4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * linux/drivers/block/loop.c
36 * Written by Theodore Ts'o, 3/29/93
38 * Copyright 1993 by Theodore Ts'o. Redistribution of this file is
39 * permitted under the GNU General Public License.
41 * Modularized and updated for 1.1.16 kernel - Mitch Dsouza 28th May 1994
42 * Adapted for 1.3.59 kernel - Andries Brouwer, 1 Feb 1996
44 * Fixed do_loop_request() re-entrancy - Vincent.Renardias@waw.com Mar 20, 1997
46 * Added devfs support - Richard Gooch <rgooch@atnf.csiro.au> 16-Jan-1998
48 * Handle sparse backing files correctly - Kenn Humborg, Jun 28, 1998
50 * Loadable modules and other fixes by AK, 1998
52 * Maximum number of loop devices now dynamic via max_loop module parameter.
53 * Russell Kroll <rkroll@exploits.org> 19990701
55 * Maximum number of loop devices when compiled-in now selectable by passing
56 * max_loop=<1-255> to the kernel on boot.
57 * Erik I. Bols?, <eriki@himolde.no>, Oct 31, 1999
59 * Completely rewrite request handling to be make_request_fn style and
60 * non blocking, pushing work to a helper thread. Lots of fixes from
62 * Jens Axboe <axboe@suse.de>, Nov 2000
64 * Support up to 256 loop devices
65 * Heinz Mauelshagen <mge@sistina.com>, Feb 2002
67 * Support for falling back on the write file operation when the address space
68 * operations prepare_write and/or commit_write are not available on the
70 * Anton Altaparmakov, 16 Feb 2005
73 * - Advisory locking is ignored here.
74 * - Should use an own CAP_* category instead of CAP_SYS_ADMIN
78 #include <linux/module.h>
80 #include <linux/sched.h>
81 #include <linux/kthread.h>
83 #include <linux/file.h>
84 #include <linux/stat.h>
85 #include <linux/errno.h>
86 #include <linux/major.h>
87 #include <linux/wait.h>
88 #include <linux/blkdev.h>
89 #include <linux/blkpg.h>
90 #include <linux/init.h>
91 #include <linux/swap.h>
92 #include <linux/slab.h>
93 #include <linux/suspend.h>
94 #include <linux/writeback.h>
95 #include <linux/buffer_head.h> /* for invalidate_bdev() */
96 #include <linux/completion.h>
97 #include <linux/highmem.h>
98 #include <linux/gfp.h>
99 #include <linux/swap.h>
100 #include <linux/pagevec.h>
102 #include <asm/uaccess.h>
104 #include <lustre_lib.h>
105 #include "llite_internal.h"
107 #define LLOOP_MAX_SEGMENTS LNET_MAX_IOV
109 /* Possible states of device */
116 struct lloop_device {
122 struct file *lo_backing_file;
123 struct block_device *lo_device;
124 unsigned lo_blocksize;
130 struct bio *lo_biotail;
132 struct semaphore lo_sem;
133 struct mutex lo_ctl_mutex;
135 wait_queue_head_t lo_bh_wait;
137 struct request_queue *lo_queue;
139 const struct lu_env *lo_env;
141 struct ll_dio_pages lo_pvec;
143 /* data to handle bio for lustre. */
144 struct lo_request_data {
145 struct page *lrd_pages[LLOOP_MAX_SEGMENTS];
146 loff_t lrd_offsets[LLOOP_MAX_SEGMENTS];
154 LO_FLAGS_READ_ONLY = 1,
157 static int lloop_major;
158 #define MAX_LOOP_DEFAULT 16
159 static int max_loop = MAX_LOOP_DEFAULT;
160 static struct lloop_device *loop_dev;
161 static struct gendisk **disks;
162 static struct mutex lloop_mutex;
163 static void *ll_iocontrol_magic = NULL;
165 static loff_t get_loop_size(struct lloop_device *lo, struct file *file)
167 loff_t size, offset, loopsize;
169 /* Compute loopsize in bytes */
170 size = i_size_read(file->f_mapping->host);
171 offset = lo->lo_offset;
172 loopsize = size - offset;
173 if (lo->lo_sizelimit > 0 && lo->lo_sizelimit < loopsize)
174 loopsize = lo->lo_sizelimit;
177 * Unfortunately, if we want to do I/O on the device,
178 * the number of 512-byte sectors has to fit into a sector_t.
180 return loopsize >> 9;
183 static int do_bio_lustrebacked(struct lloop_device *lo, struct bio *head)
185 const struct lu_env *env = lo->lo_env;
186 struct cl_io *io = &lo->lo_io;
187 struct dentry *de = file_dentry(lo->lo_backing_file);
188 struct inode *inode = de->d_inode;
189 struct cl_object *obj = ll_i2info(inode)->lli_clob;
193 struct bio_vec *bvec;
195 size_t page_count = 0;
199 struct ll_dio_pages *pvec = &lo->lo_pvec;
200 struct page **pages = pvec->ldp_pages;
201 loff_t *offsets = pvec->ldp_offsets;
203 truncate_inode_pages(inode->i_mapping, 0);
205 /* initialize the IO */
206 memset(io, 0, sizeof(*io));
208 ret = cl_io_init(env, io, CIT_MISC, obj);
210 return io->ci_result;
211 io->ci_lockreq = CILR_NEVER;
213 LASSERT(head != NULL);
215 for (bio = head; bio != NULL; bio = bio->bi_next) {
216 LASSERT(rw == bio->bi_rw);
218 #ifdef HAVE_BVEC_ITER
219 offset = (pgoff_t)(bio->bi_iter.bi_sector << 9) + lo->lo_offset;
221 offset = (pgoff_t)(bio->bi_sector << 9) + lo->lo_offset;
223 bio_for_each_segment_all(bvec, bio, iter) {
224 BUG_ON(bvec->bv_offset != 0);
225 BUG_ON(bvec->bv_len != PAGE_SIZE);
227 pages[page_count] = bvec->bv_page;
228 offsets[page_count] = offset;
230 offset += bvec->bv_len;
232 LASSERT(page_count <= LLOOP_MAX_SEGMENTS);
235 ll_stats_ops_tally(ll_i2sbi(inode),
236 (rw == WRITE) ? LPROC_LL_BRW_WRITE : LPROC_LL_BRW_READ,
239 pvec->ldp_size = page_count << PAGE_SHIFT;
240 pvec->ldp_nr = page_count;
242 /* FIXME: in ll_direct_rw_pages, it has to allocate many cl_page{}s to
243 * write those pages into OST. Even worse case is that more pages
244 * would be asked to write out to swap space, and then finally get here
246 * Unfortunately this is NOT easy to fix.
247 * Thoughts on solution:
248 * 0. Define a reserved pool for cl_pages, which could be a list of
249 * pre-allocated cl_pages;
250 * 1. Define a new operation in cl_object_operations{}, says clo_depth,
251 * which measures how many layers for this lustre object. Generally
252 * speaking, the depth would be 2, one for llite, and one for lovsub.
253 * However, for SNS, there will be more since we need additional page
255 * 2. Reserve the # of (page_count * depth) cl_pages from the reserved
256 * pool. Afterwards, the clio would allocate the pages from reserved
257 * pool, this guarantees we neeedn't allocate the cl_pages from
258 * generic cl_page slab cache.
259 * Of course, if there is NOT enough pages in the pool, we might
260 * be asked to write less pages once, this purely depends on
261 * implementation. Anyway, we should be careful to avoid deadlocking.
264 bytes = ll_direct_rw_pages(env, io, rw, inode, pvec);
267 return (bytes == pvec->ldp_size) ? 0 : (int)bytes;
271 * Add bio to back of pending list
273 static void loop_add_bio(struct lloop_device *lo, struct bio *bio)
277 spin_lock_irqsave(&lo->lo_lock, flags);
278 if (lo->lo_biotail) {
279 lo->lo_biotail->bi_next = bio;
280 lo->lo_biotail = bio;
282 lo->lo_bio = lo->lo_biotail = bio;
283 spin_unlock_irqrestore(&lo->lo_lock, flags);
285 atomic_inc(&lo->lo_pending);
286 if (waitqueue_active(&lo->lo_bh_wait))
287 wake_up(&lo->lo_bh_wait);
291 * Grab first pending buffer
293 static unsigned int loop_get_bio(struct lloop_device *lo, struct bio **req)
297 unsigned int count = 0;
298 unsigned int page_count = 0;
301 spin_lock_irq(&lo->lo_lock);
303 if (unlikely(first == NULL)) {
304 spin_unlock_irq(&lo->lo_lock);
308 /* TODO: need to split the bio, too bad. */
309 LASSERT(first->bi_vcnt <= LLOOP_MAX_SEGMENTS);
313 while (*bio && (*bio)->bi_rw == rw) {
314 #ifdef HAVE_BVEC_ITER
315 CDEBUG(D_INFO, "bio sector %llu size %u count %u vcnt%u \n",
316 (unsigned long long)(*bio)->bi_iter.bi_sector,
317 (*bio)->bi_iter.bi_size, page_count, (*bio)->bi_vcnt);
319 CDEBUG(D_INFO, "bio sector %llu size %u count %u vcnt%u \n",
320 (unsigned long long)(*bio)->bi_sector, (*bio)->bi_size,
321 page_count, (*bio)->bi_vcnt);
323 if (page_count + (*bio)->bi_vcnt > LLOOP_MAX_SEGMENTS)
326 page_count += (*bio)->bi_vcnt;
328 bio = &(*bio)->bi_next;
331 /* Some of bios can't be mergable. */
335 /* Hit the end of queue */
336 lo->lo_biotail = NULL;
340 spin_unlock_irq(&lo->lo_lock);
345 loop_make_request(struct request_queue *q, struct bio *old_bio)
347 struct lloop_device *lo = q->queuedata;
348 int rw = bio_rw(old_bio);
354 #ifdef HAVE_BVEC_ITER
355 CDEBUG(D_INFO, "submit bio sector %llu size %u\n",
356 (unsigned long long)old_bio->bi_iter.bi_sector,
357 old_bio->bi_iter.bi_size);
359 CDEBUG(D_INFO, "submit bio sector %llu size %u\n",
360 (unsigned long long)old_bio->bi_sector, old_bio->bi_size);
363 spin_lock_irq(&lo->lo_lock);
364 inactive = (lo->lo_state != LLOOP_BOUND);
365 spin_unlock_irq(&lo->lo_lock);
370 if (lo->lo_flags & LO_FLAGS_READ_ONLY)
372 } else if (rw == READA) {
374 } else if (rw != READ) {
375 CERROR("lloop: unknown command (%x)\n", rw);
378 loop_add_bio(lo, old_bio);
381 bio_io_error(old_bio);
385 #ifdef HAVE_REQUEST_QUEUE_UNPLUG_FN
387 * kick off io on the underlying address space
389 static void loop_unplug(struct request_queue *q)
391 struct lloop_device *lo = q->queuedata;
393 clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags);
394 blk_run_address_space(lo->lo_backing_file->f_mapping);
398 static inline void loop_handle_bio(struct lloop_device *lo, struct bio *bio)
402 ret = do_bio_lustrebacked(lo, bio);
404 struct bio *tmp = bio->bi_next;
407 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
417 static inline int loop_active(struct lloop_device *lo)
419 return atomic_read(&lo->lo_pending) ||
420 (lo->lo_state == LLOOP_RUNDOWN);
424 * worker thread that handles reads/writes to file backed loop devices,
425 * to avoid blocking in our make_request_fn.
427 static int loop_thread(void *data)
429 struct lloop_device *lo = data;
432 unsigned long times = 0;
433 unsigned long total_count = 0;
439 set_user_nice(current, -20);
441 lo->lo_state = LLOOP_BOUND;
443 env = cl_env_get(&refcheck);
445 GOTO(out, ret = PTR_ERR(env));
448 memset(&lo->lo_pvec, 0, sizeof(lo->lo_pvec));
449 lo->lo_pvec.ldp_pages = lo->lo_requests[0].lrd_pages;
450 lo->lo_pvec.ldp_offsets = lo->lo_requests[0].lrd_offsets;
453 * up sem, we are running
458 wait_event(lo->lo_bh_wait, loop_active(lo));
459 if (!atomic_read(&lo->lo_pending)) {
461 spin_lock_irq(&lo->lo_lock);
462 exiting = (lo->lo_state == LLOOP_RUNDOWN);
463 spin_unlock_irq(&lo->lo_lock);
469 count = loop_get_bio(lo, &bio);
471 CWARN("lloop(minor: %d): missing bio\n", lo->lo_number);
475 total_count += count;
476 if (total_count < count) { /* overflow */
482 if ((times & 127) == 0) {
483 CDEBUG(D_INFO, "total: %lu, count: %lu, avg: %lu\n",
484 total_count, times, total_count / times);
487 LASSERT(bio != NULL);
488 LASSERT(count <= atomic_read(&lo->lo_pending));
489 loop_handle_bio(lo, bio);
490 atomic_sub(count, &lo->lo_pending);
492 cl_env_put(env, &refcheck);
499 static int loop_set_fd(struct lloop_device *lo, struct file *unused,
500 struct block_device *bdev, struct file *file)
503 struct address_space *mapping;
508 if (!try_module_get(THIS_MODULE))
512 if (lo->lo_state != LLOOP_UNBOUND)
515 mapping = file->f_mapping;
516 inode = mapping->host;
519 if (!S_ISREG(inode->i_mode) || inode->i_sb->s_magic != LL_SUPER_MAGIC)
522 if (!(file->f_mode & FMODE_WRITE))
523 lo_flags |= LO_FLAGS_READ_ONLY;
525 size = get_loop_size(lo, file);
527 if ((loff_t)(sector_t)size != size) {
532 /* remove all pages in cache so as dirty pages not to be existent. */
533 truncate_inode_pages(mapping, 0);
535 set_device_ro(bdev, (lo_flags & LO_FLAGS_READ_ONLY) != 0);
537 lo->lo_blocksize = PAGE_SIZE;
538 lo->lo_device = bdev;
539 lo->lo_flags = lo_flags;
540 lo->lo_backing_file = file;
541 lo->lo_sizelimit = 0;
542 lo->old_gfp_mask = mapping_gfp_mask(mapping);
543 mapping_set_gfp_mask(mapping, lo->old_gfp_mask & ~(__GFP_IO|__GFP_FS));
545 lo->lo_bio = lo->lo_biotail = NULL;
548 * set queue make_request_fn, and add limits based on lower level
551 blk_queue_make_request(lo->lo_queue, loop_make_request);
552 lo->lo_queue->queuedata = lo;
553 #ifdef HAVE_REQUEST_QUEUE_UNPLUG_FN
554 lo->lo_queue->unplug_fn = loop_unplug;
557 /* queue parameters */
558 blk_queue_max_hw_sectors(lo->lo_queue,
559 LLOOP_MAX_SEGMENTS << (PAGE_SHIFT - 9));
560 blk_queue_max_segments(lo->lo_queue, LLOOP_MAX_SEGMENTS);
562 set_capacity(disks[lo->lo_number], size);
563 bd_set_size(bdev, size << 9);
565 set_blocksize(bdev, lo->lo_blocksize);
567 kthread_run(loop_thread, lo, "lloop%d", lo->lo_number);
572 /* This is safe: open() is still holding a reference. */
573 module_put(THIS_MODULE);
577 static int loop_clr_fd(struct lloop_device *lo, struct block_device *bdev,
580 struct file *filp = lo->lo_backing_file;
581 gfp_t gfp = lo->old_gfp_mask;
583 if (lo->lo_state != LLOOP_BOUND)
586 if (lo->lo_refcnt > count) /* we needed one fd for the ioctl */
592 spin_lock_irq(&lo->lo_lock);
593 lo->lo_state = LLOOP_RUNDOWN;
594 spin_unlock_irq(&lo->lo_lock);
595 wake_up(&lo->lo_bh_wait);
598 lo->lo_backing_file = NULL;
599 lo->lo_device = NULL;
601 lo->lo_sizelimit = 0;
603 invalidate_bdev(bdev);
604 set_capacity(disks[lo->lo_number], 0);
605 bd_set_size(bdev, 0);
606 mapping_set_gfp_mask(filp->f_mapping, gfp);
607 lo->lo_state = LLOOP_UNBOUND;
609 /* This is safe: open() is still holding a reference. */
610 module_put(THIS_MODULE);
614 static int lo_open(struct block_device *bdev, fmode_t mode)
616 struct lloop_device *lo = bdev->bd_disk->private_data;
618 mutex_lock(&lo->lo_ctl_mutex);
620 mutex_unlock(&lo->lo_ctl_mutex);
625 #ifdef HAVE_BLKDEV_RELEASE_RETURN_INT
630 lo_release(struct gendisk *disk, fmode_t mode)
632 struct lloop_device *lo = disk->private_data;
634 mutex_lock(&lo->lo_ctl_mutex);
636 mutex_unlock(&lo->lo_ctl_mutex);
637 #ifdef HAVE_BLKDEV_RELEASE_RETURN_INT
642 /* lloop device node's ioctl function. */
643 static int lo_ioctl(struct block_device *bdev, fmode_t mode,
644 unsigned int cmd, unsigned long arg)
646 struct lloop_device *lo = bdev->bd_disk->private_data;
649 mutex_lock(&lloop_mutex);
651 case LL_IOC_LLOOP_DETACH: {
652 err = loop_clr_fd(lo, bdev, 2);
654 blkdev_put(bdev, 0); /* grabbed in LLOOP_ATTACH */
658 case LL_IOC_LLOOP_INFO: {
662 if (lo->lo_backing_file == NULL) {
666 inode = file_inode(lo->lo_backing_file);
667 if (inode != NULL && lo->lo_state == LLOOP_BOUND)
668 fid = ll_i2info(inode)->lli_fid;
672 if (copy_to_user((struct lu_fid __user *)arg,
682 mutex_unlock(&lloop_mutex);
687 static struct block_device_operations lo_fops = {
688 .owner = THIS_MODULE,
690 .release = lo_release,
694 /* dynamic iocontrol callback.
695 * This callback is registered in lloop_init and will be called by
698 * This is a llite regular file ioctl function. It takes the responsibility
699 * of attaching or detaching a file by a lloop's device numner.
701 static enum llioc_iter lloop_ioctl(struct inode *unused, struct file *file,
702 unsigned int cmd, unsigned long arg,
703 void *magic, int *rcp)
705 struct lloop_device *lo = NULL;
706 struct block_device *bdev = NULL;
710 if (magic != ll_iocontrol_magic)
714 GOTO(out1, err = -ENODEV);
716 CWARN("Enter llop_ioctl\n");
718 mutex_lock(&lloop_mutex);
720 case LL_IOC_LLOOP_ATTACH: {
721 struct inode *inode = file_inode(file);
722 struct lloop_device *lo_free = NULL;
725 for (i = 0; i < max_loop; i++, lo = NULL) {
727 if (lo->lo_state == LLOOP_UNBOUND) {
732 if (file_inode(lo->lo_backing_file) == inode)
736 GOTO(out, err = -EBUSY);
739 dev = MKDEV(lloop_major, lo->lo_number);
741 /* quit if the used pointer is writable */
742 if (put_user((long)old_encode_dev(dev), (long __user *)arg))
743 GOTO(out, err = -EFAULT);
745 bdev = blkdev_get_by_dev(dev, file->f_mode, NULL);
747 GOTO(out, err = PTR_ERR(bdev));
750 err = loop_set_fd(lo, NULL, bdev, file);
759 case LL_IOC_LLOOP_DETACH_BYDEV: {
762 dev = old_decode_dev(arg);
763 if (MAJOR(dev) != lloop_major)
764 GOTO(out, err = -EINVAL);
767 if (minor > max_loop - 1)
768 GOTO(out, err = -EINVAL);
770 lo = &loop_dev[minor];
771 if (lo->lo_state != LLOOP_BOUND)
772 GOTO(out, err = -EINVAL);
774 bdev = lo->lo_device;
775 err = loop_clr_fd(lo, bdev, 1);
777 blkdev_put(bdev, 0); /* grabbed in LLOOP_ATTACH */
788 mutex_unlock(&lloop_mutex);
795 static int __init lloop_init(void)
798 unsigned int cmdlist[] = {
800 LL_IOC_LLOOP_DETACH_BYDEV,
803 if (max_loop < 1 || max_loop > 256) {
804 max_loop = MAX_LOOP_DEFAULT;
805 CWARN("lloop: invalid max_loop (must be between"
806 " 1 and 256), using default (%u)\n", max_loop);
809 lloop_major = register_blkdev(0, "lloop");
813 CDEBUG(D_CONFIG, "registered lloop major %d with %u minors\n",
814 lloop_major, max_loop);
816 ll_iocontrol_magic = ll_iocontrol_register(lloop_ioctl, 2, cmdlist);
817 if (ll_iocontrol_magic == NULL)
820 OBD_ALLOC_WAIT(loop_dev, max_loop * sizeof(*loop_dev));
824 OBD_ALLOC_WAIT(disks, max_loop * sizeof(*disks));
828 for (i = 0; i < max_loop; i++) {
829 disks[i] = alloc_disk(1);
834 mutex_init(&lloop_mutex);
836 for (i = 0; i < max_loop; i++) {
837 struct lloop_device *lo = &loop_dev[i];
838 struct gendisk *disk = disks[i];
840 lo->lo_queue = blk_alloc_queue(GFP_KERNEL);
844 mutex_init(&lo->lo_ctl_mutex);
845 sema_init(&lo->lo_sem, 0);
846 init_waitqueue_head(&lo->lo_bh_wait);
848 spin_lock_init(&lo->lo_lock);
849 disk->major = lloop_major;
850 disk->first_minor = i;
851 disk->fops = &lo_fops;
852 sprintf(disk->disk_name, "lloop%d", i);
853 disk->private_data = lo;
854 disk->queue = lo->lo_queue;
857 /* We cannot fail after we call this, so another loop!*/
858 for (i = 0; i < max_loop; i++)
864 blk_cleanup_queue(loop_dev[i].lo_queue);
869 OBD_FREE(disks, max_loop * sizeof(*disks));
871 OBD_FREE(loop_dev, max_loop * sizeof(*loop_dev));
873 unregister_blkdev(lloop_major, "lloop");
874 ll_iocontrol_unregister(ll_iocontrol_magic);
875 CERROR("lloop: ran out of memory\n");
879 static void __exit lloop_exit(void)
883 ll_iocontrol_unregister(ll_iocontrol_magic);
884 for (i = 0; i < max_loop; i++) {
885 del_gendisk(disks[i]);
886 blk_cleanup_queue(loop_dev[i].lo_queue);
889 unregister_blkdev(lloop_major, "lloop");
891 OBD_FREE(disks, max_loop * sizeof(*disks));
892 OBD_FREE(loop_dev, max_loop * sizeof(*loop_dev));
895 module_param(max_loop, int, 0444);
896 MODULE_PARM_DESC(max_loop, "maximum of lloop_device");
897 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
898 MODULE_DESCRIPTION("Lustre virtual block device");
899 MODULE_VERSION(LUSTRE_VERSION_STRING);
900 MODULE_LICENSE("GPL");
902 module_init(lloop_init);
903 module_exit(lloop_exit);