?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'trace-v5.11-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt...
[linux/fpc-iii.git] / drivers / md / dm-writecache.c
blobd5223a0e5cc51e309c93ab035b2f54fd7ce7d69c
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) \
44 do { \
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 IS_ENABLED(CONFIG_ARCH_HAS_COPY_MC) && 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
59 struct wc_memory_entry {
60 __le64 original_sector;
61 __le64 seq_count;
62 };
63
64 struct 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
79 struct 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
109 struct 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
195 struct 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
204 struct 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
212 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
213 "A percentage of time allocated for data copying");
214
215 static void wc_lock(struct dm_writecache *wc)
216 {
217 mutex_lock(&wc->lock);
218 }
219
220 static void wc_unlock(struct dm_writecache *wc)
221 {
222 mutex_unlock(&wc->lock);
223 }
224
225 #ifdef DM_WRITECACHE_HAS_PMEM
226 static 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;
312 err3:
313 kvfree(pages);
314 err2:
315 dax_read_unlock(id);
316 err1:
317 return r;
318 }
319 #else
320 static int persistent_memory_claim(struct dm_writecache *wc)
321 {
322 return -EOPNOTSUPP;
323 }
324 #endif
325
326 static 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
332 static 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
340 static unsigned persistent_memory_page_offset(void *addr)
341 {
342 return (unsigned long)addr & (PAGE_SIZE - 1);
343 }
344
345 static 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
351 static 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
357 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
358 {
359 return wc->memory_map;
360 }
361
362 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
363 {
364 return &sb(wc)->entries[e->index];
365 }
366
367 static 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
372 static 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
378 static 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
387 static 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
396 static 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
404 static 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...) \
418 do { \
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
426 static 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
432 static 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
439 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
440
441 struct io_notify {
442 struct dm_writecache *wc;
443 struct completion c;
444 atomic_t count;
445 };
446
447 static 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
458 static 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
464 static 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, &region, 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
518 static 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, &region, NULL);
542 if (unlikely(r))
543 writecache_error(wc, r, "error writing superblock");
544 }
545
546 static 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
554 static 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, &region, 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
578 static 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
624 static 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
643 static 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
649 static 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
670 static 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
676 static 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
686 static 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
717 static 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
727 static 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
738 static 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
750 static 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
757 static 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
762 static 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
826 static 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
835 static 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
842 static 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
848 static 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
883 static 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
892 static 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
925 static 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
944 static 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, &region, NULL);
960 }
961
962 static 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 = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
996 sizeof(uint64_t));
997 if (r) {
998 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
999 sb_seq_count = cpu_to_le64(0);
1000 }
1001 wc->seq_count = le64_to_cpu(sb_seq_count);
1002
1003 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1004 for (b = 0; b < wc->n_blocks; b++) {
1005 struct wc_entry *e = &wc->entries[b];
1006 struct wc_memory_entry wme;
1007 if (writecache_has_error(wc)) {
1008 e->original_sector = -1;
1009 e->seq_count = -1;
1010 continue;
1011 }
1012 r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1013 sizeof(struct wc_memory_entry));
1014 if (r) {
1015 writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1016 (unsigned long)b, r);
1017 e->original_sector = -1;
1018 e->seq_count = -1;
1019 } else {
1020 e->original_sector = le64_to_cpu(wme.original_sector);
1021 e->seq_count = le64_to_cpu(wme.seq_count);
1022 }
1023 cond_resched();
1024 }
1025 #endif
1026 for (b = 0; b < wc->n_blocks; b++) {
1027 struct wc_entry *e = &wc->entries[b];
1028 if (!writecache_entry_is_committed(wc, e)) {
1029 if (read_seq_count(wc, e) != -1) {
1030 erase_this:
1031 clear_seq_count(wc, e);
1032 need_flush = true;
1033 }
1034 writecache_add_to_freelist(wc, e);
1035 } else {
1036 struct wc_entry *old;
1037
1038 old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1039 if (!old) {
1040 writecache_insert_entry(wc, e);
1041 } else {
1042 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1043 writecache_error(wc, -EINVAL,
1044 "two identical entries, position %llu, sector %llu, sequence %llu",
1045 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1046 (unsigned long long)read_seq_count(wc, e));
1047 }
1048 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1049 goto erase_this;
1050 } else {
1051 writecache_free_entry(wc, old);
1052 writecache_insert_entry(wc, e);
1053 need_flush = true;
1054 }
1055 }
1056 }
1057 cond_resched();
1058 }
1059
1060 if (need_flush) {
1061 writecache_flush_all_metadata(wc);
1062 writecache_commit_flushed(wc, false);
1063 }
1064
1065 writecache_verify_watermark(wc);
1066
1067 if (wc->max_age != MAX_AGE_UNSPECIFIED)
1068 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1069
1070 wc_unlock(wc);
1071 }
1072
1073 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1074 {
1075 if (argc != 1)
1076 return -EINVAL;
1077
1078 wc_lock(wc);
1079 if (dm_suspended(wc->ti)) {
1080 wc_unlock(wc);
1081 return -EBUSY;
1082 }
1083 if (writecache_has_error(wc)) {
1084 wc_unlock(wc);
1085 return -EIO;
1086 }
1087
1088 writecache_flush(wc);
1089 wc->writeback_all++;
1090 queue_work(wc->writeback_wq, &wc->writeback_work);
1091 wc_unlock(wc);
1092
1093 flush_workqueue(wc->writeback_wq);
1094
1095 wc_lock(wc);
1096 wc->writeback_all--;
1097 if (writecache_has_error(wc)) {
1098 wc_unlock(wc);
1099 return -EIO;
1100 }
1101 wc_unlock(wc);
1102
1103 return 0;
1104 }
1105
1106 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1107 {
1108 if (argc != 1)
1109 return -EINVAL;
1110
1111 wc_lock(wc);
1112 wc->flush_on_suspend = true;
1113 wc_unlock(wc);
1114
1115 return 0;
1116 }
1117
1118 static void activate_cleaner(struct dm_writecache *wc)
1119 {
1120 wc->flush_on_suspend = true;
1121 wc->cleaner = true;
1122 wc->freelist_high_watermark = wc->n_blocks;
1123 wc->freelist_low_watermark = wc->n_blocks;
1124 }
1125
1126 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1127 {
1128 if (argc != 1)
1129 return -EINVAL;
1130
1131 wc_lock(wc);
1132 activate_cleaner(wc);
1133 if (!dm_suspended(wc->ti))
1134 writecache_verify_watermark(wc);
1135 wc_unlock(wc);
1136
1137 return 0;
1138 }
1139
1140 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1141 char *result, unsigned maxlen)
1142 {
1143 int r = -EINVAL;
1144 struct dm_writecache *wc = ti->private;
1145
1146 if (!strcasecmp(argv[0], "flush"))
1147 r = process_flush_mesg(argc, argv, wc);
1148 else if (!strcasecmp(argv[0], "flush_on_suspend"))
1149 r = process_flush_on_suspend_mesg(argc, argv, wc);
1150 else if (!strcasecmp(argv[0], "cleaner"))
1151 r = process_cleaner_mesg(argc, argv, wc);
1152 else
1153 DMERR("unrecognised message received: %s", argv[0]);
1154
1155 return r;
1156 }
1157
1158 static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1159 {
1160 /*
1161 * clflushopt performs better with block size 1024, 2048, 4096
1162 * non-temporal stores perform better with block size 512
1163 *
1164 * block size 512 1024 2048 4096
1165 * movnti 496 MB/s 642 MB/s 725 MB/s 744 MB/s
1166 * clflushopt 373 MB/s 688 MB/s 1.1 GB/s 1.2 GB/s
1167 *
1168 * We see that movnti performs better for 512-byte blocks, and
1169 * clflushopt performs better for 1024-byte and larger blocks. So, we
1170 * prefer clflushopt for sizes >= 768.
1171 *
1172 * NOTE: this happens to be the case now (with dm-writecache's single
1173 * threaded model) but re-evaluate this once memcpy_flushcache() is
1174 * enabled to use movdir64b which might invalidate this performance
1175 * advantage seen with cache-allocating-writes plus flushing.
1176 */
1177 #ifdef CONFIG_X86
1178 if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1179 likely(boot_cpu_data.x86_clflush_size == 64) &&
1180 likely(size >= 768)) {
1181 do {
1182 memcpy((void *)dest, (void *)source, 64);
1183 clflushopt((void *)dest);
1184 dest += 64;
1185 source += 64;
1186 size -= 64;
1187 } while (size >= 64);
1188 return;
1189 }
1190 #endif
1191 memcpy_flushcache(dest, source, size);
1192 }
1193
1194 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1195 {
1196 void *buf;
1197 unsigned long flags;
1198 unsigned size;
1199 int rw = bio_data_dir(bio);
1200 unsigned remaining_size = wc->block_size;
1201
1202 do {
1203 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1204 buf = bvec_kmap_irq(&bv, &flags);
1205 size = bv.bv_len;
1206 if (unlikely(size > remaining_size))
1207 size = remaining_size;
1208
1209 if (rw == READ) {
1210 int r;
1211 r = copy_mc_to_kernel(buf, data, size);
1212 flush_dcache_page(bio_page(bio));
1213 if (unlikely(r)) {
1214 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1215 bio->bi_status = BLK_STS_IOERR;
1216 }
1217 } else {
1218 flush_dcache_page(bio_page(bio));
1219 memcpy_flushcache_optimized(data, buf, size);
1220 }
1221
1222 bvec_kunmap_irq(buf, &flags);
1223
1224 data = (char *)data + size;
1225 remaining_size -= size;
1226 bio_advance(bio, size);
1227 } while (unlikely(remaining_size));
1228 }
1229
1230 static int writecache_flush_thread(void *data)
1231 {
1232 struct dm_writecache *wc = data;
1233
1234 while (1) {
1235 struct bio *bio;
1236
1237 wc_lock(wc);
1238 bio = bio_list_pop(&wc->flush_list);
1239 if (!bio) {
1240 set_current_state(TASK_INTERRUPTIBLE);
1241 wc_unlock(wc);
1242
1243 if (unlikely(kthread_should_stop())) {
1244 set_current_state(TASK_RUNNING);
1245 break;
1246 }
1247
1248 schedule();
1249 continue;
1250 }
1251
1252 if (bio_op(bio) == REQ_OP_DISCARD) {
1253 writecache_discard(wc, bio->bi_iter.bi_sector,
1254 bio_end_sector(bio));
1255 wc_unlock(wc);
1256 bio_set_dev(bio, wc->dev->bdev);
1257 submit_bio_noacct(bio);
1258 } else {
1259 writecache_flush(wc);
1260 wc_unlock(wc);
1261 if (writecache_has_error(wc))
1262 bio->bi_status = BLK_STS_IOERR;
1263 bio_endio(bio);
1264 }
1265 }
1266
1267 return 0;
1268 }
1269
1270 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1271 {
1272 if (bio_list_empty(&wc->flush_list))
1273 wake_up_process(wc->flush_thread);
1274 bio_list_add(&wc->flush_list, bio);
1275 }
1276
1277 static int writecache_map(struct dm_target *ti, struct bio *bio)
1278 {
1279 struct wc_entry *e;
1280 struct dm_writecache *wc = ti->private;
1281
1282 bio->bi_private = NULL;
1283
1284 wc_lock(wc);
1285
1286 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1287 if (writecache_has_error(wc))
1288 goto unlock_error;
1289 if (WC_MODE_PMEM(wc)) {
1290 writecache_flush(wc);
1291 if (writecache_has_error(wc))
1292 goto unlock_error;
1293 goto unlock_submit;
1294 } else {
1295 writecache_offload_bio(wc, bio);
1296 goto unlock_return;
1297 }
1298 }
1299
1300 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1301
1302 if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1303 (wc->block_size / 512 - 1)) != 0)) {
1304 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1305 (unsigned long long)bio->bi_iter.bi_sector,
1306 bio->bi_iter.bi_size, wc->block_size);
1307 goto unlock_error;
1308 }
1309
1310 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1311 if (writecache_has_error(wc))
1312 goto unlock_error;
1313 if (WC_MODE_PMEM(wc)) {
1314 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1315 goto unlock_remap_origin;
1316 } else {
1317 writecache_offload_bio(wc, bio);
1318 goto unlock_return;
1319 }
1320 }
1321
1322 if (bio_data_dir(bio) == READ) {
1323 read_next_block:
1324 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1325 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1326 if (WC_MODE_PMEM(wc)) {
1327 bio_copy_block(wc, bio, memory_data(wc, e));
1328 if (bio->bi_iter.bi_size)
1329 goto read_next_block;
1330 goto unlock_submit;
1331 } else {
1332 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1333 bio_set_dev(bio, wc->ssd_dev->bdev);
1334 bio->bi_iter.bi_sector = cache_sector(wc, e);
1335 if (!writecache_entry_is_committed(wc, e))
1336 writecache_wait_for_ios(wc, WRITE);
1337 goto unlock_remap;
1338 }
1339 } else {
1340 if (e) {
1341 sector_t next_boundary =
1342 read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1343 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1344 dm_accept_partial_bio(bio, next_boundary);
1345 }
1346 }
1347 goto unlock_remap_origin;
1348 }
1349 } else {
1350 do {
1351 bool found_entry = false;
1352 if (writecache_has_error(wc))
1353 goto unlock_error;
1354 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1355 if (e) {
1356 if (!writecache_entry_is_committed(wc, e))
1357 goto bio_copy;
1358 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1359 wc->overwrote_committed = true;
1360 goto bio_copy;
1361 }
1362 found_entry = true;
1363 } else {
1364 if (unlikely(wc->cleaner))
1365 goto direct_write;
1366 }
1367 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1368 if (unlikely(!e)) {
1369 if (!found_entry) {
1370 direct_write:
1371 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1372 if (e) {
1373 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1374 BUG_ON(!next_boundary);
1375 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1376 dm_accept_partial_bio(bio, next_boundary);
1377 }
1378 }
1379 goto unlock_remap_origin;
1380 }
1381 writecache_wait_on_freelist(wc);
1382 continue;
1383 }
1384 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1385 writecache_insert_entry(wc, e);
1386 wc->uncommitted_blocks++;
1387 bio_copy:
1388 if (WC_MODE_PMEM(wc)) {
1389 bio_copy_block(wc, bio, memory_data(wc, e));
1390 } else {
1391 unsigned bio_size = wc->block_size;
1392 sector_t start_cache_sec = cache_sector(wc, e);
1393 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1394
1395 while (bio_size < bio->bi_iter.bi_size) {
1396 struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1397 if (!f)
1398 break;
1399 write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1400 (bio_size >> SECTOR_SHIFT), wc->seq_count);
1401 writecache_insert_entry(wc, f);
1402 wc->uncommitted_blocks++;
1403 bio_size += wc->block_size;
1404 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1405 }
1406
1407 bio_set_dev(bio, wc->ssd_dev->bdev);
1408 bio->bi_iter.bi_sector = start_cache_sec;
1409 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1410
1411 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1412 wc->uncommitted_blocks = 0;
1413 queue_work(wc->writeback_wq, &wc->flush_work);
1414 } else {
1415 writecache_schedule_autocommit(wc);
1416 }
1417 goto unlock_remap;
1418 }
1419 } while (bio->bi_iter.bi_size);
1420
1421 if (unlikely(bio->bi_opf & REQ_FUA ||
1422 wc->uncommitted_blocks >= wc->autocommit_blocks))
1423 writecache_flush(wc);
1424 else
1425 writecache_schedule_autocommit(wc);
1426 goto unlock_submit;
1427 }
1428
1429 unlock_remap_origin:
1430 bio_set_dev(bio, wc->dev->bdev);
1431 wc_unlock(wc);
1432 return DM_MAPIO_REMAPPED;
1433
1434 unlock_remap:
1435 /* make sure that writecache_end_io decrements bio_in_progress: */
1436 bio->bi_private = (void *)1;
1437 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1438 wc_unlock(wc);
1439 return DM_MAPIO_REMAPPED;
1440
1441 unlock_submit:
1442 wc_unlock(wc);
1443 bio_endio(bio);
1444 return DM_MAPIO_SUBMITTED;
1445
1446 unlock_return:
1447 wc_unlock(wc);
1448 return DM_MAPIO_SUBMITTED;
1449
1450 unlock_error:
1451 wc_unlock(wc);
1452 bio_io_error(bio);
1453 return DM_MAPIO_SUBMITTED;
1454 }
1455
1456 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1457 {
1458 struct dm_writecache *wc = ti->private;
1459
1460 if (bio->bi_private != NULL) {
1461 int dir = bio_data_dir(bio);
1462 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1463 if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1464 wake_up(&wc->bio_in_progress_wait[dir]);
1465 }
1466 return 0;
1467 }
1468
1469 static int writecache_iterate_devices(struct dm_target *ti,
1470 iterate_devices_callout_fn fn, void *data)
1471 {
1472 struct dm_writecache *wc = ti->private;
1473
1474 return fn(ti, wc->dev, 0, ti->len, data);
1475 }
1476
1477 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1478 {
1479 struct dm_writecache *wc = ti->private;
1480
1481 if (limits->logical_block_size < wc->block_size)
1482 limits->logical_block_size = wc->block_size;
1483
1484 if (limits->physical_block_size < wc->block_size)
1485 limits->physical_block_size = wc->block_size;
1486
1487 if (limits->io_min < wc->block_size)
1488 limits->io_min = wc->block_size;
1489 }
1490
1491
1492 static void writecache_writeback_endio(struct bio *bio)
1493 {
1494 struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1495 struct dm_writecache *wc = wb->wc;
1496 unsigned long flags;
1497
1498 raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1499 if (unlikely(list_empty(&wc->endio_list)))
1500 wake_up_process(wc->endio_thread);
1501 list_add_tail(&wb->endio_entry, &wc->endio_list);
1502 raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1503 }
1504
1505 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1506 {
1507 struct copy_struct *c = ptr;
1508 struct dm_writecache *wc = c->wc;
1509
1510 c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1511
1512 raw_spin_lock_irq(&wc->endio_list_lock);
1513 if (unlikely(list_empty(&wc->endio_list)))
1514 wake_up_process(wc->endio_thread);
1515 list_add_tail(&c->endio_entry, &wc->endio_list);
1516 raw_spin_unlock_irq(&wc->endio_list_lock);
1517 }
1518
1519 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1520 {
1521 unsigned i;
1522 struct writeback_struct *wb;
1523 struct wc_entry *e;
1524 unsigned long n_walked = 0;
1525
1526 do {
1527 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1528 list_del(&wb->endio_entry);
1529
1530 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1531 writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1532 "write error %d", wb->bio.bi_status);
1533 i = 0;
1534 do {
1535 e = wb->wc_list[i];
1536 BUG_ON(!e->write_in_progress);
1537 e->write_in_progress = false;
1538 INIT_LIST_HEAD(&e->lru);
1539 if (!writecache_has_error(wc))
1540 writecache_free_entry(wc, e);
1541 BUG_ON(!wc->writeback_size);
1542 wc->writeback_size--;
1543 n_walked++;
1544 if (unlikely(n_walked >= ENDIO_LATENCY)) {
1545 writecache_commit_flushed(wc, false);
1546 wc_unlock(wc);
1547 wc_lock(wc);
1548 n_walked = 0;
1549 }
1550 } while (++i < wb->wc_list_n);
1551
1552 if (wb->wc_list != wb->wc_list_inline)
1553 kfree(wb->wc_list);
1554 bio_put(&wb->bio);
1555 } while (!list_empty(list));
1556 }
1557
1558 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1559 {
1560 struct copy_struct *c;
1561 struct wc_entry *e;
1562
1563 do {
1564 c = list_entry(list->next, struct copy_struct, endio_entry);
1565 list_del(&c->endio_entry);
1566
1567 if (unlikely(c->error))
1568 writecache_error(wc, c->error, "copy error");
1569
1570 e = c->e;
1571 do {
1572 BUG_ON(!e->write_in_progress);
1573 e->write_in_progress = false;
1574 INIT_LIST_HEAD(&e->lru);
1575 if (!writecache_has_error(wc))
1576 writecache_free_entry(wc, e);
1577
1578 BUG_ON(!wc->writeback_size);
1579 wc->writeback_size--;
1580 e++;
1581 } while (--c->n_entries);
1582 mempool_free(c, &wc->copy_pool);
1583 } while (!list_empty(list));
1584 }
1585
1586 static int writecache_endio_thread(void *data)
1587 {
1588 struct dm_writecache *wc = data;
1589
1590 while (1) {
1591 struct list_head list;
1592
1593 raw_spin_lock_irq(&wc->endio_list_lock);
1594 if (!list_empty(&wc->endio_list))
1595 goto pop_from_list;
1596 set_current_state(TASK_INTERRUPTIBLE);
1597 raw_spin_unlock_irq(&wc->endio_list_lock);
1598
1599 if (unlikely(kthread_should_stop())) {
1600 set_current_state(TASK_RUNNING);
1601 break;
1602 }
1603
1604 schedule();
1605
1606 continue;
1607
1608 pop_from_list:
1609 list = wc->endio_list;
1610 list.next->prev = list.prev->next = &list;
1611 INIT_LIST_HEAD(&wc->endio_list);
1612 raw_spin_unlock_irq(&wc->endio_list_lock);
1613
1614 if (!WC_MODE_FUA(wc))
1615 writecache_disk_flush(wc, wc->dev);
1616
1617 wc_lock(wc);
1618
1619 if (WC_MODE_PMEM(wc)) {
1620 __writecache_endio_pmem(wc, &list);
1621 } else {
1622 __writecache_endio_ssd(wc, &list);
1623 writecache_wait_for_ios(wc, READ);
1624 }
1625
1626 writecache_commit_flushed(wc, false);
1627
1628 wc_unlock(wc);
1629 }
1630
1631 return 0;
1632 }
1633
1634 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1635 {
1636 struct dm_writecache *wc = wb->wc;
1637 unsigned block_size = wc->block_size;
1638 void *address = memory_data(wc, e);
1639
1640 persistent_memory_flush_cache(address, block_size);
1641 return bio_add_page(&wb->bio, persistent_memory_page(address),
1642 block_size, persistent_memory_page_offset(address)) != 0;
1643 }
1644
1645 struct writeback_list {
1646 struct list_head list;
1647 size_t size;
1648 };
1649
1650 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1651 {
1652 if (unlikely(wc->max_writeback_jobs)) {
1653 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1654 wc_lock(wc);
1655 while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1656 writecache_wait_on_freelist(wc);
1657 wc_unlock(wc);
1658 }
1659 }
1660 cond_resched();
1661 }
1662
1663 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1664 {
1665 struct wc_entry *e, *f;
1666 struct bio *bio;
1667 struct writeback_struct *wb;
1668 unsigned max_pages;
1669
1670 while (wbl->size) {
1671 wbl->size--;
1672 e = container_of(wbl->list.prev, struct wc_entry, lru);
1673 list_del(&e->lru);
1674
1675 max_pages = e->wc_list_contiguous;
1676
1677 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1678 wb = container_of(bio, struct writeback_struct, bio);
1679 wb->wc = wc;
1680 bio->bi_end_io = writecache_writeback_endio;
1681 bio_set_dev(bio, wc->dev->bdev);
1682 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1683 if (max_pages <= WB_LIST_INLINE ||
1684 unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1685 GFP_NOIO | __GFP_NORETRY |
1686 __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1687 wb->wc_list = wb->wc_list_inline;
1688 max_pages = WB_LIST_INLINE;
1689 }
1690
1691 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1692
1693 wb->wc_list[0] = e;
1694 wb->wc_list_n = 1;
1695
1696 while (wbl->size && wb->wc_list_n < max_pages) {
1697 f = container_of(wbl->list.prev, struct wc_entry, lru);
1698 if (read_original_sector(wc, f) !=
1699 read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1700 break;
1701 if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1702 break;
1703 wbl->size--;
1704 list_del(&f->lru);
1705 wb->wc_list[wb->wc_list_n++] = f;
1706 e = f;
1707 }
1708 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1709 if (writecache_has_error(wc)) {
1710 bio->bi_status = BLK_STS_IOERR;
1711 bio_endio(bio);
1712 } else {
1713 submit_bio(bio);
1714 }
1715
1716 __writeback_throttle(wc, wbl);
1717 }
1718 }
1719
1720 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1721 {
1722 struct wc_entry *e, *f;
1723 struct dm_io_region from, to;
1724 struct copy_struct *c;
1725
1726 while (wbl->size) {
1727 unsigned n_sectors;
1728
1729 wbl->size--;
1730 e = container_of(wbl->list.prev, struct wc_entry, lru);
1731 list_del(&e->lru);
1732
1733 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1734
1735 from.bdev = wc->ssd_dev->bdev;
1736 from.sector = cache_sector(wc, e);
1737 from.count = n_sectors;
1738 to.bdev = wc->dev->bdev;
1739 to.sector = read_original_sector(wc, e);
1740 to.count = n_sectors;
1741
1742 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1743 c->wc = wc;
1744 c->e = e;
1745 c->n_entries = e->wc_list_contiguous;
1746
1747 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1748 wbl->size--;
1749 f = container_of(wbl->list.prev, struct wc_entry, lru);
1750 BUG_ON(f != e + 1);
1751 list_del(&f->lru);
1752 e = f;
1753 }
1754
1755 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1756
1757 __writeback_throttle(wc, wbl);
1758 }
1759 }
1760
1761 static void writecache_writeback(struct work_struct *work)
1762 {
1763 struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1764 struct blk_plug plug;
1765 struct wc_entry *f, *g, *e = NULL;
1766 struct rb_node *node, *next_node;
1767 struct list_head skipped;
1768 struct writeback_list wbl;
1769 unsigned long n_walked;
1770
1771 wc_lock(wc);
1772 restart:
1773 if (writecache_has_error(wc)) {
1774 wc_unlock(wc);
1775 return;
1776 }
1777
1778 if (unlikely(wc->writeback_all)) {
1779 if (writecache_wait_for_writeback(wc))
1780 goto restart;
1781 }
1782
1783 if (wc->overwrote_committed) {
1784 writecache_wait_for_ios(wc, WRITE);
1785 }
1786
1787 n_walked = 0;
1788 INIT_LIST_HEAD(&skipped);
1789 INIT_LIST_HEAD(&wbl.list);
1790 wbl.size = 0;
1791 while (!list_empty(&wc->lru) &&
1792 (wc->writeback_all ||
1793 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1794 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1795 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1796
1797 n_walked++;
1798 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1799 likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1800 queue_work(wc->writeback_wq, &wc->writeback_work);
1801 break;
1802 }
1803
1804 if (unlikely(wc->writeback_all)) {
1805 if (unlikely(!e)) {
1806 writecache_flush(wc);
1807 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1808 } else
1809 e = g;
1810 } else
1811 e = container_of(wc->lru.prev, struct wc_entry, lru);
1812 BUG_ON(e->write_in_progress);
1813 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1814 writecache_flush(wc);
1815 }
1816 node = rb_prev(&e->rb_node);
1817 if (node) {
1818 f = container_of(node, struct wc_entry, rb_node);
1819 if (unlikely(read_original_sector(wc, f) ==
1820 read_original_sector(wc, e))) {
1821 BUG_ON(!f->write_in_progress);
1822 list_del(&e->lru);
1823 list_add(&e->lru, &skipped);
1824 cond_resched();
1825 continue;
1826 }
1827 }
1828 wc->writeback_size++;
1829 list_del(&e->lru);
1830 list_add(&e->lru, &wbl.list);
1831 wbl.size++;
1832 e->write_in_progress = true;
1833 e->wc_list_contiguous = 1;
1834
1835 f = e;
1836
1837 while (1) {
1838 next_node = rb_next(&f->rb_node);
1839 if (unlikely(!next_node))
1840 break;
1841 g = container_of(next_node, struct wc_entry, rb_node);
1842 if (unlikely(read_original_sector(wc, g) ==
1843 read_original_sector(wc, f))) {
1844 f = g;
1845 continue;
1846 }
1847 if (read_original_sector(wc, g) !=
1848 read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1849 break;
1850 if (unlikely(g->write_in_progress))
1851 break;
1852 if (unlikely(!writecache_entry_is_committed(wc, g)))
1853 break;
1854
1855 if (!WC_MODE_PMEM(wc)) {
1856 if (g != f + 1)
1857 break;
1858 }
1859
1860 n_walked++;
1861 //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1862 // break;
1863
1864 wc->writeback_size++;
1865 list_del(&g->lru);
1866 list_add(&g->lru, &wbl.list);
1867 wbl.size++;
1868 g->write_in_progress = true;
1869 g->wc_list_contiguous = BIO_MAX_PAGES;
1870 f = g;
1871 e->wc_list_contiguous++;
1872 if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1873 if (unlikely(wc->writeback_all)) {
1874 next_node = rb_next(&f->rb_node);
1875 if (likely(next_node))
1876 g = container_of(next_node, struct wc_entry, rb_node);
1877 }
1878 break;
1879 }
1880 }
1881 cond_resched();
1882 }
1883
1884 if (!list_empty(&skipped)) {
1885 list_splice_tail(&skipped, &wc->lru);
1886 /*
1887 * If we didn't do any progress, we must wait until some
1888 * writeback finishes to avoid burning CPU in a loop
1889 */
1890 if (unlikely(!wbl.size))
1891 writecache_wait_for_writeback(wc);
1892 }
1893
1894 wc_unlock(wc);
1895
1896 blk_start_plug(&plug);
1897
1898 if (WC_MODE_PMEM(wc))
1899 __writecache_writeback_pmem(wc, &wbl);
1900 else
1901 __writecache_writeback_ssd(wc, &wbl);
1902
1903 blk_finish_plug(&plug);
1904
1905 if (unlikely(wc->writeback_all)) {
1906 wc_lock(wc);
1907 while (writecache_wait_for_writeback(wc));
1908 wc_unlock(wc);
1909 }
1910 }
1911
1912 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1913 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1914 {
1915 uint64_t n_blocks, offset;
1916 struct wc_entry e;
1917
1918 n_blocks = device_size;
1919 do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1920
1921 while (1) {
1922 if (!n_blocks)
1923 return -ENOSPC;
1924 /* Verify the following entries[n_blocks] won't overflow */
1925 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1926 sizeof(struct wc_memory_entry)))
1927 return -EFBIG;
1928 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1929 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1930 if (offset + n_blocks * block_size <= device_size)
1931 break;
1932 n_blocks--;
1933 }
1934
1935 /* check if the bit field overflows */
1936 e.index = n_blocks;
1937 if (e.index != n_blocks)
1938 return -EFBIG;
1939
1940 if (n_blocks_p)
1941 *n_blocks_p = n_blocks;
1942 if (n_metadata_blocks_p)
1943 *n_metadata_blocks_p = offset >> __ffs(block_size);
1944 return 0;
1945 }
1946
1947 static int init_memory(struct dm_writecache *wc)
1948 {
1949 size_t b;
1950 int r;
1951
1952 r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1953 if (r)
1954 return r;
1955
1956 r = writecache_alloc_entries(wc);
1957 if (r)
1958 return r;
1959
1960 for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1961 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1962 pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1963 pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1964 pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1965 pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1966
1967 for (b = 0; b < wc->n_blocks; b++) {
1968 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1969 cond_resched();
1970 }
1971
1972 writecache_flush_all_metadata(wc);
1973 writecache_commit_flushed(wc, false);
1974 pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1975 writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1976 writecache_commit_flushed(wc, false);
1977
1978 return 0;
1979 }
1980
1981 static void writecache_dtr(struct dm_target *ti)
1982 {
1983 struct dm_writecache *wc = ti->private;
1984
1985 if (!wc)
1986 return;
1987
1988 if (wc->endio_thread)
1989 kthread_stop(wc->endio_thread);
1990
1991 if (wc->flush_thread)
1992 kthread_stop(wc->flush_thread);
1993
1994 bioset_exit(&wc->bio_set);
1995
1996 mempool_exit(&wc->copy_pool);
1997
1998 if (wc->writeback_wq)
1999 destroy_workqueue(wc->writeback_wq);
2000
2001 if (wc->dev)
2002 dm_put_device(ti, wc->dev);
2003
2004 if (wc->ssd_dev)
2005 dm_put_device(ti, wc->ssd_dev);
2006
2007 if (wc->entries)
2008 vfree(wc->entries);
2009
2010 if (wc->memory_map) {
2011 if (WC_MODE_PMEM(wc))
2012 persistent_memory_release(wc);
2013 else
2014 vfree(wc->memory_map);
2015 }
2016
2017 if (wc->dm_kcopyd)
2018 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2019
2020 if (wc->dm_io)
2021 dm_io_client_destroy(wc->dm_io);
2022
2023 if (wc->dirty_bitmap)
2024 vfree(wc->dirty_bitmap);
2025
2026 kfree(wc);
2027 }
2028
2029 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2030 {
2031 struct dm_writecache *wc;
2032 struct dm_arg_set as;
2033 const char *string;
2034 unsigned opt_params;
2035 size_t offset, data_size;
2036 int i, r;
2037 char dummy;
2038 int high_wm_percent = HIGH_WATERMARK;
2039 int low_wm_percent = LOW_WATERMARK;
2040 uint64_t x;
2041 struct wc_memory_superblock s;
2042
2043 static struct dm_arg _args[] = {
2044 {0, 16, "Invalid number of feature args"},
2045 };
2046
2047 as.argc = argc;
2048 as.argv = argv;
2049
2050 wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2051 if (!wc) {
2052 ti->error = "Cannot allocate writecache structure";
2053 r = -ENOMEM;
2054 goto bad;
2055 }
2056 ti->private = wc;
2057 wc->ti = ti;
2058
2059 mutex_init(&wc->lock);
2060 wc->max_age = MAX_AGE_UNSPECIFIED;
2061 writecache_poison_lists(wc);
2062 init_waitqueue_head(&wc->freelist_wait);
2063 timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2064 timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2065
2066 for (i = 0; i < 2; i++) {
2067 atomic_set(&wc->bio_in_progress[i], 0);
2068 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2069 }
2070
2071 wc->dm_io = dm_io_client_create();
2072 if (IS_ERR(wc->dm_io)) {
2073 r = PTR_ERR(wc->dm_io);
2074 ti->error = "Unable to allocate dm-io client";
2075 wc->dm_io = NULL;
2076 goto bad;
2077 }
2078
2079 wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2080 if (!wc->writeback_wq) {
2081 r = -ENOMEM;
2082 ti->error = "Could not allocate writeback workqueue";
2083 goto bad;
2084 }
2085 INIT_WORK(&wc->writeback_work, writecache_writeback);
2086 INIT_WORK(&wc->flush_work, writecache_flush_work);
2087
2088 raw_spin_lock_init(&wc->endio_list_lock);
2089 INIT_LIST_HEAD(&wc->endio_list);
2090 wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2091 if (IS_ERR(wc->endio_thread)) {
2092 r = PTR_ERR(wc->endio_thread);
2093 wc->endio_thread = NULL;
2094 ti->error = "Couldn't spawn endio thread";
2095 goto bad;
2096 }
2097 wake_up_process(wc->endio_thread);
2098
2099 /*
2100 * Parse the mode (pmem or ssd)
2101 */
2102 string = dm_shift_arg(&as);
2103 if (!string)
2104 goto bad_arguments;
2105
2106 if (!strcasecmp(string, "s")) {
2107 wc->pmem_mode = false;
2108 } else if (!strcasecmp(string, "p")) {
2109 #ifdef DM_WRITECACHE_HAS_PMEM
2110 wc->pmem_mode = true;
2111 wc->writeback_fua = true;
2112 #else
2113 /*
2114 * If the architecture doesn't support persistent memory or
2115 * the kernel doesn't support any DAX drivers, this driver can
2116 * only be used in SSD-only mode.
2117 */
2118 r = -EOPNOTSUPP;
2119 ti->error = "Persistent memory or DAX not supported on this system";
2120 goto bad;
2121 #endif
2122 } else {
2123 goto bad_arguments;
2124 }
2125
2126 if (WC_MODE_PMEM(wc)) {
2127 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2128 offsetof(struct writeback_struct, bio),
2129 BIOSET_NEED_BVECS);
2130 if (r) {
2131 ti->error = "Could not allocate bio set";
2132 goto bad;
2133 }
2134 } else {
2135 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2136 if (r) {
2137 ti->error = "Could not allocate mempool";
2138 goto bad;
2139 }
2140 }
2141
2142 /*
2143 * Parse the origin data device
2144 */
2145 string = dm_shift_arg(&as);
2146 if (!string)
2147 goto bad_arguments;
2148 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2149 if (r) {
2150 ti->error = "Origin data device lookup failed";
2151 goto bad;
2152 }
2153
2154 /*
2155 * Parse cache data device (be it pmem or ssd)
2156 */
2157 string = dm_shift_arg(&as);
2158 if (!string)
2159 goto bad_arguments;
2160
2161 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2162 if (r) {
2163 ti->error = "Cache data device lookup failed";
2164 goto bad;
2165 }
2166 wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2167
2168 /*
2169 * Parse the cache block size
2170 */
2171 string = dm_shift_arg(&as);
2172 if (!string)
2173 goto bad_arguments;
2174 if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2175 wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2176 (wc->block_size & (wc->block_size - 1))) {
2177 r = -EINVAL;
2178 ti->error = "Invalid block size";
2179 goto bad;
2180 }
2181 if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2182 wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2183 r = -EINVAL;
2184 ti->error = "Block size is smaller than device logical block size";
2185 goto bad;
2186 }
2187 wc->block_size_bits = __ffs(wc->block_size);
2188
2189 wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2190 wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2191 wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2192
2193 /*
2194 * Parse optional arguments
2195 */
2196 r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2197 if (r)
2198 goto bad;
2199
2200 while (opt_params) {
2201 string = dm_shift_arg(&as), opt_params--;
2202 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2203 unsigned long long start_sector;
2204 string = dm_shift_arg(&as), opt_params--;
2205 if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2206 goto invalid_optional;
2207 wc->start_sector = start_sector;
2208 if (wc->start_sector != start_sector ||
2209 wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2210 goto invalid_optional;
2211 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2212 string = dm_shift_arg(&as), opt_params--;
2213 if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2214 goto invalid_optional;
2215 if (high_wm_percent < 0 || high_wm_percent > 100)
2216 goto invalid_optional;
2217 wc->high_wm_percent_set = true;
2218 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2219 string = dm_shift_arg(&as), opt_params--;
2220 if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2221 goto invalid_optional;
2222 if (low_wm_percent < 0 || low_wm_percent > 100)
2223 goto invalid_optional;
2224 wc->low_wm_percent_set = true;
2225 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2226 string = dm_shift_arg(&as), opt_params--;
2227 if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2228 goto invalid_optional;
2229 wc->max_writeback_jobs_set = true;
2230 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2231 string = dm_shift_arg(&as), opt_params--;
2232 if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2233 goto invalid_optional;
2234 wc->autocommit_blocks_set = true;
2235 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2236 unsigned autocommit_msecs;
2237 string = dm_shift_arg(&as), opt_params--;
2238 if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2239 goto invalid_optional;
2240 if (autocommit_msecs > 3600000)
2241 goto invalid_optional;
2242 wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2243 wc->autocommit_time_set = true;
2244 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2245 unsigned max_age_msecs;
2246 string = dm_shift_arg(&as), opt_params--;
2247 if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2248 goto invalid_optional;
2249 if (max_age_msecs > 86400000)
2250 goto invalid_optional;
2251 wc->max_age = msecs_to_jiffies(max_age_msecs);
2252 } else if (!strcasecmp(string, "cleaner")) {
2253 wc->cleaner = true;
2254 } else if (!strcasecmp(string, "fua")) {
2255 if (WC_MODE_PMEM(wc)) {
2256 wc->writeback_fua = true;
2257 wc->writeback_fua_set = true;
2258 } else goto invalid_optional;
2259 } else if (!strcasecmp(string, "nofua")) {
2260 if (WC_MODE_PMEM(wc)) {
2261 wc->writeback_fua = false;
2262 wc->writeback_fua_set = true;
2263 } else goto invalid_optional;
2264 } else {
2265 invalid_optional:
2266 r = -EINVAL;
2267 ti->error = "Invalid optional argument";
2268 goto bad;
2269 }
2270 }
2271
2272 if (high_wm_percent < low_wm_percent) {
2273 r = -EINVAL;
2274 ti->error = "High watermark must be greater than or equal to low watermark";
2275 goto bad;
2276 }
2277
2278 if (WC_MODE_PMEM(wc)) {
2279 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2280 r = -EOPNOTSUPP;
2281 ti->error = "Asynchronous persistent memory not supported as pmem cache";
2282 goto bad;
2283 }
2284
2285 r = persistent_memory_claim(wc);
2286 if (r) {
2287 ti->error = "Unable to map persistent memory for cache";
2288 goto bad;
2289 }
2290 } else {
2291 size_t n_blocks, n_metadata_blocks;
2292 uint64_t n_bitmap_bits;
2293
2294 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2295
2296 bio_list_init(&wc->flush_list);
2297 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2298 if (IS_ERR(wc->flush_thread)) {
2299 r = PTR_ERR(wc->flush_thread);
2300 wc->flush_thread = NULL;
2301 ti->error = "Couldn't spawn flush thread";
2302 goto bad;
2303 }
2304 wake_up_process(wc->flush_thread);
2305
2306 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2307 &n_blocks, &n_metadata_blocks);
2308 if (r) {
2309 ti->error = "Invalid device size";
2310 goto bad;
2311 }
2312
2313 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2314 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2315 /* this is limitation of test_bit functions */
2316 if (n_bitmap_bits > 1U << 31) {
2317 r = -EFBIG;
2318 ti->error = "Invalid device size";
2319 goto bad;
2320 }
2321
2322 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2323 if (!wc->memory_map) {
2324 r = -ENOMEM;
2325 ti->error = "Unable to allocate memory for metadata";
2326 goto bad;
2327 }
2328
2329 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2330 if (IS_ERR(wc->dm_kcopyd)) {
2331 r = PTR_ERR(wc->dm_kcopyd);
2332 ti->error = "Unable to allocate dm-kcopyd client";
2333 wc->dm_kcopyd = NULL;
2334 goto bad;
2335 }
2336
2337 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2338 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2339 BITS_PER_LONG * sizeof(unsigned long);
2340 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2341 if (!wc->dirty_bitmap) {
2342 r = -ENOMEM;
2343 ti->error = "Unable to allocate dirty bitmap";
2344 goto bad;
2345 }
2346
2347 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2348 if (r) {
2349 ti->error = "Unable to read first block of metadata";
2350 goto bad;
2351 }
2352 }
2353
2354 r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2355 if (r) {
2356 ti->error = "Hardware memory error when reading superblock";
2357 goto bad;
2358 }
2359 if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2360 r = init_memory(wc);
2361 if (r) {
2362 ti->error = "Unable to initialize device";
2363 goto bad;
2364 }
2365 r = copy_mc_to_kernel(&s, sb(wc),
2366 sizeof(struct wc_memory_superblock));
2367 if (r) {
2368 ti->error = "Hardware memory error when reading superblock";
2369 goto bad;
2370 }
2371 }
2372
2373 if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2374 ti->error = "Invalid magic in the superblock";
2375 r = -EINVAL;
2376 goto bad;
2377 }
2378
2379 if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2380 ti->error = "Invalid version in the superblock";
2381 r = -EINVAL;
2382 goto bad;
2383 }
2384
2385 if (le32_to_cpu(s.block_size) != wc->block_size) {
2386 ti->error = "Block size does not match superblock";
2387 r = -EINVAL;
2388 goto bad;
2389 }
2390
2391 wc->n_blocks = le64_to_cpu(s.n_blocks);
2392
2393 offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2394 if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2395 overflow:
2396 ti->error = "Overflow in size calculation";
2397 r = -EINVAL;
2398 goto bad;
2399 }
2400 offset += sizeof(struct wc_memory_superblock);
2401 if (offset < sizeof(struct wc_memory_superblock))
2402 goto overflow;
2403 offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2404 data_size = wc->n_blocks * (size_t)wc->block_size;
2405 if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2406 (offset + data_size < offset))
2407 goto overflow;
2408 if (offset + data_size > wc->memory_map_size) {
2409 ti->error = "Memory area is too small";
2410 r = -EINVAL;
2411 goto bad;
2412 }
2413
2414 wc->metadata_sectors = offset >> SECTOR_SHIFT;
2415 wc->block_start = (char *)sb(wc) + offset;
2416
2417 x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2418 x += 50;
2419 do_div(x, 100);
2420 wc->freelist_high_watermark = x;
2421 x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2422 x += 50;
2423 do_div(x, 100);
2424 wc->freelist_low_watermark = x;
2425
2426 if (wc->cleaner)
2427 activate_cleaner(wc);
2428
2429 r = writecache_alloc_entries(wc);
2430 if (r) {
2431 ti->error = "Cannot allocate memory";
2432 goto bad;
2433 }
2434
2435 ti->num_flush_bios = 1;
2436 ti->flush_supported = true;
2437 ti->num_discard_bios = 1;
2438
2439 if (WC_MODE_PMEM(wc))
2440 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2441
2442 return 0;
2443
2444 bad_arguments:
2445 r = -EINVAL;
2446 ti->error = "Bad arguments";
2447 bad:
2448 writecache_dtr(ti);
2449 return r;
2450 }
2451
2452 static void writecache_status(struct dm_target *ti, status_type_t type,
2453 unsigned status_flags, char *result, unsigned maxlen)
2454 {
2455 struct dm_writecache *wc = ti->private;
2456 unsigned extra_args;
2457 unsigned sz = 0;
2458 uint64_t x;
2459
2460 switch (type) {
2461 case STATUSTYPE_INFO:
2462 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2463 (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2464 (unsigned long long)wc->writeback_size);
2465 break;
2466 case STATUSTYPE_TABLE:
2467 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2468 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2469 extra_args = 0;
2470 if (wc->start_sector)
2471 extra_args += 2;
2472 if (wc->high_wm_percent_set && !wc->cleaner)
2473 extra_args += 2;
2474 if (wc->low_wm_percent_set && !wc->cleaner)
2475 extra_args += 2;
2476 if (wc->max_writeback_jobs_set)
2477 extra_args += 2;
2478 if (wc->autocommit_blocks_set)
2479 extra_args += 2;
2480 if (wc->autocommit_time_set)
2481 extra_args += 2;
2482 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2483 extra_args += 2;
2484 if (wc->cleaner)
2485 extra_args++;
2486 if (wc->writeback_fua_set)
2487 extra_args++;
2488
2489 DMEMIT("%u", extra_args);
2490 if (wc->start_sector)
2491 DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2492 if (wc->high_wm_percent_set && !wc->cleaner) {
2493 x = (uint64_t)wc->freelist_high_watermark * 100;
2494 x += wc->n_blocks / 2;
2495 do_div(x, (size_t)wc->n_blocks);
2496 DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2497 }
2498 if (wc->low_wm_percent_set && !wc->cleaner) {
2499 x = (uint64_t)wc->freelist_low_watermark * 100;
2500 x += wc->n_blocks / 2;
2501 do_div(x, (size_t)wc->n_blocks);
2502 DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2503 }
2504 if (wc->max_writeback_jobs_set)
2505 DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2506 if (wc->autocommit_blocks_set)
2507 DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2508 if (wc->autocommit_time_set)
2509 DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2510 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2511 DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2512 if (wc->cleaner)
2513 DMEMIT(" cleaner");
2514 if (wc->writeback_fua_set)
2515 DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2516 break;
2517 }
2518 }
2519
2520 static struct target_type writecache_target = {
2521 .name = "writecache",
2522 .version = {1, 3, 0},
2523 .module = THIS_MODULE,
2524 .ctr = writecache_ctr,
2525 .dtr = writecache_dtr,
2526 .status = writecache_status,
2527 .postsuspend = writecache_suspend,
2528 .resume = writecache_resume,
2529 .message = writecache_message,
2530 .map = writecache_map,
2531 .end_io = writecache_end_io,
2532 .iterate_devices = writecache_iterate_devices,
2533 .io_hints = writecache_io_hints,
2534 };
2535
2536 static int __init dm_writecache_init(void)
2537 {
2538 int r;
2539
2540 r = dm_register_target(&writecache_target);
2541 if (r < 0) {
2542 DMERR("register failed %d", r);
2543 return r;
2544 }
2545
2546 return 0;
2547 }
2548
2549 static void __exit dm_writecache_exit(void)
2550 {
2551 dm_unregister_target(&writecache_target);
2552 }
2553
2554 module_init(dm_writecache_init);
2555 module_exit(dm_writecache_exit);
2556
2557 MODULE_DESCRIPTION(DM_NAME " writecache target");
2558 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2559 MODULE_LICENSE("GPL");
Linux with added FPC-III support