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) 2013 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 struct completion dr_comp; /* Completion for sync IO */
45 atomic_t dr_ref; /* References */
46 zio_t *dr_zio; /* Parent ZIO */
47 int dr_rw; /* Read/Write */
48 int dr_error; /* Bio error */
49 int dr_bio_count; /* Count of bio's */
50 struct bio *dr_bio[0]; /* Attached bio's */
54 #ifdef HAVE_OPEN_BDEV_EXCLUSIVE
56 vdev_bdev_mode(int smode)
60 ASSERT3S(smode & (FREAD | FWRITE), !=, 0);
72 vdev_bdev_mode(int smode)
76 ASSERT3S(smode & (FREAD | FWRITE), !=, 0);
78 if ((smode & FREAD) && !(smode & FWRITE))
83 #endif /* HAVE_OPEN_BDEV_EXCLUSIVE */
86 bdev_capacity(struct block_device *bdev)
88 struct hd_struct *part = bdev->bd_part;
90 /* The partition capacity referenced by the block device */
92 return (part->nr_sects << 9);
94 /* Otherwise assume the full device capacity */
95 return (get_capacity(bdev->bd_disk) << 9);
99 vdev_disk_error(zio_t *zio)
102 printk("ZFS: zio error=%d type=%d offset=%llu size=%llu "
103 "flags=%x delay=%llu\n", zio->io_error, zio->io_type,
104 (u_longlong_t)zio->io_offset, (u_longlong_t)zio->io_size,
105 zio->io_flags, (u_longlong_t)zio->io_delay);
110 * Use the Linux 'noop' elevator for zfs managed block devices. This
111 * strikes the ideal balance by allowing the zfs elevator to do all
112 * request ordering and prioritization. While allowing the Linux
113 * elevator to do the maximum front/back merging allowed by the
114 * physical device. This yields the largest possible requests for
115 * the device with the lowest total overhead.
118 vdev_elevator_switch(vdev_t *v, char *elevator)
120 vdev_disk_t *vd = v->vdev_tsd;
121 struct block_device *bdev = vd->vd_bdev;
122 struct request_queue *q = bdev_get_queue(bdev);
123 char *device = bdev->bd_disk->disk_name;
127 * Skip devices which are not whole disks (partitions).
128 * Device-mapper devices are excepted since they may be whole
129 * disks despite the vdev_wholedisk flag, in which case we can
130 * and should switch the elevator. If the device-mapper device
131 * does not have an elevator (i.e. dm-raid, dm-crypt, etc.) the
132 * "Skip devices without schedulers" check below will fail.
134 if (!v->vdev_wholedisk && strncmp(device, "dm-", 3) != 0)
137 /* Skip devices without schedulers (loop, ram, dm, etc) */
138 if (!q->elevator || !blk_queue_stackable(q))
141 /* Leave existing scheduler when set to "none" */
142 if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4))
145 #ifdef HAVE_ELEVATOR_CHANGE
146 error = elevator_change(q, elevator);
148 /* For pre-2.6.36 kernels elevator_change() is not available.
149 * Therefore we fall back to using a usermodehelper to echo the
150 * elevator into sysfs; This requires /bin/echo and sysfs to be
151 * mounted which may not be true early in the boot process.
153 # define SET_SCHEDULER_CMD \
154 "exec 0</dev/null " \
155 " 1>/sys/block/%s/queue/scheduler " \
160 char *argv[] = { "/bin/sh", "-c", NULL, NULL };
161 char *envp[] = { NULL };
163 argv[2] = kmem_asprintf(SET_SCHEDULER_CMD, device, elevator);
164 error = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
167 #endif /* HAVE_ELEVATOR_CHANGE */
169 printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
170 elevator, v->vdev_path, device, error);
176 * Expanding a whole disk vdev involves invoking BLKRRPART on the
177 * whole disk device. This poses a problem, because BLKRRPART will
178 * return EBUSY if one of the disk's partitions is open. That's why
179 * we have to do it here, just before opening the data partition.
180 * Unfortunately, BLKRRPART works by dropping all partitions and
181 * recreating them, which means that for a short time window, all
182 * /dev/sdxN device files disappear (until udev recreates them).
183 * This means two things:
184 * - When we open the data partition just after a BLKRRPART, we
185 * can't do it using the normal device file path because of the
186 * obvious race condition with udev. Instead, we use reliable
187 * kernel APIs to get a handle to the new partition device from
188 * the whole disk device.
189 * - Because vdev_disk_open() initially needs to find the device
190 * using its path, multiple vdev_disk_open() invocations in
191 * short succession on the same disk with BLKRRPARTs in the
192 * middle have a high probability of failure (because of the
193 * race condition with udev). A typical situation where this
194 * might happen is when the zpool userspace tool does a
195 * TRYIMPORT immediately followed by an IMPORT. For this
196 * reason, we only invoke BLKRRPART in the module when strictly
197 * necessary (zpool online -e case), and rely on userspace to
198 * do it when possible.
200 static struct block_device *
201 vdev_disk_rrpart(const char *path, int mode, vdev_disk_t *vd)
203 #if defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK)
204 struct block_device *bdev, *result = ERR_PTR(-ENXIO);
205 struct gendisk *disk;
208 bdev = vdev_bdev_open(path, vdev_bdev_mode(mode), zfs_vdev_holder);
212 disk = get_gendisk(bdev->bd_dev, &partno);
213 vdev_bdev_close(bdev, vdev_bdev_mode(mode));
216 bdev = bdget(disk_devt(disk));
218 error = blkdev_get(bdev, vdev_bdev_mode(mode), vd);
220 error = ioctl_by_bdev(bdev, BLKRRPART, 0);
221 vdev_bdev_close(bdev, vdev_bdev_mode(mode));
224 bdev = bdget_disk(disk, partno);
226 error = blkdev_get(bdev,
227 vdev_bdev_mode(mode) | FMODE_EXCL, vd);
236 return ERR_PTR(-EOPNOTSUPP);
237 #endif /* defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) */
241 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
244 struct block_device *bdev = ERR_PTR(-ENXIO);
246 int mode, block_size;
248 /* Must have a pathname and it must be absolute. */
249 if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
250 v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
255 * Reopen the device if it's not currently open. Otherwise,
256 * just update the physical size of the device.
258 if (v->vdev_tsd != NULL) {
259 ASSERT(v->vdev_reopening);
264 vd = kmem_zalloc(sizeof(vdev_disk_t), KM_PUSHPAGE);
269 * Devices are always opened by the path provided at configuration
270 * time. This means that if the provided path is a udev by-id path
271 * then drives may be recabled without an issue. If the provided
272 * path is a udev by-path path, then the physical location information
273 * will be preserved. This can be critical for more complicated
274 * configurations where drives are located in specific physical
275 * locations to maximize the systems tolerence to component failure.
276 * Alternatively, you can provide your own udev rule to flexibly map
277 * the drives as you see fit. It is not advised that you use the
278 * /dev/[hd]d devices which may be reordered due to probing order.
279 * Devices in the wrong locations will be detected by the higher
280 * level vdev validation.
282 mode = spa_mode(v->vdev_spa);
283 if (v->vdev_wholedisk && v->vdev_expanding)
284 bdev = vdev_disk_rrpart(v->vdev_path, mode, vd);
286 bdev = vdev_bdev_open(v->vdev_path,
287 vdev_bdev_mode(mode), zfs_vdev_holder);
289 kmem_free(vd, sizeof(vdev_disk_t));
290 return -PTR_ERR(bdev);
297 /* Determine the physical block size */
298 block_size = vdev_bdev_block_size(vd->vd_bdev);
300 /* Clear the nowritecache bit, causes vdev_reopen() to try again. */
301 v->vdev_nowritecache = B_FALSE;
303 /* Physical volume size in bytes */
304 *psize = bdev_capacity(vd->vd_bdev);
306 /* TODO: report possible expansion size */
309 /* Based on the minimum sector size set the block size */
310 *ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
312 /* Try to set the io scheduler elevator algorithm */
313 (void) vdev_elevator_switch(v, zfs_vdev_scheduler);
319 vdev_disk_close(vdev_t *v)
321 vdev_disk_t *vd = v->vdev_tsd;
323 if (v->vdev_reopening || vd == NULL)
326 if (vd->vd_bdev != NULL)
327 vdev_bdev_close(vd->vd_bdev,
328 vdev_bdev_mode(spa_mode(v->vdev_spa)));
330 kmem_free(vd, sizeof(vdev_disk_t));
334 static dio_request_t *
335 vdev_disk_dio_alloc(int bio_count)
340 dr = kmem_zalloc(sizeof(dio_request_t) +
341 sizeof(struct bio *) * bio_count, KM_PUSHPAGE);
343 init_completion(&dr->dr_comp);
344 atomic_set(&dr->dr_ref, 0);
345 dr->dr_bio_count = bio_count;
348 for (i = 0; i < dr->dr_bio_count; i++)
349 dr->dr_bio[i] = NULL;
356 vdev_disk_dio_free(dio_request_t *dr)
360 for (i = 0; i < dr->dr_bio_count; i++)
362 bio_put(dr->dr_bio[i]);
364 kmem_free(dr, sizeof(dio_request_t) +
365 sizeof(struct bio *) * dr->dr_bio_count);
369 vdev_disk_dio_is_sync(dio_request_t *dr)
371 #ifdef HAVE_BIO_RW_SYNC
372 /* BIO_RW_SYNC preferred interface from 2.6.12-2.6.29 */
373 return (dr->dr_rw & (1 << BIO_RW_SYNC));
375 # ifdef HAVE_BIO_RW_SYNCIO
376 /* BIO_RW_SYNCIO preferred interface from 2.6.30-2.6.35 */
377 return (dr->dr_rw & (1 << BIO_RW_SYNCIO));
379 # ifdef HAVE_REQ_SYNC
380 /* REQ_SYNC preferred interface from 2.6.36-2.6.xx */
381 return (dr->dr_rw & REQ_SYNC);
383 # error "Unable to determine bio sync flag"
384 # endif /* HAVE_REQ_SYNC */
385 # endif /* HAVE_BIO_RW_SYNC */
386 #endif /* HAVE_BIO_RW_SYNCIO */
390 vdev_disk_dio_get(dio_request_t *dr)
392 atomic_inc(&dr->dr_ref);
396 vdev_disk_dio_put(dio_request_t *dr)
398 int rc = atomic_dec_return(&dr->dr_ref);
401 * Free the dio_request when the last reference is dropped and
402 * ensure zio_interpret is called only once with the correct zio
405 zio_t *zio = dr->dr_zio;
406 int error = dr->dr_error;
408 vdev_disk_dio_free(dr);
411 zio->io_delay = jiffies_64 - zio->io_delay;
412 zio->io_error = error;
413 ASSERT3S(zio->io_error, >=, 0);
415 vdev_disk_error(zio);
423 BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, size, error)
425 dio_request_t *dr = bio->bi_private;
428 /* Fatal error but print some useful debugging before asserting */
430 PANIC("dr == NULL, bio->bi_private == NULL\n"
431 "bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d\n"
432 "bi_idx: %d, bi_size: %d, bi_end_io: %p, bi_cnt: %d\n",
433 bio->bi_next, bio->bi_flags, bio->bi_rw, bio->bi_vcnt,
434 bio->bi_idx, bio->bi_size, bio->bi_end_io,
435 atomic_read(&bio->bi_cnt));
437 #ifndef HAVE_2ARGS_BIO_END_IO_T
440 #endif /* HAVE_2ARGS_BIO_END_IO_T */
442 if (error == 0 && !test_bit(BIO_UPTODATE, &bio->bi_flags))
445 if (dr->dr_error == 0)
446 dr->dr_error = -error;
448 /* Drop reference aquired by __vdev_disk_physio */
449 rc = vdev_disk_dio_put(dr);
451 /* Wake up synchronous waiter this is the last outstanding bio */
452 if ((rc == 1) && vdev_disk_dio_is_sync(dr))
453 complete(&dr->dr_comp);
455 BIO_END_IO_RETURN(0);
458 static inline unsigned long
459 bio_nr_pages(void *bio_ptr, unsigned int bio_size)
461 return ((((unsigned long)bio_ptr + bio_size + PAGE_SIZE - 1) >>
462 PAGE_SHIFT) - ((unsigned long)bio_ptr >> PAGE_SHIFT));
466 bio_map(struct bio *bio, void *bio_ptr, unsigned int bio_size)
468 unsigned int offset, size, i;
471 offset = offset_in_page(bio_ptr);
472 for (i = 0; i < bio->bi_max_vecs; i++) {
473 size = PAGE_SIZE - offset;
481 if (kmem_virt(bio_ptr))
482 page = vmalloc_to_page(bio_ptr);
484 page = virt_to_page(bio_ptr);
486 if (bio_add_page(bio, page, size, offset) != size)
498 __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
499 size_t kbuf_size, uint64_t kbuf_offset, int flags)
504 int bio_size, bio_count = 16;
505 int i = 0, error = 0;
507 ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);
510 dr = vdev_disk_dio_alloc(bio_count);
514 if (zio && !(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))
515 bio_set_flags_failfast(bdev, &flags);
521 * When the IO size exceeds the maximum bio size for the request
522 * queue we are forced to break the IO in multiple bio's and wait
523 * for them all to complete. Ideally, all pool users will set
524 * their volume block size to match the maximum request size and
525 * the common case will be one bio per vdev IO request.
528 bio_offset = kbuf_offset;
529 bio_size = kbuf_size;
530 for (i = 0; i <= dr->dr_bio_count; i++) {
532 /* Finished constructing bio's for given buffer */
537 * By default only 'bio_count' bio's per dio are allowed.
538 * However, if we find ourselves in a situation where more
539 * are needed we allocate a larger dio and warn the user.
541 if (dr->dr_bio_count == i) {
542 vdev_disk_dio_free(dr);
547 dr->dr_bio[i] = bio_alloc(GFP_NOIO,
548 bio_nr_pages(bio_ptr, bio_size));
549 if (dr->dr_bio[i] == NULL) {
550 vdev_disk_dio_free(dr);
554 /* Matching put called by vdev_disk_physio_completion */
555 vdev_disk_dio_get(dr);
557 dr->dr_bio[i]->bi_bdev = bdev;
558 dr->dr_bio[i]->bi_sector = bio_offset >> 9;
559 dr->dr_bio[i]->bi_rw = dr->dr_rw;
560 dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion;
561 dr->dr_bio[i]->bi_private = dr;
563 /* Remaining size is returned to become the new size */
564 bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size);
566 /* Advance in buffer and construct another bio if needed */
567 bio_ptr += dr->dr_bio[i]->bi_size;
568 bio_offset += dr->dr_bio[i]->bi_size;
571 /* Extra reference to protect dio_request during submit_bio */
572 vdev_disk_dio_get(dr);
574 zio->io_delay = jiffies_64;
576 /* Submit all bio's associated with this dio */
577 for (i = 0; i < dr->dr_bio_count; i++)
579 submit_bio(dr->dr_rw, dr->dr_bio[i]);
582 * On synchronous blocking requests we wait for all bio the completion
583 * callbacks to run. We will be woken when the last callback runs
584 * for this dio. We are responsible for putting the last dio_request
585 * reference will in turn put back the last bio references. The
586 * only synchronous consumer is vdev_disk_read_rootlabel() all other
587 * IO originating from vdev_disk_io_start() is asynchronous.
589 if (vdev_disk_dio_is_sync(dr)) {
590 wait_for_completion(&dr->dr_comp);
591 error = dr->dr_error;
592 ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
595 (void)vdev_disk_dio_put(dr);
601 vdev_disk_physio(struct block_device *bdev, caddr_t kbuf,
602 size_t size, uint64_t offset, int flags)
604 bio_set_flags_failfast(bdev, &flags);
605 return __vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags);
608 BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, size, rc)
610 zio_t *zio = bio->bi_private;
612 zio->io_delay = jiffies_64 - zio->io_delay;
614 if (rc && (rc == -EOPNOTSUPP))
615 zio->io_vd->vdev_nowritecache = B_TRUE;
618 ASSERT3S(zio->io_error, >=, 0);
620 vdev_disk_error(zio);
623 BIO_END_IO_RETURN(0);
627 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
629 struct request_queue *q;
632 q = bdev_get_queue(bdev);
636 bio = bio_alloc(GFP_NOIO, 0);
640 bio->bi_end_io = vdev_disk_io_flush_completion;
641 bio->bi_private = zio;
643 zio->io_delay = jiffies_64;
644 submit_bio(VDEV_WRITE_FLUSH_FUA, bio);
650 vdev_disk_io_start(zio_t *zio)
652 vdev_t *v = zio->io_vd;
653 vdev_disk_t *vd = v->vdev_tsd;
656 switch (zio->io_type) {
659 if (!vdev_readable(v)) {
660 zio->io_error = SET_ERROR(ENXIO);
661 return ZIO_PIPELINE_CONTINUE;
664 switch (zio->io_cmd) {
665 case DKIOCFLUSHWRITECACHE:
667 if (zfs_nocacheflush)
670 if (v->vdev_nowritecache) {
671 zio->io_error = SET_ERROR(ENOTSUP);
675 error = vdev_disk_io_flush(vd->vd_bdev, zio);
677 return ZIO_PIPELINE_STOP;
679 zio->io_error = error;
680 if (error == ENOTSUP)
681 v->vdev_nowritecache = B_TRUE;
686 zio->io_error = SET_ERROR(ENOTSUP);
689 return ZIO_PIPELINE_CONTINUE;
700 zio->io_error = SET_ERROR(ENOTSUP);
701 return ZIO_PIPELINE_CONTINUE;
704 error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,
705 zio->io_size, zio->io_offset, flags);
707 zio->io_error = error;
708 return ZIO_PIPELINE_CONTINUE;
711 return ZIO_PIPELINE_STOP;
715 vdev_disk_io_done(zio_t *zio)
718 * If the device returned EIO, we revalidate the media. If it is
719 * determined the media has changed this triggers the asynchronous
720 * removal of the device from the configuration.
722 if (zio->io_error == EIO) {
723 vdev_t *v = zio->io_vd;
724 vdev_disk_t *vd = v->vdev_tsd;
726 if (check_disk_change(vd->vd_bdev)) {
727 vdev_bdev_invalidate(vd->vd_bdev);
728 v->vdev_remove_wanted = B_TRUE;
729 spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
735 vdev_disk_hold(vdev_t *vd)
737 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
739 /* We must have a pathname, and it must be absolute. */
740 if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
744 * Only prefetch path and devid info if the device has
747 if (vd->vdev_tsd != NULL)
750 /* XXX: Implement me as a vnode lookup for the device */
751 vd->vdev_name_vp = NULL;
752 vd->vdev_devid_vp = NULL;
756 vdev_disk_rele(vdev_t *vd)
758 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
760 /* XXX: Implement me as a vnode rele for the device */
763 vdev_ops_t vdev_disk_ops = {
772 VDEV_TYPE_DISK, /* name of this vdev type */
773 B_TRUE /* leaf vdev */
777 * Given the root disk device devid or pathname, read the label from
778 * the device, and construct a configuration nvlist.
781 vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
783 struct block_device *bdev;
788 bdev = vdev_bdev_open(devpath, vdev_bdev_mode(FREAD), zfs_vdev_holder);
790 return -PTR_ERR(bdev);
792 s = bdev_capacity(bdev);
794 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
798 size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t);
799 label = vmem_alloc(sizeof(vdev_label_t), KM_PUSHPAGE);
801 for (i = 0; i < VDEV_LABELS; i++) {
802 uint64_t offset, state, txg = 0;
804 /* read vdev label */
805 offset = vdev_label_offset(size, i, 0);
806 if (vdev_disk_physio(bdev, (caddr_t)label,
807 VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, READ_SYNC) != 0)
810 if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist,
811 sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) {
816 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
817 &state) != 0 || state >= POOL_STATE_DESTROYED) {
818 nvlist_free(*config);
823 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
824 &txg) != 0 || txg == 0) {
825 nvlist_free(*config);
833 vmem_free(label, sizeof(vdev_label_t));
834 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
839 module_param(zfs_vdev_scheduler, charp, 0644);
840 MODULE_PARM_DESC(zfs_vdev_scheduler, "I/O scheduler");