]>
Commit | Line | Data |
---|---|---|
60101509 BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
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. | |
7 | * | |
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. | |
12 | * | |
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] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
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>. | |
25 | * LLNL-CODE-403049. | |
26 | */ | |
27 | ||
28 | #include <sys/zfs_context.h> | |
29 | #include <sys/spa.h> | |
30 | #include <sys/vdev_disk.h> | |
31 | #include <sys/vdev_impl.h> | |
32 | #include <sys/fs/zfs.h> | |
33 | #include <sys/zio.h> | |
34 | #include <sys/sunldi.h> | |
35 | ||
6839eed2 BB |
36 | char *zfs_vdev_scheduler = VDEV_SCHEDULER; |
37 | ||
60101509 BB |
38 | /* |
39 | * Virtual device vector for disks. | |
40 | */ | |
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 */ | |
49 | } dio_request_t; | |
50 | ||
51 | ||
52 | #ifdef HAVE_OPEN_BDEV_EXCLUSIVE | |
53 | static fmode_t | |
54 | vdev_bdev_mode(int smode) | |
55 | { | |
56 | fmode_t mode = 0; | |
57 | ||
58 | ASSERT3S(smode & (FREAD | FWRITE), !=, 0); | |
59 | ||
60 | if (smode & FREAD) | |
61 | mode |= FMODE_READ; | |
62 | ||
63 | if (smode & FWRITE) | |
64 | mode |= FMODE_WRITE; | |
65 | ||
66 | return mode; | |
67 | } | |
68 | #else | |
69 | static int | |
70 | vdev_bdev_mode(int smode) | |
71 | { | |
72 | int mode = 0; | |
73 | ||
74 | ASSERT3S(smode & (FREAD | FWRITE), !=, 0); | |
75 | ||
76 | if ((smode & FREAD) && !(smode & FWRITE)) | |
77 | mode = MS_RDONLY; | |
78 | ||
79 | return mode; | |
80 | } | |
81 | #endif /* HAVE_OPEN_BDEV_EXCLUSIVE */ | |
82 | ||
83 | static uint64_t | |
84 | bdev_capacity(struct block_device *bdev) | |
85 | { | |
86 | struct hd_struct *part = bdev->bd_part; | |
87 | ||
88 | /* The partition capacity referenced by the block device */ | |
89 | if (part) | |
f74fae8b | 90 | return (part->nr_sects << 9); |
60101509 BB |
91 | |
92 | /* Otherwise assume the full device capacity */ | |
f74fae8b | 93 | return (get_capacity(bdev->bd_disk) << 9); |
60101509 BB |
94 | } |
95 | ||
d148e951 BB |
96 | static void |
97 | vdev_disk_error(zio_t *zio) | |
98 | { | |
99 | #ifdef ZFS_DEBUG | |
a69052be BB |
100 | printk("ZFS: zio error=%d type=%d offset=%llu size=%llu " |
101 | "flags=%x delay=%llu\n", zio->io_error, zio->io_type, | |
d148e951 | 102 | (u_longlong_t)zio->io_offset, (u_longlong_t)zio->io_size, |
a69052be | 103 | zio->io_flags, (u_longlong_t)zio->io_delay); |
d148e951 BB |
104 | #endif |
105 | } | |
106 | ||
6839eed2 BB |
107 | /* |
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. | |
6839eed2 BB |
114 | */ |
115 | static int | |
fdcd952b | 116 | vdev_elevator_switch(vdev_t *v, char *elevator) |
6839eed2 | 117 | { |
fdcd952b BB |
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; | |
e2448b0e | 122 | int error; |
fdcd952b | 123 | |
84daadde PS |
124 | /* |
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. | |
131 | */ | |
132 | if (!v->vdev_wholedisk && strncmp(device, "dm-", 3) != 0) | |
04516a45 BB |
133 | return (0); |
134 | ||
fdcd952b BB |
135 | /* Skip devices without schedulers (loop, ram, dm, etc) */ |
136 | if (!q->elevator || !blk_queue_stackable(q)) | |
137 | return (0); | |
6839eed2 | 138 | |
fdcd952b | 139 | /* Leave existing scheduler when set to "none" */ |
6839eed2 BB |
140 | if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4)) |
141 | return (0); | |
142 | ||
6d1d976b BB |
143 | #ifdef HAVE_ELEVATOR_CHANGE |
144 | error = elevator_change(q, elevator); | |
145 | #else | |
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. | |
150 | */ | |
151 | # define SET_SCHEDULER_CMD \ | |
152 | "exec 0</dev/null " \ | |
153 | " 1>/sys/block/%s/queue/scheduler " \ | |
154 | " 2>/dev/null; " \ | |
155 | "echo %s" | |
156 | ||
157 | { | |
158 | char *argv[] = { "/bin/sh", "-c", NULL, NULL }; | |
159 | char *envp[] = { NULL }; | |
160 | ||
161 | argv[2] = kmem_asprintf(SET_SCHEDULER_CMD, device, elevator); | |
761394b3 | 162 | error = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC); |
6d1d976b BB |
163 | strfree(argv[2]); |
164 | } | |
165 | #endif /* HAVE_ELEVATOR_CHANGE */ | |
6839eed2 BB |
166 | if (error) |
167 | printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n", | |
fdcd952b | 168 | elevator, v->vdev_path, device, error); |
6839eed2 BB |
169 | |
170 | return (error); | |
171 | } | |
172 | ||
b5a28807 ED |
173 | /* |
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. | |
197 | */ | |
198 | static struct block_device * | |
199 | vdev_disk_rrpart(const char *path, int mode, vdev_disk_t *vd) | |
200 | { | |
201 | #if defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) | |
202 | struct block_device *bdev, *result = ERR_PTR(-ENXIO); | |
203 | struct gendisk *disk; | |
204 | int error, partno; | |
205 | ||
206 | bdev = vdev_bdev_open(path, vdev_bdev_mode(mode), vd); | |
207 | if (IS_ERR(bdev)) | |
208 | return bdev; | |
209 | ||
210 | disk = get_gendisk(bdev->bd_dev, &partno); | |
211 | vdev_bdev_close(bdev, vdev_bdev_mode(mode)); | |
212 | ||
213 | if (disk) { | |
214 | bdev = bdget(disk_devt(disk)); | |
215 | if (bdev) { | |
216 | error = blkdev_get(bdev, vdev_bdev_mode(mode), vd); | |
217 | if (error == 0) | |
218 | error = ioctl_by_bdev(bdev, BLKRRPART, 0); | |
219 | vdev_bdev_close(bdev, vdev_bdev_mode(mode)); | |
220 | } | |
221 | ||
222 | bdev = bdget_disk(disk, partno); | |
223 | if (bdev) { | |
224 | error = blkdev_get(bdev, | |
225 | vdev_bdev_mode(mode) | FMODE_EXCL, vd); | |
226 | if (error == 0) | |
227 | result = bdev; | |
228 | } | |
229 | put_disk(disk); | |
230 | } | |
231 | ||
232 | return result; | |
233 | #else | |
234 | return ERR_PTR(-EOPNOTSUPP); | |
235 | #endif /* defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) */ | |
236 | } | |
237 | ||
60101509 | 238 | static int |
1bd201e7 CS |
239 | vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize, |
240 | uint64_t *ashift) | |
60101509 | 241 | { |
b5a28807 | 242 | struct block_device *bdev = ERR_PTR(-ENXIO); |
60101509 BB |
243 | vdev_disk_t *vd; |
244 | int mode, block_size; | |
245 | ||
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; | |
249 | return EINVAL; | |
250 | } | |
251 | ||
b8d06fca | 252 | vd = kmem_zalloc(sizeof(vdev_disk_t), KM_PUSHPAGE); |
60101509 BB |
253 | if (vd == NULL) |
254 | return ENOMEM; | |
255 | ||
256 | /* | |
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. | |
269 | */ | |
270 | mode = spa_mode(v->vdev_spa); | |
b5a28807 ED |
271 | if (v->vdev_wholedisk && v->vdev_expanding) |
272 | bdev = vdev_disk_rrpart(v->vdev_path, mode, vd); | |
273 | if (IS_ERR(bdev)) | |
274 | bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd); | |
60101509 BB |
275 | if (IS_ERR(bdev)) { |
276 | kmem_free(vd, sizeof(vdev_disk_t)); | |
277 | return -PTR_ERR(bdev); | |
278 | } | |
279 | ||
280 | v->vdev_tsd = vd; | |
281 | vd->vd_bdev = bdev; | |
282 | block_size = vdev_bdev_block_size(bdev); | |
283 | ||
3a7381e5 NB |
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 | |
288 | * removed. */ | |
289 | ASSERT3S(v->vdev_wholedisk, !=, -1ULL); | |
290 | ||
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 | |
294 | * a partition. */ | |
295 | if (v->vdev_wholedisk == -1ULL) | |
296 | v->vdev_wholedisk = (bdev->bd_contains == bdev); | |
60101509 BB |
297 | |
298 | /* Clear the nowritecache bit, causes vdev_reopen() to try again. */ | |
299 | v->vdev_nowritecache = B_FALSE; | |
300 | ||
301 | /* Physical volume size in bytes */ | |
f74fae8b | 302 | *psize = bdev_capacity(bdev); |
60101509 | 303 | |
1bd201e7 CS |
304 | /* TODO: report possible expansion size */ |
305 | *max_psize = *psize; | |
306 | ||
60101509 BB |
307 | /* Based on the minimum sector size set the block size */ |
308 | *ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1; | |
309 | ||
6839eed2 | 310 | /* Try to set the io scheduler elevator algorithm */ |
fdcd952b | 311 | (void) vdev_elevator_switch(v, zfs_vdev_scheduler); |
6839eed2 | 312 | |
60101509 BB |
313 | return 0; |
314 | } | |
315 | ||
316 | static void | |
317 | vdev_disk_close(vdev_t *v) | |
318 | { | |
319 | vdev_disk_t *vd = v->vdev_tsd; | |
320 | ||
321 | if (vd == NULL) | |
322 | return; | |
323 | ||
324 | if (vd->vd_bdev != NULL) | |
325 | vdev_bdev_close(vd->vd_bdev, | |
326 | vdev_bdev_mode(spa_mode(v->vdev_spa))); | |
327 | ||
328 | kmem_free(vd, sizeof(vdev_disk_t)); | |
329 | v->vdev_tsd = NULL; | |
330 | } | |
331 | ||
332 | static dio_request_t * | |
333 | vdev_disk_dio_alloc(int bio_count) | |
334 | { | |
335 | dio_request_t *dr; | |
336 | int i; | |
337 | ||
338 | dr = kmem_zalloc(sizeof(dio_request_t) + | |
b8d06fca | 339 | sizeof(struct bio *) * bio_count, KM_PUSHPAGE); |
60101509 BB |
340 | if (dr) { |
341 | init_completion(&dr->dr_comp); | |
342 | atomic_set(&dr->dr_ref, 0); | |
343 | dr->dr_bio_count = bio_count; | |
344 | dr->dr_error = 0; | |
345 | ||
346 | for (i = 0; i < dr->dr_bio_count; i++) | |
347 | dr->dr_bio[i] = NULL; | |
348 | } | |
349 | ||
350 | return dr; | |
351 | } | |
352 | ||
353 | static void | |
354 | vdev_disk_dio_free(dio_request_t *dr) | |
355 | { | |
356 | int i; | |
357 | ||
358 | for (i = 0; i < dr->dr_bio_count; i++) | |
359 | if (dr->dr_bio[i]) | |
360 | bio_put(dr->dr_bio[i]); | |
361 | ||
362 | kmem_free(dr, sizeof(dio_request_t) + | |
363 | sizeof(struct bio *) * dr->dr_bio_count); | |
364 | } | |
365 | ||
675de5aa BB |
366 | static int |
367 | vdev_disk_dio_is_sync(dio_request_t *dr) | |
368 | { | |
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)); | |
372 | #else | |
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)); | |
376 | # else | |
377 | # ifdef HAVE_REQ_SYNC | |
378 | /* REQ_SYNC preferred interface from 2.6.36-2.6.xx */ | |
379 | return (dr->dr_rw & REQ_SYNC); | |
380 | # else | |
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 */ | |
385 | } | |
386 | ||
60101509 BB |
387 | static void |
388 | vdev_disk_dio_get(dio_request_t *dr) | |
389 | { | |
390 | atomic_inc(&dr->dr_ref); | |
391 | } | |
392 | ||
393 | static int | |
394 | vdev_disk_dio_put(dio_request_t *dr) | |
395 | { | |
396 | int rc = atomic_dec_return(&dr->dr_ref); | |
397 | ||
398 | /* | |
399 | * Free the dio_request when the last reference is dropped and | |
400 | * ensure zio_interpret is called only once with the correct zio | |
401 | */ | |
402 | if (rc == 0) { | |
403 | zio_t *zio = dr->dr_zio; | |
404 | int error = dr->dr_error; | |
405 | ||
406 | vdev_disk_dio_free(dr); | |
407 | ||
408 | if (zio) { | |
a69052be BB |
409 | zio->io_delay = jiffies_to_msecs( |
410 | jiffies_64 - zio->io_delay); | |
60101509 | 411 | zio->io_error = error; |
d148e951 BB |
412 | ASSERT3S(zio->io_error, >=, 0); |
413 | if (zio->io_error) | |
414 | vdev_disk_error(zio); | |
60101509 BB |
415 | zio_interrupt(zio); |
416 | } | |
417 | } | |
418 | ||
419 | return rc; | |
420 | } | |
421 | ||
422 | BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, size, error) | |
423 | { | |
424 | dio_request_t *dr = bio->bi_private; | |
425 | int rc; | |
426 | ||
427 | /* Fatal error but print some useful debugging before asserting */ | |
428 | if (dr == NULL) | |
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)); | |
435 | ||
436 | #ifndef HAVE_2ARGS_BIO_END_IO_T | |
437 | if (bio->bi_size) | |
438 | return 1; | |
439 | #endif /* HAVE_2ARGS_BIO_END_IO_T */ | |
440 | ||
441 | if (error == 0 && !test_bit(BIO_UPTODATE, &bio->bi_flags)) | |
d148e951 | 442 | error = -EIO; |
60101509 BB |
443 | |
444 | if (dr->dr_error == 0) | |
d148e951 | 445 | dr->dr_error = -error; |
60101509 BB |
446 | |
447 | /* Drop reference aquired by __vdev_disk_physio */ | |
448 | rc = vdev_disk_dio_put(dr); | |
449 | ||
450 | /* Wake up synchronous waiter this is the last outstanding bio */ | |
675de5aa | 451 | if ((rc == 1) && vdev_disk_dio_is_sync(dr)) |
60101509 BB |
452 | complete(&dr->dr_comp); |
453 | ||
454 | BIO_END_IO_RETURN(0); | |
455 | } | |
456 | ||
457 | static inline unsigned long | |
458 | bio_nr_pages(void *bio_ptr, unsigned int bio_size) | |
459 | { | |
460 | return ((((unsigned long)bio_ptr + bio_size + PAGE_SIZE - 1) >> | |
461 | PAGE_SHIFT) - ((unsigned long)bio_ptr >> PAGE_SHIFT)); | |
462 | } | |
463 | ||
464 | static unsigned int | |
465 | bio_map(struct bio *bio, void *bio_ptr, unsigned int bio_size) | |
466 | { | |
467 | unsigned int offset, size, i; | |
468 | struct page *page; | |
469 | ||
470 | offset = offset_in_page(bio_ptr); | |
471 | for (i = 0; i < bio->bi_max_vecs; i++) { | |
472 | size = PAGE_SIZE - offset; | |
473 | ||
474 | if (bio_size <= 0) | |
475 | break; | |
476 | ||
477 | if (size > bio_size) | |
478 | size = bio_size; | |
479 | ||
480 | if (kmem_virt(bio_ptr)) | |
481 | page = vmalloc_to_page(bio_ptr); | |
482 | else | |
483 | page = virt_to_page(bio_ptr); | |
484 | ||
485 | if (bio_add_page(bio, page, size, offset) != size) | |
486 | break; | |
487 | ||
488 | bio_ptr += size; | |
489 | bio_size -= size; | |
490 | offset = 0; | |
491 | } | |
492 | ||
493 | return bio_size; | |
494 | } | |
495 | ||
496 | static int | |
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) | |
499 | { | |
500 | dio_request_t *dr; | |
501 | caddr_t bio_ptr; | |
502 | uint64_t bio_offset; | |
503 | int bio_size, bio_count = 16; | |
f74fae8b | 504 | int i = 0, error = 0; |
60101509 | 505 | |
e06be586 NB |
506 | ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size); |
507 | ||
60101509 BB |
508 | retry: |
509 | dr = vdev_disk_dio_alloc(bio_count); | |
510 | if (dr == NULL) | |
511 | return ENOMEM; | |
512 | ||
2959d94a BB |
513 | if (zio && !(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD))) |
514 | bio_set_flags_failfast(bdev, &flags); | |
515 | ||
60101509 BB |
516 | dr->dr_zio = zio; |
517 | dr->dr_rw = flags; | |
60101509 | 518 | |
60101509 BB |
519 | /* |
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. | |
525 | */ | |
526 | bio_ptr = kbuf_ptr; | |
527 | bio_offset = kbuf_offset; | |
528 | bio_size = kbuf_size; | |
529 | for (i = 0; i <= dr->dr_bio_count; i++) { | |
530 | ||
531 | /* Finished constructing bio's for given buffer */ | |
532 | if (bio_size <= 0) | |
533 | break; | |
534 | ||
535 | /* | |
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. | |
539 | */ | |
540 | if (dr->dr_bio_count == i) { | |
541 | vdev_disk_dio_free(dr); | |
542 | bio_count *= 2; | |
60101509 BB |
543 | goto retry; |
544 | } | |
545 | ||
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); | |
550 | return ENOMEM; | |
551 | } | |
552 | ||
553 | /* Matching put called by vdev_disk_physio_completion */ | |
554 | vdev_disk_dio_get(dr); | |
555 | ||
556 | dr->dr_bio[i]->bi_bdev = bdev; | |
f74fae8b | 557 | dr->dr_bio[i]->bi_sector = bio_offset >> 9; |
60101509 BB |
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; | |
561 | ||
562 | /* Remaining size is returned to become the new size */ | |
563 | bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size); | |
564 | ||
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; | |
568 | } | |
569 | ||
570 | /* Extra reference to protect dio_request during submit_bio */ | |
571 | vdev_disk_dio_get(dr); | |
a69052be BB |
572 | if (zio) |
573 | zio->io_delay = jiffies_64; | |
60101509 BB |
574 | |
575 | /* Submit all bio's associated with this dio */ | |
576 | for (i = 0; i < dr->dr_bio_count; i++) | |
577 | if (dr->dr_bio[i]) | |
578 | submit_bio(dr->dr_rw, dr->dr_bio[i]); | |
579 | ||
580 | /* | |
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. | |
587 | */ | |
675de5aa | 588 | if (vdev_disk_dio_is_sync(dr)) { |
60101509 BB |
589 | wait_for_completion(&dr->dr_comp); |
590 | error = dr->dr_error; | |
591 | ASSERT3S(atomic_read(&dr->dr_ref), ==, 1); | |
592 | } | |
593 | ||
594 | (void)vdev_disk_dio_put(dr); | |
595 | ||
596 | return error; | |
597 | } | |
598 | ||
599 | int | |
600 | vdev_disk_physio(struct block_device *bdev, caddr_t kbuf, | |
601 | size_t size, uint64_t offset, int flags) | |
602 | { | |
2959d94a | 603 | bio_set_flags_failfast(bdev, &flags); |
60101509 BB |
604 | return __vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags); |
605 | } | |
606 | ||
607 | /* 2.6.24 API change */ | |
608 | #ifdef HAVE_BIO_EMPTY_BARRIER | |
609 | BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, size, rc) | |
610 | { | |
611 | zio_t *zio = bio->bi_private; | |
612 | ||
a69052be | 613 | zio->io_delay = jiffies_to_msecs(jiffies_64 - zio->io_delay); |
60101509 BB |
614 | zio->io_error = -rc; |
615 | if (rc && (rc == -EOPNOTSUPP)) | |
616 | zio->io_vd->vdev_nowritecache = B_TRUE; | |
617 | ||
618 | bio_put(bio); | |
d148e951 BB |
619 | ASSERT3S(zio->io_error, >=, 0); |
620 | if (zio->io_error) | |
621 | vdev_disk_error(zio); | |
60101509 BB |
622 | zio_interrupt(zio); |
623 | ||
624 | BIO_END_IO_RETURN(0); | |
625 | } | |
626 | ||
627 | static int | |
628 | vdev_disk_io_flush(struct block_device *bdev, zio_t *zio) | |
629 | { | |
630 | struct request_queue *q; | |
631 | struct bio *bio; | |
632 | ||
633 | q = bdev_get_queue(bdev); | |
634 | if (!q) | |
635 | return ENXIO; | |
636 | ||
637 | bio = bio_alloc(GFP_KERNEL, 0); | |
638 | if (!bio) | |
639 | return ENOMEM; | |
640 | ||
641 | bio->bi_end_io = vdev_disk_io_flush_completion; | |
642 | bio->bi_private = zio; | |
643 | bio->bi_bdev = bdev; | |
a69052be | 644 | zio->io_delay = jiffies_64; |
96801d29 | 645 | submit_bio(VDEV_WRITE_FLUSH_FUA, bio); |
60101509 BB |
646 | |
647 | return 0; | |
648 | } | |
649 | #else | |
650 | static int | |
651 | vdev_disk_io_flush(struct block_device *bdev, zio_t *zio) | |
652 | { | |
653 | return ENOTSUP; | |
654 | } | |
655 | #endif /* HAVE_BIO_EMPTY_BARRIER */ | |
656 | ||
657 | static int | |
658 | vdev_disk_io_start(zio_t *zio) | |
659 | { | |
660 | vdev_t *v = zio->io_vd; | |
661 | vdev_disk_t *vd = v->vdev_tsd; | |
662 | int flags, error; | |
663 | ||
664 | switch (zio->io_type) { | |
665 | case ZIO_TYPE_IOCTL: | |
666 | ||
667 | if (!vdev_readable(v)) { | |
668 | zio->io_error = ENXIO; | |
669 | return ZIO_PIPELINE_CONTINUE; | |
670 | } | |
671 | ||
672 | switch (zio->io_cmd) { | |
673 | case DKIOCFLUSHWRITECACHE: | |
674 | ||
675 | if (zfs_nocacheflush) | |
676 | break; | |
677 | ||
678 | if (v->vdev_nowritecache) { | |
679 | zio->io_error = ENOTSUP; | |
680 | break; | |
681 | } | |
682 | ||
683 | error = vdev_disk_io_flush(vd->vd_bdev, zio); | |
684 | if (error == 0) | |
685 | return ZIO_PIPELINE_STOP; | |
686 | ||
687 | zio->io_error = error; | |
688 | if (error == ENOTSUP) | |
689 | v->vdev_nowritecache = B_TRUE; | |
690 | ||
691 | break; | |
692 | ||
693 | default: | |
694 | zio->io_error = ENOTSUP; | |
695 | } | |
696 | ||
697 | return ZIO_PIPELINE_CONTINUE; | |
698 | ||
699 | case ZIO_TYPE_WRITE: | |
700 | flags = WRITE; | |
701 | break; | |
702 | ||
703 | case ZIO_TYPE_READ: | |
704 | flags = READ; | |
705 | break; | |
706 | ||
707 | default: | |
708 | zio->io_error = ENOTSUP; | |
709 | return ZIO_PIPELINE_CONTINUE; | |
710 | } | |
711 | ||
60101509 BB |
712 | error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data, |
713 | zio->io_size, zio->io_offset, flags); | |
714 | if (error) { | |
715 | zio->io_error = error; | |
716 | return ZIO_PIPELINE_CONTINUE; | |
717 | } | |
718 | ||
719 | return ZIO_PIPELINE_STOP; | |
720 | } | |
721 | ||
722 | static void | |
723 | vdev_disk_io_done(zio_t *zio) | |
724 | { | |
725 | /* | |
726 | * If the device returned EIO, we revalidate the media. If it is | |
727 | * determined the media has changed this triggers the asynchronous | |
728 | * removal of the device from the configuration. | |
729 | */ | |
730 | if (zio->io_error == EIO) { | |
731 | vdev_t *v = zio->io_vd; | |
732 | vdev_disk_t *vd = v->vdev_tsd; | |
733 | ||
734 | if (check_disk_change(vd->vd_bdev)) { | |
735 | vdev_bdev_invalidate(vd->vd_bdev); | |
736 | v->vdev_remove_wanted = B_TRUE; | |
737 | spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE); | |
738 | } | |
739 | } | |
740 | } | |
741 | ||
742 | static void | |
743 | vdev_disk_hold(vdev_t *vd) | |
744 | { | |
745 | ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER)); | |
746 | ||
747 | /* We must have a pathname, and it must be absolute. */ | |
748 | if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') | |
749 | return; | |
750 | ||
751 | /* | |
752 | * Only prefetch path and devid info if the device has | |
753 | * never been opened. | |
754 | */ | |
755 | if (vd->vdev_tsd != NULL) | |
756 | return; | |
757 | ||
758 | /* XXX: Implement me as a vnode lookup for the device */ | |
759 | vd->vdev_name_vp = NULL; | |
760 | vd->vdev_devid_vp = NULL; | |
761 | } | |
762 | ||
763 | static void | |
764 | vdev_disk_rele(vdev_t *vd) | |
765 | { | |
766 | ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER)); | |
767 | ||
768 | /* XXX: Implement me as a vnode rele for the device */ | |
769 | } | |
770 | ||
771 | vdev_ops_t vdev_disk_ops = { | |
772 | vdev_disk_open, | |
773 | vdev_disk_close, | |
774 | vdev_default_asize, | |
775 | vdev_disk_io_start, | |
776 | vdev_disk_io_done, | |
777 | NULL, | |
778 | vdev_disk_hold, | |
779 | vdev_disk_rele, | |
780 | VDEV_TYPE_DISK, /* name of this vdev type */ | |
781 | B_TRUE /* leaf vdev */ | |
782 | }; | |
783 | ||
784 | /* | |
785 | * Given the root disk device devid or pathname, read the label from | |
786 | * the device, and construct a configuration nvlist. | |
787 | */ | |
788 | int | |
789 | vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config) | |
790 | { | |
791 | struct block_device *bdev; | |
792 | vdev_label_t *label; | |
793 | uint64_t s, size; | |
794 | int i; | |
795 | ||
796 | bdev = vdev_bdev_open(devpath, vdev_bdev_mode(FREAD), NULL); | |
797 | if (IS_ERR(bdev)) | |
798 | return -PTR_ERR(bdev); | |
799 | ||
f74fae8b | 800 | s = bdev_capacity(bdev); |
60101509 BB |
801 | if (s == 0) { |
802 | vdev_bdev_close(bdev, vdev_bdev_mode(FREAD)); | |
803 | return EIO; | |
804 | } | |
805 | ||
806 | size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t); | |
b8d06fca | 807 | label = vmem_alloc(sizeof(vdev_label_t), KM_PUSHPAGE); |
60101509 BB |
808 | |
809 | for (i = 0; i < VDEV_LABELS; i++) { | |
810 | uint64_t offset, state, txg = 0; | |
811 | ||
812 | /* read vdev label */ | |
813 | offset = vdev_label_offset(size, i, 0); | |
814 | if (vdev_disk_physio(bdev, (caddr_t)label, | |
815 | VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, READ_SYNC) != 0) | |
816 | continue; | |
817 | ||
818 | if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist, | |
819 | sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) { | |
820 | *config = NULL; | |
821 | continue; | |
822 | } | |
823 | ||
824 | if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE, | |
825 | &state) != 0 || state >= POOL_STATE_DESTROYED) { | |
826 | nvlist_free(*config); | |
827 | *config = NULL; | |
828 | continue; | |
829 | } | |
830 | ||
831 | if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG, | |
832 | &txg) != 0 || txg == 0) { | |
833 | nvlist_free(*config); | |
834 | *config = NULL; | |
835 | continue; | |
836 | } | |
837 | ||
838 | break; | |
839 | } | |
840 | ||
841 | vmem_free(label, sizeof(vdev_label_t)); | |
842 | vdev_bdev_close(bdev, vdev_bdev_mode(FREAD)); | |
843 | ||
844 | return 0; | |
845 | } | |
6839eed2 BB |
846 | |
847 | module_param(zfs_vdev_scheduler, charp, 0644); | |
c409e464 | 848 | MODULE_PARM_DESC(zfs_vdev_scheduler, "I/O scheduler"); |