Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2017 Intel Corporation.
3 : * All rights reserved.
4 : * Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
5 : */
6 :
7 : #include "bdev_aio.h"
8 :
9 : #include "spdk/stdinc.h"
10 :
11 : #include "spdk/barrier.h"
12 : #include "spdk/bdev.h"
13 : #include "spdk/bdev_module.h"
14 : #include "spdk/env.h"
15 : #include "spdk/fd.h"
16 : #include "spdk/likely.h"
17 : #include "spdk/thread.h"
18 : #include "spdk/json.h"
19 : #include "spdk/util.h"
20 : #include "spdk/string.h"
21 :
22 : #include "spdk/log.h"
23 :
24 : #include <sys/eventfd.h>
25 :
26 : #ifndef __FreeBSD__
27 : #include <libaio.h>
28 : #endif
29 :
30 : struct bdev_aio_io_channel {
31 : uint64_t io_inflight;
32 : #ifdef __FreeBSD__
33 : int kqfd;
34 : #else
35 : io_context_t io_ctx;
36 : #endif
37 : struct bdev_aio_group_channel *group_ch;
38 : TAILQ_ENTRY(bdev_aio_io_channel) link;
39 : };
40 :
41 : struct bdev_aio_group_channel {
42 : /* eventfd for io completion notification in interrupt mode.
43 : * Negative value like '-1' indicates it is invalid or unused.
44 : */
45 : int efd;
46 : struct spdk_interrupt *intr;
47 : struct spdk_poller *poller;
48 : TAILQ_HEAD(, bdev_aio_io_channel) io_ch_head;
49 : };
50 :
51 : struct bdev_aio_task {
52 : #ifdef __FreeBSD__
53 : struct aiocb aiocb;
54 : #else
55 : struct iocb iocb;
56 : #endif
57 : uint64_t len;
58 : struct bdev_aio_io_channel *ch;
59 : };
60 :
61 : struct file_disk {
62 : struct bdev_aio_task *reset_task;
63 : struct spdk_poller *reset_retry_timer;
64 : struct spdk_bdev disk;
65 : char *filename;
66 : int fd;
67 : #ifdef RWF_NOWAIT
68 : bool use_nowait;
69 : #endif
70 : TAILQ_ENTRY(file_disk) link;
71 : bool block_size_override;
72 : bool readonly;
73 : bool fallocate;
74 : };
75 :
76 : /* For user space reaping of completions */
77 : struct spdk_aio_ring {
78 : uint32_t id;
79 : uint32_t size;
80 : uint32_t head;
81 : uint32_t tail;
82 :
83 : uint32_t version;
84 : uint32_t compat_features;
85 : uint32_t incompat_features;
86 : uint32_t header_length;
87 : };
88 :
89 : #define SPDK_AIO_RING_VERSION 0xa10a10a1
90 :
91 : static int bdev_aio_initialize(void);
92 : static void bdev_aio_fini(void);
93 : static void aio_free_disk(struct file_disk *fdisk);
94 : static TAILQ_HEAD(, file_disk) g_aio_disk_head = TAILQ_HEAD_INITIALIZER(g_aio_disk_head);
95 :
96 : #define SPDK_AIO_QUEUE_DEPTH 128
97 : #define MAX_EVENTS_PER_POLL 32
98 :
99 : static int
100 0 : bdev_aio_get_ctx_size(void)
101 : {
102 0 : return sizeof(struct bdev_aio_task);
103 : }
104 :
105 : static struct spdk_bdev_module aio_if = {
106 : .name = "aio",
107 : .module_init = bdev_aio_initialize,
108 : .module_fini = bdev_aio_fini,
109 : .get_ctx_size = bdev_aio_get_ctx_size,
110 : };
111 :
112 0 : SPDK_BDEV_MODULE_REGISTER(aio, &aio_if)
113 :
114 : static int
115 0 : bdev_aio_open(struct file_disk *disk)
116 : {
117 0 : int fd;
118 0 : int io_flag = disk->readonly ? O_RDONLY : O_RDWR;
119 : #ifdef RWF_NOWAIT
120 0 : struct stat st;
121 : #endif
122 :
123 0 : fd = open(disk->filename, io_flag | O_DIRECT);
124 0 : if (fd < 0) {
125 : /* Try without O_DIRECT for non-disk files */
126 0 : fd = open(disk->filename, io_flag);
127 0 : if (fd < 0) {
128 0 : SPDK_ERRLOG("open() failed (file:%s), errno %d: %s\n",
129 : disk->filename, errno, spdk_strerror(errno));
130 0 : disk->fd = -1;
131 0 : return -1;
132 : }
133 0 : }
134 :
135 0 : disk->fd = fd;
136 :
137 : #ifdef RWF_NOWAIT
138 : /* Some aio operations can block, for example if number outstanding
139 : * I/O exceeds number of block layer tags. But not all files can
140 : * support RWF_NOWAIT flag. So use RWF_NOWAIT on block devices only.
141 : */
142 0 : disk->use_nowait = fstat(fd, &st) == 0 && S_ISBLK(st.st_mode);
143 : #endif
144 :
145 0 : return 0;
146 0 : }
147 :
148 : static int
149 0 : bdev_aio_close(struct file_disk *disk)
150 : {
151 0 : int rc;
152 :
153 0 : if (disk->fd == -1) {
154 0 : return 0;
155 : }
156 :
157 0 : rc = close(disk->fd);
158 0 : if (rc < 0) {
159 0 : SPDK_ERRLOG("close() failed (fd=%d), errno %d: %s\n",
160 : disk->fd, errno, spdk_strerror(errno));
161 0 : return -1;
162 : }
163 :
164 0 : disk->fd = -1;
165 :
166 0 : return 0;
167 0 : }
168 :
169 : #ifdef __FreeBSD__
170 : static int
171 : bdev_aio_submit_io(enum spdk_bdev_io_type type, struct file_disk *fdisk,
172 : struct spdk_io_channel *ch, struct bdev_aio_task *aio_task,
173 : struct iovec *iov, int iovcnt, uint64_t nbytes, uint64_t offset)
174 : {
175 : struct aiocb *aiocb = &aio_task->aiocb;
176 : struct bdev_aio_io_channel *aio_ch = spdk_io_channel_get_ctx(ch);
177 :
178 : memset(aiocb, 0, sizeof(struct aiocb));
179 : aiocb->aio_fildes = fdisk->fd;
180 : aiocb->aio_iov = iov;
181 : aiocb->aio_iovcnt = iovcnt;
182 : aiocb->aio_offset = offset;
183 : aiocb->aio_sigevent.sigev_notify_kqueue = aio_ch->kqfd;
184 : aiocb->aio_sigevent.sigev_value.sival_ptr = aio_task;
185 : aiocb->aio_sigevent.sigev_notify = SIGEV_KEVENT;
186 :
187 : aio_task->len = nbytes;
188 : aio_task->ch = aio_ch;
189 :
190 : if (type == SPDK_BDEV_IO_TYPE_READ) {
191 : return aio_readv(aiocb);
192 : }
193 :
194 : return aio_writev(aiocb);
195 : }
196 : #else
197 : static int
198 0 : bdev_aio_submit_io(enum spdk_bdev_io_type type, struct file_disk *fdisk,
199 : struct spdk_io_channel *ch, struct bdev_aio_task *aio_task,
200 : struct iovec *iov, int iovcnt, uint64_t nbytes, uint64_t offset)
201 : {
202 0 : struct iocb *iocb = &aio_task->iocb;
203 0 : struct bdev_aio_io_channel *aio_ch = spdk_io_channel_get_ctx(ch);
204 :
205 0 : if (type == SPDK_BDEV_IO_TYPE_READ) {
206 0 : io_prep_preadv(iocb, fdisk->fd, iov, iovcnt, offset);
207 0 : } else {
208 0 : io_prep_pwritev(iocb, fdisk->fd, iov, iovcnt, offset);
209 : }
210 :
211 0 : if (aio_ch->group_ch->efd >= 0) {
212 0 : io_set_eventfd(iocb, aio_ch->group_ch->efd);
213 0 : }
214 0 : iocb->data = aio_task;
215 : #ifdef RWF_NOWAIT
216 0 : if (fdisk->use_nowait) {
217 0 : iocb->aio_rw_flags = RWF_NOWAIT;
218 0 : }
219 : #endif
220 0 : aio_task->len = nbytes;
221 0 : aio_task->ch = aio_ch;
222 :
223 0 : return io_submit(aio_ch->io_ctx, 1, &iocb);
224 0 : }
225 : #endif
226 :
227 : static void
228 0 : bdev_aio_rw(enum spdk_bdev_io_type type, struct file_disk *fdisk,
229 : struct spdk_io_channel *ch, struct bdev_aio_task *aio_task,
230 : struct iovec *iov, int iovcnt, uint64_t nbytes, uint64_t offset)
231 : {
232 0 : struct bdev_aio_io_channel *aio_ch = spdk_io_channel_get_ctx(ch);
233 0 : int rc;
234 :
235 0 : if (type == SPDK_BDEV_IO_TYPE_READ) {
236 0 : SPDK_DEBUGLOG(aio, "read %d iovs size %lu to off: %#lx\n",
237 : iovcnt, nbytes, offset);
238 0 : } else {
239 0 : SPDK_DEBUGLOG(aio, "write %d iovs size %lu from off: %#lx\n",
240 : iovcnt, nbytes, offset);
241 : }
242 :
243 0 : rc = bdev_aio_submit_io(type, fdisk, ch, aio_task, iov, iovcnt, nbytes, offset);
244 0 : if (spdk_unlikely(rc < 0)) {
245 0 : if (rc == -EAGAIN) {
246 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_NOMEM);
247 0 : } else {
248 0 : spdk_bdev_io_complete_aio_status(spdk_bdev_io_from_ctx(aio_task), rc);
249 0 : SPDK_ERRLOG("%s: io_submit returned %d\n", __func__, rc);
250 : }
251 0 : } else {
252 0 : aio_ch->io_inflight++;
253 : }
254 0 : }
255 :
256 : static void
257 0 : bdev_aio_flush(struct file_disk *fdisk, struct bdev_aio_task *aio_task)
258 : {
259 0 : int rc = fsync(fdisk->fd);
260 :
261 0 : if (rc == 0) {
262 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_SUCCESS);
263 0 : } else {
264 0 : spdk_bdev_io_complete_aio_status(spdk_bdev_io_from_ctx(aio_task), -errno);
265 : }
266 0 : }
267 :
268 : #ifndef __FreeBSD__
269 : static void
270 0 : bdev_aio_fallocate(struct spdk_bdev_io *bdev_io, int mode)
271 : {
272 0 : struct file_disk *fdisk = (struct file_disk *)bdev_io->bdev->ctxt;
273 0 : struct bdev_aio_task *aio_task = (struct bdev_aio_task *)bdev_io->driver_ctx;
274 0 : uint64_t offset_bytes = bdev_io->u.bdev.offset_blocks * bdev_io->bdev->blocklen;
275 0 : uint64_t length_bytes = bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen;
276 0 : int rc;
277 :
278 0 : if (!fdisk->fallocate) {
279 0 : spdk_bdev_io_complete_aio_status(spdk_bdev_io_from_ctx(aio_task), -ENOTSUP);
280 0 : return;
281 : }
282 :
283 0 : rc = fallocate(fdisk->fd, mode, offset_bytes, length_bytes);
284 0 : if (rc == 0) {
285 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_SUCCESS);
286 0 : } else {
287 0 : spdk_bdev_io_complete_aio_status(spdk_bdev_io_from_ctx(aio_task), -errno);
288 : }
289 0 : }
290 :
291 : static void
292 0 : bdev_aio_unmap(struct spdk_bdev_io *bdev_io)
293 : {
294 0 : int mode = FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE;
295 :
296 0 : bdev_aio_fallocate(bdev_io, mode);
297 0 : }
298 :
299 :
300 : static void
301 0 : bdev_aio_write_zeros(struct spdk_bdev_io *bdev_io)
302 : {
303 0 : int mode = FALLOC_FL_ZERO_RANGE;
304 :
305 0 : bdev_aio_fallocate(bdev_io, mode);
306 0 : }
307 : #endif
308 :
309 : static void
310 0 : bdev_aio_destruct_cb(void *io_device)
311 : {
312 0 : struct file_disk *fdisk = io_device;
313 0 : int rc = 0;
314 :
315 0 : TAILQ_REMOVE(&g_aio_disk_head, fdisk, link);
316 0 : rc = bdev_aio_close(fdisk);
317 0 : if (rc < 0) {
318 0 : SPDK_ERRLOG("bdev_aio_close() failed\n");
319 0 : }
320 0 : aio_free_disk(fdisk);
321 0 : }
322 :
323 : static int
324 0 : bdev_aio_destruct(void *ctx)
325 : {
326 0 : struct file_disk *fdisk = ctx;
327 :
328 0 : spdk_io_device_unregister(fdisk, bdev_aio_destruct_cb);
329 :
330 0 : return 0;
331 0 : }
332 :
333 : #ifdef __FreeBSD__
334 : static int
335 : bdev_user_io_getevents(int kq, unsigned int max, struct kevent *events)
336 : {
337 : struct timespec ts;
338 : int count;
339 :
340 : memset(events, 0, max * sizeof(struct kevent));
341 : memset(&ts, 0, sizeof(ts));
342 :
343 : count = kevent(kq, NULL, 0, events, max, &ts);
344 : if (count < 0) {
345 : SPDK_ERRLOG("failed to get kevents: %s.\n", spdk_strerror(errno));
346 : return -errno;
347 : }
348 :
349 : return count;
350 : }
351 :
352 : static int
353 : bdev_aio_io_channel_poll(struct bdev_aio_io_channel *io_ch)
354 : {
355 : int nr, i, res = 0;
356 : struct bdev_aio_task *aio_task;
357 : struct kevent events[SPDK_AIO_QUEUE_DEPTH];
358 :
359 : nr = bdev_user_io_getevents(io_ch->kqfd, SPDK_AIO_QUEUE_DEPTH, events);
360 : if (nr < 0) {
361 : return 0;
362 : }
363 :
364 : for (i = 0; i < nr; i++) {
365 : aio_task = events[i].udata;
366 : aio_task->ch->io_inflight--;
367 : if (aio_task == NULL) {
368 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_FAILED);
369 : break;
370 : } else if ((uint64_t)aio_return(&aio_task->aiocb) == aio_task->len) {
371 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_SUCCESS);
372 : } else {
373 : SPDK_ERRLOG("failed to complete aio: rc %d\n", aio_error(&aio_task->aiocb));
374 : res = aio_error(&aio_task->aiocb);
375 : if (res != 0) {
376 : spdk_bdev_io_complete_aio_status(spdk_bdev_io_from_ctx(aio_task), res);
377 : } else {
378 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_FAILED);
379 : }
380 : }
381 : }
382 :
383 : return nr;
384 : }
385 : #else
386 : static int
387 0 : bdev_user_io_getevents(io_context_t io_ctx, unsigned int max, struct io_event *uevents)
388 : {
389 0 : uint32_t head, tail, count;
390 0 : struct spdk_aio_ring *ring;
391 0 : struct timespec timeout;
392 0 : struct io_event *kevents;
393 :
394 0 : ring = (struct spdk_aio_ring *)io_ctx;
395 :
396 0 : if (spdk_unlikely(ring->version != SPDK_AIO_RING_VERSION || ring->incompat_features != 0)) {
397 0 : timeout.tv_sec = 0;
398 0 : timeout.tv_nsec = 0;
399 :
400 0 : return io_getevents(io_ctx, 0, max, uevents, &timeout);
401 : }
402 :
403 : /* Read the current state out of the ring */
404 0 : head = ring->head;
405 0 : tail = ring->tail;
406 :
407 : /* This memory barrier is required to prevent the loads above
408 : * from being re-ordered with stores to the events array
409 : * potentially occurring on other threads. */
410 0 : spdk_smp_rmb();
411 :
412 : /* Calculate how many items are in the circular ring */
413 0 : count = tail - head;
414 0 : if (tail < head) {
415 0 : count += ring->size;
416 0 : }
417 :
418 : /* Reduce the count to the limit provided by the user */
419 0 : count = spdk_min(max, count);
420 :
421 : /* Grab the memory location of the event array */
422 0 : kevents = (struct io_event *)((uintptr_t)ring + ring->header_length);
423 :
424 : /* Copy the events out of the ring. */
425 0 : if ((head + count) <= ring->size) {
426 : /* Only one copy is required */
427 0 : memcpy(uevents, &kevents[head], count * sizeof(struct io_event));
428 0 : } else {
429 0 : uint32_t first_part = ring->size - head;
430 : /* Two copies are required */
431 0 : memcpy(uevents, &kevents[head], first_part * sizeof(struct io_event));
432 0 : memcpy(&uevents[first_part], &kevents[0], (count - first_part) * sizeof(struct io_event));
433 0 : }
434 :
435 : /* Update the head pointer. On x86, stores will not be reordered with older loads,
436 : * so the copies out of the event array will always be complete prior to this
437 : * update becoming visible. On other architectures this is not guaranteed, so
438 : * add a barrier. */
439 : #if defined(__i386__) || defined(__x86_64__)
440 0 : spdk_compiler_barrier();
441 : #else
442 : spdk_smp_mb();
443 : #endif
444 0 : ring->head = (head + count) % ring->size;
445 :
446 0 : return count;
447 0 : }
448 :
449 : static int
450 0 : bdev_aio_io_channel_poll(struct bdev_aio_io_channel *io_ch)
451 : {
452 0 : int nr, i, res = 0;
453 0 : struct bdev_aio_task *aio_task;
454 0 : struct io_event events[SPDK_AIO_QUEUE_DEPTH];
455 :
456 0 : nr = bdev_user_io_getevents(io_ch->io_ctx, SPDK_AIO_QUEUE_DEPTH, events);
457 0 : if (nr < 0) {
458 0 : return 0;
459 : }
460 :
461 0 : for (i = 0; i < nr; i++) {
462 0 : aio_task = events[i].data;
463 0 : aio_task->ch->io_inflight--;
464 0 : if (events[i].res == aio_task->len) {
465 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_SUCCESS);
466 0 : } else {
467 : /* From aio_abi.h, io_event.res is defined __s64, negative errno
468 : * will be assigned to io_event.res for error situation.
469 : * But from libaio.h, io_event.res is defined unsigned long, so
470 : * convert it to signed value for error detection.
471 : */
472 0 : res = (int)events[i].res;
473 0 : if (res < 0) {
474 0 : if (res == -EAGAIN) {
475 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_NOMEM);
476 0 : } else {
477 0 : SPDK_ERRLOG("failed to complete aio: rc %"PRId64"\n", events[i].res);
478 0 : spdk_bdev_io_complete_aio_status(spdk_bdev_io_from_ctx(aio_task), res);
479 : }
480 0 : } else {
481 0 : SPDK_ERRLOG("failed to complete aio: rc %"PRId64"\n", events[i].res);
482 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(aio_task), SPDK_BDEV_IO_STATUS_FAILED);
483 : }
484 : }
485 0 : }
486 :
487 0 : return nr;
488 0 : }
489 : #endif
490 :
491 : static int
492 0 : bdev_aio_group_poll(void *arg)
493 : {
494 0 : struct bdev_aio_group_channel *group_ch = arg;
495 0 : struct bdev_aio_io_channel *io_ch;
496 0 : int nr = 0;
497 :
498 0 : TAILQ_FOREACH(io_ch, &group_ch->io_ch_head, link) {
499 0 : nr += bdev_aio_io_channel_poll(io_ch);
500 0 : }
501 :
502 0 : return nr > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
503 0 : }
504 :
505 : static int
506 0 : bdev_aio_group_interrupt(void *arg)
507 : {
508 0 : struct bdev_aio_group_channel *group_ch = arg;
509 0 : int rc;
510 0 : uint64_t num_events;
511 :
512 0 : assert(group_ch->efd >= 0);
513 :
514 : /* if completed IO number is larger than SPDK_AIO_QUEUE_DEPTH,
515 : * io_getevent should be called again to ensure all completed IO are processed.
516 : */
517 0 : rc = read(group_ch->efd, &num_events, sizeof(num_events));
518 0 : if (rc < 0) {
519 0 : SPDK_ERRLOG("failed to acknowledge aio group: %s.\n", spdk_strerror(errno));
520 0 : return -errno;
521 : }
522 :
523 0 : if (num_events > SPDK_AIO_QUEUE_DEPTH) {
524 0 : num_events -= SPDK_AIO_QUEUE_DEPTH;
525 0 : rc = write(group_ch->efd, &num_events, sizeof(num_events));
526 0 : if (rc < 0) {
527 0 : SPDK_ERRLOG("failed to notify aio group: %s.\n", spdk_strerror(errno));
528 0 : }
529 0 : }
530 :
531 0 : return bdev_aio_group_poll(group_ch);
532 0 : }
533 :
534 : static void
535 0 : _bdev_aio_get_io_inflight(struct spdk_io_channel_iter *i)
536 : {
537 0 : struct spdk_io_channel *ch = spdk_io_channel_iter_get_channel(i);
538 0 : struct bdev_aio_io_channel *aio_ch = spdk_io_channel_get_ctx(ch);
539 :
540 0 : if (aio_ch->io_inflight) {
541 0 : spdk_for_each_channel_continue(i, -1);
542 0 : return;
543 : }
544 :
545 0 : spdk_for_each_channel_continue(i, 0);
546 0 : }
547 :
548 : static int bdev_aio_reset_retry_timer(void *arg);
549 :
550 : static void
551 0 : _bdev_aio_get_io_inflight_done(struct spdk_io_channel_iter *i, int status)
552 : {
553 0 : struct file_disk *fdisk = spdk_io_channel_iter_get_ctx(i);
554 :
555 0 : if (status == -1) {
556 0 : fdisk->reset_retry_timer = SPDK_POLLER_REGISTER(bdev_aio_reset_retry_timer, fdisk, 500);
557 0 : return;
558 : }
559 :
560 0 : spdk_bdev_io_complete(spdk_bdev_io_from_ctx(fdisk->reset_task), SPDK_BDEV_IO_STATUS_SUCCESS);
561 0 : }
562 :
563 : static int
564 0 : bdev_aio_reset_retry_timer(void *arg)
565 : {
566 0 : struct file_disk *fdisk = arg;
567 :
568 0 : if (fdisk->reset_retry_timer) {
569 0 : spdk_poller_unregister(&fdisk->reset_retry_timer);
570 0 : }
571 :
572 0 : spdk_for_each_channel(fdisk,
573 : _bdev_aio_get_io_inflight,
574 0 : fdisk,
575 : _bdev_aio_get_io_inflight_done);
576 :
577 0 : return SPDK_POLLER_BUSY;
578 0 : }
579 :
580 : static void
581 0 : bdev_aio_reset(struct file_disk *fdisk, struct bdev_aio_task *aio_task)
582 : {
583 0 : fdisk->reset_task = aio_task;
584 :
585 0 : bdev_aio_reset_retry_timer(fdisk);
586 0 : }
587 :
588 : static void
589 0 : bdev_aio_get_buf_cb(struct spdk_io_channel *ch, struct spdk_bdev_io *bdev_io,
590 : bool success)
591 : {
592 0 : if (!success) {
593 0 : spdk_bdev_io_complete(bdev_io, SPDK_BDEV_IO_STATUS_FAILED);
594 0 : return;
595 : }
596 :
597 0 : switch (bdev_io->type) {
598 : case SPDK_BDEV_IO_TYPE_READ:
599 : case SPDK_BDEV_IO_TYPE_WRITE:
600 0 : bdev_aio_rw(bdev_io->type,
601 0 : (struct file_disk *)bdev_io->bdev->ctxt,
602 0 : ch,
603 0 : (struct bdev_aio_task *)bdev_io->driver_ctx,
604 0 : bdev_io->u.bdev.iovs,
605 0 : bdev_io->u.bdev.iovcnt,
606 0 : bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen,
607 0 : bdev_io->u.bdev.offset_blocks * bdev_io->bdev->blocklen);
608 0 : break;
609 : default:
610 0 : SPDK_ERRLOG("Wrong io type\n");
611 0 : break;
612 : }
613 0 : }
614 :
615 : static int
616 0 : _bdev_aio_submit_request(struct spdk_io_channel *ch, struct spdk_bdev_io *bdev_io)
617 : {
618 0 : struct file_disk *fdisk = (struct file_disk *)bdev_io->bdev->ctxt;
619 :
620 0 : switch (bdev_io->type) {
621 : /* Read and write operations must be performed on buffers aligned to
622 : * bdev->required_alignment. If user specified unaligned buffers,
623 : * get the aligned buffer from the pool by calling spdk_bdev_io_get_buf. */
624 : case SPDK_BDEV_IO_TYPE_READ:
625 0 : spdk_bdev_io_get_buf(bdev_io, bdev_aio_get_buf_cb,
626 0 : bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen);
627 0 : return 0;
628 : case SPDK_BDEV_IO_TYPE_WRITE:
629 0 : if (fdisk->readonly) {
630 0 : spdk_bdev_io_complete(bdev_io, SPDK_BDEV_IO_STATUS_FAILED);
631 0 : } else {
632 0 : spdk_bdev_io_get_buf(bdev_io, bdev_aio_get_buf_cb,
633 0 : bdev_io->u.bdev.num_blocks * bdev_io->bdev->blocklen);
634 : }
635 0 : return 0;
636 :
637 : case SPDK_BDEV_IO_TYPE_FLUSH:
638 0 : bdev_aio_flush((struct file_disk *)bdev_io->bdev->ctxt,
639 0 : (struct bdev_aio_task *)bdev_io->driver_ctx);
640 0 : return 0;
641 :
642 : case SPDK_BDEV_IO_TYPE_RESET:
643 0 : bdev_aio_reset((struct file_disk *)bdev_io->bdev->ctxt,
644 0 : (struct bdev_aio_task *)bdev_io->driver_ctx);
645 0 : return 0;
646 :
647 : #ifndef __FreeBSD__
648 : case SPDK_BDEV_IO_TYPE_UNMAP:
649 0 : bdev_aio_unmap(bdev_io);
650 0 : return 0;
651 :
652 : case SPDK_BDEV_IO_TYPE_WRITE_ZEROES:
653 0 : bdev_aio_write_zeros(bdev_io);
654 0 : return 0;
655 : #endif
656 :
657 : default:
658 0 : return -1;
659 : }
660 0 : }
661 :
662 : static void
663 0 : bdev_aio_submit_request(struct spdk_io_channel *ch, struct spdk_bdev_io *bdev_io)
664 : {
665 0 : if (_bdev_aio_submit_request(ch, bdev_io) < 0) {
666 0 : spdk_bdev_io_complete(bdev_io, SPDK_BDEV_IO_STATUS_FAILED);
667 0 : }
668 0 : }
669 :
670 : static bool
671 0 : bdev_aio_io_type_supported(void *ctx, enum spdk_bdev_io_type io_type)
672 : {
673 0 : struct file_disk *fdisk = ctx;
674 :
675 0 : switch (io_type) {
676 : case SPDK_BDEV_IO_TYPE_READ:
677 : case SPDK_BDEV_IO_TYPE_WRITE:
678 : case SPDK_BDEV_IO_TYPE_FLUSH:
679 : case SPDK_BDEV_IO_TYPE_RESET:
680 0 : return true;
681 :
682 : case SPDK_BDEV_IO_TYPE_UNMAP:
683 : case SPDK_BDEV_IO_TYPE_WRITE_ZEROES:
684 0 : return fdisk->fallocate;
685 :
686 : default:
687 0 : return false;
688 : }
689 0 : }
690 :
691 : #ifdef __FreeBSD__
692 : static int
693 : bdev_aio_create_io(struct bdev_aio_io_channel *ch)
694 : {
695 : ch->kqfd = kqueue();
696 : if (ch->kqfd < 0) {
697 : SPDK_ERRLOG("async I/O context setup failure: %s.\n", spdk_strerror(errno));
698 : return -1;
699 : }
700 :
701 : return 0;
702 : }
703 :
704 : static void
705 : bdev_aio_destroy_io(struct bdev_aio_io_channel *ch)
706 : {
707 : close(ch->kqfd);
708 : }
709 : #else
710 : static int
711 0 : bdev_aio_create_io(struct bdev_aio_io_channel *ch)
712 : {
713 0 : if (io_setup(SPDK_AIO_QUEUE_DEPTH, &ch->io_ctx) < 0) {
714 0 : SPDK_ERRLOG("Async I/O context setup failure, likely due to exceeding kernel limit.\n");
715 0 : SPDK_ERRLOG("This limit may be increased using 'sysctl -w fs.aio-max-nr'.\n");
716 0 : return -1;
717 : }
718 :
719 0 : return 0;
720 0 : }
721 :
722 : static void
723 0 : bdev_aio_destroy_io(struct bdev_aio_io_channel *ch)
724 : {
725 0 : io_destroy(ch->io_ctx);
726 0 : }
727 : #endif
728 :
729 : static int
730 0 : bdev_aio_create_cb(void *io_device, void *ctx_buf)
731 : {
732 0 : struct bdev_aio_io_channel *ch = ctx_buf;
733 0 : int rc;
734 :
735 0 : rc = bdev_aio_create_io(ch);
736 0 : if (rc < 0) {
737 0 : return rc;
738 : }
739 :
740 0 : ch->group_ch = spdk_io_channel_get_ctx(spdk_get_io_channel(&aio_if));
741 0 : TAILQ_INSERT_TAIL(&ch->group_ch->io_ch_head, ch, link);
742 :
743 0 : return 0;
744 0 : }
745 :
746 : static void
747 0 : bdev_aio_destroy_cb(void *io_device, void *ctx_buf)
748 : {
749 0 : struct bdev_aio_io_channel *ch = ctx_buf;
750 :
751 0 : bdev_aio_destroy_io(ch);
752 :
753 0 : assert(ch->group_ch);
754 0 : TAILQ_REMOVE(&ch->group_ch->io_ch_head, ch, link);
755 :
756 0 : spdk_put_io_channel(spdk_io_channel_from_ctx(ch->group_ch));
757 0 : }
758 :
759 : static struct spdk_io_channel *
760 0 : bdev_aio_get_io_channel(void *ctx)
761 : {
762 0 : struct file_disk *fdisk = ctx;
763 :
764 0 : return spdk_get_io_channel(fdisk);
765 0 : }
766 :
767 :
768 : static int
769 0 : bdev_aio_dump_info_json(void *ctx, struct spdk_json_write_ctx *w)
770 : {
771 0 : struct file_disk *fdisk = ctx;
772 :
773 0 : spdk_json_write_named_object_begin(w, "aio");
774 :
775 0 : spdk_json_write_named_string(w, "filename", fdisk->filename);
776 :
777 0 : spdk_json_write_named_bool(w, "block_size_override", fdisk->block_size_override);
778 :
779 0 : spdk_json_write_named_bool(w, "readonly", fdisk->readonly);
780 :
781 0 : spdk_json_write_named_bool(w, "fallocate", fdisk->fallocate);
782 :
783 0 : spdk_json_write_object_end(w);
784 :
785 0 : return 0;
786 0 : }
787 :
788 : static void
789 0 : bdev_aio_write_json_config(struct spdk_bdev *bdev, struct spdk_json_write_ctx *w)
790 : {
791 0 : struct file_disk *fdisk = bdev->ctxt;
792 0 : const struct spdk_uuid *uuid = spdk_bdev_get_uuid(bdev);
793 :
794 0 : spdk_json_write_object_begin(w);
795 :
796 0 : spdk_json_write_named_string(w, "method", "bdev_aio_create");
797 :
798 0 : spdk_json_write_named_object_begin(w, "params");
799 0 : spdk_json_write_named_string(w, "name", bdev->name);
800 0 : if (fdisk->block_size_override) {
801 0 : spdk_json_write_named_uint32(w, "block_size", bdev->blocklen);
802 0 : }
803 0 : spdk_json_write_named_string(w, "filename", fdisk->filename);
804 0 : spdk_json_write_named_bool(w, "readonly", fdisk->readonly);
805 0 : spdk_json_write_named_bool(w, "fallocate", fdisk->fallocate);
806 0 : if (!spdk_uuid_is_null(uuid)) {
807 0 : spdk_json_write_named_uuid(w, "uuid", uuid);
808 0 : }
809 0 : spdk_json_write_object_end(w);
810 :
811 0 : spdk_json_write_object_end(w);
812 0 : }
813 :
814 : static const struct spdk_bdev_fn_table aio_fn_table = {
815 : .destruct = bdev_aio_destruct,
816 : .submit_request = bdev_aio_submit_request,
817 : .io_type_supported = bdev_aio_io_type_supported,
818 : .get_io_channel = bdev_aio_get_io_channel,
819 : .dump_info_json = bdev_aio_dump_info_json,
820 : .write_config_json = bdev_aio_write_json_config,
821 : };
822 :
823 : static void
824 0 : aio_free_disk(struct file_disk *fdisk)
825 : {
826 0 : if (fdisk == NULL) {
827 0 : return;
828 : }
829 0 : free(fdisk->filename);
830 0 : free(fdisk->disk.name);
831 0 : free(fdisk);
832 0 : }
833 :
834 : static int
835 0 : bdev_aio_register_interrupt(struct bdev_aio_group_channel *ch)
836 : {
837 0 : int efd;
838 :
839 0 : efd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
840 0 : if (efd < 0) {
841 0 : return -1;
842 : }
843 :
844 0 : ch->intr = SPDK_INTERRUPT_REGISTER(efd, bdev_aio_group_interrupt, ch);
845 0 : if (ch->intr == NULL) {
846 0 : close(efd);
847 0 : return -1;
848 : }
849 0 : ch->efd = efd;
850 :
851 0 : return 0;
852 0 : }
853 :
854 : static void
855 0 : bdev_aio_unregister_interrupt(struct bdev_aio_group_channel *ch)
856 : {
857 0 : spdk_interrupt_unregister(&ch->intr);
858 0 : close(ch->efd);
859 0 : ch->efd = -1;
860 0 : }
861 :
862 : static int
863 0 : bdev_aio_group_create_cb(void *io_device, void *ctx_buf)
864 : {
865 0 : struct bdev_aio_group_channel *ch = ctx_buf;
866 0 : int rc;
867 :
868 0 : TAILQ_INIT(&ch->io_ch_head);
869 : /* Initialize ch->efd to be invalid and unused. */
870 0 : ch->efd = -1;
871 0 : if (spdk_interrupt_mode_is_enabled()) {
872 0 : rc = bdev_aio_register_interrupt(ch);
873 0 : if (rc < 0) {
874 0 : SPDK_ERRLOG("Failed to prepare intr resource to bdev_aio\n");
875 0 : return rc;
876 : }
877 0 : }
878 :
879 0 : ch->poller = SPDK_POLLER_REGISTER(bdev_aio_group_poll, ch, 0);
880 0 : spdk_poller_register_interrupt(ch->poller, NULL, NULL);
881 :
882 0 : return 0;
883 0 : }
884 :
885 : static void
886 0 : bdev_aio_group_destroy_cb(void *io_device, void *ctx_buf)
887 : {
888 0 : struct bdev_aio_group_channel *ch = ctx_buf;
889 :
890 0 : if (!TAILQ_EMPTY(&ch->io_ch_head)) {
891 0 : SPDK_ERRLOG("Group channel of bdev aio has uncleared io channel\n");
892 0 : }
893 :
894 0 : spdk_poller_unregister(&ch->poller);
895 0 : if (spdk_interrupt_mode_is_enabled()) {
896 0 : bdev_aio_unregister_interrupt(ch);
897 0 : }
898 0 : }
899 :
900 : int
901 0 : create_aio_bdev(const char *name, const char *filename, uint32_t block_size, bool readonly,
902 : bool fallocate, const struct spdk_uuid *uuid)
903 : {
904 0 : struct file_disk *fdisk;
905 0 : uint32_t detected_block_size;
906 0 : uint64_t disk_size;
907 0 : int rc;
908 :
909 : #ifdef __FreeBSD__
910 : if (fallocate) {
911 : SPDK_ERRLOG("Unable to support fallocate on this platform\n");
912 : return -ENOTSUP;
913 : }
914 : #endif
915 :
916 0 : fdisk = calloc(1, sizeof(*fdisk));
917 0 : if (!fdisk) {
918 0 : SPDK_ERRLOG("Unable to allocate enough memory for aio backend\n");
919 0 : return -ENOMEM;
920 : }
921 0 : fdisk->readonly = readonly;
922 0 : fdisk->fallocate = fallocate;
923 :
924 0 : fdisk->filename = strdup(filename);
925 0 : if (!fdisk->filename) {
926 0 : rc = -ENOMEM;
927 0 : goto error_return;
928 : }
929 :
930 0 : if (bdev_aio_open(fdisk)) {
931 0 : SPDK_ERRLOG("Unable to open file %s. fd: %d errno: %d\n", filename, fdisk->fd, errno);
932 0 : rc = -errno;
933 0 : goto error_return;
934 : }
935 :
936 0 : disk_size = spdk_fd_get_size(fdisk->fd);
937 :
938 0 : fdisk->disk.name = strdup(name);
939 0 : if (!fdisk->disk.name) {
940 0 : rc = -ENOMEM;
941 0 : goto error_return;
942 : }
943 0 : fdisk->disk.product_name = "AIO disk";
944 0 : fdisk->disk.module = &aio_if;
945 :
946 0 : fdisk->disk.write_cache = 1;
947 :
948 0 : detected_block_size = spdk_fd_get_blocklen(fdisk->fd);
949 0 : if (block_size == 0) {
950 : /* User did not specify block size - use autodetected block size. */
951 0 : if (detected_block_size == 0) {
952 0 : SPDK_ERRLOG("Block size could not be auto-detected\n");
953 0 : rc = -EINVAL;
954 0 : goto error_return;
955 : }
956 0 : fdisk->block_size_override = false;
957 0 : block_size = detected_block_size;
958 0 : } else {
959 0 : if (block_size < detected_block_size) {
960 0 : SPDK_ERRLOG("Specified block size %" PRIu32 " is smaller than "
961 : "auto-detected block size %" PRIu32 "\n",
962 : block_size, detected_block_size);
963 0 : rc = -EINVAL;
964 0 : goto error_return;
965 0 : } else if (detected_block_size != 0 && block_size != detected_block_size) {
966 0 : SPDK_WARNLOG("Specified block size %" PRIu32 " does not match "
967 : "auto-detected block size %" PRIu32 "\n",
968 : block_size, detected_block_size);
969 0 : }
970 0 : fdisk->block_size_override = true;
971 : }
972 :
973 0 : if (block_size < 512) {
974 0 : SPDK_ERRLOG("Invalid block size %" PRIu32 " (must be at least 512).\n", block_size);
975 0 : rc = -EINVAL;
976 0 : goto error_return;
977 : }
978 :
979 0 : if (!spdk_u32_is_pow2(block_size)) {
980 0 : SPDK_ERRLOG("Invalid block size %" PRIu32 " (must be a power of 2.)\n", block_size);
981 0 : rc = -EINVAL;
982 0 : goto error_return;
983 : }
984 :
985 0 : fdisk->disk.blocklen = block_size;
986 0 : if (fdisk->block_size_override && detected_block_size) {
987 0 : fdisk->disk.required_alignment = spdk_u32log2(detected_block_size);
988 0 : } else {
989 0 : fdisk->disk.required_alignment = spdk_u32log2(block_size);
990 : }
991 :
992 0 : if (disk_size % fdisk->disk.blocklen != 0) {
993 0 : SPDK_ERRLOG("Disk size %" PRIu64 " is not a multiple of block size %" PRIu32 "\n",
994 : disk_size, fdisk->disk.blocklen);
995 0 : rc = -EINVAL;
996 0 : goto error_return;
997 : }
998 :
999 0 : fdisk->disk.blockcnt = disk_size / fdisk->disk.blocklen;
1000 0 : fdisk->disk.ctxt = fdisk;
1001 0 : spdk_uuid_copy(&fdisk->disk.uuid, uuid);
1002 :
1003 0 : fdisk->disk.fn_table = &aio_fn_table;
1004 :
1005 0 : spdk_io_device_register(fdisk, bdev_aio_create_cb, bdev_aio_destroy_cb,
1006 : sizeof(struct bdev_aio_io_channel),
1007 0 : fdisk->disk.name);
1008 0 : rc = spdk_bdev_register(&fdisk->disk);
1009 0 : if (rc) {
1010 0 : spdk_io_device_unregister(fdisk, NULL);
1011 0 : goto error_return;
1012 : }
1013 :
1014 0 : TAILQ_INSERT_TAIL(&g_aio_disk_head, fdisk, link);
1015 0 : return 0;
1016 :
1017 : error_return:
1018 0 : bdev_aio_close(fdisk);
1019 0 : aio_free_disk(fdisk);
1020 0 : return rc;
1021 0 : }
1022 :
1023 : static void
1024 0 : dummy_bdev_event_cb(enum spdk_bdev_event_type type, struct spdk_bdev *bdev, void *ctx)
1025 : {
1026 0 : }
1027 :
1028 : int
1029 0 : bdev_aio_rescan(const char *name)
1030 : {
1031 0 : struct spdk_bdev_desc *desc;
1032 0 : struct spdk_bdev *bdev;
1033 0 : struct file_disk *fdisk;
1034 0 : uint64_t disk_size, blockcnt;
1035 0 : int rc;
1036 :
1037 0 : rc = spdk_bdev_open_ext(name, false, dummy_bdev_event_cb, NULL, &desc);
1038 0 : if (rc != 0) {
1039 0 : return rc;
1040 : }
1041 :
1042 0 : bdev = spdk_bdev_desc_get_bdev(desc);
1043 0 : if (bdev->module != &aio_if) {
1044 0 : rc = -ENODEV;
1045 0 : goto exit;
1046 : }
1047 :
1048 0 : fdisk = SPDK_CONTAINEROF(bdev, struct file_disk, disk);
1049 0 : disk_size = spdk_fd_get_size(fdisk->fd);
1050 0 : blockcnt = disk_size / bdev->blocklen;
1051 :
1052 0 : if (bdev->blockcnt != blockcnt) {
1053 0 : SPDK_NOTICELOG("AIO device is resized: bdev name %s, old block count %" PRIu64 ", new block count %"
1054 : PRIu64 "\n",
1055 : fdisk->filename,
1056 : bdev->blockcnt,
1057 : blockcnt);
1058 0 : rc = spdk_bdev_notify_blockcnt_change(bdev, blockcnt);
1059 0 : if (rc != 0) {
1060 0 : SPDK_ERRLOG("Could not change num blocks for aio bdev: name %s, errno: %d.\n",
1061 : fdisk->filename, rc);
1062 0 : goto exit;
1063 : }
1064 0 : }
1065 :
1066 : exit:
1067 0 : spdk_bdev_close(desc);
1068 0 : return rc;
1069 0 : }
1070 :
1071 : struct delete_aio_bdev_ctx {
1072 : delete_aio_bdev_complete cb_fn;
1073 : void *cb_arg;
1074 : };
1075 :
1076 : static void
1077 0 : aio_bdev_unregister_cb(void *arg, int bdeverrno)
1078 : {
1079 0 : struct delete_aio_bdev_ctx *ctx = arg;
1080 :
1081 0 : ctx->cb_fn(ctx->cb_arg, bdeverrno);
1082 0 : free(ctx);
1083 0 : }
1084 :
1085 : void
1086 0 : bdev_aio_delete(const char *name, delete_aio_bdev_complete cb_fn, void *cb_arg)
1087 : {
1088 0 : struct delete_aio_bdev_ctx *ctx;
1089 0 : int rc;
1090 :
1091 0 : ctx = calloc(1, sizeof(*ctx));
1092 0 : if (ctx == NULL) {
1093 0 : cb_fn(cb_arg, -ENOMEM);
1094 0 : return;
1095 : }
1096 :
1097 0 : ctx->cb_fn = cb_fn;
1098 0 : ctx->cb_arg = cb_arg;
1099 0 : rc = spdk_bdev_unregister_by_name(name, &aio_if, aio_bdev_unregister_cb, ctx);
1100 0 : if (rc != 0) {
1101 0 : aio_bdev_unregister_cb(ctx, rc);
1102 0 : }
1103 0 : }
1104 :
1105 : static int
1106 0 : bdev_aio_initialize(void)
1107 : {
1108 0 : spdk_io_device_register(&aio_if, bdev_aio_group_create_cb, bdev_aio_group_destroy_cb,
1109 : sizeof(struct bdev_aio_group_channel), "aio_module");
1110 :
1111 0 : return 0;
1112 : }
1113 :
1114 : static void
1115 0 : bdev_aio_fini(void)
1116 : {
1117 0 : spdk_io_device_unregister(&aio_if, NULL);
1118 0 : }
1119 :
1120 0 : SPDK_LOG_REGISTER_COMPONENT(aio)
|
