1 // SPDX-License-Identifier: GPL-2.0-only
3 * Partial Parity Log for closing the RAID5 write hole
4 * Copyright (c) 2017, Intel Corporation.
7 #include <linux/kernel.h>
8 #include <linux/blkdev.h>
9 #include <linux/slab.h>
10 #include <linux/crc32c.h>
11 #include <linux/async_tx.h>
12 #include <linux/raid/md_p.h>
15 #include "raid5-log.h"
18 * PPL consists of a 4KB header (struct ppl_header) and at least 128KB for
19 * partial parity data. The header contains an array of entries
20 * (struct ppl_header_entry) which describe the logged write requests.
21 * Partial parity for the entries comes after the header, written in the same
22 * sequence as the entries:
33 * An entry describes one or more consecutive stripe_heads, up to a full
34 * stripe. The modifed raid data chunks form an m-by-n matrix, where m is the
35 * number of stripe_heads in the entry and n is the number of modified data
36 * disks. Every stripe_head in the entry must write to the same data disks.
37 * An example of a valid case described by a single entry (writes to the first
38 * stripe of a 4 disk array, 16k chunk size):
40 * sh->sector dd0 dd1 dd2 ppl
42 * 0 | --- | --- | --- | +----+
43 * 8 | -W- | -W- | --- | | pp | data_sector = 8
44 * 16 | -W- | -W- | --- | | pp | data_size = 3 * 2 * 4k
45 * 24 | -W- | -W- | --- | | pp | pp_size = 3 * 4k
46 * +-----+-----+-----+ +----+
48 * data_sector is the first raid sector of the modified data, data_size is the
49 * total size of modified data and pp_size is the size of partial parity for
50 * this entry. Entries for full stripe writes contain no partial parity
51 * (pp_size = 0), they only mark the stripes for which parity should be
52 * recalculated after an unclean shutdown. Every entry holds a checksum of its
53 * partial parity, the header also has a checksum of the header itself.
55 * A write request is always logged to the PPL instance stored on the parity
56 * disk of the corresponding stripe. For each member disk there is one ppl_log
57 * used to handle logging for this disk, independently from others. They are
58 * grouped in child_logs array in struct ppl_conf, which is assigned to
59 * r5conf->log_private.
61 * ppl_io_unit represents a full PPL write, header_page contains the ppl_header.
62 * PPL entries for logged stripes are added in ppl_log_stripe(). A stripe_head
63 * can be appended to the last entry if it meets the conditions for a valid
64 * entry described above, otherwise a new entry is added. Checksums of entries
65 * are calculated incrementally as stripes containing partial parity are being
66 * added. ppl_submit_iounit() calculates the checksum of the header and submits
67 * a bio containing the header page and partial parity pages (sh->ppl_page) for
68 * all stripes of the io_unit. When the PPL write completes, the stripes
69 * associated with the io_unit are released and raid5d starts writing their data
70 * and parity. When all stripes are written, the io_unit is freed and the next
73 * An io_unit is used to gather stripes until it is submitted or becomes full
74 * (if the maximum number of entries or size of PPL is reached). Another io_unit
75 * can't be submitted until the previous has completed (PPL and stripe
76 * data+parity is written). The log->io_list tracks all io_units of a log
77 * (for a single member disk). New io_units are added to the end of the list
78 * and the first io_unit is submitted, if it is not submitted already.
79 * The current io_unit accepting new stripes is always at the end of the list.
81 * If write-back cache is enabled for any of the disks in the array, its data
82 * must be flushed before next io_unit is submitted.
85 #define PPL_SPACE_SIZE (128 * 1024)
90 /* array of child logs, one for each raid disk */
91 struct ppl_log *child_logs;
94 int block_size; /* the logical block size used for data_sector
95 * in ppl_header_entry */
96 u32 signature; /* raid array identifier */
97 atomic64_t seq; /* current log write sequence number */
99 struct kmem_cache *io_kc;
102 struct bio_set flush_bs;
104 /* used only for recovery */
105 int recovered_entries;
108 /* stripes to retry if failed to allocate io_unit */
109 struct list_head no_mem_stripes;
110 spinlock_t no_mem_stripes_lock;
112 unsigned short write_hint;
116 struct ppl_conf *ppl_conf; /* shared between all log instances */
118 struct md_rdev *rdev; /* array member disk associated with
119 * this log instance */
120 struct mutex io_mutex;
121 struct ppl_io_unit *current_io; /* current io_unit accepting new data
122 * always at the end of io_list */
123 spinlock_t io_list_lock;
124 struct list_head io_list; /* all io_units of this log */
126 sector_t next_io_sector;
127 unsigned int entry_space;
130 unsigned long disk_flush_bitmap;
133 #define PPL_IO_INLINE_BVECS 32
138 struct page *header_page; /* for ppl_header */
140 unsigned int entries_count; /* number of entries in ppl_header */
141 unsigned int pp_size; /* total size current of partial parity */
143 u64 seq; /* sequence number of this log write */
144 struct list_head log_sibling; /* log->io_list */
146 struct list_head stripe_list; /* stripes added to the io_unit */
147 atomic_t pending_stripes; /* how many stripes not written to raid */
148 atomic_t pending_flushes; /* how many disk flushes are in progress */
150 bool submitted; /* true if write to log started */
152 /* inline bio and its biovec for submitting the iounit */
154 struct bio_vec biovec[PPL_IO_INLINE_BVECS];
157 struct dma_async_tx_descriptor *
158 ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu,
159 struct dma_async_tx_descriptor *tx)
161 int disks = sh->disks;
162 struct page **srcs = percpu->scribble;
163 int count = 0, pd_idx = sh->pd_idx, i;
164 struct async_submit_ctl submit;
166 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
169 * Partial parity is the XOR of stripe data chunks that are not changed
170 * during the write request. Depending on available data
171 * (read-modify-write vs. reconstruct-write case) we calculate it
174 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
176 * rmw: xor old data and parity from updated disks
177 * This is calculated earlier by ops_run_prexor5() so just copy
178 * the parity dev page.
180 srcs[count++] = sh->dev[pd_idx].page;
181 } else if (sh->reconstruct_state == reconstruct_state_drain_run) {
182 /* rcw: xor data from all not updated disks */
183 for (i = disks; i--;) {
184 struct r5dev *dev = &sh->dev[i];
185 if (test_bit(R5_UPTODATE, &dev->flags))
186 srcs[count++] = dev->page;
192 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, tx,
193 NULL, sh, (void *) (srcs + sh->disks + 2));
196 tx = async_memcpy(sh->ppl_page, srcs[0], 0, 0, PAGE_SIZE,
199 tx = async_xor(sh->ppl_page, srcs, 0, count, PAGE_SIZE,
205 static void *ppl_io_pool_alloc(gfp_t gfp_mask, void *pool_data)
207 struct kmem_cache *kc = pool_data;
208 struct ppl_io_unit *io;
210 io = kmem_cache_alloc(kc, gfp_mask);
214 io->header_page = alloc_page(gfp_mask);
215 if (!io->header_page) {
216 kmem_cache_free(kc, io);
223 static void ppl_io_pool_free(void *element, void *pool_data)
225 struct kmem_cache *kc = pool_data;
226 struct ppl_io_unit *io = element;
228 __free_page(io->header_page);
229 kmem_cache_free(kc, io);
232 static struct ppl_io_unit *ppl_new_iounit(struct ppl_log *log,
233 struct stripe_head *sh)
235 struct ppl_conf *ppl_conf = log->ppl_conf;
236 struct ppl_io_unit *io;
237 struct ppl_header *pplhdr;
238 struct page *header_page;
240 io = mempool_alloc(&ppl_conf->io_pool, GFP_NOWAIT);
244 header_page = io->header_page;
245 memset(io, 0, sizeof(*io));
246 io->header_page = header_page;
249 INIT_LIST_HEAD(&io->log_sibling);
250 INIT_LIST_HEAD(&io->stripe_list);
251 atomic_set(&io->pending_stripes, 0);
252 atomic_set(&io->pending_flushes, 0);
253 bio_init(&io->bio, NULL, io->biovec, PPL_IO_INLINE_BVECS, 0);
255 pplhdr = page_address(io->header_page);
257 memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
258 pplhdr->signature = cpu_to_le32(ppl_conf->signature);
260 io->seq = atomic64_add_return(1, &ppl_conf->seq);
261 pplhdr->generation = cpu_to_le64(io->seq);
266 static int ppl_log_stripe(struct ppl_log *log, struct stripe_head *sh)
268 struct ppl_io_unit *io = log->current_io;
269 struct ppl_header_entry *e = NULL;
270 struct ppl_header *pplhdr;
272 sector_t data_sector = 0;
274 struct r5conf *conf = sh->raid_conf;
276 pr_debug("%s: stripe: %llu\n", __func__, (unsigned long long)sh->sector);
278 /* check if current io_unit is full */
279 if (io && (io->pp_size == log->entry_space ||
280 io->entries_count == PPL_HDR_MAX_ENTRIES)) {
281 pr_debug("%s: add io_unit blocked by seq: %llu\n",
286 /* add a new unit if there is none or the current is full */
288 io = ppl_new_iounit(log, sh);
291 spin_lock_irq(&log->io_list_lock);
292 list_add_tail(&io->log_sibling, &log->io_list);
293 spin_unlock_irq(&log->io_list_lock);
295 log->current_io = io;
298 for (i = 0; i < sh->disks; i++) {
299 struct r5dev *dev = &sh->dev[i];
301 if (i != sh->pd_idx && test_bit(R5_Wantwrite, &dev->flags)) {
302 if (!data_disks || dev->sector < data_sector)
303 data_sector = dev->sector;
309 pr_debug("%s: seq: %llu data_sector: %llu data_disks: %d\n", __func__,
310 io->seq, (unsigned long long)data_sector, data_disks);
312 pplhdr = page_address(io->header_page);
314 if (io->entries_count > 0) {
315 struct ppl_header_entry *last =
316 &pplhdr->entries[io->entries_count - 1];
317 struct stripe_head *sh_last = list_last_entry(
318 &io->stripe_list, struct stripe_head, log_list);
319 u64 data_sector_last = le64_to_cpu(last->data_sector);
320 u32 data_size_last = le32_to_cpu(last->data_size);
323 * Check if we can append the stripe to the last entry. It must
324 * be just after the last logged stripe and write to the same
325 * disks. Use bit shift and logarithm to avoid 64-bit division.
327 if ((sh->sector == sh_last->sector + RAID5_STRIPE_SECTORS(conf)) &&
328 (data_sector >> ilog2(conf->chunk_sectors) ==
329 data_sector_last >> ilog2(conf->chunk_sectors)) &&
330 ((data_sector - data_sector_last) * data_disks ==
331 data_size_last >> 9))
336 e = &pplhdr->entries[io->entries_count++];
337 e->data_sector = cpu_to_le64(data_sector);
338 e->parity_disk = cpu_to_le32(sh->pd_idx);
339 e->checksum = cpu_to_le32(~0);
342 le32_add_cpu(&e->data_size, data_disks << PAGE_SHIFT);
344 /* don't write any PP if full stripe write */
345 if (!test_bit(STRIPE_FULL_WRITE, &sh->state)) {
346 le32_add_cpu(&e->pp_size, PAGE_SIZE);
347 io->pp_size += PAGE_SIZE;
348 e->checksum = cpu_to_le32(crc32c_le(le32_to_cpu(e->checksum),
349 page_address(sh->ppl_page),
353 list_add_tail(&sh->log_list, &io->stripe_list);
354 atomic_inc(&io->pending_stripes);
360 int ppl_write_stripe(struct r5conf *conf, struct stripe_head *sh)
362 struct ppl_conf *ppl_conf = conf->log_private;
363 struct ppl_io_unit *io = sh->ppl_io;
366 if (io || test_bit(STRIPE_SYNCING, &sh->state) || !sh->ppl_page ||
367 !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
368 !test_bit(R5_Insync, &sh->dev[sh->pd_idx].flags)) {
369 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
373 log = &ppl_conf->child_logs[sh->pd_idx];
375 mutex_lock(&log->io_mutex);
377 if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
378 mutex_unlock(&log->io_mutex);
382 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
383 clear_bit(STRIPE_DELAYED, &sh->state);
384 atomic_inc(&sh->count);
386 if (ppl_log_stripe(log, sh)) {
387 spin_lock_irq(&ppl_conf->no_mem_stripes_lock);
388 list_add_tail(&sh->log_list, &ppl_conf->no_mem_stripes);
389 spin_unlock_irq(&ppl_conf->no_mem_stripes_lock);
392 mutex_unlock(&log->io_mutex);
397 static void ppl_log_endio(struct bio *bio)
399 struct ppl_io_unit *io = bio->bi_private;
400 struct ppl_log *log = io->log;
401 struct ppl_conf *ppl_conf = log->ppl_conf;
402 struct stripe_head *sh, *next;
404 pr_debug("%s: seq: %llu\n", __func__, io->seq);
407 md_error(ppl_conf->mddev, log->rdev);
409 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
410 list_del_init(&sh->log_list);
412 set_bit(STRIPE_HANDLE, &sh->state);
413 raid5_release_stripe(sh);
417 static void ppl_submit_iounit_bio(struct ppl_io_unit *io, struct bio *bio)
419 pr_debug("%s: seq: %llu size: %u sector: %llu dev: %pg\n",
420 __func__, io->seq, bio->bi_iter.bi_size,
421 (unsigned long long)bio->bi_iter.bi_sector,
427 static void ppl_submit_iounit(struct ppl_io_unit *io)
429 struct ppl_log *log = io->log;
430 struct ppl_conf *ppl_conf = log->ppl_conf;
431 struct ppl_header *pplhdr = page_address(io->header_page);
432 struct bio *bio = &io->bio;
433 struct stripe_head *sh;
436 bio->bi_private = io;
438 if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
443 for (i = 0; i < io->entries_count; i++) {
444 struct ppl_header_entry *e = &pplhdr->entries[i];
446 pr_debug("%s: seq: %llu entry: %d data_sector: %llu pp_size: %u data_size: %u\n",
447 __func__, io->seq, i, le64_to_cpu(e->data_sector),
448 le32_to_cpu(e->pp_size), le32_to_cpu(e->data_size));
450 e->data_sector = cpu_to_le64(le64_to_cpu(e->data_sector) >>
451 ilog2(ppl_conf->block_size >> 9));
452 e->checksum = cpu_to_le32(~le32_to_cpu(e->checksum));
455 pplhdr->entries_count = cpu_to_le32(io->entries_count);
456 pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PPL_HEADER_SIZE));
458 /* Rewind the buffer if current PPL is larger then remaining space */
459 if (log->use_multippl &&
460 log->rdev->ppl.sector + log->rdev->ppl.size - log->next_io_sector <
461 (PPL_HEADER_SIZE + io->pp_size) >> 9)
462 log->next_io_sector = log->rdev->ppl.sector;
465 bio->bi_end_io = ppl_log_endio;
466 bio->bi_opf = REQ_OP_WRITE | REQ_FUA;
467 bio_set_dev(bio, log->rdev->bdev);
468 bio->bi_iter.bi_sector = log->next_io_sector;
469 bio_add_page(bio, io->header_page, PAGE_SIZE, 0);
471 pr_debug("%s: log->current_io_sector: %llu\n", __func__,
472 (unsigned long long)log->next_io_sector);
474 if (log->use_multippl)
475 log->next_io_sector += (PPL_HEADER_SIZE + io->pp_size) >> 9;
477 WARN_ON(log->disk_flush_bitmap != 0);
479 list_for_each_entry(sh, &io->stripe_list, log_list) {
480 for (i = 0; i < sh->disks; i++) {
481 struct r5dev *dev = &sh->dev[i];
483 if ((ppl_conf->child_logs[i].wb_cache_on) &&
484 (test_bit(R5_Wantwrite, &dev->flags))) {
485 set_bit(i, &log->disk_flush_bitmap);
489 /* entries for full stripe writes have no partial parity */
490 if (test_bit(STRIPE_FULL_WRITE, &sh->state))
493 if (!bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0)) {
494 struct bio *prev = bio;
496 bio = bio_alloc_bioset(prev->bi_bdev, BIO_MAX_VECS,
497 prev->bi_opf, GFP_NOIO,
499 bio->bi_iter.bi_sector = bio_end_sector(prev);
500 bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0);
502 bio_chain(bio, prev);
503 ppl_submit_iounit_bio(io, prev);
507 ppl_submit_iounit_bio(io, bio);
510 static void ppl_submit_current_io(struct ppl_log *log)
512 struct ppl_io_unit *io;
514 spin_lock_irq(&log->io_list_lock);
516 io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
518 if (io && io->submitted)
521 spin_unlock_irq(&log->io_list_lock);
524 io->submitted = true;
526 if (io == log->current_io)
527 log->current_io = NULL;
529 ppl_submit_iounit(io);
533 void ppl_write_stripe_run(struct r5conf *conf)
535 struct ppl_conf *ppl_conf = conf->log_private;
539 for (i = 0; i < ppl_conf->count; i++) {
540 log = &ppl_conf->child_logs[i];
542 mutex_lock(&log->io_mutex);
543 ppl_submit_current_io(log);
544 mutex_unlock(&log->io_mutex);
548 static void ppl_io_unit_finished(struct ppl_io_unit *io)
550 struct ppl_log *log = io->log;
551 struct ppl_conf *ppl_conf = log->ppl_conf;
552 struct r5conf *conf = ppl_conf->mddev->private;
555 pr_debug("%s: seq: %llu\n", __func__, io->seq);
557 local_irq_save(flags);
559 spin_lock(&log->io_list_lock);
560 list_del(&io->log_sibling);
561 spin_unlock(&log->io_list_lock);
563 mempool_free(io, &ppl_conf->io_pool);
565 spin_lock(&ppl_conf->no_mem_stripes_lock);
566 if (!list_empty(&ppl_conf->no_mem_stripes)) {
567 struct stripe_head *sh;
569 sh = list_first_entry(&ppl_conf->no_mem_stripes,
570 struct stripe_head, log_list);
571 list_del_init(&sh->log_list);
572 set_bit(STRIPE_HANDLE, &sh->state);
573 raid5_release_stripe(sh);
575 spin_unlock(&ppl_conf->no_mem_stripes_lock);
577 local_irq_restore(flags);
579 wake_up(&conf->wait_for_quiescent);
582 static void ppl_flush_endio(struct bio *bio)
584 struct ppl_io_unit *io = bio->bi_private;
585 struct ppl_log *log = io->log;
586 struct ppl_conf *ppl_conf = log->ppl_conf;
587 struct r5conf *conf = ppl_conf->mddev->private;
589 pr_debug("%s: dev: %pg\n", __func__, bio->bi_bdev);
591 if (bio->bi_status) {
592 struct md_rdev *rdev;
595 rdev = md_find_rdev_rcu(conf->mddev, bio_dev(bio));
597 md_error(rdev->mddev, rdev);
603 if (atomic_dec_and_test(&io->pending_flushes)) {
604 ppl_io_unit_finished(io);
605 md_wakeup_thread(conf->mddev->thread);
609 static void ppl_do_flush(struct ppl_io_unit *io)
611 struct ppl_log *log = io->log;
612 struct ppl_conf *ppl_conf = log->ppl_conf;
613 struct r5conf *conf = ppl_conf->mddev->private;
614 int raid_disks = conf->raid_disks;
615 int flushed_disks = 0;
618 atomic_set(&io->pending_flushes, raid_disks);
620 for_each_set_bit(i, &log->disk_flush_bitmap, raid_disks) {
621 struct md_rdev *rdev;
622 struct block_device *bdev = NULL;
625 rdev = rcu_dereference(conf->disks[i].rdev);
626 if (rdev && !test_bit(Faulty, &rdev->flags))
633 bio = bio_alloc_bioset(bdev, 0, GFP_NOIO,
634 REQ_OP_WRITE | REQ_PREFLUSH,
635 &ppl_conf->flush_bs);
636 bio->bi_private = io;
637 bio->bi_end_io = ppl_flush_endio;
639 pr_debug("%s: dev: %ps\n", __func__, bio->bi_bdev);
646 log->disk_flush_bitmap = 0;
648 for (i = flushed_disks ; i < raid_disks; i++) {
649 if (atomic_dec_and_test(&io->pending_flushes))
650 ppl_io_unit_finished(io);
654 static inline bool ppl_no_io_unit_submitted(struct r5conf *conf,
657 struct ppl_io_unit *io;
659 io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
662 return !io || !io->submitted;
665 void ppl_quiesce(struct r5conf *conf, int quiesce)
667 struct ppl_conf *ppl_conf = conf->log_private;
671 for (i = 0; i < ppl_conf->count; i++) {
672 struct ppl_log *log = &ppl_conf->child_logs[i];
674 spin_lock_irq(&log->io_list_lock);
675 wait_event_lock_irq(conf->wait_for_quiescent,
676 ppl_no_io_unit_submitted(conf, log),
678 spin_unlock_irq(&log->io_list_lock);
683 int ppl_handle_flush_request(struct r5l_log *log, struct bio *bio)
685 if (bio->bi_iter.bi_size == 0) {
689 bio->bi_opf &= ~REQ_PREFLUSH;
693 void ppl_stripe_write_finished(struct stripe_head *sh)
695 struct ppl_io_unit *io;
700 if (io && atomic_dec_and_test(&io->pending_stripes)) {
701 if (io->log->disk_flush_bitmap)
704 ppl_io_unit_finished(io);
708 static void ppl_xor(int size, struct page *page1, struct page *page2)
710 struct async_submit_ctl submit;
711 struct dma_async_tx_descriptor *tx;
712 struct page *xor_srcs[] = { page1, page2 };
714 init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
715 NULL, NULL, NULL, NULL);
716 tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
718 async_tx_quiesce(&tx);
722 * PPL recovery strategy: xor partial parity and data from all modified data
723 * disks within a stripe and write the result as the new stripe parity. If all
724 * stripe data disks are modified (full stripe write), no partial parity is
725 * available, so just xor the data disks.
727 * Recovery of a PPL entry shall occur only if all modified data disks are
728 * available and read from all of them succeeds.
730 * A PPL entry applies to a stripe, partial parity size for an entry is at most
731 * the size of the chunk. Examples of possible cases for a single entry:
733 * case 0: single data disk write:
734 * data0 data1 data2 ppl parity
735 * +--------+--------+--------+ +--------------------+
736 * | ------ | ------ | ------ | +----+ | (no change) |
737 * | ------ | -data- | ------ | | pp | -> | data1 ^ pp |
738 * | ------ | -data- | ------ | | pp | -> | data1 ^ pp |
739 * | ------ | ------ | ------ | +----+ | (no change) |
740 * +--------+--------+--------+ +--------------------+
741 * pp_size = data_size
743 * case 1: more than one data disk write:
744 * data0 data1 data2 ppl parity
745 * +--------+--------+--------+ +--------------------+
746 * | ------ | ------ | ------ | +----+ | (no change) |
747 * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
748 * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
749 * | ------ | ------ | ------ | +----+ | (no change) |
750 * +--------+--------+--------+ +--------------------+
751 * pp_size = data_size / modified_data_disks
753 * case 2: write to all data disks (also full stripe write):
754 * data0 data1 data2 parity
755 * +--------+--------+--------+ +--------------------+
756 * | ------ | ------ | ------ | | (no change) |
757 * | -data- | -data- | -data- | --------> | xor all data |
758 * | ------ | ------ | ------ | --------> | (no change) |
759 * | ------ | ------ | ------ | | (no change) |
760 * +--------+--------+--------+ +--------------------+
763 * The following cases are possible only in other implementations. The recovery
764 * code can handle them, but they are not generated at runtime because they can
765 * be reduced to cases 0, 1 and 2:
768 * data0 data1 data2 ppl parity
769 * +--------+--------+--------+ +----+ +--------------------+
770 * | ------ | -data- | -data- | | pp | | data1 ^ data2 ^ pp |
771 * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp |
772 * | -data- | -data- | -data- | | -- | -> | xor all data |
773 * | -data- | -data- | ------ | | pp | | data0 ^ data1 ^ pp |
774 * +--------+--------+--------+ +----+ +--------------------+
775 * pp_size = chunk_size
778 * data0 data1 data2 ppl parity
779 * +--------+--------+--------+ +----+ +--------------------+
780 * | ------ | -data- | ------ | | pp | | data1 ^ pp |
781 * | ------ | ------ | ------ | | -- | -> | (no change) |
782 * | ------ | ------ | ------ | | -- | -> | (no change) |
783 * | -data- | ------ | ------ | | pp | | data0 ^ pp |
784 * +--------+--------+--------+ +----+ +--------------------+
785 * pp_size = chunk_size
787 static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e,
790 struct ppl_conf *ppl_conf = log->ppl_conf;
791 struct mddev *mddev = ppl_conf->mddev;
792 struct r5conf *conf = mddev->private;
793 int block_size = ppl_conf->block_size;
796 sector_t r_sector_first;
797 sector_t r_sector_last;
802 char b[BDEVNAME_SIZE];
803 unsigned int pp_size = le32_to_cpu(e->pp_size);
804 unsigned int data_size = le32_to_cpu(e->data_size);
806 page1 = alloc_page(GFP_KERNEL);
807 page2 = alloc_page(GFP_KERNEL);
809 if (!page1 || !page2) {
814 r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9);
816 if ((pp_size >> 9) < conf->chunk_sectors) {
818 data_disks = data_size / pp_size;
819 strip_sectors = pp_size >> 9;
821 data_disks = conf->raid_disks - conf->max_degraded;
822 strip_sectors = (data_size >> 9) / data_disks;
824 r_sector_last = r_sector_first +
825 (data_disks - 1) * conf->chunk_sectors +
828 data_disks = conf->raid_disks - conf->max_degraded;
829 strip_sectors = conf->chunk_sectors;
830 r_sector_last = r_sector_first + (data_size >> 9);
833 pr_debug("%s: array sector first: %llu last: %llu\n", __func__,
834 (unsigned long long)r_sector_first,
835 (unsigned long long)r_sector_last);
837 /* if start and end is 4k aligned, use a 4k block */
838 if (block_size == 512 &&
839 (r_sector_first & (RAID5_STRIPE_SECTORS(conf) - 1)) == 0 &&
840 (r_sector_last & (RAID5_STRIPE_SECTORS(conf) - 1)) == 0)
841 block_size = RAID5_STRIPE_SIZE(conf);
843 /* iterate through blocks in strip */
844 for (i = 0; i < strip_sectors; i += (block_size >> 9)) {
845 bool update_parity = false;
846 sector_t parity_sector;
847 struct md_rdev *parity_rdev;
848 struct stripe_head sh;
852 pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i);
855 memset(page_address(page1), 0, PAGE_SIZE);
857 /* iterate through data member disks */
858 for (disk = 0; disk < data_disks; disk++) {
860 struct md_rdev *rdev;
862 sector_t r_sector = r_sector_first + i +
863 (disk * conf->chunk_sectors);
865 pr_debug("%s:%*s data member disk %d start\n",
866 __func__, indent, "", disk);
869 if (r_sector >= r_sector_last) {
870 pr_debug("%s:%*s array sector %llu doesn't need parity update\n",
871 __func__, indent, "",
872 (unsigned long long)r_sector);
877 update_parity = true;
879 /* map raid sector to member disk */
880 sector = raid5_compute_sector(conf, r_sector, 0,
882 pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n",
883 __func__, indent, "",
884 (unsigned long long)r_sector, dd_idx,
885 (unsigned long long)sector);
887 rdev = conf->disks[dd_idx].rdev;
888 if (!rdev || (!test_bit(In_sync, &rdev->flags) &&
889 sector >= rdev->recovery_offset)) {
890 pr_debug("%s:%*s data member disk %d missing\n",
891 __func__, indent, "", dd_idx);
892 update_parity = false;
896 pr_debug("%s:%*s reading data member disk %s sector %llu\n",
897 __func__, indent, "", bdevname(rdev->bdev, b),
898 (unsigned long long)sector);
899 if (!sync_page_io(rdev, sector, block_size, page2,
900 REQ_OP_READ, 0, false)) {
901 md_error(mddev, rdev);
902 pr_debug("%s:%*s read failed!\n", __func__,
908 ppl_xor(block_size, page1, page2);
917 pr_debug("%s:%*s reading pp disk sector %llu\n",
918 __func__, indent, "",
919 (unsigned long long)(ppl_sector + i));
920 if (!sync_page_io(log->rdev,
921 ppl_sector - log->rdev->data_offset + i,
922 block_size, page2, REQ_OP_READ, 0,
924 pr_debug("%s:%*s read failed!\n", __func__,
926 md_error(mddev, log->rdev);
931 ppl_xor(block_size, page1, page2);
934 /* map raid sector to parity disk */
935 parity_sector = raid5_compute_sector(conf, r_sector_first + i,
937 BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk));
938 parity_rdev = conf->disks[sh.pd_idx].rdev;
940 BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev);
941 pr_debug("%s:%*s write parity at sector %llu, disk %s\n",
942 __func__, indent, "",
943 (unsigned long long)parity_sector,
944 bdevname(parity_rdev->bdev, b));
945 if (!sync_page_io(parity_rdev, parity_sector, block_size,
946 page1, REQ_OP_WRITE, 0, false)) {
947 pr_debug("%s:%*s parity write error!\n", __func__,
949 md_error(mddev, parity_rdev);
962 static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr,
965 struct ppl_conf *ppl_conf = log->ppl_conf;
966 struct md_rdev *rdev = log->rdev;
967 struct mddev *mddev = rdev->mddev;
968 sector_t ppl_sector = rdev->ppl.sector + offset +
969 (PPL_HEADER_SIZE >> 9);
974 page = alloc_page(GFP_KERNEL);
978 /* iterate through all PPL entries saved */
979 for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) {
980 struct ppl_header_entry *e = &pplhdr->entries[i];
981 u32 pp_size = le32_to_cpu(e->pp_size);
982 sector_t sector = ppl_sector;
983 int ppl_entry_sectors = pp_size >> 9;
986 pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n",
987 __func__, rdev->raid_disk, i,
988 (unsigned long long)ppl_sector, pp_size);
991 crc_stored = le32_to_cpu(e->checksum);
993 /* read parial parity for this entry and calculate its checksum */
995 int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size;
997 if (!sync_page_io(rdev, sector - rdev->data_offset,
998 s, page, REQ_OP_READ, 0, false)) {
999 md_error(mddev, rdev);
1004 crc = crc32c_le(crc, page_address(page), s);
1012 if (crc != crc_stored) {
1014 * Don't recover this entry if the checksum does not
1015 * match, but keep going and try to recover other
1018 pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n",
1019 __func__, crc_stored, crc);
1020 ppl_conf->mismatch_count++;
1022 ret = ppl_recover_entry(log, e, ppl_sector);
1025 ppl_conf->recovered_entries++;
1028 ppl_sector += ppl_entry_sectors;
1031 /* flush the disk cache after recovery if necessary */
1032 ret = blkdev_issue_flush(rdev->bdev);
1038 static int ppl_write_empty_header(struct ppl_log *log)
1041 struct ppl_header *pplhdr;
1042 struct md_rdev *rdev = log->rdev;
1045 pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__,
1046 rdev->raid_disk, (unsigned long long)rdev->ppl.sector);
1048 page = alloc_page(GFP_NOIO | __GFP_ZERO);
1052 pplhdr = page_address(page);
1053 /* zero out PPL space to avoid collision with old PPLs */
1054 blkdev_issue_zeroout(rdev->bdev, rdev->ppl.sector,
1055 log->rdev->ppl.size, GFP_NOIO, 0);
1056 memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
1057 pplhdr->signature = cpu_to_le32(log->ppl_conf->signature);
1058 pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE));
1060 if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
1061 PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_SYNC |
1062 REQ_FUA, 0, false)) {
1063 md_error(rdev->mddev, rdev);
1071 static int ppl_load_distributed(struct ppl_log *log)
1073 struct ppl_conf *ppl_conf = log->ppl_conf;
1074 struct md_rdev *rdev = log->rdev;
1075 struct mddev *mddev = rdev->mddev;
1076 struct page *page, *page2;
1077 struct ppl_header *pplhdr = NULL, *prev_pplhdr = NULL;
1078 u32 crc, crc_stored;
1081 sector_t pplhdr_offset = 0, prev_pplhdr_offset = 0;
1083 pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk);
1084 /* read PPL headers, find the recent one */
1085 page = alloc_page(GFP_KERNEL);
1089 page2 = alloc_page(GFP_KERNEL);
1095 /* searching ppl area for latest ppl */
1096 while (pplhdr_offset < rdev->ppl.size - (PPL_HEADER_SIZE >> 9)) {
1097 if (!sync_page_io(rdev,
1098 rdev->ppl.sector - rdev->data_offset +
1099 pplhdr_offset, PAGE_SIZE, page, REQ_OP_READ,
1101 md_error(mddev, rdev);
1103 /* if not able to read - don't recover any PPL */
1107 pplhdr = page_address(page);
1109 /* check header validity */
1110 crc_stored = le32_to_cpu(pplhdr->checksum);
1111 pplhdr->checksum = 0;
1112 crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE);
1114 if (crc_stored != crc) {
1115 pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x (offset: %llu)\n",
1116 __func__, crc_stored, crc,
1117 (unsigned long long)pplhdr_offset);
1118 pplhdr = prev_pplhdr;
1119 pplhdr_offset = prev_pplhdr_offset;
1123 signature = le32_to_cpu(pplhdr->signature);
1125 if (mddev->external) {
1127 * For external metadata the header signature is set and
1128 * validated in userspace.
1130 ppl_conf->signature = signature;
1131 } else if (ppl_conf->signature != signature) {
1132 pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x (offset: %llu)\n",
1133 __func__, signature, ppl_conf->signature,
1134 (unsigned long long)pplhdr_offset);
1135 pplhdr = prev_pplhdr;
1136 pplhdr_offset = prev_pplhdr_offset;
1140 if (prev_pplhdr && le64_to_cpu(prev_pplhdr->generation) >
1141 le64_to_cpu(pplhdr->generation)) {
1142 /* previous was newest */
1143 pplhdr = prev_pplhdr;
1144 pplhdr_offset = prev_pplhdr_offset;
1148 prev_pplhdr_offset = pplhdr_offset;
1149 prev_pplhdr = pplhdr;
1153 /* calculate next potential ppl offset */
1154 for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++)
1156 le32_to_cpu(pplhdr->entries[i].pp_size) >> 9;
1157 pplhdr_offset += PPL_HEADER_SIZE >> 9;
1160 /* no valid ppl found */
1162 ppl_conf->mismatch_count++;
1164 pr_debug("%s: latest PPL found at offset: %llu, with generation: %llu\n",
1165 __func__, (unsigned long long)pplhdr_offset,
1166 le64_to_cpu(pplhdr->generation));
1168 /* attempt to recover from log if we are starting a dirty array */
1169 if (pplhdr && !mddev->pers && mddev->recovery_cp != MaxSector)
1170 ret = ppl_recover(log, pplhdr, pplhdr_offset);
1172 /* write empty header if we are starting the array */
1173 if (!ret && !mddev->pers)
1174 ret = ppl_write_empty_header(log);
1179 pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1180 __func__, ret, ppl_conf->mismatch_count,
1181 ppl_conf->recovered_entries);
1185 static int ppl_load(struct ppl_conf *ppl_conf)
1189 bool signature_set = false;
1192 for (i = 0; i < ppl_conf->count; i++) {
1193 struct ppl_log *log = &ppl_conf->child_logs[i];
1195 /* skip missing drive */
1199 ret = ppl_load_distributed(log);
1204 * For external metadata we can't check if the signature is
1205 * correct on a single drive, but we can check if it is the same
1208 if (ppl_conf->mddev->external) {
1209 if (!signature_set) {
1210 signature = ppl_conf->signature;
1211 signature_set = true;
1212 } else if (signature != ppl_conf->signature) {
1213 pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n",
1214 mdname(ppl_conf->mddev));
1221 pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1222 __func__, ret, ppl_conf->mismatch_count,
1223 ppl_conf->recovered_entries);
1227 static void __ppl_exit_log(struct ppl_conf *ppl_conf)
1229 clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1230 clear_bit(MD_HAS_MULTIPLE_PPLS, &ppl_conf->mddev->flags);
1232 kfree(ppl_conf->child_logs);
1234 bioset_exit(&ppl_conf->bs);
1235 bioset_exit(&ppl_conf->flush_bs);
1236 mempool_exit(&ppl_conf->io_pool);
1237 kmem_cache_destroy(ppl_conf->io_kc);
1242 void ppl_exit_log(struct r5conf *conf)
1244 struct ppl_conf *ppl_conf = conf->log_private;
1247 __ppl_exit_log(ppl_conf);
1248 conf->log_private = NULL;
1252 static int ppl_validate_rdev(struct md_rdev *rdev)
1254 char b[BDEVNAME_SIZE];
1255 int ppl_data_sectors;
1259 * The configured PPL size must be enough to store
1260 * the header and (at the very least) partial parity
1261 * for one stripe. Round it down to ensure the data
1262 * space is cleanly divisible by stripe size.
1264 ppl_data_sectors = rdev->ppl.size - (PPL_HEADER_SIZE >> 9);
1266 if (ppl_data_sectors > 0)
1267 ppl_data_sectors = rounddown(ppl_data_sectors,
1268 RAID5_STRIPE_SECTORS((struct r5conf *)rdev->mddev->private));
1270 if (ppl_data_sectors <= 0) {
1271 pr_warn("md/raid:%s: PPL space too small on %s\n",
1272 mdname(rdev->mddev), bdevname(rdev->bdev, b));
1276 ppl_size_new = ppl_data_sectors + (PPL_HEADER_SIZE >> 9);
1278 if ((rdev->ppl.sector < rdev->data_offset &&
1279 rdev->ppl.sector + ppl_size_new > rdev->data_offset) ||
1280 (rdev->ppl.sector >= rdev->data_offset &&
1281 rdev->data_offset + rdev->sectors > rdev->ppl.sector)) {
1282 pr_warn("md/raid:%s: PPL space overlaps with data on %s\n",
1283 mdname(rdev->mddev), bdevname(rdev->bdev, b));
1287 if (!rdev->mddev->external &&
1288 ((rdev->ppl.offset > 0 && rdev->ppl.offset < (rdev->sb_size >> 9)) ||
1289 (rdev->ppl.offset <= 0 && rdev->ppl.offset + ppl_size_new > 0))) {
1290 pr_warn("md/raid:%s: PPL space overlaps with superblock on %s\n",
1291 mdname(rdev->mddev), bdevname(rdev->bdev, b));
1295 rdev->ppl.size = ppl_size_new;
1300 static void ppl_init_child_log(struct ppl_log *log, struct md_rdev *rdev)
1302 struct request_queue *q;
1304 if ((rdev->ppl.size << 9) >= (PPL_SPACE_SIZE +
1305 PPL_HEADER_SIZE) * 2) {
1306 log->use_multippl = true;
1307 set_bit(MD_HAS_MULTIPLE_PPLS,
1308 &log->ppl_conf->mddev->flags);
1309 log->entry_space = PPL_SPACE_SIZE;
1311 log->use_multippl = false;
1312 log->entry_space = (log->rdev->ppl.size << 9) -
1315 log->next_io_sector = rdev->ppl.sector;
1317 q = bdev_get_queue(rdev->bdev);
1318 if (test_bit(QUEUE_FLAG_WC, &q->queue_flags))
1319 log->wb_cache_on = true;
1322 int ppl_init_log(struct r5conf *conf)
1324 struct ppl_conf *ppl_conf;
1325 struct mddev *mddev = conf->mddev;
1330 pr_debug("md/raid:%s: enabling distributed Partial Parity Log\n",
1331 mdname(conf->mddev));
1333 if (PAGE_SIZE != 4096)
1336 if (mddev->level != 5) {
1337 pr_warn("md/raid:%s PPL is not compatible with raid level %d\n",
1338 mdname(mddev), mddev->level);
1342 if (mddev->bitmap_info.file || mddev->bitmap_info.offset) {
1343 pr_warn("md/raid:%s PPL is not compatible with bitmap\n",
1348 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
1349 pr_warn("md/raid:%s PPL is not compatible with journal\n",
1354 max_disks = sizeof_field(struct ppl_log, disk_flush_bitmap) *
1356 if (conf->raid_disks > max_disks) {
1357 pr_warn("md/raid:%s PPL doesn't support over %d disks in the array\n",
1358 mdname(mddev), max_disks);
1362 ppl_conf = kzalloc(sizeof(struct ppl_conf), GFP_KERNEL);
1366 ppl_conf->mddev = mddev;
1368 ppl_conf->io_kc = KMEM_CACHE(ppl_io_unit, 0);
1369 if (!ppl_conf->io_kc) {
1374 ret = mempool_init(&ppl_conf->io_pool, conf->raid_disks, ppl_io_pool_alloc,
1375 ppl_io_pool_free, ppl_conf->io_kc);
1379 ret = bioset_init(&ppl_conf->bs, conf->raid_disks, 0, BIOSET_NEED_BVECS);
1383 ret = bioset_init(&ppl_conf->flush_bs, conf->raid_disks, 0, 0);
1387 ppl_conf->count = conf->raid_disks;
1388 ppl_conf->child_logs = kcalloc(ppl_conf->count, sizeof(struct ppl_log),
1390 if (!ppl_conf->child_logs) {
1395 atomic64_set(&ppl_conf->seq, 0);
1396 INIT_LIST_HEAD(&ppl_conf->no_mem_stripes);
1397 spin_lock_init(&ppl_conf->no_mem_stripes_lock);
1399 if (!mddev->external) {
1400 ppl_conf->signature = ~crc32c_le(~0, mddev->uuid, sizeof(mddev->uuid));
1401 ppl_conf->block_size = 512;
1403 ppl_conf->block_size = queue_logical_block_size(mddev->queue);
1406 for (i = 0; i < ppl_conf->count; i++) {
1407 struct ppl_log *log = &ppl_conf->child_logs[i];
1408 struct md_rdev *rdev = conf->disks[i].rdev;
1410 mutex_init(&log->io_mutex);
1411 spin_lock_init(&log->io_list_lock);
1412 INIT_LIST_HEAD(&log->io_list);
1414 log->ppl_conf = ppl_conf;
1418 ret = ppl_validate_rdev(rdev);
1422 ppl_init_child_log(log, rdev);
1426 /* load and possibly recover the logs from the member disks */
1427 ret = ppl_load(ppl_conf);
1431 } else if (!mddev->pers && mddev->recovery_cp == 0 &&
1432 ppl_conf->recovered_entries > 0 &&
1433 ppl_conf->mismatch_count == 0) {
1435 * If we are starting a dirty array and the recovery succeeds
1436 * without any issues, set the array as clean.
1438 mddev->recovery_cp = MaxSector;
1439 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
1440 } else if (mddev->pers && ppl_conf->mismatch_count > 0) {
1441 /* no mismatch allowed when enabling PPL for a running array */
1446 conf->log_private = ppl_conf;
1447 set_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1451 __ppl_exit_log(ppl_conf);
1455 int ppl_modify_log(struct r5conf *conf, struct md_rdev *rdev, bool add)
1457 struct ppl_conf *ppl_conf = conf->log_private;
1458 struct ppl_log *log;
1460 char b[BDEVNAME_SIZE];
1465 pr_debug("%s: disk: %d operation: %s dev: %s\n",
1466 __func__, rdev->raid_disk, add ? "add" : "remove",
1467 bdevname(rdev->bdev, b));
1469 if (rdev->raid_disk < 0)
1472 if (rdev->raid_disk >= ppl_conf->count)
1475 log = &ppl_conf->child_logs[rdev->raid_disk];
1477 mutex_lock(&log->io_mutex);
1479 ret = ppl_validate_rdev(rdev);
1482 ret = ppl_write_empty_header(log);
1483 ppl_init_child_log(log, rdev);
1488 mutex_unlock(&log->io_mutex);
1494 ppl_write_hint_show(struct mddev *mddev, char *buf)
1496 return sprintf(buf, "%d\n", 0);
1500 ppl_write_hint_store(struct mddev *mddev, const char *page, size_t len)
1502 struct r5conf *conf;
1506 if (len >= PAGE_SIZE)
1508 if (kstrtou16(page, 10, &new))
1511 err = mddev_lock(mddev);
1515 conf = mddev->private;
1518 else if (!raid5_has_ppl(conf) || !conf->log_private)
1521 mddev_unlock(mddev);
1526 struct md_sysfs_entry
1527 ppl_write_hint = __ATTR(ppl_write_hint, S_IRUGO | S_IWUSR,
1528 ppl_write_hint_show,
1529 ppl_write_hint_store);