Commit | Line | Data |
---|---|---|
3418d036 AP |
1 | /* |
2 | * Partial Parity Log for closing the RAID5 write hole | |
3 | * Copyright (c) 2017, Intel Corporation. | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify it | |
6 | * under the terms and conditions of the GNU General Public License, | |
7 | * version 2, as published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope it will be useful, but WITHOUT | |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
12 | * more details. | |
13 | */ | |
14 | ||
15 | #include <linux/kernel.h> | |
16 | #include <linux/blkdev.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/crc32c.h> | |
19 | #include <linux/flex_array.h> | |
20 | #include <linux/async_tx.h> | |
21 | #include <linux/raid/md_p.h> | |
22 | #include "md.h" | |
23 | #include "raid5.h" | |
24 | ||
25 | /* | |
26 | * PPL consists of a 4KB header (struct ppl_header) and at least 128KB for | |
27 | * partial parity data. The header contains an array of entries | |
28 | * (struct ppl_header_entry) which describe the logged write requests. | |
29 | * Partial parity for the entries comes after the header, written in the same | |
30 | * sequence as the entries: | |
31 | * | |
32 | * Header | |
33 | * entry0 | |
34 | * ... | |
35 | * entryN | |
36 | * PP data | |
37 | * PP for entry0 | |
38 | * ... | |
39 | * PP for entryN | |
40 | * | |
41 | * An entry describes one or more consecutive stripe_heads, up to a full | |
42 | * stripe. The modifed raid data chunks form an m-by-n matrix, where m is the | |
43 | * number of stripe_heads in the entry and n is the number of modified data | |
44 | * disks. Every stripe_head in the entry must write to the same data disks. | |
45 | * An example of a valid case described by a single entry (writes to the first | |
46 | * stripe of a 4 disk array, 16k chunk size): | |
47 | * | |
48 | * sh->sector dd0 dd1 dd2 ppl | |
49 | * +-----+-----+-----+ | |
50 | * 0 | --- | --- | --- | +----+ | |
51 | * 8 | -W- | -W- | --- | | pp | data_sector = 8 | |
52 | * 16 | -W- | -W- | --- | | pp | data_size = 3 * 2 * 4k | |
53 | * 24 | -W- | -W- | --- | | pp | pp_size = 3 * 4k | |
54 | * +-----+-----+-----+ +----+ | |
55 | * | |
56 | * data_sector is the first raid sector of the modified data, data_size is the | |
57 | * total size of modified data and pp_size is the size of partial parity for | |
58 | * this entry. Entries for full stripe writes contain no partial parity | |
59 | * (pp_size = 0), they only mark the stripes for which parity should be | |
60 | * recalculated after an unclean shutdown. Every entry holds a checksum of its | |
61 | * partial parity, the header also has a checksum of the header itself. | |
62 | * | |
63 | * A write request is always logged to the PPL instance stored on the parity | |
64 | * disk of the corresponding stripe. For each member disk there is one ppl_log | |
65 | * used to handle logging for this disk, independently from others. They are | |
66 | * grouped in child_logs array in struct ppl_conf, which is assigned to | |
67 | * r5conf->log_private. | |
68 | * | |
69 | * ppl_io_unit represents a full PPL write, header_page contains the ppl_header. | |
70 | * PPL entries for logged stripes are added in ppl_log_stripe(). A stripe_head | |
71 | * can be appended to the last entry if it meets the conditions for a valid | |
72 | * entry described above, otherwise a new entry is added. Checksums of entries | |
73 | * are calculated incrementally as stripes containing partial parity are being | |
74 | * added. ppl_submit_iounit() calculates the checksum of the header and submits | |
75 | * a bio containing the header page and partial parity pages (sh->ppl_page) for | |
76 | * all stripes of the io_unit. When the PPL write completes, the stripes | |
77 | * associated with the io_unit are released and raid5d starts writing their data | |
78 | * and parity. When all stripes are written, the io_unit is freed and the next | |
79 | * can be submitted. | |
80 | * | |
81 | * An io_unit is used to gather stripes until it is submitted or becomes full | |
82 | * (if the maximum number of entries or size of PPL is reached). Another io_unit | |
83 | * can't be submitted until the previous has completed (PPL and stripe | |
84 | * data+parity is written). The log->io_list tracks all io_units of a log | |
85 | * (for a single member disk). New io_units are added to the end of the list | |
86 | * and the first io_unit is submitted, if it is not submitted already. | |
87 | * The current io_unit accepting new stripes is always at the end of the list. | |
88 | */ | |
89 | ||
90 | struct ppl_conf { | |
91 | struct mddev *mddev; | |
92 | ||
93 | /* array of child logs, one for each raid disk */ | |
94 | struct ppl_log *child_logs; | |
95 | int count; | |
96 | ||
97 | int block_size; /* the logical block size used for data_sector | |
98 | * in ppl_header_entry */ | |
99 | u32 signature; /* raid array identifier */ | |
100 | atomic64_t seq; /* current log write sequence number */ | |
101 | ||
102 | struct kmem_cache *io_kc; | |
103 | mempool_t *io_pool; | |
104 | struct bio_set *bs; | |
105 | mempool_t *meta_pool; | |
4536bf9b AP |
106 | |
107 | /* used only for recovery */ | |
108 | int recovered_entries; | |
109 | int mismatch_count; | |
3418d036 AP |
110 | }; |
111 | ||
112 | struct ppl_log { | |
113 | struct ppl_conf *ppl_conf; /* shared between all log instances */ | |
114 | ||
115 | struct md_rdev *rdev; /* array member disk associated with | |
116 | * this log instance */ | |
117 | struct mutex io_mutex; | |
118 | struct ppl_io_unit *current_io; /* current io_unit accepting new data | |
119 | * always at the end of io_list */ | |
120 | spinlock_t io_list_lock; | |
121 | struct list_head io_list; /* all io_units of this log */ | |
122 | struct list_head no_mem_stripes;/* stripes to retry if failed to | |
123 | * allocate io_unit */ | |
124 | }; | |
125 | ||
126 | #define PPL_IO_INLINE_BVECS 32 | |
127 | ||
128 | struct ppl_io_unit { | |
129 | struct ppl_log *log; | |
130 | ||
131 | struct page *header_page; /* for ppl_header */ | |
132 | ||
133 | unsigned int entries_count; /* number of entries in ppl_header */ | |
134 | unsigned int pp_size; /* total size current of partial parity */ | |
135 | ||
136 | u64 seq; /* sequence number of this log write */ | |
137 | struct list_head log_sibling; /* log->io_list */ | |
138 | ||
139 | struct list_head stripe_list; /* stripes added to the io_unit */ | |
140 | atomic_t pending_stripes; /* how many stripes not written to raid */ | |
141 | ||
142 | bool submitted; /* true if write to log started */ | |
143 | ||
144 | /* inline bio and its biovec for submitting the iounit */ | |
145 | struct bio bio; | |
146 | struct bio_vec biovec[PPL_IO_INLINE_BVECS]; | |
147 | }; | |
148 | ||
149 | struct dma_async_tx_descriptor * | |
150 | ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu, | |
151 | struct dma_async_tx_descriptor *tx) | |
152 | { | |
153 | int disks = sh->disks; | |
154 | struct page **xor_srcs = flex_array_get(percpu->scribble, 0); | |
155 | int count = 0, pd_idx = sh->pd_idx, i; | |
156 | struct async_submit_ctl submit; | |
157 | ||
158 | pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector); | |
159 | ||
160 | /* | |
161 | * Partial parity is the XOR of stripe data chunks that are not changed | |
162 | * during the write request. Depending on available data | |
163 | * (read-modify-write vs. reconstruct-write case) we calculate it | |
164 | * differently. | |
165 | */ | |
166 | if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) { | |
167 | /* rmw: xor old data and parity from updated disks */ | |
168 | for (i = disks; i--;) { | |
169 | struct r5dev *dev = &sh->dev[i]; | |
170 | if (test_bit(R5_Wantdrain, &dev->flags) || i == pd_idx) | |
171 | xor_srcs[count++] = dev->page; | |
172 | } | |
173 | } else if (sh->reconstruct_state == reconstruct_state_drain_run) { | |
174 | /* rcw: xor data from all not updated disks */ | |
175 | for (i = disks; i--;) { | |
176 | struct r5dev *dev = &sh->dev[i]; | |
177 | if (test_bit(R5_UPTODATE, &dev->flags)) | |
178 | xor_srcs[count++] = dev->page; | |
179 | } | |
180 | } else { | |
181 | return tx; | |
182 | } | |
183 | ||
184 | init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, tx, | |
185 | NULL, sh, flex_array_get(percpu->scribble, 0) | |
186 | + sizeof(struct page *) * (sh->disks + 2)); | |
187 | ||
188 | if (count == 1) | |
189 | tx = async_memcpy(sh->ppl_page, xor_srcs[0], 0, 0, PAGE_SIZE, | |
190 | &submit); | |
191 | else | |
192 | tx = async_xor(sh->ppl_page, xor_srcs, 0, count, PAGE_SIZE, | |
193 | &submit); | |
194 | ||
195 | return tx; | |
196 | } | |
197 | ||
198 | static struct ppl_io_unit *ppl_new_iounit(struct ppl_log *log, | |
199 | struct stripe_head *sh) | |
200 | { | |
201 | struct ppl_conf *ppl_conf = log->ppl_conf; | |
202 | struct ppl_io_unit *io; | |
203 | struct ppl_header *pplhdr; | |
204 | ||
205 | io = mempool_alloc(ppl_conf->io_pool, GFP_ATOMIC); | |
206 | if (!io) | |
207 | return NULL; | |
208 | ||
209 | memset(io, 0, sizeof(*io)); | |
210 | io->log = log; | |
211 | INIT_LIST_HEAD(&io->log_sibling); | |
212 | INIT_LIST_HEAD(&io->stripe_list); | |
213 | atomic_set(&io->pending_stripes, 0); | |
214 | bio_init(&io->bio, io->biovec, PPL_IO_INLINE_BVECS); | |
215 | ||
216 | io->header_page = mempool_alloc(ppl_conf->meta_pool, GFP_NOIO); | |
217 | pplhdr = page_address(io->header_page); | |
218 | clear_page(pplhdr); | |
219 | memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED); | |
220 | pplhdr->signature = cpu_to_le32(ppl_conf->signature); | |
221 | ||
222 | io->seq = atomic64_add_return(1, &ppl_conf->seq); | |
223 | pplhdr->generation = cpu_to_le64(io->seq); | |
224 | ||
225 | return io; | |
226 | } | |
227 | ||
228 | static int ppl_log_stripe(struct ppl_log *log, struct stripe_head *sh) | |
229 | { | |
230 | struct ppl_io_unit *io = log->current_io; | |
231 | struct ppl_header_entry *e = NULL; | |
232 | struct ppl_header *pplhdr; | |
233 | int i; | |
234 | sector_t data_sector = 0; | |
235 | int data_disks = 0; | |
236 | unsigned int entry_space = (log->rdev->ppl.size << 9) - PPL_HEADER_SIZE; | |
237 | struct r5conf *conf = sh->raid_conf; | |
238 | ||
239 | pr_debug("%s: stripe: %llu\n", __func__, (unsigned long long)sh->sector); | |
240 | ||
241 | /* check if current io_unit is full */ | |
242 | if (io && (io->pp_size == entry_space || | |
243 | io->entries_count == PPL_HDR_MAX_ENTRIES)) { | |
244 | pr_debug("%s: add io_unit blocked by seq: %llu\n", | |
245 | __func__, io->seq); | |
246 | io = NULL; | |
247 | } | |
248 | ||
249 | /* add a new unit if there is none or the current is full */ | |
250 | if (!io) { | |
251 | io = ppl_new_iounit(log, sh); | |
252 | if (!io) | |
253 | return -ENOMEM; | |
254 | spin_lock_irq(&log->io_list_lock); | |
255 | list_add_tail(&io->log_sibling, &log->io_list); | |
256 | spin_unlock_irq(&log->io_list_lock); | |
257 | ||
258 | log->current_io = io; | |
259 | } | |
260 | ||
261 | for (i = 0; i < sh->disks; i++) { | |
262 | struct r5dev *dev = &sh->dev[i]; | |
263 | ||
264 | if (i != sh->pd_idx && test_bit(R5_Wantwrite, &dev->flags)) { | |
265 | if (!data_disks || dev->sector < data_sector) | |
266 | data_sector = dev->sector; | |
267 | data_disks++; | |
268 | } | |
269 | } | |
270 | BUG_ON(!data_disks); | |
271 | ||
272 | pr_debug("%s: seq: %llu data_sector: %llu data_disks: %d\n", __func__, | |
273 | io->seq, (unsigned long long)data_sector, data_disks); | |
274 | ||
275 | pplhdr = page_address(io->header_page); | |
276 | ||
277 | if (io->entries_count > 0) { | |
278 | struct ppl_header_entry *last = | |
279 | &pplhdr->entries[io->entries_count - 1]; | |
280 | struct stripe_head *sh_last = list_last_entry( | |
281 | &io->stripe_list, struct stripe_head, log_list); | |
282 | u64 data_sector_last = le64_to_cpu(last->data_sector); | |
283 | u32 data_size_last = le32_to_cpu(last->data_size); | |
284 | ||
285 | /* | |
286 | * Check if we can append the stripe to the last entry. It must | |
287 | * be just after the last logged stripe and write to the same | |
288 | * disks. Use bit shift and logarithm to avoid 64-bit division. | |
289 | */ | |
290 | if ((sh->sector == sh_last->sector + STRIPE_SECTORS) && | |
291 | (data_sector >> ilog2(conf->chunk_sectors) == | |
292 | data_sector_last >> ilog2(conf->chunk_sectors)) && | |
293 | ((data_sector - data_sector_last) * data_disks == | |
294 | data_size_last >> 9)) | |
295 | e = last; | |
296 | } | |
297 | ||
298 | if (!e) { | |
299 | e = &pplhdr->entries[io->entries_count++]; | |
300 | e->data_sector = cpu_to_le64(data_sector); | |
301 | e->parity_disk = cpu_to_le32(sh->pd_idx); | |
302 | e->checksum = cpu_to_le32(~0); | |
303 | } | |
304 | ||
305 | le32_add_cpu(&e->data_size, data_disks << PAGE_SHIFT); | |
306 | ||
307 | /* don't write any PP if full stripe write */ | |
308 | if (!test_bit(STRIPE_FULL_WRITE, &sh->state)) { | |
309 | le32_add_cpu(&e->pp_size, PAGE_SIZE); | |
310 | io->pp_size += PAGE_SIZE; | |
311 | e->checksum = cpu_to_le32(crc32c_le(le32_to_cpu(e->checksum), | |
312 | page_address(sh->ppl_page), | |
313 | PAGE_SIZE)); | |
314 | } | |
315 | ||
316 | list_add_tail(&sh->log_list, &io->stripe_list); | |
317 | atomic_inc(&io->pending_stripes); | |
318 | sh->ppl_io = io; | |
319 | ||
320 | return 0; | |
321 | } | |
322 | ||
323 | int ppl_write_stripe(struct r5conf *conf, struct stripe_head *sh) | |
324 | { | |
325 | struct ppl_conf *ppl_conf = conf->log_private; | |
326 | struct ppl_io_unit *io = sh->ppl_io; | |
327 | struct ppl_log *log; | |
328 | ||
329 | if (io || test_bit(STRIPE_SYNCING, &sh->state) || | |
330 | !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || | |
331 | !test_bit(R5_Insync, &sh->dev[sh->pd_idx].flags)) { | |
332 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
333 | return -EAGAIN; | |
334 | } | |
335 | ||
336 | log = &ppl_conf->child_logs[sh->pd_idx]; | |
337 | ||
338 | mutex_lock(&log->io_mutex); | |
339 | ||
340 | if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) { | |
341 | mutex_unlock(&log->io_mutex); | |
342 | return -EAGAIN; | |
343 | } | |
344 | ||
345 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
346 | clear_bit(STRIPE_DELAYED, &sh->state); | |
347 | atomic_inc(&sh->count); | |
348 | ||
349 | if (ppl_log_stripe(log, sh)) { | |
350 | spin_lock_irq(&log->io_list_lock); | |
351 | list_add_tail(&sh->log_list, &log->no_mem_stripes); | |
352 | spin_unlock_irq(&log->io_list_lock); | |
353 | } | |
354 | ||
355 | mutex_unlock(&log->io_mutex); | |
356 | ||
357 | return 0; | |
358 | } | |
359 | ||
360 | static void ppl_log_endio(struct bio *bio) | |
361 | { | |
362 | struct ppl_io_unit *io = bio->bi_private; | |
363 | struct ppl_log *log = io->log; | |
364 | struct ppl_conf *ppl_conf = log->ppl_conf; | |
365 | struct stripe_head *sh, *next; | |
366 | ||
367 | pr_debug("%s: seq: %llu\n", __func__, io->seq); | |
368 | ||
369 | if (bio->bi_error) | |
370 | md_error(ppl_conf->mddev, log->rdev); | |
371 | ||
372 | mempool_free(io->header_page, ppl_conf->meta_pool); | |
373 | ||
374 | list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { | |
375 | list_del_init(&sh->log_list); | |
376 | ||
377 | set_bit(STRIPE_HANDLE, &sh->state); | |
378 | raid5_release_stripe(sh); | |
379 | } | |
380 | } | |
381 | ||
382 | static void ppl_submit_iounit_bio(struct ppl_io_unit *io, struct bio *bio) | |
383 | { | |
384 | char b[BDEVNAME_SIZE]; | |
385 | ||
386 | pr_debug("%s: seq: %llu size: %u sector: %llu dev: %s\n", | |
387 | __func__, io->seq, bio->bi_iter.bi_size, | |
388 | (unsigned long long)bio->bi_iter.bi_sector, | |
389 | bdevname(bio->bi_bdev, b)); | |
390 | ||
391 | submit_bio(bio); | |
392 | } | |
393 | ||
394 | static void ppl_submit_iounit(struct ppl_io_unit *io) | |
395 | { | |
396 | struct ppl_log *log = io->log; | |
397 | struct ppl_conf *ppl_conf = log->ppl_conf; | |
398 | struct ppl_header *pplhdr = page_address(io->header_page); | |
399 | struct bio *bio = &io->bio; | |
400 | struct stripe_head *sh; | |
401 | int i; | |
402 | ||
6358c239 AP |
403 | bio->bi_private = io; |
404 | ||
405 | if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) { | |
406 | ppl_log_endio(bio); | |
407 | return; | |
408 | } | |
409 | ||
3418d036 AP |
410 | for (i = 0; i < io->entries_count; i++) { |
411 | struct ppl_header_entry *e = &pplhdr->entries[i]; | |
412 | ||
413 | pr_debug("%s: seq: %llu entry: %d data_sector: %llu pp_size: %u data_size: %u\n", | |
414 | __func__, io->seq, i, le64_to_cpu(e->data_sector), | |
415 | le32_to_cpu(e->pp_size), le32_to_cpu(e->data_size)); | |
416 | ||
417 | e->data_sector = cpu_to_le64(le64_to_cpu(e->data_sector) >> | |
418 | ilog2(ppl_conf->block_size >> 9)); | |
419 | e->checksum = cpu_to_le32(~le32_to_cpu(e->checksum)); | |
420 | } | |
421 | ||
422 | pplhdr->entries_count = cpu_to_le32(io->entries_count); | |
423 | pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PPL_HEADER_SIZE)); | |
424 | ||
3418d036 AP |
425 | bio->bi_end_io = ppl_log_endio; |
426 | bio->bi_opf = REQ_OP_WRITE | REQ_FUA; | |
427 | bio->bi_bdev = log->rdev->bdev; | |
428 | bio->bi_iter.bi_sector = log->rdev->ppl.sector; | |
429 | bio_add_page(bio, io->header_page, PAGE_SIZE, 0); | |
430 | ||
431 | list_for_each_entry(sh, &io->stripe_list, log_list) { | |
432 | /* entries for full stripe writes have no partial parity */ | |
433 | if (test_bit(STRIPE_FULL_WRITE, &sh->state)) | |
434 | continue; | |
435 | ||
436 | if (!bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0)) { | |
437 | struct bio *prev = bio; | |
438 | ||
439 | bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, | |
440 | ppl_conf->bs); | |
441 | bio->bi_opf = prev->bi_opf; | |
442 | bio->bi_bdev = prev->bi_bdev; | |
443 | bio->bi_iter.bi_sector = bio_end_sector(prev); | |
444 | bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0); | |
445 | ||
446 | bio_chain(bio, prev); | |
447 | ppl_submit_iounit_bio(io, prev); | |
448 | } | |
449 | } | |
450 | ||
451 | ppl_submit_iounit_bio(io, bio); | |
452 | } | |
453 | ||
454 | static void ppl_submit_current_io(struct ppl_log *log) | |
455 | { | |
456 | struct ppl_io_unit *io; | |
457 | ||
458 | spin_lock_irq(&log->io_list_lock); | |
459 | ||
460 | io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit, | |
461 | log_sibling); | |
462 | if (io && io->submitted) | |
463 | io = NULL; | |
464 | ||
465 | spin_unlock_irq(&log->io_list_lock); | |
466 | ||
467 | if (io) { | |
468 | io->submitted = true; | |
469 | ||
470 | if (io == log->current_io) | |
471 | log->current_io = NULL; | |
472 | ||
473 | ppl_submit_iounit(io); | |
474 | } | |
475 | } | |
476 | ||
477 | void ppl_write_stripe_run(struct r5conf *conf) | |
478 | { | |
479 | struct ppl_conf *ppl_conf = conf->log_private; | |
480 | struct ppl_log *log; | |
481 | int i; | |
482 | ||
483 | for (i = 0; i < ppl_conf->count; i++) { | |
484 | log = &ppl_conf->child_logs[i]; | |
485 | ||
486 | mutex_lock(&log->io_mutex); | |
487 | ppl_submit_current_io(log); | |
488 | mutex_unlock(&log->io_mutex); | |
489 | } | |
490 | } | |
491 | ||
492 | static void ppl_io_unit_finished(struct ppl_io_unit *io) | |
493 | { | |
494 | struct ppl_log *log = io->log; | |
495 | unsigned long flags; | |
496 | ||
497 | pr_debug("%s: seq: %llu\n", __func__, io->seq); | |
498 | ||
499 | spin_lock_irqsave(&log->io_list_lock, flags); | |
500 | ||
501 | list_del(&io->log_sibling); | |
502 | mempool_free(io, log->ppl_conf->io_pool); | |
503 | ||
504 | if (!list_empty(&log->no_mem_stripes)) { | |
505 | struct stripe_head *sh = list_first_entry(&log->no_mem_stripes, | |
506 | struct stripe_head, | |
507 | log_list); | |
508 | list_del_init(&sh->log_list); | |
509 | set_bit(STRIPE_HANDLE, &sh->state); | |
510 | raid5_release_stripe(sh); | |
511 | } | |
512 | ||
513 | spin_unlock_irqrestore(&log->io_list_lock, flags); | |
514 | } | |
515 | ||
516 | void ppl_stripe_write_finished(struct stripe_head *sh) | |
517 | { | |
518 | struct ppl_io_unit *io; | |
519 | ||
520 | io = sh->ppl_io; | |
521 | sh->ppl_io = NULL; | |
522 | ||
523 | if (io && atomic_dec_and_test(&io->pending_stripes)) | |
524 | ppl_io_unit_finished(io); | |
525 | } | |
526 | ||
4536bf9b AP |
527 | static void ppl_xor(int size, struct page *page1, struct page *page2) |
528 | { | |
529 | struct async_submit_ctl submit; | |
530 | struct dma_async_tx_descriptor *tx; | |
531 | struct page *xor_srcs[] = { page1, page2 }; | |
532 | ||
533 | init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST, | |
534 | NULL, NULL, NULL, NULL); | |
535 | tx = async_xor(page1, xor_srcs, 0, 2, size, &submit); | |
536 | ||
537 | async_tx_quiesce(&tx); | |
538 | } | |
539 | ||
540 | /* | |
541 | * PPL recovery strategy: xor partial parity and data from all modified data | |
542 | * disks within a stripe and write the result as the new stripe parity. If all | |
543 | * stripe data disks are modified (full stripe write), no partial parity is | |
544 | * available, so just xor the data disks. | |
545 | * | |
546 | * Recovery of a PPL entry shall occur only if all modified data disks are | |
547 | * available and read from all of them succeeds. | |
548 | * | |
549 | * A PPL entry applies to a stripe, partial parity size for an entry is at most | |
550 | * the size of the chunk. Examples of possible cases for a single entry: | |
551 | * | |
552 | * case 0: single data disk write: | |
553 | * data0 data1 data2 ppl parity | |
554 | * +--------+--------+--------+ +--------------------+ | |
555 | * | ------ | ------ | ------ | +----+ | (no change) | | |
556 | * | ------ | -data- | ------ | | pp | -> | data1 ^ pp | | |
557 | * | ------ | -data- | ------ | | pp | -> | data1 ^ pp | | |
558 | * | ------ | ------ | ------ | +----+ | (no change) | | |
559 | * +--------+--------+--------+ +--------------------+ | |
560 | * pp_size = data_size | |
561 | * | |
562 | * case 1: more than one data disk write: | |
563 | * data0 data1 data2 ppl parity | |
564 | * +--------+--------+--------+ +--------------------+ | |
565 | * | ------ | ------ | ------ | +----+ | (no change) | | |
566 | * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp | | |
567 | * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp | | |
568 | * | ------ | ------ | ------ | +----+ | (no change) | | |
569 | * +--------+--------+--------+ +--------------------+ | |
570 | * pp_size = data_size / modified_data_disks | |
571 | * | |
572 | * case 2: write to all data disks (also full stripe write): | |
573 | * data0 data1 data2 parity | |
574 | * +--------+--------+--------+ +--------------------+ | |
575 | * | ------ | ------ | ------ | | (no change) | | |
576 | * | -data- | -data- | -data- | --------> | xor all data | | |
577 | * | ------ | ------ | ------ | --------> | (no change) | | |
578 | * | ------ | ------ | ------ | | (no change) | | |
579 | * +--------+--------+--------+ +--------------------+ | |
580 | * pp_size = 0 | |
581 | * | |
582 | * The following cases are possible only in other implementations. The recovery | |
583 | * code can handle them, but they are not generated at runtime because they can | |
584 | * be reduced to cases 0, 1 and 2: | |
585 | * | |
586 | * case 3: | |
587 | * data0 data1 data2 ppl parity | |
588 | * +--------+--------+--------+ +----+ +--------------------+ | |
589 | * | ------ | -data- | -data- | | pp | | data1 ^ data2 ^ pp | | |
590 | * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp | | |
591 | * | -data- | -data- | -data- | | -- | -> | xor all data | | |
592 | * | -data- | -data- | ------ | | pp | | data0 ^ data1 ^ pp | | |
593 | * +--------+--------+--------+ +----+ +--------------------+ | |
594 | * pp_size = chunk_size | |
595 | * | |
596 | * case 4: | |
597 | * data0 data1 data2 ppl parity | |
598 | * +--------+--------+--------+ +----+ +--------------------+ | |
599 | * | ------ | -data- | ------ | | pp | | data1 ^ pp | | |
600 | * | ------ | ------ | ------ | | -- | -> | (no change) | | |
601 | * | ------ | ------ | ------ | | -- | -> | (no change) | | |
602 | * | -data- | ------ | ------ | | pp | | data0 ^ pp | | |
603 | * +--------+--------+--------+ +----+ +--------------------+ | |
604 | * pp_size = chunk_size | |
605 | */ | |
606 | static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e, | |
607 | sector_t ppl_sector) | |
608 | { | |
609 | struct ppl_conf *ppl_conf = log->ppl_conf; | |
610 | struct mddev *mddev = ppl_conf->mddev; | |
611 | struct r5conf *conf = mddev->private; | |
612 | int block_size = ppl_conf->block_size; | |
613 | struct page *page1; | |
614 | struct page *page2; | |
615 | sector_t r_sector_first; | |
616 | sector_t r_sector_last; | |
617 | int strip_sectors; | |
618 | int data_disks; | |
619 | int i; | |
620 | int ret = 0; | |
621 | char b[BDEVNAME_SIZE]; | |
622 | unsigned int pp_size = le32_to_cpu(e->pp_size); | |
623 | unsigned int data_size = le32_to_cpu(e->data_size); | |
624 | ||
625 | page1 = alloc_page(GFP_KERNEL); | |
626 | page2 = alloc_page(GFP_KERNEL); | |
627 | ||
628 | if (!page1 || !page2) { | |
629 | ret = -ENOMEM; | |
630 | goto out; | |
631 | } | |
632 | ||
633 | r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9); | |
634 | ||
635 | if ((pp_size >> 9) < conf->chunk_sectors) { | |
636 | if (pp_size > 0) { | |
637 | data_disks = data_size / pp_size; | |
638 | strip_sectors = pp_size >> 9; | |
639 | } else { | |
640 | data_disks = conf->raid_disks - conf->max_degraded; | |
641 | strip_sectors = (data_size >> 9) / data_disks; | |
642 | } | |
643 | r_sector_last = r_sector_first + | |
644 | (data_disks - 1) * conf->chunk_sectors + | |
645 | strip_sectors; | |
646 | } else { | |
647 | data_disks = conf->raid_disks - conf->max_degraded; | |
648 | strip_sectors = conf->chunk_sectors; | |
649 | r_sector_last = r_sector_first + (data_size >> 9); | |
650 | } | |
651 | ||
652 | pr_debug("%s: array sector first: %llu last: %llu\n", __func__, | |
653 | (unsigned long long)r_sector_first, | |
654 | (unsigned long long)r_sector_last); | |
655 | ||
656 | /* if start and end is 4k aligned, use a 4k block */ | |
657 | if (block_size == 512 && | |
658 | (r_sector_first & (STRIPE_SECTORS - 1)) == 0 && | |
659 | (r_sector_last & (STRIPE_SECTORS - 1)) == 0) | |
660 | block_size = STRIPE_SIZE; | |
661 | ||
662 | /* iterate through blocks in strip */ | |
663 | for (i = 0; i < strip_sectors; i += (block_size >> 9)) { | |
664 | bool update_parity = false; | |
665 | sector_t parity_sector; | |
666 | struct md_rdev *parity_rdev; | |
667 | struct stripe_head sh; | |
668 | int disk; | |
669 | int indent = 0; | |
670 | ||
671 | pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i); | |
672 | indent += 2; | |
673 | ||
674 | memset(page_address(page1), 0, PAGE_SIZE); | |
675 | ||
676 | /* iterate through data member disks */ | |
677 | for (disk = 0; disk < data_disks; disk++) { | |
678 | int dd_idx; | |
679 | struct md_rdev *rdev; | |
680 | sector_t sector; | |
681 | sector_t r_sector = r_sector_first + i + | |
682 | (disk * conf->chunk_sectors); | |
683 | ||
684 | pr_debug("%s:%*s data member disk %d start\n", | |
685 | __func__, indent, "", disk); | |
686 | indent += 2; | |
687 | ||
688 | if (r_sector >= r_sector_last) { | |
689 | pr_debug("%s:%*s array sector %llu doesn't need parity update\n", | |
690 | __func__, indent, "", | |
691 | (unsigned long long)r_sector); | |
692 | indent -= 2; | |
693 | continue; | |
694 | } | |
695 | ||
696 | update_parity = true; | |
697 | ||
698 | /* map raid sector to member disk */ | |
699 | sector = raid5_compute_sector(conf, r_sector, 0, | |
700 | &dd_idx, NULL); | |
701 | pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n", | |
702 | __func__, indent, "", | |
703 | (unsigned long long)r_sector, dd_idx, | |
704 | (unsigned long long)sector); | |
705 | ||
706 | rdev = conf->disks[dd_idx].rdev; | |
707 | if (!rdev) { | |
708 | pr_debug("%s:%*s data member disk %d missing\n", | |
709 | __func__, indent, "", dd_idx); | |
710 | update_parity = false; | |
711 | break; | |
712 | } | |
713 | ||
714 | pr_debug("%s:%*s reading data member disk %s sector %llu\n", | |
715 | __func__, indent, "", bdevname(rdev->bdev, b), | |
716 | (unsigned long long)sector); | |
717 | if (!sync_page_io(rdev, sector, block_size, page2, | |
718 | REQ_OP_READ, 0, false)) { | |
719 | md_error(mddev, rdev); | |
720 | pr_debug("%s:%*s read failed!\n", __func__, | |
721 | indent, ""); | |
722 | ret = -EIO; | |
723 | goto out; | |
724 | } | |
725 | ||
726 | ppl_xor(block_size, page1, page2); | |
727 | ||
728 | indent -= 2; | |
729 | } | |
730 | ||
731 | if (!update_parity) | |
732 | continue; | |
733 | ||
734 | if (pp_size > 0) { | |
735 | pr_debug("%s:%*s reading pp disk sector %llu\n", | |
736 | __func__, indent, "", | |
737 | (unsigned long long)(ppl_sector + i)); | |
738 | if (!sync_page_io(log->rdev, | |
739 | ppl_sector - log->rdev->data_offset + i, | |
740 | block_size, page2, REQ_OP_READ, 0, | |
741 | false)) { | |
742 | pr_debug("%s:%*s read failed!\n", __func__, | |
743 | indent, ""); | |
744 | md_error(mddev, log->rdev); | |
745 | ret = -EIO; | |
746 | goto out; | |
747 | } | |
748 | ||
749 | ppl_xor(block_size, page1, page2); | |
750 | } | |
751 | ||
752 | /* map raid sector to parity disk */ | |
753 | parity_sector = raid5_compute_sector(conf, r_sector_first + i, | |
754 | 0, &disk, &sh); | |
755 | BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk)); | |
756 | parity_rdev = conf->disks[sh.pd_idx].rdev; | |
757 | ||
758 | BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev); | |
759 | pr_debug("%s:%*s write parity at sector %llu, disk %s\n", | |
760 | __func__, indent, "", | |
761 | (unsigned long long)parity_sector, | |
762 | bdevname(parity_rdev->bdev, b)); | |
763 | if (!sync_page_io(parity_rdev, parity_sector, block_size, | |
764 | page1, REQ_OP_WRITE, 0, false)) { | |
765 | pr_debug("%s:%*s parity write error!\n", __func__, | |
766 | indent, ""); | |
767 | md_error(mddev, parity_rdev); | |
768 | ret = -EIO; | |
769 | goto out; | |
770 | } | |
771 | } | |
772 | out: | |
773 | if (page1) | |
774 | __free_page(page1); | |
775 | if (page2) | |
776 | __free_page(page2); | |
777 | return ret; | |
778 | } | |
779 | ||
780 | static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr) | |
781 | { | |
782 | struct ppl_conf *ppl_conf = log->ppl_conf; | |
783 | struct md_rdev *rdev = log->rdev; | |
784 | struct mddev *mddev = rdev->mddev; | |
785 | sector_t ppl_sector = rdev->ppl.sector + (PPL_HEADER_SIZE >> 9); | |
786 | struct page *page; | |
787 | int i; | |
788 | int ret = 0; | |
789 | ||
790 | page = alloc_page(GFP_KERNEL); | |
791 | if (!page) | |
792 | return -ENOMEM; | |
793 | ||
794 | /* iterate through all PPL entries saved */ | |
795 | for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) { | |
796 | struct ppl_header_entry *e = &pplhdr->entries[i]; | |
797 | u32 pp_size = le32_to_cpu(e->pp_size); | |
798 | sector_t sector = ppl_sector; | |
799 | int ppl_entry_sectors = pp_size >> 9; | |
800 | u32 crc, crc_stored; | |
801 | ||
802 | pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n", | |
803 | __func__, rdev->raid_disk, i, | |
804 | (unsigned long long)ppl_sector, pp_size); | |
805 | ||
806 | crc = ~0; | |
807 | crc_stored = le32_to_cpu(e->checksum); | |
808 | ||
809 | /* read parial parity for this entry and calculate its checksum */ | |
810 | while (pp_size) { | |
811 | int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size; | |
812 | ||
813 | if (!sync_page_io(rdev, sector - rdev->data_offset, | |
814 | s, page, REQ_OP_READ, 0, false)) { | |
815 | md_error(mddev, rdev); | |
816 | ret = -EIO; | |
817 | goto out; | |
818 | } | |
819 | ||
820 | crc = crc32c_le(crc, page_address(page), s); | |
821 | ||
822 | pp_size -= s; | |
823 | sector += s >> 9; | |
824 | } | |
825 | ||
826 | crc = ~crc; | |
827 | ||
828 | if (crc != crc_stored) { | |
829 | /* | |
830 | * Don't recover this entry if the checksum does not | |
831 | * match, but keep going and try to recover other | |
832 | * entries. | |
833 | */ | |
834 | pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n", | |
835 | __func__, crc_stored, crc); | |
836 | ppl_conf->mismatch_count++; | |
837 | } else { | |
838 | ret = ppl_recover_entry(log, e, ppl_sector); | |
839 | if (ret) | |
840 | goto out; | |
841 | ppl_conf->recovered_entries++; | |
842 | } | |
843 | ||
844 | ppl_sector += ppl_entry_sectors; | |
845 | } | |
846 | ||
847 | /* flush the disk cache after recovery if necessary */ | |
848 | ret = blkdev_issue_flush(rdev->bdev, GFP_KERNEL, NULL); | |
849 | out: | |
850 | __free_page(page); | |
851 | return ret; | |
852 | } | |
853 | ||
854 | static int ppl_write_empty_header(struct ppl_log *log) | |
855 | { | |
856 | struct page *page; | |
857 | struct ppl_header *pplhdr; | |
858 | struct md_rdev *rdev = log->rdev; | |
859 | int ret = 0; | |
860 | ||
861 | pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__, | |
862 | rdev->raid_disk, (unsigned long long)rdev->ppl.sector); | |
863 | ||
864 | page = alloc_page(GFP_NOIO | __GFP_ZERO); | |
865 | if (!page) | |
866 | return -ENOMEM; | |
867 | ||
868 | pplhdr = page_address(page); | |
869 | memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED); | |
870 | pplhdr->signature = cpu_to_le32(log->ppl_conf->signature); | |
871 | pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE)); | |
872 | ||
873 | if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset, | |
874 | PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_FUA, 0, | |
875 | false)) { | |
876 | md_error(rdev->mddev, rdev); | |
877 | ret = -EIO; | |
878 | } | |
879 | ||
880 | __free_page(page); | |
881 | return ret; | |
882 | } | |
883 | ||
884 | static int ppl_load_distributed(struct ppl_log *log) | |
885 | { | |
886 | struct ppl_conf *ppl_conf = log->ppl_conf; | |
887 | struct md_rdev *rdev = log->rdev; | |
888 | struct mddev *mddev = rdev->mddev; | |
889 | struct page *page; | |
890 | struct ppl_header *pplhdr; | |
891 | u32 crc, crc_stored; | |
892 | u32 signature; | |
893 | int ret = 0; | |
894 | ||
895 | pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk); | |
896 | ||
897 | /* read PPL header */ | |
898 | page = alloc_page(GFP_KERNEL); | |
899 | if (!page) | |
900 | return -ENOMEM; | |
901 | ||
902 | if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset, | |
903 | PAGE_SIZE, page, REQ_OP_READ, 0, false)) { | |
904 | md_error(mddev, rdev); | |
905 | ret = -EIO; | |
906 | goto out; | |
907 | } | |
908 | pplhdr = page_address(page); | |
909 | ||
910 | /* check header validity */ | |
911 | crc_stored = le32_to_cpu(pplhdr->checksum); | |
912 | pplhdr->checksum = 0; | |
913 | crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE); | |
914 | ||
915 | if (crc_stored != crc) { | |
916 | pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x\n", | |
917 | __func__, crc_stored, crc); | |
918 | ppl_conf->mismatch_count++; | |
919 | goto out; | |
920 | } | |
921 | ||
922 | signature = le32_to_cpu(pplhdr->signature); | |
923 | ||
924 | if (mddev->external) { | |
925 | /* | |
926 | * For external metadata the header signature is set and | |
927 | * validated in userspace. | |
928 | */ | |
929 | ppl_conf->signature = signature; | |
930 | } else if (ppl_conf->signature != signature) { | |
931 | pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x\n", | |
932 | __func__, signature, ppl_conf->signature); | |
933 | ppl_conf->mismatch_count++; | |
934 | goto out; | |
935 | } | |
936 | ||
937 | /* attempt to recover from log if we are starting a dirty array */ | |
938 | if (!mddev->pers && mddev->recovery_cp != MaxSector) | |
939 | ret = ppl_recover(log, pplhdr); | |
940 | out: | |
941 | /* write empty header if we are starting the array */ | |
942 | if (!ret && !mddev->pers) | |
943 | ret = ppl_write_empty_header(log); | |
944 | ||
945 | __free_page(page); | |
946 | ||
947 | pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n", | |
948 | __func__, ret, ppl_conf->mismatch_count, | |
949 | ppl_conf->recovered_entries); | |
950 | return ret; | |
951 | } | |
952 | ||
953 | static int ppl_load(struct ppl_conf *ppl_conf) | |
954 | { | |
955 | int ret = 0; | |
956 | u32 signature = 0; | |
957 | bool signature_set = false; | |
958 | int i; | |
959 | ||
960 | for (i = 0; i < ppl_conf->count; i++) { | |
961 | struct ppl_log *log = &ppl_conf->child_logs[i]; | |
962 | ||
963 | /* skip missing drive */ | |
964 | if (!log->rdev) | |
965 | continue; | |
966 | ||
967 | ret = ppl_load_distributed(log); | |
968 | if (ret) | |
969 | break; | |
970 | ||
971 | /* | |
972 | * For external metadata we can't check if the signature is | |
973 | * correct on a single drive, but we can check if it is the same | |
974 | * on all drives. | |
975 | */ | |
976 | if (ppl_conf->mddev->external) { | |
977 | if (!signature_set) { | |
978 | signature = ppl_conf->signature; | |
979 | signature_set = true; | |
980 | } else if (signature != ppl_conf->signature) { | |
981 | pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n", | |
982 | mdname(ppl_conf->mddev)); | |
983 | ret = -EINVAL; | |
984 | break; | |
985 | } | |
986 | } | |
987 | } | |
988 | ||
989 | pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n", | |
990 | __func__, ret, ppl_conf->mismatch_count, | |
991 | ppl_conf->recovered_entries); | |
992 | return ret; | |
993 | } | |
994 | ||
3418d036 AP |
995 | static void __ppl_exit_log(struct ppl_conf *ppl_conf) |
996 | { | |
997 | clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags); | |
998 | ||
999 | kfree(ppl_conf->child_logs); | |
1000 | ||
1001 | mempool_destroy(ppl_conf->meta_pool); | |
1002 | if (ppl_conf->bs) | |
1003 | bioset_free(ppl_conf->bs); | |
1004 | mempool_destroy(ppl_conf->io_pool); | |
1005 | kmem_cache_destroy(ppl_conf->io_kc); | |
1006 | ||
1007 | kfree(ppl_conf); | |
1008 | } | |
1009 | ||
1010 | void ppl_exit_log(struct r5conf *conf) | |
1011 | { | |
1012 | struct ppl_conf *ppl_conf = conf->log_private; | |
1013 | ||
1014 | if (ppl_conf) { | |
1015 | __ppl_exit_log(ppl_conf); | |
1016 | conf->log_private = NULL; | |
1017 | } | |
1018 | } | |
1019 | ||
1020 | static int ppl_validate_rdev(struct md_rdev *rdev) | |
1021 | { | |
1022 | char b[BDEVNAME_SIZE]; | |
1023 | int ppl_data_sectors; | |
1024 | int ppl_size_new; | |
1025 | ||
1026 | /* | |
1027 | * The configured PPL size must be enough to store | |
1028 | * the header and (at the very least) partial parity | |
1029 | * for one stripe. Round it down to ensure the data | |
1030 | * space is cleanly divisible by stripe size. | |
1031 | */ | |
1032 | ppl_data_sectors = rdev->ppl.size - (PPL_HEADER_SIZE >> 9); | |
1033 | ||
1034 | if (ppl_data_sectors > 0) | |
1035 | ppl_data_sectors = rounddown(ppl_data_sectors, STRIPE_SECTORS); | |
1036 | ||
1037 | if (ppl_data_sectors <= 0) { | |
1038 | pr_warn("md/raid:%s: PPL space too small on %s\n", | |
1039 | mdname(rdev->mddev), bdevname(rdev->bdev, b)); | |
1040 | return -ENOSPC; | |
1041 | } | |
1042 | ||
1043 | ppl_size_new = ppl_data_sectors + (PPL_HEADER_SIZE >> 9); | |
1044 | ||
1045 | if ((rdev->ppl.sector < rdev->data_offset && | |
1046 | rdev->ppl.sector + ppl_size_new > rdev->data_offset) || | |
1047 | (rdev->ppl.sector >= rdev->data_offset && | |
1048 | rdev->data_offset + rdev->sectors > rdev->ppl.sector)) { | |
1049 | pr_warn("md/raid:%s: PPL space overlaps with data on %s\n", | |
1050 | mdname(rdev->mddev), bdevname(rdev->bdev, b)); | |
1051 | return -EINVAL; | |
1052 | } | |
1053 | ||
1054 | if (!rdev->mddev->external && | |
1055 | ((rdev->ppl.offset > 0 && rdev->ppl.offset < (rdev->sb_size >> 9)) || | |
1056 | (rdev->ppl.offset <= 0 && rdev->ppl.offset + ppl_size_new > 0))) { | |
1057 | pr_warn("md/raid:%s: PPL space overlaps with superblock on %s\n", | |
1058 | mdname(rdev->mddev), bdevname(rdev->bdev, b)); | |
1059 | return -EINVAL; | |
1060 | } | |
1061 | ||
1062 | rdev->ppl.size = ppl_size_new; | |
1063 | ||
1064 | return 0; | |
1065 | } | |
1066 | ||
1067 | int ppl_init_log(struct r5conf *conf) | |
1068 | { | |
1069 | struct ppl_conf *ppl_conf; | |
1070 | struct mddev *mddev = conf->mddev; | |
1071 | int ret = 0; | |
1072 | int i; | |
1073 | bool need_cache_flush; | |
1074 | ||
1075 | pr_debug("md/raid:%s: enabling distributed Partial Parity Log\n", | |
1076 | mdname(conf->mddev)); | |
1077 | ||
1078 | if (PAGE_SIZE != 4096) | |
1079 | return -EINVAL; | |
1080 | ||
1081 | if (mddev->level != 5) { | |
1082 | pr_warn("md/raid:%s PPL is not compatible with raid level %d\n", | |
1083 | mdname(mddev), mddev->level); | |
1084 | return -EINVAL; | |
1085 | } | |
1086 | ||
1087 | if (mddev->bitmap_info.file || mddev->bitmap_info.offset) { | |
1088 | pr_warn("md/raid:%s PPL is not compatible with bitmap\n", | |
1089 | mdname(mddev)); | |
1090 | return -EINVAL; | |
1091 | } | |
1092 | ||
1093 | if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) { | |
1094 | pr_warn("md/raid:%s PPL is not compatible with journal\n", | |
1095 | mdname(mddev)); | |
1096 | return -EINVAL; | |
1097 | } | |
1098 | ||
1099 | ppl_conf = kzalloc(sizeof(struct ppl_conf), GFP_KERNEL); | |
1100 | if (!ppl_conf) | |
1101 | return -ENOMEM; | |
1102 | ||
1103 | ppl_conf->mddev = mddev; | |
1104 | ||
1105 | ppl_conf->io_kc = KMEM_CACHE(ppl_io_unit, 0); | |
1106 | if (!ppl_conf->io_kc) { | |
1107 | ret = -EINVAL; | |
1108 | goto err; | |
1109 | } | |
1110 | ||
1111 | ppl_conf->io_pool = mempool_create_slab_pool(conf->raid_disks, ppl_conf->io_kc); | |
1112 | if (!ppl_conf->io_pool) { | |
1113 | ret = -EINVAL; | |
1114 | goto err; | |
1115 | } | |
1116 | ||
1117 | ppl_conf->bs = bioset_create(conf->raid_disks, 0); | |
1118 | if (!ppl_conf->bs) { | |
1119 | ret = -EINVAL; | |
1120 | goto err; | |
1121 | } | |
1122 | ||
1123 | ppl_conf->meta_pool = mempool_create_page_pool(conf->raid_disks, 0); | |
1124 | if (!ppl_conf->meta_pool) { | |
1125 | ret = -EINVAL; | |
1126 | goto err; | |
1127 | } | |
1128 | ||
1129 | ppl_conf->count = conf->raid_disks; | |
1130 | ppl_conf->child_logs = kcalloc(ppl_conf->count, sizeof(struct ppl_log), | |
1131 | GFP_KERNEL); | |
1132 | if (!ppl_conf->child_logs) { | |
1133 | ret = -ENOMEM; | |
1134 | goto err; | |
1135 | } | |
1136 | ||
1137 | atomic64_set(&ppl_conf->seq, 0); | |
1138 | ||
1139 | if (!mddev->external) { | |
1140 | ppl_conf->signature = ~crc32c_le(~0, mddev->uuid, sizeof(mddev->uuid)); | |
1141 | ppl_conf->block_size = 512; | |
1142 | } else { | |
1143 | ppl_conf->block_size = queue_logical_block_size(mddev->queue); | |
1144 | } | |
1145 | ||
1146 | for (i = 0; i < ppl_conf->count; i++) { | |
1147 | struct ppl_log *log = &ppl_conf->child_logs[i]; | |
1148 | struct md_rdev *rdev = conf->disks[i].rdev; | |
1149 | ||
1150 | mutex_init(&log->io_mutex); | |
1151 | spin_lock_init(&log->io_list_lock); | |
1152 | INIT_LIST_HEAD(&log->io_list); | |
1153 | INIT_LIST_HEAD(&log->no_mem_stripes); | |
1154 | ||
1155 | log->ppl_conf = ppl_conf; | |
1156 | log->rdev = rdev; | |
1157 | ||
1158 | if (rdev) { | |
1159 | struct request_queue *q; | |
1160 | ||
1161 | ret = ppl_validate_rdev(rdev); | |
1162 | if (ret) | |
1163 | goto err; | |
1164 | ||
1165 | q = bdev_get_queue(rdev->bdev); | |
1166 | if (test_bit(QUEUE_FLAG_WC, &q->queue_flags)) | |
1167 | need_cache_flush = true; | |
1168 | } | |
1169 | } | |
1170 | ||
1171 | if (need_cache_flush) | |
1172 | pr_warn("md/raid:%s: Volatile write-back cache should be disabled on all member drives when using PPL!\n", | |
1173 | mdname(mddev)); | |
1174 | ||
4536bf9b AP |
1175 | /* load and possibly recover the logs from the member disks */ |
1176 | ret = ppl_load(ppl_conf); | |
1177 | ||
1178 | if (ret) { | |
1179 | goto err; | |
1180 | } else if (!mddev->pers && | |
1181 | mddev->recovery_cp == 0 && !mddev->degraded && | |
1182 | ppl_conf->recovered_entries > 0 && | |
1183 | ppl_conf->mismatch_count == 0) { | |
1184 | /* | |
1185 | * If we are starting a dirty array and the recovery succeeds | |
1186 | * without any issues, set the array as clean. | |
1187 | */ | |
1188 | mddev->recovery_cp = MaxSector; | |
1189 | set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags); | |
1190 | } | |
1191 | ||
3418d036 AP |
1192 | conf->log_private = ppl_conf; |
1193 | ||
1194 | return 0; | |
1195 | err: | |
1196 | __ppl_exit_log(ppl_conf); | |
1197 | return ret; | |
1198 | } | |
6358c239 AP |
1199 | |
1200 | int ppl_modify_log(struct r5conf *conf, struct md_rdev *rdev, bool add) | |
1201 | { | |
1202 | struct ppl_conf *ppl_conf = conf->log_private; | |
1203 | struct ppl_log *log; | |
1204 | int ret = 0; | |
1205 | char b[BDEVNAME_SIZE]; | |
1206 | ||
1207 | if (!rdev) | |
1208 | return -EINVAL; | |
1209 | ||
1210 | pr_debug("%s: disk: %d operation: %s dev: %s\n", | |
1211 | __func__, rdev->raid_disk, add ? "add" : "remove", | |
1212 | bdevname(rdev->bdev, b)); | |
1213 | ||
1214 | if (rdev->raid_disk < 0) | |
1215 | return 0; | |
1216 | ||
1217 | if (rdev->raid_disk >= ppl_conf->count) | |
1218 | return -ENODEV; | |
1219 | ||
1220 | log = &ppl_conf->child_logs[rdev->raid_disk]; | |
1221 | ||
1222 | mutex_lock(&log->io_mutex); | |
1223 | if (add) { | |
1224 | ret = ppl_validate_rdev(rdev); | |
1225 | if (!ret) { | |
1226 | log->rdev = rdev; | |
1227 | ret = ppl_write_empty_header(log); | |
1228 | } | |
1229 | } else { | |
1230 | log->rdev = NULL; | |
1231 | } | |
1232 | mutex_unlock(&log->io_mutex); | |
1233 | ||
1234 | return ret; | |
1235 | } |