?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dm writecache: add cond_resched to loop in persistent_memory_claim()
[linux/fpc-iii.git] / drivers / md / dm-writecache.c
blob5cc94f57421caf91f4ac3bd8215b77ebecc5b491
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 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
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
235 wc->memory_vmapped = false;
236
237 if (!wc->ssd_dev->dax_dev) {
238 r = -EOPNOTSUPP;
239 goto err1;
240 }
241 s = wc->memory_map_size;
242 p = s >> PAGE_SHIFT;
243 if (!p) {
244 r = -EINVAL;
245 goto err1;
246 }
247 if (p != s >> PAGE_SHIFT) {
248 r = -EOVERFLOW;
249 goto err1;
250 }
251
252 id = dax_read_lock();
253
254 da = dax_direct_access(wc->ssd_dev->dax_dev, 0, p, &wc->memory_map, &pfn);
255 if (da < 0) {
256 wc->memory_map = NULL;
257 r = da;
258 goto err2;
259 }
260 if (!pfn_t_has_page(pfn)) {
261 wc->memory_map = NULL;
262 r = -EOPNOTSUPP;
263 goto err2;
264 }
265 if (da != p) {
266 long i;
267 wc->memory_map = NULL;
268 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
269 if (!pages) {
270 r = -ENOMEM;
271 goto err2;
272 }
273 i = 0;
274 do {
275 long daa;
276 daa = dax_direct_access(wc->ssd_dev->dax_dev, i, p - i,
277 NULL, &pfn);
278 if (daa <= 0) {
279 r = daa ? daa : -EINVAL;
280 goto err3;
281 }
282 if (!pfn_t_has_page(pfn)) {
283 r = -EOPNOTSUPP;
284 goto err3;
285 }
286 while (daa-- && i < p) {
287 pages[i++] = pfn_t_to_page(pfn);
288 pfn.val++;
289 if (!(i & 15))
290 cond_resched();
291 }
292 } while (i < p);
293 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
294 if (!wc->memory_map) {
295 r = -ENOMEM;
296 goto err3;
297 }
298 kvfree(pages);
299 wc->memory_vmapped = true;
300 }
301
302 dax_read_unlock(id);
303
304 wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
305 wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
306
307 return 0;
308 err3:
309 kvfree(pages);
310 err2:
311 dax_read_unlock(id);
312 err1:
313 return r;
314 }
315 #else
316 static int persistent_memory_claim(struct dm_writecache *wc)
317 {
318 BUG();
319 }
320 #endif
321
322 static void persistent_memory_release(struct dm_writecache *wc)
323 {
324 if (wc->memory_vmapped)
325 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
326 }
327
328 static struct page *persistent_memory_page(void *addr)
329 {
330 if (is_vmalloc_addr(addr))
331 return vmalloc_to_page(addr);
332 else
333 return virt_to_page(addr);
334 }
335
336 static unsigned persistent_memory_page_offset(void *addr)
337 {
338 return (unsigned long)addr & (PAGE_SIZE - 1);
339 }
340
341 static void persistent_memory_flush_cache(void *ptr, size_t size)
342 {
343 if (is_vmalloc_addr(ptr))
344 flush_kernel_vmap_range(ptr, size);
345 }
346
347 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
348 {
349 if (is_vmalloc_addr(ptr))
350 invalidate_kernel_vmap_range(ptr, size);
351 }
352
353 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
354 {
355 return wc->memory_map;
356 }
357
358 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
359 {
360 return &sb(wc)->entries[e->index];
361 }
362
363 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
364 {
365 return (char *)wc->block_start + (e->index << wc->block_size_bits);
366 }
367
368 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
369 {
370 return wc->start_sector + wc->metadata_sectors +
371 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
372 }
373
374 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
375 {
376 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
377 return e->original_sector;
378 #else
379 return le64_to_cpu(memory_entry(wc, e)->original_sector);
380 #endif
381 }
382
383 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
384 {
385 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
386 return e->seq_count;
387 #else
388 return le64_to_cpu(memory_entry(wc, e)->seq_count);
389 #endif
390 }
391
392 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
393 {
394 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
395 e->seq_count = -1;
396 #endif
397 pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
398 }
399
400 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
401 uint64_t original_sector, uint64_t seq_count)
402 {
403 struct wc_memory_entry me;
404 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
405 e->original_sector = original_sector;
406 e->seq_count = seq_count;
407 #endif
408 me.original_sector = cpu_to_le64(original_sector);
409 me.seq_count = cpu_to_le64(seq_count);
410 pmem_assign(*memory_entry(wc, e), me);
411 }
412
413 #define writecache_error(wc, err, msg, arg...) \
414 do { \
415 if (!cmpxchg(&(wc)->error, 0, err)) \
416 DMERR(msg, ##arg); \
417 wake_up(&(wc)->freelist_wait); \
418 } while (0)
419
420 #define writecache_has_error(wc) (unlikely(READ_ONCE((wc)->error)))
421
422 static void writecache_flush_all_metadata(struct dm_writecache *wc)
423 {
424 if (!WC_MODE_PMEM(wc))
425 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
426 }
427
428 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
429 {
430 if (!WC_MODE_PMEM(wc))
431 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
432 wc->dirty_bitmap);
433 }
434
435 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
436
437 struct io_notify {
438 struct dm_writecache *wc;
439 struct completion c;
440 atomic_t count;
441 };
442
443 static void writecache_notify_io(unsigned long error, void *context)
444 {
445 struct io_notify *endio = context;
446
447 if (unlikely(error != 0))
448 writecache_error(endio->wc, -EIO, "error writing metadata");
449 BUG_ON(atomic_read(&endio->count) <= 0);
450 if (atomic_dec_and_test(&endio->count))
451 complete(&endio->c);
452 }
453
454 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
455 {
456 wait_event(wc->bio_in_progress_wait[direction],
457 !atomic_read(&wc->bio_in_progress[direction]));
458 }
459
460 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
461 {
462 struct dm_io_region region;
463 struct dm_io_request req;
464 struct io_notify endio = {
465 wc,
466 COMPLETION_INITIALIZER_ONSTACK(endio.c),
467 ATOMIC_INIT(1),
468 };
469 unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
470 unsigned i = 0;
471
472 while (1) {
473 unsigned j;
474 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
475 if (unlikely(i == bitmap_bits))
476 break;
477 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
478
479 region.bdev = wc->ssd_dev->bdev;
480 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
481 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
482
483 if (unlikely(region.sector >= wc->metadata_sectors))
484 break;
485 if (unlikely(region.sector + region.count > wc->metadata_sectors))
486 region.count = wc->metadata_sectors - region.sector;
487
488 region.sector += wc->start_sector;
489 atomic_inc(&endio.count);
490 req.bi_op = REQ_OP_WRITE;
491 req.bi_op_flags = REQ_SYNC;
492 req.mem.type = DM_IO_VMA;
493 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
494 req.client = wc->dm_io;
495 req.notify.fn = writecache_notify_io;
496 req.notify.context = &endio;
497
498 /* writing via async dm-io (implied by notify.fn above) won't return an error */
499 (void) dm_io(&req, 1, &region, NULL);
500 i = j;
501 }
502
503 writecache_notify_io(0, &endio);
504 wait_for_completion_io(&endio.c);
505
506 if (wait_for_ios)
507 writecache_wait_for_ios(wc, WRITE);
508
509 writecache_disk_flush(wc, wc->ssd_dev);
510
511 memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
512 }
513
514 static void ssd_commit_superblock(struct dm_writecache *wc)
515 {
516 int r;
517 struct dm_io_region region;
518 struct dm_io_request req;
519
520 region.bdev = wc->ssd_dev->bdev;
521 region.sector = 0;
522 region.count = PAGE_SIZE;
523
524 if (unlikely(region.sector + region.count > wc->metadata_sectors))
525 region.count = wc->metadata_sectors - region.sector;
526
527 region.sector += wc->start_sector;
528
529 req.bi_op = REQ_OP_WRITE;
530 req.bi_op_flags = REQ_SYNC | REQ_FUA;
531 req.mem.type = DM_IO_VMA;
532 req.mem.ptr.vma = (char *)wc->memory_map;
533 req.client = wc->dm_io;
534 req.notify.fn = NULL;
535 req.notify.context = NULL;
536
537 r = dm_io(&req, 1, &region, NULL);
538 if (unlikely(r))
539 writecache_error(wc, r, "error writing superblock");
540 }
541
542 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
543 {
544 if (WC_MODE_PMEM(wc))
545 wmb();
546 else
547 ssd_commit_flushed(wc, wait_for_ios);
548 }
549
550 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
551 {
552 int r;
553 struct dm_io_region region;
554 struct dm_io_request req;
555
556 region.bdev = dev->bdev;
557 region.sector = 0;
558 region.count = 0;
559 req.bi_op = REQ_OP_WRITE;
560 req.bi_op_flags = REQ_PREFLUSH;
561 req.mem.type = DM_IO_KMEM;
562 req.mem.ptr.addr = NULL;
563 req.client = wc->dm_io;
564 req.notify.fn = NULL;
565
566 r = dm_io(&req, 1, &region, NULL);
567 if (unlikely(r))
568 writecache_error(wc, r, "error flushing metadata: %d", r);
569 }
570
571 #define WFE_RETURN_FOLLOWING 1
572 #define WFE_LOWEST_SEQ 2
573
574 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
575 uint64_t block, int flags)
576 {
577 struct wc_entry *e;
578 struct rb_node *node = wc->tree.rb_node;
579
580 if (unlikely(!node))
581 return NULL;
582
583 while (1) {
584 e = container_of(node, struct wc_entry, rb_node);
585 if (read_original_sector(wc, e) == block)
586 break;
587
588 node = (read_original_sector(wc, e) >= block ?
589 e->rb_node.rb_left : e->rb_node.rb_right);
590 if (unlikely(!node)) {
591 if (!(flags & WFE_RETURN_FOLLOWING))
592 return NULL;
593 if (read_original_sector(wc, e) >= block) {
594 return e;
595 } else {
596 node = rb_next(&e->rb_node);
597 if (unlikely(!node))
598 return NULL;
599 e = container_of(node, struct wc_entry, rb_node);
600 return e;
601 }
602 }
603 }
604
605 while (1) {
606 struct wc_entry *e2;
607 if (flags & WFE_LOWEST_SEQ)
608 node = rb_prev(&e->rb_node);
609 else
610 node = rb_next(&e->rb_node);
611 if (unlikely(!node))
612 return e;
613 e2 = container_of(node, struct wc_entry, rb_node);
614 if (read_original_sector(wc, e2) != block)
615 return e;
616 e = e2;
617 }
618 }
619
620 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
621 {
622 struct wc_entry *e;
623 struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
624
625 while (*node) {
626 e = container_of(*node, struct wc_entry, rb_node);
627 parent = &e->rb_node;
628 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
629 node = &parent->rb_left;
630 else
631 node = &parent->rb_right;
632 }
633 rb_link_node(&ins->rb_node, parent, node);
634 rb_insert_color(&ins->rb_node, &wc->tree);
635 list_add(&ins->lru, &wc->lru);
636 ins->age = jiffies;
637 }
638
639 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
640 {
641 list_del(&e->lru);
642 rb_erase(&e->rb_node, &wc->tree);
643 }
644
645 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
646 {
647 if (WC_MODE_SORT_FREELIST(wc)) {
648 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
649 if (unlikely(!*node))
650 wc->current_free = e;
651 while (*node) {
652 parent = *node;
653 if (&e->rb_node < *node)
654 node = &parent->rb_left;
655 else
656 node = &parent->rb_right;
657 }
658 rb_link_node(&e->rb_node, parent, node);
659 rb_insert_color(&e->rb_node, &wc->freetree);
660 } else {
661 list_add_tail(&e->lru, &wc->freelist);
662 }
663 wc->freelist_size++;
664 }
665
666 static inline void writecache_verify_watermark(struct dm_writecache *wc)
667 {
668 if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
669 queue_work(wc->writeback_wq, &wc->writeback_work);
670 }
671
672 static void writecache_max_age_timer(struct timer_list *t)
673 {
674 struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
675
676 if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
677 queue_work(wc->writeback_wq, &wc->writeback_work);
678 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
679 }
680 }
681
682 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
683 {
684 struct wc_entry *e;
685
686 if (WC_MODE_SORT_FREELIST(wc)) {
687 struct rb_node *next;
688 if (unlikely(!wc->current_free))
689 return NULL;
690 e = wc->current_free;
691 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
692 return NULL;
693 next = rb_next(&e->rb_node);
694 rb_erase(&e->rb_node, &wc->freetree);
695 if (unlikely(!next))
696 next = rb_first(&wc->freetree);
697 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
698 } else {
699 if (unlikely(list_empty(&wc->freelist)))
700 return NULL;
701 e = container_of(wc->freelist.next, struct wc_entry, lru);
702 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
703 return NULL;
704 list_del(&e->lru);
705 }
706 wc->freelist_size--;
707
708 writecache_verify_watermark(wc);
709
710 return e;
711 }
712
713 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
714 {
715 writecache_unlink(wc, e);
716 writecache_add_to_freelist(wc, e);
717 clear_seq_count(wc, e);
718 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
719 if (unlikely(waitqueue_active(&wc->freelist_wait)))
720 wake_up(&wc->freelist_wait);
721 }
722
723 static void writecache_wait_on_freelist(struct dm_writecache *wc)
724 {
725 DEFINE_WAIT(wait);
726
727 prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
728 wc_unlock(wc);
729 io_schedule();
730 finish_wait(&wc->freelist_wait, &wait);
731 wc_lock(wc);
732 }
733
734 static void writecache_poison_lists(struct dm_writecache *wc)
735 {
736 /*
737 * Catch incorrect access to these values while the device is suspended.
738 */
739 memset(&wc->tree, -1, sizeof wc->tree);
740 wc->lru.next = LIST_POISON1;
741 wc->lru.prev = LIST_POISON2;
742 wc->freelist.next = LIST_POISON1;
743 wc->freelist.prev = LIST_POISON2;
744 }
745
746 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
747 {
748 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
749 if (WC_MODE_PMEM(wc))
750 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
751 }
752
753 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
754 {
755 return read_seq_count(wc, e) < wc->seq_count;
756 }
757
758 static void writecache_flush(struct dm_writecache *wc)
759 {
760 struct wc_entry *e, *e2;
761 bool need_flush_after_free;
762
763 wc->uncommitted_blocks = 0;
764 del_timer(&wc->autocommit_timer);
765
766 if (list_empty(&wc->lru))
767 return;
768
769 e = container_of(wc->lru.next, struct wc_entry, lru);
770 if (writecache_entry_is_committed(wc, e)) {
771 if (wc->overwrote_committed) {
772 writecache_wait_for_ios(wc, WRITE);
773 writecache_disk_flush(wc, wc->ssd_dev);
774 wc->overwrote_committed = false;
775 }
776 return;
777 }
778 while (1) {
779 writecache_flush_entry(wc, e);
780 if (unlikely(e->lru.next == &wc->lru))
781 break;
782 e2 = container_of(e->lru.next, struct wc_entry, lru);
783 if (writecache_entry_is_committed(wc, e2))
784 break;
785 e = e2;
786 cond_resched();
787 }
788 writecache_commit_flushed(wc, true);
789
790 wc->seq_count++;
791 pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
792 if (WC_MODE_PMEM(wc))
793 writecache_commit_flushed(wc, false);
794 else
795 ssd_commit_superblock(wc);
796
797 wc->overwrote_committed = false;
798
799 need_flush_after_free = false;
800 while (1) {
801 /* Free another committed entry with lower seq-count */
802 struct rb_node *rb_node = rb_prev(&e->rb_node);
803
804 if (rb_node) {
805 e2 = container_of(rb_node, struct wc_entry, rb_node);
806 if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
807 likely(!e2->write_in_progress)) {
808 writecache_free_entry(wc, e2);
809 need_flush_after_free = true;
810 }
811 }
812 if (unlikely(e->lru.prev == &wc->lru))
813 break;
814 e = container_of(e->lru.prev, struct wc_entry, lru);
815 cond_resched();
816 }
817
818 if (need_flush_after_free)
819 writecache_commit_flushed(wc, false);
820 }
821
822 static void writecache_flush_work(struct work_struct *work)
823 {
824 struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
825
826 wc_lock(wc);
827 writecache_flush(wc);
828 wc_unlock(wc);
829 }
830
831 static void writecache_autocommit_timer(struct timer_list *t)
832 {
833 struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
834 if (!writecache_has_error(wc))
835 queue_work(wc->writeback_wq, &wc->flush_work);
836 }
837
838 static void writecache_schedule_autocommit(struct dm_writecache *wc)
839 {
840 if (!timer_pending(&wc->autocommit_timer))
841 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
842 }
843
844 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
845 {
846 struct wc_entry *e;
847 bool discarded_something = false;
848
849 e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
850 if (unlikely(!e))
851 return;
852
853 while (read_original_sector(wc, e) < end) {
854 struct rb_node *node = rb_next(&e->rb_node);
855
856 if (likely(!e->write_in_progress)) {
857 if (!discarded_something) {
858 writecache_wait_for_ios(wc, READ);
859 writecache_wait_for_ios(wc, WRITE);
860 discarded_something = true;
861 }
862 if (!writecache_entry_is_committed(wc, e))
863 wc->uncommitted_blocks--;
864 writecache_free_entry(wc, e);
865 }
866
867 if (unlikely(!node))
868 break;
869
870 e = container_of(node, struct wc_entry, rb_node);
871 }
872
873 if (discarded_something)
874 writecache_commit_flushed(wc, false);
875 }
876
877 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
878 {
879 if (wc->writeback_size) {
880 writecache_wait_on_freelist(wc);
881 return true;
882 }
883 return false;
884 }
885
886 static void writecache_suspend(struct dm_target *ti)
887 {
888 struct dm_writecache *wc = ti->private;
889 bool flush_on_suspend;
890
891 del_timer_sync(&wc->autocommit_timer);
892 del_timer_sync(&wc->max_age_timer);
893
894 wc_lock(wc);
895 writecache_flush(wc);
896 flush_on_suspend = wc->flush_on_suspend;
897 if (flush_on_suspend) {
898 wc->flush_on_suspend = false;
899 wc->writeback_all++;
900 queue_work(wc->writeback_wq, &wc->writeback_work);
901 }
902 wc_unlock(wc);
903
904 drain_workqueue(wc->writeback_wq);
905
906 wc_lock(wc);
907 if (flush_on_suspend)
908 wc->writeback_all--;
909 while (writecache_wait_for_writeback(wc));
910
911 if (WC_MODE_PMEM(wc))
912 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
913
914 writecache_poison_lists(wc);
915
916 wc_unlock(wc);
917 }
918
919 static int writecache_alloc_entries(struct dm_writecache *wc)
920 {
921 size_t b;
922
923 if (wc->entries)
924 return 0;
925 wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
926 if (!wc->entries)
927 return -ENOMEM;
928 for (b = 0; b < wc->n_blocks; b++) {
929 struct wc_entry *e = &wc->entries[b];
930 e->index = b;
931 e->write_in_progress = false;
932 cond_resched();
933 }
934
935 return 0;
936 }
937
938 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
939 {
940 struct dm_io_region region;
941 struct dm_io_request req;
942
943 region.bdev = wc->ssd_dev->bdev;
944 region.sector = wc->start_sector;
945 region.count = n_sectors;
946 req.bi_op = REQ_OP_READ;
947 req.bi_op_flags = REQ_SYNC;
948 req.mem.type = DM_IO_VMA;
949 req.mem.ptr.vma = (char *)wc->memory_map;
950 req.client = wc->dm_io;
951 req.notify.fn = NULL;
952
953 return dm_io(&req, 1, &region, NULL);
954 }
955
956 static void writecache_resume(struct dm_target *ti)
957 {
958 struct dm_writecache *wc = ti->private;
959 size_t b;
960 bool need_flush = false;
961 __le64 sb_seq_count;
962 int r;
963
964 wc_lock(wc);
965
966 if (WC_MODE_PMEM(wc)) {
967 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
968 } else {
969 r = writecache_read_metadata(wc, wc->metadata_sectors);
970 if (r) {
971 size_t sb_entries_offset;
972 writecache_error(wc, r, "unable to read metadata: %d", r);
973 sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
974 memset((char *)wc->memory_map + sb_entries_offset, -1,
975 (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
976 }
977 }
978
979 wc->tree = RB_ROOT;
980 INIT_LIST_HEAD(&wc->lru);
981 if (WC_MODE_SORT_FREELIST(wc)) {
982 wc->freetree = RB_ROOT;
983 wc->current_free = NULL;
984 } else {
985 INIT_LIST_HEAD(&wc->freelist);
986 }
987 wc->freelist_size = 0;
988
989 r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
990 if (r) {
991 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
992 sb_seq_count = cpu_to_le64(0);
993 }
994 wc->seq_count = le64_to_cpu(sb_seq_count);
995
996 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
997 for (b = 0; b < wc->n_blocks; b++) {
998 struct wc_entry *e = &wc->entries[b];
999 struct wc_memory_entry wme;
1000 if (writecache_has_error(wc)) {
1001 e->original_sector = -1;
1002 e->seq_count = -1;
1003 continue;
1004 }
1005 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
1006 if (r) {
1007 writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1008 (unsigned long)b, r);
1009 e->original_sector = -1;
1010 e->seq_count = -1;
1011 } else {
1012 e->original_sector = le64_to_cpu(wme.original_sector);
1013 e->seq_count = le64_to_cpu(wme.seq_count);
1014 }
1015 cond_resched();
1016 }
1017 #endif
1018 for (b = 0; b < wc->n_blocks; b++) {
1019 struct wc_entry *e = &wc->entries[b];
1020 if (!writecache_entry_is_committed(wc, e)) {
1021 if (read_seq_count(wc, e) != -1) {
1022 erase_this:
1023 clear_seq_count(wc, e);
1024 need_flush = true;
1025 }
1026 writecache_add_to_freelist(wc, e);
1027 } else {
1028 struct wc_entry *old;
1029
1030 old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1031 if (!old) {
1032 writecache_insert_entry(wc, e);
1033 } else {
1034 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1035 writecache_error(wc, -EINVAL,
1036 "two identical entries, position %llu, sector %llu, sequence %llu",
1037 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1038 (unsigned long long)read_seq_count(wc, e));
1039 }
1040 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1041 goto erase_this;
1042 } else {
1043 writecache_free_entry(wc, old);
1044 writecache_insert_entry(wc, e);
1045 need_flush = true;
1046 }
1047 }
1048 }
1049 cond_resched();
1050 }
1051
1052 if (need_flush) {
1053 writecache_flush_all_metadata(wc);
1054 writecache_commit_flushed(wc, false);
1055 }
1056
1057 writecache_verify_watermark(wc);
1058
1059 if (wc->max_age != MAX_AGE_UNSPECIFIED)
1060 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1061
1062 wc_unlock(wc);
1063 }
1064
1065 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1066 {
1067 if (argc != 1)
1068 return -EINVAL;
1069
1070 wc_lock(wc);
1071 if (dm_suspended(wc->ti)) {
1072 wc_unlock(wc);
1073 return -EBUSY;
1074 }
1075 if (writecache_has_error(wc)) {
1076 wc_unlock(wc);
1077 return -EIO;
1078 }
1079
1080 writecache_flush(wc);
1081 wc->writeback_all++;
1082 queue_work(wc->writeback_wq, &wc->writeback_work);
1083 wc_unlock(wc);
1084
1085 flush_workqueue(wc->writeback_wq);
1086
1087 wc_lock(wc);
1088 wc->writeback_all--;
1089 if (writecache_has_error(wc)) {
1090 wc_unlock(wc);
1091 return -EIO;
1092 }
1093 wc_unlock(wc);
1094
1095 return 0;
1096 }
1097
1098 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1099 {
1100 if (argc != 1)
1101 return -EINVAL;
1102
1103 wc_lock(wc);
1104 wc->flush_on_suspend = true;
1105 wc_unlock(wc);
1106
1107 return 0;
1108 }
1109
1110 static void activate_cleaner(struct dm_writecache *wc)
1111 {
1112 wc->flush_on_suspend = true;
1113 wc->cleaner = true;
1114 wc->freelist_high_watermark = wc->n_blocks;
1115 wc->freelist_low_watermark = wc->n_blocks;
1116 }
1117
1118 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1119 {
1120 if (argc != 1)
1121 return -EINVAL;
1122
1123 wc_lock(wc);
1124 activate_cleaner(wc);
1125 if (!dm_suspended(wc->ti))
1126 writecache_verify_watermark(wc);
1127 wc_unlock(wc);
1128
1129 return 0;
1130 }
1131
1132 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1133 char *result, unsigned maxlen)
1134 {
1135 int r = -EINVAL;
1136 struct dm_writecache *wc = ti->private;
1137
1138 if (!strcasecmp(argv[0], "flush"))
1139 r = process_flush_mesg(argc, argv, wc);
1140 else if (!strcasecmp(argv[0], "flush_on_suspend"))
1141 r = process_flush_on_suspend_mesg(argc, argv, wc);
1142 else if (!strcasecmp(argv[0], "cleaner"))
1143 r = process_cleaner_mesg(argc, argv, wc);
1144 else
1145 DMERR("unrecognised message received: %s", argv[0]);
1146
1147 return r;
1148 }
1149
1150 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1151 {
1152 void *buf;
1153 unsigned long flags;
1154 unsigned size;
1155 int rw = bio_data_dir(bio);
1156 unsigned remaining_size = wc->block_size;
1157
1158 do {
1159 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1160 buf = bvec_kmap_irq(&bv, &flags);
1161 size = bv.bv_len;
1162 if (unlikely(size > remaining_size))
1163 size = remaining_size;
1164
1165 if (rw == READ) {
1166 int r;
1167 r = memcpy_mcsafe(buf, data, size);
1168 flush_dcache_page(bio_page(bio));
1169 if (unlikely(r)) {
1170 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1171 bio->bi_status = BLK_STS_IOERR;
1172 }
1173 } else {
1174 flush_dcache_page(bio_page(bio));
1175 memcpy_flushcache(data, buf, size);
1176 }
1177
1178 bvec_kunmap_irq(buf, &flags);
1179
1180 data = (char *)data + size;
1181 remaining_size -= size;
1182 bio_advance(bio, size);
1183 } while (unlikely(remaining_size));
1184 }
1185
1186 static int writecache_flush_thread(void *data)
1187 {
1188 struct dm_writecache *wc = data;
1189
1190 while (1) {
1191 struct bio *bio;
1192
1193 wc_lock(wc);
1194 bio = bio_list_pop(&wc->flush_list);
1195 if (!bio) {
1196 set_current_state(TASK_INTERRUPTIBLE);
1197 wc_unlock(wc);
1198
1199 if (unlikely(kthread_should_stop())) {
1200 set_current_state(TASK_RUNNING);
1201 break;
1202 }
1203
1204 schedule();
1205 continue;
1206 }
1207
1208 if (bio_op(bio) == REQ_OP_DISCARD) {
1209 writecache_discard(wc, bio->bi_iter.bi_sector,
1210 bio_end_sector(bio));
1211 wc_unlock(wc);
1212 bio_set_dev(bio, wc->dev->bdev);
1213 generic_make_request(bio);
1214 } else {
1215 writecache_flush(wc);
1216 wc_unlock(wc);
1217 if (writecache_has_error(wc))
1218 bio->bi_status = BLK_STS_IOERR;
1219 bio_endio(bio);
1220 }
1221 }
1222
1223 return 0;
1224 }
1225
1226 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1227 {
1228 if (bio_list_empty(&wc->flush_list))
1229 wake_up_process(wc->flush_thread);
1230 bio_list_add(&wc->flush_list, bio);
1231 }
1232
1233 static int writecache_map(struct dm_target *ti, struct bio *bio)
1234 {
1235 struct wc_entry *e;
1236 struct dm_writecache *wc = ti->private;
1237
1238 bio->bi_private = NULL;
1239
1240 wc_lock(wc);
1241
1242 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1243 if (writecache_has_error(wc))
1244 goto unlock_error;
1245 if (WC_MODE_PMEM(wc)) {
1246 writecache_flush(wc);
1247 if (writecache_has_error(wc))
1248 goto unlock_error;
1249 goto unlock_submit;
1250 } else {
1251 writecache_offload_bio(wc, bio);
1252 goto unlock_return;
1253 }
1254 }
1255
1256 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1257
1258 if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1259 (wc->block_size / 512 - 1)) != 0)) {
1260 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1261 (unsigned long long)bio->bi_iter.bi_sector,
1262 bio->bi_iter.bi_size, wc->block_size);
1263 goto unlock_error;
1264 }
1265
1266 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1267 if (writecache_has_error(wc))
1268 goto unlock_error;
1269 if (WC_MODE_PMEM(wc)) {
1270 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1271 goto unlock_remap_origin;
1272 } else {
1273 writecache_offload_bio(wc, bio);
1274 goto unlock_return;
1275 }
1276 }
1277
1278 if (bio_data_dir(bio) == READ) {
1279 read_next_block:
1280 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1281 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1282 if (WC_MODE_PMEM(wc)) {
1283 bio_copy_block(wc, bio, memory_data(wc, e));
1284 if (bio->bi_iter.bi_size)
1285 goto read_next_block;
1286 goto unlock_submit;
1287 } else {
1288 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1289 bio_set_dev(bio, wc->ssd_dev->bdev);
1290 bio->bi_iter.bi_sector = cache_sector(wc, e);
1291 if (!writecache_entry_is_committed(wc, e))
1292 writecache_wait_for_ios(wc, WRITE);
1293 goto unlock_remap;
1294 }
1295 } else {
1296 if (e) {
1297 sector_t next_boundary =
1298 read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1299 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1300 dm_accept_partial_bio(bio, next_boundary);
1301 }
1302 }
1303 goto unlock_remap_origin;
1304 }
1305 } else {
1306 do {
1307 bool found_entry = false;
1308 if (writecache_has_error(wc))
1309 goto unlock_error;
1310 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1311 if (e) {
1312 if (!writecache_entry_is_committed(wc, e))
1313 goto bio_copy;
1314 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1315 wc->overwrote_committed = true;
1316 goto bio_copy;
1317 }
1318 found_entry = true;
1319 } else {
1320 if (unlikely(wc->cleaner))
1321 goto direct_write;
1322 }
1323 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1324 if (unlikely(!e)) {
1325 if (!found_entry) {
1326 direct_write:
1327 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1328 if (e) {
1329 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1330 BUG_ON(!next_boundary);
1331 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1332 dm_accept_partial_bio(bio, next_boundary);
1333 }
1334 }
1335 goto unlock_remap_origin;
1336 }
1337 writecache_wait_on_freelist(wc);
1338 continue;
1339 }
1340 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1341 writecache_insert_entry(wc, e);
1342 wc->uncommitted_blocks++;
1343 bio_copy:
1344 if (WC_MODE_PMEM(wc)) {
1345 bio_copy_block(wc, bio, memory_data(wc, e));
1346 } else {
1347 unsigned bio_size = wc->block_size;
1348 sector_t start_cache_sec = cache_sector(wc, e);
1349 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1350
1351 while (bio_size < bio->bi_iter.bi_size) {
1352 struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1353 if (!f)
1354 break;
1355 write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1356 (bio_size >> SECTOR_SHIFT), wc->seq_count);
1357 writecache_insert_entry(wc, f);
1358 wc->uncommitted_blocks++;
1359 bio_size += wc->block_size;
1360 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1361 }
1362
1363 bio_set_dev(bio, wc->ssd_dev->bdev);
1364 bio->bi_iter.bi_sector = start_cache_sec;
1365 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1366
1367 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1368 wc->uncommitted_blocks = 0;
1369 queue_work(wc->writeback_wq, &wc->flush_work);
1370 } else {
1371 writecache_schedule_autocommit(wc);
1372 }
1373 goto unlock_remap;
1374 }
1375 } while (bio->bi_iter.bi_size);
1376
1377 if (unlikely(bio->bi_opf & REQ_FUA ||
1378 wc->uncommitted_blocks >= wc->autocommit_blocks))
1379 writecache_flush(wc);
1380 else
1381 writecache_schedule_autocommit(wc);
1382 goto unlock_submit;
1383 }
1384
1385 unlock_remap_origin:
1386 bio_set_dev(bio, wc->dev->bdev);
1387 wc_unlock(wc);
1388 return DM_MAPIO_REMAPPED;
1389
1390 unlock_remap:
1391 /* make sure that writecache_end_io decrements bio_in_progress: */
1392 bio->bi_private = (void *)1;
1393 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1394 wc_unlock(wc);
1395 return DM_MAPIO_REMAPPED;
1396
1397 unlock_submit:
1398 wc_unlock(wc);
1399 bio_endio(bio);
1400 return DM_MAPIO_SUBMITTED;
1401
1402 unlock_return:
1403 wc_unlock(wc);
1404 return DM_MAPIO_SUBMITTED;
1405
1406 unlock_error:
1407 wc_unlock(wc);
1408 bio_io_error(bio);
1409 return DM_MAPIO_SUBMITTED;
1410 }
1411
1412 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1413 {
1414 struct dm_writecache *wc = ti->private;
1415
1416 if (bio->bi_private != NULL) {
1417 int dir = bio_data_dir(bio);
1418 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1419 if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1420 wake_up(&wc->bio_in_progress_wait[dir]);
1421 }
1422 return 0;
1423 }
1424
1425 static int writecache_iterate_devices(struct dm_target *ti,
1426 iterate_devices_callout_fn fn, void *data)
1427 {
1428 struct dm_writecache *wc = ti->private;
1429
1430 return fn(ti, wc->dev, 0, ti->len, data);
1431 }
1432
1433 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1434 {
1435 struct dm_writecache *wc = ti->private;
1436
1437 if (limits->logical_block_size < wc->block_size)
1438 limits->logical_block_size = wc->block_size;
1439
1440 if (limits->physical_block_size < wc->block_size)
1441 limits->physical_block_size = wc->block_size;
1442
1443 if (limits->io_min < wc->block_size)
1444 limits->io_min = wc->block_size;
1445 }
1446
1447
1448 static void writecache_writeback_endio(struct bio *bio)
1449 {
1450 struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1451 struct dm_writecache *wc = wb->wc;
1452 unsigned long flags;
1453
1454 raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1455 if (unlikely(list_empty(&wc->endio_list)))
1456 wake_up_process(wc->endio_thread);
1457 list_add_tail(&wb->endio_entry, &wc->endio_list);
1458 raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1459 }
1460
1461 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1462 {
1463 struct copy_struct *c = ptr;
1464 struct dm_writecache *wc = c->wc;
1465
1466 c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1467
1468 raw_spin_lock_irq(&wc->endio_list_lock);
1469 if (unlikely(list_empty(&wc->endio_list)))
1470 wake_up_process(wc->endio_thread);
1471 list_add_tail(&c->endio_entry, &wc->endio_list);
1472 raw_spin_unlock_irq(&wc->endio_list_lock);
1473 }
1474
1475 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1476 {
1477 unsigned i;
1478 struct writeback_struct *wb;
1479 struct wc_entry *e;
1480 unsigned long n_walked = 0;
1481
1482 do {
1483 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1484 list_del(&wb->endio_entry);
1485
1486 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1487 writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1488 "write error %d", wb->bio.bi_status);
1489 i = 0;
1490 do {
1491 e = wb->wc_list[i];
1492 BUG_ON(!e->write_in_progress);
1493 e->write_in_progress = false;
1494 INIT_LIST_HEAD(&e->lru);
1495 if (!writecache_has_error(wc))
1496 writecache_free_entry(wc, e);
1497 BUG_ON(!wc->writeback_size);
1498 wc->writeback_size--;
1499 n_walked++;
1500 if (unlikely(n_walked >= ENDIO_LATENCY)) {
1501 writecache_commit_flushed(wc, false);
1502 wc_unlock(wc);
1503 wc_lock(wc);
1504 n_walked = 0;
1505 }
1506 } while (++i < wb->wc_list_n);
1507
1508 if (wb->wc_list != wb->wc_list_inline)
1509 kfree(wb->wc_list);
1510 bio_put(&wb->bio);
1511 } while (!list_empty(list));
1512 }
1513
1514 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1515 {
1516 struct copy_struct *c;
1517 struct wc_entry *e;
1518
1519 do {
1520 c = list_entry(list->next, struct copy_struct, endio_entry);
1521 list_del(&c->endio_entry);
1522
1523 if (unlikely(c->error))
1524 writecache_error(wc, c->error, "copy error");
1525
1526 e = c->e;
1527 do {
1528 BUG_ON(!e->write_in_progress);
1529 e->write_in_progress = false;
1530 INIT_LIST_HEAD(&e->lru);
1531 if (!writecache_has_error(wc))
1532 writecache_free_entry(wc, e);
1533
1534 BUG_ON(!wc->writeback_size);
1535 wc->writeback_size--;
1536 e++;
1537 } while (--c->n_entries);
1538 mempool_free(c, &wc->copy_pool);
1539 } while (!list_empty(list));
1540 }
1541
1542 static int writecache_endio_thread(void *data)
1543 {
1544 struct dm_writecache *wc = data;
1545
1546 while (1) {
1547 struct list_head list;
1548
1549 raw_spin_lock_irq(&wc->endio_list_lock);
1550 if (!list_empty(&wc->endio_list))
1551 goto pop_from_list;
1552 set_current_state(TASK_INTERRUPTIBLE);
1553 raw_spin_unlock_irq(&wc->endio_list_lock);
1554
1555 if (unlikely(kthread_should_stop())) {
1556 set_current_state(TASK_RUNNING);
1557 break;
1558 }
1559
1560 schedule();
1561
1562 continue;
1563
1564 pop_from_list:
1565 list = wc->endio_list;
1566 list.next->prev = list.prev->next = &list;
1567 INIT_LIST_HEAD(&wc->endio_list);
1568 raw_spin_unlock_irq(&wc->endio_list_lock);
1569
1570 if (!WC_MODE_FUA(wc))
1571 writecache_disk_flush(wc, wc->dev);
1572
1573 wc_lock(wc);
1574
1575 if (WC_MODE_PMEM(wc)) {
1576 __writecache_endio_pmem(wc, &list);
1577 } else {
1578 __writecache_endio_ssd(wc, &list);
1579 writecache_wait_for_ios(wc, READ);
1580 }
1581
1582 writecache_commit_flushed(wc, false);
1583
1584 wc_unlock(wc);
1585 }
1586
1587 return 0;
1588 }
1589
1590 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1591 {
1592 struct dm_writecache *wc = wb->wc;
1593 unsigned block_size = wc->block_size;
1594 void *address = memory_data(wc, e);
1595
1596 persistent_memory_flush_cache(address, block_size);
1597 return bio_add_page(&wb->bio, persistent_memory_page(address),
1598 block_size, persistent_memory_page_offset(address)) != 0;
1599 }
1600
1601 struct writeback_list {
1602 struct list_head list;
1603 size_t size;
1604 };
1605
1606 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1607 {
1608 if (unlikely(wc->max_writeback_jobs)) {
1609 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1610 wc_lock(wc);
1611 while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1612 writecache_wait_on_freelist(wc);
1613 wc_unlock(wc);
1614 }
1615 }
1616 cond_resched();
1617 }
1618
1619 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1620 {
1621 struct wc_entry *e, *f;
1622 struct bio *bio;
1623 struct writeback_struct *wb;
1624 unsigned max_pages;
1625
1626 while (wbl->size) {
1627 wbl->size--;
1628 e = container_of(wbl->list.prev, struct wc_entry, lru);
1629 list_del(&e->lru);
1630
1631 max_pages = e->wc_list_contiguous;
1632
1633 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1634 wb = container_of(bio, struct writeback_struct, bio);
1635 wb->wc = wc;
1636 bio->bi_end_io = writecache_writeback_endio;
1637 bio_set_dev(bio, wc->dev->bdev);
1638 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1639 if (max_pages <= WB_LIST_INLINE ||
1640 unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1641 GFP_NOIO | __GFP_NORETRY |
1642 __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1643 wb->wc_list = wb->wc_list_inline;
1644 max_pages = WB_LIST_INLINE;
1645 }
1646
1647 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1648
1649 wb->wc_list[0] = e;
1650 wb->wc_list_n = 1;
1651
1652 while (wbl->size && wb->wc_list_n < max_pages) {
1653 f = container_of(wbl->list.prev, struct wc_entry, lru);
1654 if (read_original_sector(wc, f) !=
1655 read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1656 break;
1657 if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1658 break;
1659 wbl->size--;
1660 list_del(&f->lru);
1661 wb->wc_list[wb->wc_list_n++] = f;
1662 e = f;
1663 }
1664 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1665 if (writecache_has_error(wc)) {
1666 bio->bi_status = BLK_STS_IOERR;
1667 bio_endio(bio);
1668 } else {
1669 submit_bio(bio);
1670 }
1671
1672 __writeback_throttle(wc, wbl);
1673 }
1674 }
1675
1676 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1677 {
1678 struct wc_entry *e, *f;
1679 struct dm_io_region from, to;
1680 struct copy_struct *c;
1681
1682 while (wbl->size) {
1683 unsigned n_sectors;
1684
1685 wbl->size--;
1686 e = container_of(wbl->list.prev, struct wc_entry, lru);
1687 list_del(&e->lru);
1688
1689 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1690
1691 from.bdev = wc->ssd_dev->bdev;
1692 from.sector = cache_sector(wc, e);
1693 from.count = n_sectors;
1694 to.bdev = wc->dev->bdev;
1695 to.sector = read_original_sector(wc, e);
1696 to.count = n_sectors;
1697
1698 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1699 c->wc = wc;
1700 c->e = e;
1701 c->n_entries = e->wc_list_contiguous;
1702
1703 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1704 wbl->size--;
1705 f = container_of(wbl->list.prev, struct wc_entry, lru);
1706 BUG_ON(f != e + 1);
1707 list_del(&f->lru);
1708 e = f;
1709 }
1710
1711 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1712
1713 __writeback_throttle(wc, wbl);
1714 }
1715 }
1716
1717 static void writecache_writeback(struct work_struct *work)
1718 {
1719 struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1720 struct blk_plug plug;
1721 struct wc_entry *f, *uninitialized_var(g), *e = NULL;
1722 struct rb_node *node, *next_node;
1723 struct list_head skipped;
1724 struct writeback_list wbl;
1725 unsigned long n_walked;
1726
1727 wc_lock(wc);
1728 restart:
1729 if (writecache_has_error(wc)) {
1730 wc_unlock(wc);
1731 return;
1732 }
1733
1734 if (unlikely(wc->writeback_all)) {
1735 if (writecache_wait_for_writeback(wc))
1736 goto restart;
1737 }
1738
1739 if (wc->overwrote_committed) {
1740 writecache_wait_for_ios(wc, WRITE);
1741 }
1742
1743 n_walked = 0;
1744 INIT_LIST_HEAD(&skipped);
1745 INIT_LIST_HEAD(&wbl.list);
1746 wbl.size = 0;
1747 while (!list_empty(&wc->lru) &&
1748 (wc->writeback_all ||
1749 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1750 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1751 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1752
1753 n_walked++;
1754 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1755 likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1756 queue_work(wc->writeback_wq, &wc->writeback_work);
1757 break;
1758 }
1759
1760 if (unlikely(wc->writeback_all)) {
1761 if (unlikely(!e)) {
1762 writecache_flush(wc);
1763 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1764 } else
1765 e = g;
1766 } else
1767 e = container_of(wc->lru.prev, struct wc_entry, lru);
1768 BUG_ON(e->write_in_progress);
1769 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1770 writecache_flush(wc);
1771 }
1772 node = rb_prev(&e->rb_node);
1773 if (node) {
1774 f = container_of(node, struct wc_entry, rb_node);
1775 if (unlikely(read_original_sector(wc, f) ==
1776 read_original_sector(wc, e))) {
1777 BUG_ON(!f->write_in_progress);
1778 list_del(&e->lru);
1779 list_add(&e->lru, &skipped);
1780 cond_resched();
1781 continue;
1782 }
1783 }
1784 wc->writeback_size++;
1785 list_del(&e->lru);
1786 list_add(&e->lru, &wbl.list);
1787 wbl.size++;
1788 e->write_in_progress = true;
1789 e->wc_list_contiguous = 1;
1790
1791 f = e;
1792
1793 while (1) {
1794 next_node = rb_next(&f->rb_node);
1795 if (unlikely(!next_node))
1796 break;
1797 g = container_of(next_node, struct wc_entry, rb_node);
1798 if (unlikely(read_original_sector(wc, g) ==
1799 read_original_sector(wc, f))) {
1800 f = g;
1801 continue;
1802 }
1803 if (read_original_sector(wc, g) !=
1804 read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1805 break;
1806 if (unlikely(g->write_in_progress))
1807 break;
1808 if (unlikely(!writecache_entry_is_committed(wc, g)))
1809 break;
1810
1811 if (!WC_MODE_PMEM(wc)) {
1812 if (g != f + 1)
1813 break;
1814 }
1815
1816 n_walked++;
1817 //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1818 // break;
1819
1820 wc->writeback_size++;
1821 list_del(&g->lru);
1822 list_add(&g->lru, &wbl.list);
1823 wbl.size++;
1824 g->write_in_progress = true;
1825 g->wc_list_contiguous = BIO_MAX_PAGES;
1826 f = g;
1827 e->wc_list_contiguous++;
1828 if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1829 if (unlikely(wc->writeback_all)) {
1830 next_node = rb_next(&f->rb_node);
1831 if (likely(next_node))
1832 g = container_of(next_node, struct wc_entry, rb_node);
1833 }
1834 break;
1835 }
1836 }
1837 cond_resched();
1838 }
1839
1840 if (!list_empty(&skipped)) {
1841 list_splice_tail(&skipped, &wc->lru);
1842 /*
1843 * If we didn't do any progress, we must wait until some
1844 * writeback finishes to avoid burning CPU in a loop
1845 */
1846 if (unlikely(!wbl.size))
1847 writecache_wait_for_writeback(wc);
1848 }
1849
1850 wc_unlock(wc);
1851
1852 blk_start_plug(&plug);
1853
1854 if (WC_MODE_PMEM(wc))
1855 __writecache_writeback_pmem(wc, &wbl);
1856 else
1857 __writecache_writeback_ssd(wc, &wbl);
1858
1859 blk_finish_plug(&plug);
1860
1861 if (unlikely(wc->writeback_all)) {
1862 wc_lock(wc);
1863 while (writecache_wait_for_writeback(wc));
1864 wc_unlock(wc);
1865 }
1866 }
1867
1868 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1869 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1870 {
1871 uint64_t n_blocks, offset;
1872 struct wc_entry e;
1873
1874 n_blocks = device_size;
1875 do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1876
1877 while (1) {
1878 if (!n_blocks)
1879 return -ENOSPC;
1880 /* Verify the following entries[n_blocks] won't overflow */
1881 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1882 sizeof(struct wc_memory_entry)))
1883 return -EFBIG;
1884 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1885 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1886 if (offset + n_blocks * block_size <= device_size)
1887 break;
1888 n_blocks--;
1889 }
1890
1891 /* check if the bit field overflows */
1892 e.index = n_blocks;
1893 if (e.index != n_blocks)
1894 return -EFBIG;
1895
1896 if (n_blocks_p)
1897 *n_blocks_p = n_blocks;
1898 if (n_metadata_blocks_p)
1899 *n_metadata_blocks_p = offset >> __ffs(block_size);
1900 return 0;
1901 }
1902
1903 static int init_memory(struct dm_writecache *wc)
1904 {
1905 size_t b;
1906 int r;
1907
1908 r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1909 if (r)
1910 return r;
1911
1912 r = writecache_alloc_entries(wc);
1913 if (r)
1914 return r;
1915
1916 for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1917 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1918 pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1919 pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1920 pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1921 pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1922
1923 for (b = 0; b < wc->n_blocks; b++) {
1924 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1925 cond_resched();
1926 }
1927
1928 writecache_flush_all_metadata(wc);
1929 writecache_commit_flushed(wc, false);
1930 pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1931 writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1932 writecache_commit_flushed(wc, false);
1933
1934 return 0;
1935 }
1936
1937 static void writecache_dtr(struct dm_target *ti)
1938 {
1939 struct dm_writecache *wc = ti->private;
1940
1941 if (!wc)
1942 return;
1943
1944 if (wc->endio_thread)
1945 kthread_stop(wc->endio_thread);
1946
1947 if (wc->flush_thread)
1948 kthread_stop(wc->flush_thread);
1949
1950 bioset_exit(&wc->bio_set);
1951
1952 mempool_exit(&wc->copy_pool);
1953
1954 if (wc->writeback_wq)
1955 destroy_workqueue(wc->writeback_wq);
1956
1957 if (wc->dev)
1958 dm_put_device(ti, wc->dev);
1959
1960 if (wc->ssd_dev)
1961 dm_put_device(ti, wc->ssd_dev);
1962
1963 if (wc->entries)
1964 vfree(wc->entries);
1965
1966 if (wc->memory_map) {
1967 if (WC_MODE_PMEM(wc))
1968 persistent_memory_release(wc);
1969 else
1970 vfree(wc->memory_map);
1971 }
1972
1973 if (wc->dm_kcopyd)
1974 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1975
1976 if (wc->dm_io)
1977 dm_io_client_destroy(wc->dm_io);
1978
1979 if (wc->dirty_bitmap)
1980 vfree(wc->dirty_bitmap);
1981
1982 kfree(wc);
1983 }
1984
1985 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1986 {
1987 struct dm_writecache *wc;
1988 struct dm_arg_set as;
1989 const char *string;
1990 unsigned opt_params;
1991 size_t offset, data_size;
1992 int i, r;
1993 char dummy;
1994 int high_wm_percent = HIGH_WATERMARK;
1995 int low_wm_percent = LOW_WATERMARK;
1996 uint64_t x;
1997 struct wc_memory_superblock s;
1998
1999 static struct dm_arg _args[] = {
2000 {0, 10, "Invalid number of feature args"},
2001 };
2002
2003 as.argc = argc;
2004 as.argv = argv;
2005
2006 wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2007 if (!wc) {
2008 ti->error = "Cannot allocate writecache structure";
2009 r = -ENOMEM;
2010 goto bad;
2011 }
2012 ti->private = wc;
2013 wc->ti = ti;
2014
2015 mutex_init(&wc->lock);
2016 wc->max_age = MAX_AGE_UNSPECIFIED;
2017 writecache_poison_lists(wc);
2018 init_waitqueue_head(&wc->freelist_wait);
2019 timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2020 timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2021
2022 for (i = 0; i < 2; i++) {
2023 atomic_set(&wc->bio_in_progress[i], 0);
2024 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2025 }
2026
2027 wc->dm_io = dm_io_client_create();
2028 if (IS_ERR(wc->dm_io)) {
2029 r = PTR_ERR(wc->dm_io);
2030 ti->error = "Unable to allocate dm-io client";
2031 wc->dm_io = NULL;
2032 goto bad;
2033 }
2034
2035 wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2036 if (!wc->writeback_wq) {
2037 r = -ENOMEM;
2038 ti->error = "Could not allocate writeback workqueue";
2039 goto bad;
2040 }
2041 INIT_WORK(&wc->writeback_work, writecache_writeback);
2042 INIT_WORK(&wc->flush_work, writecache_flush_work);
2043
2044 raw_spin_lock_init(&wc->endio_list_lock);
2045 INIT_LIST_HEAD(&wc->endio_list);
2046 wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2047 if (IS_ERR(wc->endio_thread)) {
2048 r = PTR_ERR(wc->endio_thread);
2049 wc->endio_thread = NULL;
2050 ti->error = "Couldn't spawn endio thread";
2051 goto bad;
2052 }
2053 wake_up_process(wc->endio_thread);
2054
2055 /*
2056 * Parse the mode (pmem or ssd)
2057 */
2058 string = dm_shift_arg(&as);
2059 if (!string)
2060 goto bad_arguments;
2061
2062 if (!strcasecmp(string, "s")) {
2063 wc->pmem_mode = false;
2064 } else if (!strcasecmp(string, "p")) {
2065 #ifdef DM_WRITECACHE_HAS_PMEM
2066 wc->pmem_mode = true;
2067 wc->writeback_fua = true;
2068 #else
2069 /*
2070 * If the architecture doesn't support persistent memory or
2071 * the kernel doesn't support any DAX drivers, this driver can
2072 * only be used in SSD-only mode.
2073 */
2074 r = -EOPNOTSUPP;
2075 ti->error = "Persistent memory or DAX not supported on this system";
2076 goto bad;
2077 #endif
2078 } else {
2079 goto bad_arguments;
2080 }
2081
2082 if (WC_MODE_PMEM(wc)) {
2083 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2084 offsetof(struct writeback_struct, bio),
2085 BIOSET_NEED_BVECS);
2086 if (r) {
2087 ti->error = "Could not allocate bio set";
2088 goto bad;
2089 }
2090 } else {
2091 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2092 if (r) {
2093 ti->error = "Could not allocate mempool";
2094 goto bad;
2095 }
2096 }
2097
2098 /*
2099 * Parse the origin data device
2100 */
2101 string = dm_shift_arg(&as);
2102 if (!string)
2103 goto bad_arguments;
2104 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2105 if (r) {
2106 ti->error = "Origin data device lookup failed";
2107 goto bad;
2108 }
2109
2110 /*
2111 * Parse cache data device (be it pmem or ssd)
2112 */
2113 string = dm_shift_arg(&as);
2114 if (!string)
2115 goto bad_arguments;
2116
2117 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2118 if (r) {
2119 ti->error = "Cache data device lookup failed";
2120 goto bad;
2121 }
2122 wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2123
2124 /*
2125 * Parse the cache block size
2126 */
2127 string = dm_shift_arg(&as);
2128 if (!string)
2129 goto bad_arguments;
2130 if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2131 wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2132 (wc->block_size & (wc->block_size - 1))) {
2133 r = -EINVAL;
2134 ti->error = "Invalid block size";
2135 goto bad;
2136 }
2137 if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2138 wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2139 r = -EINVAL;
2140 ti->error = "Block size is smaller than device logical block size";
2141 goto bad;
2142 }
2143 wc->block_size_bits = __ffs(wc->block_size);
2144
2145 wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2146 wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2147 wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2148
2149 /*
2150 * Parse optional arguments
2151 */
2152 r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2153 if (r)
2154 goto bad;
2155
2156 while (opt_params) {
2157 string = dm_shift_arg(&as), opt_params--;
2158 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2159 unsigned long long start_sector;
2160 string = dm_shift_arg(&as), opt_params--;
2161 if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2162 goto invalid_optional;
2163 wc->start_sector = start_sector;
2164 if (wc->start_sector != start_sector ||
2165 wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2166 goto invalid_optional;
2167 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2168 string = dm_shift_arg(&as), opt_params--;
2169 if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2170 goto invalid_optional;
2171 if (high_wm_percent < 0 || high_wm_percent > 100)
2172 goto invalid_optional;
2173 wc->high_wm_percent_set = true;
2174 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2175 string = dm_shift_arg(&as), opt_params--;
2176 if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2177 goto invalid_optional;
2178 if (low_wm_percent < 0 || low_wm_percent > 100)
2179 goto invalid_optional;
2180 wc->low_wm_percent_set = true;
2181 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2182 string = dm_shift_arg(&as), opt_params--;
2183 if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2184 goto invalid_optional;
2185 wc->max_writeback_jobs_set = true;
2186 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2187 string = dm_shift_arg(&as), opt_params--;
2188 if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2189 goto invalid_optional;
2190 wc->autocommit_blocks_set = true;
2191 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2192 unsigned autocommit_msecs;
2193 string = dm_shift_arg(&as), opt_params--;
2194 if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2195 goto invalid_optional;
2196 if (autocommit_msecs > 3600000)
2197 goto invalid_optional;
2198 wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2199 wc->autocommit_time_set = true;
2200 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2201 unsigned max_age_msecs;
2202 string = dm_shift_arg(&as), opt_params--;
2203 if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2204 goto invalid_optional;
2205 if (max_age_msecs > 86400000)
2206 goto invalid_optional;
2207 wc->max_age = msecs_to_jiffies(max_age_msecs);
2208 } else if (!strcasecmp(string, "cleaner")) {
2209 wc->cleaner = true;
2210 } else if (!strcasecmp(string, "fua")) {
2211 if (WC_MODE_PMEM(wc)) {
2212 wc->writeback_fua = true;
2213 wc->writeback_fua_set = true;
2214 } else goto invalid_optional;
2215 } else if (!strcasecmp(string, "nofua")) {
2216 if (WC_MODE_PMEM(wc)) {
2217 wc->writeback_fua = false;
2218 wc->writeback_fua_set = true;
2219 } else goto invalid_optional;
2220 } else {
2221 invalid_optional:
2222 r = -EINVAL;
2223 ti->error = "Invalid optional argument";
2224 goto bad;
2225 }
2226 }
2227
2228 if (high_wm_percent < low_wm_percent) {
2229 r = -EINVAL;
2230 ti->error = "High watermark must be greater than or equal to low watermark";
2231 goto bad;
2232 }
2233
2234 if (WC_MODE_PMEM(wc)) {
2235 r = persistent_memory_claim(wc);
2236 if (r) {
2237 ti->error = "Unable to map persistent memory for cache";
2238 goto bad;
2239 }
2240 } else {
2241 size_t n_blocks, n_metadata_blocks;
2242 uint64_t n_bitmap_bits;
2243
2244 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2245
2246 bio_list_init(&wc->flush_list);
2247 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2248 if (IS_ERR(wc->flush_thread)) {
2249 r = PTR_ERR(wc->flush_thread);
2250 wc->flush_thread = NULL;
2251 ti->error = "Couldn't spawn flush thread";
2252 goto bad;
2253 }
2254 wake_up_process(wc->flush_thread);
2255
2256 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2257 &n_blocks, &n_metadata_blocks);
2258 if (r) {
2259 ti->error = "Invalid device size";
2260 goto bad;
2261 }
2262
2263 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2264 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2265 /* this is limitation of test_bit functions */
2266 if (n_bitmap_bits > 1U << 31) {
2267 r = -EFBIG;
2268 ti->error = "Invalid device size";
2269 goto bad;
2270 }
2271
2272 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2273 if (!wc->memory_map) {
2274 r = -ENOMEM;
2275 ti->error = "Unable to allocate memory for metadata";
2276 goto bad;
2277 }
2278
2279 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2280 if (IS_ERR(wc->dm_kcopyd)) {
2281 r = PTR_ERR(wc->dm_kcopyd);
2282 ti->error = "Unable to allocate dm-kcopyd client";
2283 wc->dm_kcopyd = NULL;
2284 goto bad;
2285 }
2286
2287 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2288 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2289 BITS_PER_LONG * sizeof(unsigned long);
2290 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2291 if (!wc->dirty_bitmap) {
2292 r = -ENOMEM;
2293 ti->error = "Unable to allocate dirty bitmap";
2294 goto bad;
2295 }
2296
2297 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2298 if (r) {
2299 ti->error = "Unable to read first block of metadata";
2300 goto bad;
2301 }
2302 }
2303
2304 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2305 if (r) {
2306 ti->error = "Hardware memory error when reading superblock";
2307 goto bad;
2308 }
2309 if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2310 r = init_memory(wc);
2311 if (r) {
2312 ti->error = "Unable to initialize device";
2313 goto bad;
2314 }
2315 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2316 if (r) {
2317 ti->error = "Hardware memory error when reading superblock";
2318 goto bad;
2319 }
2320 }
2321
2322 if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2323 ti->error = "Invalid magic in the superblock";
2324 r = -EINVAL;
2325 goto bad;
2326 }
2327
2328 if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2329 ti->error = "Invalid version in the superblock";
2330 r = -EINVAL;
2331 goto bad;
2332 }
2333
2334 if (le32_to_cpu(s.block_size) != wc->block_size) {
2335 ti->error = "Block size does not match superblock";
2336 r = -EINVAL;
2337 goto bad;
2338 }
2339
2340 wc->n_blocks = le64_to_cpu(s.n_blocks);
2341
2342 offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2343 if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2344 overflow:
2345 ti->error = "Overflow in size calculation";
2346 r = -EINVAL;
2347 goto bad;
2348 }
2349 offset += sizeof(struct wc_memory_superblock);
2350 if (offset < sizeof(struct wc_memory_superblock))
2351 goto overflow;
2352 offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2353 data_size = wc->n_blocks * (size_t)wc->block_size;
2354 if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2355 (offset + data_size < offset))
2356 goto overflow;
2357 if (offset + data_size > wc->memory_map_size) {
2358 ti->error = "Memory area is too small";
2359 r = -EINVAL;
2360 goto bad;
2361 }
2362
2363 wc->metadata_sectors = offset >> SECTOR_SHIFT;
2364 wc->block_start = (char *)sb(wc) + offset;
2365
2366 x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2367 x += 50;
2368 do_div(x, 100);
2369 wc->freelist_high_watermark = x;
2370 x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2371 x += 50;
2372 do_div(x, 100);
2373 wc->freelist_low_watermark = x;
2374
2375 if (wc->cleaner)
2376 activate_cleaner(wc);
2377
2378 r = writecache_alloc_entries(wc);
2379 if (r) {
2380 ti->error = "Cannot allocate memory";
2381 goto bad;
2382 }
2383
2384 ti->num_flush_bios = 1;
2385 ti->flush_supported = true;
2386 ti->num_discard_bios = 1;
2387
2388 if (WC_MODE_PMEM(wc))
2389 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2390
2391 return 0;
2392
2393 bad_arguments:
2394 r = -EINVAL;
2395 ti->error = "Bad arguments";
2396 bad:
2397 writecache_dtr(ti);
2398 return r;
2399 }
2400
2401 static void writecache_status(struct dm_target *ti, status_type_t type,
2402 unsigned status_flags, char *result, unsigned maxlen)
2403 {
2404 struct dm_writecache *wc = ti->private;
2405 unsigned extra_args;
2406 unsigned sz = 0;
2407 uint64_t x;
2408
2409 switch (type) {
2410 case STATUSTYPE_INFO:
2411 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2412 (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2413 (unsigned long long)wc->writeback_size);
2414 break;
2415 case STATUSTYPE_TABLE:
2416 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2417 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2418 extra_args = 0;
2419 if (wc->start_sector)
2420 extra_args += 2;
2421 if (wc->high_wm_percent_set && !wc->cleaner)
2422 extra_args += 2;
2423 if (wc->low_wm_percent_set && !wc->cleaner)
2424 extra_args += 2;
2425 if (wc->max_writeback_jobs_set)
2426 extra_args += 2;
2427 if (wc->autocommit_blocks_set)
2428 extra_args += 2;
2429 if (wc->autocommit_time_set)
2430 extra_args += 2;
2431 if (wc->cleaner)
2432 extra_args++;
2433 if (wc->writeback_fua_set)
2434 extra_args++;
2435
2436 DMEMIT("%u", extra_args);
2437 if (wc->start_sector)
2438 DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2439 if (wc->high_wm_percent_set && !wc->cleaner) {
2440 x = (uint64_t)wc->freelist_high_watermark * 100;
2441 x += wc->n_blocks / 2;
2442 do_div(x, (size_t)wc->n_blocks);
2443 DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2444 }
2445 if (wc->low_wm_percent_set && !wc->cleaner) {
2446 x = (uint64_t)wc->freelist_low_watermark * 100;
2447 x += wc->n_blocks / 2;
2448 do_div(x, (size_t)wc->n_blocks);
2449 DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2450 }
2451 if (wc->max_writeback_jobs_set)
2452 DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2453 if (wc->autocommit_blocks_set)
2454 DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2455 if (wc->autocommit_time_set)
2456 DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2457 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2458 DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2459 if (wc->cleaner)
2460 DMEMIT(" cleaner");
2461 if (wc->writeback_fua_set)
2462 DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2463 break;
2464 }
2465 }
2466
2467 static struct target_type writecache_target = {
2468 .name = "writecache",
2469 .version = {1, 3, 0},
2470 .module = THIS_MODULE,
2471 .ctr = writecache_ctr,
2472 .dtr = writecache_dtr,
2473 .status = writecache_status,
2474 .postsuspend = writecache_suspend,
2475 .resume = writecache_resume,
2476 .message = writecache_message,
2477 .map = writecache_map,
2478 .end_io = writecache_end_io,
2479 .iterate_devices = writecache_iterate_devices,
2480 .io_hints = writecache_io_hints,
2481 };
2482
2483 static int __init dm_writecache_init(void)
2484 {
2485 int r;
2486
2487 r = dm_register_target(&writecache_target);
2488 if (r < 0) {
2489 DMERR("register failed %d", r);
2490 return r;
2491 }
2492
2493 return 0;
2494 }
2495
2496 static void __exit dm_writecache_exit(void)
2497 {
2498 dm_unregister_target(&writecache_target);
2499 }
2500
2501 module_init(dm_writecache_init);
2502 module_exit(dm_writecache_exit);
2503
2504 MODULE_DESCRIPTION(DM_NAME " writecache target");
2505 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2506 MODULE_LICENSE("GPL");
Linux with added FPC-III support