xref: /freebsd/sys/contrib/openzfs/module/os/linux/zfs/zvol_os.c (revision 61145dc2b94f12f6a47344fb9aac702321880e43)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
24  * Copyright (c) 2024, Rob Norris <robn@despairlabs.com>
25  * Copyright (c) 2024, Klara, Inc.
26  */
27 
28 #include <sys/dataset_kstats.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_traverse.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/zap.h>
35 #include <sys/zfeature.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/zio.h>
39 #include <sys/zfs_rlock.h>
40 #include <sys/spa_impl.h>
41 #include <sys/zvol.h>
42 #include <sys/zvol_impl.h>
43 #include <cityhash.h>
44 
45 #include <linux/blkdev_compat.h>
46 #include <linux/task_io_accounting_ops.h>
47 #include <linux/workqueue.h>
48 #include <linux/blk-mq.h>
49 
50 static void zvol_request_impl(zvol_state_t *zv, struct bio *bio,
51     struct request *rq, boolean_t force_sync);
52 
53 static unsigned int zvol_major = ZVOL_MAJOR;
54 static unsigned int zvol_request_sync = 0;
55 static unsigned int zvol_prefetch_bytes = (128 * 1024);
56 static unsigned long zvol_max_discard_blocks = 16384;
57 
58 /*
59  * Switch taskq at multiple of 512 MB offset. This can be set to a lower value
60  * to utilize more threads for small files but may affect prefetch hits.
61  */
62 #define	ZVOL_TASKQ_OFFSET_SHIFT 29
63 
64 #ifndef HAVE_BLKDEV_GET_ERESTARTSYS
65 static unsigned int zvol_open_timeout_ms = 1000;
66 #endif
67 
68 static unsigned int zvol_threads = 0;
69 static unsigned int zvol_blk_mq_threads = 0;
70 static unsigned int zvol_blk_mq_actual_threads;
71 static boolean_t zvol_use_blk_mq = B_FALSE;
72 
73 /*
74  * The maximum number of volblocksize blocks to process per thread.  Typically,
75  * write heavy workloads preform better with higher values here, and read
76  * heavy workloads preform better with lower values, but that's not a hard
77  * and fast rule.  It's basically a knob to tune between "less overhead with
78  * less parallelism" and "more overhead, but more parallelism".
79  *
80  * '8' was chosen as a reasonable, balanced, default based off of sequential
81  * read and write tests to a zvol in an NVMe pool (with 16 CPUs).
82  */
83 static unsigned int zvol_blk_mq_blocks_per_thread = 8;
84 
85 static unsigned int zvol_num_taskqs = 0;
86 
87 #ifndef	BLKDEV_DEFAULT_RQ
88 /* BLKDEV_MAX_RQ was renamed to BLKDEV_DEFAULT_RQ in the 5.16 kernel */
89 #define	BLKDEV_DEFAULT_RQ BLKDEV_MAX_RQ
90 #endif
91 
92 /*
93  * Finalize our BIO or request.
94  */
95 static inline void
zvol_end_io(struct bio * bio,struct request * rq,int error)96 zvol_end_io(struct bio *bio, struct request *rq, int error)
97 {
98 	if (bio) {
99 		bio->bi_status = errno_to_bi_status(-error);
100 		bio_endio(bio);
101 	} else {
102 		blk_mq_end_request(rq, errno_to_bi_status(error));
103 	}
104 }
105 
106 static unsigned int zvol_blk_mq_queue_depth = BLKDEV_DEFAULT_RQ;
107 static unsigned int zvol_actual_blk_mq_queue_depth;
108 
109 struct zvol_state_os {
110 	struct gendisk		*zvo_disk;	/* generic disk */
111 	struct request_queue	*zvo_queue;	/* request queue */
112 	dev_t			zvo_dev;	/* device id */
113 
114 	struct blk_mq_tag_set tag_set;
115 
116 	/* Set from the global 'zvol_use_blk_mq' at zvol load */
117 	boolean_t use_blk_mq;
118 };
119 
120 typedef struct zv_taskq {
121 	uint_t tqs_cnt;
122 	taskq_t **tqs_taskq;
123 } zv_taskq_t;
124 static zv_taskq_t zvol_taskqs;
125 static struct ida zvol_ida;
126 
127 typedef struct zv_request_stack {
128 	zvol_state_t	*zv;
129 	struct bio	*bio;
130 	struct request *rq;
131 } zv_request_t;
132 
133 typedef struct zv_work {
134 	struct request  *rq;
135 	struct work_struct work;
136 } zv_work_t;
137 
138 typedef struct zv_request_task {
139 	zv_request_t zvr;
140 	taskq_ent_t	ent;
141 } zv_request_task_t;
142 
143 static zv_request_task_t *
zv_request_task_create(zv_request_t zvr)144 zv_request_task_create(zv_request_t zvr)
145 {
146 	zv_request_task_t *task;
147 	task = kmem_alloc(sizeof (zv_request_task_t), KM_SLEEP);
148 	taskq_init_ent(&task->ent);
149 	task->zvr = zvr;
150 	return (task);
151 }
152 
153 static void
zv_request_task_free(zv_request_task_t * task)154 zv_request_task_free(zv_request_task_t *task)
155 {
156 	kmem_free(task, sizeof (*task));
157 }
158 
159 /*
160  * This is called when a new block multiqueue request comes in.  A request
161  * contains one or more BIOs.
162  */
zvol_mq_queue_rq(struct blk_mq_hw_ctx * hctx,const struct blk_mq_queue_data * bd)163 static blk_status_t zvol_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
164     const struct blk_mq_queue_data *bd)
165 {
166 	struct request *rq = bd->rq;
167 	zvol_state_t *zv = rq->q->queuedata;
168 
169 	/* Tell the kernel that we are starting to process this request */
170 	blk_mq_start_request(rq);
171 
172 	if (blk_rq_is_passthrough(rq)) {
173 		/* Skip non filesystem request */
174 		blk_mq_end_request(rq, BLK_STS_IOERR);
175 		return (BLK_STS_IOERR);
176 	}
177 
178 	zvol_request_impl(zv, NULL, rq, 0);
179 
180 	/* Acknowledge to the kernel that we got this request */
181 	return (BLK_STS_OK);
182 }
183 
184 static struct blk_mq_ops zvol_blk_mq_queue_ops = {
185 	.queue_rq = zvol_mq_queue_rq,
186 };
187 
188 /* Initialize our blk-mq struct */
zvol_blk_mq_alloc_tag_set(zvol_state_t * zv)189 static int zvol_blk_mq_alloc_tag_set(zvol_state_t *zv)
190 {
191 	struct zvol_state_os *zso = zv->zv_zso;
192 
193 	memset(&zso->tag_set, 0, sizeof (zso->tag_set));
194 
195 	/* Initialize tag set. */
196 	zso->tag_set.ops = &zvol_blk_mq_queue_ops;
197 	zso->tag_set.nr_hw_queues = zvol_blk_mq_actual_threads;
198 	zso->tag_set.queue_depth = zvol_actual_blk_mq_queue_depth;
199 	zso->tag_set.numa_node = NUMA_NO_NODE;
200 	zso->tag_set.cmd_size = 0;
201 
202 	/*
203 	 * We need BLK_MQ_F_BLOCKING here since we do blocking calls in
204 	 * zvol_request_impl()
205 	 */
206 	zso->tag_set.flags = BLK_MQ_F_BLOCKING;
207 
208 #ifdef BLK_MQ_F_SHOULD_MERGE
209 	/*
210 	 * Linux 6.14 removed BLK_MQ_F_SHOULD_MERGE and made it implicit.
211 	 * For older kernels, we set it.
212 	 */
213 	zso->tag_set.flags |= BLK_MQ_F_SHOULD_MERGE;
214 #endif
215 
216 	zso->tag_set.driver_data = zv;
217 
218 	return (blk_mq_alloc_tag_set(&zso->tag_set));
219 }
220 
221 /*
222  * Given a path, return TRUE if path is a ZVOL.
223  */
224 boolean_t
zvol_os_is_zvol(const char * path)225 zvol_os_is_zvol(const char *path)
226 {
227 	dev_t dev = 0;
228 
229 	if (vdev_lookup_bdev(path, &dev) != 0)
230 		return (B_FALSE);
231 
232 	if (MAJOR(dev) == zvol_major)
233 		return (B_TRUE);
234 
235 	return (B_FALSE);
236 }
237 
238 static void
zvol_write(zv_request_t * zvr)239 zvol_write(zv_request_t *zvr)
240 {
241 	struct bio *bio = zvr->bio;
242 	struct request *rq = zvr->rq;
243 	int error = 0;
244 	zfs_uio_t uio;
245 	zvol_state_t *zv = zvr->zv;
246 	struct request_queue *q;
247 	struct gendisk *disk;
248 	unsigned long start_time = 0;
249 	boolean_t acct = B_FALSE;
250 
251 	ASSERT3P(zv, !=, NULL);
252 	ASSERT3U(zv->zv_open_count, >, 0);
253 	ASSERT3P(zv->zv_zilog, !=, NULL);
254 
255 	q = zv->zv_zso->zvo_queue;
256 	disk = zv->zv_zso->zvo_disk;
257 
258 	/* bio marked as FLUSH need to flush before write */
259 	if (io_is_flush(bio, rq))
260 		zil_commit(zv->zv_zilog, ZVOL_OBJ);
261 
262 	/* Some requests are just for flush and nothing else. */
263 	if (io_size(bio, rq) == 0) {
264 		rw_exit(&zv->zv_suspend_lock);
265 		zvol_end_io(bio, rq, 0);
266 		return;
267 	}
268 
269 	zfs_uio_bvec_init(&uio, bio, rq);
270 
271 	ssize_t start_resid = uio.uio_resid;
272 
273 	/*
274 	 * With use_blk_mq, accounting is done by blk_mq_start_request()
275 	 * and blk_mq_end_request(), so we can skip it here.
276 	 */
277 	if (bio) {
278 		acct = blk_queue_io_stat(q);
279 		if (acct) {
280 			start_time = blk_generic_start_io_acct(q, disk, WRITE,
281 			    bio);
282 		}
283 	}
284 
285 	boolean_t sync =
286 	    io_is_fua(bio, rq) || zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS;
287 
288 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zv->zv_rangelock,
289 	    uio.uio_loffset, uio.uio_resid, RL_WRITER);
290 
291 	uint64_t volsize = zv->zv_volsize;
292 	while (uio.uio_resid > 0 && uio.uio_loffset < volsize) {
293 		uint64_t bytes = MIN(uio.uio_resid, DMU_MAX_ACCESS >> 1);
294 		uint64_t off = uio.uio_loffset;
295 		dmu_tx_t *tx = dmu_tx_create(zv->zv_objset);
296 
297 		if (bytes > volsize - off)	/* don't write past the end */
298 			bytes = volsize - off;
299 
300 		dmu_tx_hold_write_by_dnode(tx, zv->zv_dn, off, bytes);
301 
302 		/* This will only fail for ENOSPC */
303 		error = dmu_tx_assign(tx, DMU_TX_WAIT);
304 		if (error) {
305 			dmu_tx_abort(tx);
306 			break;
307 		}
308 		error = dmu_write_uio_dnode(zv->zv_dn, &uio, bytes, tx);
309 		if (error == 0) {
310 			zvol_log_write(zv, tx, off, bytes, sync);
311 		}
312 		dmu_tx_commit(tx);
313 
314 		if (error)
315 			break;
316 	}
317 	zfs_rangelock_exit(lr);
318 
319 	int64_t nwritten = start_resid - uio.uio_resid;
320 	dataset_kstats_update_write_kstats(&zv->zv_kstat, nwritten);
321 	task_io_account_write(nwritten);
322 
323 	if (sync)
324 		zil_commit(zv->zv_zilog, ZVOL_OBJ);
325 
326 	rw_exit(&zv->zv_suspend_lock);
327 
328 	if (bio && acct) {
329 		blk_generic_end_io_acct(q, disk, WRITE, bio, start_time);
330 	}
331 
332 	zvol_end_io(bio, rq, -error);
333 }
334 
335 static void
zvol_write_task(void * arg)336 zvol_write_task(void *arg)
337 {
338 	zv_request_task_t *task = arg;
339 	zvol_write(&task->zvr);
340 	zv_request_task_free(task);
341 }
342 
343 static void
zvol_discard(zv_request_t * zvr)344 zvol_discard(zv_request_t *zvr)
345 {
346 	struct bio *bio = zvr->bio;
347 	struct request *rq = zvr->rq;
348 	zvol_state_t *zv = zvr->zv;
349 	uint64_t start = io_offset(bio, rq);
350 	uint64_t size = io_size(bio, rq);
351 	uint64_t end = start + size;
352 	boolean_t sync;
353 	int error = 0;
354 	dmu_tx_t *tx;
355 	struct request_queue *q = zv->zv_zso->zvo_queue;
356 	struct gendisk *disk = zv->zv_zso->zvo_disk;
357 	unsigned long start_time = 0;
358 	boolean_t acct = B_FALSE;
359 
360 	ASSERT3P(zv, !=, NULL);
361 	ASSERT3U(zv->zv_open_count, >, 0);
362 	ASSERT3P(zv->zv_zilog, !=, NULL);
363 
364 	if (bio) {
365 		acct = blk_queue_io_stat(q);
366 		if (acct) {
367 			start_time = blk_generic_start_io_acct(q, disk, WRITE,
368 			    bio);
369 		}
370 	}
371 
372 	sync = io_is_fua(bio, rq) || zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS;
373 
374 	if (end > zv->zv_volsize) {
375 		error = SET_ERROR(EIO);
376 		goto unlock;
377 	}
378 
379 	/*
380 	 * Align the request to volume block boundaries when a secure erase is
381 	 * not required.  This will prevent dnode_free_range() from zeroing out
382 	 * the unaligned parts which is slow (read-modify-write) and useless
383 	 * since we are not freeing any space by doing so.
384 	 */
385 	if (!io_is_secure_erase(bio, rq)) {
386 		start = P2ROUNDUP(start, zv->zv_volblocksize);
387 		end = P2ALIGN_TYPED(end, zv->zv_volblocksize, uint64_t);
388 		size = end - start;
389 	}
390 
391 	if (start >= end)
392 		goto unlock;
393 
394 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zv->zv_rangelock,
395 	    start, size, RL_WRITER);
396 
397 	tx = dmu_tx_create(zv->zv_objset);
398 	dmu_tx_mark_netfree(tx);
399 	error = dmu_tx_assign(tx, DMU_TX_WAIT);
400 	if (error != 0) {
401 		dmu_tx_abort(tx);
402 	} else {
403 		zvol_log_truncate(zv, tx, start, size);
404 		dmu_tx_commit(tx);
405 		error = dmu_free_long_range(zv->zv_objset,
406 		    ZVOL_OBJ, start, size);
407 	}
408 	zfs_rangelock_exit(lr);
409 
410 	if (error == 0 && sync)
411 		zil_commit(zv->zv_zilog, ZVOL_OBJ);
412 
413 unlock:
414 	rw_exit(&zv->zv_suspend_lock);
415 
416 	if (bio && acct) {
417 		blk_generic_end_io_acct(q, disk, WRITE, bio,
418 		    start_time);
419 	}
420 
421 	zvol_end_io(bio, rq, -error);
422 }
423 
424 static void
zvol_discard_task(void * arg)425 zvol_discard_task(void *arg)
426 {
427 	zv_request_task_t *task = arg;
428 	zvol_discard(&task->zvr);
429 	zv_request_task_free(task);
430 }
431 
432 static void
zvol_read(zv_request_t * zvr)433 zvol_read(zv_request_t *zvr)
434 {
435 	struct bio *bio = zvr->bio;
436 	struct request *rq = zvr->rq;
437 	int error = 0;
438 	zfs_uio_t uio;
439 	boolean_t acct = B_FALSE;
440 	zvol_state_t *zv = zvr->zv;
441 	struct request_queue *q;
442 	struct gendisk *disk;
443 	unsigned long start_time = 0;
444 
445 	ASSERT3P(zv, !=, NULL);
446 	ASSERT3U(zv->zv_open_count, >, 0);
447 
448 	zfs_uio_bvec_init(&uio, bio, rq);
449 
450 	q = zv->zv_zso->zvo_queue;
451 	disk = zv->zv_zso->zvo_disk;
452 
453 	ssize_t start_resid = uio.uio_resid;
454 
455 	/*
456 	 * When blk-mq is being used, accounting is done by
457 	 * blk_mq_start_request() and blk_mq_end_request().
458 	 */
459 	if (bio) {
460 		acct = blk_queue_io_stat(q);
461 		if (acct)
462 			start_time = blk_generic_start_io_acct(q, disk, READ,
463 			    bio);
464 	}
465 
466 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zv->zv_rangelock,
467 	    uio.uio_loffset, uio.uio_resid, RL_READER);
468 
469 	uint64_t volsize = zv->zv_volsize;
470 
471 	while (uio.uio_resid > 0 && uio.uio_loffset < volsize) {
472 		uint64_t bytes = MIN(uio.uio_resid, DMU_MAX_ACCESS >> 1);
473 
474 		/* don't read past the end */
475 		if (bytes > volsize - uio.uio_loffset)
476 			bytes = volsize - uio.uio_loffset;
477 
478 		error = dmu_read_uio_dnode(zv->zv_dn, &uio, bytes);
479 		if (error) {
480 			/* convert checksum errors into IO errors */
481 			if (error == ECKSUM)
482 				error = SET_ERROR(EIO);
483 			break;
484 		}
485 	}
486 	zfs_rangelock_exit(lr);
487 
488 	int64_t nread = start_resid - uio.uio_resid;
489 	dataset_kstats_update_read_kstats(&zv->zv_kstat, nread);
490 	task_io_account_read(nread);
491 
492 	rw_exit(&zv->zv_suspend_lock);
493 
494 	if (bio && acct) {
495 		blk_generic_end_io_acct(q, disk, READ, bio, start_time);
496 	}
497 
498 	zvol_end_io(bio, rq, -error);
499 }
500 
501 static void
zvol_read_task(void * arg)502 zvol_read_task(void *arg)
503 {
504 	zv_request_task_t *task = arg;
505 	zvol_read(&task->zvr);
506 	zv_request_task_free(task);
507 }
508 
509 
510 /*
511  * Process a BIO or request
512  *
513  * Either 'bio' or 'rq' should be set depending on if we are processing a
514  * bio or a request (both should not be set).
515  *
516  * force_sync:	Set to 0 to defer processing to a background taskq
517  *			Set to 1 to process data synchronously
518  */
519 static void
zvol_request_impl(zvol_state_t * zv,struct bio * bio,struct request * rq,boolean_t force_sync)520 zvol_request_impl(zvol_state_t *zv, struct bio *bio, struct request *rq,
521     boolean_t force_sync)
522 {
523 	fstrans_cookie_t cookie = spl_fstrans_mark();
524 	uint64_t offset = io_offset(bio, rq);
525 	uint64_t size = io_size(bio, rq);
526 	int rw = io_data_dir(bio, rq);
527 
528 	if (unlikely(zv->zv_flags & ZVOL_REMOVING)) {
529 		zvol_end_io(bio, rq, -SET_ERROR(ENXIO));
530 		goto out;
531 	}
532 
533 	if (zvol_request_sync || zv->zv_threading == B_FALSE)
534 		force_sync = 1;
535 
536 	zv_request_t zvr = {
537 		.zv = zv,
538 		.bio = bio,
539 		.rq = rq,
540 	};
541 
542 	if (io_has_data(bio, rq) && offset + size > zv->zv_volsize) {
543 		printk(KERN_INFO "%s: bad access: offset=%llu, size=%lu\n",
544 		    zv->zv_zso->zvo_disk->disk_name,
545 		    (long long unsigned)offset,
546 		    (long unsigned)size);
547 
548 		zvol_end_io(bio, rq, -SET_ERROR(EIO));
549 		goto out;
550 	}
551 
552 	zv_request_task_t *task;
553 	zv_taskq_t *ztqs = &zvol_taskqs;
554 	uint_t blk_mq_hw_queue = 0;
555 	uint_t tq_idx;
556 	uint_t taskq_hash;
557 	if (rq)
558 #ifdef HAVE_BLK_MQ_RQ_HCTX
559 		blk_mq_hw_queue = rq->mq_hctx->queue_num;
560 #else
561 		blk_mq_hw_queue =
562 		    rq->q->queue_hw_ctx[rq->q->mq_map[rq->cpu]]->queue_num;
563 #endif
564 	taskq_hash = cityhash3((uintptr_t)zv, offset >> ZVOL_TASKQ_OFFSET_SHIFT,
565 	    blk_mq_hw_queue);
566 	tq_idx = taskq_hash % ztqs->tqs_cnt;
567 
568 	if (rw == WRITE) {
569 		if (unlikely(zv->zv_flags & ZVOL_RDONLY)) {
570 			zvol_end_io(bio, rq, -SET_ERROR(EROFS));
571 			goto out;
572 		}
573 
574 		/*
575 		 * Prevents the zvol from being suspended, or the ZIL being
576 		 * concurrently opened.  Will be released after the i/o
577 		 * completes.
578 		 */
579 		rw_enter(&zv->zv_suspend_lock, RW_READER);
580 
581 		/*
582 		 * Open a ZIL if this is the first time we have written to this
583 		 * zvol. We protect zv->zv_zilog with zv_suspend_lock rather
584 		 * than zv_state_lock so that we don't need to acquire an
585 		 * additional lock in this path.
586 		 */
587 		if (zv->zv_zilog == NULL) {
588 			rw_exit(&zv->zv_suspend_lock);
589 			rw_enter(&zv->zv_suspend_lock, RW_WRITER);
590 			if (zv->zv_zilog == NULL) {
591 				zv->zv_zilog = zil_open(zv->zv_objset,
592 				    zvol_get_data, &zv->zv_kstat.dk_zil_sums);
593 				zv->zv_flags |= ZVOL_WRITTEN_TO;
594 				/* replay / destroy done in zvol_create_minor */
595 				VERIFY0((zv->zv_zilog->zl_header->zh_flags &
596 				    ZIL_REPLAY_NEEDED));
597 			}
598 			rw_downgrade(&zv->zv_suspend_lock);
599 		}
600 
601 		/*
602 		 * We don't want this thread to be blocked waiting for i/o to
603 		 * complete, so we instead wait from a taskq callback. The
604 		 * i/o may be a ZIL write (via zil_commit()), or a read of an
605 		 * indirect block, or a read of a data block (if this is a
606 		 * partial-block write).  We will indicate that the i/o is
607 		 * complete by calling END_IO() from the taskq callback.
608 		 *
609 		 * This design allows the calling thread to continue and
610 		 * initiate more concurrent operations by calling
611 		 * zvol_request() again. There are typically only a small
612 		 * number of threads available to call zvol_request() (e.g.
613 		 * one per iSCSI target), so keeping the latency of
614 		 * zvol_request() low is important for performance.
615 		 *
616 		 * The zvol_request_sync module parameter allows this
617 		 * behavior to be altered, for performance evaluation
618 		 * purposes.  If the callback blocks, setting
619 		 * zvol_request_sync=1 will result in much worse performance.
620 		 *
621 		 * We can have up to zvol_threads concurrent i/o's being
622 		 * processed for all zvols on the system.  This is typically
623 		 * a vast improvement over the zvol_request_sync=1 behavior
624 		 * of one i/o at a time per zvol.  However, an even better
625 		 * design would be for zvol_request() to initiate the zio
626 		 * directly, and then be notified by the zio_done callback,
627 		 * which would call END_IO().  Unfortunately, the DMU/ZIL
628 		 * interfaces lack this functionality (they block waiting for
629 		 * the i/o to complete).
630 		 */
631 		if (io_is_discard(bio, rq) || io_is_secure_erase(bio, rq)) {
632 			if (force_sync) {
633 				zvol_discard(&zvr);
634 			} else {
635 				task = zv_request_task_create(zvr);
636 				taskq_dispatch_ent(ztqs->tqs_taskq[tq_idx],
637 				    zvol_discard_task, task, 0, &task->ent);
638 			}
639 		} else {
640 			if (force_sync) {
641 				zvol_write(&zvr);
642 			} else {
643 				task = zv_request_task_create(zvr);
644 				taskq_dispatch_ent(ztqs->tqs_taskq[tq_idx],
645 				    zvol_write_task, task, 0, &task->ent);
646 			}
647 		}
648 	} else {
649 		/*
650 		 * The SCST driver, and possibly others, may issue READ I/Os
651 		 * with a length of zero bytes.  These empty I/Os contain no
652 		 * data and require no additional handling.
653 		 */
654 		if (size == 0) {
655 			zvol_end_io(bio, rq, 0);
656 			goto out;
657 		}
658 
659 		rw_enter(&zv->zv_suspend_lock, RW_READER);
660 
661 		/* See comment in WRITE case above. */
662 		if (force_sync) {
663 			zvol_read(&zvr);
664 		} else {
665 			task = zv_request_task_create(zvr);
666 			taskq_dispatch_ent(ztqs->tqs_taskq[tq_idx],
667 			    zvol_read_task, task, 0, &task->ent);
668 		}
669 	}
670 
671 out:
672 	spl_fstrans_unmark(cookie);
673 }
674 
675 #ifdef HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS
676 #ifdef HAVE_BDEV_SUBMIT_BIO_RETURNS_VOID
677 static void
zvol_submit_bio(struct bio * bio)678 zvol_submit_bio(struct bio *bio)
679 #else
680 static blk_qc_t
681 zvol_submit_bio(struct bio *bio)
682 #endif
683 #else
684 static MAKE_REQUEST_FN_RET
685 zvol_request(struct request_queue *q, struct bio *bio)
686 #endif
687 {
688 #ifdef HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS
689 #if defined(HAVE_BIO_BDEV_DISK)
690 	struct request_queue *q = bio->bi_bdev->bd_disk->queue;
691 #else
692 	struct request_queue *q = bio->bi_disk->queue;
693 #endif
694 #endif
695 	zvol_state_t *zv = q->queuedata;
696 
697 	zvol_request_impl(zv, bio, NULL, 0);
698 #if defined(HAVE_MAKE_REQUEST_FN_RET_QC) || \
699 	defined(HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS) && \
700 	!defined(HAVE_BDEV_SUBMIT_BIO_RETURNS_VOID)
701 	return (BLK_QC_T_NONE);
702 #endif
703 }
704 
705 static int
706 #ifdef HAVE_BLK_MODE_T
zvol_open(struct gendisk * disk,blk_mode_t flag)707 zvol_open(struct gendisk *disk, blk_mode_t flag)
708 #else
709 zvol_open(struct block_device *bdev, fmode_t flag)
710 #endif
711 {
712 	zvol_state_t *zv;
713 	int error = 0;
714 	boolean_t drop_suspend = B_FALSE;
715 #ifndef HAVE_BLKDEV_GET_ERESTARTSYS
716 	hrtime_t timeout = MSEC2NSEC(zvol_open_timeout_ms);
717 	hrtime_t start = gethrtime();
718 
719 retry:
720 #endif
721 	rw_enter(&zvol_state_lock, RW_READER);
722 	/*
723 	 * Obtain a copy of private_data under the zvol_state_lock to make
724 	 * sure that either the result of zvol free code path setting
725 	 * disk->private_data to NULL is observed, or zvol_os_free()
726 	 * is not called on this zv because of the positive zv_open_count.
727 	 */
728 #ifdef HAVE_BLK_MODE_T
729 	zv = disk->private_data;
730 #else
731 	zv = bdev->bd_disk->private_data;
732 #endif
733 	if (zv == NULL) {
734 		rw_exit(&zvol_state_lock);
735 		return (-SET_ERROR(ENXIO));
736 	}
737 
738 	mutex_enter(&zv->zv_state_lock);
739 
740 	if (unlikely(zv->zv_flags & ZVOL_REMOVING)) {
741 		mutex_exit(&zv->zv_state_lock);
742 		rw_exit(&zvol_state_lock);
743 		return (-SET_ERROR(ENXIO));
744 	}
745 
746 	/*
747 	 * Make sure zvol is not suspended during first open
748 	 * (hold zv_suspend_lock) and respect proper lock acquisition
749 	 * ordering - zv_suspend_lock before zv_state_lock
750 	 */
751 	if (zv->zv_open_count == 0) {
752 		if (!rw_tryenter(&zv->zv_suspend_lock, RW_READER)) {
753 			mutex_exit(&zv->zv_state_lock);
754 			rw_enter(&zv->zv_suspend_lock, RW_READER);
755 			mutex_enter(&zv->zv_state_lock);
756 			/* check to see if zv_suspend_lock is needed */
757 			if (zv->zv_open_count != 0) {
758 				rw_exit(&zv->zv_suspend_lock);
759 			} else {
760 				drop_suspend = B_TRUE;
761 			}
762 		} else {
763 			drop_suspend = B_TRUE;
764 		}
765 	}
766 	rw_exit(&zvol_state_lock);
767 
768 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
769 
770 	if (zv->zv_open_count == 0) {
771 		boolean_t drop_namespace = B_FALSE;
772 
773 		ASSERT(RW_READ_HELD(&zv->zv_suspend_lock));
774 
775 		/*
776 		 * In all other call paths the spa_namespace_lock is taken
777 		 * before the bdev->bd_mutex lock.  However, on open(2)
778 		 * the __blkdev_get() function calls fops->open() with the
779 		 * bdev->bd_mutex lock held.  This can result in a deadlock
780 		 * when zvols from one pool are used as vdevs in another.
781 		 *
782 		 * To prevent a lock inversion deadlock we preemptively
783 		 * take the spa_namespace_lock.  Normally the lock will not
784 		 * be contended and this is safe because spa_open_common()
785 		 * handles the case where the caller already holds the
786 		 * spa_namespace_lock.
787 		 *
788 		 * When the lock cannot be aquired after multiple retries
789 		 * this must be the vdev on zvol deadlock case and we have
790 		 * no choice but to return an error.  For 5.12 and older
791 		 * kernels returning -ERESTARTSYS will result in the
792 		 * bdev->bd_mutex being dropped, then reacquired, and
793 		 * fops->open() being called again.  This process can be
794 		 * repeated safely until both locks are acquired.  For 5.13
795 		 * and newer the -ERESTARTSYS retry logic was removed from
796 		 * the kernel so the only option is to return the error for
797 		 * the caller to handle it.
798 		 */
799 		if (!mutex_owned(&spa_namespace_lock)) {
800 			if (!mutex_tryenter(&spa_namespace_lock)) {
801 				mutex_exit(&zv->zv_state_lock);
802 				rw_exit(&zv->zv_suspend_lock);
803 				drop_suspend = B_FALSE;
804 
805 #ifdef HAVE_BLKDEV_GET_ERESTARTSYS
806 				schedule();
807 				return (-SET_ERROR(ERESTARTSYS));
808 #else
809 				if ((gethrtime() - start) > timeout)
810 					return (-SET_ERROR(ERESTARTSYS));
811 
812 				schedule_timeout_interruptible(
813 					MSEC_TO_TICK(10));
814 				goto retry;
815 #endif
816 			} else {
817 				drop_namespace = B_TRUE;
818 			}
819 		}
820 
821 		error = -zvol_first_open(zv, !(blk_mode_is_open_write(flag)));
822 
823 		if (drop_namespace)
824 			mutex_exit(&spa_namespace_lock);
825 	}
826 
827 	if (error == 0) {
828 		if ((blk_mode_is_open_write(flag)) &&
829 		    (zv->zv_flags & ZVOL_RDONLY)) {
830 			if (zv->zv_open_count == 0)
831 				zvol_last_close(zv);
832 
833 			error = -SET_ERROR(EROFS);
834 		} else {
835 			zv->zv_open_count++;
836 		}
837 	}
838 
839 	mutex_exit(&zv->zv_state_lock);
840 	if (drop_suspend)
841 		rw_exit(&zv->zv_suspend_lock);
842 
843 	if (error == 0)
844 #ifdef HAVE_BLK_MODE_T
845 		disk_check_media_change(disk);
846 #else
847 		zfs_check_media_change(bdev);
848 #endif
849 
850 	return (error);
851 }
852 
853 static void
854 #ifdef HAVE_BLOCK_DEVICE_OPERATIONS_RELEASE_1ARG
zvol_release(struct gendisk * disk)855 zvol_release(struct gendisk *disk)
856 #else
857 zvol_release(struct gendisk *disk, fmode_t unused)
858 #endif
859 {
860 #if !defined(HAVE_BLOCK_DEVICE_OPERATIONS_RELEASE_1ARG)
861 	(void) unused;
862 #endif
863 	zvol_state_t *zv;
864 	boolean_t drop_suspend = B_TRUE;
865 
866 	rw_enter(&zvol_state_lock, RW_READER);
867 	zv = disk->private_data;
868 
869 	mutex_enter(&zv->zv_state_lock);
870 	ASSERT3U(zv->zv_open_count, >, 0);
871 	/*
872 	 * make sure zvol is not suspended during last close
873 	 * (hold zv_suspend_lock) and respect proper lock acquisition
874 	 * ordering - zv_suspend_lock before zv_state_lock
875 	 */
876 	if (zv->zv_open_count == 1) {
877 		if (!rw_tryenter(&zv->zv_suspend_lock, RW_READER)) {
878 			mutex_exit(&zv->zv_state_lock);
879 			rw_enter(&zv->zv_suspend_lock, RW_READER);
880 			mutex_enter(&zv->zv_state_lock);
881 			/* check to see if zv_suspend_lock is needed */
882 			if (zv->zv_open_count != 1) {
883 				rw_exit(&zv->zv_suspend_lock);
884 				drop_suspend = B_FALSE;
885 			}
886 		}
887 	} else {
888 		drop_suspend = B_FALSE;
889 	}
890 	rw_exit(&zvol_state_lock);
891 
892 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
893 
894 	zv->zv_open_count--;
895 	if (zv->zv_open_count == 0) {
896 		ASSERT(RW_READ_HELD(&zv->zv_suspend_lock));
897 		zvol_last_close(zv);
898 	}
899 
900 	mutex_exit(&zv->zv_state_lock);
901 
902 	if (drop_suspend)
903 		rw_exit(&zv->zv_suspend_lock);
904 }
905 
906 static int
zvol_ioctl(struct block_device * bdev,fmode_t mode,unsigned int cmd,unsigned long arg)907 zvol_ioctl(struct block_device *bdev, fmode_t mode,
908     unsigned int cmd, unsigned long arg)
909 {
910 	zvol_state_t *zv = bdev->bd_disk->private_data;
911 	int error = 0;
912 
913 	ASSERT3U(zv->zv_open_count, >, 0);
914 
915 	switch (cmd) {
916 	case BLKFLSBUF:
917 #ifdef HAVE_FSYNC_BDEV
918 		fsync_bdev(bdev);
919 #elif defined(HAVE_SYNC_BLOCKDEV)
920 		sync_blockdev(bdev);
921 #else
922 #error "Neither fsync_bdev() nor sync_blockdev() found"
923 #endif
924 		invalidate_bdev(bdev);
925 		rw_enter(&zv->zv_suspend_lock, RW_READER);
926 
927 		if (!(zv->zv_flags & ZVOL_RDONLY))
928 			txg_wait_synced(dmu_objset_pool(zv->zv_objset), 0);
929 
930 		rw_exit(&zv->zv_suspend_lock);
931 		break;
932 
933 	case BLKZNAME:
934 		mutex_enter(&zv->zv_state_lock);
935 		error = copy_to_user((void *)arg, zv->zv_name, MAXNAMELEN);
936 		mutex_exit(&zv->zv_state_lock);
937 		break;
938 
939 	default:
940 		error = -ENOTTY;
941 		break;
942 	}
943 
944 	return (SET_ERROR(error));
945 }
946 
947 #ifdef CONFIG_COMPAT
948 static int
zvol_compat_ioctl(struct block_device * bdev,fmode_t mode,unsigned cmd,unsigned long arg)949 zvol_compat_ioctl(struct block_device *bdev, fmode_t mode,
950     unsigned cmd, unsigned long arg)
951 {
952 	return (zvol_ioctl(bdev, mode, cmd, arg));
953 }
954 #else
955 #define	zvol_compat_ioctl	NULL
956 #endif
957 
958 static unsigned int
zvol_check_events(struct gendisk * disk,unsigned int clearing)959 zvol_check_events(struct gendisk *disk, unsigned int clearing)
960 {
961 	unsigned int mask = 0;
962 
963 	rw_enter(&zvol_state_lock, RW_READER);
964 
965 	zvol_state_t *zv = disk->private_data;
966 	if (zv != NULL) {
967 		mutex_enter(&zv->zv_state_lock);
968 		mask = zv->zv_changed ? DISK_EVENT_MEDIA_CHANGE : 0;
969 		zv->zv_changed = 0;
970 		mutex_exit(&zv->zv_state_lock);
971 	}
972 
973 	rw_exit(&zvol_state_lock);
974 
975 	return (mask);
976 }
977 
978 static int
zvol_revalidate_disk(struct gendisk * disk)979 zvol_revalidate_disk(struct gendisk *disk)
980 {
981 	rw_enter(&zvol_state_lock, RW_READER);
982 
983 	zvol_state_t *zv = disk->private_data;
984 	if (zv != NULL) {
985 		mutex_enter(&zv->zv_state_lock);
986 		set_capacity(zv->zv_zso->zvo_disk,
987 		    zv->zv_volsize >> SECTOR_BITS);
988 		mutex_exit(&zv->zv_state_lock);
989 	}
990 
991 	rw_exit(&zvol_state_lock);
992 
993 	return (0);
994 }
995 
996 int
zvol_os_update_volsize(zvol_state_t * zv,uint64_t volsize)997 zvol_os_update_volsize(zvol_state_t *zv, uint64_t volsize)
998 {
999 	struct gendisk *disk = zv->zv_zso->zvo_disk;
1000 
1001 #if defined(HAVE_REVALIDATE_DISK_SIZE)
1002 	revalidate_disk_size(disk, zvol_revalidate_disk(disk) == 0);
1003 #elif defined(HAVE_REVALIDATE_DISK)
1004 	revalidate_disk(disk);
1005 #else
1006 	zvol_revalidate_disk(disk);
1007 #endif
1008 	return (0);
1009 }
1010 
1011 void
zvol_os_clear_private(zvol_state_t * zv)1012 zvol_os_clear_private(zvol_state_t *zv)
1013 {
1014 	/*
1015 	 * Cleared while holding zvol_state_lock as a writer
1016 	 * which will prevent zvol_open() from opening it.
1017 	 */
1018 	zv->zv_zso->zvo_disk->private_data = NULL;
1019 }
1020 
1021 /*
1022  * Provide a simple virtual geometry for legacy compatibility.  For devices
1023  * smaller than 1 MiB a small head and sector count is used to allow very
1024  * tiny devices.  For devices over 1 Mib a standard head and sector count
1025  * is used to keep the cylinders count reasonable.
1026  */
1027 static int
zvol_getgeo(struct block_device * bdev,struct hd_geometry * geo)1028 zvol_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1029 {
1030 	zvol_state_t *zv = bdev->bd_disk->private_data;
1031 	sector_t sectors;
1032 
1033 	ASSERT3U(zv->zv_open_count, >, 0);
1034 
1035 	sectors = get_capacity(zv->zv_zso->zvo_disk);
1036 
1037 	if (sectors > 2048) {
1038 		geo->heads = 16;
1039 		geo->sectors = 63;
1040 	} else {
1041 		geo->heads = 2;
1042 		geo->sectors = 4;
1043 	}
1044 
1045 	geo->start = 0;
1046 	geo->cylinders = sectors / (geo->heads * geo->sectors);
1047 
1048 	return (0);
1049 }
1050 
1051 /*
1052  * Why have two separate block_device_operations structs?
1053  *
1054  * Normally we'd just have one, and assign 'submit_bio' as needed.  However,
1055  * it's possible the user's kernel is built with CONSTIFY_PLUGIN, meaning we
1056  * can't just change submit_bio dynamically at runtime.  So just create two
1057  * separate structs to get around this.
1058  */
1059 static const struct block_device_operations zvol_ops_blk_mq = {
1060 	.open			= zvol_open,
1061 	.release		= zvol_release,
1062 	.ioctl			= zvol_ioctl,
1063 	.compat_ioctl		= zvol_compat_ioctl,
1064 	.check_events		= zvol_check_events,
1065 #ifdef HAVE_BLOCK_DEVICE_OPERATIONS_REVALIDATE_DISK
1066 	.revalidate_disk	= zvol_revalidate_disk,
1067 #endif
1068 	.getgeo			= zvol_getgeo,
1069 	.owner			= THIS_MODULE,
1070 };
1071 
1072 static const struct block_device_operations zvol_ops = {
1073 	.open			= zvol_open,
1074 	.release		= zvol_release,
1075 	.ioctl			= zvol_ioctl,
1076 	.compat_ioctl		= zvol_compat_ioctl,
1077 	.check_events		= zvol_check_events,
1078 #ifdef HAVE_BLOCK_DEVICE_OPERATIONS_REVALIDATE_DISK
1079 	.revalidate_disk	= zvol_revalidate_disk,
1080 #endif
1081 	.getgeo			= zvol_getgeo,
1082 	.owner			= THIS_MODULE,
1083 #ifdef HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS
1084 	.submit_bio		= zvol_submit_bio,
1085 #endif
1086 };
1087 
1088 /*
1089  * Since 6.9, Linux has been removing queue limit setters in favour of an
1090  * initial queue_limits struct applied when the device is open. Since 6.11,
1091  * queue_limits is being extended to allow more things to be applied when the
1092  * device is open. Setters are also being removed for this.
1093  *
1094  * For OpenZFS, this means that depending on kernel version, some options may
1095  * be set up before the device is open, and some applied to an open device
1096  * (queue) after the fact.
1097  *
1098  * We manage this complexity by having our own limits struct,
1099  * zvol_queue_limits_t, in which we carry any queue config that we're
1100  * interested in setting. This structure is the same on all kernels.
1101  *
1102  * These limits are then applied to the queue at device open time by the most
1103  * appropriate method for the kernel.
1104  *
1105  * zvol_queue_limits_convert() is used on 6.9+ (where the two-arg form of
1106  * blk_alloc_disk() exists). This converts our limits struct to a proper Linux
1107  * struct queue_limits, and passes it in. Any fields added in later kernels are
1108  * (obviously) not set up here.
1109  *
1110  * zvol_queue_limits_apply() is called on all kernel versions after the queue
1111  * is created, and applies any remaining config. Before 6.9 that will be
1112  * everything, via setter methods. After 6.9 that will be whatever couldn't be
1113  * put into struct queue_limits. (This implies that zvol_queue_limits_apply()
1114  * will always be a no-op on the latest kernel we support).
1115  */
1116 typedef struct zvol_queue_limits {
1117 	unsigned int	zql_max_hw_sectors;
1118 	unsigned short	zql_max_segments;
1119 	unsigned int	zql_max_segment_size;
1120 	unsigned int	zql_io_opt;
1121 	unsigned int	zql_physical_block_size;
1122 	unsigned int	zql_max_discard_sectors;
1123 	unsigned int	zql_discard_granularity;
1124 } zvol_queue_limits_t;
1125 
1126 static void
zvol_queue_limits_init(zvol_queue_limits_t * limits,zvol_state_t * zv,boolean_t use_blk_mq)1127 zvol_queue_limits_init(zvol_queue_limits_t *limits, zvol_state_t *zv,
1128     boolean_t use_blk_mq)
1129 {
1130 	limits->zql_max_hw_sectors = (DMU_MAX_ACCESS / 4) >> 9;
1131 
1132 	if (use_blk_mq) {
1133 		/*
1134 		 * IO requests can be really big (1MB).  When an IO request
1135 		 * comes in, it is passed off to zvol_read() or zvol_write()
1136 		 * in a new thread, where it is chunked up into 'volblocksize'
1137 		 * sized pieces and processed.  So for example, if the request
1138 		 * is a 1MB write and your volblocksize is 128k, one zvol_write
1139 		 * thread will take that request and sequentially do ten 128k
1140 		 * IOs.  This is due to the fact that the thread needs to lock
1141 		 * each volblocksize sized block.  So you might be wondering:
1142 		 * "instead of passing the whole 1MB request to one thread,
1143 		 * why not pass ten individual 128k chunks to ten threads and
1144 		 * process the whole write in parallel?"  The short answer is
1145 		 * that there's a sweet spot number of chunks that balances
1146 		 * the greater parallelism with the added overhead of more
1147 		 * threads. The sweet spot can be different depending on if you
1148 		 * have a read or write  heavy workload.  Writes typically want
1149 		 * high chunk counts while reads typically want lower ones.  On
1150 		 * a test pool with 6 NVMe drives in a 3x 2-disk mirror
1151 		 * configuration, with volblocksize=8k, the sweet spot for good
1152 		 * sequential reads and writes was at 8 chunks.
1153 		 */
1154 
1155 		/*
1156 		 * Below we tell the kernel how big we want our requests
1157 		 * to be.  You would think that blk_queue_io_opt() would be
1158 		 * used to do this since it is used to "set optimal request
1159 		 * size for the queue", but that doesn't seem to do
1160 		 * anything - the kernel still gives you huge requests
1161 		 * with tons of little PAGE_SIZE segments contained within it.
1162 		 *
1163 		 * Knowing that the kernel will just give you PAGE_SIZE segments
1164 		 * no matter what, you can say "ok, I want PAGE_SIZE byte
1165 		 * segments, and I want 'N' of them per request", where N is
1166 		 * the correct number of segments for the volblocksize and
1167 		 * number of chunks you want.
1168 		 */
1169 		if (zvol_blk_mq_blocks_per_thread != 0) {
1170 			unsigned int chunks;
1171 			chunks = MIN(zvol_blk_mq_blocks_per_thread, UINT16_MAX);
1172 
1173 			limits->zql_max_segment_size = PAGE_SIZE;
1174 			limits->zql_max_segments =
1175 			    (zv->zv_volblocksize * chunks) / PAGE_SIZE;
1176 		} else {
1177 			/*
1178 			 * Special case: zvol_blk_mq_blocks_per_thread = 0
1179 			 * Max everything out.
1180 			 */
1181 			limits->zql_max_segments = UINT16_MAX;
1182 			limits->zql_max_segment_size = UINT_MAX;
1183 		}
1184 	} else {
1185 		limits->zql_max_segments = UINT16_MAX;
1186 		limits->zql_max_segment_size = UINT_MAX;
1187 	}
1188 
1189 	limits->zql_io_opt = DMU_MAX_ACCESS / 2;
1190 
1191 	limits->zql_physical_block_size = zv->zv_volblocksize;
1192 	limits->zql_max_discard_sectors =
1193 	    (zvol_max_discard_blocks * zv->zv_volblocksize) >> 9;
1194 	limits->zql_discard_granularity = zv->zv_volblocksize;
1195 }
1196 
1197 #ifdef HAVE_BLK_ALLOC_DISK_2ARG
1198 static void
zvol_queue_limits_convert(zvol_queue_limits_t * limits,struct queue_limits * qlimits)1199 zvol_queue_limits_convert(zvol_queue_limits_t *limits,
1200     struct queue_limits *qlimits)
1201 {
1202 	memset(qlimits, 0, sizeof (struct queue_limits));
1203 	qlimits->max_hw_sectors = limits->zql_max_hw_sectors;
1204 	qlimits->max_segments = limits->zql_max_segments;
1205 	qlimits->max_segment_size = limits->zql_max_segment_size;
1206 	qlimits->io_opt = limits->zql_io_opt;
1207 	qlimits->physical_block_size = limits->zql_physical_block_size;
1208 	qlimits->max_discard_sectors = limits->zql_max_discard_sectors;
1209 	qlimits->max_hw_discard_sectors = limits->zql_max_discard_sectors;
1210 	qlimits->discard_granularity = limits->zql_discard_granularity;
1211 #ifdef HAVE_BLKDEV_QUEUE_LIMITS_FEATURES
1212 	qlimits->features =
1213 	    BLK_FEAT_WRITE_CACHE | BLK_FEAT_FUA | BLK_FEAT_IO_STAT;
1214 #endif
1215 }
1216 #endif
1217 
1218 static void
zvol_queue_limits_apply(zvol_queue_limits_t * limits,struct request_queue * queue)1219 zvol_queue_limits_apply(zvol_queue_limits_t *limits,
1220     struct request_queue *queue)
1221 {
1222 #ifndef HAVE_BLK_ALLOC_DISK_2ARG
1223 	blk_queue_max_hw_sectors(queue, limits->zql_max_hw_sectors);
1224 	blk_queue_max_segments(queue, limits->zql_max_segments);
1225 	blk_queue_max_segment_size(queue, limits->zql_max_segment_size);
1226 	blk_queue_io_opt(queue, limits->zql_io_opt);
1227 	blk_queue_physical_block_size(queue, limits->zql_physical_block_size);
1228 	blk_queue_max_discard_sectors(queue, limits->zql_max_discard_sectors);
1229 	blk_queue_discard_granularity(queue, limits->zql_discard_granularity);
1230 #endif
1231 #ifndef HAVE_BLKDEV_QUEUE_LIMITS_FEATURES
1232 	blk_queue_set_write_cache(queue, B_TRUE);
1233 	blk_queue_flag_set(QUEUE_FLAG_IO_STAT, queue);
1234 #endif
1235 }
1236 
1237 static int
zvol_alloc_non_blk_mq(struct zvol_state_os * zso,zvol_queue_limits_t * limits)1238 zvol_alloc_non_blk_mq(struct zvol_state_os *zso, zvol_queue_limits_t *limits)
1239 {
1240 #if defined(HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS)
1241 #if defined(HAVE_BLK_ALLOC_DISK)
1242 	zso->zvo_disk = blk_alloc_disk(NUMA_NO_NODE);
1243 	if (zso->zvo_disk == NULL)
1244 		return (1);
1245 
1246 	zso->zvo_disk->minors = ZVOL_MINORS;
1247 	zso->zvo_queue = zso->zvo_disk->queue;
1248 #elif defined(HAVE_BLK_ALLOC_DISK_2ARG)
1249 	struct queue_limits qlimits;
1250 	zvol_queue_limits_convert(limits, &qlimits);
1251 	struct gendisk *disk = blk_alloc_disk(&qlimits, NUMA_NO_NODE);
1252 	if (IS_ERR(disk)) {
1253 		zso->zvo_disk = NULL;
1254 		return (1);
1255 	}
1256 
1257 	zso->zvo_disk = disk;
1258 	zso->zvo_disk->minors = ZVOL_MINORS;
1259 	zso->zvo_queue = zso->zvo_disk->queue;
1260 
1261 #else
1262 	zso->zvo_queue = blk_alloc_queue(NUMA_NO_NODE);
1263 	if (zso->zvo_queue == NULL)
1264 		return (1);
1265 
1266 	zso->zvo_disk = alloc_disk(ZVOL_MINORS);
1267 	if (zso->zvo_disk == NULL) {
1268 		blk_cleanup_queue(zso->zvo_queue);
1269 		return (1);
1270 	}
1271 
1272 	zso->zvo_disk->queue = zso->zvo_queue;
1273 #endif /* HAVE_BLK_ALLOC_DISK */
1274 #else
1275 	zso->zvo_queue = blk_generic_alloc_queue(zvol_request, NUMA_NO_NODE);
1276 	if (zso->zvo_queue == NULL)
1277 		return (1);
1278 
1279 	zso->zvo_disk = alloc_disk(ZVOL_MINORS);
1280 	if (zso->zvo_disk == NULL) {
1281 		blk_cleanup_queue(zso->zvo_queue);
1282 		return (1);
1283 	}
1284 
1285 	zso->zvo_disk->queue = zso->zvo_queue;
1286 #endif /* HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS */
1287 
1288 	zvol_queue_limits_apply(limits, zso->zvo_queue);
1289 
1290 	return (0);
1291 
1292 }
1293 
1294 static int
zvol_alloc_blk_mq(zvol_state_t * zv,zvol_queue_limits_t * limits)1295 zvol_alloc_blk_mq(zvol_state_t *zv, zvol_queue_limits_t *limits)
1296 {
1297 	struct zvol_state_os *zso = zv->zv_zso;
1298 
1299 	/* Allocate our blk-mq tag_set */
1300 	if (zvol_blk_mq_alloc_tag_set(zv) != 0)
1301 		return (1);
1302 
1303 #if defined(HAVE_BLK_ALLOC_DISK)
1304 	zso->zvo_disk = blk_mq_alloc_disk(&zso->tag_set, zv);
1305 	if (zso->zvo_disk == NULL) {
1306 		blk_mq_free_tag_set(&zso->tag_set);
1307 		return (1);
1308 	}
1309 	zso->zvo_queue = zso->zvo_disk->queue;
1310 	zso->zvo_disk->minors = ZVOL_MINORS;
1311 #elif defined(HAVE_BLK_ALLOC_DISK_2ARG)
1312 	struct queue_limits qlimits;
1313 	zvol_queue_limits_convert(limits, &qlimits);
1314 	struct gendisk *disk = blk_mq_alloc_disk(&zso->tag_set, &qlimits, zv);
1315 	if (IS_ERR(disk)) {
1316 		zso->zvo_disk = NULL;
1317 		blk_mq_free_tag_set(&zso->tag_set);
1318 		return (1);
1319 	}
1320 
1321 	zso->zvo_disk = disk;
1322 	zso->zvo_queue = zso->zvo_disk->queue;
1323 	zso->zvo_disk->minors = ZVOL_MINORS;
1324 #else
1325 	zso->zvo_disk = alloc_disk(ZVOL_MINORS);
1326 	if (zso->zvo_disk == NULL) {
1327 		blk_cleanup_queue(zso->zvo_queue);
1328 		blk_mq_free_tag_set(&zso->tag_set);
1329 		return (1);
1330 	}
1331 	/* Allocate queue */
1332 	zso->zvo_queue = blk_mq_init_queue(&zso->tag_set);
1333 	if (IS_ERR(zso->zvo_queue)) {
1334 		blk_mq_free_tag_set(&zso->tag_set);
1335 		return (1);
1336 	}
1337 
1338 	/* Our queue is now created, assign it to our disk */
1339 	zso->zvo_disk->queue = zso->zvo_queue;
1340 #endif
1341 
1342 	zvol_queue_limits_apply(limits, zso->zvo_queue);
1343 
1344 	return (0);
1345 }
1346 
1347 /*
1348  * Allocate memory for a new zvol_state_t and setup the required
1349  * request queue and generic disk structures for the block device.
1350  */
1351 static zvol_state_t *
zvol_alloc(dev_t dev,const char * name,uint64_t volblocksize)1352 zvol_alloc(dev_t dev, const char *name, uint64_t volblocksize)
1353 {
1354 	zvol_state_t *zv;
1355 	struct zvol_state_os *zso;
1356 	uint64_t volmode;
1357 	int ret;
1358 
1359 	if (dsl_prop_get_integer(name, "volmode", &volmode, NULL) != 0)
1360 		return (NULL);
1361 
1362 	if (volmode == ZFS_VOLMODE_DEFAULT)
1363 		volmode = zvol_volmode;
1364 
1365 	if (volmode == ZFS_VOLMODE_NONE)
1366 		return (NULL);
1367 
1368 	zv = kmem_zalloc(sizeof (zvol_state_t), KM_SLEEP);
1369 	zso = kmem_zalloc(sizeof (struct zvol_state_os), KM_SLEEP);
1370 	zv->zv_zso = zso;
1371 	zv->zv_volmode = volmode;
1372 	zv->zv_volblocksize = volblocksize;
1373 
1374 	list_link_init(&zv->zv_next);
1375 	mutex_init(&zv->zv_state_lock, NULL, MUTEX_DEFAULT, NULL);
1376 	cv_init(&zv->zv_removing_cv, NULL, CV_DEFAULT, NULL);
1377 
1378 	zv->zv_zso->use_blk_mq = zvol_use_blk_mq;
1379 
1380 	zvol_queue_limits_t limits;
1381 	zvol_queue_limits_init(&limits, zv, zv->zv_zso->use_blk_mq);
1382 
1383 	/*
1384 	 * The block layer has 3 interfaces for getting BIOs:
1385 	 *
1386 	 * 1. blk-mq request queues (new)
1387 	 * 2. submit_bio() (oldest)
1388 	 * 3. regular request queues (old).
1389 	 *
1390 	 * Each of those interfaces has two permutations:
1391 	 *
1392 	 * a) We have blk_alloc_disk()/blk_mq_alloc_disk(), which allocates
1393 	 *    both the disk and its queue (5.14 kernel or newer)
1394 	 *
1395 	 * b) We don't have blk_*alloc_disk(), and have to allocate the
1396 	 *    disk and the queue separately. (5.13 kernel or older)
1397 	 */
1398 	if (zv->zv_zso->use_blk_mq) {
1399 		ret = zvol_alloc_blk_mq(zv, &limits);
1400 		zso->zvo_disk->fops = &zvol_ops_blk_mq;
1401 	} else {
1402 		ret = zvol_alloc_non_blk_mq(zso, &limits);
1403 		zso->zvo_disk->fops = &zvol_ops;
1404 	}
1405 	if (ret != 0)
1406 		goto out_kmem;
1407 
1408 	/* Limit read-ahead to a single page to prevent over-prefetching. */
1409 	blk_queue_set_read_ahead(zso->zvo_queue, 1);
1410 
1411 	if (!zv->zv_zso->use_blk_mq) {
1412 		/* Disable write merging in favor of the ZIO pipeline. */
1413 		blk_queue_flag_set(QUEUE_FLAG_NOMERGES, zso->zvo_queue);
1414 	}
1415 
1416 	zso->zvo_queue->queuedata = zv;
1417 	zso->zvo_dev = dev;
1418 	zv->zv_open_count = 0;
1419 	strlcpy(zv->zv_name, name, sizeof (zv->zv_name));
1420 
1421 	zfs_rangelock_init(&zv->zv_rangelock, NULL, NULL);
1422 	rw_init(&zv->zv_suspend_lock, NULL, RW_DEFAULT, NULL);
1423 
1424 	zso->zvo_disk->major = zvol_major;
1425 	zso->zvo_disk->events = DISK_EVENT_MEDIA_CHANGE;
1426 
1427 	/*
1428 	 * Setting ZFS_VOLMODE_DEV disables partitioning on ZVOL devices.
1429 	 * This is accomplished by limiting the number of minors for the
1430 	 * device to one and explicitly disabling partition scanning.
1431 	 */
1432 	if (volmode == ZFS_VOLMODE_DEV) {
1433 		zso->zvo_disk->minors = 1;
1434 		zso->zvo_disk->flags &= ~GENHD_FL_EXT_DEVT;
1435 		zso->zvo_disk->flags |= GENHD_FL_NO_PART;
1436 	}
1437 
1438 	zso->zvo_disk->first_minor = (dev & MINORMASK);
1439 	zso->zvo_disk->private_data = zv;
1440 	snprintf(zso->zvo_disk->disk_name, DISK_NAME_LEN, "%s%d",
1441 	    ZVOL_DEV_NAME, (dev & MINORMASK));
1442 
1443 	return (zv);
1444 
1445 out_kmem:
1446 	kmem_free(zso, sizeof (struct zvol_state_os));
1447 	kmem_free(zv, sizeof (zvol_state_t));
1448 	return (NULL);
1449 }
1450 
1451 /*
1452  * Cleanup then free a zvol_state_t which was created by zvol_alloc().
1453  * At this time, the structure is not opened by anyone, is taken off
1454  * the zvol_state_list, and has its private data set to NULL.
1455  * The zvol_state_lock is dropped.
1456  *
1457  * This function may take many milliseconds to complete (e.g. we've seen
1458  * it take over 256ms), due to the calls to "blk_cleanup_queue" and
1459  * "del_gendisk". Thus, consumers need to be careful to account for this
1460  * latency when calling this function.
1461  */
1462 void
zvol_os_free(zvol_state_t * zv)1463 zvol_os_free(zvol_state_t *zv)
1464 {
1465 
1466 	ASSERT(!RW_LOCK_HELD(&zv->zv_suspend_lock));
1467 	ASSERT(!MUTEX_HELD(&zv->zv_state_lock));
1468 	ASSERT0(zv->zv_open_count);
1469 	ASSERT3P(zv->zv_zso->zvo_disk->private_data, ==, NULL);
1470 
1471 	rw_destroy(&zv->zv_suspend_lock);
1472 	zfs_rangelock_fini(&zv->zv_rangelock);
1473 
1474 	del_gendisk(zv->zv_zso->zvo_disk);
1475 #if defined(HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS) && \
1476 	(defined(HAVE_BLK_ALLOC_DISK) || defined(HAVE_BLK_ALLOC_DISK_2ARG))
1477 #if defined(HAVE_BLK_CLEANUP_DISK)
1478 	blk_cleanup_disk(zv->zv_zso->zvo_disk);
1479 #else
1480 	put_disk(zv->zv_zso->zvo_disk);
1481 #endif
1482 #else
1483 	blk_cleanup_queue(zv->zv_zso->zvo_queue);
1484 	put_disk(zv->zv_zso->zvo_disk);
1485 #endif
1486 
1487 	if (zv->zv_zso->use_blk_mq)
1488 		blk_mq_free_tag_set(&zv->zv_zso->tag_set);
1489 
1490 	ida_simple_remove(&zvol_ida,
1491 	    MINOR(zv->zv_zso->zvo_dev) >> ZVOL_MINOR_BITS);
1492 
1493 	cv_destroy(&zv->zv_removing_cv);
1494 	mutex_destroy(&zv->zv_state_lock);
1495 	dataset_kstats_destroy(&zv->zv_kstat);
1496 
1497 	kmem_free(zv->zv_zso, sizeof (struct zvol_state_os));
1498 	kmem_free(zv, sizeof (zvol_state_t));
1499 }
1500 
1501 void
zvol_wait_close(zvol_state_t * zv)1502 zvol_wait_close(zvol_state_t *zv)
1503 {
1504 }
1505 
1506 struct add_disk_work {
1507 	struct delayed_work work;
1508 	struct gendisk *disk;
1509 	int error;
1510 };
1511 
1512 static int
__zvol_os_add_disk(struct gendisk * disk)1513 __zvol_os_add_disk(struct gendisk *disk)
1514 {
1515 	int error = 0;
1516 #ifdef HAVE_ADD_DISK_RET
1517 	error = add_disk(disk);
1518 #else
1519 	add_disk(disk);
1520 #endif
1521 	return (error);
1522 }
1523 
1524 #if defined(HAVE_BDEV_FILE_OPEN_BY_PATH)
1525 static void
zvol_os_add_disk_work(struct work_struct * work)1526 zvol_os_add_disk_work(struct work_struct *work)
1527 {
1528 	struct add_disk_work *add_disk_work;
1529 	add_disk_work = container_of(work, struct add_disk_work, work.work);
1530 	add_disk_work->error = __zvol_os_add_disk(add_disk_work->disk);
1531 }
1532 #endif
1533 
1534 /*
1535  * SPECIAL CASE:
1536  *
1537  * This function basically calls add_disk() from a workqueue.   You may be
1538  * thinking: why not just call add_disk() directly?
1539  *
1540  * When you call add_disk(), the zvol appears to the world.  When this happens,
1541  * the kernel calls disk_scan_partitions() on the zvol, which behaves
1542  * differently on the 6.9+ kernels:
1543  *
1544  * - 6.8 and older kernels -
1545  * disk_scan_partitions()
1546  *	handle = bdev_open_by_dev(
1547  *		zvol_open()
1548  *	bdev_release(handle);
1549  *		zvol_release()
1550  *
1551  *
1552  * - 6.9+ kernels -
1553  * disk_scan_partitions()
1554  * 	file = bdev_file_open_by_dev()
1555  *		zvol_open()
1556  *	fput(file)
1557  *	< wait for return to userspace >
1558  *		zvol_release()
1559  *
1560  * The difference is that the bdev_release() from the 6.8 kernel is synchronous
1561  * while the fput() from the 6.9 kernel is async.  Or more specifically it's
1562  * async that has to wait until we return to userspace (since it adds the fput
1563  * into the caller's work queue with the TWA_RESUME flag set).  This is not the
1564  * behavior we want, since we want do things like create+destroy a zvol within
1565  * a single ZFS_IOC_CREATE ioctl, and the "create" part needs to release the
1566  * reference to the zvol while we're in the IOCTL, which can't wait until we
1567  * return to userspace.
1568  *
1569  * We can get around this since fput() has a special codepath for when it's
1570  * running in a kernel thread or interrupt.  In those cases, it just puts the
1571  * fput into the system workqueue, which we can force to run with
1572  * __flush_workqueue().  That is why we call add_disk() from a workqueue - so it
1573  * run from a kernel thread and "tricks" the fput() codepaths.
1574  *
1575  * Note that __flush_workqueue() is slowly getting deprecated.  This may be ok
1576  * though, since our IOCTL will spin on EBUSY waiting for the zvol release (via
1577  * fput) to happen, which it eventually, naturally, will from the system_wq
1578  * without us explicitly calling __flush_workqueue().
1579  */
1580 static int
zvol_os_add_disk(struct gendisk * disk)1581 zvol_os_add_disk(struct gendisk *disk)
1582 {
1583 #if defined(HAVE_BDEV_FILE_OPEN_BY_PATH)	/* 6.9+ kernel */
1584 	struct add_disk_work add_disk_work;
1585 
1586 	INIT_DELAYED_WORK(&add_disk_work.work, zvol_os_add_disk_work);
1587 	add_disk_work.disk = disk;
1588 	add_disk_work.error = 0;
1589 
1590 	/* Use *_delayed_work functions since they're not GPL'd */
1591 	schedule_delayed_work(&add_disk_work.work, 0);
1592 	flush_delayed_work(&add_disk_work.work);
1593 
1594 	__flush_workqueue(system_wq);
1595 	return (add_disk_work.error);
1596 #else	/* <= 6.8 kernel */
1597 	return (__zvol_os_add_disk(disk));
1598 #endif
1599 }
1600 
1601 /*
1602  * Create a block device minor node and setup the linkage between it
1603  * and the specified volume.  Once this function returns the block
1604  * device is live and ready for use.
1605  */
1606 int
zvol_os_create_minor(const char * name)1607 zvol_os_create_minor(const char *name)
1608 {
1609 	zvol_state_t *zv;
1610 	objset_t *os;
1611 	dmu_object_info_t *doi;
1612 	uint64_t volsize;
1613 	uint64_t len;
1614 	unsigned minor = 0;
1615 	int error = 0;
1616 	int idx;
1617 	uint64_t hash = zvol_name_hash(name);
1618 	uint64_t volthreading;
1619 	bool replayed_zil = B_FALSE;
1620 
1621 	if (zvol_inhibit_dev)
1622 		return (0);
1623 
1624 	idx = ida_simple_get(&zvol_ida, 0, 0, kmem_flags_convert(KM_SLEEP));
1625 	if (idx < 0)
1626 		return (SET_ERROR(-idx));
1627 	minor = idx << ZVOL_MINOR_BITS;
1628 	if (MINOR(minor) != minor) {
1629 		/* too many partitions can cause an overflow */
1630 		zfs_dbgmsg("zvol: create minor overflow: %s, minor %u/%u",
1631 		    name, minor, MINOR(minor));
1632 		ida_simple_remove(&zvol_ida, idx);
1633 		return (SET_ERROR(EINVAL));
1634 	}
1635 
1636 	zv = zvol_find_by_name_hash(name, hash, RW_NONE);
1637 	if (zv) {
1638 		ASSERT(MUTEX_HELD(&zv->zv_state_lock));
1639 		mutex_exit(&zv->zv_state_lock);
1640 		ida_simple_remove(&zvol_ida, idx);
1641 		return (SET_ERROR(EEXIST));
1642 	}
1643 
1644 	doi = kmem_alloc(sizeof (dmu_object_info_t), KM_SLEEP);
1645 
1646 	error = dmu_objset_own(name, DMU_OST_ZVOL, B_TRUE, B_TRUE, FTAG, &os);
1647 	if (error)
1648 		goto out_doi;
1649 
1650 	error = dmu_object_info(os, ZVOL_OBJ, doi);
1651 	if (error)
1652 		goto out_dmu_objset_disown;
1653 
1654 	error = zap_lookup(os, ZVOL_ZAP_OBJ, "size", 8, 1, &volsize);
1655 	if (error)
1656 		goto out_dmu_objset_disown;
1657 
1658 	zv = zvol_alloc(MKDEV(zvol_major, minor), name,
1659 	    doi->doi_data_block_size);
1660 	if (zv == NULL) {
1661 		error = SET_ERROR(EAGAIN);
1662 		goto out_dmu_objset_disown;
1663 	}
1664 	zv->zv_hash = hash;
1665 
1666 	if (dmu_objset_is_snapshot(os))
1667 		zv->zv_flags |= ZVOL_RDONLY;
1668 
1669 	zv->zv_volsize = volsize;
1670 	zv->zv_objset = os;
1671 
1672 	/* Default */
1673 	zv->zv_threading = B_TRUE;
1674 	if (dsl_prop_get_integer(name, "volthreading", &volthreading, NULL)
1675 	    == 0)
1676 		zv->zv_threading = volthreading;
1677 
1678 	set_capacity(zv->zv_zso->zvo_disk, zv->zv_volsize >> 9);
1679 
1680 #ifdef QUEUE_FLAG_DISCARD
1681 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, zv->zv_zso->zvo_queue);
1682 #endif
1683 #ifdef QUEUE_FLAG_NONROT
1684 	blk_queue_flag_set(QUEUE_FLAG_NONROT, zv->zv_zso->zvo_queue);
1685 #endif
1686 #ifdef QUEUE_FLAG_ADD_RANDOM
1687 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, zv->zv_zso->zvo_queue);
1688 #endif
1689 	/* This flag was introduced in kernel version 4.12. */
1690 #ifdef QUEUE_FLAG_SCSI_PASSTHROUGH
1691 	blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, zv->zv_zso->zvo_queue);
1692 #endif
1693 
1694 	ASSERT3P(zv->zv_kstat.dk_kstats, ==, NULL);
1695 	error = dataset_kstats_create(&zv->zv_kstat, zv->zv_objset);
1696 	if (error)
1697 		goto out_dmu_objset_disown;
1698 	ASSERT3P(zv->zv_zilog, ==, NULL);
1699 	zv->zv_zilog = zil_open(os, zvol_get_data, &zv->zv_kstat.dk_zil_sums);
1700 	if (spa_writeable(dmu_objset_spa(os))) {
1701 		if (zil_replay_disable)
1702 			replayed_zil = zil_destroy(zv->zv_zilog, B_FALSE);
1703 		else
1704 			replayed_zil = zil_replay(os, zv, zvol_replay_vector);
1705 	}
1706 	if (replayed_zil)
1707 		zil_close(zv->zv_zilog);
1708 	zv->zv_zilog = NULL;
1709 
1710 	/*
1711 	 * When udev detects the addition of the device it will immediately
1712 	 * invoke blkid(8) to determine the type of content on the device.
1713 	 * Prefetching the blocks commonly scanned by blkid(8) will speed
1714 	 * up this process.
1715 	 */
1716 	len = MIN(zvol_prefetch_bytes, SPA_MAXBLOCKSIZE);
1717 	if (len > 0) {
1718 		dmu_prefetch(os, ZVOL_OBJ, 0, 0, len, ZIO_PRIORITY_SYNC_READ);
1719 		dmu_prefetch(os, ZVOL_OBJ, 0, volsize - len, len,
1720 		    ZIO_PRIORITY_SYNC_READ);
1721 	}
1722 
1723 	zv->zv_objset = NULL;
1724 out_dmu_objset_disown:
1725 	dmu_objset_disown(os, B_TRUE, FTAG);
1726 out_doi:
1727 	kmem_free(doi, sizeof (dmu_object_info_t));
1728 
1729 	/*
1730 	 * Keep in mind that once add_disk() is called, the zvol is
1731 	 * announced to the world, and zvol_open()/zvol_release() can
1732 	 * be called at any time. Incidentally, add_disk() itself calls
1733 	 * zvol_open()->zvol_first_open() and zvol_release()->zvol_last_close()
1734 	 * directly as well.
1735 	 */
1736 	if (error == 0) {
1737 		rw_enter(&zvol_state_lock, RW_WRITER);
1738 		zvol_insert(zv);
1739 		rw_exit(&zvol_state_lock);
1740 		error = zvol_os_add_disk(zv->zv_zso->zvo_disk);
1741 	} else {
1742 		ida_simple_remove(&zvol_ida, idx);
1743 	}
1744 
1745 	return (error);
1746 }
1747 
1748 void
zvol_os_rename_minor(zvol_state_t * zv,const char * newname)1749 zvol_os_rename_minor(zvol_state_t *zv, const char *newname)
1750 {
1751 	int readonly = get_disk_ro(zv->zv_zso->zvo_disk);
1752 
1753 	ASSERT(RW_LOCK_HELD(&zvol_state_lock));
1754 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
1755 
1756 	strlcpy(zv->zv_name, newname, sizeof (zv->zv_name));
1757 
1758 	/* move to new hashtable entry  */
1759 	zv->zv_hash = zvol_name_hash(newname);
1760 	hlist_del(&zv->zv_hlink);
1761 	hlist_add_head(&zv->zv_hlink, ZVOL_HT_HEAD(zv->zv_hash));
1762 
1763 	/*
1764 	 * The block device's read-only state is briefly changed causing
1765 	 * a KOBJ_CHANGE uevent to be issued.  This ensures udev detects
1766 	 * the name change and fixes the symlinks.  This does not change
1767 	 * ZVOL_RDONLY in zv->zv_flags so the actual read-only state never
1768 	 * changes.  This would normally be done using kobject_uevent() but
1769 	 * that is a GPL-only symbol which is why we need this workaround.
1770 	 */
1771 	set_disk_ro(zv->zv_zso->zvo_disk, !readonly);
1772 	set_disk_ro(zv->zv_zso->zvo_disk, readonly);
1773 
1774 	dataset_kstats_rename(&zv->zv_kstat, newname);
1775 }
1776 
1777 void
zvol_os_set_disk_ro(zvol_state_t * zv,int flags)1778 zvol_os_set_disk_ro(zvol_state_t *zv, int flags)
1779 {
1780 
1781 	set_disk_ro(zv->zv_zso->zvo_disk, flags);
1782 }
1783 
1784 void
zvol_os_set_capacity(zvol_state_t * zv,uint64_t capacity)1785 zvol_os_set_capacity(zvol_state_t *zv, uint64_t capacity)
1786 {
1787 
1788 	set_capacity(zv->zv_zso->zvo_disk, capacity);
1789 }
1790 
1791 int
zvol_init(void)1792 zvol_init(void)
1793 {
1794 	int error;
1795 
1796 	/*
1797 	 * zvol_threads is the module param the user passes in.
1798 	 *
1799 	 * zvol_actual_threads is what we use internally, since the user can
1800 	 * pass zvol_thread = 0 to mean "use all the CPUs" (the default).
1801 	 */
1802 	static unsigned int zvol_actual_threads;
1803 
1804 	if (zvol_threads == 0) {
1805 		/*
1806 		 * See dde9380a1 for why 32 was chosen here.  This should
1807 		 * probably be refined to be some multiple of the number
1808 		 * of CPUs.
1809 		 */
1810 		zvol_actual_threads = MAX(num_online_cpus(), 32);
1811 	} else {
1812 		zvol_actual_threads = MIN(MAX(zvol_threads, 1), 1024);
1813 	}
1814 
1815 	/*
1816 	 * Use atleast 32 zvol_threads but for many core system,
1817 	 * prefer 6 threads per taskq, but no more taskqs
1818 	 * than threads in them on large systems.
1819 	 *
1820 	 *                 taskq   total
1821 	 * cpus    taskqs  threads threads
1822 	 * ------- ------- ------- -------
1823 	 * 1       1       32       32
1824 	 * 2       1       32       32
1825 	 * 4       1       32       32
1826 	 * 8       2       16       32
1827 	 * 16      3       11       33
1828 	 * 32      5       7        35
1829 	 * 64      8       8        64
1830 	 * 128     11      12       132
1831 	 * 256     16      16       256
1832 	 */
1833 	zv_taskq_t *ztqs = &zvol_taskqs;
1834 	uint_t num_tqs = MIN(num_online_cpus(), zvol_num_taskqs);
1835 	if (num_tqs == 0) {
1836 		num_tqs = 1 + num_online_cpus() / 6;
1837 		while (num_tqs * num_tqs > zvol_actual_threads)
1838 			num_tqs--;
1839 	}
1840 	uint_t per_tq_thread = zvol_actual_threads / num_tqs;
1841 	if (per_tq_thread * num_tqs < zvol_actual_threads)
1842 		per_tq_thread++;
1843 	ztqs->tqs_cnt = num_tqs;
1844 	ztqs->tqs_taskq = kmem_alloc(num_tqs * sizeof (taskq_t *), KM_SLEEP);
1845 	error = register_blkdev(zvol_major, ZVOL_DRIVER);
1846 	if (error) {
1847 		kmem_free(ztqs->tqs_taskq, ztqs->tqs_cnt * sizeof (taskq_t *));
1848 		ztqs->tqs_taskq = NULL;
1849 		printk(KERN_INFO "ZFS: register_blkdev() failed %d\n", error);
1850 		return (error);
1851 	}
1852 
1853 	if (zvol_blk_mq_queue_depth == 0) {
1854 		zvol_actual_blk_mq_queue_depth = BLKDEV_DEFAULT_RQ;
1855 	} else {
1856 		zvol_actual_blk_mq_queue_depth =
1857 		    MAX(zvol_blk_mq_queue_depth, BLKDEV_MIN_RQ);
1858 	}
1859 
1860 	if (zvol_blk_mq_threads == 0) {
1861 		zvol_blk_mq_actual_threads = num_online_cpus();
1862 	} else {
1863 		zvol_blk_mq_actual_threads = MIN(MAX(zvol_blk_mq_threads, 1),
1864 		    1024);
1865 	}
1866 
1867 	for (uint_t i = 0; i < num_tqs; i++) {
1868 		char name[32];
1869 		(void) snprintf(name, sizeof (name), "%s_tq-%u",
1870 		    ZVOL_DRIVER, i);
1871 		ztqs->tqs_taskq[i] = taskq_create(name, per_tq_thread,
1872 		    maxclsyspri, per_tq_thread, INT_MAX,
1873 		    TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
1874 		if (ztqs->tqs_taskq[i] == NULL) {
1875 			for (int j = i - 1; j >= 0; j--)
1876 				taskq_destroy(ztqs->tqs_taskq[j]);
1877 			unregister_blkdev(zvol_major, ZVOL_DRIVER);
1878 			kmem_free(ztqs->tqs_taskq, ztqs->tqs_cnt *
1879 			    sizeof (taskq_t *));
1880 			ztqs->tqs_taskq = NULL;
1881 			return (-ENOMEM);
1882 		}
1883 	}
1884 
1885 	zvol_init_impl();
1886 	ida_init(&zvol_ida);
1887 	return (0);
1888 }
1889 
1890 void
zvol_fini(void)1891 zvol_fini(void)
1892 {
1893 	zv_taskq_t *ztqs = &zvol_taskqs;
1894 	zvol_fini_impl();
1895 	unregister_blkdev(zvol_major, ZVOL_DRIVER);
1896 
1897 	if (ztqs->tqs_taskq == NULL) {
1898 		ASSERT3U(ztqs->tqs_cnt, ==, 0);
1899 	} else {
1900 		for (uint_t i = 0; i < ztqs->tqs_cnt; i++) {
1901 			ASSERT3P(ztqs->tqs_taskq[i], !=, NULL);
1902 			taskq_destroy(ztqs->tqs_taskq[i]);
1903 		}
1904 		kmem_free(ztqs->tqs_taskq, ztqs->tqs_cnt *
1905 		    sizeof (taskq_t *));
1906 		ztqs->tqs_taskq = NULL;
1907 	}
1908 
1909 	ida_destroy(&zvol_ida);
1910 }
1911 
1912 module_param(zvol_inhibit_dev, uint, 0644);
1913 MODULE_PARM_DESC(zvol_inhibit_dev, "Do not create zvol device nodes");
1914 
1915 module_param(zvol_major, uint, 0444);
1916 MODULE_PARM_DESC(zvol_major, "Major number for zvol device");
1917 
1918 module_param(zvol_threads, uint, 0444);
1919 MODULE_PARM_DESC(zvol_threads, "Number of threads to handle I/O requests. Set"
1920 	"to 0 to use all active CPUs");
1921 
1922 module_param(zvol_request_sync, uint, 0644);
1923 MODULE_PARM_DESC(zvol_request_sync, "Synchronously handle bio requests");
1924 
1925 module_param(zvol_max_discard_blocks, ulong, 0444);
1926 MODULE_PARM_DESC(zvol_max_discard_blocks, "Max number of blocks to discard");
1927 
1928 module_param(zvol_num_taskqs, uint, 0444);
1929 MODULE_PARM_DESC(zvol_num_taskqs, "Number of zvol taskqs");
1930 
1931 module_param(zvol_prefetch_bytes, uint, 0644);
1932 MODULE_PARM_DESC(zvol_prefetch_bytes, "Prefetch N bytes at zvol start+end");
1933 
1934 module_param(zvol_volmode, uint, 0644);
1935 MODULE_PARM_DESC(zvol_volmode, "Default volmode property value");
1936 
1937 module_param(zvol_blk_mq_queue_depth, uint, 0644);
1938 MODULE_PARM_DESC(zvol_blk_mq_queue_depth, "Default blk-mq queue depth");
1939 
1940 module_param(zvol_use_blk_mq, uint, 0644);
1941 MODULE_PARM_DESC(zvol_use_blk_mq, "Use the blk-mq API for zvols");
1942 
1943 module_param(zvol_blk_mq_blocks_per_thread, uint, 0644);
1944 MODULE_PARM_DESC(zvol_blk_mq_blocks_per_thread,
1945 	"Process volblocksize blocks per thread");
1946 
1947 #ifndef HAVE_BLKDEV_GET_ERESTARTSYS
1948 module_param(zvol_open_timeout_ms, uint, 0644);
1949 MODULE_PARM_DESC(zvol_open_timeout_ms, "Timeout for ZVOL open retries");
1950 #endif
1951