xref: /freebsd/sys/contrib/openzfs/module/os/linux/zfs/vdev_disk.c (revision 22649d4dba730d46244fd2dff4fd174903c8379f)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * This file and its contents are supplied under the terms of the
4  * Common Development and Distribution License ("CDDL"), version 1.0.
5  * You may only use this file in accordance with the terms of version
6  * 1.0 of the CDDL.
7  *
8  * A full copy of the text of the CDDL should have accompanied this
9  * source.  A copy of the CDDL is also available via the Internet at
10  * https://opensource.org/license/CDDL-1.0.
11  */
12 /*
13  * Copyright (C) 2008-2010 Lawrence Livermore National Security, LLC.
14  * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
15  * Rewritten for Linux by Brian Behlendorf <behlendorf1@llnl.gov>.
16  * LLNL-CODE-403049.
17  * Copyright (c) 2012, 2019 by Delphix. All rights reserved.
18  * Copyright (c) 2023, 2024, 2025, Klara, Inc.
19  * Copyright (c) 2026, TrueNAS.
20  */
21 
22 #include <sys/zfs_context.h>
23 #include <sys/spa_impl.h>
24 #include <sys/vdev_disk.h>
25 #include <sys/vdev_impl.h>
26 #include <sys/vdev_trim.h>
27 #include <sys/abd.h>
28 #include <sys/fs/zfs.h>
29 #include <sys/zio.h>
30 #include <linux/blkpg.h>
31 #include <linux/msdos_fs.h>
32 #include <linux/vfs_compat.h>
33 #include <linux/blk-cgroup.h>
34 
35 /*
36  * Linux 6.8.x uses a bdev_handle as an instance/refcount for an underlying
37  * block_device. Since it carries the block_device inside, its convenient to
38  * just use the handle as a proxy.
39  *
40  * Linux 6.9.x uses a file for the same purpose.
41  *
42  * For pre-6.8, we just emulate this with a cast, since we don't need any of
43  * the other fields inside the handle.
44  */
45 #if defined(HAVE_BDEV_OPEN_BY_PATH)
46 typedef struct bdev_handle zfs_bdev_handle_t;
47 #define	BDH_BDEV(bdh)		((bdh)->bdev)
48 #define	BDH_IS_ERR(bdh)		(IS_ERR(bdh))
49 #define	BDH_PTR_ERR(bdh)	(PTR_ERR(bdh))
50 #define	BDH_ERR_PTR(err)	(ERR_PTR(err))
51 #elif defined(HAVE_BDEV_FILE_OPEN_BY_PATH)
52 typedef struct file zfs_bdev_handle_t;
53 #define	BDH_BDEV(bdh)		(file_bdev(bdh))
54 #define	BDH_IS_ERR(bdh)		(IS_ERR(bdh))
55 #define	BDH_PTR_ERR(bdh)	(PTR_ERR(bdh))
56 #define	BDH_ERR_PTR(err)	(ERR_PTR(err))
57 #else
58 typedef void zfs_bdev_handle_t;
59 #define	BDH_BDEV(bdh)		((struct block_device *)bdh)
60 #define	BDH_IS_ERR(bdh)		(IS_ERR(BDH_BDEV(bdh)))
61 #define	BDH_PTR_ERR(bdh)	(PTR_ERR(BDH_BDEV(bdh)))
62 #define	BDH_ERR_PTR(err)	(ERR_PTR(err))
63 #endif
64 
65 typedef struct vdev_disk {
66 	zfs_bdev_handle_t		*vd_bdh;
67 	krwlock_t			vd_lock;
68 } vdev_disk_t;
69 
70 /*
71  * Maximum number of segments to add to a bio (min 4). If this is higher than
72  * the maximum allowed by the device queue or the kernel itself, it will be
73  * clamped. Setting it to zero will cause the kernel's ideal size to be used.
74  */
75 uint_t zfs_vdev_disk_max_segs = 0;
76 
77 /*
78  * Unique identifier for the exclusive vdev holder.
79  */
80 static void *zfs_vdev_holder = VDEV_HOLDER;
81 
82 /*
83  * Wait up to zfs_vdev_open_timeout_ms milliseconds before determining the
84  * device is missing. The missing path may be transient since the links
85  * can be briefly removed and recreated in response to udev events.
86  */
87 static uint_t zfs_vdev_open_timeout_ms = 1000;
88 
89 /*
90  * Size of the "reserved" partition, in blocks.
91  */
92 #define	EFI_MIN_RESV_SIZE	(16 * 1024)
93 
94 /*
95  * BIO request failfast mask.
96  */
97 
98 static unsigned int zfs_vdev_failfast_mask = 1;
99 
100 /*
101  * Whether we wait for bio to complete. Also requires that
102  * zio has bypassed the vdev queue. May lead to performance
103  * improvements when backing vdev devices are fast and low
104  * latency. May impact performance with certain workloads
105  * when enabled on raidz or draid zpool configurations.
106  */
107 static unsigned int zfs_vdev_disk_calling_thread_io = 0;
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_path_backing_permission(struct path * path,int mask)284 vdev_path_backing_permission(struct path *path, int mask)
285 {
286 #if defined(HAVE_IDMAP_MNTIDMAP)
287 	return (inode_permission(mnt_idmap(path->mnt),
288 	    d_backing_inode(path->dentry), mask));
289 #elif defined(HAVE_IDMAP_USERNS)
290 	return (inode_permission(mnt_user_ns(path->mnt),
291 	    d_backing_inode(path->dentry), mask));
292 #else
293 	return (inode_permission(d_backing_inode(path->dentry), mask));
294 #endif
295 }
296 
297 static int
vdev_disk_open(vdev_t * v,uint64_t * psize,uint64_t * max_psize,uint64_t * logical_ashift,uint64_t * physical_ashift,cred_t * cr)298 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
299     uint64_t *logical_ashift, uint64_t *physical_ashift, cred_t *cr)
300 {
301 	zfs_bdev_handle_t *bdh;
302 	spa_mode_t smode = spa_mode(v->vdev_spa);
303 	hrtime_t timeout = MSEC2NSEC(zfs_vdev_open_timeout_ms);
304 	vdev_disk_t *vd;
305 	const cred_t *oldcr = NULL;
306 
307 	/* Must have a pathname and it must be absolute. */
308 	if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
309 		v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
310 		vdev_dbgmsg(v, "invalid vdev_path");
311 		return (SET_ERROR(EINVAL));
312 	}
313 
314 	/*
315 	 * Reopen the device if it is currently open.  When expanding a
316 	 * partition force re-scanning the partition table if userland
317 	 * did not take care of this already. We need to do this while closed
318 	 * in order to get an accurate updated block device size.  Then
319 	 * since udev may need to recreate the device links increase the
320 	 * open retry timeout before reporting the device as unavailable.
321 	 */
322 	vd = v->vdev_tsd;
323 	if (vd) {
324 		char disk_name[BDEVNAME_SIZE + 6] = "/dev/";
325 		boolean_t reread_part = B_FALSE;
326 
327 		/*
328 		 * Reopening an already-open device, so the caller credential
329 		 * is irrelevant - we had access to it before, we can assume
330 		 * we still do.
331 		 */
332 		oldcr = override_creds(kcred);
333 
334 		rw_enter(&vd->vd_lock, RW_WRITER);
335 		bdh = vd->vd_bdh;
336 		vd->vd_bdh = NULL;
337 
338 		if (bdh) {
339 			struct block_device *bdev = BDH_BDEV(bdh);
340 			if (v->vdev_expanding && bdev != bdev_whole(bdev)) {
341 				vdev_bdevname(bdev_whole(bdev), disk_name + 5);
342 				/*
343 				 * If userland has BLKPG_RESIZE_PARTITION,
344 				 * then it should have updated the partition
345 				 * table already. We can detect this by
346 				 * comparing our current physical size
347 				 * with that of the device. If they are
348 				 * the same, then we must not have
349 				 * BLKPG_RESIZE_PARTITION or it failed to
350 				 * update the partition table online. We
351 				 * fallback to rescanning the partition
352 				 * table from the kernel below. However,
353 				 * if the capacity already reflects the
354 				 * updated partition, then we skip
355 				 * rescanning the partition table here.
356 				 */
357 				if (v->vdev_psize == bdev_capacity(bdev))
358 					reread_part = B_TRUE;
359 			}
360 
361 			vdev_blkdev_put(bdh, smode, zfs_vdev_holder);
362 		}
363 
364 		if (reread_part) {
365 			bdh = vdev_blkdev_get_by_path(disk_name, smode,
366 			    zfs_vdev_holder);
367 			if (!BDH_IS_ERR(bdh)) {
368 				int error =
369 				    vdev_bdev_reread_part(BDH_BDEV(bdh));
370 				vdev_blkdev_put(bdh, smode, zfs_vdev_holder);
371 				if (error == 0) {
372 					timeout = MSEC2NSEC(
373 					    zfs_vdev_open_timeout_ms * 2);
374 				}
375 			}
376 		}
377 	} else {
378 		vd = kmem_zalloc(sizeof (vdev_disk_t), KM_SLEEP);
379 
380 		rw_init(&vd->vd_lock, NULL, RW_DEFAULT, NULL);
381 		rw_enter(&vd->vd_lock, RW_WRITER);
382 
383 		/* Restrict device access below to caller's permissions. */
384 		oldcr = override_creds(cr);
385 	}
386 
387 	/*
388 	 * Devices are always opened by the path provided at configuration
389 	 * time.  This means that if the provided path is a udev by-id path
390 	 * then drives may be re-cabled without an issue.  If the provided
391 	 * path is a udev by-path path, then the physical location information
392 	 * will be preserved.  This can be critical for more complicated
393 	 * configurations where drives are located in specific physical
394 	 * locations to maximize the systems tolerance to component failure.
395 	 *
396 	 * Alternatively, you can provide your own udev rule to flexibly map
397 	 * the drives as you see fit.  It is not advised that you use the
398 	 * /dev/[hd]d devices which may be reordered due to probing order.
399 	 * Devices in the wrong locations will be detected by the higher
400 	 * level vdev validation.
401 	 *
402 	 * The specified paths may be briefly removed and recreated in
403 	 * response to udev events.  This should be exceptionally unlikely
404 	 * because the zpool command makes every effort to verify these paths
405 	 * have already settled prior to reaching this point.  Therefore,
406 	 * a ENOENT failure at this point is highly likely to be transient
407 	 * and it is reasonable to sleep and retry before giving up.  In
408 	 * practice delays have been observed to be on the order of 100ms.
409 	 *
410 	 * When ERESTARTSYS is returned it indicates the block device is
411 	 * a zvol which could not be opened due to the deadlock detection
412 	 * logic in zvol_open().  Extend the timeout and retry the open
413 	 * subsequent attempts are expected to eventually succeed.
414 	 */
415 
416 	hrtime_t start = gethrtime();
417 	int err = ENXIO;
418 	while (err != 0 && ((gethrtime() - start) < timeout)) {
419 
420 		/*
421 		 * Ensure the caller credential (made live by override_creds()
422 		 * above) has access to the device node. This will include
423 		 * checking file permissions and considering the
424 		 * CAP_DAC_OVERRIDE capability.
425 		 */
426 		struct path devpath;
427 		err = kern_path(v->vdev_path, LOOKUP_FOLLOW, &devpath);
428 		if (err == 0) {
429 			int mask =
430 			    (smode & SPA_MODE_READ ? MAY_READ : 0) |
431 			    (smode & SPA_MODE_WRITE ? MAY_WRITE : 0);
432 			err = vdev_path_backing_permission(&devpath, mask);
433 			path_put(&devpath);
434 		}
435 
436 		if (err == 0) {
437 			bdh = vdev_blkdev_get_by_path(v->vdev_path, smode,
438 			    zfs_vdev_holder);
439 			err = BDH_IS_ERR(bdh) ? BDH_PTR_ERR(bdh) : 0;
440 		}
441 
442 		if (err == 0)
443 			break;
444 
445 		if (unlikely(err == -ENOENT)) {
446 			/*
447 			 * There is no point of waiting since device is removed
448 			 * explicitly
449 			 */
450 			if (v->vdev_removed)
451 				break;
452 
453 			schedule_timeout_interruptible(MSEC_TO_TICK(10));
454 		} else if (unlikely(err == -ERESTARTSYS)) {
455 			timeout = MSEC2NSEC(zfs_vdev_open_timeout_ms * 10);
456 			continue;
457 		}
458 
459 		break;
460 	}
461 
462 	revert_creds(oldcr);
463 
464 	if (err != 0) {
465 		vdev_dbgmsg(v, "open error=%d timeout=%llu/%llu", -err,
466 		    (u_longlong_t)(gethrtime() - start),
467 		    (u_longlong_t)timeout);
468 		vd->vd_bdh = NULL;
469 		v->vdev_tsd = vd;
470 		rw_exit(&vd->vd_lock);
471 		return (SET_ERROR(-err));
472 	}
473 
474 	vd->vd_bdh = bdh;
475 	v->vdev_tsd = vd;
476 	rw_exit(&vd->vd_lock);
477 
478 	struct block_device *bdev = BDH_BDEV(vd->vd_bdh);
479 
480 	/*  Determine the physical block size */
481 	int physical_block_size = bdev_physical_block_size(bdev);
482 
483 	/*  Determine the logical block size */
484 	int logical_block_size = bdev_logical_block_size(bdev);
485 
486 	/*
487 	 * If the device has a write cache, clear the nowritecache flag,
488 	 * so that we start issuing flush requests again.
489 	 */
490 	v->vdev_nowritecache = !zfs_bdev_has_write_cache(bdev);
491 
492 	/* Set when device reports it supports TRIM. */
493 	v->vdev_has_trim = bdev_discard_supported(bdev);
494 
495 	/* Set when device reports it supports secure TRIM. */
496 	v->vdev_has_securetrim = bdev_secure_discard_supported(bdev);
497 
498 	/* Inform the ZIO pipeline that we are non-rotational */
499 #ifdef HAVE_BLK_QUEUE_ROT
500 	v->vdev_nonrot = !blk_queue_rot(bdev_get_queue(bdev));
501 #else
502 	v->vdev_nonrot = blk_queue_nonrot(bdev_get_queue(bdev));
503 #endif
504 
505 	/* Is backed by a block device. */
506 	v->vdev_is_blkdev = B_TRUE;
507 
508 	/* Physical volume size in bytes for the partition */
509 	*psize = bdev_capacity(bdev);
510 
511 	/* Physical volume size in bytes including possible expansion space */
512 	*max_psize = bdev_max_capacity(bdev, v->vdev_wholedisk);
513 
514 	/* Based on the minimum sector size set the block size */
515 	*physical_ashift = highbit64(MAX(physical_block_size,
516 	    SPA_MINBLOCKSIZE)) - 1;
517 
518 	*logical_ashift = highbit64(MAX(logical_block_size,
519 	    SPA_MINBLOCKSIZE)) - 1;
520 
521 	return (0);
522 }
523 
524 static void
vdev_disk_close(vdev_t * v)525 vdev_disk_close(vdev_t *v)
526 {
527 	vdev_disk_t *vd = v->vdev_tsd;
528 
529 	if (v->vdev_reopening || vd == NULL)
530 		return;
531 
532 	rw_enter(&vd->vd_lock, RW_WRITER);
533 
534 	if (vd->vd_bdh != NULL)
535 		vdev_blkdev_put(vd->vd_bdh, spa_mode(v->vdev_spa),
536 		    zfs_vdev_holder);
537 
538 	v->vdev_tsd = NULL;
539 
540 	rw_exit(&vd->vd_lock);
541 	rw_destroy(&vd->vd_lock);
542 	kmem_free(vd, sizeof (vdev_disk_t));
543 }
544 
545 /*
546  * preempt_schedule_notrace is GPL-only which breaks the ZFS build, so
547  * replace it with preempt_schedule under the following condition:
548  */
549 #if defined(CONFIG_ARM64) && \
550     defined(CONFIG_PREEMPTION) && \
551     defined(CONFIG_BLK_CGROUP)
552 #define	preempt_schedule_notrace(x) preempt_schedule(x)
553 #endif
554 
555 /*
556  * As for the Linux 5.18 kernel bio_alloc() expects a block_device struct
557  * as an argument removing the need to set it with bio_set_dev().  This
558  * removes the need for all of the following compatibility code.
559  */
560 #if !defined(HAVE_BIO_ALLOC_4ARG)
561 
562 #if defined(CONFIG_BLK_CGROUP) && defined(HAVE_BIO_SET_DEV_GPL_ONLY)
563 /*
564  * The Linux 5.5 kernel updated percpu_ref_tryget() which is inlined by
565  * blkg_tryget() to use rcu_read_lock() instead of rcu_read_lock_sched().
566  * As a side effect the function was converted to GPL-only.  Define our
567  * own version when needed which uses rcu_read_lock_sched().
568  *
569  * The Linux 5.17 kernel split linux/blk-cgroup.h into a private and a public
570  * part, moving blkg_tryget into the private one. Define our own version.
571  */
572 #if defined(HAVE_BLKG_TRYGET_GPL_ONLY) || !defined(HAVE_BLKG_TRYGET)
573 static inline bool
vdev_blkg_tryget(struct blkcg_gq * blkg)574 vdev_blkg_tryget(struct blkcg_gq *blkg)
575 {
576 	struct percpu_ref *ref = &blkg->refcnt;
577 	unsigned long __percpu *count;
578 	bool rc;
579 
580 	rcu_read_lock_sched();
581 
582 	if (__ref_is_percpu(ref, &count)) {
583 		this_cpu_inc(*count);
584 		rc = true;
585 	} else {
586 #ifdef ZFS_PERCPU_REF_COUNT_IN_DATA
587 		rc = atomic_long_inc_not_zero(&ref->data->count);
588 #else
589 		rc = atomic_long_inc_not_zero(&ref->count);
590 #endif
591 	}
592 
593 	rcu_read_unlock_sched();
594 
595 	return (rc);
596 }
597 #else
598 #define	vdev_blkg_tryget(bg)	blkg_tryget(bg)
599 #endif
600 #ifdef HAVE_BIO_SET_DEV_MACRO
601 /*
602  * The Linux 5.0 kernel updated the bio_set_dev() macro so it calls the
603  * GPL-only bio_associate_blkg() symbol thus inadvertently converting
604  * the entire macro.  Provide a minimal version which always assigns the
605  * request queue's root_blkg to the bio.
606  */
607 static inline void
vdev_bio_associate_blkg(struct bio * bio)608 vdev_bio_associate_blkg(struct bio *bio)
609 {
610 #if defined(HAVE_BIO_BDEV_DISK)
611 	struct request_queue *q = bio->bi_bdev->bd_disk->queue;
612 #else
613 	struct request_queue *q = bio->bi_disk->queue;
614 #endif
615 
616 	ASSERT3P(q, !=, NULL);
617 	ASSERT0P(bio->bi_blkg);
618 
619 	if (q->root_blkg && vdev_blkg_tryget(q->root_blkg))
620 		bio->bi_blkg = q->root_blkg;
621 }
622 
623 #define	bio_associate_blkg vdev_bio_associate_blkg
624 #else
625 static inline void
vdev_bio_set_dev(struct bio * bio,struct block_device * bdev)626 vdev_bio_set_dev(struct bio *bio, struct block_device *bdev)
627 {
628 #if defined(HAVE_BIO_BDEV_DISK)
629 	struct request_queue *q = bdev->bd_disk->queue;
630 #else
631 	struct request_queue *q = bio->bi_disk->queue;
632 #endif
633 	bio_clear_flag(bio, BIO_REMAPPED);
634 	if (bio->bi_bdev != bdev)
635 		bio_clear_flag(bio, BIO_THROTTLED);
636 	bio->bi_bdev = bdev;
637 
638 	ASSERT3P(q, !=, NULL);
639 	ASSERT0P(bio->bi_blkg);
640 
641 	if (q->root_blkg && vdev_blkg_tryget(q->root_blkg))
642 		bio->bi_blkg = q->root_blkg;
643 }
644 #define	bio_set_dev		vdev_bio_set_dev
645 #endif
646 #endif
647 #endif /* !HAVE_BIO_ALLOC_4ARG */
648 
649 static inline void
vdev_submit_bio(struct bio * bio)650 vdev_submit_bio(struct bio *bio)
651 {
652 	struct bio_list *bio_list = current->bio_list;
653 	current->bio_list = NULL;
654 	(void) submit_bio(bio);
655 	current->bio_list = bio_list;
656 }
657 
658 static inline void
vdev_submit_bio_wait(struct bio * bio)659 vdev_submit_bio_wait(struct bio *bio)
660 {
661 	struct bio_list *bio_list = current->bio_list;
662 	current->bio_list = NULL;
663 	(void) submit_bio_wait(bio);
664 	current->bio_list = bio_list;
665 }
666 
667 static inline struct bio *
vdev_bio_alloc(struct block_device * bdev,gfp_t gfp_mask,unsigned short nr_vecs)668 vdev_bio_alloc(struct block_device *bdev, gfp_t gfp_mask,
669     unsigned short nr_vecs)
670 {
671 	struct bio *bio;
672 
673 #ifdef HAVE_BIO_ALLOC_4ARG
674 	bio = bio_alloc(bdev, nr_vecs, 0, gfp_mask);
675 #else
676 	bio = bio_alloc(gfp_mask, nr_vecs);
677 	if (likely(bio != NULL))
678 		bio_set_dev(bio, bdev);
679 #endif
680 
681 	return (bio);
682 }
683 
684 static inline uint_t
vdev_bio_max_segs(struct block_device * bdev)685 vdev_bio_max_segs(struct block_device *bdev)
686 {
687 	/*
688 	 * Smallest of the device max segs and the tunable max segs. Minimum
689 	 * 4, so there's room to finish split pages if they come up.
690 	 */
691 	const uint_t dev_max_segs = queue_max_segments(bdev_get_queue(bdev));
692 	const uint_t tune_max_segs = (zfs_vdev_disk_max_segs > 0) ?
693 	    MAX(4, zfs_vdev_disk_max_segs) : dev_max_segs;
694 	const uint_t max_segs = MIN(tune_max_segs, dev_max_segs);
695 
696 #ifdef HAVE_BIO_MAX_SEGS
697 	return (bio_max_segs(max_segs));
698 #else
699 	return (MIN(max_segs, BIO_MAX_PAGES));
700 #endif
701 }
702 
703 static inline uint_t
vdev_bio_max_bytes(struct block_device * bdev)704 vdev_bio_max_bytes(struct block_device *bdev)
705 {
706 	return (queue_max_sectors(bdev_get_queue(bdev)) << 9);
707 }
708 
709 
710 /*
711  * Virtual block IO object (VBIO)
712  *
713  * Linux block IO (BIO) objects have a limit on how many data segments (pages)
714  * they can hold. Depending on how they're allocated and structured, a large
715  * ZIO can require more than one BIO to be submitted to the kernel, which then
716  * all have to complete before we can return the completed ZIO back to ZFS.
717  *
718  * A VBIO is a wrapper around multiple BIOs, carrying everything needed to
719  * translate a ZIO down into the kernel block layer and back again.
720  *
721  * Note that these are only used for data ZIOs (read/write). Meta-operations
722  * (flush/trim) don't need multiple BIOs and so can just make the call
723  * directly.
724  */
725 typedef struct {
726 	zio_t		*vbio_zio;	/* parent zio */
727 
728 	struct block_device *vbio_bdev;	/* blockdev to submit bios to */
729 
730 	abd_t		*vbio_abd;	/* abd carrying borrowed linear buf */
731 
732 	uint_t		vbio_max_segs;	/* max segs per bio */
733 
734 	uint_t		vbio_max_bytes;	/* max bytes per bio */
735 	uint_t		vbio_lbs_mask;	/* logical block size mask */
736 
737 	uint64_t	vbio_offset;	/* start offset of next bio */
738 
739 	struct bio	*vbio_bio;	/* pointer to the current bio */
740 	int		vbio_flags;	/* bio flags */
741 	boolean_t	vbio_wait;	/* wait for completion */
742 } vbio_t;
743 
744 static vbio_t *
vbio_alloc(zio_t * zio,struct block_device * bdev,int flags)745 vbio_alloc(zio_t *zio, struct block_device *bdev, int flags)
746 {
747 	vbio_t *vbio = kmem_zalloc(sizeof (vbio_t), KM_SLEEP);
748 
749 	vbio->vbio_zio = zio;
750 	vbio->vbio_bdev = bdev;
751 	vbio->vbio_abd = NULL;
752 	vbio->vbio_max_segs = vdev_bio_max_segs(bdev);
753 	vbio->vbio_max_bytes = vdev_bio_max_bytes(bdev);
754 	vbio->vbio_lbs_mask = ~(bdev_logical_block_size(bdev)-1);
755 	vbio->vbio_offset = zio->io_offset;
756 	vbio->vbio_bio = NULL;
757 	vbio->vbio_flags = flags;
758 	vbio->vbio_wait = B_FALSE;
759 
760 	return (vbio);
761 }
762 
763 static void vbio_completion(struct bio *bio);
764 
765 static int
vbio_add_page(vbio_t * vbio,struct page * page,uint_t size,uint_t offset)766 vbio_add_page(vbio_t *vbio, struct page *page, uint_t size, uint_t offset)
767 {
768 	struct bio *bio = vbio->vbio_bio;
769 	uint_t ssize;
770 
771 	while (size > 0) {
772 		if (bio == NULL) {
773 			/* New BIO, allocate and set up */
774 			bio = vdev_bio_alloc(vbio->vbio_bdev, GFP_NOIO,
775 			    vbio->vbio_max_segs);
776 			VERIFY(bio);
777 
778 			BIO_BI_SECTOR(bio) = vbio->vbio_offset >> 9;
779 			bio_set_op_attrs(bio,
780 			    vbio->vbio_zio->io_type == ZIO_TYPE_WRITE ?
781 			    WRITE : READ, vbio->vbio_flags);
782 
783 			if (vbio->vbio_bio) {
784 				bio_chain(vbio->vbio_bio, bio);
785 				vdev_submit_bio(vbio->vbio_bio);
786 			}
787 			vbio->vbio_bio = bio;
788 		}
789 
790 		/*
791 		 * Only load as much of the current page data as will fit in
792 		 * the space left in the BIO, respecting lbs alignment. Older
793 		 * kernels will error if we try to overfill the BIO, while
794 		 * newer ones will accept it and split the BIO. This ensures
795 		 * everything works on older kernels, and avoids an additional
796 		 * overhead on the new.
797 		 */
798 		ssize = MIN(size, (vbio->vbio_max_bytes - BIO_BI_SIZE(bio)) &
799 		    vbio->vbio_lbs_mask);
800 		if (ssize > 0 &&
801 		    bio_add_page(bio, page, ssize, offset) == ssize) {
802 			/* Accepted, adjust and load any remaining. */
803 			size -= ssize;
804 			offset += ssize;
805 			continue;
806 		}
807 
808 		/* No room, set up for a new BIO and loop */
809 		vbio->vbio_offset += BIO_BI_SIZE(bio);
810 
811 		/* Signal new BIO allocation wanted */
812 		bio = NULL;
813 	}
814 
815 	return (0);
816 }
817 
818 /* Iterator callback to submit ABD pages to the vbio. */
819 static int
vbio_fill_cb(struct page * page,size_t off,size_t len,void * priv)820 vbio_fill_cb(struct page *page, size_t off, size_t len, void *priv)
821 {
822 	vbio_t *vbio = priv;
823 	return (vbio_add_page(vbio, page, len, off));
824 }
825 
826 /* Create some BIOs, fill them with data and submit them */
827 static void
vbio_submit(vbio_t * vbio,abd_t * abd,uint64_t size)828 vbio_submit(vbio_t *vbio, abd_t *abd, uint64_t size)
829 {
830 	/*
831 	 * We plug so we can submit the BIOs as we go and only unplug them when
832 	 * they are fully created and submitted. This is important; if we don't
833 	 * plug, then the kernel may start executing earlier BIOs while we're
834 	 * still creating and executing later ones, and if the device goes
835 	 * away while that's happening, older kernels can get confused and
836 	 * trample memory.
837 	 */
838 	struct blk_plug plug;
839 	blk_start_plug(&plug);
840 
841 	(void) abd_iterate_page_func(abd, 0, size, vbio_fill_cb, vbio);
842 	ASSERT(vbio->vbio_bio);
843 
844 	/*
845 	 * Once submitted, vbio_bio now owns vbio (through bi_private) and we
846 	 * can't touch it again. The bio may complete and vbio_completion() be
847 	 * called and free the vbio before this task is run again, so we must
848 	 * consider it invalid from this point.
849 	 */
850 
851 	if (vbio->vbio_wait) {
852 		vdev_submit_bio_wait(vbio->vbio_bio);
853 	} else {
854 		vbio->vbio_bio->bi_end_io = vbio_completion;
855 		vbio->vbio_bio->bi_private = vbio;
856 		vdev_submit_bio(vbio->vbio_bio);
857 	}
858 
859 	blk_finish_plug(&plug);
860 }
861 
862 /* IO completion callback */
863 static void
vbio_completion(struct bio * bio)864 vbio_completion(struct bio *bio)
865 {
866 	vbio_t *vbio = bio->bi_private;
867 	zio_t *zio = vbio->vbio_zio;
868 
869 	ASSERT(zio);
870 
871 	/* Capture and log any errors */
872 	zio->io_error = bi_status_to_errno(bio->bi_status);
873 	ASSERT3U(zio->io_error, >=, 0);
874 
875 	if (zio->io_error)
876 		vdev_disk_error(zio);
877 
878 	/* Return the BIO to the kernel */
879 	bio_put(bio);
880 
881 	/*
882 	 * We're likely in an interrupt context so we can't do ABD/memory work
883 	 * here; instead we stash vbio on the zio and take care of it in the
884 	 * done callback.
885 	 */
886 	ASSERT0P(zio->io_bio);
887 	zio->io_bio = vbio;
888 
889 	/* Using calling thread io, don't dispatch zio. */
890 	if (vbio->vbio_wait)
891 		zio_execute(zio);
892 	else
893 		zio_delay_interrupt(zio);
894 
895 }
896 
897 /*
898  * Iterator callback to count ABD pages and check their size & alignment.
899  *
900  * On Linux, each BIO segment can take a page pointer, and an offset+length of
901  * the data within that page. A page can be arbitrarily large ("compound"
902  * pages) but we still have to ensure the data portion is correctly sized and
903  * aligned to the logical block size, to ensure that if the kernel wants to
904  * split the BIO, the two halves will still be properly aligned.
905  *
906  * NOTE: if you change this function, change the copy in
907  * tests/zfs-tests/tests/functional/vdev_disk/page_alignment.c, and add test
908  * data there to validate the change you're making.
909  */
910 typedef struct {
911 	size_t	blocksize;
912 	int	seen_first;
913 	int	seen_last;
914 } vdev_disk_check_alignment_t;
915 
916 static int
vdev_disk_check_alignment_cb(struct page * page,size_t off,size_t len,void * priv)917 vdev_disk_check_alignment_cb(struct page *page, size_t off, size_t len,
918     void *priv)
919 {
920 	(void) page;
921 	vdev_disk_check_alignment_t *s = priv;
922 
923 	/*
924 	 * The cardinal rule: a single on-disk block must never cross an
925 	 * physical (order-0) page boundary, as the kernel expects to be able
926 	 * to split at both LBS and page boundaries.
927 	 *
928 	 * This implies various alignment rules for the blocks in this
929 	 * (possibly compound) page, which we can check for.
930 	 */
931 
932 	/*
933 	 * If the previous page did not end on a page boundary, then we
934 	 * can't proceed without creating a hole.
935 	 */
936 	if (s->seen_last)
937 		return (1);
938 
939 	/* This page must contain only whole LBS-sized blocks. */
940 	if (!IS_P2ALIGNED(len, s->blocksize))
941 		return (1);
942 
943 	/*
944 	 * If this is not the first page in the ABD, then the data must start
945 	 * on a page-aligned boundary (so the kernel can split on page
946 	 * boundaries without having to deal with a hole). If it is, then
947 	 * it can start on LBS-alignment.
948 	 */
949 	if (s->seen_first) {
950 		if (!IS_P2ALIGNED(off, PAGESIZE))
951 			return (1);
952 	} else {
953 		if (!IS_P2ALIGNED(off, s->blocksize))
954 			return (1);
955 		s->seen_first = 1;
956 	}
957 
958 	/*
959 	 * If this data does not end on a page-aligned boundary, then this
960 	 * must be the last page in the ABD, for the same reason.
961 	 */
962 	s->seen_last = !IS_P2ALIGNED(off+len, PAGESIZE);
963 
964 	return (0);
965 }
966 
967 /*
968  * Check if we can submit the pages in this ABD to the kernel as-is. Returns
969  * the number of pages, or 0 if it can't be submitted like this.
970  */
971 static boolean_t
vdev_disk_check_alignment(abd_t * abd,uint64_t size,struct block_device * bdev)972 vdev_disk_check_alignment(abd_t *abd, uint64_t size, struct block_device *bdev)
973 {
974 	vdev_disk_check_alignment_t s = {
975 	    .blocksize = bdev_logical_block_size(bdev),
976 	};
977 
978 	if (abd_iterate_page_func(abd, 0, size,
979 	    vdev_disk_check_alignment_cb, &s))
980 		return (B_FALSE);
981 
982 	return (B_TRUE);
983 }
984 
985 static int
vdev_disk_io_rw(zio_t * zio)986 vdev_disk_io_rw(zio_t *zio)
987 {
988 	vdev_t *v = zio->io_vd;
989 	vdev_disk_t *vd = v->vdev_tsd;
990 	struct block_device *bdev = BDH_BDEV(vd->vd_bdh);
991 	int flags = 0;
992 
993 	/*
994 	 * Accessing outside the block device is never allowed.
995 	 */
996 	if (zio->io_offset + zio->io_size > bdev_capacity(bdev)) {
997 		vdev_dbgmsg(zio->io_vd,
998 		    "Illegal access %llu size %llu, device size %llu",
999 		    (u_longlong_t)zio->io_offset,
1000 		    (u_longlong_t)zio->io_size,
1001 		    (u_longlong_t)bdev_capacity(bdev));
1002 		return (SET_ERROR(EIO));
1003 	}
1004 
1005 	vdev_t *iter = v;
1006 	while (iter != NULL && iter->vdev_failfast == ZPROP_BOOLEAN_INHERIT)
1007 		iter = iter->vdev_parent;
1008 
1009 	boolean_t failfast = iter ? iter->vdev_failfast == 1 :
1010 	    vdev_prop_default_numeric(VDEV_PROP_FAILFAST);
1011 	if (!(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)) &&
1012 	    failfast) {
1013 		bio_set_flags_failfast(bdev, &flags, zfs_vdev_failfast_mask & 1,
1014 		    zfs_vdev_failfast_mask & 2, zfs_vdev_failfast_mask & 4);
1015 	}
1016 
1017 	/*
1018 	 * Check alignment of the incoming ABD. If any part of it would require
1019 	 * submitting a page that is not aligned to both the logical block size
1020 	 * and the page size, then we take a copy into a new memory region with
1021 	 * correct alignment.  This should be impossible on a 512b LBS. On
1022 	 * larger blocks, this can happen at least when a small number of
1023 	 * blocks (usually 1) are allocated from a shared slab, or when
1024 	 * abnormally-small data regions (eg gang headers) are mixed into the
1025 	 * same ABD as larger allocations (eg aggregations).
1026 	 */
1027 	abd_t *abd = zio->io_abd;
1028 	if (!vdev_disk_check_alignment(abd, zio->io_size, bdev)) {
1029 		/* Allocate a new memory region with guaranteed alignment */
1030 		abd = abd_alloc_for_io(zio->io_size,
1031 		    zio->io_abd->abd_flags & ABD_FLAG_META);
1032 
1033 		/* If we're writing copy our data into it */
1034 		if (zio->io_type == ZIO_TYPE_WRITE)
1035 			abd_copy(abd, zio->io_abd, zio->io_size);
1036 
1037 		/*
1038 		 * False here would mean the new allocation has an invalid
1039 		 * alignment too, which would mean that abd_alloc() is not
1040 		 * guaranteeing this, or our logic in
1041 		 * vdev_disk_check_alignment() is wrong. In either case,
1042 		 * something in seriously wrong and its not safe to continue.
1043 		 */
1044 		VERIFY(vdev_disk_check_alignment(abd, zio->io_size, bdev));
1045 	}
1046 
1047 	/* Allocate vbio, with a pointer to the borrowed ABD if necessary */
1048 	vbio_t *vbio = vbio_alloc(zio, bdev, flags);
1049 	if (abd != zio->io_abd)
1050 		vbio->vbio_abd = abd;
1051 
1052 	boolean_t bio_wait = B_FALSE;
1053 	if (zfs_vdev_disk_calling_thread_io &&
1054 	    (zio->io_flags & ZIO_FLAG_BYPASSED_QUEUE)) {
1055 		vbio->vbio_wait = bio_wait = B_TRUE;
1056 	}
1057 	/* Fill it with data pages and submit it to the kernel */
1058 	vbio_submit(vbio, abd, zio->io_size);
1059 
1060 	if (bio_wait) {
1061 		vbio->vbio_bio->bi_private = vbio;
1062 		vbio_completion(vbio->vbio_bio);
1063 	}
1064 
1065 	return (0);
1066 }
1067 
1068 static void
vdev_disk_io_flush_completion(struct bio * bio)1069 vdev_disk_io_flush_completion(struct bio *bio)
1070 {
1071 	zio_t *zio = bio->bi_private;
1072 	zio->io_error = bi_status_to_errno(bio->bi_status);
1073 	if (zio->io_error == EOPNOTSUPP || zio->io_error == ENOTTY)
1074 		zio->io_error = SET_ERROR(ENOTSUP);
1075 
1076 	bio_put(bio);
1077 	ASSERT3S(zio->io_error, >=, 0);
1078 	if (zio->io_error)
1079 		vdev_disk_error(zio);
1080 	zio_interrupt(zio);
1081 }
1082 
1083 static int
vdev_disk_io_flush(struct block_device * bdev,zio_t * zio)1084 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
1085 {
1086 	struct request_queue *q;
1087 	struct bio *bio;
1088 
1089 	q = bdev_get_queue(bdev);
1090 	if (!q)
1091 		return (SET_ERROR(ENXIO));
1092 
1093 	bio = vdev_bio_alloc(bdev, GFP_NOIO, 0);
1094 	if (unlikely(bio == NULL))
1095 		return (SET_ERROR(ENOMEM));
1096 
1097 	bio->bi_end_io = vdev_disk_io_flush_completion;
1098 	bio->bi_private = zio;
1099 	bio_set_flush(bio);
1100 	vdev_submit_bio(bio);
1101 	invalidate_bdev(bdev);
1102 
1103 	return (0);
1104 }
1105 
1106 static void
vdev_disk_discard_end_io(struct bio * bio)1107 vdev_disk_discard_end_io(struct bio *bio)
1108 {
1109 	zio_t *zio = bio->bi_private;
1110 	zio->io_error = bi_status_to_errno(bio->bi_status);
1111 
1112 	bio_put(bio);
1113 	if (zio->io_error)
1114 		vdev_disk_error(zio);
1115 	zio_interrupt(zio);
1116 }
1117 
1118 /*
1119  * Wrappers for the different secure erase and discard APIs. We use async
1120  * when available; in this case, *biop is set to the last bio in the chain.
1121  */
1122 static int
vdev_bdev_issue_secure_erase(zfs_bdev_handle_t * bdh,sector_t sector,sector_t nsect,struct bio ** biop)1123 vdev_bdev_issue_secure_erase(zfs_bdev_handle_t *bdh, sector_t sector,
1124     sector_t nsect, struct bio **biop)
1125 {
1126 	*biop = NULL;
1127 	int error;
1128 
1129 #if defined(HAVE_BLKDEV_ISSUE_SECURE_ERASE)
1130 	error = blkdev_issue_secure_erase(BDH_BDEV(bdh),
1131 	    sector, nsect, GFP_NOFS);
1132 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_ASYNC_FLAGS)
1133 	error = __blkdev_issue_discard(BDH_BDEV(bdh),
1134 	    sector, nsect, GFP_NOFS, BLKDEV_DISCARD_SECURE, biop);
1135 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_FLAGS)
1136 	error = blkdev_issue_discard(BDH_BDEV(bdh),
1137 	    sector, nsect, GFP_NOFS, BLKDEV_DISCARD_SECURE);
1138 #else
1139 #error "unsupported kernel"
1140 #endif
1141 
1142 	return (error);
1143 }
1144 
1145 static int
vdev_bdev_issue_discard(zfs_bdev_handle_t * bdh,sector_t sector,sector_t nsect,struct bio ** biop)1146 vdev_bdev_issue_discard(zfs_bdev_handle_t *bdh, sector_t sector,
1147     sector_t nsect, struct bio **biop)
1148 {
1149 	*biop = NULL;
1150 	int error;
1151 
1152 #if defined(HAVE_BLKDEV_ISSUE_DISCARD_ASYNC_FLAGS)
1153 	error = __blkdev_issue_discard(BDH_BDEV(bdh),
1154 	    sector, nsect, GFP_NOFS, 0, biop);
1155 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_ASYNC_NOFLAGS)
1156 	error = __blkdev_issue_discard(BDH_BDEV(bdh),
1157 	    sector, nsect, GFP_NOFS, biop);
1158 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_FLAGS)
1159 	error = blkdev_issue_discard(BDH_BDEV(bdh),
1160 	    sector, nsect, GFP_NOFS, 0);
1161 #elif defined(HAVE_BLKDEV_ISSUE_DISCARD_NOFLAGS)
1162 	error = blkdev_issue_discard(BDH_BDEV(bdh),
1163 	    sector, nsect, GFP_NOFS);
1164 #else
1165 #error "unsupported kernel"
1166 #endif
1167 
1168 	return (error);
1169 }
1170 
1171 /*
1172  * Entry point for TRIM ops. This calls the right wrapper for secure erase or
1173  * discard, and then does the appropriate finishing work for error vs success
1174  * and async vs sync.
1175  */
1176 static int
vdev_disk_io_trim(zio_t * zio)1177 vdev_disk_io_trim(zio_t *zio)
1178 {
1179 	int error;
1180 	struct bio *bio;
1181 
1182 	zfs_bdev_handle_t *bdh = ((vdev_disk_t *)zio->io_vd->vdev_tsd)->vd_bdh;
1183 	sector_t sector = zio->io_offset >> 9;
1184 	sector_t nsects = zio->io_size >> 9;
1185 
1186 	if (zio->io_trim_flags & ZIO_TRIM_SECURE)
1187 		error = vdev_bdev_issue_secure_erase(bdh, sector, nsects, &bio);
1188 	else
1189 		error = vdev_bdev_issue_discard(bdh, sector, nsects, &bio);
1190 
1191 	if (error != 0)
1192 		return (SET_ERROR(-error));
1193 
1194 	if (bio == NULL) {
1195 		/*
1196 		 * This was a synchronous op that completed successfully, so
1197 		 * return it to ZFS immediately.
1198 		 */
1199 		zio_interrupt(zio);
1200 	} else {
1201 		/*
1202 		 * This was an asynchronous op; set up completion callback and
1203 		 * issue it.
1204 		 */
1205 		bio->bi_private = zio;
1206 		bio->bi_end_io = vdev_disk_discard_end_io;
1207 		vdev_submit_bio(bio);
1208 	}
1209 
1210 	return (0);
1211 }
1212 
1213 static void
vdev_disk_io_start(zio_t * zio)1214 vdev_disk_io_start(zio_t *zio)
1215 {
1216 	vdev_t *v = zio->io_vd;
1217 	vdev_disk_t *vd = v->vdev_tsd;
1218 	int error;
1219 
1220 	/*
1221 	 * If the vdev is closed, it's likely in the REMOVED or FAULTED state.
1222 	 * Nothing to be done here but return failure.
1223 	 */
1224 	if (vd == NULL) {
1225 		zio->io_error = ENXIO;
1226 		zio_interrupt(zio);
1227 		return;
1228 	}
1229 
1230 	rw_enter(&vd->vd_lock, RW_READER);
1231 
1232 	/*
1233 	 * If the vdev is closed, it's likely due to a failed reopen and is
1234 	 * in the UNAVAIL state.  Nothing to be done here but return failure.
1235 	 */
1236 	if (vd->vd_bdh == NULL) {
1237 		rw_exit(&vd->vd_lock);
1238 		zio->io_error = ENXIO;
1239 		zio_interrupt(zio);
1240 		return;
1241 	}
1242 
1243 	switch (zio->io_type) {
1244 	case ZIO_TYPE_FLUSH:
1245 
1246 		if (!vdev_readable(v)) {
1247 			/* Drive not there, can't flush */
1248 			error = SET_ERROR(ENXIO);
1249 		} else if (zfs_nocacheflush) {
1250 			/* Flushing disabled by operator, declare success */
1251 			error = 0;
1252 		} else if (v->vdev_nowritecache) {
1253 			/* This vdev not capable of flushing */
1254 			error = SET_ERROR(ENOTSUP);
1255 		} else {
1256 			/*
1257 			 * Issue the flush. If successful, the response will
1258 			 * be handled in the completion callback, so we're done.
1259 			 */
1260 			error = vdev_disk_io_flush(BDH_BDEV(vd->vd_bdh), zio);
1261 			if (error == 0) {
1262 				rw_exit(&vd->vd_lock);
1263 				return;
1264 			}
1265 		}
1266 
1267 		/* Couldn't issue the flush, so set the error and return it */
1268 		rw_exit(&vd->vd_lock);
1269 		zio->io_error = error;
1270 		zio_execute(zio);
1271 		return;
1272 
1273 	case ZIO_TYPE_TRIM:
1274 		error = vdev_disk_io_trim(zio);
1275 		rw_exit(&vd->vd_lock);
1276 		if (error) {
1277 			zio->io_error = error;
1278 			zio_execute(zio);
1279 		}
1280 		return;
1281 
1282 	case ZIO_TYPE_READ:
1283 	case ZIO_TYPE_WRITE:
1284 		zio->io_target_timestamp = zio_handle_io_delay(zio);
1285 		error = vdev_disk_io_rw(zio);
1286 		rw_exit(&vd->vd_lock);
1287 		if (error) {
1288 			zio->io_error = error;
1289 			zio_interrupt(zio);
1290 		}
1291 		return;
1292 
1293 	default:
1294 		/*
1295 		 * Getting here means our parent vdev has made a very strange
1296 		 * request of us, and shouldn't happen. Assert here to force a
1297 		 * crash in dev builds, but in production return the IO
1298 		 * unhandled. The pool will likely suspend anyway but that's
1299 		 * nicer than crashing the kernel.
1300 		 */
1301 		ASSERT3S(zio->io_type, ==, -1);
1302 
1303 		rw_exit(&vd->vd_lock);
1304 		zio->io_error = SET_ERROR(ENOTSUP);
1305 		zio_interrupt(zio);
1306 		return;
1307 	}
1308 
1309 	__builtin_unreachable();
1310 }
1311 
1312 static void
vdev_disk_io_done(zio_t * zio)1313 vdev_disk_io_done(zio_t *zio)
1314 {
1315 	/* If this was a read or write, we need to clean up the vbio */
1316 	if (zio->io_bio != NULL) {
1317 		vbio_t *vbio = zio->io_bio;
1318 		zio->io_bio = NULL;
1319 
1320 		/*
1321 		 * If we copied the ABD before issuing it, clean up and return
1322 		 * the copy to the ADB, with changes if appropriate.
1323 		 */
1324 		if (vbio->vbio_abd != NULL) {
1325 			if (zio->io_type == ZIO_TYPE_READ)
1326 				abd_copy(zio->io_abd, vbio->vbio_abd,
1327 				    zio->io_size);
1328 
1329 			abd_free(vbio->vbio_abd);
1330 			vbio->vbio_abd = NULL;
1331 		}
1332 
1333 		/* Final cleanup */
1334 		kmem_free(vbio, sizeof (vbio_t));
1335 	}
1336 
1337 	/*
1338 	 * If the device returned EIO, we revalidate the media.  If it is
1339 	 * determined the media has changed this triggers the asynchronous
1340 	 * removal of the device from the configuration.
1341 	 */
1342 	if (zio->io_error == EIO) {
1343 		vdev_t *v = zio->io_vd;
1344 		vdev_disk_t *vd = v->vdev_tsd;
1345 
1346 		if (!zfs_check_disk_status(BDH_BDEV(vd->vd_bdh))) {
1347 			invalidate_bdev(BDH_BDEV(vd->vd_bdh));
1348 			v->vdev_remove_wanted = B_TRUE;
1349 			spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
1350 		}
1351 	}
1352 }
1353 
1354 static void
vdev_disk_hold(vdev_t * vd)1355 vdev_disk_hold(vdev_t *vd)
1356 {
1357 	ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
1358 
1359 	/* We must have a pathname, and it must be absolute. */
1360 	if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
1361 		return;
1362 
1363 	/*
1364 	 * Only prefetch path and devid info if the device has
1365 	 * never been opened.
1366 	 */
1367 	if (vd->vdev_tsd != NULL)
1368 		return;
1369 
1370 }
1371 
1372 static void
vdev_disk_rele(vdev_t * vd)1373 vdev_disk_rele(vdev_t *vd)
1374 {
1375 	ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
1376 
1377 	/* XXX: Implement me as a vnode rele for the device */
1378 }
1379 
1380 vdev_ops_t vdev_disk_ops = {
1381 	.vdev_op_init = NULL,
1382 	.vdev_op_fini = NULL,
1383 	.vdev_op_open = vdev_disk_open,
1384 	.vdev_op_close = vdev_disk_close,
1385 	.vdev_op_asize_to_psize = vdev_default_psize,
1386 	.vdev_op_psize_to_asize = vdev_default_asize,
1387 	.vdev_op_min_asize = vdev_default_min_asize,
1388 	.vdev_op_min_alloc = NULL,
1389 	.vdev_op_io_start = vdev_disk_io_start,
1390 	.vdev_op_io_done = vdev_disk_io_done,
1391 	.vdev_op_state_change = NULL,
1392 	.vdev_op_need_resilver = NULL,
1393 	.vdev_op_hold = vdev_disk_hold,
1394 	.vdev_op_rele = vdev_disk_rele,
1395 	.vdev_op_remap = NULL,
1396 	.vdev_op_xlate = vdev_default_xlate,
1397 	.vdev_op_rebuild_asize = NULL,
1398 	.vdev_op_metaslab_init = NULL,
1399 	.vdev_op_config_generate = NULL,
1400 	.vdev_op_nparity = NULL,
1401 	.vdev_op_ndisks = NULL,
1402 	.vdev_op_type = VDEV_TYPE_DISK,		/* name of this vdev type */
1403 	.vdev_op_leaf = B_TRUE,			/* leaf vdev */
1404 	.vdev_op_kobj_evt_post = vdev_disk_kobj_evt_post
1405 };
1406 
1407 int
param_set_min_auto_ashift(const char * buf,zfs_kernel_param_t * kp)1408 param_set_min_auto_ashift(const char *buf, zfs_kernel_param_t *kp)
1409 {
1410 	uint_t val;
1411 	int error;
1412 
1413 	error = kstrtouint(buf, 0, &val);
1414 	if (error < 0)
1415 		return (SET_ERROR(error));
1416 
1417 	if (val < ASHIFT_MIN || val > zfs_vdev_max_auto_ashift)
1418 		return (SET_ERROR(-EINVAL));
1419 
1420 	error = param_set_uint(buf, kp);
1421 	if (error < 0)
1422 		return (SET_ERROR(error));
1423 
1424 	return (0);
1425 }
1426 
1427 int
param_set_max_auto_ashift(const char * buf,zfs_kernel_param_t * kp)1428 param_set_max_auto_ashift(const char *buf, zfs_kernel_param_t *kp)
1429 {
1430 	uint_t val;
1431 	int error;
1432 
1433 	error = kstrtouint(buf, 0, &val);
1434 	if (error < 0)
1435 		return (SET_ERROR(error));
1436 
1437 	if (val > ASHIFT_MAX || val < zfs_vdev_min_auto_ashift)
1438 		return (SET_ERROR(-EINVAL));
1439 
1440 	error = param_set_uint(buf, kp);
1441 	if (error < 0)
1442 		return (SET_ERROR(error));
1443 
1444 	return (0);
1445 }
1446 
1447 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, open_timeout_ms, UINT, ZMOD_RW,
1448 	"Timeout before determining that a device is missing");
1449 
1450 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, failfast_mask, UINT, ZMOD_RW,
1451 	"Defines failfast mask: 1 - device, 2 - transport, 4 - driver");
1452 
1453 ZFS_MODULE_PARAM(zfs_vdev_disk, zfs_vdev_disk_, max_segs, UINT, ZMOD_RW,
1454 	"Maximum number of data segments to add to an IO request (min 4)");
1455 
1456 ZFS_MODULE_PARAM(zfs_vdev_disk, zfs_vdev_disk_, calling_thread_io, UINT,
1457 	ZMOD_RW, "Enable calling thread io");
1458