74977203fc85310d0511887b755e4648defbcc4b
[muen/linux.git] / fs / btrfs / volumes.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/bio.h>
8 #include <linux/slab.h>
9 #include <linux/buffer_head.h>
10 #include <linux/blkdev.h>
11 #include <linux/ratelimit.h>
12 #include <linux/kthread.h>
13 #include <linux/raid/pq.h>
14 #include <linux/semaphore.h>
15 #include <linux/uuid.h>
16 #include <linux/list_sort.h>
17 #include "ctree.h"
18 #include "extent_map.h"
19 #include "disk-io.h"
20 #include "transaction.h"
21 #include "print-tree.h"
22 #include "volumes.h"
23 #include "raid56.h"
24 #include "async-thread.h"
25 #include "check-integrity.h"
26 #include "rcu-string.h"
27 #include "math.h"
28 #include "dev-replace.h"
29 #include "sysfs.h"
30
31 const struct btrfs_raid_attr btrfs_raid_array[BTRFS_NR_RAID_TYPES] = {
32         [BTRFS_RAID_RAID10] = {
33                 .sub_stripes    = 2,
34                 .dev_stripes    = 1,
35                 .devs_max       = 0,    /* 0 == as many as possible */
36                 .devs_min       = 4,
37                 .tolerated_failures = 1,
38                 .devs_increment = 2,
39                 .ncopies        = 2,
40                 .raid_name      = "raid10",
41                 .bg_flag        = BTRFS_BLOCK_GROUP_RAID10,
42                 .mindev_error   = BTRFS_ERROR_DEV_RAID10_MIN_NOT_MET,
43         },
44         [BTRFS_RAID_RAID1] = {
45                 .sub_stripes    = 1,
46                 .dev_stripes    = 1,
47                 .devs_max       = 2,
48                 .devs_min       = 2,
49                 .tolerated_failures = 1,
50                 .devs_increment = 2,
51                 .ncopies        = 2,
52                 .raid_name      = "raid1",
53                 .bg_flag        = BTRFS_BLOCK_GROUP_RAID1,
54                 .mindev_error   = BTRFS_ERROR_DEV_RAID1_MIN_NOT_MET,
55         },
56         [BTRFS_RAID_DUP] = {
57                 .sub_stripes    = 1,
58                 .dev_stripes    = 2,
59                 .devs_max       = 1,
60                 .devs_min       = 1,
61                 .tolerated_failures = 0,
62                 .devs_increment = 1,
63                 .ncopies        = 2,
64                 .raid_name      = "dup",
65                 .bg_flag        = BTRFS_BLOCK_GROUP_DUP,
66                 .mindev_error   = 0,
67         },
68         [BTRFS_RAID_RAID0] = {
69                 .sub_stripes    = 1,
70                 .dev_stripes    = 1,
71                 .devs_max       = 0,
72                 .devs_min       = 2,
73                 .tolerated_failures = 0,
74                 .devs_increment = 1,
75                 .ncopies        = 1,
76                 .raid_name      = "raid0",
77                 .bg_flag        = BTRFS_BLOCK_GROUP_RAID0,
78                 .mindev_error   = 0,
79         },
80         [BTRFS_RAID_SINGLE] = {
81                 .sub_stripes    = 1,
82                 .dev_stripes    = 1,
83                 .devs_max       = 1,
84                 .devs_min       = 1,
85                 .tolerated_failures = 0,
86                 .devs_increment = 1,
87                 .ncopies        = 1,
88                 .raid_name      = "single",
89                 .bg_flag        = 0,
90                 .mindev_error   = 0,
91         },
92         [BTRFS_RAID_RAID5] = {
93                 .sub_stripes    = 1,
94                 .dev_stripes    = 1,
95                 .devs_max       = 0,
96                 .devs_min       = 2,
97                 .tolerated_failures = 1,
98                 .devs_increment = 1,
99                 .ncopies        = 2,
100                 .raid_name      = "raid5",
101                 .bg_flag        = BTRFS_BLOCK_GROUP_RAID5,
102                 .mindev_error   = BTRFS_ERROR_DEV_RAID5_MIN_NOT_MET,
103         },
104         [BTRFS_RAID_RAID6] = {
105                 .sub_stripes    = 1,
106                 .dev_stripes    = 1,
107                 .devs_max       = 0,
108                 .devs_min       = 3,
109                 .tolerated_failures = 2,
110                 .devs_increment = 1,
111                 .ncopies        = 3,
112                 .raid_name      = "raid6",
113                 .bg_flag        = BTRFS_BLOCK_GROUP_RAID6,
114                 .mindev_error   = BTRFS_ERROR_DEV_RAID6_MIN_NOT_MET,
115         },
116 };
117
118 const char *get_raid_name(enum btrfs_raid_types type)
119 {
120         if (type >= BTRFS_NR_RAID_TYPES)
121                 return NULL;
122
123         return btrfs_raid_array[type].raid_name;
124 }
125
126 static int init_first_rw_device(struct btrfs_trans_handle *trans,
127                                 struct btrfs_fs_info *fs_info);
128 static int btrfs_relocate_sys_chunks(struct btrfs_fs_info *fs_info);
129 static void __btrfs_reset_dev_stats(struct btrfs_device *dev);
130 static void btrfs_dev_stat_print_on_error(struct btrfs_device *dev);
131 static void btrfs_dev_stat_print_on_load(struct btrfs_device *device);
132 static int __btrfs_map_block(struct btrfs_fs_info *fs_info,
133                              enum btrfs_map_op op,
134                              u64 logical, u64 *length,
135                              struct btrfs_bio **bbio_ret,
136                              int mirror_num, int need_raid_map);
137
138 /*
139  * Device locking
140  * ==============
141  *
142  * There are several mutexes that protect manipulation of devices and low-level
143  * structures like chunks but not block groups, extents or files
144  *
145  * uuid_mutex (global lock)
146  * ------------------------
147  * protects the fs_uuids list that tracks all per-fs fs_devices, resulting from
148  * the SCAN_DEV ioctl registration or from mount either implicitly (the first
149  * device) or requested by the device= mount option
150  *
151  * the mutex can be very coarse and can cover long-running operations
152  *
153  * protects: updates to fs_devices counters like missing devices, rw devices,
154  * seeding, structure cloning, openning/closing devices at mount/umount time
155  *
156  * global::fs_devs - add, remove, updates to the global list
157  *
158  * does not protect: manipulation of the fs_devices::devices list!
159  *
160  * btrfs_device::name - renames (write side), read is RCU
161  *
162  * fs_devices::device_list_mutex (per-fs, with RCU)
163  * ------------------------------------------------
164  * protects updates to fs_devices::devices, ie. adding and deleting
165  *
166  * simple list traversal with read-only actions can be done with RCU protection
167  *
168  * may be used to exclude some operations from running concurrently without any
169  * modifications to the list (see write_all_supers)
170  *
171  * balance_mutex
172  * -------------
173  * protects balance structures (status, state) and context accessed from
174  * several places (internally, ioctl)
175  *
176  * chunk_mutex
177  * -----------
178  * protects chunks, adding or removing during allocation, trim or when a new
179  * device is added/removed
180  *
181  * cleaner_mutex
182  * -------------
183  * a big lock that is held by the cleaner thread and prevents running subvolume
184  * cleaning together with relocation or delayed iputs
185  *
186  *
187  * Lock nesting
188  * ============
189  *
190  * uuid_mutex
191  *   volume_mutex
192  *     device_list_mutex
193  *       chunk_mutex
194  *     balance_mutex
195  *
196  *
197  * Exclusive operations, BTRFS_FS_EXCL_OP
198  * ======================================
199  *
200  * Maintains the exclusivity of the following operations that apply to the
201  * whole filesystem and cannot run in parallel.
202  *
203  * - Balance (*)
204  * - Device add
205  * - Device remove
206  * - Device replace (*)
207  * - Resize
208  *
209  * The device operations (as above) can be in one of the following states:
210  *
211  * - Running state
212  * - Paused state
213  * - Completed state
214  *
215  * Only device operations marked with (*) can go into the Paused state for the
216  * following reasons:
217  *
218  * - ioctl (only Balance can be Paused through ioctl)
219  * - filesystem remounted as read-only
220  * - filesystem unmounted and mounted as read-only
221  * - system power-cycle and filesystem mounted as read-only
222  * - filesystem or device errors leading to forced read-only
223  *
224  * BTRFS_FS_EXCL_OP flag is set and cleared using atomic operations.
225  * During the course of Paused state, the BTRFS_FS_EXCL_OP remains set.
226  * A device operation in Paused or Running state can be canceled or resumed
227  * either by ioctl (Balance only) or when remounted as read-write.
228  * BTRFS_FS_EXCL_OP flag is cleared when the device operation is canceled or
229  * completed.
230  */
231
232 DEFINE_MUTEX(uuid_mutex);
233 static LIST_HEAD(fs_uuids);
234 struct list_head *btrfs_get_fs_uuids(void)
235 {
236         return &fs_uuids;
237 }
238
239 /*
240  * alloc_fs_devices - allocate struct btrfs_fs_devices
241  * @fsid:       if not NULL, copy the uuid to fs_devices::fsid
242  *
243  * Return a pointer to a new struct btrfs_fs_devices on success, or ERR_PTR().
244  * The returned struct is not linked onto any lists and can be destroyed with
245  * kfree() right away.
246  */
247 static struct btrfs_fs_devices *alloc_fs_devices(const u8 *fsid)
248 {
249         struct btrfs_fs_devices *fs_devs;
250
251         fs_devs = kzalloc(sizeof(*fs_devs), GFP_KERNEL);
252         if (!fs_devs)
253                 return ERR_PTR(-ENOMEM);
254
255         mutex_init(&fs_devs->device_list_mutex);
256
257         INIT_LIST_HEAD(&fs_devs->devices);
258         INIT_LIST_HEAD(&fs_devs->resized_devices);
259         INIT_LIST_HEAD(&fs_devs->alloc_list);
260         INIT_LIST_HEAD(&fs_devs->fs_list);
261         if (fsid)
262                 memcpy(fs_devs->fsid, fsid, BTRFS_FSID_SIZE);
263
264         return fs_devs;
265 }
266
267 void btrfs_free_device(struct btrfs_device *device)
268 {
269         rcu_string_free(device->name);
270         bio_put(device->flush_bio);
271         kfree(device);
272 }
273
274 static void free_fs_devices(struct btrfs_fs_devices *fs_devices)
275 {
276         struct btrfs_device *device;
277         WARN_ON(fs_devices->opened);
278         while (!list_empty(&fs_devices->devices)) {
279                 device = list_entry(fs_devices->devices.next,
280                                     struct btrfs_device, dev_list);
281                 list_del(&device->dev_list);
282                 btrfs_free_device(device);
283         }
284         kfree(fs_devices);
285 }
286
287 static void btrfs_kobject_uevent(struct block_device *bdev,
288                                  enum kobject_action action)
289 {
290         int ret;
291
292         ret = kobject_uevent(&disk_to_dev(bdev->bd_disk)->kobj, action);
293         if (ret)
294                 pr_warn("BTRFS: Sending event '%d' to kobject: '%s' (%p): failed\n",
295                         action,
296                         kobject_name(&disk_to_dev(bdev->bd_disk)->kobj),
297                         &disk_to_dev(bdev->bd_disk)->kobj);
298 }
299
300 void __exit btrfs_cleanup_fs_uuids(void)
301 {
302         struct btrfs_fs_devices *fs_devices;
303
304         while (!list_empty(&fs_uuids)) {
305                 fs_devices = list_entry(fs_uuids.next,
306                                         struct btrfs_fs_devices, fs_list);
307                 list_del(&fs_devices->fs_list);
308                 free_fs_devices(fs_devices);
309         }
310 }
311
312 /*
313  * Returns a pointer to a new btrfs_device on success; ERR_PTR() on error.
314  * Returned struct is not linked onto any lists and must be destroyed using
315  * btrfs_free_device.
316  */
317 static struct btrfs_device *__alloc_device(void)
318 {
319         struct btrfs_device *dev;
320
321         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
322         if (!dev)
323                 return ERR_PTR(-ENOMEM);
324
325         /*
326          * Preallocate a bio that's always going to be used for flushing device
327          * barriers and matches the device lifespan
328          */
329         dev->flush_bio = bio_alloc_bioset(GFP_KERNEL, 0, NULL);
330         if (!dev->flush_bio) {
331                 kfree(dev);
332                 return ERR_PTR(-ENOMEM);
333         }
334
335         INIT_LIST_HEAD(&dev->dev_list);
336         INIT_LIST_HEAD(&dev->dev_alloc_list);
337         INIT_LIST_HEAD(&dev->resized_list);
338
339         spin_lock_init(&dev->io_lock);
340
341         atomic_set(&dev->reada_in_flight, 0);
342         atomic_set(&dev->dev_stats_ccnt, 0);
343         btrfs_device_data_ordered_init(dev);
344         INIT_RADIX_TREE(&dev->reada_zones, GFP_NOFS & ~__GFP_DIRECT_RECLAIM);
345         INIT_RADIX_TREE(&dev->reada_extents, GFP_NOFS & ~__GFP_DIRECT_RECLAIM);
346
347         return dev;
348 }
349
350 /*
351  * Find a device specified by @devid or @uuid in the list of @fs_devices, or
352  * return NULL.
353  *
354  * If devid and uuid are both specified, the match must be exact, otherwise
355  * only devid is used.
356  */
357 static struct btrfs_device *find_device(struct btrfs_fs_devices *fs_devices,
358                 u64 devid, const u8 *uuid)
359 {
360         struct btrfs_device *dev;
361
362         list_for_each_entry(dev, &fs_devices->devices, dev_list) {
363                 if (dev->devid == devid &&
364                     (!uuid || !memcmp(dev->uuid, uuid, BTRFS_UUID_SIZE))) {
365                         return dev;
366                 }
367         }
368         return NULL;
369 }
370
371 static noinline struct btrfs_fs_devices *find_fsid(u8 *fsid)
372 {
373         struct btrfs_fs_devices *fs_devices;
374
375         list_for_each_entry(fs_devices, &fs_uuids, fs_list) {
376                 if (memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE) == 0)
377                         return fs_devices;
378         }
379         return NULL;
380 }
381
382 static int
383 btrfs_get_bdev_and_sb(const char *device_path, fmode_t flags, void *holder,
384                       int flush, struct block_device **bdev,
385                       struct buffer_head **bh)
386 {
387         int ret;
388
389         *bdev = blkdev_get_by_path(device_path, flags, holder);
390
391         if (IS_ERR(*bdev)) {
392                 ret = PTR_ERR(*bdev);
393                 goto error;
394         }
395
396         if (flush)
397                 filemap_write_and_wait((*bdev)->bd_inode->i_mapping);
398         ret = set_blocksize(*bdev, BTRFS_BDEV_BLOCKSIZE);
399         if (ret) {
400                 blkdev_put(*bdev, flags);
401                 goto error;
402         }
403         invalidate_bdev(*bdev);
404         *bh = btrfs_read_dev_super(*bdev);
405         if (IS_ERR(*bh)) {
406                 ret = PTR_ERR(*bh);
407                 blkdev_put(*bdev, flags);
408                 goto error;
409         }
410
411         return 0;
412
413 error:
414         *bdev = NULL;
415         *bh = NULL;
416         return ret;
417 }
418
419 static void requeue_list(struct btrfs_pending_bios *pending_bios,
420                         struct bio *head, struct bio *tail)
421 {
422
423         struct bio *old_head;
424
425         old_head = pending_bios->head;
426         pending_bios->head = head;
427         if (pending_bios->tail)
428                 tail->bi_next = old_head;
429         else
430                 pending_bios->tail = tail;
431 }
432
433 /*
434  * we try to collect pending bios for a device so we don't get a large
435  * number of procs sending bios down to the same device.  This greatly
436  * improves the schedulers ability to collect and merge the bios.
437  *
438  * But, it also turns into a long list of bios to process and that is sure
439  * to eventually make the worker thread block.  The solution here is to
440  * make some progress and then put this work struct back at the end of
441  * the list if the block device is congested.  This way, multiple devices
442  * can make progress from a single worker thread.
443  */
444 static noinline void run_scheduled_bios(struct btrfs_device *device)
445 {
446         struct btrfs_fs_info *fs_info = device->fs_info;
447         struct bio *pending;
448         struct backing_dev_info *bdi;
449         struct btrfs_pending_bios *pending_bios;
450         struct bio *tail;
451         struct bio *cur;
452         int again = 0;
453         unsigned long num_run;
454         unsigned long batch_run = 0;
455         unsigned long last_waited = 0;
456         int force_reg = 0;
457         int sync_pending = 0;
458         struct blk_plug plug;
459
460         /*
461          * this function runs all the bios we've collected for
462          * a particular device.  We don't want to wander off to
463          * another device without first sending all of these down.
464          * So, setup a plug here and finish it off before we return
465          */
466         blk_start_plug(&plug);
467
468         bdi = device->bdev->bd_bdi;
469
470 loop:
471         spin_lock(&device->io_lock);
472
473 loop_lock:
474         num_run = 0;
475
476         /* take all the bios off the list at once and process them
477          * later on (without the lock held).  But, remember the
478          * tail and other pointers so the bios can be properly reinserted
479          * into the list if we hit congestion
480          */
481         if (!force_reg && device->pending_sync_bios.head) {
482                 pending_bios = &device->pending_sync_bios;
483                 force_reg = 1;
484         } else {
485                 pending_bios = &device->pending_bios;
486                 force_reg = 0;
487         }
488
489         pending = pending_bios->head;
490         tail = pending_bios->tail;
491         WARN_ON(pending && !tail);
492
493         /*
494          * if pending was null this time around, no bios need processing
495          * at all and we can stop.  Otherwise it'll loop back up again
496          * and do an additional check so no bios are missed.
497          *
498          * device->running_pending is used to synchronize with the
499          * schedule_bio code.
500          */
501         if (device->pending_sync_bios.head == NULL &&
502             device->pending_bios.head == NULL) {
503                 again = 0;
504                 device->running_pending = 0;
505         } else {
506                 again = 1;
507                 device->running_pending = 1;
508         }
509
510         pending_bios->head = NULL;
511         pending_bios->tail = NULL;
512
513         spin_unlock(&device->io_lock);
514
515         while (pending) {
516
517                 rmb();
518                 /* we want to work on both lists, but do more bios on the
519                  * sync list than the regular list
520                  */
521                 if ((num_run > 32 &&
522                     pending_bios != &device->pending_sync_bios &&
523                     device->pending_sync_bios.head) ||
524                    (num_run > 64 && pending_bios == &device->pending_sync_bios &&
525                     device->pending_bios.head)) {
526                         spin_lock(&device->io_lock);
527                         requeue_list(pending_bios, pending, tail);
528                         goto loop_lock;
529                 }
530
531                 cur = pending;
532                 pending = pending->bi_next;
533                 cur->bi_next = NULL;
534
535                 BUG_ON(atomic_read(&cur->__bi_cnt) == 0);
536
537                 /*
538                  * if we're doing the sync list, record that our
539                  * plug has some sync requests on it
540                  *
541                  * If we're doing the regular list and there are
542                  * sync requests sitting around, unplug before
543                  * we add more
544                  */
545                 if (pending_bios == &device->pending_sync_bios) {
546                         sync_pending = 1;
547                 } else if (sync_pending) {
548                         blk_finish_plug(&plug);
549                         blk_start_plug(&plug);
550                         sync_pending = 0;
551                 }
552
553                 btrfsic_submit_bio(cur);
554                 num_run++;
555                 batch_run++;
556
557                 cond_resched();
558
559                 /*
560                  * we made progress, there is more work to do and the bdi
561                  * is now congested.  Back off and let other work structs
562                  * run instead
563                  */
564                 if (pending && bdi_write_congested(bdi) && batch_run > 8 &&
565                     fs_info->fs_devices->open_devices > 1) {
566                         struct io_context *ioc;
567
568                         ioc = current->io_context;
569
570                         /*
571                          * the main goal here is that we don't want to
572                          * block if we're going to be able to submit
573                          * more requests without blocking.
574                          *
575                          * This code does two great things, it pokes into
576                          * the elevator code from a filesystem _and_
577                          * it makes assumptions about how batching works.
578                          */
579                         if (ioc && ioc->nr_batch_requests > 0 &&
580                             time_before(jiffies, ioc->last_waited + HZ/50UL) &&
581                             (last_waited == 0 ||
582                              ioc->last_waited == last_waited)) {
583                                 /*
584                                  * we want to go through our batch of
585                                  * requests and stop.  So, we copy out
586                                  * the ioc->last_waited time and test
587                                  * against it before looping
588                                  */
589                                 last_waited = ioc->last_waited;
590                                 cond_resched();
591                                 continue;
592                         }
593                         spin_lock(&device->io_lock);
594                         requeue_list(pending_bios, pending, tail);
595                         device->running_pending = 1;
596
597                         spin_unlock(&device->io_lock);
598                         btrfs_queue_work(fs_info->submit_workers,
599                                          &device->work);
600                         goto done;
601                 }
602         }
603
604         cond_resched();
605         if (again)
606                 goto loop;
607
608         spin_lock(&device->io_lock);
609         if (device->pending_bios.head || device->pending_sync_bios.head)
610                 goto loop_lock;
611         spin_unlock(&device->io_lock);
612
613 done:
614         blk_finish_plug(&plug);
615 }
616
617 static void pending_bios_fn(struct btrfs_work *work)
618 {
619         struct btrfs_device *device;
620
621         device = container_of(work, struct btrfs_device, work);
622         run_scheduled_bios(device);
623 }
624
625 /*
626  *  Search and remove all stale (devices which are not mounted) devices.
627  *  When both inputs are NULL, it will search and release all stale devices.
628  *  path:       Optional. When provided will it release all unmounted devices
629  *              matching this path only.
630  *  skip_dev:   Optional. Will skip this device when searching for the stale
631  *              devices.
632  */
633 static void btrfs_free_stale_devices(const char *path,
634                                      struct btrfs_device *skip_device)
635 {
636         struct btrfs_fs_devices *fs_devices, *tmp_fs_devices;
637         struct btrfs_device *device, *tmp_device;
638
639         list_for_each_entry_safe(fs_devices, tmp_fs_devices, &fs_uuids, fs_list) {
640                 mutex_lock(&fs_devices->device_list_mutex);
641                 if (fs_devices->opened) {
642                         mutex_unlock(&fs_devices->device_list_mutex);
643                         continue;
644                 }
645
646                 list_for_each_entry_safe(device, tmp_device,
647                                          &fs_devices->devices, dev_list) {
648                         int not_found = 0;
649
650                         if (skip_device && skip_device == device)
651                                 continue;
652                         if (path && !device->name)
653                                 continue;
654
655                         rcu_read_lock();
656                         if (path)
657                                 not_found = strcmp(rcu_str_deref(device->name),
658                                                    path);
659                         rcu_read_unlock();
660                         if (not_found)
661                                 continue;
662
663                         /* delete the stale device */
664                         fs_devices->num_devices--;
665                         list_del(&device->dev_list);
666                         btrfs_free_device(device);
667
668                         if (fs_devices->num_devices == 0)
669                                 break;
670                 }
671                 mutex_unlock(&fs_devices->device_list_mutex);
672                 if (fs_devices->num_devices == 0) {
673                         btrfs_sysfs_remove_fsid(fs_devices);
674                         list_del(&fs_devices->fs_list);
675                         free_fs_devices(fs_devices);
676                 }
677         }
678 }
679
680 static int btrfs_open_one_device(struct btrfs_fs_devices *fs_devices,
681                         struct btrfs_device *device, fmode_t flags,
682                         void *holder)
683 {
684         struct request_queue *q;
685         struct block_device *bdev;
686         struct buffer_head *bh;
687         struct btrfs_super_block *disk_super;
688         u64 devid;
689         int ret;
690
691         if (device->bdev)
692                 return -EINVAL;
693         if (!device->name)
694                 return -EINVAL;
695
696         ret = btrfs_get_bdev_and_sb(device->name->str, flags, holder, 1,
697                                     &bdev, &bh);
698         if (ret)
699                 return ret;
700
701         disk_super = (struct btrfs_super_block *)bh->b_data;
702         devid = btrfs_stack_device_id(&disk_super->dev_item);
703         if (devid != device->devid)
704                 goto error_brelse;
705
706         if (memcmp(device->uuid, disk_super->dev_item.uuid, BTRFS_UUID_SIZE))
707                 goto error_brelse;
708
709         device->generation = btrfs_super_generation(disk_super);
710
711         if (btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_SEEDING) {
712                 clear_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
713                 fs_devices->seeding = 1;
714         } else {
715                 if (bdev_read_only(bdev))
716                         clear_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
717                 else
718                         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
719         }
720
721         q = bdev_get_queue(bdev);
722         if (!blk_queue_nonrot(q))
723                 fs_devices->rotating = 1;
724
725         device->bdev = bdev;
726         clear_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
727         device->mode = flags;
728
729         fs_devices->open_devices++;
730         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state) &&
731             device->devid != BTRFS_DEV_REPLACE_DEVID) {
732                 fs_devices->rw_devices++;
733                 list_add_tail(&device->dev_alloc_list, &fs_devices->alloc_list);
734         }
735         brelse(bh);
736
737         return 0;
738
739 error_brelse:
740         brelse(bh);
741         blkdev_put(bdev, flags);
742
743         return -EINVAL;
744 }
745
746 /*
747  * Add new device to list of registered devices
748  *
749  * Returns:
750  * device pointer which was just added or updated when successful
751  * error pointer when failed
752  */
753 static noinline struct btrfs_device *device_list_add(const char *path,
754                            struct btrfs_super_block *disk_super,
755                            bool *new_device_added)
756 {
757         struct btrfs_device *device;
758         struct btrfs_fs_devices *fs_devices;
759         struct rcu_string *name;
760         u64 found_transid = btrfs_super_generation(disk_super);
761         u64 devid = btrfs_stack_device_id(&disk_super->dev_item);
762
763         fs_devices = find_fsid(disk_super->fsid);
764         if (!fs_devices) {
765                 fs_devices = alloc_fs_devices(disk_super->fsid);
766                 if (IS_ERR(fs_devices))
767                         return ERR_CAST(fs_devices);
768
769                 mutex_lock(&fs_devices->device_list_mutex);
770                 list_add(&fs_devices->fs_list, &fs_uuids);
771
772                 device = NULL;
773         } else {
774                 mutex_lock(&fs_devices->device_list_mutex);
775                 device = find_device(fs_devices, devid,
776                                 disk_super->dev_item.uuid);
777         }
778
779         if (!device) {
780                 if (fs_devices->opened) {
781                         mutex_unlock(&fs_devices->device_list_mutex);
782                         return ERR_PTR(-EBUSY);
783                 }
784
785                 device = btrfs_alloc_device(NULL, &devid,
786                                             disk_super->dev_item.uuid);
787                 if (IS_ERR(device)) {
788                         mutex_unlock(&fs_devices->device_list_mutex);
789                         /* we can safely leave the fs_devices entry around */
790                         return device;
791                 }
792
793                 name = rcu_string_strdup(path, GFP_NOFS);
794                 if (!name) {
795                         btrfs_free_device(device);
796                         mutex_unlock(&fs_devices->device_list_mutex);
797                         return ERR_PTR(-ENOMEM);
798                 }
799                 rcu_assign_pointer(device->name, name);
800
801                 list_add_rcu(&device->dev_list, &fs_devices->devices);
802                 fs_devices->num_devices++;
803
804                 device->fs_devices = fs_devices;
805                 *new_device_added = true;
806
807                 if (disk_super->label[0])
808                         pr_info("BTRFS: device label %s devid %llu transid %llu %s\n",
809                                 disk_super->label, devid, found_transid, path);
810                 else
811                         pr_info("BTRFS: device fsid %pU devid %llu transid %llu %s\n",
812                                 disk_super->fsid, devid, found_transid, path);
813
814         } else if (!device->name || strcmp(device->name->str, path)) {
815                 /*
816                  * When FS is already mounted.
817                  * 1. If you are here and if the device->name is NULL that
818                  *    means this device was missing at time of FS mount.
819                  * 2. If you are here and if the device->name is different
820                  *    from 'path' that means either
821                  *      a. The same device disappeared and reappeared with
822                  *         different name. or
823                  *      b. The missing-disk-which-was-replaced, has
824                  *         reappeared now.
825                  *
826                  * We must allow 1 and 2a above. But 2b would be a spurious
827                  * and unintentional.
828                  *
829                  * Further in case of 1 and 2a above, the disk at 'path'
830                  * would have missed some transaction when it was away and
831                  * in case of 2a the stale bdev has to be updated as well.
832                  * 2b must not be allowed at all time.
833                  */
834
835                 /*
836                  * For now, we do allow update to btrfs_fs_device through the
837                  * btrfs dev scan cli after FS has been mounted.  We're still
838                  * tracking a problem where systems fail mount by subvolume id
839                  * when we reject replacement on a mounted FS.
840                  */
841                 if (!fs_devices->opened && found_transid < device->generation) {
842                         /*
843                          * That is if the FS is _not_ mounted and if you
844                          * are here, that means there is more than one
845                          * disk with same uuid and devid.We keep the one
846                          * with larger generation number or the last-in if
847                          * generation are equal.
848                          */
849                         mutex_unlock(&fs_devices->device_list_mutex);
850                         return ERR_PTR(-EEXIST);
851                 }
852
853                 name = rcu_string_strdup(path, GFP_NOFS);
854                 if (!name) {
855                         mutex_unlock(&fs_devices->device_list_mutex);
856                         return ERR_PTR(-ENOMEM);
857                 }
858                 rcu_string_free(device->name);
859                 rcu_assign_pointer(device->name, name);
860                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state)) {
861                         fs_devices->missing_devices--;
862                         clear_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state);
863                 }
864         }
865
866         /*
867          * Unmount does not free the btrfs_device struct but would zero
868          * generation along with most of the other members. So just update
869          * it back. We need it to pick the disk with largest generation
870          * (as above).
871          */
872         if (!fs_devices->opened)
873                 device->generation = found_transid;
874
875         fs_devices->total_devices = btrfs_super_num_devices(disk_super);
876
877         mutex_unlock(&fs_devices->device_list_mutex);
878         return device;
879 }
880
881 static struct btrfs_fs_devices *clone_fs_devices(struct btrfs_fs_devices *orig)
882 {
883         struct btrfs_fs_devices *fs_devices;
884         struct btrfs_device *device;
885         struct btrfs_device *orig_dev;
886
887         fs_devices = alloc_fs_devices(orig->fsid);
888         if (IS_ERR(fs_devices))
889                 return fs_devices;
890
891         mutex_lock(&orig->device_list_mutex);
892         fs_devices->total_devices = orig->total_devices;
893
894         /* We have held the volume lock, it is safe to get the devices. */
895         list_for_each_entry(orig_dev, &orig->devices, dev_list) {
896                 struct rcu_string *name;
897
898                 device = btrfs_alloc_device(NULL, &orig_dev->devid,
899                                             orig_dev->uuid);
900                 if (IS_ERR(device))
901                         goto error;
902
903                 /*
904                  * This is ok to do without rcu read locked because we hold the
905                  * uuid mutex so nothing we touch in here is going to disappear.
906                  */
907                 if (orig_dev->name) {
908                         name = rcu_string_strdup(orig_dev->name->str,
909                                         GFP_KERNEL);
910                         if (!name) {
911                                 btrfs_free_device(device);
912                                 goto error;
913                         }
914                         rcu_assign_pointer(device->name, name);
915                 }
916
917                 list_add(&device->dev_list, &fs_devices->devices);
918                 device->fs_devices = fs_devices;
919                 fs_devices->num_devices++;
920         }
921         mutex_unlock(&orig->device_list_mutex);
922         return fs_devices;
923 error:
924         mutex_unlock(&orig->device_list_mutex);
925         free_fs_devices(fs_devices);
926         return ERR_PTR(-ENOMEM);
927 }
928
929 /*
930  * After we have read the system tree and know devids belonging to
931  * this filesystem, remove the device which does not belong there.
932  */
933 void btrfs_free_extra_devids(struct btrfs_fs_devices *fs_devices, int step)
934 {
935         struct btrfs_device *device, *next;
936         struct btrfs_device *latest_dev = NULL;
937
938         mutex_lock(&uuid_mutex);
939 again:
940         /* This is the initialized path, it is safe to release the devices. */
941         list_for_each_entry_safe(device, next, &fs_devices->devices, dev_list) {
942                 if (test_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
943                                                         &device->dev_state)) {
944                         if (!test_bit(BTRFS_DEV_STATE_REPLACE_TGT,
945                              &device->dev_state) &&
946                              (!latest_dev ||
947                               device->generation > latest_dev->generation)) {
948                                 latest_dev = device;
949                         }
950                         continue;
951                 }
952
953                 if (device->devid == BTRFS_DEV_REPLACE_DEVID) {
954                         /*
955                          * In the first step, keep the device which has
956                          * the correct fsid and the devid that is used
957                          * for the dev_replace procedure.
958                          * In the second step, the dev_replace state is
959                          * read from the device tree and it is known
960                          * whether the procedure is really active or
961                          * not, which means whether this device is
962                          * used or whether it should be removed.
963                          */
964                         if (step == 0 || test_bit(BTRFS_DEV_STATE_REPLACE_TGT,
965                                                   &device->dev_state)) {
966                                 continue;
967                         }
968                 }
969                 if (device->bdev) {
970                         blkdev_put(device->bdev, device->mode);
971                         device->bdev = NULL;
972                         fs_devices->open_devices--;
973                 }
974                 if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
975                         list_del_init(&device->dev_alloc_list);
976                         clear_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
977                         if (!test_bit(BTRFS_DEV_STATE_REPLACE_TGT,
978                                       &device->dev_state))
979                                 fs_devices->rw_devices--;
980                 }
981                 list_del_init(&device->dev_list);
982                 fs_devices->num_devices--;
983                 btrfs_free_device(device);
984         }
985
986         if (fs_devices->seed) {
987                 fs_devices = fs_devices->seed;
988                 goto again;
989         }
990
991         fs_devices->latest_bdev = latest_dev->bdev;
992
993         mutex_unlock(&uuid_mutex);
994 }
995
996 static void free_device_rcu(struct rcu_head *head)
997 {
998         struct btrfs_device *device;
999
1000         device = container_of(head, struct btrfs_device, rcu);
1001         btrfs_free_device(device);
1002 }
1003
1004 static void btrfs_close_bdev(struct btrfs_device *device)
1005 {
1006         if (!device->bdev)
1007                 return;
1008
1009         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
1010                 sync_blockdev(device->bdev);
1011                 invalidate_bdev(device->bdev);
1012         }
1013
1014         blkdev_put(device->bdev, device->mode);
1015 }
1016
1017 static void btrfs_close_one_device(struct btrfs_device *device)
1018 {
1019         struct btrfs_fs_devices *fs_devices = device->fs_devices;
1020         struct btrfs_device *new_device;
1021         struct rcu_string *name;
1022
1023         if (device->bdev)
1024                 fs_devices->open_devices--;
1025
1026         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state) &&
1027             device->devid != BTRFS_DEV_REPLACE_DEVID) {
1028                 list_del_init(&device->dev_alloc_list);
1029                 fs_devices->rw_devices--;
1030         }
1031
1032         if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
1033                 fs_devices->missing_devices--;
1034
1035         btrfs_close_bdev(device);
1036
1037         new_device = btrfs_alloc_device(NULL, &device->devid,
1038                                         device->uuid);
1039         BUG_ON(IS_ERR(new_device)); /* -ENOMEM */
1040
1041         /* Safe because we are under uuid_mutex */
1042         if (device->name) {
1043                 name = rcu_string_strdup(device->name->str, GFP_NOFS);
1044                 BUG_ON(!name); /* -ENOMEM */
1045                 rcu_assign_pointer(new_device->name, name);
1046         }
1047
1048         list_replace_rcu(&device->dev_list, &new_device->dev_list);
1049         new_device->fs_devices = device->fs_devices;
1050
1051         call_rcu(&device->rcu, free_device_rcu);
1052 }
1053
1054 static int close_fs_devices(struct btrfs_fs_devices *fs_devices)
1055 {
1056         struct btrfs_device *device, *tmp;
1057
1058         if (--fs_devices->opened > 0)
1059                 return 0;
1060
1061         mutex_lock(&fs_devices->device_list_mutex);
1062         list_for_each_entry_safe(device, tmp, &fs_devices->devices, dev_list) {
1063                 btrfs_close_one_device(device);
1064         }
1065         mutex_unlock(&fs_devices->device_list_mutex);
1066
1067         WARN_ON(fs_devices->open_devices);
1068         WARN_ON(fs_devices->rw_devices);
1069         fs_devices->opened = 0;
1070         fs_devices->seeding = 0;
1071
1072         return 0;
1073 }
1074
1075 int btrfs_close_devices(struct btrfs_fs_devices *fs_devices)
1076 {
1077         struct btrfs_fs_devices *seed_devices = NULL;
1078         int ret;
1079
1080         mutex_lock(&uuid_mutex);
1081         ret = close_fs_devices(fs_devices);
1082         if (!fs_devices->opened) {
1083                 seed_devices = fs_devices->seed;
1084                 fs_devices->seed = NULL;
1085         }
1086         mutex_unlock(&uuid_mutex);
1087
1088         while (seed_devices) {
1089                 fs_devices = seed_devices;
1090                 seed_devices = fs_devices->seed;
1091                 close_fs_devices(fs_devices);
1092                 free_fs_devices(fs_devices);
1093         }
1094         return ret;
1095 }
1096
1097 static int open_fs_devices(struct btrfs_fs_devices *fs_devices,
1098                                 fmode_t flags, void *holder)
1099 {
1100         struct btrfs_device *device;
1101         struct btrfs_device *latest_dev = NULL;
1102         int ret = 0;
1103
1104         flags |= FMODE_EXCL;
1105
1106         list_for_each_entry(device, &fs_devices->devices, dev_list) {
1107                 /* Just open everything we can; ignore failures here */
1108                 if (btrfs_open_one_device(fs_devices, device, flags, holder))
1109                         continue;
1110
1111                 if (!latest_dev ||
1112                     device->generation > latest_dev->generation)
1113                         latest_dev = device;
1114         }
1115         if (fs_devices->open_devices == 0) {
1116                 ret = -EINVAL;
1117                 goto out;
1118         }
1119         fs_devices->opened = 1;
1120         fs_devices->latest_bdev = latest_dev->bdev;
1121         fs_devices->total_rw_bytes = 0;
1122 out:
1123         return ret;
1124 }
1125
1126 static int devid_cmp(void *priv, struct list_head *a, struct list_head *b)
1127 {
1128         struct btrfs_device *dev1, *dev2;
1129
1130         dev1 = list_entry(a, struct btrfs_device, dev_list);
1131         dev2 = list_entry(b, struct btrfs_device, dev_list);
1132
1133         if (dev1->devid < dev2->devid)
1134                 return -1;
1135         else if (dev1->devid > dev2->devid)
1136                 return 1;
1137         return 0;
1138 }
1139
1140 int btrfs_open_devices(struct btrfs_fs_devices *fs_devices,
1141                        fmode_t flags, void *holder)
1142 {
1143         int ret;
1144
1145         lockdep_assert_held(&uuid_mutex);
1146
1147         mutex_lock(&fs_devices->device_list_mutex);
1148         if (fs_devices->opened) {
1149                 fs_devices->opened++;
1150                 ret = 0;
1151         } else {
1152                 list_sort(NULL, &fs_devices->devices, devid_cmp);
1153                 ret = open_fs_devices(fs_devices, flags, holder);
1154         }
1155         mutex_unlock(&fs_devices->device_list_mutex);
1156
1157         return ret;
1158 }
1159
1160 static void btrfs_release_disk_super(struct page *page)
1161 {
1162         kunmap(page);
1163         put_page(page);
1164 }
1165
1166 static int btrfs_read_disk_super(struct block_device *bdev, u64 bytenr,
1167                                  struct page **page,
1168                                  struct btrfs_super_block **disk_super)
1169 {
1170         void *p;
1171         pgoff_t index;
1172
1173         /* make sure our super fits in the device */
1174         if (bytenr + PAGE_SIZE >= i_size_read(bdev->bd_inode))
1175                 return 1;
1176
1177         /* make sure our super fits in the page */
1178         if (sizeof(**disk_super) > PAGE_SIZE)
1179                 return 1;
1180
1181         /* make sure our super doesn't straddle pages on disk */
1182         index = bytenr >> PAGE_SHIFT;
1183         if ((bytenr + sizeof(**disk_super) - 1) >> PAGE_SHIFT != index)
1184                 return 1;
1185
1186         /* pull in the page with our super */
1187         *page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
1188                                    index, GFP_KERNEL);
1189
1190         if (IS_ERR_OR_NULL(*page))
1191                 return 1;
1192
1193         p = kmap(*page);
1194
1195         /* align our pointer to the offset of the super block */
1196         *disk_super = p + (bytenr & ~PAGE_MASK);
1197
1198         if (btrfs_super_bytenr(*disk_super) != bytenr ||
1199             btrfs_super_magic(*disk_super) != BTRFS_MAGIC) {
1200                 btrfs_release_disk_super(*page);
1201                 return 1;
1202         }
1203
1204         if ((*disk_super)->label[0] &&
1205                 (*disk_super)->label[BTRFS_LABEL_SIZE - 1])
1206                 (*disk_super)->label[BTRFS_LABEL_SIZE - 1] = '\0';
1207
1208         return 0;
1209 }
1210
1211 /*
1212  * Look for a btrfs signature on a device. This may be called out of the mount path
1213  * and we are not allowed to call set_blocksize during the scan. The superblock
1214  * is read via pagecache
1215  */
1216 struct btrfs_device *btrfs_scan_one_device(const char *path, fmode_t flags,
1217                                            void *holder)
1218 {
1219         struct btrfs_super_block *disk_super;
1220         bool new_device_added = false;
1221         struct btrfs_device *device = NULL;
1222         struct block_device *bdev;
1223         struct page *page;
1224         u64 bytenr;
1225
1226         lockdep_assert_held(&uuid_mutex);
1227
1228         /*
1229          * we would like to check all the supers, but that would make
1230          * a btrfs mount succeed after a mkfs from a different FS.
1231          * So, we need to add a special mount option to scan for
1232          * later supers, using BTRFS_SUPER_MIRROR_MAX instead
1233          */
1234         bytenr = btrfs_sb_offset(0);
1235         flags |= FMODE_EXCL;
1236
1237         bdev = blkdev_get_by_path(path, flags, holder);
1238         if (IS_ERR(bdev))
1239                 return ERR_CAST(bdev);
1240
1241         if (btrfs_read_disk_super(bdev, bytenr, &page, &disk_super)) {
1242                 device = ERR_PTR(-EINVAL);
1243                 goto error_bdev_put;
1244         }
1245
1246         device = device_list_add(path, disk_super, &new_device_added);
1247         if (!IS_ERR(device)) {
1248                 if (new_device_added)
1249                         btrfs_free_stale_devices(path, device);
1250         }
1251
1252         btrfs_release_disk_super(page);
1253
1254 error_bdev_put:
1255         blkdev_put(bdev, flags);
1256
1257         return device;
1258 }
1259
1260 static int contains_pending_extent(struct btrfs_transaction *transaction,
1261                                    struct btrfs_device *device,
1262                                    u64 *start, u64 len)
1263 {
1264         struct btrfs_fs_info *fs_info = device->fs_info;
1265         struct extent_map *em;
1266         struct list_head *search_list = &fs_info->pinned_chunks;
1267         int ret = 0;
1268         u64 physical_start = *start;
1269
1270         if (transaction)
1271                 search_list = &transaction->pending_chunks;
1272 again:
1273         list_for_each_entry(em, search_list, list) {
1274                 struct map_lookup *map;
1275                 int i;
1276
1277                 map = em->map_lookup;
1278                 for (i = 0; i < map->num_stripes; i++) {
1279                         u64 end;
1280
1281                         if (map->stripes[i].dev != device)
1282                                 continue;
1283                         if (map->stripes[i].physical >= physical_start + len ||
1284                             map->stripes[i].physical + em->orig_block_len <=
1285                             physical_start)
1286                                 continue;
1287                         /*
1288                          * Make sure that while processing the pinned list we do
1289                          * not override our *start with a lower value, because
1290                          * we can have pinned chunks that fall within this
1291                          * device hole and that have lower physical addresses
1292                          * than the pending chunks we processed before. If we
1293                          * do not take this special care we can end up getting
1294                          * 2 pending chunks that start at the same physical
1295                          * device offsets because the end offset of a pinned
1296                          * chunk can be equal to the start offset of some
1297                          * pending chunk.
1298                          */
1299                         end = map->stripes[i].physical + em->orig_block_len;
1300                         if (end > *start) {
1301                                 *start = end;
1302                                 ret = 1;
1303                         }
1304                 }
1305         }
1306         if (search_list != &fs_info->pinned_chunks) {
1307                 search_list = &fs_info->pinned_chunks;
1308                 goto again;
1309         }
1310
1311         return ret;
1312 }
1313
1314
1315 /*
1316  * find_free_dev_extent_start - find free space in the specified device
1317  * @device:       the device which we search the free space in
1318  * @num_bytes:    the size of the free space that we need
1319  * @search_start: the position from which to begin the search
1320  * @start:        store the start of the free space.
1321  * @len:          the size of the free space. that we find, or the size
1322  *                of the max free space if we don't find suitable free space
1323  *
1324  * this uses a pretty simple search, the expectation is that it is
1325  * called very infrequently and that a given device has a small number
1326  * of extents
1327  *
1328  * @start is used to store the start of the free space if we find. But if we
1329  * don't find suitable free space, it will be used to store the start position
1330  * of the max free space.
1331  *
1332  * @len is used to store the size of the free space that we find.
1333  * But if we don't find suitable free space, it is used to store the size of
1334  * the max free space.
1335  */
1336 int find_free_dev_extent_start(struct btrfs_transaction *transaction,
1337                                struct btrfs_device *device, u64 num_bytes,
1338                                u64 search_start, u64 *start, u64 *len)
1339 {
1340         struct btrfs_fs_info *fs_info = device->fs_info;
1341         struct btrfs_root *root = fs_info->dev_root;
1342         struct btrfs_key key;
1343         struct btrfs_dev_extent *dev_extent;
1344         struct btrfs_path *path;
1345         u64 hole_size;
1346         u64 max_hole_start;
1347         u64 max_hole_size;
1348         u64 extent_end;
1349         u64 search_end = device->total_bytes;
1350         int ret;
1351         int slot;
1352         struct extent_buffer *l;
1353
1354         /*
1355          * We don't want to overwrite the superblock on the drive nor any area
1356          * used by the boot loader (grub for example), so we make sure to start
1357          * at an offset of at least 1MB.
1358          */
1359         search_start = max_t(u64, search_start, SZ_1M);
1360
1361         path = btrfs_alloc_path();
1362         if (!path)
1363                 return -ENOMEM;
1364
1365         max_hole_start = search_start;
1366         max_hole_size = 0;
1367
1368 again:
1369         if (search_start >= search_end ||
1370                 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
1371                 ret = -ENOSPC;
1372                 goto out;
1373         }
1374
1375         path->reada = READA_FORWARD;
1376         path->search_commit_root = 1;
1377         path->skip_locking = 1;
1378
1379         key.objectid = device->devid;
1380         key.offset = search_start;
1381         key.type = BTRFS_DEV_EXTENT_KEY;
1382
1383         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1384         if (ret < 0)
1385                 goto out;
1386         if (ret > 0) {
1387                 ret = btrfs_previous_item(root, path, key.objectid, key.type);
1388                 if (ret < 0)
1389                         goto out;
1390         }
1391
1392         while (1) {
1393                 l = path->nodes[0];
1394                 slot = path->slots[0];
1395                 if (slot >= btrfs_header_nritems(l)) {
1396                         ret = btrfs_next_leaf(root, path);
1397                         if (ret == 0)
1398                                 continue;
1399                         if (ret < 0)
1400                                 goto out;
1401
1402                         break;
1403                 }
1404                 btrfs_item_key_to_cpu(l, &key, slot);
1405
1406                 if (key.objectid < device->devid)
1407                         goto next;
1408
1409                 if (key.objectid > device->devid)
1410                         break;
1411
1412                 if (key.type != BTRFS_DEV_EXTENT_KEY)
1413                         goto next;
1414
1415                 if (key.offset > search_start) {
1416                         hole_size = key.offset - search_start;
1417
1418                         /*
1419                          * Have to check before we set max_hole_start, otherwise
1420                          * we could end up sending back this offset anyway.
1421                          */
1422                         if (contains_pending_extent(transaction, device,
1423                                                     &search_start,
1424                                                     hole_size)) {
1425                                 if (key.offset >= search_start) {
1426                                         hole_size = key.offset - search_start;
1427                                 } else {
1428                                         WARN_ON_ONCE(1);
1429                                         hole_size = 0;
1430                                 }
1431                         }
1432
1433                         if (hole_size > max_hole_size) {
1434                                 max_hole_start = search_start;
1435                                 max_hole_size = hole_size;
1436                         }
1437
1438                         /*
1439                          * If this free space is greater than which we need,
1440                          * it must be the max free space that we have found
1441                          * until now, so max_hole_start must point to the start
1442                          * of this free space and the length of this free space
1443                          * is stored in max_hole_size. Thus, we return
1444                          * max_hole_start and max_hole_size and go back to the
1445                          * caller.
1446                          */
1447                         if (hole_size >= num_bytes) {
1448                                 ret = 0;
1449                                 goto out;
1450                         }
1451                 }
1452
1453                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
1454                 extent_end = key.offset + btrfs_dev_extent_length(l,
1455                                                                   dev_extent);
1456                 if (extent_end > search_start)
1457                         search_start = extent_end;
1458 next:
1459                 path->slots[0]++;
1460                 cond_resched();
1461         }
1462
1463         /*
1464          * At this point, search_start should be the end of
1465          * allocated dev extents, and when shrinking the device,
1466          * search_end may be smaller than search_start.
1467          */
1468         if (search_end > search_start) {
1469                 hole_size = search_end - search_start;
1470
1471                 if (contains_pending_extent(transaction, device, &search_start,
1472                                             hole_size)) {
1473                         btrfs_release_path(path);
1474                         goto again;
1475                 }
1476
1477                 if (hole_size > max_hole_size) {
1478                         max_hole_start = search_start;
1479                         max_hole_size = hole_size;
1480                 }
1481         }
1482
1483         /* See above. */
1484         if (max_hole_size < num_bytes)
1485                 ret = -ENOSPC;
1486         else
1487                 ret = 0;
1488
1489 out:
1490         btrfs_free_path(path);
1491         *start = max_hole_start;
1492         if (len)
1493                 *len = max_hole_size;
1494         return ret;
1495 }
1496
1497 int find_free_dev_extent(struct btrfs_trans_handle *trans,
1498                          struct btrfs_device *device, u64 num_bytes,
1499                          u64 *start, u64 *len)
1500 {
1501         /* FIXME use last free of some kind */
1502         return find_free_dev_extent_start(trans->transaction, device,
1503                                           num_bytes, 0, start, len);
1504 }
1505
1506 static int btrfs_free_dev_extent(struct btrfs_trans_handle *trans,
1507                           struct btrfs_device *device,
1508                           u64 start, u64 *dev_extent_len)
1509 {
1510         struct btrfs_fs_info *fs_info = device->fs_info;
1511         struct btrfs_root *root = fs_info->dev_root;
1512         int ret;
1513         struct btrfs_path *path;
1514         struct btrfs_key key;
1515         struct btrfs_key found_key;
1516         struct extent_buffer *leaf = NULL;
1517         struct btrfs_dev_extent *extent = NULL;
1518
1519         path = btrfs_alloc_path();
1520         if (!path)
1521                 return -ENOMEM;
1522
1523         key.objectid = device->devid;
1524         key.offset = start;
1525         key.type = BTRFS_DEV_EXTENT_KEY;
1526 again:
1527         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1528         if (ret > 0) {
1529                 ret = btrfs_previous_item(root, path, key.objectid,
1530                                           BTRFS_DEV_EXTENT_KEY);
1531                 if (ret)
1532                         goto out;
1533                 leaf = path->nodes[0];
1534                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1535                 extent = btrfs_item_ptr(leaf, path->slots[0],
1536                                         struct btrfs_dev_extent);
1537                 BUG_ON(found_key.offset > start || found_key.offset +
1538                        btrfs_dev_extent_length(leaf, extent) < start);
1539                 key = found_key;
1540                 btrfs_release_path(path);
1541                 goto again;
1542         } else if (ret == 0) {
1543                 leaf = path->nodes[0];
1544                 extent = btrfs_item_ptr(leaf, path->slots[0],
1545                                         struct btrfs_dev_extent);
1546         } else {
1547                 btrfs_handle_fs_error(fs_info, ret, "Slot search failed");
1548                 goto out;
1549         }
1550
1551         *dev_extent_len = btrfs_dev_extent_length(leaf, extent);
1552
1553         ret = btrfs_del_item(trans, root, path);
1554         if (ret) {
1555                 btrfs_handle_fs_error(fs_info, ret,
1556                                       "Failed to remove dev extent item");
1557         } else {
1558                 set_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags);
1559         }
1560 out:
1561         btrfs_free_path(path);
1562         return ret;
1563 }
1564
1565 static int btrfs_alloc_dev_extent(struct btrfs_trans_handle *trans,
1566                                   struct btrfs_device *device,
1567                                   u64 chunk_offset, u64 start, u64 num_bytes)
1568 {
1569         int ret;
1570         struct btrfs_path *path;
1571         struct btrfs_fs_info *fs_info = device->fs_info;
1572         struct btrfs_root *root = fs_info->dev_root;
1573         struct btrfs_dev_extent *extent;
1574         struct extent_buffer *leaf;
1575         struct btrfs_key key;
1576
1577         WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
1578         WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
1579         path = btrfs_alloc_path();
1580         if (!path)
1581                 return -ENOMEM;
1582
1583         key.objectid = device->devid;
1584         key.offset = start;
1585         key.type = BTRFS_DEV_EXTENT_KEY;
1586         ret = btrfs_insert_empty_item(trans, root, path, &key,
1587                                       sizeof(*extent));
1588         if (ret)
1589                 goto out;
1590
1591         leaf = path->nodes[0];
1592         extent = btrfs_item_ptr(leaf, path->slots[0],
1593                                 struct btrfs_dev_extent);
1594         btrfs_set_dev_extent_chunk_tree(leaf, extent,
1595                                         BTRFS_CHUNK_TREE_OBJECTID);
1596         btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1597                                             BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1598         btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
1599
1600         btrfs_set_dev_extent_length(leaf, extent, num_bytes);
1601         btrfs_mark_buffer_dirty(leaf);
1602 out:
1603         btrfs_free_path(path);
1604         return ret;
1605 }
1606
1607 static u64 find_next_chunk(struct btrfs_fs_info *fs_info)
1608 {
1609         struct extent_map_tree *em_tree;
1610         struct extent_map *em;
1611         struct rb_node *n;
1612         u64 ret = 0;
1613
1614         em_tree = &fs_info->mapping_tree.map_tree;
1615         read_lock(&em_tree->lock);
1616         n = rb_last(&em_tree->map);
1617         if (n) {
1618                 em = rb_entry(n, struct extent_map, rb_node);
1619                 ret = em->start + em->len;
1620         }
1621         read_unlock(&em_tree->lock);
1622
1623         return ret;
1624 }
1625
1626 static noinline int find_next_devid(struct btrfs_fs_info *fs_info,
1627                                     u64 *devid_ret)
1628 {
1629         int ret;
1630         struct btrfs_key key;
1631         struct btrfs_key found_key;
1632         struct btrfs_path *path;
1633
1634         path = btrfs_alloc_path();
1635         if (!path)
1636                 return -ENOMEM;
1637
1638         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1639         key.type = BTRFS_DEV_ITEM_KEY;
1640         key.offset = (u64)-1;
1641
1642         ret = btrfs_search_slot(NULL, fs_info->chunk_root, &key, path, 0, 0);
1643         if (ret < 0)
1644                 goto error;
1645
1646         BUG_ON(ret == 0); /* Corruption */
1647
1648         ret = btrfs_previous_item(fs_info->chunk_root, path,
1649                                   BTRFS_DEV_ITEMS_OBJECTID,
1650                                   BTRFS_DEV_ITEM_KEY);
1651         if (ret) {
1652                 *devid_ret = 1;
1653         } else {
1654                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1655                                       path->slots[0]);
1656                 *devid_ret = found_key.offset + 1;
1657         }
1658         ret = 0;
1659 error:
1660         btrfs_free_path(path);
1661         return ret;
1662 }
1663
1664 /*
1665  * the device information is stored in the chunk root
1666  * the btrfs_device struct should be fully filled in
1667  */
1668 static int btrfs_add_dev_item(struct btrfs_trans_handle *trans,
1669                             struct btrfs_device *device)
1670 {
1671         int ret;
1672         struct btrfs_path *path;
1673         struct btrfs_dev_item *dev_item;
1674         struct extent_buffer *leaf;
1675         struct btrfs_key key;
1676         unsigned long ptr;
1677
1678         path = btrfs_alloc_path();
1679         if (!path)
1680                 return -ENOMEM;
1681
1682         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1683         key.type = BTRFS_DEV_ITEM_KEY;
1684         key.offset = device->devid;
1685
1686         ret = btrfs_insert_empty_item(trans, trans->fs_info->chunk_root, path,
1687                                       &key, sizeof(*dev_item));
1688         if (ret)
1689                 goto out;
1690
1691         leaf = path->nodes[0];
1692         dev_item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_item);
1693
1694         btrfs_set_device_id(leaf, dev_item, device->devid);
1695         btrfs_set_device_generation(leaf, dev_item, 0);
1696         btrfs_set_device_type(leaf, dev_item, device->type);
1697         btrfs_set_device_io_align(leaf, dev_item, device->io_align);
1698         btrfs_set_device_io_width(leaf, dev_item, device->io_width);
1699         btrfs_set_device_sector_size(leaf, dev_item, device->sector_size);
1700         btrfs_set_device_total_bytes(leaf, dev_item,
1701                                      btrfs_device_get_disk_total_bytes(device));
1702         btrfs_set_device_bytes_used(leaf, dev_item,
1703                                     btrfs_device_get_bytes_used(device));
1704         btrfs_set_device_group(leaf, dev_item, 0);
1705         btrfs_set_device_seek_speed(leaf, dev_item, 0);
1706         btrfs_set_device_bandwidth(leaf, dev_item, 0);
1707         btrfs_set_device_start_offset(leaf, dev_item, 0);
1708
1709         ptr = btrfs_device_uuid(dev_item);
1710         write_extent_buffer(leaf, device->uuid, ptr, BTRFS_UUID_SIZE);
1711         ptr = btrfs_device_fsid(dev_item);
1712         write_extent_buffer(leaf, trans->fs_info->fsid, ptr, BTRFS_FSID_SIZE);
1713         btrfs_mark_buffer_dirty(leaf);
1714
1715         ret = 0;
1716 out:
1717         btrfs_free_path(path);
1718         return ret;
1719 }
1720
1721 /*
1722  * Function to update ctime/mtime for a given device path.
1723  * Mainly used for ctime/mtime based probe like libblkid.
1724  */
1725 static void update_dev_time(const char *path_name)
1726 {
1727         struct file *filp;
1728
1729         filp = filp_open(path_name, O_RDWR, 0);
1730         if (IS_ERR(filp))
1731                 return;
1732         file_update_time(filp);
1733         filp_close(filp, NULL);
1734 }
1735
1736 static int btrfs_rm_dev_item(struct btrfs_fs_info *fs_info,
1737                              struct btrfs_device *device)
1738 {
1739         struct btrfs_root *root = fs_info->chunk_root;
1740         int ret;
1741         struct btrfs_path *path;
1742         struct btrfs_key key;
1743         struct btrfs_trans_handle *trans;
1744
1745         path = btrfs_alloc_path();
1746         if (!path)
1747                 return -ENOMEM;
1748
1749         trans = btrfs_start_transaction(root, 0);
1750         if (IS_ERR(trans)) {
1751                 btrfs_free_path(path);
1752                 return PTR_ERR(trans);
1753         }
1754         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1755         key.type = BTRFS_DEV_ITEM_KEY;
1756         key.offset = device->devid;
1757
1758         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1759         if (ret) {
1760                 if (ret > 0)
1761                         ret = -ENOENT;
1762                 btrfs_abort_transaction(trans, ret);
1763                 btrfs_end_transaction(trans);
1764                 goto out;
1765         }
1766
1767         ret = btrfs_del_item(trans, root, path);
1768         if (ret) {
1769                 btrfs_abort_transaction(trans, ret);
1770                 btrfs_end_transaction(trans);
1771         }
1772
1773 out:
1774         btrfs_free_path(path);
1775         if (!ret)
1776                 ret = btrfs_commit_transaction(trans);
1777         return ret;
1778 }
1779
1780 /*
1781  * Verify that @num_devices satisfies the RAID profile constraints in the whole
1782  * filesystem. It's up to the caller to adjust that number regarding eg. device
1783  * replace.
1784  */
1785 static int btrfs_check_raid_min_devices(struct btrfs_fs_info *fs_info,
1786                 u64 num_devices)
1787 {
1788         u64 all_avail;
1789         unsigned seq;
1790         int i;
1791
1792         do {
1793                 seq = read_seqbegin(&fs_info->profiles_lock);
1794
1795                 all_avail = fs_info->avail_data_alloc_bits |
1796                             fs_info->avail_system_alloc_bits |
1797                             fs_info->avail_metadata_alloc_bits;
1798         } while (read_seqretry(&fs_info->profiles_lock, seq));
1799
1800         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
1801                 if (!(all_avail & btrfs_raid_array[i].bg_flag))
1802                         continue;
1803
1804                 if (num_devices < btrfs_raid_array[i].devs_min) {
1805                         int ret = btrfs_raid_array[i].mindev_error;
1806
1807                         if (ret)
1808                                 return ret;
1809                 }
1810         }
1811
1812         return 0;
1813 }
1814
1815 static struct btrfs_device * btrfs_find_next_active_device(
1816                 struct btrfs_fs_devices *fs_devs, struct btrfs_device *device)
1817 {
1818         struct btrfs_device *next_device;
1819
1820         list_for_each_entry(next_device, &fs_devs->devices, dev_list) {
1821                 if (next_device != device &&
1822                     !test_bit(BTRFS_DEV_STATE_MISSING, &next_device->dev_state)
1823                     && next_device->bdev)
1824                         return next_device;
1825         }
1826
1827         return NULL;
1828 }
1829
1830 /*
1831  * Helper function to check if the given device is part of s_bdev / latest_bdev
1832  * and replace it with the provided or the next active device, in the context
1833  * where this function called, there should be always be another device (or
1834  * this_dev) which is active.
1835  */
1836 void btrfs_assign_next_active_device(struct btrfs_device *device,
1837                                      struct btrfs_device *this_dev)
1838 {
1839         struct btrfs_fs_info *fs_info = device->fs_info;
1840         struct btrfs_device *next_device;
1841
1842         if (this_dev)
1843                 next_device = this_dev;
1844         else
1845                 next_device = btrfs_find_next_active_device(fs_info->fs_devices,
1846                                                                 device);
1847         ASSERT(next_device);
1848
1849         if (fs_info->sb->s_bdev &&
1850                         (fs_info->sb->s_bdev == device->bdev))
1851                 fs_info->sb->s_bdev = next_device->bdev;
1852
1853         if (fs_info->fs_devices->latest_bdev == device->bdev)
1854                 fs_info->fs_devices->latest_bdev = next_device->bdev;
1855 }
1856
1857 int btrfs_rm_device(struct btrfs_fs_info *fs_info, const char *device_path,
1858                 u64 devid)
1859 {
1860         struct btrfs_device *device;
1861         struct btrfs_fs_devices *cur_devices;
1862         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
1863         u64 num_devices;
1864         int ret = 0;
1865
1866         mutex_lock(&uuid_mutex);
1867
1868         num_devices = fs_devices->num_devices;
1869         btrfs_dev_replace_read_lock(&fs_info->dev_replace);
1870         if (btrfs_dev_replace_is_ongoing(&fs_info->dev_replace)) {
1871                 WARN_ON(num_devices < 1);
1872                 num_devices--;
1873         }
1874         btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
1875
1876         ret = btrfs_check_raid_min_devices(fs_info, num_devices - 1);
1877         if (ret)
1878                 goto out;
1879
1880         ret = btrfs_find_device_by_devspec(fs_info, devid, device_path,
1881                                            &device);
1882         if (ret)
1883                 goto out;
1884
1885         if (test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
1886                 ret = BTRFS_ERROR_DEV_TGT_REPLACE;
1887                 goto out;
1888         }
1889
1890         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state) &&
1891             fs_info->fs_devices->rw_devices == 1) {
1892                 ret = BTRFS_ERROR_DEV_ONLY_WRITABLE;
1893                 goto out;
1894         }
1895
1896         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
1897                 mutex_lock(&fs_info->chunk_mutex);
1898                 list_del_init(&device->dev_alloc_list);
1899                 device->fs_devices->rw_devices--;
1900                 mutex_unlock(&fs_info->chunk_mutex);
1901         }
1902
1903         mutex_unlock(&uuid_mutex);
1904         ret = btrfs_shrink_device(device, 0);
1905         mutex_lock(&uuid_mutex);
1906         if (ret)
1907                 goto error_undo;
1908
1909         /*
1910          * TODO: the superblock still includes this device in its num_devices
1911          * counter although write_all_supers() is not locked out. This
1912          * could give a filesystem state which requires a degraded mount.
1913          */
1914         ret = btrfs_rm_dev_item(fs_info, device);
1915         if (ret)
1916                 goto error_undo;
1917
1918         clear_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
1919         btrfs_scrub_cancel_dev(fs_info, device);
1920
1921         /*
1922          * the device list mutex makes sure that we don't change
1923          * the device list while someone else is writing out all
1924          * the device supers. Whoever is writing all supers, should
1925          * lock the device list mutex before getting the number of
1926          * devices in the super block (super_copy). Conversely,
1927          * whoever updates the number of devices in the super block
1928          * (super_copy) should hold the device list mutex.
1929          */
1930
1931         /*
1932          * In normal cases the cur_devices == fs_devices. But in case
1933          * of deleting a seed device, the cur_devices should point to
1934          * its own fs_devices listed under the fs_devices->seed.
1935          */
1936         cur_devices = device->fs_devices;
1937         mutex_lock(&fs_devices->device_list_mutex);
1938         list_del_rcu(&device->dev_list);
1939
1940         cur_devices->num_devices--;
1941         cur_devices->total_devices--;
1942         /* Update total_devices of the parent fs_devices if it's seed */
1943         if (cur_devices != fs_devices)
1944                 fs_devices->total_devices--;
1945
1946         if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
1947                 cur_devices->missing_devices--;
1948
1949         btrfs_assign_next_active_device(device, NULL);
1950
1951         if (device->bdev) {
1952                 cur_devices->open_devices--;
1953                 /* remove sysfs entry */
1954                 btrfs_sysfs_rm_device_link(fs_devices, device);
1955         }
1956
1957         num_devices = btrfs_super_num_devices(fs_info->super_copy) - 1;
1958         btrfs_set_super_num_devices(fs_info->super_copy, num_devices);
1959         mutex_unlock(&fs_devices->device_list_mutex);
1960
1961         /*
1962          * at this point, the device is zero sized and detached from
1963          * the devices list.  All that's left is to zero out the old
1964          * supers and free the device.
1965          */
1966         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
1967                 btrfs_scratch_superblocks(device->bdev, device->name->str);
1968
1969         btrfs_close_bdev(device);
1970         call_rcu(&device->rcu, free_device_rcu);
1971
1972         if (cur_devices->open_devices == 0) {
1973                 while (fs_devices) {
1974                         if (fs_devices->seed == cur_devices) {
1975                                 fs_devices->seed = cur_devices->seed;
1976                                 break;
1977                         }
1978                         fs_devices = fs_devices->seed;
1979                 }
1980                 cur_devices->seed = NULL;
1981                 close_fs_devices(cur_devices);
1982                 free_fs_devices(cur_devices);
1983         }
1984
1985 out:
1986         mutex_unlock(&uuid_mutex);
1987         return ret;
1988
1989 error_undo:
1990         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
1991                 mutex_lock(&fs_info->chunk_mutex);
1992                 list_add(&device->dev_alloc_list,
1993                          &fs_devices->alloc_list);
1994                 device->fs_devices->rw_devices++;
1995                 mutex_unlock(&fs_info->chunk_mutex);
1996         }
1997         goto out;
1998 }
1999
2000 void btrfs_rm_dev_replace_remove_srcdev(struct btrfs_device *srcdev)
2001 {
2002         struct btrfs_fs_devices *fs_devices;
2003
2004         lockdep_assert_held(&srcdev->fs_info->fs_devices->device_list_mutex);
2005
2006         /*
2007          * in case of fs with no seed, srcdev->fs_devices will point
2008          * to fs_devices of fs_info. However when the dev being replaced is
2009          * a seed dev it will point to the seed's local fs_devices. In short
2010          * srcdev will have its correct fs_devices in both the cases.
2011          */
2012         fs_devices = srcdev->fs_devices;
2013
2014         list_del_rcu(&srcdev->dev_list);
2015         list_del(&srcdev->dev_alloc_list);
2016         fs_devices->num_devices--;
2017         if (test_bit(BTRFS_DEV_STATE_MISSING, &srcdev->dev_state))
2018                 fs_devices->missing_devices--;
2019
2020         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &srcdev->dev_state))
2021                 fs_devices->rw_devices--;
2022
2023         if (srcdev->bdev)
2024                 fs_devices->open_devices--;
2025 }
2026
2027 void btrfs_rm_dev_replace_free_srcdev(struct btrfs_fs_info *fs_info,
2028                                       struct btrfs_device *srcdev)
2029 {
2030         struct btrfs_fs_devices *fs_devices = srcdev->fs_devices;
2031
2032         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &srcdev->dev_state)) {
2033                 /* zero out the old super if it is writable */
2034                 btrfs_scratch_superblocks(srcdev->bdev, srcdev->name->str);
2035         }
2036
2037         btrfs_close_bdev(srcdev);
2038         call_rcu(&srcdev->rcu, free_device_rcu);
2039
2040         /* if this is no devs we rather delete the fs_devices */
2041         if (!fs_devices->num_devices) {
2042                 struct btrfs_fs_devices *tmp_fs_devices;
2043
2044                 /*
2045                  * On a mounted FS, num_devices can't be zero unless it's a
2046                  * seed. In case of a seed device being replaced, the replace
2047                  * target added to the sprout FS, so there will be no more
2048                  * device left under the seed FS.
2049                  */
2050                 ASSERT(fs_devices->seeding);
2051
2052                 tmp_fs_devices = fs_info->fs_devices;
2053                 while (tmp_fs_devices) {
2054                         if (tmp_fs_devices->seed == fs_devices) {
2055                                 tmp_fs_devices->seed = fs_devices->seed;
2056                                 break;
2057                         }
2058                         tmp_fs_devices = tmp_fs_devices->seed;
2059                 }
2060                 fs_devices->seed = NULL;
2061                 close_fs_devices(fs_devices);
2062                 free_fs_devices(fs_devices);
2063         }
2064 }
2065
2066 void btrfs_destroy_dev_replace_tgtdev(struct btrfs_device *tgtdev)
2067 {
2068         struct btrfs_fs_devices *fs_devices = tgtdev->fs_info->fs_devices;
2069
2070         WARN_ON(!tgtdev);
2071         mutex_lock(&fs_devices->device_list_mutex);
2072
2073         btrfs_sysfs_rm_device_link(fs_devices, tgtdev);
2074
2075         if (tgtdev->bdev)
2076                 fs_devices->open_devices--;
2077
2078         fs_devices->num_devices--;
2079
2080         btrfs_assign_next_active_device(tgtdev, NULL);
2081
2082         list_del_rcu(&tgtdev->dev_list);
2083
2084         mutex_unlock(&fs_devices->device_list_mutex);
2085
2086         /*
2087          * The update_dev_time() with in btrfs_scratch_superblocks()
2088          * may lead to a call to btrfs_show_devname() which will try
2089          * to hold device_list_mutex. And here this device
2090          * is already out of device list, so we don't have to hold
2091          * the device_list_mutex lock.
2092          */
2093         btrfs_scratch_superblocks(tgtdev->bdev, tgtdev->name->str);
2094
2095         btrfs_close_bdev(tgtdev);
2096         call_rcu(&tgtdev->rcu, free_device_rcu);
2097 }
2098
2099 static int btrfs_find_device_by_path(struct btrfs_fs_info *fs_info,
2100                                      const char *device_path,
2101                                      struct btrfs_device **device)
2102 {
2103         int ret = 0;
2104         struct btrfs_super_block *disk_super;
2105         u64 devid;
2106         u8 *dev_uuid;
2107         struct block_device *bdev;
2108         struct buffer_head *bh;
2109
2110         *device = NULL;
2111         ret = btrfs_get_bdev_and_sb(device_path, FMODE_READ,
2112                                     fs_info->bdev_holder, 0, &bdev, &bh);
2113         if (ret)
2114                 return ret;
2115         disk_super = (struct btrfs_super_block *)bh->b_data;
2116         devid = btrfs_stack_device_id(&disk_super->dev_item);
2117         dev_uuid = disk_super->dev_item.uuid;
2118         *device = btrfs_find_device(fs_info, devid, dev_uuid, disk_super->fsid);
2119         brelse(bh);
2120         if (!*device)
2121                 ret = -ENOENT;
2122         blkdev_put(bdev, FMODE_READ);
2123         return ret;
2124 }
2125
2126 int btrfs_find_device_missing_or_by_path(struct btrfs_fs_info *fs_info,
2127                                          const char *device_path,
2128                                          struct btrfs_device **device)
2129 {
2130         *device = NULL;
2131         if (strcmp(device_path, "missing") == 0) {
2132                 struct list_head *devices;
2133                 struct btrfs_device *tmp;
2134
2135                 devices = &fs_info->fs_devices->devices;
2136                 list_for_each_entry(tmp, devices, dev_list) {
2137                         if (test_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
2138                                         &tmp->dev_state) && !tmp->bdev) {
2139                                 *device = tmp;
2140                                 break;
2141                         }
2142                 }
2143
2144                 if (!*device)
2145                         return BTRFS_ERROR_DEV_MISSING_NOT_FOUND;
2146
2147                 return 0;
2148         } else {
2149                 return btrfs_find_device_by_path(fs_info, device_path, device);
2150         }
2151 }
2152
2153 /*
2154  * Lookup a device given by device id, or the path if the id is 0.
2155  */
2156 int btrfs_find_device_by_devspec(struct btrfs_fs_info *fs_info, u64 devid,
2157                                  const char *devpath,
2158                                  struct btrfs_device **device)
2159 {
2160         int ret;
2161
2162         if (devid) {
2163                 ret = 0;
2164                 *device = btrfs_find_device(fs_info, devid, NULL, NULL);
2165                 if (!*device)
2166                         ret = -ENOENT;
2167         } else {
2168                 if (!devpath || !devpath[0])
2169                         return -EINVAL;
2170
2171                 ret = btrfs_find_device_missing_or_by_path(fs_info, devpath,
2172                                                            device);
2173         }
2174         return ret;
2175 }
2176
2177 /*
2178  * does all the dirty work required for changing file system's UUID.
2179  */
2180 static int btrfs_prepare_sprout(struct btrfs_fs_info *fs_info)
2181 {
2182         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2183         struct btrfs_fs_devices *old_devices;
2184         struct btrfs_fs_devices *seed_devices;
2185         struct btrfs_super_block *disk_super = fs_info->super_copy;
2186         struct btrfs_device *device;
2187         u64 super_flags;
2188
2189         lockdep_assert_held(&uuid_mutex);
2190         if (!fs_devices->seeding)
2191                 return -EINVAL;
2192
2193         seed_devices = alloc_fs_devices(NULL);
2194         if (IS_ERR(seed_devices))
2195                 return PTR_ERR(seed_devices);
2196
2197         old_devices = clone_fs_devices(fs_devices);
2198         if (IS_ERR(old_devices)) {
2199                 kfree(seed_devices);
2200                 return PTR_ERR(old_devices);
2201         }
2202
2203         list_add(&old_devices->fs_list, &fs_uuids);
2204
2205         memcpy(seed_devices, fs_devices, sizeof(*seed_devices));
2206         seed_devices->opened = 1;
2207         INIT_LIST_HEAD(&seed_devices->devices);
2208         INIT_LIST_HEAD(&seed_devices->alloc_list);
2209         mutex_init(&seed_devices->device_list_mutex);
2210
2211         mutex_lock(&fs_devices->device_list_mutex);
2212         list_splice_init_rcu(&fs_devices->devices, &seed_devices->devices,
2213                               synchronize_rcu);
2214         list_for_each_entry(device, &seed_devices->devices, dev_list)
2215                 device->fs_devices = seed_devices;
2216
2217         mutex_lock(&fs_info->chunk_mutex);
2218         list_splice_init(&fs_devices->alloc_list, &seed_devices->alloc_list);
2219         mutex_unlock(&fs_info->chunk_mutex);
2220
2221         fs_devices->seeding = 0;
2222         fs_devices->num_devices = 0;
2223         fs_devices->open_devices = 0;
2224         fs_devices->missing_devices = 0;
2225         fs_devices->rotating = 0;
2226         fs_devices->seed = seed_devices;
2227
2228         generate_random_uuid(fs_devices->fsid);
2229         memcpy(fs_info->fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
2230         memcpy(disk_super->fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
2231         mutex_unlock(&fs_devices->device_list_mutex);
2232
2233         super_flags = btrfs_super_flags(disk_super) &
2234                       ~BTRFS_SUPER_FLAG_SEEDING;
2235         btrfs_set_super_flags(disk_super, super_flags);
2236
2237         return 0;
2238 }
2239
2240 /*
2241  * Store the expected generation for seed devices in device items.
2242  */
2243 static int btrfs_finish_sprout(struct btrfs_trans_handle *trans,
2244                                struct btrfs_fs_info *fs_info)
2245 {
2246         struct btrfs_root *root = fs_info->chunk_root;
2247         struct btrfs_path *path;
2248         struct extent_buffer *leaf;
2249         struct btrfs_dev_item *dev_item;
2250         struct btrfs_device *device;
2251         struct btrfs_key key;
2252         u8 fs_uuid[BTRFS_FSID_SIZE];
2253         u8 dev_uuid[BTRFS_UUID_SIZE];
2254         u64 devid;
2255         int ret;
2256
2257         path = btrfs_alloc_path();
2258         if (!path)
2259                 return -ENOMEM;
2260
2261         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
2262         key.offset = 0;
2263         key.type = BTRFS_DEV_ITEM_KEY;
2264
2265         while (1) {
2266                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2267                 if (ret < 0)
2268                         goto error;
2269
2270                 leaf = path->nodes[0];
2271 next_slot:
2272                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
2273                         ret = btrfs_next_leaf(root, path);
2274                         if (ret > 0)
2275                                 break;
2276                         if (ret < 0)
2277                                 goto error;
2278                         leaf = path->nodes[0];
2279                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2280                         btrfs_release_path(path);
2281                         continue;
2282                 }
2283
2284                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2285                 if (key.objectid != BTRFS_DEV_ITEMS_OBJECTID ||
2286                     key.type != BTRFS_DEV_ITEM_KEY)
2287                         break;
2288
2289                 dev_item = btrfs_item_ptr(leaf, path->slots[0],
2290                                           struct btrfs_dev_item);
2291                 devid = btrfs_device_id(leaf, dev_item);
2292                 read_extent_buffer(leaf, dev_uuid, btrfs_device_uuid(dev_item),
2293                                    BTRFS_UUID_SIZE);
2294                 read_extent_buffer(leaf, fs_uuid, btrfs_device_fsid(dev_item),
2295                                    BTRFS_FSID_SIZE);
2296                 device = btrfs_find_device(fs_info, devid, dev_uuid, fs_uuid);
2297                 BUG_ON(!device); /* Logic error */
2298
2299                 if (device->fs_devices->seeding) {
2300                         btrfs_set_device_generation(leaf, dev_item,
2301                                                     device->generation);
2302                         btrfs_mark_buffer_dirty(leaf);
2303                 }
2304
2305                 path->slots[0]++;
2306                 goto next_slot;
2307         }
2308         ret = 0;
2309 error:
2310         btrfs_free_path(path);
2311         return ret;
2312 }
2313
2314 int btrfs_init_new_device(struct btrfs_fs_info *fs_info, const char *device_path)
2315 {
2316         struct btrfs_root *root = fs_info->dev_root;
2317         struct request_queue *q;
2318         struct btrfs_trans_handle *trans;
2319         struct btrfs_device *device;
2320         struct block_device *bdev;
2321         struct super_block *sb = fs_info->sb;
2322         struct rcu_string *name;
2323         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2324         u64 tmp;
2325         int seeding_dev = 0;
2326         int ret = 0;
2327         bool unlocked = false;
2328
2329         if (sb_rdonly(sb) && !fs_devices->seeding)
2330                 return -EROFS;
2331
2332         bdev = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
2333                                   fs_info->bdev_holder);
2334         if (IS_ERR(bdev))
2335                 return PTR_ERR(bdev);
2336
2337         if (fs_devices->seeding) {
2338                 seeding_dev = 1;
2339                 down_write(&sb->s_umount);
2340                 mutex_lock(&uuid_mutex);
2341         }
2342
2343         filemap_write_and_wait(bdev->bd_inode->i_mapping);
2344
2345         mutex_lock(&fs_devices->device_list_mutex);
2346         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2347                 if (device->bdev == bdev) {
2348                         ret = -EEXIST;
2349                         mutex_unlock(
2350                                 &fs_devices->device_list_mutex);
2351                         goto error;
2352                 }
2353         }
2354         mutex_unlock(&fs_devices->device_list_mutex);
2355
2356         device = btrfs_alloc_device(fs_info, NULL, NULL);
2357         if (IS_ERR(device)) {
2358                 /* we can safely leave the fs_devices entry around */
2359                 ret = PTR_ERR(device);
2360                 goto error;
2361         }
2362
2363         name = rcu_string_strdup(device_path, GFP_KERNEL);
2364         if (!name) {
2365                 ret = -ENOMEM;
2366                 goto error_free_device;
2367         }
2368         rcu_assign_pointer(device->name, name);
2369
2370         trans = btrfs_start_transaction(root, 0);
2371         if (IS_ERR(trans)) {
2372                 ret = PTR_ERR(trans);
2373                 goto error_free_device;
2374         }
2375
2376         q = bdev_get_queue(bdev);
2377         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
2378         device->generation = trans->transid;
2379         device->io_width = fs_info->sectorsize;
2380         device->io_align = fs_info->sectorsize;
2381         device->sector_size = fs_info->sectorsize;
2382         device->total_bytes = round_down(i_size_read(bdev->bd_inode),
2383                                          fs_info->sectorsize);
2384         device->disk_total_bytes = device->total_bytes;
2385         device->commit_total_bytes = device->total_bytes;
2386         device->fs_info = fs_info;
2387         device->bdev = bdev;
2388         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
2389         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
2390         device->mode = FMODE_EXCL;
2391         device->dev_stats_valid = 1;
2392         set_blocksize(device->bdev, BTRFS_BDEV_BLOCKSIZE);
2393
2394         if (seeding_dev) {
2395                 sb->s_flags &= ~SB_RDONLY;
2396                 ret = btrfs_prepare_sprout(fs_info);
2397                 if (ret) {
2398                         btrfs_abort_transaction(trans, ret);
2399                         goto error_trans;
2400                 }
2401         }
2402
2403         device->fs_devices = fs_devices;
2404
2405         mutex_lock(&fs_devices->device_list_mutex);
2406         mutex_lock(&fs_info->chunk_mutex);
2407         list_add_rcu(&device->dev_list, &fs_devices->devices);
2408         list_add(&device->dev_alloc_list, &fs_devices->alloc_list);
2409         fs_devices->num_devices++;
2410         fs_devices->open_devices++;
2411         fs_devices->rw_devices++;
2412         fs_devices->total_devices++;
2413         fs_devices->total_rw_bytes += device->total_bytes;
2414
2415         atomic64_add(device->total_bytes, &fs_info->free_chunk_space);
2416
2417         if (!blk_queue_nonrot(q))
2418                 fs_devices->rotating = 1;
2419
2420         tmp = btrfs_super_total_bytes(fs_info->super_copy);
2421         btrfs_set_super_total_bytes(fs_info->super_copy,
2422                 round_down(tmp + device->total_bytes, fs_info->sectorsize));
2423
2424         tmp = btrfs_super_num_devices(fs_info->super_copy);
2425         btrfs_set_super_num_devices(fs_info->super_copy, tmp + 1);
2426
2427         /* add sysfs device entry */
2428         btrfs_sysfs_add_device_link(fs_devices, device);
2429
2430         /*
2431          * we've got more storage, clear any full flags on the space
2432          * infos
2433          */
2434         btrfs_clear_space_info_full(fs_info);
2435
2436         mutex_unlock(&fs_info->chunk_mutex);
2437         mutex_unlock(&fs_devices->device_list_mutex);
2438
2439         if (seeding_dev) {
2440                 mutex_lock(&fs_info->chunk_mutex);
2441                 ret = init_first_rw_device(trans, fs_info);
2442                 mutex_unlock(&fs_info->chunk_mutex);
2443                 if (ret) {
2444                         btrfs_abort_transaction(trans, ret);
2445                         goto error_sysfs;
2446                 }
2447         }
2448
2449         ret = btrfs_add_dev_item(trans, device);
2450         if (ret) {
2451                 btrfs_abort_transaction(trans, ret);
2452                 goto error_sysfs;
2453         }
2454
2455         if (seeding_dev) {
2456                 char fsid_buf[BTRFS_UUID_UNPARSED_SIZE];
2457
2458                 ret = btrfs_finish_sprout(trans, fs_info);
2459                 if (ret) {
2460                         btrfs_abort_transaction(trans, ret);
2461                         goto error_sysfs;
2462                 }
2463
2464                 /* Sprouting would change fsid of the mounted root,
2465                  * so rename the fsid on the sysfs
2466                  */
2467                 snprintf(fsid_buf, BTRFS_UUID_UNPARSED_SIZE, "%pU",
2468                                                 fs_info->fsid);
2469                 if (kobject_rename(&fs_devices->fsid_kobj, fsid_buf))
2470                         btrfs_warn(fs_info,
2471                                    "sysfs: failed to create fsid for sprout");
2472         }
2473
2474         ret = btrfs_commit_transaction(trans);
2475
2476         if (seeding_dev) {
2477                 mutex_unlock(&uuid_mutex);
2478                 up_write(&sb->s_umount);
2479                 unlocked = true;
2480
2481                 if (ret) /* transaction commit */
2482                         return ret;
2483
2484                 ret = btrfs_relocate_sys_chunks(fs_info);
2485                 if (ret < 0)
2486                         btrfs_handle_fs_error(fs_info, ret,
2487                                     "Failed to relocate sys chunks after device initialization. This can be fixed using the \"btrfs balance\" command.");
2488                 trans = btrfs_attach_transaction(root);
2489                 if (IS_ERR(trans)) {
2490                         if (PTR_ERR(trans) == -ENOENT)
2491                                 return 0;
2492                         ret = PTR_ERR(trans);
2493                         trans = NULL;
2494                         goto error_sysfs;
2495                 }
2496                 ret = btrfs_commit_transaction(trans);
2497         }
2498
2499         /* Update ctime/mtime for libblkid */
2500         update_dev_time(device_path);
2501         return ret;
2502
2503 error_sysfs:
2504         btrfs_sysfs_rm_device_link(fs_devices, device);
2505 error_trans:
2506         if (seeding_dev)
2507                 sb->s_flags |= SB_RDONLY;
2508         if (trans)
2509                 btrfs_end_transaction(trans);
2510 error_free_device:
2511         btrfs_free_device(device);
2512 error:
2513         blkdev_put(bdev, FMODE_EXCL);
2514         if (seeding_dev && !unlocked) {
2515                 mutex_unlock(&uuid_mutex);
2516                 up_write(&sb->s_umount);
2517         }
2518         return ret;
2519 }
2520
2521 static noinline int btrfs_update_device(struct btrfs_trans_handle *trans,
2522                                         struct btrfs_device *device)
2523 {
2524         int ret;
2525         struct btrfs_path *path;
2526         struct btrfs_root *root = device->fs_info->chunk_root;
2527         struct btrfs_dev_item *dev_item;
2528         struct extent_buffer *leaf;
2529         struct btrfs_key key;
2530
2531         path = btrfs_alloc_path();
2532         if (!path)
2533                 return -ENOMEM;
2534
2535         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
2536         key.type = BTRFS_DEV_ITEM_KEY;
2537         key.offset = device->devid;
2538
2539         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2540         if (ret < 0)
2541                 goto out;
2542
2543         if (ret > 0) {
2544                 ret = -ENOENT;
2545                 goto out;
2546         }
2547
2548         leaf = path->nodes[0];
2549         dev_item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_item);
2550
2551         btrfs_set_device_id(leaf, dev_item, device->devid);
2552         btrfs_set_device_type(leaf, dev_item, device->type);
2553         btrfs_set_device_io_align(leaf, dev_item, device->io_align);
2554         btrfs_set_device_io_width(leaf, dev_item, device->io_width);
2555         btrfs_set_device_sector_size(leaf, dev_item, device->sector_size);
2556         btrfs_set_device_total_bytes(leaf, dev_item,
2557                                      btrfs_device_get_disk_total_bytes(device));
2558         btrfs_set_device_bytes_used(leaf, dev_item,
2559                                     btrfs_device_get_bytes_used(device));
2560         btrfs_mark_buffer_dirty(leaf);
2561
2562 out:
2563         btrfs_free_path(path);
2564         return ret;
2565 }
2566
2567 int btrfs_grow_device(struct btrfs_trans_handle *trans,
2568                       struct btrfs_device *device, u64 new_size)
2569 {
2570         struct btrfs_fs_info *fs_info = device->fs_info;
2571         struct btrfs_super_block *super_copy = fs_info->super_copy;
2572         struct btrfs_fs_devices *fs_devices;
2573         u64 old_total;
2574         u64 diff;
2575
2576         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
2577                 return -EACCES;
2578
2579         new_size = round_down(new_size, fs_info->sectorsize);
2580
2581         mutex_lock(&fs_info->chunk_mutex);
2582         old_total = btrfs_super_total_bytes(super_copy);
2583         diff = round_down(new_size - device->total_bytes, fs_info->sectorsize);
2584
2585         if (new_size <= device->total_bytes ||
2586             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
2587                 mutex_unlock(&fs_info->chunk_mutex);
2588                 return -EINVAL;
2589         }
2590
2591         fs_devices = fs_info->fs_devices;
2592
2593         btrfs_set_super_total_bytes(super_copy,
2594                         round_down(old_total + diff, fs_info->sectorsize));
2595         device->fs_devices->total_rw_bytes += diff;
2596
2597         btrfs_device_set_total_bytes(device, new_size);
2598         btrfs_device_set_disk_total_bytes(device, new_size);
2599         btrfs_clear_space_info_full(device->fs_info);
2600         if (list_empty(&device->resized_list))
2601                 list_add_tail(&device->resized_list,
2602                               &fs_devices->resized_devices);
2603         mutex_unlock(&fs_info->chunk_mutex);
2604
2605         return btrfs_update_device(trans, device);
2606 }
2607
2608 static int btrfs_free_chunk(struct btrfs_trans_handle *trans, u64 chunk_offset)
2609 {
2610         struct btrfs_fs_info *fs_info = trans->fs_info;
2611         struct btrfs_root *root = fs_info->chunk_root;
2612         int ret;
2613         struct btrfs_path *path;
2614         struct btrfs_key key;
2615
2616         path = btrfs_alloc_path();
2617         if (!path)
2618                 return -ENOMEM;
2619
2620         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2621         key.offset = chunk_offset;
2622         key.type = BTRFS_CHUNK_ITEM_KEY;
2623
2624         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
2625         if (ret < 0)
2626                 goto out;
2627         else if (ret > 0) { /* Logic error or corruption */
2628                 btrfs_handle_fs_error(fs_info, -ENOENT,
2629                                       "Failed lookup while freeing chunk.");
2630                 ret = -ENOENT;
2631                 goto out;
2632         }
2633
2634         ret = btrfs_del_item(trans, root, path);
2635         if (ret < 0)
2636                 btrfs_handle_fs_error(fs_info, ret,
2637                                       "Failed to delete chunk item.");
2638 out:
2639         btrfs_free_path(path);
2640         return ret;
2641 }
2642
2643 static int btrfs_del_sys_chunk(struct btrfs_fs_info *fs_info, u64 chunk_offset)
2644 {
2645         struct btrfs_super_block *super_copy = fs_info->super_copy;
2646         struct btrfs_disk_key *disk_key;
2647         struct btrfs_chunk *chunk;
2648         u8 *ptr;
2649         int ret = 0;
2650         u32 num_stripes;
2651         u32 array_size;
2652         u32 len = 0;
2653         u32 cur;
2654         struct btrfs_key key;
2655
2656         mutex_lock(&fs_info->chunk_mutex);
2657         array_size = btrfs_super_sys_array_size(super_copy);
2658
2659         ptr = super_copy->sys_chunk_array;
2660         cur = 0;
2661
2662         while (cur < array_size) {
2663                 disk_key = (struct btrfs_disk_key *)ptr;
2664                 btrfs_disk_key_to_cpu(&key, disk_key);
2665
2666                 len = sizeof(*disk_key);
2667
2668                 if (key.type == BTRFS_CHUNK_ITEM_KEY) {
2669                         chunk = (struct btrfs_chunk *)(ptr + len);
2670                         num_stripes = btrfs_stack_chunk_num_stripes(chunk);
2671                         len += btrfs_chunk_item_size(num_stripes);
2672                 } else {
2673                         ret = -EIO;
2674                         break;
2675                 }
2676                 if (key.objectid == BTRFS_FIRST_CHUNK_TREE_OBJECTID &&
2677                     key.offset == chunk_offset) {
2678                         memmove(ptr, ptr + len, array_size - (cur + len));
2679                         array_size -= len;
2680                         btrfs_set_super_sys_array_size(super_copy, array_size);
2681                 } else {
2682                         ptr += len;
2683                         cur += len;
2684                 }
2685         }
2686         mutex_unlock(&fs_info->chunk_mutex);
2687         return ret;
2688 }
2689
2690 static struct extent_map *get_chunk_map(struct btrfs_fs_info *fs_info,
2691                                         u64 logical, u64 length)
2692 {
2693         struct extent_map_tree *em_tree;
2694         struct extent_map *em;
2695
2696         em_tree = &fs_info->mapping_tree.map_tree;
2697         read_lock(&em_tree->lock);
2698         em = lookup_extent_mapping(em_tree, logical, length);
2699         read_unlock(&em_tree->lock);
2700
2701         if (!em) {
2702                 btrfs_crit(fs_info, "unable to find logical %llu length %llu",
2703                            logical, length);
2704                 return ERR_PTR(-EINVAL);
2705         }
2706
2707         if (em->start > logical || em->start + em->len < logical) {
2708                 btrfs_crit(fs_info,
2709                            "found a bad mapping, wanted %llu-%llu, found %llu-%llu",
2710                            logical, length, em->start, em->start + em->len);
2711                 free_extent_map(em);
2712                 return ERR_PTR(-EINVAL);
2713         }
2714
2715         /* callers are responsible for dropping em's ref. */
2716         return em;
2717 }
2718
2719 int btrfs_remove_chunk(struct btrfs_trans_handle *trans, u64 chunk_offset)
2720 {
2721         struct btrfs_fs_info *fs_info = trans->fs_info;
2722         struct extent_map *em;
2723         struct map_lookup *map;
2724         u64 dev_extent_len = 0;
2725         int i, ret = 0;
2726         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2727
2728         em = get_chunk_map(fs_info, chunk_offset, 1);
2729         if (IS_ERR(em)) {
2730                 /*
2731                  * This is a logic error, but we don't want to just rely on the
2732                  * user having built with ASSERT enabled, so if ASSERT doesn't
2733                  * do anything we still error out.
2734                  */
2735                 ASSERT(0);
2736                 return PTR_ERR(em);
2737         }
2738         map = em->map_lookup;
2739         mutex_lock(&fs_info->chunk_mutex);
2740         check_system_chunk(trans, map->type);
2741         mutex_unlock(&fs_info->chunk_mutex);
2742
2743         /*
2744          * Take the device list mutex to prevent races with the final phase of
2745          * a device replace operation that replaces the device object associated
2746          * with map stripes (dev-replace.c:btrfs_dev_replace_finishing()).
2747          */
2748         mutex_lock(&fs_devices->device_list_mutex);
2749         for (i = 0; i < map->num_stripes; i++) {
2750                 struct btrfs_device *device = map->stripes[i].dev;
2751                 ret = btrfs_free_dev_extent(trans, device,
2752                                             map->stripes[i].physical,
2753                                             &dev_extent_len);
2754                 if (ret) {
2755                         mutex_unlock(&fs_devices->device_list_mutex);
2756                         btrfs_abort_transaction(trans, ret);
2757                         goto out;
2758                 }
2759
2760                 if (device->bytes_used > 0) {
2761                         mutex_lock(&fs_info->chunk_mutex);
2762                         btrfs_device_set_bytes_used(device,
2763                                         device->bytes_used - dev_extent_len);
2764                         atomic64_add(dev_extent_len, &fs_info->free_chunk_space);
2765                         btrfs_clear_space_info_full(fs_info);
2766                         mutex_unlock(&fs_info->chunk_mutex);
2767                 }
2768
2769                 if (map->stripes[i].dev) {
2770                         ret = btrfs_update_device(trans, map->stripes[i].dev);
2771                         if (ret) {
2772                                 mutex_unlock(&fs_devices->device_list_mutex);
2773                                 btrfs_abort_transaction(trans, ret);
2774                                 goto out;
2775                         }
2776                 }
2777         }
2778         mutex_unlock(&fs_devices->device_list_mutex);
2779
2780         ret = btrfs_free_chunk(trans, chunk_offset);
2781         if (ret) {
2782                 btrfs_abort_transaction(trans, ret);
2783                 goto out;
2784         }
2785
2786         trace_btrfs_chunk_free(fs_info, map, chunk_offset, em->len);
2787
2788         if (map->type & BTRFS_BLOCK_GROUP_SYSTEM) {
2789                 ret = btrfs_del_sys_chunk(fs_info, chunk_offset);
2790                 if (ret) {
2791                         btrfs_abort_transaction(trans, ret);
2792                         goto out;
2793                 }
2794         }
2795
2796         ret = btrfs_remove_block_group(trans, chunk_offset, em);
2797         if (ret) {
2798                 btrfs_abort_transaction(trans, ret);
2799                 goto out;
2800         }
2801
2802 out:
2803         /* once for us */
2804         free_extent_map(em);
2805         return ret;
2806 }
2807
2808 static int btrfs_relocate_chunk(struct btrfs_fs_info *fs_info, u64 chunk_offset)
2809 {
2810         struct btrfs_root *root = fs_info->chunk_root;
2811         struct btrfs_trans_handle *trans;
2812         int ret;
2813
2814         /*
2815          * Prevent races with automatic removal of unused block groups.
2816          * After we relocate and before we remove the chunk with offset
2817          * chunk_offset, automatic removal of the block group can kick in,
2818          * resulting in a failure when calling btrfs_remove_chunk() below.
2819          *
2820          * Make sure to acquire this mutex before doing a tree search (dev
2821          * or chunk trees) to find chunks. Otherwise the cleaner kthread might
2822          * call btrfs_remove_chunk() (through btrfs_delete_unused_bgs()) after
2823          * we release the path used to search the chunk/dev tree and before
2824          * the current task acquires this mutex and calls us.
2825          */
2826         lockdep_assert_held(&fs_info->delete_unused_bgs_mutex);
2827
2828         ret = btrfs_can_relocate(fs_info, chunk_offset);
2829         if (ret)
2830                 return -ENOSPC;
2831
2832         /* step one, relocate all the extents inside this chunk */
2833         btrfs_scrub_pause(fs_info);
2834         ret = btrfs_relocate_block_group(fs_info, chunk_offset);
2835         btrfs_scrub_continue(fs_info);
2836         if (ret)
2837                 return ret;
2838
2839         /*
2840          * We add the kobjects here (and after forcing data chunk creation)
2841          * since relocation is the only place we'll create chunks of a new
2842          * type at runtime.  The only place where we'll remove the last
2843          * chunk of a type is the call immediately below this one.  Even
2844          * so, we're protected against races with the cleaner thread since
2845          * we're covered by the delete_unused_bgs_mutex.
2846          */
2847         btrfs_add_raid_kobjects(fs_info);
2848
2849         trans = btrfs_start_trans_remove_block_group(root->fs_info,
2850                                                      chunk_offset);
2851         if (IS_ERR(trans)) {
2852                 ret = PTR_ERR(trans);
2853                 btrfs_handle_fs_error(root->fs_info, ret, NULL);
2854                 return ret;
2855         }
2856
2857         /*
2858          * step two, delete the device extents and the
2859          * chunk tree entries
2860          */
2861         ret = btrfs_remove_chunk(trans, chunk_offset);
2862         btrfs_end_transaction(trans);
2863         return ret;
2864 }
2865
2866 static int btrfs_relocate_sys_chunks(struct btrfs_fs_info *fs_info)
2867 {
2868         struct btrfs_root *chunk_root = fs_info->chunk_root;
2869         struct btrfs_path *path;
2870         struct extent_buffer *leaf;
2871         struct btrfs_chunk *chunk;
2872         struct btrfs_key key;
2873         struct btrfs_key found_key;
2874         u64 chunk_type;
2875         bool retried = false;
2876         int failed = 0;
2877         int ret;
2878
2879         path = btrfs_alloc_path();
2880         if (!path)
2881                 return -ENOMEM;
2882
2883 again:
2884         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2885         key.offset = (u64)-1;
2886         key.type = BTRFS_CHUNK_ITEM_KEY;
2887
2888         while (1) {
2889                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
2890                 ret = btrfs_search_slot(NULL, chunk_root, &key, path, 0, 0);
2891                 if (ret < 0) {
2892                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
2893                         goto error;
2894                 }
2895                 BUG_ON(ret == 0); /* Corruption */
2896
2897                 ret = btrfs_previous_item(chunk_root, path, key.objectid,
2898                                           key.type);
2899                 if (ret)
2900                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
2901                 if (ret < 0)
2902                         goto error;
2903                 if (ret > 0)
2904                         break;
2905
2906                 leaf = path->nodes[0];
2907                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2908
2909                 chunk = btrfs_item_ptr(leaf, path->slots[0],
2910                                        struct btrfs_chunk);
2911                 chunk_type = btrfs_chunk_type(leaf, chunk);
2912                 btrfs_release_path(path);
2913
2914                 if (chunk_type & BTRFS_BLOCK_GROUP_SYSTEM) {
2915                         ret = btrfs_relocate_chunk(fs_info, found_key.offset);
2916                         if (ret == -ENOSPC)
2917                                 failed++;
2918                         else
2919                                 BUG_ON(ret);
2920                 }
2921                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
2922
2923                 if (found_key.offset == 0)
2924                         break;
2925                 key.offset = found_key.offset - 1;
2926         }
2927         ret = 0;
2928         if (failed && !retried) {
2929                 failed = 0;
2930                 retried = true;
2931                 goto again;
2932         } else if (WARN_ON(failed && retried)) {
2933                 ret = -ENOSPC;
2934         }
2935 error:
2936         btrfs_free_path(path);
2937         return ret;
2938 }
2939
2940 /*
2941  * return 1 : allocate a data chunk successfully,
2942  * return <0: errors during allocating a data chunk,
2943  * return 0 : no need to allocate a data chunk.
2944  */
2945 static int btrfs_may_alloc_data_chunk(struct btrfs_fs_info *fs_info,
2946                                       u64 chunk_offset)
2947 {
2948         struct btrfs_block_group_cache *cache;
2949         u64 bytes_used;
2950         u64 chunk_type;
2951
2952         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2953         ASSERT(cache);
2954         chunk_type = cache->flags;
2955         btrfs_put_block_group(cache);
2956
2957         if (chunk_type & BTRFS_BLOCK_GROUP_DATA) {
2958                 spin_lock(&fs_info->data_sinfo->lock);
2959                 bytes_used = fs_info->data_sinfo->bytes_used;
2960                 spin_unlock(&fs_info->data_sinfo->lock);
2961
2962                 if (!bytes_used) {
2963                         struct btrfs_trans_handle *trans;
2964                         int ret;
2965
2966                         trans = btrfs_join_transaction(fs_info->tree_root);
2967                         if (IS_ERR(trans))
2968                                 return PTR_ERR(trans);
2969
2970                         ret = btrfs_force_chunk_alloc(trans,
2971                                                       BTRFS_BLOCK_GROUP_DATA);
2972                         btrfs_end_transaction(trans);
2973                         if (ret < 0)
2974                                 return ret;
2975
2976                         btrfs_add_raid_kobjects(fs_info);
2977
2978                         return 1;
2979                 }
2980         }
2981         return 0;
2982 }
2983
2984 static int insert_balance_item(struct btrfs_fs_info *fs_info,
2985                                struct btrfs_balance_control *bctl)
2986 {
2987         struct btrfs_root *root = fs_info->tree_root;
2988         struct btrfs_trans_handle *trans;
2989         struct btrfs_balance_item *item;
2990         struct btrfs_disk_balance_args disk_bargs;
2991         struct btrfs_path *path;
2992         struct extent_buffer *leaf;
2993         struct btrfs_key key;
2994         int ret, err;
2995
2996         path = btrfs_alloc_path();
2997         if (!path)
2998                 return -ENOMEM;
2999
3000         trans = btrfs_start_transaction(root, 0);
3001         if (IS_ERR(trans)) {
3002                 btrfs_free_path(path);
3003                 return PTR_ERR(trans);
3004         }
3005
3006         key.objectid = BTRFS_BALANCE_OBJECTID;
3007         key.type = BTRFS_TEMPORARY_ITEM_KEY;
3008         key.offset = 0;
3009
3010         ret = btrfs_insert_empty_item(trans, root, path, &key,
3011                                       sizeof(*item));
3012         if (ret)
3013                 goto out;
3014
3015         leaf = path->nodes[0];
3016         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_balance_item);
3017
3018         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3019
3020         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->data);
3021         btrfs_set_balance_data(leaf, item, &disk_bargs);
3022         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->meta);
3023         btrfs_set_balance_meta(leaf, item, &disk_bargs);
3024         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->sys);
3025         btrfs_set_balance_sys(leaf, item, &disk_bargs);
3026
3027         btrfs_set_balance_flags(leaf, item, bctl->flags);
3028
3029         btrfs_mark_buffer_dirty(leaf);
3030 out:
3031         btrfs_free_path(path);
3032         err = btrfs_commit_transaction(trans);
3033         if (err && !ret)
3034                 ret = err;
3035         return ret;
3036 }
3037
3038 static int del_balance_item(struct btrfs_fs_info *fs_info)
3039 {
3040         struct btrfs_root *root = fs_info->tree_root;
3041         struct btrfs_trans_handle *trans;
3042         struct btrfs_path *path;
3043         struct btrfs_key key;
3044         int ret, err;
3045
3046         path = btrfs_alloc_path();
3047         if (!path)
3048                 return -ENOMEM;
3049
3050         trans = btrfs_start_transaction(root, 0);
3051         if (IS_ERR(trans)) {
3052                 btrfs_free_path(path);
3053                 return PTR_ERR(trans);
3054         }
3055
3056         key.objectid = BTRFS_BALANCE_OBJECTID;
3057         key.type = BTRFS_TEMPORARY_ITEM_KEY;
3058         key.offset = 0;
3059
3060         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3061         if (ret < 0)
3062                 goto out;
3063         if (ret > 0) {
3064                 ret = -ENOENT;
3065                 goto out;
3066         }
3067
3068         ret = btrfs_del_item(trans, root, path);
3069 out:
3070         btrfs_free_path(path);
3071         err = btrfs_commit_transaction(trans);
3072         if (err && !ret)
3073                 ret = err;
3074         return ret;
3075 }
3076
3077 /*
3078  * This is a heuristic used to reduce the number of chunks balanced on
3079  * resume after balance was interrupted.
3080  */
3081 static void update_balance_args(struct btrfs_balance_control *bctl)
3082 {
3083         /*
3084          * Turn on soft mode for chunk types that were being converted.
3085          */
3086         if (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT)
3087                 bctl->data.flags |= BTRFS_BALANCE_ARGS_SOFT;
3088         if (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT)
3089                 bctl->sys.flags |= BTRFS_BALANCE_ARGS_SOFT;
3090         if (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT)
3091                 bctl->meta.flags |= BTRFS_BALANCE_ARGS_SOFT;
3092
3093         /*
3094          * Turn on usage filter if is not already used.  The idea is
3095          * that chunks that we have already balanced should be
3096          * reasonably full.  Don't do it for chunks that are being
3097          * converted - that will keep us from relocating unconverted
3098          * (albeit full) chunks.
3099          */
3100         if (!(bctl->data.flags & BTRFS_BALANCE_ARGS_USAGE) &&
3101             !(bctl->data.flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3102             !(bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
3103                 bctl->data.flags |= BTRFS_BALANCE_ARGS_USAGE;
3104                 bctl->data.usage = 90;
3105         }
3106         if (!(bctl->sys.flags & BTRFS_BALANCE_ARGS_USAGE) &&
3107             !(bctl->sys.flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3108             !(bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
3109                 bctl->sys.flags |= BTRFS_BALANCE_ARGS_USAGE;
3110                 bctl->sys.usage = 90;
3111         }
3112         if (!(bctl->meta.flags & BTRFS_BALANCE_ARGS_USAGE) &&
3113             !(bctl->meta.flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3114             !(bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
3115                 bctl->meta.flags |= BTRFS_BALANCE_ARGS_USAGE;
3116                 bctl->meta.usage = 90;
3117         }
3118 }
3119
3120 /*
3121  * Clear the balance status in fs_info and delete the balance item from disk.
3122  */
3123 static void reset_balance_state(struct btrfs_fs_info *fs_info)
3124 {
3125         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3126         int ret;
3127
3128         BUG_ON(!fs_info->balance_ctl);
3129
3130         spin_lock(&fs_info->balance_lock);
3131         fs_info->balance_ctl = NULL;
3132         spin_unlock(&fs_info->balance_lock);
3133
3134         kfree(bctl);
3135         ret = del_balance_item(fs_info);
3136         if (ret)
3137                 btrfs_handle_fs_error(fs_info, ret, NULL);
3138 }
3139
3140 /*
3141  * Balance filters.  Return 1 if chunk should be filtered out
3142  * (should not be balanced).
3143  */
3144 static int chunk_profiles_filter(u64 chunk_type,
3145                                  struct btrfs_balance_args *bargs)
3146 {
3147         chunk_type = chunk_to_extended(chunk_type) &
3148                                 BTRFS_EXTENDED_PROFILE_MASK;
3149
3150         if (bargs->profiles & chunk_type)
3151                 return 0;
3152
3153         return 1;
3154 }
3155
3156 static int chunk_usage_range_filter(struct btrfs_fs_info *fs_info, u64 chunk_offset,
3157                               struct btrfs_balance_args *bargs)
3158 {
3159         struct btrfs_block_group_cache *cache;
3160         u64 chunk_used;
3161         u64 user_thresh_min;
3162         u64 user_thresh_max;
3163         int ret = 1;
3164
3165         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3166         chunk_used = btrfs_block_group_used(&cache->item);
3167
3168         if (bargs->usage_min == 0)
3169                 user_thresh_min = 0;
3170         else
3171                 user_thresh_min = div_factor_fine(cache->key.offset,
3172                                         bargs->usage_min);
3173
3174         if (bargs->usage_max == 0)
3175                 user_thresh_max = 1;
3176         else if (bargs->usage_max > 100)
3177                 user_thresh_max = cache->key.offset;
3178         else
3179                 user_thresh_max = div_factor_fine(cache->key.offset,
3180                                         bargs->usage_max);
3181
3182         if (user_thresh_min <= chunk_used && chunk_used < user_thresh_max)
3183                 ret = 0;
3184
3185         btrfs_put_block_group(cache);
3186         return ret;
3187 }
3188
3189 static int chunk_usage_filter(struct btrfs_fs_info *fs_info,
3190                 u64 chunk_offset, struct btrfs_balance_args *bargs)
3191 {
3192         struct btrfs_block_group_cache *cache;
3193         u64 chunk_used, user_thresh;
3194         int ret = 1;
3195
3196         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3197         chunk_used = btrfs_block_group_used(&cache->item);
3198
3199         if (bargs->usage_min == 0)
3200                 user_thresh = 1;
3201         else if (bargs->usage > 100)
3202                 user_thresh = cache->key.offset;
3203         else
3204                 user_thresh = div_factor_fine(cache->key.offset,
3205                                               bargs->usage);
3206
3207         if (chunk_used < user_thresh)
3208                 ret = 0;
3209
3210         btrfs_put_block_group(cache);
3211         return ret;
3212 }
3213
3214 static int chunk_devid_filter(struct extent_buffer *leaf,
3215                               struct btrfs_chunk *chunk,
3216                               struct btrfs_balance_args *bargs)
3217 {
3218         struct btrfs_stripe *stripe;
3219         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
3220         int i;
3221
3222         for (i = 0; i < num_stripes; i++) {
3223                 stripe = btrfs_stripe_nr(chunk, i);
3224                 if (btrfs_stripe_devid(leaf, stripe) == bargs->devid)
3225                         return 0;
3226         }
3227
3228         return 1;
3229 }
3230
3231 /* [pstart, pend) */
3232 static int chunk_drange_filter(struct extent_buffer *leaf,
3233                                struct btrfs_chunk *chunk,
3234                                struct btrfs_balance_args *bargs)
3235 {
3236         struct btrfs_stripe *stripe;
3237         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
3238         u64 stripe_offset;
3239         u64 stripe_length;
3240         int factor;
3241         int i;
3242
3243         if (!(bargs->flags & BTRFS_BALANCE_ARGS_DEVID))
3244                 return 0;
3245
3246         if (btrfs_chunk_type(leaf, chunk) & (BTRFS_BLOCK_GROUP_DUP |
3247              BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10)) {
3248                 factor = num_stripes / 2;
3249         } else if (btrfs_chunk_type(leaf, chunk) & BTRFS_BLOCK_GROUP_RAID5) {
3250                 factor = num_stripes - 1;
3251         } else if (btrfs_chunk_type(leaf, chunk) & BTRFS_BLOCK_GROUP_RAID6) {
3252                 factor = num_stripes - 2;
3253         } else {
3254                 factor = num_stripes;
3255         }
3256
3257         for (i = 0; i < num_stripes; i++) {
3258                 stripe = btrfs_stripe_nr(chunk, i);
3259                 if (btrfs_stripe_devid(leaf, stripe) != bargs->devid)
3260                         continue;
3261
3262                 stripe_offset = btrfs_stripe_offset(leaf, stripe);
3263                 stripe_length = btrfs_chunk_length(leaf, chunk);
3264                 stripe_length = div_u64(stripe_length, factor);
3265
3266                 if (stripe_offset < bargs->pend &&
3267                     stripe_offset + stripe_length > bargs->pstart)
3268                         return 0;
3269         }
3270
3271         return 1;
3272 }
3273
3274 /* [vstart, vend) */
3275 static int chunk_vrange_filter(struct extent_buffer *leaf,
3276                                struct btrfs_chunk *chunk,
3277                                u64 chunk_offset,
3278                                struct btrfs_balance_args *bargs)
3279 {
3280         if (chunk_offset < bargs->vend &&
3281             chunk_offset + btrfs_chunk_length(leaf, chunk) > bargs->vstart)
3282                 /* at least part of the chunk is inside this vrange */
3283                 return 0;
3284
3285         return 1;
3286 }
3287
3288 static int chunk_stripes_range_filter(struct extent_buffer *leaf,
3289                                struct btrfs_chunk *chunk,
3290                                struct btrfs_balance_args *bargs)
3291 {
3292         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
3293
3294         if (bargs->stripes_min <= num_stripes
3295                         && num_stripes <= bargs->stripes_max)
3296                 return 0;
3297
3298         return 1;
3299 }
3300
3301 static int chunk_soft_convert_filter(u64 chunk_type,
3302                                      struct btrfs_balance_args *bargs)
3303 {
3304         if (!(bargs->flags & BTRFS_BALANCE_ARGS_CONVERT))
3305                 return 0;
3306
3307         chunk_type = chunk_to_extended(chunk_type) &
3308                                 BTRFS_EXTENDED_PROFILE_MASK;
3309
3310         if (bargs->target == chunk_type)
3311                 return 1;
3312
3313         return 0;
3314 }
3315
3316 static int should_balance_chunk(struct btrfs_fs_info *fs_info,
3317                                 struct extent_buffer *leaf,
3318                                 struct btrfs_chunk *chunk, u64 chunk_offset)
3319 {
3320         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3321         struct btrfs_balance_args *bargs = NULL;
3322         u64 chunk_type = btrfs_chunk_type(leaf, chunk);
3323
3324         /* type filter */
3325         if (!((chunk_type & BTRFS_BLOCK_GROUP_TYPE_MASK) &
3326               (bctl->flags & BTRFS_BALANCE_TYPE_MASK))) {
3327                 return 0;
3328         }
3329
3330         if (chunk_type & BTRFS_BLOCK_GROUP_DATA)
3331                 bargs = &bctl->data;
3332         else if (chunk_type & BTRFS_BLOCK_GROUP_SYSTEM)
3333                 bargs = &bctl->sys;
3334         else if (chunk_type & BTRFS_BLOCK_GROUP_METADATA)
3335                 bargs = &bctl->meta;
3336
3337         /* profiles filter */
3338         if ((bargs->flags & BTRFS_BALANCE_ARGS_PROFILES) &&
3339             chunk_profiles_filter(chunk_type, bargs)) {
3340                 return 0;
3341         }
3342
3343         /* usage filter */
3344         if ((bargs->flags & BTRFS_BALANCE_ARGS_USAGE) &&
3345             chunk_usage_filter(fs_info, chunk_offset, bargs)) {
3346                 return 0;
3347         } else if ((bargs->flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3348             chunk_usage_range_filter(fs_info, chunk_offset, bargs)) {
3349                 return 0;
3350         }
3351
3352         /* devid filter */
3353         if ((bargs->flags & BTRFS_BALANCE_ARGS_DEVID) &&
3354             chunk_devid_filter(leaf, chunk, bargs)) {
3355                 return 0;
3356         }
3357
3358         /* drange filter, makes sense only with devid filter */
3359         if ((bargs->flags & BTRFS_BALANCE_ARGS_DRANGE) &&
3360             chunk_drange_filter(leaf, chunk, bargs)) {
3361                 return 0;
3362         }
3363
3364         /* vrange filter */
3365         if ((bargs->flags & BTRFS_BALANCE_ARGS_VRANGE) &&
3366             chunk_vrange_filter(leaf, chunk, chunk_offset, bargs)) {
3367                 return 0;
3368         }
3369
3370         /* stripes filter */
3371         if ((bargs->flags & BTRFS_BALANCE_ARGS_STRIPES_RANGE) &&
3372             chunk_stripes_range_filter(leaf, chunk, bargs)) {
3373                 return 0;
3374         }
3375
3376         /* soft profile changing mode */
3377         if ((bargs->flags & BTRFS_BALANCE_ARGS_SOFT) &&
3378             chunk_soft_convert_filter(chunk_type, bargs)) {
3379                 return 0;
3380         }
3381
3382         /*
3383          * limited by count, must be the last filter
3384          */
3385         if ((bargs->flags & BTRFS_BALANCE_ARGS_LIMIT)) {
3386                 if (bargs->limit == 0)
3387                         return 0;
3388                 else
3389                         bargs->limit--;
3390         } else if ((bargs->flags & BTRFS_BALANCE_ARGS_LIMIT_RANGE)) {
3391                 /*
3392                  * Same logic as the 'limit' filter; the minimum cannot be
3393                  * determined here because we do not have the global information
3394                  * about the count of all chunks that satisfy the filters.
3395                  */
3396                 if (bargs->limit_max == 0)
3397                         return 0;
3398                 else
3399                         bargs->limit_max--;
3400         }
3401
3402         return 1;
3403 }
3404
3405 static int __btrfs_balance(struct btrfs_fs_info *fs_info)
3406 {
3407         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3408         struct btrfs_root *chunk_root = fs_info->chunk_root;
3409         struct btrfs_root *dev_root = fs_info->dev_root;
3410         struct list_head *devices;
3411         struct btrfs_device *device;
3412         u64 old_size;
3413         u64 size_to_free;
3414         u64 chunk_type;
3415         struct btrfs_chunk *chunk;
3416         struct btrfs_path *path = NULL;
3417         struct btrfs_key key;
3418         struct btrfs_key found_key;
3419         struct btrfs_trans_handle *trans;
3420         struct extent_buffer *leaf;
3421         int slot;
3422         int ret;
3423         int enospc_errors = 0;
3424         bool counting = true;
3425         /* The single value limit and min/max limits use the same bytes in the */
3426         u64 limit_data = bctl->data.limit;
3427         u64 limit_meta = bctl->meta.limit;
3428         u64 limit_sys = bctl->sys.limit;
3429         u32 count_data = 0;
3430         u32 count_meta = 0;
3431         u32 count_sys = 0;
3432         int chunk_reserved = 0;
3433
3434         /* step one make some room on all the devices */
3435         devices = &fs_info->fs_devices->devices;
3436         list_for_each_entry(device, devices, dev_list) {
3437                 old_size = btrfs_device_get_total_bytes(device);
3438                 size_to_free = div_factor(old_size, 1);
3439                 size_to_free = min_t(u64, size_to_free, SZ_1M);
3440                 if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state) ||
3441                     btrfs_device_get_total_bytes(device) -
3442                     btrfs_device_get_bytes_used(device) > size_to_free ||
3443                     test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state))
3444                         continue;
3445
3446                 ret = btrfs_shrink_device(device, old_size - size_to_free);
3447                 if (ret == -ENOSPC)
3448                         break;
3449                 if (ret) {
3450                         /* btrfs_shrink_device never returns ret > 0 */
3451                         WARN_ON(ret > 0);
3452                         goto error;
3453                 }
3454
3455                 trans = btrfs_start_transaction(dev_root, 0);
3456                 if (IS_ERR(trans)) {
3457                         ret = PTR_ERR(trans);
3458                         btrfs_info_in_rcu(fs_info,
3459                  "resize: unable to start transaction after shrinking device %s (error %d), old size %llu, new size %llu",
3460                                           rcu_str_deref(device->name), ret,
3461                                           old_size, old_size - size_to_free);
3462                         goto error;
3463                 }
3464
3465                 ret = btrfs_grow_device(trans, device, old_size);
3466                 if (ret) {
3467                         btrfs_end_transaction(trans);
3468                         /* btrfs_grow_device never returns ret > 0 */
3469                         WARN_ON(ret > 0);
3470                         btrfs_info_in_rcu(fs_info,
3471                  "resize: unable to grow device after shrinking device %s (error %d), old size %llu, new size %llu",
3472                                           rcu_str_deref(device->name), ret,
3473                                           old_size, old_size - size_to_free);
3474                         goto error;
3475                 }
3476
3477                 btrfs_end_transaction(trans);
3478         }
3479
3480         /* step two, relocate all the chunks */
3481         path = btrfs_alloc_path();
3482         if (!path) {
3483                 ret = -ENOMEM;
3484                 goto error;
3485         }
3486
3487         /* zero out stat counters */
3488         spin_lock(&fs_info->balance_lock);
3489         memset(&bctl->stat, 0, sizeof(bctl->stat));
3490         spin_unlock(&fs_info->balance_lock);
3491 again:
3492         if (!counting) {
3493                 /*
3494                  * The single value limit and min/max limits use the same bytes
3495                  * in the
3496                  */
3497                 bctl->data.limit = limit_data;
3498                 bctl->meta.limit = limit_meta;
3499                 bctl->sys.limit = limit_sys;
3500         }
3501         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
3502         key.offset = (u64)-1;
3503         key.type = BTRFS_CHUNK_ITEM_KEY;
3504
3505         while (1) {
3506                 if ((!counting && atomic_read(&fs_info->balance_pause_req)) ||
3507                     atomic_read(&fs_info->balance_cancel_req)) {
3508                         ret = -ECANCELED;
3509                         goto error;
3510                 }
3511
3512                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
3513                 ret = btrfs_search_slot(NULL, chunk_root, &key, path, 0, 0);
3514                 if (ret < 0) {
3515                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3516                         goto error;
3517                 }
3518
3519                 /*
3520                  * this shouldn't happen, it means the last relocate
3521                  * failed
3522                  */
3523                 if (ret == 0)
3524                         BUG(); /* FIXME break ? */
3525
3526                 ret = btrfs_previous_item(chunk_root, path, 0,
3527                                           BTRFS_CHUNK_ITEM_KEY);
3528                 if (ret) {
3529                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3530                         ret = 0;
3531                         break;
3532                 }
3533
3534                 leaf = path->nodes[0];
3535                 slot = path->slots[0];
3536                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3537
3538                 if (found_key.objectid != key.objectid) {
3539                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3540                         break;
3541                 }
3542
3543                 chunk = btrfs_item_ptr(leaf, slot, struct btrfs_chunk);
3544                 chunk_type = btrfs_chunk_type(leaf, chunk);
3545
3546                 if (!counting) {
3547                         spin_lock(&fs_info->balance_lock);
3548                         bctl->stat.considered++;
3549                         spin_unlock(&fs_info->balance_lock);
3550                 }
3551
3552                 ret = should_balance_chunk(fs_info, leaf, chunk,
3553                                            found_key.offset);
3554
3555                 btrfs_release_path(path);
3556                 if (!ret) {
3557                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3558                         goto loop;
3559                 }
3560
3561                 if (counting) {
3562                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3563                         spin_lock(&fs_info->balance_lock);
3564                         bctl->stat.expected++;
3565                         spin_unlock(&fs_info->balance_lock);
3566
3567                         if (chunk_type & BTRFS_BLOCK_GROUP_DATA)
3568                                 count_data++;
3569                         else if (chunk_type & BTRFS_BLOCK_GROUP_SYSTEM)
3570                                 count_sys++;
3571                         else if (chunk_type & BTRFS_BLOCK_GROUP_METADATA)
3572                                 count_meta++;
3573
3574                         goto loop;
3575                 }
3576
3577                 /*
3578                  * Apply limit_min filter, no need to check if the LIMITS
3579                  * filter is used, limit_min is 0 by default
3580                  */
3581                 if (((chunk_type & BTRFS_BLOCK_GROUP_DATA) &&
3582                                         count_data < bctl->data.limit_min)
3583                                 || ((chunk_type & BTRFS_BLOCK_GROUP_METADATA) &&
3584                                         count_meta < bctl->meta.limit_min)
3585                                 || ((chunk_type & BTRFS_BLOCK_GROUP_SYSTEM) &&
3586                                         count_sys < bctl->sys.limit_min)) {
3587                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3588                         goto loop;
3589                 }
3590
3591                 if (!chunk_reserved) {
3592                         /*
3593                          * We may be relocating the only data chunk we have,
3594                          * which could potentially end up with losing data's
3595                          * raid profile, so lets allocate an empty one in
3596                          * advance.
3597                          */
3598                         ret = btrfs_may_alloc_data_chunk(fs_info,
3599                                                          found_key.offset);
3600                         if (ret < 0) {
3601                                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3602                                 goto error;
3603                         } else if (ret == 1) {
3604                                 chunk_reserved = 1;
3605                         }
3606                 }
3607
3608                 ret = btrfs_relocate_chunk(fs_info, found_key.offset);
3609                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
3610                 if (ret && ret != -ENOSPC)
3611                         goto error;
3612                 if (ret == -ENOSPC) {
3613                         enospc_errors++;
3614                 } else {
3615                         spin_lock(&fs_info->balance_lock);
3616                         bctl->stat.completed++;
3617                         spin_unlock(&fs_info->balance_lock);
3618                 }
3619 loop:
3620                 if (found_key.offset == 0)
3621                         break;
3622                 key.offset = found_key.offset - 1;
3623         }
3624
3625         if (counting) {
3626                 btrfs_release_path(path);
3627                 counting = false;
3628                 goto again;
3629         }
3630 error:
3631         btrfs_free_path(path);
3632         if (enospc_errors) {
3633                 btrfs_info(fs_info, "%d enospc errors during balance",
3634                            enospc_errors);
3635                 if (!ret)
3636                         ret = -ENOSPC;
3637         }
3638
3639         return ret;
3640 }
3641
3642 /**
3643  * alloc_profile_is_valid - see if a given profile is valid and reduced
3644  * @flags: profile to validate
3645  * @extended: if true @flags is treated as an extended profile
3646  */
3647 static int alloc_profile_is_valid(u64 flags, int extended)
3648 {
3649         u64 mask = (extended ? BTRFS_EXTENDED_PROFILE_MASK :
3650                                BTRFS_BLOCK_GROUP_PROFILE_MASK);
3651
3652         flags &= ~BTRFS_BLOCK_GROUP_TYPE_MASK;
3653
3654         /* 1) check that all other bits are zeroed */
3655         if (flags & ~mask)
3656                 return 0;
3657
3658         /* 2) see if profile is reduced */
3659         if (flags == 0)
3660                 return !extended; /* "0" is valid for usual profiles */
3661
3662         /* true if exactly one bit set */
3663         return (flags & (flags - 1)) == 0;
3664 }
3665
3666 static inline int balance_need_close(struct btrfs_fs_info *fs_info)
3667 {
3668         /* cancel requested || normal exit path */
3669         return atomic_read(&fs_info->balance_cancel_req) ||
3670                 (atomic_read(&fs_info->balance_pause_req) == 0 &&
3671                  atomic_read(&fs_info->balance_cancel_req) == 0);
3672 }
3673
3674 /* Non-zero return value signifies invalidity */
3675 static inline int validate_convert_profile(struct btrfs_balance_args *bctl_arg,
3676                 u64 allowed)
3677 {
3678         return ((bctl_arg->flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3679                 (!alloc_profile_is_valid(bctl_arg->target, 1) ||
3680                  (bctl_arg->target & ~allowed)));
3681 }
3682
3683 /*
3684  * Should be called with balance mutexe held
3685  */
3686 int btrfs_balance(struct btrfs_fs_info *fs_info,
3687                   struct btrfs_balance_control *bctl,
3688                   struct btrfs_ioctl_balance_args *bargs)
3689 {
3690         u64 meta_target, data_target;
3691         u64 allowed;
3692         int mixed = 0;
3693         int ret;
3694         u64 num_devices;
3695         unsigned seq;
3696
3697         if (btrfs_fs_closing(fs_info) ||
3698             atomic_read(&fs_info->balance_pause_req) ||
3699             atomic_read(&fs_info->balance_cancel_req)) {
3700                 ret = -EINVAL;
3701                 goto out;
3702         }
3703
3704         allowed = btrfs_super_incompat_flags(fs_info->super_copy);
3705         if (allowed & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
3706                 mixed = 1;
3707
3708         /*
3709          * In case of mixed groups both data and meta should be picked,
3710          * and identical options should be given for both of them.
3711          */
3712         allowed = BTRFS_BALANCE_DATA | BTRFS_BALANCE_METADATA;
3713         if (mixed && (bctl->flags & allowed)) {
3714                 if (!(bctl->flags & BTRFS_BALANCE_DATA) ||
3715                     !(bctl->flags & BTRFS_BALANCE_METADATA) ||
3716                     memcmp(&bctl->data, &bctl->meta, sizeof(bctl->data))) {
3717                         btrfs_err(fs_info,
3718           "balance: mixed groups data and metadata options must be the same");
3719                         ret = -EINVAL;
3720                         goto out;
3721                 }
3722         }
3723
3724         num_devices = fs_info->fs_devices->num_devices;
3725         btrfs_dev_replace_read_lock(&fs_info->dev_replace);
3726         if (btrfs_dev_replace_is_ongoing(&fs_info->dev_replace)) {
3727                 BUG_ON(num_devices < 1);
3728                 num_devices--;
3729         }
3730         btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
3731         allowed = BTRFS_AVAIL_ALLOC_BIT_SINGLE | BTRFS_BLOCK_GROUP_DUP;
3732         if (num_devices > 1)
3733                 allowed |= (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID1);
3734         if (num_devices > 2)
3735                 allowed |= BTRFS_BLOCK_GROUP_RAID5;
3736         if (num_devices > 3)
3737                 allowed |= (BTRFS_BLOCK_GROUP_RAID10 |
3738                             BTRFS_BLOCK_GROUP_RAID6);
3739         if (validate_convert_profile(&bctl->data, allowed)) {
3740                 int index = btrfs_bg_flags_to_raid_index(bctl->data.target);
3741
3742                 btrfs_err(fs_info,
3743                           "balance: invalid convert data profile %s",
3744                           get_raid_name(index));
3745                 ret = -EINVAL;
3746                 goto out;
3747         }
3748         if (validate_convert_profile(&bctl->meta, allowed)) {
3749                 int index = btrfs_bg_flags_to_raid_index(bctl->meta.target);
3750
3751                 btrfs_err(fs_info,
3752                           "balance: invalid convert metadata profile %s",
3753                           get_raid_name(index));
3754                 ret = -EINVAL;
3755                 goto out;
3756         }
3757         if (validate_convert_profile(&bctl->sys, allowed)) {
3758                 int index = btrfs_bg_flags_to_raid_index(bctl->sys.target);
3759
3760                 btrfs_err(fs_info,
3761                           "balance: invalid convert system profile %s",
3762                           get_raid_name(index));
3763                 ret = -EINVAL;
3764                 goto out;
3765         }
3766
3767         /* allow to reduce meta or sys integrity only if force set */
3768         allowed = BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3769                         BTRFS_BLOCK_GROUP_RAID10 |
3770                         BTRFS_BLOCK_GROUP_RAID5 |
3771                         BTRFS_BLOCK_GROUP_RAID6;
3772         do {
3773                 seq = read_seqbegin(&fs_info->profiles_lock);
3774
3775                 if (((bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3776                      (fs_info->avail_system_alloc_bits & allowed) &&
3777                      !(bctl->sys.target & allowed)) ||
3778                     ((bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3779                      (fs_info->avail_metadata_alloc_bits & allowed) &&
3780                      !(bctl->meta.target & allowed))) {
3781                         if (bctl->flags & BTRFS_BALANCE_FORCE) {
3782                                 btrfs_info(fs_info,
3783                                 "balance: force reducing metadata integrity");
3784                         } else {
3785                                 btrfs_err(fs_info,
3786         "balance: reduces metadata integrity, use --force if you want this");
3787                                 ret = -EINVAL;
3788                                 goto out;
3789                         }
3790                 }
3791         } while (read_seqretry(&fs_info->profiles_lock, seq));
3792
3793         /* if we're not converting, the target field is uninitialized */
3794         meta_target = (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) ?
3795                 bctl->meta.target : fs_info->avail_metadata_alloc_bits;
3796         data_target = (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) ?
3797                 bctl->data.target : fs_info->avail_data_alloc_bits;
3798         if (btrfs_get_num_tolerated_disk_barrier_failures(meta_target) <
3799                 btrfs_get_num_tolerated_disk_barrier_failures(data_target)) {
3800                 int meta_index = btrfs_bg_flags_to_raid_index(meta_target);
3801                 int data_index = btrfs_bg_flags_to_raid_index(data_target);
3802
3803                 btrfs_warn(fs_info,
3804         "balance: metadata profile %s has lower redundancy than data profile %s",
3805                            get_raid_name(meta_index), get_raid_name(data_index));
3806         }
3807
3808         ret = insert_balance_item(fs_info, bctl);
3809         if (ret && ret != -EEXIST)
3810                 goto out;
3811
3812         if (!(bctl->flags & BTRFS_BALANCE_RESUME)) {
3813                 BUG_ON(ret == -EEXIST);
3814                 BUG_ON(fs_info->balance_ctl);
3815                 spin_lock(&fs_info->balance_lock);
3816                 fs_info->balance_ctl = bctl;
3817                 spin_unlock(&fs_info->balance_lock);
3818         } else {
3819                 BUG_ON(ret != -EEXIST);
3820                 spin_lock(&fs_info->balance_lock);
3821                 update_balance_args(bctl);
3822                 spin_unlock(&fs_info->balance_lock);
3823         }
3824
3825         ASSERT(!test_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags));
3826         set_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags);
3827         mutex_unlock(&fs_info->balance_mutex);
3828
3829         ret = __btrfs_balance(fs_info);
3830
3831         mutex_lock(&fs_info->balance_mutex);
3832         clear_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags);
3833
3834         if (bargs) {
3835                 memset(bargs, 0, sizeof(*bargs));
3836                 btrfs_update_ioctl_balance_args(fs_info, bargs);
3837         }
3838
3839         if ((ret && ret != -ECANCELED && ret != -ENOSPC) ||
3840             balance_need_close(fs_info)) {
3841                 reset_balance_state(fs_info);
3842                 clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
3843         }
3844
3845         wake_up(&fs_info->balance_wait_q);
3846
3847         return ret;
3848 out:
3849         if (bctl->flags & BTRFS_BALANCE_RESUME)
3850                 reset_balance_state(fs_info);
3851         else
3852                 kfree(bctl);
3853         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
3854
3855         return ret;
3856 }
3857
3858 static int balance_kthread(void *data)
3859 {
3860         struct btrfs_fs_info *fs_info = data;
3861         int ret = 0;
3862
3863         mutex_lock(&fs_info->balance_mutex);
3864         if (fs_info->balance_ctl) {
3865                 btrfs_info(fs_info, "balance: resuming");
3866                 ret = btrfs_balance(fs_info, fs_info->balance_ctl, NULL);
3867         }
3868         mutex_unlock(&fs_info->balance_mutex);
3869
3870         return ret;
3871 }
3872
3873 int btrfs_resume_balance_async(struct btrfs_fs_info *fs_info)
3874 {
3875         struct task_struct *tsk;
3876
3877         mutex_lock(&fs_info->balance_mutex);
3878         if (!fs_info->balance_ctl) {
3879                 mutex_unlock(&fs_info->balance_mutex);
3880                 return 0;
3881         }
3882         mutex_unlock(&fs_info->balance_mutex);
3883
3884         if (btrfs_test_opt(fs_info, SKIP_BALANCE)) {
3885                 btrfs_info(fs_info, "balance: resume skipped");
3886                 return 0;
3887         }
3888
3889         /*
3890          * A ro->rw remount sequence should continue with the paused balance
3891          * regardless of who pauses it, system or the user as of now, so set
3892          * the resume flag.
3893          */
3894         spin_lock(&fs_info->balance_lock);
3895         fs_info->balance_ctl->flags |= BTRFS_BALANCE_RESUME;
3896         spin_unlock(&fs_info->balance_lock);
3897
3898         tsk = kthread_run(balance_kthread, fs_info, "btrfs-balance");
3899         return PTR_ERR_OR_ZERO(tsk);
3900 }
3901
3902 int btrfs_recover_balance(struct btrfs_fs_info *fs_info)
3903 {
3904         struct btrfs_balance_control *bctl;
3905         struct btrfs_balance_item *item;
3906         struct btrfs_disk_balance_args disk_bargs;
3907         struct btrfs_path *path;
3908         struct extent_buffer *leaf;
3909         struct btrfs_key key;
3910         int ret;
3911
3912         path = btrfs_alloc_path();
3913         if (!path)
3914                 return -ENOMEM;
3915
3916         key.objectid = BTRFS_BALANCE_OBJECTID;
3917         key.type = BTRFS_TEMPORARY_ITEM_KEY;
3918         key.offset = 0;
3919
3920         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
3921         if (ret < 0)
3922                 goto out;
3923         if (ret > 0) { /* ret = -ENOENT; */
3924                 ret = 0;
3925                 goto out;
3926         }
3927
3928         bctl = kzalloc(sizeof(*bctl), GFP_NOFS);
3929         if (!bctl) {
3930                 ret = -ENOMEM;
3931                 goto out;
3932         }
3933
3934         leaf = path->nodes[0];
3935         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_balance_item);
3936
3937         bctl->flags = btrfs_balance_flags(leaf, item);
3938         bctl->flags |= BTRFS_BALANCE_RESUME;
3939
3940         btrfs_balance_data(leaf, item, &disk_bargs);
3941         btrfs_disk_balance_args_to_cpu(&bctl->data, &disk_bargs);
3942         btrfs_balance_meta(leaf, item, &disk_bargs);
3943         btrfs_disk_balance_args_to_cpu(&bctl->meta, &disk_bargs);
3944         btrfs_balance_sys(leaf, item, &disk_bargs);
3945         btrfs_disk_balance_args_to_cpu(&bctl->sys, &disk_bargs);
3946
3947         /*
3948          * This should never happen, as the paused balance state is recovered
3949          * during mount without any chance of other exclusive ops to collide.
3950          *
3951          * This gives the exclusive op status to balance and keeps in paused
3952          * state until user intervention (cancel or umount). If the ownership
3953          * cannot be assigned, show a message but do not fail. The balance
3954          * is in a paused state and must have fs_info::balance_ctl properly
3955          * set up.
3956          */
3957         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags))
3958                 btrfs_warn(fs_info,
3959         "balance: cannot set exclusive op status, resume manually");
3960
3961         mutex_lock(&fs_info->balance_mutex);
3962         BUG_ON(fs_info->balance_ctl);
3963         spin_lock(&fs_info->balance_lock);
3964         fs_info->balance_ctl = bctl;
3965         spin_unlock(&fs_info->balance_lock);
3966         mutex_unlock(&fs_info->balance_mutex);
3967 out:
3968         btrfs_free_path(path);
3969         return ret;
3970 }
3971
3972 int btrfs_pause_balance(struct btrfs_fs_info *fs_info)
3973 {
3974         in