?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
gpio: rcar: Fix runtime PM imbalance on error
[linux/fpc-iii.git] / drivers / md / dm-writecache.c
blob114927da9cc9a70689e1badf04ebad06dd9a52c8
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 }
290 } while (i < p);
291 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
292 if (!wc->memory_map) {
293 r = -ENOMEM;
294 goto err3;
295 }
296 kvfree(pages);
297 wc->memory_vmapped = true;
298 }
299
300 dax_read_unlock(id);
301
302 wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
303 wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
304
305 return 0;
306 err3:
307 kvfree(pages);
308 err2:
309 dax_read_unlock(id);
310 err1:
311 return r;
312 }
313 #else
314 static int persistent_memory_claim(struct dm_writecache *wc)
315 {
316 BUG();
317 }
318 #endif
319
320 static void persistent_memory_release(struct dm_writecache *wc)
321 {
322 if (wc->memory_vmapped)
323 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
324 }
325
326 static struct page *persistent_memory_page(void *addr)
327 {
328 if (is_vmalloc_addr(addr))
329 return vmalloc_to_page(addr);
330 else
331 return virt_to_page(addr);
332 }
333
334 static unsigned persistent_memory_page_offset(void *addr)
335 {
336 return (unsigned long)addr & (PAGE_SIZE - 1);
337 }
338
339 static void persistent_memory_flush_cache(void *ptr, size_t size)
340 {
341 if (is_vmalloc_addr(ptr))
342 flush_kernel_vmap_range(ptr, size);
343 }
344
345 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
346 {
347 if (is_vmalloc_addr(ptr))
348 invalidate_kernel_vmap_range(ptr, size);
349 }
350
351 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
352 {
353 return wc->memory_map;
354 }
355
356 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
357 {
358 return &sb(wc)->entries[e->index];
359 }
360
361 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
362 {
363 return (char *)wc->block_start + (e->index << wc->block_size_bits);
364 }
365
366 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
367 {
368 return wc->start_sector + wc->metadata_sectors +
369 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
370 }
371
372 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
373 {
374 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
375 return e->original_sector;
376 #else
377 return le64_to_cpu(memory_entry(wc, e)->original_sector);
378 #endif
379 }
380
381 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
382 {
383 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
384 return e->seq_count;
385 #else
386 return le64_to_cpu(memory_entry(wc, e)->seq_count);
387 #endif
388 }
389
390 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
391 {
392 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
393 e->seq_count = -1;
394 #endif
395 pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
396 }
397
398 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
399 uint64_t original_sector, uint64_t seq_count)
400 {
401 struct wc_memory_entry me;
402 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
403 e->original_sector = original_sector;
404 e->seq_count = seq_count;
405 #endif
406 me.original_sector = cpu_to_le64(original_sector);
407 me.seq_count = cpu_to_le64(seq_count);
408 pmem_assign(*memory_entry(wc, e), me);
409 }
410
411 #define writecache_error(wc, err, msg, arg...) \
412 do { \
413 if (!cmpxchg(&(wc)->error, 0, err)) \
414 DMERR(msg, ##arg); \
415 wake_up(&(wc)->freelist_wait); \
416 } while (0)
417
418 #define writecache_has_error(wc) (unlikely(READ_ONCE((wc)->error)))
419
420 static void writecache_flush_all_metadata(struct dm_writecache *wc)
421 {
422 if (!WC_MODE_PMEM(wc))
423 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
424 }
425
426 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
427 {
428 if (!WC_MODE_PMEM(wc))
429 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
430 wc->dirty_bitmap);
431 }
432
433 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
434
435 struct io_notify {
436 struct dm_writecache *wc;
437 struct completion c;
438 atomic_t count;
439 };
440
441 static void writecache_notify_io(unsigned long error, void *context)
442 {
443 struct io_notify *endio = context;
444
445 if (unlikely(error != 0))
446 writecache_error(endio->wc, -EIO, "error writing metadata");
447 BUG_ON(atomic_read(&endio->count) <= 0);
448 if (atomic_dec_and_test(&endio->count))
449 complete(&endio->c);
450 }
451
452 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
453 {
454 wait_event(wc->bio_in_progress_wait[direction],
455 !atomic_read(&wc->bio_in_progress[direction]));
456 }
457
458 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
459 {
460 struct dm_io_region region;
461 struct dm_io_request req;
462 struct io_notify endio = {
463 wc,
464 COMPLETION_INITIALIZER_ONSTACK(endio.c),
465 ATOMIC_INIT(1),
466 };
467 unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
468 unsigned i = 0;
469
470 while (1) {
471 unsigned j;
472 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
473 if (unlikely(i == bitmap_bits))
474 break;
475 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
476
477 region.bdev = wc->ssd_dev->bdev;
478 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
479 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
480
481 if (unlikely(region.sector >= wc->metadata_sectors))
482 break;
483 if (unlikely(region.sector + region.count > wc->metadata_sectors))
484 region.count = wc->metadata_sectors - region.sector;
485
486 region.sector += wc->start_sector;
487 atomic_inc(&endio.count);
488 req.bi_op = REQ_OP_WRITE;
489 req.bi_op_flags = REQ_SYNC;
490 req.mem.type = DM_IO_VMA;
491 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
492 req.client = wc->dm_io;
493 req.notify.fn = writecache_notify_io;
494 req.notify.context = &endio;
495
496 /* writing via async dm-io (implied by notify.fn above) won't return an error */
497 (void) dm_io(&req, 1, &region, NULL);
498 i = j;
499 }
500
501 writecache_notify_io(0, &endio);
502 wait_for_completion_io(&endio.c);
503
504 if (wait_for_ios)
505 writecache_wait_for_ios(wc, WRITE);
506
507 writecache_disk_flush(wc, wc->ssd_dev);
508
509 memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
510 }
511
512 static void ssd_commit_superblock(struct dm_writecache *wc)
513 {
514 int r;
515 struct dm_io_region region;
516 struct dm_io_request req;
517
518 region.bdev = wc->ssd_dev->bdev;
519 region.sector = 0;
520 region.count = PAGE_SIZE;
521
522 if (unlikely(region.sector + region.count > wc->metadata_sectors))
523 region.count = wc->metadata_sectors - region.sector;
524
525 region.sector += wc->start_sector;
526
527 req.bi_op = REQ_OP_WRITE;
528 req.bi_op_flags = REQ_SYNC | REQ_FUA;
529 req.mem.type = DM_IO_VMA;
530 req.mem.ptr.vma = (char *)wc->memory_map;
531 req.client = wc->dm_io;
532 req.notify.fn = NULL;
533 req.notify.context = NULL;
534
535 r = dm_io(&req, 1, &region, NULL);
536 if (unlikely(r))
537 writecache_error(wc, r, "error writing superblock");
538 }
539
540 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
541 {
542 if (WC_MODE_PMEM(wc))
543 wmb();
544 else
545 ssd_commit_flushed(wc, wait_for_ios);
546 }
547
548 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
549 {
550 int r;
551 struct dm_io_region region;
552 struct dm_io_request req;
553
554 region.bdev = dev->bdev;
555 region.sector = 0;
556 region.count = 0;
557 req.bi_op = REQ_OP_WRITE;
558 req.bi_op_flags = REQ_PREFLUSH;
559 req.mem.type = DM_IO_KMEM;
560 req.mem.ptr.addr = NULL;
561 req.client = wc->dm_io;
562 req.notify.fn = NULL;
563
564 r = dm_io(&req, 1, &region, NULL);
565 if (unlikely(r))
566 writecache_error(wc, r, "error flushing metadata: %d", r);
567 }
568
569 #define WFE_RETURN_FOLLOWING 1
570 #define WFE_LOWEST_SEQ 2
571
572 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
573 uint64_t block, int flags)
574 {
575 struct wc_entry *e;
576 struct rb_node *node = wc->tree.rb_node;
577
578 if (unlikely(!node))
579 return NULL;
580
581 while (1) {
582 e = container_of(node, struct wc_entry, rb_node);
583 if (read_original_sector(wc, e) == block)
584 break;
585
586 node = (read_original_sector(wc, e) >= block ?
587 e->rb_node.rb_left : e->rb_node.rb_right);
588 if (unlikely(!node)) {
589 if (!(flags & WFE_RETURN_FOLLOWING))
590 return NULL;
591 if (read_original_sector(wc, e) >= block) {
592 return e;
593 } else {
594 node = rb_next(&e->rb_node);
595 if (unlikely(!node))
596 return NULL;
597 e = container_of(node, struct wc_entry, rb_node);
598 return e;
599 }
600 }
601 }
602
603 while (1) {
604 struct wc_entry *e2;
605 if (flags & WFE_LOWEST_SEQ)
606 node = rb_prev(&e->rb_node);
607 else
608 node = rb_next(&e->rb_node);
609 if (unlikely(!node))
610 return e;
611 e2 = container_of(node, struct wc_entry, rb_node);
612 if (read_original_sector(wc, e2) != block)
613 return e;
614 e = e2;
615 }
616 }
617
618 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
619 {
620 struct wc_entry *e;
621 struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
622
623 while (*node) {
624 e = container_of(*node, struct wc_entry, rb_node);
625 parent = &e->rb_node;
626 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
627 node = &parent->rb_left;
628 else
629 node = &parent->rb_right;
630 }
631 rb_link_node(&ins->rb_node, parent, node);
632 rb_insert_color(&ins->rb_node, &wc->tree);
633 list_add(&ins->lru, &wc->lru);
634 ins->age = jiffies;
635 }
636
637 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
638 {
639 list_del(&e->lru);
640 rb_erase(&e->rb_node, &wc->tree);
641 }
642
643 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
644 {
645 if (WC_MODE_SORT_FREELIST(wc)) {
646 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
647 if (unlikely(!*node))
648 wc->current_free = e;
649 while (*node) {
650 parent = *node;
651 if (&e->rb_node < *node)
652 node = &parent->rb_left;
653 else
654 node = &parent->rb_right;
655 }
656 rb_link_node(&e->rb_node, parent, node);
657 rb_insert_color(&e->rb_node, &wc->freetree);
658 } else {
659 list_add_tail(&e->lru, &wc->freelist);
660 }
661 wc->freelist_size++;
662 }
663
664 static inline void writecache_verify_watermark(struct dm_writecache *wc)
665 {
666 if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
667 queue_work(wc->writeback_wq, &wc->writeback_work);
668 }
669
670 static void writecache_max_age_timer(struct timer_list *t)
671 {
672 struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
673
674 if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
675 queue_work(wc->writeback_wq, &wc->writeback_work);
676 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
677 }
678 }
679
680 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
681 {
682 struct wc_entry *e;
683
684 if (WC_MODE_SORT_FREELIST(wc)) {
685 struct rb_node *next;
686 if (unlikely(!wc->current_free))
687 return NULL;
688 e = wc->current_free;
689 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
690 return NULL;
691 next = rb_next(&e->rb_node);
692 rb_erase(&e->rb_node, &wc->freetree);
693 if (unlikely(!next))
694 next = rb_first(&wc->freetree);
695 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
696 } else {
697 if (unlikely(list_empty(&wc->freelist)))
698 return NULL;
699 e = container_of(wc->freelist.next, struct wc_entry, lru);
700 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
701 return NULL;
702 list_del(&e->lru);
703 }
704 wc->freelist_size--;
705
706 writecache_verify_watermark(wc);
707
708 return e;
709 }
710
711 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
712 {
713 writecache_unlink(wc, e);
714 writecache_add_to_freelist(wc, e);
715 clear_seq_count(wc, e);
716 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
717 if (unlikely(waitqueue_active(&wc->freelist_wait)))
718 wake_up(&wc->freelist_wait);
719 }
720
721 static void writecache_wait_on_freelist(struct dm_writecache *wc)
722 {
723 DEFINE_WAIT(wait);
724
725 prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
726 wc_unlock(wc);
727 io_schedule();
728 finish_wait(&wc->freelist_wait, &wait);
729 wc_lock(wc);
730 }
731
732 static void writecache_poison_lists(struct dm_writecache *wc)
733 {
734 /*
735 * Catch incorrect access to these values while the device is suspended.
736 */
737 memset(&wc->tree, -1, sizeof wc->tree);
738 wc->lru.next = LIST_POISON1;
739 wc->lru.prev = LIST_POISON2;
740 wc->freelist.next = LIST_POISON1;
741 wc->freelist.prev = LIST_POISON2;
742 }
743
744 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
745 {
746 writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
747 if (WC_MODE_PMEM(wc))
748 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
749 }
750
751 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
752 {
753 return read_seq_count(wc, e) < wc->seq_count;
754 }
755
756 static void writecache_flush(struct dm_writecache *wc)
757 {
758 struct wc_entry *e, *e2;
759 bool need_flush_after_free;
760
761 wc->uncommitted_blocks = 0;
762 del_timer(&wc->autocommit_timer);
763
764 if (list_empty(&wc->lru))
765 return;
766
767 e = container_of(wc->lru.next, struct wc_entry, lru);
768 if (writecache_entry_is_committed(wc, e)) {
769 if (wc->overwrote_committed) {
770 writecache_wait_for_ios(wc, WRITE);
771 writecache_disk_flush(wc, wc->ssd_dev);
772 wc->overwrote_committed = false;
773 }
774 return;
775 }
776 while (1) {
777 writecache_flush_entry(wc, e);
778 if (unlikely(e->lru.next == &wc->lru))
779 break;
780 e2 = container_of(e->lru.next, struct wc_entry, lru);
781 if (writecache_entry_is_committed(wc, e2))
782 break;
783 e = e2;
784 cond_resched();
785 }
786 writecache_commit_flushed(wc, true);
787
788 wc->seq_count++;
789 pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
790 if (WC_MODE_PMEM(wc))
791 writecache_commit_flushed(wc, false);
792 else
793 ssd_commit_superblock(wc);
794
795 wc->overwrote_committed = false;
796
797 need_flush_after_free = false;
798 while (1) {
799 /* Free another committed entry with lower seq-count */
800 struct rb_node *rb_node = rb_prev(&e->rb_node);
801
802 if (rb_node) {
803 e2 = container_of(rb_node, struct wc_entry, rb_node);
804 if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
805 likely(!e2->write_in_progress)) {
806 writecache_free_entry(wc, e2);
807 need_flush_after_free = true;
808 }
809 }
810 if (unlikely(e->lru.prev == &wc->lru))
811 break;
812 e = container_of(e->lru.prev, struct wc_entry, lru);
813 cond_resched();
814 }
815
816 if (need_flush_after_free)
817 writecache_commit_flushed(wc, false);
818 }
819
820 static void writecache_flush_work(struct work_struct *work)
821 {
822 struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
823
824 wc_lock(wc);
825 writecache_flush(wc);
826 wc_unlock(wc);
827 }
828
829 static void writecache_autocommit_timer(struct timer_list *t)
830 {
831 struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
832 if (!writecache_has_error(wc))
833 queue_work(wc->writeback_wq, &wc->flush_work);
834 }
835
836 static void writecache_schedule_autocommit(struct dm_writecache *wc)
837 {
838 if (!timer_pending(&wc->autocommit_timer))
839 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
840 }
841
842 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
843 {
844 struct wc_entry *e;
845 bool discarded_something = false;
846
847 e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
848 if (unlikely(!e))
849 return;
850
851 while (read_original_sector(wc, e) < end) {
852 struct rb_node *node = rb_next(&e->rb_node);
853
854 if (likely(!e->write_in_progress)) {
855 if (!discarded_something) {
856 writecache_wait_for_ios(wc, READ);
857 writecache_wait_for_ios(wc, WRITE);
858 discarded_something = true;
859 }
860 writecache_free_entry(wc, e);
861 }
862
863 if (unlikely(!node))
864 break;
865
866 e = container_of(node, struct wc_entry, rb_node);
867 }
868
869 if (discarded_something)
870 writecache_commit_flushed(wc, false);
871 }
872
873 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
874 {
875 if (wc->writeback_size) {
876 writecache_wait_on_freelist(wc);
877 return true;
878 }
879 return false;
880 }
881
882 static void writecache_suspend(struct dm_target *ti)
883 {
884 struct dm_writecache *wc = ti->private;
885 bool flush_on_suspend;
886
887 del_timer_sync(&wc->autocommit_timer);
888 del_timer_sync(&wc->max_age_timer);
889
890 wc_lock(wc);
891 writecache_flush(wc);
892 flush_on_suspend = wc->flush_on_suspend;
893 if (flush_on_suspend) {
894 wc->flush_on_suspend = false;
895 wc->writeback_all++;
896 queue_work(wc->writeback_wq, &wc->writeback_work);
897 }
898 wc_unlock(wc);
899
900 drain_workqueue(wc->writeback_wq);
901
902 wc_lock(wc);
903 if (flush_on_suspend)
904 wc->writeback_all--;
905 while (writecache_wait_for_writeback(wc));
906
907 if (WC_MODE_PMEM(wc))
908 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
909
910 writecache_poison_lists(wc);
911
912 wc_unlock(wc);
913 }
914
915 static int writecache_alloc_entries(struct dm_writecache *wc)
916 {
917 size_t b;
918
919 if (wc->entries)
920 return 0;
921 wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
922 if (!wc->entries)
923 return -ENOMEM;
924 for (b = 0; b < wc->n_blocks; b++) {
925 struct wc_entry *e = &wc->entries[b];
926 e->index = b;
927 e->write_in_progress = false;
928 cond_resched();
929 }
930
931 return 0;
932 }
933
934 static void writecache_resume(struct dm_target *ti)
935 {
936 struct dm_writecache *wc = ti->private;
937 size_t b;
938 bool need_flush = false;
939 __le64 sb_seq_count;
940 int r;
941
942 wc_lock(wc);
943
944 if (WC_MODE_PMEM(wc))
945 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
946
947 wc->tree = RB_ROOT;
948 INIT_LIST_HEAD(&wc->lru);
949 if (WC_MODE_SORT_FREELIST(wc)) {
950 wc->freetree = RB_ROOT;
951 wc->current_free = NULL;
952 } else {
953 INIT_LIST_HEAD(&wc->freelist);
954 }
955 wc->freelist_size = 0;
956
957 r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
958 if (r) {
959 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
960 sb_seq_count = cpu_to_le64(0);
961 }
962 wc->seq_count = le64_to_cpu(sb_seq_count);
963
964 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
965 for (b = 0; b < wc->n_blocks; b++) {
966 struct wc_entry *e = &wc->entries[b];
967 struct wc_memory_entry wme;
968 if (writecache_has_error(wc)) {
969 e->original_sector = -1;
970 e->seq_count = -1;
971 continue;
972 }
973 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
974 if (r) {
975 writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
976 (unsigned long)b, r);
977 e->original_sector = -1;
978 e->seq_count = -1;
979 } else {
980 e->original_sector = le64_to_cpu(wme.original_sector);
981 e->seq_count = le64_to_cpu(wme.seq_count);
982 }
983 cond_resched();
984 }
985 #endif
986 for (b = 0; b < wc->n_blocks; b++) {
987 struct wc_entry *e = &wc->entries[b];
988 if (!writecache_entry_is_committed(wc, e)) {
989 if (read_seq_count(wc, e) != -1) {
990 erase_this:
991 clear_seq_count(wc, e);
992 need_flush = true;
993 }
994 writecache_add_to_freelist(wc, e);
995 } else {
996 struct wc_entry *old;
997
998 old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
999 if (!old) {
1000 writecache_insert_entry(wc, e);
1001 } else {
1002 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1003 writecache_error(wc, -EINVAL,
1004 "two identical entries, position %llu, sector %llu, sequence %llu",
1005 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1006 (unsigned long long)read_seq_count(wc, e));
1007 }
1008 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1009 goto erase_this;
1010 } else {
1011 writecache_free_entry(wc, old);
1012 writecache_insert_entry(wc, e);
1013 need_flush = true;
1014 }
1015 }
1016 }
1017 cond_resched();
1018 }
1019
1020 if (need_flush) {
1021 writecache_flush_all_metadata(wc);
1022 writecache_commit_flushed(wc, false);
1023 }
1024
1025 writecache_verify_watermark(wc);
1026
1027 if (wc->max_age != MAX_AGE_UNSPECIFIED)
1028 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1029
1030 wc_unlock(wc);
1031 }
1032
1033 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1034 {
1035 if (argc != 1)
1036 return -EINVAL;
1037
1038 wc_lock(wc);
1039 if (dm_suspended(wc->ti)) {
1040 wc_unlock(wc);
1041 return -EBUSY;
1042 }
1043 if (writecache_has_error(wc)) {
1044 wc_unlock(wc);
1045 return -EIO;
1046 }
1047
1048 writecache_flush(wc);
1049 wc->writeback_all++;
1050 queue_work(wc->writeback_wq, &wc->writeback_work);
1051 wc_unlock(wc);
1052
1053 flush_workqueue(wc->writeback_wq);
1054
1055 wc_lock(wc);
1056 wc->writeback_all--;
1057 if (writecache_has_error(wc)) {
1058 wc_unlock(wc);
1059 return -EIO;
1060 }
1061 wc_unlock(wc);
1062
1063 return 0;
1064 }
1065
1066 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1067 {
1068 if (argc != 1)
1069 return -EINVAL;
1070
1071 wc_lock(wc);
1072 wc->flush_on_suspend = true;
1073 wc_unlock(wc);
1074
1075 return 0;
1076 }
1077
1078 static void activate_cleaner(struct dm_writecache *wc)
1079 {
1080 wc->flush_on_suspend = true;
1081 wc->cleaner = true;
1082 wc->freelist_high_watermark = wc->n_blocks;
1083 wc->freelist_low_watermark = wc->n_blocks;
1084 }
1085
1086 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1087 {
1088 if (argc != 1)
1089 return -EINVAL;
1090
1091 wc_lock(wc);
1092 activate_cleaner(wc);
1093 if (!dm_suspended(wc->ti))
1094 writecache_verify_watermark(wc);
1095 wc_unlock(wc);
1096
1097 return 0;
1098 }
1099
1100 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1101 char *result, unsigned maxlen)
1102 {
1103 int r = -EINVAL;
1104 struct dm_writecache *wc = ti->private;
1105
1106 if (!strcasecmp(argv[0], "flush"))
1107 r = process_flush_mesg(argc, argv, wc);
1108 else if (!strcasecmp(argv[0], "flush_on_suspend"))
1109 r = process_flush_on_suspend_mesg(argc, argv, wc);
1110 else if (!strcasecmp(argv[0], "cleaner"))
1111 r = process_cleaner_mesg(argc, argv, wc);
1112 else
1113 DMERR("unrecognised message received: %s", argv[0]);
1114
1115 return r;
1116 }
1117
1118 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1119 {
1120 void *buf;
1121 unsigned long flags;
1122 unsigned size;
1123 int rw = bio_data_dir(bio);
1124 unsigned remaining_size = wc->block_size;
1125
1126 do {
1127 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1128 buf = bvec_kmap_irq(&bv, &flags);
1129 size = bv.bv_len;
1130 if (unlikely(size > remaining_size))
1131 size = remaining_size;
1132
1133 if (rw == READ) {
1134 int r;
1135 r = memcpy_mcsafe(buf, data, size);
1136 flush_dcache_page(bio_page(bio));
1137 if (unlikely(r)) {
1138 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1139 bio->bi_status = BLK_STS_IOERR;
1140 }
1141 } else {
1142 flush_dcache_page(bio_page(bio));
1143 memcpy_flushcache(data, buf, size);
1144 }
1145
1146 bvec_kunmap_irq(buf, &flags);
1147
1148 data = (char *)data + size;
1149 remaining_size -= size;
1150 bio_advance(bio, size);
1151 } while (unlikely(remaining_size));
1152 }
1153
1154 static int writecache_flush_thread(void *data)
1155 {
1156 struct dm_writecache *wc = data;
1157
1158 while (1) {
1159 struct bio *bio;
1160
1161 wc_lock(wc);
1162 bio = bio_list_pop(&wc->flush_list);
1163 if (!bio) {
1164 set_current_state(TASK_INTERRUPTIBLE);
1165 wc_unlock(wc);
1166
1167 if (unlikely(kthread_should_stop())) {
1168 set_current_state(TASK_RUNNING);
1169 break;
1170 }
1171
1172 schedule();
1173 continue;
1174 }
1175
1176 if (bio_op(bio) == REQ_OP_DISCARD) {
1177 writecache_discard(wc, bio->bi_iter.bi_sector,
1178 bio_end_sector(bio));
1179 wc_unlock(wc);
1180 bio_set_dev(bio, wc->dev->bdev);
1181 generic_make_request(bio);
1182 } else {
1183 writecache_flush(wc);
1184 wc_unlock(wc);
1185 if (writecache_has_error(wc))
1186 bio->bi_status = BLK_STS_IOERR;
1187 bio_endio(bio);
1188 }
1189 }
1190
1191 return 0;
1192 }
1193
1194 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1195 {
1196 if (bio_list_empty(&wc->flush_list))
1197 wake_up_process(wc->flush_thread);
1198 bio_list_add(&wc->flush_list, bio);
1199 }
1200
1201 static int writecache_map(struct dm_target *ti, struct bio *bio)
1202 {
1203 struct wc_entry *e;
1204 struct dm_writecache *wc = ti->private;
1205
1206 bio->bi_private = NULL;
1207
1208 wc_lock(wc);
1209
1210 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1211 if (writecache_has_error(wc))
1212 goto unlock_error;
1213 if (WC_MODE_PMEM(wc)) {
1214 writecache_flush(wc);
1215 if (writecache_has_error(wc))
1216 goto unlock_error;
1217 goto unlock_submit;
1218 } else {
1219 writecache_offload_bio(wc, bio);
1220 goto unlock_return;
1221 }
1222 }
1223
1224 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1225
1226 if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1227 (wc->block_size / 512 - 1)) != 0)) {
1228 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1229 (unsigned long long)bio->bi_iter.bi_sector,
1230 bio->bi_iter.bi_size, wc->block_size);
1231 goto unlock_error;
1232 }
1233
1234 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1235 if (writecache_has_error(wc))
1236 goto unlock_error;
1237 if (WC_MODE_PMEM(wc)) {
1238 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1239 goto unlock_remap_origin;
1240 } else {
1241 writecache_offload_bio(wc, bio);
1242 goto unlock_return;
1243 }
1244 }
1245
1246 if (bio_data_dir(bio) == READ) {
1247 read_next_block:
1248 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1249 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1250 if (WC_MODE_PMEM(wc)) {
1251 bio_copy_block(wc, bio, memory_data(wc, e));
1252 if (bio->bi_iter.bi_size)
1253 goto read_next_block;
1254 goto unlock_submit;
1255 } else {
1256 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1257 bio_set_dev(bio, wc->ssd_dev->bdev);
1258 bio->bi_iter.bi_sector = cache_sector(wc, e);
1259 if (!writecache_entry_is_committed(wc, e))
1260 writecache_wait_for_ios(wc, WRITE);
1261 goto unlock_remap;
1262 }
1263 } else {
1264 if (e) {
1265 sector_t next_boundary =
1266 read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1267 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1268 dm_accept_partial_bio(bio, next_boundary);
1269 }
1270 }
1271 goto unlock_remap_origin;
1272 }
1273 } else {
1274 do {
1275 bool found_entry = false;
1276 if (writecache_has_error(wc))
1277 goto unlock_error;
1278 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1279 if (e) {
1280 if (!writecache_entry_is_committed(wc, e))
1281 goto bio_copy;
1282 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1283 wc->overwrote_committed = true;
1284 goto bio_copy;
1285 }
1286 found_entry = true;
1287 } else {
1288 if (unlikely(wc->cleaner))
1289 goto direct_write;
1290 }
1291 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1292 if (unlikely(!e)) {
1293 if (!found_entry) {
1294 direct_write:
1295 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1296 if (e) {
1297 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1298 BUG_ON(!next_boundary);
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 writecache_wait_on_freelist(wc);
1306 continue;
1307 }
1308 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1309 writecache_insert_entry(wc, e);
1310 wc->uncommitted_blocks++;
1311 bio_copy:
1312 if (WC_MODE_PMEM(wc)) {
1313 bio_copy_block(wc, bio, memory_data(wc, e));
1314 } else {
1315 unsigned bio_size = wc->block_size;
1316 sector_t start_cache_sec = cache_sector(wc, e);
1317 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1318
1319 while (bio_size < bio->bi_iter.bi_size) {
1320 struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1321 if (!f)
1322 break;
1323 write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1324 (bio_size >> SECTOR_SHIFT), wc->seq_count);
1325 writecache_insert_entry(wc, f);
1326 wc->uncommitted_blocks++;
1327 bio_size += wc->block_size;
1328 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1329 }
1330
1331 bio_set_dev(bio, wc->ssd_dev->bdev);
1332 bio->bi_iter.bi_sector = start_cache_sec;
1333 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1334
1335 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1336 wc->uncommitted_blocks = 0;
1337 queue_work(wc->writeback_wq, &wc->flush_work);
1338 } else {
1339 writecache_schedule_autocommit(wc);
1340 }
1341 goto unlock_remap;
1342 }
1343 } while (bio->bi_iter.bi_size);
1344
1345 if (unlikely(bio->bi_opf & REQ_FUA ||
1346 wc->uncommitted_blocks >= wc->autocommit_blocks))
1347 writecache_flush(wc);
1348 else
1349 writecache_schedule_autocommit(wc);
1350 goto unlock_submit;
1351 }
1352
1353 unlock_remap_origin:
1354 bio_set_dev(bio, wc->dev->bdev);
1355 wc_unlock(wc);
1356 return DM_MAPIO_REMAPPED;
1357
1358 unlock_remap:
1359 /* make sure that writecache_end_io decrements bio_in_progress: */
1360 bio->bi_private = (void *)1;
1361 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1362 wc_unlock(wc);
1363 return DM_MAPIO_REMAPPED;
1364
1365 unlock_submit:
1366 wc_unlock(wc);
1367 bio_endio(bio);
1368 return DM_MAPIO_SUBMITTED;
1369
1370 unlock_return:
1371 wc_unlock(wc);
1372 return DM_MAPIO_SUBMITTED;
1373
1374 unlock_error:
1375 wc_unlock(wc);
1376 bio_io_error(bio);
1377 return DM_MAPIO_SUBMITTED;
1378 }
1379
1380 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1381 {
1382 struct dm_writecache *wc = ti->private;
1383
1384 if (bio->bi_private != NULL) {
1385 int dir = bio_data_dir(bio);
1386 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1387 if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1388 wake_up(&wc->bio_in_progress_wait[dir]);
1389 }
1390 return 0;
1391 }
1392
1393 static int writecache_iterate_devices(struct dm_target *ti,
1394 iterate_devices_callout_fn fn, void *data)
1395 {
1396 struct dm_writecache *wc = ti->private;
1397
1398 return fn(ti, wc->dev, 0, ti->len, data);
1399 }
1400
1401 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1402 {
1403 struct dm_writecache *wc = ti->private;
1404
1405 if (limits->logical_block_size < wc->block_size)
1406 limits->logical_block_size = wc->block_size;
1407
1408 if (limits->physical_block_size < wc->block_size)
1409 limits->physical_block_size = wc->block_size;
1410
1411 if (limits->io_min < wc->block_size)
1412 limits->io_min = wc->block_size;
1413 }
1414
1415
1416 static void writecache_writeback_endio(struct bio *bio)
1417 {
1418 struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1419 struct dm_writecache *wc = wb->wc;
1420 unsigned long flags;
1421
1422 raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1423 if (unlikely(list_empty(&wc->endio_list)))
1424 wake_up_process(wc->endio_thread);
1425 list_add_tail(&wb->endio_entry, &wc->endio_list);
1426 raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1427 }
1428
1429 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1430 {
1431 struct copy_struct *c = ptr;
1432 struct dm_writecache *wc = c->wc;
1433
1434 c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1435
1436 raw_spin_lock_irq(&wc->endio_list_lock);
1437 if (unlikely(list_empty(&wc->endio_list)))
1438 wake_up_process(wc->endio_thread);
1439 list_add_tail(&c->endio_entry, &wc->endio_list);
1440 raw_spin_unlock_irq(&wc->endio_list_lock);
1441 }
1442
1443 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1444 {
1445 unsigned i;
1446 struct writeback_struct *wb;
1447 struct wc_entry *e;
1448 unsigned long n_walked = 0;
1449
1450 do {
1451 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1452 list_del(&wb->endio_entry);
1453
1454 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1455 writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1456 "write error %d", wb->bio.bi_status);
1457 i = 0;
1458 do {
1459 e = wb->wc_list[i];
1460 BUG_ON(!e->write_in_progress);
1461 e->write_in_progress = false;
1462 INIT_LIST_HEAD(&e->lru);
1463 if (!writecache_has_error(wc))
1464 writecache_free_entry(wc, e);
1465 BUG_ON(!wc->writeback_size);
1466 wc->writeback_size--;
1467 n_walked++;
1468 if (unlikely(n_walked >= ENDIO_LATENCY)) {
1469 writecache_commit_flushed(wc, false);
1470 wc_unlock(wc);
1471 wc_lock(wc);
1472 n_walked = 0;
1473 }
1474 } while (++i < wb->wc_list_n);
1475
1476 if (wb->wc_list != wb->wc_list_inline)
1477 kfree(wb->wc_list);
1478 bio_put(&wb->bio);
1479 } while (!list_empty(list));
1480 }
1481
1482 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1483 {
1484 struct copy_struct *c;
1485 struct wc_entry *e;
1486
1487 do {
1488 c = list_entry(list->next, struct copy_struct, endio_entry);
1489 list_del(&c->endio_entry);
1490
1491 if (unlikely(c->error))
1492 writecache_error(wc, c->error, "copy error");
1493
1494 e = c->e;
1495 do {
1496 BUG_ON(!e->write_in_progress);
1497 e->write_in_progress = false;
1498 INIT_LIST_HEAD(&e->lru);
1499 if (!writecache_has_error(wc))
1500 writecache_free_entry(wc, e);
1501
1502 BUG_ON(!wc->writeback_size);
1503 wc->writeback_size--;
1504 e++;
1505 } while (--c->n_entries);
1506 mempool_free(c, &wc->copy_pool);
1507 } while (!list_empty(list));
1508 }
1509
1510 static int writecache_endio_thread(void *data)
1511 {
1512 struct dm_writecache *wc = data;
1513
1514 while (1) {
1515 struct list_head list;
1516
1517 raw_spin_lock_irq(&wc->endio_list_lock);
1518 if (!list_empty(&wc->endio_list))
1519 goto pop_from_list;
1520 set_current_state(TASK_INTERRUPTIBLE);
1521 raw_spin_unlock_irq(&wc->endio_list_lock);
1522
1523 if (unlikely(kthread_should_stop())) {
1524 set_current_state(TASK_RUNNING);
1525 break;
1526 }
1527
1528 schedule();
1529
1530 continue;
1531
1532 pop_from_list:
1533 list = wc->endio_list;
1534 list.next->prev = list.prev->next = &list;
1535 INIT_LIST_HEAD(&wc->endio_list);
1536 raw_spin_unlock_irq(&wc->endio_list_lock);
1537
1538 if (!WC_MODE_FUA(wc))
1539 writecache_disk_flush(wc, wc->dev);
1540
1541 wc_lock(wc);
1542
1543 if (WC_MODE_PMEM(wc)) {
1544 __writecache_endio_pmem(wc, &list);
1545 } else {
1546 __writecache_endio_ssd(wc, &list);
1547 writecache_wait_for_ios(wc, READ);
1548 }
1549
1550 writecache_commit_flushed(wc, false);
1551
1552 wc_unlock(wc);
1553 }
1554
1555 return 0;
1556 }
1557
1558 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1559 {
1560 struct dm_writecache *wc = wb->wc;
1561 unsigned block_size = wc->block_size;
1562 void *address = memory_data(wc, e);
1563
1564 persistent_memory_flush_cache(address, block_size);
1565 return bio_add_page(&wb->bio, persistent_memory_page(address),
1566 block_size, persistent_memory_page_offset(address)) != 0;
1567 }
1568
1569 struct writeback_list {
1570 struct list_head list;
1571 size_t size;
1572 };
1573
1574 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1575 {
1576 if (unlikely(wc->max_writeback_jobs)) {
1577 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1578 wc_lock(wc);
1579 while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1580 writecache_wait_on_freelist(wc);
1581 wc_unlock(wc);
1582 }
1583 }
1584 cond_resched();
1585 }
1586
1587 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1588 {
1589 struct wc_entry *e, *f;
1590 struct bio *bio;
1591 struct writeback_struct *wb;
1592 unsigned max_pages;
1593
1594 while (wbl->size) {
1595 wbl->size--;
1596 e = container_of(wbl->list.prev, struct wc_entry, lru);
1597 list_del(&e->lru);
1598
1599 max_pages = e->wc_list_contiguous;
1600
1601 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1602 wb = container_of(bio, struct writeback_struct, bio);
1603 wb->wc = wc;
1604 bio->bi_end_io = writecache_writeback_endio;
1605 bio_set_dev(bio, wc->dev->bdev);
1606 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1607 if (max_pages <= WB_LIST_INLINE ||
1608 unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1609 GFP_NOIO | __GFP_NORETRY |
1610 __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1611 wb->wc_list = wb->wc_list_inline;
1612 max_pages = WB_LIST_INLINE;
1613 }
1614
1615 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1616
1617 wb->wc_list[0] = e;
1618 wb->wc_list_n = 1;
1619
1620 while (wbl->size && wb->wc_list_n < max_pages) {
1621 f = container_of(wbl->list.prev, struct wc_entry, lru);
1622 if (read_original_sector(wc, f) !=
1623 read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1624 break;
1625 if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1626 break;
1627 wbl->size--;
1628 list_del(&f->lru);
1629 wb->wc_list[wb->wc_list_n++] = f;
1630 e = f;
1631 }
1632 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1633 if (writecache_has_error(wc)) {
1634 bio->bi_status = BLK_STS_IOERR;
1635 bio_endio(bio);
1636 } else {
1637 submit_bio(bio);
1638 }
1639
1640 __writeback_throttle(wc, wbl);
1641 }
1642 }
1643
1644 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1645 {
1646 struct wc_entry *e, *f;
1647 struct dm_io_region from, to;
1648 struct copy_struct *c;
1649
1650 while (wbl->size) {
1651 unsigned n_sectors;
1652
1653 wbl->size--;
1654 e = container_of(wbl->list.prev, struct wc_entry, lru);
1655 list_del(&e->lru);
1656
1657 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1658
1659 from.bdev = wc->ssd_dev->bdev;
1660 from.sector = cache_sector(wc, e);
1661 from.count = n_sectors;
1662 to.bdev = wc->dev->bdev;
1663 to.sector = read_original_sector(wc, e);
1664 to.count = n_sectors;
1665
1666 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1667 c->wc = wc;
1668 c->e = e;
1669 c->n_entries = e->wc_list_contiguous;
1670
1671 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1672 wbl->size--;
1673 f = container_of(wbl->list.prev, struct wc_entry, lru);
1674 BUG_ON(f != e + 1);
1675 list_del(&f->lru);
1676 e = f;
1677 }
1678
1679 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1680
1681 __writeback_throttle(wc, wbl);
1682 }
1683 }
1684
1685 static void writecache_writeback(struct work_struct *work)
1686 {
1687 struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1688 struct blk_plug plug;
1689 struct wc_entry *f, *uninitialized_var(g), *e = NULL;
1690 struct rb_node *node, *next_node;
1691 struct list_head skipped;
1692 struct writeback_list wbl;
1693 unsigned long n_walked;
1694
1695 wc_lock(wc);
1696 restart:
1697 if (writecache_has_error(wc)) {
1698 wc_unlock(wc);
1699 return;
1700 }
1701
1702 if (unlikely(wc->writeback_all)) {
1703 if (writecache_wait_for_writeback(wc))
1704 goto restart;
1705 }
1706
1707 if (wc->overwrote_committed) {
1708 writecache_wait_for_ios(wc, WRITE);
1709 }
1710
1711 n_walked = 0;
1712 INIT_LIST_HEAD(&skipped);
1713 INIT_LIST_HEAD(&wbl.list);
1714 wbl.size = 0;
1715 while (!list_empty(&wc->lru) &&
1716 (wc->writeback_all ||
1717 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1718 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1719 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1720
1721 n_walked++;
1722 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1723 likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1724 queue_work(wc->writeback_wq, &wc->writeback_work);
1725 break;
1726 }
1727
1728 if (unlikely(wc->writeback_all)) {
1729 if (unlikely(!e)) {
1730 writecache_flush(wc);
1731 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1732 } else
1733 e = g;
1734 } else
1735 e = container_of(wc->lru.prev, struct wc_entry, lru);
1736 BUG_ON(e->write_in_progress);
1737 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1738 writecache_flush(wc);
1739 }
1740 node = rb_prev(&e->rb_node);
1741 if (node) {
1742 f = container_of(node, struct wc_entry, rb_node);
1743 if (unlikely(read_original_sector(wc, f) ==
1744 read_original_sector(wc, e))) {
1745 BUG_ON(!f->write_in_progress);
1746 list_del(&e->lru);
1747 list_add(&e->lru, &skipped);
1748 cond_resched();
1749 continue;
1750 }
1751 }
1752 wc->writeback_size++;
1753 list_del(&e->lru);
1754 list_add(&e->lru, &wbl.list);
1755 wbl.size++;
1756 e->write_in_progress = true;
1757 e->wc_list_contiguous = 1;
1758
1759 f = e;
1760
1761 while (1) {
1762 next_node = rb_next(&f->rb_node);
1763 if (unlikely(!next_node))
1764 break;
1765 g = container_of(next_node, struct wc_entry, rb_node);
1766 if (unlikely(read_original_sector(wc, g) ==
1767 read_original_sector(wc, f))) {
1768 f = g;
1769 continue;
1770 }
1771 if (read_original_sector(wc, g) !=
1772 read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1773 break;
1774 if (unlikely(g->write_in_progress))
1775 break;
1776 if (unlikely(!writecache_entry_is_committed(wc, g)))
1777 break;
1778
1779 if (!WC_MODE_PMEM(wc)) {
1780 if (g != f + 1)
1781 break;
1782 }
1783
1784 n_walked++;
1785 //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1786 // break;
1787
1788 wc->writeback_size++;
1789 list_del(&g->lru);
1790 list_add(&g->lru, &wbl.list);
1791 wbl.size++;
1792 g->write_in_progress = true;
1793 g->wc_list_contiguous = BIO_MAX_PAGES;
1794 f = g;
1795 e->wc_list_contiguous++;
1796 if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1797 if (unlikely(wc->writeback_all)) {
1798 next_node = rb_next(&f->rb_node);
1799 if (likely(next_node))
1800 g = container_of(next_node, struct wc_entry, rb_node);
1801 }
1802 break;
1803 }
1804 }
1805 cond_resched();
1806 }
1807
1808 if (!list_empty(&skipped)) {
1809 list_splice_tail(&skipped, &wc->lru);
1810 /*
1811 * If we didn't do any progress, we must wait until some
1812 * writeback finishes to avoid burning CPU in a loop
1813 */
1814 if (unlikely(!wbl.size))
1815 writecache_wait_for_writeback(wc);
1816 }
1817
1818 wc_unlock(wc);
1819
1820 blk_start_plug(&plug);
1821
1822 if (WC_MODE_PMEM(wc))
1823 __writecache_writeback_pmem(wc, &wbl);
1824 else
1825 __writecache_writeback_ssd(wc, &wbl);
1826
1827 blk_finish_plug(&plug);
1828
1829 if (unlikely(wc->writeback_all)) {
1830 wc_lock(wc);
1831 while (writecache_wait_for_writeback(wc));
1832 wc_unlock(wc);
1833 }
1834 }
1835
1836 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1837 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1838 {
1839 uint64_t n_blocks, offset;
1840 struct wc_entry e;
1841
1842 n_blocks = device_size;
1843 do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1844
1845 while (1) {
1846 if (!n_blocks)
1847 return -ENOSPC;
1848 /* Verify the following entries[n_blocks] won't overflow */
1849 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1850 sizeof(struct wc_memory_entry)))
1851 return -EFBIG;
1852 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1853 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1854 if (offset + n_blocks * block_size <= device_size)
1855 break;
1856 n_blocks--;
1857 }
1858
1859 /* check if the bit field overflows */
1860 e.index = n_blocks;
1861 if (e.index != n_blocks)
1862 return -EFBIG;
1863
1864 if (n_blocks_p)
1865 *n_blocks_p = n_blocks;
1866 if (n_metadata_blocks_p)
1867 *n_metadata_blocks_p = offset >> __ffs(block_size);
1868 return 0;
1869 }
1870
1871 static int init_memory(struct dm_writecache *wc)
1872 {
1873 size_t b;
1874 int r;
1875
1876 r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1877 if (r)
1878 return r;
1879
1880 r = writecache_alloc_entries(wc);
1881 if (r)
1882 return r;
1883
1884 for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1885 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1886 pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1887 pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1888 pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1889 pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1890
1891 for (b = 0; b < wc->n_blocks; b++) {
1892 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1893 cond_resched();
1894 }
1895
1896 writecache_flush_all_metadata(wc);
1897 writecache_commit_flushed(wc, false);
1898 pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1899 writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1900 writecache_commit_flushed(wc, false);
1901
1902 return 0;
1903 }
1904
1905 static void writecache_dtr(struct dm_target *ti)
1906 {
1907 struct dm_writecache *wc = ti->private;
1908
1909 if (!wc)
1910 return;
1911
1912 if (wc->endio_thread)
1913 kthread_stop(wc->endio_thread);
1914
1915 if (wc->flush_thread)
1916 kthread_stop(wc->flush_thread);
1917
1918 bioset_exit(&wc->bio_set);
1919
1920 mempool_exit(&wc->copy_pool);
1921
1922 if (wc->writeback_wq)
1923 destroy_workqueue(wc->writeback_wq);
1924
1925 if (wc->dev)
1926 dm_put_device(ti, wc->dev);
1927
1928 if (wc->ssd_dev)
1929 dm_put_device(ti, wc->ssd_dev);
1930
1931 if (wc->entries)
1932 vfree(wc->entries);
1933
1934 if (wc->memory_map) {
1935 if (WC_MODE_PMEM(wc))
1936 persistent_memory_release(wc);
1937 else
1938 vfree(wc->memory_map);
1939 }
1940
1941 if (wc->dm_kcopyd)
1942 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1943
1944 if (wc->dm_io)
1945 dm_io_client_destroy(wc->dm_io);
1946
1947 if (wc->dirty_bitmap)
1948 vfree(wc->dirty_bitmap);
1949
1950 kfree(wc);
1951 }
1952
1953 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1954 {
1955 struct dm_writecache *wc;
1956 struct dm_arg_set as;
1957 const char *string;
1958 unsigned opt_params;
1959 size_t offset, data_size;
1960 int i, r;
1961 char dummy;
1962 int high_wm_percent = HIGH_WATERMARK;
1963 int low_wm_percent = LOW_WATERMARK;
1964 uint64_t x;
1965 struct wc_memory_superblock s;
1966
1967 static struct dm_arg _args[] = {
1968 {0, 10, "Invalid number of feature args"},
1969 };
1970
1971 as.argc = argc;
1972 as.argv = argv;
1973
1974 wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
1975 if (!wc) {
1976 ti->error = "Cannot allocate writecache structure";
1977 r = -ENOMEM;
1978 goto bad;
1979 }
1980 ti->private = wc;
1981 wc->ti = ti;
1982
1983 mutex_init(&wc->lock);
1984 wc->max_age = MAX_AGE_UNSPECIFIED;
1985 writecache_poison_lists(wc);
1986 init_waitqueue_head(&wc->freelist_wait);
1987 timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
1988 timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
1989
1990 for (i = 0; i < 2; i++) {
1991 atomic_set(&wc->bio_in_progress[i], 0);
1992 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
1993 }
1994
1995 wc->dm_io = dm_io_client_create();
1996 if (IS_ERR(wc->dm_io)) {
1997 r = PTR_ERR(wc->dm_io);
1998 ti->error = "Unable to allocate dm-io client";
1999 wc->dm_io = NULL;
2000 goto bad;
2001 }
2002
2003 wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2004 if (!wc->writeback_wq) {
2005 r = -ENOMEM;
2006 ti->error = "Could not allocate writeback workqueue";
2007 goto bad;
2008 }
2009 INIT_WORK(&wc->writeback_work, writecache_writeback);
2010 INIT_WORK(&wc->flush_work, writecache_flush_work);
2011
2012 raw_spin_lock_init(&wc->endio_list_lock);
2013 INIT_LIST_HEAD(&wc->endio_list);
2014 wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2015 if (IS_ERR(wc->endio_thread)) {
2016 r = PTR_ERR(wc->endio_thread);
2017 wc->endio_thread = NULL;
2018 ti->error = "Couldn't spawn endio thread";
2019 goto bad;
2020 }
2021 wake_up_process(wc->endio_thread);
2022
2023 /*
2024 * Parse the mode (pmem or ssd)
2025 */
2026 string = dm_shift_arg(&as);
2027 if (!string)
2028 goto bad_arguments;
2029
2030 if (!strcasecmp(string, "s")) {
2031 wc->pmem_mode = false;
2032 } else if (!strcasecmp(string, "p")) {
2033 #ifdef DM_WRITECACHE_HAS_PMEM
2034 wc->pmem_mode = true;
2035 wc->writeback_fua = true;
2036 #else
2037 /*
2038 * If the architecture doesn't support persistent memory or
2039 * the kernel doesn't support any DAX drivers, this driver can
2040 * only be used in SSD-only mode.
2041 */
2042 r = -EOPNOTSUPP;
2043 ti->error = "Persistent memory or DAX not supported on this system";
2044 goto bad;
2045 #endif
2046 } else {
2047 goto bad_arguments;
2048 }
2049
2050 if (WC_MODE_PMEM(wc)) {
2051 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2052 offsetof(struct writeback_struct, bio),
2053 BIOSET_NEED_BVECS);
2054 if (r) {
2055 ti->error = "Could not allocate bio set";
2056 goto bad;
2057 }
2058 } else {
2059 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2060 if (r) {
2061 ti->error = "Could not allocate mempool";
2062 goto bad;
2063 }
2064 }
2065
2066 /*
2067 * Parse the origin data device
2068 */
2069 string = dm_shift_arg(&as);
2070 if (!string)
2071 goto bad_arguments;
2072 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2073 if (r) {
2074 ti->error = "Origin data device lookup failed";
2075 goto bad;
2076 }
2077
2078 /*
2079 * Parse cache data device (be it pmem or ssd)
2080 */
2081 string = dm_shift_arg(&as);
2082 if (!string)
2083 goto bad_arguments;
2084
2085 r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2086 if (r) {
2087 ti->error = "Cache data device lookup failed";
2088 goto bad;
2089 }
2090 wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2091
2092 /*
2093 * Parse the cache block size
2094 */
2095 string = dm_shift_arg(&as);
2096 if (!string)
2097 goto bad_arguments;
2098 if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2099 wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2100 (wc->block_size & (wc->block_size - 1))) {
2101 r = -EINVAL;
2102 ti->error = "Invalid block size";
2103 goto bad;
2104 }
2105 wc->block_size_bits = __ffs(wc->block_size);
2106
2107 wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2108 wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2109 wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2110
2111 /*
2112 * Parse optional arguments
2113 */
2114 r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2115 if (r)
2116 goto bad;
2117
2118 while (opt_params) {
2119 string = dm_shift_arg(&as), opt_params--;
2120 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2121 unsigned long long start_sector;
2122 string = dm_shift_arg(&as), opt_params--;
2123 if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2124 goto invalid_optional;
2125 wc->start_sector = start_sector;
2126 if (wc->start_sector != start_sector ||
2127 wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2128 goto invalid_optional;
2129 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2130 string = dm_shift_arg(&as), opt_params--;
2131 if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2132 goto invalid_optional;
2133 if (high_wm_percent < 0 || high_wm_percent > 100)
2134 goto invalid_optional;
2135 wc->high_wm_percent_set = true;
2136 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2137 string = dm_shift_arg(&as), opt_params--;
2138 if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2139 goto invalid_optional;
2140 if (low_wm_percent < 0 || low_wm_percent > 100)
2141 goto invalid_optional;
2142 wc->low_wm_percent_set = true;
2143 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2144 string = dm_shift_arg(&as), opt_params--;
2145 if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2146 goto invalid_optional;
2147 wc->max_writeback_jobs_set = true;
2148 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2149 string = dm_shift_arg(&as), opt_params--;
2150 if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2151 goto invalid_optional;
2152 wc->autocommit_blocks_set = true;
2153 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2154 unsigned autocommit_msecs;
2155 string = dm_shift_arg(&as), opt_params--;
2156 if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2157 goto invalid_optional;
2158 if (autocommit_msecs > 3600000)
2159 goto invalid_optional;
2160 wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2161 wc->autocommit_time_set = true;
2162 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2163 unsigned max_age_msecs;
2164 string = dm_shift_arg(&as), opt_params--;
2165 if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2166 goto invalid_optional;
2167 if (max_age_msecs > 86400000)
2168 goto invalid_optional;
2169 wc->max_age = msecs_to_jiffies(max_age_msecs);
2170 } else if (!strcasecmp(string, "cleaner")) {
2171 wc->cleaner = true;
2172 } else if (!strcasecmp(string, "fua")) {
2173 if (WC_MODE_PMEM(wc)) {
2174 wc->writeback_fua = true;
2175 wc->writeback_fua_set = true;
2176 } else goto invalid_optional;
2177 } else if (!strcasecmp(string, "nofua")) {
2178 if (WC_MODE_PMEM(wc)) {
2179 wc->writeback_fua = false;
2180 wc->writeback_fua_set = true;
2181 } else goto invalid_optional;
2182 } else {
2183 invalid_optional:
2184 r = -EINVAL;
2185 ti->error = "Invalid optional argument";
2186 goto bad;
2187 }
2188 }
2189
2190 if (high_wm_percent < low_wm_percent) {
2191 r = -EINVAL;
2192 ti->error = "High watermark must be greater than or equal to low watermark";
2193 goto bad;
2194 }
2195
2196 if (WC_MODE_PMEM(wc)) {
2197 r = persistent_memory_claim(wc);
2198 if (r) {
2199 ti->error = "Unable to map persistent memory for cache";
2200 goto bad;
2201 }
2202 } else {
2203 struct dm_io_region region;
2204 struct dm_io_request req;
2205 size_t n_blocks, n_metadata_blocks;
2206 uint64_t n_bitmap_bits;
2207
2208 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2209
2210 bio_list_init(&wc->flush_list);
2211 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2212 if (IS_ERR(wc->flush_thread)) {
2213 r = PTR_ERR(wc->flush_thread);
2214 wc->flush_thread = NULL;
2215 ti->error = "Couldn't spawn flush thread";
2216 goto bad;
2217 }
2218 wake_up_process(wc->flush_thread);
2219
2220 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2221 &n_blocks, &n_metadata_blocks);
2222 if (r) {
2223 ti->error = "Invalid device size";
2224 goto bad;
2225 }
2226
2227 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2228 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2229 /* this is limitation of test_bit functions */
2230 if (n_bitmap_bits > 1U << 31) {
2231 r = -EFBIG;
2232 ti->error = "Invalid device size";
2233 goto bad;
2234 }
2235
2236 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2237 if (!wc->memory_map) {
2238 r = -ENOMEM;
2239 ti->error = "Unable to allocate memory for metadata";
2240 goto bad;
2241 }
2242
2243 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2244 if (IS_ERR(wc->dm_kcopyd)) {
2245 r = PTR_ERR(wc->dm_kcopyd);
2246 ti->error = "Unable to allocate dm-kcopyd client";
2247 wc->dm_kcopyd = NULL;
2248 goto bad;
2249 }
2250
2251 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2252 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2253 BITS_PER_LONG * sizeof(unsigned long);
2254 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2255 if (!wc->dirty_bitmap) {
2256 r = -ENOMEM;
2257 ti->error = "Unable to allocate dirty bitmap";
2258 goto bad;
2259 }
2260
2261 region.bdev = wc->ssd_dev->bdev;
2262 region.sector = wc->start_sector;
2263 region.count = wc->metadata_sectors;
2264 req.bi_op = REQ_OP_READ;
2265 req.bi_op_flags = REQ_SYNC;
2266 req.mem.type = DM_IO_VMA;
2267 req.mem.ptr.vma = (char *)wc->memory_map;
2268 req.client = wc->dm_io;
2269 req.notify.fn = NULL;
2270
2271 r = dm_io(&req, 1, &region, NULL);
2272 if (r) {
2273 ti->error = "Unable to read metadata";
2274 goto bad;
2275 }
2276 }
2277
2278 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2279 if (r) {
2280 ti->error = "Hardware memory error when reading superblock";
2281 goto bad;
2282 }
2283 if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2284 r = init_memory(wc);
2285 if (r) {
2286 ti->error = "Unable to initialize device";
2287 goto bad;
2288 }
2289 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2290 if (r) {
2291 ti->error = "Hardware memory error when reading superblock";
2292 goto bad;
2293 }
2294 }
2295
2296 if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2297 ti->error = "Invalid magic in the superblock";
2298 r = -EINVAL;
2299 goto bad;
2300 }
2301
2302 if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2303 ti->error = "Invalid version in the superblock";
2304 r = -EINVAL;
2305 goto bad;
2306 }
2307
2308 if (le32_to_cpu(s.block_size) != wc->block_size) {
2309 ti->error = "Block size does not match superblock";
2310 r = -EINVAL;
2311 goto bad;
2312 }
2313
2314 wc->n_blocks = le64_to_cpu(s.n_blocks);
2315
2316 offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2317 if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2318 overflow:
2319 ti->error = "Overflow in size calculation";
2320 r = -EINVAL;
2321 goto bad;
2322 }
2323 offset += sizeof(struct wc_memory_superblock);
2324 if (offset < sizeof(struct wc_memory_superblock))
2325 goto overflow;
2326 offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2327 data_size = wc->n_blocks * (size_t)wc->block_size;
2328 if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2329 (offset + data_size < offset))
2330 goto overflow;
2331 if (offset + data_size > wc->memory_map_size) {
2332 ti->error = "Memory area is too small";
2333 r = -EINVAL;
2334 goto bad;
2335 }
2336
2337 wc->metadata_sectors = offset >> SECTOR_SHIFT;
2338 wc->block_start = (char *)sb(wc) + offset;
2339
2340 x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2341 x += 50;
2342 do_div(x, 100);
2343 wc->freelist_high_watermark = x;
2344 x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2345 x += 50;
2346 do_div(x, 100);
2347 wc->freelist_low_watermark = x;
2348
2349 if (wc->cleaner)
2350 activate_cleaner(wc);
2351
2352 r = writecache_alloc_entries(wc);
2353 if (r) {
2354 ti->error = "Cannot allocate memory";
2355 goto bad;
2356 }
2357
2358 ti->num_flush_bios = 1;
2359 ti->flush_supported = true;
2360 ti->num_discard_bios = 1;
2361
2362 if (WC_MODE_PMEM(wc))
2363 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2364
2365 return 0;
2366
2367 bad_arguments:
2368 r = -EINVAL;
2369 ti->error = "Bad arguments";
2370 bad:
2371 writecache_dtr(ti);
2372 return r;
2373 }
2374
2375 static void writecache_status(struct dm_target *ti, status_type_t type,
2376 unsigned status_flags, char *result, unsigned maxlen)
2377 {
2378 struct dm_writecache *wc = ti->private;
2379 unsigned extra_args;
2380 unsigned sz = 0;
2381 uint64_t x;
2382
2383 switch (type) {
2384 case STATUSTYPE_INFO:
2385 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2386 (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2387 (unsigned long long)wc->writeback_size);
2388 break;
2389 case STATUSTYPE_TABLE:
2390 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2391 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2392 extra_args = 0;
2393 if (wc->start_sector)
2394 extra_args += 2;
2395 if (wc->high_wm_percent_set && !wc->cleaner)
2396 extra_args += 2;
2397 if (wc->low_wm_percent_set && !wc->cleaner)
2398 extra_args += 2;
2399 if (wc->max_writeback_jobs_set)
2400 extra_args += 2;
2401 if (wc->autocommit_blocks_set)
2402 extra_args += 2;
2403 if (wc->autocommit_time_set)
2404 extra_args += 2;
2405 if (wc->cleaner)
2406 extra_args++;
2407 if (wc->writeback_fua_set)
2408 extra_args++;
2409
2410 DMEMIT("%u", extra_args);
2411 if (wc->start_sector)
2412 DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2413 if (wc->high_wm_percent_set && !wc->cleaner) {
2414 x = (uint64_t)wc->freelist_high_watermark * 100;
2415 x += wc->n_blocks / 2;
2416 do_div(x, (size_t)wc->n_blocks);
2417 DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2418 }
2419 if (wc->low_wm_percent_set && !wc->cleaner) {
2420 x = (uint64_t)wc->freelist_low_watermark * 100;
2421 x += wc->n_blocks / 2;
2422 do_div(x, (size_t)wc->n_blocks);
2423 DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2424 }
2425 if (wc->max_writeback_jobs_set)
2426 DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2427 if (wc->autocommit_blocks_set)
2428 DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2429 if (wc->autocommit_time_set)
2430 DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2431 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2432 DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2433 if (wc->cleaner)
2434 DMEMIT(" cleaner");
2435 if (wc->writeback_fua_set)
2436 DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2437 break;
2438 }
2439 }
2440
2441 static struct target_type writecache_target = {
2442 .name = "writecache",
2443 .version = {1, 3, 0},
2444 .module = THIS_MODULE,
2445 .ctr = writecache_ctr,
2446 .dtr = writecache_dtr,
2447 .status = writecache_status,
2448 .postsuspend = writecache_suspend,
2449 .resume = writecache_resume,
2450 .message = writecache_message,
2451 .map = writecache_map,
2452 .end_io = writecache_end_io,
2453 .iterate_devices = writecache_iterate_devices,
2454 .io_hints = writecache_io_hints,
2455 };
2456
2457 static int __init dm_writecache_init(void)
2458 {
2459 int r;
2460
2461 r = dm_register_target(&writecache_target);
2462 if (r < 0) {
2463 DMERR("register failed %d", r);
2464 return r;
2465 }
2466
2467 return 0;
2468 }
2469
2470 static void __exit dm_writecache_exit(void)
2471 {
2472 dm_unregister_target(&writecache_target);
2473 }
2474
2475 module_init(dm_writecache_init);
2476 module_exit(dm_writecache_exit);
2477
2478 MODULE_DESCRIPTION(DM_NAME " writecache target");
2479 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2480 MODULE_LICENSE("GPL");
Linux with added FPC-III support