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;
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) && (strlen(elevator) == 4))
143 #ifdef HAVE_ELEVATOR_CHANGE
144 error = elevator_change(q, elevator);
146 /* For pre-2.6.36 kernels elevator_change() is not available.
147 * Therefore we fall back to using a usermodehelper to echo the
148 * elevator into sysfs; This requires /bin/echo and sysfs to be
149 * mounted which may not be true early in the boot process.
151 # define SET_SCHEDULER_CMD \
152 "exec 0</dev/null " \
153 " 1>/sys/block/%s/queue/scheduler " \
158 char *argv[] = { "/bin/sh", "-c", NULL, NULL };
159 char *envp[] = { NULL };
161 argv[2] = kmem_asprintf(SET_SCHEDULER_CMD, device, elevator);
162 error = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
165 #endif /* HAVE_ELEVATOR_CHANGE */
167 printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
168 elevator, v->vdev_path, device, error);
174 * Expanding a whole disk vdev involves invoking BLKRRPART on the
175 * whole disk device. This poses a problem, because BLKRRPART will
176 * return EBUSY if one of the disk's partitions is open. That's why
177 * we have to do it here, just before opening the data partition.
178 * Unfortunately, BLKRRPART works by dropping all partitions and
179 * recreating them, which means that for a short time window, all
180 * /dev/sdxN device files disappear (until udev recreates them).
181 * This means two things:
182 * - When we open the data partition just after a BLKRRPART, we
183 * can't do it using the normal device file path because of the
184 * obvious race condition with udev. Instead, we use reliable
185 * kernel APIs to get a handle to the new partition device from
186 * the whole disk device.
187 * - Because vdev_disk_open() initially needs to find the device
188 * using its path, multiple vdev_disk_open() invocations in
189 * short succession on the same disk with BLKRRPARTs in the
190 * middle have a high probability of failure (because of the
191 * race condition with udev). A typical situation where this
192 * might happen is when the zpool userspace tool does a
193 * TRYIMPORT immediately followed by an IMPORT. For this
194 * reason, we only invoke BLKRRPART in the module when strictly
195 * necessary (zpool online -e case), and rely on userspace to
196 * do it when possible.
198 static struct block_device *
199 vdev_disk_rrpart(const char *path, int mode, vdev_disk_t *vd)
201 #if defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK)
202 struct block_device *bdev, *result = ERR_PTR(-ENXIO);
203 struct gendisk *disk;
206 bdev = vdev_bdev_open(path, vdev_bdev_mode(mode), vd);
210 disk = get_gendisk(bdev->bd_dev, &partno);
211 vdev_bdev_close(bdev, vdev_bdev_mode(mode));
214 bdev = bdget(disk_devt(disk));
216 error = blkdev_get(bdev, vdev_bdev_mode(mode), vd);
218 error = ioctl_by_bdev(bdev, BLKRRPART, 0);
219 vdev_bdev_close(bdev, vdev_bdev_mode(mode));
222 bdev = bdget_disk(disk, partno);
224 error = blkdev_get(bdev,
225 vdev_bdev_mode(mode) | FMODE_EXCL, vd);
234 return ERR_PTR(-EOPNOTSUPP);
235 #endif /* defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) */
239 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
242 struct block_device *bdev = ERR_PTR(-ENXIO);
244 int mode, block_size;
246 /* Must have a pathname and it must be absolute. */
247 if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
248 v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
252 vd = kmem_zalloc(sizeof(vdev_disk_t), KM_PUSHPAGE);
257 * Devices are always opened by the path provided at configuration
258 * time. This means that if the provided path is a udev by-id path
259 * then drives may be recabled without an issue. If the provided
260 * path is a udev by-path path then the physical location information
261 * will be preserved. This can be critical for more complicated
262 * configurations where drives are located in specific physical
263 * locations to maximize the systems tolerence to component failure.
264 * Alternately you can provide your own udev rule to flexibly map
265 * the drives as you see fit. It is not advised that you use the
266 * /dev/[hd]d devices which may be reorder due to probing order.
267 * Devices in the wrong locations will be detected by the higher
268 * level vdev validation.
270 mode = spa_mode(v->vdev_spa);
271 if (v->vdev_wholedisk && v->vdev_expanding)
272 bdev = vdev_disk_rrpart(v->vdev_path, mode, vd);
274 bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd);
276 kmem_free(vd, sizeof(vdev_disk_t));
277 return -PTR_ERR(bdev);
282 block_size = vdev_bdev_block_size(bdev);
284 /* We think the wholedisk property should always be set when this
285 * function is called. ASSERT here so if any legitimate cases exist
286 * where it's not set, we'll find them during debugging. If we never
287 * hit the ASSERT, this and the following conditional statement can be
289 ASSERT3S(v->vdev_wholedisk, !=, -1ULL);
291 /* The wholedisk property was initialized to -1 in vdev_alloc() if it
292 * was unspecified. In that case, check if this is a whole device.
293 * When bdev->bd_contains == bdev we have a whole device and not simply
295 if (v->vdev_wholedisk == -1ULL)
296 v->vdev_wholedisk = (bdev->bd_contains == bdev);
298 /* Clear the nowritecache bit, causes vdev_reopen() to try again. */
299 v->vdev_nowritecache = B_FALSE;
301 /* Physical volume size in bytes */
302 *psize = bdev_capacity(bdev);
304 /* TODO: report possible expansion size */
307 /* Based on the minimum sector size set the block size */
308 *ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
310 /* Try to set the io scheduler elevator algorithm */
311 (void) vdev_elevator_switch(v, zfs_vdev_scheduler);
317 vdev_disk_close(vdev_t *v)
319 vdev_disk_t *vd = v->vdev_tsd;
324 if (vd->vd_bdev != NULL)
325 vdev_bdev_close(vd->vd_bdev,
326 vdev_bdev_mode(spa_mode(v->vdev_spa)));
328 kmem_free(vd, sizeof(vdev_disk_t));
332 static dio_request_t *
333 vdev_disk_dio_alloc(int bio_count)
338 dr = kmem_zalloc(sizeof(dio_request_t) +
339 sizeof(struct bio *) * bio_count, KM_PUSHPAGE);
341 init_completion(&dr->dr_comp);
342 atomic_set(&dr->dr_ref, 0);
343 dr->dr_bio_count = bio_count;
346 for (i = 0; i < dr->dr_bio_count; i++)
347 dr->dr_bio[i] = NULL;
354 vdev_disk_dio_free(dio_request_t *dr)
358 for (i = 0; i < dr->dr_bio_count; i++)
360 bio_put(dr->dr_bio[i]);
362 kmem_free(dr, sizeof(dio_request_t) +
363 sizeof(struct bio *) * dr->dr_bio_count);
367 vdev_disk_dio_is_sync(dio_request_t *dr)
369 #ifdef HAVE_BIO_RW_SYNC
370 /* BIO_RW_SYNC preferred interface from 2.6.12-2.6.29 */
371 return (dr->dr_rw & (1 << BIO_RW_SYNC));
373 # ifdef HAVE_BIO_RW_SYNCIO
374 /* BIO_RW_SYNCIO preferred interface from 2.6.30-2.6.35 */
375 return (dr->dr_rw & (1 << BIO_RW_SYNCIO));
377 # ifdef HAVE_REQ_SYNC
378 /* REQ_SYNC preferred interface from 2.6.36-2.6.xx */
379 return (dr->dr_rw & REQ_SYNC);
381 # error "Unable to determine bio sync flag"
382 # endif /* HAVE_REQ_SYNC */
383 # endif /* HAVE_BIO_RW_SYNC */
384 #endif /* HAVE_BIO_RW_SYNCIO */
388 vdev_disk_dio_get(dio_request_t *dr)
390 atomic_inc(&dr->dr_ref);
394 vdev_disk_dio_put(dio_request_t *dr)
396 int rc = atomic_dec_return(&dr->dr_ref);
399 * Free the dio_request when the last reference is dropped and
400 * ensure zio_interpret is called only once with the correct zio
403 zio_t *zio = dr->dr_zio;
404 int error = dr->dr_error;
406 vdev_disk_dio_free(dr);
409 zio->io_delay = jiffies_to_msecs(
410 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, size, error)
424 dio_request_t *dr = bio->bi_private;
427 /* Fatal error but print some useful debugging before asserting */
429 PANIC("dr == NULL, bio->bi_private == NULL\n"
430 "bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d\n"
431 "bi_idx: %d, bi_size: %d, bi_end_io: %p, bi_cnt: %d\n",
432 bio->bi_next, bio->bi_flags, bio->bi_rw, bio->bi_vcnt,
433 bio->bi_idx, bio->bi_size, bio->bi_end_io,
434 atomic_read(&bio->bi_cnt));
436 #ifndef HAVE_2ARGS_BIO_END_IO_T
439 #endif /* HAVE_2ARGS_BIO_END_IO_T */
441 if (error == 0 && !test_bit(BIO_UPTODATE, &bio->bi_flags))
444 if (dr->dr_error == 0)
445 dr->dr_error = -error;
447 /* Drop reference aquired by __vdev_disk_physio */
448 rc = vdev_disk_dio_put(dr);
450 /* Wake up synchronous waiter this is the last outstanding bio */
451 if ((rc == 1) && vdev_disk_dio_is_sync(dr))
452 complete(&dr->dr_comp);
454 BIO_END_IO_RETURN(0);
457 static inline unsigned long
458 bio_nr_pages(void *bio_ptr, unsigned int bio_size)
460 return ((((unsigned long)bio_ptr + bio_size + PAGE_SIZE - 1) >>
461 PAGE_SHIFT) - ((unsigned long)bio_ptr >> PAGE_SHIFT));
465 bio_map(struct bio *bio, void *bio_ptr, unsigned int bio_size)
467 unsigned int offset, size, i;
470 offset = offset_in_page(bio_ptr);
471 for (i = 0; i < bio->bi_max_vecs; i++) {
472 size = PAGE_SIZE - offset;
480 if (kmem_virt(bio_ptr))
481 page = vmalloc_to_page(bio_ptr);
483 page = virt_to_page(bio_ptr);
485 if (bio_add_page(bio, page, size, offset) != size)
497 __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
498 size_t kbuf_size, uint64_t kbuf_offset, int flags)
503 int bio_size, bio_count = 16;
504 int i = 0, error = 0;
506 ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);
509 dr = vdev_disk_dio_alloc(bio_count);
513 if (zio && !(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))
514 bio_set_flags_failfast(bdev, &flags);
520 * When the IO size exceeds the maximum bio size for the request
521 * queue we are forced to break the IO in multiple bio's and wait
522 * for them all to complete. Ideally, all pool users will set
523 * their volume block size to match the maximum request size and
524 * the common case will be one bio per vdev IO request.
527 bio_offset = kbuf_offset;
528 bio_size = kbuf_size;
529 for (i = 0; i <= dr->dr_bio_count; i++) {
531 /* Finished constructing bio's for given buffer */
536 * By default only 'bio_count' bio's per dio are allowed.
537 * However, if we find ourselves in a situation where more
538 * are needed we allocate a larger dio and warn the user.
540 if (dr->dr_bio_count == i) {
541 vdev_disk_dio_free(dr);
546 dr->dr_bio[i] = bio_alloc(GFP_NOIO,
547 bio_nr_pages(bio_ptr, bio_size));
548 if (dr->dr_bio[i] == NULL) {
549 vdev_disk_dio_free(dr);
553 /* Matching put called by vdev_disk_physio_completion */
554 vdev_disk_dio_get(dr);
556 dr->dr_bio[i]->bi_bdev = bdev;
557 dr->dr_bio[i]->bi_sector = bio_offset >> 9;
558 dr->dr_bio[i]->bi_rw = dr->dr_rw;
559 dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion;
560 dr->dr_bio[i]->bi_private = dr;
562 /* Remaining size is returned to become the new size */
563 bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size);
565 /* Advance in buffer and construct another bio if needed */
566 bio_ptr += dr->dr_bio[i]->bi_size;
567 bio_offset += dr->dr_bio[i]->bi_size;
570 /* Extra reference to protect dio_request during submit_bio */
571 vdev_disk_dio_get(dr);
573 zio->io_delay = jiffies_64;
575 /* Submit all bio's associated with this dio */
576 for (i = 0; i < dr->dr_bio_count; i++)
578 submit_bio(dr->dr_rw, dr->dr_bio[i]);
581 * On synchronous blocking requests we wait for all bio the completion
582 * callbacks to run. We will be woken when the last callback runs
583 * for this dio. We are responsible for putting the last dio_request
584 * reference will in turn put back the last bio references. The
585 * only synchronous consumer is vdev_disk_read_rootlabel() all other
586 * IO originating from vdev_disk_io_start() is asynchronous.
588 if (vdev_disk_dio_is_sync(dr)) {
589 wait_for_completion(&dr->dr_comp);
590 error = dr->dr_error;
591 ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
594 (void)vdev_disk_dio_put(dr);
600 vdev_disk_physio(struct block_device *bdev, caddr_t kbuf,
601 size_t size, uint64_t offset, int flags)
603 bio_set_flags_failfast(bdev, &flags);
604 return __vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags);
607 BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, size, rc)
609 zio_t *zio = bio->bi_private;
611 zio->io_delay = jiffies_to_msecs(jiffies_64 - zio->io_delay);
613 if (rc && (rc == -EOPNOTSUPP))
614 zio->io_vd->vdev_nowritecache = B_TRUE;
617 ASSERT3S(zio->io_error, >=, 0);
619 vdev_disk_error(zio);
622 BIO_END_IO_RETURN(0);
626 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
628 struct request_queue *q;
631 q = bdev_get_queue(bdev);
635 bio = bio_alloc(GFP_KERNEL, 0);
639 bio->bi_end_io = vdev_disk_io_flush_completion;
640 bio->bi_private = zio;
642 zio->io_delay = jiffies_64;
643 submit_bio(VDEV_WRITE_FLUSH_FUA, bio);
649 vdev_disk_io_start(zio_t *zio)
651 vdev_t *v = zio->io_vd;
652 vdev_disk_t *vd = v->vdev_tsd;
655 switch (zio->io_type) {
658 if (!vdev_readable(v)) {
659 zio->io_error = ENXIO;
660 return ZIO_PIPELINE_CONTINUE;
663 switch (zio->io_cmd) {
664 case DKIOCFLUSHWRITECACHE:
666 if (zfs_nocacheflush)
669 if (v->vdev_nowritecache) {
670 zio->io_error = ENOTSUP;
674 error = vdev_disk_io_flush(vd->vd_bdev, zio);
676 return ZIO_PIPELINE_STOP;
678 zio->io_error = error;
679 if (error == ENOTSUP)
680 v->vdev_nowritecache = B_TRUE;
685 zio->io_error = ENOTSUP;
688 return ZIO_PIPELINE_CONTINUE;
699 zio->io_error = ENOTSUP;
700 return ZIO_PIPELINE_CONTINUE;
703 error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,
704 zio->io_size, zio->io_offset, flags);
706 zio->io_error = error;
707 return ZIO_PIPELINE_CONTINUE;
710 return ZIO_PIPELINE_STOP;
714 vdev_disk_io_done(zio_t *zio)
717 * If the device returned EIO, we revalidate the media. If it is
718 * determined the media has changed this triggers the asynchronous
719 * removal of the device from the configuration.
721 if (zio->io_error == EIO) {
722 vdev_t *v = zio->io_vd;
723 vdev_disk_t *vd = v->vdev_tsd;
725 if (check_disk_change(vd->vd_bdev)) {
726 vdev_bdev_invalidate(vd->vd_bdev);
727 v->vdev_remove_wanted = B_TRUE;
728 spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
734 vdev_disk_hold(vdev_t *vd)
736 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
738 /* We must have a pathname, and it must be absolute. */
739 if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
743 * Only prefetch path and devid info if the device has
746 if (vd->vdev_tsd != NULL)
749 /* XXX: Implement me as a vnode lookup for the device */
750 vd->vdev_name_vp = NULL;
751 vd->vdev_devid_vp = NULL;
755 vdev_disk_rele(vdev_t *vd)
757 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
759 /* XXX: Implement me as a vnode rele for the device */
762 vdev_ops_t vdev_disk_ops = {
771 VDEV_TYPE_DISK, /* name of this vdev type */
772 B_TRUE /* leaf vdev */
776 * Given the root disk device devid or pathname, read the label from
777 * the device, and construct a configuration nvlist.
780 vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
782 struct block_device *bdev;
787 bdev = vdev_bdev_open(devpath, vdev_bdev_mode(FREAD), NULL);
789 return -PTR_ERR(bdev);
791 s = bdev_capacity(bdev);
793 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
797 size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t);
798 label = vmem_alloc(sizeof(vdev_label_t), KM_PUSHPAGE);
800 for (i = 0; i < VDEV_LABELS; i++) {
801 uint64_t offset, state, txg = 0;
803 /* read vdev label */
804 offset = vdev_label_offset(size, i, 0);
805 if (vdev_disk_physio(bdev, (caddr_t)label,
806 VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, READ_SYNC) != 0)
809 if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist,
810 sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) {
815 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
816 &state) != 0 || state >= POOL_STATE_DESTROYED) {
817 nvlist_free(*config);
822 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
823 &txg) != 0 || txg == 0) {
824 nvlist_free(*config);
832 vmem_free(label, sizeof(vdev_label_t));
833 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
838 module_param(zfs_vdev_scheduler, charp, 0644);
839 MODULE_PARM_DESC(zfs_vdev_scheduler, "I/O scheduler");