btrfs: validate type when reading a chunk
[sfrench/cifs-2.6.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_fs_info *fs_info,
1670                             struct btrfs_device *device)
1671 {
1672         struct btrfs_root *root = fs_info->chunk_root;
1673         int ret;
1674         struct btrfs_path *path;
1675         struct btrfs_dev_item *dev_item;
1676         struct extent_buffer *leaf;
1677         struct btrfs_key key;
1678         unsigned long ptr;
1679
1680         path = btrfs_alloc_path();
1681         if (!path)
1682                 return -ENOMEM;
1683
1684         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1685         key.type = BTRFS_DEV_ITEM_KEY;
1686         key.offset = device->devid;
1687
1688         ret = btrfs_insert_empty_item(trans, root, path, &key,
1689                                       sizeof(*dev_item));
1690         if (ret)
1691                 goto out;
1692
1693         leaf = path->nodes[0];
1694         dev_item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_item);
1695
1696         btrfs_set_device_id(leaf, dev_item, device->devid);
1697         btrfs_set_device_generation(leaf, dev_item, 0);
1698         btrfs_set_device_type(leaf, dev_item, device->type);
1699         btrfs_set_device_io_align(leaf, dev_item, device->io_align);
1700         btrfs_set_device_io_width(leaf, dev_item, device->io_width);
1701         btrfs_set_device_sector_size(leaf, dev_item, device->sector_size);
1702         btrfs_set_device_total_bytes(leaf, dev_item,
1703                                      btrfs_device_get_disk_total_bytes(device));
1704         btrfs_set_device_bytes_used(leaf, dev_item,
1705                                     btrfs_device_get_bytes_used(device));
1706         btrfs_set_device_group(leaf, dev_item, 0);
1707         btrfs_set_device_seek_speed(leaf, dev_item, 0);
1708         btrfs_set_device_bandwidth(leaf, dev_item, 0);
1709         btrfs_set_device_start_offset(leaf, dev_item, 0);
1710
1711         ptr = btrfs_device_uuid(dev_item);
1712         write_extent_buffer(leaf, device->uuid, ptr, BTRFS_UUID_SIZE);
1713         ptr = btrfs_device_fsid(dev_item);
1714         write_extent_buffer(leaf, fs_info->fsid, ptr, BTRFS_FSID_SIZE);
1715         btrfs_mark_buffer_dirty(leaf);
1716
1717         ret = 0;
1718 out:
1719         btrfs_free_path(path);
1720         return ret;
1721 }
1722
1723 /*
1724  * Function to update ctime/mtime for a given device path.
1725  * Mainly used for ctime/mtime based probe like libblkid.
1726  */
1727 static void update_dev_time(const char *path_name)
1728 {
1729         struct file *filp;
1730
1731         filp = filp_open(path_name, O_RDWR, 0);
1732         if (IS_ERR(filp))
1733                 return;
1734         file_update_time(filp);
1735         filp_close(filp, NULL);
1736 }
1737
1738 static int btrfs_rm_dev_item(struct btrfs_fs_info *fs_info,
1739                              struct btrfs_device *device)
1740 {
1741         struct btrfs_root *root = fs_info->chunk_root;
1742         int ret;
1743         struct btrfs_path *path;
1744         struct btrfs_key key;
1745         struct btrfs_trans_handle *trans;
1746
1747         path = btrfs_alloc_path();
1748         if (!path)
1749                 return -ENOMEM;
1750
1751         trans = btrfs_start_transaction(root, 0);
1752         if (IS_ERR(trans)) {
1753                 btrfs_free_path(path);
1754                 return PTR_ERR(trans);
1755         }
1756         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
1757         key.type = BTRFS_DEV_ITEM_KEY;
1758         key.offset = device->devid;
1759
1760         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1761         if (ret) {
1762                 if (ret > 0)
1763                         ret = -ENOENT;
1764                 btrfs_abort_transaction(trans, ret);
1765                 btrfs_end_transaction(trans);
1766                 goto out;
1767         }
1768
1769         ret = btrfs_del_item(trans, root, path);
1770         if (ret) {
1771                 btrfs_abort_transaction(trans, ret);
1772                 btrfs_end_transaction(trans);
1773         }
1774
1775 out:
1776         btrfs_free_path(path);
1777         if (!ret)
1778                 ret = btrfs_commit_transaction(trans);
1779         return ret;
1780 }
1781
1782 /*
1783  * Verify that @num_devices satisfies the RAID profile constraints in the whole
1784  * filesystem. It's up to the caller to adjust that number regarding eg. device
1785  * replace.
1786  */
1787 static int btrfs_check_raid_min_devices(struct btrfs_fs_info *fs_info,
1788                 u64 num_devices)
1789 {
1790         u64 all_avail;
1791         unsigned seq;
1792         int i;
1793
1794         do {
1795                 seq = read_seqbegin(&fs_info->profiles_lock);
1796
1797                 all_avail = fs_info->avail_data_alloc_bits |
1798                             fs_info->avail_system_alloc_bits |
1799                             fs_info->avail_metadata_alloc_bits;
1800         } while (read_seqretry(&fs_info->profiles_lock, seq));
1801
1802         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
1803                 if (!(all_avail & btrfs_raid_array[i].bg_flag))
1804                         continue;
1805
1806                 if (num_devices < btrfs_raid_array[i].devs_min) {
1807                         int ret = btrfs_raid_array[i].mindev_error;
1808
1809                         if (ret)
1810                                 return ret;
1811                 }
1812         }
1813
1814         return 0;
1815 }
1816
1817 static struct btrfs_device * btrfs_find_next_active_device(
1818                 struct btrfs_fs_devices *fs_devs, struct btrfs_device *device)
1819 {
1820         struct btrfs_device *next_device;
1821
1822         list_for_each_entry(next_device, &fs_devs->devices, dev_list) {
1823                 if (next_device != device &&
1824                     !test_bit(BTRFS_DEV_STATE_MISSING, &next_device->dev_state)
1825                     && next_device->bdev)
1826                         return next_device;
1827         }
1828
1829         return NULL;
1830 }
1831
1832 /*
1833  * Helper function to check if the given device is part of s_bdev / latest_bdev
1834  * and replace it with the provided or the next active device, in the context
1835  * where this function called, there should be always be another device (or
1836  * this_dev) which is active.
1837  */
1838 void btrfs_assign_next_active_device(struct btrfs_fs_info *fs_info,
1839                 struct btrfs_device *device, struct btrfs_device *this_dev)
1840 {
1841         struct btrfs_device *next_device;
1842
1843         if (this_dev)
1844                 next_device = this_dev;
1845         else
1846                 next_device = btrfs_find_next_active_device(fs_info->fs_devices,
1847                                                                 device);
1848         ASSERT(next_device);
1849
1850         if (fs_info->sb->s_bdev &&
1851                         (fs_info->sb->s_bdev == device->bdev))
1852                 fs_info->sb->s_bdev = next_device->bdev;
1853
1854         if (fs_info->fs_devices->latest_bdev == device->bdev)
1855                 fs_info->fs_devices->latest_bdev = next_device->bdev;
1856 }
1857
1858 int btrfs_rm_device(struct btrfs_fs_info *fs_info, const char *device_path,
1859                 u64 devid)
1860 {
1861         struct btrfs_device *device;
1862         struct btrfs_fs_devices *cur_devices;
1863         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
1864         u64 num_devices;
1865         int ret = 0;
1866
1867         mutex_lock(&uuid_mutex);
1868
1869         num_devices = fs_devices->num_devices;
1870         btrfs_dev_replace_read_lock(&fs_info->dev_replace);
1871         if (btrfs_dev_replace_is_ongoing(&fs_info->dev_replace)) {
1872                 WARN_ON(num_devices < 1);
1873                 num_devices--;
1874         }
1875         btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
1876
1877         ret = btrfs_check_raid_min_devices(fs_info, num_devices - 1);
1878         if (ret)
1879                 goto out;
1880
1881         ret = btrfs_find_device_by_devspec(fs_info, devid, device_path,
1882                                            &device);
1883         if (ret)
1884                 goto out;
1885
1886         if (test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
1887                 ret = BTRFS_ERROR_DEV_TGT_REPLACE;
1888                 goto out;
1889         }
1890
1891         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state) &&
1892             fs_info->fs_devices->rw_devices == 1) {
1893                 ret = BTRFS_ERROR_DEV_ONLY_WRITABLE;
1894                 goto out;
1895         }
1896
1897         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
1898                 mutex_lock(&fs_info->chunk_mutex);
1899                 list_del_init(&device->dev_alloc_list);
1900                 device->fs_devices->rw_devices--;
1901                 mutex_unlock(&fs_info->chunk_mutex);
1902         }
1903
1904         mutex_unlock(&uuid_mutex);
1905         ret = btrfs_shrink_device(device, 0);
1906         mutex_lock(&uuid_mutex);
1907         if (ret)
1908                 goto error_undo;
1909
1910         /*
1911          * TODO: the superblock still includes this device in its num_devices
1912          * counter although write_all_supers() is not locked out. This
1913          * could give a filesystem state which requires a degraded mount.
1914          */
1915         ret = btrfs_rm_dev_item(fs_info, device);
1916         if (ret)
1917                 goto error_undo;
1918
1919         clear_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
1920         btrfs_scrub_cancel_dev(fs_info, device);
1921
1922         /*
1923          * the device list mutex makes sure that we don't change
1924          * the device list while someone else is writing out all
1925          * the device supers. Whoever is writing all supers, should
1926          * lock the device list mutex before getting the number of
1927          * devices in the super block (super_copy). Conversely,
1928          * whoever updates the number of devices in the super block
1929          * (super_copy) should hold the device list mutex.
1930          */
1931
1932         /*
1933          * In normal cases the cur_devices == fs_devices. But in case
1934          * of deleting a seed device, the cur_devices should point to
1935          * its own fs_devices listed under the fs_devices->seed.
1936          */
1937         cur_devices = device->fs_devices;
1938         mutex_lock(&fs_devices->device_list_mutex);
1939         list_del_rcu(&device->dev_list);
1940
1941         cur_devices->num_devices--;
1942         cur_devices->total_devices--;
1943         /* Update total_devices of the parent fs_devices if it's seed */
1944         if (cur_devices != fs_devices)
1945                 fs_devices->total_devices--;
1946
1947         if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
1948                 cur_devices->missing_devices--;
1949
1950         btrfs_assign_next_active_device(fs_info, device, NULL);
1951
1952         if (device->bdev) {
1953                 cur_devices->open_devices--;
1954                 /* remove sysfs entry */
1955                 btrfs_sysfs_rm_device_link(fs_devices, device);
1956         }
1957
1958         num_devices = btrfs_super_num_devices(fs_info->super_copy) - 1;
1959         btrfs_set_super_num_devices(fs_info->super_copy, num_devices);
1960         mutex_unlock(&fs_devices->device_list_mutex);
1961
1962         /*
1963          * at this point, the device is zero sized and detached from
1964          * the devices list.  All that's left is to zero out the old
1965          * supers and free the device.
1966          */
1967         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
1968                 btrfs_scratch_superblocks(device->bdev, device->name->str);
1969
1970         btrfs_close_bdev(device);
1971         call_rcu(&device->rcu, free_device_rcu);
1972
1973         if (cur_devices->open_devices == 0) {
1974                 while (fs_devices) {
1975                         if (fs_devices->seed == cur_devices) {
1976                                 fs_devices->seed = cur_devices->seed;
1977                                 break;
1978                         }
1979                         fs_devices = fs_devices->seed;
1980                 }
1981                 cur_devices->seed = NULL;
1982                 close_fs_devices(cur_devices);
1983                 free_fs_devices(cur_devices);
1984         }
1985
1986 out:
1987         mutex_unlock(&uuid_mutex);
1988         return ret;
1989
1990 error_undo:
1991         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
1992                 mutex_lock(&fs_info->chunk_mutex);
1993                 list_add(&device->dev_alloc_list,
1994                          &fs_devices->alloc_list);
1995                 device->fs_devices->rw_devices++;
1996                 mutex_unlock(&fs_info->chunk_mutex);
1997         }
1998         goto out;
1999 }
2000
2001 void btrfs_rm_dev_replace_remove_srcdev(struct btrfs_fs_info *fs_info,
2002                                         struct btrfs_device *srcdev)
2003 {
2004         struct btrfs_fs_devices *fs_devices;
2005
2006         lockdep_assert_held(&fs_info->fs_devices->device_list_mutex);
2007
2008         /*
2009          * in case of fs with no seed, srcdev->fs_devices will point
2010          * to fs_devices of fs_info. However when the dev being replaced is
2011          * a seed dev it will point to the seed's local fs_devices. In short
2012          * srcdev will have its correct fs_devices in both the cases.
2013          */
2014         fs_devices = srcdev->fs_devices;
2015
2016         list_del_rcu(&srcdev->dev_list);
2017         list_del(&srcdev->dev_alloc_list);
2018         fs_devices->num_devices--;
2019         if (test_bit(BTRFS_DEV_STATE_MISSING, &srcdev->dev_state))
2020                 fs_devices->missing_devices--;
2021
2022         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &srcdev->dev_state))
2023                 fs_devices->rw_devices--;
2024
2025         if (srcdev->bdev)
2026                 fs_devices->open_devices--;
2027 }
2028
2029 void btrfs_rm_dev_replace_free_srcdev(struct btrfs_fs_info *fs_info,
2030                                       struct btrfs_device *srcdev)
2031 {
2032         struct btrfs_fs_devices *fs_devices = srcdev->fs_devices;
2033
2034         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &srcdev->dev_state)) {
2035                 /* zero out the old super if it is writable */
2036                 btrfs_scratch_superblocks(srcdev->bdev, srcdev->name->str);
2037         }
2038
2039         btrfs_close_bdev(srcdev);
2040         call_rcu(&srcdev->rcu, free_device_rcu);
2041
2042         /* if this is no devs we rather delete the fs_devices */
2043         if (!fs_devices->num_devices) {
2044                 struct btrfs_fs_devices *tmp_fs_devices;
2045
2046                 /*
2047                  * On a mounted FS, num_devices can't be zero unless it's a
2048                  * seed. In case of a seed device being replaced, the replace
2049                  * target added to the sprout FS, so there will be no more
2050                  * device left under the seed FS.
2051                  */
2052                 ASSERT(fs_devices->seeding);
2053
2054                 tmp_fs_devices = fs_info->fs_devices;
2055                 while (tmp_fs_devices) {
2056                         if (tmp_fs_devices->seed == fs_devices) {
2057                                 tmp_fs_devices->seed = fs_devices->seed;
2058                                 break;
2059                         }
2060                         tmp_fs_devices = tmp_fs_devices->seed;
2061                 }
2062                 fs_devices->seed = NULL;
2063                 close_fs_devices(fs_devices);
2064                 free_fs_devices(fs_devices);
2065         }
2066 }
2067
2068 void btrfs_destroy_dev_replace_tgtdev(struct btrfs_fs_info *fs_info,
2069                                       struct btrfs_device *tgtdev)
2070 {
2071         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2072
2073         WARN_ON(!tgtdev);
2074         mutex_lock(&fs_devices->device_list_mutex);
2075
2076         btrfs_sysfs_rm_device_link(fs_devices, tgtdev);
2077
2078         if (tgtdev->bdev)
2079                 fs_devices->open_devices--;
2080
2081         fs_devices->num_devices--;
2082
2083         btrfs_assign_next_active_device(fs_info, tgtdev, NULL);
2084
2085         list_del_rcu(&tgtdev->dev_list);
2086
2087         mutex_unlock(&fs_devices->device_list_mutex);
2088
2089         /*
2090          * The update_dev_time() with in btrfs_scratch_superblocks()
2091          * may lead to a call to btrfs_show_devname() which will try
2092          * to hold device_list_mutex. And here this device
2093          * is already out of device list, so we don't have to hold
2094          * the device_list_mutex lock.
2095          */
2096         btrfs_scratch_superblocks(tgtdev->bdev, tgtdev->name->str);
2097
2098         btrfs_close_bdev(tgtdev);
2099         call_rcu(&tgtdev->rcu, free_device_rcu);
2100 }
2101
2102 static int btrfs_find_device_by_path(struct btrfs_fs_info *fs_info,
2103                                      const char *device_path,
2104                                      struct btrfs_device **device)
2105 {
2106         int ret = 0;
2107         struct btrfs_super_block *disk_super;
2108         u64 devid;
2109         u8 *dev_uuid;
2110         struct block_device *bdev;
2111         struct buffer_head *bh;
2112
2113         *device = NULL;
2114         ret = btrfs_get_bdev_and_sb(device_path, FMODE_READ,
2115                                     fs_info->bdev_holder, 0, &bdev, &bh);
2116         if (ret)
2117                 return ret;
2118         disk_super = (struct btrfs_super_block *)bh->b_data;
2119         devid = btrfs_stack_device_id(&disk_super->dev_item);
2120         dev_uuid = disk_super->dev_item.uuid;
2121         *device = btrfs_find_device(fs_info, devid, dev_uuid, disk_super->fsid);
2122         brelse(bh);
2123         if (!*device)
2124                 ret = -ENOENT;
2125         blkdev_put(bdev, FMODE_READ);
2126         return ret;
2127 }
2128
2129 int btrfs_find_device_missing_or_by_path(struct btrfs_fs_info *fs_info,
2130                                          const char *device_path,
2131                                          struct btrfs_device **device)
2132 {
2133         *device = NULL;
2134         if (strcmp(device_path, "missing") == 0) {
2135                 struct list_head *devices;
2136                 struct btrfs_device *tmp;
2137
2138                 devices = &fs_info->fs_devices->devices;
2139                 list_for_each_entry(tmp, devices, dev_list) {
2140                         if (test_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
2141                                         &tmp->dev_state) && !tmp->bdev) {
2142                                 *device = tmp;
2143                                 break;
2144                         }
2145                 }
2146
2147                 if (!*device)
2148                         return BTRFS_ERROR_DEV_MISSING_NOT_FOUND;
2149
2150                 return 0;
2151         } else {
2152                 return btrfs_find_device_by_path(fs_info, device_path, device);
2153         }
2154 }
2155
2156 /*
2157  * Lookup a device given by device id, or the path if the id is 0.
2158  */
2159 int btrfs_find_device_by_devspec(struct btrfs_fs_info *fs_info, u64 devid,
2160                                  const char *devpath,
2161                                  struct btrfs_device **device)
2162 {
2163         int ret;
2164
2165         if (devid) {
2166                 ret = 0;
2167                 *device = btrfs_find_device(fs_info, devid, NULL, NULL);
2168                 if (!*device)
2169                         ret = -ENOENT;
2170         } else {
2171                 if (!devpath || !devpath[0])
2172                         return -EINVAL;
2173
2174                 ret = btrfs_find_device_missing_or_by_path(fs_info, devpath,
2175                                                            device);
2176         }
2177         return ret;
2178 }
2179
2180 /*
2181  * does all the dirty work required for changing file system's UUID.
2182  */
2183 static int btrfs_prepare_sprout(struct btrfs_fs_info *fs_info)
2184 {
2185         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2186         struct btrfs_fs_devices *old_devices;
2187         struct btrfs_fs_devices *seed_devices;
2188         struct btrfs_super_block *disk_super = fs_info->super_copy;
2189         struct btrfs_device *device;
2190         u64 super_flags;
2191
2192         lockdep_assert_held(&uuid_mutex);
2193         if (!fs_devices->seeding)
2194                 return -EINVAL;
2195
2196         seed_devices = alloc_fs_devices(NULL);
2197         if (IS_ERR(seed_devices))
2198                 return PTR_ERR(seed_devices);
2199
2200         old_devices = clone_fs_devices(fs_devices);
2201         if (IS_ERR(old_devices)) {
2202                 kfree(seed_devices);
2203                 return PTR_ERR(old_devices);
2204         }
2205
2206         list_add(&old_devices->fs_list, &fs_uuids);
2207
2208         memcpy(seed_devices, fs_devices, sizeof(*seed_devices));
2209         seed_devices->opened = 1;
2210         INIT_LIST_HEAD(&seed_devices->devices);
2211         INIT_LIST_HEAD(&seed_devices->alloc_list);
2212         mutex_init(&seed_devices->device_list_mutex);
2213
2214         mutex_lock(&fs_devices->device_list_mutex);
2215         list_splice_init_rcu(&fs_devices->devices, &seed_devices->devices,
2216                               synchronize_rcu);
2217         list_for_each_entry(device, &seed_devices->devices, dev_list)
2218                 device->fs_devices = seed_devices;
2219
2220         mutex_lock(&fs_info->chunk_mutex);
2221         list_splice_init(&fs_devices->alloc_list, &seed_devices->alloc_list);
2222         mutex_unlock(&fs_info->chunk_mutex);
2223
2224         fs_devices->seeding = 0;
2225         fs_devices->num_devices = 0;
2226         fs_devices->open_devices = 0;
2227         fs_devices->missing_devices = 0;
2228         fs_devices->rotating = 0;
2229         fs_devices->seed = seed_devices;
2230
2231         generate_random_uuid(fs_devices->fsid);
2232         memcpy(fs_info->fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
2233         memcpy(disk_super->fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
2234         mutex_unlock(&fs_devices->device_list_mutex);
2235
2236         super_flags = btrfs_super_flags(disk_super) &
2237                       ~BTRFS_SUPER_FLAG_SEEDING;
2238         btrfs_set_super_flags(disk_super, super_flags);
2239
2240         return 0;
2241 }
2242
2243 /*
2244  * Store the expected generation for seed devices in device items.
2245  */
2246 static int btrfs_finish_sprout(struct btrfs_trans_handle *trans,
2247                                struct btrfs_fs_info *fs_info)
2248 {
2249         struct btrfs_root *root = fs_info->chunk_root;
2250         struct btrfs_path *path;
2251         struct extent_buffer *leaf;
2252         struct btrfs_dev_item *dev_item;
2253         struct btrfs_device *device;
2254         struct btrfs_key key;
2255         u8 fs_uuid[BTRFS_FSID_SIZE];
2256         u8 dev_uuid[BTRFS_UUID_SIZE];
2257         u64 devid;
2258         int ret;
2259
2260         path = btrfs_alloc_path();
2261         if (!path)
2262                 return -ENOMEM;
2263
2264         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
2265         key.offset = 0;
2266         key.type = BTRFS_DEV_ITEM_KEY;
2267
2268         while (1) {
2269                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2270                 if (ret < 0)
2271                         goto error;
2272
2273                 leaf = path->nodes[0];
2274 next_slot:
2275                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
2276                         ret = btrfs_next_leaf(root, path);
2277                         if (ret > 0)
2278                                 break;
2279                         if (ret < 0)
2280                                 goto error;
2281                         leaf = path->nodes[0];
2282                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2283                         btrfs_release_path(path);
2284                         continue;
2285                 }
2286
2287                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2288                 if (key.objectid != BTRFS_DEV_ITEMS_OBJECTID ||
2289                     key.type != BTRFS_DEV_ITEM_KEY)
2290                         break;
2291
2292                 dev_item = btrfs_item_ptr(leaf, path->slots[0],
2293                                           struct btrfs_dev_item);
2294                 devid = btrfs_device_id(leaf, dev_item);
2295                 read_extent_buffer(leaf, dev_uuid, btrfs_device_uuid(dev_item),
2296                                    BTRFS_UUID_SIZE);
2297                 read_extent_buffer(leaf, fs_uuid, btrfs_device_fsid(dev_item),
2298                                    BTRFS_FSID_SIZE);
2299                 device = btrfs_find_device(fs_info, devid, dev_uuid, fs_uuid);
2300                 BUG_ON(!device); /* Logic error */
2301
2302                 if (device->fs_devices->seeding) {
2303                         btrfs_set_device_generation(leaf, dev_item,
2304                                                     device->generation);
2305                         btrfs_mark_buffer_dirty(leaf);
2306                 }
2307
2308                 path->slots[0]++;
2309                 goto next_slot;
2310         }
2311         ret = 0;
2312 error:
2313         btrfs_free_path(path);
2314         return ret;
2315 }
2316
2317 int btrfs_init_new_device(struct btrfs_fs_info *fs_info, const char *device_path)
2318 {
2319         struct btrfs_root *root = fs_info->dev_root;
2320         struct request_queue *q;
2321         struct btrfs_trans_handle *trans;
2322         struct btrfs_device *device;
2323         struct block_device *bdev;
2324         struct super_block *sb = fs_info->sb;
2325         struct rcu_string *name;
2326         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2327         u64 tmp;
2328         int seeding_dev = 0;
2329         int ret = 0;
2330         bool unlocked = false;
2331
2332         if (sb_rdonly(sb) && !fs_devices->seeding)
2333                 return -EROFS;
2334
2335         bdev = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
2336                                   fs_info->bdev_holder);
2337         if (IS_ERR(bdev))
2338                 return PTR_ERR(bdev);
2339
2340         if (fs_devices->seeding) {
2341                 seeding_dev = 1;
2342                 down_write(&sb->s_umount);
2343                 mutex_lock(&uuid_mutex);
2344         }
2345
2346         filemap_write_and_wait(bdev->bd_inode->i_mapping);
2347
2348         mutex_lock(&fs_devices->device_list_mutex);
2349         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2350                 if (device->bdev == bdev) {
2351                         ret = -EEXIST;
2352                         mutex_unlock(
2353                                 &fs_devices->device_list_mutex);
2354                         goto error;
2355                 }
2356         }
2357         mutex_unlock(&fs_devices->device_list_mutex);
2358
2359         device = btrfs_alloc_device(fs_info, NULL, NULL);
2360         if (IS_ERR(device)) {
2361                 /* we can safely leave the fs_devices entry around */
2362                 ret = PTR_ERR(device);
2363                 goto error;
2364         }
2365
2366         name = rcu_string_strdup(device_path, GFP_KERNEL);
2367         if (!name) {
2368                 ret = -ENOMEM;
2369                 goto error_free_device;
2370         }
2371         rcu_assign_pointer(device->name, name);
2372
2373         trans = btrfs_start_transaction(root, 0);
2374         if (IS_ERR(trans)) {
2375                 ret = PTR_ERR(trans);
2376                 goto error_free_device;
2377         }
2378
2379         q = bdev_get_queue(bdev);
2380         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
2381         device->generation = trans->transid;
2382         device->io_width = fs_info->sectorsize;
2383         device->io_align = fs_info->sectorsize;
2384         device->sector_size = fs_info->sectorsize;
2385         device->total_bytes = round_down(i_size_read(bdev->bd_inode),
2386                                          fs_info->sectorsize);
2387         device->disk_total_bytes = device->total_bytes;
2388         device->commit_total_bytes = device->total_bytes;
2389         device->fs_info = fs_info;
2390         device->bdev = bdev;
2391         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
2392         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
2393         device->mode = FMODE_EXCL;
2394         device->dev_stats_valid = 1;
2395         set_blocksize(device->bdev, BTRFS_BDEV_BLOCKSIZE);
2396
2397         if (seeding_dev) {
2398                 sb->s_flags &= ~SB_RDONLY;
2399                 ret = btrfs_prepare_sprout(fs_info);
2400                 if (ret) {
2401                         btrfs_abort_transaction(trans, ret);
2402                         goto error_trans;
2403                 }
2404         }
2405
2406         device->fs_devices = fs_devices;
2407
2408         mutex_lock(&fs_devices->device_list_mutex);
2409         mutex_lock(&fs_info->chunk_mutex);
2410         list_add_rcu(&device->dev_list, &fs_devices->devices);
2411         list_add(&device->dev_alloc_list, &fs_devices->alloc_list);
2412         fs_devices->num_devices++;
2413         fs_devices->open_devices++;
2414         fs_devices->rw_devices++;
2415         fs_devices->total_devices++;
2416         fs_devices->total_rw_bytes += device->total_bytes;
2417
2418         atomic64_add(device->total_bytes, &fs_info->free_chunk_space);
2419
2420         if (!blk_queue_nonrot(q))
2421                 fs_devices->rotating = 1;
2422
2423         tmp = btrfs_super_total_bytes(fs_info->super_copy);
2424         btrfs_set_super_total_bytes(fs_info->super_copy,
2425                 round_down(tmp + device->total_bytes, fs_info->sectorsize));
2426
2427         tmp = btrfs_super_num_devices(fs_info->super_copy);
2428         btrfs_set_super_num_devices(fs_info->super_copy, tmp + 1);
2429
2430         /* add sysfs device entry */
2431         btrfs_sysfs_add_device_link(fs_devices, device);
2432
2433         /*
2434          * we've got more storage, clear any full flags on the space
2435          * infos
2436          */
2437         btrfs_clear_space_info_full(fs_info);
2438
2439         mutex_unlock(&fs_info->chunk_mutex);
2440         mutex_unlock(&fs_devices->device_list_mutex);
2441
2442         if (seeding_dev) {
2443                 mutex_lock(&fs_info->chunk_mutex);
2444                 ret = init_first_rw_device(trans, fs_info);
2445                 mutex_unlock(&fs_info->chunk_mutex);
2446                 if (ret) {
2447                         btrfs_abort_transaction(trans, ret);
2448                         goto error_sysfs;
2449                 }
2450         }
2451
2452         ret = btrfs_add_dev_item(trans, fs_info, device);
2453         if (ret) {
2454                 btrfs_abort_transaction(trans, ret);
2455                 goto error_sysfs;
2456         }
2457
2458         if (seeding_dev) {
2459                 char fsid_buf[BTRFS_UUID_UNPARSED_SIZE];
2460
2461                 ret = btrfs_finish_sprout(trans, fs_info);
2462                 if (ret) {
2463                         btrfs_abort_transaction(trans, ret);
2464                         goto error_sysfs;
2465                 }
2466
2467                 /* Sprouting would change fsid of the mounted root,
2468                  * so rename the fsid on the sysfs
2469                  */
2470                 snprintf(fsid_buf, BTRFS_UUID_UNPARSED_SIZE, "%pU",
2471                                                 fs_info->fsid);
2472                 if (kobject_rename(&fs_devices->fsid_kobj, fsid_buf))
2473                         btrfs_warn(fs_info,
2474                                    "sysfs: failed to create fsid for sprout");
2475         }
2476
2477         ret = btrfs_commit_transaction(trans);
2478
2479         if (seeding_dev) {
2480                 mutex_unlock(&uuid_mutex);
2481                 up_write(&sb->s_umount);
2482                 unlocked = true;
2483
2484                 if (ret) /* transaction commit */
2485                         return ret;
2486
2487                 ret = btrfs_relocate_sys_chunks(fs_info);
2488                 if (ret < 0)
2489                         btrfs_handle_fs_error(fs_info, ret,
2490                                     "Failed to relocate sys chunks after device initialization. This can be fixed using the \"btrfs balance\" command.");
2491                 trans = btrfs_attach_transaction(root);
2492                 if (IS_ERR(trans)) {
2493                         if (PTR_ERR(trans) == -ENOENT)
2494                                 return 0;
2495                         ret = PTR_ERR(trans);
2496                         trans = NULL;
2497                         goto error_sysfs;
2498                 }
2499                 ret = btrfs_commit_transaction(trans);
2500         }
2501
2502         /* Update ctime/mtime for libblkid */
2503         update_dev_time(device_path);
2504         return ret;
2505
2506 error_sysfs:
2507         btrfs_sysfs_rm_device_link(fs_devices, device);
2508 error_trans:
2509         if (seeding_dev)
2510                 sb->s_flags |= SB_RDONLY;
2511         if (trans)
2512                 btrfs_end_transaction(trans);
2513 error_free_device:
2514         btrfs_free_device(device);
2515 error:
2516         blkdev_put(bdev, FMODE_EXCL);
2517         if (seeding_dev && !unlocked) {
2518                 mutex_unlock(&uuid_mutex);
2519                 up_write(&sb->s_umount);
2520         }
2521         return ret;
2522 }
2523
2524 static noinline int btrfs_update_device(struct btrfs_trans_handle *trans,
2525                                         struct btrfs_device *device)
2526 {
2527         int ret;
2528         struct btrfs_path *path;
2529         struct btrfs_root *root = device->fs_info->chunk_root;
2530         struct btrfs_dev_item *dev_item;
2531         struct extent_buffer *leaf;
2532         struct btrfs_key key;
2533
2534         path = btrfs_alloc_path();
2535         if (!path)
2536                 return -ENOMEM;
2537
2538         key.objectid = BTRFS_DEV_ITEMS_OBJECTID;
2539         key.type = BTRFS_DEV_ITEM_KEY;
2540         key.offset = device->devid;
2541
2542         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2543         if (ret < 0)
2544                 goto out;
2545
2546         if (ret > 0) {
2547                 ret = -ENOENT;
2548                 goto out;
2549         }
2550
2551         leaf = path->nodes[0];
2552         dev_item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_item);
2553
2554         btrfs_set_device_id(leaf, dev_item, device->devid);
2555         btrfs_set_device_type(leaf, dev_item, device->type);
2556         btrfs_set_device_io_align(leaf, dev_item, device->io_align);
2557         btrfs_set_device_io_width(leaf, dev_item, device->io_width);
2558         btrfs_set_device_sector_size(leaf, dev_item, device->sector_size);
2559         btrfs_set_device_total_bytes(leaf, dev_item,
2560                                      btrfs_device_get_disk_total_bytes(device));
2561         btrfs_set_device_bytes_used(leaf, dev_item,
2562                                     btrfs_device_get_bytes_used(device));
2563         btrfs_mark_buffer_dirty(leaf);
2564
2565 out:
2566         btrfs_free_path(path);
2567         return ret;
2568 }
2569
2570 int btrfs_grow_device(struct btrfs_trans_handle *trans,
2571                       struct btrfs_device *device, u64 new_size)
2572 {
2573         struct btrfs_fs_info *fs_info = device->fs_info;
2574         struct btrfs_super_block *super_copy = fs_info->super_copy;
2575         struct btrfs_fs_devices *fs_devices;
2576         u64 old_total;
2577         u64 diff;
2578
2579         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
2580                 return -EACCES;
2581
2582         new_size = round_down(new_size, fs_info->sectorsize);
2583
2584         mutex_lock(&fs_info->chunk_mutex);
2585         old_total = btrfs_super_total_bytes(super_copy);
2586         diff = round_down(new_size - device->total_bytes, fs_info->sectorsize);
2587
2588         if (new_size <= device->total_bytes ||
2589             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
2590                 mutex_unlock(&fs_info->chunk_mutex);
2591                 return -EINVAL;
2592         }
2593
2594         fs_devices = fs_info->fs_devices;
2595
2596         btrfs_set_super_total_bytes(super_copy,
2597                         round_down(old_total + diff, fs_info->sectorsize));
2598         device->fs_devices->total_rw_bytes += diff;
2599
2600         btrfs_device_set_total_bytes(device, new_size);
2601         btrfs_device_set_disk_total_bytes(device, new_size);
2602         btrfs_clear_space_info_full(device->fs_info);
2603         if (list_empty(&device->resized_list))
2604                 list_add_tail(&device->resized_list,
2605                               &fs_devices->resized_devices);
2606         mutex_unlock(&fs_info->chunk_mutex);
2607
2608         return btrfs_update_device(trans, device);
2609 }
2610
2611 static int btrfs_free_chunk(struct btrfs_trans_handle *trans,
2612                             struct btrfs_fs_info *fs_info, u64 chunk_offset)
2613 {
2614         struct btrfs_root *root = fs_info->chunk_root;
2615         int ret;
2616         struct btrfs_path *path;
2617         struct btrfs_key key;
2618
2619         path = btrfs_alloc_path();
2620         if (!path)
2621                 return -ENOMEM;
2622
2623         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2624         key.offset = chunk_offset;
2625         key.type = BTRFS_CHUNK_ITEM_KEY;
2626
2627         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
2628         if (ret < 0)
2629                 goto out;
2630         else if (ret > 0) { /* Logic error or corruption */
2631                 btrfs_handle_fs_error(fs_info, -ENOENT,
2632                                       "Failed lookup while freeing chunk.");
2633                 ret = -ENOENT;
2634                 goto out;
2635         }
2636
2637         ret = btrfs_del_item(trans, root, path);
2638         if (ret < 0)
2639                 btrfs_handle_fs_error(fs_info, ret,
2640                                       "Failed to delete chunk item.");
2641 out:
2642         btrfs_free_path(path);
2643         return ret;
2644 }
2645
2646 static int btrfs_del_sys_chunk(struct btrfs_fs_info *fs_info, u64 chunk_offset)
2647 {
2648         struct btrfs_super_block *super_copy = fs_info->super_copy;
2649         struct btrfs_disk_key *disk_key;
2650         struct btrfs_chunk *chunk;
2651         u8 *ptr;
2652         int ret = 0;
2653         u32 num_stripes;
2654         u32 array_size;
2655         u32 len = 0;
2656         u32 cur;
2657         struct btrfs_key key;
2658
2659         mutex_lock(&fs_info->chunk_mutex);
2660         array_size = btrfs_super_sys_array_size(super_copy);
2661
2662         ptr = super_copy->sys_chunk_array;
2663         cur = 0;
2664
2665         while (cur < array_size) {
2666                 disk_key = (struct btrfs_disk_key *)ptr;
2667                 btrfs_disk_key_to_cpu(&key, disk_key);
2668
2669                 len = sizeof(*disk_key);
2670
2671                 if (key.type == BTRFS_CHUNK_ITEM_KEY) {
2672                         chunk = (struct btrfs_chunk *)(ptr + len);
2673                         num_stripes = btrfs_stack_chunk_num_stripes(chunk);
2674                         len += btrfs_chunk_item_size(num_stripes);
2675                 } else {
2676                         ret = -EIO;
2677                         break;
2678                 }
2679                 if (key.objectid == BTRFS_FIRST_CHUNK_TREE_OBJECTID &&
2680                     key.offset == chunk_offset) {
2681                         memmove(ptr, ptr + len, array_size - (cur + len));
2682                         array_size -= len;
2683                         btrfs_set_super_sys_array_size(super_copy, array_size);
2684                 } else {
2685                         ptr += len;
2686                         cur += len;
2687                 }
2688         }
2689         mutex_unlock(&fs_info->chunk_mutex);
2690         return ret;
2691 }
2692
2693 static struct extent_map *get_chunk_map(struct btrfs_fs_info *fs_info,
2694                                         u64 logical, u64 length)
2695 {
2696         struct extent_map_tree *em_tree;
2697         struct extent_map *em;
2698
2699         em_tree = &fs_info->mapping_tree.map_tree;
2700         read_lock(&em_tree->lock);
2701         em = lookup_extent_mapping(em_tree, logical, length);
2702         read_unlock(&em_tree->lock);
2703
2704         if (!em) {
2705                 btrfs_crit(fs_info, "unable to find logical %llu length %llu",
2706                            logical, length);
2707                 return ERR_PTR(-EINVAL);
2708         }
2709
2710         if (em->start > logical || em->start + em->len < logical) {
2711                 btrfs_crit(fs_info,
2712                            "found a bad mapping, wanted %llu-%llu, found %llu-%llu",
2713                            logical, length, em->start, em->start + em->len);
2714                 free_extent_map(em);
2715                 return ERR_PTR(-EINVAL);
2716         }
2717
2718         /* callers are responsible for dropping em's ref. */
2719         return em;
2720 }
2721
2722 int btrfs_remove_chunk(struct btrfs_trans_handle *trans,
2723                        struct btrfs_fs_info *fs_info, u64 chunk_offset)
2724 {
2725         struct extent_map *em;
2726         struct map_lookup *map;
2727         u64 dev_extent_len = 0;
2728         int i, ret = 0;
2729         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2730
2731         em = get_chunk_map(fs_info, chunk_offset, 1);
2732         if (IS_ERR(em)) {
2733                 /*
2734                  * This is a logic error, but we don't want to just rely on the
2735                  * user having built with ASSERT enabled, so if ASSERT doesn't
2736                  * do anything we still error out.
2737                  */
2738                 ASSERT(0);
2739                 return PTR_ERR(em);
2740         }
2741         map = em->map_lookup;
2742         mutex_lock(&fs_info->chunk_mutex);
2743         check_system_chunk(trans, map->type);
2744         mutex_unlock(&fs_info->chunk_mutex);
2745
2746         /*
2747          * Take the device list mutex to prevent races with the final phase of
2748          * a device replace operation that replaces the device object associated
2749          * with map stripes (dev-replace.c:btrfs_dev_replace_finishing()).
2750          */
2751         mutex_lock(&fs_devices->device_list_mutex);
2752         for (i = 0; i < map->num_stripes; i++) {
2753                 struct btrfs_device *device = map->stripes[i].dev;
2754                 ret = btrfs_free_dev_extent(trans, device,
2755                                             map->stripes[i].physical,
2756                                             &dev_extent_len);
2757                 if (ret) {
2758                         mutex_unlock(&fs_devices->device_list_mutex);
2759                         btrfs_abort_transaction(trans, ret);
2760                         goto out;
2761                 }
2762
2763                 if (device->bytes_used > 0) {
2764                         mutex_lock(&fs_info->chunk_mutex);
2765                         btrfs_device_set_bytes_used(device,
2766                                         device->bytes_used - dev_extent_len);
2767                         atomic64_add(dev_extent_len, &fs_info->free_chunk_space);
2768                         btrfs_clear_space_info_full(fs_info);
2769                         mutex_unlock(&fs_info->chunk_mutex);
2770                 }
2771
2772                 if (map->stripes[i].dev) {
2773                         ret = btrfs_update_device(trans, map->stripes[i].dev);
2774                         if (ret) {
2775                                 mutex_unlock(&fs_devices->device_list_mutex);
2776                                 btrfs_abort_transaction(trans, ret);
2777                                 goto out;
2778                         }
2779                 }
2780         }
2781         mutex_unlock(&fs_devices->device_list_mutex);
2782
2783         ret = btrfs_free_chunk(trans, fs_info, chunk_offset);
2784         if (ret) {
2785                 btrfs_abort_transaction(trans, ret);
2786                 goto out;
2787         }
2788
2789         trace_btrfs_chunk_free(fs_info, map, chunk_offset, em->len);
2790
2791         if (map->type & BTRFS_BLOCK_GROUP_SYSTEM) {
2792                 ret = btrfs_del_sys_chunk(fs_info, chunk_offset);
2793                 if (ret) {
2794                         btrfs_abort_transaction(trans, ret);
2795                         goto out;
2796                 }
2797         }
2798
2799         ret = btrfs_remove_block_group(trans, chunk_offset, em);
2800         if (ret) {
2801                 btrfs_abort_transaction(trans, ret);
2802                 goto out;
2803         }
2804
2805 out:
2806         /* once for us */
2807         free_extent_map(em);
2808         return ret;
2809 }
2810
2811 static int btrfs_relocate_chunk(struct btrfs_fs_info *fs_info, u64 chunk_offset)
2812 {
2813         struct btrfs_root *root = fs_info->chunk_root;
2814         struct btrfs_trans_handle *trans;
2815         int ret;
2816
2817         /*
2818          * Prevent races with automatic removal of unused block groups.
2819          * After we relocate and before we remove the chunk with offset
2820          * chunk_offset, automatic removal of the block group can kick in,
2821          * resulting in a failure when calling btrfs_remove_chunk() below.
2822          *
2823          * Make sure to acquire this mutex before doing a tree search (dev
2824          * or chunk trees) to find chunks. Otherwise the cleaner kthread might
2825          * call btrfs_remove_chunk() (through btrfs_delete_unused_bgs()) after
2826          * we release the path used to search the chunk/dev tree and before
2827          * the current task acquires this mutex and calls us.
2828          */
2829         lockdep_assert_held(&fs_info->delete_unused_bgs_mutex);
2830
2831         ret = btrfs_can_relocate(fs_info, chunk_offset);
2832         if (ret)
2833                 return -ENOSPC;
2834
2835         /* step one, relocate all the extents inside this chunk */
2836         btrfs_scrub_pause(fs_info);
2837         ret = btrfs_relocate_block_group(fs_info, chunk_offset);
2838         btrfs_scrub_continue(fs_info);
2839         if (ret)
2840                 return ret;
2841
2842         /*
2843          * We add the kobjects here (and after forcing data chunk creation)
2844          * since relocation is the only place we'll create chunks of a new
2845          * type at runtime.  The only place where we'll remove the last
2846          * chunk of a type is the call immediately below this one.  Even
2847          * so, we're protected against races with the cleaner thread since
2848          * we're covered by the delete_unused_bgs_mutex.
2849          */
2850         btrfs_add_raid_kobjects(fs_info);
2851
2852         trans = btrfs_start_trans_remove_block_group(root->fs_info,
2853                                                      chunk_offset);
2854         if (IS_ERR(trans)) {
2855                 ret = PTR_ERR(trans);
2856                 btrfs_handle_fs_error(root->fs_info, ret, NULL);
2857                 return ret;
2858         }
2859
2860         /*
2861          * step two, delete the device extents and the
2862          * chunk tree entries
2863          */
2864         ret = btrfs_remove_chunk(trans, fs_info, chunk_offset);
2865         btrfs_end_transaction(trans);
2866         return ret;
2867 }
2868
2869 static int btrfs_relocate_sys_chunks(struct btrfs_fs_info *fs_info)
2870 {
2871         struct btrfs_root *chunk_root = fs_info->chunk_root;
2872         struct btrfs_path *path;
2873         struct extent_buffer *leaf;
2874         struct btrfs_chunk *chunk;
2875         struct btrfs_key key;
2876         struct btrfs_key found_key;
2877         u64 chunk_type;
2878         bool retried = false;
2879         int failed = 0;
2880         int ret;
2881
2882         path = btrfs_alloc_path();
2883         if (!path)
2884                 return -ENOMEM;
2885
2886 again:
2887         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2888         key.offset = (u64)-1;
2889         key.type = BTRFS_CHUNK_ITEM_KEY;
2890
2891         while (1) {
2892                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
2893                 ret = btrfs_search_slot(NULL, chunk_root, &key, path, 0, 0);
2894                 if (ret < 0) {
2895                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
2896                         goto error;
2897                 }
2898                 BUG_ON(ret == 0); /* Corruption */
2899
2900                 ret = btrfs_previous_item(chunk_root, path, key.objectid,
2901                                           key.type);
2902                 if (ret)
2903                         mutex_unlock(&fs_info->delete_unused_bgs_mutex);
2904                 if (ret < 0)
2905                         goto error;
2906                 if (ret > 0)
2907                         break;
2908
2909                 leaf = path->nodes[0];
2910                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2911
2912                 chunk = btrfs_item_ptr(leaf, path->slots[0],
2913                                        struct btrfs_chunk);
2914                 chunk_type = btrfs_chunk_type(leaf, chunk);
2915                 btrfs_release_path(path);
2916
2917                 if (chunk_type & BTRFS_BLOCK_GROUP_SYSTEM) {
2918                         ret = btrfs_relocate_chunk(fs_info, found_key.offset);
2919                         if (ret == -ENOSPC)
2920                                 failed++;
2921                         else
2922                                 BUG_ON(ret);
2923                 }
2924                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
2925
2926                 if (found_key.offset == 0)
2927                         break;
2928                 key.offset = found_key.offset - 1;
2929         }
2930         ret = 0;
2931         if (failed && !retried) {
2932                 failed = 0;
2933                 retried = true;
2934                 goto again;
2935         } else if (WARN_ON(failed && retried)) {
2936                 ret = -ENOSPC;
2937         }
2938 error:
2939         btrfs_free_path(path);
2940         return ret;
2941 }
2942
2943 /*
2944  * return 1 : allocate a data chunk successfully,
2945  * return <0: errors during allocating a data chunk,
2946  * return 0 : no need to allocate a data chunk.
2947  */
2948 static int btrfs_may_alloc_data_chunk(struct btrfs_fs_info *fs_info,
2949                                       u64 chunk_offset)
2950 {
2951         struct btrfs_block_group_cache *cache;
2952         u64 bytes_used;
2953         u64 chunk_type;
2954
2955         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2956         ASSERT(cache);
2957         chunk_type = cache->flags;
2958         btrfs_put_block_group(cache);
2959
2960         if (chunk_type & BTRFS_BLOCK_GROUP_DATA) {
2961                 spin_lock(&fs_info->data_sinfo->lock);
2962                 bytes_used = fs_info->data_sinfo->bytes_used;
2963                 spin_unlock(&fs_info->data_sinfo->lock);
2964
2965                 if (!bytes_used) {
2966                         struct btrfs_trans_handle *trans;
2967                         int ret;
2968
2969                         trans = btrfs_join_transaction(fs_info->tree_root);
2970                         if (IS_ERR(trans))
2971                                 return PTR_ERR(trans);
2972
2973                         ret = btrfs_force_chunk_alloc(trans,
2974                                                       BTRFS_BLOCK_GROUP_DATA);
2975                         btrfs_end_transaction(trans);
2976                         if (ret < 0)
2977                                 return ret;
2978
2979                         btrfs_add_raid_kobjects(fs_info);
2980
2981                         return 1;
2982                 }
2983         }
2984         return 0;
2985 }
2986
2987 static int insert_balance_item(struct btrfs_fs_info *fs_info,
2988                                struct btrfs_balance_control *bctl)
2989 {
2990         struct btrfs_root *root = fs_info->tree_root;
2991         struct btrfs_trans_handle *trans;
2992         struct btrfs_balance_item *item;
2993         struct btrfs_disk_balance_args disk_bargs;
2994         struct btrfs_path *path;
2995         struct extent_buffer *leaf;
2996         struct btrfs_key key;
2997         int ret, err;
2998
2999         path = btrfs_alloc_path();
3000         if (!path)
3001                 return -ENOMEM;
3002
3003         trans = btrfs_start_transaction(root, 0);
3004         if (IS_ERR(trans)) {
3005                 btrfs_free_path(path);
3006                 return PTR_ERR(trans);
3007         }
3008
3009         key.objectid = BTRFS_BALANCE_OBJECTID;
3010         key.type = BTRFS_TEMPORARY_ITEM_KEY;
3011         key.offset = 0;
3012
3013         ret = btrfs_insert_empty_item(trans, root, path, &key,
3014                                       sizeof(*item));
3015         if (ret)
3016                 goto out;
3017
3018         leaf = path->nodes[0];
3019         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_balance_item);
3020
3021         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3022
3023         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->data);
3024         btrfs_set_balance_data(leaf, item, &disk_bargs);
3025         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->meta);
3026         btrfs_set_balance_meta(leaf, item, &disk_bargs);
3027         btrfs_cpu_balance_args_to_disk(&disk_bargs, &bctl->sys);
3028         btrfs_set_balance_sys(leaf, item, &disk_bargs);
3029
3030         btrfs_set_balance_flags(leaf, item, bctl->flags);
3031
3032         btrfs_mark_buffer_dirty(leaf);
3033 out:
3034         btrfs_free_path(path);
3035         err = btrfs_commit_transaction(trans);
3036         if (err && !ret)
3037                 ret = err;
3038         return ret;
3039 }
3040
3041 static int del_balance_item(struct btrfs_fs_info *fs_info)
3042 {
3043         struct btrfs_root *root = fs_info->tree_root;
3044         struct btrfs_trans_handle *trans;
3045         struct btrfs_path *path;
3046         struct btrfs_key key;
3047         int ret, err;
3048
3049         path = btrfs_alloc_path();
3050         if (!path)
3051                 return -ENOMEM;
3052
3053         trans = btrfs_start_transaction(root, 0);
3054         if (IS_ERR(trans)) {
3055                 btrfs_free_path(path);
3056                 return PTR_ERR(trans);
3057         }
3058
3059         key.objectid = BTRFS_BALANCE_OBJECTID;
3060         key.type = BTRFS_TEMPORARY_ITEM_KEY;
3061         key.offset = 0;
3062
3063         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3064         if (ret < 0)
3065                 goto out;
3066         if (ret > 0) {
3067                 ret = -ENOENT;
3068                 goto out;
3069         }
3070
3071         ret = btrfs_del_item(trans, root, path);
3072 out:
3073         btrfs_free_path(path);
3074         err = btrfs_commit_transaction(trans);
3075         if (err && !ret)
3076                 ret = err;
3077         return ret;
3078 }
3079
3080 /*
3081  * This is a heuristic used to reduce the number of chunks balanced on
3082  * resume after balance was interrupted.
3083  */
3084 static void update_balance_args(struct btrfs_balance_control *bctl)
3085 {
3086         /*
3087          * Turn on soft mode for chunk types that were being converted.
3088          */
3089         if (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT)
3090                 bctl->data.flags |= BTRFS_BALANCE_ARGS_SOFT;
3091         if (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT)
3092                 bctl->sys.flags |= BTRFS_BALANCE_ARGS_SOFT;
3093         if (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT)
3094                 bctl->meta.flags |= BTRFS_BALANCE_ARGS_SOFT;
3095
3096         /*
3097          * Turn on usage filter if is not already used.  The idea is
3098          * that chunks that we have already balanced should be
3099          * reasonably full.  Don't do it for chunks that are being
3100          * converted - that will keep us from relocating unconverted
3101          * (albeit full) chunks.
3102          */
3103         if (!(bctl->data.flags & BTRFS_BALANCE_ARGS_USAGE) &&
3104             !(bctl->data.flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3105             !(bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
3106                 bctl->data.flags |= BTRFS_BALANCE_ARGS_USAGE;
3107                 bctl->data.usage = 90;
3108         }
3109         if (!(bctl->sys.flags & BTRFS_BALANCE_ARGS_USAGE) &&
3110             !(bctl->sys.flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3111             !(bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
3112                 bctl->sys.flags |= BTRFS_BALANCE_ARGS_USAGE;
3113                 bctl->sys.usage = 90;
3114         }
3115         if (!(bctl->meta.flags & BTRFS_BALANCE_ARGS_USAGE) &&
3116             !(bctl->meta.flags & BTRFS_BALANCE_ARGS_USAGE_RANGE) &&
3117             !(bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT)) {
3118                 bctl->meta.flags |= BTRFS_BALANCE_ARGS_USAGE;
3119                 bctl->meta.usage = 90;
3120         }
3121 }
3122
3123 /*
3124  * Clear the balance status in fs_info and delete the balance item from disk.
3125  */
3126 static void reset_balance_state(struct btrfs_fs_info *fs_info)
3127 {
3128         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3129         int ret;
3130
3131         BUG_ON(!fs_info->balance_ctl);
3132
3133         spin_lock(&fs_info->balance_lock);
3134         fs_info->balance_ctl = NULL;
3135         spin_unlock(&fs_info->balance_lock);
3136
3137         kfree(bctl);
3138         ret = del_balance_item(fs_info);
3139         if (ret)
3140                 btrfs_handle_fs_error(fs_info, ret, NULL);
3141 }
3142
3143 /*
3144  * Balance filters.  Return 1 if chunk should be filtered out
3145  * (should not be balanced).
3146  */
3147 static int chunk_profiles_filter(u64 chunk_type,
3148                                  struct btrfs_balance_args *bargs)
3149 {
3150         chunk_type = chunk_to_extended(chunk_type) &
3151                                 BTRFS_EXTENDED_PROFILE_MASK;
3152
3153         if (bargs->profiles & chunk_type)
3154                 return 0;
3155
3156         return 1;
3157 }
3158
3159 static int chunk_usage_range_filter(struct btrfs_fs_info *fs_info, u64 chunk_offset,
3160                               struct btrfs_balance_args *bargs)
3161 {
3162         struct btrfs_block_group_cache *cache;
3163         u64 chunk_used;
3164         u64 user_thresh_min;
3165         u64 user_thresh_max;
3166         int ret = 1;
3167
3168         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3169         chunk_used = btrfs_block_group_used(&cache->item);
3170
3171         if (bargs->usage_min == 0)
3172                 user_thresh_min = 0;
3173         else
3174                 user_thresh_min = div_factor_fine(cache->key.offset,
3175                                         bargs->usage_min);
3176
3177         if (bargs->usage_max == 0)
3178                 user_thresh_max = 1;
3179         else if (bargs->usage_max > 100)
3180                 user_thresh_max = cache->key.offset;
3181         else
3182                 user_thresh_max = div_factor_fine(cache->key.offset,
3183                                         bargs->usage_max);
3184
3185         if (user_thresh_min <= chunk_used && chunk_used < user_thresh_max)
3186                 ret = 0;
3187
3188         btrfs_put_block_group(cache);
3189         return ret;
3190 }
3191
3192 static int chunk_usage_filter(struct btrfs_fs_info *fs_info,
3193                 u64 chunk_offset, struct btrfs_balance_args *bargs)
3194 {
3195         struct btrfs_block_group_cache *cache;
3196         u64 chunk_used, user_thresh;
3197         int ret = 1;
3198
3199         cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3200         chunk_used = btrfs_block_group_used(&cache->item);
3201
3202         if (bargs->usage_min == 0)
3203                 user_thresh = 1;
3204         else if (bargs->usage > 100)
3205                 user_thresh = cache->key.offset;
3206         else
3207                 user_thresh = div_factor_fine(cache->key.offset,
3208                                               bargs->usage);
3209
3210         if (chunk_used < user_thresh)
3211                 ret = 0;
3212
3213         btrfs_put_block_group(cache);
3214         return ret;
3215 }
3216
3217 static int chunk_devid_filter(struct extent_buffer *leaf,
3218                               struct btrfs_chunk *chunk,
3219                               struct btrfs_balance_args *bargs)
3220 {
3221         struct btrfs_stripe *stripe;
3222         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
3223         int i;
3224
3225         for (i = 0; i < num_stripes; i++) {
3226                 stripe = btrfs_stripe_nr(chunk, i);
3227                 if (btrfs_stripe_devid(leaf, stripe) == bargs->devid)
3228                         return 0;
3229         }
3230
3231         return 1;
3232 }
3233
3234 /* [pstart, pend) */
3235 static int chunk_drange_filter(struct extent_buffer *leaf,
3236                                struct btrfs_chunk *chunk,
3237                                struct btrfs_balance_args *bargs)
3238 {
3239         struct btrfs_stripe *stripe;
3240         int num_stripes = btrfs_chunk_num_stripes(leaf, chunk);
3241         u64 stripe_offset;
3242         u64 stripe_length;
3243         int factor;
3244         int i;
3245
3246         if (!(bargs->flags & BTRFS_BALANCE_ARGS_DEVID))
3247                 return 0;
3248
3249         if (btrfs_chunk_type(leaf, c