blob: 66e02ebdd340890affc90063b12dd687ad995a0a [file] [log] [blame]
David Sterbac1d7c512018-04-03 19:23:33 +02001// SPDX-License-Identifier: GPL-2.0
Chris Mason925baed2008-06-25 16:01:30 -04002/*
3 * Copyright (C) 2008 Oracle. All rights reserved.
Chris Mason925baed2008-06-25 16:01:30 -04004 */
David Sterbac1d7c512018-04-03 19:23:33 +02005
Chris Mason925baed2008-06-25 16:01:30 -04006#include <linux/sched.h>
Chris Mason925baed2008-06-25 16:01:30 -04007#include <linux/pagemap.h>
8#include <linux/spinlock.h>
9#include <linux/page-flags.h>
Chris Mason4881ee52008-07-24 09:51:08 -040010#include <asm/bug.h>
David Sterba602cbe92019-08-21 18:48:25 +020011#include "misc.h"
Chris Mason925baed2008-06-25 16:01:30 -040012#include "ctree.h"
13#include "extent_io.h"
14#include "locking.h"
15
David Sterbad4e253b2019-10-16 18:29:10 +020016/*
17 * Extent buffer locking
18 * =====================
19 *
20 * The locks use a custom scheme that allows to do more operations than are
21 * available fromt current locking primitives. The building blocks are still
22 * rwlock and wait queues.
23 *
24 * Required semantics:
25 *
26 * - reader/writer exclusion
27 * - writer/writer exclusion
28 * - reader/reader sharing
29 * - spinning lock semantics
30 * - blocking lock semantics
31 * - try-lock semantics for readers and writers
32 * - one level nesting, allowing read lock to be taken by the same thread that
33 * already has write lock
34 *
35 * The extent buffer locks (also called tree locks) manage access to eb data
36 * related to the storage in the b-tree (keys, items, but not the individual
37 * members of eb).
38 * We want concurrency of many readers and safe updates. The underlying locking
39 * is done by read-write spinlock and the blocking part is implemented using
40 * counters and wait queues.
41 *
42 * spinning semantics - the low-level rwlock is held so all other threads that
43 * want to take it are spinning on it.
44 *
45 * blocking semantics - the low-level rwlock is not held but the counter
46 * denotes how many times the blocking lock was held;
47 * sleeping is possible
48 *
49 * Write lock always allows only one thread to access the data.
50 *
51 *
52 * Debugging
53 * ---------
54 *
55 * There are additional state counters that are asserted in various contexts,
56 * removed from non-debug build to reduce extent_buffer size and for
57 * performance reasons.
58 *
59 *
Josef Bacik329ced72020-08-20 11:46:00 -040060 * Lock recursion
61 * --------------
David Sterbad4e253b2019-10-16 18:29:10 +020062 *
63 * A write operation on a tree might indirectly start a look up on the same
64 * tree. This can happen when btrfs_cow_block locks the tree and needs to
65 * lookup free extents.
66 *
67 * btrfs_cow_block
68 * ..
69 * alloc_tree_block_no_bg_flush
70 * btrfs_alloc_tree_block
71 * btrfs_reserve_extent
72 * ..
73 * load_free_space_cache
74 * ..
75 * btrfs_lookup_file_extent
76 * btrfs_search_slot
77 *
78 *
79 * Locking pattern - spinning
80 * --------------------------
81 *
82 * The simple locking scenario, the +--+ denotes the spinning section.
83 *
84 * +- btrfs_tree_lock
85 * | - extent_buffer::rwlock is held
86 * | - no heavy operations should happen, eg. IO, memory allocations, large
87 * | structure traversals
88 * +- btrfs_tree_unock
89*
90*
91 * Locking pattern - blocking
92 * --------------------------
93 *
94 * The blocking write uses the following scheme. The +--+ denotes the spinning
95 * section.
96 *
97 * +- btrfs_tree_lock
98 * |
99 * +- btrfs_set_lock_blocking_write
100 *
101 * - allowed: IO, memory allocations, etc.
102 *
103 * -- btrfs_tree_unlock - note, no explicit unblocking necessary
104 *
105 *
106 * Blocking read is similar.
107 *
108 * +- btrfs_tree_read_lock
109 * |
110 * +- btrfs_set_lock_blocking_read
111 *
112 * - heavy operations allowed
113 *
114 * +- btrfs_tree_read_unlock_blocking
115 * |
116 * +- btrfs_tree_read_unlock
117 *
118 */
119
David Sterbae4e9fd02018-08-24 14:45:20 +0200120#ifdef CONFIG_BTRFS_DEBUG
David Sterbad6156212019-09-24 18:29:10 +0200121static inline void btrfs_assert_spinning_writers_get(struct extent_buffer *eb)
David Sterbae4e9fd02018-08-24 14:45:20 +0200122{
David Sterbaf3dc24c2019-05-02 16:51:53 +0200123 WARN_ON(eb->spinning_writers);
124 eb->spinning_writers++;
David Sterbae4e9fd02018-08-24 14:45:20 +0200125}
126
David Sterbad6156212019-09-24 18:29:10 +0200127static inline void btrfs_assert_spinning_writers_put(struct extent_buffer *eb)
David Sterbae4e9fd02018-08-24 14:45:20 +0200128{
David Sterbaf3dc24c2019-05-02 16:51:53 +0200129 WARN_ON(eb->spinning_writers != 1);
130 eb->spinning_writers--;
David Sterbae4e9fd02018-08-24 14:45:20 +0200131}
132
David Sterbad6156212019-09-24 18:29:10 +0200133static inline void btrfs_assert_no_spinning_writers(struct extent_buffer *eb)
David Sterbae4e9fd02018-08-24 14:45:20 +0200134{
David Sterbaf3dc24c2019-05-02 16:51:53 +0200135 WARN_ON(eb->spinning_writers);
David Sterbae4e9fd02018-08-24 14:45:20 +0200136}
137
David Sterbad6156212019-09-24 18:29:10 +0200138static inline void btrfs_assert_spinning_readers_get(struct extent_buffer *eb)
David Sterba225948d2018-08-24 15:53:42 +0200139{
140 atomic_inc(&eb->spinning_readers);
141}
142
David Sterbad6156212019-09-24 18:29:10 +0200143static inline void btrfs_assert_spinning_readers_put(struct extent_buffer *eb)
David Sterba225948d2018-08-24 15:53:42 +0200144{
145 WARN_ON(atomic_read(&eb->spinning_readers) == 0);
146 atomic_dec(&eb->spinning_readers);
147}
148
David Sterbad6156212019-09-24 18:29:10 +0200149static inline void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb)
David Sterba58a2ddae2018-08-24 16:13:41 +0200150{
151 atomic_inc(&eb->read_locks);
152}
153
David Sterbad6156212019-09-24 18:29:10 +0200154static inline void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb)
David Sterba58a2ddae2018-08-24 16:13:41 +0200155{
156 atomic_dec(&eb->read_locks);
157}
158
David Sterbad6156212019-09-24 18:29:10 +0200159static inline void btrfs_assert_tree_read_locked(struct extent_buffer *eb)
David Sterba58a2ddae2018-08-24 16:13:41 +0200160{
161 BUG_ON(!atomic_read(&eb->read_locks));
162}
163
David Sterbad6156212019-09-24 18:29:10 +0200164static inline void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb)
David Sterbae3f15382018-08-24 16:20:02 +0200165{
David Sterba00801ae2019-05-02 16:53:47 +0200166 eb->write_locks++;
David Sterbae3f15382018-08-24 16:20:02 +0200167}
168
David Sterbad6156212019-09-24 18:29:10 +0200169static inline void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb)
David Sterbae3f15382018-08-24 16:20:02 +0200170{
David Sterba00801ae2019-05-02 16:53:47 +0200171 eb->write_locks--;
David Sterbae3f15382018-08-24 16:20:02 +0200172}
173
David Sterbae4e9fd02018-08-24 14:45:20 +0200174#else
175static void btrfs_assert_spinning_writers_get(struct extent_buffer *eb) { }
176static void btrfs_assert_spinning_writers_put(struct extent_buffer *eb) { }
177static void btrfs_assert_no_spinning_writers(struct extent_buffer *eb) { }
David Sterba225948d2018-08-24 15:53:42 +0200178static void btrfs_assert_spinning_readers_put(struct extent_buffer *eb) { }
179static void btrfs_assert_spinning_readers_get(struct extent_buffer *eb) { }
David Sterba58a2ddae2018-08-24 16:13:41 +0200180static void btrfs_assert_tree_read_locked(struct extent_buffer *eb) { }
181static void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb) { }
182static void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb) { }
David Sterbae3f15382018-08-24 16:20:02 +0200183static void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb) { }
184static void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb) { }
David Sterbae4e9fd02018-08-24 14:45:20 +0200185#endif
186
David Sterbad4e253b2019-10-16 18:29:10 +0200187/*
188 * Mark already held read lock as blocking. Can be nested in write lock by the
189 * same thread.
190 *
191 * Use when there are potentially long operations ahead so other thread waiting
192 * on the lock will not actively spin but sleep instead.
193 *
194 * The rwlock is released and blocking reader counter is increased.
195 */
David Sterbab95be2d2018-04-04 01:43:05 +0200196void btrfs_set_lock_blocking_read(struct extent_buffer *eb)
Chris Masonb4ce94d2009-02-04 09:25:08 -0500197{
Qu Wenruo31aab402019-04-15 21:15:25 +0800198 trace_btrfs_set_lock_blocking_read(eb);
Chris Masonea4ebde2014-06-19 14:16:52 -0700199 /*
David Sterbab95be2d2018-04-04 01:43:05 +0200200 * No lock is required. The lock owner may change if we have a read
201 * lock, but it won't change to or away from us. If we have the write
202 * lock, we are the owner and it'll never change.
Chris Masonea4ebde2014-06-19 14:16:52 -0700203 */
Josef Bacik329ced72020-08-20 11:46:00 -0400204 if (eb->lock_recursed && current->pid == eb->lock_owner)
Chris Masonea4ebde2014-06-19 14:16:52 -0700205 return;
David Sterbab95be2d2018-04-04 01:43:05 +0200206 btrfs_assert_tree_read_locked(eb);
207 atomic_inc(&eb->blocking_readers);
David Sterbaafd495a2018-08-24 15:57:38 +0200208 btrfs_assert_spinning_readers_put(eb);
David Sterbab95be2d2018-04-04 01:43:05 +0200209 read_unlock(&eb->lock);
210}
211
David Sterbad4e253b2019-10-16 18:29:10 +0200212/*
213 * Mark already held write lock as blocking.
214 *
215 * Use when there are potentially long operations ahead so other threads
216 * waiting on the lock will not actively spin but sleep instead.
217 *
218 * The rwlock is released and blocking writers is set.
219 */
David Sterbab95be2d2018-04-04 01:43:05 +0200220void btrfs_set_lock_blocking_write(struct extent_buffer *eb)
221{
Qu Wenruo31aab402019-04-15 21:15:25 +0800222 trace_btrfs_set_lock_blocking_write(eb);
David Sterbab95be2d2018-04-04 01:43:05 +0200223 /*
224 * No lock is required. The lock owner may change if we have a read
225 * lock, but it won't change to or away from us. If we have the write
226 * lock, we are the owner and it'll never change.
227 */
Josef Bacik329ced72020-08-20 11:46:00 -0400228 if (eb->lock_recursed && current->pid == eb->lock_owner)
David Sterbab95be2d2018-04-04 01:43:05 +0200229 return;
David Sterba06297d82019-05-02 16:47:23 +0200230 if (eb->blocking_writers == 0) {
David Sterba843ccf92018-08-24 14:56:28 +0200231 btrfs_assert_spinning_writers_put(eb);
David Sterbab95be2d2018-04-04 01:43:05 +0200232 btrfs_assert_tree_locked(eb);
David Sterbaa44779882019-10-11 00:03:14 +0200233 WRITE_ONCE(eb->blocking_writers, 1);
David Sterbab95be2d2018-04-04 01:43:05 +0200234 write_unlock(&eb->lock);
Chris Masonbd681512011-07-16 15:23:14 -0400235 }
Chris Masonb4ce94d2009-02-04 09:25:08 -0500236}
Chris Masond3977122009-01-05 21:25:51 -0500237
Chris Masonb4ce94d2009-02-04 09:25:08 -0500238/*
David Sterbad4e253b2019-10-16 18:29:10 +0200239 * Lock the extent buffer for read. Wait for any writers (spinning or blocking).
240 * Can be nested in write lock by the same thread.
241 *
242 * Use when the locked section does only lightweight actions and busy waiting
243 * would be cheaper than making other threads do the wait/wake loop.
244 *
245 * The rwlock is held upon exit.
Chris Masonb4ce94d2009-02-04 09:25:08 -0500246 */
Josef Bacikfd7ba1c2020-08-20 11:46:02 -0400247void __btrfs_tree_read_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest,
248 bool recurse)
Chris Masonb4ce94d2009-02-04 09:25:08 -0500249{
Qu Wenruo34e73cc2019-04-15 21:15:24 +0800250 u64 start_ns = 0;
251
252 if (trace_btrfs_tree_read_lock_enabled())
253 start_ns = ktime_get_ns();
Chris Masonbd681512011-07-16 15:23:14 -0400254again:
Arne Jansen5b25f702011-09-13 10:55:48 +0200255 read_lock(&eb->lock);
David Sterba06297d82019-05-02 16:47:23 +0200256 BUG_ON(eb->blocking_writers == 0 &&
257 current->pid == eb->lock_owner);
David Sterba06297d82019-05-02 16:47:23 +0200258 if (eb->blocking_writers) {
David Sterbaf5c2a522019-10-10 23:29:21 +0200259 if (current->pid == eb->lock_owner) {
260 /*
261 * This extent is already write-locked by our thread.
262 * We allow an additional read lock to be added because
263 * it's for the same thread. btrfs_find_all_roots()
264 * depends on this as it may be called on a partly
265 * (write-)locked tree.
266 */
Josef Bacik51899412020-08-20 11:46:01 -0400267 WARN_ON(!recurse);
Josef Bacik329ced72020-08-20 11:46:00 -0400268 BUG_ON(eb->lock_recursed);
269 eb->lock_recursed = true;
David Sterbaf5c2a522019-10-10 23:29:21 +0200270 read_unlock(&eb->lock);
271 trace_btrfs_tree_read_lock(eb, start_ns);
272 return;
273 }
Chris Masonbd681512011-07-16 15:23:14 -0400274 read_unlock(&eb->lock);
Liu Bo39f9d022012-12-27 09:01:22 +0000275 wait_event(eb->write_lock_wq,
David Sterbaa44779882019-10-11 00:03:14 +0200276 READ_ONCE(eb->blocking_writers) == 0);
Chris Masonbd681512011-07-16 15:23:14 -0400277 goto again;
Chris Masonb4ce94d2009-02-04 09:25:08 -0500278 }
David Sterba5c9c7992018-08-24 16:15:51 +0200279 btrfs_assert_tree_read_locks_get(eb);
David Sterbaafd495a2018-08-24 15:57:38 +0200280 btrfs_assert_spinning_readers_get(eb);
Qu Wenruo34e73cc2019-04-15 21:15:24 +0800281 trace_btrfs_tree_read_lock(eb, start_ns);
Chris Masonb4ce94d2009-02-04 09:25:08 -0500282}
283
Josef Bacik51899412020-08-20 11:46:01 -0400284void btrfs_tree_read_lock(struct extent_buffer *eb)
285{
Josef Bacikfd7ba1c2020-08-20 11:46:02 -0400286 __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL, false);
Josef Bacik51899412020-08-20 11:46:01 -0400287}
288
Chris Masonb4ce94d2009-02-04 09:25:08 -0500289/*
David Sterbad4e253b2019-10-16 18:29:10 +0200290 * Lock extent buffer for read, optimistically expecting that there are no
291 * contending blocking writers. If there are, don't wait.
292 *
293 * Return 1 if the rwlock has been taken, 0 otherwise
Chris Masonf82c4582014-11-19 10:25:09 -0800294 */
295int btrfs_tree_read_lock_atomic(struct extent_buffer *eb)
296{
David Sterbaa44779882019-10-11 00:03:14 +0200297 if (READ_ONCE(eb->blocking_writers))
Chris Masonf82c4582014-11-19 10:25:09 -0800298 return 0;
299
300 read_lock(&eb->lock);
David Sterbaa44779882019-10-11 00:03:14 +0200301 /* Refetch value after lock */
302 if (READ_ONCE(eb->blocking_writers)) {
Chris Masonf82c4582014-11-19 10:25:09 -0800303 read_unlock(&eb->lock);
304 return 0;
305 }
David Sterba5c9c7992018-08-24 16:15:51 +0200306 btrfs_assert_tree_read_locks_get(eb);
David Sterbaafd495a2018-08-24 15:57:38 +0200307 btrfs_assert_spinning_readers_get(eb);
Qu Wenruo31aab402019-04-15 21:15:25 +0800308 trace_btrfs_tree_read_lock_atomic(eb);
Chris Masonf82c4582014-11-19 10:25:09 -0800309 return 1;
310}
311
312/*
David Sterbad4e253b2019-10-16 18:29:10 +0200313 * Try-lock for read. Don't block or wait for contending writers.
314 *
315 * Retrun 1 if the rwlock has been taken, 0 otherwise
Chris Masonb4ce94d2009-02-04 09:25:08 -0500316 */
Chris Masonbd681512011-07-16 15:23:14 -0400317int btrfs_try_tree_read_lock(struct extent_buffer *eb)
Chris Mason925baed2008-06-25 16:01:30 -0400318{
David Sterbaa44779882019-10-11 00:03:14 +0200319 if (READ_ONCE(eb->blocking_writers))
Chris Masonbd681512011-07-16 15:23:14 -0400320 return 0;
Chris Mason66d7e852009-03-12 20:12:45 -0400321
Chris Masonea4ebde2014-06-19 14:16:52 -0700322 if (!read_trylock(&eb->lock))
323 return 0;
324
David Sterbaa44779882019-10-11 00:03:14 +0200325 /* Refetch value after lock */
326 if (READ_ONCE(eb->blocking_writers)) {
Chris Masonbd681512011-07-16 15:23:14 -0400327 read_unlock(&eb->lock);
328 return 0;
Chris Masonf9efa9c2008-06-25 16:14:04 -0400329 }
David Sterba5c9c7992018-08-24 16:15:51 +0200330 btrfs_assert_tree_read_locks_get(eb);
David Sterbaafd495a2018-08-24 15:57:38 +0200331 btrfs_assert_spinning_readers_get(eb);
Qu Wenruo31aab402019-04-15 21:15:25 +0800332 trace_btrfs_try_tree_read_lock(eb);
Chris Masonb4ce94d2009-02-04 09:25:08 -0500333 return 1;
334}
335
336/*
David Sterbad4e253b2019-10-16 18:29:10 +0200337 * Try-lock for write. May block until the lock is uncontended, but does not
338 * wait until it is free.
339 *
340 * Retrun 1 if the rwlock has been taken, 0 otherwise
Chris Masonbd681512011-07-16 15:23:14 -0400341 */
342int btrfs_try_tree_write_lock(struct extent_buffer *eb)
343{
David Sterbaa44779882019-10-11 00:03:14 +0200344 if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers))
Chris Masonbd681512011-07-16 15:23:14 -0400345 return 0;
Chris Masonea4ebde2014-06-19 14:16:52 -0700346
Chris Masonf82c4582014-11-19 10:25:09 -0800347 write_lock(&eb->lock);
David Sterbaa44779882019-10-11 00:03:14 +0200348 /* Refetch value after lock */
349 if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers)) {
Chris Masonbd681512011-07-16 15:23:14 -0400350 write_unlock(&eb->lock);
351 return 0;
352 }
David Sterbac79adfc2018-08-24 16:24:26 +0200353 btrfs_assert_tree_write_locks_get(eb);
David Sterba843ccf92018-08-24 14:56:28 +0200354 btrfs_assert_spinning_writers_get(eb);
Arne Jansen5b25f702011-09-13 10:55:48 +0200355 eb->lock_owner = current->pid;
Qu Wenruo31aab402019-04-15 21:15:25 +0800356 trace_btrfs_try_tree_write_lock(eb);
Chris Masonbd681512011-07-16 15:23:14 -0400357 return 1;
358}
359
360/*
David Sterbad4e253b2019-10-16 18:29:10 +0200361 * Release read lock. Must be used only if the lock is in spinning mode. If
362 * the read lock is nested, must pair with read lock before the write unlock.
363 *
364 * The rwlock is not held upon exit.
Chris Masonbd681512011-07-16 15:23:14 -0400365 */
366void btrfs_tree_read_unlock(struct extent_buffer *eb)
367{
Qu Wenruo31aab402019-04-15 21:15:25 +0800368 trace_btrfs_tree_read_unlock(eb);
Chris Masonea4ebde2014-06-19 14:16:52 -0700369 /*
370 * if we're nested, we have the write lock. No new locking
371 * is needed as long as we are the lock owner.
Josef Bacik329ced72020-08-20 11:46:00 -0400372 * The write unlock will do a barrier for us, and the lock_recursed
Chris Masonea4ebde2014-06-19 14:16:52 -0700373 * field only matters to the lock owner.
374 */
Josef Bacik329ced72020-08-20 11:46:00 -0400375 if (eb->lock_recursed && current->pid == eb->lock_owner) {
376 eb->lock_recursed = false;
Chris Masonea4ebde2014-06-19 14:16:52 -0700377 return;
Arne Jansen5b25f702011-09-13 10:55:48 +0200378 }
Chris Masonbd681512011-07-16 15:23:14 -0400379 btrfs_assert_tree_read_locked(eb);
David Sterbaafd495a2018-08-24 15:57:38 +0200380 btrfs_assert_spinning_readers_put(eb);
David Sterba5c9c7992018-08-24 16:15:51 +0200381 btrfs_assert_tree_read_locks_put(eb);
Chris Masonbd681512011-07-16 15:23:14 -0400382 read_unlock(&eb->lock);
383}
384
385/*
David Sterbad4e253b2019-10-16 18:29:10 +0200386 * Release read lock, previously set to blocking by a pairing call to
387 * btrfs_set_lock_blocking_read(). Can be nested in write lock by the same
388 * thread.
389 *
390 * State of rwlock is unchanged, last reader wakes waiting threads.
Chris Masonbd681512011-07-16 15:23:14 -0400391 */
392void btrfs_tree_read_unlock_blocking(struct extent_buffer *eb)
393{
Qu Wenruo31aab402019-04-15 21:15:25 +0800394 trace_btrfs_tree_read_unlock_blocking(eb);
Chris Masonea4ebde2014-06-19 14:16:52 -0700395 /*
396 * if we're nested, we have the write lock. No new locking
397 * is needed as long as we are the lock owner.
Josef Bacik329ced72020-08-20 11:46:00 -0400398 * The write unlock will do a barrier for us, and the lock_recursed
Chris Masonea4ebde2014-06-19 14:16:52 -0700399 * field only matters to the lock owner.
400 */
Josef Bacik329ced72020-08-20 11:46:00 -0400401 if (eb->lock_recursed && current->pid == eb->lock_owner) {
402 eb->lock_recursed = false;
Chris Masonea4ebde2014-06-19 14:16:52 -0700403 return;
Arne Jansen5b25f702011-09-13 10:55:48 +0200404 }
Chris Masonbd681512011-07-16 15:23:14 -0400405 btrfs_assert_tree_read_locked(eb);
406 WARN_ON(atomic_read(&eb->blocking_readers) == 0);
David Sterba093258e62018-02-26 16:15:17 +0100407 /* atomic_dec_and_test implies a barrier */
408 if (atomic_dec_and_test(&eb->blocking_readers))
409 cond_wake_up_nomb(&eb->read_lock_wq);
David Sterba5c9c7992018-08-24 16:15:51 +0200410 btrfs_assert_tree_read_locks_put(eb);
Chris Masonbd681512011-07-16 15:23:14 -0400411}
412
413/*
David Sterbad4e253b2019-10-16 18:29:10 +0200414 * Lock for write. Wait for all blocking and spinning readers and writers. This
415 * starts context where reader lock could be nested by the same thread.
416 *
417 * The rwlock is held for write upon exit.
Chris Masonb4ce94d2009-02-04 09:25:08 -0500418 */
Josef Bacikfd7ba1c2020-08-20 11:46:02 -0400419void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
Jules Irenge78d933c2020-03-31 21:46:42 +0100420 __acquires(&eb->lock)
Chris Masonb4ce94d2009-02-04 09:25:08 -0500421{
Qu Wenruo34e73cc2019-04-15 21:15:24 +0800422 u64 start_ns = 0;
423
424 if (trace_btrfs_tree_lock_enabled())
425 start_ns = ktime_get_ns();
426
Zhaolei166f66d2015-08-06 22:39:36 +0800427 WARN_ON(eb->lock_owner == current->pid);
Chris Masonbd681512011-07-16 15:23:14 -0400428again:
429 wait_event(eb->read_lock_wq, atomic_read(&eb->blocking_readers) == 0);
David Sterbaa44779882019-10-11 00:03:14 +0200430 wait_event(eb->write_lock_wq, READ_ONCE(eb->blocking_writers) == 0);
Chris Masonbd681512011-07-16 15:23:14 -0400431 write_lock(&eb->lock);
David Sterbaa44779882019-10-11 00:03:14 +0200432 /* Refetch value after lock */
433 if (atomic_read(&eb->blocking_readers) ||
434 READ_ONCE(eb->blocking_writers)) {
Chris Masonbd681512011-07-16 15:23:14 -0400435 write_unlock(&eb->lock);
Chris Masonbd681512011-07-16 15:23:14 -0400436 goto again;
437 }
David Sterba843ccf92018-08-24 14:56:28 +0200438 btrfs_assert_spinning_writers_get(eb);
David Sterbac79adfc2018-08-24 16:24:26 +0200439 btrfs_assert_tree_write_locks_get(eb);
Arne Jansen5b25f702011-09-13 10:55:48 +0200440 eb->lock_owner = current->pid;
Qu Wenruo34e73cc2019-04-15 21:15:24 +0800441 trace_btrfs_tree_lock(eb, start_ns);
Chris Masonb4ce94d2009-02-04 09:25:08 -0500442}
443
Josef Bacikfd7ba1c2020-08-20 11:46:02 -0400444void btrfs_tree_lock(struct extent_buffer *eb)
445{
446 __btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
447}
448
Chris Masonbd681512011-07-16 15:23:14 -0400449/*
David Sterbad4e253b2019-10-16 18:29:10 +0200450 * Release the write lock, either blocking or spinning (ie. there's no need
451 * for an explicit blocking unlock, like btrfs_tree_read_unlock_blocking).
452 * This also ends the context for nesting, the read lock must have been
453 * released already.
454 *
455 * Tasks blocked and waiting are woken, rwlock is not held upon exit.
Chris Masonbd681512011-07-16 15:23:14 -0400456 */
Jeff Mahoney143bede2012-03-01 14:56:26 +0100457void btrfs_tree_unlock(struct extent_buffer *eb)
Chris Mason925baed2008-06-25 16:01:30 -0400458{
David Sterbaa44779882019-10-11 00:03:14 +0200459 /*
460 * This is read both locked and unlocked but always by the same thread
461 * that already owns the lock so we don't need to use READ_ONCE
462 */
David Sterba06297d82019-05-02 16:47:23 +0200463 int blockers = eb->blocking_writers;
Chris Masonb4ce94d2009-02-04 09:25:08 -0500464
Chris Masonbd681512011-07-16 15:23:14 -0400465 BUG_ON(blockers > 1);
466
467 btrfs_assert_tree_locked(eb);
Qu Wenruo31aab402019-04-15 21:15:25 +0800468 trace_btrfs_tree_unlock(eb);
Chris Masonea4ebde2014-06-19 14:16:52 -0700469 eb->lock_owner = 0;
David Sterbac79adfc2018-08-24 16:24:26 +0200470 btrfs_assert_tree_write_locks_put(eb);
Chris Masonbd681512011-07-16 15:23:14 -0400471
472 if (blockers) {
David Sterba843ccf92018-08-24 14:56:28 +0200473 btrfs_assert_no_spinning_writers(eb);
David Sterbaa44779882019-10-11 00:03:14 +0200474 /* Unlocked write */
475 WRITE_ONCE(eb->blocking_writers, 0);
Nikolay Borisov6e7ca092019-07-25 11:27:29 +0300476 /*
477 * We need to order modifying blocking_writers above with
478 * actually waking up the sleepers to ensure they see the
479 * updated value of blocking_writers
480 */
481 cond_wake_up(&eb->write_lock_wq);
Chris Masonbd681512011-07-16 15:23:14 -0400482 } else {
David Sterba843ccf92018-08-24 14:56:28 +0200483 btrfs_assert_spinning_writers_put(eb);
Chris Masonbd681512011-07-16 15:23:14 -0400484 write_unlock(&eb->lock);
485 }
Chris Mason925baed2008-06-25 16:01:30 -0400486}
David Sterbaed2b1d32019-09-24 19:17:17 +0200487
488/*
489 * Set all locked nodes in the path to blocking locks. This should be done
490 * before scheduling
491 */
492void btrfs_set_path_blocking(struct btrfs_path *p)
493{
494 int i;
495
496 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
497 if (!p->nodes[i] || !p->locks[i])
498 continue;
499 /*
500 * If we currently have a spinning reader or writer lock this
501 * will bump the count of blocking holders and drop the
502 * spinlock.
503 */
504 if (p->locks[i] == BTRFS_READ_LOCK) {
505 btrfs_set_lock_blocking_read(p->nodes[i]);
506 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
507 } else if (p->locks[i] == BTRFS_WRITE_LOCK) {
508 btrfs_set_lock_blocking_write(p->nodes[i]);
509 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
510 }
511 }
512}
David Sterba1f95ec02019-09-24 19:17:17 +0200513
514/*
515 * This releases any locks held in the path starting at level and going all the
516 * way up to the root.
517 *
518 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
519 * cases, such as COW of the block at slot zero in the node. This ignores
520 * those rules, and it should only be called when there are no more updates to
521 * be done higher up in the tree.
522 */
523void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
524{
525 int i;
526
527 if (path->keep_locks)
528 return;
529
530 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
531 if (!path->nodes[i])
532 continue;
533 if (!path->locks[i])
534 continue;
535 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
536 path->locks[i] = 0;
537 }
538}
David Sterbab908c332020-02-05 17:26:51 +0100539
540/*
541 * Loop around taking references on and locking the root node of the tree until
542 * we end up with a lock on the root node.
543 *
544 * Return: root extent buffer with write lock held
545 */
546struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
547{
548 struct extent_buffer *eb;
549
550 while (1) {
551 eb = btrfs_root_node(root);
552 btrfs_tree_lock(eb);
553 if (eb == root->node)
554 break;
555 btrfs_tree_unlock(eb);
556 free_extent_buffer(eb);
557 }
558 return eb;
559}
560
561/*
562 * Loop around taking references on and locking the root node of the tree until
563 * we end up with a lock on the root node.
564 *
565 * Return: root extent buffer with read lock held
566 */
Josef Bacik51899412020-08-20 11:46:01 -0400567struct extent_buffer *__btrfs_read_lock_root_node(struct btrfs_root *root,
568 bool recurse)
David Sterbab908c332020-02-05 17:26:51 +0100569{
570 struct extent_buffer *eb;
571
572 while (1) {
573 eb = btrfs_root_node(root);
Josef Bacikfd7ba1c2020-08-20 11:46:02 -0400574 __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL, recurse);
David Sterbab908c332020-02-05 17:26:51 +0100575 if (eb == root->node)
576 break;
577 btrfs_tree_read_unlock(eb);
578 free_extent_buffer(eb);
579 }
580 return eb;
581}
Nikolay Borisov2992df72020-01-30 14:59:44 +0200582
583/*
584 * DREW locks
585 * ==========
586 *
587 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
588 * where you want to provide A-B exclusion but not AA or BB.
589 *
590 * Currently implementation gives more priority to reader. If a reader and a
591 * writer both race to acquire their respective sides of the lock the writer
592 * would yield its lock as soon as it detects a concurrent reader. Additionally
593 * if there are pending readers no new writers would be allowed to come in and
594 * acquire the lock.
595 */
596
597int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
598{
599 int ret;
600
601 ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
602 if (ret)
603 return ret;
604
605 atomic_set(&lock->readers, 0);
606 init_waitqueue_head(&lock->pending_readers);
607 init_waitqueue_head(&lock->pending_writers);
608
609 return 0;
610}
611
612void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
613{
614 percpu_counter_destroy(&lock->writers);
615}
616
617/* Return true if acquisition is successful, false otherwise */
618bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
619{
620 if (atomic_read(&lock->readers))
621 return false;
622
623 percpu_counter_inc(&lock->writers);
624
625 /* Ensure writers count is updated before we check for pending readers */
626 smp_mb();
627 if (atomic_read(&lock->readers)) {
628 btrfs_drew_write_unlock(lock);
629 return false;
630 }
631
632 return true;
633}
634
635void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
636{
637 while (true) {
638 if (btrfs_drew_try_write_lock(lock))
639 return;
640 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
641 }
642}
643
644void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
645{
646 percpu_counter_dec(&lock->writers);
647 cond_wake_up(&lock->pending_readers);
648}
649
650void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
651{
652 atomic_inc(&lock->readers);
653
654 /*
655 * Ensure the pending reader count is perceieved BEFORE this reader
656 * goes to sleep in case of active writers. This guarantees new writers
657 * won't be allowed and that the current reader will be woken up when
658 * the last active writer finishes its jobs.
659 */
660 smp_mb__after_atomic();
661
662 wait_event(lock->pending_readers,
663 percpu_counter_sum(&lock->writers) == 0);
664}
665
666void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
667{
668 /*
669 * atomic_dec_and_test implies a full barrier, so woken up writers
670 * are guaranteed to see the decrement
671 */
672 if (atomic_dec_and_test(&lock->readers))
673 wake_up(&lock->pending_writers);
674}