locking.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (C) 2008 Oracle.  All rights reserved.
  4 */
  5
  6#include <linux/sched.h>
  7#include <linux/pagemap.h>
  8#include <linux/spinlock.h>
  9#include <linux/page-flags.h>
 10#include <asm/bug.h>
 11#include <trace/events/btrfs.h>
 12#include "ctree.h"
 13#include "extent_io.h"
 14#include "locking.h"
 15
 16/*
 17 * Lockdep class keys for extent_buffer->lock's in this root.  For a given
 18 * eb, the lockdep key is determined by the btrfs_root it belongs to and
 19 * the level the eb occupies in the tree.
 20 *
 21 * Different roots are used for different purposes and may nest inside each
 22 * other and they require separate keysets.  As lockdep keys should be
 23 * static, assign keysets according to the purpose of the root as indicated
 24 * by btrfs_root->root_key.objectid.  This ensures that all special purpose
 25 * roots have separate keysets.
 26 *
 27 * Lock-nesting across peer nodes is always done with the immediate parent
 28 * node locked thus preventing deadlock.  As lockdep doesn't know this, use
 29 * subclass to avoid triggering lockdep warning in such cases.
 30 *
 31 * The key is set by the readpage_end_io_hook after the buffer has passed
 32 * csum validation but before the pages are unlocked.  It is also set by
 33 * btrfs_init_new_buffer on freshly allocated blocks.
 34 *
 35 * We also add a check to make sure the highest level of the tree is the
 36 * same as our lockdep setup here.  If BTRFS_MAX_LEVEL changes, this code
 37 * needs update as well.
 38 */
 39#ifdef CONFIG_DEBUG_LOCK_ALLOC
 40#if BTRFS_MAX_LEVEL != 8
 41#error
 42#endif
 43
 44#define DEFINE_LEVEL(stem, level)					\
 45	.names[level] = "btrfs-" stem "-0" #level,
 46
 47#define DEFINE_NAME(stem)						\
 48	DEFINE_LEVEL(stem, 0)						\
 49	DEFINE_LEVEL(stem, 1)						\
 50	DEFINE_LEVEL(stem, 2)						\
 51	DEFINE_LEVEL(stem, 3)						\
 52	DEFINE_LEVEL(stem, 4)						\
 53	DEFINE_LEVEL(stem, 5)						\
 54	DEFINE_LEVEL(stem, 6)						\
 55	DEFINE_LEVEL(stem, 7)
 56
 57static struct btrfs_lockdep_keyset {
 58	u64			id;		/* root objectid */
 59	/* Longest entry: btrfs-block-group-00 */
 60	char			names[BTRFS_MAX_LEVEL][24];
 61	struct lock_class_key	keys[BTRFS_MAX_LEVEL];
 62} btrfs_lockdep_keysets[] = {
 63	{ .id = BTRFS_ROOT_TREE_OBJECTID,	DEFINE_NAME("root")	},
 64	{ .id = BTRFS_EXTENT_TREE_OBJECTID,	DEFINE_NAME("extent")	},
 65	{ .id = BTRFS_CHUNK_TREE_OBJECTID,	DEFINE_NAME("chunk")	},
 66	{ .id = BTRFS_DEV_TREE_OBJECTID,	DEFINE_NAME("dev")	},
 67	{ .id = BTRFS_CSUM_TREE_OBJECTID,	DEFINE_NAME("csum")	},
 68	{ .id = BTRFS_QUOTA_TREE_OBJECTID,	DEFINE_NAME("quota")	},
 69	{ .id = BTRFS_TREE_LOG_OBJECTID,	DEFINE_NAME("log")	},
 70	{ .id = BTRFS_TREE_RELOC_OBJECTID,	DEFINE_NAME("treloc")	},
 71	{ .id = BTRFS_DATA_RELOC_TREE_OBJECTID,	DEFINE_NAME("dreloc")	},
 72	{ .id = BTRFS_UUID_TREE_OBJECTID,	DEFINE_NAME("uuid")	},
 73	{ .id = BTRFS_FREE_SPACE_TREE_OBJECTID,	DEFINE_NAME("free-space") },
 74	{ .id = BTRFS_BLOCK_GROUP_TREE_OBJECTID, DEFINE_NAME("block-group") },
 75	{ .id = BTRFS_RAID_STRIPE_TREE_OBJECTID, DEFINE_NAME("raid-stripe") },
 76	{ .id = 0,				DEFINE_NAME("tree")	},
 77};
 78
 79#undef DEFINE_LEVEL
 80#undef DEFINE_NAME
 81
 82void btrfs_set_buffer_lockdep_class(u64 objectid, struct extent_buffer *eb, int level)
 83{
 84	struct btrfs_lockdep_keyset *ks;
 85
 86	ASSERT(level < ARRAY_SIZE(ks->keys));
 87
 88	/* Find the matching keyset, id 0 is the default entry */
 89	for (ks = btrfs_lockdep_keysets; ks->id; ks++)
 90		if (ks->id == objectid)
 91			break;
 92
 93	lockdep_set_class_and_name(&eb->lock, &ks->keys[level], ks->names[level]);
 94}
 95
 96void btrfs_maybe_reset_lockdep_class(struct btrfs_root *root, struct extent_buffer *eb)
 97{
 98	if (test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
 99		btrfs_set_buffer_lockdep_class(btrfs_root_id(root),
100					       eb, btrfs_header_level(eb));
101}
102
103#endif
104
105#ifdef CONFIG_BTRFS_DEBUG
106static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner)
107{
108	eb->lock_owner = owner;
109}
110#else
111static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner) { }
112#endif
113
114/*
115 * Extent buffer locking
116 * =====================
117 *
118 * We use a rw_semaphore for tree locking, and the semantics are exactly the
119 * same:
120 *
121 * - reader/writer exclusion
122 * - writer/writer exclusion
123 * - reader/reader sharing
124 * - try-lock semantics for readers and writers
125 *
126 * The rwsem implementation does opportunistic spinning which reduces number of
127 * times the locking task needs to sleep.
128 */
129
130/*
131 * btrfs_tree_read_lock_nested - lock extent buffer for read
132 * @eb:		the eb to be locked
133 * @nest:	the nesting level to be used for lockdep
134 *
135 * This takes the read lock on the extent buffer, using the specified nesting
136 * level for lockdep purposes.
137 */
138void btrfs_tree_read_lock_nested(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
139{
140	u64 start_ns = 0;
141
142	if (trace_btrfs_tree_read_lock_enabled())
143		start_ns = ktime_get_ns();
144
145	down_read_nested(&eb->lock, nest);
146	trace_btrfs_tree_read_lock(eb, start_ns);
147}
148
149/*
150 * Try-lock for read.
151 *
152 * Return true if the rwlock has been taken, false otherwise
153 */
154bool btrfs_try_tree_read_lock(struct extent_buffer *eb)
155{
156	if (down_read_trylock(&eb->lock)) {
157		trace_btrfs_try_tree_read_lock(eb);
158		return true;
159	}
160	return false;
161}
162
163/*
164 * Release read lock.
165 */
166void btrfs_tree_read_unlock(struct extent_buffer *eb)
167{
168	trace_btrfs_tree_read_unlock(eb);
169	up_read(&eb->lock);
170}
171
172/*
173 * Lock eb for write.
174 *
175 * @eb:		the eb to lock
176 * @nest:	the nesting to use for the lock
177 *
178 * Returns with the eb->lock write locked.
179 */
180void btrfs_tree_lock_nested(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
181	__acquires(&eb->lock)
182{
183	u64 start_ns = 0;
184
185	if (trace_btrfs_tree_lock_enabled())
186		start_ns = ktime_get_ns();
187
188	down_write_nested(&eb->lock, nest);
189	btrfs_set_eb_lock_owner(eb, current->pid);
190	trace_btrfs_tree_lock(eb, start_ns);
191}
192
193/*
194 * Release the write lock.
195 */
196void btrfs_tree_unlock(struct extent_buffer *eb)
197{
198	trace_btrfs_tree_unlock(eb);
199	btrfs_set_eb_lock_owner(eb, 0);
200	up_write(&eb->lock);
201}
202
203/*
204 * This releases any locks held in the path starting at level and going all the
205 * way up to the root.
206 *
207 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
208 * cases, such as COW of the block at slot zero in the node.  This ignores
209 * those rules, and it should only be called when there are no more updates to
210 * be done higher up in the tree.
211 */
212void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
213{
214	int i;
215
216	if (path->keep_locks)
217		return;
218
219	for (i = level; i < BTRFS_MAX_LEVEL; i++) {
220		if (!path->nodes[i])
221			continue;
222		if (!path->locks[i])
223			continue;
224		btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
225		path->locks[i] = 0;
226	}
227}
228
229/*
230 * Loop around taking references on and locking the root node of the tree until
231 * we end up with a lock on the root node.
232 *
233 * Return: root extent buffer with write lock held
234 */
235struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
236{
237	struct extent_buffer *eb;
238
239	while (1) {
240		eb = btrfs_root_node(root);
241
242		btrfs_maybe_reset_lockdep_class(root, eb);
243		btrfs_tree_lock(eb);
244		if (eb == root->node)
245			break;
246		btrfs_tree_unlock(eb);
247		free_extent_buffer(eb);
248	}
249	return eb;
250}
251
252/*
253 * Loop around taking references on and locking the root node of the tree until
254 * we end up with a lock on the root node.
255 *
256 * Return: root extent buffer with read lock held
257 */
258struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
259{
260	struct extent_buffer *eb;
261
262	while (1) {
263		eb = btrfs_root_node(root);
264
265		btrfs_maybe_reset_lockdep_class(root, eb);
266		btrfs_tree_read_lock(eb);
267		if (eb == root->node)
268			break;
269		btrfs_tree_read_unlock(eb);
270		free_extent_buffer(eb);
271	}
272	return eb;
273}
274
275/*
276 * Loop around taking references on and locking the root node of the tree in
277 * nowait mode until we end up with a lock on the root node or returning to
278 * avoid blocking.
279 *
280 * Return: root extent buffer with read lock held or -EAGAIN.
281 */
282struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
283{
284	struct extent_buffer *eb;
285
286	while (1) {
287		eb = btrfs_root_node(root);
288		if (!btrfs_try_tree_read_lock(eb)) {
289			free_extent_buffer(eb);
290			return ERR_PTR(-EAGAIN);
291		}
292		if (eb == root->node)
293			break;
294		btrfs_tree_read_unlock(eb);
295		free_extent_buffer(eb);
296	}
297	return eb;
298}
299
300/*
301 * DREW locks
302 * ==========
303 *
304 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
305 * where you want to provide A-B exclusion but not AA or BB.
306 *
307 * Currently implementation gives more priority to reader. If a reader and a
308 * writer both race to acquire their respective sides of the lock the writer
309 * would yield its lock as soon as it detects a concurrent reader. Additionally
310 * if there are pending readers no new writers would be allowed to come in and
311 * acquire the lock.
312 */
313
314void btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
315{
316	atomic_set(&lock->readers, 0);
317	atomic_set(&lock->writers, 0);
318	init_waitqueue_head(&lock->pending_readers);
319	init_waitqueue_head(&lock->pending_writers);
320}
321
322/* Return true if acquisition is successful, false otherwise */
323bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
324{
325	if (atomic_read(&lock->readers))
326		return false;
327
328	atomic_inc(&lock->writers);
329
330	/* Ensure writers count is updated before we check for pending readers */
331	smp_mb__after_atomic();
332	if (atomic_read(&lock->readers)) {
333		btrfs_drew_write_unlock(lock);
334		return false;
335	}
336
337	return true;
338}
339
340void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
341{
342	while (true) {
343		if (btrfs_drew_try_write_lock(lock))
344			return;
345		wait_event(lock->pending_writers, !atomic_read(&lock->readers));
346	}
347}
348
349void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
350{
351	/*
352	 * atomic_dec_and_test() implies a full barrier, so woken up readers are
353	 * guaranteed to see the decrement.
354	 */
355	if (atomic_dec_and_test(&lock->writers))
356		wake_up(&lock->pending_readers);
357}
358
359void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
360{
361	atomic_inc(&lock->readers);
362
363	/*
364	 * Ensure the pending reader count is perceived BEFORE this reader
365	 * goes to sleep in case of active writers. This guarantees new writers
366	 * won't be allowed and that the current reader will be woken up when
367	 * the last active writer finishes its jobs.
368	 */
369	smp_mb__after_atomic();
370
371	wait_event(lock->pending_readers, atomic_read(&lock->writers) == 0);
372}
373
374void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
375{
376	/*
377	 * atomic_dec_and_test implies a full barrier, so woken up writers
378	 * are guaranteed to see the decrement
379	 */
380	if (atomic_dec_and_test(&lock->readers))
381		wake_up(&lock->pending_writers);
382}