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;
92 /* Otherwise assume the full device capacity */
93 return get_capacity(bdev->bd_disk);
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.
115 * Unfortunately we cannot directly call the elevator_switch() function
116 * because it is not exported from the block layer. This means we have
117 * to use the sysfs interface and a user space upcall. Pools will be
118 * automatically imported on module load so we must do this at device
119 * open time from the kernel.
122 vdev_elevator_switch(vdev_t *v, char *elevator)
124 vdev_disk_t *vd = v->vdev_tsd;
125 struct block_device *bdev = vd->vd_bdev;
126 struct request_queue *q = bdev_get_queue(bdev);
127 char *device = bdev->bd_disk->disk_name;
128 char sh_path[] = "/bin/sh";
130 char *argv[] = { sh_path, "-c", sh_cmd };
131 char *envp[] = { NULL };
132 int count = 0, error;
134 /* Skip devices which are not whole disks (partitions) */
135 if (!v->vdev_wholedisk)
138 /* Skip devices without schedulers (loop, ram, dm, etc) */
139 if (!q->elevator || !blk_queue_stackable(q))
142 /* Leave existing scheduler when set to "none" */
143 if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4))
147 * Set the desired scheduler with a three attempt retry for
148 * -EFAULT which has been observed to occur spuriously.
150 sprintf(sh_cmd, "%s \"%s\" >/sys/block/%s/queue/scheduler",
151 "/bin/echo", elevator, device);
153 while (++count <= 3) {
154 error = call_usermodehelper(sh_path, argv, envp, 1);
155 if ((error == 0) || (error != -EFAULT))
160 printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
161 elevator, v->vdev_path, device, error);
167 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
169 struct block_device *bdev;
171 int mode, block_size;
173 /* Must have a pathname and it must be absolute. */
174 if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
175 v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
179 vd = kmem_zalloc(sizeof(vdev_disk_t), KM_SLEEP);
184 * Devices are always opened by the path provided at configuration
185 * time. This means that if the provided path is a udev by-id path
186 * then drives may be recabled without an issue. If the provided
187 * path is a udev by-path path then the physical location information
188 * will be preserved. This can be critical for more complicated
189 * configurations where drives are located in specific physical
190 * locations to maximize the systems tolerence to component failure.
191 * Alternately you can provide your own udev rule to flexibly map
192 * the drives as you see fit. It is not advised that you use the
193 * /dev/[hd]d devices which may be reorder due to probing order.
194 * Devices in the wrong locations will be detected by the higher
195 * level vdev validation.
197 mode = spa_mode(v->vdev_spa);
198 bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd);
200 kmem_free(vd, sizeof(vdev_disk_t));
201 return -PTR_ERR(bdev);
206 block_size = vdev_bdev_block_size(bdev);
208 /* We think the wholedisk property should always be set when this
209 * function is called. ASSERT here so if any legitimate cases exist
210 * where it's not set, we'll find them during debugging. If we never
211 * hit the ASSERT, this and the following conditional statement can be
213 ASSERT3S(v->vdev_wholedisk, !=, -1ULL);
215 /* The wholedisk property was initialized to -1 in vdev_alloc() if it
216 * was unspecified. In that case, check if this is a whole device.
217 * When bdev->bd_contains == bdev we have a whole device and not simply
219 if (v->vdev_wholedisk == -1ULL)
220 v->vdev_wholedisk = (bdev->bd_contains == bdev);
222 /* Clear the nowritecache bit, causes vdev_reopen() to try again. */
223 v->vdev_nowritecache = B_FALSE;
225 /* Physical volume size in bytes */
226 *psize = bdev_capacity(bdev) * block_size;
228 /* Based on the minimum sector size set the block size */
229 *ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
231 /* Try to set the io scheduler elevator algorithm */
232 (void) vdev_elevator_switch(v, zfs_vdev_scheduler);
238 vdev_disk_close(vdev_t *v)
240 vdev_disk_t *vd = v->vdev_tsd;
245 if (vd->vd_bdev != NULL)
246 vdev_bdev_close(vd->vd_bdev,
247 vdev_bdev_mode(spa_mode(v->vdev_spa)));
249 kmem_free(vd, sizeof(vdev_disk_t));
253 static dio_request_t *
254 vdev_disk_dio_alloc(int bio_count)
259 dr = kmem_zalloc(sizeof(dio_request_t) +
260 sizeof(struct bio *) * bio_count, KM_SLEEP);
262 init_completion(&dr->dr_comp);
263 atomic_set(&dr->dr_ref, 0);
264 dr->dr_bio_count = bio_count;
267 for (i = 0; i < dr->dr_bio_count; i++)
268 dr->dr_bio[i] = NULL;
275 vdev_disk_dio_free(dio_request_t *dr)
279 for (i = 0; i < dr->dr_bio_count; i++)
281 bio_put(dr->dr_bio[i]);
283 kmem_free(dr, sizeof(dio_request_t) +
284 sizeof(struct bio *) * dr->dr_bio_count);
288 vdev_disk_dio_is_sync(dio_request_t *dr)
290 #ifdef HAVE_BIO_RW_SYNC
291 /* BIO_RW_SYNC preferred interface from 2.6.12-2.6.29 */
292 return (dr->dr_rw & (1 << BIO_RW_SYNC));
294 # ifdef HAVE_BIO_RW_SYNCIO
295 /* BIO_RW_SYNCIO preferred interface from 2.6.30-2.6.35 */
296 return (dr->dr_rw & (1 << BIO_RW_SYNCIO));
298 # ifdef HAVE_REQ_SYNC
299 /* REQ_SYNC preferred interface from 2.6.36-2.6.xx */
300 return (dr->dr_rw & REQ_SYNC);
302 # error "Unable to determine bio sync flag"
303 # endif /* HAVE_REQ_SYNC */
304 # endif /* HAVE_BIO_RW_SYNC */
305 #endif /* HAVE_BIO_RW_SYNCIO */
309 vdev_disk_dio_get(dio_request_t *dr)
311 atomic_inc(&dr->dr_ref);
315 vdev_disk_dio_put(dio_request_t *dr)
317 int rc = atomic_dec_return(&dr->dr_ref);
320 * Free the dio_request when the last reference is dropped and
321 * ensure zio_interpret is called only once with the correct zio
324 zio_t *zio = dr->dr_zio;
325 int error = dr->dr_error;
327 vdev_disk_dio_free(dr);
330 zio->io_delay = jiffies_to_msecs(
331 jiffies_64 - zio->io_delay);
332 zio->io_error = error;
333 ASSERT3S(zio->io_error, >=, 0);
335 vdev_disk_error(zio);
343 BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, size, error)
345 dio_request_t *dr = bio->bi_private;
348 /* Fatal error but print some useful debugging before asserting */
350 PANIC("dr == NULL, bio->bi_private == NULL\n"
351 "bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d\n"
352 "bi_idx: %d, bi_size: %d, bi_end_io: %p, bi_cnt: %d\n",
353 bio->bi_next, bio->bi_flags, bio->bi_rw, bio->bi_vcnt,
354 bio->bi_idx, bio->bi_size, bio->bi_end_io,
355 atomic_read(&bio->bi_cnt));
357 #ifndef HAVE_2ARGS_BIO_END_IO_T
360 #endif /* HAVE_2ARGS_BIO_END_IO_T */
362 if (error == 0 && !test_bit(BIO_UPTODATE, &bio->bi_flags))
365 if (dr->dr_error == 0)
366 dr->dr_error = -error;
368 /* Drop reference aquired by __vdev_disk_physio */
369 rc = vdev_disk_dio_put(dr);
371 /* Wake up synchronous waiter this is the last outstanding bio */
372 if ((rc == 1) && vdev_disk_dio_is_sync(dr))
373 complete(&dr->dr_comp);
375 BIO_END_IO_RETURN(0);
378 static inline unsigned long
379 bio_nr_pages(void *bio_ptr, unsigned int bio_size)
381 return ((((unsigned long)bio_ptr + bio_size + PAGE_SIZE - 1) >>
382 PAGE_SHIFT) - ((unsigned long)bio_ptr >> PAGE_SHIFT));
386 bio_map(struct bio *bio, void *bio_ptr, unsigned int bio_size)
388 unsigned int offset, size, i;
391 offset = offset_in_page(bio_ptr);
392 for (i = 0; i < bio->bi_max_vecs; i++) {
393 size = PAGE_SIZE - offset;
401 if (kmem_virt(bio_ptr))
402 page = vmalloc_to_page(bio_ptr);
404 page = virt_to_page(bio_ptr);
406 if (bio_add_page(bio, page, size, offset) != size)
418 __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
419 size_t kbuf_size, uint64_t kbuf_offset, int flags)
424 int bio_size, bio_count = 16;
425 int i = 0, error = 0, block_size;
427 ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);
430 dr = vdev_disk_dio_alloc(bio_count);
434 if (zio && !(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))
435 bio_set_flags_failfast(bdev, &flags);
439 block_size = vdev_bdev_block_size(bdev);
442 * When the IO size exceeds the maximum bio size for the request
443 * queue we are forced to break the IO in multiple bio's and wait
444 * for them all to complete. Ideally, all pool users will set
445 * their volume block size to match the maximum request size and
446 * the common case will be one bio per vdev IO request.
449 bio_offset = kbuf_offset;
450 bio_size = kbuf_size;
451 for (i = 0; i <= dr->dr_bio_count; i++) {
453 /* Finished constructing bio's for given buffer */
458 * By default only 'bio_count' bio's per dio are allowed.
459 * However, if we find ourselves in a situation where more
460 * are needed we allocate a larger dio and warn the user.
462 if (dr->dr_bio_count == i) {
463 vdev_disk_dio_free(dr);
465 printk("WARNING: Resized bio's/dio to %d\n",bio_count);
469 dr->dr_bio[i] = bio_alloc(GFP_NOIO,
470 bio_nr_pages(bio_ptr, bio_size));
471 if (dr->dr_bio[i] == NULL) {
472 vdev_disk_dio_free(dr);
476 /* Matching put called by vdev_disk_physio_completion */
477 vdev_disk_dio_get(dr);
479 dr->dr_bio[i]->bi_bdev = bdev;
480 dr->dr_bio[i]->bi_sector = bio_offset / block_size;
481 dr->dr_bio[i]->bi_rw = dr->dr_rw;
482 dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion;
483 dr->dr_bio[i]->bi_private = dr;
485 /* Remaining size is returned to become the new size */
486 bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size);
488 /* Advance in buffer and construct another bio if needed */
489 bio_ptr += dr->dr_bio[i]->bi_size;
490 bio_offset += dr->dr_bio[i]->bi_size;
493 /* Extra reference to protect dio_request during submit_bio */
494 vdev_disk_dio_get(dr);
496 zio->io_delay = jiffies_64;
498 /* Submit all bio's associated with this dio */
499 for (i = 0; i < dr->dr_bio_count; i++)
501 submit_bio(dr->dr_rw, dr->dr_bio[i]);
504 * On synchronous blocking requests we wait for all bio the completion
505 * callbacks to run. We will be woken when the last callback runs
506 * for this dio. We are responsible for putting the last dio_request
507 * reference will in turn put back the last bio references. The
508 * only synchronous consumer is vdev_disk_read_rootlabel() all other
509 * IO originating from vdev_disk_io_start() is asynchronous.
511 if (vdev_disk_dio_is_sync(dr)) {
512 wait_for_completion(&dr->dr_comp);
513 error = dr->dr_error;
514 ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
517 (void)vdev_disk_dio_put(dr);
523 vdev_disk_physio(struct block_device *bdev, caddr_t kbuf,
524 size_t size, uint64_t offset, int flags)
526 bio_set_flags_failfast(bdev, &flags);
527 return __vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags);
530 /* 2.6.24 API change */
531 #ifdef HAVE_BIO_EMPTY_BARRIER
532 BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, size, rc)
534 zio_t *zio = bio->bi_private;
536 zio->io_delay = jiffies_to_msecs(jiffies_64 - zio->io_delay);
538 if (rc && (rc == -EOPNOTSUPP))
539 zio->io_vd->vdev_nowritecache = B_TRUE;
542 ASSERT3S(zio->io_error, >=, 0);
544 vdev_disk_error(zio);
547 BIO_END_IO_RETURN(0);
551 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
553 struct request_queue *q;
556 q = bdev_get_queue(bdev);
560 bio = bio_alloc(GFP_KERNEL, 0);
564 bio->bi_end_io = vdev_disk_io_flush_completion;
565 bio->bi_private = zio;
567 zio->io_delay = jiffies_64;
568 submit_bio(WRITE_BARRIER, bio);
574 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
578 #endif /* HAVE_BIO_EMPTY_BARRIER */
581 vdev_disk_io_start(zio_t *zio)
583 vdev_t *v = zio->io_vd;
584 vdev_disk_t *vd = v->vdev_tsd;
587 switch (zio->io_type) {
590 if (!vdev_readable(v)) {
591 zio->io_error = ENXIO;
592 return ZIO_PIPELINE_CONTINUE;
595 switch (zio->io_cmd) {
596 case DKIOCFLUSHWRITECACHE:
598 if (zfs_nocacheflush)
601 if (v->vdev_nowritecache) {
602 zio->io_error = ENOTSUP;
606 error = vdev_disk_io_flush(vd->vd_bdev, zio);
608 return ZIO_PIPELINE_STOP;
610 zio->io_error = error;
611 if (error == ENOTSUP)
612 v->vdev_nowritecache = B_TRUE;
617 zio->io_error = ENOTSUP;
620 return ZIO_PIPELINE_CONTINUE;
631 zio->io_error = ENOTSUP;
632 return ZIO_PIPELINE_CONTINUE;
635 error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,
636 zio->io_size, zio->io_offset, flags);
638 zio->io_error = error;
639 return ZIO_PIPELINE_CONTINUE;
642 return ZIO_PIPELINE_STOP;
646 vdev_disk_io_done(zio_t *zio)
649 * If the device returned EIO, we revalidate the media. If it is
650 * determined the media has changed this triggers the asynchronous
651 * removal of the device from the configuration.
653 if (zio->io_error == EIO) {
654 vdev_t *v = zio->io_vd;
655 vdev_disk_t *vd = v->vdev_tsd;
657 if (check_disk_change(vd->vd_bdev)) {
658 vdev_bdev_invalidate(vd->vd_bdev);
659 v->vdev_remove_wanted = B_TRUE;
660 spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
666 vdev_disk_hold(vdev_t *vd)
668 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
670 /* We must have a pathname, and it must be absolute. */
671 if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
675 * Only prefetch path and devid info if the device has
678 if (vd->vdev_tsd != NULL)
681 /* XXX: Implement me as a vnode lookup for the device */
682 vd->vdev_name_vp = NULL;
683 vd->vdev_devid_vp = NULL;
687 vdev_disk_rele(vdev_t *vd)
689 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
691 /* XXX: Implement me as a vnode rele for the device */
694 vdev_ops_t vdev_disk_ops = {
703 VDEV_TYPE_DISK, /* name of this vdev type */
704 B_TRUE /* leaf vdev */
708 * Given the root disk device devid or pathname, read the label from
709 * the device, and construct a configuration nvlist.
712 vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
714 struct block_device *bdev;
719 bdev = vdev_bdev_open(devpath, vdev_bdev_mode(FREAD), NULL);
721 return -PTR_ERR(bdev);
723 s = bdev_capacity(bdev) * vdev_bdev_block_size(bdev);
725 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
729 size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t);
730 label = vmem_alloc(sizeof(vdev_label_t), KM_SLEEP);
732 for (i = 0; i < VDEV_LABELS; i++) {
733 uint64_t offset, state, txg = 0;
735 /* read vdev label */
736 offset = vdev_label_offset(size, i, 0);
737 if (vdev_disk_physio(bdev, (caddr_t)label,
738 VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, READ_SYNC) != 0)
741 if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist,
742 sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) {
747 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
748 &state) != 0 || state >= POOL_STATE_DESTROYED) {
749 nvlist_free(*config);
754 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
755 &txg) != 0 || txg == 0) {
756 nvlist_free(*config);
764 vmem_free(label, sizeof(vdev_label_t));
765 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
770 module_param(zfs_vdev_scheduler, charp, 0644);
771 MODULE_PARM_DESC(zfs_vdev_scheduler, "IO Scheduler (noop)");