Loading...
1/*
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/mutex.h>
19#include <linux/radix-tree.h>
20#include <linux/buffer_head.h> /* invalidate_bh_lrus() */
21#include <linux/slab.h>
22
23#include <asm/uaccess.h>
24
25#define SECTOR_SHIFT 9
26#define PAGE_SECTORS_SHIFT (PAGE_SHIFT - SECTOR_SHIFT)
27#define PAGE_SECTORS (1 << PAGE_SECTORS_SHIFT)
28
29/*
30 * Each block ramdisk device has a radix_tree brd_pages of pages that stores
31 * the pages containing the block device's contents. A brd page's ->index is
32 * its offset in PAGE_SIZE units. This is similar to, but in no way connected
33 * with, the kernel's pagecache or buffer cache (which sit above our block
34 * device).
35 */
36struct brd_device {
37 int brd_number;
38
39 struct request_queue *brd_queue;
40 struct gendisk *brd_disk;
41 struct list_head brd_list;
42
43 /*
44 * Backing store of pages and lock to protect it. This is the contents
45 * of the block device.
46 */
47 spinlock_t brd_lock;
48 struct radix_tree_root brd_pages;
49};
50
51/*
52 * Look up and return a brd's page for a given sector.
53 */
54static DEFINE_MUTEX(brd_mutex);
55static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
56{
57 pgoff_t idx;
58 struct page *page;
59
60 /*
61 * The page lifetime is protected by the fact that we have opened the
62 * device node -- brd pages will never be deleted under us, so we
63 * don't need any further locking or refcounting.
64 *
65 * This is strictly true for the radix-tree nodes as well (ie. we
66 * don't actually need the rcu_read_lock()), however that is not a
67 * documented feature of the radix-tree API so it is better to be
68 * safe here (we don't have total exclusion from radix tree updates
69 * here, only deletes).
70 */
71 rcu_read_lock();
72 idx = sector >> PAGE_SECTORS_SHIFT; /* sector to page index */
73 page = radix_tree_lookup(&brd->brd_pages, idx);
74 rcu_read_unlock();
75
76 BUG_ON(page && page->index != idx);
77
78 return page;
79}
80
81/*
82 * Look up and return a brd's page for a given sector.
83 * If one does not exist, allocate an empty page, and insert that. Then
84 * return it.
85 */
86static struct page *brd_insert_page(struct brd_device *brd, sector_t sector)
87{
88 pgoff_t idx;
89 struct page *page;
90 gfp_t gfp_flags;
91
92 page = brd_lookup_page(brd, sector);
93 if (page)
94 return page;
95
96 /*
97 * Must use NOIO because we don't want to recurse back into the
98 * block or filesystem layers from page reclaim.
99 *
100 * Cannot support XIP and highmem, because our ->direct_access
101 * routine for XIP must return memory that is always addressable.
102 * If XIP was reworked to use pfns and kmap throughout, this
103 * restriction might be able to be lifted.
104 */
105 gfp_flags = GFP_NOIO | __GFP_ZERO;
106#ifndef CONFIG_BLK_DEV_XIP
107 gfp_flags |= __GFP_HIGHMEM;
108#endif
109 page = alloc_page(gfp_flags);
110 if (!page)
111 return NULL;
112
113 if (radix_tree_preload(GFP_NOIO)) {
114 __free_page(page);
115 return NULL;
116 }
117
118 spin_lock(&brd->brd_lock);
119 idx = sector >> PAGE_SECTORS_SHIFT;
120 if (radix_tree_insert(&brd->brd_pages, idx, page)) {
121 __free_page(page);
122 page = radix_tree_lookup(&brd->brd_pages, idx);
123 BUG_ON(!page);
124 BUG_ON(page->index != idx);
125 } else
126 page->index = idx;
127 spin_unlock(&brd->brd_lock);
128
129 radix_tree_preload_end();
130
131 return page;
132}
133
134static void brd_free_page(struct brd_device *brd, sector_t sector)
135{
136 struct page *page;
137 pgoff_t idx;
138
139 spin_lock(&brd->brd_lock);
140 idx = sector >> PAGE_SECTORS_SHIFT;
141 page = radix_tree_delete(&brd->brd_pages, idx);
142 spin_unlock(&brd->brd_lock);
143 if (page)
144 __free_page(page);
145}
146
147static void brd_zero_page(struct brd_device *brd, sector_t sector)
148{
149 struct page *page;
150
151 page = brd_lookup_page(brd, sector);
152 if (page)
153 clear_highpage(page);
154}
155
156/*
157 * Free all backing store pages and radix tree. This must only be called when
158 * there are no other users of the device.
159 */
160#define FREE_BATCH 16
161static void brd_free_pages(struct brd_device *brd)
162{
163 unsigned long pos = 0;
164 struct page *pages[FREE_BATCH];
165 int nr_pages;
166
167 do {
168 int i;
169
170 nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
171 (void **)pages, pos, FREE_BATCH);
172
173 for (i = 0; i < nr_pages; i++) {
174 void *ret;
175
176 BUG_ON(pages[i]->index < pos);
177 pos = pages[i]->index;
178 ret = radix_tree_delete(&brd->brd_pages, pos);
179 BUG_ON(!ret || ret != pages[i]);
180 __free_page(pages[i]);
181 }
182
183 pos++;
184
185 /*
186 * This assumes radix_tree_gang_lookup always returns as
187 * many pages as possible. If the radix-tree code changes,
188 * so will this have to.
189 */
190 } while (nr_pages == FREE_BATCH);
191}
192
193/*
194 * copy_to_brd_setup must be called before copy_to_brd. It may sleep.
195 */
196static int copy_to_brd_setup(struct brd_device *brd, sector_t sector, size_t n)
197{
198 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
199 size_t copy;
200
201 copy = min_t(size_t, n, PAGE_SIZE - offset);
202 if (!brd_insert_page(brd, sector))
203 return -ENOMEM;
204 if (copy < n) {
205 sector += copy >> SECTOR_SHIFT;
206 if (!brd_insert_page(brd, sector))
207 return -ENOMEM;
208 }
209 return 0;
210}
211
212static void discard_from_brd(struct brd_device *brd,
213 sector_t sector, size_t n)
214{
215 while (n >= PAGE_SIZE) {
216 /*
217 * Don't want to actually discard pages here because
218 * re-allocating the pages can result in writeback
219 * deadlocks under heavy load.
220 */
221 if (0)
222 brd_free_page(brd, sector);
223 else
224 brd_zero_page(brd, sector);
225 sector += PAGE_SIZE >> SECTOR_SHIFT;
226 n -= PAGE_SIZE;
227 }
228}
229
230/*
231 * Copy n bytes from src to the brd starting at sector. Does not sleep.
232 */
233static void copy_to_brd(struct brd_device *brd, const void *src,
234 sector_t sector, size_t n)
235{
236 struct page *page;
237 void *dst;
238 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
239 size_t copy;
240
241 copy = min_t(size_t, n, PAGE_SIZE - offset);
242 page = brd_lookup_page(brd, sector);
243 BUG_ON(!page);
244
245 dst = kmap_atomic(page, KM_USER1);
246 memcpy(dst + offset, src, copy);
247 kunmap_atomic(dst, KM_USER1);
248
249 if (copy < n) {
250 src += copy;
251 sector += copy >> SECTOR_SHIFT;
252 copy = n - copy;
253 page = brd_lookup_page(brd, sector);
254 BUG_ON(!page);
255
256 dst = kmap_atomic(page, KM_USER1);
257 memcpy(dst, src, copy);
258 kunmap_atomic(dst, KM_USER1);
259 }
260}
261
262/*
263 * Copy n bytes to dst from the brd starting at sector. Does not sleep.
264 */
265static void copy_from_brd(void *dst, struct brd_device *brd,
266 sector_t sector, size_t n)
267{
268 struct page *page;
269 void *src;
270 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
271 size_t copy;
272
273 copy = min_t(size_t, n, PAGE_SIZE - offset);
274 page = brd_lookup_page(brd, sector);
275 if (page) {
276 src = kmap_atomic(page, KM_USER1);
277 memcpy(dst, src + offset, copy);
278 kunmap_atomic(src, KM_USER1);
279 } else
280 memset(dst, 0, copy);
281
282 if (copy < n) {
283 dst += copy;
284 sector += copy >> SECTOR_SHIFT;
285 copy = n - copy;
286 page = brd_lookup_page(brd, sector);
287 if (page) {
288 src = kmap_atomic(page, KM_USER1);
289 memcpy(dst, src, copy);
290 kunmap_atomic(src, KM_USER1);
291 } else
292 memset(dst, 0, copy);
293 }
294}
295
296/*
297 * Process a single bvec of a bio.
298 */
299static int brd_do_bvec(struct brd_device *brd, struct page *page,
300 unsigned int len, unsigned int off, int rw,
301 sector_t sector)
302{
303 void *mem;
304 int err = 0;
305
306 if (rw != READ) {
307 err = copy_to_brd_setup(brd, sector, len);
308 if (err)
309 goto out;
310 }
311
312 mem = kmap_atomic(page, KM_USER0);
313 if (rw == READ) {
314 copy_from_brd(mem + off, brd, sector, len);
315 flush_dcache_page(page);
316 } else {
317 flush_dcache_page(page);
318 copy_to_brd(brd, mem + off, sector, len);
319 }
320 kunmap_atomic(mem, KM_USER0);
321
322out:
323 return err;
324}
325
326static int brd_make_request(struct request_queue *q, struct bio *bio)
327{
328 struct block_device *bdev = bio->bi_bdev;
329 struct brd_device *brd = bdev->bd_disk->private_data;
330 int rw;
331 struct bio_vec *bvec;
332 sector_t sector;
333 int i;
334 int err = -EIO;
335
336 sector = bio->bi_sector;
337 if (sector + (bio->bi_size >> SECTOR_SHIFT) >
338 get_capacity(bdev->bd_disk))
339 goto out;
340
341 if (unlikely(bio->bi_rw & REQ_DISCARD)) {
342 err = 0;
343 discard_from_brd(brd, sector, bio->bi_size);
344 goto out;
345 }
346
347 rw = bio_rw(bio);
348 if (rw == READA)
349 rw = READ;
350
351 bio_for_each_segment(bvec, bio, i) {
352 unsigned int len = bvec->bv_len;
353 err = brd_do_bvec(brd, bvec->bv_page, len,
354 bvec->bv_offset, rw, sector);
355 if (err)
356 break;
357 sector += len >> SECTOR_SHIFT;
358 }
359
360out:
361 bio_endio(bio, err);
362
363 return 0;
364}
365
366#ifdef CONFIG_BLK_DEV_XIP
367static int brd_direct_access(struct block_device *bdev, sector_t sector,
368 void **kaddr, unsigned long *pfn)
369{
370 struct brd_device *brd = bdev->bd_disk->private_data;
371 struct page *page;
372
373 if (!brd)
374 return -ENODEV;
375 if (sector & (PAGE_SECTORS-1))
376 return -EINVAL;
377 if (sector + PAGE_SECTORS > get_capacity(bdev->bd_disk))
378 return -ERANGE;
379 page = brd_insert_page(brd, sector);
380 if (!page)
381 return -ENOMEM;
382 *kaddr = page_address(page);
383 *pfn = page_to_pfn(page);
384
385 return 0;
386}
387#endif
388
389static int brd_ioctl(struct block_device *bdev, fmode_t mode,
390 unsigned int cmd, unsigned long arg)
391{
392 int error;
393 struct brd_device *brd = bdev->bd_disk->private_data;
394
395 if (cmd != BLKFLSBUF)
396 return -ENOTTY;
397
398 /*
399 * ram device BLKFLSBUF has special semantics, we want to actually
400 * release and destroy the ramdisk data.
401 */
402 mutex_lock(&brd_mutex);
403 mutex_lock(&bdev->bd_mutex);
404 error = -EBUSY;
405 if (bdev->bd_openers <= 1) {
406 /*
407 * Invalidate the cache first, so it isn't written
408 * back to the device.
409 *
410 * Another thread might instantiate more buffercache here,
411 * but there is not much we can do to close that race.
412 */
413 invalidate_bh_lrus();
414 truncate_inode_pages(bdev->bd_inode->i_mapping, 0);
415 brd_free_pages(brd);
416 error = 0;
417 }
418 mutex_unlock(&bdev->bd_mutex);
419 mutex_unlock(&brd_mutex);
420
421 return error;
422}
423
424static const struct block_device_operations brd_fops = {
425 .owner = THIS_MODULE,
426 .ioctl = brd_ioctl,
427#ifdef CONFIG_BLK_DEV_XIP
428 .direct_access = brd_direct_access,
429#endif
430};
431
432/*
433 * And now the modules code and kernel interface.
434 */
435static int rd_nr;
436int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
437static int max_part;
438static int part_shift;
439module_param(rd_nr, int, S_IRUGO);
440MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
441module_param(rd_size, int, S_IRUGO);
442MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
443module_param(max_part, int, S_IRUGO);
444MODULE_PARM_DESC(max_part, "Maximum number of partitions per RAM disk");
445MODULE_LICENSE("GPL");
446MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
447MODULE_ALIAS("rd");
448
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}
456__setup("ramdisk_size=", ramdisk_size);
457#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);
482 blk_queue_max_hw_sectors(brd->brd_queue, 1024);
483 blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
484
485 brd->brd_queue->limits.discard_granularity = PAGE_SIZE;
486 brd->brd_queue->limits.max_discard_sectors = UINT_MAX;
487 brd->brd_queue->limits.discard_zeroes_data = 1;
488 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, brd->brd_queue);
489
490 disk = brd->brd_disk = alloc_disk(1 << part_shift);
491 if (!disk)
492 goto out_free_queue;
493 disk->major = RAMDISK_MAJOR;
494 disk->first_minor = i << part_shift;
495 disk->fops = &brd_fops;
496 disk->private_data = brd;
497 disk->queue = brd->brd_queue;
498 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
499 sprintf(disk->disk_name, "ram%d", i);
500 set_capacity(disk, rd_size * 2);
501
502 return brd;
503
504out_free_queue:
505 blk_cleanup_queue(brd->brd_queue);
506out_free_dev:
507 kfree(brd);
508out:
509 return NULL;
510}
511
512static void brd_free(struct brd_device *brd)
513{
514 put_disk(brd->brd_disk);
515 blk_cleanup_queue(brd->brd_queue);
516 brd_free_pages(brd);
517 kfree(brd);
518}
519
520static struct brd_device *brd_init_one(int i)
521{
522 struct brd_device *brd;
523
524 list_for_each_entry(brd, &brd_devices, brd_list) {
525 if (brd->brd_number == i)
526 goto out;
527 }
528
529 brd = brd_alloc(i);
530 if (brd) {
531 add_disk(brd->brd_disk);
532 list_add_tail(&brd->brd_list, &brd_devices);
533 }
534out:
535 return brd;
536}
537
538static void brd_del_one(struct brd_device *brd)
539{
540 list_del(&brd->brd_list);
541 del_gendisk(brd->brd_disk);
542 brd_free(brd);
543}
544
545static struct kobject *brd_probe(dev_t dev, int *part, void *data)
546{
547 struct brd_device *brd;
548 struct kobject *kobj;
549
550 mutex_lock(&brd_devices_mutex);
551 brd = brd_init_one(MINOR(dev) >> part_shift);
552 kobj = brd ? get_disk(brd->brd_disk) : ERR_PTR(-ENOMEM);
553 mutex_unlock(&brd_devices_mutex);
554
555 *part = 0;
556 return kobj;
557}
558
559static int __init brd_init(void)
560{
561 int i, nr;
562 unsigned long range;
563 struct brd_device *brd, *next;
564
565 /*
566 * brd module now has a feature to instantiate underlying device
567 * structure on-demand, provided that there is an access dev node.
568 * However, this will not work well with user space tool that doesn't
569 * know about such "feature". In order to not break any existing
570 * tool, we do the following:
571 *
572 * (1) if rd_nr is specified, create that many upfront, and this
573 * also becomes a hard limit.
574 * (2) if rd_nr is not specified, create CONFIG_BLK_DEV_RAM_COUNT
575 * (default 16) rd device on module load, user can further
576 * extend brd device by create dev node themselves and have
577 * kernel automatically instantiate actual device on-demand.
578 */
579
580 part_shift = 0;
581 if (max_part > 0) {
582 part_shift = fls(max_part);
583
584 /*
585 * Adjust max_part according to part_shift as it is exported
586 * to user space so that user can decide correct minor number
587 * if [s]he want to create more devices.
588 *
589 * Note that -1 is required because partition 0 is reserved
590 * for the whole disk.
591 */
592 max_part = (1UL << part_shift) - 1;
593 }
594
595 if ((1UL << part_shift) > DISK_MAX_PARTS)
596 return -EINVAL;
597
598 if (rd_nr > 1UL << (MINORBITS - part_shift))
599 return -EINVAL;
600
601 if (rd_nr) {
602 nr = rd_nr;
603 range = rd_nr << part_shift;
604 } else {
605 nr = CONFIG_BLK_DEV_RAM_COUNT;
606 range = 1UL << MINORBITS;
607 }
608
609 if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
610 return -EIO;
611
612 for (i = 0; i < nr; i++) {
613 brd = brd_alloc(i);
614 if (!brd)
615 goto out_free;
616 list_add_tail(&brd->brd_list, &brd_devices);
617 }
618
619 /* point of no return */
620
621 list_for_each_entry(brd, &brd_devices, brd_list)
622 add_disk(brd->brd_disk);
623
624 blk_register_region(MKDEV(RAMDISK_MAJOR, 0), range,
625 THIS_MODULE, brd_probe, NULL, NULL);
626
627 printk(KERN_INFO "brd: module loaded\n");
628 return 0;
629
630out_free:
631 list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
632 list_del(&brd->brd_list);
633 brd_free(brd);
634 }
635 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
636
637 return -ENOMEM;
638}
639
640static void __exit brd_exit(void)
641{
642 unsigned long range;
643 struct brd_device *brd, *next;
644
645 range = rd_nr ? rd_nr << part_shift : 1UL << MINORBITS;
646
647 list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
648 brd_del_one(brd);
649
650 blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), range);
651 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
652}
653
654module_init(brd_init);
655module_exit(brd_exit);
656
1/*
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/mutex.h>
19#include <linux/radix-tree.h>
20#include <linux/fs.h>
21#include <linux/slab.h>
22
23#include <asm/uaccess.h>
24
25#define SECTOR_SHIFT 9
26#define PAGE_SECTORS_SHIFT (PAGE_SHIFT - SECTOR_SHIFT)
27#define PAGE_SECTORS (1 << PAGE_SECTORS_SHIFT)
28
29/*
30 * Each block ramdisk device has a radix_tree brd_pages of pages that stores
31 * the pages containing the block device's contents. A brd page's ->index is
32 * its offset in PAGE_SIZE units. This is similar to, but in no way connected
33 * with, the kernel's pagecache or buffer cache (which sit above our block
34 * device).
35 */
36struct brd_device {
37 int brd_number;
38
39 struct request_queue *brd_queue;
40 struct gendisk *brd_disk;
41 struct list_head brd_list;
42
43 /*
44 * Backing store of pages and lock to protect it. This is the contents
45 * of the block device.
46 */
47 spinlock_t brd_lock;
48 struct radix_tree_root brd_pages;
49};
50
51/*
52 * Look up and return a brd's page for a given sector.
53 */
54static DEFINE_MUTEX(brd_mutex);
55static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
56{
57 pgoff_t idx;
58 struct page *page;
59
60 /*
61 * The page lifetime is protected by the fact that we have opened the
62 * device node -- brd pages will never be deleted under us, so we
63 * don't need any further locking or refcounting.
64 *
65 * This is strictly true for the radix-tree nodes as well (ie. we
66 * don't actually need the rcu_read_lock()), however that is not a
67 * documented feature of the radix-tree API so it is better to be
68 * safe here (we don't have total exclusion from radix tree updates
69 * here, only deletes).
70 */
71 rcu_read_lock();
72 idx = sector >> PAGE_SECTORS_SHIFT; /* sector to page index */
73 page = radix_tree_lookup(&brd->brd_pages, idx);
74 rcu_read_unlock();
75
76 BUG_ON(page && page->index != idx);
77
78 return page;
79}
80
81/*
82 * Look up and return a brd's page for a given sector.
83 * If one does not exist, allocate an empty page, and insert that. Then
84 * return it.
85 */
86static struct page *brd_insert_page(struct brd_device *brd, sector_t sector)
87{
88 pgoff_t idx;
89 struct page *page;
90 gfp_t gfp_flags;
91
92 page = brd_lookup_page(brd, sector);
93 if (page)
94 return page;
95
96 /*
97 * Must use NOIO because we don't want to recurse back into the
98 * block or filesystem layers from page reclaim.
99 *
100 * Cannot support XIP and highmem, because our ->direct_access
101 * routine for XIP must return memory that is always addressable.
102 * If XIP was reworked to use pfns and kmap throughout, this
103 * restriction might be able to be lifted.
104 */
105 gfp_flags = GFP_NOIO | __GFP_ZERO;
106#ifndef CONFIG_BLK_DEV_XIP
107 gfp_flags |= __GFP_HIGHMEM;
108#endif
109 page = alloc_page(gfp_flags);
110 if (!page)
111 return NULL;
112
113 if (radix_tree_preload(GFP_NOIO)) {
114 __free_page(page);
115 return NULL;
116 }
117
118 spin_lock(&brd->brd_lock);
119 idx = sector >> PAGE_SECTORS_SHIFT;
120 page->index = idx;
121 if (radix_tree_insert(&brd->brd_pages, idx, page)) {
122 __free_page(page);
123 page = radix_tree_lookup(&brd->brd_pages, idx);
124 BUG_ON(!page);
125 BUG_ON(page->index != idx);
126 }
127 spin_unlock(&brd->brd_lock);
128
129 radix_tree_preload_end();
130
131 return page;
132}
133
134static void brd_free_page(struct brd_device *brd, sector_t sector)
135{
136 struct page *page;
137 pgoff_t idx;
138
139 spin_lock(&brd->brd_lock);
140 idx = sector >> PAGE_SECTORS_SHIFT;
141 page = radix_tree_delete(&brd->brd_pages, idx);
142 spin_unlock(&brd->brd_lock);
143 if (page)
144 __free_page(page);
145}
146
147static void brd_zero_page(struct brd_device *brd, sector_t sector)
148{
149 struct page *page;
150
151 page = brd_lookup_page(brd, sector);
152 if (page)
153 clear_highpage(page);
154}
155
156/*
157 * Free all backing store pages and radix tree. This must only be called when
158 * there are no other users of the device.
159 */
160#define FREE_BATCH 16
161static void brd_free_pages(struct brd_device *brd)
162{
163 unsigned long pos = 0;
164 struct page *pages[FREE_BATCH];
165 int nr_pages;
166
167 do {
168 int i;
169
170 nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
171 (void **)pages, pos, FREE_BATCH);
172
173 for (i = 0; i < nr_pages; i++) {
174 void *ret;
175
176 BUG_ON(pages[i]->index < pos);
177 pos = pages[i]->index;
178 ret = radix_tree_delete(&brd->brd_pages, pos);
179 BUG_ON(!ret || ret != pages[i]);
180 __free_page(pages[i]);
181 }
182
183 pos++;
184
185 /*
186 * This assumes radix_tree_gang_lookup always returns as
187 * many pages as possible. If the radix-tree code changes,
188 * so will this have to.
189 */
190 } while (nr_pages == FREE_BATCH);
191}
192
193/*
194 * copy_to_brd_setup must be called before copy_to_brd. It may sleep.
195 */
196static int copy_to_brd_setup(struct brd_device *brd, sector_t sector, size_t n)
197{
198 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
199 size_t copy;
200
201 copy = min_t(size_t, n, PAGE_SIZE - offset);
202 if (!brd_insert_page(brd, sector))
203 return -ENOMEM;
204 if (copy < n) {
205 sector += copy >> SECTOR_SHIFT;
206 if (!brd_insert_page(brd, sector))
207 return -ENOMEM;
208 }
209 return 0;
210}
211
212static void discard_from_brd(struct brd_device *brd,
213 sector_t sector, size_t n)
214{
215 while (n >= PAGE_SIZE) {
216 /*
217 * Don't want to actually discard pages here because
218 * re-allocating the pages can result in writeback
219 * deadlocks under heavy load.
220 */
221 if (0)
222 brd_free_page(brd, sector);
223 else
224 brd_zero_page(brd, sector);
225 sector += PAGE_SIZE >> SECTOR_SHIFT;
226 n -= PAGE_SIZE;
227 }
228}
229
230/*
231 * Copy n bytes from src to the brd starting at sector. Does not sleep.
232 */
233static void copy_to_brd(struct brd_device *brd, const void *src,
234 sector_t sector, size_t n)
235{
236 struct page *page;
237 void *dst;
238 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
239 size_t copy;
240
241 copy = min_t(size_t, n, PAGE_SIZE - offset);
242 page = brd_lookup_page(brd, sector);
243 BUG_ON(!page);
244
245 dst = kmap_atomic(page);
246 memcpy(dst + offset, src, copy);
247 kunmap_atomic(dst);
248
249 if (copy < n) {
250 src += copy;
251 sector += copy >> SECTOR_SHIFT;
252 copy = n - copy;
253 page = brd_lookup_page(brd, sector);
254 BUG_ON(!page);
255
256 dst = kmap_atomic(page);
257 memcpy(dst, src, copy);
258 kunmap_atomic(dst);
259 }
260}
261
262/*
263 * Copy n bytes to dst from the brd starting at sector. Does not sleep.
264 */
265static void copy_from_brd(void *dst, struct brd_device *brd,
266 sector_t sector, size_t n)
267{
268 struct page *page;
269 void *src;
270 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
271 size_t copy;
272
273 copy = min_t(size_t, n, PAGE_SIZE - offset);
274 page = brd_lookup_page(brd, sector);
275 if (page) {
276 src = kmap_atomic(page);
277 memcpy(dst, src + offset, copy);
278 kunmap_atomic(src);
279 } else
280 memset(dst, 0, copy);
281
282 if (copy < n) {
283 dst += copy;
284 sector += copy >> SECTOR_SHIFT;
285 copy = n - copy;
286 page = brd_lookup_page(brd, sector);
287 if (page) {
288 src = kmap_atomic(page);
289 memcpy(dst, src, copy);
290 kunmap_atomic(src);
291 } else
292 memset(dst, 0, copy);
293 }
294}
295
296/*
297 * Process a single bvec of a bio.
298 */
299static int brd_do_bvec(struct brd_device *brd, struct page *page,
300 unsigned int len, unsigned int off, int rw,
301 sector_t sector)
302{
303 void *mem;
304 int err = 0;
305
306 if (rw != READ) {
307 err = copy_to_brd_setup(brd, sector, len);
308 if (err)
309 goto out;
310 }
311
312 mem = kmap_atomic(page);
313 if (rw == READ) {
314 copy_from_brd(mem + off, brd, sector, len);
315 flush_dcache_page(page);
316 } else {
317 flush_dcache_page(page);
318 copy_to_brd(brd, mem + off, sector, len);
319 }
320 kunmap_atomic(mem);
321
322out:
323 return err;
324}
325
326static void brd_make_request(struct request_queue *q, struct bio *bio)
327{
328 struct block_device *bdev = bio->bi_bdev;
329 struct brd_device *brd = bdev->bd_disk->private_data;
330 int rw;
331 struct bio_vec bvec;
332 sector_t sector;
333 struct bvec_iter iter;
334 int err = -EIO;
335
336 sector = bio->bi_iter.bi_sector;
337 if (bio_end_sector(bio) > get_capacity(bdev->bd_disk))
338 goto out;
339
340 if (unlikely(bio->bi_rw & REQ_DISCARD)) {
341 err = 0;
342 discard_from_brd(brd, sector, bio->bi_iter.bi_size);
343 goto out;
344 }
345
346 rw = bio_rw(bio);
347 if (rw == READA)
348 rw = READ;
349
350 bio_for_each_segment(bvec, bio, iter) {
351 unsigned int len = bvec.bv_len;
352 err = brd_do_bvec(brd, bvec.bv_page, len,
353 bvec.bv_offset, rw, sector);
354 if (err)
355 break;
356 sector += len >> SECTOR_SHIFT;
357 }
358
359out:
360 bio_endio(bio, err);
361}
362
363#ifdef CONFIG_BLK_DEV_XIP
364static int brd_direct_access(struct block_device *bdev, sector_t sector,
365 void **kaddr, unsigned long *pfn)
366{
367 struct brd_device *brd = bdev->bd_disk->private_data;
368 struct page *page;
369
370 if (!brd)
371 return -ENODEV;
372 if (sector & (PAGE_SECTORS-1))
373 return -EINVAL;
374 if (sector + PAGE_SECTORS > get_capacity(bdev->bd_disk))
375 return -ERANGE;
376 page = brd_insert_page(brd, sector);
377 if (!page)
378 return -ENOMEM;
379 *kaddr = page_address(page);
380 *pfn = page_to_pfn(page);
381
382 return 0;
383}
384#endif
385
386static int brd_ioctl(struct block_device *bdev, fmode_t mode,
387 unsigned int cmd, unsigned long arg)
388{
389 int error;
390 struct brd_device *brd = bdev->bd_disk->private_data;
391
392 if (cmd != BLKFLSBUF)
393 return -ENOTTY;
394
395 /*
396 * ram device BLKFLSBUF has special semantics, we want to actually
397 * release and destroy the ramdisk data.
398 */
399 mutex_lock(&brd_mutex);
400 mutex_lock(&bdev->bd_mutex);
401 error = -EBUSY;
402 if (bdev->bd_openers <= 1) {
403 /*
404 * Kill the cache first, so it isn't written back to the
405 * device.
406 *
407 * Another thread might instantiate more buffercache here,
408 * but there is not much we can do to close that race.
409 */
410 kill_bdev(bdev);
411 brd_free_pages(brd);
412 error = 0;
413 }
414 mutex_unlock(&bdev->bd_mutex);
415 mutex_unlock(&brd_mutex);
416
417 return error;
418}
419
420static const struct block_device_operations brd_fops = {
421 .owner = THIS_MODULE,
422 .ioctl = brd_ioctl,
423#ifdef CONFIG_BLK_DEV_XIP
424 .direct_access = brd_direct_access,
425#endif
426};
427
428/*
429 * And now the modules code and kernel interface.
430 */
431static int rd_nr;
432int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
433static int max_part;
434static int part_shift;
435module_param(rd_nr, int, S_IRUGO);
436MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
437module_param(rd_size, int, S_IRUGO);
438MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
439module_param(max_part, int, S_IRUGO);
440MODULE_PARM_DESC(max_part, "Maximum number of partitions per RAM disk");
441MODULE_LICENSE("GPL");
442MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
443MODULE_ALIAS("rd");
444
445#ifndef MODULE
446/* Legacy boot options - nonmodular */
447static int __init ramdisk_size(char *str)
448{
449 rd_size = simple_strtol(str, NULL, 0);
450 return 1;
451}
452__setup("ramdisk_size=", ramdisk_size);
453#endif
454
455/*
456 * The device scheme is derived from loop.c. Keep them in synch where possible
457 * (should share code eventually).
458 */
459static LIST_HEAD(brd_devices);
460static DEFINE_MUTEX(brd_devices_mutex);
461
462static struct brd_device *brd_alloc(int i)
463{
464 struct brd_device *brd;
465 struct gendisk *disk;
466
467 brd = kzalloc(sizeof(*brd), GFP_KERNEL);
468 if (!brd)
469 goto out;
470 brd->brd_number = i;
471 spin_lock_init(&brd->brd_lock);
472 INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
473
474 brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
475 if (!brd->brd_queue)
476 goto out_free_dev;
477 blk_queue_make_request(brd->brd_queue, brd_make_request);
478 blk_queue_max_hw_sectors(brd->brd_queue, 1024);
479 blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
480
481 brd->brd_queue->limits.discard_granularity = PAGE_SIZE;
482 brd->brd_queue->limits.max_discard_sectors = UINT_MAX;
483 brd->brd_queue->limits.discard_zeroes_data = 1;
484 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, brd->brd_queue);
485
486 disk = brd->brd_disk = alloc_disk(1 << part_shift);
487 if (!disk)
488 goto out_free_queue;
489 disk->major = RAMDISK_MAJOR;
490 disk->first_minor = i << part_shift;
491 disk->fops = &brd_fops;
492 disk->private_data = brd;
493 disk->queue = brd->brd_queue;
494 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
495 sprintf(disk->disk_name, "ram%d", i);
496 set_capacity(disk, rd_size * 2);
497
498 return brd;
499
500out_free_queue:
501 blk_cleanup_queue(brd->brd_queue);
502out_free_dev:
503 kfree(brd);
504out:
505 return NULL;
506}
507
508static void brd_free(struct brd_device *brd)
509{
510 put_disk(brd->brd_disk);
511 blk_cleanup_queue(brd->brd_queue);
512 brd_free_pages(brd);
513 kfree(brd);
514}
515
516static struct brd_device *brd_init_one(int i)
517{
518 struct brd_device *brd;
519
520 list_for_each_entry(brd, &brd_devices, brd_list) {
521 if (brd->brd_number == i)
522 goto out;
523 }
524
525 brd = brd_alloc(i);
526 if (brd) {
527 add_disk(brd->brd_disk);
528 list_add_tail(&brd->brd_list, &brd_devices);
529 }
530out:
531 return brd;
532}
533
534static void brd_del_one(struct brd_device *brd)
535{
536 list_del(&brd->brd_list);
537 del_gendisk(brd->brd_disk);
538 brd_free(brd);
539}
540
541static struct kobject *brd_probe(dev_t dev, int *part, void *data)
542{
543 struct brd_device *brd;
544 struct kobject *kobj;
545
546 mutex_lock(&brd_devices_mutex);
547 brd = brd_init_one(MINOR(dev) >> part_shift);
548 kobj = brd ? get_disk(brd->brd_disk) : NULL;
549 mutex_unlock(&brd_devices_mutex);
550
551 *part = 0;
552 return kobj;
553}
554
555static int __init brd_init(void)
556{
557 int i, nr;
558 unsigned long range;
559 struct brd_device *brd, *next;
560
561 /*
562 * brd module now has a feature to instantiate underlying device
563 * structure on-demand, provided that there is an access dev node.
564 * However, this will not work well with user space tool that doesn't
565 * know about such "feature". In order to not break any existing
566 * tool, we do the following:
567 *
568 * (1) if rd_nr is specified, create that many upfront, and this
569 * also becomes a hard limit.
570 * (2) if rd_nr is not specified, create CONFIG_BLK_DEV_RAM_COUNT
571 * (default 16) rd device on module load, user can further
572 * extend brd device by create dev node themselves and have
573 * kernel automatically instantiate actual device on-demand.
574 */
575
576 part_shift = 0;
577 if (max_part > 0) {
578 part_shift = fls(max_part);
579
580 /*
581 * Adjust max_part according to part_shift as it is exported
582 * to user space so that user can decide correct minor number
583 * if [s]he want to create more devices.
584 *
585 * Note that -1 is required because partition 0 is reserved
586 * for the whole disk.
587 */
588 max_part = (1UL << part_shift) - 1;
589 }
590
591 if ((1UL << part_shift) > DISK_MAX_PARTS)
592 return -EINVAL;
593
594 if (rd_nr > 1UL << (MINORBITS - part_shift))
595 return -EINVAL;
596
597 if (rd_nr) {
598 nr = rd_nr;
599 range = rd_nr << part_shift;
600 } else {
601 nr = CONFIG_BLK_DEV_RAM_COUNT;
602 range = 1UL << MINORBITS;
603 }
604
605 if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
606 return -EIO;
607
608 for (i = 0; i < nr; i++) {
609 brd = brd_alloc(i);
610 if (!brd)
611 goto out_free;
612 list_add_tail(&brd->brd_list, &brd_devices);
613 }
614
615 /* point of no return */
616
617 list_for_each_entry(brd, &brd_devices, brd_list)
618 add_disk(brd->brd_disk);
619
620 blk_register_region(MKDEV(RAMDISK_MAJOR, 0), range,
621 THIS_MODULE, brd_probe, NULL, NULL);
622
623 printk(KERN_INFO "brd: module loaded\n");
624 return 0;
625
626out_free:
627 list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
628 list_del(&brd->brd_list);
629 brd_free(brd);
630 }
631 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
632
633 return -ENOMEM;
634}
635
636static void __exit brd_exit(void)
637{
638 unsigned long range;
639 struct brd_device *brd, *next;
640
641 range = rd_nr ? rd_nr << part_shift : 1UL << MINORBITS;
642
643 list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
644 brd_del_one(brd);
645
646 blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), range);
647 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
648}
649
650module_init(brd_init);
651module_exit(brd_exit);
652