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#ifdef CONFIG_BLK_DEV_RAM_DAX
23#include <linux/pfn_t.h>
24#endif
25
26#include <asm/uaccess.h>
27
28#define SECTOR_SHIFT 9
29#define PAGE_SECTORS_SHIFT (PAGE_SHIFT - SECTOR_SHIFT)
30#define PAGE_SECTORS (1 << PAGE_SECTORS_SHIFT)
31
32/*
33 * Each block ramdisk device has a radix_tree brd_pages of pages that stores
34 * the pages containing the block device's contents. A brd page's ->index is
35 * its offset in PAGE_SIZE units. This is similar to, but in no way connected
36 * with, the kernel's pagecache or buffer cache (which sit above our block
37 * device).
38 */
39struct brd_device {
40 int brd_number;
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 DEFINE_MUTEX(brd_mutex);
58static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
59{
60 pgoff_t idx;
61 struct page *page;
62
63 /*
64 * The page lifetime is protected by the fact that we have opened the
65 * device node -- brd pages will never be deleted under us, so we
66 * don't need any further locking or refcounting.
67 *
68 * This is strictly true for the radix-tree nodes as well (ie. we
69 * don't actually need the rcu_read_lock()), however that is not a
70 * documented feature of the radix-tree API so it is better to be
71 * safe here (we don't have total exclusion from radix tree updates
72 * here, only deletes).
73 */
74 rcu_read_lock();
75 idx = sector >> PAGE_SECTORS_SHIFT; /* sector to page index */
76 page = radix_tree_lookup(&brd->brd_pages, idx);
77 rcu_read_unlock();
78
79 BUG_ON(page && page->index != idx);
80
81 return page;
82}
83
84/*
85 * Look up and return a brd's page for a given sector.
86 * If one does not exist, allocate an empty page, and insert that. Then
87 * return it.
88 */
89static struct page *brd_insert_page(struct brd_device *brd, sector_t sector)
90{
91 pgoff_t idx;
92 struct page *page;
93 gfp_t gfp_flags;
94
95 page = brd_lookup_page(brd, sector);
96 if (page)
97 return page;
98
99 /*
100 * Must use NOIO because we don't want to recurse back into the
101 * block or filesystem layers from page reclaim.
102 *
103 * Cannot support DAX and highmem, because our ->direct_access
104 * routine for DAX must return memory that is always addressable.
105 * If DAX was reworked to use pfns and kmap throughout, this
106 * restriction might be able to be lifted.
107 */
108 gfp_flags = GFP_NOIO | __GFP_ZERO;
109#ifndef CONFIG_BLK_DEV_RAM_DAX
110 gfp_flags |= __GFP_HIGHMEM;
111#endif
112 page = alloc_page(gfp_flags);
113 if (!page)
114 return NULL;
115
116 if (radix_tree_preload(GFP_NOIO)) {
117 __free_page(page);
118 return NULL;
119 }
120
121 spin_lock(&brd->brd_lock);
122 idx = sector >> PAGE_SECTORS_SHIFT;
123 page->index = idx;
124 if (radix_tree_insert(&brd->brd_pages, idx, page)) {
125 __free_page(page);
126 page = radix_tree_lookup(&brd->brd_pages, idx);
127 BUG_ON(!page);
128 BUG_ON(page->index != idx);
129 }
130 spin_unlock(&brd->brd_lock);
131
132 radix_tree_preload_end();
133
134 return page;
135}
136
137static void brd_free_page(struct brd_device *brd, sector_t sector)
138{
139 struct page *page;
140 pgoff_t idx;
141
142 spin_lock(&brd->brd_lock);
143 idx = sector >> PAGE_SECTORS_SHIFT;
144 page = radix_tree_delete(&brd->brd_pages, idx);
145 spin_unlock(&brd->brd_lock);
146 if (page)
147 __free_page(page);
148}
149
150static void brd_zero_page(struct brd_device *brd, sector_t sector)
151{
152 struct page *page;
153
154 page = brd_lookup_page(brd, sector);
155 if (page)
156 clear_highpage(page);
157}
158
159/*
160 * Free all backing store pages and radix tree. This must only be called when
161 * there are no other users of the device.
162 */
163#define FREE_BATCH 16
164static void brd_free_pages(struct brd_device *brd)
165{
166 unsigned long pos = 0;
167 struct page *pages[FREE_BATCH];
168 int nr_pages;
169
170 do {
171 int i;
172
173 nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
174 (void **)pages, pos, FREE_BATCH);
175
176 for (i = 0; i < nr_pages; i++) {
177 void *ret;
178
179 BUG_ON(pages[i]->index < pos);
180 pos = pages[i]->index;
181 ret = radix_tree_delete(&brd->brd_pages, pos);
182 BUG_ON(!ret || ret != pages[i]);
183 __free_page(pages[i]);
184 }
185
186 pos++;
187
188 /*
189 * This assumes radix_tree_gang_lookup always returns as
190 * many pages as possible. If the radix-tree code changes,
191 * so will this have to.
192 */
193 } while (nr_pages == FREE_BATCH);
194}
195
196/*
197 * copy_to_brd_setup must be called before copy_to_brd. It may sleep.
198 */
199static int copy_to_brd_setup(struct brd_device *brd, sector_t sector, size_t n)
200{
201 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
202 size_t copy;
203
204 copy = min_t(size_t, n, PAGE_SIZE - offset);
205 if (!brd_insert_page(brd, sector))
206 return -ENOSPC;
207 if (copy < n) {
208 sector += copy >> SECTOR_SHIFT;
209 if (!brd_insert_page(brd, sector))
210 return -ENOSPC;
211 }
212 return 0;
213}
214
215static void discard_from_brd(struct brd_device *brd,
216 sector_t sector, size_t n)
217{
218 while (n >= PAGE_SIZE) {
219 /*
220 * Don't want to actually discard pages here because
221 * re-allocating the pages can result in writeback
222 * deadlocks under heavy load.
223 */
224 if (0)
225 brd_free_page(brd, sector);
226 else
227 brd_zero_page(brd, sector);
228 sector += PAGE_SIZE >> SECTOR_SHIFT;
229 n -= PAGE_SIZE;
230 }
231}
232
233/*
234 * Copy n bytes from src to the brd starting at sector. Does not sleep.
235 */
236static void copy_to_brd(struct brd_device *brd, const void *src,
237 sector_t sector, size_t n)
238{
239 struct page *page;
240 void *dst;
241 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
242 size_t copy;
243
244 copy = min_t(size_t, n, PAGE_SIZE - offset);
245 page = brd_lookup_page(brd, sector);
246 BUG_ON(!page);
247
248 dst = kmap_atomic(page);
249 memcpy(dst + offset, src, copy);
250 kunmap_atomic(dst);
251
252 if (copy < n) {
253 src += copy;
254 sector += copy >> SECTOR_SHIFT;
255 copy = n - copy;
256 page = brd_lookup_page(brd, sector);
257 BUG_ON(!page);
258
259 dst = kmap_atomic(page);
260 memcpy(dst, src, copy);
261 kunmap_atomic(dst);
262 }
263}
264
265/*
266 * Copy n bytes to dst from the brd starting at sector. Does not sleep.
267 */
268static void copy_from_brd(void *dst, struct brd_device *brd,
269 sector_t sector, size_t n)
270{
271 struct page *page;
272 void *src;
273 unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
274 size_t copy;
275
276 copy = min_t(size_t, n, PAGE_SIZE - offset);
277 page = brd_lookup_page(brd, sector);
278 if (page) {
279 src = kmap_atomic(page);
280 memcpy(dst, src + offset, copy);
281 kunmap_atomic(src);
282 } else
283 memset(dst, 0, copy);
284
285 if (copy < n) {
286 dst += copy;
287 sector += copy >> SECTOR_SHIFT;
288 copy = n - copy;
289 page = brd_lookup_page(brd, sector);
290 if (page) {
291 src = kmap_atomic(page);
292 memcpy(dst, src, copy);
293 kunmap_atomic(src);
294 } else
295 memset(dst, 0, copy);
296 }
297}
298
299/*
300 * Process a single bvec of a bio.
301 */
302static int brd_do_bvec(struct brd_device *brd, struct page *page,
303 unsigned int len, unsigned int off, int rw,
304 sector_t sector)
305{
306 void *mem;
307 int err = 0;
308
309 if (rw != READ) {
310 err = copy_to_brd_setup(brd, sector, len);
311 if (err)
312 goto out;
313 }
314
315 mem = kmap_atomic(page);
316 if (rw == READ) {
317 copy_from_brd(mem + off, brd, sector, len);
318 flush_dcache_page(page);
319 } else {
320 flush_dcache_page(page);
321 copy_to_brd(brd, mem + off, sector, len);
322 }
323 kunmap_atomic(mem);
324
325out:
326 return err;
327}
328
329static blk_qc_t brd_make_request(struct request_queue *q, struct bio *bio)
330{
331 struct block_device *bdev = bio->bi_bdev;
332 struct brd_device *brd = bdev->bd_disk->private_data;
333 int rw;
334 struct bio_vec bvec;
335 sector_t sector;
336 struct bvec_iter iter;
337
338 sector = bio->bi_iter.bi_sector;
339 if (bio_end_sector(bio) > get_capacity(bdev->bd_disk))
340 goto io_error;
341
342 if (unlikely(bio->bi_rw & REQ_DISCARD)) {
343 if (sector & ((PAGE_SIZE >> SECTOR_SHIFT) - 1) ||
344 bio->bi_iter.bi_size & ~PAGE_MASK)
345 goto io_error;
346 discard_from_brd(brd, sector, bio->bi_iter.bi_size);
347 goto out;
348 }
349
350 rw = bio_rw(bio);
351 if (rw == READA)
352 rw = READ;
353
354 bio_for_each_segment(bvec, bio, iter) {
355 unsigned int len = bvec.bv_len;
356 int err;
357
358 err = brd_do_bvec(brd, bvec.bv_page, len,
359 bvec.bv_offset, rw, sector);
360 if (err)
361 goto io_error;
362 sector += len >> SECTOR_SHIFT;
363 }
364
365out:
366 bio_endio(bio);
367 return BLK_QC_T_NONE;
368io_error:
369 bio_io_error(bio);
370 return BLK_QC_T_NONE;
371}
372
373static int brd_rw_page(struct block_device *bdev, sector_t sector,
374 struct page *page, int rw)
375{
376 struct brd_device *brd = bdev->bd_disk->private_data;
377 int err = brd_do_bvec(brd, page, PAGE_SIZE, 0, rw, sector);
378 page_endio(page, rw & WRITE, err);
379 return err;
380}
381
382#ifdef CONFIG_BLK_DEV_RAM_DAX
383static long brd_direct_access(struct block_device *bdev, sector_t sector,
384 void __pmem **kaddr, pfn_t *pfn)
385{
386 struct brd_device *brd = bdev->bd_disk->private_data;
387 struct page *page;
388
389 if (!brd)
390 return -ENODEV;
391 page = brd_insert_page(brd, sector);
392 if (!page)
393 return -ENOSPC;
394 *kaddr = (void __pmem *)page_address(page);
395 *pfn = page_to_pfn_t(page);
396
397 return PAGE_SIZE;
398}
399#else
400#define brd_direct_access NULL
401#endif
402
403static int brd_ioctl(struct block_device *bdev, fmode_t mode,
404 unsigned int cmd, unsigned long arg)
405{
406 int error;
407 struct brd_device *brd = bdev->bd_disk->private_data;
408
409 if (cmd != BLKFLSBUF)
410 return -ENOTTY;
411
412 /*
413 * ram device BLKFLSBUF has special semantics, we want to actually
414 * release and destroy the ramdisk data.
415 */
416 mutex_lock(&brd_mutex);
417 mutex_lock(&bdev->bd_mutex);
418 error = -EBUSY;
419 if (bdev->bd_openers <= 1) {
420 /*
421 * Kill the cache first, so it isn't written back to the
422 * device.
423 *
424 * Another thread might instantiate more buffercache here,
425 * but there is not much we can do to close that race.
426 */
427 kill_bdev(bdev);
428 brd_free_pages(brd);
429 error = 0;
430 }
431 mutex_unlock(&bdev->bd_mutex);
432 mutex_unlock(&brd_mutex);
433
434 return error;
435}
436
437static const struct block_device_operations brd_fops = {
438 .owner = THIS_MODULE,
439 .rw_page = brd_rw_page,
440 .ioctl = brd_ioctl,
441 .direct_access = brd_direct_access,
442};
443
444/*
445 * And now the modules code and kernel interface.
446 */
447static int rd_nr = CONFIG_BLK_DEV_RAM_COUNT;
448module_param(rd_nr, int, S_IRUGO);
449MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
450
451int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
452module_param(rd_size, int, S_IRUGO);
453MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
454
455static int max_part = 1;
456module_param(max_part, int, S_IRUGO);
457MODULE_PARM_DESC(max_part, "Num Minors to reserve between devices");
458
459MODULE_LICENSE("GPL");
460MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
461MODULE_ALIAS("rd");
462
463#ifndef MODULE
464/* Legacy boot options - nonmodular */
465static int __init ramdisk_size(char *str)
466{
467 rd_size = simple_strtol(str, NULL, 0);
468 return 1;
469}
470__setup("ramdisk_size=", ramdisk_size);
471#endif
472
473/*
474 * The device scheme is derived from loop.c. Keep them in synch where possible
475 * (should share code eventually).
476 */
477static LIST_HEAD(brd_devices);
478static DEFINE_MUTEX(brd_devices_mutex);
479
480static struct brd_device *brd_alloc(int i)
481{
482 struct brd_device *brd;
483 struct gendisk *disk;
484
485 brd = kzalloc(sizeof(*brd), GFP_KERNEL);
486 if (!brd)
487 goto out;
488 brd->brd_number = i;
489 spin_lock_init(&brd->brd_lock);
490 INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
491
492 brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
493 if (!brd->brd_queue)
494 goto out_free_dev;
495
496 blk_queue_make_request(brd->brd_queue, brd_make_request);
497 blk_queue_max_hw_sectors(brd->brd_queue, 1024);
498 blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
499
500 /* This is so fdisk will align partitions on 4k, because of
501 * direct_access API needing 4k alignment, returning a PFN
502 * (This is only a problem on very small devices <= 4M,
503 * otherwise fdisk will align on 1M. Regardless this call
504 * is harmless)
505 */
506 blk_queue_physical_block_size(brd->brd_queue, PAGE_SIZE);
507
508 brd->brd_queue->limits.discard_granularity = PAGE_SIZE;
509 blk_queue_max_discard_sectors(brd->brd_queue, UINT_MAX);
510 brd->brd_queue->limits.discard_zeroes_data = 1;
511 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, brd->brd_queue);
512
513 disk = brd->brd_disk = alloc_disk(max_part);
514 if (!disk)
515 goto out_free_queue;
516 disk->major = RAMDISK_MAJOR;
517 disk->first_minor = i * max_part;
518 disk->fops = &brd_fops;
519 disk->private_data = brd;
520 disk->queue = brd->brd_queue;
521 disk->flags = GENHD_FL_EXT_DEVT;
522 sprintf(disk->disk_name, "ram%d", i);
523 set_capacity(disk, rd_size * 2);
524
525 return brd;
526
527out_free_queue:
528 blk_cleanup_queue(brd->brd_queue);
529out_free_dev:
530 kfree(brd);
531out:
532 return NULL;
533}
534
535static void brd_free(struct brd_device *brd)
536{
537 put_disk(brd->brd_disk);
538 blk_cleanup_queue(brd->brd_queue);
539 brd_free_pages(brd);
540 kfree(brd);
541}
542
543static struct brd_device *brd_init_one(int i, bool *new)
544{
545 struct brd_device *brd;
546
547 *new = false;
548 list_for_each_entry(brd, &brd_devices, brd_list) {
549 if (brd->brd_number == i)
550 goto out;
551 }
552
553 brd = brd_alloc(i);
554 if (brd) {
555 add_disk(brd->brd_disk);
556 list_add_tail(&brd->brd_list, &brd_devices);
557 }
558 *new = true;
559out:
560 return brd;
561}
562
563static void brd_del_one(struct brd_device *brd)
564{
565 list_del(&brd->brd_list);
566 del_gendisk(brd->brd_disk);
567 brd_free(brd);
568}
569
570static struct kobject *brd_probe(dev_t dev, int *part, void *data)
571{
572 struct brd_device *brd;
573 struct kobject *kobj;
574 bool new;
575
576 mutex_lock(&brd_devices_mutex);
577 brd = brd_init_one(MINOR(dev) / max_part, &new);
578 kobj = brd ? get_disk(brd->brd_disk) : NULL;
579 mutex_unlock(&brd_devices_mutex);
580
581 if (new)
582 *part = 0;
583
584 return kobj;
585}
586
587static int __init brd_init(void)
588{
589 struct brd_device *brd, *next;
590 int i;
591
592 /*
593 * brd module now has a feature to instantiate underlying device
594 * structure on-demand, provided that there is an access dev node.
595 *
596 * (1) if rd_nr is specified, create that many upfront. else
597 * it defaults to CONFIG_BLK_DEV_RAM_COUNT
598 * (2) User can further extend brd devices by create dev node themselves
599 * and have kernel automatically instantiate actual device
600 * on-demand. Example:
601 * mknod /path/devnod_name b 1 X # 1 is the rd major
602 * fdisk -l /path/devnod_name
603 * If (X / max_part) was not already created it will be created
604 * dynamically.
605 */
606
607 if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
608 return -EIO;
609
610 if (unlikely(!max_part))
611 max_part = 1;
612
613 for (i = 0; i < rd_nr; i++) {
614 brd = brd_alloc(i);
615 if (!brd)
616 goto out_free;
617 list_add_tail(&brd->brd_list, &brd_devices);
618 }
619
620 /* point of no return */
621
622 list_for_each_entry(brd, &brd_devices, brd_list)
623 add_disk(brd->brd_disk);
624
625 blk_register_region(MKDEV(RAMDISK_MAJOR, 0), 1UL << MINORBITS,
626 THIS_MODULE, brd_probe, NULL, NULL);
627
628 pr_info("brd: module loaded\n");
629 return 0;
630
631out_free:
632 list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
633 list_del(&brd->brd_list);
634 brd_free(brd);
635 }
636 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
637
638 pr_info("brd: module NOT loaded !!!\n");
639 return -ENOMEM;
640}
641
642static void __exit brd_exit(void)
643{
644 struct brd_device *brd, *next;
645
646 list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
647 brd_del_one(brd);
648
649 blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), 1UL << MINORBITS);
650 unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
651
652 pr_info("brd: module unloaded\n");
653}
654
655module_init(brd_init);
656module_exit(brd_exit);
657