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>.
28 #include <sys/zfs_context.h>
30 #include <sys/vdev_disk.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/fs/zfs.h>
34 #include <sys/sunldi.h>
36 char *zfs_vdev_scheduler = VDEV_SCHEDULER;
39 * Virtual device vector for disks.
41 typedef struct dio_request {
42 struct completion dr_comp; /* Completion for sync IO */
43 atomic_t dr_ref; /* References */
44 zio_t *dr_zio; /* Parent ZIO */
45 int dr_rw; /* Read/Write */
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;
124 /* Skip devices which are not whole disks (partitions) */
125 if (!v->vdev_wholedisk)
128 /* Skip devices without schedulers (loop, ram, dm, etc) */
129 if (!q->elevator || !blk_queue_stackable(q))
132 /* Leave existing scheduler when set to "none" */
133 if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4))
136 #ifdef HAVE_ELEVATOR_CHANGE
137 error = elevator_change(q, elevator);
139 /* For pre-2.6.36 kernels elevator_change() is not available.
140 * Therefore we fall back to using a usermodehelper to echo the
141 * elevator into sysfs; This requires /bin/echo and sysfs to be
142 * mounted which may not be true early in the boot process.
144 # define SET_SCHEDULER_CMD \
145 "exec 0</dev/null " \
146 " 1>/sys/block/%s/queue/scheduler " \
151 char *argv[] = { "/bin/sh", "-c", NULL, NULL };
152 char *envp[] = { NULL };
154 argv[2] = kmem_asprintf(SET_SCHEDULER_CMD, device, elevator);
155 error = call_usermodehelper(argv[0], argv, envp, 1);
158 #endif /* HAVE_ELEVATOR_CHANGE */
160 printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
161 elevator, v->vdev_path, device, error);
167 * Expanding a whole disk vdev involves invoking BLKRRPART on the
168 * whole disk device. This poses a problem, because BLKRRPART will
169 * return EBUSY if one of the disk's partitions is open. That's why
170 * we have to do it here, just before opening the data partition.
171 * Unfortunately, BLKRRPART works by dropping all partitions and
172 * recreating them, which means that for a short time window, all
173 * /dev/sdxN device files disappear (until udev recreates them).
174 * This means two things:
175 * - When we open the data partition just after a BLKRRPART, we
176 * can't do it using the normal device file path because of the
177 * obvious race condition with udev. Instead, we use reliable
178 * kernel APIs to get a handle to the new partition device from
179 * the whole disk device.
180 * - Because vdev_disk_open() initially needs to find the device
181 * using its path, multiple vdev_disk_open() invocations in
182 * short succession on the same disk with BLKRRPARTs in the
183 * middle have a high probability of failure (because of the
184 * race condition with udev). A typical situation where this
185 * might happen is when the zpool userspace tool does a
186 * TRYIMPORT immediately followed by an IMPORT. For this
187 * reason, we only invoke BLKRRPART in the module when strictly
188 * necessary (zpool online -e case), and rely on userspace to
189 * do it when possible.
191 static struct block_device *
192 vdev_disk_rrpart(const char *path, int mode, vdev_disk_t *vd)
194 #if defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK)
195 struct block_device *bdev, *result = ERR_PTR(-ENXIO);
196 struct gendisk *disk;
199 bdev = vdev_bdev_open(path, vdev_bdev_mode(mode), vd);
203 disk = get_gendisk(bdev->bd_dev, &partno);
204 vdev_bdev_close(bdev, vdev_bdev_mode(mode));
207 bdev = bdget(disk_devt(disk));
209 error = blkdev_get(bdev, vdev_bdev_mode(mode), vd);
211 error = ioctl_by_bdev(bdev, BLKRRPART, 0);
212 vdev_bdev_close(bdev, vdev_bdev_mode(mode));
215 bdev = bdget_disk(disk, partno);
217 error = blkdev_get(bdev,
218 vdev_bdev_mode(mode) | FMODE_EXCL, vd);
227 return ERR_PTR(-EOPNOTSUPP);
228 #endif /* defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) */
232 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
235 struct block_device *bdev = ERR_PTR(-ENXIO);
237 int mode, block_size;
239 /* Must have a pathname and it must be absolute. */
240 if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
241 v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
245 vd = kmem_zalloc(sizeof(vdev_disk_t), KM_PUSHPAGE);
250 * Devices are always opened by the path provided at configuration
251 * time. This means that if the provided path is a udev by-id path
252 * then drives may be recabled without an issue. If the provided
253 * path is a udev by-path path then the physical location information
254 * will be preserved. This can be critical for more complicated
255 * configurations where drives are located in specific physical
256 * locations to maximize the systems tolerence to component failure.
257 * Alternately you can provide your own udev rule to flexibly map
258 * the drives as you see fit. It is not advised that you use the
259 * /dev/[hd]d devices which may be reorder due to probing order.
260 * Devices in the wrong locations will be detected by the higher
261 * level vdev validation.
263 mode = spa_mode(v->vdev_spa);
264 if (v->vdev_wholedisk && v->vdev_expanding)
265 bdev = vdev_disk_rrpart(v->vdev_path, mode, vd);
267 bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd);
269 kmem_free(vd, sizeof(vdev_disk_t));
270 return -PTR_ERR(bdev);
275 block_size = vdev_bdev_block_size(bdev);
277 /* We think the wholedisk property should always be set when this
278 * function is called. ASSERT here so if any legitimate cases exist
279 * where it's not set, we'll find them during debugging. If we never
280 * hit the ASSERT, this and the following conditional statement can be
282 ASSERT3S(v->vdev_wholedisk, !=, -1ULL);
284 /* The wholedisk property was initialized to -1 in vdev_alloc() if it
285 * was unspecified. In that case, check if this is a whole device.
286 * When bdev->bd_contains == bdev we have a whole device and not simply
288 if (v->vdev_wholedisk == -1ULL)
289 v->vdev_wholedisk = (bdev->bd_contains == bdev);
291 /* Clear the nowritecache bit, causes vdev_reopen() to try again. */
292 v->vdev_nowritecache = B_FALSE;
294 /* Physical volume size in bytes */
295 *psize = bdev_capacity(bdev);
297 /* TODO: report possible expansion size */
300 /* Based on the minimum sector size set the block size */
301 *ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
303 /* Try to set the io scheduler elevator algorithm */
304 (void) vdev_elevator_switch(v, zfs_vdev_scheduler);
310 vdev_disk_close(vdev_t *v)
312 vdev_disk_t *vd = v->vdev_tsd;
317 if (vd->vd_bdev != NULL)
318 vdev_bdev_close(vd->vd_bdev,
319 vdev_bdev_mode(spa_mode(v->vdev_spa)));
321 kmem_free(vd, sizeof(vdev_disk_t));
325 static dio_request_t *
326 vdev_disk_dio_alloc(int bio_count)
331 dr = kmem_zalloc(sizeof(dio_request_t) +
332 sizeof(struct bio *) * bio_count, KM_PUSHPAGE);
334 init_completion(&dr->dr_comp);
335 atomic_set(&dr->dr_ref, 0);
336 dr->dr_bio_count = bio_count;
339 for (i = 0; i < dr->dr_bio_count; i++)
340 dr->dr_bio[i] = NULL;
347 vdev_disk_dio_free(dio_request_t *dr)
351 for (i = 0; i < dr->dr_bio_count; i++)
353 bio_put(dr->dr_bio[i]);
355 kmem_free(dr, sizeof(dio_request_t) +
356 sizeof(struct bio *) * dr->dr_bio_count);
360 vdev_disk_dio_is_sync(dio_request_t *dr)
362 #ifdef HAVE_BIO_RW_SYNC
363 /* BIO_RW_SYNC preferred interface from 2.6.12-2.6.29 */
364 return (dr->dr_rw & (1 << BIO_RW_SYNC));
366 # ifdef HAVE_BIO_RW_SYNCIO
367 /* BIO_RW_SYNCIO preferred interface from 2.6.30-2.6.35 */
368 return (dr->dr_rw & (1 << BIO_RW_SYNCIO));
370 # ifdef HAVE_REQ_SYNC
371 /* REQ_SYNC preferred interface from 2.6.36-2.6.xx */
372 return (dr->dr_rw & REQ_SYNC);
374 # error "Unable to determine bio sync flag"
375 # endif /* HAVE_REQ_SYNC */
376 # endif /* HAVE_BIO_RW_SYNC */
377 #endif /* HAVE_BIO_RW_SYNCIO */
381 vdev_disk_dio_get(dio_request_t *dr)
383 atomic_inc(&dr->dr_ref);
387 vdev_disk_dio_put(dio_request_t *dr)
389 int rc = atomic_dec_return(&dr->dr_ref);
392 * Free the dio_request when the last reference is dropped and
393 * ensure zio_interpret is called only once with the correct zio
396 zio_t *zio = dr->dr_zio;
397 int error = dr->dr_error;
399 vdev_disk_dio_free(dr);
402 zio->io_delay = jiffies_to_msecs(
403 jiffies_64 - zio->io_delay);
404 zio->io_error = error;
405 ASSERT3S(zio->io_error, >=, 0);
407 vdev_disk_error(zio);
415 BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, size, error)
417 dio_request_t *dr = bio->bi_private;
420 /* Fatal error but print some useful debugging before asserting */
422 PANIC("dr == NULL, bio->bi_private == NULL\n"
423 "bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d\n"
424 "bi_idx: %d, bi_size: %d, bi_end_io: %p, bi_cnt: %d\n",
425 bio->bi_next, bio->bi_flags, bio->bi_rw, bio->bi_vcnt,
426 bio->bi_idx, bio->bi_size, bio->bi_end_io,
427 atomic_read(&bio->bi_cnt));
429 #ifndef HAVE_2ARGS_BIO_END_IO_T
432 #endif /* HAVE_2ARGS_BIO_END_IO_T */
434 if (error == 0 && !test_bit(BIO_UPTODATE, &bio->bi_flags))
437 if (dr->dr_error == 0)
438 dr->dr_error = -error;
440 /* Drop reference aquired by __vdev_disk_physio */
441 rc = vdev_disk_dio_put(dr);
443 /* Wake up synchronous waiter this is the last outstanding bio */
444 if ((rc == 1) && vdev_disk_dio_is_sync(dr))
445 complete(&dr->dr_comp);
447 BIO_END_IO_RETURN(0);
450 static inline unsigned long
451 bio_nr_pages(void *bio_ptr, unsigned int bio_size)
453 return ((((unsigned long)bio_ptr + bio_size + PAGE_SIZE - 1) >>
454 PAGE_SHIFT) - ((unsigned long)bio_ptr >> PAGE_SHIFT));
458 bio_map(struct bio *bio, void *bio_ptr, unsigned int bio_size)
460 unsigned int offset, size, i;
463 offset = offset_in_page(bio_ptr);
464 for (i = 0; i < bio->bi_max_vecs; i++) {
465 size = PAGE_SIZE - offset;
473 if (kmem_virt(bio_ptr))
474 page = vmalloc_to_page(bio_ptr);
476 page = virt_to_page(bio_ptr);
478 if (bio_add_page(bio, page, size, offset) != size)
490 __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
491 size_t kbuf_size, uint64_t kbuf_offset, int flags)
496 int bio_size, bio_count = 16;
497 int i = 0, error = 0;
499 ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);
502 dr = vdev_disk_dio_alloc(bio_count);
506 if (zio && !(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))
507 bio_set_flags_failfast(bdev, &flags);
513 * When the IO size exceeds the maximum bio size for the request
514 * queue we are forced to break the IO in multiple bio's and wait
515 * for them all to complete. Ideally, all pool users will set
516 * their volume block size to match the maximum request size and
517 * the common case will be one bio per vdev IO request.
520 bio_offset = kbuf_offset;
521 bio_size = kbuf_size;
522 for (i = 0; i <= dr->dr_bio_count; i++) {
524 /* Finished constructing bio's for given buffer */
529 * By default only 'bio_count' bio's per dio are allowed.
530 * However, if we find ourselves in a situation where more
531 * are needed we allocate a larger dio and warn the user.
533 if (dr->dr_bio_count == i) {
534 vdev_disk_dio_free(dr);
539 dr->dr_bio[i] = bio_alloc(GFP_NOIO,
540 bio_nr_pages(bio_ptr, bio_size));
541 if (dr->dr_bio[i] == NULL) {
542 vdev_disk_dio_free(dr);
546 /* Matching put called by vdev_disk_physio_completion */
547 vdev_disk_dio_get(dr);
549 dr->dr_bio[i]->bi_bdev = bdev;
550 dr->dr_bio[i]->bi_sector = bio_offset >> 9;
551 dr->dr_bio[i]->bi_rw = dr->dr_rw;
552 dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion;
553 dr->dr_bio[i]->bi_private = dr;
555 /* Remaining size is returned to become the new size */
556 bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size);
558 /* Advance in buffer and construct another bio if needed */
559 bio_ptr += dr->dr_bio[i]->bi_size;
560 bio_offset += dr->dr_bio[i]->bi_size;
563 /* Extra reference to protect dio_request during submit_bio */
564 vdev_disk_dio_get(dr);
566 zio->io_delay = jiffies_64;
568 /* Submit all bio's associated with this dio */
569 for (i = 0; i < dr->dr_bio_count; i++)
571 submit_bio(dr->dr_rw, dr->dr_bio[i]);
574 * On synchronous blocking requests we wait for all bio the completion
575 * callbacks to run. We will be woken when the last callback runs
576 * for this dio. We are responsible for putting the last dio_request
577 * reference will in turn put back the last bio references. The
578 * only synchronous consumer is vdev_disk_read_rootlabel() all other
579 * IO originating from vdev_disk_io_start() is asynchronous.
581 if (vdev_disk_dio_is_sync(dr)) {
582 wait_for_completion(&dr->dr_comp);
583 error = dr->dr_error;
584 ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
587 (void)vdev_disk_dio_put(dr);
593 vdev_disk_physio(struct block_device *bdev, caddr_t kbuf,
594 size_t size, uint64_t offset, int flags)
596 bio_set_flags_failfast(bdev, &flags);
597 return __vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags);
600 /* 2.6.24 API change */
601 #ifdef HAVE_BIO_EMPTY_BARRIER
602 BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, size, rc)
604 zio_t *zio = bio->bi_private;
606 zio->io_delay = jiffies_to_msecs(jiffies_64 - zio->io_delay);
608 if (rc && (rc == -EOPNOTSUPP))
609 zio->io_vd->vdev_nowritecache = B_TRUE;
612 ASSERT3S(zio->io_error, >=, 0);
614 vdev_disk_error(zio);
617 BIO_END_IO_RETURN(0);
621 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
623 struct request_queue *q;
626 q = bdev_get_queue(bdev);
630 bio = bio_alloc(GFP_KERNEL, 0);
634 bio->bi_end_io = vdev_disk_io_flush_completion;
635 bio->bi_private = zio;
637 zio->io_delay = jiffies_64;
638 submit_bio(VDEV_WRITE_FLUSH_FUA, bio);
644 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
648 #endif /* HAVE_BIO_EMPTY_BARRIER */
651 vdev_disk_io_start(zio_t *zio)
653 vdev_t *v = zio->io_vd;
654 vdev_disk_t *vd = v->vdev_tsd;
657 switch (zio->io_type) {
660 if (!vdev_readable(v)) {
661 zio->io_error = ENXIO;
662 return ZIO_PIPELINE_CONTINUE;
665 switch (zio->io_cmd) {
666 case DKIOCFLUSHWRITECACHE:
668 if (zfs_nocacheflush)
671 if (v->vdev_nowritecache) {
672 zio->io_error = ENOTSUP;
676 error = vdev_disk_io_flush(vd->vd_bdev, zio);
678 return ZIO_PIPELINE_STOP;
680 zio->io_error = error;
681 if (error == ENOTSUP)
682 v->vdev_nowritecache = B_TRUE;
687 zio->io_error = ENOTSUP;
690 return ZIO_PIPELINE_CONTINUE;
701 zio->io_error = ENOTSUP;
702 return ZIO_PIPELINE_CONTINUE;
705 error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,
706 zio->io_size, zio->io_offset, flags);
708 zio->io_error = error;
709 return ZIO_PIPELINE_CONTINUE;
712 return ZIO_PIPELINE_STOP;
716 vdev_disk_io_done(zio_t *zio)
719 * If the device returned EIO, we revalidate the media. If it is
720 * determined the media has changed this triggers the asynchronous
721 * removal of the device from the configuration.
723 if (zio->io_error == EIO) {
724 vdev_t *v = zio->io_vd;
725 vdev_disk_t *vd = v->vdev_tsd;
727 if (check_disk_change(vd->vd_bdev)) {
728 vdev_bdev_invalidate(vd->vd_bdev);
729 v->vdev_remove_wanted = B_TRUE;
730 spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
736 vdev_disk_hold(vdev_t *vd)
738 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
740 /* We must have a pathname, and it must be absolute. */
741 if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
745 * Only prefetch path and devid info if the device has
748 if (vd->vdev_tsd != NULL)
751 /* XXX: Implement me as a vnode lookup for the device */
752 vd->vdev_name_vp = NULL;
753 vd->vdev_devid_vp = NULL;
757 vdev_disk_rele(vdev_t *vd)
759 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
761 /* XXX: Implement me as a vnode rele for the device */
764 vdev_ops_t vdev_disk_ops = {
773 VDEV_TYPE_DISK, /* name of this vdev type */
774 B_TRUE /* leaf vdev */
778 * Given the root disk device devid or pathname, read the label from
779 * the device, and construct a configuration nvlist.
782 vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
784 struct block_device *bdev;
789 bdev = vdev_bdev_open(devpath, vdev_bdev_mode(FREAD), NULL);
791 return -PTR_ERR(bdev);
793 s = bdev_capacity(bdev);
795 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
799 size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t);
800 label = vmem_alloc(sizeof(vdev_label_t), KM_PUSHPAGE);
802 for (i = 0; i < VDEV_LABELS; i++) {
803 uint64_t offset, state, txg = 0;
805 /* read vdev label */
806 offset = vdev_label_offset(size, i, 0);
807 if (vdev_disk_physio(bdev, (caddr_t)label,
808 VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, READ_SYNC) != 0)
811 if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist,
812 sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) {
817 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
818 &state) != 0 || state >= POOL_STATE_DESTROYED) {
819 nvlist_free(*config);
824 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
825 &txg) != 0 || txg == 0) {
826 nvlist_free(*config);
834 vmem_free(label, sizeof(vdev_label_t));
835 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
840 module_param(zfs_vdev_scheduler, charp, 0644);
841 MODULE_PARM_DESC(zfs_vdev_scheduler, "I/O scheduler");