xref: /freebsd/sys/contrib/openzfs/module/os/linux/zfs/vdev_disk.c (revision 113e60742ef6ba5c069aa737ee57ba3c2f88b248)
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) 2008-2010 Lawrence Livermore National Security, LLC.
24  * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
25  * Rewritten for Linux by Brian Behlendorf <behlendorf1@llnl.gov>.
26  * LLNL-CODE-403049.
27  * Copyright (c) 2012, 2019 by Delphix. All rights reserved.
28  * Copyright (c) 2023, 2024, 2025, Klara, Inc.
29  */
30 
31 #include <sys/zfs_context.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_disk.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/vdev_trim.h>
36 #include <sys/abd.h>
37 #include <sys/fs/zfs.h>
38 #include <sys/zio.h>
39 #include <linux/blkpg.h>
40 #include <linux/msdos_fs.h>
41 #include <linux/vfs_compat.h>
42 #include <linux/blk-cgroup.h>
43 
44 /*
45  * Linux 6.8.x uses a bdev_handle as an instance/refcount for an underlying
46  * block_device. Since it carries the block_device inside, its convenient to
47  * just use the handle as a proxy.
48  *
49  * Linux 6.9.x uses a file for the same purpose.
50  *
51  * For pre-6.8, we just emulate this with a cast, since we don't need any of
52  * the other fields inside the handle.
53  */
54 #if defined(HAVE_BDEV_OPEN_BY_PATH)
55 typedef struct bdev_handle zfs_bdev_handle_t;
56 #define	BDH_BDEV(bdh)		((bdh)->bdev)
57 #define	BDH_IS_ERR(bdh)		(IS_ERR(bdh))
58 #define	BDH_PTR_ERR(bdh)	(PTR_ERR(bdh))
59 #define	BDH_ERR_PTR(err)	(ERR_PTR(err))
60 #elif defined(HAVE_BDEV_FILE_OPEN_BY_PATH)
61 typedef struct file zfs_bdev_handle_t;
62 #define	BDH_BDEV(bdh)		(file_bdev(bdh))
63 #define	BDH_IS_ERR(bdh)		(IS_ERR(bdh))
64 #define	BDH_PTR_ERR(bdh)	(PTR_ERR(bdh))
65 #define	BDH_ERR_PTR(err)	(ERR_PTR(err))
66 #else
67 typedef void zfs_bdev_handle_t;
68 #define	BDH_BDEV(bdh)		((struct block_device *)bdh)
69 #define	BDH_IS_ERR(bdh)		(IS_ERR(BDH_BDEV(bdh)))
70 #define	BDH_PTR_ERR(bdh)	(PTR_ERR(BDH_BDEV(bdh)))
71 #define	BDH_ERR_PTR(err)	(ERR_PTR(err))
72 #endif
73 
74 typedef struct vdev_disk {
75 	zfs_bdev_handle_t		*vd_bdh;
76 	krwlock_t			vd_lock;
77 } vdev_disk_t;
78 
79 /*
80  * Maximum number of segments to add to a bio (min 4). If this is higher than
81  * the maximum allowed by the device queue or the kernel itself, it will be
82  * clamped. Setting it to zero will cause the kernel's ideal size to be used.
83  */
84 uint_t zfs_vdev_disk_max_segs = 0;
85 
86 /*
87  * Unique identifier for the exclusive vdev holder.
88  */
89 static void *zfs_vdev_holder = VDEV_HOLDER;
90 
91 /*
92  * Wait up to zfs_vdev_open_timeout_ms milliseconds before determining the
93  * device is missing. The missing path may be transient since the links
94  * can be briefly removed and recreated in response to udev events.
95  */
96 static uint_t zfs_vdev_open_timeout_ms = 1000;
97 
98 /*
99  * Size of the "reserved" partition, in blocks.
100  */
101 #define	EFI_MIN_RESV_SIZE	(16 * 1024)
102 
103 /*
104  * BIO request failfast mask.
105  */
106 
107 static unsigned int zfs_vdev_failfast_mask = 1;
108 
109 /*
110  * Convert SPA mode flags into bdev open mode flags.
111  */
112 #ifdef HAVE_BLK_MODE_T
113 typedef blk_mode_t vdev_bdev_mode_t;
114 #define	VDEV_BDEV_MODE_READ	BLK_OPEN_READ
115 #define	VDEV_BDEV_MODE_WRITE	BLK_OPEN_WRITE
116 #define	VDEV_BDEV_MODE_EXCL	BLK_OPEN_EXCL
117 #define	VDEV_BDEV_MODE_MASK	(BLK_OPEN_READ|BLK_OPEN_WRITE|BLK_OPEN_EXCL)
118 #else
119 typedef fmode_t vdev_bdev_mode_t;
120 #define	VDEV_BDEV_MODE_READ	FMODE_READ
121 #define	VDEV_BDEV_MODE_WRITE	FMODE_WRITE
122 #define	VDEV_BDEV_MODE_EXCL	FMODE_EXCL
123 #define	VDEV_BDEV_MODE_MASK	(FMODE_READ|FMODE_WRITE|FMODE_EXCL)
124 #endif
125 
126 static vdev_bdev_mode_t
vdev_bdev_mode(spa_mode_t smode)127 vdev_bdev_mode(spa_mode_t smode)
128 {
129 	ASSERT3U(smode, !=, SPA_MODE_UNINIT);
130 	ASSERT0(smode & ~(SPA_MODE_READ|SPA_MODE_WRITE));
131 
132 	vdev_bdev_mode_t bmode = VDEV_BDEV_MODE_EXCL;
133 
134 	if (smode & SPA_MODE_READ)
135 		bmode |= VDEV_BDEV_MODE_READ;
136 
137 	if (smode & SPA_MODE_WRITE)
138 		bmode |= VDEV_BDEV_MODE_WRITE;
139 
140 	ASSERT(bmode & VDEV_BDEV_MODE_MASK);
141 	ASSERT0(bmode & ~VDEV_BDEV_MODE_MASK);
142 
143 	return (bmode);
144 }
145 
146 /*
147  * Returns the usable capacity (in bytes) for the partition or disk.
148  */
149 static uint64_t
bdev_capacity(struct block_device * bdev)150 bdev_capacity(struct block_device *bdev)
151 {
152 #ifdef HAVE_BDEV_NR_BYTES
153 	return (bdev_nr_bytes(bdev));
154 #else
155 	return (i_size_read(bdev->bd_inode));
156 #endif
157 }
158 
159 #if !defined(HAVE_BDEV_WHOLE)
160 static inline struct block_device *
bdev_whole(struct block_device * bdev)161 bdev_whole(struct block_device *bdev)
162 {
163 	return (bdev->bd_contains);
164 }
165 #endif
166 
167 #if defined(HAVE_BDEVNAME)
168 #define	vdev_bdevname(bdev, name)	bdevname(bdev, name)
169 #else
170 static inline void
vdev_bdevname(struct block_device * bdev,char * name)171 vdev_bdevname(struct block_device *bdev, char *name)
172 {
173 	snprintf(name, BDEVNAME_SIZE, "%pg", bdev);
174 }
175 #endif
176 
177 /*
178  * Returns the maximum expansion capacity of the block device (in bytes).
179  *
180  * It is possible to expand a vdev when it has been created as a wholedisk
181  * and the containing block device has increased in capacity.  Or when the
182  * partition containing the pool has been manually increased in size.
183  *
184  * This function is only responsible for calculating the potential expansion
185  * size so it can be reported by 'zpool list'.  The efi_use_whole_disk() is
186  * responsible for verifying the expected partition layout in the wholedisk
187  * case, and updating the partition table if appropriate.  Once the partition
188  * size has been increased the additional capacity will be visible using
189  * bdev_capacity().
190  *
191  * The returned maximum expansion capacity is always expected to be larger, or
192  * at the very least equal, to its usable capacity to prevent overestimating
193  * the pool expandsize.
194  */
195 static uint64_t
bdev_max_capacity(struct block_device * bdev,uint64_t wholedisk)196 bdev_max_capacity(struct block_device *bdev, uint64_t wholedisk)
197 {
198 	uint64_t psize;
199 	int64_t available;
200 
201 	if (wholedisk && bdev != bdev_whole(bdev)) {
202 		/*
203 		 * When reporting maximum expansion capacity for a wholedisk
204 		 * deduct any capacity which is expected to be lost due to
205 		 * alignment restrictions.  Over reporting this value isn't
206 		 * harmful and would only result in slightly less capacity
207 		 * than expected post expansion.
208 		 * The estimated available space may be slightly smaller than
209 		 * bdev_capacity() for devices where the number of sectors is
210 		 * not a multiple of the alignment size and the partition layout
211 		 * is keeping less than PARTITION_END_ALIGNMENT bytes after the
212 		 * "reserved" EFI partition: in such cases return the device
213 		 * usable capacity.
214 		 */
215 		available = bdev_capacity(bdev_whole(bdev)) -
216 		    ((EFI_MIN_RESV_SIZE + NEW_START_BLOCK +
217 		    PARTITION_END_ALIGNMENT) << SECTOR_BITS);
218 		psize = MAX(available, bdev_capacity(bdev));
219 	} else {
220 		psize = bdev_capacity(bdev);
221 	}
222 
223 	return (psize);
224 }
225 
226 static void
vdev_disk_error(zio_t * zio)227 vdev_disk_error(zio_t *zio)
228 {
229 	/*
230 	 * This function can be called in interrupt context, for instance while
231 	 * handling IRQs coming from a misbehaving disk device; use printk()
232 	 * which is safe from any context.
233 	 */
234 	printk(KERN_WARNING "zio pool=%s vdev=%s error=%d type=%d "
235 	    "offset=%llu size=%llu flags=%llu\n", spa_name(zio->io_spa),
236 	    zio->io_vd->vdev_path, zio->io_error, zio->io_type,
237 	    (u_longlong_t)zio->io_offset, (u_longlong_t)zio->io_size,
238 	    zio->io_flags);
239 }
240 
241 static void
vdev_disk_kobj_evt_post(vdev_t * v)242 vdev_disk_kobj_evt_post(vdev_t *v)
243 {
244 	vdev_disk_t *vd = v->vdev_tsd;
245 	if (vd && vd->vd_bdh) {
246 		spl_signal_kobj_evt(BDH_BDEV(vd->vd_bdh));
247 	} else {
248 		vdev_dbgmsg(v, "vdev_disk_t is NULL for VDEV:%s\n",
249 		    v->vdev_path);
250 	}
251 }
252 
253 static zfs_bdev_handle_t *
vdev_blkdev_get_by_path(const char * path,spa_mode_t smode,void * holder)254 vdev_blkdev_get_by_path(const char *path, spa_mode_t smode, void *holder)
255 {
256 	vdev_bdev_mode_t bmode = vdev_bdev_mode(smode);
257 
258 #if defined(HAVE_BDEV_FILE_OPEN_BY_PATH)
259 	return (bdev_file_open_by_path(path, bmode, holder, NULL));
260 #elif defined(HAVE_BDEV_OPEN_BY_PATH)
261 	return (bdev_open_by_path(path, bmode, holder, NULL));
262 #elif defined(HAVE_BLKDEV_GET_BY_PATH_4ARG)
263 	return (blkdev_get_by_path(path, bmode, holder, NULL));
264 #else
265 	return (blkdev_get_by_path(path, bmode, holder));
266 #endif
267 }
268 
269 static void
vdev_blkdev_put(zfs_bdev_handle_t * bdh,spa_mode_t smode,void * holder)270 vdev_blkdev_put(zfs_bdev_handle_t *bdh, spa_mode_t smode, void *holder)
271 {
272 #if defined(HAVE_BDEV_RELEASE)
273 	return (bdev_release(bdh));
274 #elif defined(HAVE_BLKDEV_PUT_HOLDER)
275 	return (blkdev_put(BDH_BDEV(bdh), holder));
276 #elif defined(HAVE_BLKDEV_PUT)
277 	return (blkdev_put(BDH_BDEV(bdh), vdev_bdev_mode(smode)));
278 #else
279 	fput(bdh);
280 #endif
281 }
282 
283 static int
vdev_disk_open(vdev_t * v,uint64_t * psize,uint64_t * max_psize,uint64_t * logical_ashift,uint64_t * physical_ashift)284 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
285     uint64_t *logical_ashift, uint64_t *physical_ashift)
286 {
287 	zfs_bdev_handle_t *bdh;
288 	spa_mode_t smode = spa_mode(v->vdev_spa);
289 	hrtime_t timeout = MSEC2NSEC(zfs_vdev_open_timeout_ms);
290 	vdev_disk_t *vd;
291 
292 	/* Must have a pathname and it must be absolute. */
293 	if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
294 		v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
295 		vdev_dbgmsg(v, "invalid vdev_path");
296 		return (SET_ERROR(EINVAL));
297 	}
298 
299 	/*
300 	 * Reopen the device if it is currently open.  When expanding a
301 	 * partition force re-scanning the partition table if userland
302 	 * did not take care of this already. We need to do this while closed
303 	 * in order to get an accurate updated block device size.  Then
304 	 * since udev may need to recreate the device links increase the
305 	 * open retry timeout before reporting the device as unavailable.
306 	 */
307 	vd = v->vdev_tsd;
308 	if (vd) {
309 		char disk_name[BDEVNAME_SIZE + 6] = "/dev/";
310 		boolean_t reread_part = B_FALSE;
311 
312 		rw_enter(&vd->vd_lock, RW_WRITER);
313 		bdh = vd->vd_bdh;
314 		vd->vd_bdh = NULL;
315 
316 		if (bdh) {
317 			struct block_device *bdev = BDH_BDEV(bdh);
318 			if (v->vdev_expanding && bdev != bdev_whole(bdev)) {
319 				vdev_bdevname(bdev_whole(bdev), disk_name + 5);
320 				/*
321 				 * If userland has BLKPG_RESIZE_PARTITION,
322 				 * then it should have updated the partition
323 				 * table already. We can detect this by
324 				 * comparing our current physical size
325 				 * with that of the device. If they are
326 				 * the same, then we must not have
327 				 * BLKPG_RESIZE_PARTITION or it failed to
328 				 * update the partition table online. We
329 				 * fallback to rescanning the partition
330 				 * table from the kernel below. However,
331 				 * if the capacity already reflects the
332 				 * updated partition, then we skip
333 				 * rescanning the partition table here.
334 				 */
335 				if (v->vdev_psize == bdev_capacity(bdev))
336 					reread_part = B_TRUE;
337 			}
338 
339 			vdev_blkdev_put(bdh, smode, zfs_vdev_holder);
340 		}
341 
342 		if (reread_part) {
343 			bdh = vdev_blkdev_get_by_path(disk_name, smode,
344 			    zfs_vdev_holder);
345 			if (!BDH_IS_ERR(bdh)) {
346 				int error =
347 				    vdev_bdev_reread_part(BDH_BDEV(bdh));
348 				vdev_blkdev_put(bdh, smode, zfs_vdev_holder);
349 				if (error == 0) {
350 					timeout = MSEC2NSEC(
351 					    zfs_vdev_open_timeout_ms * 2);
352 				}
353 			}
354 		}
355 	} else {
356 		vd = kmem_zalloc(sizeof (vdev_disk_t), KM_SLEEP);
357 
358 		rw_init(&vd->vd_lock, NULL, RW_DEFAULT, NULL);
359 		rw_enter(&vd->vd_lock, RW_WRITER);
360 	}
361 
362 	/*
363 	 * Devices are always opened by the path provided at configuration
364 	 * time.  This means that if the provided path is a udev by-id path
365 	 * then drives may be re-cabled without an issue.  If the provided
366 	 * path is a udev by-path path, then the physical location information
367 	 * will be preserved.  This can be critical for more complicated
368 	 * configurations where drives are located in specific physical
369 	 * locations to maximize the systems tolerance to component failure.
370 	 *
371 	 * Alternatively, you can provide your own udev rule to flexibly map
372 	 * the drives as you see fit.  It is not advised that you use the
373 	 * /dev/[hd]d devices which may be reordered due to probing order.
374 	 * Devices in the wrong locations will be detected by the higher
375 	 * level vdev validation.
376 	 *
377 	 * The specified paths may be briefly removed and recreated in
378 	 * response to udev events.  This should be exceptionally unlikely
379 	 * because the zpool command makes every effort to verify these paths
380 	 * have already settled prior to reaching this point.  Therefore,
381 	 * a ENOENT failure at this point is highly likely to be transient
382 	 * and it is reasonable to sleep and retry before giving up.  In
383 	 * practice delays have been observed to be on the order of 100ms.
384 	 *
385 	 * When ERESTARTSYS is returned it indicates the block device is
386 	 * a zvol which could not be opened due to the deadlock detection
387 	 * logic in zvol_open().  Extend the timeout and retry the open
388 	 * subsequent attempts are expected to eventually succeed.
389 	 */
390 	hrtime_t start = gethrtime();
391 	bdh = BDH_ERR_PTR(-ENXIO);
392 	while (BDH_IS_ERR(bdh) && ((gethrtime() - start) < timeout)) {
393 		bdh = vdev_blkdev_get_by_path(v->vdev_path, smode,
394 		    zfs_vdev_holder);
395 		if (unlikely(BDH_PTR_ERR(bdh) == -ENOENT)) {
396 			/*
397 			 * There is no point of waiting since device is removed
398 			 * explicitly
399 			 */
400 			if (v->vdev_removed)
401 				break;
402 
403 			schedule_timeout_interruptible(MSEC_TO_TICK(10));
404 		} else if (unlikely(BDH_PTR_ERR(bdh) == -ERESTARTSYS)) {
405 			timeout = MSEC2NSEC(zfs_vdev_open_timeout_ms * 10);
406 			continue;
407 		} else if (BDH_IS_ERR(bdh)) {
408 			break;
409 		}
410 	}
411 
412 	if (BDH_IS_ERR(bdh)) {
413 		int error = -BDH_PTR_ERR(bdh);
414 		vdev_dbgmsg(v, "open error=%d timeout=%llu/%llu", error,
415 		    (u_longlong_t)(gethrtime() - start),
416 		    (u_longlong_t)timeout);
417 		vd->vd_bdh = NULL;
418 		v->vdev_tsd = vd;
419 		rw_exit(&vd->vd_lock);
420 		return (SET_ERROR(error));
421 	} else {
422 		vd->vd_bdh = bdh;
423 		v->vdev_tsd = vd;
424 		rw_exit(&vd->vd_lock);
425 	}
426 
427 	struct block_device *bdev = BDH_BDEV(vd->vd_bdh);
428 
429 	/*  Determine the physical block size */
430 	int physical_block_size = bdev_physical_block_size(bdev);
431 
432 	/*  Determine the logical block size */
433 	int logical_block_size = bdev_logical_block_size(bdev);
434 
435 	/*
436 	 * If the device has a write cache, clear the nowritecache flag,
437 	 * so that we start issuing flush requests again.
438 	 */
439 	v->vdev_nowritecache = !zfs_bdev_has_write_cache(bdev);
440 
441 	/* Set when device reports it supports TRIM. */
442 	v->vdev_has_trim = bdev_discard_supported(bdev);
443 
444 	/* Set when device reports it supports secure TRIM. */
445 	v->vdev_has_securetrim = bdev_secure_discard_supported(bdev);
446 
447 	/* Inform the ZIO pipeline that we are non-rotational */
448 	v->vdev_nonrot = blk_queue_nonrot(bdev_get_queue(bdev));
449 
450 	/* Physical volume size in bytes for the partition */
451 	*psize = bdev_capacity(bdev);
452 
453 	/* Physical volume size in bytes including possible expansion space */
454 	*max_psize = bdev_max_capacity(bdev, v->vdev_wholedisk);
455 
456 	/* Based on the minimum sector size set the block size */
457 	*physical_ashift = highbit64(MAX(physical_block_size,
458 	    SPA_MINBLOCKSIZE)) - 1;
459 
460 	*logical_ashift = highbit64(MAX(logical_block_size,
461 	    SPA_MINBLOCKSIZE)) - 1;
462 
463 	return (0);
464 }
465 
466 static void
vdev_disk_close(vdev_t * v)467 vdev_disk_close(vdev_t *v)
468 {
469 	vdev_disk_t *vd = v->vdev_tsd;
470 
471 	if (v->vdev_reopening || vd == NULL)
472 		return;
473 
474 	rw_enter(&vd->vd_lock, RW_WRITER);
475 
476 	if (vd->vd_bdh != NULL)
477 		vdev_blkdev_put(vd->vd_bdh, spa_mode(v->vdev_spa),
478 		    zfs_vdev_holder);
479 
480 	v->vdev_tsd = NULL;
481 
482 	rw_exit(&vd->vd_lock);
483 	rw_destroy(&vd->vd_lock);
484 	kmem_free(vd, sizeof (vdev_disk_t));
485 }
486 
487 /*
488  * preempt_schedule_notrace is GPL-only which breaks the ZFS build, so
489  * replace it with preempt_schedule under the following condition:
490  */
491 #if defined(CONFIG_ARM64) && \
492     defined(CONFIG_PREEMPTION) && \
493     defined(CONFIG_BLK_CGROUP)
494 #define	preempt_schedule_notrace(x) preempt_schedule(x)
495 #endif
496 
497 /*
498  * As for the Linux 5.18 kernel bio_alloc() expects a block_device struct
499  * as an argument removing the need to set it with bio_set_dev().  This
500  * removes the need for all of the following compatibility code.
501  */
502 #if !defined(HAVE_BIO_ALLOC_4ARG)
503 
504 #if defined(CONFIG_BLK_CGROUP) && defined(HAVE_BIO_SET_DEV_GPL_ONLY)
505 /*
506  * The Linux 5.5 kernel updated percpu_ref_tryget() which is inlined by
507  * blkg_tryget() to use rcu_read_lock() instead of rcu_read_lock_sched().
508  * As a side effect the function was converted to GPL-only.  Define our
509  * own version when needed which uses rcu_read_lock_sched().
510  *
511  * The Linux 5.17 kernel split linux/blk-cgroup.h into a private and a public
512  * part, moving blkg_tryget into the private one. Define our own version.
513  */
514 #if defined(HAVE_BLKG_TRYGET_GPL_ONLY) || !defined(HAVE_BLKG_TRYGET)
515 static inline bool
vdev_blkg_tryget(struct blkcg_gq * blkg)516 vdev_blkg_tryget(struct blkcg_gq *blkg)
517 {
518 	struct percpu_ref *ref = &blkg->refcnt;
519 	unsigned long __percpu *count;
520 	bool rc;
521 
522 	rcu_read_lock_sched();
523 
524 	if (__ref_is_percpu(ref, &count)) {
525 		this_cpu_inc(*count);
526 		rc = true;
527 	} else {
528 #ifdef ZFS_PERCPU_REF_COUNT_IN_DATA
529 		rc = atomic_long_inc_not_zero(&ref->data->count);
530 #else
531 		rc = atomic_long_inc_not_zero(&ref->count);
532 #endif
533 	}
534 
535 	rcu_read_unlock_sched();
536 
537 	return (rc);
538 }
539 #else
540 #define	vdev_blkg_tryget(bg)	blkg_tryget(bg)
541 #endif
542 #ifdef HAVE_BIO_SET_DEV_MACRO
543 /*
544  * The Linux 5.0 kernel updated the bio_set_dev() macro so it calls the
545  * GPL-only bio_associate_blkg() symbol thus inadvertently converting
546  * the entire macro.  Provide a minimal version which always assigns the
547  * request queue's root_blkg to the bio.
548  */
549 static inline void
vdev_bio_associate_blkg(struct bio * bio)550 vdev_bio_associate_blkg(struct bio *bio)
551 {
552 #if defined(HAVE_BIO_BDEV_DISK)
553 	struct request_queue *q = bio->bi_bdev->bd_disk->queue;
554 #else
555 	struct request_queue *q = bio->bi_disk->queue;
556 #endif
557 
558 	ASSERT3P(q, !=, NULL);
559 	ASSERT0P(bio->bi_blkg);
560 
561 	if (q->root_blkg && vdev_blkg_tryget(q->root_blkg))
562 		bio->bi_blkg = q->root_blkg;
563 }
564 
565 #define	bio_associate_blkg vdev_bio_associate_blkg
566 #else
567 static inline void
vdev_bio_set_dev(struct bio * bio,struct block_device * bdev)568 vdev_bio_set_dev(struct bio *bio, struct block_device *bdev)
569 {
570 #if defined(HAVE_BIO_BDEV_DISK)
571 	struct request_queue *q = bdev->bd_disk->queue;
572 #else
573 	struct request_queue *q = bio->bi_disk->queue;
574 #endif
575 	bio_clear_flag(bio, BIO_REMAPPED);
576 	if (bio->bi_bdev != bdev)
577 		bio_clear_flag(bio, BIO_THROTTLED);
578 	bio->bi_bdev = bdev;
579 
580 	ASSERT3P(q, !=, NULL);
581 	ASSERT0P(bio->bi_blkg);
582 
583 	if (q->root_blkg && vdev_blkg_tryget(q->root_blkg))
584 		bio->bi_blkg = q->root_blkg;
585 }
586 #define	bio_set_dev		vdev_bio_set_dev
587 #endif
588 #endif
589 #endif /* !HAVE_BIO_ALLOC_4ARG */
590 
591 static inline void
vdev_submit_bio(struct bio * bio)592 vdev_submit_bio(struct bio *bio)
593 {
594 	struct bio_list *bio_list = current->bio_list;
595 	current->bio_list = NULL;
596 	(void) submit_bio(bio);
597 	current->bio_list = bio_list;
598 }
599 
600 static inline struct bio *
vdev_bio_alloc(struct block_device * bdev,gfp_t gfp_mask,unsigned short nr_vecs)601 vdev_bio_alloc(struct block_device *bdev, gfp_t gfp_mask,
602     unsigned short nr_vecs)
603 {
604 	struct bio *bio;
605 
606 #ifdef HAVE_BIO_ALLOC_4ARG
607 	bio = bio_alloc(bdev, nr_vecs, 0, gfp_mask);
608 #else
609 	bio = bio_alloc(gfp_mask, nr_vecs);
610 	if (likely(bio != NULL))
611 		bio_set_dev(bio, bdev);
612 #endif
613 
614 	return (bio);
615 }
616 
617 static inline uint_t
vdev_bio_max_segs(struct block_device * bdev)618 vdev_bio_max_segs(struct block_device *bdev)
619 {
620 	/*
621 	 * Smallest of the device max segs and the tunable max segs. Minimum
622 	 * 4, so there's room to finish split pages if they come up.
623 	 */
624 	const uint_t dev_max_segs = queue_max_segments(bdev_get_queue(bdev));
625 	const uint_t tune_max_segs = (zfs_vdev_disk_max_segs > 0) ?
626 	    MAX(4, zfs_vdev_disk_max_segs) : dev_max_segs;
627 	const uint_t max_segs = MIN(tune_max_segs, dev_max_segs);
628 
629 #ifdef HAVE_BIO_MAX_SEGS
630 	return (bio_max_segs(max_segs));
631 #else
632 	return (MIN(max_segs, BIO_MAX_PAGES));
633 #endif
634 }
635 
636 static inline uint_t
vdev_bio_max_bytes(struct block_device * bdev)637 vdev_bio_max_bytes(struct block_device *bdev)
638 {
639 	return (queue_max_sectors(bdev_get_queue(bdev)) << 9);
640 }
641 
642 
643 /*
644  * Virtual block IO object (VBIO)
645  *
646  * Linux block IO (BIO) objects have a limit on how many data segments (pages)
647  * they can hold. Depending on how they're allocated and structured, a large
648  * ZIO can require more than one BIO to be submitted to the kernel, which then
649  * all have to complete before we can return the completed ZIO back to ZFS.
650  *
651  * A VBIO is a wrapper around multiple BIOs, carrying everything needed to
652  * translate a ZIO down into the kernel block layer and back again.
653  *
654  * Note that these are only used for data ZIOs (read/write). Meta-operations
655  * (flush/trim) don't need multiple BIOs and so can just make the call
656  * directly.
657  */
658 typedef struct {
659 	zio_t		*vbio_zio;	/* parent zio */
660 
661 	struct block_device *vbio_bdev;	/* blockdev to submit bios to */
662 
663 	abd_t		*vbio_abd;	/* abd carrying borrowed linear buf */
664 
665 	uint_t		vbio_max_segs;	/* max segs per bio */
666 
667 	uint_t		vbio_max_bytes;	/* max bytes per bio */
668 	uint_t		vbio_lbs_mask;	/* logical block size mask */
669 
670 	uint64_t	vbio_offset;	/* start offset of next bio */
671 
672 	struct bio	*vbio_bio;	/* pointer to the current bio */
673 	int		vbio_flags;	/* bio flags */
674 } vbio_t;
675 
676 static vbio_t *
vbio_alloc(zio_t * zio,struct block_device * bdev,int flags)677 vbio_alloc(zio_t *zio, struct block_device *bdev, int flags)
678 {
679 	vbio_t *vbio = kmem_zalloc(sizeof (vbio_t), KM_SLEEP);
680 
681 	vbio->vbio_zio = zio;
682 	vbio->vbio_bdev = bdev;
683 	vbio->vbio_abd = NULL;
684 	vbio->vbio_max_segs = vdev_bio_max_segs(bdev);
685 	vbio->vbio_max_bytes = vdev_bio_max_bytes(bdev);
686 	vbio->vbio_lbs_mask = ~(bdev_logical_block_size(bdev)-1);
687 	vbio->vbio_offset = zio->io_offset;
688 	vbio->vbio_bio = NULL;
689 	vbio->vbio_flags = flags;
690 
691 	return (vbio);
692 }
693 
694 static void vbio_completion(struct bio *bio);
695 
696 static int
vbio_add_page(vbio_t * vbio,struct page * page,uint_t size,uint_t offset)697 vbio_add_page(vbio_t *vbio, struct page *page, uint_t size, uint_t offset)
698 {
699 	struct bio *bio = vbio->vbio_bio;
700 	uint_t ssize;
701 
702 	while (size > 0) {
703 		if (bio == NULL) {
704 			/* New BIO, allocate and set up */
705 			bio = vdev_bio_alloc(vbio->vbio_bdev, GFP_NOIO,
706 			    vbio->vbio_max_segs);
707 			VERIFY(bio);
708 
709 			BIO_BI_SECTOR(bio) = vbio->vbio_offset >> 9;
710 			bio_set_op_attrs(bio,
711 			    vbio->vbio_zio->io_type == ZIO_TYPE_WRITE ?
712 			    WRITE : READ, vbio->vbio_flags);
713 
714 			if (vbio->vbio_bio) {
715 				bio_chain(vbio->vbio_bio, bio);
716 				vdev_submit_bio(vbio->vbio_bio);
717 			}
718 			vbio->vbio_bio = bio;
719 		}
720 
721 		/*
722 		 * Only load as much of the current page data as will fit in
723 		 * the space left in the BIO, respecting lbs alignment. Older
724 		 * kernels will error if we try to overfill the BIO, while
725 		 * newer ones will accept it and split the BIO. This ensures
726 		 * everything works on older kernels, and avoids an additional
727 		 * overhead on the new.
728 		 */
729 		ssize = MIN(size, (vbio->vbio_max_bytes - BIO_BI_SIZE(bio)) &
730 		    vbio->vbio_lbs_mask);
731 		if (ssize > 0 &&
732 		    bio_add_page(bio, page, ssize, offset) == ssize) {
733 			/* Accepted, adjust and load any remaining. */
734 			size -= ssize;
735 			offset += ssize;
736 			continue;
737 		}
738 
739 		/* No room, set up for a new BIO and loop */
740 		vbio->vbio_offset += BIO_BI_SIZE(bio);
741 
742 		/* Signal new BIO allocation wanted */
743 		bio = NULL;
744 	}
745 
746 	return (0);
747 }
748 
749 /* Iterator callback to submit ABD pages to the vbio. */
750 static int
vbio_fill_cb(struct page * page,size_t off,size_t len,void * priv)751 vbio_fill_cb(struct page *page, size_t off, size_t len, void *priv)
752 {
753 	vbio_t *vbio = priv;
754 	return (vbio_add_page(vbio, page, len, off));
755 }
756 
757 /* Create some BIOs, fill them with data and submit them */
758 static void
vbio_submit(vbio_t * vbio,abd_t * abd,uint64_t size)759 vbio_submit(vbio_t *vbio, abd_t *abd, uint64_t size)
760 {
761 	/*
762 	 * We plug so we can submit the BIOs as we go and only unplug them when
763 	 * they are fully created and submitted. This is important; if we don't
764 	 * plug, then the kernel may start executing earlier BIOs while we're
765 	 * still creating and executing later ones, and if the device goes
766 	 * away while that's happening, older kernels can get confused and
767 	 * trample memory.
768 	 */
769 	struct blk_plug plug;
770 	blk_start_plug(&plug);
771 
772 	(void) abd_iterate_page_func(abd, 0, size, vbio_fill_cb, vbio);
773 	ASSERT(vbio->vbio_bio);
774 
775 	vbio->vbio_bio->bi_end_io = vbio_completion;
776 	vbio->vbio_bio->bi_private = vbio;
777 
778 	/*
779 	 * Once submitted, vbio_bio now owns vbio (through bi_private) and we
780 	 * can't touch it again. The bio may complete and vbio_completion() be
781 	 * called and free the vbio before this task is run again, so we must
782 	 * consider it invalid from this point.
783 	 */
784 	vdev_submit_bio(vbio->vbio_bio);
785 
786 	blk_finish_plug(&plug);
787 }
788 
789 /* IO completion callback */
790 static void
vbio_completion(struct bio * bio)791 vbio_completion(struct bio *bio)
792 {
793 	vbio_t *vbio = bio->bi_private;
794 	zio_t *zio = vbio->vbio_zio;
795 
796 	ASSERT(zio);
797 
798 	/* Capture and log any errors */
799 	zio->io_error = bi_status_to_errno(bio->bi_status);
800 	ASSERT3U(zio->io_error, >=, 0);
801 
802 	if (zio->io_error)
803 		vdev_disk_error(zio);
804 
805 	/* Return the BIO to the kernel */
806 	bio_put(bio);
807 
808 	/*
809 	 * We're likely in an interrupt context so we can't do ABD/memory work
810 	 * here; instead we stash vbio on the zio and take care of it in the
811 	 * done callback.
812 	 */
813 	ASSERT0P(zio->io_bio);
814 	zio->io_bio = vbio;
815 
816 	zio_delay_interrupt(zio);
817 }
818 
819 /*
820  * Iterator callback to count ABD pages and check their size & alignment.
821  *
822  * On Linux, each BIO segment can take a page pointer, and an offset+length of
823  * the data within that page. A page can be arbitrarily large ("compound"
824  * pages) but we still have to ensure the data portion is correctly sized and
825  * aligned to the logical block size, to ensure that if the kernel wants to
826  * split the BIO, the two halves will still be properly aligned.
827  *
828  * NOTE: if you change this function, change the copy in
829  * tests/zfs-tests/tests/functional/vdev_disk/page_alignment.c, and add test
830  * data there to validate the change you're making.
831  */
832 typedef struct {
833 	size_t	blocksize;
834 	int	seen_first;
835 	int	seen_last;
836 } vdev_disk_check_alignment_t;
837 
838 static int
vdev_disk_check_alignment_cb(struct page * page,size_t off,size_t len,void * priv)839 vdev_disk_check_alignment_cb(struct page *page, size_t off, size_t len,
840     void *priv)
841 {
842 	(void) page;
843 	vdev_disk_check_alignment_t *s = priv;
844 
845 	/*
846 	 * The cardinal rule: a single on-disk block must never cross an
847 	 * physical (order-0) page boundary, as the kernel expects to be able
848 	 * to split at both LBS and page boundaries.
849 	 *
850 	 * This implies various alignment rules for the blocks in this
851 	 * (possibly compound) page, which we can check for.
852 	 */
853 
854 	/*
855 	 * If the previous page did not end on a page boundary, then we
856 	 * can't proceed without creating a hole.
857 	 */
858 	if (s->seen_last)
859 		return (1);
860 
861 	/* This page must contain only whole LBS-sized blocks. */
862 	if (!IS_P2ALIGNED(len, s->blocksize))
863 		return (1);
864 
865 	/*
866 	 * If this is not the first page in the ABD, then the data must start
867 	 * on a page-aligned boundary (so the kernel can split on page
868 	 * boundaries without having to deal with a hole). If it is, then
869 	 * it can start on LBS-alignment.
870 	 */
871 	if (s->seen_first) {
872 		if (!IS_P2ALIGNED(off, PAGESIZE))
873 			return (1);
874 	} else {
875 		if (!IS_P2ALIGNED(off, s->blocksize))
876 			return (1);
877 		s->seen_first = 1;
878 	}
879 
880 	/*
881 	 * If this data does not end on a page-aligned boundary, then this
882 	 * must be the last page in the ABD, for the same reason.
883 	 */
884 	s->seen_last = !IS_P2ALIGNED(off+len, PAGESIZE);
885 
886 	return (0);
887 }
888 
889 /*
890  * Check if we can submit the pages in this ABD to the kernel as-is. Returns
891  * the number of pages, or 0 if it can't be submitted like this.
892  */
893 static boolean_t
vdev_disk_check_alignment(abd_t * abd,uint64_t size,struct block_device * bdev)894 vdev_disk_check_alignment(abd_t *abd, uint64_t size, struct block_device *bdev)
895 {
896 	vdev_disk_check_alignment_t s = {
897 	    .blocksize = bdev_logical_block_size(bdev),
898 	};
899 
900 	if (abd_iterate_page_func(abd, 0, size,
901 	    vdev_disk_check_alignment_cb, &s))
902 		return (B_FALSE);
903 
904 	return (B_TRUE);
905 }
906 
907 static int
vdev_disk_io_rw(zio_t * zio)908 vdev_disk_io_rw(zio_t *zio)
909 {
910 	vdev_t *v = zio->io_vd;
911 	vdev_disk_t *vd = v->vdev_tsd;
912 	struct block_device *bdev = BDH_BDEV(vd->vd_bdh);
913 	int flags = 0;
914 
915 	/*
916 	 * Accessing outside the block device is never allowed.
917 	 */
918 	if (zio->io_offset + zio->io_size > bdev_capacity(bdev)) {
919 		vdev_dbgmsg(zio->io_vd,
920 		    "Illegal access %llu size %llu, device size %llu",
921 		    (u_longlong_t)zio->io_offset,
922 		    (u_longlong_t)zio->io_size,
923 		    (u_longlong_t)bdev_capacity(bdev));
924 		return (SET_ERROR(EIO));
925 	}
926 
927 	if (!(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)) &&
928 	    v->vdev_failfast == B_TRUE) {
929 		bio_set_flags_failfast(bdev, &flags, zfs_vdev_failfast_mask & 1,
930 		    zfs_vdev_failfast_mask & 2, zfs_vdev_failfast_mask & 4);
931 	}
932 
933 	/*
934 	 * Check alignment of the incoming ABD. If any part of it would require
935 	 * submitting a page that is not aligned to both the logical block size
936 	 * and the page size, then we take a copy into a new memory region with
937 	 * correct alignment.  This should be impossible on a 512b LBS. On
938 	 * larger blocks, this can happen at least when a small number of
939 	 * blocks (usually 1) are allocated from a shared slab, or when
940 	 * abnormally-small data regions (eg gang headers) are mixed into the
941 	 * same ABD as larger allocations (eg aggregations).
942 	 */
943 	abd_t *abd = zio->io_abd;
944 	if (!vdev_disk_check_alignment(abd, zio->io_size, bdev)) {
945 		/* Allocate a new memory region with guaranteed alignment */
946 		abd = abd_alloc_for_io(zio->io_size,
947 		    zio->io_abd->abd_flags & ABD_FLAG_META);
948 
949 		/* If we're writing copy our data into it */
950 		if (zio->io_type == ZIO_TYPE_WRITE)
951 			abd_copy(abd, zio->io_abd, zio->io_size);
952 
953 		/*
954 		 * False here would mean the new allocation has an invalid
955 		 * alignment too, which would mean that abd_alloc() is not
956 		 * guaranteeing this, or our logic in
957 		 * vdev_disk_check_alignment() is wrong. In either case,
958 		 * something in seriously wrong and its not safe to continue.
959 		 */
960 		VERIFY(vdev_disk_check_alignment(abd, zio->io_size, bdev));
961 	}
962 
963 	/* Allocate vbio, with a pointer to the borrowed ABD if necessary */
964 	vbio_t *vbio = vbio_alloc(zio, bdev, flags);
965 	if (abd != zio->io_abd)
966 		vbio->vbio_abd = abd;
967 
968 	/* Fill it with data pages and submit it to the kernel */
969 	vbio_submit(vbio, abd, zio->io_size);
970 	return (0);
971 }
972 
973 static void
vdev_disk_io_flush_completion(struct bio * bio)974 vdev_disk_io_flush_completion(struct bio *bio)
975 {
976 	zio_t *zio = bio->bi_private;
977 	zio->io_error = bi_status_to_errno(bio->bi_status);
978 	if (zio->io_error == EOPNOTSUPP || zio->io_error == ENOTTY)
979 		zio->io_error = SET_ERROR(ENOTSUP);
980 
981 	bio_put(bio);
982 	ASSERT3S(zio->io_error, >=, 0);
983 	if (zio->io_error)
984 		vdev_disk_error(zio);
985 	zio_interrupt(zio);
986 }
987 
988 static int
vdev_disk_io_flush(struct block_device * bdev,zio_t * zio)989 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
990 {
991 	struct request_queue *q;
992 	struct bio *bio;
993 
994 	q = bdev_get_queue(bdev);
995 	if (!q)
996 		return (SET_ERROR(ENXIO));
997 
998 	bio = vdev_bio_alloc(bdev, GFP_NOIO, 0);
999 	if (unlikely(bio == NULL))
1000 		return (SET_ERROR(ENOMEM));
1001 
1002 	bio->bi_end_io = vdev_disk_io_flush_completion;
1003 	bio->bi_private = zio;
1004 	bio_set_flush(bio);
1005 	vdev_submit_bio(bio);
1006 	invalidate_bdev(bdev);
1007 
1008 	return (0);
1009 }
1010 
1011 static void
vdev_disk_discard_end_io(struct bio * bio)1012 vdev_disk_discard_end_io(struct bio *bio)
1013 {
1014 	zio_t *zio = bio->bi_private;
1015 	zio->io_error = bi_status_to_errno(bio->bi_status);
1016 
1017 	bio_put(bio);
1018 	if (zio->io_error)
1019 		vdev_disk_error(zio);
1020 	zio_interrupt(zio);
1021 }
1022 
1023 /*
1024  * Wrappers for the different secure erase and discard APIs. We use async
1025  * when available; in this case, *biop is set to the last bio in the chain.
1026  */
1027 static int
vdev_bdev_issue_secure_erase(zfs_bdev_handle_t * bdh,sector_t sector,sector_t nsect,struct bio ** biop)1028 vdev_bdev_issue_secure_erase(zfs_bdev_handle_t *bdh, sector_t sector,
1029     sector_t nsect, struct bio **biop)
1030 {
1031 	*biop = NULL;
1032 	int error;
1033 
1034 #if defined(HAVE_BLKDEV_ISSUE_SECURE_ERASE)
1035 	error = blkdev_issue_secure_erase(BDH_BDEV(bdh),
1036 	    sector, nsect, GFP_NOFS);
1037 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_ASYNC_FLAGS)
1038 	error = __blkdev_issue_discard(BDH_BDEV(bdh),
1039 	    sector, nsect, GFP_NOFS, BLKDEV_DISCARD_SECURE, biop);
1040 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_FLAGS)
1041 	error = blkdev_issue_discard(BDH_BDEV(bdh),
1042 	    sector, nsect, GFP_NOFS, BLKDEV_DISCARD_SECURE);
1043 #else
1044 #error "unsupported kernel"
1045 #endif
1046 
1047 	return (error);
1048 }
1049 
1050 static int
vdev_bdev_issue_discard(zfs_bdev_handle_t * bdh,sector_t sector,sector_t nsect,struct bio ** biop)1051 vdev_bdev_issue_discard(zfs_bdev_handle_t *bdh, sector_t sector,
1052     sector_t nsect, struct bio **biop)
1053 {
1054 	*biop = NULL;
1055 	int error;
1056 
1057 #if defined(HAVE_BLKDEV_ISSUE_DISCARD_ASYNC_FLAGS)
1058 	error = __blkdev_issue_discard(BDH_BDEV(bdh),
1059 	    sector, nsect, GFP_NOFS, 0, biop);
1060 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_ASYNC_NOFLAGS)
1061 	error = __blkdev_issue_discard(BDH_BDEV(bdh),
1062 	    sector, nsect, GFP_NOFS, biop);
1063 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_FLAGS)
1064 	error = blkdev_issue_discard(BDH_BDEV(bdh),
1065 	    sector, nsect, GFP_NOFS, 0);
1066 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_NOFLAGS)
1067 	error = blkdev_issue_discard(BDH_BDEV(bdh),
1068 	    sector, nsect, GFP_NOFS);
1069 #else
1070 #error "unsupported kernel"
1071 #endif
1072 
1073 	return (error);
1074 }
1075 
1076 /*
1077  * Entry point for TRIM ops. This calls the right wrapper for secure erase or
1078  * discard, and then does the appropriate finishing work for error vs success
1079  * and async vs sync.
1080  */
1081 static int
vdev_disk_io_trim(zio_t * zio)1082 vdev_disk_io_trim(zio_t *zio)
1083 {
1084 	int error;
1085 	struct bio *bio;
1086 
1087 	zfs_bdev_handle_t *bdh = ((vdev_disk_t *)zio->io_vd->vdev_tsd)->vd_bdh;
1088 	sector_t sector = zio->io_offset >> 9;
1089 	sector_t nsects = zio->io_size >> 9;
1090 
1091 	if (zio->io_trim_flags & ZIO_TRIM_SECURE)
1092 		error = vdev_bdev_issue_secure_erase(bdh, sector, nsects, &bio);
1093 	else
1094 		error = vdev_bdev_issue_discard(bdh, sector, nsects, &bio);
1095 
1096 	if (error != 0)
1097 		return (SET_ERROR(-error));
1098 
1099 	if (bio == NULL) {
1100 		/*
1101 		 * This was a synchronous op that completed successfully, so
1102 		 * return it to ZFS immediately.
1103 		 */
1104 		zio_interrupt(zio);
1105 	} else {
1106 		/*
1107 		 * This was an asynchronous op; set up completion callback and
1108 		 * issue it.
1109 		 */
1110 		bio->bi_private = zio;
1111 		bio->bi_end_io = vdev_disk_discard_end_io;
1112 		vdev_submit_bio(bio);
1113 	}
1114 
1115 	return (0);
1116 }
1117 
1118 static void
vdev_disk_io_start(zio_t * zio)1119 vdev_disk_io_start(zio_t *zio)
1120 {
1121 	vdev_t *v = zio->io_vd;
1122 	vdev_disk_t *vd = v->vdev_tsd;
1123 	int error;
1124 
1125 	/*
1126 	 * If the vdev is closed, it's likely in the REMOVED or FAULTED state.
1127 	 * Nothing to be done here but return failure.
1128 	 */
1129 	if (vd == NULL) {
1130 		zio->io_error = ENXIO;
1131 		zio_interrupt(zio);
1132 		return;
1133 	}
1134 
1135 	rw_enter(&vd->vd_lock, RW_READER);
1136 
1137 	/*
1138 	 * If the vdev is closed, it's likely due to a failed reopen and is
1139 	 * in the UNAVAIL state.  Nothing to be done here but return failure.
1140 	 */
1141 	if (vd->vd_bdh == NULL) {
1142 		rw_exit(&vd->vd_lock);
1143 		zio->io_error = ENXIO;
1144 		zio_interrupt(zio);
1145 		return;
1146 	}
1147 
1148 	switch (zio->io_type) {
1149 	case ZIO_TYPE_FLUSH:
1150 
1151 		if (!vdev_readable(v)) {
1152 			/* Drive not there, can't flush */
1153 			error = SET_ERROR(ENXIO);
1154 		} else if (zfs_nocacheflush) {
1155 			/* Flushing disabled by operator, declare success */
1156 			error = 0;
1157 		} else if (v->vdev_nowritecache) {
1158 			/* This vdev not capable of flushing */
1159 			error = SET_ERROR(ENOTSUP);
1160 		} else {
1161 			/*
1162 			 * Issue the flush. If successful, the response will
1163 			 * be handled in the completion callback, so we're done.
1164 			 */
1165 			error = vdev_disk_io_flush(BDH_BDEV(vd->vd_bdh), zio);
1166 			if (error == 0) {
1167 				rw_exit(&vd->vd_lock);
1168 				return;
1169 			}
1170 		}
1171 
1172 		/* Couldn't issue the flush, so set the error and return it */
1173 		rw_exit(&vd->vd_lock);
1174 		zio->io_error = error;
1175 		zio_execute(zio);
1176 		return;
1177 
1178 	case ZIO_TYPE_TRIM:
1179 		error = vdev_disk_io_trim(zio);
1180 		rw_exit(&vd->vd_lock);
1181 		if (error) {
1182 			zio->io_error = error;
1183 			zio_execute(zio);
1184 		}
1185 		return;
1186 
1187 	case ZIO_TYPE_READ:
1188 	case ZIO_TYPE_WRITE:
1189 		zio->io_target_timestamp = zio_handle_io_delay(zio);
1190 		error = vdev_disk_io_rw(zio);
1191 		rw_exit(&vd->vd_lock);
1192 		if (error) {
1193 			zio->io_error = error;
1194 			zio_interrupt(zio);
1195 		}
1196 		return;
1197 
1198 	default:
1199 		/*
1200 		 * Getting here means our parent vdev has made a very strange
1201 		 * request of us, and shouldn't happen. Assert here to force a
1202 		 * crash in dev builds, but in production return the IO
1203 		 * unhandled. The pool will likely suspend anyway but that's
1204 		 * nicer than crashing the kernel.
1205 		 */
1206 		ASSERT3S(zio->io_type, ==, -1);
1207 
1208 		rw_exit(&vd->vd_lock);
1209 		zio->io_error = SET_ERROR(ENOTSUP);
1210 		zio_interrupt(zio);
1211 		return;
1212 	}
1213 
1214 	__builtin_unreachable();
1215 }
1216 
1217 static void
vdev_disk_io_done(zio_t * zio)1218 vdev_disk_io_done(zio_t *zio)
1219 {
1220 	/* If this was a read or write, we need to clean up the vbio */
1221 	if (zio->io_bio != NULL) {
1222 		vbio_t *vbio = zio->io_bio;
1223 		zio->io_bio = NULL;
1224 
1225 		/*
1226 		 * If we copied the ABD before issuing it, clean up and return
1227 		 * the copy to the ADB, with changes if appropriate.
1228 		 */
1229 		if (vbio->vbio_abd != NULL) {
1230 			if (zio->io_type == ZIO_TYPE_READ)
1231 				abd_copy(zio->io_abd, vbio->vbio_abd,
1232 				    zio->io_size);
1233 
1234 			abd_free(vbio->vbio_abd);
1235 			vbio->vbio_abd = NULL;
1236 		}
1237 
1238 		/* Final cleanup */
1239 		kmem_free(vbio, sizeof (vbio_t));
1240 	}
1241 
1242 	/*
1243 	 * If the device returned EIO, we revalidate the media.  If it is
1244 	 * determined the media has changed this triggers the asynchronous
1245 	 * removal of the device from the configuration.
1246 	 */
1247 	if (zio->io_error == EIO) {
1248 		vdev_t *v = zio->io_vd;
1249 		vdev_disk_t *vd = v->vdev_tsd;
1250 
1251 		if (!zfs_check_disk_status(BDH_BDEV(vd->vd_bdh))) {
1252 			invalidate_bdev(BDH_BDEV(vd->vd_bdh));
1253 			v->vdev_remove_wanted = B_TRUE;
1254 			spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
1255 		}
1256 	}
1257 }
1258 
1259 static void
vdev_disk_hold(vdev_t * vd)1260 vdev_disk_hold(vdev_t *vd)
1261 {
1262 	ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
1263 
1264 	/* We must have a pathname, and it must be absolute. */
1265 	if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
1266 		return;
1267 
1268 	/*
1269 	 * Only prefetch path and devid info if the device has
1270 	 * never been opened.
1271 	 */
1272 	if (vd->vdev_tsd != NULL)
1273 		return;
1274 
1275 }
1276 
1277 static void
vdev_disk_rele(vdev_t * vd)1278 vdev_disk_rele(vdev_t *vd)
1279 {
1280 	ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
1281 
1282 	/* XXX: Implement me as a vnode rele for the device */
1283 }
1284 
1285 vdev_ops_t vdev_disk_ops = {
1286 	.vdev_op_init = NULL,
1287 	.vdev_op_fini = NULL,
1288 	.vdev_op_open = vdev_disk_open,
1289 	.vdev_op_close = vdev_disk_close,
1290 	.vdev_op_asize_to_psize = vdev_default_psize,
1291 	.vdev_op_psize_to_asize = vdev_default_asize,
1292 	.vdev_op_min_asize = vdev_default_min_asize,
1293 	.vdev_op_min_alloc = NULL,
1294 	.vdev_op_io_start = vdev_disk_io_start,
1295 	.vdev_op_io_done = vdev_disk_io_done,
1296 	.vdev_op_state_change = NULL,
1297 	.vdev_op_need_resilver = NULL,
1298 	.vdev_op_hold = vdev_disk_hold,
1299 	.vdev_op_rele = vdev_disk_rele,
1300 	.vdev_op_remap = NULL,
1301 	.vdev_op_xlate = vdev_default_xlate,
1302 	.vdev_op_rebuild_asize = NULL,
1303 	.vdev_op_metaslab_init = NULL,
1304 	.vdev_op_config_generate = NULL,
1305 	.vdev_op_nparity = NULL,
1306 	.vdev_op_ndisks = NULL,
1307 	.vdev_op_type = VDEV_TYPE_DISK,		/* name of this vdev type */
1308 	.vdev_op_leaf = B_TRUE,			/* leaf vdev */
1309 	.vdev_op_kobj_evt_post = vdev_disk_kobj_evt_post
1310 };
1311 
1312 int
param_set_min_auto_ashift(const char * buf,zfs_kernel_param_t * kp)1313 param_set_min_auto_ashift(const char *buf, zfs_kernel_param_t *kp)
1314 {
1315 	uint_t val;
1316 	int error;
1317 
1318 	error = kstrtouint(buf, 0, &val);
1319 	if (error < 0)
1320 		return (SET_ERROR(error));
1321 
1322 	if (val < ASHIFT_MIN || val > zfs_vdev_max_auto_ashift)
1323 		return (SET_ERROR(-EINVAL));
1324 
1325 	error = param_set_uint(buf, kp);
1326 	if (error < 0)
1327 		return (SET_ERROR(error));
1328 
1329 	return (0);
1330 }
1331 
1332 int
param_set_max_auto_ashift(const char * buf,zfs_kernel_param_t * kp)1333 param_set_max_auto_ashift(const char *buf, zfs_kernel_param_t *kp)
1334 {
1335 	uint_t val;
1336 	int error;
1337 
1338 	error = kstrtouint(buf, 0, &val);
1339 	if (error < 0)
1340 		return (SET_ERROR(error));
1341 
1342 	if (val > ASHIFT_MAX || val < zfs_vdev_min_auto_ashift)
1343 		return (SET_ERROR(-EINVAL));
1344 
1345 	error = param_set_uint(buf, kp);
1346 	if (error < 0)
1347 		return (SET_ERROR(error));
1348 
1349 	return (0);
1350 }
1351 
1352 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, open_timeout_ms, UINT, ZMOD_RW,
1353 	"Timeout before determining that a device is missing");
1354 
1355 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, failfast_mask, UINT, ZMOD_RW,
1356 	"Defines failfast mask: 1 - device, 2 - transport, 4 - driver");
1357 
1358 ZFS_MODULE_PARAM(zfs_vdev_disk, zfs_vdev_disk_, max_segs, UINT, ZMOD_RW,
1359 	"Maximum number of data segments to add to an IO request (min 4)");
1360