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 DAX and highmem, because our ->direct_access
101	 * routine for DAX must return memory that is always addressable.
102	 * If DAX 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_RAM_DAX
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 -ENOSPC;
204	if (copy < n) {
205		sector += copy >> SECTOR_SHIFT;
206		if (!brd_insert_page(brd, sector))
207			return -ENOSPC;
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 blk_qc_t 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
335	sector = bio->bi_iter.bi_sector;
336	if (bio_end_sector(bio) > get_capacity(bdev->bd_disk))
337		goto io_error;
338
339	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
340		if (sector & ((PAGE_SIZE >> SECTOR_SHIFT) - 1) ||
341		    bio->bi_iter.bi_size & ~PAGE_MASK)
342			goto io_error;
343		discard_from_brd(brd, sector, bio->bi_iter.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, iter) {
352		unsigned int len = bvec.bv_len;
353		int err;
354
355		err = brd_do_bvec(brd, bvec.bv_page, len,
356					bvec.bv_offset, rw, sector);
357		if (err)
358			goto io_error;
359		sector += len >> SECTOR_SHIFT;
360	}
361
362out:
363	bio_endio(bio);
364	return BLK_QC_T_NONE;
365io_error:
366	bio_io_error(bio);
367	return BLK_QC_T_NONE;
368}
369
370static int brd_rw_page(struct block_device *bdev, sector_t sector,
371		       struct page *page, int rw)
372{
373	struct brd_device *brd = bdev->bd_disk->private_data;
374	int err = brd_do_bvec(brd, page, PAGE_CACHE_SIZE, 0, rw, sector);
375	page_endio(page, rw & WRITE, err);
376	return err;
377}
378
379#ifdef CONFIG_BLK_DEV_RAM_DAX
380static long brd_direct_access(struct block_device *bdev, sector_t sector,
381			void __pmem **kaddr, unsigned long *pfn)
382{
383	struct brd_device *brd = bdev->bd_disk->private_data;
384	struct page *page;
385
386	if (!brd)
387		return -ENODEV;
388	page = brd_insert_page(brd, sector);
389	if (!page)
390		return -ENOSPC;
391	*kaddr = (void __pmem *)page_address(page);
392	*pfn = page_to_pfn(page);
393
394	return PAGE_SIZE;
395}
396#else
397#define brd_direct_access NULL
398#endif
399
400static int brd_ioctl(struct block_device *bdev, fmode_t mode,
401			unsigned int cmd, unsigned long arg)
402{
403	int error;
404	struct brd_device *brd = bdev->bd_disk->private_data;
405
406	if (cmd != BLKFLSBUF)
407		return -ENOTTY;
408
409	/*
410	 * ram device BLKFLSBUF has special semantics, we want to actually
411	 * release and destroy the ramdisk data.
412	 */
413	mutex_lock(&brd_mutex);
414	mutex_lock(&bdev->bd_mutex);
415	error = -EBUSY;
416	if (bdev->bd_openers <= 1) {
417		/*
418		 * Kill the cache first, so it isn't written back to the
419		 * device.
420		 *
421		 * Another thread might instantiate more buffercache here,
422		 * but there is not much we can do to close that race.
423		 */
424		kill_bdev(bdev);
425		brd_free_pages(brd);
426		error = 0;
427	}
428	mutex_unlock(&bdev->bd_mutex);
429	mutex_unlock(&brd_mutex);
430
431	return error;
432}
433
434static const struct block_device_operations brd_fops = {
435	.owner =		THIS_MODULE,
436	.rw_page =		brd_rw_page,
437	.ioctl =		brd_ioctl,
438	.direct_access =	brd_direct_access,
439};
440
441/*
442 * And now the modules code and kernel interface.
443 */
444static int rd_nr = CONFIG_BLK_DEV_RAM_COUNT;
445module_param(rd_nr, int, S_IRUGO);
446MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
447
448int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
449module_param(rd_size, int, S_IRUGO);
450MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
451
452static int max_part = 1;
453module_param(max_part, int, S_IRUGO);
454MODULE_PARM_DESC(max_part, "Num Minors to reserve between devices");
455
456MODULE_LICENSE("GPL");
457MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
458MODULE_ALIAS("rd");
459
460#ifndef MODULE
461/* Legacy boot options - nonmodular */
462static int __init ramdisk_size(char *str)
463{
464	rd_size = simple_strtol(str, NULL, 0);
465	return 1;
466}
467__setup("ramdisk_size=", ramdisk_size);
468#endif
469
470/*
471 * The device scheme is derived from loop.c. Keep them in synch where possible
472 * (should share code eventually).
473 */
474static LIST_HEAD(brd_devices);
475static DEFINE_MUTEX(brd_devices_mutex);
476
477static struct brd_device *brd_alloc(int i)
478{
479	struct brd_device *brd;
480	struct gendisk *disk;
481
482	brd = kzalloc(sizeof(*brd), GFP_KERNEL);
483	if (!brd)
484		goto out;
485	brd->brd_number		= i;
486	spin_lock_init(&brd->brd_lock);
487	INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
488
489	brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
490	if (!brd->brd_queue)
491		goto out_free_dev;
492
493	blk_queue_make_request(brd->brd_queue, brd_make_request);
494	blk_queue_max_hw_sectors(brd->brd_queue, 1024);
495	blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
496
497	/* This is so fdisk will align partitions on 4k, because of
498	 * direct_access API needing 4k alignment, returning a PFN
499	 * (This is only a problem on very small devices <= 4M,
500	 *  otherwise fdisk will align on 1M. Regardless this call
501	 *  is harmless)
502	 */
503	blk_queue_physical_block_size(brd->brd_queue, PAGE_SIZE);
504
505	brd->brd_queue->limits.discard_granularity = PAGE_SIZE;
506	blk_queue_max_discard_sectors(brd->brd_queue, UINT_MAX);
507	brd->brd_queue->limits.discard_zeroes_data = 1;
508	queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, brd->brd_queue);
509
510	disk = brd->brd_disk = alloc_disk(max_part);
511	if (!disk)
512		goto out_free_queue;
513	disk->major		= RAMDISK_MAJOR;
514	disk->first_minor	= i * max_part;
515	disk->fops		= &brd_fops;
516	disk->private_data	= brd;
517	disk->queue		= brd->brd_queue;
518	disk->flags		= GENHD_FL_EXT_DEVT;
519	sprintf(disk->disk_name, "ram%d", i);
520	set_capacity(disk, rd_size * 2);
521
522	return brd;
523
524out_free_queue:
525	blk_cleanup_queue(brd->brd_queue);
526out_free_dev:
527	kfree(brd);
528out:
529	return NULL;
530}
531
532static void brd_free(struct brd_device *brd)
533{
534	put_disk(brd->brd_disk);
535	blk_cleanup_queue(brd->brd_queue);
536	brd_free_pages(brd);
537	kfree(brd);
538}
539
540static struct brd_device *brd_init_one(int i, bool *new)
541{
542	struct brd_device *brd;
543
544	*new = false;
545	list_for_each_entry(brd, &brd_devices, brd_list) {
546		if (brd->brd_number == i)
547			goto out;
548	}
549
550	brd = brd_alloc(i);
551	if (brd) {
552		add_disk(brd->brd_disk);
553		list_add_tail(&brd->brd_list, &brd_devices);
554	}
555	*new = true;
556out:
557	return brd;
558}
559
560static void brd_del_one(struct brd_device *brd)
561{
562	list_del(&brd->brd_list);
563	del_gendisk(brd->brd_disk);
564	brd_free(brd);
565}
566
567static struct kobject *brd_probe(dev_t dev, int *part, void *data)
568{
569	struct brd_device *brd;
570	struct kobject *kobj;
571	bool new;
572
573	mutex_lock(&brd_devices_mutex);
574	brd = brd_init_one(MINOR(dev) / max_part, &new);
575	kobj = brd ? get_disk(brd->brd_disk) : NULL;
576	mutex_unlock(&brd_devices_mutex);
577
578	if (new)
579		*part = 0;
580
581	return kobj;
582}
583
584static int __init brd_init(void)
585{
586	struct brd_device *brd, *next;
587	int i;
588
589	/*
590	 * brd module now has a feature to instantiate underlying device
591	 * structure on-demand, provided that there is an access dev node.
592	 *
593	 * (1) if rd_nr is specified, create that many upfront. else
594	 *     it defaults to CONFIG_BLK_DEV_RAM_COUNT
595	 * (2) User can further extend brd devices by create dev node themselves
596	 *     and have kernel automatically instantiate actual device
597	 *     on-demand. Example:
598	 *		mknod /path/devnod_name b 1 X	# 1 is the rd major
599	 *		fdisk -l /path/devnod_name
600	 *	If (X / max_part) was not already created it will be created
601	 *	dynamically.
602	 */
603
604	if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
605		return -EIO;
606
607	if (unlikely(!max_part))
608		max_part = 1;
609
610	for (i = 0; i < rd_nr; i++) {
611		brd = brd_alloc(i);
612		if (!brd)
613			goto out_free;
614		list_add_tail(&brd->brd_list, &brd_devices);
615	}
616
617	/* point of no return */
618
619	list_for_each_entry(brd, &brd_devices, brd_list)
620		add_disk(brd->brd_disk);
621
622	blk_register_region(MKDEV(RAMDISK_MAJOR, 0), 1UL << MINORBITS,
623				  THIS_MODULE, brd_probe, NULL, NULL);
624
625	pr_info("brd: module loaded\n");
626	return 0;
627
628out_free:
629	list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
630		list_del(&brd->brd_list);
631		brd_free(brd);
632	}
633	unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
634
635	pr_info("brd: module NOT loaded !!!\n");
636	return -ENOMEM;
637}
638
639static void __exit brd_exit(void)
640{
641	struct brd_device *brd, *next;
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), 1UL << MINORBITS);
647	unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
648
649	pr_info("brd: module unloaded\n");
650}
651
652module_init(brd_init);
653module_exit(brd_exit);
654
655