1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
4 */
5
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_shared.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_extent_busy.h"
14 #include "xfs_trans.h"
15 #include "xfs_trans_priv.h"
16 #include "xfs_log.h"
17 #include "xfs_log_priv.h"
18 #include "xfs_trace.h"
19 #include "xfs_discard.h"
20
21 /*
22 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
23 * recover, so we don't allow failure here. Also, we allocate in a context that
24 * we don't want to be issuing transactions from, so we need to tell the
25 * allocation code this as well.
26 *
27 * We don't reserve any space for the ticket - we are going to steal whatever
28 * space we require from transactions as they commit. To ensure we reserve all
29 * the space required, we need to set the current reservation of the ticket to
30 * zero so that we know to steal the initial transaction overhead from the
31 * first transaction commit.
32 */
33 static struct xlog_ticket *
34 xlog_cil_ticket_alloc(
35 struct xlog *log)
36 {
37 struct xlog_ticket *tic;
38
39 tic = xlog_ticket_alloc(log, 0, 1, 0);
40
41 /*
42 * set the current reservation to zero so we know to steal the basic
43 * transaction overhead reservation from the first transaction commit.
44 */
45 tic->t_curr_res = 0;
46 tic->t_iclog_hdrs = 0;
47 return tic;
48 }
49
50 static inline void
51 xlog_cil_set_iclog_hdr_count(struct xfs_cil *cil)
52 {
53 struct xlog *log = cil->xc_log;
54
55 atomic_set(&cil->xc_iclog_hdrs,
56 (XLOG_CIL_BLOCKING_SPACE_LIMIT(log) /
57 (log->l_iclog_size - log->l_iclog_hsize)));
58 }
59
60 /*
61 * Check if the current log item was first committed in this sequence.
62 * We can't rely on just the log item being in the CIL, we have to check
63 * the recorded commit sequence number.
64 *
65 * Note: for this to be used in a non-racy manner, it has to be called with
66 * CIL flushing locked out. As a result, it should only be used during the
67 * transaction commit process when deciding what to format into the item.
68 */
69 static bool
70 xlog_item_in_current_chkpt(
71 struct xfs_cil *cil,
72 struct xfs_log_item *lip)
73 {
74 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags))
75 return false;
76
77 /*
78 * li_seq is written on the first commit of a log item to record the
79 * first checkpoint it is written to. Hence if it is different to the
80 * current sequence, we're in a new checkpoint.
81 */
82 return lip->li_seq == READ_ONCE(cil->xc_current_sequence);
83 }
84
85 bool
86 xfs_log_item_in_current_chkpt(
87 struct xfs_log_item *lip)
88 {
89 return xlog_item_in_current_chkpt(lip->li_log->l_cilp, lip);
90 }
91
92 /*
93 * Unavoidable forward declaration - xlog_cil_push_work() calls
94 * xlog_cil_ctx_alloc() itself.
95 */
96 static void xlog_cil_push_work(struct work_struct *work);
97
98 static struct xfs_cil_ctx *
99 xlog_cil_ctx_alloc(void)
100 {
101 struct xfs_cil_ctx *ctx;
102
103 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL | __GFP_NOFAIL);
104 INIT_LIST_HEAD(&ctx->committing);
105 INIT_LIST_HEAD(&ctx->busy_extents.extent_list);
106 INIT_LIST_HEAD(&ctx->log_items);
107 INIT_LIST_HEAD(&ctx->lv_chain);
108 INIT_WORK(&ctx->push_work, xlog_cil_push_work);
109 return ctx;
110 }
111
112 /*
113 * Aggregate the CIL per cpu structures into global counts, lists, etc and
114 * clear the percpu state ready for the next context to use. This is called
115 * from the push code with the context lock held exclusively, hence nothing else
116 * will be accessing or modifying the per-cpu counters.
117 */
118 static void
119 xlog_cil_push_pcp_aggregate(
120 struct xfs_cil *cil,
121 struct xfs_cil_ctx *ctx)
122 {
123 struct xlog_cil_pcp *cilpcp;
124 int cpu;
125
126 for_each_cpu(cpu, &ctx->cil_pcpmask) {
127 cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);
128
129 ctx->ticket->t_curr_res += cilpcp->space_reserved;
130 cilpcp->space_reserved = 0;
131
132 if (!list_empty(&cilpcp->busy_extents)) {
133 list_splice_init(&cilpcp->busy_extents,
134 &ctx->busy_extents.extent_list);
135 }
136 if (!list_empty(&cilpcp->log_items))
137 list_splice_init(&cilpcp->log_items, &ctx->log_items);
138
139 /*
140 * We're in the middle of switching cil contexts. Reset the
141 * counter we use to detect when the current context is nearing
142 * full.
143 */
144 cilpcp->space_used = 0;
145 }
146 }
147
148 /*
149 * Aggregate the CIL per-cpu space used counters into the global atomic value.
150 * This is called when the per-cpu counter aggregation will first pass the soft
151 * limit threshold so we can switch to atomic counter aggregation for accurate
152 * detection of hard limit traversal.
153 */
154 static void
155 xlog_cil_insert_pcp_aggregate(
156 struct xfs_cil *cil,
157 struct xfs_cil_ctx *ctx)
158 {
159 struct xlog_cil_pcp *cilpcp;
160 int cpu;
161 int count = 0;
162
163 /* Trigger atomic updates then aggregate only for the first caller */
164 if (!test_and_clear_bit(XLOG_CIL_PCP_SPACE, &cil->xc_flags))
165 return;
166
167 /*
168 * We can race with other cpus setting cil_pcpmask. However, we've
169 * atomically cleared PCP_SPACE which forces other threads to add to
170 * the global space used count. cil_pcpmask is a superset of cilpcp
171 * structures that could have a nonzero space_used.
172 */
173 for_each_cpu(cpu, &ctx->cil_pcpmask) {
174 int old, prev;
175
176 cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);
177 do {
178 old = cilpcp->space_used;
179 prev = cmpxchg(&cilpcp->space_used, old, 0);
180 } while (old != prev);
181 count += old;
182 }
183 atomic_add(count, &ctx->space_used);
184 }
185
186 static void
187 xlog_cil_ctx_switch(
188 struct xfs_cil *cil,
189 struct xfs_cil_ctx *ctx)
190 {
191 xlog_cil_set_iclog_hdr_count(cil);
192 set_bit(XLOG_CIL_EMPTY, &cil->xc_flags);
193 set_bit(XLOG_CIL_PCP_SPACE, &cil->xc_flags);
194 ctx->sequence = ++cil->xc_current_sequence;
195 ctx->cil = cil;
196 cil->xc_ctx = ctx;
197 }
198
199 /*
200 * After the first stage of log recovery is done, we know where the head and
201 * tail of the log are. We need this log initialisation done before we can
202 * initialise the first CIL checkpoint context.
203 *
204 * Here we allocate a log ticket to track space usage during a CIL push. This
205 * ticket is passed to xlog_write() directly so that we don't slowly leak log
206 * space by failing to account for space used by log headers and additional
207 * region headers for split regions.
208 */
209 void
210 xlog_cil_init_post_recovery(
211 struct xlog *log)
212 {
213 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
214 log->l_cilp->xc_ctx->sequence = 1;
215 xlog_cil_set_iclog_hdr_count(log->l_cilp);
216 }
217
218 static inline int
219 xlog_cil_iovec_space(
220 uint niovecs)
221 {
222 return round_up((sizeof(struct xfs_log_vec) +
223 niovecs * sizeof(struct xfs_log_iovec)),
224 sizeof(uint64_t));
225 }
226
227 /*
228 * Allocate or pin log vector buffers for CIL insertion.
229 *
230 * The CIL currently uses disposable buffers for copying a snapshot of the
231 * modified items into the log during a push. The biggest problem with this is
232 * the requirement to allocate the disposable buffer during the commit if:
233 * a) does not exist; or
234 * b) it is too small
235 *
236 * If we do this allocation within xlog_cil_insert_format_items(), it is done
237 * under the xc_ctx_lock, which means that a CIL push cannot occur during
238 * the memory allocation. This means that we have a potential deadlock situation
239 * under low memory conditions when we have lots of dirty metadata pinned in
240 * the CIL and we need a CIL commit to occur to free memory.
241 *
242 * To avoid this, we need to move the memory allocation outside the
243 * xc_ctx_lock, but because the log vector buffers are disposable, that opens
244 * up a TOCTOU race condition w.r.t. the CIL committing and removing the log
245 * vector buffers between the check and the formatting of the item into the
246 * log vector buffer within the xc_ctx_lock.
247 *
248 * Because the log vector buffer needs to be unchanged during the CIL push
249 * process, we cannot share the buffer between the transaction commit (which
250 * modifies the buffer) and the CIL push context that is writing the changes
251 * into the log. This means skipping preallocation of buffer space is
252 * unreliable, but we most definitely do not want to be allocating and freeing
253 * buffers unnecessarily during commits when overwrites can be done safely.
254 *
255 * The simplest solution to this problem is to allocate a shadow buffer when a
256 * log item is committed for the second time, and then to only use this buffer
257 * if necessary. The buffer can remain attached to the log item until such time
258 * it is needed, and this is the buffer that is reallocated to match the size of
259 * the incoming modification. Then during the formatting of the item we can swap
260 * the active buffer with the new one if we can't reuse the existing buffer. We
261 * don't free the old buffer as it may be reused on the next modification if
262 * it's size is right, otherwise we'll free and reallocate it at that point.
263 *
264 * This function builds a vector for the changes in each log item in the
265 * transaction. It then works out the length of the buffer needed for each log
266 * item, allocates them and attaches the vector to the log item in preparation
267 * for the formatting step which occurs under the xc_ctx_lock.
268 *
269 * While this means the memory footprint goes up, it avoids the repeated
270 * alloc/free pattern that repeated modifications of an item would otherwise
271 * cause, and hence minimises the CPU overhead of such behaviour.
272 */
273 static void
274 xlog_cil_alloc_shadow_bufs(
275 struct xlog *log,
276 struct xfs_trans *tp)
277 {
278 struct xfs_log_item *lip;
279
280 list_for_each_entry(lip, &tp->t_items, li_trans) {
281 struct xfs_log_vec *lv;
282 int niovecs = 0;
283 int nbytes = 0;
284 int buf_size;
285 bool ordered = false;
286
287 /* Skip items which aren't dirty in this transaction. */
288 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
289 continue;
290
291 /* get number of vecs and size of data to be stored */
292 lip->li_ops->iop_size(lip, &niovecs, &nbytes);
293
294 /*
295 * Ordered items need to be tracked but we do not wish to write
296 * them. We need a logvec to track the object, but we do not
297 * need an iovec or buffer to be allocated for copying data.
298 */
299 if (niovecs == XFS_LOG_VEC_ORDERED) {
300 ordered = true;
301 niovecs = 0;
302 nbytes = 0;
303 }
304
305 /*
306 * We 64-bit align the length of each iovec so that the start of
307 * the next one is naturally aligned. We'll need to account for
308 * that slack space here.
309 *
310 * We also add the xlog_op_header to each region when
311 * formatting, but that's not accounted to the size of the item
312 * at this point. Hence we'll need an addition number of bytes
313 * for each vector to hold an opheader.
314 *
315 * Then round nbytes up to 64-bit alignment so that the initial
316 * buffer alignment is easy to calculate and verify.
317 */
318 nbytes += niovecs *
319 (sizeof(uint64_t) + sizeof(struct xlog_op_header));
320 nbytes = round_up(nbytes, sizeof(uint64_t));
321
322 /*
323 * The data buffer needs to start 64-bit aligned, so round up
324 * that space to ensure we can align it appropriately and not
325 * overrun the buffer.
326 */
327 buf_size = nbytes + xlog_cil_iovec_space(niovecs);
328
329 /*
330 * if we have no shadow buffer, or it is too small, we need to
331 * reallocate it.
332 */
333 if (!lip->li_lv_shadow ||
334 buf_size > lip->li_lv_shadow->lv_size) {
335 /*
336 * We free and allocate here as a realloc would copy
337 * unnecessary data. We don't use kvzalloc() for the
338 * same reason - we don't need to zero the data area in
339 * the buffer, only the log vector header and the iovec
340 * storage.
341 */
342 kvfree(lip->li_lv_shadow);
343 lv = xlog_kvmalloc(buf_size);
344
345 memset(lv, 0, xlog_cil_iovec_space(niovecs));
346
347 INIT_LIST_HEAD(&lv->lv_list);
348 lv->lv_item = lip;
349 lv->lv_size = buf_size;
350 if (ordered)
351 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
352 else
353 lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
354 lip->li_lv_shadow = lv;
355 } else {
356 /* same or smaller, optimise common overwrite case */
357 lv = lip->li_lv_shadow;
358 if (ordered)
359 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
360 else
361 lv->lv_buf_len = 0;
362 lv->lv_bytes = 0;
363 }
364
365 /* Ensure the lv is set up according to ->iop_size */
366 lv->lv_niovecs = niovecs;
367
368 /* The allocated data region lies beyond the iovec region */
369 lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs);
370 }
371
372 }
373
374 /*
375 * Prepare the log item for insertion into the CIL. Calculate the difference in
376 * log space it will consume, and if it is a new item pin it as well.
377 */
378 STATIC void
379 xfs_cil_prepare_item(
380 struct xlog *log,
381 struct xfs_log_vec *lv,
382 struct xfs_log_vec *old_lv,
383 int *diff_len)
384 {
385 /* Account for the new LV being passed in */
386 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED)
387 *diff_len += lv->lv_bytes;
388
389 /*
390 * If there is no old LV, this is the first time we've seen the item in
391 * this CIL context and so we need to pin it. If we are replacing the
392 * old_lv, then remove the space it accounts for and make it the shadow
393 * buffer for later freeing. In both cases we are now switching to the
394 * shadow buffer, so update the pointer to it appropriately.
395 */
396 if (!old_lv) {
397 if (lv->lv_item->li_ops->iop_pin)
398 lv->lv_item->li_ops->iop_pin(lv->lv_item);
399 lv->lv_item->li_lv_shadow = NULL;
400 } else if (old_lv != lv) {
401 ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
402
403 *diff_len -= old_lv->lv_bytes;
404 lv->lv_item->li_lv_shadow = old_lv;
405 }
406
407 /* attach new log vector to log item */
408 lv->lv_item->li_lv = lv;
409
410 /*
411 * If this is the first time the item is being committed to the
412 * CIL, store the sequence number on the log item so we can
413 * tell in future commits whether this is the first checkpoint
414 * the item is being committed into.
415 */
416 if (!lv->lv_item->li_seq)
417 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
418 }
419
420 /*
421 * Format log item into a flat buffers
422 *
423 * For delayed logging, we need to hold a formatted buffer containing all the
424 * changes on the log item. This enables us to relog the item in memory and
425 * write it out asynchronously without needing to relock the object that was
426 * modified at the time it gets written into the iclog.
427 *
428 * This function takes the prepared log vectors attached to each log item, and
429 * formats the changes into the log vector buffer. The buffer it uses is
430 * dependent on the current state of the vector in the CIL - the shadow lv is
431 * guaranteed to be large enough for the current modification, but we will only
432 * use that if we can't reuse the existing lv. If we can't reuse the existing
433 * lv, then simple swap it out for the shadow lv. We don't free it - that is
434 * done lazily either by th enext modification or the freeing of the log item.
435 *
436 * We don't set up region headers during this process; we simply copy the
437 * regions into the flat buffer. We can do this because we still have to do a
438 * formatting step to write the regions into the iclog buffer. Writing the
439 * ophdrs during the iclog write means that we can support splitting large
440 * regions across iclog boundares without needing a change in the format of the
441 * item/region encapsulation.
442 *
443 * Hence what we need to do now is change the rewrite the vector array to point
444 * to the copied region inside the buffer we just allocated. This allows us to
445 * format the regions into the iclog as though they are being formatted
446 * directly out of the objects themselves.
447 */
448 static void
449 xlog_cil_insert_format_items(
450 struct xlog *log,
451 struct xfs_trans *tp,
452 int *diff_len)
453 {
454 struct xfs_log_item *lip;
455
456 /* Bail out if we didn't find a log item. */
457 if (list_empty(&tp->t_items)) {
458 ASSERT(0);
459 return;
460 }
461
462 list_for_each_entry(lip, &tp->t_items, li_trans) {
463 struct xfs_log_vec *lv;
464 struct xfs_log_vec *old_lv = NULL;
465 struct xfs_log_vec *shadow;
466 bool ordered = false;
467
468 /* Skip items which aren't dirty in this transaction. */
469 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
470 continue;
471
472 /*
473 * The formatting size information is already attached to
474 * the shadow lv on the log item.
475 */
476 shadow = lip->li_lv_shadow;
477 if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED)
478 ordered = true;
479
480 /* Skip items that do not have any vectors for writing */
481 if (!shadow->lv_niovecs && !ordered)
482 continue;
483
484 /* compare to existing item size */
485 old_lv = lip->li_lv;
486 if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) {
487 /* same or smaller, optimise common overwrite case */
488 lv = lip->li_lv;
489
490 if (ordered)
491 goto insert;
492
493 /*
494 * set the item up as though it is a new insertion so
495 * that the space reservation accounting is correct.
496 */
497 *diff_len -= lv->lv_bytes;
498
499 /* Ensure the lv is set up according to ->iop_size */
500 lv->lv_niovecs = shadow->lv_niovecs;
501
502 /* reset the lv buffer information for new formatting */
503 lv->lv_buf_len = 0;
504 lv->lv_bytes = 0;
505 lv->lv_buf = (char *)lv +
506 xlog_cil_iovec_space(lv->lv_niovecs);
507 } else {
508 /* switch to shadow buffer! */
509 lv = shadow;
510 lv->lv_item = lip;
511 if (ordered) {
512 /* track as an ordered logvec */
513 ASSERT(lip->li_lv == NULL);
514 goto insert;
515 }
516 }
517
518 ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));
519 lip->li_ops->iop_format(lip, lv);
520 insert:
521 xfs_cil_prepare_item(log, lv, old_lv, diff_len);
522 }
523 }
524
525 /*
526 * The use of lockless waitqueue_active() requires that the caller has
527 * serialised itself against the wakeup call in xlog_cil_push_work(). That
528 * can be done by either holding the push lock or the context lock.
529 */
530 static inline bool
531 xlog_cil_over_hard_limit(
532 struct xlog *log,
533 int32_t space_used)
534 {
535 if (waitqueue_active(&log->l_cilp->xc_push_wait))
536 return true;
537 if (space_used >= XLOG_CIL_BLOCKING_SPACE_LIMIT(log))
538 return true;
539 return false;
540 }
541
542 /*
543 * Insert the log items into the CIL and calculate the difference in space
544 * consumed by the item. Add the space to the checkpoint ticket and calculate
545 * if the change requires additional log metadata. If it does, take that space
546 * as well. Remove the amount of space we added to the checkpoint ticket from
547 * the current transaction ticket so that the accounting works out correctly.
548 */
549 static void
550 xlog_cil_insert_items(
551 struct xlog *log,
552 struct xfs_trans *tp,
553 uint32_t released_space)
554 {
555 struct xfs_cil *cil = log->l_cilp;
556 struct xfs_cil_ctx *ctx = cil->xc_ctx;
557 struct xfs_log_item *lip;
558 int len = 0;
559 int iovhdr_res = 0, split_res = 0, ctx_res = 0;
560 int space_used;
561 int order;
562 unsigned int cpu_nr;
563 struct xlog_cil_pcp *cilpcp;
564
565 ASSERT(tp);
566
567 /*
568 * We can do this safely because the context can't checkpoint until we
569 * are done so it doesn't matter exactly how we update the CIL.
570 */
571 xlog_cil_insert_format_items(log, tp, &len);
572
573 /*
574 * Subtract the space released by intent cancelation from the space we
575 * consumed so that we remove it from the CIL space and add it back to
576 * the current transaction reservation context.
577 */
578 len -= released_space;
579
580 /*
581 * Grab the per-cpu pointer for the CIL before we start any accounting.
582 * That ensures that we are running with pre-emption disabled and so we
583 * can't be scheduled away between split sample/update operations that
584 * are done without outside locking to serialise them.
585 */
586 cpu_nr = get_cpu();
587 cilpcp = this_cpu_ptr(cil->xc_pcp);
588
589 /* Tell the future push that there was work added by this CPU. */
590 if (!cpumask_test_cpu(cpu_nr, &ctx->cil_pcpmask))
591 cpumask_test_and_set_cpu(cpu_nr, &ctx->cil_pcpmask);
592
593 /*
594 * We need to take the CIL checkpoint unit reservation on the first
595 * commit into the CIL. Test the XLOG_CIL_EMPTY bit first so we don't
596 * unnecessarily do an atomic op in the fast path here. We can clear the
597 * XLOG_CIL_EMPTY bit as we are under the xc_ctx_lock here and that
598 * needs to be held exclusively to reset the XLOG_CIL_EMPTY bit.
599 */
600 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags) &&
601 test_and_clear_bit(XLOG_CIL_EMPTY, &cil->xc_flags))
602 ctx_res = ctx->ticket->t_unit_res;
603
604 /*
605 * Check if we need to steal iclog headers. atomic_read() is not a
606 * locked atomic operation, so we can check the value before we do any
607 * real atomic ops in the fast path. If we've already taken the CIL unit
608 * reservation from this commit, we've already got one iclog header
609 * space reserved so we have to account for that otherwise we risk
610 * overrunning the reservation on this ticket.
611 *
612 * If the CIL is already at the hard limit, we might need more header
613 * space that originally reserved. So steal more header space from every
614 * commit that occurs once we are over the hard limit to ensure the CIL
615 * push won't run out of reservation space.
616 *
617 * This can steal more than we need, but that's OK.
618 *
619 * The cil->xc_ctx_lock provides the serialisation necessary for safely
620 * calling xlog_cil_over_hard_limit() in this context.
621 */
622 space_used = atomic_read(&ctx->space_used) + cilpcp->space_used + len;
623 if (atomic_read(&cil->xc_iclog_hdrs) > 0 ||
624 xlog_cil_over_hard_limit(log, space_used)) {
625 split_res = log->l_iclog_hsize +
626 sizeof(struct xlog_op_header);
627 if (ctx_res)
628 ctx_res += split_res * (tp->t_ticket->t_iclog_hdrs - 1);
629 else
630 ctx_res = split_res * tp->t_ticket->t_iclog_hdrs;
631 atomic_sub(tp->t_ticket->t_iclog_hdrs, &cil->xc_iclog_hdrs);
632 }
633 cilpcp->space_reserved += ctx_res;
634
635 /*
636 * Accurately account when over the soft limit, otherwise fold the
637 * percpu count into the global count if over the per-cpu threshold.
638 */
639 if (!test_bit(XLOG_CIL_PCP_SPACE, &cil->xc_flags)) {
640 atomic_add(len, &ctx->space_used);
641 } else if (cilpcp->space_used + len >
642 (XLOG_CIL_SPACE_LIMIT(log) / num_online_cpus())) {
643 space_used = atomic_add_return(cilpcp->space_used + len,
644 &ctx->space_used);
645 cilpcp->space_used = 0;
646
647 /*
648 * If we just transitioned over the soft limit, we need to
649 * transition to the global atomic counter.
650 */
651 if (space_used >= XLOG_CIL_SPACE_LIMIT(log))
652 xlog_cil_insert_pcp_aggregate(cil, ctx);
653 } else {
654 cilpcp->space_used += len;
655 }
656 /* attach the transaction to the CIL if it has any busy extents */
657 if (!list_empty(&tp->t_busy))
658 list_splice_init(&tp->t_busy, &cilpcp->busy_extents);
659
660 /*
661 * Now update the order of everything modified in the transaction
662 * and insert items into the CIL if they aren't already there.
663 * We do this here so we only need to take the CIL lock once during
664 * the transaction commit.
665 */
666 order = atomic_inc_return(&ctx->order_id);
667 list_for_each_entry(lip, &tp->t_items, li_trans) {
668 /* Skip items which aren't dirty in this transaction. */
669 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
670 continue;
671
672 lip->li_order_id = order;
673 if (!list_empty(&lip->li_cil))
674 continue;
675 list_add_tail(&lip->li_cil, &cilpcp->log_items);
676 }
677 put_cpu();
678
679 /*
680 * If we've overrun the reservation, dump the tx details before we move
681 * the log items. Shutdown is imminent...
682 */
683 tp->t_ticket->t_curr_res -= ctx_res + len;
684 if (WARN_ON(tp->t_ticket->t_curr_res < 0)) {
685 xfs_warn(log->l_mp, "Transaction log reservation overrun:");
686 xfs_warn(log->l_mp,
687 " log items: %d bytes (iov hdrs: %d bytes)",
688 len, iovhdr_res);
689 xfs_warn(log->l_mp, " split region headers: %d bytes",
690 split_res);
691 xfs_warn(log->l_mp, " ctx ticket: %d bytes", ctx_res);
692 xlog_print_trans(tp);
693 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
694 }
695 }
696
697 static inline void
698 xlog_cil_ail_insert_batch(
699 struct xfs_ail *ailp,
700 struct xfs_ail_cursor *cur,
701 struct xfs_log_item **log_items,
702 int nr_items,
703 xfs_lsn_t commit_lsn)
704 {
705 int i;
706
707 spin_lock(&ailp->ail_lock);
708 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
709 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
710
711 for (i = 0; i < nr_items; i++) {
712 struct xfs_log_item *lip = log_items[i];
713
714 if (lip->li_ops->iop_unpin)
715 lip->li_ops->iop_unpin(lip, 0);
716 }
717 }
718
719 /*
720 * Take the checkpoint's log vector chain of items and insert the attached log
721 * items into the AIL. This uses bulk insertion techniques to minimise AIL lock
722 * traffic.
723 *
724 * The AIL tracks log items via the start record LSN of the checkpoint,
725 * not the commit record LSN. This is because we can pipeline multiple
726 * checkpoints, and so the start record of checkpoint N+1 can be
727 * written before the commit record of checkpoint N. i.e:
728 *
729 * start N commit N
730 * +-------------+------------+----------------+
731 * start N+1 commit N+1
732 *
733 * The tail of the log cannot be moved to the LSN of commit N when all
734 * the items of that checkpoint are written back, because then the
735 * start record for N+1 is no longer in the active portion of the log
736 * and recovery will fail/corrupt the filesystem.
737 *
738 * Hence when all the log items in checkpoint N are written back, the
739 * tail of the log most now only move as far forwards as the start LSN
740 * of checkpoint N+1.
741 *
742 * If we are called with the aborted flag set, it is because a log write during
743 * a CIL checkpoint commit has failed. In this case, all the items in the
744 * checkpoint have already gone through iop_committed and iop_committing, which
745 * means that checkpoint commit abort handling is treated exactly the same as an
746 * iclog write error even though we haven't started any IO yet. Hence in this
747 * case all we need to do is iop_committed processing, followed by an
748 * iop_unpin(aborted) call.
749 *
750 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
751 * at the end of the AIL, the insert cursor avoids the need to walk the AIL to
752 * find the insertion point on every xfs_log_item_batch_insert() call. This
753 * saves a lot of needless list walking and is a net win, even though it
754 * slightly increases that amount of AIL lock traffic to set it up and tear it
755 * down.
756 */
757 static void
758 xlog_cil_ail_insert(
759 struct xfs_cil_ctx *ctx,
760 bool aborted)
761 {
762 #define LOG_ITEM_BATCH_SIZE 32
763 struct xfs_ail *ailp = ctx->cil->xc_log->l_ailp;
764 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
765 struct xfs_log_vec *lv;
766 struct xfs_ail_cursor cur;
767 xfs_lsn_t old_head;
768 int i = 0;
769
770 /*
771 * Update the AIL head LSN with the commit record LSN of this
772 * checkpoint. As iclogs are always completed in order, this should
773 * always be the same (as iclogs can contain multiple commit records) or
774 * higher LSN than the current head. We do this before insertion of the
775 * items so that log space checks during insertion will reflect the
776 * space that this checkpoint has already consumed. We call
777 * xfs_ail_update_finish() so that tail space and space-based wakeups
778 * will be recalculated appropriately.
779 */
780 ASSERT(XFS_LSN_CMP(ctx->commit_lsn, ailp->ail_head_lsn) >= 0 ||
781 aborted);
782 spin_lock(&ailp->ail_lock);
783 xfs_trans_ail_cursor_last(ailp, &cur, ctx->start_lsn);
784 old_head = ailp->ail_head_lsn;
785 ailp->ail_head_lsn = ctx->commit_lsn;
786 /* xfs_ail_update_finish() drops the ail_lock */
787 xfs_ail_update_finish(ailp, NULLCOMMITLSN);
788
789 /*
790 * We move the AIL head forwards to account for the space used in the
791 * log before we remove that space from the grant heads. This prevents a
792 * transient condition where reservation space appears to become
793 * available on return, only for it to disappear again immediately as
794 * the AIL head update accounts in the log tail space.
795 */
796 smp_wmb(); /* paired with smp_rmb in xlog_grant_space_left */
797 xlog_grant_return_space(ailp->ail_log, old_head, ailp->ail_head_lsn);
798
799 /* unpin all the log items */
800 list_for_each_entry(lv, &ctx->lv_chain, lv_list) {
801 struct xfs_log_item *lip = lv->lv_item;
802 xfs_lsn_t item_lsn;
803
804 if (aborted)
805 set_bit(XFS_LI_ABORTED, &lip->li_flags);
806
807 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
808 lip->li_ops->iop_release(lip);
809 continue;
810 }
811
812 if (lip->li_ops->iop_committed)
813 item_lsn = lip->li_ops->iop_committed(lip,
814 ctx->start_lsn);
815 else
816 item_lsn = ctx->start_lsn;
817
818 /* item_lsn of -1 means the item needs no further processing */
819 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
820 continue;
821
822 /*
823 * if we are aborting the operation, no point in inserting the
824 * object into the AIL as we are in a shutdown situation.
825 */
826 if (aborted) {
827 ASSERT(xlog_is_shutdown(ailp->ail_log));
828 if (lip->li_ops->iop_unpin)
829 lip->li_ops->iop_unpin(lip, 1);
830 continue;
831 }
832
833 if (item_lsn != ctx->start_lsn) {
834
835 /*
836 * Not a bulk update option due to unusual item_lsn.
837 * Push into AIL immediately, rechecking the lsn once
838 * we have the ail lock. Then unpin the item. This does
839 * not affect the AIL cursor the bulk insert path is
840 * using.
841 */
842 spin_lock(&ailp->ail_lock);
843 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
844 xfs_trans_ail_update(ailp, lip, item_lsn);
845 else
846 spin_unlock(&ailp->ail_lock);
847 if (lip->li_ops->iop_unpin)
848 lip->li_ops->iop_unpin(lip, 0);
849 continue;
850 }
851
852 /* Item is a candidate for bulk AIL insert. */
853 log_items[i++] = lv->lv_item;
854 if (i >= LOG_ITEM_BATCH_SIZE) {
855 xlog_cil_ail_insert_batch(ailp, &cur, log_items,
856 LOG_ITEM_BATCH_SIZE, ctx->start_lsn);
857 i = 0;
858 }
859 }
860
861 /* make sure we insert the remainder! */
862 if (i)
863 xlog_cil_ail_insert_batch(ailp, &cur, log_items, i,
864 ctx->start_lsn);
865
866 spin_lock(&ailp->ail_lock);
867 xfs_trans_ail_cursor_done(&cur);
868 spin_unlock(&ailp->ail_lock);
869 }
870
871 static void
872 xlog_cil_free_logvec(
873 struct list_head *lv_chain)
874 {
875 struct xfs_log_vec *lv;
876
877 while (!list_empty(lv_chain)) {
878 lv = list_first_entry(lv_chain, struct xfs_log_vec, lv_list);
879 list_del_init(&lv->lv_list);
880 kvfree(lv);
881 }
882 }
883
884 /*
885 * Mark all items committed and clear busy extents. We free the log vector
886 * chains in a separate pass so that we unpin the log items as quickly as
887 * possible.
888 */
889 static void
890 xlog_cil_committed(
891 struct xfs_cil_ctx *ctx)
892 {
893 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
894 bool abort = xlog_is_shutdown(ctx->cil->xc_log);
895
896 /*
897 * If the I/O failed, we're aborting the commit and already shutdown.
898 * Wake any commit waiters before aborting the log items so we don't
899 * block async log pushers on callbacks. Async log pushers explicitly do
900 * not wait on log force completion because they may be holding locks
901 * required to unpin items.
902 */
903 if (abort) {
904 spin_lock(&ctx->cil->xc_push_lock);
905 wake_up_all(&ctx->cil->xc_start_wait);
906 wake_up_all(&ctx->cil->xc_commit_wait);
907 spin_unlock(&ctx->cil->xc_push_lock);
908 }
909
910 xlog_cil_ail_insert(ctx, abort);
911
912 xfs_extent_busy_sort(&ctx->busy_extents.extent_list);
913 xfs_extent_busy_clear(mp, &ctx->busy_extents.extent_list,
914 xfs_has_discard(mp) && !abort);
915
916 spin_lock(&ctx->cil->xc_push_lock);
917 list_del(&ctx->committing);
918 spin_unlock(&ctx->cil->xc_push_lock);
919
920 xlog_cil_free_logvec(&ctx->lv_chain);
921
922 if (!list_empty(&ctx->busy_extents.extent_list)) {
923 ctx->busy_extents.mount = mp;
924 ctx->busy_extents.owner = ctx;
925 xfs_discard_extents(mp, &ctx->busy_extents);
926 return;
927 }
928
929 kfree(ctx);
930 }
931
932 void
933 xlog_cil_process_committed(
934 struct list_head *list)
935 {
936 struct xfs_cil_ctx *ctx;
937
938 while ((ctx = list_first_entry_or_null(list,
939 struct xfs_cil_ctx, iclog_entry))) {
940 list_del(&ctx->iclog_entry);
941 xlog_cil_committed(ctx);
942 }
943 }
944
945 /*
946 * Record the LSN of the iclog we were just granted space to start writing into.
947 * If the context doesn't have a start_lsn recorded, then this iclog will
948 * contain the start record for the checkpoint. Otherwise this write contains
949 * the commit record for the checkpoint.
950 */
951 void
952 xlog_cil_set_ctx_write_state(
953 struct xfs_cil_ctx *ctx,
954 struct xlog_in_core *iclog)
955 {
956 struct xfs_cil *cil = ctx->cil;
957 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
958
959 ASSERT(!ctx->commit_lsn);
960 if (!ctx->start_lsn) {
961 spin_lock(&cil->xc_push_lock);
962 /*
963 * The LSN we need to pass to the log items on transaction
964 * commit is the LSN reported by the first log vector write, not
965 * the commit lsn. If we use the commit record lsn then we can
966 * move the grant write head beyond the tail LSN and overwrite
967 * it.
968 */
969 ctx->start_lsn = lsn;
970 wake_up_all(&cil->xc_start_wait);
971 spin_unlock(&cil->xc_push_lock);
972
973 /*
974 * Make sure the metadata we are about to overwrite in the log
975 * has been flushed to stable storage before this iclog is
976 * issued.
977 */
978 spin_lock(&cil->xc_log->l_icloglock);
979 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
980 spin_unlock(&cil->xc_log->l_icloglock);
981 return;
982 }
983
984 /*
985 * Take a reference to the iclog for the context so that we still hold
986 * it when xlog_write is done and has released it. This means the
987 * context controls when the iclog is released for IO.
988 */
989 atomic_inc(&iclog->ic_refcnt);
990
991 /*
992 * xlog_state_get_iclog_space() guarantees there is enough space in the
993 * iclog for an entire commit record, so we can attach the context
994 * callbacks now. This needs to be done before we make the commit_lsn
995 * visible to waiters so that checkpoints with commit records in the
996 * same iclog order their IO completion callbacks in the same order that
997 * the commit records appear in the iclog.
998 */
999 spin_lock(&cil->xc_log->l_icloglock);
1000 list_add_tail(&ctx->iclog_entry, &iclog->ic_callbacks);
1001 spin_unlock(&cil->xc_log->l_icloglock);
1002
1003 /*
1004 * Now we can record the commit LSN and wake anyone waiting for this
1005 * sequence to have the ordered commit record assigned to a physical
1006 * location in the log.
1007 */
1008 spin_lock(&cil->xc_push_lock);
1009 ctx->commit_iclog = iclog;
1010 ctx->commit_lsn = lsn;
1011 wake_up_all(&cil->xc_commit_wait);
1012 spin_unlock(&cil->xc_push_lock);
1013 }
1014
1015
1016 /*
1017 * Ensure that the order of log writes follows checkpoint sequence order. This
1018 * relies on the context LSN being zero until the log write has guaranteed the
1019 * LSN that the log write will start at via xlog_state_get_iclog_space().
1020 */
1021 enum _record_type {
1022 _START_RECORD,
1023 _COMMIT_RECORD,
1024 };
1025
1026 static int
1027 xlog_cil_order_write(
1028 struct xfs_cil *cil,
1029 xfs_csn_t sequence,
1030 enum _record_type record)
1031 {
1032 struct xfs_cil_ctx *ctx;
1033
1034 restart:
1035 spin_lock(&cil->xc_push_lock);
1036 list_for_each_entry(ctx, &cil->xc_committing, committing) {
1037 /*
1038 * Avoid getting stuck in this loop because we were woken by the
1039 * shutdown, but then went back to sleep once already in the
1040 * shutdown state.
1041 */
1042 if (xlog_is_shutdown(cil->xc_log)) {
1043 spin_unlock(&cil->xc_push_lock);
1044 return -EIO;
1045 }
1046
1047 /*
1048 * Higher sequences will wait for this one so skip them.
1049 * Don't wait for our own sequence, either.
1050 */
1051 if (ctx->sequence >= sequence)
1052 continue;
1053
1054 /* Wait until the LSN for the record has been recorded. */
1055 switch (record) {
1056 case _START_RECORD:
1057 if (!ctx->start_lsn) {
1058 xlog_wait(&cil->xc_start_wait, &cil->xc_push_lock);
1059 goto restart;
1060 }
1061 break;
1062 case _COMMIT_RECORD:
1063 if (!ctx->commit_lsn) {
1064 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
1065 goto restart;
1066 }
1067 break;
1068 }
1069 }
1070 spin_unlock(&cil->xc_push_lock);
1071 return 0;
1072 }
1073
1074 /*
1075 * Write out the log vector change now attached to the CIL context. This will
1076 * write a start record that needs to be strictly ordered in ascending CIL
1077 * sequence order so that log recovery will always use in-order start LSNs when
1078 * replaying checkpoints.
1079 */
1080 static int
1081 xlog_cil_write_chain(
1082 struct xfs_cil_ctx *ctx,
1083 uint32_t chain_len)
1084 {
1085 struct xlog *log = ctx->cil->xc_log;
1086 int error;
1087
1088 error = xlog_cil_order_write(ctx->cil, ctx->sequence, _START_RECORD);
1089 if (error)
1090 return error;
1091 return xlog_write(log, ctx, &ctx->lv_chain, ctx->ticket, chain_len);
1092 }
1093
1094 /*
1095 * Write out the commit record of a checkpoint transaction to close off a
1096 * running log write. These commit records are strictly ordered in ascending CIL
1097 * sequence order so that log recovery will always replay the checkpoints in the
1098 * correct order.
1099 */
1100 static int
1101 xlog_cil_write_commit_record(
1102 struct xfs_cil_ctx *ctx)
1103 {
1104 struct xlog *log = ctx->cil->xc_log;
1105 struct xlog_op_header ophdr = {
1106 .oh_clientid = XFS_TRANSACTION,
1107 .oh_tid = cpu_to_be32(ctx->ticket->t_tid),
1108 .oh_flags = XLOG_COMMIT_TRANS,
1109 };
1110 struct xfs_log_iovec reg = {
1111 .i_addr = &ophdr,
1112 .i_len = sizeof(struct xlog_op_header),
1113 .i_type = XLOG_REG_TYPE_COMMIT,
1114 };
1115 struct xfs_log_vec vec = {
1116 .lv_niovecs = 1,
1117 .lv_iovecp = ®,
1118 };
1119 int error;
1120 LIST_HEAD(lv_chain);
1121 list_add(&vec.lv_list, &lv_chain);
1122
1123 if (xlog_is_shutdown(log))
1124 return -EIO;
1125
1126 error = xlog_cil_order_write(ctx->cil, ctx->sequence, _COMMIT_RECORD);
1127 if (error)
1128 return error;
1129
1130 /* account for space used by record data */
1131 ctx->ticket->t_curr_res -= reg.i_len;
1132 error = xlog_write(log, ctx, &lv_chain, ctx->ticket, reg.i_len);
1133 if (error)
1134 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1135 return error;
1136 }
1137
1138 struct xlog_cil_trans_hdr {
1139 struct xlog_op_header oph[2];
1140 struct xfs_trans_header thdr;
1141 struct xfs_log_iovec lhdr[2];
1142 };
1143
1144 /*
1145 * Build a checkpoint transaction header to begin the journal transaction. We
1146 * need to account for the space used by the transaction header here as it is
1147 * not accounted for in xlog_write().
1148 *
1149 * This is the only place we write a transaction header, so we also build the
1150 * log opheaders that indicate the start of a log transaction and wrap the
1151 * transaction header. We keep the start record in it's own log vector rather
1152 * than compacting them into a single region as this ends up making the logic
1153 * in xlog_write() for handling empty opheaders for start, commit and unmount
1154 * records much simpler.
1155 */
1156 static void
1157 xlog_cil_build_trans_hdr(
1158 struct xfs_cil_ctx *ctx,
1159 struct xlog_cil_trans_hdr *hdr,
1160 struct xfs_log_vec *lvhdr,
1161 int num_iovecs)
1162 {
1163 struct xlog_ticket *tic = ctx->ticket;
1164 __be32 tid = cpu_to_be32(tic->t_tid);
1165
1166 memset(hdr, 0, sizeof(*hdr));
1167
1168 /* Log start record */
1169 hdr->oph[0].oh_tid = tid;
1170 hdr->oph[0].oh_clientid = XFS_TRANSACTION;
1171 hdr->oph[0].oh_flags = XLOG_START_TRANS;
1172
1173 /* log iovec region pointer */
1174 hdr->lhdr[0].i_addr = &hdr->oph[0];
1175 hdr->lhdr[0].i_len = sizeof(struct xlog_op_header);
1176 hdr->lhdr[0].i_type = XLOG_REG_TYPE_LRHEADER;
1177
1178 /* log opheader */
1179 hdr->oph[1].oh_tid = tid;
1180 hdr->oph[1].oh_clientid = XFS_TRANSACTION;
1181 hdr->oph[1].oh_len = cpu_to_be32(sizeof(struct xfs_trans_header));
1182
1183 /* transaction header in host byte order format */
1184 hdr->thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
1185 hdr->thdr.th_type = XFS_TRANS_CHECKPOINT;
1186 hdr->thdr.th_tid = tic->t_tid;
1187 hdr->thdr.th_num_items = num_iovecs;
1188
1189 /* log iovec region pointer */
1190 hdr->lhdr[1].i_addr = &hdr->oph[1];
1191 hdr->lhdr[1].i_len = sizeof(struct xlog_op_header) +
1192 sizeof(struct xfs_trans_header);
1193 hdr->lhdr[1].i_type = XLOG_REG_TYPE_TRANSHDR;
1194
1195 lvhdr->lv_niovecs = 2;
1196 lvhdr->lv_iovecp = &hdr->lhdr[0];
1197 lvhdr->lv_bytes = hdr->lhdr[0].i_len + hdr->lhdr[1].i_len;
1198
1199 tic->t_curr_res -= lvhdr->lv_bytes;
1200 }
1201
1202 /*
1203 * CIL item reordering compare function. We want to order in ascending ID order,
1204 * but we want to leave items with the same ID in the order they were added to
1205 * the list. This is important for operations like reflink where we log 4 order
1206 * dependent intents in a single transaction when we overwrite an existing
1207 * shared extent with a new shared extent. i.e. BUI(unmap), CUI(drop),
1208 * CUI (inc), BUI(remap)...
1209 */
1210 static int
1211 xlog_cil_order_cmp(
1212 void *priv,
1213 const struct list_head *a,
1214 const struct list_head *b)
1215 {
1216 struct xfs_log_vec *l1 = container_of(a, struct xfs_log_vec, lv_list);
1217 struct xfs_log_vec *l2 = container_of(b, struct xfs_log_vec, lv_list);
1218
1219 return l1->lv_order_id > l2->lv_order_id;
1220 }
1221
1222 /*
1223 * Pull all the log vectors off the items in the CIL, and remove the items from
1224 * the CIL. We don't need the CIL lock here because it's only needed on the
1225 * transaction commit side which is currently locked out by the flush lock.
1226 *
1227 * If a log item is marked with a whiteout, we do not need to write it to the
1228 * journal and so we just move them to the whiteout list for the caller to
1229 * dispose of appropriately.
1230 */
1231 static void
1232 xlog_cil_build_lv_chain(
1233 struct xfs_cil_ctx *ctx,
1234 struct list_head *whiteouts,
1235 uint32_t *num_iovecs,
1236 uint32_t *num_bytes)
1237 {
1238 while (!list_empty(&ctx->log_items)) {
1239 struct xfs_log_item *item;
1240 struct xfs_log_vec *lv;
1241
1242 item = list_first_entry(&ctx->log_items,
1243 struct xfs_log_item, li_cil);
1244
1245 if (test_bit(XFS_LI_WHITEOUT, &item->li_flags)) {
1246 list_move(&item->li_cil, whiteouts);
1247 trace_xfs_cil_whiteout_skip(item);
1248 continue;
1249 }
1250
1251 lv = item->li_lv;
1252 lv->lv_order_id = item->li_order_id;
1253
1254 /* we don't write ordered log vectors */
1255 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED)
1256 *num_bytes += lv->lv_bytes;
1257 *num_iovecs += lv->lv_niovecs;
1258 list_add_tail(&lv->lv_list, &ctx->lv_chain);
1259
1260 list_del_init(&item->li_cil);
1261 item->li_order_id = 0;
1262 item->li_lv = NULL;
1263 }
1264 }
1265
1266 static void
1267 xlog_cil_cleanup_whiteouts(
1268 struct list_head *whiteouts)
1269 {
1270 while (!list_empty(whiteouts)) {
1271 struct xfs_log_item *item = list_first_entry(whiteouts,
1272 struct xfs_log_item, li_cil);
1273 list_del_init(&item->li_cil);
1274 trace_xfs_cil_whiteout_unpin(item);
1275 item->li_ops->iop_unpin(item, 1);
1276 }
1277 }
1278
1279 /*
1280 * Push the Committed Item List to the log.
1281 *
1282 * If the current sequence is the same as xc_push_seq we need to do a flush. If
1283 * xc_push_seq is less than the current sequence, then it has already been
1284 * flushed and we don't need to do anything - the caller will wait for it to
1285 * complete if necessary.
1286 *
1287 * xc_push_seq is checked unlocked against the sequence number for a match.
1288 * Hence we can allow log forces to run racily and not issue pushes for the
1289 * same sequence twice. If we get a race between multiple pushes for the same
1290 * sequence they will block on the first one and then abort, hence avoiding
1291 * needless pushes.
1292 *
1293 * This runs from a workqueue so it does not inherent any specific memory
1294 * allocation context. However, we do not want to block on memory reclaim
1295 * recursing back into the filesystem because this push may have been triggered
1296 * by memory reclaim itself. Hence we really need to run under full GFP_NOFS
1297 * contraints here.
1298 */
1299 static void
1300 xlog_cil_push_work(
1301 struct work_struct *work)
1302 {
1303 unsigned int nofs_flags = memalloc_nofs_save();
1304 struct xfs_cil_ctx *ctx =
1305 container_of(work, struct xfs_cil_ctx, push_work);
1306 struct xfs_cil *cil = ctx->cil;
1307 struct xlog *log = cil->xc_log;
1308 struct xfs_cil_ctx *new_ctx;
1309 int num_iovecs = 0;
1310 int num_bytes = 0;
1311 int error = 0;
1312 struct xlog_cil_trans_hdr thdr;
1313 struct xfs_log_vec lvhdr = {};
1314 xfs_csn_t push_seq;
1315 bool push_commit_stable;
1316 LIST_HEAD (whiteouts);
1317 struct xlog_ticket *ticket;
1318
1319 new_ctx = xlog_cil_ctx_alloc();
1320 new_ctx->ticket = xlog_cil_ticket_alloc(log);
1321
1322 down_write(&cil->xc_ctx_lock);
1323
1324 spin_lock(&cil->xc_push_lock);
1325 push_seq = cil->xc_push_seq;
1326 ASSERT(push_seq <= ctx->sequence);
1327 push_commit_stable = cil->xc_push_commit_stable;
1328 cil->xc_push_commit_stable = false;
1329
1330 /*
1331 * As we are about to switch to a new, empty CIL context, we no longer
1332 * need to throttle tasks on CIL space overruns. Wake any waiters that
1333 * the hard push throttle may have caught so they can start committing
1334 * to the new context. The ctx->xc_push_lock provides the serialisation
1335 * necessary for safely using the lockless waitqueue_active() check in
1336 * this context.
1337 */
1338 if (waitqueue_active(&cil->xc_push_wait))
1339 wake_up_all(&cil->xc_push_wait);
1340
1341 xlog_cil_push_pcp_aggregate(cil, ctx);
1342
1343 /*
1344 * Check if we've anything to push. If there is nothing, then we don't
1345 * move on to a new sequence number and so we have to be able to push
1346 * this sequence again later.
1347 */
1348 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags)) {
1349 cil->xc_push_seq = 0;
1350 spin_unlock(&cil->xc_push_lock);
1351 goto out_skip;
1352 }
1353
1354
1355 /* check for a previously pushed sequence */
1356 if (push_seq < ctx->sequence) {
1357 spin_unlock(&cil->xc_push_lock);
1358 goto out_skip;
1359 }
1360
1361 /*
1362 * We are now going to push this context, so add it to the committing
1363 * list before we do anything else. This ensures that anyone waiting on
1364 * this push can easily detect the difference between a "push in
1365 * progress" and "CIL is empty, nothing to do".
1366 *
1367 * IOWs, a wait loop can now check for:
1368 * the current sequence not being found on the committing list;
1369 * an empty CIL; and
1370 * an unchanged sequence number
1371 * to detect a push that had nothing to do and therefore does not need
1372 * waiting on. If the CIL is not empty, we get put on the committing
1373 * list before emptying the CIL and bumping the sequence number. Hence
1374 * an empty CIL and an unchanged sequence number means we jumped out
1375 * above after doing nothing.
1376 *
1377 * Hence the waiter will either find the commit sequence on the
1378 * committing list or the sequence number will be unchanged and the CIL
1379 * still dirty. In that latter case, the push has not yet started, and
1380 * so the waiter will have to continue trying to check the CIL
1381 * committing list until it is found. In extreme cases of delay, the
1382 * sequence may fully commit between the attempts the wait makes to wait
1383 * on the commit sequence.
1384 */
1385 list_add(&ctx->committing, &cil->xc_committing);
1386 spin_unlock(&cil->xc_push_lock);
1387
1388 xlog_cil_build_lv_chain(ctx, &whiteouts, &num_iovecs, &num_bytes);
1389
1390 /*
1391 * Switch the contexts so we can drop the context lock and move out
1392 * of a shared context. We can't just go straight to the commit record,
1393 * though - we need to synchronise with previous and future commits so
1394 * that the commit records are correctly ordered in the log to ensure
1395 * that we process items during log IO completion in the correct order.
1396 *
1397 * For example, if we get an EFI in one checkpoint and the EFD in the
1398 * next (e.g. due to log forces), we do not want the checkpoint with
1399 * the EFD to be committed before the checkpoint with the EFI. Hence
1400 * we must strictly order the commit records of the checkpoints so
1401 * that: a) the checkpoint callbacks are attached to the iclogs in the
1402 * correct order; and b) the checkpoints are replayed in correct order
1403 * in log recovery.
1404 *
1405 * Hence we need to add this context to the committing context list so
1406 * that higher sequences will wait for us to write out a commit record
1407 * before they do.
1408 *
1409 * xfs_log_force_seq requires us to mirror the new sequence into the cil
1410 * structure atomically with the addition of this sequence to the
1411 * committing list. This also ensures that we can do unlocked checks
1412 * against the current sequence in log forces without risking
1413 * deferencing a freed context pointer.
1414 */
1415 spin_lock(&cil->xc_push_lock);
1416 xlog_cil_ctx_switch(cil, new_ctx);
1417 spin_unlock(&cil->xc_push_lock);
1418 up_write(&cil->xc_ctx_lock);
1419
1420 /*
1421 * Sort the log vector chain before we add the transaction headers.
1422 * This ensures we always have the transaction headers at the start
1423 * of the chain.
1424 */
1425 list_sort(NULL, &ctx->lv_chain, xlog_cil_order_cmp);
1426
1427 /*
1428 * Build a checkpoint transaction header and write it to the log to
1429 * begin the transaction. We need to account for the space used by the
1430 * transaction header here as it is not accounted for in xlog_write().
1431 * Add the lvhdr to the head of the lv chain we pass to xlog_write() so
1432 * it gets written into the iclog first.
1433 */
1434 xlog_cil_build_trans_hdr(ctx, &thdr, &lvhdr, num_iovecs);
1435 num_bytes += lvhdr.lv_bytes;
1436 list_add(&lvhdr.lv_list, &ctx->lv_chain);
1437
1438 /*
1439 * Take the lvhdr back off the lv_chain immediately after calling
1440 * xlog_cil_write_chain() as it should not be passed to log IO
1441 * completion.
1442 */
1443 error = xlog_cil_write_chain(ctx, num_bytes);
1444 list_del(&lvhdr.lv_list);
1445 if (error)
1446 goto out_abort_free_ticket;
1447
1448 error = xlog_cil_write_commit_record(ctx);
1449 if (error)
1450 goto out_abort_free_ticket;
1451
1452 /*
1453 * Grab the ticket from the ctx so we can ungrant it after releasing the
1454 * commit_iclog. The ctx may be freed by the time we return from
1455 * releasing the commit_iclog (i.e. checkpoint has been completed and
1456 * callback run) so we can't reference the ctx after the call to
1457 * xlog_state_release_iclog().
1458 */
1459 ticket = ctx->ticket;
1460
1461 /*
1462 * If the checkpoint spans multiple iclogs, wait for all previous iclogs
1463 * to complete before we submit the commit_iclog. We can't use state
1464 * checks for this - ACTIVE can be either a past completed iclog or a
1465 * future iclog being filled, while WANT_SYNC through SYNC_DONE can be a
1466 * past or future iclog awaiting IO or ordered IO completion to be run.
1467 * In the latter case, if it's a future iclog and we wait on it, the we
1468 * will hang because it won't get processed through to ic_force_wait
1469 * wakeup until this commit_iclog is written to disk. Hence we use the
1470 * iclog header lsn and compare it to the commit lsn to determine if we
1471 * need to wait on iclogs or not.
1472 */
1473 spin_lock(&log->l_icloglock);
1474 if (ctx->start_lsn != ctx->commit_lsn) {
1475 xfs_lsn_t plsn;
1476
1477 plsn = be64_to_cpu(ctx->commit_iclog->ic_prev->ic_header.h_lsn);
1478 if (plsn && XFS_LSN_CMP(plsn, ctx->commit_lsn) < 0) {
1479 /*
1480 * Waiting on ic_force_wait orders the completion of
1481 * iclogs older than ic_prev. Hence we only need to wait
1482 * on the most recent older iclog here.
1483 */
1484 xlog_wait_on_iclog(ctx->commit_iclog->ic_prev);
1485 spin_lock(&log->l_icloglock);
1486 }
1487
1488 /*
1489 * We need to issue a pre-flush so that the ordering for this
1490 * checkpoint is correctly preserved down to stable storage.
1491 */
1492 ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
1493 }
1494
1495 /*
1496 * The commit iclog must be written to stable storage to guarantee
1497 * journal IO vs metadata writeback IO is correctly ordered on stable
1498 * storage.
1499 *
1500 * If the push caller needs the commit to be immediately stable and the
1501 * commit_iclog is not yet marked as XLOG_STATE_WANT_SYNC to indicate it
1502 * will be written when released, switch it's state to WANT_SYNC right
1503 * now.
1504 */
1505 ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA;
1506 if (push_commit_stable &&
1507 ctx->commit_iclog->ic_state == XLOG_STATE_ACTIVE)
1508 xlog_state_switch_iclogs(log, ctx->commit_iclog, 0);
1509 ticket = ctx->ticket;
1510 xlog_state_release_iclog(log, ctx->commit_iclog, ticket);
1511
1512 /* Not safe to reference ctx now! */
1513
1514 spin_unlock(&log->l_icloglock);
1515 xlog_cil_cleanup_whiteouts(&whiteouts);
1516 xfs_log_ticket_ungrant(log, ticket);
1517 memalloc_nofs_restore(nofs_flags);
1518 return;
1519
1520 out_skip:
1521 up_write(&cil->xc_ctx_lock);
1522 xfs_log_ticket_put(new_ctx->ticket);
1523 kfree(new_ctx);
1524 memalloc_nofs_restore(nofs_flags);
1525 return;
1526
1527 out_abort_free_ticket:
1528 ASSERT(xlog_is_shutdown(log));
1529 xlog_cil_cleanup_whiteouts(&whiteouts);
1530 if (!ctx->commit_iclog) {
1531 xfs_log_ticket_ungrant(log, ctx->ticket);
1532 xlog_cil_committed(ctx);
1533 memalloc_nofs_restore(nofs_flags);
1534 return;
1535 }
1536 spin_lock(&log->l_icloglock);
1537 ticket = ctx->ticket;
1538 xlog_state_release_iclog(log, ctx->commit_iclog, ticket);
1539 /* Not safe to reference ctx now! */
1540 spin_unlock(&log->l_icloglock);
1541 xfs_log_ticket_ungrant(log, ticket);
1542 memalloc_nofs_restore(nofs_flags);
1543 }
1544
1545 /*
1546 * We need to push CIL every so often so we don't cache more than we can fit in
1547 * the log. The limit really is that a checkpoint can't be more than half the
1548 * log (the current checkpoint is not allowed to overwrite the previous
1549 * checkpoint), but commit latency and memory usage limit this to a smaller
1550 * size.
1551 */
1552 static void
1553 xlog_cil_push_background(
1554 struct xlog *log)
1555 {
1556 struct xfs_cil *cil = log->l_cilp;
1557 int space_used = atomic_read(&cil->xc_ctx->space_used);
1558
1559 /*
1560 * The cil won't be empty because we are called while holding the
1561 * context lock so whatever we added to the CIL will still be there.
1562 */
1563 ASSERT(!test_bit(XLOG_CIL_EMPTY, &cil->xc_flags));
1564
1565 /*
1566 * We are done if:
1567 * - we haven't used up all the space available yet; or
1568 * - we've already queued up a push; and
1569 * - we're not over the hard limit; and
1570 * - nothing has been over the hard limit.
1571 *
1572 * If so, we don't need to take the push lock as there's nothing to do.
1573 */
1574 if (space_used < XLOG_CIL_SPACE_LIMIT(log) ||
1575 (cil->xc_push_seq == cil->xc_current_sequence &&
1576 space_used < XLOG_CIL_BLOCKING_SPACE_LIMIT(log) &&
1577 !waitqueue_active(&cil->xc_push_wait))) {
1578 up_read(&cil->xc_ctx_lock);
1579 return;
1580 }
1581
1582 spin_lock(&cil->xc_push_lock);
1583 if (cil->xc_push_seq < cil->xc_current_sequence) {
1584 cil->xc_push_seq = cil->xc_current_sequence;
1585 queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work);
1586 }
1587
1588 /*
1589 * Drop the context lock now, we can't hold that if we need to sleep
1590 * because we are over the blocking threshold. The push_lock is still
1591 * held, so blocking threshold sleep/wakeup is still correctly
1592 * serialised here.
1593 */
1594 up_read(&cil->xc_ctx_lock);
1595
1596 /*
1597 * If we are well over the space limit, throttle the work that is being
1598 * done until the push work on this context has begun. Enforce the hard
1599 * throttle on all transaction commits once it has been activated, even
1600 * if the committing transactions have resulted in the space usage
1601 * dipping back down under the hard limit.
1602 *
1603 * The ctx->xc_push_lock provides the serialisation necessary for safely
1604 * calling xlog_cil_over_hard_limit() in this context.
1605 */
1606 if (xlog_cil_over_hard_limit(log, space_used)) {
1607 trace_xfs_log_cil_wait(log, cil->xc_ctx->ticket);
1608 ASSERT(space_used < log->l_logsize);
1609 xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock);
1610 return;
1611 }
1612
1613 spin_unlock(&cil->xc_push_lock);
1614
1615 }
1616
1617 /*
1618 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence
1619 * number that is passed. When it returns, the work will be queued for
1620 * @push_seq, but it won't be completed.
1621 *
1622 * If the caller is performing a synchronous force, we will flush the workqueue
1623 * to get previously queued work moving to minimise the wait time they will
1624 * undergo waiting for all outstanding pushes to complete. The caller is
1625 * expected to do the required waiting for push_seq to complete.
1626 *
1627 * If the caller is performing an async push, we need to ensure that the
1628 * checkpoint is fully flushed out of the iclogs when we finish the push. If we
1629 * don't do this, then the commit record may remain sitting in memory in an
1630 * ACTIVE iclog. This then requires another full log force to push to disk,
1631 * which defeats the purpose of having an async, non-blocking CIL force
1632 * mechanism. Hence in this case we need to pass a flag to the push work to
1633 * indicate it needs to flush the commit record itself.
1634 */
1635 static void
1636 xlog_cil_push_now(
1637 struct xlog *log,
1638 xfs_lsn_t push_seq,
1639 bool async)
1640 {
1641 struct xfs_cil *cil = log->l_cilp;
1642
1643 if (!cil)
1644 return;
1645
1646 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
1647
1648 /* start on any pending background push to minimise wait time on it */
1649 if (!async)
1650 flush_workqueue(cil->xc_push_wq);
1651
1652 spin_lock(&cil->xc_push_lock);
1653
1654 /*
1655 * If this is an async flush request, we always need to set the
1656 * xc_push_commit_stable flag even if something else has already queued
1657 * a push. The flush caller is asking for the CIL to be on stable
1658 * storage when the next push completes, so regardless of who has queued
1659 * the push, the flush requires stable semantics from it.
1660 */
1661 cil->xc_push_commit_stable = async;
1662
1663 /*
1664 * If the CIL is empty or we've already pushed the sequence then
1665 * there's no more work that we need to do.
1666 */
1667 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags) ||
1668 push_seq <= cil->xc_push_seq) {
1669 spin_unlock(&cil->xc_push_lock);
1670 return;
1671 }
1672
1673 cil->xc_push_seq = push_seq;
1674 queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work);
1675 spin_unlock(&cil->xc_push_lock);
1676 }
1677
1678 bool
1679 xlog_cil_empty(
1680 struct xlog *log)
1681 {
1682 struct xfs_cil *cil = log->l_cilp;
1683 bool empty = false;
1684
1685 spin_lock(&cil->xc_push_lock);
1686 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags))
1687 empty = true;
1688 spin_unlock(&cil->xc_push_lock);
1689 return empty;
1690 }
1691
1692 /*
1693 * If there are intent done items in this transaction and the related intent was
1694 * committed in the current (same) CIL checkpoint, we don't need to write either
1695 * the intent or intent done item to the journal as the change will be
1696 * journalled atomically within this checkpoint. As we cannot remove items from
1697 * the CIL here, mark the related intent with a whiteout so that the CIL push
1698 * can remove it rather than writing it to the journal. Then remove the intent
1699 * done item from the current transaction and release it so it doesn't get put
1700 * into the CIL at all.
1701 */
1702 static uint32_t
1703 xlog_cil_process_intents(
1704 struct xfs_cil *cil,
1705 struct xfs_trans *tp)
1706 {
1707 struct xfs_log_item *lip, *ilip, *next;
1708 uint32_t len = 0;
1709
1710 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
1711 if (!(lip->li_ops->flags & XFS_ITEM_INTENT_DONE))
1712 continue;
1713
1714 ilip = lip->li_ops->iop_intent(lip);
1715 if (!ilip || !xlog_item_in_current_chkpt(cil, ilip))
1716 continue;
1717 set_bit(XFS_LI_WHITEOUT, &ilip->li_flags);
1718 trace_xfs_cil_whiteout_mark(ilip);
1719 len += ilip->li_lv->lv_bytes;
1720 kvfree(ilip->li_lv);
1721 ilip->li_lv = NULL;
1722
1723 xfs_trans_del_item(lip);
1724 lip->li_ops->iop_release(lip);
1725 }
1726 return len;
1727 }
1728
1729 /*
1730 * Commit a transaction with the given vector to the Committed Item List.
1731 *
1732 * To do this, we need to format the item, pin it in memory if required and
1733 * account for the space used by the transaction. Once we have done that we
1734 * need to release the unused reservation for the transaction, attach the
1735 * transaction to the checkpoint context so we carry the busy extents through
1736 * to checkpoint completion, and then unlock all the items in the transaction.
1737 *
1738 * Called with the context lock already held in read mode to lock out
1739 * background commit, returns without it held once background commits are
1740 * allowed again.
1741 */
1742 void
1743 xlog_cil_commit(
1744 struct xlog *log,
1745 struct xfs_trans *tp,
1746 xfs_csn_t *commit_seq,
1747 bool regrant)
1748 {
1749 struct xfs_cil *cil = log->l_cilp;
1750 struct xfs_log_item *lip, *next;
1751 uint32_t released_space = 0;
1752
1753 /*
1754 * Do all necessary memory allocation before we lock the CIL.
1755 * This ensures the allocation does not deadlock with a CIL
1756 * push in memory reclaim (e.g. from kswapd).
1757 */
1758 xlog_cil_alloc_shadow_bufs(log, tp);
1759
1760 /* lock out background commit */
1761 down_read(&cil->xc_ctx_lock);
1762
1763 if (tp->t_flags & XFS_TRANS_HAS_INTENT_DONE)
1764 released_space = xlog_cil_process_intents(cil, tp);
1765
1766 xlog_cil_insert_items(log, tp, released_space);
1767
1768 if (regrant && !xlog_is_shutdown(log))
1769 xfs_log_ticket_regrant(log, tp->t_ticket);
1770 else
1771 xfs_log_ticket_ungrant(log, tp->t_ticket);
1772 tp->t_ticket = NULL;
1773 xfs_trans_unreserve_and_mod_sb(tp);
1774
1775 /*
1776 * Once all the items of the transaction have been copied to the CIL,
1777 * the items can be unlocked and possibly freed.
1778 *
1779 * This needs to be done before we drop the CIL context lock because we
1780 * have to update state in the log items and unlock them before they go
1781 * to disk. If we don't, then the CIL checkpoint can race with us and
1782 * we can run checkpoint completion before we've updated and unlocked
1783 * the log items. This affects (at least) processing of stale buffers,
1784 * inodes and EFIs.
1785 */
1786 trace_xfs_trans_commit_items(tp, _RET_IP_);
1787 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
1788 xfs_trans_del_item(lip);
1789 if (lip->li_ops->iop_committing)
1790 lip->li_ops->iop_committing(lip, cil->xc_ctx->sequence);
1791 }
1792 if (commit_seq)
1793 *commit_seq = cil->xc_ctx->sequence;
1794
1795 /* xlog_cil_push_background() releases cil->xc_ctx_lock */
1796 xlog_cil_push_background(log);
1797 }
1798
1799 /*
1800 * Flush the CIL to stable storage but don't wait for it to complete. This
1801 * requires the CIL push to ensure the commit record for the push hits the disk,
1802 * but otherwise is no different to a push done from a log force.
1803 */
1804 void
1805 xlog_cil_flush(
1806 struct xlog *log)
1807 {
1808 xfs_csn_t seq = log->l_cilp->xc_current_sequence;
1809
1810 trace_xfs_log_force(log->l_mp, seq, _RET_IP_);
1811 xlog_cil_push_now(log, seq, true);
1812
1813 /*
1814 * If the CIL is empty, make sure that any previous checkpoint that may
1815 * still be in an active iclog is pushed to stable storage.
1816 */
1817 if (test_bit(XLOG_CIL_EMPTY, &log->l_cilp->xc_flags))
1818 xfs_log_force(log->l_mp, 0);
1819 }
1820
1821 /*
1822 * Conditionally push the CIL based on the sequence passed in.
1823 *
1824 * We only need to push if we haven't already pushed the sequence number given.
1825 * Hence the only time we will trigger a push here is if the push sequence is
1826 * the same as the current context.
1827 *
1828 * We return the current commit lsn to allow the callers to determine if a
1829 * iclog flush is necessary following this call.
1830 */
1831 xfs_lsn_t
1832 xlog_cil_force_seq(
1833 struct xlog *log,
1834 xfs_csn_t sequence)
1835 {
1836 struct xfs_cil *cil = log->l_cilp;
1837 struct xfs_cil_ctx *ctx;
1838 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
1839
1840 ASSERT(sequence <= cil->xc_current_sequence);
1841
1842 if (!sequence)
1843 sequence = cil->xc_current_sequence;
1844 trace_xfs_log_force(log->l_mp, sequence, _RET_IP_);
1845
1846 /*
1847 * check to see if we need to force out the current context.
1848 * xlog_cil_push() handles racing pushes for the same sequence,
1849 * so no need to deal with it here.
1850 */
1851 restart:
1852 xlog_cil_push_now(log, sequence, false);
1853
1854 /*
1855 * See if we can find a previous sequence still committing.
1856 * We need to wait for all previous sequence commits to complete
1857 * before allowing the force of push_seq to go ahead. Hence block
1858 * on commits for those as well.
1859 */
1860 spin_lock(&cil->xc_push_lock);
1861 list_for_each_entry(ctx, &cil->xc_committing, committing) {
1862 /*
1863 * Avoid getting stuck in this loop because we were woken by the
1864 * shutdown, but then went back to sleep once already in the
1865 * shutdown state.
1866 */
1867 if (xlog_is_shutdown(log))
1868 goto out_shutdown;
1869 if (ctx->sequence > sequence)
1870 continue;
1871 if (!ctx->commit_lsn) {
1872 /*
1873 * It is still being pushed! Wait for the push to
1874 * complete, then start again from the beginning.
1875 */
1876 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
1877 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
1878 goto restart;
1879 }
1880 if (ctx->sequence != sequence)
1881 continue;
1882 /* found it! */
1883 commit_lsn = ctx->commit_lsn;
1884 }
1885
1886 /*
1887 * The call to xlog_cil_push_now() executes the push in the background.
1888 * Hence by the time we have got here it our sequence may not have been
1889 * pushed yet. This is true if the current sequence still matches the
1890 * push sequence after the above wait loop and the CIL still contains
1891 * dirty objects. This is guaranteed by the push code first adding the
1892 * context to the committing list before emptying the CIL.
1893 *
1894 * Hence if we don't find the context in the committing list and the
1895 * current sequence number is unchanged then the CIL contents are
1896 * significant. If the CIL is empty, if means there was nothing to push
1897 * and that means there is nothing to wait for. If the CIL is not empty,
1898 * it means we haven't yet started the push, because if it had started
1899 * we would have found the context on the committing list.
1900 */
1901 if (sequence == cil->xc_current_sequence &&
1902 !test_bit(XLOG_CIL_EMPTY, &cil->xc_flags)) {
1903 spin_unlock(&cil->xc_push_lock);
1904 goto restart;
1905 }
1906
1907 spin_unlock(&cil->xc_push_lock);
1908 return commit_lsn;
1909
1910 /*
1911 * We detected a shutdown in progress. We need to trigger the log force
1912 * to pass through it's iclog state machine error handling, even though
1913 * we are already in a shutdown state. Hence we can't return
1914 * NULLCOMMITLSN here as that has special meaning to log forces (i.e.
1915 * LSN is already stable), so we return a zero LSN instead.
1916 */
1917 out_shutdown:
1918 spin_unlock(&cil->xc_push_lock);
1919 return 0;
1920 }
1921
1922 /*
1923 * Perform initial CIL structure initialisation.
1924 */
1925 int
1926 xlog_cil_init(
1927 struct xlog *log)
1928 {
1929 struct xfs_cil *cil;
1930 struct xfs_cil_ctx *ctx;
1931 struct xlog_cil_pcp *cilpcp;
1932 int cpu;
1933
1934 cil = kzalloc(sizeof(*cil), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1935 if (!cil)
1936 return -ENOMEM;
1937 /*
1938 * Limit the CIL pipeline depth to 4 concurrent works to bound the
1939 * concurrency the log spinlocks will be exposed to.
1940 */
1941 cil->xc_push_wq = alloc_workqueue("xfs-cil/%s",
1942 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | WQ_UNBOUND),
1943 4, log->l_mp->m_super->s_id);
1944 if (!cil->xc_push_wq)
1945 goto out_destroy_cil;
1946
1947 cil->xc_log = log;
1948 cil->xc_pcp = alloc_percpu(struct xlog_cil_pcp);
1949 if (!cil->xc_pcp)
1950 goto out_destroy_wq;
1951
1952 for_each_possible_cpu(cpu) {
1953 cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);
1954 INIT_LIST_HEAD(&cilpcp->busy_extents);
1955 INIT_LIST_HEAD(&cilpcp->log_items);
1956 }
1957
1958 INIT_LIST_HEAD(&cil->xc_committing);
1959 spin_lock_init(&cil->xc_push_lock);
1960 init_waitqueue_head(&cil->xc_push_wait);
1961 init_rwsem(&cil->xc_ctx_lock);
1962 init_waitqueue_head(&cil->xc_start_wait);
1963 init_waitqueue_head(&cil->xc_commit_wait);
1964 log->l_cilp = cil;
1965
1966 ctx = xlog_cil_ctx_alloc();
1967 xlog_cil_ctx_switch(cil, ctx);
1968 return 0;
1969
1970 out_destroy_wq:
1971 destroy_workqueue(cil->xc_push_wq);
1972 out_destroy_cil:
1973 kfree(cil);
1974 return -ENOMEM;
1975 }
1976
1977 void
1978 xlog_cil_destroy(
1979 struct xlog *log)
1980 {
1981 struct xfs_cil *cil = log->l_cilp;
1982
1983 if (cil->xc_ctx) {
1984 if (cil->xc_ctx->ticket)
1985 xfs_log_ticket_put(cil->xc_ctx->ticket);
1986 kfree(cil->xc_ctx);
1987 }
1988
1989 ASSERT(test_bit(XLOG_CIL_EMPTY, &cil->xc_flags));
1990 free_percpu(cil->xc_pcp);
1991 destroy_workqueue(cil->xc_push_wq);
1992 kfree(cil);
1993 }
1994
1995
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.