blob: 85364804364fc3a058114b30379927dcd974621f [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>
18#include <linux/gfp.h>
19#include <linux/radix-tree.h>
20#include <linux/buffer_head.h> /* invalidate_bh_lrus() */
21
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
Nick Piggin9db55792008-02-08 04:19:49 -0800111 page = alloc_page(GFP_NOIO | __GFP_HIGHMEM | __GFP_ZERO);
112 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
136/*
137 * Free all backing store pages and radix tree. This must only be called when
138 * there are no other users of the device.
139 */
140#define FREE_BATCH 16
141static void brd_free_pages(struct brd_device *brd)
142{
143 unsigned long pos = 0;
144 struct page *pages[FREE_BATCH];
145 int nr_pages;
146
147 do {
148 int i;
149
150 nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
151 (void **)pages, pos, FREE_BATCH);
152
153 for (i = 0; i < nr_pages; i++) {
154 void *ret;
155
156 BUG_ON(pages[i]->index < pos);
157 pos = pages[i]->index;
158 ret = radix_tree_delete(&brd->brd_pages, pos);
159 BUG_ON(!ret || ret != pages[i]);
160 __free_page(pages[i]);
161 }
162
163 pos++;
164
165 /*
166 * This assumes radix_tree_gang_lookup always returns as
167 * many pages as possible. If the radix-tree code changes,
168 * so will this have to.
169 */
170 } while (nr_pages == FREE_BATCH);
171}
172
173/*
174 * copy_to_brd_setup must be called before copy_to_brd. It may sleep.
175 */
176static int copy_to_brd_setup(struct brd_device *brd, sector_t sector, size_t n)
177{
178 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
179 size_t copy;
180
181 copy = min_t(size_t, n, PAGE_SIZE - offset);
182 if (!brd_insert_page(brd, sector))
183 return -ENOMEM;
184 if (copy < n) {
185 sector += copy >> SECTOR_SHIFT;
186 if (!brd_insert_page(brd, sector))
187 return -ENOMEM;
188 }
189 return 0;
190}
191
192/*
193 * Copy n bytes from src to the brd starting at sector. Does not sleep.
194 */
195static void copy_to_brd(struct brd_device *brd, const void *src,
196 sector_t sector, size_t n)
197{
198 struct page *page;
199 void *dst;
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 page = brd_lookup_page(brd, sector);
205 BUG_ON(!page);
206
207 dst = kmap_atomic(page, KM_USER1);
208 memcpy(dst + offset, src, copy);
209 kunmap_atomic(dst, KM_USER1);
210
211 if (copy < n) {
212 src += copy;
213 sector += copy >> SECTOR_SHIFT;
214 copy = n - copy;
215 page = brd_lookup_page(brd, sector);
216 BUG_ON(!page);
217
218 dst = kmap_atomic(page, KM_USER1);
219 memcpy(dst, src, copy);
220 kunmap_atomic(dst, KM_USER1);
221 }
222}
223
224/*
225 * Copy n bytes to dst from the brd starting at sector. Does not sleep.
226 */
227static void copy_from_brd(void *dst, struct brd_device *brd,
228 sector_t sector, size_t n)
229{
230 struct page *page;
231 void *src;
232 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
233 size_t copy;
234
235 copy = min_t(size_t, n, PAGE_SIZE - offset);
236 page = brd_lookup_page(brd, sector);
237 if (page) {
238 src = kmap_atomic(page, KM_USER1);
239 memcpy(dst, src + offset, copy);
240 kunmap_atomic(src, KM_USER1);
241 } else
242 memset(dst, 0, copy);
243
244 if (copy < n) {
245 dst += copy;
246 sector += copy >> SECTOR_SHIFT;
247 copy = n - copy;
248 page = brd_lookup_page(brd, sector);
249 if (page) {
250 src = kmap_atomic(page, KM_USER1);
251 memcpy(dst, src, copy);
252 kunmap_atomic(src, KM_USER1);
253 } else
254 memset(dst, 0, copy);
255 }
256}
257
258/*
259 * Process a single bvec of a bio.
260 */
261static int brd_do_bvec(struct brd_device *brd, struct page *page,
262 unsigned int len, unsigned int off, int rw,
263 sector_t sector)
264{
265 void *mem;
266 int err = 0;
267
268 if (rw != READ) {
269 err = copy_to_brd_setup(brd, sector, len);
270 if (err)
271 goto out;
272 }
273
274 mem = kmap_atomic(page, KM_USER0);
275 if (rw == READ) {
276 copy_from_brd(mem + off, brd, sector, len);
277 flush_dcache_page(page);
278 } else
279 copy_to_brd(brd, mem + off, sector, len);
280 kunmap_atomic(mem, KM_USER0);
281
282out:
283 return err;
284}
285
286static int brd_make_request(struct request_queue *q, struct bio *bio)
287{
288 struct block_device *bdev = bio->bi_bdev;
289 struct brd_device *brd = bdev->bd_disk->private_data;
290 int rw;
291 struct bio_vec *bvec;
292 sector_t sector;
293 int i;
294 int err = -EIO;
295
296 sector = bio->bi_sector;
297 if (sector + (bio->bi_size >> SECTOR_SHIFT) >
298 get_capacity(bdev->bd_disk))
299 goto out;
300
301 rw = bio_rw(bio);
302 if (rw == READA)
303 rw = READ;
304
305 bio_for_each_segment(bvec, bio, i) {
306 unsigned int len = bvec->bv_len;
307 err = brd_do_bvec(brd, bvec->bv_page, len,
308 bvec->bv_offset, rw, sector);
309 if (err)
310 break;
311 sector += len >> SECTOR_SHIFT;
312 }
313
314out:
315 bio_endio(bio, err);
316
317 return 0;
318}
319
Nick Piggin75acb9c2008-02-08 04:19:50 -0800320#ifdef CONFIG_BLK_DEV_XIP
321static int brd_direct_access (struct block_device *bdev, sector_t sector,
322 unsigned long *data)
323{
324 struct brd_device *brd = bdev->bd_disk->private_data;
325 struct page *page;
326
327 if (!brd)
328 return -ENODEV;
329 if (sector & (PAGE_SECTORS-1))
330 return -EINVAL;
331 if (sector + PAGE_SECTORS > get_capacity(bdev->bd_disk))
332 return -ERANGE;
333 page = brd_insert_page(brd, sector);
334 if (!page)
335 return -ENOMEM;
336 *data = (unsigned long)page_address(page);
337
338 return 0;
339}
340#endif
341
Nick Piggin9db55792008-02-08 04:19:49 -0800342static int brd_ioctl(struct inode *inode, struct file *file,
343 unsigned int cmd, unsigned long arg)
344{
345 int error;
346 struct block_device *bdev = inode->i_bdev;
347 struct brd_device *brd = bdev->bd_disk->private_data;
348
349 if (cmd != BLKFLSBUF)
350 return -ENOTTY;
351
352 /*
353 * ram device BLKFLSBUF has special semantics, we want to actually
354 * release and destroy the ramdisk data.
355 */
356 mutex_lock(&bdev->bd_mutex);
357 error = -EBUSY;
358 if (bdev->bd_openers <= 1) {
359 /*
360 * Invalidate the cache first, so it isn't written
361 * back to the device.
362 *
363 * Another thread might instantiate more buffercache here,
364 * but there is not much we can do to close that race.
365 */
366 invalidate_bh_lrus();
367 truncate_inode_pages(bdev->bd_inode->i_mapping, 0);
368 brd_free_pages(brd);
369 error = 0;
370 }
371 mutex_unlock(&bdev->bd_mutex);
372
373 return error;
374}
375
376static struct block_device_operations brd_fops = {
Nick Piggin75acb9c2008-02-08 04:19:50 -0800377 .owner = THIS_MODULE,
378 .ioctl = brd_ioctl,
379#ifdef CONFIG_BLK_DEV_XIP
380 .direct_access = brd_direct_access,
381#endif
Nick Piggin9db55792008-02-08 04:19:49 -0800382};
383
384/*
385 * And now the modules code and kernel interface.
386 */
387static int rd_nr;
388int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
389module_param(rd_nr, int, 0);
390MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
391module_param(rd_size, int, 0);
392MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
393MODULE_LICENSE("GPL");
394MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
395
396#ifndef MODULE
397/* Legacy boot options - nonmodular */
398static int __init ramdisk_size(char *str)
399{
400 rd_size = simple_strtol(str, NULL, 0);
401 return 1;
402}
403static int __init ramdisk_size2(char *str)
404{
405 return ramdisk_size(str);
406}
407__setup("ramdisk=", ramdisk_size);
408__setup("ramdisk_size=", ramdisk_size2);
409#endif
410
411/*
412 * The device scheme is derived from loop.c. Keep them in synch where possible
413 * (should share code eventually).
414 */
415static LIST_HEAD(brd_devices);
416static DEFINE_MUTEX(brd_devices_mutex);
417
418static struct brd_device *brd_alloc(int i)
419{
420 struct brd_device *brd;
421 struct gendisk *disk;
422
423 brd = kzalloc(sizeof(*brd), GFP_KERNEL);
424 if (!brd)
425 goto out;
426 brd->brd_number = i;
427 spin_lock_init(&brd->brd_lock);
428 INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
429
430 brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
431 if (!brd->brd_queue)
432 goto out_free_dev;
433 blk_queue_make_request(brd->brd_queue, brd_make_request);
434 blk_queue_max_sectors(brd->brd_queue, 1024);
435 blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
436
437 disk = brd->brd_disk = alloc_disk(1);
438 if (!disk)
439 goto out_free_queue;
440 disk->major = RAMDISK_MAJOR;
441 disk->first_minor = i;
442 disk->fops = &brd_fops;
443 disk->private_data = brd;
444 disk->queue = brd->brd_queue;
445 sprintf(disk->disk_name, "ram%d", i);
446 set_capacity(disk, rd_size * 2);
447
448 return brd;
449
450out_free_queue:
451 blk_cleanup_queue(brd->brd_queue);
452out_free_dev:
453 kfree(brd);
454out:
455 return NULL;
456}
457
458static void brd_free(struct brd_device *brd)
459{
460 put_disk(brd->brd_disk);
461 blk_cleanup_queue(brd->brd_queue);
462 brd_free_pages(brd);
463 kfree(brd);
464}
465
466static struct brd_device *brd_init_one(int i)
467{
468 struct brd_device *brd;
469
470 list_for_each_entry(brd, &brd_devices, brd_list) {
471 if (brd->brd_number == i)
472 goto out;
473 }
474
475 brd = brd_alloc(i);
476 if (brd) {
477 add_disk(brd->brd_disk);
478 list_add_tail(&brd->brd_list, &brd_devices);
479 }
480out:
481 return brd;
482}
483
484static void brd_del_one(struct brd_device *brd)
485{
486 list_del(&brd->brd_list);
487 del_gendisk(brd->brd_disk);
488 brd_free(brd);
489}
490
491static struct kobject *brd_probe(dev_t dev, int *part, void *data)
492{
493 struct brd_device *brd;
494 struct kobject *kobj;
495
496 mutex_lock(&brd_devices_mutex);
497 brd = brd_init_one(dev & MINORMASK);
498 kobj = brd ? get_disk(brd->brd_disk) : ERR_PTR(-ENOMEM);
499 mutex_unlock(&brd_devices_mutex);
500
501 *part = 0;
502 return kobj;
503}
504
505static int __init brd_init(void)
506{
507 int i, nr;
508 unsigned long range;
509 struct brd_device *brd, *next;
510
511 /*
512 * brd module now has a feature to instantiate underlying device
513 * structure on-demand, provided that there is an access dev node.
514 * However, this will not work well with user space tool that doesn't
515 * know about such "feature". In order to not break any existing
516 * tool, we do the following:
517 *
518 * (1) if rd_nr is specified, create that many upfront, and this
519 * also becomes a hard limit.
520 * (2) if rd_nr is not specified, create 1 rd device on module
521 * load, user can further extend brd device by create dev node
522 * themselves and have kernel automatically instantiate actual
523 * device on-demand.
524 */
525 if (rd_nr > 1UL << MINORBITS)
526 return -EINVAL;
527
528 if (rd_nr) {
529 nr = rd_nr;
530 range = rd_nr;
531 } else {
532 nr = CONFIG_BLK_DEV_RAM_COUNT;
533 range = 1UL << MINORBITS;
534 }
535
536 if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
537 return -EIO;
538
539 for (i = 0; i < nr; i++) {
540 brd = brd_alloc(i);
541 if (!brd)
542 goto out_free;
543 list_add_tail(&brd->brd_list, &brd_devices);
544 }
545
546 /* point of no return */
547
548 list_for_each_entry(brd, &brd_devices, brd_list)
549 add_disk(brd->brd_disk);
550
551 blk_register_region(MKDEV(RAMDISK_MAJOR, 0), range,
552 THIS_MODULE, brd_probe, NULL, NULL);
553
554 printk(KERN_INFO "brd: module loaded\n");
555 return 0;
556
557out_free:
558 list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
559 list_del(&brd->brd_list);
560 brd_free(brd);
561 }
562
563 unregister_blkdev(RAMDISK_MAJOR, "brd");
564 return -ENOMEM;
565}
566
567static void __exit brd_exit(void)
568{
569 unsigned long range;
570 struct brd_device *brd, *next;
571
572 range = rd_nr ? rd_nr : 1UL << MINORBITS;
573
574 list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
575 brd_del_one(brd);
576
577 blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), range);
578 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
579}
580
581module_init(brd_init);
582module_exit(brd_exit);
583