Index: head/sys/kern/vfs_bio.c =================================================================== --- head/sys/kern/vfs_bio.c (revision 157318) +++ head/sys/kern/vfs_bio.c (revision 157319) @@ -1,3900 +1,3919 @@ /*- * Copyright (c) 2004 Poul-Henning Kamp * Copyright (c) 1994,1997 John S. Dyson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * this file contains a new buffer I/O scheme implementing a coherent * VM object and buffer cache scheme. Pains have been taken to make * sure that the performance degradation associated with schemes such * as this is not realized. * * Author: John S. Dyson * Significant help during the development and debugging phases * had been provided by David Greenman, also of the FreeBSD core team. * * see man buf(9) for more info. */ #include __FBSDID("$FreeBSD$"); #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 "opt_directio.h" #include "opt_swap.h" static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); struct bio_ops bioops; /* I/O operation notification */ struct buf_ops buf_ops_bio = { .bop_name = "buf_ops_bio", .bop_write = bufwrite, .bop_strategy = bufstrategy, .bop_sync = bufsync, }; /* * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. */ struct buf *buf; /* buffer header pool */ static struct proc *bufdaemonproc; static int inmem(struct vnode *vp, daddr_t blkno); static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to); static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to); static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m); static void vfs_clean_pages(struct buf *bp); static void vfs_setdirty(struct buf *bp); static void vfs_vmio_release(struct buf *bp); static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno); -static int flushbufqueues(int flushdeps); +static int flushbufqueues(int, int); static void buf_daemon(void); static void bremfreel(struct buf *bp); int vmiodirenable = TRUE; SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, "Use the VM system for directory writes"); int runningbufspace; SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, "Amount of presently outstanding async buffer io"); static int bufspace; SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, "KVA memory used for bufs"); static int maxbufspace; SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, "Maximum allowed value of bufspace (including buf_daemon)"); static int bufmallocspace; SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, "Amount of malloced memory for buffers"); static int maxbufmallocspace; SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, "Maximum amount of malloced memory for buffers"); static int lobufspace; SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, "Minimum amount of buffers we want to have"); int hibufspace; SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, "Maximum allowed value of bufspace (excluding buf_daemon)"); static int bufreusecnt; SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, "Number of times we have reused a buffer"); static int buffreekvacnt; SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, "Number of times we have freed the KVA space from some buffer"); static int bufdefragcnt; SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, "Number of times we have had to repeat buffer allocation to defragment"); static int lorunningspace; SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, "Minimum preferred space used for in-progress I/O"); static int hirunningspace; SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, "Maximum amount of space to use for in-progress I/O"); static int dirtybufferflushes; SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); static int altbufferflushes; SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); static int recursiveflushes; SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); static int numdirtybuffers; SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, "Number of buffers that are dirty (has unwritten changes) at the moment"); static int lodirtybuffers; SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, "How many buffers we want to have free before bufdaemon can sleep"); static int hidirtybuffers; SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, "When the number of dirty buffers is considered severe"); static int dirtybufthresh; SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); static int numfreebuffers; SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, "Number of free buffers"); static int lofreebuffers; SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, "XXX Unused"); static int hifreebuffers; SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, "XXX Complicatedly unused"); static int getnewbufcalls; SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, "Number of calls to getnewbuf"); static int getnewbufrestarts; SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, "Number of times getnewbuf has had to restart a buffer aquisition"); /* * Wakeup point for bufdaemon, as well as indicator of whether it is already * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it * is idling. */ static int bd_request; /* * This lock synchronizes access to bd_request. */ static struct mtx bdlock; /* * bogus page -- for I/O to/from partially complete buffers * this is a temporary solution to the problem, but it is not * really that bad. it would be better to split the buffer * for input in the case of buffers partially already in memory, * but the code is intricate enough already. */ vm_page_t bogus_page; /* * Synchronization (sleep/wakeup) variable for active buffer space requests. * Set when wait starts, cleared prior to wakeup(). * Used in runningbufwakeup() and waitrunningbufspace(). */ static int runningbufreq; /* * This lock protects the runningbufreq and synchronizes runningbufwakeup and * waitrunningbufspace(). */ static struct mtx rbreqlock; /* * Synchronization (sleep/wakeup) variable for buffer requests. * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done * by and/or. * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), * getnewbuf(), and getblk(). */ static int needsbuffer; /* * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. */ static struct mtx nblock; /* * Lock that protects against bwait()/bdone()/B_DONE races. */ static struct mtx bdonelock; /* * Lock that protects against bwait()/bdone()/B_DONE races. */ static struct mtx bpinlock; /* * Definitions for the buffer free lists. */ -#define BUFFER_QUEUES 5 /* number of free buffer queues */ +#define BUFFER_QUEUES 6 /* number of free buffer queues */ #define QUEUE_NONE 0 /* on no queue */ #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ -#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ -#define QUEUE_EMPTY 4 /* empty buffer headers */ +#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ +#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ +#define QUEUE_EMPTY 5 /* empty buffer headers */ /* Queues for free buffers with various properties */ static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; /* Lock for the bufqueues */ static struct mtx bqlock; /* * Single global constant for BUF_WMESG, to avoid getting multiple references. * buf_wmesg is referred from macros. */ const char *buf_wmesg = BUF_WMESG; #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ #ifdef DIRECTIO extern void ffs_rawread_setup(void); #endif /* DIRECTIO */ /* * numdirtywakeup: * * If someone is blocked due to there being too many dirty buffers, * and numdirtybuffers is now reasonable, wake them up. */ static __inline void numdirtywakeup(int level) { if (numdirtybuffers <= level) { mtx_lock(&nblock); if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; wakeup(&needsbuffer); } mtx_unlock(&nblock); } } /* * bufspacewakeup: * * Called when buffer space is potentially available for recovery. * getnewbuf() will block on this flag when it is unable to free * sufficient buffer space. Buffer space becomes recoverable when * bp's get placed back in the queues. */ static __inline void bufspacewakeup(void) { /* * If someone is waiting for BUF space, wake them up. Even * though we haven't freed the kva space yet, the waiting * process will be able to now. */ mtx_lock(&nblock); if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; wakeup(&needsbuffer); } mtx_unlock(&nblock); } /* * runningbufwakeup() - in-progress I/O accounting. * */ void runningbufwakeup(struct buf *bp) { if (bp->b_runningbufspace) { atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); bp->b_runningbufspace = 0; mtx_lock(&rbreqlock); if (runningbufreq && runningbufspace <= lorunningspace) { runningbufreq = 0; wakeup(&runningbufreq); } mtx_unlock(&rbreqlock); } } /* * bufcountwakeup: * * Called when a buffer has been added to one of the free queues to * account for the buffer and to wakeup anyone waiting for free buffers. * This typically occurs when large amounts of metadata are being handled * by the buffer cache ( else buffer space runs out first, usually ). */ static __inline void bufcountwakeup(void) { atomic_add_int(&numfreebuffers, 1); mtx_lock(&nblock); if (needsbuffer) { needsbuffer &= ~VFS_BIO_NEED_ANY; if (numfreebuffers >= hifreebuffers) needsbuffer &= ~VFS_BIO_NEED_FREE; wakeup(&needsbuffer); } mtx_unlock(&nblock); } /* * waitrunningbufspace() * * runningbufspace is a measure of the amount of I/O currently * running. This routine is used in async-write situations to * prevent creating huge backups of pending writes to a device. * Only asynchronous writes are governed by this function. * * Reads will adjust runningbufspace, but will not block based on it. * The read load has a side effect of reducing the allowed write load. * * This does NOT turn an async write into a sync write. It waits * for earlier writes to complete and generally returns before the * caller's write has reached the device. */ void waitrunningbufspace(void) { mtx_lock(&rbreqlock); while (runningbufspace > hirunningspace) { ++runningbufreq; msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); } mtx_unlock(&rbreqlock); } /* * vfs_buf_test_cache: * * Called when a buffer is extended. This function clears the B_CACHE * bit if the newly extended portion of the buffer does not contain * valid data. */ static __inline void vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, vm_page_t m) { VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); if (bp->b_flags & B_CACHE) { int base = (foff + off) & PAGE_MASK; if (vm_page_is_valid(m, base, size) == 0) bp->b_flags &= ~B_CACHE; } } /* Wake up the buffer deamon if necessary */ static __inline void bd_wakeup(int dirtybuflevel) { mtx_lock(&bdlock); if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { bd_request = 1; wakeup(&bd_request); } mtx_unlock(&bdlock); } /* * bd_speedup - speedup the buffer cache flushing code */ static __inline void bd_speedup(void) { bd_wakeup(1); } /* * Calculating buffer cache scaling values and reserve space for buffer * headers. This is called during low level kernel initialization and * may be called more then once. We CANNOT write to the memory area * being reserved at this time. */ caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) { /* * physmem_est is in pages. Convert it to kilobytes (assumes * PAGE_SIZE is >= 1K) */ physmem_est = physmem_est * (PAGE_SIZE / 1024); /* * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. * For the first 64MB of ram nominally allocate sufficient buffers to * cover 1/4 of our ram. Beyond the first 64MB allocate additional * buffers to cover 1/20 of our ram over 64MB. When auto-sizing * the buffer cache we limit the eventual kva reservation to * maxbcache bytes. * * factor represents the 1/4 x ram conversion. */ if (nbuf == 0) { int factor = 4 * BKVASIZE / 1024; nbuf = 50; if (physmem_est > 4096) nbuf += min((physmem_est - 4096) / factor, 65536 / factor); if (physmem_est > 65536) nbuf += (physmem_est - 65536) * 2 / (factor * 5); if (maxbcache && nbuf > maxbcache / BKVASIZE) nbuf = maxbcache / BKVASIZE; } #if 0 /* * Do not allow the buffer_map to be more then 1/2 the size of the * kernel_map. */ if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / (BKVASIZE * 2)) { nbuf = (kernel_map->max_offset - kernel_map->min_offset) / (BKVASIZE * 2); printf("Warning: nbufs capped at %d\n", nbuf); } #endif /* * swbufs are used as temporary holders for I/O, such as paging I/O. * We have no less then 16 and no more then 256. */ nswbuf = max(min(nbuf/4, 256), 16); #ifdef NSWBUF_MIN if (nswbuf < NSWBUF_MIN) nswbuf = NSWBUF_MIN; #endif #ifdef DIRECTIO ffs_rawread_setup(); #endif /* * Reserve space for the buffer cache buffers */ swbuf = (void *)v; v = (caddr_t)(swbuf + nswbuf); buf = (void *)v; v = (caddr_t)(buf + nbuf); return(v); } /* Initialize the buffer subsystem. Called before use of any buffers. */ void bufinit(void) { struct buf *bp; int i; mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF); /* next, make a null set of free lists */ for (i = 0; i < BUFFER_QUEUES; i++) TAILQ_INIT(&bufqueues[i]); /* finally, initialize each buffer header and stick on empty q */ for (i = 0; i < nbuf; i++) { bp = &buf[i]; bzero(bp, sizeof *bp); bp->b_flags = B_INVAL; /* we're just an empty header */ bp->b_rcred = NOCRED; bp->b_wcred = NOCRED; bp->b_qindex = QUEUE_EMPTY; bp->b_vflags = 0; bp->b_xflags = 0; LIST_INIT(&bp->b_dep); BUF_LOCKINIT(bp); TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); } /* * maxbufspace is the absolute maximum amount of buffer space we are * allowed to reserve in KVM and in real terms. The absolute maximum * is nominally used by buf_daemon. hibufspace is the nominal maximum * used by most other processes. The differential is required to * ensure that buf_daemon is able to run when other processes might * be blocked waiting for buffer space. * * maxbufspace is based on BKVASIZE. Allocating buffers larger then * this may result in KVM fragmentation which is not handled optimally * by the system. */ maxbufspace = nbuf * BKVASIZE; hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); lobufspace = hibufspace - MAXBSIZE; lorunningspace = 512 * 1024; hirunningspace = 1024 * 1024; /* * Limit the amount of malloc memory since it is wired permanently into * the kernel space. Even though this is accounted for in the buffer * allocation, we don't want the malloced region to grow uncontrolled. * The malloc scheme improves memory utilization significantly on average * (small) directories. */ maxbufmallocspace = hibufspace / 20; /* * Reduce the chance of a deadlock occuring by limiting the number * of delayed-write dirty buffers we allow to stack up. */ hidirtybuffers = nbuf / 4 + 20; dirtybufthresh = hidirtybuffers * 9 / 10; numdirtybuffers = 0; /* * To support extreme low-memory systems, make sure hidirtybuffers cannot * eat up all available buffer space. This occurs when our minimum cannot * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming * BKVASIZE'd (8K) buffers. */ while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { hidirtybuffers >>= 1; } lodirtybuffers = hidirtybuffers / 2; /* * Try to keep the number of free buffers in the specified range, * and give special processes (e.g. like buf_daemon) access to an * emergency reserve. */ lofreebuffers = nbuf / 18 + 5; hifreebuffers = 2 * lofreebuffers; numfreebuffers = nbuf; /* * Maximum number of async ops initiated per buf_daemon loop. This is * somewhat of a hack at the moment, we really need to limit ourselves * based on the number of bytes of I/O in-transit that were initiated * from buf_daemon. */ bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED); } /* * bfreekva() - free the kva allocation for a buffer. * * Since this call frees up buffer space, we call bufspacewakeup(). */ static void bfreekva(struct buf *bp) { if (bp->b_kvasize) { atomic_add_int(&buffreekvacnt, 1); atomic_subtract_int(&bufspace, bp->b_kvasize); vm_map_lock(buffer_map); vm_map_delete(buffer_map, (vm_offset_t) bp->b_kvabase, (vm_offset_t) bp->b_kvabase + bp->b_kvasize ); vm_map_unlock(buffer_map); bp->b_kvasize = 0; bufspacewakeup(); } } /* * bremfree: * * Mark the buffer for removal from the appropriate free list in brelse. * */ void bremfree(struct buf *bp) { CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); KASSERT((bp->b_flags & B_REMFREE) == 0, ("bremfree: buffer %p already marked for delayed removal.", bp)); KASSERT(bp->b_qindex != QUEUE_NONE, ("bremfree: buffer %p not on a queue.", bp)); bp->b_flags |= B_REMFREE; /* Fixup numfreebuffers count. */ if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) atomic_subtract_int(&numfreebuffers, 1); } /* * bremfreef: * * Force an immediate removal from a free list. Used only in nfs when * it abuses the b_freelist pointer. */ void bremfreef(struct buf *bp) { mtx_lock(&bqlock); bremfreel(bp); mtx_unlock(&bqlock); } /* * bremfreel: * * Removes a buffer from the free list, must be called with the * bqlock held. */ static void bremfreel(struct buf *bp) { CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp)); KASSERT(bp->b_qindex != QUEUE_NONE, ("bremfreel: buffer %p not on a queue.", bp)); mtx_assert(&bqlock, MA_OWNED); TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); bp->b_qindex = QUEUE_NONE; /* * If this was a delayed bremfree() we only need to remove the buffer * from the queue and return the stats are already done. */ if (bp->b_flags & B_REMFREE) { bp->b_flags &= ~B_REMFREE; return; } /* * Fixup numfreebuffers count. If the buffer is invalid or not * delayed-write, the buffer was free and we must decrement * numfreebuffers. */ if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) atomic_subtract_int(&numfreebuffers, 1); } /* * Get a buffer with the specified data. Look in the cache first. We * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE * is set, the buffer is valid and we do not have to do anything ( see * getblk() ). This is really just a special case of breadn(). */ int bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, struct buf **bpp) { return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. We must * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, * the buffer is valid and we do not have to do anything. */ void breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, struct ucred * cred) { struct buf *rabp; int i; for (i = 0; i < cnt; i++, rablkno++, rabsize++) { if (inmem(vp, *rablkno)) continue; rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); if ((rabp->b_flags & B_CACHE) == 0) { if (curthread != PCPU_GET(idlethread)) curthread->td_proc->p_stats->p_ru.ru_inblock++; rabp->b_flags |= B_ASYNC; rabp->b_flags &= ~B_INVAL; rabp->b_ioflags &= ~BIO_ERROR; rabp->b_iocmd = BIO_READ; if (rabp->b_rcred == NOCRED && cred != NOCRED) rabp->b_rcred = crhold(cred); vfs_busy_pages(rabp, 0); BUF_KERNPROC(rabp); rabp->b_iooffset = dbtob(rabp->b_blkno); bstrategy(rabp); } else { brelse(rabp); } } } /* * Operates like bread, but also starts asynchronous I/O on * read-ahead blocks. */ int breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno, int *rabsize, int cnt, struct ucred * cred, struct buf **bpp) { struct buf *bp; int rv = 0, readwait = 0; CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); /* if not found in cache, do some I/O */ if ((bp->b_flags & B_CACHE) == 0) { if (curthread != PCPU_GET(idlethread)) curthread->td_proc->p_stats->p_ru.ru_inblock++; bp->b_iocmd = BIO_READ; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); vfs_busy_pages(bp, 0); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); ++readwait; } breada(vp, rablkno, rabsize, cnt, cred); if (readwait) { rv = bufwait(bp); } return (rv); } /* * Write, release buffer on completion. (Done by iodone * if async). Do not bother writing anything if the buffer * is invalid. * * Note that we set B_CACHE here, indicating that buffer is * fully valid and thus cacheable. This is true even of NFS * now so we set it generally. This could be set either here * or in biodone() since the I/O is synchronous. We put it * here. */ int bufwrite(struct buf *bp) { int oldflags; CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if (bp->b_flags & B_INVAL) { brelse(bp); return (0); } oldflags = bp->b_flags; if (BUF_REFCNT(bp) == 0) panic("bufwrite: buffer is not busy???"); if (bp->b_pin_count > 0) bunpin_wait(bp); KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), ("FFS background buffer should not get here %p", bp)); /* Mark the buffer clean */ bundirty(bp); bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_flags |= B_CACHE; bp->b_iocmd = BIO_WRITE; bufobj_wref(bp->b_bufobj); vfs_busy_pages(bp, 1); /* * Normal bwrites pipeline writes */ bp->b_runningbufspace = bp->b_bufsize; atomic_add_int(&runningbufspace, bp->b_runningbufspace); if (curthread != PCPU_GET(idlethread)) curthread->td_proc->p_stats->p_ru.ru_oublock++; if (oldflags & B_ASYNC) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); if ((oldflags & B_ASYNC) == 0) { int rtval = bufwait(bp); brelse(bp); return (rtval); } else { /* * don't allow the async write to saturate the I/O * system. We will not deadlock here because * we are blocking waiting for I/O that is already in-progress * to complete. We do not block here if it is the update * or syncer daemon trying to clean up as that can lead * to deadlock. */ if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0) waitrunningbufspace(); } return (0); } /* * Delayed write. (Buffer is marked dirty). Do not bother writing * anything if the buffer is marked invalid. * * Note that since the buffer must be completely valid, we can safely * set B_CACHE. In fact, we have to set B_CACHE here rather then in * biodone() in order to prevent getblk from writing the buffer * out synchronously. */ void bdwrite(struct buf *bp) { struct thread *td = curthread; struct vnode *vp; struct buf *nbp; struct bufobj *bo; CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); if (bp->b_flags & B_INVAL) { brelse(bp); return; } /* * If we have too many dirty buffers, don't create any more. * If we are wildly over our limit, then force a complete * cleanup. Otherwise, just keep the situation from getting * out of control. Note that we have to avoid a recursive * disaster and not try to clean up after our own cleanup! */ vp = bp->b_vp; bo = bp->b_bufobj; if ((td->td_pflags & TDP_COWINPROGRESS) == 0) { BO_LOCK(bo); if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { BO_UNLOCK(bo); (void) VOP_FSYNC(vp, MNT_NOWAIT, td); altbufferflushes++; } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { /* * Try to find a buffer to flush. */ TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { if ((nbp->b_vflags & BV_BKGRDINPROG) || BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) continue; if (bp == nbp) panic("bdwrite: found ourselves"); BO_UNLOCK(bo); /* Don't countdeps with the bo lock held. */ if (buf_countdeps(nbp, 0)) { BO_LOCK(bo); BUF_UNLOCK(nbp); continue; } if (nbp->b_flags & B_CLUSTEROK) { vfs_bio_awrite(nbp); } else { bremfree(nbp); bawrite(nbp); } dirtybufferflushes++; break; } if (nbp == NULL) BO_UNLOCK(bo); } else BO_UNLOCK(bo); } else recursiveflushes++; bdirty(bp); /* * Set B_CACHE, indicating that the buffer is fully valid. This is * true even of NFS now. */ bp->b_flags |= B_CACHE; /* * This bmap keeps the system from needing to do the bmap later, * perhaps when the system is attempting to do a sync. Since it * is likely that the indirect block -- or whatever other datastructure * that the filesystem needs is still in memory now, it is a good * thing to do this. Note also, that if the pageout daemon is * requesting a sync -- there might not be enough memory to do * the bmap then... So, this is important to do. */ if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } /* * Set the *dirty* buffer range based upon the VM system dirty pages. */ vfs_setdirty(bp); /* * We need to do this here to satisfy the vnode_pager and the * pageout daemon, so that it thinks that the pages have been * "cleaned". Note that since the pages are in a delayed write * buffer -- the VFS layer "will" see that the pages get written * out on the next sync, or perhaps the cluster will be completed. */ vfs_clean_pages(bp); bqrelse(bp); /* * Wakeup the buffer flushing daemon if we have a lot of dirty * buffers (midpoint between our recovery point and our stall * point). */ bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); /* * note: we cannot initiate I/O from a bdwrite even if we wanted to, * due to the softdep code. */ } /* * bdirty: * * Turn buffer into delayed write request. We must clear BIO_READ and * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to * itself to properly update it in the dirty/clean lists. We mark it * B_DONE to ensure that any asynchronization of the buffer properly * clears B_DONE ( else a panic will occur later ). * * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() * should only be called if the buffer is known-good. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bdirty(struct buf *bp) { CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); bp->b_flags &= ~(B_RELBUF); bp->b_iocmd = BIO_WRITE; if ((bp->b_flags & B_DELWRI) == 0) { bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; reassignbuf(bp); atomic_add_int(&numdirtybuffers, 1); bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); } } /* * bundirty: * * Clear B_DELWRI for buffer. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bundirty(struct buf *bp) { CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); if (bp->b_flags & B_DELWRI) { bp->b_flags &= ~B_DELWRI; reassignbuf(bp); atomic_subtract_int(&numdirtybuffers, 1); numdirtywakeup(lodirtybuffers); } /* * Since it is now being written, we can clear its deferred write flag. */ bp->b_flags &= ~B_DEFERRED; } /* * bawrite: * * Asynchronous write. Start output on a buffer, but do not wait for * it to complete. The buffer is released when the output completes. * * bwrite() ( or the VOP routine anyway ) is responsible for handling * B_INVAL buffers. Not us. */ void bawrite(struct buf *bp) { bp->b_flags |= B_ASYNC; (void) bwrite(bp); } /* * bwillwrite: * * Called prior to the locking of any vnodes when we are expecting to * write. We do not want to starve the buffer cache with too many * dirty buffers so we block here. By blocking prior to the locking * of any vnodes we attempt to avoid the situation where a locked vnode * prevents the various system daemons from flushing related buffers. */ void bwillwrite(void) { if (numdirtybuffers >= hidirtybuffers) { mtx_lock(&nblock); while (numdirtybuffers >= hidirtybuffers) { bd_wakeup(1); needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; msleep(&needsbuffer, &nblock, (PRIBIO + 4), "flswai", 0); } mtx_unlock(&nblock); } } /* * Return true if we have too many dirty buffers. */ int buf_dirty_count_severe(void) { return(numdirtybuffers >= hidirtybuffers); } /* * brelse: * * Release a busy buffer and, if requested, free its resources. The * buffer will be stashed in the appropriate bufqueue[] allowing it * to be accessed later as a cache entity or reused for other purposes. */ void brelse(struct buf *bp) { CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); if (bp->b_flags & B_MANAGED) { bqrelse(bp); return; } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && !(bp->b_flags & B_INVAL)) { /* * Failed write, redirty. Must clear BIO_ERROR to prevent * pages from being scrapped. If B_INVAL is set then * this case is not run and the next case is run to * destroy the buffer. B_INVAL can occur if the buffer * is outside the range supported by the underlying device. */ bp->b_ioflags &= ~BIO_ERROR; bdirty(bp); } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { /* * Either a failed I/O or we were asked to free or not * cache the buffer. */ bp->b_flags |= B_INVAL; if (LIST_FIRST(&bp->b_dep) != NULL) buf_deallocate(bp); if (bp->b_flags & B_DELWRI) { atomic_subtract_int(&numdirtybuffers, 1); numdirtywakeup(lodirtybuffers); } bp->b_flags &= ~(B_DELWRI | B_CACHE); if ((bp->b_flags & B_VMIO) == 0) { if (bp->b_bufsize) allocbuf(bp, 0); if (bp->b_vp) brelvp(bp); } } /* * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() * is called with B_DELWRI set, the underlying pages may wind up * getting freed causing a previous write (bdwrite()) to get 'lost' * because pages associated with a B_DELWRI bp are marked clean. * * We still allow the B_INVAL case to call vfs_vmio_release(), even * if B_DELWRI is set. * * If B_DELWRI is not set we may have to set B_RELBUF if we are low * on pages to return pages to the VM page queues. */ if (bp->b_flags & B_DELWRI) bp->b_flags &= ~B_RELBUF; else if (vm_page_count_severe()) { /* * XXX This lock may not be necessary since BKGRDINPROG * cannot be set while we hold the buf lock, it can only be * cleared if it is already pending. */ if (bp->b_vp) { BO_LOCK(bp->b_bufobj); if (!(bp->b_vflags & BV_BKGRDINPROG)) bp->b_flags |= B_RELBUF; BO_UNLOCK(bp->b_bufobj); } else bp->b_flags |= B_RELBUF; } /* * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer * constituted, not even NFS buffers now. Two flags effect this. If * B_INVAL, the struct buf is invalidated but the VM object is kept * around ( i.e. so it is trivial to reconstitute the buffer later ). * * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be * invalidated. BIO_ERROR cannot be set for a failed write unless the * buffer is also B_INVAL because it hits the re-dirtying code above. * * Normally we can do this whether a buffer is B_DELWRI or not. If * the buffer is an NFS buffer, it is tracking piecemeal writes or * the commit state and we cannot afford to lose the buffer. If the * buffer has a background write in progress, we need to keep it * around to prevent it from being reconstituted and starting a second * background write. */ if ((bp->b_flags & B_VMIO) && !(bp->b_vp->v_mount != NULL && (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI)) ) { int i, j, resid; vm_page_t m; off_t foff; vm_pindex_t poff; vm_object_t obj; obj = bp->b_bufobj->bo_object; /* * Get the base offset and length of the buffer. Note that * in the VMIO case if the buffer block size is not * page-aligned then b_data pointer may not be page-aligned. * But our b_pages[] array *IS* page aligned. * * block sizes less then DEV_BSIZE (usually 512) are not * supported due to the page granularity bits (m->valid, * m->dirty, etc...). * * See man buf(9) for more information */ resid = bp->b_bufsize; foff = bp->b_offset; VM_OBJECT_LOCK(obj); for (i = 0; i < bp->b_npages; i++) { int had_bogus = 0; m = bp->b_pages[i]; /* * If we hit a bogus page, fixup *all* the bogus pages * now. */ if (m == bogus_page) { poff = OFF_TO_IDX(bp->b_offset); had_bogus = 1; for (j = i; j < bp->b_npages; j++) { vm_page_t mtmp; mtmp = bp->b_pages[j]; if (mtmp == bogus_page) { mtmp = vm_page_lookup(obj, poff + j); if (!mtmp) { panic("brelse: page missing\n"); } bp->b_pages[j] = mtmp; } } if ((bp->b_flags & B_INVAL) == 0) { pmap_qenter( trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } m = bp->b_pages[i]; } if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { int poffset = foff & PAGE_MASK; int presid = resid > (PAGE_SIZE - poffset) ? (PAGE_SIZE - poffset) : resid; KASSERT(presid >= 0, ("brelse: extra page")); vm_page_lock_queues(); vm_page_set_invalid(m, poffset, presid); vm_page_unlock_queues(); if (had_bogus) printf("avoided corruption bug in bogus_page/brelse code\n"); } resid -= PAGE_SIZE - (foff & PAGE_MASK); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } VM_OBJECT_UNLOCK(obj); if (bp->b_flags & (B_INVAL | B_RELBUF)) vfs_vmio_release(bp); } else if (bp->b_flags & B_VMIO) { if (bp->b_flags & (B_INVAL | B_RELBUF)) { vfs_vmio_release(bp); } } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { if (bp->b_bufsize != 0) allocbuf(bp, 0); if (bp->b_vp != NULL) brelvp(bp); } if (BUF_REFCNT(bp) > 1) { /* do not release to free list */ BUF_UNLOCK(bp); return; } /* enqueue */ mtx_lock(&bqlock); /* Handle delayed bremfree() processing. */ if (bp->b_flags & B_REMFREE) bremfreel(bp); if (bp->b_qindex != QUEUE_NONE) panic("brelse: free buffer onto another queue???"); /* buffers with no memory */ if (bp->b_bufsize == 0) { bp->b_flags |= B_INVAL; bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 1"); if (bp->b_kvasize) { bp->b_qindex = QUEUE_EMPTYKVA; } else { bp->b_qindex = QUEUE_EMPTY; } TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); /* buffers with junk contents */ } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { bp->b_flags |= B_INVAL; bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 2"); bp->b_qindex = QUEUE_CLEAN; TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); /* remaining buffers */ } else { + if (bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) + bp->b_qindex = QUEUE_DIRTY_GIANT; if (bp->b_flags & B_DELWRI) bp->b_qindex = QUEUE_DIRTY; else bp->b_qindex = QUEUE_CLEAN; if (bp->b_flags & B_AGE) TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); else TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); } mtx_unlock(&bqlock); /* * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already * placed the buffer on the correct queue. We must also disassociate * the device and vnode for a B_INVAL buffer so gbincore() doesn't * find it. */ if (bp->b_flags & B_INVAL) { if (bp->b_flags & B_DELWRI) bundirty(bp); if (bp->b_vp) brelvp(bp); } /* * Fixup numfreebuffers count. The bp is on an appropriate queue * unless locked. We then bump numfreebuffers if it is not B_DELWRI. * We've already handled the B_INVAL case ( B_DELWRI will be clear * if B_INVAL is set ). */ if (!(bp->b_flags & B_DELWRI)) bufcountwakeup(); /* * Something we can maybe free or reuse */ if (bp->b_bufsize || bp->b_kvasize) bufspacewakeup(); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("brelse: not dirty"); /* unlock */ BUF_UNLOCK(bp); } /* * Release a buffer back to the appropriate queue but do not try to free * it. The buffer is expected to be used again soon. * * bqrelse() is used by bdwrite() to requeue a delayed write, and used by * biodone() to requeue an async I/O on completion. It is also used when * known good buffers need to be requeued but we think we may need the data * again soon. * * XXX we should be able to leave the B_RELBUF hint set on completion. */ void bqrelse(struct buf *bp) { CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); if (BUF_REFCNT(bp) > 1) { /* do not release to free list */ BUF_UNLOCK(bp); return; } if (bp->b_flags & B_MANAGED) { if (bp->b_flags & B_REMFREE) { mtx_lock(&bqlock); bremfreel(bp); mtx_unlock(&bqlock); } bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); BUF_UNLOCK(bp); return; } mtx_lock(&bqlock); /* Handle delayed bremfree() processing. */ if (bp->b_flags & B_REMFREE) bremfreel(bp); if (bp->b_qindex != QUEUE_NONE) panic("bqrelse: free buffer onto another queue???"); /* buffers with stale but valid contents */ if (bp->b_flags & B_DELWRI) { - bp->b_qindex = QUEUE_DIRTY; - TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); + if (bp->b_flags & B_NEEDSGIANT) + bp->b_qindex = QUEUE_DIRTY_GIANT; + else + bp->b_qindex = QUEUE_DIRTY; + TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); } else { /* * XXX This lock may not be necessary since BKGRDINPROG * cannot be set while we hold the buf lock, it can only be * cleared if it is already pending. */ BO_LOCK(bp->b_bufobj); if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { BO_UNLOCK(bp->b_bufobj); bp->b_qindex = QUEUE_CLEAN; TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); } else { /* * We are too low on memory, we have to try to free * the buffer (most importantly: the wired pages * making up its backing store) *now*. */ BO_UNLOCK(bp->b_bufobj); mtx_unlock(&bqlock); brelse(bp); return; } } mtx_unlock(&bqlock); if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) bufcountwakeup(); /* * Something we can maybe free or reuse. */ if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) bufspacewakeup(); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("bqrelse: not dirty"); /* unlock */ BUF_UNLOCK(bp); } /* Give pages used by the bp back to the VM system (where possible) */ static void vfs_vmio_release(struct buf *bp) { int i; vm_page_t m; VM_OBJECT_LOCK(bp->b_bufobj->bo_object); vm_page_lock_queues(); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; bp->b_pages[i] = NULL; /* * In order to keep page LRU ordering consistent, put * everything on the inactive queue. */ vm_page_unwire(m, 0); /* * We don't mess with busy pages, it is * the responsibility of the process that * busied the pages to deal with them. */ if ((m->flags & PG_BUSY) || (m->busy != 0)) continue; if (m->wire_count == 0) { /* * Might as well free the page if we can and it has * no valid data. We also free the page if the * buffer was used for direct I/O */ if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { vm_page_free(m); } else if (bp->b_flags & B_DIRECT) { vm_page_try_to_free(m); } else if (vm_page_count_severe()) { vm_page_try_to_cache(m); } } } vm_page_unlock_queues(); VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); if (bp->b_bufsize) { bufspacewakeup(); bp->b_bufsize = 0; } bp->b_npages = 0; bp->b_flags &= ~B_VMIO; if (bp->b_vp) brelvp(bp); } /* * Check to see if a block at a particular lbn is available for a clustered * write. */ static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) { struct buf *bpa; int match; match = 0; /* If the buf isn't in core skip it */ if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) return (0); /* If the buf is busy we don't want to wait for it */ if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) return (0); /* Only cluster with valid clusterable delayed write buffers */ if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != (B_DELWRI | B_CLUSTEROK)) goto done; if (bpa->b_bufsize != size) goto done; /* * Check to see if it is in the expected place on disk and that the * block has been mapped. */ if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) match = 1; done: BUF_UNLOCK(bpa); return (match); } /* * vfs_bio_awrite: * * Implement clustered async writes for clearing out B_DELWRI buffers. * This is much better then the old way of writing only one buffer at * a time. Note that we may not be presented with the buffers in the * correct order, so we search for the cluster in both directions. */ int vfs_bio_awrite(struct buf *bp) { int i; int j; daddr_t lblkno = bp->b_lblkno; struct vnode *vp = bp->b_vp; int ncl; int nwritten; int size; int maxcl; /* * right now we support clustered writing only to regular files. If * we find a clusterable block we could be in the middle of a cluster * rather then at the beginning. */ if ((vp->v_type == VREG) && (vp->v_mount != 0) && /* Only on nodes that have the size info */ (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { size = vp->v_mount->mnt_stat.f_iosize; maxcl = MAXPHYS / size; VI_LOCK(vp); for (i = 1; i < maxcl; i++) if (vfs_bio_clcheck(vp, size, lblkno + i, bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) break; for (j = 1; i + j <= maxcl && j <= lblkno; j++) if (vfs_bio_clcheck(vp, size, lblkno - j, bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) break; VI_UNLOCK(vp); --j; ncl = i + j; /* * this is a possible cluster write */ if (ncl != 1) { BUF_UNLOCK(bp); nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); return nwritten; } } bremfree(bp); bp->b_flags |= B_ASYNC; /* * default (old) behavior, writing out only one block * * XXX returns b_bufsize instead of b_bcount for nwritten? */ nwritten = bp->b_bufsize; (void) bwrite(bp); return nwritten; } /* * getnewbuf: * * Find and initialize a new buffer header, freeing up existing buffers * in the bufqueues as necessary. The new buffer is returned locked. * * Important: B_INVAL is not set. If the caller wishes to throw the * buffer away, the caller must set B_INVAL prior to calling brelse(). * * We block if: * We have insufficient buffer headers * We have insufficient buffer space * buffer_map is too fragmented ( space reservation fails ) * If we have to flush dirty buffers ( but we try to avoid this ) * * To avoid VFS layer recursion we do not flush dirty buffers ourselves. * Instead we ask the buf daemon to do it for us. We attempt to * avoid piecemeal wakeups of the pageout daemon. */ static struct buf * getnewbuf(int slpflag, int slptimeo, int size, int maxsize) { struct buf *bp; struct buf *nbp; int defrag = 0; int nqindex; static int flushingbufs; /* * We can't afford to block since we might be holding a vnode lock, * which may prevent system daemons from running. We deal with * low-memory situations by proactively returning memory and running * async I/O rather then sync I/O. */ atomic_add_int(&getnewbufcalls, 1); atomic_subtract_int(&getnewbufrestarts, 1); restart: atomic_add_int(&getnewbufrestarts, 1); /* * Setup for scan. If we do not have enough free buffers, * we setup a degenerate case that immediately fails. Note * that if we are specially marked process, we are allowed to * dip into our reserves. * * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN * * We start with EMPTYKVA. If the list is empty we backup to EMPTY. * However, there are a number of cases (defragging, reusing, ...) * where we cannot backup. */ mtx_lock(&bqlock); nqindex = QUEUE_EMPTYKVA; nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); if (nbp == NULL) { /* * If no EMPTYKVA buffers and we are either * defragging or reusing, locate a CLEAN buffer * to free or reuse. If bufspace useage is low * skip this step so we can allocate a new buffer. */ if (defrag || bufspace >= lobufspace) { nqindex = QUEUE_CLEAN; nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); } /* * If we could not find or were not allowed to reuse a * CLEAN buffer, check to see if it is ok to use an EMPTY * buffer. We can only use an EMPTY buffer if allocating * its KVA would not otherwise run us out of buffer space. */ if (nbp == NULL && defrag == 0 && bufspace + maxsize < hibufspace) { nqindex = QUEUE_EMPTY; nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); } } /* * Run scan, possibly freeing data and/or kva mappings on the fly * depending. */ while ((bp = nbp) != NULL) { int qindex = nqindex; /* * Calculate next bp ( we can only use it if we do not block * or do other fancy things ). */ if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { switch(qindex) { case QUEUE_EMPTY: nqindex = QUEUE_EMPTYKVA; if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) break; /* FALLTHROUGH */ case QUEUE_EMPTYKVA: nqindex = QUEUE_CLEAN; if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) break; /* FALLTHROUGH */ case QUEUE_CLEAN: /* * nbp is NULL. */ break; } } /* * If we are defragging then we need a buffer with * b_kvasize != 0. XXX this situation should no longer * occur, if defrag is non-zero the buffer's b_kvasize * should also be non-zero at this point. XXX */ if (defrag && bp->b_kvasize == 0) { printf("Warning: defrag empty buffer %p\n", bp); continue; } /* * Start freeing the bp. This is somewhat involved. nbp * remains valid only for QUEUE_EMPTY[KVA] bp's. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) continue; if (bp->b_vp) { BO_LOCK(bp->b_bufobj); if (bp->b_vflags & BV_BKGRDINPROG) { BO_UNLOCK(bp->b_bufobj); BUF_UNLOCK(bp); continue; } BO_UNLOCK(bp->b_bufobj); } CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, bp->b_kvasize, bp->b_bufsize, qindex); /* * Sanity Checks */ KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); /* * Note: we no longer distinguish between VMIO and non-VMIO * buffers. */ KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); bremfreel(bp); mtx_unlock(&bqlock); if (qindex == QUEUE_CLEAN) { if (bp->b_flags & B_VMIO) { bp->b_flags &= ~B_ASYNC; vfs_vmio_release(bp); } if (bp->b_vp) brelvp(bp); } /* * NOTE: nbp is now entirely invalid. We can only restart * the scan from this point on. * * Get the rest of the buffer freed up. b_kva* is still * valid after this operation. */ if (bp->b_rcred != NOCRED) { crfree(bp->b_rcred); bp->b_rcred = NOCRED; } if (bp->b_wcred != NOCRED) { crfree(bp->b_wcred); bp->b_wcred = NOCRED; } if (LIST_FIRST(&bp->b_dep) != NULL) buf_deallocate(bp); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 3"); KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", bp, bp->b_vp, qindex)); KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); if (bp->b_bufsize) allocbuf(bp, 0); bp->b_flags = 0; bp->b_ioflags = 0; bp->b_xflags = 0; bp->b_vflags = 0; bp->b_vp = NULL; bp->b_blkno = bp->b_lblkno = 0; bp->b_offset = NOOFFSET; bp->b_iodone = 0; bp->b_error = 0; bp->b_resid = 0; bp->b_bcount = 0; bp->b_npages = 0; bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_bufobj = NULL; bp->b_pin_count = 0; bp->b_fsprivate1 = NULL; bp->b_fsprivate2 = NULL; bp->b_fsprivate3 = NULL; LIST_INIT(&bp->b_dep); /* * If we are defragging then free the buffer. */ if (defrag) { bp->b_flags |= B_INVAL; bfreekva(bp); brelse(bp); defrag = 0; goto restart; } /* * If we are overcomitted then recover the buffer and its * KVM space. This occurs in rare situations when multiple * processes are blocked in getnewbuf() or allocbuf(). */ if (bufspace >= hibufspace) flushingbufs = 1; if (flushingbufs && bp->b_kvasize != 0) { bp->b_flags |= B_INVAL; bfreekva(bp); brelse(bp); goto restart; } if (bufspace < lobufspace) flushingbufs = 0; break; } /* * If we exhausted our list, sleep as appropriate. We may have to * wakeup various daemons and write out some dirty buffers. * * Generally we are sleeping due to insufficient buffer space. */ if (bp == NULL) { int flags; char *waitmsg; if (defrag) { flags = VFS_BIO_NEED_BUFSPACE; waitmsg = "nbufkv"; } else if (bufspace >= hibufspace) { waitmsg = "nbufbs"; flags = VFS_BIO_NEED_BUFSPACE; } else { waitmsg = "newbuf"; flags = VFS_BIO_NEED_ANY; } mtx_lock(&nblock); needsbuffer |= flags; mtx_unlock(&nblock); mtx_unlock(&bqlock); bd_speedup(); /* heeeelp */ mtx_lock(&nblock); while (needsbuffer & flags) { if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { mtx_unlock(&nblock); return (NULL); } } mtx_unlock(&nblock); } else { /* * We finally have a valid bp. We aren't quite out of the * woods, we still have to reserve kva space. In order * to keep fragmentation sane we only allocate kva in * BKVASIZE chunks. */ maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; if (maxsize != bp->b_kvasize) { vm_offset_t addr = 0; bfreekva(bp); vm_map_lock(buffer_map); if (vm_map_findspace(buffer_map, vm_map_min(buffer_map), maxsize, &addr)) { /* * Uh oh. Buffer map is to fragmented. We * must defragment the map. */ atomic_add_int(&bufdefragcnt, 1); vm_map_unlock(buffer_map); defrag = 1; bp->b_flags |= B_INVAL; brelse(bp); goto restart; } if (addr) { vm_map_insert(buffer_map, NULL, 0, addr, addr + maxsize, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); bp->b_kvabase = (caddr_t) addr; bp->b_kvasize = maxsize; atomic_add_int(&bufspace, bp->b_kvasize); atomic_add_int(&bufreusecnt, 1); } vm_map_unlock(buffer_map); } bp->b_saveaddr = bp->b_kvabase; bp->b_data = bp->b_saveaddr; } return(bp); } /* * buf_daemon: * * buffer flushing daemon. Buffers are normally flushed by the * update daemon but if it cannot keep up this process starts to * take the load in an attempt to prevent getnewbuf() from blocking. */ static struct kproc_desc buf_kp = { "bufdaemon", buf_daemon, &bufdaemonproc }; SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) static void buf_daemon() { - mtx_lock(&Giant); /* * This process needs to be suspended prior to shutdown sync. */ EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, SHUTDOWN_PRI_LAST); /* * This process is allowed to take the buffer cache to the limit */ curthread->td_pflags |= TDP_NORUNNINGBUF; mtx_lock(&bdlock); for (;;) { bd_request = 0; mtx_unlock(&bdlock); kthread_suspend_check(bufdaemonproc); /* * Do the flush. Limit the amount of in-transit I/O we * allow to build up, otherwise we would completely saturate * the I/O system. Wakeup any waiting processes before we * normally would so they can run in parallel with our drain. */ while (numdirtybuffers > lodirtybuffers) { - if (flushbufqueues(0) == 0) { + int flushed; + + flushed = flushbufqueues(QUEUE_DIRTY, 0); + /* The list empty check here is slightly racy */ + if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { + mtx_lock(&Giant); + flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); + mtx_unlock(&Giant); + } + if (flushed == 0) { /* * Could not find any buffers without rollback * dependencies, so just write the first one * in the hopes of eventually making progress. */ - flushbufqueues(1); + flushbufqueues(QUEUE_DIRTY, 1); + if (!TAILQ_EMPTY( + &bufqueues[QUEUE_DIRTY_GIANT])) { + mtx_lock(&Giant); + flushbufqueues(QUEUE_DIRTY_GIANT, 1); + mtx_unlock(&Giant); + } break; } uio_yield(); } /* * Only clear bd_request if we have reached our low water * mark. The buf_daemon normally waits 1 second and * then incrementally flushes any dirty buffers that have * built up, within reason. * * If we were unable to hit our low water mark and couldn't * find any flushable buffers, we sleep half a second. * Otherwise we loop immediately. */ mtx_lock(&bdlock); if (numdirtybuffers <= lodirtybuffers) { /* * We reached our low water mark, reset the * request and sleep until we are needed again. * The sleep is just so the suspend code works. */ bd_request = 0; msleep(&bd_request, &bdlock, PVM, "psleep", hz); } else { /* * We couldn't find any flushable dirty buffers but * still have too many dirty buffers, we * have to sleep and try again. (rare) */ msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); } } } /* * flushbufqueues: * * Try to flush a buffer in the dirty queue. We must be careful to * free up B_INVAL buffers instead of write them, which NFS is * particularly sensitive to. */ static int flushwithdeps = 0; SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 0, "Number of buffers flushed with dependecies that require rollbacks"); static int -flushbufqueues(int flushdeps) +flushbufqueues(int queue, int flushdeps) { struct thread *td = curthread; struct buf sentinel; struct vnode *vp; struct mount *mp; struct buf *bp; int hasdeps; int flushed; int target; target = numdirtybuffers - lodirtybuffers; if (flushdeps && target > 2) target /= 2; flushed = 0; bp = NULL; mtx_lock(&bqlock); - TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], &sentinel, b_freelist); + TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); while (flushed != target) { - bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); + bp = TAILQ_FIRST(&bufqueues[queue]); if (bp == &sentinel) break; - TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], bp, b_freelist); - TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); + TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); + TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) continue; if (bp->b_pin_count > 0) { BUF_UNLOCK(bp); continue; } BO_LOCK(bp->b_bufobj); if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || (bp->b_flags & B_DELWRI) == 0) { BO_UNLOCK(bp->b_bufobj); BUF_UNLOCK(bp); continue; } BO_UNLOCK(bp->b_bufobj); if (bp->b_flags & B_INVAL) { bremfreel(bp); mtx_unlock(&bqlock); brelse(bp); flushed++; numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); mtx_lock(&bqlock); continue; } if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { if (flushdeps == 0) { BUF_UNLOCK(bp); continue; } hasdeps = 1; } else hasdeps = 0; /* * We must hold the lock on a vnode before writing * one of its buffers. Otherwise we may confuse, or * in the case of a snapshot vnode, deadlock the * system. * * The lock order here is the reverse of the normal * of vnode followed by buf lock. This is ok because * the NOWAIT will prevent deadlock. */ vp = bp->b_vp; if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { BUF_UNLOCK(bp); continue; } if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { mtx_unlock(&bqlock); CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); vfs_bio_awrite(bp); vn_finished_write(mp); VOP_UNLOCK(vp, 0, td); flushwithdeps += hasdeps; flushed++; waitrunningbufspace(); numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); mtx_lock(&bqlock); continue; } vn_finished_write(mp); BUF_UNLOCK(bp); } - TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], &sentinel, b_freelist); + TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); mtx_unlock(&bqlock); return (flushed); } /* * Check to see if a block is currently memory resident. */ struct buf * incore(struct bufobj *bo, daddr_t blkno) { struct buf *bp; BO_LOCK(bo); bp = gbincore(bo, blkno); BO_UNLOCK(bo); return (bp); } /* * Returns true if no I/O is needed to access the * associated VM object. This is like incore except * it also hunts around in the VM system for the data. */ static int inmem(struct vnode * vp, daddr_t blkno) { vm_object_t obj; vm_offset_t toff, tinc, size; vm_page_t m; vm_ooffset_t off; ASSERT_VOP_LOCKED(vp, "inmem"); if (incore(&vp->v_bufobj, blkno)) return 1; if (vp->v_mount == NULL) return 0; obj = vp->v_object; if (obj == NULL) return (0); size = PAGE_SIZE; if (size > vp->v_mount->mnt_stat.f_iosize) size = vp->v_mount->mnt_stat.f_iosize; off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; VM_OBJECT_LOCK(obj); for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); if (!m) goto notinmem; tinc = size; if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); if (vm_page_is_valid(m, (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) goto notinmem; } VM_OBJECT_UNLOCK(obj); return 1; notinmem: VM_OBJECT_UNLOCK(obj); return (0); } /* * vfs_setdirty: * * Sets the dirty range for a buffer based on the status of the dirty * bits in the pages comprising the buffer. * * The range is limited to the size of the buffer. * * This routine is primarily used by NFS, but is generalized for the * B_VMIO case. */ static void vfs_setdirty(struct buf *bp) { int i; vm_object_t object; /* * Degenerate case - empty buffer */ if (bp->b_bufsize == 0) return; /* * We qualify the scan for modified pages on whether the * object has been flushed yet. The OBJ_WRITEABLE flag * is not cleared simply by protecting pages off. */ if ((bp->b_flags & B_VMIO) == 0) return; object = bp->b_pages[0]->object; VM_OBJECT_LOCK(object); if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) printf("Warning: object %p writeable but not mightbedirty\n", object); if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) printf("Warning: object %p mightbedirty but not writeable\n", object); if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { vm_offset_t boffset; vm_offset_t eoffset; vm_page_lock_queues(); /* * test the pages to see if they have been modified directly * by users through the VM system. */ for (i = 0; i < bp->b_npages; i++) vm_page_test_dirty(bp->b_pages[i]); /* * Calculate the encompassing dirty range, boffset and eoffset, * (eoffset - boffset) bytes. */ for (i = 0; i < bp->b_npages; i++) { if (bp->b_pages[i]->dirty) break; } boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); for (i = bp->b_npages - 1; i >= 0; --i) { if (bp->b_pages[i]->dirty) { break; } } eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); vm_page_unlock_queues(); /* * Fit it to the buffer. */ if (eoffset > bp->b_bcount) eoffset = bp->b_bcount; /* * If we have a good dirty range, merge with the existing * dirty range. */ if (boffset < eoffset) { if (bp->b_dirtyoff > boffset) bp->b_dirtyoff = boffset; if (bp->b_dirtyend < eoffset) bp->b_dirtyend = eoffset; } } VM_OBJECT_UNLOCK(object); } /* * getblk: * * Get a block given a specified block and offset into a file/device. * The buffers B_DONE bit will be cleared on return, making it almost * ready for an I/O initiation. B_INVAL may or may not be set on * return. The caller should clear B_INVAL prior to initiating a * READ. * * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for * an existing buffer. * * For a VMIO buffer, B_CACHE is modified according to the backing VM. * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set * and then cleared based on the backing VM. If the previous buffer is * non-0-sized but invalid, B_CACHE will be cleared. * * If getblk() must create a new buffer, the new buffer is returned with * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which * case it is returned with B_INVAL clear and B_CACHE set based on the * backing VM. * * getblk() also forces a bwrite() for any B_DELWRI buffer whos * B_CACHE bit is clear. * * What this means, basically, is that the caller should use B_CACHE to * determine whether the buffer is fully valid or not and should clear * B_INVAL prior to issuing a read. If the caller intends to validate * the buffer by loading its data area with something, the caller needs * to clear B_INVAL. If the caller does this without issuing an I/O, * the caller should set B_CACHE ( as an optimization ), else the caller * should issue the I/O and biodone() will set B_CACHE if the I/O was * a write attempt or if it was a successfull read. If the caller * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR * prior to issuing the READ. biodone() will *not* clear B_INVAL. */ struct buf * getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags) { struct buf *bp; struct bufobj *bo; int error; CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); ASSERT_VOP_LOCKED(vp, "getblk"); if (size > MAXBSIZE) panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); bo = &vp->v_bufobj; loop: /* * Block if we are low on buffers. Certain processes are allowed * to completely exhaust the buffer cache. * * If this check ever becomes a bottleneck it may be better to * move it into the else, when gbincore() fails. At the moment * it isn't a problem. * * XXX remove if 0 sections (clean this up after its proven) */ if (numfreebuffers == 0) { if (curthread == PCPU_GET(idlethread)) return NULL; mtx_lock(&nblock); needsbuffer |= VFS_BIO_NEED_ANY; mtx_unlock(&nblock); } VI_LOCK(vp); bp = gbincore(bo, blkno); if (bp != NULL) { int lockflags; /* * Buffer is in-core. If the buffer is not busy, it must * be on a queue. */ lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; if (flags & GB_LOCK_NOWAIT) lockflags |= LK_NOWAIT; error = BUF_TIMELOCK(bp, lockflags, VI_MTX(vp), "getblk", slpflag, slptimeo); /* * If we slept and got the lock we have to restart in case * the buffer changed identities. */ if (error == ENOLCK) goto loop; /* We timed out or were interrupted. */ else if (error) return (NULL); /* * The buffer is locked. B_CACHE is cleared if the buffer is * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set * and for a VMIO buffer B_CACHE is adjusted according to the * backing VM cache. */ if (bp->b_flags & B_INVAL) bp->b_flags &= ~B_CACHE; else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) bp->b_flags |= B_CACHE; bremfree(bp); /* * check for size inconsistancies for non-VMIO case. */ if (bp->b_bcount != size) { if ((bp->b_flags & B_VMIO) == 0 || (size > bp->b_kvasize)) { if (bp->b_flags & B_DELWRI) { /* * If buffer is pinned and caller does * not want sleep waiting for it to be * unpinned, bail out * */ if (bp->b_pin_count > 0) { if (flags & GB_LOCK_NOWAIT) { bqrelse(bp); return (NULL); } else { bunpin_wait(bp); } } bp->b_flags |= B_NOCACHE; bwrite(bp); } else { if (LIST_FIRST(&bp->b_dep) == NULL) { bp->b_flags |= B_RELBUF; brelse(bp); } else { bp->b_flags |= B_NOCACHE; bwrite(bp); } } goto loop; } } /* * If the size is inconsistant in the VMIO case, we can resize * the buffer. This might lead to B_CACHE getting set or * cleared. If the size has not changed, B_CACHE remains * unchanged from its previous state. */ if (bp->b_bcount != size) allocbuf(bp, size); KASSERT(bp->b_offset != NOOFFSET, ("getblk: no buffer offset")); /* * A buffer with B_DELWRI set and B_CACHE clear must * be committed before we can return the buffer in * order to prevent the caller from issuing a read * ( due to B_CACHE not being set ) and overwriting * it. * * Most callers, including NFS and FFS, need this to * operate properly either because they assume they * can issue a read if B_CACHE is not set, or because * ( for example ) an uncached B_DELWRI might loop due * to softupdates re-dirtying the buffer. In the latter * case, B_CACHE is set after the first write completes, * preventing further loops. * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE * above while extending the buffer, we cannot allow the * buffer to remain with B_CACHE set after the write * completes or it will represent a corrupt state. To * deal with this we set B_NOCACHE to scrap the buffer * after the write. * * We might be able to do something fancy, like setting * B_CACHE in bwrite() except if B_DELWRI is already set, * so the below call doesn't set B_CACHE, but that gets real * confusing. This is much easier. */ if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); goto loop; } bp->b_flags &= ~B_DONE; } else { int bsize, maxsize, vmio; off_t offset; /* * Buffer is not in-core, create new buffer. The buffer * returned by getnewbuf() is locked. Note that the returned * buffer is also considered valid (not marked B_INVAL). */ VI_UNLOCK(vp); /* * If the user does not want us to create the buffer, bail out * here. */ if (flags & GB_NOCREAT) return NULL; bsize = bo->bo_bsize; offset = blkno * bsize; vmio = vp->v_object != NULL; maxsize = vmio ? size + (offset & PAGE_MASK) : size; maxsize = imax(maxsize, bsize); bp = getnewbuf(slpflag, slptimeo, size, maxsize); if (bp == NULL) { if (slpflag || slptimeo) return NULL; goto loop; } /* * This code is used to make sure that a buffer is not * created while the getnewbuf routine is blocked. * This can be a problem whether the vnode is locked or not. * If the buffer is created out from under us, we have to * throw away the one we just created. * * Note: this must occur before we associate the buffer * with the vp especially considering limitations in * the splay tree implementation when dealing with duplicate * lblkno's. */ BO_LOCK(bo); if (gbincore(bo, blkno)) { BO_UNLOCK(bo); bp->b_flags |= B_INVAL; brelse(bp); goto loop; } /* * Insert the buffer into the hash, so that it can * be found by incore. */ bp->b_blkno = bp->b_lblkno = blkno; bp->b_offset = offset; - bgetvp(vp, bp); BO_UNLOCK(bo); /* * set B_VMIO bit. allocbuf() the buffer bigger. Since the * buffer size starts out as 0, B_CACHE will be set by * allocbuf() for the VMIO case prior to it testing the * backing store for validity. */ if (vmio) { bp->b_flags |= B_VMIO; #if defined(VFS_BIO_DEBUG) if (vn_canvmio(vp) != TRUE) printf("getblk: VMIO on vnode type %d\n", vp->v_type); #endif KASSERT(vp->v_object == bp->b_bufobj->bo_object, ("ARGH! different b_bufobj->bo_object %p %p %p\n", bp, vp->v_object, bp->b_bufobj->bo_object)); } else { bp->b_flags &= ~B_VMIO; KASSERT(bp->b_bufobj->bo_object == NULL, ("ARGH! has b_bufobj->bo_object %p %p\n", bp, bp->b_bufobj->bo_object)); } allocbuf(bp, size); bp->b_flags &= ~B_DONE; } CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); return (bp); } /* * Get an empty, disassociated buffer of given size. The buffer is initially * set to B_INVAL. */ struct buf * geteblk(int size) { struct buf *bp; int maxsize; maxsize = (size + BKVAMASK) & ~BKVAMASK; while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) continue; allocbuf(bp, size); bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); return (bp); } /* * This code constitutes the buffer memory from either anonymous system * memory (in the case of non-VMIO operations) or from an associated * VM object (in the case of VMIO operations). This code is able to * resize a buffer up or down. * * Note that this code is tricky, and has many complications to resolve * deadlock or inconsistant data situations. Tread lightly!!! * There are B_CACHE and B_DELWRI interactions that must be dealt with by * the caller. Calling this code willy nilly can result in the loss of data. * * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with * B_CACHE for the non-VMIO case. */ int allocbuf(struct buf *bp, int size) { int newbsize, mbsize; int i; if (BUF_REFCNT(bp) == 0) panic("allocbuf: buffer not busy"); if (bp->b_kvasize < size) panic("allocbuf: buffer too small"); if ((bp->b_flags & B_VMIO) == 0) { caddr_t origbuf; int origbufsize; /* * Just get anonymous memory from the kernel. Don't * mess with B_CACHE. */ mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); if (bp->b_flags & B_MALLOC) newbsize = mbsize; else newbsize = round_page(size); if (newbsize < bp->b_bufsize) { /* * malloced buffers are not shrunk */ if (bp->b_flags & B_MALLOC) { if (newbsize) { bp->b_bcount = size; } else { free(bp->b_data, M_BIOBUF); if (bp->b_bufsize) { atomic_subtract_int( &bufmallocspace, bp->b_bufsize); bufspacewakeup(); bp->b_bufsize = 0; } bp->b_saveaddr = bp->b_kvabase; bp->b_data = bp->b_saveaddr; bp->b_bcount = 0; bp->b_flags &= ~B_MALLOC; } return 1; } vm_hold_free_pages( bp, (vm_offset_t) bp->b_data + newbsize, (vm_offset_t) bp->b_data + bp->b_bufsize); } else if (newbsize > bp->b_bufsize) { /* * We only use malloced memory on the first allocation. * and revert to page-allocated memory when the buffer * grows. */ /* * There is a potential smp race here that could lead * to bufmallocspace slightly passing the max. It * is probably extremely rare and not worth worrying * over. */ if ( (bufmallocspace < maxbufmallocspace) && (bp->b_bufsize == 0) && (mbsize <= PAGE_SIZE/2)) { bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); bp->b_bufsize = mbsize; bp->b_bcount = size; bp->b_flags |= B_MALLOC; atomic_add_int(&bufmallocspace, mbsize); return 1; } origbuf = NULL; origbufsize = 0; /* * If the buffer is growing on its other-than-first allocation, * then we revert to the page-allocation scheme. */ if (bp->b_flags & B_MALLOC) { origbuf = bp->b_data; origbufsize = bp->b_bufsize; bp->b_data = bp->b_kvabase; if (bp->b_bufsize) { atomic_subtract_int(&bufmallocspace, bp->b_bufsize); bufspacewakeup(); bp->b_bufsize = 0; } bp->b_flags &= ~B_MALLOC; newbsize = round_page(newbsize); } vm_hold_load_pages( bp, (vm_offset_t) bp->b_data + bp->b_bufsize, (vm_offset_t) bp->b_data + newbsize); if (origbuf) { bcopy(origbuf, bp->b_data, origbufsize); free(origbuf, M_BIOBUF); } } } else { int desiredpages; newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); desiredpages = (size == 0) ? 0 : num_pages((bp->b_offset & PAGE_MASK) + newbsize); if (bp->b_flags & B_MALLOC) panic("allocbuf: VMIO buffer can't be malloced"); /* * Set B_CACHE initially if buffer is 0 length or will become * 0-length. */ if (size == 0 || bp->b_bufsize == 0) bp->b_flags |= B_CACHE; if (newbsize < bp->b_bufsize) { /* * DEV_BSIZE aligned new buffer size is less then the * DEV_BSIZE aligned existing buffer size. Figure out * if we have to remove any pages. */ if (desiredpages < bp->b_npages) { vm_page_t m; VM_OBJECT_LOCK(bp->b_bufobj->bo_object); vm_page_lock_queues(); for (i = desiredpages; i < bp->b_npages; i++) { /* * the page is not freed here -- it * is the responsibility of * vnode_pager_setsize */ m = bp->b_pages[i]; KASSERT(m != bogus_page, ("allocbuf: bogus page found")); while (vm_page_sleep_if_busy(m, TRUE, "biodep")) vm_page_lock_queues(); bp->b_pages[i] = NULL; vm_page_unwire(m, 0); } vm_page_unlock_queues(); VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); bp->b_npages = desiredpages; } } else if (size > bp->b_bcount) { /* * We are growing the buffer, possibly in a * byte-granular fashion. */ struct vnode *vp; vm_object_t obj; vm_offset_t toff; vm_offset_t tinc; /* * Step 1, bring in the VM pages from the object, * allocating them if necessary. We must clear * B_CACHE if these pages are not valid for the * range covered by the buffer. */ vp = bp->b_vp; obj = bp->b_bufobj->bo_object; VM_OBJECT_LOCK(obj); while (bp->b_npages < desiredpages) { vm_page_t m; vm_pindex_t pi; pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; if ((m = vm_page_lookup(obj, pi)) == NULL) { /* * note: must allocate system pages * since blocking here could intefere * with paging I/O, no matter which * process we are. */ m = vm_page_alloc(obj, pi, VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); if (m == NULL) { atomic_add_int(&vm_pageout_deficit, desiredpages - bp->b_npages); VM_OBJECT_UNLOCK(obj); VM_WAIT; VM_OBJECT_LOCK(obj); } else { bp->b_flags &= ~B_CACHE; bp->b_pages[bp->b_npages] = m; ++bp->b_npages; } continue; } /* * We found a page. If we have to sleep on it, * retry because it might have gotten freed out * from under us. * * We can only test PG_BUSY here. Blocking on * m->busy might lead to a deadlock: * * vm_fault->getpages->cluster_read->allocbuf * */ vm_page_lock_queues(); if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) continue; /* * We have a good page. Should we wakeup the * page daemon? */ if ((curproc != pageproc) && (VM_PAGE_INQUEUE1(m, PQ_CACHE)) && ((cnt.v_free_count + cnt.v_cache_count) < (cnt.v_free_min + cnt.v_cache_min))) { pagedaemon_wakeup(); } vm_page_wire(m); vm_page_unlock_queues(); bp->b_pages[bp->b_npages] = m; ++bp->b_npages; } /* * Step 2. We've loaded the pages into the buffer, * we have to figure out if we can still have B_CACHE * set. Note that B_CACHE is set according to the * byte-granular range ( bcount and size ), new the * aligned range ( newbsize ). * * The VM test is against m->valid, which is DEV_BSIZE * aligned. Needless to say, the validity of the data * needs to also be DEV_BSIZE aligned. Note that this * fails with NFS if the server or some other client * extends the file's EOF. If our buffer is resized, * B_CACHE may remain set! XXX */ toff = bp->b_bcount; tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); while ((bp->b_flags & B_CACHE) && toff < size) { vm_pindex_t pi; if (tinc > (size - toff)) tinc = size - toff; pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; vfs_buf_test_cache( bp, bp->b_offset, toff, tinc, bp->b_pages[pi] ); toff += tinc; tinc = PAGE_SIZE; } VM_OBJECT_UNLOCK(obj); /* * Step 3, fixup the KVM pmap. Remember that * bp->b_data is relative to bp->b_offset, but * bp->b_offset may be offset into the first page. */ bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); pmap_qenter( (vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages ); bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | (vm_offset_t)(bp->b_offset & PAGE_MASK)); } } if (newbsize < bp->b_bufsize) bufspacewakeup(); bp->b_bufsize = newbsize; /* actual buffer allocation */ bp->b_bcount = size; /* requested buffer size */ return 1; } void biodone(struct bio *bp) { void (*done)(struct bio *); mtx_lock(&bdonelock); bp->bio_flags |= BIO_DONE; done = bp->bio_done; if (done == NULL) wakeup(bp); mtx_unlock(&bdonelock); if (done != NULL) done(bp); } /* * Wait for a BIO to finish. * * XXX: resort to a timeout for now. The optimal locking (if any) for this * case is not yet clear. */ int biowait(struct bio *bp, const char *wchan) { mtx_lock(&bdonelock); while ((bp->bio_flags & BIO_DONE) == 0) msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); mtx_unlock(&bdonelock); if (bp->bio_error != 0) return (bp->bio_error); if (!(bp->bio_flags & BIO_ERROR)) return (0); return (EIO); } void biofinish(struct bio *bp, struct devstat *stat, int error) { if (error) { bp->bio_error = error; bp->bio_flags |= BIO_ERROR; } if (stat != NULL) devstat_end_transaction_bio(stat, bp); biodone(bp); } /* * bufwait: * * Wait for buffer I/O completion, returning error status. The buffer * is left locked and B_DONE on return. B_EINTR is converted into an EINTR * error and cleared. */ int bufwait(struct buf *bp) { if (bp->b_iocmd == BIO_READ) bwait(bp, PRIBIO, "biord"); else bwait(bp, PRIBIO, "biowr"); if (bp->b_flags & B_EINTR) { bp->b_flags &= ~B_EINTR; return (EINTR); } if (bp->b_ioflags & BIO_ERROR) { return (bp->b_error ? bp->b_error : EIO); } else { return (0); } } /* * Call back function from struct bio back up to struct buf. */ static void bufdonebio(struct bio *bip) { struct buf *bp; bp = bip->bio_caller2; bp->b_resid = bp->b_bcount - bip->bio_completed; bp->b_resid = bip->bio_resid; /* XXX: remove */ bp->b_ioflags = bip->bio_flags; bp->b_error = bip->bio_error; if (bp->b_error) bp->b_ioflags |= BIO_ERROR; bufdone(bp); g_destroy_bio(bip); } void dev_strategy(struct cdev *dev, struct buf *bp) { struct cdevsw *csw; struct bio *bip; if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) panic("b_iocmd botch"); for (;;) { bip = g_new_bio(); if (bip != NULL) break; /* Try again later */ tsleep(&bp, PRIBIO, "dev_strat", hz/10); } bip->bio_cmd = bp->b_iocmd; bip->bio_offset = bp->b_iooffset; bip->bio_length = bp->b_bcount; bip->bio_bcount = bp->b_bcount; /* XXX: remove */ bip->bio_data = bp->b_data; bip->bio_done = bufdonebio; bip->bio_caller2 = bp; bip->bio_dev = dev; KASSERT(dev->si_refcount > 0, ("dev_strategy on un-referenced struct cdev *(%s)", devtoname(dev))); csw = dev_refthread(dev); if (csw == NULL) { g_destroy_bio(bip); bp->b_error = ENXIO; bp->b_ioflags = BIO_ERROR; bufdone(bp); return; } (*csw->d_strategy)(bip); dev_relthread(dev); } /* * bufdone: * * Finish I/O on a buffer, optionally calling a completion function. * This is usually called from an interrupt so process blocking is * not allowed. * * biodone is also responsible for setting B_CACHE in a B_VMIO bp. * In a non-VMIO bp, B_CACHE will be set on the next getblk() * assuming B_INVAL is clear. * * For the VMIO case, we set B_CACHE if the op was a read and no * read error occured, or if the op was a write. B_CACHE is never * set if the buffer is invalid or otherwise uncacheable. * * biodone does not mess with B_INVAL, allowing the I/O routine or the * initiator to leave B_INVAL set to brelse the buffer out of existance * in the biodone routine. */ void bufdone(struct buf *bp) { struct bufobj *dropobj; void (*biodone)(struct buf *); CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); dropobj = NULL; KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); runningbufwakeup(bp); if (bp->b_iocmd == BIO_WRITE) dropobj = bp->b_bufobj; /* call optional completion function if requested */ if (bp->b_iodone != NULL) { biodone = bp->b_iodone; bp->b_iodone = NULL; (*biodone) (bp); if (dropobj) bufobj_wdrop(dropobj); return; } bufdone_finish(bp); if (dropobj) bufobj_wdrop(dropobj); } void bufdone_finish(struct buf *bp) { KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); if (LIST_FIRST(&bp->b_dep) != NULL) buf_complete(bp); if (bp->b_flags & B_VMIO) { int i; vm_ooffset_t foff; vm_page_t m; vm_object_t obj; int iosize; struct vnode *vp = bp->b_vp; obj = bp->b_bufobj->bo_object; #if defined(VFS_BIO_DEBUG) mp_fixme("usecount and vflag accessed without locks."); if (vp->v_usecount == 0) { panic("biodone: zero vnode ref count"); } KASSERT(vp->v_object != NULL, ("biodone: vnode %p has no vm_object", vp)); #endif foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("biodone: no buffer offset")); VM_OBJECT_LOCK(obj); #if defined(VFS_BIO_DEBUG) if (obj->paging_in_progress < bp->b_npages) { printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", obj->paging_in_progress, bp->b_npages); } #endif /* * Set B_CACHE if the op was a normal read and no error * occured. B_CACHE is set for writes in the b*write() * routines. */ iosize = bp->b_bcount - bp->b_resid; if (bp->b_iocmd == BIO_READ && !(bp->b_flags & (B_INVAL|B_NOCACHE)) && !(bp->b_ioflags & BIO_ERROR)) { bp->b_flags |= B_CACHE; } vm_page_lock_queues(); for (i = 0; i < bp->b_npages; i++) { int bogusflag = 0; int resid; resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; if (resid > iosize) resid = iosize; /* * cleanup bogus pages, restoring the originals */ m = bp->b_pages[i]; if (m == bogus_page) { bogusflag = 1; m = vm_page_lookup(obj, OFF_TO_IDX(foff)); if (m == NULL) panic("biodone: page disappeared!"); bp->b_pages[i] = m; pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } #if defined(VFS_BIO_DEBUG) if (OFF_TO_IDX(foff) != m->pindex) { printf( "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", (intmax_t)foff, (uintmax_t)m->pindex); } #endif /* * In the write case, the valid and clean bits are * already changed correctly ( see bdwrite() ), so we * only need to do this here in the read case. */ if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { vfs_page_set_valid(bp, foff, i, m); } /* * when debugging new filesystems or buffer I/O methods, this * is the most common error that pops up. if you see this, you * have not set the page busy flag correctly!!! */ if (m->busy == 0) { printf("biodone: page busy < 0, " "pindex: %d, foff: 0x(%x,%x), " "resid: %d, index: %d\n", (int) m->pindex, (int)(foff >> 32), (int) foff & 0xffffffff, resid, i); if (!vn_isdisk(vp, NULL)) printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, (intmax_t) bp->b_lblkno, bp->b_flags, bp->b_npages); else printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", (intmax_t) bp->b_lblkno, bp->b_flags, bp->b_npages); printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", (u_long)m->valid, (u_long)m->dirty, m->wire_count); panic("biodone: page busy < 0\n"); } vm_page_io_finish(m); vm_object_pip_subtract(obj, 1); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; iosize -= resid; } vm_page_unlock_queues(); vm_object_pip_wakeupn(obj, 0); VM_OBJECT_UNLOCK(obj); } /* * For asynchronous completions, release the buffer now. The brelse * will do a wakeup there if necessary - so no need to do a wakeup * here in the async case. The sync case always needs to do a wakeup. */ if (bp->b_flags & B_ASYNC) { if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) brelse(bp); else bqrelse(bp); } else bdone(bp); } /* * This routine is called in lieu of iodone in the case of * incomplete I/O. This keeps the busy status for pages * consistant. */ void vfs_unbusy_pages(struct buf *bp) { int i; vm_object_t obj; vm_page_t m; runningbufwakeup(bp); if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; VM_OBJECT_LOCK(obj); vm_page_lock_queues(); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) { m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); if (!m) panic("vfs_unbusy_pages: page missing\n"); bp->b_pages[i] = m; pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } vm_object_pip_subtract(obj, 1); vm_page_io_finish(m); } vm_page_unlock_queues(); vm_object_pip_wakeupn(obj, 0); VM_OBJECT_UNLOCK(obj); } /* * vfs_page_set_valid: * * Set the valid bits in a page based on the supplied offset. The * range is restricted to the buffer's size. * * This routine is typically called after a read completes. */ static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) { vm_ooffset_t soff, eoff; mtx_assert(&vm_page_queue_mtx, MA_OWNED); /* * Start and end offsets in buffer. eoff - soff may not cross a * page boundry or cross the end of the buffer. The end of the * buffer, in this case, is our file EOF, not the allocation size * of the buffer. */ soff = off; eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > soff) { vm_page_set_validclean( m, (vm_offset_t) (soff & PAGE_MASK), (vm_offset_t) (eoff - soff) ); } } /* * This routine is called before a device strategy routine. * It is used to tell the VM system that paging I/O is in * progress, and treat the pages associated with the buffer * almost as being PG_BUSY. Also the object paging_in_progress * flag is handled to make sure that the object doesn't become * inconsistant. * * Since I/O has not been initiated yet, certain buffer flags * such as BIO_ERROR or B_INVAL may be in an inconsistant state * and should be ignored. */ void vfs_busy_pages(struct buf *bp, int clear_modify) { int i, bogus; vm_object_t obj; vm_ooffset_t foff; vm_page_t m; if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_busy_pages: no buffer offset")); vfs_setdirty(bp); VM_OBJECT_LOCK(obj); retry: vm_page_lock_queues(); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) goto retry; } bogus = 0; for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if ((bp->b_flags & B_CLUSTER) == 0) { vm_object_pip_add(obj, 1); vm_page_io_start(m); } /* * When readying a buffer for a read ( i.e * clear_modify == 0 ), it is important to do * bogus_page replacement for valid pages in * partially instantiated buffers. Partially * instantiated buffers can, in turn, occur when * reconstituting a buffer from its VM backing store * base. We only have to do this if B_CACHE is * clear ( which causes the I/O to occur in the * first place ). The replacement prevents the read * I/O from overwriting potentially dirty VM-backed * pages. XXX bogus page replacement is, uh, bogus. * It may not work properly with small-block devices. * We need to find a better way. */ pmap_remove_all(m); if (clear_modify) vfs_page_set_valid(bp, foff, i, m); else if (m->valid == VM_PAGE_BITS_ALL && (bp->b_flags & B_CACHE) == 0) { bp->b_pages[i] = bogus_page; bogus++; } foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } vm_page_unlock_queues(); VM_OBJECT_UNLOCK(obj); if (bogus) pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } /* * Tell the VM system that the pages associated with this buffer * are clean. This is used for delayed writes where the data is * going to go to disk eventually without additional VM intevention. * * Note that while we only really need to clean through to b_bcount, we * just go ahead and clean through to b_bufsize. */ static void vfs_clean_pages(struct buf *bp) { int i; vm_ooffset_t foff, noff, eoff; vm_page_t m; if (!(bp->b_flags & B_VMIO)) return; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_clean_pages: no buffer offset")); VM_OBJECT_LOCK(bp->b_bufobj->bo_object); vm_page_lock_queues(); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; eoff = noff; if (eoff > bp->b_offset + bp->b_bufsize) eoff = bp->b_offset + bp->b_bufsize; vfs_page_set_valid(bp, foff, i, m); /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ foff = noff; } vm_page_unlock_queues(); VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); } /* * vfs_bio_set_validclean: * * Set the range within the buffer to valid and clean. The range is * relative to the beginning of the buffer, b_offset. Note that b_offset * itself may be offset from the beginning of the first page. * */ void vfs_bio_set_validclean(struct buf *bp, int base, int size) { int i, n; vm_page_t m; if (!(bp->b_flags & B_VMIO)) return; /* * Fixup base to be relative to beginning of first page. * Set initial n to be the maximum number of bytes in the * first page that can be validated. */ base += (bp->b_offset & PAGE_MASK); n = PAGE_SIZE - (base & PAGE_MASK); VM_OBJECT_LOCK(bp->b_bufobj->bo_object); vm_page_lock_queues(); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; vm_page_set_validclean(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } vm_page_unlock_queues(); VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); } /* * vfs_bio_clrbuf: * * clear a buffer. This routine essentially fakes an I/O, so we need * to clear BIO_ERROR and B_INVAL. * * Note that while we only theoretically need to clear through b_bcount, * we go ahead and clear through b_bufsize. */ void vfs_bio_clrbuf(struct buf *bp) { int i, j, mask = 0; caddr_t sa, ea; if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { clrbuf(bp); return; } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; VM_OBJECT_LOCK(bp->b_bufobj->bo_object); if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && (bp->b_offset & PAGE_MASK) == 0) { if (bp->b_pages[0] == bogus_page) goto unlock; mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); if ((bp->b_pages[0]->valid & mask) == mask) goto unlock; if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && ((bp->b_pages[0]->valid & mask) == 0)) { bzero(bp->b_data, bp->b_bufsize); bp->b_pages[0]->valid |= mask; goto unlock; } } ea = sa = bp->b_data; for(i = 0; i < bp->b_npages; i++, sa = ea) { ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); ea = (caddr_t)(vm_offset_t)ulmin( (u_long)(vm_offset_t)ea, (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); if (bp->b_pages[i] == bogus_page) continue; j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); if ((bp->b_pages[i]->valid & mask) == mask) continue; if ((bp->b_pages[i]->valid & mask) == 0) { if ((bp->b_pages[i]->flags & PG_ZERO) == 0) bzero(sa, ea - sa); } else { for (; sa < ea; sa += DEV_BSIZE, j++) { if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && (bp->b_pages[i]->valid & (1 << j)) == 0) bzero(sa, DEV_BSIZE); } } bp->b_pages[i]->valid |= mask; } unlock: VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); bp->b_resid = 0; } /* * vm_hold_load_pages and vm_hold_free_pages get pages into * a buffers address space. The pages are anonymous and are * not associated with a file object. */ static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index; to = round_page(to); from = round_page(from); index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; VM_OBJECT_LOCK(kernel_object); for (pg = from; pg < to; pg += PAGE_SIZE, index++) { tryagain: /* * note: must allocate system pages since blocking here * could intefere with paging I/O, no matter which * process we are. */ p = vm_page_alloc(kernel_object, ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); if (!p) { atomic_add_int(&vm_pageout_deficit, (to - pg) >> PAGE_SHIFT); VM_OBJECT_UNLOCK(kernel_object); VM_WAIT; VM_OBJECT_LOCK(kernel_object); goto tryagain; } p->valid = VM_PAGE_BITS_ALL; pmap_qenter(pg, &p, 1); bp->b_pages[index] = p; } VM_OBJECT_UNLOCK(kernel_object); bp->b_npages = index; } /* Return pages associated with this buf to the vm system */ static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index, newnpages; from = round_page(from); to = round_page(to); newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; VM_OBJECT_LOCK(kernel_object); for (pg = from; pg < to; pg += PAGE_SIZE, index++) { p = bp->b_pages[index]; if (p && (index < bp->b_npages)) { if (p->busy) { printf( "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); } bp->b_pages[index] = NULL; pmap_qremove(pg, 1); vm_page_lock_queues(); vm_page_unwire(p, 0); vm_page_free(p); vm_page_unlock_queues(); } } VM_OBJECT_UNLOCK(kernel_object); bp->b_npages = newnpages; } /* * Map an IO request into kernel virtual address space. * * All requests are (re)mapped into kernel VA space. * Notice that we use b_bufsize for the size of the buffer * to be mapped. b_bcount might be modified by the driver. * * Note that even if the caller determines that the address space should * be valid, a race or a smaller-file mapped into a larger space may * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST * check the return value. */ int vmapbuf(struct buf *bp) { caddr_t addr, kva; vm_prot_t prot; int pidx, i; struct vm_page *m; struct pmap *pmap = &curproc->p_vmspace->vm_pmap; if (bp->b_bufsize < 0) return (-1); prot = VM_PROT_READ; if (bp->b_iocmd == BIO_READ) prot |= VM_PROT_WRITE; /* Less backwards than it looks */ for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; addr < bp->b_data + bp->b_bufsize; addr += PAGE_SIZE, pidx++) { /* * Do the vm_fault if needed; do the copy-on-write thing * when reading stuff off device into memory. * * NOTE! Must use pmap_extract() because addr may be in * the userland address space, and kextract is only guarenteed * to work for the kernland address space (see: sparc64 port). */ retry: if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, prot) < 0) { vm_page_lock_queues(); for (i = 0; i < pidx; ++i) { vm_page_unhold(bp->b_pages[i]); bp->b_pages[i] = NULL; } vm_page_unlock_queues(); return(-1); } m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); if (m == NULL) goto retry; bp->b_pages[pidx] = m; } if (pidx > btoc(MAXPHYS)) panic("vmapbuf: mapped more than MAXPHYS"); pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); kva = bp->b_saveaddr; bp->b_npages = pidx; bp->b_saveaddr = bp->b_data; bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); return(0); } /* * Free the io map PTEs associated with this IO operation. * We also invalidate the TLB entries and restore the original b_addr. */ void vunmapbuf(struct buf *bp) { int pidx; int npages; npages = bp->b_npages; pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); vm_page_lock_queues(); for (pidx = 0; pidx < npages; pidx++) vm_page_unhold(bp->b_pages[pidx]); vm_page_unlock_queues(); bp->b_data = bp->b_saveaddr; } void bdone(struct buf *bp) { mtx_lock(&bdonelock); bp->b_flags |= B_DONE; wakeup(bp); mtx_unlock(&bdonelock); } void bwait(struct buf *bp, u_char pri, const char *wchan) { mtx_lock(&bdonelock); while ((bp->b_flags & B_DONE) == 0) msleep(bp, &bdonelock, pri, wchan, 0); mtx_unlock(&bdonelock); } int bufsync(struct bufobj *bo, int waitfor, struct thread *td) { return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); } void bufstrategy(struct bufobj *bo, struct buf *bp) { int i = 0; struct vnode *vp; vp = bp->b_vp; KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); i = VOP_STRATEGY(vp, bp); KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); } void bufobj_wrefl(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); ASSERT_BO_LOCKED(bo); bo->bo_numoutput++; } void bufobj_wref(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); BO_LOCK(bo); bo->bo_numoutput++; BO_UNLOCK(bo); } void bufobj_wdrop(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); BO_LOCK(bo); KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { bo->bo_flag &= ~BO_WWAIT; wakeup(&bo->bo_numoutput); } BO_UNLOCK(bo); } int bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) { int error; KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); ASSERT_BO_LOCKED(bo); error = 0; while (bo->bo_numoutput) { bo->bo_flag |= BO_WWAIT; error = msleep(&bo->bo_numoutput, BO_MTX(bo), slpflag | (PRIBIO + 1), "bo_wwait", timeo); if (error) break; } return (error); } void bpin(struct buf *bp) { mtx_lock(&bpinlock); bp->b_pin_count++; mtx_unlock(&bpinlock); } void bunpin(struct buf *bp) { mtx_lock(&bpinlock); if (--bp->b_pin_count == 0) wakeup(bp); mtx_unlock(&bpinlock); } void bunpin_wait(struct buf *bp) { mtx_lock(&bpinlock); while (bp->b_pin_count > 0) msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0); mtx_unlock(&bpinlock); } #include "opt_ddb.h" #ifdef DDB #include /* DDB command to show buffer data */ DB_SHOW_COMMAND(buffer, db_show_buffer) { /* get args */ struct buf *bp = (struct buf *)addr; if (!have_addr) { db_printf("usage: show buffer \n"); return; } db_printf("buf at %p\n", bp); db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); db_printf( "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); if (bp->b_npages) { int i; db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); for (i = 0; i < bp->b_npages; i++) { vm_page_t m; m = bp->b_pages[i]; db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); if ((i + 1) < bp->b_npages) db_printf(","); } db_printf("\n"); } lockmgr_printinfo(&bp->b_lock); } DB_SHOW_COMMAND(lockedbufs, lockedbufs) { struct buf *bp; int i; for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (lockcount(&bp->b_lock)) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } } } #endif /* DDB */ Index: head/sys/kern/vfs_subr.c =================================================================== --- head/sys/kern/vfs_subr.c (revision 157318) +++ head/sys/kern/vfs_subr.c (revision 157319) @@ -1,3883 +1,3886 @@ /*- * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 */ /* * External virtual filesystem routines */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_mac.h" #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 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure"); static void delmntque(struct vnode *vp); static void insmntque(struct vnode *vp, struct mount *mp); static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, int slptimeo); static void syncer_shutdown(void *arg, int howto); static int vtryrecycle(struct vnode *vp); static void vbusy(struct vnode *vp); static void vdropl(struct vnode *vp); static void vinactive(struct vnode *, struct thread *); static void v_incr_usecount(struct vnode *); static void v_decr_usecount(struct vnode *); static void v_decr_useonly(struct vnode *); static void v_upgrade_usecount(struct vnode *); static void vfree(struct vnode *); static void vnlru_free(int); static void vdestroy(struct vnode *); static void vgonel(struct vnode *); static void vfs_knllock(void *arg); static void vfs_knlunlock(void *arg); static int vfs_knllocked(void *arg); /* * Enable Giant pushdown based on whether or not the vm is mpsafe in this * build. Without mpsafevm the buffer cache can not run Giant free. */ #if defined(__alpha__) || defined(__amd64__) || defined(__i386__) || \ defined(__ia64__) || defined(__sparc64__) int mpsafe_vfs = 1; #else int mpsafe_vfs; #endif TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs); SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0, "MPSAFE VFS"); /* * Number of vnodes in existence. Increased whenever getnewvnode() * allocates a new vnode, never decreased. */ static unsigned long numvnodes; SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); /* * Conversion tables for conversion from vnode types to inode formats * and back. */ enum vtype iftovt_tab[16] = { VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, }; int vttoif_tab[10] = { 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT }; /* * List of vnodes that are ready for recycling. */ static TAILQ_HEAD(freelst, vnode) vnode_free_list; /* * Free vnode target. Free vnodes may simply be files which have been stat'd * but not read. This is somewhat common, and a small cache of such files * should be kept to avoid recreation costs. */ static u_long wantfreevnodes; SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); /* Number of vnodes in the free list. */ static u_long freevnodes; SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); /* * Various variables used for debugging the new implementation of * reassignbuf(). * XXX these are probably of (very) limited utility now. */ static int reassignbufcalls; SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); /* * Cache for the mount type id assigned to NFS. This is used for * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. */ int nfs_mount_type = -1; /* To keep more than one thread at a time from running vfs_getnewfsid */ static struct mtx mntid_mtx; /* * Lock for any access to the following: * vnode_free_list * numvnodes * freevnodes */ static struct mtx vnode_free_list_mtx; /* Publicly exported FS */ struct nfs_public nfs_pub; /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ static uma_zone_t vnode_zone; static uma_zone_t vnodepoll_zone; /* Set to 1 to print out reclaim of active vnodes */ int prtactive; /* * The workitem queue. * * It is useful to delay writes of file data and filesystem metadata * for tens of seconds so that quickly created and deleted files need * not waste disk bandwidth being created and removed. To realize this, * we append vnodes to a "workitem" queue. When running with a soft * updates implementation, most pending metadata dependencies should * not wait for more than a few seconds. Thus, mounted on block devices * are delayed only about a half the time that file data is delayed. * Similarly, directory updates are more critical, so are only delayed * about a third the time that file data is delayed. Thus, there are * SYNCER_MAXDELAY queues that are processed round-robin at a rate of * one each second (driven off the filesystem syncer process). The * syncer_delayno variable indicates the next queue that is to be processed. * Items that need to be processed soon are placed in this queue: * * syncer_workitem_pending[syncer_delayno] * * A delay of fifteen seconds is done by placing the request fifteen * entries later in the queue: * * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] * */ static int syncer_delayno; static long syncer_mask; LIST_HEAD(synclist, bufobj); static struct synclist *syncer_workitem_pending; /* * The sync_mtx protects: * bo->bo_synclist * sync_vnode_count * syncer_delayno * syncer_state * syncer_workitem_pending * syncer_worklist_len * rushjob */ static struct mtx sync_mtx; #define SYNCER_MAXDELAY 32 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ static int syncdelay = 30; /* max time to delay syncing data */ static int filedelay = 30; /* time to delay syncing files */ SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); static int dirdelay = 29; /* time to delay syncing directories */ SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); static int metadelay = 28; /* time to delay syncing metadata */ SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); static int rushjob; /* number of slots to run ASAP */ static int stat_rush_requests; /* number of times I/O speeded up */ SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); /* * When shutting down the syncer, run it at four times normal speed. */ #define SYNCER_SHUTDOWN_SPEEDUP 4 static int sync_vnode_count; static int syncer_worklist_len; static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } syncer_state; /* * Number of vnodes we want to exist at any one time. This is mostly used * to size hash tables in vnode-related code. It is normally not used in * getnewvnode(), as wantfreevnodes is normally nonzero.) * * XXX desiredvnodes is historical cruft and should not exist. */ int desiredvnodes; SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, &desiredvnodes, 0, "Maximum number of vnodes"); SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, &wantfreevnodes, 0, "Minimum number of vnodes (legacy)"); static int vnlru_nowhere; SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); /* * Macros to control when a vnode is freed and recycled. All require * the vnode interlock. */ #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt) /* * Initialize the vnode management data structures. */ #ifndef MAXVNODES_MAX #define MAXVNODES_MAX 100000 #endif static void vntblinit(void *dummy __unused) { /* * Desiredvnodes is a function of the physical memory size and * the kernel's heap size. Specifically, desiredvnodes scales * in proportion to the physical memory size until two fifths * of the kernel's heap size is consumed by vnodes and vm * objects. */ desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); if (desiredvnodes > MAXVNODES_MAX) { if (bootverbose) printf("Reducing kern.maxvnodes %d -> %d\n", desiredvnodes, MAXVNODES_MAX); desiredvnodes = MAXVNODES_MAX; } wantfreevnodes = desiredvnodes / 4; mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); TAILQ_INIT(&vnode_free_list); mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); /* * Initialize the filesystem syncer. */ syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, &syncer_mask); syncer_maxdelay = syncer_mask + 1; mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) /* * Mark a mount point as busy. Used to synchronize access and to delay * unmounting. Interlock is not released on failure. */ int vfs_busy(struct mount *mp, int flags, struct mtx *interlkp, struct thread *td) { int lkflags; MNT_ILOCK(mp); MNT_REF(mp); if (mp->mnt_kern_flag & MNTK_UNMOUNT) { if (flags & LK_NOWAIT) { MNT_REL(mp); MNT_IUNLOCK(mp); return (ENOENT); } if (interlkp) mtx_unlock(interlkp); mp->mnt_kern_flag |= MNTK_MWAIT; /* * Since all busy locks are shared except the exclusive * lock granted when unmounting, the only place that a * wakeup needs to be done is at the release of the * exclusive lock at the end of dounmount. */ msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0); MNT_REL(mp); MNT_IUNLOCK(mp); if (interlkp) mtx_lock(interlkp); return (ENOENT); } if (interlkp) mtx_unlock(interlkp); lkflags = LK_SHARED | LK_INTERLOCK; if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td)) panic("vfs_busy: unexpected lock failure"); vfs_rel(mp); return (0); } /* * Free a busy filesystem. */ void vfs_unbusy(struct mount *mp, struct thread *td) { lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); } /* * Lookup a mount point by filesystem identifier. */ struct mount * vfs_getvfs(fsid_t *fsid) { struct mount *mp; mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { mtx_unlock(&mountlist_mtx); return (mp); } } mtx_unlock(&mountlist_mtx); return ((struct mount *) 0); } /* * Check if a user can access priveledged mount options. */ int vfs_suser(struct mount *mp, struct thread *td) { int error; if ((mp->mnt_flag & MNT_USER) == 0 || mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { if ((error = suser(td)) != 0) return (error); } return (0); } /* * Get a new unique fsid. Try to make its val[0] unique, since this value * will be used to create fake device numbers for stat(). Also try (but * not so hard) make its val[0] unique mod 2^16, since some emulators only * support 16-bit device numbers. We end up with unique val[0]'s for the * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. * * Keep in mind that several mounts may be running in parallel. Starting * the search one past where the previous search terminated is both a * micro-optimization and a defense against returning the same fsid to * different mounts. */ void vfs_getnewfsid(struct mount *mp) { static u_int16_t mntid_base; fsid_t tfsid; int mtype; mtx_lock(&mntid_mtx); mtype = mp->mnt_vfc->vfc_typenum; tfsid.val[1] = mtype; mtype = (mtype & 0xFF) << 24; for (;;) { tfsid.val[0] = makedev(255, mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); mntid_base++; if (vfs_getvfs(&tfsid) == NULL) break; } mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; mtx_unlock(&mntid_mtx); } /* * Knob to control the precision of file timestamps: * * 0 = seconds only; nanoseconds zeroed. * 1 = seconds and nanoseconds, accurate within 1/HZ. * 2 = seconds and nanoseconds, truncated to microseconds. * >=3 = seconds and nanoseconds, maximum precision. */ enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; static int timestamp_precision = TSP_SEC; SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, ×tamp_precision, 0, ""); /* * Get a current timestamp. */ void vfs_timestamp(struct timespec *tsp) { struct timeval tv; switch (timestamp_precision) { case TSP_SEC: tsp->tv_sec = time_second; tsp->tv_nsec = 0; break; case TSP_HZ: getnanotime(tsp); break; case TSP_USEC: microtime(&tv); TIMEVAL_TO_TIMESPEC(&tv, tsp); break; case TSP_NSEC: default: nanotime(tsp); break; } } /* * Set vnode attributes to VNOVAL */ void vattr_null(struct vattr *vap) { vap->va_type = VNON; vap->va_size = VNOVAL; vap->va_bytes = VNOVAL; vap->va_mode = VNOVAL; vap->va_nlink = VNOVAL; vap->va_uid = VNOVAL; vap->va_gid = VNOVAL; vap->va_fsid = VNOVAL; vap->va_fileid = VNOVAL; vap->va_blocksize = VNOVAL; vap->va_rdev = VNOVAL; vap->va_atime.tv_sec = VNOVAL; vap->va_atime.tv_nsec = VNOVAL; vap->va_mtime.tv_sec = VNOVAL; vap->va_mtime.tv_nsec = VNOVAL; vap->va_ctime.tv_sec = VNOVAL; vap->va_ctime.tv_nsec = VNOVAL; vap->va_birthtime.tv_sec = VNOVAL; vap->va_birthtime.tv_nsec = VNOVAL; vap->va_flags = VNOVAL; vap->va_gen = VNOVAL; vap->va_vaflags = 0; } /* * This routine is called when we have too many vnodes. It attempts * to free vnodes and will potentially free vnodes that still * have VM backing store (VM backing store is typically the cause * of a vnode blowout so we want to do this). Therefore, this operation * is not considered cheap. * * A number of conditions may prevent a vnode from being reclaimed. * the buffer cache may have references on the vnode, a directory * vnode may still have references due to the namei cache representing * underlying files, or the vnode may be in active use. It is not * desireable to reuse such vnodes. These conditions may cause the * number of vnodes to reach some minimum value regardless of what * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. */ static int vlrureclaim(struct mount *mp) { struct thread *td; struct vnode *vp; int done; int trigger; int usevnodes; int count; /* * Calculate the trigger point, don't allow user * screwups to blow us up. This prevents us from * recycling vnodes with lots of resident pages. We * aren't trying to free memory, we are trying to * free vnodes. */ usevnodes = desiredvnodes; if (usevnodes <= 0) usevnodes = 1; trigger = cnt.v_page_count * 2 / usevnodes; done = 0; td = curthread; vn_start_write(NULL, &mp, V_WAIT); MNT_ILOCK(mp); count = mp->mnt_nvnodelistsize / 10 + 1; while (count != 0) { vp = TAILQ_FIRST(&mp->mnt_nvnodelist); while (vp != NULL && vp->v_type == VMARKER) vp = TAILQ_NEXT(vp, v_nmntvnodes); if (vp == NULL) break; TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); --count; if (!VI_TRYLOCK(vp)) goto next_iter; /* * If it's been deconstructed already, it's still * referenced, or it exceeds the trigger, skip it. */ if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && vp->v_object->resident_page_count > trigger)) { VI_UNLOCK(vp); goto next_iter; } MNT_IUNLOCK(mp); vholdl(vp); if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) { vdrop(vp); goto next_iter_mntunlocked; } VI_LOCK(vp); /* * v_usecount may have been bumped after VOP_LOCK() dropped * the vnode interlock and before it was locked again. * * It is not necessary to recheck VI_DOOMED because it can * only be set by another thread that holds both the vnode * lock and vnode interlock. If another thread has the * vnode lock before we get to VOP_LOCK() and obtains the * vnode interlock after VOP_LOCK() drops the vnode * interlock, the other thread will be unable to drop the * vnode lock before our VOP_LOCK() call fails. */ if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || (vp->v_object != NULL && vp->v_object->resident_page_count > trigger)) { VOP_UNLOCK(vp, LK_INTERLOCK, td); goto next_iter_mntunlocked; } KASSERT((vp->v_iflag & VI_DOOMED) == 0, ("VI_DOOMED unexpectedly detected in vlrureclaim()")); vgonel(vp); VOP_UNLOCK(vp, 0, td); vdropl(vp); done++; next_iter_mntunlocked: if ((count % 256) != 0) goto relock_mnt; goto yield; next_iter: if ((count % 256) != 0) continue; MNT_IUNLOCK(mp); yield: uio_yield(); relock_mnt: MNT_ILOCK(mp); } MNT_IUNLOCK(mp); vn_finished_write(mp); return done; } /* * Attempt to keep the free list at wantfreevnodes length. */ static void vnlru_free(int count) { struct vnode *vp; mtx_assert(&vnode_free_list_mtx, MA_OWNED); for (; count > 0; count--) { vp = TAILQ_FIRST(&vnode_free_list); /* * The list can be modified while the free_list_mtx * has been dropped and vp could be NULL here. */ if (!vp) break; VNASSERT(vp->v_op != NULL, vp, ("vnlru_free: vnode already reclaimed.")); TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); /* * Don't recycle if we can't get the interlock. */ if (!VI_TRYLOCK(vp)) { TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); continue; } VNASSERT(VCANRECYCLE(vp), vp, ("vp inconsistent on freelist")); freevnodes--; vp->v_iflag &= ~VI_FREE; vholdl(vp); mtx_unlock(&vnode_free_list_mtx); VI_UNLOCK(vp); vtryrecycle(vp); /* * If the recycled succeeded this vdrop will actually free * the vnode. If not it will simply place it back on * the free list. */ vdrop(vp); mtx_lock(&vnode_free_list_mtx); } } /* * Attempt to recycle vnodes in a context that is always safe to block. * Calling vlrurecycle() from the bowels of filesystem code has some * interesting deadlock problems. */ static struct proc *vnlruproc; static int vnlruproc_sig; static void vnlru_proc(void) { struct mount *mp, *nmp; int done; struct proc *p = vnlruproc; struct thread *td = FIRST_THREAD_IN_PROC(p); mtx_lock(&Giant); EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, SHUTDOWN_PRI_FIRST); for (;;) { kthread_suspend_check(p); mtx_lock(&vnode_free_list_mtx); if (freevnodes > wantfreevnodes) vnlru_free(freevnodes - wantfreevnodes); if (numvnodes <= desiredvnodes * 9 / 10) { vnlruproc_sig = 0; wakeup(&vnlruproc_sig); msleep(vnlruproc, &vnode_free_list_mtx, PVFS|PDROP, "vlruwt", hz); continue; } mtx_unlock(&vnode_free_list_mtx); done = 0; mtx_lock(&mountlist_mtx); for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { int vfsunlocked; if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { nmp = TAILQ_NEXT(mp, mnt_list); continue; } if (!VFS_NEEDSGIANT(mp)) { mtx_unlock(&Giant); vfsunlocked = 1; } else vfsunlocked = 0; done += vlrureclaim(mp); if (vfsunlocked) mtx_lock(&Giant); mtx_lock(&mountlist_mtx); nmp = TAILQ_NEXT(mp, mnt_list); vfs_unbusy(mp, td); } mtx_unlock(&mountlist_mtx); if (done == 0) { #if 0 /* These messages are temporary debugging aids */ if (vnlru_nowhere < 5) printf("vnlru process getting nowhere..\n"); else if (vnlru_nowhere == 5) printf("vnlru process messages stopped.\n"); #endif vnlru_nowhere++; tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); } else uio_yield(); } } static struct kproc_desc vnlru_kp = { "vnlru", vnlru_proc, &vnlruproc }; SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) /* * Routines having to do with the management of the vnode table. */ static void vdestroy(struct vnode *vp) { struct bufobj *bo; CTR1(KTR_VFS, "vdestroy vp %p", vp); mtx_lock(&vnode_free_list_mtx); numvnodes--; mtx_unlock(&vnode_free_list_mtx); bo = &vp->v_bufobj; VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("cleaned vnode still on the free list.")); VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL")); VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL")); VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); VI_UNLOCK(vp); #ifdef MAC mac_destroy_vnode(vp); #endif if (vp->v_pollinfo != NULL) { knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note); mtx_destroy(&vp->v_pollinfo->vpi_lock); uma_zfree(vnodepoll_zone, vp->v_pollinfo); } #ifdef INVARIANTS /* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */ vp->v_op = NULL; #endif lockdestroy(vp->v_vnlock); mtx_destroy(&vp->v_interlock); uma_zfree(vnode_zone, vp); } /* * Try to recycle a freed vnode. We abort if anyone picks up a reference * before we actually vgone(). This function must be called with the vnode * held to prevent the vnode from being returned to the free list midway * through vgone(). */ static int vtryrecycle(struct vnode *vp) { struct thread *td = curthread; struct mount *vnmp; CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp); VNASSERT(vp->v_holdcnt, vp, ("vtryrecycle: Recycling vp %p without a reference.", vp)); /* * This vnode may found and locked via some other list, if so we * can't recycle it yet. */ if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) return (EWOULDBLOCK); /* * Don't recycle if its filesystem is being suspended. */ if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { VOP_UNLOCK(vp, 0, td); return (EBUSY); } /* * If we got this far, we need to acquire the interlock and see if * anyone picked up this vnode from another list. If not, we will * mark it with DOOMED via vgonel() so that anyone who does find it * will skip over it. */ VI_LOCK(vp); if (vp->v_usecount) { VOP_UNLOCK(vp, LK_INTERLOCK, td); vn_finished_write(vnmp); return (EBUSY); } if ((vp->v_iflag & VI_DOOMED) == 0) vgonel(vp); VOP_UNLOCK(vp, LK_INTERLOCK, td); vn_finished_write(vnmp); CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp); return (0); } /* * Return the next vnode from the free list. */ int getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, struct vnode **vpp) { struct vnode *vp = NULL; struct bufobj *bo; mtx_lock(&vnode_free_list_mtx); /* * Lend our context to reclaim vnodes if they've exceeded the max. */ if (freevnodes > wantfreevnodes) vnlru_free(1); /* * Wait for available vnodes. */ if (numvnodes > desiredvnodes) { if (vnlruproc_sig == 0) { vnlruproc_sig = 1; /* avoid unnecessary wakeups */ wakeup(vnlruproc); } msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, "vlruwk", hz); #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ if (numvnodes > desiredvnodes) { mtx_unlock(&vnode_free_list_mtx); return (ENFILE); } #endif } numvnodes++; mtx_unlock(&vnode_free_list_mtx); vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); /* * Setup locks. */ vp->v_vnlock = &vp->v_lock; mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); /* * By default, don't allow shared locks unless filesystems * opt-in. */ lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE); /* * Initialize bufobj. */ bo = &vp->v_bufobj; bo->__bo_vnode = vp; bo->bo_mtx = &vp->v_interlock; bo->bo_ops = &buf_ops_bio; bo->bo_private = vp; TAILQ_INIT(&bo->bo_clean.bv_hd); TAILQ_INIT(&bo->bo_dirty.bv_hd); /* * Initialize namecache. */ LIST_INIT(&vp->v_cache_src); TAILQ_INIT(&vp->v_cache_dst); /* * Finalize various vnode identity bits. */ vp->v_type = VNON; vp->v_tag = tag; vp->v_op = vops; v_incr_usecount(vp); vp->v_data = 0; #ifdef MAC mac_init_vnode(vp); if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) mac_associate_vnode_singlelabel(mp, vp); else if (mp == NULL) printf("NULL mp in getnewvnode()\n"); #endif if (mp != NULL) { insmntque(vp, mp); bo->bo_bsize = mp->mnt_stat.f_iosize; if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) vp->v_vflag |= VV_NOKNOTE; } CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp); *vpp = vp; return (0); } /* * Delete from old mount point vnode list, if on one. */ static void delmntque(struct vnode *vp) { struct mount *mp; mp = vp->v_mount; if (mp == NULL) return; MNT_ILOCK(mp); vp->v_mount = NULL; VNASSERT(mp->mnt_nvnodelistsize > 0, vp, ("bad mount point vnode list size")); TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); mp->mnt_nvnodelistsize--; MNT_REL(mp); MNT_IUNLOCK(mp); } /* * Insert into list of vnodes for the new mount point, if available. */ static void insmntque(struct vnode *vp, struct mount *mp) { vp->v_mount = mp; VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); MNT_ILOCK(mp); MNT_REF(mp); TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, ("neg mount point vnode list size")); mp->mnt_nvnodelistsize++; MNT_IUNLOCK(mp); } /* * Flush out and invalidate all buffers associated with a bufobj * Called with the underlying object locked. */ int bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, int slptimeo) { int error; BO_LOCK(bo); if (flags & V_SAVE) { error = bufobj_wwait(bo, slpflag, slptimeo); if (error) { BO_UNLOCK(bo); return (error); } if (bo->bo_dirty.bv_cnt > 0) { BO_UNLOCK(bo); if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0) return (error); /* * XXX We could save a lock/unlock if this was only * enabled under INVARIANTS */ BO_LOCK(bo); if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) panic("vinvalbuf: dirty bufs"); } } /* * If you alter this loop please notice that interlock is dropped and * reacquired in flushbuflist. Special care is needed to ensure that * no race conditions occur from this. */ do { error = flushbuflist(&bo->bo_clean, flags, bo, slpflag, slptimeo); if (error == 0) error = flushbuflist(&bo->bo_dirty, flags, bo, slpflag, slptimeo); if (error != 0 && error != EAGAIN) { BO_UNLOCK(bo); return (error); } } while (error != 0); /* * Wait for I/O to complete. XXX needs cleaning up. The vnode can * have write I/O in-progress but if there is a VM object then the * VM object can also have read-I/O in-progress. */ do { bufobj_wwait(bo, 0, 0); BO_UNLOCK(bo); if (bo->bo_object != NULL) { VM_OBJECT_LOCK(bo->bo_object); vm_object_pip_wait(bo->bo_object, "bovlbx"); VM_OBJECT_UNLOCK(bo->bo_object); } BO_LOCK(bo); } while (bo->bo_numoutput > 0); BO_UNLOCK(bo); /* * Destroy the copy in the VM cache, too. */ if (bo->bo_object != NULL) { VM_OBJECT_LOCK(bo->bo_object); vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? TRUE : FALSE); VM_OBJECT_UNLOCK(bo->bo_object); } #ifdef INVARIANTS BO_LOCK(bo); if ((flags & (V_ALT | V_NORMAL)) == 0 && (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) panic("vinvalbuf: flush failed"); BO_UNLOCK(bo); #endif return (0); } /* * Flush out and invalidate all buffers associated with a vnode. * Called with the underlying object locked. */ int vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, int slptimeo) { CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags); ASSERT_VOP_LOCKED(vp, "vinvalbuf"); return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo)); } /* * Flush out buffers on the specified list. * */ static int flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, int slptimeo) { struct buf *bp, *nbp; int retval, error; daddr_t lblkno; b_xflags_t xflags; ASSERT_BO_LOCKED(bo); retval = 0; TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { continue; } lblkno = 0; xflags = 0; if (nbp != NULL) { lblkno = nbp->b_lblkno; xflags = nbp->b_xflags & (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN); } retval = EAGAIN; error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), "flushbuf", slpflag, slptimeo); if (error) { BO_LOCK(bo); return (error != ENOLCK ? error : EAGAIN); } KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); if (bp->b_bufobj != bo) { /* XXX: necessary ? */ BUF_UNLOCK(bp); BO_LOCK(bo); return (EAGAIN); } /* * XXX Since there are no node locks for NFS, I * believe there is a slight chance that a delayed * write will occur while sleeping just above, so * check for it. */ if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && (flags & V_SAVE)) { bremfree(bp); bp->b_flags |= B_ASYNC; bwrite(bp); BO_LOCK(bo); return (EAGAIN); /* XXX: why not loop ? */ } bremfree(bp); bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); BO_LOCK(bo); if (nbp != NULL && (nbp->b_bufobj != bo || nbp->b_lblkno != lblkno || (nbp->b_xflags & (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags)) break; /* nbp invalid */ } return (retval); } /* * Truncate a file's buffer and pages to a specified length. This * is in lieu of the old vinvalbuf mechanism, which performed unneeded * sync activity. */ int vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize) { struct buf *bp, *nbp; int anyfreed; int trunclbn; struct bufobj *bo; CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length); /* * Round up to the *next* lbn. */ trunclbn = (length + blksize - 1) / blksize; ASSERT_VOP_LOCKED(vp, "vtruncbuf"); restart: VI_LOCK(vp); bo = &vp->v_bufobj; anyfreed = 1; for (;anyfreed;) { anyfreed = 0; TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno < trunclbn) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp)) == ENOLCK) goto restart; bremfree(bp); bp->b_flags |= (B_INVAL | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = 1; if (nbp != NULL && (((nbp->b_xflags & BX_VNCLEAN) == 0) || (nbp->b_vp != vp) || (nbp->b_flags & B_DELWRI))) { goto restart; } VI_LOCK(vp); } TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno < trunclbn) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp)) == ENOLCK) goto restart; bremfree(bp); bp->b_flags |= (B_INVAL | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = 1; if (nbp != NULL && (((nbp->b_xflags & BX_VNDIRTY) == 0) || (nbp->b_vp != vp) || (nbp->b_flags & B_DELWRI) == 0)) { goto restart; } VI_LOCK(vp); } } if (length > 0) { restartsync: TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno > 0) continue; /* * Since we hold the vnode lock this should only * fail if we're racing with the buf daemon. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp)) == ENOLCK) { goto restart; } VNASSERT((bp->b_flags & B_DELWRI), vp, ("buf(%p) on dirty queue without DELWRI", bp)); bremfree(bp); bawrite(bp); VI_LOCK(vp); goto restartsync; } } bufobj_wwait(bo, 0, 0); VI_UNLOCK(vp); vnode_pager_setsize(vp, length); return (0); } /* * buf_splay() - splay tree core for the clean/dirty list of buffers in * a vnode. * * NOTE: We have to deal with the special case of a background bitmap * buffer, a situation where two buffers will have the same logical * block offset. We want (1) only the foreground buffer to be accessed * in a lookup and (2) must differentiate between the foreground and * background buffer in the splay tree algorithm because the splay * tree cannot normally handle multiple entities with the same 'index'. * We accomplish this by adding differentiating flags to the splay tree's * numerical domain. */ static struct buf * buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) { struct buf dummy; struct buf *lefttreemax, *righttreemin, *y; if (root == NULL) return (NULL); lefttreemax = righttreemin = &dummy; for (;;) { if (lblkno < root->b_lblkno || (lblkno == root->b_lblkno && (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { if ((y = root->b_left) == NULL) break; if (lblkno < y->b_lblkno) { /* Rotate right. */ root->b_left = y->b_right; y->b_right = root; root = y; if ((y = root->b_left) == NULL) break; } /* Link into the new root's right tree. */ righttreemin->b_left = root; righttreemin = root; } else if (lblkno > root->b_lblkno || (lblkno == root->b_lblkno && (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { if ((y = root->b_right) == NULL) break; if (lblkno > y->b_lblkno) { /* Rotate left. */ root->b_right = y->b_left; y->b_left = root; root = y; if ((y = root->b_right) == NULL) break; } /* Link into the new root's left tree. */ lefttreemax->b_right = root; lefttreemax = root; } else { break; } root = y; } /* Assemble the new root. */ lefttreemax->b_right = root->b_left; righttreemin->b_left = root->b_right; root->b_left = dummy.b_right; root->b_right = dummy.b_left; return (root); } static void buf_vlist_remove(struct buf *bp) { struct buf *root; struct bufv *bv; KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); ASSERT_BO_LOCKED(bp->b_bufobj); KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != (BX_VNDIRTY|BX_VNCLEAN), ("buf_vlist_remove: Buf %p is on two lists", bp)); if (bp->b_xflags & BX_VNDIRTY) bv = &bp->b_bufobj->bo_dirty; else bv = &bp->b_bufobj->bo_clean; if (bp != bv->bv_root) { root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); KASSERT(root == bp, ("splay lookup failed in remove")); } if (bp->b_left == NULL) { root = bp->b_right; } else { root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); root->b_right = bp->b_right; } bv->bv_root = root; TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); bv->bv_cnt--; bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); } /* * Add the buffer to the sorted clean or dirty block list using a * splay tree algorithm. * * NOTE: xflags is passed as a constant, optimizing this inline function! */ static void buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) { struct buf *root; struct bufv *bv; ASSERT_BO_LOCKED(bo); KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); bp->b_xflags |= xflags; if (xflags & BX_VNDIRTY) bv = &bo->bo_dirty; else bv = &bo->bo_clean; root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); if (root == NULL) { bp->b_left = NULL; bp->b_right = NULL; TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); } else if (bp->b_lblkno < root->b_lblkno || (bp->b_lblkno == root->b_lblkno && (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { bp->b_left = root->b_left; bp->b_right = root; root->b_left = NULL; TAILQ_INSERT_BEFORE(root, bp, b_bobufs); } else { bp->b_right = root->b_right; bp->b_left = root; root->b_right = NULL; TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); } bv->bv_cnt++; bv->bv_root = bp; } /* * Lookup a buffer using the splay tree. Note that we specifically avoid * shadow buffers used in background bitmap writes. * * This code isn't quite efficient as it could be because we are maintaining * two sorted lists and do not know which list the block resides in. * * During a "make buildworld" the desired buffer is found at one of * the roots more than 60% of the time. Thus, checking both roots * before performing either splay eliminates unnecessary splays on the * first tree splayed. */ struct buf * gbincore(struct bufobj *bo, daddr_t lblkno) { struct buf *bp; ASSERT_BO_LOCKED(bo); if ((bp = bo->bo_clean.bv_root) != NULL && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) return (bp); if ((bp = bo->bo_dirty.bv_root) != NULL && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) return (bp); if ((bp = bo->bo_clean.bv_root) != NULL) { bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) return (bp); } if ((bp = bo->bo_dirty.bv_root) != NULL) { bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) return (bp); } return (NULL); } /* * Associate a buffer with a vnode. */ void bgetvp(struct vnode *vp, struct buf *bp) { VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, ("bgetvp: bp already attached! %p", bp)); ASSERT_VI_LOCKED(vp, "bgetvp"); vholdl(vp); + if (VFS_NEEDSGIANT(vp->v_mount)) + bp->b_flags |= B_NEEDSGIANT; bp->b_vp = vp; bp->b_bufobj = &vp->v_bufobj; /* * Insert onto list for new vnode. */ buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); } /* * Disassociate a buffer from a vnode. */ void brelvp(struct buf *bp) { struct bufobj *bo; struct vnode *vp; CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); /* * Delete from old vnode list, if on one. */ vp = bp->b_vp; /* XXX */ bo = bp->b_bufobj; BO_LOCK(bo); if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) buf_vlist_remove(bp); else panic("brelvp: Buffer %p not on queue.", bp); if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { bo->bo_flag &= ~BO_ONWORKLST; mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; mtx_unlock(&sync_mtx); } + bp->b_flags &= ~B_NEEDSGIANT; bp->b_vp = NULL; bp->b_bufobj = NULL; vdropl(vp); } /* * Add an item to the syncer work queue. */ static void vn_syncer_add_to_worklist(struct bufobj *bo, int delay) { int slot; ASSERT_BO_LOCKED(bo); mtx_lock(&sync_mtx); if (bo->bo_flag & BO_ONWORKLST) LIST_REMOVE(bo, bo_synclist); else { bo->bo_flag |= BO_ONWORKLST; syncer_worklist_len++; } if (delay > syncer_maxdelay - 2) delay = syncer_maxdelay - 2; slot = (syncer_delayno + delay) & syncer_mask; LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); mtx_unlock(&sync_mtx); } static int sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) { int error, len; mtx_lock(&sync_mtx); len = syncer_worklist_len - sync_vnode_count; mtx_unlock(&sync_mtx); error = SYSCTL_OUT(req, &len, sizeof(len)); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); static struct proc *updateproc; static void sched_sync(void); static struct kproc_desc up_kp = { "syncer", sched_sync, &updateproc }; SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) static int sync_vnode(struct bufobj *bo, struct thread *td) { struct vnode *vp; struct mount *mp; vp = bo->__bo_vnode; /* XXX */ if (VOP_ISLOCKED(vp, NULL) != 0) return (1); if (VI_TRYLOCK(vp) == 0) return (1); /* * We use vhold in case the vnode does not * successfully sync. vhold prevents the vnode from * going away when we unlock the sync_mtx so that * we can acquire the vnode interlock. */ vholdl(vp); mtx_unlock(&sync_mtx); VI_UNLOCK(vp); if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { vdrop(vp); mtx_lock(&sync_mtx); return (1); } vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); (void) VOP_FSYNC(vp, MNT_LAZY, td); VOP_UNLOCK(vp, 0, td); vn_finished_write(mp); VI_LOCK(vp); if ((bo->bo_flag & BO_ONWORKLST) != 0) { /* * Put us back on the worklist. The worklist * routine will remove us from our current * position and then add us back in at a later * position. */ vn_syncer_add_to_worklist(bo, syncdelay); } vdropl(vp); mtx_lock(&sync_mtx); return (0); } /* * System filesystem synchronizer daemon. */ static void sched_sync(void) { struct synclist *next; struct synclist *slp; struct bufobj *bo; long starttime; struct thread *td = FIRST_THREAD_IN_PROC(updateproc); static int dummychan; int last_work_seen; int net_worklist_len; int syncer_final_iter; int first_printf; int error; mtx_lock(&Giant); last_work_seen = 0; syncer_final_iter = 0; first_printf = 1; syncer_state = SYNCER_RUNNING; starttime = time_uptime; td->td_pflags |= TDP_NORUNNINGBUF; EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, SHUTDOWN_PRI_LAST); for (;;) { mtx_lock(&sync_mtx); if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter == 0) { mtx_unlock(&sync_mtx); kthread_suspend_check(td->td_proc); mtx_lock(&sync_mtx); } net_worklist_len = syncer_worklist_len - sync_vnode_count; if (syncer_state != SYNCER_RUNNING && starttime != time_uptime) { if (first_printf) { printf("\nSyncing disks, vnodes remaining..."); first_printf = 0; } printf("%d ", net_worklist_len); } starttime = time_uptime; /* * Push files whose dirty time has expired. Be careful * of interrupt race on slp queue. * * Skip over empty worklist slots when shutting down. */ do { slp = &syncer_workitem_pending[syncer_delayno]; syncer_delayno += 1; if (syncer_delayno == syncer_maxdelay) syncer_delayno = 0; next = &syncer_workitem_pending[syncer_delayno]; /* * If the worklist has wrapped since the * it was emptied of all but syncer vnodes, * switch to the FINAL_DELAY state and run * for one more second. */ if (syncer_state == SYNCER_SHUTTING_DOWN && net_worklist_len == 0 && last_work_seen == syncer_delayno) { syncer_state = SYNCER_FINAL_DELAY; syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; } } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && syncer_worklist_len > 0); /* * Keep track of the last time there was anything * on the worklist other than syncer vnodes. * Return to the SHUTTING_DOWN state if any * new work appears. */ if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) last_work_seen = syncer_delayno; if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) syncer_state = SYNCER_SHUTTING_DOWN; while ((bo = LIST_FIRST(slp)) != NULL) { error = sync_vnode(bo, td); if (error == 1) { LIST_REMOVE(bo, bo_synclist); LIST_INSERT_HEAD(next, bo, bo_synclist); continue; } } if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) syncer_final_iter--; mtx_unlock(&sync_mtx); /* * The variable rushjob allows the kernel to speed up the * processing of the filesystem syncer process. A rushjob * value of N tells the filesystem syncer to process the next * N seconds worth of work on its queue ASAP. Currently rushjob * is used by the soft update code to speed up the filesystem * syncer process when the incore state is getting so far * ahead of the disk that the kernel memory pool is being * threatened with exhaustion. */ mtx_lock(&sync_mtx); if (rushjob > 0) { rushjob -= 1; mtx_unlock(&sync_mtx); continue; } mtx_unlock(&sync_mtx); /* * Just sleep for a short period if time between * iterations when shutting down to allow some I/O * to happen. * * If it has taken us less than a second to process the * current work, then wait. Otherwise start right over * again. We can still lose time if any single round * takes more than two seconds, but it does not really * matter as we are just trying to generally pace the * filesystem activity. */ if (syncer_state != SYNCER_RUNNING) tsleep(&dummychan, PPAUSE, "syncfnl", hz / SYNCER_SHUTDOWN_SPEEDUP); else if (time_uptime == starttime) tsleep(&lbolt, PPAUSE, "syncer", 0); } } /* * Request the syncer daemon to speed up its work. * We never push it to speed up more than half of its * normal turn time, otherwise it could take over the cpu. */ int speedup_syncer() { struct thread *td; int ret = 0; td = FIRST_THREAD_IN_PROC(updateproc); sleepq_remove(td, &lbolt); mtx_lock(&sync_mtx); if (rushjob < syncdelay / 2) { rushjob += 1; stat_rush_requests += 1; ret = 1; } mtx_unlock(&sync_mtx); return (ret); } /* * Tell the syncer to speed up its work and run though its work * list several times, then tell it to shut down. */ static void syncer_shutdown(void *arg, int howto) { struct thread *td; if (howto & RB_NOSYNC) return; td = FIRST_THREAD_IN_PROC(updateproc); sleepq_remove(td, &lbolt); mtx_lock(&sync_mtx); syncer_state = SYNCER_SHUTTING_DOWN; rushjob = 0; mtx_unlock(&sync_mtx); kproc_shutdown(arg, howto); } /* * Reassign a buffer from one vnode to another. * Used to assign file specific control information * (indirect blocks) to the vnode to which they belong. */ void reassignbuf(struct buf *bp) { struct vnode *vp; struct bufobj *bo; int delay; #ifdef INVARIANTS struct bufv *bv; #endif vp = bp->b_vp; bo = bp->b_bufobj; ++reassignbufcalls; CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); /* * B_PAGING flagged buffers cannot be reassigned because their vp * is not fully linked in. */ if (bp->b_flags & B_PAGING) panic("cannot reassign paging buffer"); /* * Delete from old vnode list, if on one. */ VI_LOCK(vp); if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) buf_vlist_remove(bp); else panic("reassignbuf: Buffer %p not on queue.", bp); /* * If dirty, put on list of dirty buffers; otherwise insert onto list * of clean buffers. */ if (bp->b_flags & B_DELWRI) { if ((bo->bo_flag & BO_ONWORKLST) == 0) { switch (vp->v_type) { case VDIR: delay = dirdelay; break; case VCHR: delay = metadelay; break; default: delay = filedelay; } vn_syncer_add_to_worklist(bo, delay); } buf_vlist_add(bp, bo, BX_VNDIRTY); } else { buf_vlist_add(bp, bo, BX_VNCLEAN); if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; mtx_unlock(&sync_mtx); bo->bo_flag &= ~BO_ONWORKLST; } } #ifdef INVARIANTS bv = &bo->bo_clean; bp = TAILQ_FIRST(&bv->bv_hd); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bp = TAILQ_LAST(&bv->bv_hd, buflists); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bv = &bo->bo_dirty; bp = TAILQ_FIRST(&bv->bv_hd); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bp = TAILQ_LAST(&bv->bv_hd, buflists); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); #endif VI_UNLOCK(vp); } /* * Increment the use and hold counts on the vnode, taking care to reference * the driver's usecount if this is a chardev. The vholdl() will remove * the vnode from the free list if it is presently free. Requires the * vnode interlock and returns with it held. */ static void v_incr_usecount(struct vnode *vp) { CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n", vp, vp->v_holdcnt, vp->v_usecount); vp->v_usecount++; if (vp->v_type == VCHR && vp->v_rdev != NULL) { dev_lock(); vp->v_rdev->si_usecount++; dev_unlock(); } vholdl(vp); } /* * Turn a holdcnt into a use+holdcnt such that only one call to * v_decr_usecount is needed. */ static void v_upgrade_usecount(struct vnode *vp) { CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n", vp, vp->v_holdcnt, vp->v_usecount); vp->v_usecount++; if (vp->v_type == VCHR && vp->v_rdev != NULL) { dev_lock(); vp->v_rdev->si_usecount++; dev_unlock(); } } /* * Decrement the vnode use and hold count along with the driver's usecount * if this is a chardev. The vdropl() below releases the vnode interlock * as it may free the vnode. */ static void v_decr_usecount(struct vnode *vp) { CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n", vp, vp->v_holdcnt, vp->v_usecount); ASSERT_VI_LOCKED(vp, __FUNCTION__); VNASSERT(vp->v_usecount > 0, vp, ("v_decr_usecount: negative usecount")); vp->v_usecount--; if (vp->v_type == VCHR && vp->v_rdev != NULL) { dev_lock(); vp->v_rdev->si_usecount--; dev_unlock(); } vdropl(vp); } /* * Decrement only the use count and driver use count. This is intended to * be paired with a follow on vdropl() to release the remaining hold count. * In this way we may vgone() a vnode with a 0 usecount without risk of * having it end up on a free list because the hold count is kept above 0. */ static void v_decr_useonly(struct vnode *vp) { CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n", vp, vp->v_holdcnt, vp->v_usecount); ASSERT_VI_LOCKED(vp, __FUNCTION__); VNASSERT(vp->v_usecount > 0, vp, ("v_decr_useonly: negative usecount")); vp->v_usecount--; if (vp->v_type == VCHR && vp->v_rdev != NULL) { dev_lock(); vp->v_rdev->si_usecount--; dev_unlock(); } } /* * Grab a particular vnode from the free list, increment its * reference count and lock it. The vnode lock bit is set if the * vnode is being eliminated in vgone. The process is awakened * when the transition is completed, and an error returned to * indicate that the vnode is no longer usable (possibly having * been changed to a new filesystem type). */ int vget(struct vnode *vp, int flags, struct thread *td) { int oweinact; int oldflags; int error; error = 0; oldflags = flags; oweinact = 0; VFS_ASSERT_GIANT(vp->v_mount); if ((flags & LK_INTERLOCK) == 0) VI_LOCK(vp); /* * If the inactive call was deferred because vput() was called * with a shared lock, we have to do it here before another thread * gets a reference to data that should be dead. */ if (vp->v_iflag & VI_OWEINACT) { if (flags & LK_NOWAIT) { VI_UNLOCK(vp); return (EBUSY); } flags &= ~LK_TYPE_MASK; flags |= LK_EXCLUSIVE; oweinact = 1; } vholdl(vp); if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { vdrop(vp); return (error); } VI_LOCK(vp); /* Upgrade our holdcnt to a usecount. */ v_upgrade_usecount(vp); if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) panic("vget: vn_lock failed to return ENOENT\n"); if (oweinact) { if (vp->v_iflag & VI_OWEINACT) vinactive(vp, td); VI_UNLOCK(vp); if ((oldflags & LK_TYPE_MASK) == 0) VOP_UNLOCK(vp, 0, td); } else VI_UNLOCK(vp); return (0); } /* * Increase the reference count of a vnode. */ void vref(struct vnode *vp) { VI_LOCK(vp); v_incr_usecount(vp); VI_UNLOCK(vp); } /* * Return reference count of a vnode. * * The results of this call are only guaranteed when some mechanism other * than the VI lock is used to stop other processes from gaining references * to the vnode. This may be the case if the caller holds the only reference. * This is also useful when stale data is acceptable as race conditions may * be accounted for by some other means. */ int vrefcnt(struct vnode *vp) { int usecnt; VI_LOCK(vp); usecnt = vp->v_usecount; VI_UNLOCK(vp); return (usecnt); } /* * Vnode put/release. * If count drops to zero, call inactive routine and return to freelist. */ void vrele(struct vnode *vp) { struct thread *td = curthread; /* XXX */ KASSERT(vp != NULL, ("vrele: null vp")); VFS_ASSERT_GIANT(vp->v_mount); VI_LOCK(vp); /* Skip this v_writecount check if we're going to panic below. */ VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, ("vrele: missed vn_close")); if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && vp->v_usecount == 1)) { v_decr_usecount(vp); return; } if (vp->v_usecount != 1) { #ifdef DIAGNOSTIC vprint("vrele: negative ref count", vp); #endif VI_UNLOCK(vp); panic("vrele: negative ref cnt"); } /* * We want to hold the vnode until the inactive finishes to * prevent vgone() races. We drop the use count here and the * hold count below when we're done. */ v_decr_useonly(vp); /* * We must call VOP_INACTIVE with the node locked. Mark * as VI_DOINGINACT to avoid recursion. */ vp->v_iflag |= VI_OWEINACT; if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { VI_LOCK(vp); if (vp->v_usecount > 0) vp->v_iflag &= ~VI_OWEINACT; if (vp->v_iflag & VI_OWEINACT) vinactive(vp, td); VOP_UNLOCK(vp, 0, td); } else { VI_LOCK(vp); if (vp->v_usecount > 0) vp->v_iflag &= ~VI_OWEINACT; } vdropl(vp); } /* * Release an already locked vnode. This give the same effects as * unlock+vrele(), but takes less time and avoids releasing and * re-aquiring the lock (as vrele() aquires the lock internally.) */ void vput(struct vnode *vp) { struct thread *td = curthread; /* XXX */ int error; KASSERT(vp != NULL, ("vput: null vp")); ASSERT_VOP_LOCKED(vp, "vput"); VFS_ASSERT_GIANT(vp->v_mount); VI_LOCK(vp); /* Skip this v_writecount check if we're going to panic below. */ VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, ("vput: missed vn_close")); error = 0; if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && vp->v_usecount == 1)) { VOP_UNLOCK(vp, 0, td); v_decr_usecount(vp); return; } if (vp->v_usecount != 1) { #ifdef DIAGNOSTIC vprint("vput: negative ref count", vp); #endif panic("vput: negative ref cnt"); } /* * We want to hold the vnode until the inactive finishes to * prevent vgone() races. We drop the use count here and the * hold count below when we're done. */ v_decr_useonly(vp); vp->v_iflag |= VI_OWEINACT; if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) { error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td); VI_LOCK(vp); if (error) { if (vp->v_usecount > 0) vp->v_iflag &= ~VI_OWEINACT; goto done; } } if (vp->v_usecount > 0) vp->v_iflag &= ~VI_OWEINACT; if (vp->v_iflag & VI_OWEINACT) vinactive(vp, td); VOP_UNLOCK(vp, 0, td); done: vdropl(vp); } /* * Somebody doesn't want the vnode recycled. */ void vhold(struct vnode *vp) { VI_LOCK(vp); vholdl(vp); VI_UNLOCK(vp); } void vholdl(struct vnode *vp) { vp->v_holdcnt++; if (VSHOULDBUSY(vp)) vbusy(vp); } /* * Note that there is one less who cares about this vnode. vdrop() is the * opposite of vhold(). */ void vdrop(struct vnode *vp) { VI_LOCK(vp); vdropl(vp); } /* * Drop the hold count of the vnode. If this is the last reference to * the vnode we will free it if it has been vgone'd otherwise it is * placed on the free list. */ static void vdropl(struct vnode *vp) { if (vp->v_holdcnt <= 0) panic("vdrop: holdcnt %d", vp->v_holdcnt); vp->v_holdcnt--; if (vp->v_holdcnt == 0) { if (vp->v_iflag & VI_DOOMED) { vdestroy(vp); return; } else vfree(vp); } VI_UNLOCK(vp); } /* * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT * flags. DOINGINACT prevents us from recursing in calls to vinactive. * OWEINACT tracks whether a vnode missed a call to inactive due to a * failed lock upgrade. */ static void vinactive(struct vnode *vp, struct thread *td) { ASSERT_VOP_LOCKED(vp, "vinactive"); ASSERT_VI_LOCKED(vp, "vinactive"); VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, ("vinactive: recursed on VI_DOINGINACT")); vp->v_iflag |= VI_DOINGINACT; vp->v_iflag &= ~VI_OWEINACT; VI_UNLOCK(vp); VOP_INACTIVE(vp, td); VI_LOCK(vp); VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, ("vinactive: lost VI_DOINGINACT")); vp->v_iflag &= ~VI_DOINGINACT; } /* * Remove any vnodes in the vnode table belonging to mount point mp. * * If FORCECLOSE is not specified, there should not be any active ones, * return error if any are found (nb: this is a user error, not a * system error). If FORCECLOSE is specified, detach any active vnodes * that are found. * * If WRITECLOSE is set, only flush out regular file vnodes open for * writing. * * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. * * `rootrefs' specifies the base reference count for the root vnode * of this filesystem. The root vnode is considered busy if its * v_usecount exceeds this value. On a successful return, vflush(, td) * will call vrele() on the root vnode exactly rootrefs times. * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must * be zero. */ #ifdef DIAGNOSTIC static int busyprt = 0; /* print out busy vnodes */ SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); #endif int vflush( struct mount *mp, int rootrefs, int flags, struct thread *td) { struct vnode *vp, *mvp, *rootvp = NULL; struct vattr vattr; int busy = 0, error; CTR1(KTR_VFS, "vflush: mp %p", mp); if (rootrefs > 0) { KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, ("vflush: bad args")); /* * Get the filesystem root vnode. We can vput() it * immediately, since with rootrefs > 0, it won't go away. */ if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0) return (error); vput(rootvp); } MNT_ILOCK(mp); loop: MNT_VNODE_FOREACH(vp, mp, mvp) { VI_LOCK(vp); vholdl(vp); MNT_IUNLOCK(mp); error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); if (error) { vdrop(vp); MNT_ILOCK(mp); MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp); goto loop; } /* * Skip over a vnodes marked VV_SYSTEM. */ if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { VOP_UNLOCK(vp, 0, td); vdrop(vp); MNT_ILOCK(mp); continue; } /* * If WRITECLOSE is set, flush out unlinked but still open * files (even if open only for reading) and regular file * vnodes open for writing. */ if (flags & WRITECLOSE) { error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); VI_LOCK(vp); if ((vp->v_type == VNON || (error == 0 && vattr.va_nlink > 0)) && (vp->v_writecount == 0 || vp->v_type != VREG)) { VOP_UNLOCK(vp, 0, td); vdropl(vp); MNT_ILOCK(mp); continue; } } else VI_LOCK(vp); /* * With v_usecount == 0, all we need to do is clear out the * vnode data structures and we are done. * * If FORCECLOSE is set, forcibly close the vnode. */ if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { VNASSERT(vp->v_usecount == 0 || (vp->v_type != VCHR && vp->v_type != VBLK), vp, ("device VNODE %p is FORCECLOSED", vp)); vgonel(vp); } else { busy++; #ifdef DIAGNOSTIC if (busyprt) vprint("vflush: busy vnode", vp); #endif } VOP_UNLOCK(vp, 0, td); vdropl(vp); MNT_ILOCK(mp); } MNT_IUNLOCK(mp); if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { /* * If just the root vnode is busy, and if its refcount * is equal to `rootrefs', then go ahead and kill it. */ VI_LOCK(rootvp); KASSERT(busy > 0, ("vflush: not busy")); VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, ("vflush: usecount %d < rootrefs %d", rootvp->v_usecount, rootrefs)); if (busy == 1 && rootvp->v_usecount == rootrefs) { VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td); vgone(rootvp); VOP_UNLOCK(rootvp, 0, td); busy = 0; } else VI_UNLOCK(rootvp); } if (busy) return (EBUSY); for (; rootrefs > 0; rootrefs--) vrele(rootvp); return (0); } /* * Recycle an unused vnode to the front of the free list. */ int vrecycle(struct vnode *vp, struct thread *td) { int recycled; ASSERT_VOP_LOCKED(vp, "vrecycle"); recycled = 0; VI_LOCK(vp); if (vp->v_usecount == 0) { recycled = 1; vgonel(vp); } VI_UNLOCK(vp); return (recycled); } /* * Eliminate all activity associated with a vnode * in preparation for reuse. */ void vgone(struct vnode *vp) { VI_LOCK(vp); vgonel(vp); VI_UNLOCK(vp); } /* * vgone, with the vp interlock held. */ void vgonel(struct vnode *vp) { struct thread *td; int oweinact; int active; struct mount *mp; CTR1(KTR_VFS, "vgonel: vp %p", vp); ASSERT_VOP_LOCKED(vp, "vgonel"); ASSERT_VI_LOCKED(vp, "vgonel"); #if 0 /* XXX Need to fix ttyvp before I enable this. */ VNASSERT(vp->v_holdcnt, vp, ("vgonel: vp %p has no reference.", vp)); #endif td = curthread; /* * Don't vgonel if we're already doomed. */ if (vp->v_iflag & VI_DOOMED) return; vp->v_iflag |= VI_DOOMED; /* * Check to see if the vnode is in use. If so, we have to call * VOP_CLOSE() and VOP_INACTIVE(). */ active = vp->v_usecount; oweinact = (vp->v_iflag & VI_OWEINACT); VI_UNLOCK(vp); /* * Clean out any buffers associated with the vnode. * If the flush fails, just toss the buffers. */ mp = NULL; if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) (void) vn_start_secondary_write(vp, &mp, V_WAIT); if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0) vinvalbuf(vp, 0, td, 0, 0); /* * If purging an active vnode, it must be closed and * deactivated before being reclaimed. */ if (active) VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); if (oweinact || active) { VI_LOCK(vp); if ((vp->v_iflag & VI_DOINGINACT) == 0) vinactive(vp, td); VI_UNLOCK(vp); } /* * Reclaim the vnode. */ if (VOP_RECLAIM(vp, td)) panic("vgone: cannot reclaim"); if (mp != NULL) vn_finished_secondary_write(mp); VNASSERT(vp->v_object == NULL, vp, ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); /* * Delete from old mount point vnode list. */ delmntque(vp); cache_purge(vp); /* * Done with purge, reset to the standard lock and invalidate * the vnode. */ VI_LOCK(vp); vp->v_vnlock = &vp->v_lock; vp->v_op = &dead_vnodeops; vp->v_tag = "none"; vp->v_type = VBAD; } /* * Calculate the total number of references to a special device. */ int vcount(struct vnode *vp) { int count; dev_lock(); count = vp->v_rdev->si_usecount; dev_unlock(); return (count); } /* * Same as above, but using the struct cdev *as argument */ int count_dev(struct cdev *dev) { int count; dev_lock(); count = dev->si_usecount; dev_unlock(); return(count); } /* * Print out a description of a vnode. */ static char *typename[] = {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", "VMARKER"}; void vn_printf(struct vnode *vp, const char *fmt, ...) { va_list ap; char buf[96]; va_start(ap, fmt); vprintf(fmt, ap); va_end(ap); printf("%p: ", (void *)vp); printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); buf[0] = '\0'; buf[1] = '\0'; if (vp->v_vflag & VV_ROOT) strcat(buf, "|VV_ROOT"); if (vp->v_vflag & VV_TEXT) strcat(buf, "|VV_TEXT"); if (vp->v_vflag & VV_SYSTEM) strcat(buf, "|VV_SYSTEM"); if (vp->v_iflag & VI_DOOMED) strcat(buf, "|VI_DOOMED"); if (vp->v_iflag & VI_FREE) strcat(buf, "|VI_FREE"); printf(" flags (%s)\n", buf + 1); if (mtx_owned(VI_MTX(vp))) printf(" VI_LOCKed"); if (vp->v_object != NULL) printf(" v_object %p ref %d pages %d\n", vp->v_object, vp->v_object->ref_count, vp->v_object->resident_page_count); printf(" "); lockmgr_printinfo(vp->v_vnlock); printf("\n"); if (vp->v_data != NULL) VOP_PRINT(vp); } #ifdef DDB #include /* * List all of the locked vnodes in the system. * Called when debugging the kernel. */ DB_SHOW_COMMAND(lockedvnods, lockedvnodes) { struct mount *mp, *nmp; struct vnode *vp; /* * Note: because this is DDB, we can't obey the locking semantics * for these structures, which means we could catch an inconsistent * state and dereference a nasty pointer. Not much to be done * about that. */ printf("Locked vnodes\n"); for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { nmp = TAILQ_NEXT(mp, mnt_list); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL)) vprint("", vp); } nmp = TAILQ_NEXT(mp, mnt_list); } } #endif /* * Fill in a struct xvfsconf based on a struct vfsconf. */ static void vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) { strcpy(xvfsp->vfc_name, vfsp->vfc_name); xvfsp->vfc_typenum = vfsp->vfc_typenum; xvfsp->vfc_refcount = vfsp->vfc_refcount; xvfsp->vfc_flags = vfsp->vfc_flags; /* * These are unused in userland, we keep them * to not break binary compatibility. */ xvfsp->vfc_vfsops = NULL; xvfsp->vfc_next = NULL; } /* * Top level filesystem related information gathering. */ static int sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) { struct vfsconf *vfsp; struct xvfsconf xvfsp; int error; error = 0; TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { bzero(&xvfsp, sizeof(xvfsp)); vfsconf2x(vfsp, &xvfsp); error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); if (error) break; } return (error); } SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, "S,xvfsconf", "List of all configured filesystems"); #ifndef BURN_BRIDGES static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); static int vfs_sysctl(SYSCTL_HANDLER_ARGS) { int *name = (int *)arg1 - 1; /* XXX */ u_int namelen = arg2 + 1; /* XXX */ struct vfsconf *vfsp; struct xvfsconf xvfsp; printf("WARNING: userland calling deprecated sysctl, " "please rebuild world\n"); #if 1 || defined(COMPAT_PRELITE2) /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ if (namelen == 1) return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); #endif switch (name[1]) { case VFS_MAXTYPENUM: if (namelen != 2) return (ENOTDIR); return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); case VFS_CONF: if (namelen != 3) return (ENOTDIR); /* overloaded */ TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) if (vfsp->vfc_typenum == name[2]) break; if (vfsp == NULL) return (EOPNOTSUPP); bzero(&xvfsp, sizeof(xvfsp)); vfsconf2x(vfsp, &xvfsp); return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); } return (EOPNOTSUPP); } static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, "Generic filesystem"); #if 1 || defined(COMPAT_PRELITE2) static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) { int error; struct vfsconf *vfsp; struct ovfsconf ovfs; TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { bzero(&ovfs, sizeof(ovfs)); ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ strcpy(ovfs.vfc_name, vfsp->vfc_name); ovfs.vfc_index = vfsp->vfc_typenum; ovfs.vfc_refcount = vfsp->vfc_refcount; ovfs.vfc_flags = vfsp->vfc_flags; error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); if (error) return error; } return 0; } #endif /* 1 || COMPAT_PRELITE2 */ #endif /* !BURN_BRIDGES */ #define KINFO_VNODESLOP 10 #ifdef notyet /* * Dump vnode list (via sysctl). */ /* ARGSUSED */ static int sysctl_vnode(SYSCTL_HANDLER_ARGS) { struct xvnode *xvn; struct thread *td = req->td; struct mount *mp; struct vnode *vp; int error, len, n; /* * Stale numvnodes access is not fatal here. */ req->lock = 0; len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; if (!req->oldptr) /* Make an estimate */ return (SYSCTL_OUT(req, 0, len)); error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); n = 0; mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) continue; MNT_ILOCK(mp); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (n == len) break; vref(vp); xvn[n].xv_size = sizeof *xvn; xvn[n].xv_vnode = vp; xvn[n].xv_id = 0; /* XXX compat */ #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field XV_COPY(usecount); XV_COPY(writecount); XV_COPY(holdcnt); XV_COPY(mount); XV_COPY(numoutput); XV_COPY(type); #undef XV_COPY xvn[n].xv_flag = vp->v_vflag; switch (vp->v_type) { case VREG: case VDIR: case VLNK: break; case VBLK: case VCHR: if (vp->v_rdev == NULL) { vrele(vp); continue; } xvn[n].xv_dev = dev2udev(vp->v_rdev); break; case VSOCK: xvn[n].xv_socket = vp->v_socket; break; case VFIFO: xvn[n].xv_fifo = vp->v_fifoinfo; break; case VNON: case VBAD: default: /* shouldn't happen? */ vrele(vp); continue; } vrele(vp); ++n; } MNT_IUNLOCK(mp); mtx_lock(&mountlist_mtx); vfs_unbusy(mp, td); if (n == len) break; } mtx_unlock(&mountlist_mtx); error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); free(xvn, M_TEMP); return (error); } SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 0, 0, sysctl_vnode, "S,xvnode", ""); #endif /* * Unmount all filesystems. The list is traversed in reverse order * of mounting to avoid dependencies. */ void vfs_unmountall(void) { struct mount *mp; struct thread *td; int error; KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread")); td = curthread; /* * Since this only runs when rebooting, it is not interlocked. */ while(!TAILQ_EMPTY(&mountlist)) { mp = TAILQ_LAST(&mountlist, mntlist); error = dounmount(mp, MNT_FORCE, td); if (error) { TAILQ_REMOVE(&mountlist, mp, mnt_list); /* * XXX: Due to the way in which we mount the root * file system off of devfs, devfs will generate a * "busy" warning when we try to unmount it before * the root. Don't print a warning as a result in * order to avoid false positive errors that may * cause needless upset. */ if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) { printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); if (error == EBUSY) printf("BUSY)\n"); else printf("%d)\n", error); } } else { /* The unmount has removed mp from the mountlist */ } } } /* * perform msync on all vnodes under a mount point * the mount point must be locked. */ void vfs_msync(struct mount *mp, int flags) { struct vnode *vp, *mvp; struct vm_object *obj; (void) vn_start_write(NULL, &mp, V_WAIT); MNT_ILOCK(mp); MNT_VNODE_FOREACH(vp, mp, mvp) { VI_LOCK(vp); if ((vp->v_iflag & VI_OBJDIRTY) && (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { MNT_IUNLOCK(mp); if (!vget(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, curthread)) { if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ vput(vp); MNT_ILOCK(mp); continue; } obj = vp->v_object; if (obj != NULL) { VM_OBJECT_LOCK(obj); vm_object_page_clean(obj, 0, 0, flags == MNT_WAIT ? OBJPC_SYNC : OBJPC_NOSYNC); VM_OBJECT_UNLOCK(obj); } vput(vp); } MNT_ILOCK(mp); } else VI_UNLOCK(vp); } MNT_IUNLOCK(mp); vn_finished_write(mp); } /* * Mark a vnode as free, putting it up for recycling. */ static void vfree(struct vnode *vp) { CTR1(KTR_VFS, "vfree vp %p", vp); ASSERT_VI_LOCKED(vp, "vfree"); mtx_lock(&vnode_free_list_mtx); VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed.")); VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free")); VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't")); VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp, ("vfree: Freeing doomed vnode")); if (vp->v_iflag & VI_AGE) { TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); } else { TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); } freevnodes++; vp->v_iflag &= ~VI_AGE; vp->v_iflag |= VI_FREE; mtx_unlock(&vnode_free_list_mtx); } /* * Opposite of vfree() - mark a vnode as in use. */ static void vbusy(struct vnode *vp) { CTR1(KTR_VFS, "vbusy vp %p", vp); ASSERT_VI_LOCKED(vp, "vbusy"); VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free")); VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed.")); mtx_lock(&vnode_free_list_mtx); TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); freevnodes--; vp->v_iflag &= ~(VI_FREE|VI_AGE); mtx_unlock(&vnode_free_list_mtx); } /* * Initalize per-vnode helper structure to hold poll-related state. */ void v_addpollinfo(struct vnode *vp) { struct vpollinfo *vi; vi = uma_zalloc(vnodepoll_zone, M_WAITOK); if (vp->v_pollinfo != NULL) { uma_zfree(vnodepoll_zone, vi); return; } vp->v_pollinfo = vi; mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock, vfs_knlunlock, vfs_knllocked); } /* * Record a process's interest in events which might happen to * a vnode. Because poll uses the historic select-style interface * internally, this routine serves as both the ``check for any * pending events'' and the ``record my interest in future events'' * functions. (These are done together, while the lock is held, * to avoid race conditions.) */ int vn_pollrecord(struct vnode *vp, struct thread *td, int events) { if (vp->v_pollinfo == NULL) v_addpollinfo(vp); mtx_lock(&vp->v_pollinfo->vpi_lock); if (vp->v_pollinfo->vpi_revents & events) { /* * This leaves events we are not interested * in available for the other process which * which presumably had requested them * (otherwise they would never have been * recorded). */ events &= vp->v_pollinfo->vpi_revents; vp->v_pollinfo->vpi_revents &= ~events; mtx_unlock(&vp->v_pollinfo->vpi_lock); return events; } vp->v_pollinfo->vpi_events |= events; selrecord(td, &vp->v_pollinfo->vpi_selinfo); mtx_unlock(&vp->v_pollinfo->vpi_lock); return 0; } /* * Routine to create and manage a filesystem syncer vnode. */ #define sync_close ((int (*)(struct vop_close_args *))nullop) static int sync_fsync(struct vop_fsync_args *); static int sync_inactive(struct vop_inactive_args *); static int sync_reclaim(struct vop_reclaim_args *); static struct vop_vector sync_vnodeops = { .vop_bypass = VOP_EOPNOTSUPP, .vop_close = sync_close, /* close */ .vop_fsync = sync_fsync, /* fsync */ .vop_inactive = sync_inactive, /* inactive */ .vop_reclaim = sync_reclaim, /* reclaim */ .vop_lock = vop_stdlock, /* lock */ .vop_unlock = vop_stdunlock, /* unlock */ .vop_islocked = vop_stdislocked, /* islocked */ }; /* * Create a new filesystem syncer vnode for the specified mount point. */ int vfs_allocate_syncvnode(struct mount *mp) { struct vnode *vp; static long start, incr, next; int error; /* Allocate a new vnode */ if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) { mp->mnt_syncer = NULL; return (error); } vp->v_type = VNON; /* * Place the vnode onto the syncer worklist. We attempt to * scatter them about on the list so that they will go off * at evenly distributed times even if all the filesystems * are mounted at once. */ next += incr; if (next == 0 || next > syncer_maxdelay) { start /= 2; incr /= 2; if (start == 0) { start = syncer_maxdelay / 2; incr = syncer_maxdelay; } next = start; } VI_LOCK(vp); vn_syncer_add_to_worklist(&vp->v_bufobj, syncdelay > 0 ? next % syncdelay : 0); /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ mtx_lock(&sync_mtx); sync_vnode_count++; mtx_unlock(&sync_mtx); VI_UNLOCK(vp); mp->mnt_syncer = vp; return (0); } /* * Do a lazy sync of the filesystem. */ static int sync_fsync(struct vop_fsync_args *ap) { struct vnode *syncvp = ap->a_vp; struct mount *mp = syncvp->v_mount; struct thread *td = ap->a_td; int error, asyncflag; struct bufobj *bo; /* * We only need to do something if this is a lazy evaluation. */ if (ap->a_waitfor != MNT_LAZY) return (0); /* * Move ourselves to the back of the sync list. */ bo = &syncvp->v_bufobj; BO_LOCK(bo); vn_syncer_add_to_worklist(bo, syncdelay); BO_UNLOCK(bo); /* * Walk the list of vnodes pushing all that are dirty and * not already on the sync list. */ mtx_lock(&mountlist_mtx); if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { mtx_unlock(&mountlist_mtx); return (0); } if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { vfs_unbusy(mp, td); return (0); } asyncflag = mp->mnt_flag & MNT_ASYNC; mp->mnt_flag &= ~MNT_ASYNC; vfs_msync(mp, MNT_NOWAIT); error = VFS_SYNC(mp, MNT_LAZY, td); if (asyncflag) mp->mnt_flag |= MNT_ASYNC; vn_finished_write(mp); vfs_unbusy(mp, td); return (error); } /* * The syncer vnode is no referenced. */ static int sync_inactive(struct vop_inactive_args *ap) { vgone(ap->a_vp); return (0); } /* * The syncer vnode is no longer needed and is being decommissioned. * * Modifications to the worklist must be protected by sync_mtx. */ static int sync_reclaim(struct vop_reclaim_args *ap) { struct vnode *vp = ap->a_vp; struct bufobj *bo; VI_LOCK(vp); bo = &vp->v_bufobj; vp->v_mount->mnt_syncer = NULL; if (bo->bo_flag & BO_ONWORKLST) { mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; sync_vnode_count--; mtx_unlock(&sync_mtx); bo->bo_flag &= ~BO_ONWORKLST; } VI_UNLOCK(vp); return (0); } /* * Check if vnode represents a disk device */ int vn_isdisk(struct vnode *vp, int *errp) { int error; error = 0; dev_lock(); if (vp->v_type != VCHR) error = ENOTBLK; else if (vp->v_rdev == NULL) error = ENXIO; else if (vp->v_rdev->si_devsw == NULL) error = ENXIO; else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) error = ENOTBLK; dev_unlock(); if (errp != NULL) *errp = error; return (error == 0); } /* * Common filesystem object access control check routine. Accepts a * vnode's type, "mode", uid and gid, requested access mode, credentials, * and optional call-by-reference privused argument allowing vaccess() * to indicate to the caller whether privilege was used to satisfy the * request (obsoleted). Returns 0 on success, or an errno on failure. */ int vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, mode_t acc_mode, struct ucred *cred, int *privused) { mode_t dac_granted; #ifdef CAPABILITIES mode_t cap_granted; #endif /* * Look for a normal, non-privileged way to access the file/directory * as requested. If it exists, go with that. */ if (privused != NULL) *privused = 0; dac_granted = 0; /* Check the owner. */ if (cred->cr_uid == file_uid) { dac_granted |= VADMIN; if (file_mode & S_IXUSR) dac_granted |= VEXEC; if (file_mode & S_IRUSR) dac_granted |= VREAD; if (file_mode & S_IWUSR) dac_granted |= (VWRITE | VAPPEND); if ((acc_mode & dac_granted) == acc_mode) return (0); goto privcheck; } /* Otherwise, check the groups (first match) */ if (groupmember(file_gid, cred)) { if (file_mode & S_IXGRP) dac_granted |= VEXEC; if (file_mode & S_IRGRP) dac_granted |= VREAD; if (file_mode & S_IWGRP) dac_granted |= (VWRITE | VAPPEND); if ((acc_mode & dac_granted) == acc_mode) return (0); goto privcheck; } /* Otherwise, check everyone else. */ if (file_mode & S_IXOTH) dac_granted |= VEXEC; if (file_mode & S_IROTH) dac_granted |= VREAD; if (file_mode & S_IWOTH) dac_granted |= (VWRITE | VAPPEND); if ((acc_mode & dac_granted) == acc_mode) return (0); privcheck: if (!suser_cred(cred, SUSER_ALLOWJAIL)) { /* XXX audit: privilege used */ if (privused != NULL) *privused = 1; return (0); } #ifdef CAPABILITIES /* * Build a capability mask to determine if the set of capabilities * satisfies the requirements when combined with the granted mask * from above. * For each capability, if the capability is required, bitwise * or the request type onto the cap_granted mask. */ cap_granted = 0; if (type == VDIR) { /* * For directories, use CAP_DAC_READ_SEARCH to satisfy * VEXEC requests, instead of CAP_DAC_EXECUTE. */ if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) cap_granted |= VEXEC; } else { if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL)) cap_granted |= VEXEC; } if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) cap_granted |= VREAD; if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL)) cap_granted |= (VWRITE | VAPPEND); if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL)) cap_granted |= VADMIN; if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { /* XXX audit: privilege used */ if (privused != NULL) *privused = 1; return (0); } #endif return ((acc_mode & VADMIN) ? EPERM : EACCES); } /* * Credential check based on process requesting service, and per-attribute * permissions. */ int extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, struct thread *td, int access) { /* * Kernel-invoked always succeeds. */ if (cred == NOCRED) return (0); /* * Do not allow privileged processes in jail to directly * manipulate system attributes. * * XXX What capability should apply here? * Probably CAP_SYS_SETFFLAG. */ switch (attrnamespace) { case EXTATTR_NAMESPACE_SYSTEM: /* Potentially should be: return (EPERM); */ return (suser_cred(cred, 0)); case EXTATTR_NAMESPACE_USER: return (VOP_ACCESS(vp, access, cred, td)); default: return (EPERM); } } #ifdef DEBUG_VFS_LOCKS /* * This only exists to supress warnings from unlocked specfs accesses. It is * no longer ok to have an unlocked VFS. */ #define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); int vfs_badlock_print = 1; /* Print lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); #ifdef KDB int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); #endif static void vfs_badlock(const char *msg, const char *str, struct vnode *vp) { #ifdef KDB if (vfs_badlock_backtrace) kdb_backtrace(); #endif if (vfs_badlock_print) printf("%s: %p %s\n", str, (void *)vp, msg); if (vfs_badlock_ddb) kdb_enter("lock violation"); } void assert_vi_locked(struct vnode *vp, const char *str) { if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) vfs_badlock("interlock is not locked but should be", str, vp); } void assert_vi_unlocked(struct vnode *vp, const char *str) { if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) vfs_badlock("interlock is locked but should not be", str, vp); } void assert_vop_locked(struct vnode *vp, const char *str) { if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) vfs_badlock("is not locked but should be", str, vp); } void assert_vop_unlocked(struct vnode *vp, const char *str) { if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) vfs_badlock("is locked but should not be", str, vp); } void assert_vop_elocked(struct vnode *vp, const char *str) { if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) vfs_badlock("is not exclusive locked but should be", str, vp); } #if 0 void assert_vop_elocked_other(struct vnode *vp, const char *str) { if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) vfs_badlock("is not exclusive locked by another thread", str, vp); } void assert_vop_slocked(struct vnode *vp, const char *str) { if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, curthread) != LK_SHARED) vfs_badlock("is not locked shared but should be", str, vp); } #endif /* 0 */ #endif /* DEBUG_VFS_LOCKS */ void vop_rename_pre(void *ap) { struct vop_rename_args *a = ap; #ifdef DEBUG_VFS_LOCKS if (a->a_tvp) ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); /* Check the source (from). */ if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp) ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); if (a->a_tvp != a->a_fvp) ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); /* Check the target. */ if (a->a_tvp) ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); #endif if (a->a_tdvp != a->a_fdvp) vhold(a->a_fdvp); if (a->a_tvp != a->a_fvp) vhold(a->a_fvp); vhold(a->a_tdvp); if (a->a_tvp) vhold(a->a_tvp); } void vop_strategy_pre(void *ap) { #ifdef DEBUG_VFS_LOCKS struct vop_strategy_args *a; struct buf *bp; a = ap; bp = a->a_bp; /* * Cluster ops lock their component buffers but not the IO container. */ if ((bp->b_flags & B_CLUSTER) != 0) return; if (BUF_REFCNT(bp) < 1) { if (vfs_badlock_print) printf( "VOP_STRATEGY: bp is not locked but should be\n"); if (vfs_badlock_ddb) kdb_enter("lock violation"); } #endif } void vop_lookup_pre(void *ap) { #ifdef DEBUG_VFS_LOCKS struct vop_lookup_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); #endif } void vop_lookup_post(void *ap, int rc) { #ifdef DEBUG_VFS_LOCKS struct vop_lookup_args *a; struct vnode *dvp; struct vnode *vp; a = ap; dvp = a->a_dvp; vp = *(a->a_vpp); ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); if (!rc) ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)"); #endif } void vop_lock_pre(void *ap) { #ifdef DEBUG_VFS_LOCKS struct vop_lock_args *a = ap; if ((a->a_flags & LK_INTERLOCK) == 0) ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); else ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); #endif } void vop_lock_post(void *ap, int rc) { #ifdef DEBUG_VFS_LOCKS struct vop_lock_args *a = ap; ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); if (rc == 0) ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); #endif } void vop_unlock_pre(void *ap) { #ifdef DEBUG_VFS_LOCKS struct vop_unlock_args *a = ap; if (a->a_flags & LK_INTERLOCK) ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); #endif } void vop_unlock_post(void *ap, int rc) { #ifdef DEBUG_VFS_LOCKS struct vop_unlock_args *a = ap; if (a->a_flags & LK_INTERLOCK) ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); #endif } void vop_create_post(void *ap, int rc) { struct vop_create_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); } void vop_link_post(void *ap, int rc) { struct vop_link_args *a = ap; if (!rc) { VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); } } void vop_mkdir_post(void *ap, int rc) { struct vop_mkdir_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); } void vop_mknod_post(void *ap, int rc) { struct vop_mknod_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); } void vop_remove_post(void *ap, int rc) { struct vop_remove_args *a = ap; if (!rc) { VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); } } void vop_rename_post(void *ap, int rc) { struct vop_rename_args *a = ap; if (!rc) { VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE); VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE); VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); if (a->a_tvp) VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); } if (a->a_tdvp != a->a_fdvp) vdrop(a->a_fdvp); if (a->a_tvp != a->a_fvp) vdrop(a->a_fvp); vdrop(a->a_tdvp); if (a->a_tvp) vdrop(a->a_tvp); } void vop_rmdir_post(void *ap, int rc) { struct vop_rmdir_args *a = ap; if (!rc) { VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); } } void vop_setattr_post(void *ap, int rc) { struct vop_setattr_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); } void vop_symlink_post(void *ap, int rc) { struct vop_symlink_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); } static struct knlist fs_knlist; static void vfs_event_init(void *arg) { knlist_init(&fs_knlist, NULL, NULL, NULL, NULL); } /* XXX - correct order? */ SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); void vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) { KNOTE_UNLOCKED(&fs_knlist, event); } static int filt_fsattach(struct knote *kn); static void filt_fsdetach(struct knote *kn); static int filt_fsevent(struct knote *kn, long hint); struct filterops fs_filtops = { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; static int filt_fsattach(struct knote *kn) { kn->kn_flags |= EV_CLEAR; knlist_add(&fs_knlist, kn, 0); return (0); } static void filt_fsdetach(struct knote *kn) { knlist_remove(&fs_knlist, kn, 0); } static int filt_fsevent(struct knote *kn, long hint) { kn->kn_fflags |= hint; return (kn->kn_fflags != 0); } static int sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) { struct vfsidctl vc; int error; struct mount *mp; error = SYSCTL_IN(req, &vc, sizeof(vc)); if (error) return (error); if (vc.vc_vers != VFS_CTL_VERS1) return (EINVAL); mp = vfs_getvfs(&vc.vc_fsid); if (mp == NULL) return (ENOENT); /* ensure that a specific sysctl goes to the right filesystem. */ if (strcmp(vc.vc_fstypename, "*") != 0 && strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { return (EINVAL); } VCTLTOREQ(&vc, req); return (VFS_SYSCTL(mp, vc.vc_op, req)); } SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); /* * Function to initialize a va_filerev field sensibly. * XXX: Wouldn't a random number make a lot more sense ?? */ u_quad_t init_va_filerev(void) { struct bintime bt; getbinuptime(&bt); return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); } static int filt_vfsread(struct knote *kn, long hint); static int filt_vfswrite(struct knote *kn, long hint); static int filt_vfsvnode(struct knote *kn, long hint); static void filt_vfsdetach(struct knote *kn); static struct filterops vfsread_filtops = { 1, NULL, filt_vfsdetach, filt_vfsread }; static struct filterops vfswrite_filtops = { 1, NULL, filt_vfsdetach, filt_vfswrite }; static struct filterops vfsvnode_filtops = { 1, NULL, filt_vfsdetach, filt_vfsvnode }; static void vfs_knllock(void *arg) { struct vnode *vp = arg; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread); } static void vfs_knlunlock(void *arg) { struct vnode *vp = arg; VOP_UNLOCK(vp, 0, curthread); } static int vfs_knllocked(void *arg) { struct vnode *vp = arg; return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE); } int vfs_kqfilter(struct vop_kqfilter_args *ap) { struct vnode *vp = ap->a_vp; struct knote *kn = ap->a_kn; struct knlist *knl; switch (kn->kn_filter) { case EVFILT_READ: kn->kn_fop = &vfsread_filtops; break; case EVFILT_WRITE: kn->kn_fop = &vfswrite_filtops; break; case EVFILT_VNODE: kn->kn_fop = &vfsvnode_filtops; break; default: return (EINVAL); } kn->kn_hook = (caddr_t)vp; if (vp->v_pollinfo == NULL) v_addpollinfo(vp); if (vp->v_pollinfo == NULL) return (ENOMEM); knl = &vp->v_pollinfo->vpi_selinfo.si_note; knlist_add(knl, kn, 0); return (0); } /* * Detach knote from vnode */ static void filt_vfsdetach(struct knote *kn) { struct vnode *vp = (struct vnode *)kn->kn_hook; KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); } /*ARGSUSED*/ static int filt_vfsread(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; struct vattr va; /* * filesystem is gone, so set the EOF flag and schedule * the knote for deletion. */ if (hint == NOTE_REVOKE) { kn->kn_flags |= (EV_EOF | EV_ONESHOT); return (1); } if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread)) return (0); kn->kn_data = va.va_size - kn->kn_fp->f_offset; return (kn->kn_data != 0); } /*ARGSUSED*/ static int filt_vfswrite(struct knote *kn, long hint) { /* * filesystem is gone, so set the EOF flag and schedule * the knote for deletion. */ if (hint == NOTE_REVOKE) kn->kn_flags |= (EV_EOF | EV_ONESHOT); kn->kn_data = 0; return (1); } static int filt_vfsvnode(struct knote *kn, long hint) { if (kn->kn_sfflags & hint) kn->kn_fflags |= hint; if (hint == NOTE_REVOKE) { kn->kn_flags |= EV_EOF; return (1); } return (kn->kn_fflags != 0); } int vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) { int error; if (dp->d_reclen > ap->a_uio->uio_resid) return (ENAMETOOLONG); error = uiomove(dp, dp->d_reclen, ap->a_uio); if (error) { if (ap->a_ncookies != NULL) { if (ap->a_cookies != NULL) free(ap->a_cookies, M_TEMP); ap->a_cookies = NULL; *ap->a_ncookies = 0; } return (error); } if (ap->a_ncookies == NULL) return (0); KASSERT(ap->a_cookies, ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); *ap->a_cookies = realloc(*ap->a_cookies, (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); (*ap->a_cookies)[*ap->a_ncookies] = off; return (0); } /* * Mark for update the access time of the file if the filesystem * supports VA_MARK_ATIME. This functionality is used by execve * and mmap, so we want to avoid the synchronous I/O implied by * directly setting va_atime for the sake of efficiency. */ void vfs_mark_atime(struct vnode *vp, struct thread *td) { struct vattr atimeattr; if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { VATTR_NULL(&atimeattr); atimeattr.va_vaflags |= VA_MARK_ATIME; (void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td); } } Index: head/sys/sys/buf.h =================================================================== --- head/sys/sys/buf.h (revision 157318) +++ head/sys/sys/buf.h (revision 157319) @@ -1,542 +1,542 @@ /*- * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)buf.h 8.9 (Berkeley) 3/30/95 * $FreeBSD$ */ #ifndef _SYS_BUF_H_ #define _SYS_BUF_H_ #include #include #include #include struct bio; struct buf; struct bufobj; struct mount; struct vnode; struct uio; /* * To avoid including */ LIST_HEAD(workhead, worklist); /* * These are currently used only by the soft dependency code, hence * are stored once in a global variable. If other subsystems wanted * to use these hooks, a pointer to a set of bio_ops could be added * to each buffer. */ extern struct bio_ops { void (*io_start)(struct buf *); void (*io_complete)(struct buf *); void (*io_deallocate)(struct buf *); int (*io_countdeps)(struct buf *, int); } bioops; struct vm_object; typedef unsigned char b_xflags_t; /* * The buffer header describes an I/O operation in the kernel. * * NOTES: * b_bufsize, b_bcount. b_bufsize is the allocation size of the * buffer, either DEV_BSIZE or PAGE_SIZE aligned. b_bcount is the * originally requested buffer size and can serve as a bounds check * against EOF. For most, but not all uses, b_bcount == b_bufsize. * * b_dirtyoff, b_dirtyend. Buffers support piecemeal, unaligned * ranges of dirty data that need to be written to backing store. * The range is typically clipped at b_bcount ( not b_bufsize ). * * b_resid. Number of bytes remaining in I/O. After an I/O operation * completes, b_resid is usually 0 indicating 100% success. * * All fields are protected by the buffer lock except those marked: * V - Protected by owning vnode lock * Q - Protected by the buf queue lock * D - Protected by an dependency implementation specific lock */ struct buf { struct bufobj *b_bufobj; long b_bcount; void *b_caller1; caddr_t b_data; int b_error; uint8_t b_iocmd; uint8_t b_ioflags; off_t b_iooffset; long b_resid; void (*b_iodone)(struct buf *); daddr_t b_blkno; /* Underlying physical block number. */ off_t b_offset; /* Offset into file. */ TAILQ_ENTRY(buf) b_bobufs; /* (V) Buffer's associated vnode. */ struct buf *b_left; /* (V) splay tree link */ struct buf *b_right; /* (V) splay tree link */ uint32_t b_vflags; /* (V) BV_* flags */ TAILQ_ENTRY(buf) b_freelist; /* (Q) Free list position inactive. */ unsigned short b_qindex; /* (Q) buffer queue index */ uint32_t b_flags; /* B_* flags. */ b_xflags_t b_xflags; /* extra flags */ struct lock b_lock; /* Buffer lock */ long b_bufsize; /* Allocated buffer size. */ long b_runningbufspace; /* when I/O is running, pipelining */ caddr_t b_kvabase; /* base kva for buffer */ int b_kvasize; /* size of kva for buffer */ daddr_t b_lblkno; /* Logical block number. */ struct vnode *b_vp; /* Device vnode. */ int b_dirtyoff; /* Offset in buffer of dirty region. */ int b_dirtyend; /* Offset of end of dirty region. */ struct ucred *b_rcred; /* Read credentials reference. */ struct ucred *b_wcred; /* Write credentials reference. */ void *b_saveaddr; /* Original b_addr for physio. */ union pager_info { int pg_reqpage; } b_pager; union cluster_info { TAILQ_HEAD(cluster_list_head, buf) cluster_head; TAILQ_ENTRY(buf) cluster_entry; } b_cluster; struct vm_page *b_pages[btoc(MAXPHYS)]; int b_npages; struct workhead b_dep; /* (D) List of filesystem dependencies. */ void *b_fsprivate1; void *b_fsprivate2; void *b_fsprivate3; int b_pin_count; }; #define b_object b_bufobj->bo_object /* * These flags are kept in b_flags. * * Notes: * * B_ASYNC VOP calls on bp's are usually async whether or not * B_ASYNC is set, but some subsystems, such as NFS, like * to know what is best for the caller so they can * optimize the I/O. * * B_PAGING Indicates that bp is being used by the paging system or * some paging system and that the bp is not linked into * the b_vp's clean/dirty linked lists or ref counts. * Buffer vp reassignments are illegal in this case. * * B_CACHE This may only be set if the buffer is entirely valid. * The situation where B_DELWRI is set and B_CACHE is * clear MUST be committed to disk by getblk() so * B_DELWRI can also be cleared. See the comments for * getblk() in kern/vfs_bio.c. If B_CACHE is clear, * the caller is expected to clear BIO_ERROR and B_INVAL, * set BIO_READ, and initiate an I/O. * * The 'entire buffer' is defined to be the range from * 0 through b_bcount. * * B_MALLOC Request that the buffer be allocated from the malloc * pool, DEV_BSIZE aligned instead of PAGE_SIZE aligned. * * B_CLUSTEROK This flag is typically set for B_DELWRI buffers * by filesystems that allow clustering when the buffer * is fully dirty and indicates that it may be clustered * with other adjacent dirty buffers. Note the clustering * may not be used with the stage 1 data write under NFS * but may be used for the commit rpc portion. * * B_VMIO Indicates that the buffer is tied into an VM object. * The buffer's data is always PAGE_SIZE aligned even * if b_bufsize and b_bcount are not. ( b_bufsize is * always at least DEV_BSIZE aligned, though ). * * B_DIRECT Hint that we should attempt to completely free * the pages underlying the buffer. B_DIRECT is * sticky until the buffer is released and typically * only has an effect when B_RELBUF is also set. * */ #define B_AGE 0x00000001 /* Move to age queue when I/O done. */ #define B_NEEDCOMMIT 0x00000002 /* Append-write in progress. */ #define B_ASYNC 0x00000004 /* Start I/O, do not wait. */ #define B_DIRECT 0x00000008 /* direct I/O flag (pls free vmio) */ #define B_DEFERRED 0x00000010 /* Skipped over for cleaning */ #define B_CACHE 0x00000020 /* Bread found us in the cache. */ #define B_VALIDSUSPWRT 0x00000040 /* Valid write during suspension. */ #define B_DELWRI 0x00000080 /* Delay I/O until buffer reused. */ #define B_PERSISTENT 0x00000100 /* Perm. ref'ed while EXT2FS mounted. */ #define B_DONE 0x00000200 /* I/O completed. */ #define B_EINTR 0x00000400 /* I/O was interrupted */ #define B_00000800 0x00000800 /* Available flag. */ #define B_00001000 0x00001000 /* Available flag. */ #define B_INVAL 0x00002000 /* Does not contain valid info. */ #define B_00004000 0x00004000 /* Available flag. */ #define B_NOCACHE 0x00008000 /* Do not cache block after use. */ #define B_MALLOC 0x00010000 /* malloced b_data */ #define B_CLUSTEROK 0x00020000 /* Pagein op, so swap() can count it. */ #define B_000400000 0x00040000 /* Available flag. */ #define B_000800000 0x00080000 /* Available flag. */ #define B_00100000 0x00100000 /* Available flag. */ #define B_DIRTY 0x00200000 /* Needs writing later (in EXT2FS). */ #define B_RELBUF 0x00400000 /* Release VMIO buffer. */ #define B_00800000 0x00800000 /* Available flag. */ #define B_01000000 0x01000000 /* Available flag. */ -#define B_02000000 0x02000000 /* Available flag. */ +#define B_NEEDSGIANT 0x02000000 /* Buffer's vnode needs giant. */ #define B_PAGING 0x04000000 /* volatile paging I/O -- bypass VMIO */ #define B_MANAGED 0x08000000 /* Managed by FS. */ #define B_RAM 0x10000000 /* Read ahead mark (flag) */ #define B_VMIO 0x20000000 /* VMIO flag */ #define B_CLUSTER 0x40000000 /* pagein op, so swap() can count it */ #define B_REMFREE 0x80000000 /* Delayed bremfree */ #define PRINT_BUF_FLAGS "\20\40remfree\37cluster\36vmio\35ram\34b27" \ "\33paging\32b25\31b24\30b23\27relbuf\26dirty\25b20" \ "\24b19\23b18\22clusterok\21malloc\20nocache\17b14\16inval" \ "\15b12\14b11\13eintr\12done\11persist\10delwri\7validsuspwrt" \ "\6cache\5deferred\4direct\3async\2needcommit\1age" /* * These flags are kept in b_xflags. */ #define BX_VNDIRTY 0x00000001 /* On vnode dirty list */ #define BX_VNCLEAN 0x00000002 /* On vnode clean list */ #define BX_BKGRDWRITE 0x00000010 /* Do writes in background */ #define BX_BKGRDMARKER 0x00000020 /* Mark buffer for splay tree */ #define BX_ALTDATA 0x00000040 /* Holds extended data */ #define NOOFFSET (-1LL) /* No buffer offset calculated yet */ /* * These flags are kept in b_vflags. */ #define BV_SCANNED 0x00000001 /* VOP_FSYNC funcs mark written bufs */ #define BV_BKGRDINPROG 0x00000002 /* Background write in progress */ #define BV_BKGRDWAIT 0x00000004 /* Background write waiting */ #ifdef _KERNEL /* * Buffer locking */ extern const char *buf_wmesg; /* Default buffer lock message */ #define BUF_WMESG "bufwait" #include /* XXX for curthread */ #include /* * Initialize a lock. */ #define BUF_LOCKINIT(bp) \ lockinit(&(bp)->b_lock, PRIBIO + 4, buf_wmesg, 0, 0) /* * * Get a lock sleeping non-interruptably until it becomes available. */ static __inline int BUF_LOCK(struct buf *, int, struct mtx *); static __inline int BUF_LOCK(struct buf *bp, int locktype, struct mtx *interlock) { int s, ret; s = splbio(); mtx_lock(bp->b_lock.lk_interlock); locktype |= LK_INTERNAL; bp->b_lock.lk_wmesg = buf_wmesg; bp->b_lock.lk_prio = PRIBIO + 4; ret = lockmgr(&(bp)->b_lock, locktype, interlock, curthread); splx(s); return ret; } /* * Get a lock sleeping with specified interruptably and timeout. */ static __inline int BUF_TIMELOCK(struct buf *, int, struct mtx *, char *, int, int); static __inline int BUF_TIMELOCK(struct buf *bp, int locktype, struct mtx *interlock, char *wmesg, int catch, int timo) { int s, ret; s = splbio(); mtx_lock(bp->b_lock.lk_interlock); locktype |= LK_INTERNAL | LK_TIMELOCK; bp->b_lock.lk_wmesg = wmesg; bp->b_lock.lk_prio = (PRIBIO + 4) | catch; bp->b_lock.lk_timo = timo; ret = lockmgr(&(bp)->b_lock, (locktype), interlock, curthread); splx(s); return ret; } /* * Release a lock. Only the acquiring process may free the lock unless * it has been handed off to biodone. */ static __inline void BUF_UNLOCK(struct buf *); static __inline void BUF_UNLOCK(struct buf *bp) { int s; s = splbio(); KASSERT((bp->b_flags & B_REMFREE) == 0, ("BUF_UNLOCK %p while B_REMFREE is still set.", bp)); lockmgr(&(bp)->b_lock, LK_RELEASE, NULL, curthread); splx(s); } /* * Free a buffer lock. */ #define BUF_LOCKFREE(bp) \ do { \ if (BUF_REFCNT(bp) > 0) \ panic("free locked buf"); \ lockdestroy(&(bp)->b_lock); \ } while (0) #ifdef _SYS_PROC_H_ /* Avoid #include pollution */ /* * When initiating asynchronous I/O, change ownership of the lock to the * kernel. Once done, the lock may legally released by biodone. The * original owning process can no longer acquire it recursively, but must * wait until the I/O is completed and the lock has been freed by biodone. */ static __inline void BUF_KERNPROC(struct buf *); static __inline void BUF_KERNPROC(struct buf *bp) { struct thread *td = curthread; if ((td != PCPU_GET(idlethread)) && bp->b_lock.lk_lockholder == td) td->td_locks--; bp->b_lock.lk_lockholder = LK_KERNPROC; } #endif /* * Find out the number of references to a lock. */ static __inline int BUF_REFCNT(struct buf *); static __inline int BUF_REFCNT(struct buf *bp) { int s, ret; /* * When the system is panicing, the lock manager grants all lock * requests whether or not the lock is available. To avoid "unlocked * buffer" panics after a crash, we just claim that all buffers * are locked when cleaning up after a system panic. */ if (panicstr != NULL) return (1); s = splbio(); ret = lockcount(&(bp)->b_lock); splx(s); return ret; } #endif /* _KERNEL */ struct buf_queue_head { TAILQ_HEAD(buf_queue, buf) queue; daddr_t last_pblkno; struct buf *insert_point; struct buf *switch_point; }; /* * This structure describes a clustered I/O. It is stored in the b_saveaddr * field of the buffer on which I/O is done. At I/O completion, cluster * callback uses the structure to parcel I/O's to individual buffers, and * then free's this structure. */ struct cluster_save { long bs_bcount; /* Saved b_bcount. */ long bs_bufsize; /* Saved b_bufsize. */ void *bs_saveaddr; /* Saved b_addr. */ int bs_nchildren; /* Number of associated buffers. */ struct buf **bs_children; /* List of associated buffers. */ }; #ifdef _KERNEL static __inline int bwrite(struct buf *bp) { KASSERT(bp->b_bufobj != NULL, ("bwrite: no bufobj bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops != NULL, ("bwrite: no bo_ops bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops->bop_write != NULL, ("bwrite: no bop_write bp=%p", bp)); return (BO_WRITE(bp->b_bufobj, bp)); } static __inline void bstrategy(struct buf *bp) { KASSERT(bp->b_bufobj != NULL, ("bstrategy: no bufobj bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops != NULL, ("bstrategy: no bo_ops bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops->bop_strategy != NULL, ("bstrategy: no bop_strategy bp=%p", bp)); BO_STRATEGY(bp->b_bufobj, bp); } static __inline void buf_start(struct buf *bp) { if (bioops.io_start) (*bioops.io_start)(bp); } static __inline void buf_complete(struct buf *bp) { if (bioops.io_complete) (*bioops.io_complete)(bp); } static __inline void buf_deallocate(struct buf *bp) { if (bioops.io_deallocate) (*bioops.io_deallocate)(bp); BUF_LOCKFREE(bp); } static __inline int buf_countdeps(struct buf *bp, int i) { if (bioops.io_countdeps) return ((*bioops.io_countdeps)(bp, i)); else return (0); } #endif /* _KERNEL */ /* * Zero out the buffer's data area. */ #define clrbuf(bp) { \ bzero((bp)->b_data, (u_int)(bp)->b_bcount); \ (bp)->b_resid = 0; \ } /* * Flags for getblk's last parameter. */ #define GB_LOCK_NOWAIT 0x0001 /* Fail if we block on a buf lock. */ #define GB_NOCREAT 0x0002 /* Don't create a buf if not found. */ #ifdef _KERNEL extern int nbuf; /* The number of buffer headers */ extern int maxswzone; /* Max KVA for swap structures */ extern int maxbcache; /* Max KVA for buffer cache */ extern int runningbufspace; extern int hibufspace; extern int buf_maxio; /* nominal maximum I/O for buffer */ extern struct buf *buf; /* The buffer headers. */ extern char *buffers; /* The buffer contents. */ extern int bufpages; /* Number of memory pages in the buffer pool. */ extern struct buf *swbuf; /* Swap I/O buffer headers. */ extern int nswbuf; /* Number of swap I/O buffer headers. */ extern int cluster_pbuf_freecnt; /* Number of pbufs for clusters */ extern int vnode_pbuf_freecnt; /* Number of pbufs for vnode pager */ void runningbufwakeup(struct buf *); void waitrunningbufspace(void); caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est); void bufinit(void); void bwillwrite(void); int buf_dirty_count_severe(void); void bremfree(struct buf *); void bremfreef(struct buf *); /* XXX Force bremfree, only for nfs. */ int bread(struct vnode *, daddr_t, int, struct ucred *, struct buf **); void breada(struct vnode *, daddr_t *, int *, int, struct ucred *); int breadn(struct vnode *, daddr_t, int, daddr_t *, int *, int, struct ucred *, struct buf **); void bdwrite(struct buf *); void bawrite(struct buf *); void bdirty(struct buf *); void bundirty(struct buf *); void bufstrategy(struct bufobj *, struct buf *); void brelse(struct buf *); void bqrelse(struct buf *); int vfs_bio_awrite(struct buf *); struct buf * getpbuf(int *); struct buf *incore(struct bufobj *, daddr_t); struct buf *gbincore(struct bufobj *, daddr_t); struct buf *getblk(struct vnode *, daddr_t, int, int, int, int); struct buf *geteblk(int); int bufwait(struct buf *); int bufwrite(struct buf *); void bufdone(struct buf *); void bufdone_finish(struct buf *); int cluster_read(struct vnode *, u_quad_t, daddr_t, long, struct ucred *, long, int, struct buf **); int cluster_wbuild(struct vnode *, long, daddr_t, int); void cluster_write(struct vnode *, struct buf *, u_quad_t, int); void vfs_bio_set_validclean(struct buf *, int base, int size); void vfs_bio_clrbuf(struct buf *); void vfs_busy_pages(struct buf *, int clear_modify); void vfs_unbusy_pages(struct buf *); int vmapbuf(struct buf *); void vunmapbuf(struct buf *); void relpbuf(struct buf *, int *); void brelvp(struct buf *); void bgetvp(struct vnode *, struct buf *); void pbgetbo(struct bufobj *bo, struct buf *bp); void pbgetvp(struct vnode *, struct buf *); void pbrelbo(struct buf *); void pbrelvp(struct buf *); int allocbuf(struct buf *bp, int size); void reassignbuf(struct buf *); struct buf *trypbuf(int *); void bwait(struct buf *, u_char, const char *); void bdone(struct buf *); void bpin(struct buf *); void bunpin(struct buf *); void bunpin_wait(struct buf *); #endif /* _KERNEL */ #endif /* !_SYS_BUF_H_ */