Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2022 Intel Corporation.
3 : * All rights reserved.
4 : */
5 :
6 : #include "bdev_raid.h"
7 :
8 : #include "spdk/env.h"
9 : #include "spdk/thread.h"
10 : #include "spdk/string.h"
11 : #include "spdk/util.h"
12 : #include "spdk/likely.h"
13 : #include "spdk/log.h"
14 : #include "spdk/accel.h"
15 :
16 : /* Maximum concurrent full stripe writes per io channel */
17 : #define RAID5F_MAX_STRIPES 32
18 :
19 : struct chunk {
20 : /* Corresponds to base_bdev index */
21 : uint8_t index;
22 :
23 : /* Array of iovecs */
24 : struct iovec *iovs;
25 :
26 : /* Number of used iovecs */
27 : int iovcnt;
28 :
29 : /* Total number of available iovecs in the array */
30 : int iovcnt_max;
31 :
32 : /* Pointer to buffer with I/O metadata */
33 : void *md_buf;
34 : };
35 :
36 : struct stripe_request;
37 : typedef void (*stripe_req_xor_cb)(struct stripe_request *stripe_req, int status);
38 :
39 : struct stripe_request {
40 : enum stripe_request_type {
41 : STRIPE_REQ_WRITE,
42 : STRIPE_REQ_RECONSTRUCT,
43 : } type;
44 :
45 : struct raid5f_io_channel *r5ch;
46 :
47 : /* The associated raid_bdev_io */
48 : struct raid_bdev_io *raid_io;
49 :
50 : /* The stripe's index in the raid array. */
51 : uint64_t stripe_index;
52 :
53 : /* The stripe's parity chunk */
54 : struct chunk *parity_chunk;
55 :
56 : union {
57 : struct {
58 : /* Buffer for stripe parity */
59 : void *parity_buf;
60 :
61 : /* Buffer for stripe io metadata parity */
62 : void *parity_md_buf;
63 : } write;
64 :
65 : struct {
66 : /* Array of buffers for reading chunk data */
67 : void **chunk_buffers;
68 :
69 : /* Array of buffers for reading chunk metadata */
70 : void **chunk_md_buffers;
71 :
72 : /* Chunk to reconstruct from parity */
73 : struct chunk *chunk;
74 :
75 : /* Offset from chunk start */
76 : uint64_t chunk_offset;
77 : } reconstruct;
78 : };
79 :
80 : /* Array of iovec iterators for each chunk */
81 : struct spdk_ioviter *chunk_iov_iters;
82 :
83 : /* Array of source buffer pointers for parity calculation */
84 : void **chunk_xor_buffers;
85 :
86 : /* Array of source buffer pointers for parity calculation of io metadata */
87 : void **chunk_xor_md_buffers;
88 :
89 : struct {
90 : size_t len;
91 : size_t remaining;
92 : size_t remaining_md;
93 : int status;
94 : stripe_req_xor_cb cb;
95 : } xor;
96 :
97 : TAILQ_ENTRY(stripe_request) link;
98 :
99 : /* Array of chunks corresponding to base_bdevs */
100 : struct chunk chunks[0];
101 : };
102 :
103 : struct raid5f_info {
104 : /* The parent raid bdev */
105 : struct raid_bdev *raid_bdev;
106 :
107 : /* Number of data blocks in a stripe (without parity) */
108 : uint64_t stripe_blocks;
109 :
110 : /* Number of stripes on this array */
111 : uint64_t total_stripes;
112 :
113 : /* Alignment for buffer allocation */
114 : size_t buf_alignment;
115 :
116 : /* block length bit shift for optimized calculation, only valid when no interleaved md */
117 : uint32_t blocklen_shift;
118 : };
119 :
120 : struct raid5f_io_channel {
121 : /* All available stripe requests on this channel */
122 : struct {
123 : TAILQ_HEAD(, stripe_request) write;
124 : TAILQ_HEAD(, stripe_request) reconstruct;
125 : } free_stripe_requests;
126 :
127 : /* accel_fw channel */
128 : struct spdk_io_channel *accel_ch;
129 :
130 : /* For retrying xor if accel_ch runs out of resources */
131 : TAILQ_HEAD(, stripe_request) xor_retry_queue;
132 :
133 : /* For iterating over chunk iovecs during xor calculation */
134 : void **chunk_xor_buffers;
135 : struct iovec **chunk_xor_iovs;
136 : size_t *chunk_xor_iovcnt;
137 : };
138 :
139 : #define __CHUNK_IN_RANGE(req, c) \
140 : c < req->chunks + raid5f_ch_to_r5f_info(req->r5ch)->raid_bdev->num_base_bdevs
141 :
142 : #define FOR_EACH_CHUNK_FROM(req, c, from) \
143 : for (c = from; __CHUNK_IN_RANGE(req, c); c++)
144 :
145 : #define FOR_EACH_CHUNK(req, c) \
146 : FOR_EACH_CHUNK_FROM(req, c, req->chunks)
147 :
148 : #define __NEXT_DATA_CHUNK(req, c) \
149 : c == req->parity_chunk ? c+1 : c
150 :
151 : #define FOR_EACH_DATA_CHUNK(req, c) \
152 : for (c = __NEXT_DATA_CHUNK(req, req->chunks); __CHUNK_IN_RANGE(req, c); \
153 : c = __NEXT_DATA_CHUNK(req, c+1))
154 :
155 : static inline struct raid5f_info *
156 653960 : raid5f_ch_to_r5f_info(struct raid5f_io_channel *r5ch)
157 : {
158 653960 : return spdk_io_channel_get_io_device(spdk_io_channel_from_ctx(r5ch));
159 : }
160 :
161 : static inline struct stripe_request *
162 105666 : raid5f_chunk_stripe_req(struct chunk *chunk)
163 : {
164 105666 : return SPDK_CONTAINEROF((chunk - chunk->index), struct stripe_request, chunks);
165 : }
166 :
167 : static inline uint8_t
168 338831 : raid5f_stripe_data_chunks_num(const struct raid_bdev *raid_bdev)
169 : {
170 338831 : return raid_bdev->min_base_bdevs_operational;
171 : }
172 :
173 : static inline uint8_t
174 17244 : raid5f_stripe_parity_chunk_index(const struct raid_bdev *raid_bdev, uint64_t stripe_index)
175 : {
176 17244 : return raid5f_stripe_data_chunks_num(raid_bdev) - stripe_index % raid_bdev->num_base_bdevs;
177 : }
178 :
179 : static inline void
180 8604 : raid5f_stripe_request_release(struct stripe_request *stripe_req)
181 : {
182 8604 : if (spdk_likely(stripe_req->type == STRIPE_REQ_WRITE)) {
183 7605 : TAILQ_INSERT_HEAD(&stripe_req->r5ch->free_stripe_requests.write, stripe_req, link);
184 8604 : } else if (stripe_req->type == STRIPE_REQ_RECONSTRUCT) {
185 999 : TAILQ_INSERT_HEAD(&stripe_req->r5ch->free_stripe_requests.reconstruct, stripe_req, link);
186 999 : } else {
187 0 : assert(false);
188 : }
189 8604 : }
190 :
191 : static void raid5f_xor_stripe_retry(struct stripe_request *stripe_req);
192 :
193 : static void
194 8517 : raid5f_xor_stripe_done(struct stripe_request *stripe_req)
195 : {
196 8517 : struct raid5f_io_channel *r5ch = stripe_req->r5ch;
197 :
198 8517 : if (stripe_req->xor.status != 0) {
199 0 : SPDK_ERRLOG("stripe xor failed: %s\n", spdk_strerror(-stripe_req->xor.status));
200 0 : }
201 :
202 8517 : stripe_req->xor.cb(stripe_req, stripe_req->xor.status);
203 :
204 8517 : if (!TAILQ_EMPTY(&r5ch->xor_retry_queue)) {
205 0 : stripe_req = TAILQ_FIRST(&r5ch->xor_retry_queue);
206 0 : TAILQ_REMOVE(&r5ch->xor_retry_queue, stripe_req, link);
207 0 : raid5f_xor_stripe_retry(stripe_req);
208 0 : }
209 8517 : }
210 :
211 : static void raid5f_xor_stripe_continue(struct stripe_request *stripe_req);
212 :
213 : static void
214 62458 : _raid5f_xor_stripe_cb(struct stripe_request *stripe_req, int status)
215 : {
216 62458 : if (status != 0) {
217 0 : stripe_req->xor.status = status;
218 0 : }
219 :
220 62458 : if (stripe_req->xor.remaining + stripe_req->xor.remaining_md == 0) {
221 8517 : raid5f_xor_stripe_done(stripe_req);
222 8517 : }
223 62458 : }
224 :
225 : static void
226 59619 : raid5f_xor_stripe_cb(void *_stripe_req, int status)
227 : {
228 59619 : struct stripe_request *stripe_req = _stripe_req;
229 :
230 59619 : stripe_req->xor.remaining -= stripe_req->xor.len;
231 :
232 59619 : if (stripe_req->xor.remaining > 0) {
233 102204 : stripe_req->xor.len = spdk_ioviter_nextv(stripe_req->chunk_iov_iters,
234 51102 : stripe_req->r5ch->chunk_xor_buffers);
235 51102 : raid5f_xor_stripe_continue(stripe_req);
236 51102 : }
237 :
238 59619 : _raid5f_xor_stripe_cb(stripe_req, status);
239 59619 : }
240 :
241 : static void
242 2839 : raid5f_xor_stripe_md_cb(void *_stripe_req, int status)
243 : {
244 2839 : struct stripe_request *stripe_req = _stripe_req;
245 :
246 2839 : stripe_req->xor.remaining_md = 0;
247 :
248 2839 : _raid5f_xor_stripe_cb(stripe_req, status);
249 2839 : }
250 :
251 : static void
252 59619 : raid5f_xor_stripe_continue(struct stripe_request *stripe_req)
253 : {
254 59619 : struct raid5f_io_channel *r5ch = stripe_req->r5ch;
255 59619 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
256 59619 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
257 59619 : uint8_t n_src = raid5f_stripe_data_chunks_num(raid_bdev);
258 59619 : uint8_t i;
259 59619 : int ret;
260 :
261 59619 : assert(stripe_req->xor.len > 0);
262 :
263 256494 : for (i = 0; i < n_src; i++) {
264 196875 : stripe_req->chunk_xor_buffers[i] = r5ch->chunk_xor_buffers[i];
265 196875 : }
266 :
267 119238 : ret = spdk_accel_submit_xor(r5ch->accel_ch, r5ch->chunk_xor_buffers[n_src],
268 59619 : stripe_req->chunk_xor_buffers, n_src, stripe_req->xor.len,
269 59619 : raid5f_xor_stripe_cb, stripe_req);
270 59619 : if (spdk_unlikely(ret)) {
271 0 : if (ret == -ENOMEM) {
272 0 : TAILQ_INSERT_HEAD(&r5ch->xor_retry_queue, stripe_req, link);
273 0 : } else {
274 0 : stripe_req->xor.status = ret;
275 0 : raid5f_xor_stripe_done(stripe_req);
276 : }
277 0 : }
278 59619 : }
279 :
280 : static void
281 8517 : raid5f_xor_stripe(struct stripe_request *stripe_req, stripe_req_xor_cb cb)
282 : {
283 8517 : struct raid5f_io_channel *r5ch = stripe_req->r5ch;
284 8517 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
285 8517 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
286 8517 : struct chunk *chunk;
287 8517 : struct chunk *dest_chunk = NULL;
288 8517 : uint64_t num_blocks = 0;
289 8517 : uint8_t c;
290 :
291 8517 : assert(cb != NULL);
292 :
293 8517 : if (spdk_likely(stripe_req->type == STRIPE_REQ_WRITE)) {
294 7518 : num_blocks = raid_bdev->strip_size;
295 7518 : dest_chunk = stripe_req->parity_chunk;
296 8517 : } else if (stripe_req->type == STRIPE_REQ_RECONSTRUCT) {
297 999 : num_blocks = raid_io->num_blocks;
298 999 : dest_chunk = stripe_req->reconstruct.chunk;
299 999 : } else {
300 0 : assert(false);
301 : }
302 :
303 8517 : c = 0;
304 45159 : FOR_EACH_CHUNK(stripe_req, chunk) {
305 36642 : if (chunk == dest_chunk) {
306 8517 : continue;
307 : }
308 28125 : r5ch->chunk_xor_iovs[c] = chunk->iovs;
309 28125 : r5ch->chunk_xor_iovcnt[c] = chunk->iovcnt;
310 28125 : c++;
311 28125 : }
312 8517 : r5ch->chunk_xor_iovs[c] = dest_chunk->iovs;
313 8517 : r5ch->chunk_xor_iovcnt[c] = dest_chunk->iovcnt;
314 :
315 17034 : stripe_req->xor.len = spdk_ioviter_firstv(stripe_req->chunk_iov_iters,
316 8517 : raid_bdev->num_base_bdevs,
317 8517 : r5ch->chunk_xor_iovs,
318 8517 : r5ch->chunk_xor_iovcnt,
319 8517 : r5ch->chunk_xor_buffers);
320 8517 : stripe_req->xor.remaining = num_blocks * raid_bdev->bdev.blocklen;
321 8517 : stripe_req->xor.status = 0;
322 8517 : stripe_req->xor.cb = cb;
323 :
324 8517 : if (raid_io->md_buf != NULL) {
325 2839 : uint8_t n_src = raid5f_stripe_data_chunks_num(raid_bdev);
326 2839 : uint64_t len = num_blocks * raid_bdev->bdev.md_len;
327 2839 : int ret;
328 :
329 2839 : stripe_req->xor.remaining_md = len;
330 :
331 2839 : c = 0;
332 15053 : FOR_EACH_CHUNK(stripe_req, chunk) {
333 12214 : if (chunk != dest_chunk) {
334 9375 : stripe_req->chunk_xor_md_buffers[c] = chunk->md_buf;
335 9375 : c++;
336 9375 : }
337 12214 : }
338 :
339 5678 : ret = spdk_accel_submit_xor(stripe_req->r5ch->accel_ch, dest_chunk->md_buf,
340 2839 : stripe_req->chunk_xor_md_buffers, n_src, len,
341 2839 : raid5f_xor_stripe_md_cb, stripe_req);
342 2839 : if (spdk_unlikely(ret)) {
343 0 : if (ret == -ENOMEM) {
344 0 : TAILQ_INSERT_HEAD(&stripe_req->r5ch->xor_retry_queue, stripe_req, link);
345 0 : } else {
346 0 : stripe_req->xor.status = ret;
347 0 : raid5f_xor_stripe_done(stripe_req);
348 : }
349 0 : return;
350 : }
351 2839 : }
352 :
353 8517 : raid5f_xor_stripe_continue(stripe_req);
354 8517 : }
355 :
356 : static void
357 0 : raid5f_xor_stripe_retry(struct stripe_request *stripe_req)
358 : {
359 0 : if (stripe_req->xor.remaining_md) {
360 0 : raid5f_xor_stripe(stripe_req, stripe_req->xor.cb);
361 0 : } else {
362 0 : raid5f_xor_stripe_continue(stripe_req);
363 : }
364 0 : }
365 :
366 : static void
367 27216 : raid5f_stripe_request_chunk_write_complete(struct stripe_request *stripe_req,
368 : enum spdk_bdev_io_status status)
369 : {
370 27216 : if (raid_bdev_io_complete_part(stripe_req->raid_io, 1, status)) {
371 7239 : raid5f_stripe_request_release(stripe_req);
372 7239 : }
373 27216 : }
374 :
375 : static void
376 3213 : raid5f_stripe_request_chunk_read_complete(struct stripe_request *stripe_req,
377 : enum spdk_bdev_io_status status)
378 : {
379 3213 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
380 :
381 3213 : raid_bdev_io_complete_part(raid_io, 1, status);
382 3213 : }
383 :
384 : static void
385 30429 : raid5f_chunk_complete_bdev_io(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg)
386 : {
387 30429 : struct chunk *chunk = cb_arg;
388 30429 : struct stripe_request *stripe_req = raid5f_chunk_stripe_req(chunk);
389 30429 : enum spdk_bdev_io_status status = success ? SPDK_BDEV_IO_STATUS_SUCCESS :
390 : SPDK_BDEV_IO_STATUS_FAILED;
391 :
392 30429 : spdk_bdev_free_io(bdev_io);
393 :
394 30429 : if (spdk_likely(stripe_req->type == STRIPE_REQ_WRITE)) {
395 27216 : raid5f_stripe_request_chunk_write_complete(stripe_req, status);
396 30429 : } else if (stripe_req->type == STRIPE_REQ_RECONSTRUCT) {
397 3213 : raid5f_stripe_request_chunk_read_complete(stripe_req, status);
398 3213 : } else {
399 0 : assert(false);
400 : }
401 30429 : }
402 :
403 : static void raid5f_stripe_request_submit_chunks(struct stripe_request *stripe_req);
404 :
405 : static void
406 3831 : raid5f_chunk_submit_retry(void *_raid_io)
407 : {
408 3831 : struct raid_bdev_io *raid_io = _raid_io;
409 3831 : struct stripe_request *stripe_req = raid_io->module_private;
410 :
411 3831 : raid5f_stripe_request_submit_chunks(stripe_req);
412 3831 : }
413 :
414 : static inline void
415 46575 : raid5f_init_ext_io_opts(struct spdk_bdev_ext_io_opts *opts, struct raid_bdev_io *raid_io)
416 : {
417 46575 : memset(opts, 0, sizeof(*opts));
418 46575 : opts->size = sizeof(*opts);
419 46575 : opts->memory_domain = raid_io->memory_domain;
420 46575 : opts->memory_domain_ctx = raid_io->memory_domain_ctx;
421 46575 : opts->metadata = raid_io->md_buf;
422 46575 : }
423 :
424 : static int
425 38301 : raid5f_chunk_submit(struct chunk *chunk)
426 : {
427 38301 : struct stripe_request *stripe_req = raid5f_chunk_stripe_req(chunk);
428 38301 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
429 38301 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
430 38301 : struct raid_base_bdev_info *base_info = &raid_bdev->base_bdev_info[chunk->index];
431 76602 : struct spdk_io_channel *base_ch = raid_bdev_channel_get_base_channel(raid_io->raid_ch,
432 38301 : chunk->index);
433 38301 : uint64_t base_offset_blocks = (stripe_req->stripe_index << raid_bdev->strip_size_shift);
434 38301 : struct spdk_bdev_ext_io_opts io_opts;
435 38301 : int ret;
436 :
437 38301 : raid5f_init_ext_io_opts(&io_opts, raid_io);
438 38301 : io_opts.metadata = chunk->md_buf;
439 :
440 38301 : raid_io->base_bdev_io_submitted++;
441 :
442 38301 : switch (stripe_req->type) {
443 : case STRIPE_REQ_WRITE:
444 34089 : if (base_ch == NULL) {
445 366 : raid_bdev_io_complete_part(raid_io, 1, SPDK_BDEV_IO_STATUS_SUCCESS);
446 366 : return 0;
447 : }
448 :
449 67446 : ret = raid_bdev_writev_blocks_ext(base_info, base_ch, chunk->iovs, chunk->iovcnt,
450 33723 : base_offset_blocks, raid_bdev->strip_size,
451 33723 : raid5f_chunk_complete_bdev_io, chunk, &io_opts);
452 33723 : break;
453 : case STRIPE_REQ_RECONSTRUCT:
454 4212 : if (chunk == stripe_req->reconstruct.chunk) {
455 999 : raid_bdev_io_complete_part(raid_io, 1, SPDK_BDEV_IO_STATUS_SUCCESS);
456 999 : return 0;
457 : }
458 :
459 3213 : base_offset_blocks += stripe_req->reconstruct.chunk_offset;
460 :
461 6426 : ret = raid_bdev_readv_blocks_ext(base_info, base_ch, chunk->iovs, chunk->iovcnt,
462 3213 : base_offset_blocks, raid_io->num_blocks,
463 3213 : raid5f_chunk_complete_bdev_io, chunk, &io_opts);
464 3213 : break;
465 : default:
466 0 : assert(false);
467 : ret = -EINVAL;
468 : break;
469 : }
470 :
471 36936 : if (spdk_unlikely(ret)) {
472 6507 : raid_io->base_bdev_io_submitted--;
473 6507 : if (ret == -ENOMEM) {
474 3831 : raid_bdev_queue_io_wait(raid_io, spdk_bdev_desc_get_bdev(base_info->desc),
475 3831 : base_ch, raid5f_chunk_submit_retry);
476 3831 : } else {
477 : /*
478 : * Implicitly complete any I/Os not yet submitted as FAILED. If completing
479 : * these means there are no more to complete for the stripe request, we can
480 : * release the stripe request as well.
481 : */
482 2676 : uint64_t base_bdev_io_not_submitted;
483 :
484 2676 : if (stripe_req->type == STRIPE_REQ_WRITE) {
485 5352 : base_bdev_io_not_submitted = raid_bdev->num_base_bdevs -
486 2676 : raid_io->base_bdev_io_submitted;
487 2676 : } else {
488 0 : base_bdev_io_not_submitted = raid5f_stripe_data_chunks_num(raid_bdev) -
489 0 : raid_io->base_bdev_io_submitted;
490 : }
491 :
492 2676 : if (raid_bdev_io_complete_part(raid_io, base_bdev_io_not_submitted,
493 : SPDK_BDEV_IO_STATUS_FAILED)) {
494 366 : raid5f_stripe_request_release(stripe_req);
495 366 : }
496 2676 : }
497 6507 : }
498 :
499 36936 : return ret;
500 38301 : }
501 :
502 : static int
503 26466 : raid5f_chunk_set_iovcnt(struct chunk *chunk, int iovcnt)
504 : {
505 26466 : if (iovcnt > chunk->iovcnt_max) {
506 99 : struct iovec *iovs = chunk->iovs;
507 :
508 99 : iovs = realloc(iovs, iovcnt * sizeof(*iovs));
509 99 : if (!iovs) {
510 0 : return -ENOMEM;
511 : }
512 99 : chunk->iovs = iovs;
513 99 : chunk->iovcnt_max = iovcnt;
514 99 : }
515 26466 : chunk->iovcnt = iovcnt;
516 :
517 26466 : return 0;
518 26466 : }
519 :
520 : static int
521 7704 : raid5f_stripe_request_map_iovecs(struct stripe_request *stripe_req)
522 : {
523 7704 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
524 7704 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
525 7704 : struct raid5f_info *r5f_info = raid_bdev->module_private;
526 7704 : struct chunk *chunk;
527 7704 : int raid_io_iov_idx = 0;
528 7704 : size_t raid_io_offset = 0;
529 7704 : size_t raid_io_iov_offset = 0;
530 7704 : int i;
531 :
532 33171 : FOR_EACH_DATA_CHUNK(stripe_req, chunk) {
533 25467 : int chunk_iovcnt = 0;
534 25467 : uint64_t len = raid_bdev->strip_size * raid_bdev->bdev.blocklen;
535 25467 : size_t off = raid_io_iov_offset;
536 25467 : int ret;
537 :
538 71229 : for (i = raid_io_iov_idx; i < raid_io->iovcnt; i++) {
539 71229 : chunk_iovcnt++;
540 71229 : off += raid_io->iovs[i].iov_len;
541 71229 : if (off >= raid_io_offset + len) {
542 25467 : break;
543 : }
544 45762 : }
545 :
546 25467 : assert(raid_io_iov_idx + chunk_iovcnt <= raid_io->iovcnt);
547 :
548 25467 : ret = raid5f_chunk_set_iovcnt(chunk, chunk_iovcnt);
549 25467 : if (ret) {
550 0 : return ret;
551 : }
552 :
553 25467 : if (raid_io->md_buf != NULL) {
554 16780 : chunk->md_buf = raid_io->md_buf +
555 8390 : (raid_io_offset >> r5f_info->blocklen_shift) * raid_bdev->bdev.md_len;
556 8390 : }
557 :
558 96696 : for (i = 0; i < chunk_iovcnt; i++) {
559 71229 : struct iovec *chunk_iov = &chunk->iovs[i];
560 71229 : const struct iovec *raid_io_iov = &raid_io->iovs[raid_io_iov_idx];
561 71229 : size_t chunk_iov_offset = raid_io_offset - raid_io_iov_offset;
562 :
563 71229 : chunk_iov->iov_base = raid_io_iov->iov_base + chunk_iov_offset;
564 71229 : chunk_iov->iov_len = spdk_min(len, raid_io_iov->iov_len - chunk_iov_offset);
565 71229 : raid_io_offset += chunk_iov->iov_len;
566 71229 : len -= chunk_iov->iov_len;
567 :
568 71229 : if (raid_io_offset >= raid_io_iov_offset + raid_io_iov->iov_len) {
569 53466 : raid_io_iov_idx++;
570 53466 : raid_io_iov_offset += raid_io_iov->iov_len;
571 53466 : }
572 71229 : }
573 :
574 25467 : if (spdk_unlikely(len > 0)) {
575 0 : return -EINVAL;
576 : }
577 25467 : }
578 :
579 7704 : stripe_req->parity_chunk->iovs[0].iov_base = stripe_req->write.parity_buf;
580 7704 : stripe_req->parity_chunk->iovs[0].iov_len = raid_bdev->strip_size * raid_bdev->bdev.blocklen;
581 7704 : stripe_req->parity_chunk->iovcnt = 1;
582 7704 : stripe_req->parity_chunk->md_buf = stripe_req->write.parity_md_buf;
583 :
584 7704 : return 0;
585 7704 : }
586 :
587 : static void
588 12435 : raid5f_stripe_request_submit_chunks(struct stripe_request *stripe_req)
589 : {
590 12435 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
591 12435 : struct chunk *start = &stripe_req->chunks[raid_io->base_bdev_io_submitted];
592 12435 : struct chunk *chunk;
593 :
594 44229 : FOR_EACH_CHUNK_FROM(stripe_req, chunk, start) {
595 38301 : if (spdk_unlikely(raid5f_chunk_submit(chunk) != 0)) {
596 6507 : break;
597 : }
598 31794 : }
599 12435 : }
600 :
601 : static inline void
602 8604 : raid5f_stripe_request_init(struct stripe_request *stripe_req, struct raid_bdev_io *raid_io,
603 : uint64_t stripe_index)
604 : {
605 8604 : stripe_req->raid_io = raid_io;
606 8604 : stripe_req->stripe_index = stripe_index;
607 8604 : stripe_req->parity_chunk = &stripe_req->chunks[raid5f_stripe_parity_chunk_index(raid_io->raid_bdev,
608 8604 : stripe_index)];
609 8604 : }
610 :
611 : static void
612 7605 : raid5f_stripe_write_request_xor_done(struct stripe_request *stripe_req, int status)
613 : {
614 7605 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
615 :
616 7605 : if (status != 0) {
617 0 : raid5f_stripe_request_release(stripe_req);
618 0 : raid_bdev_io_complete(raid_io, SPDK_BDEV_IO_STATUS_FAILED);
619 0 : } else {
620 7605 : raid5f_stripe_request_submit_chunks(stripe_req);
621 : }
622 7605 : }
623 :
624 : static int
625 7605 : raid5f_submit_write_request(struct raid_bdev_io *raid_io, uint64_t stripe_index)
626 : {
627 7605 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
628 7605 : struct raid5f_io_channel *r5ch = raid_bdev_channel_get_module_ctx(raid_io->raid_ch);
629 7605 : struct stripe_request *stripe_req;
630 7605 : int ret;
631 :
632 7605 : stripe_req = TAILQ_FIRST(&r5ch->free_stripe_requests.write);
633 7605 : if (!stripe_req) {
634 0 : return -ENOMEM;
635 : }
636 :
637 7605 : raid5f_stripe_request_init(stripe_req, raid_io, stripe_index);
638 :
639 7605 : ret = raid5f_stripe_request_map_iovecs(stripe_req);
640 7605 : if (spdk_unlikely(ret)) {
641 0 : return ret;
642 : }
643 :
644 7605 : TAILQ_REMOVE(&r5ch->free_stripe_requests.write, stripe_req, link);
645 :
646 7605 : raid_io->module_private = stripe_req;
647 7605 : raid_io->base_bdev_io_remaining = raid_bdev->num_base_bdevs;
648 :
649 7605 : if (raid_bdev_channel_get_base_channel(raid_io->raid_ch, stripe_req->parity_chunk->index) != NULL) {
650 7518 : raid5f_xor_stripe(stripe_req, raid5f_stripe_write_request_xor_done);
651 7518 : } else {
652 87 : raid5f_stripe_write_request_xor_done(stripe_req, 0);
653 : }
654 :
655 7605 : return 0;
656 7605 : }
657 :
658 : static void
659 7275 : raid5f_chunk_read_complete(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg)
660 : {
661 7275 : struct raid_bdev_io *raid_io = cb_arg;
662 :
663 7275 : spdk_bdev_free_io(bdev_io);
664 :
665 7275 : raid_bdev_io_complete(raid_io, success ? SPDK_BDEV_IO_STATUS_SUCCESS :
666 : SPDK_BDEV_IO_STATUS_FAILED);
667 7275 : }
668 :
669 : static void raid5f_submit_rw_request(struct raid_bdev_io *raid_io);
670 :
671 : static void
672 0 : _raid5f_submit_rw_request(void *_raid_io)
673 : {
674 0 : struct raid_bdev_io *raid_io = _raid_io;
675 :
676 0 : raid5f_submit_rw_request(raid_io);
677 0 : }
678 :
679 : static void
680 999 : raid5f_stripe_request_reconstruct_xor_done(struct stripe_request *stripe_req, int status)
681 : {
682 999 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
683 :
684 999 : raid5f_stripe_request_release(stripe_req);
685 :
686 1998 : raid_bdev_io_complete(raid_io,
687 999 : status == 0 ? SPDK_BDEV_IO_STATUS_SUCCESS : SPDK_BDEV_IO_STATUS_FAILED);
688 999 : }
689 :
690 : static void
691 999 : raid5f_reconstruct_reads_completed_cb(struct raid_bdev_io *raid_io, enum spdk_bdev_io_status status)
692 : {
693 999 : struct stripe_request *stripe_req = raid_io->module_private;
694 :
695 999 : raid_io->completion_cb = NULL;
696 :
697 999 : if (status != SPDK_BDEV_IO_STATUS_SUCCESS) {
698 0 : stripe_req->xor.cb(stripe_req, -EIO);
699 0 : return;
700 : }
701 :
702 999 : raid5f_xor_stripe(stripe_req, stripe_req->xor.cb);
703 999 : }
704 :
705 : static int
706 999 : raid5f_submit_reconstruct_read(struct raid_bdev_io *raid_io, uint64_t stripe_index,
707 : uint8_t chunk_idx, uint64_t chunk_offset, stripe_req_xor_cb cb)
708 : {
709 999 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
710 999 : struct raid5f_io_channel *r5ch = raid_bdev_channel_get_module_ctx(raid_io->raid_ch);
711 999 : void *raid_io_md = raid_io->md_buf;
712 999 : struct stripe_request *stripe_req;
713 999 : struct chunk *chunk;
714 999 : int buf_idx;
715 :
716 999 : assert(cb != NULL);
717 :
718 999 : stripe_req = TAILQ_FIRST(&r5ch->free_stripe_requests.reconstruct);
719 999 : if (!stripe_req) {
720 0 : return -ENOMEM;
721 : }
722 :
723 999 : raid5f_stripe_request_init(stripe_req, raid_io, stripe_index);
724 :
725 999 : stripe_req->reconstruct.chunk = &stripe_req->chunks[chunk_idx];
726 999 : stripe_req->reconstruct.chunk_offset = chunk_offset;
727 999 : stripe_req->xor.cb = cb;
728 999 : buf_idx = 0;
729 :
730 5211 : FOR_EACH_CHUNK(stripe_req, chunk) {
731 4212 : if (chunk == stripe_req->reconstruct.chunk) {
732 999 : int i;
733 999 : int ret;
734 :
735 999 : ret = raid5f_chunk_set_iovcnt(chunk, raid_io->iovcnt);
736 999 : if (ret) {
737 0 : return ret;
738 : }
739 :
740 7992 : for (i = 0; i < raid_io->iovcnt; i++) {
741 6993 : chunk->iovs[i] = raid_io->iovs[i];
742 6993 : }
743 :
744 999 : chunk->md_buf = raid_io_md;
745 999 : } else {
746 3213 : struct iovec *iov = &chunk->iovs[0];
747 :
748 3213 : iov->iov_base = stripe_req->reconstruct.chunk_buffers[buf_idx];
749 3213 : iov->iov_len = raid_io->num_blocks * raid_bdev->bdev.blocklen;
750 3213 : chunk->iovcnt = 1;
751 :
752 3213 : if (raid_io_md) {
753 1071 : chunk->md_buf = stripe_req->reconstruct.chunk_md_buffers[buf_idx];
754 1071 : }
755 :
756 3213 : buf_idx++;
757 3213 : }
758 4212 : }
759 :
760 999 : raid_io->module_private = stripe_req;
761 999 : raid_io->base_bdev_io_remaining = raid_bdev->num_base_bdevs;
762 999 : raid_io->completion_cb = raid5f_reconstruct_reads_completed_cb;
763 :
764 999 : TAILQ_REMOVE(&r5ch->free_stripe_requests.reconstruct, stripe_req, link);
765 :
766 999 : raid5f_stripe_request_submit_chunks(stripe_req);
767 :
768 999 : return 0;
769 999 : }
770 :
771 : static int
772 8274 : raid5f_submit_read_request(struct raid_bdev_io *raid_io, uint64_t stripe_index,
773 : uint64_t stripe_offset)
774 : {
775 8274 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
776 8274 : uint8_t chunk_data_idx = stripe_offset >> raid_bdev->strip_size_shift;
777 8274 : uint8_t p_idx = raid5f_stripe_parity_chunk_index(raid_bdev, stripe_index);
778 8274 : uint8_t chunk_idx = chunk_data_idx < p_idx ? chunk_data_idx : chunk_data_idx + 1;
779 8274 : struct raid_base_bdev_info *base_info = &raid_bdev->base_bdev_info[chunk_idx];
780 8274 : struct spdk_io_channel *base_ch = raid_bdev_channel_get_base_channel(raid_io->raid_ch, chunk_idx);
781 8274 : uint64_t chunk_offset = stripe_offset - (chunk_data_idx << raid_bdev->strip_size_shift);
782 8274 : uint64_t base_offset_blocks = (stripe_index << raid_bdev->strip_size_shift) + chunk_offset;
783 8274 : struct spdk_bdev_ext_io_opts io_opts;
784 8274 : int ret;
785 :
786 8274 : raid5f_init_ext_io_opts(&io_opts, raid_io);
787 8274 : if (base_ch == NULL) {
788 999 : return raid5f_submit_reconstruct_read(raid_io, stripe_index, chunk_idx, chunk_offset,
789 : raid5f_stripe_request_reconstruct_xor_done);
790 : }
791 :
792 14550 : ret = raid_bdev_readv_blocks_ext(base_info, base_ch, raid_io->iovs, raid_io->iovcnt,
793 7275 : base_offset_blocks, raid_io->num_blocks,
794 7275 : raid5f_chunk_read_complete, raid_io, &io_opts);
795 7275 : if (spdk_unlikely(ret == -ENOMEM)) {
796 0 : raid_bdev_queue_io_wait(raid_io, spdk_bdev_desc_get_bdev(base_info->desc),
797 0 : base_ch, _raid5f_submit_rw_request);
798 0 : return 0;
799 : }
800 :
801 7275 : return ret;
802 8274 : }
803 :
804 : static void
805 15879 : raid5f_submit_rw_request(struct raid_bdev_io *raid_io)
806 : {
807 15879 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
808 15879 : struct raid5f_info *r5f_info = raid_bdev->module_private;
809 15879 : uint64_t stripe_index = raid_io->offset_blocks / r5f_info->stripe_blocks;
810 15879 : uint64_t stripe_offset = raid_io->offset_blocks % r5f_info->stripe_blocks;
811 15879 : int ret;
812 :
813 15879 : switch (raid_io->type) {
814 : case SPDK_BDEV_IO_TYPE_READ:
815 8274 : assert(raid_io->num_blocks <= raid_bdev->strip_size);
816 8274 : ret = raid5f_submit_read_request(raid_io, stripe_index, stripe_offset);
817 8274 : break;
818 : case SPDK_BDEV_IO_TYPE_WRITE:
819 7605 : assert(stripe_offset == 0);
820 7605 : assert(raid_io->num_blocks == r5f_info->stripe_blocks);
821 7605 : ret = raid5f_submit_write_request(raid_io, stripe_index);
822 7605 : break;
823 : default:
824 0 : ret = -EINVAL;
825 0 : break;
826 : }
827 :
828 15879 : if (spdk_unlikely(ret)) {
829 0 : raid_bdev_io_complete(raid_io, ret == -ENOMEM ? SPDK_BDEV_IO_STATUS_NOMEM :
830 : SPDK_BDEV_IO_STATUS_FAILED);
831 0 : }
832 15879 : }
833 :
834 : static void
835 44451 : raid5f_stripe_request_free(struct stripe_request *stripe_req)
836 : {
837 44451 : struct chunk *chunk;
838 :
839 222255 : FOR_EACH_CHUNK(stripe_req, chunk) {
840 177804 : free(chunk->iovs);
841 177804 : }
842 :
843 44451 : if (stripe_req->type == STRIPE_REQ_WRITE) {
844 22275 : spdk_dma_free(stripe_req->write.parity_buf);
845 22275 : spdk_dma_free(stripe_req->write.parity_md_buf);
846 44451 : } else if (stripe_req->type == STRIPE_REQ_RECONSTRUCT) {
847 22176 : struct raid5f_info *r5f_info = raid5f_ch_to_r5f_info(stripe_req->r5ch);
848 22176 : struct raid_bdev *raid_bdev = r5f_info->raid_bdev;
849 22176 : uint8_t i;
850 :
851 22176 : if (stripe_req->reconstruct.chunk_buffers) {
852 88704 : for (i = 0; i < raid5f_stripe_data_chunks_num(raid_bdev); i++) {
853 66528 : spdk_dma_free(stripe_req->reconstruct.chunk_buffers[i]);
854 66528 : }
855 22176 : free(stripe_req->reconstruct.chunk_buffers);
856 22176 : }
857 :
858 22176 : if (stripe_req->reconstruct.chunk_md_buffers) {
859 29568 : for (i = 0; i < raid5f_stripe_data_chunks_num(raid_bdev); i++) {
860 22176 : spdk_dma_free(stripe_req->reconstruct.chunk_md_buffers[i]);
861 22176 : }
862 7392 : free(stripe_req->reconstruct.chunk_md_buffers);
863 7392 : }
864 22176 : } else {
865 0 : assert(false);
866 : }
867 :
868 44451 : free(stripe_req->chunk_xor_buffers);
869 44451 : free(stripe_req->chunk_xor_md_buffers);
870 44451 : free(stripe_req->chunk_iov_iters);
871 :
872 44451 : free(stripe_req);
873 44451 : }
874 :
875 : static struct stripe_request *
876 44451 : raid5f_stripe_request_alloc(struct raid5f_io_channel *r5ch, enum stripe_request_type type)
877 : {
878 44451 : struct raid5f_info *r5f_info = raid5f_ch_to_r5f_info(r5ch);
879 44451 : struct raid_bdev *raid_bdev = r5f_info->raid_bdev;
880 44451 : uint32_t raid_io_md_size = raid_bdev->bdev.md_interleave ? 0 : raid_bdev->bdev.md_len;
881 44451 : struct stripe_request *stripe_req;
882 44451 : struct chunk *chunk;
883 44451 : size_t chunk_len;
884 :
885 44451 : stripe_req = calloc(1, sizeof(*stripe_req) + sizeof(*chunk) * raid_bdev->num_base_bdevs);
886 44451 : if (!stripe_req) {
887 0 : return NULL;
888 : }
889 :
890 44451 : stripe_req->r5ch = r5ch;
891 44451 : stripe_req->type = type;
892 :
893 222255 : FOR_EACH_CHUNK(stripe_req, chunk) {
894 177804 : chunk->index = chunk - stripe_req->chunks;
895 177804 : chunk->iovcnt_max = 4;
896 177804 : chunk->iovs = calloc(chunk->iovcnt_max, sizeof(chunk->iovs[0]));
897 177804 : if (!chunk->iovs) {
898 0 : goto err;
899 : }
900 177804 : }
901 :
902 44451 : chunk_len = raid_bdev->strip_size * raid_bdev->bdev.blocklen;
903 :
904 44451 : if (type == STRIPE_REQ_WRITE) {
905 22275 : stripe_req->write.parity_buf = spdk_dma_malloc(chunk_len, r5f_info->buf_alignment, NULL);
906 22275 : if (!stripe_req->write.parity_buf) {
907 0 : goto err;
908 : }
909 :
910 22275 : if (raid_io_md_size != 0) {
911 14850 : stripe_req->write.parity_md_buf = spdk_dma_malloc(raid_bdev->strip_size * raid_io_md_size,
912 7425 : r5f_info->buf_alignment, NULL);
913 7425 : if (!stripe_req->write.parity_md_buf) {
914 0 : goto err;
915 : }
916 7425 : }
917 44451 : } else if (type == STRIPE_REQ_RECONSTRUCT) {
918 22176 : uint8_t n = raid5f_stripe_data_chunks_num(raid_bdev);
919 22176 : void *buf;
920 22176 : uint8_t i;
921 :
922 22176 : stripe_req->reconstruct.chunk_buffers = calloc(n, sizeof(void *));
923 22176 : if (!stripe_req->reconstruct.chunk_buffers) {
924 0 : goto err;
925 : }
926 :
927 88704 : for (i = 0; i < n; i++) {
928 66528 : buf = spdk_dma_malloc(chunk_len, r5f_info->buf_alignment, NULL);
929 66528 : if (!buf) {
930 0 : goto err;
931 : }
932 66528 : stripe_req->reconstruct.chunk_buffers[i] = buf;
933 66528 : }
934 :
935 22176 : if (raid_io_md_size != 0) {
936 7392 : stripe_req->reconstruct.chunk_md_buffers = calloc(n, sizeof(void *));
937 7392 : if (!stripe_req->reconstruct.chunk_md_buffers) {
938 0 : goto err;
939 : }
940 :
941 29568 : for (i = 0; i < n; i++) {
942 22176 : buf = spdk_dma_malloc(raid_bdev->strip_size * raid_io_md_size, r5f_info->buf_alignment, NULL);
943 22176 : if (!buf) {
944 0 : goto err;
945 : }
946 22176 : stripe_req->reconstruct.chunk_md_buffers[i] = buf;
947 22176 : }
948 7392 : }
949 22176 : } else {
950 0 : assert(false);
951 : return NULL;
952 : }
953 :
954 44451 : stripe_req->chunk_iov_iters = malloc(SPDK_IOVITER_SIZE(raid_bdev->num_base_bdevs));
955 44451 : if (!stripe_req->chunk_iov_iters) {
956 0 : goto err;
957 : }
958 :
959 44451 : stripe_req->chunk_xor_buffers = calloc(raid5f_stripe_data_chunks_num(raid_bdev),
960 : sizeof(stripe_req->chunk_xor_buffers[0]));
961 44451 : if (!stripe_req->chunk_xor_buffers) {
962 0 : goto err;
963 : }
964 :
965 44451 : stripe_req->chunk_xor_md_buffers = calloc(raid5f_stripe_data_chunks_num(raid_bdev),
966 : sizeof(stripe_req->chunk_xor_md_buffers[0]));
967 44451 : if (!stripe_req->chunk_xor_md_buffers) {
968 0 : goto err;
969 : }
970 :
971 44451 : return stripe_req;
972 : err:
973 0 : raid5f_stripe_request_free(stripe_req);
974 0 : return NULL;
975 44451 : }
976 :
977 : static void
978 693 : raid5f_ioch_destroy(void *io_device, void *ctx_buf)
979 : {
980 693 : struct raid5f_io_channel *r5ch = ctx_buf;
981 693 : struct stripe_request *stripe_req;
982 :
983 693 : assert(TAILQ_EMPTY(&r5ch->xor_retry_queue));
984 :
985 22869 : while ((stripe_req = TAILQ_FIRST(&r5ch->free_stripe_requests.write))) {
986 22176 : TAILQ_REMOVE(&r5ch->free_stripe_requests.write, stripe_req, link);
987 22176 : raid5f_stripe_request_free(stripe_req);
988 : }
989 :
990 22869 : while ((stripe_req = TAILQ_FIRST(&r5ch->free_stripe_requests.reconstruct))) {
991 22176 : TAILQ_REMOVE(&r5ch->free_stripe_requests.reconstruct, stripe_req, link);
992 22176 : raid5f_stripe_request_free(stripe_req);
993 : }
994 :
995 693 : if (r5ch->accel_ch) {
996 693 : spdk_put_io_channel(r5ch->accel_ch);
997 693 : }
998 :
999 693 : free(r5ch->chunk_xor_buffers);
1000 693 : free(r5ch->chunk_xor_iovs);
1001 693 : free(r5ch->chunk_xor_iovcnt);
1002 693 : }
1003 :
1004 : static int
1005 693 : raid5f_ioch_create(void *io_device, void *ctx_buf)
1006 : {
1007 693 : struct raid5f_io_channel *r5ch = ctx_buf;
1008 693 : struct raid5f_info *r5f_info = io_device;
1009 693 : struct raid_bdev *raid_bdev = r5f_info->raid_bdev;
1010 693 : struct stripe_request *stripe_req;
1011 693 : int i;
1012 :
1013 693 : TAILQ_INIT(&r5ch->free_stripe_requests.write);
1014 693 : TAILQ_INIT(&r5ch->free_stripe_requests.reconstruct);
1015 693 : TAILQ_INIT(&r5ch->xor_retry_queue);
1016 :
1017 22869 : for (i = 0; i < RAID5F_MAX_STRIPES; i++) {
1018 22176 : stripe_req = raid5f_stripe_request_alloc(r5ch, STRIPE_REQ_WRITE);
1019 22176 : if (!stripe_req) {
1020 0 : goto err;
1021 : }
1022 :
1023 22176 : TAILQ_INSERT_HEAD(&r5ch->free_stripe_requests.write, stripe_req, link);
1024 22176 : }
1025 :
1026 22869 : for (i = 0; i < RAID5F_MAX_STRIPES; i++) {
1027 22176 : stripe_req = raid5f_stripe_request_alloc(r5ch, STRIPE_REQ_RECONSTRUCT);
1028 22176 : if (!stripe_req) {
1029 0 : goto err;
1030 : }
1031 :
1032 22176 : TAILQ_INSERT_HEAD(&r5ch->free_stripe_requests.reconstruct, stripe_req, link);
1033 22176 : }
1034 :
1035 693 : r5ch->accel_ch = spdk_accel_get_io_channel();
1036 693 : if (!r5ch->accel_ch) {
1037 0 : SPDK_ERRLOG("Failed to get accel framework's IO channel\n");
1038 0 : goto err;
1039 : }
1040 :
1041 693 : r5ch->chunk_xor_buffers = calloc(raid_bdev->num_base_bdevs, sizeof(*r5ch->chunk_xor_buffers));
1042 693 : if (!r5ch->chunk_xor_buffers) {
1043 0 : goto err;
1044 : }
1045 :
1046 693 : r5ch->chunk_xor_iovs = calloc(raid_bdev->num_base_bdevs, sizeof(*r5ch->chunk_xor_iovs));
1047 693 : if (!r5ch->chunk_xor_iovs) {
1048 0 : goto err;
1049 : }
1050 :
1051 693 : r5ch->chunk_xor_iovcnt = calloc(raid_bdev->num_base_bdevs, sizeof(*r5ch->chunk_xor_iovcnt));
1052 693 : if (!r5ch->chunk_xor_iovcnt) {
1053 0 : goto err;
1054 : }
1055 :
1056 693 : return 0;
1057 : err:
1058 0 : SPDK_ERRLOG("Failed to initialize io channel\n");
1059 0 : raid5f_ioch_destroy(r5f_info, r5ch);
1060 0 : return -ENOMEM;
1061 693 : }
1062 :
1063 : static int
1064 792 : raid5f_start(struct raid_bdev *raid_bdev)
1065 : {
1066 792 : uint64_t min_blockcnt = UINT64_MAX;
1067 792 : uint64_t base_bdev_data_size;
1068 792 : struct raid_base_bdev_info *base_info;
1069 792 : struct spdk_bdev *base_bdev;
1070 792 : struct raid5f_info *r5f_info;
1071 792 : size_t alignment = 0;
1072 :
1073 792 : r5f_info = calloc(1, sizeof(*r5f_info));
1074 792 : if (!r5f_info) {
1075 0 : SPDK_ERRLOG("Failed to allocate r5f_info\n");
1076 0 : return -ENOMEM;
1077 : }
1078 792 : r5f_info->raid_bdev = raid_bdev;
1079 :
1080 3960 : RAID_FOR_EACH_BASE_BDEV(raid_bdev, base_info) {
1081 3168 : min_blockcnt = spdk_min(min_blockcnt, base_info->data_size);
1082 3168 : if (base_info->desc) {
1083 3168 : base_bdev = spdk_bdev_desc_get_bdev(base_info->desc);
1084 3168 : alignment = spdk_max(alignment, spdk_bdev_get_buf_align(base_bdev));
1085 3168 : }
1086 3168 : }
1087 :
1088 792 : base_bdev_data_size = (min_blockcnt / raid_bdev->strip_size) * raid_bdev->strip_size;
1089 :
1090 3960 : RAID_FOR_EACH_BASE_BDEV(raid_bdev, base_info) {
1091 3168 : base_info->data_size = base_bdev_data_size;
1092 3168 : }
1093 :
1094 792 : r5f_info->total_stripes = min_blockcnt / raid_bdev->strip_size;
1095 792 : r5f_info->stripe_blocks = raid_bdev->strip_size * raid5f_stripe_data_chunks_num(raid_bdev);
1096 792 : r5f_info->buf_alignment = alignment;
1097 792 : if (!raid_bdev->bdev.md_interleave) {
1098 528 : r5f_info->blocklen_shift = spdk_u32log2(raid_bdev->bdev.blocklen);
1099 528 : }
1100 :
1101 792 : raid_bdev->bdev.blockcnt = r5f_info->stripe_blocks * r5f_info->total_stripes;
1102 792 : raid_bdev->bdev.optimal_io_boundary = raid_bdev->strip_size;
1103 792 : raid_bdev->bdev.split_on_optimal_io_boundary = true;
1104 792 : raid_bdev->bdev.write_unit_size = r5f_info->stripe_blocks;
1105 792 : raid_bdev->bdev.split_on_write_unit = true;
1106 :
1107 792 : raid_bdev->module_private = r5f_info;
1108 :
1109 792 : spdk_io_device_register(r5f_info, raid5f_ioch_create, raid5f_ioch_destroy,
1110 : sizeof(struct raid5f_io_channel), NULL);
1111 :
1112 792 : return 0;
1113 792 : }
1114 :
1115 : static void
1116 792 : raid5f_io_device_unregister_done(void *io_device)
1117 : {
1118 792 : struct raid5f_info *r5f_info = io_device;
1119 :
1120 792 : raid_bdev_module_stop_done(r5f_info->raid_bdev);
1121 :
1122 792 : free(r5f_info);
1123 792 : }
1124 :
1125 : static bool
1126 792 : raid5f_stop(struct raid_bdev *raid_bdev)
1127 : {
1128 792 : struct raid5f_info *r5f_info = raid_bdev->module_private;
1129 :
1130 792 : spdk_io_device_unregister(r5f_info, raid5f_io_device_unregister_done);
1131 :
1132 792 : return false;
1133 792 : }
1134 :
1135 : static struct spdk_io_channel *
1136 693 : raid5f_get_io_channel(struct raid_bdev *raid_bdev)
1137 : {
1138 693 : struct raid5f_info *r5f_info = raid_bdev->module_private;
1139 :
1140 1386 : return spdk_get_io_channel(r5f_info);
1141 693 : }
1142 :
1143 : static void
1144 0 : raid5f_process_write_completed(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg)
1145 : {
1146 0 : struct raid_bdev_process_request *process_req = cb_arg;
1147 :
1148 0 : spdk_bdev_free_io(bdev_io);
1149 :
1150 0 : raid_bdev_process_request_complete(process_req, success ? 0 : -EIO);
1151 0 : }
1152 :
1153 : static void raid5f_process_submit_write(struct raid_bdev_process_request *process_req);
1154 :
1155 : static void
1156 0 : _raid5f_process_submit_write(void *ctx)
1157 : {
1158 0 : struct raid_bdev_process_request *process_req = ctx;
1159 :
1160 0 : raid5f_process_submit_write(process_req);
1161 0 : }
1162 :
1163 : static void
1164 0 : raid5f_process_submit_write(struct raid_bdev_process_request *process_req)
1165 : {
1166 0 : struct raid_bdev_io *raid_io = &process_req->raid_io;
1167 0 : struct raid_bdev *raid_bdev = raid_io->raid_bdev;
1168 0 : struct raid5f_info *r5f_info = raid_bdev->module_private;
1169 0 : uint64_t stripe_index = process_req->offset_blocks / r5f_info->stripe_blocks;
1170 0 : struct spdk_bdev_ext_io_opts io_opts;
1171 0 : int ret;
1172 :
1173 0 : raid5f_init_ext_io_opts(&io_opts, raid_io);
1174 0 : ret = raid_bdev_writev_blocks_ext(process_req->target, process_req->target_ch,
1175 0 : raid_io->iovs, raid_io->iovcnt,
1176 0 : stripe_index << raid_bdev->strip_size_shift, raid_bdev->strip_size,
1177 0 : raid5f_process_write_completed, process_req, &io_opts);
1178 0 : if (spdk_unlikely(ret != 0)) {
1179 0 : if (ret == -ENOMEM) {
1180 0 : raid_bdev_queue_io_wait(raid_io, spdk_bdev_desc_get_bdev(process_req->target->desc),
1181 0 : process_req->target_ch, _raid5f_process_submit_write);
1182 0 : } else {
1183 0 : raid_bdev_process_request_complete(process_req, ret);
1184 : }
1185 0 : }
1186 0 : }
1187 :
1188 : static void
1189 0 : raid5f_process_stripe_request_reconstruct_xor_done(struct stripe_request *stripe_req, int status)
1190 : {
1191 0 : struct raid_bdev_io *raid_io = stripe_req->raid_io;
1192 0 : struct raid_bdev_process_request *process_req = SPDK_CONTAINEROF(raid_io,
1193 : struct raid_bdev_process_request, raid_io);
1194 :
1195 0 : raid5f_stripe_request_release(stripe_req);
1196 :
1197 0 : if (status != 0) {
1198 0 : raid_bdev_process_request_complete(process_req, status);
1199 0 : return;
1200 : }
1201 :
1202 0 : raid5f_process_submit_write(process_req);
1203 0 : }
1204 :
1205 : static int
1206 0 : raid5f_submit_process_request(struct raid_bdev_process_request *process_req,
1207 : struct raid_bdev_io_channel *raid_ch)
1208 : {
1209 0 : struct spdk_io_channel *ch = spdk_io_channel_from_ctx(raid_ch);
1210 0 : struct raid_bdev *raid_bdev = spdk_io_channel_get_io_device(ch);
1211 0 : struct raid5f_info *r5f_info = raid_bdev->module_private;
1212 0 : struct raid_bdev_io *raid_io = &process_req->raid_io;
1213 0 : uint8_t chunk_idx = raid_bdev_base_bdev_slot(process_req->target);
1214 0 : uint64_t stripe_index = process_req->offset_blocks / r5f_info->stripe_blocks;
1215 0 : int ret;
1216 :
1217 0 : assert((process_req->offset_blocks % r5f_info->stripe_blocks) == 0);
1218 :
1219 0 : if (process_req->num_blocks < r5f_info->stripe_blocks) {
1220 0 : return 0;
1221 : }
1222 :
1223 0 : raid_bdev_io_init(raid_io, raid_ch, SPDK_BDEV_IO_TYPE_READ,
1224 0 : process_req->offset_blocks, raid_bdev->strip_size,
1225 0 : &process_req->iov, 1, process_req->md_buf, NULL, NULL);
1226 :
1227 0 : ret = raid5f_submit_reconstruct_read(raid_io, stripe_index, chunk_idx, 0,
1228 : raid5f_process_stripe_request_reconstruct_xor_done);
1229 0 : if (spdk_likely(ret == 0)) {
1230 0 : return r5f_info->stripe_blocks;
1231 0 : } else if (ret < 0) {
1232 0 : return ret;
1233 : } else {
1234 0 : return -EINVAL;
1235 : }
1236 0 : }
1237 :
1238 : static struct raid_bdev_module g_raid5f_module = {
1239 : .level = RAID5F,
1240 : .base_bdevs_min = 3,
1241 : .base_bdevs_constraint = {CONSTRAINT_MAX_BASE_BDEVS_REMOVED, 1},
1242 : .start = raid5f_start,
1243 : .stop = raid5f_stop,
1244 : .submit_rw_request = raid5f_submit_rw_request,
1245 : .get_io_channel = raid5f_get_io_channel,
1246 : .submit_process_request = raid5f_submit_process_request,
1247 : };
1248 1 : RAID_MODULE_REGISTER(&g_raid5f_module)
1249 :
1250 1 : SPDK_LOG_REGISTER_COMPONENT(bdev_raid5f)
|
