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;
jbd2_journal_init_transaction_cache(void)37 int __init jbd2_journal_init_transaction_cache(void)
38 {
39 	J_ASSERT(!transaction_cache);
40 	transaction_cache = kmem_cache_create("jbd2_transaction_s",
41 					sizeof(transaction_t),
42 					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 
jbd2_journal_destroy_transaction_cache(void)52 void jbd2_journal_destroy_transaction_cache(void)
53 {
54 	kmem_cache_destroy(transaction_cache);
55 	transaction_cache = NULL;
56 }
57 
jbd2_journal_free_transaction(transaction_t * transaction)58 void jbd2_journal_free_transaction(transaction_t *transaction)
59 {
60 	if (unlikely(ZERO_OR_NULL_PTR(transaction)))
61 		return;
62 	kmem_cache_free(transaction_cache, transaction);
63 }
64 
65 /*
66  * Base amount of descriptor blocks we reserve for each transaction.
67  */
jbd2_descriptor_blocks_per_trans(journal_t * journal)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 
jbd2_get_transaction(journal_t * journal,transaction_t * transaction)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(&transaction->t_updates, 0);
111 	atomic_set(&transaction->t_outstanding_credits,
112 		   jbd2_descriptor_blocks_per_trans(journal) +
113 		   atomic_read(&journal->j_reserved_credits));
114 	atomic_set(&transaction->t_outstanding_revokes, 0);
115 	atomic_set(&transaction->t_handle_count, 0);
116 	INIT_LIST_HEAD(&transaction->t_inode_list);
117 	INIT_LIST_HEAD(&transaction->t_private_list);
118 
119 	/* Set up the commit timer for the new transaction. */
120 	journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
121 	add_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  */
update_t_max_wait(transaction_t * transaction,unsigned long ts)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(ts, 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  */
wait_transaction_locked(journal_t * journal)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(&journal->j_wait_transaction_locked, &wait,
172 			TASK_UNINTERRUPTIBLE);
173 	need_to_start = !tid_geq(journal->j_commit_request, 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(&journal->j_wait_transaction_locked, &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  */
wait_transaction_switching(journal_t * journal)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(&journal->j_wait_transaction_locked, &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(&journal->j_wait_transaction_locked, &wait);
208 }
209 
sub_reserved_credits(journal_t * journal,int blocks)210 static void sub_reserved_credits(journal_t *journal, int blocks)
211 {
212 	atomic_sub(blocks, &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  */
add_transaction_credits(journal_t * journal,int blocks,int rsv_blocks)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(total, &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(total, &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(&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(total, &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(rsv_blocks, &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, rsv_blocks);
312 		atomic_sub(total, &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 
start_this_handle(journal_t * journal,handle_t * handle,gfp_t gfp_mask)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(transaction_cache,
371 						    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(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, 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(&transaction->t_updates);
455 	atomic_inc(&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(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... */
new_handle(int nblocks)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 
jbd2__journal_start(journal_t * journal,int nblocks,int rsv_blocks,int revoke_records,gfp_t gfp_mask,unsigned int type,unsigned int line_no)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(-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(-ENOMEM);
506 	if (rsv_blocks) {
507 		handle_t *rsv_handle;
508 
509 		rsv_handle = new_handle(rsv_blocks);
510 		if (!rsv_handle) {
511 			jbd2_free_handle(handle);
512 			return ERR_PTR(-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->h_rsv_handle);
524 		jbd2_free_handle(handle);
525 		return ERR_PTR(err);
526 	}
527 	handle->h_type = type;
528 	handle->h_line_no = line_no;
529 	trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
530 				handle->h_transaction->t_tid, type,
531 				line_no, 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  */
jbd2_journal_start(journal_t * journal,int nblocks)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 
__jbd2_journal_unreserve_handle(handle_t * handle,transaction_t * t)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, handle->h_total_credits);
569 	if (t)
570 		atomic_sub(handle->h_total_credits, &t->t_outstanding_credits);
571 }
572 
jbd2_journal_free_reserved(handle_t * handle)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, 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  */
jbd2_journal_start_reserved(handle_t * handle,unsigned int type,unsigned int line_no)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(journal->j_fs_dev->bd_dev,
633 				handle->h_transaction->t_tid, type,
634 				line_no, 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  */
jbd2_journal_extend(handle_t * handle,int nblocks,int revoke_records)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(nblocks,
689 				   &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(nblocks, &transaction->t_outstanding_credits);
695 		goto error_out;
696 	}
697 
698 	trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
699 				 transaction->t_tid,
700 				 handle->h_type, handle->h_line_no,
701 				 handle->h_total_credits,
702 				 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 
stop_this_handle(handle_t * handle)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(revokes,
739 				&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(handle->h_total_credits,
746 		   &transaction->t_outstanding_credits);
747 	if (handle->h_rsv_handle)
748 		__jbd2_journal_unreserve_handle(handle->h_rsv_handle,
749 						transaction);
750 	if (atomic_dec_and_test(&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(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  */
jbd2__journal_restart(handle_t * handle,int nblocks,int revoke_records,gfp_t gfp_mask)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(journal->j_commit_request, 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(journal->j_fs_dev->bd_dev,
817 				 ret ? 0 : handle->h_transaction->t_tid,
818 				 handle->h_type, handle->h_line_no,
819 				 handle->h_total_credits);
820 	return ret;
821 }
822 EXPORT_SYMBOL(jbd2__journal_restart);
823 
824 
jbd2_journal_restart(handle_t * handle,int nblocks)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  */
jbd2_journal_wait_updates(journal_t * journal)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(&journal->j_wait_updates, &wait,
856 				TASK_UNINTERRUPTIBLE);
857 		if (!atomic_read(&transaction->t_updates)) {
858 			finish_wait(&journal->j_wait_updates, &wait);
859 			break;
860 		}
861 		write_unlock(&journal->j_state_lock);
862 		schedule();
863 		finish_wait(&journal->j_wait_updates, &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  */
jbd2_journal_lock_updates(journal_t * journal)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(&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  */
jbd2_journal_unlock_updates(journal_t * journal)915 void jbd2_journal_unlock_updates (journal_t *journal)
916 {
917 	J_ASSERT(journal->j_barrier_count != 0);
918 
919 	mutex_unlock(&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 
warn_dirty_buffer(struct buffer_head * bh)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. */
jbd2_freeze_jh_data(struct journal_head * jh)936 static void jbd2_freeze_jh_data(struct journal_head *jh)
937 {
938 	struct page *page;
939 	int offset;
940 	char *source;
941 	struct buffer_head *bh = jh2bh(jh);
942 
943 	J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
944 	page = bh->b_page;
945 	offset = offset_in_page(bh->b_data);
946 	source = kmap_atomic(page);
947 	/* Fire data frozen trigger just before we copy the data */
948 	jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
949 	memcpy(jh->b_frozen_data, source + offset, bh->b_size);
950 	kunmap_atomic(source);
951 
952 	/*
953 	 * Now that the frozen data is saved off, we need to store any matching
954 	 * triggers.
955 	 */
956 	jh->b_frozen_triggers = jh->b_triggers;
957 }
958 
959 /*
960  * If the buffer is already part of the current transaction, then there
961  * is nothing we need to do.  If it is already part of a prior
962  * transaction which we are still committing to disk, then we need to
963  * make sure that we do not overwrite the old copy: we do copy-out to
964  * preserve the copy going to disk.  We also account the buffer against
965  * the handle's metadata buffer credits (unless the buffer is already
966  * part of the transaction, that is).
967  *
968  */
969 static int
do_get_write_access(handle_t * handle,struct journal_head * jh,int force_copy)970 do_get_write_access(handle_t *handle, struct journal_head *jh,
971 			int force_copy)
972 {
973 	struct buffer_head *bh;
974 	transaction_t *transaction = handle->h_transaction;
975 	journal_t *journal;
976 	int error;
977 	char *frozen_buffer = NULL;
978 	unsigned long start_lock, time_lock;
979 
980 	journal = transaction->t_journal;
981 
982 	jbd2_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
983 
984 	JBUFFER_TRACE(jh, "entry");
985 repeat:
986 	bh = jh2bh(jh);
987 
988 	/* @@@ Need to check for errors here at some point. */
989 
990  	start_lock = jiffies;
991 	lock_buffer(bh);
992 	spin_lock(&jh->b_state_lock);
993 
994 	/* If it takes too long to lock the buffer, trace it */
995 	time_lock = jbd2_time_diff(start_lock, jiffies);
996 	if (time_lock > HZ/10)
997 		trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
998 			jiffies_to_msecs(time_lock));
999 
1000 	/* We now hold the buffer lock so it is safe to query the buffer
1001 	 * state.  Is the buffer dirty?
1002 	 *
1003 	 * If so, there are two possibilities.  The buffer may be
1004 	 * non-journaled, and undergoing a quite legitimate writeback.
1005 	 * Otherwise, it is journaled, and we don't expect dirty buffers
1006 	 * in that state (the buffers should be marked JBD_Dirty
1007 	 * instead.)  So either the IO is being done under our own
1008 	 * control and this is a bug, or it's a third party IO such as
1009 	 * dump(8) (which may leave the buffer scheduled for read ---
1010 	 * ie. locked but not dirty) or tune2fs (which may actually have
1011 	 * the buffer dirtied, ugh.)  */
1012 
1013 	if (buffer_dirty(bh)) {
1014 		/*
1015 		 * First question: is this buffer already part of the current
1016 		 * transaction or the existing committing transaction?
1017 		 */
1018 		if (jh->b_transaction) {
1019 			J_ASSERT_JH(jh,
1020 				jh->b_transaction == transaction ||
1021 				jh->b_transaction ==
1022 					journal->j_committing_transaction);
1023 			if (jh->b_next_transaction)
1024 				J_ASSERT_JH(jh, jh->b_next_transaction ==
1025 							transaction);
1026 			warn_dirty_buffer(bh);
1027 		}
1028 		/*
1029 		 * In any case we need to clean the dirty flag and we must
1030 		 * do it under the buffer lock to be sure we don't race
1031 		 * with running write-out.
1032 		 */
1033 		JBUFFER_TRACE(jh, "Journalling dirty buffer");
1034 		clear_buffer_dirty(bh);
1035 		set_buffer_jbddirty(bh);
1036 	}
1037 
1038 	unlock_buffer(bh);
1039 
1040 	error = -EROFS;
1041 	if (is_handle_aborted(handle)) {
1042 		spin_unlock(&jh->b_state_lock);
1043 		goto out;
1044 	}
1045 	error = 0;
1046 
1047 	/*
1048 	 * The buffer is already part of this transaction if b_transaction or
1049 	 * b_next_transaction points to it
1050 	 */
1051 	if (jh->b_transaction == transaction ||
1052 	    jh->b_next_transaction == transaction)
1053 		goto done;
1054 
1055 	/*
1056 	 * this is the first time this transaction is touching this buffer,
1057 	 * reset the modified flag
1058 	 */
1059 	jh->b_modified = 0;
1060 
1061 	/*
1062 	 * If the buffer is not journaled right now, we need to make sure it
1063 	 * doesn't get written to disk before the caller actually commits the
1064 	 * new data
1065 	 */
1066 	if (!jh->b_transaction) {
1067 		JBUFFER_TRACE(jh, "no transaction");
1068 		J_ASSERT_JH(jh, !jh->b_next_transaction);
1069 		JBUFFER_TRACE(jh, "file as BJ_Reserved");
1070 		/*
1071 		 * Make sure all stores to jh (b_modified, b_frozen_data) are
1072 		 * visible before attaching it to the running transaction.
1073 		 * Paired with barrier in jbd2_write_access_granted()
1074 		 */
1075 		smp_wmb();
1076 		spin_lock(&journal->j_list_lock);
1077 		__jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1078 		spin_unlock(&journal->j_list_lock);
1079 		goto done;
1080 	}
1081 	/*
1082 	 * If there is already a copy-out version of this buffer, then we don't
1083 	 * need to make another one
1084 	 */
1085 	if (jh->b_frozen_data) {
1086 		JBUFFER_TRACE(jh, "has frozen data");
1087 		J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1088 		goto attach_next;
1089 	}
1090 
1091 	JBUFFER_TRACE(jh, "owned by older transaction");
1092 	J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1093 	J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
1094 
1095 	/*
1096 	 * There is one case we have to be very careful about.  If the
1097 	 * committing transaction is currently writing this buffer out to disk
1098 	 * and has NOT made a copy-out, then we cannot modify the buffer
1099 	 * contents at all right now.  The essence of copy-out is that it is
1100 	 * the extra copy, not the primary copy, which gets journaled.  If the
1101 	 * primary copy is already going to disk then we cannot do copy-out
1102 	 * here.
1103 	 */
1104 	if (buffer_shadow(bh)) {
1105 		JBUFFER_TRACE(jh, "on shadow: sleep");
1106 		spin_unlock(&jh->b_state_lock);
1107 		wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
1108 		goto repeat;
1109 	}
1110 
1111 	/*
1112 	 * Only do the copy if the currently-owning transaction still needs it.
1113 	 * If buffer isn't on BJ_Metadata list, the committing transaction is
1114 	 * past that stage (here we use the fact that BH_Shadow is set under
1115 	 * bh_state lock together with refiling to BJ_Shadow list and at this
1116 	 * point we know the buffer doesn't have BH_Shadow set).
1117 	 *
1118 	 * Subtle point, though: if this is a get_undo_access, then we will be
1119 	 * relying on the frozen_data to contain the new value of the
1120 	 * committed_data record after the transaction, so we HAVE to force the
1121 	 * frozen_data copy in that case.
1122 	 */
1123 	if (jh->b_jlist == BJ_Metadata || force_copy) {
1124 		JBUFFER_TRACE(jh, "generate frozen data");
1125 		if (!frozen_buffer) {
1126 			JBUFFER_TRACE(jh, "allocate memory for buffer");
1127 			spin_unlock(&jh->b_state_lock);
1128 			frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
1129 						   GFP_NOFS | __GFP_NOFAIL);
1130 			goto repeat;
1131 		}
1132 		jh->b_frozen_data = frozen_buffer;
1133 		frozen_buffer = NULL;
1134 		jbd2_freeze_jh_data(jh);
1135 	}
1136 attach_next:
1137 	/*
1138 	 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1139 	 * before attaching it to the running transaction. Paired with barrier
1140 	 * in jbd2_write_access_granted()
1141 	 */
1142 	smp_wmb();
1143 	jh->b_next_transaction = transaction;
1144 
1145 done:
1146 	spin_unlock(&jh->b_state_lock);
1147 
1148 	/*
1149 	 * If we are about to journal a buffer, then any revoke pending on it is
1150 	 * no longer valid
1151 	 */
1152 	jbd2_journal_cancel_revoke(handle, jh);
1153 
1154 out:
1155 	if (unlikely(frozen_buffer))	/* It's usually NULL */
1156 		jbd2_free(frozen_buffer, bh->b_size);
1157 
1158 	JBUFFER_TRACE(jh, "exit");
1159 	return error;
1160 }
1161 
1162 /* Fast check whether buffer is already attached to the required transaction */
jbd2_write_access_granted(handle_t * handle,struct buffer_head * bh,bool undo)1163 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1164 							bool undo)
1165 {
1166 	struct journal_head *jh;
1167 	bool ret = false;
1168 
1169 	/* Dirty buffers require special handling... */
1170 	if (buffer_dirty(bh))
1171 		return false;
1172 
1173 	/*
1174 	 * RCU protects us from dereferencing freed pages. So the checks we do
1175 	 * are guaranteed not to oops. However the jh slab object can get freed
1176 	 * & reallocated while we work with it. So we have to be careful. When
1177 	 * we see jh attached to the running transaction, we know it must stay
1178 	 * so until the transaction is committed. Thus jh won't be freed and
1179 	 * will be attached to the same bh while we run.  However it can
1180 	 * happen jh gets freed, reallocated, and attached to the transaction
1181 	 * just after we get pointer to it from bh. So we have to be careful
1182 	 * and recheck jh still belongs to our bh before we return success.
1183 	 */
1184 	rcu_read_lock();
1185 	if (!buffer_jbd(bh))
1186 		goto out;
1187 	/* This should be bh2jh() but that doesn't work with inline functions */
1188 	jh = READ_ONCE(bh->b_private);
1189 	if (!jh)
1190 		goto out;
1191 	/* For undo access buffer must have data copied */
1192 	if (undo && !jh->b_committed_data)
1193 		goto out;
1194 	if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
1195 	    READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
1196 		goto out;
1197 	/*
1198 	 * There are two reasons for the barrier here:
1199 	 * 1) Make sure to fetch b_bh after we did previous checks so that we
1200 	 * detect when jh went through free, realloc, attach to transaction
1201 	 * while we were checking. Paired with implicit barrier in that path.
1202 	 * 2) So that access to bh done after jbd2_write_access_granted()
1203 	 * doesn't get reordered and see inconsistent state of concurrent
1204 	 * do_get_write_access().
1205 	 */
1206 	smp_mb();
1207 	if (unlikely(jh->b_bh != bh))
1208 		goto out;
1209 	ret = true;
1210 out:
1211 	rcu_read_unlock();
1212 	return ret;
1213 }
1214 
1215 /**
1216  * jbd2_journal_get_write_access() - notify intent to modify a buffer
1217  *				     for metadata (not data) update.
1218  * @handle: transaction to add buffer modifications to
1219  * @bh:     bh to be used for metadata writes
1220  *
1221  * Returns: error code or 0 on success.
1222  *
1223  * In full data journalling mode the buffer may be of type BJ_AsyncData,
1224  * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1225  */
1226 
jbd2_journal_get_write_access(handle_t * handle,struct buffer_head * bh)1227 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1228 {
1229 	struct journal_head *jh;
1230 	int rc;
1231 
1232 	if (is_handle_aborted(handle))
1233 		return -EROFS;
1234 
1235 	if (jbd2_write_access_granted(handle, bh, false))
1236 		return 0;
1237 
1238 	jh = jbd2_journal_add_journal_head(bh);
1239 	/* We do not want to get caught playing with fields which the
1240 	 * log thread also manipulates.  Make sure that the buffer
1241 	 * completes any outstanding IO before proceeding. */
1242 	rc = do_get_write_access(handle, jh, 0);
1243 	jbd2_journal_put_journal_head(jh);
1244 	return rc;
1245 }
1246 
1247 
1248 /*
1249  * When the user wants to journal a newly created buffer_head
1250  * (ie. getblk() returned a new buffer and we are going to populate it
1251  * manually rather than reading off disk), then we need to keep the
1252  * buffer_head locked until it has been completely filled with new
1253  * data.  In this case, we should be able to make the assertion that
1254  * the bh is not already part of an existing transaction.
1255  *
1256  * The buffer should already be locked by the caller by this point.
1257  * There is no lock ranking violation: it was a newly created,
1258  * unlocked buffer beforehand. */
1259 
1260 /**
1261  * jbd2_journal_get_create_access () - notify intent to use newly created bh
1262  * @handle: transaction to new buffer to
1263  * @bh: new buffer.
1264  *
1265  * Call this if you create a new bh.
1266  */
jbd2_journal_get_create_access(handle_t * handle,struct buffer_head * bh)1267 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1268 {
1269 	transaction_t *transaction = handle->h_transaction;
1270 	journal_t *journal;
1271 	struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1272 	int err;
1273 
1274 	jbd2_debug(5, "journal_head %p\n", jh);
1275 	err = -EROFS;
1276 	if (is_handle_aborted(handle))
1277 		goto out;
1278 	journal = transaction->t_journal;
1279 	err = 0;
1280 
1281 	JBUFFER_TRACE(jh, "entry");
1282 	/*
1283 	 * The buffer may already belong to this transaction due to pre-zeroing
1284 	 * in the filesystem's new_block code.  It may also be on the previous,
1285 	 * committing transaction's lists, but it HAS to be in Forget state in
1286 	 * that case: the transaction must have deleted the buffer for it to be
1287 	 * reused here.
1288 	 */
1289 	spin_lock(&jh->b_state_lock);
1290 	J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1291 		jh->b_transaction == NULL ||
1292 		(jh->b_transaction == journal->j_committing_transaction &&
1293 			  jh->b_jlist == BJ_Forget)));
1294 
1295 	J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1296 	J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1297 
1298 	if (jh->b_transaction == NULL) {
1299 		/*
1300 		 * Previous jbd2_journal_forget() could have left the buffer
1301 		 * with jbddirty bit set because it was being committed. When
1302 		 * the commit finished, we've filed the buffer for
1303 		 * checkpointing and marked it dirty. Now we are reallocating
1304 		 * the buffer so the transaction freeing it must have
1305 		 * committed and so it's safe to clear the dirty bit.
1306 		 */
1307 		clear_buffer_dirty(jh2bh(jh));
1308 		/* first access by this transaction */
1309 		jh->b_modified = 0;
1310 
1311 		JBUFFER_TRACE(jh, "file as BJ_Reserved");
1312 		spin_lock(&journal->j_list_lock);
1313 		__jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1314 		spin_unlock(&journal->j_list_lock);
1315 	} else if (jh->b_transaction == journal->j_committing_transaction) {
1316 		/* first access by this transaction */
1317 		jh->b_modified = 0;
1318 
1319 		JBUFFER_TRACE(jh, "set next transaction");
1320 		spin_lock(&journal->j_list_lock);
1321 		jh->b_next_transaction = transaction;
1322 		spin_unlock(&journal->j_list_lock);
1323 	}
1324 	spin_unlock(&jh->b_state_lock);
1325 
1326 	/*
1327 	 * akpm: I added this.  ext3_alloc_branch can pick up new indirect
1328 	 * blocks which contain freed but then revoked metadata.  We need
1329 	 * to cancel the revoke in case we end up freeing it yet again
1330 	 * and the reallocating as data - this would cause a second revoke,
1331 	 * which hits an assertion error.
1332 	 */
1333 	JBUFFER_TRACE(jh, "cancelling revoke");
1334 	jbd2_journal_cancel_revoke(handle, jh);
1335 out:
1336 	jbd2_journal_put_journal_head(jh);
1337 	return err;
1338 }
1339 
1340 /**
1341  * jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
1342  *     non-rewindable consequences
1343  * @handle: transaction
1344  * @bh: buffer to undo
1345  *
1346  * Sometimes there is a need to distinguish between metadata which has
1347  * been committed to disk and that which has not.  The ext3fs code uses
1348  * this for freeing and allocating space, we have to make sure that we
1349  * do not reuse freed space until the deallocation has been committed,
1350  * since if we overwrote that space we would make the delete
1351  * un-rewindable in case of a crash.
1352  *
1353  * To deal with that, jbd2_journal_get_undo_access requests write access to a
1354  * buffer for parts of non-rewindable operations such as delete
1355  * operations on the bitmaps.  The journaling code must keep a copy of
1356  * the buffer's contents prior to the undo_access call until such time
1357  * as we know that the buffer has definitely been committed to disk.
1358  *
1359  * We never need to know which transaction the committed data is part
1360  * of, buffers touched here are guaranteed to be dirtied later and so
1361  * will be committed to a new transaction in due course, at which point
1362  * we can discard the old committed data pointer.
1363  *
1364  * Returns error number or 0 on success.
1365  */
jbd2_journal_get_undo_access(handle_t * handle,struct buffer_head * bh)1366 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1367 {
1368 	int err;
1369 	struct journal_head *jh;
1370 	char *committed_data = NULL;
1371 
1372 	if (is_handle_aborted(handle))
1373 		return -EROFS;
1374 
1375 	if (jbd2_write_access_granted(handle, bh, true))
1376 		return 0;
1377 
1378 	jh = jbd2_journal_add_journal_head(bh);
1379 	JBUFFER_TRACE(jh, "entry");
1380 
1381 	/*
1382 	 * Do this first --- it can drop the journal lock, so we want to
1383 	 * make sure that obtaining the committed_data is done
1384 	 * atomically wrt. completion of any outstanding commits.
1385 	 */
1386 	err = do_get_write_access(handle, jh, 1);
1387 	if (err)
1388 		goto out;
1389 
1390 repeat:
1391 	if (!jh->b_committed_data)
1392 		committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1393 					    GFP_NOFS|__GFP_NOFAIL);
1394 
1395 	spin_lock(&jh->b_state_lock);
1396 	if (!jh->b_committed_data) {
1397 		/* Copy out the current buffer contents into the
1398 		 * preserved, committed copy. */
1399 		JBUFFER_TRACE(jh, "generate b_committed data");
1400 		if (!committed_data) {
1401 			spin_unlock(&jh->b_state_lock);
1402 			goto repeat;
1403 		}
1404 
1405 		jh->b_committed_data = committed_data;
1406 		committed_data = NULL;
1407 		memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1408 	}
1409 	spin_unlock(&jh->b_state_lock);
1410 out:
1411 	jbd2_journal_put_journal_head(jh);
1412 	if (unlikely(committed_data))
1413 		jbd2_free(committed_data, bh->b_size);
1414 	return err;
1415 }
1416 
1417 /**
1418  * jbd2_journal_set_triggers() - Add triggers for commit writeout
1419  * @bh: buffer to trigger on
1420  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1421  *
1422  * Set any triggers on this journal_head.  This is always safe, because
1423  * triggers for a committing buffer will be saved off, and triggers for
1424  * a running transaction will match the buffer in that transaction.
1425  *
1426  * Call with NULL to clear the triggers.
1427  */
jbd2_journal_set_triggers(struct buffer_head * bh,struct jbd2_buffer_trigger_type * type)1428 void jbd2_journal_set_triggers(struct buffer_head *bh,
1429 			       struct jbd2_buffer_trigger_type *type)
1430 {
1431 	struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1432 
1433 	if (WARN_ON_ONCE(!jh))
1434 		return;
1435 	jh->b_triggers = type;
1436 	jbd2_journal_put_journal_head(jh);
1437 }
1438 
jbd2_buffer_frozen_trigger(struct journal_head * jh,void * mapped_data,struct jbd2_buffer_trigger_type * triggers)1439 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1440 				struct jbd2_buffer_trigger_type *triggers)
1441 {
1442 	struct buffer_head *bh = jh2bh(jh);
1443 
1444 	if (!triggers || !triggers->t_frozen)
1445 		return;
1446 
1447 	triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1448 }
1449 
jbd2_buffer_abort_trigger(struct journal_head * jh,struct jbd2_buffer_trigger_type * triggers)1450 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1451 			       struct jbd2_buffer_trigger_type *triggers)
1452 {
1453 	if (!triggers || !triggers->t_abort)
1454 		return;
1455 
1456 	triggers->t_abort(triggers, jh2bh(jh));
1457 }
1458 
1459 /**
1460  * jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1461  * @handle: transaction to add buffer to.
1462  * @bh: buffer to mark
1463  *
1464  * mark dirty metadata which needs to be journaled as part of the current
1465  * transaction.
1466  *
1467  * The buffer must have previously had jbd2_journal_get_write_access()
1468  * called so that it has a valid journal_head attached to the buffer
1469  * head.
1470  *
1471  * The buffer is placed on the transaction's metadata list and is marked
1472  * as belonging to the transaction.
1473  *
1474  * Returns error number or 0 on success.
1475  *
1476  * Special care needs to be taken if the buffer already belongs to the
1477  * current committing transaction (in which case we should have frozen
1478  * data present for that commit).  In that case, we don't relink the
1479  * buffer: that only gets done when the old transaction finally
1480  * completes its commit.
1481  */
jbd2_journal_dirty_metadata(handle_t * handle,struct buffer_head * bh)1482 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1483 {
1484 	transaction_t *transaction = handle->h_transaction;
1485 	journal_t *journal;
1486 	struct journal_head *jh;
1487 	int ret = 0;
1488 
1489 	if (!buffer_jbd(bh))
1490 		return -EUCLEAN;
1491 
1492 	/*
1493 	 * We don't grab jh reference here since the buffer must be part
1494 	 * of the running transaction.
1495 	 */
1496 	jh = bh2jh(bh);
1497 	jbd2_debug(5, "journal_head %p\n", jh);
1498 	JBUFFER_TRACE(jh, "entry");
1499 
1500 	/*
1501 	 * This and the following assertions are unreliable since we may see jh
1502 	 * in inconsistent state unless we grab bh_state lock. But this is
1503 	 * crucial to catch bugs so let's do a reliable check until the
1504 	 * lockless handling is fully proven.
1505 	 */
1506 	if (data_race(jh->b_transaction != transaction &&
1507 	    jh->b_next_transaction != transaction)) {
1508 		spin_lock(&jh->b_state_lock);
1509 		J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1510 				jh->b_next_transaction == transaction);
1511 		spin_unlock(&jh->b_state_lock);
1512 	}
1513 	if (jh->b_modified == 1) {
1514 		/* If it's in our transaction it must be in BJ_Metadata list. */
1515 		if (data_race(jh->b_transaction == transaction &&
1516 		    jh->b_jlist != BJ_Metadata)) {
1517 			spin_lock(&jh->b_state_lock);
1518 			if (jh->b_transaction == transaction &&
1519 			    jh->b_jlist != BJ_Metadata)
1520 				pr_err("JBD2: assertion failure: h_type=%u "
1521 				       "h_line_no=%u block_no=%llu jlist=%u\n",
1522 				       handle->h_type, handle->h_line_no,
1523 				       (unsigned long long) bh->b_blocknr,
1524 				       jh->b_jlist);
1525 			J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1526 					jh->b_jlist == BJ_Metadata);
1527 			spin_unlock(&jh->b_state_lock);
1528 		}
1529 		goto out;
1530 	}
1531 
1532 	journal = transaction->t_journal;
1533 	spin_lock(&jh->b_state_lock);
1534 
1535 	if (is_handle_aborted(handle)) {
1536 		/*
1537 		 * Check journal aborting with @jh->b_state_lock locked,
1538 		 * since 'jh->b_transaction' could be replaced with
1539 		 * 'jh->b_next_transaction' during old transaction
1540 		 * committing if journal aborted, which may fail
1541 		 * assertion on 'jh->b_frozen_data == NULL'.
1542 		 */
1543 		ret = -EROFS;
1544 		goto out_unlock_bh;
1545 	}
1546 
1547 	if (jh->b_modified == 0) {
1548 		/*
1549 		 * This buffer's got modified and becoming part
1550 		 * of the transaction. This needs to be done
1551 		 * once a transaction -bzzz
1552 		 */
1553 		if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle) <= 0)) {
1554 			ret = -ENOSPC;
1555 			goto out_unlock_bh;
1556 		}
1557 		jh->b_modified = 1;
1558 		handle->h_total_credits--;
1559 	}
1560 
1561 	/*
1562 	 * fastpath, to avoid expensive locking.  If this buffer is already
1563 	 * on the running transaction's metadata list there is nothing to do.
1564 	 * Nobody can take it off again because there is a handle open.
1565 	 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1566 	 * result in this test being false, so we go in and take the locks.
1567 	 */
1568 	if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1569 		JBUFFER_TRACE(jh, "fastpath");
1570 		if (unlikely(jh->b_transaction !=
1571 			     journal->j_running_transaction)) {
1572 			printk(KERN_ERR "JBD2: %s: "
1573 			       "jh->b_transaction (%llu, %p, %u) != "
1574 			       "journal->j_running_transaction (%p, %u)\n",
1575 			       journal->j_devname,
1576 			       (unsigned long long) bh->b_blocknr,
1577 			       jh->b_transaction,
1578 			       jh->b_transaction ? jh->b_transaction->t_tid : 0,
1579 			       journal->j_running_transaction,
1580 			       journal->j_running_transaction ?
1581 			       journal->j_running_transaction->t_tid : 0);
1582 			ret = -EINVAL;
1583 		}
1584 		goto out_unlock_bh;
1585 	}
1586 
1587 	set_buffer_jbddirty(bh);
1588 
1589 	/*
1590 	 * Metadata already on the current transaction list doesn't
1591 	 * need to be filed.  Metadata on another transaction's list must
1592 	 * be committing, and will be refiled once the commit completes:
1593 	 * leave it alone for now.
1594 	 */
1595 	if (jh->b_transaction != transaction) {
1596 		JBUFFER_TRACE(jh, "already on other transaction");
1597 		if (unlikely(((jh->b_transaction !=
1598 			       journal->j_committing_transaction)) ||
1599 			     (jh->b_next_transaction != transaction))) {
1600 			printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1601 			       "bad jh for block %llu: "
1602 			       "transaction (%p, %u), "
1603 			       "jh->b_transaction (%p, %u), "
1604 			       "jh->b_next_transaction (%p, %u), jlist %u\n",
1605 			       journal->j_devname,
1606 			       (unsigned long long) bh->b_blocknr,
1607 			       transaction, transaction->t_tid,
1608 			       jh->b_transaction,
1609 			       jh->b_transaction ?
1610 			       jh->b_transaction->t_tid : 0,
1611 			       jh->b_next_transaction,
1612 			       jh->b_next_transaction ?
1613 			       jh->b_next_transaction->t_tid : 0,
1614 			       jh->b_jlist);
1615 			WARN_ON(1);
1616 			ret = -EINVAL;
1617 		}
1618 		/* And this case is illegal: we can't reuse another
1619 		 * transaction's data buffer, ever. */
1620 		goto out_unlock_bh;
1621 	}
1622 
1623 	/* That test should have eliminated the following case: */
1624 	J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1625 
1626 	JBUFFER_TRACE(jh, "file as BJ_Metadata");
1627 	spin_lock(&journal->j_list_lock);
1628 	__jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1629 	spin_unlock(&journal->j_list_lock);
1630 out_unlock_bh:
1631 	spin_unlock(&jh->b_state_lock);
1632 out:
1633 	JBUFFER_TRACE(jh, "exit");
1634 	return ret;
1635 }
1636 
1637 /**
1638  * jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1639  * @handle: transaction handle
1640  * @bh:     bh to 'forget'
1641  *
1642  * We can only do the bforget if there are no commits pending against the
1643  * buffer.  If the buffer is dirty in the current running transaction we
1644  * can safely unlink it.
1645  *
1646  * bh may not be a journalled buffer at all - it may be a non-JBD
1647  * buffer which came off the hashtable.  Check for this.
1648  *
1649  * Decrements bh->b_count by one.
1650  *
1651  * Allow this call even if the handle has aborted --- it may be part of
1652  * the caller's cleanup after an abort.
1653  */
jbd2_journal_forget(handle_t * handle,struct buffer_head * bh)1654 int jbd2_journal_forget(handle_t *handle, struct buffer_head *bh)
1655 {
1656 	transaction_t *transaction = handle->h_transaction;
1657 	journal_t *journal;
1658 	struct journal_head *jh;
1659 	int drop_reserve = 0;
1660 	int err = 0;
1661 	int was_modified = 0;
1662 
1663 	if (is_handle_aborted(handle))
1664 		return -EROFS;
1665 	journal = transaction->t_journal;
1666 
1667 	BUFFER_TRACE(bh, "entry");
1668 
1669 	jh = jbd2_journal_grab_journal_head(bh);
1670 	if (!jh) {
1671 		__bforget(bh);
1672 		return 0;
1673 	}
1674 
1675 	spin_lock(&jh->b_state_lock);
1676 
1677 	/* Critical error: attempting to delete a bitmap buffer, maybe?
1678 	 * Don't do any jbd operations, and return an error. */
1679 	if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1680 			 "inconsistent data on disk")) {
1681 		err = -EIO;
1682 		goto drop;
1683 	}
1684 
1685 	/* keep track of whether or not this transaction modified us */
1686 	was_modified = jh->b_modified;
1687 
1688 	/*
1689 	 * The buffer's going from the transaction, we must drop
1690 	 * all references -bzzz
1691 	 */
1692 	jh->b_modified = 0;
1693 
1694 	if (jh->b_transaction == transaction) {
1695 		J_ASSERT_JH(jh, !jh->b_frozen_data);
1696 
1697 		/* If we are forgetting a buffer which is already part
1698 		 * of this transaction, then we can just drop it from
1699 		 * the transaction immediately. */
1700 		clear_buffer_dirty(bh);
1701 		clear_buffer_jbddirty(bh);
1702 
1703 		JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1704 
1705 		/*
1706 		 * we only want to drop a reference if this transaction
1707 		 * modified the buffer
1708 		 */
1709 		if (was_modified)
1710 			drop_reserve = 1;
1711 
1712 		/*
1713 		 * We are no longer going to journal this buffer.
1714 		 * However, the commit of this transaction is still
1715 		 * important to the buffer: the delete that we are now
1716 		 * processing might obsolete an old log entry, so by
1717 		 * committing, we can satisfy the buffer's checkpoint.
1718 		 *
1719 		 * So, if we have a checkpoint on the buffer, we should
1720 		 * now refile the buffer on our BJ_Forget list so that
1721 		 * we know to remove the checkpoint after we commit.
1722 		 */
1723 
1724 		spin_lock(&journal->j_list_lock);
1725 		if (jh->b_cp_transaction) {
1726 			__jbd2_journal_temp_unlink_buffer(jh);
1727 			__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1728 		} else {
1729 			__jbd2_journal_unfile_buffer(jh);
1730 			jbd2_journal_put_journal_head(jh);
1731 		}
1732 		spin_unlock(&journal->j_list_lock);
1733 	} else if (jh->b_transaction) {
1734 		J_ASSERT_JH(jh, (jh->b_transaction ==
1735 				 journal->j_committing_transaction));
1736 		/* However, if the buffer is still owned by a prior
1737 		 * (committing) transaction, we can't drop it yet... */
1738 		JBUFFER_TRACE(jh, "belongs to older transaction");
1739 		/* ... but we CAN drop it from the new transaction through
1740 		 * marking the buffer as freed and set j_next_transaction to
1741 		 * the new transaction, so that not only the commit code
1742 		 * knows it should clear dirty bits when it is done with the
1743 		 * buffer, but also the buffer can be checkpointed only
1744 		 * after the new transaction commits. */
1745 
1746 		set_buffer_freed(bh);
1747 
1748 		if (!jh->b_next_transaction) {
1749 			spin_lock(&journal->j_list_lock);
1750 			jh->b_next_transaction = transaction;
1751 			spin_unlock(&journal->j_list_lock);
1752 		} else {
1753 			J_ASSERT(jh->b_next_transaction == transaction);
1754 
1755 			/*
1756 			 * only drop a reference if this transaction modified
1757 			 * the buffer
1758 			 */
1759 			if (was_modified)
1760 				drop_reserve = 1;
1761 		}
1762 	} else {
1763 		/*
1764 		 * Finally, if the buffer is not belongs to any
1765 		 * transaction, we can just drop it now if it has no
1766 		 * checkpoint.
1767 		 */
1768 		spin_lock(&journal->j_list_lock);
1769 		if (!jh->b_cp_transaction) {
1770 			JBUFFER_TRACE(jh, "belongs to none transaction");
1771 			spin_unlock(&journal->j_list_lock);
1772 			goto drop;
1773 		}
1774 
1775 		/*
1776 		 * Otherwise, if the buffer has been written to disk,
1777 		 * it is safe to remove the checkpoint and drop it.
1778 		 */
1779 		if (!buffer_dirty(bh)) {
1780 			__jbd2_journal_remove_checkpoint(jh);
1781 			spin_unlock(&journal->j_list_lock);
1782 			goto drop;
1783 		}
1784 
1785 		/*
1786 		 * The buffer is still not written to disk, we should
1787 		 * attach this buffer to current transaction so that the
1788 		 * buffer can be checkpointed only after the current
1789 		 * transaction commits.
1790 		 */
1791 		clear_buffer_dirty(bh);
1792 		__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1793 		spin_unlock(&journal->j_list_lock);
1794 	}
1795 drop:
1796 	__brelse(bh);
1797 	spin_unlock(&jh->b_state_lock);
1798 	jbd2_journal_put_journal_head(jh);
1799 	if (drop_reserve) {
1800 		/* no need to reserve log space for this block -bzzz */
1801 		handle->h_total_credits++;
1802 	}
1803 	return err;
1804 }
1805 
1806 /**
1807  * jbd2_journal_stop() - complete a transaction
1808  * @handle: transaction to complete.
1809  *
1810  * All done for a particular handle.
1811  *
1812  * There is not much action needed here.  We just return any remaining
1813  * buffer credits to the transaction and remove the handle.  The only
1814  * complication is that we need to start a commit operation if the
1815  * filesystem is marked for synchronous update.
1816  *
1817  * jbd2_journal_stop itself will not usually return an error, but it may
1818  * do so in unusual circumstances.  In particular, expect it to
1819  * return -EIO if a jbd2_journal_abort has been executed since the
1820  * transaction began.
1821  */
jbd2_journal_stop(handle_t * handle)1822 int jbd2_journal_stop(handle_t *handle)
1823 {
1824 	transaction_t *transaction = handle->h_transaction;
1825 	journal_t *journal;
1826 	int err = 0, wait_for_commit = 0;
1827 	tid_t tid;
1828 	pid_t pid;
1829 
1830 	if (--handle->h_ref > 0) {
1831 		jbd2_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1832 						 handle->h_ref);
1833 		if (is_handle_aborted(handle))
1834 			return -EIO;
1835 		return 0;
1836 	}
1837 	if (!transaction) {
1838 		/*
1839 		 * Handle is already detached from the transaction so there is
1840 		 * nothing to do other than free the handle.
1841 		 */
1842 		memalloc_nofs_restore(handle->saved_alloc_context);
1843 		goto free_and_exit;
1844 	}
1845 	journal = transaction->t_journal;
1846 	tid = transaction->t_tid;
1847 
1848 	if (is_handle_aborted(handle))
1849 		err = -EIO;
1850 
1851 	jbd2_debug(4, "Handle %p going down\n", handle);
1852 	trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1853 				tid, handle->h_type, handle->h_line_no,
1854 				jiffies - handle->h_start_jiffies,
1855 				handle->h_sync, handle->h_requested_credits,
1856 				(handle->h_requested_credits -
1857 				 handle->h_total_credits));
1858 
1859 	/*
1860 	 * Implement synchronous transaction batching.  If the handle
1861 	 * was synchronous, don't force a commit immediately.  Let's
1862 	 * yield and let another thread piggyback onto this
1863 	 * transaction.  Keep doing that while new threads continue to
1864 	 * arrive.  It doesn't cost much - we're about to run a commit
1865 	 * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1866 	 * operations by 30x or more...
1867 	 *
1868 	 * We try and optimize the sleep time against what the
1869 	 * underlying disk can do, instead of having a static sleep
1870 	 * time.  This is useful for the case where our storage is so
1871 	 * fast that it is more optimal to go ahead and force a flush
1872 	 * and wait for the transaction to be committed than it is to
1873 	 * wait for an arbitrary amount of time for new writers to
1874 	 * join the transaction.  We achieve this by measuring how
1875 	 * long it takes to commit a transaction, and compare it with
1876 	 * how long this transaction has been running, and if run time
1877 	 * < commit time then we sleep for the delta and commit.  This
1878 	 * greatly helps super fast disks that would see slowdowns as
1879 	 * more threads started doing fsyncs.
1880 	 *
1881 	 * But don't do this if this process was the most recent one
1882 	 * to perform a synchronous write.  We do this to detect the
1883 	 * case where a single process is doing a stream of sync
1884 	 * writes.  No point in waiting for joiners in that case.
1885 	 *
1886 	 * Setting max_batch_time to 0 disables this completely.
1887 	 */
1888 	pid = current->pid;
1889 	if (handle->h_sync && journal->j_last_sync_writer != pid &&
1890 	    journal->j_max_batch_time) {
1891 		u64 commit_time, trans_time;
1892 
1893 		journal->j_last_sync_writer = pid;
1894 
1895 		read_lock(&journal->j_state_lock);
1896 		commit_time = journal->j_average_commit_time;
1897 		read_unlock(&journal->j_state_lock);
1898 
1899 		trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1900 						   transaction->t_start_time));
1901 
1902 		commit_time = max_t(u64, commit_time,
1903 				    1000*journal->j_min_batch_time);
1904 		commit_time = min_t(u64, commit_time,
1905 				    1000*journal->j_max_batch_time);
1906 
1907 		if (trans_time < commit_time) {
1908 			ktime_t expires = ktime_add_ns(ktime_get(),
1909 						       commit_time);
1910 			set_current_state(TASK_UNINTERRUPTIBLE);
1911 			schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1912 		}
1913 	}
1914 
1915 	if (handle->h_sync)
1916 		transaction->t_synchronous_commit = 1;
1917 
1918 	/*
1919 	 * If the handle is marked SYNC, we need to set another commit
1920 	 * going!  We also want to force a commit if the transaction is too
1921 	 * old now.
1922 	 */
1923 	if (handle->h_sync ||
1924 	    time_after_eq(jiffies, transaction->t_expires)) {
1925 		/* Do this even for aborted journals: an abort still
1926 		 * completes the commit thread, it just doesn't write
1927 		 * anything to disk. */
1928 
1929 		jbd2_debug(2, "transaction too old, requesting commit for "
1930 					"handle %p\n", handle);
1931 		/* This is non-blocking */
1932 		jbd2_log_start_commit(journal, tid);
1933 
1934 		/*
1935 		 * Special case: JBD2_SYNC synchronous updates require us
1936 		 * to wait for the commit to complete.
1937 		 */
1938 		if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1939 			wait_for_commit = 1;
1940 	}
1941 
1942 	/*
1943 	 * Once stop_this_handle() drops t_updates, the transaction could start
1944 	 * committing on us and eventually disappear.  So we must not
1945 	 * dereference transaction pointer again after calling
1946 	 * stop_this_handle().
1947 	 */
1948 	stop_this_handle(handle);
1949 
1950 	if (wait_for_commit)
1951 		err = jbd2_log_wait_commit(journal, tid);
1952 
1953 free_and_exit:
1954 	if (handle->h_rsv_handle)
1955 		jbd2_free_handle(handle->h_rsv_handle);
1956 	jbd2_free_handle(handle);
1957 	return err;
1958 }
1959 
1960 /*
1961  *
1962  * List management code snippets: various functions for manipulating the
1963  * transaction buffer lists.
1964  *
1965  */
1966 
1967 /*
1968  * Append a buffer to a transaction list, given the transaction's list head
1969  * pointer.
1970  *
1971  * j_list_lock is held.
1972  *
1973  * jh->b_state_lock is held.
1974  */
1975 
1976 static inline void
__blist_add_buffer(struct journal_head ** list,struct journal_head * jh)1977 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1978 {
1979 	if (!*list) {
1980 		jh->b_tnext = jh->b_tprev = jh;
1981 		*list = jh;
1982 	} else {
1983 		/* Insert at the tail of the list to preserve order */
1984 		struct journal_head *first = *list, *last = first->b_tprev;
1985 		jh->b_tprev = last;
1986 		jh->b_tnext = first;
1987 		last->b_tnext = first->b_tprev = jh;
1988 	}
1989 }
1990 
1991 /*
1992  * Remove a buffer from a transaction list, given the transaction's list
1993  * head pointer.
1994  *
1995  * Called with j_list_lock held, and the journal may not be locked.
1996  *
1997  * jh->b_state_lock is held.
1998  */
1999 
2000 static inline void
__blist_del_buffer(struct journal_head ** list,struct journal_head * jh)2001 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
2002 {
2003 	if (*list == jh) {
2004 		*list = jh->b_tnext;
2005 		if (*list == jh)
2006 			*list = NULL;
2007 	}
2008 	jh->b_tprev->b_tnext = jh->b_tnext;
2009 	jh->b_tnext->b_tprev = jh->b_tprev;
2010 }
2011 
2012 /*
2013  * Remove a buffer from the appropriate transaction list.
2014  *
2015  * Note that this function can *change* the value of
2016  * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
2017  * t_reserved_list.  If the caller is holding onto a copy of one of these
2018  * pointers, it could go bad.  Generally the caller needs to re-read the
2019  * pointer from the transaction_t.
2020  *
2021  * Called under j_list_lock.
2022  */
__jbd2_journal_temp_unlink_buffer(struct journal_head * jh)2023 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
2024 {
2025 	struct journal_head **list = NULL;
2026 	transaction_t *transaction;
2027 	struct buffer_head *bh = jh2bh(jh);
2028 
2029 	lockdep_assert_held(&jh->b_state_lock);
2030 	transaction = jh->b_transaction;
2031 	if (transaction)
2032 		assert_spin_locked(&transaction->t_journal->j_list_lock);
2033 
2034 	J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2035 	if (jh->b_jlist != BJ_None)
2036 		J_ASSERT_JH(jh, transaction != NULL);
2037 
2038 	switch (jh->b_jlist) {
2039 	case BJ_None:
2040 		return;
2041 	case BJ_Metadata:
2042 		transaction->t_nr_buffers--;
2043 		J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
2044 		list = &transaction->t_buffers;
2045 		break;
2046 	case BJ_Forget:
2047 		list = &transaction->t_forget;
2048 		break;
2049 	case BJ_Shadow:
2050 		list = &transaction->t_shadow_list;
2051 		break;
2052 	case BJ_Reserved:
2053 		list = &transaction->t_reserved_list;
2054 		break;
2055 	}
2056 
2057 	__blist_del_buffer(list, jh);
2058 	jh->b_jlist = BJ_None;
2059 	if (transaction && is_journal_aborted(transaction->t_journal))
2060 		clear_buffer_jbddirty(bh);
2061 	else if (test_clear_buffer_jbddirty(bh))
2062 		mark_buffer_dirty(bh);	/* Expose it to the VM */
2063 }
2064 
2065 /*
2066  * Remove buffer from all transactions. The caller is responsible for dropping
2067  * the jh reference that belonged to the transaction.
2068  *
2069  * Called with bh_state lock and j_list_lock
2070  */
__jbd2_journal_unfile_buffer(struct journal_head * jh)2071 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
2072 {
2073 	J_ASSERT_JH(jh, jh->b_transaction != NULL);
2074 	J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2075 
2076 	__jbd2_journal_temp_unlink_buffer(jh);
2077 	jh->b_transaction = NULL;
2078 }
2079 
jbd2_journal_unfile_buffer(journal_t * journal,struct journal_head * jh)2080 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
2081 {
2082 	struct buffer_head *bh = jh2bh(jh);
2083 
2084 	/* Get reference so that buffer cannot be freed before we unlock it */
2085 	get_bh(bh);
2086 	spin_lock(&jh->b_state_lock);
2087 	spin_lock(&journal->j_list_lock);
2088 	__jbd2_journal_unfile_buffer(jh);
2089 	spin_unlock(&journal->j_list_lock);
2090 	spin_unlock(&jh->b_state_lock);
2091 	jbd2_journal_put_journal_head(jh);
2092 	__brelse(bh);
2093 }
2094 
2095 /*
2096  * Called from jbd2_journal_try_to_free_buffers().
2097  *
2098  * Called under jh->b_state_lock
2099  */
2100 static void
__journal_try_to_free_buffer(journal_t * journal,struct buffer_head * bh)2101 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
2102 {
2103 	struct journal_head *jh;
2104 
2105 	jh = bh2jh(bh);
2106 
2107 	if (buffer_locked(bh) || buffer_dirty(bh))
2108 		goto out;
2109 
2110 	if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
2111 		goto out;
2112 
2113 	spin_lock(&journal->j_list_lock);
2114 	if (jh->b_cp_transaction != NULL) {
2115 		/* written-back checkpointed metadata buffer */
2116 		JBUFFER_TRACE(jh, "remove from checkpoint list");
2117 		__jbd2_journal_remove_checkpoint(jh);
2118 	}
2119 	spin_unlock(&journal->j_list_lock);
2120 out:
2121 	return;
2122 }
2123 
2124 /**
2125  * jbd2_journal_try_to_free_buffers() - try to free page buffers.
2126  * @journal: journal for operation
2127  * @folio: Folio to detach data from.
2128  *
2129  * For all the buffers on this page,
2130  * if they are fully written out ordered data, move them onto BUF_CLEAN
2131  * so try_to_free_buffers() can reap them.
2132  *
2133  * This function returns non-zero if we wish try_to_free_buffers()
2134  * to be called. We do this if the page is releasable by try_to_free_buffers().
2135  * We also do it if the page has locked or dirty buffers and the caller wants
2136  * us to perform sync or async writeout.
2137  *
2138  * This complicates JBD locking somewhat.  We aren't protected by the
2139  * BKL here.  We wish to remove the buffer from its committing or
2140  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2141  *
2142  * This may *change* the value of transaction_t->t_datalist, so anyone
2143  * who looks at t_datalist needs to lock against this function.
2144  *
2145  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2146  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
2147  * will come out of the lock with the buffer dirty, which makes it
2148  * ineligible for release here.
2149  *
2150  * Who else is affected by this?  hmm...  Really the only contender
2151  * is do_get_write_access() - it could be looking at the buffer while
2152  * journal_try_to_free_buffer() is changing its state.  But that
2153  * cannot happen because we never reallocate freed data as metadata
2154  * while the data is part of a transaction.  Yes?
2155  *
2156  * Return false on failure, true on success
2157  */
jbd2_journal_try_to_free_buffers(journal_t * journal,struct folio * folio)2158 bool jbd2_journal_try_to_free_buffers(journal_t *journal, struct folio *folio)
2159 {
2160 	struct buffer_head *head;
2161 	struct buffer_head *bh;
2162 	bool ret = false;
2163 
2164 	J_ASSERT(folio_test_locked(folio));
2165 
2166 	head = folio_buffers(folio);
2167 	bh = head;
2168 	do {
2169 		struct journal_head *jh;
2170 
2171 		/*
2172 		 * We take our own ref against the journal_head here to avoid
2173 		 * having to add tons of locking around each instance of
2174 		 * jbd2_journal_put_journal_head().
2175 		 */
2176 		jh = jbd2_journal_grab_journal_head(bh);
2177 		if (!jh)
2178 			continue;
2179 
2180 		spin_lock(&jh->b_state_lock);
2181 		__journal_try_to_free_buffer(journal, bh);
2182 		spin_unlock(&jh->b_state_lock);
2183 		jbd2_journal_put_journal_head(jh);
2184 		if (buffer_jbd(bh))
2185 			goto busy;
2186 	} while ((bh = bh->b_this_page) != head);
2187 
2188 	ret = try_to_free_buffers(folio);
2189 busy:
2190 	return ret;
2191 }
2192 
2193 /*
2194  * This buffer is no longer needed.  If it is on an older transaction's
2195  * checkpoint list we need to record it on this transaction's forget list
2196  * to pin this buffer (and hence its checkpointing transaction) down until
2197  * this transaction commits.  If the buffer isn't on a checkpoint list, we
2198  * release it.
2199  * Returns non-zero if JBD no longer has an interest in the buffer.
2200  *
2201  * Called under j_list_lock.
2202  *
2203  * Called under jh->b_state_lock.
2204  */
__dispose_buffer(struct journal_head * jh,transaction_t * transaction)2205 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2206 {
2207 	int may_free = 1;
2208 	struct buffer_head *bh = jh2bh(jh);
2209 
2210 	if (jh->b_cp_transaction) {
2211 		JBUFFER_TRACE(jh, "on running+cp transaction");
2212 		__jbd2_journal_temp_unlink_buffer(jh);
2213 		/*
2214 		 * We don't want to write the buffer anymore, clear the
2215 		 * bit so that we don't confuse checks in
2216 		 * __journal_file_buffer
2217 		 */
2218 		clear_buffer_dirty(bh);
2219 		__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2220 		may_free = 0;
2221 	} else {
2222 		JBUFFER_TRACE(jh, "on running transaction");
2223 		__jbd2_journal_unfile_buffer(jh);
2224 		jbd2_journal_put_journal_head(jh);
2225 	}
2226 	return may_free;
2227 }
2228 
2229 /*
2230  * jbd2_journal_invalidate_folio
2231  *
2232  * This code is tricky.  It has a number of cases to deal with.
2233  *
2234  * There are two invariants which this code relies on:
2235  *
2236  * i_size must be updated on disk before we start calling invalidate_folio
2237  * on the data.
2238  *
2239  *  This is done in ext3 by defining an ext3_setattr method which
2240  *  updates i_size before truncate gets going.  By maintaining this
2241  *  invariant, we can be sure that it is safe to throw away any buffers
2242  *  attached to the current transaction: once the transaction commits,
2243  *  we know that the data will not be needed.
2244  *
2245  *  Note however that we can *not* throw away data belonging to the
2246  *  previous, committing transaction!
2247  *
2248  * Any disk blocks which *are* part of the previous, committing
2249  * transaction (and which therefore cannot be discarded immediately) are
2250  * not going to be reused in the new running transaction
2251  *
2252  *  The bitmap committed_data images guarantee this: any block which is
2253  *  allocated in one transaction and removed in the next will be marked
2254  *  as in-use in the committed_data bitmap, so cannot be reused until
2255  *  the next transaction to delete the block commits.  This means that
2256  *  leaving committing buffers dirty is quite safe: the disk blocks
2257  *  cannot be reallocated to a different file and so buffer aliasing is
2258  *  not possible.
2259  *
2260  *
2261  * The above applies mainly to ordered data mode.  In writeback mode we
2262  * don't make guarantees about the order in which data hits disk --- in
2263  * particular we don't guarantee that new dirty data is flushed before
2264  * transaction commit --- so it is always safe just to discard data
2265  * immediately in that mode.  --sct
2266  */
2267 
2268 /*
2269  * The journal_unmap_buffer helper function returns zero if the buffer
2270  * concerned remains pinned as an anonymous buffer belonging to an older
2271  * transaction.
2272  *
2273  * We're outside-transaction here.  Either or both of j_running_transaction
2274  * and j_committing_transaction may be NULL.
2275  */
journal_unmap_buffer(journal_t * journal,struct buffer_head * bh,int partial_page)2276 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2277 				int partial_page)
2278 {
2279 	transaction_t *transaction;
2280 	struct journal_head *jh;
2281 	int may_free = 1;
2282 
2283 	BUFFER_TRACE(bh, "entry");
2284 
2285 	/*
2286 	 * It is safe to proceed here without the j_list_lock because the
2287 	 * buffers cannot be stolen by try_to_free_buffers as long as we are
2288 	 * holding the page lock. --sct
2289 	 */
2290 
2291 	jh = jbd2_journal_grab_journal_head(bh);
2292 	if (!jh)
2293 		goto zap_buffer_unlocked;
2294 
2295 	/* OK, we have data buffer in journaled mode */
2296 	write_lock(&journal->j_state_lock);
2297 	spin_lock(&jh->b_state_lock);
2298 	spin_lock(&journal->j_list_lock);
2299 
2300 	/*
2301 	 * We cannot remove the buffer from checkpoint lists until the
2302 	 * transaction adding inode to orphan list (let's call it T)
2303 	 * is committed.  Otherwise if the transaction changing the
2304 	 * buffer would be cleaned from the journal before T is
2305 	 * committed, a crash will cause that the correct contents of
2306 	 * the buffer will be lost.  On the other hand we have to
2307 	 * clear the buffer dirty bit at latest at the moment when the
2308 	 * transaction marking the buffer as freed in the filesystem
2309 	 * structures is committed because from that moment on the
2310 	 * block can be reallocated and used by a different page.
2311 	 * Since the block hasn't been freed yet but the inode has
2312 	 * already been added to orphan list, it is safe for us to add
2313 	 * the buffer to BJ_Forget list of the newest transaction.
2314 	 *
2315 	 * Also we have to clear buffer_mapped flag of a truncated buffer
2316 	 * because the buffer_head may be attached to the page straddling
2317 	 * i_size (can happen only when blocksize < pagesize) and thus the
2318 	 * buffer_head can be reused when the file is extended again. So we end
2319 	 * up keeping around invalidated buffers attached to transactions'
2320 	 * BJ_Forget list just to stop checkpointing code from cleaning up
2321 	 * the transaction this buffer was modified in.
2322 	 */
2323 	transaction = jh->b_transaction;
2324 	if (transaction == NULL) {
2325 		/* First case: not on any transaction.  If it
2326 		 * has no checkpoint link, then we can zap it:
2327 		 * it's a writeback-mode buffer so we don't care
2328 		 * if it hits disk safely. */
2329 		if (!jh->b_cp_transaction) {
2330 			JBUFFER_TRACE(jh, "not on any transaction: zap");
2331 			goto zap_buffer;
2332 		}
2333 
2334 		if (!buffer_dirty(bh)) {
2335 			/* bdflush has written it.  We can drop it now */
2336 			__jbd2_journal_remove_checkpoint(jh);
2337 			goto zap_buffer;
2338 		}
2339 
2340 		/* OK, it must be in the journal but still not
2341 		 * written fully to disk: it's metadata or
2342 		 * journaled data... */
2343 
2344 		if (journal->j_running_transaction) {
2345 			/* ... and once the current transaction has
2346 			 * committed, the buffer won't be needed any
2347 			 * longer. */
2348 			JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2349 			may_free = __dispose_buffer(jh,
2350 					journal->j_running_transaction);
2351 			goto zap_buffer;
2352 		} else {
2353 			/* There is no currently-running transaction. So the
2354 			 * orphan record which we wrote for this file must have
2355 			 * passed into commit.  We must attach this buffer to
2356 			 * the committing transaction, if it exists. */
2357 			if (journal->j_committing_transaction) {
2358 				JBUFFER_TRACE(jh, "give to committing trans");
2359 				may_free = __dispose_buffer(jh,
2360 					journal->j_committing_transaction);
2361 				goto zap_buffer;
2362 			} else {
2363 				/* The orphan record's transaction has
2364 				 * committed.  We can cleanse this buffer */
2365 				clear_buffer_jbddirty(bh);
2366 				__jbd2_journal_remove_checkpoint(jh);
2367 				goto zap_buffer;
2368 			}
2369 		}
2370 	} else if (transaction == journal->j_committing_transaction) {
2371 		JBUFFER_TRACE(jh, "on committing transaction");
2372 		/*
2373 		 * The buffer is committing, we simply cannot touch
2374 		 * it. If the page is straddling i_size we have to wait
2375 		 * for commit and try again.
2376 		 */
2377 		if (partial_page) {
2378 			spin_unlock(&journal->j_list_lock);
2379 			spin_unlock(&jh->b_state_lock);
2380 			write_unlock(&journal->j_state_lock);
2381 			jbd2_journal_put_journal_head(jh);
2382 			return -EBUSY;
2383 		}
2384 		/*
2385 		 * OK, buffer won't be reachable after truncate. We just clear
2386 		 * b_modified to not confuse transaction credit accounting, and
2387 		 * set j_next_transaction to the running transaction (if there
2388 		 * is one) and mark buffer as freed so that commit code knows
2389 		 * it should clear dirty bits when it is done with the buffer.
2390 		 */
2391 		set_buffer_freed(bh);
2392 		if (journal->j_running_transaction && buffer_jbddirty(bh))
2393 			jh->b_next_transaction = journal->j_running_transaction;
2394 		jh->b_modified = 0;
2395 		spin_unlock(&journal->j_list_lock);
2396 		spin_unlock(&jh->b_state_lock);
2397 		write_unlock(&journal->j_state_lock);
2398 		jbd2_journal_put_journal_head(jh);
2399 		return 0;
2400 	} else {
2401 		/* Good, the buffer belongs to the running transaction.
2402 		 * We are writing our own transaction's data, not any
2403 		 * previous one's, so it is safe to throw it away
2404 		 * (remember that we expect the filesystem to have set
2405 		 * i_size already for this truncate so recovery will not
2406 		 * expose the disk blocks we are discarding here.) */
2407 		J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2408 		JBUFFER_TRACE(jh, "on running transaction");
2409 		may_free = __dispose_buffer(jh, transaction);
2410 	}
2411 
2412 zap_buffer:
2413 	/*
2414 	 * This is tricky. Although the buffer is truncated, it may be reused
2415 	 * if blocksize < pagesize and it is attached to the page straddling
2416 	 * EOF. Since the buffer might have been added to BJ_Forget list of the
2417 	 * running transaction, journal_get_write_access() won't clear
2418 	 * b_modified and credit accounting gets confused. So clear b_modified
2419 	 * here.
2420 	 */
2421 	jh->b_modified = 0;
2422 	spin_unlock(&journal->j_list_lock);
2423 	spin_unlock(&jh->b_state_lock);
2424 	write_unlock(&journal->j_state_lock);
2425 	jbd2_journal_put_journal_head(jh);
2426 zap_buffer_unlocked:
2427 	clear_buffer_dirty(bh);
2428 	J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2429 	clear_buffer_mapped(bh);
2430 	clear_buffer_req(bh);
2431 	clear_buffer_new(bh);
2432 	clear_buffer_delay(bh);
2433 	clear_buffer_unwritten(bh);
2434 	bh->b_bdev = NULL;
2435 	return may_free;
2436 }
2437 
2438 /**
2439  * jbd2_journal_invalidate_folio()
2440  * @journal: journal to use for flush...
2441  * @folio:    folio to flush
2442  * @offset:  start of the range to invalidate
2443  * @length:  length of the range to invalidate
2444  *
2445  * Reap page buffers containing data after in the specified range in page.
2446  * Can return -EBUSY if buffers are part of the committing transaction and
2447  * the page is straddling i_size. Caller then has to wait for current commit
2448  * and try again.
2449  */
jbd2_journal_invalidate_folio(journal_t * journal,struct folio * folio,size_t offset,size_t length)2450 int jbd2_journal_invalidate_folio(journal_t *journal, struct folio *folio,
2451 				size_t offset, size_t length)
2452 {
2453 	struct buffer_head *head, *bh, *next;
2454 	unsigned int stop = offset + length;
2455 	unsigned int curr_off = 0;
2456 	int partial_page = (offset || length < folio_size(folio));
2457 	int may_free = 1;
2458 	int ret = 0;
2459 
2460 	if (!folio_test_locked(folio))
2461 		BUG();
2462 	head = folio_buffers(folio);
2463 	if (!head)
2464 		return 0;
2465 
2466 	BUG_ON(stop > folio_size(folio) || stop < length);
2467 
2468 	/* We will potentially be playing with lists other than just the
2469 	 * data lists (especially for journaled data mode), so be
2470 	 * cautious in our locking. */
2471 
2472 	bh = head;
2473 	do {
2474 		unsigned int next_off = curr_off + bh->b_size;
2475 		next = bh->b_this_page;
2476 
2477 		if (next_off > stop)
2478 			return 0;
2479 
2480 		if (offset <= curr_off) {
2481 			/* This block is wholly outside the truncation point */
2482 			lock_buffer(bh);
2483 			ret = journal_unmap_buffer(journal, bh, partial_page);
2484 			unlock_buffer(bh);
2485 			if (ret < 0)
2486 				return ret;
2487 			may_free &= ret;
2488 		}
2489 		curr_off = next_off;
2490 		bh = next;
2491 
2492 	} while (bh != head);
2493 
2494 	if (!partial_page) {
2495 		if (may_free && try_to_free_buffers(folio))
2496 			J_ASSERT(!folio_buffers(folio));
2497 	}
2498 	return 0;
2499 }
2500 
2501 /*
2502  * File a buffer on the given transaction list.
2503  */
__jbd2_journal_file_buffer(struct journal_head * jh,transaction_t * transaction,int jlist)2504 void __jbd2_journal_file_buffer(struct journal_head *jh,
2505 			transaction_t *transaction, int jlist)
2506 {
2507 	struct journal_head **list = NULL;
2508 	int was_dirty = 0;
2509 	struct buffer_head *bh = jh2bh(jh);
2510 
2511 	lockdep_assert_held(&jh->b_state_lock);
2512 	assert_spin_locked(&transaction->t_journal->j_list_lock);
2513 
2514 	J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2515 	J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2516 				jh->b_transaction == NULL);
2517 
2518 	if (jh->b_transaction && jh->b_jlist == jlist)
2519 		return;
2520 
2521 	if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2522 	    jlist == BJ_Shadow || jlist == BJ_Forget) {
2523 		/*
2524 		 * For metadata buffers, we track dirty bit in buffer_jbddirty
2525 		 * instead of buffer_dirty. We should not see a dirty bit set
2526 		 * here because we clear it in do_get_write_access but e.g.
2527 		 * tune2fs can modify the sb and set the dirty bit at any time
2528 		 * so we try to gracefully handle that.
2529 		 */
2530 		if (buffer_dirty(bh))
2531 			warn_dirty_buffer(bh);
2532 		if (test_clear_buffer_dirty(bh) ||
2533 		    test_clear_buffer_jbddirty(bh))
2534 			was_dirty = 1;
2535 	}
2536 
2537 	if (jh->b_transaction)
2538 		__jbd2_journal_temp_unlink_buffer(jh);
2539 	else
2540 		jbd2_journal_grab_journal_head(bh);
2541 	jh->b_transaction = transaction;
2542 
2543 	switch (jlist) {
2544 	case BJ_None:
2545 		J_ASSERT_JH(jh, !jh->b_committed_data);
2546 		J_ASSERT_JH(jh, !jh->b_frozen_data);
2547 		return;
2548 	case BJ_Metadata:
2549 		transaction->t_nr_buffers++;
2550 		list = &transaction->t_buffers;
2551 		break;
2552 	case BJ_Forget:
2553 		list = &transaction->t_forget;
2554 		break;
2555 	case BJ_Shadow:
2556 		list = &transaction->t_shadow_list;
2557 		break;
2558 	case BJ_Reserved:
2559 		list = &transaction->t_reserved_list;
2560 		break;
2561 	}
2562 
2563 	__blist_add_buffer(list, jh);
2564 	jh->b_jlist = jlist;
2565 
2566 	if (was_dirty)
2567 		set_buffer_jbddirty(bh);
2568 }
2569 
jbd2_journal_file_buffer(struct journal_head * jh,transaction_t * transaction,int jlist)2570 void jbd2_journal_file_buffer(struct journal_head *jh,
2571 				transaction_t *transaction, int jlist)
2572 {
2573 	spin_lock(&jh->b_state_lock);
2574 	spin_lock(&transaction->t_journal->j_list_lock);
2575 	__jbd2_journal_file_buffer(jh, transaction, jlist);
2576 	spin_unlock(&transaction->t_journal->j_list_lock);
2577 	spin_unlock(&jh->b_state_lock);
2578 }
2579 
2580 /*
2581  * Remove a buffer from its current buffer list in preparation for
2582  * dropping it from its current transaction entirely.  If the buffer has
2583  * already started to be used by a subsequent transaction, refile the
2584  * buffer on that transaction's metadata list.
2585  *
2586  * Called under j_list_lock
2587  * Called under jh->b_state_lock
2588  *
2589  * When this function returns true, there's no next transaction to refile to
2590  * and the caller has to drop jh reference through
2591  * jbd2_journal_put_journal_head().
2592  */
__jbd2_journal_refile_buffer(struct journal_head * jh)2593 bool __jbd2_journal_refile_buffer(struct journal_head *jh)
2594 {
2595 	int was_dirty, jlist;
2596 	struct buffer_head *bh = jh2bh(jh);
2597 
2598 	lockdep_assert_held(&jh->b_state_lock);
2599 	if (jh->b_transaction)
2600 		assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2601 
2602 	/* If the buffer is now unused, just drop it. */
2603 	if (jh->b_next_transaction == NULL) {
2604 		__jbd2_journal_unfile_buffer(jh);
2605 		return true;
2606 	}
2607 
2608 	/*
2609 	 * It has been modified by a later transaction: add it to the new
2610 	 * transaction's metadata list.
2611 	 */
2612 
2613 	was_dirty = test_clear_buffer_jbddirty(bh);
2614 	__jbd2_journal_temp_unlink_buffer(jh);
2615 
2616 	/*
2617 	 * b_transaction must be set, otherwise the new b_transaction won't
2618 	 * be holding jh reference
2619 	 */
2620 	J_ASSERT_JH(jh, jh->b_transaction != NULL);
2621 
2622 	/*
2623 	 * We set b_transaction here because b_next_transaction will inherit
2624 	 * our jh reference and thus __jbd2_journal_file_buffer() must not
2625 	 * take a new one.
2626 	 */
2627 	WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
2628 	WRITE_ONCE(jh->b_next_transaction, NULL);
2629 	if (buffer_freed(bh))
2630 		jlist = BJ_Forget;
2631 	else if (jh->b_modified)
2632 		jlist = BJ_Metadata;
2633 	else
2634 		jlist = BJ_Reserved;
2635 	__jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2636 	J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2637 
2638 	if (was_dirty)
2639 		set_buffer_jbddirty(bh);
2640 	return false;
2641 }
2642 
2643 /*
2644  * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2645  * bh reference so that we can safely unlock bh.
2646  *
2647  * The jh and bh may be freed by this call.
2648  */
jbd2_journal_refile_buffer(journal_t * journal,struct journal_head * jh)2649 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2650 {
2651 	bool drop;
2652 
2653 	spin_lock(&jh->b_state_lock);
2654 	spin_lock(&journal->j_list_lock);
2655 	drop = __jbd2_journal_refile_buffer(jh);
2656 	spin_unlock(&jh->b_state_lock);
2657 	spin_unlock(&journal->j_list_lock);
2658 	if (drop)
2659 		jbd2_journal_put_journal_head(jh);
2660 }
2661 
2662 /*
2663  * File inode in the inode list of the handle's transaction
2664  */
jbd2_journal_file_inode(handle_t * handle,struct jbd2_inode * jinode,unsigned long flags,loff_t start_byte,loff_t end_byte)2665 static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2666 		unsigned long flags, loff_t start_byte, loff_t end_byte)
2667 {
2668 	transaction_t *transaction = handle->h_transaction;
2669 	journal_t *journal;
2670 
2671 	if (is_handle_aborted(handle))
2672 		return -EROFS;
2673 	journal = transaction->t_journal;
2674 
2675 	jbd2_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2676 			transaction->t_tid);
2677 
2678 	spin_lock(&journal->j_list_lock);
2679 	jinode->i_flags |= flags;
2680 
2681 	if (jinode->i_dirty_end) {
2682 		jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2683 		jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2684 	} else {
2685 		jinode->i_dirty_start = start_byte;
2686 		jinode->i_dirty_end = end_byte;
2687 	}
2688 
2689 	/* Is inode already attached where we need it? */
2690 	if (jinode->i_transaction == transaction ||
2691 	    jinode->i_next_transaction == transaction)
2692 		goto done;
2693 
2694 	/*
2695 	 * We only ever set this variable to 1 so the test is safe. Since
2696 	 * t_need_data_flush is likely to be set, we do the test to save some
2697 	 * cacheline bouncing
2698 	 */
2699 	if (!transaction->t_need_data_flush)
2700 		transaction->t_need_data_flush = 1;
2701 	/* On some different transaction's list - should be
2702 	 * the committing one */
2703 	if (jinode->i_transaction) {
2704 		J_ASSERT(jinode->i_next_transaction == NULL);
2705 		J_ASSERT(jinode->i_transaction ==
2706 					journal->j_committing_transaction);
2707 		jinode->i_next_transaction = transaction;
2708 		goto done;
2709 	}
2710 	/* Not on any transaction list... */
2711 	J_ASSERT(!jinode->i_next_transaction);
2712 	jinode->i_transaction = transaction;
2713 	list_add(&jinode->i_list, &transaction->t_inode_list);
2714 done:
2715 	spin_unlock(&journal->j_list_lock);
2716 
2717 	return 0;
2718 }
2719 
jbd2_journal_inode_ranged_write(handle_t * handle,struct jbd2_inode * jinode,loff_t start_byte,loff_t length)2720 int jbd2_journal_inode_ranged_write(handle_t *handle,
2721 		struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2722 {
2723 	return jbd2_journal_file_inode(handle, jinode,
2724 			JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2725 			start_byte + length - 1);
2726 }
2727 
jbd2_journal_inode_ranged_wait(handle_t * handle,struct jbd2_inode * jinode,loff_t start_byte,loff_t length)2728 int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2729 		loff_t start_byte, loff_t length)
2730 {
2731 	return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2732 			start_byte, start_byte + length - 1);
2733 }
2734 
2735 /*
2736  * File truncate and transaction commit interact with each other in a
2737  * non-trivial way.  If a transaction writing data block A is
2738  * committing, we cannot discard the data by truncate until we have
2739  * written them.  Otherwise if we crashed after the transaction with
2740  * write has committed but before the transaction with truncate has
2741  * committed, we could see stale data in block A.  This function is a
2742  * helper to solve this problem.  It starts writeout of the truncated
2743  * part in case it is in the committing transaction.
2744  *
2745  * Filesystem code must call this function when inode is journaled in
2746  * ordered mode before truncation happens and after the inode has been
2747  * placed on orphan list with the new inode size. The second condition
2748  * avoids the race that someone writes new data and we start
2749  * committing the transaction after this function has been called but
2750  * before a transaction for truncate is started (and furthermore it
2751  * allows us to optimize the case where the addition to orphan list
2752  * happens in the same transaction as write --- we don't have to write
2753  * any data in such case).
2754  */
jbd2_journal_begin_ordered_truncate(journal_t * journal,struct jbd2_inode * jinode,loff_t new_size)2755 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2756 					struct jbd2_inode *jinode,
2757 					loff_t new_size)
2758 {
2759 	transaction_t *inode_trans, *commit_trans;
2760 	int ret = 0;
2761 
2762 	/* This is a quick check to avoid locking if not necessary */
2763 	if (!jinode->i_transaction)
2764 		goto out;
2765 	/* Locks are here just to force reading of recent values, it is
2766 	 * enough that the transaction was not committing before we started
2767 	 * a transaction adding the inode to orphan list */
2768 	read_lock(&journal->j_state_lock);
2769 	commit_trans = journal->j_committing_transaction;
2770 	read_unlock(&journal->j_state_lock);
2771 	spin_lock(&journal->j_list_lock);
2772 	inode_trans = jinode->i_transaction;
2773 	spin_unlock(&journal->j_list_lock);
2774 	if (inode_trans == commit_trans) {
2775 		ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2776 			new_size, LLONG_MAX);
2777 		if (ret)
2778 			jbd2_journal_abort(journal, ret);
2779 	}
2780 out:
2781 	return ret;
2782 }
2783