diff --git a/sys/arm64/arm64/pmap.c b/sys/arm64/arm64/pmap.c index 4bd1e86ffd5f..3166b3d7959b 100644 --- a/sys/arm64/arm64/pmap.c +++ b/sys/arm64/arm64/pmap.c @@ -1,8180 +1,8176 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2003 Peter Wemm * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * Copyright (c) 2014 Andrew Turner * All rights reserved. * Copyright (c) 2014-2016 The FreeBSD Foundation * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * This software was developed by Andrew Turner under sponsorship from * the FreeBSD Foundation. * * 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. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 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. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * 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. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_vm.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 #include #include #include #include #include #include #include #include #ifdef NUMA #define PMAP_MEMDOM MAXMEMDOM #else #define PMAP_MEMDOM 1 #endif #define PMAP_ASSERT_STAGE1(pmap) MPASS((pmap)->pm_stage == PM_STAGE1) #define PMAP_ASSERT_STAGE2(pmap) MPASS((pmap)->pm_stage == PM_STAGE2) #define NL0PG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NL1PG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NL2PG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NL3PG (PAGE_SIZE/(sizeof (pt_entry_t))) #define NUL0E L0_ENTRIES #define NUL1E (NUL0E * NL1PG) #define NUL2E (NUL1E * NL2PG) #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #define __pvused #else #define PV_STAT(x) do { } while (0) #define __pvused __unused #endif #define pmap_l0_pindex(v) (NUL2E + NUL1E + ((v) >> L0_SHIFT)) #define pmap_l1_pindex(v) (NUL2E + ((v) >> L1_SHIFT)) #define pmap_l2_pindex(v) ((v) >> L2_SHIFT) #define PMAP_SAN_PTE_BITS (ATTR_DEFAULT | ATTR_S1_XN | \ ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | ATTR_S1_AP(ATTR_S1_AP_RW)) struct pmap_large_md_page { struct rwlock pv_lock; struct md_page pv_page; /* Pad to a power of 2, see pmap_init_pv_table(). */ int pv_pad[2]; }; __exclusive_cache_line static struct pmap_large_md_page pv_dummy_large; #define pv_dummy pv_dummy_large.pv_page __read_mostly static struct pmap_large_md_page *pv_table; static struct pmap_large_md_page * _pa_to_pmdp(vm_paddr_t pa) { struct vm_phys_seg *seg; - int segind; - for (segind = 0; segind < vm_phys_nsegs; segind++) { - seg = &vm_phys_segs[segind]; - if (pa >= seg->start && pa < seg->end) - return ((struct pmap_large_md_page *)seg->md_first + - pmap_l2_pindex(pa) - pmap_l2_pindex(seg->start)); - } + if ((seg = vm_phys_paddr_to_seg(pa)) != NULL) + return ((struct pmap_large_md_page *)seg->md_first + + pmap_l2_pindex(pa) - pmap_l2_pindex(seg->start)); return (NULL); } static struct pmap_large_md_page * pa_to_pmdp(vm_paddr_t pa) { struct pmap_large_md_page *pvd; pvd = _pa_to_pmdp(pa); if (pvd == NULL) panic("pa 0x%jx not within vm_phys_segs", (uintmax_t)pa); return (pvd); } static struct pmap_large_md_page * page_to_pmdp(vm_page_t m) { struct vm_phys_seg *seg; seg = &vm_phys_segs[m->segind]; return ((struct pmap_large_md_page *)seg->md_first + pmap_l2_pindex(VM_PAGE_TO_PHYS(m)) - pmap_l2_pindex(seg->start)); } #define pa_to_pvh(pa) (&(pa_to_pmdp(pa)->pv_page)) #define page_to_pvh(m) (&(page_to_pmdp(m)->pv_page)) #define PHYS_TO_PV_LIST_LOCK(pa) ({ \ struct pmap_large_md_page *_pvd; \ struct rwlock *_lock; \ _pvd = _pa_to_pmdp(pa); \ if (__predict_false(_pvd == NULL)) \ _lock = &pv_dummy_large.pv_lock; \ else \ _lock = &(_pvd->pv_lock); \ _lock; \ }) static struct rwlock * VM_PAGE_TO_PV_LIST_LOCK(vm_page_t m) { if ((m->flags & PG_FICTITIOUS) == 0) return (&page_to_pmdp(m)->pv_lock); else return (&pv_dummy_large.pv_lock); } #define CHANGE_PV_LIST_LOCK(lockp, new_lock) do { \ struct rwlock **_lockp = (lockp); \ struct rwlock *_new_lock = (new_lock); \ \ if (_new_lock != *_lockp) { \ if (*_lockp != NULL) \ rw_wunlock(*_lockp); \ *_lockp = _new_lock; \ rw_wlock(*_lockp); \ } \ } while (0) #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) \ CHANGE_PV_LIST_LOCK(lockp, PHYS_TO_PV_LIST_LOCK(pa)) #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \ CHANGE_PV_LIST_LOCK(lockp, VM_PAGE_TO_PV_LIST_LOCK(m)) #define RELEASE_PV_LIST_LOCK(lockp) do { \ struct rwlock **_lockp = (lockp); \ \ if (*_lockp != NULL) { \ rw_wunlock(*_lockp); \ *_lockp = NULL; \ } \ } while (0) /* * The presence of this flag indicates that the mapping is writeable. * If the ATTR_S1_AP_RO bit is also set, then the mapping is clean, otherwise * it is dirty. This flag may only be set on managed mappings. * * The DBM bit is reserved on ARMv8.0 but it seems we can safely treat it * as a software managed bit. */ #define ATTR_SW_DBM ATTR_DBM struct pmap kernel_pmap_store; /* Used for mapping ACPI memory before VM is initialized */ #define PMAP_PREINIT_MAPPING_COUNT 32 #define PMAP_PREINIT_MAPPING_SIZE (PMAP_PREINIT_MAPPING_COUNT * L2_SIZE) static vm_offset_t preinit_map_va; /* Start VA of pre-init mapping space */ static int vm_initialized = 0; /* No need to use pre-init maps when set */ /* * Reserve a few L2 blocks starting from 'preinit_map_va' pointer. * Always map entire L2 block for simplicity. * VA of L2 block = preinit_map_va + i * L2_SIZE */ static struct pmap_preinit_mapping { vm_paddr_t pa; vm_offset_t va; vm_size_t size; } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT]; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ vm_offset_t kernel_vm_end = 0; /* * Data for the pv entry allocation mechanism. */ #ifdef NUMA static __inline int pc_to_domain(struct pv_chunk *pc) { return (vm_phys_domain(DMAP_TO_PHYS((vm_offset_t)pc))); } #else static __inline int pc_to_domain(struct pv_chunk *pc __unused) { return (0); } #endif struct pv_chunks_list { struct mtx pvc_lock; TAILQ_HEAD(pch, pv_chunk) pvc_list; int active_reclaims; } __aligned(CACHE_LINE_SIZE); struct pv_chunks_list __exclusive_cache_line pv_chunks[PMAP_MEMDOM]; vm_paddr_t dmap_phys_base; /* The start of the dmap region */ vm_paddr_t dmap_phys_max; /* The limit of the dmap region */ vm_offset_t dmap_max_addr; /* The virtual address limit of the dmap */ extern pt_entry_t pagetable_l0_ttbr1[]; #define PHYSMAP_SIZE (2 * (VM_PHYSSEG_MAX - 1)) static vm_paddr_t physmap[PHYSMAP_SIZE]; static u_int physmap_idx; static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "VM/pmap parameters"); #if PAGE_SIZE == PAGE_SIZE_4K #define L1_BLOCKS_SUPPORTED 1 #else /* TODO: Make this dynamic when we support FEAT_LPA2 (TCR_EL1.DS == 1) */ #define L1_BLOCKS_SUPPORTED 0 #endif #define PMAP_ASSERT_L1_BLOCKS_SUPPORTED MPASS(L1_BLOCKS_SUPPORTED) /* * This ASID allocator uses a bit vector ("asid_set") to remember which ASIDs * that it has currently allocated to a pmap, a cursor ("asid_next") to * optimize its search for a free ASID in the bit vector, and an epoch number * ("asid_epoch") to indicate when it has reclaimed all previously allocated * ASIDs that are not currently active on a processor. * * The current epoch number is always in the range [0, INT_MAX). Negative * numbers and INT_MAX are reserved for special cases that are described * below. */ struct asid_set { int asid_bits; bitstr_t *asid_set; int asid_set_size; int asid_next; int asid_epoch; struct mtx asid_set_mutex; }; static struct asid_set asids; static struct asid_set vmids; static SYSCTL_NODE(_vm_pmap, OID_AUTO, asid, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "ASID allocator"); SYSCTL_INT(_vm_pmap_asid, OID_AUTO, bits, CTLFLAG_RD, &asids.asid_bits, 0, "The number of bits in an ASID"); SYSCTL_INT(_vm_pmap_asid, OID_AUTO, next, CTLFLAG_RD, &asids.asid_next, 0, "The last allocated ASID plus one"); SYSCTL_INT(_vm_pmap_asid, OID_AUTO, epoch, CTLFLAG_RD, &asids.asid_epoch, 0, "The current epoch number"); static SYSCTL_NODE(_vm_pmap, OID_AUTO, vmid, CTLFLAG_RD, 0, "VMID allocator"); SYSCTL_INT(_vm_pmap_vmid, OID_AUTO, bits, CTLFLAG_RD, &vmids.asid_bits, 0, "The number of bits in an VMID"); SYSCTL_INT(_vm_pmap_vmid, OID_AUTO, next, CTLFLAG_RD, &vmids.asid_next, 0, "The last allocated VMID plus one"); SYSCTL_INT(_vm_pmap_vmid, OID_AUTO, epoch, CTLFLAG_RD, &vmids.asid_epoch, 0, "The current epoch number"); void (*pmap_clean_stage2_tlbi)(void); void (*pmap_invalidate_vpipt_icache)(void); void (*pmap_stage2_invalidate_range)(uint64_t, vm_offset_t, vm_offset_t, bool); void (*pmap_stage2_invalidate_all)(uint64_t); /* * A pmap's cookie encodes an ASID and epoch number. Cookies for reserved * ASIDs have a negative epoch number, specifically, INT_MIN. Cookies for * dynamically allocated ASIDs have a non-negative epoch number. * * An invalid ASID is represented by -1. * * There are two special-case cookie values: (1) COOKIE_FROM(-1, INT_MIN), * which indicates that an ASID should never be allocated to the pmap, and * (2) COOKIE_FROM(-1, INT_MAX), which indicates that an ASID should be * allocated when the pmap is next activated. */ #define COOKIE_FROM(asid, epoch) ((long)((u_int)(asid) | \ ((u_long)(epoch) << 32))) #define COOKIE_TO_ASID(cookie) ((int)(cookie)) #define COOKIE_TO_EPOCH(cookie) ((int)((u_long)(cookie) >> 32)) #define TLBI_VA_SHIFT 12 #define TLBI_VA_MASK ((1ul << 44) - 1) #define TLBI_VA(addr) (((addr) >> TLBI_VA_SHIFT) & TLBI_VA_MASK) #define TLBI_VA_L3_INCR (L3_SIZE >> TLBI_VA_SHIFT) static int superpages_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, superpages_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &superpages_enabled, 0, "Are large page mappings enabled?"); /* * Internal flags for pmap_enter()'s helper functions. */ #define PMAP_ENTER_NORECLAIM 0x1000000 /* Don't reclaim PV entries. */ #define PMAP_ENTER_NOREPLACE 0x2000000 /* Don't replace mappings. */ TAILQ_HEAD(pv_chunklist, pv_chunk); static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_chunk_batch(struct pv_chunklist *batch); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp); static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp); static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte); static bool pmap_activate_int(pmap_t pmap); static void pmap_alloc_asid(pmap_t pmap); static int pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, int mode, bool skip_unmapped); static pt_entry_t *pmap_demote_l1(pmap_t pmap, pt_entry_t *l1, vm_offset_t va); static pt_entry_t *pmap_demote_l2_locked(pmap_t pmap, pt_entry_t *l2, vm_offset_t va, struct rwlock **lockp); static pt_entry_t *pmap_demote_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t va); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static int pmap_enter_l2(pmap_t pmap, vm_offset_t va, pd_entry_t new_l2, u_int flags, vm_page_t m, struct rwlock **lockp); static int pmap_remove_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t sva, pd_entry_t l1e, struct spglist *free, struct rwlock **lockp); static int pmap_remove_l3(pmap_t pmap, pt_entry_t *l3, vm_offset_t sva, pd_entry_t l2e, struct spglist *free, struct rwlock **lockp); static void pmap_reset_asid_set(pmap_t pmap); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp); static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp); static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free); static int pmap_unuse_pt(pmap_t, vm_offset_t, pd_entry_t, struct spglist *); static __inline vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va); /* * These load the old table data and store the new value. * They need to be atomic as the System MMU may write to the table at * the same time as the CPU. */ #define pmap_clear(table) atomic_store_64(table, 0) #define pmap_clear_bits(table, bits) atomic_clear_64(table, bits) #define pmap_load(table) (*table) #define pmap_load_clear(table) atomic_swap_64(table, 0) #define pmap_load_store(table, entry) atomic_swap_64(table, entry) #define pmap_set_bits(table, bits) atomic_set_64(table, bits) #define pmap_store(table, entry) atomic_store_64(table, entry) /********************/ /* Inline functions */ /********************/ static __inline void pagecopy(void *s, void *d) { memcpy(d, s, PAGE_SIZE); } static __inline pd_entry_t * pmap_l0(pmap_t pmap, vm_offset_t va) { return (&pmap->pm_l0[pmap_l0_index(va)]); } static __inline pd_entry_t * pmap_l0_to_l1(pd_entry_t *l0, vm_offset_t va) { pd_entry_t *l1; l1 = (pd_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(pmap_load(l0))); return (&l1[pmap_l1_index(va)]); } static __inline pd_entry_t * pmap_l1(pmap_t pmap, vm_offset_t va) { pd_entry_t *l0; l0 = pmap_l0(pmap, va); if ((pmap_load(l0) & ATTR_DESCR_MASK) != L0_TABLE) return (NULL); return (pmap_l0_to_l1(l0, va)); } static __inline pd_entry_t * pmap_l1_to_l2(pd_entry_t *l1p, vm_offset_t va) { pd_entry_t l1, *l2p; l1 = pmap_load(l1p); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); /* * The valid bit may be clear if pmap_update_entry() is concurrently * modifying the entry, so for KVA only the entry type may be checked. */ KASSERT(ADDR_IS_KERNEL(va) || (l1 & ATTR_DESCR_VALID) != 0, ("%s: L1 entry %#lx for %#lx is invalid", __func__, l1, va)); KASSERT((l1 & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_TABLE, ("%s: L1 entry %#lx for %#lx is a leaf", __func__, l1, va)); l2p = (pd_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(l1)); return (&l2p[pmap_l2_index(va)]); } static __inline pd_entry_t * pmap_l2(pmap_t pmap, vm_offset_t va) { pd_entry_t *l1; l1 = pmap_l1(pmap, va); if ((pmap_load(l1) & ATTR_DESCR_MASK) != L1_TABLE) return (NULL); return (pmap_l1_to_l2(l1, va)); } static __inline pt_entry_t * pmap_l2_to_l3(pd_entry_t *l2p, vm_offset_t va) { pd_entry_t l2; pt_entry_t *l3p; l2 = pmap_load(l2p); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); /* * The valid bit may be clear if pmap_update_entry() is concurrently * modifying the entry, so for KVA only the entry type may be checked. */ KASSERT(ADDR_IS_KERNEL(va) || (l2 & ATTR_DESCR_VALID) != 0, ("%s: L2 entry %#lx for %#lx is invalid", __func__, l2, va)); KASSERT((l2 & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_TABLE, ("%s: L2 entry %#lx for %#lx is a leaf", __func__, l2, va)); l3p = (pt_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(l2)); return (&l3p[pmap_l3_index(va)]); } /* * Returns the lowest valid pde for a given virtual address. * The next level may or may not point to a valid page or block. */ static __inline pd_entry_t * pmap_pde(pmap_t pmap, vm_offset_t va, int *level) { pd_entry_t *l0, *l1, *l2, desc; l0 = pmap_l0(pmap, va); desc = pmap_load(l0) & ATTR_DESCR_MASK; if (desc != L0_TABLE) { *level = -1; return (NULL); } l1 = pmap_l0_to_l1(l0, va); desc = pmap_load(l1) & ATTR_DESCR_MASK; if (desc != L1_TABLE) { *level = 0; return (l0); } l2 = pmap_l1_to_l2(l1, va); desc = pmap_load(l2) & ATTR_DESCR_MASK; if (desc != L2_TABLE) { *level = 1; return (l1); } *level = 2; return (l2); } /* * Returns the lowest valid pte block or table entry for a given virtual * address. If there are no valid entries return NULL and set the level to * the first invalid level. */ static __inline pt_entry_t * pmap_pte(pmap_t pmap, vm_offset_t va, int *level) { pd_entry_t *l1, *l2, desc; pt_entry_t *l3; l1 = pmap_l1(pmap, va); if (l1 == NULL) { *level = 0; return (NULL); } desc = pmap_load(l1) & ATTR_DESCR_MASK; if (desc == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; *level = 1; return (l1); } if (desc != L1_TABLE) { *level = 1; return (NULL); } l2 = pmap_l1_to_l2(l1, va); desc = pmap_load(l2) & ATTR_DESCR_MASK; if (desc == L2_BLOCK) { *level = 2; return (l2); } if (desc != L2_TABLE) { *level = 2; return (NULL); } *level = 3; l3 = pmap_l2_to_l3(l2, va); if ((pmap_load(l3) & ATTR_DESCR_MASK) != L3_PAGE) return (NULL); return (l3); } /* * If the given pmap has an L{1,2}_BLOCK or L3_PAGE entry at the specified * level that maps the specified virtual address, then a pointer to that entry * is returned. Otherwise, NULL is returned, unless INVARIANTS are enabled * and a diagnostic message is provided, in which case this function panics. */ static __always_inline pt_entry_t * pmap_pte_exists(pmap_t pmap, vm_offset_t va, int level, const char *diag) { pd_entry_t *l0p, *l1p, *l2p; pt_entry_t desc, *l3p; int walk_level __diagused; KASSERT(level >= 0 && level < 4, ("%s: %s passed an out-of-range level (%d)", __func__, diag, level)); l0p = pmap_l0(pmap, va); desc = pmap_load(l0p) & ATTR_DESCR_MASK; if (desc == L0_TABLE && level > 0) { l1p = pmap_l0_to_l1(l0p, va); desc = pmap_load(l1p) & ATTR_DESCR_MASK; if (desc == L1_BLOCK && level == 1) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; return (l1p); } if (desc == L1_TABLE && level > 1) { l2p = pmap_l1_to_l2(l1p, va); desc = pmap_load(l2p) & ATTR_DESCR_MASK; if (desc == L2_BLOCK && level == 2) return (l2p); else if (desc == L2_TABLE && level > 2) { l3p = pmap_l2_to_l3(l2p, va); desc = pmap_load(l3p) & ATTR_DESCR_MASK; if (desc == L3_PAGE && level == 3) return (l3p); else walk_level = 3; } else walk_level = 2; } else walk_level = 1; } else walk_level = 0; KASSERT(diag == NULL, ("%s: va %#lx not mapped at level %d, desc %ld at level %d", diag, va, level, desc, walk_level)); return (NULL); } bool pmap_ps_enabled(pmap_t pmap) { /* * Promotion requires a hypervisor call when the kernel is running * in EL1. To stop this disable superpage support on non-stage 1 * pmaps for now. */ if (pmap->pm_stage != PM_STAGE1) return (false); return (superpages_enabled != 0); } bool pmap_get_tables(pmap_t pmap, vm_offset_t va, pd_entry_t **l0, pd_entry_t **l1, pd_entry_t **l2, pt_entry_t **l3) { pd_entry_t *l0p, *l1p, *l2p; if (pmap->pm_l0 == NULL) return (false); l0p = pmap_l0(pmap, va); *l0 = l0p; if ((pmap_load(l0p) & ATTR_DESCR_MASK) != L0_TABLE) return (false); l1p = pmap_l0_to_l1(l0p, va); *l1 = l1p; if ((pmap_load(l1p) & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; *l2 = NULL; *l3 = NULL; return (true); } if ((pmap_load(l1p) & ATTR_DESCR_MASK) != L1_TABLE) return (false); l2p = pmap_l1_to_l2(l1p, va); *l2 = l2p; if ((pmap_load(l2p) & ATTR_DESCR_MASK) == L2_BLOCK) { *l3 = NULL; return (true); } if ((pmap_load(l2p) & ATTR_DESCR_MASK) != L2_TABLE) return (false); *l3 = pmap_l2_to_l3(l2p, va); return (true); } static __inline int pmap_l3_valid(pt_entry_t l3) { return ((l3 & ATTR_DESCR_MASK) == L3_PAGE); } CTASSERT(L1_BLOCK == L2_BLOCK); static pt_entry_t pmap_pte_memattr(pmap_t pmap, vm_memattr_t memattr) { pt_entry_t val; if (pmap->pm_stage == PM_STAGE1) { val = ATTR_S1_IDX(memattr); if (memattr == VM_MEMATTR_DEVICE) val |= ATTR_S1_XN; return (val); } val = 0; switch (memattr) { case VM_MEMATTR_DEVICE: return (ATTR_S2_MEMATTR(ATTR_S2_MEMATTR_DEVICE_nGnRnE) | ATTR_S2_XN(ATTR_S2_XN_ALL)); case VM_MEMATTR_UNCACHEABLE: return (ATTR_S2_MEMATTR(ATTR_S2_MEMATTR_NC)); case VM_MEMATTR_WRITE_BACK: return (ATTR_S2_MEMATTR(ATTR_S2_MEMATTR_WB)); case VM_MEMATTR_WRITE_THROUGH: return (ATTR_S2_MEMATTR(ATTR_S2_MEMATTR_WT)); default: panic("%s: invalid memory attribute %x", __func__, memattr); } } static pt_entry_t pmap_pte_prot(pmap_t pmap, vm_prot_t prot) { pt_entry_t val; val = 0; if (pmap->pm_stage == PM_STAGE1) { if ((prot & VM_PROT_EXECUTE) == 0) val |= ATTR_S1_XN; if ((prot & VM_PROT_WRITE) == 0) val |= ATTR_S1_AP(ATTR_S1_AP_RO); } else { if ((prot & VM_PROT_WRITE) != 0) val |= ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE); if ((prot & VM_PROT_READ) != 0) val |= ATTR_S2_S2AP(ATTR_S2_S2AP_READ); if ((prot & VM_PROT_EXECUTE) == 0) val |= ATTR_S2_XN(ATTR_S2_XN_ALL); } return (val); } /* * Checks if the PTE is dirty. */ static inline int pmap_pte_dirty(pmap_t pmap, pt_entry_t pte) { KASSERT((pte & ATTR_SW_MANAGED) != 0, ("pte %#lx is unmanaged", pte)); if (pmap->pm_stage == PM_STAGE1) { KASSERT((pte & (ATTR_S1_AP_RW_BIT | ATTR_SW_DBM)) != 0, ("pte %#lx is writeable and missing ATTR_SW_DBM", pte)); return ((pte & (ATTR_S1_AP_RW_BIT | ATTR_SW_DBM)) == (ATTR_S1_AP(ATTR_S1_AP_RW) | ATTR_SW_DBM)); } return ((pte & ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE)) == ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE)); } static __inline void pmap_resident_count_inc(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); pmap->pm_stats.resident_count += count; } static __inline void pmap_resident_count_dec(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(pmap->pm_stats.resident_count >= count, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count -= count; } static vm_paddr_t pmap_early_vtophys(vm_offset_t va) { vm_paddr_t pa_page; pa_page = arm64_address_translate_s1e1r(va) & PAR_PA_MASK; return (pa_page | (va & PAR_LOW_MASK)); } /* State of the bootstrapped DMAP page tables */ struct pmap_bootstrap_state { pt_entry_t *l1; pt_entry_t *l2; pt_entry_t *l3; vm_offset_t freemempos; vm_offset_t va; vm_paddr_t pa; pt_entry_t table_attrs; u_int l0_slot; u_int l1_slot; u_int l2_slot; bool dmap_valid; }; /* The bootstrap state */ static struct pmap_bootstrap_state bs_state = { .l1 = NULL, .l2 = NULL, .l3 = NULL, .table_attrs = TATTR_PXN_TABLE, .l0_slot = L0_ENTRIES, .l1_slot = Ln_ENTRIES, .l2_slot = Ln_ENTRIES, .dmap_valid = false, }; static void pmap_bootstrap_l0_table(struct pmap_bootstrap_state *state) { vm_paddr_t l1_pa; pd_entry_t l0e; u_int l0_slot; /* Link the level 0 table to a level 1 table */ l0_slot = pmap_l0_index(state->va); if (l0_slot != state->l0_slot) { /* * Make sure we move from a low address to high address * before the DMAP region is ready. This ensures we never * modify an existing mapping until we can map from a * physical address to a virtual address. */ MPASS(state->l0_slot < l0_slot || state->l0_slot == L0_ENTRIES || state->dmap_valid); /* Reset lower levels */ state->l2 = NULL; state->l3 = NULL; state->l1_slot = Ln_ENTRIES; state->l2_slot = Ln_ENTRIES; /* Check the existing L0 entry */ state->l0_slot = l0_slot; if (state->dmap_valid) { l0e = pagetable_l0_ttbr1[l0_slot]; if ((l0e & ATTR_DESCR_VALID) != 0) { MPASS((l0e & ATTR_DESCR_MASK) == L0_TABLE); l1_pa = PTE_TO_PHYS(l0e); state->l1 = (pt_entry_t *)PHYS_TO_DMAP(l1_pa); return; } } /* Create a new L0 table entry */ state->l1 = (pt_entry_t *)state->freemempos; memset(state->l1, 0, PAGE_SIZE); state->freemempos += PAGE_SIZE; l1_pa = pmap_early_vtophys((vm_offset_t)state->l1); MPASS((l1_pa & Ln_TABLE_MASK) == 0); MPASS(pagetable_l0_ttbr1[l0_slot] == 0); pmap_store(&pagetable_l0_ttbr1[l0_slot], PHYS_TO_PTE(l1_pa) | TATTR_UXN_TABLE | TATTR_AP_TABLE_NO_EL0 | L0_TABLE); } KASSERT(state->l1 != NULL, ("%s: NULL l1", __func__)); } static void pmap_bootstrap_l1_table(struct pmap_bootstrap_state *state) { vm_paddr_t l2_pa; pd_entry_t l1e; u_int l1_slot; /* Make sure there is a valid L0 -> L1 table */ pmap_bootstrap_l0_table(state); /* Link the level 1 table to a level 2 table */ l1_slot = pmap_l1_index(state->va); if (l1_slot != state->l1_slot) { /* See pmap_bootstrap_l0_table for a description */ MPASS(state->l1_slot < l1_slot || state->l1_slot == Ln_ENTRIES || state->dmap_valid); /* Reset lower levels */ state->l3 = NULL; state->l2_slot = Ln_ENTRIES; /* Check the existing L1 entry */ state->l1_slot = l1_slot; if (state->dmap_valid) { l1e = state->l1[l1_slot]; if ((l1e & ATTR_DESCR_VALID) != 0) { MPASS((l1e & ATTR_DESCR_MASK) == L1_TABLE); l2_pa = PTE_TO_PHYS(l1e); state->l2 = (pt_entry_t *)PHYS_TO_DMAP(l2_pa); return; } } /* Create a new L1 table entry */ state->l2 = (pt_entry_t *)state->freemempos; memset(state->l2, 0, PAGE_SIZE); state->freemempos += PAGE_SIZE; l2_pa = pmap_early_vtophys((vm_offset_t)state->l2); MPASS((l2_pa & Ln_TABLE_MASK) == 0); MPASS(state->l1[l1_slot] == 0); pmap_store(&state->l1[l1_slot], PHYS_TO_PTE(l2_pa) | state->table_attrs | L1_TABLE); } KASSERT(state->l2 != NULL, ("%s: NULL l2", __func__)); } static void pmap_bootstrap_l2_table(struct pmap_bootstrap_state *state) { vm_paddr_t l3_pa; pd_entry_t l2e; u_int l2_slot; /* Make sure there is a valid L1 -> L2 table */ pmap_bootstrap_l1_table(state); /* Link the level 2 table to a level 3 table */ l2_slot = pmap_l2_index(state->va); if (l2_slot != state->l2_slot) { /* See pmap_bootstrap_l0_table for a description */ MPASS(state->l2_slot < l2_slot || state->l2_slot == Ln_ENTRIES || state->dmap_valid); /* Check the existing L2 entry */ state->l2_slot = l2_slot; if (state->dmap_valid) { l2e = state->l2[l2_slot]; if ((l2e & ATTR_DESCR_VALID) != 0) { MPASS((l2e & ATTR_DESCR_MASK) == L2_TABLE); l3_pa = PTE_TO_PHYS(l2e); state->l3 = (pt_entry_t *)PHYS_TO_DMAP(l3_pa); return; } } /* Create a new L2 table entry */ state->l3 = (pt_entry_t *)state->freemempos; memset(state->l3, 0, PAGE_SIZE); state->freemempos += PAGE_SIZE; l3_pa = pmap_early_vtophys((vm_offset_t)state->l3); MPASS((l3_pa & Ln_TABLE_MASK) == 0); MPASS(state->l2[l2_slot] == 0); pmap_store(&state->l2[l2_slot], PHYS_TO_PTE(l3_pa) | state->table_attrs | L2_TABLE); } KASSERT(state->l3 != NULL, ("%s: NULL l3", __func__)); } static void pmap_bootstrap_l2_block(struct pmap_bootstrap_state *state, int i) { u_int l2_slot; bool first; if ((physmap[i + 1] - state->pa) < L2_SIZE) return; /* Make sure there is a valid L1 table */ pmap_bootstrap_l1_table(state); MPASS((state->va & L2_OFFSET) == 0); for (first = true; state->va < DMAP_MAX_ADDRESS && (physmap[i + 1] - state->pa) >= L2_SIZE; state->va += L2_SIZE, state->pa += L2_SIZE) { /* * Stop if we are about to walk off the end of what the * current L1 slot can address. */ if (!first && (state->pa & L1_OFFSET) == 0) break; first = false; l2_slot = pmap_l2_index(state->va); MPASS((state->pa & L2_OFFSET) == 0); MPASS(state->l2[l2_slot] == 0); pmap_store(&state->l2[l2_slot], PHYS_TO_PTE(state->pa) | ATTR_DEFAULT | ATTR_S1_XN | ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | L2_BLOCK); } MPASS(state->va == (state->pa - dmap_phys_base + DMAP_MIN_ADDRESS)); } static void pmap_bootstrap_l3_page(struct pmap_bootstrap_state *state, int i) { u_int l3_slot; bool first; if ((physmap[i + 1] - state->pa) < L3_SIZE) return; /* Make sure there is a valid L2 table */ pmap_bootstrap_l2_table(state); MPASS((state->va & L3_OFFSET) == 0); for (first = true; state->va < DMAP_MAX_ADDRESS && (physmap[i + 1] - state->pa) >= L3_SIZE; state->va += L3_SIZE, state->pa += L3_SIZE) { /* * Stop if we are about to walk off the end of what the * current L2 slot can address. */ if (!first && (state->pa & L2_OFFSET) == 0) break; first = false; l3_slot = pmap_l3_index(state->va); MPASS((state->pa & L3_OFFSET) == 0); MPASS(state->l3[l3_slot] == 0); pmap_store(&state->l3[l3_slot], PHYS_TO_PTE(state->pa) | ATTR_DEFAULT | ATTR_S1_XN | ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | L3_PAGE); } MPASS(state->va == (state->pa - dmap_phys_base + DMAP_MIN_ADDRESS)); } static void pmap_bootstrap_dmap(vm_paddr_t min_pa) { int i; dmap_phys_base = min_pa & ~L1_OFFSET; dmap_phys_max = 0; dmap_max_addr = 0; for (i = 0; i < (physmap_idx * 2); i += 2) { bs_state.pa = physmap[i] & ~L3_OFFSET; bs_state.va = bs_state.pa - dmap_phys_base + DMAP_MIN_ADDRESS; /* Create L3 mappings at the start of the region */ if ((bs_state.pa & L2_OFFSET) != 0) pmap_bootstrap_l3_page(&bs_state, i); MPASS(bs_state.pa <= physmap[i + 1]); if (L1_BLOCKS_SUPPORTED) { /* Create L2 mappings at the start of the region */ if ((bs_state.pa & L1_OFFSET) != 0) pmap_bootstrap_l2_block(&bs_state, i); MPASS(bs_state.pa <= physmap[i + 1]); /* Create the main L1 block mappings */ for (; bs_state.va < DMAP_MAX_ADDRESS && (physmap[i + 1] - bs_state.pa) >= L1_SIZE; bs_state.va += L1_SIZE, bs_state.pa += L1_SIZE) { /* Make sure there is a valid L1 table */ pmap_bootstrap_l0_table(&bs_state); MPASS((bs_state.pa & L1_OFFSET) == 0); pmap_store( &bs_state.l1[pmap_l1_index(bs_state.va)], PHYS_TO_PTE(bs_state.pa) | ATTR_DEFAULT | ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | ATTR_S1_XN | L1_BLOCK); } MPASS(bs_state.pa <= physmap[i + 1]); /* Create L2 mappings at the end of the region */ pmap_bootstrap_l2_block(&bs_state, i); } else { while (bs_state.va < DMAP_MAX_ADDRESS && (physmap[i + 1] - bs_state.pa) >= L2_SIZE) { pmap_bootstrap_l2_block(&bs_state, i); } } MPASS(bs_state.pa <= physmap[i + 1]); /* Create L3 mappings at the end of the region */ pmap_bootstrap_l3_page(&bs_state, i); MPASS(bs_state.pa == physmap[i + 1]); if (bs_state.pa > dmap_phys_max) { dmap_phys_max = bs_state.pa; dmap_max_addr = bs_state.va; } } cpu_tlb_flushID(); } static void pmap_bootstrap_l2(vm_offset_t va) { KASSERT((va & L1_OFFSET) == 0, ("Invalid virtual address")); /* Leave bs_state.pa as it's only needed to bootstrap blocks and pages*/ bs_state.va = va; for (; bs_state.va < VM_MAX_KERNEL_ADDRESS; bs_state.va += L1_SIZE) pmap_bootstrap_l1_table(&bs_state); } static void pmap_bootstrap_l3(vm_offset_t va) { KASSERT((va & L2_OFFSET) == 0, ("Invalid virtual address")); /* Leave bs_state.pa as it's only needed to bootstrap blocks and pages*/ bs_state.va = va; for (; bs_state.va < VM_MAX_KERNEL_ADDRESS; bs_state.va += L2_SIZE) pmap_bootstrap_l2_table(&bs_state); } #ifdef KASAN static void pmap_bootstrap_allocate_kasan_l2(vm_paddr_t start_pa, vm_paddr_t end_pa, vm_offset_t *start_va, int *nkasan_l2) { int i; vm_paddr_t pa; vm_offset_t va; pd_entry_t *l2; va = *start_va; pa = rounddown2(end_pa - L2_SIZE, L2_SIZE); l2 = pmap_l2(kernel_pmap, va); for (i = 0; pa >= start_pa && i < *nkasan_l2; i++, va += L2_SIZE, pa -= L2_SIZE, l2++) { /* * KASAN stack checking results in us having already allocated * part of our shadow map, so we can just skip those segments. */ if ((pmap_load(l2) & ATTR_DESCR_VALID) != 0) { pa += L2_SIZE; continue; } pmap_store(l2, PHYS_TO_PTE(pa) | PMAP_SAN_PTE_BITS | L2_BLOCK); } /* * Ended the allocation due to start_pa constraint, rather than because * we allocated everything. Adjust back up to the start_pa and remove * the invalid L2 block from our accounting. */ if (pa < start_pa) { va += L2_SIZE; i--; pa = start_pa; } bzero((void *)PHYS_TO_DMAP(pa), i * L2_SIZE); physmem_exclude_region(pa, i * L2_SIZE, EXFLAG_NOALLOC); *nkasan_l2 -= i; *start_va = va; } #endif /* * Bootstrap the system enough to run with virtual memory. */ void pmap_bootstrap(vm_paddr_t kernstart, vm_size_t kernlen) { vm_offset_t dpcpu, msgbufpv; vm_paddr_t start_pa, pa, min_pa; uint64_t kern_delta; int i; /* Verify that the ASID is set through TTBR0. */ KASSERT((READ_SPECIALREG(tcr_el1) & TCR_A1) == 0, ("pmap_bootstrap: TCR_EL1.A1 != 0")); kern_delta = KERNBASE - kernstart; printf("pmap_bootstrap %lx %lx\n", kernstart, kernlen); printf("%lx\n", (KERNBASE >> L1_SHIFT) & Ln_ADDR_MASK); /* Set this early so we can use the pagetable walking functions */ kernel_pmap_store.pm_l0 = pagetable_l0_ttbr1; PMAP_LOCK_INIT(kernel_pmap); kernel_pmap->pm_l0_paddr = pmap_early_vtophys((vm_offset_t)kernel_pmap_store.pm_l0); kernel_pmap->pm_cookie = COOKIE_FROM(-1, INT_MIN); kernel_pmap->pm_stage = PM_STAGE1; kernel_pmap->pm_levels = 4; kernel_pmap->pm_ttbr = kernel_pmap->pm_l0_paddr; kernel_pmap->pm_asid_set = &asids; /* Assume the address we were loaded to is a valid physical address */ min_pa = KERNBASE - kern_delta; physmap_idx = physmem_avail(physmap, nitems(physmap)); physmap_idx /= 2; /* * Find the minimum physical address. physmap is sorted, * but may contain empty ranges. */ for (i = 0; i < physmap_idx * 2; i += 2) { if (physmap[i] == physmap[i + 1]) continue; if (physmap[i] <= min_pa) min_pa = physmap[i]; } bs_state.freemempos = KERNBASE + kernlen; bs_state.freemempos = roundup2(bs_state.freemempos, PAGE_SIZE); /* Create a direct map region early so we can use it for pa -> va */ pmap_bootstrap_dmap(min_pa); bs_state.dmap_valid = true; /* * We only use PXN when we know nothing will be executed from it, e.g. * the DMAP region. */ bs_state.table_attrs &= ~TATTR_PXN_TABLE; start_pa = pa = KERNBASE - kern_delta; /* * Create the l2 tables up to VM_MAX_KERNEL_ADDRESS. We assume that the * loader allocated the first and only l2 page table page used to map * the kernel, preloaded files and module metadata. */ pmap_bootstrap_l2(KERNBASE + L1_SIZE); /* And the l3 tables for the early devmap */ pmap_bootstrap_l3(VM_MAX_KERNEL_ADDRESS - (PMAP_MAPDEV_EARLY_SIZE)); cpu_tlb_flushID(); #define alloc_pages(var, np) \ (var) = bs_state.freemempos; \ bs_state.freemempos += (np * PAGE_SIZE); \ memset((char *)(var), 0, ((np) * PAGE_SIZE)); /* Allocate dynamic per-cpu area. */ alloc_pages(dpcpu, DPCPU_SIZE / PAGE_SIZE); dpcpu_init((void *)dpcpu, 0); /* Allocate memory for the msgbuf, e.g. for /sbin/dmesg */ alloc_pages(msgbufpv, round_page(msgbufsize) / PAGE_SIZE); msgbufp = (void *)msgbufpv; /* Reserve some VA space for early BIOS/ACPI mapping */ preinit_map_va = roundup2(bs_state.freemempos, L2_SIZE); virtual_avail = preinit_map_va + PMAP_PREINIT_MAPPING_SIZE; virtual_avail = roundup2(virtual_avail, L1_SIZE); virtual_end = VM_MAX_KERNEL_ADDRESS - (PMAP_MAPDEV_EARLY_SIZE); kernel_vm_end = virtual_avail; pa = pmap_early_vtophys(bs_state.freemempos); physmem_exclude_region(start_pa, pa - start_pa, EXFLAG_NOALLOC); cpu_tlb_flushID(); } #if defined(KASAN) /* * Finish constructing the initial shadow map: * - Count how many pages from KERNBASE to virtual_avail (scaled for * shadow map) * - Map that entire range using L2 superpages. */ void pmap_bootstrap_san(vm_paddr_t kernstart) { vm_offset_t va; int i, shadow_npages, nkasan_l2; /* * Rebuild physmap one more time, we may have excluded more regions from * allocation since pmap_bootstrap(). */ bzero(physmap, sizeof(physmap)); physmap_idx = physmem_avail(physmap, nitems(physmap)); physmap_idx /= 2; shadow_npages = (virtual_avail - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE; shadow_npages = howmany(shadow_npages, KASAN_SHADOW_SCALE); nkasan_l2 = howmany(shadow_npages, Ln_ENTRIES); /* Map the valid KVA up to this point. */ va = KASAN_MIN_ADDRESS; /* * Find a slot in the physmap large enough for what we needed. We try to put * the shadow map as high up as we can to avoid depleting the lower 4GB in case * it's needed for, e.g., an xhci controller that can only do 32-bit DMA. */ for (i = (physmap_idx * 2) - 2; i >= 0 && nkasan_l2 > 0; i -= 2) { vm_paddr_t plow, phigh; /* L2 mappings must be backed by memory that is L2-aligned */ plow = roundup2(physmap[i], L2_SIZE); phigh = physmap[i + 1]; if (plow >= phigh) continue; if (kernstart >= plow && kernstart < phigh) phigh = kernstart; if (phigh - plow >= L2_SIZE) pmap_bootstrap_allocate_kasan_l2(plow, phigh, &va, &nkasan_l2); } if (nkasan_l2 != 0) panic("Could not find phys region for shadow map"); /* * Done. We should now have a valid shadow address mapped for all KVA * that has been mapped so far, i.e., KERNBASE to virtual_avail. Thus, * shadow accesses by the kasan(9) runtime will succeed for this range. * When the kernel virtual address range is later expanded, as will * happen in vm_mem_init(), the shadow map will be grown as well. This * is handled by pmap_san_enter(). */ } #endif /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pv_memattr = VM_MEMATTR_WRITE_BACK; } static void pmap_init_asids(struct asid_set *set, int bits) { int i; set->asid_bits = bits; /* * We may be too early in the overall initialization process to use * bit_alloc(). */ set->asid_set_size = 1 << set->asid_bits; set->asid_set = kmem_malloc(bitstr_size(set->asid_set_size), M_WAITOK | M_ZERO); for (i = 0; i < ASID_FIRST_AVAILABLE; i++) bit_set(set->asid_set, i); set->asid_next = ASID_FIRST_AVAILABLE; mtx_init(&set->asid_set_mutex, "asid set", NULL, MTX_SPIN); } static void pmap_init_pv_table(void) { struct vm_phys_seg *seg, *next_seg; struct pmap_large_md_page *pvd; vm_size_t s; int domain, i, j, pages; /* * We strongly depend on the size being a power of two, so the assert * is overzealous. However, should the struct be resized to a * different power of two, the code below needs to be revisited. */ CTASSERT((sizeof(*pvd) == 64)); /* * Calculate the size of the array. */ s = 0; for (i = 0; i < vm_phys_nsegs; i++) { seg = &vm_phys_segs[i]; pages = pmap_l2_pindex(roundup2(seg->end, L2_SIZE)) - pmap_l2_pindex(seg->start); s += round_page(pages * sizeof(*pvd)); } pv_table = (struct pmap_large_md_page *)kva_alloc(s); if (pv_table == NULL) panic("%s: kva_alloc failed\n", __func__); /* * Iterate physical segments to allocate domain-local memory for PV * list headers. */ pvd = pv_table; for (i = 0; i < vm_phys_nsegs; i++) { seg = &vm_phys_segs[i]; pages = pmap_l2_pindex(roundup2(seg->end, L2_SIZE)) - pmap_l2_pindex(seg->start); domain = seg->domain; s = round_page(pages * sizeof(*pvd)); for (j = 0; j < s; j += PAGE_SIZE) { vm_page_t m = vm_page_alloc_noobj_domain(domain, VM_ALLOC_ZERO); if (m == NULL) panic("failed to allocate PV table page"); pmap_qenter((vm_offset_t)pvd + j, &m, 1); } for (j = 0; j < s / sizeof(*pvd); j++) { rw_init_flags(&pvd->pv_lock, "pmap pv list", RW_NEW); TAILQ_INIT(&pvd->pv_page.pv_list); pvd++; } } pvd = &pv_dummy_large; memset(pvd, 0, sizeof(*pvd)); rw_init_flags(&pvd->pv_lock, "pmap pv list dummy", RW_NEW); TAILQ_INIT(&pvd->pv_page.pv_list); /* * Set pointers from vm_phys_segs to pv_table. */ for (i = 0, pvd = pv_table; i < vm_phys_nsegs; i++) { seg = &vm_phys_segs[i]; seg->md_first = pvd; pvd += pmap_l2_pindex(roundup2(seg->end, L2_SIZE)) - pmap_l2_pindex(seg->start); /* * If there is a following segment, and the final * superpage of this segment and the initial superpage * of the next segment are the same then adjust the * pv_table entry for that next segment down by one so * that the pv_table entries will be shared. */ if (i + 1 < vm_phys_nsegs) { next_seg = &vm_phys_segs[i + 1]; if (pmap_l2_pindex(roundup2(seg->end, L2_SIZE)) - 1 == pmap_l2_pindex(next_seg->start)) { pvd--; } } } } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { uint64_t mmfr1; int i, vmid_bits; /* * Are large page mappings enabled? */ TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled); if (superpages_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = L2_SIZE; if (L1_BLOCKS_SUPPORTED) { KASSERT(MAXPAGESIZES > 2 && pagesizes[2] == 0, ("pmap_init: can't assign to pagesizes[2]")); pagesizes[2] = L1_SIZE; } } /* * Initialize the ASID allocator. */ pmap_init_asids(&asids, (READ_SPECIALREG(tcr_el1) & TCR_ASID_16) != 0 ? 16 : 8); if (has_hyp()) { mmfr1 = READ_SPECIALREG(id_aa64mmfr1_el1); vmid_bits = 8; if (ID_AA64MMFR1_VMIDBits_VAL(mmfr1) == ID_AA64MMFR1_VMIDBits_16) vmid_bits = 16; pmap_init_asids(&vmids, vmid_bits); } /* * Initialize pv chunk lists. */ for (i = 0; i < PMAP_MEMDOM; i++) { mtx_init(&pv_chunks[i].pvc_lock, "pmap pv chunk list", NULL, MTX_DEF); TAILQ_INIT(&pv_chunks[i].pvc_list); } pmap_init_pv_table(); vm_initialized = 1; } static SYSCTL_NODE(_vm_pmap, OID_AUTO, l2, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "2MB page mapping counters"); static u_long pmap_l2_demotions; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, demotions, CTLFLAG_RD, &pmap_l2_demotions, 0, "2MB page demotions"); static u_long pmap_l2_mappings; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, mappings, CTLFLAG_RD, &pmap_l2_mappings, 0, "2MB page mappings"); static u_long pmap_l2_p_failures; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_l2_p_failures, 0, "2MB page promotion failures"); static u_long pmap_l2_promotions; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, promotions, CTLFLAG_RD, &pmap_l2_promotions, 0, "2MB page promotions"); /* * If the given value for "final_only" is false, then any cached intermediate- * level entries, i.e., L{0,1,2}_TABLE entries, are invalidated in addition to * any cached final-level entry, i.e., either an L{1,2}_BLOCK or L3_PAGE entry. * Otherwise, just the cached final-level entry is invalidated. */ static __inline void pmap_s1_invalidate_kernel(uint64_t r, bool final_only) { if (final_only) __asm __volatile("tlbi vaale1is, %0" : : "r" (r)); else __asm __volatile("tlbi vaae1is, %0" : : "r" (r)); } static __inline void pmap_s1_invalidate_user(uint64_t r, bool final_only) { if (final_only) __asm __volatile("tlbi vale1is, %0" : : "r" (r)); else __asm __volatile("tlbi vae1is, %0" : : "r" (r)); } /* * Invalidates any cached final- and optionally intermediate-level TLB entries * for the specified virtual address in the given virtual address space. */ static __inline void pmap_s1_invalidate_page(pmap_t pmap, vm_offset_t va, bool final_only) { uint64_t r; PMAP_ASSERT_STAGE1(pmap); dsb(ishst); r = TLBI_VA(va); if (pmap == kernel_pmap) { pmap_s1_invalidate_kernel(r, final_only); } else { r |= ASID_TO_OPERAND(COOKIE_TO_ASID(pmap->pm_cookie)); pmap_s1_invalidate_user(r, final_only); } dsb(ish); isb(); } static __inline void pmap_s2_invalidate_page(pmap_t pmap, vm_offset_t va, bool final_only) { PMAP_ASSERT_STAGE2(pmap); MPASS(pmap_stage2_invalidate_range != NULL); pmap_stage2_invalidate_range(pmap_to_ttbr0(pmap), va, va + PAGE_SIZE, final_only); } static __inline void pmap_invalidate_page(pmap_t pmap, vm_offset_t va, bool final_only) { if (pmap->pm_stage == PM_STAGE1) pmap_s1_invalidate_page(pmap, va, final_only); else pmap_s2_invalidate_page(pmap, va, final_only); } /* * Invalidates any cached final- and optionally intermediate-level TLB entries * for the specified virtual address range in the given virtual address space. */ static __inline void pmap_s1_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, bool final_only) { uint64_t end, r, start; PMAP_ASSERT_STAGE1(pmap); dsb(ishst); if (pmap == kernel_pmap) { start = TLBI_VA(sva); end = TLBI_VA(eva); for (r = start; r < end; r += TLBI_VA_L3_INCR) pmap_s1_invalidate_kernel(r, final_only); } else { start = end = ASID_TO_OPERAND(COOKIE_TO_ASID(pmap->pm_cookie)); start |= TLBI_VA(sva); end |= TLBI_VA(eva); for (r = start; r < end; r += TLBI_VA_L3_INCR) pmap_s1_invalidate_user(r, final_only); } dsb(ish); isb(); } static __inline void pmap_s2_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, bool final_only) { PMAP_ASSERT_STAGE2(pmap); MPASS(pmap_stage2_invalidate_range != NULL); pmap_stage2_invalidate_range(pmap_to_ttbr0(pmap), sva, eva, final_only); } static __inline void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, bool final_only) { if (pmap->pm_stage == PM_STAGE1) pmap_s1_invalidate_range(pmap, sva, eva, final_only); else pmap_s2_invalidate_range(pmap, sva, eva, final_only); } /* * Invalidates all cached intermediate- and final-level TLB entries for the * given virtual address space. */ static __inline void pmap_s1_invalidate_all(pmap_t pmap) { uint64_t r; PMAP_ASSERT_STAGE1(pmap); dsb(ishst); if (pmap == kernel_pmap) { __asm __volatile("tlbi vmalle1is"); } else { r = ASID_TO_OPERAND(COOKIE_TO_ASID(pmap->pm_cookie)); __asm __volatile("tlbi aside1is, %0" : : "r" (r)); } dsb(ish); isb(); } static __inline void pmap_s2_invalidate_all(pmap_t pmap) { PMAP_ASSERT_STAGE2(pmap); MPASS(pmap_stage2_invalidate_all != NULL); pmap_stage2_invalidate_all(pmap_to_ttbr0(pmap)); } static __inline void pmap_invalidate_all(pmap_t pmap) { if (pmap->pm_stage == PM_STAGE1) pmap_s1_invalidate_all(pmap); else pmap_s2_invalidate_all(pmap); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { pt_entry_t *pte, tpte; vm_paddr_t pa; int lvl; pa = 0; PMAP_LOCK(pmap); /* * Find the block or page map for this virtual address. pmap_pte * will return either a valid block/page entry, or NULL. */ pte = pmap_pte(pmap, va, &lvl); if (pte != NULL) { tpte = pmap_load(pte); pa = PTE_TO_PHYS(tpte); switch(lvl) { case 1: PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT((tpte & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_extract: Invalid L1 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L1_OFFSET); break; case 2: KASSERT((tpte & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_extract: Invalid L2 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L2_OFFSET); break; case 3: KASSERT((tpte & ATTR_DESCR_MASK) == L3_PAGE, ("pmap_extract: Invalid L3 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L3_OFFSET); break; } } PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pt_entry_t *pte, tpte; vm_offset_t off; vm_page_t m; int lvl; bool use; m = NULL; PMAP_LOCK(pmap); pte = pmap_pte(pmap, va, &lvl); if (pte != NULL) { tpte = pmap_load(pte); KASSERT(lvl > 0 && lvl <= 3, ("pmap_extract_and_hold: Invalid level %d", lvl)); /* * Check that the pte is either a L3 page, or a L1 or L2 block * entry. We can assume L1_BLOCK == L2_BLOCK. */ KASSERT((lvl == 3 && (tpte & ATTR_DESCR_MASK) == L3_PAGE) || (lvl < 3 && (tpte & ATTR_DESCR_MASK) == L1_BLOCK), ("pmap_extract_and_hold: Invalid pte at L%d: %lx", lvl, tpte & ATTR_DESCR_MASK)); use = false; if ((prot & VM_PROT_WRITE) == 0) use = true; else if (pmap->pm_stage == PM_STAGE1 && (tpte & ATTR_S1_AP_RW_BIT) == ATTR_S1_AP(ATTR_S1_AP_RW)) use = true; else if (pmap->pm_stage == PM_STAGE2 && ((tpte & ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE)) == ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE))) use = true; if (use) { switch (lvl) { case 1: off = va & L1_OFFSET; break; case 2: off = va & L2_OFFSET; break; case 3: default: off = 0; } m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tpte) | off); if (m != NULL && !vm_page_wire_mapped(m)) m = NULL; } } PMAP_UNLOCK(pmap); return (m); } /* * Walks the page tables to translate a kernel virtual address to a * physical address. Returns true if the kva is valid and stores the * physical address in pa if it is not NULL. * * See the comment above data_abort() for the rationale for specifying * NO_PERTHREAD_SSP here. */ bool NO_PERTHREAD_SSP pmap_klookup(vm_offset_t va, vm_paddr_t *pa) { pt_entry_t *pte, tpte; register_t intr; uint64_t par; /* * Disable interrupts so we don't get interrupted between asking * for address translation, and getting the result back. */ intr = intr_disable(); par = arm64_address_translate_s1e1r(va); intr_restore(intr); if (PAR_SUCCESS(par)) { if (pa != NULL) *pa = (par & PAR_PA_MASK) | (va & PAR_LOW_MASK); return (true); } /* * Fall back to walking the page table. The address translation * instruction may fail when the page is in a break-before-make * sequence. As we only clear the valid bit in said sequence we * can walk the page table to find the physical address. */ pte = pmap_l1(kernel_pmap, va); if (pte == NULL) return (false); /* * A concurrent pmap_update_entry() will clear the entry's valid bit * but leave the rest of the entry unchanged. Therefore, we treat a * non-zero entry as being valid, and we ignore the valid bit when * determining whether the entry maps a block, page, or table. */ tpte = pmap_load(pte); if (tpte == 0) return (false); if ((tpte & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_BLOCK) { if (pa != NULL) *pa = PTE_TO_PHYS(tpte) | (va & L1_OFFSET); return (true); } pte = pmap_l1_to_l2(&tpte, va); tpte = pmap_load(pte); if (tpte == 0) return (false); if ((tpte & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_BLOCK) { if (pa != NULL) *pa = PTE_TO_PHYS(tpte) | (va & L2_OFFSET); return (true); } pte = pmap_l2_to_l3(&tpte, va); tpte = pmap_load(pte); if (tpte == 0) return (false); if (pa != NULL) *pa = PTE_TO_PHYS(tpte) | (va & L3_OFFSET); return (true); } vm_paddr_t pmap_kextract(vm_offset_t va) { vm_paddr_t pa; if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) return (DMAP_TO_PHYS(va)); if (pmap_klookup(va, &pa) == false) return (0); return (pa); } /*************************************************** * Low level mapping routines..... ***************************************************/ void pmap_kenter(vm_offset_t sva, vm_size_t size, vm_paddr_t pa, int mode) { pd_entry_t *pde; pt_entry_t *pte, attr; vm_offset_t va; int lvl; KASSERT((pa & L3_OFFSET) == 0, ("pmap_kenter: Invalid physical address")); KASSERT((sva & L3_OFFSET) == 0, ("pmap_kenter: Invalid virtual address")); KASSERT((size & PAGE_MASK) == 0, ("pmap_kenter: Mapping is not page-sized")); attr = ATTR_DEFAULT | ATTR_S1_AP(ATTR_S1_AP_RW) | ATTR_S1_XN | ATTR_S1_IDX(mode) | L3_PAGE; va = sva; while (size != 0) { pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_kenter: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_kenter: Invalid level %d", lvl)); pte = pmap_l2_to_l3(pde, va); pmap_load_store(pte, PHYS_TO_PTE(pa) | attr); va += PAGE_SIZE; pa += PAGE_SIZE; size -= PAGE_SIZE; } pmap_s1_invalidate_range(kernel_pmap, sva, va, true); } void pmap_kenter_device(vm_offset_t sva, vm_size_t size, vm_paddr_t pa) { pmap_kenter(sva, size, pa, VM_MEMATTR_DEVICE); } /* * Remove a page from the kernel pagetables. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *pte; pte = pmap_pte_exists(kernel_pmap, va, 3, __func__); pmap_clear(pte); pmap_s1_invalidate_page(kernel_pmap, va, true); } void pmap_kremove_device(vm_offset_t sva, vm_size_t size) { pt_entry_t *pte; vm_offset_t va; KASSERT((sva & L3_OFFSET) == 0, ("pmap_kremove_device: Invalid virtual address")); KASSERT((size & PAGE_MASK) == 0, ("pmap_kremove_device: Mapping is not page-sized")); va = sva; while (size != 0) { pte = pmap_pte_exists(kernel_pmap, va, 3, __func__); pmap_clear(pte); va += PAGE_SIZE; size -= PAGE_SIZE; } pmap_s1_invalidate_range(kernel_pmap, sva, va, true); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return PHYS_TO_DMAP(start); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pd_entry_t *pde; pt_entry_t *pte, pa, attr; vm_offset_t va; vm_page_t m; int i, lvl; va = sva; for (i = 0; i < count; i++) { pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_qenter: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_qenter: Invalid level %d", lvl)); m = ma[i]; pa = VM_PAGE_TO_PHYS(m); attr = ATTR_DEFAULT | ATTR_S1_AP(ATTR_S1_AP_RW) | ATTR_S1_XN | ATTR_S1_IDX(m->md.pv_memattr) | L3_PAGE; pte = pmap_l2_to_l3(pde, va); pmap_load_store(pte, PHYS_TO_PTE(pa) | attr); va += L3_SIZE; } pmap_s1_invalidate_range(kernel_pmap, sva, va, true); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. */ void pmap_qremove(vm_offset_t sva, int count) { pt_entry_t *pte; vm_offset_t va; KASSERT(ADDR_IS_CANONICAL(sva), ("%s: Address not in canonical form: %lx", __func__, sva)); KASSERT(ADDR_IS_KERNEL(sva), ("usermode va %lx", sva)); va = sva; while (count-- > 0) { pte = pmap_pte_exists(kernel_pmap, va, 3, NULL); if (pte != NULL) { pmap_clear(pte); } va += PAGE_SIZE; } pmap_s1_invalidate_range(kernel_pmap, sva, va, true); } /*************************************************** * Page table page management routines..... ***************************************************/ /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Decrements a page table page's reference count, which is used to record the * number of valid page table entries within the page. If the reference count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { --m->ref_count; if (m->ref_count == 0) { _pmap_unwire_l3(pmap, va, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * unmap the page table page */ if (m->pindex >= (NUL2E + NUL1E)) { /* l1 page */ pd_entry_t *l0; l0 = pmap_l0(pmap, va); pmap_clear(l0); } else if (m->pindex >= NUL2E) { /* l2 page */ pd_entry_t *l1; l1 = pmap_l1(pmap, va); pmap_clear(l1); } else { /* l3 page */ pd_entry_t *l2; l2 = pmap_l2(pmap, va); pmap_clear(l2); } pmap_resident_count_dec(pmap, 1); if (m->pindex < NUL2E) { /* We just released an l3, unhold the matching l2 */ pd_entry_t *l1, tl1; vm_page_t l2pg; l1 = pmap_l1(pmap, va); tl1 = pmap_load(l1); l2pg = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tl1)); pmap_unwire_l3(pmap, va, l2pg, free); } else if (m->pindex < (NUL2E + NUL1E)) { /* We just released an l2, unhold the matching l1 */ pd_entry_t *l0, tl0; vm_page_t l1pg; l0 = pmap_l0(pmap, va); tl0 = pmap_load(l0); l1pg = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tl0)); pmap_unwire_l3(pmap, va, l1pg, free); } pmap_invalidate_page(pmap, va, false); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing a page table entry, this routine is used to * conditionally free the page, and manage the reference count. */ static int pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde, struct spglist *free) { vm_page_t mpte; KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); if (ADDR_IS_KERNEL(va)) return (0); KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0")); mpte = PHYS_TO_VM_PAGE(PTE_TO_PHYS(ptepde)); return (pmap_unwire_l3(pmap, va, mpte, free)); } /* * Release a page table page reference after a failed attempt to create a * mapping. */ static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte) { struct spglist free; SLIST_INIT(&free); if (pmap_unwire_l3(pmap, va, mpte, &free)) vm_page_free_pages_toq(&free, true); } void pmap_pinit0(pmap_t pmap) { PMAP_LOCK_INIT(pmap); bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); pmap->pm_l0_paddr = READ_SPECIALREG(ttbr0_el1); pmap->pm_l0 = (pd_entry_t *)PHYS_TO_DMAP(pmap->pm_l0_paddr); vm_radix_init(&pmap->pm_root); pmap->pm_cookie = COOKIE_FROM(ASID_RESERVED_FOR_PID_0, INT_MIN); pmap->pm_stage = PM_STAGE1; pmap->pm_levels = 4; pmap->pm_ttbr = pmap->pm_l0_paddr; pmap->pm_asid_set = &asids; PCPU_SET(curpmap, pmap); } int pmap_pinit_stage(pmap_t pmap, enum pmap_stage stage, int levels) { vm_page_t m; /* * allocate the l0 page */ m = vm_page_alloc_noobj(VM_ALLOC_WAITOK | VM_ALLOC_WIRED | VM_ALLOC_ZERO); pmap->pm_l0_paddr = VM_PAGE_TO_PHYS(m); pmap->pm_l0 = (pd_entry_t *)PHYS_TO_DMAP(pmap->pm_l0_paddr); vm_radix_init(&pmap->pm_root); bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); pmap->pm_cookie = COOKIE_FROM(-1, INT_MAX); MPASS(levels == 3 || levels == 4); pmap->pm_levels = levels; pmap->pm_stage = stage; switch (stage) { case PM_STAGE1: pmap->pm_asid_set = &asids; break; case PM_STAGE2: pmap->pm_asid_set = &vmids; break; default: panic("%s: Invalid pmap type %d", __func__, stage); break; } /* XXX Temporarily disable deferred ASID allocation. */ pmap_alloc_asid(pmap); /* * Allocate the level 1 entry to use as the root. This will increase * the refcount on the level 1 page so it won't be removed until * pmap_release() is called. */ if (pmap->pm_levels == 3) { PMAP_LOCK(pmap); m = _pmap_alloc_l3(pmap, NUL2E + NUL1E, NULL); PMAP_UNLOCK(pmap); } pmap->pm_ttbr = VM_PAGE_TO_PHYS(m); return (1); } int pmap_pinit(pmap_t pmap) { return (pmap_pinit_stage(pmap, PM_STAGE1, 4)); } /* * This routine is called if the desired page table page does not exist. * * If page table page allocation fails, this routine may sleep before * returning NULL. It sleeps only if a lock pointer was given. * * Note: If a page allocation fails at page table level two or three, * one or two pages may be held during the wait, only to be released * afterwards. This conservative approach is easily argued to avoid * race conditions. */ static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp) { vm_page_t m, l1pg, l2pg; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Allocate a page table page. */ if ((m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { if (lockp != NULL) { RELEASE_PV_LIST_LOCK(lockp); PMAP_UNLOCK(pmap); vm_wait(NULL); PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, the page table * page may have been allocated. */ return (NULL); } m->pindex = ptepindex; /* * Because of AArch64's weak memory consistency model, we must have a * barrier here to ensure that the stores for zeroing "m", whether by * pmap_zero_page() or an earlier function, are visible before adding * "m" to the page table. Otherwise, a page table walk by another * processor's MMU could see the mapping to "m" and a stale, non-zero * PTE within "m". */ dmb(ishst); /* * Map the pagetable page into the process address space, if * it isn't already there. */ if (ptepindex >= (NUL2E + NUL1E)) { pd_entry_t *l0p, l0e; vm_pindex_t l0index; l0index = ptepindex - (NUL2E + NUL1E); l0p = &pmap->pm_l0[l0index]; KASSERT((pmap_load(l0p) & ATTR_DESCR_VALID) == 0, ("%s: L0 entry %#lx is valid", __func__, pmap_load(l0p))); l0e = PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | L0_TABLE; /* * Mark all kernel memory as not accessible from userspace * and userspace memory as not executable from the kernel. * This has been done for the bootstrap L0 entries in * locore.S. */ if (pmap == kernel_pmap) l0e |= TATTR_UXN_TABLE | TATTR_AP_TABLE_NO_EL0; else l0e |= TATTR_PXN_TABLE; pmap_store(l0p, l0e); } else if (ptepindex >= NUL2E) { vm_pindex_t l0index, l1index; pd_entry_t *l0, *l1; pd_entry_t tl0; l1index = ptepindex - NUL2E; l0index = l1index >> Ln_ENTRIES_SHIFT; l0 = &pmap->pm_l0[l0index]; tl0 = pmap_load(l0); if (tl0 == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUL2E + NUL1E + l0index, lockp) == NULL) { vm_page_unwire_noq(m); vm_page_free_zero(m); return (NULL); } } else { l1pg = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tl0)); l1pg->ref_count++; } l1 = (pd_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(pmap_load(l0))); l1 = &l1[ptepindex & Ln_ADDR_MASK]; KASSERT((pmap_load(l1) & ATTR_DESCR_VALID) == 0, ("%s: L1 entry %#lx is valid", __func__, pmap_load(l1))); pmap_store(l1, PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | L1_TABLE); } else { vm_pindex_t l0index, l1index; pd_entry_t *l0, *l1, *l2; pd_entry_t tl0, tl1; l1index = ptepindex >> Ln_ENTRIES_SHIFT; l0index = l1index >> Ln_ENTRIES_SHIFT; l0 = &pmap->pm_l0[l0index]; tl0 = pmap_load(l0); if (tl0 == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUL2E + l1index, lockp) == NULL) { vm_page_unwire_noq(m); vm_page_free_zero(m); return (NULL); } tl0 = pmap_load(l0); l1 = (pd_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(tl0)); l1 = &l1[l1index & Ln_ADDR_MASK]; } else { l1 = (pd_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(tl0)); l1 = &l1[l1index & Ln_ADDR_MASK]; tl1 = pmap_load(l1); if (tl1 == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUL2E + l1index, lockp) == NULL) { vm_page_unwire_noq(m); vm_page_free_zero(m); return (NULL); } } else { l2pg = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tl1)); l2pg->ref_count++; } } l2 = (pd_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(pmap_load(l1))); l2 = &l2[ptepindex & Ln_ADDR_MASK]; KASSERT((pmap_load(l2) & ATTR_DESCR_VALID) == 0, ("%s: L2 entry %#lx is valid", __func__, pmap_load(l2))); pmap_store(l2, PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | L2_TABLE); } pmap_resident_count_inc(pmap, 1); return (m); } static pd_entry_t * pmap_alloc_l2(pmap_t pmap, vm_offset_t va, vm_page_t *l2pgp, struct rwlock **lockp) { pd_entry_t *l1, *l2; vm_page_t l2pg; vm_pindex_t l2pindex; KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); retry: l1 = pmap_l1(pmap, va); if (l1 != NULL && (pmap_load(l1) & ATTR_DESCR_MASK) == L1_TABLE) { l2 = pmap_l1_to_l2(l1, va); if (!ADDR_IS_KERNEL(va)) { /* Add a reference to the L2 page. */ l2pg = PHYS_TO_VM_PAGE(PTE_TO_PHYS(pmap_load(l1))); l2pg->ref_count++; } else l2pg = NULL; } else if (!ADDR_IS_KERNEL(va)) { /* Allocate a L2 page. */ l2pindex = pmap_l2_pindex(va) >> Ln_ENTRIES_SHIFT; l2pg = _pmap_alloc_l3(pmap, NUL2E + l2pindex, lockp); if (l2pg == NULL) { if (lockp != NULL) goto retry; else return (NULL); } l2 = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(l2pg)); l2 = &l2[pmap_l2_index(va)]; } else panic("pmap_alloc_l2: missing page table page for va %#lx", va); *l2pgp = l2pg; return (l2); } static vm_page_t pmap_alloc_l3(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t ptepindex; pd_entry_t *pde, tpde; #ifdef INVARIANTS pt_entry_t *pte; #endif vm_page_t m; int lvl; /* * Calculate pagetable page index */ ptepindex = pmap_l2_pindex(va); retry: /* * Get the page directory entry */ pde = pmap_pde(pmap, va, &lvl); /* * If the page table page is mapped, we just increment the hold count, * and activate it. If we get a level 2 pde it will point to a level 3 * table. */ switch (lvl) { case -1: break; case 0: #ifdef INVARIANTS pte = pmap_l0_to_l1(pde, va); KASSERT(pmap_load(pte) == 0, ("pmap_alloc_l3: TODO: l0 superpages")); #endif break; case 1: #ifdef INVARIANTS pte = pmap_l1_to_l2(pde, va); KASSERT(pmap_load(pte) == 0, ("pmap_alloc_l3: TODO: l1 superpages")); #endif break; case 2: tpde = pmap_load(pde); if (tpde != 0) { m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tpde)); m->ref_count++; return (m); } break; default: panic("pmap_alloc_l3: Invalid level %d", lvl); } /* * Here if the pte page isn't mapped, or if it has been deallocated. */ m = _pmap_alloc_l3(pmap, ptepindex, lockp); if (m == NULL && lockp != NULL) goto retry; return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { boolean_t rv __diagused; struct spglist free; struct asid_set *set; vm_page_t m; int asid; if (pmap->pm_levels != 4) { PMAP_ASSERT_STAGE2(pmap); KASSERT(pmap->pm_stats.resident_count == 1, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); KASSERT((pmap->pm_l0[0] & ATTR_DESCR_VALID) == ATTR_DESCR_VALID, ("pmap_release: Invalid l0 entry: %lx", pmap->pm_l0[0])); SLIST_INIT(&free); m = PHYS_TO_VM_PAGE(pmap->pm_ttbr); PMAP_LOCK(pmap); rv = pmap_unwire_l3(pmap, 0, m, &free); PMAP_UNLOCK(pmap); MPASS(rv == TRUE); vm_page_free_pages_toq(&free, true); } KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap has reserved page table page(s)")); set = pmap->pm_asid_set; KASSERT(set != NULL, ("%s: NULL asid set", __func__)); /* * Allow the ASID to be reused. In stage 2 VMIDs we don't invalidate * the entries when removing them so rely on a later tlb invalidation. * this will happen when updating the VMID generation. Because of this * we don't reuse VMIDs within a generation. */ if (pmap->pm_stage == PM_STAGE1) { mtx_lock_spin(&set->asid_set_mutex); if (COOKIE_TO_EPOCH(pmap->pm_cookie) == set->asid_epoch) { asid = COOKIE_TO_ASID(pmap->pm_cookie); KASSERT(asid >= ASID_FIRST_AVAILABLE && asid < set->asid_set_size, ("pmap_release: pmap cookie has out-of-range asid")); bit_clear(set->asid_set, asid); } mtx_unlock_spin(&set->asid_set_mutex); } m = PHYS_TO_VM_PAGE(pmap->pm_l0_paddr); vm_page_unwire_noq(m); vm_page_free_zero(m); } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; return sysctl_handle_long(oidp, &ksize, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, kvm_size, "LU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return sysctl_handle_long(oidp, &kfree, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, kvm_free, "LU", "Amount of KVM free"); /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t paddr; vm_page_t nkpg; pd_entry_t *l0, *l1, *l2; mtx_assert(&kernel_map->system_mtx, MA_OWNED); addr = roundup2(addr, L2_SIZE); if (addr - 1 >= vm_map_max(kernel_map)) addr = vm_map_max(kernel_map); if (kernel_vm_end < addr) kasan_shadow_map(kernel_vm_end, addr - kernel_vm_end); while (kernel_vm_end < addr) { l0 = pmap_l0(kernel_pmap, kernel_vm_end); KASSERT(pmap_load(l0) != 0, ("pmap_growkernel: No level 0 kernel entry")); l1 = pmap_l0_to_l1(l0, kernel_vm_end); if (pmap_load(l1) == 0) { /* We need a new PDP entry */ nkpg = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); nkpg->pindex = kernel_vm_end >> L1_SHIFT; /* See the dmb() in _pmap_alloc_l3(). */ dmb(ishst); paddr = VM_PAGE_TO_PHYS(nkpg); pmap_store(l1, PHYS_TO_PTE(paddr) | L1_TABLE); continue; /* try again */ } l2 = pmap_l1_to_l2(l1, kernel_vm_end); if (pmap_load(l2) != 0) { kernel_vm_end = (kernel_vm_end + L2_SIZE) & ~L2_OFFSET; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } continue; } nkpg = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); nkpg->pindex = kernel_vm_end >> L2_SHIFT; /* See the dmb() in _pmap_alloc_l3(). */ dmb(ishst); paddr = VM_PAGE_TO_PHYS(nkpg); pmap_store(l2, PHYS_TO_PTE(paddr) | L2_TABLE); kernel_vm_end = (kernel_vm_end + L2_SIZE) & ~L2_OFFSET; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } } } /*************************************************** * page management routines. ***************************************************/ static const uint64_t pc_freemask[_NPCM] = { [0 ... _NPCM - 2] = PC_FREEN, [_NPCM - 1] = PC_FREEL }; #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, "Current number of pv entry chunks"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, "Current number of pv entry chunks allocated"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, "Current number of pv entry chunks frees"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, "Number of times tried to get a chunk page but failed."); static long pv_entry_frees, pv_entry_allocs, pv_entry_count; static int pv_entry_spare; SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, "Current number of pv entry frees"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, "Current number of pv entry allocs"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, "Current number of pv entries"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, "Current number of spare pv entries"); #endif /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. * * Returns NULL if PV entries were reclaimed from the specified pmap. * * We do not, however, unmap 2mpages because subsequent accesses will * allocate per-page pv entries until repromotion occurs, thereby * exacerbating the shortage of free pv entries. */ static vm_page_t reclaim_pv_chunk_domain(pmap_t locked_pmap, struct rwlock **lockp, int domain) { struct pv_chunks_list *pvc; struct pv_chunk *pc, *pc_marker, *pc_marker_end; struct pv_chunk_header pc_marker_b, pc_marker_end_b; struct md_page *pvh; pd_entry_t *pde; pmap_t next_pmap, pmap; pt_entry_t *pte, tpte; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint64_t inuse; int bit, field, freed, lvl; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL")); pmap = NULL; m_pc = NULL; SLIST_INIT(&free); bzero(&pc_marker_b, sizeof(pc_marker_b)); bzero(&pc_marker_end_b, sizeof(pc_marker_end_b)); pc_marker = (struct pv_chunk *)&pc_marker_b; pc_marker_end = (struct pv_chunk *)&pc_marker_end_b; pvc = &pv_chunks[domain]; mtx_lock(&pvc->pvc_lock); pvc->active_reclaims++; TAILQ_INSERT_HEAD(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc_marker_end, pc_lru); while ((pc = TAILQ_NEXT(pc_marker, pc_lru)) != pc_marker_end && SLIST_EMPTY(&free)) { next_pmap = pc->pc_pmap; if (next_pmap == NULL) { /* * The next chunk is a marker. However, it is * not our marker, so active_reclaims must be * > 1. Consequently, the next_chunk code * will not rotate the pv_chunks list. */ goto next_chunk; } mtx_unlock(&pvc->pvc_lock); /* * A pv_chunk can only be removed from the pc_lru list * when both pvc->pvc_lock is owned and the * corresponding pmap is locked. */ if (pmap != next_pmap) { if (pmap != NULL && pmap != locked_pmap) PMAP_UNLOCK(pmap); pmap = next_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) { RELEASE_PV_LIST_LOCK(lockp); PMAP_LOCK(pmap); mtx_lock(&pvc->pvc_lock); continue; } else if (pmap != locked_pmap) { if (PMAP_TRYLOCK(pmap)) { mtx_lock(&pvc->pvc_lock); continue; } else { pmap = NULL; /* pmap is not locked */ mtx_lock(&pvc->pvc_lock); pc = TAILQ_NEXT(pc_marker, pc_lru); if (pc == NULL || pc->pc_pmap != next_pmap) continue; goto next_chunk; } } } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = ffsl(inuse) - 1; pv = &pc->pc_pventry[field * 64 + bit]; va = pv->pv_va; pde = pmap_pde(pmap, va, &lvl); if (lvl != 2) continue; pte = pmap_l2_to_l3(pde, va); tpte = pmap_load(pte); if ((tpte & ATTR_SW_WIRED) != 0) continue; tpte = pmap_load_clear(pte); m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(tpte)); if (pmap_pte_dirty(pmap, tpte)) vm_page_dirty(m); if ((tpte & ATTR_AF) != 0) { pmap_s1_invalidate_page(pmap, va, true); vm_page_aflag_set(m, PGA_REFERENCED); } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt(pmap, va, pmap_load(pde), &free); freed++; } } if (freed == 0) { mtx_lock(&pvc->pvc_lock); goto next_chunk; } /* Every freed mapping is for a 4 KB page. */ pmap_resident_count_dec(pmap, freed); PV_STAT(atomic_add_long(&pv_entry_frees, freed)); PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); if (pc_is_free(pc)) { PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m_pc->phys_addr); mtx_lock(&pvc->pvc_lock); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); break; } TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); mtx_lock(&pvc->pvc_lock); /* One freed pv entry in locked_pmap is sufficient. */ if (pmap == locked_pmap) break; next_chunk: TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_INSERT_AFTER(&pvc->pvc_list, pc, pc_marker, pc_lru); if (pvc->active_reclaims == 1 && pmap != NULL) { /* * Rotate the pv chunks list so that we do not * scan the same pv chunks that could not be * freed (because they contained a wired * and/or superpage mapping) on every * invocation of reclaim_pv_chunk(). */ while ((pc = TAILQ_FIRST(&pvc->pvc_list)) != pc_marker){ MPASS(pc->pc_pmap != NULL); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru); } } } TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_REMOVE(&pvc->pvc_list, pc_marker_end, pc_lru); pvc->active_reclaims--; mtx_unlock(&pvc->pvc_lock); if (pmap != NULL && pmap != locked_pmap) PMAP_UNLOCK(pmap); if (m_pc == NULL && !SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->ref_count = 1; } vm_page_free_pages_toq(&free, true); return (m_pc); } static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) { vm_page_t m; int i, domain; domain = PCPU_GET(domain); for (i = 0; i < vm_ndomains; i++) { m = reclaim_pv_chunk_domain(locked_pmap, lockp, domain); if (m != NULL) break; domain = (domain + 1) % vm_ndomains; } return (m); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_frees, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, 1)); PV_STAT(atomic_subtract_long(&pv_entry_count, 1)); pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; if (!pc_is_free(pc)) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk_dequeued(struct pv_chunk *pc) { vm_page_t m; PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m->phys_addr); vm_page_unwire_noq(m); vm_page_free(m); } static void free_pv_chunk(struct pv_chunk *pc) { struct pv_chunks_list *pvc; pvc = &pv_chunks[pc_to_domain(pc)]; mtx_lock(&pvc->pvc_lock); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); mtx_unlock(&pvc->pvc_lock); free_pv_chunk_dequeued(pc); } static void free_pv_chunk_batch(struct pv_chunklist *batch) { struct pv_chunks_list *pvc; struct pv_chunk *pc, *npc; int i; for (i = 0; i < vm_ndomains; i++) { if (TAILQ_EMPTY(&batch[i])) continue; pvc = &pv_chunks[i]; mtx_lock(&pvc->pvc_lock); TAILQ_FOREACH(pc, &batch[i], pc_list) { TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); } mtx_unlock(&pvc->pvc_lock); } for (i = 0; i < vm_ndomains; i++) { TAILQ_FOREACH_SAFE(pc, &batch[i], pc_list, npc) { free_pv_chunk_dequeued(pc); } } } /* * Returns a new PV entry, allocating a new PV chunk from the system when * needed. If this PV chunk allocation fails and a PV list lock pointer was * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is * returned. * * The given PV list lock may be released. */ static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp) { struct pv_chunks_list *pvc; int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_allocs, 1)); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = ffsl(pc->pc_map[field]) - 1; break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 64 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ if (pc_is_full(pc)) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, 1)); return (pv); } } /* No free items, allocate another chunk */ m = vm_page_alloc_noobj(VM_ALLOC_WIRED); if (m == NULL) { if (lockp == NULL) { PV_STAT(pc_chunk_tryfail++); return (NULL); } m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; memcpy(pc->pc_map, pc_freemask, sizeof(pc_freemask)); pc->pc_map[0] &= ~1ul; /* preallocated bit 0 */ pvc = &pv_chunks[vm_page_domain(m)]; mtx_lock(&pvc->pvc_lock); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru); mtx_unlock(&pvc->pvc_lock); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1)); return (pv); } /* * Ensure that the number of spare PV entries in the specified pmap meets or * exceeds the given count, "needed". * * The given PV list lock may be released. */ static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp) { struct pv_chunks_list *pvc; struct pch new_tail[PMAP_MEMDOM]; struct pv_chunk *pc; vm_page_t m; int avail, free, i; bool reclaimed; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL")); /* * Newly allocated PV chunks must be stored in a private list until * the required number of PV chunks have been allocated. Otherwise, * reclaim_pv_chunk() could recycle one of these chunks. In * contrast, these chunks must be added to the pmap upon allocation. */ for (i = 0; i < PMAP_MEMDOM; i++) TAILQ_INIT(&new_tail[i]); retry: avail = 0; TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) { bit_count((bitstr_t *)pc->pc_map, 0, sizeof(pc->pc_map) * NBBY, &free); if (free == 0) break; avail += free; if (avail >= needed) break; } for (reclaimed = false; avail < needed; avail += _NPCPV) { m = vm_page_alloc_noobj(VM_ALLOC_WIRED); if (m == NULL) { m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; reclaimed = true; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; memcpy(pc->pc_map, pc_freemask, sizeof(pc_freemask)); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail[vm_page_domain(m)], pc, pc_lru); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV)); /* * The reclaim might have freed a chunk from the current pmap. * If that chunk contained available entries, we need to * re-count the number of available entries. */ if (reclaimed) goto retry; } for (i = 0; i < vm_ndomains; i++) { if (TAILQ_EMPTY(&new_tail[i])) continue; pvc = &pv_chunks[i]; mtx_lock(&pvc->pvc_lock); TAILQ_CONCAT(&pvc->pvc_list, &new_tail[i], pc_lru); mtx_unlock(&pvc->pvc_lock); } } /* * First find and then remove the pv entry for the specified pmap and virtual * address from the specified pv list. Returns the pv entry if found and NULL * otherwise. This operation can be performed on pv lists for either 4KB or * 2MB page mappings. */ static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; break; } } return (pv); } /* * After demotion from a 2MB page mapping to 512 4KB page mappings, * destroy the pv entry for the 2MB page mapping and reinstantiate the pv * entries for each of the 4KB page mappings. */ static void pmap_pv_demote_l2(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; struct pv_chunk *pc; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; int bit, field; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((va & L2_OFFSET) == 0, ("pmap_pv_demote_l2: va is not 2mpage aligned")); KASSERT((pa & L2_OFFSET) == 0, ("pmap_pv_demote_l2: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the 2mpage's pv entry for this mapping to the first * page's pv list. Once this transfer begins, the pv list lock * must not be released until the last pv entry is reinstantiated. */ pvh = pa_to_pvh(pa); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_l2: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; /* Instantiate the remaining Ln_ENTRIES - 1 pv entries. */ PV_STAT(atomic_add_long(&pv_entry_allocs, Ln_ENTRIES - 1)); va_last = va + L2_SIZE - PAGE_SIZE; for (;;) { pc = TAILQ_FIRST(&pmap->pm_pvchunk); KASSERT(!pc_is_full(pc), ("pmap_pv_demote_l2: missing spare")); for (field = 0; field < _NPCM; field++) { while (pc->pc_map[field]) { bit = ffsl(pc->pc_map[field]) - 1; pc->pc_map[field] &= ~(1ul << bit); pv = &pc->pc_pventry[field * 64 + bit]; va += PAGE_SIZE; pv->pv_va = va; m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_l2: page %p is not managed", m)); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (va == va_last) goto out; } } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } out: if (pc_is_full(pc)) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, Ln_ENTRIES - 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, Ln_ENTRIES - 1)); } /* * First find and then destroy the pv entry for the specified pmap and virtual * address. This operation can be performed on pv lists for either 4KB or 2MB * page mappings. */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } /* * Conditionally create the PV entry for a 4KB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp) { pv_entry_t pv; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; return (TRUE); } else return (FALSE); } /* * Create the PV entry for a 2MB page mapping. Always returns true unless the * flag PMAP_ENTER_NORECLAIM is specified. If that flag is specified, returns * false if the PV entry cannot be allocated without resorting to reclamation. */ static bool pmap_pv_insert_l2(pmap_t pmap, vm_offset_t va, pd_entry_t l2e, u_int flags, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_paddr_t pa; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, (flags & PMAP_ENTER_NORECLAIM) != 0 ? NULL : lockp)) == NULL) return (false); pv->pv_va = va; pa = PTE_TO_PHYS(l2e); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; return (true); } static void pmap_remove_kernel_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t va) { pt_entry_t newl2, oldl2 __diagused; vm_page_t ml3; vm_paddr_t ml3pa; KASSERT(!VIRT_IN_DMAP(va), ("removing direct mapping of %#lx", va)); KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); ml3 = pmap_remove_pt_page(pmap, va); if (ml3 == NULL) panic("pmap_remove_kernel_l2: Missing pt page"); ml3pa = VM_PAGE_TO_PHYS(ml3); newl2 = PHYS_TO_PTE(ml3pa) | L2_TABLE; /* * If this page table page was unmapped by a promotion, then it * contains valid mappings. Zero it to invalidate those mappings. */ if (vm_page_any_valid(ml3)) pagezero((void *)PHYS_TO_DMAP(ml3pa)); /* * Demote the mapping. The caller must have already invalidated the * mapping (i.e., the "break" in break-before-make). */ oldl2 = pmap_load_store(l2, newl2); KASSERT(oldl2 == 0, ("%s: found existing mapping at %p: %#lx", __func__, l2, oldl2)); } /* * pmap_remove_l2: Do the things to unmap a level 2 superpage. */ static int pmap_remove_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t sva, pd_entry_t l1e, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pt_entry_t old_l2; vm_page_t m, ml3, mt; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & L2_OFFSET) == 0, ("pmap_remove_l2: sva is not aligned")); old_l2 = pmap_load_clear(l2); KASSERT((old_l2 & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_remove_l2: L2e %lx is not a block mapping", old_l2)); /* * Since a promotion must break the 4KB page mappings before making * the 2MB page mapping, a pmap_s1_invalidate_page() suffices. */ pmap_s1_invalidate_page(pmap, sva, true); if (old_l2 & ATTR_SW_WIRED) pmap->pm_stats.wired_count -= L2_SIZE / PAGE_SIZE; pmap_resident_count_dec(pmap, L2_SIZE / PAGE_SIZE); if (old_l2 & ATTR_SW_MANAGED) { m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(old_l2)); pvh = page_to_pvh(m); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(pvh, pmap, sva); for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) { if (pmap_pte_dirty(pmap, old_l2)) vm_page_dirty(mt); if (old_l2 & ATTR_AF) vm_page_aflag_set(mt, PGA_REFERENCED); if (TAILQ_EMPTY(&mt->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } } if (pmap == kernel_pmap) { pmap_remove_kernel_l2(pmap, l2, sva); } else { ml3 = pmap_remove_pt_page(pmap, sva); if (ml3 != NULL) { KASSERT(vm_page_any_valid(ml3), ("pmap_remove_l2: l3 page not promoted")); pmap_resident_count_dec(pmap, 1); KASSERT(ml3->ref_count == NL3PG, ("pmap_remove_l2: l3 page ref count error")); ml3->ref_count = 0; pmap_add_delayed_free_list(ml3, free, FALSE); } } return (pmap_unuse_pt(pmap, sva, l1e, free)); } /* * pmap_remove_l3: do the things to unmap a page in a process */ static int pmap_remove_l3(pmap_t pmap, pt_entry_t *l3, vm_offset_t va, pd_entry_t l2e, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pt_entry_t old_l3; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); old_l3 = pmap_load_clear(l3); pmap_s1_invalidate_page(pmap, va, true); if (old_l3 & ATTR_SW_WIRED) pmap->pm_stats.wired_count -= 1; pmap_resident_count_dec(pmap, 1); if (old_l3 & ATTR_SW_MANAGED) { m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(old_l3)); if (pmap_pte_dirty(pmap, old_l3)) vm_page_dirty(m); if (old_l3 & ATTR_AF) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } return (pmap_unuse_pt(pmap, va, l2e, free)); } /* * Remove the specified range of addresses from the L3 page table that is * identified by the given L2 entry. */ static void pmap_remove_l3_range(pmap_t pmap, pd_entry_t l2e, vm_offset_t sva, vm_offset_t eva, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; struct rwlock *new_lock; pt_entry_t *l3, old_l3; vm_offset_t va; vm_page_t l3pg, m; KASSERT(ADDR_IS_CANONICAL(sva), ("%s: Start address not in canonical form: %lx", __func__, sva)); KASSERT(ADDR_IS_CANONICAL(eva) || eva == VM_MAX_USER_ADDRESS, ("%s: End address not in canonical form: %lx", __func__, eva)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(rounddown2(sva, L2_SIZE) + L2_SIZE == roundup2(eva, L2_SIZE), ("pmap_remove_l3_range: range crosses an L3 page table boundary")); l3pg = !ADDR_IS_KERNEL(sva) ? PHYS_TO_VM_PAGE(PTE_TO_PHYS(l2e)) : NULL; va = eva; for (l3 = pmap_l2_to_l3(&l2e, sva); sva != eva; l3++, sva += L3_SIZE) { if (!pmap_l3_valid(pmap_load(l3))) { if (va != eva) { pmap_invalidate_range(pmap, va, sva, true); va = eva; } continue; } old_l3 = pmap_load_clear(l3); if ((old_l3 & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count--; pmap_resident_count_dec(pmap, 1); if ((old_l3 & ATTR_SW_MANAGED) != 0) { m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(old_l3)); if (pmap_pte_dirty(pmap, old_l3)) vm_page_dirty(m); if ((old_l3 & ATTR_AF) != 0) vm_page_aflag_set(m, PGA_REFERENCED); new_lock = VM_PAGE_TO_PV_LIST_LOCK(m); if (new_lock != *lockp) { if (*lockp != NULL) { /* * Pending TLB invalidations must be * performed before the PV list lock is * released. Otherwise, a concurrent * pmap_remove_all() on a physical page * could return while a stale TLB entry * still provides access to that page. */ if (va != eva) { pmap_invalidate_range(pmap, va, sva, true); va = eva; } rw_wunlock(*lockp); } *lockp = new_lock; rw_wlock(*lockp); } pmap_pvh_free(&m->md, pmap, sva); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } if (l3pg != NULL && pmap_unwire_l3(pmap, sva, l3pg, free)) { /* * _pmap_unwire_l3() has already invalidated the TLB * entries at all levels for "sva". So, we need not * perform "sva += L3_SIZE;" here. Moreover, we need * not perform "va = sva;" if "sva" is at the start * of a new valid range consisting of a single page. */ break; } if (va == eva) va = sva; } if (va != eva) pmap_invalidate_range(pmap, va, sva, true); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct rwlock *lock; vm_offset_t va_next; pd_entry_t *l0, *l1, *l2; pt_entry_t l3_paddr; struct spglist free; /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; SLIST_INIT(&free); PMAP_LOCK(pmap); lock = NULL; for (; sva < eva; sva = va_next) { if (pmap->pm_stats.resident_count == 0) break; l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; l1 = pmap_l0_to_l1(l0, sva); if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", pmap_load(l1), sva, eva, va_next)); MPASS(pmap != kernel_pmap); MPASS((pmap_load(l1) & ATTR_SW_MANAGED) == 0); pmap_clear(l1); pmap_s1_invalidate_page(pmap, sva, true); pmap_resident_count_dec(pmap, L1_SIZE / PAGE_SIZE); pmap_unuse_pt(pmap, sva, pmap_load(l0), &free); continue; } /* * Calculate index for next page table. */ va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (l2 == NULL) continue; l3_paddr = pmap_load(l2); if ((l3_paddr & ATTR_DESCR_MASK) == L2_BLOCK) { if (sva + L2_SIZE == va_next && eva >= va_next) { pmap_remove_l2(pmap, l2, sva, pmap_load(l1), &free, &lock); continue; } else if (pmap_demote_l2_locked(pmap, l2, sva, &lock) == NULL) continue; l3_paddr = pmap_load(l2); } /* * Weed out invalid mappings. */ if ((l3_paddr & ATTR_DESCR_MASK) != L2_TABLE) continue; /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (va_next > eva) va_next = eva; pmap_remove_l3_range(pmap, l3_paddr, sva, va_next, &free, &lock); } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); vm_page_free_pages_toq(&free, true); } /* * Remove the given range of addresses as part of a logical unmap * operation. This has the effect of calling pmap_remove(), but * also clears any metadata that should persist for the lifetime * of a logical mapping. */ void pmap_map_delete(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { pmap_remove(pmap, sva, eva); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; struct rwlock *lock; pd_entry_t *pde, tpde; pt_entry_t *pte, tpte; vm_offset_t va; struct spglist free; int lvl, pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : page_to_pvh(m); rw_wlock(lock); retry: while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pte = pmap_pte_exists(pmap, va, 2, __func__); pmap_demote_l2_locked(pmap, pte, va, &lock); PMAP_UNLOCK(pmap); } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } pmap_resident_count_dec(pmap, 1); pde = pmap_pde(pmap, pv->pv_va, &lvl); KASSERT(pde != NULL, ("pmap_remove_all: no page directory entry found")); KASSERT(lvl == 2, ("pmap_remove_all: invalid pde level %d", lvl)); tpde = pmap_load(pde); pte = pmap_l2_to_l3(pde, pv->pv_va); tpte = pmap_load_clear(pte); if (tpte & ATTR_SW_WIRED) pmap->pm_stats.wired_count--; if ((tpte & ATTR_AF) != 0) { pmap_invalidate_page(pmap, pv->pv_va, true); vm_page_aflag_set(m, PGA_REFERENCED); } /* * Update the vm_page_t clean and reference bits. */ if (pmap_pte_dirty(pmap, tpte)) vm_page_dirty(m); pmap_unuse_pt(pmap, pv->pv_va, tpde, &free); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(lock); vm_page_free_pages_toq(&free, true); } /* * Masks and sets bits in a level 2 page table entries in the specified pmap */ static void pmap_protect_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t sva, pt_entry_t mask, pt_entry_t nbits) { pd_entry_t old_l2; vm_page_t m, mt; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PMAP_ASSERT_STAGE1(pmap); KASSERT((sva & L2_OFFSET) == 0, ("pmap_protect_l2: sva is not 2mpage aligned")); old_l2 = pmap_load(l2); KASSERT((old_l2 & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_protect_l2: L2e %lx is not a block mapping", old_l2)); /* * Return if the L2 entry already has the desired access restrictions * in place. */ if ((old_l2 & mask) == nbits) return; while (!atomic_fcmpset_64(l2, &old_l2, (old_l2 & ~mask) | nbits)) cpu_spinwait(); /* * When a dirty read/write superpage mapping is write protected, * update the dirty field of each of the superpage's constituent 4KB * pages. */ if ((old_l2 & ATTR_SW_MANAGED) != 0 && (nbits & ATTR_S1_AP(ATTR_S1_AP_RO)) != 0 && pmap_pte_dirty(pmap, old_l2)) { m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(old_l2)); for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) vm_page_dirty(mt); } /* * Since a promotion must break the 4KB page mappings before making * the 2MB page mapping, a pmap_s1_invalidate_page() suffices. */ pmap_s1_invalidate_page(pmap, sva, true); } /* * Masks and sets bits in last level page table entries in the specified * pmap and range */ static void pmap_mask_set_locked(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pt_entry_t mask, pt_entry_t nbits, bool invalidate) { vm_offset_t va, va_next; pd_entry_t *l0, *l1, *l2; pt_entry_t *l3p, l3; PMAP_LOCK_ASSERT(pmap, MA_OWNED); for (; sva < eva; sva = va_next) { l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; l1 = pmap_l0_to_l1(l0, sva); if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", pmap_load(l1), sva, eva, va_next)); MPASS((pmap_load(l1) & ATTR_SW_MANAGED) == 0); if ((pmap_load(l1) & mask) != nbits) { pmap_store(l1, (pmap_load(l1) & ~mask) | nbits); if (invalidate) pmap_s1_invalidate_page(pmap, sva, true); } continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (pmap_load(l2) == 0) continue; if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) { if (sva + L2_SIZE == va_next && eva >= va_next) { pmap_protect_l2(pmap, l2, sva, mask, nbits); continue; } else if (pmap_demote_l2(pmap, l2, sva) == NULL) continue; } KASSERT((pmap_load(l2) & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_protect: Invalid L2 entry after demotion")); if (va_next > eva) va_next = eva; va = va_next; for (l3p = pmap_l2_to_l3(l2, sva); sva != va_next; l3p++, sva += L3_SIZE) { l3 = pmap_load(l3p); /* * Go to the next L3 entry if the current one is * invalid or already has the desired access * restrictions in place. (The latter case occurs * frequently. For example, in a "buildworld" * workload, almost 1 out of 4 L3 entries already * have the desired restrictions.) */ if (!pmap_l3_valid(l3) || (l3 & mask) == nbits) { if (va != va_next) { if (invalidate) pmap_s1_invalidate_range(pmap, va, sva, true); va = va_next; } continue; } while (!atomic_fcmpset_64(l3p, &l3, (l3 & ~mask) | nbits)) cpu_spinwait(); /* * When a dirty read/write mapping is write protected, * update the page's dirty field. */ if ((l3 & ATTR_SW_MANAGED) != 0 && (nbits & ATTR_S1_AP(ATTR_S1_AP_RO)) != 0 && pmap_pte_dirty(pmap, l3)) vm_page_dirty(PHYS_TO_VM_PAGE(PTE_TO_PHYS(l3))); if (va == va_next) va = sva; } if (va != va_next && invalidate) pmap_s1_invalidate_range(pmap, va, sva, true); } } static void pmap_mask_set(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pt_entry_t mask, pt_entry_t nbits, bool invalidate) { PMAP_LOCK(pmap); pmap_mask_set_locked(pmap, sva, eva, mask, nbits, invalidate); PMAP_UNLOCK(pmap); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { pt_entry_t mask, nbits; PMAP_ASSERT_STAGE1(pmap); KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } mask = nbits = 0; if ((prot & VM_PROT_WRITE) == 0) { mask |= ATTR_S1_AP_RW_BIT | ATTR_SW_DBM; nbits |= ATTR_S1_AP(ATTR_S1_AP_RO); } if ((prot & VM_PROT_EXECUTE) == 0) { mask |= ATTR_S1_XN; nbits |= ATTR_S1_XN; } if (mask == 0) return; pmap_mask_set(pmap, sva, eva, mask, nbits, true); } void pmap_disable_promotion(vm_offset_t sva, vm_size_t size) { MPASS((sva & L3_OFFSET) == 0); MPASS(((sva + size) & L3_OFFSET) == 0); pmap_mask_set(kernel_pmap, sva, sva + size, ATTR_SW_NO_PROMOTE, ATTR_SW_NO_PROMOTE, false); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. * * If "promoted" is false, then the page table page "mpte" must be zero filled; * "mpte"'s valid field will be set to 0. * * If "promoted" is true and "all_l3e_AF_set" is false, then "mpte" must * contain valid mappings with identical attributes except for ATTR_AF; * "mpte"'s valid field will be set to 1. * * If "promoted" and "all_l3e_AF_set" are both true, then "mpte" must contain * valid mappings with identical attributes including ATTR_AF; "mpte"'s valid * field will be set to VM_PAGE_BITS_ALL. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted, bool all_l3e_AF_set) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(promoted || !all_l3e_AF_set, ("a zero-filled PTP can't have ATTR_AF set in every PTE")); mpte->valid = promoted ? (all_l3e_AF_set ? VM_PAGE_BITS_ALL : 1) : 0; return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Removes the page table page mapping the specified virtual address from the * specified pmap's collection of idle page table pages, and returns it. * Otherwise, returns NULL if there is no page table page corresponding to the * specified virtual address. */ static __inline vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_remove(&pmap->pm_root, pmap_l2_pindex(va))); } /* * Performs a break-before-make update of a pmap entry. This is needed when * either promoting or demoting pages to ensure the TLB doesn't get into an * inconsistent state. */ static void pmap_update_entry(pmap_t pmap, pd_entry_t *pte, pd_entry_t newpte, vm_offset_t va, vm_size_t size) { register_t intr; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((newpte & ATTR_SW_NO_PROMOTE) != 0) panic("%s: Updating non-promote pte", __func__); /* * Ensure we don't get switched out with the page table in an * inconsistent state. We also need to ensure no interrupts fire * as they may make use of an address we are about to invalidate. */ intr = intr_disable(); /* * Clear the old mapping's valid bit, but leave the rest of the entry * unchanged, so that a lockless, concurrent pmap_kextract() can still * lookup the physical address. */ pmap_clear_bits(pte, ATTR_DESCR_VALID); /* * When promoting, the L{1,2}_TABLE entry that is being replaced might * be cached, so we invalidate intermediate entries as well as final * entries. */ pmap_s1_invalidate_range(pmap, va, va + size, false); /* Create the new mapping */ pmap_store(pte, newpte); dsb(ishst); intr_restore(intr); } #if VM_NRESERVLEVEL > 0 /* * After promotion from 512 4KB page mappings to a single 2MB page mapping, * replace the many pv entries for the 4KB page mappings by a single pv entry * for the 2MB page mapping. */ static void pmap_pv_promote_l2(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; KASSERT((pa & L2_OFFSET) == 0, ("pmap_pv_promote_l2: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the first page's pv entry for this mapping to the 2mpage's * pv list. Aside from avoiding the cost of a call to get_pv_entry(), * a transfer avoids the possibility that get_pv_entry() calls * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the * mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = va & ~L2_OFFSET; pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_l2: pv not found")); pvh = page_to_pvh(m); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + L2_SIZE - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } /* * Tries to promote the 512, contiguous 4KB page mappings that are within a * single level 2 table entry to a single 2MB page mapping. For promotion * to occur, two conditions must be met: (1) the 4KB page mappings must map * aligned, contiguous physical memory and (2) the 4KB page mappings must have * identical characteristics. */ static void pmap_promote_l2(pmap_t pmap, pd_entry_t *l2, vm_offset_t va, vm_page_t mpte, struct rwlock **lockp) { pt_entry_t all_l3e_AF, *firstl3, *l3, newl2, oldl3, pa; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PMAP_ASSERT_STAGE1(pmap); /* * Examine the first L3E in the specified PTP. Abort if this L3E is * ineligible for promotion... */ firstl3 = (pt_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(pmap_load(l2))); newl2 = pmap_load(firstl3); if ((newl2 & ATTR_SW_NO_PROMOTE) != 0) return; /* ... is not the first physical page within an L2 block */ if ((PTE_TO_PHYS(newl2) & L2_OFFSET) != 0 || ((newl2 & ATTR_DESCR_MASK) != L3_PAGE)) { /* ... or is invalid */ atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx" " in pmap %p", va, pmap); return; } /* * Both here and in the below "for" loop, to allow for repromotion * after MADV_FREE, conditionally write protect a clean L3E before * possibly aborting the promotion due to other L3E attributes. Why? * Suppose that MADV_FREE is applied to a part of a superpage, the * address range [S, E). pmap_advise() will demote the superpage * mapping, destroy the 4KB page mapping at the end of [S, E), and * set AP_RO and clear AF in the L3Es for the rest of [S, E). Later, * imagine that the memory in [S, E) is recycled, but the last 4KB * page in [S, E) is not the last to be rewritten, or simply accessed. * In other words, there is still a 4KB page in [S, E), call it P, * that is writeable but AP_RO is set and AF is clear in P's L3E. * Unless we write protect P before aborting the promotion, if and * when P is finally rewritten, there won't be a page fault to trigger * repromotion. */ setl2: if ((newl2 & (ATTR_S1_AP_RW_BIT | ATTR_SW_DBM)) == (ATTR_S1_AP(ATTR_S1_AP_RO) | ATTR_SW_DBM)) { /* * When the mapping is clean, i.e., ATTR_S1_AP_RO is set, * ATTR_SW_DBM can be cleared without a TLB invalidation. */ if (!atomic_fcmpset_64(firstl3, &newl2, newl2 & ~ATTR_SW_DBM)) goto setl2; newl2 &= ~ATTR_SW_DBM; CTR2(KTR_PMAP, "pmap_promote_l2: protect for va %#lx" " in pmap %p", va & ~L2_OFFSET, pmap); } /* * Examine each of the other L3Es in the specified PTP. Abort if this * L3E maps an unexpected 4KB physical page or does not have identical * characteristics to the first L3E. If ATTR_AF is not set in every * PTE, then request that the PTP be refilled on demotion. */ all_l3e_AF = newl2 & ATTR_AF; pa = (PTE_TO_PHYS(newl2) | (newl2 & ATTR_DESCR_MASK)) + L2_SIZE - PAGE_SIZE; for (l3 = firstl3 + NL3PG - 1; l3 > firstl3; l3--) { oldl3 = pmap_load(l3); if ((PTE_TO_PHYS(oldl3) | (oldl3 & ATTR_DESCR_MASK)) != pa) { atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx" " in pmap %p", va, pmap); return; } setl3: if ((oldl3 & (ATTR_S1_AP_RW_BIT | ATTR_SW_DBM)) == (ATTR_S1_AP(ATTR_S1_AP_RO) | ATTR_SW_DBM)) { /* * When the mapping is clean, i.e., ATTR_S1_AP_RO is * set, ATTR_SW_DBM can be cleared without a TLB * invalidation. */ if (!atomic_fcmpset_64(l3, &oldl3, oldl3 & ~ATTR_SW_DBM)) goto setl3; oldl3 &= ~ATTR_SW_DBM; } if ((oldl3 & (ATTR_MASK & ~ATTR_AF)) != (newl2 & (ATTR_MASK & ~ATTR_AF))) { atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx" " in pmap %p", va, pmap); return; } all_l3e_AF &= oldl3; pa -= PAGE_SIZE; } /* * Unless all PTEs have ATTR_AF set, clear it from the superpage * mapping, so that promotions triggered by speculative mappings, * such as pmap_enter_quick(), don't automatically mark the * underlying pages as referenced. */ newl2 &= ~ATTR_AF | all_l3e_AF; /* * Save the page table page in its current state until the L2 * mapping the superpage is demoted by pmap_demote_l2() or * destroyed by pmap_remove_l3(). */ if (mpte == NULL) mpte = PHYS_TO_VM_PAGE(PTE_TO_PHYS(pmap_load(l2))); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_l2: page table page is out of range")); KASSERT(mpte->pindex == pmap_l2_pindex(va), ("pmap_promote_l2: page table page's pindex is wrong")); if (pmap_insert_pt_page(pmap, mpte, true, all_l3e_AF != 0)) { atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx in pmap %p", va, pmap); return; } if ((newl2 & ATTR_SW_MANAGED) != 0) pmap_pv_promote_l2(pmap, va, PTE_TO_PHYS(newl2), lockp); newl2 &= ~ATTR_DESCR_MASK; newl2 |= L2_BLOCK; pmap_update_entry(pmap, l2, newl2, va & ~L2_OFFSET, L2_SIZE); atomic_add_long(&pmap_l2_promotions, 1); CTR2(KTR_PMAP, "pmap_promote_l2: success for va %#lx in pmap %p", va, pmap); } #endif /* VM_NRESERVLEVEL > 0 */ static int pmap_enter_largepage(pmap_t pmap, vm_offset_t va, pt_entry_t newpte, int flags, int psind) { pd_entry_t *l0p, *l1p, *l2p, origpte; vm_page_t mp; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(psind > 0 && psind < MAXPAGESIZES, ("psind %d unexpected", psind)); KASSERT((PTE_TO_PHYS(newpte) & (pagesizes[psind] - 1)) == 0, ("unaligned phys address %#lx newpte %#lx psind %d", PTE_TO_PHYS(newpte), newpte, psind)); restart: if (psind == 2) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; l0p = pmap_l0(pmap, va); if ((pmap_load(l0p) & ATTR_DESCR_VALID) == 0) { mp = _pmap_alloc_l3(pmap, pmap_l0_pindex(va), NULL); if (mp == NULL) { if ((flags & PMAP_ENTER_NOSLEEP) != 0) return (KERN_RESOURCE_SHORTAGE); PMAP_UNLOCK(pmap); vm_wait(NULL); PMAP_LOCK(pmap); goto restart; } l1p = pmap_l0_to_l1(l0p, va); KASSERT(l1p != NULL, ("va %#lx lost l1 entry", va)); origpte = pmap_load(l1p); } else { l1p = pmap_l0_to_l1(l0p, va); KASSERT(l1p != NULL, ("va %#lx lost l1 entry", va)); origpte = pmap_load(l1p); if ((origpte & ATTR_DESCR_VALID) == 0) { mp = PHYS_TO_VM_PAGE( PTE_TO_PHYS(pmap_load(l0p))); mp->ref_count++; } } KASSERT((PTE_TO_PHYS(origpte) == PTE_TO_PHYS(newpte) && (origpte & ATTR_DESCR_MASK) == L1_BLOCK) || (origpte & ATTR_DESCR_VALID) == 0, ("va %#lx changing 1G phys page l1 %#lx newpte %#lx", va, origpte, newpte)); pmap_store(l1p, newpte); } else /* (psind == 1) */ { l2p = pmap_l2(pmap, va); if (l2p == NULL) { mp = _pmap_alloc_l3(pmap, pmap_l1_pindex(va), NULL); if (mp == NULL) { if ((flags & PMAP_ENTER_NOSLEEP) != 0) return (KERN_RESOURCE_SHORTAGE); PMAP_UNLOCK(pmap); vm_wait(NULL); PMAP_LOCK(pmap); goto restart; } l2p = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mp)); l2p = &l2p[pmap_l2_index(va)]; origpte = pmap_load(l2p); } else { l1p = pmap_l1(pmap, va); origpte = pmap_load(l2p); if ((origpte & ATTR_DESCR_VALID) == 0) { mp = PHYS_TO_VM_PAGE( PTE_TO_PHYS(pmap_load(l1p))); mp->ref_count++; } } KASSERT((origpte & ATTR_DESCR_VALID) == 0 || ((origpte & ATTR_DESCR_MASK) == L2_BLOCK && PTE_TO_PHYS(origpte) == PTE_TO_PHYS(newpte)), ("va %#lx changing 2M phys page l2 %#lx newpte %#lx", va, origpte, newpte)); pmap_store(l2p, newpte); } dsb(ishst); if ((origpte & ATTR_DESCR_VALID) == 0) pmap_resident_count_inc(pmap, pagesizes[psind] / PAGE_SIZE); if ((newpte & ATTR_SW_WIRED) != 0 && (origpte & ATTR_SW_WIRED) == 0) pmap->pm_stats.wired_count += pagesizes[psind] / PAGE_SIZE; else if ((newpte & ATTR_SW_WIRED) == 0 && (origpte & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count -= pagesizes[psind] / PAGE_SIZE; return (KERN_SUCCESS); } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { struct rwlock *lock; pd_entry_t *pde; pt_entry_t new_l3, orig_l3; pt_entry_t *l2, *l3; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte, om; boolean_t nosleep; int lvl, rv; KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); va = trunc_page(va); if ((m->oflags & VPO_UNMANAGED) == 0) VM_PAGE_OBJECT_BUSY_ASSERT(m); pa = VM_PAGE_TO_PHYS(m); new_l3 = (pt_entry_t)(PHYS_TO_PTE(pa) | ATTR_DEFAULT | L3_PAGE); new_l3 |= pmap_pte_memattr(pmap, m->md.pv_memattr); new_l3 |= pmap_pte_prot(pmap, prot); if ((flags & PMAP_ENTER_WIRED) != 0) new_l3 |= ATTR_SW_WIRED; if (pmap->pm_stage == PM_STAGE1) { if (!ADDR_IS_KERNEL(va)) new_l3 |= ATTR_S1_AP(ATTR_S1_AP_USER) | ATTR_S1_PXN; else new_l3 |= ATTR_S1_UXN; if (pmap != kernel_pmap) new_l3 |= ATTR_S1_nG; } else { /* * Clear the access flag on executable mappings, this will be * set later when the page is accessed. The fault handler is * required to invalidate the I-cache. * * TODO: Switch to the valid flag to allow hardware management * of the access flag. Much of the pmap code assumes the * valid flag is set and fails to destroy the old page tables * correctly if it is clear. */ if (prot & VM_PROT_EXECUTE) new_l3 &= ~ATTR_AF; } if ((m->oflags & VPO_UNMANAGED) == 0) { new_l3 |= ATTR_SW_MANAGED; if ((prot & VM_PROT_WRITE) != 0) { new_l3 |= ATTR_SW_DBM; if ((flags & VM_PROT_WRITE) == 0) { if (pmap->pm_stage == PM_STAGE1) new_l3 |= ATTR_S1_AP(ATTR_S1_AP_RO); else new_l3 &= ~ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE); } } } CTR2(KTR_PMAP, "pmap_enter: %.16lx -> %.16lx", va, pa); lock = NULL; PMAP_LOCK(pmap); if ((flags & PMAP_ENTER_LARGEPAGE) != 0) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed largepage va %#lx flags %#x", va, flags)); new_l3 &= ~L3_PAGE; if (psind == 2) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; new_l3 |= L1_BLOCK; } else /* (psind == 1) */ new_l3 |= L2_BLOCK; rv = pmap_enter_largepage(pmap, va, new_l3, flags, psind); goto out; } if (psind == 1) { /* Assert the required virtual and physical alignment. */ KASSERT((va & L2_OFFSET) == 0, ("pmap_enter: va unaligned")); KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind")); rv = pmap_enter_l2(pmap, va, (new_l3 & ~L3_PAGE) | L2_BLOCK, flags, m, &lock); goto out; } mpte = NULL; /* * In the case that a page table page is not * resident, we are creating it here. */ retry: pde = pmap_pde(pmap, va, &lvl); if (pde != NULL && lvl == 2) { l3 = pmap_l2_to_l3(pde, va); if (!ADDR_IS_KERNEL(va) && mpte == NULL) { mpte = PHYS_TO_VM_PAGE(PTE_TO_PHYS(pmap_load(pde))); mpte->ref_count++; } goto havel3; } else if (pde != NULL && lvl == 1) { l2 = pmap_l1_to_l2(pde, va); if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK && (l3 = pmap_demote_l2_locked(pmap, l2, va, &lock)) != NULL) { l3 = &l3[pmap_l3_index(va)]; if (!ADDR_IS_KERNEL(va)) { mpte = PHYS_TO_VM_PAGE( PTE_TO_PHYS(pmap_load(l2))); mpte->ref_count++; } goto havel3; } /* We need to allocate an L3 table. */ } if (!ADDR_IS_KERNEL(va)) { nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0; /* * We use _pmap_alloc_l3() instead of pmap_alloc_l3() in order * to handle the possibility that a superpage mapping for "va" * was created while we slept. */ mpte = _pmap_alloc_l3(pmap, pmap_l2_pindex(va), nosleep ? NULL : &lock); if (mpte == NULL && nosleep) { CTR0(KTR_PMAP, "pmap_enter: mpte == NULL"); rv = KERN_RESOURCE_SHORTAGE; goto out; } goto retry; } else panic("pmap_enter: missing L3 table for kernel va %#lx", va); havel3: orig_l3 = pmap_load(l3); opa = PTE_TO_PHYS(orig_l3); pv = NULL; /* * Is the specified virtual address already mapped? */ if (pmap_l3_valid(orig_l3)) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((flags & PMAP_ENTER_WIRED) != 0 && (orig_l3 & ATTR_SW_WIRED) == 0) pmap->pm_stats.wired_count++; else if ((flags & PMAP_ENTER_WIRED) == 0 && (orig_l3 & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->ref_count--; KASSERT(mpte->ref_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%lx", va)); } /* * Has the physical page changed? */ if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((orig_l3 & ATTR_SW_MANAGED) != 0 && (new_l3 & ATTR_SW_DBM) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); goto validate; } /* * The physical page has changed. Temporarily invalidate * the mapping. */ orig_l3 = pmap_load_clear(l3); KASSERT(PTE_TO_PHYS(orig_l3) == opa, ("pmap_enter: unexpected pa update for %#lx", va)); if ((orig_l3 & ATTR_SW_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); /* * The pmap lock is sufficient to synchronize with * concurrent calls to pmap_page_test_mappings() and * pmap_ts_referenced(). */ if (pmap_pte_dirty(pmap, orig_l3)) vm_page_dirty(om); if ((orig_l3 & ATTR_AF) != 0) { pmap_invalidate_page(pmap, va, true); vm_page_aflag_set(om, PGA_REFERENCED); } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, om); pv = pmap_pvh_remove(&om->md, pmap, va); if ((m->oflags & VPO_UNMANAGED) != 0) free_pv_entry(pmap, pv); if ((om->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&page_to_pvh(om)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } else { KASSERT((orig_l3 & ATTR_AF) != 0, ("pmap_enter: unmanaged mapping lacks ATTR_AF")); pmap_invalidate_page(pmap, va, true); } orig_l3 = 0; } else { /* * Increment the counters. */ if ((new_l3 & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count++; pmap_resident_count_inc(pmap, 1); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { if (pv == NULL) { pv = get_pv_entry(pmap, &lock); pv->pv_va = va; } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((new_l3 & ATTR_SW_DBM) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } validate: if (pmap->pm_stage == PM_STAGE1) { /* * Sync icache if exec permission and attribute * VM_MEMATTR_WRITE_BACK is set. Do it now, before the mapping * is stored and made valid for hardware table walk. If done * later, then other can access this page before caches are * properly synced. Don't do it for kernel memory which is * mapped with exec permission even if the memory isn't going * to hold executable code. The only time when icache sync is * needed is after kernel module is loaded and the relocation * info is processed. And it's done in elf_cpu_load_file(). */ if ((prot & VM_PROT_EXECUTE) && pmap != kernel_pmap && m->md.pv_memattr == VM_MEMATTR_WRITE_BACK && (opa != pa || (orig_l3 & ATTR_S1_XN))) { PMAP_ASSERT_STAGE1(pmap); cpu_icache_sync_range(PHYS_TO_DMAP(pa), PAGE_SIZE); } } else { cpu_dcache_wb_range(PHYS_TO_DMAP(pa), PAGE_SIZE); } /* * Update the L3 entry */ if (pmap_l3_valid(orig_l3)) { KASSERT(opa == pa, ("pmap_enter: invalid update")); if ((orig_l3 & ~ATTR_AF) != (new_l3 & ~ATTR_AF)) { /* same PA, different attributes */ orig_l3 = pmap_load_store(l3, new_l3); pmap_invalidate_page(pmap, va, true); if ((orig_l3 & ATTR_SW_MANAGED) != 0 && pmap_pte_dirty(pmap, orig_l3)) vm_page_dirty(m); } else { /* * orig_l3 == new_l3 * This can happens if multiple threads simultaneously * access not yet mapped page. This bad for performance * since this can cause full demotion-NOP-promotion * cycle. * Another possible reasons are: * - VM and pmap memory layout are diverged * - tlb flush is missing somewhere and CPU doesn't see * actual mapping. */ CTR4(KTR_PMAP, "%s: already mapped page - " "pmap %p va 0x%#lx pte 0x%lx", __func__, pmap, va, new_l3); } } else { /* New mapping */ pmap_store(l3, new_l3); dsb(ishst); } #if VM_NRESERVLEVEL > 0 /* * Try to promote from level 3 pages to a level 2 superpage. This * currently only works on stage 1 pmaps as pmap_promote_l2 looks at * stage 1 specific fields and performs a break-before-make sequence * that is incorrect a stage 2 pmap. */ if ((mpte == NULL || mpte->ref_count == NL3PG) && pmap_ps_enabled(pmap) && pmap->pm_stage == PM_STAGE1 && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) { pmap_promote_l2(pmap, pde, va, mpte, &lock); } #endif rv = KERN_SUCCESS; out: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (rv); } /* * Tries to create a read- and/or execute-only 2MB page mapping. Returns * KERN_SUCCESS if the mapping was created. Otherwise, returns an error * value. See pmap_enter_l2() for the possible error values when "no sleep", * "no replace", and "no reclaim" are specified. */ static int pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp) { pd_entry_t new_l2; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PMAP_ASSERT_STAGE1(pmap); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); new_l2 = (pd_entry_t)(PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | ATTR_DEFAULT | ATTR_S1_IDX(m->md.pv_memattr) | ATTR_S1_AP(ATTR_S1_AP_RO) | L2_BLOCK); if ((m->oflags & VPO_UNMANAGED) == 0) { new_l2 |= ATTR_SW_MANAGED; new_l2 &= ~ATTR_AF; } if ((prot & VM_PROT_EXECUTE) == 0 || m->md.pv_memattr == VM_MEMATTR_DEVICE) new_l2 |= ATTR_S1_XN; if (!ADDR_IS_KERNEL(va)) new_l2 |= ATTR_S1_AP(ATTR_S1_AP_USER) | ATTR_S1_PXN; else new_l2 |= ATTR_S1_UXN; if (pmap != kernel_pmap) new_l2 |= ATTR_S1_nG; return (pmap_enter_l2(pmap, va, new_l2, PMAP_ENTER_NOSLEEP | PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, m, lockp)); } /* * Returns true if every page table entry in the specified page table is * zero. */ static bool pmap_every_pte_zero(vm_paddr_t pa) { pt_entry_t *pt_end, *pte; KASSERT((pa & PAGE_MASK) == 0, ("pa is misaligned")); pte = (pt_entry_t *)PHYS_TO_DMAP(pa); for (pt_end = pte + Ln_ENTRIES; pte < pt_end; pte++) { if (*pte != 0) return (false); } return (true); } /* * Tries to create the specified 2MB page mapping. Returns KERN_SUCCESS if * the mapping was created, and one of KERN_FAILURE, KERN_NO_SPACE, or * KERN_RESOURCE_SHORTAGE otherwise. Returns KERN_FAILURE if * PMAP_ENTER_NOREPLACE was specified and a 4KB page mapping already exists * within the 2MB virtual address range starting at the specified virtual * address. Returns KERN_NO_SPACE if PMAP_ENTER_NOREPLACE was specified and a * 2MB page mapping already exists at the specified virtual address. Returns * KERN_RESOURCE_SHORTAGE if either (1) PMAP_ENTER_NOSLEEP was specified and a * page table page allocation failed or (2) PMAP_ENTER_NORECLAIM was specified * and a PV entry allocation failed. */ static int pmap_enter_l2(pmap_t pmap, vm_offset_t va, pd_entry_t new_l2, u_int flags, vm_page_t m, struct rwlock **lockp) { struct spglist free; pd_entry_t *l2, old_l2; vm_page_t l2pg, mt; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); if ((l2 = pmap_alloc_l2(pmap, va, &l2pg, (flags & PMAP_ENTER_NOSLEEP) != 0 ? NULL : lockp)) == NULL) { CTR2(KTR_PMAP, "pmap_enter_l2: failure for va %#lx in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } /* * If there are existing mappings, either abort or remove them. */ if ((old_l2 = pmap_load(l2)) != 0) { KASSERT(l2pg == NULL || l2pg->ref_count > 1, ("pmap_enter_l2: l2pg's ref count is too low")); if ((flags & PMAP_ENTER_NOREPLACE) != 0) { if ((old_l2 & ATTR_DESCR_MASK) == L2_BLOCK) { if (l2pg != NULL) l2pg->ref_count--; CTR2(KTR_PMAP, "pmap_enter_l2: no space for va %#lx" " in pmap %p", va, pmap); return (KERN_NO_SPACE); } else if (!ADDR_IS_KERNEL(va) || !pmap_every_pte_zero(PTE_TO_PHYS(old_l2))) { if (l2pg != NULL) l2pg->ref_count--; CTR2(KTR_PMAP, "pmap_enter_l2: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_FAILURE); } } SLIST_INIT(&free); if ((old_l2 & ATTR_DESCR_MASK) == L2_BLOCK) (void)pmap_remove_l2(pmap, l2, va, pmap_load(pmap_l1(pmap, va)), &free, lockp); else pmap_remove_l3_range(pmap, old_l2, va, va + L2_SIZE, &free, lockp); if (!ADDR_IS_KERNEL(va)) { vm_page_free_pages_toq(&free, true); KASSERT(pmap_load(l2) == 0, ("pmap_enter_l2: non-zero L2 entry %p", l2)); } else { KASSERT(SLIST_EMPTY(&free), ("pmap_enter_l2: freed kernel page table page")); /* * Both pmap_remove_l2() and pmap_remove_l3_range() * will leave the kernel page table page zero filled. * Nonetheless, the TLB could have an intermediate * entry for the kernel page table page, so request * an invalidation at all levels after clearing * the L2_TABLE entry. */ mt = PHYS_TO_VM_PAGE(PTE_TO_PHYS(pmap_load(l2))); if (pmap_insert_pt_page(pmap, mt, false, false)) panic("pmap_enter_l2: trie insert failed"); pmap_clear(l2); pmap_s1_invalidate_page(pmap, va, false); } } if ((new_l2 & ATTR_SW_MANAGED) != 0) { /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_l2(pmap, va, new_l2, flags, lockp)) { if (l2pg != NULL) pmap_abort_ptp(pmap, va, l2pg); CTR2(KTR_PMAP, "pmap_enter_l2: failure for va %#lx in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } if ((new_l2 & ATTR_SW_DBM) != 0) for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) vm_page_aflag_set(mt, PGA_WRITEABLE); } /* * Increment counters. */ if ((new_l2 & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count += L2_SIZE / PAGE_SIZE; pmap->pm_stats.resident_count += L2_SIZE / PAGE_SIZE; /* * Conditionally sync the icache. See pmap_enter() for details. */ if ((new_l2 & ATTR_S1_XN) == 0 && (PTE_TO_PHYS(new_l2) != PTE_TO_PHYS(old_l2) || (old_l2 & ATTR_S1_XN) != 0) && pmap != kernel_pmap && m->md.pv_memattr == VM_MEMATTR_WRITE_BACK) { cpu_icache_sync_range(PHYS_TO_DMAP(PTE_TO_PHYS(new_l2)), L2_SIZE); } /* * Map the superpage. */ pmap_store(l2, new_l2); dsb(ishst); atomic_add_long(&pmap_l2_mappings, 1); CTR2(KTR_PMAP, "pmap_enter_l2: success for va %#lx in pmap %p", va, pmap); return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { struct rwlock *lock; vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; int rv; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; lock = NULL; PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & L2_OFFSET) == 0 && va + L2_SIZE <= end && m->psind == 1 && pmap_ps_enabled(pmap) && ((rv = pmap_enter_2mpage(pmap, va, m, prot, &lock)) == KERN_SUCCESS || rv == KERN_NO_SPACE)) m = &m[L2_SIZE / PAGE_SIZE - 1]; else mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte, &lock); m = TAILQ_NEXT(m, listq); } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { struct rwlock *lock; lock = NULL; PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp) { pd_entry_t *pde; pt_entry_t *l1, *l2, *l3, l3_val; vm_paddr_t pa; int lvl; KASSERT(!VA_IS_CLEANMAP(va) || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PMAP_ASSERT_STAGE1(pmap); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); l2 = NULL; CTR2(KTR_PMAP, "pmap_enter_quick_locked: %p %lx", pmap, va); /* * In the case that a page table page is not * resident, we are creating it here. */ if (!ADDR_IS_KERNEL(va)) { vm_pindex_t l2pindex; /* * Calculate pagetable page index */ l2pindex = pmap_l2_pindex(va); if (mpte && (mpte->pindex == l2pindex)) { mpte->ref_count++; } else { /* * If the page table page is mapped, we just increment * the hold count, and activate it. Otherwise, we * attempt to allocate a page table page, passing NULL * instead of the PV list lock pointer because we don't * intend to sleep. If this attempt fails, we don't * retry. Instead, we give up. */ l1 = pmap_l1(pmap, va); if (l1 != NULL && pmap_load(l1) != 0) { if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) return (NULL); l2 = pmap_l1_to_l2(l1, va); if (pmap_load(l2) != 0) { if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) return (NULL); mpte = PHYS_TO_VM_PAGE( PTE_TO_PHYS(pmap_load(l2))); mpte->ref_count++; } else { mpte = _pmap_alloc_l3(pmap, l2pindex, NULL); if (mpte == NULL) return (mpte); } } else { mpte = _pmap_alloc_l3(pmap, l2pindex, NULL); if (mpte == NULL) return (mpte); } } l3 = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte)); l3 = &l3[pmap_l3_index(va)]; } else { mpte = NULL; pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_enter_quick_locked: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_enter_quick_locked: Invalid level %d", lvl)); l3 = pmap_l2_to_l3(pde, va); } /* * Abort if a mapping already exists. */ if (pmap_load(l3) != 0) { if (mpte != NULL) mpte->ref_count--; return (NULL); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m, lockp)) { if (mpte != NULL) pmap_abort_ptp(pmap, va, mpte); return (NULL); } /* * Increment counters */ pmap_resident_count_inc(pmap, 1); pa = VM_PAGE_TO_PHYS(m); l3_val = PHYS_TO_PTE(pa) | ATTR_DEFAULT | ATTR_S1_IDX(m->md.pv_memattr) | ATTR_S1_AP(ATTR_S1_AP_RO) | L3_PAGE; if ((prot & VM_PROT_EXECUTE) == 0 || m->md.pv_memattr == VM_MEMATTR_DEVICE) l3_val |= ATTR_S1_XN; if (!ADDR_IS_KERNEL(va)) l3_val |= ATTR_S1_AP(ATTR_S1_AP_USER) | ATTR_S1_PXN; else l3_val |= ATTR_S1_UXN; if (pmap != kernel_pmap) l3_val |= ATTR_S1_nG; /* * Now validate mapping with RO protection */ if ((m->oflags & VPO_UNMANAGED) == 0) { l3_val |= ATTR_SW_MANAGED; l3_val &= ~ATTR_AF; } /* Sync icache before the mapping is stored to PTE */ if ((prot & VM_PROT_EXECUTE) && pmap != kernel_pmap && m->md.pv_memattr == VM_MEMATTR_WRITE_BACK) cpu_icache_sync_range(PHYS_TO_DMAP(pa), PAGE_SIZE); pmap_store(l3, l3_val); dsb(ishst); #if VM_NRESERVLEVEL > 0 /* * If both the PTP and the reservation are fully populated, then * attempt promotion. */ if ((mpte == NULL || mpte->ref_count == NL3PG) && pmap_ps_enabled(pmap) && pmap->pm_stage == PM_STAGE1 && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) { if (l2 == NULL) l2 = pmap_pde(pmap, va, &lvl); pmap_promote_l2(pmap, l2, va, mpte, lockp); /* * If promotion succeeds, then the next call to this function * should not be given the unmapped PTP as a hint. */ mpte = NULL; } #endif return (mpte); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, * so there is no need to invalidate any TLB entries. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va_next; pd_entry_t *l0, *l1, *l2; pt_entry_t *l3; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } l1 = pmap_l0_to_l1(l0, sva); va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", pmap_load(l1), sva, eva, va_next)); MPASS(pmap != kernel_pmap); MPASS((pmap_load(l1) & (ATTR_SW_MANAGED | ATTR_SW_WIRED)) == ATTR_SW_WIRED); pmap_clear_bits(l1, ATTR_SW_WIRED); pmap->pm_stats.wired_count -= L1_SIZE / PAGE_SIZE; continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (pmap_load(l2) == 0) continue; if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) { if ((pmap_load(l2) & ATTR_SW_WIRED) == 0) panic("pmap_unwire: l2 %#jx is missing " "ATTR_SW_WIRED", (uintmax_t)pmap_load(l2)); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + L2_SIZE == va_next && eva >= va_next) { pmap_clear_bits(l2, ATTR_SW_WIRED); pmap->pm_stats.wired_count -= L2_SIZE / PAGE_SIZE; continue; } else if (pmap_demote_l2(pmap, l2, sva) == NULL) panic("pmap_unwire: demotion failed"); } KASSERT((pmap_load(l2) & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_unwire: Invalid l2 entry after demotion")); if (va_next > eva) va_next = eva; for (l3 = pmap_l2_to_l3(l2, sva); sva != va_next; l3++, sva += L3_SIZE) { if (pmap_load(l3) == 0) continue; if ((pmap_load(l3) & ATTR_SW_WIRED) == 0) panic("pmap_unwire: l3 %#jx is missing " "ATTR_SW_WIRED", (uintmax_t)pmap_load(l3)); /* * ATTR_SW_WIRED must be cleared atomically. Although * the pmap lock synchronizes access to ATTR_SW_WIRED, * the System MMU may write to the entry concurrently. */ pmap_clear_bits(l3, ATTR_SW_WIRED); pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. * * Because the executable mappings created by this routine are copied, * it should not have to flush the instruction cache. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct rwlock *lock; pd_entry_t *l0, *l1, *l2, srcptepaddr; pt_entry_t *dst_pte, mask, nbits, ptetemp, *src_pte; vm_offset_t addr, end_addr, va_next; vm_page_t dst_m, dstmpte, srcmpte; PMAP_ASSERT_STAGE1(dst_pmap); PMAP_ASSERT_STAGE1(src_pmap); if (dst_addr != src_addr) return; end_addr = src_addr + len; lock = NULL; if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } for (addr = src_addr; addr < end_addr; addr = va_next) { l0 = pmap_l0(src_pmap, addr); if (pmap_load(l0) == 0) { va_next = (addr + L0_SIZE) & ~L0_OFFSET; if (va_next < addr) va_next = end_addr; continue; } va_next = (addr + L1_SIZE) & ~L1_OFFSET; if (va_next < addr) va_next = end_addr; l1 = pmap_l0_to_l1(l0, addr); if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= end_addr, ("partial update of non-transparent 1G page " "l1 %#lx addr %#lx end_addr %#lx va_next %#lx", pmap_load(l1), addr, end_addr, va_next)); srcptepaddr = pmap_load(l1); l1 = pmap_l1(dst_pmap, addr); if (l1 == NULL) { if (_pmap_alloc_l3(dst_pmap, pmap_l0_pindex(addr), NULL) == NULL) break; l1 = pmap_l1(dst_pmap, addr); } else { l0 = pmap_l0(dst_pmap, addr); dst_m = PHYS_TO_VM_PAGE( PTE_TO_PHYS(pmap_load(l0))); dst_m->ref_count++; } KASSERT(pmap_load(l1) == 0, ("1G mapping present in dst pmap " "l1 %#lx addr %#lx end_addr %#lx va_next %#lx", pmap_load(l1), addr, end_addr, va_next)); pmap_store(l1, srcptepaddr & ~ATTR_SW_WIRED); pmap_resident_count_inc(dst_pmap, L1_SIZE / PAGE_SIZE); continue; } va_next = (addr + L2_SIZE) & ~L2_OFFSET; if (va_next < addr) va_next = end_addr; l2 = pmap_l1_to_l2(l1, addr); srcptepaddr = pmap_load(l2); if (srcptepaddr == 0) continue; if ((srcptepaddr & ATTR_DESCR_MASK) == L2_BLOCK) { /* * We can only virtual copy whole superpages. */ if ((addr & L2_OFFSET) != 0 || addr + L2_SIZE > end_addr) continue; l2 = pmap_alloc_l2(dst_pmap, addr, &dst_m, NULL); if (l2 == NULL) break; if (pmap_load(l2) == 0 && ((srcptepaddr & ATTR_SW_MANAGED) == 0 || pmap_pv_insert_l2(dst_pmap, addr, srcptepaddr, PMAP_ENTER_NORECLAIM, &lock))) { /* * We leave the dirty bit unchanged because * managed read/write superpage mappings are * required to be dirty. However, managed * superpage mappings are not required to * have their accessed bit set, so we clear * it because we don't know if this mapping * will be used. */ srcptepaddr &= ~ATTR_SW_WIRED; if ((srcptepaddr & ATTR_SW_MANAGED) != 0) srcptepaddr &= ~ATTR_AF; pmap_store(l2, srcptepaddr); pmap_resident_count_inc(dst_pmap, L2_SIZE / PAGE_SIZE); atomic_add_long(&pmap_l2_mappings, 1); } else pmap_abort_ptp(dst_pmap, addr, dst_m); continue; } KASSERT((srcptepaddr & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_copy: invalid L2 entry")); srcmpte = PHYS_TO_VM_PAGE(PTE_TO_PHYS(srcptepaddr)); KASSERT(srcmpte->ref_count > 0, ("pmap_copy: source page table page is unused")); if (va_next > end_addr) va_next = end_addr; src_pte = (pt_entry_t *)PHYS_TO_DMAP(PTE_TO_PHYS(srcptepaddr)); src_pte = &src_pte[pmap_l3_index(addr)]; dstmpte = NULL; for (; addr < va_next; addr += PAGE_SIZE, src_pte++) { ptetemp = pmap_load(src_pte); /* * We only virtual copy managed pages. */ if ((ptetemp & ATTR_SW_MANAGED) == 0) continue; if (dstmpte != NULL) { KASSERT(dstmpte->pindex == pmap_l2_pindex(addr), ("dstmpte pindex/addr mismatch")); dstmpte->ref_count++; } else if ((dstmpte = pmap_alloc_l3(dst_pmap, addr, NULL)) == NULL) goto out; dst_pte = (pt_entry_t *) PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte)); dst_pte = &dst_pte[pmap_l3_index(addr)]; if (pmap_load(dst_pte) == 0 && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(PTE_TO_PHYS(ptetemp)), &lock)) { /* * Clear the wired, modified, and accessed * (referenced) bits during the copy. */ mask = ATTR_AF | ATTR_SW_WIRED; nbits = 0; if ((ptetemp & ATTR_SW_DBM) != 0) nbits |= ATTR_S1_AP_RW_BIT; pmap_store(dst_pte, (ptetemp & ~mask) | nbits); pmap_resident_count_inc(dst_pmap, 1); } else { pmap_abort_ptp(dst_pmap, addr, dstmpte); goto out; } /* Have we copied all of the valid mappings? */ if (dstmpte->ref_count >= srcmpte->ref_count) break; } } out: /* * XXX This barrier may not be needed because the destination pmap is * not active. */ dsb(ishst); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } /* * pmap_zero_page zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. */ void pmap_zero_page(vm_page_t m) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); pagezero((void *)va); } /* * pmap_zero_page_area zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. * * off and size may not cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); if (off == 0 && size == PAGE_SIZE) pagezero((void *)va); else bzero((char *)va + off, size); } /* * pmap_copy_page copies the specified (machine independent) * page by mapping the page into virtual memory and using * bcopy to copy the page, one machine dependent page at a * time. */ void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc)); vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst)); pagecopy((void *)src, (void *)dst); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_page_t m_a, m_b; vm_paddr_t p_a, p_b; vm_offset_t a_pg_offset, b_pg_offset; int cnt; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; m_a = ma[a_offset >> PAGE_SHIFT]; p_a = m_a->phys_addr; b_pg_offset = b_offset & PAGE_MASK; m_b = mb[b_offset >> PAGE_SHIFT]; p_b = m_b->phys_addr; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); cnt = min(cnt, PAGE_SIZE - b_pg_offset); if (__predict_false(!PHYS_IN_DMAP(p_a))) { panic("!DMAP a %lx", p_a); } else { a_cp = (char *)PHYS_TO_DMAP(p_a) + a_pg_offset; } if (__predict_false(!PHYS_IN_DMAP(p_b))) { panic("!DMAP b %lx", p_b); } else { b_cp = (char *)PHYS_TO_DMAP(p_b) + b_pg_offset; } bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { return (PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m))); } void pmap_quick_remove_page(vm_offset_t addr) { } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; struct rwlock *lock; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_runlock(lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct rwlock *lock; struct md_page *pvh; pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; int count, md_gen, pvh_gen; if ((m->oflags & VPO_UNMANAGED) != 0) return (0); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: count = 0; TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte_exists(pmap, pv->pv_va, 3, __func__); if ((pmap_load(pte) & ATTR_SW_WIRED) != 0) count++; PMAP_UNLOCK(pmap); } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte_exists(pmap, pv->pv_va, 2, __func__); if ((pmap_load(pte) & ATTR_SW_WIRED) != 0) count++; PMAP_UNLOCK(pmap); } } rw_runlock(lock); return (count); } /* * Returns true if the given page is mapped individually or as part of * a 2mpage. Otherwise, returns false. */ bool pmap_page_is_mapped(vm_page_t m) { struct rwlock *lock; bool rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (false); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&page_to_pvh(m)->pv_list)); rw_runlock(lock); return (rv); } /* * Destroy all managed, non-wired mappings in the given user-space * pmap. This pmap cannot be active on any processor besides the * caller. * * This function cannot be applied to the kernel pmap. Moreover, it * is not intended for general use. It is only to be used during * process termination. Consequently, it can be implemented in ways * that make it faster than pmap_remove(). First, it can more quickly * destroy mappings by iterating over the pmap's collection of PV * entries, rather than searching the page table. Second, it doesn't * have to test and clear the page table entries atomically, because * no processor is currently accessing the user address space. In * particular, a page table entry's dirty bit won't change state once * this function starts. */ void pmap_remove_pages(pmap_t pmap) { pd_entry_t *pde; pt_entry_t *pte, tpte; struct spglist free; struct pv_chunklist free_chunks[PMAP_MEMDOM]; vm_page_t m, ml3, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct rwlock *lock; int64_t bit; uint64_t inuse, bitmask; int allfree, field, i, idx, lvl; int freed __pvused; vm_paddr_t pa; lock = NULL; for (i = 0; i < PMAP_MEMDOM; i++) TAILQ_INIT(&free_chunks[i]); SLIST_INIT(&free); PMAP_LOCK(pmap); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { allfree = 1; freed = 0; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = ffsl(inuse) - 1; bitmask = 1UL << bit; idx = field * 64 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pde = pmap_pde(pmap, pv->pv_va, &lvl); KASSERT(pde != NULL, ("Attempting to remove an unmapped page")); switch(lvl) { case 1: pte = pmap_l1_to_l2(pde, pv->pv_va); tpte = pmap_load(pte); KASSERT((tpte & ATTR_DESCR_MASK) == L2_BLOCK, ("Attempting to remove an invalid " "block: %lx", tpte)); break; case 2: pte = pmap_l2_to_l3(pde, pv->pv_va); tpte = pmap_load(pte); KASSERT((tpte & ATTR_DESCR_MASK) == L3_PAGE, ("Attempting to remove an invalid " "page: %lx", tpte)); break; default: panic( "Invalid page directory level: %d", lvl); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & ATTR_SW_WIRED) { allfree = 0; continue; } /* Mark free */ pc->pc_map[field] |= bitmask; /* * Because this pmap is not active on other * processors, the dirty bit cannot have * changed state since we last loaded pte. */ pmap_clear(pte); pa = PTE_TO_PHYS(tpte); m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad pte %#jx", (uintmax_t)tpte)); /* * Update the vm_page_t clean/reference bits. */ if (pmap_pte_dirty(pmap, tpte)) { switch (lvl) { case 1: for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) vm_page_dirty(mt); break; case 2: vm_page_dirty(m); break; } } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); switch (lvl) { case 1: pmap_resident_count_dec(pmap, L2_SIZE / PAGE_SIZE); pvh = page_to_pvh(m); TAILQ_REMOVE(&pvh->pv_list, pv,pv_next); pvh->pv_gen++; if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) if ((mt->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } ml3 = pmap_remove_pt_page(pmap, pv->pv_va); if (ml3 != NULL) { KASSERT(vm_page_any_valid(ml3), ("pmap_remove_pages: l3 page not promoted")); pmap_resident_count_dec(pmap,1); KASSERT(ml3->ref_count == NL3PG, ("pmap_remove_pages: l3 page ref count error")); ml3->ref_count = 0; pmap_add_delayed_free_list(ml3, &free, FALSE); } break; case 2: pmap_resident_count_dec(pmap, 1); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((m->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } break; } pmap_unuse_pt(pmap, pv->pv_va, pmap_load(pde), &free); freed++; } } PV_STAT(atomic_add_long(&pv_entry_frees, freed)); PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&free_chunks[pc_to_domain(pc)], pc, pc_list); } } if (lock != NULL) rw_wunlock(lock); pmap_invalidate_all(pmap); free_pv_chunk_batch(free_chunks); PMAP_UNLOCK(pmap); vm_page_free_pages_toq(&free, true); } /* * This is used to check if a page has been accessed or modified. */ static boolean_t pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified) { struct rwlock *lock; pv_entry_t pv; struct md_page *pvh; pt_entry_t *pte, mask, value; pmap_t pmap; int md_gen, pvh_gen; boolean_t rv; rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_ASSERT_STAGE1(pmap); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte_exists(pmap, pv->pv_va, 3, __func__); mask = 0; value = 0; if (modified) { mask |= ATTR_S1_AP_RW_BIT; value |= ATTR_S1_AP(ATTR_S1_AP_RW); } if (accessed) { mask |= ATTR_AF | ATTR_DESCR_MASK; value |= ATTR_AF | L3_PAGE; } rv = (pmap_load(pte) & mask) == value; PMAP_UNLOCK(pmap); if (rv) goto out; } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = page_to_pvh(m); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_ASSERT_STAGE1(pmap); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte_exists(pmap, pv->pv_va, 2, __func__); mask = 0; value = 0; if (modified) { mask |= ATTR_S1_AP_RW_BIT; value |= ATTR_S1_AP(ATTR_S1_AP_RW); } if (accessed) { mask |= ATTR_AF | ATTR_DESCR_MASK; value |= ATTR_AF | L2_BLOCK; } rv = (pmap_load(pte) & mask) == value; PMAP_UNLOCK(pmap); if (rv) goto out; } } out: rw_runlock(lock); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not busied then this check is racy. */ if (!pmap_page_is_write_mapped(m)) return (FALSE); return (pmap_page_test_mappings(m, FALSE, TRUE)); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pd_entry_t *pde; pt_entry_t *pte; boolean_t rv; int lvl; /* * Return TRUE if and only if the L3 entry for the specified virtual * address is allocated but invalid. */ rv = FALSE; PMAP_LOCK(pmap); pde = pmap_pde(pmap, addr, &lvl); if (pde != NULL && lvl == 2) { pte = pmap_l2_to_l3(pde, addr); rv = pmap_load(pte) == 0; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); return (pmap_page_test_mappings(m, TRUE, FALSE)); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pmap_t pmap; struct rwlock *lock; pv_entry_t next_pv, pv; pt_entry_t oldpte, *pte, set, clear, mask, val; vm_offset_t va; int md_gen, pvh_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : page_to_pvh(m); rw_wlock(lock); retry: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); PMAP_ASSERT_STAGE1(pmap); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pte = pmap_pte_exists(pmap, va, 2, __func__); if ((pmap_load(pte) & ATTR_SW_DBM) != 0) (void)pmap_demote_l2_locked(pmap, pte, va, &lock); KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } pte = pmap_pte_exists(pmap, pv->pv_va, 3, __func__); oldpte = pmap_load(pte); if ((oldpte & ATTR_SW_DBM) != 0) { if (pmap->pm_stage == PM_STAGE1) { set = ATTR_S1_AP_RW_BIT; clear = 0; mask = ATTR_S1_AP_RW_BIT; val = ATTR_S1_AP(ATTR_S1_AP_RW); } else { set = 0; clear = ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE); mask = ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE); val = ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE); } clear |= ATTR_SW_DBM; while (!atomic_fcmpset_64(pte, &oldpte, (oldpte | set) & ~clear)) cpu_spinwait(); if ((oldpte & mask) == val) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va, true); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); vm_page_aflag_clear(m, PGA_WRITEABLE); } /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; struct rwlock *lock; pt_entry_t *pte, tpte; vm_offset_t va; vm_paddr_t pa; int cleared, md_gen, not_cleared, pvh_gen; struct spglist free; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); SLIST_INIT(&free); cleared = 0; pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : page_to_pvh(m); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_wlock(lock); retry: not_cleared = 0; if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pte = pmap_pte_exists(pmap, va, 2, __func__); tpte = pmap_load(pte); if (pmap_pte_dirty(pmap, tpte)) { /* * Although "tpte" is mapping a 2MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } if ((tpte & ATTR_AF) != 0) { pa = VM_PAGE_TO_PHYS(m); /* * Since this reference bit is shared by 512 4KB pages, * it should not be cleared every time it is tested. * Apply a simple "hash" function on the physical page * number, the virtual superpage number, and the pmap * address to select one 4KB page out of the 512 on * which testing the reference bit will result in * clearing that reference bit. This function is * designed to avoid the selection of the same 4KB page * for every 2MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (va >> L2_SHIFT) ^ (uintptr_t)pmap) & (Ln_ENTRIES - 1)) == 0 && (tpte & ATTR_SW_WIRED) == 0) { pmap_clear_bits(pte, ATTR_AF); pmap_invalidate_page(pmap, va, true); cleared++; } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; } if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } pte = pmap_pte_exists(pmap, pv->pv_va, 3, __func__); tpte = pmap_load(pte); if (pmap_pte_dirty(pmap, tpte)) vm_page_dirty(m); if ((tpte & ATTR_AF) != 0) { if ((tpte & ATTR_SW_WIRED) == 0) { pmap_clear_bits(pte, ATTR_AF); pmap_invalidate_page(pmap, pv->pv_va, true); cleared++; } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared + not_cleared < PMAP_TS_REFERENCED_MAX); out: rw_wunlock(lock); vm_page_free_pages_toq(&free, true); return (cleared + not_cleared); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { struct rwlock *lock; vm_offset_t va, va_next; vm_page_t m; pd_entry_t *l0, *l1, *l2, oldl2; pt_entry_t *l3, oldl3; PMAP_ASSERT_STAGE1(pmap); if (advice != MADV_DONTNEED && advice != MADV_FREE) return; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; l1 = pmap_l0_to_l1(l0, sva); if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); oldl2 = pmap_load(l2); if (oldl2 == 0) continue; if ((oldl2 & ATTR_DESCR_MASK) == L2_BLOCK) { if ((oldl2 & ATTR_SW_MANAGED) == 0) continue; lock = NULL; if (!pmap_demote_l2_locked(pmap, l2, sva, &lock)) { if (lock != NULL) rw_wunlock(lock); /* * The 2MB page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Choosing the last page * within the address range [sva, min(va_next, eva)) * generally results in more repromotions. Since the * underlying page table page is fully populated, this * removal never frees a page table page. */ if ((oldl2 & ATTR_SW_WIRED) == 0) { va = eva; if (va > va_next) va = va_next; va -= PAGE_SIZE; KASSERT(va >= sva, ("pmap_advise: no address gap")); l3 = pmap_l2_to_l3(l2, va); KASSERT(pmap_load(l3) != 0, ("pmap_advise: invalid PTE")); pmap_remove_l3(pmap, l3, va, pmap_load(l2), NULL, &lock); } if (lock != NULL) rw_wunlock(lock); } KASSERT((pmap_load(l2) & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_advise: invalid L2 entry after demotion")); if (va_next > eva) va_next = eva; va = va_next; for (l3 = pmap_l2_to_l3(l2, sva); sva != va_next; l3++, sva += L3_SIZE) { oldl3 = pmap_load(l3); if ((oldl3 & (ATTR_SW_MANAGED | ATTR_DESCR_MASK)) != (ATTR_SW_MANAGED | L3_PAGE)) goto maybe_invlrng; else if (pmap_pte_dirty(pmap, oldl3)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(PTE_TO_PHYS(oldl3)); vm_page_dirty(m); } while (!atomic_fcmpset_long(l3, &oldl3, (oldl3 & ~ATTR_AF) | ATTR_S1_AP(ATTR_S1_AP_RO))) cpu_spinwait(); } else if ((oldl3 & ATTR_AF) != 0) pmap_clear_bits(l3, ATTR_AF); else goto maybe_invlrng; if (va == va_next) va = sva; continue; maybe_invlrng: if (va != va_next) { pmap_s1_invalidate_range(pmap, va, sva, true); va = va_next; } } if (va != va_next) pmap_s1_invalidate_range(pmap, va, sva, true); } PMAP_UNLOCK(pmap); } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { struct md_page *pvh; struct rwlock *lock; pmap_t pmap; pv_entry_t next_pv, pv; pd_entry_t *l2, oldl2; pt_entry_t *l3, oldl3; vm_offset_t va; int md_gen, pvh_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : page_to_pvh(m); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_wlock(lock); restart: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); PMAP_ASSERT_STAGE1(pmap); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } va = pv->pv_va; l2 = pmap_l2(pmap, va); oldl2 = pmap_load(l2); /* If oldl2 has ATTR_SW_DBM set, then it is also dirty. */ if ((oldl2 & ATTR_SW_DBM) != 0 && pmap_demote_l2_locked(pmap, l2, va, &lock) && (oldl2 & ATTR_SW_WIRED) == 0) { /* * Write protect the mapping to a single page so that * a subsequent write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - PTE_TO_PHYS(oldl2); l3 = pmap_l2_to_l3(l2, va); oldl3 = pmap_load(l3); while (!atomic_fcmpset_long(l3, &oldl3, (oldl3 & ~ATTR_SW_DBM) | ATTR_S1_AP(ATTR_S1_AP_RO))) cpu_spinwait(); vm_page_dirty(m); pmap_s1_invalidate_page(pmap, va, true); } PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_ASSERT_STAGE1(pmap); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } l2 = pmap_l2(pmap, pv->pv_va); l3 = pmap_l2_to_l3(l2, pv->pv_va); oldl3 = pmap_load(l3); if ((oldl3 & (ATTR_S1_AP_RW_BIT | ATTR_SW_DBM)) == ATTR_SW_DBM){ pmap_set_bits(l3, ATTR_S1_AP(ATTR_S1_AP_RO)); pmap_s1_invalidate_page(pmap, pv->pv_va, true); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t va, offset; pd_entry_t *pde; pt_entry_t *l2; int i, lvl, l2_blocks, free_l2_count, start_idx; if (!vm_initialized) { /* * No L3 ptables so map entire L2 blocks where start VA is: * preinit_map_va + start_idx * L2_SIZE * There may be duplicate mappings (multiple VA -> same PA) but * ARM64 dcache is always PIPT so that's acceptable. */ if (size == 0) return (NULL); /* Calculate how many L2 blocks are needed for the mapping */ l2_blocks = (roundup2(pa + size, L2_SIZE) - rounddown2(pa, L2_SIZE)) >> L2_SHIFT; offset = pa & L2_OFFSET; if (preinit_map_va == 0) return (NULL); /* Map 2MiB L2 blocks from reserved VA space */ free_l2_count = 0; start_idx = -1; /* Find enough free contiguous VA space */ for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (free_l2_count > 0 && ppim->pa != 0) { /* Not enough space here */ free_l2_count = 0; start_idx = -1; continue; } if (ppim->pa == 0) { /* Free L2 block */ if (start_idx == -1) start_idx = i; free_l2_count++; if (free_l2_count == l2_blocks) break; } } if (free_l2_count != l2_blocks) panic("%s: too many preinit mappings", __func__); va = preinit_map_va + (start_idx * L2_SIZE); for (i = start_idx; i < start_idx + l2_blocks; i++) { /* Mark entries as allocated */ ppim = pmap_preinit_mapping + i; ppim->pa = pa; ppim->va = va + offset; ppim->size = size; } /* Map L2 blocks */ pa = rounddown2(pa, L2_SIZE); for (i = 0; i < l2_blocks; i++) { pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_mapbios: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 1, ("pmap_mapbios: Invalid level %d", lvl)); /* Insert L2_BLOCK */ l2 = pmap_l1_to_l2(pde, va); pmap_load_store(l2, PHYS_TO_PTE(pa) | ATTR_DEFAULT | ATTR_S1_XN | ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | L2_BLOCK); va += L2_SIZE; pa += L2_SIZE; } pmap_s1_invalidate_all(kernel_pmap); va = preinit_map_va + (start_idx * L2_SIZE); } else { /* kva_alloc may be used to map the pages */ offset = pa & PAGE_MASK; size = round_page(offset + size); va = kva_alloc(size); if (va == 0) panic("%s: Couldn't allocate KVA", __func__); pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(lvl == 2, ("pmap_mapbios: Invalid level %d", lvl)); /* L3 table is linked */ va = trunc_page(va); pa = trunc_page(pa); pmap_kenter(va, size, pa, memory_mapping_mode(pa)); } return ((void *)(va + offset)); } void pmap_unmapbios(void *p, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t offset, tmpsize, va, va_trunc; pd_entry_t *pde; pt_entry_t *l2; int i, lvl, l2_blocks, block; bool preinit_map; va = (vm_offset_t)p; l2_blocks = (roundup2(va + size, L2_SIZE) - rounddown2(va, L2_SIZE)) >> L2_SHIFT; KASSERT(l2_blocks > 0, ("pmap_unmapbios: invalid size %lx", size)); /* Remove preinit mapping */ preinit_map = false; block = 0; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == va) { KASSERT(ppim->size == size, ("pmap_unmapbios: size mismatch")); ppim->va = 0; ppim->pa = 0; ppim->size = 0; preinit_map = true; offset = block * L2_SIZE; va_trunc = rounddown2(va, L2_SIZE) + offset; /* Remove L2_BLOCK */ pde = pmap_pde(kernel_pmap, va_trunc, &lvl); KASSERT(pde != NULL, ("pmap_unmapbios: Invalid page entry, va: 0x%lx", va_trunc)); l2 = pmap_l1_to_l2(pde, va_trunc); pmap_clear(l2); if (block == (l2_blocks - 1)) break; block++; } } if (preinit_map) { pmap_s1_invalidate_all(kernel_pmap); return; } /* Unmap the pages reserved with kva_alloc. */ if (vm_initialized) { offset = va & PAGE_MASK; size = round_page(offset + size); va = trunc_page(va); pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_unmapbios: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_unmapbios: Invalid level %d", lvl)); /* Unmap and invalidate the pages */ for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) pmap_kremove(va + tmpsize); kva_free(va, size); } } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pv_memattr = ma; /* * If "m" is a normal page, update its direct mapping. This update * can be relied upon to perform any cache operations that are * required for data coherence. */ if ((m->flags & PG_FICTITIOUS) == 0 && pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, m->md.pv_memattr) != 0) panic("memory attribute change on the direct map failed"); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the direct map or the kernel map. If * the virtual address range is contained within the kernel map, then the * memory type for each of the corresponding ranges of the direct map is also * changed. (The corresponding ranges of the direct map are those ranges that * map the same physical pages as the specified virtual address range.) These * changes to the direct map are necessary because Intel describes the * behavior of their processors as "undefined" if two or more mappings to the * same physical page have different memory types. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. In the * latter case, the memory type may have been changed on some part of the * virtual address range or the direct map. */ int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) { int error; PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(va, size, PROT_NONE, mode, false); PMAP_UNLOCK(kernel_pmap); return (error); } /* * Changes the specified virtual address range's protections to those * specified by "prot". Like pmap_change_attr(), protections for aliases * in the direct map are updated as well. Protections on aliasing mappings may * be a subset of the requested protections; for example, mappings in the direct * map are never executable. */ int pmap_change_prot(vm_offset_t va, vm_size_t size, vm_prot_t prot) { int error; /* Only supported within the kernel map. */ if (va < VM_MIN_KERNEL_ADDRESS) return (EINVAL); PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(va, size, prot, -1, false); PMAP_UNLOCK(kernel_pmap); return (error); } static int pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, int mode, bool skip_unmapped) { vm_offset_t base, offset, tmpva; vm_size_t pte_size; vm_paddr_t pa; pt_entry_t pte, *ptep, *newpte; pt_entry_t bits, mask; int lvl, rv; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); if (!VIRT_IN_DMAP(base) && !(base >= VM_MIN_KERNEL_ADDRESS && base < VM_MAX_KERNEL_ADDRESS)) return (EINVAL); bits = 0; mask = 0; if (mode != -1) { bits = ATTR_S1_IDX(mode); mask = ATTR_S1_IDX_MASK; if (mode == VM_MEMATTR_DEVICE) { mask |= ATTR_S1_XN; bits |= ATTR_S1_XN; } } if (prot != VM_PROT_NONE) { /* Don't mark the DMAP as executable. It never is on arm64. */ if (VIRT_IN_DMAP(base)) { prot &= ~VM_PROT_EXECUTE; /* * XXX Mark the DMAP as writable for now. We rely * on this in ddb & dtrace to insert breakpoint * instructions. */ prot |= VM_PROT_WRITE; } if ((prot & VM_PROT_WRITE) == 0) { bits |= ATTR_S1_AP(ATTR_S1_AP_RO); } if ((prot & VM_PROT_EXECUTE) == 0) { bits |= ATTR_S1_PXN; } bits |= ATTR_S1_UXN; mask |= ATTR_S1_AP_MASK | ATTR_S1_XN; } for (tmpva = base; tmpva < base + size; ) { ptep = pmap_pte(kernel_pmap, tmpva, &lvl); if (ptep == NULL && !skip_unmapped) { return (EINVAL); } else if ((ptep == NULL && skip_unmapped) || (pmap_load(ptep) & mask) == bits) { /* * We already have the correct attribute or there * is no memory mapped at this address and we are * skipping unmapped memory. */ switch (lvl) { default: panic("Invalid DMAP table level: %d\n", lvl); case 1: tmpva = (tmpva & ~L1_OFFSET) + L1_SIZE; break; case 2: tmpva = (tmpva & ~L2_OFFSET) + L2_SIZE; break; case 3: tmpva += PAGE_SIZE; break; } } else { /* We can't demote/promote this entry */ MPASS((pmap_load(ptep) & ATTR_SW_NO_PROMOTE) == 0); /* * Split the entry to an level 3 table, then * set the new attribute. */ switch (lvl) { default: panic("Invalid DMAP table level: %d\n", lvl); case 1: PMAP_ASSERT_L1_BLOCKS_SUPPORTED; if ((tmpva & L1_OFFSET) == 0 && (base + size - tmpva) >= L1_SIZE) { pte_size = L1_SIZE; break; } newpte = pmap_demote_l1(kernel_pmap, ptep, tmpva & ~L1_OFFSET); if (newpte == NULL) return (EINVAL); ptep = pmap_l1_to_l2(ptep, tmpva); /* FALLTHROUGH */ case 2: if ((tmpva & L2_OFFSET) == 0 && (base + size - tmpva) >= L2_SIZE) { pte_size = L2_SIZE; break; } newpte = pmap_demote_l2(kernel_pmap, ptep, tmpva); if (newpte == NULL) return (EINVAL); ptep = pmap_l2_to_l3(ptep, tmpva); /* FALLTHROUGH */ case 3: pte_size = PAGE_SIZE; break; } /* Update the entry */ pte = pmap_load(ptep); pte &= ~mask; pte |= bits; pmap_update_entry(kernel_pmap, ptep, pte, tmpva, pte_size); pa = PTE_TO_PHYS(pte); if (!VIRT_IN_DMAP(tmpva) && PHYS_IN_DMAP(pa)) { /* * Keep the DMAP memory in sync. */ rv = pmap_change_props_locked( PHYS_TO_DMAP(pa), pte_size, prot, mode, true); if (rv != 0) return (rv); } /* * If moving to a non-cacheable entry flush * the cache. */ if (mode == VM_MEMATTR_UNCACHEABLE) cpu_dcache_wbinv_range(tmpva, pte_size); tmpva += pte_size; } } return (0); } /* * Create an L2 table to map all addresses within an L1 mapping. */ static pt_entry_t * pmap_demote_l1(pmap_t pmap, pt_entry_t *l1, vm_offset_t va) { pt_entry_t *l2, newl2, oldl1; vm_offset_t tmpl1; vm_paddr_t l2phys, phys; vm_page_t ml2; int i; PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldl1 = pmap_load(l1); PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT((oldl1 & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_demote_l1: Demoting a non-block entry")); KASSERT((va & L1_OFFSET) == 0, ("pmap_demote_l1: Invalid virtual address %#lx", va)); KASSERT((oldl1 & ATTR_SW_MANAGED) == 0, ("pmap_demote_l1: Level 1 table shouldn't be managed")); KASSERT((oldl1 & ATTR_SW_NO_PROMOTE) == 0, ("pmap_demote_l1: Demoting entry with no-demote flag set")); tmpl1 = 0; if (va <= (vm_offset_t)l1 && va + L1_SIZE > (vm_offset_t)l1) { tmpl1 = kva_alloc(PAGE_SIZE); if (tmpl1 == 0) return (NULL); } if ((ml2 = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED)) == NULL) { CTR2(KTR_PMAP, "pmap_demote_l1: failure for va %#lx" " in pmap %p", va, pmap); l2 = NULL; goto fail; } l2phys = VM_PAGE_TO_PHYS(ml2); l2 = (pt_entry_t *)PHYS_TO_DMAP(l2phys); /* Address the range points at */ phys = PTE_TO_PHYS(oldl1); /* The attributed from the old l1 table to be copied */ newl2 = oldl1 & ATTR_MASK; /* Create the new entries */ for (i = 0; i < Ln_ENTRIES; i++) { l2[i] = newl2 | phys; phys += L2_SIZE; } KASSERT(l2[0] == ((oldl1 & ~ATTR_DESCR_MASK) | L2_BLOCK), ("Invalid l2 page (%lx != %lx)", l2[0], (oldl1 & ~ATTR_DESCR_MASK) | L2_BLOCK)); if (tmpl1 != 0) { pmap_kenter(tmpl1, PAGE_SIZE, DMAP_TO_PHYS((vm_offset_t)l1) & ~L3_OFFSET, VM_MEMATTR_WRITE_BACK); l1 = (pt_entry_t *)(tmpl1 + ((vm_offset_t)l1 & PAGE_MASK)); } pmap_update_entry(pmap, l1, l2phys | L1_TABLE, va, PAGE_SIZE); fail: if (tmpl1 != 0) { pmap_kremove(tmpl1); kva_free(tmpl1, PAGE_SIZE); } return (l2); } static void pmap_fill_l3(pt_entry_t *firstl3, pt_entry_t newl3) { pt_entry_t *l3; for (l3 = firstl3; l3 - firstl3 < Ln_ENTRIES; l3++) { *l3 = newl3; newl3 += L3_SIZE; } } static void pmap_demote_l2_check(pt_entry_t *firstl3p __unused, pt_entry_t newl3e __unused) { #ifdef INVARIANTS #ifdef DIAGNOSTIC pt_entry_t *xl3p, *yl3p; for (xl3p = firstl3p; xl3p < firstl3p + Ln_ENTRIES; xl3p++, newl3e += PAGE_SIZE) { if (PTE_TO_PHYS(pmap_load(xl3p)) != PTE_TO_PHYS(newl3e)) { printf("pmap_demote_l2: xl3e %zd and newl3e map " "different pages: found %#lx, expected %#lx\n", xl3p - firstl3p, pmap_load(xl3p), newl3e); printf("page table dump\n"); for (yl3p = firstl3p; yl3p < firstl3p + Ln_ENTRIES; yl3p++) { printf("%zd %#lx\n", yl3p - firstl3p, pmap_load(yl3p)); } panic("firstpte"); } } #else KASSERT(PTE_TO_PHYS(pmap_load(firstl3p)) == PTE_TO_PHYS(newl3e), ("pmap_demote_l2: firstl3 and newl3e map different physical" " addresses")); #endif #endif } static void pmap_demote_l2_abort(pmap_t pmap, vm_offset_t va, pt_entry_t *l2, struct rwlock **lockp) { struct spglist free; SLIST_INIT(&free); (void)pmap_remove_l2(pmap, l2, va, pmap_load(pmap_l1(pmap, va)), &free, lockp); vm_page_free_pages_toq(&free, true); } /* * Create an L3 table to map all addresses within an L2 mapping. */ static pt_entry_t * pmap_demote_l2_locked(pmap_t pmap, pt_entry_t *l2, vm_offset_t va, struct rwlock **lockp) { pt_entry_t *l3, newl3, oldl2; vm_offset_t tmpl2; vm_paddr_t l3phys; vm_page_t ml3; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PMAP_ASSERT_STAGE1(pmap); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); l3 = NULL; oldl2 = pmap_load(l2); KASSERT((oldl2 & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_demote_l2: Demoting a non-block entry")); KASSERT((oldl2 & ATTR_SW_NO_PROMOTE) == 0, ("pmap_demote_l2: Demoting entry with no-demote flag set")); va &= ~L2_OFFSET; tmpl2 = 0; if (va <= (vm_offset_t)l2 && va + L2_SIZE > (vm_offset_t)l2) { tmpl2 = kva_alloc(PAGE_SIZE); if (tmpl2 == 0) return (NULL); } /* * Invalidate the 2MB page mapping and return "failure" if the * mapping was never accessed. */ if ((oldl2 & ATTR_AF) == 0) { KASSERT((oldl2 & ATTR_SW_WIRED) == 0, ("pmap_demote_l2: a wired mapping is missing ATTR_AF")); pmap_demote_l2_abort(pmap, va, l2, lockp); CTR2(KTR_PMAP, "pmap_demote_l2: failure for va %#lx in pmap %p", va, pmap); goto fail; } if ((ml3 = pmap_remove_pt_page(pmap, va)) == NULL) { KASSERT((oldl2 & ATTR_SW_WIRED) == 0, ("pmap_demote_l2: page table page for a wired mapping" " is missing")); /* * If the page table page is missing and the mapping * is for a kernel address, the mapping must belong to * either the direct map or the early kernel memory. * Page table pages are preallocated for every other * part of the kernel address space, so the direct map * region and early kernel memory are the only parts of the * kernel address space that must be handled here. */ KASSERT(!ADDR_IS_KERNEL(va) || VIRT_IN_DMAP(va) || (va >= VM_MIN_KERNEL_ADDRESS && va < kernel_vm_end), ("pmap_demote_l2: No saved mpte for va %#lx", va)); /* * If the 2MB page mapping belongs to the direct map * region of the kernel's address space, then the page * allocation request specifies the highest possible * priority (VM_ALLOC_INTERRUPT). Otherwise, the * priority is normal. */ ml3 = vm_page_alloc_noobj( (VIRT_IN_DMAP(va) ? VM_ALLOC_INTERRUPT : 0) | VM_ALLOC_WIRED); /* * If the allocation of the new page table page fails, * invalidate the 2MB page mapping and return "failure". */ if (ml3 == NULL) { pmap_demote_l2_abort(pmap, va, l2, lockp); CTR2(KTR_PMAP, "pmap_demote_l2: failure for va %#lx" " in pmap %p", va, pmap); goto fail; } ml3->pindex = pmap_l2_pindex(va); if (!ADDR_IS_KERNEL(va)) { ml3->ref_count = NL3PG; pmap_resident_count_inc(pmap, 1); } } l3phys = VM_PAGE_TO_PHYS(ml3); l3 = (pt_entry_t *)PHYS_TO_DMAP(l3phys); newl3 = (oldl2 & ~ATTR_DESCR_MASK) | L3_PAGE; KASSERT((oldl2 & (ATTR_S1_AP_RW_BIT | ATTR_SW_DBM)) != (ATTR_S1_AP(ATTR_S1_AP_RO) | ATTR_SW_DBM), ("pmap_demote_l2: L2 entry is writeable but not dirty")); /* * If the PTP is not leftover from an earlier promotion or it does not * have ATTR_AF set in every L3E, then fill it. The new L3Es will all * have ATTR_AF set. * * When pmap_update_entry() clears the old L2 mapping, it (indirectly) * performs a dsb(). That dsb() ensures that the stores for filling * "l3" are visible before "l3" is added to the page table. */ if (!vm_page_all_valid(ml3)) pmap_fill_l3(l3, newl3); pmap_demote_l2_check(l3, newl3); /* * If the mapping has changed attributes, update the L3Es. */ if ((pmap_load(l3) & (ATTR_MASK & ~ATTR_AF)) != (newl3 & (ATTR_MASK & ~ATTR_AF))) pmap_fill_l3(l3, newl3); /* * Map the temporary page so we don't lose access to the l2 table. */ if (tmpl2 != 0) { pmap_kenter(tmpl2, PAGE_SIZE, DMAP_TO_PHYS((vm_offset_t)l2) & ~L3_OFFSET, VM_MEMATTR_WRITE_BACK); l2 = (pt_entry_t *)(tmpl2 + ((vm_offset_t)l2 & PAGE_MASK)); } /* * The spare PV entries must be reserved prior to demoting the * mapping, that is, prior to changing the PDE. Otherwise, the state * of the L2 and the PV lists will be inconsistent, which can result * in reclaim_pv_chunk() attempting to remove a PV entry from the * wrong PV list and pmap_pv_demote_l2() failing to find the expected * PV entry for the 2MB page mapping that is being demoted. */ if ((oldl2 & ATTR_SW_MANAGED) != 0) reserve_pv_entries(pmap, Ln_ENTRIES - 1, lockp); /* * Pass PAGE_SIZE so that a single TLB invalidation is performed on * the 2MB page mapping. */ pmap_update_entry(pmap, l2, l3phys | L2_TABLE, va, PAGE_SIZE); /* * Demote the PV entry. */ if ((oldl2 & ATTR_SW_MANAGED) != 0) pmap_pv_demote_l2(pmap, va, PTE_TO_PHYS(oldl2), lockp); atomic_add_long(&pmap_l2_demotions, 1); CTR3(KTR_PMAP, "pmap_demote_l2: success for va %#lx" " in pmap %p %lx", va, pmap, l3[0]); fail: if (tmpl2 != 0) { pmap_kremove(tmpl2); kva_free(tmpl2, PAGE_SIZE); } return (l3); } static pt_entry_t * pmap_demote_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t va) { struct rwlock *lock; pt_entry_t *l3; lock = NULL; l3 = pmap_demote_l2_locked(pmap, l2, va, &lock); if (lock != NULL) rw_wunlock(lock); return (l3); } /* * Perform the pmap work for mincore(2). If the page is not both referenced and * modified by this pmap, returns its physical address so that the caller can * find other mappings. */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap) { pt_entry_t *pte, tpte; vm_paddr_t mask, pa; int lvl, val; bool managed; PMAP_ASSERT_STAGE1(pmap); PMAP_LOCK(pmap); pte = pmap_pte(pmap, addr, &lvl); if (pte != NULL) { tpte = pmap_load(pte); switch (lvl) { case 3: mask = L3_OFFSET; break; case 2: mask = L2_OFFSET; break; case 1: mask = L1_OFFSET; break; default: panic("pmap_mincore: invalid level %d", lvl); } managed = (tpte & ATTR_SW_MANAGED) != 0; val = MINCORE_INCORE; if (lvl != 3) val |= MINCORE_PSIND(3 - lvl); if ((managed && pmap_pte_dirty(pmap, tpte)) || (!managed && (tpte & ATTR_S1_AP_RW_BIT) == ATTR_S1_AP(ATTR_S1_AP_RW))) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((tpte & ATTR_AF) == ATTR_AF) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; pa = PTE_TO_PHYS(tpte) | (addr & mask); } else { managed = false; val = 0; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && managed) { *pap = pa; } PMAP_UNLOCK(pmap); return (val); } /* * Garbage collect every ASID that is neither active on a processor nor * reserved. */ static void pmap_reset_asid_set(pmap_t pmap) { pmap_t curpmap; int asid, cpuid, epoch; struct asid_set *set; enum pmap_stage stage; set = pmap->pm_asid_set; stage = pmap->pm_stage; set = pmap->pm_asid_set; KASSERT(set != NULL, ("%s: NULL asid set", __func__)); mtx_assert(&set->asid_set_mutex, MA_OWNED); /* * Ensure that the store to asid_epoch is globally visible before the * loads from pc_curpmap are performed. */ epoch = set->asid_epoch + 1; if (epoch == INT_MAX) epoch = 0; set->asid_epoch = epoch; dsb(ishst); if (stage == PM_STAGE1) { __asm __volatile("tlbi vmalle1is"); } else { KASSERT(pmap_clean_stage2_tlbi != NULL, ("%s: Unset stage 2 tlb invalidation callback\n", __func__)); pmap_clean_stage2_tlbi(); } dsb(ish); bit_nclear(set->asid_set, ASID_FIRST_AVAILABLE, set->asid_set_size - 1); CPU_FOREACH(cpuid) { if (cpuid == curcpu) continue; if (stage == PM_STAGE1) { curpmap = pcpu_find(cpuid)->pc_curpmap; PMAP_ASSERT_STAGE1(pmap); } else { curpmap = pcpu_find(cpuid)->pc_curvmpmap; if (curpmap == NULL) continue; PMAP_ASSERT_STAGE2(pmap); } KASSERT(curpmap->pm_asid_set == set, ("Incorrect set")); asid = COOKIE_TO_ASID(curpmap->pm_cookie); if (asid == -1) continue; bit_set(set->asid_set, asid); curpmap->pm_cookie = COOKIE_FROM(asid, epoch); } } /* * Allocate a new ASID for the specified pmap. */ static void pmap_alloc_asid(pmap_t pmap) { struct asid_set *set; int new_asid; set = pmap->pm_asid_set; KASSERT(set != NULL, ("%s: NULL asid set", __func__)); mtx_lock_spin(&set->asid_set_mutex); /* * While this processor was waiting to acquire the asid set mutex, * pmap_reset_asid_set() running on another processor might have * updated this pmap's cookie to the current epoch. In which case, we * don't need to allocate a new ASID. */ if (COOKIE_TO_EPOCH(pmap->pm_cookie) == set->asid_epoch) goto out; bit_ffc_at(set->asid_set, set->asid_next, set->asid_set_size, &new_asid); if (new_asid == -1) { bit_ffc_at(set->asid_set, ASID_FIRST_AVAILABLE, set->asid_next, &new_asid); if (new_asid == -1) { pmap_reset_asid_set(pmap); bit_ffc_at(set->asid_set, ASID_FIRST_AVAILABLE, set->asid_set_size, &new_asid); KASSERT(new_asid != -1, ("ASID allocation failure")); } } bit_set(set->asid_set, new_asid); set->asid_next = new_asid + 1; pmap->pm_cookie = COOKIE_FROM(new_asid, set->asid_epoch); out: mtx_unlock_spin(&set->asid_set_mutex); } static uint64_t __read_mostly ttbr_flags; /* * Compute the value that should be stored in ttbr0 to activate the specified * pmap. This value may change from time to time. */ uint64_t pmap_to_ttbr0(pmap_t pmap) { uint64_t ttbr; ttbr = pmap->pm_ttbr; ttbr |= ASID_TO_OPERAND(COOKIE_TO_ASID(pmap->pm_cookie)); ttbr |= ttbr_flags; return (ttbr); } static void pmap_set_cnp(void *arg) { uint64_t ttbr0, ttbr1; u_int cpuid; cpuid = *(u_int *)arg; if (cpuid == curcpu) { /* * Set the flags while all CPUs are handling the * smp_rendezvous so will not call pmap_to_ttbr0. Any calls * to pmap_to_ttbr0 after this will have the CnP flag set. * The dsb after invalidating the TLB will act as a barrier * to ensure all CPUs can observe this change. */ ttbr_flags |= TTBR_CnP; } ttbr0 = READ_SPECIALREG(ttbr0_el1); ttbr0 |= TTBR_CnP; ttbr1 = READ_SPECIALREG(ttbr1_el1); ttbr1 |= TTBR_CnP; /* Update ttbr{0,1}_el1 with the CnP flag */ WRITE_SPECIALREG(ttbr0_el1, ttbr0); WRITE_SPECIALREG(ttbr1_el1, ttbr1); isb(); __asm __volatile("tlbi vmalle1is"); dsb(ish); isb(); } /* * Defer enabling CnP until we have read the ID registers to know if it's * supported on all CPUs. */ static void pmap_init_cnp(void *dummy __unused) { uint64_t reg; u_int cpuid; if (!get_kernel_reg(ID_AA64MMFR2_EL1, ®)) return; if (ID_AA64MMFR2_CnP_VAL(reg) != ID_AA64MMFR2_CnP_NONE) { if (bootverbose) printf("Enabling CnP\n"); cpuid = curcpu; smp_rendezvous(NULL, pmap_set_cnp, NULL, &cpuid); } } SYSINIT(pmap_init_cnp, SI_SUB_SMP, SI_ORDER_ANY, pmap_init_cnp, NULL); static bool pmap_activate_int(pmap_t pmap) { struct asid_set *set; int epoch; KASSERT(PCPU_GET(curpmap) != NULL, ("no active pmap")); KASSERT(pmap != kernel_pmap, ("kernel pmap activation")); if ((pmap->pm_stage == PM_STAGE1 && pmap == PCPU_GET(curpmap)) || (pmap->pm_stage == PM_STAGE2 && pmap == PCPU_GET(curvmpmap))) { /* * Handle the possibility that the old thread was preempted * after an "ic" or "tlbi" instruction but before it performed * a "dsb" instruction. If the old thread migrates to a new * processor, its completion of a "dsb" instruction on that * new processor does not guarantee that the "ic" or "tlbi" * instructions performed on the old processor have completed. */ dsb(ish); return (false); } set = pmap->pm_asid_set; KASSERT(set != NULL, ("%s: NULL asid set", __func__)); /* * Ensure that the store to curpmap is globally visible before the * load from asid_epoch is performed. */ if (pmap->pm_stage == PM_STAGE1) PCPU_SET(curpmap, pmap); else PCPU_SET(curvmpmap, pmap); dsb(ish); epoch = COOKIE_TO_EPOCH(pmap->pm_cookie); if (epoch >= 0 && epoch != set->asid_epoch) pmap_alloc_asid(pmap); if (pmap->pm_stage == PM_STAGE1) { set_ttbr0(pmap_to_ttbr0(pmap)); if (PCPU_GET(bcast_tlbi_workaround) != 0) invalidate_local_icache(); } return (true); } void pmap_activate_vm(pmap_t pmap) { PMAP_ASSERT_STAGE2(pmap); (void)pmap_activate_int(pmap); } void pmap_activate(struct thread *td) { pmap_t pmap; pmap = vmspace_pmap(td->td_proc->p_vmspace); PMAP_ASSERT_STAGE1(pmap); critical_enter(); (void)pmap_activate_int(pmap); critical_exit(); } /* * Activate the thread we are switching to. * To simplify the assembly in cpu_throw return the new threads pcb. */ struct pcb * pmap_switch(struct thread *new) { pcpu_bp_harden bp_harden; struct pcb *pcb; /* Store the new curthread */ PCPU_SET(curthread, new); /* And the new pcb */ pcb = new->td_pcb; PCPU_SET(curpcb, pcb); /* * TODO: We may need to flush the cache here if switching * to a user process. */ if (pmap_activate_int(vmspace_pmap(new->td_proc->p_vmspace))) { /* * Stop userspace from training the branch predictor against * other processes. This will call into a CPU specific * function that clears the branch predictor state. */ bp_harden = PCPU_GET(bp_harden); if (bp_harden != NULL) bp_harden(); } return (pcb); } void pmap_sync_icache(pmap_t pmap, vm_offset_t va, vm_size_t sz) { PMAP_ASSERT_STAGE1(pmap); KASSERT(ADDR_IS_CANONICAL(va), ("%s: Address not in canonical form: %lx", __func__, va)); if (ADDR_IS_KERNEL(va)) { cpu_icache_sync_range(va, sz); } else { u_int len, offset; vm_paddr_t pa; /* Find the length of data in this page to flush */ offset = va & PAGE_MASK; len = imin(PAGE_SIZE - offset, sz); while (sz != 0) { /* Extract the physical address & find it in the DMAP */ pa = pmap_extract(pmap, va); if (pa != 0) cpu_icache_sync_range(PHYS_TO_DMAP(pa), len); /* Move to the next page */ sz -= len; va += len; /* Set the length for the next iteration */ len = imin(PAGE_SIZE, sz); } } } static int pmap_stage2_fault(pmap_t pmap, uint64_t esr, uint64_t far) { pd_entry_t *pdep; pt_entry_t *ptep, pte; int rv, lvl, dfsc; PMAP_ASSERT_STAGE2(pmap); rv = KERN_FAILURE; /* Data and insn aborts use same encoding for FSC field. */ dfsc = esr & ISS_DATA_DFSC_MASK; switch (dfsc) { case ISS_DATA_DFSC_TF_L0: case ISS_DATA_DFSC_TF_L1: case ISS_DATA_DFSC_TF_L2: case ISS_DATA_DFSC_TF_L3: PMAP_LOCK(pmap); pdep = pmap_pde(pmap, far, &lvl); if (pdep == NULL || lvl != (dfsc - ISS_DATA_DFSC_TF_L1)) { PMAP_UNLOCK(pmap); break; } switch (lvl) { case 0: ptep = pmap_l0_to_l1(pdep, far); break; case 1: ptep = pmap_l1_to_l2(pdep, far); break; case 2: ptep = pmap_l2_to_l3(pdep, far); break; default: panic("%s: Invalid pde level %d", __func__,lvl); } goto fault_exec; case ISS_DATA_DFSC_AFF_L1: case ISS_DATA_DFSC_AFF_L2: case ISS_DATA_DFSC_AFF_L3: PMAP_LOCK(pmap); ptep = pmap_pte(pmap, far, &lvl); fault_exec: if (ptep != NULL && (pte = pmap_load(ptep)) != 0) { if (icache_vmid) { pmap_invalidate_vpipt_icache(); } else { /* * If accessing an executable page invalidate * the I-cache so it will be valid when we * continue execution in the guest. The D-cache * is assumed to already be clean to the Point * of Coherency. */ if ((pte & ATTR_S2_XN_MASK) != ATTR_S2_XN(ATTR_S2_XN_NONE)) { invalidate_icache(); } } pmap_set_bits(ptep, ATTR_AF | ATTR_DESCR_VALID); rv = KERN_SUCCESS; } PMAP_UNLOCK(pmap); break; } return (rv); } int pmap_fault(pmap_t pmap, uint64_t esr, uint64_t far) { pt_entry_t pte, *ptep; register_t intr; uint64_t ec, par; int lvl, rv; rv = KERN_FAILURE; ec = ESR_ELx_EXCEPTION(esr); switch (ec) { case EXCP_INSN_ABORT_L: case EXCP_INSN_ABORT: case EXCP_DATA_ABORT_L: case EXCP_DATA_ABORT: break; default: return (rv); } if (pmap->pm_stage == PM_STAGE2) return (pmap_stage2_fault(pmap, esr, far)); /* Data and insn aborts use same encoding for FSC field. */ switch (esr & ISS_DATA_DFSC_MASK) { case ISS_DATA_DFSC_AFF_L1: case ISS_DATA_DFSC_AFF_L2: case ISS_DATA_DFSC_AFF_L3: PMAP_LOCK(pmap); ptep = pmap_pte(pmap, far, &lvl); if (ptep != NULL) { pmap_set_bits(ptep, ATTR_AF); rv = KERN_SUCCESS; /* * XXXMJ as an optimization we could mark the entry * dirty if this is a write fault. */ } PMAP_UNLOCK(pmap); break; case ISS_DATA_DFSC_PF_L1: case ISS_DATA_DFSC_PF_L2: case ISS_DATA_DFSC_PF_L3: if ((ec != EXCP_DATA_ABORT_L && ec != EXCP_DATA_ABORT) || (esr & ISS_DATA_WnR) == 0) return (rv); PMAP_LOCK(pmap); ptep = pmap_pte(pmap, far, &lvl); if (ptep != NULL && ((pte = pmap_load(ptep)) & ATTR_SW_DBM) != 0) { if ((pte & ATTR_S1_AP_RW_BIT) == ATTR_S1_AP(ATTR_S1_AP_RO)) { pmap_clear_bits(ptep, ATTR_S1_AP_RW_BIT); pmap_s1_invalidate_page(pmap, far, true); } rv = KERN_SUCCESS; } PMAP_UNLOCK(pmap); break; case ISS_DATA_DFSC_TF_L0: case ISS_DATA_DFSC_TF_L1: case ISS_DATA_DFSC_TF_L2: case ISS_DATA_DFSC_TF_L3: /* * Retry the translation. A break-before-make sequence can * produce a transient fault. */ if (pmap == kernel_pmap) { /* * The translation fault may have occurred within a * critical section. Therefore, we must check the * address without acquiring the kernel pmap's lock. */ if (pmap_klookup(far, NULL)) rv = KERN_SUCCESS; } else { PMAP_LOCK(pmap); /* Ask the MMU to check the address. */ intr = intr_disable(); par = arm64_address_translate_s1e0r(far); intr_restore(intr); PMAP_UNLOCK(pmap); /* * If the translation was successful, then we can * return success to the trap handler. */ if (PAR_SUCCESS(par)) rv = KERN_SUCCESS; } break; } return (rv); } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < L2_SIZE) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & L2_OFFSET; if (size - ((L2_SIZE - superpage_offset) & L2_OFFSET) < L2_SIZE || (*addr & L2_OFFSET) == superpage_offset) return; if ((*addr & L2_OFFSET) < superpage_offset) *addr = (*addr & ~L2_OFFSET) + superpage_offset; else *addr = ((*addr + L2_OFFSET) & ~L2_OFFSET) + superpage_offset; } /** * Get the kernel virtual address of a set of physical pages. If there are * physical addresses not covered by the DMAP perform a transient mapping * that will be removed when calling pmap_unmap_io_transient. * * \param page The pages the caller wishes to obtain the virtual * address on the kernel memory map. * \param vaddr On return contains the kernel virtual memory address * of the pages passed in the page parameter. * \param count Number of pages passed in. * \param can_fault true if the thread using the mapped pages can take * page faults, false otherwise. * * \returns true if the caller must call pmap_unmap_io_transient when * finished or false otherwise. * */ bool pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, bool can_fault) { vm_paddr_t paddr; bool needs_mapping; int error __diagused, i; /* * Allocate any KVA space that we need, this is done in a separate * loop to prevent calling vmem_alloc while pinned. */ needs_mapping = false; for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (__predict_false(!PHYS_IN_DMAP(paddr))) { error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr[i]); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); needs_mapping = true; } else { vaddr[i] = PHYS_TO_DMAP(paddr); } } /* Exit early if everything is covered by the DMAP */ if (!needs_mapping) return (false); if (!can_fault) sched_pin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (!PHYS_IN_DMAP(paddr)) { panic( "pmap_map_io_transient: TODO: Map out of DMAP data"); } } return (needs_mapping); } void pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, bool can_fault) { vm_paddr_t paddr; int i; if (!can_fault) sched_unpin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (!PHYS_IN_DMAP(paddr)) { panic("ARM64TODO: pmap_unmap_io_transient: Unmap data"); } } } boolean_t pmap_is_valid_memattr(pmap_t pmap __unused, vm_memattr_t mode) { return (mode >= VM_MEMATTR_DEVICE && mode <= VM_MEMATTR_WRITE_THROUGH); } #if defined(KASAN) static vm_paddr_t pmap_san_early_kernstart; static pd_entry_t *pmap_san_early_l2; void __nosanitizeaddress pmap_san_bootstrap(struct arm64_bootparams *abp) { pmap_san_early_kernstart = KERNBASE - abp->kern_delta; kasan_init_early(abp->kern_stack, KSTACK_PAGES * PAGE_SIZE); } #define SAN_BOOTSTRAP_L2_SIZE (1 * L2_SIZE) #define SAN_BOOTSTRAP_SIZE (2 * PAGE_SIZE) static vm_offset_t __nosanitizeaddress pmap_san_enter_bootstrap_alloc_l2(void) { static uint8_t bootstrap_data[SAN_BOOTSTRAP_L2_SIZE] __aligned(L2_SIZE); static size_t offset = 0; vm_offset_t addr; if (offset + L2_SIZE > sizeof(bootstrap_data)) { panic("%s: out of memory for the bootstrap shadow map L2 entries", __func__); } addr = (uintptr_t)&bootstrap_data[offset]; offset += L2_SIZE; return (addr); } /* * SAN L1 + L2 pages, maybe L3 entries later? */ static vm_offset_t __nosanitizeaddress pmap_san_enter_bootstrap_alloc_pages(int npages) { static uint8_t bootstrap_data[SAN_BOOTSTRAP_SIZE] __aligned(PAGE_SIZE); static size_t offset = 0; vm_offset_t addr; if (offset + (npages * PAGE_SIZE) > sizeof(bootstrap_data)) { panic("%s: out of memory for the bootstrap shadow map", __func__); } addr = (uintptr_t)&bootstrap_data[offset]; offset += (npages * PAGE_SIZE); return (addr); } static void __nosanitizeaddress pmap_san_enter_bootstrap(void) { vm_offset_t freemempos; /* L1, L2 */ freemempos = pmap_san_enter_bootstrap_alloc_pages(2); bs_state.freemempos = freemempos; bs_state.va = KASAN_MIN_ADDRESS; pmap_bootstrap_l1_table(&bs_state); pmap_san_early_l2 = bs_state.l2; } static vm_page_t pmap_san_enter_alloc_l3(void) { vm_page_t m; m = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) panic("%s: no memory to grow shadow map", __func__); return (m); } static vm_page_t pmap_san_enter_alloc_l2(void) { return (vm_page_alloc_noobj_contig(VM_ALLOC_WIRED | VM_ALLOC_ZERO, Ln_ENTRIES, 0, ~0ul, L2_SIZE, 0, VM_MEMATTR_DEFAULT)); } void __nosanitizeaddress pmap_san_enter(vm_offset_t va) { pd_entry_t *l1, *l2; pt_entry_t *l3; vm_page_t m; if (virtual_avail == 0) { vm_offset_t block; int slot; bool first; /* Temporary shadow map prior to pmap_bootstrap(). */ first = pmap_san_early_l2 == NULL; if (first) pmap_san_enter_bootstrap(); l2 = pmap_san_early_l2; slot = pmap_l2_index(va); if ((pmap_load(&l2[slot]) & ATTR_DESCR_VALID) == 0) { MPASS(first); block = pmap_san_enter_bootstrap_alloc_l2(); pmap_store(&l2[slot], PHYS_TO_PTE(pmap_early_vtophys(block)) | PMAP_SAN_PTE_BITS | L2_BLOCK); dmb(ishst); } return; } mtx_assert(&kernel_map->system_mtx, MA_OWNED); l1 = pmap_l1(kernel_pmap, va); MPASS(l1 != NULL); if ((pmap_load(l1) & ATTR_DESCR_VALID) == 0) { m = pmap_san_enter_alloc_l3(); pmap_store(l1, PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | L1_TABLE); } l2 = pmap_l1_to_l2(l1, va); if ((pmap_load(l2) & ATTR_DESCR_VALID) == 0) { m = pmap_san_enter_alloc_l2(); if (m != NULL) { pmap_store(l2, PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | PMAP_SAN_PTE_BITS | L2_BLOCK); } else { m = pmap_san_enter_alloc_l3(); pmap_store(l2, PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | L2_TABLE); } dmb(ishst); } if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) return; l3 = pmap_l2_to_l3(l2, va); if ((pmap_load(l3) & ATTR_DESCR_VALID) != 0) return; m = pmap_san_enter_alloc_l3(); pmap_store(l3, PHYS_TO_PTE(VM_PAGE_TO_PHYS(m)) | PMAP_SAN_PTE_BITS | L3_PAGE); dmb(ishst); } #endif /* KASAN */ /* * Track a range of the kernel's virtual address space that is contiguous * in various mapping attributes. */ struct pmap_kernel_map_range { vm_offset_t sva; pt_entry_t attrs; int l3pages; int l3contig; int l2blocks; int l1blocks; }; static void sysctl_kmaps_dump(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t eva) { const char *mode; int index; if (eva <= range->sva) return; index = range->attrs & ATTR_S1_IDX_MASK; switch (index) { case ATTR_S1_IDX(VM_MEMATTR_DEVICE_NP): mode = "DEV-NP"; break; case ATTR_S1_IDX(VM_MEMATTR_DEVICE): mode = "DEV"; break; case ATTR_S1_IDX(VM_MEMATTR_UNCACHEABLE): mode = "UC"; break; case ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK): mode = "WB"; break; case ATTR_S1_IDX(VM_MEMATTR_WRITE_THROUGH): mode = "WT"; break; default: printf( "%s: unknown memory type %x for range 0x%016lx-0x%016lx\n", __func__, index, range->sva, eva); mode = "??"; break; } sbuf_printf(sb, "0x%016lx-0x%016lx r%c%c%c%c %6s %d %d %d %d\n", range->sva, eva, (range->attrs & ATTR_S1_AP_RW_BIT) == ATTR_S1_AP_RW ? 'w' : '-', (range->attrs & ATTR_S1_PXN) != 0 ? '-' : 'x', (range->attrs & ATTR_S1_UXN) != 0 ? '-' : 'X', (range->attrs & ATTR_S1_AP(ATTR_S1_AP_USER)) != 0 ? 'u' : 's', mode, range->l1blocks, range->l2blocks, range->l3contig, range->l3pages); /* Reset to sentinel value. */ range->sva = 0xfffffffffffffffful; } /* * Determine whether the attributes specified by a page table entry match those * being tracked by the current range. */ static bool sysctl_kmaps_match(struct pmap_kernel_map_range *range, pt_entry_t attrs) { return (range->attrs == attrs); } static void sysctl_kmaps_reinit(struct pmap_kernel_map_range *range, vm_offset_t va, pt_entry_t attrs) { memset(range, 0, sizeof(*range)); range->sva = va; range->attrs = attrs; } /* Get the block/page attributes that correspond to the table attributes */ static pt_entry_t sysctl_kmaps_table_attrs(pd_entry_t table) { pt_entry_t attrs; attrs = 0; if ((table & TATTR_UXN_TABLE) != 0) attrs |= ATTR_S1_UXN; if ((table & TATTR_PXN_TABLE) != 0) attrs |= ATTR_S1_PXN; if ((table & TATTR_AP_TABLE_RO) != 0) attrs |= ATTR_S1_AP(ATTR_S1_AP_RO); return (attrs); } /* Read the block/page attributes we care about */ static pt_entry_t sysctl_kmaps_block_attrs(pt_entry_t block) { return (block & (ATTR_S1_AP_MASK | ATTR_S1_XN | ATTR_S1_IDX_MASK)); } /* * Given a leaf PTE, derive the mapping's attributes. If they do not match * those of the current run, dump the address range and its attributes, and * begin a new run. */ static void sysctl_kmaps_check(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t va, pd_entry_t l0e, pd_entry_t l1e, pd_entry_t l2e, pt_entry_t l3e) { pt_entry_t attrs; attrs = sysctl_kmaps_table_attrs(l0e); if ((l1e & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_BLOCK) { attrs |= sysctl_kmaps_block_attrs(l1e); goto done; } attrs |= sysctl_kmaps_table_attrs(l1e); if ((l2e & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_BLOCK) { attrs |= sysctl_kmaps_block_attrs(l2e); goto done; } attrs |= sysctl_kmaps_table_attrs(l2e); attrs |= sysctl_kmaps_block_attrs(l3e); done: if (range->sva > va || !sysctl_kmaps_match(range, attrs)) { sysctl_kmaps_dump(sb, range, va); sysctl_kmaps_reinit(range, va, attrs); } } static int sysctl_kmaps(SYSCTL_HANDLER_ARGS) { struct pmap_kernel_map_range range; struct sbuf sbuf, *sb; pd_entry_t l0e, *l1, l1e, *l2, l2e; pt_entry_t *l3, l3e; vm_offset_t sva; vm_paddr_t pa; int error, i, j, k, l; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sb = &sbuf; sbuf_new_for_sysctl(sb, NULL, PAGE_SIZE, req); /* Sentinel value. */ range.sva = 0xfffffffffffffffful; /* * Iterate over the kernel page tables without holding the kernel pmap * lock. Kernel page table pages are never freed, so at worst we will * observe inconsistencies in the output. */ for (sva = 0xffff000000000000ul, i = pmap_l0_index(sva); i < Ln_ENTRIES; i++) { if (i == pmap_l0_index(DMAP_MIN_ADDRESS)) sbuf_printf(sb, "\nDirect map:\n"); else if (i == pmap_l0_index(VM_MIN_KERNEL_ADDRESS)) sbuf_printf(sb, "\nKernel map:\n"); #ifdef KASAN else if (i == pmap_l0_index(KASAN_MIN_ADDRESS)) sbuf_printf(sb, "\nKASAN shadow map:\n"); #endif l0e = kernel_pmap->pm_l0[i]; if ((l0e & ATTR_DESCR_VALID) == 0) { sysctl_kmaps_dump(sb, &range, sva); sva += L0_SIZE; continue; } pa = PTE_TO_PHYS(l0e); l1 = (pd_entry_t *)PHYS_TO_DMAP(pa); for (j = pmap_l1_index(sva); j < Ln_ENTRIES; j++) { l1e = l1[j]; if ((l1e & ATTR_DESCR_VALID) == 0) { sysctl_kmaps_dump(sb, &range, sva); sva += L1_SIZE; continue; } if ((l1e & ATTR_DESCR_MASK) == L1_BLOCK) { PMAP_ASSERT_L1_BLOCKS_SUPPORTED; sysctl_kmaps_check(sb, &range, sva, l0e, l1e, 0, 0); range.l1blocks++; sva += L1_SIZE; continue; } pa = PTE_TO_PHYS(l1e); l2 = (pd_entry_t *)PHYS_TO_DMAP(pa); for (k = pmap_l2_index(sva); k < Ln_ENTRIES; k++) { l2e = l2[k]; if ((l2e & ATTR_DESCR_VALID) == 0) { sysctl_kmaps_dump(sb, &range, sva); sva += L2_SIZE; continue; } if ((l2e & ATTR_DESCR_MASK) == L2_BLOCK) { sysctl_kmaps_check(sb, &range, sva, l0e, l1e, l2e, 0); range.l2blocks++; sva += L2_SIZE; continue; } pa = PTE_TO_PHYS(l2e); l3 = (pt_entry_t *)PHYS_TO_DMAP(pa); for (l = pmap_l3_index(sva); l < Ln_ENTRIES; l++, sva += L3_SIZE) { l3e = l3[l]; if ((l3e & ATTR_DESCR_VALID) == 0) { sysctl_kmaps_dump(sb, &range, sva); continue; } sysctl_kmaps_check(sb, &range, sva, l0e, l1e, l2e, l3e); if ((l3e & ATTR_CONTIGUOUS) != 0) range.l3contig += l % 16 == 0 ? 1 : 0; else range.l3pages++; } } } } error = sbuf_finish(sb); sbuf_delete(sb); return (error); } SYSCTL_OID(_vm_pmap, OID_AUTO, kernel_maps, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE | CTLFLAG_SKIP, NULL, 0, sysctl_kmaps, "A", "Dump kernel address layout"); diff --git a/sys/vm/vm_page.c b/sys/vm/vm_page.c index 5d822d34ed7c..5e613ff4db4c 100644 --- a/sys/vm/vm_page.c +++ b/sys/vm/vm_page.c @@ -1,5661 +1,5661 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * 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. * 3. 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. * * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 */ /*- * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Resident memory management module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.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 #include #include #include #include #include #include struct vm_domain vm_dom[MAXMEMDOM]; DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; struct mtx_padalign __exclusive_cache_line vm_domainset_lock; /* The following fields are protected by the domainset lock. */ domainset_t __exclusive_cache_line vm_min_domains; domainset_t __exclusive_cache_line vm_severe_domains; static int vm_min_waiters; static int vm_severe_waiters; static int vm_pageproc_waiters; static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "VM page statistics"); static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, CTLFLAG_RD, &pqstate_commit_retries, "Number of failed per-page atomic queue state updates"); static COUNTER_U64_DEFINE_EARLY(queue_ops); SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, CTLFLAG_RD, &queue_ops, "Number of batched queue operations"); static COUNTER_U64_DEFINE_EARLY(queue_nops); SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, CTLFLAG_RD, &queue_nops, "Number of batched queue operations with no effects"); /* * bogus page -- for I/O to/from partially complete buffers, * or for paging into sparsely invalid regions. */ vm_page_t bogus_page; vm_page_t vm_page_array; long vm_page_array_size; long first_page; struct bitset *vm_page_dump; long vm_page_dump_pages; static TAILQ_HEAD(, vm_page) blacklist_head; static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); static uma_zone_t fakepg_zone; static void vm_page_alloc_check(vm_page_t m); static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_enqueue(vm_page_t m, uint8_t queue); static bool vm_page_free_prep(vm_page_t m); static void vm_page_free_toq(vm_page_t m); static void vm_page_init(void *dummy); static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred); static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred); static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, const uint16_t nflag); static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high); static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req); static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags); static void vm_page_zone_release(void *arg, void **store, int cnt); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); static void vm_page_init(void *dummy) { fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED); } /* * The cache page zone is initialized later since we need to be able to allocate * pages before UMA is fully initialized. */ static void vm_page_init_cache_zones(void *dummy __unused) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int cache, domain, maxcache, pool; maxcache = 0; TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache); maxcache *= mp_ncpus; for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (pool = 0; pool < VM_NFREEPOOL; pool++) { pgcache = &vmd->vmd_pgcache[pool]; pgcache->domain = domain; pgcache->pool = pool; pgcache->zone = uma_zcache_create("vm pgcache", PAGE_SIZE, NULL, NULL, NULL, NULL, vm_page_zone_import, vm_page_zone_release, pgcache, UMA_ZONE_VM); /* * Limit each pool's zone to 0.1% of the pages in the * domain. */ cache = maxcache != 0 ? maxcache : vmd->vmd_page_count / 1000; uma_zone_set_maxcache(pgcache->zone, cache); } } } SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ #if PAGE_SIZE == 32768 #ifdef CTASSERT CTASSERT(sizeof(u_long) >= 8); #endif #endif /* * vm_set_page_size: * * Sets the page size, perhaps based upon the memory * size. Must be called before any use of page-size * dependent functions. */ void vm_set_page_size(void) { if (vm_cnt.v_page_size == 0) vm_cnt.v_page_size = PAGE_SIZE; if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) panic("vm_set_page_size: page size not a power of two"); } /* * vm_page_blacklist_next: * * Find the next entry in the provided string of blacklist * addresses. Entries are separated by space, comma, or newline. * If an invalid integer is encountered then the rest of the * string is skipped. Updates the list pointer to the next * character, or NULL if the string is exhausted or invalid. */ static vm_paddr_t vm_page_blacklist_next(char **list, char *end) { vm_paddr_t bad; char *cp, *pos; if (list == NULL || *list == NULL) return (0); if (**list =='\0') { *list = NULL; return (0); } /* * If there's no end pointer then the buffer is coming from * the kenv and we know it's null-terminated. */ if (end == NULL) end = *list + strlen(*list); /* Ensure that strtoq() won't walk off the end */ if (*end != '\0') { if (*end == '\n' || *end == ' ' || *end == ',') *end = '\0'; else { printf("Blacklist not terminated, skipping\n"); *list = NULL; return (0); } } for (pos = *list; *pos != '\0'; pos = cp) { bad = strtoq(pos, &cp, 0); if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { if (bad == 0) { if (++cp < end) continue; else break; } } else break; if (*cp == '\0' || ++cp >= end) *list = NULL; else *list = cp; return (trunc_page(bad)); } printf("Garbage in RAM blacklist, skipping\n"); *list = NULL; return (0); } bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose) { struct vm_domain *vmd; vm_page_t m; bool found; m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) return (true); /* page does not exist, no failure */ vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); found = vm_phys_unfree_page(m); vm_domain_free_unlock(vmd); if (found) { vm_domain_freecnt_inc(vmd, -1); TAILQ_INSERT_TAIL(&blacklist_head, m, listq); if (verbose) printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); } return (found); } /* * vm_page_blacklist_check: * * Iterate through the provided string of blacklist addresses, pulling * each entry out of the physical allocator free list and putting it * onto a list for reporting via the vm.page_blacklist sysctl. */ static void vm_page_blacklist_check(char *list, char *end) { vm_paddr_t pa; char *next; next = list; while (next != NULL) { if ((pa = vm_page_blacklist_next(&next, end)) == 0) continue; vm_page_blacklist_add(pa, bootverbose); } } /* * vm_page_blacklist_load: * * Search for a special module named "ram_blacklist". It'll be a * plain text file provided by the user via the loader directive * of the same name. */ static void vm_page_blacklist_load(char **list, char **end) { void *mod; u_char *ptr; u_int len; mod = NULL; ptr = NULL; mod = preload_search_by_type("ram_blacklist"); if (mod != NULL) { ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); } *list = ptr; if (ptr != NULL) *end = ptr + len; else *end = NULL; return; } static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) { vm_page_t m; struct sbuf sbuf; int error, first; first = 1; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); TAILQ_FOREACH(m, &blacklist_head, listq) { sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", (uintmax_t)m->phys_addr); first = 0; } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Initialize a dummy page for use in scans of the specified paging queue. * In principle, this function only needs to set the flag PG_MARKER. * Nonetheless, it write busies the page as a safety precaution. */ void vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->a.flags = aflags; marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; marker->a.queue = queue; } static void vm_page_domain_init(int domain) { struct vm_domain *vmd; struct vm_pagequeue *pq; int i; vmd = VM_DOMAIN(domain); bzero(vmd, sizeof(*vmd)); *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = "vm inactive pagequeue"; *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = "vm active pagequeue"; *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = "vm laundry pagequeue"; *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = "vm unswappable pagequeue"; vmd->vmd_domain = domain; vmd->vmd_page_count = 0; vmd->vmd_free_count = 0; vmd->vmd_segs = 0; vmd->vmd_oom = FALSE; for (i = 0; i < PQ_COUNT; i++) { pq = &vmd->vmd_pagequeues[i]; TAILQ_INIT(&pq->pq_pl); mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", MTX_DEF | MTX_DUPOK); pq->pq_pdpages = 0; vm_page_init_marker(&vmd->vmd_markers[i], i, 0); } mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); /* * inacthead is used to provide FIFO ordering for LRU-bypassing * insertions. */ vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, &vmd->vmd_inacthead, plinks.q); /* * The clock pages are used to implement active queue scanning without * requeues. Scans start at clock[0], which is advanced after the scan * ends. When the two clock hands meet, they are reset and scanning * resumes from the head of the queue. */ vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[1], plinks.q); } /* * Initialize a physical page in preparation for adding it to the free * lists. */ void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) { m->object = NULL; m->ref_count = 0; m->busy_lock = VPB_FREED; m->flags = m->a.flags = 0; m->phys_addr = pa; m->a.queue = PQ_NONE; m->psind = 0; m->segind = segind; m->order = VM_NFREEORDER; m->pool = VM_FREEPOOL_DEFAULT; m->valid = m->dirty = 0; pmap_page_init(m); } #ifndef PMAP_HAS_PAGE_ARRAY static vm_paddr_t vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) { vm_paddr_t new_end; /* * Reserve an unmapped guard page to trap access to vm_page_array[-1]. * However, because this page is allocated from KVM, out-of-bounds * accesses using the direct map will not be trapped. */ *vaddr += PAGE_SIZE; /* * Allocate physical memory for the page structures, and map it. */ new_end = trunc_page(end - page_range * sizeof(struct vm_page)); vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array_size = page_range; return (new_end); } #endif /* * vm_page_startup: * * Initializes the resident memory module. Allocates physical memory for * bootstrapping UMA and some data structures that are used to manage * physical pages. Initializes these structures, and populates the free * page queues. */ vm_offset_t vm_page_startup(vm_offset_t vaddr) { struct vm_phys_seg *seg; struct vm_domain *vmd; vm_page_t m; char *list, *listend; vm_paddr_t end, high_avail, low_avail, new_end, size; vm_paddr_t page_range __unused; vm_paddr_t last_pa, pa, startp, endp; u_long pagecount; #if MINIDUMP_PAGE_TRACKING u_long vm_page_dump_size; #endif int biggestone, i, segind; #ifdef WITNESS vm_offset_t mapped; int witness_size; #endif #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) long ii; #endif vaddr = round_page(vaddr); vm_phys_early_startup(); biggestone = vm_phys_avail_largest(); end = phys_avail[biggestone+1]; /* * Initialize the page and queue locks. */ mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); for (i = 0; i < PA_LOCK_COUNT; i++) mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); for (i = 0; i < vm_ndomains; i++) vm_page_domain_init(i); new_end = end; #ifdef WITNESS witness_size = round_page(witness_startup_count()); new_end -= witness_size; mapped = pmap_map(&vaddr, new_end, new_end + witness_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, witness_size); witness_startup((void *)mapped); #endif #if MINIDUMP_PAGE_TRACKING /* * Allocate a bitmap to indicate that a random physical page * needs to be included in a minidump. * * The amd64 port needs this to indicate which direct map pages * need to be dumped, via calls to dump_add_page()/dump_drop_page(). * * However, i386 still needs this workspace internally within the * minidump code. In theory, they are not needed on i386, but are * included should the sf_buf code decide to use them. */ last_pa = 0; vm_page_dump_pages = 0; for (i = 0; dump_avail[i + 1] != 0; i += 2) { vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - dump_avail[i] / PAGE_SIZE; if (dump_avail[i + 1] > last_pa) last_pa = dump_avail[i + 1]; } vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); new_end -= vm_page_dump_size; vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)vm_page_dump, vm_page_dump_size); #else (void)last_pa; #endif #if defined(__aarch64__) || defined(__amd64__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include the UMA bootstrap pages, witness pages and vm_page_dump * in a crash dump. When pmap_map() uses the direct map, they are * not automatically included. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; #ifdef __amd64__ /* * Request that the physical pages underlying the message buffer be * included in a crash dump. Since the message buffer is accessed * through the direct map, they are not automatically included. */ pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); last_pa = pa + round_page(msgbufsize); while (pa < last_pa) { dump_add_page(pa); pa += PAGE_SIZE; } #endif /* * Compute the number of pages of memory that will be available for * use, taking into account the overhead of a page structure per page. * In other words, solve * "available physical memory" - round_page(page_range * * sizeof(struct vm_page)) = page_range * PAGE_SIZE * for page_range. */ low_avail = phys_avail[0]; high_avail = phys_avail[1]; for (i = 0; i < vm_phys_nsegs; i++) { if (vm_phys_segs[i].start < low_avail) low_avail = vm_phys_segs[i].start; if (vm_phys_segs[i].end > high_avail) high_avail = vm_phys_segs[i].end; } /* Skip the first chunk. It is already accounted for. */ for (i = 2; phys_avail[i + 1] != 0; i += 2) { if (phys_avail[i] < low_avail) low_avail = phys_avail[i]; if (phys_avail[i + 1] > high_avail) high_avail = phys_avail[i + 1]; } first_page = low_avail / PAGE_SIZE; #ifdef VM_PHYSSEG_SPARSE size = 0; for (i = 0; i < vm_phys_nsegs; i++) size += vm_phys_segs[i].end - vm_phys_segs[i].start; for (i = 0; phys_avail[i + 1] != 0; i += 2) size += phys_avail[i + 1] - phys_avail[i]; #elif defined(VM_PHYSSEG_DENSE) size = high_avail - low_avail; #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif #ifdef PMAP_HAS_PAGE_ARRAY pmap_page_array_startup(size / PAGE_SIZE); biggestone = vm_phys_avail_largest(); end = new_end = phys_avail[biggestone + 1]; #else #ifdef VM_PHYSSEG_DENSE /* * In the VM_PHYSSEG_DENSE case, the number of pages can account for * the overhead of a page structure per page only if vm_page_array is * allocated from the last physical memory chunk. Otherwise, we must * allocate page structures representing the physical memory * underlying vm_page_array, even though they will not be used. */ if (new_end != high_avail) page_range = size / PAGE_SIZE; else #endif { page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); /* * If the partial bytes remaining are large enough for * a page (PAGE_SIZE) without a corresponding * 'struct vm_page', then new_end will contain an * extra page after subtracting the length of the VM * page array. Compensate by subtracting an extra * page from new_end. */ if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { if (new_end == high_avail) high_avail -= PAGE_SIZE; new_end -= PAGE_SIZE; } } end = new_end; new_end = vm_page_array_alloc(&vaddr, end, page_range); #endif #if VM_NRESERVLEVEL > 0 /* * Allocate physical memory for the reservation management system's * data structures, and map it. */ new_end = vm_reserv_startup(&vaddr, new_end); #endif #if defined(__aarch64__) || defined(__amd64__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include vm_page_array and vm_reserv_array in a crash dump. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; /* * Add physical memory segments corresponding to the available * physical pages. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) if (vm_phys_avail_size(i) != 0) vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); /* * Initialize the physical memory allocator. */ vm_phys_init(); /* * Initialize the page structures and add every available page to the * physical memory allocator's free lists. */ #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) for (ii = 0; ii < vm_page_array_size; ii++) { m = &vm_page_array[ii]; vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); m->flags = PG_FICTITIOUS; } #endif vm_cnt.v_page_count = 0; for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; for (m = seg->first_page, pa = seg->start; pa < seg->end; m++, pa += PAGE_SIZE) vm_page_init_page(m, pa, segind); /* * Add the segment's pages that are covered by one of * phys_avail's ranges to the free lists. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) { if (seg->end <= phys_avail[i] || seg->start >= phys_avail[i + 1]) continue; startp = MAX(seg->start, phys_avail[i]); endp = MIN(seg->end, phys_avail[i + 1]); pagecount = (u_long)atop(endp - startp); if (pagecount == 0) continue; m = seg->first_page + atop(startp - seg->start); vmd = VM_DOMAIN(seg->domain); vm_domain_free_lock(vmd); vm_phys_enqueue_contig(m, pagecount); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, pagecount); vm_cnt.v_page_count += (u_int)pagecount; vmd->vmd_page_count += (u_int)pagecount; vmd->vmd_segs |= 1UL << segind; } } /* * Remove blacklisted pages from the physical memory allocator. */ TAILQ_INIT(&blacklist_head); vm_page_blacklist_load(&list, &listend); vm_page_blacklist_check(list, listend); list = kern_getenv("vm.blacklist"); vm_page_blacklist_check(list, NULL); freeenv(list); #if VM_NRESERVLEVEL > 0 /* * Initialize the reservation management system. */ vm_reserv_init(); #endif return (vaddr); } void vm_page_reference(vm_page_t m) { vm_page_aflag_set(m, PGA_REFERENCED); } /* * vm_page_trybusy * * Helper routine for grab functions to trylock busy. * * Returns true on success and false on failure. */ static bool vm_page_trybusy(vm_page_t m, int allocflags) { if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) return (vm_page_trysbusy(m)); else return (vm_page_tryxbusy(m)); } /* * vm_page_tryacquire * * Helper routine for grab functions to trylock busy and wire. * * Returns true on success and false on failure. */ static inline bool vm_page_tryacquire(vm_page_t m, int allocflags) { bool locked; locked = vm_page_trybusy(m, allocflags); if (locked && (allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); return (locked); } /* * vm_page_busy_acquire: * * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop * and drop the object lock if necessary. */ bool vm_page_busy_acquire(vm_page_t m, int allocflags) { vm_object_t obj; bool locked; /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = atomic_load_ptr(&m->object); for (;;) { if (vm_page_tryacquire(m, allocflags)) return (true); if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (false); if (obj != NULL) locked = VM_OBJECT_WOWNED(obj); else locked = false; MPASS(locked || vm_page_wired(m)); if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, locked) && locked) VM_OBJECT_WLOCK(obj); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (false); KASSERT(m->object == obj || m->object == NULL, ("vm_page_busy_acquire: page %p does not belong to %p", m, obj)); } } /* * vm_page_busy_downgrade: * * Downgrade an exclusive busy page into a single shared busy page. */ void vm_page_busy_downgrade(vm_page_t m) { u_int x; vm_page_assert_xbusied(m); x = vm_page_busy_fetch(m); for (;;) { if (atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_SHARERS_WORD(1))) break; } if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* * * vm_page_busy_tryupgrade: * * Attempt to upgrade a single shared busy into an exclusive busy. */ int vm_page_busy_tryupgrade(vm_page_t m) { u_int ce, x; vm_page_assert_sbusied(m); x = vm_page_busy_fetch(m); ce = VPB_CURTHREAD_EXCLUSIVE; for (;;) { if (VPB_SHARERS(x) > 1) return (0); KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_busy_tryupgrade: invalid lock state")); if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, ce | (x & VPB_BIT_WAITERS))) continue; return (1); } } /* * vm_page_sbusied: * * Return a positive value if the page is shared busied, 0 otherwise. */ int vm_page_sbusied(vm_page_t m) { u_int x; x = vm_page_busy_fetch(m); return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); } /* * vm_page_sunbusy: * * Shared unbusy a page. */ void vm_page_sunbusy(vm_page_t m) { u_int x; vm_page_assert_sbusied(m); x = vm_page_busy_fetch(m); for (;;) { KASSERT(x != VPB_FREED, ("vm_page_sunbusy: Unlocking freed page.")); if (VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_sunbusy: invalid lock state")); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) == 0) break; wakeup(m); break; } } /* * vm_page_busy_sleep: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of the busying mechanism. * * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function * will not sleep if the page is shared-busy. * * The object lock must be held on entry. * * Returns true if it slept and dropped the object lock, or false * if there was no sleep and the lock is still held. */ bool vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) { vm_object_t obj; obj = m->object; VM_OBJECT_ASSERT_LOCKED(obj); return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, true)); } /* * vm_page_busy_sleep_unlocked: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of busying mechanism. * * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function * will not sleep if the page is shared-busy. * * The object lock must not be held on entry. The operation will * return if the page changes identity. */ void vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags) { VM_OBJECT_ASSERT_UNLOCKED(obj); (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); } /* * _vm_page_busy_sleep: * * Internal busy sleep function. Verifies the page identity and * lockstate against parameters. Returns true if it sleeps and * false otherwise. * * allocflags uses VM_ALLOC_* flags to specify the lock required. * * If locked is true the lock will be dropped for any true returns * and held for any false returns. */ static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked) { bool xsleep; u_int x; /* * If the object is busy we must wait for that to drain to zero * before trying the page again. */ if (obj != NULL && vm_object_busied(obj)) { if (locked) VM_OBJECT_DROP(obj); vm_object_busy_wait(obj, wmesg); return (true); } if (!vm_page_busied(m)) return (false); xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; sleepq_lock(m); x = vm_page_busy_fetch(m); do { /* * If the page changes objects or becomes unlocked we can * simply return. */ if (x == VPB_UNBUSIED || (xsleep && (x & VPB_BIT_SHARED) != 0) || m->object != obj || m->pindex != pindex) { sleepq_release(m); return (false); } if ((x & VPB_BIT_WAITERS) != 0) break; } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); if (locked) VM_OBJECT_DROP(obj); DROP_GIANT(); sleepq_add(m, NULL, wmesg, 0, 0); sleepq_wait(m, PVM); PICKUP_GIANT(); return (true); } /* * vm_page_trysbusy: * * Try to shared busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_trysbusy(vm_page_t m) { vm_object_t obj; u_int x; obj = m->object; x = vm_page_busy_fetch(m); for (;;) { if ((x & VPB_BIT_SHARED) == 0) return (0); /* * Reduce the window for transient busies that will trigger * false negatives in vm_page_ps_test(). */ if (obj != NULL && vm_object_busied(obj)) return (0); if (atomic_fcmpset_acq_int(&m->busy_lock, &x, x + VPB_ONE_SHARER)) break; } /* Refetch the object now that we're guaranteed that it is stable. */ obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_sunbusy(m); return (0); } return (1); } /* * vm_page_tryxbusy: * * Try to exclusive busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_tryxbusy(vm_page_t m) { vm_object_t obj; if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, VPB_CURTHREAD_EXCLUSIVE) == 0) return (0); obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_xunbusy(m); return (0); } return (1); } static void vm_page_xunbusy_hard_tail(vm_page_t m) { atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); /* Wake the waiter. */ wakeup(m); } /* * vm_page_xunbusy_hard: * * Called when unbusy has failed because there is a waiter. */ void vm_page_xunbusy_hard(vm_page_t m) { vm_page_assert_xbusied(m); vm_page_xunbusy_hard_tail(m); } void vm_page_xunbusy_hard_unchecked(vm_page_t m) { vm_page_assert_xbusied_unchecked(m); vm_page_xunbusy_hard_tail(m); } static void vm_page_busy_free(vm_page_t m) { u_int x; atomic_thread_fence_rel(); x = atomic_swap_int(&m->busy_lock, VPB_FREED); if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* * vm_page_unhold_pages: * * Unhold each of the pages that is referenced by the given array. */ void vm_page_unhold_pages(vm_page_t *ma, int count) { for (; count != 0; count--) { vm_page_unwire(*ma, PQ_ACTIVE); ma++; } } vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa) { vm_page_t m; #ifdef VM_PHYSSEG_SPARSE m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) m = vm_phys_fictitious_to_vm_page(pa); return (m); #elif defined(VM_PHYSSEG_DENSE) long pi; pi = atop(pa); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { m = &vm_page_array[pi - first_page]; return (m); } return (vm_phys_fictitious_to_vm_page(pa)); #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif } /* * vm_page_getfake: * * Create a fictitious page with the specified physical address and * memory attribute. The memory attribute is the only the machine- * dependent aspect of a fictitious page that must be initialized. */ vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) { vm_page_t m; m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); vm_page_initfake(m, paddr, memattr); return (m); } void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { if ((m->flags & PG_FICTITIOUS) != 0) { /* * The page's memattr might have changed since the * previous initialization. Update the pmap to the * new memattr. */ goto memattr; } m->phys_addr = paddr; m->a.queue = PQ_NONE; /* Fictitious pages don't use "segind". */ m->flags = PG_FICTITIOUS; /* Fictitious pages don't use "order" or "pool". */ m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; /* Fictitious pages are unevictable. */ m->ref_count = 1; pmap_page_init(m); memattr: pmap_page_set_memattr(m, memattr); } /* * vm_page_putfake: * * Release a fictitious page. */ void vm_page_putfake(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_putfake: bad page %p", m)); vm_page_assert_xbusied(m); vm_page_busy_free(m); uma_zfree(fakepg_zone, m); } /* * vm_page_updatefake: * * Update the given fictitious page to the specified physical address and * memory attribute. */ void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_updatefake: bad page %p", m)); m->phys_addr = paddr; pmap_page_set_memattr(m, memattr); } /* * vm_page_free: * * Free a page. */ void vm_page_free(vm_page_t m) { m->flags &= ~PG_ZERO; vm_page_free_toq(m); } /* * vm_page_free_zero: * * Free a page to the zerod-pages queue */ void vm_page_free_zero(vm_page_t m) { m->flags |= PG_ZERO; vm_page_free_toq(m); } /* * Unbusy and handle the page queueing for a page from a getpages request that * was optionally read ahead or behind. */ void vm_page_readahead_finish(vm_page_t m) { /* We shouldn't put invalid pages on queues. */ KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); /* * Since the page is not the actually needed one, whether it should * be activated or deactivated is not obvious. Empirical results * have shown that deactivating the page is usually the best choice, * unless the page is wanted by another thread. */ if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_xunbusy_unchecked(m); } /* * Destroy the identity of an invalid page and free it if possible. * This is intended to be used when reading a page from backing store fails. */ void vm_page_free_invalid(vm_page_t m) { KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); KASSERT(m->object != NULL, ("page %p has no object", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); /* * We may be attempting to free the page as part of the handling for an * I/O error, in which case the page was xbusied by a different thread. */ vm_page_xbusy_claim(m); /* * If someone has wired this page while the object lock * was not held, then the thread that unwires is responsible * for freeing the page. Otherwise just free the page now. * The wire count of this unmapped page cannot change while * we have the page xbusy and the page's object wlocked. */ if (vm_page_remove(m)) vm_page_free(m); } /* * vm_page_dirty_KBI: [ internal use only ] * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). * * This function should only be called by vm_page_dirty(). */ void vm_page_dirty_KBI(vm_page_t m) { /* Refer to this operation by its public name. */ KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); m->dirty = VM_PAGE_BITS_ALL; } /* * vm_page_insert: [ internal use only ] * * Inserts the given mem entry into the object and object list. * * The object must be locked. */ int vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) { vm_page_t mpred; VM_OBJECT_ASSERT_WLOCKED(object); mpred = vm_radix_lookup_le(&object->rtree, pindex); return (vm_page_insert_after(m, object, pindex, mpred)); } /* * vm_page_insert_after: * * Inserts the page "m" into the specified object at offset "pindex". * * The page "mpred" must immediately precede the offset "pindex" within * the specified object. * * The object must be locked. */ static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(m->object == NULL, ("vm_page_insert_after: page already inserted")); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); msucc = TAILQ_NEXT(mpred, listq); } else msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, ("vm_page_insert_after: msucc doesn't succeed pindex")); /* * Record the object/offset pair in this page. */ m->object = object; m->pindex = pindex; m->ref_count |= VPRC_OBJREF; /* * Now link into the object's ordered list of backed pages. */ if (vm_radix_insert(&object->rtree, m)) { m->object = NULL; m->pindex = 0; m->ref_count &= ~VPRC_OBJREF; return (1); } vm_page_insert_radixdone(m, object, mpred); vm_pager_page_inserted(object, m); return (0); } /* * vm_page_insert_radixdone: * * Complete page "m" insertion into the specified object after the * radix trie hooking. * * The page "mpred" must precede the offset "m->pindex" within the * specified object. * * The object must be locked. */ static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object != NULL && m->object == object, ("vm_page_insert_radixdone: page %p has inconsistent object", m)); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("vm_page_insert_radixdone: page %p is missing object ref", m)); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_radixdone: object doesn't contain mpred")); KASSERT(mpred->pindex < m->pindex, ("vm_page_insert_radixdone: mpred doesn't precede pindex")); } if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); /* * Show that the object has one more resident page. */ object->resident_page_count++; /* * Hold the vnode until the last page is released. */ if (object->resident_page_count == 1 && object->type == OBJT_VNODE) vhold(object->handle); /* * Since we are inserting a new and possibly dirty page, * update the object's generation count. */ if (pmap_page_is_write_mapped(m)) vm_object_set_writeable_dirty(object); } /* * Do the work to remove a page from its object. The caller is responsible for * updating the page's fields to reflect this removal. */ static void vm_page_object_remove(vm_page_t m) { vm_object_t object; vm_page_t mrem __diagused; vm_page_assert_xbusied(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("page %p is missing its object ref", m)); /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0) vm_pager_page_unswapped(m); vm_pager_page_removed(object, m); m->object = NULL; mrem = vm_radix_remove(&object->rtree, m->pindex); KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); /* * Now remove from the object's list of backed pages. */ TAILQ_REMOVE(&object->memq, m, listq); /* * And show that the object has one fewer resident page. */ object->resident_page_count--; /* * The vnode may now be recycled. */ if (object->resident_page_count == 0 && object->type == OBJT_VNODE) vdrop(object->handle); } /* * vm_page_remove: * * Removes the specified page from its containing object, but does not * invalidate any backing storage. Returns true if the object's reference * was the last reference to the page, and false otherwise. * * The object must be locked and the page must be exclusively busied. * The exclusive busy will be released on return. If this is not the * final ref and the caller does not hold a wire reference it may not * continue to access the page. */ bool vm_page_remove(vm_page_t m) { bool dropped; dropped = vm_page_remove_xbusy(m); vm_page_xunbusy(m); return (dropped); } /* * vm_page_remove_xbusy * * Removes the page but leaves the xbusy held. Returns true if this * removed the final ref and false otherwise. */ bool vm_page_remove_xbusy(vm_page_t m) { vm_page_object_remove(m); return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); } /* * vm_page_lookup: * * Returns the page associated with the object/offset * pair specified; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_lookup(vm_object_t object, vm_pindex_t pindex) { VM_OBJECT_ASSERT_LOCKED(object); return (vm_radix_lookup(&object->rtree, pindex)); } /* * vm_page_lookup_unlocked: * * Returns the page associated with the object/offset pair specified; * if none is found, NULL is returned. The page may be no longer be * present in the object at the time that this function returns. Only * useful for opportunistic checks such as inmem(). */ vm_page_t vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) { return (vm_radix_lookup_unlocked(&object->rtree, pindex)); } /* * vm_page_relookup: * * Returns a page that must already have been busied by * the caller. Used for bogus page replacement. */ vm_page_t vm_page_relookup(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; m = vm_radix_lookup_unlocked(&object->rtree, pindex); KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && m->object == object && m->pindex == pindex, ("vm_page_relookup: Invalid page %p", m)); return (m); } /* * This should only be used by lockless functions for releasing transient * incorrect acquires. The page may have been freed after we acquired a * busy lock. In this case busy_lock == VPB_FREED and we have nothing * further to do. */ static void vm_page_busy_release(vm_page_t m) { u_int x; x = vm_page_busy_fetch(m); for (;;) { if (x == VPB_FREED) break; if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & VPB_BIT_SHARED) != 0 || (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, ("vm_page_busy_release: %p xbusy not owned.", m)); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); break; } } /* * vm_page_find_least: * * Returns the page associated with the object with least pindex * greater than or equal to the parameter pindex, or NULL. * * The object must be locked. */ vm_page_t vm_page_find_least(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; VM_OBJECT_ASSERT_LOCKED(object); if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) m = vm_radix_lookup_ge(&object->rtree, pindex); return (m); } /* * Returns the given page's successor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_next(vm_page_t m) { vm_page_t next; VM_OBJECT_ASSERT_LOCKED(m->object); if ((next = TAILQ_NEXT(m, listq)) != NULL) { MPASS(next->object == m->object); if (next->pindex != m->pindex + 1) next = NULL; } return (next); } /* * Returns the given page's predecessor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_prev(vm_page_t m) { vm_page_t prev; VM_OBJECT_ASSERT_LOCKED(m->object); if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { MPASS(prev->object == m->object); if (prev->pindex != m->pindex - 1) prev = NULL; } return (prev); } /* * Uses the page mnew as a replacement for an existing page at index * pindex which must be already present in the object. * * Both pages must be exclusively busied on enter. The old page is * unbusied on exit. * * A return value of true means mold is now free. If this is not the * final ref and the caller does not hold a wire reference it may not * continue to access the page. */ static bool vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_t mret __diagused; bool dropped; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_assert_xbusied(mold); KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, ("vm_page_replace: page %p already in object", mnew)); /* * This function mostly follows vm_page_insert() and * vm_page_remove() without the radix, object count and vnode * dance. Double check such functions for more comments. */ mnew->object = object; mnew->pindex = pindex; atomic_set_int(&mnew->ref_count, VPRC_OBJREF); mret = vm_radix_replace(&object->rtree, mnew); KASSERT(mret == mold, ("invalid page replacement, mold=%p, mret=%p", mold, mret)); KASSERT((mold->oflags & VPO_UNMANAGED) == (mnew->oflags & VPO_UNMANAGED), ("vm_page_replace: mismatched VPO_UNMANAGED")); /* Keep the resident page list in sorted order. */ TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); TAILQ_REMOVE(&object->memq, mold, listq); mold->object = NULL; /* * The object's resident_page_count does not change because we have * swapped one page for another, but the generation count should * change if the page is dirty. */ if (pmap_page_is_write_mapped(mnew)) vm_object_set_writeable_dirty(object); dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; vm_page_xunbusy(mold); return (dropped); } void vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_assert_xbusied(mnew); if (vm_page_replace_hold(mnew, object, pindex, mold)) vm_page_free(mold); } /* * vm_page_rename: * * Move the given memory entry from its * current object to the specified target object/offset. * * Note: swap associated with the page must be invalidated by the move. We * have to do this for several reasons: (1) we aren't freeing the * page, (2) we are dirtying the page, (3) the VM system is probably * moving the page from object A to B, and will then later move * the backing store from A to B and we can't have a conflict. * * Note: we *always* dirty the page. It is necessary both for the * fact that we moved it, and because we may be invalidating * swap. * * The objects must be locked. */ int vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) { vm_page_t mpred; vm_pindex_t opidx; VM_OBJECT_ASSERT_WLOCKED(new_object); KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); KASSERT(mpred == NULL || mpred->pindex != new_pindex, ("vm_page_rename: pindex already renamed")); /* * Create a custom version of vm_page_insert() which does not depend * by m_prev and can cheat on the implementation aspects of the * function. */ opidx = m->pindex; m->pindex = new_pindex; if (vm_radix_insert(&new_object->rtree, m)) { m->pindex = opidx; return (1); } /* * The operation cannot fail anymore. The removal must happen before * the listq iterator is tainted. */ m->pindex = opidx; vm_page_object_remove(m); /* Return back to the new pindex to complete vm_page_insert(). */ m->pindex = new_pindex; m->object = new_object; vm_page_insert_radixdone(m, new_object, mpred); vm_page_dirty(m); vm_pager_page_inserted(new_object, m); return (0); } /* * vm_page_alloc: * * Allocate and return a page that is associated with the specified * object and offset pair. By default, this page is exclusive busied. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) { return (vm_page_alloc_after(object, pindex, req, vm_radix_lookup_le(&object->rtree, pindex))); } vm_page_t vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req) { return (vm_page_alloc_domain_after(object, pindex, domain, req, vm_radix_lookup_le(&object->rtree, pindex))); } /* * Allocate a page in the specified object with the given page index. To * optimize insertion of the page into the object, the caller must also specifiy * the resident page in the object with largest index smaller than the given * page index, or NULL if no such page exists. */ vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, vm_page_t mpred) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_domain_after(object, pindex, domain, req, mpred); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } /* * Returns true if the number of free pages exceeds the minimum * for the request class and false otherwise. */ static int _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) { u_int limit, old, new; if (req_class == VM_ALLOC_INTERRUPT) limit = 0; else if (req_class == VM_ALLOC_SYSTEM) limit = vmd->vmd_interrupt_free_min; else limit = vmd->vmd_free_reserved; /* * Attempt to reserve the pages. Fail if we're below the limit. */ limit += npages; old = vmd->vmd_free_count; do { if (old < limit) return (0); new = old - npages; } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); /* Wake the page daemon if we've crossed the threshold. */ if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) pagedaemon_wakeup(vmd->vmd_domain); /* Only update bitsets on transitions. */ if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) vm_domain_set(vmd); return (1); } int vm_domain_allocate(struct vm_domain *vmd, int req, int npages) { int req_class; /* * The page daemon is allowed to dig deeper into the free page list. */ req_class = req & VM_ALLOC_CLASS_MASK; if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; return (_vm_domain_allocate(vmd, req_class, npages)); } vm_page_t vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, int req, vm_page_t mpred) { struct vm_domain *vmd; vm_page_t m; int flags; #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) KASSERT((req & ~VPA_FLAGS) == 0, ("invalid request %#x", req)); KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("invalid request %#x", req)); KASSERT(mpred == NULL || mpred->pindex < pindex, ("mpred %p doesn't precede pindex 0x%jx", mpred, (uintmax_t)pindex)); VM_OBJECT_ASSERT_WLOCKED(object); flags = 0; m = NULL; if (!vm_pager_can_alloc_page(object, pindex)) return (NULL); again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the page from a reservation? */ if (vm_object_reserv(object) && (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != NULL) { goto found; } #endif vmd = VM_DOMAIN(domain); if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, M_NOWAIT | M_NOVM); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { /* * If not, allocate it from the free page queues. */ vm_domain_free_lock(vmd); m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { /* * Not allocatable, give up. */ if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } /* * At this point we had better have found a good page. */ found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ flags |= m->flags & PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; m->flags = flags; m->a.flags = 0; m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; else if ((req & VM_ALLOC_SBUSY) != 0) m->busy_lock = VPB_SHARERS_WORD(1); else m->busy_lock = VPB_UNBUSIED; if (req & VM_ALLOC_WIRED) { vm_wire_add(1); m->ref_count = 1; } m->a.act_count = 0; if (vm_page_insert_after(m, object, pindex, mpred)) { if (req & VM_ALLOC_WIRED) { vm_wire_sub(1); m->ref_count = 0; } KASSERT(m->object == NULL, ("page %p has object", m)); m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } /* Ignore device objects; the pager sets "memattr" for them. */ if (object->memattr != VM_MEMATTR_DEFAULT && (object->flags & OBJ_FICTITIOUS) == 0) pmap_page_set_memattr(m, object->memattr); return (m); } /* * vm_page_alloc_contig: * * Allocate a contiguous set of physical pages of the given size "npages" * from the free lists. All of the physical pages must be at or above * the given physical address "low" and below the given physical address * "high". The given value "alignment" determines the alignment of the * first physical page in the set. If the given value "boundary" is * non-zero, then the set of physical pages cannot cross any physical * address boundary that is a multiple of that value. Both "alignment" * and "boundary" must be a power of two. * * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, * then the memory attribute setting for the physical pages is configured * to the object's memory attribute setting. Otherwise, the memory * attribute setting for the physical pages is configured to "memattr", * overriding the object's memory attribute setting. However, if the * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the * memory attribute setting for the physical pages cannot be configured * to VM_MEMATTR_DEFAULT. * * The specified object may not contain fictitious pages. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_contig_domain(object, pindex, domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } static vm_page_t vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domain *vmd; vm_page_t m_ret; /* * Can we allocate the pages without the number of free pages falling * below the lower bound for the allocation class? */ vmd = VM_DOMAIN(domain); if (!vm_domain_allocate(vmd, req, npages)) return (NULL); /* * Try to allocate the pages from the free page queues. */ vm_domain_free_lock(vmd); m_ret = vm_phys_alloc_contig(domain, npages, low, high, alignment, boundary); vm_domain_free_unlock(vmd); if (m_ret != NULL) return (m_ret); #if VM_NRESERVLEVEL > 0 /* * Try to break a reservation to allocate the pages. */ if ((req & VM_ALLOC_NORECLAIM) == 0) { m_ret = vm_reserv_reclaim_contig(domain, npages, low, high, alignment, boundary); if (m_ret != NULL) return (m_ret); } #endif vm_domain_freecnt_inc(vmd, npages); return (NULL); } vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vm_page_t m, m_ret, mpred; u_int busy_lock, flags, oflags; #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) KASSERT((req & ~VPAC_FLAGS) == 0, ("invalid request %#x", req)); KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("invalid request %#x", req)); KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), ("invalid request %#x", req)); VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_FICTITIOUS) == 0, ("vm_page_alloc_contig: object %p has fictitious pages", object)); KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); mpred = vm_radix_lookup_le(&object->rtree, pindex); KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc_contig: pindex already allocated")); for (;;) { #if VM_NRESERVLEVEL > 0 /* * Can we allocate the pages from a reservation? */ if (vm_object_reserv(object) && (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, mpred, npages, low, high, alignment, boundary)) != NULL) { break; } #endif if ((m_ret = vm_page_find_contig_domain(domain, req, npages, low, high, alignment, boundary)) != NULL) break; if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) return (NULL); } for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) busy_lock = VPB_CURTHREAD_EXCLUSIVE; else if ((req & VM_ALLOC_SBUSY) != 0) busy_lock = VPB_SHARERS_WORD(1); else busy_lock = VPB_UNBUSIED; if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); if (object->memattr != VM_MEMATTR_DEFAULT && memattr == VM_MEMATTR_DEFAULT) memattr = object->memattr; for (m = m_ret; m < &m_ret[npages]; m++) { m->a.flags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = busy_lock; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->a.act_count = 0; m->oflags = oflags; if (vm_page_insert_after(m, object, pindex, mpred)) { if ((req & VM_ALLOC_WIRED) != 0) vm_wire_sub(npages); KASSERT(m->object == NULL, ("page %p has object", m)); mpred = m; for (m = m_ret; m < &m_ret[npages]; m++) { if (m <= mpred && (req & VM_ALLOC_WIRED) != 0) m->ref_count = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); } if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } mpred = m; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); pindex++; } return (m_ret); } /* * Allocate a physical page that is not intended to be inserted into a VM * object. If the "freelist" parameter is not equal to VM_NFREELIST, then only * pages from the specified vm_phys freelist will be returned. */ static __always_inline vm_page_t _vm_page_alloc_noobj_domain(int domain, const int freelist, int req) { struct vm_domain *vmd; vm_page_t m; int flags; #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) KASSERT((req & ~VPAN_FLAGS) == 0, ("invalid request %#x", req)); flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0; vmd = VM_DOMAIN(domain); again: if (freelist == VM_NFREELIST && vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, M_NOWAIT | M_NOVM); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { vm_domain_free_lock(vmd); if (freelist == VM_NFREELIST) m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); else m = vm_phys_alloc_freelist_pages(domain, freelist, VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (freelist == VM_NFREELIST && vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { if (vm_domain_alloc_fail(vmd, NULL, req)) goto again; return (NULL); } found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Consumers should not rely on a useful default pindex value. */ m->pindex = 0xdeadc0dedeadc0de; m->flags = (m->flags & PG_ZERO) | flags; m->a.flags = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; if ((req & VM_ALLOC_WIRED) != 0) { vm_wire_add(1); m->ref_count = 1; } if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); return (m); } vm_page_t vm_page_alloc_freelist(int freelist, int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_freelist_domain(domain, freelist, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_freelist_domain(int domain, int freelist, int req) { KASSERT(freelist >= 0 && freelist < VM_NFREELIST, ("%s: invalid freelist %d", __func__, freelist)); return (_vm_page_alloc_noobj_domain(domain, freelist, req)); } vm_page_t vm_page_alloc_noobj(int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_noobj_domain(domain, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_noobj_domain(int domain, int req) { return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req)); } vm_page_t vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vm_page_t m, m_ret; u_int flags; #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) KASSERT((req & ~VPANC_FLAGS) == 0, ("invalid request %#x", req)); KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), ("invalid request %#x", req)); KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("invalid request %#x", req)); KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); while ((m_ret = vm_page_find_contig_domain(domain, req, npages, low, high, alignment, boundary)) == NULL) { if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) return (NULL); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); /* * Consumers should not rely on a useful default pindex value. */ m->pindex = 0xdeadc0dedeadc0de; m->a.flags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = VPB_UNBUSIED; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->a.act_count = 0; m->oflags = VPO_UNMANAGED; /* * Zero the page before updating any mappings since the page is * not yet shared with any devices which might require the * non-default memory attribute. pmap_page_set_memattr() * flushes data caches before returning. */ if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); } return (m_ret); } /* * Check a page that has been freshly dequeued from a freelist. */ static void vm_page_alloc_check(vm_page_t m) { KASSERT(m->object == NULL, ("page %p has object", m)); KASSERT(m->a.queue == PQ_NONE && (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has unexpected queue %d, flags %#x", m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); KASSERT(m->ref_count == 0, ("page %p has references", m)); KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); KASSERT(m->dirty == 0, ("page %p is dirty", m)); KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has unexpected memattr %d", m, pmap_page_get_memattr(m))); KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); pmap_vm_page_alloc_check(m); } static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); /* * The page daemon should avoid creating extra memory pressure since its * main purpose is to replenish the store of free pages. */ if (vmd->vmd_severeset || curproc == pageproc || !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) return (0); domain = vmd->vmd_domain; vm_domain_free_lock(vmd); i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, (vm_page_t *)store); vm_domain_free_unlock(vmd); if (cnt != i) vm_domain_freecnt_inc(vmd, cnt - i); return (i); } static void vm_page_zone_release(void *arg, void **store, int cnt) { struct vm_domain *vmd; struct vm_pgcache *pgcache; vm_page_t m; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); vm_domain_free_lock(vmd); for (i = 0; i < cnt; i++) { m = (vm_page_t)store[i]; vm_phys_free_pages(m, 0); } vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } #define VPSC_ANY 0 /* No restrictions. */ #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ #define VPSC_NOSUPER 2 /* Skip superpages. */ /* * vm_page_scan_contig: * * Scan vm_page_array[] between the specified entries "m_start" and * "m_end" for a run of contiguous physical pages that satisfy the * specified conditions, and return the lowest page in the run. The * specified "alignment" determines the alignment of the lowest physical * page in the run. If the specified "boundary" is non-zero, then the * run of physical pages cannot span a physical address that is a * multiple of "boundary". * * "m_end" is never dereferenced, so it need not point to a vm_page * structure within vm_page_array[]. * * "npages" must be greater than zero. "m_start" and "m_end" must not * span a hole (or discontiguity) in the physical address space. Both * "alignment" and "boundary" must be a power of two. */ -vm_page_t +static vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options) { vm_object_t object; vm_paddr_t pa; vm_page_t m, m_run; #if VM_NRESERVLEVEL > 0 int level; #endif int m_inc, order, run_ext, run_len; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); m_run = NULL; run_len = 0; for (m = m_start; m < m_end && run_len < npages; m += m_inc) { KASSERT((m->flags & PG_MARKER) == 0, ("page %p is PG_MARKER", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, ("fictitious page %p has invalid ref count", m)); /* * If the current page would be the start of a run, check its * physical address against the end, alignment, and boundary * conditions. If it doesn't satisfy these conditions, either * terminate the scan or advance to the next page that * satisfies the failed condition. */ if (run_len == 0) { KASSERT(m_run == NULL, ("m_run != NULL")); if (m + npages > m_end) break; pa = VM_PAGE_TO_PHYS(m); if (!vm_addr_align_ok(pa, alignment)) { m_inc = atop(roundup2(pa, alignment) - pa); continue; } if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { m_inc = atop(roundup2(pa, boundary) - pa); continue; } } else KASSERT(m_run != NULL, ("m_run == NULL")); retry: m_inc = 1; if (vm_page_wired(m)) run_ext = 0; #if VM_NRESERVLEVEL > 0 else if ((level = vm_reserv_level(m)) >= 0 && (options & VPSC_NORESERV) != 0) { run_ext = 0; /* Advance to the end of the reservation. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); } #endif else if ((object = atomic_load_ptr(&m->object)) != NULL) { /* * The page is considered eligible for relocation if * and only if it could be laundered or reclaimed by * the page daemon. */ VM_OBJECT_RLOCK(object); if (object != m->object) { VM_OBJECT_RUNLOCK(object); goto retry; } /* Don't care: PG_NODUMP, PG_ZERO. */ if ((object->flags & OBJ_SWAP) == 0 && object->type != OBJT_VNODE) { run_ext = 0; #if VM_NRESERVLEVEL > 0 } else if ((options & VPSC_NOSUPER) != 0 && (level = vm_reserv_level_iffullpop(m)) >= 0) { run_ext = 0; /* Advance to the end of the superpage. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); #endif } else if (object->memattr == VM_MEMATTR_DEFAULT && vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { /* * The page is allocated but eligible for * relocation. Extend the current run by one * page. */ KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: PGA_NOSYNC. */ run_ext = 1; } else run_ext = 0; VM_OBJECT_RUNLOCK(object); #if VM_NRESERVLEVEL > 0 } else if (level >= 0) { /* * The page is reserved but not yet allocated. In * other words, it is still free. Extend the current * run by one page. */ run_ext = 1; #endif } else if ((order = m->order) < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it is the * first page in a power-of-two-sized run of * contiguous free pages. Add these pages to the end * of the current run, and jump ahead. */ run_ext = 1 << order; m_inc = 1 << order; } else { /* * Skip the page for one of the following reasons: (1) * It is enqueued in the physical memory allocator's * free page queues. However, it is not the first * page in a run of contiguous free pages. (This case * rarely occurs because the scan is performed in * ascending order.) (2) It is not reserved, and it is * transitioning from free to allocated. (Conversely, * the transition from allocated to free for managed * pages is blocked by the page busy lock.) (3) It is * allocated but not contained by an object and not * wired, e.g., allocated by Xen's balloon driver. */ run_ext = 0; } /* * Extend or reset the current run of pages. */ if (run_ext > 0) { if (run_len == 0) m_run = m; run_len += run_ext; } else { if (run_len > 0) { m_run = NULL; run_len = 0; } } } if (run_len >= npages) return (m_run); return (NULL); } /* * vm_page_reclaim_run: * * Try to relocate each of the allocated virtual pages within the * specified run of physical pages to a new physical address. Free the * physical pages underlying the relocated virtual pages. A virtual page * is relocatable if and only if it could be laundered or reclaimed by * the page daemon. Whenever possible, a virtual page is relocated to a * physical address above "high". * * Returns 0 if every physical page within the run was already free or * just freed by a successful relocation. Otherwise, returns a non-zero * value indicating why the last attempt to relocate a virtual page was * unsuccessful. * * "req_class" must be an allocation class. */ static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high) { struct vm_domain *vmd; struct spglist free; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_end, m_new; int error, order, req; KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, ("req_class is not an allocation class")); SLIST_INIT(&free); error = 0; m = m_run; m_end = m_run + npages; for (; error == 0 && m < m_end; m++) { KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, ("page %p is PG_FICTITIOUS or PG_MARKER", m)); /* * Racily check for wirings. Races are handled once the object * lock is held and the page is unmapped. */ if (vm_page_wired(m)) error = EBUSY; else if ((object = atomic_load_ptr(&m->object)) != NULL) { /* * The page is relocated if and only if it could be * laundered or reclaimed by the page daemon. */ VM_OBJECT_WLOCK(object); /* Don't care: PG_NODUMP, PG_ZERO. */ if (m->object != object || ((object->flags & OBJ_SWAP) == 0 && object->type != OBJT_VNODE)) error = EINVAL; else if (object->memattr != VM_MEMATTR_DEFAULT) error = EINVAL; else if (vm_page_queue(m) != PQ_NONE && vm_page_tryxbusy(m) != 0) { if (vm_page_wired(m)) { vm_page_xunbusy(m); error = EBUSY; goto unlock; } KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT(m->oflags == 0, ("page %p has unexpected oflags", m)); /* Don't care: PGA_NOSYNC. */ if (!vm_page_none_valid(m)) { /* * First, try to allocate a new page * that is above "high". Failing * that, try to allocate a new page * that is below "m_run". Allocate * the new page between the end of * "m_run" and "high" only as a last * resort. */ req = req_class; if ((m->flags & PG_NODUMP) != 0) req |= VM_ALLOC_NODUMP; if (trunc_page(high) != ~(vm_paddr_t)PAGE_MASK) { m_new = vm_page_alloc_noobj_contig( req, 1, round_page(high), ~(vm_paddr_t)0, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } else m_new = NULL; if (m_new == NULL) { pa = VM_PAGE_TO_PHYS(m_run); m_new = vm_page_alloc_noobj_contig( req, 1, 0, pa - 1, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { pa += ptoa(npages); m_new = vm_page_alloc_noobj_contig( req, 1, pa, high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { vm_page_xunbusy(m); error = ENOMEM; goto unlock; } /* * Unmap the page and check for new * wirings that may have been acquired * through a pmap lookup. */ if (object->ref_count != 0 && !vm_page_try_remove_all(m)) { vm_page_xunbusy(m); vm_page_free(m_new); error = EBUSY; goto unlock; } /* * Replace "m" with the new page. For * vm_page_replace(), "m" must be busy * and dequeued. Finally, change "m" * as if vm_page_free() was called. */ m_new->a.flags = m->a.flags & ~PGA_QUEUE_STATE_MASK; KASSERT(m_new->oflags == VPO_UNMANAGED, ("page %p is managed", m_new)); m_new->oflags = 0; pmap_copy_page(m, m_new); m_new->valid = m->valid; m_new->dirty = m->dirty; m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_replace_hold(m_new, object, m->pindex, m) && vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); /* * The new page must be deactivated * before the object is unlocked. */ vm_page_deactivate(m_new); } else { m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); KASSERT(m->dirty == 0, ("page %p is dirty", m)); } } else error = EBUSY; unlock: VM_OBJECT_WUNLOCK(object); } else { MPASS(vm_page_domain(m) == domain); vmd = VM_DOMAIN(domain); vm_domain_free_lock(vmd); order = m->order; if (order < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it * is the first page in a power-of-two-sized * run of contiguous free pages. Jump ahead * to the last page within that run, and * continue from there. */ m += (1 << order) - 1; } #if VM_NRESERVLEVEL > 0 else if (vm_reserv_is_page_free(m)) order = 0; #endif vm_domain_free_unlock(vmd); if (order == VM_NFREEORDER) error = EINVAL; } } if ((m = SLIST_FIRST(&free)) != NULL) { int cnt; vmd = VM_DOMAIN(domain); cnt = 0; vm_domain_free_lock(vmd); do { MPASS(vm_page_domain(m) == domain); SLIST_REMOVE_HEAD(&free, plinks.s.ss); vm_phys_free_pages(m, 0); cnt++; } while ((m = SLIST_FIRST(&free)) != NULL); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } return (error); } #define NRUNS 16 #define RUN_INDEX(count, nruns) ((count) % (nruns)) #define MIN_RECLAIM 8 /* * vm_page_reclaim_contig: * * Reclaim allocated, contiguous physical memory satisfying the specified * conditions by relocating the virtual pages using that physical memory. * Returns true if reclamation is successful and false otherwise. Since * relocation requires the allocation of physical pages, reclamation may * fail due to a shortage of free pages. When reclamation fails, callers * are expected to perform vm_wait() before retrying a failed allocation * operation, e.g., vm_page_alloc_contig(). * * The caller must always specify an allocation class through "req". * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * The optional allocation flags are ignored. * * "npages" must be greater than zero. Both "alignment" and "boundary" * must be a power of two. */ bool vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, int desired_runs) { struct vm_domain *vmd; - vm_paddr_t curr_low; - vm_page_t m_run, _m_runs[NRUNS], *m_runs; + vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; u_long count, minalign, reclaimed; - int error, i, min_reclaim, nruns, options, req_class; + int error, i, min_reclaim, nruns, options, req_class, segind; bool ret; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); ret = false; /* * If the caller wants to reclaim multiple runs, try to allocate * space to store the runs. If that fails, fall back to the old * behavior of just reclaiming MIN_RECLAIM pages. */ if (desired_runs > 1) m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), M_TEMP, M_NOWAIT); else m_runs = NULL; if (m_runs == NULL) { m_runs = _m_runs; nruns = NRUNS; } else { nruns = NRUNS + desired_runs - 1; } min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); /* * The caller will attempt an allocation after some runs have been * reclaimed and added to the vm_phys buddy lists. Due to limitations * of vm_phys_alloc_contig(), round up the requested length to the next * power of two or maximum chunk size, and ensure that each run is * suitably aligned. */ minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); npages = roundup2(npages, minalign); if (alignment < ptoa(minalign)) alignment = ptoa(minalign); /* * The page daemon is allowed to dig deeper into the free page list. */ req_class = req & VM_ALLOC_CLASS_MASK; if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; /* * Return if the number of free pages cannot satisfy the requested * allocation. */ vmd = VM_DOMAIN(domain); count = vmd->vmd_free_count; if (count < npages + vmd->vmd_free_reserved || (count < npages + vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || (count < npages && req_class == VM_ALLOC_INTERRUPT)) goto done; /* * Scan up to three times, relaxing the restrictions ("options") on * the reclamation of reservations and superpages each time. */ for (options = VPSC_NORESERV;;) { /* * Find the highest runs that satisfy the given constraints * and restrictions, and record them in "m_runs". */ - curr_low = low; count = 0; - for (;;) { - m_run = vm_phys_scan_contig(domain, npages, curr_low, - high, alignment, boundary, options); - if (m_run == NULL) - break; - curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); - m_runs[RUN_INDEX(count, nruns)] = m_run; - count++; + segind = vm_phys_lookup_segind(low); + while ((segind = vm_phys_find_range(bounds, segind, domain, + npages, low, high)) != -1) { + while ((m_run = vm_page_scan_contig(npages, bounds[0], + bounds[1], alignment, boundary, options))) { + bounds[0] = m_run + npages; + m_runs[RUN_INDEX(count, nruns)] = m_run; + count++; + } + segind++; } /* * Reclaim the highest runs in LIFO (descending) order until * the number of reclaimed pages, "reclaimed", is at least * "min_reclaim". Reset "reclaimed" each time because each * reclamation is idempotent, and runs will (likely) recur * from one scan to the next as restrictions are relaxed. */ reclaimed = 0; for (i = 0; count > 0 && i < nruns; i++) { count--; m_run = m_runs[RUN_INDEX(count, nruns)]; error = vm_page_reclaim_run(req_class, domain, npages, m_run, high); if (error == 0) { reclaimed += npages; if (reclaimed >= min_reclaim) { ret = true; goto done; } } } /* * Either relax the restrictions on the next scan or return if * the last scan had no restrictions. */ if (options == VPSC_NORESERV) options = VPSC_NOSUPER; else if (options == VPSC_NOSUPER) options = VPSC_ANY; else if (options == VPSC_ANY) { ret = reclaimed != 0; goto done; } } done: if (m_runs != _m_runs) free(m_runs, M_TEMP); return (ret); } bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high, alignment, boundary, 1)); } bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domainset_iter di; int domain; bool ret; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { ret = vm_page_reclaim_contig_domain(domain, req, npages, low, high, alignment, boundary); if (ret) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (ret); } /* * Set the domain in the appropriate page level domainset. */ void vm_domain_set(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (!vmd->vmd_minset && vm_paging_min(vmd)) { vmd->vmd_minset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); } if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { vmd->vmd_severeset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); } mtx_unlock(&vm_domainset_lock); } /* * Clear the domain from the appropriate page level domainset. */ void vm_domain_clear(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_minset && !vm_paging_min(vmd)) { vmd->vmd_minset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); if (vm_min_waiters != 0) { vm_min_waiters = 0; wakeup(&vm_min_domains); } } if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { vmd->vmd_severeset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); if (vm_severe_waiters != 0) { vm_severe_waiters = 0; wakeup(&vm_severe_domains); } } /* * If pageout daemon needs pages, then tell it that there are * some free. */ if (vmd->vmd_pageout_pages_needed && vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { wakeup(&vmd->vmd_pageout_pages_needed); vmd->vmd_pageout_pages_needed = 0; } /* See comments in vm_wait_doms(). */ if (vm_pageproc_waiters) { vm_pageproc_waiters = 0; wakeup(&vm_pageproc_waiters); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the min threshold globally. */ void vm_wait_min(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_min()) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the severe threshold globally. */ void vm_wait_severe(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_severe()) { vm_severe_waiters++; msleep(&vm_severe_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } u_int vm_wait_count(void) { return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); } int vm_wait_doms(const domainset_t *wdoms, int mflags) { int error; error = 0; /* * We use racey wakeup synchronization to avoid expensive global * locking for the pageproc when sleeping with a non-specific vm_wait. * To handle this, we only sleep for one tick in this instance. It * is expected that most allocations for the pageproc will come from * kmem or vm_page_grab* which will use the more specific and * race-free vm_wait_domain(). */ if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); vm_pageproc_waiters++; error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP | mflags, "pageprocwait", 1); } else { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(wdoms)) { if (pageproc == NULL) panic("vm_wait in early boot"); vm_min_waiters++; error = msleep(&vm_min_domains, &vm_domainset_lock, PVM | PDROP | mflags, "vmwait", 0); } else mtx_unlock(&vm_domainset_lock); } return (error); } /* * vm_wait_domain: * * Sleep until free pages are available for allocation. * - Called in various places after failed memory allocations. */ void vm_wait_domain(int domain) { struct vm_domain *vmd; domainset_t wdom; vmd = VM_DOMAIN(domain); vm_domain_free_assert_unlocked(vmd); if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { vmd->vmd_pageout_pages_needed = 1; msleep(&vmd->vmd_pageout_pages_needed, &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); } else mtx_unlock(&vm_domainset_lock); } else { DOMAINSET_ZERO(&wdom); DOMAINSET_SET(vmd->vmd_domain, &wdom); vm_wait_doms(&wdom, 0); } } static int vm_wait_flags(vm_object_t obj, int mflags) { struct domainset *d; d = NULL; /* * Carefully fetch pointers only once: the struct domainset * itself is ummutable but the pointer might change. */ if (obj != NULL) d = obj->domain.dr_policy; if (d == NULL) d = curthread->td_domain.dr_policy; return (vm_wait_doms(&d->ds_mask, mflags)); } /* * vm_wait: * * Sleep until free pages are available for allocation in the * affinity domains of the obj. If obj is NULL, the domain set * for the calling thread is used. * Called in various places after failed memory allocations. */ void vm_wait(vm_object_t obj) { (void)vm_wait_flags(obj, 0); } int vm_wait_intr(vm_object_t obj) { return (vm_wait_flags(obj, PCATCH)); } /* * vm_domain_alloc_fail: * * Called when a page allocation function fails. Informs the * pagedaemon and performs the requested wait. Requires the * domain_free and object lock on entry. Returns with the * object lock held and free lock released. Returns an error when * retry is necessary. * */ static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) { vm_domain_free_assert_unlocked(vmd); atomic_add_int(&vmd->vmd_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_wait_domain(vmd->vmd_domain); if (object != NULL) VM_OBJECT_WLOCK(object); if (req & VM_ALLOC_WAITOK) return (EAGAIN); } return (0); } /* * vm_waitpfault: * * Sleep until free pages are available for allocation. * - Called only in vm_fault so that processes page faulting * can be easily tracked. * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing * processes will be able to grab memory first. Do not change * this balance without careful testing first. */ void vm_waitpfault(struct domainset *dset, int timo) { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(&dset->ds_mask)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, "pfault", timo); } else mtx_unlock(&vm_domainset_lock); } static struct vm_pagequeue * _vm_page_pagequeue(vm_page_t m, uint8_t queue) { return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); } #ifdef INVARIANTS static struct vm_pagequeue * vm_page_pagequeue(vm_page_t m) { return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); } #endif static __always_inline bool vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { vm_page_astate_t tmp; tmp = *old; do { if (__predict_true(vm_page_astate_fcmpset(m, old, new))) return (true); counter_u64_add(pqstate_commit_retries, 1); } while (old->_bits == tmp._bits); return (false); } /* * Do the work of committing a queue state update that moves the page out of * its current queue. */ static bool _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { vm_page_t next; vm_pagequeue_assert_locked(pq); KASSERT(vm_page_pagequeue(m) == pq, ("%s: queue %p does not match page %p", __func__, pq, m)); KASSERT(old->queue != PQ_NONE && new.queue != old->queue, ("%s: invalid queue indices %d %d", __func__, old->queue, new.queue)); /* * Once the queue index of the page changes there is nothing * synchronizing with further updates to the page's physical * queue state. Therefore we must speculatively remove the page * from the queue now and be prepared to roll back if the queue * state update fails. If the page is not physically enqueued then * we just update its queue index. */ if ((old->flags & PGA_ENQUEUED) != 0) { new.flags &= ~PGA_ENQUEUED; next = TAILQ_NEXT(m, plinks.q); TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_pagequeue_cnt_dec(pq); if (!vm_page_pqstate_fcmpset(m, old, new)) { if (next == NULL) TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); else TAILQ_INSERT_BEFORE(next, m, plinks.q); vm_pagequeue_cnt_inc(pq); return (false); } else { return (true); } } else { return (vm_page_pqstate_fcmpset(m, old, new)); } } static bool vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { struct vm_pagequeue *pq; vm_page_astate_t as; bool ret; pq = _vm_page_pagequeue(m, old->queue); /* * The queue field and PGA_ENQUEUED flag are stable only so long as the * corresponding page queue lock is held. */ vm_pagequeue_lock(pq); as = vm_page_astate_load(m); if (__predict_false(as._bits != old->_bits)) { *old = as; ret = false; } else { ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); } vm_pagequeue_unlock(pq); return (ret); } /* * Commit a queue state update that enqueues or requeues a page. */ static bool _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { struct vm_domain *vmd; vm_pagequeue_assert_locked(pq); KASSERT(old->queue != PQ_NONE && new.queue == old->queue, ("%s: invalid queue indices %d %d", __func__, old->queue, new.queue)); new.flags |= PGA_ENQUEUED; if (!vm_page_pqstate_fcmpset(m, old, new)) return (false); if ((old->flags & PGA_ENQUEUED) != 0) TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); else vm_pagequeue_cnt_inc(pq); /* * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if * both flags are set in close succession, only PGA_REQUEUE_HEAD will be * applied, even if it was set first. */ if ((old->flags & PGA_REQUEUE_HEAD) != 0) { vmd = vm_pagequeue_domain(m); KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], ("%s: invalid page queue for page %p", __func__, m)); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); } else { TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); } return (true); } /* * Commit a queue state update that encodes a request for a deferred queue * operation. */ static bool vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { KASSERT(old->queue == new.queue || new.queue != PQ_NONE, ("%s: invalid state, queue %d flags %x", __func__, new.queue, new.flags)); if (old->_bits != new._bits && !vm_page_pqstate_fcmpset(m, old, new)) return (false); vm_page_pqbatch_submit(m, new.queue); return (true); } /* * A generic queue state update function. This handles more cases than the * specialized functions above. */ bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { if (old->_bits == new._bits) return (true); if (old->queue != PQ_NONE && new.queue != old->queue) { if (!vm_page_pqstate_commit_dequeue(m, old, new)) return (false); if (new.queue != PQ_NONE) vm_page_pqbatch_submit(m, new.queue); } else { if (!vm_page_pqstate_fcmpset(m, old, new)) return (false); if (new.queue != PQ_NONE && ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) vm_page_pqbatch_submit(m, new.queue); } return (true); } /* * Apply deferred queue state updates to a page. */ static inline void vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) { vm_page_astate_t new, old; CRITICAL_ASSERT(curthread); vm_pagequeue_assert_locked(pq); KASSERT(queue < PQ_COUNT, ("%s: invalid queue index %d", __func__, queue)); KASSERT(pq == _vm_page_pagequeue(m, queue), ("%s: page %p does not belong to queue %p", __func__, m, pq)); for (old = vm_page_astate_load(m);;) { if (__predict_false(old.queue != queue || (old.flags & PGA_QUEUE_OP_MASK) == 0)) { counter_u64_add(queue_nops, 1); break; } KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is unmanaged", __func__, m)); new = old; if ((old.flags & PGA_DEQUEUE) != 0) { new.flags &= ~PGA_QUEUE_OP_MASK; new.queue = PQ_NONE; if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, m, &old, new))) { counter_u64_add(queue_ops, 1); break; } } else { new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); if (__predict_true(_vm_page_pqstate_commit_requeue(pq, m, &old, new))) { counter_u64_add(queue_ops, 1); break; } } } } static void vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, uint8_t queue) { int i; for (i = 0; i < bq->bq_cnt; i++) vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); vm_batchqueue_init(bq); } /* * vm_page_pqbatch_submit: [ internal use only ] * * Enqueue a page in the specified page queue's batched work queue. * The caller must have encoded the requested operation in the page * structure's a.flags field. */ void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) { struct vm_batchqueue *bq; struct vm_pagequeue *pq; int domain, slots_remaining; KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); domain = vm_page_domain(m); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); slots_remaining = vm_batchqueue_insert(bq, m); if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { /* keep building the bq */ critical_exit(); return; } else if (slots_remaining > 0 ) { /* Try to process the bq if we can get the lock */ pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; if (vm_pagequeue_trylock(pq)) { vm_pqbatch_process(pq, bq, queue); vm_pagequeue_unlock(pq); } critical_exit(); return; } critical_exit(); /* if we make it here, the bq is full so wait for the lock */ pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); vm_pqbatch_process(pq, bq, queue); vm_pqbatch_process_page(pq, m, queue); vm_pagequeue_unlock(pq); critical_exit(); } /* * vm_page_pqbatch_drain: [ internal use only ] * * Force all per-CPU page queue batch queues to be drained. This is * intended for use in severe memory shortages, to ensure that pages * do not remain stuck in the batch queues. */ void vm_page_pqbatch_drain(void) { struct thread *td; struct vm_domain *vmd; struct vm_pagequeue *pq; int cpu, domain, queue; td = curthread; CPU_FOREACH(cpu) { thread_lock(td); sched_bind(td, cpu); thread_unlock(td); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (queue = 0; queue < PQ_COUNT; queue++) { pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); vm_pqbatch_process(pq, DPCPU_PTR(pqbatch[domain][queue]), queue); critical_exit(); vm_pagequeue_unlock(pq); } } } thread_lock(td); sched_unbind(td); thread_unlock(td); } /* * vm_page_dequeue_deferred: [ internal use only ] * * Request removal of the given page from its current page * queue. Physical removal from the queue may be deferred * indefinitely. */ void vm_page_dequeue_deferred(vm_page_t m) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue == PQ_NONE) { KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); break; } new = old; new.flags |= PGA_DEQUEUE; } while (!vm_page_pqstate_commit_request(m, &old, new)); } /* * vm_page_dequeue: * * Remove the page from whichever page queue it's in, if any, before * returning. */ void vm_page_dequeue(vm_page_t m) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue == PQ_NONE) { KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); break; } new = old; new.flags &= ~PGA_QUEUE_OP_MASK; new.queue = PQ_NONE; } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); } /* * Schedule the given page for insertion into the specified page queue. * Physical insertion of the page may be deferred indefinitely. */ static void vm_page_enqueue(vm_page_t m, uint8_t queue) { KASSERT(m->a.queue == PQ_NONE && (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p is already enqueued", __func__, m)); KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); m->a.queue = queue; if ((m->a.flags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, queue); } /* * vm_page_free_prep: * * Prepares the given page to be put on the free list, * disassociating it from any VM object. The caller may return * the page to the free list only if this function returns true. * * The object, if it exists, must be locked, and then the page must * be xbusy. Otherwise the page must be not busied. A managed * page must be unmapped. */ static bool vm_page_free_prep(vm_page_t m) { /* * Synchronize with threads that have dropped a reference to this * page. */ atomic_thread_fence_acq(); #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { uint64_t *p; int i; p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", m, i, (uintmax_t)*p)); } #endif if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(!pmap_page_is_mapped(m), ("vm_page_free_prep: freeing mapped page %p", m)); KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, ("vm_page_free_prep: mapping flags set in page %p", m)); } else { KASSERT(m->a.queue == PQ_NONE, ("vm_page_free_prep: unmanaged page %p is queued", m)); } VM_CNT_INC(v_tfree); if (m->object != NULL) { KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == ((m->object->flags & OBJ_UNMANAGED) != 0), ("vm_page_free_prep: managed flag mismatch for page %p", m)); vm_page_assert_xbusied(m); /* * The object reference can be released without an atomic * operation. */ KASSERT((m->flags & PG_FICTITIOUS) != 0 || m->ref_count == VPRC_OBJREF, ("vm_page_free_prep: page %p has unexpected ref_count %u", m, m->ref_count)); vm_page_object_remove(m); m->ref_count -= VPRC_OBJREF; } else vm_page_assert_unbusied(m); vm_page_busy_free(m); /* * If fictitious remove object association and * return. */ if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->ref_count == 1, ("fictitious page %p is referenced", m)); KASSERT(m->a.queue == PQ_NONE, ("fictitious page %p is queued", m)); return (false); } /* * Pages need not be dequeued before they are returned to the physical * memory allocator, but they must at least be marked for a deferred * dequeue. */ if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_dequeue_deferred(m); m->valid = 0; vm_page_undirty(m); if (m->ref_count != 0) panic("vm_page_free_prep: page %p has references", m); /* * Restore the default memory attribute to the page. */ if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); #if VM_NRESERVLEVEL > 0 /* * Determine whether the page belongs to a reservation. If the page was * allocated from a per-CPU cache, it cannot belong to a reservation, so * as an optimization, we avoid the check in that case. */ if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) return (false); #endif return (true); } /* * vm_page_free_toq: * * Returns the given page to the free list, disassociating it * from any VM object. * * The object must be locked. The page must be exclusively busied if it * belongs to an object. */ static void vm_page_free_toq(vm_page_t m) { struct vm_domain *vmd; uma_zone_t zone; if (!vm_page_free_prep(m)) return; vmd = vm_pagequeue_domain(m); zone = vmd->vmd_pgcache[m->pool].zone; if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { uma_zfree(zone, m); return; } vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); } /* * vm_page_free_pages_toq: * * Returns a list of pages to the free list, disassociating it * from any VM object. In other words, this is equivalent to * calling vm_page_free_toq() for each page of a list of VM objects. */ void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) { vm_page_t m; int count; if (SLIST_EMPTY(free)) return; count = 0; while ((m = SLIST_FIRST(free)) != NULL) { count++; SLIST_REMOVE_HEAD(free, plinks.s.ss); vm_page_free_toq(m); } if (update_wire_count) vm_wire_sub(count); } /* * Mark this page as wired down. For managed pages, this prevents reclamation * by the page daemon, or when the containing object, if any, is destroyed. */ void vm_page_wire(vm_page_t m) { u_int old; #ifdef INVARIANTS if (m->object != NULL && !vm_page_busied(m) && !vm_object_busied(m->object)) VM_OBJECT_ASSERT_LOCKED(m->object); #endif KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(m->ref_count) >= 1, ("vm_page_wire: fictitious page %p has zero wirings", m)); old = atomic_fetchadd_int(&m->ref_count, 1); KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, ("vm_page_wire: counter overflow for page %p", m)); if (VPRC_WIRE_COUNT(old) == 0) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_wire_add(1); } } /* * Attempt to wire a mapped page following a pmap lookup of that page. * This may fail if a thread is concurrently tearing down mappings of the page. * The transient failure is acceptable because it translates to the * failure of the caller pmap_extract_and_hold(), which should be then * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). */ bool vm_page_wire_mapped(vm_page_t m) { u_int old; old = m->ref_count; do { KASSERT(old > 0, ("vm_page_wire_mapped: wiring unreferenced page %p", m)); if ((old & VPRC_BLOCKED) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); if (VPRC_WIRE_COUNT(old) == 0) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_wire_add(1); } return (true); } /* * Release a wiring reference to a managed page. If the page still belongs to * an object, update its position in the page queues to reflect the reference. * If the wiring was the last reference to the page, free the page. */ static void vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) { u_int old; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is unmanaged", __func__, m)); /* * Update LRU state before releasing the wiring reference. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = m->ref_count; do { KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); if (old > VPRC_OBJREF + 1) { /* * The page has at least one other wiring reference. An * earlier iteration of this loop may have called * vm_page_release_toq() and cleared PGA_DEQUEUE, so * re-set it if necessary. */ if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); } else if (old == VPRC_OBJREF + 1) { /* * This is the last wiring. Clear PGA_DEQUEUE and * update the page's queue state to reflect the * reference. If the page does not belong to an object * (i.e., the VPRC_OBJREF bit is clear), we only need to * clear leftover queue state. */ vm_page_release_toq(m, nqueue, noreuse); } else if (old == 1) { vm_page_aflag_clear(m, PGA_DEQUEUE); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* * Release one wiring of the specified page, potentially allowing it to be * paged out. * * Only managed pages belonging to an object can be paged out. If the number * of wirings transitions to zero and the page is eligible for page out, then * the page is added to the specified paging queue. If the released wiring * represented the last reference to the page, the page is freed. */ void vm_page_unwire(vm_page_t m, uint8_t nqueue) { KASSERT(nqueue < PQ_COUNT, ("vm_page_unwire: invalid queue %u request for page %p", nqueue, m)); if ((m->oflags & VPO_UNMANAGED) != 0) { if (vm_page_unwire_noq(m) && m->ref_count == 0) vm_page_free(m); return; } vm_page_unwire_managed(m, nqueue, false); } /* * Unwire a page without (re-)inserting it into a page queue. It is up * to the caller to enqueue, requeue, or free the page as appropriate. * In most cases involving managed pages, vm_page_unwire() should be used * instead. */ bool vm_page_unwire_noq(vm_page_t m) { u_int old; old = vm_page_drop(m, 1); KASSERT(VPRC_WIRE_COUNT(old) != 0, ("%s: counter underflow for page %p", __func__, m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, ("%s: missing ref on fictitious page %p", __func__, m)); if (VPRC_WIRE_COUNT(old) > 1) return (false); if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_clear(m, PGA_DEQUEUE); vm_wire_sub(1); return (true); } /* * Ensure that the page ends up in the specified page queue. If the page is * active or being moved to the active queue, ensure that its act_count is * at least ACT_INIT but do not otherwise mess with it. */ static __always_inline void vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) { vm_page_astate_t old, new; KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, ("%s: invalid flags %x", __func__, nflag)); if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; old = vm_page_astate_load(m); do { if ((old.flags & PGA_DEQUEUE) != 0) break; new = old; new.flags &= ~PGA_QUEUE_OP_MASK; if (nqueue == PQ_ACTIVE) new.act_count = max(old.act_count, ACT_INIT); if (old.queue == nqueue) { /* * There is no need to requeue pages already in the * active queue. */ if (nqueue != PQ_ACTIVE || (old.flags & PGA_ENQUEUED) == 0) new.flags |= nflag; } else { new.flags |= nflag; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); } /* * Put the specified page on the active list (if appropriate). */ void vm_page_activate(vm_page_t m) { vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); } /* * Move the specified page to the tail of the inactive queue, or requeue * the page if it is already in the inactive queue. */ void vm_page_deactivate(vm_page_t m) { vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); } void vm_page_deactivate_noreuse(vm_page_t m) { vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); } /* * Put a page in the laundry, or requeue it if it is already there. */ void vm_page_launder(vm_page_t m) { vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); } /* * Put a page in the PQ_UNSWAPPABLE holding queue. */ void vm_page_unswappable(vm_page_t m) { VM_OBJECT_ASSERT_LOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p already unswappable", m)); vm_page_dequeue(m); vm_page_enqueue(m, PQ_UNSWAPPABLE); } /* * Release a page back to the page queues in preparation for unwiring. */ static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) { vm_page_astate_t old, new; uint16_t nflag; /* * Use a check of the valid bits to determine whether we should * accelerate reclamation of the page. The object lock might not be * held here, in which case the check is racy. At worst we will either * accelerate reclamation of a valid page and violate LRU, or * unnecessarily defer reclamation of an invalid page. * * If we were asked to not cache the page, place it near the head of the * inactive queue so that is reclaimed sooner. */ if (noreuse || vm_page_none_valid(m)) { nqueue = PQ_INACTIVE; nflag = PGA_REQUEUE_HEAD; } else { nflag = PGA_REQUEUE; } old = vm_page_astate_load(m); do { new = old; /* * If the page is already in the active queue and we are not * trying to accelerate reclamation, simply mark it as * referenced and avoid any queue operations. */ new.flags &= ~PGA_QUEUE_OP_MASK; if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && (old.flags & PGA_ENQUEUED) != 0) new.flags |= PGA_REFERENCED; else { new.flags |= nflag; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); } /* * Unwire a page and either attempt to free it or re-add it to the page queues. */ void vm_page_release(vm_page_t m, int flags) { vm_object_t object; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release: page %p is unmanaged", m)); if ((flags & VPR_TRYFREE) != 0) { for (;;) { object = atomic_load_ptr(&m->object); if (object == NULL) break; /* Depends on type-stability. */ if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) break; if (object == m->object) { vm_page_release_locked(m, flags); VM_OBJECT_WUNLOCK(object); return; } VM_OBJECT_WUNLOCK(object); } } vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); } /* See vm_page_release(). */ void vm_page_release_locked(vm_page_t m, int flags) { VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release_locked: page %p is unmanaged", m)); if (vm_page_unwire_noq(m)) { if ((flags & VPR_TRYFREE) != 0 && (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && m->dirty == 0 && vm_page_tryxbusy(m)) { /* * An unlocked lookup may have wired the page before the * busy lock was acquired, in which case the page must * not be freed. */ if (__predict_true(!vm_page_wired(m))) { vm_page_free(m); return; } vm_page_xunbusy(m); } else { vm_page_release_toq(m, PQ_INACTIVE, flags != 0); } } } static bool vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) { u_int old; KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, ("vm_page_try_blocked_op: page %p has no object", m)); KASSERT(vm_page_busied(m), ("vm_page_try_blocked_op: page %p is not busy", m)); VM_OBJECT_ASSERT_LOCKED(m->object); old = m->ref_count; do { KASSERT(old != 0, ("vm_page_try_blocked_op: page %p has no references", m)); if (VPRC_WIRE_COUNT(old) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); (op)(m); /* * If the object is read-locked, new wirings may be created via an * object lookup. */ old = vm_page_drop(m, VPRC_BLOCKED); KASSERT(!VM_OBJECT_WOWNED(m->object) || old == (VPRC_BLOCKED | VPRC_OBJREF), ("vm_page_try_blocked_op: unexpected refcount value %u for %p", old, m)); return (true); } /* * Atomically check for wirings and remove all mappings of the page. */ bool vm_page_try_remove_all(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_all)); } /* * Atomically check for wirings and remove all writeable mappings of the page. */ bool vm_page_try_remove_write(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_write)); } /* * vm_page_advise * * Apply the specified advice to the given page. */ void vm_page_advise(vm_page_t m, int advice) { VM_OBJECT_ASSERT_WLOCKED(m->object); vm_page_assert_xbusied(m); if (advice == MADV_FREE) /* * Mark the page clean. This will allow the page to be freed * without first paging it out. MADV_FREE pages are often * quickly reused by malloc(3), so we do not do anything that * would result in a page fault on a later access. */ vm_page_undirty(m); else if (advice != MADV_DONTNEED) { if (advice == MADV_WILLNEED) vm_page_activate(m); return; } if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) vm_page_dirty(m); /* * Clear any references to the page. Otherwise, the page daemon will * immediately reactivate the page. */ vm_page_aflag_clear(m, PGA_REFERENCED); /* * Place clean pages near the head of the inactive queue rather than * the tail, thus defeating the queue's LRU operation and ensuring that * the page will be reused quickly. Dirty pages not already in the * laundry are moved there. */ if (m->dirty == 0) vm_page_deactivate_noreuse(m); else if (!vm_page_in_laundry(m)) vm_page_launder(m); } /* * vm_page_grab_release * * Helper routine for grab functions to release busy on return. */ static inline void vm_page_grab_release(vm_page_t m, int allocflags) { if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) vm_page_sunbusy(m); else vm_page_xunbusy(m); } } /* * vm_page_grab_sleep * * Sleep for busy according to VM_ALLOC_ parameters. Returns true * if the caller should retry and false otherwise. * * If the object is locked on entry the object will be unlocked with * false returns and still locked but possibly having been dropped * with true returns. */ static bool vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (false); /* * Reference the page before unlocking and sleeping so that * the page daemon is less likely to reclaim it. */ if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_reference(m); if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && locked) VM_OBJECT_WLOCK(object); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (false); return (true); } /* * Assert that the grab flags are valid. */ static inline void vm_page_grab_check(int allocflags) { KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || (allocflags & VM_ALLOC_WIRED) != 0, ("vm_page_grab*: the pages must be busied or wired")); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); } /* * Calculate the page allocation flags for grab. */ static inline int vm_page_grab_pflags(int allocflags) { int pflags; pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) pflags |= VM_ALLOC_SBUSY; return (pflags); } /* * Grab a page, waiting until we are waken up due to the page * changing state. We keep on waiting, if the page continues * to be in the object. If the page doesn't exist, first allocate it * and then conditionally zero it. * * This routine may sleep. * * The object must be locked on entry. The lock will, however, be released * and reacquired if the routine sleeps. */ vm_page_t vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_grab_check(allocflags); retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { if (!vm_page_tryacquire(m, allocflags)) { if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", allocflags, true)) goto retrylookup; return (NULL); } goto out; } if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); if (m == NULL) { if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) return (NULL); goto retrylookup; } if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); out: vm_page_grab_release(m, allocflags); return (m); } /* * Locklessly attempt to acquire a page given a (object, pindex) tuple * and an optional previous page to avoid the radix lookup. The resulting * page will be validated against the identity tuple and busied or wired * as requested. A NULL *mp return guarantees that the page was not in * radix at the time of the call but callers must perform higher level * synchronization or retry the operation under a lock if they require * an atomic answer. This is the only lock free validation routine, * other routines can depend on the resulting page state. * * The return value indicates whether the operation failed due to caller * flags. The return is tri-state with mp: * * (true, *mp != NULL) - The operation was successful. * (true, *mp == NULL) - The page was not found in tree. * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. */ static bool vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t prev, vm_page_t *mp, int allocflags) { vm_page_t m; vm_page_grab_check(allocflags); MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); *mp = NULL; for (;;) { /* * We may see a false NULL here because the previous page * has been removed or just inserted and the list is loaded * without barriers. Switch to radix to verify. */ if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex || atomic_load_ptr(&m->object) != object) { prev = NULL; /* * This guarantees the result is instantaneously * correct. */ m = vm_radix_lookup_unlocked(&object->rtree, pindex); } if (m == NULL) return (true); if (vm_page_trybusy(m, allocflags)) { if (m->object == object && m->pindex == pindex) break; /* relookup. */ vm_page_busy_release(m); cpu_spinwait(); continue; } if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", allocflags, false)) return (false); } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); vm_page_grab_release(m, allocflags); *mp = m; return (true); } /* * Try to locklessly grab a page and fall back to the object lock if NOCREAT * is not set. */ vm_page_t vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; vm_page_grab_check(allocflags); if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) return (NULL); if (m != NULL) return (m); /* * The radix lockless lookup should never return a false negative * errors. If the user specifies NOCREAT they are guaranteed there * was no page present at the instant of the call. A NOCREAT caller * must handle create races gracefully. */ if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); VM_OBJECT_WLOCK(object); m = vm_page_grab(object, pindex, allocflags); VM_OBJECT_WUNLOCK(object); return (m); } /* * Grab a page and make it valid, paging in if necessary. Pages missing from * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought * in simultaneously. Additional pages will be left on a paging queue but * will neither be wired nor busy regardless of allocflags. */ int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; vm_page_t ma[VM_INITIAL_PAGEIN]; int after, i, pflags, rv; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); VM_OBJECT_ASSERT_WLOCKED(object); pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); pflags |= VM_ALLOC_WAITFAIL; retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { /* * If the page is fully valid it can only become invalid * with the object lock held. If it is not valid it can * become valid with the busy lock held. Therefore, we * may unnecessarily lock the exclusive busy here if we * race with I/O completion not using the object lock. * However, we will not end up with an invalid page and a * shared lock. */ if (!vm_page_trybusy(m, vm_page_all_valid(m) ? allocflags : 0)) { (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", allocflags, true); goto retrylookup; } if (vm_page_all_valid(m)) goto out; if ((allocflags & VM_ALLOC_NOCREAT) != 0) { vm_page_busy_release(m); *mp = NULL; return (VM_PAGER_FAIL); } } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { if (!vm_pager_can_alloc_page(object, pindex)) { *mp = NULL; return (VM_PAGER_AGAIN); } goto retrylookup; } vm_page_assert_xbusied(m); if (vm_pager_has_page(object, pindex, NULL, &after)) { after = MIN(after, VM_INITIAL_PAGEIN); after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); after = MAX(after, 1); ma[0] = m; for (i = 1; i < after; i++) { if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { if (vm_page_any_valid(ma[i]) || !vm_page_tryxbusy(ma[i])) break; } else { ma[i] = vm_page_alloc(object, m->pindex + i, VM_ALLOC_NORMAL); if (ma[i] == NULL) break; } } after = i; vm_object_pip_add(object, after); VM_OBJECT_WUNLOCK(object); rv = vm_pager_get_pages(object, ma, after, NULL, NULL); VM_OBJECT_WLOCK(object); vm_object_pip_wakeupn(object, after); /* Pager may have replaced a page. */ m = ma[0]; if (rv != VM_PAGER_OK) { for (i = 0; i < after; i++) { if (!vm_page_wired(ma[i])) vm_page_free(ma[i]); else vm_page_xunbusy(ma[i]); } *mp = NULL; return (rv); } for (i = 1; i < after; i++) vm_page_readahead_finish(ma[i]); MPASS(vm_page_all_valid(m)); } else { vm_page_zero_invalid(m, TRUE); } out: if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) vm_page_busy_downgrade(m); else if ((allocflags & VM_ALLOC_NOBUSY) != 0) vm_page_busy_release(m); *mp = m; return (VM_PAGER_OK); } /* * Locklessly grab a valid page. If the page is not valid or not yet * allocated this will fall back to the object lock method. */ int vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; int flags; int error; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " "mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); /* * Attempt a lockless lookup and busy. We need at least an sbusy * before we can inspect the valid field and return a wired page. */ flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) return (VM_PAGER_FAIL); if ((m = *mp) != NULL) { if (vm_page_all_valid(m)) { if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); vm_page_grab_release(m, allocflags); return (VM_PAGER_OK); } vm_page_busy_release(m); } if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } VM_OBJECT_WLOCK(object); error = vm_page_grab_valid(mp, object, pindex, allocflags); VM_OBJECT_WUNLOCK(object); return (error); } /* * Return the specified range of pages from the given object. For each * page offset within the range, if a page already exists within the object * at that offset and it is busy, then wait for it to change state. If, * instead, the page doesn't exist, then allocate it. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs the pages * * The caller must always specify that the pages are to be busied and/or * wired. * * optional allocation flags: * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NOWAIT do not sleep * VM_ALLOC_SBUSY set page to sbusy state * VM_ALLOC_WIRED wire the pages * VM_ALLOC_ZERO zero and validate any invalid pages * * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it * may return a partial prefix of the requested range. */ int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, mpred; int pflags; int i; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); KASSERT(count > 0, ("vm_page_grab_pages: invalid page count %d", count)); vm_page_grab_check(allocflags); pflags = vm_page_grab_pflags(allocflags); i = 0; retrylookup: m = vm_radix_lookup_le(&object->rtree, pindex + i); if (m == NULL || m->pindex != pindex + i) { mpred = m; m = NULL; } else mpred = TAILQ_PREV(m, pglist, listq); for (; i < count; i++) { if (m != NULL) { if (!vm_page_tryacquire(m, allocflags)) { if (vm_page_grab_sleep(object, m, pindex + i, "grbmaw", allocflags, true)) goto retrylookup; break; } } else { if ((allocflags & VM_ALLOC_NOCREAT) != 0) break; m = vm_page_alloc_after(object, pindex + i, pflags | VM_ALLOC_COUNT(count - i), mpred); if (m == NULL) { if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) break; goto retrylookup; } } if (vm_page_none_valid(m) && (allocflags & VM_ALLOC_ZERO) != 0) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); } vm_page_grab_release(m, allocflags); ma[i] = mpred = m; m = vm_page_next(m); } return (i); } /* * Unlocked variant of vm_page_grab_pages(). This accepts the same flags * and will fall back to the locked variant to handle allocation. */ int vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, pred; int flags; int i; KASSERT(count > 0, ("vm_page_grab_pages_unlocked: invalid page count %d", count)); vm_page_grab_check(allocflags); /* * Modify flags for lockless acquire to hold the page until we * set it valid if necessary. */ flags = allocflags & ~VM_ALLOC_NOBUSY; pred = NULL; for (i = 0; i < count; i++, pindex++) { if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) return (i); if (m == NULL) break; if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); } /* m will still be wired or busy according to flags. */ vm_page_grab_release(m, allocflags); pred = ma[i] = m; } if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) return (i); count -= i; VM_OBJECT_WLOCK(object); i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); VM_OBJECT_WUNLOCK(object); return (i); } /* * Mapping function for valid or dirty bits in a page. * * Inputs are required to range within a page. */ vm_page_bits_t vm_page_bits(int base, int size) { int first_bit; int last_bit; KASSERT( base + size <= PAGE_SIZE, ("vm_page_bits: illegal base/size %d/%d", base, size) ); if (size == 0) /* handle degenerate case */ return (0); first_bit = base >> DEV_BSHIFT; last_bit = (base + size - 1) >> DEV_BSHIFT; return (((vm_page_bits_t)2 << last_bit) - ((vm_page_bits_t)1 << first_bit)); } void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) { #if PAGE_SIZE == 32768 atomic_set_64((uint64_t *)bits, set); #elif PAGE_SIZE == 16384 atomic_set_32((uint32_t *)bits, set); #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) atomic_set_16((uint16_t *)bits, set); #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) atomic_set_8((uint8_t *)bits, set); #else /* PAGE_SIZE <= 8192 */ uintptr_t addr; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_set_32((uint32_t *)addr, set << shift); #endif /* PAGE_SIZE */ } static inline void vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) { #if PAGE_SIZE == 32768 atomic_clear_64((uint64_t *)bits, clear); #elif PAGE_SIZE == 16384 atomic_clear_32((uint32_t *)bits, clear); #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) atomic_clear_16((uint16_t *)bits, clear); #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) atomic_clear_8((uint8_t *)bits, clear); #else /* PAGE_SIZE <= 8192 */ uintptr_t addr; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_clear_32((uint32_t *)addr, clear << shift); #endif /* PAGE_SIZE */ } static inline vm_page_bits_t vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) { #if PAGE_SIZE == 32768 uint64_t old; old = *bits; while (atomic_fcmpset_64(bits, &old, newbits) == 0); return (old); #elif PAGE_SIZE == 16384 uint32_t old; old = *bits; while (atomic_fcmpset_32(bits, &old, newbits) == 0); return (old); #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) uint16_t old; old = *bits; while (atomic_fcmpset_16(bits, &old, newbits) == 0); return (old); #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) uint8_t old; old = *bits; while (atomic_fcmpset_8(bits, &old, newbits) == 0); return (old); #else /* PAGE_SIZE <= 4096*/ uintptr_t addr; uint32_t old, new, mask; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, swap, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); mask = VM_PAGE_BITS_ALL << shift; old = *bits; do { new = old & ~mask; new |= newbits << shift; } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); return (old >> shift); #endif /* PAGE_SIZE */ } /* * vm_page_set_valid_range: * * Sets portions of a page valid. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zeroed. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_valid_range(vm_page_t m, int base, int size) { int endoff, frag; vm_page_bits_t pagebits; vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Assert that no previously invalid block that is now being validated * is already dirty. */ KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, ("vm_page_set_valid_range: page %p is dirty", m)); /* * Set valid bits inclusive of any overlap. */ pagebits = vm_page_bits(base, size); if (vm_page_xbusied(m)) m->valid |= pagebits; else vm_page_bits_set(m, &m->valid, pagebits); } /* * Set the page dirty bits and free the invalid swap space if * present. Returns the previous dirty bits. */ vm_page_bits_t vm_page_set_dirty(vm_page_t m) { vm_page_bits_t old; VM_PAGE_OBJECT_BUSY_ASSERT(m); if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { old = m->dirty; m->dirty = VM_PAGE_BITS_ALL; } else old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) vm_pager_page_unswapped(m); return (old); } /* * Clear the given bits from the specified page's dirty field. */ static __inline void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) { vm_page_assert_busied(m); /* * If the page is xbusied and not write mapped we are the * only thread that can modify dirty bits. Otherwise, The pmap * layer can call vm_page_dirty() without holding a distinguished * lock. The combination of page busy and atomic operations * suffice to guarantee consistency of the page dirty field. */ if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) m->dirty &= ~pagebits; else vm_page_bits_clear(m, &m->dirty, pagebits); } /* * vm_page_set_validclean: * * Sets portions of a page valid and clean. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zero'd. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_validclean(vm_page_t m, int base, int size) { vm_page_bits_t oldvalid, pagebits; int endoff, frag; vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Set valid, clear dirty bits. If validating the entire * page we can safely clear the pmap modify bit. We also * use this opportunity to clear the PGA_NOSYNC flag. If a process * takes a write fault on a MAP_NOSYNC memory area the flag will * be set again. * * We set valid bits inclusive of any overlap, but we can only * clear dirty bits for DEV_BSIZE chunks that are fully within * the range. */ oldvalid = m->valid; pagebits = vm_page_bits(base, size); if (vm_page_xbusied(m)) m->valid |= pagebits; else vm_page_bits_set(m, &m->valid, pagebits); #if 0 /* NOT YET */ if ((frag = base & (DEV_BSIZE - 1)) != 0) { frag = DEV_BSIZE - frag; base += frag; size -= frag; if (size < 0) size = 0; } pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); #endif if (base == 0 && size == PAGE_SIZE) { /* * The page can only be modified within the pmap if it is * mapped, and it can only be mapped if it was previously * fully valid. */ if (oldvalid == VM_PAGE_BITS_ALL) /* * Perform the pmap_clear_modify() first. Otherwise, * a concurrent pmap operation, such as * pmap_protect(), could clear a modification in the * pmap and set the dirty field on the page before * pmap_clear_modify() had begun and after the dirty * field was cleared here. */ pmap_clear_modify(m); m->dirty = 0; vm_page_aflag_clear(m, PGA_NOSYNC); } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) m->dirty &= ~pagebits; else vm_page_clear_dirty_mask(m, pagebits); } void vm_page_clear_dirty(vm_page_t m, int base, int size) { vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); } /* * vm_page_set_invalid: * * Invalidates DEV_BSIZE'd chunks within a page. Both the * valid and dirty bits for the effected areas are cleared. */ void vm_page_set_invalid(vm_page_t m, int base, int size) { vm_page_bits_t bits; vm_object_t object; /* * The object lock is required so that pages can't be mapped * read-only while we're in the process of invalidating them. */ object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_assert_busied(m); if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + size >= object->un_pager.vnp.vnp_size) bits = VM_PAGE_BITS_ALL; else bits = vm_page_bits(base, size); if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) pmap_remove_all(m); KASSERT((bits == 0 && vm_page_all_valid(m)) || !pmap_page_is_mapped(m), ("vm_page_set_invalid: page %p is mapped", m)); if (vm_page_xbusied(m)) { m->valid &= ~bits; m->dirty &= ~bits; } else { vm_page_bits_clear(m, &m->valid, bits); vm_page_bits_clear(m, &m->dirty, bits); } } /* * vm_page_invalid: * * Invalidates the entire page. The page must be busy, unmapped, and * the enclosing object must be locked. The object locks protects * against concurrent read-only pmap enter which is done without * busy. */ void vm_page_invalid(vm_page_t m) { vm_page_assert_busied(m); VM_OBJECT_ASSERT_WLOCKED(m->object); MPASS(!pmap_page_is_mapped(m)); if (vm_page_xbusied(m)) m->valid = 0; else vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); } /* * vm_page_zero_invalid() * * The kernel assumes that the invalid portions of a page contain * garbage, but such pages can be mapped into memory by user code. * When this occurs, we must zero out the non-valid portions of the * page so user code sees what it expects. * * Pages are most often semi-valid when the end of a file is mapped * into memory and the file's size is not page aligned. */ void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) { int b; int i; /* * Scan the valid bits looking for invalid sections that * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the * valid bit may be set ) have already been zeroed by * vm_page_set_validclean(). */ for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { if (i == (PAGE_SIZE / DEV_BSIZE) || (m->valid & ((vm_page_bits_t)1 << i))) { if (i > b) { pmap_zero_page_area(m, b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); } b = i + 1; } } /* * setvalid is TRUE when we can safely set the zero'd areas * as being valid. We can do this if there are no cache consistency * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. */ if (setvalid) vm_page_valid(m); } /* * vm_page_is_valid: * * Is (partial) page valid? Note that the case where size == 0 * will return FALSE in the degenerate case where the page is * entirely invalid, and TRUE otherwise. * * Some callers envoke this routine without the busy lock held and * handle races via higher level locks. Typical callers should * hold a busy lock to prevent invalidation. */ int vm_page_is_valid(vm_page_t m, int base, int size) { vm_page_bits_t bits; bits = vm_page_bits(base, size); return (vm_page_any_valid(m) && (m->valid & bits) == bits); } /* * Returns true if all of the specified predicates are true for the entire * (super)page and false otherwise. */ bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) { vm_object_t object; int i, npages; object = m->object; if (skip_m != NULL && skip_m->object != object) return (false); VM_OBJECT_ASSERT_LOCKED(object); npages = atop(pagesizes[m->psind]); /* * The physically contiguous pages that make up a superpage, i.e., a * page with a page size index ("psind") greater than zero, will * occupy adjacent entries in vm_page_array[]. */ for (i = 0; i < npages; i++) { /* Always test object consistency, including "skip_m". */ if (m[i].object != object) return (false); if (&m[i] == skip_m) continue; if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) return (false); if ((flags & PS_ALL_DIRTY) != 0) { /* * Calling vm_page_test_dirty() or pmap_is_modified() * might stop this case from spuriously returning * "false". However, that would require a write lock * on the object containing "m[i]". */ if (m[i].dirty != VM_PAGE_BITS_ALL) return (false); } if ((flags & PS_ALL_VALID) != 0 && m[i].valid != VM_PAGE_BITS_ALL) return (false); } return (true); } /* * Set the page's dirty bits if the page is modified. */ void vm_page_test_dirty(vm_page_t m) { vm_page_assert_busied(m); if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) vm_page_dirty(m); } void vm_page_valid(vm_page_t m) { vm_page_assert_busied(m); if (vm_page_xbusied(m)) m->valid = VM_PAGE_BITS_ALL; else vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); } void vm_page_lock_KBI(vm_page_t m, const char *file, int line) { mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); } void vm_page_unlock_KBI(vm_page_t m, const char *file, int line) { mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); } int vm_page_trylock_KBI(vm_page_t m, const char *file, int line) { return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); } #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) { vm_page_lock_assert_KBI(m, MA_OWNED, file, line); } void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) { mtx_assert_(vm_page_lockptr(m), a, file, line); } #endif #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m) { /* * Certain of the page's fields may only be modified by the * holder of a page or object busy. */ if (m->object != NULL && !vm_page_busied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) { if ((bits & PGA_WRITEABLE) == 0) return; /* * The PGA_WRITEABLE flag can only be set if the page is * managed, is exclusively busied or the object is locked. * Currently, this flag is only set by pmap_enter(). */ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("PGA_WRITEABLE on unmanaged page")); if (!vm_page_xbusied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } #endif #include "opt_ddb.h" #ifdef DDB #include #include DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) { db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); } DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) { int dom; db_printf("pq_free %d\n", vm_free_count()); for (dom = 0; dom < vm_ndomains; dom++) { db_printf( "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", dom, vm_dom[dom].vmd_page_count, vm_dom[dom].vmd_free_count, vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); } } DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) { vm_page_t m; boolean_t phys, virt; if (!have_addr) { db_printf("show pginfo addr\n"); return; } phys = strchr(modif, 'p') != NULL; virt = strchr(modif, 'v') != NULL; if (virt) m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); else if (phys) m = PHYS_TO_VM_PAGE(addr); else m = (vm_page_t)addr; db_printf( "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, m->a.queue, m->ref_count, m->a.flags, m->oflags, m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); } #endif /* DDB */ diff --git a/sys/vm/vm_page.h b/sys/vm/vm_page.h index 824a853fb0f7..8ac99da21c59 100644 --- a/sys/vm/vm_page.h +++ b/sys/vm/vm_page.h @@ -1,1030 +1,1028 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * 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. * 3. 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. * * from: @(#)vm_page.h 8.2 (Berkeley) 12/13/93 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. * * $FreeBSD$ */ /* * Resident memory system definitions. */ #ifndef _VM_PAGE_ #define _VM_PAGE_ #include #include /* * Management of resident (logical) pages. * * A small structure is kept for each resident * page, indexed by page number. Each structure * is an element of several collections: * * A radix tree used to quickly * perform object/offset lookups * * A list of all pages for a given object, * so they can be quickly deactivated at * time of deallocation. * * An ordered list of pages due for pageout. * * In addition, the structure contains the object * and offset to which this page belongs (for pageout), * and sundry status bits. * * In general, operations on this structure's mutable fields are * synchronized using either one of or a combination of locks. If a * field is annotated with two of these locks then holding either is * sufficient for read access but both are required for write access. * The queue lock for a page depends on the value of its queue field and is * described in detail below. * * The following annotations are possible: * (A) the field must be accessed using atomic(9) and may require * additional synchronization. * (B) the page busy lock. * (C) the field is immutable. * (F) the per-domain lock for the free queues. * (M) Machine dependent, defined by pmap layer. * (O) the object that the page belongs to. * (Q) the page's queue lock. * * The busy lock is an embedded reader-writer lock that protects the * page's contents and identity (i.e., its tuple) as * well as certain valid/dirty modifications. To avoid bloating the * the page structure, the busy lock lacks some of the features available * the kernel's general-purpose synchronization primitives. As a result, * busy lock ordering rules are not verified, lock recursion is not * detected, and an attempt to xbusy a busy page or sbusy an xbusy page * results will trigger a panic rather than causing the thread to block. * vm_page_sleep_if_busy() can be used to sleep until the page's busy * state changes, after which the caller must re-lookup the page and * re-evaluate its state. vm_page_busy_acquire() will block until * the lock is acquired. * * The valid field is protected by the page busy lock (B) and object * lock (O). Transitions from invalid to valid are generally done * via I/O or zero filling and do not require the object lock. * These must be protected with the busy lock to prevent page-in or * creation races. Page invalidation generally happens as a result * of truncate or msync. When invalidated, pages must not be present * in pmap and must hold the object lock to prevent concurrent * speculative read-only mappings that do not require busy. I/O * routines may check for validity without a lock if they are prepared * to handle invalidation races with higher level locks (vnode) or are * unconcerned with races so long as they hold a reference to prevent * recycling. When a valid bit is set while holding a shared busy * lock (A) atomic operations are used to protect against concurrent * modification. * * In contrast, the synchronization of accesses to the page's * dirty field is a mix of machine dependent (M) and busy (B). In * the machine-independent layer, the page busy must be held to * operate on the field. However, the pmap layer is permitted to * set all bits within the field without holding that lock. If the * underlying architecture does not support atomic read-modify-write * operations on the field's type, then the machine-independent * layer uses a 32-bit atomic on the aligned 32-bit word that * contains the dirty field. In the machine-independent layer, * the implementation of read-modify-write operations on the * field is encapsulated in vm_page_clear_dirty_mask(). An * exclusive busy lock combined with pmap_remove_{write/all}() is the * only way to ensure a page can not become dirty. I/O generally * removes the page from pmap to ensure exclusive access and atomic * writes. * * The ref_count field tracks references to the page. References that * prevent the page from being reclaimable are called wirings and are * counted in the low bits of ref_count. The containing object's * reference, if one exists, is counted using the VPRC_OBJREF bit in the * ref_count field. Additionally, the VPRC_BLOCKED bit is used to * atomically check for wirings and prevent new wirings via * pmap_extract_and_hold(). When a page belongs to an object, it may be * wired only when the object is locked, or the page is busy, or by * pmap_extract_and_hold(). As a result, if the object is locked and the * page is not busy (or is exclusively busied by the current thread), and * the page is unmapped, its wire count will not increase. The ref_count * field is updated using atomic operations in most cases, except when it * is known that no other references to the page exist, such as in the page * allocator. A page may be present in the page queues, or even actively * scanned by the page daemon, without an explicitly counted referenced. * The page daemon must therefore handle the possibility of a concurrent * free of the page. * * The queue state of a page consists of the queue and act_count fields of * its atomically updated state, and the subset of atomic flags specified * by PGA_QUEUE_STATE_MASK. The queue field contains the page's page queue * index, or PQ_NONE if it does not belong to a page queue. To modify the * queue field, the page queue lock corresponding to the old value must be * held, unless that value is PQ_NONE, in which case the queue index must * be updated using an atomic RMW operation. There is one exception to * this rule: the page daemon may transition the queue field from * PQ_INACTIVE to PQ_NONE immediately prior to freeing the page during an * inactive queue scan. At that point the page is already dequeued and no * other references to that vm_page structure can exist. The PGA_ENQUEUED * flag, when set, indicates that the page structure is physically inserted * into the queue corresponding to the page's queue index, and may only be * set or cleared with the corresponding page queue lock held. * * To avoid contention on page queue locks, page queue operations (enqueue, * dequeue, requeue) are batched using fixed-size per-CPU queues. A * deferred operation is requested by setting one of the flags in * PGA_QUEUE_OP_MASK and inserting an entry into a batch queue. When a * queue is full, an attempt to insert a new entry will lock the page * queues and trigger processing of the pending entries. The * type-stability of vm_page structures is crucial to this scheme since the * processing of entries in a given batch queue may be deferred * indefinitely. In particular, a page may be freed with pending batch * queue entries. The page queue operation flags must be set using atomic * RWM operations. */ #if PAGE_SIZE == 4096 #define VM_PAGE_BITS_ALL 0xffu typedef uint8_t vm_page_bits_t; #elif PAGE_SIZE == 8192 #define VM_PAGE_BITS_ALL 0xffffu typedef uint16_t vm_page_bits_t; #elif PAGE_SIZE == 16384 #define VM_PAGE_BITS_ALL 0xffffffffu typedef uint32_t vm_page_bits_t; #elif PAGE_SIZE == 32768 #define VM_PAGE_BITS_ALL 0xfffffffffffffffflu typedef uint64_t vm_page_bits_t; #endif typedef union vm_page_astate { struct { uint16_t flags; uint8_t queue; uint8_t act_count; }; uint32_t _bits; } vm_page_astate_t; struct vm_page { union { TAILQ_ENTRY(vm_page) q; /* page queue or free list (Q) */ struct { SLIST_ENTRY(vm_page) ss; /* private slists */ } s; struct { u_long p; u_long v; } memguard; struct { void *slab; void *zone; } uma; } plinks; TAILQ_ENTRY(vm_page) listq; /* pages in same object (O) */ vm_object_t object; /* which object am I in (O) */ vm_pindex_t pindex; /* offset into object (O,P) */ vm_paddr_t phys_addr; /* physical address of page (C) */ struct md_page md; /* machine dependent stuff */ u_int ref_count; /* page references (A) */ u_int busy_lock; /* busy owners lock (A) */ union vm_page_astate a; /* state accessed atomically (A) */ uint8_t order; /* index of the buddy queue (F) */ uint8_t pool; /* vm_phys freepool index (F) */ uint8_t flags; /* page PG_* flags (P) */ uint8_t oflags; /* page VPO_* flags (O) */ int8_t psind; /* pagesizes[] index (O) */ int8_t segind; /* vm_phys segment index (C) */ /* NOTE that these must support one bit per DEV_BSIZE in a page */ /* so, on normal X86 kernels, they must be at least 8 bits wide */ vm_page_bits_t valid; /* valid DEV_BSIZE chunk map (O,B) */ vm_page_bits_t dirty; /* dirty DEV_BSIZE chunk map (M,B) */ }; /* * Special bits used in the ref_count field. * * ref_count is normally used to count wirings that prevent the page from being * reclaimed, but also supports several special types of references that do not * prevent reclamation. Accesses to the ref_count field must be atomic unless * the page is unallocated. * * VPRC_OBJREF is the reference held by the containing object. It can set or * cleared only when the corresponding object's write lock is held. * * VPRC_BLOCKED is used to atomically block wirings via pmap lookups while * attempting to tear down all mappings of a given page. The page busy lock and * object write lock must both be held in order to set or clear this bit. */ #define VPRC_BLOCKED 0x40000000u /* mappings are being removed */ #define VPRC_OBJREF 0x80000000u /* object reference, cleared with (O) */ #define VPRC_WIRE_COUNT(c) ((c) & ~(VPRC_BLOCKED | VPRC_OBJREF)) #define VPRC_WIRE_COUNT_MAX (~(VPRC_BLOCKED | VPRC_OBJREF)) /* * Page flags stored in oflags: * * Access to these page flags is synchronized by the lock on the object * containing the page (O). * * Note: VPO_UNMANAGED (used by OBJT_DEVICE, OBJT_PHYS and OBJT_SG) * indicates that the page is not under PV management but * otherwise should be treated as a normal page. Pages not * under PV management cannot be paged out via the * object/vm_page_t because there is no knowledge of their pte * mappings, and such pages are also not on any PQ queue. * */ #define VPO_KMEM_EXEC 0x01 /* kmem mapping allows execution */ #define VPO_SWAPSLEEP 0x02 /* waiting for swap to finish */ #define VPO_UNMANAGED 0x04 /* no PV management for page */ #define VPO_SWAPINPROG 0x08 /* swap I/O in progress on page */ /* * Busy page implementation details. * The algorithm is taken mostly by rwlock(9) and sx(9) locks implementation, * even if the support for owner identity is removed because of size * constraints. Checks on lock recursion are then not possible, while the * lock assertions effectiveness is someway reduced. */ #define VPB_BIT_SHARED 0x01 #define VPB_BIT_EXCLUSIVE 0x02 #define VPB_BIT_WAITERS 0x04 #define VPB_BIT_FLAGMASK \ (VPB_BIT_SHARED | VPB_BIT_EXCLUSIVE | VPB_BIT_WAITERS) #define VPB_SHARERS_SHIFT 3 #define VPB_SHARERS(x) \ (((x) & ~VPB_BIT_FLAGMASK) >> VPB_SHARERS_SHIFT) #define VPB_SHARERS_WORD(x) ((x) << VPB_SHARERS_SHIFT | VPB_BIT_SHARED) #define VPB_ONE_SHARER (1 << VPB_SHARERS_SHIFT) #define VPB_SINGLE_EXCLUSIVE VPB_BIT_EXCLUSIVE #ifdef INVARIANTS #define VPB_CURTHREAD_EXCLUSIVE \ (VPB_BIT_EXCLUSIVE | ((u_int)(uintptr_t)curthread & ~VPB_BIT_FLAGMASK)) #else #define VPB_CURTHREAD_EXCLUSIVE VPB_SINGLE_EXCLUSIVE #endif #define VPB_UNBUSIED VPB_SHARERS_WORD(0) /* Freed lock blocks both shared and exclusive. */ #define VPB_FREED (0xffffffff - VPB_BIT_SHARED) #define PQ_NONE 255 #define PQ_INACTIVE 0 #define PQ_ACTIVE 1 #define PQ_LAUNDRY 2 #define PQ_UNSWAPPABLE 3 #define PQ_COUNT 4 #ifndef VM_PAGE_HAVE_PGLIST TAILQ_HEAD(pglist, vm_page); #define VM_PAGE_HAVE_PGLIST #endif SLIST_HEAD(spglist, vm_page); #ifdef _KERNEL extern vm_page_t bogus_page; #endif /* _KERNEL */ extern struct mtx_padalign pa_lock[]; #if defined(__arm__) #define PDRSHIFT PDR_SHIFT #elif !defined(PDRSHIFT) #define PDRSHIFT 21 #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define PA_LOCKPTR(pa) ((struct mtx *)(&pa_lock[pa_index(pa) % PA_LOCK_COUNT])) #define PA_LOCKOBJPTR(pa) ((struct lock_object *)PA_LOCKPTR((pa))) #define PA_LOCK(pa) mtx_lock(PA_LOCKPTR(pa)) #define PA_TRYLOCK(pa) mtx_trylock(PA_LOCKPTR(pa)) #define PA_UNLOCK(pa) mtx_unlock(PA_LOCKPTR(pa)) #define PA_UNLOCK_COND(pa) \ do { \ if ((pa) != 0) { \ PA_UNLOCK((pa)); \ (pa) = 0; \ } \ } while (0) #define PA_LOCK_ASSERT(pa, a) mtx_assert(PA_LOCKPTR(pa), (a)) #if defined(KLD_MODULE) && !defined(KLD_TIED) #define vm_page_lock(m) vm_page_lock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_unlock(m) vm_page_unlock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_trylock(m) vm_page_trylock_KBI((m), LOCK_FILE, LOCK_LINE) #else /* !KLD_MODULE */ #define vm_page_lockptr(m) (PA_LOCKPTR(VM_PAGE_TO_PHYS((m)))) #define vm_page_lock(m) mtx_lock(vm_page_lockptr((m))) #define vm_page_unlock(m) mtx_unlock(vm_page_lockptr((m))) #define vm_page_trylock(m) mtx_trylock(vm_page_lockptr((m))) #endif #if defined(INVARIANTS) #define vm_page_assert_locked(m) \ vm_page_assert_locked_KBI((m), __FILE__, __LINE__) #define vm_page_lock_assert(m, a) \ vm_page_lock_assert_KBI((m), (a), __FILE__, __LINE__) #else #define vm_page_assert_locked(m) #define vm_page_lock_assert(m, a) #endif /* * The vm_page's aflags are updated using atomic operations. To set or clear * these flags, the functions vm_page_aflag_set() and vm_page_aflag_clear() * must be used. Neither these flags nor these functions are part of the KBI. * * PGA_REFERENCED may be cleared only if the page is locked. It is set by * both the MI and MD VM layers. However, kernel loadable modules should not * directly set this flag. They should call vm_page_reference() instead. * * PGA_WRITEABLE is set exclusively on managed pages by pmap_enter(). * When it does so, the object must be locked, or the page must be * exclusive busied. The MI VM layer must never access this flag * directly. Instead, it should call pmap_page_is_write_mapped(). * * PGA_EXECUTABLE may be set by pmap routines, and indicates that a page has * at least one executable mapping. It is not consumed by the MI VM layer. * * PGA_NOSYNC must be set and cleared with the page busy lock held. * * PGA_ENQUEUED is set and cleared when a page is inserted into or removed * from a page queue, respectively. It determines whether the plinks.q field * of the page is valid. To set or clear this flag, page's "queue" field must * be a valid queue index, and the corresponding page queue lock must be held. * * PGA_DEQUEUE is set when the page is scheduled to be dequeued from a page * queue, and cleared when the dequeue request is processed. A page may * have PGA_DEQUEUE set and PGA_ENQUEUED cleared, for instance if a dequeue * is requested after the page is scheduled to be enqueued but before it is * actually inserted into the page queue. * * PGA_REQUEUE is set when the page is scheduled to be enqueued or requeued * in its page queue. * * PGA_REQUEUE_HEAD is a special flag for enqueuing pages near the head of * the inactive queue, thus bypassing LRU. * * The PGA_DEQUEUE, PGA_REQUEUE and PGA_REQUEUE_HEAD flags must be set using an * atomic RMW operation to ensure that the "queue" field is a valid queue index, * and the corresponding page queue lock must be held when clearing any of the * flags. * * PGA_SWAP_FREE is used to defer freeing swap space to the pageout daemon * when the context that dirties the page does not have the object write lock * held. */ #define PGA_WRITEABLE 0x0001 /* page may be mapped writeable */ #define PGA_REFERENCED 0x0002 /* page has been referenced */ #define PGA_EXECUTABLE 0x0004 /* page may be mapped executable */ #define PGA_ENQUEUED 0x0008 /* page is enqueued in a page queue */ #define PGA_DEQUEUE 0x0010 /* page is due to be dequeued */ #define PGA_REQUEUE 0x0020 /* page is due to be requeued */ #define PGA_REQUEUE_HEAD 0x0040 /* page requeue should bypass LRU */ #define PGA_NOSYNC 0x0080 /* do not collect for syncer */ #define PGA_SWAP_FREE 0x0100 /* page with swap space was dirtied */ #define PGA_SWAP_SPACE 0x0200 /* page has allocated swap space */ #define PGA_QUEUE_OP_MASK (PGA_DEQUEUE | PGA_REQUEUE | PGA_REQUEUE_HEAD) #define PGA_QUEUE_STATE_MASK (PGA_ENQUEUED | PGA_QUEUE_OP_MASK) /* * Page flags. Updates to these flags are not synchronized, and thus they must * be set during page allocation or free to avoid races. * * The PG_PCPU_CACHE flag is set at allocation time if the page was * allocated from a per-CPU cache. It is cleared the next time that the * page is allocated from the physical memory allocator. */ #define PG_PCPU_CACHE 0x01 /* was allocated from per-CPU caches */ #define PG_FICTITIOUS 0x02 /* physical page doesn't exist */ #define PG_ZERO 0x04 /* page is zeroed */ #define PG_MARKER 0x08 /* special queue marker page */ #define PG_NODUMP 0x10 /* don't include this page in a dump */ /* * Misc constants. */ #define ACT_DECLINE 1 #define ACT_ADVANCE 3 #define ACT_INIT 5 #define ACT_MAX 64 #ifdef _KERNEL #include #include /* * Each pageable resident page falls into one of five lists: * * free * Available for allocation now. * * inactive * Low activity, candidates for reclamation. * This list is approximately LRU ordered. * * laundry * This is the list of pages that should be * paged out next. * * unswappable * Dirty anonymous pages that cannot be paged * out because no swap device is configured. * * active * Pages that are "active", i.e., they have been * recently referenced. * */ extern vm_page_t vm_page_array; /* First resident page in table */ extern long vm_page_array_size; /* number of vm_page_t's */ extern long first_page; /* first physical page number */ #define VM_PAGE_TO_PHYS(entry) ((entry)->phys_addr) /* * PHYS_TO_VM_PAGE() returns the vm_page_t object that represents a memory * page to which the given physical address belongs. The correct vm_page_t * object is returned for addresses that are not page-aligned. */ vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa); /* * Page allocation parameters for vm_page for the functions * vm_page_alloc(), vm_page_grab(), vm_page_alloc_contig() and * vm_page_alloc_freelist(). Some functions support only a subset * of the flags, and ignore others, see the flags legend. * * The meaning of VM_ALLOC_ZERO differs slightly between the vm_page_alloc*() * and the vm_page_grab*() functions. See these functions for details. * * Bits 0 - 1 define class. * Bits 2 - 15 dedicated for flags. * Legend: * (a) - vm_page_alloc() supports the flag. * (c) - vm_page_alloc_contig() supports the flag. * (g) - vm_page_grab() supports the flag. * (n) - vm_page_alloc_noobj() and vm_page_alloc_freelist() support the flag. * (p) - vm_page_grab_pages() supports the flag. * Bits above 15 define the count of additional pages that the caller * intends to allocate. */ #define VM_ALLOC_NORMAL 0 #define VM_ALLOC_INTERRUPT 1 #define VM_ALLOC_SYSTEM 2 #define VM_ALLOC_CLASS_MASK 3 #define VM_ALLOC_WAITOK 0x0008 /* (acn) Sleep and retry */ #define VM_ALLOC_WAITFAIL 0x0010 /* (acn) Sleep and return error */ #define VM_ALLOC_WIRED 0x0020 /* (acgnp) Allocate a wired page */ #define VM_ALLOC_ZERO 0x0040 /* (acgnp) Allocate a zeroed page */ #define VM_ALLOC_NORECLAIM 0x0080 /* (c) Do not reclaim after failure */ #define VM_ALLOC_AVAIL0 0x0100 #define VM_ALLOC_NOBUSY 0x0200 /* (acgp) Do not excl busy the page */ #define VM_ALLOC_NOCREAT 0x0400 /* (gp) Don't create a page */ #define VM_ALLOC_AVAIL1 0x0800 #define VM_ALLOC_IGN_SBUSY 0x1000 /* (gp) Ignore shared busy flag */ #define VM_ALLOC_NODUMP 0x2000 /* (ag) don't include in dump */ #define VM_ALLOC_SBUSY 0x4000 /* (acgp) Shared busy the page */ #define VM_ALLOC_NOWAIT 0x8000 /* (acgnp) Do not sleep */ #define VM_ALLOC_COUNT_MAX 0xffff #define VM_ALLOC_COUNT_SHIFT 16 #define VM_ALLOC_COUNT_MASK (VM_ALLOC_COUNT(VM_ALLOC_COUNT_MAX)) #define VM_ALLOC_COUNT(count) ({ \ KASSERT((count) <= VM_ALLOC_COUNT_MAX, \ ("%s: invalid VM_ALLOC_COUNT value", __func__)); \ (count) << VM_ALLOC_COUNT_SHIFT; \ }) #ifdef M_NOWAIT static inline int malloc2vm_flags(int malloc_flags) { int pflags; KASSERT((malloc_flags & M_USE_RESERVE) == 0 || (malloc_flags & M_NOWAIT) != 0, ("M_USE_RESERVE requires M_NOWAIT")); pflags = (malloc_flags & M_USE_RESERVE) != 0 ? VM_ALLOC_INTERRUPT : VM_ALLOC_SYSTEM; if ((malloc_flags & M_ZERO) != 0) pflags |= VM_ALLOC_ZERO; if ((malloc_flags & M_NODUMP) != 0) pflags |= VM_ALLOC_NODUMP; if ((malloc_flags & M_NOWAIT)) pflags |= VM_ALLOC_NOWAIT; if ((malloc_flags & M_WAITOK)) pflags |= VM_ALLOC_WAITOK; if ((malloc_flags & M_NORECLAIM)) pflags |= VM_ALLOC_NORECLAIM; return (pflags); } #endif /* * Predicates supported by vm_page_ps_test(): * * PS_ALL_DIRTY is true only if the entire (super)page is dirty. * However, it can be spuriously false when the (super)page has become * dirty in the pmap but that information has not been propagated to the * machine-independent layer. */ #define PS_ALL_DIRTY 0x1 #define PS_ALL_VALID 0x2 #define PS_NONE_BUSY 0x4 bool vm_page_busy_acquire(vm_page_t m, int allocflags); void vm_page_busy_downgrade(vm_page_t m); int vm_page_busy_tryupgrade(vm_page_t m); bool vm_page_busy_sleep(vm_page_t m, const char *msg, int allocflags); void vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, const char *wmesg, int allocflags); void vm_page_free(vm_page_t m); void vm_page_free_zero(vm_page_t m); void vm_page_activate (vm_page_t); void vm_page_advise(vm_page_t m, int advice); vm_page_t vm_page_alloc(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_alloc_domain(vm_object_t, vm_pindex_t, int, int); vm_page_t vm_page_alloc_after(vm_object_t, vm_pindex_t, int, vm_page_t); vm_page_t vm_page_alloc_domain_after(vm_object_t, vm_pindex_t, int, int, vm_page_t); vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_freelist(int, int); vm_page_t vm_page_alloc_freelist_domain(int, int, int); vm_page_t vm_page_alloc_noobj(int); vm_page_t vm_page_alloc_noobj_domain(int, int); vm_page_t vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set); bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose); vm_page_t vm_page_grab(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_grab_unlocked(vm_object_t, vm_pindex_t, int); int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); int vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); void vm_page_deactivate(vm_page_t); void vm_page_deactivate_noreuse(vm_page_t); void vm_page_dequeue(vm_page_t m); void vm_page_dequeue_deferred(vm_page_t m); vm_page_t vm_page_find_least(vm_object_t, vm_pindex_t); void vm_page_free_invalid(vm_page_t); vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags); void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind); int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t); void vm_page_invalid(vm_page_t m); void vm_page_launder(vm_page_t m); vm_page_t vm_page_lookup(vm_object_t, vm_pindex_t); vm_page_t vm_page_lookup_unlocked(vm_object_t, vm_pindex_t); vm_page_t vm_page_next(vm_page_t m); void vm_page_pqbatch_drain(void); void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue); bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new); vm_page_t vm_page_prev(vm_page_t m); bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m); void vm_page_putfake(vm_page_t m); void vm_page_readahead_finish(vm_page_t m); bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); bool vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, int desired_runs); void vm_page_reference(vm_page_t m); #define VPR_TRYFREE 0x01 #define VPR_NOREUSE 0x02 void vm_page_release(vm_page_t m, int flags); void vm_page_release_locked(vm_page_t m, int flags); vm_page_t vm_page_relookup(vm_object_t, vm_pindex_t); bool vm_page_remove(vm_page_t); bool vm_page_remove_xbusy(vm_page_t); int vm_page_rename(vm_page_t, vm_object_t, vm_pindex_t); void vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold); int vm_page_sbusied(vm_page_t m); -vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, - vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options); vm_page_bits_t vm_page_set_dirty(vm_page_t m); void vm_page_set_valid_range(vm_page_t m, int base, int size); vm_offset_t vm_page_startup(vm_offset_t vaddr); void vm_page_sunbusy(vm_page_t m); bool vm_page_try_remove_all(vm_page_t m); bool vm_page_try_remove_write(vm_page_t m); int vm_page_trysbusy(vm_page_t m); int vm_page_tryxbusy(vm_page_t m); void vm_page_unhold_pages(vm_page_t *ma, int count); void vm_page_unswappable(vm_page_t m); void vm_page_unwire(vm_page_t m, uint8_t queue); bool vm_page_unwire_noq(vm_page_t m); void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_wire(vm_page_t); bool vm_page_wire_mapped(vm_page_t m); void vm_page_xunbusy_hard(vm_page_t m); void vm_page_xunbusy_hard_unchecked(vm_page_t m); void vm_page_set_validclean (vm_page_t, int, int); void vm_page_clear_dirty(vm_page_t, int, int); void vm_page_set_invalid(vm_page_t, int, int); void vm_page_valid(vm_page_t m); int vm_page_is_valid(vm_page_t, int, int); void vm_page_test_dirty(vm_page_t); vm_page_bits_t vm_page_bits(int base, int size); void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid); void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count); void vm_page_dirty_KBI(vm_page_t m); void vm_page_lock_KBI(vm_page_t m, const char *file, int line); void vm_page_unlock_KBI(vm_page_t m, const char *file, int line); int vm_page_trylock_KBI(vm_page_t m, const char *file, int line); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line); void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line); #endif #define vm_page_busy_fetch(m) atomic_load_int(&(m)->busy_lock) #define vm_page_assert_busied(m) \ KASSERT(vm_page_busied(m), \ ("vm_page_assert_busied: page %p not busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_sbusied(m) \ KASSERT(vm_page_sbusied(m), \ ("vm_page_assert_sbusied: page %p not shared busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_unbusied(m) \ KASSERT((vm_page_busy_fetch(m) & ~VPB_BIT_WAITERS) != \ VPB_CURTHREAD_EXCLUSIVE, \ ("vm_page_assert_xbusied: page %p busy_lock %#x owned" \ " by me @ %s:%d", \ (m), (m)->busy_lock, __FILE__, __LINE__)); \ #define vm_page_assert_xbusied_unchecked(m) do { \ KASSERT(vm_page_xbusied(m), \ ("vm_page_assert_xbusied: page %p not exclusive busy @ %s:%d", \ (m), __FILE__, __LINE__)); \ } while (0) #define vm_page_assert_xbusied(m) do { \ vm_page_assert_xbusied_unchecked(m); \ KASSERT((vm_page_busy_fetch(m) & ~VPB_BIT_WAITERS) == \ VPB_CURTHREAD_EXCLUSIVE, \ ("vm_page_assert_xbusied: page %p busy_lock %#x not owned" \ " by me @ %s:%d", \ (m), (m)->busy_lock, __FILE__, __LINE__)); \ } while (0) #define vm_page_busied(m) \ (vm_page_busy_fetch(m) != VPB_UNBUSIED) #define vm_page_xbusied(m) \ ((vm_page_busy_fetch(m) & VPB_SINGLE_EXCLUSIVE) != 0) #define vm_page_busy_freed(m) \ (vm_page_busy_fetch(m) == VPB_FREED) /* Note: page m's lock must not be owned by the caller. */ #define vm_page_xunbusy(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED)) \ vm_page_xunbusy_hard(m); \ } while (0) #define vm_page_xunbusy_unchecked(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED)) \ vm_page_xunbusy_hard_unchecked(m); \ } while (0) #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m); #define VM_PAGE_OBJECT_BUSY_ASSERT(m) vm_page_object_busy_assert(m) void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits); #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) \ vm_page_assert_pga_writeable(m, bits) /* * Claim ownership of a page's xbusy state. In non-INVARIANTS kernels this * operation is a no-op since ownership is not tracked. In particular * this macro does not provide any synchronization with the previous owner. */ #define vm_page_xbusy_claim(m) do { \ u_int _busy_lock; \ \ vm_page_assert_xbusied_unchecked((m)); \ do { \ _busy_lock = vm_page_busy_fetch(m); \ } while (!atomic_cmpset_int(&(m)->busy_lock, _busy_lock, \ (_busy_lock & VPB_BIT_FLAGMASK) | VPB_CURTHREAD_EXCLUSIVE)); \ } while (0) #else #define VM_PAGE_OBJECT_BUSY_ASSERT(m) (void)0 #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) (void)0 #define vm_page_xbusy_claim(m) #endif #if BYTE_ORDER == BIG_ENDIAN #define VM_PAGE_AFLAG_SHIFT 16 #else #define VM_PAGE_AFLAG_SHIFT 0 #endif /* * Load a snapshot of a page's 32-bit atomic state. */ static inline vm_page_astate_t vm_page_astate_load(vm_page_t m) { vm_page_astate_t a; a._bits = atomic_load_32(&m->a._bits); return (a); } /* * Atomically compare and set a page's atomic state. */ static inline bool vm_page_astate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { KASSERT(new.queue == PQ_INACTIVE || (new.flags & PGA_REQUEUE_HEAD) == 0, ("%s: invalid head requeue request for page %p", __func__, m)); KASSERT((new.flags & PGA_ENQUEUED) == 0 || new.queue != PQ_NONE, ("%s: setting PGA_ENQUEUED with PQ_NONE in page %p", __func__, m)); KASSERT(new._bits != old->_bits, ("%s: bits are unchanged", __func__)); return (atomic_fcmpset_32(&m->a._bits, &old->_bits, new._bits) != 0); } /* * Clear the given bits in the specified page. */ static inline void vm_page_aflag_clear(vm_page_t m, uint16_t bits) { uint32_t *addr, val; /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->a; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_clear_32(addr, val); } /* * Set the given bits in the specified page. */ static inline void vm_page_aflag_set(vm_page_t m, uint16_t bits) { uint32_t *addr, val; VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->a; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_set_32(addr, val); } /* * vm_page_dirty: * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). */ static __inline void vm_page_dirty(vm_page_t m) { /* Use vm_page_dirty_KBI() under INVARIANTS to save memory. */ #if (defined(KLD_MODULE) && !defined(KLD_TIED)) || defined(INVARIANTS) vm_page_dirty_KBI(m); #else m->dirty = VM_PAGE_BITS_ALL; #endif } /* * vm_page_undirty: * * Set page to not be dirty. Note: does not clear pmap modify bits */ static __inline void vm_page_undirty(vm_page_t m) { VM_PAGE_OBJECT_BUSY_ASSERT(m); m->dirty = 0; } static inline uint8_t _vm_page_queue(vm_page_astate_t as) { if ((as.flags & PGA_DEQUEUE) != 0) return (PQ_NONE); return (as.queue); } /* * vm_page_queue: * * Return the index of the queue containing m. */ static inline uint8_t vm_page_queue(vm_page_t m) { return (_vm_page_queue(vm_page_astate_load(m))); } static inline bool vm_page_active(vm_page_t m) { return (vm_page_queue(m) == PQ_ACTIVE); } static inline bool vm_page_inactive(vm_page_t m) { return (vm_page_queue(m) == PQ_INACTIVE); } static inline bool vm_page_in_laundry(vm_page_t m) { uint8_t queue; queue = vm_page_queue(m); return (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE); } /* * vm_page_drop: * * Release a reference to a page and return the old reference count. */ static inline u_int vm_page_drop(vm_page_t m, u_int val) { u_int old; /* * Synchronize with vm_page_free_prep(): ensure that all updates to the * page structure are visible before it is freed. */ atomic_thread_fence_rel(); old = atomic_fetchadd_int(&m->ref_count, -val); KASSERT(old != VPRC_BLOCKED, ("vm_page_drop: page %p has an invalid refcount value", m)); return (old); } /* * vm_page_wired: * * Perform a racy check to determine whether a reference prevents the page * from being reclaimable. If the page's object is locked, and the page is * unmapped and exclusively busied by the current thread, no new wirings * may be created. */ static inline bool vm_page_wired(vm_page_t m) { return (VPRC_WIRE_COUNT(m->ref_count) > 0); } static inline bool vm_page_all_valid(vm_page_t m) { return (m->valid == VM_PAGE_BITS_ALL); } static inline bool vm_page_any_valid(vm_page_t m) { return (m->valid != 0); } static inline bool vm_page_none_valid(vm_page_t m) { return (m->valid == 0); } static inline int vm_page_domain(vm_page_t m) { #ifdef NUMA int domn, segind; segind = m->segind; KASSERT(segind < vm_phys_nsegs, ("segind %d m %p", segind, m)); domn = vm_phys_segs[segind].domain; KASSERT(domn >= 0 && domn < vm_ndomains, ("domain %d m %p", domn, m)); return (domn); #else return (0); #endif } #endif /* _KERNEL */ #endif /* !_VM_PAGE_ */ diff --git a/sys/vm/vm_phys.c b/sys/vm/vm_phys.c index 8db6529b8c80..a0b53f0f7c4b 100644 --- a/sys/vm/vm_phys.c +++ b/sys/vm/vm_phys.c @@ -1,1894 +1,1869 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2002-2006 Rice University * Copyright (c) 2007 Alan L. Cox * All rights reserved. * * This software was developed for the FreeBSD Project by Alan L. Cox, * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro. * * 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 COPYRIGHT HOLDERS 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 COPYRIGHT * HOLDERS 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. */ /* * Physical memory system implementation * * Any external functions defined by this module are only to be used by the * virtual memory system. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX, "Too many physsegs."); #ifdef NUMA struct mem_affinity __read_mostly *mem_affinity; int __read_mostly *mem_locality; #endif int __read_mostly vm_ndomains = 1; domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1); struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX]; int __read_mostly vm_phys_nsegs; static struct vm_phys_seg vm_phys_early_segs[8]; static int vm_phys_early_nsegs; struct vm_phys_fictitious_seg; static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *, struct vm_phys_fictitious_seg *); RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree = RB_INITIALIZER(&vm_phys_fictitious_tree); struct vm_phys_fictitious_seg { RB_ENTRY(vm_phys_fictitious_seg) node; /* Memory region data */ vm_paddr_t start; vm_paddr_t end; vm_page_t first_page; }; RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node, vm_phys_fictitious_cmp); static struct rwlock_padalign vm_phys_fictitious_reg_lock; MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages"); static struct vm_freelist __aligned(CACHE_LINE_SIZE) vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL] [VM_NFREEORDER_MAX]; static int __read_mostly vm_nfreelists; /* * These "avail lists" are globals used to communicate boot-time physical * memory layout to other parts of the kernel. Each physically contiguous * region of memory is defined by a start address at an even index and an * end address at the following odd index. Each list is terminated by a * pair of zero entries. * * dump_avail tells the dump code what regions to include in a crash dump, and * phys_avail is all of the remaining physical memory that is available for * the vm system. * * Initially dump_avail and phys_avail are identical. Boot time memory * allocations remove extents from phys_avail that may still be included * in dumps. */ vm_paddr_t phys_avail[PHYS_AVAIL_COUNT]; vm_paddr_t dump_avail[PHYS_AVAIL_COUNT]; /* * Provides the mapping from VM_FREELIST_* to free list indices (flind). */ static int __read_mostly vm_freelist_to_flind[VM_NFREELIST]; CTASSERT(VM_FREELIST_DEFAULT == 0); #ifdef VM_FREELIST_DMA32 #define VM_DMA32_BOUNDARY ((vm_paddr_t)1 << 32) #endif /* * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about * the ordering of the free list boundaries. */ #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY) CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY); #endif static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS); SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info"); static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS); SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info"); #ifdef NUMA static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS); SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info"); #endif SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD, &vm_ndomains, 0, "Number of physical memory domains available."); static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain); static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end); static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order, int tail); /* * Red-black tree helpers for vm fictitious range management. */ static inline int vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p, struct vm_phys_fictitious_seg *range) { KASSERT(range->start != 0 && range->end != 0, ("Invalid range passed on search for vm_fictitious page")); if (p->start >= range->end) return (1); if (p->start < range->start) return (-1); return (0); } static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1, struct vm_phys_fictitious_seg *p2) { /* Check if this is a search for a page */ if (p1->end == 0) return (vm_phys_fictitious_in_range(p1, p2)); KASSERT(p2->end != 0, ("Invalid range passed as second parameter to vm fictitious comparison")); /* Searching to add a new range */ if (p1->end <= p2->start) return (-1); if (p1->start >= p2->end) return (1); panic("Trying to add overlapping vm fictitious ranges:\n" "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start, (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end); } int vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high) { #ifdef NUMA domainset_t mask; int i; if (vm_ndomains == 1 || mem_affinity == NULL) return (0); DOMAINSET_ZERO(&mask); /* * Check for any memory that overlaps low, high. */ for (i = 0; mem_affinity[i].end != 0; i++) if (mem_affinity[i].start <= high && mem_affinity[i].end >= low) DOMAINSET_SET(mem_affinity[i].domain, &mask); if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask)) return (prefer); if (DOMAINSET_EMPTY(&mask)) panic("vm_phys_domain_match: Impossible constraint"); return (DOMAINSET_FFS(&mask) - 1); #else return (0); #endif } /* * Outputs the state of the physical memory allocator, specifically, * the amount of physical memory in each free list. */ static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS) { struct sbuf sbuf; struct vm_freelist *fl; int dom, error, flind, oind, pind; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req); for (dom = 0; dom < vm_ndomains; dom++) { sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom); for (flind = 0; flind < vm_nfreelists; flind++) { sbuf_printf(&sbuf, "\nFREE LIST %d:\n" "\n ORDER (SIZE) | NUMBER" "\n ", flind); for (pind = 0; pind < VM_NFREEPOOL; pind++) sbuf_printf(&sbuf, " | POOL %d", pind); sbuf_printf(&sbuf, "\n-- "); for (pind = 0; pind < VM_NFREEPOOL; pind++) sbuf_printf(&sbuf, "-- -- "); sbuf_printf(&sbuf, "--\n"); for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) { sbuf_printf(&sbuf, " %2d (%6dK)", oind, 1 << (PAGE_SHIFT - 10 + oind)); for (pind = 0; pind < VM_NFREEPOOL; pind++) { fl = vm_phys_free_queues[dom][flind][pind]; sbuf_printf(&sbuf, " | %6d", fl[oind].lcnt); } sbuf_printf(&sbuf, "\n"); } } } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Outputs the set of physical memory segments. */ static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS) { struct sbuf sbuf; struct vm_phys_seg *seg; int error, segind; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); for (segind = 0; segind < vm_phys_nsegs; segind++) { sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind); seg = &vm_phys_segs[segind]; sbuf_printf(&sbuf, "start: %#jx\n", (uintmax_t)seg->start); sbuf_printf(&sbuf, "end: %#jx\n", (uintmax_t)seg->end); sbuf_printf(&sbuf, "domain: %d\n", seg->domain); sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues); } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Return affinity, or -1 if there's no affinity information. */ int vm_phys_mem_affinity(int f, int t) { #ifdef NUMA if (mem_locality == NULL) return (-1); if (f >= vm_ndomains || t >= vm_ndomains) return (-1); return (mem_locality[f * vm_ndomains + t]); #else return (-1); #endif } #ifdef NUMA /* * Outputs the VM locality table. */ static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS) { struct sbuf sbuf; int error, i, j; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); sbuf_printf(&sbuf, "\n"); for (i = 0; i < vm_ndomains; i++) { sbuf_printf(&sbuf, "%d: ", i); for (j = 0; j < vm_ndomains; j++) { sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j)); } sbuf_printf(&sbuf, "\n"); } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } #endif static void vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail) { m->order = order; if (tail) TAILQ_INSERT_TAIL(&fl[order].pl, m, listq); else TAILQ_INSERT_HEAD(&fl[order].pl, m, listq); fl[order].lcnt++; } static void vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order) { TAILQ_REMOVE(&fl[order].pl, m, listq); fl[order].lcnt--; m->order = VM_NFREEORDER; } /* * Create a physical memory segment. */ static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain) { struct vm_phys_seg *seg; KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX, ("vm_phys_create_seg: increase VM_PHYSSEG_MAX")); KASSERT(domain >= 0 && domain < vm_ndomains, ("vm_phys_create_seg: invalid domain provided")); seg = &vm_phys_segs[vm_phys_nsegs++]; while (seg > vm_phys_segs && (seg - 1)->start >= end) { *seg = *(seg - 1); seg--; } seg->start = start; seg->end = end; seg->domain = domain; } static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end) { #ifdef NUMA int i; if (mem_affinity == NULL) { _vm_phys_create_seg(start, end, 0); return; } for (i = 0;; i++) { if (mem_affinity[i].end == 0) panic("Reached end of affinity info"); if (mem_affinity[i].end <= start) continue; if (mem_affinity[i].start > start) panic("No affinity info for start %jx", (uintmax_t)start); if (mem_affinity[i].end >= end) { _vm_phys_create_seg(start, end, mem_affinity[i].domain); break; } _vm_phys_create_seg(start, mem_affinity[i].end, mem_affinity[i].domain); start = mem_affinity[i].end; } #else _vm_phys_create_seg(start, end, 0); #endif } /* * Add a physical memory segment. */ void vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end) { vm_paddr_t paddr; KASSERT((start & PAGE_MASK) == 0, ("vm_phys_define_seg: start is not page aligned")); KASSERT((end & PAGE_MASK) == 0, ("vm_phys_define_seg: end is not page aligned")); /* * Split the physical memory segment if it spans two or more free * list boundaries. */ paddr = start; #ifdef VM_FREELIST_LOWMEM if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) { vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY); paddr = VM_LOWMEM_BOUNDARY; } #endif #ifdef VM_FREELIST_DMA32 if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) { vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY); paddr = VM_DMA32_BOUNDARY; } #endif vm_phys_create_seg(paddr, end); } /* * Initialize the physical memory allocator. * * Requires that vm_page_array is initialized! */ void vm_phys_init(void) { struct vm_freelist *fl; struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg; #if defined(VM_DMA32_NPAGES_THRESHOLD) || defined(VM_PHYSSEG_SPARSE) u_long npages; #endif int dom, flind, freelist, oind, pind, segind; /* * Compute the number of free lists, and generate the mapping from the * manifest constants VM_FREELIST_* to the free list indices. * * Initially, the entries of vm_freelist_to_flind[] are set to either * 0 or 1 to indicate which free lists should be created. */ #ifdef VM_DMA32_NPAGES_THRESHOLD npages = 0; #endif for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) { seg = &vm_phys_segs[segind]; #ifdef VM_FREELIST_LOWMEM if (seg->end <= VM_LOWMEM_BOUNDARY) vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1; else #endif #ifdef VM_FREELIST_DMA32 if ( #ifdef VM_DMA32_NPAGES_THRESHOLD /* * Create the DMA32 free list only if the amount of * physical memory above physical address 4G exceeds the * given threshold. */ npages > VM_DMA32_NPAGES_THRESHOLD && #endif seg->end <= VM_DMA32_BOUNDARY) vm_freelist_to_flind[VM_FREELIST_DMA32] = 1; else #endif { #ifdef VM_DMA32_NPAGES_THRESHOLD npages += atop(seg->end - seg->start); #endif vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1; } } /* Change each entry into a running total of the free lists. */ for (freelist = 1; freelist < VM_NFREELIST; freelist++) { vm_freelist_to_flind[freelist] += vm_freelist_to_flind[freelist - 1]; } vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1]; KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists")); /* Change each entry into a free list index. */ for (freelist = 0; freelist < VM_NFREELIST; freelist++) vm_freelist_to_flind[freelist]--; /* * Initialize the first_page and free_queues fields of each physical * memory segment. */ #ifdef VM_PHYSSEG_SPARSE npages = 0; #endif for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; #ifdef VM_PHYSSEG_SPARSE seg->first_page = &vm_page_array[npages]; npages += atop(seg->end - seg->start); #else seg->first_page = PHYS_TO_VM_PAGE(seg->start); #endif #ifdef VM_FREELIST_LOWMEM if (seg->end <= VM_LOWMEM_BOUNDARY) { flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM]; KASSERT(flind >= 0, ("vm_phys_init: LOWMEM flind < 0")); } else #endif #ifdef VM_FREELIST_DMA32 if (seg->end <= VM_DMA32_BOUNDARY) { flind = vm_freelist_to_flind[VM_FREELIST_DMA32]; KASSERT(flind >= 0, ("vm_phys_init: DMA32 flind < 0")); } else #endif { flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT]; KASSERT(flind >= 0, ("vm_phys_init: DEFAULT flind < 0")); } seg->free_queues = &vm_phys_free_queues[seg->domain][flind]; } /* * Coalesce physical memory segments that are contiguous and share the * same per-domain free queues. */ prev_seg = vm_phys_segs; seg = &vm_phys_segs[1]; end_seg = &vm_phys_segs[vm_phys_nsegs]; while (seg < end_seg) { if (prev_seg->end == seg->start && prev_seg->free_queues == seg->free_queues) { prev_seg->end = seg->end; KASSERT(prev_seg->domain == seg->domain, ("vm_phys_init: free queues cannot span domains")); vm_phys_nsegs--; end_seg--; for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++) *tmp_seg = *(tmp_seg + 1); } else { prev_seg = seg; seg++; } } /* * Initialize the free queues. */ for (dom = 0; dom < vm_ndomains; dom++) { for (flind = 0; flind < vm_nfreelists; flind++) { for (pind = 0; pind < VM_NFREEPOOL; pind++) { fl = vm_phys_free_queues[dom][flind][pind]; for (oind = 0; oind < VM_NFREEORDER; oind++) TAILQ_INIT(&fl[oind].pl); } } } rw_init(&vm_phys_fictitious_reg_lock, "vmfctr"); } /* * Register info about the NUMA topology of the system. * * Invoked by platform-dependent code prior to vm_phys_init(). */ void vm_phys_register_domains(int ndomains, struct mem_affinity *affinity, int *locality) { #ifdef NUMA int d, i; /* * For now the only override value that we support is 1, which * effectively disables NUMA-awareness in the allocators. */ d = 0; TUNABLE_INT_FETCH("vm.numa.disabled", &d); if (d) ndomains = 1; if (ndomains > 1) { vm_ndomains = ndomains; mem_affinity = affinity; mem_locality = locality; } for (i = 0; i < vm_ndomains; i++) DOMAINSET_SET(i, &all_domains); #else (void)ndomains; (void)affinity; (void)locality; #endif } /* * Split a contiguous, power of two-sized set of physical pages. * * When this function is called by a page allocation function, the caller * should request insertion at the head unless the order [order, oind) queues * are known to be empty. The objective being to reduce the likelihood of * long-term fragmentation by promoting contemporaneous allocation and * (hopefully) deallocation. */ static __inline void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order, int tail) { vm_page_t m_buddy; while (oind > order) { oind--; m_buddy = &m[1 << oind]; KASSERT(m_buddy->order == VM_NFREEORDER, ("vm_phys_split_pages: page %p has unexpected order %d", m_buddy, m_buddy->order)); vm_freelist_add(fl, m_buddy, oind, tail); } } /* * Add the physical pages [m, m + npages) at the end of a power-of-two aligned * and sized set to the specified free list. * * When this function is called by a page allocation function, the caller * should request insertion at the head unless the lower-order queues are * known to be empty. The objective being to reduce the likelihood of long- * term fragmentation by promoting contemporaneous allocation and (hopefully) * deallocation. * * The physical page m's buddy must not be free. */ static void vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail) { u_int n; int order; KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0")); KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) & ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0, ("vm_phys_enq_range: page %p and npages %u are misaligned", m, npages)); do { KASSERT(m->order == VM_NFREEORDER, ("vm_phys_enq_range: page %p has unexpected order %d", m, m->order)); order = ffs(npages) - 1; KASSERT(order < VM_NFREEORDER, ("vm_phys_enq_range: order %d is out of range", order)); vm_freelist_add(fl, m, order, tail); n = 1 << order; m += n; npages -= n; } while (npages > 0); } /* * Set the pool for a contiguous, power of two-sized set of physical pages. */ static void vm_phys_set_pool(int pool, vm_page_t m, int order) { vm_page_t m_tmp; for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++) m_tmp->pool = pool; } /* * Tries to allocate the specified number of pages from the specified pool * within the specified domain. Returns the actual number of allocated pages * and a pointer to each page through the array ma[]. * * The returned pages may not be physically contiguous. However, in contrast * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0), * calling this function once to allocate the desired number of pages will * avoid wasted time in vm_phys_split_pages(). * * The free page queues for the specified domain must be locked. */ int vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[]) { struct vm_freelist *alt, *fl; vm_page_t m; int avail, end, flind, freelist, i, need, oind, pind; KASSERT(domain >= 0 && domain < vm_ndomains, ("vm_phys_alloc_npages: domain %d is out of range", domain)); KASSERT(pool < VM_NFREEPOOL, ("vm_phys_alloc_npages: pool %d is out of range", pool)); KASSERT(npages <= 1 << (VM_NFREEORDER - 1), ("vm_phys_alloc_npages: npages %d is out of range", npages)); vm_domain_free_assert_locked(VM_DOMAIN(domain)); i = 0; for (freelist = 0; freelist < VM_NFREELIST; freelist++) { flind = vm_freelist_to_flind[freelist]; if (flind < 0) continue; fl = vm_phys_free_queues[domain][flind][pool]; for (oind = 0; oind < VM_NFREEORDER; oind++) { while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) { vm_freelist_rem(fl, m, oind); avail = 1 << oind; need = imin(npages - i, avail); for (end = i + need; i < end;) ma[i++] = m++; if (need < avail) { /* * Return excess pages to fl. Its * order [0, oind) queues are empty. */ vm_phys_enq_range(m, avail - need, fl, 1); return (npages); } else if (i == npages) return (npages); } } for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) { for (pind = 0; pind < VM_NFREEPOOL; pind++) { alt = vm_phys_free_queues[domain][flind][pind]; while ((m = TAILQ_FIRST(&alt[oind].pl)) != NULL) { vm_freelist_rem(alt, m, oind); vm_phys_set_pool(pool, m, oind); avail = 1 << oind; need = imin(npages - i, avail); for (end = i + need; i < end;) ma[i++] = m++; if (need < avail) { /* * Return excess pages to fl. * Its order [0, oind) queues * are empty. */ vm_phys_enq_range(m, avail - need, fl, 1); return (npages); } else if (i == npages) return (npages); } } } } return (i); } /* * Allocate a contiguous, power of two-sized set of physical pages * from the free lists. * * The free page queues must be locked. */ vm_page_t vm_phys_alloc_pages(int domain, int pool, int order) { vm_page_t m; int freelist; for (freelist = 0; freelist < VM_NFREELIST; freelist++) { m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order); if (m != NULL) return (m); } return (NULL); } /* * Allocate a contiguous, power of two-sized set of physical pages from the * specified free list. The free list must be specified using one of the * manifest constants VM_FREELIST_*. * * The free page queues must be locked. */ vm_page_t vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order) { struct vm_freelist *alt, *fl; vm_page_t m; int oind, pind, flind; KASSERT(domain >= 0 && domain < vm_ndomains, ("vm_phys_alloc_freelist_pages: domain %d is out of range", domain)); KASSERT(freelist < VM_NFREELIST, ("vm_phys_alloc_freelist_pages: freelist %d is out of range", freelist)); KASSERT(pool < VM_NFREEPOOL, ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool)); KASSERT(order < VM_NFREEORDER, ("vm_phys_alloc_freelist_pages: order %d is out of range", order)); flind = vm_freelist_to_flind[freelist]; /* Check if freelist is present */ if (flind < 0) return (NULL); vm_domain_free_assert_locked(VM_DOMAIN(domain)); fl = &vm_phys_free_queues[domain][flind][pool][0]; for (oind = order; oind < VM_NFREEORDER; oind++) { m = TAILQ_FIRST(&fl[oind].pl); if (m != NULL) { vm_freelist_rem(fl, m, oind); /* The order [order, oind) queues are empty. */ vm_phys_split_pages(m, oind, fl, order, 1); return (m); } } /* * The given pool was empty. Find the largest * contiguous, power-of-two-sized set of pages in any * pool. Transfer these pages to the given pool, and * use them to satisfy the allocation. */ for (oind = VM_NFREEORDER - 1; oind >= order; oind--) { for (pind = 0; pind < VM_NFREEPOOL; pind++) { alt = &vm_phys_free_queues[domain][flind][pind][0]; m = TAILQ_FIRST(&alt[oind].pl); if (m != NULL) { vm_freelist_rem(alt, m, oind); vm_phys_set_pool(pool, m, oind); /* The order [order, oind) queues are empty. */ vm_phys_split_pages(m, oind, fl, order, 1); return (m); } } } return (NULL); } /* * Find the vm_page corresponding to the given physical address. */ vm_page_t vm_phys_paddr_to_vm_page(vm_paddr_t pa) { struct vm_phys_seg *seg; - int segind; - for (segind = 0; segind < vm_phys_nsegs; segind++) { - seg = &vm_phys_segs[segind]; - if (pa >= seg->start && pa < seg->end) - return (&seg->first_page[atop(pa - seg->start)]); - } + if ((seg = vm_phys_paddr_to_seg(pa)) != NULL) + return (&seg->first_page[atop(pa - seg->start)]); return (NULL); } vm_page_t vm_phys_fictitious_to_vm_page(vm_paddr_t pa) { struct vm_phys_fictitious_seg tmp, *seg; vm_page_t m; m = NULL; tmp.start = pa; tmp.end = 0; rw_rlock(&vm_phys_fictitious_reg_lock); seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp); rw_runlock(&vm_phys_fictitious_reg_lock); if (seg == NULL) return (NULL); m = &seg->first_page[atop(pa - seg->start)]; KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m)); return (m); } static inline void vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start, long page_count, vm_memattr_t memattr) { long i; bzero(range, page_count * sizeof(*range)); for (i = 0; i < page_count; i++) { vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr); range[i].oflags &= ~VPO_UNMANAGED; range[i].busy_lock = VPB_UNBUSIED; } } int vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end, vm_memattr_t memattr) { struct vm_phys_fictitious_seg *seg; vm_page_t fp; long page_count; #ifdef VM_PHYSSEG_DENSE long pi, pe; long dpage_count; #endif KASSERT(start < end, ("Start of segment isn't less than end (start: %jx end: %jx)", (uintmax_t)start, (uintmax_t)end)); page_count = (end - start) / PAGE_SIZE; #ifdef VM_PHYSSEG_DENSE pi = atop(start); pe = atop(end); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { fp = &vm_page_array[pi - first_page]; if ((pe - first_page) > vm_page_array_size) { /* * We have a segment that starts inside * of vm_page_array, but ends outside of it. * * Use vm_page_array pages for those that are * inside of the vm_page_array range, and * allocate the remaining ones. */ dpage_count = vm_page_array_size - (pi - first_page); vm_phys_fictitious_init_range(fp, start, dpage_count, memattr); page_count -= dpage_count; start += ptoa(dpage_count); goto alloc; } /* * We can allocate the full range from vm_page_array, * so there's no need to register the range in the tree. */ vm_phys_fictitious_init_range(fp, start, page_count, memattr); return (0); } else if (pe > first_page && (pe - first_page) < vm_page_array_size) { /* * We have a segment that ends inside of vm_page_array, * but starts outside of it. */ fp = &vm_page_array[0]; dpage_count = pe - first_page; vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count, memattr); end -= ptoa(dpage_count); page_count -= dpage_count; goto alloc; } else if (pi < first_page && pe > (first_page + vm_page_array_size)) { /* * Trying to register a fictitious range that expands before * and after vm_page_array. */ return (EINVAL); } else { alloc: #endif fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES, M_WAITOK); #ifdef VM_PHYSSEG_DENSE } #endif vm_phys_fictitious_init_range(fp, start, page_count, memattr); seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO); seg->start = start; seg->end = end; seg->first_page = fp; rw_wlock(&vm_phys_fictitious_reg_lock); RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg); rw_wunlock(&vm_phys_fictitious_reg_lock); return (0); } void vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end) { struct vm_phys_fictitious_seg *seg, tmp; #ifdef VM_PHYSSEG_DENSE long pi, pe; #endif KASSERT(start < end, ("Start of segment isn't less than end (start: %jx end: %jx)", (uintmax_t)start, (uintmax_t)end)); #ifdef VM_PHYSSEG_DENSE pi = atop(start); pe = atop(end); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { if ((pe - first_page) <= vm_page_array_size) { /* * This segment was allocated using vm_page_array * only, there's nothing to do since those pages * were never added to the tree. */ return; } /* * We have a segment that starts inside * of vm_page_array, but ends outside of it. * * Calculate how many pages were added to the * tree and free them. */ start = ptoa(first_page + vm_page_array_size); } else if (pe > first_page && (pe - first_page) < vm_page_array_size) { /* * We have a segment that ends inside of vm_page_array, * but starts outside of it. */ end = ptoa(first_page); } else if (pi < first_page && pe > (first_page + vm_page_array_size)) { /* Since it's not possible to register such a range, panic. */ panic( "Unregistering not registered fictitious range [%#jx:%#jx]", (uintmax_t)start, (uintmax_t)end); } #endif tmp.start = start; tmp.end = 0; rw_wlock(&vm_phys_fictitious_reg_lock); seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp); if (seg->start != start || seg->end != end) { rw_wunlock(&vm_phys_fictitious_reg_lock); panic( "Unregistering not registered fictitious range [%#jx:%#jx]", (uintmax_t)start, (uintmax_t)end); } RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg); rw_wunlock(&vm_phys_fictitious_reg_lock); free(seg->first_page, M_FICT_PAGES); free(seg, M_FICT_PAGES); } /* * Free a contiguous, power of two-sized set of physical pages. * * The free page queues must be locked. */ void vm_phys_free_pages(vm_page_t m, int order) { struct vm_freelist *fl; struct vm_phys_seg *seg; vm_paddr_t pa; vm_page_t m_buddy; KASSERT(m->order == VM_NFREEORDER, ("vm_phys_free_pages: page %p has unexpected order %d", m, m->order)); KASSERT(m->pool < VM_NFREEPOOL, ("vm_phys_free_pages: page %p has unexpected pool %d", m, m->pool)); KASSERT(order < VM_NFREEORDER, ("vm_phys_free_pages: order %d is out of range", order)); seg = &vm_phys_segs[m->segind]; vm_domain_free_assert_locked(VM_DOMAIN(seg->domain)); if (order < VM_NFREEORDER - 1) { pa = VM_PAGE_TO_PHYS(m); do { pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order)); if (pa < seg->start || pa >= seg->end) break; m_buddy = &seg->first_page[atop(pa - seg->start)]; if (m_buddy->order != order) break; fl = (*seg->free_queues)[m_buddy->pool]; vm_freelist_rem(fl, m_buddy, order); if (m_buddy->pool != m->pool) vm_phys_set_pool(m->pool, m_buddy, order); order++; pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1); m = &seg->first_page[atop(pa - seg->start)]; } while (order < VM_NFREEORDER - 1); } fl = (*seg->free_queues)[m->pool]; vm_freelist_add(fl, m, order, 1); } /* * Return the largest possible order of a set of pages starting at m. */ static int max_order(vm_page_t m) { /* * Unsigned "min" is used here so that "order" is assigned * "VM_NFREEORDER - 1" when "m"'s physical address is zero * or the low-order bits of its physical address are zero * because the size of a physical address exceeds the size of * a long. */ return (min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1, VM_NFREEORDER - 1)); } /* * Free a contiguous, arbitrarily sized set of physical pages, without * merging across set boundaries. * * The free page queues must be locked. */ void vm_phys_enqueue_contig(vm_page_t m, u_long npages) { struct vm_freelist *fl; struct vm_phys_seg *seg; vm_page_t m_end; int order; /* * Avoid unnecessary coalescing by freeing the pages in the largest * possible power-of-two-sized subsets. */ vm_domain_free_assert_locked(vm_pagequeue_domain(m)); seg = &vm_phys_segs[m->segind]; fl = (*seg->free_queues)[m->pool]; m_end = m + npages; /* Free blocks of increasing size. */ while ((order = max_order(m)) < VM_NFREEORDER - 1 && m + (1 << order) <= m_end) { KASSERT(seg == &vm_phys_segs[m->segind], ("%s: page range [%p,%p) spans multiple segments", __func__, m_end - npages, m)); vm_freelist_add(fl, m, order, 1); m += 1 << order; } /* Free blocks of maximum size. */ while (m + (1 << order) <= m_end) { KASSERT(seg == &vm_phys_segs[m->segind], ("%s: page range [%p,%p) spans multiple segments", __func__, m_end - npages, m)); vm_freelist_add(fl, m, order, 1); m += 1 << order; } /* Free blocks of diminishing size. */ while (m < m_end) { KASSERT(seg == &vm_phys_segs[m->segind], ("%s: page range [%p,%p) spans multiple segments", __func__, m_end - npages, m)); order = flsl(m_end - m) - 1; vm_freelist_add(fl, m, order, 1); m += 1 << order; } } /* * Free a contiguous, arbitrarily sized set of physical pages. * * The free page queues must be locked. */ void vm_phys_free_contig(vm_page_t m, u_long npages) { int order_start, order_end; vm_page_t m_start, m_end; vm_domain_free_assert_locked(vm_pagequeue_domain(m)); m_start = m; order_start = max_order(m_start); if (order_start < VM_NFREEORDER - 1) m_start += 1 << order_start; m_end = m + npages; order_end = max_order(m_end); if (order_end < VM_NFREEORDER - 1) m_end -= 1 << order_end; /* * Avoid unnecessary coalescing by freeing the pages at the start and * end of the range last. */ if (m_start < m_end) vm_phys_enqueue_contig(m_start, m_end - m_start); if (order_start < VM_NFREEORDER - 1) vm_phys_free_pages(m, order_start); if (order_end < VM_NFREEORDER - 1) vm_phys_free_pages(m_end, order_end); } /* - * Scan physical memory between the specified addresses "low" and "high" for a - * run of contiguous physical pages that satisfy the specified conditions, and - * return the lowest page in the run. The specified "alignment" determines - * the alignment of the lowest physical page in the run. If the specified - * "boundary" is non-zero, then the run of physical pages cannot span a - * physical address that is a multiple of "boundary". - * - * "npages" must be greater than zero. Both "alignment" and "boundary" must - * be a power of two. + * Identify the first address range within segment segind or greater + * that matches the domain, lies within the low/high range, and has + * enough pages. Return -1 if there is none. */ -vm_page_t -vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high, - u_long alignment, vm_paddr_t boundary, int options) +int +vm_phys_find_range(vm_page_t bounds[], int segind, int domain, + u_long npages, vm_paddr_t low, vm_paddr_t high) { - vm_paddr_t pa_end; - vm_page_t m_end, m_run, m_start; - struct vm_phys_seg *seg; - int segind; + vm_paddr_t pa_end, pa_start; + struct vm_phys_seg *end_seg, *seg; - KASSERT(npages > 0, ("npages is 0")); - KASSERT(powerof2(alignment), ("alignment is not a power of 2")); - KASSERT(powerof2(boundary), ("boundary is not a power of 2")); - if (low >= high) - return (NULL); - for (segind = 0; segind < vm_phys_nsegs; segind++) { - seg = &vm_phys_segs[segind]; + KASSERT(npages > 0, ("npages is zero")); + KASSERT(domain >= 0 && domain < vm_ndomain, ("domain out of range")); + end_seg = &vm_phys_segs[vm_phys_nsegs]; + for (seg = &vm_phys_segs[segind]; seg < end_seg; seg++) { if (seg->domain != domain) continue; if (seg->start >= high) - break; - if (low >= seg->end) - continue; - if (low <= seg->start) - m_start = seg->first_page; - else - m_start = &seg->first_page[atop(low - seg->start)]; - if (high < seg->end) - pa_end = high; - else - pa_end = seg->end; - if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages)) + return (-1); + pa_start = MAX(low, seg->start); + pa_end = MIN(high, seg->end); + if (pa_end - pa_start < ptoa(npages)) continue; - m_end = &seg->first_page[atop(pa_end - seg->start)]; - m_run = vm_page_scan_contig(npages, m_start, m_end, - alignment, boundary, options); - if (m_run != NULL) - return (m_run); + bounds[0] = &seg->first_page[atop(pa_start - seg->start)]; + bounds[1] = &seg->first_page[atop(pa_end - seg->start)]; + return (seg - vm_phys_segs); } - return (NULL); + return (-1); } /* * Search for the given physical page "m" in the free lists. If the search * succeeds, remove "m" from the free lists and return true. Otherwise, return * false, indicating that "m" is not in the free lists. * * The free page queues must be locked. */ bool vm_phys_unfree_page(vm_page_t m) { struct vm_freelist *fl; struct vm_phys_seg *seg; vm_paddr_t pa, pa_half; vm_page_t m_set, m_tmp; int order; /* * First, find the contiguous, power of two-sized set of free * physical pages containing the given physical page "m" and * assign it to "m_set". */ seg = &vm_phys_segs[m->segind]; vm_domain_free_assert_locked(VM_DOMAIN(seg->domain)); for (m_set = m, order = 0; m_set->order == VM_NFREEORDER && order < VM_NFREEORDER - 1; ) { order++; pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order)); if (pa >= seg->start) m_set = &seg->first_page[atop(pa - seg->start)]; else return (false); } if (m_set->order < order) return (false); if (m_set->order == VM_NFREEORDER) return (false); KASSERT(m_set->order < VM_NFREEORDER, ("vm_phys_unfree_page: page %p has unexpected order %d", m_set, m_set->order)); /* * Next, remove "m_set" from the free lists. Finally, extract * "m" from "m_set" using an iterative algorithm: While "m_set" * is larger than a page, shrink "m_set" by returning the half * of "m_set" that does not contain "m" to the free lists. */ fl = (*seg->free_queues)[m_set->pool]; order = m_set->order; vm_freelist_rem(fl, m_set, order); while (order > 0) { order--; pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order)); if (m->phys_addr < pa_half) m_tmp = &seg->first_page[atop(pa_half - seg->start)]; else { m_tmp = m_set; m_set = &seg->first_page[atop(pa_half - seg->start)]; } vm_freelist_add(fl, m_tmp, order, 0); } KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency")); return (true); } /* * Find a run of contiguous physical pages from the specified page list. */ static vm_page_t vm_phys_find_freelist_contig(struct vm_freelist *fl, int oind, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_phys_seg *seg; vm_paddr_t frag, lbound, pa, page_size, pa_end, pa_pre, size; vm_page_t m, m_listed, m_ret; int order; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); /* Search for a run satisfying the specified conditions. */ page_size = PAGE_SIZE; size = npages << PAGE_SHIFT; frag = (npages & ~(~0UL << oind)) << PAGE_SHIFT; TAILQ_FOREACH(m_listed, &fl[oind].pl, listq) { /* * Determine if the address range starting at pa is * too low. */ pa = VM_PAGE_TO_PHYS(m_listed); if (pa < low) continue; /* * If this is not the first free oind-block in this range, bail * out. We have seen the first free block already, or will see * it before failing to find an appropriate range. */ seg = &vm_phys_segs[m_listed->segind]; lbound = low > seg->start ? low : seg->start; pa_pre = pa - (page_size << oind); m = &seg->first_page[atop(pa_pre - seg->start)]; if (pa != 0 && pa_pre >= lbound && m->order == oind) continue; if (!vm_addr_align_ok(pa, alignment)) /* Advance to satisfy alignment condition. */ pa = roundup2(pa, alignment); else if (frag != 0 && lbound + frag <= pa) { /* * Back up to the first aligned free block in this * range, without moving below lbound. */ pa_end = pa; for (order = oind - 1; order >= 0; order--) { pa_pre = pa_end - (page_size << order); if (!vm_addr_align_ok(pa_pre, alignment)) break; m = &seg->first_page[atop(pa_pre - seg->start)]; if (pa_pre >= lbound && m->order == order) pa_end = pa_pre; } /* * If the extra small blocks are enough to complete the * fragment, use them. Otherwise, look to allocate the * fragment at the other end. */ if (pa_end + frag <= pa) pa = pa_end; } /* Advance as necessary to satisfy boundary conditions. */ if (!vm_addr_bound_ok(pa, size, boundary)) pa = roundup2(pa + 1, boundary); pa_end = pa + size; /* * Determine if the address range is valid (without overflow in * pa_end calculation), and fits within the segment. */ if (pa_end < pa || seg->end < pa_end) continue; m_ret = &seg->first_page[atop(pa - seg->start)]; /* * Determine whether there are enough free oind-blocks here to * satisfy the allocation request. */ pa = VM_PAGE_TO_PHYS(m_listed); do { pa += page_size << oind; if (pa >= pa_end) return (m_ret); m = &seg->first_page[atop(pa - seg->start)]; } while (oind == m->order); /* * Determine if an additional series of free blocks of * diminishing size can help to satisfy the allocation request. */ while (m->order < oind && pa + 2 * (page_size << m->order) > pa_end) { pa += page_size << m->order; if (pa >= pa_end) return (m_ret); m = &seg->first_page[atop(pa - seg->start)]; } } return (NULL); } /* * Find a run of contiguous physical pages from the specified free list * table. */ static vm_page_t vm_phys_find_queues_contig( struct vm_freelist (*queues)[VM_NFREEPOOL][VM_NFREEORDER_MAX], u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_freelist *fl; vm_page_t m_ret; vm_paddr_t pa, pa_end, size; int oind, order, pind; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); /* Compute the queue that is the best fit for npages. */ order = flsl(npages - 1); /* Search for a large enough free block. */ size = npages << PAGE_SHIFT; for (oind = order; oind < VM_NFREEORDER; oind++) { for (pind = 0; pind < VM_NFREEPOOL; pind++) { fl = (*queues)[pind]; TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) { /* * Determine if the address range starting at pa * is within the given range, satisfies the * given alignment, and does not cross the given * boundary. */ pa = VM_PAGE_TO_PHYS(m_ret); pa_end = pa + size; if (low <= pa && pa_end <= high && vm_addr_ok(pa, size, alignment, boundary)) return (m_ret); } } } if (order < VM_NFREEORDER) return (NULL); /* Search for a long-enough sequence of small blocks. */ oind = VM_NFREEORDER - 1; for (pind = 0; pind < VM_NFREEPOOL; pind++) { fl = (*queues)[pind]; m_ret = vm_phys_find_freelist_contig(fl, oind, npages, low, high, alignment, boundary); if (m_ret != NULL) return (m_ret); } return (NULL); } /* * Allocate a contiguous set of physical pages of the given size * "npages" from the free lists. All of the physical pages must be at * or above the given physical address "low" and below the given * physical address "high". The given value "alignment" determines the * alignment of the first physical page in the set. If the given value * "boundary" is non-zero, then the set of physical pages cannot cross * any physical address boundary that is a multiple of that value. Both * "alignment" and "boundary" must be a power of two. */ vm_page_t vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { vm_paddr_t pa_end, pa_start; struct vm_freelist *fl; vm_page_t m, m_run; struct vm_phys_seg *seg; struct vm_freelist (*queues)[VM_NFREEPOOL][VM_NFREEORDER_MAX]; int oind, segind; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); vm_domain_free_assert_locked(VM_DOMAIN(domain)); if (low >= high) return (NULL); queues = NULL; m_run = NULL; for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) { seg = &vm_phys_segs[segind]; if (seg->start >= high || seg->domain != domain) continue; if (low >= seg->end) break; if (low <= seg->start) pa_start = seg->start; else pa_start = low; if (high < seg->end) pa_end = high; else pa_end = seg->end; if (pa_end - pa_start < ptoa(npages)) continue; /* * If a previous segment led to a search using * the same free lists as would this segment, then * we've actually already searched within this * too. So skip it. */ if (seg->free_queues == queues) continue; queues = seg->free_queues; m_run = vm_phys_find_queues_contig(queues, npages, low, high, alignment, boundary); if (m_run != NULL) break; } if (m_run == NULL) return (NULL); /* Allocate pages from the page-range found. */ for (m = m_run; m < &m_run[npages]; m = &m[1 << oind]) { fl = (*queues)[m->pool]; oind = m->order; vm_freelist_rem(fl, m, oind); if (m->pool != VM_FREEPOOL_DEFAULT) vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind); } /* Return excess pages to the free lists. */ if (&m_run[npages] < m) { fl = (*queues)[VM_FREEPOOL_DEFAULT]; vm_phys_enq_range(&m_run[npages], m - &m_run[npages], fl, 0); } return (m_run); } /* * Return the index of the first unused slot which may be the terminating * entry. */ static int vm_phys_avail_count(void) { int i; for (i = 0; phys_avail[i + 1]; i += 2) continue; if (i > PHYS_AVAIL_ENTRIES) panic("Improperly terminated phys_avail %d entries", i); return (i); } /* * Assert that a phys_avail entry is valid. */ static void vm_phys_avail_check(int i) { if (phys_avail[i] & PAGE_MASK) panic("Unaligned phys_avail[%d]: %#jx", i, (intmax_t)phys_avail[i]); if (phys_avail[i+1] & PAGE_MASK) panic("Unaligned phys_avail[%d + 1]: %#jx", i, (intmax_t)phys_avail[i]); if (phys_avail[i + 1] < phys_avail[i]) panic("phys_avail[%d] start %#jx < end %#jx", i, (intmax_t)phys_avail[i], (intmax_t)phys_avail[i+1]); } /* * Return the index of an overlapping phys_avail entry or -1. */ #ifdef NUMA static int vm_phys_avail_find(vm_paddr_t pa) { int i; for (i = 0; phys_avail[i + 1]; i += 2) if (phys_avail[i] <= pa && phys_avail[i + 1] > pa) return (i); return (-1); } #endif /* * Return the index of the largest entry. */ int vm_phys_avail_largest(void) { vm_paddr_t sz, largesz; int largest; int i; largest = 0; largesz = 0; for (i = 0; phys_avail[i + 1]; i += 2) { sz = vm_phys_avail_size(i); if (sz > largesz) { largesz = sz; largest = i; } } return (largest); } vm_paddr_t vm_phys_avail_size(int i) { return (phys_avail[i + 1] - phys_avail[i]); } /* * Split an entry at the address 'pa'. Return zero on success or errno. */ static int vm_phys_avail_split(vm_paddr_t pa, int i) { int cnt; vm_phys_avail_check(i); if (pa <= phys_avail[i] || pa >= phys_avail[i + 1]) panic("vm_phys_avail_split: invalid address"); cnt = vm_phys_avail_count(); if (cnt >= PHYS_AVAIL_ENTRIES) return (ENOSPC); memmove(&phys_avail[i + 2], &phys_avail[i], (cnt - i) * sizeof(phys_avail[0])); phys_avail[i + 1] = pa; phys_avail[i + 2] = pa; vm_phys_avail_check(i); vm_phys_avail_check(i+2); return (0); } /* * Check if a given physical address can be included as part of a crash dump. */ bool vm_phys_is_dumpable(vm_paddr_t pa) { vm_page_t m; int i; if ((m = vm_phys_paddr_to_vm_page(pa)) != NULL) return ((m->flags & PG_NODUMP) == 0); for (i = 0; dump_avail[i] != 0 || dump_avail[i + 1] != 0; i += 2) { if (pa >= dump_avail[i] && pa < dump_avail[i + 1]) return (true); } return (false); } void vm_phys_early_add_seg(vm_paddr_t start, vm_paddr_t end) { struct vm_phys_seg *seg; if (vm_phys_early_nsegs == -1) panic("%s: called after initialization", __func__); if (vm_phys_early_nsegs == nitems(vm_phys_early_segs)) panic("%s: ran out of early segments", __func__); seg = &vm_phys_early_segs[vm_phys_early_nsegs++]; seg->start = start; seg->end = end; } /* * This routine allocates NUMA node specific memory before the page * allocator is bootstrapped. */ vm_paddr_t vm_phys_early_alloc(int domain, size_t alloc_size) { #ifdef NUMA int mem_index; #endif int i, biggestone; vm_paddr_t pa, mem_start, mem_end, size, biggestsize, align; KASSERT(domain == -1 || (domain >= 0 && domain < vm_ndomains), ("%s: invalid domain index %d", __func__, domain)); /* * Search the mem_affinity array for the biggest address * range in the desired domain. This is used to constrain * the phys_avail selection below. */ biggestsize = 0; mem_start = 0; mem_end = -1; #ifdef NUMA mem_index = 0; if (mem_affinity != NULL) { for (i = 0;; i++) { size = mem_affinity[i].end - mem_affinity[i].start; if (size == 0) break; if (domain != -1 && mem_affinity[i].domain != domain) continue; if (size > biggestsize) { mem_index = i; biggestsize = size; } } mem_start = mem_affinity[mem_index].start; mem_end = mem_affinity[mem_index].end; } #endif /* * Now find biggest physical segment in within the desired * numa domain. */ biggestsize = 0; biggestone = 0; for (i = 0; phys_avail[i + 1] != 0; i += 2) { /* skip regions that are out of range */ if (phys_avail[i+1] - alloc_size < mem_start || phys_avail[i+1] > mem_end) continue; size = vm_phys_avail_size(i); if (size > biggestsize) { biggestone = i; biggestsize = size; } } alloc_size = round_page(alloc_size); /* * Grab single pages from the front to reduce fragmentation. */ if (alloc_size == PAGE_SIZE) { pa = phys_avail[biggestone]; phys_avail[biggestone] += PAGE_SIZE; vm_phys_avail_check(biggestone); return (pa); } /* * Naturally align large allocations. */ align = phys_avail[biggestone + 1] & (alloc_size - 1); if (alloc_size + align > biggestsize) panic("cannot find a large enough size\n"); if (align != 0 && vm_phys_avail_split(phys_avail[biggestone + 1] - align, biggestone) != 0) /* Wasting memory. */ phys_avail[biggestone + 1] -= align; phys_avail[biggestone + 1] -= alloc_size; vm_phys_avail_check(biggestone); pa = phys_avail[biggestone + 1]; return (pa); } void vm_phys_early_startup(void) { struct vm_phys_seg *seg; int i; for (i = 0; phys_avail[i + 1] != 0; i += 2) { phys_avail[i] = round_page(phys_avail[i]); phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); } for (i = 0; i < vm_phys_early_nsegs; i++) { seg = &vm_phys_early_segs[i]; vm_phys_add_seg(seg->start, seg->end); } vm_phys_early_nsegs = -1; #ifdef NUMA /* Force phys_avail to be split by domain. */ if (mem_affinity != NULL) { int idx; for (i = 0; mem_affinity[i].end != 0; i++) { idx = vm_phys_avail_find(mem_affinity[i].start); if (idx != -1 && phys_avail[idx] != mem_affinity[i].start) vm_phys_avail_split(mem_affinity[i].start, idx); idx = vm_phys_avail_find(mem_affinity[i].end); if (idx != -1 && phys_avail[idx] != mem_affinity[i].end) vm_phys_avail_split(mem_affinity[i].end, idx); } } #endif } #ifdef DDB /* * Show the number of physical pages in each of the free lists. */ DB_SHOW_COMMAND_FLAGS(freepages, db_show_freepages, DB_CMD_MEMSAFE) { struct vm_freelist *fl; int flind, oind, pind, dom; for (dom = 0; dom < vm_ndomains; dom++) { db_printf("DOMAIN: %d\n", dom); for (flind = 0; flind < vm_nfreelists; flind++) { db_printf("FREE LIST %d:\n" "\n ORDER (SIZE) | NUMBER" "\n ", flind); for (pind = 0; pind < VM_NFREEPOOL; pind++) db_printf(" | POOL %d", pind); db_printf("\n-- "); for (pind = 0; pind < VM_NFREEPOOL; pind++) db_printf("-- -- "); db_printf("--\n"); for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) { db_printf(" %2.2d (%6.6dK)", oind, 1 << (PAGE_SHIFT - 10 + oind)); for (pind = 0; pind < VM_NFREEPOOL; pind++) { fl = vm_phys_free_queues[dom][flind][pind]; db_printf(" | %6.6d", fl[oind].lcnt); } db_printf("\n"); } db_printf("\n"); } db_printf("\n"); } } #endif diff --git a/sys/vm/vm_phys.h b/sys/vm/vm_phys.h index a294bbaad80a..fed20bbaae1e 100644 --- a/sys/vm/vm_phys.h +++ b/sys/vm/vm_phys.h @@ -1,110 +1,154 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2002-2006 Rice University * Copyright (c) 2007 Alan L. Cox * All rights reserved. * * This software was developed for the FreeBSD Project by Alan L. Cox, * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro. * * 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 COPYRIGHT HOLDERS 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 COPYRIGHT * HOLDERS 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. * * $FreeBSD$ */ /* * Physical memory system definitions */ #ifndef _VM_PHYS_H_ #define _VM_PHYS_H_ #ifdef _KERNEL +#include + extern vm_paddr_t phys_avail[]; /* Domains must be dense (non-sparse) and zero-based. */ struct mem_affinity { vm_paddr_t start; vm_paddr_t end; int domain; }; #ifdef NUMA extern struct mem_affinity *mem_affinity; extern int *mem_locality; #endif /* * The following functions are only to be used by the virtual memory system. */ void vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end); vm_page_t vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); vm_page_t vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order); int vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[]); vm_page_t vm_phys_alloc_pages(int domain, int pool, int order); int vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high); void vm_phys_enqueue_contig(vm_page_t m, u_long npages); int vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end, vm_memattr_t memattr); void vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end); vm_page_t vm_phys_fictitious_to_vm_page(vm_paddr_t pa); +int vm_phys_find_range(vm_page_t bounds[], int segind, int domain, + u_long npages, vm_paddr_t low, vm_paddr_t high); void vm_phys_free_contig(vm_page_t m, u_long npages); void vm_phys_free_pages(vm_page_t m, int order); void vm_phys_init(void); vm_page_t vm_phys_paddr_to_vm_page(vm_paddr_t pa); void vm_phys_register_domains(int ndomains, struct mem_affinity *affinity, int *locality); -vm_page_t vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, - vm_paddr_t high, u_long alignment, vm_paddr_t boundary, int options); bool vm_phys_unfree_page(vm_page_t m); int vm_phys_mem_affinity(int f, int t); void vm_phys_early_add_seg(vm_paddr_t start, vm_paddr_t end); vm_paddr_t vm_phys_early_alloc(int domain, size_t alloc_size); void vm_phys_early_startup(void); int vm_phys_avail_largest(void); vm_paddr_t vm_phys_avail_size(int i); bool vm_phys_is_dumpable(vm_paddr_t pa); static inline int vm_phys_domain(vm_paddr_t pa) { #ifdef NUMA int i; if (vm_ndomains == 1) return (0); for (i = 0; mem_affinity[i].end != 0; i++) if (mem_affinity[i].start <= pa && mem_affinity[i].end >= pa) return (mem_affinity[i].domain); return (-1); #else return (0); #endif } +/* + * Find the segind for the first segment at or after the given physical address. + */ +static inline int +vm_phys_lookup_segind(vm_paddr_t pa) +{ + u_int hi, lo, mid; + + lo = 0; + hi = vm_phys_nsegs; + while (lo != hi) { + /* + * for i in [0, lo), segs[i].end <= pa + * for i in [hi, nsegs), segs[i].end > pa + */ + mid = lo + (hi - lo) / 2; + if (vm_phys_segs[mid].end <= pa) + lo = mid + 1; + else + hi = mid; + } + return (lo); +} + +/* + * Find the segment corresponding to the given physical address. + */ +static inline struct vm_phys_seg * +vm_phys_paddr_to_seg(vm_paddr_t pa) +{ + struct vm_phys_seg *seg; + int segind; + + segind = vm_phys_lookup_segind(pa); + if (segind < vm_phys_nsegs) { + seg = &vm_phys_segs[segind]; + if (pa >= seg->start) + return (seg); + } + return (NULL); +} + #endif /* _KERNEL */ #endif /* !_VM_PHYS_H_ */