X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/04b8595b18b1b41ac7a206e4b3d51a635f8413d7..HEAD:/osfmk/vm/vm_compressor.c?ds=inline diff --git a/osfmk/vm/vm_compressor.c b/osfmk/vm/vm_compressor.c index 224858282..315c8b429 100644 --- a/osfmk/vm/vm_compressor.c +++ b/osfmk/vm/vm_compressor.c @@ -1,8 +1,8 @@ /* - * Copyright (c) 2000-2013 Apple Inc. All rights reserved. + * Copyright (c) 2000-2020 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ - * + * * This file contains Original Code and/or Modifications of Original Code * as defined in and that are subject to the Apple Public Source License * Version 2.0 (the 'License'). You may not use this file except in @@ -11,10 +11,10 @@ * unlawful or unlicensed copies of an Apple operating system, or to * circumvent, violate, or enable the circumvention or violation of, any * terms of an Apple operating system software license agreement. - * + * * Please obtain a copy of the License at * http://www.opensource.apple.com/apsl/ and read it before using this file. - * + * * The Original Code and all software distributed under the License are * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, @@ -22,7 +22,7 @@ * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. * Please see the License for the specific language governing rights and * limitations under the License. - * + * * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ @@ -35,14 +35,82 @@ #include #include #include -#include /* for host_info() */ +#include +#include +#include +#include /* for host_info() */ #include +#include +#include +#include -#include -#include +#if defined(__x86_64__) +#include +#endif +#if defined(__arm64__) +#include +#endif #include +extern boolean_t vm_darkwake_mode; +extern zone_t vm_page_zone; + +#if DEVELOPMENT || DEBUG +/* sysctl defined in bsd/dev/arm64/sysctl.c */ +int do_cseg_wedge_thread(void); +int do_cseg_unwedge_thread(void); +static event_t debug_cseg_wait_event = NULL; +#endif /* DEVELOPMENT || DEBUG */ + +#if CONFIG_FREEZE +bool freezer_incore_cseg_acct = TRUE; /* Only count incore compressed memory for jetsams. */ +void task_disown_frozen_csegs(task_t owner_task); +#endif /* CONFIG_FREEZE */ + +#if POPCOUNT_THE_COMPRESSED_DATA +boolean_t popcount_c_segs = TRUE; + +static inline uint32_t +vmc_pop(uintptr_t ins, int sz) +{ + uint32_t rv = 0; + + if (__probable(popcount_c_segs == FALSE)) { + return 0xDEAD707C; + } + + while (sz >= 16) { + uint32_t rv1, rv2; + uint64_t *ins64 = (uint64_t *) ins; + uint64_t *ins642 = (uint64_t *) (ins + 8); + rv1 = __builtin_popcountll(*ins64); + rv2 = __builtin_popcountll(*ins642); + rv += rv1 + rv2; + sz -= 16; + ins += 16; + } + + while (sz >= 4) { + uint32_t *ins32 = (uint32_t *) ins; + rv += __builtin_popcount(*ins32); + sz -= 4; + ins += 4; + } + + while (sz > 0) { + char *ins8 = (char *)ins; + rv += __builtin_popcount(*ins8); + sz--; + ins++; + } + return rv; +} +#endif + +#if VALIDATE_C_SEGMENTS +boolean_t validate_c_segs = TRUE; +#endif /* * vm_compressor_mode has a heirarchy of control to set its value. * boot-args are checked first, then device-tree, and finally @@ -50,165 +118,225 @@ * the boot-arg & device-tree code. */ -extern ipc_port_t min_pages_trigger_port; -extern lck_mtx_t paging_segments_lock; -#define PSL_LOCK() lck_mtx_lock(&paging_segments_lock) -#define PSL_UNLOCK() lck_mtx_unlock(&paging_segments_lock) +#if !XNU_TARGET_OS_OSX +#if CONFIG_FREEZE +int vm_compressor_mode = VM_PAGER_FREEZER_DEFAULT; +struct freezer_context freezer_context_global; +#else /* CONFIG_FREEZE */ +int vm_compressor_mode = VM_PAGER_NOT_CONFIGURED; +#endif /* CONFIG_FREEZE */ -int vm_compressor_mode = VM_PAGER_COMPRESSOR_WITH_SWAP; -int vm_scale = 16; +#else /* !XNU_TARGET_OS_OSX */ +int vm_compressor_mode = VM_PAGER_COMPRESSOR_WITH_SWAP; +#endif /* !XNU_TARGET_OS_OSX */ -int vm_compressor_is_active = 0; -int vm_compression_limit = 0; +TUNABLE(uint32_t, vm_compression_limit, "vm_compression_limit", 0); +int vm_compressor_is_active = 0; +int vm_compressor_available = 0; -extern boolean_t vm_swap_up; -extern void vm_pageout_io_throttle(void); +extern uint64_t vm_swap_get_max_configured_space(void); +extern void vm_pageout_io_throttle(void); #if CHECKSUM_THE_DATA || CHECKSUM_THE_SWAP || CHECKSUM_THE_COMPRESSED_DATA extern unsigned int hash_string(char *cp, int len); -#endif - +static unsigned int vmc_hash(char *, int); +boolean_t checksum_c_segs = TRUE; -struct c_slot { - uint64_t c_offset:C_SEG_OFFSET_BITS, - c_size:12, - c_packed_ptr:36; -#if CHECKSUM_THE_DATA - unsigned int c_hash_data; -#endif -#if CHECKSUM_THE_COMPRESSED_DATA - unsigned int c_hash_compressed_data; +unsigned int +vmc_hash(char *cp, int len) +{ + if (__probable(checksum_c_segs == FALSE)) { + return 0xDEAD7A37; + } + return hash_string(cp, len); +} #endif -}; - -#define UNPACK_C_SIZE(cs) ((cs->c_size == (PAGE_SIZE-1)) ? PAGE_SIZE : cs->c_size) -#define PACK_C_SIZE(cs, size) (cs->c_size = ((size == PAGE_SIZE) ? PAGE_SIZE - 1 : size)) +#define UNPACK_C_SIZE(cs) ((cs->c_size == (PAGE_SIZE-1)) ? PAGE_SIZE : cs->c_size) +#define PACK_C_SIZE(cs, size) (cs->c_size = ((size == PAGE_SIZE) ? PAGE_SIZE - 1 : size)) -struct c_slot_mapping { - uint32_t s_cseg:22, /* segment number + 1 */ - s_cindx:10; /* index in the segment */ +struct c_sv_hash_entry { + union { + struct { + uint32_t c_sv_he_ref; + uint32_t c_sv_he_data; + } c_sv_he; + uint64_t c_sv_he_record; + } c_sv_he_un; }; -typedef struct c_slot_mapping *c_slot_mapping_t; +#define he_ref c_sv_he_un.c_sv_he.c_sv_he_ref +#define he_data c_sv_he_un.c_sv_he.c_sv_he_data +#define he_record c_sv_he_un.c_sv_he_record + +#define C_SV_HASH_MAX_MISS 32 +#define C_SV_HASH_SIZE ((1 << 10)) +#define C_SV_HASH_MASK ((1 << 10) - 1) +#define C_SV_CSEG_ID ((1 << 22) - 1) union c_segu { - c_segment_t c_seg; - uint32_t c_segno; + c_segment_t c_seg; + uintptr_t c_segno; }; +#define C_SLOT_ASSERT_PACKABLE(ptr) \ + VM_ASSERT_POINTER_PACKABLE((vm_offset_t)(ptr), C_SLOT_PACKED_PTR); +#define C_SLOT_PACK_PTR(ptr) \ + VM_PACK_POINTER((vm_offset_t)(ptr), C_SLOT_PACKED_PTR) -#define C_SLOT_PACK_PTR(ptr) (((uintptr_t)ptr - (uintptr_t) VM_MIN_KERNEL_AND_KEXT_ADDRESS) >> 2) -#define C_SLOT_UNPACK_PTR(cslot) ((uintptr_t)(cslot->c_packed_ptr << 2) + (uintptr_t) VM_MIN_KERNEL_AND_KEXT_ADDRESS) +#define C_SLOT_UNPACK_PTR(cslot) \ + (c_slot_mapping_t)VM_UNPACK_POINTER((cslot)->c_packed_ptr, C_SLOT_PACKED_PTR) +/* for debugging purposes */ +SECURITY_READ_ONLY_EARLY(vm_packing_params_t) c_slot_packing_params = + VM_PACKING_PARAMS(C_SLOT_PACKED_PTR); -uint32_t c_segment_count = 0; +uint32_t c_segment_count = 0; +uint32_t c_segment_count_max = 0; -uint64_t c_generation_id = 0; -uint64_t c_generation_id_flush_barrier; +uint64_t c_generation_id = 0; +uint64_t c_generation_id_flush_barrier; -#define HIBERNATE_FLUSHING_SECS_TO_COMPLETE 120 +#define HIBERNATE_FLUSHING_SECS_TO_COMPLETE 120 -boolean_t hibernate_no_swapspace = FALSE; -clock_sec_t hibernate_flushing_deadline = 0; +boolean_t hibernate_no_swapspace = FALSE; +clock_sec_t hibernate_flushing_deadline = 0; -#if TRACK_BAD_C_SEGMENTS -queue_head_t c_bad_list_head; -uint32_t c_bad_count = 0; +#if RECORD_THE_COMPRESSED_DATA +char *c_compressed_record_sbuf; +char *c_compressed_record_ebuf; +char *c_compressed_record_cptr; #endif -queue_head_t c_age_list_head; -queue_head_t c_swapout_list_head; -queue_head_t c_swappedin_list_head; -queue_head_t c_swappedout_list_head; -queue_head_t c_swappedout_sparse_list_head; -uint32_t c_age_count = 0; -uint32_t c_swapout_count = 0; -uint32_t c_swappedin_count = 0; -uint32_t c_swappedout_count = 0; -uint32_t c_swappedout_sparse_count = 0; +queue_head_t c_age_list_head; +queue_head_t c_swappedin_list_head; +queue_head_t c_swapout_list_head; +queue_head_t c_swapio_list_head; +queue_head_t c_swappedout_list_head; +queue_head_t c_swappedout_sparse_list_head; +queue_head_t c_major_list_head; +queue_head_t c_filling_list_head; +queue_head_t c_bad_list_head; + +uint32_t c_age_count = 0; +uint32_t c_swappedin_count = 0; +uint32_t c_swapout_count = 0; +uint32_t c_swapio_count = 0; +uint32_t c_swappedout_count = 0; +uint32_t c_swappedout_sparse_count = 0; +uint32_t c_major_count = 0; +uint32_t c_filling_count = 0; +uint32_t c_empty_count = 0; +uint32_t c_bad_count = 0; + + +queue_head_t c_minor_list_head; +uint32_t c_minor_count = 0; + +int c_overage_swapped_count = 0; +int c_overage_swapped_limit = 0; + +int c_seg_fixed_array_len; +union c_segu *c_segments; +vm_offset_t c_buffers; +vm_size_t c_buffers_size; +caddr_t c_segments_next_page; +boolean_t c_segments_busy; +uint32_t c_segments_available; +uint32_t c_segments_limit; +uint32_t c_segments_nearing_limit; + +uint32_t c_segment_svp_in_hash; +uint32_t c_segment_svp_hash_succeeded; +uint32_t c_segment_svp_hash_failed; +uint32_t c_segment_svp_zero_compressions; +uint32_t c_segment_svp_nonzero_compressions; +uint32_t c_segment_svp_zero_decompressions; +uint32_t c_segment_svp_nonzero_decompressions; + +uint32_t c_segment_noncompressible_pages; + +uint32_t c_segment_pages_compressed = 0; /* Tracks # of uncompressed pages fed into the compressor */ +#if CONFIG_FREEZE +int32_t c_segment_pages_compressed_incore = 0; /* Tracks # of uncompressed pages fed into the compressor that are in memory */ +uint32_t c_segments_incore_limit = 0; /* Tracks # of segments allowed to be in-core. Based on compressor pool size */ +#endif /* CONFIG_FREEZE */ -queue_head_t c_minor_list_head; -uint32_t c_minor_count = 0; +uint32_t c_segment_pages_compressed_limit; +uint32_t c_segment_pages_compressed_nearing_limit; +uint32_t c_free_segno_head = (uint32_t)-1; -union c_segu *c_segments; -caddr_t c_segments_next_page; -boolean_t c_segments_busy; -uint32_t c_segments_available; -uint32_t c_segments_limit; -uint32_t c_segments_nearing_limit; -uint32_t c_segment_pages_compressed; -uint32_t c_segment_pages_compressed_limit; -uint32_t c_segment_pages_compressed_nearing_limit; -uint32_t c_free_segno_head = (uint32_t)-1; +uint32_t vm_compressor_minorcompact_threshold_divisor = 10; +uint32_t vm_compressor_majorcompact_threshold_divisor = 10; +uint32_t vm_compressor_unthrottle_threshold_divisor = 10; +uint32_t vm_compressor_catchup_threshold_divisor = 10; -uint32_t vm_compressor_minorcompact_threshold_divisor = 10; -uint32_t vm_compressor_majorcompact_threshold_divisor = 10; -uint32_t vm_compressor_unthrottle_threshold_divisor = 10; -uint32_t vm_compressor_catchup_threshold_divisor = 10; +uint32_t vm_compressor_minorcompact_threshold_divisor_overridden = 0; +uint32_t vm_compressor_majorcompact_threshold_divisor_overridden = 0; +uint32_t vm_compressor_unthrottle_threshold_divisor_overridden = 0; +uint32_t vm_compressor_catchup_threshold_divisor_overridden = 0; -#define C_SEGMENTS_PER_PAGE (PAGE_SIZE / sizeof(union c_segu)) +#define C_SEGMENTS_PER_PAGE (PAGE_SIZE / sizeof(union c_segu)) +LCK_GRP_DECLARE(vm_compressor_lck_grp, "vm_compressor"); +LCK_RW_DECLARE(c_master_lock, &vm_compressor_lck_grp); +LCK_MTX_DECLARE(c_list_lock_storage, &vm_compressor_lck_grp); -lck_grp_attr_t vm_compressor_lck_grp_attr; -lck_attr_t vm_compressor_lck_attr; -lck_grp_t vm_compressor_lck_grp; +boolean_t decompressions_blocked = FALSE; +zone_t compressor_segment_zone; +int c_compressor_swap_trigger = 0; -#if __i386__ || __x86_64__ -lck_mtx_t *c_list_lock; -#else /* __i386__ || __x86_64__ */ -lck_spin_t *c_list_lock; -#endif /* __i386__ || __x86_64__ */ +uint32_t compressor_cpus; +char *compressor_scratch_bufs; +char *kdp_compressor_scratch_buf; +char *kdp_compressor_decompressed_page; +addr64_t kdp_compressor_decompressed_page_paddr; +ppnum_t kdp_compressor_decompressed_page_ppnum; -lck_rw_t c_master_lock; -boolean_t decompressions_blocked = FALSE; +clock_sec_t start_of_sample_period_sec = 0; +clock_nsec_t start_of_sample_period_nsec = 0; +clock_sec_t start_of_eval_period_sec = 0; +clock_nsec_t start_of_eval_period_nsec = 0; +uint32_t sample_period_decompression_count = 0; +uint32_t sample_period_compression_count = 0; +uint32_t last_eval_decompression_count = 0; +uint32_t last_eval_compression_count = 0; -zone_t compressor_segment_zone; -int c_compressor_swap_trigger = 0; +#define DECOMPRESSION_SAMPLE_MAX_AGE (60 * 30) -uint32_t compressor_cpus; -char *compressor_scratch_bufs; +boolean_t vm_swapout_ripe_segments = FALSE; +uint32_t vm_ripe_target_age = (60 * 60 * 48); +uint32_t swapout_target_age = 0; +uint32_t age_of_decompressions_during_sample_period[DECOMPRESSION_SAMPLE_MAX_AGE]; +uint32_t overage_decompressions_during_sample_period = 0; -clock_sec_t start_of_sample_period_sec = 0; -clock_nsec_t start_of_sample_period_nsec = 0; -clock_sec_t start_of_eval_period_sec = 0; -clock_nsec_t start_of_eval_period_nsec = 0; -uint32_t sample_period_decompression_count = 0; -uint32_t sample_period_compression_count = 0; -uint32_t last_eval_decompression_count = 0; -uint32_t last_eval_compression_count = 0; -#define DECOMPRESSION_SAMPLE_MAX_AGE (60 * 30) +void do_fastwake_warmup(queue_head_t *, boolean_t); +boolean_t fastwake_warmup = FALSE; +boolean_t fastwake_recording_in_progress = FALSE; +clock_sec_t dont_trim_until_ts = 0; -uint32_t swapout_target_age = 0; -uint32_t age_of_decompressions_during_sample_period[DECOMPRESSION_SAMPLE_MAX_AGE]; -uint32_t overage_decompressions_during_sample_period = 0; +uint64_t c_segment_warmup_count; +uint64_t first_c_segment_to_warm_generation_id = 0; +uint64_t last_c_segment_to_warm_generation_id = 0; +boolean_t hibernate_flushing = FALSE; -void do_fastwake_warmup(void); -boolean_t fastwake_warmup = FALSE; -boolean_t fastwake_recording_in_progress = FALSE; -clock_sec_t dont_trim_until_ts = 0; +int64_t c_segment_input_bytes __attribute__((aligned(8))) = 0; +int64_t c_segment_compressed_bytes __attribute__((aligned(8))) = 0; +int64_t compressor_bytes_used __attribute__((aligned(8))) = 0; -uint64_t c_segment_warmup_count; -uint64_t first_c_segment_to_warm_generation_id = 0; -uint64_t last_c_segment_to_warm_generation_id = 0; -boolean_t hibernate_flushing = FALSE; -int64_t c_segment_input_bytes __attribute__((aligned(8))) = 0; -int64_t c_segment_compressed_bytes __attribute__((aligned(8))) = 0; -int64_t compressor_bytes_used __attribute__((aligned(8))) = 0; -uint64_t compressor_kvspace_used __attribute__((aligned(8))) = 0; -uint64_t compressor_kvwaste_limit = 0; +struct c_sv_hash_entry c_segment_sv_hash_table[C_SV_HASH_SIZE] __attribute__ ((aligned(8))); static boolean_t compressor_needs_to_swap(void); static void vm_compressor_swap_trigger_thread(void); @@ -216,7 +344,9 @@ static void vm_compressor_do_delayed_compactions(boolean_t); static void vm_compressor_compact_and_swap(boolean_t); static void vm_compressor_age_swapped_in_segments(boolean_t); -boolean_t vm_compressor_low_on_space(void); +#if XNU_TARGET_OS_OSX +static void vm_compressor_take_paging_space_action(void); +#endif /* XNU_TARGET_OS_OSX */ void compute_swapout_target_age(void); @@ -226,87 +356,161 @@ boolean_t c_seg_major_compact_ok(c_segment_t, c_segment_t); int c_seg_minor_compaction_and_unlock(c_segment_t, boolean_t); int c_seg_do_minor_compaction_and_unlock(c_segment_t, boolean_t, boolean_t, boolean_t); void c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg); -void c_seg_need_delayed_compaction(c_segment_t); void c_seg_move_to_sparse_list(c_segment_t); void c_seg_insert_into_q(queue_head_t *, c_segment_t); -boolean_t c_seg_try_free(c_segment_t); -void c_seg_free(c_segment_t); -void c_seg_free_locked(c_segment_t); - - uint64_t vm_available_memory(void); uint64_t vm_compressor_pages_compressed(void); -extern unsigned int dp_pages_free, dp_pages_reserve; +/* + * indicate the need to do a major compaction if + * the overall set of in-use compression segments + * becomes sparse... on systems that support pressure + * driven swapping, this will also cause swapouts to + * be initiated. + */ +static inline boolean_t +vm_compressor_needs_to_major_compact() +{ + uint32_t incore_seg_count; + + incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count; + + if ((c_segment_count >= (c_segments_nearing_limit / 8)) && + ((incore_seg_count * C_SEG_MAX_PAGES) - VM_PAGE_COMPRESSOR_COUNT) > + ((incore_seg_count / 8) * C_SEG_MAX_PAGES)) { + return 1; + } + return 0; +} + uint64_t vm_available_memory(void) { - return (((uint64_t)AVAILABLE_NON_COMPRESSED_MEMORY) * PAGE_SIZE_64); + return ((uint64_t)AVAILABLE_NON_COMPRESSED_MEMORY) * PAGE_SIZE_64; } uint64_t vm_compressor_pages_compressed(void) { - return (c_segment_pages_compressed * PAGE_SIZE_64); + return c_segment_pages_compressed * PAGE_SIZE_64; } boolean_t -vm_compression_available(void) +vm_compressor_low_on_space(void) { - if ( !(COMPRESSED_PAGER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE)) - return (FALSE); - - if (c_segments_available >= c_segments_limit || c_segment_pages_compressed >= c_segment_pages_compressed_limit) - return (FALSE); +#if CONFIG_FREEZE + uint64_t incore_seg_count; + uint32_t incore_compressed_pages; + if (freezer_incore_cseg_acct) { + incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count; + incore_compressed_pages = c_segment_pages_compressed_incore; + } else { + incore_seg_count = c_segment_count; + incore_compressed_pages = c_segment_pages_compressed; + } - return (TRUE); + if ((incore_compressed_pages > c_segment_pages_compressed_nearing_limit) || + (incore_seg_count > c_segments_nearing_limit)) { + return TRUE; + } +#else /* CONFIG_FREEZE */ + if ((c_segment_pages_compressed > c_segment_pages_compressed_nearing_limit) || + (c_segment_count > c_segments_nearing_limit)) { + return TRUE; + } +#endif /* CONFIG_FREEZE */ + return FALSE; } boolean_t -vm_compressor_low_on_space(void) +vm_compressor_out_of_space(void) { - if ((c_segment_pages_compressed > c_segment_pages_compressed_nearing_limit) || - (c_segment_count > c_segments_nearing_limit)) - return (TRUE); +#if CONFIG_FREEZE + uint64_t incore_seg_count; + uint32_t incore_compressed_pages; + if (freezer_incore_cseg_acct) { + incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count; + incore_compressed_pages = c_segment_pages_compressed_incore; + } else { + incore_seg_count = c_segment_count; + incore_compressed_pages = c_segment_pages_compressed; + } - return (FALSE); + if ((incore_compressed_pages >= c_segment_pages_compressed_limit) || + (incore_seg_count > c_segments_incore_limit)) { + return TRUE; + } +#else /* CONFIG_FREEZE */ + if ((c_segment_pages_compressed >= c_segment_pages_compressed_limit) || + (c_segment_count >= c_segments_limit)) { + return TRUE; + } +#endif /* CONFIG_FREEZE */ + return FALSE; } - + int vm_wants_task_throttled(task_t task) { - if (task == kernel_task) - return (0); + if (task == kernel_task) { + return 0; + } - if (vm_compressor_mode == COMPRESSED_PAGER_IS_ACTIVE || vm_compressor_mode == DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) { + if (VM_CONFIG_SWAP_IS_ACTIVE) { if ((vm_compressor_low_on_space() || HARD_THROTTLE_LIMIT_REACHED()) && - (unsigned int)pmap_compressed(task->map->pmap) > (c_segment_pages_compressed / 4)) - return (1); - } else { - if (((dp_pages_free + dp_pages_reserve < 2000) && VM_DYNAMIC_PAGING_ENABLED(memory_manager_default)) && - get_task_resident_size(task) > (((AVAILABLE_NON_COMPRESSED_MEMORY) * PAGE_SIZE) / 5)) - return (1); + (unsigned int)pmap_compressed(task->map->pmap) > (c_segment_pages_compressed / 4)) { + return 1; + } } - return (0); + return 0; } -void -vm_compressor_init_locks(void) +#if DEVELOPMENT || DEBUG +/* + * On compressor/swap exhaustion, kill the largest process regardless of + * its chosen process policy. + */ +TUNABLE(bool, kill_on_no_paging_space, "-kill_on_no_paging_space", false); +#endif /* DEVELOPMENT || DEBUG */ + +#if XNU_TARGET_OS_OSX + +static uint32_t no_paging_space_action_in_progress = 0; +extern void memorystatus_send_low_swap_note(void); + +static void +vm_compressor_take_paging_space_action(void) { - lck_grp_attr_setdefault(&vm_compressor_lck_grp_attr); - lck_grp_init(&vm_compressor_lck_grp, "vm_compressor", &vm_compressor_lck_grp_attr); - lck_attr_setdefault(&vm_compressor_lck_attr); + if (no_paging_space_action_in_progress == 0) { + if (OSCompareAndSwap(0, 1, (UInt32 *)&no_paging_space_action_in_progress)) { + if (no_paging_space_action()) { +#if DEVELOPMENT || DEBUG + if (kill_on_no_paging_space) { + /* + * Since we are choosing to always kill a process, we don't need the + * "out of application memory" dialog box in this mode. And, hence we won't + * send the knote. + */ + no_paging_space_action_in_progress = 0; + return; + } +#endif /* DEVELOPMENT || DEBUG */ + memorystatus_send_low_swap_note(); + } - lck_rw_init(&c_master_lock, &vm_compressor_lck_grp, &vm_compressor_lck_attr); + no_paging_space_action_in_progress = 0; + } + } } +#endif /* XNU_TARGET_OS_OSX */ void @@ -315,7 +519,7 @@ vm_decompressor_lock(void) PAGE_REPLACEMENT_ALLOWED(TRUE); decompressions_blocked = TRUE; - + PAGE_REPLACEMENT_ALLOWED(FALSE); } @@ -331,34 +535,100 @@ vm_decompressor_unlock(void) thread_wakeup((event_t)&decompressions_blocked); } +static inline void +cslot_copy(c_slot_t cdst, c_slot_t csrc) +{ +#if CHECKSUM_THE_DATA + cdst->c_hash_data = csrc->c_hash_data; +#endif +#if CHECKSUM_THE_COMPRESSED_DATA + cdst->c_hash_compressed_data = csrc->c_hash_compressed_data; +#endif +#if POPCOUNT_THE_COMPRESSED_DATA + cdst->c_pop_cdata = csrc->c_pop_cdata; +#endif + cdst->c_size = csrc->c_size; + cdst->c_packed_ptr = csrc->c_packed_ptr; +#if defined(__arm__) || defined(__arm64__) + cdst->c_codec = csrc->c_codec; +#endif +#if __ARM_WKDM_POPCNT__ + cdst->c_inline_popcount = csrc->c_inline_popcount; +#endif +} + +vm_map_t compressor_map; +uint64_t compressor_pool_max_size; +uint64_t compressor_pool_size; +uint32_t compressor_pool_multiplier; +#if DEVELOPMENT || DEBUG +/* + * Compressor segments are write-protected in development/debug + * kernels to help debug memory corruption. + * In cases where performance is a concern, this can be disabled + * via the boot-arg "-disable_cseg_write_protection". + */ +boolean_t write_protect_c_segs = TRUE; +int vm_compressor_test_seg_wp; +uint32_t vm_ktrace_enabled; +#endif /* DEVELOPMENT || DEBUG */ void vm_compressor_init(void) { - thread_t thread; - struct c_slot cs_dummy; - c_slot_t cs = &cs_dummy; - - /* - * ensure that any pointer that gets created from - * the vm_page zone can be packed properly - */ - cs->c_packed_ptr = C_SLOT_PACK_PTR(zone_map_min_address); - - if (C_SLOT_UNPACK_PTR(cs) != (uintptr_t)zone_map_min_address) - panic("C_SLOT_UNPACK_PTR failed on zone_map_min_address - %p", (void *)zone_map_min_address); - - cs->c_packed_ptr = C_SLOT_PACK_PTR(zone_map_max_address); + thread_t thread; + int attempts = 1; + kern_return_t retval = KERN_SUCCESS; + vm_offset_t start_addr = 0; + vm_size_t c_segments_arr_size = 0, compressor_submap_size = 0; + vm_map_kernel_flags_t vmk_flags; +#if RECORD_THE_COMPRESSED_DATA + vm_size_t c_compressed_record_sbuf_size = 0; +#endif /* RECORD_THE_COMPRESSED_DATA */ + +#if DEVELOPMENT || DEBUG || CONFIG_FREEZE + char bootarg_name[32]; +#endif /* DEVELOPMENT || DEBUG || CONFIG_FREEZE */ + +#if DEVELOPMENT || DEBUG + if (PE_parse_boot_argn("-disable_cseg_write_protection", bootarg_name, sizeof(bootarg_name))) { + write_protect_c_segs = FALSE; + } + int vmcval = 1; + PE_parse_boot_argn("vm_compressor_validation", &vmcval, sizeof(vmcval)); - if (C_SLOT_UNPACK_PTR(cs) != (uintptr_t)zone_map_max_address) - panic("C_SLOT_UNPACK_PTR failed on zone_map_max_address - %p", (void *)zone_map_max_address); + if (kern_feature_override(KF_COMPRSV_OVRD)) { + vmcval = 0; + } + if (vmcval == 0) { +#if POPCOUNT_THE_COMPRESSED_DATA + popcount_c_segs = FALSE; +#endif +#if CHECKSUM_THE_DATA || CHECKSUM_THE_COMPRESSED_DATA + checksum_c_segs = FALSE; +#endif +#if VALIDATE_C_SEGMENTS + validate_c_segs = FALSE; +#endif + write_protect_c_segs = FALSE; + } +#endif /* DEVELOPMENT || DEBUG */ +#if CONFIG_FREEZE + if (PE_parse_boot_argn("-disable_freezer_cseg_acct", bootarg_name, sizeof(bootarg_name))) { + freezer_incore_cseg_acct = FALSE; + } +#endif /* CONFIG_FREEZE */ assert((C_SEGMENTS_PER_PAGE * sizeof(union c_segu)) == PAGE_SIZE); - PE_parse_boot_argn("vm_compression_limit", &vm_compression_limit, sizeof (vm_compression_limit)); - +#if !XNU_TARGET_OS_OSX + vm_compressor_minorcompact_threshold_divisor = 20; + vm_compressor_majorcompact_threshold_divisor = 30; + vm_compressor_unthrottle_threshold_divisor = 40; + vm_compressor_catchup_threshold_divisor = 60; +#else /* !XNU_TARGET_OS_OSX */ if (max_mem <= (3ULL * 1024ULL * 1024ULL * 1024ULL)) { vm_compressor_minorcompact_threshold_divisor = 11; vm_compressor_majorcompact_threshold_divisor = 13; @@ -370,109 +640,281 @@ vm_compressor_init(void) vm_compressor_unthrottle_threshold_divisor = 35; vm_compressor_catchup_threshold_divisor = 50; } - /* - * vm_page_init_lck_grp is now responsible for calling vm_compressor_init_locks - * c_master_lock needs to be available early so that "vm_page_find_contiguous" can - * use PAGE_REPLACEMENT_ALLOWED to coordinate with the compressor. - */ +#endif /* !XNU_TARGET_OS_OSX */ -#if __i386__ || __x86_64__ - c_list_lock = lck_mtx_alloc_init(&vm_compressor_lck_grp, &vm_compressor_lck_attr); -#else /* __i386__ || __x86_64__ */ - c_list_lock = lck_spin_alloc_init(&vm_compressor_lck_grp, &vm_compressor_lck_attr); -#endif /* __i386__ || __x86_64__ */ - -#if TRACK_BAD_C_SEGMENTS queue_init(&c_bad_list_head); -#endif queue_init(&c_age_list_head); queue_init(&c_minor_list_head); + queue_init(&c_major_list_head); + queue_init(&c_filling_list_head); queue_init(&c_swapout_list_head); + queue_init(&c_swapio_list_head); queue_init(&c_swappedin_list_head); queue_init(&c_swappedout_list_head); queue_init(&c_swappedout_sparse_list_head); - compressor_segment_zone = zinit(sizeof (struct c_segment), - 128000 * sizeof (struct c_segment), - 8192, "compressor_segment"); - zone_change(compressor_segment_zone, Z_CALLERACCT, FALSE); - zone_change(compressor_segment_zone, Z_NOENCRYPT, TRUE); - - c_free_segno_head = -1; c_segments_available = 0; - if (vm_compression_limit == 0) { - c_segment_pages_compressed_limit = (uint32_t)((max_mem / PAGE_SIZE)) * vm_scale; + if (vm_compression_limit) { + compressor_pool_size = ptoa_64(vm_compression_limit); + } + + compressor_pool_max_size = C_SEG_MAX_LIMIT; + compressor_pool_max_size *= C_SEG_BUFSIZE; -#define OLD_SWAP_LIMIT (1024 * 1024 * 16) -#define MAX_SWAP_LIMIT (1024 * 1024 * 128) - - if (c_segment_pages_compressed_limit > (OLD_SWAP_LIMIT)) - c_segment_pages_compressed_limit = OLD_SWAP_LIMIT; +#if XNU_TARGET_OS_OSX - if (c_segment_pages_compressed_limit < (uint32_t)(max_mem / PAGE_SIZE_64)) - c_segment_pages_compressed_limit = (uint32_t)(max_mem / PAGE_SIZE_64); + if (vm_compression_limit == 0) { + if (max_mem <= (4ULL * 1024ULL * 1024ULL * 1024ULL)) { + compressor_pool_size = 16ULL * max_mem; + } else if (max_mem <= (8ULL * 1024ULL * 1024ULL * 1024ULL)) { + compressor_pool_size = 8ULL * max_mem; + } else if (max_mem <= (32ULL * 1024ULL * 1024ULL * 1024ULL)) { + compressor_pool_size = 4ULL * max_mem; + } else { + compressor_pool_size = 2ULL * max_mem; + } + } + if (max_mem <= (8ULL * 1024ULL * 1024ULL * 1024ULL)) { + compressor_pool_multiplier = 1; + } else if (max_mem <= (32ULL * 1024ULL * 1024ULL * 1024ULL)) { + compressor_pool_multiplier = 2; } else { - if (vm_compression_limit < MAX_SWAP_LIMIT) - c_segment_pages_compressed_limit = vm_compression_limit; - else - c_segment_pages_compressed_limit = MAX_SWAP_LIMIT; + compressor_pool_multiplier = 4; } - if ((c_segments_limit = c_segment_pages_compressed_limit / (C_SEG_BUFSIZE / PAGE_SIZE)) > C_SEG_MAX_LIMIT) - c_segments_limit = C_SEG_MAX_LIMIT; - c_segment_pages_compressed_nearing_limit = (c_segment_pages_compressed_limit * 98) / 100; - c_segments_nearing_limit = (c_segments_limit * 98) / 100; +#elif defined(__arm__) - compressor_kvwaste_limit = (vm_map_max(kernel_map) - vm_map_min(kernel_map)) / 16; +#define VM_RESERVE_SIZE (1024 * 1024 * 256) +#define MAX_COMPRESSOR_POOL_SIZE (1024 * 1024 * 450) - c_segments_busy = FALSE; + if (compressor_pool_max_size > MAX_COMPRESSOR_POOL_SIZE) { + compressor_pool_max_size = MAX_COMPRESSOR_POOL_SIZE; + } - if (kernel_memory_allocate(kernel_map, (vm_offset_t *)(&c_segments), (sizeof(union c_segu) * c_segments_limit), 0, KMA_KOBJECT | KMA_VAONLY) != KERN_SUCCESS) - panic("vm_compressor_init: kernel_memory_allocate failed\n"); + if (vm_compression_limit == 0) { + compressor_pool_size = ((kernel_map->max_offset - kernel_map->min_offset) - kernel_map->size) - VM_RESERVE_SIZE; + } + compressor_pool_multiplier = 1; + +#elif defined(__arm64__) && defined(XNU_TARGET_OS_WATCH) + + /* + * On M9 watches the compressor can become big and can lead to + * churn in workingset resulting in audio drops. Setting a cap + * on the compressor size favors reclaiming unused memory + * sitting in idle band via jetsams + */ + +#define COMPRESSOR_CAP_PERCENTAGE 37ULL + + if (compressor_pool_max_size > max_mem) { + compressor_pool_max_size = max_mem; + } + + if (vm_compression_limit == 0) { + compressor_pool_size = (max_mem * COMPRESSOR_CAP_PERCENTAGE) / 100ULL; + } + compressor_pool_multiplier = 1; + +#else + + if (compressor_pool_max_size > max_mem) { + compressor_pool_max_size = max_mem; + } + + if (vm_compression_limit == 0) { + compressor_pool_size = max_mem; + } + compressor_pool_multiplier = 1; +#endif + if (compressor_pool_size > compressor_pool_max_size) { + compressor_pool_size = compressor_pool_max_size; + } + +try_again: + c_segments_limit = (uint32_t)(compressor_pool_size / (vm_size_t)(C_SEG_ALLOCSIZE)); + c_segments_nearing_limit = (uint32_t)(((uint64_t)c_segments_limit * 98ULL) / 100ULL); + + c_segment_pages_compressed_limit = (c_segments_limit * (C_SEG_BUFSIZE / PAGE_SIZE) * compressor_pool_multiplier); + + if (c_segment_pages_compressed_limit < (uint32_t)(max_mem / PAGE_SIZE)) { + if (!vm_compression_limit) { + c_segment_pages_compressed_limit = (uint32_t)(max_mem / PAGE_SIZE); + } + } + + c_segment_pages_compressed_nearing_limit = (uint32_t)(((uint64_t)c_segment_pages_compressed_limit * 98ULL) / 100ULL); + +#if CONFIG_FREEZE + /* + * Our in-core limits are based on the size of the compressor pool. + * The c_segments_nearing_limit is also based on the compressor pool + * size and calculated above. + */ + c_segments_incore_limit = c_segments_limit; + + if (freezer_incore_cseg_acct) { + /* + * Add enough segments to track all frozen c_segs that can be stored in swap. + */ + c_segments_limit += (uint32_t)(vm_swap_get_max_configured_space() / (vm_size_t)(C_SEG_ALLOCSIZE)); + } +#endif + /* + * Submap needs space for: + * - c_segments + * - c_buffers + * - swap reclaimations -- C_SEG_BUFSIZE + */ + c_segments_arr_size = vm_map_round_page((sizeof(union c_segu) * c_segments_limit), VM_MAP_PAGE_MASK(kernel_map)); + c_buffers_size = vm_map_round_page(((vm_size_t)C_SEG_ALLOCSIZE * (vm_size_t)c_segments_limit), VM_MAP_PAGE_MASK(kernel_map)); + + compressor_submap_size = c_segments_arr_size + c_buffers_size + C_SEG_BUFSIZE; + +#if RECORD_THE_COMPRESSED_DATA + c_compressed_record_sbuf_size = (vm_size_t)C_SEG_ALLOCSIZE + (PAGE_SIZE * 2); + compressor_submap_size += c_compressed_record_sbuf_size; +#endif /* RECORD_THE_COMPRESSED_DATA */ + + vmk_flags = VM_MAP_KERNEL_FLAGS_NONE; + vmk_flags.vmkf_permanent = TRUE; + retval = kmem_suballoc(kernel_map, &start_addr, compressor_submap_size, + FALSE, VM_FLAGS_ANYWHERE, vmk_flags, VM_KERN_MEMORY_COMPRESSOR, + &compressor_map); + + if (retval != KERN_SUCCESS) { + if (++attempts > 3) { + panic("vm_compressor_init: kmem_suballoc failed - 0x%llx", (uint64_t)compressor_submap_size); + } + + compressor_pool_size = compressor_pool_size / 2; + + kprintf("retrying creation of the compressor submap at 0x%llx bytes\n", compressor_pool_size); + goto try_again; + } + if (kernel_memory_allocate(compressor_map, (vm_offset_t *)(&c_segments), + (sizeof(union c_segu) * c_segments_limit), 0, + KMA_KOBJECT | KMA_VAONLY | KMA_PERMANENT, VM_KERN_MEMORY_COMPRESSOR) != KERN_SUCCESS) { + panic("vm_compressor_init: kernel_memory_allocate failed - c_segments\n"); + } + if (kernel_memory_allocate(compressor_map, &c_buffers, c_buffers_size, 0, + KMA_COMPRESSOR | KMA_VAONLY | KMA_PERMANENT, VM_KERN_MEMORY_COMPRESSOR) != KERN_SUCCESS) { + panic("vm_compressor_init: kernel_memory_allocate failed - c_buffers\n"); + } + + + /* + * Pick a good size that will minimize fragmentation in zalloc + * by minimizing the fragmentation in a 16k run. + * + * C_SEG_SLOT_VAR_ARRAY_MIN_LEN is larger on 4k systems than 16k ones, + * making the fragmentation in a 4k page terrible. Using 16k for all + * systems matches zalloc() and will minimize fragmentation. + */ + uint32_t c_segment_size = sizeof(struct c_segment) + (C_SEG_SLOT_VAR_ARRAY_MIN_LEN * sizeof(struct c_slot)); + uint32_t cnt = (16 << 10) / c_segment_size; + uint32_t frag = (16 << 10) % c_segment_size; + + c_seg_fixed_array_len = C_SEG_SLOT_VAR_ARRAY_MIN_LEN; + + while (cnt * sizeof(struct c_slot) < frag) { + c_segment_size += sizeof(struct c_slot); + c_seg_fixed_array_len++; + frag -= cnt * sizeof(struct c_slot); + } + + compressor_segment_zone = zone_create("compressor_segment", + c_segment_size, ZC_NOENCRYPT | ZC_ZFREE_CLEARMEM); + + c_segments_busy = FALSE; c_segments_next_page = (caddr_t)c_segments; + vm_compressor_algorithm_init(); { host_basic_info_data_t hinfo; mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT; + size_t bufsize; + char *buf; #define BSD_HOST 1 host_info((host_t)BSD_HOST, HOST_BASIC_INFO, (host_info_t)&hinfo, &count); compressor_cpus = hinfo.max_cpus; - compressor_scratch_bufs = kalloc(compressor_cpus * WKdm_SCRATCH_BUF_SIZE); + bufsize = PAGE_SIZE; + bufsize += compressor_cpus * vm_compressor_get_decode_scratch_size(); + bufsize += vm_compressor_get_decode_scratch_size(); +#if CONFIG_FREEZE + bufsize += vm_compressor_get_encode_scratch_size(); +#endif +#if RECORD_THE_COMPRESSED_DATA + bufsize += c_compressed_record_sbuf_size; +#endif + + if (kernel_memory_allocate(kernel_map, (vm_offset_t *)&buf, bufsize, + PAGE_MASK, KMA_KOBJECT | KMA_PERMANENT, VM_KERN_MEMORY_COMPRESSOR)) { + panic("vm_compressor_init: Unable to allocate %zd bytes", bufsize); + } + + /* + * kdp_compressor_decompressed_page must be page aligned because we access + * it through the physical apperture by page number. + */ + kdp_compressor_decompressed_page = buf; + kdp_compressor_decompressed_page_paddr = kvtophys((vm_offset_t)kdp_compressor_decompressed_page); + kdp_compressor_decompressed_page_ppnum = (ppnum_t) atop(kdp_compressor_decompressed_page_paddr); + buf += PAGE_SIZE; + bufsize -= PAGE_SIZE; + + compressor_scratch_bufs = buf; + buf += compressor_cpus * vm_compressor_get_decode_scratch_size(); + bufsize -= compressor_cpus * vm_compressor_get_decode_scratch_size(); + + kdp_compressor_scratch_buf = buf; + buf += vm_compressor_get_decode_scratch_size(); + bufsize -= vm_compressor_get_decode_scratch_size(); + +#if CONFIG_FREEZE + freezer_context_global.freezer_ctx_compressor_scratch_buf = buf; + buf += vm_compressor_get_encode_scratch_size(); + bufsize -= vm_compressor_get_encode_scratch_size(); +#endif + +#if RECORD_THE_COMPRESSED_DATA + c_compressed_record_sbuf = buf; + c_compressed_record_cptr = buf; + c_compressed_record_ebuf = c_compressed_record_sbuf + c_compressed_record_sbuf_size; + buf += c_compressed_record_sbuf_size; + bufsize -= c_compressed_record_sbuf_size; +#endif + assert(bufsize == 0); } if (kernel_thread_start_priority((thread_continue_t)vm_compressor_swap_trigger_thread, NULL, - BASEPRI_PREEMPT - 1, &thread) != KERN_SUCCESS) { + BASEPRI_VM, &thread) != KERN_SUCCESS) { panic("vm_compressor_swap_trigger_thread: create failed"); } - thread->options |= TH_OPT_VMPRIV; - thread_deallocate(thread); - assert(default_pager_init_flag == 0); - if (vm_pageout_internal_start() != KERN_SUCCESS) { panic("vm_compressor_init: Failed to start the internal pageout thread.\n"); } - - if ((vm_compressor_mode == VM_PAGER_COMPRESSOR_WITH_SWAP) || - (vm_compressor_mode == VM_PAGER_FREEZER_COMPRESSOR_WITH_SWAP)) { + if (VM_CONFIG_SWAP_IS_PRESENT) { vm_compressor_swap_init(); } - if (COMPRESSED_PAGER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) + if (VM_CONFIG_COMPRESSOR_IS_ACTIVE) { vm_compressor_is_active = 1; + } #if CONFIG_FREEZE memorystatus_freeze_enabled = TRUE; #endif /* CONFIG_FREEZE */ - default_pager_init_flag = 1; + vm_compressor_available = 1; vm_page_reactivate_all_throttled(); } @@ -483,28 +925,30 @@ vm_compressor_init(void) static void c_seg_validate(c_segment_t c_seg, boolean_t must_be_compact) { - int c_indx; - int32_t bytes_used; - int32_t bytes_unused; - uint32_t c_rounded_size; - uint32_t c_size; - c_slot_t cs; + uint16_t c_indx; + int32_t bytes_used; + uint32_t c_rounded_size; + uint32_t c_size; + c_slot_t cs; + if (__probable(validate_c_segs == FALSE)) { + return; + } if (c_seg->c_firstemptyslot < c_seg->c_nextslot) { c_indx = c_seg->c_firstemptyslot; cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx); - if (cs == NULL) + if (cs == NULL) { panic("c_seg_validate: no slot backing c_firstemptyslot"); - - if (cs->c_size) + } + + if (cs->c_size) { panic("c_seg_validate: c_firstemptyslot has non-zero size (%d)\n", cs->c_size); + } } bytes_used = 0; - bytes_unused = 0; for (c_indx = 0; c_indx < c_seg->c_nextslot; c_indx++) { - cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx); c_size = UNPACK_C_SIZE(cs); @@ -514,22 +958,37 @@ c_seg_validate(c_segment_t c_seg, boolean_t must_be_compact) bytes_used += c_rounded_size; #if CHECKSUM_THE_COMPRESSED_DATA - if (c_size && cs->c_hash_compressed_data != hash_string((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size)) - panic("compressed data doesn't match original"); + unsigned csvhash; + if (c_size && cs->c_hash_compressed_data != (csvhash = vmc_hash((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size))) { + addr64_t csvphys = kvtophys((vm_offset_t)&c_seg->c_store.c_buffer[cs->c_offset]); + panic("Compressed data doesn't match original %p phys: 0x%llx %d %p %d %d 0x%x 0x%x", c_seg, csvphys, cs->c_offset, cs, c_indx, c_size, cs->c_hash_compressed_data, csvhash); + } +#endif +#if POPCOUNT_THE_COMPRESSED_DATA + unsigned csvpop; + if (c_size) { + uintptr_t csvaddr = (uintptr_t) &c_seg->c_store.c_buffer[cs->c_offset]; + if (cs->c_pop_cdata != (csvpop = vmc_pop(csvaddr, c_size))) { + panic("Compressed data popcount doesn't match original, bit distance: %d %p (phys: %p) %p %p 0x%llx 0x%x 0x%x 0x%x", (csvpop - cs->c_pop_cdata), (void *)csvaddr, (void *) kvtophys(csvaddr), c_seg, cs, (uint64_t)cs->c_offset, c_size, csvpop, cs->c_pop_cdata); + } + } #endif } - if (bytes_used != c_seg->c_bytes_used) + if (bytes_used != c_seg->c_bytes_used) { panic("c_seg_validate: bytes_used mismatch - found %d, segment has %d\n", bytes_used, c_seg->c_bytes_used); + } - if (c_seg->c_bytes_used > C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset)) + if (c_seg->c_bytes_used > C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset)) { panic("c_seg_validate: c_bytes_used > c_nextoffset - c_nextoffset = %d, c_bytes_used = %d\n", - (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used); + (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used); + } if (must_be_compact) { - if (c_seg->c_bytes_used != C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset)) + if (c_seg->c_bytes_used != C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset)) { panic("c_seg_validate: c_bytes_used doesn't match c_nextoffset - c_nextoffset = %d, c_bytes_used = %d\n", - (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used); + (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used); + } } } @@ -537,28 +996,35 @@ c_seg_validate(c_segment_t c_seg, boolean_t must_be_compact) void -c_seg_need_delayed_compaction(c_segment_t c_seg) +c_seg_need_delayed_compaction(c_segment_t c_seg, boolean_t c_list_lock_held) { - boolean_t clear_busy = FALSE; + boolean_t clear_busy = FALSE; - if ( !lck_mtx_try_lock_spin_always(c_list_lock)) { - C_SEG_BUSY(c_seg); - - lck_mtx_unlock_always(&c_seg->c_lock); - lck_mtx_lock_spin_always(c_list_lock); - lck_mtx_lock_spin_always(&c_seg->c_lock); + if (c_list_lock_held == FALSE) { + if (!lck_mtx_try_lock_spin_always(c_list_lock)) { + C_SEG_BUSY(c_seg); - clear_busy = TRUE; + lck_mtx_unlock_always(&c_seg->c_lock); + lck_mtx_lock_spin_always(c_list_lock); + lck_mtx_lock_spin_always(&c_seg->c_lock); + + clear_busy = TRUE; + } } - if (!c_seg->c_on_minorcompact_q && !c_seg->c_ondisk && !c_seg->c_on_swapout_q) { + assert(c_seg->c_state != C_IS_FILLING); + + if (!c_seg->c_on_minorcompact_q && !(C_SEG_IS_ON_DISK_OR_SOQ(c_seg))) { queue_enter(&c_minor_list_head, c_seg, c_segment_t, c_list); c_seg->c_on_minorcompact_q = 1; c_minor_count++; } - lck_mtx_unlock_always(c_list_lock); - - if (clear_busy == TRUE) + if (c_list_lock_held == FALSE) { + lck_mtx_unlock_always(c_list_lock); + } + + if (clear_busy == TRUE) { C_SEG_WAKEUP_DONE(c_seg); + } } @@ -567,35 +1033,26 @@ unsigned int c_seg_moved_to_sparse_list = 0; void c_seg_move_to_sparse_list(c_segment_t c_seg) { - boolean_t clear_busy = FALSE; + boolean_t clear_busy = FALSE; - if ( !lck_mtx_try_lock_spin_always(c_list_lock)) { + if (!lck_mtx_try_lock_spin_always(c_list_lock)) { C_SEG_BUSY(c_seg); lck_mtx_unlock_always(&c_seg->c_lock); lck_mtx_lock_spin_always(c_list_lock); lck_mtx_lock_spin_always(&c_seg->c_lock); - + clear_busy = TRUE; } - assert(c_seg->c_ondisk); - assert(c_seg->c_on_swappedout_q); - assert(!c_seg->c_on_swappedout_sparse_q); - - queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedout_q = 0; - c_swappedout_count--; - - c_seg_insert_into_q(&c_swappedout_sparse_list_head, c_seg); - c_seg->c_on_swappedout_sparse_q = 1; - c_swappedout_sparse_count++; + c_seg_switch_state(c_seg, C_ON_SWAPPEDOUTSPARSE_Q, FALSE); c_seg_moved_to_sparse_list++; lck_mtx_unlock_always(c_list_lock); - if (clear_busy == TRUE) + if (clear_busy == TRUE) { C_SEG_WAKEUP_DONE(c_seg); + } } @@ -610,13 +1067,12 @@ c_seg_insert_into_q(queue_head_t *qhead, c_segment_t c_seg) c_seg_next = (c_segment_t)queue_first(qhead); while (TRUE) { - if (c_seg->c_generation_id < c_seg_next->c_generation_id) { queue_insert_before(qhead, c_seg, c_seg_next, c_segment_t, c_age_list); break; } c_seg_next = (c_segment_t) queue_next(&c_seg_next->c_age_list); - + if (queue_end(qhead, (queue_entry_t) c_seg_next)) { queue_enter(qhead, c_seg, c_segment_t, c_age_list); break; @@ -632,7 +1088,6 @@ int try_minor_compaction_succeeded = 0; void c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg) { - assert(c_seg->c_on_minorcompact_q); /* * c_seg is currently on the delayed minor compaction @@ -641,7 +1096,7 @@ c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg) * because the lock order is c_list_lock then c_seg's lock) * we'll pull it from the delayed list and free it directly */ - if ( !lck_mtx_try_lock_spin_always(c_list_lock)) { + if (!lck_mtx_try_lock_spin_always(c_list_lock)) { /* * c_list_lock is held, we need to bail */ @@ -660,22 +1115,36 @@ c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg) int c_seg_do_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy, boolean_t need_list_lock, boolean_t disallow_page_replacement) { - int c_seg_freed; + int c_seg_freed; assert(c_seg->c_busy); + assert(!C_SEG_IS_ON_DISK_OR_SOQ(c_seg)); + /* + * check for the case that can occur when we are not swapping + * and this segment has been major compacted in the past + * and moved to the majorcompact q to remove it from further + * consideration... if the occupancy falls too low we need + * to put it back on the age_q so that it will be considered + * in the next major compaction sweep... if we don't do this + * we will eventually run into the c_segments_limit + */ + if (c_seg->c_state == C_ON_MAJORCOMPACT_Q && C_SEG_SHOULD_MAJORCOMPACT_NOW(c_seg)) { + c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE); + } if (!c_seg->c_on_minorcompact_q) { - if (clear_busy == TRUE) + if (clear_busy == TRUE) { C_SEG_WAKEUP_DONE(c_seg); + } lck_mtx_unlock_always(&c_seg->c_lock); - return (0); + return 0; } queue_remove(&c_minor_list_head, c_seg, c_segment_t, c_list); c_seg->c_on_minorcompact_q = 0; c_minor_count--; - + lck_mtx_unlock_always(c_list_lock); if (disallow_page_replacement == TRUE) { @@ -687,199 +1156,495 @@ c_seg_do_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy, bo } c_seg_freed = c_seg_minor_compaction_and_unlock(c_seg, clear_busy); - if (disallow_page_replacement == TRUE) + if (disallow_page_replacement == TRUE) { PAGE_REPLACEMENT_DISALLOWED(FALSE); + } - if (need_list_lock == TRUE) + if (need_list_lock == TRUE) { lck_mtx_lock_spin_always(c_list_lock); + } - return (c_seg_freed); + return c_seg_freed; } - void -c_seg_wait_on_busy(c_segment_t c_seg) +kdp_compressor_busy_find_owner(event64_t wait_event, thread_waitinfo_t *waitinfo) { - c_seg->c_wanted = 1; - assert_wait((event_t) (c_seg), THREAD_UNINT); + c_segment_t c_seg = (c_segment_t) wait_event; - lck_mtx_unlock_always(&c_seg->c_lock); - thread_block(THREAD_CONTINUE_NULL); + waitinfo->owner = thread_tid(c_seg->c_busy_for_thread); + waitinfo->context = VM_KERNEL_UNSLIDE_OR_PERM(c_seg); } +#if DEVELOPMENT || DEBUG +int +do_cseg_wedge_thread(void) +{ + struct c_segment c_seg; + c_seg.c_busy_for_thread = current_thread(); + debug_cseg_wait_event = (event_t) &c_seg; -int try_free_succeeded = 0; -int try_free_failed = 0; + thread_set_pending_block_hint(current_thread(), kThreadWaitCompressor); + assert_wait((event_t) (&c_seg), THREAD_INTERRUPTIBLE); -boolean_t -c_seg_try_free(c_segment_t c_seg) + thread_block(THREAD_CONTINUE_NULL); + + return 0; +} + +int +do_cseg_unwedge_thread(void) { - /* - * c_seg is currently on the delayed minor compaction - * or the spapped out sparse queue and we have c_seg locked... - * if we can get the c_list_lock w/o blocking (if we blocked we - * could deadlock because the lock order is c_list_lock then c_seg's lock) - * we'll pull it from the appropriate queue and free it - */ - if ( !lck_mtx_try_lock_spin_always(c_list_lock)) { - /* - * c_list_lock is held, we need to bail - */ - try_free_failed++; - return (FALSE); - } - if (c_seg->c_on_minorcompact_q) { - queue_remove(&c_minor_list_head, c_seg, c_segment_t, c_list); - c_seg->c_on_minorcompact_q = 0; - c_minor_count--; - } else { - assert(c_seg->c_on_swappedout_sparse_q); + thread_wakeup(debug_cseg_wait_event); + debug_cseg_wait_event = NULL; - /* - * c_seg_free_locked will remove it from the swappedout sparse list - */ - } - if (!c_seg->c_busy_swapping) - C_SEG_BUSY(c_seg); + return 0; +} +#endif /* DEVELOPMENT || DEBUG */ - c_seg_free_locked(c_seg); +void +c_seg_wait_on_busy(c_segment_t c_seg) +{ + c_seg->c_wanted = 1; - try_free_succeeded++; + thread_set_pending_block_hint(current_thread(), kThreadWaitCompressor); + assert_wait((event_t) (c_seg), THREAD_UNINT); - return (TRUE); + lck_mtx_unlock_always(&c_seg->c_lock); + thread_block(THREAD_CONTINUE_NULL); } +#if CONFIG_FREEZE +/* + * We don't have the task lock held while updating the task's + * c_seg queues. We can do that because of the following restrictions: + * + * - SINGLE FREEZER CONTEXT: + * We 'insert' c_segs into the task list on the task_freeze path. + * There can only be one such freeze in progress and the task + * isn't disappearing because we have the VM map lock held throughout + * and we have a reference on the proc too. + * + * - SINGLE TASK DISOWN CONTEXT: + * We 'disown' c_segs of a task ONLY from the task_terminate context. So + * we don't need the task lock but we need the c_list_lock and the + * compressor master lock (shared). We also hold the individual + * c_seg locks (exclusive). + * + * If we either: + * - can't get the c_seg lock on a try, then we start again because maybe + * the c_seg is part of a compaction and might get freed. So we can't trust + * that linkage and need to restart our queue traversal. + * - OR, we run into a busy c_seg (say being swapped in or free-ing) we + * drop all locks again and wait and restart our queue traversal. + * + * - The new_owner_task below is currently only the kernel or NULL. + * + */ +void +c_seg_update_task_owner(c_segment_t c_seg, task_t new_owner_task) +{ + task_t owner_task = c_seg->c_task_owner; + uint64_t uncompressed_bytes = ((c_seg->c_slots_used) * PAGE_SIZE_64); + + LCK_MTX_ASSERT(c_list_lock, LCK_MTX_ASSERT_OWNED); + LCK_MTX_ASSERT(&c_seg->c_lock, LCK_MTX_ASSERT_OWNED); + + if (owner_task) { + task_update_frozen_to_swap_acct(owner_task, uncompressed_bytes, DEBIT_FROM_SWAP); + queue_remove(&owner_task->task_frozen_cseg_q, c_seg, + c_segment_t, c_task_list_next_cseg); + } + + if (new_owner_task) { + queue_enter(&new_owner_task->task_frozen_cseg_q, c_seg, + c_segment_t, c_task_list_next_cseg); + task_update_frozen_to_swap_acct(new_owner_task, uncompressed_bytes, CREDIT_TO_SWAP); + } + + c_seg->c_task_owner = new_owner_task; +} void -c_seg_free(c_segment_t c_seg) +task_disown_frozen_csegs(task_t owner_task) { - assert(c_seg->c_busy); + c_segment_t c_seg = NULL, next_cseg = NULL; - lck_mtx_unlock_always(&c_seg->c_lock); +again: + PAGE_REPLACEMENT_DISALLOWED(TRUE); lck_mtx_lock_spin_always(c_list_lock); - lck_mtx_lock_spin_always(&c_seg->c_lock); - c_seg_free_locked(c_seg); -} + for (c_seg = (c_segment_t) queue_first(&owner_task->task_frozen_cseg_q); + !queue_end(&owner_task->task_frozen_cseg_q, (queue_entry_t) c_seg); + c_seg = next_cseg) { + next_cseg = (c_segment_t) queue_next(&c_seg->c_task_list_next_cseg);; + + if (!lck_mtx_try_lock_spin_always(&c_seg->c_lock)) { + lck_mtx_unlock(c_list_lock); + PAGE_REPLACEMENT_DISALLOWED(FALSE); + goto again; + } + + if (c_seg->c_busy) { + lck_mtx_unlock(c_list_lock); + PAGE_REPLACEMENT_DISALLOWED(FALSE); + + c_seg_wait_on_busy(c_seg); + + goto again; + } + assert(c_seg->c_task_owner == owner_task); + c_seg_update_task_owner(c_seg, kernel_task); + lck_mtx_unlock_always(&c_seg->c_lock); + } + lck_mtx_unlock(c_list_lock); + PAGE_REPLACEMENT_DISALLOWED(FALSE); +} +#endif /* CONFIG_FREEZE */ void -c_seg_free_locked(c_segment_t c_seg) +c_seg_switch_state(c_segment_t c_seg, int new_state, boolean_t insert_head) { - int segno, i; - int pages_populated = 0; - int32_t *c_buffer = NULL; - uint64_t c_swap_handle = 0; + int old_state = c_seg->c_state; - assert(!c_seg->c_on_minorcompact_q); +#if XNU_TARGET_OS_OSX +#if DEVELOPMENT || DEBUG + if (new_state != C_IS_FILLING) { + LCK_MTX_ASSERT(&c_seg->c_lock, LCK_MTX_ASSERT_OWNED); + } + LCK_MTX_ASSERT(c_list_lock, LCK_MTX_ASSERT_OWNED); +#endif +#endif /* XNU_TARGET_OS_OSX */ + switch (old_state) { + case C_IS_EMPTY: + assert(new_state == C_IS_FILLING || new_state == C_IS_FREE); + + c_empty_count--; + break; + + case C_IS_FILLING: + assert(new_state == C_ON_AGE_Q || new_state == C_ON_SWAPOUT_Q); + + queue_remove(&c_filling_list_head, c_seg, c_segment_t, c_age_list); + c_filling_count--; + break; + + case C_ON_AGE_Q: + assert(new_state == C_ON_SWAPOUT_Q || new_state == C_ON_MAJORCOMPACT_Q || + new_state == C_IS_FREE); - if (c_seg->c_on_age_q) { queue_remove(&c_age_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_age_q = 0; c_age_count--; - } else if (c_seg->c_on_swappedin_q) { + break; + + case C_ON_SWAPPEDIN_Q: + assert(new_state == C_ON_AGE_Q || new_state == C_IS_FREE); + queue_remove(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedin_q = 0; c_swappedin_count--; - } else if (c_seg->c_on_swapout_q) { + break; + + case C_ON_SWAPOUT_Q: + assert(new_state == C_ON_AGE_Q || new_state == C_IS_FREE || new_state == C_IS_EMPTY || new_state == C_ON_SWAPIO_Q); + +#if CONFIG_FREEZE + if (c_seg->c_task_owner && (new_state != C_ON_SWAPIO_Q)) { + c_seg_update_task_owner(c_seg, NULL); + } +#endif /* CONFIG_FREEZE */ + queue_remove(&c_swapout_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swapout_q = 0; - c_swapout_count--; thread_wakeup((event_t)&compaction_swapper_running); - } else if (c_seg->c_on_swappedout_q) { + c_swapout_count--; + break; + + case C_ON_SWAPIO_Q: + assert(new_state == C_ON_SWAPPEDOUT_Q || new_state == C_ON_SWAPPEDOUTSPARSE_Q || new_state == C_ON_AGE_Q); + + queue_remove(&c_swapio_list_head, c_seg, c_segment_t, c_age_list); + c_swapio_count--; + break; + + case C_ON_SWAPPEDOUT_Q: + assert(new_state == C_ON_SWAPPEDIN_Q || new_state == C_ON_AGE_Q || + new_state == C_ON_SWAPPEDOUTSPARSE_Q || + new_state == C_ON_BAD_Q || new_state == C_IS_EMPTY || new_state == C_IS_FREE); + queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedout_q = 0; c_swappedout_count--; - } else if (c_seg->c_on_swappedout_sparse_q) { + break; + + case C_ON_SWAPPEDOUTSPARSE_Q: + assert(new_state == C_ON_SWAPPEDIN_Q || new_state == C_ON_AGE_Q || + new_state == C_ON_BAD_Q || new_state == C_IS_EMPTY || new_state == C_IS_FREE); + queue_remove(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedout_sparse_q = 0; c_swappedout_sparse_count--; - } -#if TRACK_BAD_C_SEGMENTS - else if (c_seg->c_on_bad_q) { + break; + + case C_ON_MAJORCOMPACT_Q: + assert(new_state == C_ON_AGE_Q || new_state == C_IS_FREE); + + queue_remove(&c_major_list_head, c_seg, c_segment_t, c_age_list); + c_major_count--; + break; + + case C_ON_BAD_Q: + assert(new_state == C_IS_FREE); + queue_remove(&c_bad_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_bad_q = 0; c_bad_count--; + break; + + default: + panic("c_seg %p has bad c_state = %d\n", c_seg, old_state); } -#endif - segno = c_seg->c_mysegno; - c_segments[segno].c_segno = c_free_segno_head; - c_free_segno_head = segno; - c_segment_count--; - lck_mtx_unlock_always(c_list_lock); + switch (new_state) { + case C_IS_FREE: + assert(old_state != C_IS_FILLING); - if (c_seg->c_wanted) { - thread_wakeup((event_t) (c_seg)); - c_seg->c_wanted = 0; - } - if (c_seg->c_busy_swapping) { - c_seg->c_must_free = 1; + break; - lck_mtx_unlock_always(&c_seg->c_lock); - return; + case C_IS_EMPTY: + assert(old_state == C_ON_SWAPOUT_Q || old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q); + + c_empty_count++; + break; + + case C_IS_FILLING: + assert(old_state == C_IS_EMPTY); + + queue_enter(&c_filling_list_head, c_seg, c_segment_t, c_age_list); + c_filling_count++; + break; + + case C_ON_AGE_Q: + assert(old_state == C_IS_FILLING || old_state == C_ON_SWAPPEDIN_Q || + old_state == C_ON_SWAPOUT_Q || old_state == C_ON_SWAPIO_Q || + old_state == C_ON_MAJORCOMPACT_Q || old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q); + + if (old_state == C_IS_FILLING) { + queue_enter(&c_age_list_head, c_seg, c_segment_t, c_age_list); + } else { + if (!queue_empty(&c_age_list_head)) { + c_segment_t c_first; + + c_first = (c_segment_t)queue_first(&c_age_list_head); + c_seg->c_creation_ts = c_first->c_creation_ts; + } + queue_enter_first(&c_age_list_head, c_seg, c_segment_t, c_age_list); + } + c_age_count++; + break; + + case C_ON_SWAPPEDIN_Q: + assert(old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q); + + if (insert_head == TRUE) { + queue_enter_first(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list); + } + c_swappedin_count++; + break; + + case C_ON_SWAPOUT_Q: + assert(old_state == C_ON_AGE_Q || old_state == C_IS_FILLING); + + if (insert_head == TRUE) { + queue_enter_first(&c_swapout_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_swapout_list_head, c_seg, c_segment_t, c_age_list); + } + c_swapout_count++; + break; + + case C_ON_SWAPIO_Q: + assert(old_state == C_ON_SWAPOUT_Q); + + if (insert_head == TRUE) { + queue_enter_first(&c_swapio_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_swapio_list_head, c_seg, c_segment_t, c_age_list); + } + c_swapio_count++; + break; + + case C_ON_SWAPPEDOUT_Q: + assert(old_state == C_ON_SWAPIO_Q); + + if (insert_head == TRUE) { + queue_enter_first(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list); + } + c_swappedout_count++; + break; + + case C_ON_SWAPPEDOUTSPARSE_Q: + assert(old_state == C_ON_SWAPIO_Q || old_state == C_ON_SWAPPEDOUT_Q); + + if (insert_head == TRUE) { + queue_enter_first(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list); + } + + c_swappedout_sparse_count++; + break; + + case C_ON_MAJORCOMPACT_Q: + assert(old_state == C_ON_AGE_Q); + + if (insert_head == TRUE) { + queue_enter_first(&c_major_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_major_list_head, c_seg, c_segment_t, c_age_list); + } + c_major_count++; + break; + + case C_ON_BAD_Q: + assert(old_state == C_ON_SWAPPEDOUT_Q || old_state == C_ON_SWAPPEDOUTSPARSE_Q); + + if (insert_head == TRUE) { + queue_enter_first(&c_bad_list_head, c_seg, c_segment_t, c_age_list); + } else { + queue_enter(&c_bad_list_head, c_seg, c_segment_t, c_age_list); + } + c_bad_count++; + break; + + default: + panic("c_seg %p requesting bad c_state = %d\n", c_seg, new_state); } - if (c_seg->c_ondisk == 0) { - pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE; + c_seg->c_state = new_state; +} + + + +void +c_seg_free(c_segment_t c_seg) +{ + assert(c_seg->c_busy); + + lck_mtx_unlock_always(&c_seg->c_lock); + lck_mtx_lock_spin_always(c_list_lock); + lck_mtx_lock_spin_always(&c_seg->c_lock); + + c_seg_free_locked(c_seg); +} + + +void +c_seg_free_locked(c_segment_t c_seg) +{ + int segno; + int pages_populated = 0; + int32_t *c_buffer = NULL; + uint64_t c_swap_handle = 0; + + assert(c_seg->c_busy); + assert(c_seg->c_slots_used == 0); + assert(!c_seg->c_on_minorcompact_q); + assert(!c_seg->c_busy_swapping); + if (c_seg->c_overage_swap == TRUE) { + c_overage_swapped_count--; + c_seg->c_overage_swap = FALSE; + } + if (!(C_SEG_IS_ONDISK(c_seg))) { c_buffer = c_seg->c_store.c_buffer; - c_seg->c_store.c_buffer = NULL; } else { - /* - * Free swap space on disk. - */ c_swap_handle = c_seg->c_store.c_swap_handle; + } + + c_seg_switch_state(c_seg, C_IS_FREE, FALSE); + + if (c_buffer) { + pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE; + c_seg->c_store.c_buffer = NULL; + } else { +#if CONFIG_FREEZE + c_seg_update_task_owner(c_seg, NULL); +#endif /* CONFIG_FREEZE */ + c_seg->c_store.c_swap_handle = (uint64_t)-1; } + lck_mtx_unlock_always(&c_seg->c_lock); + lck_mtx_unlock_always(c_list_lock); + if (c_buffer) { - if (pages_populated) - kernel_memory_depopulate(kernel_map, (vm_offset_t) c_buffer, pages_populated * PAGE_SIZE, KMA_COMPRESSOR); + if (pages_populated) { + kernel_memory_depopulate(compressor_map, (vm_offset_t)c_buffer, + pages_populated * PAGE_SIZE, KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR); + } + } else if (c_swap_handle) { + /* + * Free swap space on disk. + */ + vm_swap_free(c_swap_handle); + } + lck_mtx_lock_spin_always(&c_seg->c_lock); + /* + * c_seg must remain busy until + * after the call to vm_swap_free + */ + C_SEG_WAKEUP_DONE(c_seg); + lck_mtx_unlock_always(&c_seg->c_lock); - kmem_free(kernel_map, (vm_offset_t) c_buffer, C_SEG_ALLOCSIZE); - OSAddAtomic64(-C_SEG_ALLOCSIZE, &compressor_kvspace_used); + segno = c_seg->c_mysegno; - } else if (c_swap_handle) - vm_swap_free(c_swap_handle); + lck_mtx_lock_spin_always(c_list_lock); + /* + * because the c_buffer is now associated with the segno, + * we can't put the segno back on the free list until + * after we have depopulated the c_buffer range, or + * we run the risk of depopulating a range that is + * now being used in one of the compressor heads + */ + c_segments[segno].c_segno = c_free_segno_head; + c_free_segno_head = segno; + c_segment_count--; + lck_mtx_unlock_always(c_list_lock); -#if __i386__ || __x86_64__ lck_mtx_destroy(&c_seg->c_lock, &vm_compressor_lck_grp); -#else /* __i386__ || __x86_64__ */ - lck_spin_destroy(&c_seg->c_lock, &vm_compressor_lck_grp); -#endif /* __i386__ || __x86_64__ */ - - for (i = 0; i < C_SEG_SLOT_ARRAYS; i++) { - if (c_seg->c_slots[i] == 0) - break; - kfree((char *)c_seg->c_slots[i], sizeof(struct c_slot) * C_SEG_SLOT_ARRAY_SIZE); + if (c_seg->c_slot_var_array_len) { + kheap_free(KHEAP_DATA_BUFFERS, c_seg->c_slot_var_array, + sizeof(struct c_slot) * c_seg->c_slot_var_array_len); } + zfree(compressor_segment_zone, c_seg); } - +#if DEVELOPMENT || DEBUG int c_seg_trim_page_count = 0; +#endif void c_seg_trim_tail(c_segment_t c_seg) { - c_slot_t cs; - uint32_t c_size; - uint32_t c_offset; - uint32_t c_rounded_size; - uint16_t current_nextslot; - uint32_t current_populated_offset; + c_slot_t cs; + uint32_t c_size; + uint32_t c_offset; + uint32_t c_rounded_size; + uint16_t current_nextslot; + uint32_t current_populated_offset; - if (c_seg->c_bytes_used == 0) + if (c_seg->c_bytes_used == 0) { return; + } current_nextslot = c_seg->c_nextslot; current_populated_offset = c_seg->c_populated_offset; - - while (c_seg->c_nextslot) { + while (c_seg->c_nextslot) { cs = C_SEG_SLOT_FROM_INDEX(c_seg, (c_seg->c_nextslot - 1)); c_size = UNPACK_C_SIZE(cs); @@ -890,16 +1655,20 @@ c_seg_trim_tail(c_segment_t c_seg) c_offset = cs->c_offset + C_SEG_BYTES_TO_OFFSET(c_rounded_size); c_seg->c_nextoffset = c_offset; - c_seg->c_populated_offset = (c_offset + (C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1)) & ~(C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1); + c_seg->c_populated_offset = (c_offset + (C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1)) & + ~(C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1); - if (c_seg->c_firstemptyslot > c_seg->c_nextslot) + if (c_seg->c_firstemptyslot > c_seg->c_nextslot) { c_seg->c_firstemptyslot = c_seg->c_nextslot; - + } +#if DEVELOPMENT || DEBUG c_seg_trim_page_count += ((round_page_32(C_SEG_OFFSET_TO_BYTES(current_populated_offset)) - - round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE); + round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / + PAGE_SIZE); +#endif } break; - } + } c_seg->c_nextslot--; } assert(c_seg->c_nextslot); @@ -910,15 +1679,14 @@ int c_seg_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy) { c_slot_mapping_t slot_ptr; - uint32_t c_offset = 0; - uint32_t old_populated_offset; - uint32_t c_rounded_size; - uint32_t c_size; - int c_indx = 0; - int i; - c_slot_t c_dst; - c_slot_t c_src; - boolean_t need_unlock = TRUE; + uint32_t c_offset = 0; + uint32_t old_populated_offset; + uint32_t c_rounded_size; + uint32_t c_size; + uint16_t c_indx = 0; + int i; + c_slot_t c_dst; + c_slot_t c_src; assert(c_seg->c_busy); @@ -927,46 +1695,48 @@ c_seg_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy) #endif if (c_seg->c_bytes_used == 0) { c_seg_free(c_seg); - return (1); + return 1; } - if (c_seg->c_firstemptyslot >= c_seg->c_nextslot || C_SEG_UNUSED_BYTES(c_seg) < PAGE_SIZE) + lck_mtx_unlock_always(&c_seg->c_lock); + + if (c_seg->c_firstemptyslot >= c_seg->c_nextslot || C_SEG_UNUSED_BYTES(c_seg) < PAGE_SIZE) { goto done; - + } + +/* TODO: assert first emptyslot's c_size is actually 0 */ + +#if DEVELOPMENT || DEBUG + C_SEG_MAKE_WRITEABLE(c_seg); +#endif + #if VALIDATE_C_SEGMENTS c_seg->c_was_minor_compacted++; #endif c_indx = c_seg->c_firstemptyslot; c_dst = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx); - + old_populated_offset = c_seg->c_populated_offset; c_offset = c_dst->c_offset; for (i = c_indx + 1; i < c_seg->c_nextslot && c_offset < c_seg->c_nextoffset; i++) { - c_src = C_SEG_SLOT_FROM_INDEX(c_seg, i); c_size = UNPACK_C_SIZE(c_src); - if (c_size == 0) + if (c_size == 0) { continue; + } - memcpy(&c_seg->c_store.c_buffer[c_offset], &c_seg->c_store.c_buffer[c_src->c_offset], c_size); + c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK; +/* N.B.: This memcpy may be an overlapping copy */ + memcpy(&c_seg->c_store.c_buffer[c_offset], &c_seg->c_store.c_buffer[c_src->c_offset], c_rounded_size); -#if CHECKSUM_THE_DATA - c_dst->c_hash_data = c_src->c_hash_data; -#endif -#if CHECKSUM_THE_COMPRESSED_DATA - c_dst->c_hash_compressed_data = c_src->c_hash_compressed_data; -#endif - c_dst->c_size = c_src->c_size; - c_dst->c_packed_ptr = c_src->c_packed_ptr; + cslot_copy(c_dst, c_src); c_dst->c_offset = c_offset; - slot_ptr = (c_slot_mapping_t)C_SLOT_UNPACK_PTR(c_dst); + slot_ptr = C_SLOT_UNPACK_PTR(c_dst); slot_ptr->s_cindx = c_indx; - c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK; - c_offset += C_SEG_BYTES_TO_OFFSET(c_rounded_size); PACK_C_SIZE(c_src, 0); c_indx++; @@ -982,46 +1752,94 @@ c_seg_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy) #if VALIDATE_C_SEGMENTS c_seg_validate(c_seg, TRUE); #endif - if (old_populated_offset > c_seg->c_populated_offset) { - uint32_t gc_size; - int32_t *gc_ptr; + uint32_t gc_size; + int32_t *gc_ptr; gc_size = C_SEG_OFFSET_TO_BYTES(old_populated_offset - c_seg->c_populated_offset); gc_ptr = &c_seg->c_store.c_buffer[c_seg->c_populated_offset]; - lck_mtx_unlock_always(&c_seg->c_lock); + kernel_memory_depopulate(compressor_map, (vm_offset_t)gc_ptr, gc_size, + KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR); + } - kernel_memory_depopulate(kernel_map, (vm_offset_t)gc_ptr, gc_size, KMA_COMPRESSOR); +#if DEVELOPMENT || DEBUG + C_SEG_WRITE_PROTECT(c_seg); +#endif - if (clear_busy == TRUE) - lck_mtx_lock_spin_always(&c_seg->c_lock); - else - need_unlock = FALSE; - } done: - if (need_unlock == TRUE) { - if (clear_busy == TRUE) - C_SEG_WAKEUP_DONE(c_seg); - + if (clear_busy == TRUE) { + lck_mtx_lock_spin_always(&c_seg->c_lock); + C_SEG_WAKEUP_DONE(c_seg); lck_mtx_unlock_always(&c_seg->c_lock); } - return (0); + return 0; } +static void +c_seg_alloc_nextslot(c_segment_t c_seg) +{ + struct c_slot *old_slot_array = NULL; + struct c_slot *new_slot_array = NULL; + int newlen; + int oldlen; -struct { - uint64_t asked_permission; - uint64_t compactions; + if (c_seg->c_nextslot < c_seg_fixed_array_len) { + return; + } + + if ((c_seg->c_nextslot - c_seg_fixed_array_len) >= c_seg->c_slot_var_array_len) { + oldlen = c_seg->c_slot_var_array_len; + old_slot_array = c_seg->c_slot_var_array; + + if (oldlen == 0) { + newlen = C_SEG_SLOT_VAR_ARRAY_MIN_LEN; + } else { + newlen = oldlen * 2; + } + + new_slot_array = kheap_alloc(KHEAP_DATA_BUFFERS, + sizeof(struct c_slot) * newlen, Z_WAITOK); + + lck_mtx_lock_spin_always(&c_seg->c_lock); + + if (old_slot_array) { + memcpy(new_slot_array, old_slot_array, + sizeof(struct c_slot) * oldlen); + } + + c_seg->c_slot_var_array_len = newlen; + c_seg->c_slot_var_array = new_slot_array; + + lck_mtx_unlock_always(&c_seg->c_lock); + + if (old_slot_array) { + kheap_free(KHEAP_DATA_BUFFERS, old_slot_array, + sizeof(struct c_slot) * oldlen); + } + } +} + + +#define C_SEG_MAJOR_COMPACT_STATS_MAX (30) + +struct { + uint64_t asked_permission; + uint64_t compactions; uint64_t moved_slots; uint64_t moved_bytes; uint64_t wasted_space_in_swapouts; uint64_t count_of_swapouts; -} c_seg_major_compact_stats; + uint64_t count_of_freed_segs; + uint64_t bailed_compactions; + uint64_t bytes_freed_rate_us; +} c_seg_major_compact_stats[C_SEG_MAJOR_COMPACT_STATS_MAX]; +int c_seg_major_compact_stats_now = 0; -#define C_MAJOR_COMPACTION_SIZE_APPROPRIATE ((C_SEG_BUFSIZE * 90) / 100) + +#define C_MAJOR_COMPACTION_SIZE_APPROPRIATE ((C_SEG_BUFSIZE * 90) / 100) boolean_t @@ -1029,27 +1847,21 @@ c_seg_major_compact_ok( c_segment_t c_seg_dst, c_segment_t c_seg_src) { + c_seg_major_compact_stats[c_seg_major_compact_stats_now].asked_permission++; - c_seg_major_compact_stats.asked_permission++; - - if (c_seg_src->c_filling) { - /* - * we're at or near the head... don't compact - */ - return (FALSE); - } if (c_seg_src->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE && - c_seg_dst->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE) - return (FALSE); + c_seg_dst->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE) { + return FALSE; + } - if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX) { + if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX_INDEX) { /* * destination segment is full... can't compact */ - return (FALSE); + return FALSE; } - return (TRUE); + return TRUE; } @@ -1059,15 +1871,14 @@ c_seg_major_compact( c_segment_t c_seg_src) { c_slot_mapping_t slot_ptr; - uint32_t c_rounded_size; - uint32_t c_size; - uint16_t dst_slot; - int i; - c_slot_t c_dst; - c_slot_t c_src; - int slotarray; - boolean_t keep_compacting = TRUE; - + uint32_t c_rounded_size; + uint32_t c_size; + uint16_t dst_slot; + int i; + c_slot_t c_dst; + c_slot_t c_src; + boolean_t keep_compacting = TRUE; + /* * segments are not locked but they are both marked c_busy * which keeps c_decompress from working on them... @@ -1075,17 +1886,19 @@ c_seg_major_compact( * from c_seg_src to c_seg_dst and update both c_segment's * state w/o holding the master lock */ +#if DEVELOPMENT || DEBUG + C_SEG_MAKE_WRITEABLE(c_seg_dst); +#endif #if VALIDATE_C_SEGMENTS c_seg_dst->c_was_major_compacted++; c_seg_src->c_was_major_donor++; #endif - c_seg_major_compact_stats.compactions++; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].compactions++; dst_slot = c_seg_dst->c_nextslot; for (i = 0; i < c_seg_src->c_nextslot; i++) { - c_src = C_SEG_SLOT_FROM_INDEX(c_seg_src, i); c_size = UNPACK_C_SIZE(c_src); @@ -1096,51 +1909,47 @@ c_seg_major_compact( } if (C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset - c_seg_dst->c_nextoffset) < (unsigned) c_size) { + int size_to_populate; + /* doesn't fit */ - if ((C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset) == C_SEG_BUFSIZE)) { + size_to_populate = C_SEG_BUFSIZE - C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset); + + if (size_to_populate == 0) { /* can't fit */ keep_compacting = FALSE; break; } - kernel_memory_populate(kernel_map, - (vm_offset_t) &c_seg_dst->c_store.c_buffer[c_seg_dst->c_populated_offset], - PAGE_SIZE, - KMA_COMPRESSOR); + if (size_to_populate > C_SEG_MAX_POPULATE_SIZE) { + size_to_populate = C_SEG_MAX_POPULATE_SIZE; + } - c_seg_dst->c_populated_offset += C_SEG_BYTES_TO_OFFSET(PAGE_SIZE); + kernel_memory_populate(compressor_map, + (vm_offset_t) &c_seg_dst->c_store.c_buffer[c_seg_dst->c_populated_offset], + size_to_populate, + KMA_COMPRESSOR, + VM_KERN_MEMORY_COMPRESSOR); + + c_seg_dst->c_populated_offset += C_SEG_BYTES_TO_OFFSET(size_to_populate); assert(C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset) <= C_SEG_BUFSIZE); } + c_seg_alloc_nextslot(c_seg_dst); - slotarray = C_SEG_SLOTARRAY_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot); - - if (c_seg_dst->c_slots[slotarray] == 0) { - KERNEL_DEBUG(0xe0400008 | DBG_FUNC_START, 0, 0, 0, 0, 0); - c_seg_dst->c_slots[slotarray] = (struct c_slot *) - kalloc(sizeof(struct c_slot) * - C_SEG_SLOT_ARRAY_SIZE); - KERNEL_DEBUG(0xe0400008 | DBG_FUNC_END, 0, 0, 0, 0, 0); - } c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot); memcpy(&c_seg_dst->c_store.c_buffer[c_seg_dst->c_nextoffset], &c_seg_src->c_store.c_buffer[c_src->c_offset], c_size); c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK; - c_seg_major_compact_stats.moved_slots++; - c_seg_major_compact_stats.moved_bytes += c_size; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].moved_slots++; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].moved_bytes += c_size; -#if CHECKSUM_THE_DATA - c_dst->c_hash_data = c_src->c_hash_data; -#endif -#if CHECKSUM_THE_COMPRESSED_DATA - c_dst->c_hash_compressed_data = c_src->c_hash_compressed_data; -#endif - c_dst->c_size = c_src->c_size; - c_dst->c_packed_ptr = c_src->c_packed_ptr; + cslot_copy(c_dst, c_src); c_dst->c_offset = c_seg_dst->c_nextoffset; - if (c_seg_dst->c_firstemptyslot == c_seg_dst->c_nextslot) + if (c_seg_dst->c_firstemptyslot == c_seg_dst->c_nextslot) { c_seg_dst->c_firstemptyslot++; + } + c_seg_dst->c_slots_used++; c_seg_dst->c_nextslot++; c_seg_dst->c_bytes_used += c_rounded_size; c_seg_dst->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size); @@ -1151,47 +1960,51 @@ c_seg_major_compact( c_seg_src->c_bytes_unused += c_rounded_size; c_seg_src->c_firstemptyslot = 0; - if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX) { + assert(c_seg_src->c_slots_used); + c_seg_src->c_slots_used--; + + if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX_INDEX) { /* dest segment is now full */ keep_compacting = FALSE; break; } } +#if DEVELOPMENT || DEBUG + C_SEG_WRITE_PROTECT(c_seg_dst); +#endif if (dst_slot < c_seg_dst->c_nextslot) { - PAGE_REPLACEMENT_ALLOWED(TRUE); /* - * we've now locked out c_decompress from + * we've now locked out c_decompress from * converting the slot passed into it into - * a c_segment_t which allows us to use + * a c_segment_t which allows us to use * the backptr to change which c_segment and * index the slot points to */ while (dst_slot < c_seg_dst->c_nextslot) { - c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, dst_slot); - - slot_ptr = (c_slot_mapping_t)C_SLOT_UNPACK_PTR(c_dst); + + slot_ptr = C_SLOT_UNPACK_PTR(c_dst); /* would mean "empty slot", so use csegno+1 */ slot_ptr->s_cseg = c_seg_dst->c_mysegno + 1; slot_ptr->s_cindx = dst_slot++; } PAGE_REPLACEMENT_ALLOWED(FALSE); } - return (keep_compacting); + return keep_compacting; } uint64_t vm_compressor_compute_elapsed_msecs(clock_sec_t end_sec, clock_nsec_t end_nsec, clock_sec_t start_sec, clock_nsec_t start_nsec) { - uint64_t end_msecs; - uint64_t start_msecs; - + uint64_t end_msecs; + uint64_t start_msecs; + end_msecs = (end_sec * 1000) + end_nsec / 1000000; start_msecs = (start_sec * 1000) + start_nsec / 1000000; - return (end_msecs - start_msecs); + return end_msecs - start_msecs; } @@ -1203,20 +2016,18 @@ uint32_t compressor_thrashing_threshold_per_10msecs = 50; uint32_t compressor_thrashing_min_per_10msecs = 20; /* When true, reset sample data next chance we get. */ -static boolean_t compressor_need_sample_reset = FALSE; - -extern uint32_t vm_page_filecache_min; +static boolean_t compressor_need_sample_reset = FALSE; void compute_swapout_target_age(void) { - clock_sec_t cur_ts_sec; - clock_nsec_t cur_ts_nsec; - uint32_t min_operations_needed_in_this_sample; - uint64_t elapsed_msecs_in_eval; - uint64_t elapsed_msecs_in_sample; - boolean_t need_eval_reset = FALSE; + clock_sec_t cur_ts_sec; + clock_nsec_t cur_ts_nsec; + uint32_t min_operations_needed_in_this_sample; + uint64_t elapsed_msecs_in_eval; + uint64_t elapsed_msecs_in_sample; + boolean_t need_eval_reset = FALSE; clock_get_system_nanotime(&cur_ts_sec, &cur_ts_nsec); @@ -1229,9 +2040,10 @@ compute_swapout_target_age(void) goto done; } elapsed_msecs_in_eval = vm_compressor_compute_elapsed_msecs(cur_ts_sec, cur_ts_nsec, start_of_eval_period_sec, start_of_eval_period_nsec); - - if (elapsed_msecs_in_eval < compressor_eval_period_in_msecs) + + if (elapsed_msecs_in_eval < compressor_eval_period_in_msecs) { goto done; + } need_eval_reset = TRUE; KERNEL_DEBUG(0xe0400020 | DBG_FUNC_START, elapsed_msecs_in_eval, sample_period_compression_count, sample_period_decompression_count, 0, 0); @@ -1240,9 +2052,8 @@ compute_swapout_target_age(void) if ((sample_period_compression_count - last_eval_compression_count) < min_operations_needed_in_this_sample || (sample_period_decompression_count - last_eval_decompression_count) < min_operations_needed_in_this_sample) { - KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, sample_period_compression_count - last_eval_compression_count, - sample_period_decompression_count - last_eval_decompression_count, 0, 1, 0); + sample_period_decompression_count - last_eval_decompression_count, 0, 1, 0); swapout_target_age = 0; @@ -1254,46 +2065,43 @@ compute_swapout_target_age(void) last_eval_decompression_count = sample_period_decompression_count; if (elapsed_msecs_in_sample < compressor_sample_min_in_msecs) { - KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, swapout_target_age, 0, 0, 5, 0); goto done; } if (sample_period_decompression_count > ((compressor_thrashing_threshold_per_10msecs * elapsed_msecs_in_sample) / 10)) { - - uint64_t running_total; - uint64_t working_target; - uint64_t aging_target; - uint32_t oldest_age_of_csegs_sampled = 0; - uint64_t working_set_approximation = 0; + uint64_t running_total; + uint64_t working_target; + uint64_t aging_target; + uint32_t oldest_age_of_csegs_sampled = 0; + uint64_t working_set_approximation = 0; swapout_target_age = 0; - working_target = (sample_period_decompression_count / 100) * 95; /* 95 percent */ - aging_target = (sample_period_decompression_count / 100) * 1; /* 1 percent */ + working_target = (sample_period_decompression_count / 100) * 95; /* 95 percent */ + aging_target = (sample_period_decompression_count / 100) * 1; /* 1 percent */ running_total = 0; for (oldest_age_of_csegs_sampled = 0; oldest_age_of_csegs_sampled < DECOMPRESSION_SAMPLE_MAX_AGE; oldest_age_of_csegs_sampled++) { - running_total += age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled]; working_set_approximation += oldest_age_of_csegs_sampled * age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled]; - if (running_total >= working_target) + if (running_total >= working_target) { break; + } } if (oldest_age_of_csegs_sampled < DECOMPRESSION_SAMPLE_MAX_AGE) { - working_set_approximation = (working_set_approximation * 1000) / elapsed_msecs_in_sample; if (working_set_approximation < VM_PAGE_COMPRESSOR_COUNT) { - running_total = overage_decompressions_during_sample_period; for (oldest_age_of_csegs_sampled = DECOMPRESSION_SAMPLE_MAX_AGE - 1; oldest_age_of_csegs_sampled; oldest_age_of_csegs_sampled--) { running_total += age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled]; - if (running_total >= aging_target) + if (running_total >= aging_target) { break; + } } swapout_target_age = (uint32_t)cur_ts_sec - oldest_age_of_csegs_sampled; @@ -1301,13 +2109,15 @@ compute_swapout_target_age(void) } else { KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, working_set_approximation, VM_PAGE_COMPRESSOR_COUNT, 0, 3, 0); } - } else + } else { KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, working_target, running_total, 0, 4, 0); + } compressor_need_sample_reset = TRUE; need_eval_reset = TRUE; - } else + } else { KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, sample_period_decompression_count, (compressor_thrashing_threshold_per_10msecs * elapsed_msecs_in_sample) / 10, 0, 6, 0); + } done: if (compressor_need_sample_reset == TRUE) { bzero(age_of_decompressions_during_sample_period, sizeof(age_of_decompressions_during_sample_period)); @@ -1328,85 +2138,144 @@ done: } -int compaction_swapper_inited = 0; -int compaction_swapper_init_now = 0; -int compaction_swapper_running = 0; -int compaction_swapper_abort = 0; +int compaction_swapper_init_now = 0; +int compaction_swapper_running = 0; +int compaction_swapper_awakened = 0; +int compaction_swapper_abort = 0; #if CONFIG_JETSAM -boolean_t memorystatus_kill_on_VM_thrashing(boolean_t); -boolean_t memorystatus_kill_on_FC_thrashing(boolean_t); -int compressor_thrashing_induced_jetsam = 0; -int filecache_thrashing_induced_jetsam = 0; -static boolean_t vm_compressor_thrashing_detected = FALSE; +boolean_t memorystatus_kill_on_VM_compressor_thrashing(boolean_t); +boolean_t memorystatus_kill_on_VM_compressor_space_shortage(boolean_t); +boolean_t memorystatus_kill_on_FC_thrashing(boolean_t); +int compressor_thrashing_induced_jetsam = 0; +int filecache_thrashing_induced_jetsam = 0; +static boolean_t vm_compressor_thrashing_detected = FALSE; #endif /* CONFIG_JETSAM */ static boolean_t compressor_needs_to_swap(void) { - boolean_t should_swap = FALSE; + boolean_t should_swap = FALSE; - if (vm_swap_up == TRUE) { + if (vm_swapout_ripe_segments == TRUE && c_overage_swapped_count < c_overage_swapped_limit) { + c_segment_t c_seg; + clock_sec_t now; + clock_sec_t age; + clock_nsec_t nsec; + + clock_get_system_nanotime(&now, &nsec); + age = 0; + + lck_mtx_lock_spin_always(c_list_lock); + + if (!queue_empty(&c_age_list_head)) { + c_seg = (c_segment_t) queue_first(&c_age_list_head); + + age = now - c_seg->c_creation_ts; + } + lck_mtx_unlock_always(c_list_lock); + + if (age >= vm_ripe_target_age) { + should_swap = TRUE; + goto check_if_low_space; + } + } + if (VM_CONFIG_SWAP_IS_ACTIVE) { if (COMPRESSOR_NEEDS_TO_SWAP()) { - return (TRUE); + should_swap = TRUE; + goto check_if_low_space; } if (VM_PAGE_Q_THROTTLED(&vm_pageout_queue_external) && vm_page_anonymous_count < (vm_page_inactive_count / 20)) { - return (TRUE); + should_swap = TRUE; + goto check_if_low_space; + } + if (vm_page_free_count < (vm_page_free_reserved - (COMPRESSOR_FREE_RESERVED_LIMIT * 2))) { + should_swap = TRUE; + goto check_if_low_space; } - if (vm_page_free_count < (vm_page_free_reserved - COMPRESSOR_FREE_RESERVED_LIMIT)) - return (TRUE); } + +#if (XNU_TARGET_OS_OSX && __arm64__) + /* + * Thrashing detection disabled. + */ +#else /* (XNU_TARGET_OS_OSX && __arm64__) */ + compute_swapout_target_age(); - + if (swapout_target_age) { - c_segment_t c_seg; + c_segment_t c_seg; lck_mtx_lock_spin_always(c_list_lock); if (!queue_empty(&c_age_list_head)) { - c_seg = (c_segment_t) queue_first(&c_age_list_head); - if (c_seg->c_creation_ts > swapout_target_age) + if (c_seg->c_creation_ts > swapout_target_age) { swapout_target_age = 0; + } } lck_mtx_unlock_always(c_list_lock); } #if CONFIG_PHANTOM_CACHE - if (vm_phantom_cache_check_pressure()) + if (vm_phantom_cache_check_pressure()) { should_swap = TRUE; + } #endif - if (swapout_target_age) + if (swapout_target_age) { should_swap = TRUE; + } +#endif /* (XNU_TARGET_OS_OSX && __arm64__) */ - if (vm_swap_up == FALSE) { +check_if_low_space: - if (should_swap) { #if CONFIG_JETSAM - if (vm_compressor_thrashing_detected == FALSE) { - vm_compressor_thrashing_detected = TRUE; - - if (swapout_target_age) { - memorystatus_kill_on_VM_thrashing(TRUE /* async */); - compressor_thrashing_induced_jetsam++; - } else { - memorystatus_kill_on_FC_thrashing(TRUE /* async */); - filecache_thrashing_induced_jetsam++; - } - /* - * let the jetsam take precedence over - * any major compactions we might have - * been able to do... otherwise we run - * the risk of doing major compactions - * on segments we're about to free up - * due to the jetsam activity. - */ - should_swap = FALSE; + if (should_swap || vm_compressor_low_on_space() == TRUE) { + if (vm_compressor_thrashing_detected == FALSE) { + vm_compressor_thrashing_detected = TRUE; + + if (swapout_target_age) { + /* The compressor is thrashing. */ + memorystatus_kill_on_VM_compressor_thrashing(TRUE /* async */); + compressor_thrashing_induced_jetsam++; + } else if (vm_compressor_low_on_space() == TRUE) { + /* The compressor is running low on space. */ + memorystatus_kill_on_VM_compressor_space_shortage(TRUE /* async */); + compressor_thrashing_induced_jetsam++; + } else { + memorystatus_kill_on_FC_thrashing(TRUE /* async */); + filecache_thrashing_induced_jetsam++; } + } + /* + * let the jetsam take precedence over + * any major compactions we might have + * been able to do... otherwise we run + * the risk of doing major compactions + * on segments we're about to free up + * due to the jetsam activity. + */ + should_swap = FALSE; + } + +#else /* CONFIG_JETSAM */ + if (should_swap && vm_swap_low_on_space()) { + vm_compressor_take_paging_space_action(); + } #endif /* CONFIG_JETSAM */ - } else - should_swap = COMPRESSOR_NEEDS_TO_MAJOR_COMPACT(); + + if (should_swap == FALSE) { + /* + * vm_compressor_needs_to_major_compact returns true only if we're + * about to run out of available compressor segments... in this + * case, we absolutely need to run a major compaction even if + * we've just kicked off a jetsam or we don't otherwise need to + * swap... terminating objects releases + * pages back to the uncompressed cache, but does not guarantee + * that we will free up even a single compression segment + */ + should_swap = vm_compressor_needs_to_major_compact(); } /* @@ -1414,10 +2283,9 @@ compressor_needs_to_swap(void) * will cause the major compaction engine to * run, but will not trigger any swapping... * segments that have been major compacted - * will be moved to the swapped_out_q - * but will not have the c_ondisk flag set + * will be moved to the majorcompact queue */ - return (should_swap); + return should_swap; } #if CONFIG_JETSAM @@ -1448,40 +2316,154 @@ vm_thrashing_jetsam_done(void) #endif /* CONFIG_JETSAM */ uint32_t vm_wake_compactor_swapper_calls = 0; +uint32_t vm_run_compactor_already_running = 0; +uint32_t vm_run_compactor_empty_minor_q = 0; +uint32_t vm_run_compactor_did_compact = 0; +uint32_t vm_run_compactor_waited = 0; void -vm_wake_compactor_swapper(void) +vm_run_compactor(void) { - boolean_t need_major_compaction = FALSE; + if (c_segment_count == 0) { + return; + } + + lck_mtx_lock_spin_always(c_list_lock); + + if (c_minor_count == 0) { + vm_run_compactor_empty_minor_q++; + + lck_mtx_unlock_always(c_list_lock); + return; + } + if (compaction_swapper_running) { + if (vm_pageout_state.vm_restricted_to_single_processor == FALSE) { + vm_run_compactor_already_running++; + + lck_mtx_unlock_always(c_list_lock); + return; + } + vm_run_compactor_waited++; + + assert_wait((event_t)&compaction_swapper_running, THREAD_UNINT); + + lck_mtx_unlock_always(c_list_lock); + + thread_block(THREAD_CONTINUE_NULL); - if (compaction_swapper_running) return; + } + vm_run_compactor_did_compact++; + + fastwake_warmup = FALSE; + compaction_swapper_running = 1; + + vm_compressor_do_delayed_compactions(FALSE); + + compaction_swapper_running = 0; + + lck_mtx_unlock_always(c_list_lock); + + thread_wakeup((event_t)&compaction_swapper_running); +} - if (c_minor_count == 0 && need_major_compaction == FALSE) + +void +vm_wake_compactor_swapper(void) +{ + if (compaction_swapper_running || compaction_swapper_awakened || c_segment_count == 0) { return; + } + + if (c_minor_count || vm_compressor_needs_to_major_compact()) { + lck_mtx_lock_spin_always(c_list_lock); + + fastwake_warmup = FALSE; + + if (compaction_swapper_running == 0 && compaction_swapper_awakened == 0) { + vm_wake_compactor_swapper_calls++; + + compaction_swapper_awakened = 1; + thread_wakeup((event_t)&c_compressor_swap_trigger); + } + lck_mtx_unlock_always(c_list_lock); + } +} + + +void +vm_consider_swapping() +{ + c_segment_t c_seg, c_seg_next; + clock_sec_t now; + clock_nsec_t nsec; + + assert(VM_CONFIG_SWAP_IS_PRESENT); lck_mtx_lock_spin_always(c_list_lock); - fastwake_warmup = FALSE; + compaction_swapper_abort = 1; - if (compaction_swapper_running == 0) { - vm_wake_compactor_swapper_calls++; + while (compaction_swapper_running) { + assert_wait((event_t)&compaction_swapper_running, THREAD_UNINT); - thread_wakeup((event_t)&c_compressor_swap_trigger); - - compaction_swapper_running = 1; + lck_mtx_unlock_always(c_list_lock); + + thread_block(THREAD_CONTINUE_NULL); + + lck_mtx_lock_spin_always(c_list_lock); + } + compaction_swapper_abort = 0; + compaction_swapper_running = 1; + + vm_swapout_ripe_segments = TRUE; + + if (!queue_empty(&c_major_list_head)) { + clock_get_system_nanotime(&now, &nsec); + + c_seg = (c_segment_t)queue_first(&c_major_list_head); + + while (!queue_end(&c_major_list_head, (queue_entry_t)c_seg)) { + if (c_overage_swapped_count >= c_overage_swapped_limit) { + break; + } + + c_seg_next = (c_segment_t) queue_next(&c_seg->c_age_list); + + if ((now - c_seg->c_creation_ts) >= vm_ripe_target_age) { + lck_mtx_lock_spin_always(&c_seg->c_lock); + + c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE); + + lck_mtx_unlock_always(&c_seg->c_lock); + } + c_seg = c_seg_next; + } } + vm_compressor_compact_and_swap(FALSE); + + compaction_swapper_running = 0; + + vm_swapout_ripe_segments = FALSE; + lck_mtx_unlock_always(c_list_lock); + + thread_wakeup((event_t)&compaction_swapper_running); } void vm_consider_waking_compactor_swapper(void) { - boolean_t need_wakeup = FALSE; + boolean_t need_wakeup = FALSE; + + if (c_segment_count == 0) { + return; + } - if (compaction_swapper_running) + if (compaction_swapper_running || compaction_swapper_awakened) { return; + } if (!compaction_swapper_inited && !compaction_swapper_init_now) { compaction_swapper_init_now = 1; @@ -1489,60 +2471,57 @@ vm_consider_waking_compactor_swapper(void) } if (c_minor_count && (COMPRESSOR_NEEDS_TO_MINOR_COMPACT())) { - need_wakeup = TRUE; - } else if (compressor_needs_to_swap()) { - need_wakeup = TRUE; - } else if (c_minor_count) { - uint64_t total_bytes; + uint64_t total_bytes; total_bytes = compressor_object->resident_page_count * PAGE_SIZE_64; - if ((total_bytes - compressor_bytes_used) > total_bytes / 10) + if ((total_bytes - compressor_bytes_used) > total_bytes / 10) { need_wakeup = TRUE; + } } if (need_wakeup == TRUE) { - lck_mtx_lock_spin_always(c_list_lock); fastwake_warmup = FALSE; - if (compaction_swapper_running == 0) { + if (compaction_swapper_running == 0 && compaction_swapper_awakened == 0) { memoryshot(VM_WAKEUP_COMPACTOR_SWAPPER, DBG_FUNC_NONE); + compaction_swapper_awakened = 1; thread_wakeup((event_t)&c_compressor_swap_trigger); - - compaction_swapper_running = 1; } lck_mtx_unlock_always(c_list_lock); } } -#define C_SWAPOUT_LIMIT 4 -#define DELAYED_COMPACTIONS_PER_PASS 30 +#define C_SWAPOUT_LIMIT 4 +#define DELAYED_COMPACTIONS_PER_PASS 30 void vm_compressor_do_delayed_compactions(boolean_t flush_all) { - c_segment_t c_seg; - int number_compacted = 0; - boolean_t needs_to_swap = FALSE; + c_segment_t c_seg; + int number_compacted = 0; + boolean_t needs_to_swap = FALSE; + + VM_DEBUG_CONSTANT_EVENT(vm_compressor_do_delayed_compactions, VM_COMPRESSOR_DO_DELAYED_COMPACTIONS, DBG_FUNC_START, c_minor_count, flush_all, 0, 0); - lck_mtx_assert(c_list_lock, LCK_MTX_ASSERT_OWNED); +#if XNU_TARGET_OS_OSX + LCK_MTX_ASSERT(c_list_lock, LCK_MTX_ASSERT_OWNED); +#endif /* XNU_TARGET_OS_OSX */ while (!queue_empty(&c_minor_list_head) && needs_to_swap == FALSE) { - c_seg = (c_segment_t)queue_first(&c_minor_list_head); - + lck_mtx_lock_spin_always(&c_seg->c_lock); if (c_seg->c_busy) { - lck_mtx_unlock_always(c_list_lock); c_seg_wait_on_busy(c_seg); lck_mtx_lock_spin_always(c_list_lock); @@ -1553,60 +2532,64 @@ vm_compressor_do_delayed_compactions(boolean_t flush_all) c_seg_do_minor_compaction_and_unlock(c_seg, TRUE, FALSE, TRUE); - if (vm_swap_up == TRUE && (number_compacted++ > DELAYED_COMPACTIONS_PER_PASS)) { - - if ((flush_all == TRUE || compressor_needs_to_swap() == TRUE) && c_swapout_count < C_SWAPOUT_LIMIT) + if (VM_CONFIG_SWAP_IS_ACTIVE && (number_compacted++ > DELAYED_COMPACTIONS_PER_PASS)) { + if ((flush_all == TRUE || compressor_needs_to_swap() == TRUE) && c_swapout_count < C_SWAPOUT_LIMIT) { needs_to_swap = TRUE; + } number_compacted = 0; } lck_mtx_lock_spin_always(c_list_lock); } + + VM_DEBUG_CONSTANT_EVENT(vm_compressor_do_delayed_compactions, VM_COMPRESSOR_DO_DELAYED_COMPACTIONS, DBG_FUNC_END, c_minor_count, number_compacted, needs_to_swap, 0); } -#define C_SEGMENT_SWAPPEDIN_AGE_LIMIT 10 +#define C_SEGMENT_SWAPPEDIN_AGE_LIMIT 10 static void vm_compressor_age_swapped_in_segments(boolean_t flush_all) { - c_segment_t c_seg; - clock_sec_t now; - clock_nsec_t nsec; + c_segment_t c_seg; + clock_sec_t now; + clock_nsec_t nsec; - clock_get_system_nanotime(&now, &nsec); - - while (!queue_empty(&c_swappedin_list_head)) { + clock_get_system_nanotime(&now, &nsec); + while (!queue_empty(&c_swappedin_list_head)) { c_seg = (c_segment_t)queue_first(&c_swappedin_list_head); - if (flush_all == FALSE && (now - c_seg->c_swappedin_ts) < C_SEGMENT_SWAPPEDIN_AGE_LIMIT) + if (flush_all == FALSE && (now - c_seg->c_swappedin_ts) < C_SEGMENT_SWAPPEDIN_AGE_LIMIT) { break; - - lck_mtx_lock_spin_always(&c_seg->c_lock); + } - queue_remove(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedin_q = 0; - c_swappedin_count--; + lck_mtx_lock_spin_always(&c_seg->c_lock); - c_seg_insert_into_q(&c_age_list_head, c_seg); - c_seg->c_on_age_q = 1; - c_age_count++; + c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE); lck_mtx_unlock_always(&c_seg->c_lock); } } +extern int vm_num_swap_files; +extern int vm_num_pinned_swap_files; +extern int vm_swappin_enabled; + +extern unsigned int vm_swapfile_total_segs_used; +extern unsigned int vm_swapfile_total_segs_alloced; + + void vm_compressor_flush(void) { - uint64_t vm_swap_put_failures_at_start; - wait_result_t wait_result = 0; - AbsoluteTime startTime, endTime; - clock_sec_t now_sec; - clock_nsec_t now_nsec; - uint64_t nsec; + uint64_t vm_swap_put_failures_at_start; + wait_result_t wait_result = 0; + AbsoluteTime startTime, endTime; + clock_sec_t now_sec; + clock_nsec_t now_nsec; + uint64_t nsec; HIBLOG("vm_compressor_flush - starting\n"); @@ -1621,7 +2604,7 @@ vm_compressor_flush(void) assert_wait((event_t)&compaction_swapper_running, THREAD_UNINT); lck_mtx_unlock_always(c_list_lock); - + thread_block(THREAD_CONTINUE_NULL); lck_mtx_lock_spin_always(c_list_lock); @@ -1641,73 +2624,90 @@ vm_compressor_flush(void) vm_compressor_compact_and_swap(TRUE); while (!queue_empty(&c_swapout_list_head)) { - - assert_wait_timeout((event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, 5000, 1000*NSEC_PER_USEC); + assert_wait_timeout((event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, 5000, 1000 * NSEC_PER_USEC); lck_mtx_unlock_always(c_list_lock); - + wait_result = thread_block(THREAD_CONTINUE_NULL); lck_mtx_lock_spin_always(c_list_lock); - if (wait_result == THREAD_TIMED_OUT) + if (wait_result == THREAD_TIMED_OUT) { break; + } } hibernate_flushing = FALSE; compaction_swapper_running = 0; - if (vm_swap_put_failures > vm_swap_put_failures_at_start) + if (vm_swap_put_failures > vm_swap_put_failures_at_start) { HIBLOG("vm_compressor_flush failed to clean %llu segments - vm_page_compressor_count(%d)\n", - vm_swap_put_failures - vm_swap_put_failures_at_start, VM_PAGE_COMPRESSOR_COUNT); - + vm_swap_put_failures - vm_swap_put_failures_at_start, VM_PAGE_COMPRESSOR_COUNT); + } + lck_mtx_unlock_always(c_list_lock); - clock_get_uptime(&endTime); - SUB_ABSOLUTETIME(&endTime, &startTime); - absolutetime_to_nanoseconds(endTime, &nsec); + thread_wakeup((event_t)&compaction_swapper_running); + + clock_get_uptime(&endTime); + SUB_ABSOLUTETIME(&endTime, &startTime); + absolutetime_to_nanoseconds(endTime, &nsec); - HIBLOG("vm_compressor_flush completed - took %qd msecs\n", nsec / 1000000ULL); + HIBLOG("vm_compressor_flush completed - took %qd msecs - vm_num_swap_files = %d, vm_num_pinned_swap_files = %d, vm_swappin_enabled = %d\n", + nsec / 1000000ULL, vm_num_swap_files, vm_num_pinned_swap_files, vm_swappin_enabled); } -extern void vm_swap_file_set_tuneables(void); -int compaction_swap_trigger_thread_awakened = 0; - +int compaction_swap_trigger_thread_awakened = 0; static void vm_compressor_swap_trigger_thread(void) { + current_thread()->options |= TH_OPT_VMPRIV; + /* * compaction_swapper_init_now is set when the first call to - * vm_consider_waking_compactor_swapper is made from - * vm_pageout_scan... since this function is called upon + * vm_consider_waking_compactor_swapper is made from + * vm_pageout_scan... since this function is called upon * thread creation, we want to make sure to delay adjusting * the tuneables until we are awakened via vm_pageout_scan * so that we are at a point where the vm_swapfile_open will * be operating on the correct directory (in case the default - * of /var/vm/ is overridden by the dymanic_pager + * of using the VM volume is overridden by the dynamic_pager) */ - if (compaction_swapper_init_now && !compaction_swapper_inited) { - if (vm_compressor_mode == VM_PAGER_COMPRESSOR_WITH_SWAP) - vm_swap_file_set_tuneables(); + if (compaction_swapper_init_now) { + vm_compaction_swapper_do_init(); - compaction_swapper_inited = 1; + if (vm_pageout_state.vm_restricted_to_single_processor == TRUE) { + thread_vm_bind_group_add(); + } +#if CONFIG_THREAD_GROUPS + thread_group_vm_add(); +#endif + thread_set_thread_name(current_thread(), "VM_cswap_trigger"); + compaction_swapper_init_now = 0; } lck_mtx_lock_spin_always(c_list_lock); compaction_swap_trigger_thread_awakened++; + compaction_swapper_awakened = 0; - vm_compressor_compact_and_swap(FALSE); + if (compaction_swapper_running == 0) { + compaction_swapper_running = 1; + vm_compressor_compact_and_swap(FALSE); + + compaction_swapper_running = 0; + } assert_wait((event_t)&c_compressor_swap_trigger, THREAD_UNINT); - compaction_swapper_running = 0; - thread_wakeup((event_t)&compaction_swapper_running); + if (compaction_swapper_running == 0) { + thread_wakeup((event_t)&compaction_swapper_running); + } lck_mtx_unlock_always(c_list_lock); - + thread_block((thread_continue_t)vm_compressor_swap_trigger_thread); - + /* NOTREACHED */ } @@ -1715,18 +2715,18 @@ vm_compressor_swap_trigger_thread(void) void vm_compressor_record_warmup_start(void) { - c_segment_t c_seg; + c_segment_t c_seg; lck_mtx_lock_spin_always(c_list_lock); if (first_c_segment_to_warm_generation_id == 0) { if (!queue_empty(&c_age_list_head)) { - c_seg = (c_segment_t)queue_last(&c_age_list_head); first_c_segment_to_warm_generation_id = c_seg->c_generation_id; - } else + } else { first_c_segment_to_warm_generation_id = 0; + } fastwake_recording_in_progress = TRUE; } @@ -1734,22 +2734,21 @@ vm_compressor_record_warmup_start(void) } -void +void vm_compressor_record_warmup_end(void) { - c_segment_t c_seg; + c_segment_t c_seg; lck_mtx_lock_spin_always(c_list_lock); if (fastwake_recording_in_progress == TRUE) { - if (!queue_empty(&c_age_list_head)) { - c_seg = (c_segment_t)queue_last(&c_age_list_head); last_c_segment_to_warm_generation_id = c_seg->c_generation_id; - } else + } else { last_c_segment_to_warm_generation_id = first_c_segment_to_warm_generation_id; + } fastwake_recording_in_progress = FALSE; @@ -1759,13 +2758,13 @@ vm_compressor_record_warmup_end(void) } -#define DELAY_TRIM_ON_WAKE_SECS 4 +#define DELAY_TRIM_ON_WAKE_SECS 25 void vm_compressor_delay_trim(void) { - clock_sec_t sec; - clock_nsec_t nsec; + clock_sec_t sec; + clock_nsec_t nsec; clock_get_system_nanotime(&sec, &nsec); dont_trim_until_ts = sec + DELAY_TRIM_ON_WAKE_SECS; @@ -1776,7 +2775,7 @@ void vm_compressor_do_warmup(void) { lck_mtx_lock_spin_always(c_list_lock); - + if (first_c_segment_to_warm_generation_id == last_c_segment_to_warm_generation_id) { first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0; @@ -1784,58 +2783,82 @@ vm_compressor_do_warmup(void) return; } - if (compaction_swapper_running == 0) { - + if (compaction_swapper_running == 0 && compaction_swapper_awakened == 0) { fastwake_warmup = TRUE; - compaction_swapper_running = 1; + + compaction_swapper_awakened = 1; thread_wakeup((event_t)&c_compressor_swap_trigger); } lck_mtx_unlock_always(c_list_lock); } - void -do_fastwake_warmup(void) +do_fastwake_warmup_all(void) { - uint64_t my_thread_id; - c_segment_t c_seg = NULL; - AbsoluteTime startTime, endTime; - uint64_t nsec; - + lck_mtx_lock_spin_always(c_list_lock); - HIBLOG("vm_compressor_fastwake_warmup (%qd - %qd) - starting\n", first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id); + if (queue_empty(&c_swappedout_list_head) && queue_empty(&c_swappedout_sparse_list_head)) { + lck_mtx_unlock_always(c_list_lock); + return; + } + + fastwake_warmup = TRUE; + + do_fastwake_warmup(&c_swappedout_list_head, TRUE); + + do_fastwake_warmup(&c_swappedout_sparse_list_head, TRUE); + + fastwake_warmup = FALSE; + + lck_mtx_unlock_always(c_list_lock); +} + +void +do_fastwake_warmup(queue_head_t *c_queue, boolean_t consider_all_cseg) +{ + c_segment_t c_seg = NULL; + AbsoluteTime startTime, endTime; + uint64_t nsec; + + + HIBLOG("vm_compressor_fastwake_warmup (%qd - %qd) - starting\n", first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id); clock_get_uptime(&startTime); lck_mtx_unlock_always(c_list_lock); - my_thread_id = current_thread()->thread_id; - proc_set_task_policy_thread(kernel_task, my_thread_id, - TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER2); + proc_set_thread_policy(current_thread(), + TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER2); PAGE_REPLACEMENT_DISALLOWED(TRUE); lck_mtx_lock_spin_always(c_list_lock); - while (!queue_empty(&c_swappedout_list_head) && fastwake_warmup == TRUE) { + while (!queue_empty(c_queue) && fastwake_warmup == TRUE) { + c_seg = (c_segment_t) queue_first(c_queue); - c_seg = (c_segment_t) queue_first(&c_swappedout_list_head); + if (consider_all_cseg == FALSE) { + if (c_seg->c_generation_id < first_c_segment_to_warm_generation_id || + c_seg->c_generation_id > last_c_segment_to_warm_generation_id) { + break; + } - if (c_seg->c_generation_id < first_c_segment_to_warm_generation_id || - c_seg->c_generation_id > last_c_segment_to_warm_generation_id) - break; + if (vm_page_free_count < (AVAILABLE_MEMORY / 4)) { + break; + } + } lck_mtx_lock_spin_always(&c_seg->c_lock); lck_mtx_unlock_always(c_list_lock); - + if (c_seg->c_busy) { PAGE_REPLACEMENT_DISALLOWED(FALSE); c_seg_wait_on_busy(c_seg); PAGE_REPLACEMENT_DISALLOWED(TRUE); } else { - c_seg_swapin(c_seg, TRUE); - - lck_mtx_unlock_always(&c_seg->c_lock); + if (c_seg_swapin(c_seg, TRUE, FALSE) == 0) { + lck_mtx_unlock_always(&c_seg->c_lock); + } c_segment_warmup_count++; PAGE_REPLACEMENT_DISALLOWED(FALSE); @@ -1848,36 +2871,47 @@ do_fastwake_warmup(void) PAGE_REPLACEMENT_DISALLOWED(FALSE); - proc_set_task_policy_thread(kernel_task, my_thread_id, - TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER0); + proc_set_thread_policy(current_thread(), + TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER0); - clock_get_uptime(&endTime); - SUB_ABSOLUTETIME(&endTime, &startTime); - absolutetime_to_nanoseconds(endTime, &nsec); + clock_get_uptime(&endTime); + SUB_ABSOLUTETIME(&endTime, &startTime); + absolutetime_to_nanoseconds(endTime, &nsec); HIBLOG("vm_compressor_fastwake_warmup completed - took %qd msecs\n", nsec / 1000000ULL); lck_mtx_lock_spin_always(c_list_lock); - first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0; + if (consider_all_cseg == FALSE) { + first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0; + } } +int min_csegs_per_major_compaction = DELAYED_COMPACTIONS_PER_PASS; +extern bool vm_swapout_thread_running; +extern boolean_t compressor_store_stop_compaction; void vm_compressor_compact_and_swap(boolean_t flush_all) { - c_segment_t c_seg, c_seg_next; - boolean_t keep_compacting; + c_segment_t c_seg, c_seg_next; + boolean_t keep_compacting, switch_state; + clock_sec_t now; + clock_nsec_t nsec; + mach_timespec_t start_ts, end_ts; + unsigned int number_considered, wanted_cseg_found, yield_after_considered_per_pass, number_yields; + uint64_t bytes_to_free, bytes_freed, delta_usec; + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_START, c_age_count, c_minor_count, c_major_count, vm_page_free_count); if (fastwake_warmup == TRUE) { - uint64_t starting_warmup_count; + uint64_t starting_warmup_count; starting_warmup_count = c_segment_warmup_count; KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 11) | DBG_FUNC_START, c_segment_warmup_count, - first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id, 0, 0); - do_fastwake_warmup(); + first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id, 0, 0); + do_fastwake_warmup(&c_swappedout_list_head, FALSE); KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 11) | DBG_FUNC_END, c_segment_warmup_count, c_segment_warmup_count - starting_warmup_count, 0, 0, 0); fastwake_warmup = FALSE; @@ -1894,12 +2928,27 @@ vm_compressor_compact_and_swap(boolean_t flush_all) vm_compressor_age_swapped_in_segments(flush_all); - - while (!queue_empty(&c_age_list_head) && compaction_swapper_abort == 0) { - + /* + * we only need to grab the timestamp once per + * invocation of this function since the + * timescale we're interested in is measured + * in days + */ + clock_get_system_nanotime(&now, &nsec); + + start_ts.tv_sec = (int) now; + start_ts.tv_nsec = nsec; + delta_usec = 0; + number_considered = 0; + wanted_cseg_found = 0; + number_yields = 0; + bytes_to_free = 0; + bytes_freed = 0; + yield_after_considered_per_pass = MAX(min_csegs_per_major_compaction, DELAYED_COMPACTIONS_PER_PASS); + + while (!queue_empty(&c_age_list_head) && !compaction_swapper_abort && !compressor_store_stop_compaction) { if (hibernate_flushing == TRUE) { - clock_sec_t sec; - clock_nsec_t nsec; + clock_sec_t sec; if (hibernate_should_abort()) { HIBLOG("vm_compressor_flush - hibernate_should_abort returned TRUE\n"); @@ -1909,19 +2958,33 @@ vm_compressor_compact_and_swap(boolean_t flush_all) HIBLOG("vm_compressor_flush - out of swap space\n"); break; } + if (vm_swap_files_pinned() == FALSE) { + HIBLOG("vm_compressor_flush - unpinned swap files\n"); + break; + } + if (hibernate_in_progress_with_pinned_swap == TRUE && + (vm_swapfile_total_segs_alloced == vm_swapfile_total_segs_used)) { + HIBLOG("vm_compressor_flush - out of pinned swap space\n"); + break; + } clock_get_system_nanotime(&sec, &nsec); - + if (sec > hibernate_flushing_deadline) { HIBLOG("vm_compressor_flush - failed to finish before deadline\n"); break; } } - if (c_swapout_count >= C_SWAPOUT_LIMIT) { + if (!vm_swap_out_of_space() && c_swapout_count >= C_SWAPOUT_LIMIT) { + assert_wait_timeout((event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, 100, 1000 * NSEC_PER_USEC); - assert_wait_timeout((event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, 100, 1000*NSEC_PER_USEC); + if (!vm_swapout_thread_running) { + thread_wakeup((event_t)&c_swapout_list_head); + } lck_mtx_unlock_always(c_list_lock); + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 1, c_swapout_count, 0, 0); + thread_block(THREAD_CONTINUE_NULL); lck_mtx_lock_spin_always(c_list_lock); @@ -1933,7 +2996,7 @@ vm_compressor_compact_and_swap(boolean_t flush_all) vm_compressor_age_swapped_in_segments(flush_all); - if (c_swapout_count >= C_SWAPOUT_LIMIT) { + if (!vm_swap_out_of_space() && c_swapout_count >= C_SWAPOUT_LIMIT) { /* * we timed out on the above thread_block * let's loop around and try again @@ -1941,6 +3004,8 @@ vm_compressor_compact_and_swap(boolean_t flush_all) * to do minor compactions to make * more memory available */ + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 2, c_swapout_count, 0, 0); + continue; } @@ -1948,33 +3013,37 @@ vm_compressor_compact_and_swap(boolean_t flush_all) * Swap out segments? */ if (flush_all == FALSE) { - boolean_t needs_to_swap; + boolean_t needs_to_swap; lck_mtx_unlock_always(c_list_lock); needs_to_swap = compressor_needs_to_swap(); lck_mtx_lock_spin_always(c_list_lock); - - if (needs_to_swap == FALSE) + + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 3, needs_to_swap, 0, 0); + + if (needs_to_swap == FALSE) { break; + } } - if (queue_empty(&c_age_list_head)) + if (queue_empty(&c_age_list_head)) { + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 4, c_age_count, 0, 0); break; + } c_seg = (c_segment_t) queue_first(&c_age_list_head); - if (flush_all == TRUE && c_seg->c_generation_id > c_generation_id_flush_barrier) - break; + assert(c_seg->c_state == C_ON_AGE_Q); - if (c_seg->c_filling) { - /* - * we're at or near the head... no more work to do - */ + if (flush_all == TRUE && c_seg->c_generation_id > c_generation_id_flush_barrier) { + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 5, 0, 0, 0); break; } + lck_mtx_lock_spin_always(&c_seg->c_lock); if (c_seg->c_busy) { + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 6, (void*) VM_KERNEL_ADDRPERM(c_seg), 0, 0); lck_mtx_unlock_always(c_list_lock); c_seg_wait_on_busy(c_seg); @@ -1989,32 +3058,56 @@ vm_compressor_compact_and_swap(boolean_t flush_all) * found an empty c_segment and freed it * so go grab the next guy in the queue */ + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 7, 0, 0, 0); + c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_freed_segs++; continue; } /* * Major compaction */ keep_compacting = TRUE; + switch_state = TRUE; while (keep_compacting == TRUE) { - assert(c_seg->c_busy); /* look for another segment to consolidate */ c_seg_next = (c_segment_t) queue_next(&c_seg->c_age_list); - - if (queue_end(&c_age_list_head, (queue_entry_t)c_seg_next)) + + if (queue_end(&c_age_list_head, (queue_entry_t)c_seg_next)) { break; + } + + assert(c_seg_next->c_state == C_ON_AGE_Q); - if (c_seg_major_compact_ok(c_seg, c_seg_next) == FALSE) + number_considered++; + + if (c_seg_major_compact_ok(c_seg, c_seg_next) == FALSE) { break; + } lck_mtx_lock_spin_always(&c_seg_next->c_lock); if (c_seg_next->c_busy) { + /* + * We are going to block for our neighbor. + * If our c_seg is wanted, we should unbusy + * it because we don't know how long we might + * have to block here. + */ + if (c_seg->c_wanted) { + lck_mtx_unlock_always(&c_seg_next->c_lock); + switch_state = FALSE; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].bailed_compactions++; + wanted_cseg_found++; + break; + } lck_mtx_unlock_always(c_list_lock); + + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 8, (void*) VM_KERNEL_ADDRPERM(c_seg_next), 0, 0); + c_seg_wait_on_busy(c_seg_next); lck_mtx_lock_spin_always(c_list_lock); @@ -2023,11 +3116,14 @@ vm_compressor_compact_and_swap(boolean_t flush_all) /* grab that segment */ C_SEG_BUSY(c_seg_next); + bytes_to_free = C_SEG_OFFSET_TO_BYTES(c_seg_next->c_populated_offset); if (c_seg_do_minor_compaction_and_unlock(c_seg_next, FALSE, TRUE, TRUE)) { /* * found an empty c_segment and freed it * so we can't continue to use c_seg_next */ + bytes_freed += bytes_to_free; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_freed_segs++; continue; } @@ -2038,12 +3134,14 @@ vm_compressor_compact_and_swap(boolean_t flush_all) keep_compacting = c_seg_major_compact(c_seg, c_seg_next); + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 9, keep_compacting, 0, 0); + PAGE_REPLACEMENT_DISALLOWED(TRUE); lck_mtx_lock_spin_always(&c_seg_next->c_lock); /* * run a minor compaction on the donor segment - * since we pulled at least some of it's + * since we pulled at least some of it's * data into our target... if we've emptied * it, now is a good time to free it which * c_seg_minor_compaction_and_unlock also takes care of @@ -2051,88 +3149,165 @@ vm_compressor_compact_and_swap(boolean_t flush_all) * by passing TRUE, we ask for c_busy to be cleared * and c_wanted to be taken care of */ - c_seg_minor_compaction_and_unlock(c_seg_next, TRUE); + bytes_to_free = C_SEG_OFFSET_TO_BYTES(c_seg_next->c_populated_offset); + if (c_seg_minor_compaction_and_unlock(c_seg_next, TRUE)) { + bytes_freed += bytes_to_free; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_freed_segs++; + } else { + bytes_to_free -= C_SEG_OFFSET_TO_BYTES(c_seg_next->c_populated_offset); + bytes_freed += bytes_to_free; + } PAGE_REPLACEMENT_DISALLOWED(FALSE); /* relock the list */ lck_mtx_lock_spin_always(c_list_lock); + if (c_seg->c_wanted) { + /* + * Our c_seg is in demand. Let's + * unbusy it and wakeup the waiters + * instead of continuing the compaction + * because we could be in this loop + * for a while. + */ + switch_state = FALSE; + wanted_cseg_found++; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].bailed_compactions++; + break; + } } /* major compaction */ - c_seg_major_compact_stats.wasted_space_in_swapouts += C_SEG_BUFSIZE - c_seg->c_bytes_used; - c_seg_major_compact_stats.count_of_swapouts++; + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 10, number_considered, wanted_cseg_found, 0); lck_mtx_lock_spin_always(&c_seg->c_lock); assert(c_seg->c_busy); - assert(c_seg->c_on_age_q); assert(!c_seg->c_on_minorcompact_q); - queue_remove(&c_age_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_age_q = 0; - c_age_count--; - - if (vm_swap_up == TRUE) { - queue_enter(&c_swapout_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swapout_q = 1; - c_swapout_count++; - } else { - queue_enter(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedout_q = 1; - c_swappedout_count++; + if (switch_state) { + if (VM_CONFIG_SWAP_IS_ACTIVE) { + /* + * This mode of putting a generic c_seg on the swapout list is + * only supported when we have general swapping enabled + */ + c_seg_switch_state(c_seg, C_ON_SWAPOUT_Q, FALSE); + } else { + if ((vm_swapout_ripe_segments == TRUE && c_overage_swapped_count < c_overage_swapped_limit)) { + assert(VM_CONFIG_SWAP_IS_PRESENT); + /* + * we are running compressor sweeps with swap-behind + * make sure the c_seg has aged enough before swapping it + * out... + */ + if ((now - c_seg->c_creation_ts) >= vm_ripe_target_age) { + c_seg->c_overage_swap = TRUE; + c_overage_swapped_count++; + c_seg_switch_state(c_seg, C_ON_SWAPOUT_Q, FALSE); + } + } + } + if (c_seg->c_state == C_ON_AGE_Q) { + /* + * this c_seg didn't get moved to the swapout queue + * so we need to move it out of the way... + * we just did a major compaction on it so put it + * on that queue + */ + c_seg_switch_state(c_seg, C_ON_MAJORCOMPACT_Q, FALSE); + } else { + c_seg_major_compact_stats[c_seg_major_compact_stats_now].wasted_space_in_swapouts += C_SEG_BUFSIZE - c_seg->c_bytes_used; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].count_of_swapouts++; + } } + C_SEG_WAKEUP_DONE(c_seg); lck_mtx_unlock_always(&c_seg->c_lock); if (c_swapout_count) { - lck_mtx_unlock_always(c_list_lock); + /* + * We don't pause/yield here because we will either + * yield below or at the top of the loop with the + * assert_wait_timeout. + */ + if (!vm_swapout_thread_running) { + thread_wakeup((event_t)&c_swapout_list_head); + } + } - thread_wakeup((event_t)&c_swapout_list_head); - - lck_mtx_lock_spin_always(c_list_lock); + if (number_considered >= yield_after_considered_per_pass) { + if (wanted_cseg_found) { + /* + * We stopped major compactions on a c_seg + * that is wanted. We don't know the priority + * of the waiter unfortunately but we are at + * a very high priority and so, just in case + * the waiter is a critical system daemon or + * UI thread, let's give up the CPU in case + * the system is running a few CPU intensive + * tasks. + */ + lck_mtx_unlock_always(c_list_lock); + + mutex_pause(2); /* 100us yield */ + + number_yields++; + + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_NONE, 11, number_considered, number_yields, 0); + + lck_mtx_lock_spin_always(c_list_lock); + } + + number_considered = 0; + wanted_cseg_found = 0; } } -} + clock_get_system_nanotime(&now, &nsec); + end_ts.tv_sec = (int) now; + end_ts.tv_nsec = nsec; + SUB_MACH_TIMESPEC(&end_ts, &start_ts); -static uint32_t no_paging_space_action_in_progress = 0; -extern void memorystatus_send_low_swap_note(void); + delta_usec = (end_ts.tv_sec * USEC_PER_SEC) + (end_ts.tv_nsec / NSEC_PER_USEC) - (number_yields * 100); + delta_usec = MAX(1, delta_usec); /* we could have 0 usec run if conditions weren't right */ -static c_segment_t -c_seg_allocate(c_segment_t *current_chead) -{ - clock_sec_t sec; - clock_nsec_t nsec; - c_segment_t c_seg; - int slotarray; + c_seg_major_compact_stats[c_seg_major_compact_stats_now].bytes_freed_rate_us = (bytes_freed / delta_usec); - if ( (c_seg = *current_chead) == NULL ) { - uint32_t c_segno; + if ((c_seg_major_compact_stats_now + 1) == C_SEG_MAJOR_COMPACT_STATS_MAX) { + c_seg_major_compact_stats_now = 0; + } else { + c_seg_major_compact_stats_now++; + } - if (vm_compressor_low_on_space() || vm_swap_low_on_space()) { + assert(c_seg_major_compact_stats_now < C_SEG_MAJOR_COMPACT_STATS_MAX); - if (no_paging_space_action_in_progress == 0) { + VM_DEBUG_CONSTANT_EVENT(vm_compressor_compact_and_swap, VM_COMPRESSOR_COMPACT_AND_SWAP, DBG_FUNC_END, c_age_count, c_minor_count, c_major_count, vm_page_free_count); +} - if (OSCompareAndSwap(0, 1, (UInt32 *)&no_paging_space_action_in_progress)) { - if (no_paging_space_action()) { - memorystatus_send_low_swap_note(); - } +static c_segment_t +c_seg_allocate(c_segment_t *current_chead) +{ + c_segment_t c_seg; + int min_needed; + int size_to_populate; - no_paging_space_action_in_progress = 0; - } - } - } - KERNEL_DEBUG(0xe0400004 | DBG_FUNC_START, 0, 0, 0, 0, 0); +#if XNU_TARGET_OS_OSX + if (vm_compressor_low_on_space()) { + vm_compressor_take_paging_space_action(); + } +#endif /* XNU_TARGET_OS_OSX */ + + if ((c_seg = *current_chead) == NULL) { + uint32_t c_segno; lck_mtx_lock_spin_always(c_list_lock); while (c_segments_busy == TRUE) { assert_wait((event_t) (&c_segments_busy), THREAD_UNINT); - + lck_mtx_unlock_always(c_list_lock); thread_block(THREAD_CONTINUE_NULL); @@ -2140,117 +3315,143 @@ c_seg_allocate(c_segment_t *current_chead) lck_mtx_lock_spin_always(c_list_lock); } if (c_free_segno_head == (uint32_t)-1) { + uint32_t c_segments_available_new; + uint32_t compressed_pages; + +#if CONFIG_FREEZE + if (freezer_incore_cseg_acct) { + compressed_pages = c_segment_pages_compressed_incore; + } else { + compressed_pages = c_segment_pages_compressed; + } +#else + compressed_pages = c_segment_pages_compressed; +#endif /* CONFIG_FREEZE */ - if (c_segments_available >= c_segments_limit || c_segment_pages_compressed >= c_segment_pages_compressed_limit) { + if (c_segments_available >= c_segments_limit || compressed_pages >= c_segment_pages_compressed_limit) { lck_mtx_unlock_always(c_list_lock); - KERNEL_DEBUG(0xe0400004 | DBG_FUNC_END, 0, 0, 0, 1, 0); - return (NULL); + return NULL; } c_segments_busy = TRUE; lck_mtx_unlock_always(c_list_lock); - kernel_memory_populate(kernel_map, (vm_offset_t)c_segments_next_page, PAGE_SIZE, KMA_KOBJECT); + kernel_memory_populate(compressor_map, (vm_offset_t)c_segments_next_page, + PAGE_SIZE, KMA_KOBJECT, VM_KERN_MEMORY_COMPRESSOR); c_segments_next_page += PAGE_SIZE; - for (c_segno = c_segments_available + 1; c_segno < (c_segments_available + C_SEGMENTS_PER_PAGE); c_segno++) + c_segments_available_new = c_segments_available + C_SEGMENTS_PER_PAGE; + + if (c_segments_available_new > c_segments_limit) { + c_segments_available_new = c_segments_limit; + } + + for (c_segno = c_segments_available + 1; c_segno < c_segments_available_new; c_segno++) { c_segments[c_segno - 1].c_segno = c_segno; + } lck_mtx_lock_spin_always(c_list_lock); c_segments[c_segno - 1].c_segno = c_free_segno_head; c_free_segno_head = c_segments_available; - c_segments_available += C_SEGMENTS_PER_PAGE; + c_segments_available = c_segments_available_new; c_segments_busy = FALSE; thread_wakeup((event_t) (&c_segments_busy)); } c_segno = c_free_segno_head; - c_free_segno_head = c_segments[c_segno].c_segno; + assert(c_segno >= 0 && c_segno < c_segments_limit); - lck_mtx_unlock_always(c_list_lock); - - c_seg = (c_segment_t)zalloc(compressor_segment_zone); - bzero((char *)c_seg, sizeof(struct c_segment)); - - if (kernel_memory_allocate(kernel_map, (vm_offset_t *)(&c_seg->c_store.c_buffer), C_SEG_ALLOCSIZE, 0, KMA_COMPRESSOR | KMA_VAONLY) != KERN_SUCCESS) { - zfree(compressor_segment_zone, c_seg); - - lck_mtx_lock_spin_always(c_list_lock); + c_free_segno_head = (uint32_t)c_segments[c_segno].c_segno; - c_segments[c_segno].c_segno = c_free_segno_head; - c_free_segno_head = c_segno; + /* + * do the rest of the bookkeeping now while we're still behind + * the list lock and grab our generation id now into a local + * so that we can install it once we have the c_seg allocated + */ + c_segment_count++; + if (c_segment_count > c_segment_count_max) { + c_segment_count_max = c_segment_count; + } - lck_mtx_unlock_always(c_list_lock); + lck_mtx_unlock_always(c_list_lock); - KERNEL_DEBUG(0xe0400004 | DBG_FUNC_END, 0, 0, 0, 2, 0); + c_seg = zalloc_flags(compressor_segment_zone, Z_WAITOK | Z_ZERO); - return (NULL); - } - OSAddAtomic64(C_SEG_ALLOCSIZE, &compressor_kvspace_used); + c_seg->c_store.c_buffer = (int32_t *)C_SEG_BUFFER_ADDRESS(c_segno); -#if __i386__ || __x86_64__ - lck_mtx_init(&c_seg->c_lock, &vm_compressor_lck_grp, &vm_compressor_lck_attr); -#else /* __i386__ || __x86_64__ */ - lck_spin_init(&c_seg->c_lock, &vm_compressor_lck_grp, &vm_compressor_lck_attr); -#endif /* __i386__ || __x86_64__ */ - - kernel_memory_populate(kernel_map, (vm_offset_t)(c_seg->c_store.c_buffer), 3 * PAGE_SIZE, KMA_COMPRESSOR); + lck_mtx_init(&c_seg->c_lock, &vm_compressor_lck_grp, LCK_ATTR_NULL); - c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(3 * PAGE_SIZE); - c_seg->c_firstemptyslot = C_SLOT_MAX; + c_seg->c_state = C_IS_EMPTY; + c_seg->c_firstemptyslot = C_SLOT_MAX_INDEX; c_seg->c_mysegno = c_segno; - c_seg->c_filling = 1; lck_mtx_lock_spin_always(c_list_lock); - - c_segment_count++; + c_empty_count++; + c_seg_switch_state(c_seg, C_IS_FILLING, FALSE); c_segments[c_segno].c_seg = c_seg; - - c_seg->c_generation_id = c_generation_id++; - - queue_enter(&c_age_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_age_q = 1; - c_age_count++; - + assert(c_segments[c_segno].c_segno > c_segments_available); lck_mtx_unlock_always(c_list_lock); - clock_get_system_nanotime(&sec, &nsec); - c_seg->c_creation_ts = (uint32_t)sec; - *current_chead = c_seg; - KERNEL_DEBUG(0xe0400004 | DBG_FUNC_END, c_seg, 0, 0, 3, 0); +#if DEVELOPMENT || DEBUG + C_SEG_MAKE_WRITEABLE(c_seg); +#endif } - slotarray = C_SEG_SLOTARRAY_FROM_INDEX(c_seg, c_seg->c_nextslot); + c_seg_alloc_nextslot(c_seg); + + size_to_populate = C_SEG_ALLOCSIZE - C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset); - if (c_seg->c_slots[slotarray] == 0) { - KERNEL_DEBUG(0xe0400008 | DBG_FUNC_START, 0, 0, 0, 0, 0); + if (size_to_populate) { + min_needed = PAGE_SIZE + (C_SEG_ALLOCSIZE - C_SEG_BUFSIZE); - c_seg->c_slots[slotarray] = (struct c_slot *)kalloc(sizeof(struct c_slot) * C_SEG_SLOT_ARRAY_SIZE); + if (C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset - c_seg->c_nextoffset) < (unsigned) min_needed) { + if (size_to_populate > C_SEG_MAX_POPULATE_SIZE) { + size_to_populate = C_SEG_MAX_POPULATE_SIZE; + } + + OSAddAtomic64(size_to_populate / PAGE_SIZE, &vm_pageout_vminfo.vm_compressor_pages_grabbed); - KERNEL_DEBUG(0xe0400008 | DBG_FUNC_END, 0, 0, 0, 0, 0); + kernel_memory_populate(compressor_map, + (vm_offset_t) &c_seg->c_store.c_buffer[c_seg->c_populated_offset], + size_to_populate, + KMA_COMPRESSOR, + VM_KERN_MEMORY_COMPRESSOR); + } else { + size_to_populate = 0; + } } - PAGE_REPLACEMENT_DISALLOWED(TRUE); lck_mtx_lock_spin_always(&c_seg->c_lock); - return (c_seg); -} + if (size_to_populate) { + c_seg->c_populated_offset += C_SEG_BYTES_TO_OFFSET(size_to_populate); + } + return c_seg; +} +#if DEVELOPMENT || DEBUG +#if CONFIG_FREEZE +extern boolean_t memorystatus_freeze_to_memory; +#endif /* CONFIG_FREEZE */ +#endif /* DEVELOPMENT || DEBUG */ static void c_current_seg_filled(c_segment_t c_seg, c_segment_t *current_chead) { - uint32_t unused_bytes; - uint32_t offset_to_depopulate; + uint32_t unused_bytes; + uint32_t offset_to_depopulate; + int new_state = C_ON_AGE_Q; + clock_sec_t sec; + clock_nsec_t nsec; + boolean_t head_insert = FALSE; unused_bytes = trunc_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset - c_seg->c_nextoffset)); if (unused_bytes) { - offset_to_depopulate = C_SEG_BYTES_TO_OFFSET(round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_nextoffset))); /* @@ -2259,21 +3460,96 @@ c_current_seg_filled(c_segment_t c_seg, c_segment_t *current_chead) lck_mtx_unlock_always(&c_seg->c_lock); kernel_memory_depopulate( - kernel_map, + compressor_map, (vm_offset_t) &c_seg->c_store.c_buffer[offset_to_depopulate], unused_bytes, - KMA_COMPRESSOR); + KMA_COMPRESSOR, + VM_KERN_MEMORY_COMPRESSOR); lck_mtx_lock_spin_always(&c_seg->c_lock); c_seg->c_populated_offset = offset_to_depopulate; } - c_seg->c_filling = 0; + assert(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset) <= C_SEG_BUFSIZE); - if (C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) - c_seg_need_delayed_compaction(c_seg); +#if DEVELOPMENT || DEBUG + { + boolean_t c_seg_was_busy = FALSE; - lck_mtx_unlock_always(&c_seg->c_lock); + if (!c_seg->c_busy) { + C_SEG_BUSY(c_seg); + } else { + c_seg_was_busy = TRUE; + } + + lck_mtx_unlock_always(&c_seg->c_lock); + + C_SEG_WRITE_PROTECT(c_seg); + + lck_mtx_lock_spin_always(&c_seg->c_lock); + + if (c_seg_was_busy == FALSE) { + C_SEG_WAKEUP_DONE(c_seg); + } + } +#endif + +#if CONFIG_FREEZE + if (current_chead == (c_segment_t*) &(freezer_context_global.freezer_ctx_chead) && + VM_CONFIG_SWAP_IS_PRESENT && + VM_CONFIG_FREEZER_SWAP_IS_ACTIVE +#if DEVELOPMENT || DEBUG + && !memorystatus_freeze_to_memory +#endif /* DEVELOPMENT || DEBUG */ + ) { + new_state = C_ON_SWAPOUT_Q; + } +#endif /* CONFIG_FREEZE */ + + if (vm_darkwake_mode == TRUE) { + new_state = C_ON_SWAPOUT_Q; + head_insert = TRUE; + } + + clock_get_system_nanotime(&sec, &nsec); + c_seg->c_creation_ts = (uint32_t)sec; + + lck_mtx_lock_spin_always(c_list_lock); + + c_seg->c_generation_id = c_generation_id++; + c_seg_switch_state(c_seg, new_state, head_insert); + +#if CONFIG_FREEZE + if (c_seg->c_state == C_ON_SWAPOUT_Q) { + /* + * darkwake and freezer can't co-exist together + * We'll need to fix this accounting as a start. + */ + assert(vm_darkwake_mode == FALSE); + c_seg_update_task_owner(c_seg, freezer_context_global.freezer_ctx_task); + freezer_context_global.freezer_ctx_swapped_bytes += c_seg->c_bytes_used; + } +#endif /* CONFIG_FREEZE */ + + if (c_seg->c_state == C_ON_AGE_Q && C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) { +#if CONFIG_FREEZE + assert(c_seg->c_task_owner == NULL); +#endif /* CONFIG_FREEZE */ + c_seg_need_delayed_compaction(c_seg, TRUE); + } + + lck_mtx_unlock_always(c_list_lock); + + if (c_seg->c_state == C_ON_SWAPOUT_Q) { + /* + * Darkwake and Freeze configs always + * wake up the swapout thread because + * the compactor thread that normally handles + * it may not be running as much in these + * configs. + */ + thread_wakeup((event_t)&c_swapout_list_head); + } *current_chead = NULL; } @@ -2283,42 +3559,42 @@ c_current_seg_filled(c_segment_t c_seg, c_segment_t *current_chead) * returns with c_seg locked */ void -c_seg_swapin_requeue(c_segment_t c_seg) +c_seg_swapin_requeue(c_segment_t c_seg, boolean_t has_data, boolean_t minor_compact_ok, boolean_t age_on_swapin_q) { - clock_sec_t sec; - clock_nsec_t nsec; + clock_sec_t sec; + clock_nsec_t nsec; clock_get_system_nanotime(&sec, &nsec); lck_mtx_lock_spin_always(c_list_lock); lck_mtx_lock_spin_always(&c_seg->c_lock); - if (c_seg->c_on_swappedout_q) { - queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedout_q = 0; - c_swappedout_count--; - } else { - assert(c_seg->c_on_swappedout_sparse_q); + assert(c_seg->c_busy_swapping); + assert(c_seg->c_busy); - queue_remove(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedout_sparse_q = 0; - c_swappedout_sparse_count--; - } - if (c_seg->c_store.c_buffer) { - queue_enter(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_swappedin_q = 1; - c_swappedin_count++; + c_seg->c_busy_swapping = 0; + + if (c_seg->c_overage_swap == TRUE) { + c_overage_swapped_count--; + c_seg->c_overage_swap = FALSE; } -#if TRACK_BAD_C_SEGMENTS - else { - queue_enter(&c_bad_list_head, c_seg, c_segment_t, c_age_list); - c_seg->c_on_bad_q = 1; - c_bad_count++; + if (has_data == TRUE) { + if (age_on_swapin_q == TRUE) { + c_seg_switch_state(c_seg, C_ON_SWAPPEDIN_Q, FALSE); + } else { + c_seg_switch_state(c_seg, C_ON_AGE_Q, FALSE); + } + + if (minor_compact_ok == TRUE && !c_seg->c_on_minorcompact_q && C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) { + c_seg_need_delayed_compaction(c_seg, TRUE); + } + } else { + c_seg->c_store.c_buffer = (int32_t*) NULL; + c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0); + + c_seg_switch_state(c_seg, C_ON_BAD_Q, FALSE); } -#endif c_seg->c_swappedin_ts = (uint32_t)sec; - c_seg->c_ondisk = 0; - c_seg->c_was_swapped_in = 1; lck_mtx_unlock_always(c_list_lock); } @@ -2326,192 +3602,482 @@ c_seg_swapin_requeue(c_segment_t c_seg) /* - * c_seg has to be locked and is returned locked. + * c_seg has to be locked and is returned locked if the c_seg isn't freed * PAGE_REPLACMENT_DISALLOWED has to be TRUE on entry and is returned TRUE + * c_seg_swapin returns 1 if the c_seg was freed, 0 otherwise */ -void -c_seg_swapin(c_segment_t c_seg, boolean_t force_minor_compaction) +int +c_seg_swapin(c_segment_t c_seg, boolean_t force_minor_compaction, boolean_t age_on_swapin_q) { - vm_offset_t addr = 0; - uint32_t io_size = 0; - uint64_t f_offset; + vm_offset_t addr = 0; + uint32_t io_size = 0; + uint64_t f_offset; + + assert(C_SEG_IS_ONDISK(c_seg)); #if !CHECKSUM_THE_SWAP - if (c_seg->c_ondisk) - c_seg_trim_tail(c_seg); + c_seg_trim_tail(c_seg); #endif io_size = round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset)); f_offset = c_seg->c_store.c_swap_handle; C_SEG_BUSY(c_seg); + c_seg->c_busy_swapping = 1; + + /* + * This thread is likely going to block for I/O. + * Make sure it is ready to run when the I/O completes because + * it needs to clear the busy bit on the c_seg so that other + * waiting threads can make progress too. To do that, boost + * the rwlock_count so that the priority is boosted. + */ + set_thread_rwlock_boost(); lck_mtx_unlock_always(&c_seg->c_lock); - - if (c_seg->c_ondisk) { - PAGE_REPLACEMENT_DISALLOWED(FALSE); + PAGE_REPLACEMENT_DISALLOWED(FALSE); - if (kernel_memory_allocate(kernel_map, &addr, C_SEG_ALLOCSIZE, 0, KMA_COMPRESSOR | KMA_VAONLY) != KERN_SUCCESS) - panic("c_seg_swapin: kernel_memory_allocate failed\n"); + addr = (vm_offset_t)C_SEG_BUFFER_ADDRESS(c_seg->c_mysegno); + c_seg->c_store.c_buffer = (int32_t*) addr; - kernel_memory_populate(kernel_map, addr, io_size, KMA_COMPRESSOR); + kernel_memory_populate(compressor_map, addr, io_size, KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR); - if (vm_swap_get(addr, f_offset, io_size) != KERN_SUCCESS) { - PAGE_REPLACEMENT_DISALLOWED(TRUE); + if (vm_swap_get(c_seg, f_offset, io_size) != KERN_SUCCESS) { + PAGE_REPLACEMENT_DISALLOWED(TRUE); - kernel_memory_depopulate(kernel_map, addr, io_size, KMA_COMPRESSOR); - kmem_free(kernel_map, addr, C_SEG_ALLOCSIZE); + kernel_memory_depopulate(compressor_map, addr, io_size, KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR); - c_seg->c_store.c_buffer = (int32_t*) NULL; - c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0); - } else { - c_seg->c_store.c_buffer = (int32_t*) addr; + c_seg_swapin_requeue(c_seg, FALSE, TRUE, age_on_swapin_q); + } else { #if ENCRYPTED_SWAP - vm_swap_decrypt(c_seg); + vm_swap_decrypt(c_seg); #endif /* ENCRYPTED_SWAP */ #if CHECKSUM_THE_SWAP - if (c_seg->cseg_swap_size != io_size) - panic("swapin size doesn't match swapout size"); + if (c_seg->cseg_swap_size != io_size) { + panic("swapin size doesn't match swapout size"); + } - if (c_seg->cseg_hash != hash_string((char*) c_seg->c_store.c_buffer, (int)io_size)) { - panic("c_seg_swapin - Swap hash mismatch\n"); - } + if (c_seg->cseg_hash != vmc_hash((char*) c_seg->c_store.c_buffer, (int)io_size)) { + panic("c_seg_swapin - Swap hash mismatch\n"); + } #endif /* CHECKSUM_THE_SWAP */ - PAGE_REPLACEMENT_DISALLOWED(TRUE); + PAGE_REPLACEMENT_DISALLOWED(TRUE); - if (force_minor_compaction == TRUE) { - lck_mtx_lock_spin_always(&c_seg->c_lock); - - c_seg_minor_compaction_and_unlock(c_seg, FALSE); - } - OSAddAtomic64(c_seg->c_bytes_used, &compressor_bytes_used); - OSAddAtomic64(C_SEG_ALLOCSIZE, &compressor_kvspace_used); + c_seg_swapin_requeue(c_seg, TRUE, force_minor_compaction == TRUE ? FALSE : TRUE, age_on_swapin_q); + +#if CONFIG_FREEZE + /* + * c_seg_swapin_requeue() returns with the c_seg lock held. + */ + if (!lck_mtx_try_lock_spin_always(c_list_lock)) { + assert(c_seg->c_busy); + + lck_mtx_unlock_always(&c_seg->c_lock); + lck_mtx_lock_spin_always(c_list_lock); + lck_mtx_lock_spin_always(&c_seg->c_lock); } - } - c_seg_swapin_requeue(c_seg); - C_SEG_WAKEUP_DONE(c_seg); -} + if (c_seg->c_task_owner) { + c_seg_update_task_owner(c_seg, NULL); + } + lck_mtx_unlock_always(c_list_lock); -static int -c_compress_page(char *src, c_slot_mapping_t slot_ptr, c_segment_t *current_chead, char *scratch_buf) -{ - int c_size; - int c_rounded_size = 0; - int max_csize; - c_slot_t cs; - c_segment_t c_seg; + OSAddAtomic(c_seg->c_slots_used, &c_segment_pages_compressed_incore); +#endif /* CONFIG_FREEZE */ - KERNEL_DEBUG(0xe0400000 | DBG_FUNC_START, *current_chead, 0, 0, 0, 0); -retry: - if ((c_seg = c_seg_allocate(current_chead)) == NULL) - return (1); + OSAddAtomic64(c_seg->c_bytes_used, &compressor_bytes_used); + + if (force_minor_compaction == TRUE) { + if (c_seg_minor_compaction_and_unlock(c_seg, FALSE)) { + /* + * c_seg was completely empty so it was freed, + * so be careful not to reference it again + * + * Drop the rwlock_count so that the thread priority + * is returned back to where it is supposed to be. + */ + clear_thread_rwlock_boost(); + return 1; + } + + lck_mtx_lock_spin_always(&c_seg->c_lock); + } + } + C_SEG_WAKEUP_DONE(c_seg); + + /* + * Drop the rwlock_count so that the thread priority + * is returned back to where it is supposed to be. + */ + clear_thread_rwlock_boost(); + + return 0; +} + + +static void +c_segment_sv_hash_drop_ref(int hash_indx) +{ + struct c_sv_hash_entry o_sv_he, n_sv_he; + + while (1) { + o_sv_he.he_record = c_segment_sv_hash_table[hash_indx].he_record; + + n_sv_he.he_ref = o_sv_he.he_ref - 1; + n_sv_he.he_data = o_sv_he.he_data; + + if (OSCompareAndSwap64((UInt64)o_sv_he.he_record, (UInt64)n_sv_he.he_record, (UInt64 *) &c_segment_sv_hash_table[hash_indx].he_record) == TRUE) { + if (n_sv_he.he_ref == 0) { + OSAddAtomic(-1, &c_segment_svp_in_hash); + } + break; + } + } +} + + +static int +c_segment_sv_hash_insert(uint32_t data) +{ + int hash_sindx; + int misses; + struct c_sv_hash_entry o_sv_he, n_sv_he; + boolean_t got_ref = FALSE; + + if (data == 0) { + OSAddAtomic(1, &c_segment_svp_zero_compressions); + } else { + OSAddAtomic(1, &c_segment_svp_nonzero_compressions); + } + + hash_sindx = data & C_SV_HASH_MASK; + + for (misses = 0; misses < C_SV_HASH_MAX_MISS; misses++) { + o_sv_he.he_record = c_segment_sv_hash_table[hash_sindx].he_record; + + while (o_sv_he.he_data == data || o_sv_he.he_ref == 0) { + n_sv_he.he_ref = o_sv_he.he_ref + 1; + n_sv_he.he_data = data; + + if (OSCompareAndSwap64((UInt64)o_sv_he.he_record, (UInt64)n_sv_he.he_record, (UInt64 *) &c_segment_sv_hash_table[hash_sindx].he_record) == TRUE) { + if (n_sv_he.he_ref == 1) { + OSAddAtomic(1, &c_segment_svp_in_hash); + } + got_ref = TRUE; + break; + } + o_sv_he.he_record = c_segment_sv_hash_table[hash_sindx].he_record; + } + if (got_ref == TRUE) { + break; + } + hash_sindx++; + + if (hash_sindx == C_SV_HASH_SIZE) { + hash_sindx = 0; + } + } + if (got_ref == FALSE) { + return -1; + } + + return hash_sindx; +} + + +#if RECORD_THE_COMPRESSED_DATA + +static void +c_compressed_record_data(char *src, int c_size) +{ + if ((c_compressed_record_cptr + c_size + 4) >= c_compressed_record_ebuf) { + panic("c_compressed_record_cptr >= c_compressed_record_ebuf"); + } + + *(int *)((void *)c_compressed_record_cptr) = c_size; + + c_compressed_record_cptr += 4; + + memcpy(c_compressed_record_cptr, src, c_size); + c_compressed_record_cptr += c_size; +} +#endif + + +static int +c_compress_page(char *src, c_slot_mapping_t slot_ptr, c_segment_t *current_chead, char *scratch_buf) +{ + int c_size; + int c_rounded_size = 0; + int max_csize; + c_slot_t cs; + c_segment_t c_seg; + + KERNEL_DEBUG(0xe0400000 | DBG_FUNC_START, *current_chead, 0, 0, 0, 0); +retry: + if ((c_seg = c_seg_allocate(current_chead)) == NULL) { + return 1; + } /* * returns with c_seg lock held - * and PAGE_REPLACEMENT_DISALLOWED(TRUE) + * and PAGE_REPLACEMENT_DISALLOWED(TRUE)... + * c_nextslot has been allocated and + * c_store.c_buffer populated */ + assert(c_seg->c_state == C_IS_FILLING); + cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_seg->c_nextslot); + C_SLOT_ASSERT_PACKABLE(slot_ptr); cs->c_packed_ptr = C_SLOT_PACK_PTR(slot_ptr); - assert(slot_ptr == (c_slot_mapping_t)C_SLOT_UNPACK_PTR(cs)); cs->c_offset = c_seg->c_nextoffset; max_csize = C_SEG_BUFSIZE - C_SEG_OFFSET_TO_BYTES((int32_t)cs->c_offset); - if (max_csize > PAGE_SIZE) + if (max_csize > PAGE_SIZE) { max_csize = PAGE_SIZE; - - if (C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset - - c_seg->c_nextoffset) - < (unsigned) max_csize + PAGE_SIZE && - (C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset) - < C_SEG_ALLOCSIZE)) { - lck_mtx_unlock_always(&c_seg->c_lock); - - kernel_memory_populate(kernel_map, - (vm_offset_t) &c_seg->c_store.c_buffer[c_seg->c_populated_offset], - PAGE_SIZE, - KMA_COMPRESSOR); - - lck_mtx_lock_spin_always(&c_seg->c_lock); - - c_seg->c_populated_offset += C_SEG_BYTES_TO_OFFSET(PAGE_SIZE); } #if CHECKSUM_THE_DATA - cs->c_hash_data = hash_string(src, PAGE_SIZE); + cs->c_hash_data = vmc_hash(src, PAGE_SIZE); +#endif + boolean_t incomp_copy = FALSE; + int max_csize_adj = (max_csize - 4); + + if (vm_compressor_algorithm() != VM_COMPRESSOR_DEFAULT_CODEC) { +#if defined(__arm__) || defined(__arm64__) + uint16_t ccodec = CINVALID; + uint32_t inline_popcount; + if (max_csize >= C_SEG_OFFSET_ALIGNMENT_BOUNDARY) { + c_size = metacompressor((const uint8_t *) src, + (uint8_t *) &c_seg->c_store.c_buffer[cs->c_offset], + max_csize_adj, &ccodec, + scratch_buf, &incomp_copy, &inline_popcount); +#if __ARM_WKDM_POPCNT__ + cs->c_inline_popcount = inline_popcount; +#else + assert(inline_popcount == C_SLOT_NO_POPCOUNT); #endif - c_size = WKdm_compress_new((WK_word *)(uintptr_t)src, (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], - (WK_word *)(uintptr_t)scratch_buf, max_csize - 4); - assert(c_size <= (max_csize - 4) && c_size >= -1); +#if C_SEG_OFFSET_ALIGNMENT_BOUNDARY > 4 + if (c_size > max_csize_adj) { + c_size = -1; + } +#endif + } else { + c_size = -1; + } + assert(ccodec == CCWK || ccodec == CCLZ4); + cs->c_codec = ccodec; +#endif + } else { +#if defined(__arm__) || defined(__arm64__) + cs->c_codec = CCWK; +#endif +#if defined(__arm64__) + __unreachable_ok_push + if (PAGE_SIZE == 4096) { + c_size = WKdm_compress_4k((WK_word *)(uintptr_t)src, (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], + (WK_word *)(uintptr_t)scratch_buf, max_csize_adj); + } else { + c_size = WKdm_compress_16k((WK_word *)(uintptr_t)src, (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], + (WK_word *)(uintptr_t)scratch_buf, max_csize_adj); + } + __unreachable_ok_pop +#else + c_size = WKdm_compress_new((const WK_word *)(uintptr_t)src, (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], + (WK_word *)(uintptr_t)scratch_buf, max_csize_adj); +#endif + } + assertf(((c_size <= max_csize_adj) && (c_size >= -1)), + "c_size invalid (%d, %d), cur compressions: %d", c_size, max_csize_adj, c_segment_pages_compressed); if (c_size == -1) { - if (max_csize < PAGE_SIZE) { c_current_seg_filled(c_seg, current_chead); + assert(*current_chead == NULL); + lck_mtx_unlock_always(&c_seg->c_lock); + /* TODO: it may be worth requiring codecs to distinguish + * between incompressible inputs and failures due to + * budget exhaustion. + */ PAGE_REPLACEMENT_DISALLOWED(FALSE); - goto retry; } c_size = PAGE_SIZE; + if (incomp_copy == FALSE) { + memcpy(&c_seg->c_store.c_buffer[cs->c_offset], src, c_size); + } + + OSAddAtomic(1, &c_segment_noncompressible_pages); + } else if (c_size == 0) { + int hash_index; + + /* + * special case - this is a page completely full of a single 32 bit value + */ + hash_index = c_segment_sv_hash_insert(*(uint32_t *)(uintptr_t)src); + + if (hash_index != -1) { + slot_ptr->s_cindx = hash_index; + slot_ptr->s_cseg = C_SV_CSEG_ID; + + OSAddAtomic(1, &c_segment_svp_hash_succeeded); +#if RECORD_THE_COMPRESSED_DATA + c_compressed_record_data(src, 4); +#endif + goto sv_compression; + } + c_size = 4; + memcpy(&c_seg->c_store.c_buffer[cs->c_offset], src, c_size); + + OSAddAtomic(1, &c_segment_svp_hash_failed); } + +#if RECORD_THE_COMPRESSED_DATA + c_compressed_record_data((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size); +#endif #if CHECKSUM_THE_COMPRESSED_DATA - cs->c_hash_compressed_data = hash_string((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size); + cs->c_hash_compressed_data = vmc_hash((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size); +#endif +#if POPCOUNT_THE_COMPRESSED_DATA + cs->c_pop_cdata = vmc_pop((uintptr_t) &c_seg->c_store.c_buffer[cs->c_offset], c_size); #endif c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK; PACK_C_SIZE(cs, c_size); c_seg->c_bytes_used += c_rounded_size; c_seg->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size); + c_seg->c_slots_used++; slot_ptr->s_cindx = c_seg->c_nextslot++; /* would mean "empty slot", so use csegno+1 */ - slot_ptr->s_cseg = c_seg->c_mysegno + 1; + slot_ptr->s_cseg = c_seg->c_mysegno + 1; - if (c_seg->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg->c_nextslot >= C_SLOT_MAX) +sv_compression: + if (c_seg->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg->c_nextslot >= C_SLOT_MAX_INDEX) { c_current_seg_filled(c_seg, current_chead); - else - lck_mtx_unlock_always(&c_seg->c_lock); + assert(*current_chead == NULL); + } + lck_mtx_unlock_always(&c_seg->c_lock); PAGE_REPLACEMENT_DISALLOWED(FALSE); - OSAddAtomic64(c_rounded_size, &compressor_bytes_used); +#if RECORD_THE_COMPRESSED_DATA + if ((c_compressed_record_cptr - c_compressed_record_sbuf) >= C_SEG_ALLOCSIZE) { + c_compressed_record_write(c_compressed_record_sbuf, (int)(c_compressed_record_cptr - c_compressed_record_sbuf)); + c_compressed_record_cptr = c_compressed_record_sbuf; + } +#endif + if (c_size) { + OSAddAtomic64(c_size, &c_segment_compressed_bytes); + OSAddAtomic64(c_rounded_size, &compressor_bytes_used); + } OSAddAtomic64(PAGE_SIZE, &c_segment_input_bytes); - OSAddAtomic64(c_size, &c_segment_compressed_bytes); OSAddAtomic(1, &c_segment_pages_compressed); +#if CONFIG_FREEZE + OSAddAtomic(1, &c_segment_pages_compressed_incore); +#endif /* CONFIG_FREEZE */ OSAddAtomic(1, &sample_period_compression_count); KERNEL_DEBUG(0xe0400000 | DBG_FUNC_END, *current_chead, c_size, c_segment_input_bytes, c_segment_compressed_bytes, 0); - return (0); + return 0; } +static inline void +sv_decompress(int32_t *ddst, int32_t pattern) +{ +// assert(__builtin_constant_p(PAGE_SIZE) != 0); +#if defined(__x86_64__) + memset_word(ddst, pattern, PAGE_SIZE / sizeof(int32_t)); +#elif defined(__arm64__) + assert((PAGE_SIZE % 128) == 0); + if (pattern == 0) { + fill32_dczva((addr64_t)ddst, PAGE_SIZE); + } else { + fill32_nt((addr64_t)ddst, PAGE_SIZE, pattern); + } +#else + size_t i; + + /* Unroll the pattern fill loop 4x to encourage the + * compiler to emit NEON stores, cf. + * Loop autovectorization + * anomalies. + */ + /* * We use separate loops for each PAGE_SIZE + * to allow the autovectorizer to engage, as PAGE_SIZE + * may not be a constant. + */ + + __unreachable_ok_push + if (PAGE_SIZE == 4096) { + for (i = 0; i < (4096U / sizeof(int32_t)); i += 4) { + *ddst++ = pattern; + *ddst++ = pattern; + *ddst++ = pattern; + *ddst++ = pattern; + } + } else { + assert(PAGE_SIZE == 16384); + for (i = 0; i < (int)(16384U / sizeof(int32_t)); i += 4) { + *ddst++ = pattern; + *ddst++ = pattern; + *ddst++ = pattern; + *ddst++ = pattern; + } + } + __unreachable_ok_pop +#endif +} static int c_decompress_page(char *dst, volatile c_slot_mapping_t slot_ptr, int flags, int *zeroslot) { - c_slot_t cs; - c_segment_t c_seg; - int c_indx; - int c_rounded_size; - uint32_t c_size; - int retval = 0; - boolean_t c_seg_has_data = TRUE; - boolean_t c_seg_swappedin = FALSE; - boolean_t need_unlock = TRUE; - boolean_t consider_defragmenting = FALSE; + c_slot_t cs; + c_segment_t c_seg; + uint32_t c_segno; + uint16_t c_indx; + int c_rounded_size; + uint32_t c_size; + int retval = 0; + boolean_t need_unlock = TRUE; + boolean_t consider_defragmenting = FALSE; + boolean_t kdp_mode = FALSE; + + if (__improbable(flags & C_KDP)) { + if (not_in_kdp) { + panic("C_KDP passed to decompress page from outside of debugger context"); + } + + assert((flags & C_KEEP) == C_KEEP); + assert((flags & C_DONT_BLOCK) == C_DONT_BLOCK); + + if ((flags & (C_DONT_BLOCK | C_KEEP)) != (C_DONT_BLOCK | C_KEEP)) { + return -2; + } + + kdp_mode = TRUE; + *zeroslot = 0; + } ReTry: - PAGE_REPLACEMENT_DISALLOWED(TRUE); + if (__probable(!kdp_mode)) { + PAGE_REPLACEMENT_DISALLOWED(TRUE); + } else { + if (kdp_lck_rw_lock_is_acquired_exclusive(&c_master_lock)) { + return -2; + } + } #if HIBERNATION /* @@ -2523,13 +4089,14 @@ ReTry: * to disk... in this state we allow freeing of compressed * pages and must honor the C_DONT_BLOCK case */ - if (dst && decompressions_blocked == TRUE) { + if (__improbable(dst && decompressions_blocked == TRUE)) { if (flags & C_DONT_BLOCK) { - - PAGE_REPLACEMENT_DISALLOWED(FALSE); + if (__probable(!kdp_mode)) { + PAGE_REPLACEMENT_DISALLOWED(FALSE); + } *zeroslot = 0; - return (-2); + return -2; } /* * it's safe to atomically assert and block behind the @@ -2547,272 +4114,611 @@ ReTry: } #endif /* s_cseg is actually "segno+1" */ - c_seg = c_segments[slot_ptr->s_cseg - 1].c_seg; + c_segno = slot_ptr->s_cseg - 1; - lck_mtx_lock_spin_always(&c_seg->c_lock); + if (__improbable(c_segno >= c_segments_available)) { + panic("c_decompress_page: c_segno %d >= c_segments_available %d, slot_ptr(%p), slot_data(%x)", + c_segno, c_segments_available, slot_ptr, *(int *)((void *)slot_ptr)); + } - if (flags & C_DONT_BLOCK) { - if (c_seg->c_busy || (c_seg->c_ondisk && dst)) { + if (__improbable(c_segments[c_segno].c_segno < c_segments_available)) { + panic("c_decompress_page: c_segno %d is free, slot_ptr(%p), slot_data(%x)", + c_segno, slot_ptr, *(int *)((void *)slot_ptr)); + } - retval = -2; + c_seg = c_segments[c_segno].c_seg; + + if (__probable(!kdp_mode)) { + lck_mtx_lock_spin_always(&c_seg->c_lock); + } else { + if (kdp_lck_mtx_lock_spin_is_acquired(&c_seg->c_lock)) { + return -2; + } + } + + assert(c_seg->c_state != C_IS_EMPTY && c_seg->c_state != C_IS_FREE); + + if (dst == NULL && c_seg->c_busy_swapping) { + assert(c_seg->c_busy); + + goto bypass_busy_check; + } + if (flags & C_DONT_BLOCK) { + if (c_seg->c_busy || (C_SEG_IS_ONDISK(c_seg) && dst)) { *zeroslot = 0; + retval = -2; goto done; } } if (c_seg->c_busy) { - PAGE_REPLACEMENT_DISALLOWED(FALSE); c_seg_wait_on_busy(c_seg); goto ReTry; } +bypass_busy_check: + c_indx = slot_ptr->s_cindx; + if (__improbable(c_indx >= c_seg->c_nextslot)) { + panic("c_decompress_page: c_indx %d >= c_nextslot %d, c_seg(%p), slot_ptr(%p), slot_data(%x)", + c_indx, c_seg->c_nextslot, c_seg, slot_ptr, *(int *)((void *)slot_ptr)); + } + cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx); c_size = UNPACK_C_SIZE(cs); + if (__improbable(c_size == 0)) { + panic("c_decompress_page: c_size == 0, c_seg(%p), slot_ptr(%p), slot_data(%x)", + c_seg, slot_ptr, *(int *)((void *)slot_ptr)); + } + c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK; if (dst) { - uint32_t age_of_cseg; - clock_sec_t cur_ts_sec; - clock_nsec_t cur_ts_nsec; - - if (c_seg->c_on_swappedout_q || c_seg->c_on_swappedout_sparse_q) { - if (c_seg->c_ondisk) - c_seg_swappedin = TRUE; - c_seg_swapin(c_seg, FALSE); - } - if (c_seg->c_store.c_buffer == NULL) { - c_seg_has_data = FALSE; - goto c_seg_invalid_data; + uint32_t age_of_cseg; + clock_sec_t cur_ts_sec; + clock_nsec_t cur_ts_nsec; + + if (C_SEG_IS_ONDISK(c_seg)) { +#if CONFIG_FREEZE + if (freezer_incore_cseg_acct) { + if ((c_seg->c_slots_used + c_segment_pages_compressed_incore) >= c_segment_pages_compressed_nearing_limit) { + PAGE_REPLACEMENT_DISALLOWED(FALSE); + lck_mtx_unlock_always(&c_seg->c_lock); + + memorystatus_kill_on_VM_compressor_space_shortage(FALSE /* async */); + + goto ReTry; + } + + uint32_t incore_seg_count = c_segment_count - c_swappedout_count - c_swappedout_sparse_count; + if ((incore_seg_count + 1) >= c_segments_nearing_limit) { + PAGE_REPLACEMENT_DISALLOWED(FALSE); + lck_mtx_unlock_always(&c_seg->c_lock); + + memorystatus_kill_on_VM_compressor_space_shortage(FALSE /* async */); + + goto ReTry; + } + } +#endif /* CONFIG_FREEZE */ + assert(kdp_mode == FALSE); + retval = c_seg_swapin(c_seg, FALSE, TRUE); + assert(retval == 0); + + retval = 1; } + if (c_seg->c_state == C_ON_BAD_Q) { + assert(c_seg->c_store.c_buffer == NULL); + *zeroslot = 0; + + retval = -1; + goto done; + } + +#if POPCOUNT_THE_COMPRESSED_DATA + unsigned csvpop; + uintptr_t csvaddr = (uintptr_t) &c_seg->c_store.c_buffer[cs->c_offset]; + if (cs->c_pop_cdata != (csvpop = vmc_pop(csvaddr, c_size))) { + panic("Compressed data popcount doesn't match original, bit distance: %d %p (phys: %p) %p %p 0x%x 0x%x 0x%x 0x%x", (csvpop - cs->c_pop_cdata), (void *)csvaddr, (void *) kvtophys(csvaddr), c_seg, cs, cs->c_offset, c_size, csvpop, cs->c_pop_cdata); + } +#endif + #if CHECKSUM_THE_COMPRESSED_DATA - if (cs->c_hash_compressed_data != hash_string((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size)) - panic("compressed data doesn't match original"); + unsigned csvhash; + if (cs->c_hash_compressed_data != (csvhash = vmc_hash((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size))) { + panic("Compressed data doesn't match original %p %p %u %u %u", c_seg, cs, c_size, cs->c_hash_compressed_data, csvhash); + } #endif if (c_rounded_size == PAGE_SIZE) { /* * page wasn't compressible... just copy it out */ memcpy(dst, &c_seg->c_store.c_buffer[cs->c_offset], PAGE_SIZE); - } else { - uint32_t my_cpu_no; - char *scratch_buf; + } else if (c_size == 4) { + int32_t data; + int32_t *dptr; /* - * we're behind the c_seg lock held in spin mode - * which means pre-emption is disabled... therefore - * the following sequence is atomic and safe + * page was populated with a single value + * that didn't fit into our fast hash + * so we packed it in as a single non-compressed value + * that we need to populate the page with */ - my_cpu_no = cpu_number(); + dptr = (int32_t *)(uintptr_t)dst; + data = *(int32_t *)(&c_seg->c_store.c_buffer[cs->c_offset]); + sv_decompress(dptr, data); + } else { + uint32_t my_cpu_no; + char *scratch_buf; - assert(my_cpu_no < compressor_cpus); + if (__probable(!kdp_mode)) { + /* + * we're behind the c_seg lock held in spin mode + * which means pre-emption is disabled... therefore + * the following sequence is atomic and safe + */ + my_cpu_no = cpu_number(); + + assert(my_cpu_no < compressor_cpus); + + scratch_buf = &compressor_scratch_bufs[my_cpu_no * vm_compressor_get_decode_scratch_size()]; + } else { + scratch_buf = kdp_compressor_scratch_buf; + } - scratch_buf = &compressor_scratch_bufs[my_cpu_no * WKdm_SCRATCH_BUF_SIZE]; - WKdm_decompress_new((WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], + if (vm_compressor_algorithm() != VM_COMPRESSOR_DEFAULT_CODEC) { +#if defined(__arm__) || defined(__arm64__) + uint16_t c_codec = cs->c_codec; + uint32_t inline_popcount; + if (!metadecompressor((const uint8_t *) &c_seg->c_store.c_buffer[cs->c_offset], + (uint8_t *)dst, c_size, c_codec, (void *)scratch_buf, &inline_popcount)) { + retval = -1; + } else { +#if __ARM_WKDM_POPCNT__ + if (inline_popcount != cs->c_inline_popcount) { + /* + * The codec choice in compression and + * decompression must agree, so there + * should never be a disagreement in + * whether an inline population count + * was performed. + */ + assert(inline_popcount != C_SLOT_NO_POPCOUNT); + assert(cs->c_inline_popcount != C_SLOT_NO_POPCOUNT); + printf("decompression failure from physical region %llx+%05x: popcount mismatch (%d != %d)\n", + (unsigned long long)kvtophys((uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset]), c_size, + inline_popcount, + cs->c_inline_popcount); + retval = -1; + } +#else + assert(inline_popcount == C_SLOT_NO_POPCOUNT); +#endif /* __ARM_WKDM_POPCNT__ */ + } +#endif + } else { +#if defined(__arm64__) + __unreachable_ok_push + if (PAGE_SIZE == 4096) { + WKdm_decompress_4k((WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], (WK_word *)(uintptr_t)dst, (WK_word *)(uintptr_t)scratch_buf, c_size); + } else { + WKdm_decompress_16k((WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], + (WK_word *)(uintptr_t)dst, (WK_word *)(uintptr_t)scratch_buf, c_size); + } + __unreachable_ok_pop +#else + WKdm_decompress_new((WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset], + (WK_word *)(uintptr_t)dst, (WK_word *)(uintptr_t)scratch_buf, c_size); +#endif + } } #if CHECKSUM_THE_DATA - if (cs->c_hash_data != hash_string(dst, PAGE_SIZE)) - panic("decompressed data doesn't match original"); + if (cs->c_hash_data != vmc_hash(dst, PAGE_SIZE)) { +#if defined(__arm__) || defined(__arm64__) + int32_t *dinput = &c_seg->c_store.c_buffer[cs->c_offset]; + panic("decompressed data doesn't match original cs: %p, hash: 0x%x, offset: %d, c_size: %d, c_rounded_size: %d, codec: %d, header: 0x%x 0x%x 0x%x", cs, cs->c_hash_data, cs->c_offset, c_size, c_rounded_size, cs->c_codec, *dinput, *(dinput + 1), *(dinput + 2)); +#else + panic("decompressed data doesn't match original cs: %p, hash: %d, offset: 0x%x, c_size: %d", cs, cs->c_hash_data, cs->c_offset, c_size); #endif - if (!c_seg->c_was_swapped_in) { - + } +#endif + if (c_seg->c_swappedin_ts == 0 && !kdp_mode) { clock_get_system_nanotime(&cur_ts_sec, &cur_ts_nsec); age_of_cseg = (uint32_t)cur_ts_sec - c_seg->c_creation_ts; - - if (age_of_cseg < DECOMPRESSION_SAMPLE_MAX_AGE) + if (age_of_cseg < DECOMPRESSION_SAMPLE_MAX_AGE) { OSAddAtomic(1, &age_of_decompressions_during_sample_period[age_of_cseg]); - else + } else { OSAddAtomic(1, &overage_decompressions_during_sample_period); + } OSAddAtomic(1, &sample_period_decompression_count); } - } else { - if (c_seg->c_store.c_buffer == NULL) - c_seg_has_data = FALSE; } -c_seg_invalid_data: +#if CONFIG_FREEZE + else { + /* + * We are freeing an uncompressed page from this c_seg and so balance the ledgers. + */ + if (C_SEG_IS_ONDISK(c_seg)) { + /* + * The compression sweep feature will push out anonymous pages to disk + * without going through the freezer path and so those c_segs, while + * swapped out, won't have an owner. + */ + if (c_seg->c_task_owner) { + task_update_frozen_to_swap_acct(c_seg->c_task_owner, PAGE_SIZE_64, DEBIT_FROM_SWAP); + } - if (c_seg_has_data == TRUE) { - if (c_seg_swappedin == TRUE) - retval = 1; - else - retval = 0; - } else - retval = -1; + /* + * We are freeing a page in swap without swapping it in. We bump the in-core + * count here to simulate a swapin of a page so that we can accurately + * decrement it below. + */ + OSAddAtomic(1, &c_segment_pages_compressed_incore); + } + } +#endif /* CONFIG_FREEZE */ if (flags & C_KEEP) { *zeroslot = 0; goto done; } + assert(kdp_mode == FALSE); + c_seg->c_bytes_unused += c_rounded_size; c_seg->c_bytes_used -= c_rounded_size; + + assert(c_seg->c_slots_used); + c_seg->c_slots_used--; + PACK_C_SIZE(cs, 0); - if (c_indx < c_seg->c_firstemptyslot) + if (c_indx < c_seg->c_firstemptyslot) { c_seg->c_firstemptyslot = c_indx; + } OSAddAtomic(-1, &c_segment_pages_compressed); +#if CONFIG_FREEZE + OSAddAtomic(-1, &c_segment_pages_compressed_incore); + assertf(c_segment_pages_compressed_incore >= 0, "-ve incore count %p 0x%x", c_seg, c_segment_pages_compressed_incore); +#endif /* CONFIG_FREEZE */ - if (c_seg_has_data == TRUE && !c_seg->c_ondisk) { + if (c_seg->c_state != C_ON_BAD_Q && !(C_SEG_IS_ONDISK(c_seg))) { /* - * c_ondisk == TRUE can occur when we're doing a + * C_SEG_IS_ONDISK == TRUE can occur when we're doing a * free of a compressed page (i.e. dst == NULL) */ OSAddAtomic64(-c_rounded_size, &compressor_bytes_used); } - if (!c_seg->c_filling) { + if (c_seg->c_busy_swapping) { + /* + * bypass case for c_busy_swapping... + * let the swapin/swapout paths deal with putting + * the c_seg on the minor compaction queue if needed + */ + assert(c_seg->c_busy); + goto done; + } + assert(!c_seg->c_busy); + + if (c_seg->c_state != C_IS_FILLING) { if (c_seg->c_bytes_used == 0) { - if (!c_seg->c_ondisk) { - int pages_populated; + if (!(C_SEG_IS_ONDISK(c_seg))) { + int pages_populated; pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE; c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0); if (pages_populated) { + assert(c_seg->c_state != C_ON_BAD_Q); assert(c_seg->c_store.c_buffer != NULL); C_SEG_BUSY(c_seg); lck_mtx_unlock_always(&c_seg->c_lock); - kernel_memory_depopulate(kernel_map, (vm_offset_t) c_seg->c_store.c_buffer, pages_populated * PAGE_SIZE, KMA_COMPRESSOR); + kernel_memory_depopulate(compressor_map, + (vm_offset_t) c_seg->c_store.c_buffer, + pages_populated * PAGE_SIZE, KMA_COMPRESSOR, VM_KERN_MEMORY_COMPRESSOR); lck_mtx_lock_spin_always(&c_seg->c_lock); C_SEG_WAKEUP_DONE(c_seg); } - if (!c_seg->c_on_minorcompact_q && !c_seg->c_on_swapout_q) - c_seg_need_delayed_compaction(c_seg); - } else - assert(c_seg->c_on_swappedout_sparse_q); - + if (!c_seg->c_on_minorcompact_q && c_seg->c_state != C_ON_SWAPOUT_Q && c_seg->c_state != C_ON_SWAPIO_Q) { + c_seg_need_delayed_compaction(c_seg, FALSE); + } + } else { + if (c_seg->c_state != C_ON_SWAPPEDOUTSPARSE_Q) { + c_seg_move_to_sparse_list(c_seg); + consider_defragmenting = TRUE; + } + } } else if (c_seg->c_on_minorcompact_q) { + assert(c_seg->c_state != C_ON_BAD_Q); + assert(!C_SEG_IS_ON_DISK_OR_SOQ(c_seg)); - if (C_SEG_INCORE_IS_SPARSE(c_seg)) { + if (C_SEG_SHOULD_MINORCOMPACT_NOW(c_seg)) { c_seg_try_minor_compaction_and_unlock(c_seg); need_unlock = FALSE; } - } else if (!c_seg->c_ondisk) { - - if (c_seg_has_data == TRUE && !c_seg->c_on_swapout_q && C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) { - c_seg_need_delayed_compaction(c_seg); + } else if (!(C_SEG_IS_ONDISK(c_seg))) { + if (c_seg->c_state != C_ON_BAD_Q && c_seg->c_state != C_ON_SWAPOUT_Q && c_seg->c_state != C_ON_SWAPIO_Q && + C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) { + c_seg_need_delayed_compaction(c_seg, FALSE); } - } else if (!c_seg->c_on_swappedout_sparse_q && C_SEG_ONDISK_IS_SPARSE(c_seg)) { - + } else if (c_seg->c_state != C_ON_SWAPPEDOUTSPARSE_Q && C_SEG_ONDISK_IS_SPARSE(c_seg)) { c_seg_move_to_sparse_list(c_seg); consider_defragmenting = TRUE; } } done: - if (need_unlock == TRUE) + if (__improbable(kdp_mode)) { + return retval; + } + + if (need_unlock == TRUE) { lck_mtx_unlock_always(&c_seg->c_lock); + } PAGE_REPLACEMENT_DISALLOWED(FALSE); - if (consider_defragmenting == TRUE) - vm_swap_consider_defragmenting(); + if (consider_defragmenting == TRUE) { + vm_swap_consider_defragmenting(VM_SWAP_FLAGS_NONE); + } +#if !XNU_TARGET_OS_OSX + if ((c_minor_count && COMPRESSOR_NEEDS_TO_MINOR_COMPACT()) || vm_compressor_needs_to_major_compact()) { + vm_wake_compactor_swapper(); + } +#endif /* !XNU_TARGET_OS_OSX */ - return (retval); + return retval; } int vm_compressor_get(ppnum_t pn, int *slot, int flags) { - char *dst; - int zeroslot = 1; - int retval; + c_slot_mapping_t slot_ptr; + char *dst; + int zeroslot = 1; + int retval; -#if __x86_64__ - dst = PHYSMAP_PTOV((uint64_t)pn << (uint64_t)PAGE_SHIFT); -#else -#error "unsupported architecture" -#endif + dst = pmap_map_compressor_page(pn); + slot_ptr = (c_slot_mapping_t)slot; + + assert(dst != NULL); - retval = c_decompress_page(dst, (c_slot_mapping_t)slot, flags, &zeroslot); + if (slot_ptr->s_cseg == C_SV_CSEG_ID) { + int32_t data; + int32_t *dptr; + + /* + * page was populated with a single value + * that found a home in our hash table + * grab that value from the hash and populate the page + * that we need to populate the page with + */ + dptr = (int32_t *)(uintptr_t)dst; + data = c_segment_sv_hash_table[slot_ptr->s_cindx].he_data; + sv_decompress(dptr, data); + if (!(flags & C_KEEP)) { + c_segment_sv_hash_drop_ref(slot_ptr->s_cindx); + + OSAddAtomic(-1, &c_segment_pages_compressed); +#if CONFIG_FREEZE + OSAddAtomic(-1, &c_segment_pages_compressed_incore); + assertf(c_segment_pages_compressed_incore >= 0, "-ve incore count 0x%x", c_segment_pages_compressed_incore); +#endif /* CONFIG_FREEZE */ + *slot = 0; + } + if (data) { + OSAddAtomic(1, &c_segment_svp_nonzero_decompressions); + } else { + OSAddAtomic(1, &c_segment_svp_zero_decompressions); + } + + pmap_unmap_compressor_page(pn, dst); + return 0; + } + + retval = c_decompress_page(dst, slot_ptr, flags, &zeroslot); /* * zeroslot will be set to 0 by c_decompress_page if (flags & C_KEEP) - * or (flags & C_DONT_BLOCK) and we found 'c_busy' or 'c_ondisk' set + * or (flags & C_DONT_BLOCK) and we found 'c_busy' or 'C_SEG_IS_ONDISK' to be TRUE */ if (zeroslot) { *slot = 0; } + + pmap_unmap_compressor_page(pn, dst); + /* * returns 0 if we successfully decompressed a page from a segment already in memory * returns 1 if we had to first swap in the segment, before successfully decompressing the page * returns -1 if we encountered an error swapping in the segment - decompression failed - * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' or 'c_ondisk' set + * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' or 'C_SEG_IS_ONDISK' to be true */ - return (retval); + return retval; } +#if DEVELOPMENT || DEBUG + +void +vm_compressor_inject_error(int *slot) +{ + c_slot_mapping_t slot_ptr = (c_slot_mapping_t)slot; + + /* No error detection for single-value compression. */ + if (slot_ptr->s_cseg == C_SV_CSEG_ID) { + printf("%s(): cannot inject errors in SV-compressed pages\n", __func__ ); + return; + } + + /* s_cseg is actually "segno+1" */ + const uint32_t c_segno = slot_ptr->s_cseg - 1; + + assert(c_segno < c_segments_available); + assert(c_segments[c_segno].c_segno >= c_segments_available); + + const c_segment_t c_seg = c_segments[c_segno].c_seg; + + PAGE_REPLACEMENT_DISALLOWED(TRUE); + + lck_mtx_lock_spin_always(&c_seg->c_lock); + assert(c_seg->c_state != C_IS_EMPTY && c_seg->c_state != C_IS_FREE); + + const uint16_t c_indx = slot_ptr->s_cindx; + assert(c_indx < c_seg->c_nextslot); + + /* + * To safely make this segment temporarily writable, we need to mark + * the segment busy, which allows us to release the segment lock. + */ + while (c_seg->c_busy) { + c_seg_wait_on_busy(c_seg); + lck_mtx_lock_spin_always(&c_seg->c_lock); + } + C_SEG_BUSY(c_seg); + + bool already_writable = (c_seg->c_state == C_IS_FILLING); + if (!already_writable) { + /* + * Protection update must be performed preemptibly, so temporarily drop + * the lock. Having set c_busy will prevent most other concurrent + * operations. + */ + lck_mtx_unlock_always(&c_seg->c_lock); + C_SEG_MAKE_WRITEABLE(c_seg); + lck_mtx_lock_spin_always(&c_seg->c_lock); + } + + /* + * Once we've released the lock following our c_state == C_IS_FILLING check, + * c_current_seg_filled() can (re-)write-protect the segment. However, it + * will transition from C_IS_FILLING before releasing the c_seg lock, so we + * can detect this by re-checking after we've reobtained the lock. + */ + if (already_writable && c_seg->c_state != C_IS_FILLING) { + lck_mtx_unlock_always(&c_seg->c_lock); + C_SEG_MAKE_WRITEABLE(c_seg); + lck_mtx_lock_spin_always(&c_seg->c_lock); + already_writable = false; + /* Segment can't be freed while c_busy is set. */ + assert(c_seg->c_state != C_IS_FILLING); + } + + c_slot_t cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx); + int32_t *data = &c_seg->c_store.c_buffer[cs->c_offset]; + /* assume that the compressed data holds at least one int32_t */ + assert(UNPACK_C_SIZE(cs) > sizeof(*data)); + /* + * This bit is known to be in the payload of a MISS packet resulting from + * the pattern used in the test pattern from decompression_failure.c. + * Flipping it should result in many corrupted bits in the test page. + */ + data[0] ^= 0x00000100; + if (!already_writable) { + lck_mtx_unlock_always(&c_seg->c_lock); + C_SEG_WRITE_PROTECT(c_seg); + lck_mtx_lock_spin_always(&c_seg->c_lock); + } + + C_SEG_WAKEUP_DONE(c_seg); + lck_mtx_unlock_always(&c_seg->c_lock); + + PAGE_REPLACEMENT_DISALLOWED(FALSE); +} + +#endif /* DEVELOPMENT || DEBUG */ int vm_compressor_free(int *slot, int flags) { - int zeroslot = 1; - int retval; + c_slot_mapping_t slot_ptr; + int zeroslot = 1; + int retval; assert(flags == 0 || flags == C_DONT_BLOCK); - retval = c_decompress_page(NULL, (c_slot_mapping_t)slot, flags, &zeroslot); + slot_ptr = (c_slot_mapping_t)slot; + + if (slot_ptr->s_cseg == C_SV_CSEG_ID) { + c_segment_sv_hash_drop_ref(slot_ptr->s_cindx); + OSAddAtomic(-1, &c_segment_pages_compressed); +#if CONFIG_FREEZE + OSAddAtomic(-1, &c_segment_pages_compressed_incore); + assertf(c_segment_pages_compressed_incore >= 0, "-ve incore count 0x%x", c_segment_pages_compressed_incore); +#endif /* CONFIG_FREEZE */ + + *slot = 0; + return 0; + } + retval = c_decompress_page(NULL, slot_ptr, flags, &zeroslot); /* * returns 0 if we successfully freed the specified compressed page * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' set */ - if (retval == 0) + if (retval == 0) { *slot = 0; + } else { + assert(retval == -2); + } - return (retval); + return retval; } int vm_compressor_put(ppnum_t pn, int *slot, void **current_chead, char *scratch_buf) { - char *src; - int retval; + char *src; + int retval; + + src = pmap_map_compressor_page(pn); + assert(src != NULL); -#if __x86_64__ - src = PHYSMAP_PTOV((uint64_t)pn << (uint64_t)PAGE_SHIFT); -#else -#error "unsupported architecture" -#endif retval = c_compress_page(src, (c_slot_mapping_t)slot, (c_segment_t *)current_chead, scratch_buf); + pmap_unmap_compressor_page(pn, src); - return (retval); + return retval; } void vm_compressor_transfer( - int *dst_slot_p, - int *src_slot_p) + int *dst_slot_p, + int *src_slot_p) { - c_slot_mapping_t dst_slot, src_slot; - c_segment_t c_seg; - int c_indx; - c_slot_t cs; + c_slot_mapping_t dst_slot, src_slot; + c_segment_t c_seg; + uint16_t c_indx; + c_slot_t cs; - dst_slot = (c_slot_mapping_t) dst_slot_p; src_slot = (c_slot_mapping_t) src_slot_p; + if (src_slot->s_cseg == C_SV_CSEG_ID) { + *dst_slot_p = *src_slot_p; + *src_slot_p = 0; + return; + } + dst_slot = (c_slot_mapping_t) dst_slot_p; Retry: PAGE_REPLACEMENT_DISALLOWED(TRUE); /* get segment for src_slot */ - c_seg = c_segments[src_slot->s_cseg -1].c_seg; + c_seg = c_segments[src_slot->s_cseg - 1].c_seg; /* lock segment */ lck_mtx_lock_spin_always(&c_seg->c_lock); /* wait if it's busy */ - if (c_seg->c_busy) { + if (c_seg->c_busy && !c_seg->c_busy_swapping) { PAGE_REPLACEMENT_DISALLOWED(FALSE); c_seg_wait_on_busy(c_seg); goto Retry; @@ -2821,6 +4727,7 @@ Retry: c_indx = src_slot->s_cindx; cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx); /* point the c_slot back to dst_slot instead of src_slot */ + C_SLOT_ASSERT_PACKABLE(dst_slot); cs->c_packed_ptr = C_SLOT_PACK_PTR(dst_slot); /* transfer */ *dst_slot_p = *src_slot_p; @@ -2828,3 +4735,274 @@ Retry: lck_mtx_unlock_always(&c_seg->c_lock); PAGE_REPLACEMENT_DISALLOWED(FALSE); } + +#if CONFIG_FREEZE + +int freezer_finished_filling = 0; + +void +vm_compressor_finished_filling( + void **current_chead) +{ + c_segment_t c_seg; + + if ((c_seg = *(c_segment_t *)current_chead) == NULL) { + return; + } + + assert(c_seg->c_state == C_IS_FILLING); + + lck_mtx_lock_spin_always(&c_seg->c_lock); + + c_current_seg_filled(c_seg, (c_segment_t *)current_chead); + + lck_mtx_unlock_always(&c_seg->c_lock); + + freezer_finished_filling++; +} + + +/* + * This routine is used to transfer the compressed chunks from + * the c_seg/cindx pointed to by slot_p into a new c_seg headed + * by the current_chead and a new cindx within that c_seg. + * + * Currently, this routine is only used by the "freezer backed by + * compressor with swap" mode to create a series of c_segs that + * only contain compressed data belonging to one task. So, we + * move a task's previously compressed data into a set of new + * c_segs which will also hold the task's yet to be compressed data. + */ + +kern_return_t +vm_compressor_relocate( + void **current_chead, + int *slot_p) +{ + c_slot_mapping_t slot_ptr; + c_slot_mapping_t src_slot; + uint32_t c_rounded_size; + uint32_t c_size; + uint16_t dst_slot; + c_slot_t c_dst; + c_slot_t c_src; + uint16_t c_indx; + c_segment_t c_seg_dst = NULL; + c_segment_t c_seg_src = NULL; + kern_return_t kr = KERN_SUCCESS; + + + src_slot = (c_slot_mapping_t) slot_p; + + if (src_slot->s_cseg == C_SV_CSEG_ID) { + /* + * no need to relocate... this is a page full of a single + * value which is hashed to a single entry not contained + * in a c_segment_t + */ + return kr; + } + +Relookup_dst: + c_seg_dst = c_seg_allocate((c_segment_t *)current_chead); + /* + * returns with c_seg lock held + * and PAGE_REPLACEMENT_DISALLOWED(TRUE)... + * c_nextslot has been allocated and + * c_store.c_buffer populated + */ + if (c_seg_dst == NULL) { + /* + * Out of compression segments? + */ + kr = KERN_RESOURCE_SHORTAGE; + goto out; + } + + assert(c_seg_dst->c_busy == 0); + + C_SEG_BUSY(c_seg_dst); + + dst_slot = c_seg_dst->c_nextslot; + + lck_mtx_unlock_always(&c_seg_dst->c_lock); + +Relookup_src: + c_seg_src = c_segments[src_slot->s_cseg - 1].c_seg; + + assert(c_seg_dst != c_seg_src); + + lck_mtx_lock_spin_always(&c_seg_src->c_lock); + + if (C_SEG_IS_ON_DISK_OR_SOQ(c_seg_src) || + c_seg_src->c_state == C_IS_FILLING) { + /* + * Skip this page if :- + * a) the src c_seg is already on-disk (or on its way there) + * A "thaw" can mark a process as eligible for + * another freeze cycle without bringing any of + * its swapped out c_segs back from disk (because + * that is done on-demand). + * Or, this page may be mapped elsewhere in the task's map, + * and we may have marked it for swap already. + * + * b) Or, the src c_seg is being filled by the compressor + * thread. We don't want the added latency of waiting for + * this c_seg in the freeze path and so we skip it. + */ + + PAGE_REPLACEMENT_DISALLOWED(FALSE); + + lck_mtx_unlock_always(&c_seg_src->c_lock); + + c_seg_src = NULL; + + goto out; + } + + if (c_seg_src->c_busy) { + PAGE_REPLACEMENT_DISALLOWED(FALSE); + c_seg_wait_on_busy(c_seg_src); + + c_seg_src = NULL; + + PAGE_REPLACEMENT_DISALLOWED(TRUE); + + goto Relookup_src; + } + + C_SEG_BUSY(c_seg_src); + + lck_mtx_unlock_always(&c_seg_src->c_lock); + + PAGE_REPLACEMENT_DISALLOWED(FALSE); + + /* find the c_slot */ + c_indx = src_slot->s_cindx; + + c_src = C_SEG_SLOT_FROM_INDEX(c_seg_src, c_indx); + + c_size = UNPACK_C_SIZE(c_src); + + assert(c_size); + + if (c_size > (uint32_t)(C_SEG_BUFSIZE - C_SEG_OFFSET_TO_BYTES((int32_t)c_seg_dst->c_nextoffset))) { + /* + * This segment is full. We need a new one. + */ + + PAGE_REPLACEMENT_DISALLOWED(TRUE); + + lck_mtx_lock_spin_always(&c_seg_src->c_lock); + C_SEG_WAKEUP_DONE(c_seg_src); + lck_mtx_unlock_always(&c_seg_src->c_lock); + + c_seg_src = NULL; + + lck_mtx_lock_spin_always(&c_seg_dst->c_lock); + + assert(c_seg_dst->c_busy); + assert(c_seg_dst->c_state == C_IS_FILLING); + assert(!c_seg_dst->c_on_minorcompact_q); + + c_current_seg_filled(c_seg_dst, (c_segment_t *)current_chead); + assert(*current_chead == NULL); + + C_SEG_WAKEUP_DONE(c_seg_dst); + + lck_mtx_unlock_always(&c_seg_dst->c_lock); + + c_seg_dst = NULL; + + PAGE_REPLACEMENT_DISALLOWED(FALSE); + + goto Relookup_dst; + } + + c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot); + + memcpy(&c_seg_dst->c_store.c_buffer[c_seg_dst->c_nextoffset], &c_seg_src->c_store.c_buffer[c_src->c_offset], c_size); + /* + * Is platform alignment actually necessary since wkdm aligns its output? + */ + c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK; + + cslot_copy(c_dst, c_src); + c_dst->c_offset = c_seg_dst->c_nextoffset; + + if (c_seg_dst->c_firstemptyslot == c_seg_dst->c_nextslot) { + c_seg_dst->c_firstemptyslot++; + } + + c_seg_dst->c_slots_used++; + c_seg_dst->c_nextslot++; + c_seg_dst->c_bytes_used += c_rounded_size; + c_seg_dst->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size); + + + PACK_C_SIZE(c_src, 0); + + c_seg_src->c_bytes_used -= c_rounded_size; + c_seg_src->c_bytes_unused += c_rounded_size; + + assert(c_seg_src->c_slots_used); + c_seg_src->c_slots_used--; + + if (c_indx < c_seg_src->c_firstemptyslot) { + c_seg_src->c_firstemptyslot = c_indx; + } + + c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, dst_slot); + + PAGE_REPLACEMENT_ALLOWED(TRUE); + slot_ptr = C_SLOT_UNPACK_PTR(c_dst); + /* would mean "empty slot", so use csegno+1 */ + slot_ptr->s_cseg = c_seg_dst->c_mysegno + 1; + slot_ptr->s_cindx = dst_slot; + + PAGE_REPLACEMENT_ALLOWED(FALSE); + +out: + if (c_seg_src) { + lck_mtx_lock_spin_always(&c_seg_src->c_lock); + + C_SEG_WAKEUP_DONE(c_seg_src); + + if (c_seg_src->c_bytes_used == 0 && c_seg_src->c_state != C_IS_FILLING) { + if (!c_seg_src->c_on_minorcompact_q) { + c_seg_need_delayed_compaction(c_seg_src, FALSE); + } + } + + lck_mtx_unlock_always(&c_seg_src->c_lock); + } + + if (c_seg_dst) { + PAGE_REPLACEMENT_DISALLOWED(TRUE); + + lck_mtx_lock_spin_always(&c_seg_dst->c_lock); + + if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX_INDEX) { + /* + * Nearing or exceeded maximum slot and offset capacity. + */ + assert(c_seg_dst->c_busy); + assert(c_seg_dst->c_state == C_IS_FILLING); + assert(!c_seg_dst->c_on_minorcompact_q); + + c_current_seg_filled(c_seg_dst, (c_segment_t *)current_chead); + assert(*current_chead == NULL); + } + + C_SEG_WAKEUP_DONE(c_seg_dst); + + lck_mtx_unlock_always(&c_seg_dst->c_lock); + + c_seg_dst = NULL; + + PAGE_REPLACEMENT_DISALLOWED(FALSE); + } + + return kr; +} +#endif /* CONFIG_FREEZE */