Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2018 Red Hat. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include <linux/device-mapper.h>
9#include <linux/module.h>
10#include <linux/init.h>
11#include <linux/vmalloc.h>
12#include <linux/kthread.h>
13#include <linux/dm-io.h>
14#include <linux/dm-kcopyd.h>
15#include <linux/dax.h>
16#include <linux/pfn_t.h>
17#include <linux/libnvdimm.h>
18#include <linux/delay.h>
19#include "dm-io-tracker.h"
20
21#define DM_MSG_PREFIX "writecache"
22
23#define HIGH_WATERMARK 50
24#define LOW_WATERMARK 45
25#define MAX_WRITEBACK_JOBS min(0x10000000 / PAGE_SIZE, totalram_pages() / 16)
26#define ENDIO_LATENCY 16
27#define WRITEBACK_LATENCY 64
28#define AUTOCOMMIT_BLOCKS_SSD 65536
29#define AUTOCOMMIT_BLOCKS_PMEM 64
30#define AUTOCOMMIT_MSEC 1000
31#define MAX_AGE_DIV 16
32#define MAX_AGE_UNSPECIFIED -1UL
33#define PAUSE_WRITEBACK (HZ * 3)
34
35#define BITMAP_GRANULARITY 65536
36#if BITMAP_GRANULARITY < PAGE_SIZE
37#undef BITMAP_GRANULARITY
38#define BITMAP_GRANULARITY PAGE_SIZE
39#endif
40
41#if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_FS_DAX)
42#define DM_WRITECACHE_HAS_PMEM
43#endif
44
45#ifdef DM_WRITECACHE_HAS_PMEM
46#define pmem_assign(dest, src) \
47do { \
48 typeof(dest) uniq = (src); \
49 memcpy_flushcache(&(dest), &uniq, sizeof(dest)); \
50} while (0)
51#else
52#define pmem_assign(dest, src) ((dest) = (src))
53#endif
54
55#if IS_ENABLED(CONFIG_ARCH_HAS_COPY_MC) && defined(DM_WRITECACHE_HAS_PMEM)
56#define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
57#endif
58
59#define MEMORY_SUPERBLOCK_MAGIC 0x23489321
60#define MEMORY_SUPERBLOCK_VERSION 1
61
62struct wc_memory_entry {
63 __le64 original_sector;
64 __le64 seq_count;
65};
66
67struct wc_memory_superblock {
68 union {
69 struct {
70 __le32 magic;
71 __le32 version;
72 __le32 block_size;
73 __le32 pad;
74 __le64 n_blocks;
75 __le64 seq_count;
76 };
77 __le64 padding[8];
78 };
79 struct wc_memory_entry entries[];
80};
81
82struct wc_entry {
83 struct rb_node rb_node;
84 struct list_head lru;
85 unsigned short wc_list_contiguous;
86#if BITS_PER_LONG == 64
87 bool write_in_progress : 1;
88 unsigned long index : 47;
89#else
90 bool write_in_progress;
91 unsigned long index;
92#endif
93 unsigned long age;
94#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
95 uint64_t original_sector;
96 uint64_t seq_count;
97#endif
98};
99
100#ifdef DM_WRITECACHE_HAS_PMEM
101#define WC_MODE_PMEM(wc) ((wc)->pmem_mode)
102#define WC_MODE_FUA(wc) ((wc)->writeback_fua)
103#else
104#define WC_MODE_PMEM(wc) false
105#define WC_MODE_FUA(wc) false
106#endif
107#define WC_MODE_SORT_FREELIST(wc) (!WC_MODE_PMEM(wc))
108
109struct dm_writecache {
110 struct mutex lock;
111 struct list_head lru;
112 union {
113 struct list_head freelist;
114 struct {
115 struct rb_root freetree;
116 struct wc_entry *current_free;
117 };
118 };
119 struct rb_root tree;
120
121 size_t freelist_size;
122 size_t writeback_size;
123 size_t freelist_high_watermark;
124 size_t freelist_low_watermark;
125 unsigned long max_age;
126 unsigned long pause;
127
128 unsigned int uncommitted_blocks;
129 unsigned int autocommit_blocks;
130 unsigned int max_writeback_jobs;
131
132 int error;
133
134 unsigned long autocommit_jiffies;
135 struct timer_list autocommit_timer;
136 struct wait_queue_head freelist_wait;
137
138 struct timer_list max_age_timer;
139
140 atomic_t bio_in_progress[2];
141 struct wait_queue_head bio_in_progress_wait[2];
142
143 struct dm_target *ti;
144 struct dm_dev *dev;
145 struct dm_dev *ssd_dev;
146 sector_t start_sector;
147 void *memory_map;
148 uint64_t memory_map_size;
149 size_t metadata_sectors;
150 size_t n_blocks;
151 uint64_t seq_count;
152 sector_t data_device_sectors;
153 void *block_start;
154 struct wc_entry *entries;
155 unsigned int block_size;
156 unsigned char block_size_bits;
157
158 bool pmem_mode:1;
159 bool writeback_fua:1;
160
161 bool overwrote_committed:1;
162 bool memory_vmapped:1;
163
164 bool start_sector_set:1;
165 bool high_wm_percent_set:1;
166 bool low_wm_percent_set:1;
167 bool max_writeback_jobs_set:1;
168 bool autocommit_blocks_set:1;
169 bool autocommit_time_set:1;
170 bool max_age_set:1;
171 bool writeback_fua_set:1;
172 bool flush_on_suspend:1;
173 bool cleaner:1;
174 bool cleaner_set:1;
175 bool metadata_only:1;
176 bool pause_set:1;
177
178 unsigned int high_wm_percent_value;
179 unsigned int low_wm_percent_value;
180 unsigned int autocommit_time_value;
181 unsigned int max_age_value;
182 unsigned int pause_value;
183
184 unsigned int writeback_all;
185 struct workqueue_struct *writeback_wq;
186 struct work_struct writeback_work;
187 struct work_struct flush_work;
188
189 struct dm_io_tracker iot;
190
191 struct dm_io_client *dm_io;
192
193 raw_spinlock_t endio_list_lock;
194 struct list_head endio_list;
195 struct task_struct *endio_thread;
196
197 struct task_struct *flush_thread;
198 struct bio_list flush_list;
199
200 struct dm_kcopyd_client *dm_kcopyd;
201 unsigned long *dirty_bitmap;
202 unsigned int dirty_bitmap_size;
203
204 struct bio_set bio_set;
205 mempool_t copy_pool;
206
207 struct {
208 unsigned long long reads;
209 unsigned long long read_hits;
210 unsigned long long writes;
211 unsigned long long write_hits_uncommitted;
212 unsigned long long write_hits_committed;
213 unsigned long long writes_around;
214 unsigned long long writes_allocate;
215 unsigned long long writes_blocked_on_freelist;
216 unsigned long long flushes;
217 unsigned long long discards;
218 } stats;
219};
220
221#define WB_LIST_INLINE 16
222
223struct writeback_struct {
224 struct list_head endio_entry;
225 struct dm_writecache *wc;
226 struct wc_entry **wc_list;
227 unsigned int wc_list_n;
228 struct wc_entry *wc_list_inline[WB_LIST_INLINE];
229 struct bio bio;
230};
231
232struct copy_struct {
233 struct list_head endio_entry;
234 struct dm_writecache *wc;
235 struct wc_entry *e;
236 unsigned int n_entries;
237 int error;
238};
239
240DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
241 "A percentage of time allocated for data copying");
242
243static void wc_lock(struct dm_writecache *wc)
244{
245 mutex_lock(&wc->lock);
246}
247
248static void wc_unlock(struct dm_writecache *wc)
249{
250 mutex_unlock(&wc->lock);
251}
252
253#ifdef DM_WRITECACHE_HAS_PMEM
254static int persistent_memory_claim(struct dm_writecache *wc)
255{
256 int r;
257 loff_t s;
258 long p, da;
259 pfn_t pfn;
260 int id;
261 struct page **pages;
262 sector_t offset;
263
264 wc->memory_vmapped = false;
265
266 s = wc->memory_map_size;
267 p = s >> PAGE_SHIFT;
268 if (!p) {
269 r = -EINVAL;
270 goto err1;
271 }
272 if (p != s >> PAGE_SHIFT) {
273 r = -EOVERFLOW;
274 goto err1;
275 }
276
277 offset = get_start_sect(wc->ssd_dev->bdev);
278 if (offset & (PAGE_SIZE / 512 - 1)) {
279 r = -EINVAL;
280 goto err1;
281 }
282 offset >>= PAGE_SHIFT - 9;
283
284 id = dax_read_lock();
285
286 da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, DAX_ACCESS,
287 &wc->memory_map, &pfn);
288 if (da < 0) {
289 wc->memory_map = NULL;
290 r = da;
291 goto err2;
292 }
293 if (!pfn_t_has_page(pfn)) {
294 wc->memory_map = NULL;
295 r = -EOPNOTSUPP;
296 goto err2;
297 }
298 if (da != p) {
299 long i;
300
301 wc->memory_map = NULL;
302 pages = vmalloc_array(p, sizeof(struct page *));
303 if (!pages) {
304 r = -ENOMEM;
305 goto err2;
306 }
307 i = 0;
308 do {
309 long daa;
310
311 daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i,
312 p - i, DAX_ACCESS, NULL, &pfn);
313 if (daa <= 0) {
314 r = daa ? daa : -EINVAL;
315 goto err3;
316 }
317 if (!pfn_t_has_page(pfn)) {
318 r = -EOPNOTSUPP;
319 goto err3;
320 }
321 while (daa-- && i < p) {
322 pages[i++] = pfn_t_to_page(pfn);
323 pfn.val++;
324 if (!(i & 15))
325 cond_resched();
326 }
327 } while (i < p);
328 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
329 if (!wc->memory_map) {
330 r = -ENOMEM;
331 goto err3;
332 }
333 vfree(pages);
334 wc->memory_vmapped = true;
335 }
336
337 dax_read_unlock(id);
338
339 wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
340 wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
341
342 return 0;
343err3:
344 vfree(pages);
345err2:
346 dax_read_unlock(id);
347err1:
348 return r;
349}
350#else
351static int persistent_memory_claim(struct dm_writecache *wc)
352{
353 return -EOPNOTSUPP;
354}
355#endif
356
357static void persistent_memory_release(struct dm_writecache *wc)
358{
359 if (wc->memory_vmapped)
360 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
361}
362
363static struct page *persistent_memory_page(void *addr)
364{
365 if (is_vmalloc_addr(addr))
366 return vmalloc_to_page(addr);
367 else
368 return virt_to_page(addr);
369}
370
371static unsigned int persistent_memory_page_offset(void *addr)
372{
373 return (unsigned long)addr & (PAGE_SIZE - 1);
374}
375
376static void persistent_memory_flush_cache(void *ptr, size_t size)
377{
378 if (is_vmalloc_addr(ptr))
379 flush_kernel_vmap_range(ptr, size);
380}
381
382static void persistent_memory_invalidate_cache(void *ptr, size_t size)
383{
384 if (is_vmalloc_addr(ptr))
385 invalidate_kernel_vmap_range(ptr, size);
386}
387
388static struct wc_memory_superblock *sb(struct dm_writecache *wc)
389{
390 return wc->memory_map;
391}
392
393static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
394{
395 return &sb(wc)->entries[e->index];
396}
397
398static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
399{
400 return (char *)wc->block_start + (e->index << wc->block_size_bits);
401}
402
403static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
404{
405 return wc->start_sector + wc->metadata_sectors +
406 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
407}
408
409static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
410{
411#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
412 return e->original_sector;
413#else
414 return le64_to_cpu(memory_entry(wc, e)->original_sector);
415#endif
416}
417
418static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
419{
420#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
421 return e->seq_count;
422#else
423 return le64_to_cpu(memory_entry(wc, e)->seq_count);
424#endif
425}
426
427static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
428{
429#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
430 e->seq_count = -1;
431#endif
432 pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
433}
434
435static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
436 uint64_t original_sector, uint64_t seq_count)
437{
438 struct wc_memory_entry me;
439#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
440 e->original_sector = original_sector;
441 e->seq_count = seq_count;
442#endif
443 me.original_sector = cpu_to_le64(original_sector);
444 me.seq_count = cpu_to_le64(seq_count);
445 pmem_assign(*memory_entry(wc, e), me);
446}
447
448#define writecache_error(wc, err, msg, arg...) \
449do { \
450 if (!cmpxchg(&(wc)->error, 0, err)) \
451 DMERR(msg, ##arg); \
452 wake_up(&(wc)->freelist_wait); \
453} while (0)
454
455#define writecache_has_error(wc) (unlikely(READ_ONCE((wc)->error)))
456
457static void writecache_flush_all_metadata(struct dm_writecache *wc)
458{
459 if (!WC_MODE_PMEM(wc))
460 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
461}
462
463static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
464{
465 if (!WC_MODE_PMEM(wc))
466 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
467 wc->dirty_bitmap);
468}
469
470static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
471
472struct io_notify {
473 struct dm_writecache *wc;
474 struct completion c;
475 atomic_t count;
476};
477
478static void writecache_notify_io(unsigned long error, void *context)
479{
480 struct io_notify *endio = context;
481
482 if (unlikely(error != 0))
483 writecache_error(endio->wc, -EIO, "error writing metadata");
484 BUG_ON(atomic_read(&endio->count) <= 0);
485 if (atomic_dec_and_test(&endio->count))
486 complete(&endio->c);
487}
488
489static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
490{
491 wait_event(wc->bio_in_progress_wait[direction],
492 !atomic_read(&wc->bio_in_progress[direction]));
493}
494
495static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
496{
497 struct dm_io_region region;
498 struct dm_io_request req;
499 struct io_notify endio = {
500 wc,
501 COMPLETION_INITIALIZER_ONSTACK(endio.c),
502 ATOMIC_INIT(1),
503 };
504 unsigned int bitmap_bits = wc->dirty_bitmap_size * 8;
505 unsigned int i = 0;
506
507 while (1) {
508 unsigned int j;
509
510 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
511 if (unlikely(i == bitmap_bits))
512 break;
513 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
514
515 region.bdev = wc->ssd_dev->bdev;
516 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
517 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
518
519 if (unlikely(region.sector >= wc->metadata_sectors))
520 break;
521 if (unlikely(region.sector + region.count > wc->metadata_sectors))
522 region.count = wc->metadata_sectors - region.sector;
523
524 region.sector += wc->start_sector;
525 atomic_inc(&endio.count);
526 req.bi_opf = REQ_OP_WRITE | REQ_SYNC;
527 req.mem.type = DM_IO_VMA;
528 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
529 req.client = wc->dm_io;
530 req.notify.fn = writecache_notify_io;
531 req.notify.context = &endio;
532
533 /* writing via async dm-io (implied by notify.fn above) won't return an error */
534 (void) dm_io(&req, 1, ®ion, NULL);
535 i = j;
536 }
537
538 writecache_notify_io(0, &endio);
539 wait_for_completion_io(&endio.c);
540
541 if (wait_for_ios)
542 writecache_wait_for_ios(wc, WRITE);
543
544 writecache_disk_flush(wc, wc->ssd_dev);
545
546 memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
547}
548
549static void ssd_commit_superblock(struct dm_writecache *wc)
550{
551 int r;
552 struct dm_io_region region;
553 struct dm_io_request req;
554
555 region.bdev = wc->ssd_dev->bdev;
556 region.sector = 0;
557 region.count = max(4096U, wc->block_size) >> SECTOR_SHIFT;
558
559 if (unlikely(region.sector + region.count > wc->metadata_sectors))
560 region.count = wc->metadata_sectors - region.sector;
561
562 region.sector += wc->start_sector;
563
564 req.bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_FUA;
565 req.mem.type = DM_IO_VMA;
566 req.mem.ptr.vma = (char *)wc->memory_map;
567 req.client = wc->dm_io;
568 req.notify.fn = NULL;
569 req.notify.context = NULL;
570
571 r = dm_io(&req, 1, ®ion, NULL);
572 if (unlikely(r))
573 writecache_error(wc, r, "error writing superblock");
574}
575
576static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
577{
578 if (WC_MODE_PMEM(wc))
579 pmem_wmb();
580 else
581 ssd_commit_flushed(wc, wait_for_ios);
582}
583
584static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
585{
586 int r;
587 struct dm_io_region region;
588 struct dm_io_request req;
589
590 region.bdev = dev->bdev;
591 region.sector = 0;
592 region.count = 0;
593 req.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
594 req.mem.type = DM_IO_KMEM;
595 req.mem.ptr.addr = NULL;
596 req.client = wc->dm_io;
597 req.notify.fn = NULL;
598
599 r = dm_io(&req, 1, ®ion, NULL);
600 if (unlikely(r))
601 writecache_error(wc, r, "error flushing metadata: %d", r);
602}
603
604#define WFE_RETURN_FOLLOWING 1
605#define WFE_LOWEST_SEQ 2
606
607static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
608 uint64_t block, int flags)
609{
610 struct wc_entry *e;
611 struct rb_node *node = wc->tree.rb_node;
612
613 if (unlikely(!node))
614 return NULL;
615
616 while (1) {
617 e = container_of(node, struct wc_entry, rb_node);
618 if (read_original_sector(wc, e) == block)
619 break;
620
621 node = (read_original_sector(wc, e) >= block ?
622 e->rb_node.rb_left : e->rb_node.rb_right);
623 if (unlikely(!node)) {
624 if (!(flags & WFE_RETURN_FOLLOWING))
625 return NULL;
626 if (read_original_sector(wc, e) >= block)
627 return e;
628
629 node = rb_next(&e->rb_node);
630 if (unlikely(!node))
631 return NULL;
632
633 e = container_of(node, struct wc_entry, rb_node);
634 return e;
635 }
636 }
637
638 while (1) {
639 struct wc_entry *e2;
640
641 if (flags & WFE_LOWEST_SEQ)
642 node = rb_prev(&e->rb_node);
643 else
644 node = rb_next(&e->rb_node);
645 if (unlikely(!node))
646 return e;
647 e2 = container_of(node, struct wc_entry, rb_node);
648 if (read_original_sector(wc, e2) != block)
649 return e;
650 e = e2;
651 }
652}
653
654static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
655{
656 struct wc_entry *e;
657 struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
658
659 while (*node) {
660 e = container_of(*node, struct wc_entry, rb_node);
661 parent = &e->rb_node;
662 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
663 node = &parent->rb_left;
664 else
665 node = &parent->rb_right;
666 }
667 rb_link_node(&ins->rb_node, parent, node);
668 rb_insert_color(&ins->rb_node, &wc->tree);
669 list_add(&ins->lru, &wc->lru);
670 ins->age = jiffies;
671}
672
673static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
674{
675 list_del(&e->lru);
676 rb_erase(&e->rb_node, &wc->tree);
677}
678
679static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
680{
681 if (WC_MODE_SORT_FREELIST(wc)) {
682 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
683
684 if (unlikely(!*node))
685 wc->current_free = e;
686 while (*node) {
687 parent = *node;
688 if (&e->rb_node < *node)
689 node = &parent->rb_left;
690 else
691 node = &parent->rb_right;
692 }
693 rb_link_node(&e->rb_node, parent, node);
694 rb_insert_color(&e->rb_node, &wc->freetree);
695 } else {
696 list_add_tail(&e->lru, &wc->freelist);
697 }
698 wc->freelist_size++;
699}
700
701static inline void writecache_verify_watermark(struct dm_writecache *wc)
702{
703 if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
704 queue_work(wc->writeback_wq, &wc->writeback_work);
705}
706
707static void writecache_max_age_timer(struct timer_list *t)
708{
709 struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
710
711 if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
712 queue_work(wc->writeback_wq, &wc->writeback_work);
713 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
714 }
715}
716
717static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
718{
719 struct wc_entry *e;
720
721 if (WC_MODE_SORT_FREELIST(wc)) {
722 struct rb_node *next;
723
724 if (unlikely(!wc->current_free))
725 return NULL;
726 e = wc->current_free;
727 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
728 return NULL;
729 next = rb_next(&e->rb_node);
730 rb_erase(&e->rb_node, &wc->freetree);
731 if (unlikely(!next))
732 next = rb_first(&wc->freetree);
733 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
734 } else {
735 if (unlikely(list_empty(&wc->freelist)))
736 return NULL;
737 e = container_of(wc->freelist.next, struct wc_entry, lru);
738 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
739 return NULL;
740 list_del(&e->lru);
741 }
742 wc->freelist_size--;
743
744 writecache_verify_watermark(wc);
745
746 return e;
747}
748
749static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
750{
751 writecache_unlink(wc, e);
752 writecache_add_to_freelist(wc, e);
753 clear_seq_count(wc, e);
754 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
755 if (unlikely(waitqueue_active(&wc->freelist_wait)))
756 wake_up(&wc->freelist_wait);
757}
758
759static void writecache_wait_on_freelist(struct dm_writecache *wc)
760{
761 DEFINE_WAIT(wait);
762
763 prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
764 wc_unlock(wc);
765 io_schedule();
766 finish_wait(&wc->freelist_wait, &wait);
767 wc_lock(wc);
768}
769
770static void writecache_poison_lists(struct dm_writecache *wc)
771{
772 /*
773 * Catch incorrect access to these values while the device is suspended.
774 */
775 memset(&wc->tree, -1, sizeof(wc->tree));
776 wc->lru.next = LIST_POISON1;
777 wc->lru.prev = LIST_POISON2;
778 wc->freelist.next = LIST_POISON1;
779 wc->freelist.prev = LIST_POISON2;
780}
781
782static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
783{
784 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
785 if (WC_MODE_PMEM(wc))
786 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
787}
788
789static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
790{
791 return read_seq_count(wc, e) < wc->seq_count;
792}
793
794static void writecache_flush(struct dm_writecache *wc)
795{
796 struct wc_entry *e, *e2;
797 bool need_flush_after_free;
798
799 wc->uncommitted_blocks = 0;
800 del_timer(&wc->autocommit_timer);
801
802 if (list_empty(&wc->lru))
803 return;
804
805 e = container_of(wc->lru.next, struct wc_entry, lru);
806 if (writecache_entry_is_committed(wc, e)) {
807 if (wc->overwrote_committed) {
808 writecache_wait_for_ios(wc, WRITE);
809 writecache_disk_flush(wc, wc->ssd_dev);
810 wc->overwrote_committed = false;
811 }
812 return;
813 }
814 while (1) {
815 writecache_flush_entry(wc, e);
816 if (unlikely(e->lru.next == &wc->lru))
817 break;
818 e2 = container_of(e->lru.next, struct wc_entry, lru);
819 if (writecache_entry_is_committed(wc, e2))
820 break;
821 e = e2;
822 cond_resched();
823 }
824 writecache_commit_flushed(wc, true);
825
826 wc->seq_count++;
827 pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
828 if (WC_MODE_PMEM(wc))
829 writecache_commit_flushed(wc, false);
830 else
831 ssd_commit_superblock(wc);
832
833 wc->overwrote_committed = false;
834
835 need_flush_after_free = false;
836 while (1) {
837 /* Free another committed entry with lower seq-count */
838 struct rb_node *rb_node = rb_prev(&e->rb_node);
839
840 if (rb_node) {
841 e2 = container_of(rb_node, struct wc_entry, rb_node);
842 if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
843 likely(!e2->write_in_progress)) {
844 writecache_free_entry(wc, e2);
845 need_flush_after_free = true;
846 }
847 }
848 if (unlikely(e->lru.prev == &wc->lru))
849 break;
850 e = container_of(e->lru.prev, struct wc_entry, lru);
851 cond_resched();
852 }
853
854 if (need_flush_after_free)
855 writecache_commit_flushed(wc, false);
856}
857
858static void writecache_flush_work(struct work_struct *work)
859{
860 struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
861
862 wc_lock(wc);
863 writecache_flush(wc);
864 wc_unlock(wc);
865}
866
867static void writecache_autocommit_timer(struct timer_list *t)
868{
869 struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
870
871 if (!writecache_has_error(wc))
872 queue_work(wc->writeback_wq, &wc->flush_work);
873}
874
875static void writecache_schedule_autocommit(struct dm_writecache *wc)
876{
877 if (!timer_pending(&wc->autocommit_timer))
878 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
879}
880
881static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
882{
883 struct wc_entry *e;
884 bool discarded_something = false;
885
886 e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
887 if (unlikely(!e))
888 return;
889
890 while (read_original_sector(wc, e) < end) {
891 struct rb_node *node = rb_next(&e->rb_node);
892
893 if (likely(!e->write_in_progress)) {
894 if (!discarded_something) {
895 if (!WC_MODE_PMEM(wc)) {
896 writecache_wait_for_ios(wc, READ);
897 writecache_wait_for_ios(wc, WRITE);
898 }
899 discarded_something = true;
900 }
901 if (!writecache_entry_is_committed(wc, e))
902 wc->uncommitted_blocks--;
903 writecache_free_entry(wc, e);
904 }
905
906 if (unlikely(!node))
907 break;
908
909 e = container_of(node, struct wc_entry, rb_node);
910 }
911
912 if (discarded_something)
913 writecache_commit_flushed(wc, false);
914}
915
916static bool writecache_wait_for_writeback(struct dm_writecache *wc)
917{
918 if (wc->writeback_size) {
919 writecache_wait_on_freelist(wc);
920 return true;
921 }
922 return false;
923}
924
925static void writecache_suspend(struct dm_target *ti)
926{
927 struct dm_writecache *wc = ti->private;
928 bool flush_on_suspend;
929
930 del_timer_sync(&wc->autocommit_timer);
931 del_timer_sync(&wc->max_age_timer);
932
933 wc_lock(wc);
934 writecache_flush(wc);
935 flush_on_suspend = wc->flush_on_suspend;
936 if (flush_on_suspend) {
937 wc->flush_on_suspend = false;
938 wc->writeback_all++;
939 queue_work(wc->writeback_wq, &wc->writeback_work);
940 }
941 wc_unlock(wc);
942
943 drain_workqueue(wc->writeback_wq);
944
945 wc_lock(wc);
946 if (flush_on_suspend)
947 wc->writeback_all--;
948 while (writecache_wait_for_writeback(wc))
949 ;
950
951 if (WC_MODE_PMEM(wc))
952 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
953
954 writecache_poison_lists(wc);
955
956 wc_unlock(wc);
957}
958
959static int writecache_alloc_entries(struct dm_writecache *wc)
960{
961 size_t b;
962
963 if (wc->entries)
964 return 0;
965 wc->entries = vmalloc_array(wc->n_blocks, sizeof(struct wc_entry));
966 if (!wc->entries)
967 return -ENOMEM;
968 for (b = 0; b < wc->n_blocks; b++) {
969 struct wc_entry *e = &wc->entries[b];
970
971 e->index = b;
972 e->write_in_progress = false;
973 cond_resched();
974 }
975
976 return 0;
977}
978
979static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
980{
981 struct dm_io_region region;
982 struct dm_io_request req;
983
984 region.bdev = wc->ssd_dev->bdev;
985 region.sector = wc->start_sector;
986 region.count = n_sectors;
987 req.bi_opf = REQ_OP_READ | REQ_SYNC;
988 req.mem.type = DM_IO_VMA;
989 req.mem.ptr.vma = (char *)wc->memory_map;
990 req.client = wc->dm_io;
991 req.notify.fn = NULL;
992
993 return dm_io(&req, 1, ®ion, NULL);
994}
995
996static void writecache_resume(struct dm_target *ti)
997{
998 struct dm_writecache *wc = ti->private;
999 size_t b;
1000 bool need_flush = false;
1001 __le64 sb_seq_count;
1002 int r;
1003
1004 wc_lock(wc);
1005
1006 wc->data_device_sectors = bdev_nr_sectors(wc->dev->bdev);
1007
1008 if (WC_MODE_PMEM(wc)) {
1009 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
1010 } else {
1011 r = writecache_read_metadata(wc, wc->metadata_sectors);
1012 if (r) {
1013 size_t sb_entries_offset;
1014
1015 writecache_error(wc, r, "unable to read metadata: %d", r);
1016 sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
1017 memset((char *)wc->memory_map + sb_entries_offset, -1,
1018 (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
1019 }
1020 }
1021
1022 wc->tree = RB_ROOT;
1023 INIT_LIST_HEAD(&wc->lru);
1024 if (WC_MODE_SORT_FREELIST(wc)) {
1025 wc->freetree = RB_ROOT;
1026 wc->current_free = NULL;
1027 } else {
1028 INIT_LIST_HEAD(&wc->freelist);
1029 }
1030 wc->freelist_size = 0;
1031
1032 r = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
1033 sizeof(uint64_t));
1034 if (r) {
1035 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
1036 sb_seq_count = cpu_to_le64(0);
1037 }
1038 wc->seq_count = le64_to_cpu(sb_seq_count);
1039
1040#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1041 for (b = 0; b < wc->n_blocks; b++) {
1042 struct wc_entry *e = &wc->entries[b];
1043 struct wc_memory_entry wme;
1044
1045 if (writecache_has_error(wc)) {
1046 e->original_sector = -1;
1047 e->seq_count = -1;
1048 continue;
1049 }
1050 r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1051 sizeof(struct wc_memory_entry));
1052 if (r) {
1053 writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1054 (unsigned long)b, r);
1055 e->original_sector = -1;
1056 e->seq_count = -1;
1057 } else {
1058 e->original_sector = le64_to_cpu(wme.original_sector);
1059 e->seq_count = le64_to_cpu(wme.seq_count);
1060 }
1061 cond_resched();
1062 }
1063#endif
1064 for (b = 0; b < wc->n_blocks; b++) {
1065 struct wc_entry *e = &wc->entries[b];
1066
1067 if (!writecache_entry_is_committed(wc, e)) {
1068 if (read_seq_count(wc, e) != -1) {
1069erase_this:
1070 clear_seq_count(wc, e);
1071 need_flush = true;
1072 }
1073 writecache_add_to_freelist(wc, e);
1074 } else {
1075 struct wc_entry *old;
1076
1077 old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1078 if (!old) {
1079 writecache_insert_entry(wc, e);
1080 } else {
1081 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1082 writecache_error(wc, -EINVAL,
1083 "two identical entries, position %llu, sector %llu, sequence %llu",
1084 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1085 (unsigned long long)read_seq_count(wc, e));
1086 }
1087 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1088 goto erase_this;
1089 } else {
1090 writecache_free_entry(wc, old);
1091 writecache_insert_entry(wc, e);
1092 need_flush = true;
1093 }
1094 }
1095 }
1096 cond_resched();
1097 }
1098
1099 if (need_flush) {
1100 writecache_flush_all_metadata(wc);
1101 writecache_commit_flushed(wc, false);
1102 }
1103
1104 writecache_verify_watermark(wc);
1105
1106 if (wc->max_age != MAX_AGE_UNSPECIFIED)
1107 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1108
1109 wc_unlock(wc);
1110}
1111
1112static int process_flush_mesg(unsigned int argc, char **argv, struct dm_writecache *wc)
1113{
1114 if (argc != 1)
1115 return -EINVAL;
1116
1117 wc_lock(wc);
1118 if (dm_suspended(wc->ti)) {
1119 wc_unlock(wc);
1120 return -EBUSY;
1121 }
1122 if (writecache_has_error(wc)) {
1123 wc_unlock(wc);
1124 return -EIO;
1125 }
1126
1127 writecache_flush(wc);
1128 wc->writeback_all++;
1129 queue_work(wc->writeback_wq, &wc->writeback_work);
1130 wc_unlock(wc);
1131
1132 flush_workqueue(wc->writeback_wq);
1133
1134 wc_lock(wc);
1135 wc->writeback_all--;
1136 if (writecache_has_error(wc)) {
1137 wc_unlock(wc);
1138 return -EIO;
1139 }
1140 wc_unlock(wc);
1141
1142 return 0;
1143}
1144
1145static int process_flush_on_suspend_mesg(unsigned int argc, char **argv, struct dm_writecache *wc)
1146{
1147 if (argc != 1)
1148 return -EINVAL;
1149
1150 wc_lock(wc);
1151 wc->flush_on_suspend = true;
1152 wc_unlock(wc);
1153
1154 return 0;
1155}
1156
1157static void activate_cleaner(struct dm_writecache *wc)
1158{
1159 wc->flush_on_suspend = true;
1160 wc->cleaner = true;
1161 wc->freelist_high_watermark = wc->n_blocks;
1162 wc->freelist_low_watermark = wc->n_blocks;
1163}
1164
1165static int process_cleaner_mesg(unsigned int argc, char **argv, struct dm_writecache *wc)
1166{
1167 if (argc != 1)
1168 return -EINVAL;
1169
1170 wc_lock(wc);
1171 activate_cleaner(wc);
1172 if (!dm_suspended(wc->ti))
1173 writecache_verify_watermark(wc);
1174 wc_unlock(wc);
1175
1176 return 0;
1177}
1178
1179static int process_clear_stats_mesg(unsigned int argc, char **argv, struct dm_writecache *wc)
1180{
1181 if (argc != 1)
1182 return -EINVAL;
1183
1184 wc_lock(wc);
1185 memset(&wc->stats, 0, sizeof(wc->stats));
1186 wc_unlock(wc);
1187
1188 return 0;
1189}
1190
1191static int writecache_message(struct dm_target *ti, unsigned int argc, char **argv,
1192 char *result, unsigned int maxlen)
1193{
1194 int r = -EINVAL;
1195 struct dm_writecache *wc = ti->private;
1196
1197 if (!strcasecmp(argv[0], "flush"))
1198 r = process_flush_mesg(argc, argv, wc);
1199 else if (!strcasecmp(argv[0], "flush_on_suspend"))
1200 r = process_flush_on_suspend_mesg(argc, argv, wc);
1201 else if (!strcasecmp(argv[0], "cleaner"))
1202 r = process_cleaner_mesg(argc, argv, wc);
1203 else if (!strcasecmp(argv[0], "clear_stats"))
1204 r = process_clear_stats_mesg(argc, argv, wc);
1205 else
1206 DMERR("unrecognised message received: %s", argv[0]);
1207
1208 return r;
1209}
1210
1211static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1212{
1213 /*
1214 * clflushopt performs better with block size 1024, 2048, 4096
1215 * non-temporal stores perform better with block size 512
1216 *
1217 * block size 512 1024 2048 4096
1218 * movnti 496 MB/s 642 MB/s 725 MB/s 744 MB/s
1219 * clflushopt 373 MB/s 688 MB/s 1.1 GB/s 1.2 GB/s
1220 *
1221 * We see that movnti performs better for 512-byte blocks, and
1222 * clflushopt performs better for 1024-byte and larger blocks. So, we
1223 * prefer clflushopt for sizes >= 768.
1224 *
1225 * NOTE: this happens to be the case now (with dm-writecache's single
1226 * threaded model) but re-evaluate this once memcpy_flushcache() is
1227 * enabled to use movdir64b which might invalidate this performance
1228 * advantage seen with cache-allocating-writes plus flushing.
1229 */
1230#ifdef CONFIG_X86
1231 if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1232 likely(boot_cpu_data.x86_clflush_size == 64) &&
1233 likely(size >= 768)) {
1234 do {
1235 memcpy((void *)dest, (void *)source, 64);
1236 clflushopt((void *)dest);
1237 dest += 64;
1238 source += 64;
1239 size -= 64;
1240 } while (size >= 64);
1241 return;
1242 }
1243#endif
1244 memcpy_flushcache(dest, source, size);
1245}
1246
1247static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1248{
1249 void *buf;
1250 unsigned int size;
1251 int rw = bio_data_dir(bio);
1252 unsigned int remaining_size = wc->block_size;
1253
1254 do {
1255 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1256
1257 buf = bvec_kmap_local(&bv);
1258 size = bv.bv_len;
1259 if (unlikely(size > remaining_size))
1260 size = remaining_size;
1261
1262 if (rw == READ) {
1263 int r;
1264
1265 r = copy_mc_to_kernel(buf, data, size);
1266 flush_dcache_page(bio_page(bio));
1267 if (unlikely(r)) {
1268 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1269 bio->bi_status = BLK_STS_IOERR;
1270 }
1271 } else {
1272 flush_dcache_page(bio_page(bio));
1273 memcpy_flushcache_optimized(data, buf, size);
1274 }
1275
1276 kunmap_local(buf);
1277
1278 data = (char *)data + size;
1279 remaining_size -= size;
1280 bio_advance(bio, size);
1281 } while (unlikely(remaining_size));
1282}
1283
1284static int writecache_flush_thread(void *data)
1285{
1286 struct dm_writecache *wc = data;
1287
1288 while (1) {
1289 struct bio *bio;
1290
1291 wc_lock(wc);
1292 bio = bio_list_pop(&wc->flush_list);
1293 if (!bio) {
1294 set_current_state(TASK_INTERRUPTIBLE);
1295 wc_unlock(wc);
1296
1297 if (unlikely(kthread_should_stop())) {
1298 set_current_state(TASK_RUNNING);
1299 break;
1300 }
1301
1302 schedule();
1303 continue;
1304 }
1305
1306 if (bio_op(bio) == REQ_OP_DISCARD) {
1307 writecache_discard(wc, bio->bi_iter.bi_sector,
1308 bio_end_sector(bio));
1309 wc_unlock(wc);
1310 bio_set_dev(bio, wc->dev->bdev);
1311 submit_bio_noacct(bio);
1312 } else {
1313 writecache_flush(wc);
1314 wc_unlock(wc);
1315 if (writecache_has_error(wc))
1316 bio->bi_status = BLK_STS_IOERR;
1317 bio_endio(bio);
1318 }
1319 }
1320
1321 return 0;
1322}
1323
1324static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1325{
1326 if (bio_list_empty(&wc->flush_list))
1327 wake_up_process(wc->flush_thread);
1328 bio_list_add(&wc->flush_list, bio);
1329}
1330
1331enum wc_map_op {
1332 WC_MAP_SUBMIT,
1333 WC_MAP_REMAP,
1334 WC_MAP_REMAP_ORIGIN,
1335 WC_MAP_RETURN,
1336 WC_MAP_ERROR,
1337};
1338
1339static void writecache_map_remap_origin(struct dm_writecache *wc, struct bio *bio,
1340 struct wc_entry *e)
1341{
1342 if (e) {
1343 sector_t next_boundary =
1344 read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1345 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT)
1346 dm_accept_partial_bio(bio, next_boundary);
1347 }
1348}
1349
1350static enum wc_map_op writecache_map_read(struct dm_writecache *wc, struct bio *bio)
1351{
1352 enum wc_map_op map_op;
1353 struct wc_entry *e;
1354
1355read_next_block:
1356 wc->stats.reads++;
1357 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1358 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1359 wc->stats.read_hits++;
1360 if (WC_MODE_PMEM(wc)) {
1361 bio_copy_block(wc, bio, memory_data(wc, e));
1362 if (bio->bi_iter.bi_size)
1363 goto read_next_block;
1364 map_op = WC_MAP_SUBMIT;
1365 } else {
1366 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1367 bio_set_dev(bio, wc->ssd_dev->bdev);
1368 bio->bi_iter.bi_sector = cache_sector(wc, e);
1369 if (!writecache_entry_is_committed(wc, e))
1370 writecache_wait_for_ios(wc, WRITE);
1371 map_op = WC_MAP_REMAP;
1372 }
1373 } else {
1374 writecache_map_remap_origin(wc, bio, e);
1375 wc->stats.reads += (bio->bi_iter.bi_size - wc->block_size) >> wc->block_size_bits;
1376 map_op = WC_MAP_REMAP_ORIGIN;
1377 }
1378
1379 return map_op;
1380}
1381
1382static void writecache_bio_copy_ssd(struct dm_writecache *wc, struct bio *bio,
1383 struct wc_entry *e, bool search_used)
1384{
1385 unsigned int bio_size = wc->block_size;
1386 sector_t start_cache_sec = cache_sector(wc, e);
1387 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1388
1389 while (bio_size < bio->bi_iter.bi_size) {
1390 if (!search_used) {
1391 struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1392
1393 if (!f)
1394 break;
1395 write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1396 (bio_size >> SECTOR_SHIFT), wc->seq_count);
1397 writecache_insert_entry(wc, f);
1398 wc->uncommitted_blocks++;
1399 } else {
1400 struct wc_entry *f;
1401 struct rb_node *next = rb_next(&e->rb_node);
1402
1403 if (!next)
1404 break;
1405 f = container_of(next, struct wc_entry, rb_node);
1406 if (f != e + 1)
1407 break;
1408 if (read_original_sector(wc, f) !=
1409 read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1410 break;
1411 if (unlikely(f->write_in_progress))
1412 break;
1413 if (writecache_entry_is_committed(wc, f))
1414 wc->overwrote_committed = true;
1415 e = f;
1416 }
1417 bio_size += wc->block_size;
1418 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1419 }
1420
1421 bio_set_dev(bio, wc->ssd_dev->bdev);
1422 bio->bi_iter.bi_sector = start_cache_sec;
1423 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1424
1425 wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1426 wc->stats.writes_allocate += (bio->bi_iter.bi_size - wc->block_size) >> wc->block_size_bits;
1427
1428 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1429 wc->uncommitted_blocks = 0;
1430 queue_work(wc->writeback_wq, &wc->flush_work);
1431 } else {
1432 writecache_schedule_autocommit(wc);
1433 }
1434}
1435
1436static enum wc_map_op writecache_map_write(struct dm_writecache *wc, struct bio *bio)
1437{
1438 struct wc_entry *e;
1439
1440 do {
1441 bool found_entry = false;
1442 bool search_used = false;
1443
1444 if (writecache_has_error(wc)) {
1445 wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1446 return WC_MAP_ERROR;
1447 }
1448 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1449 if (e) {
1450 if (!writecache_entry_is_committed(wc, e)) {
1451 wc->stats.write_hits_uncommitted++;
1452 search_used = true;
1453 goto bio_copy;
1454 }
1455 wc->stats.write_hits_committed++;
1456 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1457 wc->overwrote_committed = true;
1458 search_used = true;
1459 goto bio_copy;
1460 }
1461 found_entry = true;
1462 } else {
1463 if (unlikely(wc->cleaner) ||
1464 (wc->metadata_only && !(bio->bi_opf & REQ_META)))
1465 goto direct_write;
1466 }
1467 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1468 if (unlikely(!e)) {
1469 if (!WC_MODE_PMEM(wc) && !found_entry) {
1470direct_write:
1471 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1472 writecache_map_remap_origin(wc, bio, e);
1473 wc->stats.writes_around += bio->bi_iter.bi_size >> wc->block_size_bits;
1474 wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1475 return WC_MAP_REMAP_ORIGIN;
1476 }
1477 wc->stats.writes_blocked_on_freelist++;
1478 writecache_wait_on_freelist(wc);
1479 continue;
1480 }
1481 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1482 writecache_insert_entry(wc, e);
1483 wc->uncommitted_blocks++;
1484 wc->stats.writes_allocate++;
1485bio_copy:
1486 if (WC_MODE_PMEM(wc)) {
1487 bio_copy_block(wc, bio, memory_data(wc, e));
1488 wc->stats.writes++;
1489 } else {
1490 writecache_bio_copy_ssd(wc, bio, e, search_used);
1491 return WC_MAP_REMAP;
1492 }
1493 } while (bio->bi_iter.bi_size);
1494
1495 if (unlikely(bio->bi_opf & REQ_FUA || wc->uncommitted_blocks >= wc->autocommit_blocks))
1496 writecache_flush(wc);
1497 else
1498 writecache_schedule_autocommit(wc);
1499
1500 return WC_MAP_SUBMIT;
1501}
1502
1503static enum wc_map_op writecache_map_flush(struct dm_writecache *wc, struct bio *bio)
1504{
1505 if (writecache_has_error(wc))
1506 return WC_MAP_ERROR;
1507
1508 if (WC_MODE_PMEM(wc)) {
1509 wc->stats.flushes++;
1510 writecache_flush(wc);
1511 if (writecache_has_error(wc))
1512 return WC_MAP_ERROR;
1513 else if (unlikely(wc->cleaner) || unlikely(wc->metadata_only))
1514 return WC_MAP_REMAP_ORIGIN;
1515 return WC_MAP_SUBMIT;
1516 }
1517 /* SSD: */
1518 if (dm_bio_get_target_bio_nr(bio))
1519 return WC_MAP_REMAP_ORIGIN;
1520 wc->stats.flushes++;
1521 writecache_offload_bio(wc, bio);
1522 return WC_MAP_RETURN;
1523}
1524
1525static enum wc_map_op writecache_map_discard(struct dm_writecache *wc, struct bio *bio)
1526{
1527 wc->stats.discards += bio->bi_iter.bi_size >> wc->block_size_bits;
1528
1529 if (writecache_has_error(wc))
1530 return WC_MAP_ERROR;
1531
1532 if (WC_MODE_PMEM(wc)) {
1533 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1534 return WC_MAP_REMAP_ORIGIN;
1535 }
1536 /* SSD: */
1537 writecache_offload_bio(wc, bio);
1538 return WC_MAP_RETURN;
1539}
1540
1541static int writecache_map(struct dm_target *ti, struct bio *bio)
1542{
1543 struct dm_writecache *wc = ti->private;
1544 enum wc_map_op map_op;
1545
1546 bio->bi_private = NULL;
1547
1548 wc_lock(wc);
1549
1550 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1551 map_op = writecache_map_flush(wc, bio);
1552 goto done;
1553 }
1554
1555 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1556
1557 if (unlikely((((unsigned int)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1558 (wc->block_size / 512 - 1)) != 0)) {
1559 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1560 (unsigned long long)bio->bi_iter.bi_sector,
1561 bio->bi_iter.bi_size, wc->block_size);
1562 map_op = WC_MAP_ERROR;
1563 goto done;
1564 }
1565
1566 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1567 map_op = writecache_map_discard(wc, bio);
1568 goto done;
1569 }
1570
1571 if (bio_data_dir(bio) == READ)
1572 map_op = writecache_map_read(wc, bio);
1573 else
1574 map_op = writecache_map_write(wc, bio);
1575done:
1576 switch (map_op) {
1577 case WC_MAP_REMAP_ORIGIN:
1578 if (likely(wc->pause != 0)) {
1579 if (bio_op(bio) == REQ_OP_WRITE) {
1580 dm_iot_io_begin(&wc->iot, 1);
1581 bio->bi_private = (void *)2;
1582 }
1583 }
1584 bio_set_dev(bio, wc->dev->bdev);
1585 wc_unlock(wc);
1586 return DM_MAPIO_REMAPPED;
1587
1588 case WC_MAP_REMAP:
1589 /* make sure that writecache_end_io decrements bio_in_progress: */
1590 bio->bi_private = (void *)1;
1591 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1592 wc_unlock(wc);
1593 return DM_MAPIO_REMAPPED;
1594
1595 case WC_MAP_SUBMIT:
1596 wc_unlock(wc);
1597 bio_endio(bio);
1598 return DM_MAPIO_SUBMITTED;
1599
1600 case WC_MAP_RETURN:
1601 wc_unlock(wc);
1602 return DM_MAPIO_SUBMITTED;
1603
1604 case WC_MAP_ERROR:
1605 wc_unlock(wc);
1606 bio_io_error(bio);
1607 return DM_MAPIO_SUBMITTED;
1608
1609 default:
1610 BUG();
1611 wc_unlock(wc);
1612 return DM_MAPIO_KILL;
1613 }
1614}
1615
1616static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1617{
1618 struct dm_writecache *wc = ti->private;
1619
1620 if (bio->bi_private == (void *)1) {
1621 int dir = bio_data_dir(bio);
1622
1623 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1624 if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1625 wake_up(&wc->bio_in_progress_wait[dir]);
1626 } else if (bio->bi_private == (void *)2) {
1627 dm_iot_io_end(&wc->iot, 1);
1628 }
1629 return 0;
1630}
1631
1632static int writecache_iterate_devices(struct dm_target *ti,
1633 iterate_devices_callout_fn fn, void *data)
1634{
1635 struct dm_writecache *wc = ti->private;
1636
1637 return fn(ti, wc->dev, 0, ti->len, data);
1638}
1639
1640static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1641{
1642 struct dm_writecache *wc = ti->private;
1643
1644 if (limits->logical_block_size < wc->block_size)
1645 limits->logical_block_size = wc->block_size;
1646
1647 if (limits->physical_block_size < wc->block_size)
1648 limits->physical_block_size = wc->block_size;
1649
1650 if (limits->io_min < wc->block_size)
1651 limits->io_min = wc->block_size;
1652}
1653
1654
1655static void writecache_writeback_endio(struct bio *bio)
1656{
1657 struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1658 struct dm_writecache *wc = wb->wc;
1659 unsigned long flags;
1660
1661 raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1662 if (unlikely(list_empty(&wc->endio_list)))
1663 wake_up_process(wc->endio_thread);
1664 list_add_tail(&wb->endio_entry, &wc->endio_list);
1665 raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1666}
1667
1668static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1669{
1670 struct copy_struct *c = ptr;
1671 struct dm_writecache *wc = c->wc;
1672
1673 c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1674
1675 raw_spin_lock_irq(&wc->endio_list_lock);
1676 if (unlikely(list_empty(&wc->endio_list)))
1677 wake_up_process(wc->endio_thread);
1678 list_add_tail(&c->endio_entry, &wc->endio_list);
1679 raw_spin_unlock_irq(&wc->endio_list_lock);
1680}
1681
1682static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1683{
1684 unsigned int i;
1685 struct writeback_struct *wb;
1686 struct wc_entry *e;
1687 unsigned long n_walked = 0;
1688
1689 do {
1690 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1691 list_del(&wb->endio_entry);
1692
1693 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1694 writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1695 "write error %d", wb->bio.bi_status);
1696 i = 0;
1697 do {
1698 e = wb->wc_list[i];
1699 BUG_ON(!e->write_in_progress);
1700 e->write_in_progress = false;
1701 INIT_LIST_HEAD(&e->lru);
1702 if (!writecache_has_error(wc))
1703 writecache_free_entry(wc, e);
1704 BUG_ON(!wc->writeback_size);
1705 wc->writeback_size--;
1706 n_walked++;
1707 if (unlikely(n_walked >= ENDIO_LATENCY)) {
1708 writecache_commit_flushed(wc, false);
1709 wc_unlock(wc);
1710 wc_lock(wc);
1711 n_walked = 0;
1712 }
1713 } while (++i < wb->wc_list_n);
1714
1715 if (wb->wc_list != wb->wc_list_inline)
1716 kfree(wb->wc_list);
1717 bio_put(&wb->bio);
1718 } while (!list_empty(list));
1719}
1720
1721static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1722{
1723 struct copy_struct *c;
1724 struct wc_entry *e;
1725
1726 do {
1727 c = list_entry(list->next, struct copy_struct, endio_entry);
1728 list_del(&c->endio_entry);
1729
1730 if (unlikely(c->error))
1731 writecache_error(wc, c->error, "copy error");
1732
1733 e = c->e;
1734 do {
1735 BUG_ON(!e->write_in_progress);
1736 e->write_in_progress = false;
1737 INIT_LIST_HEAD(&e->lru);
1738 if (!writecache_has_error(wc))
1739 writecache_free_entry(wc, e);
1740
1741 BUG_ON(!wc->writeback_size);
1742 wc->writeback_size--;
1743 e++;
1744 } while (--c->n_entries);
1745 mempool_free(c, &wc->copy_pool);
1746 } while (!list_empty(list));
1747}
1748
1749static int writecache_endio_thread(void *data)
1750{
1751 struct dm_writecache *wc = data;
1752
1753 while (1) {
1754 struct list_head list;
1755
1756 raw_spin_lock_irq(&wc->endio_list_lock);
1757 if (!list_empty(&wc->endio_list))
1758 goto pop_from_list;
1759 set_current_state(TASK_INTERRUPTIBLE);
1760 raw_spin_unlock_irq(&wc->endio_list_lock);
1761
1762 if (unlikely(kthread_should_stop())) {
1763 set_current_state(TASK_RUNNING);
1764 break;
1765 }
1766
1767 schedule();
1768
1769 continue;
1770
1771pop_from_list:
1772 list = wc->endio_list;
1773 list.next->prev = list.prev->next = &list;
1774 INIT_LIST_HEAD(&wc->endio_list);
1775 raw_spin_unlock_irq(&wc->endio_list_lock);
1776
1777 if (!WC_MODE_FUA(wc))
1778 writecache_disk_flush(wc, wc->dev);
1779
1780 wc_lock(wc);
1781
1782 if (WC_MODE_PMEM(wc)) {
1783 __writecache_endio_pmem(wc, &list);
1784 } else {
1785 __writecache_endio_ssd(wc, &list);
1786 writecache_wait_for_ios(wc, READ);
1787 }
1788
1789 writecache_commit_flushed(wc, false);
1790
1791 wc_unlock(wc);
1792 }
1793
1794 return 0;
1795}
1796
1797static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e)
1798{
1799 struct dm_writecache *wc = wb->wc;
1800 unsigned int block_size = wc->block_size;
1801 void *address = memory_data(wc, e);
1802
1803 persistent_memory_flush_cache(address, block_size);
1804
1805 if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1806 return true;
1807
1808 return bio_add_page(&wb->bio, persistent_memory_page(address),
1809 block_size, persistent_memory_page_offset(address)) != 0;
1810}
1811
1812struct writeback_list {
1813 struct list_head list;
1814 size_t size;
1815};
1816
1817static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1818{
1819 if (unlikely(wc->max_writeback_jobs)) {
1820 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1821 wc_lock(wc);
1822 while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1823 writecache_wait_on_freelist(wc);
1824 wc_unlock(wc);
1825 }
1826 }
1827 cond_resched();
1828}
1829
1830static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1831{
1832 struct wc_entry *e, *f;
1833 struct bio *bio;
1834 struct writeback_struct *wb;
1835 unsigned int max_pages;
1836
1837 while (wbl->size) {
1838 wbl->size--;
1839 e = container_of(wbl->list.prev, struct wc_entry, lru);
1840 list_del(&e->lru);
1841
1842 max_pages = e->wc_list_contiguous;
1843
1844 bio = bio_alloc_bioset(wc->dev->bdev, max_pages, REQ_OP_WRITE,
1845 GFP_NOIO, &wc->bio_set);
1846 wb = container_of(bio, struct writeback_struct, bio);
1847 wb->wc = wc;
1848 bio->bi_end_io = writecache_writeback_endio;
1849 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1850
1851 if (unlikely(max_pages > WB_LIST_INLINE))
1852 wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1853 GFP_NOIO | __GFP_NORETRY |
1854 __GFP_NOMEMALLOC | __GFP_NOWARN);
1855
1856 if (likely(max_pages <= WB_LIST_INLINE) || unlikely(!wb->wc_list)) {
1857 wb->wc_list = wb->wc_list_inline;
1858 max_pages = WB_LIST_INLINE;
1859 }
1860
1861 BUG_ON(!wc_add_block(wb, e));
1862
1863 wb->wc_list[0] = e;
1864 wb->wc_list_n = 1;
1865
1866 while (wbl->size && wb->wc_list_n < max_pages) {
1867 f = container_of(wbl->list.prev, struct wc_entry, lru);
1868 if (read_original_sector(wc, f) !=
1869 read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1870 break;
1871 if (!wc_add_block(wb, f))
1872 break;
1873 wbl->size--;
1874 list_del(&f->lru);
1875 wb->wc_list[wb->wc_list_n++] = f;
1876 e = f;
1877 }
1878 if (WC_MODE_FUA(wc))
1879 bio->bi_opf |= REQ_FUA;
1880 if (writecache_has_error(wc)) {
1881 bio->bi_status = BLK_STS_IOERR;
1882 bio_endio(bio);
1883 } else if (unlikely(!bio_sectors(bio))) {
1884 bio->bi_status = BLK_STS_OK;
1885 bio_endio(bio);
1886 } else {
1887 submit_bio(bio);
1888 }
1889
1890 __writeback_throttle(wc, wbl);
1891 }
1892}
1893
1894static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1895{
1896 struct wc_entry *e, *f;
1897 struct dm_io_region from, to;
1898 struct copy_struct *c;
1899
1900 while (wbl->size) {
1901 unsigned int n_sectors;
1902
1903 wbl->size--;
1904 e = container_of(wbl->list.prev, struct wc_entry, lru);
1905 list_del(&e->lru);
1906
1907 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1908
1909 from.bdev = wc->ssd_dev->bdev;
1910 from.sector = cache_sector(wc, e);
1911 from.count = n_sectors;
1912 to.bdev = wc->dev->bdev;
1913 to.sector = read_original_sector(wc, e);
1914 to.count = n_sectors;
1915
1916 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1917 c->wc = wc;
1918 c->e = e;
1919 c->n_entries = e->wc_list_contiguous;
1920
1921 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1922 wbl->size--;
1923 f = container_of(wbl->list.prev, struct wc_entry, lru);
1924 BUG_ON(f != e + 1);
1925 list_del(&f->lru);
1926 e = f;
1927 }
1928
1929 if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1930 if (to.sector >= wc->data_device_sectors) {
1931 writecache_copy_endio(0, 0, c);
1932 continue;
1933 }
1934 from.count = to.count = wc->data_device_sectors - to.sector;
1935 }
1936
1937 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1938
1939 __writeback_throttle(wc, wbl);
1940 }
1941}
1942
1943static void writecache_writeback(struct work_struct *work)
1944{
1945 struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1946 struct blk_plug plug;
1947 struct wc_entry *f, *g, *e = NULL;
1948 struct rb_node *node, *next_node;
1949 struct list_head skipped;
1950 struct writeback_list wbl;
1951 unsigned long n_walked;
1952
1953 if (!WC_MODE_PMEM(wc)) {
1954 /* Wait for any active kcopyd work on behalf of ssd writeback */
1955 dm_kcopyd_client_flush(wc->dm_kcopyd);
1956 }
1957
1958 if (likely(wc->pause != 0)) {
1959 while (1) {
1960 unsigned long idle;
1961
1962 if (unlikely(wc->cleaner) || unlikely(wc->writeback_all) ||
1963 unlikely(dm_suspended(wc->ti)))
1964 break;
1965 idle = dm_iot_idle_time(&wc->iot);
1966 if (idle >= wc->pause)
1967 break;
1968 idle = wc->pause - idle;
1969 if (idle > HZ)
1970 idle = HZ;
1971 schedule_timeout_idle(idle);
1972 }
1973 }
1974
1975 wc_lock(wc);
1976restart:
1977 if (writecache_has_error(wc)) {
1978 wc_unlock(wc);
1979 return;
1980 }
1981
1982 if (unlikely(wc->writeback_all)) {
1983 if (writecache_wait_for_writeback(wc))
1984 goto restart;
1985 }
1986
1987 if (wc->overwrote_committed)
1988 writecache_wait_for_ios(wc, WRITE);
1989
1990 n_walked = 0;
1991 INIT_LIST_HEAD(&skipped);
1992 INIT_LIST_HEAD(&wbl.list);
1993 wbl.size = 0;
1994 while (!list_empty(&wc->lru) &&
1995 (wc->writeback_all ||
1996 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1997 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1998 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1999
2000 n_walked++;
2001 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
2002 likely(!wc->writeback_all)) {
2003 if (likely(!dm_suspended(wc->ti)))
2004 queue_work(wc->writeback_wq, &wc->writeback_work);
2005 break;
2006 }
2007
2008 if (unlikely(wc->writeback_all)) {
2009 if (unlikely(!e)) {
2010 writecache_flush(wc);
2011 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
2012 } else
2013 e = g;
2014 } else
2015 e = container_of(wc->lru.prev, struct wc_entry, lru);
2016 BUG_ON(e->write_in_progress);
2017 if (unlikely(!writecache_entry_is_committed(wc, e)))
2018 writecache_flush(wc);
2019
2020 node = rb_prev(&e->rb_node);
2021 if (node) {
2022 f = container_of(node, struct wc_entry, rb_node);
2023 if (unlikely(read_original_sector(wc, f) ==
2024 read_original_sector(wc, e))) {
2025 BUG_ON(!f->write_in_progress);
2026 list_move(&e->lru, &skipped);
2027 cond_resched();
2028 continue;
2029 }
2030 }
2031 wc->writeback_size++;
2032 list_move(&e->lru, &wbl.list);
2033 wbl.size++;
2034 e->write_in_progress = true;
2035 e->wc_list_contiguous = 1;
2036
2037 f = e;
2038
2039 while (1) {
2040 next_node = rb_next(&f->rb_node);
2041 if (unlikely(!next_node))
2042 break;
2043 g = container_of(next_node, struct wc_entry, rb_node);
2044 if (unlikely(read_original_sector(wc, g) ==
2045 read_original_sector(wc, f))) {
2046 f = g;
2047 continue;
2048 }
2049 if (read_original_sector(wc, g) !=
2050 read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
2051 break;
2052 if (unlikely(g->write_in_progress))
2053 break;
2054 if (unlikely(!writecache_entry_is_committed(wc, g)))
2055 break;
2056
2057 if (!WC_MODE_PMEM(wc)) {
2058 if (g != f + 1)
2059 break;
2060 }
2061
2062 n_walked++;
2063 //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
2064 // break;
2065
2066 wc->writeback_size++;
2067 list_move(&g->lru, &wbl.list);
2068 wbl.size++;
2069 g->write_in_progress = true;
2070 g->wc_list_contiguous = BIO_MAX_VECS;
2071 f = g;
2072 e->wc_list_contiguous++;
2073 if (unlikely(e->wc_list_contiguous == BIO_MAX_VECS)) {
2074 if (unlikely(wc->writeback_all)) {
2075 next_node = rb_next(&f->rb_node);
2076 if (likely(next_node))
2077 g = container_of(next_node, struct wc_entry, rb_node);
2078 }
2079 break;
2080 }
2081 }
2082 cond_resched();
2083 }
2084
2085 if (!list_empty(&skipped)) {
2086 list_splice_tail(&skipped, &wc->lru);
2087 /*
2088 * If we didn't do any progress, we must wait until some
2089 * writeback finishes to avoid burning CPU in a loop
2090 */
2091 if (unlikely(!wbl.size))
2092 writecache_wait_for_writeback(wc);
2093 }
2094
2095 wc_unlock(wc);
2096
2097 blk_start_plug(&plug);
2098
2099 if (WC_MODE_PMEM(wc))
2100 __writecache_writeback_pmem(wc, &wbl);
2101 else
2102 __writecache_writeback_ssd(wc, &wbl);
2103
2104 blk_finish_plug(&plug);
2105
2106 if (unlikely(wc->writeback_all)) {
2107 wc_lock(wc);
2108 while (writecache_wait_for_writeback(wc))
2109 ;
2110 wc_unlock(wc);
2111 }
2112}
2113
2114static int calculate_memory_size(uint64_t device_size, unsigned int block_size,
2115 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
2116{
2117 uint64_t n_blocks, offset;
2118 struct wc_entry e;
2119
2120 n_blocks = device_size;
2121 do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
2122
2123 while (1) {
2124 if (!n_blocks)
2125 return -ENOSPC;
2126 /* Verify the following entries[n_blocks] won't overflow */
2127 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
2128 sizeof(struct wc_memory_entry)))
2129 return -EFBIG;
2130 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
2131 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
2132 if (offset + n_blocks * block_size <= device_size)
2133 break;
2134 n_blocks--;
2135 }
2136
2137 /* check if the bit field overflows */
2138 e.index = n_blocks;
2139 if (e.index != n_blocks)
2140 return -EFBIG;
2141
2142 if (n_blocks_p)
2143 *n_blocks_p = n_blocks;
2144 if (n_metadata_blocks_p)
2145 *n_metadata_blocks_p = offset >> __ffs(block_size);
2146 return 0;
2147}
2148
2149static int init_memory(struct dm_writecache *wc)
2150{
2151 size_t b;
2152 int r;
2153
2154 r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
2155 if (r)
2156 return r;
2157
2158 r = writecache_alloc_entries(wc);
2159 if (r)
2160 return r;
2161
2162 for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
2163 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
2164 pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
2165 pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
2166 pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
2167 pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
2168
2169 for (b = 0; b < wc->n_blocks; b++) {
2170 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
2171 cond_resched();
2172 }
2173
2174 writecache_flush_all_metadata(wc);
2175 writecache_commit_flushed(wc, false);
2176 pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
2177 writecache_flush_region(wc, &sb(wc)->magic, sizeof(sb(wc)->magic));
2178 writecache_commit_flushed(wc, false);
2179
2180 return 0;
2181}
2182
2183static void writecache_dtr(struct dm_target *ti)
2184{
2185 struct dm_writecache *wc = ti->private;
2186
2187 if (!wc)
2188 return;
2189
2190 if (wc->endio_thread)
2191 kthread_stop(wc->endio_thread);
2192
2193 if (wc->flush_thread)
2194 kthread_stop(wc->flush_thread);
2195
2196 bioset_exit(&wc->bio_set);
2197
2198 mempool_exit(&wc->copy_pool);
2199
2200 if (wc->writeback_wq)
2201 destroy_workqueue(wc->writeback_wq);
2202
2203 if (wc->dev)
2204 dm_put_device(ti, wc->dev);
2205
2206 if (wc->ssd_dev)
2207 dm_put_device(ti, wc->ssd_dev);
2208
2209 vfree(wc->entries);
2210
2211 if (wc->memory_map) {
2212 if (WC_MODE_PMEM(wc))
2213 persistent_memory_release(wc);
2214 else
2215 vfree(wc->memory_map);
2216 }
2217
2218 if (wc->dm_kcopyd)
2219 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2220
2221 if (wc->dm_io)
2222 dm_io_client_destroy(wc->dm_io);
2223
2224 vfree(wc->dirty_bitmap);
2225
2226 kfree(wc);
2227}
2228
2229static int writecache_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2230{
2231 struct dm_writecache *wc;
2232 struct dm_arg_set as;
2233 const char *string;
2234 unsigned int opt_params;
2235 size_t offset, data_size;
2236 int i, r;
2237 char dummy;
2238 int high_wm_percent = HIGH_WATERMARK;
2239 int low_wm_percent = LOW_WATERMARK;
2240 uint64_t x;
2241 struct wc_memory_superblock s;
2242
2243 static struct dm_arg _args[] = {
2244 {0, 18, "Invalid number of feature args"},
2245 };
2246
2247 as.argc = argc;
2248 as.argv = argv;
2249
2250 wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2251 if (!wc) {
2252 ti->error = "Cannot allocate writecache structure";
2253 r = -ENOMEM;
2254 goto bad;
2255 }
2256 ti->private = wc;
2257 wc->ti = ti;
2258
2259 mutex_init(&wc->lock);
2260 wc->max_age = MAX_AGE_UNSPECIFIED;
2261 writecache_poison_lists(wc);
2262 init_waitqueue_head(&wc->freelist_wait);
2263 timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2264 timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2265
2266 for (i = 0; i < 2; i++) {
2267 atomic_set(&wc->bio_in_progress[i], 0);
2268 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2269 }
2270
2271 wc->dm_io = dm_io_client_create();
2272 if (IS_ERR(wc->dm_io)) {
2273 r = PTR_ERR(wc->dm_io);
2274 ti->error = "Unable to allocate dm-io client";
2275 wc->dm_io = NULL;
2276 goto bad;
2277 }
2278
2279 wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2280 if (!wc->writeback_wq) {
2281 r = -ENOMEM;
2282 ti->error = "Could not allocate writeback workqueue";
2283 goto bad;
2284 }
2285 INIT_WORK(&wc->writeback_work, writecache_writeback);
2286 INIT_WORK(&wc->flush_work, writecache_flush_work);
2287
2288 dm_iot_init(&wc->iot);
2289
2290 raw_spin_lock_init(&wc->endio_list_lock);
2291 INIT_LIST_HEAD(&wc->endio_list);
2292 wc->endio_thread = kthread_run(writecache_endio_thread, wc, "writecache_endio");
2293 if (IS_ERR(wc->endio_thread)) {
2294 r = PTR_ERR(wc->endio_thread);
2295 wc->endio_thread = NULL;
2296 ti->error = "Couldn't spawn endio thread";
2297 goto bad;
2298 }
2299
2300 /*
2301 * Parse the mode (pmem or ssd)
2302 */
2303 string = dm_shift_arg(&as);
2304 if (!string)
2305 goto bad_arguments;
2306
2307 if (!strcasecmp(string, "s")) {
2308 wc->pmem_mode = false;
2309 } else if (!strcasecmp(string, "p")) {
2310#ifdef DM_WRITECACHE_HAS_PMEM
2311 wc->pmem_mode = true;
2312 wc->writeback_fua = true;
2313#else
2314 /*
2315 * If the architecture doesn't support persistent memory or
2316 * the kernel doesn't support any DAX drivers, this driver can
2317 * only be used in SSD-only mode.
2318 */
2319 r = -EOPNOTSUPP;
2320 ti->error = "Persistent memory or DAX not supported on this system";
2321 goto bad;
2322#endif
2323 } else {
2324 goto bad_arguments;
2325 }
2326
2327 if (WC_MODE_PMEM(wc)) {
2328 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2329 offsetof(struct writeback_struct, bio),
2330 BIOSET_NEED_BVECS);
2331 if (r) {
2332 ti->error = "Could not allocate bio set";
2333 goto bad;
2334 }
2335 } else {
2336 wc->pause = PAUSE_WRITEBACK;
2337 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2338 if (r) {
2339 ti->error = "Could not allocate mempool";
2340 goto bad;
2341 }
2342 }
2343
2344 /*
2345 * Parse the origin data device
2346 */
2347 string = dm_shift_arg(&as);
2348 if (!string)
2349 goto bad_arguments;
2350 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2351 if (r) {
2352 ti->error = "Origin data device lookup failed";
2353 goto bad;
2354 }
2355
2356 /*
2357 * Parse cache data device (be it pmem or ssd)
2358 */
2359 string = dm_shift_arg(&as);
2360 if (!string)
2361 goto bad_arguments;
2362
2363 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2364 if (r) {
2365 ti->error = "Cache data device lookup failed";
2366 goto bad;
2367 }
2368 wc->memory_map_size = bdev_nr_bytes(wc->ssd_dev->bdev);
2369
2370 /*
2371 * Parse the cache block size
2372 */
2373 string = dm_shift_arg(&as);
2374 if (!string)
2375 goto bad_arguments;
2376 if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2377 wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2378 (wc->block_size & (wc->block_size - 1))) {
2379 r = -EINVAL;
2380 ti->error = "Invalid block size";
2381 goto bad;
2382 }
2383 if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2384 wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2385 r = -EINVAL;
2386 ti->error = "Block size is smaller than device logical block size";
2387 goto bad;
2388 }
2389 wc->block_size_bits = __ffs(wc->block_size);
2390
2391 wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2392 wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2393 wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2394
2395 /*
2396 * Parse optional arguments
2397 */
2398 r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2399 if (r)
2400 goto bad;
2401
2402 while (opt_params) {
2403 string = dm_shift_arg(&as), opt_params--;
2404 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2405 unsigned long long start_sector;
2406
2407 string = dm_shift_arg(&as), opt_params--;
2408 if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2409 goto invalid_optional;
2410 wc->start_sector = start_sector;
2411 wc->start_sector_set = true;
2412 if (wc->start_sector != start_sector ||
2413 wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2414 goto invalid_optional;
2415 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2416 string = dm_shift_arg(&as), opt_params--;
2417 if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2418 goto invalid_optional;
2419 if (high_wm_percent < 0 || high_wm_percent > 100)
2420 goto invalid_optional;
2421 wc->high_wm_percent_value = high_wm_percent;
2422 wc->high_wm_percent_set = true;
2423 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2424 string = dm_shift_arg(&as), opt_params--;
2425 if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2426 goto invalid_optional;
2427 if (low_wm_percent < 0 || low_wm_percent > 100)
2428 goto invalid_optional;
2429 wc->low_wm_percent_value = low_wm_percent;
2430 wc->low_wm_percent_set = true;
2431 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2432 string = dm_shift_arg(&as), opt_params--;
2433 if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2434 goto invalid_optional;
2435 wc->max_writeback_jobs_set = true;
2436 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2437 string = dm_shift_arg(&as), opt_params--;
2438 if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2439 goto invalid_optional;
2440 wc->autocommit_blocks_set = true;
2441 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2442 unsigned int autocommit_msecs;
2443
2444 string = dm_shift_arg(&as), opt_params--;
2445 if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2446 goto invalid_optional;
2447 if (autocommit_msecs > 3600000)
2448 goto invalid_optional;
2449 wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2450 wc->autocommit_time_value = autocommit_msecs;
2451 wc->autocommit_time_set = true;
2452 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2453 unsigned int max_age_msecs;
2454
2455 string = dm_shift_arg(&as), opt_params--;
2456 if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2457 goto invalid_optional;
2458 if (max_age_msecs > 86400000)
2459 goto invalid_optional;
2460 wc->max_age = msecs_to_jiffies(max_age_msecs);
2461 wc->max_age_set = true;
2462 wc->max_age_value = max_age_msecs;
2463 } else if (!strcasecmp(string, "cleaner")) {
2464 wc->cleaner_set = true;
2465 wc->cleaner = true;
2466 } else if (!strcasecmp(string, "fua")) {
2467 if (WC_MODE_PMEM(wc)) {
2468 wc->writeback_fua = true;
2469 wc->writeback_fua_set = true;
2470 } else
2471 goto invalid_optional;
2472 } else if (!strcasecmp(string, "nofua")) {
2473 if (WC_MODE_PMEM(wc)) {
2474 wc->writeback_fua = false;
2475 wc->writeback_fua_set = true;
2476 } else
2477 goto invalid_optional;
2478 } else if (!strcasecmp(string, "metadata_only")) {
2479 wc->metadata_only = true;
2480 } else if (!strcasecmp(string, "pause_writeback") && opt_params >= 1) {
2481 unsigned int pause_msecs;
2482
2483 if (WC_MODE_PMEM(wc))
2484 goto invalid_optional;
2485 string = dm_shift_arg(&as), opt_params--;
2486 if (sscanf(string, "%u%c", &pause_msecs, &dummy) != 1)
2487 goto invalid_optional;
2488 if (pause_msecs > 60000)
2489 goto invalid_optional;
2490 wc->pause = msecs_to_jiffies(pause_msecs);
2491 wc->pause_set = true;
2492 wc->pause_value = pause_msecs;
2493 } else {
2494invalid_optional:
2495 r = -EINVAL;
2496 ti->error = "Invalid optional argument";
2497 goto bad;
2498 }
2499 }
2500
2501 if (high_wm_percent < low_wm_percent) {
2502 r = -EINVAL;
2503 ti->error = "High watermark must be greater than or equal to low watermark";
2504 goto bad;
2505 }
2506
2507 if (WC_MODE_PMEM(wc)) {
2508 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2509 r = -EOPNOTSUPP;
2510 ti->error = "Asynchronous persistent memory not supported as pmem cache";
2511 goto bad;
2512 }
2513
2514 r = persistent_memory_claim(wc);
2515 if (r) {
2516 ti->error = "Unable to map persistent memory for cache";
2517 goto bad;
2518 }
2519 } else {
2520 size_t n_blocks, n_metadata_blocks;
2521 uint64_t n_bitmap_bits;
2522
2523 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2524
2525 bio_list_init(&wc->flush_list);
2526 wc->flush_thread = kthread_run(writecache_flush_thread, wc, "dm_writecache_flush");
2527 if (IS_ERR(wc->flush_thread)) {
2528 r = PTR_ERR(wc->flush_thread);
2529 wc->flush_thread = NULL;
2530 ti->error = "Couldn't spawn flush thread";
2531 goto bad;
2532 }
2533
2534 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2535 &n_blocks, &n_metadata_blocks);
2536 if (r) {
2537 ti->error = "Invalid device size";
2538 goto bad;
2539 }
2540
2541 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2542 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2543 /* this is limitation of test_bit functions */
2544 if (n_bitmap_bits > 1U << 31) {
2545 r = -EFBIG;
2546 ti->error = "Invalid device size";
2547 goto bad;
2548 }
2549
2550 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2551 if (!wc->memory_map) {
2552 r = -ENOMEM;
2553 ti->error = "Unable to allocate memory for metadata";
2554 goto bad;
2555 }
2556
2557 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2558 if (IS_ERR(wc->dm_kcopyd)) {
2559 r = PTR_ERR(wc->dm_kcopyd);
2560 ti->error = "Unable to allocate dm-kcopyd client";
2561 wc->dm_kcopyd = NULL;
2562 goto bad;
2563 }
2564
2565 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2566 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2567 BITS_PER_LONG * sizeof(unsigned long);
2568 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2569 if (!wc->dirty_bitmap) {
2570 r = -ENOMEM;
2571 ti->error = "Unable to allocate dirty bitmap";
2572 goto bad;
2573 }
2574
2575 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2576 if (r) {
2577 ti->error = "Unable to read first block of metadata";
2578 goto bad;
2579 }
2580 }
2581
2582 r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2583 if (r) {
2584 ti->error = "Hardware memory error when reading superblock";
2585 goto bad;
2586 }
2587 if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2588 r = init_memory(wc);
2589 if (r) {
2590 ti->error = "Unable to initialize device";
2591 goto bad;
2592 }
2593 r = copy_mc_to_kernel(&s, sb(wc),
2594 sizeof(struct wc_memory_superblock));
2595 if (r) {
2596 ti->error = "Hardware memory error when reading superblock";
2597 goto bad;
2598 }
2599 }
2600
2601 if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2602 ti->error = "Invalid magic in the superblock";
2603 r = -EINVAL;
2604 goto bad;
2605 }
2606
2607 if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2608 ti->error = "Invalid version in the superblock";
2609 r = -EINVAL;
2610 goto bad;
2611 }
2612
2613 if (le32_to_cpu(s.block_size) != wc->block_size) {
2614 ti->error = "Block size does not match superblock";
2615 r = -EINVAL;
2616 goto bad;
2617 }
2618
2619 wc->n_blocks = le64_to_cpu(s.n_blocks);
2620
2621 offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2622 if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2623overflow:
2624 ti->error = "Overflow in size calculation";
2625 r = -EINVAL;
2626 goto bad;
2627 }
2628 offset += sizeof(struct wc_memory_superblock);
2629 if (offset < sizeof(struct wc_memory_superblock))
2630 goto overflow;
2631 offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2632 data_size = wc->n_blocks * (size_t)wc->block_size;
2633 if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2634 (offset + data_size < offset))
2635 goto overflow;
2636 if (offset + data_size > wc->memory_map_size) {
2637 ti->error = "Memory area is too small";
2638 r = -EINVAL;
2639 goto bad;
2640 }
2641
2642 wc->metadata_sectors = offset >> SECTOR_SHIFT;
2643 wc->block_start = (char *)sb(wc) + offset;
2644
2645 x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2646 x += 50;
2647 do_div(x, 100);
2648 wc->freelist_high_watermark = x;
2649 x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2650 x += 50;
2651 do_div(x, 100);
2652 wc->freelist_low_watermark = x;
2653
2654 if (wc->cleaner)
2655 activate_cleaner(wc);
2656
2657 r = writecache_alloc_entries(wc);
2658 if (r) {
2659 ti->error = "Cannot allocate memory";
2660 goto bad;
2661 }
2662
2663 ti->num_flush_bios = WC_MODE_PMEM(wc) ? 1 : 2;
2664 ti->flush_supported = true;
2665 ti->num_discard_bios = 1;
2666
2667 if (WC_MODE_PMEM(wc))
2668 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2669
2670 return 0;
2671
2672bad_arguments:
2673 r = -EINVAL;
2674 ti->error = "Bad arguments";
2675bad:
2676 writecache_dtr(ti);
2677 return r;
2678}
2679
2680static void writecache_status(struct dm_target *ti, status_type_t type,
2681 unsigned int status_flags, char *result, unsigned int maxlen)
2682{
2683 struct dm_writecache *wc = ti->private;
2684 unsigned int extra_args;
2685 unsigned int sz = 0;
2686
2687 switch (type) {
2688 case STATUSTYPE_INFO:
2689 DMEMIT("%ld %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu",
2690 writecache_has_error(wc),
2691 (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2692 (unsigned long long)wc->writeback_size,
2693 wc->stats.reads,
2694 wc->stats.read_hits,
2695 wc->stats.writes,
2696 wc->stats.write_hits_uncommitted,
2697 wc->stats.write_hits_committed,
2698 wc->stats.writes_around,
2699 wc->stats.writes_allocate,
2700 wc->stats.writes_blocked_on_freelist,
2701 wc->stats.flushes,
2702 wc->stats.discards);
2703 break;
2704 case STATUSTYPE_TABLE:
2705 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2706 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2707 extra_args = 0;
2708 if (wc->start_sector_set)
2709 extra_args += 2;
2710 if (wc->high_wm_percent_set)
2711 extra_args += 2;
2712 if (wc->low_wm_percent_set)
2713 extra_args += 2;
2714 if (wc->max_writeback_jobs_set)
2715 extra_args += 2;
2716 if (wc->autocommit_blocks_set)
2717 extra_args += 2;
2718 if (wc->autocommit_time_set)
2719 extra_args += 2;
2720 if (wc->max_age_set)
2721 extra_args += 2;
2722 if (wc->cleaner_set)
2723 extra_args++;
2724 if (wc->writeback_fua_set)
2725 extra_args++;
2726 if (wc->metadata_only)
2727 extra_args++;
2728 if (wc->pause_set)
2729 extra_args += 2;
2730
2731 DMEMIT("%u", extra_args);
2732 if (wc->start_sector_set)
2733 DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2734 if (wc->high_wm_percent_set)
2735 DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2736 if (wc->low_wm_percent_set)
2737 DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2738 if (wc->max_writeback_jobs_set)
2739 DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2740 if (wc->autocommit_blocks_set)
2741 DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2742 if (wc->autocommit_time_set)
2743 DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2744 if (wc->max_age_set)
2745 DMEMIT(" max_age %u", wc->max_age_value);
2746 if (wc->cleaner_set)
2747 DMEMIT(" cleaner");
2748 if (wc->writeback_fua_set)
2749 DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2750 if (wc->metadata_only)
2751 DMEMIT(" metadata_only");
2752 if (wc->pause_set)
2753 DMEMIT(" pause_writeback %u", wc->pause_value);
2754 break;
2755 case STATUSTYPE_IMA:
2756 *result = '\0';
2757 break;
2758 }
2759}
2760
2761static struct target_type writecache_target = {
2762 .name = "writecache",
2763 .version = {1, 6, 0},
2764 .module = THIS_MODULE,
2765 .ctr = writecache_ctr,
2766 .dtr = writecache_dtr,
2767 .status = writecache_status,
2768 .postsuspend = writecache_suspend,
2769 .resume = writecache_resume,
2770 .message = writecache_message,
2771 .map = writecache_map,
2772 .end_io = writecache_end_io,
2773 .iterate_devices = writecache_iterate_devices,
2774 .io_hints = writecache_io_hints,
2775};
2776module_dm(writecache);
2777
2778MODULE_DESCRIPTION(DM_NAME " writecache target");
2779MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2780MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2018 Red Hat. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include <linux/device-mapper.h>
9#include <linux/module.h>
10#include <linux/init.h>
11#include <linux/vmalloc.h>
12#include <linux/kthread.h>
13#include <linux/dm-io.h>
14#include <linux/dm-kcopyd.h>
15#include <linux/dax.h>
16#include <linux/pfn_t.h>
17#include <linux/libnvdimm.h>
18
19#define DM_MSG_PREFIX "writecache"
20
21#define HIGH_WATERMARK 50
22#define LOW_WATERMARK 45
23#define MAX_WRITEBACK_JOBS 0
24#define ENDIO_LATENCY 16
25#define WRITEBACK_LATENCY 64
26#define AUTOCOMMIT_BLOCKS_SSD 65536
27#define AUTOCOMMIT_BLOCKS_PMEM 64
28#define AUTOCOMMIT_MSEC 1000
29#define MAX_AGE_DIV 16
30#define MAX_AGE_UNSPECIFIED -1UL
31
32#define BITMAP_GRANULARITY 65536
33#if BITMAP_GRANULARITY < PAGE_SIZE
34#undef BITMAP_GRANULARITY
35#define BITMAP_GRANULARITY PAGE_SIZE
36#endif
37
38#if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_DAX_DRIVER)
39#define DM_WRITECACHE_HAS_PMEM
40#endif
41
42#ifdef DM_WRITECACHE_HAS_PMEM
43#define pmem_assign(dest, src) \
44do { \
45 typeof(dest) uniq = (src); \
46 memcpy_flushcache(&(dest), &uniq, sizeof(dest)); \
47} while (0)
48#else
49#define pmem_assign(dest, src) ((dest) = (src))
50#endif
51
52#if defined(__HAVE_ARCH_MEMCPY_MCSAFE) && defined(DM_WRITECACHE_HAS_PMEM)
53#define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
54#endif
55
56#define MEMORY_SUPERBLOCK_MAGIC 0x23489321
57#define MEMORY_SUPERBLOCK_VERSION 1
58
59struct wc_memory_entry {
60 __le64 original_sector;
61 __le64 seq_count;
62};
63
64struct wc_memory_superblock {
65 union {
66 struct {
67 __le32 magic;
68 __le32 version;
69 __le32 block_size;
70 __le32 pad;
71 __le64 n_blocks;
72 __le64 seq_count;
73 };
74 __le64 padding[8];
75 };
76 struct wc_memory_entry entries[0];
77};
78
79struct wc_entry {
80 struct rb_node rb_node;
81 struct list_head lru;
82 unsigned short wc_list_contiguous;
83 bool write_in_progress
84#if BITS_PER_LONG == 64
85 :1
86#endif
87 ;
88 unsigned long index
89#if BITS_PER_LONG == 64
90 :47
91#endif
92 ;
93 unsigned long age;
94#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
95 uint64_t original_sector;
96 uint64_t seq_count;
97#endif
98};
99
100#ifdef DM_WRITECACHE_HAS_PMEM
101#define WC_MODE_PMEM(wc) ((wc)->pmem_mode)
102#define WC_MODE_FUA(wc) ((wc)->writeback_fua)
103#else
104#define WC_MODE_PMEM(wc) false
105#define WC_MODE_FUA(wc) false
106#endif
107#define WC_MODE_SORT_FREELIST(wc) (!WC_MODE_PMEM(wc))
108
109struct dm_writecache {
110 struct mutex lock;
111 struct list_head lru;
112 union {
113 struct list_head freelist;
114 struct {
115 struct rb_root freetree;
116 struct wc_entry *current_free;
117 };
118 };
119 struct rb_root tree;
120
121 size_t freelist_size;
122 size_t writeback_size;
123 size_t freelist_high_watermark;
124 size_t freelist_low_watermark;
125 unsigned long max_age;
126
127 unsigned uncommitted_blocks;
128 unsigned autocommit_blocks;
129 unsigned max_writeback_jobs;
130
131 int error;
132
133 unsigned long autocommit_jiffies;
134 struct timer_list autocommit_timer;
135 struct wait_queue_head freelist_wait;
136
137 struct timer_list max_age_timer;
138
139 atomic_t bio_in_progress[2];
140 struct wait_queue_head bio_in_progress_wait[2];
141
142 struct dm_target *ti;
143 struct dm_dev *dev;
144 struct dm_dev *ssd_dev;
145 sector_t start_sector;
146 void *memory_map;
147 uint64_t memory_map_size;
148 size_t metadata_sectors;
149 size_t n_blocks;
150 uint64_t seq_count;
151 void *block_start;
152 struct wc_entry *entries;
153 unsigned block_size;
154 unsigned char block_size_bits;
155
156 bool pmem_mode:1;
157 bool writeback_fua:1;
158
159 bool overwrote_committed:1;
160 bool memory_vmapped:1;
161
162 bool high_wm_percent_set:1;
163 bool low_wm_percent_set:1;
164 bool max_writeback_jobs_set:1;
165 bool autocommit_blocks_set:1;
166 bool autocommit_time_set:1;
167 bool writeback_fua_set:1;
168 bool flush_on_suspend:1;
169 bool cleaner:1;
170
171 unsigned writeback_all;
172 struct workqueue_struct *writeback_wq;
173 struct work_struct writeback_work;
174 struct work_struct flush_work;
175
176 struct dm_io_client *dm_io;
177
178 raw_spinlock_t endio_list_lock;
179 struct list_head endio_list;
180 struct task_struct *endio_thread;
181
182 struct task_struct *flush_thread;
183 struct bio_list flush_list;
184
185 struct dm_kcopyd_client *dm_kcopyd;
186 unsigned long *dirty_bitmap;
187 unsigned dirty_bitmap_size;
188
189 struct bio_set bio_set;
190 mempool_t copy_pool;
191};
192
193#define WB_LIST_INLINE 16
194
195struct writeback_struct {
196 struct list_head endio_entry;
197 struct dm_writecache *wc;
198 struct wc_entry **wc_list;
199 unsigned wc_list_n;
200 struct wc_entry *wc_list_inline[WB_LIST_INLINE];
201 struct bio bio;
202};
203
204struct copy_struct {
205 struct list_head endio_entry;
206 struct dm_writecache *wc;
207 struct wc_entry *e;
208 unsigned n_entries;
209 int error;
210};
211
212DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
213 "A percentage of time allocated for data copying");
214
215static void wc_lock(struct dm_writecache *wc)
216{
217 mutex_lock(&wc->lock);
218}
219
220static void wc_unlock(struct dm_writecache *wc)
221{
222 mutex_unlock(&wc->lock);
223}
224
225#ifdef DM_WRITECACHE_HAS_PMEM
226static int persistent_memory_claim(struct dm_writecache *wc)
227{
228 int r;
229 loff_t s;
230 long p, da;
231 pfn_t pfn;
232 int id;
233 struct page **pages;
234 sector_t offset;
235
236 wc->memory_vmapped = false;
237
238 s = wc->memory_map_size;
239 p = s >> PAGE_SHIFT;
240 if (!p) {
241 r = -EINVAL;
242 goto err1;
243 }
244 if (p != s >> PAGE_SHIFT) {
245 r = -EOVERFLOW;
246 goto err1;
247 }
248
249 offset = get_start_sect(wc->ssd_dev->bdev);
250 if (offset & (PAGE_SIZE / 512 - 1)) {
251 r = -EINVAL;
252 goto err1;
253 }
254 offset >>= PAGE_SHIFT - 9;
255
256 id = dax_read_lock();
257
258 da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, &wc->memory_map, &pfn);
259 if (da < 0) {
260 wc->memory_map = NULL;
261 r = da;
262 goto err2;
263 }
264 if (!pfn_t_has_page(pfn)) {
265 wc->memory_map = NULL;
266 r = -EOPNOTSUPP;
267 goto err2;
268 }
269 if (da != p) {
270 long i;
271 wc->memory_map = NULL;
272 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
273 if (!pages) {
274 r = -ENOMEM;
275 goto err2;
276 }
277 i = 0;
278 do {
279 long daa;
280 daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i, p - i,
281 NULL, &pfn);
282 if (daa <= 0) {
283 r = daa ? daa : -EINVAL;
284 goto err3;
285 }
286 if (!pfn_t_has_page(pfn)) {
287 r = -EOPNOTSUPP;
288 goto err3;
289 }
290 while (daa-- && i < p) {
291 pages[i++] = pfn_t_to_page(pfn);
292 pfn.val++;
293 if (!(i & 15))
294 cond_resched();
295 }
296 } while (i < p);
297 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
298 if (!wc->memory_map) {
299 r = -ENOMEM;
300 goto err3;
301 }
302 kvfree(pages);
303 wc->memory_vmapped = true;
304 }
305
306 dax_read_unlock(id);
307
308 wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
309 wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
310
311 return 0;
312err3:
313 kvfree(pages);
314err2:
315 dax_read_unlock(id);
316err1:
317 return r;
318}
319#else
320static int persistent_memory_claim(struct dm_writecache *wc)
321{
322 BUG();
323}
324#endif
325
326static void persistent_memory_release(struct dm_writecache *wc)
327{
328 if (wc->memory_vmapped)
329 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
330}
331
332static struct page *persistent_memory_page(void *addr)
333{
334 if (is_vmalloc_addr(addr))
335 return vmalloc_to_page(addr);
336 else
337 return virt_to_page(addr);
338}
339
340static unsigned persistent_memory_page_offset(void *addr)
341{
342 return (unsigned long)addr & (PAGE_SIZE - 1);
343}
344
345static void persistent_memory_flush_cache(void *ptr, size_t size)
346{
347 if (is_vmalloc_addr(ptr))
348 flush_kernel_vmap_range(ptr, size);
349}
350
351static void persistent_memory_invalidate_cache(void *ptr, size_t size)
352{
353 if (is_vmalloc_addr(ptr))
354 invalidate_kernel_vmap_range(ptr, size);
355}
356
357static struct wc_memory_superblock *sb(struct dm_writecache *wc)
358{
359 return wc->memory_map;
360}
361
362static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
363{
364 return &sb(wc)->entries[e->index];
365}
366
367static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
368{
369 return (char *)wc->block_start + (e->index << wc->block_size_bits);
370}
371
372static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
373{
374 return wc->start_sector + wc->metadata_sectors +
375 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
376}
377
378static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
379{
380#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
381 return e->original_sector;
382#else
383 return le64_to_cpu(memory_entry(wc, e)->original_sector);
384#endif
385}
386
387static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
388{
389#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
390 return e->seq_count;
391#else
392 return le64_to_cpu(memory_entry(wc, e)->seq_count);
393#endif
394}
395
396static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
397{
398#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
399 e->seq_count = -1;
400#endif
401 pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
402}
403
404static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
405 uint64_t original_sector, uint64_t seq_count)
406{
407 struct wc_memory_entry me;
408#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
409 e->original_sector = original_sector;
410 e->seq_count = seq_count;
411#endif
412 me.original_sector = cpu_to_le64(original_sector);
413 me.seq_count = cpu_to_le64(seq_count);
414 pmem_assign(*memory_entry(wc, e), me);
415}
416
417#define writecache_error(wc, err, msg, arg...) \
418do { \
419 if (!cmpxchg(&(wc)->error, 0, err)) \
420 DMERR(msg, ##arg); \
421 wake_up(&(wc)->freelist_wait); \
422} while (0)
423
424#define writecache_has_error(wc) (unlikely(READ_ONCE((wc)->error)))
425
426static void writecache_flush_all_metadata(struct dm_writecache *wc)
427{
428 if (!WC_MODE_PMEM(wc))
429 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
430}
431
432static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
433{
434 if (!WC_MODE_PMEM(wc))
435 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
436 wc->dirty_bitmap);
437}
438
439static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
440
441struct io_notify {
442 struct dm_writecache *wc;
443 struct completion c;
444 atomic_t count;
445};
446
447static void writecache_notify_io(unsigned long error, void *context)
448{
449 struct io_notify *endio = context;
450
451 if (unlikely(error != 0))
452 writecache_error(endio->wc, -EIO, "error writing metadata");
453 BUG_ON(atomic_read(&endio->count) <= 0);
454 if (atomic_dec_and_test(&endio->count))
455 complete(&endio->c);
456}
457
458static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
459{
460 wait_event(wc->bio_in_progress_wait[direction],
461 !atomic_read(&wc->bio_in_progress[direction]));
462}
463
464static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
465{
466 struct dm_io_region region;
467 struct dm_io_request req;
468 struct io_notify endio = {
469 wc,
470 COMPLETION_INITIALIZER_ONSTACK(endio.c),
471 ATOMIC_INIT(1),
472 };
473 unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
474 unsigned i = 0;
475
476 while (1) {
477 unsigned j;
478 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
479 if (unlikely(i == bitmap_bits))
480 break;
481 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
482
483 region.bdev = wc->ssd_dev->bdev;
484 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
485 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
486
487 if (unlikely(region.sector >= wc->metadata_sectors))
488 break;
489 if (unlikely(region.sector + region.count > wc->metadata_sectors))
490 region.count = wc->metadata_sectors - region.sector;
491
492 region.sector += wc->start_sector;
493 atomic_inc(&endio.count);
494 req.bi_op = REQ_OP_WRITE;
495 req.bi_op_flags = REQ_SYNC;
496 req.mem.type = DM_IO_VMA;
497 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
498 req.client = wc->dm_io;
499 req.notify.fn = writecache_notify_io;
500 req.notify.context = &endio;
501
502 /* writing via async dm-io (implied by notify.fn above) won't return an error */
503 (void) dm_io(&req, 1, ®ion, NULL);
504 i = j;
505 }
506
507 writecache_notify_io(0, &endio);
508 wait_for_completion_io(&endio.c);
509
510 if (wait_for_ios)
511 writecache_wait_for_ios(wc, WRITE);
512
513 writecache_disk_flush(wc, wc->ssd_dev);
514
515 memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
516}
517
518static void ssd_commit_superblock(struct dm_writecache *wc)
519{
520 int r;
521 struct dm_io_region region;
522 struct dm_io_request req;
523
524 region.bdev = wc->ssd_dev->bdev;
525 region.sector = 0;
526 region.count = PAGE_SIZE;
527
528 if (unlikely(region.sector + region.count > wc->metadata_sectors))
529 region.count = wc->metadata_sectors - region.sector;
530
531 region.sector += wc->start_sector;
532
533 req.bi_op = REQ_OP_WRITE;
534 req.bi_op_flags = REQ_SYNC | REQ_FUA;
535 req.mem.type = DM_IO_VMA;
536 req.mem.ptr.vma = (char *)wc->memory_map;
537 req.client = wc->dm_io;
538 req.notify.fn = NULL;
539 req.notify.context = NULL;
540
541 r = dm_io(&req, 1, ®ion, NULL);
542 if (unlikely(r))
543 writecache_error(wc, r, "error writing superblock");
544}
545
546static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
547{
548 if (WC_MODE_PMEM(wc))
549 pmem_wmb();
550 else
551 ssd_commit_flushed(wc, wait_for_ios);
552}
553
554static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
555{
556 int r;
557 struct dm_io_region region;
558 struct dm_io_request req;
559
560 region.bdev = dev->bdev;
561 region.sector = 0;
562 region.count = 0;
563 req.bi_op = REQ_OP_WRITE;
564 req.bi_op_flags = REQ_PREFLUSH;
565 req.mem.type = DM_IO_KMEM;
566 req.mem.ptr.addr = NULL;
567 req.client = wc->dm_io;
568 req.notify.fn = NULL;
569
570 r = dm_io(&req, 1, ®ion, NULL);
571 if (unlikely(r))
572 writecache_error(wc, r, "error flushing metadata: %d", r);
573}
574
575#define WFE_RETURN_FOLLOWING 1
576#define WFE_LOWEST_SEQ 2
577
578static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
579 uint64_t block, int flags)
580{
581 struct wc_entry *e;
582 struct rb_node *node = wc->tree.rb_node;
583
584 if (unlikely(!node))
585 return NULL;
586
587 while (1) {
588 e = container_of(node, struct wc_entry, rb_node);
589 if (read_original_sector(wc, e) == block)
590 break;
591
592 node = (read_original_sector(wc, e) >= block ?
593 e->rb_node.rb_left : e->rb_node.rb_right);
594 if (unlikely(!node)) {
595 if (!(flags & WFE_RETURN_FOLLOWING))
596 return NULL;
597 if (read_original_sector(wc, e) >= block) {
598 return e;
599 } else {
600 node = rb_next(&e->rb_node);
601 if (unlikely(!node))
602 return NULL;
603 e = container_of(node, struct wc_entry, rb_node);
604 return e;
605 }
606 }
607 }
608
609 while (1) {
610 struct wc_entry *e2;
611 if (flags & WFE_LOWEST_SEQ)
612 node = rb_prev(&e->rb_node);
613 else
614 node = rb_next(&e->rb_node);
615 if (unlikely(!node))
616 return e;
617 e2 = container_of(node, struct wc_entry, rb_node);
618 if (read_original_sector(wc, e2) != block)
619 return e;
620 e = e2;
621 }
622}
623
624static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
625{
626 struct wc_entry *e;
627 struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
628
629 while (*node) {
630 e = container_of(*node, struct wc_entry, rb_node);
631 parent = &e->rb_node;
632 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
633 node = &parent->rb_left;
634 else
635 node = &parent->rb_right;
636 }
637 rb_link_node(&ins->rb_node, parent, node);
638 rb_insert_color(&ins->rb_node, &wc->tree);
639 list_add(&ins->lru, &wc->lru);
640 ins->age = jiffies;
641}
642
643static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
644{
645 list_del(&e->lru);
646 rb_erase(&e->rb_node, &wc->tree);
647}
648
649static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
650{
651 if (WC_MODE_SORT_FREELIST(wc)) {
652 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
653 if (unlikely(!*node))
654 wc->current_free = e;
655 while (*node) {
656 parent = *node;
657 if (&e->rb_node < *node)
658 node = &parent->rb_left;
659 else
660 node = &parent->rb_right;
661 }
662 rb_link_node(&e->rb_node, parent, node);
663 rb_insert_color(&e->rb_node, &wc->freetree);
664 } else {
665 list_add_tail(&e->lru, &wc->freelist);
666 }
667 wc->freelist_size++;
668}
669
670static inline void writecache_verify_watermark(struct dm_writecache *wc)
671{
672 if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
673 queue_work(wc->writeback_wq, &wc->writeback_work);
674}
675
676static void writecache_max_age_timer(struct timer_list *t)
677{
678 struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
679
680 if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
681 queue_work(wc->writeback_wq, &wc->writeback_work);
682 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
683 }
684}
685
686static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
687{
688 struct wc_entry *e;
689
690 if (WC_MODE_SORT_FREELIST(wc)) {
691 struct rb_node *next;
692 if (unlikely(!wc->current_free))
693 return NULL;
694 e = wc->current_free;
695 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
696 return NULL;
697 next = rb_next(&e->rb_node);
698 rb_erase(&e->rb_node, &wc->freetree);
699 if (unlikely(!next))
700 next = rb_first(&wc->freetree);
701 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
702 } else {
703 if (unlikely(list_empty(&wc->freelist)))
704 return NULL;
705 e = container_of(wc->freelist.next, struct wc_entry, lru);
706 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
707 return NULL;
708 list_del(&e->lru);
709 }
710 wc->freelist_size--;
711
712 writecache_verify_watermark(wc);
713
714 return e;
715}
716
717static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
718{
719 writecache_unlink(wc, e);
720 writecache_add_to_freelist(wc, e);
721 clear_seq_count(wc, e);
722 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
723 if (unlikely(waitqueue_active(&wc->freelist_wait)))
724 wake_up(&wc->freelist_wait);
725}
726
727static void writecache_wait_on_freelist(struct dm_writecache *wc)
728{
729 DEFINE_WAIT(wait);
730
731 prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
732 wc_unlock(wc);
733 io_schedule();
734 finish_wait(&wc->freelist_wait, &wait);
735 wc_lock(wc);
736}
737
738static void writecache_poison_lists(struct dm_writecache *wc)
739{
740 /*
741 * Catch incorrect access to these values while the device is suspended.
742 */
743 memset(&wc->tree, -1, sizeof wc->tree);
744 wc->lru.next = LIST_POISON1;
745 wc->lru.prev = LIST_POISON2;
746 wc->freelist.next = LIST_POISON1;
747 wc->freelist.prev = LIST_POISON2;
748}
749
750static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
751{
752 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
753 if (WC_MODE_PMEM(wc))
754 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
755}
756
757static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
758{
759 return read_seq_count(wc, e) < wc->seq_count;
760}
761
762static void writecache_flush(struct dm_writecache *wc)
763{
764 struct wc_entry *e, *e2;
765 bool need_flush_after_free;
766
767 wc->uncommitted_blocks = 0;
768 del_timer(&wc->autocommit_timer);
769
770 if (list_empty(&wc->lru))
771 return;
772
773 e = container_of(wc->lru.next, struct wc_entry, lru);
774 if (writecache_entry_is_committed(wc, e)) {
775 if (wc->overwrote_committed) {
776 writecache_wait_for_ios(wc, WRITE);
777 writecache_disk_flush(wc, wc->ssd_dev);
778 wc->overwrote_committed = false;
779 }
780 return;
781 }
782 while (1) {
783 writecache_flush_entry(wc, e);
784 if (unlikely(e->lru.next == &wc->lru))
785 break;
786 e2 = container_of(e->lru.next, struct wc_entry, lru);
787 if (writecache_entry_is_committed(wc, e2))
788 break;
789 e = e2;
790 cond_resched();
791 }
792 writecache_commit_flushed(wc, true);
793
794 wc->seq_count++;
795 pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
796 if (WC_MODE_PMEM(wc))
797 writecache_commit_flushed(wc, false);
798 else
799 ssd_commit_superblock(wc);
800
801 wc->overwrote_committed = false;
802
803 need_flush_after_free = false;
804 while (1) {
805 /* Free another committed entry with lower seq-count */
806 struct rb_node *rb_node = rb_prev(&e->rb_node);
807
808 if (rb_node) {
809 e2 = container_of(rb_node, struct wc_entry, rb_node);
810 if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
811 likely(!e2->write_in_progress)) {
812 writecache_free_entry(wc, e2);
813 need_flush_after_free = true;
814 }
815 }
816 if (unlikely(e->lru.prev == &wc->lru))
817 break;
818 e = container_of(e->lru.prev, struct wc_entry, lru);
819 cond_resched();
820 }
821
822 if (need_flush_after_free)
823 writecache_commit_flushed(wc, false);
824}
825
826static void writecache_flush_work(struct work_struct *work)
827{
828 struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
829
830 wc_lock(wc);
831 writecache_flush(wc);
832 wc_unlock(wc);
833}
834
835static void writecache_autocommit_timer(struct timer_list *t)
836{
837 struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
838 if (!writecache_has_error(wc))
839 queue_work(wc->writeback_wq, &wc->flush_work);
840}
841
842static void writecache_schedule_autocommit(struct dm_writecache *wc)
843{
844 if (!timer_pending(&wc->autocommit_timer))
845 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
846}
847
848static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
849{
850 struct wc_entry *e;
851 bool discarded_something = false;
852
853 e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
854 if (unlikely(!e))
855 return;
856
857 while (read_original_sector(wc, e) < end) {
858 struct rb_node *node = rb_next(&e->rb_node);
859
860 if (likely(!e->write_in_progress)) {
861 if (!discarded_something) {
862 if (!WC_MODE_PMEM(wc)) {
863 writecache_wait_for_ios(wc, READ);
864 writecache_wait_for_ios(wc, WRITE);
865 }
866 discarded_something = true;
867 }
868 if (!writecache_entry_is_committed(wc, e))
869 wc->uncommitted_blocks--;
870 writecache_free_entry(wc, e);
871 }
872
873 if (unlikely(!node))
874 break;
875
876 e = container_of(node, struct wc_entry, rb_node);
877 }
878
879 if (discarded_something)
880 writecache_commit_flushed(wc, false);
881}
882
883static bool writecache_wait_for_writeback(struct dm_writecache *wc)
884{
885 if (wc->writeback_size) {
886 writecache_wait_on_freelist(wc);
887 return true;
888 }
889 return false;
890}
891
892static void writecache_suspend(struct dm_target *ti)
893{
894 struct dm_writecache *wc = ti->private;
895 bool flush_on_suspend;
896
897 del_timer_sync(&wc->autocommit_timer);
898 del_timer_sync(&wc->max_age_timer);
899
900 wc_lock(wc);
901 writecache_flush(wc);
902 flush_on_suspend = wc->flush_on_suspend;
903 if (flush_on_suspend) {
904 wc->flush_on_suspend = false;
905 wc->writeback_all++;
906 queue_work(wc->writeback_wq, &wc->writeback_work);
907 }
908 wc_unlock(wc);
909
910 drain_workqueue(wc->writeback_wq);
911
912 wc_lock(wc);
913 if (flush_on_suspend)
914 wc->writeback_all--;
915 while (writecache_wait_for_writeback(wc));
916
917 if (WC_MODE_PMEM(wc))
918 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
919
920 writecache_poison_lists(wc);
921
922 wc_unlock(wc);
923}
924
925static int writecache_alloc_entries(struct dm_writecache *wc)
926{
927 size_t b;
928
929 if (wc->entries)
930 return 0;
931 wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
932 if (!wc->entries)
933 return -ENOMEM;
934 for (b = 0; b < wc->n_blocks; b++) {
935 struct wc_entry *e = &wc->entries[b];
936 e->index = b;
937 e->write_in_progress = false;
938 cond_resched();
939 }
940
941 return 0;
942}
943
944static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
945{
946 struct dm_io_region region;
947 struct dm_io_request req;
948
949 region.bdev = wc->ssd_dev->bdev;
950 region.sector = wc->start_sector;
951 region.count = n_sectors;
952 req.bi_op = REQ_OP_READ;
953 req.bi_op_flags = REQ_SYNC;
954 req.mem.type = DM_IO_VMA;
955 req.mem.ptr.vma = (char *)wc->memory_map;
956 req.client = wc->dm_io;
957 req.notify.fn = NULL;
958
959 return dm_io(&req, 1, ®ion, NULL);
960}
961
962static void writecache_resume(struct dm_target *ti)
963{
964 struct dm_writecache *wc = ti->private;
965 size_t b;
966 bool need_flush = false;
967 __le64 sb_seq_count;
968 int r;
969
970 wc_lock(wc);
971
972 if (WC_MODE_PMEM(wc)) {
973 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
974 } else {
975 r = writecache_read_metadata(wc, wc->metadata_sectors);
976 if (r) {
977 size_t sb_entries_offset;
978 writecache_error(wc, r, "unable to read metadata: %d", r);
979 sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
980 memset((char *)wc->memory_map + sb_entries_offset, -1,
981 (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
982 }
983 }
984
985 wc->tree = RB_ROOT;
986 INIT_LIST_HEAD(&wc->lru);
987 if (WC_MODE_SORT_FREELIST(wc)) {
988 wc->freetree = RB_ROOT;
989 wc->current_free = NULL;
990 } else {
991 INIT_LIST_HEAD(&wc->freelist);
992 }
993 wc->freelist_size = 0;
994
995 r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
996 if (r) {
997 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
998 sb_seq_count = cpu_to_le64(0);
999 }
1000 wc->seq_count = le64_to_cpu(sb_seq_count);
1001
1002#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1003 for (b = 0; b < wc->n_blocks; b++) {
1004 struct wc_entry *e = &wc->entries[b];
1005 struct wc_memory_entry wme;
1006 if (writecache_has_error(wc)) {
1007 e->original_sector = -1;
1008 e->seq_count = -1;
1009 continue;
1010 }
1011 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
1012 if (r) {
1013 writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1014 (unsigned long)b, r);
1015 e->original_sector = -1;
1016 e->seq_count = -1;
1017 } else {
1018 e->original_sector = le64_to_cpu(wme.original_sector);
1019 e->seq_count = le64_to_cpu(wme.seq_count);
1020 }
1021 cond_resched();
1022 }
1023#endif
1024 for (b = 0; b < wc->n_blocks; b++) {
1025 struct wc_entry *e = &wc->entries[b];
1026 if (!writecache_entry_is_committed(wc, e)) {
1027 if (read_seq_count(wc, e) != -1) {
1028erase_this:
1029 clear_seq_count(wc, e);
1030 need_flush = true;
1031 }
1032 writecache_add_to_freelist(wc, e);
1033 } else {
1034 struct wc_entry *old;
1035
1036 old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1037 if (!old) {
1038 writecache_insert_entry(wc, e);
1039 } else {
1040 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1041 writecache_error(wc, -EINVAL,
1042 "two identical entries, position %llu, sector %llu, sequence %llu",
1043 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1044 (unsigned long long)read_seq_count(wc, e));
1045 }
1046 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1047 goto erase_this;
1048 } else {
1049 writecache_free_entry(wc, old);
1050 writecache_insert_entry(wc, e);
1051 need_flush = true;
1052 }
1053 }
1054 }
1055 cond_resched();
1056 }
1057
1058 if (need_flush) {
1059 writecache_flush_all_metadata(wc);
1060 writecache_commit_flushed(wc, false);
1061 }
1062
1063 writecache_verify_watermark(wc);
1064
1065 if (wc->max_age != MAX_AGE_UNSPECIFIED)
1066 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1067
1068 wc_unlock(wc);
1069}
1070
1071static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1072{
1073 if (argc != 1)
1074 return -EINVAL;
1075
1076 wc_lock(wc);
1077 if (dm_suspended(wc->ti)) {
1078 wc_unlock(wc);
1079 return -EBUSY;
1080 }
1081 if (writecache_has_error(wc)) {
1082 wc_unlock(wc);
1083 return -EIO;
1084 }
1085
1086 writecache_flush(wc);
1087 wc->writeback_all++;
1088 queue_work(wc->writeback_wq, &wc->writeback_work);
1089 wc_unlock(wc);
1090
1091 flush_workqueue(wc->writeback_wq);
1092
1093 wc_lock(wc);
1094 wc->writeback_all--;
1095 if (writecache_has_error(wc)) {
1096 wc_unlock(wc);
1097 return -EIO;
1098 }
1099 wc_unlock(wc);
1100
1101 return 0;
1102}
1103
1104static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1105{
1106 if (argc != 1)
1107 return -EINVAL;
1108
1109 wc_lock(wc);
1110 wc->flush_on_suspend = true;
1111 wc_unlock(wc);
1112
1113 return 0;
1114}
1115
1116static void activate_cleaner(struct dm_writecache *wc)
1117{
1118 wc->flush_on_suspend = true;
1119 wc->cleaner = true;
1120 wc->freelist_high_watermark = wc->n_blocks;
1121 wc->freelist_low_watermark = wc->n_blocks;
1122}
1123
1124static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1125{
1126 if (argc != 1)
1127 return -EINVAL;
1128
1129 wc_lock(wc);
1130 activate_cleaner(wc);
1131 if (!dm_suspended(wc->ti))
1132 writecache_verify_watermark(wc);
1133 wc_unlock(wc);
1134
1135 return 0;
1136}
1137
1138static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1139 char *result, unsigned maxlen)
1140{
1141 int r = -EINVAL;
1142 struct dm_writecache *wc = ti->private;
1143
1144 if (!strcasecmp(argv[0], "flush"))
1145 r = process_flush_mesg(argc, argv, wc);
1146 else if (!strcasecmp(argv[0], "flush_on_suspend"))
1147 r = process_flush_on_suspend_mesg(argc, argv, wc);
1148 else if (!strcasecmp(argv[0], "cleaner"))
1149 r = process_cleaner_mesg(argc, argv, wc);
1150 else
1151 DMERR("unrecognised message received: %s", argv[0]);
1152
1153 return r;
1154}
1155
1156static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1157{
1158 /*
1159 * clflushopt performs better with block size 1024, 2048, 4096
1160 * non-temporal stores perform better with block size 512
1161 *
1162 * block size 512 1024 2048 4096
1163 * movnti 496 MB/s 642 MB/s 725 MB/s 744 MB/s
1164 * clflushopt 373 MB/s 688 MB/s 1.1 GB/s 1.2 GB/s
1165 *
1166 * We see that movnti performs better for 512-byte blocks, and
1167 * clflushopt performs better for 1024-byte and larger blocks. So, we
1168 * prefer clflushopt for sizes >= 768.
1169 *
1170 * NOTE: this happens to be the case now (with dm-writecache's single
1171 * threaded model) but re-evaluate this once memcpy_flushcache() is
1172 * enabled to use movdir64b which might invalidate this performance
1173 * advantage seen with cache-allocating-writes plus flushing.
1174 */
1175#ifdef CONFIG_X86
1176 if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1177 likely(boot_cpu_data.x86_clflush_size == 64) &&
1178 likely(size >= 768)) {
1179 do {
1180 memcpy((void *)dest, (void *)source, 64);
1181 clflushopt((void *)dest);
1182 dest += 64;
1183 source += 64;
1184 size -= 64;
1185 } while (size >= 64);
1186 return;
1187 }
1188#endif
1189 memcpy_flushcache(dest, source, size);
1190}
1191
1192static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1193{
1194 void *buf;
1195 unsigned long flags;
1196 unsigned size;
1197 int rw = bio_data_dir(bio);
1198 unsigned remaining_size = wc->block_size;
1199
1200 do {
1201 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1202 buf = bvec_kmap_irq(&bv, &flags);
1203 size = bv.bv_len;
1204 if (unlikely(size > remaining_size))
1205 size = remaining_size;
1206
1207 if (rw == READ) {
1208 int r;
1209 r = memcpy_mcsafe(buf, data, size);
1210 flush_dcache_page(bio_page(bio));
1211 if (unlikely(r)) {
1212 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1213 bio->bi_status = BLK_STS_IOERR;
1214 }
1215 } else {
1216 flush_dcache_page(bio_page(bio));
1217 memcpy_flushcache_optimized(data, buf, size);
1218 }
1219
1220 bvec_kunmap_irq(buf, &flags);
1221
1222 data = (char *)data + size;
1223 remaining_size -= size;
1224 bio_advance(bio, size);
1225 } while (unlikely(remaining_size));
1226}
1227
1228static int writecache_flush_thread(void *data)
1229{
1230 struct dm_writecache *wc = data;
1231
1232 while (1) {
1233 struct bio *bio;
1234
1235 wc_lock(wc);
1236 bio = bio_list_pop(&wc->flush_list);
1237 if (!bio) {
1238 set_current_state(TASK_INTERRUPTIBLE);
1239 wc_unlock(wc);
1240
1241 if (unlikely(kthread_should_stop())) {
1242 set_current_state(TASK_RUNNING);
1243 break;
1244 }
1245
1246 schedule();
1247 continue;
1248 }
1249
1250 if (bio_op(bio) == REQ_OP_DISCARD) {
1251 writecache_discard(wc, bio->bi_iter.bi_sector,
1252 bio_end_sector(bio));
1253 wc_unlock(wc);
1254 bio_set_dev(bio, wc->dev->bdev);
1255 submit_bio_noacct(bio);
1256 } else {
1257 writecache_flush(wc);
1258 wc_unlock(wc);
1259 if (writecache_has_error(wc))
1260 bio->bi_status = BLK_STS_IOERR;
1261 bio_endio(bio);
1262 }
1263 }
1264
1265 return 0;
1266}
1267
1268static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1269{
1270 if (bio_list_empty(&wc->flush_list))
1271 wake_up_process(wc->flush_thread);
1272 bio_list_add(&wc->flush_list, bio);
1273}
1274
1275static int writecache_map(struct dm_target *ti, struct bio *bio)
1276{
1277 struct wc_entry *e;
1278 struct dm_writecache *wc = ti->private;
1279
1280 bio->bi_private = NULL;
1281
1282 wc_lock(wc);
1283
1284 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1285 if (writecache_has_error(wc))
1286 goto unlock_error;
1287 if (WC_MODE_PMEM(wc)) {
1288 writecache_flush(wc);
1289 if (writecache_has_error(wc))
1290 goto unlock_error;
1291 goto unlock_submit;
1292 } else {
1293 writecache_offload_bio(wc, bio);
1294 goto unlock_return;
1295 }
1296 }
1297
1298 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1299
1300 if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1301 (wc->block_size / 512 - 1)) != 0)) {
1302 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1303 (unsigned long long)bio->bi_iter.bi_sector,
1304 bio->bi_iter.bi_size, wc->block_size);
1305 goto unlock_error;
1306 }
1307
1308 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1309 if (writecache_has_error(wc))
1310 goto unlock_error;
1311 if (WC_MODE_PMEM(wc)) {
1312 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1313 goto unlock_remap_origin;
1314 } else {
1315 writecache_offload_bio(wc, bio);
1316 goto unlock_return;
1317 }
1318 }
1319
1320 if (bio_data_dir(bio) == READ) {
1321read_next_block:
1322 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1323 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1324 if (WC_MODE_PMEM(wc)) {
1325 bio_copy_block(wc, bio, memory_data(wc, e));
1326 if (bio->bi_iter.bi_size)
1327 goto read_next_block;
1328 goto unlock_submit;
1329 } else {
1330 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1331 bio_set_dev(bio, wc->ssd_dev->bdev);
1332 bio->bi_iter.bi_sector = cache_sector(wc, e);
1333 if (!writecache_entry_is_committed(wc, e))
1334 writecache_wait_for_ios(wc, WRITE);
1335 goto unlock_remap;
1336 }
1337 } else {
1338 if (e) {
1339 sector_t next_boundary =
1340 read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1341 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1342 dm_accept_partial_bio(bio, next_boundary);
1343 }
1344 }
1345 goto unlock_remap_origin;
1346 }
1347 } else {
1348 do {
1349 bool found_entry = false;
1350 if (writecache_has_error(wc))
1351 goto unlock_error;
1352 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1353 if (e) {
1354 if (!writecache_entry_is_committed(wc, e))
1355 goto bio_copy;
1356 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1357 wc->overwrote_committed = true;
1358 goto bio_copy;
1359 }
1360 found_entry = true;
1361 } else {
1362 if (unlikely(wc->cleaner))
1363 goto direct_write;
1364 }
1365 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1366 if (unlikely(!e)) {
1367 if (!found_entry) {
1368direct_write:
1369 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1370 if (e) {
1371 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1372 BUG_ON(!next_boundary);
1373 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1374 dm_accept_partial_bio(bio, next_boundary);
1375 }
1376 }
1377 goto unlock_remap_origin;
1378 }
1379 writecache_wait_on_freelist(wc);
1380 continue;
1381 }
1382 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1383 writecache_insert_entry(wc, e);
1384 wc->uncommitted_blocks++;
1385bio_copy:
1386 if (WC_MODE_PMEM(wc)) {
1387 bio_copy_block(wc, bio, memory_data(wc, e));
1388 } else {
1389 unsigned bio_size = wc->block_size;
1390 sector_t start_cache_sec = cache_sector(wc, e);
1391 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1392
1393 while (bio_size < bio->bi_iter.bi_size) {
1394 struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1395 if (!f)
1396 break;
1397 write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1398 (bio_size >> SECTOR_SHIFT), wc->seq_count);
1399 writecache_insert_entry(wc, f);
1400 wc->uncommitted_blocks++;
1401 bio_size += wc->block_size;
1402 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1403 }
1404
1405 bio_set_dev(bio, wc->ssd_dev->bdev);
1406 bio->bi_iter.bi_sector = start_cache_sec;
1407 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1408
1409 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1410 wc->uncommitted_blocks = 0;
1411 queue_work(wc->writeback_wq, &wc->flush_work);
1412 } else {
1413 writecache_schedule_autocommit(wc);
1414 }
1415 goto unlock_remap;
1416 }
1417 } while (bio->bi_iter.bi_size);
1418
1419 if (unlikely(bio->bi_opf & REQ_FUA ||
1420 wc->uncommitted_blocks >= wc->autocommit_blocks))
1421 writecache_flush(wc);
1422 else
1423 writecache_schedule_autocommit(wc);
1424 goto unlock_submit;
1425 }
1426
1427unlock_remap_origin:
1428 bio_set_dev(bio, wc->dev->bdev);
1429 wc_unlock(wc);
1430 return DM_MAPIO_REMAPPED;
1431
1432unlock_remap:
1433 /* make sure that writecache_end_io decrements bio_in_progress: */
1434 bio->bi_private = (void *)1;
1435 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1436 wc_unlock(wc);
1437 return DM_MAPIO_REMAPPED;
1438
1439unlock_submit:
1440 wc_unlock(wc);
1441 bio_endio(bio);
1442 return DM_MAPIO_SUBMITTED;
1443
1444unlock_return:
1445 wc_unlock(wc);
1446 return DM_MAPIO_SUBMITTED;
1447
1448unlock_error:
1449 wc_unlock(wc);
1450 bio_io_error(bio);
1451 return DM_MAPIO_SUBMITTED;
1452}
1453
1454static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1455{
1456 struct dm_writecache *wc = ti->private;
1457
1458 if (bio->bi_private != NULL) {
1459 int dir = bio_data_dir(bio);
1460 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1461 if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1462 wake_up(&wc->bio_in_progress_wait[dir]);
1463 }
1464 return 0;
1465}
1466
1467static int writecache_iterate_devices(struct dm_target *ti,
1468 iterate_devices_callout_fn fn, void *data)
1469{
1470 struct dm_writecache *wc = ti->private;
1471
1472 return fn(ti, wc->dev, 0, ti->len, data);
1473}
1474
1475static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1476{
1477 struct dm_writecache *wc = ti->private;
1478
1479 if (limits->logical_block_size < wc->block_size)
1480 limits->logical_block_size = wc->block_size;
1481
1482 if (limits->physical_block_size < wc->block_size)
1483 limits->physical_block_size = wc->block_size;
1484
1485 if (limits->io_min < wc->block_size)
1486 limits->io_min = wc->block_size;
1487}
1488
1489
1490static void writecache_writeback_endio(struct bio *bio)
1491{
1492 struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1493 struct dm_writecache *wc = wb->wc;
1494 unsigned long flags;
1495
1496 raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1497 if (unlikely(list_empty(&wc->endio_list)))
1498 wake_up_process(wc->endio_thread);
1499 list_add_tail(&wb->endio_entry, &wc->endio_list);
1500 raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1501}
1502
1503static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1504{
1505 struct copy_struct *c = ptr;
1506 struct dm_writecache *wc = c->wc;
1507
1508 c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1509
1510 raw_spin_lock_irq(&wc->endio_list_lock);
1511 if (unlikely(list_empty(&wc->endio_list)))
1512 wake_up_process(wc->endio_thread);
1513 list_add_tail(&c->endio_entry, &wc->endio_list);
1514 raw_spin_unlock_irq(&wc->endio_list_lock);
1515}
1516
1517static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1518{
1519 unsigned i;
1520 struct writeback_struct *wb;
1521 struct wc_entry *e;
1522 unsigned long n_walked = 0;
1523
1524 do {
1525 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1526 list_del(&wb->endio_entry);
1527
1528 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1529 writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1530 "write error %d", wb->bio.bi_status);
1531 i = 0;
1532 do {
1533 e = wb->wc_list[i];
1534 BUG_ON(!e->write_in_progress);
1535 e->write_in_progress = false;
1536 INIT_LIST_HEAD(&e->lru);
1537 if (!writecache_has_error(wc))
1538 writecache_free_entry(wc, e);
1539 BUG_ON(!wc->writeback_size);
1540 wc->writeback_size--;
1541 n_walked++;
1542 if (unlikely(n_walked >= ENDIO_LATENCY)) {
1543 writecache_commit_flushed(wc, false);
1544 wc_unlock(wc);
1545 wc_lock(wc);
1546 n_walked = 0;
1547 }
1548 } while (++i < wb->wc_list_n);
1549
1550 if (wb->wc_list != wb->wc_list_inline)
1551 kfree(wb->wc_list);
1552 bio_put(&wb->bio);
1553 } while (!list_empty(list));
1554}
1555
1556static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1557{
1558 struct copy_struct *c;
1559 struct wc_entry *e;
1560
1561 do {
1562 c = list_entry(list->next, struct copy_struct, endio_entry);
1563 list_del(&c->endio_entry);
1564
1565 if (unlikely(c->error))
1566 writecache_error(wc, c->error, "copy error");
1567
1568 e = c->e;
1569 do {
1570 BUG_ON(!e->write_in_progress);
1571 e->write_in_progress = false;
1572 INIT_LIST_HEAD(&e->lru);
1573 if (!writecache_has_error(wc))
1574 writecache_free_entry(wc, e);
1575
1576 BUG_ON(!wc->writeback_size);
1577 wc->writeback_size--;
1578 e++;
1579 } while (--c->n_entries);
1580 mempool_free(c, &wc->copy_pool);
1581 } while (!list_empty(list));
1582}
1583
1584static int writecache_endio_thread(void *data)
1585{
1586 struct dm_writecache *wc = data;
1587
1588 while (1) {
1589 struct list_head list;
1590
1591 raw_spin_lock_irq(&wc->endio_list_lock);
1592 if (!list_empty(&wc->endio_list))
1593 goto pop_from_list;
1594 set_current_state(TASK_INTERRUPTIBLE);
1595 raw_spin_unlock_irq(&wc->endio_list_lock);
1596
1597 if (unlikely(kthread_should_stop())) {
1598 set_current_state(TASK_RUNNING);
1599 break;
1600 }
1601
1602 schedule();
1603
1604 continue;
1605
1606pop_from_list:
1607 list = wc->endio_list;
1608 list.next->prev = list.prev->next = &list;
1609 INIT_LIST_HEAD(&wc->endio_list);
1610 raw_spin_unlock_irq(&wc->endio_list_lock);
1611
1612 if (!WC_MODE_FUA(wc))
1613 writecache_disk_flush(wc, wc->dev);
1614
1615 wc_lock(wc);
1616
1617 if (WC_MODE_PMEM(wc)) {
1618 __writecache_endio_pmem(wc, &list);
1619 } else {
1620 __writecache_endio_ssd(wc, &list);
1621 writecache_wait_for_ios(wc, READ);
1622 }
1623
1624 writecache_commit_flushed(wc, false);
1625
1626 wc_unlock(wc);
1627 }
1628
1629 return 0;
1630}
1631
1632static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1633{
1634 struct dm_writecache *wc = wb->wc;
1635 unsigned block_size = wc->block_size;
1636 void *address = memory_data(wc, e);
1637
1638 persistent_memory_flush_cache(address, block_size);
1639 return bio_add_page(&wb->bio, persistent_memory_page(address),
1640 block_size, persistent_memory_page_offset(address)) != 0;
1641}
1642
1643struct writeback_list {
1644 struct list_head list;
1645 size_t size;
1646};
1647
1648static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1649{
1650 if (unlikely(wc->max_writeback_jobs)) {
1651 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1652 wc_lock(wc);
1653 while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1654 writecache_wait_on_freelist(wc);
1655 wc_unlock(wc);
1656 }
1657 }
1658 cond_resched();
1659}
1660
1661static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1662{
1663 struct wc_entry *e, *f;
1664 struct bio *bio;
1665 struct writeback_struct *wb;
1666 unsigned max_pages;
1667
1668 while (wbl->size) {
1669 wbl->size--;
1670 e = container_of(wbl->list.prev, struct wc_entry, lru);
1671 list_del(&e->lru);
1672
1673 max_pages = e->wc_list_contiguous;
1674
1675 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1676 wb = container_of(bio, struct writeback_struct, bio);
1677 wb->wc = wc;
1678 bio->bi_end_io = writecache_writeback_endio;
1679 bio_set_dev(bio, wc->dev->bdev);
1680 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1681 if (max_pages <= WB_LIST_INLINE ||
1682 unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1683 GFP_NOIO | __GFP_NORETRY |
1684 __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1685 wb->wc_list = wb->wc_list_inline;
1686 max_pages = WB_LIST_INLINE;
1687 }
1688
1689 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1690
1691 wb->wc_list[0] = e;
1692 wb->wc_list_n = 1;
1693
1694 while (wbl->size && wb->wc_list_n < max_pages) {
1695 f = container_of(wbl->list.prev, struct wc_entry, lru);
1696 if (read_original_sector(wc, f) !=
1697 read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1698 break;
1699 if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1700 break;
1701 wbl->size--;
1702 list_del(&f->lru);
1703 wb->wc_list[wb->wc_list_n++] = f;
1704 e = f;
1705 }
1706 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1707 if (writecache_has_error(wc)) {
1708 bio->bi_status = BLK_STS_IOERR;
1709 bio_endio(bio);
1710 } else {
1711 submit_bio(bio);
1712 }
1713
1714 __writeback_throttle(wc, wbl);
1715 }
1716}
1717
1718static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1719{
1720 struct wc_entry *e, *f;
1721 struct dm_io_region from, to;
1722 struct copy_struct *c;
1723
1724 while (wbl->size) {
1725 unsigned n_sectors;
1726
1727 wbl->size--;
1728 e = container_of(wbl->list.prev, struct wc_entry, lru);
1729 list_del(&e->lru);
1730
1731 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1732
1733 from.bdev = wc->ssd_dev->bdev;
1734 from.sector = cache_sector(wc, e);
1735 from.count = n_sectors;
1736 to.bdev = wc->dev->bdev;
1737 to.sector = read_original_sector(wc, e);
1738 to.count = n_sectors;
1739
1740 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1741 c->wc = wc;
1742 c->e = e;
1743 c->n_entries = e->wc_list_contiguous;
1744
1745 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1746 wbl->size--;
1747 f = container_of(wbl->list.prev, struct wc_entry, lru);
1748 BUG_ON(f != e + 1);
1749 list_del(&f->lru);
1750 e = f;
1751 }
1752
1753 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1754
1755 __writeback_throttle(wc, wbl);
1756 }
1757}
1758
1759static void writecache_writeback(struct work_struct *work)
1760{
1761 struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1762 struct blk_plug plug;
1763 struct wc_entry *f, *g, *e = NULL;
1764 struct rb_node *node, *next_node;
1765 struct list_head skipped;
1766 struct writeback_list wbl;
1767 unsigned long n_walked;
1768
1769 wc_lock(wc);
1770restart:
1771 if (writecache_has_error(wc)) {
1772 wc_unlock(wc);
1773 return;
1774 }
1775
1776 if (unlikely(wc->writeback_all)) {
1777 if (writecache_wait_for_writeback(wc))
1778 goto restart;
1779 }
1780
1781 if (wc->overwrote_committed) {
1782 writecache_wait_for_ios(wc, WRITE);
1783 }
1784
1785 n_walked = 0;
1786 INIT_LIST_HEAD(&skipped);
1787 INIT_LIST_HEAD(&wbl.list);
1788 wbl.size = 0;
1789 while (!list_empty(&wc->lru) &&
1790 (wc->writeback_all ||
1791 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1792 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1793 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1794
1795 n_walked++;
1796 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1797 likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1798 queue_work(wc->writeback_wq, &wc->writeback_work);
1799 break;
1800 }
1801
1802 if (unlikely(wc->writeback_all)) {
1803 if (unlikely(!e)) {
1804 writecache_flush(wc);
1805 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1806 } else
1807 e = g;
1808 } else
1809 e = container_of(wc->lru.prev, struct wc_entry, lru);
1810 BUG_ON(e->write_in_progress);
1811 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1812 writecache_flush(wc);
1813 }
1814 node = rb_prev(&e->rb_node);
1815 if (node) {
1816 f = container_of(node, struct wc_entry, rb_node);
1817 if (unlikely(read_original_sector(wc, f) ==
1818 read_original_sector(wc, e))) {
1819 BUG_ON(!f->write_in_progress);
1820 list_del(&e->lru);
1821 list_add(&e->lru, &skipped);
1822 cond_resched();
1823 continue;
1824 }
1825 }
1826 wc->writeback_size++;
1827 list_del(&e->lru);
1828 list_add(&e->lru, &wbl.list);
1829 wbl.size++;
1830 e->write_in_progress = true;
1831 e->wc_list_contiguous = 1;
1832
1833 f = e;
1834
1835 while (1) {
1836 next_node = rb_next(&f->rb_node);
1837 if (unlikely(!next_node))
1838 break;
1839 g = container_of(next_node, struct wc_entry, rb_node);
1840 if (unlikely(read_original_sector(wc, g) ==
1841 read_original_sector(wc, f))) {
1842 f = g;
1843 continue;
1844 }
1845 if (read_original_sector(wc, g) !=
1846 read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1847 break;
1848 if (unlikely(g->write_in_progress))
1849 break;
1850 if (unlikely(!writecache_entry_is_committed(wc, g)))
1851 break;
1852
1853 if (!WC_MODE_PMEM(wc)) {
1854 if (g != f + 1)
1855 break;
1856 }
1857
1858 n_walked++;
1859 //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1860 // break;
1861
1862 wc->writeback_size++;
1863 list_del(&g->lru);
1864 list_add(&g->lru, &wbl.list);
1865 wbl.size++;
1866 g->write_in_progress = true;
1867 g->wc_list_contiguous = BIO_MAX_PAGES;
1868 f = g;
1869 e->wc_list_contiguous++;
1870 if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1871 if (unlikely(wc->writeback_all)) {
1872 next_node = rb_next(&f->rb_node);
1873 if (likely(next_node))
1874 g = container_of(next_node, struct wc_entry, rb_node);
1875 }
1876 break;
1877 }
1878 }
1879 cond_resched();
1880 }
1881
1882 if (!list_empty(&skipped)) {
1883 list_splice_tail(&skipped, &wc->lru);
1884 /*
1885 * If we didn't do any progress, we must wait until some
1886 * writeback finishes to avoid burning CPU in a loop
1887 */
1888 if (unlikely(!wbl.size))
1889 writecache_wait_for_writeback(wc);
1890 }
1891
1892 wc_unlock(wc);
1893
1894 blk_start_plug(&plug);
1895
1896 if (WC_MODE_PMEM(wc))
1897 __writecache_writeback_pmem(wc, &wbl);
1898 else
1899 __writecache_writeback_ssd(wc, &wbl);
1900
1901 blk_finish_plug(&plug);
1902
1903 if (unlikely(wc->writeback_all)) {
1904 wc_lock(wc);
1905 while (writecache_wait_for_writeback(wc));
1906 wc_unlock(wc);
1907 }
1908}
1909
1910static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1911 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1912{
1913 uint64_t n_blocks, offset;
1914 struct wc_entry e;
1915
1916 n_blocks = device_size;
1917 do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1918
1919 while (1) {
1920 if (!n_blocks)
1921 return -ENOSPC;
1922 /* Verify the following entries[n_blocks] won't overflow */
1923 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1924 sizeof(struct wc_memory_entry)))
1925 return -EFBIG;
1926 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1927 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1928 if (offset + n_blocks * block_size <= device_size)
1929 break;
1930 n_blocks--;
1931 }
1932
1933 /* check if the bit field overflows */
1934 e.index = n_blocks;
1935 if (e.index != n_blocks)
1936 return -EFBIG;
1937
1938 if (n_blocks_p)
1939 *n_blocks_p = n_blocks;
1940 if (n_metadata_blocks_p)
1941 *n_metadata_blocks_p = offset >> __ffs(block_size);
1942 return 0;
1943}
1944
1945static int init_memory(struct dm_writecache *wc)
1946{
1947 size_t b;
1948 int r;
1949
1950 r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1951 if (r)
1952 return r;
1953
1954 r = writecache_alloc_entries(wc);
1955 if (r)
1956 return r;
1957
1958 for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1959 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1960 pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1961 pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1962 pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1963 pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1964
1965 for (b = 0; b < wc->n_blocks; b++) {
1966 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1967 cond_resched();
1968 }
1969
1970 writecache_flush_all_metadata(wc);
1971 writecache_commit_flushed(wc, false);
1972 pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1973 writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1974 writecache_commit_flushed(wc, false);
1975
1976 return 0;
1977}
1978
1979static void writecache_dtr(struct dm_target *ti)
1980{
1981 struct dm_writecache *wc = ti->private;
1982
1983 if (!wc)
1984 return;
1985
1986 if (wc->endio_thread)
1987 kthread_stop(wc->endio_thread);
1988
1989 if (wc->flush_thread)
1990 kthread_stop(wc->flush_thread);
1991
1992 bioset_exit(&wc->bio_set);
1993
1994 mempool_exit(&wc->copy_pool);
1995
1996 if (wc->writeback_wq)
1997 destroy_workqueue(wc->writeback_wq);
1998
1999 if (wc->dev)
2000 dm_put_device(ti, wc->dev);
2001
2002 if (wc->ssd_dev)
2003 dm_put_device(ti, wc->ssd_dev);
2004
2005 if (wc->entries)
2006 vfree(wc->entries);
2007
2008 if (wc->memory_map) {
2009 if (WC_MODE_PMEM(wc))
2010 persistent_memory_release(wc);
2011 else
2012 vfree(wc->memory_map);
2013 }
2014
2015 if (wc->dm_kcopyd)
2016 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2017
2018 if (wc->dm_io)
2019 dm_io_client_destroy(wc->dm_io);
2020
2021 if (wc->dirty_bitmap)
2022 vfree(wc->dirty_bitmap);
2023
2024 kfree(wc);
2025}
2026
2027static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2028{
2029 struct dm_writecache *wc;
2030 struct dm_arg_set as;
2031 const char *string;
2032 unsigned opt_params;
2033 size_t offset, data_size;
2034 int i, r;
2035 char dummy;
2036 int high_wm_percent = HIGH_WATERMARK;
2037 int low_wm_percent = LOW_WATERMARK;
2038 uint64_t x;
2039 struct wc_memory_superblock s;
2040
2041 static struct dm_arg _args[] = {
2042 {0, 10, "Invalid number of feature args"},
2043 };
2044
2045 as.argc = argc;
2046 as.argv = argv;
2047
2048 wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2049 if (!wc) {
2050 ti->error = "Cannot allocate writecache structure";
2051 r = -ENOMEM;
2052 goto bad;
2053 }
2054 ti->private = wc;
2055 wc->ti = ti;
2056
2057 mutex_init(&wc->lock);
2058 wc->max_age = MAX_AGE_UNSPECIFIED;
2059 writecache_poison_lists(wc);
2060 init_waitqueue_head(&wc->freelist_wait);
2061 timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2062 timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2063
2064 for (i = 0; i < 2; i++) {
2065 atomic_set(&wc->bio_in_progress[i], 0);
2066 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2067 }
2068
2069 wc->dm_io = dm_io_client_create();
2070 if (IS_ERR(wc->dm_io)) {
2071 r = PTR_ERR(wc->dm_io);
2072 ti->error = "Unable to allocate dm-io client";
2073 wc->dm_io = NULL;
2074 goto bad;
2075 }
2076
2077 wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2078 if (!wc->writeback_wq) {
2079 r = -ENOMEM;
2080 ti->error = "Could not allocate writeback workqueue";
2081 goto bad;
2082 }
2083 INIT_WORK(&wc->writeback_work, writecache_writeback);
2084 INIT_WORK(&wc->flush_work, writecache_flush_work);
2085
2086 raw_spin_lock_init(&wc->endio_list_lock);
2087 INIT_LIST_HEAD(&wc->endio_list);
2088 wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2089 if (IS_ERR(wc->endio_thread)) {
2090 r = PTR_ERR(wc->endio_thread);
2091 wc->endio_thread = NULL;
2092 ti->error = "Couldn't spawn endio thread";
2093 goto bad;
2094 }
2095 wake_up_process(wc->endio_thread);
2096
2097 /*
2098 * Parse the mode (pmem or ssd)
2099 */
2100 string = dm_shift_arg(&as);
2101 if (!string)
2102 goto bad_arguments;
2103
2104 if (!strcasecmp(string, "s")) {
2105 wc->pmem_mode = false;
2106 } else if (!strcasecmp(string, "p")) {
2107#ifdef DM_WRITECACHE_HAS_PMEM
2108 wc->pmem_mode = true;
2109 wc->writeback_fua = true;
2110#else
2111 /*
2112 * If the architecture doesn't support persistent memory or
2113 * the kernel doesn't support any DAX drivers, this driver can
2114 * only be used in SSD-only mode.
2115 */
2116 r = -EOPNOTSUPP;
2117 ti->error = "Persistent memory or DAX not supported on this system";
2118 goto bad;
2119#endif
2120 } else {
2121 goto bad_arguments;
2122 }
2123
2124 if (WC_MODE_PMEM(wc)) {
2125 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2126 offsetof(struct writeback_struct, bio),
2127 BIOSET_NEED_BVECS);
2128 if (r) {
2129 ti->error = "Could not allocate bio set";
2130 goto bad;
2131 }
2132 } else {
2133 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2134 if (r) {
2135 ti->error = "Could not allocate mempool";
2136 goto bad;
2137 }
2138 }
2139
2140 /*
2141 * Parse the origin data device
2142 */
2143 string = dm_shift_arg(&as);
2144 if (!string)
2145 goto bad_arguments;
2146 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2147 if (r) {
2148 ti->error = "Origin data device lookup failed";
2149 goto bad;
2150 }
2151
2152 /*
2153 * Parse cache data device (be it pmem or ssd)
2154 */
2155 string = dm_shift_arg(&as);
2156 if (!string)
2157 goto bad_arguments;
2158
2159 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2160 if (r) {
2161 ti->error = "Cache data device lookup failed";
2162 goto bad;
2163 }
2164 wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2165
2166 /*
2167 * Parse the cache block size
2168 */
2169 string = dm_shift_arg(&as);
2170 if (!string)
2171 goto bad_arguments;
2172 if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2173 wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2174 (wc->block_size & (wc->block_size - 1))) {
2175 r = -EINVAL;
2176 ti->error = "Invalid block size";
2177 goto bad;
2178 }
2179 if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2180 wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2181 r = -EINVAL;
2182 ti->error = "Block size is smaller than device logical block size";
2183 goto bad;
2184 }
2185 wc->block_size_bits = __ffs(wc->block_size);
2186
2187 wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2188 wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2189 wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2190
2191 /*
2192 * Parse optional arguments
2193 */
2194 r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2195 if (r)
2196 goto bad;
2197
2198 while (opt_params) {
2199 string = dm_shift_arg(&as), opt_params--;
2200 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2201 unsigned long long start_sector;
2202 string = dm_shift_arg(&as), opt_params--;
2203 if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2204 goto invalid_optional;
2205 wc->start_sector = start_sector;
2206 if (wc->start_sector != start_sector ||
2207 wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2208 goto invalid_optional;
2209 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2210 string = dm_shift_arg(&as), opt_params--;
2211 if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2212 goto invalid_optional;
2213 if (high_wm_percent < 0 || high_wm_percent > 100)
2214 goto invalid_optional;
2215 wc->high_wm_percent_set = true;
2216 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2217 string = dm_shift_arg(&as), opt_params--;
2218 if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2219 goto invalid_optional;
2220 if (low_wm_percent < 0 || low_wm_percent > 100)
2221 goto invalid_optional;
2222 wc->low_wm_percent_set = true;
2223 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2224 string = dm_shift_arg(&as), opt_params--;
2225 if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2226 goto invalid_optional;
2227 wc->max_writeback_jobs_set = true;
2228 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2229 string = dm_shift_arg(&as), opt_params--;
2230 if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2231 goto invalid_optional;
2232 wc->autocommit_blocks_set = true;
2233 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2234 unsigned autocommit_msecs;
2235 string = dm_shift_arg(&as), opt_params--;
2236 if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2237 goto invalid_optional;
2238 if (autocommit_msecs > 3600000)
2239 goto invalid_optional;
2240 wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2241 wc->autocommit_time_set = true;
2242 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2243 unsigned max_age_msecs;
2244 string = dm_shift_arg(&as), opt_params--;
2245 if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2246 goto invalid_optional;
2247 if (max_age_msecs > 86400000)
2248 goto invalid_optional;
2249 wc->max_age = msecs_to_jiffies(max_age_msecs);
2250 } else if (!strcasecmp(string, "cleaner")) {
2251 wc->cleaner = true;
2252 } else if (!strcasecmp(string, "fua")) {
2253 if (WC_MODE_PMEM(wc)) {
2254 wc->writeback_fua = true;
2255 wc->writeback_fua_set = true;
2256 } else goto invalid_optional;
2257 } else if (!strcasecmp(string, "nofua")) {
2258 if (WC_MODE_PMEM(wc)) {
2259 wc->writeback_fua = false;
2260 wc->writeback_fua_set = true;
2261 } else goto invalid_optional;
2262 } else {
2263invalid_optional:
2264 r = -EINVAL;
2265 ti->error = "Invalid optional argument";
2266 goto bad;
2267 }
2268 }
2269
2270 if (high_wm_percent < low_wm_percent) {
2271 r = -EINVAL;
2272 ti->error = "High watermark must be greater than or equal to low watermark";
2273 goto bad;
2274 }
2275
2276 if (WC_MODE_PMEM(wc)) {
2277 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2278 r = -EOPNOTSUPP;
2279 ti->error = "Asynchronous persistent memory not supported as pmem cache";
2280 goto bad;
2281 }
2282
2283 r = persistent_memory_claim(wc);
2284 if (r) {
2285 ti->error = "Unable to map persistent memory for cache";
2286 goto bad;
2287 }
2288 } else {
2289 size_t n_blocks, n_metadata_blocks;
2290 uint64_t n_bitmap_bits;
2291
2292 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2293
2294 bio_list_init(&wc->flush_list);
2295 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2296 if (IS_ERR(wc->flush_thread)) {
2297 r = PTR_ERR(wc->flush_thread);
2298 wc->flush_thread = NULL;
2299 ti->error = "Couldn't spawn flush thread";
2300 goto bad;
2301 }
2302 wake_up_process(wc->flush_thread);
2303
2304 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2305 &n_blocks, &n_metadata_blocks);
2306 if (r) {
2307 ti->error = "Invalid device size";
2308 goto bad;
2309 }
2310
2311 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2312 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2313 /* this is limitation of test_bit functions */
2314 if (n_bitmap_bits > 1U << 31) {
2315 r = -EFBIG;
2316 ti->error = "Invalid device size";
2317 goto bad;
2318 }
2319
2320 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2321 if (!wc->memory_map) {
2322 r = -ENOMEM;
2323 ti->error = "Unable to allocate memory for metadata";
2324 goto bad;
2325 }
2326
2327 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2328 if (IS_ERR(wc->dm_kcopyd)) {
2329 r = PTR_ERR(wc->dm_kcopyd);
2330 ti->error = "Unable to allocate dm-kcopyd client";
2331 wc->dm_kcopyd = NULL;
2332 goto bad;
2333 }
2334
2335 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2336 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2337 BITS_PER_LONG * sizeof(unsigned long);
2338 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2339 if (!wc->dirty_bitmap) {
2340 r = -ENOMEM;
2341 ti->error = "Unable to allocate dirty bitmap";
2342 goto bad;
2343 }
2344
2345 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2346 if (r) {
2347 ti->error = "Unable to read first block of metadata";
2348 goto bad;
2349 }
2350 }
2351
2352 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2353 if (r) {
2354 ti->error = "Hardware memory error when reading superblock";
2355 goto bad;
2356 }
2357 if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2358 r = init_memory(wc);
2359 if (r) {
2360 ti->error = "Unable to initialize device";
2361 goto bad;
2362 }
2363 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2364 if (r) {
2365 ti->error = "Hardware memory error when reading superblock";
2366 goto bad;
2367 }
2368 }
2369
2370 if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2371 ti->error = "Invalid magic in the superblock";
2372 r = -EINVAL;
2373 goto bad;
2374 }
2375
2376 if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2377 ti->error = "Invalid version in the superblock";
2378 r = -EINVAL;
2379 goto bad;
2380 }
2381
2382 if (le32_to_cpu(s.block_size) != wc->block_size) {
2383 ti->error = "Block size does not match superblock";
2384 r = -EINVAL;
2385 goto bad;
2386 }
2387
2388 wc->n_blocks = le64_to_cpu(s.n_blocks);
2389
2390 offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2391 if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2392overflow:
2393 ti->error = "Overflow in size calculation";
2394 r = -EINVAL;
2395 goto bad;
2396 }
2397 offset += sizeof(struct wc_memory_superblock);
2398 if (offset < sizeof(struct wc_memory_superblock))
2399 goto overflow;
2400 offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2401 data_size = wc->n_blocks * (size_t)wc->block_size;
2402 if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2403 (offset + data_size < offset))
2404 goto overflow;
2405 if (offset + data_size > wc->memory_map_size) {
2406 ti->error = "Memory area is too small";
2407 r = -EINVAL;
2408 goto bad;
2409 }
2410
2411 wc->metadata_sectors = offset >> SECTOR_SHIFT;
2412 wc->block_start = (char *)sb(wc) + offset;
2413
2414 x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2415 x += 50;
2416 do_div(x, 100);
2417 wc->freelist_high_watermark = x;
2418 x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2419 x += 50;
2420 do_div(x, 100);
2421 wc->freelist_low_watermark = x;
2422
2423 if (wc->cleaner)
2424 activate_cleaner(wc);
2425
2426 r = writecache_alloc_entries(wc);
2427 if (r) {
2428 ti->error = "Cannot allocate memory";
2429 goto bad;
2430 }
2431
2432 ti->num_flush_bios = 1;
2433 ti->flush_supported = true;
2434 ti->num_discard_bios = 1;
2435
2436 if (WC_MODE_PMEM(wc))
2437 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2438
2439 return 0;
2440
2441bad_arguments:
2442 r = -EINVAL;
2443 ti->error = "Bad arguments";
2444bad:
2445 writecache_dtr(ti);
2446 return r;
2447}
2448
2449static void writecache_status(struct dm_target *ti, status_type_t type,
2450 unsigned status_flags, char *result, unsigned maxlen)
2451{
2452 struct dm_writecache *wc = ti->private;
2453 unsigned extra_args;
2454 unsigned sz = 0;
2455 uint64_t x;
2456
2457 switch (type) {
2458 case STATUSTYPE_INFO:
2459 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2460 (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2461 (unsigned long long)wc->writeback_size);
2462 break;
2463 case STATUSTYPE_TABLE:
2464 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2465 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2466 extra_args = 0;
2467 if (wc->start_sector)
2468 extra_args += 2;
2469 if (wc->high_wm_percent_set && !wc->cleaner)
2470 extra_args += 2;
2471 if (wc->low_wm_percent_set && !wc->cleaner)
2472 extra_args += 2;
2473 if (wc->max_writeback_jobs_set)
2474 extra_args += 2;
2475 if (wc->autocommit_blocks_set)
2476 extra_args += 2;
2477 if (wc->autocommit_time_set)
2478 extra_args += 2;
2479 if (wc->cleaner)
2480 extra_args++;
2481 if (wc->writeback_fua_set)
2482 extra_args++;
2483
2484 DMEMIT("%u", extra_args);
2485 if (wc->start_sector)
2486 DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2487 if (wc->high_wm_percent_set && !wc->cleaner) {
2488 x = (uint64_t)wc->freelist_high_watermark * 100;
2489 x += wc->n_blocks / 2;
2490 do_div(x, (size_t)wc->n_blocks);
2491 DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2492 }
2493 if (wc->low_wm_percent_set && !wc->cleaner) {
2494 x = (uint64_t)wc->freelist_low_watermark * 100;
2495 x += wc->n_blocks / 2;
2496 do_div(x, (size_t)wc->n_blocks);
2497 DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2498 }
2499 if (wc->max_writeback_jobs_set)
2500 DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2501 if (wc->autocommit_blocks_set)
2502 DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2503 if (wc->autocommit_time_set)
2504 DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2505 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2506 DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2507 if (wc->cleaner)
2508 DMEMIT(" cleaner");
2509 if (wc->writeback_fua_set)
2510 DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2511 break;
2512 }
2513}
2514
2515static struct target_type writecache_target = {
2516 .name = "writecache",
2517 .version = {1, 3, 0},
2518 .module = THIS_MODULE,
2519 .ctr = writecache_ctr,
2520 .dtr = writecache_dtr,
2521 .status = writecache_status,
2522 .postsuspend = writecache_suspend,
2523 .resume = writecache_resume,
2524 .message = writecache_message,
2525 .map = writecache_map,
2526 .end_io = writecache_end_io,
2527 .iterate_devices = writecache_iterate_devices,
2528 .io_hints = writecache_io_hints,
2529};
2530
2531static int __init dm_writecache_init(void)
2532{
2533 int r;
2534
2535 r = dm_register_target(&writecache_target);
2536 if (r < 0) {
2537 DMERR("register failed %d", r);
2538 return r;
2539 }
2540
2541 return 0;
2542}
2543
2544static void __exit dm_writecache_exit(void)
2545{
2546 dm_unregister_target(&writecache_target);
2547}
2548
2549module_init(dm_writecache_init);
2550module_exit(dm_writecache_exit);
2551
2552MODULE_DESCRIPTION(DM_NAME " writecache target");
2553MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2554MODULE_LICENSE("GPL");