xref: /openbmc/linux/drivers/block/brd.c (revision 732a675a)
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/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  */
35 struct 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  */
57 static 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  */
88 static struct page *brd_insert_page(struct brd_device *brd, sector_t sector)
89 {
90 	pgoff_t idx;
91 	struct page *page;
92 	gfp_t gfp_flags;
93 
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.
101 	 *
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.
106 	 */
107 	gfp_flags = GFP_NOIO | __GFP_ZERO;
108 #ifndef CONFIG_BLK_DEV_XIP
109 	gfp_flags |= __GFP_HIGHMEM;
110 #endif
111 	page = alloc_page(gfp_flags);
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
141 static 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  */
176 static 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  */
195 static 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  */
227 static 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  */
261 static 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 
282 out:
283 	return err;
284 }
285 
286 static 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 
314 out:
315 	bio_endio(bio, err);
316 
317 	return 0;
318 }
319 
320 #ifdef CONFIG_BLK_DEV_XIP
321 static int brd_direct_access (struct block_device *bdev, sector_t sector,
322 			void **kaddr, unsigned long *pfn)
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 	*kaddr = page_address(page);
337 	*pfn = page_to_pfn(page);
338 
339 	return 0;
340 }
341 #endif
342 
343 static int brd_ioctl(struct inode *inode, struct file *file,
344 			unsigned int cmd, unsigned long arg)
345 {
346 	int error;
347 	struct block_device *bdev = inode->i_bdev;
348 	struct brd_device *brd = bdev->bd_disk->private_data;
349 
350 	if (cmd != BLKFLSBUF)
351 		return -ENOTTY;
352 
353 	/*
354 	 * ram device BLKFLSBUF has special semantics, we want to actually
355 	 * release and destroy the ramdisk data.
356 	 */
357 	mutex_lock(&bdev->bd_mutex);
358 	error = -EBUSY;
359 	if (bdev->bd_openers <= 1) {
360 		/*
361 		 * Invalidate the cache first, so it isn't written
362 		 * back to the device.
363 		 *
364 		 * Another thread might instantiate more buffercache here,
365 		 * but there is not much we can do to close that race.
366 		 */
367 		invalidate_bh_lrus();
368 		truncate_inode_pages(bdev->bd_inode->i_mapping, 0);
369 		brd_free_pages(brd);
370 		error = 0;
371 	}
372 	mutex_unlock(&bdev->bd_mutex);
373 
374 	return error;
375 }
376 
377 static struct block_device_operations brd_fops = {
378 	.owner =		THIS_MODULE,
379 	.ioctl =		brd_ioctl,
380 #ifdef CONFIG_BLK_DEV_XIP
381 	.direct_access =	brd_direct_access,
382 #endif
383 };
384 
385 /*
386  * And now the modules code and kernel interface.
387  */
388 static int rd_nr;
389 int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
390 static int max_part;
391 static int part_shift;
392 module_param(rd_nr, int, 0);
393 MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
394 module_param(rd_size, int, 0);
395 MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
396 module_param(max_part, int, 0);
397 MODULE_PARM_DESC(max_part, "Maximum number of partitions per RAM disk");
398 MODULE_LICENSE("GPL");
399 MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
400 MODULE_ALIAS("rd");
401 
402 #ifndef MODULE
403 /* Legacy boot options - nonmodular */
404 static int __init ramdisk_size(char *str)
405 {
406 	rd_size = simple_strtol(str, NULL, 0);
407 	return 1;
408 }
409 static int __init ramdisk_size2(char *str)
410 {
411 	return ramdisk_size(str);
412 }
413 __setup("ramdisk=", ramdisk_size);
414 __setup("ramdisk_size=", ramdisk_size2);
415 #endif
416 
417 /*
418  * The device scheme is derived from loop.c. Keep them in synch where possible
419  * (should share code eventually).
420  */
421 static LIST_HEAD(brd_devices);
422 static DEFINE_MUTEX(brd_devices_mutex);
423 
424 static struct brd_device *brd_alloc(int i)
425 {
426 	struct brd_device *brd;
427 	struct gendisk *disk;
428 
429 	brd = kzalloc(sizeof(*brd), GFP_KERNEL);
430 	if (!brd)
431 		goto out;
432 	brd->brd_number		= i;
433 	spin_lock_init(&brd->brd_lock);
434 	INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
435 
436 	brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
437 	if (!brd->brd_queue)
438 		goto out_free_dev;
439 	blk_queue_make_request(brd->brd_queue, brd_make_request);
440 	blk_queue_max_sectors(brd->brd_queue, 1024);
441 	blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
442 
443 	disk = brd->brd_disk = alloc_disk(1 << part_shift);
444 	if (!disk)
445 		goto out_free_queue;
446 	disk->major		= RAMDISK_MAJOR;
447 	disk->first_minor	= i << part_shift;
448 	disk->fops		= &brd_fops;
449 	disk->private_data	= brd;
450 	disk->queue		= brd->brd_queue;
451 	disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
452 	sprintf(disk->disk_name, "ram%d", i);
453 	set_capacity(disk, rd_size * 2);
454 
455 	return brd;
456 
457 out_free_queue:
458 	blk_cleanup_queue(brd->brd_queue);
459 out_free_dev:
460 	kfree(brd);
461 out:
462 	return NULL;
463 }
464 
465 static void brd_free(struct brd_device *brd)
466 {
467 	put_disk(brd->brd_disk);
468 	blk_cleanup_queue(brd->brd_queue);
469 	brd_free_pages(brd);
470 	kfree(brd);
471 }
472 
473 static struct brd_device *brd_init_one(int i)
474 {
475 	struct brd_device *brd;
476 
477 	list_for_each_entry(brd, &brd_devices, brd_list) {
478 		if (brd->brd_number == i)
479 			goto out;
480 	}
481 
482 	brd = brd_alloc(i);
483 	if (brd) {
484 		add_disk(brd->brd_disk);
485 		list_add_tail(&brd->brd_list, &brd_devices);
486 	}
487 out:
488 	return brd;
489 }
490 
491 static void brd_del_one(struct brd_device *brd)
492 {
493 	list_del(&brd->brd_list);
494 	del_gendisk(brd->brd_disk);
495 	brd_free(brd);
496 }
497 
498 static struct kobject *brd_probe(dev_t dev, int *part, void *data)
499 {
500 	struct brd_device *brd;
501 	struct kobject *kobj;
502 
503 	mutex_lock(&brd_devices_mutex);
504 	brd = brd_init_one(dev & MINORMASK);
505 	kobj = brd ? get_disk(brd->brd_disk) : ERR_PTR(-ENOMEM);
506 	mutex_unlock(&brd_devices_mutex);
507 
508 	*part = 0;
509 	return kobj;
510 }
511 
512 static int __init brd_init(void)
513 {
514 	int i, nr;
515 	unsigned long range;
516 	struct brd_device *brd, *next;
517 
518 	/*
519 	 * brd module now has a feature to instantiate underlying device
520 	 * structure on-demand, provided that there is an access dev node.
521 	 * However, this will not work well with user space tool that doesn't
522 	 * know about such "feature".  In order to not break any existing
523 	 * tool, we do the following:
524 	 *
525 	 * (1) if rd_nr is specified, create that many upfront, and this
526 	 *     also becomes a hard limit.
527 	 * (2) if rd_nr is not specified, create 1 rd device on module
528 	 *     load, user can further extend brd device by create dev node
529 	 *     themselves and have kernel automatically instantiate actual
530 	 *     device on-demand.
531 	 */
532 
533 	part_shift = 0;
534 	if (max_part > 0)
535 		part_shift = fls(max_part);
536 
537 	if (rd_nr > 1UL << (MINORBITS - part_shift))
538 		return -EINVAL;
539 
540 	if (rd_nr) {
541 		nr = rd_nr;
542 		range = rd_nr;
543 	} else {
544 		nr = CONFIG_BLK_DEV_RAM_COUNT;
545 		range = 1UL << (MINORBITS - part_shift);
546 	}
547 
548 	if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
549 		return -EIO;
550 
551 	for (i = 0; i < nr; i++) {
552 		brd = brd_alloc(i);
553 		if (!brd)
554 			goto out_free;
555 		list_add_tail(&brd->brd_list, &brd_devices);
556 	}
557 
558 	/* point of no return */
559 
560 	list_for_each_entry(brd, &brd_devices, brd_list)
561 		add_disk(brd->brd_disk);
562 
563 	blk_register_region(MKDEV(RAMDISK_MAJOR, 0), range,
564 				  THIS_MODULE, brd_probe, NULL, NULL);
565 
566 	printk(KERN_INFO "brd: module loaded\n");
567 	return 0;
568 
569 out_free:
570 	list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
571 		list_del(&brd->brd_list);
572 		brd_free(brd);
573 	}
574 
575 	unregister_blkdev(RAMDISK_MAJOR, "brd");
576 	return -ENOMEM;
577 }
578 
579 static void __exit brd_exit(void)
580 {
581 	unsigned long range;
582 	struct brd_device *brd, *next;
583 
584 	range = rd_nr ? rd_nr :  1UL << (MINORBITS - part_shift);
585 
586 	list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
587 		brd_del_one(brd);
588 
589 	blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), range);
590 	unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
591 }
592 
593 module_init(brd_init);
594 module_exit(brd_exit);
595 
596