X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/89b3af67bb32e691275bf6fa803d1834b2284115..22ba694c5857e62b5a553b1505dcf2e509177f28:/osfmk/vm/vm_resident.c diff --git a/osfmk/vm/vm_resident.c b/osfmk/vm/vm_resident.c index 91b359f80..a9fd4d0b2 100644 --- a/osfmk/vm/vm_resident.c +++ b/osfmk/vm/vm_resident.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2000-2005 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2009 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * @@ -63,14 +63,17 @@ */ #include +#include #include #include #include +#include #include #include #include #include +#include #include #include #include @@ -82,17 +85,43 @@ #include #include #include -#include /* (BRINGUP) */ -#include /* (BRINGUP) */ +#include #include +#include +#include +#include + +#include + +#include + +boolean_t hibernate_cleaning_in_progress = FALSE; +boolean_t vm_page_free_verify = TRUE; + +uint32_t vm_lopage_free_count = 0; +uint32_t vm_lopage_free_limit = 0; +uint32_t vm_lopage_lowater = 0; +boolean_t vm_lopage_refill = FALSE; +boolean_t vm_lopage_needed = FALSE; + +lck_mtx_ext_t vm_page_queue_lock_ext; +lck_mtx_ext_t vm_page_queue_free_lock_ext; +lck_mtx_ext_t vm_purgeable_queue_lock_ext; + +int speculative_age_index = 0; +int speculative_steal_index = 0; +struct vm_speculative_age_q vm_page_queue_speculative[VM_PAGE_MAX_SPECULATIVE_AGE_Q + 1]; + + +__private_extern__ void vm_page_init_lck_grp(void); + +static void vm_page_free_prepare(vm_page_t page); +static vm_page_t vm_page_grab_fictitious_common(ppnum_t phys_addr); + + -/* Variables used to indicate the relative age of pages in the - * inactive list - */ -unsigned int vm_page_ticket_roll = 0; -unsigned int vm_page_ticket = 0; /* * Associated with page of user-allocatable memory is a * page structure. @@ -106,7 +135,7 @@ unsigned int vm_page_ticket = 0; vm_offset_t virtual_space_start; vm_offset_t virtual_space_end; -int vm_page_pages; +uint32_t vm_page_pages; /* * The vm_page_lookup() routine, which provides for fast @@ -124,16 +153,25 @@ typedef struct { #endif /* MACH_PAGE_HASH_STATS */ } vm_page_bucket_t; + +#define BUCKETS_PER_LOCK 16 + vm_page_bucket_t *vm_page_buckets; /* Array of buckets */ unsigned int vm_page_bucket_count = 0; /* How big is array? */ unsigned int vm_page_hash_mask; /* Mask for hash function */ unsigned int vm_page_hash_shift; /* Shift for hash function */ -uint32_t vm_page_bucket_hash; /* Basic bucket hash */ -decl_simple_lock_data(,vm_page_bucket_lock) +uint32_t vm_page_bucket_hash; /* Basic bucket hash */ +unsigned int vm_page_bucket_lock_count = 0; /* How big is array of locks? */ -vm_page_t -vm_page_lookup_nohint(vm_object_t object, vm_object_offset_t offset); +lck_spin_t *vm_page_bucket_locks; +#if VM_PAGE_BUCKETS_CHECK +boolean_t vm_page_buckets_check_ready = FALSE; +#if VM_PAGE_FAKE_BUCKETS +vm_page_bucket_t *vm_page_fake_buckets; /* decoy buckets */ +vm_map_offset_t vm_page_fake_buckets_start, vm_page_fake_buckets_end; +#endif /* VM_PAGE_FAKE_BUCKETS */ +#endif /* VM_PAGE_BUCKETS_CHECK */ #if MACH_PAGE_HASH_STATS /* This routine is only for debug. It is intended to be called by @@ -183,7 +221,7 @@ hash_debug(void) */ vm_size_t page_size = PAGE_SIZE; vm_size_t page_mask = PAGE_MASK; -int page_shift = PAGE_SHIFT; +int page_shift = PAGE_SHIFT; /* * Resident page structures are initialized from @@ -195,13 +233,21 @@ int page_shift = PAGE_SHIFT; */ struct vm_page vm_page_template; +vm_page_t vm_pages = VM_PAGE_NULL; +unsigned int vm_pages_count = 0; +ppnum_t vm_page_lowest = 0; + /* * Resident pages that represent real memory - * are allocated from a free list. + * are allocated from a set of free lists, + * one per color. */ -vm_page_t vm_page_queue_free; -vm_page_t vm_page_queue_fictitious; +unsigned int vm_colors; +unsigned int vm_color_mask; /* mask is == (vm_colors-1) */ +unsigned int vm_cache_geometry_colors = 0; /* set by hw dependent code during startup */ +queue_head_t vm_page_queue_free[MAX_COLORS]; unsigned int vm_page_free_wanted; +unsigned int vm_page_free_wanted_privileged; unsigned int vm_page_free_count; unsigned int vm_page_fictitious_count; @@ -217,16 +263,37 @@ unsigned int vm_page_free_count_minimum; /* debugging */ * most other kernel structures are. */ zone_t vm_page_zone; -decl_mutex_data(,vm_page_alloc_lock) +vm_locks_array_t vm_page_locks; +decl_lck_mtx_data(,vm_page_alloc_lock) +lck_mtx_ext_t vm_page_alloc_lock_ext; + unsigned int io_throttle_zero_fill; +unsigned int vm_page_local_q_count = 0; +unsigned int vm_page_local_q_soft_limit = 250; +unsigned int vm_page_local_q_hard_limit = 500; +struct vplq *vm_page_local_q = NULL; + +/* N.B. Guard and fictitious pages must not + * be assigned a zero phys_page value. + */ /* * Fictitious pages don't have a physical address, * but we must initialize phys_page to something. * For debugging, this should be a strange value * that the pmap module can recognize in assertions. */ -vm_offset_t vm_page_fictitious_addr = (vm_offset_t) -1; +ppnum_t vm_page_fictitious_addr = (ppnum_t) -1; + +/* + * Guard pages are not accessible so they don't + * need a physical address, but we need to enter + * one in the pmap. + * Let's make it recognizable and make sure that + * we don't use a real physical page with that + * physical address. + */ +ppnum_t vm_page_guard_addr = (ppnum_t) -2; /* * Resident page structures are also chained on @@ -234,26 +301,48 @@ vm_offset_t vm_page_fictitious_addr = (vm_offset_t) -1; * system (pageout daemon). These queues are * defined here, but are shared by the pageout * module. The inactive queue is broken into - * inactive and zf for convenience as the + * file backed and anonymous for convenience as the * pageout daemon often assignes a higher - * affinity to zf pages + * importance to anonymous pages (less likely to pick) */ queue_head_t vm_page_queue_active; queue_head_t vm_page_queue_inactive; +queue_head_t vm_page_queue_anonymous; /* inactive memory queue for anonymous pages */ +queue_head_t vm_page_queue_throttled; + unsigned int vm_page_active_count; unsigned int vm_page_inactive_count; +unsigned int vm_page_anonymous_count; +unsigned int vm_page_throttled_count; +unsigned int vm_page_speculative_count; unsigned int vm_page_wire_count; +unsigned int vm_page_wire_count_initial; unsigned int vm_page_gobble_count = 0; unsigned int vm_page_wire_count_warning = 0; unsigned int vm_page_gobble_count_warning = 0; unsigned int vm_page_purgeable_count = 0; /* # of pages purgeable now */ +unsigned int vm_page_purgeable_wired_count = 0; /* # of purgeable pages that are wired now */ uint64_t vm_page_purged_count = 0; /* total count of purged pages */ -ppnum_t vm_lopage_poolstart = 0; -ppnum_t vm_lopage_poolend = 0; -int vm_lopage_poolsize = 0; +unsigned int vm_page_external_count = 0; +unsigned int vm_page_internal_count = 0; +unsigned int vm_page_pageable_external_count = 0; +unsigned int vm_page_pageable_internal_count = 0; + +#if DEVELOPMENT || DEBUG +unsigned int vm_page_speculative_recreated = 0; +unsigned int vm_page_speculative_created = 0; +unsigned int vm_page_speculative_used = 0; +#endif + +queue_head_t vm_page_queue_cleaned; + +unsigned int vm_page_cleaned_count = 0; +unsigned int vm_pageout_enqueued_cleaned = 0; + uint64_t max_valid_dma_address = 0xffffffffffffffffULL; +ppnum_t max_valid_low_ppnum = 0xffffffff; /* @@ -264,9 +353,14 @@ uint64_t max_valid_dma_address = 0xffffffffffffffffULL; */ unsigned int vm_page_free_target = 0; unsigned int vm_page_free_min = 0; +unsigned int vm_page_throttle_limit = 0; +uint32_t vm_page_creation_throttle = 0; unsigned int vm_page_inactive_target = 0; +unsigned int vm_page_anonymous_min = 0; +unsigned int vm_page_inactive_min = 0; unsigned int vm_page_free_reserved = 0; -unsigned int vm_page_throttled_count = 0; +unsigned int vm_page_throttle_count = 0; + /* * The VM system has a couple of heuristics for deciding @@ -278,6 +372,8 @@ unsigned int vm_page_throttled_count = 0; boolean_t vm_page_deactivate_hint = TRUE; +struct vm_page_stats_reusable vm_page_stats_reusable; + /* * vm_set_page_size: * @@ -300,6 +396,95 @@ vm_set_page_size(void) break; } + +/* Called once during statup, once the cache geometry is known. + */ +static void +vm_page_set_colors( void ) +{ + unsigned int n, override; + + if ( PE_parse_boot_argn("colors", &override, sizeof (override)) ) /* colors specified as a boot-arg? */ + n = override; + else if ( vm_cache_geometry_colors ) /* do we know what the cache geometry is? */ + n = vm_cache_geometry_colors; + else n = DEFAULT_COLORS; /* use default if all else fails */ + + if ( n == 0 ) + n = 1; + if ( n > MAX_COLORS ) + n = MAX_COLORS; + + /* the count must be a power of 2 */ + if ( ( n & (n - 1)) != 0 ) + panic("vm_page_set_colors"); + + vm_colors = n; + vm_color_mask = n - 1; +} + + +lck_grp_t vm_page_lck_grp_free; +lck_grp_t vm_page_lck_grp_queue; +lck_grp_t vm_page_lck_grp_local; +lck_grp_t vm_page_lck_grp_purge; +lck_grp_t vm_page_lck_grp_alloc; +lck_grp_t vm_page_lck_grp_bucket; +lck_grp_attr_t vm_page_lck_grp_attr; +lck_attr_t vm_page_lck_attr; + + +__private_extern__ void +vm_page_init_lck_grp(void) +{ + /* + * initialze the vm_page lock world + */ + lck_grp_attr_setdefault(&vm_page_lck_grp_attr); + lck_grp_init(&vm_page_lck_grp_free, "vm_page_free", &vm_page_lck_grp_attr); + lck_grp_init(&vm_page_lck_grp_queue, "vm_page_queue", &vm_page_lck_grp_attr); + lck_grp_init(&vm_page_lck_grp_local, "vm_page_queue_local", &vm_page_lck_grp_attr); + lck_grp_init(&vm_page_lck_grp_purge, "vm_page_purge", &vm_page_lck_grp_attr); + lck_grp_init(&vm_page_lck_grp_alloc, "vm_page_alloc", &vm_page_lck_grp_attr); + lck_grp_init(&vm_page_lck_grp_bucket, "vm_page_bucket", &vm_page_lck_grp_attr); + lck_attr_setdefault(&vm_page_lck_attr); + lck_mtx_init_ext(&vm_page_alloc_lock, &vm_page_alloc_lock_ext, &vm_page_lck_grp_alloc, &vm_page_lck_attr); + + vm_compressor_init_locks(); +} + +void +vm_page_init_local_q() +{ + unsigned int num_cpus; + unsigned int i; + struct vplq *t_local_q; + + num_cpus = ml_get_max_cpus(); + + /* + * no point in this for a uni-processor system + */ + if (num_cpus >= 2) { + t_local_q = (struct vplq *)kalloc(num_cpus * sizeof(struct vplq)); + + for (i = 0; i < num_cpus; i++) { + struct vpl *lq; + + lq = &t_local_q[i].vpl_un.vpl; + VPL_LOCK_INIT(lq, &vm_page_lck_grp_local, &vm_page_lck_attr); + queue_init(&lq->vpl_queue); + lq->vpl_count = 0; + lq->vpl_internal_count = 0; + lq->vpl_external_count = 0; + } + vm_page_local_q_count = num_cpus; + + vm_page_local_q = (struct vplq *)t_local_q; + } +} + + /* * vm_page_bootstrap: * @@ -327,69 +512,115 @@ vm_page_bootstrap( */ m = &vm_page_template; - m->object = VM_OBJECT_NULL; /* reset later */ - m->offset = (vm_object_offset_t) -1; /* reset later */ - m->wire_count = 0; + bzero(m, sizeof (*m)); m->pageq.next = NULL; m->pageq.prev = NULL; m->listq.next = NULL; m->listq.prev = NULL; + m->next = VM_PAGE_NULL; + + m->object = VM_OBJECT_NULL; /* reset later */ + m->offset = (vm_object_offset_t) -1; /* reset later */ + m->wire_count = 0; + m->local = FALSE; m->inactive = FALSE; m->active = FALSE; + m->pageout_queue = FALSE; + m->speculative = FALSE; m->laundry = FALSE; m->free = FALSE; - m->no_isync = TRUE; m->reference = FALSE; - m->pageout = FALSE; - m->dump_cleaning = FALSE; - m->list_req_pending = FALSE; + m->gobbled = FALSE; + m->private = FALSE; + m->throttled = FALSE; + m->__unused_pageq_bits = 0; + + m->phys_page = 0; /* reset later */ m->busy = TRUE; m->wanted = FALSE; m->tabled = FALSE; + m->hashed = FALSE; m->fictitious = FALSE; - m->private = FALSE; + m->pmapped = FALSE; + m->wpmapped = FALSE; + m->pageout = FALSE; m->absent = FALSE; m->error = FALSE; m->dirty = FALSE; m->cleaning = FALSE; m->precious = FALSE; m->clustered = FALSE; - m->lock_supplied = FALSE; - m->unusual = FALSE; + m->overwriting = FALSE; m->restart = FALSE; - m->zero_fill = FALSE; + m->unusual = FALSE; m->encrypted = FALSE; - - m->phys_page = 0; /* reset later */ - - m->page_lock = VM_PROT_NONE; - m->unlock_request = VM_PROT_NONE; - m->page_error = KERN_SUCCESS; + m->encrypted_cleaning = FALSE; + m->cs_validated = FALSE; + m->cs_tainted = FALSE; + m->no_cache = FALSE; + m->reusable = FALSE; + m->slid = FALSE; + m->was_dirty = FALSE; + m->xpmapped = FALSE; + m->compressor = FALSE; + m->written_by_kernel = FALSE; + m->__unused_object_bits = 0; /* * Initialize the page queues. */ + vm_page_init_lck_grp(); + + lck_mtx_init_ext(&vm_page_queue_free_lock, &vm_page_queue_free_lock_ext, &vm_page_lck_grp_free, &vm_page_lck_attr); + lck_mtx_init_ext(&vm_page_queue_lock, &vm_page_queue_lock_ext, &vm_page_lck_grp_queue, &vm_page_lck_attr); + lck_mtx_init_ext(&vm_purgeable_queue_lock, &vm_purgeable_queue_lock_ext, &vm_page_lck_grp_purge, &vm_page_lck_attr); + + for (i = 0; i < PURGEABLE_Q_TYPE_MAX; i++) { + int group; + + purgeable_queues[i].token_q_head = 0; + purgeable_queues[i].token_q_tail = 0; + for (group = 0; group < NUM_VOLATILE_GROUPS; group++) + queue_init(&purgeable_queues[i].objq[group]); + + purgeable_queues[i].type = i; + purgeable_queues[i].new_pages = 0; +#if MACH_ASSERT + purgeable_queues[i].debug_count_tokens = 0; + purgeable_queues[i].debug_count_objects = 0; +#endif + }; + + for (i = 0; i < MAX_COLORS; i++ ) + queue_init(&vm_page_queue_free[i]); - mutex_init(&vm_page_queue_free_lock, 0); - mutex_init(&vm_page_queue_lock, 0); - - vm_page_queue_free = VM_PAGE_NULL; - vm_page_queue_fictitious = VM_PAGE_NULL; + queue_init(&vm_lopage_queue_free); queue_init(&vm_page_queue_active); queue_init(&vm_page_queue_inactive); - queue_init(&vm_page_queue_zf); + queue_init(&vm_page_queue_cleaned); + queue_init(&vm_page_queue_throttled); + queue_init(&vm_page_queue_anonymous); + + for ( i = 0; i <= VM_PAGE_MAX_SPECULATIVE_AGE_Q; i++ ) { + queue_init(&vm_page_queue_speculative[i].age_q); + vm_page_queue_speculative[i].age_ts.tv_sec = 0; + vm_page_queue_speculative[i].age_ts.tv_nsec = 0; + } vm_page_free_wanted = 0; + vm_page_free_wanted_privileged = 0; + + vm_page_set_colors(); + /* * Steal memory for the map and zone subsystems. */ - - vm_map_steal_memory(); zone_steal_memory(); + vm_map_steal_memory(); /* * Allocate (and initialize) the virtual-to-physical @@ -401,8 +632,6 @@ vm_page_bootstrap( * than the number of physical pages in the system. */ - simple_lock_init(&vm_page_bucket_lock, 0); - if (vm_page_bucket_count == 0) { unsigned int npages = pmap_free_pages(); @@ -410,6 +639,7 @@ vm_page_bootstrap( while (vm_page_bucket_count < npages) vm_page_bucket_count <<= 1; } + vm_page_bucket_lock_count = (vm_page_bucket_count + BUCKETS_PER_LOCK - 1) / BUCKETS_PER_LOCK; vm_page_hash_mask = vm_page_bucket_count - 1; @@ -435,10 +665,38 @@ vm_page_bootstrap( if (vm_page_hash_mask & vm_page_bucket_count) printf("vm_page_bootstrap: WARNING -- strange page hash\n"); +#if VM_PAGE_BUCKETS_CHECK +#if VM_PAGE_FAKE_BUCKETS + /* + * Allocate a decoy set of page buckets, to detect + * any stomping there. + */ + vm_page_fake_buckets = (vm_page_bucket_t *) + pmap_steal_memory(vm_page_bucket_count * + sizeof(vm_page_bucket_t)); + vm_page_fake_buckets_start = (vm_map_offset_t) vm_page_fake_buckets; + vm_page_fake_buckets_end = + vm_map_round_page((vm_page_fake_buckets_start + + (vm_page_bucket_count * + sizeof (vm_page_bucket_t))), + PAGE_MASK); + char *cp; + for (cp = (char *)vm_page_fake_buckets_start; + cp < (char *)vm_page_fake_buckets_end; + cp++) { + *cp = 0x5a; + } +#endif /* VM_PAGE_FAKE_BUCKETS */ +#endif /* VM_PAGE_BUCKETS_CHECK */ + vm_page_buckets = (vm_page_bucket_t *) pmap_steal_memory(vm_page_bucket_count * sizeof(vm_page_bucket_t)); + vm_page_bucket_locks = (lck_spin_t *) + pmap_steal_memory(vm_page_bucket_lock_count * + sizeof(lck_spin_t)); + for (i = 0; i < vm_page_bucket_count; i++) { register vm_page_bucket_t *bucket = &vm_page_buckets[i]; @@ -449,6 +707,13 @@ vm_page_bootstrap( #endif /* MACH_PAGE_HASH_STATS */ } + for (i = 0; i < vm_page_bucket_lock_count; i++) + lck_spin_init(&vm_page_bucket_locks[i], &vm_page_lck_grp_bucket, &vm_page_lck_attr); + +#if VM_PAGE_BUCKETS_CHECK + vm_page_buckets_check_ready = TRUE; +#endif /* VM_PAGE_BUCKETS_CHECK */ + /* * Machine-dependent code allocates the resident page table. * It uses vm_page_init to initialize the page frames. @@ -471,11 +736,14 @@ vm_page_bootstrap( * wired, they nonetheless can't be moved. At this moment, * all VM managed pages are "free", courtesy of pmap_startup. */ - vm_page_wire_count = atop_64(max_mem) - vm_page_free_count; /* initial value */ - - printf("vm_page_bootstrap: %d free pages\n", vm_page_free_count); + assert((unsigned int) atop_64(max_mem) == atop_64(max_mem)); + vm_page_wire_count = ((unsigned int) atop_64(max_mem)) - vm_page_free_count - vm_lopage_free_count; /* initial value */ + vm_page_wire_count_initial = vm_page_wire_count; vm_page_free_count_minimum = vm_page_free_count; + printf("vm_page_bootstrap: %d free pages and %d wired pages\n", + vm_page_free_count, vm_page_wire_count); + simple_lock_init(&vm_paging_lock, 0); } @@ -522,7 +790,7 @@ pmap_steal_memory( addr = virtual_space_start; virtual_space_start += size; - kprintf("pmap_steal_memory: %08X - %08X; size=%08X\n", addr, virtual_space_start, size); /* (TEST/DEBUG) */ + //kprintf("pmap_steal_memory: %08lX - %08lX; size=%08lX\n", (long)addr, (long)virtual_space_start, (long)size); /* (TEST/DEBUG) */ /* * Allocate and map physical pages to back new virtual pages. @@ -531,16 +799,20 @@ pmap_steal_memory( for (vaddr = round_page(addr); vaddr < addr + size; vaddr += PAGE_SIZE) { - if (!pmap_next_page(&phys_page)) + + if (!pmap_next_page_hi(&phys_page)) panic("pmap_steal_memory"); /* * XXX Logically, these mappings should be wired, * but some pmap modules barf if they are. */ +#if defined(__LP64__) + pmap_pre_expand(kernel_pmap, vaddr); +#endif pmap_enter(kernel_pmap, vaddr, phys_page, - VM_PROT_READ|VM_PROT_WRITE, + VM_PROT_READ|VM_PROT_WRITE, VM_PROT_NONE, VM_WIMG_USE_DEFAULT, FALSE); /* * Account for newly stolen memory @@ -558,11 +830,8 @@ pmap_startup( vm_offset_t *endp) { unsigned int i, npages, pages_initialized, fill, fillval; - vm_page_t pages; ppnum_t phys_page; addr64_t tmpaddr; - unsigned int num_of_lopages = 0; - unsigned int last_index; /* * We calculate how many page frames we will have @@ -570,10 +839,10 @@ pmap_startup( */ tmpaddr = (addr64_t)pmap_free_pages() * (addr64_t)PAGE_SIZE; /* Get the amount of memory left */ - tmpaddr = tmpaddr + (addr64_t)(round_page_32(virtual_space_start) - virtual_space_start); /* Account for any slop */ - npages = (unsigned int)(tmpaddr / (addr64_t)(PAGE_SIZE + sizeof(*pages))); /* Figure size of all vm_page_ts, including enough to hold the vm_page_ts */ + tmpaddr = tmpaddr + (addr64_t)(round_page(virtual_space_start) - virtual_space_start); /* Account for any slop */ + npages = (unsigned int)(tmpaddr / (addr64_t)(PAGE_SIZE + sizeof(*vm_pages))); /* Figure size of all vm_page_ts, including enough to hold the vm_page_ts */ - pages = (vm_page_t) pmap_steal_memory(npages * sizeof *pages); + vm_pages = (vm_page_t) pmap_steal_memory(npages * sizeof *vm_pages); /* * Initialize the page frames. @@ -581,59 +850,38 @@ pmap_startup( for (i = 0, pages_initialized = 0; i < npages; i++) { if (!pmap_next_page(&phys_page)) break; + if (pages_initialized == 0 || phys_page < vm_page_lowest) + vm_page_lowest = phys_page; - vm_page_init(&pages[i], phys_page); + vm_page_init(&vm_pages[i], phys_page, FALSE); vm_page_pages++; pages_initialized++; } + vm_pages_count = pages_initialized; /* * Check if we want to initialize pages to a known value */ fill = 0; /* Assume no fill */ - if (PE_parse_boot_arg("fill", &fillval)) fill = 1; /* Set fill */ - - /* - * if vm_lopage_poolsize is non-zero, than we need to reserve - * a pool of pages whose addresess are less than 4G... this pool - * is used by drivers whose hardware can't DMA beyond 32 bits... - * - * note that I'm assuming that the page list is ascending and - * ordered w/r to the physical address + if (PE_parse_boot_argn("fill", &fillval, sizeof (fillval))) fill = 1; /* Set fill */ +#if DEBUG + /* This slows down booting the DEBUG kernel, particularly on + * large memory systems, but is worthwhile in deterministically + * trapping uninitialized memory usage. */ - for (i = 0, num_of_lopages = vm_lopage_poolsize; num_of_lopages && i < pages_initialized; num_of_lopages--, i++) { - vm_page_t m; - - m = &pages[i]; - - if (m->phys_page >= (1 << (32 - PAGE_SHIFT))) - panic("couldn't reserve the lopage pool: not enough lo pages\n"); - - if (m->phys_page < vm_lopage_poolend) - panic("couldn't reserve the lopage pool: page list out of order\n"); - - vm_lopage_poolend = m->phys_page; - - if (vm_lopage_poolstart == 0) - vm_lopage_poolstart = m->phys_page; - else { - if (m->phys_page < vm_lopage_poolstart) - panic("couldn't reserve the lopage pool: page list out of order\n"); - } - - if (fill) - fillPage(m->phys_page, fillval); /* Fill the page with a know value if requested at boot */ - - vm_page_release(m); - } - last_index = i; - + if (fill == 0) { + fill = 1; + fillval = 0xDEB8F177; + } +#endif + if (fill) + kprintf("Filling vm_pages with pattern: 0x%x\n", fillval); // -debug code remove if (2 == vm_himemory_mode) { // free low -> high so high is preferred - for (i = last_index + 1; i <= pages_initialized; i++) { - if(fill) fillPage(pages[i - 1].phys_page, fillval); /* Fill the page with a know value if requested at boot */ - vm_page_release(&pages[i - 1]); + for (i = 1; i <= pages_initialized; i++) { + if(fill) fillPage(vm_pages[i - 1].phys_page, fillval); /* Fill the page with a know value if requested at boot */ + vm_page_release(&vm_pages[i - 1]); } } else @@ -645,36 +893,43 @@ pmap_startup( * the devices (which must address physical memory) happy if * they require several consecutive pages. */ - for (i = pages_initialized; i > last_index; i--) { - if(fill) fillPage(pages[i - 1].phys_page, fillval); /* Fill the page with a know value if requested at boot */ - vm_page_release(&pages[i - 1]); + for (i = pages_initialized; i > 0; i--) { + if(fill) fillPage(vm_pages[i - 1].phys_page, fillval); /* Fill the page with a know value if requested at boot */ + vm_page_release(&vm_pages[i - 1]); } #if 0 { vm_page_t xx, xxo, xxl; - int j, k, l; + int i, j, k, l; j = 0; /* (BRINGUP) */ xxl = 0; - for(xx = vm_page_queue_free; xx; xxl = xx, xx = xx->pageq.next) { /* (BRINGUP) */ - j++; /* (BRINGUP) */ - if(j > vm_page_free_count) { /* (BRINGUP) */ - panic("pmap_startup: too many pages, xx = %08X, xxl = %08X\n", xx, xxl); - } - - l = vm_page_free_count - j; /* (BRINGUP) */ - k = 0; /* (BRINGUP) */ - - if(((j - 1) & 0xFFFF) == 0) kprintf("checking number %d of %d\n", j, vm_page_free_count); - - for(xxo = xx->pageq.next; xxo; xxo = xxo->pageq.next) { /* (BRINGUP) */ - k++; - if(k > l) panic("pmap_startup: too many in secondary check %d %d\n", k, l); - if((xx->phys_page & 0xFFFFFFFF) == (xxo->phys_page & 0xFFFFFFFF)) { /* (BRINGUP) */ - panic("pmap_startup: duplicate physaddr, xx = %08X, xxo = %08X\n", xx, xxo); + for( i = 0; i < vm_colors; i++ ) { + queue_iterate(&vm_page_queue_free[i], + xx, + vm_page_t, + pageq) { /* BRINGUP */ + j++; /* (BRINGUP) */ + if(j > vm_page_free_count) { /* (BRINGUP) */ + panic("pmap_startup: too many pages, xx = %08X, xxl = %08X\n", xx, xxl); } + + l = vm_page_free_count - j; /* (BRINGUP) */ + k = 0; /* (BRINGUP) */ + + if(((j - 1) & 0xFFFF) == 0) kprintf("checking number %d of %d\n", j, vm_page_free_count); + + for(xxo = xx->pageq.next; xxo != &vm_page_queue_free[i]; xxo = xxo->pageq.next) { /* (BRINGUP) */ + k++; + if(k > l) panic("pmap_startup: too many in secondary check %d %d\n", k, l); + if((xx->phys_page & 0xFFFFFFFF) == (xxo->phys_page & 0xFFFFFFFF)) { /* (BRINGUP) */ + panic("pmap_startup: duplicate physaddr, xx = %08X, xxo = %08X\n", xx, xxo); + } + } + + xxl = xx; } } @@ -690,7 +945,7 @@ pmap_startup( * because pmap_steal_memory has been using it. */ - virtual_space_start = round_page_32(virtual_space_start); + virtual_space_start = round_page(virtual_space_start); *startp = virtual_space_start; *endp = virtual_space_end; @@ -713,18 +968,18 @@ vm_page_module_init(void) zone_debug_disable(vm_page_zone); #endif /* ZONE_DEBUG */ + zone_change(vm_page_zone, Z_CALLERACCT, FALSE); zone_change(vm_page_zone, Z_EXPAND, FALSE); zone_change(vm_page_zone, Z_EXHAUST, TRUE); zone_change(vm_page_zone, Z_FOREIGN, TRUE); - + zone_change(vm_page_zone, Z_GZALLOC_EXEMPT, TRUE); /* * Adjust zone statistics to account for the real pages allocated * in vm_page_create(). [Q: is this really what we want?] */ vm_page_zone->count += vm_page_pages; + vm_page_zone->sum_count += vm_page_pages; vm_page_zone->cur_size += vm_page_pages * vm_page_zone->elem_size; - - mutex_init(&vm_page_alloc_lock, 0); } /* @@ -747,11 +1002,13 @@ vm_page_create( for (phys_page = start; phys_page < end; phys_page++) { - while ((m = (vm_page_t) vm_page_grab_fictitious()) + while ((m = (vm_page_t) vm_page_grab_fictitious_common(phys_page)) == VM_PAGE_NULL) vm_page_more_fictitious(); - vm_page_init(m, phys_page); + m->fictitious = FALSE; + pmap_clear_noencrypt(phys_page); + vm_page_pages++; vm_page_release(m); } @@ -765,9 +1022,10 @@ vm_page_create( * NOTE: The bucket count must be a power of 2 */ #define vm_page_hash(object, offset) (\ - ( (natural_t)((uint32_t)object * vm_page_bucket_hash) + ((uint32_t)atop_64(offset) ^ vm_page_bucket_hash))\ + ( (natural_t)((uintptr_t)object * vm_page_bucket_hash) + ((uint32_t)atop_64(offset) ^ vm_page_bucket_hash))\ & vm_page_hash_mask) + /* * vm_page_insert: [ internal use only ] * @@ -776,61 +1034,106 @@ vm_page_create( * * The object must be locked. */ - void vm_page_insert( - register vm_page_t mem, - register vm_object_t object, - register vm_object_offset_t offset) + vm_page_t mem, + vm_object_t object, + vm_object_offset_t offset) +{ + vm_page_insert_internal(mem, object, offset, FALSE, TRUE, FALSE); +} + +void +vm_page_insert_internal( + vm_page_t mem, + vm_object_t object, + vm_object_offset_t offset, + boolean_t queues_lock_held, + boolean_t insert_in_hash, + boolean_t batch_pmap_op) { - register vm_page_bucket_t *bucket; + vm_page_bucket_t *bucket; + lck_spin_t *bucket_lock; + int hash_id; XPR(XPR_VM_PAGE, "vm_page_insert, object 0x%X offset 0x%X page 0x%X\n", - (integer_t)object, (integer_t)offset, (integer_t)mem, 0,0); - + object, offset, mem, 0,0); +#if 0 + /* + * we may not hold the page queue lock + * so this check isn't safe to make + */ VM_PAGE_CHECK(mem); -#if DEBUG - _mutex_assert(&object->Lock, MA_OWNED); - - if (mem->tabled || mem->object != VM_OBJECT_NULL) - panic("vm_page_insert: page %p for (obj=%p,off=0x%llx) " - "already in (obj=%p,off=0x%llx)", - mem, object, offset, mem->object, mem->offset); #endif - assert(!object->internal || offset < object->size); - /* only insert "pageout" pages into "pageout" objects, - * and normal pages into normal objects */ - assert(object->pageout == mem->pageout); + assert(page_aligned(offset)); - assert(vm_page_lookup(object, offset) == VM_PAGE_NULL); + if (object == vm_submap_object) { + /* the vm_submap_object is only a placeholder for submaps */ + panic("vm_page_insert(vm_submap_object,0x%llx)\n", offset); + } - /* - * Record the object/offset pair in this page - */ + vm_object_lock_assert_exclusive(object); +#if DEBUG + lck_mtx_assert(&vm_page_queue_lock, + queues_lock_held ? LCK_MTX_ASSERT_OWNED + : LCK_MTX_ASSERT_NOTOWNED); +#endif /* DEBUG */ + + if (insert_in_hash == TRUE) { +#if DEBUG || VM_PAGE_CHECK_BUCKETS + if (mem->tabled || mem->object != VM_OBJECT_NULL) + panic("vm_page_insert: page %p for (obj=%p,off=0x%llx) " + "already in (obj=%p,off=0x%llx)", + mem, object, offset, mem->object, mem->offset); +#endif + assert(!object->internal || offset < object->vo_size); - mem->object = object; - mem->offset = offset; + /* only insert "pageout" pages into "pageout" objects, + * and normal pages into normal objects */ + assert(object->pageout == mem->pageout); - /* - * Insert it into the object_object/offset hash table - */ + assert(vm_page_lookup(object, offset) == VM_PAGE_NULL); + + /* + * Record the object/offset pair in this page + */ - bucket = &vm_page_buckets[vm_page_hash(object, offset)]; - simple_lock(&vm_page_bucket_lock); - mem->next = bucket->pages; - bucket->pages = mem; + mem->object = object; + mem->offset = offset; + + /* + * Insert it into the object_object/offset hash table + */ + hash_id = vm_page_hash(object, offset); + bucket = &vm_page_buckets[hash_id]; + bucket_lock = &vm_page_bucket_locks[hash_id / BUCKETS_PER_LOCK]; + + lck_spin_lock(bucket_lock); + + mem->next = bucket->pages; + bucket->pages = mem; #if MACH_PAGE_HASH_STATS - if (++bucket->cur_count > bucket->hi_count) - bucket->hi_count = bucket->cur_count; + if (++bucket->cur_count > bucket->hi_count) + bucket->hi_count = bucket->cur_count; #endif /* MACH_PAGE_HASH_STATS */ - simple_unlock(&vm_page_bucket_lock); + mem->hashed = TRUE; + lck_spin_unlock(bucket_lock); + } + { + unsigned int cache_attr; + + cache_attr = object->wimg_bits & VM_WIMG_MASK; + + if (cache_attr != VM_WIMG_USE_DEFAULT) { + PMAP_SET_CACHE_ATTR(mem, object, cache_attr, batch_pmap_op); + } + } /* * Now link into the object's list of backed pages. */ - VM_PAGE_INSERT(mem, object); mem->tabled = TRUE; @@ -839,12 +1142,55 @@ vm_page_insert( */ object->resident_page_count++; + if (VM_PAGE_WIRED(mem)) { + object->wired_page_count++; + } + assert(object->resident_page_count >= object->wired_page_count); - if (object->purgable == VM_OBJECT_PURGABLE_VOLATILE || - object->purgable == VM_OBJECT_PURGABLE_EMPTY) { - vm_page_lock_queues(); - vm_page_purgeable_count++; - vm_page_unlock_queues(); + if (object->internal) { + OSAddAtomic(1, &vm_page_internal_count); + } else { + OSAddAtomic(1, &vm_page_external_count); + } + + /* + * It wouldn't make sense to insert a "reusable" page in + * an object (the page would have been marked "reusable" only + * at the time of a madvise(MADV_FREE_REUSABLE) if it was already + * in the object at that time). + * But a page could be inserted in a "all_reusable" object, if + * something faults it in (a vm_read() from another task or a + * "use-after-free" issue in user space, for example). It can + * also happen if we're relocating a page from that object to + * a different physical page during a physically-contiguous + * allocation. + */ + assert(!mem->reusable); + if (mem->object->all_reusable) { + OSAddAtomic(+1, &vm_page_stats_reusable.reusable_count); + } + + if (object->purgable == VM_PURGABLE_VOLATILE) { + if (VM_PAGE_WIRED(mem)) { + OSAddAtomic(1, &vm_page_purgeable_wired_count); + } else { + OSAddAtomic(1, &vm_page_purgeable_count); + } + } else if (object->purgable == VM_PURGABLE_EMPTY && + mem->throttled) { + /* + * This page belongs to a purged VM object but hasn't + * been purged (because it was "busy"). + * It's in the "throttled" queue and hence not + * visible to vm_pageout_scan(). Move it to a pageable + * queue, so that it can eventually be reclaimed, instead + * of lingering in the "empty" object. + */ + if (queues_lock_held == FALSE) + vm_page_lockspin_queues(); + vm_page_deactivate(mem); + if (queues_lock_held == FALSE) + vm_page_unlock_queues(); } } @@ -854,9 +1200,8 @@ vm_page_insert( * Exactly like vm_page_insert, except that we first * remove any existing page at the given offset in object. * - * The object and page queues must be locked. + * The object must be locked. */ - void vm_page_replace( register vm_page_t mem, @@ -865,16 +1210,23 @@ vm_page_replace( { vm_page_bucket_t *bucket; vm_page_t found_m = VM_PAGE_NULL; + lck_spin_t *bucket_lock; + int hash_id; +#if 0 + /* + * we don't hold the page queue lock + * so this check isn't safe to make + */ VM_PAGE_CHECK(mem); -#if DEBUG - _mutex_assert(&object->Lock, MA_OWNED); - _mutex_assert(&vm_page_queue_lock, MA_OWNED); - +#endif + vm_object_lock_assert_exclusive(object); +#if DEBUG || VM_PAGE_CHECK_BUCKETS if (mem->tabled || mem->object != VM_OBJECT_NULL) panic("vm_page_replace: page %p for (obj=%p,off=0x%llx) " "already in (obj=%p,off=0x%llx)", mem, object, offset, mem->object, mem->offset); + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_NOTOWNED); #endif /* * Record the object/offset pair in this page @@ -888,12 +1240,15 @@ vm_page_replace( * replacing any page that might have been there. */ - bucket = &vm_page_buckets[vm_page_hash(object, offset)]; - simple_lock(&vm_page_bucket_lock); + hash_id = vm_page_hash(object, offset); + bucket = &vm_page_buckets[hash_id]; + bucket_lock = &vm_page_bucket_locks[hash_id / BUCKETS_PER_LOCK]; + + lck_spin_lock(bucket_lock); if (bucket->pages) { vm_page_t *mp = &bucket->pages; - register vm_page_t m = *mp; + vm_page_t m = *mp; do { if (m->object == object && m->offset == offset) { @@ -901,6 +1256,7 @@ vm_page_replace( * Remove old page from hash list */ *mp = m->next; + m->hashed = FALSE; found_m = m; break; @@ -916,8 +1272,9 @@ vm_page_replace( * insert new page at head of hash list */ bucket->pages = mem; + mem->hashed = TRUE; - simple_unlock(&vm_page_bucket_lock); + lck_spin_unlock(bucket_lock); if (found_m) { /* @@ -925,43 +1282,9 @@ vm_page_replace( * offset for this object... remove it from * the object and free it back to the free list */ - VM_PAGE_REMOVE(found_m); - found_m->tabled = FALSE; - - found_m->object = VM_OBJECT_NULL; - found_m->offset = (vm_object_offset_t) -1; - object->resident_page_count--; - - if (object->purgable == VM_OBJECT_PURGABLE_VOLATILE || - object->purgable == VM_OBJECT_PURGABLE_EMPTY) { - assert(vm_page_purgeable_count > 0); - vm_page_purgeable_count--; - } - - /* - * Return page to the free list. - * Note the page is not tabled now - */ - vm_page_free(found_m); - } - /* - * Now link into the object's list of backed pages. - */ - - VM_PAGE_INSERT(mem, object); - mem->tabled = TRUE; - - /* - * And show that the object has one more resident - * page. - */ - - object->resident_page_count++; - - if (object->purgable == VM_OBJECT_PURGABLE_VOLATILE || - object->purgable == VM_OBJECT_PURGABLE_EMPTY) { - vm_page_purgeable_count++; + vm_page_free_unlocked(found_m, FALSE); } + vm_page_insert_internal(mem, object, offset, FALSE, FALSE, FALSE); } /* @@ -970,53 +1293,64 @@ vm_page_replace( * Removes the given mem entry from the object/offset-page * table and the object page list. * - * The object and page queues must be locked. + * The object must be locked. */ void vm_page_remove( - register vm_page_t mem) + vm_page_t mem, + boolean_t remove_from_hash) { - register vm_page_bucket_t *bucket; - register vm_page_t this; + vm_page_bucket_t *bucket; + vm_page_t this; + lck_spin_t *bucket_lock; + int hash_id; XPR(XPR_VM_PAGE, "vm_page_remove, object 0x%X offset 0x%X page 0x%X\n", - (integer_t)mem->object, (integer_t)mem->offset, - (integer_t)mem, 0,0); -#if DEBUG - _mutex_assert(&vm_page_queue_lock, MA_OWNED); - _mutex_assert(&mem->object->Lock, MA_OWNED); -#endif + mem->object, mem->offset, + mem, 0,0); + + vm_object_lock_assert_exclusive(mem->object); assert(mem->tabled); assert(!mem->cleaning); - VM_PAGE_CHECK(mem); - - + assert(!mem->laundry); +#if 0 /* - * Remove from the object_object/offset hash table + * we don't hold the page queue lock + * so this check isn't safe to make */ + VM_PAGE_CHECK(mem); +#endif + if (remove_from_hash == TRUE) { + /* + * Remove from the object_object/offset hash table + */ + hash_id = vm_page_hash(mem->object, mem->offset); + bucket = &vm_page_buckets[hash_id]; + bucket_lock = &vm_page_bucket_locks[hash_id / BUCKETS_PER_LOCK]; - bucket = &vm_page_buckets[vm_page_hash(mem->object, mem->offset)]; - simple_lock(&vm_page_bucket_lock); - if ((this = bucket->pages) == mem) { - /* optimize for common case */ + lck_spin_lock(bucket_lock); - bucket->pages = mem->next; - } else { - register vm_page_t *prev; + if ((this = bucket->pages) == mem) { + /* optimize for common case */ - for (prev = &this->next; - (this = *prev) != mem; - prev = &this->next) - continue; - *prev = this->next; - } + bucket->pages = mem->next; + } else { + vm_page_t *prev; + + for (prev = &this->next; + (this = *prev) != mem; + prev = &this->next) + continue; + *prev = this->next; + } #if MACH_PAGE_HASH_STATS - bucket->cur_count--; + bucket->cur_count--; #endif /* MACH_PAGE_HASH_STATS */ - simple_unlock(&vm_page_bucket_lock); - + mem->hashed = FALSE; + lck_spin_unlock(bucket_lock); + } /* * Now remove from the object's list of backed pages. */ @@ -1028,18 +1362,57 @@ vm_page_remove( * page. */ + assert(mem->object->resident_page_count > 0); mem->object->resident_page_count--; - if (mem->object->purgable == VM_OBJECT_PURGABLE_VOLATILE || - mem->object->purgable == VM_OBJECT_PURGABLE_EMPTY) { - assert(vm_page_purgeable_count > 0); - vm_page_purgeable_count--; + if (mem->object->internal) { + assert(vm_page_internal_count); + OSAddAtomic(-1, &vm_page_internal_count); + } else { + assert(vm_page_external_count); + OSAddAtomic(-1, &vm_page_external_count); + } + if (!mem->object->internal && (mem->object->objq.next || mem->object->objq.prev)) { + if (mem->object->resident_page_count == 0) + vm_object_cache_remove(mem->object); } - mem->tabled = FALSE; - mem->object = VM_OBJECT_NULL; - mem->offset = (vm_object_offset_t) -1; -} + if (VM_PAGE_WIRED(mem)) { + assert(mem->object->wired_page_count > 0); + mem->object->wired_page_count--; + } + assert(mem->object->resident_page_count >= + mem->object->wired_page_count); + if (mem->reusable) { + assert(mem->object->reusable_page_count > 0); + mem->object->reusable_page_count--; + assert(mem->object->reusable_page_count <= + mem->object->resident_page_count); + mem->reusable = FALSE; + OSAddAtomic(-1, &vm_page_stats_reusable.reusable_count); + vm_page_stats_reusable.reused_remove++; + } else if (mem->object->all_reusable) { + OSAddAtomic(-1, &vm_page_stats_reusable.reusable_count); + vm_page_stats_reusable.reused_remove++; + } + + if (mem->object->purgable == VM_PURGABLE_VOLATILE) { + if (VM_PAGE_WIRED(mem)) { + assert(vm_page_purgeable_wired_count > 0); + OSAddAtomic(-1, &vm_page_purgeable_wired_count); + } else { + assert(vm_page_purgeable_count > 0); + OSAddAtomic(-1, &vm_page_purgeable_count); + } + } + if (mem->object->set_cache_attr == TRUE) + pmap_set_cache_attributes(mem->phys_page, 0); + + mem->tabled = FALSE; + mem->object = VM_OBJECT_NULL; + mem->offset = (vm_object_offset_t) -1; +} + /* * vm_page_lookup: @@ -1054,32 +1427,39 @@ unsigned long vm_page_lookup_hint = 0; unsigned long vm_page_lookup_hint_next = 0; unsigned long vm_page_lookup_hint_prev = 0; unsigned long vm_page_lookup_hint_miss = 0; +unsigned long vm_page_lookup_bucket_NULL = 0; +unsigned long vm_page_lookup_miss = 0; + vm_page_t vm_page_lookup( - register vm_object_t object, - register vm_object_offset_t offset) + vm_object_t object, + vm_object_offset_t offset) { - register vm_page_t mem; - register vm_page_bucket_t *bucket; - queue_entry_t qe; -#if 0 - _mutex_assert(&object->Lock, MA_OWNED); -#endif + vm_page_t mem; + vm_page_bucket_t *bucket; + queue_entry_t qe; + lck_spin_t *bucket_lock; + int hash_id; + vm_object_lock_assert_held(object); mem = object->memq_hint; + if (mem != VM_PAGE_NULL) { assert(mem->object == object); + if (mem->offset == offset) { vm_page_lookup_hint++; return mem; } qe = queue_next(&mem->listq); + if (! queue_end(&object->memq, qe)) { vm_page_t next_page; next_page = (vm_page_t) qe; assert(next_page->object == object); + if (next_page->offset == offset) { vm_page_lookup_hint_next++; object->memq_hint = next_page; /* new hint */ @@ -1087,11 +1467,13 @@ vm_page_lookup( } } qe = queue_prev(&mem->listq); + if (! queue_end(&object->memq, qe)) { vm_page_t prev_page; prev_page = (vm_page_t) qe; assert(prev_page->object == object); + if (prev_page->offset == offset) { vm_page_lookup_hint_prev++; object->memq_hint = prev_page; /* new hint */ @@ -1099,32 +1481,41 @@ vm_page_lookup( } } } + /* + * Search the hash table for this object/offset pair + */ + hash_id = vm_page_hash(object, offset); + bucket = &vm_page_buckets[hash_id]; /* - * Search the hash table for this object/offset pair + * since we hold the object lock, we are guaranteed that no + * new pages can be inserted into this object... this in turn + * guarantess that the page we're looking for can't exist + * if the bucket it hashes to is currently NULL even when looked + * at outside the scope of the hash bucket lock... this is a + * really cheap optimiztion to avoid taking the lock */ + if (bucket->pages == VM_PAGE_NULL) { + vm_page_lookup_bucket_NULL++; - bucket = &vm_page_buckets[vm_page_hash(object, offset)]; + return (VM_PAGE_NULL); + } + bucket_lock = &vm_page_bucket_locks[hash_id / BUCKETS_PER_LOCK]; - /* - * since we hold the object lock, we are guaranteed that no - * new pages can be inserted into this object... this in turn - * guarantess that the page we're looking for can't exist - * if the bucket it hashes to is currently NULL even when looked - * at outside the scope of the hash bucket lock... this is a - * really cheap optimiztion to avoid taking the lock - */ - if (bucket->pages == VM_PAGE_NULL) { - return (VM_PAGE_NULL); - } - simple_lock(&vm_page_bucket_lock); + lck_spin_lock(bucket_lock); for (mem = bucket->pages; mem != VM_PAGE_NULL; mem = mem->next) { +#if 0 + /* + * we don't hold the page queue lock + * so this check isn't safe to make + */ VM_PAGE_CHECK(mem); +#endif if ((mem->object == object) && (mem->offset == offset)) break; } - simple_unlock(&vm_page_bucket_lock); + lck_spin_unlock(bucket_lock); if (mem != VM_PAGE_NULL) { if (object->memq_hint != VM_PAGE_NULL) { @@ -1132,40 +1523,13 @@ vm_page_lookup( } assert(mem->object == object); object->memq_hint = mem; - } + } else + vm_page_lookup_miss++; return(mem); } -vm_page_t -vm_page_lookup_nohint( - vm_object_t object, - vm_object_offset_t offset) -{ - register vm_page_t mem; - register vm_page_bucket_t *bucket; - -#if 0 - _mutex_assert(&object->Lock, MA_OWNED); -#endif - /* - * Search the hash table for this object/offset pair - */ - - bucket = &vm_page_buckets[vm_page_hash(object, offset)]; - - simple_lock(&vm_page_bucket_lock); - for (mem = bucket->pages; mem != VM_PAGE_NULL; mem = mem->next) { - VM_PAGE_CHECK(mem); - if ((mem->object == object) && (mem->offset == offset)) - break; - } - simple_unlock(&vm_page_bucket_lock); - - return(mem); -} - /* * vm_page_rename: * @@ -1178,32 +1542,74 @@ void vm_page_rename( register vm_page_t mem, register vm_object_t new_object, - vm_object_offset_t new_offset) + vm_object_offset_t new_offset, + boolean_t encrypted_ok) { + boolean_t internal_to_external, external_to_internal; + assert(mem->object != new_object); + /* * ENCRYPTED SWAP: * The encryption key is based on the page's memory object * (aka "pager") and paging offset. Moving the page to * another VM object changes its "pager" and "paging_offset" - * so it has to be decrypted first. + * so it has to be decrypted first, or we would lose the key. + * + * One exception is VM object collapsing, where we transfer pages + * from one backing object to its parent object. This operation also + * transfers the paging information, so the info + * should remain consistent. The caller (vm_object_do_collapse()) + * sets "encrypted_ok" in this case. */ - if (mem->encrypted) { + if (!encrypted_ok && mem->encrypted) { panic("vm_page_rename: page %p is encrypted\n", mem); } + + XPR(XPR_VM_PAGE, + "vm_page_rename, new object 0x%X, offset 0x%X page 0x%X\n", + new_object, new_offset, + mem, 0,0); + /* * Changes to mem->object require the page lock because * the pageout daemon uses that lock to get the object. */ + vm_page_lockspin_queues(); - XPR(XPR_VM_PAGE, - "vm_page_rename, new object 0x%X, offset 0x%X page 0x%X\n", - (integer_t)new_object, (integer_t)new_offset, - (integer_t)mem, 0,0); + internal_to_external = FALSE; + external_to_internal = FALSE; + + if (mem->local) { + /* + * it's much easier to get the vm_page_pageable_xxx accounting correct + * if we first move the page to the active queue... it's going to end + * up there anyway, and we don't do vm_page_rename's frequently enough + * for this to matter. + */ + VM_PAGE_QUEUES_REMOVE(mem); + vm_page_activate(mem); + } + if (mem->active || mem->inactive || mem->speculative) { + if (mem->object->internal && !new_object->internal) { + internal_to_external = TRUE; + } + if (!mem->object->internal && new_object->internal) { + external_to_internal = TRUE; + } + } + + vm_page_remove(mem, TRUE); + vm_page_insert_internal(mem, new_object, new_offset, TRUE, TRUE, FALSE); + + if (internal_to_external) { + vm_page_pageable_internal_count--; + vm_page_pageable_external_count++; + } else if (external_to_internal) { + vm_page_pageable_external_count--; + vm_page_pageable_internal_count++; + } - vm_page_lock_queues(); - vm_page_remove(mem); - vm_page_insert(mem, new_object, new_offset); vm_page_unlock_queues(); } @@ -1217,11 +1623,40 @@ vm_page_rename( void vm_page_init( vm_page_t mem, - ppnum_t phys_page) + ppnum_t phys_page, + boolean_t lopage) { assert(phys_page); + +#if DEBUG + if ((phys_page != vm_page_fictitious_addr) && (phys_page != vm_page_guard_addr)) { + if (!(pmap_valid_page(phys_page))) { + panic("vm_page_init: non-DRAM phys_page 0x%x\n", phys_page); + } + } +#endif *mem = vm_page_template; mem->phys_page = phys_page; +#if 0 + /* + * we're leaving this turned off for now... currently pages + * come off the free list and are either immediately dirtied/referenced + * due to zero-fill or COW faults, or are used to read or write files... + * in the file I/O case, the UPL mechanism takes care of clearing + * the state of the HW ref/mod bits in a somewhat fragile way. + * Since we may change the way this works in the future (to toughen it up), + * I'm leaving this as a reminder of where these bits could get cleared + */ + + /* + * make sure both the h/w referenced and modified bits are + * clear at this point... we are especially dependent on + * not finding a 'stale' h/w modified in a number of spots + * once this page goes back into use + */ + pmap_clear_refmod(phys_page, VM_MEM_MODIFIED | VM_MEM_REFERENCED); +#endif + mem->lopage = lopage; } /* @@ -1231,52 +1666,64 @@ vm_page_init( * Returns VM_PAGE_NULL if there are no free pages. */ int c_vm_page_grab_fictitious = 0; +int c_vm_page_grab_fictitious_failed = 0; int c_vm_page_release_fictitious = 0; int c_vm_page_more_fictitious = 0; vm_page_t -vm_page_grab_fictitious(void) +vm_page_grab_fictitious_common( + ppnum_t phys_addr) { - register vm_page_t m; + vm_page_t m; - m = (vm_page_t)zget(vm_page_zone); - if (m) { - vm_page_init(m, vm_page_fictitious_addr); + if ((m = (vm_page_t)zget(vm_page_zone))) { + + vm_page_init(m, phys_addr, FALSE); m->fictitious = TRUE; - } - c_vm_page_grab_fictitious++; + c_vm_page_grab_fictitious++; + } else + c_vm_page_grab_fictitious_failed++; + return m; } +vm_page_t +vm_page_grab_fictitious(void) +{ + return vm_page_grab_fictitious_common(vm_page_fictitious_addr); +} + +vm_page_t +vm_page_grab_guard(void) +{ + return vm_page_grab_fictitious_common(vm_page_guard_addr); +} + + /* * vm_page_release_fictitious: * - * Release a fictitious page to the free list. + * Release a fictitious page to the zone pool */ - void vm_page_release_fictitious( - register vm_page_t m) + vm_page_t m) { assert(!m->free); - assert(m->busy); assert(m->fictitious); - assert(m->phys_page == vm_page_fictitious_addr); + assert(m->phys_page == vm_page_fictitious_addr || + m->phys_page == vm_page_guard_addr); c_vm_page_release_fictitious++; -#if DEBUG - if (m->free) - panic("vm_page_release_fictitious"); -#endif - m->free = TRUE; + zfree(vm_page_zone, m); } /* * vm_page_more_fictitious: * - * Add more fictitious pages to the free list. + * Add more fictitious pages to the zone. * Allowed to block. This routine is way intimate * with the zones code, for several reasons: * 1. we need to carve some page structures out of physical @@ -1290,23 +1737,13 @@ vm_page_release_fictitious( * permanent allocation of a resource. * 3. To smooth allocation humps, we allocate single pages * with kernel_memory_allocate(), and cram them into the - * zone. This also allows us to initialize the vm_page_t's - * on the way into the zone, so that zget() always returns - * an initialized structure. The zone free element pointer - * and the free page pointer are both the first item in the - * vm_page_t. - * 4. By having the pages in the zone pre-initialized, we need - * not keep 2 levels of lists. The garbage collector simply - * scans our list, and reduces physical memory usage as it - * sees fit. + * zone. */ void vm_page_more_fictitious(void) { - register vm_page_t m; - vm_offset_t addr; - kern_return_t retval; - int i; + vm_offset_t addr; + kern_return_t retval; c_vm_page_more_fictitious++; @@ -1319,7 +1756,7 @@ void vm_page_more_fictitious(void) * If winner is not vm-privileged, then the page allocation will fail, * and it will temporarily block here in the vm_page_wait(). */ - mutex_lock(&vm_page_alloc_lock); + lck_mtx_lock(&vm_page_alloc_lock); /* * If another thread allocated space, just bail out now. */ @@ -1336,7 +1773,7 @@ void vm_page_more_fictitious(void) * of fictitious pages required in this manner is 2. 5 is * simply a somewhat larger number. */ - mutex_unlock(&vm_page_alloc_lock); + lck_mtx_unlock(&vm_page_alloc_lock); return; } @@ -1345,67 +1782,24 @@ void vm_page_more_fictitious(void) KMA_KOBJECT|KMA_NOPAGEWAIT); if (retval != KERN_SUCCESS) { /* - * No page was available. Tell the pageout daemon, drop the + * No page was available. Drop the * lock to give another thread a chance at it, and * wait for the pageout daemon to make progress. */ - mutex_unlock(&vm_page_alloc_lock); + lck_mtx_unlock(&vm_page_alloc_lock); vm_page_wait(THREAD_UNINT); return; } - /* - * Initialize as many vm_page_t's as will fit on this page. This - * depends on the zone code disturbing ONLY the first item of - * each zone element. - */ - m = (vm_page_t)addr; - for (i = PAGE_SIZE/sizeof(struct vm_page); i > 0; i--) { - vm_page_init(m, vm_page_fictitious_addr); - m->fictitious = TRUE; - m++; - } - zcram(vm_page_zone, (void *) addr, PAGE_SIZE); - mutex_unlock(&vm_page_alloc_lock); -} -/* - * vm_page_convert: - * - * Attempt to convert a fictitious page into a real page. - */ - -boolean_t -vm_page_convert( - register vm_page_t m) -{ - register vm_page_t real_m; - - assert(m->busy); - assert(m->fictitious); - assert(!m->dirty); + /* Increment zone page count. We account for all memory managed by the zone in z->page_count */ + OSAddAtomic64(1, &(vm_page_zone->page_count)); - real_m = vm_page_grab(); - if (real_m == VM_PAGE_NULL) - return FALSE; + zcram(vm_page_zone, addr, PAGE_SIZE); - m->phys_page = real_m->phys_page; - m->fictitious = FALSE; - m->no_isync = TRUE; - - vm_page_lock_queues(); - if (m->active) - vm_page_active_count++; - else if (m->inactive) - vm_page_inactive_count++; - vm_page_unlock_queues(); - - real_m->phys_page = vm_page_fictitious_addr; - real_m->fictitious = TRUE; - - vm_page_release_fictitious(real_m); - return TRUE; + lck_mtx_unlock(&vm_page_alloc_lock); } + /* * vm_pool_low(): * @@ -1417,7 +1811,7 @@ int vm_pool_low(void) { /* No locking, at worst we will fib. */ - return( vm_page_free_count < vm_page_free_reserved ); + return( vm_page_free_count <= vm_page_free_reserved ); } @@ -1434,57 +1828,130 @@ int vm_himemory_mode = 0; * incapable of generating DMAs with more than 32 bits * of address on platforms with physical memory > 4G... */ -unsigned int vm_lopage_free_count = 0; -unsigned int vm_lopage_max_count = 0; -vm_page_t vm_lopage_queue_free = VM_PAGE_NULL; +unsigned int vm_lopages_allocated_q = 0; +unsigned int vm_lopages_allocated_cpm_success = 0; +unsigned int vm_lopages_allocated_cpm_failed = 0; +queue_head_t vm_lopage_queue_free; vm_page_t vm_page_grablo(void) { - register vm_page_t mem; - unsigned int vm_lopage_alloc_count; + vm_page_t mem; - if (vm_lopage_poolsize == 0) + if (vm_lopage_needed == FALSE) return (vm_page_grab()); - mutex_lock(&vm_page_queue_free_lock); + lck_mtx_lock_spin(&vm_page_queue_free_lock); - if ((mem = vm_lopage_queue_free) != VM_PAGE_NULL) { + if ( !queue_empty(&vm_lopage_queue_free)) { + queue_remove_first(&vm_lopage_queue_free, + mem, + vm_page_t, + pageq); + assert(vm_lopage_free_count); - vm_lopage_queue_free = (vm_page_t) mem->pageq.next; - mem->pageq.next = NULL; - mem->pageq.prev = NULL; - mem->free = FALSE; - mem->no_isync = TRUE; + vm_lopage_free_count--; + vm_lopages_allocated_q++; + + if (vm_lopage_free_count < vm_lopage_lowater) + vm_lopage_refill = TRUE; + + lck_mtx_unlock(&vm_page_queue_free_lock); + } else { + lck_mtx_unlock(&vm_page_queue_free_lock); + + if (cpm_allocate(PAGE_SIZE, &mem, atop(0xffffffff), 0, FALSE, KMA_LOMEM) != KERN_SUCCESS) { + + lck_mtx_lock_spin(&vm_page_queue_free_lock); + vm_lopages_allocated_cpm_failed++; + lck_mtx_unlock(&vm_page_queue_free_lock); + + return (VM_PAGE_NULL); + } + mem->busy = TRUE; + + vm_page_lockspin_queues(); + + mem->gobbled = FALSE; + vm_page_gobble_count--; + vm_page_wire_count--; - vm_lopage_free_count--; - vm_lopage_alloc_count = (vm_lopage_poolend - vm_lopage_poolstart) - vm_lopage_free_count; - if (vm_lopage_alloc_count > vm_lopage_max_count) - vm_lopage_max_count = vm_lopage_alloc_count; + vm_lopages_allocated_cpm_success++; + vm_page_unlock_queues(); } - mutex_unlock(&vm_page_queue_free_lock); + assert(mem->busy); + assert(!mem->free); + assert(!mem->pmapped); + assert(!mem->wpmapped); + assert(!pmap_is_noencrypt(mem->phys_page)); + + mem->pageq.next = NULL; + mem->pageq.prev = NULL; return (mem); } - /* * vm_page_grab: * - * Remove a page from the free list. - * Returns VM_PAGE_NULL if the free list is too small. + * first try to grab a page from the per-cpu free list... + * this must be done while pre-emption is disabled... if + * a page is available, we're done... + * if no page is available, grab the vm_page_queue_free_lock + * and see if current number of free pages would allow us + * to grab at least 1... if not, return VM_PAGE_NULL as before... + * if there are pages available, disable preemption and + * recheck the state of the per-cpu free list... we could + * have been preempted and moved to a different cpu, or + * some other thread could have re-filled it... if still + * empty, figure out how many pages we can steal from the + * global free queue and move to the per-cpu queue... + * return 1 of these pages when done... only wakeup the + * pageout_scan thread if we moved pages from the global + * list... no need for the wakeup if we've satisfied the + * request from the per-cpu queue. */ -unsigned long vm_page_grab_count = 0; /* measure demand */ +#define COLOR_GROUPS_TO_STEAL 4 + vm_page_t -vm_page_grab(void) +vm_page_grab( void ) { - register vm_page_t mem; + vm_page_t mem; + + + disable_preemption(); + + if ((mem = PROCESSOR_DATA(current_processor(), free_pages))) { +return_page_from_cpu_list: + PROCESSOR_DATA(current_processor(), page_grab_count) += 1; + PROCESSOR_DATA(current_processor(), free_pages) = mem->pageq.next; + mem->pageq.next = NULL; + + enable_preemption(); + + assert(mem->listq.next == NULL && mem->listq.prev == NULL); + assert(mem->tabled == FALSE); + assert(mem->object == VM_OBJECT_NULL); + assert(!mem->laundry); + assert(!mem->free); + assert(pmap_verify_free(mem->phys_page)); + assert(mem->busy); + assert(!mem->encrypted); + assert(!mem->pmapped); + assert(!mem->wpmapped); + assert(!mem->active); + assert(!mem->inactive); + assert(!mem->throttled); + assert(!mem->speculative); + assert(!pmap_is_noencrypt(mem->phys_page)); + + return mem; + } + enable_preemption(); - mutex_lock(&vm_page_queue_free_lock); - vm_page_grab_count++; /* * Optionally produce warnings if the wire or gobble @@ -1503,40 +1970,115 @@ vm_page_grab(void) assert(vm_page_gobble_count < vm_page_gobble_count_warning); } + lck_mtx_lock_spin(&vm_page_queue_free_lock); + /* * Only let privileged threads (involved in pageout) * dip into the reserved pool. */ - if ((vm_page_free_count < vm_page_free_reserved) && !(current_thread()->options & TH_OPT_VMPRIV)) { - mutex_unlock(&vm_page_queue_free_lock); + lck_mtx_unlock(&vm_page_queue_free_lock); mem = VM_PAGE_NULL; - goto wakeup_pageout; } + else { + vm_page_t head; + vm_page_t tail; + unsigned int pages_to_steal; + unsigned int color; - while (vm_page_queue_free == VM_PAGE_NULL) { - mutex_unlock(&vm_page_queue_free_lock); - VM_PAGE_WAIT(); - mutex_lock(&vm_page_queue_free_lock); - } + while ( vm_page_free_count == 0 ) { - if (--vm_page_free_count < vm_page_free_count_minimum) - vm_page_free_count_minimum = vm_page_free_count; - mem = vm_page_queue_free; - vm_page_queue_free = (vm_page_t) mem->pageq.next; - mem->pageq.next = NULL; - mem->pageq.prev = NULL; - assert(mem->listq.next == NULL && mem->listq.prev == NULL); - assert(mem->tabled == FALSE); - assert(mem->object == VM_OBJECT_NULL); - assert(!mem->laundry); - mem->free = FALSE; - mem->no_isync = TRUE; - mutex_unlock(&vm_page_queue_free_lock); + lck_mtx_unlock(&vm_page_queue_free_lock); + /* + * must be a privileged thread to be + * in this state since a non-privileged + * thread would have bailed if we were + * under the vm_page_free_reserved mark + */ + VM_PAGE_WAIT(); + lck_mtx_lock_spin(&vm_page_queue_free_lock); + } + + disable_preemption(); + + if ((mem = PROCESSOR_DATA(current_processor(), free_pages))) { + lck_mtx_unlock(&vm_page_queue_free_lock); + + /* + * we got preempted and moved to another processor + * or we got preempted and someone else ran and filled the cache + */ + goto return_page_from_cpu_list; + } + if (vm_page_free_count <= vm_page_free_reserved) + pages_to_steal = 1; + else { + pages_to_steal = COLOR_GROUPS_TO_STEAL * vm_colors; + + if (pages_to_steal > (vm_page_free_count - vm_page_free_reserved)) + pages_to_steal = (vm_page_free_count - vm_page_free_reserved); + } + color = PROCESSOR_DATA(current_processor(), start_color); + head = tail = NULL; + + while (pages_to_steal--) { + if (--vm_page_free_count < vm_page_free_count_minimum) + vm_page_free_count_minimum = vm_page_free_count; + + while (queue_empty(&vm_page_queue_free[color])) + color = (color + 1) & vm_color_mask; + + queue_remove_first(&vm_page_queue_free[color], + mem, + vm_page_t, + pageq); + mem->pageq.next = NULL; + mem->pageq.prev = NULL; + + assert(!mem->active); + assert(!mem->inactive); + assert(!mem->throttled); + assert(!mem->speculative); + + color = (color + 1) & vm_color_mask; + + if (head == NULL) + head = mem; + else + tail->pageq.next = (queue_t)mem; + tail = mem; + + mem->pageq.prev = NULL; + assert(mem->listq.next == NULL && mem->listq.prev == NULL); + assert(mem->tabled == FALSE); + assert(mem->object == VM_OBJECT_NULL); + assert(!mem->laundry); + assert(mem->free); + mem->free = FALSE; + + assert(pmap_verify_free(mem->phys_page)); + assert(mem->busy); + assert(!mem->free); + assert(!mem->encrypted); + assert(!mem->pmapped); + assert(!mem->wpmapped); + assert(!pmap_is_noencrypt(mem->phys_page)); + } + PROCESSOR_DATA(current_processor(), free_pages) = head->pageq.next; + PROCESSOR_DATA(current_processor(), start_color) = color; + + /* + * satisfy this request + */ + PROCESSOR_DATA(current_processor(), page_grab_count) += 1; + mem = head; + mem->pageq.next = NULL; - assert(pmap_verify_free(mem->phys_page)); + lck_mtx_unlock(&vm_page_queue_free_lock); + enable_preemption(); + } /* * Decide if we should poke the pageout daemon. * We do this if the free count is less than the low @@ -1547,13 +2089,13 @@ vm_page_grab(void) * We don't have the counts locked ... if they change a little, * it doesn't really matter. */ - -wakeup_pageout: if ((vm_page_free_count < vm_page_free_min) || - ((vm_page_free_count < vm_page_free_target) && - (vm_page_inactive_count < vm_page_inactive_target))) - thread_wakeup((event_t) &vm_page_free_wanted); + ((vm_page_free_count < vm_page_free_target) && + ((vm_page_inactive_count + vm_page_speculative_count) < vm_page_inactive_min))) + thread_wakeup((event_t) &vm_page_free_wanted); + VM_CHECK_MEMORYSTATUS; + // dbgLog(mem->phys_page, vm_page_free_count, vm_page_wire_count, 4); /* (TEST/DEBUG) */ return mem; @@ -1569,44 +2111,60 @@ void vm_page_release( register vm_page_t mem) { + unsigned int color; + int need_wakeup = 0; + int need_priv_wakeup = 0; -#if 0 - unsigned int pindex; - phys_entry *physent; - physent = mapping_phys_lookup(mem->phys_page, &pindex); /* (BRINGUP) */ - if(physent->ppLink & ppN) { /* (BRINGUP) */ - panic("vm_page_release: already released - %08X %08X\n", mem, mem->phys_page); - } - physent->ppLink = physent->ppLink | ppN; /* (BRINGUP) */ -#endif assert(!mem->private && !mem->fictitious); - + if (vm_page_free_verify) { + assert(pmap_verify_free(mem->phys_page)); + } // dbgLog(mem->phys_page, vm_page_free_count, vm_page_wire_count, 5); /* (TEST/DEBUG) */ - mutex_lock(&vm_page_queue_free_lock); + pmap_clear_noencrypt(mem->phys_page); + + lck_mtx_lock_spin(&vm_page_queue_free_lock); #if DEBUG if (mem->free) panic("vm_page_release"); #endif - mem->free = TRUE; + + assert(mem->busy); assert(!mem->laundry); assert(mem->object == VM_OBJECT_NULL); assert(mem->pageq.next == NULL && mem->pageq.prev == NULL); - - if (mem->phys_page <= vm_lopage_poolend && mem->phys_page >= vm_lopage_poolstart) { + assert(mem->listq.next == NULL && + mem->listq.prev == NULL); + + if ((mem->lopage == TRUE || vm_lopage_refill == TRUE) && + vm_lopage_free_count < vm_lopage_free_limit && + mem->phys_page < max_valid_low_ppnum) { /* * this exists to support hardware controllers * incapable of generating DMAs with more than 32 bits * of address on platforms with physical memory > 4G... */ - mem->pageq.next = (queue_entry_t) vm_lopage_queue_free; - vm_lopage_queue_free = mem; + queue_enter_first(&vm_lopage_queue_free, + mem, + vm_page_t, + pageq); vm_lopage_free_count++; + + if (vm_lopage_free_count >= vm_lopage_free_limit) + vm_lopage_refill = FALSE; + + mem->lopage = TRUE; } else { - mem->pageq.next = (queue_entry_t) vm_page_queue_free; - vm_page_queue_free = mem; + mem->lopage = FALSE; + mem->free = TRUE; + + color = mem->phys_page & vm_color_mask; + queue_enter_first(&vm_page_queue_free[color], + mem, + vm_page_t, + pageq); vm_page_free_count++; /* * Check if we should wake up someone waiting for page. @@ -1628,13 +2186,24 @@ vm_page_release( * as long as vm_page_free_wanted is non-zero. */ - if ((vm_page_free_wanted > 0) && - (vm_page_free_count >= vm_page_free_reserved)) { + assert(vm_page_free_count > 0); + if (vm_page_free_wanted_privileged > 0) { + vm_page_free_wanted_privileged--; + need_priv_wakeup = 1; + } else if (vm_page_free_wanted > 0 && + vm_page_free_count > vm_page_free_reserved) { vm_page_free_wanted--; - thread_wakeup_one((event_t) &vm_page_free_count); + need_wakeup = 1; } } - mutex_unlock(&vm_page_queue_free_lock); + lck_mtx_unlock(&vm_page_queue_free_lock); + + if (need_priv_wakeup) + thread_wakeup_one((event_t) &vm_page_free_wanted_privileged); + else if (need_wakeup) + thread_wakeup_one((event_t) &vm_page_free_count); + + VM_CHECK_MEMORYSTATUS; } /* @@ -1661,24 +2230,41 @@ vm_page_wait( */ kern_return_t wait_result; int need_wakeup = 0; + int is_privileged = current_thread()->options & TH_OPT_VMPRIV; - mutex_lock(&vm_page_queue_free_lock); + lck_mtx_lock_spin(&vm_page_queue_free_lock); + + if (is_privileged && vm_page_free_count) { + lck_mtx_unlock(&vm_page_queue_free_lock); + return TRUE; + } if (vm_page_free_count < vm_page_free_target) { - if (vm_page_free_wanted++ == 0) - need_wakeup = 1; - wait_result = assert_wait((event_t)&vm_page_free_count, interruptible); - mutex_unlock(&vm_page_queue_free_lock); + + if (is_privileged) { + if (vm_page_free_wanted_privileged++ == 0) + need_wakeup = 1; + wait_result = assert_wait((event_t)&vm_page_free_wanted_privileged, interruptible); + } else { + if (vm_page_free_wanted++ == 0) + need_wakeup = 1; + wait_result = assert_wait((event_t)&vm_page_free_count, interruptible); + } + lck_mtx_unlock(&vm_page_queue_free_lock); counter(c_vm_page_wait_block++); if (need_wakeup) thread_wakeup((event_t)&vm_page_free_wanted); - if (wait_result == THREAD_WAITING) + if (wait_result == THREAD_WAITING) { + VM_DEBUG_EVENT(vm_page_wait_block, VM_PAGE_WAIT_BLOCK, DBG_FUNC_START, + vm_page_free_wanted_privileged, vm_page_free_wanted, 0, 0); wait_result = thread_block(THREAD_CONTINUE_NULL); + VM_DEBUG_EVENT(vm_page_wait_block, VM_PAGE_WAIT_BLOCK, DBG_FUNC_END, 0, 0, 0, 0); + } return(wait_result == THREAD_AWAKENED); } else { - mutex_unlock(&vm_page_queue_free_lock); + lck_mtx_unlock(&vm_page_queue_free_lock); return TRUE; } } @@ -1699,9 +2285,7 @@ vm_page_alloc( { register vm_page_t mem; -#if DEBUG - _mutex_assert(&object->Lock, MA_OWNED); -#endif + vm_object_lock_assert_exclusive(object); mem = vm_page_grab(); if (mem == VM_PAGE_NULL) return VM_PAGE_NULL; @@ -1711,7 +2295,6 @@ vm_page_alloc( return(mem); } - vm_page_t vm_page_alloclo( vm_object_t object, @@ -1719,9 +2302,7 @@ vm_page_alloclo( { register vm_page_t mem; -#if DEBUG - _mutex_assert(&object->Lock, MA_OWNED); -#endif + vm_object_lock_assert_exclusive(object); mem = vm_page_grablo(); if (mem == VM_PAGE_NULL) return VM_PAGE_NULL; @@ -1732,50 +2313,94 @@ vm_page_alloclo( } +/* + * vm_page_alloc_guard: + * + * Allocate a fictitious page which will be used + * as a guard page. The page will be inserted into + * the object and returned to the caller. + */ + +vm_page_t +vm_page_alloc_guard( + vm_object_t object, + vm_object_offset_t offset) +{ + register vm_page_t mem; + + vm_object_lock_assert_exclusive(object); + mem = vm_page_grab_guard(); + if (mem == VM_PAGE_NULL) + return VM_PAGE_NULL; + + vm_page_insert(mem, object, offset); + + return(mem); +} + + counter(unsigned int c_laundry_pages_freed = 0;) -int vm_pagein_cluster_unused = 0; -boolean_t vm_page_free_verify = TRUE; /* - * vm_page_free: + * vm_page_free_prepare: * - * Returns the given page to the free list, - * disassociating it with any VM object. + * Removes page from any queue it may be on + * and disassociates it from its VM object. * * Object and page queues must be locked prior to entry. */ -void -vm_page_free( - register vm_page_t mem) +static void +vm_page_free_prepare( + vm_page_t mem) { - vm_object_t object = mem->object; + vm_page_free_prepare_queues(mem); + vm_page_free_prepare_object(mem, TRUE); +} + +void +vm_page_free_prepare_queues( + vm_page_t mem) +{ + VM_PAGE_CHECK(mem); assert(!mem->free); assert(!mem->cleaning); - assert(!mem->pageout); - if (vm_page_free_verify && !mem->fictitious && !mem->private) { - assert(pmap_verify_free(mem->phys_page)); - } - #if DEBUG - if (mem->object) - _mutex_assert(&mem->object->Lock, MA_OWNED); - _mutex_assert(&vm_page_queue_lock, MA_OWNED); - + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); if (mem->free) - panic("vm_page_free: freeing page on free list\n"); + panic("vm_page_free: freeing page on free list\n"); #endif - if (mem->tabled) - vm_page_remove(mem); /* clears tabled, object, offset */ - VM_PAGE_QUEUES_REMOVE(mem); /* clears active or inactive */ - - if (mem->clustered) { - mem->clustered = FALSE; - vm_pagein_cluster_unused++; + if (mem->object) { + vm_object_lock_assert_exclusive(mem->object); } - - if (mem->wire_count) { - if (!mem->private && !mem->fictitious) + if (mem->laundry) { + /* + * We may have to free a page while it's being laundered + * if we lost its pager (due to a forced unmount, for example). + * We need to call vm_pageout_steal_laundry() before removing + * the page from its VM object, so that we can remove it + * from its pageout queue and adjust the laundry accounting + */ + vm_pageout_steal_laundry(mem, TRUE); + counter(++c_laundry_pages_freed); + } + + VM_PAGE_QUEUES_REMOVE(mem); /* clears local/active/inactive/throttled/speculative */ + + if (VM_PAGE_WIRED(mem)) { + if (mem->object) { + assert(mem->object->wired_page_count > 0); + mem->object->wired_page_count--; + assert(mem->object->resident_page_count >= + mem->object->wired_page_count); + + if (mem->object->purgable == VM_PURGABLE_VOLATILE) { + OSAddAtomic(+1, &vm_page_purgeable_count); + assert(vm_page_purgeable_wired_count > 0); + OSAddAtomic(-1, &vm_page_purgeable_wired_count); + } + } + if (!mem->private && !mem->fictitious) vm_page_wire_count--; mem->wire_count = 0; assert(!mem->gobbled); @@ -1784,138 +2409,227 @@ vm_page_free( vm_page_wire_count--; vm_page_gobble_count--; } - mem->gobbled = FALSE; - - if (mem->laundry) { - vm_pageout_throttle_up(mem); - counter(++c_laundry_pages_freed); - } - - PAGE_WAKEUP(mem); /* clears wanted */ +} - if (mem->absent) - vm_object_absent_release(object); - /* Some of these may be unnecessary */ - mem->page_lock = 0; - mem->unlock_request = 0; - mem->busy = TRUE; - mem->absent = FALSE; - mem->error = FALSE; - mem->dirty = FALSE; - mem->precious = FALSE; - mem->reference = FALSE; - mem->encrypted = FALSE; +void +vm_page_free_prepare_object( + vm_page_t mem, + boolean_t remove_from_hash) +{ + if (mem->tabled) + vm_page_remove(mem, remove_from_hash); /* clears tabled, object, offset */ - mem->page_error = KERN_SUCCESS; + PAGE_WAKEUP(mem); /* clears wanted */ if (mem->private) { mem->private = FALSE; mem->fictitious = TRUE; mem->phys_page = vm_page_fictitious_addr; } + if ( !mem->fictitious) { + vm_page_init(mem, mem->phys_page, mem->lopage); + } +} + + +/* + * vm_page_free: + * + * Returns the given page to the free list, + * disassociating it with any VM object. + * + * Object and page queues must be locked prior to entry. + */ +void +vm_page_free( + vm_page_t mem) +{ + vm_page_free_prepare(mem); + if (mem->fictitious) { vm_page_release_fictitious(mem); } else { - /* depends on the queues lock */ - if(mem->zero_fill) { - vm_zf_count-=1; - mem->zero_fill = FALSE; - } - vm_page_init(mem, mem->phys_page); vm_page_release(mem); } } void -vm_page_free_list( - register vm_page_t mem) +vm_page_free_unlocked( + vm_page_t mem, + boolean_t remove_from_hash) { - register vm_page_t nxt; - register vm_page_t first = NULL; - register vm_page_t last = VM_PAGE_NULL; - register int pg_count = 0; + vm_page_lockspin_queues(); + vm_page_free_prepare_queues(mem); + vm_page_unlock_queues(); -#if DEBUG - _mutex_assert(&vm_page_queue_lock, MA_OWNED); -#endif - while (mem) { -#if DEBUG - if (mem->tabled || mem->object) - panic("vm_page_free_list: freeing tabled page\n"); - if (mem->inactive || mem->active || mem->free) - panic("vm_page_free_list: freeing page on list\n"); -#endif - assert(mem->pageq.prev == NULL); - nxt = (vm_page_t)(mem->pageq.next); + vm_page_free_prepare_object(mem, remove_from_hash); - if (mem->clustered) - vm_pagein_cluster_unused++; + if (mem->fictitious) { + vm_page_release_fictitious(mem); + } else { + vm_page_release(mem); + } +} - if (mem->laundry) { - vm_pageout_throttle_up(mem); - counter(++c_laundry_pages_freed); - } - mem->busy = TRUE; - PAGE_WAKEUP(mem); /* clears wanted */ +/* + * Free a list of pages. The list can be up to several hundred pages, + * as blocked up by vm_pageout_scan(). + * The big win is not having to take the free list lock once + * per page. + */ +void +vm_page_free_list( + vm_page_t freeq, + boolean_t prepare_object) +{ + vm_page_t mem; + vm_page_t nxt; + vm_page_t local_freeq; + int pg_count; - if (mem->private) - mem->fictitious = TRUE; + while (freeq) { - if (!mem->fictitious) { - /* depends on the queues lock */ - if (mem->zero_fill) - vm_zf_count -= 1; - assert(!mem->laundry); - vm_page_init(mem, mem->phys_page); + pg_count = 0; + local_freeq = VM_PAGE_NULL; + mem = freeq; - mem->free = TRUE; + /* + * break up the processing into smaller chunks so + * that we can 'pipeline' the pages onto the + * free list w/o introducing too much + * contention on the global free queue lock + */ + while (mem && pg_count < 64) { - if (first == NULL) - last = mem; - mem->pageq.next = (queue_t) first; - first = mem; + assert(!mem->inactive); + assert(!mem->active); + assert(!mem->throttled); + assert(!mem->free); + assert(!mem->speculative); + assert(!VM_PAGE_WIRED(mem)); + assert(mem->pageq.prev == NULL); - pg_count++; - } else { - mem->phys_page = vm_page_fictitious_addr; - vm_page_release_fictitious(mem); + nxt = (vm_page_t)(mem->pageq.next); + + if (vm_page_free_verify && !mem->fictitious && !mem->private) { + assert(pmap_verify_free(mem->phys_page)); + } + if (prepare_object == TRUE) + vm_page_free_prepare_object(mem, TRUE); + + if (!mem->fictitious) { + assert(mem->busy); + + if ((mem->lopage == TRUE || vm_lopage_refill == TRUE) && + vm_lopage_free_count < vm_lopage_free_limit && + mem->phys_page < max_valid_low_ppnum) { + mem->pageq.next = NULL; + vm_page_release(mem); + } else { + /* + * IMPORTANT: we can't set the page "free" here + * because that would make the page eligible for + * a physically-contiguous allocation (see + * vm_page_find_contiguous()) right away (we don't + * hold the vm_page_queue_free lock). That would + * cause trouble because the page is not actually + * in the free queue yet... + */ + mem->pageq.next = (queue_entry_t)local_freeq; + local_freeq = mem; + pg_count++; + + pmap_clear_noencrypt(mem->phys_page); + } + } else { + assert(mem->phys_page == vm_page_fictitious_addr || + mem->phys_page == vm_page_guard_addr); + vm_page_release_fictitious(mem); + } + mem = nxt; } - mem = nxt; - } - if (first) { - - mutex_lock(&vm_page_queue_free_lock); + freeq = mem; + + if ( (mem = local_freeq) ) { + unsigned int avail_free_count; + unsigned int need_wakeup = 0; + unsigned int need_priv_wakeup = 0; + + lck_mtx_lock_spin(&vm_page_queue_free_lock); + + while (mem) { + int color; + + nxt = (vm_page_t)(mem->pageq.next); + + assert(!mem->free); + assert(mem->busy); + mem->free = TRUE; + + color = mem->phys_page & vm_color_mask; + queue_enter_first(&vm_page_queue_free[color], + mem, + vm_page_t, + pageq); + mem = nxt; + } + vm_page_free_count += pg_count; + avail_free_count = vm_page_free_count; - last->pageq.next = (queue_entry_t) vm_page_queue_free; - vm_page_queue_free = first; + if (vm_page_free_wanted_privileged > 0 && avail_free_count > 0) { - vm_page_free_count += pg_count; + if (avail_free_count < vm_page_free_wanted_privileged) { + need_priv_wakeup = avail_free_count; + vm_page_free_wanted_privileged -= avail_free_count; + avail_free_count = 0; + } else { + need_priv_wakeup = vm_page_free_wanted_privileged; + vm_page_free_wanted_privileged = 0; + avail_free_count -= vm_page_free_wanted_privileged; + } + } + if (vm_page_free_wanted > 0 && avail_free_count > vm_page_free_reserved) { + unsigned int available_pages; - if ((vm_page_free_wanted > 0) && - (vm_page_free_count >= vm_page_free_reserved)) { - unsigned int available_pages; + available_pages = avail_free_count - vm_page_free_reserved; - if (vm_page_free_count >= vm_page_free_reserved) { - available_pages = (vm_page_free_count - - vm_page_free_reserved); - } else { - available_pages = 0; + if (available_pages >= vm_page_free_wanted) { + need_wakeup = vm_page_free_wanted; + vm_page_free_wanted = 0; + } else { + need_wakeup = available_pages; + vm_page_free_wanted -= available_pages; + } } + lck_mtx_unlock(&vm_page_queue_free_lock); - if (available_pages >= vm_page_free_wanted) { - vm_page_free_wanted = 0; + if (need_priv_wakeup != 0) { + /* + * There shouldn't be that many VM-privileged threads, + * so let's wake them all up, even if we don't quite + * have enough pages to satisfy them all. + */ + thread_wakeup((event_t)&vm_page_free_wanted_privileged); + } + if (need_wakeup != 0 && vm_page_free_wanted == 0) { + /* + * We don't expect to have any more waiters + * after this, so let's wake them all up at + * once. + */ thread_wakeup((event_t) &vm_page_free_count); - } else { - while (available_pages--) { - vm_page_free_wanted--; - thread_wakeup_one((event_t) &vm_page_free_count); - } + } else for (; need_wakeup != 0; need_wakeup--) { + /* + * Wake up one waiter per page we just released. + */ + thread_wakeup_one((event_t) &vm_page_free_count); } + + VM_CHECK_MEMORYSTATUS; } - mutex_unlock(&vm_page_queue_free_lock); } } @@ -1937,23 +2651,67 @@ vm_page_wire( // dbgLog(current_thread(), mem->offset, mem->object, 1); /* (TEST/DEBUG) */ VM_PAGE_CHECK(mem); + if (mem->object) { + vm_object_lock_assert_exclusive(mem->object); + } else { + /* + * In theory, the page should be in an object before it + * gets wired, since we need to hold the object lock + * to update some fields in the page structure. + * However, some code (i386 pmap, for example) might want + * to wire a page before it gets inserted into an object. + * That's somewhat OK, as long as nobody else can get to + * that page and update it at the same time. + */ + } #if DEBUG - if (mem->object) - _mutex_assert(&mem->object->Lock, MA_OWNED); - _mutex_assert(&vm_page_queue_lock, MA_OWNED); + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); #endif - if (mem->wire_count == 0) { + if ( !VM_PAGE_WIRED(mem)) { + + if (mem->pageout_queue) { + mem->pageout = FALSE; + vm_pageout_throttle_up(mem); + } VM_PAGE_QUEUES_REMOVE(mem); + + if (mem->object) { + mem->object->wired_page_count++; + assert(mem->object->resident_page_count >= + mem->object->wired_page_count); + if (mem->object->purgable == VM_PURGABLE_VOLATILE) { + assert(vm_page_purgeable_count > 0); + OSAddAtomic(-1, &vm_page_purgeable_count); + OSAddAtomic(1, &vm_page_purgeable_wired_count); + } + if (mem->object->all_reusable) { + /* + * Wired pages are not counted as "re-usable" + * in "all_reusable" VM objects, so nothing + * to do here. + */ + } else if (mem->reusable) { + /* + * This page is not "re-usable" when it's + * wired, so adjust its state and the + * accounting. + */ + vm_object_reuse_pages(mem->object, + mem->offset, + mem->offset+PAGE_SIZE_64, + FALSE); + } + } + assert(!mem->reusable); + if (!mem->private && !mem->fictitious && !mem->gobbled) vm_page_wire_count++; if (mem->gobbled) vm_page_gobble_count--; mem->gobbled = FALSE; - if(mem->zero_fill) { - /* depends on the queues lock */ - vm_zf_count-=1; - mem->zero_fill = FALSE; - } + + VM_CHECK_MEMORYSTATUS; + /* * ENCRYPTED SWAP: * The page could be encrypted, but @@ -1966,6 +2724,7 @@ vm_page_wire( } assert(!mem->gobbled); mem->wire_count++; + VM_PAGE_CHECK(mem); } /* @@ -1979,13 +2738,13 @@ void vm_page_gobble( register vm_page_t mem) { - vm_page_lock_queues(); + vm_page_lockspin_queues(); VM_PAGE_CHECK(mem); assert(!mem->gobbled); - assert(mem->wire_count == 0); + assert( !VM_PAGE_WIRED(mem)); - if (!mem->gobbled && mem->wire_count == 0) { + if (!mem->gobbled && !VM_PAGE_WIRED(mem)) { if (!mem->private && !mem->fictitious) vm_page_wire_count++; } @@ -2004,29 +2763,47 @@ vm_page_gobble( */ void vm_page_unwire( - register vm_page_t mem) + vm_page_t mem, + boolean_t queueit) { // dbgLog(current_thread(), mem->offset, mem->object, 0); /* (TEST/DEBUG) */ VM_PAGE_CHECK(mem); - assert(mem->wire_count > 0); + assert(VM_PAGE_WIRED(mem)); + assert(mem->object != VM_OBJECT_NULL); #if DEBUG - if (mem->object) - _mutex_assert(&mem->object->Lock, MA_OWNED); - _mutex_assert(&vm_page_queue_lock, MA_OWNED); + vm_object_lock_assert_exclusive(mem->object); + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); #endif if (--mem->wire_count == 0) { assert(!mem->private && !mem->fictitious); vm_page_wire_count--; + assert(mem->object->wired_page_count > 0); + mem->object->wired_page_count--; + assert(mem->object->resident_page_count >= + mem->object->wired_page_count); + if (mem->object->purgable == VM_PURGABLE_VOLATILE) { + OSAddAtomic(+1, &vm_page_purgeable_count); + assert(vm_page_purgeable_wired_count > 0); + OSAddAtomic(-1, &vm_page_purgeable_wired_count); + } assert(!mem->laundry); assert(mem->object != kernel_object); assert(mem->pageq.next == NULL && mem->pageq.prev == NULL); - queue_enter(&vm_page_queue_active, mem, vm_page_t, pageq); - vm_page_active_count++; - mem->active = TRUE; - mem->reference = TRUE; + + if (queueit == TRUE) { + if (mem->object->purgable == VM_PURGABLE_EMPTY) { + vm_page_deactivate(mem); + } else { + vm_page_activate(mem); + } + } + + VM_CHECK_MEMORYSTATUS; + } + VM_PAGE_CHECK(mem); } /* @@ -2040,14 +2817,25 @@ vm_page_unwire( */ void vm_page_deactivate( - register vm_page_t m) + vm_page_t m) +{ + vm_page_deactivate_internal(m, TRUE); +} + + +void +vm_page_deactivate_internal( + vm_page_t m, + boolean_t clear_hw_reference) { + VM_PAGE_CHECK(m); assert(m->object != kernel_object); + assert(m->phys_page != vm_page_guard_addr); // dbgLog(m->phys_page, vm_page_free_count, vm_page_wire_count, 6); /* (TEST/DEBUG) */ #if DEBUG - _mutex_assert(&vm_page_queue_lock, MA_OWNED); + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); #endif /* * This page is no longer very interesting. If it was @@ -2056,46 +2844,108 @@ vm_page_deactivate( * inactive queue. Note wired pages should not have * their reference bit cleared. */ + assert ( !(m->absent && !m->unusual)); + if (m->gobbled) { /* can this happen? */ - assert(m->wire_count == 0); + assert( !VM_PAGE_WIRED(m)); + if (!m->private && !m->fictitious) vm_page_wire_count--; vm_page_gobble_count--; m->gobbled = FALSE; } - if (m->private || (m->wire_count != 0)) + /* + * if this page is currently on the pageout queue, we can't do the + * VM_PAGE_QUEUES_REMOVE (which doesn't handle the pageout queue case) + * and we can't remove it manually since we would need the object lock + * (which is not required here) to decrement the activity_in_progress + * reference which is held on the object while the page is in the pageout queue... + * just let the normal laundry processing proceed + */ + if (m->pageout_queue || m->private || m->fictitious || m->compressor || (VM_PAGE_WIRED(m))) return; - if (m->active || (m->inactive && m->reference)) { - if (!m->fictitious && !m->absent) - pmap_clear_reference(m->phys_page); - m->reference = FALSE; + + if (!m->absent && clear_hw_reference == TRUE) + pmap_clear_reference(m->phys_page); + + m->reference = FALSE; + m->no_cache = FALSE; + + if (!m->inactive) { VM_PAGE_QUEUES_REMOVE(m); - } - if (m->wire_count == 0 && !m->inactive) { - m->page_ticket = vm_page_ticket; - vm_page_ticket_roll++; - if(vm_page_ticket_roll == VM_PAGE_TICKETS_IN_ROLL) { - vm_page_ticket_roll = 0; - if(vm_page_ticket == VM_PAGE_TICKET_ROLL_IDS) - vm_page_ticket= 0; - else - vm_page_ticket++; - } - - assert(!m->laundry); - assert(m->pageq.next == NULL && m->pageq.prev == NULL); - if(m->zero_fill) { - queue_enter(&vm_page_queue_zf, m, vm_page_t, pageq); + if (!VM_DYNAMIC_PAGING_ENABLED(memory_manager_default) && + m->dirty && m->object->internal && + (m->object->purgable == VM_PURGABLE_DENY || + m->object->purgable == VM_PURGABLE_NONVOLATILE || + m->object->purgable == VM_PURGABLE_VOLATILE)) { + queue_enter(&vm_page_queue_throttled, m, vm_page_t, pageq); + m->throttled = TRUE; + vm_page_throttled_count++; } else { - queue_enter(&vm_page_queue_inactive, - m, vm_page_t, pageq); + if (m->object->named && m->object->ref_count == 1) { + vm_page_speculate(m, FALSE); +#if DEVELOPMENT || DEBUG + vm_page_speculative_recreated++; +#endif + } else { + VM_PAGE_ENQUEUE_INACTIVE(m, FALSE); + } } + } +} + +/* + * vm_page_enqueue_cleaned + * + * Put the page on the cleaned queue, mark it cleaned, etc. + * Being on the cleaned queue (and having m->clean_queue set) + * does ** NOT ** guarantee that the page is clean! + * + * Call with the queues lock held. + */ + +void vm_page_enqueue_cleaned(vm_page_t m) +{ + assert(m->phys_page != vm_page_guard_addr); +#if DEBUG + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); +#endif + assert( !(m->absent && !m->unusual)); + + if (m->gobbled) { + assert( !VM_PAGE_WIRED(m)); + if (!m->private && !m->fictitious) + vm_page_wire_count--; + vm_page_gobble_count--; + m->gobbled = FALSE; + } + /* + * if this page is currently on the pageout queue, we can't do the + * VM_PAGE_QUEUES_REMOVE (which doesn't handle the pageout queue case) + * and we can't remove it manually since we would need the object lock + * (which is not required here) to decrement the activity_in_progress + * reference which is held on the object while the page is in the pageout queue... + * just let the normal laundry processing proceed + */ + if (m->clean_queue || m->pageout_queue || m->private || m->fictitious) + return; - m->inactive = TRUE; - if (!m->fictitious) - vm_page_inactive_count++; + VM_PAGE_QUEUES_REMOVE(m); + + queue_enter(&vm_page_queue_cleaned, m, vm_page_t, pageq); + m->clean_queue = TRUE; + vm_page_cleaned_count++; + + m->inactive = TRUE; + vm_page_inactive_count++; + if (m->object->internal) { + vm_page_pageable_internal_count++; + } else { + vm_page_pageable_external_count++; } + + vm_pageout_enqueued_cleaned++; } /* @@ -2106,131 +2956,552 @@ vm_page_deactivate( * The page queues must be locked. */ +#if CONFIG_JETSAM +#if LATENCY_JETSAM +extern struct vm_page jetsam_latency_page[NUM_OF_JETSAM_LATENCY_TOKENS]; +#endif /* LATENCY_JETSAM */ +#endif /* CONFIG_JETSAM */ + void vm_page_activate( register vm_page_t m) { VM_PAGE_CHECK(m); +#ifdef FIXME_4778297 assert(m->object != kernel_object); +#endif + assert(m->phys_page != vm_page_guard_addr); #if DEBUG - _mutex_assert(&vm_page_queue_lock, MA_OWNED); + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); #endif + assert( !(m->absent && !m->unusual)); + if (m->gobbled) { - assert(m->wire_count == 0); + assert( !VM_PAGE_WIRED(m)); if (!m->private && !m->fictitious) vm_page_wire_count--; vm_page_gobble_count--; m->gobbled = FALSE; } - if (m->private) + /* + * if this page is currently on the pageout queue, we can't do the + * VM_PAGE_QUEUES_REMOVE (which doesn't handle the pageout queue case) + * and we can't remove it manually since we would need the object lock + * (which is not required here) to decrement the activity_in_progress + * reference which is held on the object while the page is in the pageout queue... + * just let the normal laundry processing proceed + */ + if (m->pageout_queue || m->private || m->fictitious || m->compressor) return; - if (m->inactive) { - assert(!m->laundry); - if (m->zero_fill) { - queue_remove(&vm_page_queue_zf, m, vm_page_t, pageq); - } else { - queue_remove(&vm_page_queue_inactive, - m, vm_page_t, pageq); - } - m->pageq.next = NULL; - m->pageq.prev = NULL; - if (!m->fictitious) - vm_page_inactive_count--; - m->inactive = FALSE; - } - if (m->wire_count == 0) { #if DEBUG - if (m->active) - panic("vm_page_activate: already active"); + if (m->active) + panic("vm_page_activate: already active"); #endif - assert(!m->laundry); - assert(m->pageq.next == NULL && m->pageq.prev == NULL); - queue_enter(&vm_page_queue_active, m, vm_page_t, pageq); - m->active = TRUE; - m->reference = TRUE; - if (!m->fictitious) + + if (m->speculative) { + DTRACE_VM2(pgrec, int, 1, (uint64_t *), NULL); + DTRACE_VM2(pgfrec, int, 1, (uint64_t *), NULL); + } + + VM_PAGE_QUEUES_REMOVE(m); + + if ( !VM_PAGE_WIRED(m)) { + + if (!VM_DYNAMIC_PAGING_ENABLED(memory_manager_default) && + m->dirty && m->object->internal && + (m->object->purgable == VM_PURGABLE_DENY || + m->object->purgable == VM_PURGABLE_NONVOLATILE || + m->object->purgable == VM_PURGABLE_VOLATILE)) { + queue_enter(&vm_page_queue_throttled, m, vm_page_t, pageq); + m->throttled = TRUE; + vm_page_throttled_count++; + } else { + queue_enter(&vm_page_queue_active, m, vm_page_t, pageq); + m->active = TRUE; vm_page_active_count++; + if (m->object->internal) { + vm_page_pageable_internal_count++; + } else { + vm_page_pageable_external_count++; + } +#if LATENCY_JETSAM + if (jlp_init) { + uint64_t now = mach_absolute_time(); + uint64_t delta = now - jlp_time; + clock_sec_t jl_secs = 0; + clock_usec_t jl_usecs = 0; + vm_page_t jlp; + + absolutetime_to_microtime(delta, &jl_secs, &jl_usecs); + + jl_usecs += jl_secs * USEC_PER_SEC; + if (jl_usecs >= JETSAM_LATENCY_TOKEN_AGE) { + + jlp = &jetsam_latency_page[jlp_current]; + if (jlp->active) { + queue_remove(&vm_page_queue_active, jlp, vm_page_t, pageq); + } + queue_enter(&vm_page_queue_active, jlp, vm_page_t, pageq); + + jlp->active = TRUE; + + jlp->offset = now; + jlp_time = jlp->offset; + + if(++jlp_current == NUM_OF_JETSAM_LATENCY_TOKENS) { + jlp_current = 0; + } + + } + } +#endif /* LATENCY_JETSAM */ + } + m->reference = TRUE; + m->no_cache = FALSE; } + VM_PAGE_CHECK(m); } + /* - * vm_page_part_zero_fill: + * vm_page_speculate: * - * Zero-fill a part of the page. + * Put the specified page on the speculative list (if appropriate). + * + * The page queues must be locked. */ void -vm_page_part_zero_fill( +vm_page_speculate( vm_page_t m, - vm_offset_t m_pa, - vm_size_t len) + boolean_t new) { - vm_page_t tmp; + struct vm_speculative_age_q *aq; VM_PAGE_CHECK(m); -#ifdef PMAP_ZERO_PART_PAGE_IMPLEMENTED - pmap_zero_part_page(m->phys_page, m_pa, len); -#else - while (1) { - tmp = vm_page_grab(); - if (tmp == VM_PAGE_NULL) { - vm_page_wait(THREAD_UNINT); - continue; - } - break; - } - vm_page_zero_fill(tmp); - if(m_pa != 0) { - vm_page_part_copy(m, 0, tmp, 0, m_pa); - } - if((m_pa + len) < PAGE_SIZE) { - vm_page_part_copy(m, m_pa + len, tmp, - m_pa + len, PAGE_SIZE - (m_pa + len)); - } - vm_page_copy(tmp,m); - vm_page_lock_queues(); - vm_page_free(tmp); - vm_page_unlock_queues(); + assert(m->object != kernel_object); + assert(m->phys_page != vm_page_guard_addr); +#if DEBUG + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); #endif + assert( !(m->absent && !m->unusual)); -} + /* + * if this page is currently on the pageout queue, we can't do the + * VM_PAGE_QUEUES_REMOVE (which doesn't handle the pageout queue case) + * and we can't remove it manually since we would need the object lock + * (which is not required here) to decrement the activity_in_progress + * reference which is held on the object while the page is in the pageout queue... + * just let the normal laundry processing proceed + */ + if (m->pageout_queue || m->private || m->fictitious || m->compressor) + return; -/* - * vm_page_zero_fill: - * - * Zero-fill the specified page. - */ -void -vm_page_zero_fill( - vm_page_t m) -{ - XPR(XPR_VM_PAGE, - "vm_page_zero_fill, object 0x%X offset 0x%X page 0x%X\n", - (integer_t)m->object, (integer_t)m->offset, (integer_t)m, 0,0); + VM_PAGE_QUEUES_REMOVE(m); - VM_PAGE_CHECK(m); + if ( !VM_PAGE_WIRED(m)) { + mach_timespec_t ts; + clock_sec_t sec; + clock_nsec_t nsec; -// dbgTrace(0xAEAEAEAE, m->phys_page, 0); /* (BRINGUP) */ - pmap_zero_page(m->phys_page); -} + clock_get_system_nanotime(&sec, &nsec); + ts.tv_sec = (unsigned int) sec; + ts.tv_nsec = nsec; -/* - * vm_page_part_copy: - * - * copy part of one page to another - */ + if (vm_page_speculative_count == 0) { -void -vm_page_part_copy( - vm_page_t src_m, - vm_offset_t src_pa, - vm_page_t dst_m, + speculative_age_index = VM_PAGE_MIN_SPECULATIVE_AGE_Q; + speculative_steal_index = VM_PAGE_MIN_SPECULATIVE_AGE_Q; + + aq = &vm_page_queue_speculative[speculative_age_index]; + + /* + * set the timer to begin a new group + */ + aq->age_ts.tv_sec = vm_page_speculative_q_age_ms / 1000; + aq->age_ts.tv_nsec = (vm_page_speculative_q_age_ms % 1000) * 1000 * NSEC_PER_USEC; + + ADD_MACH_TIMESPEC(&aq->age_ts, &ts); + } else { + aq = &vm_page_queue_speculative[speculative_age_index]; + + if (CMP_MACH_TIMESPEC(&ts, &aq->age_ts) >= 0) { + + speculative_age_index++; + + if (speculative_age_index > VM_PAGE_MAX_SPECULATIVE_AGE_Q) + speculative_age_index = VM_PAGE_MIN_SPECULATIVE_AGE_Q; + if (speculative_age_index == speculative_steal_index) { + speculative_steal_index = speculative_age_index + 1; + + if (speculative_steal_index > VM_PAGE_MAX_SPECULATIVE_AGE_Q) + speculative_steal_index = VM_PAGE_MIN_SPECULATIVE_AGE_Q; + } + aq = &vm_page_queue_speculative[speculative_age_index]; + + if (!queue_empty(&aq->age_q)) + vm_page_speculate_ageit(aq); + + aq->age_ts.tv_sec = vm_page_speculative_q_age_ms / 1000; + aq->age_ts.tv_nsec = (vm_page_speculative_q_age_ms % 1000) * 1000 * NSEC_PER_USEC; + + ADD_MACH_TIMESPEC(&aq->age_ts, &ts); + } + } + enqueue_tail(&aq->age_q, &m->pageq); + m->speculative = TRUE; + vm_page_speculative_count++; + if (m->object->internal) { + vm_page_pageable_internal_count++; + } else { + vm_page_pageable_external_count++; + } + + if (new == TRUE) { + vm_object_lock_assert_exclusive(m->object); + + m->object->pages_created++; +#if DEVELOPMENT || DEBUG + vm_page_speculative_created++; +#endif + } + } + VM_PAGE_CHECK(m); +} + + +/* + * move pages from the specified aging bin to + * the speculative bin that pageout_scan claims from + * + * The page queues must be locked. + */ +void +vm_page_speculate_ageit(struct vm_speculative_age_q *aq) +{ + struct vm_speculative_age_q *sq; + vm_page_t t; + + sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q]; + + if (queue_empty(&sq->age_q)) { + sq->age_q.next = aq->age_q.next; + sq->age_q.prev = aq->age_q.prev; + + t = (vm_page_t)sq->age_q.next; + t->pageq.prev = &sq->age_q; + + t = (vm_page_t)sq->age_q.prev; + t->pageq.next = &sq->age_q; + } else { + t = (vm_page_t)sq->age_q.prev; + t->pageq.next = aq->age_q.next; + + t = (vm_page_t)aq->age_q.next; + t->pageq.prev = sq->age_q.prev; + + t = (vm_page_t)aq->age_q.prev; + t->pageq.next = &sq->age_q; + + sq->age_q.prev = aq->age_q.prev; + } + queue_init(&aq->age_q); +} + + +void +vm_page_lru( + vm_page_t m) +{ + VM_PAGE_CHECK(m); + assert(m->object != kernel_object); + assert(m->phys_page != vm_page_guard_addr); + +#if DEBUG + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); +#endif + /* + * if this page is currently on the pageout queue, we can't do the + * VM_PAGE_QUEUES_REMOVE (which doesn't handle the pageout queue case) + * and we can't remove it manually since we would need the object lock + * (which is not required here) to decrement the activity_in_progress + * reference which is held on the object while the page is in the pageout queue... + * just let the normal laundry processing proceed + */ + if (m->pageout_queue || m->private || m->compressor || (VM_PAGE_WIRED(m))) + return; + + m->no_cache = FALSE; + + VM_PAGE_QUEUES_REMOVE(m); + + VM_PAGE_ENQUEUE_INACTIVE(m, FALSE); +} + + +void +vm_page_reactivate_all_throttled(void) +{ + vm_page_t first_throttled, last_throttled; + vm_page_t first_active; + vm_page_t m; + int extra_active_count; + int extra_internal_count, extra_external_count; + + if (!VM_DYNAMIC_PAGING_ENABLED(memory_manager_default)) + return; + + extra_active_count = 0; + extra_internal_count = 0; + extra_external_count = 0; + vm_page_lock_queues(); + if (! queue_empty(&vm_page_queue_throttled)) { + /* + * Switch "throttled" pages to "active". + */ + queue_iterate(&vm_page_queue_throttled, m, vm_page_t, pageq) { + VM_PAGE_CHECK(m); + assert(m->throttled); + assert(!m->active); + assert(!m->inactive); + assert(!m->speculative); + assert(!VM_PAGE_WIRED(m)); + + extra_active_count++; + if (m->object->internal) { + extra_internal_count++; + } else { + extra_external_count++; + } + + m->throttled = FALSE; + m->active = TRUE; + VM_PAGE_CHECK(m); + } + + /* + * Transfer the entire throttled queue to a regular LRU page queues. + * We insert it at the head of the active queue, so that these pages + * get re-evaluated by the LRU algorithm first, since they've been + * completely out of it until now. + */ + first_throttled = (vm_page_t) queue_first(&vm_page_queue_throttled); + last_throttled = (vm_page_t) queue_last(&vm_page_queue_throttled); + first_active = (vm_page_t) queue_first(&vm_page_queue_active); + if (queue_empty(&vm_page_queue_active)) { + queue_last(&vm_page_queue_active) = (queue_entry_t) last_throttled; + } else { + queue_prev(&first_active->pageq) = (queue_entry_t) last_throttled; + } + queue_first(&vm_page_queue_active) = (queue_entry_t) first_throttled; + queue_prev(&first_throttled->pageq) = (queue_entry_t) &vm_page_queue_active; + queue_next(&last_throttled->pageq) = (queue_entry_t) first_active; + +#if DEBUG + printf("reactivated %d throttled pages\n", vm_page_throttled_count); +#endif + queue_init(&vm_page_queue_throttled); + /* + * Adjust the global page counts. + */ + vm_page_active_count += extra_active_count; + vm_page_pageable_internal_count += extra_internal_count; + vm_page_pageable_external_count += extra_external_count; + vm_page_throttled_count = 0; + } + assert(vm_page_throttled_count == 0); + assert(queue_empty(&vm_page_queue_throttled)); + vm_page_unlock_queues(); +} + + +/* + * move pages from the indicated local queue to the global active queue + * its ok to fail if we're below the hard limit and force == FALSE + * the nolocks == TRUE case is to allow this function to be run on + * the hibernate path + */ + +void +vm_page_reactivate_local(uint32_t lid, boolean_t force, boolean_t nolocks) +{ + struct vpl *lq; + vm_page_t first_local, last_local; + vm_page_t first_active; + vm_page_t m; + uint32_t count = 0; + + if (vm_page_local_q == NULL) + return; + + lq = &vm_page_local_q[lid].vpl_un.vpl; + + if (nolocks == FALSE) { + if (lq->vpl_count < vm_page_local_q_hard_limit && force == FALSE) { + if ( !vm_page_trylockspin_queues()) + return; + } else + vm_page_lockspin_queues(); + + VPL_LOCK(&lq->vpl_lock); + } + if (lq->vpl_count) { + /* + * Switch "local" pages to "active". + */ + assert(!queue_empty(&lq->vpl_queue)); + + queue_iterate(&lq->vpl_queue, m, vm_page_t, pageq) { + VM_PAGE_CHECK(m); + assert(m->local); + assert(!m->active); + assert(!m->inactive); + assert(!m->speculative); + assert(!VM_PAGE_WIRED(m)); + assert(!m->throttled); + assert(!m->fictitious); + + if (m->local_id != lid) + panic("vm_page_reactivate_local: found vm_page_t(%p) with wrong cpuid", m); + + m->local_id = 0; + m->local = FALSE; + m->active = TRUE; + VM_PAGE_CHECK(m); + + count++; + } + if (count != lq->vpl_count) + panic("vm_page_reactivate_local: count = %d, vm_page_local_count = %d\n", count, lq->vpl_count); + + /* + * Transfer the entire local queue to a regular LRU page queues. + */ + first_local = (vm_page_t) queue_first(&lq->vpl_queue); + last_local = (vm_page_t) queue_last(&lq->vpl_queue); + first_active = (vm_page_t) queue_first(&vm_page_queue_active); + + if (queue_empty(&vm_page_queue_active)) { + queue_last(&vm_page_queue_active) = (queue_entry_t) last_local; + } else { + queue_prev(&first_active->pageq) = (queue_entry_t) last_local; + } + queue_first(&vm_page_queue_active) = (queue_entry_t) first_local; + queue_prev(&first_local->pageq) = (queue_entry_t) &vm_page_queue_active; + queue_next(&last_local->pageq) = (queue_entry_t) first_active; + + queue_init(&lq->vpl_queue); + /* + * Adjust the global page counts. + */ + vm_page_active_count += lq->vpl_count; + vm_page_pageable_internal_count += lq->vpl_internal_count; + vm_page_pageable_external_count += lq->vpl_external_count; + lq->vpl_count = 0; + lq->vpl_internal_count = 0; + lq->vpl_external_count = 0; + } + assert(queue_empty(&lq->vpl_queue)); + + if (nolocks == FALSE) { + VPL_UNLOCK(&lq->vpl_lock); + vm_page_unlock_queues(); + } +} + +/* + * vm_page_part_zero_fill: + * + * Zero-fill a part of the page. + */ +#define PMAP_ZERO_PART_PAGE_IMPLEMENTED +void +vm_page_part_zero_fill( + vm_page_t m, + vm_offset_t m_pa, + vm_size_t len) +{ + +#if 0 + /* + * we don't hold the page queue lock + * so this check isn't safe to make + */ + VM_PAGE_CHECK(m); +#endif + +#ifdef PMAP_ZERO_PART_PAGE_IMPLEMENTED + pmap_zero_part_page(m->phys_page, m_pa, len); +#else + vm_page_t tmp; + while (1) { + tmp = vm_page_grab(); + if (tmp == VM_PAGE_NULL) { + vm_page_wait(THREAD_UNINT); + continue; + } + break; + } + vm_page_zero_fill(tmp); + if(m_pa != 0) { + vm_page_part_copy(m, 0, tmp, 0, m_pa); + } + if((m_pa + len) < PAGE_SIZE) { + vm_page_part_copy(m, m_pa + len, tmp, + m_pa + len, PAGE_SIZE - (m_pa + len)); + } + vm_page_copy(tmp,m); + VM_PAGE_FREE(tmp); +#endif + +} + +/* + * vm_page_zero_fill: + * + * Zero-fill the specified page. + */ +void +vm_page_zero_fill( + vm_page_t m) +{ + XPR(XPR_VM_PAGE, + "vm_page_zero_fill, object 0x%X offset 0x%X page 0x%X\n", + m->object, m->offset, m, 0,0); +#if 0 + /* + * we don't hold the page queue lock + * so this check isn't safe to make + */ + VM_PAGE_CHECK(m); +#endif + +// dbgTrace(0xAEAEAEAE, m->phys_page, 0); /* (BRINGUP) */ + pmap_zero_page(m->phys_page); +} + +/* + * vm_page_part_copy: + * + * copy part of one page to another + */ + +void +vm_page_part_copy( + vm_page_t src_m, + vm_offset_t src_pa, + vm_page_t dst_m, vm_offset_t dst_pa, vm_size_t len) { +#if 0 + /* + * we don't hold the page queue lock + * so this check isn't safe to make + */ VM_PAGE_CHECK(src_m); VM_PAGE_CHECK(dst_m); - +#endif pmap_copy_part_page(src_m->phys_page, src_pa, dst_m->phys_page, dst_pa, len); } @@ -2245,387 +3516,2785 @@ vm_page_part_copy( * make sure the page is decrypted first, if necessary. */ +int vm_page_copy_cs_validations = 0; +int vm_page_copy_cs_tainted = 0; + +void +vm_page_copy( + vm_page_t src_m, + vm_page_t dest_m) +{ + XPR(XPR_VM_PAGE, + "vm_page_copy, object 0x%X offset 0x%X to object 0x%X offset 0x%X\n", + src_m->object, src_m->offset, + dest_m->object, dest_m->offset, + 0); +#if 0 + /* + * we don't hold the page queue lock + * so this check isn't safe to make + */ + VM_PAGE_CHECK(src_m); + VM_PAGE_CHECK(dest_m); +#endif + vm_object_lock_assert_held(src_m->object); + + /* + * ENCRYPTED SWAP: + * The source page should not be encrypted at this point. + * The destination page will therefore not contain encrypted + * data after the copy. + */ + if (src_m->encrypted) { + panic("vm_page_copy: source page %p is encrypted\n", src_m); + } + dest_m->encrypted = FALSE; + + if (src_m->object != VM_OBJECT_NULL && + src_m->object->code_signed) { + /* + * We're copying a page from a code-signed object. + * Whoever ends up mapping the copy page might care about + * the original page's integrity, so let's validate the + * source page now. + */ + vm_page_copy_cs_validations++; + vm_page_validate_cs(src_m); + } + + if (vm_page_is_slideable(src_m)) { + boolean_t was_busy = src_m->busy; + src_m->busy = TRUE; + (void) vm_page_slide(src_m, 0); + assert(src_m->busy); + if (!was_busy) { + PAGE_WAKEUP_DONE(src_m); + } + } + + /* + * Propagate the cs_tainted bit to the copy page. Do not propagate + * the cs_validated bit. + */ + dest_m->cs_tainted = src_m->cs_tainted; + if (dest_m->cs_tainted) { + vm_page_copy_cs_tainted++; + } + dest_m->slid = src_m->slid; + dest_m->error = src_m->error; /* sliding src_m might have failed... */ + pmap_copy_page(src_m->phys_page, dest_m->phys_page); +} + +#if MACH_ASSERT +static void +_vm_page_print( + vm_page_t p) +{ + printf("vm_page %p: \n", p); + printf(" pageq: next=%p prev=%p\n", p->pageq.next, p->pageq.prev); + printf(" listq: next=%p prev=%p\n", p->listq.next, p->listq.prev); + printf(" next=%p\n", p->next); + printf(" object=%p offset=0x%llx\n", p->object, p->offset); + printf(" wire_count=%u\n", p->wire_count); + + printf(" %slocal, %sinactive, %sactive, %spageout_queue, %sspeculative, %slaundry\n", + (p->local ? "" : "!"), + (p->inactive ? "" : "!"), + (p->active ? "" : "!"), + (p->pageout_queue ? "" : "!"), + (p->speculative ? "" : "!"), + (p->laundry ? "" : "!")); + printf(" %sfree, %sref, %sgobbled, %sprivate, %sthrottled\n", + (p->free ? "" : "!"), + (p->reference ? "" : "!"), + (p->gobbled ? "" : "!"), + (p->private ? "" : "!"), + (p->throttled ? "" : "!")); + printf(" %sbusy, %swanted, %stabled, %sfictitious, %spmapped, %swpmapped\n", + (p->busy ? "" : "!"), + (p->wanted ? "" : "!"), + (p->tabled ? "" : "!"), + (p->fictitious ? "" : "!"), + (p->pmapped ? "" : "!"), + (p->wpmapped ? "" : "!")); + printf(" %spageout, %sabsent, %serror, %sdirty, %scleaning, %sprecious, %sclustered\n", + (p->pageout ? "" : "!"), + (p->absent ? "" : "!"), + (p->error ? "" : "!"), + (p->dirty ? "" : "!"), + (p->cleaning ? "" : "!"), + (p->precious ? "" : "!"), + (p->clustered ? "" : "!")); + printf(" %soverwriting, %srestart, %sunusual, %sencrypted, %sencrypted_cleaning\n", + (p->overwriting ? "" : "!"), + (p->restart ? "" : "!"), + (p->unusual ? "" : "!"), + (p->encrypted ? "" : "!"), + (p->encrypted_cleaning ? "" : "!")); + printf(" %scs_validated, %scs_tainted, %sno_cache\n", + (p->cs_validated ? "" : "!"), + (p->cs_tainted ? "" : "!"), + (p->no_cache ? "" : "!")); + + printf("phys_page=0x%x\n", p->phys_page); +} + +/* + * Check that the list of pages is ordered by + * ascending physical address and has no holes. + */ +static int +vm_page_verify_contiguous( + vm_page_t pages, + unsigned int npages) +{ + register vm_page_t m; + unsigned int page_count; + vm_offset_t prev_addr; + + prev_addr = pages->phys_page; + page_count = 1; + for (m = NEXT_PAGE(pages); m != VM_PAGE_NULL; m = NEXT_PAGE(m)) { + if (m->phys_page != prev_addr + 1) { + printf("m %p prev_addr 0x%lx, current addr 0x%x\n", + m, (long)prev_addr, m->phys_page); + printf("pages %p page_count %d npages %d\n", pages, page_count, npages); + panic("vm_page_verify_contiguous: not contiguous!"); + } + prev_addr = m->phys_page; + ++page_count; + } + if (page_count != npages) { + printf("pages %p actual count 0x%x but requested 0x%x\n", + pages, page_count, npages); + panic("vm_page_verify_contiguous: count error"); + } + return 1; +} + + +/* + * Check the free lists for proper length etc. + */ +static unsigned int +vm_page_verify_free_list( + queue_head_t *vm_page_queue, + unsigned int color, + vm_page_t look_for_page, + boolean_t expect_page) +{ + unsigned int npages; + vm_page_t m; + vm_page_t prev_m; + boolean_t found_page; + + found_page = FALSE; + npages = 0; + prev_m = (vm_page_t) vm_page_queue; + queue_iterate(vm_page_queue, + m, + vm_page_t, + pageq) { + + if (m == look_for_page) { + found_page = TRUE; + } + if ((vm_page_t) m->pageq.prev != prev_m) + panic("vm_page_verify_free_list(color=%u, npages=%u): page %p corrupted prev ptr %p instead of %p\n", + color, npages, m, m->pageq.prev, prev_m); + if ( ! m->busy ) + panic("vm_page_verify_free_list(color=%u, npages=%u): page %p not busy\n", + color, npages, m); + if (color != (unsigned int) -1) { + if ((m->phys_page & vm_color_mask) != color) + panic("vm_page_verify_free_list(color=%u, npages=%u): page %p wrong color %u instead of %u\n", + color, npages, m, m->phys_page & vm_color_mask, color); + if ( ! m->free ) + panic("vm_page_verify_free_list(color=%u, npages=%u): page %p not free\n", + color, npages, m); + } + ++npages; + prev_m = m; + } + if (look_for_page != VM_PAGE_NULL) { + unsigned int other_color; + + if (expect_page && !found_page) { + printf("vm_page_verify_free_list(color=%u, npages=%u): page %p not found phys=%u\n", + color, npages, look_for_page, look_for_page->phys_page); + _vm_page_print(look_for_page); + for (other_color = 0; + other_color < vm_colors; + other_color++) { + if (other_color == color) + continue; + vm_page_verify_free_list(&vm_page_queue_free[other_color], + other_color, look_for_page, FALSE); + } + if (color == (unsigned int) -1) { + vm_page_verify_free_list(&vm_lopage_queue_free, + (unsigned int) -1, look_for_page, FALSE); + } + panic("vm_page_verify_free_list(color=%u)\n", color); + } + if (!expect_page && found_page) { + printf("vm_page_verify_free_list(color=%u, npages=%u): page %p found phys=%u\n", + color, npages, look_for_page, look_for_page->phys_page); + } + } + return npages; +} + +static boolean_t vm_page_verify_free_lists_enabled = FALSE; +static void +vm_page_verify_free_lists( void ) +{ + unsigned int color, npages, nlopages; + + if (! vm_page_verify_free_lists_enabled) + return; + + npages = 0; + + lck_mtx_lock(&vm_page_queue_free_lock); + + for( color = 0; color < vm_colors; color++ ) { + npages += vm_page_verify_free_list(&vm_page_queue_free[color], + color, VM_PAGE_NULL, FALSE); + } + nlopages = vm_page_verify_free_list(&vm_lopage_queue_free, + (unsigned int) -1, + VM_PAGE_NULL, FALSE); + if (npages != vm_page_free_count || nlopages != vm_lopage_free_count) + panic("vm_page_verify_free_lists: " + "npages %u free_count %d nlopages %u lo_free_count %u", + npages, vm_page_free_count, nlopages, vm_lopage_free_count); + + lck_mtx_unlock(&vm_page_queue_free_lock); +} + +void +vm_page_queues_assert( + vm_page_t mem, + int val) +{ +#if DEBUG + lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED); +#endif + if (mem->free + mem->active + mem->inactive + mem->speculative + + mem->throttled + mem->pageout_queue > (val)) { + _vm_page_print(mem); + panic("vm_page_queues_assert(%p, %d)\n", mem, val); + } + if (VM_PAGE_WIRED(mem)) { + assert(!mem->active); + assert(!mem->inactive); + assert(!mem->speculative); + assert(!mem->throttled); + assert(!mem->pageout_queue); + } +} +#endif /* MACH_ASSERT */ + + +/* + * CONTIGUOUS PAGE ALLOCATION + * + * Find a region large enough to contain at least n pages + * of contiguous physical memory. + * + * This is done by traversing the vm_page_t array in a linear fashion + * we assume that the vm_page_t array has the avaiable physical pages in an + * ordered, ascending list... this is currently true of all our implementations + * and must remain so... there can be 'holes' in the array... we also can + * no longer tolerate the vm_page_t's in the list being 'freed' and reclaimed + * which use to happen via 'vm_page_convert'... that function was no longer + * being called and was removed... + * + * The basic flow consists of stabilizing some of the interesting state of + * a vm_page_t behind the vm_page_queue and vm_page_free locks... we start our + * sweep at the beginning of the array looking for pages that meet our criterea + * for a 'stealable' page... currently we are pretty conservative... if the page + * meets this criterea and is physically contiguous to the previous page in the 'run' + * we keep developing it. If we hit a page that doesn't fit, we reset our state + * and start to develop a new run... if at this point we've already considered + * at least MAX_CONSIDERED_BEFORE_YIELD pages, we'll drop the 2 locks we hold, + * and mutex_pause (which will yield the processor), to keep the latency low w/r + * to other threads trying to acquire free pages (or move pages from q to q), + * and then continue from the spot we left off... we only make 1 pass through the + * array. Once we have a 'run' that is long enough, we'll go into the loop which + * which steals the pages from the queues they're currently on... pages on the free + * queue can be stolen directly... pages that are on any of the other queues + * must be removed from the object they are tabled on... this requires taking the + * object lock... we do this as a 'try' to prevent deadlocks... if the 'try' fails + * or if the state of the page behind the vm_object lock is no longer viable, we'll + * dump the pages we've currently stolen back to the free list, and pick up our + * scan from the point where we aborted the 'current' run. + * + * + * Requirements: + * - neither vm_page_queue nor vm_free_list lock can be held on entry + * + * Returns a pointer to a list of gobbled/wired pages or VM_PAGE_NULL. + * + * Algorithm: + */ + +#define MAX_CONSIDERED_BEFORE_YIELD 1000 + + +#define RESET_STATE_OF_RUN() \ + MACRO_BEGIN \ + prevcontaddr = -2; \ + start_pnum = -1; \ + free_considered = 0; \ + substitute_needed = 0; \ + npages = 0; \ + MACRO_END + +/* + * Can we steal in-use (i.e. not free) pages when searching for + * physically-contiguous pages ? + */ +#define VM_PAGE_FIND_CONTIGUOUS_CAN_STEAL 1 + +static unsigned int vm_page_find_contiguous_last_idx = 0, vm_page_lomem_find_contiguous_last_idx = 0; +#if DEBUG +int vm_page_find_contig_debug = 0; +#endif + +static vm_page_t +vm_page_find_contiguous( + unsigned int contig_pages, + ppnum_t max_pnum, + ppnum_t pnum_mask, + boolean_t wire, + int flags) +{ + vm_page_t m = NULL; + ppnum_t prevcontaddr; + ppnum_t start_pnum; + unsigned int npages, considered, scanned; + unsigned int page_idx, start_idx, last_idx, orig_last_idx; + unsigned int idx_last_contig_page_found = 0; + int free_considered, free_available; + int substitute_needed; + boolean_t wrapped; +#if DEBUG + clock_sec_t tv_start_sec, tv_end_sec; + clock_usec_t tv_start_usec, tv_end_usec; +#endif +#if MACH_ASSERT + int yielded = 0; + int dumped_run = 0; + int stolen_pages = 0; + int compressed_pages = 0; +#endif + + if (contig_pages == 0) + return VM_PAGE_NULL; + +#if MACH_ASSERT + vm_page_verify_free_lists(); +#endif +#if DEBUG + clock_get_system_microtime(&tv_start_sec, &tv_start_usec); +#endif + PAGE_REPLACEMENT_ALLOWED(TRUE); + + vm_page_lock_queues(); + lck_mtx_lock(&vm_page_queue_free_lock); + + RESET_STATE_OF_RUN(); + + scanned = 0; + considered = 0; + free_available = vm_page_free_count - vm_page_free_reserved; + + wrapped = FALSE; + + if(flags & KMA_LOMEM) + idx_last_contig_page_found = vm_page_lomem_find_contiguous_last_idx; + else + idx_last_contig_page_found = vm_page_find_contiguous_last_idx; + + orig_last_idx = idx_last_contig_page_found; + last_idx = orig_last_idx; + + for (page_idx = last_idx, start_idx = last_idx; + npages < contig_pages && page_idx < vm_pages_count; + page_idx++) { +retry: + if (wrapped && + npages == 0 && + page_idx >= orig_last_idx) { + /* + * We're back where we started and we haven't + * found any suitable contiguous range. Let's + * give up. + */ + break; + } + scanned++; + m = &vm_pages[page_idx]; + + assert(!m->fictitious); + assert(!m->private); + + if (max_pnum && m->phys_page > max_pnum) { + /* no more low pages... */ + break; + } + if (!npages & ((m->phys_page & pnum_mask) != 0)) { + /* + * not aligned + */ + RESET_STATE_OF_RUN(); + + } else if (VM_PAGE_WIRED(m) || m->gobbled || + m->encrypted_cleaning || + m->pageout_queue || m->laundry || m->wanted || + m->cleaning || m->overwriting || m->pageout) { + /* + * page is in a transient state + * or a state we don't want to deal + * with, so don't consider it which + * means starting a new run + */ + RESET_STATE_OF_RUN(); + + } else if (!m->free && !m->active && !m->inactive && !m->speculative && !m->throttled && !m->compressor) { + /* + * page needs to be on one of our queues + * or it needs to belong to the compressor pool + * in order for it to be stable behind the + * locks we hold at this point... + * if not, don't consider it which + * means starting a new run + */ + RESET_STATE_OF_RUN(); + + } else if (!m->free && (!m->tabled || m->busy)) { + /* + * pages on the free list are always 'busy' + * so we couldn't test for 'busy' in the check + * for the transient states... pages that are + * 'free' are never 'tabled', so we also couldn't + * test for 'tabled'. So we check here to make + * sure that a non-free page is not busy and is + * tabled on an object... + * if not, don't consider it which + * means starting a new run + */ + RESET_STATE_OF_RUN(); + + } else { + if (m->phys_page != prevcontaddr + 1) { + if ((m->phys_page & pnum_mask) != 0) { + RESET_STATE_OF_RUN(); + goto did_consider; + } else { + npages = 1; + start_idx = page_idx; + start_pnum = m->phys_page; + } + } else { + npages++; + } + prevcontaddr = m->phys_page; + + VM_PAGE_CHECK(m); + if (m->free) { + free_considered++; + } else { + /* + * This page is not free. + * If we can't steal used pages, + * we have to give up this run + * and keep looking. + * Otherwise, we might need to + * move the contents of this page + * into a substitute page. + */ +#if VM_PAGE_FIND_CONTIGUOUS_CAN_STEAL + if (m->pmapped || m->dirty || m->precious) { + substitute_needed++; + } +#else + RESET_STATE_OF_RUN(); +#endif + } + + if ((free_considered + substitute_needed) > free_available) { + /* + * if we let this run continue + * we will end up dropping the vm_page_free_count + * below the reserve limit... we need to abort + * this run, but we can at least re-consider this + * page... thus the jump back to 'retry' + */ + RESET_STATE_OF_RUN(); + + if (free_available && considered <= MAX_CONSIDERED_BEFORE_YIELD) { + considered++; + goto retry; + } + /* + * free_available == 0 + * so can't consider any free pages... if + * we went to retry in this case, we'd + * get stuck looking at the same page + * w/o making any forward progress + * we also want to take this path if we've already + * reached our limit that controls the lock latency + */ + } + } +did_consider: + if (considered > MAX_CONSIDERED_BEFORE_YIELD && npages <= 1) { + + PAGE_REPLACEMENT_ALLOWED(FALSE); + + lck_mtx_unlock(&vm_page_queue_free_lock); + vm_page_unlock_queues(); + + mutex_pause(0); + + PAGE_REPLACEMENT_ALLOWED(TRUE); + + vm_page_lock_queues(); + lck_mtx_lock(&vm_page_queue_free_lock); + + RESET_STATE_OF_RUN(); + /* + * reset our free page limit since we + * dropped the lock protecting the vm_page_free_queue + */ + free_available = vm_page_free_count - vm_page_free_reserved; + considered = 0; +#if MACH_ASSERT + yielded++; +#endif + goto retry; + } + considered++; + } + m = VM_PAGE_NULL; + + if (npages != contig_pages) { + if (!wrapped) { + /* + * We didn't find a contiguous range but we didn't + * start from the very first page. + * Start again from the very first page. + */ + RESET_STATE_OF_RUN(); + if( flags & KMA_LOMEM) + idx_last_contig_page_found = vm_page_lomem_find_contiguous_last_idx = 0; + else + idx_last_contig_page_found = vm_page_find_contiguous_last_idx = 0; + last_idx = 0; + page_idx = last_idx; + wrapped = TRUE; + goto retry; + } + lck_mtx_unlock(&vm_page_queue_free_lock); + } else { + vm_page_t m1; + vm_page_t m2; + unsigned int cur_idx; + unsigned int tmp_start_idx; + vm_object_t locked_object = VM_OBJECT_NULL; + boolean_t abort_run = FALSE; + + assert(page_idx - start_idx == contig_pages); + + tmp_start_idx = start_idx; + + /* + * first pass through to pull the free pages + * off of the free queue so that in case we + * need substitute pages, we won't grab any + * of the free pages in the run... we'll clear + * the 'free' bit in the 2nd pass, and even in + * an abort_run case, we'll collect all of the + * free pages in this run and return them to the free list + */ + while (start_idx < page_idx) { + + m1 = &vm_pages[start_idx++]; + +#if !VM_PAGE_FIND_CONTIGUOUS_CAN_STEAL + assert(m1->free); +#endif + + if (m1->free) { + unsigned int color; + + color = m1->phys_page & vm_color_mask; +#if MACH_ASSERT + vm_page_verify_free_list(&vm_page_queue_free[color], color, m1, TRUE); +#endif + queue_remove(&vm_page_queue_free[color], + m1, + vm_page_t, + pageq); + m1->pageq.next = NULL; + m1->pageq.prev = NULL; +#if MACH_ASSERT + vm_page_verify_free_list(&vm_page_queue_free[color], color, VM_PAGE_NULL, FALSE); +#endif + /* + * Clear the "free" bit so that this page + * does not get considered for another + * concurrent physically-contiguous allocation. + */ + m1->free = FALSE; + assert(m1->busy); + + vm_page_free_count--; + } + } + /* + * adjust global freelist counts + */ + if (vm_page_free_count < vm_page_free_count_minimum) + vm_page_free_count_minimum = vm_page_free_count; + + if( flags & KMA_LOMEM) + vm_page_lomem_find_contiguous_last_idx = page_idx; + else + vm_page_find_contiguous_last_idx = page_idx; + + /* + * we can drop the free queue lock at this point since + * we've pulled any 'free' candidates off of the list + * we need it dropped so that we can do a vm_page_grab + * when substituing for pmapped/dirty pages + */ + lck_mtx_unlock(&vm_page_queue_free_lock); + + start_idx = tmp_start_idx; + cur_idx = page_idx - 1; + + while (start_idx++ < page_idx) { + /* + * must go through the list from back to front + * so that the page list is created in the + * correct order - low -> high phys addresses + */ + m1 = &vm_pages[cur_idx--]; + + assert(!m1->free); + + if (m1->object == VM_OBJECT_NULL) { + /* + * page has already been removed from + * the free list in the 1st pass + */ + assert(m1->offset == (vm_object_offset_t) -1); + assert(m1->busy); + assert(!m1->wanted); + assert(!m1->laundry); + } else { + vm_object_t object; + int refmod; + boolean_t disconnected, reusable; + + if (abort_run == TRUE) + continue; + + object = m1->object; + + if (object != locked_object) { + if (locked_object) { + vm_object_unlock(locked_object); + locked_object = VM_OBJECT_NULL; + } + if (vm_object_lock_try(object)) + locked_object = object; + } + if (locked_object == VM_OBJECT_NULL || + (VM_PAGE_WIRED(m1) || m1->gobbled || + m1->encrypted_cleaning || + m1->pageout_queue || m1->laundry || m1->wanted || + m1->cleaning || m1->overwriting || m1->pageout || m1->busy)) { + + if (locked_object) { + vm_object_unlock(locked_object); + locked_object = VM_OBJECT_NULL; + } + tmp_start_idx = cur_idx; + abort_run = TRUE; + continue; + } + + disconnected = FALSE; + reusable = FALSE; + + if ((m1->reusable || + m1->object->all_reusable) && + m1->inactive && + !m1->dirty && + !m1->reference) { + /* reusable page... */ + refmod = pmap_disconnect(m1->phys_page); + disconnected = TRUE; + if (refmod == 0) { + /* + * ... not reused: can steal + * without relocating contents. + */ + reusable = TRUE; + } + } + + if ((m1->pmapped && + ! reusable) || + m1->dirty || + m1->precious) { + vm_object_offset_t offset; + + m2 = vm_page_grab(); + + if (m2 == VM_PAGE_NULL) { + if (locked_object) { + vm_object_unlock(locked_object); + locked_object = VM_OBJECT_NULL; + } + tmp_start_idx = cur_idx; + abort_run = TRUE; + continue; + } + if (! disconnected) { + if (m1->pmapped) + refmod = pmap_disconnect(m1->phys_page); + else + refmod = 0; + } + + /* copy the page's contents */ + pmap_copy_page(m1->phys_page, m2->phys_page); + /* copy the page's state */ + assert(!VM_PAGE_WIRED(m1)); + assert(!m1->free); + assert(!m1->pageout_queue); + assert(!m1->laundry); + m2->reference = m1->reference; + assert(!m1->gobbled); + assert(!m1->private); + m2->no_cache = m1->no_cache; + m2->xpmapped = m1->xpmapped; + assert(!m1->busy); + assert(!m1->wanted); + assert(!m1->fictitious); + m2->pmapped = m1->pmapped; /* should flush cache ? */ + m2->wpmapped = m1->wpmapped; + assert(!m1->pageout); + m2->absent = m1->absent; + m2->error = m1->error; + m2->dirty = m1->dirty; + assert(!m1->cleaning); + m2->precious = m1->precious; + m2->clustered = m1->clustered; + assert(!m1->overwriting); + m2->restart = m1->restart; + m2->unusual = m1->unusual; + m2->encrypted = m1->encrypted; + assert(!m1->encrypted_cleaning); + m2->cs_validated = m1->cs_validated; + m2->cs_tainted = m1->cs_tainted; + + /* + * If m1 had really been reusable, + * we would have just stolen it, so + * let's not propagate it's "reusable" + * bit and assert that m2 is not + * marked as "reusable". + */ + // m2->reusable = m1->reusable; + assert(!m2->reusable); + + assert(!m1->lopage); + m2->slid = m1->slid; + m2->was_dirty = m1->was_dirty; + m2->compressor = m1->compressor; + + /* + * page may need to be flushed if + * it is marshalled into a UPL + * that is going to be used by a device + * that doesn't support coherency + */ + m2->written_by_kernel = TRUE; + + /* + * make sure we clear the ref/mod state + * from the pmap layer... else we risk + * inheriting state from the last time + * this page was used... + */ + pmap_clear_refmod(m2->phys_page, VM_MEM_MODIFIED | VM_MEM_REFERENCED); + + if (refmod & VM_MEM_REFERENCED) + m2->reference = TRUE; + if (refmod & VM_MEM_MODIFIED) { + SET_PAGE_DIRTY(m2, TRUE); + } + offset = m1->offset; + + /* + * completely cleans up the state + * of the page so that it is ready + * to be put onto the free list, or + * for this purpose it looks like it + * just came off of the free list + */ + vm_page_free_prepare(m1); + + /* + * now put the substitute page + * on the object + */ + vm_page_insert_internal(m2, locked_object, offset, TRUE, TRUE, FALSE); + + if (m2->compressor) { + m2->pmapped = TRUE; + m2->wpmapped = TRUE; + + PMAP_ENTER(kernel_pmap, m2->offset, m2, + VM_PROT_READ | VM_PROT_WRITE, VM_PROT_NONE, 0, TRUE); +#if MACH_ASSERT + compressed_pages++; +#endif + } else { + if (m2->reference) + vm_page_activate(m2); + else + vm_page_deactivate(m2); + } + PAGE_WAKEUP_DONE(m2); + + } else { + assert(!m1->compressor); + + /* + * completely cleans up the state + * of the page so that it is ready + * to be put onto the free list, or + * for this purpose it looks like it + * just came off of the free list + */ + vm_page_free_prepare(m1); + } +#if MACH_ASSERT + stolen_pages++; +#endif + } + m1->pageq.next = (queue_entry_t) m; + m1->pageq.prev = NULL; + m = m1; + } + if (locked_object) { + vm_object_unlock(locked_object); + locked_object = VM_OBJECT_NULL; + } + + if (abort_run == TRUE) { + if (m != VM_PAGE_NULL) { + vm_page_free_list(m, FALSE); + } +#if MACH_ASSERT + dumped_run++; +#endif + /* + * want the index of the last + * page in this run that was + * successfully 'stolen', so back + * it up 1 for the auto-decrement on use + * and 1 more to bump back over this page + */ + page_idx = tmp_start_idx + 2; + if (page_idx >= vm_pages_count) { + if (wrapped) + goto done_scanning; + page_idx = last_idx = 0; + wrapped = TRUE; + } + abort_run = FALSE; + + /* + * We didn't find a contiguous range but we didn't + * start from the very first page. + * Start again from the very first page. + */ + RESET_STATE_OF_RUN(); + + if( flags & KMA_LOMEM) + idx_last_contig_page_found = vm_page_lomem_find_contiguous_last_idx = page_idx; + else + idx_last_contig_page_found = vm_page_find_contiguous_last_idx = page_idx; + + last_idx = page_idx; + + lck_mtx_lock(&vm_page_queue_free_lock); + /* + * reset our free page limit since we + * dropped the lock protecting the vm_page_free_queue + */ + free_available = vm_page_free_count - vm_page_free_reserved; + goto retry; + } + + for (m1 = m; m1 != VM_PAGE_NULL; m1 = NEXT_PAGE(m1)) { + + if (wire == TRUE) + m1->wire_count++; + else + m1->gobbled = TRUE; + } + if (wire == FALSE) + vm_page_gobble_count += npages; + + /* + * gobbled pages are also counted as wired pages + */ + vm_page_wire_count += npages; + + assert(vm_page_verify_contiguous(m, npages)); + } +done_scanning: + PAGE_REPLACEMENT_ALLOWED(FALSE); + + vm_page_unlock_queues(); + +#if DEBUG + clock_get_system_microtime(&tv_end_sec, &tv_end_usec); + + tv_end_sec -= tv_start_sec; + if (tv_end_usec < tv_start_usec) { + tv_end_sec--; + tv_end_usec += 1000000; + } + tv_end_usec -= tv_start_usec; + if (tv_end_usec >= 1000000) { + tv_end_sec++; + tv_end_sec -= 1000000; + } + if (vm_page_find_contig_debug) { + printf("%s(num=%d,low=%d): found %d pages at 0x%llx in %ld.%06ds... started at %d... scanned %d pages... yielded %d times... dumped run %d times... stole %d pages... stole %d compressed pages\n", + __func__, contig_pages, max_pnum, npages, (vm_object_offset_t)start_pnum << PAGE_SHIFT, + (long)tv_end_sec, tv_end_usec, orig_last_idx, + scanned, yielded, dumped_run, stolen_pages, compressed_pages); + } + +#endif +#if MACH_ASSERT + vm_page_verify_free_lists(); +#endif + return m; +} + +/* + * Allocate a list of contiguous, wired pages. + */ +kern_return_t +cpm_allocate( + vm_size_t size, + vm_page_t *list, + ppnum_t max_pnum, + ppnum_t pnum_mask, + boolean_t wire, + int flags) +{ + vm_page_t pages; + unsigned int npages; + + if (size % PAGE_SIZE != 0) + return KERN_INVALID_ARGUMENT; + + npages = (unsigned int) (size / PAGE_SIZE); + if (npages != size / PAGE_SIZE) { + /* 32-bit overflow */ + return KERN_INVALID_ARGUMENT; + } + + /* + * Obtain a pointer to a subset of the free + * list large enough to satisfy the request; + * the region will be physically contiguous. + */ + pages = vm_page_find_contiguous(npages, max_pnum, pnum_mask, wire, flags); + + if (pages == VM_PAGE_NULL) + return KERN_NO_SPACE; + /* + * determine need for wakeups + */ + if ((vm_page_free_count < vm_page_free_min) || + ((vm_page_free_count < vm_page_free_target) && + ((vm_page_inactive_count + vm_page_speculative_count) < vm_page_inactive_min))) + thread_wakeup((event_t) &vm_page_free_wanted); + + VM_CHECK_MEMORYSTATUS; + + /* + * The CPM pages should now be available and + * ordered by ascending physical address. + */ + assert(vm_page_verify_contiguous(pages, npages)); + + *list = pages; + return KERN_SUCCESS; +} + + +unsigned int vm_max_delayed_work_limit = DEFAULT_DELAYED_WORK_LIMIT; + +/* + * when working on a 'run' of pages, it is necessary to hold + * the vm_page_queue_lock (a hot global lock) for certain operations + * on the page... however, the majority of the work can be done + * while merely holding the object lock... in fact there are certain + * collections of pages that don't require any work brokered by the + * vm_page_queue_lock... to mitigate the time spent behind the global + * lock, go to a 2 pass algorithm... collect pages up to DELAYED_WORK_LIMIT + * while doing all of the work that doesn't require the vm_page_queue_lock... + * then call vm_page_do_delayed_work to acquire the vm_page_queue_lock and do the + * necessary work for each page... we will grab the busy bit on the page + * if it's not already held so that vm_page_do_delayed_work can drop the object lock + * if it can't immediately take the vm_page_queue_lock in order to compete + * for the locks in the same order that vm_pageout_scan takes them. + * the operation names are modeled after the names of the routines that + * need to be called in order to make the changes very obvious in the + * original loop + */ + +void +vm_page_do_delayed_work( + vm_object_t object, + struct vm_page_delayed_work *dwp, + int dw_count) +{ + int j; + vm_page_t m; + vm_page_t local_free_q = VM_PAGE_NULL; + + /* + * pageout_scan takes the vm_page_lock_queues first + * then tries for the object lock... to avoid what + * is effectively a lock inversion, we'll go to the + * trouble of taking them in that same order... otherwise + * if this object contains the majority of the pages resident + * in the UBC (or a small set of large objects actively being + * worked on contain the majority of the pages), we could + * cause the pageout_scan thread to 'starve' in its attempt + * to find pages to move to the free queue, since it has to + * successfully acquire the object lock of any candidate page + * before it can steal/clean it. + */ + if (!vm_page_trylockspin_queues()) { + vm_object_unlock(object); + + vm_page_lockspin_queues(); + + for (j = 0; ; j++) { + if (!vm_object_lock_avoid(object) && + _vm_object_lock_try(object)) + break; + vm_page_unlock_queues(); + mutex_pause(j); + vm_page_lockspin_queues(); + } + } + for (j = 0; j < dw_count; j++, dwp++) { + + m = dwp->dw_m; + + if (dwp->dw_mask & DW_vm_pageout_throttle_up) + vm_pageout_throttle_up(m); + + if (dwp->dw_mask & DW_vm_page_wire) + vm_page_wire(m); + else if (dwp->dw_mask & DW_vm_page_unwire) { + boolean_t queueit; + + queueit = (dwp->dw_mask & DW_vm_page_free) ? FALSE : TRUE; + + vm_page_unwire(m, queueit); + } + if (dwp->dw_mask & DW_vm_page_free) { + vm_page_free_prepare_queues(m); + + assert(m->pageq.next == NULL && m->pageq.prev == NULL); + /* + * Add this page to our list of reclaimed pages, + * to be freed later. + */ + m->pageq.next = (queue_entry_t) local_free_q; + local_free_q = m; + } else { + if (dwp->dw_mask & DW_vm_page_deactivate_internal) + vm_page_deactivate_internal(m, FALSE); + else if (dwp->dw_mask & DW_vm_page_activate) { + if (m->active == FALSE) { + vm_page_activate(m); + } + } + else if (dwp->dw_mask & DW_vm_page_speculate) + vm_page_speculate(m, TRUE); + else if (dwp->dw_mask & DW_enqueue_cleaned) { + /* + * if we didn't hold the object lock and did this, + * we might disconnect the page, then someone might + * soft fault it back in, then we would put it on the + * cleaned queue, and so we would have a referenced (maybe even dirty) + * page on that queue, which we don't want + */ + int refmod_state = pmap_disconnect(m->phys_page); + + if ((refmod_state & VM_MEM_REFERENCED)) { + /* + * this page has been touched since it got cleaned; let's activate it + * if it hasn't already been + */ + vm_pageout_enqueued_cleaned++; + vm_pageout_cleaned_reactivated++; + vm_pageout_cleaned_commit_reactivated++; + + if (m->active == FALSE) + vm_page_activate(m); + } else { + m->reference = FALSE; + vm_page_enqueue_cleaned(m); + } + } + else if (dwp->dw_mask & DW_vm_page_lru) + vm_page_lru(m); + else if (dwp->dw_mask & DW_VM_PAGE_QUEUES_REMOVE) { + if ( !m->pageout_queue) + VM_PAGE_QUEUES_REMOVE(m); + } + if (dwp->dw_mask & DW_set_reference) + m->reference = TRUE; + else if (dwp->dw_mask & DW_clear_reference) + m->reference = FALSE; + + if (dwp->dw_mask & DW_move_page) { + if ( !m->pageout_queue) { + VM_PAGE_QUEUES_REMOVE(m); + + assert(m->object != kernel_object); + + VM_PAGE_ENQUEUE_INACTIVE(m, FALSE); + } + } + if (dwp->dw_mask & DW_clear_busy) + m->busy = FALSE; + + if (dwp->dw_mask & DW_PAGE_WAKEUP) + PAGE_WAKEUP(m); + } + } + vm_page_unlock_queues(); + + if (local_free_q) + vm_page_free_list(local_free_q, TRUE); + + VM_CHECK_MEMORYSTATUS; + +} + +kern_return_t +vm_page_alloc_list( + int page_count, + int flags, + vm_page_t *list) +{ + vm_page_t lo_page_list = VM_PAGE_NULL; + vm_page_t mem; + int i; + + if ( !(flags & KMA_LOMEM)) + panic("vm_page_alloc_list: called w/o KMA_LOMEM"); + + for (i = 0; i < page_count; i++) { + + mem = vm_page_grablo(); + + if (mem == VM_PAGE_NULL) { + if (lo_page_list) + vm_page_free_list(lo_page_list, FALSE); + + *list = VM_PAGE_NULL; + + return (KERN_RESOURCE_SHORTAGE); + } + mem->pageq.next = (queue_entry_t) lo_page_list; + lo_page_list = mem; + } + *list = lo_page_list; + + return (KERN_SUCCESS); +} + +void +vm_page_set_offset(vm_page_t page, vm_object_offset_t offset) +{ + page->offset = offset; +} + +vm_page_t +vm_page_get_next(vm_page_t page) +{ + return ((vm_page_t) page->pageq.next); +} + +vm_object_offset_t +vm_page_get_offset(vm_page_t page) +{ + return (page->offset); +} + +ppnum_t +vm_page_get_phys_page(vm_page_t page) +{ + return (page->phys_page); +} + + +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +#if HIBERNATION + +static vm_page_t hibernate_gobble_queue; + +extern boolean_t (* volatile consider_buffer_cache_collect)(int); + +static int hibernate_drain_pageout_queue(struct vm_pageout_queue *); +static int hibernate_flush_dirty_pages(int); +static int hibernate_flush_queue(queue_head_t *, int); + +void hibernate_flush_wait(void); +void hibernate_mark_in_progress(void); +void hibernate_clear_in_progress(void); + +void hibernate_free_range(int, int); +void hibernate_hash_insert_page(vm_page_t); +uint32_t hibernate_mark_as_unneeded(addr64_t, addr64_t, hibernate_page_list_t *, hibernate_page_list_t *); +void hibernate_rebuild_vm_structs(void); +uint32_t hibernate_teardown_vm_structs(hibernate_page_list_t *, hibernate_page_list_t *); +ppnum_t hibernate_lookup_paddr(unsigned int); + +struct hibernate_statistics { + int hibernate_considered; + int hibernate_reentered_on_q; + int hibernate_found_dirty; + int hibernate_skipped_cleaning; + int hibernate_skipped_transient; + int hibernate_skipped_precious; + int hibernate_skipped_external; + int hibernate_queue_nolock; + int hibernate_queue_paused; + int hibernate_throttled; + int hibernate_throttle_timeout; + int hibernate_drained; + int hibernate_drain_timeout; + int cd_lock_failed; + int cd_found_precious; + int cd_found_wired; + int cd_found_busy; + int cd_found_unusual; + int cd_found_cleaning; + int cd_found_laundry; + int cd_found_dirty; + int cd_found_xpmapped; + int cd_local_free; + int cd_total_free; + int cd_vm_page_wire_count; + int cd_vm_struct_pages_unneeded; + int cd_pages; + int cd_discarded; + int cd_count_wire; +} hibernate_stats; + + + +static int +hibernate_drain_pageout_queue(struct vm_pageout_queue *q) +{ + wait_result_t wait_result; + + vm_page_lock_queues(); + + while ( !queue_empty(&q->pgo_pending) ) { + + q->pgo_draining = TRUE; + + assert_wait_timeout((event_t) (&q->pgo_laundry+1), THREAD_INTERRUPTIBLE, 5000, 1000*NSEC_PER_USEC); + + vm_page_unlock_queues(); + + wait_result = thread_block(THREAD_CONTINUE_NULL); + + if (wait_result == THREAD_TIMED_OUT && !queue_empty(&q->pgo_pending)) { + hibernate_stats.hibernate_drain_timeout++; + + if (q == &vm_pageout_queue_external) + return (0); + + return (1); + } + vm_page_lock_queues(); + + hibernate_stats.hibernate_drained++; + } + vm_page_unlock_queues(); + + return (0); +} + + +boolean_t hibernate_skip_external = FALSE; + +static int +hibernate_flush_queue(queue_head_t *q, int qcount) +{ + vm_page_t m; + vm_object_t l_object = NULL; + vm_object_t m_object = NULL; + int refmod_state = 0; + int try_failed_count = 0; + int retval = 0; + int current_run = 0; + struct vm_pageout_queue *iq; + struct vm_pageout_queue *eq; + struct vm_pageout_queue *tq; + + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 4) | DBG_FUNC_START, q, qcount, 0, 0, 0); + + iq = &vm_pageout_queue_internal; + eq = &vm_pageout_queue_external; + + vm_page_lock_queues(); + + while (qcount && !queue_empty(q)) { + + if (current_run++ == 1000) { + if (hibernate_should_abort()) { + retval = 1; + break; + } + current_run = 0; + } + + m = (vm_page_t) queue_first(q); + m_object = m->object; + + /* + * check to see if we currently are working + * with the same object... if so, we've + * already got the lock + */ + if (m_object != l_object) { + /* + * the object associated with candidate page is + * different from the one we were just working + * with... dump the lock if we still own it + */ + if (l_object != NULL) { + vm_object_unlock(l_object); + l_object = NULL; + } + /* + * Try to lock object; since we've alread got the + * page queues lock, we can only 'try' for this one. + * if the 'try' fails, we need to do a mutex_pause + * to allow the owner of the object lock a chance to + * run... + */ + if ( !vm_object_lock_try_scan(m_object)) { + + if (try_failed_count > 20) { + hibernate_stats.hibernate_queue_nolock++; + + goto reenter_pg_on_q; + } + vm_pageout_scan_wants_object = m_object; + + vm_page_unlock_queues(); + mutex_pause(try_failed_count++); + vm_page_lock_queues(); + + hibernate_stats.hibernate_queue_paused++; + continue; + } else { + l_object = m_object; + vm_pageout_scan_wants_object = VM_OBJECT_NULL; + } + } + if ( !m_object->alive || m->encrypted_cleaning || m->cleaning || m->laundry || m->busy || m->absent || m->error) { + /* + * page is not to be cleaned + * put it back on the head of its queue + */ + if (m->cleaning) + hibernate_stats.hibernate_skipped_cleaning++; + else + hibernate_stats.hibernate_skipped_transient++; + + goto reenter_pg_on_q; + } + if (m_object->copy == VM_OBJECT_NULL) { + if (m_object->purgable == VM_PURGABLE_VOLATILE || m_object->purgable == VM_PURGABLE_EMPTY) { + /* + * let the normal hibernate image path + * deal with these + */ + goto reenter_pg_on_q; + } + } + if ( !m->dirty && m->pmapped) { + refmod_state = pmap_get_refmod(m->phys_page); + + if ((refmod_state & VM_MEM_MODIFIED)) { + SET_PAGE_DIRTY(m, FALSE); + } + } else + refmod_state = 0; + + if ( !m->dirty) { + /* + * page is not to be cleaned + * put it back on the head of its queue + */ + if (m->precious) + hibernate_stats.hibernate_skipped_precious++; + + goto reenter_pg_on_q; + } + + if (hibernate_skip_external == TRUE && !m_object->internal) { + + hibernate_stats.hibernate_skipped_external++; + + goto reenter_pg_on_q; + } + tq = NULL; + + if (m_object->internal) { + if (VM_PAGE_Q_THROTTLED(iq)) + tq = iq; + } else if (VM_PAGE_Q_THROTTLED(eq)) + tq = eq; + + if (tq != NULL) { + wait_result_t wait_result; + int wait_count = 5; + + if (l_object != NULL) { + vm_object_unlock(l_object); + l_object = NULL; + } + vm_pageout_scan_wants_object = VM_OBJECT_NULL; + + while (retval == 0) { + + tq->pgo_throttled = TRUE; + + assert_wait_timeout((event_t) &tq->pgo_laundry, THREAD_INTERRUPTIBLE, 1000, 1000*NSEC_PER_USEC); + + vm_page_unlock_queues(); + + wait_result = thread_block(THREAD_CONTINUE_NULL); + + vm_page_lock_queues(); + + if (wait_result != THREAD_TIMED_OUT) + break; + if (!VM_PAGE_Q_THROTTLED(tq)) + break; + + if (hibernate_should_abort()) + retval = 1; + + if (--wait_count == 0) { + + hibernate_stats.hibernate_throttle_timeout++; + + if (tq == eq) { + hibernate_skip_external = TRUE; + break; + } + retval = 1; + } + } + if (retval) + break; + + hibernate_stats.hibernate_throttled++; + + continue; + } + /* + * we've already factored out pages in the laundry which + * means this page can't be on the pageout queue so it's + * safe to do the VM_PAGE_QUEUES_REMOVE + */ + assert(!m->pageout_queue); + + VM_PAGE_QUEUES_REMOVE(m); + + if (COMPRESSED_PAGER_IS_ACTIVE) + pmap_disconnect(m->phys_page); + + vm_pageout_cluster(m, FALSE); + + hibernate_stats.hibernate_found_dirty++; + + goto next_pg; + +reenter_pg_on_q: + queue_remove(q, m, vm_page_t, pageq); + queue_enter(q, m, vm_page_t, pageq); + + hibernate_stats.hibernate_reentered_on_q++; +next_pg: + hibernate_stats.hibernate_considered++; + + qcount--; + try_failed_count = 0; + } + if (l_object != NULL) { + vm_object_unlock(l_object); + l_object = NULL; + } + vm_pageout_scan_wants_object = VM_OBJECT_NULL; + + vm_page_unlock_queues(); + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 4) | DBG_FUNC_END, hibernate_stats.hibernate_found_dirty, retval, 0, 0, 0); + + return (retval); +} + + +static int +hibernate_flush_dirty_pages(int pass) +{ + struct vm_speculative_age_q *aq; + uint32_t i; + + bzero(&hibernate_stats, sizeof(struct hibernate_statistics)); + + if (vm_page_local_q) { + for (i = 0; i < vm_page_local_q_count; i++) + vm_page_reactivate_local(i, TRUE, FALSE); + } + + for (i = 0; i <= VM_PAGE_MAX_SPECULATIVE_AGE_Q; i++) { + int qcount; + vm_page_t m; + + aq = &vm_page_queue_speculative[i]; + + if (queue_empty(&aq->age_q)) + continue; + qcount = 0; + + vm_page_lockspin_queues(); + + queue_iterate(&aq->age_q, + m, + vm_page_t, + pageq) + { + qcount++; + } + vm_page_unlock_queues(); + + if (qcount) { + if (hibernate_flush_queue(&aq->age_q, qcount)) + return (1); + } + } + if (hibernate_flush_queue(&vm_page_queue_inactive, vm_page_inactive_count - vm_page_anonymous_count - vm_page_cleaned_count)) + return (1); + if (hibernate_flush_queue(&vm_page_queue_anonymous, vm_page_anonymous_count)) + return (1); + if (hibernate_flush_queue(&vm_page_queue_cleaned, vm_page_cleaned_count)) + return (1); + if (hibernate_drain_pageout_queue(&vm_pageout_queue_internal)) + return (1); + + if (COMPRESSED_PAGER_IS_ACTIVE && pass == 1) + vm_compressor_record_warmup_start(); + + if (hibernate_flush_queue(&vm_page_queue_active, vm_page_active_count)) { + if (COMPRESSED_PAGER_IS_ACTIVE && pass == 1) + vm_compressor_record_warmup_end(); + return (1); + } + if (hibernate_drain_pageout_queue(&vm_pageout_queue_internal)) { + if (COMPRESSED_PAGER_IS_ACTIVE && pass == 1) + vm_compressor_record_warmup_end(); + return (1); + } + if (COMPRESSED_PAGER_IS_ACTIVE && pass == 1) + vm_compressor_record_warmup_end(); + + if (hibernate_skip_external == FALSE && hibernate_drain_pageout_queue(&vm_pageout_queue_external)) + return (1); + + return (0); +} + + +int +hibernate_flush_memory() +{ + int retval; + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 3) | DBG_FUNC_START, vm_page_free_count, 0, 0, 0, 0); + + hibernate_cleaning_in_progress = TRUE; + hibernate_skip_external = FALSE; + + if ((retval = hibernate_flush_dirty_pages(1)) == 0) { + + if (COMPRESSED_PAGER_IS_ACTIVE) { + + if ((retval = hibernate_flush_dirty_pages(2)) == 0) { + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 10) | DBG_FUNC_START, VM_PAGE_COMPRESSOR_COUNT, 0, 0, 0, 0); + + vm_compressor_flush(); + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 10) | DBG_FUNC_END, VM_PAGE_COMPRESSOR_COUNT, 0, 0, 0, 0); + } + } + if (retval == 0 && consider_buffer_cache_collect != NULL) { + unsigned int orig_wire_count; + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 7) | DBG_FUNC_START, 0, 0, 0, 0, 0); + orig_wire_count = vm_page_wire_count; + + (void)(*consider_buffer_cache_collect)(1); + consider_zone_gc(TRUE); + + HIBLOG("hibernate_flush_memory: buffer_cache_gc freed up %d wired pages\n", orig_wire_count - vm_page_wire_count); + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 7) | DBG_FUNC_END, orig_wire_count - vm_page_wire_count, 0, 0, 0, 0); + } + } + hibernate_cleaning_in_progress = FALSE; + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 3) | DBG_FUNC_END, vm_page_free_count, hibernate_stats.hibernate_found_dirty, retval, 0, 0); + + if (retval && COMPRESSED_PAGER_IS_ACTIVE) + HIBLOG("hibernate_flush_memory() failed to finish - vm_page_compressor_count(%d)\n", VM_PAGE_COMPRESSOR_COUNT); + + + HIBPRINT("hibernate_flush_memory() considered(%d) reentered_on_q(%d) found_dirty(%d)\n", + hibernate_stats.hibernate_considered, + hibernate_stats.hibernate_reentered_on_q, + hibernate_stats.hibernate_found_dirty); + HIBPRINT(" skipped_cleaning(%d) skipped_transient(%d) skipped_precious(%d) skipped_external(%d) queue_nolock(%d)\n", + hibernate_stats.hibernate_skipped_cleaning, + hibernate_stats.hibernate_skipped_transient, + hibernate_stats.hibernate_skipped_precious, + hibernate_stats.hibernate_skipped_external, + hibernate_stats.hibernate_queue_nolock); + HIBPRINT(" queue_paused(%d) throttled(%d) throttle_timeout(%d) drained(%d) drain_timeout(%d)\n", + hibernate_stats.hibernate_queue_paused, + hibernate_stats.hibernate_throttled, + hibernate_stats.hibernate_throttle_timeout, + hibernate_stats.hibernate_drained, + hibernate_stats.hibernate_drain_timeout); + + return (retval); +} + + +static void +hibernate_page_list_zero(hibernate_page_list_t *list) +{ + uint32_t bank; + hibernate_bitmap_t * bitmap; + + bitmap = &list->bank_bitmap[0]; + for (bank = 0; bank < list->bank_count; bank++) + { + uint32_t last_bit; + + bzero((void *) &bitmap->bitmap[0], bitmap->bitmapwords << 2); + // set out-of-bound bits at end of bitmap. + last_bit = ((bitmap->last_page - bitmap->first_page + 1) & 31); + if (last_bit) + bitmap->bitmap[bitmap->bitmapwords - 1] = (0xFFFFFFFF >> last_bit); + + bitmap = (hibernate_bitmap_t *) &bitmap->bitmap[bitmap->bitmapwords]; + } +} + +void +hibernate_gobble_pages(uint32_t gobble_count, uint32_t free_page_time) +{ + uint32_t i; + vm_page_t m; + uint64_t start, end, timeout, nsec; + clock_interval_to_deadline(free_page_time, 1000 * 1000 /*ms*/, &timeout); + clock_get_uptime(&start); + + for (i = 0; i < gobble_count; i++) + { + while (VM_PAGE_NULL == (m = vm_page_grab())) + { + clock_get_uptime(&end); + if (end >= timeout) + break; + VM_PAGE_WAIT(); + } + if (!m) + break; + m->busy = FALSE; + vm_page_gobble(m); + + m->pageq.next = (queue_entry_t) hibernate_gobble_queue; + hibernate_gobble_queue = m; + } + + clock_get_uptime(&end); + absolutetime_to_nanoseconds(end - start, &nsec); + HIBLOG("Gobbled %d pages, time: %qd ms\n", i, nsec / 1000000ULL); +} + +void +hibernate_free_gobble_pages(void) +{ + vm_page_t m, next; + uint32_t count = 0; + + m = (vm_page_t) hibernate_gobble_queue; + while(m) + { + next = (vm_page_t) m->pageq.next; + vm_page_free(m); + count++; + m = next; + } + hibernate_gobble_queue = VM_PAGE_NULL; + + if (count) + HIBLOG("Freed %d pages\n", count); +} + +static boolean_t +hibernate_consider_discard(vm_page_t m, boolean_t preflight) +{ + vm_object_t object = NULL; + int refmod_state; + boolean_t discard = FALSE; + + do + { + if (m->private) + panic("hibernate_consider_discard: private"); + + if (!vm_object_lock_try(m->object)) { + if (!preflight) hibernate_stats.cd_lock_failed++; + break; + } + object = m->object; + + if (VM_PAGE_WIRED(m)) { + if (!preflight) hibernate_stats.cd_found_wired++; + break; + } + if (m->precious) { + if (!preflight) hibernate_stats.cd_found_precious++; + break; + } + if (m->busy || !object->alive) { + /* + * Somebody is playing with this page. + */ + if (!preflight) hibernate_stats.cd_found_busy++; + break; + } + if (m->absent || m->unusual || m->error) { + /* + * If it's unusual in anyway, ignore it + */ + if (!preflight) hibernate_stats.cd_found_unusual++; + break; + } + if (m->cleaning) { + if (!preflight) hibernate_stats.cd_found_cleaning++; + break; + } + if (m->laundry) { + if (!preflight) hibernate_stats.cd_found_laundry++; + break; + } + if (!m->dirty) + { + refmod_state = pmap_get_refmod(m->phys_page); + + if (refmod_state & VM_MEM_REFERENCED) + m->reference = TRUE; + if (refmod_state & VM_MEM_MODIFIED) { + SET_PAGE_DIRTY(m, FALSE); + } + } + + /* + * If it's clean or purgeable we can discard the page on wakeup. + */ + discard = (!m->dirty) + || (VM_PURGABLE_VOLATILE == object->purgable) + || (VM_PURGABLE_EMPTY == object->purgable); + + + if (discard == FALSE) { + if (!preflight) + hibernate_stats.cd_found_dirty++; + } else if (m->xpmapped && m->reference) { + if (!preflight) + hibernate_stats.cd_found_xpmapped++; + discard = FALSE; + } + } + while (FALSE); + + if (object) + vm_object_unlock(object); + + return (discard); +} + + +static void +hibernate_discard_page(vm_page_t m) +{ + if (m->absent || m->unusual || m->error) + /* + * If it's unusual in anyway, ignore + */ + return; + +#if DEBUG + vm_object_t object = m->object; + if (!vm_object_lock_try(m->object)) + panic("hibernate_discard_page(%p) !vm_object_lock_try", m); +#else + /* No need to lock page queue for token delete, hibernate_vm_unlock() + makes sure these locks are uncontended before sleep */ +#endif /* !DEBUG */ + + if (m->pmapped == TRUE) + { + __unused int refmod_state = pmap_disconnect(m->phys_page); + } + + if (m->laundry) + panic("hibernate_discard_page(%p) laundry", m); + if (m->private) + panic("hibernate_discard_page(%p) private", m); + if (m->fictitious) + panic("hibernate_discard_page(%p) fictitious", m); + + if (VM_PURGABLE_VOLATILE == m->object->purgable) + { + /* object should be on a queue */ + assert((m->object->objq.next != NULL) && (m->object->objq.prev != NULL)); + purgeable_q_t old_queue = vm_purgeable_object_remove(m->object); + assert(old_queue); + if (m->object->purgeable_when_ripe) { + vm_purgeable_token_delete_first(old_queue); + } + m->object->purgable = VM_PURGABLE_EMPTY; + } + + vm_page_free(m); + +#if DEBUG + vm_object_unlock(object); +#endif /* DEBUG */ +} + +/* + Grab locks for hibernate_page_list_setall() +*/ +void +hibernate_vm_lock_queues(void) +{ + vm_object_lock(compressor_object); + vm_page_lock_queues(); + lck_mtx_lock(&vm_page_queue_free_lock); + + if (vm_page_local_q) { + uint32_t i; + for (i = 0; i < vm_page_local_q_count; i++) { + struct vpl *lq; + lq = &vm_page_local_q[i].vpl_un.vpl; + VPL_LOCK(&lq->vpl_lock); + } + } +} + +void +hibernate_vm_unlock_queues(void) +{ + if (vm_page_local_q) { + uint32_t i; + for (i = 0; i < vm_page_local_q_count; i++) { + struct vpl *lq; + lq = &vm_page_local_q[i].vpl_un.vpl; + VPL_UNLOCK(&lq->vpl_lock); + } + } + lck_mtx_unlock(&vm_page_queue_free_lock); + vm_page_unlock_queues(); + vm_object_unlock(compressor_object); +} + +/* + Bits zero in the bitmaps => page needs to be saved. All pages default to be saved, + pages known to VM to not need saving are subtracted. + Wired pages to be saved are present in page_list_wired, pageable in page_list. +*/ + +void +hibernate_page_list_setall(hibernate_page_list_t * page_list, + hibernate_page_list_t * page_list_wired, + hibernate_page_list_t * page_list_pal, + boolean_t preflight, + boolean_t will_discard, + uint32_t * pagesOut) +{ + uint64_t start, end, nsec; + vm_page_t m; + vm_page_t next; + uint32_t pages = page_list->page_count; + uint32_t count_anonymous = 0, count_throttled = 0, count_compressor = 0; + uint32_t count_inactive = 0, count_active = 0, count_speculative = 0, count_cleaned = 0; + uint32_t count_wire = pages; + uint32_t count_discard_active = 0; + uint32_t count_discard_inactive = 0; + uint32_t count_discard_cleaned = 0; + uint32_t count_discard_purgeable = 0; + uint32_t count_discard_speculative = 0; + uint32_t count_discard_vm_struct_pages = 0; + uint32_t i; + uint32_t bank; + hibernate_bitmap_t * bitmap; + hibernate_bitmap_t * bitmap_wired; + boolean_t discard_all; + boolean_t discard; + + HIBLOG("hibernate_page_list_setall(preflight %d) start %p, %p\n", preflight, page_list, page_list_wired); + + if (preflight) { + page_list = NULL; + page_list_wired = NULL; + page_list_pal = NULL; + discard_all = FALSE; + } else { + discard_all = will_discard; + } + +#if DEBUG + if (!preflight) + { + vm_page_lock_queues(); + if (vm_page_local_q) { + for (i = 0; i < vm_page_local_q_count; i++) { + struct vpl *lq; + lq = &vm_page_local_q[i].vpl_un.vpl; + VPL_LOCK(&lq->vpl_lock); + } + } + } +#endif /* DEBUG */ + + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 8) | DBG_FUNC_START, count_wire, 0, 0, 0, 0); + + clock_get_uptime(&start); + + if (!preflight) { + hibernate_page_list_zero(page_list); + hibernate_page_list_zero(page_list_wired); + hibernate_page_list_zero(page_list_pal); + + hibernate_stats.cd_vm_page_wire_count = vm_page_wire_count; + hibernate_stats.cd_pages = pages; + } + + if (vm_page_local_q) { + for (i = 0; i < vm_page_local_q_count; i++) + vm_page_reactivate_local(i, TRUE, !preflight); + } + + if (preflight) { + vm_object_lock(compressor_object); + vm_page_lock_queues(); + lck_mtx_lock(&vm_page_queue_free_lock); + } + + m = (vm_page_t) hibernate_gobble_queue; + while (m) + { + pages--; + count_wire--; + if (!preflight) { + hibernate_page_bitset(page_list, TRUE, m->phys_page); + hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + } + m = (vm_page_t) m->pageq.next; + } + + if (!preflight) for( i = 0; i < real_ncpus; i++ ) + { + if (cpu_data_ptr[i] && cpu_data_ptr[i]->cpu_processor) + { + for (m = PROCESSOR_DATA(cpu_data_ptr[i]->cpu_processor, free_pages); m; m = (vm_page_t)m->pageq.next) + { + pages--; + count_wire--; + hibernate_page_bitset(page_list, TRUE, m->phys_page); + hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + + hibernate_stats.cd_local_free++; + hibernate_stats.cd_total_free++; + } + } + } + + for( i = 0; i < vm_colors; i++ ) + { + queue_iterate(&vm_page_queue_free[i], + m, + vm_page_t, + pageq) + { + pages--; + count_wire--; + if (!preflight) { + hibernate_page_bitset(page_list, TRUE, m->phys_page); + hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + + hibernate_stats.cd_total_free++; + } + } + } + + queue_iterate(&vm_lopage_queue_free, + m, + vm_page_t, + pageq) + { + pages--; + count_wire--; + if (!preflight) { + hibernate_page_bitset(page_list, TRUE, m->phys_page); + hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + + hibernate_stats.cd_total_free++; + } + } + + m = (vm_page_t) queue_first(&vm_page_queue_throttled); + while (m && !queue_end(&vm_page_queue_throttled, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + discard = FALSE; + if ((kIOHibernateModeDiscardCleanInactive & gIOHibernateMode) + && hibernate_consider_discard(m, preflight)) + { + if (!preflight) hibernate_page_bitset(page_list, TRUE, m->phys_page); + count_discard_inactive++; + discard = discard_all; + } + else + count_throttled++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + + if (discard) hibernate_discard_page(m); + m = next; + } + + m = (vm_page_t) queue_first(&vm_page_queue_anonymous); + while (m && !queue_end(&vm_page_queue_anonymous, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + discard = FALSE; + if ((kIOHibernateModeDiscardCleanInactive & gIOHibernateMode) + && hibernate_consider_discard(m, preflight)) + { + if (!preflight) hibernate_page_bitset(page_list, TRUE, m->phys_page); + if (m->dirty) + count_discard_purgeable++; + else + count_discard_inactive++; + discard = discard_all; + } + else + count_anonymous++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + if (discard) hibernate_discard_page(m); + m = next; + } + + m = (vm_page_t) queue_first(&vm_page_queue_inactive); + while (m && !queue_end(&vm_page_queue_inactive, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + discard = FALSE; + if ((kIOHibernateModeDiscardCleanInactive & gIOHibernateMode) + && hibernate_consider_discard(m, preflight)) + { + if (!preflight) hibernate_page_bitset(page_list, TRUE, m->phys_page); + if (m->dirty) + count_discard_purgeable++; + else + count_discard_inactive++; + discard = discard_all; + } + else + count_inactive++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + if (discard) hibernate_discard_page(m); + m = next; + } + + m = (vm_page_t) queue_first(&vm_page_queue_cleaned); + while (m && !queue_end(&vm_page_queue_cleaned, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + discard = FALSE; + if ((kIOHibernateModeDiscardCleanInactive & gIOHibernateMode) + && hibernate_consider_discard(m, preflight)) + { + if (!preflight) hibernate_page_bitset(page_list, TRUE, m->phys_page); + if (m->dirty) + count_discard_purgeable++; + else + count_discard_cleaned++; + discard = discard_all; + } + else + count_cleaned++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + if (discard) hibernate_discard_page(m); + m = next; + } + + for( i = 0; i <= VM_PAGE_MAX_SPECULATIVE_AGE_Q; i++ ) + { + m = (vm_page_t) queue_first(&vm_page_queue_speculative[i].age_q); + while (m && !queue_end(&vm_page_queue_speculative[i].age_q, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + discard = FALSE; + if ((kIOHibernateModeDiscardCleanInactive & gIOHibernateMode) + && hibernate_consider_discard(m, preflight)) + { + if (!preflight) hibernate_page_bitset(page_list, TRUE, m->phys_page); + count_discard_speculative++; + discard = discard_all; + } + else + count_speculative++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + if (discard) hibernate_discard_page(m); + m = next; + } + } + + m = (vm_page_t) queue_first(&vm_page_queue_active); + while (m && !queue_end(&vm_page_queue_active, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + discard = FALSE; + if ((kIOHibernateModeDiscardCleanActive & gIOHibernateMode) + && hibernate_consider_discard(m, preflight)) + { + if (!preflight) hibernate_page_bitset(page_list, TRUE, m->phys_page); + if (m->dirty) + count_discard_purgeable++; + else + count_discard_active++; + discard = discard_all; + } + else + count_active++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + if (discard) hibernate_discard_page(m); + m = next; + } + + queue_iterate(&compressor_object->memq, m, vm_page_t, listq) + { + count_compressor++; + count_wire--; + if (!preflight) hibernate_page_bitset(page_list_wired, TRUE, m->phys_page); + } + + if (preflight == FALSE && discard_all == TRUE) { + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 12) | DBG_FUNC_START, 0, 0, 0, 0, 0); + + HIBLOG("hibernate_teardown started\n"); + count_discard_vm_struct_pages = hibernate_teardown_vm_structs(page_list, page_list_wired); + HIBLOG("hibernate_teardown completed - discarded %d\n", count_discard_vm_struct_pages); + + pages -= count_discard_vm_struct_pages; + count_wire -= count_discard_vm_struct_pages; + + hibernate_stats.cd_vm_struct_pages_unneeded = count_discard_vm_struct_pages; + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 13) | DBG_FUNC_END, 0, 0, 0, 0, 0); + } + + if (!preflight) { + // pull wired from hibernate_bitmap + bitmap = &page_list->bank_bitmap[0]; + bitmap_wired = &page_list_wired->bank_bitmap[0]; + for (bank = 0; bank < page_list->bank_count; bank++) + { + for (i = 0; i < bitmap->bitmapwords; i++) + bitmap->bitmap[i] = bitmap->bitmap[i] | ~bitmap_wired->bitmap[i]; + bitmap = (hibernate_bitmap_t *) &bitmap->bitmap [bitmap->bitmapwords]; + bitmap_wired = (hibernate_bitmap_t *) &bitmap_wired->bitmap[bitmap_wired->bitmapwords]; + } + } + + // machine dependent adjustments + hibernate_page_list_setall_machine(page_list, page_list_wired, preflight, &pages); + + if (!preflight) { + hibernate_stats.cd_count_wire = count_wire; + hibernate_stats.cd_discarded = count_discard_active + count_discard_inactive + count_discard_purgeable + + count_discard_speculative + count_discard_cleaned + count_discard_vm_struct_pages; + } + + clock_get_uptime(&end); + absolutetime_to_nanoseconds(end - start, &nsec); + HIBLOG("hibernate_page_list_setall time: %qd ms\n", nsec / 1000000ULL); + + HIBLOG("pages %d, wire %d, act %d, inact %d, cleaned %d spec %d, zf %d, throt %d, compr %d, xpmapped %d\n %s discard act %d inact %d purgeable %d spec %d cleaned %d\n", + pages, count_wire, count_active, count_inactive, count_cleaned, count_speculative, count_anonymous, count_throttled, count_compressor, hibernate_stats.cd_found_xpmapped, + discard_all ? "did" : "could", + count_discard_active, count_discard_inactive, count_discard_purgeable, count_discard_speculative, count_discard_cleaned); + + *pagesOut = pages - count_discard_active - count_discard_inactive - count_discard_purgeable - count_discard_speculative - count_discard_cleaned; + + if (preflight && will_discard) *pagesOut -= count_compressor + count_throttled + count_anonymous + count_inactive + count_cleaned + count_speculative + count_active; + +#if DEBUG + if (!preflight) + { + if (vm_page_local_q) { + for (i = 0; i < vm_page_local_q_count; i++) { + struct vpl *lq; + lq = &vm_page_local_q[i].vpl_un.vpl; + VPL_UNLOCK(&lq->vpl_lock); + } + } + vm_page_unlock_queues(); + } +#endif /* DEBUG */ + + if (preflight) { + lck_mtx_unlock(&vm_page_queue_free_lock); + vm_page_unlock_queues(); + vm_object_unlock(compressor_object); + } + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 8) | DBG_FUNC_END, count_wire, *pagesOut, 0, 0, 0); +} + void -vm_page_copy( - vm_page_t src_m, - vm_page_t dest_m) +hibernate_page_list_discard(hibernate_page_list_t * page_list) { - XPR(XPR_VM_PAGE, - "vm_page_copy, object 0x%X offset 0x%X to object 0x%X offset 0x%X\n", - (integer_t)src_m->object, src_m->offset, - (integer_t)dest_m->object, dest_m->offset, - 0); + uint64_t start, end, nsec; + vm_page_t m; + vm_page_t next; + uint32_t i; + uint32_t count_discard_active = 0; + uint32_t count_discard_inactive = 0; + uint32_t count_discard_purgeable = 0; + uint32_t count_discard_cleaned = 0; + uint32_t count_discard_speculative = 0; - VM_PAGE_CHECK(src_m); - VM_PAGE_CHECK(dest_m); - /* - * ENCRYPTED SWAP: - * The source page should not be encrypted at this point. - * The destination page will therefore not contain encrypted - * data after the copy. - */ - if (src_m->encrypted) { - panic("vm_page_copy: source page %p is encrypted\n", src_m); +#if DEBUG + vm_page_lock_queues(); + if (vm_page_local_q) { + for (i = 0; i < vm_page_local_q_count; i++) { + struct vpl *lq; + lq = &vm_page_local_q[i].vpl_un.vpl; + VPL_LOCK(&lq->vpl_lock); + } } - dest_m->encrypted = FALSE; +#endif /* DEBUG */ + + clock_get_uptime(&start); + + m = (vm_page_t) queue_first(&vm_page_queue_anonymous); + while (m && !queue_end(&vm_page_queue_anonymous, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + if (hibernate_page_bittst(page_list, m->phys_page)) + { + if (m->dirty) + count_discard_purgeable++; + else + count_discard_inactive++; + hibernate_discard_page(m); + } + m = next; + } + + for( i = 0; i <= VM_PAGE_MAX_SPECULATIVE_AGE_Q; i++ ) + { + m = (vm_page_t) queue_first(&vm_page_queue_speculative[i].age_q); + while (m && !queue_end(&vm_page_queue_speculative[i].age_q, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + if (hibernate_page_bittst(page_list, m->phys_page)) + { + count_discard_speculative++; + hibernate_discard_page(m); + } + m = next; + } + } + + m = (vm_page_t) queue_first(&vm_page_queue_inactive); + while (m && !queue_end(&vm_page_queue_inactive, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + if (hibernate_page_bittst(page_list, m->phys_page)) + { + if (m->dirty) + count_discard_purgeable++; + else + count_discard_inactive++; + hibernate_discard_page(m); + } + m = next; + } + + m = (vm_page_t) queue_first(&vm_page_queue_active); + while (m && !queue_end(&vm_page_queue_active, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + if (hibernate_page_bittst(page_list, m->phys_page)) + { + if (m->dirty) + count_discard_purgeable++; + else + count_discard_active++; + hibernate_discard_page(m); + } + m = next; + } + + m = (vm_page_t) queue_first(&vm_page_queue_cleaned); + while (m && !queue_end(&vm_page_queue_cleaned, (queue_entry_t)m)) + { + next = (vm_page_t) m->pageq.next; + if (hibernate_page_bittst(page_list, m->phys_page)) + { + if (m->dirty) + count_discard_purgeable++; + else + count_discard_cleaned++; + hibernate_discard_page(m); + } + m = next; + } - pmap_copy_page(src_m->phys_page, dest_m->phys_page); +#if DEBUG + if (vm_page_local_q) { + for (i = 0; i < vm_page_local_q_count; i++) { + struct vpl *lq; + lq = &vm_page_local_q[i].vpl_un.vpl; + VPL_UNLOCK(&lq->vpl_lock); + } + } + vm_page_unlock_queues(); +#endif /* DEBUG */ + + clock_get_uptime(&end); + absolutetime_to_nanoseconds(end - start, &nsec); + HIBLOG("hibernate_page_list_discard time: %qd ms, discarded act %d inact %d purgeable %d spec %d cleaned %d\n", + nsec / 1000000ULL, + count_discard_active, count_discard_inactive, count_discard_purgeable, count_discard_speculative, count_discard_cleaned); } -/* - * Currently, this is a primitive allocator that grabs - * free pages from the system, sorts them by physical - * address, then searches for a region large enough to - * satisfy the user's request. - * - * Additional levels of effort: - * + steal clean active/inactive pages - * + force pageouts of dirty pages - * + maintain a map of available physical - * memory - */ +boolean_t hibernate_paddr_map_inited = FALSE; +boolean_t hibernate_rebuild_needed = FALSE; +unsigned int hibernate_teardown_last_valid_compact_indx = -1; +vm_page_t hibernate_rebuild_hash_list = NULL; -#if MACH_ASSERT -/* - * Check that the list of pages is ordered by - * ascending physical address and has no holes. - */ -int vm_page_verify_contiguous( - vm_page_t pages, - unsigned int npages); +unsigned int hibernate_teardown_found_tabled_pages = 0; +unsigned int hibernate_teardown_found_created_pages = 0; +unsigned int hibernate_teardown_found_free_pages = 0; +unsigned int hibernate_teardown_vm_page_free_count; -int -vm_page_verify_contiguous( - vm_page_t pages, - unsigned int npages) + +struct ppnum_mapping { + struct ppnum_mapping *ppnm_next; + ppnum_t ppnm_base_paddr; + unsigned int ppnm_sindx; + unsigned int ppnm_eindx; +}; + +struct ppnum_mapping *ppnm_head; +struct ppnum_mapping *ppnm_last_found = NULL; + + +void +hibernate_create_paddr_map() { - register vm_page_t m; - unsigned int page_count; - vm_offset_t prev_addr; + unsigned int i; + ppnum_t next_ppnum_in_run = 0; + struct ppnum_mapping *ppnm = NULL; - prev_addr = pages->phys_page; - page_count = 1; - for (m = NEXT_PAGE(pages); m != VM_PAGE_NULL; m = NEXT_PAGE(m)) { - if (m->phys_page != prev_addr + 1) { - printf("m 0x%x prev_addr 0x%x, current addr 0x%x\n", - m, prev_addr, m->phys_page); - printf("pages 0x%x page_count %d\n", pages, page_count); - panic("vm_page_verify_contiguous: not contiguous!"); + if (hibernate_paddr_map_inited == FALSE) { + + for (i = 0; i < vm_pages_count; i++) { + + if (ppnm) + ppnm->ppnm_eindx = i; + + if (ppnm == NULL || vm_pages[i].phys_page != next_ppnum_in_run) { + + ppnm = kalloc(sizeof(struct ppnum_mapping)); + + ppnm->ppnm_next = ppnm_head; + ppnm_head = ppnm; + + ppnm->ppnm_sindx = i; + ppnm->ppnm_base_paddr = vm_pages[i].phys_page; + } + next_ppnum_in_run = vm_pages[i].phys_page + 1; } - prev_addr = m->phys_page; - ++page_count; + ppnm->ppnm_eindx++; + + hibernate_paddr_map_inited = TRUE; } - if (page_count != npages) { - printf("pages 0x%x actual count 0x%x but requested 0x%x\n", - pages, page_count, npages); - panic("vm_page_verify_contiguous: count error"); +} + +ppnum_t +hibernate_lookup_paddr(unsigned int indx) +{ + struct ppnum_mapping *ppnm = NULL; + + ppnm = ppnm_last_found; + + if (ppnm) { + if (indx >= ppnm->ppnm_sindx && indx < ppnm->ppnm_eindx) + goto done; } - return 1; + for (ppnm = ppnm_head; ppnm; ppnm = ppnm->ppnm_next) { + + if (indx >= ppnm->ppnm_sindx && indx < ppnm->ppnm_eindx) { + ppnm_last_found = ppnm; + break; + } + } + if (ppnm == NULL) + panic("hibernate_lookup_paddr of %d failed\n", indx); +done: + return (ppnm->ppnm_base_paddr + (indx - ppnm->ppnm_sindx)); } -#endif /* MACH_ASSERT */ -cpm_counter(unsigned int vpfls_pages_handled = 0;) -cpm_counter(unsigned int vpfls_head_insertions = 0;) -cpm_counter(unsigned int vpfls_tail_insertions = 0;) -cpm_counter(unsigned int vpfls_general_insertions = 0;) -cpm_counter(unsigned int vpfc_failed = 0;) -cpm_counter(unsigned int vpfc_satisfied = 0;) +uint32_t +hibernate_mark_as_unneeded(addr64_t saddr, addr64_t eaddr, hibernate_page_list_t *page_list, hibernate_page_list_t *page_list_wired) +{ + addr64_t saddr_aligned; + addr64_t eaddr_aligned; + addr64_t addr; + ppnum_t paddr; + unsigned int mark_as_unneeded_pages = 0; -/* - * Find a region large enough to contain at least npages - * of contiguous physical memory. - * - * Requirements: - * - Called while holding vm_page_queue_free_lock. - * - Doesn't respect vm_page_free_reserved; caller - * must not ask for more pages than are legal to grab. - * - * Returns a pointer to a list of gobbled pages or VM_PAGE_NULL. - * - * Algorithm: - * Loop over the free list, extracting one page at a time and - * inserting those into a sorted sub-list. We stop as soon as - * there's a contiguous range within the sorted list that can - * satisfy the contiguous memory request. This contiguous sub- - * list is chopped out of the sorted sub-list and the remainder - * of the sorted sub-list is put back onto the beginning of the - * free list. - */ -static vm_page_t -vm_page_find_contiguous( - unsigned int contig_pages) + saddr_aligned = (saddr + PAGE_MASK_64) & ~PAGE_MASK_64; + eaddr_aligned = eaddr & ~PAGE_MASK_64; + + for (addr = saddr_aligned; addr < eaddr_aligned; addr += PAGE_SIZE_64) { + + paddr = pmap_find_phys(kernel_pmap, addr); + + assert(paddr); + + hibernate_page_bitset(page_list, TRUE, paddr); + hibernate_page_bitset(page_list_wired, TRUE, paddr); + + mark_as_unneeded_pages++; + } + return (mark_as_unneeded_pages); +} + + +void +hibernate_hash_insert_page(vm_page_t mem) { - vm_page_t sort_list; - vm_page_t *contfirstprev, contlast; - vm_page_t m, m1; - ppnum_t prevcontaddr; - ppnum_t nextcontaddr; - unsigned int npages; + vm_page_bucket_t *bucket; + int hash_id; + + assert(mem->hashed); + assert(mem->object); + assert(mem->offset != (vm_object_offset_t) -1); - m = NULL; -#if DEBUG - _mutex_assert(&vm_page_queue_free_lock, MA_OWNED); -#endif -#if MACH_ASSERT /* - * Verify pages in the free list.. + * Insert it into the object_object/offset hash table */ - npages = 0; - for (m = vm_page_queue_free; m != VM_PAGE_NULL; m = NEXT_PAGE(m)) - ++npages; - if (npages != vm_page_free_count) - panic("vm_sort_free_list: prelim: npages %u free_count %d", - npages, vm_page_free_count); -#endif /* MACH_ASSERT */ + hash_id = vm_page_hash(mem->object, mem->offset); + bucket = &vm_page_buckets[hash_id]; - if (contig_pages == 0 || vm_page_queue_free == VM_PAGE_NULL) - return VM_PAGE_NULL; + mem->next = bucket->pages; + bucket->pages = mem; +} -#define PPNUM_PREV(x) (((x) > 0) ? ((x) - 1) : 0) -#define PPNUM_NEXT(x) (((x) < PPNUM_MAX) ? ((x) + 1) : PPNUM_MAX) -#define SET_NEXT_PAGE(m,n) ((m)->pageq.next = (struct queue_entry *) (n)) - npages = 1; - contfirstprev = &sort_list; - contlast = sort_list = vm_page_queue_free; - vm_page_queue_free = NEXT_PAGE(sort_list); - SET_NEXT_PAGE(sort_list, VM_PAGE_NULL); - prevcontaddr = PPNUM_PREV(sort_list->phys_page); - nextcontaddr = PPNUM_NEXT(sort_list->phys_page); +void +hibernate_free_range(int sindx, int eindx) +{ + vm_page_t mem; + unsigned int color; - while (npages < contig_pages && - (m = vm_page_queue_free) != VM_PAGE_NULL) - { - cpm_counter(++vpfls_pages_handled); + while (sindx < eindx) { + mem = &vm_pages[sindx]; - /* prepend to existing run? */ - if (m->phys_page == prevcontaddr) - { - vm_page_queue_free = NEXT_PAGE(m); - cpm_counter(++vpfls_head_insertions); - prevcontaddr = PPNUM_PREV(prevcontaddr); - SET_NEXT_PAGE(m, *contfirstprev); - *contfirstprev = m; - npages++; - continue; /* no tail expansion check needed */ - } - - /* append to tail of existing run? */ - else if (m->phys_page == nextcontaddr) - { - vm_page_queue_free = NEXT_PAGE(m); - cpm_counter(++vpfls_tail_insertions); - nextcontaddr = PPNUM_NEXT(nextcontaddr); - SET_NEXT_PAGE(m, NEXT_PAGE(contlast)); - SET_NEXT_PAGE(contlast, m); - contlast = m; - npages++; - } + vm_page_init(mem, hibernate_lookup_paddr(sindx), FALSE); - /* prepend to the very front of sorted list? */ - else if (m->phys_page < sort_list->phys_page) - { - vm_page_queue_free = NEXT_PAGE(m); - cpm_counter(++vpfls_general_insertions); - prevcontaddr = PPNUM_PREV(m->phys_page); - nextcontaddr = PPNUM_NEXT(m->phys_page); - SET_NEXT_PAGE(m, sort_list); - contfirstprev = &sort_list; - contlast = sort_list = m; - npages = 1; - } + mem->lopage = FALSE; + mem->free = TRUE; - else /* get to proper place for insertion */ - { - if (m->phys_page < nextcontaddr) - { - prevcontaddr = PPNUM_PREV(sort_list->phys_page); - nextcontaddr = PPNUM_NEXT(sort_list->phys_page); - contfirstprev = &sort_list; - contlast = sort_list; - npages = 1; - } - for (m1 = NEXT_PAGE(contlast); - npages < contig_pages && - m1 != VM_PAGE_NULL && m1->phys_page < m->phys_page; - m1 = NEXT_PAGE(m1)) - { - if (m1->phys_page != nextcontaddr) { - prevcontaddr = PPNUM_PREV(m1->phys_page); - contfirstprev = NEXT_PAGE_PTR(contlast); - npages = 1; - } else { - npages++; - } - nextcontaddr = PPNUM_NEXT(m1->phys_page); - contlast = m1; - } + color = mem->phys_page & vm_color_mask; + queue_enter_first(&vm_page_queue_free[color], + mem, + vm_page_t, + pageq); + vm_page_free_count++; - /* - * We may actually already have enough. - * This could happen if a previous prepend - * joined up two runs to meet our needs. - * If so, bail before we take the current - * page off the free queue. - */ - if (npages == contig_pages) - break; + sindx++; + } +} - if (m->phys_page != nextcontaddr) - { - contfirstprev = NEXT_PAGE_PTR(contlast); - prevcontaddr = PPNUM_PREV(m->phys_page); - nextcontaddr = PPNUM_NEXT(m->phys_page); - npages = 1; - } else { - nextcontaddr = PPNUM_NEXT(nextcontaddr); - npages++; - } - vm_page_queue_free = NEXT_PAGE(m); - cpm_counter(++vpfls_general_insertions); - SET_NEXT_PAGE(m, NEXT_PAGE(contlast)); - SET_NEXT_PAGE(contlast, m); - contlast = m; - } + +extern void hibernate_rebuild_pmap_structs(void); + +void +hibernate_rebuild_vm_structs(void) +{ + int cindx, sindx, eindx; + vm_page_t mem, tmem, mem_next; + AbsoluteTime startTime, endTime; + uint64_t nsec; + + if (hibernate_rebuild_needed == FALSE) + return; + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 13) | DBG_FUNC_START, 0, 0, 0, 0, 0); + HIBLOG("hibernate_rebuild started\n"); + + clock_get_uptime(&startTime); + + hibernate_rebuild_pmap_structs(); + + bzero(&vm_page_buckets[0], vm_page_bucket_count * sizeof(vm_page_bucket_t)); + eindx = vm_pages_count; + + for (cindx = hibernate_teardown_last_valid_compact_indx; cindx >= 0; cindx--) { - /* See how many pages are now contiguous after the insertion */ - for (m1 = NEXT_PAGE(m); - npages < contig_pages && - m1 != VM_PAGE_NULL && m1->phys_page == nextcontaddr; - m1 = NEXT_PAGE(m1)) - { - nextcontaddr = PPNUM_NEXT(nextcontaddr); - contlast = m1; - npages++; + mem = &vm_pages[cindx]; + /* + * hibernate_teardown_vm_structs leaves the location where + * this vm_page_t must be located in "next". + */ + tmem = mem->next; + mem->next = NULL; + + sindx = (int)(tmem - &vm_pages[0]); + + if (mem != tmem) { + /* + * this vm_page_t was moved by hibernate_teardown_vm_structs, + * so move it back to its real location + */ + *tmem = *mem; + mem = tmem; } - } + if (mem->hashed) + hibernate_hash_insert_page(mem); + /* + * the 'hole' between this vm_page_t and the previous + * vm_page_t we moved needs to be initialized as + * a range of free vm_page_t's + */ + hibernate_free_range(sindx + 1, eindx); - /* how did we do? */ - if (npages == contig_pages) - { - cpm_counter(++vpfc_satisfied); + eindx = sindx; + } + if (sindx) + hibernate_free_range(0, sindx); - /* remove the contiguous range from the sorted list */ - m = *contfirstprev; - *contfirstprev = NEXT_PAGE(contlast); - SET_NEXT_PAGE(contlast, VM_PAGE_NULL); - assert(vm_page_verify_contiguous(m, npages)); + assert(vm_page_free_count == hibernate_teardown_vm_page_free_count); - /* inline vm_page_gobble() for each returned page */ - for (m1 = m; m1 != VM_PAGE_NULL; m1 = NEXT_PAGE(m1)) { - assert(m1->free); - assert(!m1->wanted); - assert(!m1->laundry); - m1->free = FALSE; - m1->no_isync = TRUE; - m1->gobbled = TRUE; - } - vm_page_wire_count += npages; - vm_page_gobble_count += npages; - vm_page_free_count -= npages; + /* + * process the list of vm_page_t's that were entered in the hash, + * but were not located in the vm_pages arrary... these are + * vm_page_t's that were created on the fly (i.e. fictitious) + */ + for (mem = hibernate_rebuild_hash_list; mem; mem = mem_next) { + mem_next = mem->next; - /* stick free list at the tail of the sorted list */ - while ((m1 = *contfirstprev) != VM_PAGE_NULL) - contfirstprev = (vm_page_t *)&m1->pageq.next; - *contfirstprev = vm_page_queue_free; + mem->next = NULL; + hibernate_hash_insert_page(mem); } + hibernate_rebuild_hash_list = NULL; - vm_page_queue_free = sort_list; - return m; + clock_get_uptime(&endTime); + SUB_ABSOLUTETIME(&endTime, &startTime); + absolutetime_to_nanoseconds(endTime, &nsec); + + HIBLOG("hibernate_rebuild completed - took %qd msecs\n", nsec / 1000000ULL); + + hibernate_rebuild_needed = FALSE; + + KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 13) | DBG_FUNC_END, 0, 0, 0, 0, 0); } -/* - * Allocate a list of contiguous, wired pages. - */ -kern_return_t -cpm_allocate( - vm_size_t size, - vm_page_t *list, - boolean_t wire) + +extern void hibernate_teardown_pmap_structs(addr64_t *, addr64_t *); + +uint32_t +hibernate_teardown_vm_structs(hibernate_page_list_t *page_list, hibernate_page_list_t *page_list_wired) { - register vm_page_t m; - vm_page_t pages; - unsigned int npages; - unsigned int vm_pages_available; - boolean_t wakeup; + unsigned int i; + unsigned int compact_target_indx; + vm_page_t mem, mem_next; + vm_page_bucket_t *bucket; + unsigned int mark_as_unneeded_pages = 0; + unsigned int unneeded_vm_page_bucket_pages = 0; + unsigned int unneeded_vm_pages_pages = 0; + unsigned int unneeded_pmap_pages = 0; + addr64_t start_of_unneeded = 0; + addr64_t end_of_unneeded = 0; - if (size % page_size != 0) - return KERN_INVALID_ARGUMENT; + + if (hibernate_should_abort()) + return (0); - vm_page_lock_queues(); - mutex_lock(&vm_page_queue_free_lock); + HIBLOG("hibernate_teardown: wired_pages %d, free_pages %d, active_pages %d, inactive_pages %d, speculative_pages %d, cleaned_pages %d, compressor_pages %d\n", + vm_page_wire_count, vm_page_free_count, vm_page_active_count, vm_page_inactive_count, vm_page_speculative_count, + vm_page_cleaned_count, compressor_object->resident_page_count); - /* - * Should also take active and inactive pages - * into account... One day... - */ - npages = size / page_size; - vm_pages_available = vm_page_free_count - vm_page_free_reserved; + for (i = 0; i < vm_page_bucket_count; i++) { - if (npages > vm_pages_available) { - mutex_unlock(&vm_page_queue_free_lock); - vm_page_unlock_queues(); - return KERN_RESOURCE_SHORTAGE; + bucket = &vm_page_buckets[i]; + + for (mem = bucket->pages; mem != VM_PAGE_NULL; mem = mem_next) { + assert(mem->hashed); + + mem_next = mem->next; + + if (mem < &vm_pages[0] || mem >= &vm_pages[vm_pages_count]) { + mem->next = hibernate_rebuild_hash_list; + hibernate_rebuild_hash_list = mem; + } + } } + unneeded_vm_page_bucket_pages = hibernate_mark_as_unneeded((addr64_t)&vm_page_buckets[0], (addr64_t)&vm_page_buckets[vm_page_bucket_count], page_list, page_list_wired); + mark_as_unneeded_pages += unneeded_vm_page_bucket_pages; - /* - * Obtain a pointer to a subset of the free - * list large enough to satisfy the request; - * the region will be physically contiguous. - */ - pages = vm_page_find_contiguous(npages); + hibernate_teardown_vm_page_free_count = vm_page_free_count; - /* adjust global freelist counts and determine need for wakeups */ - if (vm_page_free_count < vm_page_free_count_minimum) - vm_page_free_count_minimum = vm_page_free_count; + compact_target_indx = 0; - wakeup = ((vm_page_free_count < vm_page_free_min) || - ((vm_page_free_count < vm_page_free_target) && - (vm_page_inactive_count < vm_page_inactive_target))); - - mutex_unlock(&vm_page_queue_free_lock); + for (i = 0; i < vm_pages_count; i++) { - if (pages == VM_PAGE_NULL) { - vm_page_unlock_queues(); - return KERN_NO_SPACE; - } + mem = &vm_pages[i]; - /* - * Walk the returned list, wiring the pages. - */ - if (wire == TRUE) - for (m = pages; m != VM_PAGE_NULL; m = NEXT_PAGE(m)) { + if (mem->free) { + unsigned int color; + + assert(mem->busy); + assert(!mem->lopage); + + color = mem->phys_page & vm_color_mask; + + queue_remove(&vm_page_queue_free[color], + mem, + vm_page_t, + pageq); + mem->pageq.next = NULL; + mem->pageq.prev = NULL; + + vm_page_free_count--; + + hibernate_teardown_found_free_pages++; + + if ( !vm_pages[compact_target_indx].free) + compact_target_indx = i; + } else { /* - * Essentially inlined vm_page_wire. + * record this vm_page_t's original location + * we need this even if it doesn't get moved + * as an indicator to the rebuild function that + * we don't have to move it */ - assert(!m->active); - assert(!m->inactive); - assert(!m->private); - assert(!m->fictitious); - assert(m->wire_count == 0); - assert(m->gobbled); - m->gobbled = FALSE; - m->wire_count++; - --vm_page_gobble_count; + mem->next = mem; + + if (vm_pages[compact_target_indx].free) { + /* + * we've got a hole to fill, so + * move this vm_page_t to it's new home + */ + vm_pages[compact_target_indx] = *mem; + mem->free = TRUE; + + hibernate_teardown_last_valid_compact_indx = compact_target_indx; + compact_target_indx++; + } else + hibernate_teardown_last_valid_compact_indx = i; } - vm_page_unlock_queues(); + } + unneeded_vm_pages_pages = hibernate_mark_as_unneeded((addr64_t)&vm_pages[hibernate_teardown_last_valid_compact_indx+1], + (addr64_t)&vm_pages[vm_pages_count-1], page_list, page_list_wired); + mark_as_unneeded_pages += unneeded_vm_pages_pages; - if (wakeup) - thread_wakeup((event_t) &vm_page_free_wanted); + hibernate_teardown_pmap_structs(&start_of_unneeded, &end_of_unneeded); - /* - * The CPM pages should now be available and - * ordered by ascending physical address. - */ - assert(vm_page_verify_contiguous(pages, npages)); + if (start_of_unneeded) { + unneeded_pmap_pages = hibernate_mark_as_unneeded(start_of_unneeded, end_of_unneeded, page_list, page_list_wired); + mark_as_unneeded_pages += unneeded_pmap_pages; + } + HIBLOG("hibernate_teardown: mark_as_unneeded_pages %d, %d, %d\n", unneeded_vm_page_bucket_pages, unneeded_vm_pages_pages, unneeded_pmap_pages); - *list = pages; - return KERN_SUCCESS; + hibernate_rebuild_needed = TRUE; + + return (mark_as_unneeded_pages); } +#endif /* HIBERNATION */ + +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + #include #if MACH_VM_DEBUG @@ -2649,6 +6318,7 @@ vm_page_info( unsigned int count) { unsigned int i; + lck_spin_t *bucket_lock; if (vm_page_bucket_count < count) count = vm_page_bucket_count; @@ -2658,10 +6328,13 @@ vm_page_info( unsigned int bucket_count = 0; vm_page_t m; - simple_lock(&vm_page_bucket_lock); + bucket_lock = &vm_page_bucket_locks[i / BUCKETS_PER_LOCK]; + lck_spin_lock(bucket_lock); + for (m = bucket->pages; m != VM_PAGE_NULL; m = m->next) bucket_count++; - simple_unlock(&vm_page_bucket_lock); + + lck_spin_unlock(bucket_lock); /* don't touch pageable memory while holding locks */ info[i].hib_count = bucket_count; @@ -2671,65 +6344,75 @@ vm_page_info( } #endif /* MACH_VM_DEBUG */ -#include -#if MACH_KDB - -#include -#include -#define printf kdbprintf - -/* - * Routine: vm_page_print [exported] - */ +#if VM_PAGE_BUCKETS_CHECK void -vm_page_print( - db_addr_t db_addr) +vm_page_buckets_check(void) { - vm_page_t p; + unsigned int i; + vm_page_t p; + unsigned int p_hash; + vm_page_bucket_t *bucket; + lck_spin_t *bucket_lock; - p = (vm_page_t) (long) db_addr; + if (!vm_page_buckets_check_ready) { + return; + } - iprintf("page 0x%x\n", p); +#if HIBERNATION + if (hibernate_rebuild_needed || + hibernate_rebuild_hash_list) { + panic("BUCKET_CHECK: hibernation in progress: " + "rebuild_needed=%d rebuild_hash_list=%p\n", + hibernate_rebuild_needed, + hibernate_rebuild_hash_list); + } +#endif /* HIBERNATION */ + +#if VM_PAGE_FAKE_BUCKETS + char *cp; + for (cp = (char *) vm_page_fake_buckets_start; + cp < (char *) vm_page_fake_buckets_end; + cp++) { + if (*cp != 0x5a) { + panic("BUCKET_CHECK: corruption at %p in fake buckets " + "[0x%llx:0x%llx]\n", + cp, + vm_page_fake_buckets_start, + vm_page_fake_buckets_end); + } + } +#endif /* VM_PAGE_FAKE_BUCKETS */ - db_indent += 2; + for (i = 0; i < vm_page_bucket_count; i++) { + bucket = &vm_page_buckets[i]; + if (bucket->pages == VM_PAGE_NULL) { + continue; + } - iprintf("object=0x%x", p->object); - printf(", offset=0x%x", p->offset); - printf(", wire_count=%d", p->wire_count); + bucket_lock = &vm_page_bucket_locks[i / BUCKETS_PER_LOCK]; + lck_spin_lock(bucket_lock); + p = bucket->pages; + while (p != VM_PAGE_NULL) { + if (!p->hashed) { + panic("BUCKET_CHECK: page %p (%p,0x%llx) " + "hash %d in bucket %d at %p " + "is not hashed\n", + p, p->object, p->offset, + p_hash, i, bucket); + } + p_hash = vm_page_hash(p->object, p->offset); + if (p_hash != i) { + panic("BUCKET_CHECK: corruption in bucket %d " + "at %p: page %p object %p offset 0x%llx " + "hash %d\n", + i, bucket, p, p->object, p->offset, + p_hash); + } + p = p->next; + } + lck_spin_unlock(bucket_lock); + } - iprintf("%sinactive, %sactive, %sgobbled, %slaundry, %sfree, %sref, %sencrypted\n", - (p->inactive ? "" : "!"), - (p->active ? "" : "!"), - (p->gobbled ? "" : "!"), - (p->laundry ? "" : "!"), - (p->free ? "" : "!"), - (p->reference ? "" : "!"), - (p->encrypted ? "" : "!")); - iprintf("%sbusy, %swanted, %stabled, %sfictitious, %sprivate, %sprecious\n", - (p->busy ? "" : "!"), - (p->wanted ? "" : "!"), - (p->tabled ? "" : "!"), - (p->fictitious ? "" : "!"), - (p->private ? "" : "!"), - (p->precious ? "" : "!")); - iprintf("%sabsent, %serror, %sdirty, %scleaning, %spageout, %sclustered\n", - (p->absent ? "" : "!"), - (p->error ? "" : "!"), - (p->dirty ? "" : "!"), - (p->cleaning ? "" : "!"), - (p->pageout ? "" : "!"), - (p->clustered ? "" : "!")); - iprintf("%slock_supplied, %soverwriting, %srestart, %sunusual\n", - (p->lock_supplied ? "" : "!"), - (p->overwriting ? "" : "!"), - (p->restart ? "" : "!"), - (p->unusual ? "" : "!")); - - iprintf("phys_page=0x%x", p->phys_page); - printf(", page_error=0x%x", p->page_error); - printf(", page_lock=0x%x", p->page_lock); - printf(", unlock_request=%d\n", p->unlock_request); - - db_indent -= 2; -} -#endif /* MACH_KDB */ +// printf("BUCKET_CHECK: checked buckets\n"); +} +#endif /* VM_PAGE_BUCKETS_CHECK */