xref: /freebsd/sys/contrib/openzfs/module/os/linux/zfs/zvol_os.c (revision 718519f4efc71096422fc71dab90b2a3369871ff)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
23  * Copyright (c) 2024, Rob Norris <robn@despairlabs.com>
24  * Copyright (c) 2024, Klara, Inc.
25  */
26 
27 #include <sys/dataset_kstats.h>
28 #include <sys/dbuf.h>
29 #include <sys/dmu_traverse.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/zap.h>
34 #include <sys/zfeature.h>
35 #include <sys/zil_impl.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/zio.h>
38 #include <sys/zfs_rlock.h>
39 #include <sys/spa_impl.h>
40 #include <sys/zvol.h>
41 #include <sys/zvol_impl.h>
42 #include <cityhash.h>
43 
44 #include <linux/blkdev_compat.h>
45 #include <linux/task_io_accounting_ops.h>
46 #include <linux/workqueue.h>
47 #include <linux/blk-mq.h>
48 
49 static void zvol_request_impl(zvol_state_t *zv, struct bio *bio,
50     struct request *rq, boolean_t force_sync);
51 
52 static unsigned int zvol_major = ZVOL_MAJOR;
53 static unsigned int zvol_request_sync = 0;
54 static unsigned int zvol_prefetch_bytes = (128 * 1024);
55 static unsigned long zvol_max_discard_blocks = 16384;
56 
57 /*
58  * Switch taskq at multiple of 512 MB offset. This can be set to a lower value
59  * to utilize more threads for small files but may affect prefetch hits.
60  */
61 #define	ZVOL_TASKQ_OFFSET_SHIFT 29
62 
63 #ifndef HAVE_BLKDEV_GET_ERESTARTSYS
64 static unsigned int zvol_open_timeout_ms = 1000;
65 #endif
66 
67 static unsigned int zvol_threads = 0;
68 static unsigned int zvol_blk_mq_threads = 0;
69 static unsigned int zvol_blk_mq_actual_threads;
70 static boolean_t zvol_use_blk_mq = B_FALSE;
71 
72 /*
73  * The maximum number of volblocksize blocks to process per thread.  Typically,
74  * write heavy workloads preform better with higher values here, and read
75  * heavy workloads preform better with lower values, but that's not a hard
76  * and fast rule.  It's basically a knob to tune between "less overhead with
77  * less parallelism" and "more overhead, but more parallelism".
78  *
79  * '8' was chosen as a reasonable, balanced, default based off of sequential
80  * read and write tests to a zvol in an NVMe pool (with 16 CPUs).
81  */
82 static unsigned int zvol_blk_mq_blocks_per_thread = 8;
83 
84 static unsigned int zvol_num_taskqs = 0;
85 
86 #ifndef	BLKDEV_DEFAULT_RQ
87 /* BLKDEV_MAX_RQ was renamed to BLKDEV_DEFAULT_RQ in the 5.16 kernel */
88 #define	BLKDEV_DEFAULT_RQ BLKDEV_MAX_RQ
89 #endif
90 
91 /*
92  * Finalize our BIO or request.
93  */
94 static inline void
zvol_end_io(struct bio * bio,struct request * rq,int error)95 zvol_end_io(struct bio *bio, struct request *rq, int error)
96 {
97 	if (bio) {
98 		bio->bi_status = errno_to_bi_status(-error);
99 		bio_endio(bio);
100 	} else {
101 		blk_mq_end_request(rq, errno_to_bi_status(error));
102 	}
103 }
104 
105 static unsigned int zvol_blk_mq_queue_depth = BLKDEV_DEFAULT_RQ;
106 static unsigned int zvol_actual_blk_mq_queue_depth;
107 
108 struct zvol_state_os {
109 	struct gendisk		*zvo_disk;	/* generic disk */
110 	struct request_queue	*zvo_queue;	/* request queue */
111 	dev_t			zvo_dev;	/* device id */
112 
113 	struct blk_mq_tag_set tag_set;
114 
115 	/* Set from the global 'zvol_use_blk_mq' at zvol load */
116 	boolean_t use_blk_mq;
117 };
118 
119 typedef struct zv_taskq {
120 	uint_t tqs_cnt;
121 	taskq_t **tqs_taskq;
122 } zv_taskq_t;
123 static zv_taskq_t zvol_taskqs;
124 static struct ida zvol_ida;
125 
126 typedef struct zv_request_stack {
127 	zvol_state_t	*zv;
128 	struct bio	*bio;
129 	struct request *rq;
130 } zv_request_t;
131 
132 typedef struct zv_work {
133 	struct request  *rq;
134 	struct work_struct work;
135 } zv_work_t;
136 
137 typedef struct zv_request_task {
138 	zv_request_t zvr;
139 	taskq_ent_t	ent;
140 } zv_request_task_t;
141 
142 static zv_request_task_t *
zv_request_task_create(zv_request_t zvr)143 zv_request_task_create(zv_request_t zvr)
144 {
145 	zv_request_task_t *task;
146 	task = kmem_alloc(sizeof (zv_request_task_t), KM_SLEEP);
147 	taskq_init_ent(&task->ent);
148 	task->zvr = zvr;
149 	return (task);
150 }
151 
152 static void
zv_request_task_free(zv_request_task_t * task)153 zv_request_task_free(zv_request_task_t *task)
154 {
155 	kmem_free(task, sizeof (*task));
156 }
157 
158 /*
159  * This is called when a new block multiqueue request comes in.  A request
160  * contains one or more BIOs.
161  */
zvol_mq_queue_rq(struct blk_mq_hw_ctx * hctx,const struct blk_mq_queue_data * bd)162 static blk_status_t zvol_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
163     const struct blk_mq_queue_data *bd)
164 {
165 	struct request *rq = bd->rq;
166 	zvol_state_t *zv = rq->q->queuedata;
167 
168 	/* Tell the kernel that we are starting to process this request */
169 	blk_mq_start_request(rq);
170 
171 	if (blk_rq_is_passthrough(rq)) {
172 		/* Skip non filesystem request */
173 		blk_mq_end_request(rq, BLK_STS_IOERR);
174 		return (BLK_STS_IOERR);
175 	}
176 
177 	zvol_request_impl(zv, NULL, rq, 0);
178 
179 	/* Acknowledge to the kernel that we got this request */
180 	return (BLK_STS_OK);
181 }
182 
183 static struct blk_mq_ops zvol_blk_mq_queue_ops = {
184 	.queue_rq = zvol_mq_queue_rq,
185 };
186 
187 /* Initialize our blk-mq struct */
zvol_blk_mq_alloc_tag_set(zvol_state_t * zv)188 static int zvol_blk_mq_alloc_tag_set(zvol_state_t *zv)
189 {
190 	struct zvol_state_os *zso = zv->zv_zso;
191 
192 	memset(&zso->tag_set, 0, sizeof (zso->tag_set));
193 
194 	/* Initialize tag set. */
195 	zso->tag_set.ops = &zvol_blk_mq_queue_ops;
196 	zso->tag_set.nr_hw_queues = zvol_blk_mq_actual_threads;
197 	zso->tag_set.queue_depth = zvol_actual_blk_mq_queue_depth;
198 	zso->tag_set.numa_node = NUMA_NO_NODE;
199 	zso->tag_set.cmd_size = 0;
200 
201 	/*
202 	 * We need BLK_MQ_F_BLOCKING here since we do blocking calls in
203 	 * zvol_request_impl()
204 	 */
205 	zso->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
206 	zso->tag_set.driver_data = zv;
207 
208 	return (blk_mq_alloc_tag_set(&zso->tag_set));
209 }
210 
211 /*
212  * Given a path, return TRUE if path is a ZVOL.
213  */
214 boolean_t
zvol_os_is_zvol(const char * path)215 zvol_os_is_zvol(const char *path)
216 {
217 	dev_t dev = 0;
218 
219 	if (vdev_lookup_bdev(path, &dev) != 0)
220 		return (B_FALSE);
221 
222 	if (MAJOR(dev) == zvol_major)
223 		return (B_TRUE);
224 
225 	return (B_FALSE);
226 }
227 
228 static void
zvol_write(zv_request_t * zvr)229 zvol_write(zv_request_t *zvr)
230 {
231 	struct bio *bio = zvr->bio;
232 	struct request *rq = zvr->rq;
233 	int error = 0;
234 	zfs_uio_t uio;
235 	zvol_state_t *zv = zvr->zv;
236 	struct request_queue *q;
237 	struct gendisk *disk;
238 	unsigned long start_time = 0;
239 	boolean_t acct = B_FALSE;
240 
241 	ASSERT3P(zv, !=, NULL);
242 	ASSERT3U(zv->zv_open_count, >, 0);
243 	ASSERT3P(zv->zv_zilog, !=, NULL);
244 
245 	q = zv->zv_zso->zvo_queue;
246 	disk = zv->zv_zso->zvo_disk;
247 
248 	/* bio marked as FLUSH need to flush before write */
249 	if (io_is_flush(bio, rq))
250 		zil_commit(zv->zv_zilog, ZVOL_OBJ);
251 
252 	/* Some requests are just for flush and nothing else. */
253 	if (io_size(bio, rq) == 0) {
254 		rw_exit(&zv->zv_suspend_lock);
255 		zvol_end_io(bio, rq, 0);
256 		return;
257 	}
258 
259 	zfs_uio_bvec_init(&uio, bio, rq);
260 
261 	ssize_t start_resid = uio.uio_resid;
262 
263 	/*
264 	 * With use_blk_mq, accounting is done by blk_mq_start_request()
265 	 * and blk_mq_end_request(), so we can skip it here.
266 	 */
267 	if (bio) {
268 		acct = blk_queue_io_stat(q);
269 		if (acct) {
270 			start_time = blk_generic_start_io_acct(q, disk, WRITE,
271 			    bio);
272 		}
273 	}
274 
275 	boolean_t sync =
276 	    io_is_fua(bio, rq) || zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS;
277 
278 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zv->zv_rangelock,
279 	    uio.uio_loffset, uio.uio_resid, RL_WRITER);
280 
281 	uint64_t volsize = zv->zv_volsize;
282 	while (uio.uio_resid > 0 && uio.uio_loffset < volsize) {
283 		uint64_t bytes = MIN(uio.uio_resid, DMU_MAX_ACCESS >> 1);
284 		uint64_t off = uio.uio_loffset;
285 		dmu_tx_t *tx = dmu_tx_create(zv->zv_objset);
286 
287 		if (bytes > volsize - off)	/* don't write past the end */
288 			bytes = volsize - off;
289 
290 		dmu_tx_hold_write_by_dnode(tx, zv->zv_dn, off, bytes);
291 
292 		/* This will only fail for ENOSPC */
293 		error = dmu_tx_assign(tx, TXG_WAIT);
294 		if (error) {
295 			dmu_tx_abort(tx);
296 			break;
297 		}
298 		error = dmu_write_uio_dnode(zv->zv_dn, &uio, bytes, tx);
299 		if (error == 0) {
300 			zvol_log_write(zv, tx, off, bytes, sync);
301 		}
302 		dmu_tx_commit(tx);
303 
304 		if (error)
305 			break;
306 	}
307 	zfs_rangelock_exit(lr);
308 
309 	int64_t nwritten = start_resid - uio.uio_resid;
310 	dataset_kstats_update_write_kstats(&zv->zv_kstat, nwritten);
311 	task_io_account_write(nwritten);
312 
313 	if (sync)
314 		zil_commit(zv->zv_zilog, ZVOL_OBJ);
315 
316 	rw_exit(&zv->zv_suspend_lock);
317 
318 	if (bio && acct) {
319 		blk_generic_end_io_acct(q, disk, WRITE, bio, start_time);
320 	}
321 
322 	zvol_end_io(bio, rq, -error);
323 }
324 
325 static void
zvol_write_task(void * arg)326 zvol_write_task(void *arg)
327 {
328 	zv_request_task_t *task = arg;
329 	zvol_write(&task->zvr);
330 	zv_request_task_free(task);
331 }
332 
333 static void
zvol_discard(zv_request_t * zvr)334 zvol_discard(zv_request_t *zvr)
335 {
336 	struct bio *bio = zvr->bio;
337 	struct request *rq = zvr->rq;
338 	zvol_state_t *zv = zvr->zv;
339 	uint64_t start = io_offset(bio, rq);
340 	uint64_t size = io_size(bio, rq);
341 	uint64_t end = start + size;
342 	boolean_t sync;
343 	int error = 0;
344 	dmu_tx_t *tx;
345 	struct request_queue *q = zv->zv_zso->zvo_queue;
346 	struct gendisk *disk = zv->zv_zso->zvo_disk;
347 	unsigned long start_time = 0;
348 	boolean_t acct = B_FALSE;
349 
350 	ASSERT3P(zv, !=, NULL);
351 	ASSERT3U(zv->zv_open_count, >, 0);
352 	ASSERT3P(zv->zv_zilog, !=, NULL);
353 
354 	if (bio) {
355 		acct = blk_queue_io_stat(q);
356 		if (acct) {
357 			start_time = blk_generic_start_io_acct(q, disk, WRITE,
358 			    bio);
359 		}
360 	}
361 
362 	sync = io_is_fua(bio, rq) || zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS;
363 
364 	if (end > zv->zv_volsize) {
365 		error = SET_ERROR(EIO);
366 		goto unlock;
367 	}
368 
369 	/*
370 	 * Align the request to volume block boundaries when a secure erase is
371 	 * not required.  This will prevent dnode_free_range() from zeroing out
372 	 * the unaligned parts which is slow (read-modify-write) and useless
373 	 * since we are not freeing any space by doing so.
374 	 */
375 	if (!io_is_secure_erase(bio, rq)) {
376 		start = P2ROUNDUP(start, zv->zv_volblocksize);
377 		end = P2ALIGN_TYPED(end, zv->zv_volblocksize, uint64_t);
378 		size = end - start;
379 	}
380 
381 	if (start >= end)
382 		goto unlock;
383 
384 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zv->zv_rangelock,
385 	    start, size, RL_WRITER);
386 
387 	tx = dmu_tx_create(zv->zv_objset);
388 	dmu_tx_mark_netfree(tx);
389 	error = dmu_tx_assign(tx, TXG_WAIT);
390 	if (error != 0) {
391 		dmu_tx_abort(tx);
392 	} else {
393 		zvol_log_truncate(zv, tx, start, size);
394 		dmu_tx_commit(tx);
395 		error = dmu_free_long_range(zv->zv_objset,
396 		    ZVOL_OBJ, start, size);
397 	}
398 	zfs_rangelock_exit(lr);
399 
400 	if (error == 0 && sync)
401 		zil_commit(zv->zv_zilog, ZVOL_OBJ);
402 
403 unlock:
404 	rw_exit(&zv->zv_suspend_lock);
405 
406 	if (bio && acct) {
407 		blk_generic_end_io_acct(q, disk, WRITE, bio,
408 		    start_time);
409 	}
410 
411 	zvol_end_io(bio, rq, -error);
412 }
413 
414 static void
zvol_discard_task(void * arg)415 zvol_discard_task(void *arg)
416 {
417 	zv_request_task_t *task = arg;
418 	zvol_discard(&task->zvr);
419 	zv_request_task_free(task);
420 }
421 
422 static void
zvol_read(zv_request_t * zvr)423 zvol_read(zv_request_t *zvr)
424 {
425 	struct bio *bio = zvr->bio;
426 	struct request *rq = zvr->rq;
427 	int error = 0;
428 	zfs_uio_t uio;
429 	boolean_t acct = B_FALSE;
430 	zvol_state_t *zv = zvr->zv;
431 	struct request_queue *q;
432 	struct gendisk *disk;
433 	unsigned long start_time = 0;
434 
435 	ASSERT3P(zv, !=, NULL);
436 	ASSERT3U(zv->zv_open_count, >, 0);
437 
438 	zfs_uio_bvec_init(&uio, bio, rq);
439 
440 	q = zv->zv_zso->zvo_queue;
441 	disk = zv->zv_zso->zvo_disk;
442 
443 	ssize_t start_resid = uio.uio_resid;
444 
445 	/*
446 	 * When blk-mq is being used, accounting is done by
447 	 * blk_mq_start_request() and blk_mq_end_request().
448 	 */
449 	if (bio) {
450 		acct = blk_queue_io_stat(q);
451 		if (acct)
452 			start_time = blk_generic_start_io_acct(q, disk, READ,
453 			    bio);
454 	}
455 
456 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zv->zv_rangelock,
457 	    uio.uio_loffset, uio.uio_resid, RL_READER);
458 
459 	uint64_t volsize = zv->zv_volsize;
460 
461 	while (uio.uio_resid > 0 && uio.uio_loffset < volsize) {
462 		uint64_t bytes = MIN(uio.uio_resid, DMU_MAX_ACCESS >> 1);
463 
464 		/* don't read past the end */
465 		if (bytes > volsize - uio.uio_loffset)
466 			bytes = volsize - uio.uio_loffset;
467 
468 		error = dmu_read_uio_dnode(zv->zv_dn, &uio, bytes);
469 		if (error) {
470 			/* convert checksum errors into IO errors */
471 			if (error == ECKSUM)
472 				error = SET_ERROR(EIO);
473 			break;
474 		}
475 	}
476 	zfs_rangelock_exit(lr);
477 
478 	int64_t nread = start_resid - uio.uio_resid;
479 	dataset_kstats_update_read_kstats(&zv->zv_kstat, nread);
480 	task_io_account_read(nread);
481 
482 	rw_exit(&zv->zv_suspend_lock);
483 
484 	if (bio && acct) {
485 		blk_generic_end_io_acct(q, disk, READ, bio, start_time);
486 	}
487 
488 	zvol_end_io(bio, rq, -error);
489 }
490 
491 static void
zvol_read_task(void * arg)492 zvol_read_task(void *arg)
493 {
494 	zv_request_task_t *task = arg;
495 	zvol_read(&task->zvr);
496 	zv_request_task_free(task);
497 }
498 
499 
500 /*
501  * Process a BIO or request
502  *
503  * Either 'bio' or 'rq' should be set depending on if we are processing a
504  * bio or a request (both should not be set).
505  *
506  * force_sync:	Set to 0 to defer processing to a background taskq
507  *			Set to 1 to process data synchronously
508  */
509 static void
zvol_request_impl(zvol_state_t * zv,struct bio * bio,struct request * rq,boolean_t force_sync)510 zvol_request_impl(zvol_state_t *zv, struct bio *bio, struct request *rq,
511     boolean_t force_sync)
512 {
513 	fstrans_cookie_t cookie = spl_fstrans_mark();
514 	uint64_t offset = io_offset(bio, rq);
515 	uint64_t size = io_size(bio, rq);
516 	int rw = io_data_dir(bio, rq);
517 
518 	if (unlikely(zv->zv_flags & ZVOL_REMOVING)) {
519 		zvol_end_io(bio, rq, -SET_ERROR(ENXIO));
520 		goto out;
521 	}
522 
523 	if (zvol_request_sync || zv->zv_threading == B_FALSE)
524 		force_sync = 1;
525 
526 	zv_request_t zvr = {
527 		.zv = zv,
528 		.bio = bio,
529 		.rq = rq,
530 	};
531 
532 	if (io_has_data(bio, rq) && offset + size > zv->zv_volsize) {
533 		printk(KERN_INFO "%s: bad access: offset=%llu, size=%lu\n",
534 		    zv->zv_zso->zvo_disk->disk_name,
535 		    (long long unsigned)offset,
536 		    (long unsigned)size);
537 
538 		zvol_end_io(bio, rq, -SET_ERROR(EIO));
539 		goto out;
540 	}
541 
542 	zv_request_task_t *task;
543 	zv_taskq_t *ztqs = &zvol_taskqs;
544 	uint_t blk_mq_hw_queue = 0;
545 	uint_t tq_idx;
546 	uint_t taskq_hash;
547 	if (rq)
548 #ifdef HAVE_BLK_MQ_RQ_HCTX
549 		blk_mq_hw_queue = rq->mq_hctx->queue_num;
550 #else
551 		blk_mq_hw_queue =
552 		    rq->q->queue_hw_ctx[rq->q->mq_map[rq->cpu]]->queue_num;
553 #endif
554 	taskq_hash = cityhash3((uintptr_t)zv, offset >> ZVOL_TASKQ_OFFSET_SHIFT,
555 	    blk_mq_hw_queue);
556 	tq_idx = taskq_hash % ztqs->tqs_cnt;
557 
558 	if (rw == WRITE) {
559 		if (unlikely(zv->zv_flags & ZVOL_RDONLY)) {
560 			zvol_end_io(bio, rq, -SET_ERROR(EROFS));
561 			goto out;
562 		}
563 
564 		/*
565 		 * Prevents the zvol from being suspended, or the ZIL being
566 		 * concurrently opened.  Will be released after the i/o
567 		 * completes.
568 		 */
569 		rw_enter(&zv->zv_suspend_lock, RW_READER);
570 
571 		/*
572 		 * Open a ZIL if this is the first time we have written to this
573 		 * zvol. We protect zv->zv_zilog with zv_suspend_lock rather
574 		 * than zv_state_lock so that we don't need to acquire an
575 		 * additional lock in this path.
576 		 */
577 		if (zv->zv_zilog == NULL) {
578 			rw_exit(&zv->zv_suspend_lock);
579 			rw_enter(&zv->zv_suspend_lock, RW_WRITER);
580 			if (zv->zv_zilog == NULL) {
581 				zv->zv_zilog = zil_open(zv->zv_objset,
582 				    zvol_get_data, &zv->zv_kstat.dk_zil_sums);
583 				zv->zv_flags |= ZVOL_WRITTEN_TO;
584 				/* replay / destroy done in zvol_create_minor */
585 				VERIFY0((zv->zv_zilog->zl_header->zh_flags &
586 				    ZIL_REPLAY_NEEDED));
587 			}
588 			rw_downgrade(&zv->zv_suspend_lock);
589 		}
590 
591 		/*
592 		 * We don't want this thread to be blocked waiting for i/o to
593 		 * complete, so we instead wait from a taskq callback. The
594 		 * i/o may be a ZIL write (via zil_commit()), or a read of an
595 		 * indirect block, or a read of a data block (if this is a
596 		 * partial-block write).  We will indicate that the i/o is
597 		 * complete by calling END_IO() from the taskq callback.
598 		 *
599 		 * This design allows the calling thread to continue and
600 		 * initiate more concurrent operations by calling
601 		 * zvol_request() again. There are typically only a small
602 		 * number of threads available to call zvol_request() (e.g.
603 		 * one per iSCSI target), so keeping the latency of
604 		 * zvol_request() low is important for performance.
605 		 *
606 		 * The zvol_request_sync module parameter allows this
607 		 * behavior to be altered, for performance evaluation
608 		 * purposes.  If the callback blocks, setting
609 		 * zvol_request_sync=1 will result in much worse performance.
610 		 *
611 		 * We can have up to zvol_threads concurrent i/o's being
612 		 * processed for all zvols on the system.  This is typically
613 		 * a vast improvement over the zvol_request_sync=1 behavior
614 		 * of one i/o at a time per zvol.  However, an even better
615 		 * design would be for zvol_request() to initiate the zio
616 		 * directly, and then be notified by the zio_done callback,
617 		 * which would call END_IO().  Unfortunately, the DMU/ZIL
618 		 * interfaces lack this functionality (they block waiting for
619 		 * the i/o to complete).
620 		 */
621 		if (io_is_discard(bio, rq) || io_is_secure_erase(bio, rq)) {
622 			if (force_sync) {
623 				zvol_discard(&zvr);
624 			} else {
625 				task = zv_request_task_create(zvr);
626 				taskq_dispatch_ent(ztqs->tqs_taskq[tq_idx],
627 				    zvol_discard_task, task, 0, &task->ent);
628 			}
629 		} else {
630 			if (force_sync) {
631 				zvol_write(&zvr);
632 			} else {
633 				task = zv_request_task_create(zvr);
634 				taskq_dispatch_ent(ztqs->tqs_taskq[tq_idx],
635 				    zvol_write_task, task, 0, &task->ent);
636 			}
637 		}
638 	} else {
639 		/*
640 		 * The SCST driver, and possibly others, may issue READ I/Os
641 		 * with a length of zero bytes.  These empty I/Os contain no
642 		 * data and require no additional handling.
643 		 */
644 		if (size == 0) {
645 			zvol_end_io(bio, rq, 0);
646 			goto out;
647 		}
648 
649 		rw_enter(&zv->zv_suspend_lock, RW_READER);
650 
651 		/* See comment in WRITE case above. */
652 		if (force_sync) {
653 			zvol_read(&zvr);
654 		} else {
655 			task = zv_request_task_create(zvr);
656 			taskq_dispatch_ent(ztqs->tqs_taskq[tq_idx],
657 			    zvol_read_task, task, 0, &task->ent);
658 		}
659 	}
660 
661 out:
662 	spl_fstrans_unmark(cookie);
663 }
664 
665 #ifdef HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS
666 #ifdef HAVE_BDEV_SUBMIT_BIO_RETURNS_VOID
667 static void
zvol_submit_bio(struct bio * bio)668 zvol_submit_bio(struct bio *bio)
669 #else
670 static blk_qc_t
671 zvol_submit_bio(struct bio *bio)
672 #endif
673 #else
674 static MAKE_REQUEST_FN_RET
675 zvol_request(struct request_queue *q, struct bio *bio)
676 #endif
677 {
678 #ifdef HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS
679 #if defined(HAVE_BIO_BDEV_DISK)
680 	struct request_queue *q = bio->bi_bdev->bd_disk->queue;
681 #else
682 	struct request_queue *q = bio->bi_disk->queue;
683 #endif
684 #endif
685 	zvol_state_t *zv = q->queuedata;
686 
687 	zvol_request_impl(zv, bio, NULL, 0);
688 #if defined(HAVE_MAKE_REQUEST_FN_RET_QC) || \
689 	defined(HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS) && \
690 	!defined(HAVE_BDEV_SUBMIT_BIO_RETURNS_VOID)
691 	return (BLK_QC_T_NONE);
692 #endif
693 }
694 
695 static int
696 #ifdef HAVE_BLK_MODE_T
zvol_open(struct gendisk * disk,blk_mode_t flag)697 zvol_open(struct gendisk *disk, blk_mode_t flag)
698 #else
699 zvol_open(struct block_device *bdev, fmode_t flag)
700 #endif
701 {
702 	zvol_state_t *zv;
703 	int error = 0;
704 	boolean_t drop_suspend = B_FALSE;
705 #ifndef HAVE_BLKDEV_GET_ERESTARTSYS
706 	hrtime_t timeout = MSEC2NSEC(zvol_open_timeout_ms);
707 	hrtime_t start = gethrtime();
708 
709 retry:
710 #endif
711 	rw_enter(&zvol_state_lock, RW_READER);
712 	/*
713 	 * Obtain a copy of private_data under the zvol_state_lock to make
714 	 * sure that either the result of zvol free code path setting
715 	 * disk->private_data to NULL is observed, or zvol_os_free()
716 	 * is not called on this zv because of the positive zv_open_count.
717 	 */
718 #ifdef HAVE_BLK_MODE_T
719 	zv = disk->private_data;
720 #else
721 	zv = bdev->bd_disk->private_data;
722 #endif
723 	if (zv == NULL) {
724 		rw_exit(&zvol_state_lock);
725 		return (-SET_ERROR(ENXIO));
726 	}
727 
728 	mutex_enter(&zv->zv_state_lock);
729 
730 	if (unlikely(zv->zv_flags & ZVOL_REMOVING)) {
731 		mutex_exit(&zv->zv_state_lock);
732 		rw_exit(&zvol_state_lock);
733 		return (-SET_ERROR(ENXIO));
734 	}
735 
736 	/*
737 	 * Make sure zvol is not suspended during first open
738 	 * (hold zv_suspend_lock) and respect proper lock acquisition
739 	 * ordering - zv_suspend_lock before zv_state_lock
740 	 */
741 	if (zv->zv_open_count == 0) {
742 		if (!rw_tryenter(&zv->zv_suspend_lock, RW_READER)) {
743 			mutex_exit(&zv->zv_state_lock);
744 			rw_enter(&zv->zv_suspend_lock, RW_READER);
745 			mutex_enter(&zv->zv_state_lock);
746 			/* check to see if zv_suspend_lock is needed */
747 			if (zv->zv_open_count != 0) {
748 				rw_exit(&zv->zv_suspend_lock);
749 			} else {
750 				drop_suspend = B_TRUE;
751 			}
752 		} else {
753 			drop_suspend = B_TRUE;
754 		}
755 	}
756 	rw_exit(&zvol_state_lock);
757 
758 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
759 
760 	if (zv->zv_open_count == 0) {
761 		boolean_t drop_namespace = B_FALSE;
762 
763 		ASSERT(RW_READ_HELD(&zv->zv_suspend_lock));
764 
765 		/*
766 		 * In all other call paths the spa_namespace_lock is taken
767 		 * before the bdev->bd_mutex lock.  However, on open(2)
768 		 * the __blkdev_get() function calls fops->open() with the
769 		 * bdev->bd_mutex lock held.  This can result in a deadlock
770 		 * when zvols from one pool are used as vdevs in another.
771 		 *
772 		 * To prevent a lock inversion deadlock we preemptively
773 		 * take the spa_namespace_lock.  Normally the lock will not
774 		 * be contended and this is safe because spa_open_common()
775 		 * handles the case where the caller already holds the
776 		 * spa_namespace_lock.
777 		 *
778 		 * When the lock cannot be aquired after multiple retries
779 		 * this must be the vdev on zvol deadlock case and we have
780 		 * no choice but to return an error.  For 5.12 and older
781 		 * kernels returning -ERESTARTSYS will result in the
782 		 * bdev->bd_mutex being dropped, then reacquired, and
783 		 * fops->open() being called again.  This process can be
784 		 * repeated safely until both locks are acquired.  For 5.13
785 		 * and newer the -ERESTARTSYS retry logic was removed from
786 		 * the kernel so the only option is to return the error for
787 		 * the caller to handle it.
788 		 */
789 		if (!mutex_owned(&spa_namespace_lock)) {
790 			if (!mutex_tryenter(&spa_namespace_lock)) {
791 				mutex_exit(&zv->zv_state_lock);
792 				rw_exit(&zv->zv_suspend_lock);
793 				drop_suspend = B_FALSE;
794 
795 #ifdef HAVE_BLKDEV_GET_ERESTARTSYS
796 				schedule();
797 				return (-SET_ERROR(ERESTARTSYS));
798 #else
799 				if ((gethrtime() - start) > timeout)
800 					return (-SET_ERROR(ERESTARTSYS));
801 
802 				schedule_timeout_interruptible(
803 					MSEC_TO_TICK(10));
804 				goto retry;
805 #endif
806 			} else {
807 				drop_namespace = B_TRUE;
808 			}
809 		}
810 
811 		error = -zvol_first_open(zv, !(blk_mode_is_open_write(flag)));
812 
813 		if (drop_namespace)
814 			mutex_exit(&spa_namespace_lock);
815 	}
816 
817 	if (error == 0) {
818 		if ((blk_mode_is_open_write(flag)) &&
819 		    (zv->zv_flags & ZVOL_RDONLY)) {
820 			if (zv->zv_open_count == 0)
821 				zvol_last_close(zv);
822 
823 			error = -SET_ERROR(EROFS);
824 		} else {
825 			zv->zv_open_count++;
826 		}
827 	}
828 
829 	mutex_exit(&zv->zv_state_lock);
830 	if (drop_suspend)
831 		rw_exit(&zv->zv_suspend_lock);
832 
833 	if (error == 0)
834 #ifdef HAVE_BLK_MODE_T
835 		disk_check_media_change(disk);
836 #else
837 		zfs_check_media_change(bdev);
838 #endif
839 
840 	return (error);
841 }
842 
843 static void
844 #ifdef HAVE_BLOCK_DEVICE_OPERATIONS_RELEASE_1ARG
zvol_release(struct gendisk * disk)845 zvol_release(struct gendisk *disk)
846 #else
847 zvol_release(struct gendisk *disk, fmode_t unused)
848 #endif
849 {
850 #if !defined(HAVE_BLOCK_DEVICE_OPERATIONS_RELEASE_1ARG)
851 	(void) unused;
852 #endif
853 	zvol_state_t *zv;
854 	boolean_t drop_suspend = B_TRUE;
855 
856 	rw_enter(&zvol_state_lock, RW_READER);
857 	zv = disk->private_data;
858 
859 	mutex_enter(&zv->zv_state_lock);
860 	ASSERT3U(zv->zv_open_count, >, 0);
861 	/*
862 	 * make sure zvol is not suspended during last close
863 	 * (hold zv_suspend_lock) and respect proper lock acquisition
864 	 * ordering - zv_suspend_lock before zv_state_lock
865 	 */
866 	if (zv->zv_open_count == 1) {
867 		if (!rw_tryenter(&zv->zv_suspend_lock, RW_READER)) {
868 			mutex_exit(&zv->zv_state_lock);
869 			rw_enter(&zv->zv_suspend_lock, RW_READER);
870 			mutex_enter(&zv->zv_state_lock);
871 			/* check to see if zv_suspend_lock is needed */
872 			if (zv->zv_open_count != 1) {
873 				rw_exit(&zv->zv_suspend_lock);
874 				drop_suspend = B_FALSE;
875 			}
876 		}
877 	} else {
878 		drop_suspend = B_FALSE;
879 	}
880 	rw_exit(&zvol_state_lock);
881 
882 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
883 
884 	zv->zv_open_count--;
885 	if (zv->zv_open_count == 0) {
886 		ASSERT(RW_READ_HELD(&zv->zv_suspend_lock));
887 		zvol_last_close(zv);
888 	}
889 
890 	mutex_exit(&zv->zv_state_lock);
891 
892 	if (drop_suspend)
893 		rw_exit(&zv->zv_suspend_lock);
894 }
895 
896 static int
zvol_ioctl(struct block_device * bdev,fmode_t mode,unsigned int cmd,unsigned long arg)897 zvol_ioctl(struct block_device *bdev, fmode_t mode,
898     unsigned int cmd, unsigned long arg)
899 {
900 	zvol_state_t *zv = bdev->bd_disk->private_data;
901 	int error = 0;
902 
903 	ASSERT3U(zv->zv_open_count, >, 0);
904 
905 	switch (cmd) {
906 	case BLKFLSBUF:
907 #ifdef HAVE_FSYNC_BDEV
908 		fsync_bdev(bdev);
909 #elif defined(HAVE_SYNC_BLOCKDEV)
910 		sync_blockdev(bdev);
911 #else
912 #error "Neither fsync_bdev() nor sync_blockdev() found"
913 #endif
914 		invalidate_bdev(bdev);
915 		rw_enter(&zv->zv_suspend_lock, RW_READER);
916 
917 		if (!(zv->zv_flags & ZVOL_RDONLY))
918 			txg_wait_synced(dmu_objset_pool(zv->zv_objset), 0);
919 
920 		rw_exit(&zv->zv_suspend_lock);
921 		break;
922 
923 	case BLKZNAME:
924 		mutex_enter(&zv->zv_state_lock);
925 		error = copy_to_user((void *)arg, zv->zv_name, MAXNAMELEN);
926 		mutex_exit(&zv->zv_state_lock);
927 		break;
928 
929 	default:
930 		error = -ENOTTY;
931 		break;
932 	}
933 
934 	return (SET_ERROR(error));
935 }
936 
937 #ifdef CONFIG_COMPAT
938 static int
zvol_compat_ioctl(struct block_device * bdev,fmode_t mode,unsigned cmd,unsigned long arg)939 zvol_compat_ioctl(struct block_device *bdev, fmode_t mode,
940     unsigned cmd, unsigned long arg)
941 {
942 	return (zvol_ioctl(bdev, mode, cmd, arg));
943 }
944 #else
945 #define	zvol_compat_ioctl	NULL
946 #endif
947 
948 static unsigned int
zvol_check_events(struct gendisk * disk,unsigned int clearing)949 zvol_check_events(struct gendisk *disk, unsigned int clearing)
950 {
951 	unsigned int mask = 0;
952 
953 	rw_enter(&zvol_state_lock, RW_READER);
954 
955 	zvol_state_t *zv = disk->private_data;
956 	if (zv != NULL) {
957 		mutex_enter(&zv->zv_state_lock);
958 		mask = zv->zv_changed ? DISK_EVENT_MEDIA_CHANGE : 0;
959 		zv->zv_changed = 0;
960 		mutex_exit(&zv->zv_state_lock);
961 	}
962 
963 	rw_exit(&zvol_state_lock);
964 
965 	return (mask);
966 }
967 
968 static int
zvol_revalidate_disk(struct gendisk * disk)969 zvol_revalidate_disk(struct gendisk *disk)
970 {
971 	rw_enter(&zvol_state_lock, RW_READER);
972 
973 	zvol_state_t *zv = disk->private_data;
974 	if (zv != NULL) {
975 		mutex_enter(&zv->zv_state_lock);
976 		set_capacity(zv->zv_zso->zvo_disk,
977 		    zv->zv_volsize >> SECTOR_BITS);
978 		mutex_exit(&zv->zv_state_lock);
979 	}
980 
981 	rw_exit(&zvol_state_lock);
982 
983 	return (0);
984 }
985 
986 int
zvol_os_update_volsize(zvol_state_t * zv,uint64_t volsize)987 zvol_os_update_volsize(zvol_state_t *zv, uint64_t volsize)
988 {
989 	struct gendisk *disk = zv->zv_zso->zvo_disk;
990 
991 #if defined(HAVE_REVALIDATE_DISK_SIZE)
992 	revalidate_disk_size(disk, zvol_revalidate_disk(disk) == 0);
993 #elif defined(HAVE_REVALIDATE_DISK)
994 	revalidate_disk(disk);
995 #else
996 	zvol_revalidate_disk(disk);
997 #endif
998 	return (0);
999 }
1000 
1001 void
zvol_os_clear_private(zvol_state_t * zv)1002 zvol_os_clear_private(zvol_state_t *zv)
1003 {
1004 	/*
1005 	 * Cleared while holding zvol_state_lock as a writer
1006 	 * which will prevent zvol_open() from opening it.
1007 	 */
1008 	zv->zv_zso->zvo_disk->private_data = NULL;
1009 }
1010 
1011 /*
1012  * Provide a simple virtual geometry for legacy compatibility.  For devices
1013  * smaller than 1 MiB a small head and sector count is used to allow very
1014  * tiny devices.  For devices over 1 Mib a standard head and sector count
1015  * is used to keep the cylinders count reasonable.
1016  */
1017 static int
zvol_getgeo(struct block_device * bdev,struct hd_geometry * geo)1018 zvol_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1019 {
1020 	zvol_state_t *zv = bdev->bd_disk->private_data;
1021 	sector_t sectors;
1022 
1023 	ASSERT3U(zv->zv_open_count, >, 0);
1024 
1025 	sectors = get_capacity(zv->zv_zso->zvo_disk);
1026 
1027 	if (sectors > 2048) {
1028 		geo->heads = 16;
1029 		geo->sectors = 63;
1030 	} else {
1031 		geo->heads = 2;
1032 		geo->sectors = 4;
1033 	}
1034 
1035 	geo->start = 0;
1036 	geo->cylinders = sectors / (geo->heads * geo->sectors);
1037 
1038 	return (0);
1039 }
1040 
1041 /*
1042  * Why have two separate block_device_operations structs?
1043  *
1044  * Normally we'd just have one, and assign 'submit_bio' as needed.  However,
1045  * it's possible the user's kernel is built with CONSTIFY_PLUGIN, meaning we
1046  * can't just change submit_bio dynamically at runtime.  So just create two
1047  * separate structs to get around this.
1048  */
1049 static const struct block_device_operations zvol_ops_blk_mq = {
1050 	.open			= zvol_open,
1051 	.release		= zvol_release,
1052 	.ioctl			= zvol_ioctl,
1053 	.compat_ioctl		= zvol_compat_ioctl,
1054 	.check_events		= zvol_check_events,
1055 #ifdef HAVE_BLOCK_DEVICE_OPERATIONS_REVALIDATE_DISK
1056 	.revalidate_disk	= zvol_revalidate_disk,
1057 #endif
1058 	.getgeo			= zvol_getgeo,
1059 	.owner			= THIS_MODULE,
1060 };
1061 
1062 static const struct block_device_operations zvol_ops = {
1063 	.open			= zvol_open,
1064 	.release		= zvol_release,
1065 	.ioctl			= zvol_ioctl,
1066 	.compat_ioctl		= zvol_compat_ioctl,
1067 	.check_events		= zvol_check_events,
1068 #ifdef HAVE_BLOCK_DEVICE_OPERATIONS_REVALIDATE_DISK
1069 	.revalidate_disk	= zvol_revalidate_disk,
1070 #endif
1071 	.getgeo			= zvol_getgeo,
1072 	.owner			= THIS_MODULE,
1073 #ifdef HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS
1074 	.submit_bio		= zvol_submit_bio,
1075 #endif
1076 };
1077 
1078 /*
1079  * Since 6.9, Linux has been removing queue limit setters in favour of an
1080  * initial queue_limits struct applied when the device is open. Since 6.11,
1081  * queue_limits is being extended to allow more things to be applied when the
1082  * device is open. Setters are also being removed for this.
1083  *
1084  * For OpenZFS, this means that depending on kernel version, some options may
1085  * be set up before the device is open, and some applied to an open device
1086  * (queue) after the fact.
1087  *
1088  * We manage this complexity by having our own limits struct,
1089  * zvol_queue_limits_t, in which we carry any queue config that we're
1090  * interested in setting. This structure is the same on all kernels.
1091  *
1092  * These limits are then applied to the queue at device open time by the most
1093  * appropriate method for the kernel.
1094  *
1095  * zvol_queue_limits_convert() is used on 6.9+ (where the two-arg form of
1096  * blk_alloc_disk() exists). This converts our limits struct to a proper Linux
1097  * struct queue_limits, and passes it in. Any fields added in later kernels are
1098  * (obviously) not set up here.
1099  *
1100  * zvol_queue_limits_apply() is called on all kernel versions after the queue
1101  * is created, and applies any remaining config. Before 6.9 that will be
1102  * everything, via setter methods. After 6.9 that will be whatever couldn't be
1103  * put into struct queue_limits. (This implies that zvol_queue_limits_apply()
1104  * will always be a no-op on the latest kernel we support).
1105  */
1106 typedef struct zvol_queue_limits {
1107 	unsigned int	zql_max_hw_sectors;
1108 	unsigned short	zql_max_segments;
1109 	unsigned int	zql_max_segment_size;
1110 	unsigned int	zql_io_opt;
1111 	unsigned int	zql_physical_block_size;
1112 	unsigned int	zql_max_discard_sectors;
1113 	unsigned int	zql_discard_granularity;
1114 } zvol_queue_limits_t;
1115 
1116 static void
zvol_queue_limits_init(zvol_queue_limits_t * limits,zvol_state_t * zv,boolean_t use_blk_mq)1117 zvol_queue_limits_init(zvol_queue_limits_t *limits, zvol_state_t *zv,
1118     boolean_t use_blk_mq)
1119 {
1120 	limits->zql_max_hw_sectors = (DMU_MAX_ACCESS / 4) >> 9;
1121 
1122 	if (use_blk_mq) {
1123 		/*
1124 		 * IO requests can be really big (1MB).  When an IO request
1125 		 * comes in, it is passed off to zvol_read() or zvol_write()
1126 		 * in a new thread, where it is chunked up into 'volblocksize'
1127 		 * sized pieces and processed.  So for example, if the request
1128 		 * is a 1MB write and your volblocksize is 128k, one zvol_write
1129 		 * thread will take that request and sequentially do ten 128k
1130 		 * IOs.  This is due to the fact that the thread needs to lock
1131 		 * each volblocksize sized block.  So you might be wondering:
1132 		 * "instead of passing the whole 1MB request to one thread,
1133 		 * why not pass ten individual 128k chunks to ten threads and
1134 		 * process the whole write in parallel?"  The short answer is
1135 		 * that there's a sweet spot number of chunks that balances
1136 		 * the greater parallelism with the added overhead of more
1137 		 * threads. The sweet spot can be different depending on if you
1138 		 * have a read or write  heavy workload.  Writes typically want
1139 		 * high chunk counts while reads typically want lower ones.  On
1140 		 * a test pool with 6 NVMe drives in a 3x 2-disk mirror
1141 		 * configuration, with volblocksize=8k, the sweet spot for good
1142 		 * sequential reads and writes was at 8 chunks.
1143 		 */
1144 
1145 		/*
1146 		 * Below we tell the kernel how big we want our requests
1147 		 * to be.  You would think that blk_queue_io_opt() would be
1148 		 * used to do this since it is used to "set optimal request
1149 		 * size for the queue", but that doesn't seem to do
1150 		 * anything - the kernel still gives you huge requests
1151 		 * with tons of little PAGE_SIZE segments contained within it.
1152 		 *
1153 		 * Knowing that the kernel will just give you PAGE_SIZE segments
1154 		 * no matter what, you can say "ok, I want PAGE_SIZE byte
1155 		 * segments, and I want 'N' of them per request", where N is
1156 		 * the correct number of segments for the volblocksize and
1157 		 * number of chunks you want.
1158 		 */
1159 		if (zvol_blk_mq_blocks_per_thread != 0) {
1160 			unsigned int chunks;
1161 			chunks = MIN(zvol_blk_mq_blocks_per_thread, UINT16_MAX);
1162 
1163 			limits->zql_max_segment_size = PAGE_SIZE;
1164 			limits->zql_max_segments =
1165 			    (zv->zv_volblocksize * chunks) / PAGE_SIZE;
1166 		} else {
1167 			/*
1168 			 * Special case: zvol_blk_mq_blocks_per_thread = 0
1169 			 * Max everything out.
1170 			 */
1171 			limits->zql_max_segments = UINT16_MAX;
1172 			limits->zql_max_segment_size = UINT_MAX;
1173 		}
1174 	} else {
1175 		limits->zql_max_segments = UINT16_MAX;
1176 		limits->zql_max_segment_size = UINT_MAX;
1177 	}
1178 
1179 	limits->zql_io_opt = DMU_MAX_ACCESS / 2;
1180 
1181 	limits->zql_physical_block_size = zv->zv_volblocksize;
1182 	limits->zql_max_discard_sectors =
1183 	    (zvol_max_discard_blocks * zv->zv_volblocksize) >> 9;
1184 	limits->zql_discard_granularity = zv->zv_volblocksize;
1185 }
1186 
1187 #ifdef HAVE_BLK_ALLOC_DISK_2ARG
1188 static void
zvol_queue_limits_convert(zvol_queue_limits_t * limits,struct queue_limits * qlimits)1189 zvol_queue_limits_convert(zvol_queue_limits_t *limits,
1190     struct queue_limits *qlimits)
1191 {
1192 	memset(qlimits, 0, sizeof (struct queue_limits));
1193 	qlimits->max_hw_sectors = limits->zql_max_hw_sectors;
1194 	qlimits->max_segments = limits->zql_max_segments;
1195 	qlimits->max_segment_size = limits->zql_max_segment_size;
1196 	qlimits->io_opt = limits->zql_io_opt;
1197 	qlimits->physical_block_size = limits->zql_physical_block_size;
1198 	qlimits->max_discard_sectors = limits->zql_max_discard_sectors;
1199 	qlimits->max_hw_discard_sectors = limits->zql_max_discard_sectors;
1200 	qlimits->discard_granularity = limits->zql_discard_granularity;
1201 #ifdef HAVE_BLKDEV_QUEUE_LIMITS_FEATURES
1202 	qlimits->features =
1203 	    BLK_FEAT_WRITE_CACHE | BLK_FEAT_FUA | BLK_FEAT_IO_STAT;
1204 #endif
1205 }
1206 #endif
1207 
1208 static void
zvol_queue_limits_apply(zvol_queue_limits_t * limits,struct request_queue * queue)1209 zvol_queue_limits_apply(zvol_queue_limits_t *limits,
1210     struct request_queue *queue)
1211 {
1212 #ifndef HAVE_BLK_ALLOC_DISK_2ARG
1213 	blk_queue_max_hw_sectors(queue, limits->zql_max_hw_sectors);
1214 	blk_queue_max_segments(queue, limits->zql_max_segments);
1215 	blk_queue_max_segment_size(queue, limits->zql_max_segment_size);
1216 	blk_queue_io_opt(queue, limits->zql_io_opt);
1217 	blk_queue_physical_block_size(queue, limits->zql_physical_block_size);
1218 	blk_queue_max_discard_sectors(queue, limits->zql_max_discard_sectors);
1219 	blk_queue_discard_granularity(queue, limits->zql_discard_granularity);
1220 #endif
1221 #ifndef HAVE_BLKDEV_QUEUE_LIMITS_FEATURES
1222 	blk_queue_set_write_cache(queue, B_TRUE);
1223 	blk_queue_flag_set(QUEUE_FLAG_IO_STAT, queue);
1224 #endif
1225 }
1226 
1227 static int
zvol_alloc_non_blk_mq(struct zvol_state_os * zso,zvol_queue_limits_t * limits)1228 zvol_alloc_non_blk_mq(struct zvol_state_os *zso, zvol_queue_limits_t *limits)
1229 {
1230 #if defined(HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS)
1231 #if defined(HAVE_BLK_ALLOC_DISK)
1232 	zso->zvo_disk = blk_alloc_disk(NUMA_NO_NODE);
1233 	if (zso->zvo_disk == NULL)
1234 		return (1);
1235 
1236 	zso->zvo_disk->minors = ZVOL_MINORS;
1237 	zso->zvo_queue = zso->zvo_disk->queue;
1238 #elif defined(HAVE_BLK_ALLOC_DISK_2ARG)
1239 	struct queue_limits qlimits;
1240 	zvol_queue_limits_convert(limits, &qlimits);
1241 	struct gendisk *disk = blk_alloc_disk(&qlimits, NUMA_NO_NODE);
1242 	if (IS_ERR(disk)) {
1243 		zso->zvo_disk = NULL;
1244 		return (1);
1245 	}
1246 
1247 	zso->zvo_disk = disk;
1248 	zso->zvo_disk->minors = ZVOL_MINORS;
1249 	zso->zvo_queue = zso->zvo_disk->queue;
1250 
1251 #else
1252 	zso->zvo_queue = blk_alloc_queue(NUMA_NO_NODE);
1253 	if (zso->zvo_queue == NULL)
1254 		return (1);
1255 
1256 	zso->zvo_disk = alloc_disk(ZVOL_MINORS);
1257 	if (zso->zvo_disk == NULL) {
1258 		blk_cleanup_queue(zso->zvo_queue);
1259 		return (1);
1260 	}
1261 
1262 	zso->zvo_disk->queue = zso->zvo_queue;
1263 #endif /* HAVE_BLK_ALLOC_DISK */
1264 #else
1265 	zso->zvo_queue = blk_generic_alloc_queue(zvol_request, NUMA_NO_NODE);
1266 	if (zso->zvo_queue == NULL)
1267 		return (1);
1268 
1269 	zso->zvo_disk = alloc_disk(ZVOL_MINORS);
1270 	if (zso->zvo_disk == NULL) {
1271 		blk_cleanup_queue(zso->zvo_queue);
1272 		return (1);
1273 	}
1274 
1275 	zso->zvo_disk->queue = zso->zvo_queue;
1276 #endif /* HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS */
1277 
1278 	zvol_queue_limits_apply(limits, zso->zvo_queue);
1279 
1280 	return (0);
1281 
1282 }
1283 
1284 static int
zvol_alloc_blk_mq(zvol_state_t * zv,zvol_queue_limits_t * limits)1285 zvol_alloc_blk_mq(zvol_state_t *zv, zvol_queue_limits_t *limits)
1286 {
1287 	struct zvol_state_os *zso = zv->zv_zso;
1288 
1289 	/* Allocate our blk-mq tag_set */
1290 	if (zvol_blk_mq_alloc_tag_set(zv) != 0)
1291 		return (1);
1292 
1293 #if defined(HAVE_BLK_ALLOC_DISK)
1294 	zso->zvo_disk = blk_mq_alloc_disk(&zso->tag_set, zv);
1295 	if (zso->zvo_disk == NULL) {
1296 		blk_mq_free_tag_set(&zso->tag_set);
1297 		return (1);
1298 	}
1299 	zso->zvo_queue = zso->zvo_disk->queue;
1300 	zso->zvo_disk->minors = ZVOL_MINORS;
1301 #elif defined(HAVE_BLK_ALLOC_DISK_2ARG)
1302 	struct queue_limits qlimits;
1303 	zvol_queue_limits_convert(limits, &qlimits);
1304 	struct gendisk *disk = blk_mq_alloc_disk(&zso->tag_set, &qlimits, zv);
1305 	if (IS_ERR(disk)) {
1306 		zso->zvo_disk = NULL;
1307 		blk_mq_free_tag_set(&zso->tag_set);
1308 		return (1);
1309 	}
1310 
1311 	zso->zvo_disk = disk;
1312 	zso->zvo_queue = zso->zvo_disk->queue;
1313 	zso->zvo_disk->minors = ZVOL_MINORS;
1314 #else
1315 	zso->zvo_disk = alloc_disk(ZVOL_MINORS);
1316 	if (zso->zvo_disk == NULL) {
1317 		blk_cleanup_queue(zso->zvo_queue);
1318 		blk_mq_free_tag_set(&zso->tag_set);
1319 		return (1);
1320 	}
1321 	/* Allocate queue */
1322 	zso->zvo_queue = blk_mq_init_queue(&zso->tag_set);
1323 	if (IS_ERR(zso->zvo_queue)) {
1324 		blk_mq_free_tag_set(&zso->tag_set);
1325 		return (1);
1326 	}
1327 
1328 	/* Our queue is now created, assign it to our disk */
1329 	zso->zvo_disk->queue = zso->zvo_queue;
1330 #endif
1331 
1332 	zvol_queue_limits_apply(limits, zso->zvo_queue);
1333 
1334 	return (0);
1335 }
1336 
1337 /*
1338  * Allocate memory for a new zvol_state_t and setup the required
1339  * request queue and generic disk structures for the block device.
1340  */
1341 static zvol_state_t *
zvol_alloc(dev_t dev,const char * name,uint64_t volblocksize)1342 zvol_alloc(dev_t dev, const char *name, uint64_t volblocksize)
1343 {
1344 	zvol_state_t *zv;
1345 	struct zvol_state_os *zso;
1346 	uint64_t volmode;
1347 	int ret;
1348 
1349 	if (dsl_prop_get_integer(name, "volmode", &volmode, NULL) != 0)
1350 		return (NULL);
1351 
1352 	if (volmode == ZFS_VOLMODE_DEFAULT)
1353 		volmode = zvol_volmode;
1354 
1355 	if (volmode == ZFS_VOLMODE_NONE)
1356 		return (NULL);
1357 
1358 	zv = kmem_zalloc(sizeof (zvol_state_t), KM_SLEEP);
1359 	zso = kmem_zalloc(sizeof (struct zvol_state_os), KM_SLEEP);
1360 	zv->zv_zso = zso;
1361 	zv->zv_volmode = volmode;
1362 	zv->zv_volblocksize = volblocksize;
1363 
1364 	list_link_init(&zv->zv_next);
1365 	mutex_init(&zv->zv_state_lock, NULL, MUTEX_DEFAULT, NULL);
1366 	cv_init(&zv->zv_removing_cv, NULL, CV_DEFAULT, NULL);
1367 
1368 	zv->zv_zso->use_blk_mq = zvol_use_blk_mq;
1369 
1370 	zvol_queue_limits_t limits;
1371 	zvol_queue_limits_init(&limits, zv, zv->zv_zso->use_blk_mq);
1372 
1373 	/*
1374 	 * The block layer has 3 interfaces for getting BIOs:
1375 	 *
1376 	 * 1. blk-mq request queues (new)
1377 	 * 2. submit_bio() (oldest)
1378 	 * 3. regular request queues (old).
1379 	 *
1380 	 * Each of those interfaces has two permutations:
1381 	 *
1382 	 * a) We have blk_alloc_disk()/blk_mq_alloc_disk(), which allocates
1383 	 *    both the disk and its queue (5.14 kernel or newer)
1384 	 *
1385 	 * b) We don't have blk_*alloc_disk(), and have to allocate the
1386 	 *    disk and the queue separately. (5.13 kernel or older)
1387 	 */
1388 	if (zv->zv_zso->use_blk_mq) {
1389 		ret = zvol_alloc_blk_mq(zv, &limits);
1390 		zso->zvo_disk->fops = &zvol_ops_blk_mq;
1391 	} else {
1392 		ret = zvol_alloc_non_blk_mq(zso, &limits);
1393 		zso->zvo_disk->fops = &zvol_ops;
1394 	}
1395 	if (ret != 0)
1396 		goto out_kmem;
1397 
1398 	/* Limit read-ahead to a single page to prevent over-prefetching. */
1399 	blk_queue_set_read_ahead(zso->zvo_queue, 1);
1400 
1401 	if (!zv->zv_zso->use_blk_mq) {
1402 		/* Disable write merging in favor of the ZIO pipeline. */
1403 		blk_queue_flag_set(QUEUE_FLAG_NOMERGES, zso->zvo_queue);
1404 	}
1405 
1406 	zso->zvo_queue->queuedata = zv;
1407 	zso->zvo_dev = dev;
1408 	zv->zv_open_count = 0;
1409 	strlcpy(zv->zv_name, name, sizeof (zv->zv_name));
1410 
1411 	zfs_rangelock_init(&zv->zv_rangelock, NULL, NULL);
1412 	rw_init(&zv->zv_suspend_lock, NULL, RW_DEFAULT, NULL);
1413 
1414 	zso->zvo_disk->major = zvol_major;
1415 	zso->zvo_disk->events = DISK_EVENT_MEDIA_CHANGE;
1416 
1417 	/*
1418 	 * Setting ZFS_VOLMODE_DEV disables partitioning on ZVOL devices.
1419 	 * This is accomplished by limiting the number of minors for the
1420 	 * device to one and explicitly disabling partition scanning.
1421 	 */
1422 	if (volmode == ZFS_VOLMODE_DEV) {
1423 		zso->zvo_disk->minors = 1;
1424 		zso->zvo_disk->flags &= ~GENHD_FL_EXT_DEVT;
1425 		zso->zvo_disk->flags |= GENHD_FL_NO_PART;
1426 	}
1427 
1428 	zso->zvo_disk->first_minor = (dev & MINORMASK);
1429 	zso->zvo_disk->private_data = zv;
1430 	snprintf(zso->zvo_disk->disk_name, DISK_NAME_LEN, "%s%d",
1431 	    ZVOL_DEV_NAME, (dev & MINORMASK));
1432 
1433 	return (zv);
1434 
1435 out_kmem:
1436 	kmem_free(zso, sizeof (struct zvol_state_os));
1437 	kmem_free(zv, sizeof (zvol_state_t));
1438 	return (NULL);
1439 }
1440 
1441 /*
1442  * Cleanup then free a zvol_state_t which was created by zvol_alloc().
1443  * At this time, the structure is not opened by anyone, is taken off
1444  * the zvol_state_list, and has its private data set to NULL.
1445  * The zvol_state_lock is dropped.
1446  *
1447  * This function may take many milliseconds to complete (e.g. we've seen
1448  * it take over 256ms), due to the calls to "blk_cleanup_queue" and
1449  * "del_gendisk". Thus, consumers need to be careful to account for this
1450  * latency when calling this function.
1451  */
1452 void
zvol_os_free(zvol_state_t * zv)1453 zvol_os_free(zvol_state_t *zv)
1454 {
1455 
1456 	ASSERT(!RW_LOCK_HELD(&zv->zv_suspend_lock));
1457 	ASSERT(!MUTEX_HELD(&zv->zv_state_lock));
1458 	ASSERT0(zv->zv_open_count);
1459 	ASSERT3P(zv->zv_zso->zvo_disk->private_data, ==, NULL);
1460 
1461 	rw_destroy(&zv->zv_suspend_lock);
1462 	zfs_rangelock_fini(&zv->zv_rangelock);
1463 
1464 	del_gendisk(zv->zv_zso->zvo_disk);
1465 #if defined(HAVE_SUBMIT_BIO_IN_BLOCK_DEVICE_OPERATIONS) && \
1466 	(defined(HAVE_BLK_ALLOC_DISK) || defined(HAVE_BLK_ALLOC_DISK_2ARG))
1467 #if defined(HAVE_BLK_CLEANUP_DISK)
1468 	blk_cleanup_disk(zv->zv_zso->zvo_disk);
1469 #else
1470 	put_disk(zv->zv_zso->zvo_disk);
1471 #endif
1472 #else
1473 	blk_cleanup_queue(zv->zv_zso->zvo_queue);
1474 	put_disk(zv->zv_zso->zvo_disk);
1475 #endif
1476 
1477 	if (zv->zv_zso->use_blk_mq)
1478 		blk_mq_free_tag_set(&zv->zv_zso->tag_set);
1479 
1480 	ida_simple_remove(&zvol_ida,
1481 	    MINOR(zv->zv_zso->zvo_dev) >> ZVOL_MINOR_BITS);
1482 
1483 	cv_destroy(&zv->zv_removing_cv);
1484 	mutex_destroy(&zv->zv_state_lock);
1485 	dataset_kstats_destroy(&zv->zv_kstat);
1486 
1487 	kmem_free(zv->zv_zso, sizeof (struct zvol_state_os));
1488 	kmem_free(zv, sizeof (zvol_state_t));
1489 }
1490 
1491 void
zvol_wait_close(zvol_state_t * zv)1492 zvol_wait_close(zvol_state_t *zv)
1493 {
1494 }
1495 
1496 struct add_disk_work {
1497 	struct delayed_work work;
1498 	struct gendisk *disk;
1499 	int error;
1500 };
1501 
1502 static int
__zvol_os_add_disk(struct gendisk * disk)1503 __zvol_os_add_disk(struct gendisk *disk)
1504 {
1505 	int error = 0;
1506 #ifdef HAVE_ADD_DISK_RET
1507 	error = add_disk(disk);
1508 #else
1509 	add_disk(disk);
1510 #endif
1511 	return (error);
1512 }
1513 
1514 #if defined(HAVE_BDEV_FILE_OPEN_BY_PATH)
1515 static void
zvol_os_add_disk_work(struct work_struct * work)1516 zvol_os_add_disk_work(struct work_struct *work)
1517 {
1518 	struct add_disk_work *add_disk_work;
1519 	add_disk_work = container_of(work, struct add_disk_work, work.work);
1520 	add_disk_work->error = __zvol_os_add_disk(add_disk_work->disk);
1521 }
1522 #endif
1523 
1524 /*
1525  * SPECIAL CASE:
1526  *
1527  * This function basically calls add_disk() from a workqueue.   You may be
1528  * thinking: why not just call add_disk() directly?
1529  *
1530  * When you call add_disk(), the zvol appears to the world.  When this happens,
1531  * the kernel calls disk_scan_partitions() on the zvol, which behaves
1532  * differently on the 6.9+ kernels:
1533  *
1534  * - 6.8 and older kernels -
1535  * disk_scan_partitions()
1536  *	handle = bdev_open_by_dev(
1537  *		zvol_open()
1538  *	bdev_release(handle);
1539  *		zvol_release()
1540  *
1541  *
1542  * - 6.9+ kernels -
1543  * disk_scan_partitions()
1544  * 	file = bdev_file_open_by_dev()
1545  *		zvol_open()
1546  *	fput(file)
1547  *	< wait for return to userspace >
1548  *		zvol_release()
1549  *
1550  * The difference is that the bdev_release() from the 6.8 kernel is synchronous
1551  * while the fput() from the 6.9 kernel is async.  Or more specifically it's
1552  * async that has to wait until we return to userspace (since it adds the fput
1553  * into the caller's work queue with the TWA_RESUME flag set).  This is not the
1554  * behavior we want, since we want do things like create+destroy a zvol within
1555  * a single ZFS_IOC_CREATE ioctl, and the "create" part needs to release the
1556  * reference to the zvol while we're in the IOCTL, which can't wait until we
1557  * return to userspace.
1558  *
1559  * We can get around this since fput() has a special codepath for when it's
1560  * running in a kernel thread or interrupt.  In those cases, it just puts the
1561  * fput into the system workqueue, which we can force to run with
1562  * __flush_workqueue().  That is why we call add_disk() from a workqueue - so it
1563  * run from a kernel thread and "tricks" the fput() codepaths.
1564  *
1565  * Note that __flush_workqueue() is slowly getting deprecated.  This may be ok
1566  * though, since our IOCTL will spin on EBUSY waiting for the zvol release (via
1567  * fput) to happen, which it eventually, naturally, will from the system_wq
1568  * without us explicitly calling __flush_workqueue().
1569  */
1570 static int
zvol_os_add_disk(struct gendisk * disk)1571 zvol_os_add_disk(struct gendisk *disk)
1572 {
1573 #if defined(HAVE_BDEV_FILE_OPEN_BY_PATH)	/* 6.9+ kernel */
1574 	struct add_disk_work add_disk_work;
1575 
1576 	INIT_DELAYED_WORK(&add_disk_work.work, zvol_os_add_disk_work);
1577 	add_disk_work.disk = disk;
1578 	add_disk_work.error = 0;
1579 
1580 	/* Use *_delayed_work functions since they're not GPL'd */
1581 	schedule_delayed_work(&add_disk_work.work, 0);
1582 	flush_delayed_work(&add_disk_work.work);
1583 
1584 	__flush_workqueue(system_wq);
1585 	return (add_disk_work.error);
1586 #else	/* <= 6.8 kernel */
1587 	return (__zvol_os_add_disk(disk));
1588 #endif
1589 }
1590 
1591 /*
1592  * Create a block device minor node and setup the linkage between it
1593  * and the specified volume.  Once this function returns the block
1594  * device is live and ready for use.
1595  */
1596 int
zvol_os_create_minor(const char * name)1597 zvol_os_create_minor(const char *name)
1598 {
1599 	zvol_state_t *zv;
1600 	objset_t *os;
1601 	dmu_object_info_t *doi;
1602 	uint64_t volsize;
1603 	uint64_t len;
1604 	unsigned minor = 0;
1605 	int error = 0;
1606 	int idx;
1607 	uint64_t hash = zvol_name_hash(name);
1608 	uint64_t volthreading;
1609 	bool replayed_zil = B_FALSE;
1610 
1611 	if (zvol_inhibit_dev)
1612 		return (0);
1613 
1614 	idx = ida_simple_get(&zvol_ida, 0, 0, kmem_flags_convert(KM_SLEEP));
1615 	if (idx < 0)
1616 		return (SET_ERROR(-idx));
1617 	minor = idx << ZVOL_MINOR_BITS;
1618 	if (MINOR(minor) != minor) {
1619 		/* too many partitions can cause an overflow */
1620 		zfs_dbgmsg("zvol: create minor overflow: %s, minor %u/%u",
1621 		    name, minor, MINOR(minor));
1622 		ida_simple_remove(&zvol_ida, idx);
1623 		return (SET_ERROR(EINVAL));
1624 	}
1625 
1626 	zv = zvol_find_by_name_hash(name, hash, RW_NONE);
1627 	if (zv) {
1628 		ASSERT(MUTEX_HELD(&zv->zv_state_lock));
1629 		mutex_exit(&zv->zv_state_lock);
1630 		ida_simple_remove(&zvol_ida, idx);
1631 		return (SET_ERROR(EEXIST));
1632 	}
1633 
1634 	doi = kmem_alloc(sizeof (dmu_object_info_t), KM_SLEEP);
1635 
1636 	error = dmu_objset_own(name, DMU_OST_ZVOL, B_TRUE, B_TRUE, FTAG, &os);
1637 	if (error)
1638 		goto out_doi;
1639 
1640 	error = dmu_object_info(os, ZVOL_OBJ, doi);
1641 	if (error)
1642 		goto out_dmu_objset_disown;
1643 
1644 	error = zap_lookup(os, ZVOL_ZAP_OBJ, "size", 8, 1, &volsize);
1645 	if (error)
1646 		goto out_dmu_objset_disown;
1647 
1648 	zv = zvol_alloc(MKDEV(zvol_major, minor), name,
1649 	    doi->doi_data_block_size);
1650 	if (zv == NULL) {
1651 		error = SET_ERROR(EAGAIN);
1652 		goto out_dmu_objset_disown;
1653 	}
1654 	zv->zv_hash = hash;
1655 
1656 	if (dmu_objset_is_snapshot(os))
1657 		zv->zv_flags |= ZVOL_RDONLY;
1658 
1659 	zv->zv_volsize = volsize;
1660 	zv->zv_objset = os;
1661 
1662 	/* Default */
1663 	zv->zv_threading = B_TRUE;
1664 	if (dsl_prop_get_integer(name, "volthreading", &volthreading, NULL)
1665 	    == 0)
1666 		zv->zv_threading = volthreading;
1667 
1668 	set_capacity(zv->zv_zso->zvo_disk, zv->zv_volsize >> 9);
1669 
1670 #ifdef QUEUE_FLAG_DISCARD
1671 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, zv->zv_zso->zvo_queue);
1672 #endif
1673 #ifdef QUEUE_FLAG_NONROT
1674 	blk_queue_flag_set(QUEUE_FLAG_NONROT, zv->zv_zso->zvo_queue);
1675 #endif
1676 #ifdef QUEUE_FLAG_ADD_RANDOM
1677 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, zv->zv_zso->zvo_queue);
1678 #endif
1679 	/* This flag was introduced in kernel version 4.12. */
1680 #ifdef QUEUE_FLAG_SCSI_PASSTHROUGH
1681 	blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, zv->zv_zso->zvo_queue);
1682 #endif
1683 
1684 	ASSERT3P(zv->zv_kstat.dk_kstats, ==, NULL);
1685 	error = dataset_kstats_create(&zv->zv_kstat, zv->zv_objset);
1686 	if (error)
1687 		goto out_dmu_objset_disown;
1688 	ASSERT3P(zv->zv_zilog, ==, NULL);
1689 	zv->zv_zilog = zil_open(os, zvol_get_data, &zv->zv_kstat.dk_zil_sums);
1690 	if (spa_writeable(dmu_objset_spa(os))) {
1691 		if (zil_replay_disable)
1692 			replayed_zil = zil_destroy(zv->zv_zilog, B_FALSE);
1693 		else
1694 			replayed_zil = zil_replay(os, zv, zvol_replay_vector);
1695 	}
1696 	if (replayed_zil)
1697 		zil_close(zv->zv_zilog);
1698 	zv->zv_zilog = NULL;
1699 
1700 	/*
1701 	 * When udev detects the addition of the device it will immediately
1702 	 * invoke blkid(8) to determine the type of content on the device.
1703 	 * Prefetching the blocks commonly scanned by blkid(8) will speed
1704 	 * up this process.
1705 	 */
1706 	len = MIN(zvol_prefetch_bytes, SPA_MAXBLOCKSIZE);
1707 	if (len > 0) {
1708 		dmu_prefetch(os, ZVOL_OBJ, 0, 0, len, ZIO_PRIORITY_SYNC_READ);
1709 		dmu_prefetch(os, ZVOL_OBJ, 0, volsize - len, len,
1710 		    ZIO_PRIORITY_SYNC_READ);
1711 	}
1712 
1713 	zv->zv_objset = NULL;
1714 out_dmu_objset_disown:
1715 	dmu_objset_disown(os, B_TRUE, FTAG);
1716 out_doi:
1717 	kmem_free(doi, sizeof (dmu_object_info_t));
1718 
1719 	/*
1720 	 * Keep in mind that once add_disk() is called, the zvol is
1721 	 * announced to the world, and zvol_open()/zvol_release() can
1722 	 * be called at any time. Incidentally, add_disk() itself calls
1723 	 * zvol_open()->zvol_first_open() and zvol_release()->zvol_last_close()
1724 	 * directly as well.
1725 	 */
1726 	if (error == 0) {
1727 		rw_enter(&zvol_state_lock, RW_WRITER);
1728 		zvol_insert(zv);
1729 		rw_exit(&zvol_state_lock);
1730 		error = zvol_os_add_disk(zv->zv_zso->zvo_disk);
1731 	} else {
1732 		ida_simple_remove(&zvol_ida, idx);
1733 	}
1734 
1735 	return (error);
1736 }
1737 
1738 void
zvol_os_rename_minor(zvol_state_t * zv,const char * newname)1739 zvol_os_rename_minor(zvol_state_t *zv, const char *newname)
1740 {
1741 	int readonly = get_disk_ro(zv->zv_zso->zvo_disk);
1742 
1743 	ASSERT(RW_LOCK_HELD(&zvol_state_lock));
1744 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
1745 
1746 	strlcpy(zv->zv_name, newname, sizeof (zv->zv_name));
1747 
1748 	/* move to new hashtable entry  */
1749 	zv->zv_hash = zvol_name_hash(newname);
1750 	hlist_del(&zv->zv_hlink);
1751 	hlist_add_head(&zv->zv_hlink, ZVOL_HT_HEAD(zv->zv_hash));
1752 
1753 	/*
1754 	 * The block device's read-only state is briefly changed causing
1755 	 * a KOBJ_CHANGE uevent to be issued.  This ensures udev detects
1756 	 * the name change and fixes the symlinks.  This does not change
1757 	 * ZVOL_RDONLY in zv->zv_flags so the actual read-only state never
1758 	 * changes.  This would normally be done using kobject_uevent() but
1759 	 * that is a GPL-only symbol which is why we need this workaround.
1760 	 */
1761 	set_disk_ro(zv->zv_zso->zvo_disk, !readonly);
1762 	set_disk_ro(zv->zv_zso->zvo_disk, readonly);
1763 
1764 	dataset_kstats_rename(&zv->zv_kstat, newname);
1765 }
1766 
1767 void
zvol_os_set_disk_ro(zvol_state_t * zv,int flags)1768 zvol_os_set_disk_ro(zvol_state_t *zv, int flags)
1769 {
1770 
1771 	set_disk_ro(zv->zv_zso->zvo_disk, flags);
1772 }
1773 
1774 void
zvol_os_set_capacity(zvol_state_t * zv,uint64_t capacity)1775 zvol_os_set_capacity(zvol_state_t *zv, uint64_t capacity)
1776 {
1777 
1778 	set_capacity(zv->zv_zso->zvo_disk, capacity);
1779 }
1780 
1781 int
zvol_init(void)1782 zvol_init(void)
1783 {
1784 	int error;
1785 
1786 	/*
1787 	 * zvol_threads is the module param the user passes in.
1788 	 *
1789 	 * zvol_actual_threads is what we use internally, since the user can
1790 	 * pass zvol_thread = 0 to mean "use all the CPUs" (the default).
1791 	 */
1792 	static unsigned int zvol_actual_threads;
1793 
1794 	if (zvol_threads == 0) {
1795 		/*
1796 		 * See dde9380a1 for why 32 was chosen here.  This should
1797 		 * probably be refined to be some multiple of the number
1798 		 * of CPUs.
1799 		 */
1800 		zvol_actual_threads = MAX(num_online_cpus(), 32);
1801 	} else {
1802 		zvol_actual_threads = MIN(MAX(zvol_threads, 1), 1024);
1803 	}
1804 
1805 	/*
1806 	 * Use atleast 32 zvol_threads but for many core system,
1807 	 * prefer 6 threads per taskq, but no more taskqs
1808 	 * than threads in them on large systems.
1809 	 *
1810 	 *                 taskq   total
1811 	 * cpus    taskqs  threads threads
1812 	 * ------- ------- ------- -------
1813 	 * 1       1       32       32
1814 	 * 2       1       32       32
1815 	 * 4       1       32       32
1816 	 * 8       2       16       32
1817 	 * 16      3       11       33
1818 	 * 32      5       7        35
1819 	 * 64      8       8        64
1820 	 * 128     11      12       132
1821 	 * 256     16      16       256
1822 	 */
1823 	zv_taskq_t *ztqs = &zvol_taskqs;
1824 	uint_t num_tqs = MIN(num_online_cpus(), zvol_num_taskqs);
1825 	if (num_tqs == 0) {
1826 		num_tqs = 1 + num_online_cpus() / 6;
1827 		while (num_tqs * num_tqs > zvol_actual_threads)
1828 			num_tqs--;
1829 	}
1830 	uint_t per_tq_thread = zvol_actual_threads / num_tqs;
1831 	if (per_tq_thread * num_tqs < zvol_actual_threads)
1832 		per_tq_thread++;
1833 	ztqs->tqs_cnt = num_tqs;
1834 	ztqs->tqs_taskq = kmem_alloc(num_tqs * sizeof (taskq_t *), KM_SLEEP);
1835 	error = register_blkdev(zvol_major, ZVOL_DRIVER);
1836 	if (error) {
1837 		kmem_free(ztqs->tqs_taskq, ztqs->tqs_cnt * sizeof (taskq_t *));
1838 		ztqs->tqs_taskq = NULL;
1839 		printk(KERN_INFO "ZFS: register_blkdev() failed %d\n", error);
1840 		return (error);
1841 	}
1842 
1843 	if (zvol_blk_mq_queue_depth == 0) {
1844 		zvol_actual_blk_mq_queue_depth = BLKDEV_DEFAULT_RQ;
1845 	} else {
1846 		zvol_actual_blk_mq_queue_depth =
1847 		    MAX(zvol_blk_mq_queue_depth, BLKDEV_MIN_RQ);
1848 	}
1849 
1850 	if (zvol_blk_mq_threads == 0) {
1851 		zvol_blk_mq_actual_threads = num_online_cpus();
1852 	} else {
1853 		zvol_blk_mq_actual_threads = MIN(MAX(zvol_blk_mq_threads, 1),
1854 		    1024);
1855 	}
1856 
1857 	for (uint_t i = 0; i < num_tqs; i++) {
1858 		char name[32];
1859 		(void) snprintf(name, sizeof (name), "%s_tq-%u",
1860 		    ZVOL_DRIVER, i);
1861 		ztqs->tqs_taskq[i] = taskq_create(name, per_tq_thread,
1862 		    maxclsyspri, per_tq_thread, INT_MAX,
1863 		    TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
1864 		if (ztqs->tqs_taskq[i] == NULL) {
1865 			for (int j = i - 1; j >= 0; j--)
1866 				taskq_destroy(ztqs->tqs_taskq[j]);
1867 			unregister_blkdev(zvol_major, ZVOL_DRIVER);
1868 			kmem_free(ztqs->tqs_taskq, ztqs->tqs_cnt *
1869 			    sizeof (taskq_t *));
1870 			ztqs->tqs_taskq = NULL;
1871 			return (-ENOMEM);
1872 		}
1873 	}
1874 
1875 	zvol_init_impl();
1876 	ida_init(&zvol_ida);
1877 	return (0);
1878 }
1879 
1880 void
zvol_fini(void)1881 zvol_fini(void)
1882 {
1883 	zv_taskq_t *ztqs = &zvol_taskqs;
1884 	zvol_fini_impl();
1885 	unregister_blkdev(zvol_major, ZVOL_DRIVER);
1886 
1887 	if (ztqs->tqs_taskq == NULL) {
1888 		ASSERT3U(ztqs->tqs_cnt, ==, 0);
1889 	} else {
1890 		for (uint_t i = 0; i < ztqs->tqs_cnt; i++) {
1891 			ASSERT3P(ztqs->tqs_taskq[i], !=, NULL);
1892 			taskq_destroy(ztqs->tqs_taskq[i]);
1893 		}
1894 		kmem_free(ztqs->tqs_taskq, ztqs->tqs_cnt *
1895 		    sizeof (taskq_t *));
1896 		ztqs->tqs_taskq = NULL;
1897 	}
1898 
1899 	ida_destroy(&zvol_ida);
1900 }
1901 
1902 /* BEGIN CSTYLED */
1903 module_param(zvol_inhibit_dev, uint, 0644);
1904 MODULE_PARM_DESC(zvol_inhibit_dev, "Do not create zvol device nodes");
1905 
1906 module_param(zvol_major, uint, 0444);
1907 MODULE_PARM_DESC(zvol_major, "Major number for zvol device");
1908 
1909 module_param(zvol_threads, uint, 0444);
1910 MODULE_PARM_DESC(zvol_threads, "Number of threads to handle I/O requests. Set"
1911     "to 0 to use all active CPUs");
1912 
1913 module_param(zvol_request_sync, uint, 0644);
1914 MODULE_PARM_DESC(zvol_request_sync, "Synchronously handle bio requests");
1915 
1916 module_param(zvol_max_discard_blocks, ulong, 0444);
1917 MODULE_PARM_DESC(zvol_max_discard_blocks, "Max number of blocks to discard");
1918 
1919 module_param(zvol_num_taskqs, uint, 0444);
1920 MODULE_PARM_DESC(zvol_num_taskqs, "Number of zvol taskqs");
1921 
1922 module_param(zvol_prefetch_bytes, uint, 0644);
1923 MODULE_PARM_DESC(zvol_prefetch_bytes, "Prefetch N bytes at zvol start+end");
1924 
1925 module_param(zvol_volmode, uint, 0644);
1926 MODULE_PARM_DESC(zvol_volmode, "Default volmode property value");
1927 
1928 module_param(zvol_blk_mq_queue_depth, uint, 0644);
1929 MODULE_PARM_DESC(zvol_blk_mq_queue_depth, "Default blk-mq queue depth");
1930 
1931 module_param(zvol_use_blk_mq, uint, 0644);
1932 MODULE_PARM_DESC(zvol_use_blk_mq, "Use the blk-mq API for zvols");
1933 
1934 module_param(zvol_blk_mq_blocks_per_thread, uint, 0644);
1935 MODULE_PARM_DESC(zvol_blk_mq_blocks_per_thread,
1936     "Process volblocksize blocks per thread");
1937 
1938 #ifndef HAVE_BLKDEV_GET_ERESTARTSYS
1939 module_param(zvol_open_timeout_ms, uint, 0644);
1940 MODULE_PARM_DESC(zvol_open_timeout_ms, "Timeout for ZVOL open retries");
1941 #endif
1942 
1943 /* END CSTYLED */
1944