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, char *device)
124 char sh_path[] = "/bin/sh";
126 char *argv[] = { sh_path, "-c", sh_cmd };
127 char *envp[] = { NULL };
130 if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4))
133 sprintf(sh_cmd, "%s \"%s\" >/sys/block/%s/queue/scheduler",
134 "/bin/echo", elevator, device);
136 error = call_usermodehelper(sh_path, argv, envp, 1);
138 printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
139 elevator, v->vdev_path, device, error);
145 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
147 struct block_device *bdev;
149 int mode, block_size;
151 /* Must have a pathname and it must be absolute. */
152 if (v->vdev_path == NULL || v->vdev_path[0] != '/') {
153 v->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
157 vd = kmem_zalloc(sizeof(vdev_disk_t), KM_SLEEP);
162 * Devices are always opened by the path provided at configuration
163 * time. This means that if the provided path is a udev by-id path
164 * then drives may be recabled without an issue. If the provided
165 * path is a udev by-path path then the physical location information
166 * will be preserved. This can be critical for more complicated
167 * configurations where drives are located in specific physical
168 * locations to maximize the systems tolerence to component failure.
169 * Alternately you can provide your own udev rule to flexibly map
170 * the drives as you see fit. It is not advised that you use the
171 * /dev/[hd]d devices which may be reorder due to probing order.
172 * Devices in the wrong locations will be detected by the higher
173 * level vdev validation.
175 mode = spa_mode(v->vdev_spa);
176 bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd);
178 kmem_free(vd, sizeof(vdev_disk_t));
179 return -PTR_ERR(bdev);
184 block_size = vdev_bdev_block_size(bdev);
186 /* We think the wholedisk property should always be set when this
187 * function is called. ASSERT here so if any legitimate cases exist
188 * where it's not set, we'll find them during debugging. If we never
189 * hit the ASSERT, this and the following conditional statement can be
191 ASSERT3S(v->vdev_wholedisk, !=, -1ULL);
193 /* The wholedisk property was initialized to -1 in vdev_alloc() if it
194 * was unspecified. In that case, check if this is a whole device.
195 * When bdev->bd_contains == bdev we have a whole device and not simply
197 if (v->vdev_wholedisk == -1ULL)
198 v->vdev_wholedisk = (bdev->bd_contains == bdev);
200 /* Clear the nowritecache bit, causes vdev_reopen() to try again. */
201 v->vdev_nowritecache = B_FALSE;
203 /* Physical volume size in bytes */
204 *psize = bdev_capacity(bdev) * block_size;
206 /* Based on the minimum sector size set the block size */
207 *ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
209 /* Try to set the io scheduler elevator algorithm */
210 (void) vdev_elevator_switch(v, zfs_vdev_scheduler,
211 bdev->bd_disk->disk_name);
217 vdev_disk_close(vdev_t *v)
219 vdev_disk_t *vd = v->vdev_tsd;
224 if (vd->vd_bdev != NULL)
225 vdev_bdev_close(vd->vd_bdev,
226 vdev_bdev_mode(spa_mode(v->vdev_spa)));
228 kmem_free(vd, sizeof(vdev_disk_t));
232 static dio_request_t *
233 vdev_disk_dio_alloc(int bio_count)
238 dr = kmem_zalloc(sizeof(dio_request_t) +
239 sizeof(struct bio *) * bio_count, KM_SLEEP);
241 init_completion(&dr->dr_comp);
242 atomic_set(&dr->dr_ref, 0);
243 dr->dr_bio_count = bio_count;
246 for (i = 0; i < dr->dr_bio_count; i++)
247 dr->dr_bio[i] = NULL;
254 vdev_disk_dio_free(dio_request_t *dr)
258 for (i = 0; i < dr->dr_bio_count; i++)
260 bio_put(dr->dr_bio[i]);
262 kmem_free(dr, sizeof(dio_request_t) +
263 sizeof(struct bio *) * dr->dr_bio_count);
267 vdev_disk_dio_is_sync(dio_request_t *dr)
269 #ifdef HAVE_BIO_RW_SYNC
270 /* BIO_RW_SYNC preferred interface from 2.6.12-2.6.29 */
271 return (dr->dr_rw & (1 << BIO_RW_SYNC));
273 # ifdef HAVE_BIO_RW_SYNCIO
274 /* BIO_RW_SYNCIO preferred interface from 2.6.30-2.6.35 */
275 return (dr->dr_rw & (1 << BIO_RW_SYNCIO));
277 # ifdef HAVE_REQ_SYNC
278 /* REQ_SYNC preferred interface from 2.6.36-2.6.xx */
279 return (dr->dr_rw & REQ_SYNC);
281 # error "Unable to determine bio sync flag"
282 # endif /* HAVE_REQ_SYNC */
283 # endif /* HAVE_BIO_RW_SYNC */
284 #endif /* HAVE_BIO_RW_SYNCIO */
288 vdev_disk_dio_get(dio_request_t *dr)
290 atomic_inc(&dr->dr_ref);
294 vdev_disk_dio_put(dio_request_t *dr)
296 int rc = atomic_dec_return(&dr->dr_ref);
299 * Free the dio_request when the last reference is dropped and
300 * ensure zio_interpret is called only once with the correct zio
303 zio_t *zio = dr->dr_zio;
304 int error = dr->dr_error;
306 vdev_disk_dio_free(dr);
309 zio->io_delay = jiffies_to_msecs(
310 jiffies_64 - zio->io_delay);
311 zio->io_error = error;
312 ASSERT3S(zio->io_error, >=, 0);
314 vdev_disk_error(zio);
322 BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, size, error)
324 dio_request_t *dr = bio->bi_private;
327 /* Fatal error but print some useful debugging before asserting */
329 PANIC("dr == NULL, bio->bi_private == NULL\n"
330 "bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d\n"
331 "bi_idx: %d, bi_size: %d, bi_end_io: %p, bi_cnt: %d\n",
332 bio->bi_next, bio->bi_flags, bio->bi_rw, bio->bi_vcnt,
333 bio->bi_idx, bio->bi_size, bio->bi_end_io,
334 atomic_read(&bio->bi_cnt));
336 #ifndef HAVE_2ARGS_BIO_END_IO_T
339 #endif /* HAVE_2ARGS_BIO_END_IO_T */
341 if (error == 0 && !test_bit(BIO_UPTODATE, &bio->bi_flags))
344 if (dr->dr_error == 0)
345 dr->dr_error = -error;
347 /* Drop reference aquired by __vdev_disk_physio */
348 rc = vdev_disk_dio_put(dr);
350 /* Wake up synchronous waiter this is the last outstanding bio */
351 if ((rc == 1) && vdev_disk_dio_is_sync(dr))
352 complete(&dr->dr_comp);
354 BIO_END_IO_RETURN(0);
357 static inline unsigned long
358 bio_nr_pages(void *bio_ptr, unsigned int bio_size)
360 return ((((unsigned long)bio_ptr + bio_size + PAGE_SIZE - 1) >>
361 PAGE_SHIFT) - ((unsigned long)bio_ptr >> PAGE_SHIFT));
365 bio_map(struct bio *bio, void *bio_ptr, unsigned int bio_size)
367 unsigned int offset, size, i;
370 offset = offset_in_page(bio_ptr);
371 for (i = 0; i < bio->bi_max_vecs; i++) {
372 size = PAGE_SIZE - offset;
380 if (kmem_virt(bio_ptr))
381 page = vmalloc_to_page(bio_ptr);
383 page = virt_to_page(bio_ptr);
385 if (bio_add_page(bio, page, size, offset) != size)
397 __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
398 size_t kbuf_size, uint64_t kbuf_offset, int flags)
403 int bio_size, bio_count = 16;
404 int i = 0, error = 0, block_size;
406 ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);
409 dr = vdev_disk_dio_alloc(bio_count);
413 if (zio && !(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))
414 bio_set_flags_failfast(bdev, &flags);
418 block_size = vdev_bdev_block_size(bdev);
421 * When the IO size exceeds the maximum bio size for the request
422 * queue we are forced to break the IO in multiple bio's and wait
423 * for them all to complete. Ideally, all pool users will set
424 * their volume block size to match the maximum request size and
425 * the common case will be one bio per vdev IO request.
428 bio_offset = kbuf_offset;
429 bio_size = kbuf_size;
430 for (i = 0; i <= dr->dr_bio_count; i++) {
432 /* Finished constructing bio's for given buffer */
437 * By default only 'bio_count' bio's per dio are allowed.
438 * However, if we find ourselves in a situation where more
439 * are needed we allocate a larger dio and warn the user.
441 if (dr->dr_bio_count == i) {
442 vdev_disk_dio_free(dr);
444 printk("WARNING: Resized bio's/dio to %d\n",bio_count);
448 dr->dr_bio[i] = bio_alloc(GFP_NOIO,
449 bio_nr_pages(bio_ptr, bio_size));
450 if (dr->dr_bio[i] == NULL) {
451 vdev_disk_dio_free(dr);
455 /* Matching put called by vdev_disk_physio_completion */
456 vdev_disk_dio_get(dr);
458 dr->dr_bio[i]->bi_bdev = bdev;
459 dr->dr_bio[i]->bi_sector = bio_offset / block_size;
460 dr->dr_bio[i]->bi_rw = dr->dr_rw;
461 dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion;
462 dr->dr_bio[i]->bi_private = dr;
464 /* Remaining size is returned to become the new size */
465 bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size);
467 /* Advance in buffer and construct another bio if needed */
468 bio_ptr += dr->dr_bio[i]->bi_size;
469 bio_offset += dr->dr_bio[i]->bi_size;
472 /* Extra reference to protect dio_request during submit_bio */
473 vdev_disk_dio_get(dr);
475 zio->io_delay = jiffies_64;
477 /* Submit all bio's associated with this dio */
478 for (i = 0; i < dr->dr_bio_count; i++)
480 submit_bio(dr->dr_rw, dr->dr_bio[i]);
483 * On synchronous blocking requests we wait for all bio the completion
484 * callbacks to run. We will be woken when the last callback runs
485 * for this dio. We are responsible for putting the last dio_request
486 * reference will in turn put back the last bio references. The
487 * only synchronous consumer is vdev_disk_read_rootlabel() all other
488 * IO originating from vdev_disk_io_start() is asynchronous.
490 if (vdev_disk_dio_is_sync(dr)) {
491 wait_for_completion(&dr->dr_comp);
492 error = dr->dr_error;
493 ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
496 (void)vdev_disk_dio_put(dr);
502 vdev_disk_physio(struct block_device *bdev, caddr_t kbuf,
503 size_t size, uint64_t offset, int flags)
505 bio_set_flags_failfast(bdev, &flags);
506 return __vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags);
509 /* 2.6.24 API change */
510 #ifdef HAVE_BIO_EMPTY_BARRIER
511 BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, size, rc)
513 zio_t *zio = bio->bi_private;
515 zio->io_delay = jiffies_to_msecs(jiffies_64 - zio->io_delay);
517 if (rc && (rc == -EOPNOTSUPP))
518 zio->io_vd->vdev_nowritecache = B_TRUE;
521 ASSERT3S(zio->io_error, >=, 0);
523 vdev_disk_error(zio);
526 BIO_END_IO_RETURN(0);
530 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
532 struct request_queue *q;
535 q = bdev_get_queue(bdev);
539 bio = bio_alloc(GFP_KERNEL, 0);
543 bio->bi_end_io = vdev_disk_io_flush_completion;
544 bio->bi_private = zio;
546 zio->io_delay = jiffies_64;
547 submit_bio(WRITE_BARRIER, bio);
553 vdev_disk_io_flush(struct block_device *bdev, zio_t *zio)
557 #endif /* HAVE_BIO_EMPTY_BARRIER */
560 vdev_disk_io_start(zio_t *zio)
562 vdev_t *v = zio->io_vd;
563 vdev_disk_t *vd = v->vdev_tsd;
566 switch (zio->io_type) {
569 if (!vdev_readable(v)) {
570 zio->io_error = ENXIO;
571 return ZIO_PIPELINE_CONTINUE;
574 switch (zio->io_cmd) {
575 case DKIOCFLUSHWRITECACHE:
577 if (zfs_nocacheflush)
580 if (v->vdev_nowritecache) {
581 zio->io_error = ENOTSUP;
585 error = vdev_disk_io_flush(vd->vd_bdev, zio);
587 return ZIO_PIPELINE_STOP;
589 zio->io_error = error;
590 if (error == ENOTSUP)
591 v->vdev_nowritecache = B_TRUE;
596 zio->io_error = ENOTSUP;
599 return ZIO_PIPELINE_CONTINUE;
610 zio->io_error = ENOTSUP;
611 return ZIO_PIPELINE_CONTINUE;
614 error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,
615 zio->io_size, zio->io_offset, flags);
617 zio->io_error = error;
618 return ZIO_PIPELINE_CONTINUE;
621 return ZIO_PIPELINE_STOP;
625 vdev_disk_io_done(zio_t *zio)
628 * If the device returned EIO, we revalidate the media. If it is
629 * determined the media has changed this triggers the asynchronous
630 * removal of the device from the configuration.
632 if (zio->io_error == EIO) {
633 vdev_t *v = zio->io_vd;
634 vdev_disk_t *vd = v->vdev_tsd;
636 if (check_disk_change(vd->vd_bdev)) {
637 vdev_bdev_invalidate(vd->vd_bdev);
638 v->vdev_remove_wanted = B_TRUE;
639 spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
645 vdev_disk_hold(vdev_t *vd)
647 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
649 /* We must have a pathname, and it must be absolute. */
650 if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
654 * Only prefetch path and devid info if the device has
657 if (vd->vdev_tsd != NULL)
660 /* XXX: Implement me as a vnode lookup for the device */
661 vd->vdev_name_vp = NULL;
662 vd->vdev_devid_vp = NULL;
666 vdev_disk_rele(vdev_t *vd)
668 ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
670 /* XXX: Implement me as a vnode rele for the device */
673 vdev_ops_t vdev_disk_ops = {
682 VDEV_TYPE_DISK, /* name of this vdev type */
683 B_TRUE /* leaf vdev */
687 * Given the root disk device devid or pathname, read the label from
688 * the device, and construct a configuration nvlist.
691 vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
693 struct block_device *bdev;
698 bdev = vdev_bdev_open(devpath, vdev_bdev_mode(FREAD), NULL);
700 return -PTR_ERR(bdev);
702 s = bdev_capacity(bdev) * vdev_bdev_block_size(bdev);
704 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
708 size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t);
709 label = vmem_alloc(sizeof(vdev_label_t), KM_SLEEP);
711 for (i = 0; i < VDEV_LABELS; i++) {
712 uint64_t offset, state, txg = 0;
714 /* read vdev label */
715 offset = vdev_label_offset(size, i, 0);
716 if (vdev_disk_physio(bdev, (caddr_t)label,
717 VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, READ_SYNC) != 0)
720 if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist,
721 sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) {
726 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
727 &state) != 0 || state >= POOL_STATE_DESTROYED) {
728 nvlist_free(*config);
733 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
734 &txg) != 0 || txg == 0) {
735 nvlist_free(*config);
743 vmem_free(label, sizeof(vdev_label_t));
744 vdev_bdev_close(bdev, vdev_bdev_mode(FREAD));
749 module_param(zfs_vdev_scheduler, charp, 0644);
750 MODULE_PARM_DESC(zfs_vdev_scheduler, "IO Scheduler (noop)");