]> git.saurik.com Git - apple/xnu.git/blame - osfmk/vm/vm_compressor.c
xnu-2782.1.97.tar.gz
[apple/xnu.git] / osfmk / vm / vm_compressor.c
CommitLineData
39236c6e
A
1/*
2 * Copyright (c) 2000-2013 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28
29#include <vm/vm_compressor.h>
fe8ab488
A
30
31#if CONFIG_PHANTOM_CACHE
32#include <vm/vm_phantom_cache.h>
33#endif
34
39236c6e
A
35#include <vm/vm_map.h>
36#include <vm/vm_pageout.h>
37#include <vm/memory_object.h>
38#include <mach/mach_host.h> /* for host_info() */
39#include <kern/ledger.h>
40
41#include <default_pager/default_pager_alerts.h>
42#include <default_pager/default_pager_object_server.h>
43
44#include <IOKit/IOHibernatePrivate.h>
45
46/*
47 * vm_compressor_mode has a heirarchy of control to set its value.
48 * boot-args are checked first, then device-tree, and finally
49 * the default value that is defined below. See vm_fault_init() for
50 * the boot-arg & device-tree code.
51 */
52
53extern ipc_port_t min_pages_trigger_port;
54extern lck_mtx_t paging_segments_lock;
55#define PSL_LOCK() lck_mtx_lock(&paging_segments_lock)
56#define PSL_UNLOCK() lck_mtx_unlock(&paging_segments_lock)
57
58
59int vm_compressor_mode = VM_PAGER_COMPRESSOR_WITH_SWAP;
60int vm_scale = 16;
61
62
63int vm_compression_limit = 0;
64
65extern boolean_t vm_swap_up;
66extern void vm_pageout_io_throttle(void);
67
68#if CHECKSUM_THE_DATA || CHECKSUM_THE_SWAP || CHECKSUM_THE_COMPRESSED_DATA
69extern unsigned int hash_string(char *cp, int len);
70#endif
71
fe8ab488 72
39236c6e
A
73struct c_slot {
74 uint64_t c_offset:C_SEG_OFFSET_BITS,
75 c_size:12,
76 c_packed_ptr:36;
77#if CHECKSUM_THE_DATA
78 unsigned int c_hash_data;
79#endif
80#if CHECKSUM_THE_COMPRESSED_DATA
81 unsigned int c_hash_compressed_data;
82#endif
83
84};
85
fe8ab488 86#define UNPACK_C_SIZE(cs) ((cs->c_size == (PAGE_SIZE-1)) ? PAGE_SIZE : cs->c_size)
39236c6e
A
87#define PACK_C_SIZE(cs, size) (cs->c_size = ((size == PAGE_SIZE) ? PAGE_SIZE - 1 : size))
88
89
90struct c_slot_mapping {
91 uint32_t s_cseg:22, /* segment number + 1 */
92 s_cindx:10; /* index in the segment */
93};
94
95typedef struct c_slot_mapping *c_slot_mapping_t;
96
97
98union c_segu {
99 c_segment_t c_seg;
100 uint32_t c_segno;
101};
102
103
104
105#define C_SLOT_PACK_PTR(ptr) (((uintptr_t)ptr - (uintptr_t) VM_MIN_KERNEL_AND_KEXT_ADDRESS) >> 2)
106#define C_SLOT_UNPACK_PTR(cslot) ((uintptr_t)(cslot->c_packed_ptr << 2) + (uintptr_t) VM_MIN_KERNEL_AND_KEXT_ADDRESS)
107
108
109uint32_t c_segment_count = 0;
110
111uint64_t c_generation_id = 0;
112uint64_t c_generation_id_flush_barrier;
113
114
115#define HIBERNATE_FLUSHING_SECS_TO_COMPLETE 120
116
117boolean_t hibernate_no_swapspace = FALSE;
118clock_sec_t hibernate_flushing_deadline = 0;
119
120
121#if TRACK_BAD_C_SEGMENTS
122queue_head_t c_bad_list_head;
123uint32_t c_bad_count = 0;
124#endif
125
126queue_head_t c_age_list_head;
127queue_head_t c_swapout_list_head;
128queue_head_t c_swappedin_list_head;
129queue_head_t c_swappedout_list_head;
130queue_head_t c_swappedout_sparse_list_head;
131
132uint32_t c_age_count = 0;
133uint32_t c_swapout_count = 0;
134uint32_t c_swappedin_count = 0;
135uint32_t c_swappedout_count = 0;
136uint32_t c_swappedout_sparse_count = 0;
137
138queue_head_t c_minor_list_head;
139uint32_t c_minor_count = 0;
140
141union c_segu *c_segments;
142caddr_t c_segments_next_page;
143boolean_t c_segments_busy;
144uint32_t c_segments_available;
145uint32_t c_segments_limit;
fe8ab488 146uint32_t c_segments_nearing_limit;
39236c6e
A
147uint32_t c_segment_pages_compressed;
148uint32_t c_segment_pages_compressed_limit;
fe8ab488 149uint32_t c_segment_pages_compressed_nearing_limit;
39236c6e
A
150uint32_t c_free_segno_head = (uint32_t)-1;
151
152uint32_t vm_compressor_minorcompact_threshold_divisor = 10;
153uint32_t vm_compressor_majorcompact_threshold_divisor = 10;
154uint32_t vm_compressor_unthrottle_threshold_divisor = 10;
155uint32_t vm_compressor_catchup_threshold_divisor = 10;
156
157#define C_SEGMENTS_PER_PAGE (PAGE_SIZE / sizeof(union c_segu))
158
159
160lck_grp_attr_t vm_compressor_lck_grp_attr;
161lck_attr_t vm_compressor_lck_attr;
162lck_grp_t vm_compressor_lck_grp;
163
164
165#if __i386__ || __x86_64__
166lck_mtx_t *c_list_lock;
167#else /* __i386__ || __x86_64__ */
168lck_spin_t *c_list_lock;
169#endif /* __i386__ || __x86_64__ */
170
171lck_rw_t c_master_lock;
fe8ab488 172boolean_t decompressions_blocked = FALSE;
39236c6e
A
173
174zone_t compressor_segment_zone;
175int c_compressor_swap_trigger = 0;
176
177uint32_t compressor_cpus;
178char *compressor_scratch_bufs;
179
180
181clock_sec_t start_of_sample_period_sec = 0;
182clock_nsec_t start_of_sample_period_nsec = 0;
183clock_sec_t start_of_eval_period_sec = 0;
184clock_nsec_t start_of_eval_period_nsec = 0;
185uint32_t sample_period_decompression_count = 0;
186uint32_t sample_period_compression_count = 0;
187uint32_t last_eval_decompression_count = 0;
188uint32_t last_eval_compression_count = 0;
189
190#define DECOMPRESSION_SAMPLE_MAX_AGE (60 * 30)
191
192uint32_t swapout_target_age = 0;
193uint32_t age_of_decompressions_during_sample_period[DECOMPRESSION_SAMPLE_MAX_AGE];
194uint32_t overage_decompressions_during_sample_period = 0;
195
196void do_fastwake_warmup(void);
197boolean_t fastwake_warmup = FALSE;
198boolean_t fastwake_recording_in_progress = FALSE;
199clock_sec_t dont_trim_until_ts = 0;
200
201uint64_t c_segment_warmup_count;
202uint64_t first_c_segment_to_warm_generation_id = 0;
203uint64_t last_c_segment_to_warm_generation_id = 0;
204boolean_t hibernate_flushing = FALSE;
205
fe8ab488
A
206int64_t c_segment_input_bytes __attribute__((aligned(8))) = 0;
207int64_t c_segment_compressed_bytes __attribute__((aligned(8))) = 0;
208int64_t compressor_bytes_used __attribute__((aligned(8))) = 0;
209uint64_t compressor_kvspace_used __attribute__((aligned(8))) = 0;
210uint64_t compressor_kvwaste_limit = 0;
39236c6e
A
211
212static boolean_t compressor_needs_to_swap(void);
213static void vm_compressor_swap_trigger_thread(void);
214static void vm_compressor_do_delayed_compactions(boolean_t);
215static void vm_compressor_compact_and_swap(boolean_t);
216static void vm_compressor_age_swapped_in_segments(boolean_t);
39236c6e
A
217
218boolean_t vm_compressor_low_on_space(void);
219
220void compute_swapout_target_age(void);
221
222boolean_t c_seg_major_compact(c_segment_t, c_segment_t);
223boolean_t c_seg_major_compact_ok(c_segment_t, c_segment_t);
224
225int c_seg_minor_compaction_and_unlock(c_segment_t, boolean_t);
226int c_seg_do_minor_compaction_and_unlock(c_segment_t, boolean_t, boolean_t, boolean_t);
227void c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg);
228void c_seg_need_delayed_compaction(c_segment_t);
229
230void c_seg_move_to_sparse_list(c_segment_t);
231void c_seg_insert_into_q(queue_head_t *, c_segment_t);
232
233boolean_t c_seg_try_free(c_segment_t);
234void c_seg_free(c_segment_t);
235void c_seg_free_locked(c_segment_t);
236
237
238uint64_t vm_available_memory(void);
fe8ab488 239uint64_t vm_compressor_pages_compressed(void);
39236c6e
A
240
241extern unsigned int dp_pages_free, dp_pages_reserve;
242
243uint64_t
244vm_available_memory(void)
245{
246 return (((uint64_t)AVAILABLE_NON_COMPRESSED_MEMORY) * PAGE_SIZE_64);
247}
248
249
fe8ab488
A
250uint64_t
251vm_compressor_pages_compressed(void)
252{
253 return (c_segment_pages_compressed * PAGE_SIZE_64);
254}
255
256
39236c6e
A
257boolean_t
258vm_compression_available(void)
259{
260 if ( !(COMPRESSED_PAGER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE))
261 return (FALSE);
262
263 if (c_segments_available >= c_segments_limit || c_segment_pages_compressed >= c_segment_pages_compressed_limit)
264 return (FALSE);
265
266 return (TRUE);
267}
268
269
270boolean_t
271vm_compressor_low_on_space(void)
272{
fe8ab488
A
273 if ((c_segment_pages_compressed > c_segment_pages_compressed_nearing_limit) ||
274 (c_segment_count > c_segments_nearing_limit))
39236c6e
A
275 return (TRUE);
276
277 return (FALSE);
278}
279
280
281int
fe8ab488 282vm_wants_task_throttled(task_t task)
39236c6e 283{
fe8ab488
A
284 if (task == kernel_task)
285 return (0);
286
39236c6e 287 if (vm_compressor_mode == COMPRESSED_PAGER_IS_ACTIVE || vm_compressor_mode == DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) {
fe8ab488
A
288 if ((vm_compressor_low_on_space() || HARD_THROTTLE_LIMIT_REACHED()) &&
289 (unsigned int)pmap_compressed(task->map->pmap) > (c_segment_pages_compressed / 4))
39236c6e
A
290 return (1);
291 } else {
fe8ab488
A
292 if (((dp_pages_free + dp_pages_reserve < 2000) && VM_DYNAMIC_PAGING_ENABLED(memory_manager_default)) &&
293 get_task_resident_size(task) > (((AVAILABLE_NON_COMPRESSED_MEMORY) * PAGE_SIZE) / 5))
39236c6e
A
294 return (1);
295 }
296 return (0);
297}
298
299
300void
301vm_compressor_init_locks(void)
302{
303 lck_grp_attr_setdefault(&vm_compressor_lck_grp_attr);
304 lck_grp_init(&vm_compressor_lck_grp, "vm_compressor", &vm_compressor_lck_grp_attr);
305 lck_attr_setdefault(&vm_compressor_lck_attr);
306
307 lck_rw_init(&c_master_lock, &vm_compressor_lck_grp, &vm_compressor_lck_attr);
39236c6e
A
308}
309
310
311void
312vm_decompressor_lock(void)
313{
fe8ab488
A
314 PAGE_REPLACEMENT_ALLOWED(TRUE);
315
316 decompressions_blocked = TRUE;
317
318 PAGE_REPLACEMENT_ALLOWED(FALSE);
39236c6e
A
319}
320
321void
322vm_decompressor_unlock(void)
323{
fe8ab488
A
324 PAGE_REPLACEMENT_ALLOWED(TRUE);
325
326 decompressions_blocked = FALSE;
327
328 PAGE_REPLACEMENT_ALLOWED(FALSE);
39236c6e 329
fe8ab488 330 thread_wakeup((event_t)&decompressions_blocked);
39236c6e
A
331}
332
333
334
335void
336vm_compressor_init(void)
337{
338 thread_t thread;
fe8ab488
A
339 struct c_slot cs_dummy;
340 c_slot_t cs = &cs_dummy;
341
342 /*
343 * ensure that any pointer that gets created from
344 * the vm_page zone can be packed properly
345 */
346 cs->c_packed_ptr = C_SLOT_PACK_PTR(zone_map_min_address);
347
348 if (C_SLOT_UNPACK_PTR(cs) != (uintptr_t)zone_map_min_address)
349 panic("C_SLOT_UNPACK_PTR failed on zone_map_min_address - %p", (void *)zone_map_min_address);
350
351 cs->c_packed_ptr = C_SLOT_PACK_PTR(zone_map_max_address);
352
353 if (C_SLOT_UNPACK_PTR(cs) != (uintptr_t)zone_map_max_address)
354 panic("C_SLOT_UNPACK_PTR failed on zone_map_max_address - %p", (void *)zone_map_max_address);
355
39236c6e
A
356
357 assert((C_SEGMENTS_PER_PAGE * sizeof(union c_segu)) == PAGE_SIZE);
358
359 PE_parse_boot_argn("vm_compression_limit", &vm_compression_limit, sizeof (vm_compression_limit));
360
361 if (max_mem <= (3ULL * 1024ULL * 1024ULL * 1024ULL)) {
362 vm_compressor_minorcompact_threshold_divisor = 11;
363 vm_compressor_majorcompact_threshold_divisor = 13;
364 vm_compressor_unthrottle_threshold_divisor = 20;
365 vm_compressor_catchup_threshold_divisor = 35;
366 } else {
367 vm_compressor_minorcompact_threshold_divisor = 20;
368 vm_compressor_majorcompact_threshold_divisor = 25;
369 vm_compressor_unthrottle_threshold_divisor = 35;
370 vm_compressor_catchup_threshold_divisor = 50;
371 }
372 /*
373 * vm_page_init_lck_grp is now responsible for calling vm_compressor_init_locks
374 * c_master_lock needs to be available early so that "vm_page_find_contiguous" can
375 * use PAGE_REPLACEMENT_ALLOWED to coordinate with the compressor.
376 */
377
378#if __i386__ || __x86_64__
379 c_list_lock = lck_mtx_alloc_init(&vm_compressor_lck_grp, &vm_compressor_lck_attr);
380#else /* __i386__ || __x86_64__ */
381 c_list_lock = lck_spin_alloc_init(&vm_compressor_lck_grp, &vm_compressor_lck_attr);
382#endif /* __i386__ || __x86_64__ */
383
384#if TRACK_BAD_C_SEGMENTS
385 queue_init(&c_bad_list_head);
386#endif
387 queue_init(&c_age_list_head);
388 queue_init(&c_minor_list_head);
389 queue_init(&c_swapout_list_head);
390 queue_init(&c_swappedin_list_head);
391 queue_init(&c_swappedout_list_head);
392 queue_init(&c_swappedout_sparse_list_head);
393
394 compressor_segment_zone = zinit(sizeof (struct c_segment),
395 128000 * sizeof (struct c_segment),
396 8192, "compressor_segment");
397 zone_change(compressor_segment_zone, Z_CALLERACCT, FALSE);
398 zone_change(compressor_segment_zone, Z_NOENCRYPT, TRUE);
399
400
401 c_free_segno_head = -1;
402 c_segments_available = 0;
403
404 if (vm_compression_limit == 0) {
405 c_segment_pages_compressed_limit = (uint32_t)((max_mem / PAGE_SIZE)) * vm_scale;
406
407#define OLD_SWAP_LIMIT (1024 * 1024 * 16)
408#define MAX_SWAP_LIMIT (1024 * 1024 * 128)
409
410 if (c_segment_pages_compressed_limit > (OLD_SWAP_LIMIT))
411 c_segment_pages_compressed_limit = OLD_SWAP_LIMIT;
412
413 if (c_segment_pages_compressed_limit < (uint32_t)(max_mem / PAGE_SIZE_64))
414 c_segment_pages_compressed_limit = (uint32_t)(max_mem / PAGE_SIZE_64);
415 } else {
416 if (vm_compression_limit < MAX_SWAP_LIMIT)
417 c_segment_pages_compressed_limit = vm_compression_limit;
418 else
419 c_segment_pages_compressed_limit = MAX_SWAP_LIMIT;
420 }
421 if ((c_segments_limit = c_segment_pages_compressed_limit / (C_SEG_BUFSIZE / PAGE_SIZE)) > C_SEG_MAX_LIMIT)
422 c_segments_limit = C_SEG_MAX_LIMIT;
423
fe8ab488
A
424 c_segment_pages_compressed_nearing_limit = (c_segment_pages_compressed_limit * 98) / 100;
425 c_segments_nearing_limit = (c_segments_limit * 98) / 100;
426
427 compressor_kvwaste_limit = (vm_map_max(kernel_map) - vm_map_min(kernel_map)) / 16;
428
39236c6e
A
429 c_segments_busy = FALSE;
430
431 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)
432 panic("vm_compressor_init: kernel_memory_allocate failed\n");
433
434 c_segments_next_page = (caddr_t)c_segments;
435
436 {
437 host_basic_info_data_t hinfo;
438 mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT;
439
440#define BSD_HOST 1
441 host_info((host_t)BSD_HOST, HOST_BASIC_INFO, (host_info_t)&hinfo, &count);
442
443 compressor_cpus = hinfo.max_cpus;
444
445 compressor_scratch_bufs = kalloc(compressor_cpus * WKdm_SCRATCH_BUF_SIZE);
446 }
447
448 if (kernel_thread_start_priority((thread_continue_t)vm_compressor_swap_trigger_thread, NULL,
449 BASEPRI_PREEMPT - 1, &thread) != KERN_SUCCESS) {
450 panic("vm_compressor_swap_trigger_thread: create failed");
451 }
452 thread->options |= TH_OPT_VMPRIV;
453
454 thread_deallocate(thread);
455
456 assert(default_pager_init_flag == 0);
457
458 if (vm_pageout_internal_start() != KERN_SUCCESS) {
459 panic("vm_compressor_init: Failed to start the internal pageout thread.\n");
460 }
461
fe8ab488
A
462 if ((vm_compressor_mode == VM_PAGER_COMPRESSOR_WITH_SWAP) ||
463 (vm_compressor_mode == VM_PAGER_FREEZER_COMPRESSOR_WITH_SWAP)) {
464 vm_compressor_swap_init();
465 }
466
39236c6e
A
467#if CONFIG_FREEZE
468 memorystatus_freeze_enabled = TRUE;
469#endif /* CONFIG_FREEZE */
470
471 default_pager_init_flag = 1;
472
473 vm_page_reactivate_all_throttled();
474}
475
476
477#if VALIDATE_C_SEGMENTS
478
479static void
480c_seg_validate(c_segment_t c_seg, boolean_t must_be_compact)
481{
482 int c_indx;
483 int32_t bytes_used;
484 int32_t bytes_unused;
485 uint32_t c_rounded_size;
486 uint32_t c_size;
487 c_slot_t cs;
488
489 if (c_seg->c_firstemptyslot < c_seg->c_nextslot) {
490 c_indx = c_seg->c_firstemptyslot;
491 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
492
493 if (cs == NULL)
494 panic("c_seg_validate: no slot backing c_firstemptyslot");
495
496 if (cs->c_size)
497 panic("c_seg_validate: c_firstemptyslot has non-zero size (%d)\n", cs->c_size);
498 }
499 bytes_used = 0;
500 bytes_unused = 0;
501
502 for (c_indx = 0; c_indx < c_seg->c_nextslot; c_indx++) {
503
504 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
505
506 c_size = UNPACK_C_SIZE(cs);
507
508 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
509
510 bytes_used += c_rounded_size;
511
512#if CHECKSUM_THE_COMPRESSED_DATA
513 if (c_size && cs->c_hash_compressed_data != hash_string((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size))
514 panic("compressed data doesn't match original");
515#endif
516 }
517
518 if (bytes_used != c_seg->c_bytes_used)
519 panic("c_seg_validate: bytes_used mismatch - found %d, segment has %d\n", bytes_used, c_seg->c_bytes_used);
520
521 if (c_seg->c_bytes_used > C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset))
522 panic("c_seg_validate: c_bytes_used > c_nextoffset - c_nextoffset = %d, c_bytes_used = %d\n",
523 (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used);
524
525 if (must_be_compact) {
526 if (c_seg->c_bytes_used != C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset))
527 panic("c_seg_validate: c_bytes_used doesn't match c_nextoffset - c_nextoffset = %d, c_bytes_used = %d\n",
528 (int32_t)C_SEG_OFFSET_TO_BYTES((int32_t)c_seg->c_nextoffset), c_seg->c_bytes_used);
529 }
530}
531
532#endif
533
534
535void
536c_seg_need_delayed_compaction(c_segment_t c_seg)
537{
538 boolean_t clear_busy = FALSE;
539
540 if ( !lck_mtx_try_lock_spin_always(c_list_lock)) {
fe8ab488 541 C_SEG_BUSY(c_seg);
39236c6e
A
542
543 lck_mtx_unlock_always(&c_seg->c_lock);
544 lck_mtx_lock_spin_always(c_list_lock);
545 lck_mtx_lock_spin_always(&c_seg->c_lock);
546
547 clear_busy = TRUE;
548 }
549 if (!c_seg->c_on_minorcompact_q && !c_seg->c_ondisk && !c_seg->c_on_swapout_q) {
550 queue_enter(&c_minor_list_head, c_seg, c_segment_t, c_list);
551 c_seg->c_on_minorcompact_q = 1;
552 c_minor_count++;
553 }
554 lck_mtx_unlock_always(c_list_lock);
555
556 if (clear_busy == TRUE)
557 C_SEG_WAKEUP_DONE(c_seg);
558}
559
560
561unsigned int c_seg_moved_to_sparse_list = 0;
562
563void
564c_seg_move_to_sparse_list(c_segment_t c_seg)
565{
566 boolean_t clear_busy = FALSE;
567
568 if ( !lck_mtx_try_lock_spin_always(c_list_lock)) {
fe8ab488 569 C_SEG_BUSY(c_seg);
39236c6e
A
570
571 lck_mtx_unlock_always(&c_seg->c_lock);
572 lck_mtx_lock_spin_always(c_list_lock);
573 lck_mtx_lock_spin_always(&c_seg->c_lock);
574
575 clear_busy = TRUE;
576 }
577 assert(c_seg->c_ondisk);
578 assert(c_seg->c_on_swappedout_q);
579 assert(!c_seg->c_on_swappedout_sparse_q);
580
581 queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
582 c_seg->c_on_swappedout_q = 0;
583 c_swappedout_count--;
584
585 c_seg_insert_into_q(&c_swappedout_sparse_list_head, c_seg);
586 c_seg->c_on_swappedout_sparse_q = 1;
587 c_swappedout_sparse_count++;
588
589 c_seg_moved_to_sparse_list++;
590
591 lck_mtx_unlock_always(c_list_lock);
592
593 if (clear_busy == TRUE)
594 C_SEG_WAKEUP_DONE(c_seg);
595}
596
597
598void
599c_seg_insert_into_q(queue_head_t *qhead, c_segment_t c_seg)
600{
601 c_segment_t c_seg_next;
602
603 if (queue_empty(qhead)) {
604 queue_enter(qhead, c_seg, c_segment_t, c_age_list);
605 } else {
606 c_seg_next = (c_segment_t)queue_first(qhead);
607
608 while (TRUE) {
609
610 if (c_seg->c_generation_id < c_seg_next->c_generation_id) {
611 queue_insert_before(qhead, c_seg, c_seg_next, c_segment_t, c_age_list);
612 break;
613 }
614 c_seg_next = (c_segment_t) queue_next(&c_seg_next->c_age_list);
615
616 if (queue_end(qhead, (queue_entry_t) c_seg_next)) {
617 queue_enter(qhead, c_seg, c_segment_t, c_age_list);
618 break;
619 }
620 }
621 }
622}
623
624
625int try_minor_compaction_failed = 0;
626int try_minor_compaction_succeeded = 0;
627
628void
629c_seg_try_minor_compaction_and_unlock(c_segment_t c_seg)
630{
631
632 assert(c_seg->c_on_minorcompact_q);
633 /*
634 * c_seg is currently on the delayed minor compaction
635 * queue and we have c_seg locked... if we can get the
636 * c_list_lock w/o blocking (if we blocked we could deadlock
637 * because the lock order is c_list_lock then c_seg's lock)
638 * we'll pull it from the delayed list and free it directly
639 */
640 if ( !lck_mtx_try_lock_spin_always(c_list_lock)) {
641 /*
642 * c_list_lock is held, we need to bail
643 */
644 try_minor_compaction_failed++;
645
646 lck_mtx_unlock_always(&c_seg->c_lock);
647 } else {
648 try_minor_compaction_succeeded++;
649
fe8ab488 650 C_SEG_BUSY(c_seg);
39236c6e
A
651 c_seg_do_minor_compaction_and_unlock(c_seg, TRUE, FALSE, FALSE);
652 }
653}
654
655
656int
657c_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)
658{
659 int c_seg_freed;
660
661 assert(c_seg->c_busy);
662
663 if (!c_seg->c_on_minorcompact_q) {
664 if (clear_busy == TRUE)
665 C_SEG_WAKEUP_DONE(c_seg);
666
667 lck_mtx_unlock_always(&c_seg->c_lock);
668
669 return (0);
670 }
671 queue_remove(&c_minor_list_head, c_seg, c_segment_t, c_list);
672 c_seg->c_on_minorcompact_q = 0;
673 c_minor_count--;
674
675 lck_mtx_unlock_always(c_list_lock);
676
677 if (disallow_page_replacement == TRUE) {
678 lck_mtx_unlock_always(&c_seg->c_lock);
679
680 PAGE_REPLACEMENT_DISALLOWED(TRUE);
681
682 lck_mtx_lock_spin_always(&c_seg->c_lock);
683 }
684 c_seg_freed = c_seg_minor_compaction_and_unlock(c_seg, clear_busy);
685
686 if (disallow_page_replacement == TRUE)
687 PAGE_REPLACEMENT_DISALLOWED(FALSE);
688
689 if (need_list_lock == TRUE)
690 lck_mtx_lock_spin_always(c_list_lock);
691
692 return (c_seg_freed);
693}
694
695
696void
697c_seg_wait_on_busy(c_segment_t c_seg)
698{
699 c_seg->c_wanted = 1;
700 assert_wait((event_t) (c_seg), THREAD_UNINT);
701
702 lck_mtx_unlock_always(&c_seg->c_lock);
703 thread_block(THREAD_CONTINUE_NULL);
704}
705
706
707
708int try_free_succeeded = 0;
709int try_free_failed = 0;
710
711boolean_t
712c_seg_try_free(c_segment_t c_seg)
713{
714 /*
715 * c_seg is currently on the delayed minor compaction
716 * or the spapped out sparse queue and we have c_seg locked...
717 * if we can get the c_list_lock w/o blocking (if we blocked we
718 * could deadlock because the lock order is c_list_lock then c_seg's lock)
719 * we'll pull it from the appropriate queue and free it
720 */
721 if ( !lck_mtx_try_lock_spin_always(c_list_lock)) {
722 /*
723 * c_list_lock is held, we need to bail
724 */
725 try_free_failed++;
726 return (FALSE);
727 }
728 if (c_seg->c_on_minorcompact_q) {
729 queue_remove(&c_minor_list_head, c_seg, c_segment_t, c_list);
730 c_seg->c_on_minorcompact_q = 0;
731 c_minor_count--;
732 } else {
733 assert(c_seg->c_on_swappedout_sparse_q);
734
735 /*
736 * c_seg_free_locked will remove it from the swappedout sparse list
737 */
738 }
739 if (!c_seg->c_busy_swapping)
fe8ab488 740 C_SEG_BUSY(c_seg);
39236c6e
A
741
742 c_seg_free_locked(c_seg);
743
744 try_free_succeeded++;
745
746 return (TRUE);
747}
748
749
750void
751c_seg_free(c_segment_t c_seg)
752{
fe8ab488 753 assert(c_seg->c_busy);
39236c6e
A
754
755 lck_mtx_unlock_always(&c_seg->c_lock);
756 lck_mtx_lock_spin_always(c_list_lock);
757 lck_mtx_lock_spin_always(&c_seg->c_lock);
758
759 c_seg_free_locked(c_seg);
760}
761
762
763void
764c_seg_free_locked(c_segment_t c_seg)
765{
766 int segno, i;
767 int pages_populated;
768 int32_t *c_buffer = NULL;
769 uint64_t c_swap_handle;
770
771 assert(!c_seg->c_on_minorcompact_q);
772
773 if (c_seg->c_on_age_q) {
774 queue_remove(&c_age_list_head, c_seg, c_segment_t, c_age_list);
775 c_seg->c_on_age_q = 0;
776 c_age_count--;
777 } else if (c_seg->c_on_swappedin_q) {
778 queue_remove(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list);
779 c_seg->c_on_swappedin_q = 0;
780 c_swappedin_count--;
781 } else if (c_seg->c_on_swapout_q) {
782 queue_remove(&c_swapout_list_head, c_seg, c_segment_t, c_age_list);
783 c_seg->c_on_swapout_q = 0;
784 c_swapout_count--;
785 thread_wakeup((event_t)&compaction_swapper_running);
786 } else if (c_seg->c_on_swappedout_q) {
787 queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
788 c_seg->c_on_swappedout_q = 0;
789 c_swappedout_count--;
790 } else if (c_seg->c_on_swappedout_sparse_q) {
791 queue_remove(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list);
792 c_seg->c_on_swappedout_sparse_q = 0;
793 c_swappedout_sparse_count--;
794 }
795#if TRACK_BAD_C_SEGMENTS
796 else if (c_seg->c_on_bad_q) {
797 queue_remove(&c_bad_list_head, c_seg, c_segment_t, c_age_list);
798 c_seg->c_on_bad_q = 0;
799 c_bad_count--;
800 }
801#endif
802 segno = c_seg->c_mysegno;
803 c_segments[segno].c_segno = c_free_segno_head;
804 c_free_segno_head = segno;
805 c_segment_count--;
806
807 lck_mtx_unlock_always(c_list_lock);
808
809 if (c_seg->c_wanted) {
810 thread_wakeup((event_t) (c_seg));
811 c_seg->c_wanted = 0;
812 }
813 if (c_seg->c_busy_swapping) {
814 c_seg->c_must_free = 1;
815
816 lck_mtx_unlock_always(&c_seg->c_lock);
817 return;
818 }
819 if (c_seg->c_ondisk == 0) {
820 pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE;
821
822 c_buffer = c_seg->c_store.c_buffer;
823 c_seg->c_store.c_buffer = NULL;
824 } else {
825 /*
826 * Free swap space on disk.
827 */
828 c_swap_handle = c_seg->c_store.c_swap_handle;
829 c_seg->c_store.c_swap_handle = (uint64_t)-1;
830 }
831 lck_mtx_unlock_always(&c_seg->c_lock);
832
833 if (c_buffer) {
8a3053a0
A
834 if (pages_populated)
835 kernel_memory_depopulate(kernel_map, (vm_offset_t) c_buffer, pages_populated * PAGE_SIZE, KMA_COMPRESSOR);
39236c6e
A
836
837 kmem_free(kernel_map, (vm_offset_t) c_buffer, C_SEG_ALLOCSIZE);
fe8ab488
A
838 OSAddAtomic64(-C_SEG_ALLOCSIZE, &compressor_kvspace_used);
839
39236c6e
A
840 } else if (c_swap_handle)
841 vm_swap_free(c_swap_handle);
842
843
844#if __i386__ || __x86_64__
845 lck_mtx_destroy(&c_seg->c_lock, &vm_compressor_lck_grp);
846#else /* __i386__ || __x86_64__ */
847 lck_spin_destroy(&c_seg->c_lock, &vm_compressor_lck_grp);
848#endif /* __i386__ || __x86_64__ */
849
850 for (i = 0; i < C_SEG_SLOT_ARRAYS; i++) {
851 if (c_seg->c_slots[i] == 0)
852 break;
853
854 kfree((char *)c_seg->c_slots[i], sizeof(struct c_slot) * C_SEG_SLOT_ARRAY_SIZE);
855 }
856 zfree(compressor_segment_zone, c_seg);
857}
858
859
860int c_seg_trim_page_count = 0;
861
862void
863c_seg_trim_tail(c_segment_t c_seg)
864{
865 c_slot_t cs;
866 uint32_t c_size;
867 uint32_t c_offset;
868 uint32_t c_rounded_size;
869 uint16_t current_nextslot;
870 uint32_t current_populated_offset;
871
872 if (c_seg->c_bytes_used == 0)
873 return;
874 current_nextslot = c_seg->c_nextslot;
875 current_populated_offset = c_seg->c_populated_offset;
876
877 while (c_seg->c_nextslot) {
878
879 cs = C_SEG_SLOT_FROM_INDEX(c_seg, (c_seg->c_nextslot - 1));
880
881 c_size = UNPACK_C_SIZE(cs);
882
883 if (c_size) {
884 if (current_nextslot != c_seg->c_nextslot) {
885 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
886 c_offset = cs->c_offset + C_SEG_BYTES_TO_OFFSET(c_rounded_size);
887
888 c_seg->c_nextoffset = c_offset;
889 c_seg->c_populated_offset = (c_offset + (C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1)) & ~(C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1);
890
891 if (c_seg->c_firstemptyslot > c_seg->c_nextslot)
892 c_seg->c_firstemptyslot = c_seg->c_nextslot;
893
894 c_seg_trim_page_count += ((round_page_32(C_SEG_OFFSET_TO_BYTES(current_populated_offset)) -
895 round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE);
896 }
897 break;
898 }
899 c_seg->c_nextslot--;
900 }
901 assert(c_seg->c_nextslot);
902}
903
904
905int
906c_seg_minor_compaction_and_unlock(c_segment_t c_seg, boolean_t clear_busy)
907{
908 c_slot_mapping_t slot_ptr;
909 uint32_t c_offset = 0;
910 uint32_t old_populated_offset;
911 uint32_t c_rounded_size;
912 uint32_t c_size;
913 int c_indx = 0;
914 int i;
915 c_slot_t c_dst;
916 c_slot_t c_src;
917 boolean_t need_unlock = TRUE;
918
919 assert(c_seg->c_busy);
920
921#if VALIDATE_C_SEGMENTS
922 c_seg_validate(c_seg, FALSE);
923#endif
924 if (c_seg->c_bytes_used == 0) {
925 c_seg_free(c_seg);
926 return (1);
927 }
928 if (c_seg->c_firstemptyslot >= c_seg->c_nextslot || C_SEG_UNUSED_BYTES(c_seg) < PAGE_SIZE)
929 goto done;
930
931#if VALIDATE_C_SEGMENTS
932 c_seg->c_was_minor_compacted++;
933#endif
934 c_indx = c_seg->c_firstemptyslot;
935 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
936
937 old_populated_offset = c_seg->c_populated_offset;
938 c_offset = c_dst->c_offset;
939
940 for (i = c_indx + 1; i < c_seg->c_nextslot && c_offset < c_seg->c_nextoffset; i++) {
941
942 c_src = C_SEG_SLOT_FROM_INDEX(c_seg, i);
943
944 c_size = UNPACK_C_SIZE(c_src);
945
946 if (c_size == 0)
947 continue;
948
949 memcpy(&c_seg->c_store.c_buffer[c_offset], &c_seg->c_store.c_buffer[c_src->c_offset], c_size);
950
951#if CHECKSUM_THE_DATA
952 c_dst->c_hash_data = c_src->c_hash_data;
953#endif
954#if CHECKSUM_THE_COMPRESSED_DATA
955 c_dst->c_hash_compressed_data = c_src->c_hash_compressed_data;
956#endif
957 c_dst->c_size = c_src->c_size;
958 c_dst->c_packed_ptr = c_src->c_packed_ptr;
959 c_dst->c_offset = c_offset;
960
961 slot_ptr = (c_slot_mapping_t)C_SLOT_UNPACK_PTR(c_dst);
962 slot_ptr->s_cindx = c_indx;
963
964 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
965
966 c_offset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
967 PACK_C_SIZE(c_src, 0);
968 c_indx++;
969
970 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
971 }
972 c_seg->c_firstemptyslot = c_indx;
973 c_seg->c_nextslot = c_indx;
974 c_seg->c_nextoffset = c_offset;
975 c_seg->c_populated_offset = (c_offset + (C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1)) & ~(C_SEG_BYTES_TO_OFFSET(PAGE_SIZE) - 1);
976 c_seg->c_bytes_unused = 0;
977
978#if VALIDATE_C_SEGMENTS
979 c_seg_validate(c_seg, TRUE);
980#endif
981
982 if (old_populated_offset > c_seg->c_populated_offset) {
983 uint32_t gc_size;
984 int32_t *gc_ptr;
985
986 gc_size = C_SEG_OFFSET_TO_BYTES(old_populated_offset - c_seg->c_populated_offset);
987 gc_ptr = &c_seg->c_store.c_buffer[c_seg->c_populated_offset];
988
989 lck_mtx_unlock_always(&c_seg->c_lock);
990
991 kernel_memory_depopulate(kernel_map, (vm_offset_t)gc_ptr, gc_size, KMA_COMPRESSOR);
992
993 if (clear_busy == TRUE)
994 lck_mtx_lock_spin_always(&c_seg->c_lock);
995 else
996 need_unlock = FALSE;
997 }
998done:
999 if (need_unlock == TRUE) {
1000 if (clear_busy == TRUE)
1001 C_SEG_WAKEUP_DONE(c_seg);
1002
1003 lck_mtx_unlock_always(&c_seg->c_lock);
1004 }
1005 return (0);
1006}
1007
1008
1009
1010struct {
1011 uint64_t asked_permission;
1012 uint64_t compactions;
1013 uint64_t moved_slots;
1014 uint64_t moved_bytes;
1015 uint64_t wasted_space_in_swapouts;
1016 uint64_t count_of_swapouts;
1017} c_seg_major_compact_stats;
1018
1019
1020#define C_MAJOR_COMPACTION_AGE_APPROPRIATE 30
1021#define C_MAJOR_COMPACTION_OLD_ENOUGH 300
1022#define C_MAJOR_COMPACTION_SIZE_APPROPRIATE ((C_SEG_BUFSIZE * 80) / 100)
1023
1024
1025boolean_t
1026c_seg_major_compact_ok(
1027 c_segment_t c_seg_dst,
1028 c_segment_t c_seg_src)
1029{
1030
1031 c_seg_major_compact_stats.asked_permission++;
1032
1033 if (c_seg_src->c_filling) {
1034 /*
1035 * we're at or near the head... don't compact
1036 */
1037 return (FALSE);
1038 }
1039 if (c_seg_src->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE &&
1040 c_seg_dst->c_bytes_used >= C_MAJOR_COMPACTION_SIZE_APPROPRIATE)
1041 return (FALSE);
1042
1043 if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX) {
1044 /*
1045 * destination segment is full... can't compact
1046 */
1047 return (FALSE);
1048 }
1049
1050 return (TRUE);
1051}
1052
1053
1054boolean_t
1055c_seg_major_compact(
1056 c_segment_t c_seg_dst,
1057 c_segment_t c_seg_src)
1058{
1059 c_slot_mapping_t slot_ptr;
1060 uint32_t c_rounded_size;
1061 uint32_t c_size;
1062 uint16_t dst_slot;
1063 int i;
1064 c_slot_t c_dst;
1065 c_slot_t c_src;
1066 int slotarray;
1067 boolean_t keep_compacting = TRUE;
1068
1069 /*
1070 * segments are not locked but they are both marked c_busy
1071 * which keeps c_decompress from working on them...
1072 * we can safely allocate new pages, move compressed data
1073 * from c_seg_src to c_seg_dst and update both c_segment's
1074 * state w/o holding the master lock
1075 */
1076
1077#if VALIDATE_C_SEGMENTS
1078 c_seg_dst->c_was_major_compacted++;
1079 c_seg_src->c_was_major_donor++;
1080#endif
1081 c_seg_major_compact_stats.compactions++;
1082
1083 dst_slot = c_seg_dst->c_nextslot;
1084
1085 for (i = 0; i < c_seg_src->c_nextslot; i++) {
1086
1087 c_src = C_SEG_SLOT_FROM_INDEX(c_seg_src, i);
1088
1089 c_size = UNPACK_C_SIZE(c_src);
1090
1091 if (c_size == 0) {
1092 /* BATCH: move what we have so far; */
1093 continue;
1094 }
1095
1096 if (C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset - c_seg_dst->c_nextoffset) < (unsigned) c_size) {
1097 /* doesn't fit */
1098 if ((C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset) == C_SEG_BUFSIZE)) {
1099 /* can't fit */
1100 keep_compacting = FALSE;
1101 break;
1102 }
1103 kernel_memory_populate(kernel_map,
1104 (vm_offset_t) &c_seg_dst->c_store.c_buffer[c_seg_dst->c_populated_offset],
1105 PAGE_SIZE,
1106 KMA_COMPRESSOR);
1107
1108 c_seg_dst->c_populated_offset += C_SEG_BYTES_TO_OFFSET(PAGE_SIZE);
1109 assert(C_SEG_OFFSET_TO_BYTES(c_seg_dst->c_populated_offset) <= C_SEG_BUFSIZE);
1110 }
1111
1112 slotarray = C_SEG_SLOTARRAY_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot);
1113
1114 if (c_seg_dst->c_slots[slotarray] == 0) {
1115 KERNEL_DEBUG(0xe0400008 | DBG_FUNC_START, 0, 0, 0, 0, 0);
1116 c_seg_dst->c_slots[slotarray] = (struct c_slot *)
1117 kalloc(sizeof(struct c_slot) *
1118 C_SEG_SLOT_ARRAY_SIZE);
1119 KERNEL_DEBUG(0xe0400008 | DBG_FUNC_END, 0, 0, 0, 0, 0);
1120 }
1121 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, c_seg_dst->c_nextslot);
1122
1123 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);
1124
1125 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
1126
1127 c_seg_major_compact_stats.moved_slots++;
1128 c_seg_major_compact_stats.moved_bytes += c_size;
1129
1130#if CHECKSUM_THE_DATA
1131 c_dst->c_hash_data = c_src->c_hash_data;
1132#endif
1133#if CHECKSUM_THE_COMPRESSED_DATA
1134 c_dst->c_hash_compressed_data = c_src->c_hash_compressed_data;
1135#endif
1136 c_dst->c_size = c_src->c_size;
1137 c_dst->c_packed_ptr = c_src->c_packed_ptr;
1138 c_dst->c_offset = c_seg_dst->c_nextoffset;
1139
1140 if (c_seg_dst->c_firstemptyslot == c_seg_dst->c_nextslot)
1141 c_seg_dst->c_firstemptyslot++;
1142 c_seg_dst->c_nextslot++;
1143 c_seg_dst->c_bytes_used += c_rounded_size;
1144 c_seg_dst->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
1145
1146 PACK_C_SIZE(c_src, 0);
1147
1148 c_seg_src->c_bytes_used -= c_rounded_size;
1149 c_seg_src->c_bytes_unused += c_rounded_size;
1150 c_seg_src->c_firstemptyslot = 0;
1151
1152 if (c_seg_dst->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg_dst->c_nextslot >= C_SLOT_MAX) {
1153 /* dest segment is now full */
1154 keep_compacting = FALSE;
1155 break;
1156 }
1157 }
1158 if (dst_slot < c_seg_dst->c_nextslot) {
1159
1160 PAGE_REPLACEMENT_ALLOWED(TRUE);
1161 /*
1162 * we've now locked out c_decompress from
1163 * converting the slot passed into it into
1164 * a c_segment_t which allows us to use
1165 * the backptr to change which c_segment and
1166 * index the slot points to
1167 */
1168 while (dst_slot < c_seg_dst->c_nextslot) {
1169
1170 c_dst = C_SEG_SLOT_FROM_INDEX(c_seg_dst, dst_slot);
1171
1172 slot_ptr = (c_slot_mapping_t)C_SLOT_UNPACK_PTR(c_dst);
1173 /* <csegno=0,indx=0> would mean "empty slot", so use csegno+1 */
1174 slot_ptr->s_cseg = c_seg_dst->c_mysegno + 1;
1175 slot_ptr->s_cindx = dst_slot++;
1176 }
1177 PAGE_REPLACEMENT_ALLOWED(FALSE);
1178 }
1179 return (keep_compacting);
1180}
1181
1182
fe8ab488
A
1183uint64_t
1184vm_compressor_compute_elapsed_msecs(clock_sec_t end_sec, clock_nsec_t end_nsec, clock_sec_t start_sec, clock_nsec_t start_nsec)
39236c6e
A
1185{
1186 uint64_t end_msecs;
1187 uint64_t start_msecs;
1188
1189 end_msecs = (end_sec * 1000) + end_nsec / 1000000;
1190 start_msecs = (start_sec * 1000) + start_nsec / 1000000;
1191
1192 return (end_msecs - start_msecs);
1193}
1194
1195
1196
1197uint32_t compressor_eval_period_in_msecs = 250;
1198uint32_t compressor_sample_min_in_msecs = 500;
1199uint32_t compressor_sample_max_in_msecs = 10000;
1200uint32_t compressor_thrashing_threshold_per_10msecs = 50;
1201uint32_t compressor_thrashing_min_per_10msecs = 20;
1202
fe8ab488
A
1203/* When true, reset sample data next chance we get. */
1204static boolean_t compressor_need_sample_reset = FALSE;
1205
39236c6e
A
1206extern uint32_t vm_page_filecache_min;
1207
1208
1209void
1210compute_swapout_target_age(void)
1211{
1212 clock_sec_t cur_ts_sec;
1213 clock_nsec_t cur_ts_nsec;
1214 uint32_t min_operations_needed_in_this_sample;
1215 uint64_t elapsed_msecs_in_eval;
1216 uint64_t elapsed_msecs_in_sample;
39236c6e
A
1217 boolean_t need_eval_reset = FALSE;
1218
1219 clock_get_system_nanotime(&cur_ts_sec, &cur_ts_nsec);
1220
fe8ab488 1221 elapsed_msecs_in_sample = vm_compressor_compute_elapsed_msecs(cur_ts_sec, cur_ts_nsec, start_of_sample_period_sec, start_of_sample_period_nsec);
39236c6e 1222
fe8ab488
A
1223 if (compressor_need_sample_reset ||
1224 elapsed_msecs_in_sample >= compressor_sample_max_in_msecs) {
1225 compressor_need_sample_reset = TRUE;
39236c6e
A
1226 need_eval_reset = TRUE;
1227 goto done;
1228 }
fe8ab488 1229 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);
39236c6e
A
1230
1231 if (elapsed_msecs_in_eval < compressor_eval_period_in_msecs)
1232 goto done;
1233 need_eval_reset = TRUE;
1234
1235 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_START, elapsed_msecs_in_eval, sample_period_compression_count, sample_period_decompression_count, 0, 0);
1236
1237 min_operations_needed_in_this_sample = (compressor_thrashing_min_per_10msecs * (uint32_t)elapsed_msecs_in_eval) / 10;
1238
1239 if ((sample_period_compression_count - last_eval_compression_count) < min_operations_needed_in_this_sample ||
1240 (sample_period_decompression_count - last_eval_decompression_count) < min_operations_needed_in_this_sample) {
1241
1242 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, sample_period_compression_count - last_eval_compression_count,
1243 sample_period_decompression_count - last_eval_decompression_count, 0, 1, 0);
1244
1245 swapout_target_age = 0;
1246
fe8ab488 1247 compressor_need_sample_reset = TRUE;
39236c6e
A
1248 need_eval_reset = TRUE;
1249 goto done;
1250 }
1251 last_eval_compression_count = sample_period_compression_count;
1252 last_eval_decompression_count = sample_period_decompression_count;
1253
1254 if (elapsed_msecs_in_sample < compressor_sample_min_in_msecs) {
1255
1256 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, swapout_target_age, 0, 0, 5, 0);
1257 goto done;
1258 }
1259 if (sample_period_decompression_count > ((compressor_thrashing_threshold_per_10msecs * elapsed_msecs_in_sample) / 10)) {
1260
1261 uint64_t running_total;
1262 uint64_t working_target;
1263 uint64_t aging_target;
1264 uint32_t oldest_age_of_csegs_sampled = 0;
1265 uint64_t working_set_approximation = 0;
1266
1267 swapout_target_age = 0;
1268
1269 working_target = (sample_period_decompression_count / 100) * 95; /* 95 percent */
1270 aging_target = (sample_period_decompression_count / 100) * 1; /* 1 percent */
1271 running_total = 0;
1272
1273 for (oldest_age_of_csegs_sampled = 0; oldest_age_of_csegs_sampled < DECOMPRESSION_SAMPLE_MAX_AGE; oldest_age_of_csegs_sampled++) {
1274
1275 running_total += age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled];
1276
1277 working_set_approximation += oldest_age_of_csegs_sampled * age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled];
1278
1279 if (running_total >= working_target)
1280 break;
1281 }
1282 if (oldest_age_of_csegs_sampled < DECOMPRESSION_SAMPLE_MAX_AGE) {
1283
1284 working_set_approximation = (working_set_approximation * 1000) / elapsed_msecs_in_sample;
1285
1286 if (working_set_approximation < VM_PAGE_COMPRESSOR_COUNT) {
1287
1288 running_total = overage_decompressions_during_sample_period;
1289
1290 for (oldest_age_of_csegs_sampled = DECOMPRESSION_SAMPLE_MAX_AGE - 1; oldest_age_of_csegs_sampled; oldest_age_of_csegs_sampled--) {
1291 running_total += age_of_decompressions_during_sample_period[oldest_age_of_csegs_sampled];
1292
1293 if (running_total >= aging_target)
1294 break;
1295 }
1296 swapout_target_age = (uint32_t)cur_ts_sec - oldest_age_of_csegs_sampled;
1297
1298 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, swapout_target_age, working_set_approximation, VM_PAGE_COMPRESSOR_COUNT, 2, 0);
1299 } else {
1300 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, working_set_approximation, VM_PAGE_COMPRESSOR_COUNT, 0, 3, 0);
1301 }
1302 } else
1303 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, working_target, running_total, 0, 4, 0);
1304
fe8ab488 1305 compressor_need_sample_reset = TRUE;
39236c6e
A
1306 need_eval_reset = TRUE;
1307 } else
1308 KERNEL_DEBUG(0xe0400020 | DBG_FUNC_END, sample_period_decompression_count, (compressor_thrashing_threshold_per_10msecs * elapsed_msecs_in_sample) / 10, 0, 6, 0);
1309done:
fe8ab488 1310 if (compressor_need_sample_reset == TRUE) {
39236c6e
A
1311 bzero(age_of_decompressions_during_sample_period, sizeof(age_of_decompressions_during_sample_period));
1312 overage_decompressions_during_sample_period = 0;
1313
1314 start_of_sample_period_sec = cur_ts_sec;
1315 start_of_sample_period_nsec = cur_ts_nsec;
1316 sample_period_decompression_count = 0;
1317 sample_period_compression_count = 0;
1318 last_eval_decompression_count = 0;
1319 last_eval_compression_count = 0;
fe8ab488 1320 compressor_need_sample_reset = FALSE;
39236c6e
A
1321 }
1322 if (need_eval_reset == TRUE) {
1323 start_of_eval_period_sec = cur_ts_sec;
1324 start_of_eval_period_nsec = cur_ts_nsec;
1325 }
1326}
1327
1328
fe8ab488
A
1329int compaction_swapper_inited = 0;
1330int compaction_swapper_init_now = 0;
39236c6e
A
1331int compaction_swapper_running = 0;
1332int compaction_swapper_abort = 0;
1333
1334
1335#if CONFIG_JETSAM
1336boolean_t memorystatus_kill_on_VM_thrashing(boolean_t);
fe8ab488 1337boolean_t memorystatus_kill_on_FC_thrashing(boolean_t);
39236c6e 1338int compressor_thrashing_induced_jetsam = 0;
fe8ab488
A
1339int filecache_thrashing_induced_jetsam = 0;
1340static boolean_t vm_compressor_thrashing_detected = FALSE;
39236c6e
A
1341#endif /* CONFIG_JETSAM */
1342
1343static boolean_t
1344compressor_needs_to_swap(void)
1345{
1346 boolean_t should_swap = FALSE;
1347
1348 if (vm_swap_up == TRUE) {
1349 if (COMPRESSOR_NEEDS_TO_SWAP()) {
1350 return (TRUE);
1351 }
1352 if (VM_PAGE_Q_THROTTLED(&vm_pageout_queue_external) && vm_page_anonymous_count < (vm_page_inactive_count / 20)) {
1353 return (TRUE);
1354 }
1355 if (vm_page_free_count < (vm_page_free_reserved - COMPRESSOR_FREE_RESERVED_LIMIT))
1356 return (TRUE);
1357 }
1358 compute_swapout_target_age();
1359
1360 if (swapout_target_age) {
1361 c_segment_t c_seg;
1362
1363 lck_mtx_lock_spin_always(c_list_lock);
1364
1365 if (!queue_empty(&c_age_list_head)) {
1366
1367 c_seg = (c_segment_t) queue_first(&c_age_list_head);
1368
fe8ab488 1369 if (c_seg->c_creation_ts > swapout_target_age)
39236c6e
A
1370 swapout_target_age = 0;
1371 }
1372 lck_mtx_unlock_always(c_list_lock);
1373 }
fe8ab488
A
1374#if CONFIG_PHANTOM_CACHE
1375 if (vm_phantom_cache_check_pressure())
1376 should_swap = TRUE;
1377#endif
1378 if (swapout_target_age)
1379 should_swap = TRUE;
39236c6e
A
1380
1381 if (vm_swap_up == FALSE) {
fe8ab488 1382
39236c6e 1383 if (should_swap) {
fe8ab488 1384#if CONFIG_JETSAM
39236c6e
A
1385 if (vm_compressor_thrashing_detected == FALSE) {
1386 vm_compressor_thrashing_detected = TRUE;
fe8ab488
A
1387
1388 if (swapout_target_age) {
1389 memorystatus_kill_on_VM_thrashing(TRUE /* async */);
1390 compressor_thrashing_induced_jetsam++;
1391 } else {
1392 memorystatus_kill_on_FC_thrashing(TRUE /* async */);
1393 filecache_thrashing_induced_jetsam++;
1394 }
39236c6e
A
1395 /*
1396 * let the jetsam take precedence over
1397 * any major compactions we might have
1398 * been able to do... otherwise we run
1399 * the risk of doing major compactions
1400 * on segments we're about to free up
1401 * due to the jetsam activity.
1402 */
1403 should_swap = FALSE;
1404 }
39236c6e 1405#endif /* CONFIG_JETSAM */
fe8ab488
A
1406 } else
1407 should_swap = COMPRESSOR_NEEDS_TO_MAJOR_COMPACT();
39236c6e 1408 }
fe8ab488 1409
39236c6e
A
1410 /*
1411 * returning TRUE when swap_supported == FALSE
1412 * will cause the major compaction engine to
1413 * run, but will not trigger any swapping...
1414 * segments that have been major compacted
1415 * will be moved to the swapped_out_q
1416 * but will not have the c_ondisk flag set
1417 */
1418 return (should_swap);
1419}
1420
fe8ab488
A
1421#if CONFIG_JETSAM
1422/*
1423 * This function is called from the jetsam thread after killing something to
1424 * mitigate thrashing.
1425 *
1426 * We need to restart our thrashing detection heuristics since memory pressure
1427 * has potentially changed significantly, and we don't want to detect on old
1428 * data from before the jetsam.
1429 */
1430void
1431vm_thrashing_jetsam_done(void)
39236c6e 1432{
fe8ab488 1433 vm_compressor_thrashing_detected = FALSE;
39236c6e 1434
fe8ab488
A
1435 /* Were we compressor-thrashing or filecache-thrashing? */
1436 if (swapout_target_age) {
1437 swapout_target_age = 0;
1438 compressor_need_sample_reset = TRUE;
39236c6e 1439 }
fe8ab488
A
1440#if CONFIG_PHANTOM_CACHE
1441 else {
1442 vm_phantom_cache_restart_sample();
1443 }
1444#endif
39236c6e 1445}
fe8ab488 1446#endif /* CONFIG_JETSAM */
39236c6e
A
1447
1448uint32_t vm_wake_compactor_swapper_calls = 0;
1449
1450void
1451vm_wake_compactor_swapper(void)
1452{
fe8ab488
A
1453 boolean_t need_major_compaction = FALSE;
1454
39236c6e
A
1455 if (compaction_swapper_running)
1456 return;
1457
fe8ab488 1458 if (c_minor_count == 0 && need_major_compaction == FALSE)
39236c6e
A
1459 return;
1460
1461 lck_mtx_lock_spin_always(c_list_lock);
1462
1463 fastwake_warmup = FALSE;
1464
1465 if (compaction_swapper_running == 0) {
1466 vm_wake_compactor_swapper_calls++;
1467
1468 thread_wakeup((event_t)&c_compressor_swap_trigger);
1469
1470 compaction_swapper_running = 1;
1471 }
1472 lck_mtx_unlock_always(c_list_lock);
1473}
1474
fe8ab488 1475
39236c6e
A
1476void
1477vm_consider_waking_compactor_swapper(void)
1478{
1479 boolean_t need_wakeup = FALSE;
1480
fe8ab488 1481 if (compaction_swapper_running)
39236c6e 1482 return;
fe8ab488
A
1483
1484 if (!compaction_swapper_inited && !compaction_swapper_init_now) {
1485 compaction_swapper_init_now = 1;
1486 need_wakeup = TRUE;
1487 }
39236c6e
A
1488
1489 if (c_minor_count && (COMPRESSOR_NEEDS_TO_MINOR_COMPACT())) {
1490
1491 need_wakeup = TRUE;
1492
1493 } else if (compressor_needs_to_swap()) {
1494
1495 need_wakeup = TRUE;
1496
1497 } else if (c_minor_count) {
1498 uint64_t total_bytes;
1499
1500 total_bytes = compressor_object->resident_page_count * PAGE_SIZE_64;
1501
1502 if ((total_bytes - compressor_bytes_used) > total_bytes / 10)
1503 need_wakeup = TRUE;
1504 }
1505 if (need_wakeup == TRUE) {
1506
1507 lck_mtx_lock_spin_always(c_list_lock);
1508
1509 fastwake_warmup = FALSE;
1510
1511 if (compaction_swapper_running == 0) {
1512 memoryshot(VM_WAKEUP_COMPACTOR_SWAPPER, DBG_FUNC_NONE);
1513
1514 thread_wakeup((event_t)&c_compressor_swap_trigger);
1515
1516 compaction_swapper_running = 1;
1517 }
1518 lck_mtx_unlock_always(c_list_lock);
1519 }
1520}
1521
1522
1523#define C_SWAPOUT_LIMIT 4
1524#define DELAYED_COMPACTIONS_PER_PASS 30
1525
1526void
1527vm_compressor_do_delayed_compactions(boolean_t flush_all)
1528{
1529 c_segment_t c_seg;
1530 int number_compacted = 0;
1531 boolean_t needs_to_swap = FALSE;
1532
1533
1534 lck_mtx_assert(c_list_lock, LCK_MTX_ASSERT_OWNED);
1535
1536 while (!queue_empty(&c_minor_list_head) && needs_to_swap == FALSE) {
1537
1538 c_seg = (c_segment_t)queue_first(&c_minor_list_head);
1539
1540 lck_mtx_lock_spin_always(&c_seg->c_lock);
8a3053a0 1541
fe8ab488 1542 if (c_seg->c_busy) {
8a3053a0 1543
fe8ab488
A
1544 lck_mtx_unlock_always(c_list_lock);
1545 c_seg_wait_on_busy(c_seg);
1546 lck_mtx_lock_spin_always(c_list_lock);
8a3053a0 1547
fe8ab488
A
1548 continue;
1549 }
1550 C_SEG_BUSY(c_seg);
39236c6e
A
1551
1552 c_seg_do_minor_compaction_and_unlock(c_seg, TRUE, FALSE, TRUE);
1553
1554 if (vm_swap_up == TRUE && (number_compacted++ > DELAYED_COMPACTIONS_PER_PASS)) {
1555
1556 if ((flush_all == TRUE || compressor_needs_to_swap() == TRUE) && c_swapout_count < C_SWAPOUT_LIMIT)
1557 needs_to_swap = TRUE;
1558
1559 number_compacted = 0;
1560 }
1561 lck_mtx_lock_spin_always(c_list_lock);
1562 }
1563}
1564
1565
1566#define C_SEGMENT_SWAPPEDIN_AGE_LIMIT 10
1567
1568static void
1569vm_compressor_age_swapped_in_segments(boolean_t flush_all)
1570{
1571 c_segment_t c_seg;
1572 clock_sec_t now;
1573 clock_nsec_t nsec;
1574
1575 clock_get_system_nanotime(&now, &nsec);
1576
1577 while (!queue_empty(&c_swappedin_list_head)) {
1578
1579 c_seg = (c_segment_t)queue_first(&c_swappedin_list_head);
1580
1581 if (flush_all == FALSE && (now - c_seg->c_swappedin_ts) < C_SEGMENT_SWAPPEDIN_AGE_LIMIT)
1582 break;
1583
1584 lck_mtx_lock_spin_always(&c_seg->c_lock);
1585
1586 queue_remove(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list);
1587 c_seg->c_on_swappedin_q = 0;
1588 c_swappedin_count--;
1589
1590 c_seg_insert_into_q(&c_age_list_head, c_seg);
1591 c_seg->c_on_age_q = 1;
1592 c_age_count++;
1593
1594 lck_mtx_unlock_always(&c_seg->c_lock);
1595 }
1596}
1597
1598
1599void
1600vm_compressor_flush(void)
1601{
1602 uint64_t vm_swap_put_failures_at_start;
1603 wait_result_t wait_result = 0;
1604 AbsoluteTime startTime, endTime;
1605 clock_sec_t now_sec;
1606 clock_nsec_t now_nsec;
1607 uint64_t nsec;
1608
1609 HIBLOG("vm_compressor_flush - starting\n");
1610
1611 clock_get_uptime(&startTime);
1612
1613 lck_mtx_lock_spin_always(c_list_lock);
1614
1615 fastwake_warmup = FALSE;
1616 compaction_swapper_abort = 1;
1617
1618 while (compaction_swapper_running) {
1619 assert_wait((event_t)&compaction_swapper_running, THREAD_UNINT);
1620
1621 lck_mtx_unlock_always(c_list_lock);
1622
1623 thread_block(THREAD_CONTINUE_NULL);
1624
1625 lck_mtx_lock_spin_always(c_list_lock);
1626 }
1627 compaction_swapper_abort = 0;
1628 compaction_swapper_running = 1;
1629
1630 hibernate_flushing = TRUE;
1631 hibernate_no_swapspace = FALSE;
1632 c_generation_id_flush_barrier = c_generation_id + 1000;
1633
1634 clock_get_system_nanotime(&now_sec, &now_nsec);
1635 hibernate_flushing_deadline = now_sec + HIBERNATE_FLUSHING_SECS_TO_COMPLETE;
1636
1637 vm_swap_put_failures_at_start = vm_swap_put_failures;
1638
1639 vm_compressor_compact_and_swap(TRUE);
1640
1641 while (!queue_empty(&c_swapout_list_head)) {
1642
1643 assert_wait_timeout((event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, 5000, 1000*NSEC_PER_USEC);
1644
1645 lck_mtx_unlock_always(c_list_lock);
1646
1647 wait_result = thread_block(THREAD_CONTINUE_NULL);
1648
1649 lck_mtx_lock_spin_always(c_list_lock);
1650
1651 if (wait_result == THREAD_TIMED_OUT)
1652 break;
1653 }
1654 hibernate_flushing = FALSE;
1655 compaction_swapper_running = 0;
1656
1657 if (vm_swap_put_failures > vm_swap_put_failures_at_start)
1658 HIBLOG("vm_compressor_flush failed to clean %llu segments - vm_page_compressor_count(%d)\n",
1659 vm_swap_put_failures - vm_swap_put_failures_at_start, VM_PAGE_COMPRESSOR_COUNT);
1660
1661 lck_mtx_unlock_always(c_list_lock);
1662
1663 clock_get_uptime(&endTime);
1664 SUB_ABSOLUTETIME(&endTime, &startTime);
1665 absolutetime_to_nanoseconds(endTime, &nsec);
1666
1667 HIBLOG("vm_compressor_flush completed - took %qd msecs\n", nsec / 1000000ULL);
1668}
1669
1670
fe8ab488 1671extern void vm_swap_file_set_tuneables(void);
39236c6e
A
1672int compaction_swap_trigger_thread_awakened = 0;
1673
fe8ab488 1674
39236c6e
A
1675static void
1676vm_compressor_swap_trigger_thread(void)
1677{
fe8ab488
A
1678 /*
1679 * compaction_swapper_init_now is set when the first call to
1680 * vm_consider_waking_compactor_swapper is made from
1681 * vm_pageout_scan... since this function is called upon
1682 * thread creation, we want to make sure to delay adjusting
1683 * the tuneables until we are awakened via vm_pageout_scan
1684 * so that we are at a point where the vm_swapfile_open will
1685 * be operating on the correct directory (in case the default
1686 * of /var/vm/ is overridden by the dymanic_pager
1687 */
1688 if (compaction_swapper_init_now && !compaction_swapper_inited) {
1689 if (vm_compressor_mode == VM_PAGER_COMPRESSOR_WITH_SWAP)
1690 vm_swap_file_set_tuneables();
39236c6e 1691
fe8ab488
A
1692 compaction_swapper_inited = 1;
1693 }
39236c6e
A
1694 lck_mtx_lock_spin_always(c_list_lock);
1695
1696 compaction_swap_trigger_thread_awakened++;
1697
1698 vm_compressor_compact_and_swap(FALSE);
1699
1700 assert_wait((event_t)&c_compressor_swap_trigger, THREAD_UNINT);
1701
1702 compaction_swapper_running = 0;
1703 thread_wakeup((event_t)&compaction_swapper_running);
1704
1705 lck_mtx_unlock_always(c_list_lock);
1706
1707 thread_block((thread_continue_t)vm_compressor_swap_trigger_thread);
1708
1709 /* NOTREACHED */
1710}
1711
1712
1713void
1714vm_compressor_record_warmup_start(void)
1715{
1716 c_segment_t c_seg;
1717
1718 lck_mtx_lock_spin_always(c_list_lock);
1719
8a3053a0
A
1720 if (first_c_segment_to_warm_generation_id == 0) {
1721 if (!queue_empty(&c_age_list_head)) {
39236c6e 1722
8a3053a0 1723 c_seg = (c_segment_t)queue_last(&c_age_list_head);
39236c6e 1724
8a3053a0
A
1725 first_c_segment_to_warm_generation_id = c_seg->c_generation_id;
1726 } else
1727 first_c_segment_to_warm_generation_id = 0;
39236c6e 1728
8a3053a0
A
1729 fastwake_recording_in_progress = TRUE;
1730 }
39236c6e
A
1731 lck_mtx_unlock_always(c_list_lock);
1732}
1733
1734
1735void
1736vm_compressor_record_warmup_end(void)
1737{
1738 c_segment_t c_seg;
1739
1740 lck_mtx_lock_spin_always(c_list_lock);
1741
8a3053a0 1742 if (fastwake_recording_in_progress == TRUE) {
39236c6e 1743
8a3053a0 1744 if (!queue_empty(&c_age_list_head)) {
39236c6e 1745
8a3053a0
A
1746 c_seg = (c_segment_t)queue_last(&c_age_list_head);
1747
1748 last_c_segment_to_warm_generation_id = c_seg->c_generation_id;
1749 } else
1750 last_c_segment_to_warm_generation_id = first_c_segment_to_warm_generation_id;
39236c6e 1751
8a3053a0 1752 fastwake_recording_in_progress = FALSE;
39236c6e 1753
8a3053a0
A
1754 HIBLOG("vm_compressor_record_warmup (%qd - %qd)\n", first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id);
1755 }
39236c6e
A
1756 lck_mtx_unlock_always(c_list_lock);
1757}
1758
1759
1760#define DELAY_TRIM_ON_WAKE_SECS 4
1761
1762void
8a3053a0 1763vm_compressor_delay_trim(void)
39236c6e 1764{
8a3053a0 1765 clock_sec_t sec;
39236c6e
A
1766 clock_nsec_t nsec;
1767
1768 clock_get_system_nanotime(&sec, &nsec);
1769 dont_trim_until_ts = sec + DELAY_TRIM_ON_WAKE_SECS;
8a3053a0 1770}
39236c6e 1771
39236c6e 1772
8a3053a0
A
1773void
1774vm_compressor_do_warmup(void)
1775{
39236c6e
A
1776 lck_mtx_lock_spin_always(c_list_lock);
1777
8a3053a0
A
1778 if (first_c_segment_to_warm_generation_id == last_c_segment_to_warm_generation_id) {
1779 first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0;
1780
1781 lck_mtx_unlock_always(c_list_lock);
1782 return;
1783 }
1784
39236c6e
A
1785 if (compaction_swapper_running == 0) {
1786
1787 fastwake_warmup = TRUE;
1788 compaction_swapper_running = 1;
1789 thread_wakeup((event_t)&c_compressor_swap_trigger);
1790 }
1791 lck_mtx_unlock_always(c_list_lock);
1792}
1793
1794
1795void
1796do_fastwake_warmup(void)
1797{
1798 uint64_t my_thread_id;
1799 c_segment_t c_seg = NULL;
8a3053a0
A
1800 AbsoluteTime startTime, endTime;
1801 uint64_t nsec;
1802
1803
1804 HIBLOG("vm_compressor_fastwake_warmup (%qd - %qd) - starting\n", first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id);
1805
1806 clock_get_uptime(&startTime);
39236c6e
A
1807
1808 lck_mtx_unlock_always(c_list_lock);
1809
1810 my_thread_id = current_thread()->thread_id;
1811 proc_set_task_policy_thread(kernel_task, my_thread_id,
1812 TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER2);
1813
1814 PAGE_REPLACEMENT_DISALLOWED(TRUE);
1815
1816 lck_mtx_lock_spin_always(c_list_lock);
1817
1818 while (!queue_empty(&c_swappedout_list_head) && fastwake_warmup == TRUE) {
1819
1820 c_seg = (c_segment_t) queue_first(&c_swappedout_list_head);
1821
1822 if (c_seg->c_generation_id < first_c_segment_to_warm_generation_id ||
1823 c_seg->c_generation_id > last_c_segment_to_warm_generation_id)
1824 break;
1825
1826 lck_mtx_lock_spin_always(&c_seg->c_lock);
1827 lck_mtx_unlock_always(c_list_lock);
1828
8a3053a0
A
1829 if (c_seg->c_busy) {
1830 PAGE_REPLACEMENT_DISALLOWED(FALSE);
39236c6e 1831 c_seg_wait_on_busy(c_seg);
8a3053a0
A
1832 PAGE_REPLACEMENT_DISALLOWED(TRUE);
1833 } else {
39236c6e
A
1834 c_seg_swapin(c_seg, TRUE);
1835
1836 lck_mtx_unlock_always(&c_seg->c_lock);
39236c6e 1837 c_segment_warmup_count++;
8a3053a0
A
1838
1839 PAGE_REPLACEMENT_DISALLOWED(FALSE);
39236c6e 1840 vm_pageout_io_throttle();
8a3053a0 1841 PAGE_REPLACEMENT_DISALLOWED(TRUE);
39236c6e
A
1842 }
1843 lck_mtx_lock_spin_always(c_list_lock);
1844 }
1845 lck_mtx_unlock_always(c_list_lock);
1846
1847 PAGE_REPLACEMENT_DISALLOWED(FALSE);
1848
1849 proc_set_task_policy_thread(kernel_task, my_thread_id,
1850 TASK_POLICY_INTERNAL, TASK_POLICY_IO, THROTTLE_LEVEL_COMPRESSOR_TIER0);
1851
8a3053a0
A
1852 clock_get_uptime(&endTime);
1853 SUB_ABSOLUTETIME(&endTime, &startTime);
1854 absolutetime_to_nanoseconds(endTime, &nsec);
1855
1856 HIBLOG("vm_compressor_fastwake_warmup completed - took %qd msecs\n", nsec / 1000000ULL);
1857
39236c6e 1858 lck_mtx_lock_spin_always(c_list_lock);
8a3053a0
A
1859
1860 first_c_segment_to_warm_generation_id = last_c_segment_to_warm_generation_id = 0;
39236c6e
A
1861}
1862
1863
1864void
1865vm_compressor_compact_and_swap(boolean_t flush_all)
1866{
1867 c_segment_t c_seg, c_seg_next;
1868 boolean_t keep_compacting;
1869
1870
1871 if (fastwake_warmup == TRUE) {
1872 uint64_t starting_warmup_count;
1873
1874 starting_warmup_count = c_segment_warmup_count;
1875
1876 KERNEL_DEBUG_CONSTANT(IOKDBG_CODE(DBG_HIBERNATE, 11) | DBG_FUNC_START, c_segment_warmup_count,
1877 first_c_segment_to_warm_generation_id, last_c_segment_to_warm_generation_id, 0, 0);
1878 do_fastwake_warmup();
1879 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);
1880
1881 fastwake_warmup = FALSE;
1882 }
1883
8a3053a0
A
1884 /*
1885 * it's possible for the c_age_list_head to be empty if we
1886 * hit our limits for growing the compressor pool and we subsequently
1887 * hibernated... on the next hibernation we could see the queue as
1888 * empty and not proceeed even though we have a bunch of segments on
1889 * the swapped in queue that need to be dealt with.
1890 */
1891 vm_compressor_do_delayed_compactions(flush_all);
1892
1893 vm_compressor_age_swapped_in_segments(flush_all);
1894
1895
39236c6e
A
1896 while (!queue_empty(&c_age_list_head) && compaction_swapper_abort == 0) {
1897
1898 if (hibernate_flushing == TRUE) {
1899 clock_sec_t sec;
1900 clock_nsec_t nsec;
1901
1902 if (hibernate_should_abort()) {
1903 HIBLOG("vm_compressor_flush - hibernate_should_abort returned TRUE\n");
1904 break;
1905 }
1906 if (hibernate_no_swapspace == TRUE) {
1907 HIBLOG("vm_compressor_flush - out of swap space\n");
1908 break;
1909 }
1910 clock_get_system_nanotime(&sec, &nsec);
1911
1912 if (sec > hibernate_flushing_deadline) {
1913 HIBLOG("vm_compressor_flush - failed to finish before deadline\n");
1914 break;
1915 }
1916 }
1917 if (c_swapout_count >= C_SWAPOUT_LIMIT) {
1918
1919 assert_wait_timeout((event_t) &compaction_swapper_running, THREAD_INTERRUPTIBLE, 100, 1000*NSEC_PER_USEC);
1920
1921 lck_mtx_unlock_always(c_list_lock);
1922
1923 thread_block(THREAD_CONTINUE_NULL);
1924
1925 lck_mtx_lock_spin_always(c_list_lock);
1926 }
1927 /*
1928 * Minor compactions
1929 */
1930 vm_compressor_do_delayed_compactions(flush_all);
1931
1932 vm_compressor_age_swapped_in_segments(flush_all);
1933
1934 if (c_swapout_count >= C_SWAPOUT_LIMIT) {
1935 /*
1936 * we timed out on the above thread_block
1937 * let's loop around and try again
1938 * the timeout allows us to continue
1939 * to do minor compactions to make
1940 * more memory available
1941 */
1942 continue;
1943 }
1944
1945 /*
1946 * Swap out segments?
1947 */
1948 if (flush_all == FALSE) {
1949 boolean_t needs_to_swap;
1950
1951 lck_mtx_unlock_always(c_list_lock);
1952
1953 needs_to_swap = compressor_needs_to_swap();
1954
1955 lck_mtx_lock_spin_always(c_list_lock);
1956
1957 if (needs_to_swap == FALSE)
1958 break;
1959 }
1960 if (queue_empty(&c_age_list_head))
1961 break;
1962 c_seg = (c_segment_t) queue_first(&c_age_list_head);
1963
1964 if (flush_all == TRUE && c_seg->c_generation_id > c_generation_id_flush_barrier)
1965 break;
1966
1967 if (c_seg->c_filling) {
1968 /*
1969 * we're at or near the head... no more work to do
1970 */
1971 break;
1972 }
1973 lck_mtx_lock_spin_always(&c_seg->c_lock);
1974
1975 if (c_seg->c_busy) {
1976
1977 lck_mtx_unlock_always(c_list_lock);
1978 c_seg_wait_on_busy(c_seg);
1979 lck_mtx_lock_spin_always(c_list_lock);
1980
1981 continue;
1982 }
fe8ab488 1983 C_SEG_BUSY(c_seg);
39236c6e
A
1984
1985 if (c_seg_do_minor_compaction_and_unlock(c_seg, FALSE, TRUE, TRUE)) {
1986 /*
1987 * found an empty c_segment and freed it
1988 * so go grab the next guy in the queue
1989 */
1990 continue;
1991 }
1992 /*
1993 * Major compaction
1994 */
1995 keep_compacting = TRUE;
1996
1997 while (keep_compacting == TRUE) {
1998
1999 assert(c_seg->c_busy);
2000
2001 /* look for another segment to consolidate */
2002
2003 c_seg_next = (c_segment_t) queue_next(&c_seg->c_age_list);
2004
2005 if (queue_end(&c_age_list_head, (queue_entry_t)c_seg_next))
2006 break;
2007
2008 if (c_seg_major_compact_ok(c_seg, c_seg_next) == FALSE)
2009 break;
2010
2011 lck_mtx_lock_spin_always(&c_seg_next->c_lock);
2012
2013 if (c_seg_next->c_busy) {
2014
2015 lck_mtx_unlock_always(c_list_lock);
2016 c_seg_wait_on_busy(c_seg_next);
2017 lck_mtx_lock_spin_always(c_list_lock);
2018
2019 continue;
2020 }
2021 /* grab that segment */
fe8ab488 2022 C_SEG_BUSY(c_seg_next);
39236c6e
A
2023
2024 if (c_seg_do_minor_compaction_and_unlock(c_seg_next, FALSE, TRUE, TRUE)) {
2025 /*
2026 * found an empty c_segment and freed it
2027 * so we can't continue to use c_seg_next
2028 */
2029 continue;
2030 }
2031
2032 /* unlock the list ... */
2033 lck_mtx_unlock_always(c_list_lock);
2034
2035 /* do the major compaction */
2036
2037 keep_compacting = c_seg_major_compact(c_seg, c_seg_next);
2038
2039 PAGE_REPLACEMENT_DISALLOWED(TRUE);
2040
2041 lck_mtx_lock_spin_always(&c_seg_next->c_lock);
2042 /*
2043 * run a minor compaction on the donor segment
2044 * since we pulled at least some of it's
2045 * data into our target... if we've emptied
2046 * it, now is a good time to free it which
2047 * c_seg_minor_compaction_and_unlock also takes care of
2048 *
2049 * by passing TRUE, we ask for c_busy to be cleared
2050 * and c_wanted to be taken care of
2051 */
2052 c_seg_minor_compaction_and_unlock(c_seg_next, TRUE);
2053
2054 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2055
2056 /* relock the list */
2057 lck_mtx_lock_spin_always(c_list_lock);
2058
2059 } /* major compaction */
2060
2061 c_seg_major_compact_stats.wasted_space_in_swapouts += C_SEG_BUFSIZE - c_seg->c_bytes_used;
2062 c_seg_major_compact_stats.count_of_swapouts++;
2063
2064 lck_mtx_lock_spin_always(&c_seg->c_lock);
2065
2066 assert(c_seg->c_busy);
2067 assert(c_seg->c_on_age_q);
2068 assert(!c_seg->c_on_minorcompact_q);
2069
2070 queue_remove(&c_age_list_head, c_seg, c_segment_t, c_age_list);
2071 c_seg->c_on_age_q = 0;
2072 c_age_count--;
2073
2074 if (vm_swap_up == TRUE) {
2075 queue_enter(&c_swapout_list_head, c_seg, c_segment_t, c_age_list);
2076 c_seg->c_on_swapout_q = 1;
2077 c_swapout_count++;
2078 } else {
2079 queue_enter(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
2080 c_seg->c_on_swappedout_q = 1;
2081 c_swappedout_count++;
2082 }
2083 C_SEG_WAKEUP_DONE(c_seg);
2084
2085 lck_mtx_unlock_always(&c_seg->c_lock);
2086
2087 if (c_swapout_count) {
2088 lck_mtx_unlock_always(c_list_lock);
2089
2090 thread_wakeup((event_t)&c_swapout_list_head);
2091
2092 lck_mtx_lock_spin_always(c_list_lock);
2093 }
2094 }
2095}
2096
2097
fe8ab488
A
2098static uint32_t no_paging_space_action_in_progress = 0;
2099extern void memorystatus_send_low_swap_note(void);
2100
2101
39236c6e
A
2102static c_segment_t
2103c_seg_allocate(c_segment_t *current_chead)
2104{
2105 clock_sec_t sec;
2106 clock_nsec_t nsec;
2107 c_segment_t c_seg;
2108 int slotarray;
2109
2110 if ( (c_seg = *current_chead) == NULL ) {
2111 uint32_t c_segno;
2112
fe8ab488
A
2113 if (vm_compressor_low_on_space() || vm_swap_low_on_space()) {
2114
2115 if (no_paging_space_action_in_progress == 0) {
2116
2117 if (OSCompareAndSwap(0, 1, (UInt32 *)&no_paging_space_action_in_progress)) {
2118
2119 if (no_paging_space_action()) {
2120 memorystatus_send_low_swap_note();
2121 }
2122
2123 no_paging_space_action_in_progress = 0;
2124 }
2125 }
2126 }
39236c6e
A
2127 KERNEL_DEBUG(0xe0400004 | DBG_FUNC_START, 0, 0, 0, 0, 0);
2128
2129 lck_mtx_lock_spin_always(c_list_lock);
2130
2131 while (c_segments_busy == TRUE) {
2132 assert_wait((event_t) (&c_segments_busy), THREAD_UNINT);
2133
2134 lck_mtx_unlock_always(c_list_lock);
2135
2136 thread_block(THREAD_CONTINUE_NULL);
2137
2138 lck_mtx_lock_spin_always(c_list_lock);
2139 }
2140 if (c_free_segno_head == (uint32_t)-1) {
2141
2142 if (c_segments_available >= c_segments_limit || c_segment_pages_compressed >= c_segment_pages_compressed_limit) {
2143 lck_mtx_unlock_always(c_list_lock);
2144
2145 KERNEL_DEBUG(0xe0400004 | DBG_FUNC_END, 0, 0, 0, 1, 0);
2146 return (NULL);
2147 }
2148 c_segments_busy = TRUE;
2149 lck_mtx_unlock_always(c_list_lock);
2150
2151 kernel_memory_populate(kernel_map, (vm_offset_t)c_segments_next_page, PAGE_SIZE, KMA_KOBJECT);
2152 c_segments_next_page += PAGE_SIZE;
2153
2154 for (c_segno = c_segments_available + 1; c_segno < (c_segments_available + C_SEGMENTS_PER_PAGE); c_segno++)
2155 c_segments[c_segno - 1].c_segno = c_segno;
2156
2157 lck_mtx_lock_spin_always(c_list_lock);
2158
2159 c_segments[c_segno - 1].c_segno = c_free_segno_head;
2160 c_free_segno_head = c_segments_available;
2161 c_segments_available += C_SEGMENTS_PER_PAGE;
2162
2163 c_segments_busy = FALSE;
2164 thread_wakeup((event_t) (&c_segments_busy));
2165 }
2166 c_segno = c_free_segno_head;
2167 c_free_segno_head = c_segments[c_segno].c_segno;
2168
2169 lck_mtx_unlock_always(c_list_lock);
2170
2171 c_seg = (c_segment_t)zalloc(compressor_segment_zone);
2172 bzero((char *)c_seg, sizeof(struct c_segment));
2173
2174 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) {
2175 zfree(compressor_segment_zone, c_seg);
2176
2177 lck_mtx_lock_spin_always(c_list_lock);
2178
2179 c_segments[c_segno].c_segno = c_free_segno_head;
2180 c_free_segno_head = c_segno;
2181
2182 lck_mtx_unlock_always(c_list_lock);
2183
2184 KERNEL_DEBUG(0xe0400004 | DBG_FUNC_END, 0, 0, 0, 2, 0);
2185
2186 return (NULL);
2187 }
fe8ab488 2188 OSAddAtomic64(C_SEG_ALLOCSIZE, &compressor_kvspace_used);
39236c6e
A
2189
2190#if __i386__ || __x86_64__
2191 lck_mtx_init(&c_seg->c_lock, &vm_compressor_lck_grp, &vm_compressor_lck_attr);
2192#else /* __i386__ || __x86_64__ */
2193 lck_spin_init(&c_seg->c_lock, &vm_compressor_lck_grp, &vm_compressor_lck_attr);
2194#endif /* __i386__ || __x86_64__ */
2195
2196 kernel_memory_populate(kernel_map, (vm_offset_t)(c_seg->c_store.c_buffer), 3 * PAGE_SIZE, KMA_COMPRESSOR);
2197
2198 c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(3 * PAGE_SIZE);
2199 c_seg->c_firstemptyslot = C_SLOT_MAX;
2200 c_seg->c_mysegno = c_segno;
2201 c_seg->c_filling = 1;
2202
2203 lck_mtx_lock_spin_always(c_list_lock);
2204
2205 c_segment_count++;
2206 c_segments[c_segno].c_seg = c_seg;
2207
2208 c_seg->c_generation_id = c_generation_id++;
2209
2210 queue_enter(&c_age_list_head, c_seg, c_segment_t, c_age_list);
2211 c_seg->c_on_age_q = 1;
2212 c_age_count++;
2213
2214 lck_mtx_unlock_always(c_list_lock);
2215
2216 clock_get_system_nanotime(&sec, &nsec);
2217 c_seg->c_creation_ts = (uint32_t)sec;
2218
2219 *current_chead = c_seg;
2220
2221 KERNEL_DEBUG(0xe0400004 | DBG_FUNC_END, c_seg, 0, 0, 3, 0);
2222 }
2223 slotarray = C_SEG_SLOTARRAY_FROM_INDEX(c_seg, c_seg->c_nextslot);
2224
2225 if (c_seg->c_slots[slotarray] == 0) {
2226 KERNEL_DEBUG(0xe0400008 | DBG_FUNC_START, 0, 0, 0, 0, 0);
2227
2228 c_seg->c_slots[slotarray] = (struct c_slot *)kalloc(sizeof(struct c_slot) * C_SEG_SLOT_ARRAY_SIZE);
2229
2230 KERNEL_DEBUG(0xe0400008 | DBG_FUNC_END, 0, 0, 0, 0, 0);
2231 }
2232
2233 PAGE_REPLACEMENT_DISALLOWED(TRUE);
2234
2235 lck_mtx_lock_spin_always(&c_seg->c_lock);
2236
2237 return (c_seg);
2238}
2239
2240
2241
2242static void
2243c_current_seg_filled(c_segment_t c_seg, c_segment_t *current_chead)
2244{
2245 uint32_t unused_bytes;
2246 uint32_t offset_to_depopulate;
2247
2248 unused_bytes = trunc_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset - c_seg->c_nextoffset));
2249
2250 if (unused_bytes) {
2251
2252 offset_to_depopulate = C_SEG_BYTES_TO_OFFSET(round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_nextoffset)));
2253
2254 /*
2255 * release the extra physical page(s) at the end of the segment
2256 */
2257 lck_mtx_unlock_always(&c_seg->c_lock);
2258
2259 kernel_memory_depopulate(
2260 kernel_map,
2261 (vm_offset_t) &c_seg->c_store.c_buffer[offset_to_depopulate],
2262 unused_bytes,
2263 KMA_COMPRESSOR);
2264
2265 lck_mtx_lock_spin_always(&c_seg->c_lock);
2266
2267 c_seg->c_populated_offset = offset_to_depopulate;
2268 }
2269 c_seg->c_filling = 0;
2270
2271 if (C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE)
2272 c_seg_need_delayed_compaction(c_seg);
2273
2274 lck_mtx_unlock_always(&c_seg->c_lock);
2275
2276 *current_chead = NULL;
2277}
2278
2279
2280/*
2281 * returns with c_seg locked
2282 */
2283void
2284c_seg_swapin_requeue(c_segment_t c_seg)
2285{
2286 clock_sec_t sec;
2287 clock_nsec_t nsec;
2288
2289 clock_get_system_nanotime(&sec, &nsec);
2290
2291 lck_mtx_lock_spin_always(c_list_lock);
2292 lck_mtx_lock_spin_always(&c_seg->c_lock);
2293
2294 if (c_seg->c_on_swappedout_q) {
2295 queue_remove(&c_swappedout_list_head, c_seg, c_segment_t, c_age_list);
2296 c_seg->c_on_swappedout_q = 0;
2297 c_swappedout_count--;
2298 } else {
2299 assert(c_seg->c_on_swappedout_sparse_q);
2300
2301 queue_remove(&c_swappedout_sparse_list_head, c_seg, c_segment_t, c_age_list);
2302 c_seg->c_on_swappedout_sparse_q = 0;
2303 c_swappedout_sparse_count--;
2304 }
2305 if (c_seg->c_store.c_buffer) {
2306 queue_enter(&c_swappedin_list_head, c_seg, c_segment_t, c_age_list);
2307 c_seg->c_on_swappedin_q = 1;
2308 c_swappedin_count++;
2309 }
2310#if TRACK_BAD_C_SEGMENTS
2311 else {
2312 queue_enter(&c_bad_list_head, c_seg, c_segment_t, c_age_list);
2313 c_seg->c_on_bad_q = 1;
2314 c_bad_count++;
2315 }
2316#endif
2317 c_seg->c_swappedin_ts = (uint32_t)sec;
2318 c_seg->c_ondisk = 0;
2319 c_seg->c_was_swapped_in = 1;
2320
2321 lck_mtx_unlock_always(c_list_lock);
2322}
2323
2324
2325
2326/*
2327 * c_seg has to be locked and is returned locked.
2328 * PAGE_REPLACMENT_DISALLOWED has to be TRUE on entry and is returned TRUE
2329 */
2330
2331void
2332c_seg_swapin(c_segment_t c_seg, boolean_t force_minor_compaction)
2333{
2334 vm_offset_t addr = 0;
2335 uint32_t io_size = 0;
2336 uint64_t f_offset;
2337
2338#if !CHECKSUM_THE_SWAP
2339 if (c_seg->c_ondisk)
2340 c_seg_trim_tail(c_seg);
2341#endif
2342 io_size = round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset));
2343 f_offset = c_seg->c_store.c_swap_handle;
fe8ab488
A
2344
2345 C_SEG_BUSY(c_seg);
39236c6e
A
2346 lck_mtx_unlock_always(&c_seg->c_lock);
2347
2348 if (c_seg->c_ondisk) {
2349
2350 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2351
2352 if (kernel_memory_allocate(kernel_map, &addr, C_SEG_ALLOCSIZE, 0, KMA_COMPRESSOR | KMA_VAONLY) != KERN_SUCCESS)
2353 panic("c_seg_swapin: kernel_memory_allocate failed\n");
2354
2355 kernel_memory_populate(kernel_map, addr, io_size, KMA_COMPRESSOR);
2356
2357 if (vm_swap_get(addr, f_offset, io_size) != KERN_SUCCESS) {
2358 PAGE_REPLACEMENT_DISALLOWED(TRUE);
2359
2360 kernel_memory_depopulate(kernel_map, addr, io_size, KMA_COMPRESSOR);
2361 kmem_free(kernel_map, addr, C_SEG_ALLOCSIZE);
2362
2363 c_seg->c_store.c_buffer = (int32_t*) NULL;
fe8ab488 2364 c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0);
39236c6e
A
2365 } else {
2366 c_seg->c_store.c_buffer = (int32_t*) addr;
fe8ab488 2367#if ENCRYPTED_SWAP
39236c6e 2368 vm_swap_decrypt(c_seg);
fe8ab488 2369#endif /* ENCRYPTED_SWAP */
39236c6e
A
2370
2371#if CHECKSUM_THE_SWAP
2372 if (c_seg->cseg_swap_size != io_size)
2373 panic("swapin size doesn't match swapout size");
2374
2375 if (c_seg->cseg_hash != hash_string((char*) c_seg->c_store.c_buffer, (int)io_size)) {
2376 panic("c_seg_swapin - Swap hash mismatch\n");
2377 }
2378#endif /* CHECKSUM_THE_SWAP */
2379
2380 PAGE_REPLACEMENT_DISALLOWED(TRUE);
2381
2382 if (force_minor_compaction == TRUE) {
2383 lck_mtx_lock_spin_always(&c_seg->c_lock);
2384
2385 c_seg_minor_compaction_and_unlock(c_seg, FALSE);
2386 }
2387 OSAddAtomic64(c_seg->c_bytes_used, &compressor_bytes_used);
fe8ab488 2388 OSAddAtomic64(C_SEG_ALLOCSIZE, &compressor_kvspace_used);
39236c6e
A
2389 }
2390 }
2391 c_seg_swapin_requeue(c_seg);
2392
2393 C_SEG_WAKEUP_DONE(c_seg);
2394}
2395
2396
2397static int
2398c_compress_page(char *src, c_slot_mapping_t slot_ptr, c_segment_t *current_chead, char *scratch_buf)
2399{
2400 int c_size;
2401 int c_rounded_size;
2402 int max_csize;
2403 c_slot_t cs;
2404 c_segment_t c_seg;
2405
2406 KERNEL_DEBUG(0xe0400000 | DBG_FUNC_START, *current_chead, 0, 0, 0, 0);
2407retry:
2408 if ((c_seg = c_seg_allocate(current_chead)) == NULL)
2409 return (1);
2410 /*
2411 * returns with c_seg lock held
2412 * and PAGE_REPLACEMENT_DISALLOWED(TRUE)
2413 */
2414 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_seg->c_nextslot);
2415
2416 cs->c_packed_ptr = C_SLOT_PACK_PTR(slot_ptr);
fe8ab488
A
2417 assert(slot_ptr == (c_slot_mapping_t)C_SLOT_UNPACK_PTR(cs));
2418
39236c6e
A
2419 cs->c_offset = c_seg->c_nextoffset;
2420
2421 max_csize = C_SEG_BUFSIZE - C_SEG_OFFSET_TO_BYTES((int32_t)cs->c_offset);
2422
2423 if (max_csize > PAGE_SIZE)
2424 max_csize = PAGE_SIZE;
2425
2426 if (C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset -
2427 c_seg->c_nextoffset)
2428 < (unsigned) max_csize + PAGE_SIZE &&
2429 (C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset)
2430 < C_SEG_ALLOCSIZE)) {
2431 lck_mtx_unlock_always(&c_seg->c_lock);
2432
2433 kernel_memory_populate(kernel_map,
2434 (vm_offset_t) &c_seg->c_store.c_buffer[c_seg->c_populated_offset],
2435 PAGE_SIZE,
2436 KMA_COMPRESSOR);
2437
2438 lck_mtx_lock_spin_always(&c_seg->c_lock);
2439
2440 c_seg->c_populated_offset += C_SEG_BYTES_TO_OFFSET(PAGE_SIZE);
2441 }
2442
2443#if CHECKSUM_THE_DATA
2444 cs->c_hash_data = hash_string(src, PAGE_SIZE);
2445#endif
39236c6e 2446
fe8ab488
A
2447 c_size = WKdm_compress_new((WK_word *)(uintptr_t)src, (WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
2448 (WK_word *)(uintptr_t)scratch_buf, max_csize - 4);
39236c6e
A
2449 assert(c_size <= (max_csize - 4) && c_size >= -1);
2450
2451 if (c_size == -1) {
2452
2453 if (max_csize < PAGE_SIZE) {
2454 c_current_seg_filled(c_seg, current_chead);
2455
2456 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2457
2458 goto retry;
2459 }
2460 c_size = PAGE_SIZE;
2461
2462 memcpy(&c_seg->c_store.c_buffer[cs->c_offset], src, c_size);
2463 }
2464#if CHECKSUM_THE_COMPRESSED_DATA
2465 cs->c_hash_compressed_data = hash_string((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size);
2466#endif
2467 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
2468
2469 PACK_C_SIZE(cs, c_size);
2470 c_seg->c_bytes_used += c_rounded_size;
2471 c_seg->c_nextoffset += C_SEG_BYTES_TO_OFFSET(c_rounded_size);
2472
2473 slot_ptr->s_cindx = c_seg->c_nextslot++;
2474 /* <csegno=0,indx=0> would mean "empty slot", so use csegno+1 */
2475 slot_ptr->s_cseg = c_seg->c_mysegno + 1;
2476
2477 if (c_seg->c_nextoffset >= C_SEG_OFF_LIMIT || c_seg->c_nextslot >= C_SLOT_MAX)
2478 c_current_seg_filled(c_seg, current_chead);
2479 else
2480 lck_mtx_unlock_always(&c_seg->c_lock);
2481
2482 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2483
2484 OSAddAtomic64(c_rounded_size, &compressor_bytes_used);
2485 OSAddAtomic64(PAGE_SIZE, &c_segment_input_bytes);
2486 OSAddAtomic64(c_size, &c_segment_compressed_bytes);
2487
2488 OSAddAtomic(1, &c_segment_pages_compressed);
2489 OSAddAtomic(1, &sample_period_compression_count);
2490
2491 KERNEL_DEBUG(0xe0400000 | DBG_FUNC_END, *current_chead, c_size, c_segment_input_bytes, c_segment_compressed_bytes, 0);
2492
39236c6e
A
2493 return (0);
2494}
2495
2496
2497static int
2498c_decompress_page(char *dst, volatile c_slot_mapping_t slot_ptr, int flags, int *zeroslot)
2499{
2500 c_slot_t cs;
2501 c_segment_t c_seg;
2502 int c_indx;
2503 int c_rounded_size;
2504 uint32_t c_size;
2505 int retval = 0;
2506 boolean_t c_seg_has_data = TRUE;
2507 boolean_t c_seg_swappedin = FALSE;
2508 boolean_t need_unlock = TRUE;
2509 boolean_t consider_defragmenting = FALSE;
2510
2511ReTry:
fe8ab488
A
2512 PAGE_REPLACEMENT_DISALLOWED(TRUE);
2513
39236c6e 2514#if HIBERNATION
fe8ab488
A
2515 /*
2516 * if hibernation is enabled, it indicates (via a call
2517 * to 'vm_decompressor_lock' that no further
2518 * decompressions are allowed once it reaches
2519 * the point of flushing all of the currently dirty
2520 * anonymous memory through the compressor and out
2521 * to disk... in this state we allow freeing of compressed
2522 * pages and must honor the C_DONT_BLOCK case
2523 */
2524 if (dst && decompressions_blocked == TRUE) {
2525 if (flags & C_DONT_BLOCK) {
2526
2527 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2528
2529 *zeroslot = 0;
2530 return (-2);
39236c6e 2531 }
fe8ab488
A
2532 /*
2533 * it's safe to atomically assert and block behind the
2534 * lock held in shared mode because "decompressions_blocked" is
2535 * only set and cleared and the thread_wakeup done when the lock
2536 * is held exclusively
2537 */
2538 assert_wait((event_t)&decompressions_blocked, THREAD_UNINT);
2539
2540 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2541
2542 thread_block(THREAD_CONTINUE_NULL);
2543
2544 goto ReTry;
39236c6e
A
2545 }
2546#endif
39236c6e
A
2547 /* s_cseg is actually "segno+1" */
2548 c_seg = c_segments[slot_ptr->s_cseg - 1].c_seg;
2549
2550 lck_mtx_lock_spin_always(&c_seg->c_lock);
2551
2552 if (flags & C_DONT_BLOCK) {
fe8ab488 2553 if (c_seg->c_busy || (c_seg->c_ondisk && dst)) {
39236c6e
A
2554
2555 retval = -2;
2556 *zeroslot = 0;
2557
2558 goto done;
2559 }
2560 }
2561 if (c_seg->c_busy) {
2562
2563 PAGE_REPLACEMENT_DISALLOWED(FALSE);
fe8ab488 2564
39236c6e
A
2565 c_seg_wait_on_busy(c_seg);
2566
2567 goto ReTry;
2568 }
2569 c_indx = slot_ptr->s_cindx;
2570
2571 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
2572
2573 c_size = UNPACK_C_SIZE(cs);
2574
2575 c_rounded_size = (c_size + C_SEG_OFFSET_ALIGNMENT_MASK) & ~C_SEG_OFFSET_ALIGNMENT_MASK;
2576
2577 if (dst) {
2578 uint32_t age_of_cseg;
2579 clock_sec_t cur_ts_sec;
2580 clock_nsec_t cur_ts_nsec;
2581
2582 if (c_seg->c_on_swappedout_q || c_seg->c_on_swappedout_sparse_q) {
2583 if (c_seg->c_ondisk)
2584 c_seg_swappedin = TRUE;
2585 c_seg_swapin(c_seg, FALSE);
2586 }
2587 if (c_seg->c_store.c_buffer == NULL) {
2588 c_seg_has_data = FALSE;
2589 goto c_seg_invalid_data;
2590 }
2591#if CHECKSUM_THE_COMPRESSED_DATA
2592 if (cs->c_hash_compressed_data != hash_string((char *)&c_seg->c_store.c_buffer[cs->c_offset], c_size))
2593 panic("compressed data doesn't match original");
2594#endif
2595 if (c_rounded_size == PAGE_SIZE) {
2596 /*
2597 * page wasn't compressible... just copy it out
2598 */
2599 memcpy(dst, &c_seg->c_store.c_buffer[cs->c_offset], PAGE_SIZE);
2600 } else {
2601 uint32_t my_cpu_no;
2602 char *scratch_buf;
2603
2604 /*
2605 * we're behind the c_seg lock held in spin mode
2606 * which means pre-emption is disabled... therefore
2607 * the following sequence is atomic and safe
2608 */
2609 my_cpu_no = cpu_number();
2610
2611 assert(my_cpu_no < compressor_cpus);
2612
2613 scratch_buf = &compressor_scratch_bufs[my_cpu_no * WKdm_SCRATCH_BUF_SIZE];
39236c6e
A
2614 WKdm_decompress_new((WK_word *)(uintptr_t)&c_seg->c_store.c_buffer[cs->c_offset],
2615 (WK_word *)(uintptr_t)dst, (WK_word *)(uintptr_t)scratch_buf, c_size);
2616 }
2617
2618#if CHECKSUM_THE_DATA
2619 if (cs->c_hash_data != hash_string(dst, PAGE_SIZE))
2620 panic("decompressed data doesn't match original");
2621#endif
2622 if (!c_seg->c_was_swapped_in) {
2623
2624 clock_get_system_nanotime(&cur_ts_sec, &cur_ts_nsec);
2625
2626 age_of_cseg = (uint32_t)cur_ts_sec - c_seg->c_creation_ts;
2627
2628 if (age_of_cseg < DECOMPRESSION_SAMPLE_MAX_AGE)
2629 OSAddAtomic(1, &age_of_decompressions_during_sample_period[age_of_cseg]);
2630 else
2631 OSAddAtomic(1, &overage_decompressions_during_sample_period);
2632
2633 OSAddAtomic(1, &sample_period_decompression_count);
2634 }
2635 } else {
2636 if (c_seg->c_store.c_buffer == NULL)
2637 c_seg_has_data = FALSE;
2638 }
2639c_seg_invalid_data:
2640
2641 if (c_seg_has_data == TRUE) {
2642 if (c_seg_swappedin == TRUE)
2643 retval = 1;
2644 else
2645 retval = 0;
2646 } else
2647 retval = -1;
2648
2649 if (flags & C_KEEP) {
2650 *zeroslot = 0;
2651 goto done;
2652 }
2653 c_seg->c_bytes_unused += c_rounded_size;
2654 c_seg->c_bytes_used -= c_rounded_size;
2655 PACK_C_SIZE(cs, 0);
2656
2657 if (c_indx < c_seg->c_firstemptyslot)
2658 c_seg->c_firstemptyslot = c_indx;
2659
2660 OSAddAtomic(-1, &c_segment_pages_compressed);
2661
2662 if (c_seg_has_data == TRUE && !c_seg->c_ondisk) {
2663 /*
2664 * c_ondisk == TRUE can occur when we're doing a
2665 * free of a compressed page (i.e. dst == NULL)
2666 */
2667 OSAddAtomic64(-c_rounded_size, &compressor_bytes_used);
2668 }
2669 if (!c_seg->c_filling) {
2670 if (c_seg->c_bytes_used == 0) {
8a3053a0
A
2671 if (!c_seg->c_ondisk) {
2672 int pages_populated;
2673
2674 pages_populated = (round_page_32(C_SEG_OFFSET_TO_BYTES(c_seg->c_populated_offset))) / PAGE_SIZE;
2675 c_seg->c_populated_offset = C_SEG_BYTES_TO_OFFSET(0);
2676
2677 if (pages_populated) {
2678 assert(c_seg->c_store.c_buffer != NULL);
2679
fe8ab488 2680 C_SEG_BUSY(c_seg);
8a3053a0
A
2681 lck_mtx_unlock_always(&c_seg->c_lock);
2682
2683 kernel_memory_depopulate(kernel_map, (vm_offset_t) c_seg->c_store.c_buffer, pages_populated * PAGE_SIZE, KMA_COMPRESSOR);
2684
2685 lck_mtx_lock_spin_always(&c_seg->c_lock);
2686 C_SEG_WAKEUP_DONE(c_seg);
2687 }
2688 if (!c_seg->c_on_minorcompact_q && !c_seg->c_on_swapout_q)
2689 c_seg_need_delayed_compaction(c_seg);
2690 } else
2691 assert(c_seg->c_on_swappedout_sparse_q);
2692
39236c6e
A
2693 } else if (c_seg->c_on_minorcompact_q) {
2694
2695 if (C_SEG_INCORE_IS_SPARSE(c_seg)) {
2696 c_seg_try_minor_compaction_and_unlock(c_seg);
2697 need_unlock = FALSE;
2698 }
2699 } else if (!c_seg->c_ondisk) {
2700
2701 if (c_seg_has_data == TRUE && !c_seg->c_on_swapout_q && C_SEG_UNUSED_BYTES(c_seg) >= PAGE_SIZE) {
2702 c_seg_need_delayed_compaction(c_seg);
2703 }
2704 } else if (!c_seg->c_on_swappedout_sparse_q && C_SEG_ONDISK_IS_SPARSE(c_seg)) {
2705
2706 c_seg_move_to_sparse_list(c_seg);
2707 consider_defragmenting = TRUE;
2708 }
2709 }
2710done:
2711 if (need_unlock == TRUE)
2712 lck_mtx_unlock_always(&c_seg->c_lock);
2713
2714 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2715
2716 if (consider_defragmenting == TRUE)
2717 vm_swap_consider_defragmenting();
fe8ab488
A
2718
2719
39236c6e
A
2720 return (retval);
2721}
2722
2723
2724int
2725vm_compressor_get(ppnum_t pn, int *slot, int flags)
2726{
2727 char *dst;
2728 int zeroslot = 1;
2729 int retval;
2730
2731#if __x86_64__
2732 dst = PHYSMAP_PTOV((uint64_t)pn << (uint64_t)PAGE_SHIFT);
2733#else
2734#error "unsupported architecture"
2735#endif
2736
2737 retval = c_decompress_page(dst, (c_slot_mapping_t)slot, flags, &zeroslot);
2738
2739 /*
2740 * zeroslot will be set to 0 by c_decompress_page if (flags & C_KEEP)
2741 * or (flags & C_DONT_BLOCK) and we found 'c_busy' or 'c_ondisk' set
2742 */
2743 if (zeroslot) {
39236c6e
A
2744 *slot = 0;
2745 }
2746 /*
2747 * returns 0 if we successfully decompressed a page from a segment already in memory
2748 * returns 1 if we had to first swap in the segment, before successfully decompressing the page
2749 * returns -1 if we encountered an error swapping in the segment - decompression failed
2750 * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' or 'c_ondisk' set
2751 */
2752 return (retval);
2753}
2754
2755
fe8ab488
A
2756int
2757vm_compressor_free(int *slot, int flags)
39236c6e
A
2758{
2759 int zeroslot = 1;
fe8ab488 2760 int retval;
39236c6e 2761
fe8ab488 2762 assert(flags == 0 || flags == C_DONT_BLOCK);
39236c6e 2763
fe8ab488
A
2764 retval = c_decompress_page(NULL, (c_slot_mapping_t)slot, flags, &zeroslot);
2765 /*
2766 * returns 0 if we successfully freed the specified compressed page
2767 * returns -2 if (flags & C_DONT_BLOCK) and we found 'c_busy' set
2768 */
2769
2770 if (retval == 0)
2771 *slot = 0;
2772
2773 return (retval);
39236c6e
A
2774}
2775
2776
2777int
2778vm_compressor_put(ppnum_t pn, int *slot, void **current_chead, char *scratch_buf)
2779{
2780 char *src;
2781 int retval;
2782
39236c6e
A
2783#if __x86_64__
2784 src = PHYSMAP_PTOV((uint64_t)pn << (uint64_t)PAGE_SHIFT);
2785#else
2786#error "unsupported architecture"
2787#endif
2788 retval = c_compress_page(src, (c_slot_mapping_t)slot, (c_segment_t *)current_chead, scratch_buf);
2789
2790 return (retval);
2791}
fe8ab488
A
2792
2793void
2794vm_compressor_transfer(
2795 int *dst_slot_p,
2796 int *src_slot_p)
2797{
2798 c_slot_mapping_t dst_slot, src_slot;
2799 c_segment_t c_seg;
2800 int c_indx;
2801 c_slot_t cs;
2802
2803 dst_slot = (c_slot_mapping_t) dst_slot_p;
2804 src_slot = (c_slot_mapping_t) src_slot_p;
2805
2806Retry:
2807 PAGE_REPLACEMENT_DISALLOWED(TRUE);
2808 /* get segment for src_slot */
2809 c_seg = c_segments[src_slot->s_cseg -1].c_seg;
2810 /* lock segment */
2811 lck_mtx_lock_spin_always(&c_seg->c_lock);
2812 /* wait if it's busy */
2813 if (c_seg->c_busy) {
2814 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2815 c_seg_wait_on_busy(c_seg);
2816 goto Retry;
2817 }
2818 /* find the c_slot */
2819 c_indx = src_slot->s_cindx;
2820 cs = C_SEG_SLOT_FROM_INDEX(c_seg, c_indx);
2821 /* point the c_slot back to dst_slot instead of src_slot */
2822 cs->c_packed_ptr = C_SLOT_PACK_PTR(dst_slot);
2823 /* transfer */
2824 *dst_slot_p = *src_slot_p;
2825 *src_slot_p = 0;
2826 lck_mtx_unlock_always(&c_seg->c_lock);
2827 PAGE_REPLACEMENT_DISALLOWED(FALSE);
2828}