blob: 2ad91ddad8e207631c1acf2153a1ded3d4766731 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * gendisk handling
3 */
4
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/module.h>
6#include <linux/fs.h>
7#include <linux/genhd.h>
Andrew Mortonb446b602007-02-20 13:57:48 -08008#include <linux/kdev_t.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/kernel.h>
10#include <linux/blkdev.h>
11#include <linux/init.h>
12#include <linux/spinlock.h>
Alexey Dobriyanf5009752008-10-04 23:53:21 +040013#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/seq_file.h>
15#include <linux/slab.h>
16#include <linux/kmod.h>
17#include <linux/kobj_map.h>
Christoph Hellwig2ef41632005-05-05 16:15:59 -070018#include <linux/buffer_head.h>
Jes Sorensen58383af2006-02-06 14:12:43 -080019#include <linux/mutex.h>
Tejun Heobcce3de2008-08-25 19:47:22 +090020#include <linux/idr.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021
Adrian Bunkff889722008-03-04 11:23:45 +010022#include "blk.h"
23
Kay Sieversedfaa7c2007-05-21 22:08:01 +020024static DEFINE_MUTEX(block_class_lock);
25#ifndef CONFIG_SYSFS_DEPRECATED
26struct kobject *block_depr;
27#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -070028
Tejun Heobcce3de2008-08-25 19:47:22 +090029/* for extended dynamic devt allocation, currently only one major is used */
30#define MAX_EXT_DEVT (1 << MINORBITS)
31
32/* For extended devt allocation. ext_devt_mutex prevents look up
33 * results from going away underneath its user.
34 */
35static DEFINE_MUTEX(ext_devt_mutex);
36static DEFINE_IDR(ext_devt_idr);
37
Adrian Bunk1826ead2008-03-04 11:23:46 +010038static struct device_type disk_type;
39
Tejun Heoe71bf0d2008-09-03 09:03:02 +020040/**
41 * disk_get_part - get partition
42 * @disk: disk to look partition from
43 * @partno: partition number
44 *
45 * Look for partition @partno from @disk. If found, increment
46 * reference count and return it.
47 *
48 * CONTEXT:
49 * Don't care.
50 *
51 * RETURNS:
52 * Pointer to the found partition on success, NULL if not found.
53 */
54struct hd_struct *disk_get_part(struct gendisk *disk, int partno)
55{
Tejun Heo540eed52008-08-25 19:56:15 +090056 struct hd_struct *part = NULL;
57 struct disk_part_tbl *ptbl;
Tejun Heoe71bf0d2008-09-03 09:03:02 +020058
Tejun Heo540eed52008-08-25 19:56:15 +090059 if (unlikely(partno < 0))
Tejun Heoe71bf0d2008-09-03 09:03:02 +020060 return NULL;
Tejun Heo540eed52008-08-25 19:56:15 +090061
Tejun Heoe71bf0d2008-09-03 09:03:02 +020062 rcu_read_lock();
Tejun Heo540eed52008-08-25 19:56:15 +090063
64 ptbl = rcu_dereference(disk->part_tbl);
65 if (likely(partno < ptbl->len)) {
66 part = rcu_dereference(ptbl->part[partno]);
67 if (part)
68 get_device(part_to_dev(part));
69 }
70
Tejun Heoe71bf0d2008-09-03 09:03:02 +020071 rcu_read_unlock();
72
73 return part;
74}
75EXPORT_SYMBOL_GPL(disk_get_part);
76
77/**
78 * disk_part_iter_init - initialize partition iterator
79 * @piter: iterator to initialize
80 * @disk: disk to iterate over
81 * @flags: DISK_PITER_* flags
82 *
83 * Initialize @piter so that it iterates over partitions of @disk.
84 *
85 * CONTEXT:
86 * Don't care.
87 */
88void disk_part_iter_init(struct disk_part_iter *piter, struct gendisk *disk,
89 unsigned int flags)
90{
Tejun Heo540eed52008-08-25 19:56:15 +090091 struct disk_part_tbl *ptbl;
92
93 rcu_read_lock();
94 ptbl = rcu_dereference(disk->part_tbl);
95
Tejun Heoe71bf0d2008-09-03 09:03:02 +020096 piter->disk = disk;
97 piter->part = NULL;
98
99 if (flags & DISK_PITER_REVERSE)
Tejun Heo540eed52008-08-25 19:56:15 +0900100 piter->idx = ptbl->len - 1;
Tejun Heo71982a42009-04-17 08:34:48 +0200101 else if (flags & (DISK_PITER_INCL_PART0 | DISK_PITER_INCL_EMPTY_PART0))
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200102 piter->idx = 0;
Tejun Heob5d0b9d2008-09-03 09:06:42 +0200103 else
104 piter->idx = 1;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200105
106 piter->flags = flags;
Tejun Heo540eed52008-08-25 19:56:15 +0900107
108 rcu_read_unlock();
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200109}
110EXPORT_SYMBOL_GPL(disk_part_iter_init);
111
112/**
113 * disk_part_iter_next - proceed iterator to the next partition and return it
114 * @piter: iterator of interest
115 *
116 * Proceed @piter to the next partition and return it.
117 *
118 * CONTEXT:
119 * Don't care.
120 */
121struct hd_struct *disk_part_iter_next(struct disk_part_iter *piter)
122{
Tejun Heo540eed52008-08-25 19:56:15 +0900123 struct disk_part_tbl *ptbl;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200124 int inc, end;
125
126 /* put the last partition */
127 disk_put_part(piter->part);
128 piter->part = NULL;
129
Tejun Heo540eed52008-08-25 19:56:15 +0900130 /* get part_tbl */
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200131 rcu_read_lock();
Tejun Heo540eed52008-08-25 19:56:15 +0900132 ptbl = rcu_dereference(piter->disk->part_tbl);
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200133
134 /* determine iteration parameters */
135 if (piter->flags & DISK_PITER_REVERSE) {
136 inc = -1;
Tejun Heo71982a42009-04-17 08:34:48 +0200137 if (piter->flags & (DISK_PITER_INCL_PART0 |
138 DISK_PITER_INCL_EMPTY_PART0))
Tejun Heob5d0b9d2008-09-03 09:06:42 +0200139 end = -1;
140 else
141 end = 0;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200142 } else {
143 inc = 1;
Tejun Heo540eed52008-08-25 19:56:15 +0900144 end = ptbl->len;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200145 }
146
147 /* iterate to the next partition */
148 for (; piter->idx != end; piter->idx += inc) {
149 struct hd_struct *part;
150
Tejun Heo540eed52008-08-25 19:56:15 +0900151 part = rcu_dereference(ptbl->part[piter->idx]);
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200152 if (!part)
153 continue;
Tejun Heo71982a42009-04-17 08:34:48 +0200154 if (!part->nr_sects &&
155 !(piter->flags & DISK_PITER_INCL_EMPTY) &&
156 !(piter->flags & DISK_PITER_INCL_EMPTY_PART0 &&
157 piter->idx == 0))
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200158 continue;
159
Tejun Heoed9e1982008-08-25 19:56:05 +0900160 get_device(part_to_dev(part));
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200161 piter->part = part;
162 piter->idx += inc;
163 break;
164 }
165
166 rcu_read_unlock();
167
168 return piter->part;
169}
170EXPORT_SYMBOL_GPL(disk_part_iter_next);
171
172/**
173 * disk_part_iter_exit - finish up partition iteration
174 * @piter: iter of interest
175 *
176 * Called when iteration is over. Cleans up @piter.
177 *
178 * CONTEXT:
179 * Don't care.
180 */
181void disk_part_iter_exit(struct disk_part_iter *piter)
182{
183 disk_put_part(piter->part);
184 piter->part = NULL;
185}
186EXPORT_SYMBOL_GPL(disk_part_iter_exit);
187
Jens Axboea6f23652008-10-24 12:52:42 +0200188static inline int sector_in_part(struct hd_struct *part, sector_t sector)
189{
190 return part->start_sect <= sector &&
191 sector < part->start_sect + part->nr_sects;
192}
193
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200194/**
195 * disk_map_sector_rcu - map sector to partition
196 * @disk: gendisk of interest
197 * @sector: sector to map
198 *
199 * Find out which partition @sector maps to on @disk. This is
200 * primarily used for stats accounting.
201 *
202 * CONTEXT:
203 * RCU read locked. The returned partition pointer is valid only
204 * while preemption is disabled.
205 *
206 * RETURNS:
Tejun Heo074a7ac2008-08-25 19:56:14 +0900207 * Found partition on success, part0 is returned if no partition matches
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200208 */
209struct hd_struct *disk_map_sector_rcu(struct gendisk *disk, sector_t sector)
210{
Tejun Heo540eed52008-08-25 19:56:15 +0900211 struct disk_part_tbl *ptbl;
Jens Axboea6f23652008-10-24 12:52:42 +0200212 struct hd_struct *part;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200213 int i;
214
Tejun Heo540eed52008-08-25 19:56:15 +0900215 ptbl = rcu_dereference(disk->part_tbl);
216
Jens Axboea6f23652008-10-24 12:52:42 +0200217 part = rcu_dereference(ptbl->last_lookup);
218 if (part && sector_in_part(part, sector))
219 return part;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200220
Jens Axboea6f23652008-10-24 12:52:42 +0200221 for (i = 1; i < ptbl->len; i++) {
222 part = rcu_dereference(ptbl->part[i]);
223
224 if (part && sector_in_part(part, sector)) {
225 rcu_assign_pointer(ptbl->last_lookup, part);
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200226 return part;
Jens Axboea6f23652008-10-24 12:52:42 +0200227 }
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200228 }
Tejun Heo074a7ac2008-08-25 19:56:14 +0900229 return &disk->part0;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200230}
231EXPORT_SYMBOL_GPL(disk_map_sector_rcu);
232
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233/*
234 * Can be deleted altogether. Later.
235 *
236 */
237static struct blk_major_name {
238 struct blk_major_name *next;
239 int major;
240 char name[16];
Joe Korty68eef3b2006-03-31 02:30:32 -0800241} *major_names[BLKDEV_MAJOR_HASH_SIZE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
243/* index in the above - for now: assume no multimajor ranges */
244static inline int major_to_index(int major)
245{
Joe Korty68eef3b2006-03-31 02:30:32 -0800246 return major % BLKDEV_MAJOR_HASH_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247}
248
Joe Korty68eef3b2006-03-31 02:30:32 -0800249#ifdef CONFIG_PROC_FS
Tejun Heocf771cb2008-09-03 09:01:09 +0200250void blkdev_show(struct seq_file *seqf, off_t offset)
Neil Horman7170be52006-01-14 13:20:38 -0800251{
Joe Korty68eef3b2006-03-31 02:30:32 -0800252 struct blk_major_name *dp;
Neil Horman7170be52006-01-14 13:20:38 -0800253
Joe Korty68eef3b2006-03-31 02:30:32 -0800254 if (offset < BLKDEV_MAJOR_HASH_SIZE) {
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200255 mutex_lock(&block_class_lock);
Joe Korty68eef3b2006-03-31 02:30:32 -0800256 for (dp = major_names[offset]; dp; dp = dp->next)
Tejun Heocf771cb2008-09-03 09:01:09 +0200257 seq_printf(seqf, "%3d %s\n", dp->major, dp->name);
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200258 mutex_unlock(&block_class_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260}
Joe Korty68eef3b2006-03-31 02:30:32 -0800261#endif /* CONFIG_PROC_FS */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262
Márton Németh9e8c0bc2009-02-20 08:12:51 +0100263/**
264 * register_blkdev - register a new block device
265 *
266 * @major: the requested major device number [1..255]. If @major=0, try to
267 * allocate any unused major number.
268 * @name: the name of the new block device as a zero terminated string
269 *
270 * The @name must be unique within the system.
271 *
272 * The return value depends on the @major input parameter.
273 * - if a major device number was requested in range [1..255] then the
274 * function returns zero on success, or a negative error code
275 * - if any unused major number was requested with @major=0 parameter
276 * then the return value is the allocated major number in range
277 * [1..255] or a negative error code otherwise
278 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279int register_blkdev(unsigned int major, const char *name)
280{
281 struct blk_major_name **n, *p;
282 int index, ret = 0;
283
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200284 mutex_lock(&block_class_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
286 /* temporary */
287 if (major == 0) {
288 for (index = ARRAY_SIZE(major_names)-1; index > 0; index--) {
289 if (major_names[index] == NULL)
290 break;
291 }
292
293 if (index == 0) {
294 printk("register_blkdev: failed to get major for %s\n",
295 name);
296 ret = -EBUSY;
297 goto out;
298 }
299 major = index;
300 ret = major;
301 }
302
303 p = kmalloc(sizeof(struct blk_major_name), GFP_KERNEL);
304 if (p == NULL) {
305 ret = -ENOMEM;
306 goto out;
307 }
308
309 p->major = major;
310 strlcpy(p->name, name, sizeof(p->name));
311 p->next = NULL;
312 index = major_to_index(major);
313
314 for (n = &major_names[index]; *n; n = &(*n)->next) {
315 if ((*n)->major == major)
316 break;
317 }
318 if (!*n)
319 *n = p;
320 else
321 ret = -EBUSY;
322
323 if (ret < 0) {
324 printk("register_blkdev: cannot get major %d for %s\n",
325 major, name);
326 kfree(p);
327 }
328out:
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200329 mutex_unlock(&block_class_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 return ret;
331}
332
333EXPORT_SYMBOL(register_blkdev);
334
Akinobu Mitaf4480242007-07-17 04:03:47 -0700335void unregister_blkdev(unsigned int major, const char *name)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336{
337 struct blk_major_name **n;
338 struct blk_major_name *p = NULL;
339 int index = major_to_index(major);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200341 mutex_lock(&block_class_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 for (n = &major_names[index]; *n; n = &(*n)->next)
343 if ((*n)->major == major)
344 break;
Akinobu Mita294462a2007-07-17 04:03:45 -0700345 if (!*n || strcmp((*n)->name, name)) {
346 WARN_ON(1);
Akinobu Mita294462a2007-07-17 04:03:45 -0700347 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 p = *n;
349 *n = p->next;
350 }
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200351 mutex_unlock(&block_class_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 kfree(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353}
354
355EXPORT_SYMBOL(unregister_blkdev);
356
357static struct kobj_map *bdev_map;
358
Tejun Heobcce3de2008-08-25 19:47:22 +0900359/**
Tejun Heo870d6652008-08-25 19:47:25 +0900360 * blk_mangle_minor - scatter minor numbers apart
361 * @minor: minor number to mangle
362 *
363 * Scatter consecutively allocated @minor number apart if MANGLE_DEVT
364 * is enabled. Mangling twice gives the original value.
365 *
366 * RETURNS:
367 * Mangled value.
368 *
369 * CONTEXT:
370 * Don't care.
371 */
372static int blk_mangle_minor(int minor)
373{
374#ifdef CONFIG_DEBUG_BLOCK_EXT_DEVT
375 int i;
376
377 for (i = 0; i < MINORBITS / 2; i++) {
378 int low = minor & (1 << i);
379 int high = minor & (1 << (MINORBITS - 1 - i));
380 int distance = MINORBITS - 1 - 2 * i;
381
382 minor ^= low | high; /* clear both bits */
383 low <<= distance; /* swap the positions */
384 high >>= distance;
385 minor |= low | high; /* and set */
386 }
387#endif
388 return minor;
389}
390
391/**
Tejun Heobcce3de2008-08-25 19:47:22 +0900392 * blk_alloc_devt - allocate a dev_t for a partition
393 * @part: partition to allocate dev_t for
Tejun Heobcce3de2008-08-25 19:47:22 +0900394 * @devt: out parameter for resulting dev_t
395 *
396 * Allocate a dev_t for block device.
397 *
398 * RETURNS:
399 * 0 on success, allocated dev_t is returned in *@devt. -errno on
400 * failure.
401 *
402 * CONTEXT:
403 * Might sleep.
404 */
405int blk_alloc_devt(struct hd_struct *part, dev_t *devt)
406{
407 struct gendisk *disk = part_to_disk(part);
408 int idx, rc;
409
410 /* in consecutive minor range? */
411 if (part->partno < disk->minors) {
412 *devt = MKDEV(disk->major, disk->first_minor + part->partno);
413 return 0;
414 }
415
416 /* allocate ext devt */
417 do {
418 if (!idr_pre_get(&ext_devt_idr, GFP_KERNEL))
419 return -ENOMEM;
420 rc = idr_get_new(&ext_devt_idr, part, &idx);
421 } while (rc == -EAGAIN);
422
423 if (rc)
424 return rc;
425
426 if (idx > MAX_EXT_DEVT) {
427 idr_remove(&ext_devt_idr, idx);
428 return -EBUSY;
429 }
430
Tejun Heo870d6652008-08-25 19:47:25 +0900431 *devt = MKDEV(BLOCK_EXT_MAJOR, blk_mangle_minor(idx));
Tejun Heobcce3de2008-08-25 19:47:22 +0900432 return 0;
433}
434
435/**
436 * blk_free_devt - free a dev_t
437 * @devt: dev_t to free
438 *
439 * Free @devt which was allocated using blk_alloc_devt().
440 *
441 * CONTEXT:
442 * Might sleep.
443 */
444void blk_free_devt(dev_t devt)
445{
446 might_sleep();
447
448 if (devt == MKDEV(0, 0))
449 return;
450
451 if (MAJOR(devt) == BLOCK_EXT_MAJOR) {
452 mutex_lock(&ext_devt_mutex);
Tejun Heo870d6652008-08-25 19:47:25 +0900453 idr_remove(&ext_devt_idr, blk_mangle_minor(MINOR(devt)));
Tejun Heobcce3de2008-08-25 19:47:22 +0900454 mutex_unlock(&ext_devt_mutex);
455 }
456}
457
Tejun Heo1f014292008-08-25 19:47:23 +0900458static char *bdevt_str(dev_t devt, char *buf)
459{
460 if (MAJOR(devt) <= 0xff && MINOR(devt) <= 0xff) {
461 char tbuf[BDEVT_SIZE];
462 snprintf(tbuf, BDEVT_SIZE, "%02x%02x", MAJOR(devt), MINOR(devt));
463 snprintf(buf, BDEVT_SIZE, "%-9s", tbuf);
464 } else
465 snprintf(buf, BDEVT_SIZE, "%03x:%05x", MAJOR(devt), MINOR(devt));
466
467 return buf;
468}
469
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470/*
471 * Register device numbers dev..(dev+range-1)
472 * range must be nonzero
473 * The hash chain is sorted on range, so that subranges can override.
474 */
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200475void blk_register_region(dev_t devt, unsigned long range, struct module *module,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 struct kobject *(*probe)(dev_t, int *, void *),
477 int (*lock)(dev_t, void *), void *data)
478{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200479 kobj_map(bdev_map, devt, range, module, probe, lock, data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480}
481
482EXPORT_SYMBOL(blk_register_region);
483
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200484void blk_unregister_region(dev_t devt, unsigned long range)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200486 kobj_unmap(bdev_map, devt, range);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487}
488
489EXPORT_SYMBOL(blk_unregister_region);
490
Tejun Heocf771cb2008-09-03 09:01:09 +0200491static struct kobject *exact_match(dev_t devt, int *partno, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492{
493 struct gendisk *p = data;
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200494
Tejun Heoed9e1982008-08-25 19:56:05 +0900495 return &disk_to_dev(p)->kobj;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496}
497
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200498static int exact_lock(dev_t devt, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499{
500 struct gendisk *p = data;
501
502 if (!get_disk(p))
503 return -1;
504 return 0;
505}
506
507/**
508 * add_disk - add partitioning information to kernel list
509 * @disk: per-device partitioning information
510 *
511 * This function registers the partitioning information in @disk
512 * with the kernel.
Tejun Heo3e1a7ff2008-08-25 19:56:17 +0900513 *
514 * FIXME: error handling
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 */
516void add_disk(struct gendisk *disk)
517{
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700518 struct backing_dev_info *bdi;
Tejun Heo3e1a7ff2008-08-25 19:56:17 +0900519 dev_t devt;
Greg Kroah-Hartman6ffeea72008-05-22 17:21:08 -0400520 int retval;
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700521
Tejun Heo3e1a7ff2008-08-25 19:56:17 +0900522 /* minors == 0 indicates to use ext devt from part0 and should
523 * be accompanied with EXT_DEVT flag. Make sure all
524 * parameters make sense.
525 */
526 WARN_ON(disk->minors && !(disk->major || disk->first_minor));
527 WARN_ON(!disk->minors && !(disk->flags & GENHD_FL_EXT_DEVT));
528
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 disk->flags |= GENHD_FL_UP;
Tejun Heo3e1a7ff2008-08-25 19:56:17 +0900530
531 retval = blk_alloc_devt(&disk->part0, &devt);
532 if (retval) {
533 WARN_ON(1);
534 return;
535 }
536 disk_to_dev(disk)->devt = devt;
537
538 /* ->major and ->first_minor aren't supposed to be
539 * dereferenced from here on, but set them just in case.
540 */
541 disk->major = MAJOR(devt);
542 disk->first_minor = MINOR(devt);
543
Tejun Heof331c022008-09-03 09:01:48 +0200544 blk_register_region(disk_devt(disk), disk->minors, NULL,
545 exact_match, exact_lock, disk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 register_disk(disk);
547 blk_register_queue(disk);
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700548
549 bdi = &disk->queue->backing_dev_info;
Tejun Heof331c022008-09-03 09:01:48 +0200550 bdi_register_dev(bdi, disk_devt(disk));
Tejun Heoed9e1982008-08-25 19:56:05 +0900551 retval = sysfs_create_link(&disk_to_dev(disk)->kobj, &bdi->dev->kobj,
552 "bdi");
Greg Kroah-Hartman6ffeea72008-05-22 17:21:08 -0400553 WARN_ON(retval);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554}
555
556EXPORT_SYMBOL(add_disk);
557EXPORT_SYMBOL(del_gendisk); /* in partitions/check.c */
558
559void unlink_gendisk(struct gendisk *disk)
560{
Tejun Heoed9e1982008-08-25 19:56:05 +0900561 sysfs_remove_link(&disk_to_dev(disk)->kobj, "bdi");
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700562 bdi_unregister(&disk->queue->backing_dev_info);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 blk_unregister_queue(disk);
Tejun Heof331c022008-09-03 09:01:48 +0200564 blk_unregister_region(disk_devt(disk), disk->minors);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565}
566
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567/**
568 * get_gendisk - get partitioning information for a given device
Randy Dunlap710027a2008-08-19 20:13:11 +0200569 * @devt: device to get partitioning information for
Randy Dunlap496aa8a2008-10-16 07:46:23 +0200570 * @partno: returned partition index
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 *
572 * This function gets the structure containing partitioning
Randy Dunlap710027a2008-08-19 20:13:11 +0200573 * information for the given device @devt.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 */
Tejun Heocf771cb2008-09-03 09:01:09 +0200575struct gendisk *get_gendisk(dev_t devt, int *partno)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576{
Tejun Heobcce3de2008-08-25 19:47:22 +0900577 struct gendisk *disk = NULL;
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200578
Tejun Heobcce3de2008-08-25 19:47:22 +0900579 if (MAJOR(devt) != BLOCK_EXT_MAJOR) {
580 struct kobject *kobj;
581
582 kobj = kobj_lookup(bdev_map, devt, partno);
583 if (kobj)
584 disk = dev_to_disk(kobj_to_dev(kobj));
585 } else {
586 struct hd_struct *part;
587
588 mutex_lock(&ext_devt_mutex);
Tejun Heo870d6652008-08-25 19:47:25 +0900589 part = idr_find(&ext_devt_idr, blk_mangle_minor(MINOR(devt)));
Tejun Heobcce3de2008-08-25 19:47:22 +0900590 if (part && get_disk(part_to_disk(part))) {
591 *partno = part->partno;
592 disk = part_to_disk(part);
593 }
594 mutex_unlock(&ext_devt_mutex);
595 }
596
597 return disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598}
599
Tejun Heof331c022008-09-03 09:01:48 +0200600/**
601 * bdget_disk - do bdget() by gendisk and partition number
602 * @disk: gendisk of interest
603 * @partno: partition number
604 *
605 * Find partition @partno from @disk, do bdget() on it.
606 *
607 * CONTEXT:
608 * Don't care.
609 *
610 * RETURNS:
611 * Resulting block_device on success, NULL on failure.
612 */
Harvey Harrisonaeb3d3a2008-08-28 09:27:42 +0200613struct block_device *bdget_disk(struct gendisk *disk, int partno)
Tejun Heof331c022008-09-03 09:01:48 +0200614{
Tejun Heo548b10e2008-08-29 09:01:47 +0200615 struct hd_struct *part;
616 struct block_device *bdev = NULL;
Tejun Heof331c022008-09-03 09:01:48 +0200617
Tejun Heo548b10e2008-08-29 09:01:47 +0200618 part = disk_get_part(disk, partno);
Tejun Heo2bbedcb2008-08-29 11:41:51 +0200619 if (part)
Tejun Heo548b10e2008-08-29 09:01:47 +0200620 bdev = bdget(part_devt(part));
621 disk_put_part(part);
Tejun Heof331c022008-09-03 09:01:48 +0200622
Tejun Heo548b10e2008-08-29 09:01:47 +0200623 return bdev;
Tejun Heof331c022008-09-03 09:01:48 +0200624}
625EXPORT_SYMBOL(bdget_disk);
626
Dave Gilbertdd2a3452007-05-09 02:33:24 -0700627/*
628 * print a full list of all partitions - intended for places where the root
629 * filesystem can't be mounted and thus to give the victim some idea of what
630 * went wrong
631 */
632void __init printk_all_partitions(void)
633{
Tejun Heodef4e382008-09-03 08:57:12 +0200634 struct class_dev_iter iter;
635 struct device *dev;
636
637 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
638 while ((dev = class_dev_iter_next(&iter))) {
639 struct gendisk *disk = dev_to_disk(dev);
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200640 struct disk_part_iter piter;
641 struct hd_struct *part;
Tejun Heo1f014292008-08-25 19:47:23 +0900642 char name_buf[BDEVNAME_SIZE];
643 char devt_buf[BDEVT_SIZE];
Tejun Heodef4e382008-09-03 08:57:12 +0200644
645 /*
646 * Don't show empty devices or things that have been
647 * surpressed
648 */
649 if (get_capacity(disk) == 0 ||
650 (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
651 continue;
652
653 /*
654 * Note, unlike /proc/partitions, I am showing the
655 * numbers in hex - the same format as the root=
656 * option takes.
657 */
Tejun Heo074a7ac2008-08-25 19:56:14 +0900658 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
659 while ((part = disk_part_iter_next(&piter))) {
660 bool is_part0 = part == &disk->part0;
Tejun Heodef4e382008-09-03 08:57:12 +0200661
Tejun Heo074a7ac2008-08-25 19:56:14 +0900662 printk("%s%s %10llu %s", is_part0 ? "" : " ",
Tejun Heo1f014292008-08-25 19:47:23 +0900663 bdevt_str(part_devt(part), devt_buf),
Tejun Heof331c022008-09-03 09:01:48 +0200664 (unsigned long long)part->nr_sects >> 1,
Tejun Heo1f014292008-08-25 19:47:23 +0900665 disk_name(disk, part->partno, name_buf));
Tejun Heo074a7ac2008-08-25 19:56:14 +0900666 if (is_part0) {
667 if (disk->driverfs_dev != NULL &&
668 disk->driverfs_dev->driver != NULL)
669 printk(" driver: %s\n",
670 disk->driverfs_dev->driver->name);
671 else
672 printk(" (driver?)\n");
673 } else
674 printk("\n");
675 }
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200676 disk_part_iter_exit(&piter);
Tejun Heodef4e382008-09-03 08:57:12 +0200677 }
678 class_dev_iter_exit(&iter);
Dave Gilbertdd2a3452007-05-09 02:33:24 -0700679}
680
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681#ifdef CONFIG_PROC_FS
682/* iterator */
Tejun Heodef4e382008-09-03 08:57:12 +0200683static void *disk_seqf_start(struct seq_file *seqf, loff_t *pos)
Greg Kroah-Hartman68c4d4a2008-05-22 17:21:08 -0400684{
Tejun Heodef4e382008-09-03 08:57:12 +0200685 loff_t skip = *pos;
686 struct class_dev_iter *iter;
687 struct device *dev;
Greg Kroah-Hartman68c4d4a2008-05-22 17:21:08 -0400688
Harvey Harrisonaeb3d3a2008-08-28 09:27:42 +0200689 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
Tejun Heodef4e382008-09-03 08:57:12 +0200690 if (!iter)
691 return ERR_PTR(-ENOMEM);
692
693 seqf->private = iter;
694 class_dev_iter_init(iter, &block_class, NULL, &disk_type);
695 do {
696 dev = class_dev_iter_next(iter);
697 if (!dev)
698 return NULL;
699 } while (skip--);
700
701 return dev_to_disk(dev);
Greg Kroah-Hartman68c4d4a2008-05-22 17:21:08 -0400702}
703
Tejun Heodef4e382008-09-03 08:57:12 +0200704static void *disk_seqf_next(struct seq_file *seqf, void *v, loff_t *pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200706 struct device *dev;
Greg Kroah-Hartman66c64af2008-05-22 17:21:08 -0400707
Tejun Heodef4e382008-09-03 08:57:12 +0200708 (*pos)++;
709 dev = class_dev_iter_next(seqf->private);
Tejun Heo2ac3cee2008-09-03 08:53:37 +0200710 if (dev)
Greg Kroah-Hartman68c4d4a2008-05-22 17:21:08 -0400711 return dev_to_disk(dev);
Tejun Heo2ac3cee2008-09-03 08:53:37 +0200712
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 return NULL;
714}
715
Tejun Heodef4e382008-09-03 08:57:12 +0200716static void disk_seqf_stop(struct seq_file *seqf, void *v)
Greg Kroah-Hartman27f30252008-05-22 17:21:08 -0400717{
Tejun Heodef4e382008-09-03 08:57:12 +0200718 struct class_dev_iter *iter = seqf->private;
Greg Kroah-Hartman27f30252008-05-22 17:21:08 -0400719
Tejun Heodef4e382008-09-03 08:57:12 +0200720 /* stop is called even after start failed :-( */
721 if (iter) {
722 class_dev_iter_exit(iter);
723 kfree(iter);
Kay Sievers5c0ef6d2008-08-16 14:30:30 +0200724 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725}
726
Tejun Heodef4e382008-09-03 08:57:12 +0200727static void *show_partition_start(struct seq_file *seqf, loff_t *pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728{
Tejun Heodef4e382008-09-03 08:57:12 +0200729 static void *p;
730
731 p = disk_seqf_start(seqf, pos);
Tejun Heo243294d2008-09-04 09:17:31 +0200732 if (!IS_ERR(p) && p && !*pos)
Tejun Heodef4e382008-09-03 08:57:12 +0200733 seq_puts(seqf, "major minor #blocks name\n\n");
734 return p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735}
736
Tejun Heocf771cb2008-09-03 09:01:09 +0200737static int show_partition(struct seq_file *seqf, void *v)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738{
739 struct gendisk *sgp = v;
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200740 struct disk_part_iter piter;
741 struct hd_struct *part;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 char buf[BDEVNAME_SIZE];
743
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 /* Don't show non-partitionable removeable devices or empty devices */
Tejun Heob5d0b9d2008-09-03 09:06:42 +0200745 if (!get_capacity(sgp) || (!disk_partitionable(sgp) &&
Tejun Heof331c022008-09-03 09:01:48 +0200746 (sgp->flags & GENHD_FL_REMOVABLE)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 return 0;
748 if (sgp->flags & GENHD_FL_SUPPRESS_PARTITION_INFO)
749 return 0;
750
751 /* show the full disk and all non-0 size partitions of it */
Tejun Heo074a7ac2008-08-25 19:56:14 +0900752 disk_part_iter_init(&piter, sgp, DISK_PITER_INCL_PART0);
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200753 while ((part = disk_part_iter_next(&piter)))
Tejun Heo1f014292008-08-25 19:47:23 +0900754 seq_printf(seqf, "%4d %7d %10llu %s\n",
Tejun Heof331c022008-09-03 09:01:48 +0200755 MAJOR(part_devt(part)), MINOR(part_devt(part)),
756 (unsigned long long)part->nr_sects >> 1,
757 disk_name(sgp, part->partno, buf));
Tejun Heoe71bf0d2008-09-03 09:03:02 +0200758 disk_part_iter_exit(&piter);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
760 return 0;
761}
762
Alexey Dobriyanf5009752008-10-04 23:53:21 +0400763static const struct seq_operations partitions_op = {
Tejun Heodef4e382008-09-03 08:57:12 +0200764 .start = show_partition_start,
765 .next = disk_seqf_next,
766 .stop = disk_seqf_stop,
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200767 .show = show_partition
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768};
Alexey Dobriyanf5009752008-10-04 23:53:21 +0400769
770static int partitions_open(struct inode *inode, struct file *file)
771{
772 return seq_open(file, &partitions_op);
773}
774
775static const struct file_operations proc_partitions_operations = {
776 .open = partitions_open,
777 .read = seq_read,
778 .llseek = seq_lseek,
779 .release = seq_release,
780};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781#endif
782
783
Tejun Heocf771cb2008-09-03 09:01:09 +0200784static struct kobject *base_probe(dev_t devt, int *partno, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200786 if (request_module("block-major-%d-%d", MAJOR(devt), MINOR(devt)) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 /* Make old-style 2.4 aliases work */
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200788 request_module("block-major-%d", MAJOR(devt));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 return NULL;
790}
791
792static int __init genhd_device_init(void)
793{
Dan Williamse105b8b2008-04-21 10:51:07 -0700794 int error;
795
796 block_class.dev_kobj = sysfs_dev_block_kobj;
797 error = class_register(&block_class);
Roland McGrathee27a552008-03-11 17:13:15 -0700798 if (unlikely(error))
799 return error;
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200800 bdev_map = kobj_map_init(base_probe, &block_class_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 blk_dev_init();
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200802
Zhang, Yanmin561ec682008-11-14 08:26:30 +0100803 register_blkdev(BLOCK_EXT_MAJOR, "blkext");
804
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200805#ifndef CONFIG_SYSFS_DEPRECATED
806 /* create top-level block dir */
807 block_depr = kobject_create_and_add("block", NULL);
808#endif
Greg Kroah-Hartman830d3cf2007-11-06 10:36:58 -0800809 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810}
811
812subsys_initcall(genhd_device_init);
813
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200814static ssize_t disk_range_show(struct device *dev,
815 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200817 struct gendisk *disk = dev_to_disk(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200819 return sprintf(buf, "%d\n", disk->minors);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820}
821
Tejun Heo1f014292008-08-25 19:47:23 +0900822static ssize_t disk_ext_range_show(struct device *dev,
823 struct device_attribute *attr, char *buf)
824{
825 struct gendisk *disk = dev_to_disk(dev);
826
Tejun Heob5d0b9d2008-09-03 09:06:42 +0200827 return sprintf(buf, "%d\n", disk_max_parts(disk));
Tejun Heo1f014292008-08-25 19:47:23 +0900828}
829
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200830static ssize_t disk_removable_show(struct device *dev,
831 struct device_attribute *attr, char *buf)
Kay Sieversa7fd6702005-10-01 14:49:43 +0200832{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200833 struct gendisk *disk = dev_to_disk(dev);
Kay Sieversa7fd6702005-10-01 14:49:43 +0200834
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200835 return sprintf(buf, "%d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 (disk->flags & GENHD_FL_REMOVABLE ? 1 : 0));
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200837}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838
Kay Sievers1c9ce522008-06-13 09:41:00 +0200839static ssize_t disk_ro_show(struct device *dev,
840 struct device_attribute *attr, char *buf)
841{
842 struct gendisk *disk = dev_to_disk(dev);
843
Tejun Heob7db9952008-08-25 19:56:10 +0900844 return sprintf(buf, "%d\n", get_disk_ro(disk) ? 1 : 0);
Kay Sievers1c9ce522008-06-13 09:41:00 +0200845}
846
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200847static ssize_t disk_capability_show(struct device *dev,
848 struct device_attribute *attr, char *buf)
Kristen Carlson Accardi86ce18d2007-05-23 13:57:38 -0700849{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200850 struct gendisk *disk = dev_to_disk(dev);
851
852 return sprintf(buf, "%x\n", disk->flags);
Kristen Carlson Accardi86ce18d2007-05-23 13:57:38 -0700853}
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200854
Martin K. Petersenc72758f2009-05-22 17:17:53 -0400855static ssize_t disk_alignment_offset_show(struct device *dev,
856 struct device_attribute *attr,
857 char *buf)
858{
859 struct gendisk *disk = dev_to_disk(dev);
860
861 return sprintf(buf, "%d\n", queue_alignment_offset(disk->queue));
862}
863
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200864static DEVICE_ATTR(range, S_IRUGO, disk_range_show, NULL);
Tejun Heo1f014292008-08-25 19:47:23 +0900865static DEVICE_ATTR(ext_range, S_IRUGO, disk_ext_range_show, NULL);
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200866static DEVICE_ATTR(removable, S_IRUGO, disk_removable_show, NULL);
Kay Sievers1c9ce522008-06-13 09:41:00 +0200867static DEVICE_ATTR(ro, S_IRUGO, disk_ro_show, NULL);
Tejun Heoe5610522008-08-25 19:56:09 +0900868static DEVICE_ATTR(size, S_IRUGO, part_size_show, NULL);
Martin K. Petersenc72758f2009-05-22 17:17:53 -0400869static DEVICE_ATTR(alignment_offset, S_IRUGO, disk_alignment_offset_show, NULL);
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200870static DEVICE_ATTR(capability, S_IRUGO, disk_capability_show, NULL);
Tejun Heo074a7ac2008-08-25 19:56:14 +0900871static DEVICE_ATTR(stat, S_IRUGO, part_stat_show, NULL);
Nikanth Karthikesana9327ca2009-09-11 09:18:54 +0200872static DEVICE_ATTR(inflight, S_IRUGO, part_inflight_show, NULL);
Akinobu Mitac17bb492006-12-08 02:39:46 -0800873#ifdef CONFIG_FAIL_MAKE_REQUEST
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200874static struct device_attribute dev_attr_fail =
Tejun Heoeddb2e22008-08-25 19:56:13 +0900875 __ATTR(make-it-fail, S_IRUGO|S_IWUSR, part_fail_show, part_fail_store);
Akinobu Mitac17bb492006-12-08 02:39:46 -0800876#endif
Jens Axboe581d4e22008-09-14 05:56:33 -0700877#ifdef CONFIG_FAIL_IO_TIMEOUT
878static struct device_attribute dev_attr_fail_timeout =
879 __ATTR(io-timeout-fail, S_IRUGO|S_IWUSR, part_timeout_show,
880 part_timeout_store);
881#endif
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200882
883static struct attribute *disk_attrs[] = {
884 &dev_attr_range.attr,
Tejun Heo1f014292008-08-25 19:47:23 +0900885 &dev_attr_ext_range.attr,
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200886 &dev_attr_removable.attr,
Kay Sievers1c9ce522008-06-13 09:41:00 +0200887 &dev_attr_ro.attr,
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200888 &dev_attr_size.attr,
Martin K. Petersenc72758f2009-05-22 17:17:53 -0400889 &dev_attr_alignment_offset.attr,
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200890 &dev_attr_capability.attr,
891 &dev_attr_stat.attr,
Nikanth Karthikesana9327ca2009-09-11 09:18:54 +0200892 &dev_attr_inflight.attr,
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200893#ifdef CONFIG_FAIL_MAKE_REQUEST
894 &dev_attr_fail.attr,
895#endif
Jens Axboe581d4e22008-09-14 05:56:33 -0700896#ifdef CONFIG_FAIL_IO_TIMEOUT
897 &dev_attr_fail_timeout.attr,
898#endif
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200899 NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900};
901
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200902static struct attribute_group disk_attr_group = {
903 .attrs = disk_attrs,
904};
905
David Brownella4dbd672009-06-24 10:06:31 -0700906static const struct attribute_group *disk_attr_groups[] = {
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200907 &disk_attr_group,
908 NULL
909};
910
Tejun Heo540eed52008-08-25 19:56:15 +0900911static void disk_free_ptbl_rcu_cb(struct rcu_head *head)
912{
913 struct disk_part_tbl *ptbl =
914 container_of(head, struct disk_part_tbl, rcu_head);
915
916 kfree(ptbl);
917}
918
919/**
920 * disk_replace_part_tbl - replace disk->part_tbl in RCU-safe way
921 * @disk: disk to replace part_tbl for
922 * @new_ptbl: new part_tbl to install
923 *
924 * Replace disk->part_tbl with @new_ptbl in RCU-safe way. The
925 * original ptbl is freed using RCU callback.
926 *
927 * LOCKING:
928 * Matching bd_mutx locked.
929 */
930static void disk_replace_part_tbl(struct gendisk *disk,
931 struct disk_part_tbl *new_ptbl)
932{
933 struct disk_part_tbl *old_ptbl = disk->part_tbl;
934
935 rcu_assign_pointer(disk->part_tbl, new_ptbl);
Jens Axboea6f23652008-10-24 12:52:42 +0200936
937 if (old_ptbl) {
938 rcu_assign_pointer(old_ptbl->last_lookup, NULL);
Tejun Heo540eed52008-08-25 19:56:15 +0900939 call_rcu(&old_ptbl->rcu_head, disk_free_ptbl_rcu_cb);
Jens Axboea6f23652008-10-24 12:52:42 +0200940 }
Tejun Heo540eed52008-08-25 19:56:15 +0900941}
942
943/**
944 * disk_expand_part_tbl - expand disk->part_tbl
945 * @disk: disk to expand part_tbl for
946 * @partno: expand such that this partno can fit in
947 *
948 * Expand disk->part_tbl such that @partno can fit in. disk->part_tbl
949 * uses RCU to allow unlocked dereferencing for stats and other stuff.
950 *
951 * LOCKING:
952 * Matching bd_mutex locked, might sleep.
953 *
954 * RETURNS:
955 * 0 on success, -errno on failure.
956 */
957int disk_expand_part_tbl(struct gendisk *disk, int partno)
958{
959 struct disk_part_tbl *old_ptbl = disk->part_tbl;
960 struct disk_part_tbl *new_ptbl;
961 int len = old_ptbl ? old_ptbl->len : 0;
962 int target = partno + 1;
963 size_t size;
964 int i;
965
966 /* disk_max_parts() is zero during initialization, ignore if so */
967 if (disk_max_parts(disk) && target > disk_max_parts(disk))
968 return -EINVAL;
969
970 if (target <= len)
971 return 0;
972
973 size = sizeof(*new_ptbl) + target * sizeof(new_ptbl->part[0]);
974 new_ptbl = kzalloc_node(size, GFP_KERNEL, disk->node_id);
975 if (!new_ptbl)
976 return -ENOMEM;
977
978 INIT_RCU_HEAD(&new_ptbl->rcu_head);
979 new_ptbl->len = target;
980
981 for (i = 0; i < len; i++)
982 rcu_assign_pointer(new_ptbl->part[i], old_ptbl->part[i]);
983
984 disk_replace_part_tbl(disk, new_ptbl);
985 return 0;
986}
987
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200988static void disk_release(struct device *dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989{
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200990 struct gendisk *disk = dev_to_disk(dev);
991
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 kfree(disk->random);
Tejun Heo540eed52008-08-25 19:56:15 +0900993 disk_replace_part_tbl(disk, NULL);
Tejun Heo074a7ac2008-08-25 19:56:14 +0900994 free_part_stats(&disk->part0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 kfree(disk);
996}
Kay Sieversedfaa7c2007-05-21 22:08:01 +0200997struct class block_class = {
998 .name = "block",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999};
1000
Kay Sieversb03f38b2009-04-30 15:23:42 +02001001static char *block_nodename(struct device *dev)
1002{
1003 struct gendisk *disk = dev_to_disk(dev);
1004
1005 if (disk->nodename)
1006 return disk->nodename(disk);
1007 return NULL;
1008}
1009
Adrian Bunk1826ead2008-03-04 11:23:46 +01001010static struct device_type disk_type = {
Kay Sieversedfaa7c2007-05-21 22:08:01 +02001011 .name = "disk",
1012 .groups = disk_attr_groups,
1013 .release = disk_release,
Kay Sieversb03f38b2009-04-30 15:23:42 +02001014 .nodename = block_nodename,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015};
1016
Randy Dunlapa6e2ba82008-05-23 09:44:11 -07001017#ifdef CONFIG_PROC_FS
Tejun Heocf771cb2008-09-03 09:01:09 +02001018/*
1019 * aggregate disk stat collector. Uses the same stats that the sysfs
1020 * entries do, above, but makes them available through one seq_file.
1021 *
1022 * The output looks suspiciously like /proc/partitions with a bunch of
1023 * extra fields.
1024 */
1025static int diskstats_show(struct seq_file *seqf, void *v)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026{
1027 struct gendisk *gp = v;
Tejun Heoe71bf0d2008-09-03 09:03:02 +02001028 struct disk_part_iter piter;
1029 struct hd_struct *hd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 char buf[BDEVNAME_SIZE];
Tejun Heoc9959052008-08-25 19:47:21 +09001031 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032
1033 /*
Tejun Heoed9e1982008-08-25 19:56:05 +09001034 if (&disk_to_dev(gp)->kobj.entry == block_class.devices.next)
Tejun Heocf771cb2008-09-03 09:01:09 +02001035 seq_puts(seqf, "major minor name"
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 " rio rmerge rsect ruse wio wmerge "
1037 "wsect wuse running use aveq"
1038 "\n\n");
1039 */
1040
Tejun Heo71982a42009-04-17 08:34:48 +02001041 disk_part_iter_init(&piter, gp, DISK_PITER_INCL_EMPTY_PART0);
Tejun Heoe71bf0d2008-09-03 09:03:02 +02001042 while ((hd = disk_part_iter_next(&piter))) {
Tejun Heo074a7ac2008-08-25 19:56:14 +09001043 cpu = part_stat_lock();
Tejun Heoc9959052008-08-25 19:47:21 +09001044 part_round_stats(cpu, hd);
Tejun Heo074a7ac2008-08-25 19:56:14 +09001045 part_stat_unlock();
Tejun Heo1f014292008-08-25 19:47:23 +09001046 seq_printf(seqf, "%4d %7d %s %lu %lu %llu "
Jerome Marchand28f39d52008-02-08 11:04:56 +01001047 "%u %lu %lu %llu %u %u %u %u\n",
Tejun Heof331c022008-09-03 09:01:48 +02001048 MAJOR(part_devt(hd)), MINOR(part_devt(hd)),
1049 disk_name(gp, hd->partno, buf),
Jerome Marchand28f39d52008-02-08 11:04:56 +01001050 part_stat_read(hd, ios[0]),
1051 part_stat_read(hd, merges[0]),
1052 (unsigned long long)part_stat_read(hd, sectors[0]),
1053 jiffies_to_msecs(part_stat_read(hd, ticks[0])),
1054 part_stat_read(hd, ios[1]),
1055 part_stat_read(hd, merges[1]),
1056 (unsigned long long)part_stat_read(hd, sectors[1]),
1057 jiffies_to_msecs(part_stat_read(hd, ticks[1])),
Nikanth Karthikesana9327ca2009-09-11 09:18:54 +02001058 part_in_flight(hd),
Jerome Marchand28f39d52008-02-08 11:04:56 +01001059 jiffies_to_msecs(part_stat_read(hd, io_ticks)),
1060 jiffies_to_msecs(part_stat_read(hd, time_in_queue))
1061 );
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 }
Tejun Heoe71bf0d2008-09-03 09:03:02 +02001063 disk_part_iter_exit(&piter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064
1065 return 0;
1066}
1067
Alexey Dobriyan31d85ab2008-10-06 12:55:38 +04001068static const struct seq_operations diskstats_op = {
Tejun Heodef4e382008-09-03 08:57:12 +02001069 .start = disk_seqf_start,
1070 .next = disk_seqf_next,
1071 .stop = disk_seqf_stop,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 .show = diskstats_show
1073};
Alexey Dobriyanf5009752008-10-04 23:53:21 +04001074
Alexey Dobriyan31d85ab2008-10-06 12:55:38 +04001075static int diskstats_open(struct inode *inode, struct file *file)
1076{
1077 return seq_open(file, &diskstats_op);
1078}
1079
1080static const struct file_operations proc_diskstats_operations = {
1081 .open = diskstats_open,
1082 .read = seq_read,
1083 .llseek = seq_lseek,
1084 .release = seq_release,
1085};
1086
Alexey Dobriyanf5009752008-10-04 23:53:21 +04001087static int __init proc_genhd_init(void)
1088{
Alexey Dobriyan31d85ab2008-10-06 12:55:38 +04001089 proc_create("diskstats", 0, NULL, &proc_diskstats_operations);
Alexey Dobriyanf5009752008-10-04 23:53:21 +04001090 proc_create("partitions", 0, NULL, &proc_partitions_operations);
1091 return 0;
1092}
1093module_init(proc_genhd_init);
Randy Dunlapa6e2ba82008-05-23 09:44:11 -07001094#endif /* CONFIG_PROC_FS */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095
Kristen Carlson Accardi8ce7ad72007-05-23 13:57:38 -07001096static void media_change_notify_thread(struct work_struct *work)
1097{
1098 struct gendisk *gd = container_of(work, struct gendisk, async_notify);
1099 char event[] = "MEDIA_CHANGE=1";
1100 char *envp[] = { event, NULL };
1101
1102 /*
1103 * set enviroment vars to indicate which event this is for
1104 * so that user space will know to go check the media status.
1105 */
Tejun Heoed9e1982008-08-25 19:56:05 +09001106 kobject_uevent_env(&disk_to_dev(gd)->kobj, KOBJ_CHANGE, envp);
Kristen Carlson Accardi8ce7ad72007-05-23 13:57:38 -07001107 put_device(gd->driverfs_dev);
1108}
1109
Adrian Bunk1826ead2008-03-04 11:23:46 +01001110#if 0
Kristen Carlson Accardi8ce7ad72007-05-23 13:57:38 -07001111void genhd_media_change_notify(struct gendisk *disk)
1112{
1113 get_device(disk->driverfs_dev);
1114 schedule_work(&disk->async_notify);
1115}
1116EXPORT_SYMBOL_GPL(genhd_media_change_notify);
Adrian Bunk1826ead2008-03-04 11:23:46 +01001117#endif /* 0 */
Kristen Carlson Accardi8ce7ad72007-05-23 13:57:38 -07001118
Tejun Heocf771cb2008-09-03 09:01:09 +02001119dev_t blk_lookup_devt(const char *name, int partno)
Kay Sieversedfaa7c2007-05-21 22:08:01 +02001120{
Kay Sieversedfaa7c2007-05-21 22:08:01 +02001121 dev_t devt = MKDEV(0, 0);
Tejun Heodef4e382008-09-03 08:57:12 +02001122 struct class_dev_iter iter;
1123 struct device *dev;
Kay Sieversedfaa7c2007-05-21 22:08:01 +02001124
Tejun Heodef4e382008-09-03 08:57:12 +02001125 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
1126 while ((dev = class_dev_iter_next(&iter))) {
1127 struct gendisk *disk = dev_to_disk(dev);
Tejun Heo548b10e2008-08-29 09:01:47 +02001128 struct hd_struct *part;
Tejun Heodef4e382008-09-03 08:57:12 +02001129
Kay Sievers3ada8b72009-01-06 10:44:43 -08001130 if (strcmp(dev_name(dev), name))
Tejun Heof331c022008-09-03 09:01:48 +02001131 continue;
Tejun Heof331c022008-09-03 09:01:48 +02001132
Neil Brown41b8c852009-02-18 10:33:59 +01001133 if (partno < disk->minors) {
1134 /* We need to return the right devno, even
1135 * if the partition doesn't exist yet.
1136 */
1137 devt = MKDEV(MAJOR(dev->devt),
1138 MINOR(dev->devt) + partno);
1139 break;
1140 }
Tejun Heo548b10e2008-08-29 09:01:47 +02001141 part = disk_get_part(disk, partno);
Tejun Heo2bbedcb2008-08-29 11:41:51 +02001142 if (part) {
Tejun Heof331c022008-09-03 09:01:48 +02001143 devt = part_devt(part);
Tejun Heoe71bf0d2008-09-03 09:03:02 +02001144 disk_put_part(part);
Tejun Heo548b10e2008-08-29 09:01:47 +02001145 break;
Tejun Heodef4e382008-09-03 08:57:12 +02001146 }
Tejun Heo548b10e2008-08-29 09:01:47 +02001147 disk_put_part(part);
Kay Sievers5c0ef6d2008-08-16 14:30:30 +02001148 }
Tejun Heodef4e382008-09-03 08:57:12 +02001149 class_dev_iter_exit(&iter);
Kay Sieversedfaa7c2007-05-21 22:08:01 +02001150 return devt;
1151}
Kay Sieversedfaa7c2007-05-21 22:08:01 +02001152EXPORT_SYMBOL(blk_lookup_devt);
1153
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154struct gendisk *alloc_disk(int minors)
1155{
Christoph Lameter19460892005-06-23 00:08:19 -07001156 return alloc_disk_node(minors, -1);
1157}
Tejun Heo689d6fa2008-08-25 19:56:16 +09001158EXPORT_SYMBOL(alloc_disk);
Christoph Lameter19460892005-06-23 00:08:19 -07001159
1160struct gendisk *alloc_disk_node(int minors, int node_id)
1161{
1162 struct gendisk *disk;
1163
Christoph Lameter94f60302007-07-17 04:03:29 -07001164 disk = kmalloc_node(sizeof(struct gendisk),
1165 GFP_KERNEL | __GFP_ZERO, node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 if (disk) {
Tejun Heo074a7ac2008-08-25 19:56:14 +09001167 if (!init_part_stats(&disk->part0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 kfree(disk);
1169 return NULL;
1170 }
Cheng Renquanbf91db12008-11-20 08:37:37 +01001171 disk->node_id = node_id;
Tejun Heo540eed52008-08-25 19:56:15 +09001172 if (disk_expand_part_tbl(disk, 0)) {
1173 free_part_stats(&disk->part0);
Tejun Heob5d0b9d2008-09-03 09:06:42 +02001174 kfree(disk);
1175 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 }
Tejun Heo540eed52008-08-25 19:56:15 +09001177 disk->part_tbl->part[0] = &disk->part0;
Tejun Heob5d0b9d2008-09-03 09:06:42 +02001178
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 disk->minors = minors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 rand_initialize_disk(disk);
Tejun Heoed9e1982008-08-25 19:56:05 +09001181 disk_to_dev(disk)->class = &block_class;
1182 disk_to_dev(disk)->type = &disk_type;
1183 device_initialize(disk_to_dev(disk));
Kristen Carlson Accardi8ce7ad72007-05-23 13:57:38 -07001184 INIT_WORK(&disk->async_notify,
1185 media_change_notify_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 }
1187 return disk;
1188}
Christoph Lameter19460892005-06-23 00:08:19 -07001189EXPORT_SYMBOL(alloc_disk_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190
1191struct kobject *get_disk(struct gendisk *disk)
1192{
1193 struct module *owner;
1194 struct kobject *kobj;
1195
1196 if (!disk->fops)
1197 return NULL;
1198 owner = disk->fops->owner;
1199 if (owner && !try_module_get(owner))
1200 return NULL;
Tejun Heoed9e1982008-08-25 19:56:05 +09001201 kobj = kobject_get(&disk_to_dev(disk)->kobj);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 if (kobj == NULL) {
1203 module_put(owner);
1204 return NULL;
1205 }
1206 return kobj;
1207
1208}
1209
1210EXPORT_SYMBOL(get_disk);
1211
1212void put_disk(struct gendisk *disk)
1213{
1214 if (disk)
Tejun Heoed9e1982008-08-25 19:56:05 +09001215 kobject_put(&disk_to_dev(disk)->kobj);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216}
1217
1218EXPORT_SYMBOL(put_disk);
1219
Hannes Reineckee3264a42009-07-28 09:13:13 +02001220static void set_disk_ro_uevent(struct gendisk *gd, int ro)
1221{
1222 char event[] = "DISK_RO=1";
1223 char *envp[] = { event, NULL };
1224
1225 if (!ro)
1226 event[8] = '0';
1227 kobject_uevent_env(&disk_to_dev(gd)->kobj, KOBJ_CHANGE, envp);
1228}
1229
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230void set_device_ro(struct block_device *bdev, int flag)
1231{
Tejun Heob7db9952008-08-25 19:56:10 +09001232 bdev->bd_part->policy = flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233}
1234
1235EXPORT_SYMBOL(set_device_ro);
1236
1237void set_disk_ro(struct gendisk *disk, int flag)
1238{
Tejun Heoe71bf0d2008-09-03 09:03:02 +02001239 struct disk_part_iter piter;
1240 struct hd_struct *part;
1241
Hannes Reineckee3264a42009-07-28 09:13:13 +02001242 if (disk->part0.policy != flag) {
1243 set_disk_ro_uevent(disk, flag);
1244 disk->part0.policy = flag;
1245 }
1246
1247 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_EMPTY);
Tejun Heoe71bf0d2008-09-03 09:03:02 +02001248 while ((part = disk_part_iter_next(&piter)))
1249 part->policy = flag;
1250 disk_part_iter_exit(&piter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251}
1252
1253EXPORT_SYMBOL(set_disk_ro);
1254
1255int bdev_read_only(struct block_device *bdev)
1256{
1257 if (!bdev)
1258 return 0;
Tejun Heob7db9952008-08-25 19:56:10 +09001259 return bdev->bd_part->policy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260}
1261
1262EXPORT_SYMBOL(bdev_read_only);
1263
Tejun Heocf771cb2008-09-03 09:01:09 +02001264int invalidate_partition(struct gendisk *disk, int partno)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265{
1266 int res = 0;
Tejun Heocf771cb2008-09-03 09:01:09 +02001267 struct block_device *bdev = bdget_disk(disk, partno);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 if (bdev) {
Christoph Hellwig2ef41632005-05-05 16:15:59 -07001269 fsync_bdev(bdev);
1270 res = __invalidate_device(bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 bdput(bdev);
1272 }
1273 return res;
1274}
1275
1276EXPORT_SYMBOL(invalidate_partition);