blob: f1bf79d9bc0a1c65df988ef4ec0b3c1eab907d93 [file] [log] [blame]
Nick Piggin9db55792008-02-08 04:19:49 -08001/*
2 * Ram backed block device driver.
3 *
4 * Copyright (C) 2007 Nick Piggin
5 * Copyright (C) 2007 Novell Inc.
6 *
7 * Parts derived from drivers/block/rd.c, and drivers/block/loop.c, copyright
8 * of their respective owners.
9 */
10
11#include <linux/init.h>
12#include <linux/module.h>
13#include <linux/moduleparam.h>
14#include <linux/major.h>
15#include <linux/blkdev.h>
16#include <linux/bio.h>
17#include <linux/highmem.h>
Nick Piggin9db55792008-02-08 04:19:49 -080018#include <linux/radix-tree.h>
19#include <linux/buffer_head.h> /* invalidate_bh_lrus() */
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Nick Piggin9db55792008-02-08 04:19:49 -080021
22#include <asm/uaccess.h>
23
24#define SECTOR_SHIFT 9
25#define PAGE_SECTORS_SHIFT (PAGE_SHIFT - SECTOR_SHIFT)
26#define PAGE_SECTORS (1 << PAGE_SECTORS_SHIFT)
27
28/*
29 * Each block ramdisk device has a radix_tree brd_pages of pages that stores
30 * the pages containing the block device's contents. A brd page's ->index is
31 * its offset in PAGE_SIZE units. This is similar to, but in no way connected
32 * with, the kernel's pagecache or buffer cache (which sit above our block
33 * device).
34 */
35struct brd_device {
36 int brd_number;
37 int brd_refcnt;
38 loff_t brd_offset;
39 loff_t brd_sizelimit;
40 unsigned brd_blocksize;
41
42 struct request_queue *brd_queue;
43 struct gendisk *brd_disk;
44 struct list_head brd_list;
45
46 /*
47 * Backing store of pages and lock to protect it. This is the contents
48 * of the block device.
49 */
50 spinlock_t brd_lock;
51 struct radix_tree_root brd_pages;
52};
53
54/*
55 * Look up and return a brd's page for a given sector.
56 */
57static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
58{
59 pgoff_t idx;
60 struct page *page;
61
62 /*
63 * The page lifetime is protected by the fact that we have opened the
64 * device node -- brd pages will never be deleted under us, so we
65 * don't need any further locking or refcounting.
66 *
67 * This is strictly true for the radix-tree nodes as well (ie. we
68 * don't actually need the rcu_read_lock()), however that is not a
69 * documented feature of the radix-tree API so it is better to be
70 * safe here (we don't have total exclusion from radix tree updates
71 * here, only deletes).
72 */
73 rcu_read_lock();
74 idx = sector >> PAGE_SECTORS_SHIFT; /* sector to page index */
75 page = radix_tree_lookup(&brd->brd_pages, idx);
76 rcu_read_unlock();
77
78 BUG_ON(page && page->index != idx);
79
80 return page;
81}
82
83/*
84 * Look up and return a brd's page for a given sector.
85 * If one does not exist, allocate an empty page, and insert that. Then
86 * return it.
87 */
88static struct page *brd_insert_page(struct brd_device *brd, sector_t sector)
89{
90 pgoff_t idx;
91 struct page *page;
Nick Piggin75acb9c2008-02-08 04:19:50 -080092 gfp_t gfp_flags;
Nick Piggin9db55792008-02-08 04:19:49 -080093
94 page = brd_lookup_page(brd, sector);
95 if (page)
96 return page;
97
98 /*
99 * Must use NOIO because we don't want to recurse back into the
100 * block or filesystem layers from page reclaim.
Nick Piggin75acb9c2008-02-08 04:19:50 -0800101 *
102 * Cannot support XIP and highmem, because our ->direct_access
103 * routine for XIP must return memory that is always addressable.
104 * If XIP was reworked to use pfns and kmap throughout, this
105 * restriction might be able to be lifted.
Nick Piggin9db55792008-02-08 04:19:49 -0800106 */
Nick Piggin75acb9c2008-02-08 04:19:50 -0800107 gfp_flags = GFP_NOIO | __GFP_ZERO;
108#ifndef CONFIG_BLK_DEV_XIP
109 gfp_flags |= __GFP_HIGHMEM;
110#endif
Petr Tesarik26defe32008-04-22 05:36:52 +0200111 page = alloc_page(gfp_flags);
Nick Piggin9db55792008-02-08 04:19:49 -0800112 if (!page)
113 return NULL;
114
115 if (radix_tree_preload(GFP_NOIO)) {
116 __free_page(page);
117 return NULL;
118 }
119
120 spin_lock(&brd->brd_lock);
121 idx = sector >> PAGE_SECTORS_SHIFT;
122 if (radix_tree_insert(&brd->brd_pages, idx, page)) {
123 __free_page(page);
124 page = radix_tree_lookup(&brd->brd_pages, idx);
125 BUG_ON(!page);
126 BUG_ON(page->index != idx);
127 } else
128 page->index = idx;
129 spin_unlock(&brd->brd_lock);
130
131 radix_tree_preload_end();
132
133 return page;
134}
135
Nick Pigginb7c33572010-05-26 15:41:14 +0200136static void brd_free_page(struct brd_device *brd, sector_t sector)
137{
138 struct page *page;
139 pgoff_t idx;
140
141 spin_lock(&brd->brd_lock);
142 idx = sector >> PAGE_SECTORS_SHIFT;
143 page = radix_tree_delete(&brd->brd_pages, idx);
144 spin_unlock(&brd->brd_lock);
145 if (page)
146 __free_page(page);
147}
148
149static void brd_zero_page(struct brd_device *brd, sector_t sector)
150{
151 struct page *page;
152
153 page = brd_lookup_page(brd, sector);
154 if (page)
155 clear_highpage(page);
156}
157
Nick Piggin9db55792008-02-08 04:19:49 -0800158/*
159 * Free all backing store pages and radix tree. This must only be called when
160 * there are no other users of the device.
161 */
162#define FREE_BATCH 16
163static void brd_free_pages(struct brd_device *brd)
164{
165 unsigned long pos = 0;
166 struct page *pages[FREE_BATCH];
167 int nr_pages;
168
169 do {
170 int i;
171
172 nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
173 (void **)pages, pos, FREE_BATCH);
174
175 for (i = 0; i < nr_pages; i++) {
176 void *ret;
177
178 BUG_ON(pages[i]->index < pos);
179 pos = pages[i]->index;
180 ret = radix_tree_delete(&brd->brd_pages, pos);
181 BUG_ON(!ret || ret != pages[i]);
182 __free_page(pages[i]);
183 }
184
185 pos++;
186
187 /*
188 * This assumes radix_tree_gang_lookup always returns as
189 * many pages as possible. If the radix-tree code changes,
190 * so will this have to.
191 */
192 } while (nr_pages == FREE_BATCH);
193}
194
195/*
196 * copy_to_brd_setup must be called before copy_to_brd. It may sleep.
197 */
198static int copy_to_brd_setup(struct brd_device *brd, sector_t sector, size_t n)
199{
200 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
201 size_t copy;
202
203 copy = min_t(size_t, n, PAGE_SIZE - offset);
204 if (!brd_insert_page(brd, sector))
205 return -ENOMEM;
206 if (copy < n) {
207 sector += copy >> SECTOR_SHIFT;
208 if (!brd_insert_page(brd, sector))
209 return -ENOMEM;
210 }
211 return 0;
212}
213
Nick Pigginb7c33572010-05-26 15:41:14 +0200214static void discard_from_brd(struct brd_device *brd,
215 sector_t sector, size_t n)
216{
217 while (n >= PAGE_SIZE) {
218 /*
219 * Don't want to actually discard pages here because
220 * re-allocating the pages can result in writeback
221 * deadlocks under heavy load.
222 */
223 if (0)
224 brd_free_page(brd, sector);
225 else
226 brd_zero_page(brd, sector);
227 sector += PAGE_SIZE >> SECTOR_SHIFT;
228 n -= PAGE_SIZE;
229 }
230}
231
Nick Piggin9db55792008-02-08 04:19:49 -0800232/*
233 * Copy n bytes from src to the brd starting at sector. Does not sleep.
234 */
235static void copy_to_brd(struct brd_device *brd, const void *src,
236 sector_t sector, size_t n)
237{
238 struct page *page;
239 void *dst;
240 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
241 size_t copy;
242
243 copy = min_t(size_t, n, PAGE_SIZE - offset);
244 page = brd_lookup_page(brd, sector);
245 BUG_ON(!page);
246
247 dst = kmap_atomic(page, KM_USER1);
248 memcpy(dst + offset, src, copy);
249 kunmap_atomic(dst, KM_USER1);
250
251 if (copy < n) {
252 src += copy;
253 sector += copy >> SECTOR_SHIFT;
254 copy = n - copy;
255 page = brd_lookup_page(brd, sector);
256 BUG_ON(!page);
257
258 dst = kmap_atomic(page, KM_USER1);
259 memcpy(dst, src, copy);
260 kunmap_atomic(dst, KM_USER1);
261 }
262}
263
264/*
265 * Copy n bytes to dst from the brd starting at sector. Does not sleep.
266 */
267static void copy_from_brd(void *dst, struct brd_device *brd,
268 sector_t sector, size_t n)
269{
270 struct page *page;
271 void *src;
272 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
273 size_t copy;
274
275 copy = min_t(size_t, n, PAGE_SIZE - offset);
276 page = brd_lookup_page(brd, sector);
277 if (page) {
278 src = kmap_atomic(page, KM_USER1);
279 memcpy(dst, src + offset, copy);
280 kunmap_atomic(src, KM_USER1);
281 } else
282 memset(dst, 0, copy);
283
284 if (copy < n) {
285 dst += copy;
286 sector += copy >> SECTOR_SHIFT;
287 copy = n - copy;
288 page = brd_lookup_page(brd, sector);
289 if (page) {
290 src = kmap_atomic(page, KM_USER1);
291 memcpy(dst, src, copy);
292 kunmap_atomic(src, KM_USER1);
293 } else
294 memset(dst, 0, copy);
295 }
296}
297
298/*
299 * Process a single bvec of a bio.
300 */
301static int brd_do_bvec(struct brd_device *brd, struct page *page,
302 unsigned int len, unsigned int off, int rw,
303 sector_t sector)
304{
305 void *mem;
306 int err = 0;
307
308 if (rw != READ) {
309 err = copy_to_brd_setup(brd, sector, len);
310 if (err)
311 goto out;
312 }
313
314 mem = kmap_atomic(page, KM_USER0);
315 if (rw == READ) {
316 copy_from_brd(mem + off, brd, sector, len);
317 flush_dcache_page(page);
Nick Pigginc2572f22009-04-15 10:32:07 +0200318 } else {
319 flush_dcache_page(page);
Nick Piggin9db55792008-02-08 04:19:49 -0800320 copy_to_brd(brd, mem + off, sector, len);
Nick Pigginc2572f22009-04-15 10:32:07 +0200321 }
Nick Piggin9db55792008-02-08 04:19:49 -0800322 kunmap_atomic(mem, KM_USER0);
323
324out:
325 return err;
326}
327
328static int brd_make_request(struct request_queue *q, struct bio *bio)
329{
330 struct block_device *bdev = bio->bi_bdev;
331 struct brd_device *brd = bdev->bd_disk->private_data;
332 int rw;
333 struct bio_vec *bvec;
334 sector_t sector;
335 int i;
336 int err = -EIO;
337
338 sector = bio->bi_sector;
339 if (sector + (bio->bi_size >> SECTOR_SHIFT) >
340 get_capacity(bdev->bd_disk))
341 goto out;
342
Nick Pigginb7c33572010-05-26 15:41:14 +0200343 if (unlikely(bio_rw_flagged(bio, BIO_RW_DISCARD))) {
344 err = 0;
345 discard_from_brd(brd, sector, bio->bi_size);
346 goto out;
347 }
348
Nick Piggin9db55792008-02-08 04:19:49 -0800349 rw = bio_rw(bio);
350 if (rw == READA)
351 rw = READ;
352
353 bio_for_each_segment(bvec, bio, i) {
354 unsigned int len = bvec->bv_len;
355 err = brd_do_bvec(brd, bvec->bv_page, len,
356 bvec->bv_offset, rw, sector);
357 if (err)
358 break;
359 sector += len >> SECTOR_SHIFT;
360 }
361
362out:
363 bio_endio(bio, err);
364
365 return 0;
366}
367
Nick Piggin75acb9c2008-02-08 04:19:50 -0800368#ifdef CONFIG_BLK_DEV_XIP
Nick Pigginb7c33572010-05-26 15:41:14 +0200369static int brd_direct_access(struct block_device *bdev, sector_t sector,
Jared Hulbert30afcb42008-04-28 02:13:02 -0700370 void **kaddr, unsigned long *pfn)
Nick Piggin75acb9c2008-02-08 04:19:50 -0800371{
372 struct brd_device *brd = bdev->bd_disk->private_data;
373 struct page *page;
374
375 if (!brd)
376 return -ENODEV;
377 if (sector & (PAGE_SECTORS-1))
378 return -EINVAL;
379 if (sector + PAGE_SECTORS > get_capacity(bdev->bd_disk))
380 return -ERANGE;
381 page = brd_insert_page(brd, sector);
382 if (!page)
383 return -ENOMEM;
Jared Hulbert30afcb42008-04-28 02:13:02 -0700384 *kaddr = page_address(page);
385 *pfn = page_to_pfn(page);
Nick Piggin75acb9c2008-02-08 04:19:50 -0800386
387 return 0;
388}
389#endif
390
Al Viro2b9ecd02008-03-02 09:24:35 -0500391static int brd_ioctl(struct block_device *bdev, fmode_t mode,
Nick Piggin9db55792008-02-08 04:19:49 -0800392 unsigned int cmd, unsigned long arg)
393{
394 int error;
Nick Piggin9db55792008-02-08 04:19:49 -0800395 struct brd_device *brd = bdev->bd_disk->private_data;
396
397 if (cmd != BLKFLSBUF)
398 return -ENOTTY;
399
400 /*
401 * ram device BLKFLSBUF has special semantics, we want to actually
402 * release and destroy the ramdisk data.
403 */
404 mutex_lock(&bdev->bd_mutex);
405 error = -EBUSY;
406 if (bdev->bd_openers <= 1) {
407 /*
408 * Invalidate the cache first, so it isn't written
409 * back to the device.
410 *
411 * Another thread might instantiate more buffercache here,
412 * but there is not much we can do to close that race.
413 */
414 invalidate_bh_lrus();
415 truncate_inode_pages(bdev->bd_inode->i_mapping, 0);
416 brd_free_pages(brd);
417 error = 0;
418 }
419 mutex_unlock(&bdev->bd_mutex);
420
421 return error;
422}
423
Alexey Dobriyan83d5cde2009-09-21 17:01:13 -0700424static const struct block_device_operations brd_fops = {
Nick Piggin75acb9c2008-02-08 04:19:50 -0800425 .owner = THIS_MODULE,
Al Viro2b9ecd02008-03-02 09:24:35 -0500426 .locked_ioctl = brd_ioctl,
Nick Piggin75acb9c2008-02-08 04:19:50 -0800427#ifdef CONFIG_BLK_DEV_XIP
428 .direct_access = brd_direct_access,
429#endif
Nick Piggin9db55792008-02-08 04:19:49 -0800430};
431
432/*
433 * And now the modules code and kernel interface.
434 */
435static int rd_nr;
436int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
Laurent Vivierd7853d12008-04-30 00:55:06 -0700437static int max_part;
438static int part_shift;
Nick Piggin9db55792008-02-08 04:19:49 -0800439module_param(rd_nr, int, 0);
440MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
441module_param(rd_size, int, 0);
442MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
Laurent Vivierd7853d12008-04-30 00:55:06 -0700443module_param(max_part, int, 0);
444MODULE_PARM_DESC(max_part, "Maximum number of partitions per RAM disk");
Nick Piggin9db55792008-02-08 04:19:49 -0800445MODULE_LICENSE("GPL");
446MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
Nick Pigginefedf512008-06-04 17:18:42 +0200447MODULE_ALIAS("rd");
Nick Piggin9db55792008-02-08 04:19:49 -0800448
449#ifndef MODULE
450/* Legacy boot options - nonmodular */
451static int __init ramdisk_size(char *str)
452{
453 rd_size = simple_strtol(str, NULL, 0);
454 return 1;
455}
Robert P. J. Day1adbee52009-06-10 12:57:08 -0700456__setup("ramdisk_size=", ramdisk_size);
Nick Piggin9db55792008-02-08 04:19:49 -0800457#endif
458
459/*
460 * The device scheme is derived from loop.c. Keep them in synch where possible
461 * (should share code eventually).
462 */
463static LIST_HEAD(brd_devices);
464static DEFINE_MUTEX(brd_devices_mutex);
465
466static struct brd_device *brd_alloc(int i)
467{
468 struct brd_device *brd;
469 struct gendisk *disk;
470
471 brd = kzalloc(sizeof(*brd), GFP_KERNEL);
472 if (!brd)
473 goto out;
474 brd->brd_number = i;
475 spin_lock_init(&brd->brd_lock);
476 INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
477
478 brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
479 if (!brd->brd_queue)
480 goto out_free_dev;
481 blk_queue_make_request(brd->brd_queue, brd_make_request);
Nick Piggindfbc4752009-04-15 10:27:07 +0200482 blk_queue_ordered(brd->brd_queue, QUEUE_ORDERED_TAG, NULL);
Martin K. Petersen086fa5f2010-02-26 00:20:38 -0500483 blk_queue_max_hw_sectors(brd->brd_queue, 1024);
Nick Piggin9db55792008-02-08 04:19:49 -0800484 blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
485
Nick Pigginb7c33572010-05-26 15:41:14 +0200486 brd->brd_queue->limits.discard_granularity = PAGE_SIZE;
487 brd->brd_queue->limits.max_discard_sectors = UINT_MAX;
488 brd->brd_queue->limits.discard_zeroes_data = 1;
489 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, brd->brd_queue);
490
Laurent Vivierd7853d12008-04-30 00:55:06 -0700491 disk = brd->brd_disk = alloc_disk(1 << part_shift);
Nick Piggin9db55792008-02-08 04:19:49 -0800492 if (!disk)
493 goto out_free_queue;
494 disk->major = RAMDISK_MAJOR;
Laurent Vivierd7853d12008-04-30 00:55:06 -0700495 disk->first_minor = i << part_shift;
Nick Piggin9db55792008-02-08 04:19:49 -0800496 disk->fops = &brd_fops;
497 disk->private_data = brd;
498 disk->queue = brd->brd_queue;
Marcin Krol53978d02008-05-23 13:04:46 -0700499 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
Nick Piggin9db55792008-02-08 04:19:49 -0800500 sprintf(disk->disk_name, "ram%d", i);
501 set_capacity(disk, rd_size * 2);
502
503 return brd;
504
505out_free_queue:
506 blk_cleanup_queue(brd->brd_queue);
507out_free_dev:
508 kfree(brd);
509out:
510 return NULL;
511}
512
513static void brd_free(struct brd_device *brd)
514{
515 put_disk(brd->brd_disk);
516 blk_cleanup_queue(brd->brd_queue);
517 brd_free_pages(brd);
518 kfree(brd);
519}
520
521static struct brd_device *brd_init_one(int i)
522{
523 struct brd_device *brd;
524
525 list_for_each_entry(brd, &brd_devices, brd_list) {
526 if (brd->brd_number == i)
527 goto out;
528 }
529
530 brd = brd_alloc(i);
531 if (brd) {
532 add_disk(brd->brd_disk);
533 list_add_tail(&brd->brd_list, &brd_devices);
534 }
535out:
536 return brd;
537}
538
539static void brd_del_one(struct brd_device *brd)
540{
541 list_del(&brd->brd_list);
542 del_gendisk(brd->brd_disk);
543 brd_free(brd);
544}
545
546static struct kobject *brd_probe(dev_t dev, int *part, void *data)
547{
548 struct brd_device *brd;
549 struct kobject *kobj;
550
551 mutex_lock(&brd_devices_mutex);
552 brd = brd_init_one(dev & MINORMASK);
553 kobj = brd ? get_disk(brd->brd_disk) : ERR_PTR(-ENOMEM);
554 mutex_unlock(&brd_devices_mutex);
555
556 *part = 0;
557 return kobj;
558}
559
560static int __init brd_init(void)
561{
562 int i, nr;
563 unsigned long range;
564 struct brd_device *brd, *next;
565
566 /*
567 * brd module now has a feature to instantiate underlying device
568 * structure on-demand, provided that there is an access dev node.
569 * However, this will not work well with user space tool that doesn't
570 * know about such "feature". In order to not break any existing
571 * tool, we do the following:
572 *
573 * (1) if rd_nr is specified, create that many upfront, and this
574 * also becomes a hard limit.
575 * (2) if rd_nr is not specified, create 1 rd device on module
576 * load, user can further extend brd device by create dev node
577 * themselves and have kernel automatically instantiate actual
578 * device on-demand.
579 */
Laurent Vivierd7853d12008-04-30 00:55:06 -0700580
581 part_shift = 0;
582 if (max_part > 0)
583 part_shift = fls(max_part);
584
585 if (rd_nr > 1UL << (MINORBITS - part_shift))
Nick Piggin9db55792008-02-08 04:19:49 -0800586 return -EINVAL;
587
588 if (rd_nr) {
589 nr = rd_nr;
590 range = rd_nr;
591 } else {
592 nr = CONFIG_BLK_DEV_RAM_COUNT;
Laurent Vivierd7853d12008-04-30 00:55:06 -0700593 range = 1UL << (MINORBITS - part_shift);
Nick Piggin9db55792008-02-08 04:19:49 -0800594 }
595
596 if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
597 return -EIO;
598
599 for (i = 0; i < nr; i++) {
600 brd = brd_alloc(i);
601 if (!brd)
602 goto out_free;
603 list_add_tail(&brd->brd_list, &brd_devices);
604 }
605
606 /* point of no return */
607
608 list_for_each_entry(brd, &brd_devices, brd_list)
609 add_disk(brd->brd_disk);
610
611 blk_register_region(MKDEV(RAMDISK_MAJOR, 0), range,
612 THIS_MODULE, brd_probe, NULL, NULL);
613
614 printk(KERN_INFO "brd: module loaded\n");
615 return 0;
616
617out_free:
618 list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
619 list_del(&brd->brd_list);
620 brd_free(brd);
621 }
Akinobu Mitac82f2962008-08-20 14:09:09 -0700622 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
Nick Piggin9db55792008-02-08 04:19:49 -0800623
Nick Piggin9db55792008-02-08 04:19:49 -0800624 return -ENOMEM;
625}
626
627static void __exit brd_exit(void)
628{
629 unsigned long range;
630 struct brd_device *brd, *next;
631
Laurent Vivierd7853d12008-04-30 00:55:06 -0700632 range = rd_nr ? rd_nr : 1UL << (MINORBITS - part_shift);
Nick Piggin9db55792008-02-08 04:19:49 -0800633
634 list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
635 brd_del_one(brd);
636
637 blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), range);
638 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
639}
640
641module_init(brd_init);
642module_exit(brd_exit);
643