fs/block_dev.c: fix performance regression in O_DIRECT|O_SYNC writes to block devices
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / partitions / check.c
blobe238ab23a9e7ff50f712847e3660f4f67a5912d2
1 /*
2 * fs/partitions/check.c
4 * Code extracted from drivers/block/genhd.c
5 * Copyright (C) 1991-1998 Linus Torvalds
6 * Re-organised Feb 1998 Russell King
8 * We now have independent partition support from the
9 * block drivers, which allows all the partition code to
10 * be grouped in one location, and it to be mostly self
11 * contained.
13 * Added needed MAJORS for new pairs, {hdi,hdj}, {hdk,hdl}
16 #include <linux/init.h>
17 #include <linux/module.h>
18 #include <linux/fs.h>
19 #include <linux/slab.h>
20 #include <linux/kmod.h>
21 #include <linux/ctype.h>
22 #include <linux/genhd.h>
23 #include <linux/blktrace_api.h>
25 #include "check.h"
27 #include "acorn.h"
28 #include "amiga.h"
29 #include "atari.h"
30 #include "ldm.h"
31 #include "mac.h"
32 #include "msdos.h"
33 #include "osf.h"
34 #include "sgi.h"
35 #include "sun.h"
36 #include "ibm.h"
37 #include "ultrix.h"
38 #include "efi.h"
39 #include "karma.h"
40 #include "sysv68.h"
42 #ifdef CONFIG_BLK_DEV_MD
43 extern void md_autodetect_dev(dev_t dev);
44 #endif
46 int warn_no_part = 1; /*This is ugly: should make genhd removable media aware*/
48 static int (*check_part[])(struct parsed_partitions *, struct block_device *) = {
50 * Probe partition formats with tables at disk address 0
51 * that also have an ADFS boot block at 0xdc0.
53 #ifdef CONFIG_ACORN_PARTITION_ICS
54 adfspart_check_ICS,
55 #endif
56 #ifdef CONFIG_ACORN_PARTITION_POWERTEC
57 adfspart_check_POWERTEC,
58 #endif
59 #ifdef CONFIG_ACORN_PARTITION_EESOX
60 adfspart_check_EESOX,
61 #endif
64 * Now move on to formats that only have partition info at
65 * disk address 0xdc0. Since these may also have stale
66 * PC/BIOS partition tables, they need to come before
67 * the msdos entry.
69 #ifdef CONFIG_ACORN_PARTITION_CUMANA
70 adfspart_check_CUMANA,
71 #endif
72 #ifdef CONFIG_ACORN_PARTITION_ADFS
73 adfspart_check_ADFS,
74 #endif
76 #ifdef CONFIG_EFI_PARTITION
77 efi_partition, /* this must come before msdos */
78 #endif
79 #ifdef CONFIG_SGI_PARTITION
80 sgi_partition,
81 #endif
82 #ifdef CONFIG_LDM_PARTITION
83 ldm_partition, /* this must come before msdos */
84 #endif
85 #ifdef CONFIG_MSDOS_PARTITION
86 msdos_partition,
87 #endif
88 #ifdef CONFIG_OSF_PARTITION
89 osf_partition,
90 #endif
91 #ifdef CONFIG_SUN_PARTITION
92 sun_partition,
93 #endif
94 #ifdef CONFIG_AMIGA_PARTITION
95 amiga_partition,
96 #endif
97 #ifdef CONFIG_ATARI_PARTITION
98 atari_partition,
99 #endif
100 #ifdef CONFIG_MAC_PARTITION
101 mac_partition,
102 #endif
103 #ifdef CONFIG_ULTRIX_PARTITION
104 ultrix_partition,
105 #endif
106 #ifdef CONFIG_IBM_PARTITION
107 ibm_partition,
108 #endif
109 #ifdef CONFIG_KARMA_PARTITION
110 karma_partition,
111 #endif
112 #ifdef CONFIG_SYSV68_PARTITION
113 sysv68_partition,
114 #endif
115 NULL
119 * disk_name() is used by partition check code and the genhd driver.
120 * It formats the devicename of the indicated disk into
121 * the supplied buffer (of size at least 32), and returns
122 * a pointer to that same buffer (for convenience).
125 char *disk_name(struct gendisk *hd, int partno, char *buf)
127 if (!partno)
128 snprintf(buf, BDEVNAME_SIZE, "%s", hd->disk_name);
129 else if (isdigit(hd->disk_name[strlen(hd->disk_name)-1]))
130 snprintf(buf, BDEVNAME_SIZE, "%sp%d", hd->disk_name, partno);
131 else
132 snprintf(buf, BDEVNAME_SIZE, "%s%d", hd->disk_name, partno);
134 return buf;
137 const char *bdevname(struct block_device *bdev, char *buf)
139 return disk_name(bdev->bd_disk, bdev->bd_part->partno, buf);
142 EXPORT_SYMBOL(bdevname);
145 * There's very little reason to use this, you should really
146 * have a struct block_device just about everywhere and use
147 * bdevname() instead.
149 const char *__bdevname(dev_t dev, char *buffer)
151 scnprintf(buffer, BDEVNAME_SIZE, "unknown-block(%u,%u)",
152 MAJOR(dev), MINOR(dev));
153 return buffer;
156 EXPORT_SYMBOL(__bdevname);
158 static struct parsed_partitions *
159 check_partition(struct gendisk *hd, struct block_device *bdev)
161 struct parsed_partitions *state;
162 int i, res, err;
164 state = kmalloc(sizeof(struct parsed_partitions), GFP_KERNEL);
165 if (!state)
166 return NULL;
168 disk_name(hd, 0, state->name);
169 printk(KERN_INFO " %s:", state->name);
170 if (isdigit(state->name[strlen(state->name)-1]))
171 sprintf(state->name, "p");
173 state->limit = disk_max_parts(hd);
174 i = res = err = 0;
175 while (!res && check_part[i]) {
176 memset(&state->parts, 0, sizeof(state->parts));
177 res = check_part[i++](state, bdev);
178 if (res < 0) {
179 /* We have hit an I/O error which we don't report now.
180 * But record it, and let the others do their job.
182 err = res;
183 res = 0;
187 if (res > 0)
188 return state;
189 if (err)
190 /* The partition is unrecognized. So report I/O errors if there were any */
191 res = err;
192 if (!res)
193 printk(" unknown partition table\n");
194 else if (warn_no_part)
195 printk(" unable to read partition table\n");
196 kfree(state);
197 return ERR_PTR(res);
200 static ssize_t part_partition_show(struct device *dev,
201 struct device_attribute *attr, char *buf)
203 struct hd_struct *p = dev_to_part(dev);
205 return sprintf(buf, "%d\n", p->partno);
208 static ssize_t part_start_show(struct device *dev,
209 struct device_attribute *attr, char *buf)
211 struct hd_struct *p = dev_to_part(dev);
213 return sprintf(buf, "%llu\n",(unsigned long long)p->start_sect);
216 ssize_t part_size_show(struct device *dev,
217 struct device_attribute *attr, char *buf)
219 struct hd_struct *p = dev_to_part(dev);
220 return sprintf(buf, "%llu\n",(unsigned long long)p->nr_sects);
223 ssize_t part_alignment_offset_show(struct device *dev,
224 struct device_attribute *attr, char *buf)
226 struct hd_struct *p = dev_to_part(dev);
227 return sprintf(buf, "%llu\n", (unsigned long long)p->alignment_offset);
230 ssize_t part_discard_alignment_show(struct device *dev,
231 struct device_attribute *attr, char *buf)
233 struct hd_struct *p = dev_to_part(dev);
234 return sprintf(buf, "%u\n", p->discard_alignment);
237 ssize_t part_stat_show(struct device *dev,
238 struct device_attribute *attr, char *buf)
240 struct hd_struct *p = dev_to_part(dev);
241 int cpu;
243 cpu = part_stat_lock();
244 part_round_stats(cpu, p);
245 part_stat_unlock();
246 return sprintf(buf,
247 "%8lu %8lu %8llu %8u "
248 "%8lu %8lu %8llu %8u "
249 "%8u %8u %8u"
250 "\n",
251 part_stat_read(p, ios[READ]),
252 part_stat_read(p, merges[READ]),
253 (unsigned long long)part_stat_read(p, sectors[READ]),
254 jiffies_to_msecs(part_stat_read(p, ticks[READ])),
255 part_stat_read(p, ios[WRITE]),
256 part_stat_read(p, merges[WRITE]),
257 (unsigned long long)part_stat_read(p, sectors[WRITE]),
258 jiffies_to_msecs(part_stat_read(p, ticks[WRITE])),
259 part_in_flight(p),
260 jiffies_to_msecs(part_stat_read(p, io_ticks)),
261 jiffies_to_msecs(part_stat_read(p, time_in_queue)));
264 ssize_t part_inflight_show(struct device *dev,
265 struct device_attribute *attr, char *buf)
267 struct hd_struct *p = dev_to_part(dev);
269 return sprintf(buf, "%8u %8u\n", p->in_flight[0], p->in_flight[1]);
272 #ifdef CONFIG_FAIL_MAKE_REQUEST
273 ssize_t part_fail_show(struct device *dev,
274 struct device_attribute *attr, char *buf)
276 struct hd_struct *p = dev_to_part(dev);
278 return sprintf(buf, "%d\n", p->make_it_fail);
281 ssize_t part_fail_store(struct device *dev,
282 struct device_attribute *attr,
283 const char *buf, size_t count)
285 struct hd_struct *p = dev_to_part(dev);
286 int i;
288 if (count > 0 && sscanf(buf, "%d", &i) > 0)
289 p->make_it_fail = (i == 0) ? 0 : 1;
291 return count;
293 #endif
295 static DEVICE_ATTR(partition, S_IRUGO, part_partition_show, NULL);
296 static DEVICE_ATTR(start, S_IRUGO, part_start_show, NULL);
297 static DEVICE_ATTR(size, S_IRUGO, part_size_show, NULL);
298 static DEVICE_ATTR(alignment_offset, S_IRUGO, part_alignment_offset_show, NULL);
299 static DEVICE_ATTR(discard_alignment, S_IRUGO, part_discard_alignment_show,
300 NULL);
301 static DEVICE_ATTR(stat, S_IRUGO, part_stat_show, NULL);
302 static DEVICE_ATTR(inflight, S_IRUGO, part_inflight_show, NULL);
303 #ifdef CONFIG_FAIL_MAKE_REQUEST
304 static struct device_attribute dev_attr_fail =
305 __ATTR(make-it-fail, S_IRUGO|S_IWUSR, part_fail_show, part_fail_store);
306 #endif
308 static struct attribute *part_attrs[] = {
309 &dev_attr_partition.attr,
310 &dev_attr_start.attr,
311 &dev_attr_size.attr,
312 &dev_attr_alignment_offset.attr,
313 &dev_attr_discard_alignment.attr,
314 &dev_attr_stat.attr,
315 &dev_attr_inflight.attr,
316 #ifdef CONFIG_FAIL_MAKE_REQUEST
317 &dev_attr_fail.attr,
318 #endif
319 NULL
322 static struct attribute_group part_attr_group = {
323 .attrs = part_attrs,
326 static const struct attribute_group *part_attr_groups[] = {
327 &part_attr_group,
328 #ifdef CONFIG_BLK_DEV_IO_TRACE
329 &blk_trace_attr_group,
330 #endif
331 NULL
334 static void part_release(struct device *dev)
336 struct hd_struct *p = dev_to_part(dev);
337 free_part_stats(p);
338 kfree(p);
341 struct device_type part_type = {
342 .name = "partition",
343 .groups = part_attr_groups,
344 .release = part_release,
347 static void delete_partition_rcu_cb(struct rcu_head *head)
349 struct hd_struct *part = container_of(head, struct hd_struct, rcu_head);
351 part->start_sect = 0;
352 part->nr_sects = 0;
353 part_stat_set_all(part, 0);
354 put_device(part_to_dev(part));
357 void delete_partition(struct gendisk *disk, int partno)
359 struct disk_part_tbl *ptbl = disk->part_tbl;
360 struct hd_struct *part;
362 if (partno >= ptbl->len)
363 return;
365 part = ptbl->part[partno];
366 if (!part)
367 return;
369 blk_free_devt(part_devt(part));
370 rcu_assign_pointer(ptbl->part[partno], NULL);
371 rcu_assign_pointer(ptbl->last_lookup, NULL);
372 kobject_put(part->holder_dir);
373 device_del(part_to_dev(part));
375 call_rcu(&part->rcu_head, delete_partition_rcu_cb);
378 static ssize_t whole_disk_show(struct device *dev,
379 struct device_attribute *attr, char *buf)
381 return 0;
383 static DEVICE_ATTR(whole_disk, S_IRUSR | S_IRGRP | S_IROTH,
384 whole_disk_show, NULL);
386 struct hd_struct *add_partition(struct gendisk *disk, int partno,
387 sector_t start, sector_t len, int flags)
389 struct hd_struct *p;
390 dev_t devt = MKDEV(0, 0);
391 struct device *ddev = disk_to_dev(disk);
392 struct device *pdev;
393 struct disk_part_tbl *ptbl;
394 const char *dname;
395 int err;
397 err = disk_expand_part_tbl(disk, partno);
398 if (err)
399 return ERR_PTR(err);
400 ptbl = disk->part_tbl;
402 if (ptbl->part[partno])
403 return ERR_PTR(-EBUSY);
405 p = kzalloc(sizeof(*p), GFP_KERNEL);
406 if (!p)
407 return ERR_PTR(-EBUSY);
409 if (!init_part_stats(p)) {
410 err = -ENOMEM;
411 goto out_free;
413 pdev = part_to_dev(p);
415 p->start_sect = start;
416 p->alignment_offset =
417 queue_limit_alignment_offset(&disk->queue->limits, start);
418 p->discard_alignment =
419 queue_limit_discard_alignment(&disk->queue->limits, start);
420 p->nr_sects = len;
421 p->partno = partno;
422 p->policy = get_disk_ro(disk);
424 dname = dev_name(ddev);
425 if (isdigit(dname[strlen(dname) - 1]))
426 dev_set_name(pdev, "%sp%d", dname, partno);
427 else
428 dev_set_name(pdev, "%s%d", dname, partno);
430 device_initialize(pdev);
431 pdev->class = &block_class;
432 pdev->type = &part_type;
433 pdev->parent = ddev;
435 err = blk_alloc_devt(p, &devt);
436 if (err)
437 goto out_free_stats;
438 pdev->devt = devt;
440 /* delay uevent until 'holders' subdir is created */
441 dev_set_uevent_suppress(pdev, 1);
442 err = device_add(pdev);
443 if (err)
444 goto out_put;
446 err = -ENOMEM;
447 p->holder_dir = kobject_create_and_add("holders", &pdev->kobj);
448 if (!p->holder_dir)
449 goto out_del;
451 dev_set_uevent_suppress(pdev, 0);
452 if (flags & ADDPART_FLAG_WHOLEDISK) {
453 err = device_create_file(pdev, &dev_attr_whole_disk);
454 if (err)
455 goto out_del;
458 /* everything is up and running, commence */
459 INIT_RCU_HEAD(&p->rcu_head);
460 rcu_assign_pointer(ptbl->part[partno], p);
462 /* suppress uevent if the disk supresses it */
463 if (!dev_get_uevent_suppress(ddev))
464 kobject_uevent(&pdev->kobj, KOBJ_ADD);
466 return p;
468 out_free_stats:
469 free_part_stats(p);
470 out_free:
471 kfree(p);
472 return ERR_PTR(err);
473 out_del:
474 kobject_put(p->holder_dir);
475 device_del(pdev);
476 out_put:
477 put_device(pdev);
478 blk_free_devt(devt);
479 return ERR_PTR(err);
482 /* Not exported, helper to add_disk(). */
483 void register_disk(struct gendisk *disk)
485 struct device *ddev = disk_to_dev(disk);
486 struct block_device *bdev;
487 struct disk_part_iter piter;
488 struct hd_struct *part;
489 int err;
491 ddev->parent = disk->driverfs_dev;
493 dev_set_name(ddev, disk->disk_name);
495 /* delay uevents, until we scanned partition table */
496 dev_set_uevent_suppress(ddev, 1);
498 if (device_add(ddev))
499 return;
500 #ifndef CONFIG_SYSFS_DEPRECATED
501 err = sysfs_create_link(block_depr, &ddev->kobj,
502 kobject_name(&ddev->kobj));
503 if (err) {
504 device_del(ddev);
505 return;
507 #endif
508 disk->part0.holder_dir = kobject_create_and_add("holders", &ddev->kobj);
509 disk->slave_dir = kobject_create_and_add("slaves", &ddev->kobj);
511 /* No minors to use for partitions */
512 if (!disk_partitionable(disk))
513 goto exit;
515 /* No such device (e.g., media were just removed) */
516 if (!get_capacity(disk))
517 goto exit;
519 bdev = bdget_disk(disk, 0);
520 if (!bdev)
521 goto exit;
523 bdev->bd_invalidated = 1;
524 err = blkdev_get(bdev, FMODE_READ);
525 if (err < 0)
526 goto exit;
527 blkdev_put(bdev, FMODE_READ);
529 exit:
530 /* announce disk after possible partitions are created */
531 dev_set_uevent_suppress(ddev, 0);
532 kobject_uevent(&ddev->kobj, KOBJ_ADD);
534 /* announce possible partitions */
535 disk_part_iter_init(&piter, disk, 0);
536 while ((part = disk_part_iter_next(&piter)))
537 kobject_uevent(&part_to_dev(part)->kobj, KOBJ_ADD);
538 disk_part_iter_exit(&piter);
541 int rescan_partitions(struct gendisk *disk, struct block_device *bdev)
543 struct disk_part_iter piter;
544 struct hd_struct *part;
545 struct parsed_partitions *state;
546 int p, highest, res;
548 if (bdev->bd_part_count)
549 return -EBUSY;
550 res = invalidate_partition(disk, 0);
551 if (res)
552 return res;
554 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_EMPTY);
555 while ((part = disk_part_iter_next(&piter)))
556 delete_partition(disk, part->partno);
557 disk_part_iter_exit(&piter);
559 if (disk->fops->revalidate_disk)
560 disk->fops->revalidate_disk(disk);
561 check_disk_size_change(disk, bdev);
562 bdev->bd_invalidated = 0;
563 if (!get_capacity(disk) || !(state = check_partition(disk, bdev)))
564 return 0;
565 if (IS_ERR(state)) /* I/O error reading the partition table */
566 return -EIO;
568 /* tell userspace that the media / partition table may have changed */
569 kobject_uevent(&disk_to_dev(disk)->kobj, KOBJ_CHANGE);
571 /* Detect the highest partition number and preallocate
572 * disk->part_tbl. This is an optimization and not strictly
573 * necessary.
575 for (p = 1, highest = 0; p < state->limit; p++)
576 if (state->parts[p].size)
577 highest = p;
579 disk_expand_part_tbl(disk, highest);
581 /* add partitions */
582 for (p = 1; p < state->limit; p++) {
583 sector_t size, from;
584 try_scan:
585 size = state->parts[p].size;
586 if (!size)
587 continue;
589 from = state->parts[p].from;
590 if (from >= get_capacity(disk)) {
591 printk(KERN_WARNING
592 "%s: p%d ignored, start %llu is behind the end of the disk\n",
593 disk->disk_name, p, (unsigned long long) from);
594 continue;
597 if (from + size > get_capacity(disk)) {
598 const struct block_device_operations *bdops = disk->fops;
599 unsigned long long capacity;
601 printk(KERN_WARNING
602 "%s: p%d size %llu exceeds device capacity, ",
603 disk->disk_name, p, (unsigned long long) size);
605 if (bdops->set_capacity &&
606 (disk->flags & GENHD_FL_NATIVE_CAPACITY) == 0) {
607 printk(KERN_CONT "enabling native capacity\n");
608 capacity = bdops->set_capacity(disk, ~0ULL);
609 disk->flags |= GENHD_FL_NATIVE_CAPACITY;
610 if (capacity > get_capacity(disk)) {
611 set_capacity(disk, capacity);
612 check_disk_size_change(disk, bdev);
613 bdev->bd_invalidated = 0;
615 goto try_scan;
616 } else {
618 * we can not ignore partitions of broken tables
619 * created by for example camera firmware, but
620 * we limit them to the end of the disk to avoid
621 * creating invalid block devices
623 printk(KERN_CONT "limited to end of disk\n");
624 size = get_capacity(disk) - from;
627 part = add_partition(disk, p, from, size,
628 state->parts[p].flags);
629 if (IS_ERR(part)) {
630 printk(KERN_ERR " %s: p%d could not be added: %ld\n",
631 disk->disk_name, p, -PTR_ERR(part));
632 continue;
634 #ifdef CONFIG_BLK_DEV_MD
635 if (state->parts[p].flags & ADDPART_FLAG_RAID)
636 md_autodetect_dev(part_to_dev(part)->devt);
637 #endif
639 kfree(state);
640 return 0;
643 unsigned char *read_dev_sector(struct block_device *bdev, sector_t n, Sector *p)
645 struct address_space *mapping = bdev->bd_inode->i_mapping;
646 struct page *page;
648 page = read_mapping_page(mapping, (pgoff_t)(n >> (PAGE_CACHE_SHIFT-9)),
649 NULL);
650 if (!IS_ERR(page)) {
651 if (PageError(page))
652 goto fail;
653 p->v = page;
654 return (unsigned char *)page_address(page) + ((n & ((1 << (PAGE_CACHE_SHIFT - 9)) - 1)) << 9);
655 fail:
656 page_cache_release(page);
658 p->v = NULL;
659 return NULL;
662 EXPORT_SYMBOL(read_dev_sector);
664 void del_gendisk(struct gendisk *disk)
666 struct disk_part_iter piter;
667 struct hd_struct *part;
669 /* invalidate stuff */
670 disk_part_iter_init(&piter, disk,
671 DISK_PITER_INCL_EMPTY | DISK_PITER_REVERSE);
672 while ((part = disk_part_iter_next(&piter))) {
673 invalidate_partition(disk, part->partno);
674 delete_partition(disk, part->partno);
676 disk_part_iter_exit(&piter);
678 invalidate_partition(disk, 0);
679 blk_free_devt(disk_to_dev(disk)->devt);
680 set_capacity(disk, 0);
681 disk->flags &= ~GENHD_FL_UP;
682 unlink_gendisk(disk);
683 part_stat_set_all(&disk->part0, 0);
684 disk->part0.stamp = 0;
686 kobject_put(disk->part0.holder_dir);
687 kobject_put(disk->slave_dir);
688 disk->driverfs_dev = NULL;
689 #ifndef CONFIG_SYSFS_DEPRECATED
690 sysfs_remove_link(block_depr, dev_name(disk_to_dev(disk)));
691 #endif
692 device_del(disk_to_dev(disk));