brd.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Ram backed block device driver.
  4 *
  5 * Copyright (C) 2007 Nick Piggin
  6 * Copyright (C) 2007 Novell Inc.
  7 *
  8 * Parts derived from drivers/block/rd.c, and drivers/block/loop.c, copyright
  9 * of their respective owners.
 10 */
 11
 12#include <linux/init.h>
 13#include <linux/initrd.h>
 14#include <linux/module.h>
 15#include <linux/moduleparam.h>
 16#include <linux/major.h>
 17#include <linux/blkdev.h>
 18#include <linux/bio.h>
 19#include <linux/highmem.h>
 20#include <linux/mutex.h>
 21#include <linux/pagemap.h>
 22#include <linux/xarray.h>
 23#include <linux/fs.h>
 24#include <linux/slab.h>
 25#include <linux/backing-dev.h>
 26#include <linux/debugfs.h>
 27
 28#include <linux/uaccess.h>
 29
 30/*
 31 * Each block ramdisk device has a xarray brd_pages of pages that stores
 32 * the pages containing the block device's contents.
 33 */
 34struct brd_device {
 35	int			brd_number;
 36	struct gendisk		*brd_disk;
 37	struct list_head	brd_list;
 38
 39	/*
 40	 * Backing store of pages. This is the contents of the block device.
 41	 */
 42	struct xarray	        brd_pages;
 43	u64			brd_nr_pages;
 44};
 45
 46/*
 47 * Look up and return a brd's page with reference grabbed for a given sector.
 48 */
 49static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
 50{
 51	struct page *page;
 52	XA_STATE(xas, &brd->brd_pages, sector >> PAGE_SECTORS_SHIFT);
 53
 54	rcu_read_lock();
 55repeat:
 56	page = xas_load(&xas);
 57	if (xas_retry(&xas, page)) {
 58		xas_reset(&xas);
 59		goto repeat;
 60	}
 61
 62	if (!page)
 63		goto out;
 64
 65	if (!get_page_unless_zero(page)) {
 66		xas_reset(&xas);
 67		goto repeat;
 68	}
 69
 70	if (unlikely(page != xas_reload(&xas))) {
 71		put_page(page);
 72		xas_reset(&xas);
 73		goto repeat;
 74	}
 75out:
 76	rcu_read_unlock();
 77
 78	return page;
 79}
 80
 81/*
 82 * Insert a new page for a given sector, if one does not already exist.
 83 * The returned page will grab reference.
 84 */
 85static struct page *brd_insert_page(struct brd_device *brd, sector_t sector,
 86		blk_opf_t opf)
 87{
 88	gfp_t gfp = (opf & REQ_NOWAIT) ? GFP_NOWAIT : GFP_NOIO;
 89	struct page *page, *ret;
 90
 91	page = alloc_page(gfp | __GFP_ZERO | __GFP_HIGHMEM);
 92	if (!page)
 93		return ERR_PTR(-ENOMEM);
 94
 95	xa_lock(&brd->brd_pages);
 96	ret = __xa_cmpxchg(&brd->brd_pages, sector >> PAGE_SECTORS_SHIFT, NULL,
 97			page, gfp);
 98	if (!ret) {
 99		brd->brd_nr_pages++;
100		get_page(page);
101		xa_unlock(&brd->brd_pages);
102		return page;
103	}
104
105	if (!xa_is_err(ret)) {
106		get_page(ret);
107		xa_unlock(&brd->brd_pages);
108		put_page(page);
109		return ret;
110	}
111
112	xa_unlock(&brd->brd_pages);
113	put_page(page);
114	return ERR_PTR(xa_err(ret));
115}
116
117/*
118 * Free all backing store pages and xarray. This must only be called when
119 * there are no other users of the device.
120 */
121static void brd_free_pages(struct brd_device *brd)
122{
123	struct page *page;
124	pgoff_t idx;
125
126	xa_for_each(&brd->brd_pages, idx, page) {
127		put_page(page);
128		cond_resched();
129	}
130
131	xa_destroy(&brd->brd_pages);
132}
133
134/*
135 * Process a single segment.  The segment is capped to not cross page boundaries
136 * in both the bio and the brd backing memory.
137 */
138static bool brd_rw_bvec(struct brd_device *brd, struct bio *bio)
139{
140	struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
141	sector_t sector = bio->bi_iter.bi_sector;
142	u32 offset = (sector & (PAGE_SECTORS - 1)) << SECTOR_SHIFT;
143	blk_opf_t opf = bio->bi_opf;
144	struct page *page;
145	void *kaddr;
146
147	bv.bv_len = min_t(u32, bv.bv_len, PAGE_SIZE - offset);
148
149	page = brd_lookup_page(brd, sector);
150	if (!page && op_is_write(opf)) {
151		page = brd_insert_page(brd, sector, opf);
152		if (IS_ERR(page))
153			goto out_error;
154	}
155
156	kaddr = bvec_kmap_local(&bv);
157	if (op_is_write(opf)) {
158		memcpy_to_page(page, offset, kaddr, bv.bv_len);
159	} else {
160		if (page)
161			memcpy_from_page(kaddr, page, offset, bv.bv_len);
162		else
163			memset(kaddr, 0, bv.bv_len);
164	}
165	kunmap_local(kaddr);
166
167	bio_advance_iter_single(bio, &bio->bi_iter, bv.bv_len);
168	if (page)
169		put_page(page);
170	return true;
171
172out_error:
173	if (PTR_ERR(page) == -ENOMEM && (opf & REQ_NOWAIT))
174		bio_wouldblock_error(bio);
175	else
176		bio_io_error(bio);
177	return false;
178}
179
180static void brd_do_discard(struct brd_device *brd, sector_t sector, u32 size)
181{
182	sector_t aligned_sector = round_up(sector, PAGE_SECTORS);
183	sector_t aligned_end = round_down(
184			sector + (size >> SECTOR_SHIFT), PAGE_SECTORS);
185	struct page *page;
186
187	if (aligned_end <= aligned_sector)
188		return;
189
190	xa_lock(&brd->brd_pages);
191	while (aligned_sector < aligned_end && aligned_sector < rd_size * 2) {
192		page = __xa_erase(&brd->brd_pages, aligned_sector >> PAGE_SECTORS_SHIFT);
193		if (page) {
194			put_page(page);
195			brd->brd_nr_pages--;
196		}
197		aligned_sector += PAGE_SECTORS;
198	}
199	xa_unlock(&brd->brd_pages);
200}
201
202static void brd_submit_bio(struct bio *bio)
203{
204	struct brd_device *brd = bio->bi_bdev->bd_disk->private_data;
205
206	if (unlikely(op_is_discard(bio->bi_opf))) {
207		brd_do_discard(brd, bio->bi_iter.bi_sector,
208				bio->bi_iter.bi_size);
209		bio_endio(bio);
210		return;
211	}
212
213	do {
214		if (!brd_rw_bvec(brd, bio))
215			return;
216	} while (bio->bi_iter.bi_size);
217
218	bio_endio(bio);
219}
220
221static const struct block_device_operations brd_fops = {
222	.owner =		THIS_MODULE,
223	.submit_bio =		brd_submit_bio,
224};
225
226/*
227 * And now the modules code and kernel interface.
228 */
229static int rd_nr = CONFIG_BLK_DEV_RAM_COUNT;
230module_param(rd_nr, int, 0444);
231MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
232
233unsigned long rd_size = CONFIG_BLK_DEV_RAM_SIZE;
234module_param(rd_size, ulong, 0444);
235MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
236
237static int max_part = 1;
238module_param(max_part, int, 0444);
239MODULE_PARM_DESC(max_part, "Num Minors to reserve between devices");
240
241MODULE_DESCRIPTION("Ram backed block device driver");
242MODULE_LICENSE("GPL");
243MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
244MODULE_ALIAS("rd");
245
246#ifndef MODULE
247/* Legacy boot options - nonmodular */
248static int __init ramdisk_size(char *str)
249{
250	rd_size = simple_strtol(str, NULL, 0);
251	return 1;
252}
253__setup("ramdisk_size=", ramdisk_size);
254#endif
255
256/*
257 * The device scheme is derived from loop.c. Keep them in synch where possible
258 * (should share code eventually).
259 */
260static LIST_HEAD(brd_devices);
261static DEFINE_MUTEX(brd_devices_mutex);
262static struct dentry *brd_debugfs_dir;
263
264static struct brd_device *brd_find_or_alloc_device(int i)
265{
266	struct brd_device *brd;
267
268	mutex_lock(&brd_devices_mutex);
269	list_for_each_entry(brd, &brd_devices, brd_list) {
270		if (brd->brd_number == i) {
271			mutex_unlock(&brd_devices_mutex);
272			return ERR_PTR(-EEXIST);
273		}
274	}
275
276	brd = kzalloc(sizeof(*brd), GFP_KERNEL);
277	if (!brd) {
278		mutex_unlock(&brd_devices_mutex);
279		return ERR_PTR(-ENOMEM);
280	}
281	brd->brd_number	= i;
282	list_add_tail(&brd->brd_list, &brd_devices);
283	mutex_unlock(&brd_devices_mutex);
284	return brd;
285}
286
287static void brd_free_device(struct brd_device *brd)
288{
289	mutex_lock(&brd_devices_mutex);
290	list_del(&brd->brd_list);
291	mutex_unlock(&brd_devices_mutex);
292	kfree(brd);
293}
294
295static int brd_alloc(int i)
296{
297	struct brd_device *brd;
298	struct gendisk *disk;
299	char buf[DISK_NAME_LEN];
300	int err = -ENOMEM;
301	struct queue_limits lim = {
302		/*
303		 * This is so fdisk will align partitions on 4k, because of
304		 * direct_access API needing 4k alignment, returning a PFN
305		 * (This is only a problem on very small devices <= 4M,
306		 *  otherwise fdisk will align on 1M. Regardless this call
307		 *  is harmless)
308		 */
309		.physical_block_size	= PAGE_SIZE,
310		.max_hw_discard_sectors	= UINT_MAX,
311		.max_discard_segments	= 1,
312		.discard_granularity	= PAGE_SIZE,
313		.features		= BLK_FEAT_SYNCHRONOUS |
314					  BLK_FEAT_NOWAIT,
315	};
316
317	brd = brd_find_or_alloc_device(i);
318	if (IS_ERR(brd))
319		return PTR_ERR(brd);
320
321	xa_init(&brd->brd_pages);
322
323	snprintf(buf, DISK_NAME_LEN, "ram%d", i);
324	if (!IS_ERR_OR_NULL(brd_debugfs_dir))
325		debugfs_create_u64(buf, 0444, brd_debugfs_dir,
326				&brd->brd_nr_pages);
327
328	disk = brd->brd_disk = blk_alloc_disk(&lim, NUMA_NO_NODE);
329	if (IS_ERR(disk)) {
330		err = PTR_ERR(disk);
331		goto out_free_dev;
332	}
333	disk->major		= RAMDISK_MAJOR;
334	disk->first_minor	= i * max_part;
335	disk->minors		= max_part;
336	disk->fops		= &brd_fops;
337	disk->private_data	= brd;
338	strscpy(disk->disk_name, buf, DISK_NAME_LEN);
339	set_capacity(disk, rd_size * 2);
340	
341	err = add_disk(disk);
342	if (err)
343		goto out_cleanup_disk;
344
345	return 0;
346
347out_cleanup_disk:
348	put_disk(disk);
349out_free_dev:
350	brd_free_device(brd);
351	return err;
352}
353
354static void brd_probe(dev_t dev)
355{
356	brd_alloc(MINOR(dev) / max_part);
357}
358
359static void brd_cleanup(void)
360{
361	struct brd_device *brd, *next;
362
363	debugfs_remove_recursive(brd_debugfs_dir);
364
365	list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
366		del_gendisk(brd->brd_disk);
367		put_disk(brd->brd_disk);
368		brd_free_pages(brd);
369		brd_free_device(brd);
370	}
371}
372
373static inline void brd_check_and_reset_par(void)
374{
375	if (unlikely(!max_part))
376		max_part = 1;
377
378	/*
379	 * make sure 'max_part' can be divided exactly by (1U << MINORBITS),
380	 * otherwise, it is possiable to get same dev_t when adding partitions.
381	 */
382	if ((1U << MINORBITS) % max_part != 0)
383		max_part = 1UL << fls(max_part);
384
385	if (max_part > DISK_MAX_PARTS) {
386		pr_info("brd: max_part can't be larger than %d, reset max_part = %d.\n",
387			DISK_MAX_PARTS, DISK_MAX_PARTS);
388		max_part = DISK_MAX_PARTS;
389	}
390}
391
392static int __init brd_init(void)
393{
394	int err, i;
395
396	/*
397	 * brd module now has a feature to instantiate underlying device
398	 * structure on-demand, provided that there is an access dev node.
399	 *
400	 * (1) if rd_nr is specified, create that many upfront. else
401	 *     it defaults to CONFIG_BLK_DEV_RAM_COUNT
402	 * (2) User can further extend brd devices by create dev node themselves
403	 *     and have kernel automatically instantiate actual device
404	 *     on-demand. Example:
405	 *		mknod /path/devnod_name b 1 X	# 1 is the rd major
406	 *		fdisk -l /path/devnod_name
407	 *	If (X / max_part) was not already created it will be created
408	 *	dynamically.
409	 */
410
411	brd_check_and_reset_par();
412
413	brd_debugfs_dir = debugfs_create_dir("ramdisk_pages", NULL);
414
415	if (__register_blkdev(RAMDISK_MAJOR, "ramdisk", brd_probe)) {
416		err = -EIO;
417		goto out_free;
418	}
419
420	for (i = 0; i < rd_nr; i++)
421		brd_alloc(i);
422
423	pr_info("brd: module loaded\n");
424	return 0;
425
426out_free:
427	brd_cleanup();
428
429	pr_info("brd: module NOT loaded !!!\n");
430	return err;
431}
432
433static void __exit brd_exit(void)
434{
435
436	unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
437	brd_cleanup();
438
439	pr_info("brd: module unloaded\n");
440}
441
442module_init(brd_init);
443module_exit(brd_exit);
444