blob: 5fe1ca8abc70577fe28ee6fd69899d11b618d479 [file] [log] [blame]
Stefan Behrense93c89c2012-11-05 17:33:06 +01001/*
2 * Copyright (C) STRATO AG 2012. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18#include <linux/sched.h>
19#include <linux/bio.h>
20#include <linux/slab.h>
21#include <linux/buffer_head.h>
22#include <linux/blkdev.h>
23#include <linux/random.h>
24#include <linux/iocontext.h>
25#include <linux/capability.h>
26#include <linux/kthread.h>
27#include <linux/math64.h>
28#include <asm/div64.h>
Stefan Behrense93c89c2012-11-05 17:33:06 +010029#include "ctree.h"
30#include "extent_map.h"
31#include "disk-io.h"
32#include "transaction.h"
33#include "print-tree.h"
34#include "volumes.h"
35#include "async-thread.h"
36#include "check-integrity.h"
37#include "rcu-string.h"
38#include "dev-replace.h"
Anand Jain49c6f732014-06-03 11:36:02 +080039#include "sysfs.h"
Stefan Behrense93c89c2012-11-05 17:33:06 +010040
Stefan Behrense93c89c2012-11-05 17:33:06 +010041static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
42 int scrub_ret);
43static void btrfs_dev_replace_update_device_in_mapping_tree(
44 struct btrfs_fs_info *fs_info,
45 struct btrfs_device *srcdev,
46 struct btrfs_device *tgtdev);
Stefan Behrense93c89c2012-11-05 17:33:06 +010047static u64 __btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info);
48static int btrfs_dev_replace_kthread(void *data);
49static int btrfs_dev_replace_continue_on_mount(struct btrfs_fs_info *fs_info);
50
51
52int btrfs_init_dev_replace(struct btrfs_fs_info *fs_info)
53{
54 struct btrfs_key key;
55 struct btrfs_root *dev_root = fs_info->dev_root;
56 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
57 struct extent_buffer *eb;
58 int slot;
59 int ret = 0;
60 struct btrfs_path *path = NULL;
61 int item_size;
62 struct btrfs_dev_replace_item *ptr;
63 u64 src_devid;
64
65 path = btrfs_alloc_path();
66 if (!path) {
67 ret = -ENOMEM;
68 goto out;
69 }
70
71 key.objectid = 0;
72 key.type = BTRFS_DEV_REPLACE_KEY;
73 key.offset = 0;
74 ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
75 if (ret) {
76no_valid_dev_replace_entry_found:
77 ret = 0;
78 dev_replace->replace_state =
79 BTRFS_DEV_REPLACE_ITEM_STATE_NEVER_STARTED;
80 dev_replace->cont_reading_from_srcdev_mode =
81 BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_ALWAYS;
82 dev_replace->replace_state = 0;
83 dev_replace->time_started = 0;
84 dev_replace->time_stopped = 0;
85 atomic64_set(&dev_replace->num_write_errors, 0);
86 atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
87 dev_replace->cursor_left = 0;
88 dev_replace->committed_cursor_left = 0;
89 dev_replace->cursor_left_last_write_of_item = 0;
90 dev_replace->cursor_right = 0;
91 dev_replace->srcdev = NULL;
92 dev_replace->tgtdev = NULL;
93 dev_replace->is_valid = 0;
94 dev_replace->item_needs_writeback = 0;
95 goto out;
96 }
97 slot = path->slots[0];
98 eb = path->nodes[0];
99 item_size = btrfs_item_size_nr(eb, slot);
100 ptr = btrfs_item_ptr(eb, slot, struct btrfs_dev_replace_item);
101
102 if (item_size != sizeof(struct btrfs_dev_replace_item)) {
Frank Holtonefe120a2013-12-20 11:37:06 -0500103 btrfs_warn(fs_info,
104 "dev_replace entry found has unexpected size, ignore entry");
Stefan Behrense93c89c2012-11-05 17:33:06 +0100105 goto no_valid_dev_replace_entry_found;
106 }
107
108 src_devid = btrfs_dev_replace_src_devid(eb, ptr);
109 dev_replace->cont_reading_from_srcdev_mode =
110 btrfs_dev_replace_cont_reading_from_srcdev_mode(eb, ptr);
111 dev_replace->replace_state = btrfs_dev_replace_replace_state(eb, ptr);
112 dev_replace->time_started = btrfs_dev_replace_time_started(eb, ptr);
113 dev_replace->time_stopped =
114 btrfs_dev_replace_time_stopped(eb, ptr);
115 atomic64_set(&dev_replace->num_write_errors,
116 btrfs_dev_replace_num_write_errors(eb, ptr));
117 atomic64_set(&dev_replace->num_uncorrectable_read_errors,
118 btrfs_dev_replace_num_uncorrectable_read_errors(eb, ptr));
119 dev_replace->cursor_left = btrfs_dev_replace_cursor_left(eb, ptr);
120 dev_replace->committed_cursor_left = dev_replace->cursor_left;
121 dev_replace->cursor_left_last_write_of_item = dev_replace->cursor_left;
122 dev_replace->cursor_right = btrfs_dev_replace_cursor_right(eb, ptr);
123 dev_replace->is_valid = 1;
124
125 dev_replace->item_needs_writeback = 0;
126 switch (dev_replace->replace_state) {
127 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
128 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
129 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
130 dev_replace->srcdev = NULL;
131 dev_replace->tgtdev = NULL;
132 break;
133 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
134 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
135 dev_replace->srcdev = btrfs_find_device(fs_info, src_devid,
136 NULL, NULL);
137 dev_replace->tgtdev = btrfs_find_device(fs_info,
138 BTRFS_DEV_REPLACE_DEVID,
139 NULL, NULL);
140 /*
141 * allow 'btrfs dev replace_cancel' if src/tgt device is
142 * missing
143 */
144 if (!dev_replace->srcdev &&
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400145 !btrfs_test_opt(fs_info, DEGRADED)) {
Stefan Behrense93c89c2012-11-05 17:33:06 +0100146 ret = -EIO;
Frank Holtonefe120a2013-12-20 11:37:06 -0500147 btrfs_warn(fs_info,
148 "cannot mount because device replace operation is ongoing and");
149 btrfs_warn(fs_info,
150 "srcdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
151 src_devid);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100152 }
153 if (!dev_replace->tgtdev &&
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400154 !btrfs_test_opt(fs_info, DEGRADED)) {
Stefan Behrense93c89c2012-11-05 17:33:06 +0100155 ret = -EIO;
Frank Holtonefe120a2013-12-20 11:37:06 -0500156 btrfs_warn(fs_info,
157 "cannot mount because device replace operation is ongoing and");
158 btrfs_warn(fs_info,
159 "tgtdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
Geert Uytterhoeven6e71c472013-08-20 13:20:08 +0200160 BTRFS_DEV_REPLACE_DEVID);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100161 }
162 if (dev_replace->tgtdev) {
163 if (dev_replace->srcdev) {
164 dev_replace->tgtdev->total_bytes =
165 dev_replace->srcdev->total_bytes;
166 dev_replace->tgtdev->disk_total_bytes =
167 dev_replace->srcdev->disk_total_bytes;
Miao Xie935e5cc2014-09-03 21:35:33 +0800168 dev_replace->tgtdev->commit_total_bytes =
169 dev_replace->srcdev->commit_total_bytes;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100170 dev_replace->tgtdev->bytes_used =
171 dev_replace->srcdev->bytes_used;
Miao Xiece7213c2014-09-03 21:35:34 +0800172 dev_replace->tgtdev->commit_bytes_used =
173 dev_replace->srcdev->commit_bytes_used;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100174 }
175 dev_replace->tgtdev->is_tgtdev_for_dev_replace = 1;
176 btrfs_init_dev_replace_tgtdev_for_resume(fs_info,
177 dev_replace->tgtdev);
178 }
179 break;
180 }
181
182out:
Tsutomu Itoh527afb42015-08-19 14:55:00 +0900183 btrfs_free_path(path);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100184 return ret;
185}
186
187/*
188 * called from commit_transaction. Writes changed device replace state to
189 * disk.
190 */
191int btrfs_run_dev_replace(struct btrfs_trans_handle *trans,
192 struct btrfs_fs_info *fs_info)
193{
194 int ret;
195 struct btrfs_root *dev_root = fs_info->dev_root;
196 struct btrfs_path *path;
197 struct btrfs_key key;
198 struct extent_buffer *eb;
199 struct btrfs_dev_replace_item *ptr;
200 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
201
Liu Bo73beece2015-07-17 16:49:19 +0800202 btrfs_dev_replace_lock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100203 if (!dev_replace->is_valid ||
204 !dev_replace->item_needs_writeback) {
Liu Bo73beece2015-07-17 16:49:19 +0800205 btrfs_dev_replace_unlock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100206 return 0;
207 }
Liu Bo73beece2015-07-17 16:49:19 +0800208 btrfs_dev_replace_unlock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100209
210 key.objectid = 0;
211 key.type = BTRFS_DEV_REPLACE_KEY;
212 key.offset = 0;
213
214 path = btrfs_alloc_path();
215 if (!path) {
216 ret = -ENOMEM;
217 goto out;
218 }
219 ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
220 if (ret < 0) {
Jeff Mahoney5d163e02016-09-20 10:05:00 -0400221 btrfs_warn(fs_info,
222 "error %d while searching for dev_replace item!",
223 ret);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100224 goto out;
225 }
226
227 if (ret == 0 &&
228 btrfs_item_size_nr(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
229 /*
230 * need to delete old one and insert a new one.
231 * Since no attempt is made to recover any old state, if the
232 * dev_replace state is 'running', the data on the target
233 * drive is lost.
234 * It would be possible to recover the state: just make sure
235 * that the beginning of the item is never changed and always
236 * contains all the essential information. Then read this
237 * minimal set of information and use it as a base for the
238 * new state.
239 */
240 ret = btrfs_del_item(trans, dev_root, path);
241 if (ret != 0) {
Jeff Mahoney5d163e02016-09-20 10:05:00 -0400242 btrfs_warn(fs_info,
243 "delete too small dev_replace item failed %d!",
244 ret);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100245 goto out;
246 }
247 ret = 1;
248 }
249
250 if (ret == 1) {
251 /* need to insert a new item */
252 btrfs_release_path(path);
253 ret = btrfs_insert_empty_item(trans, dev_root, path,
254 &key, sizeof(*ptr));
255 if (ret < 0) {
Jeff Mahoney5d163e02016-09-20 10:05:00 -0400256 btrfs_warn(fs_info,
257 "insert dev_replace item failed %d!", ret);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100258 goto out;
259 }
260 }
261
262 eb = path->nodes[0];
263 ptr = btrfs_item_ptr(eb, path->slots[0],
264 struct btrfs_dev_replace_item);
265
Liu Bo73beece2015-07-17 16:49:19 +0800266 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100267 if (dev_replace->srcdev)
268 btrfs_set_dev_replace_src_devid(eb, ptr,
269 dev_replace->srcdev->devid);
270 else
271 btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
272 btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
273 dev_replace->cont_reading_from_srcdev_mode);
274 btrfs_set_dev_replace_replace_state(eb, ptr,
275 dev_replace->replace_state);
276 btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
277 btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
278 btrfs_set_dev_replace_num_write_errors(eb, ptr,
279 atomic64_read(&dev_replace->num_write_errors));
280 btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
281 atomic64_read(&dev_replace->num_uncorrectable_read_errors));
282 dev_replace->cursor_left_last_write_of_item =
283 dev_replace->cursor_left;
284 btrfs_set_dev_replace_cursor_left(eb, ptr,
285 dev_replace->cursor_left_last_write_of_item);
286 btrfs_set_dev_replace_cursor_right(eb, ptr,
287 dev_replace->cursor_right);
288 dev_replace->item_needs_writeback = 0;
Liu Bo73beece2015-07-17 16:49:19 +0800289 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100290
291 btrfs_mark_buffer_dirty(eb);
292
293out:
294 btrfs_free_path(path);
295
296 return ret;
297}
298
299void btrfs_after_dev_replace_commit(struct btrfs_fs_info *fs_info)
300{
301 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
302
303 dev_replace->committed_cursor_left =
304 dev_replace->cursor_left_last_write_of_item;
305}
306
David Sterbada353f62017-02-14 17:55:53 +0100307int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
308 const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
309 int read_src)
Stefan Behrense93c89c2012-11-05 17:33:06 +0100310{
Jeff Mahoney2ff7e612016-06-22 18:54:24 -0400311 struct btrfs_root *root = fs_info->dev_root;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100312 struct btrfs_trans_handle *trans;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100313 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
314 int ret;
315 struct btrfs_device *tgt_device = NULL;
316 struct btrfs_device *src_device = NULL;
317
Miao Xie935e5cc2014-09-03 21:35:33 +0800318 /* the disk copy procedure reuses the scrub code */
Stefan Behrense93c89c2012-11-05 17:33:06 +0100319 mutex_lock(&fs_info->volume_mutex);
Jeff Mahoney2ff7e612016-06-22 18:54:24 -0400320 ret = btrfs_find_device_by_devspec(fs_info, srcdevid,
Anand Jainb5255452016-03-24 18:48:14 +0800321 srcdev_name, &src_device);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100322 if (ret) {
Miao Xie1c433662014-09-03 21:35:32 +0800323 mutex_unlock(&fs_info->volume_mutex);
324 return ret;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100325 }
326
Jeff Mahoney2ff7e612016-06-22 18:54:24 -0400327 ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
Miao Xie1c433662014-09-03 21:35:32 +0800328 src_device, &tgt_device);
329 mutex_unlock(&fs_info->volume_mutex);
330 if (ret)
331 return ret;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100332
Anand Jain9e271ae2015-08-14 18:33:02 +0800333 /*
334 * Here we commit the transaction to make sure commit_total_bytes
335 * of all the devices are updated.
336 */
337 trans = btrfs_attach_transaction(root);
338 if (!IS_ERR(trans)) {
Jeff Mahoney3a45bb22016-09-09 21:39:03 -0400339 ret = btrfs_commit_transaction(trans);
Anand Jain9e271ae2015-08-14 18:33:02 +0800340 if (ret)
341 return ret;
342 } else if (PTR_ERR(trans) != -ENOENT) {
343 return PTR_ERR(trans);
344 }
345
Liu Bo73beece2015-07-17 16:49:19 +0800346 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100347 switch (dev_replace->replace_state) {
348 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
349 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
350 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
351 break;
352 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
353 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
Anand Jainb5255452016-03-24 18:48:14 +0800354 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100355 goto leave;
356 }
357
Anand Jainb5255452016-03-24 18:48:14 +0800358 dev_replace->cont_reading_from_srcdev_mode = read_src;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100359 WARN_ON(!src_device);
360 dev_replace->srcdev = src_device;
361 WARN_ON(!tgt_device);
362 dev_replace->tgtdev = tgt_device;
363
Anand Jainfc23c242016-03-24 18:48:12 +0800364 btrfs_info_in_rcu(fs_info,
David Sterbaecaeb142015-10-08 09:01:03 +0200365 "dev_replace from %s (devid %llu) to %s started",
Stefan Behrense93c89c2012-11-05 17:33:06 +0100366 src_device->missing ? "<missing disk>" :
367 rcu_str_deref(src_device->name),
368 src_device->devid,
369 rcu_str_deref(tgt_device->name));
370
Stefan Behrense93c89c2012-11-05 17:33:06 +0100371 /*
372 * from now on, the writes to the srcdev are all duplicated to
373 * go to the tgtdev as well (refer to btrfs_map_block()).
374 */
375 dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
Zach Brown4546bca2013-10-16 12:10:32 -0700376 dev_replace->time_started = get_seconds();
Stefan Behrense93c89c2012-11-05 17:33:06 +0100377 dev_replace->cursor_left = 0;
378 dev_replace->committed_cursor_left = 0;
379 dev_replace->cursor_left_last_write_of_item = 0;
380 dev_replace->cursor_right = 0;
381 dev_replace->is_valid = 1;
382 dev_replace->item_needs_writeback = 1;
Yauhen Kharuzhy7ccefb92016-03-29 14:17:48 -0700383 atomic64_set(&dev_replace->num_write_errors, 0);
384 atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
Liu Bo73beece2015-07-17 16:49:19 +0800385 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100386
Liu Bo73416da2015-08-14 18:33:07 +0800387 ret = btrfs_sysfs_add_device_link(tgt_device->fs_devices, tgt_device);
388 if (ret)
Jeff Mahoneyab8d0fc2016-09-20 10:05:02 -0400389 btrfs_err(fs_info, "kobj add dev failed %d", ret);
Liu Bo73416da2015-08-14 18:33:07 +0800390
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400391 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100392
393 /* force writing the updated state information to disk */
394 trans = btrfs_start_transaction(root, 0);
395 if (IS_ERR(trans)) {
396 ret = PTR_ERR(trans);
Liu Bo73beece2015-07-17 16:49:19 +0800397 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100398 goto leave;
399 }
400
Jeff Mahoney3a45bb22016-09-09 21:39:03 -0400401 ret = btrfs_commit_transaction(trans);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100402 WARN_ON(ret);
403
404 /* the disk copy procedure reuses the scrub code */
405 ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
Miao Xie7cc8e582014-09-03 21:35:38 +0800406 btrfs_device_get_total_bytes(src_device),
Stefan Behrense93c89c2012-11-05 17:33:06 +0100407 &dev_replace->scrub_progress, 0, 1);
408
Anand Jainfc23c242016-03-24 18:48:12 +0800409 ret = btrfs_dev_replace_finishing(fs_info, ret);
Eryu Guan2fc9f6b2014-10-13 12:42:12 +0800410 if (ret == -EINPROGRESS) {
Anand Jainb5255452016-03-24 18:48:14 +0800411 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
Eryu Guan2fc9f6b2014-10-13 12:42:12 +0800412 } else {
413 WARN_ON(ret);
414 }
Stefan Behrense93c89c2012-11-05 17:33:06 +0100415
Eryu Guan2fc9f6b2014-10-13 12:42:12 +0800416 return ret;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100417
418leave:
419 dev_replace->srcdev = NULL;
420 dev_replace->tgtdev = NULL;
Liu Bo73beece2015-07-17 16:49:19 +0800421 btrfs_dev_replace_unlock(dev_replace, 1);
Miao Xie1c433662014-09-03 21:35:32 +0800422 btrfs_destroy_dev_replace_tgtdev(fs_info, tgt_device);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100423 return ret;
424}
425
Jeff Mahoney2ff7e612016-06-22 18:54:24 -0400426int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
Anand Jainb5255452016-03-24 18:48:14 +0800427 struct btrfs_ioctl_dev_replace_args *args)
428{
429 int ret;
430
431 switch (args->start.cont_reading_from_srcdev_mode) {
432 case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
433 case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
434 break;
435 default:
436 return -EINVAL;
437 }
438
439 if ((args->start.srcdevid == 0 && args->start.srcdev_name[0] == '\0') ||
440 args->start.tgtdev_name[0] == '\0')
441 return -EINVAL;
442
Jeff Mahoney2ff7e612016-06-22 18:54:24 -0400443 ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
Anand Jainb5255452016-03-24 18:48:14 +0800444 args->start.srcdevid,
445 args->start.srcdev_name,
446 args->start.cont_reading_from_srcdev_mode);
447 args->result = ret;
448 /* don't warn if EINPROGRESS, someone else might be running scrub */
449 if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS)
450 ret = 0;
451
452 return ret;
453}
454
Miao Xiec404e0d2014-01-30 16:46:55 +0800455/*
Nicholas D Steeves01327612016-05-19 21:18:45 -0400456 * blocked until all in-flight bios operations are finished.
Miao Xiec404e0d2014-01-30 16:46:55 +0800457 */
458static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
459{
Miao Xiec404e0d2014-01-30 16:46:55 +0800460 set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
Zhao Lei76539472015-01-20 15:11:38 +0800461 wait_event(fs_info->replace_wait, !percpu_counter_sum(
462 &fs_info->bio_counter));
Miao Xiec404e0d2014-01-30 16:46:55 +0800463}
464
465/*
466 * we have removed target device, it is safe to allow new bios request.
467 */
468static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
469{
470 clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
David Sterba730d9ec2015-10-10 18:16:38 +0200471 wake_up(&fs_info->replace_wait);
Miao Xiec404e0d2014-01-30 16:46:55 +0800472}
473
Stefan Behrense93c89c2012-11-05 17:33:06 +0100474static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
475 int scrub_ret)
476{
477 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
478 struct btrfs_device *tgt_device;
479 struct btrfs_device *src_device;
480 struct btrfs_root *root = fs_info->tree_root;
481 u8 uuid_tmp[BTRFS_UUID_SIZE];
482 struct btrfs_trans_handle *trans;
483 int ret = 0;
484
485 /* don't allow cancel or unmount to disturb the finishing procedure */
486 mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
487
Liu Bo73beece2015-07-17 16:49:19 +0800488 btrfs_dev_replace_lock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100489 /* was the operation canceled, or is it finished? */
490 if (dev_replace->replace_state !=
491 BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
Liu Bo73beece2015-07-17 16:49:19 +0800492 btrfs_dev_replace_unlock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100493 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
494 return 0;
495 }
496
497 tgt_device = dev_replace->tgtdev;
498 src_device = dev_replace->srcdev;
Liu Bo73beece2015-07-17 16:49:19 +0800499 btrfs_dev_replace_unlock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100500
Stefan Behrense93c89c2012-11-05 17:33:06 +0100501 /*
502 * flush all outstanding I/O and inode extent mappings before the
503 * copy operation is declared as being finished
504 */
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400505 ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
Miao Xie3edb2a62013-01-22 10:49:33 +0000506 if (ret) {
507 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
508 return ret;
509 }
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400510 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100511
512 trans = btrfs_start_transaction(root, 0);
513 if (IS_ERR(trans)) {
514 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
515 return PTR_ERR(trans);
516 }
Jeff Mahoney3a45bb22016-09-09 21:39:03 -0400517 ret = btrfs_commit_transaction(trans);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100518 WARN_ON(ret);
519
Miao Xie67a2c452014-09-03 21:35:43 +0800520 mutex_lock(&uuid_mutex);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100521 /* keep away write_all_supers() during the finishing procedure */
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400522 mutex_lock(&fs_info->fs_devices->device_list_mutex);
523 mutex_lock(&fs_info->chunk_mutex);
Liu Bo73beece2015-07-17 16:49:19 +0800524 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100525 dev_replace->replace_state =
526 scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
527 : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
528 dev_replace->tgtdev = NULL;
529 dev_replace->srcdev = NULL;
Zach Brown4546bca2013-10-16 12:10:32 -0700530 dev_replace->time_stopped = get_seconds();
Stefan Behrense93c89c2012-11-05 17:33:06 +0100531 dev_replace->item_needs_writeback = 1;
532
Miao Xiec404e0d2014-01-30 16:46:55 +0800533 /* replace old device with new one in mapping tree */
534 if (!scrub_ret) {
535 btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
536 src_device,
537 tgt_device);
538 } else {
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400539 btrfs_err_in_rcu(fs_info,
540 "btrfs_scrub_dev(%s, %llu, %s) failed %d",
541 src_device->missing ? "<missing disk>" :
542 rcu_str_deref(src_device->name),
543 src_device->devid,
544 rcu_str_deref(tgt_device->name), scrub_ret);
Liu Bo73beece2015-07-17 16:49:19 +0800545 btrfs_dev_replace_unlock(dev_replace, 1);
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400546 mutex_unlock(&fs_info->chunk_mutex);
547 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
Miao Xie67a2c452014-09-03 21:35:43 +0800548 mutex_unlock(&uuid_mutex);
Qu Wenruoae6529c2017-03-29 09:33:21 +0800549 btrfs_rm_dev_replace_blocked(fs_info);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100550 if (tgt_device)
551 btrfs_destroy_dev_replace_tgtdev(fs_info, tgt_device);
Qu Wenruoae6529c2017-03-29 09:33:21 +0800552 btrfs_rm_dev_replace_unblocked(fs_info);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100553 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
554
Eryu Guan2fc9f6b2014-10-13 12:42:12 +0800555 return scrub_ret;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100556 }
557
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400558 btrfs_info_in_rcu(fs_info,
559 "dev_replace from %s (devid %llu) to %s finished",
560 src_device->missing ? "<missing disk>" :
561 rcu_str_deref(src_device->name),
562 src_device->devid,
563 rcu_str_deref(tgt_device->name));
Stefan Behrense93c89c2012-11-05 17:33:06 +0100564 tgt_device->is_tgtdev_for_dev_replace = 0;
565 tgt_device->devid = src_device->devid;
566 src_device->devid = BTRFS_DEV_REPLACE_DEVID;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100567 memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
568 memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
569 memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
Miao Xie7cc8e582014-09-03 21:35:38 +0800570 btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
571 btrfs_device_set_disk_total_bytes(tgt_device,
572 src_device->disk_total_bytes);
573 btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
Miao Xie935e5cc2014-09-03 21:35:33 +0800574 ASSERT(list_empty(&src_device->resized_list));
575 tgt_device->commit_total_bytes = src_device->commit_total_bytes;
Miao Xiece7213c2014-09-03 21:35:34 +0800576 tgt_device->commit_bytes_used = src_device->bytes_used;
Anand Jain88acff62016-05-03 17:44:43 +0800577
578 btrfs_assign_next_active_device(fs_info, src_device, tgt_device);
579
Stefan Behrense93c89c2012-11-05 17:33:06 +0100580 list_add(&tgt_device->dev_alloc_list, &fs_info->fs_devices->alloc_list);
Miao Xie82372bc2014-09-03 21:35:44 +0800581 fs_info->fs_devices->rw_devices++;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100582
Liu Bo73beece2015-07-17 16:49:19 +0800583 btrfs_dev_replace_unlock(dev_replace, 1);
Qu Wenruo12b894c2014-08-20 16:10:15 +0800584
Miao Xiec404e0d2014-01-30 16:46:55 +0800585 btrfs_rm_dev_replace_blocked(fs_info);
586
Qu Wenruo084b6e7c2014-10-30 16:52:31 +0800587 btrfs_rm_dev_replace_remove_srcdev(fs_info, src_device);
Ilya Dryomov13572722013-10-02 20:41:01 +0300588
Miao Xiec404e0d2014-01-30 16:46:55 +0800589 btrfs_rm_dev_replace_unblocked(fs_info);
590
Stefan Behrense93c89c2012-11-05 17:33:06 +0100591 /*
592 * this is again a consistent state where no dev_replace procedure
593 * is running, the target device is part of the filesystem, the
594 * source device is not part of the filesystem anymore and its 1st
595 * superblock is scratched out so that it is no longer marked to
596 * belong to this filesystem.
597 */
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400598 mutex_unlock(&fs_info->chunk_mutex);
599 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
Miao Xie67a2c452014-09-03 21:35:43 +0800600 mutex_unlock(&uuid_mutex);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100601
Qu Wenruo084b6e7c2014-10-30 16:52:31 +0800602 /* replace the sysfs entry */
Anand Jain32576042015-08-14 18:32:49 +0800603 btrfs_sysfs_rm_device_link(fs_info->fs_devices, src_device);
Qu Wenruo084b6e7c2014-10-30 16:52:31 +0800604 btrfs_rm_dev_replace_free_srcdev(fs_info, src_device);
605
Stefan Behrense93c89c2012-11-05 17:33:06 +0100606 /* write back the superblocks */
607 trans = btrfs_start_transaction(root, 0);
608 if (!IS_ERR(trans))
Jeff Mahoney3a45bb22016-09-09 21:39:03 -0400609 btrfs_commit_transaction(trans);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100610
611 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
612
613 return 0;
614}
615
616static void btrfs_dev_replace_update_device_in_mapping_tree(
617 struct btrfs_fs_info *fs_info,
618 struct btrfs_device *srcdev,
619 struct btrfs_device *tgtdev)
620{
621 struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
622 struct extent_map *em;
623 struct map_lookup *map;
624 u64 start = 0;
625 int i;
626
627 write_lock(&em_tree->lock);
628 do {
629 em = lookup_extent_mapping(em_tree, start, (u64)-1);
630 if (!em)
631 break;
Jeff Mahoney95617d62015-06-03 10:55:48 -0400632 map = em->map_lookup;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100633 for (i = 0; i < map->num_stripes; i++)
634 if (srcdev == map->stripes[i].dev)
635 map->stripes[i].dev = tgtdev;
636 start = em->start + em->len;
637 free_extent_map(em);
638 } while (start);
639 write_unlock(&em_tree->lock);
640}
641
Stefan Behrense93c89c2012-11-05 17:33:06 +0100642void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
643 struct btrfs_ioctl_dev_replace_args *args)
644{
645 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
Miao Xie7cc8e582014-09-03 21:35:38 +0800646 struct btrfs_device *srcdev;
Stefan Behrense93c89c2012-11-05 17:33:06 +0100647
Liu Bo73beece2015-07-17 16:49:19 +0800648 btrfs_dev_replace_lock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100649 /* even if !dev_replace_is_valid, the values are good enough for
650 * the replace_status ioctl */
651 args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
652 args->status.replace_state = dev_replace->replace_state;
653 args->status.time_started = dev_replace->time_started;
654 args->status.time_stopped = dev_replace->time_stopped;
655 args->status.num_write_errors =
656 atomic64_read(&dev_replace->num_write_errors);
657 args->status.num_uncorrectable_read_errors =
658 atomic64_read(&dev_replace->num_uncorrectable_read_errors);
659 switch (dev_replace->replace_state) {
660 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
661 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
662 args->status.progress_1000 = 0;
663 break;
664 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
665 args->status.progress_1000 = 1000;
666 break;
667 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
668 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
Miao Xie7cc8e582014-09-03 21:35:38 +0800669 srcdev = dev_replace->srcdev;
Liu Bo42c61ab2017-04-03 13:45:24 -0700670 args->status.progress_1000 = div64_u64(dev_replace->cursor_left,
David Sterbaf8c269d2015-01-16 17:21:12 +0100671 div_u64(btrfs_device_get_total_bytes(srcdev), 1000));
Stefan Behrense93c89c2012-11-05 17:33:06 +0100672 break;
673 }
Liu Bo73beece2015-07-17 16:49:19 +0800674 btrfs_dev_replace_unlock(dev_replace, 0);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100675}
676
677int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info,
678 struct btrfs_ioctl_dev_replace_args *args)
679{
680 args->result = __btrfs_dev_replace_cancel(fs_info);
681 return 0;
682}
683
684static u64 __btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
685{
686 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
687 struct btrfs_device *tgt_device = NULL;
688 struct btrfs_trans_handle *trans;
689 struct btrfs_root *root = fs_info->tree_root;
690 u64 result;
691 int ret;
692
Ilya Dryomove649e582013-10-10 20:40:21 +0300693 if (fs_info->sb->s_flags & MS_RDONLY)
694 return -EROFS;
695
Stefan Behrense93c89c2012-11-05 17:33:06 +0100696 mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
Liu Bo73beece2015-07-17 16:49:19 +0800697 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100698 switch (dev_replace->replace_state) {
699 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
700 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
701 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
702 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
Liu Bo73beece2015-07-17 16:49:19 +0800703 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100704 goto leave;
705 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
706 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
707 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
708 tgt_device = dev_replace->tgtdev;
709 dev_replace->tgtdev = NULL;
710 dev_replace->srcdev = NULL;
711 break;
712 }
713 dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
Zach Brown4546bca2013-10-16 12:10:32 -0700714 dev_replace->time_stopped = get_seconds();
Stefan Behrense93c89c2012-11-05 17:33:06 +0100715 dev_replace->item_needs_writeback = 1;
Liu Bo73beece2015-07-17 16:49:19 +0800716 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100717 btrfs_scrub_cancel(fs_info);
718
719 trans = btrfs_start_transaction(root, 0);
720 if (IS_ERR(trans)) {
721 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
722 return PTR_ERR(trans);
723 }
Jeff Mahoney3a45bb22016-09-09 21:39:03 -0400724 ret = btrfs_commit_transaction(trans);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100725 WARN_ON(ret);
726 if (tgt_device)
727 btrfs_destroy_dev_replace_tgtdev(fs_info, tgt_device);
728
729leave:
730 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
731 return result;
732}
733
734void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
735{
736 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
737
738 mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
Liu Bo73beece2015-07-17 16:49:19 +0800739 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100740 switch (dev_replace->replace_state) {
741 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
742 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
743 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
744 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
745 break;
746 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
747 dev_replace->replace_state =
748 BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
Zach Brown4546bca2013-10-16 12:10:32 -0700749 dev_replace->time_stopped = get_seconds();
Stefan Behrense93c89c2012-11-05 17:33:06 +0100750 dev_replace->item_needs_writeback = 1;
Frank Holtonefe120a2013-12-20 11:37:06 -0500751 btrfs_info(fs_info, "suspending dev_replace for unmount");
Stefan Behrense93c89c2012-11-05 17:33:06 +0100752 break;
753 }
754
Liu Bo73beece2015-07-17 16:49:19 +0800755 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100756 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
757}
758
759/* resume dev_replace procedure that was interrupted by unmount */
760int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
761{
762 struct task_struct *task;
763 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
764
Liu Bo73beece2015-07-17 16:49:19 +0800765 btrfs_dev_replace_lock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100766 switch (dev_replace->replace_state) {
767 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
768 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
769 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
Liu Bo73beece2015-07-17 16:49:19 +0800770 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100771 return 0;
772 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
773 break;
774 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
775 dev_replace->replace_state =
776 BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
777 break;
778 }
779 if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
Frank Holtonefe120a2013-12-20 11:37:06 -0500780 btrfs_info(fs_info,
Jeff Mahoney5d163e02016-09-20 10:05:00 -0400781 "cannot continue dev_replace, tgtdev is missing");
782 btrfs_info(fs_info,
783 "you may cancel the operation after 'mount -o degraded'");
Liu Bo73beece2015-07-17 16:49:19 +0800784 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100785 return 0;
786 }
Liu Bo73beece2015-07-17 16:49:19 +0800787 btrfs_dev_replace_unlock(dev_replace, 1);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100788
David Sterba171938e2017-03-28 14:44:21 +0200789 WARN_ON(test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags));
Stefan Behrense93c89c2012-11-05 17:33:06 +0100790 task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
Rusty Russell8c6ffba2013-07-15 11:20:32 +0930791 return PTR_ERR_OR_ZERO(task);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100792}
793
794static int btrfs_dev_replace_kthread(void *data)
795{
796 struct btrfs_fs_info *fs_info = data;
797 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
798 struct btrfs_ioctl_dev_replace_args *status_args;
799 u64 progress;
800
David Sterba58c4e172016-02-11 10:49:42 +0100801 status_args = kzalloc(sizeof(*status_args), GFP_KERNEL);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100802 if (status_args) {
803 btrfs_dev_replace_status(fs_info, status_args);
804 progress = status_args->status.progress_1000;
805 kfree(status_args);
David Sterbaf8c269d2015-01-16 17:21:12 +0100806 progress = div_u64(progress, 10);
David Sterbaecaeb142015-10-08 09:01:03 +0200807 btrfs_info_in_rcu(fs_info,
808 "continuing dev_replace from %s (devid %llu) to %s @%u%%",
Frank Holtonefe120a2013-12-20 11:37:06 -0500809 dev_replace->srcdev->missing ? "<missing disk>" :
810 rcu_str_deref(dev_replace->srcdev->name),
811 dev_replace->srcdev->devid,
812 dev_replace->tgtdev ?
813 rcu_str_deref(dev_replace->tgtdev->name) :
814 "<missing target disk>",
815 (unsigned int)progress);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100816 }
817 btrfs_dev_replace_continue_on_mount(fs_info);
David Sterba171938e2017-03-28 14:44:21 +0200818 clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100819
820 return 0;
821}
822
823static int btrfs_dev_replace_continue_on_mount(struct btrfs_fs_info *fs_info)
824{
825 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
826 int ret;
827
828 ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
829 dev_replace->committed_cursor_left,
Miao Xie7cc8e582014-09-03 21:35:38 +0800830 btrfs_device_get_total_bytes(dev_replace->srcdev),
Stefan Behrense93c89c2012-11-05 17:33:06 +0100831 &dev_replace->scrub_progress, 0, 1);
832 ret = btrfs_dev_replace_finishing(fs_info, ret);
833 WARN_ON(ret);
834 return 0;
835}
836
837int btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
838{
839 if (!dev_replace->is_valid)
840 return 0;
841
842 switch (dev_replace->replace_state) {
843 case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
844 case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
845 case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
846 return 0;
847 case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
848 case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
849 /*
850 * return true even if tgtdev is missing (this is
851 * something that can happen if the dev_replace
852 * procedure is suspended by an umount and then
853 * the tgtdev is missing (or "btrfs dev scan") was
854 * not called and the the filesystem is remounted
855 * in degraded state. This does not stop the
856 * dev_replace procedure. It needs to be canceled
Adam Buchbinderbb7ab3b2016-03-04 11:23:12 -0800857 * manually if the cancellation is wanted.
Stefan Behrense93c89c2012-11-05 17:33:06 +0100858 */
859 break;
860 }
861 return 1;
862}
863
Liu Bo73beece2015-07-17 16:49:19 +0800864void btrfs_dev_replace_lock(struct btrfs_dev_replace *dev_replace, int rw)
Stefan Behrense93c89c2012-11-05 17:33:06 +0100865{
Liu Bo73beece2015-07-17 16:49:19 +0800866 if (rw == 1) {
867 /* write */
868again:
869 wait_event(dev_replace->read_lock_wq,
870 atomic_read(&dev_replace->blocking_readers) == 0);
871 write_lock(&dev_replace->lock);
872 if (atomic_read(&dev_replace->blocking_readers)) {
873 write_unlock(&dev_replace->lock);
874 goto again;
875 }
876 } else {
877 read_lock(&dev_replace->lock);
878 atomic_inc(&dev_replace->read_locks);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100879 }
Stefan Behrense93c89c2012-11-05 17:33:06 +0100880}
881
Liu Bo73beece2015-07-17 16:49:19 +0800882void btrfs_dev_replace_unlock(struct btrfs_dev_replace *dev_replace, int rw)
Stefan Behrense93c89c2012-11-05 17:33:06 +0100883{
Liu Bo73beece2015-07-17 16:49:19 +0800884 if (rw == 1) {
885 /* write */
886 ASSERT(atomic_read(&dev_replace->blocking_readers) == 0);
887 write_unlock(&dev_replace->lock);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100888 } else {
Liu Bo73beece2015-07-17 16:49:19 +0800889 ASSERT(atomic_read(&dev_replace->read_locks) > 0);
890 atomic_dec(&dev_replace->read_locks);
891 read_unlock(&dev_replace->lock);
Stefan Behrense93c89c2012-11-05 17:33:06 +0100892 }
893}
Miao Xiec404e0d2014-01-30 16:46:55 +0800894
Liu Bo73beece2015-07-17 16:49:19 +0800895/* inc blocking cnt and release read lock */
896void btrfs_dev_replace_set_lock_blocking(
897 struct btrfs_dev_replace *dev_replace)
898{
899 /* only set blocking for read lock */
900 ASSERT(atomic_read(&dev_replace->read_locks) > 0);
901 atomic_inc(&dev_replace->blocking_readers);
902 read_unlock(&dev_replace->lock);
903}
904
905/* acquire read lock and dec blocking cnt */
906void btrfs_dev_replace_clear_lock_blocking(
907 struct btrfs_dev_replace *dev_replace)
908{
909 /* only set blocking for read lock */
910 ASSERT(atomic_read(&dev_replace->read_locks) > 0);
911 ASSERT(atomic_read(&dev_replace->blocking_readers) > 0);
912 read_lock(&dev_replace->lock);
913 if (atomic_dec_and_test(&dev_replace->blocking_readers) &&
914 waitqueue_active(&dev_replace->read_lock_wq))
915 wake_up(&dev_replace->read_lock_wq);
916}
917
Miao Xiec404e0d2014-01-30 16:46:55 +0800918void btrfs_bio_counter_inc_noblocked(struct btrfs_fs_info *fs_info)
919{
920 percpu_counter_inc(&fs_info->bio_counter);
921}
922
Miao Xie42452152014-11-25 16:39:28 +0800923void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
Miao Xiec404e0d2014-01-30 16:46:55 +0800924{
Miao Xie42452152014-11-25 16:39:28 +0800925 percpu_counter_sub(&fs_info->bio_counter, amount);
Miao Xiec404e0d2014-01-30 16:46:55 +0800926
927 if (waitqueue_active(&fs_info->replace_wait))
928 wake_up(&fs_info->replace_wait);
929}
930
931void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
932{
Zhao Lei09dd7a02015-01-20 15:11:37 +0800933 while (1) {
934 percpu_counter_inc(&fs_info->bio_counter);
935 if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
936 &fs_info->fs_state)))
937 break;
938
Miao Xiec404e0d2014-01-30 16:46:55 +0800939 btrfs_bio_counter_dec(fs_info);
940 wait_event(fs_info->replace_wait,
941 !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
942 &fs_info->fs_state));
Miao Xiec404e0d2014-01-30 16:46:55 +0800943 }
Miao Xiec404e0d2014-01-30 16:46:55 +0800944}