i] == NULL) break; #endif if (zio_buf_cache[i] != prev_cache) { prev_cache = zio_buf_cache[i]; kmem_cache_reap_now(zio_buf_cache[i]); } if (zio_data_buf_cache[i] != prev_data_cache) { prev_data_cache = zio_data_buf_cache[i]; kmem_cache_reap_now(zio_data_buf_cache[i]); } } kmem_cache_reap_now(buf_cache); kmem_cache_reap_now(hdr_full_cache); kmem_cache_reap_now(hdr_l2only_cache); kmem_cache_reap_now(zfs_btree_leaf_cache); abd_cache_reap_now(); } /* ARGSUSED */ static boolean_t arc_evict_cb_check(void *arg, zthr_t *zthr) { #ifdef ZFS_DEBUG /* * This is necessary in order to keep the kstat information * up to date for tools that display kstat data such as the * mdb ::arc dcmd and the Linux crash utility. These tools * typically do not call kstat's update function, but simply * dump out stats from the most recent update. Without * this call, these commands may show stale stats for the * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even * with this call, the data might be out of date if the * evict thread hasn't been woken recently; but that should * suffice. The arc_state_t structures can be queried * directly if more accurate information is needed. */ if (arc_ksp != NULL) arc_ksp->ks_update(arc_ksp, KSTAT_READ); #endif /* * We have to rely on arc_wait_for_eviction() to tell us when to * evict, rather than checking if we are overflowing here, so that we * are sure to not leave arc_wait_for_eviction() waiting on aew_cv. * If we have become "not overflowing" since arc_wait_for_eviction() * checked, we need to wake it up. We could broadcast the CV here, * but arc_wait_for_eviction() may have not yet gone to sleep. We * would need to use a mutex to ensure that this function doesn't * broadcast until arc_wait_for_eviction() has gone to sleep (e.g. * the arc_evict_lock). However, the lock ordering of such a lock * would necessarily be incorrect with respect to the zthr_lock, * which is held before this function is called, and is held by * arc_wait_for_eviction() when it calls zthr_wakeup(). */ return (arc_evict_needed); } /* * Keep arc_size under arc_c by running arc_evict which evicts data * from the ARC. */ /* ARGSUSED */ static void arc_evict_cb(void *arg, zthr_t *zthr) { uint64_t evicted = 0; fstrans_cookie_t cookie = spl_fstrans_mark(); /* Evict from cache */ evicted = arc_evict(); /* * If evicted is zero, we couldn't evict anything * via arc_evict(). This could be due to hash lock * collisions, but more likely due to the majority of * arc buffers being unevictable. Therefore, even if * arc_size is above arc_c, another pass is unlikely to * be helpful and could potentially cause us to enter an * infinite loop. Additionally, zthr_iscancelled() is * checked here so that if the arc is shutting down, the * broadcast will wake any remaining arc evict waiters. */ mutex_enter(&arc_evict_lock); arc_evict_needed = !zthr_iscancelled(arc_evict_zthr) && evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0; if (!arc_evict_needed) { /* * We're either no longer overflowing, or we * can't evict anything more, so we should wake * arc_get_data_impl() sooner. */ arc_evict_waiter_t *aw; while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) { cv_broadcast(&aw->aew_cv); } arc_set_need_free(); } mutex_exit(&arc_evict_lock); spl_fstrans_unmark(cookie); } /* ARGSUSED */ static boolean_t arc_reap_cb_check(void *arg, zthr_t *zthr) { int64_t free_memory = arc_available_memory(); static int reap_cb_check_counter = 0; /* * If a kmem reap is already active, don't schedule more. We must * check for this because kmem_cache_reap_soon() won't actually * block on the cache being reaped (this is to prevent callers from * becoming implicitly blocked by a system-wide kmem reap -- which, * on a system with many, many full magazines, can take minutes). */ if (!kmem_cache_reap_active() && free_memory < 0) { arc_no_grow = B_TRUE; arc_warm = B_TRUE; /* * Wait at least zfs_grow_retry (default 5) seconds * before considering growing. */ arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry); return (B_TRUE); } else if (free_memory < arc_c >> arc_no_grow_shift) { arc_no_grow = B_TRUE; } else if (gethrtime() >= arc_growtime) { arc_no_grow = B_FALSE; } /* * Called unconditionally every 60 seconds to reclaim unused * zstd compression and decompression context. This is done * here to avoid the need for an independent thread. */ if (!((reap_cb_check_counter++) % 60)) zfs_zstd_cache_reap_now(); return (B_FALSE); } /* * Keep enough free memory in the system by reaping the ARC's kmem * caches. To cause more slabs to be reapable, we may reduce the * target size of the cache (arc_c), causing the arc_evict_cb() * to free more buffers. */ /* ARGSUSED */ static void arc_reap_cb(void *arg, zthr_t *zthr) { int64_t free_memory; fstrans_cookie_t cookie = spl_fstrans_mark(); /* * Kick off asynchronous kmem_reap()'s of all our caches. */ arc_kmem_reap_soon(); /* * Wait at least arc_kmem_cache_reap_retry_ms between * arc_kmem_reap_soon() calls. Without this check it is possible to * end up in a situation where we spend lots of time reaping * caches, while we're near arc_c_min. Waiting here also gives the * subsequent free memory check a chance of finding that the * asynchronous reap has already freed enough memory, and we don't * need to call arc_reduce_target_size(). */ delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000); /* * Reduce the target size as needed to maintain the amount of free * memory in the system at a fraction of the arc_size (1/128th by * default). If oversubscribed (free_memory < 0) then reduce the * target arc_size by the deficit amount plus the fractional * amount. If free memory is positive but less than the fractional * amount, reduce by what is needed to hit the fractional amount. */ free_memory = arc_available_memory(); int64_t to_free = (arc_c >> arc_shrink_shift) - free_memory; if (to_free > 0) { arc_reduce_target_size(to_free); } spl_fstrans_unmark(cookie); } #ifdef _KERNEL /* * Determine the amount of memory eligible for eviction contained in the * ARC. All clean data reported by the ghost lists can always be safely * evicted. Due to arc_c_min, the same does not hold for all clean data * contained by the regular mru and mfu lists. * * In the case of the regular mru and mfu lists, we need to report as * much clean data as possible, such that evicting that same reported * data will not bring arc_size below arc_c_min. Thus, in certain * circumstances, the total amount of clean data in the mru and mfu * lists might not actually be evictable. * * The following two distinct cases are accounted for: * * 1. The sum of the amount of dirty data contained by both the mru and * mfu lists, plus the ARC's other accounting (e.g. the anon list), * is greater than or equal to arc_c_min. * (i.e. amount of dirty data >= arc_c_min) * * This is the easy case; all clean data contained by the mru and mfu * lists is evictable. Evicting all clean data can only drop arc_size * to the amount of dirty data, which is greater than arc_c_min. * * 2. The sum of the amount of dirty data contained by both the mru and * mfu lists, plus the ARC's other accounting (e.g. the anon list), * is less than arc_c_min. * (i.e. arc_c_min > amount of dirty data) * * 2.1. arc_size is greater than or equal arc_c_min. * (i.e. arc_size >= arc_c_min > amount of dirty data) * * In this case, not all clean data from the regular mru and mfu * lists is actually evictable; we must leave enough clean data * to keep arc_size above arc_c_min. Thus, the maximum amount of * evictable data from the two lists combined, is exactly the * difference between arc_size and arc_c_min. * * 2.2. arc_size is less than arc_c_min * (i.e. arc_c_min > arc_size > amount of dirty data) * * In this case, none of the data contained in the mru and mfu * lists is evictable, even if it's clean. Since arc_size is * already below arc_c_min, evicting any more would only * increase this negative difference. */ #endif /* _KERNEL */ /* * Adapt arc info given the number of bytes we are trying to add and * the state that we are coming from. This function is only called * when we are adding new content to the cache. */ static void arc_adapt(int bytes, arc_state_t *state) { int mult; uint64_t arc_p_min = (arc_c >> arc_p_min_shift); int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size); int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size); ASSERT(bytes > 0); /* * Adapt the target size of the MRU list: * - if we just hit in the MRU ghost list, then increase * the target size of the MRU list. * - if we just hit in the MFU ghost list, then increase * the target size of the MFU list by decreasing the * target size of the MRU list. */ if (state == arc_mru_ghost) { mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); if (!zfs_arc_p_dampener_disable) mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); } else if (state == arc_mfu_ghost) { uint64_t delta; mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); if (!zfs_arc_p_dampener_disable) mult = MIN(mult, 10); delta = MIN(bytes * mult, arc_p); arc_p = MAX(arc_p_min, arc_p - delta); } ASSERT((int64_t)arc_p >= 0); /* * Wake reap thread if we do not have any available memory */ if (arc_reclaim_needed()) { zthr_wakeup(arc_reap_zthr); return; } if (arc_no_grow) return; if (arc_c >= arc_c_max) return; /* * If we're within (2 * maxblocksize) bytes of the target * cache size, increment the target cache size */ ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT); if (aggsum_upper_bound(&arc_sums.arcstat_size) >= arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { atomic_add_64(&arc_c, (int64_t)bytes); if (arc_c > arc_c_max) arc_c = arc_c_max; else if (state == arc_anon) atomic_add_64(&arc_p, (int64_t)bytes); if (arc_p > arc_c) arc_p = arc_c; } ASSERT((int64_t)arc_p >= 0); } /* * Check if arc_size has grown past our upper threshold, determined by * zfs_arc_overflow_shift. */ static arc_ovf_level_t arc_is_overflowing(void) { /* Always allow at least one block of overflow */ int64_t overflow = MAX(SPA_MAXBLOCKSIZE, arc_c >> zfs_arc_overflow_shift); /* * We just compare the lower bound here for performance reasons. Our * primary goals are to make sure that the arc never grows without * bound, and that it can reach its maximum size. This check * accomplishes both goals. The maximum amount we could run over by is * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block * in the ARC. In practice, that's in the tens of MB, which is low * enough to be safe. */ int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) - arc_c - overflow / 2; return (over < 0 ? ARC_OVF_NONE : over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE); } static abd_t * arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag, boolean_t do_adapt) { arc_buf_contents_t type = arc_buf_type(hdr); arc_get_data_impl(hdr, size, tag, do_adapt); if (type == ARC_BUFC_METADATA) { return (abd_alloc(size, B_TRUE)); } else { ASSERT(type == ARC_BUFC_DATA); return (abd_alloc(size, B_FALSE)); } } static void * arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) { arc_buf_contents_t type = arc_buf_type(hdr); arc_get_data_impl(hdr, size, tag, B_TRUE); if (type == ARC_BUFC_METADATA) { return (zio_buf_alloc(size)); } else { ASSERT(type == ARC_BUFC_DATA); return (zio_data_buf_alloc(size)); } } /* * Wait for the specified amount of data (in bytes) to be evicted from the * ARC, and for there to be sufficient free memory in the system. Waiting for * eviction ensures that the memory used by the ARC decreases. Waiting for * free memory ensures that the system won't run out of free pages, regardless * of ARC behavior and settings. See arc_lowmem_init(). */ void arc_wait_for_eviction(uint64_t amount) { switch (arc_is_overflowing()) { case ARC_OVF_NONE: return; case ARC_OVF_SOME: /* * This is a bit racy without taking arc_evict_lock, but the * worst that can happen is we either call zthr_wakeup() extra * time due to race with other thread here, or the set flag * get cleared by arc_evict_cb(), which is unlikely due to * big hysteresis, but also not important since at this level * of overflow the eviction is purely advisory. Same time * taking the global lock here every time without waiting for * the actual eviction creates a significant lock contention. */ if (!arc_evict_needed) { arc_evict_needed = B_TRUE; zthr_wakeup(arc_evict_zthr); } return; case ARC_OVF_SEVERE: default: { arc_evict_waiter_t aw; list_link_init(&aw.aew_node); cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL); uint64_t last_count = 0; mutex_enter(&arc_evict_lock); if (!list_is_empty(&arc_evict_waiters)) { arc_evict_waiter_t *last = list_tail(&arc_evict_waiters); last_count = last->aew_count; } else if (!arc_evict_needed) { arc_evict_needed = B_TRUE; zthr_wakeup(arc_evict_zthr); } /* * Note, the last waiter's count may be less than * arc_evict_count if we are low on memory in which * case arc_evict_state_impl() may have deferred * wakeups (but still incremented arc_evict_count). */ aw.aew_count = MAX(last_count, arc_evict_count) + amount; list_insert_tail(&arc_evict_waiters, &aw); arc_set_need_free(); DTRACE_PROBE3(arc__wait__for__eviction, uint64_t, amount, uint64_t, arc_evict_count, uint64_t, aw.aew_count); /* * We will be woken up either when arc_evict_count reaches * aew_count, or when the ARC is no longer overflowing and * eviction completes. * In case of "false" wakeup, we will still be on the list. */ do { cv_wait(&aw.aew_cv, &arc_evict_lock); } while (list_link_active(&aw.aew_node)); mutex_exit(&arc_evict_lock); cv_destroy(&aw.aew_cv); } } } /* * Allocate a block and return it to the caller. If we are hitting the * hard limit for the cache size, we must sleep, waiting for the eviction * thread to catch up. If we're past the target size but below the hard * limit, we'll only signal the reclaim thread and continue on. */ static void arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag, boolean_t do_adapt) { arc_state_t *state = hdr->b_l1hdr.b_state; arc_buf_contents_t type = arc_buf_type(hdr); if (do_adapt) arc_adapt(size, state); /* * If arc_size is currently overflowing, we must be adding data * faster than we are evicting. To ensure we don't compound the * problem by adding more data and forcing arc_size to grow even * further past it's target size, we wait for the eviction thread to * make some progress. We also wait for there to be sufficient free * memory in the system, as measured by arc_free_memory(). * * Specifically, we wait for zfs_arc_eviction_pct percent of the * requested size to be evicted. This should be more than 100%, to * ensure that that progress is also made towards getting arc_size * under arc_c. See the comment above zfs_arc_eviction_pct. */ arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100); VERIFY3U(hdr->b_type, ==, type); if (type == ARC_BUFC_METADATA) { arc_space_consume(size, ARC_SPACE_META); } else { arc_space_consume(size, ARC_SPACE_DATA); } /* * Update the state size. Note that ghost states have a * "ghost size" and so don't need to be updated. */ if (!GHOST_STATE(state)) { (void) zfs_refcount_add_many(&state->arcs_size, size, tag); /* * If this is reached via arc_read, the link is * protected by the hash lock. If reached via * arc_buf_alloc, the header should not be accessed by * any other thread. And, if reached via arc_read_done, * the hash lock will protect it if it's found in the * hash table; otherwise no other thread should be * trying to [add|remove]_reference it. */ if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); (void) zfs_refcount_add_many(&state->arcs_esize[type], size, tag); } /* * If we are growing the cache, and we are adding anonymous * data, and we have outgrown arc_p, update arc_p */ if (aggsum_upper_bound(&arc_sums.arcstat_size) < arc_c && hdr->b_l1hdr.b_state == arc_anon && (zfs_refcount_count(&arc_anon->arcs_size) + zfs_refcount_count(&arc_mru->arcs_size) > arc_p)) arc_p = MIN(arc_c, arc_p + size); } } static void arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) { arc_free_data_impl(hdr, size, tag); abd_free(abd); } static void arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) { arc_buf_contents_t type = arc_buf_type(hdr); arc_free_data_impl(hdr, size, tag); if (type == ARC_BUFC_METADATA) { zio_buf_free(buf, size); } else { ASSERT(type == ARC_BUFC_DATA); zio_data_buf_free(buf, size); } } /* * Free the arc data buffer. */ static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) { arc_state_t *state = hdr->b_l1hdr.b_state; arc_buf_contents_t type = arc_buf_type(hdr); /* protected by hash lock, if in the hash table */ if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); ASSERT(state != arc_anon && state != arc_l2c_only); (void) zfs_refcount_remove_many(&state->arcs_esize[type], size, tag); } (void) zfs_refcount_remove_many(&state->arcs_size, size, tag); VERIFY3U(hdr->b_type, ==, type); if (type == ARC_BUFC_METADATA) { arc_space_return(size, ARC_SPACE_META); } else { ASSERT(type == ARC_BUFC_DATA); arc_space_return(size, ARC_SPACE_DATA); } } /* * This routine is called whenever a buffer is accessed. * NOTE: the hash lock is dropped in this function. */ static void arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) { clock_t now; ASSERT(MUTEX_HELD(hash_lock)); ASSERT(HDR_HAS_L1HDR(hdr)); if (hdr->b_l1hdr.b_state == arc_anon) { /* * This buffer is not in the cache, and does not * appear in our "ghost" list. Add the new buffer * to the MRU state. */ ASSERT0(hdr->b_l1hdr.b_arc_access); hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); arc_change_state(arc_mru, hdr, hash_lock); } else if (hdr->b_l1hdr.b_state == arc_mru) { now = ddi_get_lbolt(); /* * If this buffer is here because of a prefetch, then either: * - clear the flag if this is a "referencing" read * (any subsequent access will bump this into the MFU state). * or * - move the buffer to the head of the list if this is * another prefetch (to make it less likely to be evicted). */ if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { /* link protected by hash lock */ ASSERT(multilist_link_active( &hdr->b_l1hdr.b_arc_node)); } else { if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_decrement_state(hdr); arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH | ARC_FLAG_PRESCIENT_PREFETCH); atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); ARCSTAT_BUMP(arcstat_mru_hits); if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_increment_state(hdr); } hdr->b_l1hdr.b_arc_access = now; return; } /* * This buffer has been "accessed" only once so far, * but it is still in the cache. Move it to the MFU * state. */ if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access + ARC_MINTIME)) { /* * More than 125ms have passed since we * instantiated this buffer. Move it to the * most frequently used state. */ hdr->b_l1hdr.b_arc_access = now; DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); arc_change_state(arc_mfu, hdr, hash_lock); } atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); ARCSTAT_BUMP(arcstat_mru_hits); } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { arc_state_t *new_state; /* * This buffer has been "accessed" recently, but * was evicted from the cache. Move it to the * MFU state. */ if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { new_state = arc_mru; if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) { if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_decrement_state(hdr); arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH | ARC_FLAG_PRESCIENT_PREFETCH); if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_increment_state(hdr); } DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); } else { new_state = arc_mfu; DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); } hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); arc_change_state(new_state, hdr, hash_lock); atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits); ARCSTAT_BUMP(arcstat_mru_ghost_hits); } else if (hdr->b_l1hdr.b_state == arc_mfu) { /* * This buffer has been accessed more than once and is * still in the cache. Keep it in the MFU state. * * NOTE: an add_reference() that occurred when we did * the arc_read() will have kicked this off the list. * If it was a prefetch, we will explicitly move it to * the head of the list now. */ atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits); ARCSTAT_BUMP(arcstat_mfu_hits); hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { arc_state_t *new_state = arc_mfu; /* * This buffer has been accessed more than once but has * been evicted from the cache. Move it back to the * MFU state. */ if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { /* * This is a prefetch access... * move this block back to the MRU state. */ new_state = arc_mru; } hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); arc_change_state(new_state, hdr, hash_lock); atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits); ARCSTAT_BUMP(arcstat_mfu_ghost_hits); } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { /* * This buffer is on the 2nd Level ARC. */ hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); arc_change_state(arc_mfu, hdr, hash_lock); } else { cmn_err(CE_PANIC, "invalid arc state 0x%p", hdr->b_l1hdr.b_state); } } /* * This routine is called by dbuf_hold() to update the arc_access() state * which otherwise would be skipped for entries in the dbuf cache. */ void arc_buf_access(arc_buf_t *buf) { mutex_enter(&buf->b_evict_lock); arc_buf_hdr_t *hdr = buf->b_hdr; /* * Avoid taking the hash_lock when possible as an optimization. * The header must be checked again under the hash_lock in order * to handle the case where it is concurrently being released. */ if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { mutex_exit(&buf->b_evict_lock); return; } kmutex_t *hash_lock = HDR_LOCK(hdr); mutex_enter(hash_lock); if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { mutex_exit(hash_lock); mutex_exit(&buf->b_evict_lock); ARCSTAT_BUMP(arcstat_access_skip); return; } mutex_exit(&buf->b_evict_lock); ASSERT(hdr->b_l1hdr.b_state == arc_mru || hdr->b_l1hdr.b_state == arc_mfu); DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); arc_access(hdr, hash_lock); mutex_exit(hash_lock); ARCSTAT_BUMP(arcstat_hits); ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr), demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); } /* a generic arc_read_done_func_t which you can use */ /* ARGSUSED */ void arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, arc_buf_t *buf, void *arg) { if (buf == NULL) return; bcopy(buf->b_data, arg, arc_buf_size(buf)); arc_buf_destroy(buf, arg); } /* a generic arc_read_done_func_t */ /* ARGSUSED */ void arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, arc_buf_t *buf, void *arg) { arc_buf_t **bufp = arg; if (buf == NULL) { ASSERT(zio == NULL || zio->io_error != 0); *bufp = NULL; } else { ASSERT(zio == NULL || zio->io_error == 0); *bufp = buf; ASSERT(buf->b_data != NULL); } } static void arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) { if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF); } else { if (HDR_COMPRESSION_ENABLED(hdr)) { ASSERT3U(arc_hdr_get_compress(hdr), ==, BP_GET_COMPRESS(bp)); } ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp)); } } static void arc_read_done(zio_t *zio) { blkptr_t *bp = zio->io_bp; arc_buf_hdr_t *hdr = zio->io_private; kmutex_t *hash_lock = NULL; arc_callback_t *callback_list; arc_callback_t *acb; boolean_t freeable = B_FALSE; /* * The hdr was inserted into hash-table and removed from lists * prior to starting I/O. We should find this header, since * it's in the hash table, and it should be legit since it's * not possible to evict it during the I/O. The only possible * reason for it not to be found is if we were freed during the * read. */ if (HDR_IN_HASH_TABLE(hdr)) { arc_buf_hdr_t *found; ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); ASSERT3U(hdr->b_dva.dva_word[0], ==, BP_IDENTITY(zio->io_bp)->dva_word[0]); ASSERT3U(hdr->b_dva.dva_word[1], ==, BP_IDENTITY(zio->io_bp)->dva_word[1]); found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock); ASSERT((found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || (found == hdr && HDR_L2_READING(hdr))); ASSERT3P(hash_lock, !=, NULL); } if (BP_IS_PROTECTED(bp)) { hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp); hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset; zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv); if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) { void *tmpbuf; tmpbuf = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); zio_crypt_decode_mac_zil(tmpbuf, hdr->b_crypt_hdr.b_mac); abd_return_buf(zio->io_abd, tmpbuf, sizeof (zil_chain_t)); } else { zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac); } } if (zio->io_error == 0) { /* byteswap if necessary */ if (BP_SHOULD_BYTESWAP(zio->io_bp)) { if (BP_GET_LEVEL(zio->io_bp) > 0) { hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; } else { hdr->b_l1hdr.b_byteswap = DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); } } else { hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; } if (!HDR_L2_READING(hdr)) { hdr->b_complevel = zio->io_prop.zp_complevel; } } arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); if (l2arc_noprefetch && HDR_PREFETCH(hdr)) arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); callback_list = hdr->b_l1hdr.b_acb; ASSERT3P(callback_list, !=, NULL); if (hash_lock && zio->io_error == 0 && hdr->b_l1hdr.b_state == arc_anon) { /* * Only call arc_access on anonymous buffers. This is because * if we've issued an I/O for an evicted buffer, we've already * called arc_access (to prevent any simultaneous readers from * getting confused). */ arc_access(hdr, hash_lock); } /* * If a read request has a callback (i.e. acb_done is not NULL), then we * make a buf containing the data according to the parameters which were * passed in. The implementation of arc_buf_alloc_impl() ensures that we * aren't needlessly decompressing the data multiple times. */ int callback_cnt = 0; for (acb = callback_list; acb != NULL; acb = acb->acb_next) { if (!acb->acb_done || acb->acb_nobuf) continue; callback_cnt++; if (zio->io_error != 0) continue; int error = arc_buf_alloc_impl(hdr, zio->io_spa, &acb->acb_zb, acb->acb_private, acb->acb_encrypted, acb->acb_compressed, acb->acb_noauth, B_TRUE, &acb->acb_buf); /* * Assert non-speculative zios didn't fail because an * encryption key wasn't loaded */ ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) || error != EACCES); /* * If we failed to decrypt, report an error now (as the zio * layer would have done if it had done the transforms). */ if (error == ECKSUM) { ASSERT(BP_IS_PROTECTED(bp)); error = SET_ERROR(EIO); if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { spa_log_error(zio->io_spa, &acb->acb_zb); (void) zfs_ereport_post( FM_EREPORT_ZFS_AUTHENTICATION, zio->io_spa, NULL, &acb->acb_zb, zio, 0); } } if (error != 0) { /* * Decompression or decryption failed. Set * io_error so that when we call acb_done * (below), we will indicate that the read * failed. Note that in the unusual case * where one callback is compressed and another * uncompressed, we will mark all of them * as failed, even though the uncompressed * one can't actually fail. In this case, * the hdr will not be anonymous, because * if there are multiple callbacks, it's * because multiple threads found the same * arc buf in the hash table. */ zio->io_error = error; } } /* * If there are multiple callbacks, we must have the hash lock, * because the only way for multiple threads to find this hdr is * in the hash table. This ensures that if there are multiple * callbacks, the hdr is not anonymous. If it were anonymous, * we couldn't use arc_buf_destroy() in the error case below. */ ASSERT(callback_cnt < 2 || hash_lock != NULL); hdr->b_l1hdr.b_acb = NULL; arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); if (callback_cnt == 0) ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || callback_list != NULL); if (zio->io_error == 0) { arc_hdr_verify(hdr, zio->io_bp); } else { arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); if (hdr->b_l1hdr.b_state != arc_anon) arc_change_state(arc_anon, hdr, hash_lock); if (HDR_IN_HASH_TABLE(hdr)) buf_hash_remove(hdr); freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); } /* * Broadcast before we drop the hash_lock to avoid the possibility * that the hdr (and hence the cv) might be freed before we get to * the cv_broadcast(). */ cv_broadcast(&hdr->b_l1hdr.b_cv); if (hash_lock != NULL) { mutex_exit(hash_lock); } else { /* * This block was freed while we waited for the read to * complete. It has been removed from the hash table and * moved to the anonymous state (so that it won't show up * in the cache). */ ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); } /* execute each callback and free its structure */ while ((acb = callback_list) != NULL) { if (acb->acb_done != NULL) { if (zio->io_error != 0 && acb->acb_buf != NULL) { /* * If arc_buf_alloc_impl() fails during * decompression, the buf will still be * allocated, and needs to be freed here. */ arc_buf_destroy(acb->acb_buf, acb->acb_private); acb->acb_buf = NULL; } acb->acb_done(zio, &zio->io_bookmark, zio->io_bp, acb->acb_buf, acb->acb_private); } if (acb->acb_zio_dummy != NULL) { acb->acb_zio_dummy->io_error = zio->io_error; zio_nowait(acb->acb_zio_dummy); } callback_list = acb->acb_next; kmem_free(acb, sizeof (arc_callback_t)); } if (freeable) arc_hdr_destroy(hdr); } /* * "Read" the block at the specified DVA (in bp) via the * cache. If the block is found in the cache, invoke the provided * callback immediately and return. Note that the `zio' parameter * in the callback will be NULL in this case, since no IO was * required. If the block is not in the cache pass the read request * on to the spa with a substitute callback function, so that the * requested block will be added to the cache. * * If a read request arrives for a block that has a read in-progress, * either wait for the in-progress read to complete (and return the * results); or, if this is a read with a "done" func, add a record * to the read to invoke the "done" func when the read completes, * and return; or just return. * * arc_read_done() will invoke all the requested "done" functions * for readers of this block. */ int arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done, void *private, zio_priority_t priority, int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb) { arc_buf_hdr_t *hdr = NULL; kmutex_t *hash_lock = NULL; zio_t *rzio; uint64_t guid = spa_load_guid(spa); boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0; boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) && (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0; boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) && (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0; boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp); boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF; int rc = 0; ASSERT(!embedded_bp || BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); ASSERT(!BP_IS_HOLE(bp)); ASSERT(!BP_IS_REDACTED(bp)); /* * Normally SPL_FSTRANS will already be set since kernel threads which * expect to call the DMU interfaces will set it when created. System * calls are similarly handled by setting/cleaning the bit in the * registered callback (module/os/.../zfs/zpl_*). * * External consumers such as Lustre which call the exported DMU * interfaces may not have set SPL_FSTRANS. To avoid a deadlock * on the hash_lock always set and clear the bit. */ fstrans_cookie_t cookie = spl_fstrans_mark(); top: if (!embedded_bp) { /* * Embedded BP's have no DVA and require no I/O to "read". * Create an anonymous arc buf to back it. */ if (!zfs_blkptr_verify(spa, bp, zio_flags & ZIO_FLAG_CONFIG_WRITER, BLK_VERIFY_LOG)) { rc = SET_ERROR(ECKSUM); goto out; } hdr = buf_hash_find(guid, bp, &hash_lock); } /* * Determine if we have an L1 cache hit or a cache miss. For simplicity * we maintain encrypted data separately from compressed / uncompressed * data. If the user is requesting raw encrypted data and we don't have * that in the header we will read from disk to guarantee that we can * get it even if the encryption keys aren't loaded. */ if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) || (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) { arc_buf_t *buf = NULL; *arc_flags |= ARC_FLAG_CACHED; if (HDR_IO_IN_PROGRESS(hdr)) { zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head; if (*arc_flags & ARC_FLAG_CACHED_ONLY) { mutex_exit(hash_lock); ARCSTAT_BUMP(arcstat_cached_only_in_progress); rc = SET_ERROR(ENOENT); goto out; } ASSERT3P(head_zio, !=, NULL); if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && priority == ZIO_PRIORITY_SYNC_READ) { /* * This is a sync read that needs to wait for * an in-flight async read. Request that the * zio have its priority upgraded. */ zio_change_priority(head_zio, priority); DTRACE_PROBE1(arc__async__upgrade__sync, arc_buf_hdr_t *, hdr); ARCSTAT_BUMP(arcstat_async_upgrade_sync); } if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { arc_hdr_clear_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); } if (*arc_flags & ARC_FLAG_WAIT) { cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); mutex_exit(hash_lock); goto top; } ASSERT(*arc_flags & ARC_FLAG_NOWAIT); if (done) { arc_callback_t *acb = NULL; acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); acb->acb_done = done; acb->acb_private = private; acb->acb_compressed = compressed_read; acb->acb_encrypted = encrypted_read; acb->acb_noauth = noauth_read; acb->acb_nobuf = no_buf; acb->acb_zb = *zb; if (pio != NULL) acb->acb_zio_dummy = zio_null(pio, spa, NULL, NULL, NULL, zio_flags); ASSERT3P(acb->acb_done, !=, NULL); acb->acb_zio_head = head_zio; acb->acb_next = hdr->b_l1hdr.b_acb; hdr->b_l1hdr.b_acb = acb; } mutex_exit(hash_lock); goto out; } ASSERT(hdr->b_l1hdr.b_state == arc_mru || hdr->b_l1hdr.b_state == arc_mfu); if (done && !no_buf) { if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { /* * This is a demand read which does not have to * wait for i/o because we did a predictive * prefetch i/o for it, which has completed. */ DTRACE_PROBE1( arc__demand__hit__predictive__prefetch, arc_buf_hdr_t *, hdr); ARCSTAT_BUMP( arcstat_demand_hit_predictive_prefetch); arc_hdr_clear_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); } if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) { ARCSTAT_BUMP( arcstat_demand_hit_prescient_prefetch); arc_hdr_clear_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); } ASSERT(!embedded_bp || !BP_IS_HOLE(bp)); /* Get a buf with the desired data in it. */ rc = arc_buf_alloc_impl(hdr, spa, zb, private, encrypted_read, compressed_read, noauth_read, B_TRUE, &buf); if (rc == ECKSUM) { /* * Convert authentication and decryption errors * to EIO (and generate an ereport if needed) * before leaving the ARC. */ rc = SET_ERROR(EIO); if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) { spa_log_error(spa, zb); (void) zfs_ereport_post( FM_EREPORT_ZFS_AUTHENTICATION, spa, NULL, zb, NULL, 0); } } if (rc != 0) { (void) remove_reference(hdr, hash_lock, private); arc_buf_destroy_impl(buf); buf = NULL; } /* assert any errors weren't due to unloaded keys */ ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) || rc != EACCES); } else if (*arc_flags & ARC_FLAG_PREFETCH && zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_decrement_state(hdr); arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_increment_state(hdr); } DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); arc_access(hdr, hash_lock); if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); if (*arc_flags & ARC_FLAG_L2CACHE) arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); mutex_exit(hash_lock); ARCSTAT_BUMP(arcstat_hits); ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); if (done) done(NULL, zb, bp, buf, private); } else { uint64_t lsize = BP_GET_LSIZE(bp); uint64_t psize = BP_GET_PSIZE(bp); arc_callback_t *acb; vdev_t *vd = NULL; uint64_t addr = 0; boolean_t devw = B_FALSE; uint64_t size; abd_t *hdr_abd; int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0; if (*arc_flags & ARC_FLAG_CACHED_ONLY) { rc = SET_ERROR(ENOENT); if (hash_lock != NULL) mutex_exit(hash_lock); goto out; } if (hdr == NULL) { /* * This block is not in the cache or it has * embedded data. */ arc_buf_hdr_t *exists = NULL; arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type, encrypted_read); if (!embedded_bp) { hdr->b_dva = *BP_IDENTITY(bp); hdr->b_birth = BP_PHYSICAL_BIRTH(bp); exists = buf_hash_insert(hdr, &hash_lock); } if (exists != NULL) { /* somebody beat us to the hash insert */ mutex_exit(hash_lock); buf_discard_identity(hdr); arc_hdr_destroy(hdr); goto top; /* restart the IO request */ } } else { /* * This block is in the ghost cache or encrypted data * was requested and we didn't have it. If it was * L2-only (and thus didn't have an L1 hdr), * we realloc the header to add an L1 hdr. */ if (!HDR_HAS_L1HDR(hdr)) { hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, hdr_full_cache); } if (GHOST_STATE(hdr->b_l1hdr.b_state)) { ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); ASSERT(!HDR_IO_IN_PROGRESS(hdr)); ASSERT0(zfs_refcount_count( &hdr->b_l1hdr.b_refcnt)); ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); } else if (HDR_IO_IN_PROGRESS(hdr)) { /* * If this header already had an IO in progress * and we are performing another IO to fetch * encrypted data we must wait until the first * IO completes so as not to confuse * arc_read_done(). This should be very rare * and so the performance impact shouldn't * matter. */ cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); mutex_exit(hash_lock); goto top; } /* * This is a delicate dance that we play here. * This hdr might be in the ghost list so we access * it to move it out of the ghost list before we * initiate the read. If it's a prefetch then * it won't have a callback so we'll remove the * reference that arc_buf_alloc_impl() created. We * do this after we've called arc_access() to * avoid hitting an assert in remove_reference(). */ arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state); arc_access(hdr, hash_lock); arc_hdr_alloc_abd(hdr, alloc_flags); } if (encrypted_read) { ASSERT(HDR_HAS_RABD(hdr)); size = HDR_GET_PSIZE(hdr); hdr_abd = hdr->b_crypt_hdr.b_rabd; zio_flags |= ZIO_FLAG_RAW; } else { ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); size = arc_hdr_size(hdr); hdr_abd = hdr->b_l1hdr.b_pabd; if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) { zio_flags |= ZIO_FLAG_RAW_COMPRESS; } /* * For authenticated bp's, we do not ask the ZIO layer * to authenticate them since this will cause the entire * IO to fail if the key isn't loaded. Instead, we * defer authentication until arc_buf_fill(), which will * verify the data when the key is available. */ if (BP_IS_AUTHENTICATED(bp)) zio_flags |= ZIO_FLAG_RAW_ENCRYPT; } if (*arc_flags & ARC_FLAG_PREFETCH && zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_decrement_state(hdr); arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); if (HDR_HAS_L2HDR(hdr)) l2arc_hdr_arcstats_increment_state(hdr); } if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); if (*arc_flags & ARC_FLAG_L2CACHE) arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); if (BP_IS_AUTHENTICATED(bp)) arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH); if (BP_GET_LEVEL(bp) > 0) arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); acb->acb_done = done; acb->acb_private = private; acb->acb_compressed = compressed_read; acb->acb_encrypted = encrypted_read; acb->acb_noauth = noauth_read; acb->acb_zb = *zb; ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); hdr->b_l1hdr.b_acb = acb; arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); if (HDR_HAS_L2HDR(hdr) && (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { devw = hdr->b_l2hdr.b_dev->l2ad_writing; addr = hdr->b_l2hdr.b_daddr; /* * Lock out L2ARC device removal. */ if (vdev_is_dead(vd) || !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) vd = NULL; } /* * We count both async reads and scrub IOs as asynchronous so * that both can be upgraded in the event of a cache hit while * the read IO is still in-flight. */ if (priority == ZIO_PRIORITY_ASYNC_READ || priority == ZIO_PRIORITY_SCRUB) arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); else arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); /* * At this point, we have a level 1 cache miss or a blkptr * with embedded data. Try again in L2ARC if possible. */ ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); /* * Skip ARC stat bump for block pointers with embedded * data. The data are read from the blkptr itself via * decode_embedded_bp_compressed(). */ if (!embedded_bp) { DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, uint64_t, lsize, zbookmark_phys_t *, zb); ARCSTAT_BUMP(arcstat_misses); ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, misses); zfs_racct_read(size, 1); } /* Check if the spa even has l2 configured */ const boolean_t spa_has_l2 = l2arc_ndev != 0 && spa->spa_l2cache.sav_count > 0; if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) { /* * Read from the L2ARC if the following are true: * 1. The L2ARC vdev was previously cached. * 2. This buffer still has L2ARC metadata. * 3. This buffer isn't currently writing to the L2ARC. * 4. The L2ARC entry wasn't evicted, which may * also have invalidated the vdev. * 5. This isn't prefetch or l2arc_noprefetch is 0. */ if (HDR_HAS_L2HDR(hdr) && !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { l2arc_read_callback_t *cb; abd_t *abd; uint64_t asize; DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); ARCSTAT_BUMP(arcstat_l2_hits); atomic_inc_32(&hdr->b_l2hdr.b_hits); cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP); cb->l2rcb_hdr = hdr; cb->l2rcb_bp = *bp; cb->l2rcb_zb = *zb; cb->l2rcb_flags = zio_flags; /* * When Compressed ARC is disabled, but the * L2ARC block is compressed, arc_hdr_size() * will have returned LSIZE rather than PSIZE. */ if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr) && HDR_GET_PSIZE(hdr) != 0) { size = HDR_GET_PSIZE(hdr); } asize = vdev_psize_to_asize(vd, size); if (asize != size) { abd = abd_alloc_for_io(asize, HDR_ISTYPE_METADATA(hdr)); cb->l2rcb_abd = abd; } else { abd = hdr_abd; } ASSERT(addr >= VDEV_LABEL_START_SIZE && addr + asize <= vd->vdev_psize - VDEV_LABEL_END_SIZE); /* * l2arc read. The SCL_L2ARC lock will be * released by l2arc_read_done(). * Issue a null zio if the underlying buffer * was squashed to zero size by compression. */ ASSERT3U(arc_hdr_get_compress(hdr), !=, ZIO_COMPRESS_EMPTY); rzio = zio_read_phys(pio, vd, addr, asize, abd, ZIO_CHECKSUM_OFF, l2arc_read_done, cb, priority, zio_flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE); acb->acb_zio_head = rzio; if (hash_lock != NULL) mutex_exit(hash_lock); DTRACE_PROBE2(l2arc__read, vdev_t *, vd, zio_t *, rzio); ARCSTAT_INCR(arcstat_l2_read_bytes, HDR_GET_PSIZE(hdr)); if (*arc_flags & ARC_FLAG_NOWAIT) { zio_nowait(rzio); goto out; } ASSERT(*arc_flags & ARC_FLAG_WAIT); if (zio_wait(rzio) == 0) goto out; /* l2arc read error; goto zio_read() */ if (hash_lock != NULL) mutex_enter(hash_lock); } else { DTRACE_PROBE1(l2arc__miss, arc_buf_hdr_t *, hdr); ARCSTAT_BUMP(arcstat_l2_misses); if (HDR_L2_WRITING(hdr)) ARCSTAT_BUMP(arcstat_l2_rw_clash); spa_config_exit(spa, SCL_L2ARC, vd); } } else { if (vd != NULL) spa_config_exit(spa, SCL_L2ARC, vd); /* * Only a spa with l2 should contribute to l2 * miss stats. (Including the case of having a * faulted cache device - that's also a miss.) */ if (spa_has_l2) { /* * Skip ARC stat bump for block pointers with * embedded data. The data are read from the * blkptr itself via * decode_embedded_bp_compressed(). */ if (!embedded_bp) { DTRACE_PROBE1(l2arc__miss, arc_buf_hdr_t *, hdr); ARCSTAT_BUMP(arcstat_l2_misses); } } } rzio = zio_read(pio, spa, bp, hdr_abd, size, arc_read_done, hdr, priority, zio_flags, zb); acb->acb_zio_head = rzio; if (hash_lock != NULL) mutex_exit(hash_lock); if (*arc_flags & ARC_FLAG_WAIT) { rc = zio_wait(rzio); goto out; } ASSERT(*arc_flags & ARC_FLAG_NOWAIT); zio_nowait(rzio); } out: /* embedded bps don't actually go to disk */ if (!embedded_bp) spa_read_history_add(spa, zb, *arc_flags); spl_fstrans_unmark(cookie); return (rc); } arc_prune_t * arc_add_prune_callback(arc_prune_func_t *func, void *private) { arc_prune_t *p; p = kmem_alloc(sizeof (*p), KM_SLEEP); p->p_pfunc = func; p->p_private = private; list_link_init(&p->p_node); zfs_refcount_create(&p->p_refcnt); mutex_enter(&arc_prune_mtx); zfs_refcount_add(&p->p_refcnt, &arc_prune_list); list_insert_head(&arc_prune_list, p); mutex_exit(&arc_prune_mtx); return (p); } void arc_remove_prune_callback(arc_prune_t *p) { boolean_t wait = B_FALSE; mutex_enter(&arc_prune_mtx); list_remove(&arc_prune_list, p); if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0) wait = B_TRUE; mutex_exit(&arc_prune_mtx); /* wait for arc_prune_task to finish */ if (wait) taskq_wait_outstanding(arc_prune_taskq, 0); ASSERT0(zfs_refcount_count(&p->p_refcnt)); zfs_refcount_destroy(&p->p_refcnt); kmem_free(p, sizeof (*p)); } /* * Notify the arc that a block was freed, and thus will never be used again. */ void arc_freed(spa_t *spa, const blkptr_t *bp) { arc_buf_hdr_t *hdr; kmutex_t *hash_lock; uint64_t guid = spa_load_guid(spa); ASSERT(!BP_IS_EMBEDDED(bp)); hdr = buf_hash_find(guid, bp, &hash_lock); if (hdr == NULL) return; /* * We might be trying to free a block that is still doing I/O * (i.e. prefetch) or has a reference (i.e. a dedup-ed, * dmu_sync-ed block). If this block is being prefetched, then it * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr * until the I/O completes. A block may also have a reference if it is * part of a dedup-ed, dmu_synced write. The dmu_sync() function would * have written the new block to its final resting place on disk but * without the dedup flag set. This would have left the hdr in the MRU * state and discoverable. When the txg finally syncs it detects that * the block was overridden in open context and issues an override I/O. * Since this is a dedup block, the override I/O will determine if the * block is already in the DDT. If so, then it will replace the io_bp * with the bp from the DDT and allow the I/O to finish. When the I/O * reaches the done callback, dbuf_write_override_done, it will * check to see if the io_bp and io_bp_override are identical. * If they are not, then it indicates that the bp was replaced with * the bp in the DDT and the override bp is freed. This allows * us to arrive here with a reference on a block that is being * freed. So if we have an I/O in progress, or a reference to * this hdr, then we don't destroy the hdr. */ if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { arc_change_state(arc_anon, hdr, hash_lock); arc_hdr_destroy(hdr); mutex_exit(hash_lock); } else { mutex_exit(hash_lock); } } /* * Release this buffer from the cache, making it an anonymous buffer. This * must be done after a read and prior to modifying the buffer contents. * If the buffer has more than one reference, we must make * a new hdr for the buffer. */ void arc_release(arc_buf_t *buf, void *tag) { arc_buf_hdr_t *hdr = buf->b_hdr; /* * It would be nice to assert that if its DMU metadata (level > * 0 || it's the dnode file), then it must be syncing context. * But we don't know that information at this level. */ mutex_enter(&buf->b_evict_lock); ASSERT(HDR_HAS_L1HDR(hdr)); /* * We don't grab the hash lock prior to this check, because if * the buffer's header is in the arc_anon state, it won't be * linked into the hash table. */ if (hdr->b_l1hdr.b_state == arc_anon) { mutex_exit(&buf->b_evict_lock); ASSERT(!HDR_IO_IN_PROGRESS(hdr)); ASSERT(!HDR_IN_HASH_TABLE(hdr)); ASSERT(!HDR_HAS_L2HDR(hdr)); ASSERT(HDR_EMPTY(hdr)); ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); hdr->b_l1hdr.b_arc_access = 0; /* * If the buf is being overridden then it may already * have a hdr that is not empty. */ buf_discard_identity(hdr); arc_buf_thaw(buf); return; } kmutex_t *hash_lock = HDR_LOCK(hdr); mutex_enter(hash_lock); /* * This assignment is only valid as long as the hash_lock is * held, we must be careful not to reference state or the * b_state field after dropping the lock. */ arc_state_t *state = hdr->b_l1hdr.b_state; ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); ASSERT3P(state, !=, arc_anon); /* this buffer is not on any list */ ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); if (HDR_HAS_L2HDR(hdr)) { mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); /* * We have to recheck this conditional again now that * we're holding the l2ad_mtx to prevent a race with * another thread which might be concurrently calling * l2arc_evict(). In that case, l2arc_evict() might have * destroyed the header's L2 portion as we were waiting * to acquire the l2ad_mtx. */ if (HDR_HAS_L2HDR(hdr)) arc_hdr_l2hdr_destroy(hdr); mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); } /* * Do we have more than one buf? */ if (hdr->b_l1hdr.b_bufcnt > 1) { arc_buf_hdr_t *nhdr; uint64_t spa = hdr->b_spa; uint64_t psize = HDR_GET_PSIZE(hdr); uint64_t lsize = HDR_GET_LSIZE(hdr); boolean_t protected = HDR_PROTECTED(hdr); enum zio_compress compress = arc_hdr_get_compress(hdr); arc_buf_contents_t type = arc_buf_type(hdr); VERIFY3U(hdr->b_type, ==, type); ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); (void) remove_reference(hdr, hash_lock, tag); if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); ASSERT(ARC_BUF_LAST(buf)); } /* * Pull the data off of this hdr and attach it to * a new anonymous hdr. Also find the last buffer * in the hdr's buffer list. */ arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); ASSERT3P(lastbuf, !=, NULL); /* * If the current arc_buf_t and the hdr are sharing their data * buffer, then we must stop sharing that block. */ if (arc_buf_is_shared(buf)) { ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); VERIFY(!arc_buf_is_shared(lastbuf)); /* * First, sever the block sharing relationship between * buf and the arc_buf_hdr_t. */ arc_unshare_buf(hdr, buf); /* * Now we need to recreate the hdr's b_pabd. Since we * have lastbuf handy, we try to share with it, but if * we can't then we allocate a new b_pabd and copy the * data from buf into it. */ if (arc_can_share(hdr, lastbuf)) { arc_share_buf(hdr, lastbuf); } else { arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT); abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, psize); } VERIFY3P(lastbuf->b_data, !=, NULL); } else if (HDR_SHARED_DATA(hdr)) { /* * Uncompressed shared buffers are always at the end * of the list. Compressed buffers don't have the * same requirements. This makes it hard to * simply assert that the lastbuf is shared so * we rely on the hdr's compression flags to determine * if we have a compressed, shared buffer. */ ASSERT(arc_buf_is_shared(lastbuf) || arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); ASSERT(!ARC_BUF_SHARED(buf)); } ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); ASSERT3P(state, !=, arc_l2c_only); (void) zfs_refcount_remove_many(&state->arcs_size, arc_buf_size(buf), buf); if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { ASSERT3P(state, !=, arc_l2c_only); (void) zfs_refcount_remove_many( &state->arcs_esize[type], arc_buf_size(buf), buf); } hdr->b_l1hdr.b_bufcnt -= 1; if (ARC_BUF_ENCRYPTED(buf)) hdr->b_crypt_hdr.b_ebufcnt -= 1; arc_cksum_verify(buf); arc_buf_unwatch(buf); /* if this is the last uncompressed buf free the checksum */ if (!arc_hdr_has_uncompressed_buf(hdr)) arc_cksum_free(hdr); mutex_exit(hash_lock); /* * Allocate a new hdr. The new hdr will contain a b_pabd * buffer which will be freed in arc_write(). */ nhdr = arc_hdr_alloc(spa, psize, lsize, protected, compress, hdr->b_complevel, type, HDR_HAS_RABD(hdr)); ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); ASSERT0(nhdr->b_l1hdr.b_bufcnt); ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt)); VERIFY3U(nhdr->b_type, ==, type); ASSERT(!HDR_SHARED_DATA(nhdr)); nhdr->b_l1hdr.b_buf = buf; nhdr->b_l1hdr.b_bufcnt = 1; if (ARC_BUF_ENCRYPTED(buf)) nhdr->b_crypt_hdr.b_ebufcnt = 1; nhdr->b_l1hdr.b_mru_hits = 0; nhdr->b_l1hdr.b_mru_ghost_hits = 0; nhdr->b_l1hdr.b_mfu_hits = 0; nhdr->b_l1hdr.b_mfu_ghost_hits = 0; nhdr->b_l1hdr.b_l2_hits = 0; (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); buf->b_hdr = nhdr; mutex_exit(&buf->b_evict_lock); (void) zfs_refcount_add_many(&arc_anon->arcs_size, arc_buf_size(buf), buf); } else { mutex_exit(&buf->b_evict_lock); ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); /* protected by hash lock, or hdr is on arc_anon */ ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); ASSERT(!HDR_IO_IN_PROGRESS(hdr)); hdr->b_l1hdr.b_mru_hits = 0; hdr->b_l1hdr.b_mru_ghost_hits = 0; hdr->b_l1hdr.b_mfu_hits = 0; hdr->b_l1hdr.b_mfu_ghost_hits = 0; hdr->b_l1hdr.b_l2_hits = 0; arc_change_state(arc_anon, hdr, hash_lock); hdr->b_l1hdr.b_arc_access = 0; mutex_exit(hash_lock); buf_discard_identity(hdr); arc_buf_thaw(buf); } } int arc_released(arc_buf_t *buf) { int released; mutex_enter(&buf->b_evict_lock); released = (buf->b_data != NULL && buf->b_hdr->b_l1hdr.b_state == arc_anon); mutex_exit(&buf->b_evict_lock); return (released); } #ifdef ZFS_DEBUG int arc_referenced(arc_buf_t *buf) { int referenced; mutex_enter(&buf->b_evict_lock); referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); mutex_exit(&buf->b_evict_lock); return (referenced); } #endif static void arc_write_ready(zio_t *zio) { arc_write_callback_t *callback = zio->io_private; arc_buf_t *buf = callback->awcb_buf; arc_buf_hdr_t *hdr = buf->b_hdr; blkptr_t *bp = zio->io_bp; uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp); fstrans_cookie_t cookie = spl_fstrans_mark(); ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); ASSERT(hdr->b_l1hdr.b_bufcnt > 0); /* * If we're reexecuting this zio because the pool suspended, then * cleanup any state that was previously set the first time the * callback was invoked. */ if (zio->io_flags & ZIO_FLAG_REEXECUTED) { arc_cksum_free(hdr); arc_buf_unwatch(buf); if (hdr->b_l1hdr.b_pabd != NULL) { if (arc_buf_is_shared(buf)) { arc_unshare_buf(hdr, buf); } else { arc_hdr_free_abd(hdr, B_FALSE); } } if (HDR_HAS_RABD(hdr)) arc_hdr_free_abd(hdr, B_TRUE); } ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); ASSERT(!HDR_HAS_RABD(hdr)); ASSERT(!HDR_SHARED_DATA(hdr)); ASSERT(!arc_buf_is_shared(buf)); callback->awcb_ready(zio, buf, callback->awcb_private); if (HDR_IO_IN_PROGRESS(hdr)) ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr)) hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp)); if (BP_IS_PROTECTED(bp)) { /* ZIL blocks are written through zio_rewrite */ ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG); ASSERT(HDR_PROTECTED(hdr)); if (BP_SHOULD_BYTESWAP(bp)) { if (BP_GET_LEVEL(bp) > 0) { hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; } else { hdr->b_l1hdr.b_byteswap = DMU_OT_BYTESWAP(BP_GET_TYPE(bp)); } } else { hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; } hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp); hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset; zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv); zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac); } /* * If this block was written for raw encryption but the zio layer * ended up only authenticating it, adjust the buffer flags now. */ if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) { arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH); buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF) buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) { buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; } /* this must be done after the buffer flags are adjusted */ arc_cksum_compute(buf); enum zio_compress compress; if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { compress = ZIO_COMPRESS_OFF; } else { ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); compress = BP_GET_COMPRESS(bp); } HDR_SET_PSIZE(hdr, psize); arc_hdr_set_compress(hdr, compress); hdr->b_complevel = zio->io_prop.zp_complevel; if (zio->io_error != 0 || psize == 0) goto out; /* * Fill the hdr with data. If the buffer is encrypted we have no choice * but to copy the data into b_radb. If the hdr is compressed, the data * we want is available from the zio, otherwise we can take it from * the buf. * * We might be able to share the buf's data with the hdr here. However, * doing so would cause the ARC to be full of linear ABDs if we write a * lot of shareable data. As a compromise, we check whether scattered * ABDs are allowed, and assume that if they are then the user wants * the ARC to be primarily filled with them regardless of the data being * written. Therefore, if they're allowed then we allocate one and copy * the data into it; otherwise, we share the data directly if we can. */ if (ARC_BUF_ENCRYPTED(buf)) { ASSERT3U(psize, >, 0); ASSERT(ARC_BUF_COMPRESSED(buf)); arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA); abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize); } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { /* * Ideally, we would always copy the io_abd into b_pabd, but the * user may have disabled compressed ARC, thus we must check the * hdr's compression setting rather than the io_bp's. */ if (BP_IS_ENCRYPTED(bp)) { ASSERT3U(psize, >, 0); arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA); abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize); } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF && !ARC_BUF_COMPRESSED(buf)) { ASSERT3U(psize, >, 0); arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT); abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); } else { ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT); abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, arc_buf_size(buf)); } } else { ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); arc_share_buf(hdr, buf); } out: arc_hdr_verify(hdr, bp); spl_fstrans_unmark(cookie); } static void arc_write_children_ready(zio_t *zio) { arc_write_callback_t *callback = zio->io_private; arc_buf_t *buf = callback->awcb_buf; callback->awcb_children_ready(zio, buf, callback->awcb_private); } /* * The SPA calls this callback for each physical write that happens on behalf * of a logical write. See the comment in dbuf_write_physdone() for details. */ static void arc_write_physdone(zio_t *zio) { arc_write_callback_t *cb = zio->io_private; if (cb->awcb_physdone != NULL) cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); } static void arc_write_done(zio_t *zio) { arc_write_callback_t *callback = zio->io_private; arc_buf_t *buf = callback->awcb_buf; arc_buf_hdr_t *hdr = buf->b_hdr; ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); if (zio->io_error == 0) { arc_hdr_verify(hdr, zio->io_bp); if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { buf_discard_identity(hdr); } else { hdr->b_dva = *BP_IDENTITY(zio->io_bp); hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); } } else { ASSERT(HDR_EMPTY(hdr)); } /* * If the block to be written was all-zero or compressed enough to be * embedded in the BP, no write was performed so there will be no * dva/birth/checksum. The buffer must therefore remain anonymous * (and uncached). */ if (!HDR_EMPTY(hdr)) { arc_buf_hdr_t *exists; kmutex_t *hash_lock; ASSERT3U(zio->io_error, ==, 0); arc_cksum_verify(buf); exists = buf_hash_insert(hdr, &hash_lock); if (exists != NULL) { /* * This can only happen if we overwrite for * sync-to-convergence, because we remove * buffers from the hash table when we arc_free(). */ if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) panic("bad overwrite, hdr=%p exists=%p", (void *)hdr, (void *)exists); ASSERT(zfs_refcount_is_zero( &exists->b_l1hdr.b_refcnt)); arc_change_state(arc_anon, exists, hash_lock); arc_hdr_destroy(exists); mutex_exit(hash_lock); exists = buf_hash_insert(hdr, &hash_lock); ASSERT3P(exists, ==, NULL); } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { /* nopwrite */ ASSERT(zio->io_prop.zp_nopwrite); if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) panic("bad nopwrite, hdr=%p exists=%p", (void *)hdr, (void *)exists); } else { /* Dedup */ ASSERT(hdr->b_l1hdr.b_bufcnt == 1); ASSERT(hdr->b_l1hdr.b_state == arc_anon); ASSERT(BP_GET_DEDUP(zio->io_bp)); ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); } } arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); /* if it's not anon, we are doing a scrub */ if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) arc_access(hdr, hash_lock); mutex_exit(hash_lock); } else { arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); } ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); callback->awcb_done(zio, buf, callback->awcb_private); abd_free(zio->io_abd); kmem_free(callback, sizeof (arc_write_callback_t)); } zio_t * arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready, arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone, arc_write_done_func_t *done, void *private, zio_priority_t priority, int zio_flags, const zbookmark_phys_t *zb) { arc_buf_hdr_t *hdr = buf->b_hdr; arc_write_callback_t *callback; zio_t *zio; zio_prop_t localprop = *zp; ASSERT3P(ready, !=, NULL); ASSERT3P(done, !=, NULL); ASSERT(!HDR_IO_ERROR(hdr)); ASSERT(!HDR_IO_IN_PROGRESS(hdr)); ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); if (l2arc) arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); if (ARC_BUF_ENCRYPTED(buf)) { ASSERT(ARC_BUF_COMPRESSED(buf)); localprop.zp_encrypt = B_TRUE; localprop.zp_compress = HDR_GET_COMPRESS(hdr); localprop.zp_complevel = hdr->b_complevel; localprop.zp_byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ? ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER; bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt, ZIO_DATA_SALT_LEN); bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv, ZIO_DATA_IV_LEN); bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac, ZIO_DATA_MAC_LEN); if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) { localprop.zp_nopwrite = B_FALSE; localprop.zp_copies = MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1); } zio_flags |= ZIO_FLAG_RAW; } else if (ARC_BUF_COMPRESSED(buf)) { ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); localprop.zp_compress = HDR_GET_COMPRESS(hdr); localprop.zp_complevel = hdr->b_complevel; zio_flags |= ZIO_FLAG_RAW_COMPRESS; } callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); callback->awcb_ready = ready; callback->awcb_children_ready = children_ready; callback->awcb_physdone = physdone; callback->awcb_done = done; callback->awcb_private = private; callback->awcb_buf = buf; /* * The hdr's b_pabd is now stale, free it now. A new data block * will be allocated when the zio pipeline calls arc_write_ready(). */ if (hdr->b_l1hdr.b_pabd != NULL) { /* * If the buf is currently sharing the data block with * the hdr then we need to break that relationship here. * The hdr will remain with a NULL data pointer and the * buf will take sole ownership of the block. */ if (arc_buf_is_shared(buf)) { arc_unshare_buf(hdr, buf); } else { arc_hdr_free_abd(hdr, B_FALSE); } VERIFY3P(buf->b_data, !=, NULL); } if (HDR_HAS_RABD(hdr)) arc_hdr_free_abd(hdr, B_TRUE); if (!(zio_flags & ZIO_FLAG_RAW)) arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); ASSERT(!arc_buf_is_shared(buf)); ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); zio = zio_write(pio, spa, txg, bp, abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, (children_ready != NULL) ? arc_write_children_ready : NULL, arc_write_physdone, arc_write_done, callback, priority, zio_flags, zb); return (zio); } void arc_tempreserve_clear(uint64_t reserve) { atomic_add_64(&arc_tempreserve, -reserve); ASSERT((int64_t)arc_tempreserve >= 0); } int arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg) { int error; uint64_t anon_size; if (!arc_no_grow && reserve > arc_c/4 && reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT)) arc_c = MIN(arc_c_max, reserve * 4); /* * Throttle when the calculated memory footprint for the TXG * exceeds the target ARC size. */ if (reserve > arc_c) { DMU_TX_STAT_BUMP(dmu_tx_memory_reserve); return (SET_ERROR(ERESTART)); } /* * Don't count loaned bufs as in flight dirty data to prevent long * network delays from blocking transactions that are ready to be * assigned to a txg. */ /* assert that it has not wrapped around */ ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) - arc_loaned_bytes), 0); /* * Writes will, almost always, require additional memory allocations * in order to compress/encrypt/etc the data. We therefore need to * make sure that there is sufficient available memory for this. */ error = arc_memory_throttle(spa, reserve, txg); if (error != 0) return (error); /* * Throttle writes when the amount of dirty data in the cache * gets too large. We try to keep the cache less than half full * of dirty blocks so that our sync times don't grow too large. * * In the case of one pool being built on another pool, we want * to make sure we don't end up throttling the lower (backing) * pool when the upper pool is the majority contributor to dirty * data. To insure we make forward progress during throttling, we * also check the current pool's net dirty data and only throttle * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty * data in the cache. * * Note: if two requests come in concurrently, we might let them * both succeed, when one of them should fail. Not a huge deal. */ uint64_t total_dirty = reserve + arc_tempreserve + anon_size; uint64_t spa_dirty_anon = spa_dirty_data(spa); uint64_t rarc_c = arc_warm ? arc_c : arc_c_max; if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 && anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 && spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) { #ifdef ZFS_DEBUG uint64_t meta_esize = zfs_refcount_count( &arc_anon->arcs_esize[ARC_BUFC_METADATA]); uint64_t data_esize = zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n", (u_longlong_t)arc_tempreserve >> 10, (u_longlong_t)meta_esize >> 10, (u_longlong_t)data_esize >> 10, (u_longlong_t)reserve >> 10, (u_longlong_t)rarc_c >> 10); #endif DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle); return (SET_ERROR(ERESTART)); } atomic_add_64(&arc_tempreserve, reserve); return (0); } static void arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, kstat_named_t *evict_data, kstat_named_t *evict_metadata) { size->value.ui64 = zfs_refcount_count(&state->arcs_size); evict_data->value.ui64 = zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); evict_metadata->value.ui64 = zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); } static int arc_kstat_update(kstat_t *ksp, int rw) { arc_stats_t *as = ksp->ks_data; if (rw == KSTAT_WRITE) return (SET_ERROR(EACCES)); as->arcstat_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_hits); as->arcstat_misses.value.ui64 = wmsum_value(&arc_sums.arcstat_misses); as->arcstat_demand_data_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_demand_data_hits); as->arcstat_demand_data_misses.value.ui64 = wmsum_value(&arc_sums.arcstat_demand_data_misses); as->arcstat_demand_metadata_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_demand_metadata_hits); as->arcstat_demand_metadata_misses.value.ui64 = wmsum_value(&arc_sums.arcstat_demand_metadata_misses); as->arcstat_prefetch_data_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_prefetch_data_hits); as->arcstat_prefetch_data_misses.value.ui64 = wmsum_value(&arc_sums.arcstat_prefetch_data_misses); as->arcstat_prefetch_metadata_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits); as->arcstat_prefetch_metadata_misses.value.ui64 = wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses); as->arcstat_mru_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_mru_hits); as->arcstat_mru_ghost_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_mru_ghost_hits); as->arcstat_mfu_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_mfu_hits); as->arcstat_mfu_ghost_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_mfu_ghost_hits); as->arcstat_deleted.value.ui64 = wmsum_value(&arc_sums.arcstat_deleted); as->arcstat_mutex_miss.value.ui64 = wmsum_value(&arc_sums.arcstat_mutex_miss); as->arcstat_access_skip.value.ui64 = wmsum_value(&arc_sums.arcstat_access_skip); as->arcstat_evict_skip.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_skip); as->arcstat_evict_not_enough.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_not_enough); as->arcstat_evict_l2_cached.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_l2_cached); as->arcstat_evict_l2_eligible.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_l2_eligible); as->arcstat_evict_l2_eligible_mfu.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu); as->arcstat_evict_l2_eligible_mru.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru); as->arcstat_evict_l2_ineligible.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_l2_ineligible); as->arcstat_evict_l2_skip.value.ui64 = wmsum_value(&arc_sums.arcstat_evict_l2_skip); as->arcstat_hash_collisions.value.ui64 = wmsum_value(&arc_sums.arcstat_hash_collisions); as->arcstat_hash_chains.value.ui64 = wmsum_value(&arc_sums.arcstat_hash_chains); as->arcstat_size.value.ui64 = aggsum_value(&arc_sums.arcstat_size); as->arcstat_compressed_size.value.ui64 = wmsum_value(&arc_sums.arcstat_compressed_size); as->arcstat_uncompressed_size.value.ui64 = wmsum_value(&arc_sums.arcstat_uncompressed_size); as->arcstat_overhead_size.value.ui64 = wmsum_value(&arc_sums.arcstat_overhead_size); as->arcstat_hdr_size.value.ui64 = wmsum_value(&arc_sums.arcstat_hdr_size); as->arcstat_data_size.value.ui64 = wmsum_value(&arc_sums.arcstat_data_size); as->arcstat_metadata_size.value.ui64 = wmsum_value(&arc_sums.arcstat_metadata_size); as->arcstat_dbuf_size.value.ui64 = wmsum_value(&arc_sums.arcstat_dbuf_size); #if defined(COMPAT_FREEBSD11) as->arcstat_other_size.value.ui64 = wmsum_value(&arc_sums.arcstat_bonus_size) + aggsum_value(&arc_sums.arcstat_dnode_size) + wmsum_value(&arc_sums.arcstat_dbuf_size); #endif arc_kstat_update_state(arc_anon, &as->arcstat_anon_size, &as->arcstat_anon_evictable_data, &as->arcstat_anon_evictable_metadata); arc_kstat_update_state(arc_mru, &as->arcstat_mru_size, &as->arcstat_mru_evictable_data, &as->arcstat_mru_evictable_metadata); arc_kstat_update_state(arc_mru_ghost, &as->arcstat_mru_ghost_size, &as->arcstat_mru_ghost_evictable_data, &as->arcstat_mru_ghost_evictable_metadata); arc_kstat_update_state(arc_mfu, &as->arcstat_mfu_size, &as->arcstat_mfu_evictable_data, &as->arcstat_mfu_evictable_metadata); arc_kstat_update_state(arc_mfu_ghost, &as->arcstat_mfu_ghost_size, &as->arcstat_mfu_ghost_evictable_data, &as->arcstat_mfu_ghost_evictable_metadata); as->arcstat_dnode_size.value.ui64 = aggsum_value(&arc_sums.arcstat_dnode_size); as->arcstat_bonus_size.value.ui64 = wmsum_value(&arc_sums.arcstat_bonus_size); as->arcstat_l2_hits.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_hits); as->arcstat_l2_misses.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_misses); as->arcstat_l2_prefetch_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_prefetch_asize); as->arcstat_l2_mru_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_mru_asize); as->arcstat_l2_mfu_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_mfu_asize); as->arcstat_l2_bufc_data_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize); as->arcstat_l2_bufc_metadata_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize); as->arcstat_l2_feeds.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_feeds); as->arcstat_l2_rw_clash.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rw_clash); as->arcstat_l2_read_bytes.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_read_bytes); as->arcstat_l2_write_bytes.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_write_bytes); as->arcstat_l2_writes_sent.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_writes_sent); as->arcstat_l2_writes_done.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_writes_done); as->arcstat_l2_writes_error.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_writes_error); as->arcstat_l2_writes_lock_retry.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry); as->arcstat_l2_evict_lock_retry.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry); as->arcstat_l2_evict_reading.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_evict_reading); as->arcstat_l2_evict_l1cached.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_evict_l1cached); as->arcstat_l2_free_on_write.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_free_on_write); as->arcstat_l2_abort_lowmem.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_abort_lowmem); as->arcstat_l2_cksum_bad.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_cksum_bad); as->arcstat_l2_io_error.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_io_error); as->arcstat_l2_lsize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_lsize); as->arcstat_l2_psize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_psize); as->arcstat_l2_hdr_size.value.ui64 = aggsum_value(&arc_sums.arcstat_l2_hdr_size); as->arcstat_l2_log_blk_writes.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_log_blk_writes); as->arcstat_l2_log_blk_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_log_blk_asize); as->arcstat_l2_log_blk_count.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_log_blk_count); as->arcstat_l2_rebuild_success.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_success); as->arcstat_l2_rebuild_abort_unsupported.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported); as->arcstat_l2_rebuild_abort_io_errors.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors); as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors); as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors); as->arcstat_l2_rebuild_abort_lowmem.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem); as->arcstat_l2_rebuild_size.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_size); as->arcstat_l2_rebuild_asize.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_asize); as->arcstat_l2_rebuild_bufs.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs); as->arcstat_l2_rebuild_bufs_precached.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached); as->arcstat_l2_rebuild_log_blks.value.ui64 = wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks); as->arcstat_memory_throttle_count.value.ui64 = wmsum_value(&arc_sums.arcstat_memory_throttle_count); as->arcstat_memory_direct_count.value.ui64 = wmsum_value(&arc_sums.arcstat_memory_direct_count); as->arcstat_memory_indirect_count.value.ui64 = wmsum_value(&arc_sums.arcstat_memory_indirect_count); as->arcstat_memory_all_bytes.value.ui64 = arc_all_memory(); as->arcstat_memory_free_bytes.value.ui64 = arc_free_memory(); as->arcstat_memory_available_bytes.value.i64 = arc_available_memory(); as->arcstat_prune.value.ui64 = wmsum_value(&arc_sums.arcstat_prune); as->arcstat_meta_used.value.ui64 = aggsum_value(&arc_sums.arcstat_meta_used); as->arcstat_async_upgrade_sync.value.ui64 = wmsum_value(&arc_sums.arcstat_async_upgrade_sync); as->arcstat_demand_hit_predictive_prefetch.value.ui64 = wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch); as->arcstat_demand_hit_prescient_prefetch.value.ui64 = wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch); as->arcstat_raw_size.value.ui64 = wmsum_value(&arc_sums.arcstat_raw_size); as->arcstat_cached_only_in_progress.value.ui64 = wmsum_value(&arc_sums.arcstat_cached_only_in_progress); as->arcstat_abd_chunk_waste_size.value.ui64 = wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size); return (0); } /* * This function *must* return indices evenly distributed between all * sublists of the multilist. This is needed due to how the ARC eviction * code is laid out; arc_evict_state() assumes ARC buffers are evenly * distributed between all sublists and uses this assumption when * deciding which sublist to evict from and how much to evict from it. */ static unsigned int arc_state_multilist_index_func(multilist_t *ml, void *obj) { arc_buf_hdr_t *hdr = obj; /* * We rely on b_dva to generate evenly distributed index * numbers using buf_hash below. So, as an added precaution, * let's make sure we never add empty buffers to the arc lists. */ ASSERT(!HDR_EMPTY(hdr)); /* * The assumption here, is the hash value for a given * arc_buf_hdr_t will remain constant throughout its lifetime * (i.e. its b_spa, b_dva, and b_birth fields don't change). * Thus, we don't need to store the header's sublist index * on insertion, as this index can be recalculated on removal. * * Also, the low order bits of the hash value are thought to be * distributed evenly. Otherwise, in the case that the multilist * has a power of two number of sublists, each sublists' usage * would not be evenly distributed. In this context full 64bit * division would be a waste of time, so limit it to 32 bits. */ return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % multilist_get_num_sublists(ml)); } #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \ if ((do_warn) && (tuning) && ((tuning) != (value))) { \ cmn_err(CE_WARN, \ "ignoring tunable %s (using %llu instead)", \ (#tuning), (value)); \ } \ } while (0) /* * Called during module initialization and periodically thereafter to * apply reasonable changes to the exposed performance tunings. Can also be * called explicitly by param_set_arc_*() functions when ARC tunables are * updated manually. Non-zero zfs_* values which differ from the currently set * values will be applied. */ void arc_tuning_update(boolean_t verbose) { uint64_t allmem = arc_all_memory(); unsigned long limit; /* Valid range: 32M - */ if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) && (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) && (zfs_arc_min <= arc_c_max)) { arc_c_min = zfs_arc_min; arc_c = MAX(arc_c, arc_c_min); } WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose); /* Valid range: 64M - */ if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) && (zfs_arc_max >= 64 << 20) && (zfs_arc_max < allmem) && (zfs_arc_max > arc_c_min)) { arc_c_max = zfs_arc_max; arc_c = MIN(arc_c, arc_c_max); arc_p = (arc_c >> 1); if (arc_meta_limit > arc_c_max) arc_meta_limit = arc_c_max; if (arc_dnode_size_limit > arc_meta_limit) arc_dnode_size_limit = arc_meta_limit; } WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose); /* Valid range: 16M - */ if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) && (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) && (zfs_arc_meta_min <= arc_c_max)) { arc_meta_min = zfs_arc_meta_min; if (arc_meta_limit < arc_meta_min) arc_meta_limit = arc_meta_min; if (arc_dnode_size_limit < arc_meta_min) arc_dnode_size_limit = arc_meta_min; } WARN_IF_TUNING_IGNORED(zfs_arc_meta_min, arc_meta_min, verbose); /* Valid range: - */ limit = zfs_arc_meta_limit ? zfs_arc_meta_limit : MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100; if ((limit != arc_meta_limit) && (limit >= arc_meta_min) && (limit <= arc_c_max)) arc_meta_limit = limit; WARN_IF_TUNING_IGNORED(zfs_arc_meta_limit, arc_meta_limit, verbose); /* Valid range: - */ limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit : MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100; if ((limit != arc_dnode_size_limit) && (limit >= arc_meta_min) && (limit <= arc_meta_limit)) arc_dnode_size_limit = limit; WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_size_limit, verbose); /* Valid range: 1 - N */ if (zfs_arc_grow_retry) arc_grow_retry = zfs_arc_grow_retry; /* Valid range: 1 - N */ if (zfs_arc_shrink_shift) { arc_shrink_shift = zfs_arc_shrink_shift; arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1); } /* Valid range: 1 - N */ if (zfs_arc_p_min_shift) arc_p_min_shift = zfs_arc_p_min_shift; /* Valid range: 1 - N ms */ if (zfs_arc_min_prefetch_ms) arc_min_prefetch_ms = zfs_arc_min_prefetch_ms; /* Valid range: 1 - N ms */ if (zfs_arc_min_prescient_prefetch_ms) { arc_min_prescient_prefetch_ms = zfs_arc_min_prescient_prefetch_ms; } /* Valid range: 0 - 100 */ if ((zfs_arc_lotsfree_percent >= 0) && (zfs_arc_lotsfree_percent <= 100)) arc_lotsfree_percent = zfs_arc_lotsfree_percent; WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent, verbose); /* Valid range: 0 - */ if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free)) arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem); WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose); } static void arc_state_init(void) { multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), arc_state_multilist_index_func); zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_create(&arc_anon->arcs_size); zfs_refcount_create(&arc_mru->arcs_size); zfs_refcount_create(&arc_mru_ghost->arcs_size); zfs_refcount_create(&arc_mfu->arcs_size); zfs_refcount_create(&arc_mfu_ghost->arcs_size); zfs_refcount_create(&arc_l2c_only->arcs_size); wmsum_init(&arc_sums.arcstat_hits, 0); wmsum_init(&arc_sums.arcstat_misses, 0); wmsum_init(&arc_sums.arcstat_demand_data_hits, 0); wmsum_init(&arc_sums.arcstat_demand_data_misses, 0); wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0); wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0); wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0); wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0); wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0); wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0); wmsum_init(&arc_sums.arcstat_mru_hits, 0); wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0); wmsum_init(&arc_sums.arcstat_mfu_hits, 0); wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0); wmsum_init(&arc_sums.arcstat_deleted, 0); wmsum_init(&arc_sums.arcstat_mutex_miss, 0); wmsum_init(&arc_sums.arcstat_access_skip, 0); wmsum_init(&arc_sums.arcstat_evict_skip, 0); wmsum_init(&arc_sums.arcstat_evict_not_enough, 0); wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0); wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0); wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0); wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0); wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0); wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0); wmsum_init(&arc_sums.arcstat_hash_collisions, 0); wmsum_init(&arc_sums.arcstat_hash_chains, 0); aggsum_init(&arc_sums.arcstat_size, 0); wmsum_init(&arc_sums.arcstat_compressed_size, 0); wmsum_init(&arc_sums.arcstat_uncompressed_size, 0); wmsum_init(&arc_sums.arcstat_overhead_size, 0); wmsum_init(&arc_sums.arcstat_hdr_size, 0); wmsum_init(&arc_sums.arcstat_data_size, 0); wmsum_init(&arc_sums.arcstat_metadata_size, 0); wmsum_init(&arc_sums.arcstat_dbuf_size, 0); aggsum_init(&arc_sums.arcstat_dnode_size, 0); wmsum_init(&arc_sums.arcstat_bonus_size, 0); wmsum_init(&arc_sums.arcstat_l2_hits, 0); wmsum_init(&arc_sums.arcstat_l2_misses, 0); wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0); wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0); wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0); wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0); wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0); wmsum_init(&arc_sums.arcstat_l2_feeds, 0); wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0); wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0); wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0); wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0); wmsum_init(&arc_sums.arcstat_l2_writes_done, 0); wmsum_init(&arc_sums.arcstat_l2_writes_error, 0); wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0); wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0); wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0); wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0); wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0); wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0); wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0); wmsum_init(&arc_sums.arcstat_l2_io_error, 0); wmsum_init(&arc_sums.arcstat_l2_lsize, 0); wmsum_init(&arc_sums.arcstat_l2_psize, 0); aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0); wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0); wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0); wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0); wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0); wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0); wmsum_init(&arc_sums.arcstat_memory_direct_count, 0); wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0); wmsum_init(&arc_sums.arcstat_prune, 0); aggsum_init(&arc_sums.arcstat_meta_used, 0); wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0); wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0); wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0); wmsum_init(&arc_sums.arcstat_raw_size, 0); wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0); wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0); arc_anon->arcs_state = ARC_STATE_ANON; arc_mru->arcs_state = ARC_STATE_MRU; arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST; arc_mfu->arcs_state = ARC_STATE_MFU; arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST; arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY; } static void arc_state_fini(void) { zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); zfs_refcount_destroy(&arc_anon->arcs_size); zfs_refcount_destroy(&arc_mru->arcs_size); zfs_refcount_destroy(&arc_mru_ghost->arcs_size); zfs_refcount_destroy(&arc_mfu->arcs_size); zfs_refcount_destroy(&arc_mfu_ghost->arcs_size); zfs_refcount_destroy(&arc_l2c_only->arcs_size); multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]); wmsum_fini(&arc_sums.arcstat_hits); wmsum_fini(&arc_sums.arcstat_misses); wmsum_fini(&arc_sums.arcstat_demand_data_hits); wmsum_fini(&arc_sums.arcstat_demand_data_misses); wmsum_fini(&arc_sums.arcstat_demand_metadata_hits); wmsum_fini(&arc_sums.arcstat_demand_metadata_misses); wmsum_fini(&arc_sums.arcstat_prefetch_data_hits); wmsum_fini(&arc_sums.arcstat_prefetch_data_misses); wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits); wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses); wmsum_fini(&arc_sums.arcstat_mru_hits); wmsum_fini(&arc_sums.arcstat_mru_ghost_hits); wmsum_fini(&arc_sums.arcstat_mfu_hits); wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits); wmsum_fini(&arc_sums.arcstat_deleted); wmsum_fini(&arc_sums.arcstat_mutex_miss); wmsum_fini(&arc_sums.arcstat_access_skip); wmsum_fini(&arc_sums.arcstat_evict_skip); wmsum_fini(&arc_sums.arcstat_evict_not_enough); wmsum_fini(&arc_sums.arcstat_evict_l2_cached); wmsum_fini(&arc_sums.arcstat_evict_l2_eligible); wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu); wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru); wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible); wmsum_fini(&arc_sums.arcstat_evict_l2_skip); wmsum_fini(&arc_sums.arcstat_hash_collisions); wmsum_fini(&arc_sums.arcstat_hash_chains); aggsum_fini(&arc_sums.arcstat_size); wmsum_fini(&arc_sums.arcstat_compressed_size); wmsum_fini(&arc_sums.arcstat_uncompressed_size); wmsum_fini(&arc_sums.arcstat_overhead_size); wmsum_fini(&arc_sums.arcstat_hdr_size); wmsum_fini(&arc_sums.arcstat_data_size); wmsum_fini(&arc_sums.arcstat_metadata_size); wmsum_fini(&arc_sums.arcstat_dbuf_size); aggsum_fini(&arc_sums.arcstat_dnode_size); wmsum_fini(&arc_sums.arcstat_bonus_size); wmsum_fini(&arc_sums.arcstat_l2_hits); wmsum_fini(&arc_sums.arcstat_l2_misses); wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize); wmsum_fini(&arc_sums.arcstat_l2_mru_asize); wmsum_fini(&arc_sums.arcstat_l2_mfu_asize); wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize); wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize); wmsum_fini(&arc_sums.arcstat_l2_feeds); wmsum_fini(&arc_sums.arcstat_l2_rw_clash); wmsum_fini(&arc_sums.arcstat_l2_read_bytes); wmsum_fini(&arc_sums.arcstat_l2_write_bytes); wmsum_fini(&arc_sums.arcstat_l2_writes_sent); wmsum_fini(&arc_sums.arcstat_l2_writes_done); wmsum_fini(&arc_sums.arcstat_l2_writes_error); wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry); wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry); wmsum_fini(&arc_sums.arcstat_l2_evict_reading); wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached); wmsum_fini(&arc_sums.arcstat_l2_free_on_write); wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem); wmsum_fini(&arc_sums.arcstat_l2_cksum_bad); wmsum_fini(&arc_sums.arcstat_l2_io_error); wmsum_fini(&arc_sums.arcstat_l2_lsize); wmsum_fini(&arc_sums.arcstat_l2_psize); aggsum_fini(&arc_sums.arcstat_l2_hdr_size); wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes); wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize); wmsum_fini(&arc_sums.arcstat_l2_log_blk_count); wmsum_fini(&arc_sums.arcstat_l2_rebuild_success); wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported); wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors); wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors); wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors); wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem); wmsum_fini(&arc_sums.arcstat_l2_rebuild_size); wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize); wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs); wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached); wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks); wmsum_fini(&arc_sums.arcstat_memory_throttle_count); wmsum_fini(&arc_sums.arcstat_memory_direct_count); wmsum_fini(&arc_sums.arcstat_memory_indirect_count); wmsum_fini(&arc_sums.arcstat_prune); aggsum_fini(&arc_sums.arcstat_meta_used); wmsum_fini(&arc_sums.arcstat_async_upgrade_sync); wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch); wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch); wmsum_fini(&arc_sums.arcstat_raw_size); wmsum_fini(&arc_sums.arcstat_cached_only_in_progress); wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size); } uint64_t arc_target_bytes(void) { return (arc_c); } void arc_set_limits(uint64_t allmem) { /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */ arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT); /* How to set default max varies by platform. */ arc_c_max = arc_default_max(arc_c_min, allmem); } void arc_init(void) { uint64_t percent, allmem = arc_all_memory(); mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL); list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t), offsetof(arc_evict_waiter_t, aew_node)); arc_min_prefetch_ms = 1000; arc_min_prescient_prefetch_ms = 6000; #if defined(_KERNEL) arc_lowmem_init(); #endif arc_set_limits(allmem); #ifndef _KERNEL /* * In userland, there's only the memory pressure that we artificially * create (see arc_available_memory()). Don't let arc_c get too * small, because it can cause transactions to be larger than * arc_c, causing arc_tempreserve_space() to fail. */ arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT); #endif arc_c = arc_c_min; arc_p = (arc_c >> 1); /* Set min to 1/2 of arc_c_min */ arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT; /* * Set arc_meta_limit to a percent of arc_c_max with a floor of * arc_meta_min, and a ceiling of arc_c_max. */ percent = MIN(zfs_arc_meta_limit_percent, 100); arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100); percent = MIN(zfs_arc_dnode_limit_percent, 100); arc_dnode_size_limit = (percent * arc_meta_limit) / 100; /* Apply user specified tunings */ arc_tuning_update(B_TRUE); /* if kmem_flags are set, lets try to use less memory */ if (kmem_debugging()) arc_c = arc_c / 2; if (arc_c < arc_c_min) arc_c = arc_c_min; arc_register_hotplug(); arc_state_init(); buf_init(); list_create(&arc_prune_list, sizeof (arc_prune_t), offsetof(arc_prune_t, p_node)); mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL); arc_prune_taskq = taskq_create("arc_prune", 100, defclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); if (arc_ksp != NULL) { arc_ksp->ks_data = &arc_stats; arc_ksp->ks_update = arc_kstat_update; kstat_install(arc_ksp); } arc_evict_zthr = zthr_create("arc_evict", - arc_evict_cb_check, arc_evict_cb, NULL); + arc_evict_cb_check, arc_evict_cb, NULL, defclsyspri); arc_reap_zthr = zthr_create_timer("arc_reap", - arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1)); + arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri); arc_warm = B_FALSE; /* * Calculate maximum amount of dirty data per pool. * * If it has been set by a module parameter, take that. * Otherwise, use a percentage of physical memory defined by * zfs_dirty_data_max_percent (default 10%) with a cap at * zfs_dirty_data_max_max (default 4G or 25% of physical memory). */ #ifdef __LP64__ if (zfs_dirty_data_max_max == 0) zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024, allmem * zfs_dirty_data_max_max_percent / 100); #else if (zfs_dirty_data_max_max == 0) zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024, allmem * zfs_dirty_data_max_max_percent / 100); #endif if (zfs_dirty_data_max == 0) { zfs_dirty_data_max = allmem * zfs_dirty_data_max_percent / 100; zfs_dirty_data_max = MIN(zfs_dirty_data_max, zfs_dirty_data_max_max); } } void arc_fini(void) { arc_prune_t *p; #ifdef _KERNEL arc_lowmem_fini(); #endif /* _KERNEL */ /* Use B_TRUE to ensure *all* buffers are evicted */ arc_flush(NULL, B_TRUE); if (arc_ksp != NULL) { kstat_delete(arc_ksp); arc_ksp = NULL; } taskq_wait(arc_prune_taskq); taskq_destroy(arc_prune_taskq); mutex_enter(&arc_prune_mtx); while ((p = list_head(&arc_prune_list)) != NULL) { list_remove(&arc_prune_list, p); zfs_refcount_remove(&p->p_refcnt, &arc_prune_list); zfs_refcount_destroy(&p->p_refcnt); kmem_free(p, sizeof (*p)); } mutex_exit(&arc_prune_mtx); list_destroy(&arc_prune_list); mutex_destroy(&arc_prune_mtx); (void) zthr_cancel(arc_evict_zthr); (void) zthr_cancel(arc_reap_zthr); mutex_destroy(&arc_evict_lock); list_destroy(&arc_evict_waiters); /* * Free any buffers that were tagged for destruction. This needs * to occur before arc_state_fini() runs and destroys the aggsum * values which are updated when freeing scatter ABDs. */ l2arc_do_free_on_write(); /* * buf_fini() must proceed arc_state_fini() because buf_fin() may * trigger the release of kmem magazines, which can callback to * arc_space_return() which accesses aggsums freed in act_state_fini(). */ buf_fini(); arc_state_fini(); arc_unregister_hotplug(); /* * We destroy the zthrs after all the ARC state has been * torn down to avoid the case of them receiving any * wakeup() signals after they are destroyed. */ zthr_destroy(arc_evict_zthr); zthr_destroy(arc_reap_zthr); ASSERT0(arc_loaned_bytes); } /* * Level 2 ARC * * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. * It uses dedicated storage devices to hold cached data, which are populated * using large infrequent writes. The main role of this cache is to boost * the performance of random read workloads. The intended L2ARC devices * include short-stroked disks, solid state disks, and other media with * substantially faster read latency than disk. * * +-----------------------+ * | ARC | * +-----------------------+ * | ^ ^ * | | | * l2arc_feed_thread() arc_read() * | | | * | l2arc read | * V | | * +---------------+ | * | L2ARC | | * +---------------+ | * | ^ | * l2arc_write() | | * | | | * V | | * +-------+ +-------+ * | vdev | | vdev | * | cache | | cache | * +-------+ +-------+ * +=========+ .-----. * : L2ARC : |-_____-| * : devices : | Disks | * +=========+ `-_____-' * * Read requests are satisfied from the following sources, in order: * * 1) ARC * 2) vdev cache of L2ARC devices * 3) L2ARC devices * 4) vdev cache of disks * 5) disks * * Some L2ARC device types exhibit extremely slow write performance. * To accommodate for this there are some significant differences between * the L2ARC and traditional cache design: * * 1. There is no eviction path from the ARC to the L2ARC. Evictions from * the ARC behave as usual, freeing buffers and placing headers on ghost * lists. The ARC does not send buffers to the L2ARC during eviction as * this would add inflated write latencies for all ARC memory pressure. * * 2. The L2ARC attempts to cache data from the ARC before it is evicted. * It does this by periodically scanning buffers from the eviction-end of * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are * not already there. It scans until a headroom of buffers is satisfied, * which itself is a buffer for ARC eviction. If a compressible buffer is * found during scanning and selected for writing to an L2ARC device, we * temporarily boost scanning headroom during the next scan cycle to make * sure we adapt to compression effects (which might significantly reduce * the data volume we write to L2ARC). The thread that does this is * l2arc_feed_thread(), illustrated below; example sizes are included to * provide a better sense of ratio than this diagram: * * head --> tail * +---------------------+----------+ * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC * +---------------------+----------+ | o L2ARC eligible * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer * +---------------------+----------+ | * 15.9 Gbytes ^ 32 Mbytes | * headroom | * l2arc_feed_thread() * | * l2arc write hand <--[oooo]--' * | 8 Mbyte * | write max * V * +==============================+ * L2ARC dev |####|#|###|###| |####| ... | * +==============================+ * 32 Gbytes * * 3. If an ARC buffer is copied to the L2ARC but then hit instead of * evicted, then the L2ARC has cached a buffer much sooner than it probably * needed to, potentially wasting L2ARC device bandwidth and storage. It is * safe to say that this is an uncommon case, since buffers at the end of * the ARC lists have moved there due to inactivity. * * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, * then the L2ARC simply misses copying some buffers. This serves as a * pressure valve to prevent heavy read workloads from both stalling the ARC * with waits and clogging the L2ARC with writes. This also helps prevent * the potential for the L2ARC to churn if it attempts to cache content too * quickly, such as during backups of the entire pool. * * 5. After system boot and before the ARC has filled main memory, there are * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru * lists can remain mostly static. Instead of searching from tail of these * lists as pictured, the l2arc_feed_thread() will search from the list heads * for eligible buffers, greatly increasing its chance of finding them. * * The L2ARC device write speed is also boosted during this time so that * the L2ARC warms up faster. Since there have been no ARC evictions yet, * there are no L2ARC reads, and no fear of degrading read performance * through increased writes. * * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that * the vdev queue can aggregate them into larger and fewer writes. Each * device is written to in a rotor fashion, sweeping writes through * available space then repeating. * * 7. The L2ARC does not store dirty content. It never needs to flush * write buffers back to disk based storage. * * 8. If an ARC buffer is written (and dirtied) which also exists in the * L2ARC, the now stale L2ARC buffer is immediately dropped. * * The performance of the L2ARC can be tweaked by a number of tunables, which * may be necessary for different workloads: * * l2arc_write_max max write bytes per interval * l2arc_write_boost extra write bytes during device warmup * l2arc_noprefetch skip caching prefetched buffers * l2arc_headroom number of max device writes to precache * l2arc_headroom_boost when we find compressed buffers during ARC * scanning, we multiply headroom by this * percentage factor for the next scan cycle, * since more compressed buffers are likely to * be present * l2arc_feed_secs seconds between L2ARC writing * * Tunables may be removed or added as future performance improvements are * integrated, and also may become zpool properties. * * There are three key functions that control how the L2ARC warms up: * * l2arc_write_eligible() check if a buffer is eligible to cache * l2arc_write_size() calculate how much to write * l2arc_write_interval() calculate sleep delay between writes * * These three functions determine what to write, how much, and how quickly * to send writes. * * L2ARC persistence: * * When writing buffers to L2ARC, we periodically add some metadata to * make sure we can pick them up after reboot, thus dramatically reducing * the impact that any downtime has on the performance of storage systems * with large caches. * * The implementation works fairly simply by integrating the following two * modifications: * * *) When writing to the L2ARC, we occasionally write a "l2arc log block", * which is an additional piece of metadata which describes what's been * written. This allows us to rebuild the arc_buf_hdr_t structures of the * main ARC buffers. There are 2 linked-lists of log blocks headed by * dh_start_lbps[2]. We alternate which chain we append to, so they are * time-wise and offset-wise interleaved, but that is an optimization rather * than for correctness. The log block also includes a pointer to the * previous block in its chain. * * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device * for our header bookkeeping purposes. This contains a device header, * which contains our top-level reference structures. We update it each * time we write a new log block, so that we're able to locate it in the * L2ARC device. If this write results in an inconsistent device header * (e.g. due to power failure), we detect this by verifying the header's * checksum and simply fail to reconstruct the L2ARC after reboot. * * Implementation diagram: * * +=== L2ARC device (not to scale) ======================================+ * | ___two newest log block pointers__.__________ | * | / \dh_start_lbps[1] | * | / \ \dh_start_lbps[0]| * |.___/__. V V | * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---| * || hdr| ^ /^ /^ / / | * |+------+ ...--\-------/ \-----/--\------/ / | * | \--------------/ \--------------/ | * +======================================================================+ * * As can be seen on the diagram, rather than using a simple linked list, * we use a pair of linked lists with alternating elements. This is a * performance enhancement due to the fact that we only find out the * address of the next log block access once the current block has been * completely read in. Obviously, this hurts performance, because we'd be * keeping the device's I/O queue at only a 1 operation deep, thus * incurring a large amount of I/O round-trip latency. Having two lists * allows us to fetch two log blocks ahead of where we are currently * rebuilding L2ARC buffers. * * On-device data structures: * * L2ARC device header: l2arc_dev_hdr_phys_t * L2ARC log block: l2arc_log_blk_phys_t * * L2ARC reconstruction: * * When writing data, we simply write in the standard rotary fashion, * evicting buffers as we go and simply writing new data over them (writing * a new log block every now and then). This obviously means that once we * loop around the end of the device, we will start cutting into an already * committed log block (and its referenced data buffers), like so: * * current write head__ __old tail * \ / * V V * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |--> * ^ ^^^^^^^^^___________________________________ * | \ * <> may overwrite this blk and/or its bufs --' * * When importing the pool, we detect this situation and use it to stop * our scanning process (see l2arc_rebuild). * * There is one significant caveat to consider when rebuilding ARC contents * from an L2ARC device: what about invalidated buffers? Given the above * construction, we cannot update blocks which we've already written to amend * them to remove buffers which were invalidated. Thus, during reconstruction, * we might be populating the cache with buffers for data that's not on the * main pool anymore, or may have been overwritten! * * As it turns out, this isn't a problem. Every arc_read request includes * both the DVA and, crucially, the birth TXG of the BP the caller is * looking for. So even if the cache were populated by completely rotten * blocks for data that had been long deleted and/or overwritten, we'll * never actually return bad data from the cache, since the DVA with the * birth TXG uniquely identify a block in space and time - once created, * a block is immutable on disk. The worst thing we have done is wasted * some time and memory at l2arc rebuild to reconstruct outdated ARC * entries that will get dropped from the l2arc as it is being updated * with new blocks. * * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write * hand are not restored. This is done by saving the offset (in bytes) * l2arc_evict() has evicted to in the L2ARC device header and taking it * into account when restoring buffers. */ static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) { /* * A buffer is *not* eligible for the L2ARC if it: * 1. belongs to a different spa. * 2. is already cached on the L2ARC. * 3. has an I/O in progress (it may be an incomplete read). * 4. is flagged not eligible (zfs property). */ if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) return (B_FALSE); return (B_TRUE); } static uint64_t l2arc_write_size(l2arc_dev_t *dev) { uint64_t size, dev_size, tsize; /* * Make sure our globals have meaningful values in case the user * altered them. */ size = l2arc_write_max; if (size == 0) { cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " "be greater than zero, resetting it to the default (%d)", L2ARC_WRITE_SIZE); size = l2arc_write_max = L2ARC_WRITE_SIZE; } if (arc_warm == B_FALSE) size += l2arc_write_boost; /* * Make sure the write size does not exceed the size of the cache * device. This is important in l2arc_evict(), otherwise infinite * iteration can occur. */ dev_size = dev->l2ad_end - dev->l2ad_start; tsize = size + l2arc_log_blk_overhead(size, dev); if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0) tsize += MAX(64 * 1024 * 1024, (tsize * l2arc_trim_ahead) / 100); if (tsize >= dev_size) { cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost " "plus the overhead of log blocks (persistent L2ARC, " "%llu bytes) exceeds the size of the cache device " "(guid %llu), resetting them to the default (%d)", l2arc_log_blk_overhead(size, dev), dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE); size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE; if (arc_warm == B_FALSE) size += l2arc_write_boost; } return (size); } static clock_t l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) { clock_t interval, next, now; /* * If the ARC lists are busy, increase our write rate; if the * lists are stale, idle back. This is achieved by checking * how much we previously wrote - if it was more than half of * what we wanted, schedule the next write much sooner. */ if (l2arc_feed_again && wrote > (wanted / 2)) interval = (hz * l2arc_feed_min_ms) / 1000; else interval = hz * l2arc_feed_secs; now = ddi_get_lbolt(); next = MAX(now, MIN(now + interval, began + interval)); return (next); } /* * Cycle through L2ARC devices. This is how L2ARC load balances. * If a device is returned, this also returns holding the spa config lock. */ static l2arc_dev_t * l2arc_dev_get_next(void) { l2arc_dev_t *first, *next = NULL; /* * Lock out the removal of spas (spa_namespace_lock), then removal * of cache devices (l2arc_dev_mtx). Once a device has been selected, * both locks will be dropped and a spa config lock held instead. */ mutex_enter(&spa_namespace_lock); mutex_enter(&l2arc_dev_mtx); /* if there are no vdevs, there is nothing to do */ if (l2arc_ndev == 0) goto out; first = NULL; next = l2arc_dev_last; do { /* loop around the list looking for a non-faulted vdev */ if (next == NULL) { next = list_head(l2arc_dev_list); } else { next = list_next(l2arc_dev_list, next); if (next == NULL) next = list_head(l2arc_dev_list); } /* if we have come back to the start, bail out */ if (first == NULL) first = next; else if (next == first) break; } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild || next->l2ad_trim_all); /* if we were unable to find any usable vdevs, return NULL */ if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild || next->l2ad_trim_all) next = NULL; l2arc_dev_last = next; out: mutex_exit(&l2arc_dev_mtx); /* * Grab the config lock to prevent the 'next' device from being * removed while we are writing to it. */ if (next != NULL) spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); mutex_exit(&spa_namespace_lock); return (next); } /* * Free buffers that were tagged for destruction. */ static void l2arc_do_free_on_write(void) { list_t *buflist; l2arc_data_free_t *df, *df_prev; mutex_enter(&l2arc_free_on_write_mtx); buflist = l2arc_free_on_write; for (df = list_tail(buflist); df; df = df_prev) { df_prev = list_prev(buflist, df); ASSERT3P(df->l2df_abd, !=, NULL); abd_free(df->l2df_abd); list_remove(buflist, df); kmem_free(df, sizeof (l2arc_data_free_t)); } mutex_exit(&l2arc_free_on_write_mtx); } /* * A write to a cache device has completed. Update all headers to allow * reads from these buffers to begin. */ static void l2arc_write_done(zio_t *zio) { l2arc_write_callback_t *cb; l2arc_lb_abd_buf_t *abd_buf; l2arc_lb_ptr_buf_t *lb_ptr_buf; l2arc_dev_t *dev; l2arc_dev_hdr_phys_t *l2dhdr; list_t *buflist; arc_buf_hdr_t *head, *hdr, *hdr_prev; kmutex_t *hash_lock; int64_t bytes_dropped = 0; cb = zio->io_private; ASSERT3P(cb, !=, NULL); dev = cb->l2wcb_dev; l2dhdr = dev->l2ad_dev_hdr; ASSERT3P(dev, !=, NULL); head = cb->l2wcb_head; ASSERT3P(head, !=, NULL); buflist = &dev->l2ad_buflist; ASSERT3P(buflist, !=, NULL); DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, l2arc_write_callback_t *, cb); /* * All writes completed, or an error was hit. */ top: mutex_enter(&dev->l2ad_mtx); for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { hdr_prev = list_prev(buflist, hdr); hash_lock = HDR_LOCK(hdr); /* * We cannot use mutex_enter or else we can deadlock * with l2arc_write_buffers (due to swapping the order * the hash lock and l2ad_mtx are taken). */ if (!mutex_tryenter(hash_lock)) { /* * Missed the hash lock. We must retry so we * don't leave the ARC_FLAG_L2_WRITING bit set. */ ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); /* * We don't want to rescan the headers we've * already marked as having been written out, so * we reinsert the head node so we can pick up * where we left off. */ list_remove(buflist, head); list_insert_after(buflist, hdr, head); mutex_exit(&dev->l2ad_mtx); /* * We wait for the hash lock to become available * to try and prevent busy waiting, and increase * the chance we'll be able to acquire the lock * the next time around. */ mutex_enter(hash_lock); mutex_exit(hash_lock); goto top; } /* * We could not have been moved into the arc_l2c_only * state while in-flight due to our ARC_FLAG_L2_WRITING * bit being set. Let's just ensure that's being enforced. */ ASSERT(HDR_HAS_L1HDR(hdr)); /* * Skipped - drop L2ARC entry and mark the header as no * longer L2 eligibile. */ if (zio->io_error != 0) { /* * Error - drop L2ARC entry. */ list_remove(buflist, hdr); arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); uint64_t psize = HDR_GET_PSIZE(hdr); l2arc_hdr_arcstats_decrement(hdr); bytes_dropped += vdev_psize_to_asize(dev->l2ad_vdev, psize); (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); } /* * Allow ARC to begin reads and ghost list evictions to * this L2ARC entry. */ arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); mutex_exit(hash_lock); } /* * Free the allocated abd buffers for writing the log blocks. * If the zio failed reclaim the allocated space and remove the * pointers to these log blocks from the log block pointer list * of the L2ARC device. */ while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) { abd_free(abd_buf->abd); zio_buf_free(abd_buf, sizeof (*abd_buf)); if (zio->io_error != 0) { lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list); /* * L2BLK_GET_PSIZE returns aligned size for log * blocks. */ uint64_t asize = L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop); bytes_dropped += asize; ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize); ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count); zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf); zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf); kmem_free(lb_ptr_buf->lb_ptr, sizeof (l2arc_log_blkptr_t)); kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t)); } } list_destroy(&cb->l2wcb_abd_list); if (zio->io_error != 0) { ARCSTAT_BUMP(arcstat_l2_writes_error); /* * Restore the lbps array in the header to its previous state. * If the list of log block pointers is empty, zero out the * log block pointers in the device header. */ lb_ptr_buf = list_head(&dev->l2ad_lbptr_list); for (int i = 0; i < 2; i++) { if (lb_ptr_buf == NULL) { /* * If the list is empty zero out the device * header. Otherwise zero out the second log * block pointer in the header. */ if (i == 0) { bzero(l2dhdr, dev->l2ad_dev_hdr_asize); } else { bzero(&l2dhdr->dh_start_lbps[i], sizeof (l2arc_log_blkptr_t)); } break; } bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i], sizeof (l2arc_log_blkptr_t)); lb_ptr_buf = list_next(&dev->l2ad_lbptr_list, lb_ptr_buf); } } ARCSTAT_BUMP(arcstat_l2_writes_done); list_remove(buflist, head); ASSERT(!HDR_HAS_L1HDR(head)); kmem_cache_free(hdr_l2only_cache, head); mutex_exit(&dev->l2ad_mtx); ASSERT(dev->l2ad_vdev != NULL); vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); l2arc_do_free_on_write(); kmem_free(cb, sizeof (l2arc_write_callback_t)); } static int l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb) { int ret; spa_t *spa = zio->io_spa; arc_buf_hdr_t *hdr = cb->l2rcb_hdr; blkptr_t *bp = zio->io_bp; uint8_t salt[ZIO_DATA_SALT_LEN]; uint8_t iv[ZIO_DATA_IV_LEN]; uint8_t mac[ZIO_DATA_MAC_LEN]; boolean_t no_crypt = B_FALSE; /* * ZIL data is never be written to the L2ARC, so we don't need * special handling for its unique MAC storage. */ ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG); ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); /* * If the data was encrypted, decrypt it now. Note that * we must check the bp here and not the hdr, since the * hdr does not have its encryption parameters updated * until arc_read_done(). */ if (BP_IS_ENCRYPTED(bp)) { abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, B_TRUE); zio_crypt_decode_params_bp(bp, salt, iv); zio_crypt_decode_mac_bp(bp, mac); ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb, BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, HDR_GET_PSIZE(hdr), eabd, hdr->b_l1hdr.b_pabd, &no_crypt); if (ret != 0) { arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr); goto error; } /* * If we actually performed decryption, replace b_pabd * with the decrypted data. Otherwise we can just throw * our decryption buffer away. */ if (!no_crypt) { arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, arc_hdr_size(hdr), hdr); hdr->b_l1hdr.b_pabd = eabd; zio->io_abd = eabd; } else { arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr); } } /* * If the L2ARC block was compressed, but ARC compression * is disabled we decompress the data into a new buffer and * replace the existing data. */ if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) { abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, B_TRUE); void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr)); ret = zio_decompress_data(HDR_GET_COMPRESS(hdr), hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr), &hdr->b_complevel); if (ret != 0) { abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr); goto error; } abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, arc_hdr_size(hdr), hdr); hdr->b_l1hdr.b_pabd = cabd; zio->io_abd = cabd; zio->io_size = HDR_GET_LSIZE(hdr); } return (0); error: return (ret); } /* * A read to a cache device completed. Validate buffer contents before * handing over to the regular ARC routines. */ static void l2arc_read_done(zio_t *zio) { int tfm_error = 0; l2arc_read_callback_t *cb = zio->io_private; arc_buf_hdr_t *hdr; kmutex_t *hash_lock; boolean_t valid_cksum; boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) && (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT)); ASSERT3P(zio->io_vd, !=, NULL); ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); ASSERT3P(cb, !=, NULL); hdr = cb->l2rcb_hdr; ASSERT3P(hdr, !=, NULL); hash_lock = HDR_LOCK(hdr); mutex_enter(hash_lock); ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); /* * If the data was read into a temporary buffer, * move it and free the buffer. */ if (cb->l2rcb_abd != NULL) { ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); if (zio->io_error == 0) { if (using_rdata) { abd_copy(hdr->b_crypt_hdr.b_rabd, cb->l2rcb_abd, arc_hdr_size(hdr)); } else { abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, arc_hdr_size(hdr)); } } /* * The following must be done regardless of whether * there was an error: * - free the temporary buffer * - point zio to the real ARC buffer * - set zio size accordingly * These are required because zio is either re-used for * an I/O of the block in the case of the error * or the zio is passed to arc_read_done() and it * needs real data. */ abd_free(cb->l2rcb_abd); zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); if (using_rdata) { ASSERT(HDR_HAS_RABD(hdr)); zio->io_abd = zio->io_orig_abd = hdr->b_crypt_hdr.b_rabd; } else { ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; } } ASSERT3P(zio->io_abd, !=, NULL); /* * Check this survived the L2ARC journey. */ ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd || (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd)); zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ zio->io_prop.zp_complevel = hdr->b_complevel; valid_cksum = arc_cksum_is_equal(hdr, zio); /* * b_rabd will always match the data as it exists on disk if it is * being used. Therefore if we are reading into b_rabd we do not * attempt to untransform the data. */ if (valid_cksum && !using_rdata) tfm_error = l2arc_untransform(zio, cb); if (valid_cksum && tfm_error == 0 && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { mutex_exit(hash_lock); zio->io_private = hdr; arc_read_done(zio); } else { /* * Buffer didn't survive caching. Increment stats and * reissue to the original storage device. */ if (zio->io_error != 0) { ARCSTAT_BUMP(arcstat_l2_io_error); } else { zio->io_error = SET_ERROR(EIO); } if (!valid_cksum || tfm_error != 0) ARCSTAT_BUMP(arcstat_l2_cksum_bad); /* * If there's no waiter, issue an async i/o to the primary * storage now. If there *is* a waiter, the caller must * issue the i/o in a context where it's OK to block. */ if (zio->io_waiter == NULL) { zio_t *pio = zio_unique_parent(zio); void *abd = (using_rdata) ? hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd; ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); zio = zio_read(pio, zio->io_spa, zio->io_bp, abd, zio->io_size, arc_read_done, hdr, zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb); /* * Original ZIO will be freed, so we need to update * ARC header with the new ZIO pointer to be used * by zio_change_priority() in arc_read(). */ for (struct arc_callback *acb = hdr->b_l1hdr.b_acb; acb != NULL; acb = acb->acb_next) acb->acb_zio_head = zio; mutex_exit(hash_lock); zio_nowait(zio); } else { mutex_exit(hash_lock); } } kmem_free(cb, sizeof (l2arc_read_callback_t)); } /* * This is the list priority from which the L2ARC will search for pages to * cache. This is used within loops (0..3) to cycle through lists in the * desired order. This order can have a significant effect on cache * performance. * * Currently the metadata lists are hit first, MFU then MRU, followed by * the data lists. This function returns a locked list, and also returns * the lock pointer. */ static multilist_sublist_t * l2arc_sublist_lock(int list_num) { multilist_t *ml = NULL; unsigned int idx; ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES); switch (list_num) { case 0: ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; break; case 1: ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; break; case 2: ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; break; case 3: ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; break; default: return (NULL); } /* * Return a randomly-selected sublist. This is acceptable * because the caller feeds only a little bit of data for each * call (8MB). Subsequent calls will result in different * sublists being selected. */ idx = multilist_get_random_index(ml); return (multilist_sublist_lock(ml, idx)); } /* * Calculates the maximum overhead of L2ARC metadata log blocks for a given * L2ARC write size. l2arc_evict and l2arc_write_size need to include this * overhead in processing to make sure there is enough headroom available * when writing buffers. */ static inline uint64_t l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev) { if (dev->l2ad_log_entries == 0) { return (0); } else { uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT; uint64_t log_blocks = (log_entries + dev->l2ad_log_entries - 1) / dev->l2ad_log_entries; return (vdev_psize_to_asize(dev->l2ad_vdev, sizeof (l2arc_log_blk_phys_t)) * log_blocks); } } /* * Evict buffers from the device write hand to the distance specified in * bytes. This distance may span populated buffers, it may span nothing. * This is clearing a region on the L2ARC device ready for writing. * If the 'all' boolean is set, every buffer is evicted. */ static void l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) { list_t *buflist; arc_buf_hdr_t *hdr, *hdr_prev; kmutex_t *hash_lock; uint64_t taddr; l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev; vdev_t *vd = dev->l2ad_vdev; boolean_t rerun; buflist = &dev->l2ad_buflist; /* * We need to add in the worst case scenario of log block overhead. */ distance += l2arc_log_blk_overhead(distance, dev); if (vd->vdev_has_trim && l2arc_trim_ahead > 0) { /* * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100) * times the write size, whichever is greater. */ distance += MAX(64 * 1024 * 1024, (distance * l2arc_trim_ahead) / 100); } top: rerun = B_FALSE; if (dev->l2ad_hand >= (dev->l2ad_end - distance)) { /* * When there is no space to accommodate upcoming writes, * evict to the end. Then bump the write and evict hands * to the start and iterate. This iteration does not * happen indefinitely as we make sure in * l2arc_write_size() that when the write hand is reset, * the write size does not exceed the end of the device. */ rerun = B_TRUE; taddr = dev->l2ad_end; } else { taddr = dev->l2ad_hand + distance; } DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, uint64_t, taddr, boolean_t, all); if (!all) { /* * This check has to be placed after deciding whether to * iterate (rerun). */ if (dev->l2ad_first) { /* * This is the first sweep through the device. There is * nothing to evict. We have already trimmmed the * whole device. */ goto out; } else { /* * Trim the space to be evicted. */ if (vd->vdev_has_trim && dev->l2ad_evict < taddr && l2arc_trim_ahead > 0) { /* * We have to drop the spa_config lock because * vdev_trim_range() will acquire it. * l2ad_evict already accounts for the label * size. To prevent vdev_trim_ranges() from * adding it again, we subtract it from * l2ad_evict. */ spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev); vdev_trim_simple(vd, dev->l2ad_evict - VDEV_LABEL_START_SIZE, taddr - dev->l2ad_evict); spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev, RW_READER); } /* * When rebuilding L2ARC we retrieve the evict hand * from the header of the device. Of note, l2arc_evict() * does not actually delete buffers from the cache * device, but trimming may do so depending on the * hardware implementation. Thus keeping track of the * evict hand is useful. */ dev->l2ad_evict = MAX(dev->l2ad_evict, taddr); } } retry: mutex_enter(&dev->l2ad_mtx); /* * We have to account for evicted log blocks. Run vdev_space_update() * on log blocks whose offset (in bytes) is before the evicted offset * (in bytes) by searching in the list of pointers to log blocks * present in the L2ARC device. */ for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf; lb_ptr_buf = lb_ptr_buf_prev) { lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf); /* L2BLK_GET_PSIZE returns aligned size for log blocks */ uint64_t asize = L2BLK_GET_PSIZE( (lb_ptr_buf->lb_ptr)->lbp_prop); /* * We don't worry about log blocks left behind (ie * lbp_payload_start < l2ad_hand) because l2arc_write_buffers() * will never write more than l2arc_evict() evicts. */ if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) { break; } else { vdev_space_update(vd, -asize, 0, 0); ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize); ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count); zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf); zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf); list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf); kmem_free(lb_ptr_buf->lb_ptr, sizeof (l2arc_log_blkptr_t)); kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t)); } } for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { hdr_prev = list_prev(buflist, hdr); ASSERT(!HDR_EMPTY(hdr)); hash_lock = HDR_LOCK(hdr); /* * We cannot use mutex_enter or else we can deadlock * with l2arc_write_buffers (due to swapping the order * the hash lock and l2ad_mtx are taken). */ if (!mutex_tryenter(hash_lock)) { /* * Missed the hash lock. Retry. */ ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); mutex_exit(&dev->l2ad_mtx); mutex_enter(hash_lock); mutex_exit(hash_lock); goto retry; } /* * A header can't be on this list if it doesn't have L2 header. */ ASSERT(HDR_HAS_L2HDR(hdr)); /* Ensure this header has finished being written. */ ASSERT(!HDR_L2_WRITING(hdr)); ASSERT(!HDR_L2_WRITE_HEAD(hdr)); if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict || hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { /* * We've evicted to the target address, * or the end of the device. */ mutex_exit(hash_lock); break; } if (!HDR_HAS_L1HDR(hdr)) { ASSERT(!HDR_L2_READING(hdr)); /* * This doesn't exist in the ARC. Destroy. * arc_hdr_destroy() will call list_remove() * and decrement arcstat_l2_lsize. */ arc_change_state(arc_anon, hdr, hash_lock); arc_hdr_destroy(hdr); } else { ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); ARCSTAT_BUMP(arcstat_l2_evict_l1cached); /* * Invalidate issued or about to be issued * reads, since we may be about to write * over this location. */ if (HDR_L2_READING(hdr)) { ARCSTAT_BUMP(arcstat_l2_evict_reading); arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); } arc_hdr_l2hdr_destroy(hdr); } mutex_exit(hash_lock); } mutex_exit(&dev->l2ad_mtx); out: /* * We need to check if we evict all buffers, otherwise we may iterate * unnecessarily. */ if (!all && rerun) { /* * Bump device hand to the device start if it is approaching the * end. l2arc_evict() has already evicted ahead for this case. */ dev->l2ad_hand = dev->l2ad_start; dev->l2ad_evict = dev->l2ad_start; dev->l2ad_first = B_FALSE; goto top; } if (!all) { /* * In case of cache device removal (all) the following * assertions may be violated without functional consequences * as the device is about to be removed. */ ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end); if (!dev->l2ad_first) ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict); } } /* * Handle any abd transforms that might be required for writing to the L2ARC. * If successful, this function will always return an abd with the data * transformed as it is on disk in a new abd of asize bytes. */ static int l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize, abd_t **abd_out) { int ret; void *tmp = NULL; abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd; enum zio_compress compress = HDR_GET_COMPRESS(hdr); uint64_t psize = HDR_GET_PSIZE(hdr); uint64_t size = arc_hdr_size(hdr); boolean_t ismd = HDR_ISTYPE_METADATA(hdr); boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); dsl_crypto_key_t *dck = NULL; uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 }; boolean_t no_crypt = B_FALSE; ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) || HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize); ASSERT3U(psize, <=, asize); /* * If this data simply needs its own buffer, we simply allocate it * and copy the data. This may be done to eliminate a dependency on a * shared buffer or to reallocate the buffer to match asize. */ if (HDR_HAS_RABD(hdr) && asize != psize) { ASSERT3U(asize, >=, psize); to_write = abd_alloc_for_io(asize, ismd); abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize); if (psize != asize) abd_zero_off(to_write, psize, asize - psize); goto out; } if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) && !HDR_ENCRYPTED(hdr)) { ASSERT3U(size, ==, psize); to_write = abd_alloc_for_io(asize, ismd); abd_copy(to_write, hdr->b_l1hdr.b_pabd, size); if (size != asize) abd_zero_off(to_write, size, asize - size); goto out; } if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) { cabd = abd_alloc_for_io(asize, ismd); tmp = abd_borrow_buf(cabd, asize); psize = zio_compress_data(compress, to_write, tmp, size, hdr->b_complevel); if (psize >= size) { abd_return_buf(cabd, tmp, asize); HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); to_write = cabd; abd_copy(to_write, hdr->b_l1hdr.b_pabd, size); if (size != asize) abd_zero_off(to_write, size, asize - size); goto encrypt; } ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr)); if (psize < asize) bzero((char *)tmp + psize, asize - psize); psize = HDR_GET_PSIZE(hdr); abd_return_buf_copy(cabd, tmp, asize); to_write = cabd; } encrypt: if (HDR_ENCRYPTED(hdr)) { eabd = abd_alloc_for_io(asize, ismd); /* * If the dataset was disowned before the buffer * made it to this point, the key to re-encrypt * it won't be available. In this case we simply * won't write the buffer to the L2ARC. */ ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj, FTAG, &dck); if (ret != 0) goto error; ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key, hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd, &no_crypt); if (ret != 0) goto error; if (no_crypt) abd_copy(eabd, to_write, psize); if (psize != asize) abd_zero_off(eabd, psize, asize - psize); /* assert that the MAC we got here matches the one we saved */ ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN)); spa_keystore_dsl_key_rele(spa, dck, FTAG); if (to_write == cabd) abd_free(cabd); to_write = eabd; } out: ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd); *abd_out = to_write; return (0); error: if (dck != NULL) spa_keystore_dsl_key_rele(spa, dck, FTAG); if (cabd != NULL) abd_free(cabd); if (eabd != NULL) abd_free(eabd); *abd_out = NULL; return (ret); } static void l2arc_blk_fetch_done(zio_t *zio) { l2arc_read_callback_t *cb; cb = zio->io_private; if (cb->l2rcb_abd != NULL) abd_free(cb->l2rcb_abd); kmem_free(cb, sizeof (l2arc_read_callback_t)); } /* * Find and write ARC buffers to the L2ARC device. * * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid * for reading until they have completed writing. * The headroom_boost is an in-out parameter used to maintain headroom boost * state between calls to this function. * * Returns the number of bytes actually written (which may be smaller than * the delta by which the device hand has changed due to alignment and the * writing of log blocks). */ static uint64_t l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) { arc_buf_hdr_t *hdr, *hdr_prev, *head; uint64_t write_asize, write_psize, write_lsize, headroom; boolean_t full; l2arc_write_callback_t *cb = NULL; zio_t *pio, *wzio; uint64_t guid = spa_load_guid(spa); l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; ASSERT3P(dev->l2ad_vdev, !=, NULL); pio = NULL; write_lsize = write_asize = write_psize = 0; full = B_FALSE; head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); /* * Copy buffers for L2ARC writing. */ for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) { /* * If pass == 1 or 3, we cache MRU metadata and data * respectively. */ if (l2arc_mfuonly) { if (pass == 1 || pass == 3) continue; } multilist_sublist_t *mls = l2arc_sublist_lock(pass); uint64_t passed_sz = 0; VERIFY3P(mls, !=, NULL); /* * L2ARC fast warmup. * * Until the ARC is warm and starts to evict, read from the * head of the ARC lists rather than the tail. */ if (arc_warm == B_FALSE) hdr = multilist_sublist_head(mls); else hdr = multilist_sublist_tail(mls); headroom = target_sz * l2arc_headroom; if (zfs_compressed_arc_enabled) headroom = (headroom * l2arc_headroom_boost) / 100; for (; hdr; hdr = hdr_prev) { kmutex_t *hash_lock; abd_t *to_write = NULL; if (arc_warm == B_FALSE) hdr_prev = multilist_sublist_next(mls, hdr); else hdr_prev = multilist_sublist_prev(mls, hdr); hash_lock = HDR_LOCK(hdr); if (!mutex_tryenter(hash_lock)) { /* * Skip this buffer rather than waiting. */ continue; } passed_sz += HDR_GET_LSIZE(hdr); if (l2arc_headroom != 0 && passed_sz > headroom) { /* * Searched too far. */ mutex_exit(hash_lock); break; } if (!l2arc_write_eligible(guid, hdr)) { mutex_exit(hash_lock); continue; } /* * We rely on the L1 portion of the header below, so * it's invalid for this header to have been evicted out * of the ghost cache, prior to being written out. The * ARC_FLAG_L2_WRITING bit ensures this won't happen. */ ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); ASSERT3U(arc_hdr_size(hdr), >, 0); ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); uint64_t psize = HDR_GET_PSIZE(hdr); uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize); if ((write_asize + asize) > target_sz) { full = B_TRUE; mutex_exit(hash_lock); break; } /* * We rely on the L1 portion of the header below, so * it's invalid for this header to have been evicted out * of the ghost cache, prior to being written out. The * ARC_FLAG_L2_WRITING bit ensures this won't happen. */ arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING); ASSERT(HDR_HAS_L1HDR(hdr)); ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); ASSERT3U(arc_hdr_size(hdr), >, 0); /* * If this header has b_rabd, we can use this since it * must always match the data exactly as it exists on * disk. Otherwise, the L2ARC can normally use the * hdr's data, but if we're sharing data between the * hdr and one of its bufs, L2ARC needs its own copy of * the data so that the ZIO below can't race with the * buf consumer. To ensure that this copy will be * available for the lifetime of the ZIO and be cleaned * up afterwards, we add it to the l2arc_free_on_write * queue. If we need to apply any transforms to the * data (compression, encryption) we will also need the * extra buffer. */ if (HDR_HAS_RABD(hdr) && psize == asize) { to_write = hdr->b_crypt_hdr.b_rabd; } else if ((HDR_COMPRESSION_ENABLED(hdr) || HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) && !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) && psize == asize) { to_write = hdr->b_l1hdr.b_pabd; } else { int ret; arc_buf_contents_t type = arc_buf_type(hdr); ret = l2arc_apply_transforms(spa, hdr, asize, &to_write); if (ret != 0) { arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); mutex_exit(hash_lock); continue; } l2arc_free_abd_on_write(to_write, asize, type); } if (pio == NULL) { /* * Insert a dummy header on the buflist so * l2arc_write_done() can find where the * write buffers begin without searching. */ mutex_enter(&dev->l2ad_mtx); list_insert_head(&dev->l2ad_buflist, head); mutex_exit(&dev->l2ad_mtx); cb = kmem_alloc( sizeof (l2arc_write_callback_t), KM_SLEEP); cb->l2wcb_dev = dev; cb->l2wcb_head = head; /* * Create a list to save allocated abd buffers * for l2arc_log_blk_commit(). */ list_create(&cb->l2wcb_abd_list, sizeof (l2arc_lb_abd_buf_t), offsetof(l2arc_lb_abd_buf_t, node)); pio = zio_root(spa, l2arc_write_done, cb, ZIO_FLAG_CANFAIL); } hdr->b_l2hdr.b_dev = dev; hdr->b_l2hdr.b_hits = 0; hdr->b_l2hdr.b_daddr = dev->l2ad_hand; hdr->b_l2hdr.b_arcs_state = hdr->b_l1hdr.b_state->arcs_state; arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR); mutex_enter(&dev->l2ad_mtx); list_insert_head(&dev->l2ad_buflist, hdr); mutex_exit(&dev->l2ad_mtx); (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); wzio = zio_write_phys(pio, dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, asize, to_write, ZIO_CHECKSUM_OFF, NULL, hdr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE); write_lsize += HDR_GET_LSIZE(hdr); DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio); write_psize += psize; write_asize += asize; dev->l2ad_hand += asize; l2arc_hdr_arcstats_increment(hdr); vdev_space_update(dev->l2ad_vdev, asize, 0, 0); mutex_exit(hash_lock); /* * Append buf info to current log and commit if full. * arcstat_l2_{size,asize} kstats are updated * internally. */ if (l2arc_log_blk_insert(dev, hdr)) l2arc_log_blk_commit(dev, pio, cb); zio_nowait(wzio); } multilist_sublist_unlock(mls); if (full == B_TRUE) break; } /* No buffers selected for writing? */ if (pio == NULL) { ASSERT0(write_lsize); ASSERT(!HDR_HAS_L1HDR(head)); kmem_cache_free(hdr_l2only_cache, head); /* * Although we did not write any buffers l2ad_evict may * have advanced. */ if (dev->l2ad_evict != l2dhdr->dh_evict) l2arc_dev_hdr_update(dev); return (0); } if (!dev->l2ad_first) ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict); ASSERT3U(write_asize, <=, target_sz); ARCSTAT_BUMP(arcstat_l2_writes_sent); ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); dev->l2ad_writing = B_TRUE; (void) zio_wait(pio); dev->l2ad_writing = B_FALSE; /* * Update the device header after the zio completes as * l2arc_write_done() may have updated the memory holding the log block * pointers in the device header. */ l2arc_dev_hdr_update(dev); return (write_asize); } static boolean_t l2arc_hdr_limit_reached(void) { int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size); return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) || (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100)); } /* * This thread feeds the L2ARC at regular intervals. This is the beating * heart of the L2ARC. */ /* ARGSUSED */ static void l2arc_feed_thread(void *unused) { callb_cpr_t cpr; l2arc_dev_t *dev; spa_t *spa; uint64_t size, wrote; clock_t begin, next = ddi_get_lbolt(); fstrans_cookie_t cookie; CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); mutex_enter(&l2arc_feed_thr_lock); cookie = spl_fstrans_mark(); while (l2arc_thread_exit == 0) { CALLB_CPR_SAFE_BEGIN(&cpr); (void) cv_timedwait_idle(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, next); CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); next = ddi_get_lbolt() + hz; /* * Quick check for L2ARC devices. */ mutex_enter(&l2arc_dev_mtx); if (l2arc_ndev == 0) { mutex_exit(&l2arc_dev_mtx); continue; } mutex_exit(&l2arc_dev_mtx); begin = ddi_get_lbolt(); /* * This selects the next l2arc device to write to, and in * doing so the next spa to feed from: dev->l2ad_spa. This * will return NULL if there are now no l2arc devices or if * they are all faulted. * * If a device is returned, its spa's config lock is also * held to prevent device removal. l2arc_dev_get_next() * will grab and release l2arc_dev_mtx. */ if ((dev = l2arc_dev_get_next()) == NULL) continue; spa = dev->l2ad_spa; ASSERT3P(spa, !=, NULL); /* * If the pool is read-only then force the feed thread to * sleep a little longer. */ if (!spa_writeable(spa)) { next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; spa_config_exit(spa, SCL_L2ARC, dev); continue; } /* * Avoid contributing to memory pressure. */ if (l2arc_hdr_limit_reached()) { ARCSTAT_BUMP(arcstat_l2_abort_lowmem); spa_config_exit(spa, SCL_L2ARC, dev); continue; } ARCSTAT_BUMP(arcstat_l2_feeds); size = l2arc_write_size(dev); /* * Evict L2ARC buffers that will be overwritten. */ l2arc_evict(dev, size, B_FALSE); /* * Write ARC buffers. */ wrote = l2arc_write_buffers(spa, dev, size); /* * Calculate interval between writes. */ next = l2arc_write_interval(begin, size, wrote); spa_config_exit(spa, SCL_L2ARC, dev); } spl_fstrans_unmark(cookie); l2arc_thread_exit = 0; cv_broadcast(&l2arc_feed_thr_cv); CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ thread_exit(); } boolean_t l2arc_vdev_present(vdev_t *vd) { return (l2arc_vdev_get(vd) != NULL); } /* * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if * the vdev_t isn't an L2ARC device. */ l2arc_dev_t * l2arc_vdev_get(vdev_t *vd) { l2arc_dev_t *dev; mutex_enter(&l2arc_dev_mtx); for (dev = list_head(l2arc_dev_list); dev != NULL; dev = list_next(l2arc_dev_list, dev)) { if (dev->l2ad_vdev == vd) break; } mutex_exit(&l2arc_dev_mtx); return (dev); } static void l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen) { l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize; spa_t *spa = dev->l2ad_spa; /* * The L2ARC has to hold at least the payload of one log block for * them to be restored (persistent L2ARC). The payload of a log block * depends on the amount of its log entries. We always write log blocks * with 1022 entries. How many of them are committed or restored depends * on the size of the L2ARC device. Thus the maximum payload of * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device * is less than that, we reduce the amount of committed and restored * log entries per block so as to enable persistence. */ if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) { dev->l2ad_log_entries = 0; } else { dev->l2ad_log_entries = MIN((dev->l2ad_end - dev->l2ad_start) >> SPA_MAXBLOCKSHIFT, L2ARC_LOG_BLK_MAX_ENTRIES); } /* * Read the device header, if an error is returned do not rebuild L2ARC. */ if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) { /* * If we are onlining a cache device (vdev_reopen) that was * still present (l2arc_vdev_present()) and rebuild is enabled, * we should evict all ARC buffers and pointers to log blocks * and reclaim their space before restoring its contents to * L2ARC. */ if (reopen) { if (!l2arc_rebuild_enabled) { return; } else { l2arc_evict(dev, 0, B_TRUE); /* start a new log block */ dev->l2ad_log_ent_idx = 0; dev->l2ad_log_blk_payload_asize = 0; dev->l2ad_log_blk_payload_start = 0; } } /* * Just mark the device as pending for a rebuild. We won't * be starting a rebuild in line here as it would block pool * import. Instead spa_load_impl will hand that off to an * async task which will call l2arc_spa_rebuild_start. */ dev->l2ad_rebuild = B_TRUE; } else if (spa_writeable(spa)) { /* * In this case TRIM the whole device if l2arc_trim_ahead > 0, * otherwise create a new header. We zero out the memory holding * the header to reset dh_start_lbps. If we TRIM the whole * device the new header will be written by * vdev_trim_l2arc_thread() at the end of the TRIM to update the * trim_state in the header too. When reading the header, if * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0 * we opt to TRIM the whole device again. */ if (l2arc_trim_ahead > 0) { dev->l2ad_trim_all = B_TRUE; } else { bzero(l2dhdr, l2dhdr_asize); l2arc_dev_hdr_update(dev); } } } /* * Add a vdev for use by the L2ARC. By this point the spa has already * validated the vdev and opened it. */ void l2arc_add_vdev(spa_t *spa, vdev_t *vd) { l2arc_dev_t *adddev; uint64_t l2dhdr_asize; ASSERT(!l2arc_vdev_present(vd)); /* * Create a new l2arc device entry. */ adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); adddev->l2ad_spa = spa; adddev->l2ad_vdev = vd; /* leave extra size for an l2arc device header */ l2dhdr_asize = adddev->l2ad_dev_hdr_asize = MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift); adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize; adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end); adddev->l2ad_hand = adddev->l2ad_start; adddev->l2ad_evict = adddev->l2ad_start; adddev->l2ad_first = B_TRUE; adddev->l2ad_writing = B_FALSE; adddev->l2ad_trim_all = B_FALSE; list_link_init(&adddev->l2ad_node); adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP); mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); /* * This is a list of all ARC buffers that are still valid on the * device. */ list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); /* * This is a list of pointers to log blocks that are still present * on the device. */ list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t), offsetof(l2arc_lb_ptr_buf_t, node)); vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); zfs_refcount_create(&adddev->l2ad_alloc); zfs_refcount_create(&adddev->l2ad_lb_asize); zfs_refcount_create(&adddev->l2ad_lb_count); /* * Decide if dev is eligible for L2ARC rebuild or whole device * trimming. This has to happen before the device is added in the * cache device list and l2arc_dev_mtx is released. Otherwise * l2arc_feed_thread() might already start writing on the * device. */ l2arc_rebuild_dev(adddev, B_FALSE); /* * Add device to global list */ mutex_enter(&l2arc_dev_mtx); list_insert_head(l2arc_dev_list, adddev); atomic_inc_64(&l2arc_ndev); mutex_exit(&l2arc_dev_mtx); } /* * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen() * in case of onlining a cache device. */ void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen) { l2arc_dev_t *dev = NULL; dev = l2arc_vdev_get(vd); ASSERT3P(dev, !=, NULL); /* * In contrast to l2arc_add_vdev() we do not have to worry about * l2arc_feed_thread() invalidating previous content when onlining a * cache device. The device parameters (l2ad*) are not cleared when * offlining the device and writing new buffers will not invalidate * all previous content. In worst case only buffers that have not had * their log block written to the device will be lost. * When onlining the cache device (ie offline->online without exporting * the pool in between) this happens: * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev() * | | * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild * is set to B_TRUE we might write additional buffers to the device. */ l2arc_rebuild_dev(dev, reopen); } /* * Remove a vdev from the L2ARC. */ void l2arc_remove_vdev(vdev_t *vd) { l2arc_dev_t *remdev = NULL; /* * Find the device by vdev */ remdev = l2arc_vdev_get(vd); ASSERT3P(remdev, !=, NULL); /* * Cancel any ongoing or scheduled rebuild. */ mutex_enter(&l2arc_rebuild_thr_lock); if (remdev->l2ad_rebuild_began == B_TRUE) { remdev->l2ad_rebuild_cancel = B_TRUE; while (remdev->l2ad_rebuild == B_TRUE) cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock); } mutex_exit(&l2arc_rebuild_thr_lock); /* * Remove device from global list */ mutex_enter(&l2arc_dev_mtx); list_remove(l2arc_dev_list, remdev); l2arc_dev_last = NULL; /* may have been invalidated */ atomic_dec_64(&l2arc_ndev); mutex_exit(&l2arc_dev_mtx); /* * Clear all buflists and ARC references. L2ARC device flush. */ l2arc_evict(remdev, 0, B_TRUE); list_destroy(&remdev->l2ad_buflist); ASSERT(list_is_empty(&remdev->l2ad_lbptr_list)); list_destroy(&remdev->l2ad_lbptr_list); mutex_destroy(&remdev->l2ad_mtx); zfs_refcount_destroy(&remdev->l2ad_alloc); zfs_refcount_destroy(&remdev->l2ad_lb_asize); zfs_refcount_destroy(&remdev->l2ad_lb_count); kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize); vmem_free(remdev, sizeof (l2arc_dev_t)); } void l2arc_init(void) { l2arc_thread_exit = 0; l2arc_ndev = 0; mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL); cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL); mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); l2arc_dev_list = &L2ARC_dev_list; l2arc_free_on_write = &L2ARC_free_on_write; list_create(l2arc_dev_list, sizeof (l2arc_dev_t), offsetof(l2arc_dev_t, l2ad_node)); list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), offsetof(l2arc_data_free_t, l2df_list_node)); } void l2arc_fini(void) { mutex_destroy(&l2arc_feed_thr_lock); cv_destroy(&l2arc_feed_thr_cv); mutex_destroy(&l2arc_rebuild_thr_lock); cv_destroy(&l2arc_rebuild_thr_cv); mutex_destroy(&l2arc_dev_mtx); mutex_destroy(&l2arc_free_on_write_mtx); list_destroy(l2arc_dev_list); list_destroy(l2arc_free_on_write); } void l2arc_start(void) { if (!(spa_mode_global & SPA_MODE_WRITE)) return; (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, TS_RUN, defclsyspri); } void l2arc_stop(void) { if (!(spa_mode_global & SPA_MODE_WRITE)) return; mutex_enter(&l2arc_feed_thr_lock); cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ l2arc_thread_exit = 1; while (l2arc_thread_exit != 0) cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); mutex_exit(&l2arc_feed_thr_lock); } /* * Punches out rebuild threads for the L2ARC devices in a spa. This should * be called after pool import from the spa async thread, since starting * these threads directly from spa_import() will make them part of the * "zpool import" context and delay process exit (and thus pool import). */ void l2arc_spa_rebuild_start(spa_t *spa) { ASSERT(MUTEX_HELD(&spa_namespace_lock)); /* * Locate the spa's l2arc devices and kick off rebuild threads. */ for (int i = 0; i < spa->spa_l2cache.sav_count; i++) { l2arc_dev_t *dev = l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]); if (dev == NULL) { /* Don't attempt a rebuild if the vdev is UNAVAIL */ continue; } mutex_enter(&l2arc_rebuild_thr_lock); if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) { dev->l2ad_rebuild_began = B_TRUE; (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread, dev, 0, &p0, TS_RUN, minclsyspri); } mutex_exit(&l2arc_rebuild_thr_lock); } } /* * Main entry point for L2ARC rebuilding. */ static void l2arc_dev_rebuild_thread(void *arg) { l2arc_dev_t *dev = arg; VERIFY(!dev->l2ad_rebuild_cancel); VERIFY(dev->l2ad_rebuild); (void) l2arc_rebuild(dev); mutex_enter(&l2arc_rebuild_thr_lock); dev->l2ad_rebuild_began = B_FALSE; dev->l2ad_rebuild = B_FALSE; mutex_exit(&l2arc_rebuild_thr_lock); thread_exit(); } /* * This function implements the actual L2ARC metadata rebuild. It: * starts reading the log block chain and restores each block's contents * to memory (reconstructing arc_buf_hdr_t's). * * Operation stops under any of the following conditions: * * 1) We reach the end of the log block chain. * 2) We encounter *any* error condition (cksum errors, io errors) */ static int l2arc_rebuild(l2arc_dev_t *dev) { vdev_t *vd = dev->l2ad_vdev; spa_t *spa = vd->vdev_spa; int err = 0; l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; l2arc_log_blk_phys_t *this_lb, *next_lb; zio_t *this_io = NULL, *next_io = NULL; l2arc_log_blkptr_t lbps[2]; l2arc_lb_ptr_buf_t *lb_ptr_buf; boolean_t lock_held; this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP); next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP); /* * We prevent device removal while issuing reads to the device, * then during the rebuilding phases we drop this lock again so * that a spa_unload or device remove can be initiated - this is * safe, because the spa will signal us to stop before removing * our device and wait for us to stop. */ spa_config_enter(spa, SCL_L2ARC, vd, RW_READER); lock_held = B_TRUE; /* * Retrieve the persistent L2ARC device state. * L2BLK_GET_PSIZE returns aligned size for log blocks. */ dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start); dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr + L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop), dev->l2ad_start); dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST); vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time; vd->vdev_trim_state = l2dhdr->dh_trim_state; /* * In case the zfs module parameter l2arc_rebuild_enabled is false * we do not start the rebuild process. */ if (!l2arc_rebuild_enabled) goto out; /* Prepare the rebuild process */ bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps)); /* Start the rebuild process */ for (;;) { if (!l2arc_log_blkptr_valid(dev, &lbps[0])) break; if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1], this_lb, next_lb, this_io, &next_io)) != 0) goto out; /* * Our memory pressure valve. If the system is running low * on memory, rather than swamping memory with new ARC buf * hdrs, we opt not to rebuild the L2ARC. At this point, * however, we have already set up our L2ARC dev to chain in * new metadata log blocks, so the user may choose to offline/ * online the L2ARC dev at a later time (or re-import the pool) * to reconstruct it (when there's less memory pressure). */ if (l2arc_hdr_limit_reached()) { ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem); cmn_err(CE_NOTE, "System running low on memory, " "aborting L2ARC rebuild."); err = SET_ERROR(ENOMEM); goto out; } spa_config_exit(spa, SCL_L2ARC, vd); lock_held = B_FALSE; /* * Now that we know that the next_lb checks out alright, we * can start reconstruction from this log block. * L2BLK_GET_PSIZE returns aligned size for log blocks. */ uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop); l2arc_log_blk_restore(dev, this_lb, asize); /* * log block restored, include its pointer in the list of * pointers to log blocks present in the L2ARC device. */ lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP); lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP); bcopy(&lbps[0], lb_ptr_buf->lb_ptr, sizeof (l2arc_log_blkptr_t)); mutex_enter(&dev->l2ad_mtx); list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf); ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize); ARCSTAT_BUMP(arcstat_l2_log_blk_count); zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf); zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf); mutex_exit(&dev->l2ad_mtx); vdev_space_update(vd, asize, 0, 0); /* * Protection against loops of log blocks: * * l2ad_hand l2ad_evict * V V * l2ad_start |=======================================| l2ad_end * -----|||----|||---|||----||| * (3) (2) (1) (0) * ---|||---|||----|||---||| * (7) (6) (5) (4) * * In this situation the pointer of log block (4) passes * l2arc_log_blkptr_valid() but the log block should not be * restored as it is overwritten by the payload of log block * (0). Only log blocks (0)-(3) should be restored. We check * whether l2ad_evict lies in between the payload starting * offset of the next log block (lbps[1].lbp_payload_start) * and the payload starting offset of the present log block * (lbps[0].lbp_payload_start). If true and this isn't the * first pass, we are looping from the beginning and we should * stop. */ if (l2arc_range_check_overlap(lbps[1].lbp_payload_start, lbps[0].lbp_payload_start, dev->l2ad_evict) && !dev->l2ad_first) goto out; cond_resched(); for (;;) { mutex_enter(&l2arc_rebuild_thr_lock); if (dev->l2ad_rebuild_cancel) { dev->l2ad_rebuild = B_FALSE; cv_signal(&l2arc_rebuild_thr_cv); mutex_exit(&l2arc_rebuild_thr_lock); err = SET_ERROR(ECANCELED); goto out; } mutex_exit(&l2arc_rebuild_thr_lock); if (spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) { lock_held = B_TRUE; break; } /* * L2ARC config lock held by somebody in writer, * possibly due to them trying to remove us. They'll * likely to want us to shut down, so after a little * delay, we check l2ad_rebuild_cancel and retry * the lock again. */ delay(1); } /* * Continue with the next log block. */ lbps[0] = lbps[1]; lbps[1] = this_lb->lb_prev_lbp; PTR_SWAP(this_lb, next_lb); this_io = next_io; next_io = NULL; } if (this_io != NULL) l2arc_log_blk_fetch_abort(this_io); out: if (next_io != NULL) l2arc_log_blk_fetch_abort(next_io); vmem_free(this_lb, sizeof (*this_lb)); vmem_free(next_lb, sizeof (*next_lb)); if (!l2arc_rebuild_enabled) { spa_history_log_internal(spa, "L2ARC rebuild", NULL, "disabled"); } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) { ARCSTAT_BUMP(arcstat_l2_rebuild_success); spa_history_log_internal(spa, "L2ARC rebuild", NULL, "successful, restored %llu blocks", (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count)); } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) { /* * No error but also nothing restored, meaning the lbps array * in the device header points to invalid/non-present log * blocks. Reset the header. */ spa_history_log_internal(spa, "L2ARC rebuild", NULL, "no valid log blocks"); bzero(l2dhdr, dev->l2ad_dev_hdr_asize); l2arc_dev_hdr_update(dev); } else if (err == ECANCELED) { /* * In case the rebuild was canceled do not log to spa history * log as the pool may be in the process of being removed. */ zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks", (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count)); } else if (err != 0) { spa_history_log_internal(spa, "L2ARC rebuild", NULL, "aborted, restored %llu blocks", (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count)); } if (lock_held) spa_config_exit(spa, SCL_L2ARC, vd); return (err); } /* * Attempts to read the device header on the provided L2ARC device and writes * it to `hdr'. On success, this function returns 0, otherwise the appropriate * error code is returned. */ static int l2arc_dev_hdr_read(l2arc_dev_t *dev) { int err; uint64_t guid; l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize; abd_t *abd; guid = spa_guid(dev->l2ad_vdev->vdev_spa); abd = abd_get_from_buf(l2dhdr, l2dhdr_asize); err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev, VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_SPECULATIVE, B_FALSE)); abd_free(abd); if (err != 0) { ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors); zfs_dbgmsg("L2ARC IO error (%d) while reading device header, " "vdev guid: %llu", err, (u_longlong_t)dev->l2ad_vdev->vdev_guid); return (err); } if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC)) byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr)); if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC || l2dhdr->dh_spa_guid != guid || l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid || l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION || l2dhdr->dh_log_entries != dev->l2ad_log_entries || l2dhdr->dh_end != dev->l2ad_end || !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end, l2dhdr->dh_evict) || (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE && l2arc_trim_ahead > 0)) { /* * Attempt to rebuild a device containing no actual dev hdr * or containing a header from some other pool or from another * version of persistent L2ARC. */ ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported); return (SET_ERROR(ENOTSUP)); } return (0); } /* * Reads L2ARC log blocks from storage and validates their contents. * * This function implements a simple fetcher to make sure that while * we're processing one buffer the L2ARC is already fetching the next * one in the chain. * * The arguments this_lp and next_lp point to the current and next log block * address in the block chain. Similarly, this_lb and next_lb hold the * l2arc_log_blk_phys_t's of the current and next L2ARC blk. * * The `this_io' and `next_io' arguments are used for block fetching. * When issuing the first blk IO during rebuild, you should pass NULL for * `this_io'. This function will then issue a sync IO to read the block and * also issue an async IO to fetch the next block in the block chain. The * fetched IO is returned in `next_io'. On subsequent calls to this * function, pass the value returned in `next_io' from the previous call * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO. * Prior to the call, you should initialize your `next_io' pointer to be * NULL. If no fetch IO was issued, the pointer is left set at NULL. * * On success, this function returns 0, otherwise it returns an appropriate * error code. On error the fetching IO is aborted and cleared before * returning from this function. Therefore, if we return `success', the * caller can assume that we have taken care of cleanup of fetch IOs. */ static int l2arc_log_blk_read(l2arc_dev_t *dev, const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp, l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb, zio_t *this_io, zio_t **next_io) { int err = 0; zio_cksum_t cksum; abd_t *abd = NULL; uint64_t asize; ASSERT(this_lbp != NULL && next_lbp != NULL); ASSERT(this_lb != NULL && next_lb != NULL); ASSERT(next_io != NULL && *next_io == NULL); ASSERT(l2arc_log_blkptr_valid(dev, this_lbp)); /* * Check to see if we have issued the IO for this log block in a * previous run. If not, this is the first call, so issue it now. */ if (this_io == NULL) { this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp, this_lb); } /* * Peek to see if we can start issuing the next IO immediately. */ if (l2arc_log_blkptr_valid(dev, next_lbp)) { /* * Start issuing IO for the next log block early - this * should help keep the L2ARC device busy while we * decompress and restore this log block. */ *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp, next_lb); } /* Wait for the IO to read this log block to complete */ if ((err = zio_wait(this_io)) != 0) { ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors); zfs_dbgmsg("L2ARC IO error (%d) while reading log block, " "offset: %llu, vdev guid: %llu", err, (u_longlong_t)this_lbp->lbp_daddr, (u_longlong_t)dev->l2ad_vdev->vdev_guid); goto cleanup; } /* * Make sure the buffer checks out. * L2BLK_GET_PSIZE returns aligned size for log blocks. */ asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop); fletcher_4_native(this_lb, asize, NULL, &cksum); if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) { ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors); zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, " "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu", (u_longlong_t)this_lbp->lbp_daddr, (u_longlong_t)dev->l2ad_vdev->vdev_guid, (u_longlong_t)dev->l2ad_hand, (u_longlong_t)dev->l2ad_evict); err = SET_ERROR(ECKSUM); goto cleanup; } /* Now we can take our time decoding this buffer */ switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) { case ZIO_COMPRESS_OFF: break; case ZIO_COMPRESS_LZ4: abd = abd_alloc_for_io(asize, B_TRUE); abd_copy_from_buf_off(abd, this_lb, 0, asize); if ((err = zio_decompress_data( L2BLK_GET_COMPRESS((this_lbp)->lbp_prop), abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) { err = SET_ERROR(EINVAL); goto cleanup; } break; default: err = SET_ERROR(EINVAL); goto cleanup; } if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC)) byteswap_uint64_array(this_lb, sizeof (*this_lb)); if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) { err = SET_ERROR(EINVAL); goto cleanup; } cleanup: /* Abort an in-flight fetch I/O in case of error */ if (err != 0 && *next_io != NULL) { l2arc_log_blk_fetch_abort(*next_io); *next_io = NULL; } if (abd != NULL) abd_free(abd); return (err); } /* * Restores the payload of a log block to ARC. This creates empty ARC hdr * entries which only contain an l2arc hdr, essentially restoring the * buffers to their L2ARC evicted state. This function also updates space * usage on the L2ARC vdev to make sure it tracks restored buffers. */ static void l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb, uint64_t lb_asize) { uint64_t size = 0, asize = 0; uint64_t log_entries = dev->l2ad_log_entries; /* * Usually arc_adapt() is called only for data, not headers, but * since we may allocate significant amount of memory here, let ARC * grow its arc_c. */ arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only); for (int i = log_entries - 1; i >= 0; i--) { /* * Restore goes in the reverse temporal direction to preserve * correct temporal ordering of buffers in the l2ad_buflist. * l2arc_hdr_restore also does a list_insert_tail instead of * list_insert_head on the l2ad_buflist: * * LIST l2ad_buflist LIST * HEAD <------ (time) ------ TAIL * direction +-----+-----+-----+-----+-----+ direction * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild * fill +-----+-----+-----+-----+-----+ * ^ ^ * | | * | | * l2arc_feed_thread l2arc_rebuild * will place new bufs here restores bufs here * * During l2arc_rebuild() the device is not used by * l2arc_feed_thread() as dev->l2ad_rebuild is set to true. */ size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop); asize += vdev_psize_to_asize(dev->l2ad_vdev, L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop)); l2arc_hdr_restore(&lb->lb_entries[i], dev); } /* * Record rebuild stats: * size Logical size of restored buffers in the L2ARC * asize Aligned size of restored buffers in the L2ARC */ ARCSTAT_INCR(arcstat_l2_rebuild_size, size); ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize); ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries); ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize); ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize); ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks); } /* * Restores a single ARC buf hdr from a log entry. The ARC buffer is put * into a state indicating that it has been evicted to L2ARC. */ static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev) { arc_buf_hdr_t *hdr, *exists; kmutex_t *hash_lock; arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop); uint64_t asize; /* * Do all the allocation before grabbing any locks, this lets us * sleep if memory is full and we don't have to deal with failed * allocations. */ hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type, dev, le->le_dva, le->le_daddr, L2BLK_GET_PSIZE((le)->le_prop), le->le_birth, L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel, L2BLK_GET_PROTECTED((le)->le_prop), L2BLK_GET_PREFETCH((le)->le_prop), L2BLK_GET_STATE((le)->le_prop)); asize = vdev_psize_to_asize(dev->l2ad_vdev, L2BLK_GET_PSIZE((le)->le_prop)); /* * vdev_space_update() has to be called before arc_hdr_destroy() to * avoid underflow since the latter also calls vdev_space_update(). */ l2arc_hdr_arcstats_increment(hdr); vdev_space_update(dev->l2ad_vdev, asize, 0, 0); mutex_enter(&dev->l2ad_mtx); list_insert_tail(&dev->l2ad_buflist, hdr); (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); mutex_exit(&dev->l2ad_mtx); exists = buf_hash_insert(hdr, &hash_lock); if (exists) { /* Buffer was already cached, no need to restore it. */ arc_hdr_destroy(hdr); /* * If the buffer is already cached, check whether it has * L2ARC metadata. If not, enter them and update the flag. * This is important is case of onlining a cache device, since * we previously evicted all L2ARC metadata from ARC. */ if (!HDR_HAS_L2HDR(exists)) { arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR); exists->b_l2hdr.b_dev = dev; exists->b_l2hdr.b_daddr = le->le_daddr; exists->b_l2hdr.b_arcs_state = L2BLK_GET_STATE((le)->le_prop); mutex_enter(&dev->l2ad_mtx); list_insert_tail(&dev->l2ad_buflist, exists); (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(exists), exists); mutex_exit(&dev->l2ad_mtx); l2arc_hdr_arcstats_increment(exists); vdev_space_update(dev->l2ad_vdev, asize, 0, 0); } ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached); } mutex_exit(hash_lock); } /* * Starts an asynchronous read IO to read a log block. This is used in log * block reconstruction to start reading the next block before we are done * decoding and reconstructing the current block, to keep the l2arc device * nice and hot with read IO to process. * The returned zio will contain a newly allocated memory buffers for the IO * data which should then be freed by the caller once the zio is no longer * needed (i.e. due to it having completed). If you wish to abort this * zio, you should do so using l2arc_log_blk_fetch_abort, which takes * care of disposing of the allocated buffers correctly. */ static zio_t * l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp, l2arc_log_blk_phys_t *lb) { uint32_t asize; zio_t *pio; l2arc_read_callback_t *cb; /* L2BLK_GET_PSIZE returns aligned size for log blocks */ asize = L2BLK_GET_PSIZE((lbp)->lbp_prop); ASSERT(asize <= sizeof (l2arc_log_blk_phys_t)); cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP); cb->l2rcb_abd = abd_get_from_buf(lb, asize); pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY); (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize, cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE)); return (pio); } /* * Aborts a zio returned from l2arc_log_blk_fetch and frees the data * buffers allocated for it. */ static void l2arc_log_blk_fetch_abort(zio_t *zio) { (void) zio_wait(zio); } /* * Creates a zio to update the device header on an l2arc device. */ void l2arc_dev_hdr_update(l2arc_dev_t *dev) { l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize; abd_t *abd; int err; VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER)); l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC; l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION; l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa); l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid; l2dhdr->dh_log_entries = dev->l2ad_log_entries; l2dhdr->dh_evict = dev->l2ad_evict; l2dhdr->dh_start = dev->l2ad_start; l2dhdr->dh_end = dev->l2ad_end; l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize); l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count); l2dhdr->dh_flags = 0; l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time; l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state; if (dev->l2ad_first) l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST; abd = abd_get_from_buf(l2dhdr, l2dhdr_asize); err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev, VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE)); abd_free(abd); if (err != 0) { zfs_dbgmsg("L2ARC IO error (%d) while writing device header, " "vdev guid: %llu", err, (u_longlong_t)dev->l2ad_vdev->vdev_guid); } } /* * Commits a log block to the L2ARC device. This routine is invoked from * l2arc_write_buffers when the log block fills up. * This function allocates some memory to temporarily hold the serialized * buffer to be written. This is then released in l2arc_write_done. */ static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb) { l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk; l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; uint64_t psize, asize; zio_t *wzio; l2arc_lb_abd_buf_t *abd_buf; uint8_t *tmpbuf; l2arc_lb_ptr_buf_t *lb_ptr_buf; VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries); tmpbuf = zio_buf_alloc(sizeof (*lb)); abd_buf = zio_buf_alloc(sizeof (*abd_buf)); abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb)); lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP); lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP); /* link the buffer into the block chain */ lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1]; lb->lb_magic = L2ARC_LOG_BLK_MAGIC; /* * l2arc_log_blk_commit() may be called multiple times during a single * l2arc_write_buffers() call. Save the allocated abd buffers in a list * so we can free them in l2arc_write_done() later on. */ list_insert_tail(&cb->l2wcb_abd_list, abd_buf); /* try to compress the buffer */ psize = zio_compress_data(ZIO_COMPRESS_LZ4, abd_buf->abd, tmpbuf, sizeof (*lb), 0); /* a log block is never entirely zero */ ASSERT(psize != 0); asize = vdev_psize_to_asize(dev->l2ad_vdev, psize); ASSERT(asize <= sizeof (*lb)); /* * Update the start log block pointer in the device header to point * to the log block we're about to write. */ l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0]; l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand; l2dhdr->dh_start_lbps[0].lbp_payload_asize = dev->l2ad_log_blk_payload_asize; l2dhdr->dh_start_lbps[0].lbp_payload_start = dev->l2ad_log_blk_payload_start; _NOTE(CONSTCOND) L2BLK_SET_LSIZE( (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb)); L2BLK_SET_PSIZE( (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize); L2BLK_SET_CHECKSUM( (&l2dhdr->dh_start_lbps[0])->lbp_prop, ZIO_CHECKSUM_FLETCHER_4); if (asize < sizeof (*lb)) { /* compression succeeded */ bzero(tmpbuf + psize, asize - psize); L2BLK_SET_COMPRESS( (&l2dhdr->dh_start_lbps[0])->lbp_prop, ZIO_COMPRESS_LZ4); } else { /* compression failed */ bcopy(lb, tmpbuf, sizeof (*lb)); L2BLK_SET_COMPRESS( (&l2dhdr->dh_start_lbps[0])->lbp_prop, ZIO_COMPRESS_OFF); } /* checksum what we're about to write */ fletcher_4_native(tmpbuf, asize, NULL, &l2dhdr->dh_start_lbps[0].lbp_cksum); abd_free(abd_buf->abd); /* perform the write itself */ abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb)); abd_take_ownership_of_buf(abd_buf->abd, B_TRUE); wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand, asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE); DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio); (void) zio_nowait(wzio); dev->l2ad_hand += asize; /* * Include the committed log block's pointer in the list of pointers * to log blocks present in the L2ARC device. */ bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr, sizeof (l2arc_log_blkptr_t)); mutex_enter(&dev->l2ad_mtx); list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf); ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize); ARCSTAT_BUMP(arcstat_l2_log_blk_count); zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf); zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf); mutex_exit(&dev->l2ad_mtx); vdev_space_update(dev->l2ad_vdev, asize, 0, 0); /* bump the kstats */ ARCSTAT_INCR(arcstat_l2_write_bytes, asize); ARCSTAT_BUMP(arcstat_l2_log_blk_writes); ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize); ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, dev->l2ad_log_blk_payload_asize / asize); /* start a new log block */ dev->l2ad_log_ent_idx = 0; dev->l2ad_log_blk_payload_asize = 0; dev->l2ad_log_blk_payload_start = 0; } /* * Validates an L2ARC log block address to make sure that it can be read * from the provided L2ARC device. */ boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp) { /* L2BLK_GET_PSIZE returns aligned size for log blocks */ uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop); uint64_t end = lbp->lbp_daddr + asize - 1; uint64_t start = lbp->lbp_payload_start; boolean_t evicted = B_FALSE; /* * A log block is valid if all of the following conditions are true: * - it fits entirely (including its payload) between l2ad_start and * l2ad_end * - it has a valid size * - neither the log block itself nor part of its payload was evicted * by l2arc_evict(): * * l2ad_hand l2ad_evict * | | lbp_daddr * | start | | end * | | | | | * V V V V V * l2ad_start ============================================ l2ad_end * --------------------------|||| * ^ ^ * | log block * payload */ evicted = l2arc_range_check_overlap(start, end, dev->l2ad_hand) || l2arc_range_check_overlap(start, end, dev->l2ad_evict) || l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) || l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end); return (start >= dev->l2ad_start && end <= dev->l2ad_end && asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) && (!evicted || dev->l2ad_first)); } /* * Inserts ARC buffer header `hdr' into the current L2ARC log block on * the device. The buffer being inserted must be present in L2ARC. * Returns B_TRUE if the L2ARC log block is full and needs to be committed * to L2ARC, or B_FALSE if it still has room for more ARC buffers. */ static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr) { l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk; l2arc_log_ent_phys_t *le; if (dev->l2ad_log_entries == 0) return (B_FALSE); int index = dev->l2ad_log_ent_idx++; ASSERT3S(index, <, dev->l2ad_log_entries); ASSERT(HDR_HAS_L2HDR(hdr)); le = &lb->lb_entries[index]; bzero(le, sizeof (*le)); le->le_dva = hdr->b_dva; le->le_birth = hdr->b_birth; le->le_daddr = hdr->b_l2hdr.b_daddr; if (index == 0) dev->l2ad_log_blk_payload_start = le->le_daddr; L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr)); L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr)); L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr)); le->le_complevel = hdr->b_complevel; L2BLK_SET_TYPE((le)->le_prop, hdr->b_type); L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr))); L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr))); L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state); dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev, HDR_GET_PSIZE(hdr)); return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries); } /* * Checks whether a given L2ARC device address sits in a time-sequential * range. The trick here is that the L2ARC is a rotary buffer, so we can't * just do a range comparison, we need to handle the situation in which the * range wraps around the end of the L2ARC device. Arguments: * bottom -- Lower end of the range to check (written to earlier). * top -- Upper end of the range to check (written to later). * check -- The address for which we want to determine if it sits in * between the top and bottom. * * The 3-way conditional below represents the following cases: * * bottom < top : Sequentially ordered case: * --------+-------------------+ * | (overlap here?) | * L2ARC dev V V * |---------------============--------------| * * bottom > top: Looped-around case: * --------+------------------+ * | (overlap here?) | * L2ARC dev V V * |===============---------------===========| * ^ ^ * | (or here?) | * +---------------+--------- * * top == bottom : Just a single address comparison. */ boolean_t l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check) { if (bottom < top) return (bottom <= check && check <= top); else if (bottom > top) return (check <= top || bottom <= check); else return (check == top); } EXPORT_SYMBOL(arc_buf_size); EXPORT_SYMBOL(arc_write); EXPORT_SYMBOL(arc_read); EXPORT_SYMBOL(arc_buf_info); EXPORT_SYMBOL(arc_getbuf_func); EXPORT_SYMBOL(arc_add_prune_callback); EXPORT_SYMBOL(arc_remove_prune_callback); /* BEGIN CSTYLED */ ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_long, param_get_long, ZMOD_RW, "Min arc size"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_long, param_get_long, ZMOD_RW, "Max arc size"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit, param_set_arc_long, param_get_long, ZMOD_RW, "Metadata limit for arc size"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit_percent, param_set_arc_long, param_get_long, ZMOD_RW, "Percent of arc size for arc meta limit"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_min, param_set_arc_long, param_get_long, ZMOD_RW, "Min arc metadata"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_prune, INT, ZMOD_RW, "Meta objects to scan for prune"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_adjust_restarts, INT, ZMOD_RW, "Limit number of restarts in arc_evict_meta"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_strategy, INT, ZMOD_RW, "Meta reclaim strategy"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int, param_get_int, ZMOD_RW, "Seconds before growing arc size"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, p_dampener_disable, INT, ZMOD_RW, "Disable arc_p adapt dampener"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int, param_get_int, ZMOD_RW, "log2(fraction of arc to reclaim)"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW, "Percent of pagecache to reclaim arc to"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, p_min_shift, param_set_arc_int, param_get_int, ZMOD_RW, "arc_c shift to calc min/max arc_p"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, INT, ZMOD_RD, "Target average block size"); ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW, "Disable compressed arc buffers"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int, param_get_int, ZMOD_RW, "Min life of prefetch block in ms"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms, param_set_arc_int, param_get_int, ZMOD_RW, "Min life of prescient prefetched block in ms"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, ULONG, ZMOD_RW, "Max write bytes per interval"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, ULONG, ZMOD_RW, "Extra write bytes during device warmup"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, ULONG, ZMOD_RW, "Number of max device writes to precache"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, ULONG, ZMOD_RW, "Compressed l2arc_headroom multiplier"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, ULONG, ZMOD_RW, "TRIM ahead L2ARC write size multiplier"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, ULONG, ZMOD_RW, "Seconds between L2ARC writing"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, ULONG, ZMOD_RW, "Min feed interval in milliseconds"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW, "Skip caching prefetched buffers"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW, "Turbo L2ARC warmup"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW, "No reads during writes"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, INT, ZMOD_RW, "Percent of ARC size allowed for L2ARC-only headers"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW, "Rebuild the L2ARC when importing a pool"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, ULONG, ZMOD_RW, "Min size in bytes to write rebuild log blocks in L2ARC"); ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW, "Cache only MFU data from ARC into L2ARC"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int, param_get_int, ZMOD_RW, "System free memory I/O throttle in bytes"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_long, param_get_long, ZMOD_RW, "System free memory target size in bytes"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_long, param_get_long, ZMOD_RW, "Minimum bytes of dnodes in arc"); ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent, param_set_arc_long, param_get_long, ZMOD_RW, "Percent of ARC meta buffers for dnodes"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, ULONG, ZMOD_RW, "Percentage of excess dnodes to try to unpin"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, INT, ZMOD_RW, "When full, ARC allocation waits for eviction of this % of alloc size"); ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, INT, ZMOD_RW, "The number of headers to evict per sublist before moving to the next"); /* END CSTYLED */ diff --git a/module/zfs/spa.c b/module/zfs/spa.c index 8ca9b49ba3b3..55870bee47fb 100644 --- a/module/zfs/spa.c +++ b/module/zfs/spa.c @@ -1,9950 +1,9951 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2011, 2020 by Delphix. All rights reserved. * Copyright (c) 2018, Nexenta Systems, Inc. All rights reserved. * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. * Copyright 2013 Saso Kiselkov. All rights reserved. * Copyright (c) 2014 Integros [integros.com] * Copyright 2016 Toomas Soome * Copyright (c) 2016 Actifio, Inc. All rights reserved. * Copyright 2018 Joyent, Inc. * Copyright (c) 2017, 2019, Datto Inc. All rights reserved. * Copyright 2017 Joyent, Inc. * Copyright (c) 2017, Intel Corporation. * Copyright (c) 2021, Colm Buckley */ /* * SPA: Storage Pool Allocator * * This file contains all the routines used when modifying on-disk SPA state. * This includes opening, importing, destroying, exporting a pool, and syncing a * pool. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef _KERNEL #include #include #include #include #include #endif /* _KERNEL */ #include "zfs_prop.h" #include "zfs_comutil.h" /* * The interval, in seconds, at which failed configuration cache file writes * should be retried. */ int zfs_ccw_retry_interval = 300; typedef enum zti_modes { ZTI_MODE_FIXED, /* value is # of threads (min 1) */ ZTI_MODE_BATCH, /* cpu-intensive; value is ignored */ ZTI_MODE_SCALE, /* Taskqs scale with CPUs. */ ZTI_MODE_NULL, /* don't create a taskq */ ZTI_NMODES } zti_modes_t; #define ZTI_P(n, q) { ZTI_MODE_FIXED, (n), (q) } #define ZTI_PCT(n) { ZTI_MODE_ONLINE_PERCENT, (n), 1 } #define ZTI_BATCH { ZTI_MODE_BATCH, 0, 1 } #define ZTI_SCALE { ZTI_MODE_SCALE, 0, 1 } #define ZTI_NULL { ZTI_MODE_NULL, 0, 0 } #define ZTI_N(n) ZTI_P(n, 1) #define ZTI_ONE ZTI_N(1) typedef struct zio_taskq_info { zti_modes_t zti_mode; uint_t zti_value; uint_t zti_count; } zio_taskq_info_t; static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = { "iss", "iss_h", "int", "int_h" }; /* * This table defines the taskq settings for each ZFS I/O type. When * initializing a pool, we use this table to create an appropriately sized * taskq. Some operations are low volume and therefore have a small, static * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE * macros. Other operations process a large amount of data; the ZTI_BATCH * macro causes us to create a taskq oriented for throughput. Some operations * are so high frequency and short-lived that the taskq itself can become a * point of lock contention. The ZTI_P(#, #) macro indicates that we need an * additional degree of parallelism specified by the number of threads per- * taskq and the number of taskqs; when dispatching an event in this case, the * particular taskq is chosen at random. ZTI_SCALE is similar to ZTI_BATCH, * but with number of taskqs also scaling with number of CPUs. * * The different taskq priorities are to handle the different contexts (issue * and interrupt) and then to reserve threads for ZIO_PRIORITY_NOW I/Os that * need to be handled with minimum delay. */ const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = { /* ISSUE ISSUE_HIGH INTR INTR_HIGH */ { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* NULL */ { ZTI_N(8), ZTI_NULL, ZTI_SCALE, ZTI_NULL }, /* READ */ { ZTI_BATCH, ZTI_N(5), ZTI_SCALE, ZTI_N(5) }, /* WRITE */ { ZTI_SCALE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* FREE */ { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* CLAIM */ { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* IOCTL */ { ZTI_N(4), ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* TRIM */ }; static void spa_sync_version(void *arg, dmu_tx_t *tx); static void spa_sync_props(void *arg, dmu_tx_t *tx); static boolean_t spa_has_active_shared_spare(spa_t *spa); static int spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport); static void spa_vdev_resilver_done(spa_t *spa); uint_t zio_taskq_batch_pct = 80; /* 1 thread per cpu in pset */ uint_t zio_taskq_batch_tpq; /* threads per taskq */ boolean_t zio_taskq_sysdc = B_TRUE; /* use SDC scheduling class */ uint_t zio_taskq_basedc = 80; /* base duty cycle */ boolean_t spa_create_process = B_TRUE; /* no process ==> no sysdc */ /* * Report any spa_load_verify errors found, but do not fail spa_load. * This is used by zdb to analyze non-idle pools. */ boolean_t spa_load_verify_dryrun = B_FALSE; /* * This (illegal) pool name is used when temporarily importing a spa_t in order * to get the vdev stats associated with the imported devices. */ #define TRYIMPORT_NAME "$import" /* * For debugging purposes: print out vdev tree during pool import. */ int spa_load_print_vdev_tree = B_FALSE; /* * A non-zero value for zfs_max_missing_tvds means that we allow importing * pools with missing top-level vdevs. This is strictly intended for advanced * pool recovery cases since missing data is almost inevitable. Pools with * missing devices can only be imported read-only for safety reasons, and their * fail-mode will be automatically set to "continue". * * With 1 missing vdev we should be able to import the pool and mount all * datasets. User data that was not modified after the missing device has been * added should be recoverable. This means that snapshots created prior to the * addition of that device should be completely intact. * * With 2 missing vdevs, some datasets may fail to mount since there are * dataset statistics that are stored as regular metadata. Some data might be * recoverable if those vdevs were added recently. * * With 3 or more missing vdevs, the pool is severely damaged and MOS entries * may be missing entirely. Chances of data recovery are very low. Note that * there are also risks of performing an inadvertent rewind as we might be * missing all the vdevs with the latest uberblocks. */ unsigned long zfs_max_missing_tvds = 0; /* * The parameters below are similar to zfs_max_missing_tvds but are only * intended for a preliminary open of the pool with an untrusted config which * might be incomplete or out-dated. * * We are more tolerant for pools opened from a cachefile since we could have * an out-dated cachefile where a device removal was not registered. * We could have set the limit arbitrarily high but in the case where devices * are really missing we would want to return the proper error codes; we chose * SPA_DVAS_PER_BP - 1 so that some copies of the MOS would still be available * and we get a chance to retrieve the trusted config. */ uint64_t zfs_max_missing_tvds_cachefile = SPA_DVAS_PER_BP - 1; /* * In the case where config was assembled by scanning device paths (/dev/dsks * by default) we are less tolerant since all the existing devices should have * been detected and we want spa_load to return the right error codes. */ uint64_t zfs_max_missing_tvds_scan = 0; /* * Debugging aid that pauses spa_sync() towards the end. */ boolean_t zfs_pause_spa_sync = B_FALSE; /* * Variables to indicate the livelist condense zthr func should wait at certain * points for the livelist to be removed - used to test condense/destroy races */ int zfs_livelist_condense_zthr_pause = 0; int zfs_livelist_condense_sync_pause = 0; /* * Variables to track whether or not condense cancellation has been * triggered in testing. */ int zfs_livelist_condense_sync_cancel = 0; int zfs_livelist_condense_zthr_cancel = 0; /* * Variable to track whether or not extra ALLOC blkptrs were added to a * livelist entry while it was being condensed (caused by the way we track * remapped blkptrs in dbuf_remap_impl) */ int zfs_livelist_condense_new_alloc = 0; /* * ========================================================================== * SPA properties routines * ========================================================================== */ /* * Add a (source=src, propname=propval) list to an nvlist. */ static void spa_prop_add_list(nvlist_t *nvl, zpool_prop_t prop, char *strval, uint64_t intval, zprop_source_t src) { const char *propname = zpool_prop_to_name(prop); nvlist_t *propval; VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_uint64(propval, ZPROP_SOURCE, src) == 0); if (strval != NULL) VERIFY(nvlist_add_string(propval, ZPROP_VALUE, strval) == 0); else VERIFY(nvlist_add_uint64(propval, ZPROP_VALUE, intval) == 0); VERIFY(nvlist_add_nvlist(nvl, propname, propval) == 0); nvlist_free(propval); } /* * Get property values from the spa configuration. */ static void spa_prop_get_config(spa_t *spa, nvlist_t **nvp) { vdev_t *rvd = spa->spa_root_vdev; dsl_pool_t *pool = spa->spa_dsl_pool; uint64_t size, alloc, cap, version; const zprop_source_t src = ZPROP_SRC_NONE; spa_config_dirent_t *dp; metaslab_class_t *mc = spa_normal_class(spa); ASSERT(MUTEX_HELD(&spa->spa_props_lock)); if (rvd != NULL) { alloc = metaslab_class_get_alloc(mc); alloc += metaslab_class_get_alloc(spa_special_class(spa)); alloc += metaslab_class_get_alloc(spa_dedup_class(spa)); alloc += metaslab_class_get_alloc(spa_embedded_log_class(spa)); size = metaslab_class_get_space(mc); size += metaslab_class_get_space(spa_special_class(spa)); size += metaslab_class_get_space(spa_dedup_class(spa)); size += metaslab_class_get_space(spa_embedded_log_class(spa)); spa_prop_add_list(*nvp, ZPOOL_PROP_NAME, spa_name(spa), 0, src); spa_prop_add_list(*nvp, ZPOOL_PROP_SIZE, NULL, size, src); spa_prop_add_list(*nvp, ZPOOL_PROP_ALLOCATED, NULL, alloc, src); spa_prop_add_list(*nvp, ZPOOL_PROP_FREE, NULL, size - alloc, src); spa_prop_add_list(*nvp, ZPOOL_PROP_CHECKPOINT, NULL, spa->spa_checkpoint_info.sci_dspace, src); spa_prop_add_list(*nvp, ZPOOL_PROP_FRAGMENTATION, NULL, metaslab_class_fragmentation(mc), src); spa_prop_add_list(*nvp, ZPOOL_PROP_EXPANDSZ, NULL, metaslab_class_expandable_space(mc), src); spa_prop_add_list(*nvp, ZPOOL_PROP_READONLY, NULL, (spa_mode(spa) == SPA_MODE_READ), src); cap = (size == 0) ? 0 : (alloc * 100 / size); spa_prop_add_list(*nvp, ZPOOL_PROP_CAPACITY, NULL, cap, src); spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL, ddt_get_pool_dedup_ratio(spa), src); spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL, rvd->vdev_state, src); version = spa_version(spa); if (version == zpool_prop_default_numeric(ZPOOL_PROP_VERSION)) { spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, version, ZPROP_SRC_DEFAULT); } else { spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, version, ZPROP_SRC_LOCAL); } spa_prop_add_list(*nvp, ZPOOL_PROP_LOAD_GUID, NULL, spa_load_guid(spa), src); } if (pool != NULL) { /* * The $FREE directory was introduced in SPA_VERSION_DEADLISTS, * when opening pools before this version freedir will be NULL. */ if (pool->dp_free_dir != NULL) { spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL, dsl_dir_phys(pool->dp_free_dir)->dd_used_bytes, src); } else { spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL, 0, src); } if (pool->dp_leak_dir != NULL) { spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL, dsl_dir_phys(pool->dp_leak_dir)->dd_used_bytes, src); } else { spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL, 0, src); } } spa_prop_add_list(*nvp, ZPOOL_PROP_GUID, NULL, spa_guid(spa), src); if (spa->spa_comment != NULL) { spa_prop_add_list(*nvp, ZPOOL_PROP_COMMENT, spa->spa_comment, 0, ZPROP_SRC_LOCAL); } if (spa->spa_compatibility != NULL) { spa_prop_add_list(*nvp, ZPOOL_PROP_COMPATIBILITY, spa->spa_compatibility, 0, ZPROP_SRC_LOCAL); } if (spa->spa_root != NULL) spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root, 0, ZPROP_SRC_LOCAL); if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) { spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL, MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE); } else { spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL, SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE); } if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE)) { spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL, DNODE_MAX_SIZE, ZPROP_SRC_NONE); } else { spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL, DNODE_MIN_SIZE, ZPROP_SRC_NONE); } if ((dp = list_head(&spa->spa_config_list)) != NULL) { if (dp->scd_path == NULL) { spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE, "none", 0, ZPROP_SRC_LOCAL); } else if (strcmp(dp->scd_path, spa_config_path) != 0) { spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE, dp->scd_path, 0, ZPROP_SRC_LOCAL); } } } /* * Get zpool property values. */ int spa_prop_get(spa_t *spa, nvlist_t **nvp) { objset_t *mos = spa->spa_meta_objset; zap_cursor_t zc; zap_attribute_t za; dsl_pool_t *dp; int err; err = nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP); if (err) return (err); dp = spa_get_dsl(spa); dsl_pool_config_enter(dp, FTAG); mutex_enter(&spa->spa_props_lock); /* * Get properties from the spa config. */ spa_prop_get_config(spa, nvp); /* If no pool property object, no more prop to get. */ if (mos == NULL || spa->spa_pool_props_object == 0) goto out; /* * Get properties from the MOS pool property object. */ for (zap_cursor_init(&zc, mos, spa->spa_pool_props_object); (err = zap_cursor_retrieve(&zc, &za)) == 0; zap_cursor_advance(&zc)) { uint64_t intval = 0; char *strval = NULL; zprop_source_t src = ZPROP_SRC_DEFAULT; zpool_prop_t prop; if ((prop = zpool_name_to_prop(za.za_name)) == ZPOOL_PROP_INVAL) continue; switch (za.za_integer_length) { case 8: /* integer property */ if (za.za_first_integer != zpool_prop_default_numeric(prop)) src = ZPROP_SRC_LOCAL; if (prop == ZPOOL_PROP_BOOTFS) { dsl_dataset_t *ds = NULL; err = dsl_dataset_hold_obj(dp, za.za_first_integer, FTAG, &ds); if (err != 0) break; strval = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP); dsl_dataset_name(ds, strval); dsl_dataset_rele(ds, FTAG); } else { strval = NULL; intval = za.za_first_integer; } spa_prop_add_list(*nvp, prop, strval, intval, src); if (strval != NULL) kmem_free(strval, ZFS_MAX_DATASET_NAME_LEN); break; case 1: /* string property */ strval = kmem_alloc(za.za_num_integers, KM_SLEEP); err = zap_lookup(mos, spa->spa_pool_props_object, za.za_name, 1, za.za_num_integers, strval); if (err) { kmem_free(strval, za.za_num_integers); break; } spa_prop_add_list(*nvp, prop, strval, 0, src); kmem_free(strval, za.za_num_integers); break; default: break; } } zap_cursor_fini(&zc); out: mutex_exit(&spa->spa_props_lock); dsl_pool_config_exit(dp, FTAG); if (err && err != ENOENT) { nvlist_free(*nvp); *nvp = NULL; return (err); } return (0); } /* * Validate the given pool properties nvlist and modify the list * for the property values to be set. */ static int spa_prop_validate(spa_t *spa, nvlist_t *props) { nvpair_t *elem; int error = 0, reset_bootfs = 0; uint64_t objnum = 0; boolean_t has_feature = B_FALSE; elem = NULL; while ((elem = nvlist_next_nvpair(props, elem)) != NULL) { uint64_t intval; char *strval, *slash, *check, *fname; const char *propname = nvpair_name(elem); zpool_prop_t prop = zpool_name_to_prop(propname); switch (prop) { case ZPOOL_PROP_INVAL: if (!zpool_prop_feature(propname)) { error = SET_ERROR(EINVAL); break; } /* * Sanitize the input. */ if (nvpair_type(elem) != DATA_TYPE_UINT64) { error = SET_ERROR(EINVAL); break; } if (nvpair_value_uint64(elem, &intval) != 0) { error = SET_ERROR(EINVAL); break; } if (intval != 0) { error = SET_ERROR(EINVAL); break; } fname = strchr(propname, '@') + 1; if (zfeature_lookup_name(fname, NULL) != 0) { error = SET_ERROR(EINVAL); break; } has_feature = B_TRUE; break; case ZPOOL_PROP_VERSION: error = nvpair_value_uint64(elem, &intval); if (!error && (intval < spa_version(spa) || intval > SPA_VERSION_BEFORE_FEATURES || has_feature)) error = SET_ERROR(EINVAL); break; case ZPOOL_PROP_DELEGATION: case ZPOOL_PROP_AUTOREPLACE: case ZPOOL_PROP_LISTSNAPS: case ZPOOL_PROP_AUTOEXPAND: case ZPOOL_PROP_AUTOTRIM: error = nvpair_value_uint64(elem, &intval); if (!error && intval > 1) error = SET_ERROR(EINVAL); break; case ZPOOL_PROP_MULTIHOST: error = nvpair_value_uint64(elem, &intval); if (!error && intval > 1) error = SET_ERROR(EINVAL); if (!error) { uint32_t hostid = zone_get_hostid(NULL); if (hostid) spa->spa_hostid = hostid; else error = SET_ERROR(ENOTSUP); } break; case ZPOOL_PROP_BOOTFS: /* * If the pool version is less than SPA_VERSION_BOOTFS, * or the pool is still being created (version == 0), * the bootfs property cannot be set. */ if (spa_version(spa) < SPA_VERSION_BOOTFS) { error = SET_ERROR(ENOTSUP); break; } /* * Make sure the vdev config is bootable */ if (!vdev_is_bootable(spa->spa_root_vdev)) { error = SET_ERROR(ENOTSUP); break; } reset_bootfs = 1; error = nvpair_value_string(elem, &strval); if (!error) { objset_t *os; if (strval == NULL || strval[0] == '\0') { objnum = zpool_prop_default_numeric( ZPOOL_PROP_BOOTFS); break; } error = dmu_objset_hold(strval, FTAG, &os); if (error != 0) break; /* Must be ZPL. */ if (dmu_objset_type(os) != DMU_OST_ZFS) { error = SET_ERROR(ENOTSUP); } else { objnum = dmu_objset_id(os); } dmu_objset_rele(os, FTAG); } break; case ZPOOL_PROP_FAILUREMODE: error = nvpair_value_uint64(elem, &intval); if (!error && intval > ZIO_FAILURE_MODE_PANIC) error = SET_ERROR(EINVAL); /* * This is a special case which only occurs when * the pool has completely failed. This allows * the user to change the in-core failmode property * without syncing it out to disk (I/Os might * currently be blocked). We do this by returning * EIO to the caller (spa_prop_set) to trick it * into thinking we encountered a property validation * error. */ if (!error && spa_suspended(spa)) { spa->spa_failmode = intval; error = SET_ERROR(EIO); } break; case ZPOOL_PROP_CACHEFILE: if ((error = nvpair_value_string(elem, &strval)) != 0) break; if (strval[0] == '\0') break; if (strcmp(strval, "none") == 0) break; if (strval[0] != '/') { error = SET_ERROR(EINVAL); break; } slash = strrchr(strval, '/'); ASSERT(slash != NULL); if (slash[1] == '\0' || strcmp(slash, "/.") == 0 || strcmp(slash, "/..") == 0) error = SET_ERROR(EINVAL); break; case ZPOOL_PROP_COMMENT: if ((error = nvpair_value_string(elem, &strval)) != 0) break; for (check = strval; *check != '\0'; check++) { if (!isprint(*check)) { error = SET_ERROR(EINVAL); break; } } if (strlen(strval) > ZPROP_MAX_COMMENT) error = SET_ERROR(E2BIG); break; default: break; } if (error) break; } (void) nvlist_remove_all(props, zpool_prop_to_name(ZPOOL_PROP_DEDUPDITTO)); if (!error && reset_bootfs) { error = nvlist_remove(props, zpool_prop_to_name(ZPOOL_PROP_BOOTFS), DATA_TYPE_STRING); if (!error) { error = nvlist_add_uint64(props, zpool_prop_to_name(ZPOOL_PROP_BOOTFS), objnum); } } return (error); } void spa_configfile_set(spa_t *spa, nvlist_t *nvp, boolean_t need_sync) { char *cachefile; spa_config_dirent_t *dp; if (nvlist_lookup_string(nvp, zpool_prop_to_name(ZPOOL_PROP_CACHEFILE), &cachefile) != 0) return; dp = kmem_alloc(sizeof (spa_config_dirent_t), KM_SLEEP); if (cachefile[0] == '\0') dp->scd_path = spa_strdup(spa_config_path); else if (strcmp(cachefile, "none") == 0) dp->scd_path = NULL; else dp->scd_path = spa_strdup(cachefile); list_insert_head(&spa->spa_config_list, dp); if (need_sync) spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); } int spa_prop_set(spa_t *spa, nvlist_t *nvp) { int error; nvpair_t *elem = NULL; boolean_t need_sync = B_FALSE; if ((error = spa_prop_validate(spa, nvp)) != 0) return (error); while ((elem = nvlist_next_nvpair(nvp, elem)) != NULL) { zpool_prop_t prop = zpool_name_to_prop(nvpair_name(elem)); if (prop == ZPOOL_PROP_CACHEFILE || prop == ZPOOL_PROP_ALTROOT || prop == ZPOOL_PROP_READONLY) continue; if (prop == ZPOOL_PROP_VERSION || prop == ZPOOL_PROP_INVAL) { uint64_t ver; if (prop == ZPOOL_PROP_VERSION) { VERIFY(nvpair_value_uint64(elem, &ver) == 0); } else { ASSERT(zpool_prop_feature(nvpair_name(elem))); ver = SPA_VERSION_FEATURES; need_sync = B_TRUE; } /* Save time if the version is already set. */ if (ver == spa_version(spa)) continue; /* * In addition to the pool directory object, we might * create the pool properties object, the features for * read object, the features for write object, or the * feature descriptions object. */ error = dsl_sync_task(spa->spa_name, NULL, spa_sync_version, &ver, 6, ZFS_SPACE_CHECK_RESERVED); if (error) return (error); continue; } need_sync = B_TRUE; break; } if (need_sync) { return (dsl_sync_task(spa->spa_name, NULL, spa_sync_props, nvp, 6, ZFS_SPACE_CHECK_RESERVED)); } return (0); } /* * If the bootfs property value is dsobj, clear it. */ void spa_prop_clear_bootfs(spa_t *spa, uint64_t dsobj, dmu_tx_t *tx) { if (spa->spa_bootfs == dsobj && spa->spa_pool_props_object != 0) { VERIFY(zap_remove(spa->spa_meta_objset, spa->spa_pool_props_object, zpool_prop_to_name(ZPOOL_PROP_BOOTFS), tx) == 0); spa->spa_bootfs = 0; } } /*ARGSUSED*/ static int spa_change_guid_check(void *arg, dmu_tx_t *tx) { uint64_t *newguid __maybe_unused = arg; spa_t *spa = dmu_tx_pool(tx)->dp_spa; vdev_t *rvd = spa->spa_root_vdev; uint64_t vdev_state; if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { int error = (spa_has_checkpoint(spa)) ? ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; return (SET_ERROR(error)); } spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); vdev_state = rvd->vdev_state; spa_config_exit(spa, SCL_STATE, FTAG); if (vdev_state != VDEV_STATE_HEALTHY) return (SET_ERROR(ENXIO)); ASSERT3U(spa_guid(spa), !=, *newguid); return (0); } static void spa_change_guid_sync(void *arg, dmu_tx_t *tx) { uint64_t *newguid = arg; spa_t *spa = dmu_tx_pool(tx)->dp_spa; uint64_t oldguid; vdev_t *rvd = spa->spa_root_vdev; oldguid = spa_guid(spa); spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); rvd->vdev_guid = *newguid; rvd->vdev_guid_sum += (*newguid - oldguid); vdev_config_dirty(rvd); spa_config_exit(spa, SCL_STATE, FTAG); spa_history_log_internal(spa, "guid change", tx, "old=%llu new=%llu", (u_longlong_t)oldguid, (u_longlong_t)*newguid); } /* * Change the GUID for the pool. This is done so that we can later * re-import a pool built from a clone of our own vdevs. We will modify * the root vdev's guid, our own pool guid, and then mark all of our * vdevs dirty. Note that we must make sure that all our vdevs are * online when we do this, or else any vdevs that weren't present * would be orphaned from our pool. We are also going to issue a * sysevent to update any watchers. */ int spa_change_guid(spa_t *spa) { int error; uint64_t guid; mutex_enter(&spa->spa_vdev_top_lock); mutex_enter(&spa_namespace_lock); guid = spa_generate_guid(NULL); error = dsl_sync_task(spa->spa_name, spa_change_guid_check, spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED); if (error == 0) { spa_write_cachefile(spa, B_FALSE, B_TRUE); spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID); } mutex_exit(&spa_namespace_lock); mutex_exit(&spa->spa_vdev_top_lock); return (error); } /* * ========================================================================== * SPA state manipulation (open/create/destroy/import/export) * ========================================================================== */ static int spa_error_entry_compare(const void *a, const void *b) { const spa_error_entry_t *sa = (const spa_error_entry_t *)a; const spa_error_entry_t *sb = (const spa_error_entry_t *)b; int ret; ret = memcmp(&sa->se_bookmark, &sb->se_bookmark, sizeof (zbookmark_phys_t)); return (TREE_ISIGN(ret)); } /* * Utility function which retrieves copies of the current logs and * re-initializes them in the process. */ void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub) { ASSERT(MUTEX_HELD(&spa->spa_errlist_lock)); bcopy(&spa->spa_errlist_last, last, sizeof (avl_tree_t)); bcopy(&spa->spa_errlist_scrub, scrub, sizeof (avl_tree_t)); avl_create(&spa->spa_errlist_scrub, spa_error_entry_compare, sizeof (spa_error_entry_t), offsetof(spa_error_entry_t, se_avl)); avl_create(&spa->spa_errlist_last, spa_error_entry_compare, sizeof (spa_error_entry_t), offsetof(spa_error_entry_t, se_avl)); } static void spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q) { const zio_taskq_info_t *ztip = &zio_taskqs[t][q]; enum zti_modes mode = ztip->zti_mode; uint_t value = ztip->zti_value; uint_t count = ztip->zti_count; spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; uint_t cpus, flags = TASKQ_DYNAMIC; boolean_t batch = B_FALSE; switch (mode) { case ZTI_MODE_FIXED: ASSERT3U(value, >, 0); break; case ZTI_MODE_BATCH: batch = B_TRUE; flags |= TASKQ_THREADS_CPU_PCT; value = MIN(zio_taskq_batch_pct, 100); break; case ZTI_MODE_SCALE: flags |= TASKQ_THREADS_CPU_PCT; /* * We want more taskqs to reduce lock contention, but we want * less for better request ordering and CPU utilization. */ cpus = MAX(1, boot_ncpus * zio_taskq_batch_pct / 100); if (zio_taskq_batch_tpq > 0) { count = MAX(1, (cpus + zio_taskq_batch_tpq / 2) / zio_taskq_batch_tpq); } else { /* * Prefer 6 threads per taskq, but no more taskqs * than threads in them on large systems. For 80%: * * taskq taskq total * cpus taskqs percent threads threads * ------- ------- ------- ------- ------- * 1 1 80% 1 1 * 2 1 80% 1 1 * 4 1 80% 3 3 * 8 2 40% 3 6 * 16 3 27% 4 12 * 32 5 16% 5 25 * 64 7 11% 7 49 * 128 10 8% 10 100 * 256 14 6% 15 210 */ count = 1 + cpus / 6; while (count * count > cpus) count--; } /* Limit each taskq within 100% to not trigger assertion. */ count = MAX(count, (zio_taskq_batch_pct + 99) / 100); value = (zio_taskq_batch_pct + count / 2) / count; break; case ZTI_MODE_NULL: tqs->stqs_count = 0; tqs->stqs_taskq = NULL; return; default: panic("unrecognized mode for %s_%s taskq (%u:%u) in " "spa_activate()", zio_type_name[t], zio_taskq_types[q], mode, value); break; } ASSERT3U(count, >, 0); tqs->stqs_count = count; tqs->stqs_taskq = kmem_alloc(count * sizeof (taskq_t *), KM_SLEEP); for (uint_t i = 0; i < count; i++) { taskq_t *tq; char name[32]; if (count > 1) (void) snprintf(name, sizeof (name), "%s_%s_%u", zio_type_name[t], zio_taskq_types[q], i); else (void) snprintf(name, sizeof (name), "%s_%s", zio_type_name[t], zio_taskq_types[q]); if (zio_taskq_sysdc && spa->spa_proc != &p0) { if (batch) flags |= TASKQ_DC_BATCH; tq = taskq_create_sysdc(name, value, 50, INT_MAX, spa->spa_proc, zio_taskq_basedc, flags); } else { pri_t pri = maxclsyspri; /* * The write issue taskq can be extremely CPU * intensive. Run it at slightly less important * priority than the other taskqs. * * Under Linux and FreeBSD this means incrementing * the priority value as opposed to platforms like * illumos where it should be decremented. * * On FreeBSD, if priorities divided by four (RQ_PPQ) * are equal then a difference between them is * insignificant. */ if (t == ZIO_TYPE_WRITE && q == ZIO_TASKQ_ISSUE) { #if defined(__linux__) pri++; #elif defined(__FreeBSD__) pri += 4; #else #error "unknown OS" #endif } tq = taskq_create_proc(name, value, pri, 50, INT_MAX, spa->spa_proc, flags); } tqs->stqs_taskq[i] = tq; } } static void spa_taskqs_fini(spa_t *spa, zio_type_t t, zio_taskq_type_t q) { spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; if (tqs->stqs_taskq == NULL) { ASSERT3U(tqs->stqs_count, ==, 0); return; } for (uint_t i = 0; i < tqs->stqs_count; i++) { ASSERT3P(tqs->stqs_taskq[i], !=, NULL); taskq_destroy(tqs->stqs_taskq[i]); } kmem_free(tqs->stqs_taskq, tqs->stqs_count * sizeof (taskq_t *)); tqs->stqs_taskq = NULL; } /* * Dispatch a task to the appropriate taskq for the ZFS I/O type and priority. * Note that a type may have multiple discrete taskqs to avoid lock contention * on the taskq itself. In that case we choose which taskq at random by using * the low bits of gethrtime(). */ void spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q, task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent) { spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; taskq_t *tq; ASSERT3P(tqs->stqs_taskq, !=, NULL); ASSERT3U(tqs->stqs_count, !=, 0); if (tqs->stqs_count == 1) { tq = tqs->stqs_taskq[0]; } else { tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count]; } taskq_dispatch_ent(tq, func, arg, flags, ent); } /* * Same as spa_taskq_dispatch_ent() but block on the task until completion. */ void spa_taskq_dispatch_sync(spa_t *spa, zio_type_t t, zio_taskq_type_t q, task_func_t *func, void *arg, uint_t flags) { spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; taskq_t *tq; taskqid_t id; ASSERT3P(tqs->stqs_taskq, !=, NULL); ASSERT3U(tqs->stqs_count, !=, 0); if (tqs->stqs_count == 1) { tq = tqs->stqs_taskq[0]; } else { tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count]; } id = taskq_dispatch(tq, func, arg, flags); if (id) taskq_wait_id(tq, id); } static void spa_create_zio_taskqs(spa_t *spa) { for (int t = 0; t < ZIO_TYPES; t++) { for (int q = 0; q < ZIO_TASKQ_TYPES; q++) { spa_taskqs_init(spa, t, q); } } } /* * Disabled until spa_thread() can be adapted for Linux. */ #undef HAVE_SPA_THREAD #if defined(_KERNEL) && defined(HAVE_SPA_THREAD) static void spa_thread(void *arg) { psetid_t zio_taskq_psrset_bind = PS_NONE; callb_cpr_t cprinfo; spa_t *spa = arg; user_t *pu = PTOU(curproc); CALLB_CPR_INIT(&cprinfo, &spa->spa_proc_lock, callb_generic_cpr, spa->spa_name); ASSERT(curproc != &p0); (void) snprintf(pu->u_psargs, sizeof (pu->u_psargs), "zpool-%s", spa->spa_name); (void) strlcpy(pu->u_comm, pu->u_psargs, sizeof (pu->u_comm)); /* bind this thread to the requested psrset */ if (zio_taskq_psrset_bind != PS_NONE) { pool_lock(); mutex_enter(&cpu_lock); mutex_enter(&pidlock); mutex_enter(&curproc->p_lock); if (cpupart_bind_thread(curthread, zio_taskq_psrset_bind, 0, NULL, NULL) == 0) { curthread->t_bind_pset = zio_taskq_psrset_bind; } else { cmn_err(CE_WARN, "Couldn't bind process for zfs pool \"%s\" to " "pset %d\n", spa->spa_name, zio_taskq_psrset_bind); } mutex_exit(&curproc->p_lock); mutex_exit(&pidlock); mutex_exit(&cpu_lock); pool_unlock(); } if (zio_taskq_sysdc) { sysdc_thread_enter(curthread, 100, 0); } spa->spa_proc = curproc; spa->spa_did = curthread->t_did; spa_create_zio_taskqs(spa); mutex_enter(&spa->spa_proc_lock); ASSERT(spa->spa_proc_state == SPA_PROC_CREATED); spa->spa_proc_state = SPA_PROC_ACTIVE; cv_broadcast(&spa->spa_proc_cv); CALLB_CPR_SAFE_BEGIN(&cprinfo); while (spa->spa_proc_state == SPA_PROC_ACTIVE) cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); CALLB_CPR_SAFE_END(&cprinfo, &spa->spa_proc_lock); ASSERT(spa->spa_proc_state == SPA_PROC_DEACTIVATE); spa->spa_proc_state = SPA_PROC_GONE; spa->spa_proc = &p0; cv_broadcast(&spa->spa_proc_cv); CALLB_CPR_EXIT(&cprinfo); /* drops spa_proc_lock */ mutex_enter(&curproc->p_lock); lwp_exit(); } #endif /* * Activate an uninitialized pool. */ static void spa_activate(spa_t *spa, spa_mode_t mode) { ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED); spa->spa_state = POOL_STATE_ACTIVE; spa->spa_mode = mode; spa->spa_normal_class = metaslab_class_create(spa, zfs_metaslab_ops); spa->spa_log_class = metaslab_class_create(spa, zfs_metaslab_ops); spa->spa_embedded_log_class = metaslab_class_create(spa, zfs_metaslab_ops); spa->spa_special_class = metaslab_class_create(spa, zfs_metaslab_ops); spa->spa_dedup_class = metaslab_class_create(spa, zfs_metaslab_ops); /* Try to create a covering process */ mutex_enter(&spa->spa_proc_lock); ASSERT(spa->spa_proc_state == SPA_PROC_NONE); ASSERT(spa->spa_proc == &p0); spa->spa_did = 0; #ifdef HAVE_SPA_THREAD /* Only create a process if we're going to be around a while. */ if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) { if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri, NULL, 0) == 0) { spa->spa_proc_state = SPA_PROC_CREATED; while (spa->spa_proc_state == SPA_PROC_CREATED) { cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); } ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE); ASSERT(spa->spa_proc != &p0); ASSERT(spa->spa_did != 0); } else { #ifdef _KERNEL cmn_err(CE_WARN, "Couldn't create process for zfs pool \"%s\"\n", spa->spa_name); #endif } } #endif /* HAVE_SPA_THREAD */ mutex_exit(&spa->spa_proc_lock); /* If we didn't create a process, we need to create our taskqs. */ if (spa->spa_proc == &p0) { spa_create_zio_taskqs(spa); } for (size_t i = 0; i < TXG_SIZE; i++) { spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); } list_create(&spa->spa_config_dirty_list, sizeof (vdev_t), offsetof(vdev_t, vdev_config_dirty_node)); list_create(&spa->spa_evicting_os_list, sizeof (objset_t), offsetof(objset_t, os_evicting_node)); list_create(&spa->spa_state_dirty_list, sizeof (vdev_t), offsetof(vdev_t, vdev_state_dirty_node)); txg_list_create(&spa->spa_vdev_txg_list, spa, offsetof(struct vdev, vdev_txg_node)); avl_create(&spa->spa_errlist_scrub, spa_error_entry_compare, sizeof (spa_error_entry_t), offsetof(spa_error_entry_t, se_avl)); avl_create(&spa->spa_errlist_last, spa_error_entry_compare, sizeof (spa_error_entry_t), offsetof(spa_error_entry_t, se_avl)); spa_keystore_init(&spa->spa_keystore); /* * This taskq is used to perform zvol-minor-related tasks * asynchronously. This has several advantages, including easy * resolution of various deadlocks. * * The taskq must be single threaded to ensure tasks are always * processed in the order in which they were dispatched. * * A taskq per pool allows one to keep the pools independent. * This way if one pool is suspended, it will not impact another. * * The preferred location to dispatch a zvol minor task is a sync * task. In this context, there is easy access to the spa_t and minimal * error handling is required because the sync task must succeed. */ spa->spa_zvol_taskq = taskq_create("z_zvol", 1, defclsyspri, 1, INT_MAX, 0); /* * Taskq dedicated to prefetcher threads: this is used to prevent the * pool traverse code from monopolizing the global (and limited) * system_taskq by inappropriately scheduling long running tasks on it. */ spa->spa_prefetch_taskq = taskq_create("z_prefetch", 100, defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); /* * The taskq to upgrade datasets in this pool. Currently used by * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA. */ spa->spa_upgrade_taskq = taskq_create("z_upgrade", 100, defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); } /* * Opposite of spa_activate(). */ static void spa_deactivate(spa_t *spa) { ASSERT(spa->spa_sync_on == B_FALSE); ASSERT(spa->spa_dsl_pool == NULL); ASSERT(spa->spa_root_vdev == NULL); ASSERT(spa->spa_async_zio_root == NULL); ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED); spa_evicting_os_wait(spa); if (spa->spa_zvol_taskq) { taskq_destroy(spa->spa_zvol_taskq); spa->spa_zvol_taskq = NULL; } if (spa->spa_prefetch_taskq) { taskq_destroy(spa->spa_prefetch_taskq); spa->spa_prefetch_taskq = NULL; } if (spa->spa_upgrade_taskq) { taskq_destroy(spa->spa_upgrade_taskq); spa->spa_upgrade_taskq = NULL; } txg_list_destroy(&spa->spa_vdev_txg_list); list_destroy(&spa->spa_config_dirty_list); list_destroy(&spa->spa_evicting_os_list); list_destroy(&spa->spa_state_dirty_list); taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); for (int t = 0; t < ZIO_TYPES; t++) { for (int q = 0; q < ZIO_TASKQ_TYPES; q++) { spa_taskqs_fini(spa, t, q); } } for (size_t i = 0; i < TXG_SIZE; i++) { ASSERT3P(spa->spa_txg_zio[i], !=, NULL); VERIFY0(zio_wait(spa->spa_txg_zio[i])); spa->spa_txg_zio[i] = NULL; } metaslab_class_destroy(spa->spa_normal_class); spa->spa_normal_class = NULL; metaslab_class_destroy(spa->spa_log_class); spa->spa_log_class = NULL; metaslab_class_destroy(spa->spa_embedded_log_class); spa->spa_embedded_log_class = NULL; metaslab_class_destroy(spa->spa_special_class); spa->spa_special_class = NULL; metaslab_class_destroy(spa->spa_dedup_class); spa->spa_dedup_class = NULL; /* * If this was part of an import or the open otherwise failed, we may * still have errors left in the queues. Empty them just in case. */ spa_errlog_drain(spa); avl_destroy(&spa->spa_errlist_scrub); avl_destroy(&spa->spa_errlist_last); spa_keystore_fini(&spa->spa_keystore); spa->spa_state = POOL_STATE_UNINITIALIZED; mutex_enter(&spa->spa_proc_lock); if (spa->spa_proc_state != SPA_PROC_NONE) { ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE); spa->spa_proc_state = SPA_PROC_DEACTIVATE; cv_broadcast(&spa->spa_proc_cv); while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) { ASSERT(spa->spa_proc != &p0); cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); } ASSERT(spa->spa_proc_state == SPA_PROC_GONE); spa->spa_proc_state = SPA_PROC_NONE; } ASSERT(spa->spa_proc == &p0); mutex_exit(&spa->spa_proc_lock); /* * We want to make sure spa_thread() has actually exited the ZFS * module, so that the module can't be unloaded out from underneath * it. */ if (spa->spa_did != 0) { thread_join(spa->spa_did); spa->spa_did = 0; } } /* * Verify a pool configuration, and construct the vdev tree appropriately. This * will create all the necessary vdevs in the appropriate layout, with each vdev * in the CLOSED state. This will prep the pool before open/creation/import. * All vdev validation is done by the vdev_alloc() routine. */ int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id, int atype) { nvlist_t **child; uint_t children; int error; if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0) return (error); if ((*vdp)->vdev_ops->vdev_op_leaf) return (0); error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN, &child, &children); if (error == ENOENT) return (0); if (error) { vdev_free(*vdp); *vdp = NULL; return (SET_ERROR(EINVAL)); } for (int c = 0; c < children; c++) { vdev_t *vd; if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c, atype)) != 0) { vdev_free(*vdp); *vdp = NULL; return (error); } } ASSERT(*vdp != NULL); return (0); } static boolean_t spa_should_flush_logs_on_unload(spa_t *spa) { if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) return (B_FALSE); if (!spa_writeable(spa)) return (B_FALSE); if (!spa->spa_sync_on) return (B_FALSE); if (spa_state(spa) != POOL_STATE_EXPORTED) return (B_FALSE); if (zfs_keep_log_spacemaps_at_export) return (B_FALSE); return (B_TRUE); } /* * Opens a transaction that will set the flag that will instruct * spa_sync to attempt to flush all the metaslabs for that txg. */ static void spa_unload_log_sm_flush_all(spa_t *spa) { dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); ASSERT3U(spa->spa_log_flushall_txg, ==, 0); spa->spa_log_flushall_txg = dmu_tx_get_txg(tx); dmu_tx_commit(tx); txg_wait_synced(spa_get_dsl(spa), spa->spa_log_flushall_txg); } static void spa_unload_log_sm_metadata(spa_t *spa) { void *cookie = NULL; spa_log_sm_t *sls; while ((sls = avl_destroy_nodes(&spa->spa_sm_logs_by_txg, &cookie)) != NULL) { VERIFY0(sls->sls_mscount); kmem_free(sls, sizeof (spa_log_sm_t)); } for (log_summary_entry_t *e = list_head(&spa->spa_log_summary); e != NULL; e = list_head(&spa->spa_log_summary)) { VERIFY0(e->lse_mscount); list_remove(&spa->spa_log_summary, e); kmem_free(e, sizeof (log_summary_entry_t)); } spa->spa_unflushed_stats.sus_nblocks = 0; spa->spa_unflushed_stats.sus_memused = 0; spa->spa_unflushed_stats.sus_blocklimit = 0; } static void spa_destroy_aux_threads(spa_t *spa) { if (spa->spa_condense_zthr != NULL) { zthr_destroy(spa->spa_condense_zthr); spa->spa_condense_zthr = NULL; } if (spa->spa_checkpoint_discard_zthr != NULL) { zthr_destroy(spa->spa_checkpoint_discard_zthr); spa->spa_checkpoint_discard_zthr = NULL; } if (spa->spa_livelist_delete_zthr != NULL) { zthr_destroy(spa->spa_livelist_delete_zthr); spa->spa_livelist_delete_zthr = NULL; } if (spa->spa_livelist_condense_zthr != NULL) { zthr_destroy(spa->spa_livelist_condense_zthr); spa->spa_livelist_condense_zthr = NULL; } } /* * Opposite of spa_load(). */ static void spa_unload(spa_t *spa) { ASSERT(MUTEX_HELD(&spa_namespace_lock)); ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED); spa_import_progress_remove(spa_guid(spa)); spa_load_note(spa, "UNLOADING"); spa_wake_waiters(spa); /* * If the log space map feature is enabled and the pool is getting * exported (but not destroyed), we want to spend some time flushing * as many metaslabs as we can in an attempt to destroy log space * maps and save import time. */ if (spa_should_flush_logs_on_unload(spa)) spa_unload_log_sm_flush_all(spa); /* * Stop async tasks. */ spa_async_suspend(spa); if (spa->spa_root_vdev) { vdev_t *root_vdev = spa->spa_root_vdev; vdev_initialize_stop_all(root_vdev, VDEV_INITIALIZE_ACTIVE); vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE); vdev_autotrim_stop_all(spa); vdev_rebuild_stop_all(spa); } /* * Stop syncing. */ if (spa->spa_sync_on) { txg_sync_stop(spa->spa_dsl_pool); spa->spa_sync_on = B_FALSE; } /* * This ensures that there is no async metaslab prefetching * while we attempt to unload the spa. */ if (spa->spa_root_vdev != NULL) { for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++) { vdev_t *vc = spa->spa_root_vdev->vdev_child[c]; if (vc->vdev_mg != NULL) taskq_wait(vc->vdev_mg->mg_taskq); } } if (spa->spa_mmp.mmp_thread) mmp_thread_stop(spa); /* * Wait for any outstanding async I/O to complete. */ if (spa->spa_async_zio_root != NULL) { for (int i = 0; i < max_ncpus; i++) (void) zio_wait(spa->spa_async_zio_root[i]); kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *)); spa->spa_async_zio_root = NULL; } if (spa->spa_vdev_removal != NULL) { spa_vdev_removal_destroy(spa->spa_vdev_removal); spa->spa_vdev_removal = NULL; } spa_destroy_aux_threads(spa); spa_condense_fini(spa); bpobj_close(&spa->spa_deferred_bpobj); spa_config_enter(spa, SCL_ALL, spa, RW_WRITER); /* * Close all vdevs. */ if (spa->spa_root_vdev) vdev_free(spa->spa_root_vdev); ASSERT(spa->spa_root_vdev == NULL); /* * Close the dsl pool. */ if (spa->spa_dsl_pool) { dsl_pool_close(spa->spa_dsl_pool); spa->spa_dsl_pool = NULL; spa->spa_meta_objset = NULL; } ddt_unload(spa); spa_unload_log_sm_metadata(spa); /* * Drop and purge level 2 cache */ spa_l2cache_drop(spa); for (int i = 0; i < spa->spa_spares.sav_count; i++) vdev_free(spa->spa_spares.sav_vdevs[i]); if (spa->spa_spares.sav_vdevs) { kmem_free(spa->spa_spares.sav_vdevs, spa->spa_spares.sav_count * sizeof (void *)); spa->spa_spares.sav_vdevs = NULL; } if (spa->spa_spares.sav_config) { nvlist_free(spa->spa_spares.sav_config); spa->spa_spares.sav_config = NULL; } spa->spa_spares.sav_count = 0; for (int i = 0; i < spa->spa_l2cache.sav_count; i++) { vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]); vdev_free(spa->spa_l2cache.sav_vdevs[i]); } if (spa->spa_l2cache.sav_vdevs) { kmem_free(spa->spa_l2cache.sav_vdevs, spa->spa_l2cache.sav_count * sizeof (void *)); spa->spa_l2cache.sav_vdevs = NULL; } if (spa->spa_l2cache.sav_config) { nvlist_free(spa->spa_l2cache.sav_config); spa->spa_l2cache.sav_config = NULL; } spa->spa_l2cache.sav_count = 0; spa->spa_async_suspended = 0; spa->spa_indirect_vdevs_loaded = B_FALSE; if (spa->spa_comment != NULL) { spa_strfree(spa->spa_comment); spa->spa_comment = NULL; } if (spa->spa_compatibility != NULL) { spa_strfree(spa->spa_compatibility); spa->spa_compatibility = NULL; } spa_config_exit(spa, SCL_ALL, spa); } /* * Load (or re-load) the current list of vdevs describing the active spares for * this pool. When this is called, we have some form of basic information in * 'spa_spares.sav_config'. We parse this into vdevs, try to open them, and * then re-generate a more complete list including status information. */ void spa_load_spares(spa_t *spa) { nvlist_t **spares; uint_t nspares; int i; vdev_t *vd, *tvd; #ifndef _KERNEL /* * zdb opens both the current state of the pool and the * checkpointed state (if present), with a different spa_t. * * As spare vdevs are shared among open pools, we skip loading * them when we load the checkpointed state of the pool. */ if (!spa_writeable(spa)) return; #endif ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); /* * First, close and free any existing spare vdevs. */ for (i = 0; i < spa->spa_spares.sav_count; i++) { vd = spa->spa_spares.sav_vdevs[i]; /* Undo the call to spa_activate() below */ if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid, B_FALSE)) != NULL && tvd->vdev_isspare) spa_spare_remove(tvd); vdev_close(vd); vdev_free(vd); } if (spa->spa_spares.sav_vdevs) kmem_free(spa->spa_spares.sav_vdevs, spa->spa_spares.sav_count * sizeof (void *)); if (spa->spa_spares.sav_config == NULL) nspares = 0; else VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0); spa->spa_spares.sav_count = (int)nspares; spa->spa_spares.sav_vdevs = NULL; if (nspares == 0) return; /* * Construct the array of vdevs, opening them to get status in the * process. For each spare, there is potentially two different vdev_t * structures associated with it: one in the list of spares (used only * for basic validation purposes) and one in the active vdev * configuration (if it's spared in). During this phase we open and * validate each vdev on the spare list. If the vdev also exists in the * active configuration, then we also mark this vdev as an active spare. */ spa->spa_spares.sav_vdevs = kmem_zalloc(nspares * sizeof (void *), KM_SLEEP); for (i = 0; i < spa->spa_spares.sav_count; i++) { VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0, VDEV_ALLOC_SPARE) == 0); ASSERT(vd != NULL); spa->spa_spares.sav_vdevs[i] = vd; if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid, B_FALSE)) != NULL) { if (!tvd->vdev_isspare) spa_spare_add(tvd); /* * We only mark the spare active if we were successfully * able to load the vdev. Otherwise, importing a pool * with a bad active spare would result in strange * behavior, because multiple pool would think the spare * is actively in use. * * There is a vulnerability here to an equally bizarre * circumstance, where a dead active spare is later * brought back to life (onlined or otherwise). Given * the rarity of this scenario, and the extra complexity * it adds, we ignore the possibility. */ if (!vdev_is_dead(tvd)) spa_spare_activate(tvd); } vd->vdev_top = vd; vd->vdev_aux = &spa->spa_spares; if (vdev_open(vd) != 0) continue; if (vdev_validate_aux(vd) == 0) spa_spare_add(vd); } /* * Recompute the stashed list of spares, with status information * this time. */ VERIFY(nvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, DATA_TYPE_NVLIST_ARRAY) == 0); spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *), KM_SLEEP); for (i = 0; i < spa->spa_spares.sav_count; i++) spares[i] = vdev_config_generate(spa, spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE); VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, spares, spa->spa_spares.sav_count) == 0); for (i = 0; i < spa->spa_spares.sav_count; i++) nvlist_free(spares[i]); kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *)); } /* * Load (or re-load) the current list of vdevs describing the active l2cache for * this pool. When this is called, we have some form of basic information in * 'spa_l2cache.sav_config'. We parse this into vdevs, try to open them, and * then re-generate a more complete list including status information. * Devices which are already active have their details maintained, and are * not re-opened. */ void spa_load_l2cache(spa_t *spa) { nvlist_t **l2cache = NULL; uint_t nl2cache; int i, j, oldnvdevs; uint64_t guid; vdev_t *vd, **oldvdevs, **newvdevs; spa_aux_vdev_t *sav = &spa->spa_l2cache; #ifndef _KERNEL /* * zdb opens both the current state of the pool and the * checkpointed state (if present), with a different spa_t. * * As L2 caches are part of the ARC which is shared among open * pools, we skip loading them when we load the checkpointed * state of the pool. */ if (!spa_writeable(spa)) return; #endif ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); oldvdevs = sav->sav_vdevs; oldnvdevs = sav->sav_count; sav->sav_vdevs = NULL; sav->sav_count = 0; if (sav->sav_config == NULL) { nl2cache = 0; newvdevs = NULL; goto out; } VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0); newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP); /* * Process new nvlist of vdevs. */ for (i = 0; i < nl2cache; i++) { VERIFY(nvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID, &guid) == 0); newvdevs[i] = NULL; for (j = 0; j < oldnvdevs; j++) { vd = oldvdevs[j]; if (vd != NULL && guid == vd->vdev_guid) { /* * Retain previous vdev for add/remove ops. */ newvdevs[i] = vd; oldvdevs[j] = NULL; break; } } if (newvdevs[i] == NULL) { /* * Create new vdev */ VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0, VDEV_ALLOC_L2CACHE) == 0); ASSERT(vd != NULL); newvdevs[i] = vd; /* * Commit this vdev as an l2cache device, * even if it fails to open. */ spa_l2cache_add(vd); vd->vdev_top = vd; vd->vdev_aux = sav; spa_l2cache_activate(vd); if (vdev_open(vd) != 0) continue; (void) vdev_validate_aux(vd); if (!vdev_is_dead(vd)) l2arc_add_vdev(spa, vd); /* * Upon cache device addition to a pool or pool * creation with a cache device or if the header * of the device is invalid we issue an async * TRIM command for the whole device which will * execute if l2arc_trim_ahead > 0. */ spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM); } } sav->sav_vdevs = newvdevs; sav->sav_count = (int)nl2cache; /* * Recompute the stashed list of l2cache devices, with status * information this time. */ VERIFY(nvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE, DATA_TYPE_NVLIST_ARRAY) == 0); if (sav->sav_count > 0) l2cache = kmem_alloc(sav->sav_count * sizeof (void *), KM_SLEEP); for (i = 0; i < sav->sav_count; i++) l2cache[i] = vdev_config_generate(spa, sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE); VERIFY(nvlist_add_nvlist_array(sav->sav_config, ZPOOL_CONFIG_L2CACHE, l2cache, sav->sav_count) == 0); out: /* * Purge vdevs that were dropped */ for (i = 0; i < oldnvdevs; i++) { uint64_t pool; vd = oldvdevs[i]; if (vd != NULL) { ASSERT(vd->vdev_isl2cache); if (spa_l2cache_exists(vd->vdev_guid, &pool) && pool != 0ULL && l2arc_vdev_present(vd)) l2arc_remove_vdev(vd); vdev_clear_stats(vd); vdev_free(vd); } } if (oldvdevs) kmem_free(oldvdevs, oldnvdevs * sizeof (void *)); for (i = 0; i < sav->sav_count; i++) nvlist_free(l2cache[i]); if (sav->sav_count) kmem_free(l2cache, sav->sav_count * sizeof (void *)); } static int load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value) { dmu_buf_t *db; char *packed = NULL; size_t nvsize = 0; int error; *value = NULL; error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db); if (error) return (error); nvsize = *(uint64_t *)db->db_data; dmu_buf_rele(db, FTAG); packed = vmem_alloc(nvsize, KM_SLEEP); error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed, DMU_READ_PREFETCH); if (error == 0) error = nvlist_unpack(packed, nvsize, value, 0); vmem_free(packed, nvsize); return (error); } /* * Concrete top-level vdevs that are not missing and are not logs. At every * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds. */ static uint64_t spa_healthy_core_tvds(spa_t *spa) { vdev_t *rvd = spa->spa_root_vdev; uint64_t tvds = 0; for (uint64_t i = 0; i < rvd->vdev_children; i++) { vdev_t *vd = rvd->vdev_child[i]; if (vd->vdev_islog) continue; if (vdev_is_concrete(vd) && !vdev_is_dead(vd)) tvds++; } return (tvds); } /* * Checks to see if the given vdev could not be opened, in which case we post a * sysevent to notify the autoreplace code that the device has been removed. */ static void spa_check_removed(vdev_t *vd) { for (uint64_t c = 0; c < vd->vdev_children; c++) spa_check_removed(vd->vdev_child[c]); if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) && vdev_is_concrete(vd)) { zfs_post_autoreplace(vd->vdev_spa, vd); spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK); } } static int spa_check_for_missing_logs(spa_t *spa) { vdev_t *rvd = spa->spa_root_vdev; /* * If we're doing a normal import, then build up any additional * diagnostic information about missing log devices. * We'll pass this up to the user for further processing. */ if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) { nvlist_t **child, *nv; uint64_t idx = 0; child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t *), KM_SLEEP); VERIFY(nvlist_alloc(&nv, NV_UNIQUE_NAME, KM_SLEEP) == 0); for (uint64_t c = 0; c < rvd->vdev_children; c++) { vdev_t *tvd = rvd->vdev_child[c]; /* * We consider a device as missing only if it failed * to open (i.e. offline or faulted is not considered * as missing). */ if (tvd->vdev_islog && tvd->vdev_state == VDEV_STATE_CANT_OPEN) { child[idx++] = vdev_config_generate(spa, tvd, B_FALSE, VDEV_CONFIG_MISSING); } } if (idx > 0) { fnvlist_add_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN, child, idx); fnvlist_add_nvlist(spa->spa_load_info, ZPOOL_CONFIG_MISSING_DEVICES, nv); for (uint64_t i = 0; i < idx; i++) nvlist_free(child[i]); } nvlist_free(nv); kmem_free(child, rvd->vdev_children * sizeof (char **)); if (idx > 0) { spa_load_failed(spa, "some log devices are missing"); vdev_dbgmsg_print_tree(rvd, 2); return (SET_ERROR(ENXIO)); } } else { for (uint64_t c = 0; c < rvd->vdev_children; c++) { vdev_t *tvd = rvd->vdev_child[c]; if (tvd->vdev_islog && tvd->vdev_state == VDEV_STATE_CANT_OPEN) { spa_set_log_state(spa, SPA_LOG_CLEAR); spa_load_note(spa, "some log devices are " "missing, ZIL is dropped."); vdev_dbgmsg_print_tree(rvd, 2); break; } } } return (0); } /* * Check for missing log devices */ static boolean_t spa_check_logs(spa_t *spa) { boolean_t rv = B_FALSE; dsl_pool_t *dp = spa_get_dsl(spa); switch (spa->spa_log_state) { default: break; case SPA_LOG_MISSING: /* need to recheck in case slog has been restored */ case SPA_LOG_UNKNOWN: rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj, zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0); if (rv) spa_set_log_state(spa, SPA_LOG_MISSING); break; } return (rv); } /* * Passivate any log vdevs (note, does not apply to embedded log metaslabs). */ static boolean_t spa_passivate_log(spa_t *spa) { vdev_t *rvd = spa->spa_root_vdev; boolean_t slog_found = B_FALSE; ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER)); for (int c = 0; c < rvd->vdev_children; c++) { vdev_t *tvd = rvd->vdev_child[c]; if (tvd->vdev_islog) { ASSERT3P(tvd->vdev_log_mg, ==, NULL); metaslab_group_passivate(tvd->vdev_mg); slog_found = B_TRUE; } } return (slog_found); } /* * Activate any log vdevs (note, does not apply to embedded log metaslabs). */ static void spa_activate_log(spa_t *spa) { vdev_t *rvd = spa->spa_root_vdev; ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER)); for (int c = 0; c < rvd->vdev_children; c++) { vdev_t *tvd = rvd->vdev_child[c]; if (tvd->vdev_islog) { ASSERT3P(tvd->vdev_log_mg, ==, NULL); metaslab_group_activate(tvd->vdev_mg); } } } int spa_reset_logs(spa_t *spa) { int error; error = dmu_objset_find(spa_name(spa), zil_reset, NULL, DS_FIND_CHILDREN); if (error == 0) { /* * We successfully offlined the log device, sync out the * current txg so that the "stubby" block can be removed * by zil_sync(). */ txg_wait_synced(spa->spa_dsl_pool, 0); } return (error); } static void spa_aux_check_removed(spa_aux_vdev_t *sav) { for (int i = 0; i < sav->sav_count; i++) spa_check_removed(sav->sav_vdevs[i]); } void spa_claim_notify(zio_t *zio) { spa_t *spa = zio->io_spa; if (zio->io_error) return; mutex_enter(&spa->spa_props_lock); /* any mutex will do */ if (spa->spa_claim_max_txg < zio->io_bp->blk_birth) spa->spa_claim_max_txg = zio->io_bp->blk_birth; mutex_exit(&spa->spa_props_lock); } typedef struct spa_load_error { uint64_t sle_meta_count; uint64_t sle_data_count; } spa_load_error_t; static void spa_load_verify_done(zio_t *zio) { blkptr_t *bp = zio->io_bp; spa_load_error_t *sle = zio->io_private; dmu_object_type_t type = BP_GET_TYPE(bp); int error = zio->io_error; spa_t *spa = zio->io_spa; abd_free(zio->io_abd); if (error) { if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) && type != DMU_OT_INTENT_LOG) atomic_inc_64(&sle->sle_meta_count); else atomic_inc_64(&sle->sle_data_count); } mutex_enter(&spa->spa_scrub_lock); spa->spa_load_verify_bytes -= BP_GET_PSIZE(bp); cv_broadcast(&spa->spa_scrub_io_cv); mutex_exit(&spa->spa_scrub_lock); } /* * Maximum number of inflight bytes is the log2 fraction of the arc size. * By default, we set it to 1/16th of the arc. */ int spa_load_verify_shift = 4; int spa_load_verify_metadata = B_TRUE; int spa_load_verify_data = B_TRUE; /*ARGSUSED*/ static int spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) { if (zb->zb_level == ZB_DNODE_LEVEL || BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp)) return (0); /* * Note: normally this routine will not be called if * spa_load_verify_metadata is not set. However, it may be useful * to manually set the flag after the traversal has begun. */ if (!spa_load_verify_metadata) return (0); if (!BP_IS_METADATA(bp) && !spa_load_verify_data) return (0); uint64_t maxinflight_bytes = arc_target_bytes() >> spa_load_verify_shift; zio_t *rio = arg; size_t size = BP_GET_PSIZE(bp); mutex_enter(&spa->spa_scrub_lock); while (spa->spa_load_verify_bytes >= maxinflight_bytes) cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); spa->spa_load_verify_bytes += size; mutex_exit(&spa->spa_scrub_lock); zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size, spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB, ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL | ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb)); return (0); } /* ARGSUSED */ static int verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) { if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN) return (SET_ERROR(ENAMETOOLONG)); return (0); } static int spa_load_verify(spa_t *spa) { zio_t *rio; spa_load_error_t sle = { 0 }; zpool_load_policy_t policy; boolean_t verify_ok = B_FALSE; int error = 0; zpool_get_load_policy(spa->spa_config, &policy); if (policy.zlp_rewind & ZPOOL_NEVER_REWIND) return (0); dsl_pool_config_enter(spa->spa_dsl_pool, FTAG); error = dmu_objset_find_dp(spa->spa_dsl_pool, spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL, DS_FIND_CHILDREN); dsl_pool_config_exit(spa->spa_dsl_pool, FTAG); if (error != 0) return (error); rio = zio_root(spa, NULL, &sle, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE); if (spa_load_verify_metadata) { if (spa->spa_extreme_rewind) { spa_load_note(spa, "performing a complete scan of the " "pool since extreme rewind is on. This may take " "a very long time.\n (spa_load_verify_data=%u, " "spa_load_verify_metadata=%u)", spa_load_verify_data, spa_load_verify_metadata); } error = traverse_pool(spa, spa->spa_verify_min_txg, TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA | TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio); } (void) zio_wait(rio); ASSERT0(spa->spa_load_verify_bytes); spa->spa_load_meta_errors = sle.sle_meta_count; spa->spa_load_data_errors = sle.sle_data_count; if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) { spa_load_note(spa, "spa_load_verify found %llu metadata errors " "and %llu data errors", (u_longlong_t)sle.sle_meta_count, (u_longlong_t)sle.sle_data_count); } if (spa_load_verify_dryrun || (!error && sle.sle_meta_count <= policy.zlp_maxmeta && sle.sle_data_count <= policy.zlp_maxdata)) { int64_t loss = 0; verify_ok = B_TRUE; spa->spa_load_txg = spa->spa_uberblock.ub_txg; spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp; loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts; VERIFY(nvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_LOAD_TIME, spa->spa_load_txg_ts) == 0); VERIFY(nvlist_add_int64(spa->spa_load_info, ZPOOL_CONFIG_REWIND_TIME, loss) == 0); VERIFY(nvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count) == 0); } else { spa->spa_load_max_txg = spa->spa_uberblock.ub_txg; } if (spa_load_verify_dryrun) return (0); if (error) { if (error != ENXIO && error != EIO) error = SET_ERROR(EIO); return (error); } return (verify_ok ? 0 : EIO); } /* * Find a value in the pool props object. */ static void spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val) { (void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object, zpool_prop_to_name(prop), sizeof (uint64_t), 1, val); } /* * Find a value in the pool directory object. */ static int spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent) { int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, name, sizeof (uint64_t), 1, val); if (error != 0 && (error != ENOENT || log_enoent)) { spa_load_failed(spa, "couldn't get '%s' value in MOS directory " "[error=%d]", name, error); } return (error); } static int spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err) { vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux); return (SET_ERROR(err)); } boolean_t spa_livelist_delete_check(spa_t *spa) { return (spa->spa_livelists_to_delete != 0); } /* ARGSUSED */ static boolean_t spa_livelist_delete_cb_check(void *arg, zthr_t *z) { spa_t *spa = arg; return (spa_livelist_delete_check(spa)); } static int delete_blkptr_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) { spa_t *spa = arg; zio_free(spa, tx->tx_txg, bp); dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, -bp_get_dsize_sync(spa, bp), -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); return (0); } static int dsl_get_next_livelist_obj(objset_t *os, uint64_t zap_obj, uint64_t *llp) { int err; zap_cursor_t zc; zap_attribute_t za; zap_cursor_init(&zc, os, zap_obj); err = zap_cursor_retrieve(&zc, &za); zap_cursor_fini(&zc); if (err == 0) *llp = za.za_first_integer; return (err); } /* * Components of livelist deletion that must be performed in syncing * context: freeing block pointers and updating the pool-wide data * structures to indicate how much work is left to do */ typedef struct sublist_delete_arg { spa_t *spa; dsl_deadlist_t *ll; uint64_t key; bplist_t *to_free; } sublist_delete_arg_t; static void sublist_delete_sync(void *arg, dmu_tx_t *tx) { sublist_delete_arg_t *sda = arg; spa_t *spa = sda->spa; dsl_deadlist_t *ll = sda->ll; uint64_t key = sda->key; bplist_t *to_free = sda->to_free; bplist_iterate(to_free, delete_blkptr_cb, spa, tx); dsl_deadlist_remove_entry(ll, key, tx); } typedef struct livelist_delete_arg { spa_t *spa; uint64_t ll_obj; uint64_t zap_obj; } livelist_delete_arg_t; static void livelist_delete_sync(void *arg, dmu_tx_t *tx) { livelist_delete_arg_t *lda = arg; spa_t *spa = lda->spa; uint64_t ll_obj = lda->ll_obj; uint64_t zap_obj = lda->zap_obj; objset_t *mos = spa->spa_meta_objset; uint64_t count; /* free the livelist and decrement the feature count */ VERIFY0(zap_remove_int(mos, zap_obj, ll_obj, tx)); dsl_deadlist_free(mos, ll_obj, tx); spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx); VERIFY0(zap_count(mos, zap_obj, &count)); if (count == 0) { /* no more livelists to delete */ VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DELETED_CLONES, tx)); VERIFY0(zap_destroy(mos, zap_obj, tx)); spa->spa_livelists_to_delete = 0; spa_notify_waiters(spa); } } /* * Load in the value for the livelist to be removed and open it. Then, * load its first sublist and determine which block pointers should actually * be freed. Then, call a synctask which performs the actual frees and updates * the pool-wide livelist data. */ /* ARGSUSED */ static void spa_livelist_delete_cb(void *arg, zthr_t *z) { spa_t *spa = arg; uint64_t ll_obj = 0, count; objset_t *mos = spa->spa_meta_objset; uint64_t zap_obj = spa->spa_livelists_to_delete; /* * Determine the next livelist to delete. This function should only * be called if there is at least one deleted clone. */ VERIFY0(dsl_get_next_livelist_obj(mos, zap_obj, &ll_obj)); VERIFY0(zap_count(mos, ll_obj, &count)); if (count > 0) { dsl_deadlist_t *ll; dsl_deadlist_entry_t *dle; bplist_t to_free; ll = kmem_zalloc(sizeof (dsl_deadlist_t), KM_SLEEP); dsl_deadlist_open(ll, mos, ll_obj); dle = dsl_deadlist_first(ll); ASSERT3P(dle, !=, NULL); bplist_create(&to_free); int err = dsl_process_sub_livelist(&dle->dle_bpobj, &to_free, z, NULL); if (err == 0) { sublist_delete_arg_t sync_arg = { .spa = spa, .ll = ll, .key = dle->dle_mintxg, .to_free = &to_free }; zfs_dbgmsg("deleting sublist (id %llu) from" " livelist %llu, %lld remaining", (u_longlong_t)dle->dle_bpobj.bpo_object, (u_longlong_t)ll_obj, (longlong_t)count - 1); VERIFY0(dsl_sync_task(spa_name(spa), NULL, sublist_delete_sync, &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY)); } else { VERIFY3U(err, ==, EINTR); } bplist_clear(&to_free); bplist_destroy(&to_free); dsl_deadlist_close(ll); kmem_free(ll, sizeof (dsl_deadlist_t)); } else { livelist_delete_arg_t sync_arg = { .spa = spa, .ll_obj = ll_obj, .zap_obj = zap_obj }; zfs_dbgmsg("deletion of livelist %llu completed", (u_longlong_t)ll_obj); VERIFY0(dsl_sync_task(spa_name(spa), NULL, livelist_delete_sync, &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY)); } } static void spa_start_livelist_destroy_thread(spa_t *spa) { ASSERT3P(spa->spa_livelist_delete_zthr, ==, NULL); spa->spa_livelist_delete_zthr = zthr_create("z_livelist_destroy", - spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa); + spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa, + minclsyspri); } typedef struct livelist_new_arg { bplist_t *allocs; bplist_t *frees; } livelist_new_arg_t; static int livelist_track_new_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, dmu_tx_t *tx) { ASSERT(tx == NULL); livelist_new_arg_t *lna = arg; if (bp_freed) { bplist_append(lna->frees, bp); } else { bplist_append(lna->allocs, bp); zfs_livelist_condense_new_alloc++; } return (0); } typedef struct livelist_condense_arg { spa_t *spa; bplist_t to_keep; uint64_t first_size; uint64_t next_size; } livelist_condense_arg_t; static void spa_livelist_condense_sync(void *arg, dmu_tx_t *tx) { livelist_condense_arg_t *lca = arg; spa_t *spa = lca->spa; bplist_t new_frees; dsl_dataset_t *ds = spa->spa_to_condense.ds; /* Have we been cancelled? */ if (spa->spa_to_condense.cancelled) { zfs_livelist_condense_sync_cancel++; goto out; } dsl_deadlist_entry_t *first = spa->spa_to_condense.first; dsl_deadlist_entry_t *next = spa->spa_to_condense.next; dsl_deadlist_t *ll = &ds->ds_dir->dd_livelist; /* * It's possible that the livelist was changed while the zthr was * running. Therefore, we need to check for new blkptrs in the two * entries being condensed and continue to track them in the livelist. * Because of the way we handle remapped blkptrs (see dbuf_remap_impl), * it's possible that the newly added blkptrs are FREEs or ALLOCs so * we need to sort them into two different bplists. */ uint64_t first_obj = first->dle_bpobj.bpo_object; uint64_t next_obj = next->dle_bpobj.bpo_object; uint64_t cur_first_size = first->dle_bpobj.bpo_phys->bpo_num_blkptrs; uint64_t cur_next_size = next->dle_bpobj.bpo_phys->bpo_num_blkptrs; bplist_create(&new_frees); livelist_new_arg_t new_bps = { .allocs = &lca->to_keep, .frees = &new_frees, }; if (cur_first_size > lca->first_size) { VERIFY0(livelist_bpobj_iterate_from_nofree(&first->dle_bpobj, livelist_track_new_cb, &new_bps, lca->first_size)); } if (cur_next_size > lca->next_size) { VERIFY0(livelist_bpobj_iterate_from_nofree(&next->dle_bpobj, livelist_track_new_cb, &new_bps, lca->next_size)); } dsl_deadlist_clear_entry(first, ll, tx); ASSERT(bpobj_is_empty(&first->dle_bpobj)); dsl_deadlist_remove_entry(ll, next->dle_mintxg, tx); bplist_iterate(&lca->to_keep, dsl_deadlist_insert_alloc_cb, ll, tx); bplist_iterate(&new_frees, dsl_deadlist_insert_free_cb, ll, tx); bplist_destroy(&new_frees); char dsname[ZFS_MAX_DATASET_NAME_LEN]; dsl_dataset_name(ds, dsname); zfs_dbgmsg("txg %llu condensing livelist of %s (id %llu), bpobj %llu " "(%llu blkptrs) and bpobj %llu (%llu blkptrs) -> bpobj %llu " "(%llu blkptrs)", (u_longlong_t)tx->tx_txg, dsname, (u_longlong_t)ds->ds_object, (u_longlong_t)first_obj, (u_longlong_t)cur_first_size, (u_longlong_t)next_obj, (u_longlong_t)cur_next_size, (u_longlong_t)first->dle_bpobj.bpo_object, (u_longlong_t)first->dle_bpobj.bpo_phys->bpo_num_blkptrs); out: dmu_buf_rele(ds->ds_dbuf, spa); spa->spa_to_condense.ds = NULL; bplist_clear(&lca->to_keep); bplist_destroy(&lca->to_keep); kmem_free(lca, sizeof (livelist_condense_arg_t)); spa->spa_to_condense.syncing = B_FALSE; } static void spa_livelist_condense_cb(void *arg, zthr_t *t) { while (zfs_livelist_condense_zthr_pause && !(zthr_has_waiters(t) || zthr_iscancelled(t))) delay(1); spa_t *spa = arg; dsl_deadlist_entry_t *first = spa->spa_to_condense.first; dsl_deadlist_entry_t *next = spa->spa_to_condense.next; uint64_t first_size, next_size; livelist_condense_arg_t *lca = kmem_alloc(sizeof (livelist_condense_arg_t), KM_SLEEP); bplist_create(&lca->to_keep); /* * Process the livelists (matching FREEs and ALLOCs) in open context * so we have minimal work in syncing context to condense. * * We save bpobj sizes (first_size and next_size) to use later in * syncing context to determine if entries were added to these sublists * while in open context. This is possible because the clone is still * active and open for normal writes and we want to make sure the new, * unprocessed blockpointers are inserted into the livelist normally. * * Note that dsl_process_sub_livelist() both stores the size number of * blockpointers and iterates over them while the bpobj's lock held, so * the sizes returned to us are consistent which what was actually * processed. */ int err = dsl_process_sub_livelist(&first->dle_bpobj, &lca->to_keep, t, &first_size); if (err == 0) err = dsl_process_sub_livelist(&next->dle_bpobj, &lca->to_keep, t, &next_size); if (err == 0) { while (zfs_livelist_condense_sync_pause && !(zthr_has_waiters(t) || zthr_iscancelled(t))) delay(1); dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); dmu_tx_mark_netfree(tx); dmu_tx_hold_space(tx, 1); err = dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE); if (err == 0) { /* * Prevent the condense zthr restarting before * the synctask completes. */ spa->spa_to_condense.syncing = B_TRUE; lca->spa = spa; lca->first_size = first_size; lca->next_size = next_size; dsl_sync_task_nowait(spa_get_dsl(spa), spa_livelist_condense_sync, lca, tx); dmu_tx_commit(tx); return; } } /* * Condensing can not continue: either it was externally stopped or * we were unable to assign to a tx because the pool has run out of * space. In the second case, we'll just end up trying to condense * again in a later txg. */ ASSERT(err != 0); bplist_clear(&lca->to_keep); bplist_destroy(&lca->to_keep); kmem_free(lca, sizeof (livelist_condense_arg_t)); dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf, spa); spa->spa_to_condense.ds = NULL; if (err == EINTR) zfs_livelist_condense_zthr_cancel++; } /* ARGSUSED */ /* * Check that there is something to condense but that a condense is not * already in progress and that condensing has not been cancelled. */ static boolean_t spa_livelist_condense_cb_check(void *arg, zthr_t *z) { spa_t *spa = arg; if ((spa->spa_to_condense.ds != NULL) && (spa->spa_to_condense.syncing == B_FALSE) && (spa->spa_to_condense.cancelled == B_FALSE)) { return (B_TRUE); } return (B_FALSE); } static void spa_start_livelist_condensing_thread(spa_t *spa) { spa->spa_to_condense.ds = NULL; spa->spa_to_condense.first = NULL; spa->spa_to_condense.next = NULL; spa->spa_to_condense.syncing = B_FALSE; spa->spa_to_condense.cancelled = B_FALSE; ASSERT3P(spa->spa_livelist_condense_zthr, ==, NULL); spa->spa_livelist_condense_zthr = zthr_create("z_livelist_condense", spa_livelist_condense_cb_check, - spa_livelist_condense_cb, spa); + spa_livelist_condense_cb, spa, minclsyspri); } static void spa_spawn_aux_threads(spa_t *spa) { ASSERT(spa_writeable(spa)); ASSERT(MUTEX_HELD(&spa_namespace_lock)); spa_start_indirect_condensing_thread(spa); spa_start_livelist_destroy_thread(spa); spa_start_livelist_condensing_thread(spa); ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL); spa->spa_checkpoint_discard_zthr = zthr_create("z_checkpoint_discard", spa_checkpoint_discard_thread_check, - spa_checkpoint_discard_thread, spa); + spa_checkpoint_discard_thread, spa, minclsyspri); } /* * Fix up config after a partly-completed split. This is done with the * ZPOOL_CONFIG_SPLIT nvlist. Both the splitting pool and the split-off * pool have that entry in their config, but only the splitting one contains * a list of all the guids of the vdevs that are being split off. * * This function determines what to do with that list: either rejoin * all the disks to the pool, or complete the splitting process. To attempt * the rejoin, each disk that is offlined is marked online again, and * we do a reopen() call. If the vdev label for every disk that was * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL) * then we call vdev_split() on each disk, and complete the split. * * Otherwise we leave the config alone, with all the vdevs in place in * the original pool. */ static void spa_try_repair(spa_t *spa, nvlist_t *config) { uint_t extracted; uint64_t *glist; uint_t i, gcount; nvlist_t *nvl; vdev_t **vd; boolean_t attempt_reopen; if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0) return; /* check that the config is complete */ if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, &glist, &gcount) != 0) return; vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP); /* attempt to online all the vdevs & validate */ attempt_reopen = B_TRUE; for (i = 0; i < gcount; i++) { if (glist[i] == 0) /* vdev is hole */ continue; vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE); if (vd[i] == NULL) { /* * Don't bother attempting to reopen the disks; * just do the split. */ attempt_reopen = B_FALSE; } else { /* attempt to re-online it */ vd[i]->vdev_offline = B_FALSE; } } if (attempt_reopen) { vdev_reopen(spa->spa_root_vdev); /* check each device to see what state it's in */ for (extracted = 0, i = 0; i < gcount; i++) { if (vd[i] != NULL && vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL) break; ++extracted; } } /* * If every disk has been moved to the new pool, or if we never * even attempted to look at them, then we split them off for * good. */ if (!attempt_reopen || gcount == extracted) { for (i = 0; i < gcount; i++) if (vd[i] != NULL) vdev_split(vd[i]); vdev_reopen(spa->spa_root_vdev); } kmem_free(vd, gcount * sizeof (vdev_t *)); } static int spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type) { char *ereport = FM_EREPORT_ZFS_POOL; int error; spa->spa_load_state = state; (void) spa_import_progress_set_state(spa_guid(spa), spa_load_state(spa)); gethrestime(&spa->spa_loaded_ts); error = spa_load_impl(spa, type, &ereport); /* * Don't count references from objsets that are already closed * and are making their way through the eviction process. */ spa_evicting_os_wait(spa); spa->spa_minref = zfs_refcount_count(&spa->spa_refcount); if (error) { if (error != EEXIST) { spa->spa_loaded_ts.tv_sec = 0; spa->spa_loaded_ts.tv_nsec = 0; } if (error != EBADF) { (void) zfs_ereport_post(ereport, spa, NULL, NULL, NULL, 0); } } spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE; spa->spa_ena = 0; (void) spa_import_progress_set_state(spa_guid(spa), spa_load_state(spa)); return (error); } #ifdef ZFS_DEBUG /* * Count the number of per-vdev ZAPs associated with all of the vdevs in the * vdev tree rooted in the given vd, and ensure that each ZAP is present in the * spa's per-vdev ZAP list. */ static uint64_t vdev_count_verify_zaps(vdev_t *vd) { spa_t *spa = vd->vdev_spa; uint64_t total = 0; if (vd->vdev_top_zap != 0) { total++; ASSERT0(zap_lookup_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps, vd->vdev_top_zap)); } if (vd->vdev_leaf_zap != 0) { total++; ASSERT0(zap_lookup_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps, vd->vdev_leaf_zap)); } for (uint64_t i = 0; i < vd->vdev_children; i++) { total += vdev_count_verify_zaps(vd->vdev_child[i]); } return (total); } #endif /* * Determine whether the activity check is required. */ static boolean_t spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label, nvlist_t *config) { uint64_t state = 0; uint64_t hostid = 0; uint64_t tryconfig_txg = 0; uint64_t tryconfig_timestamp = 0; uint16_t tryconfig_mmp_seq = 0; nvlist_t *nvinfo; if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) { nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO); (void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG, &tryconfig_txg); (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP, &tryconfig_timestamp); (void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ, &tryconfig_mmp_seq); } (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state); /* * Disable the MMP activity check - This is used by zdb which * is intended to be used on potentially active pools. */ if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) return (B_FALSE); /* * Skip the activity check when the MMP feature is disabled. */ if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0) return (B_FALSE); /* * If the tryconfig_ values are nonzero, they are the results of an * earlier tryimport. If they all match the uberblock we just found, * then the pool has not changed and we return false so we do not test * a second time. */ if (tryconfig_txg && tryconfig_txg == ub->ub_txg && tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp && tryconfig_mmp_seq && tryconfig_mmp_seq == (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) return (B_FALSE); /* * Allow the activity check to be skipped when importing the pool * on the same host which last imported it. Since the hostid from * configuration may be stale use the one read from the label. */ if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID)) hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID); if (hostid == spa_get_hostid(spa)) return (B_FALSE); /* * Skip the activity test when the pool was cleanly exported. */ if (state != POOL_STATE_ACTIVE) return (B_FALSE); return (B_TRUE); } /* * Nanoseconds the activity check must watch for changes on-disk. */ static uint64_t spa_activity_check_duration(spa_t *spa, uberblock_t *ub) { uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1); uint64_t multihost_interval = MSEC2NSEC( MMP_INTERVAL_OK(zfs_multihost_interval)); uint64_t import_delay = MAX(NANOSEC, import_intervals * multihost_interval); /* * Local tunables determine a minimum duration except for the case * where we know when the remote host will suspend the pool if MMP * writes do not land. * * See Big Theory comment at the top of mmp.c for the reasoning behind * these cases and times. */ ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100); if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) && MMP_FAIL_INT(ub) > 0) { /* MMP on remote host will suspend pool after failed writes */ import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) * MMP_IMPORT_SAFETY_FACTOR / 100; zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp " "mmp_fails=%llu ub_mmp mmp_interval=%llu " "import_intervals=%llu", (u_longlong_t)import_delay, (u_longlong_t)MMP_FAIL_INT(ub), (u_longlong_t)MMP_INTERVAL(ub), (u_longlong_t)import_intervals); } else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) && MMP_FAIL_INT(ub) == 0) { /* MMP on remote host will never suspend pool */ import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) + ub->ub_mmp_delay) * import_intervals); zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp " "mmp_interval=%llu ub_mmp_delay=%llu " "import_intervals=%llu", (u_longlong_t)import_delay, (u_longlong_t)MMP_INTERVAL(ub), (u_longlong_t)ub->ub_mmp_delay, (u_longlong_t)import_intervals); } else if (MMP_VALID(ub)) { /* * zfs-0.7 compatibility case */ import_delay = MAX(import_delay, (multihost_interval + ub->ub_mmp_delay) * import_intervals); zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu " "import_intervals=%llu leaves=%u", (u_longlong_t)import_delay, (u_longlong_t)ub->ub_mmp_delay, (u_longlong_t)import_intervals, vdev_count_leaves(spa)); } else { /* Using local tunings is the only reasonable option */ zfs_dbgmsg("pool last imported on non-MMP aware " "host using import_delay=%llu multihost_interval=%llu " "import_intervals=%llu", (u_longlong_t)import_delay, (u_longlong_t)multihost_interval, (u_longlong_t)import_intervals); } return (import_delay); } /* * Perform the import activity check. If the user canceled the import or * we detected activity then fail. */ static int spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config) { uint64_t txg = ub->ub_txg; uint64_t timestamp = ub->ub_timestamp; uint64_t mmp_config = ub->ub_mmp_config; uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0; uint64_t import_delay; hrtime_t import_expire; nvlist_t *mmp_label = NULL; vdev_t *rvd = spa->spa_root_vdev; kcondvar_t cv; kmutex_t mtx; int error = 0; cv_init(&cv, NULL, CV_DEFAULT, NULL); mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL); mutex_enter(&mtx); /* * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed * during the earlier tryimport. If the txg recorded there is 0 then * the pool is known to be active on another host. * * Otherwise, the pool might be in use on another host. Check for * changes in the uberblocks on disk if necessary. */ if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) { nvlist_t *nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO); if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) && fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) { vdev_uberblock_load(rvd, ub, &mmp_label); error = SET_ERROR(EREMOTEIO); goto out; } } import_delay = spa_activity_check_duration(spa, ub); /* Add a small random factor in case of simultaneous imports (0-25%) */ import_delay += import_delay * random_in_range(250) / 1000; import_expire = gethrtime() + import_delay; while (gethrtime() < import_expire) { (void) spa_import_progress_set_mmp_check(spa_guid(spa), NSEC2SEC(import_expire - gethrtime())); vdev_uberblock_load(rvd, ub, &mmp_label); if (txg != ub->ub_txg || timestamp != ub->ub_timestamp || mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) { zfs_dbgmsg("multihost activity detected " "txg %llu ub_txg %llu " "timestamp %llu ub_timestamp %llu " "mmp_config %#llx ub_mmp_config %#llx", (u_longlong_t)txg, (u_longlong_t)ub->ub_txg, (u_longlong_t)timestamp, (u_longlong_t)ub->ub_timestamp, (u_longlong_t)mmp_config, (u_longlong_t)ub->ub_mmp_config); error = SET_ERROR(EREMOTEIO); break; } if (mmp_label) { nvlist_free(mmp_label); mmp_label = NULL; } error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz); if (error != -1) { error = SET_ERROR(EINTR); break; } error = 0; } out: mutex_exit(&mtx); mutex_destroy(&mtx); cv_destroy(&cv); /* * If the pool is determined to be active store the status in the * spa->spa_load_info nvlist. If the remote hostname or hostid are * available from configuration read from disk store them as well. * This allows 'zpool import' to generate a more useful message. * * ZPOOL_CONFIG_MMP_STATE - observed pool status (mandatory) * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool * ZPOOL_CONFIG_MMP_HOSTID - hostid from the active pool */ if (error == EREMOTEIO) { char *hostname = ""; uint64_t hostid = 0; if (mmp_label) { if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) { hostname = fnvlist_lookup_string(mmp_label, ZPOOL_CONFIG_HOSTNAME); fnvlist_add_string(spa->spa_load_info, ZPOOL_CONFIG_MMP_HOSTNAME, hostname); } if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) { hostid = fnvlist_lookup_uint64(mmp_label, ZPOOL_CONFIG_HOSTID); fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_HOSTID, hostid); } } fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE); fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_TXG, 0); error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO); } if (mmp_label) nvlist_free(mmp_label); return (error); } static int spa_verify_host(spa_t *spa, nvlist_t *mos_config) { uint64_t hostid; char *hostname; uint64_t myhostid = 0; if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config, ZPOOL_CONFIG_HOSTID, &hostid) == 0) { hostname = fnvlist_lookup_string(mos_config, ZPOOL_CONFIG_HOSTNAME); myhostid = zone_get_hostid(NULL); if (hostid != 0 && myhostid != 0 && hostid != myhostid) { cmn_err(CE_WARN, "pool '%s' could not be " "loaded as it was last accessed by " "another system (host: %s hostid: 0x%llx). " "See: https://openzfs.github.io/openzfs-docs/msg/" "ZFS-8000-EY", spa_name(spa), hostname, (u_longlong_t)hostid); spa_load_failed(spa, "hostid verification failed: pool " "last accessed by host: %s (hostid: 0x%llx)", hostname, (u_longlong_t)hostid); return (SET_ERROR(EBADF)); } } return (0); } static int spa_ld_parse_config(spa_t *spa, spa_import_type_t type) { int error = 0; nvlist_t *nvtree, *nvl, *config = spa->spa_config; int parse; vdev_t *rvd; uint64_t pool_guid; char *comment; char *compatibility; /* * Versioning wasn't explicitly added to the label until later, so if * it's not present treat it as the initial version. */ if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION, &spa->spa_ubsync.ub_version) != 0) spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL; if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) { spa_load_failed(spa, "invalid config provided: '%s' missing", ZPOOL_CONFIG_POOL_GUID); return (SET_ERROR(EINVAL)); } /* * If we are doing an import, ensure that the pool is not already * imported by checking if its pool guid already exists in the * spa namespace. * * The only case that we allow an already imported pool to be * imported again, is when the pool is checkpointed and we want to * look at its checkpointed state from userland tools like zdb. */ #ifdef _KERNEL if ((spa->spa_load_state == SPA_LOAD_IMPORT || spa->spa_load_state == SPA_LOAD_TRYIMPORT) && spa_guid_exists(pool_guid, 0)) { #else if ((spa->spa_load_state == SPA_LOAD_IMPORT || spa->spa_load_state == SPA_LOAD_TRYIMPORT) && spa_guid_exists(pool_guid, 0) && !spa_importing_readonly_checkpoint(spa)) { #endif spa_load_failed(spa, "a pool with guid %llu is already open", (u_longlong_t)pool_guid); return (SET_ERROR(EEXIST)); } spa->spa_config_guid = pool_guid; nvlist_free(spa->spa_load_info); spa->spa_load_info = fnvlist_alloc(); ASSERT(spa->spa_comment == NULL); if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0) spa->spa_comment = spa_strdup(comment); ASSERT(spa->spa_compatibility == NULL); if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMPATIBILITY, &compatibility) == 0) spa->spa_compatibility = spa_strdup(compatibility); (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, &spa->spa_config_txg); if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0) spa->spa_config_splitting = fnvlist_dup(nvl); if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) { spa_load_failed(spa, "invalid config provided: '%s' missing", ZPOOL_CONFIG_VDEV_TREE); return (SET_ERROR(EINVAL)); } /* * Create "The Godfather" zio to hold all async IOs */ spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *), KM_SLEEP); for (int i = 0; i < max_ncpus; i++) { spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_GODFATHER); } /* * Parse the configuration into a vdev tree. We explicitly set the * value that will be returned by spa_version() since parsing the * configuration requires knowing the version number. */ spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); parse = (type == SPA_IMPORT_EXISTING ? VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT); error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse); spa_config_exit(spa, SCL_ALL, FTAG); if (error != 0) { spa_load_failed(spa, "unable to parse config [error=%d]", error); return (error); } ASSERT(spa->spa_root_vdev == rvd); ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT); if (type != SPA_IMPORT_ASSEMBLE) { ASSERT(spa_guid(spa) == pool_guid); } return (0); } /* * Recursively open all vdevs in the vdev tree. This function is called twice: * first with the untrusted config, then with the trusted config. */ static int spa_ld_open_vdevs(spa_t *spa) { int error = 0; /* * spa_missing_tvds_allowed defines how many top-level vdevs can be * missing/unopenable for the root vdev to be still considered openable. */ if (spa->spa_trust_config) { spa->spa_missing_tvds_allowed = zfs_max_missing_tvds; } else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) { spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile; } else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) { spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan; } else { spa->spa_missing_tvds_allowed = 0; } spa->spa_missing_tvds_allowed = MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); error = vdev_open(spa->spa_root_vdev); spa_config_exit(spa, SCL_ALL, FTAG); if (spa->spa_missing_tvds != 0) { spa_load_note(spa, "vdev tree has %lld missing top-level " "vdevs.", (u_longlong_t)spa->spa_missing_tvds); if (spa->spa_trust_config && (spa->spa_mode & SPA_MODE_WRITE)) { /* * Although theoretically we could allow users to open * incomplete pools in RW mode, we'd need to add a lot * of extra logic (e.g. adjust pool space to account * for missing vdevs). * This limitation also prevents users from accidentally * opening the pool in RW mode during data recovery and * damaging it further. */ spa_load_note(spa, "pools with missing top-level " "vdevs can only be opened in read-only mode."); error = SET_ERROR(ENXIO); } else { spa_load_note(spa, "current settings allow for maximum " "%lld missing top-level vdevs at this stage.", (u_longlong_t)spa->spa_missing_tvds_allowed); } } if (error != 0) { spa_load_failed(spa, "unable to open vdev tree [error=%d]", error); } if (spa->spa_missing_tvds != 0 || error != 0) vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2); return (error); } /* * We need to validate the vdev labels against the configuration that * we have in hand. This function is called twice: first with an untrusted * config, then with a trusted config. The validation is more strict when the * config is trusted. */ static int spa_ld_validate_vdevs(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); error = vdev_validate(rvd); spa_config_exit(spa, SCL_ALL, FTAG); if (error != 0) { spa_load_failed(spa, "vdev_validate failed [error=%d]", error); return (error); } if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) { spa_load_failed(spa, "cannot open vdev tree after invalidating " "some vdevs"); vdev_dbgmsg_print_tree(rvd, 2); return (SET_ERROR(ENXIO)); } return (0); } static void spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub) { spa->spa_state = POOL_STATE_ACTIVE; spa->spa_ubsync = spa->spa_uberblock; spa->spa_verify_min_txg = spa->spa_extreme_rewind ? TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1; spa->spa_first_txg = spa->spa_last_ubsync_txg ? spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1; spa->spa_claim_max_txg = spa->spa_first_txg; spa->spa_prev_software_version = ub->ub_software_version; } static int spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type) { vdev_t *rvd = spa->spa_root_vdev; nvlist_t *label; uberblock_t *ub = &spa->spa_uberblock; boolean_t activity_check = B_FALSE; /* * If we are opening the checkpointed state of the pool by * rewinding to it, at this point we will have written the * checkpointed uberblock to the vdev labels, so searching * the labels will find the right uberblock. However, if * we are opening the checkpointed state read-only, we have * not modified the labels. Therefore, we must ignore the * labels and continue using the spa_uberblock that was set * by spa_ld_checkpoint_rewind. * * Note that it would be fine to ignore the labels when * rewinding (opening writeable) as well. However, if we * crash just after writing the labels, we will end up * searching the labels. Doing so in the common case means * that this code path gets exercised normally, rather than * just in the edge case. */ if (ub->ub_checkpoint_txg != 0 && spa_importing_readonly_checkpoint(spa)) { spa_ld_select_uberblock_done(spa, ub); return (0); } /* * Find the best uberblock. */ vdev_uberblock_load(rvd, ub, &label); /* * If we weren't able to find a single valid uberblock, return failure. */ if (ub->ub_txg == 0) { nvlist_free(label); spa_load_failed(spa, "no valid uberblock found"); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO)); } if (spa->spa_load_max_txg != UINT64_MAX) { (void) spa_import_progress_set_max_txg(spa_guid(spa), (u_longlong_t)spa->spa_load_max_txg); } spa_load_note(spa, "using uberblock with txg=%llu", (u_longlong_t)ub->ub_txg); /* * For pools which have the multihost property on determine if the * pool is truly inactive and can be safely imported. Prevent * hosts which don't have a hostid set from importing the pool. */ activity_check = spa_activity_check_required(spa, ub, label, spa->spa_config); if (activity_check) { if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay && spa_get_hostid(spa) == 0) { nvlist_free(label); fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID); return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO)); } int error = spa_activity_check(spa, ub, spa->spa_config); if (error) { nvlist_free(label); return (error); } fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE); fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_TXG, ub->ub_txg); fnvlist_add_uint16(spa->spa_load_info, ZPOOL_CONFIG_MMP_SEQ, (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)); } /* * If the pool has an unsupported version we can't open it. */ if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) { nvlist_free(label); spa_load_failed(spa, "version %llu is not supported", (u_longlong_t)ub->ub_version); return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP)); } if (ub->ub_version >= SPA_VERSION_FEATURES) { nvlist_t *features; /* * If we weren't able to find what's necessary for reading the * MOS in the label, return failure. */ if (label == NULL) { spa_load_failed(spa, "label config unavailable"); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO)); } if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ, &features) != 0) { nvlist_free(label); spa_load_failed(spa, "invalid label: '%s' missing", ZPOOL_CONFIG_FEATURES_FOR_READ); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO)); } /* * Update our in-core representation with the definitive values * from the label. */ nvlist_free(spa->spa_label_features); VERIFY(nvlist_dup(features, &spa->spa_label_features, 0) == 0); } nvlist_free(label); /* * Look through entries in the label nvlist's features_for_read. If * there is a feature listed there which we don't understand then we * cannot open a pool. */ if (ub->ub_version >= SPA_VERSION_FEATURES) { nvlist_t *unsup_feat; VERIFY(nvlist_alloc(&unsup_feat, NV_UNIQUE_NAME, KM_SLEEP) == 0); for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features, NULL); nvp != NULL; nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) { if (!zfeature_is_supported(nvpair_name(nvp))) { VERIFY(nvlist_add_string(unsup_feat, nvpair_name(nvp), "") == 0); } } if (!nvlist_empty(unsup_feat)) { VERIFY(nvlist_add_nvlist(spa->spa_load_info, ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat) == 0); nvlist_free(unsup_feat); spa_load_failed(spa, "some features are unsupported"); return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT, ENOTSUP)); } nvlist_free(unsup_feat); } if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) { spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_try_repair(spa, spa->spa_config); spa_config_exit(spa, SCL_ALL, FTAG); nvlist_free(spa->spa_config_splitting); spa->spa_config_splitting = NULL; } /* * Initialize internal SPA structures. */ spa_ld_select_uberblock_done(spa, ub); return (0); } static int spa_ld_open_rootbp(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool); if (error != 0) { spa_load_failed(spa, "unable to open rootbp in dsl_pool_init " "[error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset; return (0); } static int spa_ld_trusted_config(spa_t *spa, spa_import_type_t type, boolean_t reloading) { vdev_t *mrvd, *rvd = spa->spa_root_vdev; nvlist_t *nv, *mos_config, *policy; int error = 0, copy_error; uint64_t healthy_tvds, healthy_tvds_mos; uint64_t mos_config_txg; if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE) != 0) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); /* * If we're assembling a pool from a split, the config provided is * already trusted so there is nothing to do. */ if (type == SPA_IMPORT_ASSEMBLE) return (0); healthy_tvds = spa_healthy_core_tvds(spa); if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) { spa_load_failed(spa, "unable to retrieve MOS config"); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } /* * If we are doing an open, pool owner wasn't verified yet, thus do * the verification here. */ if (spa->spa_load_state == SPA_LOAD_OPEN) { error = spa_verify_host(spa, mos_config); if (error != 0) { nvlist_free(mos_config); return (error); } } nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); /* * Build a new vdev tree from the trusted config */ error = spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD); if (error != 0) { nvlist_free(mos_config); spa_config_exit(spa, SCL_ALL, FTAG); spa_load_failed(spa, "spa_config_parse failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); } /* * Vdev paths in the MOS may be obsolete. If the untrusted config was * obtained by scanning /dev/dsk, then it will have the right vdev * paths. We update the trusted MOS config with this information. * We first try to copy the paths with vdev_copy_path_strict, which * succeeds only when both configs have exactly the same vdev tree. * If that fails, we fall back to a more flexible method that has a * best effort policy. */ copy_error = vdev_copy_path_strict(rvd, mrvd); if (copy_error != 0 || spa_load_print_vdev_tree) { spa_load_note(spa, "provided vdev tree:"); vdev_dbgmsg_print_tree(rvd, 2); spa_load_note(spa, "MOS vdev tree:"); vdev_dbgmsg_print_tree(mrvd, 2); } if (copy_error != 0) { spa_load_note(spa, "vdev_copy_path_strict failed, falling " "back to vdev_copy_path_relaxed"); vdev_copy_path_relaxed(rvd, mrvd); } vdev_close(rvd); vdev_free(rvd); spa->spa_root_vdev = mrvd; rvd = mrvd; spa_config_exit(spa, SCL_ALL, FTAG); /* * We will use spa_config if we decide to reload the spa or if spa_load * fails and we rewind. We must thus regenerate the config using the * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to * pass settings on how to load the pool and is not stored in the MOS. * We copy it over to our new, trusted config. */ mos_config_txg = fnvlist_lookup_uint64(mos_config, ZPOOL_CONFIG_POOL_TXG); nvlist_free(mos_config); mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE); if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY, &policy) == 0) fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy); spa_config_set(spa, mos_config); spa->spa_config_source = SPA_CONFIG_SRC_MOS; /* * Now that we got the config from the MOS, we should be more strict * in checking blkptrs and can make assumptions about the consistency * of the vdev tree. spa_trust_config must be set to true before opening * vdevs in order for them to be writeable. */ spa->spa_trust_config = B_TRUE; /* * Open and validate the new vdev tree */ error = spa_ld_open_vdevs(spa); if (error != 0) return (error); error = spa_ld_validate_vdevs(spa); if (error != 0) return (error); if (copy_error != 0 || spa_load_print_vdev_tree) { spa_load_note(spa, "final vdev tree:"); vdev_dbgmsg_print_tree(rvd, 2); } if (spa->spa_load_state != SPA_LOAD_TRYIMPORT && !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) { /* * Sanity check to make sure that we are indeed loading the * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds * in the config provided and they happened to be the only ones * to have the latest uberblock, we could involuntarily perform * an extreme rewind. */ healthy_tvds_mos = spa_healthy_core_tvds(spa); if (healthy_tvds_mos - healthy_tvds >= SPA_SYNC_MIN_VDEVS) { spa_load_note(spa, "config provided misses too many " "top-level vdevs compared to MOS (%lld vs %lld). ", (u_longlong_t)healthy_tvds, (u_longlong_t)healthy_tvds_mos); spa_load_note(spa, "vdev tree:"); vdev_dbgmsg_print_tree(rvd, 2); if (reloading) { spa_load_failed(spa, "config was already " "provided from MOS. Aborting."); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } spa_load_note(spa, "spa must be reloaded using MOS " "config"); return (SET_ERROR(EAGAIN)); } } error = spa_check_for_missing_logs(spa); if (error != 0) return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO)); if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) { spa_load_failed(spa, "uberblock guid sum doesn't match MOS " "guid sum (%llu != %llu)", (u_longlong_t)spa->spa_uberblock.ub_guid_sum, (u_longlong_t)rvd->vdev_guid_sum); return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO)); } return (0); } static int spa_ld_open_indirect_vdev_metadata(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; /* * Everything that we read before spa_remove_init() must be stored * on concreted vdevs. Therefore we do this as early as possible. */ error = spa_remove_init(spa); if (error != 0) { spa_load_failed(spa, "spa_remove_init failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } /* * Retrieve information needed to condense indirect vdev mappings. */ error = spa_condense_init(spa); if (error != 0) { spa_load_failed(spa, "spa_condense_init failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); } return (0); } static int spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; if (spa_version(spa) >= SPA_VERSION_FEATURES) { boolean_t missing_feat_read = B_FALSE; nvlist_t *unsup_feat, *enabled_feat; if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ, &spa->spa_feat_for_read_obj, B_TRUE) != 0) { return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE, &spa->spa_feat_for_write_obj, B_TRUE) != 0) { return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS, &spa->spa_feat_desc_obj, B_TRUE) != 0) { return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } enabled_feat = fnvlist_alloc(); unsup_feat = fnvlist_alloc(); if (!spa_features_check(spa, B_FALSE, unsup_feat, enabled_feat)) missing_feat_read = B_TRUE; if (spa_writeable(spa) || spa->spa_load_state == SPA_LOAD_TRYIMPORT) { if (!spa_features_check(spa, B_TRUE, unsup_feat, enabled_feat)) { *missing_feat_writep = B_TRUE; } } fnvlist_add_nvlist(spa->spa_load_info, ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat); if (!nvlist_empty(unsup_feat)) { fnvlist_add_nvlist(spa->spa_load_info, ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat); } fnvlist_free(enabled_feat); fnvlist_free(unsup_feat); if (!missing_feat_read) { fnvlist_add_boolean(spa->spa_load_info, ZPOOL_CONFIG_CAN_RDONLY); } /* * If the state is SPA_LOAD_TRYIMPORT, our objective is * twofold: to determine whether the pool is available for * import in read-write mode and (if it is not) whether the * pool is available for import in read-only mode. If the pool * is available for import in read-write mode, it is displayed * as available in userland; if it is not available for import * in read-only mode, it is displayed as unavailable in * userland. If the pool is available for import in read-only * mode but not read-write mode, it is displayed as unavailable * in userland with a special note that the pool is actually * available for open in read-only mode. * * As a result, if the state is SPA_LOAD_TRYIMPORT and we are * missing a feature for write, we must first determine whether * the pool can be opened read-only before returning to * userland in order to know whether to display the * abovementioned note. */ if (missing_feat_read || (*missing_feat_writep && spa_writeable(spa))) { spa_load_failed(spa, "pool uses unsupported features"); return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT, ENOTSUP)); } /* * Load refcounts for ZFS features from disk into an in-memory * cache during SPA initialization. */ for (spa_feature_t i = 0; i < SPA_FEATURES; i++) { uint64_t refcount; error = feature_get_refcount_from_disk(spa, &spa_feature_table[i], &refcount); if (error == 0) { spa->spa_feat_refcount_cache[i] = refcount; } else if (error == ENOTSUP) { spa->spa_feat_refcount_cache[i] = SPA_FEATURE_DISABLED; } else { spa_load_failed(spa, "error getting refcount " "for feature %s [error=%d]", spa_feature_table[i].fi_guid, error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } } } if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) { if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG, &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } /* * Encryption was added before bookmark_v2, even though bookmark_v2 * is now a dependency. If this pool has encryption enabled without * bookmark_v2, trigger an errata message. */ if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) && !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) { spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION; } return (0); } static int spa_ld_load_special_directories(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; spa->spa_is_initializing = B_TRUE; error = dsl_pool_open(spa->spa_dsl_pool); spa->spa_is_initializing = B_FALSE; if (error != 0) { spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } return (0); } static int spa_ld_get_props(spa_t *spa) { int error = 0; uint64_t obj; vdev_t *rvd = spa->spa_root_vdev; /* Grab the checksum salt from the MOS. */ error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1, sizeof (spa->spa_cksum_salt.zcs_bytes), spa->spa_cksum_salt.zcs_bytes); if (error == ENOENT) { /* Generate a new salt for subsequent use */ (void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes, sizeof (spa->spa_cksum_salt.zcs_bytes)); } else if (error != 0) { spa_load_failed(spa, "unable to retrieve checksum salt from " "MOS [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj); if (error != 0) { spa_load_failed(spa, "error opening deferred-frees bpobj " "[error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } /* * Load the bit that tells us to use the new accounting function * (raid-z deflation). If we have an older pool, this will not * be present. */ error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION, &spa->spa_creation_version, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); /* * Load the persistent error log. If we have an older pool, this will * not be present. */ error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB, &spa->spa_errlog_scrub, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); /* * Load the livelist deletion field. If a livelist is queued for * deletion, indicate that in the spa */ error = spa_dir_prop(spa, DMU_POOL_DELETED_CLONES, &spa->spa_livelists_to_delete, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); /* * Load the history object. If we have an older pool, this * will not be present. */ error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); /* * Load the per-vdev ZAP map. If we have an older pool, this will not * be present; in this case, defer its creation to a later time to * avoid dirtying the MOS this early / out of sync context. See * spa_sync_config_object. */ /* The sentinel is only available in the MOS config. */ nvlist_t *mos_config; if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) { spa_load_failed(spa, "unable to retrieve MOS config"); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP, &spa->spa_all_vdev_zaps, B_FALSE); if (error == ENOENT) { VERIFY(!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)); spa->spa_avz_action = AVZ_ACTION_INITIALIZE; ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev)); } else if (error != 0) { return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) { /* * An older version of ZFS overwrote the sentinel value, so * we have orphaned per-vdev ZAPs in the MOS. Defer their * destruction to later; see spa_sync_config_object. */ spa->spa_avz_action = AVZ_ACTION_DESTROY; /* * We're assuming that no vdevs have had their ZAPs created * before this. Better be sure of it. */ ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev)); } nvlist_free(mos_config); spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION); error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object, B_FALSE); if (error && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); if (error == 0) { uint64_t autoreplace = 0; spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs); spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace); spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation); spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode); spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand); spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost); spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim); spa->spa_autoreplace = (autoreplace != 0); } /* * If we are importing a pool with missing top-level vdevs, * we enforce that the pool doesn't panic or get suspended on * error since the likelihood of missing data is extremely high. */ if (spa->spa_missing_tvds > 0 && spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE && spa->spa_load_state != SPA_LOAD_TRYIMPORT) { spa_load_note(spa, "forcing failmode to 'continue' " "as some top level vdevs are missing"); spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE; } return (0); } static int spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; /* * If we're assembling the pool from the split-off vdevs of * an existing pool, we don't want to attach the spares & cache * devices. */ /* * Load any hot spares for this pool. */ error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); if (error == 0 && type != SPA_IMPORT_ASSEMBLE) { ASSERT(spa_version(spa) >= SPA_VERSION_SPARES); if (load_nvlist(spa, spa->spa_spares.sav_object, &spa->spa_spares.sav_config) != 0) { spa_load_failed(spa, "error loading spares nvlist"); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_load_spares(spa); spa_config_exit(spa, SCL_ALL, FTAG); } else if (error == 0) { spa->spa_spares.sav_sync = B_TRUE; } /* * Load any level 2 ARC devices for this pool. */ error = spa_dir_prop(spa, DMU_POOL_L2CACHE, &spa->spa_l2cache.sav_object, B_FALSE); if (error != 0 && error != ENOENT) return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); if (error == 0 && type != SPA_IMPORT_ASSEMBLE) { ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE); if (load_nvlist(spa, spa->spa_l2cache.sav_object, &spa->spa_l2cache.sav_config) != 0) { spa_load_failed(spa, "error loading l2cache nvlist"); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_load_l2cache(spa); spa_config_exit(spa, SCL_ALL, FTAG); } else if (error == 0) { spa->spa_l2cache.sav_sync = B_TRUE; } return (0); } static int spa_ld_load_vdev_metadata(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; /* * If the 'multihost' property is set, then never allow a pool to * be imported when the system hostid is zero. The exception to * this rule is zdb which is always allowed to access pools. */ if (spa_multihost(spa) && spa_get_hostid(spa) == 0 && (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) { fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID); return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO)); } /* * If the 'autoreplace' property is set, then post a resource notifying * the ZFS DE that it should not issue any faults for unopenable * devices. We also iterate over the vdevs, and post a sysevent for any * unopenable vdevs so that the normal autoreplace handler can take * over. */ if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) { spa_check_removed(spa->spa_root_vdev); /* * For the import case, this is done in spa_import(), because * at this point we're using the spare definitions from * the MOS config, not necessarily from the userland config. */ if (spa->spa_load_state != SPA_LOAD_IMPORT) { spa_aux_check_removed(&spa->spa_spares); spa_aux_check_removed(&spa->spa_l2cache); } } /* * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc. */ error = vdev_load(rvd); if (error != 0) { spa_load_failed(spa, "vdev_load failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); } error = spa_ld_log_spacemaps(spa); if (error != 0) { spa_load_failed(spa, "spa_ld_log_sm_data failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); } /* * Propagate the leaf DTLs we just loaded all the way up the vdev tree. */ spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); vdev_dtl_reassess(rvd, 0, 0, B_FALSE, B_FALSE); spa_config_exit(spa, SCL_ALL, FTAG); return (0); } static int spa_ld_load_dedup_tables(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; error = ddt_load(spa); if (error != 0) { spa_load_failed(spa, "ddt_load failed [error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); } return (0); } static int spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, char **ereport) { vdev_t *rvd = spa->spa_root_vdev; if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) { boolean_t missing = spa_check_logs(spa); if (missing) { if (spa->spa_missing_tvds != 0) { spa_load_note(spa, "spa_check_logs failed " "so dropping the logs"); } else { *ereport = FM_EREPORT_ZFS_LOG_REPLAY; spa_load_failed(spa, "spa_check_logs failed"); return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG, ENXIO)); } } } return (0); } static int spa_ld_verify_pool_data(spa_t *spa) { int error = 0; vdev_t *rvd = spa->spa_root_vdev; /* * We've successfully opened the pool, verify that we're ready * to start pushing transactions. */ if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) { error = spa_load_verify(spa); if (error != 0) { spa_load_failed(spa, "spa_load_verify failed " "[error=%d]", error); return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); } } return (0); } static void spa_ld_claim_log_blocks(spa_t *spa) { dmu_tx_t *tx; dsl_pool_t *dp = spa_get_dsl(spa); /* * Claim log blocks that haven't been committed yet. * This must all happen in a single txg. * Note: spa_claim_max_txg is updated by spa_claim_notify(), * invoked from zil_claim_log_block()'s i/o done callback. * Price of rollback is that we abandon the log. */ spa->spa_claiming = B_TRUE; tx = dmu_tx_create_assigned(dp, spa_first_txg(spa)); (void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj, zil_claim, tx, DS_FIND_CHILDREN); dmu_tx_commit(tx); spa->spa_claiming = B_FALSE; spa_set_log_state(spa, SPA_LOG_GOOD); } static void spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg, boolean_t update_config_cache) { vdev_t *rvd = spa->spa_root_vdev; int need_update = B_FALSE; /* * If the config cache is stale, or we have uninitialized * metaslabs (see spa_vdev_add()), then update the config. * * If this is a verbatim import, trust the current * in-core spa_config and update the disk labels. */ if (update_config_cache || config_cache_txg != spa->spa_config_txg || spa->spa_load_state == SPA_LOAD_IMPORT || spa->spa_load_state == SPA_LOAD_RECOVER || (spa->spa_import_flags & ZFS_IMPORT_VERBATIM)) need_update = B_TRUE; for (int c = 0; c < rvd->vdev_children; c++) if (rvd->vdev_child[c]->vdev_ms_array == 0) need_update = B_TRUE; /* * Update the config cache asynchronously in case we're the * root pool, in which case the config cache isn't writable yet. */ if (need_update) spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); } static void spa_ld_prepare_for_reload(spa_t *spa) { spa_mode_t mode = spa->spa_mode; int async_suspended = spa->spa_async_suspended; spa_unload(spa); spa_deactivate(spa); spa_activate(spa, mode); /* * We save the value of spa_async_suspended as it gets reset to 0 by * spa_unload(). We want to restore it back to the original value before * returning as we might be calling spa_async_resume() later. */ spa->spa_async_suspended = async_suspended; } static int spa_ld_read_checkpoint_txg(spa_t *spa) { uberblock_t checkpoint; int error = 0; ASSERT0(spa->spa_checkpoint_txg); ASSERT(MUTEX_HELD(&spa_namespace_lock)); error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t), sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint); if (error == ENOENT) return (0); if (error != 0) return (error); ASSERT3U(checkpoint.ub_txg, !=, 0); ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0); ASSERT3U(checkpoint.ub_timestamp, !=, 0); spa->spa_checkpoint_txg = checkpoint.ub_txg; spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp; return (0); } static int spa_ld_mos_init(spa_t *spa, spa_import_type_t type) { int error = 0; ASSERT(MUTEX_HELD(&spa_namespace_lock)); ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE); /* * Never trust the config that is provided unless we are assembling * a pool following a split. * This means don't trust blkptrs and the vdev tree in general. This * also effectively puts the spa in read-only mode since * spa_writeable() checks for spa_trust_config to be true. * We will later load a trusted config from the MOS. */ if (type != SPA_IMPORT_ASSEMBLE) spa->spa_trust_config = B_FALSE; /* * Parse the config provided to create a vdev tree. */ error = spa_ld_parse_config(spa, type); if (error != 0) return (error); spa_import_progress_add(spa); /* * Now that we have the vdev tree, try to open each vdev. This involves * opening the underlying physical device, retrieving its geometry and * probing the vdev with a dummy I/O. The state of each vdev will be set * based on the success of those operations. After this we'll be ready * to read from the vdevs. */ error = spa_ld_open_vdevs(spa); if (error != 0) return (error); /* * Read the label of each vdev and make sure that the GUIDs stored * there match the GUIDs in the config provided. * If we're assembling a new pool that's been split off from an * existing pool, the labels haven't yet been updated so we skip * validation for now. */ if (type != SPA_IMPORT_ASSEMBLE) { error = spa_ld_validate_vdevs(spa); if (error != 0) return (error); } /* * Read all vdev labels to find the best uberblock (i.e. latest, * unless spa_load_max_txg is set) and store it in spa_uberblock. We * get the list of features required to read blkptrs in the MOS from * the vdev label with the best uberblock and verify that our version * of zfs supports them all. */ error = spa_ld_select_uberblock(spa, type); if (error != 0) return (error); /* * Pass that uberblock to the dsl_pool layer which will open the root * blkptr. This blkptr points to the latest version of the MOS and will * allow us to read its contents. */ error = spa_ld_open_rootbp(spa); if (error != 0) return (error); return (0); } static int spa_ld_checkpoint_rewind(spa_t *spa) { uberblock_t checkpoint; int error = 0; ASSERT(MUTEX_HELD(&spa_namespace_lock)); ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t), sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint); if (error != 0) { spa_load_failed(spa, "unable to retrieve checkpointed " "uberblock from the MOS config [error=%d]", error); if (error == ENOENT) error = ZFS_ERR_NO_CHECKPOINT; return (error); } ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg); ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg); /* * We need to update the txg and timestamp of the checkpointed * uberblock to be higher than the latest one. This ensures that * the checkpointed uberblock is selected if we were to close and * reopen the pool right after we've written it in the vdev labels. * (also see block comment in vdev_uberblock_compare) */ checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1; checkpoint.ub_timestamp = gethrestime_sec(); /* * Set current uberblock to be the checkpointed uberblock. */ spa->spa_uberblock = checkpoint; /* * If we are doing a normal rewind, then the pool is open for * writing and we sync the "updated" checkpointed uberblock to * disk. Once this is done, we've basically rewound the whole * pool and there is no way back. * * There are cases when we don't want to attempt and sync the * checkpointed uberblock to disk because we are opening a * pool as read-only. Specifically, verifying the checkpointed * state with zdb, and importing the checkpointed state to get * a "preview" of its content. */ if (spa_writeable(spa)) { vdev_t *rvd = spa->spa_root_vdev; spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL }; int svdcount = 0; int children = rvd->vdev_children; int c0 = random_in_range(children); for (int c = 0; c < children; c++) { vdev_t *vd = rvd->vdev_child[(c0 + c) % children]; /* Stop when revisiting the first vdev */ if (c > 0 && svd[0] == vd) break; if (vd->vdev_ms_array == 0 || vd->vdev_islog || !vdev_is_concrete(vd)) continue; svd[svdcount++] = vd; if (svdcount == SPA_SYNC_MIN_VDEVS) break; } error = vdev_config_sync(svd, svdcount, spa->spa_first_txg); if (error == 0) spa->spa_last_synced_guid = rvd->vdev_guid; spa_config_exit(spa, SCL_ALL, FTAG); if (error != 0) { spa_load_failed(spa, "failed to write checkpointed " "uberblock to the vdev labels [error=%d]", error); return (error); } } return (0); } static int spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type, boolean_t *update_config_cache) { int error; /* * Parse the config for pool, open and validate vdevs, * select an uberblock, and use that uberblock to open * the MOS. */ error = spa_ld_mos_init(spa, type); if (error != 0) return (error); /* * Retrieve the trusted config stored in the MOS and use it to create * a new, exact version of the vdev tree, then reopen all vdevs. */ error = spa_ld_trusted_config(spa, type, B_FALSE); if (error == EAGAIN) { if (update_config_cache != NULL) *update_config_cache = B_TRUE; /* * Redo the loading process with the trusted config if it is * too different from the untrusted config. */ spa_ld_prepare_for_reload(spa); spa_load_note(spa, "RELOADING"); error = spa_ld_mos_init(spa, type); if (error != 0) return (error); error = spa_ld_trusted_config(spa, type, B_TRUE); if (error != 0) return (error); } else if (error != 0) { return (error); } return (0); } /* * Load an existing storage pool, using the config provided. This config * describes which vdevs are part of the pool and is later validated against * partial configs present in each vdev's label and an entire copy of the * config stored in the MOS. */ static int spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport) { int error = 0; boolean_t missing_feat_write = B_FALSE; boolean_t checkpoint_rewind = (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); boolean_t update_config_cache = B_FALSE; ASSERT(MUTEX_HELD(&spa_namespace_lock)); ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE); spa_load_note(spa, "LOADING"); error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache); if (error != 0) return (error); /* * If we are rewinding to the checkpoint then we need to repeat * everything we've done so far in this function but this time * selecting the checkpointed uberblock and using that to open * the MOS. */ if (checkpoint_rewind) { /* * If we are rewinding to the checkpoint update config cache * anyway. */ update_config_cache = B_TRUE; /* * Extract the checkpointed uberblock from the current MOS * and use this as the pool's uberblock from now on. If the * pool is imported as writeable we also write the checkpoint * uberblock to the labels, making the rewind permanent. */ error = spa_ld_checkpoint_rewind(spa); if (error != 0) return (error); /* * Redo the loading process again with the * checkpointed uberblock. */ spa_ld_prepare_for_reload(spa); spa_load_note(spa, "LOADING checkpointed uberblock"); error = spa_ld_mos_with_trusted_config(spa, type, NULL); if (error != 0) return (error); } /* * Retrieve the checkpoint txg if the pool has a checkpoint. */ error = spa_ld_read_checkpoint_txg(spa); if (error != 0) return (error); /* * Retrieve the mapping of indirect vdevs. Those vdevs were removed * from the pool and their contents were re-mapped to other vdevs. Note * that everything that we read before this step must have been * rewritten on concrete vdevs after the last device removal was * initiated. Otherwise we could be reading from indirect vdevs before * we have loaded their mappings. */ error = spa_ld_open_indirect_vdev_metadata(spa); if (error != 0) return (error); /* * Retrieve the full list of active features from the MOS and check if * they are all supported. */ error = spa_ld_check_features(spa, &missing_feat_write); if (error != 0) return (error); /* * Load several special directories from the MOS needed by the dsl_pool * layer. */ error = spa_ld_load_special_directories(spa); if (error != 0) return (error); /* * Retrieve pool properties from the MOS. */ error = spa_ld_get_props(spa); if (error != 0) return (error); /* * Retrieve the list of auxiliary devices - cache devices and spares - * and open them. */ error = spa_ld_open_aux_vdevs(spa, type); if (error != 0) return (error); /* * Load the metadata for all vdevs. Also check if unopenable devices * should be autoreplaced. */ error = spa_ld_load_vdev_metadata(spa); if (error != 0) return (error); error = spa_ld_load_dedup_tables(spa); if (error != 0) return (error); /* * Verify the logs now to make sure we don't have any unexpected errors * when we claim log blocks later. */ error = spa_ld_verify_logs(spa, type, ereport); if (error != 0) return (error); if (missing_feat_write) { ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT); /* * At this point, we know that we can open the pool in * read-only mode but not read-write mode. We now have enough * information and can return to userland. */ return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT, ENOTSUP)); } /* * Traverse the last txgs to make sure the pool was left off in a safe * state. When performing an extreme rewind, we verify the whole pool, * which can take a very long time. */ error = spa_ld_verify_pool_data(spa); if (error != 0) return (error); /* * Calculate the deflated space for the pool. This must be done before * we write anything to the pool because we'd need to update the space * accounting using the deflated sizes. */ spa_update_dspace(spa); /* * We have now retrieved all the information we needed to open the * pool. If we are importing the pool in read-write mode, a few * additional steps must be performed to finish the import. */ if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER || spa->spa_load_max_txg == UINT64_MAX)) { uint64_t config_cache_txg = spa->spa_config_txg; ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT); /* * In case of a checkpoint rewind, log the original txg * of the checkpointed uberblock. */ if (checkpoint_rewind) { spa_history_log_internal(spa, "checkpoint rewind", NULL, "rewound state to txg=%llu", (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg); } /* * Traverse the ZIL and claim all blocks. */ spa_ld_claim_log_blocks(spa); /* * Kick-off the syncing thread. */ spa->spa_sync_on = B_TRUE; txg_sync_start(spa->spa_dsl_pool); mmp_thread_start(spa); /* * Wait for all claims to sync. We sync up to the highest * claimed log block birth time so that claimed log blocks * don't appear to be from the future. spa_claim_max_txg * will have been set for us by ZIL traversal operations * performed above. */ txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg); /* * Check if we need to request an update of the config. On the * next sync, we would update the config stored in vdev labels * and the cachefile (by default /etc/zfs/zpool.cache). */ spa_ld_check_for_config_update(spa, config_cache_txg, update_config_cache); /* * Check if a rebuild was in progress and if so resume it. * Then check all DTLs to see if anything needs resilvering. * The resilver will be deferred if a rebuild was started. */ if (vdev_rebuild_active(spa->spa_root_vdev)) { vdev_rebuild_restart(spa); } else if (!dsl_scan_resilvering(spa->spa_dsl_pool) && vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) { spa_async_request(spa, SPA_ASYNC_RESILVER); } /* * Log the fact that we booted up (so that we can detect if * we rebooted in the middle of an operation). */ spa_history_log_version(spa, "open", NULL); spa_restart_removal(spa); spa_spawn_aux_threads(spa); /* * Delete any inconsistent datasets. * * Note: * Since we may be issuing deletes for clones here, * we make sure to do so after we've spawned all the * auxiliary threads above (from which the livelist * deletion zthr is part of). */ (void) dmu_objset_find(spa_name(spa), dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN); /* * Clean up any stale temporary dataset userrefs. */ dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool); spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); vdev_initialize_restart(spa->spa_root_vdev); vdev_trim_restart(spa->spa_root_vdev); vdev_autotrim_restart(spa); spa_config_exit(spa, SCL_CONFIG, FTAG); } spa_import_progress_remove(spa_guid(spa)); spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD); spa_load_note(spa, "LOADED"); return (0); } static int spa_load_retry(spa_t *spa, spa_load_state_t state) { spa_mode_t mode = spa->spa_mode; spa_unload(spa); spa_deactivate(spa); spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1; spa_activate(spa, mode); spa_async_suspend(spa); spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu", (u_longlong_t)spa->spa_load_max_txg); return (spa_load(spa, state, SPA_IMPORT_EXISTING)); } /* * If spa_load() fails this function will try loading prior txg's. If * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this * function will not rewind the pool and will return the same error as * spa_load(). */ static int spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request, int rewind_flags) { nvlist_t *loadinfo = NULL; nvlist_t *config = NULL; int load_error, rewind_error; uint64_t safe_rewind_txg; uint64_t min_txg; if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) { spa->spa_load_max_txg = spa->spa_load_txg; spa_set_log_state(spa, SPA_LOG_CLEAR); } else { spa->spa_load_max_txg = max_request; if (max_request != UINT64_MAX) spa->spa_extreme_rewind = B_TRUE; } load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING); if (load_error == 0) return (0); if (load_error == ZFS_ERR_NO_CHECKPOINT) { /* * When attempting checkpoint-rewind on a pool with no * checkpoint, we should not attempt to load uberblocks * from previous txgs when spa_load fails. */ ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); spa_import_progress_remove(spa_guid(spa)); return (load_error); } if (spa->spa_root_vdev != NULL) config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg; spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp; if (rewind_flags & ZPOOL_NEVER_REWIND) { nvlist_free(config); spa_import_progress_remove(spa_guid(spa)); return (load_error); } if (state == SPA_LOAD_RECOVER) { /* Price of rolling back is discarding txgs, including log */ spa_set_log_state(spa, SPA_LOG_CLEAR); } else { /* * If we aren't rolling back save the load info from our first * import attempt so that we can restore it after attempting * to rewind. */ loadinfo = spa->spa_load_info; spa->spa_load_info = fnvlist_alloc(); } spa->spa_load_max_txg = spa->spa_last_ubsync_txg; safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE; min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ? TXG_INITIAL : safe_rewind_txg; /* * Continue as long as we're finding errors, we're still within * the acceptable rewind range, and we're still finding uberblocks */ while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg && spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) { if (spa->spa_load_max_txg < safe_rewind_txg) spa->spa_extreme_rewind = B_TRUE; rewind_error = spa_load_retry(spa, state); } spa->spa_extreme_rewind = B_FALSE; spa->spa_load_max_txg = UINT64_MAX; if (config && (rewind_error || state != SPA_LOAD_RECOVER)) spa_config_set(spa, config); else nvlist_free(config); if (state == SPA_LOAD_RECOVER) { ASSERT3P(loadinfo, ==, NULL); spa_import_progress_remove(spa_guid(spa)); return (rewind_error); } else { /* Store the rewind info as part of the initial load info */ fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO, spa->spa_load_info); /* Restore the initial load info */ fnvlist_free(spa->spa_load_info); spa->spa_load_info = loadinfo; spa_import_progress_remove(spa_guid(spa)); return (load_error); } } /* * Pool Open/Import * * The import case is identical to an open except that the configuration is sent * down from userland, instead of grabbed from the configuration cache. For the * case of an open, the pool configuration will exist in the * POOL_STATE_UNINITIALIZED state. * * The stats information (gen/count/ustats) is used to gather vdev statistics at * the same time open the pool, without having to keep around the spa_t in some * ambiguous state. */ static int spa_open_common(const char *pool, spa_t **spapp, void *tag, nvlist_t *nvpolicy, nvlist_t **config) { spa_t *spa; spa_load_state_t state = SPA_LOAD_OPEN; int error; int locked = B_FALSE; int firstopen = B_FALSE; *spapp = NULL; /* * As disgusting as this is, we need to support recursive calls to this * function because dsl_dir_open() is called during spa_load(), and ends * up calling spa_open() again. The real fix is to figure out how to * avoid dsl_dir_open() calling this in the first place. */ if (MUTEX_NOT_HELD(&spa_namespace_lock)) { mutex_enter(&spa_namespace_lock); locked = B_TRUE; } if ((spa = spa_lookup(pool)) == NULL) { if (locked) mutex_exit(&spa_namespace_lock); return (SET_ERROR(ENOENT)); } if (spa->spa_state == POOL_STATE_UNINITIALIZED) { zpool_load_policy_t policy; firstopen = B_TRUE; zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config, &policy); if (policy.zlp_rewind & ZPOOL_DO_REWIND) state = SPA_LOAD_RECOVER; spa_activate(spa, spa_mode_global); if (state != SPA_LOAD_RECOVER) spa->spa_last_ubsync_txg = spa->spa_load_txg = 0; spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE; zfs_dbgmsg("spa_open_common: opening %s", pool); error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind); if (error == EBADF) { /* * If vdev_validate() returns failure (indicated by * EBADF), it indicates that one of the vdevs indicates * that the pool has been exported or destroyed. If * this is the case, the config cache is out of sync and * we should remove the pool from the namespace. */ spa_unload(spa); spa_deactivate(spa); spa_write_cachefile(spa, B_TRUE, B_TRUE); spa_remove(spa); if (locked) mutex_exit(&spa_namespace_lock); return (SET_ERROR(ENOENT)); } if (error) { /* * We can't open the pool, but we still have useful * information: the state of each vdev after the * attempted vdev_open(). Return this to the user. */ if (config != NULL && spa->spa_config) { VERIFY(nvlist_dup(spa->spa_config, config, KM_SLEEP) == 0); VERIFY(nvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info) == 0); } spa_unload(spa); spa_deactivate(spa); spa->spa_last_open_failed = error; if (locked) mutex_exit(&spa_namespace_lock); *spapp = NULL; return (error); } } spa_open_ref(spa, tag); if (config != NULL) *config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); /* * If we've recovered the pool, pass back any information we * gathered while doing the load. */ if (state == SPA_LOAD_RECOVER) { VERIFY(nvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info) == 0); } if (locked) { spa->spa_last_open_failed = 0; spa->spa_last_ubsync_txg = 0; spa->spa_load_txg = 0; mutex_exit(&spa_namespace_lock); } if (firstopen) zvol_create_minors_recursive(spa_name(spa)); *spapp = spa; return (0); } int spa_open_rewind(const char *name, spa_t **spapp, void *tag, nvlist_t *policy, nvlist_t **config) { return (spa_open_common(name, spapp, tag, policy, config)); } int spa_open(const char *name, spa_t **spapp, void *tag) { return (spa_open_common(name, spapp, tag, NULL, NULL)); } /* * Lookup the given spa_t, incrementing the inject count in the process, * preventing it from being exported or destroyed. */ spa_t * spa_inject_addref(char *name) { spa_t *spa; mutex_enter(&spa_namespace_lock); if ((spa = spa_lookup(name)) == NULL) { mutex_exit(&spa_namespace_lock); return (NULL); } spa->spa_inject_ref++; mutex_exit(&spa_namespace_lock); return (spa); } void spa_inject_delref(spa_t *spa) { mutex_enter(&spa_namespace_lock); spa->spa_inject_ref--; mutex_exit(&spa_namespace_lock); } /* * Add spares device information to the nvlist. */ static void spa_add_spares(spa_t *spa, nvlist_t *config) { nvlist_t **spares; uint_t i, nspares; nvlist_t *nvroot; uint64_t guid; vdev_stat_t *vs; uint_t vsc; uint64_t pool; ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); if (spa->spa_spares.sav_count == 0) return; VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0); VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0); if (nspares != 0) { VERIFY(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, spares, nspares) == 0); VERIFY(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0); /* * Go through and find any spares which have since been * repurposed as an active spare. If this is the case, update * their status appropriately. */ for (i = 0; i < nspares; i++) { VERIFY(nvlist_lookup_uint64(spares[i], ZPOOL_CONFIG_GUID, &guid) == 0); if (spa_spare_exists(guid, &pool, NULL) && pool != 0ULL) { VERIFY(nvlist_lookup_uint64_array( spares[i], ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc) == 0); vs->vs_state = VDEV_STATE_CANT_OPEN; vs->vs_aux = VDEV_AUX_SPARED; } } } } /* * Add l2cache device information to the nvlist, including vdev stats. */ static void spa_add_l2cache(spa_t *spa, nvlist_t *config) { nvlist_t **l2cache; uint_t i, j, nl2cache; nvlist_t *nvroot; uint64_t guid; vdev_t *vd; vdev_stat_t *vs; uint_t vsc; ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); if (spa->spa_l2cache.sav_count == 0) return; VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0); VERIFY(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config, ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0); if (nl2cache != 0) { VERIFY(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0); VERIFY(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0); /* * Update level 2 cache device stats. */ for (i = 0; i < nl2cache; i++) { VERIFY(nvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID, &guid) == 0); vd = NULL; for (j = 0; j < spa->spa_l2cache.sav_count; j++) { if (guid == spa->spa_l2cache.sav_vdevs[j]->vdev_guid) { vd = spa->spa_l2cache.sav_vdevs[j]; break; } } ASSERT(vd != NULL); VERIFY(nvlist_lookup_uint64_array(l2cache[i], ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc) == 0); vdev_get_stats(vd, vs); vdev_config_generate_stats(vd, l2cache[i]); } } } static void spa_feature_stats_from_disk(spa_t *spa, nvlist_t *features) { zap_cursor_t zc; zap_attribute_t za; if (spa->spa_feat_for_read_obj != 0) { for (zap_cursor_init(&zc, spa->spa_meta_objset, spa->spa_feat_for_read_obj); zap_cursor_retrieve(&zc, &za) == 0; zap_cursor_advance(&zc)) { ASSERT(za.za_integer_length == sizeof (uint64_t) && za.za_num_integers == 1); VERIFY0(nvlist_add_uint64(features, za.za_name, za.za_first_integer)); } zap_cursor_fini(&zc); } if (spa->spa_feat_for_write_obj != 0) { for (zap_cursor_init(&zc, spa->spa_meta_objset, spa->spa_feat_for_write_obj); zap_cursor_retrieve(&zc, &za) == 0; zap_cursor_advance(&zc)) { ASSERT(za.za_integer_length == sizeof (uint64_t) && za.za_num_integers == 1); VERIFY0(nvlist_add_uint64(features, za.za_name, za.za_first_integer)); } zap_cursor_fini(&zc); } } static void spa_feature_stats_from_cache(spa_t *spa, nvlist_t *features) { int i; for (i = 0; i < SPA_FEATURES; i++) { zfeature_info_t feature = spa_feature_table[i]; uint64_t refcount; if (feature_get_refcount(spa, &feature, &refcount) != 0) continue; VERIFY0(nvlist_add_uint64(features, feature.fi_guid, refcount)); } } /* * Store a list of pool features and their reference counts in the * config. * * The first time this is called on a spa, allocate a new nvlist, fetch * the pool features and reference counts from disk, then save the list * in the spa. In subsequent calls on the same spa use the saved nvlist * and refresh its values from the cached reference counts. This * ensures we don't block here on I/O on a suspended pool so 'zpool * clear' can resume the pool. */ static void spa_add_feature_stats(spa_t *spa, nvlist_t *config) { nvlist_t *features; ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); mutex_enter(&spa->spa_feat_stats_lock); features = spa->spa_feat_stats; if (features != NULL) { spa_feature_stats_from_cache(spa, features); } else { VERIFY0(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP)); spa->spa_feat_stats = features; spa_feature_stats_from_disk(spa, features); } VERIFY0(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS, features)); mutex_exit(&spa->spa_feat_stats_lock); } int spa_get_stats(const char *name, nvlist_t **config, char *altroot, size_t buflen) { int error; spa_t *spa; *config = NULL; error = spa_open_common(name, &spa, FTAG, NULL, config); if (spa != NULL) { /* * This still leaves a window of inconsistency where the spares * or l2cache devices could change and the config would be * self-inconsistent. */ spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); if (*config != NULL) { uint64_t loadtimes[2]; loadtimes[0] = spa->spa_loaded_ts.tv_sec; loadtimes[1] = spa->spa_loaded_ts.tv_nsec; VERIFY(nvlist_add_uint64_array(*config, ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2) == 0); VERIFY(nvlist_add_uint64(*config, ZPOOL_CONFIG_ERRCOUNT, spa_get_errlog_size(spa)) == 0); if (spa_suspended(spa)) { VERIFY(nvlist_add_uint64(*config, ZPOOL_CONFIG_SUSPENDED, spa->spa_failmode) == 0); VERIFY(nvlist_add_uint64(*config, ZPOOL_CONFIG_SUSPENDED_REASON, spa->spa_suspended) == 0); } spa_add_spares(spa, *config); spa_add_l2cache(spa, *config); spa_add_feature_stats(spa, *config); } } /* * We want to get the alternate root even for faulted pools, so we cheat * and call spa_lookup() directly. */ if (altroot) { if (spa == NULL) { mutex_enter(&spa_namespace_lock); spa = spa_lookup(name); if (spa) spa_altroot(spa, altroot, buflen); else altroot[0] = '\0'; spa = NULL; mutex_exit(&spa_namespace_lock); } else { spa_altroot(spa, altroot, buflen); } } if (spa != NULL) { spa_config_exit(spa, SCL_CONFIG, FTAG); spa_close(spa, FTAG); } return (error); } /* * Validate that the auxiliary device array is well formed. We must have an * array of nvlists, each which describes a valid leaf vdev. If this is an * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be * specified, as long as they are well-formed. */ static int spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode, spa_aux_vdev_t *sav, const char *config, uint64_t version, vdev_labeltype_t label) { nvlist_t **dev; uint_t i, ndev; vdev_t *vd; int error; ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); /* * It's acceptable to have no devs specified. */ if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0) return (0); if (ndev == 0) return (SET_ERROR(EINVAL)); /* * Make sure the pool is formatted with a version that supports this * device type. */ if (spa_version(spa) < version) return (SET_ERROR(ENOTSUP)); /* * Set the pending device list so we correctly handle device in-use * checking. */ sav->sav_pending = dev; sav->sav_npending = ndev; for (i = 0; i < ndev; i++) { if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0, mode)) != 0) goto out; if (!vd->vdev_ops->vdev_op_leaf) { vdev_free(vd); error = SET_ERROR(EINVAL); goto out; } vd->vdev_top = vd; if ((error = vdev_open(vd)) == 0 && (error = vdev_label_init(vd, crtxg, label)) == 0) { VERIFY(nvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID, vd->vdev_guid) == 0); } vdev_free(vd); if (error && (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE)) goto out; else error = 0; } out: sav->sav_pending = NULL; sav->sav_npending = 0; return (error); } static int spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode) { int error; ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode, &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES, VDEV_LABEL_SPARE)) != 0) { return (error); } return (spa_validate_aux_devs(spa, nvroot, crtxg, mode, &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE, VDEV_LABEL_L2CACHE)); } static void spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs, const char *config) { int i; if (sav->sav_config != NULL) { nvlist_t **olddevs; uint_t oldndevs; nvlist_t **newdevs; /* * Generate new dev list by concatenating with the * current dev list. */ VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, config, &olddevs, &oldndevs) == 0); newdevs = kmem_alloc(sizeof (void *) * (ndevs + oldndevs), KM_SLEEP); for (i = 0; i < oldndevs; i++) VERIFY(nvlist_dup(olddevs[i], &newdevs[i], KM_SLEEP) == 0); for (i = 0; i < ndevs; i++) VERIFY(nvlist_dup(devs[i], &newdevs[i + oldndevs], KM_SLEEP) == 0); VERIFY(nvlist_remove(sav->sav_config, config, DATA_TYPE_NVLIST_ARRAY) == 0); VERIFY(nvlist_add_nvlist_array(sav->sav_config, config, newdevs, ndevs + oldndevs) == 0); for (i = 0; i < oldndevs + ndevs; i++) nvlist_free(newdevs[i]); kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *)); } else { /* * Generate a new dev list. */ VERIFY(nvlist_alloc(&sav->sav_config, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_nvlist_array(sav->sav_config, config, devs, ndevs) == 0); } } /* * Stop and drop level 2 ARC devices */ void spa_l2cache_drop(spa_t *spa) { vdev_t *vd; int i; spa_aux_vdev_t *sav = &spa->spa_l2cache; for (i = 0; i < sav->sav_count; i++) { uint64_t pool; vd = sav->sav_vdevs[i]; ASSERT(vd != NULL); if (spa_l2cache_exists(vd->vdev_guid, &pool) && pool != 0ULL && l2arc_vdev_present(vd)) l2arc_remove_vdev(vd); } } /* * Verify encryption parameters for spa creation. If we are encrypting, we must * have the encryption feature flag enabled. */ static int spa_create_check_encryption_params(dsl_crypto_params_t *dcp, boolean_t has_encryption) { if (dcp->cp_crypt != ZIO_CRYPT_OFF && dcp->cp_crypt != ZIO_CRYPT_INHERIT && !has_encryption) return (SET_ERROR(ENOTSUP)); return (dmu_objset_create_crypt_check(NULL, dcp, NULL)); } /* * Pool Creation */ int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, nvlist_t *zplprops, dsl_crypto_params_t *dcp) { spa_t *spa; char *altroot = NULL; vdev_t *rvd; dsl_pool_t *dp; dmu_tx_t *tx; int error = 0; uint64_t txg = TXG_INITIAL; nvlist_t **spares, **l2cache; uint_t nspares, nl2cache; uint64_t version, obj, ndraid = 0; boolean_t has_features; boolean_t has_encryption; boolean_t has_allocclass; spa_feature_t feat; char *feat_name; char *poolname; nvlist_t *nvl; if (props == NULL || nvlist_lookup_string(props, "tname", &poolname) != 0) poolname = (char *)pool; /* * If this pool already exists, return failure. */ mutex_enter(&spa_namespace_lock); if (spa_lookup(poolname) != NULL) { mutex_exit(&spa_namespace_lock); return (SET_ERROR(EEXIST)); } /* * Allocate a new spa_t structure. */ nvl = fnvlist_alloc(); fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool); (void) nvlist_lookup_string(props, zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); spa = spa_add(poolname, nvl, altroot); fnvlist_free(nvl); spa_activate(spa, spa_mode_global); if (props && (error = spa_prop_validate(spa, props))) { spa_deactivate(spa); spa_remove(spa); mutex_exit(&spa_namespace_lock); return (error); } /* * Temporary pool names should never be written to disk. */ if (poolname != pool) spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME; has_features = B_FALSE; has_encryption = B_FALSE; has_allocclass = B_FALSE; for (nvpair_t *elem = nvlist_next_nvpair(props, NULL); elem != NULL; elem = nvlist_next_nvpair(props, elem)) { if (zpool_prop_feature(nvpair_name(elem))) { has_features = B_TRUE; feat_name = strchr(nvpair_name(elem), '@') + 1; VERIFY0(zfeature_lookup_name(feat_name, &feat)); if (feat == SPA_FEATURE_ENCRYPTION) has_encryption = B_TRUE; if (feat == SPA_FEATURE_ALLOCATION_CLASSES) has_allocclass = B_TRUE; } } /* verify encryption params, if they were provided */ if (dcp != NULL) { error = spa_create_check_encryption_params(dcp, has_encryption); if (error != 0) { spa_deactivate(spa); spa_remove(spa); mutex_exit(&spa_namespace_lock); return (error); } } if (!has_allocclass && zfs_special_devs(nvroot, NULL)) { spa_deactivate(spa); spa_remove(spa); mutex_exit(&spa_namespace_lock); return (ENOTSUP); } if (has_features || nvlist_lookup_uint64(props, zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) { version = SPA_VERSION; } ASSERT(SPA_VERSION_IS_SUPPORTED(version)); spa->spa_first_txg = txg; spa->spa_uberblock.ub_txg = txg - 1; spa->spa_uberblock.ub_version = version; spa->spa_ubsync = spa->spa_uberblock; spa->spa_load_state = SPA_LOAD_CREATE; spa->spa_removing_phys.sr_state = DSS_NONE; spa->spa_removing_phys.sr_removing_vdev = -1; spa->spa_removing_phys.sr_prev_indirect_vdev = -1; spa->spa_indirect_vdevs_loaded = B_TRUE; /* * Create "The Godfather" zio to hold all async IOs */ spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *), KM_SLEEP); for (int i = 0; i < max_ncpus; i++) { spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_GODFATHER); } /* * Create the root vdev. */ spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD); ASSERT(error != 0 || rvd != NULL); ASSERT(error != 0 || spa->spa_root_vdev == rvd); if (error == 0 && !zfs_allocatable_devs(nvroot)) error = SET_ERROR(EINVAL); if (error == 0 && (error = vdev_create(rvd, txg, B_FALSE)) == 0 && (error = vdev_draid_spare_create(nvroot, rvd, &ndraid, 0)) == 0 && (error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) == 0) { /* * instantiate the metaslab groups (this will dirty the vdevs) * we can no longer error exit past this point */ for (int c = 0; error == 0 && c < rvd->vdev_children; c++) { vdev_t *vd = rvd->vdev_child[c]; vdev_metaslab_set_size(vd); vdev_expand(vd, txg); } } spa_config_exit(spa, SCL_ALL, FTAG); if (error != 0) { spa_unload(spa); spa_deactivate(spa); spa_remove(spa); mutex_exit(&spa_namespace_lock); return (error); } /* * Get the list of spares, if specified. */ if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0) { VERIFY(nvlist_alloc(&spa->spa_spares.sav_config, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, spares, nspares) == 0); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_load_spares(spa); spa_config_exit(spa, SCL_ALL, FTAG); spa->spa_spares.sav_sync = B_TRUE; } /* * Get the list of level 2 cache devices, if specified. */ if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0) { VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config, ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_load_l2cache(spa); spa_config_exit(spa, SCL_ALL, FTAG); spa->spa_l2cache.sav_sync = B_TRUE; } spa->spa_is_initializing = B_TRUE; spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg); spa->spa_is_initializing = B_FALSE; /* * Create DDTs (dedup tables). */ ddt_create(spa); spa_update_dspace(spa); tx = dmu_tx_create_assigned(dp, txg); /* * Create the pool's history object. */ if (version >= SPA_VERSION_ZPOOL_HISTORY && !spa->spa_history) spa_history_create_obj(spa, tx); spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE); spa_history_log_version(spa, "create", tx); /* * Create the pool config object. */ spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset, DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE, DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx); if (zap_add(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG, sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) { cmn_err(CE_PANIC, "failed to add pool config"); } if (zap_add(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION, sizeof (uint64_t), 1, &version, tx) != 0) { cmn_err(CE_PANIC, "failed to add pool version"); } /* Newly created pools with the right version are always deflated. */ if (version >= SPA_VERSION_RAIDZ_DEFLATE) { spa->spa_deflate = TRUE; if (zap_add(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE, sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) { cmn_err(CE_PANIC, "failed to add deflate"); } } /* * Create the deferred-free bpobj. Turn off compression * because sync-to-convergence takes longer if the blocksize * keeps changing. */ obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx); dmu_object_set_compress(spa->spa_meta_objset, obj, ZIO_COMPRESS_OFF, tx); if (zap_add(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ, sizeof (uint64_t), 1, &obj, tx) != 0) { cmn_err(CE_PANIC, "failed to add bpobj"); } VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj)); /* * Generate some random noise for salted checksums to operate on. */ (void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes, sizeof (spa->spa_cksum_salt.zcs_bytes)); /* * Set pool properties. */ spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS); spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION); spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE); spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND); spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST); spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM); if (props != NULL) { spa_configfile_set(spa, props, B_FALSE); spa_sync_props(props, tx); } for (int i = 0; i < ndraid; i++) spa_feature_incr(spa, SPA_FEATURE_DRAID, tx); dmu_tx_commit(tx); spa->spa_sync_on = B_TRUE; txg_sync_start(dp); mmp_thread_start(spa); txg_wait_synced(dp, txg); spa_spawn_aux_threads(spa); spa_write_cachefile(spa, B_FALSE, B_TRUE); /* * Don't count references from objsets that are already closed * and are making their way through the eviction process. */ spa_evicting_os_wait(spa); spa->spa_minref = zfs_refcount_count(&spa->spa_refcount); spa->spa_load_state = SPA_LOAD_NONE; mutex_exit(&spa_namespace_lock); return (0); } /* * Import a non-root pool into the system. */ int spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags) { spa_t *spa; char *altroot = NULL; spa_load_state_t state = SPA_LOAD_IMPORT; zpool_load_policy_t policy; spa_mode_t mode = spa_mode_global; uint64_t readonly = B_FALSE; int error; nvlist_t *nvroot; nvlist_t **spares, **l2cache; uint_t nspares, nl2cache; /* * If a pool with this name exists, return failure. */ mutex_enter(&spa_namespace_lock); if (spa_lookup(pool) != NULL) { mutex_exit(&spa_namespace_lock); return (SET_ERROR(EEXIST)); } /* * Create and initialize the spa structure. */ (void) nvlist_lookup_string(props, zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); (void) nvlist_lookup_uint64(props, zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly); if (readonly) mode = SPA_MODE_READ; spa = spa_add(pool, config, altroot); spa->spa_import_flags = flags; /* * Verbatim import - Take a pool and insert it into the namespace * as if it had been loaded at boot. */ if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) { if (props != NULL) spa_configfile_set(spa, props, B_FALSE); spa_write_cachefile(spa, B_FALSE, B_TRUE); spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT); zfs_dbgmsg("spa_import: verbatim import of %s", pool); mutex_exit(&spa_namespace_lock); return (0); } spa_activate(spa, mode); /* * Don't start async tasks until we know everything is healthy. */ spa_async_suspend(spa); zpool_get_load_policy(config, &policy); if (policy.zlp_rewind & ZPOOL_DO_REWIND) state = SPA_LOAD_RECOVER; spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT; if (state != SPA_LOAD_RECOVER) { spa->spa_last_ubsync_txg = spa->spa_load_txg = 0; zfs_dbgmsg("spa_import: importing %s", pool); } else { zfs_dbgmsg("spa_import: importing %s, max_txg=%lld " "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg); } error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind); /* * Propagate anything learned while loading the pool and pass it * back to caller (i.e. rewind info, missing devices, etc). */ VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info) == 0); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); /* * Toss any existing sparelist, as it doesn't have any validity * anymore, and conflicts with spa_has_spare(). */ if (spa->spa_spares.sav_config) { nvlist_free(spa->spa_spares.sav_config); spa->spa_spares.sav_config = NULL; spa_load_spares(spa); } if (spa->spa_l2cache.sav_config) { nvlist_free(spa->spa_l2cache.sav_config); spa->spa_l2cache.sav_config = NULL; spa_load_l2cache(spa); } VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0); spa_config_exit(spa, SCL_ALL, FTAG); if (props != NULL) spa_configfile_set(spa, props, B_FALSE); if (error != 0 || (props && spa_writeable(spa) && (error = spa_prop_set(spa, props)))) { spa_unload(spa); spa_deactivate(spa); spa_remove(spa); mutex_exit(&spa_namespace_lock); return (error); } spa_async_resume(spa); /* * Override any spares and level 2 cache devices as specified by * the user, as these may have correct device names/devids, etc. */ if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0) { if (spa->spa_spares.sav_config) VERIFY(nvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, DATA_TYPE_NVLIST_ARRAY) == 0); else VERIFY(nvlist_alloc(&spa->spa_spares.sav_config, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, spares, nspares) == 0); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_load_spares(spa); spa_config_exit(spa, SCL_ALL, FTAG); spa->spa_spares.sav_sync = B_TRUE; } if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0) { if (spa->spa_l2cache.sav_config) VERIFY(nvlist_remove(spa->spa_l2cache.sav_config, ZPOOL_CONFIG_L2CACHE, DATA_TYPE_NVLIST_ARRAY) == 0); else VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config, ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa_load_l2cache(spa); spa_config_exit(spa, SCL_ALL, FTAG); spa->spa_l2cache.sav_sync = B_TRUE; } /* * Check for any removed devices. */ if (spa->spa_autoreplace) { spa_aux_check_removed(&spa->spa_spares); spa_aux_check_removed(&spa->spa_l2cache); } if (spa_writeable(spa)) { /* * Update the config cache to include the newly-imported pool. */ spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); } /* * It's possible that the pool was expanded while it was exported. * We kick off an async task to handle this for us. */ spa_async_request(spa, SPA_ASYNC_AUTOEXPAND); spa_history_log_version(spa, "import", NULL); spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT); mutex_exit(&spa_namespace_lock); zvol_create_minors_recursive(pool); return (0); } nvlist_t * spa_tryimport(nvlist_t *tryconfig) { nvlist_t *config = NULL; char *poolname, *cachefile; spa_t *spa; uint64_t state; int error; zpool_load_policy_t policy; if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname)) return (NULL); if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state)) return (NULL); /* * Create and initialize the spa structure. */ mutex_enter(&spa_namespace_lock); spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL); spa_activate(spa, SPA_MODE_READ); /* * Rewind pool if a max txg was provided. */ zpool_get_load_policy(spa->spa_config, &policy); if (policy.zlp_txg != UINT64_MAX) { spa->spa_load_max_txg = policy.zlp_txg; spa->spa_extreme_rewind = B_TRUE; zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld", poolname, (longlong_t)policy.zlp_txg); } else { zfs_dbgmsg("spa_tryimport: importing %s", poolname); } if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile) == 0) { zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile); spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE; } else { spa->spa_config_source = SPA_CONFIG_SRC_SCAN; } error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING); /* * If 'tryconfig' was at least parsable, return the current config. */ if (spa->spa_root_vdev != NULL) { config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, poolname) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, state) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP, spa->spa_uberblock.ub_timestamp) == 0); VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_ERRATA, spa->spa_errata) == 0); /* * If the bootfs property exists on this pool then we * copy it out so that external consumers can tell which * pools are bootable. */ if ((!error || error == EEXIST) && spa->spa_bootfs) { char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP); /* * We have to play games with the name since the * pool was opened as TRYIMPORT_NAME. */ if (dsl_dsobj_to_dsname(spa_name(spa), spa->spa_bootfs, tmpname) == 0) { char *cp; char *dsname; dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP); cp = strchr(tmpname, '/'); if (cp == NULL) { (void) strlcpy(dsname, tmpname, MAXPATHLEN); } else { (void) snprintf(dsname, MAXPATHLEN, "%s/%s", poolname, ++cp); } VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_BOOTFS, dsname) == 0); kmem_free(dsname, MAXPATHLEN); } kmem_free(tmpname, MAXPATHLEN); } /* * Add the list of hot spares and level 2 cache devices. */ spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); spa_add_spares(spa, config); spa_add_l2cache(spa, config); spa_config_exit(spa, SCL_CONFIG, FTAG); } spa_unload(spa); spa_deactivate(spa); spa_remove(spa); mutex_exit(&spa_namespace_lock); return (config); } /* * Pool export/destroy * * The act of destroying or exporting a pool is very simple. We make sure there * is no more pending I/O and any references to the pool are gone. Then, we * update the pool state and sync all the labels to disk, removing the * configuration from the cache afterwards. If the 'hardforce' flag is set, then * we don't sync the labels or remove the configuration cache. */ static int spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig, boolean_t force, boolean_t hardforce) { int error; spa_t *spa; if (oldconfig) *oldconfig = NULL; if (!(spa_mode_global & SPA_MODE_WRITE)) return (SET_ERROR(EROFS)); mutex_enter(&spa_namespace_lock); if ((spa = spa_lookup(pool)) == NULL) { mutex_exit(&spa_namespace_lock); return (SET_ERROR(ENOENT)); } if (spa->spa_is_exporting) { /* the pool is being exported by another thread */ mutex_exit(&spa_namespace_lock); return (SET_ERROR(ZFS_ERR_EXPORT_IN_PROGRESS)); } spa->spa_is_exporting = B_TRUE; /* * Put a hold on the pool, drop the namespace lock, stop async tasks, * reacquire the namespace lock, and see if we can export. */ spa_open_ref(spa, FTAG); mutex_exit(&spa_namespace_lock); spa_async_suspend(spa); if (spa->spa_zvol_taskq) { zvol_remove_minors(spa, spa_name(spa), B_TRUE); taskq_wait(spa->spa_zvol_taskq); } mutex_enter(&spa_namespace_lock); spa_close(spa, FTAG); if (spa->spa_state == POOL_STATE_UNINITIALIZED) goto export_spa; /* * The pool will be in core if it's openable, in which case we can * modify its state. Objsets may be open only because they're dirty, * so we have to force it to sync before checking spa_refcnt. */ if (spa->spa_sync_on) { txg_wait_synced(spa->spa_dsl_pool, 0); spa_evicting_os_wait(spa); } /* * A pool cannot be exported or destroyed if there are active * references. If we are resetting a pool, allow references by * fault injection handlers. */ if (!spa_refcount_zero(spa) || (spa->spa_inject_ref != 0)) { error = SET_ERROR(EBUSY); goto fail; } if (spa->spa_sync_on) { /* * A pool cannot be exported if it has an active shared spare. * This is to prevent other pools stealing the active spare * from an exported pool. At user's own will, such pool can * be forcedly exported. */ if (!force && new_state == POOL_STATE_EXPORTED && spa_has_active_shared_spare(spa)) { error = SET_ERROR(EXDEV); goto fail; } /* * We're about to export or destroy this pool. Make sure * we stop all initialization and trim activity here before * we set the spa_final_txg. This will ensure that all * dirty data resulting from the initialization is * committed to disk before we unload the pool. */ if (spa->spa_root_vdev != NULL) { vdev_t *rvd = spa->spa_root_vdev; vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE); vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE); vdev_autotrim_stop_all(spa); vdev_rebuild_stop_all(spa); } /* * We want this to be reflected on every label, * so mark them all dirty. spa_unload() will do the * final sync that pushes these changes out. */ if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) { spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); spa->spa_state = new_state; spa->spa_final_txg = spa_last_synced_txg(spa) + TXG_DEFER_SIZE + 1; vdev_config_dirty(spa->spa_root_vdev); spa_config_exit(spa, SCL_ALL, FTAG); } } export_spa: if (new_state == POOL_STATE_DESTROYED) spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY); else if (new_state == POOL_STATE_EXPORTED) spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_EXPORT); if (spa->spa_state != POOL_STATE_UNINITIALIZED) { spa_unload(spa); spa_deactivate(spa); } if (oldconfig && spa->spa_config) VERIFY(nvlist_dup(spa->spa_config, oldconfig, 0) == 0); if (new_state != POOL_STATE_UNINITIALIZED) { if (!hardforce) spa_write_cachefile(spa, B_TRUE, B_TRUE); spa_remove(spa); } else { /* * If spa_remove() is not called for this spa_t and * there is any possibility that it can be reused, * we make sure to reset the exporting flag. */ spa->spa_is_exporting = B_FALSE; } mutex_exit(&spa_namespace_lock); return (0); fail: spa->spa_is_exporting = B_FALSE; spa_async_resume(spa); mutex_exit(&spa_namespace_lock); return (error); } /* * Destroy a storage pool. */ int spa_destroy(const char *pool) { return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL, B_FALSE, B_FALSE)); } /* * Export a storage pool. */ int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force, boolean_t hardforce) { return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig, force, hardforce)); } /* * Similar to spa_export(), this unloads the spa_t without actually removing it * from the namespace in any way. */ int spa_reset(const char *pool) { return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL, B_FALSE, B_FALSE)); } /* * ========================================================================== * Device manipulation * ========================================================================== */ /* * This is called as a synctask to increment the draid feature flag */ static void spa_draid_feature_incr(void *arg, dmu_tx_t *tx) { spa_t *spa = dmu_tx_pool(tx)->dp_spa; int draid = (int)(uintptr_t)arg; for (int c = 0; c < draid; c++) spa_feature_incr(spa, SPA_FEATURE_DRAID, tx); } /* * Add a device to a storage pool. */ int spa_vdev_add(spa_t *spa, nvlist_t *nvroot) { uint64_t txg, ndraid = 0; int error; vdev_t *rvd = spa->spa_root_vdev; vdev_t *vd, *tvd; nvlist_t **spares, **l2cache; uint_t nspares, nl2cache; ASSERT(spa_writeable(spa)); txg = spa_vdev_enter(spa); if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0, VDEV_ALLOC_ADD)) != 0) return (spa_vdev_exit(spa, NULL, txg, error)); spa->spa_pending_vdev = vd; /* spa_vdev_exit() will clear this */ if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0) nspares = 0; if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) != 0) nl2cache = 0; if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0) return (spa_vdev_exit(spa, vd, txg, EINVAL)); if (vd->vdev_children != 0 && (error = vdev_create(vd, txg, B_FALSE)) != 0) { return (spa_vdev_exit(spa, vd, txg, error)); } /* * The virtual dRAID spares must be added after vdev tree is created * and the vdev guids are generated. The guid of their associated * dRAID is stored in the config and used when opening the spare. */ if ((error = vdev_draid_spare_create(nvroot, vd, &ndraid, rvd->vdev_children)) == 0) { if (ndraid > 0 && nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0) nspares = 0; } else { return (spa_vdev_exit(spa, vd, txg, error)); } /* * We must validate the spares and l2cache devices after checking the * children. Otherwise, vdev_inuse() will blindly overwrite the spare. */ if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0) return (spa_vdev_exit(spa, vd, txg, error)); /* * If we are in the middle of a device removal, we can only add * devices which match the existing devices in the pool. * If we are in the middle of a removal, or have some indirect * vdevs, we can not add raidz or dRAID top levels. */ if (spa->spa_vdev_removal != NULL || spa->spa_removing_phys.sr_prev_indirect_vdev != -1) { for (int c = 0; c < vd->vdev_children; c++) { tvd = vd->vdev_child[c]; if (spa->spa_vdev_removal != NULL && tvd->vdev_ashift != spa->spa_max_ashift) { return (spa_vdev_exit(spa, vd, txg, EINVAL)); } /* Fail if top level vdev is raidz or a dRAID */ if (vdev_get_nparity(tvd) != 0) return (spa_vdev_exit(spa, vd, txg, EINVAL)); /* * Need the top level mirror to be * a mirror of leaf vdevs only */ if (tvd->vdev_ops == &vdev_mirror_ops) { for (uint64_t cid = 0; cid < tvd->vdev_children; cid++) { vdev_t *cvd = tvd->vdev_child[cid]; if (!cvd->vdev_ops->vdev_op_leaf) { return (spa_vdev_exit(spa, vd, txg, EINVAL)); } } } } } for (int c = 0; c < vd->vdev_children; c++) { tvd = vd->vdev_child[c]; vdev_remove_child(vd, tvd); tvd->vdev_id = rvd->vdev_children; vdev_add_child(rvd, tvd); vdev_config_dirty(tvd); } if (nspares != 0) { spa_set_aux_vdevs(&spa->spa_spares, spares, nspares, ZPOOL_CONFIG_SPARES); spa_load_spares(spa); spa->spa_spares.sav_sync = B_TRUE; } if (nl2cache != 0) { spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache, ZPOOL_CONFIG_L2CACHE); spa_load_l2cache(spa); spa->spa_l2cache.sav_sync = B_TRUE; } /* * We can't increment a feature while holding spa_vdev so we * have to do it in a synctask. */ if (ndraid != 0) { dmu_tx_t *tx; tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); dsl_sync_task_nowait(spa->spa_dsl_pool, spa_draid_feature_incr, (void *)(uintptr_t)ndraid, tx); dmu_tx_commit(tx); } /* * We have to be careful when adding new vdevs to an existing pool. * If other threads start allocating from these vdevs before we * sync the config cache, and we lose power, then upon reboot we may * fail to open the pool because there are DVAs that the config cache * can't translate. Therefore, we first add the vdevs without * initializing metaslabs; sync the config cache (via spa_vdev_exit()); * and then let spa_config_update() initialize the new metaslabs. * * spa_load() checks for added-but-not-initialized vdevs, so that * if we lose power at any point in this sequence, the remaining * steps will be completed the next time we load the pool. */ (void) spa_vdev_exit(spa, vd, txg, 0); mutex_enter(&spa_namespace_lock); spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD); mutex_exit(&spa_namespace_lock); return (0); } /* * Attach a device to a mirror. The arguments are the path to any device * in the mirror, and the nvroot for the new device. If the path specifies * a device that is not mirrored, we automatically insert the mirror vdev. * * If 'replacing' is specified, the new device is intended to replace the * existing device; in this case the two devices are made into their own * mirror using the 'replacing' vdev, which is functionally identical to * the mirror vdev (it actually reuses all the same ops) but has a few * extra rules: you can't attach to it after it's been created, and upon * completion of resilvering, the first disk (the one being replaced) * is automatically detached. * * If 'rebuild' is specified, then sequential reconstruction (a.ka. rebuild) * should be performed instead of traditional healing reconstruction. From * an administrators perspective these are both resilver operations. */ int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing, int rebuild) { uint64_t txg, dtl_max_txg; vdev_t *rvd = spa->spa_root_vdev; vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd; vdev_ops_t *pvops; char *oldvdpath, *newvdpath; int newvd_isspare; int error; ASSERT(spa_writeable(spa)); txg = spa_vdev_enter(spa); oldvd = spa_lookup_by_guid(spa, guid, B_FALSE); ASSERT(MUTEX_HELD(&spa_namespace_lock)); if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { error = (spa_has_checkpoint(spa)) ? ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; return (spa_vdev_exit(spa, NULL, txg, error)); } if (rebuild) { if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REBUILD)) return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); if (dsl_scan_resilvering(spa_get_dsl(spa))) return (spa_vdev_exit(spa, NULL, txg, ZFS_ERR_RESILVER_IN_PROGRESS)); } else { if (vdev_rebuild_active(rvd)) return (spa_vdev_exit(spa, NULL, txg, ZFS_ERR_REBUILD_IN_PROGRESS)); } if (spa->spa_vdev_removal != NULL) return (spa_vdev_exit(spa, NULL, txg, EBUSY)); if (oldvd == NULL) return (spa_vdev_exit(spa, NULL, txg, ENODEV)); if (!oldvd->vdev_ops->vdev_op_leaf) return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); pvd = oldvd->vdev_parent; if ((error = spa_config_parse(spa, &newrootvd, nvroot, NULL, 0, VDEV_ALLOC_ATTACH)) != 0) return (spa_vdev_exit(spa, NULL, txg, EINVAL)); if (newrootvd->vdev_children != 1) return (spa_vdev_exit(spa, newrootvd, txg, EINVAL)); newvd = newrootvd->vdev_child[0]; if (!newvd->vdev_ops->vdev_op_leaf) return (spa_vdev_exit(spa, newrootvd, txg, EINVAL)); if ((error = vdev_create(newrootvd, txg, replacing)) != 0) return (spa_vdev_exit(spa, newrootvd, txg, error)); /* * Spares can't replace logs */ if (oldvd->vdev_top->vdev_islog && newvd->vdev_isspare) return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); /* * A dRAID spare can only replace a child of its parent dRAID vdev. */ if (newvd->vdev_ops == &vdev_draid_spare_ops && oldvd->vdev_top != vdev_draid_spare_get_parent(newvd)) { return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); } if (rebuild) { /* * For rebuilds, the top vdev must support reconstruction * using only space maps. This means the only allowable * vdevs types are the root vdev, a mirror, or dRAID. */ tvd = pvd; if (pvd->vdev_top != NULL) tvd = pvd->vdev_top; if (tvd->vdev_ops != &vdev_mirror_ops && tvd->vdev_ops != &vdev_root_ops && tvd->vdev_ops != &vdev_draid_ops) { return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); } } if (!replacing) { /* * For attach, the only allowable parent is a mirror or the root * vdev. */ if (pvd->vdev_ops != &vdev_mirror_ops && pvd->vdev_ops != &vdev_root_ops) return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); pvops = &vdev_mirror_ops; } else { /* * Active hot spares can only be replaced by inactive hot * spares. */ if (pvd->vdev_ops == &vdev_spare_ops && oldvd->vdev_isspare && !spa_has_spare(spa, newvd->vdev_guid)) return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); /* * If the source is a hot spare, and the parent isn't already a * spare, then we want to create a new hot spare. Otherwise, we * want to create a replacing vdev. The user is not allowed to * attach to a spared vdev child unless the 'isspare' state is * the same (spare replaces spare, non-spare replaces * non-spare). */ if (pvd->vdev_ops == &vdev_replacing_ops && spa_version(spa) < SPA_VERSION_MULTI_REPLACE) { return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); } else if (pvd->vdev_ops == &vdev_spare_ops && newvd->vdev_isspare != oldvd->vdev_isspare) { return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); } if (newvd->vdev_isspare) pvops = &vdev_spare_ops; else pvops = &vdev_replacing_ops; } /* * Make sure the new device is big enough. */ if (newvd->vdev_asize < vdev_get_min_asize(oldvd)) return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW)); /* * The new device cannot have a higher alignment requirement * than the top-level vdev. */ if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift) return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); /* * If this is an in-place replacement, update oldvd's path and devid * to make it distinguishable from newvd, and unopenable from now on. */ if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) { spa_strfree(oldvd->vdev_path); oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5, KM_SLEEP); (void) snprintf(oldvd->vdev_path, strlen(newvd->vdev_path) + 5, "%s/%s", newvd->vdev_path, "old"); if (oldvd->vdev_devid != NULL) { spa_strfree(oldvd->vdev_devid); oldvd->vdev_devid = NULL; } } /* * If the parent is not a mirror, or if we're replacing, insert the new * mirror/replacing/spare vdev above oldvd. */ if (pvd->vdev_ops != pvops) pvd = vdev_add_parent(oldvd, pvops); ASSERT(pvd->vdev_top->vdev_parent == rvd); ASSERT(pvd->vdev_ops == pvops); ASSERT(oldvd->vdev_parent == pvd); /* * Extract the new device from its root and add it to pvd. */ vdev_remove_child(newrootvd, newvd); newvd->vdev_id = pvd->vdev_children; newvd->vdev_crtxg = oldvd->vdev_crtxg; vdev_add_child(pvd, newvd); /* * Reevaluate the parent vdev state. */ vdev_propagate_state(pvd); tvd = newvd->vdev_top; ASSERT(pvd->vdev_top == tvd); ASSERT(tvd->vdev_parent == rvd); vdev_config_dirty(tvd); /* * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account * for any dmu_sync-ed blocks. It will propagate upward when * spa_vdev_exit() calls vdev_dtl_reassess(). */ dtl_max_txg = txg + TXG_CONCURRENT_STATES; vdev_dtl_dirty(newvd, DTL_MISSING, TXG_INITIAL, dtl_max_txg - TXG_INITIAL); if (newvd->vdev_isspare) { spa_spare_activate(newvd); spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE); } oldvdpath = spa_strdup(oldvd->vdev_path); newvdpath = spa_strdup(newvd->vdev_path); newvd_isspare = newvd->vdev_isspare; /* * Mark newvd's DTL dirty in this txg. */ vdev_dirty(tvd, VDD_DTL, newvd, txg); /* * Schedule the resilver or rebuild to restart in the future. We do * this to ensure that dmu_sync-ed blocks have been stitched into the * respective datasets. */ if (rebuild) { newvd->vdev_rebuild_txg = txg; vdev_rebuild(tvd); } else { newvd->vdev_resilver_txg = txg; if (dsl_scan_resilvering(spa_get_dsl(spa)) && spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) { vdev_defer_resilver(newvd); } else { dsl_scan_restart_resilver(spa->spa_dsl_pool, dtl_max_txg); } } if (spa->spa_bootfs) spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH); spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH); /* * Commit the config */ (void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0); spa_history_log_internal(spa, "vdev attach", NULL, "%s vdev=%s %s vdev=%s", replacing && newvd_isspare ? "spare in" : replacing ? "replace" : "attach", newvdpath, replacing ? "for" : "to", oldvdpath); spa_strfree(oldvdpath); spa_strfree(newvdpath); return (0); } /* * Detach a device from a mirror or replacing vdev. * * If 'replace_done' is specified, only detach if the parent * is a replacing vdev. */ int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done) { uint64_t txg; int error; vdev_t *rvd __maybe_unused = spa->spa_root_vdev; vdev_t *vd, *pvd, *cvd, *tvd; boolean_t unspare = B_FALSE; uint64_t unspare_guid = 0; char *vdpath; ASSERT(spa_writeable(spa)); txg = spa_vdev_detach_enter(spa, guid); vd = spa_lookup_by_guid(spa, guid, B_FALSE); /* * Besides being called directly from the userland through the * ioctl interface, spa_vdev_detach() can be potentially called * at the end of spa_vdev_resilver_done(). * * In the regular case, when we have a checkpoint this shouldn't * happen as we never empty the DTLs of a vdev during the scrub * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done() * should never get here when we have a checkpoint. * * That said, even in a case when we checkpoint the pool exactly * as spa_vdev_resilver_done() calls this function everything * should be fine as the resilver will return right away. */ ASSERT(MUTEX_HELD(&spa_namespace_lock)); if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { error = (spa_has_checkpoint(spa)) ? ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; return (spa_vdev_exit(spa, NULL, txg, error)); } if (vd == NULL) return (spa_vdev_exit(spa, NULL, txg, ENODEV)); if (!vd->vdev_ops->vdev_op_leaf) return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); pvd = vd->vdev_parent; /* * If the parent/child relationship is not as expected, don't do it. * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing * vdev that's replacing B with C. The user's intent in replacing * is to go from M(A,B) to M(A,C). If the user decides to cancel * the replace by detaching C, the expected behavior is to end up * M(A,B). But suppose that right after deciding to detach C, * the replacement of B completes. We would have M(A,C), and then * ask to detach C, which would leave us with just A -- not what * the user wanted. To prevent this, we make sure that the * parent/child relationship hasn't changed -- in this example, * that C's parent is still the replacing vdev R. */ if (pvd->vdev_guid != pguid && pguid != 0) return (spa_vdev_exit(spa, NULL, txg, EBUSY)); /* * Only 'replacing' or 'spare' vdevs can be replaced. */ if (replace_done && pvd->vdev_ops != &vdev_replacing_ops && pvd->vdev_ops != &vdev_spare_ops) return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); ASSERT(pvd->vdev_ops != &vdev_spare_ops || spa_version(spa) >= SPA_VERSION_SPARES); /* * Only mirror, replacing, and spare vdevs support detach. */ if (pvd->vdev_ops != &vdev_replacing_ops && pvd->vdev_ops != &vdev_mirror_ops && pvd->vdev_ops != &vdev_spare_ops) return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); /* * If this device has the only valid copy of some data, * we cannot safely detach it. */ if (vdev_dtl_required(vd)) return (spa_vdev_exit(spa, NULL, txg, EBUSY)); ASSERT(pvd->vdev_children >= 2); /* * If we are detaching the second disk from a replacing vdev, then * check to see if we changed the original vdev's path to have "/old" * at the end in spa_vdev_attach(). If so, undo that change now. */ if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 && vd->vdev_path != NULL) { size_t len = strlen(vd->vdev_path); for (int c = 0; c < pvd->vdev_children; c++) { cvd = pvd->vdev_child[c]; if (cvd == vd || cvd->vdev_path == NULL) continue; if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 && strcmp(cvd->vdev_path + len, "/old") == 0) { spa_strfree(cvd->vdev_path); cvd->vdev_path = spa_strdup(vd->vdev_path); break; } } } /* * If we are detaching the original disk from a normal spare, then it * implies that the spare should become a real disk, and be removed * from the active spare list for the pool. dRAID spares on the * other hand are coupled to the pool and thus should never be removed * from the spares list. */ if (pvd->vdev_ops == &vdev_spare_ops && vd->vdev_id == 0) { vdev_t *last_cvd = pvd->vdev_child[pvd->vdev_children - 1]; if (last_cvd->vdev_isspare && last_cvd->vdev_ops != &vdev_draid_spare_ops) { unspare = B_TRUE; } } /* * Erase the disk labels so the disk can be used for other things. * This must be done after all other error cases are handled, * but before we disembowel vd (so we can still do I/O to it). * But if we can't do it, don't treat the error as fatal -- * it may be that the unwritability of the disk is the reason * it's being detached! */ error = vdev_label_init(vd, 0, VDEV_LABEL_REMOVE); /* * Remove vd from its parent and compact the parent's children. */ vdev_remove_child(pvd, vd); vdev_compact_children(pvd); /* * Remember one of the remaining children so we can get tvd below. */ cvd = pvd->vdev_child[pvd->vdev_children - 1]; /* * If we need to remove the remaining child from the list of hot spares, * do it now, marking the vdev as no longer a spare in the process. * We must do this before vdev_remove_parent(), because that can * change the GUID if it creates a new toplevel GUID. For a similar * reason, we must remove the spare now, in the same txg as the detach; * otherwise someone could attach a new sibling, change the GUID, and * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail. */ if (unspare) { ASSERT(cvd->vdev_isspare); spa_spare_remove(cvd); unspare_guid = cvd->vdev_guid; (void) spa_vdev_remove(spa, unspare_guid, B_TRUE); cvd->vdev_unspare = B_TRUE; } /* * If the parent mirror/replacing vdev only has one child, * the parent is no longer needed. Remove it from the tree. */ if (pvd->vdev_children == 1) { if (pvd->vdev_ops == &vdev_spare_ops) cvd->vdev_unspare = B_FALSE; vdev_remove_parent(cvd); } /* * We don't set tvd until now because the parent we just removed * may have been the previous top-level vdev. */ tvd = cvd->vdev_top; ASSERT(tvd->vdev_parent == rvd); /* * Reevaluate the parent vdev state. */ vdev_propagate_state(cvd); /* * If the 'autoexpand' property is set on the pool then automatically * try to expand the size of the pool. For example if the device we * just detached was smaller than the others, it may be possible to * add metaslabs (i.e. grow the pool). We need to reopen the vdev * first so that we can obtain the updated sizes of the leaf vdevs. */ if (spa->spa_autoexpand) { vdev_reopen(tvd); vdev_expand(tvd, txg); } vdev_config_dirty(tvd); /* * Mark vd's DTL as dirty in this txg. vdev_dtl_sync() will see that * vd->vdev_detached is set and free vd's DTL object in syncing context. * But first make sure we're not on any *other* txg's DTL list, to * prevent vd from being accessed after it's freed. */ vdpath = spa_strdup(vd->vdev_path ? vd->vdev_path : "none"); for (int t = 0; t < TXG_SIZE; t++) (void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t); vd->vdev_detached = B_TRUE; vdev_dirty(tvd, VDD_DTL, vd, txg); spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE); spa_notify_waiters(spa); /* hang on to the spa before we release the lock */ spa_open_ref(spa, FTAG); error = spa_vdev_exit(spa, vd, txg, 0); spa_history_log_internal(spa, "detach", NULL, "vdev=%s", vdpath); spa_strfree(vdpath); /* * If this was the removal of the original device in a hot spare vdev, * then we want to go through and remove the device from the hot spare * list of every other pool. */ if (unspare) { spa_t *altspa = NULL; mutex_enter(&spa_namespace_lock); while ((altspa = spa_next(altspa)) != NULL) { if (altspa->spa_state != POOL_STATE_ACTIVE || altspa == spa) continue; spa_open_ref(altspa, FTAG); mutex_exit(&spa_namespace_lock); (void) spa_vdev_remove(altspa, unspare_guid, B_TRUE); mutex_enter(&spa_namespace_lock); spa_close(altspa, FTAG); } mutex_exit(&spa_namespace_lock); /* search the rest of the vdevs for spares to remove */ spa_vdev_resilver_done(spa); } /* all done with the spa; OK to release */ mutex_enter(&spa_namespace_lock); spa_close(spa, FTAG); mutex_exit(&spa_namespace_lock); return (error); } static int spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type, list_t *vd_list) { ASSERT(MUTEX_HELD(&spa_namespace_lock)); spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); /* Look up vdev and ensure it's a leaf. */ vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE); if (vd == NULL || vd->vdev_detached) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(ENODEV)); } else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(EINVAL)); } else if (!vdev_writeable(vd)) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(EROFS)); } mutex_enter(&vd->vdev_initialize_lock); spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); /* * When we activate an initialize action we check to see * if the vdev_initialize_thread is NULL. We do this instead * of using the vdev_initialize_state since there might be * a previous initialization process which has completed but * the thread is not exited. */ if (cmd_type == POOL_INITIALIZE_START && (vd->vdev_initialize_thread != NULL || vd->vdev_top->vdev_removing)) { mutex_exit(&vd->vdev_initialize_lock); return (SET_ERROR(EBUSY)); } else if (cmd_type == POOL_INITIALIZE_CANCEL && (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE && vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) { mutex_exit(&vd->vdev_initialize_lock); return (SET_ERROR(ESRCH)); } else if (cmd_type == POOL_INITIALIZE_SUSPEND && vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) { mutex_exit(&vd->vdev_initialize_lock); return (SET_ERROR(ESRCH)); } switch (cmd_type) { case POOL_INITIALIZE_START: vdev_initialize(vd); break; case POOL_INITIALIZE_CANCEL: vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, vd_list); break; case POOL_INITIALIZE_SUSPEND: vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED, vd_list); break; default: panic("invalid cmd_type %llu", (unsigned long long)cmd_type); } mutex_exit(&vd->vdev_initialize_lock); return (0); } int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, nvlist_t *vdev_errlist) { int total_errors = 0; list_t vd_list; list_create(&vd_list, sizeof (vdev_t), offsetof(vdev_t, vdev_initialize_node)); /* * We hold the namespace lock through the whole function * to prevent any changes to the pool while we're starting or * stopping initialization. The config and state locks are held so that * we can properly assess the vdev state before we commit to * the initializing operation. */ mutex_enter(&spa_namespace_lock); for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL); pair != NULL; pair = nvlist_next_nvpair(nv, pair)) { uint64_t vdev_guid = fnvpair_value_uint64(pair); int error = spa_vdev_initialize_impl(spa, vdev_guid, cmd_type, &vd_list); if (error != 0) { char guid_as_str[MAXNAMELEN]; (void) snprintf(guid_as_str, sizeof (guid_as_str), "%llu", (unsigned long long)vdev_guid); fnvlist_add_int64(vdev_errlist, guid_as_str, error); total_errors++; } } /* Wait for all initialize threads to stop. */ vdev_initialize_stop_wait(spa, &vd_list); /* Sync out the initializing state */ txg_wait_synced(spa->spa_dsl_pool, 0); mutex_exit(&spa_namespace_lock); list_destroy(&vd_list); return (total_errors); } static int spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type, uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list) { ASSERT(MUTEX_HELD(&spa_namespace_lock)); spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); /* Look up vdev and ensure it's a leaf. */ vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE); if (vd == NULL || vd->vdev_detached) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(ENODEV)); } else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(EINVAL)); } else if (!vdev_writeable(vd)) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(EROFS)); } else if (!vd->vdev_has_trim) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(EOPNOTSUPP)); } else if (secure && !vd->vdev_has_securetrim) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (SET_ERROR(EOPNOTSUPP)); } mutex_enter(&vd->vdev_trim_lock); spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); /* * When we activate a TRIM action we check to see if the * vdev_trim_thread is NULL. We do this instead of using the * vdev_trim_state since there might be a previous TRIM process * which has completed but the thread is not exited. */ if (cmd_type == POOL_TRIM_START && (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) { mutex_exit(&vd->vdev_trim_lock); return (SET_ERROR(EBUSY)); } else if (cmd_type == POOL_TRIM_CANCEL && (vd->vdev_trim_state != VDEV_TRIM_ACTIVE && vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) { mutex_exit(&vd->vdev_trim_lock); return (SET_ERROR(ESRCH)); } else if (cmd_type == POOL_TRIM_SUSPEND && vd->vdev_trim_state != VDEV_TRIM_ACTIVE) { mutex_exit(&vd->vdev_trim_lock); return (SET_ERROR(ESRCH)); } switch (cmd_type) { case POOL_TRIM_START: vdev_trim(vd, rate, partial, secure); break; case POOL_TRIM_CANCEL: vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list); break; case POOL_TRIM_SUSPEND: vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list); break; default: panic("invalid cmd_type %llu", (unsigned long long)cmd_type); } mutex_exit(&vd->vdev_trim_lock); return (0); } /* * Initiates a manual TRIM for the requested vdevs. This kicks off individual * TRIM threads for each child vdev. These threads pass over all of the free * space in the vdev's metaslabs and issues TRIM commands for that space. */ int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist) { int total_errors = 0; list_t vd_list; list_create(&vd_list, sizeof (vdev_t), offsetof(vdev_t, vdev_trim_node)); /* * We hold the namespace lock through the whole function * to prevent any changes to the pool while we're starting or * stopping TRIM. The config and state locks are held so that * we can properly assess the vdev state before we commit to * the TRIM operation. */ mutex_enter(&spa_namespace_lock); for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL); pair != NULL; pair = nvlist_next_nvpair(nv, pair)) { uint64_t vdev_guid = fnvpair_value_uint64(pair); int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type, rate, partial, secure, &vd_list); if (error != 0) { char guid_as_str[MAXNAMELEN]; (void) snprintf(guid_as_str, sizeof (guid_as_str), "%llu", (unsigned long long)vdev_guid); fnvlist_add_int64(vdev_errlist, guid_as_str, error); total_errors++; } } /* Wait for all TRIM threads to stop. */ vdev_trim_stop_wait(spa, &vd_list); /* Sync out the TRIM state */ txg_wait_synced(spa->spa_dsl_pool, 0); mutex_exit(&spa_namespace_lock); list_destroy(&vd_list); return (total_errors); } /* * Split a set of devices from their mirrors, and create a new pool from them. */ int spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config, nvlist_t *props, boolean_t exp) { int error = 0; uint64_t txg, *glist; spa_t *newspa; uint_t c, children, lastlog; nvlist_t **child, *nvl, *tmp; dmu_tx_t *tx; char *altroot = NULL; vdev_t *rvd, **vml = NULL; /* vdev modify list */ boolean_t activate_slog; ASSERT(spa_writeable(spa)); txg = spa_vdev_enter(spa); ASSERT(MUTEX_HELD(&spa_namespace_lock)); if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { error = (spa_has_checkpoint(spa)) ? ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; return (spa_vdev_exit(spa, NULL, txg, error)); } /* clear the log and flush everything up to now */ activate_slog = spa_passivate_log(spa); (void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG); error = spa_reset_logs(spa); txg = spa_vdev_config_enter(spa); if (activate_slog) spa_activate_log(spa); if (error != 0) return (spa_vdev_exit(spa, NULL, txg, error)); /* check new spa name before going any further */ if (spa_lookup(newname) != NULL) return (spa_vdev_exit(spa, NULL, txg, EEXIST)); /* * scan through all the children to ensure they're all mirrors */ if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 || nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child, &children) != 0) return (spa_vdev_exit(spa, NULL, txg, EINVAL)); /* first, check to ensure we've got the right child count */ rvd = spa->spa_root_vdev; lastlog = 0; for (c = 0; c < rvd->vdev_children; c++) { vdev_t *vd = rvd->vdev_child[c]; /* don't count the holes & logs as children */ if (vd->vdev_islog || (vd->vdev_ops != &vdev_indirect_ops && !vdev_is_concrete(vd))) { if (lastlog == 0) lastlog = c; continue; } lastlog = 0; } if (children != (lastlog != 0 ? lastlog : rvd->vdev_children)) return (spa_vdev_exit(spa, NULL, txg, EINVAL)); /* next, ensure no spare or cache devices are part of the split */ if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 || nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0) return (spa_vdev_exit(spa, NULL, txg, EINVAL)); vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP); glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP); /* then, loop over each vdev and validate it */ for (c = 0; c < children; c++) { uint64_t is_hole = 0; (void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE, &is_hole); if (is_hole != 0) { if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole || spa->spa_root_vdev->vdev_child[c]->vdev_islog) { continue; } else { error = SET_ERROR(EINVAL); break; } } /* deal with indirect vdevs */ if (spa->spa_root_vdev->vdev_child[c]->vdev_ops == &vdev_indirect_ops) continue; /* which disk is going to be split? */ if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID, &glist[c]) != 0) { error = SET_ERROR(EINVAL); break; } /* look it up in the spa */ vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE); if (vml[c] == NULL) { error = SET_ERROR(ENODEV); break; } /* make sure there's nothing stopping the split */ if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops || vml[c]->vdev_islog || !vdev_is_concrete(vml[c]) || vml[c]->vdev_isspare || vml[c]->vdev_isl2cache || !vdev_writeable(vml[c]) || vml[c]->vdev_children != 0 || vml[c]->vdev_state != VDEV_STATE_HEALTHY || c != spa->spa_root_vdev->vdev_child[c]->vdev_id) { error = SET_ERROR(EINVAL); break; } if (vdev_dtl_required(vml[c]) || vdev_resilver_needed(vml[c], NULL, NULL)) { error = SET_ERROR(EBUSY); break; } /* we need certain info from the top level */ VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY, vml[c]->vdev_top->vdev_ms_array) == 0); VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT, vml[c]->vdev_top->vdev_ms_shift) == 0); VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE, vml[c]->vdev_top->vdev_asize) == 0); VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT, vml[c]->vdev_top->vdev_ashift) == 0); /* transfer per-vdev ZAPs */ ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0); VERIFY0(nvlist_add_uint64(child[c], ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap)); ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0); VERIFY0(nvlist_add_uint64(child[c], ZPOOL_CONFIG_VDEV_TOP_ZAP, vml[c]->vdev_parent->vdev_top_zap)); } if (error != 0) { kmem_free(vml, children * sizeof (vdev_t *)); kmem_free(glist, children * sizeof (uint64_t)); return (spa_vdev_exit(spa, NULL, txg, error)); } /* stop writers from using the disks */ for (c = 0; c < children; c++) { if (vml[c] != NULL) vml[c]->vdev_offline = B_TRUE; } vdev_reopen(spa->spa_root_vdev); /* * Temporarily record the splitting vdevs in the spa config. This * will disappear once the config is regenerated. */ VERIFY(nvlist_alloc(&nvl, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, glist, children) == 0); kmem_free(glist, children * sizeof (uint64_t)); mutex_enter(&spa->spa_props_lock); VERIFY(nvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT, nvl) == 0); mutex_exit(&spa->spa_props_lock); spa->spa_config_splitting = nvl; vdev_config_dirty(spa->spa_root_vdev); /* configure and create the new pool */ VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, spa_version(spa)) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG, spa->spa_config_txg) == 0); VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID, spa_generate_guid(NULL)) == 0); VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)); (void) nvlist_lookup_string(props, zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); /* add the new pool to the namespace */ newspa = spa_add(newname, config, altroot); newspa->spa_avz_action = AVZ_ACTION_REBUILD; newspa->spa_config_txg = spa->spa_config_txg; spa_set_log_state(newspa, SPA_LOG_CLEAR); /* release the spa config lock, retaining the namespace lock */ spa_vdev_config_exit(spa, NULL, txg, 0, FTAG); if (zio_injection_enabled) zio_handle_panic_injection(spa, FTAG, 1); spa_activate(newspa, spa_mode_global); spa_async_suspend(newspa); /* * Temporarily stop the initializing and TRIM activity. We set the * state to ACTIVE so that we know to resume initializing or TRIM * once the split has completed. */ list_t vd_initialize_list; list_create(&vd_initialize_list, sizeof (vdev_t), offsetof(vdev_t, vdev_initialize_node)); list_t vd_trim_list; list_create(&vd_trim_list, sizeof (vdev_t), offsetof(vdev_t, vdev_trim_node)); for (c = 0; c < children; c++) { if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) { mutex_enter(&vml[c]->vdev_initialize_lock); vdev_initialize_stop(vml[c], VDEV_INITIALIZE_ACTIVE, &vd_initialize_list); mutex_exit(&vml[c]->vdev_initialize_lock); mutex_enter(&vml[c]->vdev_trim_lock); vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list); mutex_exit(&vml[c]->vdev_trim_lock); } } vdev_initialize_stop_wait(spa, &vd_initialize_list); vdev_trim_stop_wait(spa, &vd_trim_list); list_destroy(&vd_initialize_list); list_destroy(&vd_trim_list); newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT; newspa->spa_is_splitting = B_TRUE; /* create the new pool from the disks of the original pool */ error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE); if (error) goto out; /* if that worked, generate a real config for the new pool */ if (newspa->spa_root_vdev != NULL) { VERIFY(nvlist_alloc(&newspa->spa_config_splitting, NV_UNIQUE_NAME, KM_SLEEP) == 0); VERIFY(nvlist_add_uint64(newspa->spa_config_splitting, ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa)) == 0); spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL, B_TRUE)); } /* set the props */ if (props != NULL) { spa_configfile_set(newspa, props, B_FALSE); error = spa_prop_set(newspa, props); if (error) goto out; } /* flush everything */ txg = spa_vdev_config_enter(newspa); vdev_config_dirty(newspa->spa_root_vdev); (void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG); if (zio_injection_enabled) zio_handle_panic_injection(spa, FTAG, 2); spa_async_resume(newspa); /* finally, update the original pool's config */ txg = spa_vdev_config_enter(spa); tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); error = dmu_tx_assign(tx, TXG_WAIT); if (error != 0) dmu_tx_abort(tx); for (c = 0; c < children; c++) { if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) { vdev_t *tvd = vml[c]->vdev_top; /* * Need to be sure the detachable VDEV is not * on any *other* txg's DTL list to prevent it * from being accessed after it's freed. */ for (int t = 0; t < TXG_SIZE; t++) { (void) txg_list_remove_this( &tvd->vdev_dtl_list, vml[c], t); } vdev_split(vml[c]); if (error == 0) spa_history_log_internal(spa, "detach", tx, "vdev=%s", vml[c]->vdev_path); vdev_free(vml[c]); } } spa->spa_avz_action = AVZ_ACTION_REBUILD; vdev_config_dirty(spa->spa_root_vdev); spa->spa_config_splitting = NULL; nvlist_free(nvl); if (error == 0) dmu_tx_commit(tx); (void) spa_vdev_exit(spa, NULL, txg, 0); if (zio_injection_enabled) zio_handle_panic_injection(spa, FTAG, 3); /* split is complete; log a history record */ spa_history_log_internal(newspa, "split", NULL, "from pool %s", spa_name(spa)); newspa->spa_is_splitting = B_FALSE; kmem_free(vml, children * sizeof (vdev_t *)); /* if we're not going to mount the filesystems in userland, export */ if (exp) error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL, B_FALSE, B_FALSE); return (error); out: spa_unload(newspa); spa_deactivate(newspa); spa_remove(newspa); txg = spa_vdev_config_enter(spa); /* re-online all offlined disks */ for (c = 0; c < children; c++) { if (vml[c] != NULL) vml[c]->vdev_offline = B_FALSE; } /* restart initializing or trimming disks as necessary */ spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART); spa_async_request(spa, SPA_ASYNC_TRIM_RESTART); spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART); vdev_reopen(spa->spa_root_vdev); nvlist_free(spa->spa_config_splitting); spa->spa_config_splitting = NULL; (void) spa_vdev_exit(spa, NULL, txg, error); kmem_free(vml, children * sizeof (vdev_t *)); return (error); } /* * Find any device that's done replacing, or a vdev marked 'unspare' that's * currently spared, so we can detach it. */ static vdev_t * spa_vdev_resilver_done_hunt(vdev_t *vd) { vdev_t *newvd, *oldvd; for (int c = 0; c < vd->vdev_children; c++) { oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]); if (oldvd != NULL) return (oldvd); } /* * Check for a completed replacement. We always consider the first * vdev in the list to be the oldest vdev, and the last one to be * the newest (see spa_vdev_attach() for how that works). In * the case where the newest vdev is faulted, we will not automatically * remove it after a resilver completes. This is OK as it will require * user intervention to determine which disk the admin wishes to keep. */ if (vd->vdev_ops == &vdev_replacing_ops) { ASSERT(vd->vdev_children > 1); newvd = vd->vdev_child[vd->vdev_children - 1]; oldvd = vd->vdev_child[0]; if (vdev_dtl_empty(newvd, DTL_MISSING) && vdev_dtl_empty(newvd, DTL_OUTAGE) && !vdev_dtl_required(oldvd)) return (oldvd); } /* * Check for a completed resilver with the 'unspare' flag set. * Also potentially update faulted state. */ if (vd->vdev_ops == &vdev_spare_ops) { vdev_t *first = vd->vdev_child[0]; vdev_t *last = vd->vdev_child[vd->vdev_children - 1]; if (last->vdev_unspare) { oldvd = first; newvd = last; } else if (first->vdev_unspare) { oldvd = last; newvd = first; } else { oldvd = NULL; } if (oldvd != NULL && vdev_dtl_empty(newvd, DTL_MISSING) && vdev_dtl_empty(newvd, DTL_OUTAGE) && !vdev_dtl_required(oldvd)) return (oldvd); vdev_propagate_state(vd); /* * If there are more than two spares attached to a disk, * and those spares are not required, then we want to * attempt to free them up now so that they can be used * by other pools. Once we're back down to a single * disk+spare, we stop removing them. */ if (vd->vdev_children > 2) { newvd = vd->vdev_child[1]; if (newvd->vdev_isspare && last->vdev_isspare && vdev_dtl_empty(last, DTL_MISSING) && vdev_dtl_empty(last, DTL_OUTAGE) && !vdev_dtl_required(newvd)) return (newvd); } } return (NULL); } static void spa_vdev_resilver_done(spa_t *spa) { vdev_t *vd, *pvd, *ppvd; uint64_t guid, sguid, pguid, ppguid; spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) { pvd = vd->vdev_parent; ppvd = pvd->vdev_parent; guid = vd->vdev_guid; pguid = pvd->vdev_guid; ppguid = ppvd->vdev_guid; sguid = 0; /* * If we have just finished replacing a hot spared device, then * we need to detach the parent's first child (the original hot * spare) as well. */ if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 && ppvd->vdev_children == 2) { ASSERT(pvd->vdev_ops == &vdev_replacing_ops); sguid = ppvd->vdev_child[1]->vdev_guid; } ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd)); spa_config_exit(spa, SCL_ALL, FTAG); if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0) return; if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0) return; spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); } spa_config_exit(spa, SCL_ALL, FTAG); /* * If a detach was not performed above replace waiters will not have * been notified. In which case we must do so now. */ spa_notify_waiters(spa); } /* * Update the stored path or FRU for this vdev. */ static int spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value, boolean_t ispath) { vdev_t *vd; boolean_t sync = B_FALSE; ASSERT(spa_writeable(spa)); spa_vdev_state_enter(spa, SCL_ALL); if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) return (spa_vdev_state_exit(spa, NULL, ENOENT)); if (!vd->vdev_ops->vdev_op_leaf) return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); if (ispath) { if (strcmp(value, vd->vdev_path) != 0) { spa_strfree(vd->vdev_path); vd->vdev_path = spa_strdup(value); sync = B_TRUE; } } else { if (vd->vdev_fru == NULL) { vd->vdev_fru = spa_strdup(value); sync = B_TRUE; } else if (strcmp(value, vd->vdev_fru) != 0) { spa_strfree(vd->vdev_fru); vd->vdev_fru = spa_strdup(value); sync = B_TRUE; } } return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0)); } int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath) { return (spa_vdev_set_common(spa, guid, newpath, B_TRUE)); } int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru) { return (spa_vdev_set_common(spa, guid, newfru, B_FALSE)); } /* * ========================================================================== * SPA Scanning * ========================================================================== */ int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd) { ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); if (dsl_scan_resilvering(spa->spa_dsl_pool)) return (SET_ERROR(EBUSY)); return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd)); } int spa_scan_stop(spa_t *spa) { ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); if (dsl_scan_resilvering(spa->spa_dsl_pool)) return (SET_ERROR(EBUSY)); return (dsl_scan_cancel(spa->spa_dsl_pool)); } int spa_scan(spa_t *spa, pool_scan_func_t func) { ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE) return (SET_ERROR(ENOTSUP)); if (func == POOL_SCAN_RESILVER && !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) return (SET_ERROR(ENOTSUP)); /* * If a resilver was requested, but there is no DTL on a * writeable leaf device, we have nothing to do. */ if (func == POOL_SCAN_RESILVER && !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) { spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); return (0); } return (dsl_scan(spa->spa_dsl_pool, func)); } /* * ========================================================================== * SPA async task processing * ========================================================================== */ static void spa_async_remove(spa_t *spa, vdev_t *vd) { if (vd->vdev_remove_wanted) { vd->vdev_remove_wanted = B_FALSE; vd->vdev_delayed_close = B_FALSE; vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE); /* * We want to clear the stats, but we don't want to do a full * vdev_clear() as that will cause us to throw away * degraded/faulted state as well as attempt to reopen the * device, all of which is a waste. */ vd->vdev_stat.vs_read_errors = 0; vd->vdev_stat.vs_write_errors = 0; vd->vdev_stat.vs_checksum_errors = 0; vdev_state_dirty(vd->vdev_top); /* Tell userspace that the vdev is gone. */ zfs_post_remove(spa, vd); } for (int c = 0; c < vd->vdev_children; c++) spa_async_remove(spa, vd->vdev_child[c]); } static void spa_async_probe(spa_t *spa, vdev_t *vd) { if (vd->vdev_probe_wanted) { vd->vdev_probe_wanted = B_FALSE; vdev_reopen(vd); /* vdev_open() does the actual probe */ } for (int c = 0; c < vd->vdev_children; c++) spa_async_probe(spa, vd->vdev_child[c]); } static void spa_async_autoexpand(spa_t *spa, vdev_t *vd) { if (!spa->spa_autoexpand) return; for (int c = 0; c < vd->vdev_children; c++) { vdev_t *cvd = vd->vdev_child[c]; spa_async_autoexpand(spa, cvd); } if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL) return; spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_AUTOEXPAND); } static void spa_async_thread(void *arg) { spa_t *spa = (spa_t *)arg; dsl_pool_t *dp = spa->spa_dsl_pool; int tasks; ASSERT(spa->spa_sync_on); mutex_enter(&spa->spa_async_lock); tasks = spa->spa_async_tasks; spa->spa_async_tasks = 0; mutex_exit(&spa->spa_async_lock); /* * See if the config needs to be updated. */ if (tasks & SPA_ASYNC_CONFIG_UPDATE) { uint64_t old_space, new_space; mutex_enter(&spa_namespace_lock); old_space = metaslab_class_get_space(spa_normal_class(spa)); old_space += metaslab_class_get_space(spa_special_class(spa)); old_space += metaslab_class_get_space(spa_dedup_class(spa)); old_space += metaslab_class_get_space( spa_embedded_log_class(spa)); spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); new_space = metaslab_class_get_space(spa_normal_class(spa)); new_space += metaslab_class_get_space(spa_special_class(spa)); new_space += metaslab_class_get_space(spa_dedup_class(spa)); new_space += metaslab_class_get_space( spa_embedded_log_class(spa)); mutex_exit(&spa_namespace_lock); /* * If the pool grew as a result of the config update, * then log an internal history event. */ if (new_space != old_space) { spa_history_log_internal(spa, "vdev online", NULL, "pool '%s' size: %llu(+%llu)", spa_name(spa), (u_longlong_t)new_space, (u_longlong_t)(new_space - old_space)); } } /* * See if any devices need to be marked REMOVED. */ if (tasks & SPA_ASYNC_REMOVE) { spa_vdev_state_enter(spa, SCL_NONE); spa_async_remove(spa, spa->spa_root_vdev); for (int i = 0; i < spa->spa_l2cache.sav_count; i++) spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]); for (int i = 0; i < spa->spa_spares.sav_count; i++) spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]); (void) spa_vdev_state_exit(spa, NULL, 0); } if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) { spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); spa_async_autoexpand(spa, spa->spa_root_vdev); spa_config_exit(spa, SCL_CONFIG, FTAG); } /* * See if any devices need to be probed. */ if (tasks & SPA_ASYNC_PROBE) { spa_vdev_state_enter(spa, SCL_NONE); spa_async_probe(spa, spa->spa_root_vdev); (void) spa_vdev_state_exit(spa, NULL, 0); } /* * If any devices are done replacing, detach them. */ if (tasks & SPA_ASYNC_RESILVER_DONE || tasks & SPA_ASYNC_REBUILD_DONE) { spa_vdev_resilver_done(spa); } /* * Kick off a resilver. */ if (tasks & SPA_ASYNC_RESILVER && !vdev_rebuild_active(spa->spa_root_vdev) && (!dsl_scan_resilvering(dp) || !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))) dsl_scan_restart_resilver(dp, 0); if (tasks & SPA_ASYNC_INITIALIZE_RESTART) { mutex_enter(&spa_namespace_lock); spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); vdev_initialize_restart(spa->spa_root_vdev); spa_config_exit(spa, SCL_CONFIG, FTAG); mutex_exit(&spa_namespace_lock); } if (tasks & SPA_ASYNC_TRIM_RESTART) { mutex_enter(&spa_namespace_lock); spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); vdev_trim_restart(spa->spa_root_vdev); spa_config_exit(spa, SCL_CONFIG, FTAG); mutex_exit(&spa_namespace_lock); } if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) { mutex_enter(&spa_namespace_lock); spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); vdev_autotrim_restart(spa); spa_config_exit(spa, SCL_CONFIG, FTAG); mutex_exit(&spa_namespace_lock); } /* * Kick off L2 cache whole device TRIM. */ if (tasks & SPA_ASYNC_L2CACHE_TRIM) { mutex_enter(&spa_namespace_lock); spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); vdev_trim_l2arc(spa); spa_config_exit(spa, SCL_CONFIG, FTAG); mutex_exit(&spa_namespace_lock); } /* * Kick off L2 cache rebuilding. */ if (tasks & SPA_ASYNC_L2CACHE_REBUILD) { mutex_enter(&spa_namespace_lock); spa_config_enter(spa, SCL_L2ARC, FTAG, RW_READER); l2arc_spa_rebuild_start(spa); spa_config_exit(spa, SCL_L2ARC, FTAG); mutex_exit(&spa_namespace_lock); } /* * Let the world know that we're done. */ mutex_enter(&spa->spa_async_lock); spa->spa_async_thread = NULL; cv_broadcast(&spa->spa_async_cv); mutex_exit(&spa->spa_async_lock); thread_exit(); } void spa_async_suspend(spa_t *spa) { mutex_enter(&spa->spa_async_lock); spa->spa_async_suspended++; while (spa->spa_async_thread != NULL) cv_wait(&spa->spa_async_cv, &spa->spa_async_lock); mutex_exit(&spa->spa_async_lock); spa_vdev_remove_suspend(spa); zthr_t *condense_thread = spa->spa_condense_zthr; if (condense_thread != NULL) zthr_cancel(condense_thread); zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr; if (discard_thread != NULL) zthr_cancel(discard_thread); zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr; if (ll_delete_thread != NULL) zthr_cancel(ll_delete_thread); zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr; if (ll_condense_thread != NULL) zthr_cancel(ll_condense_thread); } void spa_async_resume(spa_t *spa) { mutex_enter(&spa->spa_async_lock); ASSERT(spa->spa_async_suspended != 0); spa->spa_async_suspended--; mutex_exit(&spa->spa_async_lock); spa_restart_removal(spa); zthr_t *condense_thread = spa->spa_condense_zthr; if (condense_thread != NULL) zthr_resume(condense_thread); zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr; if (discard_thread != NULL) zthr_resume(discard_thread); zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr; if (ll_delete_thread != NULL) zthr_resume(ll_delete_thread); zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr; if (ll_condense_thread != NULL) zthr_resume(ll_condense_thread); } static boolean_t spa_async_tasks_pending(spa_t *spa) { uint_t non_config_tasks; uint_t config_task; boolean_t config_task_suspended; non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE; config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE; if (spa->spa_ccw_fail_time == 0) { config_task_suspended = B_FALSE; } else { config_task_suspended = (gethrtime() - spa->spa_ccw_fail_time) < ((hrtime_t)zfs_ccw_retry_interval * NANOSEC); } return (non_config_tasks || (config_task && !config_task_suspended)); } static void spa_async_dispatch(spa_t *spa) { mutex_enter(&spa->spa_async_lock); if (spa_async_tasks_pending(spa) && !spa->spa_async_suspended && spa->spa_async_thread == NULL) spa->spa_async_thread = thread_create(NULL, 0, spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri); mutex_exit(&spa->spa_async_lock); } void spa_async_request(spa_t *spa, int task) { zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task); mutex_enter(&spa->spa_async_lock); spa->spa_async_tasks |= task; mutex_exit(&spa->spa_async_lock); } int spa_async_tasks(spa_t *spa) { return (spa->spa_async_tasks); } /* * ========================================================================== * SPA syncing routines * ========================================================================== */ static int bpobj_enqueue_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, dmu_tx_t *tx) { bpobj_t *bpo = arg; bpobj_enqueue(bpo, bp, bp_freed, tx); return (0); } int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) { return (bpobj_enqueue_cb(arg, bp, B_FALSE, tx)); } int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) { return (bpobj_enqueue_cb(arg, bp, B_TRUE, tx)); } static int spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) { zio_t *pio = arg; zio_nowait(zio_free_sync(pio, pio->io_spa, dmu_tx_get_txg(tx), bp, pio->io_flags)); return (0); } static int bpobj_spa_free_sync_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, dmu_tx_t *tx) { ASSERT(!bp_freed); return (spa_free_sync_cb(arg, bp, tx)); } /* * Note: this simple function is not inlined to make it easier to dtrace the * amount of time spent syncing frees. */ static void spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx) { zio_t *zio = zio_root(spa, NULL, NULL, 0); bplist_iterate(bpl, spa_free_sync_cb, zio, tx); VERIFY(zio_wait(zio) == 0); } /* * Note: this simple function is not inlined to make it easier to dtrace the * amount of time spent syncing deferred frees. */ static void spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx) { if (spa_sync_pass(spa) != 1) return; /* * Note: * If the log space map feature is active, we stop deferring * frees to the next TXG and therefore running this function * would be considered a no-op as spa_deferred_bpobj should * not have any entries. * * That said we run this function anyway (instead of returning * immediately) for the edge-case scenario where we just * activated the log space map feature in this TXG but we have * deferred frees from the previous TXG. */ zio_t *zio = zio_root(spa, NULL, NULL, 0); VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj, bpobj_spa_free_sync_cb, zio, tx), ==, 0); VERIFY0(zio_wait(zio)); } static void spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx) { char *packed = NULL; size_t bufsize; size_t nvsize = 0; dmu_buf_t *db; VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0); /* * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration * information. This avoids the dmu_buf_will_dirty() path and * saves us a pre-read to get data we don't actually care about. */ bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE); packed = vmem_alloc(bufsize, KM_SLEEP); VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR, KM_SLEEP) == 0); bzero(packed + nvsize, bufsize - nvsize); dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx); vmem_free(packed, bufsize); VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db)); dmu_buf_will_dirty(db, tx); *(uint64_t *)db->db_data = nvsize; dmu_buf_rele(db, FTAG); } static void spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx, const char *config, const char *entry) { nvlist_t *nvroot; nvlist_t **list; int i; if (!sav->sav_sync) return; /* * Update the MOS nvlist describing the list of available devices. * spa_validate_aux() will have already made sure this nvlist is * valid and the vdevs are labeled appropriately. */ if (sav->sav_object == 0) { sav->sav_object = dmu_object_alloc(spa->spa_meta_objset, DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx); VERIFY(zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1, &sav->sav_object, tx) == 0); } VERIFY(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, KM_SLEEP) == 0); if (sav->sav_count == 0) { VERIFY(nvlist_add_nvlist_array(nvroot, config, NULL, 0) == 0); } else { list = kmem_alloc(sav->sav_count*sizeof (void *), KM_SLEEP); for (i = 0; i < sav->sav_count; i++) list[i] = vdev_config_generate(spa, sav->sav_vdevs[i], B_FALSE, VDEV_CONFIG_L2CACHE); VERIFY(nvlist_add_nvlist_array(nvroot, config, list, sav->sav_count) == 0); for (i = 0; i < sav->sav_count; i++) nvlist_free(list[i]); kmem_free(list, sav->sav_count * sizeof (void *)); } spa_sync_nvlist(spa, sav->sav_object, nvroot, tx); nvlist_free(nvroot); sav->sav_sync = B_FALSE; } /* * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t. * The all-vdev ZAP must be empty. */ static void spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx) { spa_t *spa = vd->vdev_spa; if (vd->vdev_top_zap != 0) { VERIFY0(zap_add_int(spa->spa_meta_objset, avz, vd->vdev_top_zap, tx)); } if (vd->vdev_leaf_zap != 0) { VERIFY0(zap_add_int(spa->spa_meta_objset, avz, vd->vdev_leaf_zap, tx)); } for (uint64_t i = 0; i < vd->vdev_children; i++) { spa_avz_build(vd->vdev_child[i], avz, tx); } } static void spa_sync_config_object(spa_t *spa, dmu_tx_t *tx) { nvlist_t *config; /* * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS, * its config may not be dirty but we still need to build per-vdev ZAPs. * Similarly, if the pool is being assembled (e.g. after a split), we * need to rebuild the AVZ although the config may not be dirty. */ if (list_is_empty(&spa->spa_config_dirty_list) && spa->spa_avz_action == AVZ_ACTION_NONE) return; spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE || spa->spa_avz_action == AVZ_ACTION_INITIALIZE || spa->spa_all_vdev_zaps != 0); if (spa->spa_avz_action == AVZ_ACTION_REBUILD) { /* Make and build the new AVZ */ uint64_t new_avz = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx); spa_avz_build(spa->spa_root_vdev, new_avz, tx); /* Diff old AVZ with new one */ zap_cursor_t zc; zap_attribute_t za; for (zap_cursor_init(&zc, spa->spa_meta_objset, spa->spa_all_vdev_zaps); zap_cursor_retrieve(&zc, &za) == 0; zap_cursor_advance(&zc)) { uint64_t vdzap = za.za_first_integer; if (zap_lookup_int(spa->spa_meta_objset, new_avz, vdzap) == ENOENT) { /* * ZAP is listed in old AVZ but not in new one; * destroy it */ VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap, tx)); } } zap_cursor_fini(&zc); /* Destroy the old AVZ */ VERIFY0(zap_destroy(spa->spa_meta_objset, spa->spa_all_vdev_zaps, tx)); /* Replace the old AVZ in the dir obj with the new one */ VERIFY0(zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, sizeof (new_avz), 1, &new_avz, tx)); spa->spa_all_vdev_zaps = new_avz; } else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) { zap_cursor_t zc; zap_attribute_t za; /* Walk through the AVZ and destroy all listed ZAPs */ for (zap_cursor_init(&zc, spa->spa_meta_objset, spa->spa_all_vdev_zaps); zap_cursor_retrieve(&zc, &za) == 0; zap_cursor_advance(&zc)) { uint64_t zap = za.za_first_integer; VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx)); } zap_cursor_fini(&zc); /* Destroy and unlink the AVZ itself */ VERIFY0(zap_destroy(spa->spa_meta_objset, spa->spa_all_vdev_zaps, tx)); VERIFY0(zap_remove(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx)); spa->spa_all_vdev_zaps = 0; } if (spa->spa_all_vdev_zaps == 0) { spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx); } spa->spa_avz_action = AVZ_ACTION_NONE; /* Create ZAPs for vdevs that don't have them. */ vdev_construct_zaps(spa->spa_root_vdev, tx); config = spa_config_generate(spa, spa->spa_root_vdev, dmu_tx_get_txg(tx), B_FALSE); /* * If we're upgrading the spa version then make sure that * the config object gets updated with the correct version. */ if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version) fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, spa->spa_uberblock.ub_version); spa_config_exit(spa, SCL_STATE, FTAG); nvlist_free(spa->spa_config_syncing); spa->spa_config_syncing = config; spa_sync_nvlist(spa, spa->spa_config_object, config, tx); } static void spa_sync_version(void *arg, dmu_tx_t *tx) { uint64_t *versionp = arg; uint64_t version = *versionp; spa_t *spa = dmu_tx_pool(tx)->dp_spa; /* * Setting the version is special cased when first creating the pool. */ ASSERT(tx->tx_txg != TXG_INITIAL); ASSERT(SPA_VERSION_IS_SUPPORTED(version)); ASSERT(version >= spa_version(spa)); spa->spa_uberblock.ub_version = version; vdev_config_dirty(spa->spa_root_vdev); spa_history_log_internal(spa, "set", tx, "version=%lld", (longlong_t)version); } /* * Set zpool properties. */ static void spa_sync_props(void *arg, dmu_tx_t *tx) { nvlist_t *nvp = arg; spa_t *spa = dmu_tx_pool(tx)->dp_spa; objset_t *mos = spa->spa_meta_objset; nvpair_t *elem = NULL; mutex_enter(&spa->spa_props_lock); while ((elem = nvlist_next_nvpair(nvp, elem))) { uint64_t intval; char *strval, *fname; zpool_prop_t prop; const char *propname; zprop_type_t proptype; spa_feature_t fid; switch (prop = zpool_name_to_prop(nvpair_name(elem))) { case ZPOOL_PROP_INVAL: /* * We checked this earlier in spa_prop_validate(). */ ASSERT(zpool_prop_feature(nvpair_name(elem))); fname = strchr(nvpair_name(elem), '@') + 1; VERIFY0(zfeature_lookup_name(fname, &fid)); spa_feature_enable(spa, fid, tx); spa_history_log_internal(spa, "set", tx, "%s=enabled", nvpair_name(elem)); break; case ZPOOL_PROP_VERSION: intval = fnvpair_value_uint64(elem); /* * The version is synced separately before other * properties and should be correct by now. */ ASSERT3U(spa_version(spa), >=, intval); break; case ZPOOL_PROP_ALTROOT: /* * 'altroot' is a non-persistent property. It should * have been set temporarily at creation or import time. */ ASSERT(spa->spa_root != NULL); break; case ZPOOL_PROP_READONLY: case ZPOOL_PROP_CACHEFILE: /* * 'readonly' and 'cachefile' are also non-persistent * properties. */ break; case ZPOOL_PROP_COMMENT: strval = fnvpair_value_string(elem); if (spa->spa_comment != NULL) spa_strfree(spa->spa_comment); spa->spa_comment = spa_strdup(strval); /* * We need to dirty the configuration on all the vdevs * so that their labels get updated. We also need to * update the cache file to keep it in sync with the * MOS version. It's unnecessary to do this for pool * creation since the vdev's configuration has already * been dirtied. */ if (tx->tx_txg != TXG_INITIAL) { vdev_config_dirty(spa->spa_root_vdev); spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); } spa_history_log_internal(spa, "set", tx, "%s=%s", nvpair_name(elem), strval); break; case ZPOOL_PROP_COMPATIBILITY: strval = fnvpair_value_string(elem); if (spa->spa_compatibility != NULL) spa_strfree(spa->spa_compatibility); spa->spa_compatibility = spa_strdup(strval); /* * Dirty the configuration on vdevs as above. */ if (tx->tx_txg != TXG_INITIAL) { vdev_config_dirty(spa->spa_root_vdev); spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); } spa_history_log_internal(spa, "set", tx, "%s=%s", nvpair_name(elem), strval); break; default: /* * Set pool property values in the poolprops mos object. */ if (spa->spa_pool_props_object == 0) { spa->spa_pool_props_object = zap_create_link(mos, DMU_OT_POOL_PROPS, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS, tx); } /* normalize the property name */ propname = zpool_prop_to_name(prop); proptype = zpool_prop_get_type(prop); if (nvpair_type(elem) == DATA_TYPE_STRING) { ASSERT(proptype == PROP_TYPE_STRING); strval = fnvpair_value_string(elem); VERIFY0(zap_update(mos, spa->spa_pool_props_object, propname, 1, strlen(strval) + 1, strval, tx)); spa_history_log_internal(spa, "set", tx, "%s=%s", nvpair_name(elem), strval); } else if (nvpair_type(elem) == DATA_TYPE_UINT64) { intval = fnvpair_value_uint64(elem); if (proptype == PROP_TYPE_INDEX) { const char *unused; VERIFY0(zpool_prop_index_to_string( prop, intval, &unused)); } VERIFY0(zap_update(mos, spa->spa_pool_props_object, propname, 8, 1, &intval, tx)); spa_history_log_internal(spa, "set", tx, "%s=%lld", nvpair_name(elem), (longlong_t)intval); } else { ASSERT(0); /* not allowed */ } switch (prop) { case ZPOOL_PROP_DELEGATION: spa->spa_delegation = intval; break; case ZPOOL_PROP_BOOTFS: spa->spa_bootfs = intval; break; case ZPOOL_PROP_FAILUREMODE: spa->spa_failmode = intval; break; case ZPOOL_PROP_AUTOTRIM: spa->spa_autotrim = intval; spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART); break; case ZPOOL_PROP_AUTOEXPAND: spa->spa_autoexpand = intval; if (tx->tx_txg != TXG_INITIAL) spa_async_request(spa, SPA_ASYNC_AUTOEXPAND); break; case ZPOOL_PROP_MULTIHOST: spa->spa_multihost = intval; break; default: break; } } } mutex_exit(&spa->spa_props_lock); } /* * Perform one-time upgrade on-disk changes. spa_version() does not * reflect the new version this txg, so there must be no changes this * txg to anything that the upgrade code depends on after it executes. * Therefore this must be called after dsl_pool_sync() does the sync * tasks. */ static void spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx) { if (spa_sync_pass(spa) != 1) return; dsl_pool_t *dp = spa->spa_dsl_pool; rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN && spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) { dsl_pool_create_origin(dp, tx); /* Keeping the origin open increases spa_minref */ spa->spa_minref += 3; } if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES && spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) { dsl_pool_upgrade_clones(dp, tx); } if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES && spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) { dsl_pool_upgrade_dir_clones(dp, tx); /* Keeping the freedir open increases spa_minref */ spa->spa_minref += 3; } if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES && spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) { spa_feature_create_zap_objects(spa, tx); } /* * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable * when possibility to use lz4 compression for metadata was added * Old pools that have this feature enabled must be upgraded to have * this feature active */ if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) { boolean_t lz4_en = spa_feature_is_enabled(spa, SPA_FEATURE_LZ4_COMPRESS); boolean_t lz4_ac = spa_feature_is_active(spa, SPA_FEATURE_LZ4_COMPRESS); if (lz4_en && !lz4_ac) spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx); } /* * If we haven't written the salt, do so now. Note that the * feature may not be activated yet, but that's fine since * the presence of this ZAP entry is backwards compatible. */ if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT) == ENOENT) { VERIFY0(zap_add(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1, sizeof (spa->spa_cksum_salt.zcs_bytes), spa->spa_cksum_salt.zcs_bytes, tx)); } rrw_exit(&dp->dp_config_rwlock, FTAG); } static void vdev_indirect_state_sync_verify(vdev_t *vd) { vdev_indirect_mapping_t *vim __maybe_unused = vd->vdev_indirect_mapping; vdev_indirect_births_t *vib __maybe_unused = vd->vdev_indirect_births; if (vd->vdev_ops == &vdev_indirect_ops) { ASSERT(vim != NULL); ASSERT(vib != NULL); } uint64_t obsolete_sm_object = 0; ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object)); if (obsolete_sm_object != 0) { ASSERT(vd->vdev_obsolete_sm != NULL); ASSERT(vd->vdev_removing || vd->vdev_ops == &vdev_indirect_ops); ASSERT(vdev_indirect_mapping_num_entries(vim) > 0); ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0); ASSERT3U(obsolete_sm_object, ==, space_map_object(vd->vdev_obsolete_sm)); ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=, space_map_allocated(vd->vdev_obsolete_sm)); } ASSERT(vd->vdev_obsolete_segments != NULL); /* * Since frees / remaps to an indirect vdev can only * happen in syncing context, the obsolete segments * tree must be empty when we start syncing. */ ASSERT0(range_tree_space(vd->vdev_obsolete_segments)); } /* * Set the top-level vdev's max queue depth. Evaluate each top-level's * async write queue depth in case it changed. The max queue depth will * not change in the middle of syncing out this txg. */ static void spa_sync_adjust_vdev_max_queue_depth(spa_t *spa) { ASSERT(spa_writeable(spa)); vdev_t *rvd = spa->spa_root_vdev; uint32_t max_queue_depth = zfs_vdev_async_write_max_active * zfs_vdev_queue_depth_pct / 100; metaslab_class_t *normal = spa_normal_class(spa); metaslab_class_t *special = spa_special_class(spa); metaslab_class_t *dedup = spa_dedup_class(spa); uint64_t slots_per_allocator = 0; for (int c = 0; c < rvd->vdev_children; c++) { vdev_t *tvd = rvd->vdev_child[c]; metaslab_group_t *mg = tvd->vdev_mg; if (mg == NULL || !metaslab_group_initialized(mg)) continue; metaslab_class_t *mc = mg->mg_class; if (mc != normal && mc != special && mc != dedup) continue; /* * It is safe to do a lock-free check here because only async * allocations look at mg_max_alloc_queue_depth, and async * allocations all happen from spa_sync(). */ for (int i = 0; i < mg->mg_allocators; i++) { ASSERT0(zfs_refcount_count( &(mg->mg_allocator[i].mga_alloc_queue_depth))); } mg->mg_max_alloc_queue_depth = max_queue_depth; for (int i = 0; i < mg->mg_allocators; i++) { mg->mg_allocator[i].mga_cur_max_alloc_queue_depth = zfs_vdev_def_queue_depth; } slots_per_allocator += zfs_vdev_def_queue_depth; } for (int i = 0; i < spa->spa_alloc_count; i++) { ASSERT0(zfs_refcount_count(&normal->mc_allocator[i]. mca_alloc_slots)); ASSERT0(zfs_refcount_count(&special->mc_allocator[i]. mca_alloc_slots)); ASSERT0(zfs_refcount_count(&dedup->mc_allocator[i]. mca_alloc_slots)); normal->mc_allocator[i].mca_alloc_max_slots = slots_per_allocator; special->mc_allocator[i].mca_alloc_max_slots = slots_per_allocator; dedup->mc_allocator[i].mca_alloc_max_slots = slots_per_allocator; } normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; } static void spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx) { ASSERT(spa_writeable(spa)); vdev_t *rvd = spa->spa_root_vdev; for (int c = 0; c < rvd->vdev_children; c++) { vdev_t *vd = rvd->vdev_child[c]; vdev_indirect_state_sync_verify(vd); if (vdev_indirect_should_condense(vd)) { spa_condense_indirect_start_sync(vd, tx); break; } } } static void spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx) { objset_t *mos = spa->spa_meta_objset; dsl_pool_t *dp = spa->spa_dsl_pool; uint64_t txg = tx->tx_txg; bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK]; do { int pass = ++spa->spa_sync_pass; spa_sync_config_object(spa, tx); spa_sync_aux_dev(spa, &spa->spa_spares, tx, ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES); spa_sync_aux_dev(spa, &spa->spa_l2cache, tx, ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE); spa_errlog_sync(spa, txg); dsl_pool_sync(dp, txg); if (pass < zfs_sync_pass_deferred_free || spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) { /* * If the log space map feature is active we don't * care about deferred frees and the deferred bpobj * as the log space map should effectively have the * same results (i.e. appending only to one object). */ spa_sync_frees(spa, free_bpl, tx); } else { /* * We can not defer frees in pass 1, because * we sync the deferred frees later in pass 1. */ ASSERT3U(pass, >, 1); bplist_iterate(free_bpl, bpobj_enqueue_alloc_cb, &spa->spa_deferred_bpobj, tx); } ddt_sync(spa, txg); dsl_scan_sync(dp, tx); svr_sync(spa, tx); spa_sync_upgrades(spa, tx); spa_flush_metaslabs(spa, tx); vdev_t *vd = NULL; while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg)) != NULL) vdev_sync(vd, txg); /* * Note: We need to check if the MOS is dirty because we could * have marked the MOS dirty without updating the uberblock * (e.g. if we have sync tasks but no dirty user data). We need * to check the uberblock's rootbp because it is updated if we * have synced out dirty data (though in this case the MOS will * most likely also be dirty due to second order effects, we * don't want to rely on that here). */ if (pass == 1 && spa->spa_uberblock.ub_rootbp.blk_birth < txg && !dmu_objset_is_dirty(mos, txg)) { /* * Nothing changed on the first pass, therefore this * TXG is a no-op. Avoid syncing deferred frees, so * that we can keep this TXG as a no-op. */ ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg)); ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg)); ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg)); ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg)); break; } spa_sync_deferred_frees(spa, tx); } while (dmu_objset_is_dirty(mos, txg)); } /* * Rewrite the vdev configuration (which includes the uberblock) to * commit the transaction group. * * If there are no dirty vdevs, we sync the uberblock to a few random * top-level vdevs that are known to be visible in the config cache * (see spa_vdev_add() for a complete description). If there *are* dirty * vdevs, sync the uberblock to all vdevs. */ static void spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx) { vdev_t *rvd = spa->spa_root_vdev; uint64_t txg = tx->tx_txg; for (;;) { int error = 0; /* * We hold SCL_STATE to prevent vdev open/close/etc. * while we're attempting to write the vdev labels. */ spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); if (list_is_empty(&spa->spa_config_dirty_list)) { vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL }; int svdcount = 0; int children = rvd->vdev_children; int c0 = random_in_range(children); for (int c = 0; c < children; c++) { vdev_t *vd = rvd->vdev_child[(c0 + c) % children]; /* Stop when revisiting the first vdev */ if (c > 0 && svd[0] == vd) break; if (vd->vdev_ms_array == 0 || vd->vdev_islog || !vdev_is_concrete(vd)) continue; svd[svdcount++] = vd; if (svdcount == SPA_SYNC_MIN_VDEVS) break; } error = vdev_config_sync(svd, svdcount, txg); } else { error = vdev_config_sync(rvd->vdev_child, rvd->vdev_children, txg); } if (error == 0) spa->spa_last_synced_guid = rvd->vdev_guid; spa_config_exit(spa, SCL_STATE, FTAG); if (error == 0) break; zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR); zio_resume_wait(spa); } } /* * Sync the specified transaction group. New blocks may be dirtied as * part of the process, so we iterate until it converges. */ void spa_sync(spa_t *spa, uint64_t txg) { vdev_t *vd = NULL; VERIFY(spa_writeable(spa)); /* * Wait for i/os issued in open context that need to complete * before this txg syncs. */ (void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]); spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); /* * Lock out configuration changes. */ spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); spa->spa_syncing_txg = txg; spa->spa_sync_pass = 0; for (int i = 0; i < spa->spa_alloc_count; i++) { mutex_enter(&spa->spa_allocs[i].spaa_lock); VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree)); mutex_exit(&spa->spa_allocs[i].spaa_lock); } /* * If there are any pending vdev state changes, convert them * into config changes that go out with this transaction group. */ spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); while (list_head(&spa->spa_state_dirty_list) != NULL) { /* * We need the write lock here because, for aux vdevs, * calling vdev_config_dirty() modifies sav_config. * This is ugly and will become unnecessary when we * eliminate the aux vdev wart by integrating all vdevs * into the root vdev tree. */ spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER); while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) { vdev_state_clean(vd); vdev_config_dirty(vd); } spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); } spa_config_exit(spa, SCL_STATE, FTAG); dsl_pool_t *dp = spa->spa_dsl_pool; dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); spa->spa_sync_starttime = gethrtime(); taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); spa->spa_deadman_tqid = taskq_dispatch_delay(system_delay_taskq, spa_deadman, spa, TQ_SLEEP, ddi_get_lbolt() + NSEC_TO_TICK(spa->spa_deadman_synctime)); /* * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg, * set spa_deflate if we have no raid-z vdevs. */ if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE && spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) { vdev_t *rvd = spa->spa_root_vdev; int i; for (i = 0; i < rvd->vdev_children; i++) { vd = rvd->vdev_child[i]; if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE) break; } if (i == rvd->vdev_children) { spa->spa_deflate = TRUE; VERIFY0(zap_add(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE, sizeof (uint64_t), 1, &spa->spa_deflate, tx)); } } spa_sync_adjust_vdev_max_queue_depth(spa); spa_sync_condense_indirect(spa, tx); spa_sync_iterate_to_convergence(spa, tx); #ifdef ZFS_DEBUG if (!list_is_empty(&spa->spa_config_dirty_list)) { /* * Make sure that the number of ZAPs for all the vdevs matches * the number of ZAPs in the per-vdev ZAP list. This only gets * called if the config is dirty; otherwise there may be * outstanding AVZ operations that weren't completed in * spa_sync_config_object. */ uint64_t all_vdev_zap_entry_count; ASSERT0(zap_count(spa->spa_meta_objset, spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count)); ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==, all_vdev_zap_entry_count); } #endif if (spa->spa_vdev_removal != NULL) { ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]); } spa_sync_rewrite_vdev_config(spa, tx); dmu_tx_commit(tx); taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); spa->spa_deadman_tqid = 0; /* * Clear the dirty config list. */ while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL) vdev_config_clean(vd); /* * Now that the new config has synced transactionally, * let it become visible to the config cache. */ if (spa->spa_config_syncing != NULL) { spa_config_set(spa, spa->spa_config_syncing); spa->spa_config_txg = txg; spa->spa_config_syncing = NULL; } dsl_pool_sync_done(dp, txg); for (int i = 0; i < spa->spa_alloc_count; i++) { mutex_enter(&spa->spa_allocs[i].spaa_lock); VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree)); mutex_exit(&spa->spa_allocs[i].spaa_lock); } /* * Update usable space statistics. */ while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg))) != NULL) vdev_sync_done(vd, txg); metaslab_class_evict_old(spa->spa_normal_class, txg); metaslab_class_evict_old(spa->spa_log_class, txg); spa_sync_close_syncing_log_sm(spa); spa_update_dspace(spa); /* * It had better be the case that we didn't dirty anything * since vdev_config_sync(). */ ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg)); ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg)); ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg)); while (zfs_pause_spa_sync) delay(1); spa->spa_sync_pass = 0; /* * Update the last synced uberblock here. We want to do this at * the end of spa_sync() so that consumers of spa_last_synced_txg() * will be guaranteed that all the processing associated with * that txg has been completed. */ spa->spa_ubsync = spa->spa_uberblock; spa_config_exit(spa, SCL_CONFIG, FTAG); spa_handle_ignored_writes(spa); /* * If any async tasks have been requested, kick them off. */ spa_async_dispatch(spa); } /* * Sync all pools. We don't want to hold the namespace lock across these * operations, so we take a reference on the spa_t and drop the lock during the * sync. */ void spa_sync_allpools(void) { spa_t *spa = NULL; mutex_enter(&spa_namespace_lock); while ((spa = spa_next(spa)) != NULL) { if (spa_state(spa) != POOL_STATE_ACTIVE || !spa_writeable(spa) || spa_suspended(spa)) continue; spa_open_ref(spa, FTAG); mutex_exit(&spa_namespace_lock); txg_wait_synced(spa_get_dsl(spa), 0); mutex_enter(&spa_namespace_lock); spa_close(spa, FTAG); } mutex_exit(&spa_namespace_lock); } /* * ========================================================================== * Miscellaneous routines * ========================================================================== */ /* * Remove all pools in the system. */ void spa_evict_all(void) { spa_t *spa; /* * Remove all cached state. All pools should be closed now, * so every spa in the AVL tree should be unreferenced. */ mutex_enter(&spa_namespace_lock); while ((spa = spa_next(NULL)) != NULL) { /* * Stop async tasks. The async thread may need to detach * a device that's been replaced, which requires grabbing * spa_namespace_lock, so we must drop it here. */ spa_open_ref(spa, FTAG); mutex_exit(&spa_namespace_lock); spa_async_suspend(spa); mutex_enter(&spa_namespace_lock); spa_close(spa, FTAG); if (spa->spa_state != POOL_STATE_UNINITIALIZED) { spa_unload(spa); spa_deactivate(spa); } spa_remove(spa); } mutex_exit(&spa_namespace_lock); } vdev_t * spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux) { vdev_t *vd; int i; if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL) return (vd); if (aux) { for (i = 0; i < spa->spa_l2cache.sav_count; i++) { vd = spa->spa_l2cache.sav_vdevs[i]; if (vd->vdev_guid == guid) return (vd); } for (i = 0; i < spa->spa_spares.sav_count; i++) { vd = spa->spa_spares.sav_vdevs[i]; if (vd->vdev_guid == guid) return (vd); } } return (NULL); } void spa_upgrade(spa_t *spa, uint64_t version) { ASSERT(spa_writeable(spa)); spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); /* * This should only be called for a non-faulted pool, and since a * future version would result in an unopenable pool, this shouldn't be * possible. */ ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version)); ASSERT3U(version, >=, spa->spa_uberblock.ub_version); spa->spa_uberblock.ub_version = version; vdev_config_dirty(spa->spa_root_vdev); spa_config_exit(spa, SCL_ALL, FTAG); txg_wait_synced(spa_get_dsl(spa), 0); } boolean_t spa_has_spare(spa_t *spa, uint64_t guid) { int i; uint64_t spareguid; spa_aux_vdev_t *sav = &spa->spa_spares; for (i = 0; i < sav->sav_count; i++) if (sav->sav_vdevs[i]->vdev_guid == guid) return (B_TRUE); for (i = 0; i < sav->sav_npending; i++) { if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID, &spareguid) == 0 && spareguid == guid) return (B_TRUE); } return (B_FALSE); } /* * Check if a pool has an active shared spare device. * Note: reference count of an active spare is 2, as a spare and as a replace */ static boolean_t spa_has_active_shared_spare(spa_t *spa) { int i, refcnt; uint64_t pool; spa_aux_vdev_t *sav = &spa->spa_spares; for (i = 0; i < sav->sav_count; i++) { if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool, &refcnt) && pool != 0ULL && pool == spa_guid(spa) && refcnt > 2) return (B_TRUE); } return (B_FALSE); } uint64_t spa_total_metaslabs(spa_t *spa) { vdev_t *rvd = spa->spa_root_vdev; uint64_t m = 0; for (uint64_t c = 0; c < rvd->vdev_children; c++) { vdev_t *vd = rvd->vdev_child[c]; if (!vdev_is_concrete(vd)) continue; m += vd->vdev_ms_count; } return (m); } /* * Notify any waiting threads that some activity has switched from being in- * progress to not-in-progress so that the thread can wake up and determine * whether it is finished waiting. */ void spa_notify_waiters(spa_t *spa) { /* * Acquiring spa_activities_lock here prevents the cv_broadcast from * happening between the waiting thread's check and cv_wait. */ mutex_enter(&spa->spa_activities_lock); cv_broadcast(&spa->spa_activities_cv); mutex_exit(&spa->spa_activities_lock); } /* * Notify any waiting threads that the pool is exporting, and then block until * they are finished using the spa_t. */ void spa_wake_waiters(spa_t *spa) { mutex_enter(&spa->spa_activities_lock); spa->spa_waiters_cancel = B_TRUE; cv_broadcast(&spa->spa_activities_cv); while (spa->spa_waiters != 0) cv_wait(&spa->spa_waiters_cv, &spa->spa_activities_lock); spa->spa_waiters_cancel = B_FALSE; mutex_exit(&spa->spa_activities_lock); } /* Whether the vdev or any of its descendants are being initialized/trimmed. */ static boolean_t spa_vdev_activity_in_progress_impl(vdev_t *vd, zpool_wait_activity_t activity) { spa_t *spa = vd->vdev_spa; ASSERT(spa_config_held(spa, SCL_CONFIG | SCL_STATE, RW_READER)); ASSERT(MUTEX_HELD(&spa->spa_activities_lock)); ASSERT(activity == ZPOOL_WAIT_INITIALIZE || activity == ZPOOL_WAIT_TRIM); kmutex_t *lock = activity == ZPOOL_WAIT_INITIALIZE ? &vd->vdev_initialize_lock : &vd->vdev_trim_lock; mutex_exit(&spa->spa_activities_lock); mutex_enter(lock); mutex_enter(&spa->spa_activities_lock); boolean_t in_progress = (activity == ZPOOL_WAIT_INITIALIZE) ? (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) : (vd->vdev_trim_state == VDEV_TRIM_ACTIVE); mutex_exit(lock); if (in_progress) return (B_TRUE); for (int i = 0; i < vd->vdev_children; i++) { if (spa_vdev_activity_in_progress_impl(vd->vdev_child[i], activity)) return (B_TRUE); } return (B_FALSE); } /* * If use_guid is true, this checks whether the vdev specified by guid is * being initialized/trimmed. Otherwise, it checks whether any vdev in the pool * is being initialized/trimmed. The caller must hold the config lock and * spa_activities_lock. */ static int spa_vdev_activity_in_progress(spa_t *spa, boolean_t use_guid, uint64_t guid, zpool_wait_activity_t activity, boolean_t *in_progress) { mutex_exit(&spa->spa_activities_lock); spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); mutex_enter(&spa->spa_activities_lock); vdev_t *vd; if (use_guid) { vd = spa_lookup_by_guid(spa, guid, B_FALSE); if (vd == NULL || !vd->vdev_ops->vdev_op_leaf) { spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (EINVAL); } } else { vd = spa->spa_root_vdev; } *in_progress = spa_vdev_activity_in_progress_impl(vd, activity); spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); return (0); } /* * Locking for waiting threads * --------------------------- * * Waiting threads need a way to check whether a given activity is in progress, * and then, if it is, wait for it to complete. Each activity will have some * in-memory representation of the relevant on-disk state which can be used to * determine whether or not the activity is in progress. The in-memory state and * the locking used to protect it will be different for each activity, and may * not be suitable for use with a cvar (e.g., some state is protected by the * config lock). To allow waiting threads to wait without any races, another * lock, spa_activities_lock, is used. * * When the state is checked, both the activity-specific lock (if there is one) * and spa_activities_lock are held. In some cases, the activity-specific lock * is acquired explicitly (e.g. the config lock). In others, the locking is * internal to some check (e.g. bpobj_is_empty). After checking, the waiting * thread releases the activity-specific lock and, if the activity is in * progress, then cv_waits using spa_activities_lock. * * The waiting thread is woken when another thread, one completing some * activity, updates the state of the activity and then calls * spa_notify_waiters, which will cv_broadcast. This 'completing' thread only * needs to hold its activity-specific lock when updating the state, and this * lock can (but doesn't have to) be dropped before calling spa_notify_waiters. * * Because spa_notify_waiters acquires spa_activities_lock before broadcasting, * and because it is held when the waiting thread checks the state of the * activity, it can never be the case that the completing thread both updates * the activity state and cv_broadcasts in between the waiting thread's check * and cv_wait. Thus, a waiting thread can never miss a wakeup. * * In order to prevent deadlock, when the waiting thread does its check, in some * cases it will temporarily drop spa_activities_lock in order to acquire the * activity-specific lock. The order in which spa_activities_lock and the * activity specific lock are acquired in the waiting thread is determined by * the order in which they are acquired in the completing thread; if the * completing thread calls spa_notify_waiters with the activity-specific lock * held, then the waiting thread must also acquire the activity-specific lock * first. */ static int spa_activity_in_progress(spa_t *spa, zpool_wait_activity_t activity, boolean_t use_tag, uint64_t tag, boolean_t *in_progress) { int error = 0; ASSERT(MUTEX_HELD(&spa->spa_activities_lock)); switch (activity) { case ZPOOL_WAIT_CKPT_DISCARD: *in_progress = (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT) && zap_contains(spa_meta_objset(spa), DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT) == ENOENT); break; case ZPOOL_WAIT_FREE: *in_progress = ((spa_version(spa) >= SPA_VERSION_DEADLISTS && !bpobj_is_empty(&spa->spa_dsl_pool->dp_free_bpobj)) || spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY) || spa_livelist_delete_check(spa)); break; case ZPOOL_WAIT_INITIALIZE: case ZPOOL_WAIT_TRIM: error = spa_vdev_activity_in_progress(spa, use_tag, tag, activity, in_progress); break; case ZPOOL_WAIT_REPLACE: mutex_exit(&spa->spa_activities_lock); spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); mutex_enter(&spa->spa_activities_lock); *in_progress = vdev_replace_in_progress(spa->spa_root_vdev); spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); break; case ZPOOL_WAIT_REMOVE: *in_progress = (spa->spa_removing_phys.sr_state == DSS_SCANNING); break; case ZPOOL_WAIT_RESILVER: if ((*in_progress = vdev_rebuild_active(spa->spa_root_vdev))) break; /* fall through */ case ZPOOL_WAIT_SCRUB: { boolean_t scanning, paused, is_scrub; dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan; is_scrub = (scn->scn_phys.scn_func == POOL_SCAN_SCRUB); scanning = (scn->scn_phys.scn_state == DSS_SCANNING); paused = dsl_scan_is_paused_scrub(scn); *in_progress = (scanning && !paused && is_scrub == (activity == ZPOOL_WAIT_SCRUB)); break; } default: panic("unrecognized value for activity %d", activity); } return (error); } static int spa_wait_common(const char *pool, zpool_wait_activity_t activity, boolean_t use_tag, uint64_t tag, boolean_t *waited) { /* * The tag is used to distinguish between instances of an activity. * 'initialize' and 'trim' are the only activities that we use this for. * The other activities can only have a single instance in progress in a * pool at one time, making the tag unnecessary. * * There can be multiple devices being replaced at once, but since they * all finish once resilvering finishes, we don't bother keeping track * of them individually, we just wait for them all to finish. */ if (use_tag && activity != ZPOOL_WAIT_INITIALIZE && activity != ZPOOL_WAIT_TRIM) return (EINVAL); if (activity < 0 || activity >= ZPOOL_WAIT_NUM_ACTIVITIES) return (EINVAL); spa_t *spa; int error = spa_open(pool, &spa, FTAG); if (error != 0) return (error); /* * Increment the spa's waiter count so that we can call spa_close and * still ensure that the spa_t doesn't get freed before this thread is * finished with it when the pool is exported. We want to call spa_close * before we start waiting because otherwise the additional ref would * prevent the pool from being exported or destroyed throughout the * potentially long wait. */ mutex_enter(&spa->spa_activities_lock); spa->spa_waiters++; spa_close(spa, FTAG); *waited = B_FALSE; for (;;) { boolean_t in_progress; error = spa_activity_in_progress(spa, activity, use_tag, tag, &in_progress); if (error || !in_progress || spa->spa_waiters_cancel) break; *waited = B_TRUE; if (cv_wait_sig(&spa->spa_activities_cv, &spa->spa_activities_lock) == 0) { error = EINTR; break; } } spa->spa_waiters--; cv_signal(&spa->spa_waiters_cv); mutex_exit(&spa->spa_activities_lock); return (error); } /* * Wait for a particular instance of the specified activity to complete, where * the instance is identified by 'tag' */ int spa_wait_tag(const char *pool, zpool_wait_activity_t activity, uint64_t tag, boolean_t *waited) { return (spa_wait_common(pool, activity, B_TRUE, tag, waited)); } /* * Wait for all instances of the specified activity complete */ int spa_wait(const char *pool, zpool_wait_activity_t activity, boolean_t *waited) { return (spa_wait_common(pool, activity, B_FALSE, 0, waited)); } sysevent_t * spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name) { sysevent_t *ev = NULL; #ifdef _KERNEL nvlist_t *resource; resource = zfs_event_create(spa, vd, FM_SYSEVENT_CLASS, name, hist_nvl); if (resource) { ev = kmem_alloc(sizeof (sysevent_t), KM_SLEEP); ev->resource = resource; } #endif return (ev); } void spa_event_post(sysevent_t *ev) { #ifdef _KERNEL if (ev) { zfs_zevent_post(ev->resource, NULL, zfs_zevent_post_cb); kmem_free(ev, sizeof (*ev)); } #endif } /* * Post a zevent corresponding to the given sysevent. The 'name' must be one * of the event definitions in sys/sysevent/eventdefs.h. The payload will be * filled in from the spa and (optionally) the vdev. This doesn't do anything * in the userland libzpool, as we don't want consumers to misinterpret ztest * or zdb as real changes. */ void spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name) { spa_event_post(spa_event_create(spa, vd, hist_nvl, name)); } /* state manipulation functions */ EXPORT_SYMBOL(spa_open); EXPORT_SYMBOL(spa_open_rewind); EXPORT_SYMBOL(spa_get_stats); EXPORT_SYMBOL(spa_create); EXPORT_SYMBOL(spa_import); EXPORT_SYMBOL(spa_tryimport); EXPORT_SYMBOL(spa_destroy); EXPORT_SYMBOL(spa_export); EXPORT_SYMBOL(spa_reset); EXPORT_SYMBOL(spa_async_request); EXPORT_SYMBOL(spa_async_suspend); EXPORT_SYMBOL(spa_async_resume); EXPORT_SYMBOL(spa_inject_addref); EXPORT_SYMBOL(spa_inject_delref); EXPORT_SYMBOL(spa_scan_stat_init); EXPORT_SYMBOL(spa_scan_get_stats); /* device manipulation */ EXPORT_SYMBOL(spa_vdev_add); EXPORT_SYMBOL(spa_vdev_attach); EXPORT_SYMBOL(spa_vdev_detach); EXPORT_SYMBOL(spa_vdev_setpath); EXPORT_SYMBOL(spa_vdev_setfru); EXPORT_SYMBOL(spa_vdev_split_mirror); /* spare statech is global across all pools) */ EXPORT_SYMBOL(spa_spare_add); EXPORT_SYMBOL(spa_spare_remove); EXPORT_SYMBOL(spa_spare_exists); EXPORT_SYMBOL(spa_spare_activate); /* L2ARC statech is global across all pools) */ EXPORT_SYMBOL(spa_l2cache_add); EXPORT_SYMBOL(spa_l2cache_remove); EXPORT_SYMBOL(spa_l2cache_exists); EXPORT_SYMBOL(spa_l2cache_activate); EXPORT_SYMBOL(spa_l2cache_drop); /* scanning */ EXPORT_SYMBOL(spa_scan); EXPORT_SYMBOL(spa_scan_stop); /* spa syncing */ EXPORT_SYMBOL(spa_sync); /* only for DMU use */ EXPORT_SYMBOL(spa_sync_allpools); /* properties */ EXPORT_SYMBOL(spa_prop_set); EXPORT_SYMBOL(spa_prop_get); EXPORT_SYMBOL(spa_prop_clear_bootfs); /* asynchronous event notification */ EXPORT_SYMBOL(spa_event_notify); /* BEGIN CSTYLED */ ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_shift, INT, ZMOD_RW, "log2 fraction of arc that can be used by inflight I/Os when " "verifying pool during import"); ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_metadata, INT, ZMOD_RW, "Set to traverse metadata on pool import"); ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW, "Set to traverse data on pool import"); ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW, "Print vdev tree to zfs_dbgmsg during pool import"); ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD, "Percentage of CPUs to run an IO worker thread"); ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_tpq, UINT, ZMOD_RD, "Number of threads per IO worker taskqueue"); ZFS_MODULE_PARAM(zfs, zfs_, max_missing_tvds, ULONG, ZMOD_RW, "Allow importing pool with up to this number of missing top-level " "vdevs (in read-only mode)"); ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_pause, INT, ZMOD_RW, "Set the livelist condense zthr to pause"); ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_pause, INT, ZMOD_RW, "Set the livelist condense synctask to pause"); ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_cancel, INT, ZMOD_RW, "Whether livelist condensing was canceled in the synctask"); ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_cancel, INT, ZMOD_RW, "Whether livelist condensing was canceled in the zthr function"); ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT, ZMOD_RW, "Whether extra ALLOC blkptrs were added to a livelist entry while it " "was being condensed"); /* END CSTYLED */ diff --git a/module/zfs/vdev_indirect.c b/module/zfs/vdev_indirect.c index ffe592b2a428..3237dc4021a0 100644 --- a/module/zfs/vdev_indirect.c +++ b/module/zfs/vdev_indirect.c @@ -1,1912 +1,1912 @@ /* * CDDL HEADER START * * This file and its contents are supplied under the terms of the * Common Development and Distribution License ("CDDL"), version 1.0. * You may only use this file in accordance with the terms of version * 1.0 of the CDDL. * * A full copy of the text of the CDDL should have accompanied this * source. A copy of the CDDL is also available via the Internet at * http://www.illumos.org/license/CDDL. * * CDDL HEADER END */ /* * Copyright (c) 2014, 2017 by Delphix. All rights reserved. * Copyright (c) 2019, loli10K . All rights reserved. * Copyright (c) 2014, 2020 by Delphix. All rights reserved. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * An indirect vdev corresponds to a vdev that has been removed. Since * we cannot rewrite block pointers of snapshots, etc., we keep a * mapping from old location on the removed device to the new location * on another device in the pool and use this mapping whenever we need * to access the DVA. Unfortunately, this mapping did not respect * logical block boundaries when it was first created, and so a DVA on * this indirect vdev may be "split" into multiple sections that each * map to a different location. As a consequence, not all DVAs can be * translated to an equivalent new DVA. Instead we must provide a * "vdev_remap" operation that executes a callback on each contiguous * segment of the new location. This function is used in multiple ways: * * - i/os to this vdev use the callback to determine where the * data is now located, and issue child i/os for each segment's new * location. * * - frees and claims to this vdev use the callback to free or claim * each mapped segment. (Note that we don't actually need to claim * log blocks on indirect vdevs, because we don't allocate to * removing vdevs. However, zdb uses zio_claim() for its leak * detection.) */ /* * "Big theory statement" for how we mark blocks obsolete. * * When a block on an indirect vdev is freed or remapped, a section of * that vdev's mapping may no longer be referenced (aka "obsolete"). We * keep track of how much of each mapping entry is obsolete. When * an entry becomes completely obsolete, we can remove it, thus reducing * the memory used by the mapping. The complete picture of obsolescence * is given by the following data structures, described below: * - the entry-specific obsolete count * - the vdev-specific obsolete spacemap * - the pool-specific obsolete bpobj * * == On disk data structures used == * * We track the obsolete space for the pool using several objects. Each * of these objects is created on demand and freed when no longer * needed, and is assumed to be empty if it does not exist. * SPA_FEATURE_OBSOLETE_COUNTS includes the count of these objects. * * - Each vic_mapping_object (associated with an indirect vdev) can * have a vimp_counts_object. This is an array of uint32_t's * with the same number of entries as the vic_mapping_object. When * the mapping is condensed, entries from the vic_obsolete_sm_object * (see below) are folded into the counts. Therefore, each * obsolete_counts entry tells us the number of bytes in the * corresponding mapping entry that were not referenced when the * mapping was last condensed. * * - Each indirect or removing vdev can have a vic_obsolete_sm_object. * This is a space map containing an alloc entry for every DVA that * has been obsoleted since the last time this indirect vdev was * condensed. We use this object in order to improve performance * when marking a DVA as obsolete. Instead of modifying an arbitrary * offset of the vimp_counts_object, we only need to append an entry * to the end of this object. When a DVA becomes obsolete, it is * added to the obsolete space map. This happens when the DVA is * freed, remapped and not referenced by a snapshot, or the last * snapshot referencing it is destroyed. * * - Each dataset can have a ds_remap_deadlist object. This is a * deadlist object containing all blocks that were remapped in this * dataset but referenced in a previous snapshot. Blocks can *only* * appear on this list if they were remapped (dsl_dataset_block_remapped); * blocks that were killed in a head dataset are put on the normal * ds_deadlist and marked obsolete when they are freed. * * - The pool can have a dp_obsolete_bpobj. This is a list of blocks * in the pool that need to be marked obsolete. When a snapshot is * destroyed, we move some of the ds_remap_deadlist to the obsolete * bpobj (see dsl_destroy_snapshot_handle_remaps()). We then * asynchronously process the obsolete bpobj, moving its entries to * the specific vdevs' obsolete space maps. * * == Summary of how we mark blocks as obsolete == * * - When freeing a block: if any DVA is on an indirect vdev, append to * vic_obsolete_sm_object. * - When remapping a block, add dva to ds_remap_deadlist (if prev snap * references; otherwise append to vic_obsolete_sm_object). * - When freeing a snapshot: move parts of ds_remap_deadlist to * dp_obsolete_bpobj (same algorithm as ds_deadlist). * - When syncing the spa: process dp_obsolete_bpobj, moving ranges to * individual vdev's vic_obsolete_sm_object. */ /* * "Big theory statement" for how we condense indirect vdevs. * * Condensing an indirect vdev's mapping is the process of determining * the precise counts of obsolete space for each mapping entry (by * integrating the obsolete spacemap into the obsolete counts) and * writing out a new mapping that contains only referenced entries. * * We condense a vdev when we expect the mapping to shrink (see * vdev_indirect_should_condense()), but only perform one condense at a * time to limit the memory usage. In addition, we use a separate * open-context thread (spa_condense_indirect_thread) to incrementally * create the new mapping object in a way that minimizes the impact on * the rest of the system. * * == Generating a new mapping == * * To generate a new mapping, we follow these steps: * * 1. Save the old obsolete space map and create a new mapping object * (see spa_condense_indirect_start_sync()). This initializes the * spa_condensing_indirect_phys with the "previous obsolete space map", * which is now read only. Newly obsolete DVAs will be added to a * new (initially empty) obsolete space map, and will not be * considered as part of this condense operation. * * 2. Construct in memory the precise counts of obsolete space for each * mapping entry, by incorporating the obsolete space map into the * counts. (See vdev_indirect_mapping_load_obsolete_{counts,spacemap}().) * * 3. Iterate through each mapping entry, writing to the new mapping any * entries that are not completely obsolete (i.e. which don't have * obsolete count == mapping length). (See * spa_condense_indirect_generate_new_mapping().) * * 4. Destroy the old mapping object and switch over to the new one * (spa_condense_indirect_complete_sync). * * == Restarting from failure == * * To restart the condense when we import/open the pool, we must start * at the 2nd step above: reconstruct the precise counts in memory, * based on the space map + counts. Then in the 3rd step, we start * iterating where we left off: at vimp_max_offset of the new mapping * object. */ int zfs_condense_indirect_vdevs_enable = B_TRUE; /* * Condense if at least this percent of the bytes in the mapping is * obsolete. With the default of 25%, the amount of space mapped * will be reduced to 1% of its original size after at most 16 * condenses. Higher values will condense less often (causing less * i/o); lower values will reduce the mapping size more quickly. */ int zfs_condense_indirect_obsolete_pct = 25; /* * Condense if the obsolete space map takes up more than this amount of * space on disk (logically). This limits the amount of disk space * consumed by the obsolete space map; the default of 1GB is small enough * that we typically don't mind "wasting" it. */ unsigned long zfs_condense_max_obsolete_bytes = 1024 * 1024 * 1024; /* * Don't bother condensing if the mapping uses less than this amount of * memory. The default of 128KB is considered a "trivial" amount of * memory and not worth reducing. */ unsigned long zfs_condense_min_mapping_bytes = 128 * 1024; /* * This is used by the test suite so that it can ensure that certain * actions happen while in the middle of a condense (which might otherwise * complete too quickly). If used to reduce the performance impact of * condensing in production, a maximum value of 1 should be sufficient. */ int zfs_condense_indirect_commit_entry_delay_ms = 0; /* * If an indirect split block contains more than this many possible unique * combinations when being reconstructed, consider it too computationally * expensive to check them all. Instead, try at most 100 randomly-selected * combinations each time the block is accessed. This allows all segment * copies to participate fairly in the reconstruction when all combinations * cannot be checked and prevents repeated use of one bad copy. */ int zfs_reconstruct_indirect_combinations_max = 4096; /* * Enable to simulate damaged segments and validate reconstruction. This * is intentionally not exposed as a module parameter. */ unsigned long zfs_reconstruct_indirect_damage_fraction = 0; /* * The indirect_child_t represents the vdev that we will read from, when we * need to read all copies of the data (e.g. for scrub or reconstruction). * For plain (non-mirror) top-level vdevs (i.e. is_vdev is not a mirror), * ic_vdev is the same as is_vdev. However, for mirror top-level vdevs, * ic_vdev is a child of the mirror. */ typedef struct indirect_child { abd_t *ic_data; vdev_t *ic_vdev; /* * ic_duplicate is NULL when the ic_data contents are unique, when it * is determined to be a duplicate it references the primary child. */ struct indirect_child *ic_duplicate; list_node_t ic_node; /* node on is_unique_child */ int ic_error; /* set when a child does not contain the data */ } indirect_child_t; /* * The indirect_split_t represents one mapped segment of an i/o to the * indirect vdev. For non-split (contiguously-mapped) blocks, there will be * only one indirect_split_t, with is_split_offset==0 and is_size==io_size. * For split blocks, there will be several of these. */ typedef struct indirect_split { list_node_t is_node; /* link on iv_splits */ /* * is_split_offset is the offset into the i/o. * This is the sum of the previous splits' is_size's. */ uint64_t is_split_offset; vdev_t *is_vdev; /* top-level vdev */ uint64_t is_target_offset; /* offset on is_vdev */ uint64_t is_size; int is_children; /* number of entries in is_child[] */ int is_unique_children; /* number of entries in is_unique_child */ list_t is_unique_child; /* * is_good_child is the child that we are currently using to * attempt reconstruction. */ indirect_child_t *is_good_child; indirect_child_t is_child[1]; /* variable-length */ } indirect_split_t; /* * The indirect_vsd_t is associated with each i/o to the indirect vdev. * It is the "Vdev-Specific Data" in the zio_t's io_vsd. */ typedef struct indirect_vsd { boolean_t iv_split_block; boolean_t iv_reconstruct; uint64_t iv_unique_combinations; uint64_t iv_attempts; uint64_t iv_attempts_max; list_t iv_splits; /* list of indirect_split_t's */ } indirect_vsd_t; static void vdev_indirect_map_free(zio_t *zio) { indirect_vsd_t *iv = zio->io_vsd; indirect_split_t *is; while ((is = list_head(&iv->iv_splits)) != NULL) { for (int c = 0; c < is->is_children; c++) { indirect_child_t *ic = &is->is_child[c]; if (ic->ic_data != NULL) abd_free(ic->ic_data); } list_remove(&iv->iv_splits, is); indirect_child_t *ic; while ((ic = list_head(&is->is_unique_child)) != NULL) list_remove(&is->is_unique_child, ic); list_destroy(&is->is_unique_child); kmem_free(is, offsetof(indirect_split_t, is_child[is->is_children])); } kmem_free(iv, sizeof (*iv)); } static const zio_vsd_ops_t vdev_indirect_vsd_ops = { .vsd_free = vdev_indirect_map_free, }; /* * Mark the given offset and size as being obsolete. */ void vdev_indirect_mark_obsolete(vdev_t *vd, uint64_t offset, uint64_t size) { spa_t *spa = vd->vdev_spa; ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, !=, 0); ASSERT(vd->vdev_removing || vd->vdev_ops == &vdev_indirect_ops); ASSERT(size > 0); VERIFY(vdev_indirect_mapping_entry_for_offset( vd->vdev_indirect_mapping, offset) != NULL); if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) { mutex_enter(&vd->vdev_obsolete_lock); range_tree_add(vd->vdev_obsolete_segments, offset, size); mutex_exit(&vd->vdev_obsolete_lock); vdev_dirty(vd, 0, NULL, spa_syncing_txg(spa)); } } /* * Mark the DVA vdev_id:offset:size as being obsolete in the given tx. This * wrapper is provided because the DMU does not know about vdev_t's and * cannot directly call vdev_indirect_mark_obsolete. */ void spa_vdev_indirect_mark_obsolete(spa_t *spa, uint64_t vdev_id, uint64_t offset, uint64_t size, dmu_tx_t *tx) { vdev_t *vd = vdev_lookup_top(spa, vdev_id); ASSERT(dmu_tx_is_syncing(tx)); /* The DMU can only remap indirect vdevs. */ ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); vdev_indirect_mark_obsolete(vd, offset, size); } static spa_condensing_indirect_t * spa_condensing_indirect_create(spa_t *spa) { spa_condensing_indirect_phys_t *scip = &spa->spa_condensing_indirect_phys; spa_condensing_indirect_t *sci = kmem_zalloc(sizeof (*sci), KM_SLEEP); objset_t *mos = spa->spa_meta_objset; for (int i = 0; i < TXG_SIZE; i++) { list_create(&sci->sci_new_mapping_entries[i], sizeof (vdev_indirect_mapping_entry_t), offsetof(vdev_indirect_mapping_entry_t, vime_node)); } sci->sci_new_mapping = vdev_indirect_mapping_open(mos, scip->scip_next_mapping_object); return (sci); } static void spa_condensing_indirect_destroy(spa_condensing_indirect_t *sci) { for (int i = 0; i < TXG_SIZE; i++) list_destroy(&sci->sci_new_mapping_entries[i]); if (sci->sci_new_mapping != NULL) vdev_indirect_mapping_close(sci->sci_new_mapping); kmem_free(sci, sizeof (*sci)); } boolean_t vdev_indirect_should_condense(vdev_t *vd) { vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; spa_t *spa = vd->vdev_spa; ASSERT(dsl_pool_sync_context(spa->spa_dsl_pool)); if (!zfs_condense_indirect_vdevs_enable) return (B_FALSE); /* * We can only condense one indirect vdev at a time. */ if (spa->spa_condensing_indirect != NULL) return (B_FALSE); if (spa_shutting_down(spa)) return (B_FALSE); /* * The mapping object size must not change while we are * condensing, so we can only condense indirect vdevs * (not vdevs that are still in the middle of being removed). */ if (vd->vdev_ops != &vdev_indirect_ops) return (B_FALSE); /* * If nothing new has been marked obsolete, there is no * point in condensing. */ uint64_t obsolete_sm_obj __maybe_unused; ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_obj)); if (vd->vdev_obsolete_sm == NULL) { ASSERT0(obsolete_sm_obj); return (B_FALSE); } ASSERT(vd->vdev_obsolete_sm != NULL); ASSERT3U(obsolete_sm_obj, ==, space_map_object(vd->vdev_obsolete_sm)); uint64_t bytes_mapped = vdev_indirect_mapping_bytes_mapped(vim); uint64_t bytes_obsolete = space_map_allocated(vd->vdev_obsolete_sm); uint64_t mapping_size = vdev_indirect_mapping_size(vim); uint64_t obsolete_sm_size = space_map_length(vd->vdev_obsolete_sm); ASSERT3U(bytes_obsolete, <=, bytes_mapped); /* * If a high percentage of the bytes that are mapped have become * obsolete, condense (unless the mapping is already small enough). * This has a good chance of reducing the amount of memory used * by the mapping. */ if (bytes_obsolete * 100 / bytes_mapped >= zfs_condense_indirect_obsolete_pct && mapping_size > zfs_condense_min_mapping_bytes) { zfs_dbgmsg("should condense vdev %llu because obsolete " "spacemap covers %d%% of %lluMB mapping", (u_longlong_t)vd->vdev_id, (int)(bytes_obsolete * 100 / bytes_mapped), (u_longlong_t)bytes_mapped / 1024 / 1024); return (B_TRUE); } /* * If the obsolete space map takes up too much space on disk, * condense in order to free up this disk space. */ if (obsolete_sm_size >= zfs_condense_max_obsolete_bytes) { zfs_dbgmsg("should condense vdev %llu because obsolete sm " "length %lluMB >= max size %lluMB", (u_longlong_t)vd->vdev_id, (u_longlong_t)obsolete_sm_size / 1024 / 1024, (u_longlong_t)zfs_condense_max_obsolete_bytes / 1024 / 1024); return (B_TRUE); } return (B_FALSE); } /* * This sync task completes (finishes) a condense, deleting the old * mapping and replacing it with the new one. */ static void spa_condense_indirect_complete_sync(void *arg, dmu_tx_t *tx) { spa_condensing_indirect_t *sci = arg; spa_t *spa = dmu_tx_pool(tx)->dp_spa; spa_condensing_indirect_phys_t *scip = &spa->spa_condensing_indirect_phys; vdev_t *vd = vdev_lookup_top(spa, scip->scip_vdev); vdev_indirect_config_t *vic = &vd->vdev_indirect_config; objset_t *mos = spa->spa_meta_objset; vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping; uint64_t old_count = vdev_indirect_mapping_num_entries(old_mapping); uint64_t new_count = vdev_indirect_mapping_num_entries(sci->sci_new_mapping); ASSERT(dmu_tx_is_syncing(tx)); ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); ASSERT3P(sci, ==, spa->spa_condensing_indirect); for (int i = 0; i < TXG_SIZE; i++) { ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i])); } ASSERT(vic->vic_mapping_object != 0); ASSERT3U(vd->vdev_id, ==, scip->scip_vdev); ASSERT(scip->scip_next_mapping_object != 0); ASSERT(scip->scip_prev_obsolete_sm_object != 0); /* * Reset vdev_indirect_mapping to refer to the new object. */ rw_enter(&vd->vdev_indirect_rwlock, RW_WRITER); vdev_indirect_mapping_close(vd->vdev_indirect_mapping); vd->vdev_indirect_mapping = sci->sci_new_mapping; rw_exit(&vd->vdev_indirect_rwlock); sci->sci_new_mapping = NULL; vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx); vic->vic_mapping_object = scip->scip_next_mapping_object; scip->scip_next_mapping_object = 0; space_map_free_obj(mos, scip->scip_prev_obsolete_sm_object, tx); spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); scip->scip_prev_obsolete_sm_object = 0; scip->scip_vdev = 0; VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONDENSING_INDIRECT, tx)); spa_condensing_indirect_destroy(spa->spa_condensing_indirect); spa->spa_condensing_indirect = NULL; zfs_dbgmsg("finished condense of vdev %llu in txg %llu: " "new mapping object %llu has %llu entries " "(was %llu entries)", (u_longlong_t)vd->vdev_id, (u_longlong_t)dmu_tx_get_txg(tx), (u_longlong_t)vic->vic_mapping_object, (u_longlong_t)new_count, (u_longlong_t)old_count); vdev_config_dirty(spa->spa_root_vdev); } /* * This sync task appends entries to the new mapping object. */ static void spa_condense_indirect_commit_sync(void *arg, dmu_tx_t *tx) { spa_condensing_indirect_t *sci = arg; uint64_t txg = dmu_tx_get_txg(tx); spa_t *spa __maybe_unused = dmu_tx_pool(tx)->dp_spa; ASSERT(dmu_tx_is_syncing(tx)); ASSERT3P(sci, ==, spa->spa_condensing_indirect); vdev_indirect_mapping_add_entries(sci->sci_new_mapping, &sci->sci_new_mapping_entries[txg & TXG_MASK], tx); ASSERT(list_is_empty(&sci->sci_new_mapping_entries[txg & TXG_MASK])); } /* * Open-context function to add one entry to the new mapping. The new * entry will be remembered and written from syncing context. */ static void spa_condense_indirect_commit_entry(spa_t *spa, vdev_indirect_mapping_entry_phys_t *vimep, uint32_t count) { spa_condensing_indirect_t *sci = spa->spa_condensing_indirect; ASSERT3U(count, <, DVA_GET_ASIZE(&vimep->vimep_dst)); dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); dmu_tx_hold_space(tx, sizeof (*vimep) + sizeof (count)); VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); int txgoff = dmu_tx_get_txg(tx) & TXG_MASK; /* * If we are the first entry committed this txg, kick off the sync * task to write to the MOS on our behalf. */ if (list_is_empty(&sci->sci_new_mapping_entries[txgoff])) { dsl_sync_task_nowait(dmu_tx_pool(tx), spa_condense_indirect_commit_sync, sci, tx); } vdev_indirect_mapping_entry_t *vime = kmem_alloc(sizeof (*vime), KM_SLEEP); vime->vime_mapping = *vimep; vime->vime_obsolete_count = count; list_insert_tail(&sci->sci_new_mapping_entries[txgoff], vime); dmu_tx_commit(tx); } static void spa_condense_indirect_generate_new_mapping(vdev_t *vd, uint32_t *obsolete_counts, uint64_t start_index, zthr_t *zthr) { spa_t *spa = vd->vdev_spa; uint64_t mapi = start_index; vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping; uint64_t old_num_entries = vdev_indirect_mapping_num_entries(old_mapping); ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); ASSERT3U(vd->vdev_id, ==, spa->spa_condensing_indirect_phys.scip_vdev); zfs_dbgmsg("starting condense of vdev %llu from index %llu", (u_longlong_t)vd->vdev_id, (u_longlong_t)mapi); while (mapi < old_num_entries) { if (zthr_iscancelled(zthr)) { zfs_dbgmsg("pausing condense of vdev %llu " "at index %llu", (u_longlong_t)vd->vdev_id, (u_longlong_t)mapi); break; } vdev_indirect_mapping_entry_phys_t *entry = &old_mapping->vim_entries[mapi]; uint64_t entry_size = DVA_GET_ASIZE(&entry->vimep_dst); ASSERT3U(obsolete_counts[mapi], <=, entry_size); if (obsolete_counts[mapi] < entry_size) { spa_condense_indirect_commit_entry(spa, entry, obsolete_counts[mapi]); /* * This delay may be requested for testing, debugging, * or performance reasons. */ hrtime_t now = gethrtime(); hrtime_t sleep_until = now + MSEC2NSEC( zfs_condense_indirect_commit_entry_delay_ms); zfs_sleep_until(sleep_until); } mapi++; } } /* ARGSUSED */ static boolean_t spa_condense_indirect_thread_check(void *arg, zthr_t *zthr) { spa_t *spa = arg; return (spa->spa_condensing_indirect != NULL); } /* ARGSUSED */ static void spa_condense_indirect_thread(void *arg, zthr_t *zthr) { spa_t *spa = arg; vdev_t *vd; ASSERT3P(spa->spa_condensing_indirect, !=, NULL); spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); vd = vdev_lookup_top(spa, spa->spa_condensing_indirect_phys.scip_vdev); ASSERT3P(vd, !=, NULL); spa_config_exit(spa, SCL_VDEV, FTAG); spa_condensing_indirect_t *sci = spa->spa_condensing_indirect; spa_condensing_indirect_phys_t *scip = &spa->spa_condensing_indirect_phys; uint32_t *counts; uint64_t start_index; vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping; space_map_t *prev_obsolete_sm = NULL; ASSERT3U(vd->vdev_id, ==, scip->scip_vdev); ASSERT(scip->scip_next_mapping_object != 0); ASSERT(scip->scip_prev_obsolete_sm_object != 0); ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); for (int i = 0; i < TXG_SIZE; i++) { /* * The list must start out empty in order for the * _commit_sync() sync task to be properly registered * on the first call to _commit_entry(); so it's wise * to double check and ensure we actually are starting * with empty lists. */ ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i])); } VERIFY0(space_map_open(&prev_obsolete_sm, spa->spa_meta_objset, scip->scip_prev_obsolete_sm_object, 0, vd->vdev_asize, 0)); counts = vdev_indirect_mapping_load_obsolete_counts(old_mapping); if (prev_obsolete_sm != NULL) { vdev_indirect_mapping_load_obsolete_spacemap(old_mapping, counts, prev_obsolete_sm); } space_map_close(prev_obsolete_sm); /* * Generate new mapping. Determine what index to continue from * based on the max offset that we've already written in the * new mapping. */ uint64_t max_offset = vdev_indirect_mapping_max_offset(sci->sci_new_mapping); if (max_offset == 0) { /* We haven't written anything to the new mapping yet. */ start_index = 0; } else { /* * Pick up from where we left off. _entry_for_offset() * returns a pointer into the vim_entries array. If * max_offset is greater than any of the mappings * contained in the table NULL will be returned and * that indicates we've exhausted our iteration of the * old_mapping. */ vdev_indirect_mapping_entry_phys_t *entry = vdev_indirect_mapping_entry_for_offset_or_next(old_mapping, max_offset); if (entry == NULL) { /* * We've already written the whole new mapping. * This special value will cause us to skip the * generate_new_mapping step and just do the sync * task to complete the condense. */ start_index = UINT64_MAX; } else { start_index = entry - old_mapping->vim_entries; ASSERT3U(start_index, <, vdev_indirect_mapping_num_entries(old_mapping)); } } spa_condense_indirect_generate_new_mapping(vd, counts, start_index, zthr); vdev_indirect_mapping_free_obsolete_counts(old_mapping, counts); /* * If the zthr has received a cancellation signal while running * in generate_new_mapping() or at any point after that, then bail * early. We don't want to complete the condense if the spa is * shutting down. */ if (zthr_iscancelled(zthr)) return; VERIFY0(dsl_sync_task(spa_name(spa), NULL, spa_condense_indirect_complete_sync, sci, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED)); } /* * Sync task to begin the condensing process. */ void spa_condense_indirect_start_sync(vdev_t *vd, dmu_tx_t *tx) { spa_t *spa = vd->vdev_spa; spa_condensing_indirect_phys_t *scip = &spa->spa_condensing_indirect_phys; ASSERT0(scip->scip_next_mapping_object); ASSERT0(scip->scip_prev_obsolete_sm_object); ASSERT0(scip->scip_vdev); ASSERT(dmu_tx_is_syncing(tx)); ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); ASSERT(spa_feature_is_active(spa, SPA_FEATURE_OBSOLETE_COUNTS)); ASSERT(vdev_indirect_mapping_num_entries(vd->vdev_indirect_mapping)); uint64_t obsolete_sm_obj; VERIFY0(vdev_obsolete_sm_object(vd, &obsolete_sm_obj)); ASSERT3U(obsolete_sm_obj, !=, 0); scip->scip_vdev = vd->vdev_id; scip->scip_next_mapping_object = vdev_indirect_mapping_alloc(spa->spa_meta_objset, tx); scip->scip_prev_obsolete_sm_object = obsolete_sm_obj; /* * We don't need to allocate a new space map object, since * vdev_indirect_sync_obsolete will allocate one when needed. */ space_map_close(vd->vdev_obsolete_sm); vd->vdev_obsolete_sm = NULL; VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap, VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx)); VERIFY0(zap_add(spa->spa_dsl_pool->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONDENSING_INDIRECT, sizeof (uint64_t), sizeof (*scip) / sizeof (uint64_t), scip, tx)); ASSERT3P(spa->spa_condensing_indirect, ==, NULL); spa->spa_condensing_indirect = spa_condensing_indirect_create(spa); zfs_dbgmsg("starting condense of vdev %llu in txg %llu: " "posm=%llu nm=%llu", (u_longlong_t)vd->vdev_id, (u_longlong_t)dmu_tx_get_txg(tx), (u_longlong_t)scip->scip_prev_obsolete_sm_object, (u_longlong_t)scip->scip_next_mapping_object); zthr_wakeup(spa->spa_condense_zthr); } /* * Sync to the given vdev's obsolete space map any segments that are no longer * referenced as of the given txg. * * If the obsolete space map doesn't exist yet, create and open it. */ void vdev_indirect_sync_obsolete(vdev_t *vd, dmu_tx_t *tx) { spa_t *spa = vd->vdev_spa; vdev_indirect_config_t *vic __maybe_unused = &vd->vdev_indirect_config; ASSERT3U(vic->vic_mapping_object, !=, 0); ASSERT(range_tree_space(vd->vdev_obsolete_segments) > 0); ASSERT(vd->vdev_removing || vd->vdev_ops == &vdev_indirect_ops); ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)); uint64_t obsolete_sm_object; VERIFY0(vdev_obsolete_sm_object(vd, &obsolete_sm_object)); if (obsolete_sm_object == 0) { obsolete_sm_object = space_map_alloc(spa->spa_meta_objset, zfs_vdev_standard_sm_blksz, tx); ASSERT(vd->vdev_top_zap != 0); VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset, vd->vdev_top_zap, VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, sizeof (obsolete_sm_object), 1, &obsolete_sm_object, tx)); ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object)); ASSERT3U(obsolete_sm_object, !=, 0); spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); VERIFY0(space_map_open(&vd->vdev_obsolete_sm, spa->spa_meta_objset, obsolete_sm_object, 0, vd->vdev_asize, 0)); } ASSERT(vd->vdev_obsolete_sm != NULL); ASSERT3U(obsolete_sm_object, ==, space_map_object(vd->vdev_obsolete_sm)); space_map_write(vd->vdev_obsolete_sm, vd->vdev_obsolete_segments, SM_ALLOC, SM_NO_VDEVID, tx); range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL); } int spa_condense_init(spa_t *spa) { int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONDENSING_INDIRECT, sizeof (uint64_t), sizeof (spa->spa_condensing_indirect_phys) / sizeof (uint64_t), &spa->spa_condensing_indirect_phys); if (error == 0) { if (spa_writeable(spa)) { spa->spa_condensing_indirect = spa_condensing_indirect_create(spa); } return (0); } else if (error == ENOENT) { return (0); } else { return (error); } } void spa_condense_fini(spa_t *spa) { if (spa->spa_condensing_indirect != NULL) { spa_condensing_indirect_destroy(spa->spa_condensing_indirect); spa->spa_condensing_indirect = NULL; } } void spa_start_indirect_condensing_thread(spa_t *spa) { ASSERT3P(spa->spa_condense_zthr, ==, NULL); spa->spa_condense_zthr = zthr_create("z_indirect_condense", spa_condense_indirect_thread_check, - spa_condense_indirect_thread, spa); + spa_condense_indirect_thread, spa, minclsyspri); } /* * Gets the obsolete spacemap object from the vdev's ZAP. On success sm_obj * will contain either the obsolete spacemap object or zero if none exists. * All other errors are returned to the caller. */ int vdev_obsolete_sm_object(vdev_t *vd, uint64_t *sm_obj) { ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER)); if (vd->vdev_top_zap == 0) { *sm_obj = 0; return (0); } int error = zap_lookup(vd->vdev_spa->spa_meta_objset, vd->vdev_top_zap, VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, sizeof (uint64_t), 1, sm_obj); if (error == ENOENT) { *sm_obj = 0; error = 0; } return (error); } /* * Gets the obsolete count are precise spacemap object from the vdev's ZAP. * On success are_precise will be set to reflect if the counts are precise. * All other errors are returned to the caller. */ int vdev_obsolete_counts_are_precise(vdev_t *vd, boolean_t *are_precise) { ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER)); if (vd->vdev_top_zap == 0) { *are_precise = B_FALSE; return (0); } uint64_t val = 0; int error = zap_lookup(vd->vdev_spa->spa_meta_objset, vd->vdev_top_zap, VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (val), 1, &val); if (error == 0) { *are_precise = (val != 0); } else if (error == ENOENT) { *are_precise = B_FALSE; error = 0; } return (error); } /* ARGSUSED */ static void vdev_indirect_close(vdev_t *vd) { } /* ARGSUSED */ static int vdev_indirect_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize, uint64_t *logical_ashift, uint64_t *physical_ashift) { *psize = *max_psize = vd->vdev_asize + VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE; *logical_ashift = vd->vdev_ashift; *physical_ashift = vd->vdev_physical_ashift; return (0); } typedef struct remap_segment { vdev_t *rs_vd; uint64_t rs_offset; uint64_t rs_asize; uint64_t rs_split_offset; list_node_t rs_node; } remap_segment_t; static remap_segment_t * rs_alloc(vdev_t *vd, uint64_t offset, uint64_t asize, uint64_t split_offset) { remap_segment_t *rs = kmem_alloc(sizeof (remap_segment_t), KM_SLEEP); rs->rs_vd = vd; rs->rs_offset = offset; rs->rs_asize = asize; rs->rs_split_offset = split_offset; return (rs); } /* * Given an indirect vdev and an extent on that vdev, it duplicates the * physical entries of the indirect mapping that correspond to the extent * to a new array and returns a pointer to it. In addition, copied_entries * is populated with the number of mapping entries that were duplicated. * * Note that the function assumes that the caller holds vdev_indirect_rwlock. * This ensures that the mapping won't change due to condensing as we * copy over its contents. * * Finally, since we are doing an allocation, it is up to the caller to * free the array allocated in this function. */ static vdev_indirect_mapping_entry_phys_t * vdev_indirect_mapping_duplicate_adjacent_entries(vdev_t *vd, uint64_t offset, uint64_t asize, uint64_t *copied_entries) { vdev_indirect_mapping_entry_phys_t *duplicate_mappings = NULL; vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; uint64_t entries = 0; ASSERT(RW_READ_HELD(&vd->vdev_indirect_rwlock)); vdev_indirect_mapping_entry_phys_t *first_mapping = vdev_indirect_mapping_entry_for_offset(vim, offset); ASSERT3P(first_mapping, !=, NULL); vdev_indirect_mapping_entry_phys_t *m = first_mapping; while (asize > 0) { uint64_t size = DVA_GET_ASIZE(&m->vimep_dst); ASSERT3U(offset, >=, DVA_MAPPING_GET_SRC_OFFSET(m)); ASSERT3U(offset, <, DVA_MAPPING_GET_SRC_OFFSET(m) + size); uint64_t inner_offset = offset - DVA_MAPPING_GET_SRC_OFFSET(m); uint64_t inner_size = MIN(asize, size - inner_offset); offset += inner_size; asize -= inner_size; entries++; m++; } size_t copy_length = entries * sizeof (*first_mapping); duplicate_mappings = kmem_alloc(copy_length, KM_SLEEP); bcopy(first_mapping, duplicate_mappings, copy_length); *copied_entries = entries; return (duplicate_mappings); } /* * Goes through the relevant indirect mappings until it hits a concrete vdev * and issues the callback. On the way to the concrete vdev, if any other * indirect vdevs are encountered, then the callback will also be called on * each of those indirect vdevs. For example, if the segment is mapped to * segment A on indirect vdev 1, and then segment A on indirect vdev 1 is * mapped to segment B on concrete vdev 2, then the callback will be called on * both vdev 1 and vdev 2. * * While the callback passed to vdev_indirect_remap() is called on every vdev * the function encounters, certain callbacks only care about concrete vdevs. * These types of callbacks should return immediately and explicitly when they * are called on an indirect vdev. * * Because there is a possibility that a DVA section in the indirect device * has been split into multiple sections in our mapping, we keep track * of the relevant contiguous segments of the new location (remap_segment_t) * in a stack. This way we can call the callback for each of the new sections * created by a single section of the indirect device. Note though, that in * this scenario the callbacks in each split block won't occur in-order in * terms of offset, so callers should not make any assumptions about that. * * For callbacks that don't handle split blocks and immediately return when * they encounter them (as is the case for remap_blkptr_cb), the caller can * assume that its callback will be applied from the first indirect vdev * encountered to the last one and then the concrete vdev, in that order. */ static void vdev_indirect_remap(vdev_t *vd, uint64_t offset, uint64_t asize, void (*func)(uint64_t, vdev_t *, uint64_t, uint64_t, void *), void *arg) { list_t stack; spa_t *spa = vd->vdev_spa; list_create(&stack, sizeof (remap_segment_t), offsetof(remap_segment_t, rs_node)); for (remap_segment_t *rs = rs_alloc(vd, offset, asize, 0); rs != NULL; rs = list_remove_head(&stack)) { vdev_t *v = rs->rs_vd; uint64_t num_entries = 0; ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0); ASSERT(rs->rs_asize > 0); /* * Note: As this function can be called from open context * (e.g. zio_read()), we need the following rwlock to * prevent the mapping from being changed by condensing. * * So we grab the lock and we make a copy of the entries * that are relevant to the extent that we are working on. * Once that is done, we drop the lock and iterate over * our copy of the mapping. Once we are done with the with * the remap segment and we free it, we also free our copy * of the indirect mapping entries that are relevant to it. * * This way we don't need to wait until the function is * finished with a segment, to condense it. In addition, we * don't need a recursive rwlock for the case that a call to * vdev_indirect_remap() needs to call itself (through the * codepath of its callback) for the same vdev in the middle * of its execution. */ rw_enter(&v->vdev_indirect_rwlock, RW_READER); ASSERT3P(v->vdev_indirect_mapping, !=, NULL); vdev_indirect_mapping_entry_phys_t *mapping = vdev_indirect_mapping_duplicate_adjacent_entries(v, rs->rs_offset, rs->rs_asize, &num_entries); ASSERT3P(mapping, !=, NULL); ASSERT3U(num_entries, >, 0); rw_exit(&v->vdev_indirect_rwlock); for (uint64_t i = 0; i < num_entries; i++) { /* * Note: the vdev_indirect_mapping can not change * while we are running. It only changes while the * removal is in progress, and then only from syncing * context. While a removal is in progress, this * function is only called for frees, which also only * happen from syncing context. */ vdev_indirect_mapping_entry_phys_t *m = &mapping[i]; ASSERT3P(m, !=, NULL); ASSERT3U(rs->rs_asize, >, 0); uint64_t size = DVA_GET_ASIZE(&m->vimep_dst); uint64_t dst_offset = DVA_GET_OFFSET(&m->vimep_dst); uint64_t dst_vdev = DVA_GET_VDEV(&m->vimep_dst); ASSERT3U(rs->rs_offset, >=, DVA_MAPPING_GET_SRC_OFFSET(m)); ASSERT3U(rs->rs_offset, <, DVA_MAPPING_GET_SRC_OFFSET(m) + size); ASSERT3U(dst_vdev, !=, v->vdev_id); uint64_t inner_offset = rs->rs_offset - DVA_MAPPING_GET_SRC_OFFSET(m); uint64_t inner_size = MIN(rs->rs_asize, size - inner_offset); vdev_t *dst_v = vdev_lookup_top(spa, dst_vdev); ASSERT3P(dst_v, !=, NULL); if (dst_v->vdev_ops == &vdev_indirect_ops) { list_insert_head(&stack, rs_alloc(dst_v, dst_offset + inner_offset, inner_size, rs->rs_split_offset)); } if ((zfs_flags & ZFS_DEBUG_INDIRECT_REMAP) && IS_P2ALIGNED(inner_size, 2 * SPA_MINBLOCKSIZE)) { /* * Note: This clause exists only solely for * testing purposes. We use it to ensure that * split blocks work and that the callbacks * using them yield the same result if issued * in reverse order. */ uint64_t inner_half = inner_size / 2; func(rs->rs_split_offset + inner_half, dst_v, dst_offset + inner_offset + inner_half, inner_half, arg); func(rs->rs_split_offset, dst_v, dst_offset + inner_offset, inner_half, arg); } else { func(rs->rs_split_offset, dst_v, dst_offset + inner_offset, inner_size, arg); } rs->rs_offset += inner_size; rs->rs_asize -= inner_size; rs->rs_split_offset += inner_size; } VERIFY0(rs->rs_asize); kmem_free(mapping, num_entries * sizeof (*mapping)); kmem_free(rs, sizeof (remap_segment_t)); } list_destroy(&stack); } static void vdev_indirect_child_io_done(zio_t *zio) { zio_t *pio = zio->io_private; mutex_enter(&pio->io_lock); pio->io_error = zio_worst_error(pio->io_error, zio->io_error); mutex_exit(&pio->io_lock); abd_free(zio->io_abd); } /* * This is a callback for vdev_indirect_remap() which allocates an * indirect_split_t for each split segment and adds it to iv_splits. */ static void vdev_indirect_gather_splits(uint64_t split_offset, vdev_t *vd, uint64_t offset, uint64_t size, void *arg) { zio_t *zio = arg; indirect_vsd_t *iv = zio->io_vsd; ASSERT3P(vd, !=, NULL); if (vd->vdev_ops == &vdev_indirect_ops) return; int n = 1; if (vd->vdev_ops == &vdev_mirror_ops) n = vd->vdev_children; indirect_split_t *is = kmem_zalloc(offsetof(indirect_split_t, is_child[n]), KM_SLEEP); is->is_children = n; is->is_size = size; is->is_split_offset = split_offset; is->is_target_offset = offset; is->is_vdev = vd; list_create(&is->is_unique_child, sizeof (indirect_child_t), offsetof(indirect_child_t, ic_node)); /* * Note that we only consider multiple copies of the data for * *mirror* vdevs. We don't for "replacing" or "spare" vdevs, even * though they use the same ops as mirror, because there's only one * "good" copy under the replacing/spare. */ if (vd->vdev_ops == &vdev_mirror_ops) { for (int i = 0; i < n; i++) { is->is_child[i].ic_vdev = vd->vdev_child[i]; list_link_init(&is->is_child[i].ic_node); } } else { is->is_child[0].ic_vdev = vd; } list_insert_tail(&iv->iv_splits, is); } static void vdev_indirect_read_split_done(zio_t *zio) { indirect_child_t *ic = zio->io_private; if (zio->io_error != 0) { /* * Clear ic_data to indicate that we do not have data for this * child. */ abd_free(ic->ic_data); ic->ic_data = NULL; } } /* * Issue reads for all copies (mirror children) of all splits. */ static void vdev_indirect_read_all(zio_t *zio) { indirect_vsd_t *iv = zio->io_vsd; ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { for (int i = 0; i < is->is_children; i++) { indirect_child_t *ic = &is->is_child[i]; if (!vdev_readable(ic->ic_vdev)) continue; /* * If a child is missing the data, set ic_error. Used * in vdev_indirect_repair(). We perform the read * nevertheless which provides the opportunity to * reconstruct the split block if at all possible. */ if (vdev_dtl_contains(ic->ic_vdev, DTL_MISSING, zio->io_txg, 1)) ic->ic_error = SET_ERROR(ESTALE); ic->ic_data = abd_alloc_sametype(zio->io_abd, is->is_size); ic->ic_duplicate = NULL; zio_nowait(zio_vdev_child_io(zio, NULL, ic->ic_vdev, is->is_target_offset, ic->ic_data, is->is_size, zio->io_type, zio->io_priority, 0, vdev_indirect_read_split_done, ic)); } } iv->iv_reconstruct = B_TRUE; } static void vdev_indirect_io_start(zio_t *zio) { spa_t *spa __maybe_unused = zio->io_spa; indirect_vsd_t *iv = kmem_zalloc(sizeof (*iv), KM_SLEEP); list_create(&iv->iv_splits, sizeof (indirect_split_t), offsetof(indirect_split_t, is_node)); zio->io_vsd = iv; zio->io_vsd_ops = &vdev_indirect_vsd_ops; ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0); if (zio->io_type != ZIO_TYPE_READ) { ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); /* * Note: this code can handle other kinds of writes, * but we don't expect them. */ ASSERT((zio->io_flags & (ZIO_FLAG_SELF_HEAL | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)) != 0); } vdev_indirect_remap(zio->io_vd, zio->io_offset, zio->io_size, vdev_indirect_gather_splits, zio); indirect_split_t *first = list_head(&iv->iv_splits); if (first->is_size == zio->io_size) { /* * This is not a split block; we are pointing to the entire * data, which will checksum the same as the original data. * Pass the BP down so that the child i/o can verify the * checksum, and try a different location if available * (e.g. on a mirror). * * While this special case could be handled the same as the * general (split block) case, doing it this way ensures * that the vast majority of blocks on indirect vdevs * (which are not split) are handled identically to blocks * on non-indirect vdevs. This allows us to be less strict * about performance in the general (but rare) case. */ ASSERT0(first->is_split_offset); ASSERT3P(list_next(&iv->iv_splits, first), ==, NULL); zio_nowait(zio_vdev_child_io(zio, zio->io_bp, first->is_vdev, first->is_target_offset, abd_get_offset(zio->io_abd, 0), zio->io_size, zio->io_type, zio->io_priority, 0, vdev_indirect_child_io_done, zio)); } else { iv->iv_split_block = B_TRUE; if (zio->io_type == ZIO_TYPE_READ && zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER)) { /* * Read all copies. Note that for simplicity, * we don't bother consulting the DTL in the * resilver case. */ vdev_indirect_read_all(zio); } else { /* * If this is a read zio, we read one copy of each * split segment, from the top-level vdev. Since * we don't know the checksum of each split * individually, the child zio can't ensure that * we get the right data. E.g. if it's a mirror, * it will just read from a random (healthy) leaf * vdev. We have to verify the checksum in * vdev_indirect_io_done(). * * For write zios, the vdev code will ensure we write * to all children. */ for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { zio_nowait(zio_vdev_child_io(zio, NULL, is->is_vdev, is->is_target_offset, abd_get_offset(zio->io_abd, is->is_split_offset), is->is_size, zio->io_type, zio->io_priority, 0, vdev_indirect_child_io_done, zio)); } } } zio_execute(zio); } /* * Report a checksum error for a child. */ static void vdev_indirect_checksum_error(zio_t *zio, indirect_split_t *is, indirect_child_t *ic) { vdev_t *vd = ic->ic_vdev; if (zio->io_flags & ZIO_FLAG_SPECULATIVE) return; mutex_enter(&vd->vdev_stat_lock); vd->vdev_stat.vs_checksum_errors++; mutex_exit(&vd->vdev_stat_lock); zio_bad_cksum_t zbc = {{{ 0 }}}; abd_t *bad_abd = ic->ic_data; abd_t *good_abd = is->is_good_child->ic_data; (void) zfs_ereport_post_checksum(zio->io_spa, vd, NULL, zio, is->is_target_offset, is->is_size, good_abd, bad_abd, &zbc); } /* * Issue repair i/os for any incorrect copies. We do this by comparing * each split segment's correct data (is_good_child's ic_data) with each * other copy of the data. If they differ, then we overwrite the bad data * with the good copy. The DTL is checked in vdev_indirect_read_all() and * if a vdev is missing a copy of the data we set ic_error and the read is * performed. This provides the opportunity to reconstruct the split block * if at all possible. ic_error is checked here and if set it suppresses * incrementing the checksum counter. Aside from this DTLs are not checked, * which simplifies this code and also issues the optimal number of writes * (based on which copies actually read bad data, as opposed to which we * think might be wrong). For the same reason, we always use * ZIO_FLAG_SELF_HEAL, to bypass the DTL check in zio_vdev_io_start(). */ static void vdev_indirect_repair(zio_t *zio) { indirect_vsd_t *iv = zio->io_vsd; enum zio_flag flags = ZIO_FLAG_IO_REPAIR; if (!(zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) flags |= ZIO_FLAG_SELF_HEAL; if (!spa_writeable(zio->io_spa)) return; for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { for (int c = 0; c < is->is_children; c++) { indirect_child_t *ic = &is->is_child[c]; if (ic == is->is_good_child) continue; if (ic->ic_data == NULL) continue; if (ic->ic_duplicate == is->is_good_child) continue; zio_nowait(zio_vdev_child_io(zio, NULL, ic->ic_vdev, is->is_target_offset, is->is_good_child->ic_data, is->is_size, ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_IO_REPAIR | ZIO_FLAG_SELF_HEAL, NULL, NULL)); /* * If ic_error is set the current child does not have * a copy of the data, so suppress incrementing the * checksum counter. */ if (ic->ic_error == ESTALE) continue; vdev_indirect_checksum_error(zio, is, ic); } } } /* * Report checksum errors on all children that we read from. */ static void vdev_indirect_all_checksum_errors(zio_t *zio) { indirect_vsd_t *iv = zio->io_vsd; if (zio->io_flags & ZIO_FLAG_SPECULATIVE) return; for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { for (int c = 0; c < is->is_children; c++) { indirect_child_t *ic = &is->is_child[c]; if (ic->ic_data == NULL) continue; vdev_t *vd = ic->ic_vdev; (void) zfs_ereport_post_checksum(zio->io_spa, vd, NULL, zio, is->is_target_offset, is->is_size, NULL, NULL, NULL); mutex_enter(&vd->vdev_stat_lock); vd->vdev_stat.vs_checksum_errors++; mutex_exit(&vd->vdev_stat_lock); } } } /* * Copy data from all the splits to a main zio then validate the checksum. * If then checksum is successfully validated return success. */ static int vdev_indirect_splits_checksum_validate(indirect_vsd_t *iv, zio_t *zio) { zio_bad_cksum_t zbc; for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { ASSERT3P(is->is_good_child->ic_data, !=, NULL); ASSERT3P(is->is_good_child->ic_duplicate, ==, NULL); abd_copy_off(zio->io_abd, is->is_good_child->ic_data, is->is_split_offset, 0, is->is_size); } return (zio_checksum_error(zio, &zbc)); } /* * There are relatively few possible combinations making it feasible to * deterministically check them all. We do this by setting the good_child * to the next unique split version. If we reach the end of the list then * "carry over" to the next unique split version (like counting in base * is_unique_children, but each digit can have a different base). */ static int vdev_indirect_splits_enumerate_all(indirect_vsd_t *iv, zio_t *zio) { boolean_t more = B_TRUE; iv->iv_attempts = 0; for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) is->is_good_child = list_head(&is->is_unique_child); while (more == B_TRUE) { iv->iv_attempts++; more = B_FALSE; if (vdev_indirect_splits_checksum_validate(iv, zio) == 0) return (0); for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { is->is_good_child = list_next(&is->is_unique_child, is->is_good_child); if (is->is_good_child != NULL) { more = B_TRUE; break; } is->is_good_child = list_head(&is->is_unique_child); } } ASSERT3S(iv->iv_attempts, <=, iv->iv_unique_combinations); return (SET_ERROR(ECKSUM)); } /* * There are too many combinations to try all of them in a reasonable amount * of time. So try a fixed number of random combinations from the unique * split versions, after which we'll consider the block unrecoverable. */ static int vdev_indirect_splits_enumerate_randomly(indirect_vsd_t *iv, zio_t *zio) { iv->iv_attempts = 0; while (iv->iv_attempts < iv->iv_attempts_max) { iv->iv_attempts++; for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { indirect_child_t *ic = list_head(&is->is_unique_child); int children = is->is_unique_children; for (int i = random_in_range(children); i > 0; i--) ic = list_next(&is->is_unique_child, ic); ASSERT3P(ic, !=, NULL); is->is_good_child = ic; } if (vdev_indirect_splits_checksum_validate(iv, zio) == 0) return (0); } return (SET_ERROR(ECKSUM)); } /* * This is a validation function for reconstruction. It randomly selects * a good combination, if one can be found, and then it intentionally * damages all other segment copes by zeroing them. This forces the * reconstruction algorithm to locate the one remaining known good copy. */ static int vdev_indirect_splits_damage(indirect_vsd_t *iv, zio_t *zio) { int error; /* Presume all the copies are unique for initial selection. */ for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { is->is_unique_children = 0; for (int i = 0; i < is->is_children; i++) { indirect_child_t *ic = &is->is_child[i]; if (ic->ic_data != NULL) { is->is_unique_children++; list_insert_tail(&is->is_unique_child, ic); } } if (list_is_empty(&is->is_unique_child)) { error = SET_ERROR(EIO); goto out; } } /* * Set each is_good_child to a randomly-selected child which * is known to contain validated data. */ error = vdev_indirect_splits_enumerate_randomly(iv, zio); if (error) goto out; /* * Damage all but the known good copy by zeroing it. This will * result in two or less unique copies per indirect_child_t. * Both may need to be checked in order to reconstruct the block. * Set iv->iv_attempts_max such that all unique combinations will * enumerated, but limit the damage to at most 12 indirect splits. */ iv->iv_attempts_max = 1; for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { for (int c = 0; c < is->is_children; c++) { indirect_child_t *ic = &is->is_child[c]; if (ic == is->is_good_child) continue; if (ic->ic_data == NULL) continue; abd_zero(ic->ic_data, abd_get_size(ic->ic_data)); } iv->iv_attempts_max *= 2; if (iv->iv_attempts_max >= (1ULL << 12)) { iv->iv_attempts_max = UINT64_MAX; break; } } out: /* Empty the unique children lists so they can be reconstructed. */ for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { indirect_child_t *ic; while ((ic = list_head(&is->is_unique_child)) != NULL) list_remove(&is->is_unique_child, ic); is->is_unique_children = 0; } return (error); } /* * This function is called when we have read all copies of the data and need * to try to find a combination of copies that gives us the right checksum. * * If we pointed to any mirror vdevs, this effectively does the job of the * mirror. The mirror vdev code can't do its own job because we don't know * the checksum of each split segment individually. * * We have to try every unique combination of copies of split segments, until * we find one that checksums correctly. Duplicate segment copies are first * identified and latter skipped during reconstruction. This optimization * reduces the search space and ensures that of the remaining combinations * at most one is correct. * * When the total number of combinations is small they can all be checked. * For example, if we have 3 segments in the split, and each points to a * 2-way mirror with unique copies, we will have the following pieces of data: * * | mirror child * split | [0] [1] * ======|===================== * A | data_A_0 data_A_1 * B | data_B_0 data_B_1 * C | data_C_0 data_C_1 * * We will try the following (mirror children)^(number of splits) (2^3=8) * combinations, which is similar to bitwise-little-endian counting in * binary. In general each "digit" corresponds to a split segment, and the * base of each digit is is_children, which can be different for each * digit. * * "low bit" "high bit" * v v * data_A_0 data_B_0 data_C_0 * data_A_1 data_B_0 data_C_0 * data_A_0 data_B_1 data_C_0 * data_A_1 data_B_1 data_C_0 * data_A_0 data_B_0 data_C_1 * data_A_1 data_B_0 data_C_1 * data_A_0 data_B_1 data_C_1 * data_A_1 data_B_1 data_C_1 * * Note that the split segments may be on the same or different top-level * vdevs. In either case, we may need to try lots of combinations (see * zfs_reconstruct_indirect_combinations_max). This ensures that if a mirror * has small silent errors on all of its children, we can still reconstruct * the correct data, as long as those errors are at sufficiently-separated * offsets (specifically, separated by the largest block size - default of * 128KB, but up to 16MB). */ static void vdev_indirect_reconstruct_io_done(zio_t *zio) { indirect_vsd_t *iv = zio->io_vsd; boolean_t known_good = B_FALSE; int error; iv->iv_unique_combinations = 1; iv->iv_attempts_max = UINT64_MAX; if (zfs_reconstruct_indirect_combinations_max > 0) iv->iv_attempts_max = zfs_reconstruct_indirect_combinations_max; /* * If nonzero, every 1/x blocks will be damaged, in order to validate * reconstruction when there are split segments with damaged copies. * Known_good will be TRUE when reconstruction is known to be possible. */ if (zfs_reconstruct_indirect_damage_fraction != 0 && random_in_range(zfs_reconstruct_indirect_damage_fraction) == 0) known_good = (vdev_indirect_splits_damage(iv, zio) == 0); /* * Determine the unique children for a split segment and add them * to the is_unique_child list. By restricting reconstruction * to these children, only unique combinations will be considered. * This can vastly reduce the search space when there are a large * number of indirect splits. */ for (indirect_split_t *is = list_head(&iv->iv_splits); is != NULL; is = list_next(&iv->iv_splits, is)) { is->is_unique_children = 0; for (int i = 0; i < is->is_children; i++) { indirect_child_t *ic_i = &is->is_child[i]; if (ic_i->ic_data == NULL || ic_i->ic_duplicate != NULL) continue; for (int j = i + 1; j < is->is_children; j++) { indirect_child_t *ic_j = &is->is_child[j]; if (ic_j->ic_data == NULL || ic_j->ic_duplicate != NULL) continue; if (abd_cmp(ic_i->ic_data, ic_j->ic_data) == 0) ic_j->ic_duplicate = ic_i; } is->is_unique_children++; list_insert_tail(&is->is_unique_child, ic_i); } /* Reconstruction is impossible, no valid children */ EQUIV(list_is_empty(&is->is_unique_child), is->is_unique_children == 0); if (list_is_empty(&is->is_unique_child)) { zio->io_error = EIO; vdev_indirect_all_checksum_errors(zio); zio_checksum_verified(zio); return; } iv->iv_unique_combinations *= is->is_unique_children; } if (iv->iv_unique_combinations <= iv->iv_attempts_max) error = vdev_indirect_splits_enumerate_all(iv, zio); else error = vdev_indirect_splits_enumerate_randomly(iv, zio); if (error != 0) { /* All attempted combinations failed. */ ASSERT3B(known_good, ==, B_FALSE); zio->io_error = error; vdev_indirect_all_checksum_errors(zio); } else { /* * The checksum has been successfully validated. Issue * repair I/Os to any copies of splits which don't match * the validated version. */ ASSERT0(vdev_indirect_splits_checksum_validate(iv, zio)); vdev_indirect_repair(zio); zio_checksum_verified(zio); } } static void vdev_indirect_io_done(zio_t *zio) { indirect_vsd_t *iv = zio->io_vsd; if (iv->iv_reconstruct) { /* * We have read all copies of the data (e.g. from mirrors), * either because this was a scrub/resilver, or because the * one-copy read didn't checksum correctly. */ vdev_indirect_reconstruct_io_done(zio); return; } if (!iv->iv_split_block) { /* * This was not a split block, so we passed the BP down, * and the checksum was handled by the (one) child zio. */ return; } zio_bad_cksum_t zbc; int ret = zio_checksum_error(zio, &zbc); if (ret == 0) { zio_checksum_verified(zio); return; } /* * The checksum didn't match. Read all copies of all splits, and * then we will try to reconstruct. The next time * vdev_indirect_io_done() is called, iv_reconstruct will be set. */ vdev_indirect_read_all(zio); zio_vdev_io_redone(zio); } vdev_ops_t vdev_indirect_ops = { .vdev_op_init = NULL, .vdev_op_fini = NULL, .vdev_op_open = vdev_indirect_open, .vdev_op_close = vdev_indirect_close, .vdev_op_asize = vdev_default_asize, .vdev_op_min_asize = vdev_default_min_asize, .vdev_op_min_alloc = NULL, .vdev_op_io_start = vdev_indirect_io_start, .vdev_op_io_done = vdev_indirect_io_done, .vdev_op_state_change = NULL, .vdev_op_need_resilver = NULL, .vdev_op_hold = NULL, .vdev_op_rele = NULL, .vdev_op_remap = vdev_indirect_remap, .vdev_op_xlate = NULL, .vdev_op_rebuild_asize = NULL, .vdev_op_metaslab_init = NULL, .vdev_op_config_generate = NULL, .vdev_op_nparity = NULL, .vdev_op_ndisks = NULL, .vdev_op_type = VDEV_TYPE_INDIRECT, /* name of this vdev type */ .vdev_op_leaf = B_FALSE /* leaf vdev */ }; EXPORT_SYMBOL(spa_condense_fini); EXPORT_SYMBOL(spa_start_indirect_condensing_thread); EXPORT_SYMBOL(spa_condense_indirect_start_sync); EXPORT_SYMBOL(spa_condense_init); EXPORT_SYMBOL(spa_vdev_indirect_mark_obsolete); EXPORT_SYMBOL(vdev_indirect_mark_obsolete); EXPORT_SYMBOL(vdev_indirect_should_condense); EXPORT_SYMBOL(vdev_indirect_sync_obsolete); EXPORT_SYMBOL(vdev_obsolete_counts_are_precise); EXPORT_SYMBOL(vdev_obsolete_sm_object); /* BEGIN CSTYLED */ ZFS_MODULE_PARAM(zfs_condense, zfs_condense_, indirect_vdevs_enable, INT, ZMOD_RW, "Whether to attempt condensing indirect vdev mappings"); ZFS_MODULE_PARAM(zfs_condense, zfs_condense_, indirect_obsolete_pct, INT, ZMOD_RW, "Minimum obsolete percent of bytes in the mapping to attempt condensing"); ZFS_MODULE_PARAM(zfs_condense, zfs_condense_, min_mapping_bytes, ULONG, ZMOD_RW, "Don't bother condensing if the mapping uses less than this amount of " "memory"); ZFS_MODULE_PARAM(zfs_condense, zfs_condense_, max_obsolete_bytes, ULONG, ZMOD_RW, "Minimum size obsolete spacemap to attempt condensing"); ZFS_MODULE_PARAM(zfs_condense, zfs_condense_, indirect_commit_entry_delay_ms, INT, ZMOD_RW, "Used by tests to ensure certain actions happen in the middle of a " "condense. A maximum value of 1 should be sufficient."); ZFS_MODULE_PARAM(zfs_reconstruct, zfs_reconstruct_, indirect_combinations_max, INT, ZMOD_RW, "Maximum number of combinations when reconstructing split segments"); /* END CSTYLED */ diff --git a/module/zfs/zthr.c b/module/zfs/zthr.c index 5ac2e30467e3..33fdda7b68d1 100644 --- a/module/zfs/zthr.c +++ b/module/zfs/zthr.c @@ -1,536 +1,541 @@ /* * CDDL HEADER START * * This file and its contents are supplied under the terms of the * Common Development and Distribution License ("CDDL"), version 1.0. * You may only use this file in accordance with the terms of version * 1.0 of the CDDL. * * A full copy of the text of the CDDL should have accompanied this * source. A copy of the CDDL is also available via the Internet at * http://www.illumos.org/license/CDDL. * * CDDL HEADER END */ /* * Copyright (c) 2017, 2020 by Delphix. All rights reserved. */ /* * ZTHR Infrastructure * =================== * * ZTHR threads are used for isolated operations that span multiple txgs * within a SPA. They generally exist from SPA creation/loading and until * the SPA is exported/destroyed. The ideal requirements for an operation * to be modeled with a zthr are the following: * * 1] The operation needs to run over multiple txgs. * 2] There is be a single point of reference in memory or on disk that * indicates whether the operation should run/is running or has * stopped. * * If the operation satisfies the above then the following rules guarantee * a certain level of correctness: * * 1] Any thread EXCEPT the zthr changes the work indicator from stopped * to running but not the opposite. * 2] Only the zthr can change the work indicator from running to stopped * (e.g. when it is done) but not the opposite. * * This way a normal zthr cycle should go like this: * * 1] An external thread changes the work indicator from stopped to * running and wakes up the zthr. * 2] The zthr wakes up, checks the indicator and starts working. * 3] When the zthr is done, it changes the indicator to stopped, allowing * a new cycle to start. * * Besides being awakened by other threads, a zthr can be configured * during creation to wakeup on its own after a specified interval * [see zthr_create_timer()]. * * Note: ZTHR threads are NOT a replacement for generic threads! Please * ensure that they fit your use-case well before using them. * * == ZTHR creation * * Every zthr needs four inputs to start running: * * 1] A user-defined checker function (checkfunc) that decides whether * the zthr should start working or go to sleep. The function should * return TRUE when the zthr needs to work or FALSE to let it sleep, * and should adhere to the following signature: * boolean_t checkfunc_name(void *args, zthr_t *t); * * 2] A user-defined ZTHR function (func) which the zthr executes when * it is not sleeping. The function should adhere to the following * signature type: * void func_name(void *args, zthr_t *t); * * 3] A void args pointer that will be passed to checkfunc and func * implicitly by the infrastructure. * * 4] A name for the thread. This string must be valid for the lifetime * of the zthr. * * The reason why the above API needs two different functions, * instead of one that both checks and does the work, has to do with * the zthr's internal state lock (zthr_state_lock) and the allowed * cancellation windows. We want to hold the zthr_state_lock while * running checkfunc but not while running func. This way the zthr * can be cancelled while doing work and not while checking for work. * * To start a zthr: - * zthr_t *zthr_pointer = zthr_create(checkfunc, func, args); + * zthr_t *zthr_pointer = zthr_create(checkfunc, func, args, + * pri); * or * zthr_t *zthr_pointer = zthr_create_timer(checkfunc, func, - * args, max_sleep); + * args, max_sleep, pri); * * After that you should be able to wakeup, cancel, and resume the * zthr from another thread using the zthr_pointer. * * NOTE: ZTHR threads could potentially wake up spuriously and the * user should take this into account when writing a checkfunc. * [see ZTHR state transitions] * * == ZTHR wakeup * * ZTHR wakeup should be used when new work is added for the zthr. The * sleeping zthr will wakeup, see that it has more work to complete * and proceed. This can be invoked from open or syncing context. * * To wakeup a zthr: * zthr_wakeup(zthr_t *t) * * == ZTHR cancellation and resumption * * ZTHR threads must be cancelled when their SPA is being exported * or when they need to be paused so they don't interfere with other * operations. * * To cancel a zthr: * zthr_cancel(zthr_pointer); * * To resume it: * zthr_resume(zthr_pointer); * * ZTHR cancel and resume should be invoked in open context during the * lifecycle of the pool as it is imported, exported or destroyed. * * A zthr will implicitly check if it has received a cancellation * signal every time func returns and every time it wakes up [see * ZTHR state transitions below]. * * At times, waiting for the zthr's func to finish its job may take * time. This may be very time-consuming for some operations that * need to cancel the SPA's zthrs (e.g spa_export). For this scenario * the user can explicitly make their ZTHR function aware of incoming * cancellation signals using zthr_iscancelled(). A common pattern for * that looks like this: * * int * func_name(void *args, zthr_t *t) * { * ... ... * while (!work_done && !zthr_iscancelled(t)) { * ... ... * } * } * * == ZTHR cleanup * * Cancelling a zthr doesn't clean up its metadata (internal locks, * function pointers to func and checkfunc, etc..). This is because * we want to keep them around in case we want to resume the execution * of the zthr later. Similarly for zthrs that exit themselves. * * To completely cleanup a zthr, cancel it first to ensure that it * is not running and then use zthr_destroy(). * * == ZTHR state transitions * * zthr creation * + * | * | woke up * | +--------------+ sleep * | | ^ * | | | * | | | FALSE * | | | * v v FALSE + * cancelled? +---------> checkfunc? * + ^ + * | | | * | | | TRUE * | | | * | | func returned v * | +---------------+ func * | * | TRUE * | * v * zthr stopped running * * == Implementation of ZTHR requests * * ZTHR cancel and resume are requests on a zthr to change its * internal state. These requests are serialized using the * zthr_request_lock, while changes in its internal state are * protected by the zthr_state_lock. A request will first acquire * the zthr_request_lock and then immediately acquire the * zthr_state_lock. We do this so that incoming requests are * serialized using the request lock, while still allowing us * to use the state lock for thread communication via zthr_cv. * * ZTHR wakeup broadcasts to zthr_cv, causing sleeping threads * to wakeup. It acquires the zthr_state_lock but not the * zthr_request_lock, so that a wakeup on a zthr in the middle * of being cancelled will not block. */ #include #include struct zthr { /* running thread doing the work */ kthread_t *zthr_thread; /* lock protecting internal data & invariants */ kmutex_t zthr_state_lock; /* mutex that serializes external requests */ kmutex_t zthr_request_lock; /* notification mechanism for requests */ kcondvar_t zthr_cv; /* flag set to true if we are canceling the zthr */ boolean_t zthr_cancel; /* flag set to true if we are waiting for the zthr to finish */ boolean_t zthr_haswaiters; kcondvar_t zthr_wait_cv; /* * maximum amount of time that the zthr is spent sleeping; * if this is 0, the thread doesn't wake up until it gets * signaled. */ hrtime_t zthr_sleep_timeout; + /* Thread priority */ + pri_t zthr_pri; + /* consumer-provided callbacks & data */ zthr_checkfunc_t *zthr_checkfunc; zthr_func_t *zthr_func; void *zthr_arg; const char *zthr_name; }; static void zthr_procedure(void *arg) { zthr_t *t = arg; mutex_enter(&t->zthr_state_lock); ASSERT3P(t->zthr_thread, ==, curthread); while (!t->zthr_cancel) { if (t->zthr_checkfunc(t->zthr_arg, t)) { mutex_exit(&t->zthr_state_lock); t->zthr_func(t->zthr_arg, t); mutex_enter(&t->zthr_state_lock); } else { if (t->zthr_sleep_timeout == 0) { cv_wait_idle(&t->zthr_cv, &t->zthr_state_lock); } else { (void) cv_timedwait_idle_hires(&t->zthr_cv, &t->zthr_state_lock, t->zthr_sleep_timeout, MSEC2NSEC(1), 0); } } if (t->zthr_haswaiters) { t->zthr_haswaiters = B_FALSE; cv_broadcast(&t->zthr_wait_cv); } } /* * Clear out the kernel thread metadata and notify the * zthr_cancel() thread that we've stopped running. */ t->zthr_thread = NULL; t->zthr_cancel = B_FALSE; cv_broadcast(&t->zthr_cv); mutex_exit(&t->zthr_state_lock); thread_exit(); } zthr_t * zthr_create(const char *zthr_name, zthr_checkfunc_t *checkfunc, - zthr_func_t *func, void *arg) + zthr_func_t *func, void *arg, pri_t pri) { return (zthr_create_timer(zthr_name, checkfunc, - func, arg, (hrtime_t)0)); + func, arg, (hrtime_t)0, pri)); } /* * Create a zthr with specified maximum sleep time. If the time * in sleeping state exceeds max_sleep, a wakeup(do the check and * start working if required) will be triggered. */ zthr_t * zthr_create_timer(const char *zthr_name, zthr_checkfunc_t *checkfunc, - zthr_func_t *func, void *arg, hrtime_t max_sleep) + zthr_func_t *func, void *arg, hrtime_t max_sleep, pri_t pri) { zthr_t *t = kmem_zalloc(sizeof (*t), KM_SLEEP); mutex_init(&t->zthr_state_lock, NULL, MUTEX_DEFAULT, NULL); mutex_init(&t->zthr_request_lock, NULL, MUTEX_DEFAULT, NULL); cv_init(&t->zthr_cv, NULL, CV_DEFAULT, NULL); cv_init(&t->zthr_wait_cv, NULL, CV_DEFAULT, NULL); mutex_enter(&t->zthr_state_lock); t->zthr_checkfunc = checkfunc; t->zthr_func = func; t->zthr_arg = arg; t->zthr_sleep_timeout = max_sleep; t->zthr_name = zthr_name; + t->zthr_pri = pri; t->zthr_thread = thread_create_named(zthr_name, NULL, 0, - zthr_procedure, t, 0, &p0, TS_RUN, minclsyspri); + zthr_procedure, t, 0, &p0, TS_RUN, pri); mutex_exit(&t->zthr_state_lock); return (t); } void zthr_destroy(zthr_t *t) { ASSERT(!MUTEX_HELD(&t->zthr_state_lock)); ASSERT(!MUTEX_HELD(&t->zthr_request_lock)); VERIFY3P(t->zthr_thread, ==, NULL); mutex_destroy(&t->zthr_request_lock); mutex_destroy(&t->zthr_state_lock); cv_destroy(&t->zthr_cv); cv_destroy(&t->zthr_wait_cv); kmem_free(t, sizeof (*t)); } /* * Wake up the zthr if it is sleeping. If the thread has been cancelled * or is in the process of being cancelled, this is a no-op. */ void zthr_wakeup(zthr_t *t) { mutex_enter(&t->zthr_state_lock); /* * There are 5 states that we can find the zthr when issuing * this broadcast: * * [1] The common case of the thread being asleep, at which * point the broadcast will wake it up. * [2] The thread has been cancelled. Waking up a cancelled * thread is a no-op. Any work that is still left to be * done should be handled the next time the thread is * resumed. * [3] The thread is doing work and is already up, so this * is basically a no-op. * [4] The thread was just created/resumed, in which case the * behavior is similar to [3]. * [5] The thread is in the middle of being cancelled, which * will be a no-op. */ cv_broadcast(&t->zthr_cv); mutex_exit(&t->zthr_state_lock); } /* * Sends a cancel request to the zthr and blocks until the zthr is * cancelled. If the zthr is not running (e.g. has been cancelled * already), this is a no-op. Note that this function should not be * called from syncing context as it could deadlock with the zthr_func. */ void zthr_cancel(zthr_t *t) { mutex_enter(&t->zthr_request_lock); mutex_enter(&t->zthr_state_lock); /* * Since we are holding the zthr_state_lock at this point * we can find the state in one of the following 4 states: * * [1] The thread has already been cancelled, therefore * there is nothing for us to do. * [2] The thread is sleeping so we set the flag, broadcast * the CV and wait for it to exit. * [3] The thread is doing work, in which case we just set * the flag and wait for it to finish. * [4] The thread was just created/resumed, in which case * the behavior is similar to [3]. * * Since requests are serialized, by the time that we get * control back we expect that the zthr is cancelled and * not running anymore. */ if (t->zthr_thread != NULL) { t->zthr_cancel = B_TRUE; /* broadcast in case the zthr is sleeping */ cv_broadcast(&t->zthr_cv); while (t->zthr_thread != NULL) cv_wait(&t->zthr_cv, &t->zthr_state_lock); ASSERT(!t->zthr_cancel); } mutex_exit(&t->zthr_state_lock); mutex_exit(&t->zthr_request_lock); } /* * Sends a resume request to the supplied zthr. If the zthr is already * running this is a no-op. Note that this function should not be * called from syncing context as it could deadlock with the zthr_func. */ void zthr_resume(zthr_t *t) { mutex_enter(&t->zthr_request_lock); mutex_enter(&t->zthr_state_lock); ASSERT3P(&t->zthr_checkfunc, !=, NULL); ASSERT3P(&t->zthr_func, !=, NULL); ASSERT(!t->zthr_cancel); ASSERT(!t->zthr_haswaiters); /* * There are 4 states that we find the zthr in at this point * given the locks that we hold: * * [1] The zthr was cancelled, so we spawn a new thread for * the zthr (common case). * [2] The zthr is running at which point this is a no-op. * [3] The zthr is sleeping at which point this is a no-op. * [4] The zthr was just spawned at which point this is a * no-op. */ if (t->zthr_thread == NULL) { t->zthr_thread = thread_create_named(t->zthr_name, NULL, 0, - zthr_procedure, t, 0, &p0, TS_RUN, minclsyspri); + zthr_procedure, t, 0, &p0, TS_RUN, t->zthr_pri); } mutex_exit(&t->zthr_state_lock); mutex_exit(&t->zthr_request_lock); } /* * This function is intended to be used by the zthr itself * (specifically the zthr_func callback provided) to check * if another thread has signaled it to stop running before * doing some expensive operation. * * returns TRUE if we are in the middle of trying to cancel * this thread. * * returns FALSE otherwise. */ boolean_t zthr_iscancelled(zthr_t *t) { ASSERT3P(t->zthr_thread, ==, curthread); /* * The majority of the functions here grab zthr_request_lock * first and then zthr_state_lock. This function only grabs * the zthr_state_lock. That is because this function should * only be called from the zthr_func to check if someone has * issued a zthr_cancel() on the thread. If there is a zthr_cancel() * happening concurrently, attempting to grab the request lock * here would result in a deadlock. * * By grabbing only the zthr_state_lock this function is allowed * to run concurrently with a zthr_cancel() request. */ mutex_enter(&t->zthr_state_lock); boolean_t cancelled = t->zthr_cancel; mutex_exit(&t->zthr_state_lock); return (cancelled); } /* * Wait for the zthr to finish its current function. Similar to * zthr_iscancelled, you can use zthr_has_waiters to have the zthr_func end * early. Unlike zthr_cancel, the thread is not destroyed. If the zthr was * sleeping or cancelled, return immediately. */ void zthr_wait_cycle_done(zthr_t *t) { mutex_enter(&t->zthr_state_lock); /* * Since we are holding the zthr_state_lock at this point * we can find the state in one of the following 5 states: * * [1] The thread has already cancelled, therefore * there is nothing for us to do. * [2] The thread is sleeping so we set the flag, broadcast * the CV and wait for it to exit. * [3] The thread is doing work, in which case we just set * the flag and wait for it to finish. * [4] The thread was just created/resumed, in which case * the behavior is similar to [3]. * [5] The thread is the middle of being cancelled, which is * similar to [3]. We'll wait for the cancel, which is * waiting for the zthr func. * * Since requests are serialized, by the time that we get * control back we expect that the zthr has completed it's * zthr_func. */ if (t->zthr_thread != NULL) { t->zthr_haswaiters = B_TRUE; /* broadcast in case the zthr is sleeping */ cv_broadcast(&t->zthr_cv); while ((t->zthr_haswaiters) && (t->zthr_thread != NULL)) cv_wait(&t->zthr_wait_cv, &t->zthr_state_lock); ASSERT(!t->zthr_haswaiters); } mutex_exit(&t->zthr_state_lock); } /* * This function is intended to be used by the zthr itself * to check if another thread is waiting on it to finish * * returns TRUE if we have been asked to finish. * * returns FALSE otherwise. */ boolean_t zthr_has_waiters(zthr_t *t) { ASSERT3P(t->zthr_thread, ==, curthread); mutex_enter(&t->zthr_state_lock); /* * Similarly to zthr_iscancelled(), we only grab the * zthr_state_lock so that the zthr itself can use this * to check for the request. */ boolean_t has_waiters = t->zthr_haswaiters; mutex_exit(&t->zthr_state_lock); return (has_waiters); }