Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2017 Intel Corporation.
3 : * All rights reserved.
4 : * Copyright (c) 2021 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
5 : */
6 :
7 : #include "spdk/stdinc.h"
8 :
9 : #include "bdev_malloc.h"
10 : #include "spdk/endian.h"
11 : #include "spdk/env.h"
12 : #include "spdk/accel.h"
13 : #include "spdk/dma.h"
14 : #include "spdk/likely.h"
15 : #include "spdk/string.h"
16 :
17 : #include "spdk/log.h"
18 :
19 : struct malloc_disk {
20 : struct spdk_bdev disk;
21 : void *malloc_buf;
22 : void *malloc_md_buf;
23 : TAILQ_ENTRY(malloc_disk) link;
24 : };
25 :
26 : struct malloc_task {
27 : struct iovec iov;
28 : int num_outstanding;
29 : enum spdk_bdev_io_status status;
30 : TAILQ_ENTRY(malloc_task) tailq;
31 : };
32 :
33 : struct malloc_channel {
34 : struct spdk_io_channel *accel_channel;
35 : struct spdk_poller *completion_poller;
36 : TAILQ_HEAD(, malloc_task) completed_tasks;
37 : };
38 :
39 : static int
40 0 : malloc_verify_pi(struct spdk_bdev_io *bdev_io)
41 : {
42 0 : struct spdk_bdev *bdev = bdev_io->bdev;
43 : struct spdk_dif_ctx dif_ctx;
44 : struct spdk_dif_error err_blk;
45 : int rc;
46 : struct spdk_dif_ctx_init_ext_opts dif_opts;
47 :
48 0 : assert(bdev_io->u.bdev.memory_domain == NULL);
49 0 : dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format);
50 0 : dif_opts.dif_pi_format = bdev->dif_pi_format;
51 0 : rc = spdk_dif_ctx_init(&dif_ctx,
52 0 : bdev->blocklen,
53 0 : bdev->md_len,
54 0 : bdev->md_interleave,
55 0 : bdev->dif_is_head_of_md,
56 0 : bdev->dif_type,
57 0 : bdev_io->u.bdev.dif_check_flags,
58 0 : bdev_io->u.bdev.offset_blocks & 0xFFFFFFFF,
59 : 0xFFFF, 0, 0, 0, &dif_opts);
60 0 : if (rc != 0) {
61 0 : SPDK_ERRLOG("Failed to initialize DIF/DIX context\n");
62 0 : return rc;
63 : }
64 :
65 0 : if (spdk_bdev_is_md_interleaved(bdev)) {
66 0 : rc = spdk_dif_verify(bdev_io->u.bdev.iovs,
67 0 : bdev_io->u.bdev.iovcnt,
68 0 : bdev_io->u.bdev.num_blocks,
69 : &dif_ctx,
70 : &err_blk);
71 0 : } else {
72 0 : struct iovec md_iov = {
73 0 : .iov_base = bdev_io->u.bdev.md_buf,
74 0 : .iov_len = bdev_io->u.bdev.num_blocks * bdev->md_len,
75 : };
76 :
77 0 : if (bdev_io->u.bdev.md_buf == NULL) {
78 0 : return 0;
79 : }
80 :
81 0 : rc = spdk_dix_verify(bdev_io->u.bdev.iovs,
82 0 : bdev_io->u.bdev.iovcnt,
83 : &md_iov,
84 0 : bdev_io->u.bdev.num_blocks,
85 : &dif_ctx,
86 : &err_blk);
87 : }
88 :
89 0 : if (rc != 0) {
90 0 : SPDK_ERRLOG("DIF/DIX verify failed: lba %" PRIu64 ", num_blocks %" PRIu64 ", "
91 : "err_type %u, expected %lu, actual %lu, err_offset %u\n",
92 : bdev_io->u.bdev.offset_blocks,
93 : bdev_io->u.bdev.num_blocks,
94 : err_blk.err_type,
95 : err_blk.expected,
96 : err_blk.actual,
97 : err_blk.err_offset);
98 0 : }
99 :
100 0 : return rc;
101 0 : }
102 :
103 : static int
104 0 : malloc_unmap_write_zeroes_generate_pi(struct spdk_bdev_io *bdev_io)
105 : {
106 0 : struct spdk_bdev *bdev = bdev_io->bdev;
107 0 : struct malloc_disk *mdisk = bdev_io->bdev->ctxt;
108 0 : uint32_t block_size = bdev_io->bdev->blocklen;
109 : uint32_t dif_check_flags;
110 : struct spdk_dif_ctx dif_ctx;
111 : struct spdk_dif_ctx_init_ext_opts dif_opts;
112 : int rc;
113 :
114 0 : dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format);
115 0 : dif_opts.dif_pi_format = bdev->dif_pi_format;
116 0 : dif_check_flags = bdev->dif_check_flags | SPDK_DIF_CHECK_TYPE_REFTAG |
117 : SPDK_DIF_FLAGS_APPTAG_CHECK;
118 0 : rc = spdk_dif_ctx_init(&dif_ctx,
119 0 : bdev->blocklen,
120 0 : bdev->md_len,
121 0 : bdev->md_interleave,
122 0 : bdev->dif_is_head_of_md,
123 0 : bdev->dif_type,
124 0 : dif_check_flags,
125 : SPDK_DIF_REFTAG_IGNORE,
126 : 0xFFFF, SPDK_DIF_APPTAG_IGNORE,
127 : 0, 0, &dif_opts);
128 0 : if (rc != 0) {
129 0 : SPDK_ERRLOG("Initialization of DIF/DIX context failed\n");
130 0 : return rc;
131 : }
132 :
133 0 : if (bdev->md_interleave) {
134 0 : struct iovec iov = {
135 0 : .iov_base = mdisk->malloc_buf + bdev_io->u.bdev.offset_blocks * block_size,
136 0 : .iov_len = bdev_io->u.bdev.num_blocks * block_size,
137 : };
138 :
139 0 : rc = spdk_dif_generate(&iov, 1, bdev_io->u.bdev.num_blocks, &dif_ctx);
140 0 : } else {
141 0 : struct iovec iov = {
142 0 : .iov_base = mdisk->malloc_buf + bdev_io->u.bdev.offset_blocks * block_size,
143 0 : .iov_len = bdev_io->u.bdev.num_blocks * block_size,
144 : };
145 :
146 0 : struct iovec md_iov = {
147 0 : .iov_base = mdisk->malloc_md_buf + bdev_io->u.bdev.offset_blocks * bdev->md_len,
148 0 : .iov_len = bdev_io->u.bdev.num_blocks * bdev->md_len,
149 : };
150 :
151 0 : rc = spdk_dix_generate(&iov, 1, &md_iov, bdev_io->u.bdev.num_blocks, &dif_ctx);
152 : }
153 :
154 0 : if (rc != 0) {
155 0 : SPDK_ERRLOG("Formatting by DIF/DIX failed\n");
156 0 : }
157 :
158 :
159 0 : return rc;
160 0 : }
161 :
162 : static void
163 0 : malloc_done(void *ref, int status)
164 : {
165 0 : struct malloc_task *task = (struct malloc_task *)ref;
166 0 : struct spdk_bdev_io *bdev_io = spdk_bdev_io_from_ctx(task);
167 : int rc;
168 :
169 0 : if (status != 0) {
170 0 : if (status == -ENOMEM) {
171 0 : if (task->status == SPDK_BDEV_IO_STATUS_SUCCESS) {
172 0 : task->status = SPDK_BDEV_IO_STATUS_NOMEM;
173 0 : }
174 0 : } else {
175 0 : task->status = SPDK_BDEV_IO_STATUS_FAILED;
176 : }
177 0 : }
178 :
179 0 : if (--task->num_outstanding != 0) {
180 0 : return;
181 : }
182 :
183 0 : if (bdev_io->bdev->dif_type != SPDK_DIF_DISABLE &&
184 0 : bdev_io->type == SPDK_BDEV_IO_TYPE_READ &&
185 0 : task->status == SPDK_BDEV_IO_STATUS_SUCCESS) {
186 0 : rc = malloc_verify_pi(bdev_io);
187 0 : if (rc != 0) {
188 0 : task->status = SPDK_BDEV_IO_STATUS_FAILED;
189 0 : }
190 0 : }
191 :
192 0 : if (bdev_io->bdev->dif_type != SPDK_DIF_DISABLE &&
193 0 : (bdev_io->type == SPDK_BDEV_IO_TYPE_UNMAP || bdev_io->type == SPDK_BDEV_IO_TYPE_WRITE_ZEROES) &&
194 0 : task->status == SPDK_BDEV_IO_STATUS_SUCCESS) {
195 0 : rc = malloc_unmap_write_zeroes_generate_pi(bdev_io);
196 0 : if (rc != 0) {
197 0 : task->status = SPDK_BDEV_IO_STATUS_FAILED;
198 0 : }
199 0 : }
200 :
201 0 : assert(!bdev_io->u.bdev.accel_sequence || task->status == SPDK_BDEV_IO_STATUS_NOMEM);
202 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(task), task->status);
203 0 : }
204 :
205 : static void
206 0 : malloc_complete_task(struct malloc_task *task, struct malloc_channel *mch,
207 : enum spdk_bdev_io_status status)
208 : {
209 0 : task->status = status;
210 0 : TAILQ_INSERT_TAIL(&mch->completed_tasks, task, tailq);
211 0 : }
212 :
213 : static TAILQ_HEAD(, malloc_disk) g_malloc_disks = TAILQ_HEAD_INITIALIZER(g_malloc_disks);
214 :
215 : int malloc_disk_count = 0;
216 :
217 : static int bdev_malloc_initialize(void);
218 : static void bdev_malloc_deinitialize(void);
219 :
220 : static int
221 0 : bdev_malloc_get_ctx_size(void)
222 : {
223 0 : return sizeof(struct malloc_task);
224 : }
225 :
226 : static struct spdk_bdev_module malloc_if = {
227 : .name = "malloc",
228 : .module_init = bdev_malloc_initialize,
229 : .module_fini = bdev_malloc_deinitialize,
230 : .get_ctx_size = bdev_malloc_get_ctx_size,
231 :
232 : };
233 :
234 0 : SPDK_BDEV_MODULE_REGISTER(malloc, &malloc_if)
235 :
236 : static void
237 0 : malloc_disk_free(struct malloc_disk *malloc_disk)
238 : {
239 0 : if (!malloc_disk) {
240 0 : return;
241 : }
242 :
243 0 : free(malloc_disk->disk.name);
244 0 : spdk_free(malloc_disk->malloc_buf);
245 0 : spdk_free(malloc_disk->malloc_md_buf);
246 0 : free(malloc_disk);
247 0 : }
248 :
249 : static int
250 0 : bdev_malloc_destruct(void *ctx)
251 : {
252 0 : struct malloc_disk *malloc_disk = ctx;
253 :
254 0 : TAILQ_REMOVE(&g_malloc_disks, malloc_disk, link);
255 0 : malloc_disk_free(malloc_disk);
256 0 : return 0;
257 : }
258 :
259 : static int
260 0 : bdev_malloc_check_iov_len(struct iovec *iovs, int iovcnt, size_t nbytes)
261 : {
262 : int i;
263 :
264 0 : for (i = 0; i < iovcnt; i++) {
265 0 : if (nbytes < iovs[i].iov_len) {
266 0 : return 0;
267 : }
268 :
269 0 : nbytes -= iovs[i].iov_len;
270 0 : }
271 :
272 0 : return nbytes != 0;
273 0 : }
274 :
275 : static size_t
276 0 : malloc_get_md_len(struct spdk_bdev_io *bdev_io)
277 : {
278 0 : return bdev_io->u.bdev.num_blocks * bdev_io->bdev->md_len;
279 : }
280 :
281 : static uint64_t
282 0 : malloc_get_md_offset(struct spdk_bdev_io *bdev_io)
283 : {
284 0 : return bdev_io->u.bdev.offset_blocks * bdev_io->bdev->md_len;
285 : }
286 :
287 : static void *
288 0 : malloc_get_md_buf(struct spdk_bdev_io *bdev_io)
289 : {
290 0 : struct malloc_disk *mdisk = SPDK_CONTAINEROF(bdev_io->bdev, struct malloc_disk, disk);
291 :
292 0 : assert(spdk_bdev_is_md_separate(bdev_io->bdev));
293 :
294 0 : return (char *)mdisk->malloc_md_buf + malloc_get_md_offset(bdev_io);
295 : }
296 :
297 : static void
298 0 : malloc_sequence_fail(struct malloc_task *task, int status)
299 : {
300 0 : struct spdk_bdev_io *bdev_io = spdk_bdev_io_from_ctx(task);
301 :
302 : /* For ENOMEM, the IO will be retried by the bdev layer, so we don't abort the sequence */
303 0 : if (status != -ENOMEM) {
304 0 : spdk_accel_sequence_abort(bdev_io->u.bdev.accel_sequence);
305 0 : bdev_io->u.bdev.accel_sequence = NULL;
306 0 : }
307 :
308 0 : malloc_done(task, status);
309 0 : }
310 :
311 : static void
312 0 : malloc_sequence_done(void *ctx, int status)
313 : {
314 0 : struct malloc_task *task = ctx;
315 0 : struct spdk_bdev_io *bdev_io = spdk_bdev_io_from_ctx(task);
316 :
317 0 : bdev_io->u.bdev.accel_sequence = NULL;
318 : /* Prevent bdev layer from retrying the request if the sequence failed with ENOMEM */
319 0 : malloc_done(task, status != -ENOMEM ? status : -EFAULT);
320 0 : }
321 :
322 : static void
323 0 : bdev_malloc_readv(struct malloc_disk *mdisk, struct spdk_io_channel *ch,
324 : struct malloc_task *task, struct spdk_bdev_io *bdev_io)
325 : {
326 : uint64_t len, offset;
327 0 : int res = 0;
328 :
329 0 : len = bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen;
330 0 : offset = bdev_io->u.bdev.offset_blocks * bdev_io->bdev->blocklen;
331 :
332 0 : if (bdev_malloc_check_iov_len(bdev_io->u.bdev.iovs, bdev_io->u.bdev.iovcnt, len)) {
333 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(task),
334 : SPDK_BDEV_IO_STATUS_FAILED);
335 0 : return;
336 : }
337 :
338 0 : task->status = SPDK_BDEV_IO_STATUS_SUCCESS;
339 0 : task->num_outstanding = 0;
340 0 : task->iov.iov_base = mdisk->malloc_buf + offset;
341 0 : task->iov.iov_len = len;
342 :
343 0 : SPDK_DEBUGLOG(bdev_malloc, "read %zu bytes from offset %#" PRIx64 ", iovcnt=%d\n",
344 : len, offset, bdev_io->u.bdev.iovcnt);
345 :
346 0 : task->num_outstanding++;
347 0 : res = spdk_accel_append_copy(&bdev_io->u.bdev.accel_sequence, ch,
348 0 : bdev_io->u.bdev.iovs, bdev_io->u.bdev.iovcnt,
349 0 : bdev_io->u.bdev.memory_domain,
350 0 : bdev_io->u.bdev.memory_domain_ctx,
351 0 : &task->iov, 1, NULL, NULL, NULL, NULL);
352 0 : if (spdk_unlikely(res != 0)) {
353 0 : malloc_sequence_fail(task, res);
354 0 : return;
355 : }
356 :
357 0 : spdk_accel_sequence_reverse(bdev_io->u.bdev.accel_sequence);
358 0 : spdk_accel_sequence_finish(bdev_io->u.bdev.accel_sequence, malloc_sequence_done, task);
359 :
360 0 : if (bdev_io->u.bdev.md_buf == NULL) {
361 0 : return;
362 : }
363 :
364 0 : SPDK_DEBUGLOG(bdev_malloc, "read metadata %zu bytes from offset%#" PRIx64 "\n",
365 : malloc_get_md_len(bdev_io), malloc_get_md_offset(bdev_io));
366 :
367 0 : task->num_outstanding++;
368 0 : res = spdk_accel_submit_copy(ch, bdev_io->u.bdev.md_buf, malloc_get_md_buf(bdev_io),
369 0 : malloc_get_md_len(bdev_io), malloc_done, task);
370 0 : if (res != 0) {
371 0 : malloc_done(task, res);
372 0 : }
373 0 : }
374 :
375 : static void
376 0 : bdev_malloc_writev(struct malloc_disk *mdisk, struct spdk_io_channel *ch,
377 : struct malloc_task *task, struct spdk_bdev_io *bdev_io)
378 : {
379 : uint64_t len, offset;
380 0 : int res = 0;
381 :
382 0 : len = bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen;
383 0 : offset = bdev_io->u.bdev.offset_blocks * bdev_io->bdev->blocklen;
384 :
385 0 : if (bdev_malloc_check_iov_len(bdev_io->u.bdev.iovs, bdev_io->u.bdev.iovcnt, len)) {
386 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(task),
387 : SPDK_BDEV_IO_STATUS_FAILED);
388 0 : return;
389 : }
390 :
391 0 : task->status = SPDK_BDEV_IO_STATUS_SUCCESS;
392 0 : task->num_outstanding = 0;
393 0 : task->iov.iov_base = mdisk->malloc_buf + offset;
394 0 : task->iov.iov_len = len;
395 :
396 0 : SPDK_DEBUGLOG(bdev_malloc, "write %zu bytes to offset %#" PRIx64 ", iovcnt=%d\n",
397 : len, offset, bdev_io->u.bdev.iovcnt);
398 :
399 0 : task->num_outstanding++;
400 0 : res = spdk_accel_append_copy(&bdev_io->u.bdev.accel_sequence, ch, &task->iov, 1, NULL, NULL,
401 0 : bdev_io->u.bdev.iovs, bdev_io->u.bdev.iovcnt,
402 0 : bdev_io->u.bdev.memory_domain,
403 0 : bdev_io->u.bdev.memory_domain_ctx, NULL, NULL);
404 0 : if (spdk_unlikely(res != 0)) {
405 0 : malloc_sequence_fail(task, res);
406 0 : return;
407 : }
408 :
409 0 : spdk_accel_sequence_finish(bdev_io->u.bdev.accel_sequence, malloc_sequence_done, task);
410 :
411 0 : if (bdev_io->u.bdev.md_buf == NULL) {
412 0 : return;
413 : }
414 :
415 0 : SPDK_DEBUGLOG(bdev_malloc, "write metadata %zu bytes to offset %#" PRIx64 "\n",
416 : malloc_get_md_len(bdev_io), malloc_get_md_offset(bdev_io));
417 :
418 0 : task->num_outstanding++;
419 0 : res = spdk_accel_submit_copy(ch, malloc_get_md_buf(bdev_io), bdev_io->u.bdev.md_buf,
420 0 : malloc_get_md_len(bdev_io), malloc_done, task);
421 0 : if (res != 0) {
422 0 : malloc_done(task, res);
423 0 : }
424 0 : }
425 :
426 : static int
427 0 : bdev_malloc_unmap(struct malloc_disk *mdisk,
428 : struct spdk_io_channel *ch,
429 : struct malloc_task *task,
430 : uint64_t offset,
431 : uint64_t byte_count)
432 : {
433 0 : task->status = SPDK_BDEV_IO_STATUS_SUCCESS;
434 0 : task->num_outstanding = 1;
435 :
436 0 : return spdk_accel_submit_fill(ch, mdisk->malloc_buf + offset, 0,
437 0 : byte_count, malloc_done, task);
438 : }
439 :
440 : static void
441 0 : bdev_malloc_copy(struct malloc_disk *mdisk, struct spdk_io_channel *ch,
442 : struct malloc_task *task,
443 : uint64_t dst_offset, uint64_t src_offset, size_t len)
444 : {
445 0 : int64_t res = 0;
446 0 : void *dst = mdisk->malloc_buf + dst_offset;
447 0 : void *src = mdisk->malloc_buf + src_offset;
448 :
449 0 : SPDK_DEBUGLOG(bdev_malloc, "Copy %zu bytes from offset %#" PRIx64 " to offset %#" PRIx64 "\n",
450 : len, src_offset, dst_offset);
451 :
452 0 : task->status = SPDK_BDEV_IO_STATUS_SUCCESS;
453 0 : task->num_outstanding = 1;
454 :
455 0 : res = spdk_accel_submit_copy(ch, dst, src, len, malloc_done, task);
456 0 : if (res != 0) {
457 0 : malloc_done(task, res);
458 0 : }
459 0 : }
460 :
461 : static int
462 0 : _bdev_malloc_submit_request(struct malloc_channel *mch, struct spdk_bdev_io *bdev_io)
463 : {
464 0 : struct malloc_task *task = (struct malloc_task *)bdev_io->driver_ctx;
465 0 : struct malloc_disk *disk = bdev_io->bdev->ctxt;
466 0 : uint32_t block_size = bdev_io->bdev->blocklen;
467 : int rc;
468 :
469 0 : switch (bdev_io->type) {
470 : case SPDK_BDEV_IO_TYPE_READ:
471 0 : if (bdev_io->u.bdev.iovs[0].iov_base == NULL) {
472 0 : assert(bdev_io->u.bdev.iovcnt == 1);
473 0 : assert(bdev_io->u.bdev.memory_domain == NULL);
474 0 : bdev_io->u.bdev.iovs[0].iov_base =
475 0 : disk->malloc_buf + bdev_io->u.bdev.offset_blocks * block_size;
476 0 : bdev_io->u.bdev.iovs[0].iov_len = bdev_io->u.bdev.num_blocks * block_size;
477 0 : if (spdk_bdev_is_md_separate(bdev_io->bdev)) {
478 0 : spdk_bdev_io_set_md_buf(bdev_io, malloc_get_md_buf(bdev_io),
479 0 : malloc_get_md_len(bdev_io));
480 0 : }
481 0 : malloc_complete_task(task, mch, SPDK_BDEV_IO_STATUS_SUCCESS);
482 0 : return 0;
483 : }
484 :
485 0 : bdev_malloc_readv(disk, mch->accel_channel, task, bdev_io);
486 0 : return 0;
487 :
488 : case SPDK_BDEV_IO_TYPE_WRITE:
489 0 : if (bdev_io->bdev->dif_type != SPDK_DIF_DISABLE) {
490 0 : rc = malloc_verify_pi(bdev_io);
491 0 : if (rc != 0) {
492 0 : malloc_complete_task(task, mch, SPDK_BDEV_IO_STATUS_FAILED);
493 0 : return 0;
494 : }
495 0 : }
496 :
497 0 : bdev_malloc_writev(disk, mch->accel_channel, task, bdev_io);
498 0 : return 0;
499 :
500 : case SPDK_BDEV_IO_TYPE_RESET:
501 0 : malloc_complete_task(task, mch, SPDK_BDEV_IO_STATUS_SUCCESS);
502 0 : return 0;
503 :
504 : case SPDK_BDEV_IO_TYPE_FLUSH:
505 0 : malloc_complete_task(task, mch, SPDK_BDEV_IO_STATUS_SUCCESS);
506 0 : return 0;
507 :
508 : case SPDK_BDEV_IO_TYPE_UNMAP:
509 0 : return bdev_malloc_unmap(disk, mch->accel_channel, task,
510 0 : bdev_io->u.bdev.offset_blocks * block_size,
511 0 : bdev_io->u.bdev.num_blocks * block_size);
512 :
513 : case SPDK_BDEV_IO_TYPE_WRITE_ZEROES:
514 : /* bdev_malloc_unmap is implemented with a call to mem_cpy_fill which zeroes out all of the requested bytes. */
515 0 : return bdev_malloc_unmap(disk, mch->accel_channel, task,
516 0 : bdev_io->u.bdev.offset_blocks * block_size,
517 0 : bdev_io->u.bdev.num_blocks * block_size);
518 :
519 : case SPDK_BDEV_IO_TYPE_ZCOPY:
520 0 : if (bdev_io->u.bdev.zcopy.start) {
521 : void *buf;
522 : size_t len;
523 :
524 0 : buf = disk->malloc_buf + bdev_io->u.bdev.offset_blocks * block_size;
525 0 : len = bdev_io->u.bdev.num_blocks * block_size;
526 0 : spdk_bdev_io_set_buf(bdev_io, buf, len);
527 0 : if (spdk_bdev_is_md_separate(bdev_io->bdev)) {
528 0 : spdk_bdev_io_set_md_buf(bdev_io, malloc_get_md_buf(bdev_io),
529 0 : malloc_get_md_len(bdev_io));
530 0 : }
531 0 : }
532 0 : malloc_complete_task(task, mch, SPDK_BDEV_IO_STATUS_SUCCESS);
533 0 : return 0;
534 : case SPDK_BDEV_IO_TYPE_ABORT:
535 0 : malloc_complete_task(task, mch, SPDK_BDEV_IO_STATUS_FAILED);
536 0 : return 0;
537 : case SPDK_BDEV_IO_TYPE_COPY:
538 0 : bdev_malloc_copy(disk, mch->accel_channel, task,
539 0 : bdev_io->u.bdev.offset_blocks * block_size,
540 0 : bdev_io->u.bdev.copy.src_offset_blocks * block_size,
541 0 : bdev_io->u.bdev.num_blocks * block_size);
542 0 : return 0;
543 :
544 : default:
545 0 : return -1;
546 : }
547 : return 0;
548 0 : }
549 :
550 : static void
551 0 : bdev_malloc_submit_request(struct spdk_io_channel *ch, struct spdk_bdev_io *bdev_io)
552 : {
553 0 : struct malloc_channel *mch = spdk_io_channel_get_ctx(ch);
554 :
555 0 : if (_bdev_malloc_submit_request(mch, bdev_io) != 0) {
556 0 : malloc_complete_task((struct malloc_task *)bdev_io->driver_ctx, mch,
557 : SPDK_BDEV_IO_STATUS_FAILED);
558 0 : }
559 0 : }
560 :
561 : static bool
562 0 : bdev_malloc_io_type_supported(void *ctx, enum spdk_bdev_io_type io_type)
563 : {
564 0 : switch (io_type) {
565 : case SPDK_BDEV_IO_TYPE_READ:
566 : case SPDK_BDEV_IO_TYPE_WRITE:
567 : case SPDK_BDEV_IO_TYPE_FLUSH:
568 : case SPDK_BDEV_IO_TYPE_RESET:
569 : case SPDK_BDEV_IO_TYPE_UNMAP:
570 : case SPDK_BDEV_IO_TYPE_WRITE_ZEROES:
571 : case SPDK_BDEV_IO_TYPE_ZCOPY:
572 : case SPDK_BDEV_IO_TYPE_ABORT:
573 : case SPDK_BDEV_IO_TYPE_COPY:
574 0 : return true;
575 :
576 : default:
577 0 : return false;
578 : }
579 0 : }
580 :
581 : static struct spdk_io_channel *
582 0 : bdev_malloc_get_io_channel(void *ctx)
583 : {
584 0 : return spdk_get_io_channel(&g_malloc_disks);
585 : }
586 :
587 : static void
588 0 : bdev_malloc_write_json_config(struct spdk_bdev *bdev, struct spdk_json_write_ctx *w)
589 : {
590 0 : spdk_json_write_object_begin(w);
591 :
592 0 : spdk_json_write_named_string(w, "method", "bdev_malloc_create");
593 :
594 0 : spdk_json_write_named_object_begin(w, "params");
595 0 : spdk_json_write_named_string(w, "name", bdev->name);
596 0 : spdk_json_write_named_uint64(w, "num_blocks", bdev->blockcnt);
597 0 : spdk_json_write_named_uint32(w, "block_size", bdev->blocklen);
598 0 : spdk_json_write_named_uint32(w, "physical_block_size", bdev->phys_blocklen);
599 0 : spdk_json_write_named_uuid(w, "uuid", &bdev->uuid);
600 0 : spdk_json_write_named_uint32(w, "optimal_io_boundary", bdev->optimal_io_boundary);
601 0 : spdk_json_write_named_uint32(w, "md_size", bdev->md_len);
602 0 : spdk_json_write_named_uint32(w, "dif_type", bdev->dif_type);
603 0 : spdk_json_write_named_bool(w, "dif_is_head_of_md", bdev->dif_is_head_of_md);
604 0 : spdk_json_write_named_uint32(w, "dif_pi_format", bdev->dif_pi_format);
605 :
606 0 : spdk_json_write_object_end(w);
607 :
608 0 : spdk_json_write_object_end(w);
609 0 : }
610 :
611 : static int
612 0 : bdev_malloc_get_memory_domains(void *ctx, struct spdk_memory_domain **domains, int array_size)
613 : {
614 0 : struct malloc_disk *malloc_disk = ctx;
615 : struct spdk_memory_domain *domain;
616 0 : int num_domains = 0;
617 :
618 0 : if (malloc_disk->disk.dif_type != SPDK_DIF_DISABLE) {
619 0 : return 0;
620 : }
621 :
622 : /* Report support for every memory domain */
623 0 : for (domain = spdk_memory_domain_get_first(NULL); domain != NULL;
624 0 : domain = spdk_memory_domain_get_next(domain, NULL)) {
625 0 : if (domains != NULL && num_domains < array_size) {
626 0 : domains[num_domains] = domain;
627 0 : }
628 0 : num_domains++;
629 0 : }
630 :
631 0 : return num_domains;
632 0 : }
633 :
634 : static bool
635 0 : bdev_malloc_accel_sequence_supported(void *ctx, enum spdk_bdev_io_type type)
636 : {
637 0 : struct malloc_disk *malloc_disk = ctx;
638 :
639 0 : if (malloc_disk->disk.dif_type != SPDK_DIF_DISABLE) {
640 0 : return false;
641 : }
642 :
643 0 : switch (type) {
644 : case SPDK_BDEV_IO_TYPE_READ:
645 : case SPDK_BDEV_IO_TYPE_WRITE:
646 0 : return true;
647 : default:
648 0 : return false;
649 : }
650 0 : }
651 :
652 : static const struct spdk_bdev_fn_table malloc_fn_table = {
653 : .destruct = bdev_malloc_destruct,
654 : .submit_request = bdev_malloc_submit_request,
655 : .io_type_supported = bdev_malloc_io_type_supported,
656 : .get_io_channel = bdev_malloc_get_io_channel,
657 : .write_config_json = bdev_malloc_write_json_config,
658 : .get_memory_domains = bdev_malloc_get_memory_domains,
659 : .accel_sequence_supported = bdev_malloc_accel_sequence_supported,
660 : };
661 :
662 : static int
663 0 : malloc_disk_setup_pi(struct malloc_disk *mdisk)
664 : {
665 0 : struct spdk_bdev *bdev = &mdisk->disk;
666 : struct spdk_dif_ctx dif_ctx;
667 : struct iovec iov, md_iov;
668 : uint32_t dif_check_flags;
669 : int rc;
670 : struct spdk_dif_ctx_init_ext_opts dif_opts;
671 :
672 0 : dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format);
673 0 : dif_opts.dif_pi_format = bdev->dif_pi_format;
674 : /* Set APPTAG|REFTAG_IGNORE to PI fields after creation of malloc bdev */
675 0 : dif_check_flags = bdev->dif_check_flags | SPDK_DIF_CHECK_TYPE_REFTAG |
676 : SPDK_DIF_FLAGS_APPTAG_CHECK;
677 0 : rc = spdk_dif_ctx_init(&dif_ctx,
678 0 : bdev->blocklen,
679 0 : bdev->md_len,
680 0 : bdev->md_interleave,
681 0 : bdev->dif_is_head_of_md,
682 0 : bdev->dif_type,
683 0 : dif_check_flags,
684 : SPDK_DIF_REFTAG_IGNORE,
685 : 0xFFFF, SPDK_DIF_APPTAG_IGNORE,
686 : 0, 0, &dif_opts);
687 0 : if (rc != 0) {
688 0 : SPDK_ERRLOG("Initialization of DIF/DIX context failed\n");
689 0 : return rc;
690 : }
691 :
692 0 : iov.iov_base = mdisk->malloc_buf;
693 0 : iov.iov_len = bdev->blockcnt * bdev->blocklen;
694 :
695 0 : if (mdisk->disk.md_interleave) {
696 0 : rc = spdk_dif_generate(&iov, 1, bdev->blockcnt, &dif_ctx);
697 0 : } else {
698 0 : md_iov.iov_base = mdisk->malloc_md_buf;
699 0 : md_iov.iov_len = bdev->blockcnt * bdev->md_len;
700 :
701 0 : rc = spdk_dix_generate(&iov, 1, &md_iov, bdev->blockcnt, &dif_ctx);
702 : }
703 :
704 0 : if (rc != 0) {
705 0 : SPDK_ERRLOG("Formatting by DIF/DIX failed\n");
706 0 : }
707 :
708 0 : return rc;
709 0 : }
710 :
711 : int
712 0 : create_malloc_disk(struct spdk_bdev **bdev, const struct malloc_bdev_opts *opts)
713 : {
714 : struct malloc_disk *mdisk;
715 : uint32_t block_size;
716 : int rc;
717 :
718 0 : assert(opts != NULL);
719 :
720 0 : if (opts->num_blocks == 0) {
721 0 : SPDK_ERRLOG("Disk num_blocks must be greater than 0");
722 0 : return -EINVAL;
723 : }
724 :
725 0 : if (opts->block_size % 512) {
726 0 : SPDK_ERRLOG("Data block size must be 512 bytes aligned\n");
727 0 : return -EINVAL;
728 : }
729 :
730 0 : if (opts->physical_block_size % 512) {
731 0 : SPDK_ERRLOG("Physical block must be 512 bytes aligned\n");
732 0 : return -EINVAL;
733 : }
734 :
735 0 : switch (opts->md_size) {
736 : case 0:
737 : case 8:
738 : case 16:
739 : case 32:
740 : case 64:
741 : case 128:
742 0 : break;
743 : default:
744 0 : SPDK_ERRLOG("metadata size %u is not supported\n", opts->md_size);
745 0 : return -EINVAL;
746 : }
747 :
748 0 : if (opts->md_interleave) {
749 0 : block_size = opts->block_size + opts->md_size;
750 0 : } else {
751 0 : block_size = opts->block_size;
752 : }
753 :
754 0 : mdisk = calloc(1, sizeof(*mdisk));
755 0 : if (!mdisk) {
756 0 : SPDK_ERRLOG("mdisk calloc() failed\n");
757 0 : return -ENOMEM;
758 : }
759 :
760 : /*
761 : * Allocate the large backend memory buffer from pinned memory.
762 : *
763 : * TODO: need to pass a hint so we know which socket to allocate
764 : * from on multi-socket systems.
765 : */
766 0 : mdisk->malloc_buf = spdk_zmalloc(opts->num_blocks * block_size, 2 * 1024 * 1024, NULL,
767 : SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
768 0 : if (!mdisk->malloc_buf) {
769 0 : SPDK_ERRLOG("malloc_buf spdk_zmalloc() failed\n");
770 0 : malloc_disk_free(mdisk);
771 0 : return -ENOMEM;
772 : }
773 :
774 0 : if (!opts->md_interleave && opts->md_size != 0) {
775 0 : mdisk->malloc_md_buf = spdk_zmalloc(opts->num_blocks * opts->md_size, 2 * 1024 * 1024, NULL,
776 : SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
777 0 : if (!mdisk->malloc_md_buf) {
778 0 : SPDK_ERRLOG("malloc_md_buf spdk_zmalloc() failed\n");
779 0 : malloc_disk_free(mdisk);
780 0 : return -ENOMEM;
781 : }
782 0 : }
783 :
784 0 : if (opts->name) {
785 0 : mdisk->disk.name = strdup(opts->name);
786 0 : } else {
787 : /* Auto-generate a name */
788 0 : mdisk->disk.name = spdk_sprintf_alloc("Malloc%d", malloc_disk_count);
789 0 : malloc_disk_count++;
790 : }
791 0 : if (!mdisk->disk.name) {
792 0 : malloc_disk_free(mdisk);
793 0 : return -ENOMEM;
794 : }
795 0 : mdisk->disk.product_name = "Malloc disk";
796 :
797 0 : mdisk->disk.write_cache = 1;
798 0 : mdisk->disk.blocklen = block_size;
799 0 : mdisk->disk.phys_blocklen = opts->physical_block_size;
800 0 : mdisk->disk.blockcnt = opts->num_blocks;
801 0 : mdisk->disk.md_len = opts->md_size;
802 0 : mdisk->disk.md_interleave = opts->md_interleave;
803 0 : mdisk->disk.dif_type = opts->dif_type;
804 0 : mdisk->disk.dif_is_head_of_md = opts->dif_is_head_of_md;
805 : /* Current block device layer API does not propagate
806 : * any DIF related information from user. So, we can
807 : * not generate or verify Application Tag.
808 : */
809 0 : switch (opts->dif_type) {
810 : case SPDK_DIF_TYPE1:
811 : case SPDK_DIF_TYPE2:
812 0 : mdisk->disk.dif_check_flags = SPDK_DIF_FLAGS_GUARD_CHECK |
813 : SPDK_DIF_FLAGS_REFTAG_CHECK;
814 0 : break;
815 : case SPDK_DIF_TYPE3:
816 0 : mdisk->disk.dif_check_flags = SPDK_DIF_FLAGS_GUARD_CHECK;
817 0 : break;
818 : case SPDK_DIF_DISABLE:
819 0 : break;
820 : }
821 0 : mdisk->disk.dif_pi_format = opts->dif_pi_format;
822 :
823 0 : if (opts->dif_type != SPDK_DIF_DISABLE) {
824 0 : rc = malloc_disk_setup_pi(mdisk);
825 0 : if (rc) {
826 0 : SPDK_ERRLOG("Failed to set up protection information.\n");
827 0 : malloc_disk_free(mdisk);
828 0 : return rc;
829 : }
830 0 : }
831 :
832 0 : if (opts->optimal_io_boundary) {
833 0 : mdisk->disk.optimal_io_boundary = opts->optimal_io_boundary;
834 0 : mdisk->disk.split_on_optimal_io_boundary = true;
835 0 : }
836 0 : if (!spdk_uuid_is_null(&opts->uuid)) {
837 0 : spdk_uuid_copy(&mdisk->disk.uuid, &opts->uuid);
838 0 : }
839 :
840 0 : mdisk->disk.max_copy = 0;
841 0 : mdisk->disk.ctxt = mdisk;
842 0 : mdisk->disk.fn_table = &malloc_fn_table;
843 0 : mdisk->disk.module = &malloc_if;
844 :
845 0 : rc = spdk_bdev_register(&mdisk->disk);
846 0 : if (rc) {
847 0 : malloc_disk_free(mdisk);
848 0 : return rc;
849 : }
850 :
851 0 : *bdev = &(mdisk->disk);
852 :
853 0 : TAILQ_INSERT_TAIL(&g_malloc_disks, mdisk, link);
854 :
855 0 : return rc;
856 0 : }
857 :
858 : void
859 0 : delete_malloc_disk(const char *name, spdk_delete_malloc_complete cb_fn, void *cb_arg)
860 : {
861 : int rc;
862 :
863 0 : rc = spdk_bdev_unregister_by_name(name, &malloc_if, cb_fn, cb_arg);
864 0 : if (rc != 0) {
865 0 : cb_fn(cb_arg, rc);
866 0 : }
867 0 : }
868 :
869 : static int
870 0 : malloc_completion_poller(void *ctx)
871 : {
872 0 : struct malloc_channel *ch = ctx;
873 : struct malloc_task *task;
874 : TAILQ_HEAD(, malloc_task) completed_tasks;
875 0 : uint32_t num_completions = 0;
876 :
877 0 : TAILQ_INIT(&completed_tasks);
878 0 : TAILQ_SWAP(&completed_tasks, &ch->completed_tasks, malloc_task, tailq);
879 :
880 0 : while (!TAILQ_EMPTY(&completed_tasks)) {
881 0 : task = TAILQ_FIRST(&completed_tasks);
882 0 : TAILQ_REMOVE(&completed_tasks, task, tailq);
883 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(task), task->status);
884 0 : num_completions++;
885 : }
886 :
887 0 : return num_completions > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
888 : }
889 :
890 : static int
891 0 : malloc_create_channel_cb(void *io_device, void *ctx)
892 : {
893 0 : struct malloc_channel *ch = ctx;
894 :
895 0 : ch->accel_channel = spdk_accel_get_io_channel();
896 0 : if (!ch->accel_channel) {
897 0 : SPDK_ERRLOG("Failed to get accel framework's IO channel\n");
898 0 : return -ENOMEM;
899 : }
900 :
901 0 : ch->completion_poller = SPDK_POLLER_REGISTER(malloc_completion_poller, ch, 0);
902 0 : if (!ch->completion_poller) {
903 0 : SPDK_ERRLOG("Failed to register malloc completion poller\n");
904 0 : spdk_put_io_channel(ch->accel_channel);
905 0 : return -ENOMEM;
906 : }
907 :
908 0 : TAILQ_INIT(&ch->completed_tasks);
909 :
910 0 : return 0;
911 0 : }
912 :
913 : static void
914 0 : malloc_destroy_channel_cb(void *io_device, void *ctx)
915 : {
916 0 : struct malloc_channel *ch = ctx;
917 :
918 0 : assert(TAILQ_EMPTY(&ch->completed_tasks));
919 :
920 0 : spdk_put_io_channel(ch->accel_channel);
921 0 : spdk_poller_unregister(&ch->completion_poller);
922 0 : }
923 :
924 : static int
925 0 : bdev_malloc_initialize(void)
926 : {
927 : /* This needs to be reset for each reinitialization of submodules.
928 : * Otherwise after enough devices or reinitializations the value gets too high.
929 : * TODO: Make malloc bdev name mandatory and remove this counter. */
930 0 : malloc_disk_count = 0;
931 :
932 0 : spdk_io_device_register(&g_malloc_disks, malloc_create_channel_cb,
933 : malloc_destroy_channel_cb, sizeof(struct malloc_channel),
934 : "bdev_malloc");
935 :
936 0 : return 0;
937 : }
938 :
939 : static void
940 0 : bdev_malloc_deinitialize(void)
941 : {
942 0 : spdk_io_device_unregister(&g_malloc_disks, NULL);
943 0 : }
944 :
945 0 : SPDK_LOG_REGISTER_COMPONENT(bdev_malloc)
|