1 | // SPDX-License-Identifier: GPL-2.0+ |
2 | /* |
3 | * linux/fs/jbd2/transaction.c |
4 | * |
5 | * Written by Stephen C. Tweedie <sct@redhat.com>, 1998 |
6 | * |
7 | * Copyright 1998 Red Hat corp --- All Rights Reserved |
8 | * |
9 | * Generic filesystem transaction handling code; part of the ext2fs |
10 | * journaling system. |
11 | * |
12 | * This file manages transactions (compound commits managed by the |
13 | * journaling code) and handles (individual atomic operations by the |
14 | * filesystem). |
15 | */ |
16 | |
17 | #include <linux/time.h> |
18 | #include <linux/fs.h> |
19 | #include <linux/jbd2.h> |
20 | #include <linux/errno.h> |
21 | #include <linux/slab.h> |
22 | #include <linux/timer.h> |
23 | #include <linux/mm.h> |
24 | #include <linux/highmem.h> |
25 | #include <linux/hrtimer.h> |
26 | #include <linux/backing-dev.h> |
27 | #include <linux/bug.h> |
28 | #include <linux/module.h> |
29 | #include <linux/sched/mm.h> |
30 | |
31 | #include <trace/events/jbd2.h> |
32 | |
33 | static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh); |
34 | static void __jbd2_journal_unfile_buffer(struct journal_head *jh); |
35 | |
36 | static struct kmem_cache *transaction_cache; |
37 | int __init jbd2_journal_init_transaction_cache(void) |
38 | { |
39 | J_ASSERT(!transaction_cache); |
40 | transaction_cache = kmem_cache_create(name: "jbd2_transaction_s" , |
41 | size: sizeof(transaction_t), |
42 | align: 0, |
43 | SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY, |
44 | NULL); |
45 | if (!transaction_cache) { |
46 | pr_emerg("JBD2: failed to create transaction cache\n" ); |
47 | return -ENOMEM; |
48 | } |
49 | return 0; |
50 | } |
51 | |
52 | void jbd2_journal_destroy_transaction_cache(void) |
53 | { |
54 | kmem_cache_destroy(s: transaction_cache); |
55 | transaction_cache = NULL; |
56 | } |
57 | |
58 | void jbd2_journal_free_transaction(transaction_t *transaction) |
59 | { |
60 | if (unlikely(ZERO_OR_NULL_PTR(transaction))) |
61 | return; |
62 | kmem_cache_free(s: transaction_cache, objp: transaction); |
63 | } |
64 | |
65 | /* |
66 | * Base amount of descriptor blocks we reserve for each transaction. |
67 | */ |
68 | static int jbd2_descriptor_blocks_per_trans(journal_t *journal) |
69 | { |
70 | int tag_space = journal->j_blocksize - sizeof(journal_header_t); |
71 | int tags_per_block; |
72 | |
73 | /* Subtract UUID */ |
74 | tag_space -= 16; |
75 | if (jbd2_journal_has_csum_v2or3(journal)) |
76 | tag_space -= sizeof(struct jbd2_journal_block_tail); |
77 | /* Commit code leaves a slack space of 16 bytes at the end of block */ |
78 | tags_per_block = (tag_space - 16) / journal_tag_bytes(journal); |
79 | /* |
80 | * Revoke descriptors are accounted separately so we need to reserve |
81 | * space for commit block and normal transaction descriptor blocks. |
82 | */ |
83 | return 1 + DIV_ROUND_UP(journal->j_max_transaction_buffers, |
84 | tags_per_block); |
85 | } |
86 | |
87 | /* |
88 | * jbd2_get_transaction: obtain a new transaction_t object. |
89 | * |
90 | * Simply initialise a new transaction. Initialize it in |
91 | * RUNNING state and add it to the current journal (which should not |
92 | * have an existing running transaction: we only make a new transaction |
93 | * once we have started to commit the old one). |
94 | * |
95 | * Preconditions: |
96 | * The journal MUST be locked. We don't perform atomic mallocs on the |
97 | * new transaction and we can't block without protecting against other |
98 | * processes trying to touch the journal while it is in transition. |
99 | * |
100 | */ |
101 | |
102 | static void jbd2_get_transaction(journal_t *journal, |
103 | transaction_t *transaction) |
104 | { |
105 | transaction->t_journal = journal; |
106 | transaction->t_state = T_RUNNING; |
107 | transaction->t_start_time = ktime_get(); |
108 | transaction->t_tid = journal->j_transaction_sequence++; |
109 | transaction->t_expires = jiffies + journal->j_commit_interval; |
110 | atomic_set(v: &transaction->t_updates, i: 0); |
111 | atomic_set(v: &transaction->t_outstanding_credits, |
112 | i: jbd2_descriptor_blocks_per_trans(journal) + |
113 | atomic_read(v: &journal->j_reserved_credits)); |
114 | atomic_set(v: &transaction->t_outstanding_revokes, i: 0); |
115 | atomic_set(v: &transaction->t_handle_count, i: 0); |
116 | INIT_LIST_HEAD(list: &transaction->t_inode_list); |
117 | INIT_LIST_HEAD(list: &transaction->t_private_list); |
118 | |
119 | /* Set up the commit timer for the new transaction. */ |
120 | journal->j_commit_timer.expires = round_jiffies_up(j: transaction->t_expires); |
121 | add_timer(timer: &journal->j_commit_timer); |
122 | |
123 | J_ASSERT(journal->j_running_transaction == NULL); |
124 | journal->j_running_transaction = transaction; |
125 | transaction->t_max_wait = 0; |
126 | transaction->t_start = jiffies; |
127 | transaction->t_requested = 0; |
128 | } |
129 | |
130 | /* |
131 | * Handle management. |
132 | * |
133 | * A handle_t is an object which represents a single atomic update to a |
134 | * filesystem, and which tracks all of the modifications which form part |
135 | * of that one update. |
136 | */ |
137 | |
138 | /* |
139 | * Update transaction's maximum wait time, if debugging is enabled. |
140 | * |
141 | * t_max_wait is carefully updated here with use of atomic compare exchange. |
142 | * Note that there could be multiplre threads trying to do this simultaneously |
143 | * hence using cmpxchg to avoid any use of locks in this case. |
144 | * With this t_max_wait can be updated w/o enabling jbd2_journal_enable_debug. |
145 | */ |
146 | static inline void update_t_max_wait(transaction_t *transaction, |
147 | unsigned long ts) |
148 | { |
149 | unsigned long oldts, newts; |
150 | |
151 | if (time_after(transaction->t_start, ts)) { |
152 | newts = jbd2_time_diff(start: ts, end: transaction->t_start); |
153 | oldts = READ_ONCE(transaction->t_max_wait); |
154 | while (oldts < newts) |
155 | oldts = cmpxchg(&transaction->t_max_wait, oldts, newts); |
156 | } |
157 | } |
158 | |
159 | /* |
160 | * Wait until running transaction passes to T_FLUSH state and new transaction |
161 | * can thus be started. Also starts the commit if needed. The function expects |
162 | * running transaction to exist and releases j_state_lock. |
163 | */ |
164 | static void wait_transaction_locked(journal_t *journal) |
165 | __releases(journal->j_state_lock) |
166 | { |
167 | DEFINE_WAIT(wait); |
168 | int need_to_start; |
169 | tid_t tid = journal->j_running_transaction->t_tid; |
170 | |
171 | prepare_to_wait_exclusive(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait, |
172 | TASK_UNINTERRUPTIBLE); |
173 | need_to_start = !tid_geq(x: journal->j_commit_request, y: tid); |
174 | read_unlock(&journal->j_state_lock); |
175 | if (need_to_start) |
176 | jbd2_log_start_commit(journal, tid); |
177 | jbd2_might_wait_for_commit(journal); |
178 | schedule(); |
179 | finish_wait(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait); |
180 | } |
181 | |
182 | /* |
183 | * Wait until running transaction transitions from T_SWITCH to T_FLUSH |
184 | * state and new transaction can thus be started. The function releases |
185 | * j_state_lock. |
186 | */ |
187 | static void wait_transaction_switching(journal_t *journal) |
188 | __releases(journal->j_state_lock) |
189 | { |
190 | DEFINE_WAIT(wait); |
191 | |
192 | if (WARN_ON(!journal->j_running_transaction || |
193 | journal->j_running_transaction->t_state != T_SWITCH)) { |
194 | read_unlock(&journal->j_state_lock); |
195 | return; |
196 | } |
197 | prepare_to_wait_exclusive(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait, |
198 | TASK_UNINTERRUPTIBLE); |
199 | read_unlock(&journal->j_state_lock); |
200 | /* |
201 | * We don't call jbd2_might_wait_for_commit() here as there's no |
202 | * waiting for outstanding handles happening anymore in T_SWITCH state |
203 | * and handling of reserved handles actually relies on that for |
204 | * correctness. |
205 | */ |
206 | schedule(); |
207 | finish_wait(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait); |
208 | } |
209 | |
210 | static void sub_reserved_credits(journal_t *journal, int blocks) |
211 | { |
212 | atomic_sub(i: blocks, v: &journal->j_reserved_credits); |
213 | wake_up(&journal->j_wait_reserved); |
214 | } |
215 | |
216 | /* |
217 | * Wait until we can add credits for handle to the running transaction. Called |
218 | * with j_state_lock held for reading. Returns 0 if handle joined the running |
219 | * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and |
220 | * caller must retry. |
221 | * |
222 | * Note: because j_state_lock may be dropped depending on the return |
223 | * value, we need to fake out sparse so ti doesn't complain about a |
224 | * locking imbalance. Callers of add_transaction_credits will need to |
225 | * make a similar accomodation. |
226 | */ |
227 | static int add_transaction_credits(journal_t *journal, int blocks, |
228 | int rsv_blocks) |
229 | __must_hold(&journal->j_state_lock) |
230 | { |
231 | transaction_t *t = journal->j_running_transaction; |
232 | int needed; |
233 | int total = blocks + rsv_blocks; |
234 | |
235 | /* |
236 | * If the current transaction is locked down for commit, wait |
237 | * for the lock to be released. |
238 | */ |
239 | if (t->t_state != T_RUNNING) { |
240 | WARN_ON_ONCE(t->t_state >= T_FLUSH); |
241 | wait_transaction_locked(journal); |
242 | __acquire(&journal->j_state_lock); /* fake out sparse */ |
243 | return 1; |
244 | } |
245 | |
246 | /* |
247 | * If there is not enough space left in the log to write all |
248 | * potential buffers requested by this operation, we need to |
249 | * stall pending a log checkpoint to free some more log space. |
250 | */ |
251 | needed = atomic_add_return(i: total, v: &t->t_outstanding_credits); |
252 | if (needed > journal->j_max_transaction_buffers) { |
253 | /* |
254 | * If the current transaction is already too large, |
255 | * then start to commit it: we can then go back and |
256 | * attach this handle to a new transaction. |
257 | */ |
258 | atomic_sub(i: total, v: &t->t_outstanding_credits); |
259 | |
260 | /* |
261 | * Is the number of reserved credits in the current transaction too |
262 | * big to fit this handle? Wait until reserved credits are freed. |
263 | */ |
264 | if (atomic_read(v: &journal->j_reserved_credits) + total > |
265 | journal->j_max_transaction_buffers) { |
266 | read_unlock(&journal->j_state_lock); |
267 | jbd2_might_wait_for_commit(journal); |
268 | wait_event(journal->j_wait_reserved, |
269 | atomic_read(&journal->j_reserved_credits) + total <= |
270 | journal->j_max_transaction_buffers); |
271 | __acquire(&journal->j_state_lock); /* fake out sparse */ |
272 | return 1; |
273 | } |
274 | |
275 | wait_transaction_locked(journal); |
276 | __acquire(&journal->j_state_lock); /* fake out sparse */ |
277 | return 1; |
278 | } |
279 | |
280 | /* |
281 | * The commit code assumes that it can get enough log space |
282 | * without forcing a checkpoint. This is *critical* for |
283 | * correctness: a checkpoint of a buffer which is also |
284 | * associated with a committing transaction creates a deadlock, |
285 | * so commit simply cannot force through checkpoints. |
286 | * |
287 | * We must therefore ensure the necessary space in the journal |
288 | * *before* starting to dirty potentially checkpointed buffers |
289 | * in the new transaction. |
290 | */ |
291 | if (jbd2_log_space_left(journal) < journal->j_max_transaction_buffers) { |
292 | atomic_sub(i: total, v: &t->t_outstanding_credits); |
293 | read_unlock(&journal->j_state_lock); |
294 | jbd2_might_wait_for_commit(journal); |
295 | write_lock(&journal->j_state_lock); |
296 | if (jbd2_log_space_left(journal) < |
297 | journal->j_max_transaction_buffers) |
298 | __jbd2_log_wait_for_space(journal); |
299 | write_unlock(&journal->j_state_lock); |
300 | __acquire(&journal->j_state_lock); /* fake out sparse */ |
301 | return 1; |
302 | } |
303 | |
304 | /* No reservation? We are done... */ |
305 | if (!rsv_blocks) |
306 | return 0; |
307 | |
308 | needed = atomic_add_return(i: rsv_blocks, v: &journal->j_reserved_credits); |
309 | /* We allow at most half of a transaction to be reserved */ |
310 | if (needed > journal->j_max_transaction_buffers / 2) { |
311 | sub_reserved_credits(journal, blocks: rsv_blocks); |
312 | atomic_sub(i: total, v: &t->t_outstanding_credits); |
313 | read_unlock(&journal->j_state_lock); |
314 | jbd2_might_wait_for_commit(journal); |
315 | wait_event(journal->j_wait_reserved, |
316 | atomic_read(&journal->j_reserved_credits) + rsv_blocks |
317 | <= journal->j_max_transaction_buffers / 2); |
318 | __acquire(&journal->j_state_lock); /* fake out sparse */ |
319 | return 1; |
320 | } |
321 | return 0; |
322 | } |
323 | |
324 | /* |
325 | * start_this_handle: Given a handle, deal with any locking or stalling |
326 | * needed to make sure that there is enough journal space for the handle |
327 | * to begin. Attach the handle to a transaction and set up the |
328 | * transaction's buffer credits. |
329 | */ |
330 | |
331 | static int start_this_handle(journal_t *journal, handle_t *handle, |
332 | gfp_t gfp_mask) |
333 | { |
334 | transaction_t *transaction, *new_transaction = NULL; |
335 | int blocks = handle->h_total_credits; |
336 | int rsv_blocks = 0; |
337 | unsigned long ts = jiffies; |
338 | |
339 | if (handle->h_rsv_handle) |
340 | rsv_blocks = handle->h_rsv_handle->h_total_credits; |
341 | |
342 | /* |
343 | * Limit the number of reserved credits to 1/2 of maximum transaction |
344 | * size and limit the number of total credits to not exceed maximum |
345 | * transaction size per operation. |
346 | */ |
347 | if ((rsv_blocks > journal->j_max_transaction_buffers / 2) || |
348 | (rsv_blocks + blocks > journal->j_max_transaction_buffers)) { |
349 | printk(KERN_ERR "JBD2: %s wants too many credits " |
350 | "credits:%d rsv_credits:%d max:%d\n" , |
351 | current->comm, blocks, rsv_blocks, |
352 | journal->j_max_transaction_buffers); |
353 | WARN_ON(1); |
354 | return -ENOSPC; |
355 | } |
356 | |
357 | alloc_transaction: |
358 | /* |
359 | * This check is racy but it is just an optimization of allocating new |
360 | * transaction early if there are high chances we'll need it. If we |
361 | * guess wrong, we'll retry or free unused transaction. |
362 | */ |
363 | if (!data_race(journal->j_running_transaction)) { |
364 | /* |
365 | * If __GFP_FS is not present, then we may be being called from |
366 | * inside the fs writeback layer, so we MUST NOT fail. |
367 | */ |
368 | if ((gfp_mask & __GFP_FS) == 0) |
369 | gfp_mask |= __GFP_NOFAIL; |
370 | new_transaction = kmem_cache_zalloc(k: transaction_cache, |
371 | flags: gfp_mask); |
372 | if (!new_transaction) |
373 | return -ENOMEM; |
374 | } |
375 | |
376 | jbd2_debug(3, "New handle %p going live.\n" , handle); |
377 | |
378 | /* |
379 | * We need to hold j_state_lock until t_updates has been incremented, |
380 | * for proper journal barrier handling |
381 | */ |
382 | repeat: |
383 | read_lock(&journal->j_state_lock); |
384 | BUG_ON(journal->j_flags & JBD2_UNMOUNT); |
385 | if (is_journal_aborted(journal) || |
386 | (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) { |
387 | read_unlock(&journal->j_state_lock); |
388 | jbd2_journal_free_transaction(transaction: new_transaction); |
389 | return -EROFS; |
390 | } |
391 | |
392 | /* |
393 | * Wait on the journal's transaction barrier if necessary. Specifically |
394 | * we allow reserved handles to proceed because otherwise commit could |
395 | * deadlock on page writeback not being able to complete. |
396 | */ |
397 | if (!handle->h_reserved && journal->j_barrier_count) { |
398 | read_unlock(&journal->j_state_lock); |
399 | wait_event(journal->j_wait_transaction_locked, |
400 | journal->j_barrier_count == 0); |
401 | goto repeat; |
402 | } |
403 | |
404 | if (!journal->j_running_transaction) { |
405 | read_unlock(&journal->j_state_lock); |
406 | if (!new_transaction) |
407 | goto alloc_transaction; |
408 | write_lock(&journal->j_state_lock); |
409 | if (!journal->j_running_transaction && |
410 | (handle->h_reserved || !journal->j_barrier_count)) { |
411 | jbd2_get_transaction(journal, transaction: new_transaction); |
412 | new_transaction = NULL; |
413 | } |
414 | write_unlock(&journal->j_state_lock); |
415 | goto repeat; |
416 | } |
417 | |
418 | transaction = journal->j_running_transaction; |
419 | |
420 | if (!handle->h_reserved) { |
421 | /* We may have dropped j_state_lock - restart in that case */ |
422 | if (add_transaction_credits(journal, blocks, rsv_blocks)) { |
423 | /* |
424 | * add_transaction_credits releases |
425 | * j_state_lock on a non-zero return |
426 | */ |
427 | __release(&journal->j_state_lock); |
428 | goto repeat; |
429 | } |
430 | } else { |
431 | /* |
432 | * We have handle reserved so we are allowed to join T_LOCKED |
433 | * transaction and we don't have to check for transaction size |
434 | * and journal space. But we still have to wait while running |
435 | * transaction is being switched to a committing one as it |
436 | * won't wait for any handles anymore. |
437 | */ |
438 | if (transaction->t_state == T_SWITCH) { |
439 | wait_transaction_switching(journal); |
440 | goto repeat; |
441 | } |
442 | sub_reserved_credits(journal, blocks); |
443 | handle->h_reserved = 0; |
444 | } |
445 | |
446 | /* OK, account for the buffers that this operation expects to |
447 | * use and add the handle to the running transaction. |
448 | */ |
449 | update_t_max_wait(transaction, ts); |
450 | handle->h_transaction = transaction; |
451 | handle->h_requested_credits = blocks; |
452 | handle->h_revoke_credits_requested = handle->h_revoke_credits; |
453 | handle->h_start_jiffies = jiffies; |
454 | atomic_inc(v: &transaction->t_updates); |
455 | atomic_inc(v: &transaction->t_handle_count); |
456 | jbd2_debug(4, "Handle %p given %d credits (total %d, free %lu)\n" , |
457 | handle, blocks, |
458 | atomic_read(&transaction->t_outstanding_credits), |
459 | jbd2_log_space_left(journal)); |
460 | read_unlock(&journal->j_state_lock); |
461 | current->journal_info = handle; |
462 | |
463 | rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_); |
464 | jbd2_journal_free_transaction(transaction: new_transaction); |
465 | /* |
466 | * Ensure that no allocations done while the transaction is open are |
467 | * going to recurse back to the fs layer. |
468 | */ |
469 | handle->saved_alloc_context = memalloc_nofs_save(); |
470 | return 0; |
471 | } |
472 | |
473 | /* Allocate a new handle. This should probably be in a slab... */ |
474 | static handle_t *new_handle(int nblocks) |
475 | { |
476 | handle_t *handle = jbd2_alloc_handle(GFP_NOFS); |
477 | if (!handle) |
478 | return NULL; |
479 | handle->h_total_credits = nblocks; |
480 | handle->h_ref = 1; |
481 | |
482 | return handle; |
483 | } |
484 | |
485 | handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks, |
486 | int revoke_records, gfp_t gfp_mask, |
487 | unsigned int type, unsigned int line_no) |
488 | { |
489 | handle_t *handle = journal_current_handle(); |
490 | int err; |
491 | |
492 | if (!journal) |
493 | return ERR_PTR(error: -EROFS); |
494 | |
495 | if (handle) { |
496 | J_ASSERT(handle->h_transaction->t_journal == journal); |
497 | handle->h_ref++; |
498 | return handle; |
499 | } |
500 | |
501 | nblocks += DIV_ROUND_UP(revoke_records, |
502 | journal->j_revoke_records_per_block); |
503 | handle = new_handle(nblocks); |
504 | if (!handle) |
505 | return ERR_PTR(error: -ENOMEM); |
506 | if (rsv_blocks) { |
507 | handle_t *rsv_handle; |
508 | |
509 | rsv_handle = new_handle(nblocks: rsv_blocks); |
510 | if (!rsv_handle) { |
511 | jbd2_free_handle(handle); |
512 | return ERR_PTR(error: -ENOMEM); |
513 | } |
514 | rsv_handle->h_reserved = 1; |
515 | rsv_handle->h_journal = journal; |
516 | handle->h_rsv_handle = rsv_handle; |
517 | } |
518 | handle->h_revoke_credits = revoke_records; |
519 | |
520 | err = start_this_handle(journal, handle, gfp_mask); |
521 | if (err < 0) { |
522 | if (handle->h_rsv_handle) |
523 | jbd2_free_handle(handle: handle->h_rsv_handle); |
524 | jbd2_free_handle(handle); |
525 | return ERR_PTR(error: err); |
526 | } |
527 | handle->h_type = type; |
528 | handle->h_line_no = line_no; |
529 | trace_jbd2_handle_start(dev: journal->j_fs_dev->bd_dev, |
530 | tid: handle->h_transaction->t_tid, type, |
531 | line_no, requested_blocks: nblocks); |
532 | |
533 | return handle; |
534 | } |
535 | EXPORT_SYMBOL(jbd2__journal_start); |
536 | |
537 | |
538 | /** |
539 | * jbd2_journal_start() - Obtain a new handle. |
540 | * @journal: Journal to start transaction on. |
541 | * @nblocks: number of block buffer we might modify |
542 | * |
543 | * We make sure that the transaction can guarantee at least nblocks of |
544 | * modified buffers in the log. We block until the log can guarantee |
545 | * that much space. Additionally, if rsv_blocks > 0, we also create another |
546 | * handle with rsv_blocks reserved blocks in the journal. This handle is |
547 | * stored in h_rsv_handle. It is not attached to any particular transaction |
548 | * and thus doesn't block transaction commit. If the caller uses this reserved |
549 | * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop() |
550 | * on the parent handle will dispose the reserved one. Reserved handle has to |
551 | * be converted to a normal handle using jbd2_journal_start_reserved() before |
552 | * it can be used. |
553 | * |
554 | * Return a pointer to a newly allocated handle, or an ERR_PTR() value |
555 | * on failure. |
556 | */ |
557 | handle_t *jbd2_journal_start(journal_t *journal, int nblocks) |
558 | { |
559 | return jbd2__journal_start(journal, nblocks, 0, 0, GFP_NOFS, 0, 0); |
560 | } |
561 | EXPORT_SYMBOL(jbd2_journal_start); |
562 | |
563 | static void __jbd2_journal_unreserve_handle(handle_t *handle, transaction_t *t) |
564 | { |
565 | journal_t *journal = handle->h_journal; |
566 | |
567 | WARN_ON(!handle->h_reserved); |
568 | sub_reserved_credits(journal, blocks: handle->h_total_credits); |
569 | if (t) |
570 | atomic_sub(i: handle->h_total_credits, v: &t->t_outstanding_credits); |
571 | } |
572 | |
573 | void jbd2_journal_free_reserved(handle_t *handle) |
574 | { |
575 | journal_t *journal = handle->h_journal; |
576 | |
577 | /* Get j_state_lock to pin running transaction if it exists */ |
578 | read_lock(&journal->j_state_lock); |
579 | __jbd2_journal_unreserve_handle(handle, t: journal->j_running_transaction); |
580 | read_unlock(&journal->j_state_lock); |
581 | jbd2_free_handle(handle); |
582 | } |
583 | EXPORT_SYMBOL(jbd2_journal_free_reserved); |
584 | |
585 | /** |
586 | * jbd2_journal_start_reserved() - start reserved handle |
587 | * @handle: handle to start |
588 | * @type: for handle statistics |
589 | * @line_no: for handle statistics |
590 | * |
591 | * Start handle that has been previously reserved with jbd2_journal_reserve(). |
592 | * This attaches @handle to the running transaction (or creates one if there's |
593 | * not transaction running). Unlike jbd2_journal_start() this function cannot |
594 | * block on journal commit, checkpointing, or similar stuff. It can block on |
595 | * memory allocation or frozen journal though. |
596 | * |
597 | * Return 0 on success, non-zero on error - handle is freed in that case. |
598 | */ |
599 | int jbd2_journal_start_reserved(handle_t *handle, unsigned int type, |
600 | unsigned int line_no) |
601 | { |
602 | journal_t *journal = handle->h_journal; |
603 | int ret = -EIO; |
604 | |
605 | if (WARN_ON(!handle->h_reserved)) { |
606 | /* Someone passed in normal handle? Just stop it. */ |
607 | jbd2_journal_stop(handle); |
608 | return ret; |
609 | } |
610 | /* |
611 | * Usefulness of mixing of reserved and unreserved handles is |
612 | * questionable. So far nobody seems to need it so just error out. |
613 | */ |
614 | if (WARN_ON(current->journal_info)) { |
615 | jbd2_journal_free_reserved(handle); |
616 | return ret; |
617 | } |
618 | |
619 | handle->h_journal = NULL; |
620 | /* |
621 | * GFP_NOFS is here because callers are likely from writeback or |
622 | * similarly constrained call sites |
623 | */ |
624 | ret = start_this_handle(journal, handle, GFP_NOFS); |
625 | if (ret < 0) { |
626 | handle->h_journal = journal; |
627 | jbd2_journal_free_reserved(handle); |
628 | return ret; |
629 | } |
630 | handle->h_type = type; |
631 | handle->h_line_no = line_no; |
632 | trace_jbd2_handle_start(dev: journal->j_fs_dev->bd_dev, |
633 | tid: handle->h_transaction->t_tid, type, |
634 | line_no, requested_blocks: handle->h_total_credits); |
635 | return 0; |
636 | } |
637 | EXPORT_SYMBOL(jbd2_journal_start_reserved); |
638 | |
639 | /** |
640 | * jbd2_journal_extend() - extend buffer credits. |
641 | * @handle: handle to 'extend' |
642 | * @nblocks: nr blocks to try to extend by. |
643 | * @revoke_records: number of revoke records to try to extend by. |
644 | * |
645 | * Some transactions, such as large extends and truncates, can be done |
646 | * atomically all at once or in several stages. The operation requests |
647 | * a credit for a number of buffer modifications in advance, but can |
648 | * extend its credit if it needs more. |
649 | * |
650 | * jbd2_journal_extend tries to give the running handle more buffer credits. |
651 | * It does not guarantee that allocation - this is a best-effort only. |
652 | * The calling process MUST be able to deal cleanly with a failure to |
653 | * extend here. |
654 | * |
655 | * Return 0 on success, non-zero on failure. |
656 | * |
657 | * return code < 0 implies an error |
658 | * return code > 0 implies normal transaction-full status. |
659 | */ |
660 | int jbd2_journal_extend(handle_t *handle, int nblocks, int revoke_records) |
661 | { |
662 | transaction_t *transaction = handle->h_transaction; |
663 | journal_t *journal; |
664 | int result; |
665 | int wanted; |
666 | |
667 | if (is_handle_aborted(handle)) |
668 | return -EROFS; |
669 | journal = transaction->t_journal; |
670 | |
671 | result = 1; |
672 | |
673 | read_lock(&journal->j_state_lock); |
674 | |
675 | /* Don't extend a locked-down transaction! */ |
676 | if (transaction->t_state != T_RUNNING) { |
677 | jbd2_debug(3, "denied handle %p %d blocks: " |
678 | "transaction not running\n" , handle, nblocks); |
679 | goto error_out; |
680 | } |
681 | |
682 | nblocks += DIV_ROUND_UP( |
683 | handle->h_revoke_credits_requested + revoke_records, |
684 | journal->j_revoke_records_per_block) - |
685 | DIV_ROUND_UP( |
686 | handle->h_revoke_credits_requested, |
687 | journal->j_revoke_records_per_block); |
688 | wanted = atomic_add_return(i: nblocks, |
689 | v: &transaction->t_outstanding_credits); |
690 | |
691 | if (wanted > journal->j_max_transaction_buffers) { |
692 | jbd2_debug(3, "denied handle %p %d blocks: " |
693 | "transaction too large\n" , handle, nblocks); |
694 | atomic_sub(i: nblocks, v: &transaction->t_outstanding_credits); |
695 | goto error_out; |
696 | } |
697 | |
698 | trace_jbd2_handle_extend(dev: journal->j_fs_dev->bd_dev, |
699 | tid: transaction->t_tid, |
700 | type: handle->h_type, line_no: handle->h_line_no, |
701 | buffer_credits: handle->h_total_credits, |
702 | requested_blocks: nblocks); |
703 | |
704 | handle->h_total_credits += nblocks; |
705 | handle->h_requested_credits += nblocks; |
706 | handle->h_revoke_credits += revoke_records; |
707 | handle->h_revoke_credits_requested += revoke_records; |
708 | result = 0; |
709 | |
710 | jbd2_debug(3, "extended handle %p by %d\n" , handle, nblocks); |
711 | error_out: |
712 | read_unlock(&journal->j_state_lock); |
713 | return result; |
714 | } |
715 | |
716 | static void stop_this_handle(handle_t *handle) |
717 | { |
718 | transaction_t *transaction = handle->h_transaction; |
719 | journal_t *journal = transaction->t_journal; |
720 | int revokes; |
721 | |
722 | J_ASSERT(journal_current_handle() == handle); |
723 | J_ASSERT(atomic_read(&transaction->t_updates) > 0); |
724 | current->journal_info = NULL; |
725 | /* |
726 | * Subtract necessary revoke descriptor blocks from handle credits. We |
727 | * take care to account only for revoke descriptor blocks the |
728 | * transaction will really need as large sequences of transactions with |
729 | * small numbers of revokes are relatively common. |
730 | */ |
731 | revokes = handle->h_revoke_credits_requested - handle->h_revoke_credits; |
732 | if (revokes) { |
733 | int t_revokes, revoke_descriptors; |
734 | int rr_per_blk = journal->j_revoke_records_per_block; |
735 | |
736 | WARN_ON_ONCE(DIV_ROUND_UP(revokes, rr_per_blk) |
737 | > handle->h_total_credits); |
738 | t_revokes = atomic_add_return(i: revokes, |
739 | v: &transaction->t_outstanding_revokes); |
740 | revoke_descriptors = |
741 | DIV_ROUND_UP(t_revokes, rr_per_blk) - |
742 | DIV_ROUND_UP(t_revokes - revokes, rr_per_blk); |
743 | handle->h_total_credits -= revoke_descriptors; |
744 | } |
745 | atomic_sub(i: handle->h_total_credits, |
746 | v: &transaction->t_outstanding_credits); |
747 | if (handle->h_rsv_handle) |
748 | __jbd2_journal_unreserve_handle(handle: handle->h_rsv_handle, |
749 | t: transaction); |
750 | if (atomic_dec_and_test(v: &transaction->t_updates)) |
751 | wake_up(&journal->j_wait_updates); |
752 | |
753 | rwsem_release(&journal->j_trans_commit_map, _THIS_IP_); |
754 | /* |
755 | * Scope of the GFP_NOFS context is over here and so we can restore the |
756 | * original alloc context. |
757 | */ |
758 | memalloc_nofs_restore(flags: handle->saved_alloc_context); |
759 | } |
760 | |
761 | /** |
762 | * jbd2__journal_restart() - restart a handle . |
763 | * @handle: handle to restart |
764 | * @nblocks: nr credits requested |
765 | * @revoke_records: number of revoke record credits requested |
766 | * @gfp_mask: memory allocation flags (for start_this_handle) |
767 | * |
768 | * Restart a handle for a multi-transaction filesystem |
769 | * operation. |
770 | * |
771 | * If the jbd2_journal_extend() call above fails to grant new buffer credits |
772 | * to a running handle, a call to jbd2_journal_restart will commit the |
773 | * handle's transaction so far and reattach the handle to a new |
774 | * transaction capable of guaranteeing the requested number of |
775 | * credits. We preserve reserved handle if there's any attached to the |
776 | * passed in handle. |
777 | */ |
778 | int jbd2__journal_restart(handle_t *handle, int nblocks, int revoke_records, |
779 | gfp_t gfp_mask) |
780 | { |
781 | transaction_t *transaction = handle->h_transaction; |
782 | journal_t *journal; |
783 | tid_t tid; |
784 | int need_to_start; |
785 | int ret; |
786 | |
787 | /* If we've had an abort of any type, don't even think about |
788 | * actually doing the restart! */ |
789 | if (is_handle_aborted(handle)) |
790 | return 0; |
791 | journal = transaction->t_journal; |
792 | tid = transaction->t_tid; |
793 | |
794 | /* |
795 | * First unlink the handle from its current transaction, and start the |
796 | * commit on that. |
797 | */ |
798 | jbd2_debug(2, "restarting handle %p\n" , handle); |
799 | stop_this_handle(handle); |
800 | handle->h_transaction = NULL; |
801 | |
802 | /* |
803 | * TODO: If we use READ_ONCE / WRITE_ONCE for j_commit_request we can |
804 | * get rid of pointless j_state_lock traffic like this. |
805 | */ |
806 | read_lock(&journal->j_state_lock); |
807 | need_to_start = !tid_geq(x: journal->j_commit_request, y: tid); |
808 | read_unlock(&journal->j_state_lock); |
809 | if (need_to_start) |
810 | jbd2_log_start_commit(journal, tid); |
811 | handle->h_total_credits = nblocks + |
812 | DIV_ROUND_UP(revoke_records, |
813 | journal->j_revoke_records_per_block); |
814 | handle->h_revoke_credits = revoke_records; |
815 | ret = start_this_handle(journal, handle, gfp_mask); |
816 | trace_jbd2_handle_restart(dev: journal->j_fs_dev->bd_dev, |
817 | tid: ret ? 0 : handle->h_transaction->t_tid, |
818 | type: handle->h_type, line_no: handle->h_line_no, |
819 | requested_blocks: handle->h_total_credits); |
820 | return ret; |
821 | } |
822 | EXPORT_SYMBOL(jbd2__journal_restart); |
823 | |
824 | |
825 | int jbd2_journal_restart(handle_t *handle, int nblocks) |
826 | { |
827 | return jbd2__journal_restart(handle, nblocks, 0, GFP_NOFS); |
828 | } |
829 | EXPORT_SYMBOL(jbd2_journal_restart); |
830 | |
831 | /* |
832 | * Waits for any outstanding t_updates to finish. |
833 | * This is called with write j_state_lock held. |
834 | */ |
835 | void jbd2_journal_wait_updates(journal_t *journal) |
836 | { |
837 | DEFINE_WAIT(wait); |
838 | |
839 | while (1) { |
840 | /* |
841 | * Note that the running transaction can get freed under us if |
842 | * this transaction is getting committed in |
843 | * jbd2_journal_commit_transaction() -> |
844 | * jbd2_journal_free_transaction(). This can only happen when we |
845 | * release j_state_lock -> schedule() -> acquire j_state_lock. |
846 | * Hence we should everytime retrieve new j_running_transaction |
847 | * value (after j_state_lock release acquire cycle), else it may |
848 | * lead to use-after-free of old freed transaction. |
849 | */ |
850 | transaction_t *transaction = journal->j_running_transaction; |
851 | |
852 | if (!transaction) |
853 | break; |
854 | |
855 | prepare_to_wait(wq_head: &journal->j_wait_updates, wq_entry: &wait, |
856 | TASK_UNINTERRUPTIBLE); |
857 | if (!atomic_read(v: &transaction->t_updates)) { |
858 | finish_wait(wq_head: &journal->j_wait_updates, wq_entry: &wait); |
859 | break; |
860 | } |
861 | write_unlock(&journal->j_state_lock); |
862 | schedule(); |
863 | finish_wait(wq_head: &journal->j_wait_updates, wq_entry: &wait); |
864 | write_lock(&journal->j_state_lock); |
865 | } |
866 | } |
867 | |
868 | /** |
869 | * jbd2_journal_lock_updates () - establish a transaction barrier. |
870 | * @journal: Journal to establish a barrier on. |
871 | * |
872 | * This locks out any further updates from being started, and blocks |
873 | * until all existing updates have completed, returning only once the |
874 | * journal is in a quiescent state with no updates running. |
875 | * |
876 | * The journal lock should not be held on entry. |
877 | */ |
878 | void jbd2_journal_lock_updates(journal_t *journal) |
879 | { |
880 | jbd2_might_wait_for_commit(journal); |
881 | |
882 | write_lock(&journal->j_state_lock); |
883 | ++journal->j_barrier_count; |
884 | |
885 | /* Wait until there are no reserved handles */ |
886 | if (atomic_read(v: &journal->j_reserved_credits)) { |
887 | write_unlock(&journal->j_state_lock); |
888 | wait_event(journal->j_wait_reserved, |
889 | atomic_read(&journal->j_reserved_credits) == 0); |
890 | write_lock(&journal->j_state_lock); |
891 | } |
892 | |
893 | /* Wait until there are no running t_updates */ |
894 | jbd2_journal_wait_updates(journal); |
895 | |
896 | write_unlock(&journal->j_state_lock); |
897 | |
898 | /* |
899 | * We have now established a barrier against other normal updates, but |
900 | * we also need to barrier against other jbd2_journal_lock_updates() calls |
901 | * to make sure that we serialise special journal-locked operations |
902 | * too. |
903 | */ |
904 | mutex_lock(&journal->j_barrier); |
905 | } |
906 | |
907 | /** |
908 | * jbd2_journal_unlock_updates () - release barrier |
909 | * @journal: Journal to release the barrier on. |
910 | * |
911 | * Release a transaction barrier obtained with jbd2_journal_lock_updates(). |
912 | * |
913 | * Should be called without the journal lock held. |
914 | */ |
915 | void jbd2_journal_unlock_updates (journal_t *journal) |
916 | { |
917 | J_ASSERT(journal->j_barrier_count != 0); |
918 | |
919 | mutex_unlock(lock: &journal->j_barrier); |
920 | write_lock(&journal->j_state_lock); |
921 | --journal->j_barrier_count; |
922 | write_unlock(&journal->j_state_lock); |
923 | wake_up_all(&journal->j_wait_transaction_locked); |
924 | } |
925 | |
926 | static void warn_dirty_buffer(struct buffer_head *bh) |
927 | { |
928 | printk(KERN_WARNING |
929 | "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). " |
930 | "There's a risk of filesystem corruption in case of system " |
931 | "crash.\n" , |
932 | bh->b_bdev, (unsigned long long)bh->b_blocknr); |
933 | } |
934 | |
935 | /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */ |
936 | static void jbd2_freeze_jh_data(struct journal_head *jh) |
937 | { |
938 | char *source; |
939 | struct buffer_head *bh = jh2bh(jh); |
940 | |
941 | J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n" ); |
942 | source = kmap_local_folio(folio: bh->b_folio, offset: bh_offset(bh)); |
943 | /* Fire data frozen trigger just before we copy the data */ |
944 | jbd2_buffer_frozen_trigger(jh, mapped_data: source, triggers: jh->b_triggers); |
945 | memcpy(jh->b_frozen_data, source, bh->b_size); |
946 | kunmap_local(source); |
947 | |
948 | /* |
949 | * Now that the frozen data is saved off, we need to store any matching |
950 | * triggers. |
951 | */ |
952 | jh->b_frozen_triggers = jh->b_triggers; |
953 | } |
954 | |
955 | /* |
956 | * If the buffer is already part of the current transaction, then there |
957 | * is nothing we need to do. If it is already part of a prior |
958 | * transaction which we are still committing to disk, then we need to |
959 | * make sure that we do not overwrite the old copy: we do copy-out to |
960 | * preserve the copy going to disk. We also account the buffer against |
961 | * the handle's metadata buffer credits (unless the buffer is already |
962 | * part of the transaction, that is). |
963 | * |
964 | */ |
965 | static int |
966 | do_get_write_access(handle_t *handle, struct journal_head *jh, |
967 | int force_copy) |
968 | { |
969 | struct buffer_head *bh; |
970 | transaction_t *transaction = handle->h_transaction; |
971 | journal_t *journal; |
972 | int error; |
973 | char *frozen_buffer = NULL; |
974 | unsigned long start_lock, time_lock; |
975 | |
976 | journal = transaction->t_journal; |
977 | |
978 | jbd2_debug(5, "journal_head %p, force_copy %d\n" , jh, force_copy); |
979 | |
980 | JBUFFER_TRACE(jh, "entry" ); |
981 | repeat: |
982 | bh = jh2bh(jh); |
983 | |
984 | /* @@@ Need to check for errors here at some point. */ |
985 | |
986 | start_lock = jiffies; |
987 | lock_buffer(bh); |
988 | spin_lock(lock: &jh->b_state_lock); |
989 | |
990 | /* If it takes too long to lock the buffer, trace it */ |
991 | time_lock = jbd2_time_diff(start: start_lock, end: jiffies); |
992 | if (time_lock > HZ/10) |
993 | trace_jbd2_lock_buffer_stall(dev: bh->b_bdev->bd_dev, |
994 | stall_ms: jiffies_to_msecs(j: time_lock)); |
995 | |
996 | /* We now hold the buffer lock so it is safe to query the buffer |
997 | * state. Is the buffer dirty? |
998 | * |
999 | * If so, there are two possibilities. The buffer may be |
1000 | * non-journaled, and undergoing a quite legitimate writeback. |
1001 | * Otherwise, it is journaled, and we don't expect dirty buffers |
1002 | * in that state (the buffers should be marked JBD_Dirty |
1003 | * instead.) So either the IO is being done under our own |
1004 | * control and this is a bug, or it's a third party IO such as |
1005 | * dump(8) (which may leave the buffer scheduled for read --- |
1006 | * ie. locked but not dirty) or tune2fs (which may actually have |
1007 | * the buffer dirtied, ugh.) */ |
1008 | |
1009 | if (buffer_dirty(bh) && jh->b_transaction) { |
1010 | warn_dirty_buffer(bh); |
1011 | /* |
1012 | * We need to clean the dirty flag and we must do it under the |
1013 | * buffer lock to be sure we don't race with running write-out. |
1014 | */ |
1015 | JBUFFER_TRACE(jh, "Journalling dirty buffer" ); |
1016 | clear_buffer_dirty(bh); |
1017 | /* |
1018 | * The buffer is going to be added to BJ_Reserved list now and |
1019 | * nothing guarantees jbd2_journal_dirty_metadata() will be |
1020 | * ever called for it. So we need to set jbddirty bit here to |
1021 | * make sure the buffer is dirtied and written out when the |
1022 | * journaling machinery is done with it. |
1023 | */ |
1024 | set_buffer_jbddirty(bh); |
1025 | } |
1026 | |
1027 | error = -EROFS; |
1028 | if (is_handle_aborted(handle)) { |
1029 | spin_unlock(lock: &jh->b_state_lock); |
1030 | unlock_buffer(bh); |
1031 | goto out; |
1032 | } |
1033 | error = 0; |
1034 | |
1035 | /* |
1036 | * The buffer is already part of this transaction if b_transaction or |
1037 | * b_next_transaction points to it |
1038 | */ |
1039 | if (jh->b_transaction == transaction || |
1040 | jh->b_next_transaction == transaction) { |
1041 | unlock_buffer(bh); |
1042 | goto done; |
1043 | } |
1044 | |
1045 | /* |
1046 | * this is the first time this transaction is touching this buffer, |
1047 | * reset the modified flag |
1048 | */ |
1049 | jh->b_modified = 0; |
1050 | |
1051 | /* |
1052 | * If the buffer is not journaled right now, we need to make sure it |
1053 | * doesn't get written to disk before the caller actually commits the |
1054 | * new data |
1055 | */ |
1056 | if (!jh->b_transaction) { |
1057 | JBUFFER_TRACE(jh, "no transaction" ); |
1058 | J_ASSERT_JH(jh, !jh->b_next_transaction); |
1059 | JBUFFER_TRACE(jh, "file as BJ_Reserved" ); |
1060 | /* |
1061 | * Make sure all stores to jh (b_modified, b_frozen_data) are |
1062 | * visible before attaching it to the running transaction. |
1063 | * Paired with barrier in jbd2_write_access_granted() |
1064 | */ |
1065 | smp_wmb(); |
1066 | spin_lock(lock: &journal->j_list_lock); |
1067 | if (test_clear_buffer_dirty(bh)) { |
1068 | /* |
1069 | * Execute buffer dirty clearing and jh->b_transaction |
1070 | * assignment under journal->j_list_lock locked to |
1071 | * prevent bh being removed from checkpoint list if |
1072 | * the buffer is in an intermediate state (not dirty |
1073 | * and jh->b_transaction is NULL). |
1074 | */ |
1075 | JBUFFER_TRACE(jh, "Journalling dirty buffer" ); |
1076 | set_buffer_jbddirty(bh); |
1077 | } |
1078 | __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved); |
1079 | spin_unlock(lock: &journal->j_list_lock); |
1080 | unlock_buffer(bh); |
1081 | goto done; |
1082 | } |
1083 | unlock_buffer(bh); |
1084 | |
1085 | /* |
1086 | * If there is already a copy-out version of this buffer, then we don't |
1087 | * need to make another one |
1088 | */ |
1089 | if (jh->b_frozen_data) { |
1090 | JBUFFER_TRACE(jh, "has frozen data" ); |
1091 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
1092 | goto attach_next; |
1093 | } |
1094 | |
1095 | JBUFFER_TRACE(jh, "owned by older transaction" ); |
1096 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
1097 | J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction); |
1098 | |
1099 | /* |
1100 | * There is one case we have to be very careful about. If the |
1101 | * committing transaction is currently writing this buffer out to disk |
1102 | * and has NOT made a copy-out, then we cannot modify the buffer |
1103 | * contents at all right now. The essence of copy-out is that it is |
1104 | * the extra copy, not the primary copy, which gets journaled. If the |
1105 | * primary copy is already going to disk then we cannot do copy-out |
1106 | * here. |
1107 | */ |
1108 | if (buffer_shadow(bh)) { |
1109 | JBUFFER_TRACE(jh, "on shadow: sleep" ); |
1110 | spin_unlock(lock: &jh->b_state_lock); |
1111 | wait_on_bit_io(word: &bh->b_state, bit: BH_Shadow, TASK_UNINTERRUPTIBLE); |
1112 | goto repeat; |
1113 | } |
1114 | |
1115 | /* |
1116 | * Only do the copy if the currently-owning transaction still needs it. |
1117 | * If buffer isn't on BJ_Metadata list, the committing transaction is |
1118 | * past that stage (here we use the fact that BH_Shadow is set under |
1119 | * bh_state lock together with refiling to BJ_Shadow list and at this |
1120 | * point we know the buffer doesn't have BH_Shadow set). |
1121 | * |
1122 | * Subtle point, though: if this is a get_undo_access, then we will be |
1123 | * relying on the frozen_data to contain the new value of the |
1124 | * committed_data record after the transaction, so we HAVE to force the |
1125 | * frozen_data copy in that case. |
1126 | */ |
1127 | if (jh->b_jlist == BJ_Metadata || force_copy) { |
1128 | JBUFFER_TRACE(jh, "generate frozen data" ); |
1129 | if (!frozen_buffer) { |
1130 | JBUFFER_TRACE(jh, "allocate memory for buffer" ); |
1131 | spin_unlock(lock: &jh->b_state_lock); |
1132 | frozen_buffer = jbd2_alloc(size: jh2bh(jh)->b_size, |
1133 | GFP_NOFS | __GFP_NOFAIL); |
1134 | goto repeat; |
1135 | } |
1136 | jh->b_frozen_data = frozen_buffer; |
1137 | frozen_buffer = NULL; |
1138 | jbd2_freeze_jh_data(jh); |
1139 | } |
1140 | attach_next: |
1141 | /* |
1142 | * Make sure all stores to jh (b_modified, b_frozen_data) are visible |
1143 | * before attaching it to the running transaction. Paired with barrier |
1144 | * in jbd2_write_access_granted() |
1145 | */ |
1146 | smp_wmb(); |
1147 | jh->b_next_transaction = transaction; |
1148 | |
1149 | done: |
1150 | spin_unlock(lock: &jh->b_state_lock); |
1151 | |
1152 | /* |
1153 | * If we are about to journal a buffer, then any revoke pending on it is |
1154 | * no longer valid |
1155 | */ |
1156 | jbd2_journal_cancel_revoke(handle, jh); |
1157 | |
1158 | out: |
1159 | if (unlikely(frozen_buffer)) /* It's usually NULL */ |
1160 | jbd2_free(ptr: frozen_buffer, size: bh->b_size); |
1161 | |
1162 | JBUFFER_TRACE(jh, "exit" ); |
1163 | return error; |
1164 | } |
1165 | |
1166 | /* Fast check whether buffer is already attached to the required transaction */ |
1167 | static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh, |
1168 | bool undo) |
1169 | { |
1170 | struct journal_head *jh; |
1171 | bool ret = false; |
1172 | |
1173 | /* Dirty buffers require special handling... */ |
1174 | if (buffer_dirty(bh)) |
1175 | return false; |
1176 | |
1177 | /* |
1178 | * RCU protects us from dereferencing freed pages. So the checks we do |
1179 | * are guaranteed not to oops. However the jh slab object can get freed |
1180 | * & reallocated while we work with it. So we have to be careful. When |
1181 | * we see jh attached to the running transaction, we know it must stay |
1182 | * so until the transaction is committed. Thus jh won't be freed and |
1183 | * will be attached to the same bh while we run. However it can |
1184 | * happen jh gets freed, reallocated, and attached to the transaction |
1185 | * just after we get pointer to it from bh. So we have to be careful |
1186 | * and recheck jh still belongs to our bh before we return success. |
1187 | */ |
1188 | rcu_read_lock(); |
1189 | if (!buffer_jbd(bh)) |
1190 | goto out; |
1191 | /* This should be bh2jh() but that doesn't work with inline functions */ |
1192 | jh = READ_ONCE(bh->b_private); |
1193 | if (!jh) |
1194 | goto out; |
1195 | /* For undo access buffer must have data copied */ |
1196 | if (undo && !jh->b_committed_data) |
1197 | goto out; |
1198 | if (READ_ONCE(jh->b_transaction) != handle->h_transaction && |
1199 | READ_ONCE(jh->b_next_transaction) != handle->h_transaction) |
1200 | goto out; |
1201 | /* |
1202 | * There are two reasons for the barrier here: |
1203 | * 1) Make sure to fetch b_bh after we did previous checks so that we |
1204 | * detect when jh went through free, realloc, attach to transaction |
1205 | * while we were checking. Paired with implicit barrier in that path. |
1206 | * 2) So that access to bh done after jbd2_write_access_granted() |
1207 | * doesn't get reordered and see inconsistent state of concurrent |
1208 | * do_get_write_access(). |
1209 | */ |
1210 | smp_mb(); |
1211 | if (unlikely(jh->b_bh != bh)) |
1212 | goto out; |
1213 | ret = true; |
1214 | out: |
1215 | rcu_read_unlock(); |
1216 | return ret; |
1217 | } |
1218 | |
1219 | /** |
1220 | * jbd2_journal_get_write_access() - notify intent to modify a buffer |
1221 | * for metadata (not data) update. |
1222 | * @handle: transaction to add buffer modifications to |
1223 | * @bh: bh to be used for metadata writes |
1224 | * |
1225 | * Returns: error code or 0 on success. |
1226 | * |
1227 | * In full data journalling mode the buffer may be of type BJ_AsyncData, |
1228 | * because we're ``write()ing`` a buffer which is also part of a shared mapping. |
1229 | */ |
1230 | |
1231 | int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh) |
1232 | { |
1233 | struct journal_head *jh; |
1234 | journal_t *journal; |
1235 | int rc; |
1236 | |
1237 | if (is_handle_aborted(handle)) |
1238 | return -EROFS; |
1239 | |
1240 | journal = handle->h_transaction->t_journal; |
1241 | if (jbd2_check_fs_dev_write_error(journal)) { |
1242 | /* |
1243 | * If the fs dev has writeback errors, it may have failed |
1244 | * to async write out metadata buffers in the background. |
1245 | * In this case, we could read old data from disk and write |
1246 | * it out again, which may lead to on-disk filesystem |
1247 | * inconsistency. Aborting journal can avoid it happen. |
1248 | */ |
1249 | jbd2_journal_abort(journal, -EIO); |
1250 | return -EIO; |
1251 | } |
1252 | |
1253 | if (jbd2_write_access_granted(handle, bh, undo: false)) |
1254 | return 0; |
1255 | |
1256 | jh = jbd2_journal_add_journal_head(bh); |
1257 | /* We do not want to get caught playing with fields which the |
1258 | * log thread also manipulates. Make sure that the buffer |
1259 | * completes any outstanding IO before proceeding. */ |
1260 | rc = do_get_write_access(handle, jh, force_copy: 0); |
1261 | jbd2_journal_put_journal_head(jh); |
1262 | return rc; |
1263 | } |
1264 | |
1265 | |
1266 | /* |
1267 | * When the user wants to journal a newly created buffer_head |
1268 | * (ie. getblk() returned a new buffer and we are going to populate it |
1269 | * manually rather than reading off disk), then we need to keep the |
1270 | * buffer_head locked until it has been completely filled with new |
1271 | * data. In this case, we should be able to make the assertion that |
1272 | * the bh is not already part of an existing transaction. |
1273 | * |
1274 | * The buffer should already be locked by the caller by this point. |
1275 | * There is no lock ranking violation: it was a newly created, |
1276 | * unlocked buffer beforehand. */ |
1277 | |
1278 | /** |
1279 | * jbd2_journal_get_create_access () - notify intent to use newly created bh |
1280 | * @handle: transaction to new buffer to |
1281 | * @bh: new buffer. |
1282 | * |
1283 | * Call this if you create a new bh. |
1284 | */ |
1285 | int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh) |
1286 | { |
1287 | transaction_t *transaction = handle->h_transaction; |
1288 | journal_t *journal; |
1289 | struct journal_head *jh = jbd2_journal_add_journal_head(bh); |
1290 | int err; |
1291 | |
1292 | jbd2_debug(5, "journal_head %p\n" , jh); |
1293 | err = -EROFS; |
1294 | if (is_handle_aborted(handle)) |
1295 | goto out; |
1296 | journal = transaction->t_journal; |
1297 | err = 0; |
1298 | |
1299 | JBUFFER_TRACE(jh, "entry" ); |
1300 | /* |
1301 | * The buffer may already belong to this transaction due to pre-zeroing |
1302 | * in the filesystem's new_block code. It may also be on the previous, |
1303 | * committing transaction's lists, but it HAS to be in Forget state in |
1304 | * that case: the transaction must have deleted the buffer for it to be |
1305 | * reused here. |
1306 | */ |
1307 | spin_lock(lock: &jh->b_state_lock); |
1308 | J_ASSERT_JH(jh, (jh->b_transaction == transaction || |
1309 | jh->b_transaction == NULL || |
1310 | (jh->b_transaction == journal->j_committing_transaction && |
1311 | jh->b_jlist == BJ_Forget))); |
1312 | |
1313 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
1314 | J_ASSERT_JH(jh, buffer_locked(jh2bh(jh))); |
1315 | |
1316 | if (jh->b_transaction == NULL) { |
1317 | /* |
1318 | * Previous jbd2_journal_forget() could have left the buffer |
1319 | * with jbddirty bit set because it was being committed. When |
1320 | * the commit finished, we've filed the buffer for |
1321 | * checkpointing and marked it dirty. Now we are reallocating |
1322 | * the buffer so the transaction freeing it must have |
1323 | * committed and so it's safe to clear the dirty bit. |
1324 | */ |
1325 | clear_buffer_dirty(bh: jh2bh(jh)); |
1326 | /* first access by this transaction */ |
1327 | jh->b_modified = 0; |
1328 | |
1329 | JBUFFER_TRACE(jh, "file as BJ_Reserved" ); |
1330 | spin_lock(lock: &journal->j_list_lock); |
1331 | __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved); |
1332 | spin_unlock(lock: &journal->j_list_lock); |
1333 | } else if (jh->b_transaction == journal->j_committing_transaction) { |
1334 | /* first access by this transaction */ |
1335 | jh->b_modified = 0; |
1336 | |
1337 | JBUFFER_TRACE(jh, "set next transaction" ); |
1338 | spin_lock(lock: &journal->j_list_lock); |
1339 | jh->b_next_transaction = transaction; |
1340 | spin_unlock(lock: &journal->j_list_lock); |
1341 | } |
1342 | spin_unlock(lock: &jh->b_state_lock); |
1343 | |
1344 | /* |
1345 | * akpm: I added this. ext3_alloc_branch can pick up new indirect |
1346 | * blocks which contain freed but then revoked metadata. We need |
1347 | * to cancel the revoke in case we end up freeing it yet again |
1348 | * and the reallocating as data - this would cause a second revoke, |
1349 | * which hits an assertion error. |
1350 | */ |
1351 | JBUFFER_TRACE(jh, "cancelling revoke" ); |
1352 | jbd2_journal_cancel_revoke(handle, jh); |
1353 | out: |
1354 | jbd2_journal_put_journal_head(jh); |
1355 | return err; |
1356 | } |
1357 | |
1358 | /** |
1359 | * jbd2_journal_get_undo_access() - Notify intent to modify metadata with |
1360 | * non-rewindable consequences |
1361 | * @handle: transaction |
1362 | * @bh: buffer to undo |
1363 | * |
1364 | * Sometimes there is a need to distinguish between metadata which has |
1365 | * been committed to disk and that which has not. The ext3fs code uses |
1366 | * this for freeing and allocating space, we have to make sure that we |
1367 | * do not reuse freed space until the deallocation has been committed, |
1368 | * since if we overwrote that space we would make the delete |
1369 | * un-rewindable in case of a crash. |
1370 | * |
1371 | * To deal with that, jbd2_journal_get_undo_access requests write access to a |
1372 | * buffer for parts of non-rewindable operations such as delete |
1373 | * operations on the bitmaps. The journaling code must keep a copy of |
1374 | * the buffer's contents prior to the undo_access call until such time |
1375 | * as we know that the buffer has definitely been committed to disk. |
1376 | * |
1377 | * We never need to know which transaction the committed data is part |
1378 | * of, buffers touched here are guaranteed to be dirtied later and so |
1379 | * will be committed to a new transaction in due course, at which point |
1380 | * we can discard the old committed data pointer. |
1381 | * |
1382 | * Returns error number or 0 on success. |
1383 | */ |
1384 | int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh) |
1385 | { |
1386 | int err; |
1387 | struct journal_head *jh; |
1388 | char *committed_data = NULL; |
1389 | |
1390 | if (is_handle_aborted(handle)) |
1391 | return -EROFS; |
1392 | |
1393 | if (jbd2_write_access_granted(handle, bh, undo: true)) |
1394 | return 0; |
1395 | |
1396 | jh = jbd2_journal_add_journal_head(bh); |
1397 | JBUFFER_TRACE(jh, "entry" ); |
1398 | |
1399 | /* |
1400 | * Do this first --- it can drop the journal lock, so we want to |
1401 | * make sure that obtaining the committed_data is done |
1402 | * atomically wrt. completion of any outstanding commits. |
1403 | */ |
1404 | err = do_get_write_access(handle, jh, force_copy: 1); |
1405 | if (err) |
1406 | goto out; |
1407 | |
1408 | repeat: |
1409 | if (!jh->b_committed_data) |
1410 | committed_data = jbd2_alloc(size: jh2bh(jh)->b_size, |
1411 | GFP_NOFS|__GFP_NOFAIL); |
1412 | |
1413 | spin_lock(lock: &jh->b_state_lock); |
1414 | if (!jh->b_committed_data) { |
1415 | /* Copy out the current buffer contents into the |
1416 | * preserved, committed copy. */ |
1417 | JBUFFER_TRACE(jh, "generate b_committed data" ); |
1418 | if (!committed_data) { |
1419 | spin_unlock(lock: &jh->b_state_lock); |
1420 | goto repeat; |
1421 | } |
1422 | |
1423 | jh->b_committed_data = committed_data; |
1424 | committed_data = NULL; |
1425 | memcpy(jh->b_committed_data, bh->b_data, bh->b_size); |
1426 | } |
1427 | spin_unlock(lock: &jh->b_state_lock); |
1428 | out: |
1429 | jbd2_journal_put_journal_head(jh); |
1430 | if (unlikely(committed_data)) |
1431 | jbd2_free(ptr: committed_data, size: bh->b_size); |
1432 | return err; |
1433 | } |
1434 | |
1435 | /** |
1436 | * jbd2_journal_set_triggers() - Add triggers for commit writeout |
1437 | * @bh: buffer to trigger on |
1438 | * @type: struct jbd2_buffer_trigger_type containing the trigger(s). |
1439 | * |
1440 | * Set any triggers on this journal_head. This is always safe, because |
1441 | * triggers for a committing buffer will be saved off, and triggers for |
1442 | * a running transaction will match the buffer in that transaction. |
1443 | * |
1444 | * Call with NULL to clear the triggers. |
1445 | */ |
1446 | void jbd2_journal_set_triggers(struct buffer_head *bh, |
1447 | struct jbd2_buffer_trigger_type *type) |
1448 | { |
1449 | struct journal_head *jh = jbd2_journal_grab_journal_head(bh); |
1450 | |
1451 | if (WARN_ON_ONCE(!jh)) |
1452 | return; |
1453 | jh->b_triggers = type; |
1454 | jbd2_journal_put_journal_head(jh); |
1455 | } |
1456 | |
1457 | void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data, |
1458 | struct jbd2_buffer_trigger_type *triggers) |
1459 | { |
1460 | struct buffer_head *bh = jh2bh(jh); |
1461 | |
1462 | if (!triggers || !triggers->t_frozen) |
1463 | return; |
1464 | |
1465 | triggers->t_frozen(triggers, bh, mapped_data, bh->b_size); |
1466 | } |
1467 | |
1468 | void jbd2_buffer_abort_trigger(struct journal_head *jh, |
1469 | struct jbd2_buffer_trigger_type *triggers) |
1470 | { |
1471 | if (!triggers || !triggers->t_abort) |
1472 | return; |
1473 | |
1474 | triggers->t_abort(triggers, jh2bh(jh)); |
1475 | } |
1476 | |
1477 | /** |
1478 | * jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata |
1479 | * @handle: transaction to add buffer to. |
1480 | * @bh: buffer to mark |
1481 | * |
1482 | * mark dirty metadata which needs to be journaled as part of the current |
1483 | * transaction. |
1484 | * |
1485 | * The buffer must have previously had jbd2_journal_get_write_access() |
1486 | * called so that it has a valid journal_head attached to the buffer |
1487 | * head. |
1488 | * |
1489 | * The buffer is placed on the transaction's metadata list and is marked |
1490 | * as belonging to the transaction. |
1491 | * |
1492 | * Returns error number or 0 on success. |
1493 | * |
1494 | * Special care needs to be taken if the buffer already belongs to the |
1495 | * current committing transaction (in which case we should have frozen |
1496 | * data present for that commit). In that case, we don't relink the |
1497 | * buffer: that only gets done when the old transaction finally |
1498 | * completes its commit. |
1499 | */ |
1500 | int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh) |
1501 | { |
1502 | transaction_t *transaction = handle->h_transaction; |
1503 | journal_t *journal; |
1504 | struct journal_head *jh; |
1505 | int ret = 0; |
1506 | |
1507 | if (!buffer_jbd(bh)) |
1508 | return -EUCLEAN; |
1509 | |
1510 | /* |
1511 | * We don't grab jh reference here since the buffer must be part |
1512 | * of the running transaction. |
1513 | */ |
1514 | jh = bh2jh(bh); |
1515 | jbd2_debug(5, "journal_head %p\n" , jh); |
1516 | JBUFFER_TRACE(jh, "entry" ); |
1517 | |
1518 | /* |
1519 | * This and the following assertions are unreliable since we may see jh |
1520 | * in inconsistent state unless we grab bh_state lock. But this is |
1521 | * crucial to catch bugs so let's do a reliable check until the |
1522 | * lockless handling is fully proven. |
1523 | */ |
1524 | if (data_race(jh->b_transaction != transaction && |
1525 | jh->b_next_transaction != transaction)) { |
1526 | spin_lock(lock: &jh->b_state_lock); |
1527 | J_ASSERT_JH(jh, jh->b_transaction == transaction || |
1528 | jh->b_next_transaction == transaction); |
1529 | spin_unlock(lock: &jh->b_state_lock); |
1530 | } |
1531 | if (jh->b_modified == 1) { |
1532 | /* If it's in our transaction it must be in BJ_Metadata list. */ |
1533 | if (data_race(jh->b_transaction == transaction && |
1534 | jh->b_jlist != BJ_Metadata)) { |
1535 | spin_lock(lock: &jh->b_state_lock); |
1536 | if (jh->b_transaction == transaction && |
1537 | jh->b_jlist != BJ_Metadata) |
1538 | pr_err("JBD2: assertion failure: h_type=%u " |
1539 | "h_line_no=%u block_no=%llu jlist=%u\n" , |
1540 | handle->h_type, handle->h_line_no, |
1541 | (unsigned long long) bh->b_blocknr, |
1542 | jh->b_jlist); |
1543 | J_ASSERT_JH(jh, jh->b_transaction != transaction || |
1544 | jh->b_jlist == BJ_Metadata); |
1545 | spin_unlock(lock: &jh->b_state_lock); |
1546 | } |
1547 | goto out; |
1548 | } |
1549 | |
1550 | journal = transaction->t_journal; |
1551 | spin_lock(lock: &jh->b_state_lock); |
1552 | |
1553 | if (is_handle_aborted(handle)) { |
1554 | /* |
1555 | * Check journal aborting with @jh->b_state_lock locked, |
1556 | * since 'jh->b_transaction' could be replaced with |
1557 | * 'jh->b_next_transaction' during old transaction |
1558 | * committing if journal aborted, which may fail |
1559 | * assertion on 'jh->b_frozen_data == NULL'. |
1560 | */ |
1561 | ret = -EROFS; |
1562 | goto out_unlock_bh; |
1563 | } |
1564 | |
1565 | if (jh->b_modified == 0) { |
1566 | /* |
1567 | * This buffer's got modified and becoming part |
1568 | * of the transaction. This needs to be done |
1569 | * once a transaction -bzzz |
1570 | */ |
1571 | if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle) <= 0)) { |
1572 | ret = -ENOSPC; |
1573 | goto out_unlock_bh; |
1574 | } |
1575 | jh->b_modified = 1; |
1576 | handle->h_total_credits--; |
1577 | } |
1578 | |
1579 | /* |
1580 | * fastpath, to avoid expensive locking. If this buffer is already |
1581 | * on the running transaction's metadata list there is nothing to do. |
1582 | * Nobody can take it off again because there is a handle open. |
1583 | * I _think_ we're OK here with SMP barriers - a mistaken decision will |
1584 | * result in this test being false, so we go in and take the locks. |
1585 | */ |
1586 | if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) { |
1587 | JBUFFER_TRACE(jh, "fastpath" ); |
1588 | if (unlikely(jh->b_transaction != |
1589 | journal->j_running_transaction)) { |
1590 | printk(KERN_ERR "JBD2: %s: " |
1591 | "jh->b_transaction (%llu, %p, %u) != " |
1592 | "journal->j_running_transaction (%p, %u)\n" , |
1593 | journal->j_devname, |
1594 | (unsigned long long) bh->b_blocknr, |
1595 | jh->b_transaction, |
1596 | jh->b_transaction ? jh->b_transaction->t_tid : 0, |
1597 | journal->j_running_transaction, |
1598 | journal->j_running_transaction ? |
1599 | journal->j_running_transaction->t_tid : 0); |
1600 | ret = -EINVAL; |
1601 | } |
1602 | goto out_unlock_bh; |
1603 | } |
1604 | |
1605 | set_buffer_jbddirty(bh); |
1606 | |
1607 | /* |
1608 | * Metadata already on the current transaction list doesn't |
1609 | * need to be filed. Metadata on another transaction's list must |
1610 | * be committing, and will be refiled once the commit completes: |
1611 | * leave it alone for now. |
1612 | */ |
1613 | if (jh->b_transaction != transaction) { |
1614 | JBUFFER_TRACE(jh, "already on other transaction" ); |
1615 | if (unlikely(((jh->b_transaction != |
1616 | journal->j_committing_transaction)) || |
1617 | (jh->b_next_transaction != transaction))) { |
1618 | printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: " |
1619 | "bad jh for block %llu: " |
1620 | "transaction (%p, %u), " |
1621 | "jh->b_transaction (%p, %u), " |
1622 | "jh->b_next_transaction (%p, %u), jlist %u\n" , |
1623 | journal->j_devname, |
1624 | (unsigned long long) bh->b_blocknr, |
1625 | transaction, transaction->t_tid, |
1626 | jh->b_transaction, |
1627 | jh->b_transaction ? |
1628 | jh->b_transaction->t_tid : 0, |
1629 | jh->b_next_transaction, |
1630 | jh->b_next_transaction ? |
1631 | jh->b_next_transaction->t_tid : 0, |
1632 | jh->b_jlist); |
1633 | WARN_ON(1); |
1634 | ret = -EINVAL; |
1635 | } |
1636 | /* And this case is illegal: we can't reuse another |
1637 | * transaction's data buffer, ever. */ |
1638 | goto out_unlock_bh; |
1639 | } |
1640 | |
1641 | /* That test should have eliminated the following case: */ |
1642 | J_ASSERT_JH(jh, jh->b_frozen_data == NULL); |
1643 | |
1644 | JBUFFER_TRACE(jh, "file as BJ_Metadata" ); |
1645 | spin_lock(lock: &journal->j_list_lock); |
1646 | __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata); |
1647 | spin_unlock(lock: &journal->j_list_lock); |
1648 | out_unlock_bh: |
1649 | spin_unlock(lock: &jh->b_state_lock); |
1650 | out: |
1651 | JBUFFER_TRACE(jh, "exit" ); |
1652 | return ret; |
1653 | } |
1654 | |
1655 | /** |
1656 | * jbd2_journal_forget() - bforget() for potentially-journaled buffers. |
1657 | * @handle: transaction handle |
1658 | * @bh: bh to 'forget' |
1659 | * |
1660 | * We can only do the bforget if there are no commits pending against the |
1661 | * buffer. If the buffer is dirty in the current running transaction we |
1662 | * can safely unlink it. |
1663 | * |
1664 | * bh may not be a journalled buffer at all - it may be a non-JBD |
1665 | * buffer which came off the hashtable. Check for this. |
1666 | * |
1667 | * Decrements bh->b_count by one. |
1668 | * |
1669 | * Allow this call even if the handle has aborted --- it may be part of |
1670 | * the caller's cleanup after an abort. |
1671 | */ |
1672 | int jbd2_journal_forget(handle_t *handle, struct buffer_head *bh) |
1673 | { |
1674 | transaction_t *transaction = handle->h_transaction; |
1675 | journal_t *journal; |
1676 | struct journal_head *jh; |
1677 | int drop_reserve = 0; |
1678 | int err = 0; |
1679 | int was_modified = 0; |
1680 | |
1681 | if (is_handle_aborted(handle)) |
1682 | return -EROFS; |
1683 | journal = transaction->t_journal; |
1684 | |
1685 | BUFFER_TRACE(bh, "entry" ); |
1686 | |
1687 | jh = jbd2_journal_grab_journal_head(bh); |
1688 | if (!jh) { |
1689 | __bforget(bh); |
1690 | return 0; |
1691 | } |
1692 | |
1693 | spin_lock(lock: &jh->b_state_lock); |
1694 | |
1695 | /* Critical error: attempting to delete a bitmap buffer, maybe? |
1696 | * Don't do any jbd operations, and return an error. */ |
1697 | if (!J_EXPECT_JH(jh, !jh->b_committed_data, |
1698 | "inconsistent data on disk" )) { |
1699 | err = -EIO; |
1700 | goto drop; |
1701 | } |
1702 | |
1703 | /* keep track of whether or not this transaction modified us */ |
1704 | was_modified = jh->b_modified; |
1705 | |
1706 | /* |
1707 | * The buffer's going from the transaction, we must drop |
1708 | * all references -bzzz |
1709 | */ |
1710 | jh->b_modified = 0; |
1711 | |
1712 | if (jh->b_transaction == transaction) { |
1713 | J_ASSERT_JH(jh, !jh->b_frozen_data); |
1714 | |
1715 | /* If we are forgetting a buffer which is already part |
1716 | * of this transaction, then we can just drop it from |
1717 | * the transaction immediately. */ |
1718 | clear_buffer_dirty(bh); |
1719 | clear_buffer_jbddirty(bh); |
1720 | |
1721 | JBUFFER_TRACE(jh, "belongs to current transaction: unfile" ); |
1722 | |
1723 | /* |
1724 | * we only want to drop a reference if this transaction |
1725 | * modified the buffer |
1726 | */ |
1727 | if (was_modified) |
1728 | drop_reserve = 1; |
1729 | |
1730 | /* |
1731 | * We are no longer going to journal this buffer. |
1732 | * However, the commit of this transaction is still |
1733 | * important to the buffer: the delete that we are now |
1734 | * processing might obsolete an old log entry, so by |
1735 | * committing, we can satisfy the buffer's checkpoint. |
1736 | * |
1737 | * So, if we have a checkpoint on the buffer, we should |
1738 | * now refile the buffer on our BJ_Forget list so that |
1739 | * we know to remove the checkpoint after we commit. |
1740 | */ |
1741 | |
1742 | spin_lock(lock: &journal->j_list_lock); |
1743 | if (jh->b_cp_transaction) { |
1744 | __jbd2_journal_temp_unlink_buffer(jh); |
1745 | __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); |
1746 | } else { |
1747 | __jbd2_journal_unfile_buffer(jh); |
1748 | jbd2_journal_put_journal_head(jh); |
1749 | } |
1750 | spin_unlock(lock: &journal->j_list_lock); |
1751 | } else if (jh->b_transaction) { |
1752 | J_ASSERT_JH(jh, (jh->b_transaction == |
1753 | journal->j_committing_transaction)); |
1754 | /* However, if the buffer is still owned by a prior |
1755 | * (committing) transaction, we can't drop it yet... */ |
1756 | JBUFFER_TRACE(jh, "belongs to older transaction" ); |
1757 | /* ... but we CAN drop it from the new transaction through |
1758 | * marking the buffer as freed and set j_next_transaction to |
1759 | * the new transaction, so that not only the commit code |
1760 | * knows it should clear dirty bits when it is done with the |
1761 | * buffer, but also the buffer can be checkpointed only |
1762 | * after the new transaction commits. */ |
1763 | |
1764 | set_buffer_freed(bh); |
1765 | |
1766 | if (!jh->b_next_transaction) { |
1767 | spin_lock(lock: &journal->j_list_lock); |
1768 | jh->b_next_transaction = transaction; |
1769 | spin_unlock(lock: &journal->j_list_lock); |
1770 | } else { |
1771 | J_ASSERT(jh->b_next_transaction == transaction); |
1772 | |
1773 | /* |
1774 | * only drop a reference if this transaction modified |
1775 | * the buffer |
1776 | */ |
1777 | if (was_modified) |
1778 | drop_reserve = 1; |
1779 | } |
1780 | } else { |
1781 | /* |
1782 | * Finally, if the buffer is not belongs to any |
1783 | * transaction, we can just drop it now if it has no |
1784 | * checkpoint. |
1785 | */ |
1786 | spin_lock(lock: &journal->j_list_lock); |
1787 | if (!jh->b_cp_transaction) { |
1788 | JBUFFER_TRACE(jh, "belongs to none transaction" ); |
1789 | spin_unlock(lock: &journal->j_list_lock); |
1790 | goto drop; |
1791 | } |
1792 | |
1793 | /* |
1794 | * Otherwise, if the buffer has been written to disk, |
1795 | * it is safe to remove the checkpoint and drop it. |
1796 | */ |
1797 | if (jbd2_journal_try_remove_checkpoint(jh) >= 0) { |
1798 | spin_unlock(lock: &journal->j_list_lock); |
1799 | goto drop; |
1800 | } |
1801 | |
1802 | /* |
1803 | * The buffer is still not written to disk, we should |
1804 | * attach this buffer to current transaction so that the |
1805 | * buffer can be checkpointed only after the current |
1806 | * transaction commits. |
1807 | */ |
1808 | clear_buffer_dirty(bh); |
1809 | __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); |
1810 | spin_unlock(lock: &journal->j_list_lock); |
1811 | } |
1812 | drop: |
1813 | __brelse(bh); |
1814 | spin_unlock(lock: &jh->b_state_lock); |
1815 | jbd2_journal_put_journal_head(jh); |
1816 | if (drop_reserve) { |
1817 | /* no need to reserve log space for this block -bzzz */ |
1818 | handle->h_total_credits++; |
1819 | } |
1820 | return err; |
1821 | } |
1822 | |
1823 | /** |
1824 | * jbd2_journal_stop() - complete a transaction |
1825 | * @handle: transaction to complete. |
1826 | * |
1827 | * All done for a particular handle. |
1828 | * |
1829 | * There is not much action needed here. We just return any remaining |
1830 | * buffer credits to the transaction and remove the handle. The only |
1831 | * complication is that we need to start a commit operation if the |
1832 | * filesystem is marked for synchronous update. |
1833 | * |
1834 | * jbd2_journal_stop itself will not usually return an error, but it may |
1835 | * do so in unusual circumstances. In particular, expect it to |
1836 | * return -EIO if a jbd2_journal_abort has been executed since the |
1837 | * transaction began. |
1838 | */ |
1839 | int jbd2_journal_stop(handle_t *handle) |
1840 | { |
1841 | transaction_t *transaction = handle->h_transaction; |
1842 | journal_t *journal; |
1843 | int err = 0, wait_for_commit = 0; |
1844 | tid_t tid; |
1845 | pid_t pid; |
1846 | |
1847 | if (--handle->h_ref > 0) { |
1848 | jbd2_debug(4, "h_ref %d -> %d\n" , handle->h_ref + 1, |
1849 | handle->h_ref); |
1850 | if (is_handle_aborted(handle)) |
1851 | return -EIO; |
1852 | return 0; |
1853 | } |
1854 | if (!transaction) { |
1855 | /* |
1856 | * Handle is already detached from the transaction so there is |
1857 | * nothing to do other than free the handle. |
1858 | */ |
1859 | memalloc_nofs_restore(flags: handle->saved_alloc_context); |
1860 | goto free_and_exit; |
1861 | } |
1862 | journal = transaction->t_journal; |
1863 | tid = transaction->t_tid; |
1864 | |
1865 | if (is_handle_aborted(handle)) |
1866 | err = -EIO; |
1867 | |
1868 | jbd2_debug(4, "Handle %p going down\n" , handle); |
1869 | trace_jbd2_handle_stats(dev: journal->j_fs_dev->bd_dev, |
1870 | tid, type: handle->h_type, line_no: handle->h_line_no, |
1871 | interval: jiffies - handle->h_start_jiffies, |
1872 | sync: handle->h_sync, requested_blocks: handle->h_requested_credits, |
1873 | dirtied_blocks: (handle->h_requested_credits - |
1874 | handle->h_total_credits)); |
1875 | |
1876 | /* |
1877 | * Implement synchronous transaction batching. If the handle |
1878 | * was synchronous, don't force a commit immediately. Let's |
1879 | * yield and let another thread piggyback onto this |
1880 | * transaction. Keep doing that while new threads continue to |
1881 | * arrive. It doesn't cost much - we're about to run a commit |
1882 | * and sleep on IO anyway. Speeds up many-threaded, many-dir |
1883 | * operations by 30x or more... |
1884 | * |
1885 | * We try and optimize the sleep time against what the |
1886 | * underlying disk can do, instead of having a static sleep |
1887 | * time. This is useful for the case where our storage is so |
1888 | * fast that it is more optimal to go ahead and force a flush |
1889 | * and wait for the transaction to be committed than it is to |
1890 | * wait for an arbitrary amount of time for new writers to |
1891 | * join the transaction. We achieve this by measuring how |
1892 | * long it takes to commit a transaction, and compare it with |
1893 | * how long this transaction has been running, and if run time |
1894 | * < commit time then we sleep for the delta and commit. This |
1895 | * greatly helps super fast disks that would see slowdowns as |
1896 | * more threads started doing fsyncs. |
1897 | * |
1898 | * But don't do this if this process was the most recent one |
1899 | * to perform a synchronous write. We do this to detect the |
1900 | * case where a single process is doing a stream of sync |
1901 | * writes. No point in waiting for joiners in that case. |
1902 | * |
1903 | * Setting max_batch_time to 0 disables this completely. |
1904 | */ |
1905 | pid = current->pid; |
1906 | if (handle->h_sync && journal->j_last_sync_writer != pid && |
1907 | journal->j_max_batch_time) { |
1908 | u64 commit_time, trans_time; |
1909 | |
1910 | journal->j_last_sync_writer = pid; |
1911 | |
1912 | read_lock(&journal->j_state_lock); |
1913 | commit_time = journal->j_average_commit_time; |
1914 | read_unlock(&journal->j_state_lock); |
1915 | |
1916 | trans_time = ktime_to_ns(ktime_sub(ktime_get(), |
1917 | transaction->t_start_time)); |
1918 | |
1919 | commit_time = max_t(u64, commit_time, |
1920 | 1000*journal->j_min_batch_time); |
1921 | commit_time = min_t(u64, commit_time, |
1922 | 1000*journal->j_max_batch_time); |
1923 | |
1924 | if (trans_time < commit_time) { |
1925 | ktime_t expires = ktime_add_ns(ktime_get(), |
1926 | commit_time); |
1927 | set_current_state(TASK_UNINTERRUPTIBLE); |
1928 | schedule_hrtimeout(expires: &expires, mode: HRTIMER_MODE_ABS); |
1929 | } |
1930 | } |
1931 | |
1932 | if (handle->h_sync) |
1933 | transaction->t_synchronous_commit = 1; |
1934 | |
1935 | /* |
1936 | * If the handle is marked SYNC, we need to set another commit |
1937 | * going! We also want to force a commit if the transaction is too |
1938 | * old now. |
1939 | */ |
1940 | if (handle->h_sync || |
1941 | time_after_eq(jiffies, transaction->t_expires)) { |
1942 | /* Do this even for aborted journals: an abort still |
1943 | * completes the commit thread, it just doesn't write |
1944 | * anything to disk. */ |
1945 | |
1946 | jbd2_debug(2, "transaction too old, requesting commit for " |
1947 | "handle %p\n" , handle); |
1948 | /* This is non-blocking */ |
1949 | jbd2_log_start_commit(journal, tid); |
1950 | |
1951 | /* |
1952 | * Special case: JBD2_SYNC synchronous updates require us |
1953 | * to wait for the commit to complete. |
1954 | */ |
1955 | if (handle->h_sync && !(current->flags & PF_MEMALLOC)) |
1956 | wait_for_commit = 1; |
1957 | } |
1958 | |
1959 | /* |
1960 | * Once stop_this_handle() drops t_updates, the transaction could start |
1961 | * committing on us and eventually disappear. So we must not |
1962 | * dereference transaction pointer again after calling |
1963 | * stop_this_handle(). |
1964 | */ |
1965 | stop_this_handle(handle); |
1966 | |
1967 | if (wait_for_commit) |
1968 | err = jbd2_log_wait_commit(journal, tid); |
1969 | |
1970 | free_and_exit: |
1971 | if (handle->h_rsv_handle) |
1972 | jbd2_free_handle(handle: handle->h_rsv_handle); |
1973 | jbd2_free_handle(handle); |
1974 | return err; |
1975 | } |
1976 | |
1977 | /* |
1978 | * |
1979 | * List management code snippets: various functions for manipulating the |
1980 | * transaction buffer lists. |
1981 | * |
1982 | */ |
1983 | |
1984 | /* |
1985 | * Append a buffer to a transaction list, given the transaction's list head |
1986 | * pointer. |
1987 | * |
1988 | * j_list_lock is held. |
1989 | * |
1990 | * jh->b_state_lock is held. |
1991 | */ |
1992 | |
1993 | static inline void |
1994 | __blist_add_buffer(struct journal_head **list, struct journal_head *jh) |
1995 | { |
1996 | if (!*list) { |
1997 | jh->b_tnext = jh->b_tprev = jh; |
1998 | *list = jh; |
1999 | } else { |
2000 | /* Insert at the tail of the list to preserve order */ |
2001 | struct journal_head *first = *list, *last = first->b_tprev; |
2002 | jh->b_tprev = last; |
2003 | jh->b_tnext = first; |
2004 | last->b_tnext = first->b_tprev = jh; |
2005 | } |
2006 | } |
2007 | |
2008 | /* |
2009 | * Remove a buffer from a transaction list, given the transaction's list |
2010 | * head pointer. |
2011 | * |
2012 | * Called with j_list_lock held, and the journal may not be locked. |
2013 | * |
2014 | * jh->b_state_lock is held. |
2015 | */ |
2016 | |
2017 | static inline void |
2018 | __blist_del_buffer(struct journal_head **list, struct journal_head *jh) |
2019 | { |
2020 | if (*list == jh) { |
2021 | *list = jh->b_tnext; |
2022 | if (*list == jh) |
2023 | *list = NULL; |
2024 | } |
2025 | jh->b_tprev->b_tnext = jh->b_tnext; |
2026 | jh->b_tnext->b_tprev = jh->b_tprev; |
2027 | } |
2028 | |
2029 | /* |
2030 | * Remove a buffer from the appropriate transaction list. |
2031 | * |
2032 | * Note that this function can *change* the value of |
2033 | * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or |
2034 | * t_reserved_list. If the caller is holding onto a copy of one of these |
2035 | * pointers, it could go bad. Generally the caller needs to re-read the |
2036 | * pointer from the transaction_t. |
2037 | * |
2038 | * Called under j_list_lock. |
2039 | */ |
2040 | static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh) |
2041 | { |
2042 | struct journal_head **list = NULL; |
2043 | transaction_t *transaction; |
2044 | struct buffer_head *bh = jh2bh(jh); |
2045 | |
2046 | lockdep_assert_held(&jh->b_state_lock); |
2047 | transaction = jh->b_transaction; |
2048 | if (transaction) |
2049 | assert_spin_locked(&transaction->t_journal->j_list_lock); |
2050 | |
2051 | J_ASSERT_JH(jh, jh->b_jlist < BJ_Types); |
2052 | if (jh->b_jlist != BJ_None) |
2053 | J_ASSERT_JH(jh, transaction != NULL); |
2054 | |
2055 | switch (jh->b_jlist) { |
2056 | case BJ_None: |
2057 | return; |
2058 | case BJ_Metadata: |
2059 | transaction->t_nr_buffers--; |
2060 | J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0); |
2061 | list = &transaction->t_buffers; |
2062 | break; |
2063 | case BJ_Forget: |
2064 | list = &transaction->t_forget; |
2065 | break; |
2066 | case BJ_Shadow: |
2067 | list = &transaction->t_shadow_list; |
2068 | break; |
2069 | case BJ_Reserved: |
2070 | list = &transaction->t_reserved_list; |
2071 | break; |
2072 | } |
2073 | |
2074 | __blist_del_buffer(list, jh); |
2075 | jh->b_jlist = BJ_None; |
2076 | if (transaction && is_journal_aborted(journal: transaction->t_journal)) |
2077 | clear_buffer_jbddirty(bh); |
2078 | else if (test_clear_buffer_jbddirty(bh)) |
2079 | mark_buffer_dirty(bh); /* Expose it to the VM */ |
2080 | } |
2081 | |
2082 | /* |
2083 | * Remove buffer from all transactions. The caller is responsible for dropping |
2084 | * the jh reference that belonged to the transaction. |
2085 | * |
2086 | * Called with bh_state lock and j_list_lock |
2087 | */ |
2088 | static void __jbd2_journal_unfile_buffer(struct journal_head *jh) |
2089 | { |
2090 | J_ASSERT_JH(jh, jh->b_transaction != NULL); |
2091 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
2092 | |
2093 | __jbd2_journal_temp_unlink_buffer(jh); |
2094 | jh->b_transaction = NULL; |
2095 | } |
2096 | |
2097 | void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh) |
2098 | { |
2099 | struct buffer_head *bh = jh2bh(jh); |
2100 | |
2101 | /* Get reference so that buffer cannot be freed before we unlock it */ |
2102 | get_bh(bh); |
2103 | spin_lock(lock: &jh->b_state_lock); |
2104 | spin_lock(lock: &journal->j_list_lock); |
2105 | __jbd2_journal_unfile_buffer(jh); |
2106 | spin_unlock(lock: &journal->j_list_lock); |
2107 | spin_unlock(lock: &jh->b_state_lock); |
2108 | jbd2_journal_put_journal_head(jh); |
2109 | __brelse(bh); |
2110 | } |
2111 | |
2112 | /** |
2113 | * jbd2_journal_try_to_free_buffers() - try to free page buffers. |
2114 | * @journal: journal for operation |
2115 | * @folio: Folio to detach data from. |
2116 | * |
2117 | * For all the buffers on this page, |
2118 | * if they are fully written out ordered data, move them onto BUF_CLEAN |
2119 | * so try_to_free_buffers() can reap them. |
2120 | * |
2121 | * This function returns non-zero if we wish try_to_free_buffers() |
2122 | * to be called. We do this if the page is releasable by try_to_free_buffers(). |
2123 | * We also do it if the page has locked or dirty buffers and the caller wants |
2124 | * us to perform sync or async writeout. |
2125 | * |
2126 | * This complicates JBD locking somewhat. We aren't protected by the |
2127 | * BKL here. We wish to remove the buffer from its committing or |
2128 | * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer. |
2129 | * |
2130 | * This may *change* the value of transaction_t->t_datalist, so anyone |
2131 | * who looks at t_datalist needs to lock against this function. |
2132 | * |
2133 | * Even worse, someone may be doing a jbd2_journal_dirty_data on this |
2134 | * buffer. So we need to lock against that. jbd2_journal_dirty_data() |
2135 | * will come out of the lock with the buffer dirty, which makes it |
2136 | * ineligible for release here. |
2137 | * |
2138 | * Who else is affected by this? hmm... Really the only contender |
2139 | * is do_get_write_access() - it could be looking at the buffer while |
2140 | * journal_try_to_free_buffer() is changing its state. But that |
2141 | * cannot happen because we never reallocate freed data as metadata |
2142 | * while the data is part of a transaction. Yes? |
2143 | * |
2144 | * Return false on failure, true on success |
2145 | */ |
2146 | bool jbd2_journal_try_to_free_buffers(journal_t *journal, struct folio *folio) |
2147 | { |
2148 | struct buffer_head *head; |
2149 | struct buffer_head *bh; |
2150 | bool ret = false; |
2151 | |
2152 | J_ASSERT(folio_test_locked(folio)); |
2153 | |
2154 | head = folio_buffers(folio); |
2155 | bh = head; |
2156 | do { |
2157 | struct journal_head *jh; |
2158 | |
2159 | /* |
2160 | * We take our own ref against the journal_head here to avoid |
2161 | * having to add tons of locking around each instance of |
2162 | * jbd2_journal_put_journal_head(). |
2163 | */ |
2164 | jh = jbd2_journal_grab_journal_head(bh); |
2165 | if (!jh) |
2166 | continue; |
2167 | |
2168 | spin_lock(lock: &jh->b_state_lock); |
2169 | if (!jh->b_transaction && !jh->b_next_transaction) { |
2170 | spin_lock(lock: &journal->j_list_lock); |
2171 | /* Remove written-back checkpointed metadata buffer */ |
2172 | if (jh->b_cp_transaction != NULL) |
2173 | jbd2_journal_try_remove_checkpoint(jh); |
2174 | spin_unlock(lock: &journal->j_list_lock); |
2175 | } |
2176 | spin_unlock(lock: &jh->b_state_lock); |
2177 | jbd2_journal_put_journal_head(jh); |
2178 | if (buffer_jbd(bh)) |
2179 | goto busy; |
2180 | } while ((bh = bh->b_this_page) != head); |
2181 | |
2182 | ret = try_to_free_buffers(folio); |
2183 | busy: |
2184 | return ret; |
2185 | } |
2186 | |
2187 | /* |
2188 | * This buffer is no longer needed. If it is on an older transaction's |
2189 | * checkpoint list we need to record it on this transaction's forget list |
2190 | * to pin this buffer (and hence its checkpointing transaction) down until |
2191 | * this transaction commits. If the buffer isn't on a checkpoint list, we |
2192 | * release it. |
2193 | * Returns non-zero if JBD no longer has an interest in the buffer. |
2194 | * |
2195 | * Called under j_list_lock. |
2196 | * |
2197 | * Called under jh->b_state_lock. |
2198 | */ |
2199 | static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction) |
2200 | { |
2201 | int may_free = 1; |
2202 | struct buffer_head *bh = jh2bh(jh); |
2203 | |
2204 | if (jh->b_cp_transaction) { |
2205 | JBUFFER_TRACE(jh, "on running+cp transaction" ); |
2206 | __jbd2_journal_temp_unlink_buffer(jh); |
2207 | /* |
2208 | * We don't want to write the buffer anymore, clear the |
2209 | * bit so that we don't confuse checks in |
2210 | * __journal_file_buffer |
2211 | */ |
2212 | clear_buffer_dirty(bh); |
2213 | __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); |
2214 | may_free = 0; |
2215 | } else { |
2216 | JBUFFER_TRACE(jh, "on running transaction" ); |
2217 | __jbd2_journal_unfile_buffer(jh); |
2218 | jbd2_journal_put_journal_head(jh); |
2219 | } |
2220 | return may_free; |
2221 | } |
2222 | |
2223 | /* |
2224 | * jbd2_journal_invalidate_folio |
2225 | * |
2226 | * This code is tricky. It has a number of cases to deal with. |
2227 | * |
2228 | * There are two invariants which this code relies on: |
2229 | * |
2230 | * i_size must be updated on disk before we start calling invalidate_folio |
2231 | * on the data. |
2232 | * |
2233 | * This is done in ext3 by defining an ext3_setattr method which |
2234 | * updates i_size before truncate gets going. By maintaining this |
2235 | * invariant, we can be sure that it is safe to throw away any buffers |
2236 | * attached to the current transaction: once the transaction commits, |
2237 | * we know that the data will not be needed. |
2238 | * |
2239 | * Note however that we can *not* throw away data belonging to the |
2240 | * previous, committing transaction! |
2241 | * |
2242 | * Any disk blocks which *are* part of the previous, committing |
2243 | * transaction (and which therefore cannot be discarded immediately) are |
2244 | * not going to be reused in the new running transaction |
2245 | * |
2246 | * The bitmap committed_data images guarantee this: any block which is |
2247 | * allocated in one transaction and removed in the next will be marked |
2248 | * as in-use in the committed_data bitmap, so cannot be reused until |
2249 | * the next transaction to delete the block commits. This means that |
2250 | * leaving committing buffers dirty is quite safe: the disk blocks |
2251 | * cannot be reallocated to a different file and so buffer aliasing is |
2252 | * not possible. |
2253 | * |
2254 | * |
2255 | * The above applies mainly to ordered data mode. In writeback mode we |
2256 | * don't make guarantees about the order in which data hits disk --- in |
2257 | * particular we don't guarantee that new dirty data is flushed before |
2258 | * transaction commit --- so it is always safe just to discard data |
2259 | * immediately in that mode. --sct |
2260 | */ |
2261 | |
2262 | /* |
2263 | * The journal_unmap_buffer helper function returns zero if the buffer |
2264 | * concerned remains pinned as an anonymous buffer belonging to an older |
2265 | * transaction. |
2266 | * |
2267 | * We're outside-transaction here. Either or both of j_running_transaction |
2268 | * and j_committing_transaction may be NULL. |
2269 | */ |
2270 | static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh, |
2271 | int partial_page) |
2272 | { |
2273 | transaction_t *transaction; |
2274 | struct journal_head *jh; |
2275 | int may_free = 1; |
2276 | |
2277 | BUFFER_TRACE(bh, "entry" ); |
2278 | |
2279 | /* |
2280 | * It is safe to proceed here without the j_list_lock because the |
2281 | * buffers cannot be stolen by try_to_free_buffers as long as we are |
2282 | * holding the page lock. --sct |
2283 | */ |
2284 | |
2285 | jh = jbd2_journal_grab_journal_head(bh); |
2286 | if (!jh) |
2287 | goto zap_buffer_unlocked; |
2288 | |
2289 | /* OK, we have data buffer in journaled mode */ |
2290 | write_lock(&journal->j_state_lock); |
2291 | spin_lock(lock: &jh->b_state_lock); |
2292 | spin_lock(lock: &journal->j_list_lock); |
2293 | |
2294 | /* |
2295 | * We cannot remove the buffer from checkpoint lists until the |
2296 | * transaction adding inode to orphan list (let's call it T) |
2297 | * is committed. Otherwise if the transaction changing the |
2298 | * buffer would be cleaned from the journal before T is |
2299 | * committed, a crash will cause that the correct contents of |
2300 | * the buffer will be lost. On the other hand we have to |
2301 | * clear the buffer dirty bit at latest at the moment when the |
2302 | * transaction marking the buffer as freed in the filesystem |
2303 | * structures is committed because from that moment on the |
2304 | * block can be reallocated and used by a different page. |
2305 | * Since the block hasn't been freed yet but the inode has |
2306 | * already been added to orphan list, it is safe for us to add |
2307 | * the buffer to BJ_Forget list of the newest transaction. |
2308 | * |
2309 | * Also we have to clear buffer_mapped flag of a truncated buffer |
2310 | * because the buffer_head may be attached to the page straddling |
2311 | * i_size (can happen only when blocksize < pagesize) and thus the |
2312 | * buffer_head can be reused when the file is extended again. So we end |
2313 | * up keeping around invalidated buffers attached to transactions' |
2314 | * BJ_Forget list just to stop checkpointing code from cleaning up |
2315 | * the transaction this buffer was modified in. |
2316 | */ |
2317 | transaction = jh->b_transaction; |
2318 | if (transaction == NULL) { |
2319 | /* First case: not on any transaction. If it |
2320 | * has no checkpoint link, then we can zap it: |
2321 | * it's a writeback-mode buffer so we don't care |
2322 | * if it hits disk safely. */ |
2323 | if (!jh->b_cp_transaction) { |
2324 | JBUFFER_TRACE(jh, "not on any transaction: zap" ); |
2325 | goto zap_buffer; |
2326 | } |
2327 | |
2328 | if (!buffer_dirty(bh)) { |
2329 | /* bdflush has written it. We can drop it now */ |
2330 | __jbd2_journal_remove_checkpoint(jh); |
2331 | goto zap_buffer; |
2332 | } |
2333 | |
2334 | /* OK, it must be in the journal but still not |
2335 | * written fully to disk: it's metadata or |
2336 | * journaled data... */ |
2337 | |
2338 | if (journal->j_running_transaction) { |
2339 | /* ... and once the current transaction has |
2340 | * committed, the buffer won't be needed any |
2341 | * longer. */ |
2342 | JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget" ); |
2343 | may_free = __dispose_buffer(jh, |
2344 | transaction: journal->j_running_transaction); |
2345 | goto zap_buffer; |
2346 | } else { |
2347 | /* There is no currently-running transaction. So the |
2348 | * orphan record which we wrote for this file must have |
2349 | * passed into commit. We must attach this buffer to |
2350 | * the committing transaction, if it exists. */ |
2351 | if (journal->j_committing_transaction) { |
2352 | JBUFFER_TRACE(jh, "give to committing trans" ); |
2353 | may_free = __dispose_buffer(jh, |
2354 | transaction: journal->j_committing_transaction); |
2355 | goto zap_buffer; |
2356 | } else { |
2357 | /* The orphan record's transaction has |
2358 | * committed. We can cleanse this buffer */ |
2359 | clear_buffer_jbddirty(bh); |
2360 | __jbd2_journal_remove_checkpoint(jh); |
2361 | goto zap_buffer; |
2362 | } |
2363 | } |
2364 | } else if (transaction == journal->j_committing_transaction) { |
2365 | JBUFFER_TRACE(jh, "on committing transaction" ); |
2366 | /* |
2367 | * The buffer is committing, we simply cannot touch |
2368 | * it. If the page is straddling i_size we have to wait |
2369 | * for commit and try again. |
2370 | */ |
2371 | if (partial_page) { |
2372 | spin_unlock(lock: &journal->j_list_lock); |
2373 | spin_unlock(lock: &jh->b_state_lock); |
2374 | write_unlock(&journal->j_state_lock); |
2375 | jbd2_journal_put_journal_head(jh); |
2376 | /* Already zapped buffer? Nothing to do... */ |
2377 | if (!bh->b_bdev) |
2378 | return 0; |
2379 | return -EBUSY; |
2380 | } |
2381 | /* |
2382 | * OK, buffer won't be reachable after truncate. We just clear |
2383 | * b_modified to not confuse transaction credit accounting, and |
2384 | * set j_next_transaction to the running transaction (if there |
2385 | * is one) and mark buffer as freed so that commit code knows |
2386 | * it should clear dirty bits when it is done with the buffer. |
2387 | */ |
2388 | set_buffer_freed(bh); |
2389 | if (journal->j_running_transaction && buffer_jbddirty(bh)) |
2390 | jh->b_next_transaction = journal->j_running_transaction; |
2391 | jh->b_modified = 0; |
2392 | spin_unlock(lock: &journal->j_list_lock); |
2393 | spin_unlock(lock: &jh->b_state_lock); |
2394 | write_unlock(&journal->j_state_lock); |
2395 | jbd2_journal_put_journal_head(jh); |
2396 | return 0; |
2397 | } else { |
2398 | /* Good, the buffer belongs to the running transaction. |
2399 | * We are writing our own transaction's data, not any |
2400 | * previous one's, so it is safe to throw it away |
2401 | * (remember that we expect the filesystem to have set |
2402 | * i_size already for this truncate so recovery will not |
2403 | * expose the disk blocks we are discarding here.) */ |
2404 | J_ASSERT_JH(jh, transaction == journal->j_running_transaction); |
2405 | JBUFFER_TRACE(jh, "on running transaction" ); |
2406 | may_free = __dispose_buffer(jh, transaction); |
2407 | } |
2408 | |
2409 | zap_buffer: |
2410 | /* |
2411 | * This is tricky. Although the buffer is truncated, it may be reused |
2412 | * if blocksize < pagesize and it is attached to the page straddling |
2413 | * EOF. Since the buffer might have been added to BJ_Forget list of the |
2414 | * running transaction, journal_get_write_access() won't clear |
2415 | * b_modified and credit accounting gets confused. So clear b_modified |
2416 | * here. |
2417 | */ |
2418 | jh->b_modified = 0; |
2419 | spin_unlock(lock: &journal->j_list_lock); |
2420 | spin_unlock(lock: &jh->b_state_lock); |
2421 | write_unlock(&journal->j_state_lock); |
2422 | jbd2_journal_put_journal_head(jh); |
2423 | zap_buffer_unlocked: |
2424 | clear_buffer_dirty(bh); |
2425 | J_ASSERT_BH(bh, !buffer_jbddirty(bh)); |
2426 | clear_buffer_mapped(bh); |
2427 | clear_buffer_req(bh); |
2428 | clear_buffer_new(bh); |
2429 | clear_buffer_delay(bh); |
2430 | clear_buffer_unwritten(bh); |
2431 | bh->b_bdev = NULL; |
2432 | return may_free; |
2433 | } |
2434 | |
2435 | /** |
2436 | * jbd2_journal_invalidate_folio() |
2437 | * @journal: journal to use for flush... |
2438 | * @folio: folio to flush |
2439 | * @offset: start of the range to invalidate |
2440 | * @length: length of the range to invalidate |
2441 | * |
2442 | * Reap page buffers containing data after in the specified range in page. |
2443 | * Can return -EBUSY if buffers are part of the committing transaction and |
2444 | * the page is straddling i_size. Caller then has to wait for current commit |
2445 | * and try again. |
2446 | */ |
2447 | int jbd2_journal_invalidate_folio(journal_t *journal, struct folio *folio, |
2448 | size_t offset, size_t length) |
2449 | { |
2450 | struct buffer_head *head, *bh, *next; |
2451 | unsigned int stop = offset + length; |
2452 | unsigned int curr_off = 0; |
2453 | int partial_page = (offset || length < folio_size(folio)); |
2454 | int may_free = 1; |
2455 | int ret = 0; |
2456 | |
2457 | if (!folio_test_locked(folio)) |
2458 | BUG(); |
2459 | head = folio_buffers(folio); |
2460 | if (!head) |
2461 | return 0; |
2462 | |
2463 | BUG_ON(stop > folio_size(folio) || stop < length); |
2464 | |
2465 | /* We will potentially be playing with lists other than just the |
2466 | * data lists (especially for journaled data mode), so be |
2467 | * cautious in our locking. */ |
2468 | |
2469 | bh = head; |
2470 | do { |
2471 | unsigned int next_off = curr_off + bh->b_size; |
2472 | next = bh->b_this_page; |
2473 | |
2474 | if (next_off > stop) |
2475 | return 0; |
2476 | |
2477 | if (offset <= curr_off) { |
2478 | /* This block is wholly outside the truncation point */ |
2479 | lock_buffer(bh); |
2480 | ret = journal_unmap_buffer(journal, bh, partial_page); |
2481 | unlock_buffer(bh); |
2482 | if (ret < 0) |
2483 | return ret; |
2484 | may_free &= ret; |
2485 | } |
2486 | curr_off = next_off; |
2487 | bh = next; |
2488 | |
2489 | } while (bh != head); |
2490 | |
2491 | if (!partial_page) { |
2492 | if (may_free && try_to_free_buffers(folio)) |
2493 | J_ASSERT(!folio_buffers(folio)); |
2494 | } |
2495 | return 0; |
2496 | } |
2497 | |
2498 | /* |
2499 | * File a buffer on the given transaction list. |
2500 | */ |
2501 | void __jbd2_journal_file_buffer(struct journal_head *jh, |
2502 | transaction_t *transaction, int jlist) |
2503 | { |
2504 | struct journal_head **list = NULL; |
2505 | int was_dirty = 0; |
2506 | struct buffer_head *bh = jh2bh(jh); |
2507 | |
2508 | lockdep_assert_held(&jh->b_state_lock); |
2509 | assert_spin_locked(&transaction->t_journal->j_list_lock); |
2510 | |
2511 | J_ASSERT_JH(jh, jh->b_jlist < BJ_Types); |
2512 | J_ASSERT_JH(jh, jh->b_transaction == transaction || |
2513 | jh->b_transaction == NULL); |
2514 | |
2515 | if (jh->b_transaction && jh->b_jlist == jlist) |
2516 | return; |
2517 | |
2518 | if (jlist == BJ_Metadata || jlist == BJ_Reserved || |
2519 | jlist == BJ_Shadow || jlist == BJ_Forget) { |
2520 | /* |
2521 | * For metadata buffers, we track dirty bit in buffer_jbddirty |
2522 | * instead of buffer_dirty. We should not see a dirty bit set |
2523 | * here because we clear it in do_get_write_access but e.g. |
2524 | * tune2fs can modify the sb and set the dirty bit at any time |
2525 | * so we try to gracefully handle that. |
2526 | */ |
2527 | if (buffer_dirty(bh)) |
2528 | warn_dirty_buffer(bh); |
2529 | if (test_clear_buffer_dirty(bh) || |
2530 | test_clear_buffer_jbddirty(bh)) |
2531 | was_dirty = 1; |
2532 | } |
2533 | |
2534 | if (jh->b_transaction) |
2535 | __jbd2_journal_temp_unlink_buffer(jh); |
2536 | else |
2537 | jbd2_journal_grab_journal_head(bh); |
2538 | jh->b_transaction = transaction; |
2539 | |
2540 | switch (jlist) { |
2541 | case BJ_None: |
2542 | J_ASSERT_JH(jh, !jh->b_committed_data); |
2543 | J_ASSERT_JH(jh, !jh->b_frozen_data); |
2544 | return; |
2545 | case BJ_Metadata: |
2546 | transaction->t_nr_buffers++; |
2547 | list = &transaction->t_buffers; |
2548 | break; |
2549 | case BJ_Forget: |
2550 | list = &transaction->t_forget; |
2551 | break; |
2552 | case BJ_Shadow: |
2553 | list = &transaction->t_shadow_list; |
2554 | break; |
2555 | case BJ_Reserved: |
2556 | list = &transaction->t_reserved_list; |
2557 | break; |
2558 | } |
2559 | |
2560 | __blist_add_buffer(list, jh); |
2561 | jh->b_jlist = jlist; |
2562 | |
2563 | if (was_dirty) |
2564 | set_buffer_jbddirty(bh); |
2565 | } |
2566 | |
2567 | void jbd2_journal_file_buffer(struct journal_head *jh, |
2568 | transaction_t *transaction, int jlist) |
2569 | { |
2570 | spin_lock(lock: &jh->b_state_lock); |
2571 | spin_lock(lock: &transaction->t_journal->j_list_lock); |
2572 | __jbd2_journal_file_buffer(jh, transaction, jlist); |
2573 | spin_unlock(lock: &transaction->t_journal->j_list_lock); |
2574 | spin_unlock(lock: &jh->b_state_lock); |
2575 | } |
2576 | |
2577 | /* |
2578 | * Remove a buffer from its current buffer list in preparation for |
2579 | * dropping it from its current transaction entirely. If the buffer has |
2580 | * already started to be used by a subsequent transaction, refile the |
2581 | * buffer on that transaction's metadata list. |
2582 | * |
2583 | * Called under j_list_lock |
2584 | * Called under jh->b_state_lock |
2585 | * |
2586 | * When this function returns true, there's no next transaction to refile to |
2587 | * and the caller has to drop jh reference through |
2588 | * jbd2_journal_put_journal_head(). |
2589 | */ |
2590 | bool __jbd2_journal_refile_buffer(struct journal_head *jh) |
2591 | { |
2592 | int was_dirty, jlist; |
2593 | struct buffer_head *bh = jh2bh(jh); |
2594 | |
2595 | lockdep_assert_held(&jh->b_state_lock); |
2596 | if (jh->b_transaction) |
2597 | assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock); |
2598 | |
2599 | /* If the buffer is now unused, just drop it. */ |
2600 | if (jh->b_next_transaction == NULL) { |
2601 | __jbd2_journal_unfile_buffer(jh); |
2602 | return true; |
2603 | } |
2604 | |
2605 | /* |
2606 | * It has been modified by a later transaction: add it to the new |
2607 | * transaction's metadata list. |
2608 | */ |
2609 | |
2610 | was_dirty = test_clear_buffer_jbddirty(bh); |
2611 | __jbd2_journal_temp_unlink_buffer(jh); |
2612 | |
2613 | /* |
2614 | * b_transaction must be set, otherwise the new b_transaction won't |
2615 | * be holding jh reference |
2616 | */ |
2617 | J_ASSERT_JH(jh, jh->b_transaction != NULL); |
2618 | |
2619 | /* |
2620 | * We set b_transaction here because b_next_transaction will inherit |
2621 | * our jh reference and thus __jbd2_journal_file_buffer() must not |
2622 | * take a new one. |
2623 | */ |
2624 | WRITE_ONCE(jh->b_transaction, jh->b_next_transaction); |
2625 | WRITE_ONCE(jh->b_next_transaction, NULL); |
2626 | if (buffer_freed(bh)) |
2627 | jlist = BJ_Forget; |
2628 | else if (jh->b_modified) |
2629 | jlist = BJ_Metadata; |
2630 | else |
2631 | jlist = BJ_Reserved; |
2632 | __jbd2_journal_file_buffer(jh, transaction: jh->b_transaction, jlist); |
2633 | J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING); |
2634 | |
2635 | if (was_dirty) |
2636 | set_buffer_jbddirty(bh); |
2637 | return false; |
2638 | } |
2639 | |
2640 | /* |
2641 | * __jbd2_journal_refile_buffer() with necessary locking added. We take our |
2642 | * bh reference so that we can safely unlock bh. |
2643 | * |
2644 | * The jh and bh may be freed by this call. |
2645 | */ |
2646 | void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh) |
2647 | { |
2648 | bool drop; |
2649 | |
2650 | spin_lock(lock: &jh->b_state_lock); |
2651 | spin_lock(lock: &journal->j_list_lock); |
2652 | drop = __jbd2_journal_refile_buffer(jh); |
2653 | spin_unlock(lock: &jh->b_state_lock); |
2654 | spin_unlock(lock: &journal->j_list_lock); |
2655 | if (drop) |
2656 | jbd2_journal_put_journal_head(jh); |
2657 | } |
2658 | |
2659 | /* |
2660 | * File inode in the inode list of the handle's transaction |
2661 | */ |
2662 | static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode, |
2663 | unsigned long flags, loff_t start_byte, loff_t end_byte) |
2664 | { |
2665 | transaction_t *transaction = handle->h_transaction; |
2666 | journal_t *journal; |
2667 | |
2668 | if (is_handle_aborted(handle)) |
2669 | return -EROFS; |
2670 | journal = transaction->t_journal; |
2671 | |
2672 | jbd2_debug(4, "Adding inode %lu, tid:%d\n" , jinode->i_vfs_inode->i_ino, |
2673 | transaction->t_tid); |
2674 | |
2675 | spin_lock(lock: &journal->j_list_lock); |
2676 | jinode->i_flags |= flags; |
2677 | |
2678 | if (jinode->i_dirty_end) { |
2679 | jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte); |
2680 | jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte); |
2681 | } else { |
2682 | jinode->i_dirty_start = start_byte; |
2683 | jinode->i_dirty_end = end_byte; |
2684 | } |
2685 | |
2686 | /* Is inode already attached where we need it? */ |
2687 | if (jinode->i_transaction == transaction || |
2688 | jinode->i_next_transaction == transaction) |
2689 | goto done; |
2690 | |
2691 | /* |
2692 | * We only ever set this variable to 1 so the test is safe. Since |
2693 | * t_need_data_flush is likely to be set, we do the test to save some |
2694 | * cacheline bouncing |
2695 | */ |
2696 | if (!transaction->t_need_data_flush) |
2697 | transaction->t_need_data_flush = 1; |
2698 | /* On some different transaction's list - should be |
2699 | * the committing one */ |
2700 | if (jinode->i_transaction) { |
2701 | J_ASSERT(jinode->i_next_transaction == NULL); |
2702 | J_ASSERT(jinode->i_transaction == |
2703 | journal->j_committing_transaction); |
2704 | jinode->i_next_transaction = transaction; |
2705 | goto done; |
2706 | } |
2707 | /* Not on any transaction list... */ |
2708 | J_ASSERT(!jinode->i_next_transaction); |
2709 | jinode->i_transaction = transaction; |
2710 | list_add(new: &jinode->i_list, head: &transaction->t_inode_list); |
2711 | done: |
2712 | spin_unlock(lock: &journal->j_list_lock); |
2713 | |
2714 | return 0; |
2715 | } |
2716 | |
2717 | int jbd2_journal_inode_ranged_write(handle_t *handle, |
2718 | struct jbd2_inode *jinode, loff_t start_byte, loff_t length) |
2719 | { |
2720 | return jbd2_journal_file_inode(handle, jinode, |
2721 | JI_WRITE_DATA | JI_WAIT_DATA, start_byte, |
2722 | end_byte: start_byte + length - 1); |
2723 | } |
2724 | |
2725 | int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode, |
2726 | loff_t start_byte, loff_t length) |
2727 | { |
2728 | return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA, |
2729 | start_byte, end_byte: start_byte + length - 1); |
2730 | } |
2731 | |
2732 | /* |
2733 | * File truncate and transaction commit interact with each other in a |
2734 | * non-trivial way. If a transaction writing data block A is |
2735 | * committing, we cannot discard the data by truncate until we have |
2736 | * written them. Otherwise if we crashed after the transaction with |
2737 | * write has committed but before the transaction with truncate has |
2738 | * committed, we could see stale data in block A. This function is a |
2739 | * helper to solve this problem. It starts writeout of the truncated |
2740 | * part in case it is in the committing transaction. |
2741 | * |
2742 | * Filesystem code must call this function when inode is journaled in |
2743 | * ordered mode before truncation happens and after the inode has been |
2744 | * placed on orphan list with the new inode size. The second condition |
2745 | * avoids the race that someone writes new data and we start |
2746 | * committing the transaction after this function has been called but |
2747 | * before a transaction for truncate is started (and furthermore it |
2748 | * allows us to optimize the case where the addition to orphan list |
2749 | * happens in the same transaction as write --- we don't have to write |
2750 | * any data in such case). |
2751 | */ |
2752 | int jbd2_journal_begin_ordered_truncate(journal_t *journal, |
2753 | struct jbd2_inode *jinode, |
2754 | loff_t new_size) |
2755 | { |
2756 | transaction_t *inode_trans, *commit_trans; |
2757 | int ret = 0; |
2758 | |
2759 | /* This is a quick check to avoid locking if not necessary */ |
2760 | if (!jinode->i_transaction) |
2761 | goto out; |
2762 | /* Locks are here just to force reading of recent values, it is |
2763 | * enough that the transaction was not committing before we started |
2764 | * a transaction adding the inode to orphan list */ |
2765 | read_lock(&journal->j_state_lock); |
2766 | commit_trans = journal->j_committing_transaction; |
2767 | read_unlock(&journal->j_state_lock); |
2768 | spin_lock(lock: &journal->j_list_lock); |
2769 | inode_trans = jinode->i_transaction; |
2770 | spin_unlock(lock: &journal->j_list_lock); |
2771 | if (inode_trans == commit_trans) { |
2772 | ret = filemap_fdatawrite_range(mapping: jinode->i_vfs_inode->i_mapping, |
2773 | start: new_size, LLONG_MAX); |
2774 | if (ret) |
2775 | jbd2_journal_abort(journal, ret); |
2776 | } |
2777 | out: |
2778 | return ret; |
2779 | } |
2780 | |