]>
Commit | Line | Data |
---|---|---|
fd66dbd4 SH |
1 | /* SPDX-License-Identifier: LGPL-2.1-or-later */ |
2 | /* | |
3 | * libblkio BlockDriver | |
4 | * | |
5 | * Copyright Red Hat, Inc. | |
6 | * | |
7 | * Author: | |
8 | * Stefan Hajnoczi <stefanha@redhat.com> | |
9 | */ | |
10 | ||
11 | #include "qemu/osdep.h" | |
12 | #include <blkio.h> | |
13 | #include "block/block_int.h" | |
c5640b3e SH |
14 | #include "exec/memory.h" |
15 | #include "exec/cpu-common.h" /* for qemu_ram_get_fd() */ | |
fd66dbd4 | 16 | #include "qapi/error.h" |
c5640b3e | 17 | #include "qemu/error-report.h" |
fd66dbd4 SH |
18 | #include "qapi/qmp/qdict.h" |
19 | #include "qemu/module.h" | |
c5640b3e | 20 | #include "exec/memory.h" /* for ram_block_discard_disable() */ |
fd66dbd4 SH |
21 | |
22 | /* | |
23 | * Keep the QEMU BlockDriver names identical to the libblkio driver names. | |
24 | * Using macros instead of typing out the string literals avoids typos. | |
25 | */ | |
26 | #define DRIVER_IO_URING "io_uring" | |
27 | #define DRIVER_NVME_IO_URING "nvme-io_uring" | |
03d9e4c0 | 28 | #define DRIVER_VIRTIO_BLK_VFIO_PCI "virtio-blk-vfio-pci" |
fd66dbd4 SH |
29 | #define DRIVER_VIRTIO_BLK_VHOST_USER "virtio-blk-vhost-user" |
30 | #define DRIVER_VIRTIO_BLK_VHOST_VDPA "virtio-blk-vhost-vdpa" | |
31 | ||
32 | /* | |
33 | * Allocated bounce buffers are kept in a list sorted by buffer address. | |
34 | */ | |
35 | typedef struct BlkioBounceBuf { | |
36 | QLIST_ENTRY(BlkioBounceBuf) next; | |
37 | ||
38 | /* The bounce buffer */ | |
39 | struct iovec buf; | |
40 | } BlkioBounceBuf; | |
41 | ||
42 | typedef struct { | |
43 | /* | |
44 | * libblkio is not thread-safe so this lock protects ->blkio and | |
45 | * ->blkioq. | |
46 | */ | |
47 | QemuMutex blkio_lock; | |
48 | struct blkio *blkio; | |
49 | struct blkioq *blkioq; /* make this multi-queue in the future... */ | |
50 | int completion_fd; | |
51 | ||
52 | /* | |
53 | * Polling fetches the next completion into this field. | |
54 | * | |
55 | * No lock is necessary since only one thread calls aio_poll() and invokes | |
56 | * fd and poll handlers. | |
57 | */ | |
58 | struct blkio_completion poll_completion; | |
59 | ||
60 | /* | |
61 | * Protects ->bounce_pool, ->bounce_bufs, ->bounce_available. | |
62 | * | |
63 | * Lock ordering: ->bounce_lock before ->blkio_lock. | |
64 | */ | |
65 | CoMutex bounce_lock; | |
66 | ||
67 | /* Bounce buffer pool */ | |
68 | struct blkio_mem_region bounce_pool; | |
69 | ||
70 | /* Sorted list of allocated bounce buffers */ | |
71 | QLIST_HEAD(, BlkioBounceBuf) bounce_bufs; | |
72 | ||
73 | /* Queue for coroutines waiting for bounce buffer space */ | |
74 | CoQueue bounce_available; | |
75 | ||
76 | /* The value of the "mem-region-alignment" property */ | |
77 | size_t mem_region_alignment; | |
78 | ||
79 | /* Can we skip adding/deleting blkio_mem_regions? */ | |
80 | bool needs_mem_regions; | |
c5640b3e SH |
81 | |
82 | /* Are file descriptors necessary for blkio_mem_regions? */ | |
83 | bool needs_mem_region_fd; | |
84 | ||
85 | /* Are madvise(MADV_DONTNEED)-style operations unavailable? */ | |
86 | bool may_pin_mem_regions; | |
fd66dbd4 SH |
87 | } BDRVBlkioState; |
88 | ||
89 | /* Called with s->bounce_lock held */ | |
90 | static int blkio_resize_bounce_pool(BDRVBlkioState *s, int64_t bytes) | |
91 | { | |
92 | /* There can be no allocated bounce buffers during resize */ | |
93 | assert(QLIST_EMPTY(&s->bounce_bufs)); | |
94 | ||
95 | /* Pad size to reduce frequency of resize calls */ | |
96 | bytes += 128 * 1024; | |
97 | ||
98 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
99 | int ret; | |
100 | ||
101 | if (s->bounce_pool.addr) { | |
102 | blkio_unmap_mem_region(s->blkio, &s->bounce_pool); | |
103 | blkio_free_mem_region(s->blkio, &s->bounce_pool); | |
104 | memset(&s->bounce_pool, 0, sizeof(s->bounce_pool)); | |
105 | } | |
106 | ||
107 | /* Automatically freed when s->blkio is destroyed */ | |
108 | ret = blkio_alloc_mem_region(s->blkio, &s->bounce_pool, bytes); | |
109 | if (ret < 0) { | |
110 | return ret; | |
111 | } | |
112 | ||
113 | ret = blkio_map_mem_region(s->blkio, &s->bounce_pool); | |
114 | if (ret < 0) { | |
115 | blkio_free_mem_region(s->blkio, &s->bounce_pool); | |
116 | memset(&s->bounce_pool, 0, sizeof(s->bounce_pool)); | |
117 | return ret; | |
118 | } | |
119 | } | |
120 | ||
121 | return 0; | |
122 | } | |
123 | ||
124 | /* Called with s->bounce_lock held */ | |
125 | static bool | |
126 | blkio_do_alloc_bounce_buffer(BDRVBlkioState *s, BlkioBounceBuf *bounce, | |
127 | int64_t bytes) | |
128 | { | |
129 | void *addr = s->bounce_pool.addr; | |
130 | BlkioBounceBuf *cur = NULL; | |
131 | BlkioBounceBuf *prev = NULL; | |
132 | ptrdiff_t space; | |
133 | ||
134 | /* | |
135 | * This is just a linear search over the holes between requests. An | |
136 | * efficient allocator would be nice. | |
137 | */ | |
138 | QLIST_FOREACH(cur, &s->bounce_bufs, next) { | |
139 | space = cur->buf.iov_base - addr; | |
140 | if (bytes <= space) { | |
141 | QLIST_INSERT_BEFORE(cur, bounce, next); | |
142 | bounce->buf.iov_base = addr; | |
143 | bounce->buf.iov_len = bytes; | |
144 | return true; | |
145 | } | |
146 | ||
147 | addr = cur->buf.iov_base + cur->buf.iov_len; | |
148 | prev = cur; | |
149 | } | |
150 | ||
151 | /* Is there space after the last request? */ | |
152 | space = s->bounce_pool.addr + s->bounce_pool.len - addr; | |
153 | if (bytes > space) { | |
154 | return false; | |
155 | } | |
156 | if (prev) { | |
157 | QLIST_INSERT_AFTER(prev, bounce, next); | |
158 | } else { | |
159 | QLIST_INSERT_HEAD(&s->bounce_bufs, bounce, next); | |
160 | } | |
161 | bounce->buf.iov_base = addr; | |
162 | bounce->buf.iov_len = bytes; | |
163 | return true; | |
164 | } | |
165 | ||
166 | static int coroutine_fn | |
167 | blkio_alloc_bounce_buffer(BDRVBlkioState *s, BlkioBounceBuf *bounce, | |
168 | int64_t bytes) | |
169 | { | |
170 | /* | |
171 | * Ensure fairness: first time around we join the back of the queue, | |
172 | * subsequently we join the front so we don't lose our place. | |
173 | */ | |
174 | CoQueueWaitFlags wait_flags = 0; | |
175 | ||
176 | QEMU_LOCK_GUARD(&s->bounce_lock); | |
177 | ||
178 | /* Ensure fairness: don't even try if other requests are already waiting */ | |
179 | if (!qemu_co_queue_empty(&s->bounce_available)) { | |
180 | qemu_co_queue_wait_flags(&s->bounce_available, &s->bounce_lock, | |
181 | wait_flags); | |
182 | wait_flags = CO_QUEUE_WAIT_FRONT; | |
183 | } | |
184 | ||
185 | while (true) { | |
186 | if (blkio_do_alloc_bounce_buffer(s, bounce, bytes)) { | |
187 | /* Kick the next queued request since there may be space */ | |
188 | qemu_co_queue_next(&s->bounce_available); | |
189 | return 0; | |
190 | } | |
191 | ||
192 | /* | |
193 | * If there are no in-flight requests then the pool was simply too | |
194 | * small. | |
195 | */ | |
196 | if (QLIST_EMPTY(&s->bounce_bufs)) { | |
197 | bool ok; | |
198 | int ret; | |
199 | ||
200 | ret = blkio_resize_bounce_pool(s, bytes); | |
201 | if (ret < 0) { | |
202 | /* Kick the next queued request since that may fail too */ | |
203 | qemu_co_queue_next(&s->bounce_available); | |
204 | return ret; | |
205 | } | |
206 | ||
207 | ok = blkio_do_alloc_bounce_buffer(s, bounce, bytes); | |
208 | assert(ok); /* must have space this time */ | |
209 | return 0; | |
210 | } | |
211 | ||
212 | qemu_co_queue_wait_flags(&s->bounce_available, &s->bounce_lock, | |
213 | wait_flags); | |
214 | wait_flags = CO_QUEUE_WAIT_FRONT; | |
215 | } | |
216 | } | |
217 | ||
218 | static void coroutine_fn blkio_free_bounce_buffer(BDRVBlkioState *s, | |
219 | BlkioBounceBuf *bounce) | |
220 | { | |
221 | QEMU_LOCK_GUARD(&s->bounce_lock); | |
222 | ||
223 | QLIST_REMOVE(bounce, next); | |
224 | ||
225 | /* Wake up waiting coroutines since space may now be available */ | |
226 | qemu_co_queue_next(&s->bounce_available); | |
227 | } | |
228 | ||
229 | /* For async to .bdrv_co_*() conversion */ | |
230 | typedef struct { | |
231 | Coroutine *coroutine; | |
232 | int ret; | |
233 | } BlkioCoData; | |
234 | ||
235 | static void blkio_completion_fd_read(void *opaque) | |
236 | { | |
237 | BlockDriverState *bs = opaque; | |
238 | BDRVBlkioState *s = bs->opaque; | |
239 | uint64_t val; | |
240 | int ret; | |
241 | ||
242 | /* Polling may have already fetched a completion */ | |
243 | if (s->poll_completion.user_data != NULL) { | |
244 | BlkioCoData *cod = s->poll_completion.user_data; | |
245 | cod->ret = s->poll_completion.ret; | |
246 | ||
247 | /* Clear it in case aio_co_wake() enters a nested event loop */ | |
248 | s->poll_completion.user_data = NULL; | |
249 | ||
250 | aio_co_wake(cod->coroutine); | |
251 | } | |
252 | ||
253 | /* Reset completion fd status */ | |
254 | ret = read(s->completion_fd, &val, sizeof(val)); | |
255 | ||
256 | /* Ignore errors, there's nothing we can do */ | |
257 | (void)ret; | |
258 | ||
259 | /* | |
260 | * Reading one completion at a time makes nested event loop re-entrancy | |
261 | * simple. Change this loop to get multiple completions in one go if it | |
262 | * becomes a performance bottleneck. | |
263 | */ | |
264 | while (true) { | |
265 | struct blkio_completion completion; | |
266 | ||
267 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
268 | ret = blkioq_do_io(s->blkioq, &completion, 0, 1, NULL); | |
269 | } | |
270 | if (ret != 1) { | |
271 | break; | |
272 | } | |
273 | ||
274 | BlkioCoData *cod = completion.user_data; | |
275 | cod->ret = completion.ret; | |
276 | aio_co_wake(cod->coroutine); | |
277 | } | |
278 | } | |
279 | ||
280 | static bool blkio_completion_fd_poll(void *opaque) | |
281 | { | |
282 | BlockDriverState *bs = opaque; | |
283 | BDRVBlkioState *s = bs->opaque; | |
284 | int ret; | |
285 | ||
286 | /* Just in case we already fetched a completion */ | |
287 | if (s->poll_completion.user_data != NULL) { | |
288 | return true; | |
289 | } | |
290 | ||
291 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
292 | ret = blkioq_do_io(s->blkioq, &s->poll_completion, 0, 1, NULL); | |
293 | } | |
294 | return ret == 1; | |
295 | } | |
296 | ||
297 | static void blkio_completion_fd_poll_ready(void *opaque) | |
298 | { | |
299 | blkio_completion_fd_read(opaque); | |
300 | } | |
301 | ||
302 | static void blkio_attach_aio_context(BlockDriverState *bs, | |
303 | AioContext *new_context) | |
304 | { | |
305 | BDRVBlkioState *s = bs->opaque; | |
306 | ||
307 | aio_set_fd_handler(new_context, | |
308 | s->completion_fd, | |
309 | false, | |
310 | blkio_completion_fd_read, | |
311 | NULL, | |
312 | blkio_completion_fd_poll, | |
313 | blkio_completion_fd_poll_ready, | |
314 | bs); | |
315 | } | |
316 | ||
317 | static void blkio_detach_aio_context(BlockDriverState *bs) | |
318 | { | |
319 | BDRVBlkioState *s = bs->opaque; | |
320 | ||
321 | aio_set_fd_handler(bdrv_get_aio_context(bs), | |
322 | s->completion_fd, | |
323 | false, NULL, NULL, NULL, NULL, NULL); | |
324 | } | |
325 | ||
326 | /* Call with s->blkio_lock held to submit I/O after enqueuing a new request */ | |
327 | static void blkio_submit_io(BlockDriverState *bs) | |
328 | { | |
329 | if (qatomic_read(&bs->io_plugged) == 0) { | |
330 | BDRVBlkioState *s = bs->opaque; | |
331 | ||
332 | blkioq_do_io(s->blkioq, NULL, 0, 0, NULL); | |
333 | } | |
334 | } | |
335 | ||
336 | static int coroutine_fn | |
337 | blkio_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes) | |
338 | { | |
339 | BDRVBlkioState *s = bs->opaque; | |
340 | BlkioCoData cod = { | |
341 | .coroutine = qemu_coroutine_self(), | |
342 | }; | |
343 | ||
344 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
345 | blkioq_discard(s->blkioq, offset, bytes, &cod, 0); | |
346 | blkio_submit_io(bs); | |
347 | } | |
348 | ||
349 | qemu_coroutine_yield(); | |
350 | return cod.ret; | |
351 | } | |
352 | ||
353 | static int coroutine_fn | |
354 | blkio_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, | |
355 | QEMUIOVector *qiov, BdrvRequestFlags flags) | |
356 | { | |
357 | BlkioCoData cod = { | |
358 | .coroutine = qemu_coroutine_self(), | |
359 | }; | |
360 | BDRVBlkioState *s = bs->opaque; | |
c5640b3e SH |
361 | bool use_bounce_buffer = |
362 | s->needs_mem_regions && !(flags & BDRV_REQ_REGISTERED_BUF); | |
fd66dbd4 SH |
363 | BlkioBounceBuf bounce; |
364 | struct iovec *iov = qiov->iov; | |
365 | int iovcnt = qiov->niov; | |
366 | ||
367 | if (use_bounce_buffer) { | |
368 | int ret = blkio_alloc_bounce_buffer(s, &bounce, bytes); | |
369 | if (ret < 0) { | |
370 | return ret; | |
371 | } | |
372 | ||
373 | iov = &bounce.buf; | |
374 | iovcnt = 1; | |
375 | } | |
376 | ||
377 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
378 | blkioq_readv(s->blkioq, offset, iov, iovcnt, &cod, 0); | |
379 | blkio_submit_io(bs); | |
380 | } | |
381 | ||
382 | qemu_coroutine_yield(); | |
383 | ||
384 | if (use_bounce_buffer) { | |
385 | if (cod.ret == 0) { | |
386 | qemu_iovec_from_buf(qiov, 0, | |
387 | bounce.buf.iov_base, | |
388 | bounce.buf.iov_len); | |
389 | } | |
390 | ||
391 | blkio_free_bounce_buffer(s, &bounce); | |
392 | } | |
393 | ||
394 | return cod.ret; | |
395 | } | |
396 | ||
397 | static int coroutine_fn blkio_co_pwritev(BlockDriverState *bs, int64_t offset, | |
398 | int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags) | |
399 | { | |
400 | uint32_t blkio_flags = (flags & BDRV_REQ_FUA) ? BLKIO_REQ_FUA : 0; | |
401 | BlkioCoData cod = { | |
402 | .coroutine = qemu_coroutine_self(), | |
403 | }; | |
404 | BDRVBlkioState *s = bs->opaque; | |
c5640b3e SH |
405 | bool use_bounce_buffer = |
406 | s->needs_mem_regions && !(flags & BDRV_REQ_REGISTERED_BUF); | |
fd66dbd4 SH |
407 | BlkioBounceBuf bounce; |
408 | struct iovec *iov = qiov->iov; | |
409 | int iovcnt = qiov->niov; | |
410 | ||
411 | if (use_bounce_buffer) { | |
412 | int ret = blkio_alloc_bounce_buffer(s, &bounce, bytes); | |
413 | if (ret < 0) { | |
414 | return ret; | |
415 | } | |
416 | ||
417 | qemu_iovec_to_buf(qiov, 0, bounce.buf.iov_base, bytes); | |
418 | iov = &bounce.buf; | |
419 | iovcnt = 1; | |
420 | } | |
421 | ||
422 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
423 | blkioq_writev(s->blkioq, offset, iov, iovcnt, &cod, blkio_flags); | |
424 | blkio_submit_io(bs); | |
425 | } | |
426 | ||
427 | qemu_coroutine_yield(); | |
428 | ||
429 | if (use_bounce_buffer) { | |
430 | blkio_free_bounce_buffer(s, &bounce); | |
431 | } | |
432 | ||
433 | return cod.ret; | |
434 | } | |
435 | ||
436 | static int coroutine_fn blkio_co_flush(BlockDriverState *bs) | |
437 | { | |
438 | BDRVBlkioState *s = bs->opaque; | |
439 | BlkioCoData cod = { | |
440 | .coroutine = qemu_coroutine_self(), | |
441 | }; | |
442 | ||
443 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
444 | blkioq_flush(s->blkioq, &cod, 0); | |
445 | blkio_submit_io(bs); | |
446 | } | |
447 | ||
448 | qemu_coroutine_yield(); | |
449 | return cod.ret; | |
450 | } | |
451 | ||
452 | static int coroutine_fn blkio_co_pwrite_zeroes(BlockDriverState *bs, | |
453 | int64_t offset, int64_t bytes, BdrvRequestFlags flags) | |
454 | { | |
455 | BDRVBlkioState *s = bs->opaque; | |
456 | BlkioCoData cod = { | |
457 | .coroutine = qemu_coroutine_self(), | |
458 | }; | |
459 | uint32_t blkio_flags = 0; | |
460 | ||
461 | if (flags & BDRV_REQ_FUA) { | |
462 | blkio_flags |= BLKIO_REQ_FUA; | |
463 | } | |
464 | if (!(flags & BDRV_REQ_MAY_UNMAP)) { | |
465 | blkio_flags |= BLKIO_REQ_NO_UNMAP; | |
466 | } | |
467 | if (flags & BDRV_REQ_NO_FALLBACK) { | |
468 | blkio_flags |= BLKIO_REQ_NO_FALLBACK; | |
469 | } | |
470 | ||
471 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
472 | blkioq_write_zeroes(s->blkioq, offset, bytes, &cod, blkio_flags); | |
473 | blkio_submit_io(bs); | |
474 | } | |
475 | ||
476 | qemu_coroutine_yield(); | |
477 | return cod.ret; | |
478 | } | |
479 | ||
480 | static void blkio_io_unplug(BlockDriverState *bs) | |
481 | { | |
482 | BDRVBlkioState *s = bs->opaque; | |
483 | ||
484 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
485 | blkio_submit_io(bs); | |
486 | } | |
487 | } | |
488 | ||
c5640b3e SH |
489 | typedef enum { |
490 | BMRR_OK, | |
491 | BMRR_SKIP, | |
492 | BMRR_FAIL, | |
493 | } BlkioMemRegionResult; | |
494 | ||
495 | /* | |
496 | * Produce a struct blkio_mem_region for a given address and size. | |
497 | * | |
498 | * This function produces identical results when called multiple times with the | |
499 | * same arguments. This property is necessary because blkio_unmap_mem_region() | |
500 | * must receive the same struct blkio_mem_region field values that were passed | |
501 | * to blkio_map_mem_region(). | |
502 | */ | |
503 | static BlkioMemRegionResult | |
504 | blkio_mem_region_from_host(BlockDriverState *bs, | |
505 | void *host, size_t size, | |
506 | struct blkio_mem_region *region, | |
507 | Error **errp) | |
508 | { | |
509 | BDRVBlkioState *s = bs->opaque; | |
510 | int fd = -1; | |
511 | ram_addr_t fd_offset = 0; | |
512 | ||
513 | if (((uintptr_t)host | size) % s->mem_region_alignment) { | |
514 | error_setg(errp, "unaligned buf %p with size %zu", host, size); | |
515 | return BMRR_FAIL; | |
516 | } | |
517 | ||
518 | /* Attempt to find the fd for the underlying memory */ | |
519 | if (s->needs_mem_region_fd) { | |
520 | RAMBlock *ram_block; | |
521 | RAMBlock *end_block; | |
522 | ram_addr_t offset; | |
523 | ||
524 | /* | |
525 | * bdrv_register_buf() is called with the BQL held so mr lives at least | |
526 | * until this function returns. | |
527 | */ | |
528 | ram_block = qemu_ram_block_from_host(host, false, &fd_offset); | |
529 | if (ram_block) { | |
530 | fd = qemu_ram_get_fd(ram_block); | |
531 | } | |
532 | if (fd == -1) { | |
533 | /* | |
534 | * Ideally every RAMBlock would have an fd. pc-bios and other | |
535 | * things don't. Luckily they are usually not I/O buffers and we | |
536 | * can just ignore them. | |
537 | */ | |
538 | return BMRR_SKIP; | |
539 | } | |
540 | ||
541 | /* Make sure the fd covers the entire range */ | |
542 | end_block = qemu_ram_block_from_host(host + size - 1, false, &offset); | |
543 | if (ram_block != end_block) { | |
544 | error_setg(errp, "registered buffer at %p with size %zu extends " | |
545 | "beyond RAMBlock", host, size); | |
546 | return BMRR_FAIL; | |
547 | } | |
548 | } | |
549 | ||
550 | *region = (struct blkio_mem_region){ | |
551 | .addr = host, | |
552 | .len = size, | |
553 | .fd = fd, | |
554 | .fd_offset = fd_offset, | |
555 | }; | |
556 | return BMRR_OK; | |
557 | } | |
558 | ||
559 | static bool blkio_register_buf(BlockDriverState *bs, void *host, size_t size, | |
560 | Error **errp) | |
561 | { | |
562 | BDRVBlkioState *s = bs->opaque; | |
563 | struct blkio_mem_region region; | |
564 | BlkioMemRegionResult region_result; | |
565 | int ret; | |
566 | ||
567 | /* | |
568 | * Mapping memory regions conflicts with RAM discard (virtio-mem) when | |
569 | * there is pinning, so only do it when necessary. | |
570 | */ | |
571 | if (!s->needs_mem_regions && s->may_pin_mem_regions) { | |
572 | return true; | |
573 | } | |
574 | ||
575 | region_result = blkio_mem_region_from_host(bs, host, size, ®ion, errp); | |
576 | if (region_result == BMRR_SKIP) { | |
577 | return true; | |
578 | } else if (region_result != BMRR_OK) { | |
579 | return false; | |
580 | } | |
581 | ||
582 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
583 | ret = blkio_map_mem_region(s->blkio, ®ion); | |
584 | } | |
585 | ||
586 | if (ret < 0) { | |
587 | error_setg(errp, "Failed to add blkio mem region %p with size %zu: %s", | |
588 | host, size, blkio_get_error_msg()); | |
589 | return false; | |
590 | } | |
591 | return true; | |
592 | } | |
593 | ||
594 | static void blkio_unregister_buf(BlockDriverState *bs, void *host, size_t size) | |
595 | { | |
596 | BDRVBlkioState *s = bs->opaque; | |
597 | struct blkio_mem_region region; | |
598 | ||
599 | /* See blkio_register_buf() */ | |
600 | if (!s->needs_mem_regions && s->may_pin_mem_regions) { | |
601 | return; | |
602 | } | |
603 | ||
604 | if (blkio_mem_region_from_host(bs, host, size, ®ion, NULL) != BMRR_OK) { | |
605 | return; | |
606 | } | |
607 | ||
608 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
609 | blkio_unmap_mem_region(s->blkio, ®ion); | |
610 | } | |
611 | } | |
612 | ||
fd66dbd4 SH |
613 | static int blkio_io_uring_open(BlockDriverState *bs, QDict *options, int flags, |
614 | Error **errp) | |
615 | { | |
616 | const char *filename = qdict_get_str(options, "filename"); | |
617 | BDRVBlkioState *s = bs->opaque; | |
618 | int ret; | |
619 | ||
620 | ret = blkio_set_str(s->blkio, "path", filename); | |
621 | qdict_del(options, "filename"); | |
622 | if (ret < 0) { | |
623 | error_setg_errno(errp, -ret, "failed to set path: %s", | |
624 | blkio_get_error_msg()); | |
625 | return ret; | |
626 | } | |
627 | ||
628 | if (flags & BDRV_O_NOCACHE) { | |
629 | ret = blkio_set_bool(s->blkio, "direct", true); | |
630 | if (ret < 0) { | |
631 | error_setg_errno(errp, -ret, "failed to set direct: %s", | |
632 | blkio_get_error_msg()); | |
633 | return ret; | |
634 | } | |
635 | } | |
636 | ||
637 | return 0; | |
638 | } | |
639 | ||
640 | static int blkio_nvme_io_uring(BlockDriverState *bs, QDict *options, int flags, | |
641 | Error **errp) | |
642 | { | |
6c32fc0d | 643 | const char *path = qdict_get_try_str(options, "path"); |
fd66dbd4 SH |
644 | BDRVBlkioState *s = bs->opaque; |
645 | int ret; | |
646 | ||
6c32fc0d AF |
647 | if (!path) { |
648 | error_setg(errp, "missing 'path' option"); | |
649 | return -EINVAL; | |
650 | } | |
651 | ||
652 | ret = blkio_set_str(s->blkio, "path", path); | |
653 | qdict_del(options, "path"); | |
fd66dbd4 SH |
654 | if (ret < 0) { |
655 | error_setg_errno(errp, -ret, "failed to set path: %s", | |
656 | blkio_get_error_msg()); | |
657 | return ret; | |
658 | } | |
659 | ||
660 | if (!(flags & BDRV_O_NOCACHE)) { | |
661 | error_setg(errp, "cache.direct=off is not supported"); | |
662 | return -EINVAL; | |
663 | } | |
664 | ||
665 | return 0; | |
666 | } | |
667 | ||
668 | static int blkio_virtio_blk_common_open(BlockDriverState *bs, | |
669 | QDict *options, int flags, Error **errp) | |
670 | { | |
671 | const char *path = qdict_get_try_str(options, "path"); | |
672 | BDRVBlkioState *s = bs->opaque; | |
673 | int ret; | |
674 | ||
675 | if (!path) { | |
676 | error_setg(errp, "missing 'path' option"); | |
677 | return -EINVAL; | |
678 | } | |
679 | ||
680 | ret = blkio_set_str(s->blkio, "path", path); | |
681 | qdict_del(options, "path"); | |
682 | if (ret < 0) { | |
683 | error_setg_errno(errp, -ret, "failed to set path: %s", | |
684 | blkio_get_error_msg()); | |
685 | return ret; | |
686 | } | |
687 | ||
688 | if (!(flags & BDRV_O_NOCACHE)) { | |
689 | error_setg(errp, "cache.direct=off is not supported"); | |
690 | return -EINVAL; | |
691 | } | |
692 | return 0; | |
693 | } | |
694 | ||
695 | static int blkio_file_open(BlockDriverState *bs, QDict *options, int flags, | |
696 | Error **errp) | |
697 | { | |
698 | const char *blkio_driver = bs->drv->protocol_name; | |
699 | BDRVBlkioState *s = bs->opaque; | |
700 | int ret; | |
701 | ||
702 | ret = blkio_create(blkio_driver, &s->blkio); | |
703 | if (ret < 0) { | |
704 | error_setg_errno(errp, -ret, "blkio_create failed: %s", | |
705 | blkio_get_error_msg()); | |
706 | return ret; | |
707 | } | |
708 | ||
709 | if (strcmp(blkio_driver, DRIVER_IO_URING) == 0) { | |
710 | ret = blkio_io_uring_open(bs, options, flags, errp); | |
711 | } else if (strcmp(blkio_driver, DRIVER_NVME_IO_URING) == 0) { | |
712 | ret = blkio_nvme_io_uring(bs, options, flags, errp); | |
03d9e4c0 AF |
713 | } else if (strcmp(blkio_driver, DRIVER_VIRTIO_BLK_VFIO_PCI) == 0) { |
714 | ret = blkio_virtio_blk_common_open(bs, options, flags, errp); | |
fd66dbd4 SH |
715 | } else if (strcmp(blkio_driver, DRIVER_VIRTIO_BLK_VHOST_USER) == 0) { |
716 | ret = blkio_virtio_blk_common_open(bs, options, flags, errp); | |
717 | } else if (strcmp(blkio_driver, DRIVER_VIRTIO_BLK_VHOST_VDPA) == 0) { | |
718 | ret = blkio_virtio_blk_common_open(bs, options, flags, errp); | |
719 | } else { | |
720 | g_assert_not_reached(); | |
721 | } | |
722 | if (ret < 0) { | |
723 | blkio_destroy(&s->blkio); | |
724 | return ret; | |
725 | } | |
726 | ||
727 | if (!(flags & BDRV_O_RDWR)) { | |
728 | ret = blkio_set_bool(s->blkio, "read-only", true); | |
729 | if (ret < 0) { | |
730 | error_setg_errno(errp, -ret, "failed to set read-only: %s", | |
731 | blkio_get_error_msg()); | |
732 | blkio_destroy(&s->blkio); | |
733 | return ret; | |
734 | } | |
735 | } | |
736 | ||
737 | ret = blkio_connect(s->blkio); | |
738 | if (ret < 0) { | |
739 | error_setg_errno(errp, -ret, "blkio_connect failed: %s", | |
740 | blkio_get_error_msg()); | |
741 | blkio_destroy(&s->blkio); | |
742 | return ret; | |
743 | } | |
744 | ||
745 | ret = blkio_get_bool(s->blkio, | |
746 | "needs-mem-regions", | |
747 | &s->needs_mem_regions); | |
748 | if (ret < 0) { | |
749 | error_setg_errno(errp, -ret, | |
750 | "failed to get needs-mem-regions: %s", | |
751 | blkio_get_error_msg()); | |
752 | blkio_destroy(&s->blkio); | |
753 | return ret; | |
754 | } | |
755 | ||
c5640b3e SH |
756 | ret = blkio_get_bool(s->blkio, |
757 | "needs-mem-region-fd", | |
758 | &s->needs_mem_region_fd); | |
759 | if (ret < 0) { | |
760 | error_setg_errno(errp, -ret, | |
761 | "failed to get needs-mem-region-fd: %s", | |
762 | blkio_get_error_msg()); | |
763 | blkio_destroy(&s->blkio); | |
764 | return ret; | |
765 | } | |
766 | ||
fd66dbd4 SH |
767 | ret = blkio_get_uint64(s->blkio, |
768 | "mem-region-alignment", | |
769 | &s->mem_region_alignment); | |
770 | if (ret < 0) { | |
771 | error_setg_errno(errp, -ret, | |
772 | "failed to get mem-region-alignment: %s", | |
773 | blkio_get_error_msg()); | |
774 | blkio_destroy(&s->blkio); | |
775 | return ret; | |
776 | } | |
777 | ||
c5640b3e SH |
778 | ret = blkio_get_bool(s->blkio, |
779 | "may-pin-mem-regions", | |
780 | &s->may_pin_mem_regions); | |
781 | if (ret < 0) { | |
782 | /* Be conservative (assume pinning) if the property is not supported */ | |
783 | s->may_pin_mem_regions = s->needs_mem_regions; | |
784 | } | |
785 | ||
786 | /* | |
787 | * Notify if libblkio drivers pin memory and prevent features like | |
788 | * virtio-mem from working. | |
789 | */ | |
790 | if (s->may_pin_mem_regions) { | |
791 | ret = ram_block_discard_disable(true); | |
792 | if (ret < 0) { | |
793 | error_setg_errno(errp, -ret, "ram_block_discard_disable() failed"); | |
794 | blkio_destroy(&s->blkio); | |
795 | return ret; | |
796 | } | |
797 | } | |
798 | ||
fd66dbd4 SH |
799 | ret = blkio_start(s->blkio); |
800 | if (ret < 0) { | |
801 | error_setg_errno(errp, -ret, "blkio_start failed: %s", | |
802 | blkio_get_error_msg()); | |
803 | blkio_destroy(&s->blkio); | |
c5640b3e SH |
804 | if (s->may_pin_mem_regions) { |
805 | ram_block_discard_disable(false); | |
806 | } | |
fd66dbd4 SH |
807 | return ret; |
808 | } | |
809 | ||
c5640b3e | 810 | bs->supported_write_flags = BDRV_REQ_FUA | BDRV_REQ_REGISTERED_BUF; |
fd66dbd4 SH |
811 | bs->supported_zero_flags = BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | |
812 | BDRV_REQ_NO_FALLBACK; | |
813 | ||
814 | qemu_mutex_init(&s->blkio_lock); | |
815 | qemu_co_mutex_init(&s->bounce_lock); | |
816 | qemu_co_queue_init(&s->bounce_available); | |
817 | QLIST_INIT(&s->bounce_bufs); | |
818 | s->blkioq = blkio_get_queue(s->blkio, 0); | |
819 | s->completion_fd = blkioq_get_completion_fd(s->blkioq); | |
820 | ||
821 | blkio_attach_aio_context(bs, bdrv_get_aio_context(bs)); | |
822 | return 0; | |
823 | } | |
824 | ||
825 | static void blkio_close(BlockDriverState *bs) | |
826 | { | |
827 | BDRVBlkioState *s = bs->opaque; | |
828 | ||
829 | /* There is no destroy() API for s->bounce_lock */ | |
830 | ||
831 | qemu_mutex_destroy(&s->blkio_lock); | |
832 | blkio_detach_aio_context(bs); | |
833 | blkio_destroy(&s->blkio); | |
c5640b3e SH |
834 | |
835 | if (s->may_pin_mem_regions) { | |
836 | ram_block_discard_disable(false); | |
837 | } | |
fd66dbd4 SH |
838 | } |
839 | ||
840 | static int64_t blkio_getlength(BlockDriverState *bs) | |
841 | { | |
842 | BDRVBlkioState *s = bs->opaque; | |
843 | uint64_t capacity; | |
844 | int ret; | |
845 | ||
846 | WITH_QEMU_LOCK_GUARD(&s->blkio_lock) { | |
847 | ret = blkio_get_uint64(s->blkio, "capacity", &capacity); | |
848 | } | |
849 | if (ret < 0) { | |
850 | return -ret; | |
851 | } | |
852 | ||
853 | return capacity; | |
854 | } | |
855 | ||
4c8f4fda AF |
856 | static int coroutine_fn blkio_truncate(BlockDriverState *bs, int64_t offset, |
857 | bool exact, PreallocMode prealloc, | |
858 | BdrvRequestFlags flags, Error **errp) | |
859 | { | |
860 | int64_t current_length; | |
861 | ||
862 | if (prealloc != PREALLOC_MODE_OFF) { | |
863 | error_setg(errp, "Unsupported preallocation mode '%s'", | |
864 | PreallocMode_str(prealloc)); | |
865 | return -ENOTSUP; | |
866 | } | |
867 | ||
868 | current_length = blkio_getlength(bs); | |
869 | ||
870 | if (offset > current_length) { | |
871 | error_setg(errp, "Cannot grow device"); | |
872 | return -EINVAL; | |
873 | } else if (exact && offset != current_length) { | |
874 | error_setg(errp, "Cannot resize device"); | |
875 | return -ENOTSUP; | |
876 | } | |
877 | ||
878 | return 0; | |
879 | } | |
880 | ||
fd66dbd4 SH |
881 | static int blkio_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) |
882 | { | |
883 | return 0; | |
884 | } | |
885 | ||
886 | static void blkio_refresh_limits(BlockDriverState *bs, Error **errp) | |
887 | { | |
888 | BDRVBlkioState *s = bs->opaque; | |
889 | QEMU_LOCK_GUARD(&s->blkio_lock); | |
890 | int value; | |
891 | int ret; | |
892 | ||
893 | ret = blkio_get_int(s->blkio, "request-alignment", &value); | |
894 | if (ret < 0) { | |
895 | error_setg_errno(errp, -ret, "failed to get \"request-alignment\": %s", | |
896 | blkio_get_error_msg()); | |
897 | return; | |
898 | } | |
899 | bs->bl.request_alignment = value; | |
900 | if (bs->bl.request_alignment < 1 || | |
901 | bs->bl.request_alignment >= INT_MAX || | |
902 | !is_power_of_2(bs->bl.request_alignment)) { | |
903 | error_setg(errp, "invalid \"request-alignment\" value %" PRIu32 ", " | |
904 | "must be a power of 2 less than INT_MAX", | |
905 | bs->bl.request_alignment); | |
906 | return; | |
907 | } | |
908 | ||
909 | ret = blkio_get_int(s->blkio, "optimal-io-size", &value); | |
910 | if (ret < 0) { | |
911 | error_setg_errno(errp, -ret, "failed to get \"optimal-io-size\": %s", | |
912 | blkio_get_error_msg()); | |
913 | return; | |
914 | } | |
915 | bs->bl.opt_transfer = value; | |
916 | if (bs->bl.opt_transfer > INT_MAX || | |
917 | (bs->bl.opt_transfer % bs->bl.request_alignment)) { | |
918 | error_setg(errp, "invalid \"optimal-io-size\" value %" PRIu32 ", must " | |
919 | "be a multiple of %" PRIu32, bs->bl.opt_transfer, | |
920 | bs->bl.request_alignment); | |
921 | return; | |
922 | } | |
923 | ||
924 | ret = blkio_get_int(s->blkio, "max-transfer", &value); | |
925 | if (ret < 0) { | |
926 | error_setg_errno(errp, -ret, "failed to get \"max-transfer\": %s", | |
927 | blkio_get_error_msg()); | |
928 | return; | |
929 | } | |
930 | bs->bl.max_transfer = value; | |
931 | if ((bs->bl.max_transfer % bs->bl.request_alignment) || | |
932 | (bs->bl.opt_transfer && (bs->bl.max_transfer % bs->bl.opt_transfer))) { | |
933 | error_setg(errp, "invalid \"max-transfer\" value %" PRIu32 ", must be " | |
934 | "a multiple of %" PRIu32 " and %" PRIu32 " (if non-zero)", | |
935 | bs->bl.max_transfer, bs->bl.request_alignment, | |
936 | bs->bl.opt_transfer); | |
937 | return; | |
938 | } | |
939 | ||
940 | ret = blkio_get_int(s->blkio, "buf-alignment", &value); | |
941 | if (ret < 0) { | |
942 | error_setg_errno(errp, -ret, "failed to get \"buf-alignment\": %s", | |
943 | blkio_get_error_msg()); | |
944 | return; | |
945 | } | |
946 | if (value < 1) { | |
947 | error_setg(errp, "invalid \"buf-alignment\" value %d, must be " | |
948 | "positive", value); | |
949 | return; | |
950 | } | |
951 | bs->bl.min_mem_alignment = value; | |
952 | ||
953 | ret = blkio_get_int(s->blkio, "optimal-buf-alignment", &value); | |
954 | if (ret < 0) { | |
955 | error_setg_errno(errp, -ret, | |
956 | "failed to get \"optimal-buf-alignment\": %s", | |
957 | blkio_get_error_msg()); | |
958 | return; | |
959 | } | |
960 | if (value < 1) { | |
961 | error_setg(errp, "invalid \"optimal-buf-alignment\" value %d, " | |
962 | "must be positive", value); | |
963 | return; | |
964 | } | |
965 | bs->bl.opt_mem_alignment = value; | |
966 | ||
967 | ret = blkio_get_int(s->blkio, "max-segments", &value); | |
968 | if (ret < 0) { | |
969 | error_setg_errno(errp, -ret, "failed to get \"max-segments\": %s", | |
970 | blkio_get_error_msg()); | |
971 | return; | |
972 | } | |
973 | if (value < 1) { | |
974 | error_setg(errp, "invalid \"max-segments\" value %d, must be positive", | |
975 | value); | |
976 | return; | |
977 | } | |
978 | bs->bl.max_iov = value; | |
979 | } | |
980 | ||
981 | /* | |
982 | * TODO | |
983 | * Missing libblkio APIs: | |
984 | * - block_status | |
985 | * - co_invalidate_cache | |
986 | * | |
987 | * Out of scope? | |
988 | * - create | |
989 | * - truncate | |
990 | */ | |
991 | ||
992 | #define BLKIO_DRIVER(name, ...) \ | |
993 | { \ | |
994 | .format_name = name, \ | |
995 | .protocol_name = name, \ | |
996 | .instance_size = sizeof(BDRVBlkioState), \ | |
997 | .bdrv_file_open = blkio_file_open, \ | |
998 | .bdrv_close = blkio_close, \ | |
999 | .bdrv_getlength = blkio_getlength, \ | |
4c8f4fda | 1000 | .bdrv_co_truncate = blkio_truncate, \ |
fd66dbd4 SH |
1001 | .bdrv_get_info = blkio_get_info, \ |
1002 | .bdrv_attach_aio_context = blkio_attach_aio_context, \ | |
1003 | .bdrv_detach_aio_context = blkio_detach_aio_context, \ | |
1004 | .bdrv_co_pdiscard = blkio_co_pdiscard, \ | |
1005 | .bdrv_co_preadv = blkio_co_preadv, \ | |
1006 | .bdrv_co_pwritev = blkio_co_pwritev, \ | |
1007 | .bdrv_co_flush_to_disk = blkio_co_flush, \ | |
1008 | .bdrv_co_pwrite_zeroes = blkio_co_pwrite_zeroes, \ | |
1009 | .bdrv_io_unplug = blkio_io_unplug, \ | |
1010 | .bdrv_refresh_limits = blkio_refresh_limits, \ | |
c5640b3e SH |
1011 | .bdrv_register_buf = blkio_register_buf, \ |
1012 | .bdrv_unregister_buf = blkio_unregister_buf, \ | |
fd66dbd4 SH |
1013 | __VA_ARGS__ \ |
1014 | } | |
1015 | ||
1016 | static BlockDriver bdrv_io_uring = BLKIO_DRIVER( | |
1017 | DRIVER_IO_URING, | |
1018 | .bdrv_needs_filename = true, | |
1019 | ); | |
1020 | ||
1021 | static BlockDriver bdrv_nvme_io_uring = BLKIO_DRIVER( | |
1022 | DRIVER_NVME_IO_URING, | |
fd66dbd4 SH |
1023 | ); |
1024 | ||
03d9e4c0 AF |
1025 | static BlockDriver bdrv_virtio_blk_vfio_pci = BLKIO_DRIVER( |
1026 | DRIVER_VIRTIO_BLK_VFIO_PCI | |
1027 | ); | |
1028 | ||
fd66dbd4 SH |
1029 | static BlockDriver bdrv_virtio_blk_vhost_user = BLKIO_DRIVER( |
1030 | DRIVER_VIRTIO_BLK_VHOST_USER | |
1031 | ); | |
1032 | ||
1033 | static BlockDriver bdrv_virtio_blk_vhost_vdpa = BLKIO_DRIVER( | |
1034 | DRIVER_VIRTIO_BLK_VHOST_VDPA | |
1035 | ); | |
1036 | ||
1037 | static void bdrv_blkio_init(void) | |
1038 | { | |
1039 | bdrv_register(&bdrv_io_uring); | |
1040 | bdrv_register(&bdrv_nvme_io_uring); | |
03d9e4c0 | 1041 | bdrv_register(&bdrv_virtio_blk_vfio_pci); |
fd66dbd4 SH |
1042 | bdrv_register(&bdrv_virtio_blk_vhost_user); |
1043 | bdrv_register(&bdrv_virtio_blk_vhost_vdpa); | |
1044 | } | |
1045 | ||
1046 | block_init(bdrv_blkio_init); |