4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (C) 2008-2010 Lawrence Livermore National Security, LLC.
23 * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
24 * Rewritten for Linux by Brian Behlendorf <behlendorf1@llnl.gov>.
26 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
29 #include <sys/zfs_context.h>
31 #include <sys/vdev_disk.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/fs/zfs.h>
35 #include <sys/sunldi.h>
37 char *zfs_vdev_scheduler
= VDEV_SCHEDULER
;
38 static void *zfs_vdev_holder
= VDEV_HOLDER
;
41 * Virtual device vector for disks.
43 typedef struct dio_request
{
44 zio_t
*dr_zio
; /* Parent ZIO */
45 atomic_t dr_ref
; /* References */
46 int dr_error
; /* Bio error */
47 int dr_bio_count
; /* Count of bio's */
48 struct bio
*dr_bio
[0]; /* Attached bio's */
52 #ifdef HAVE_OPEN_BDEV_EXCLUSIVE
54 vdev_bdev_mode(int smode
)
58 ASSERT3S(smode
& (FREAD
| FWRITE
), !=, 0);
70 vdev_bdev_mode(int smode
)
74 ASSERT3S(smode
& (FREAD
| FWRITE
), !=, 0);
76 if ((smode
& FREAD
) && !(smode
& FWRITE
))
81 #endif /* HAVE_OPEN_BDEV_EXCLUSIVE */
84 bdev_capacity(struct block_device
*bdev
)
86 struct hd_struct
*part
= bdev
->bd_part
;
88 /* The partition capacity referenced by the block device */
90 return (part
->nr_sects
<< 9);
92 /* Otherwise assume the full device capacity */
93 return (get_capacity(bdev
->bd_disk
) << 9);
97 vdev_disk_error(zio_t
*zio
)
100 printk("ZFS: zio error=%d type=%d offset=%llu size=%llu "
101 "flags=%x delay=%llu\n", zio
->io_error
, zio
->io_type
,
102 (u_longlong_t
)zio
->io_offset
, (u_longlong_t
)zio
->io_size
,
103 zio
->io_flags
, (u_longlong_t
)zio
->io_delay
);
108 * Use the Linux 'noop' elevator for zfs managed block devices. This
109 * strikes the ideal balance by allowing the zfs elevator to do all
110 * request ordering and prioritization. While allowing the Linux
111 * elevator to do the maximum front/back merging allowed by the
112 * physical device. This yields the largest possible requests for
113 * the device with the lowest total overhead.
116 vdev_elevator_switch(vdev_t
*v
, char *elevator
)
118 vdev_disk_t
*vd
= v
->vdev_tsd
;
119 struct block_device
*bdev
= vd
->vd_bdev
;
120 struct request_queue
*q
= bdev_get_queue(bdev
);
121 char *device
= bdev
->bd_disk
->disk_name
;
125 * Skip devices which are not whole disks (partitions).
126 * Device-mapper devices are excepted since they may be whole
127 * disks despite the vdev_wholedisk flag, in which case we can
128 * and should switch the elevator. If the device-mapper device
129 * does not have an elevator (i.e. dm-raid, dm-crypt, etc.) the
130 * "Skip devices without schedulers" check below will fail.
132 if (!v
->vdev_wholedisk
&& strncmp(device
, "dm-", 3) != 0)
135 /* Skip devices without schedulers (loop, ram, dm, etc) */
136 if (!q
->elevator
|| !blk_queue_stackable(q
))
139 /* Leave existing scheduler when set to "none" */
140 if ((strncmp(elevator
, "none", 4) == 0) && (strlen(elevator
) == 4))
143 #ifdef HAVE_ELEVATOR_CHANGE
144 error
= elevator_change(q
, elevator
);
147 * For pre-2.6.36 kernels elevator_change() is not available.
148 * Therefore we fall back to using a usermodehelper to echo the
149 * elevator into sysfs; This requires /bin/echo and sysfs to be
150 * mounted which may not be true early in the boot process.
152 #define SET_SCHEDULER_CMD \
153 "exec 0</dev/null " \
154 " 1>/sys/block/%s/queue/scheduler " \
159 char *argv
[] = { "/bin/sh", "-c", NULL
, NULL
};
160 char *envp
[] = { NULL
};
162 argv
[2] = kmem_asprintf(SET_SCHEDULER_CMD
, device
, elevator
);
163 error
= call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_PROC
);
166 #endif /* HAVE_ELEVATOR_CHANGE */
168 printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
169 elevator
, v
->vdev_path
, device
, error
);
175 * Expanding a whole disk vdev involves invoking BLKRRPART on the
176 * whole disk device. This poses a problem, because BLKRRPART will
177 * return EBUSY if one of the disk's partitions is open. That's why
178 * we have to do it here, just before opening the data partition.
179 * Unfortunately, BLKRRPART works by dropping all partitions and
180 * recreating them, which means that for a short time window, all
181 * /dev/sdxN device files disappear (until udev recreates them).
182 * This means two things:
183 * - When we open the data partition just after a BLKRRPART, we
184 * can't do it using the normal device file path because of the
185 * obvious race condition with udev. Instead, we use reliable
186 * kernel APIs to get a handle to the new partition device from
187 * the whole disk device.
188 * - Because vdev_disk_open() initially needs to find the device
189 * using its path, multiple vdev_disk_open() invocations in
190 * short succession on the same disk with BLKRRPARTs in the
191 * middle have a high probability of failure (because of the
192 * race condition with udev). A typical situation where this
193 * might happen is when the zpool userspace tool does a
194 * TRYIMPORT immediately followed by an IMPORT. For this
195 * reason, we only invoke BLKRRPART in the module when strictly
196 * necessary (zpool online -e case), and rely on userspace to
197 * do it when possible.
199 static struct block_device
*
200 vdev_disk_rrpart(const char *path
, int mode
, vdev_disk_t
*vd
)
202 #if defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK)
203 struct block_device
*bdev
, *result
= ERR_PTR(-ENXIO
);
204 struct gendisk
*disk
;
207 bdev
= vdev_bdev_open(path
, vdev_bdev_mode(mode
), zfs_vdev_holder
);
211 disk
= get_gendisk(bdev
->bd_dev
, &partno
);
212 vdev_bdev_close(bdev
, vdev_bdev_mode(mode
));
215 bdev
= bdget(disk_devt(disk
));
217 error
= blkdev_get(bdev
, vdev_bdev_mode(mode
), vd
);
219 error
= ioctl_by_bdev(bdev
, BLKRRPART
, 0);
220 vdev_bdev_close(bdev
, vdev_bdev_mode(mode
));
223 bdev
= bdget_disk(disk
, partno
);
225 error
= blkdev_get(bdev
,
226 vdev_bdev_mode(mode
) | FMODE_EXCL
, vd
);
235 return (ERR_PTR(-EOPNOTSUPP
));
236 #endif /* defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) */
240 vdev_disk_open(vdev_t
*v
, uint64_t *psize
, uint64_t *max_psize
,
243 struct block_device
*bdev
= ERR_PTR(-ENXIO
);
245 int count
= 0, mode
, block_size
;
247 /* Must have a pathname and it must be absolute. */
248 if (v
->vdev_path
== NULL
|| v
->vdev_path
[0] != '/') {
249 v
->vdev_stat
.vs_aux
= VDEV_AUX_BAD_LABEL
;
250 return (SET_ERROR(EINVAL
));
254 * Reopen the device if it's not currently open. Otherwise,
255 * just update the physical size of the device.
257 if (v
->vdev_tsd
!= NULL
) {
258 ASSERT(v
->vdev_reopening
);
263 vd
= kmem_zalloc(sizeof (vdev_disk_t
), KM_SLEEP
);
265 return (SET_ERROR(ENOMEM
));
268 * Devices are always opened by the path provided at configuration
269 * time. This means that if the provided path is a udev by-id path
270 * then drives may be recabled without an issue. If the provided
271 * path is a udev by-path path, then the physical location information
272 * will be preserved. This can be critical for more complicated
273 * configurations where drives are located in specific physical
274 * locations to maximize the systems tolerence to component failure.
275 * Alternatively, you can provide your own udev rule to flexibly map
276 * the drives as you see fit. It is not advised that you use the
277 * /dev/[hd]d devices which may be reordered due to probing order.
278 * Devices in the wrong locations will be detected by the higher
279 * level vdev validation.
281 * The specified paths may be briefly removed and recreated in
282 * response to udev events. This should be exceptionally unlikely
283 * because the zpool command makes every effort to verify these paths
284 * have already settled prior to reaching this point. Therefore,
285 * a ENOENT failure at this point is highly likely to be transient
286 * and it is reasonable to sleep and retry before giving up. In
287 * practice delays have been observed to be on the order of 100ms.
289 mode
= spa_mode(v
->vdev_spa
);
290 if (v
->vdev_wholedisk
&& v
->vdev_expanding
)
291 bdev
= vdev_disk_rrpart(v
->vdev_path
, mode
, vd
);
293 while (IS_ERR(bdev
) && count
< 50) {
294 bdev
= vdev_bdev_open(v
->vdev_path
,
295 vdev_bdev_mode(mode
), zfs_vdev_holder
);
296 if (unlikely(PTR_ERR(bdev
) == -ENOENT
)) {
299 } else if (IS_ERR(bdev
)) {
305 dprintf("failed open v->vdev_path=%s, error=%d count=%d\n",
306 v
->vdev_path
, -PTR_ERR(bdev
), count
);
307 kmem_free(vd
, sizeof (vdev_disk_t
));
308 return (SET_ERROR(-PTR_ERR(bdev
)));
315 /* Determine the physical block size */
316 block_size
= vdev_bdev_block_size(vd
->vd_bdev
);
318 /* Clear the nowritecache bit, causes vdev_reopen() to try again. */
319 v
->vdev_nowritecache
= B_FALSE
;
321 /* Inform the ZIO pipeline that we are non-rotational */
322 v
->vdev_nonrot
= blk_queue_nonrot(bdev_get_queue(vd
->vd_bdev
));
324 /* Physical volume size in bytes */
325 *psize
= bdev_capacity(vd
->vd_bdev
);
327 /* TODO: report possible expansion size */
330 /* Based on the minimum sector size set the block size */
331 *ashift
= highbit64(MAX(block_size
, SPA_MINBLOCKSIZE
)) - 1;
333 /* Try to set the io scheduler elevator algorithm */
334 (void) vdev_elevator_switch(v
, zfs_vdev_scheduler
);
340 vdev_disk_close(vdev_t
*v
)
342 vdev_disk_t
*vd
= v
->vdev_tsd
;
344 if (v
->vdev_reopening
|| vd
== NULL
)
347 if (vd
->vd_bdev
!= NULL
)
348 vdev_bdev_close(vd
->vd_bdev
,
349 vdev_bdev_mode(spa_mode(v
->vdev_spa
)));
351 kmem_free(vd
, sizeof (vdev_disk_t
));
355 static dio_request_t
*
356 vdev_disk_dio_alloc(int bio_count
)
361 dr
= kmem_zalloc(sizeof (dio_request_t
) +
362 sizeof (struct bio
*) * bio_count
, KM_SLEEP
);
364 atomic_set(&dr
->dr_ref
, 0);
365 dr
->dr_bio_count
= bio_count
;
368 for (i
= 0; i
< dr
->dr_bio_count
; i
++)
369 dr
->dr_bio
[i
] = NULL
;
376 vdev_disk_dio_free(dio_request_t
*dr
)
380 for (i
= 0; i
< dr
->dr_bio_count
; i
++)
382 bio_put(dr
->dr_bio
[i
]);
384 kmem_free(dr
, sizeof (dio_request_t
) +
385 sizeof (struct bio
*) * dr
->dr_bio_count
);
389 vdev_disk_dio_get(dio_request_t
*dr
)
391 atomic_inc(&dr
->dr_ref
);
395 vdev_disk_dio_put(dio_request_t
*dr
)
397 int rc
= atomic_dec_return(&dr
->dr_ref
);
400 * Free the dio_request when the last reference is dropped and
401 * ensure zio_interpret is called only once with the correct zio
404 zio_t
*zio
= dr
->dr_zio
;
405 int error
= dr
->dr_error
;
407 vdev_disk_dio_free(dr
);
410 zio
->io_delay
= jiffies_64
- zio
->io_delay
;
411 zio
->io_error
= error
;
412 ASSERT3S(zio
->io_error
, >=, 0);
414 vdev_disk_error(zio
);
422 BIO_END_IO_PROTO(vdev_disk_physio_completion
, bio
, error
)
424 dio_request_t
*dr
= bio
->bi_private
;
427 if (dr
->dr_error
== 0) {
428 #ifdef HAVE_1ARG_BIO_END_IO_T
429 dr
->dr_error
= BIO_END_IO_ERROR(bio
);
432 dr
->dr_error
= -(error
);
433 else if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
438 /* Drop reference aquired by __vdev_disk_physio */
439 rc
= vdev_disk_dio_put(dr
);
442 static inline unsigned long
443 bio_nr_pages(void *bio_ptr
, unsigned int bio_size
)
445 return ((((unsigned long)bio_ptr
+ bio_size
+ PAGE_SIZE
- 1) >>
446 PAGE_SHIFT
) - ((unsigned long)bio_ptr
>> PAGE_SHIFT
));
450 bio_map(struct bio
*bio
, void *bio_ptr
, unsigned int bio_size
)
452 unsigned int offset
, size
, i
;
455 offset
= offset_in_page(bio_ptr
);
456 for (i
= 0; i
< bio
->bi_max_vecs
; i
++) {
457 size
= PAGE_SIZE
- offset
;
465 if (is_vmalloc_addr(bio_ptr
))
466 page
= vmalloc_to_page(bio_ptr
);
468 page
= virt_to_page(bio_ptr
);
471 * Some network related block device uses tcp_sendpage, which
472 * doesn't behave well when using 0-count page, this is a
473 * safety net to catch them.
475 ASSERT3S(page_count(page
), >, 0);
477 if (bio_add_page(bio
, page
, size
, offset
) != size
)
489 vdev_submit_bio_impl(struct bio
*bio
)
491 #ifdef HAVE_1ARG_SUBMIT_BIO
499 vdev_submit_bio(struct bio
*bio
)
501 #ifdef HAVE_CURRENT_BIO_TAIL
502 struct bio
**bio_tail
= current
->bio_tail
;
503 current
->bio_tail
= NULL
;
504 vdev_submit_bio_impl(bio
);
505 current
->bio_tail
= bio_tail
;
507 struct bio_list
*bio_list
= current
->bio_list
;
508 current
->bio_list
= NULL
;
509 vdev_submit_bio_impl(bio
);
510 current
->bio_list
= bio_list
;
515 __vdev_disk_physio(struct block_device
*bdev
, zio_t
*zio
, caddr_t kbuf_ptr
,
516 size_t kbuf_size
, uint64_t kbuf_offset
, int rw
, int flags
)
521 int bio_size
, bio_count
= 16;
522 int i
= 0, error
= 0;
523 #if defined(HAVE_BLK_QUEUE_HAVE_BLK_PLUG)
524 struct blk_plug plug
;
527 ASSERT3U(kbuf_offset
+ kbuf_size
, <=, bdev
->bd_inode
->i_size
);
530 dr
= vdev_disk_dio_alloc(bio_count
);
534 if (zio
&& !(zio
->io_flags
& (ZIO_FLAG_IO_RETRY
| ZIO_FLAG_TRYHARD
)))
535 bio_set_flags_failfast(bdev
, &flags
);
540 * When the IO size exceeds the maximum bio size for the request
541 * queue we are forced to break the IO in multiple bio's and wait
542 * for them all to complete. Ideally, all pool users will set
543 * their volume block size to match the maximum request size and
544 * the common case will be one bio per vdev IO request.
547 bio_offset
= kbuf_offset
;
548 bio_size
= kbuf_size
;
549 for (i
= 0; i
<= dr
->dr_bio_count
; i
++) {
551 /* Finished constructing bio's for given buffer */
556 * By default only 'bio_count' bio's per dio are allowed.
557 * However, if we find ourselves in a situation where more
558 * are needed we allocate a larger dio and warn the user.
560 if (dr
->dr_bio_count
== i
) {
561 vdev_disk_dio_free(dr
);
566 /* bio_alloc() with __GFP_WAIT never returns NULL */
567 dr
->dr_bio
[i
] = bio_alloc(GFP_NOIO
,
568 MIN(bio_nr_pages(bio_ptr
, bio_size
), BIO_MAX_PAGES
));
569 if (unlikely(dr
->dr_bio
[i
] == NULL
)) {
570 vdev_disk_dio_free(dr
);
574 /* Matching put called by vdev_disk_physio_completion */
575 vdev_disk_dio_get(dr
);
577 dr
->dr_bio
[i
]->bi_bdev
= bdev
;
578 BIO_BI_SECTOR(dr
->dr_bio
[i
]) = bio_offset
>> 9;
579 dr
->dr_bio
[i
]->bi_end_io
= vdev_disk_physio_completion
;
580 dr
->dr_bio
[i
]->bi_private
= dr
;
581 bio_set_op_attrs(dr
->dr_bio
[i
], rw
, flags
);
583 /* Remaining size is returned to become the new size */
584 bio_size
= bio_map(dr
->dr_bio
[i
], bio_ptr
, bio_size
);
586 /* Advance in buffer and construct another bio if needed */
587 bio_ptr
+= BIO_BI_SIZE(dr
->dr_bio
[i
]);
588 bio_offset
+= BIO_BI_SIZE(dr
->dr_bio
[i
]);
591 /* Extra reference to protect dio_request during vdev_submit_bio */
592 vdev_disk_dio_get(dr
);
594 zio
->io_delay
= jiffies_64
;
596 #if defined(HAVE_BLK_QUEUE_HAVE_BLK_PLUG)
597 if (dr
->dr_bio_count
> 1)
598 blk_start_plug(&plug
);
601 /* Submit all bio's associated with this dio */
602 for (i
= 0; i
< dr
->dr_bio_count
; i
++)
604 vdev_submit_bio(dr
->dr_bio
[i
]);
606 #if defined(HAVE_BLK_QUEUE_HAVE_BLK_PLUG)
607 if (dr
->dr_bio_count
> 1)
608 blk_finish_plug(&plug
);
611 (void) vdev_disk_dio_put(dr
);
616 BIO_END_IO_PROTO(vdev_disk_io_flush_completion
, bio
, error
)
618 zio_t
*zio
= bio
->bi_private
;
619 #ifdef HAVE_1ARG_BIO_END_IO_T
620 zio
->io_error
= BIO_END_IO_ERROR(bio
);
622 zio
->io_error
= -error
;
625 zio
->io_delay
= jiffies_64
- zio
->io_delay
;
626 if (zio
->io_error
&& (zio
->io_error
== EOPNOTSUPP
))
627 zio
->io_vd
->vdev_nowritecache
= B_TRUE
;
630 ASSERT3S(zio
->io_error
, >=, 0);
632 vdev_disk_error(zio
);
637 vdev_disk_io_flush(struct block_device
*bdev
, zio_t
*zio
)
639 struct request_queue
*q
;
642 q
= bdev_get_queue(bdev
);
646 bio
= bio_alloc(GFP_NOIO
, 0);
647 /* bio_alloc() with __GFP_WAIT never returns NULL */
648 if (unlikely(bio
== NULL
))
651 bio
->bi_end_io
= vdev_disk_io_flush_completion
;
652 bio
->bi_private
= zio
;
654 zio
->io_delay
= jiffies_64
;
656 vdev_submit_bio(bio
);
657 invalidate_bdev(bdev
);
663 vdev_disk_io_start(zio_t
*zio
)
665 vdev_t
*v
= zio
->io_vd
;
666 vdev_disk_t
*vd
= v
->vdev_tsd
;
667 int rw
, flags
, error
;
669 switch (zio
->io_type
) {
672 if (!vdev_readable(v
)) {
673 zio
->io_error
= SET_ERROR(ENXIO
);
678 switch (zio
->io_cmd
) {
679 case DKIOCFLUSHWRITECACHE
:
681 if (zfs_nocacheflush
)
684 if (v
->vdev_nowritecache
) {
685 zio
->io_error
= SET_ERROR(ENOTSUP
);
689 error
= vdev_disk_io_flush(vd
->vd_bdev
, zio
);
693 zio
->io_error
= error
;
694 if (error
== ENOTSUP
)
695 v
->vdev_nowritecache
= B_TRUE
;
700 zio
->io_error
= SET_ERROR(ENOTSUP
);
707 #if defined(HAVE_BLK_QUEUE_HAVE_BIO_RW_UNPLUG)
708 flags
= (1 << BIO_RW_UNPLUG
);
709 #elif defined(REQ_UNPLUG)
718 #if defined(HAVE_BLK_QUEUE_HAVE_BIO_RW_UNPLUG)
719 flags
= (1 << BIO_RW_UNPLUG
);
720 #elif defined(REQ_UNPLUG)
728 zio
->io_error
= SET_ERROR(ENOTSUP
);
733 error
= __vdev_disk_physio(vd
->vd_bdev
, zio
, zio
->io_data
,
734 zio
->io_size
, zio
->io_offset
, rw
, flags
);
736 zio
->io_error
= error
;
743 vdev_disk_io_done(zio_t
*zio
)
746 * If the device returned EIO, we revalidate the media. If it is
747 * determined the media has changed this triggers the asynchronous
748 * removal of the device from the configuration.
750 if (zio
->io_error
== EIO
) {
751 vdev_t
*v
= zio
->io_vd
;
752 vdev_disk_t
*vd
= v
->vdev_tsd
;
754 if (check_disk_change(vd
->vd_bdev
)) {
755 vdev_bdev_invalidate(vd
->vd_bdev
);
756 v
->vdev_remove_wanted
= B_TRUE
;
757 spa_async_request(zio
->io_spa
, SPA_ASYNC_REMOVE
);
763 vdev_disk_hold(vdev_t
*vd
)
765 ASSERT(spa_config_held(vd
->vdev_spa
, SCL_STATE
, RW_WRITER
));
767 /* We must have a pathname, and it must be absolute. */
768 if (vd
->vdev_path
== NULL
|| vd
->vdev_path
[0] != '/')
772 * Only prefetch path and devid info if the device has
775 if (vd
->vdev_tsd
!= NULL
)
778 /* XXX: Implement me as a vnode lookup for the device */
779 vd
->vdev_name_vp
= NULL
;
780 vd
->vdev_devid_vp
= NULL
;
784 vdev_disk_rele(vdev_t
*vd
)
786 ASSERT(spa_config_held(vd
->vdev_spa
, SCL_STATE
, RW_WRITER
));
788 /* XXX: Implement me as a vnode rele for the device */
791 vdev_ops_t vdev_disk_ops
= {
800 VDEV_TYPE_DISK
, /* name of this vdev type */
801 B_TRUE
/* leaf vdev */
804 module_param(zfs_vdev_scheduler
, charp
, 0644);
805 MODULE_PARM_DESC(zfs_vdev_scheduler
, "I/O scheduler");