Commit | Line | Data |
---|---|---|
0b61f8a4 | 1 | // SPDX-License-Identifier: GPL-2.0 |
71e330b5 DC |
2 | /* |
3 | * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved. | |
71e330b5 DC |
4 | */ |
5 | ||
6 | #include "xfs.h" | |
7 | #include "xfs_fs.h" | |
4fb6e8ad | 8 | #include "xfs_format.h" |
239880ef | 9 | #include "xfs_log_format.h" |
70a9883c | 10 | #include "xfs_shared.h" |
239880ef | 11 | #include "xfs_trans_resv.h" |
71e330b5 | 12 | #include "xfs_mount.h" |
efc27b52 | 13 | #include "xfs_extent_busy.h" |
239880ef DC |
14 | #include "xfs_trans.h" |
15 | #include "xfs_trans_priv.h" | |
16 | #include "xfs_log.h" | |
17 | #include "xfs_log_priv.h" | |
4560e78f CH |
18 | #include "xfs_trace.h" |
19 | ||
20 | struct workqueue_struct *xfs_discard_wq; | |
71e330b5 | 21 | |
71e330b5 DC |
22 | /* |
23 | * Allocate a new ticket. Failing to get a new ticket makes it really hard to | |
24 | * recover, so we don't allow failure here. Also, we allocate in a context that | |
25 | * we don't want to be issuing transactions from, so we need to tell the | |
26 | * allocation code this as well. | |
27 | * | |
28 | * We don't reserve any space for the ticket - we are going to steal whatever | |
29 | * space we require from transactions as they commit. To ensure we reserve all | |
30 | * the space required, we need to set the current reservation of the ticket to | |
31 | * zero so that we know to steal the initial transaction overhead from the | |
32 | * first transaction commit. | |
33 | */ | |
34 | static struct xlog_ticket * | |
35 | xlog_cil_ticket_alloc( | |
f7bdf03a | 36 | struct xlog *log) |
71e330b5 DC |
37 | { |
38 | struct xlog_ticket *tic; | |
39 | ||
ca4f2589 | 40 | tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0); |
71e330b5 DC |
41 | |
42 | /* | |
43 | * set the current reservation to zero so we know to steal the basic | |
44 | * transaction overhead reservation from the first transaction commit. | |
45 | */ | |
46 | tic->t_curr_res = 0; | |
47 | return tic; | |
48 | } | |
49 | ||
39823d0f DC |
50 | /* |
51 | * Unavoidable forward declaration - xlog_cil_push_work() calls | |
52 | * xlog_cil_ctx_alloc() itself. | |
53 | */ | |
54 | static void xlog_cil_push_work(struct work_struct *work); | |
55 | ||
56 | static struct xfs_cil_ctx * | |
57 | xlog_cil_ctx_alloc(void) | |
58 | { | |
59 | struct xfs_cil_ctx *ctx; | |
60 | ||
61 | ctx = kmem_zalloc(sizeof(*ctx), KM_NOFS); | |
62 | INIT_LIST_HEAD(&ctx->committing); | |
63 | INIT_LIST_HEAD(&ctx->busy_extents); | |
64 | INIT_WORK(&ctx->push_work, xlog_cil_push_work); | |
65 | return ctx; | |
66 | } | |
67 | ||
68 | static void | |
69 | xlog_cil_ctx_switch( | |
70 | struct xfs_cil *cil, | |
71 | struct xfs_cil_ctx *ctx) | |
72 | { | |
73 | ctx->sequence = ++cil->xc_current_sequence; | |
74 | ctx->cil = cil; | |
75 | cil->xc_ctx = ctx; | |
76 | } | |
77 | ||
71e330b5 DC |
78 | /* |
79 | * After the first stage of log recovery is done, we know where the head and | |
80 | * tail of the log are. We need this log initialisation done before we can | |
81 | * initialise the first CIL checkpoint context. | |
82 | * | |
83 | * Here we allocate a log ticket to track space usage during a CIL push. This | |
84 | * ticket is passed to xlog_write() directly so that we don't slowly leak log | |
85 | * space by failing to account for space used by log headers and additional | |
86 | * region headers for split regions. | |
87 | */ | |
88 | void | |
89 | xlog_cil_init_post_recovery( | |
f7bdf03a | 90 | struct xlog *log) |
71e330b5 | 91 | { |
71e330b5 DC |
92 | log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log); |
93 | log->l_cilp->xc_ctx->sequence = 1; | |
71e330b5 DC |
94 | } |
95 | ||
b1c5ebb2 DC |
96 | static inline int |
97 | xlog_cil_iovec_space( | |
98 | uint niovecs) | |
99 | { | |
100 | return round_up((sizeof(struct xfs_log_vec) + | |
101 | niovecs * sizeof(struct xfs_log_iovec)), | |
102 | sizeof(uint64_t)); | |
103 | } | |
104 | ||
8dc9384b DC |
105 | /* |
106 | * shadow buffers can be large, so we need to use kvmalloc() here to ensure | |
107 | * success. Unfortunately, kvmalloc() only allows GFP_KERNEL contexts to fall | |
108 | * back to vmalloc, so we can't actually do anything useful with gfp flags to | |
109 | * control the kmalloc() behaviour within kvmalloc(). Hence kmalloc() will do | |
110 | * direct reclaim and compaction in the slow path, both of which are | |
111 | * horrendously expensive. We just want kmalloc to fail fast and fall back to | |
112 | * vmalloc if it can't get somethign straight away from the free lists or buddy | |
113 | * allocator. Hence we have to open code kvmalloc outselves here. | |
114 | * | |
115 | * Also, we are in memalloc_nofs_save task context here, so despite the use of | |
116 | * GFP_KERNEL here, we are actually going to be doing GFP_NOFS allocations. This | |
117 | * is actually the only way to make vmalloc() do GFP_NOFS allocations, so lets | |
118 | * just all pretend this is a GFP_KERNEL context operation.... | |
119 | */ | |
120 | static inline void * | |
121 | xlog_cil_kvmalloc( | |
122 | size_t buf_size) | |
123 | { | |
124 | gfp_t flags = GFP_KERNEL; | |
125 | void *p; | |
126 | ||
127 | flags &= ~__GFP_DIRECT_RECLAIM; | |
128 | flags |= __GFP_NOWARN | __GFP_NORETRY; | |
129 | do { | |
130 | p = kmalloc(buf_size, flags); | |
131 | if (!p) | |
132 | p = vmalloc(buf_size); | |
133 | } while (!p); | |
134 | ||
135 | return p; | |
136 | } | |
137 | ||
b1c5ebb2 DC |
138 | /* |
139 | * Allocate or pin log vector buffers for CIL insertion. | |
140 | * | |
141 | * The CIL currently uses disposable buffers for copying a snapshot of the | |
142 | * modified items into the log during a push. The biggest problem with this is | |
143 | * the requirement to allocate the disposable buffer during the commit if: | |
144 | * a) does not exist; or | |
145 | * b) it is too small | |
146 | * | |
147 | * If we do this allocation within xlog_cil_insert_format_items(), it is done | |
148 | * under the xc_ctx_lock, which means that a CIL push cannot occur during | |
149 | * the memory allocation. This means that we have a potential deadlock situation | |
150 | * under low memory conditions when we have lots of dirty metadata pinned in | |
151 | * the CIL and we need a CIL commit to occur to free memory. | |
152 | * | |
153 | * To avoid this, we need to move the memory allocation outside the | |
154 | * xc_ctx_lock, but because the log vector buffers are disposable, that opens | |
155 | * up a TOCTOU race condition w.r.t. the CIL committing and removing the log | |
156 | * vector buffers between the check and the formatting of the item into the | |
157 | * log vector buffer within the xc_ctx_lock. | |
158 | * | |
159 | * Because the log vector buffer needs to be unchanged during the CIL push | |
160 | * process, we cannot share the buffer between the transaction commit (which | |
161 | * modifies the buffer) and the CIL push context that is writing the changes | |
162 | * into the log. This means skipping preallocation of buffer space is | |
163 | * unreliable, but we most definitely do not want to be allocating and freeing | |
164 | * buffers unnecessarily during commits when overwrites can be done safely. | |
165 | * | |
166 | * The simplest solution to this problem is to allocate a shadow buffer when a | |
167 | * log item is committed for the second time, and then to only use this buffer | |
168 | * if necessary. The buffer can remain attached to the log item until such time | |
169 | * it is needed, and this is the buffer that is reallocated to match the size of | |
170 | * the incoming modification. Then during the formatting of the item we can swap | |
171 | * the active buffer with the new one if we can't reuse the existing buffer. We | |
172 | * don't free the old buffer as it may be reused on the next modification if | |
173 | * it's size is right, otherwise we'll free and reallocate it at that point. | |
174 | * | |
175 | * This function builds a vector for the changes in each log item in the | |
176 | * transaction. It then works out the length of the buffer needed for each log | |
177 | * item, allocates them and attaches the vector to the log item in preparation | |
178 | * for the formatting step which occurs under the xc_ctx_lock. | |
179 | * | |
180 | * While this means the memory footprint goes up, it avoids the repeated | |
181 | * alloc/free pattern that repeated modifications of an item would otherwise | |
182 | * cause, and hence minimises the CPU overhead of such behaviour. | |
183 | */ | |
184 | static void | |
185 | xlog_cil_alloc_shadow_bufs( | |
186 | struct xlog *log, | |
187 | struct xfs_trans *tp) | |
188 | { | |
e6631f85 | 189 | struct xfs_log_item *lip; |
b1c5ebb2 | 190 | |
e6631f85 | 191 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
b1c5ebb2 DC |
192 | struct xfs_log_vec *lv; |
193 | int niovecs = 0; | |
194 | int nbytes = 0; | |
195 | int buf_size; | |
196 | bool ordered = false; | |
197 | ||
198 | /* Skip items which aren't dirty in this transaction. */ | |
e6631f85 | 199 | if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) |
b1c5ebb2 DC |
200 | continue; |
201 | ||
202 | /* get number of vecs and size of data to be stored */ | |
203 | lip->li_ops->iop_size(lip, &niovecs, &nbytes); | |
204 | ||
205 | /* | |
206 | * Ordered items need to be tracked but we do not wish to write | |
207 | * them. We need a logvec to track the object, but we do not | |
208 | * need an iovec or buffer to be allocated for copying data. | |
209 | */ | |
210 | if (niovecs == XFS_LOG_VEC_ORDERED) { | |
211 | ordered = true; | |
212 | niovecs = 0; | |
213 | nbytes = 0; | |
214 | } | |
215 | ||
216 | /* | |
217 | * We 64-bit align the length of each iovec so that the start | |
218 | * of the next one is naturally aligned. We'll need to | |
219 | * account for that slack space here. Then round nbytes up | |
220 | * to 64-bit alignment so that the initial buffer alignment is | |
221 | * easy to calculate and verify. | |
222 | */ | |
223 | nbytes += niovecs * sizeof(uint64_t); | |
224 | nbytes = round_up(nbytes, sizeof(uint64_t)); | |
225 | ||
226 | /* | |
227 | * The data buffer needs to start 64-bit aligned, so round up | |
228 | * that space to ensure we can align it appropriately and not | |
229 | * overrun the buffer. | |
230 | */ | |
231 | buf_size = nbytes + xlog_cil_iovec_space(niovecs); | |
232 | ||
233 | /* | |
234 | * if we have no shadow buffer, or it is too small, we need to | |
235 | * reallocate it. | |
236 | */ | |
237 | if (!lip->li_lv_shadow || | |
238 | buf_size > lip->li_lv_shadow->lv_size) { | |
b1c5ebb2 DC |
239 | /* |
240 | * We free and allocate here as a realloc would copy | |
8dc9384b | 241 | * unnecessary data. We don't use kvzalloc() for the |
b1c5ebb2 DC |
242 | * same reason - we don't need to zero the data area in |
243 | * the buffer, only the log vector header and the iovec | |
244 | * storage. | |
245 | */ | |
246 | kmem_free(lip->li_lv_shadow); | |
8dc9384b | 247 | lv = xlog_cil_kvmalloc(buf_size); |
b1c5ebb2 | 248 | |
b1c5ebb2 DC |
249 | memset(lv, 0, xlog_cil_iovec_space(niovecs)); |
250 | ||
251 | lv->lv_item = lip; | |
252 | lv->lv_size = buf_size; | |
253 | if (ordered) | |
254 | lv->lv_buf_len = XFS_LOG_VEC_ORDERED; | |
255 | else | |
256 | lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1]; | |
257 | lip->li_lv_shadow = lv; | |
258 | } else { | |
259 | /* same or smaller, optimise common overwrite case */ | |
260 | lv = lip->li_lv_shadow; | |
261 | if (ordered) | |
262 | lv->lv_buf_len = XFS_LOG_VEC_ORDERED; | |
263 | else | |
264 | lv->lv_buf_len = 0; | |
265 | lv->lv_bytes = 0; | |
266 | lv->lv_next = NULL; | |
267 | } | |
268 | ||
269 | /* Ensure the lv is set up according to ->iop_size */ | |
270 | lv->lv_niovecs = niovecs; | |
271 | ||
272 | /* The allocated data region lies beyond the iovec region */ | |
273 | lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs); | |
274 | } | |
275 | ||
276 | } | |
277 | ||
991aaf65 DC |
278 | /* |
279 | * Prepare the log item for insertion into the CIL. Calculate the difference in | |
280 | * log space and vectors it will consume, and if it is a new item pin it as | |
281 | * well. | |
282 | */ | |
283 | STATIC void | |
284 | xfs_cil_prepare_item( | |
285 | struct xlog *log, | |
286 | struct xfs_log_vec *lv, | |
287 | struct xfs_log_vec *old_lv, | |
288 | int *diff_len, | |
289 | int *diff_iovecs) | |
290 | { | |
291 | /* Account for the new LV being passed in */ | |
292 | if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) { | |
110dc24a | 293 | *diff_len += lv->lv_bytes; |
991aaf65 DC |
294 | *diff_iovecs += lv->lv_niovecs; |
295 | } | |
296 | ||
297 | /* | |
298 | * If there is no old LV, this is the first time we've seen the item in | |
299 | * this CIL context and so we need to pin it. If we are replacing the | |
b1c5ebb2 DC |
300 | * old_lv, then remove the space it accounts for and make it the shadow |
301 | * buffer for later freeing. In both cases we are now switching to the | |
b63da6c8 | 302 | * shadow buffer, so update the pointer to it appropriately. |
991aaf65 | 303 | */ |
b1c5ebb2 | 304 | if (!old_lv) { |
e8b78db7 CH |
305 | if (lv->lv_item->li_ops->iop_pin) |
306 | lv->lv_item->li_ops->iop_pin(lv->lv_item); | |
b1c5ebb2 DC |
307 | lv->lv_item->li_lv_shadow = NULL; |
308 | } else if (old_lv != lv) { | |
991aaf65 DC |
309 | ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED); |
310 | ||
110dc24a | 311 | *diff_len -= old_lv->lv_bytes; |
991aaf65 | 312 | *diff_iovecs -= old_lv->lv_niovecs; |
b1c5ebb2 | 313 | lv->lv_item->li_lv_shadow = old_lv; |
991aaf65 DC |
314 | } |
315 | ||
316 | /* attach new log vector to log item */ | |
317 | lv->lv_item->li_lv = lv; | |
318 | ||
319 | /* | |
320 | * If this is the first time the item is being committed to the | |
321 | * CIL, store the sequence number on the log item so we can | |
322 | * tell in future commits whether this is the first checkpoint | |
323 | * the item is being committed into. | |
324 | */ | |
325 | if (!lv->lv_item->li_seq) | |
326 | lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence; | |
327 | } | |
328 | ||
71e330b5 DC |
329 | /* |
330 | * Format log item into a flat buffers | |
331 | * | |
332 | * For delayed logging, we need to hold a formatted buffer containing all the | |
333 | * changes on the log item. This enables us to relog the item in memory and | |
334 | * write it out asynchronously without needing to relock the object that was | |
335 | * modified at the time it gets written into the iclog. | |
336 | * | |
b1c5ebb2 DC |
337 | * This function takes the prepared log vectors attached to each log item, and |
338 | * formats the changes into the log vector buffer. The buffer it uses is | |
339 | * dependent on the current state of the vector in the CIL - the shadow lv is | |
340 | * guaranteed to be large enough for the current modification, but we will only | |
341 | * use that if we can't reuse the existing lv. If we can't reuse the existing | |
342 | * lv, then simple swap it out for the shadow lv. We don't free it - that is | |
343 | * done lazily either by th enext modification or the freeing of the log item. | |
71e330b5 DC |
344 | * |
345 | * We don't set up region headers during this process; we simply copy the | |
346 | * regions into the flat buffer. We can do this because we still have to do a | |
347 | * formatting step to write the regions into the iclog buffer. Writing the | |
348 | * ophdrs during the iclog write means that we can support splitting large | |
349 | * regions across iclog boundares without needing a change in the format of the | |
350 | * item/region encapsulation. | |
351 | * | |
352 | * Hence what we need to do now is change the rewrite the vector array to point | |
353 | * to the copied region inside the buffer we just allocated. This allows us to | |
354 | * format the regions into the iclog as though they are being formatted | |
355 | * directly out of the objects themselves. | |
356 | */ | |
991aaf65 DC |
357 | static void |
358 | xlog_cil_insert_format_items( | |
359 | struct xlog *log, | |
360 | struct xfs_trans *tp, | |
361 | int *diff_len, | |
362 | int *diff_iovecs) | |
71e330b5 | 363 | { |
e6631f85 | 364 | struct xfs_log_item *lip; |
71e330b5 | 365 | |
0244b960 CH |
366 | |
367 | /* Bail out if we didn't find a log item. */ | |
368 | if (list_empty(&tp->t_items)) { | |
369 | ASSERT(0); | |
991aaf65 | 370 | return; |
0244b960 CH |
371 | } |
372 | ||
e6631f85 | 373 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
7492c5b4 | 374 | struct xfs_log_vec *lv; |
b1c5ebb2 DC |
375 | struct xfs_log_vec *old_lv = NULL; |
376 | struct xfs_log_vec *shadow; | |
fd63875c | 377 | bool ordered = false; |
71e330b5 | 378 | |
0244b960 | 379 | /* Skip items which aren't dirty in this transaction. */ |
e6631f85 | 380 | if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) |
0244b960 CH |
381 | continue; |
382 | ||
fd63875c | 383 | /* |
b1c5ebb2 DC |
384 | * The formatting size information is already attached to |
385 | * the shadow lv on the log item. | |
fd63875c | 386 | */ |
b1c5ebb2 DC |
387 | shadow = lip->li_lv_shadow; |
388 | if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED) | |
fd63875c | 389 | ordered = true; |
fd63875c | 390 | |
b1c5ebb2 DC |
391 | /* Skip items that do not have any vectors for writing */ |
392 | if (!shadow->lv_niovecs && !ordered) | |
393 | continue; | |
0244b960 | 394 | |
f5baac35 | 395 | /* compare to existing item size */ |
b1c5ebb2 DC |
396 | old_lv = lip->li_lv; |
397 | if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) { | |
f5baac35 DC |
398 | /* same or smaller, optimise common overwrite case */ |
399 | lv = lip->li_lv; | |
400 | lv->lv_next = NULL; | |
401 | ||
402 | if (ordered) | |
403 | goto insert; | |
404 | ||
991aaf65 DC |
405 | /* |
406 | * set the item up as though it is a new insertion so | |
407 | * that the space reservation accounting is correct. | |
408 | */ | |
409 | *diff_iovecs -= lv->lv_niovecs; | |
110dc24a | 410 | *diff_len -= lv->lv_bytes; |
b1c5ebb2 DC |
411 | |
412 | /* Ensure the lv is set up according to ->iop_size */ | |
413 | lv->lv_niovecs = shadow->lv_niovecs; | |
414 | ||
415 | /* reset the lv buffer information for new formatting */ | |
416 | lv->lv_buf_len = 0; | |
417 | lv->lv_bytes = 0; | |
418 | lv->lv_buf = (char *)lv + | |
419 | xlog_cil_iovec_space(lv->lv_niovecs); | |
9597df6b | 420 | } else { |
b1c5ebb2 DC |
421 | /* switch to shadow buffer! */ |
422 | lv = shadow; | |
9597df6b | 423 | lv->lv_item = lip; |
9597df6b CH |
424 | if (ordered) { |
425 | /* track as an ordered logvec */ | |
426 | ASSERT(lip->li_lv == NULL); | |
9597df6b CH |
427 | goto insert; |
428 | } | |
f5baac35 DC |
429 | } |
430 | ||
3895e51f | 431 | ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t))); |
bde7cff6 | 432 | lip->li_ops->iop_format(lip, lv); |
7492c5b4 | 433 | insert: |
991aaf65 | 434 | xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs); |
3b93c7aa | 435 | } |
d1583a38 DC |
436 | } |
437 | ||
438 | /* | |
439 | * Insert the log items into the CIL and calculate the difference in space | |
440 | * consumed by the item. Add the space to the checkpoint ticket and calculate | |
441 | * if the change requires additional log metadata. If it does, take that space | |
42b2aa86 | 442 | * as well. Remove the amount of space we added to the checkpoint ticket from |
d1583a38 DC |
443 | * the current transaction ticket so that the accounting works out correctly. |
444 | */ | |
3b93c7aa DC |
445 | static void |
446 | xlog_cil_insert_items( | |
f7bdf03a | 447 | struct xlog *log, |
991aaf65 | 448 | struct xfs_trans *tp) |
3b93c7aa | 449 | { |
d1583a38 DC |
450 | struct xfs_cil *cil = log->l_cilp; |
451 | struct xfs_cil_ctx *ctx = cil->xc_ctx; | |
e6631f85 | 452 | struct xfs_log_item *lip; |
d1583a38 DC |
453 | int len = 0; |
454 | int diff_iovecs = 0; | |
455 | int iclog_space; | |
e2f23426 | 456 | int iovhdr_res = 0, split_res = 0, ctx_res = 0; |
3b93c7aa | 457 | |
991aaf65 | 458 | ASSERT(tp); |
d1583a38 DC |
459 | |
460 | /* | |
d1583a38 DC |
461 | * We can do this safely because the context can't checkpoint until we |
462 | * are done so it doesn't matter exactly how we update the CIL. | |
463 | */ | |
991aaf65 DC |
464 | xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs); |
465 | ||
d1583a38 | 466 | spin_lock(&cil->xc_cil_lock); |
d1583a38 | 467 | |
fd63875c | 468 | /* account for space used by new iovec headers */ |
e2f23426 BF |
469 | iovhdr_res = diff_iovecs * sizeof(xlog_op_header_t); |
470 | len += iovhdr_res; | |
d1583a38 DC |
471 | ctx->nvecs += diff_iovecs; |
472 | ||
991aaf65 DC |
473 | /* attach the transaction to the CIL if it has any busy extents */ |
474 | if (!list_empty(&tp->t_busy)) | |
475 | list_splice_init(&tp->t_busy, &ctx->busy_extents); | |
476 | ||
d1583a38 DC |
477 | /* |
478 | * Now transfer enough transaction reservation to the context ticket | |
479 | * for the checkpoint. The context ticket is special - the unit | |
480 | * reservation has to grow as well as the current reservation as we | |
481 | * steal from tickets so we can correctly determine the space used | |
482 | * during the transaction commit. | |
483 | */ | |
484 | if (ctx->ticket->t_curr_res == 0) { | |
e2f23426 BF |
485 | ctx_res = ctx->ticket->t_unit_res; |
486 | ctx->ticket->t_curr_res = ctx_res; | |
487 | tp->t_ticket->t_curr_res -= ctx_res; | |
d1583a38 DC |
488 | } |
489 | ||
490 | /* do we need space for more log record headers? */ | |
491 | iclog_space = log->l_iclog_size - log->l_iclog_hsize; | |
492 | if (len > 0 && (ctx->space_used / iclog_space != | |
493 | (ctx->space_used + len) / iclog_space)) { | |
e2f23426 | 494 | split_res = (len + iclog_space - 1) / iclog_space; |
d1583a38 | 495 | /* need to take into account split region headers, too */ |
e2f23426 BF |
496 | split_res *= log->l_iclog_hsize + sizeof(struct xlog_op_header); |
497 | ctx->ticket->t_unit_res += split_res; | |
498 | ctx->ticket->t_curr_res += split_res; | |
499 | tp->t_ticket->t_curr_res -= split_res; | |
991aaf65 | 500 | ASSERT(tp->t_ticket->t_curr_res >= len); |
d1583a38 | 501 | } |
991aaf65 | 502 | tp->t_ticket->t_curr_res -= len; |
d1583a38 DC |
503 | ctx->space_used += len; |
504 | ||
d4ca1d55 BF |
505 | /* |
506 | * If we've overrun the reservation, dump the tx details before we move | |
507 | * the log items. Shutdown is imminent... | |
508 | */ | |
509 | if (WARN_ON(tp->t_ticket->t_curr_res < 0)) { | |
510 | xfs_warn(log->l_mp, "Transaction log reservation overrun:"); | |
511 | xfs_warn(log->l_mp, | |
512 | " log items: %d bytes (iov hdrs: %d bytes)", | |
513 | len, iovhdr_res); | |
514 | xfs_warn(log->l_mp, " split region headers: %d bytes", | |
515 | split_res); | |
516 | xfs_warn(log->l_mp, " ctx ticket: %d bytes", ctx_res); | |
517 | xlog_print_trans(tp); | |
518 | } | |
519 | ||
e2f23426 BF |
520 | /* |
521 | * Now (re-)position everything modified at the tail of the CIL. | |
522 | * We do this here so we only need to take the CIL lock once during | |
523 | * the transaction commit. | |
524 | */ | |
e6631f85 | 525 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
e2f23426 BF |
526 | |
527 | /* Skip items which aren't dirty in this transaction. */ | |
e6631f85 | 528 | if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) |
e2f23426 BF |
529 | continue; |
530 | ||
531 | /* | |
532 | * Only move the item if it isn't already at the tail. This is | |
533 | * to prevent a transient list_empty() state when reinserting | |
534 | * an item that is already the only item in the CIL. | |
535 | */ | |
536 | if (!list_is_last(&lip->li_cil, &cil->xc_cil)) | |
537 | list_move_tail(&lip->li_cil, &cil->xc_cil); | |
538 | } | |
539 | ||
d1583a38 | 540 | spin_unlock(&cil->xc_cil_lock); |
d4ca1d55 BF |
541 | |
542 | if (tp->t_ticket->t_curr_res < 0) | |
b5f17bec | 543 | xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); |
71e330b5 DC |
544 | } |
545 | ||
546 | static void | |
547 | xlog_cil_free_logvec( | |
548 | struct xfs_log_vec *log_vector) | |
549 | { | |
550 | struct xfs_log_vec *lv; | |
551 | ||
552 | for (lv = log_vector; lv; ) { | |
553 | struct xfs_log_vec *next = lv->lv_next; | |
71e330b5 DC |
554 | kmem_free(lv); |
555 | lv = next; | |
556 | } | |
557 | } | |
558 | ||
4560e78f CH |
559 | static void |
560 | xlog_discard_endio_work( | |
561 | struct work_struct *work) | |
562 | { | |
563 | struct xfs_cil_ctx *ctx = | |
564 | container_of(work, struct xfs_cil_ctx, discard_endio_work); | |
565 | struct xfs_mount *mp = ctx->cil->xc_log->l_mp; | |
566 | ||
567 | xfs_extent_busy_clear(mp, &ctx->busy_extents, false); | |
568 | kmem_free(ctx); | |
569 | } | |
570 | ||
571 | /* | |
572 | * Queue up the actual completion to a thread to avoid IRQ-safe locking for | |
573 | * pagb_lock. Note that we need a unbounded workqueue, otherwise we might | |
574 | * get the execution delayed up to 30 seconds for weird reasons. | |
575 | */ | |
576 | static void | |
577 | xlog_discard_endio( | |
578 | struct bio *bio) | |
579 | { | |
580 | struct xfs_cil_ctx *ctx = bio->bi_private; | |
581 | ||
582 | INIT_WORK(&ctx->discard_endio_work, xlog_discard_endio_work); | |
583 | queue_work(xfs_discard_wq, &ctx->discard_endio_work); | |
ea7bd56f | 584 | bio_put(bio); |
4560e78f CH |
585 | } |
586 | ||
587 | static void | |
588 | xlog_discard_busy_extents( | |
589 | struct xfs_mount *mp, | |
590 | struct xfs_cil_ctx *ctx) | |
591 | { | |
592 | struct list_head *list = &ctx->busy_extents; | |
593 | struct xfs_extent_busy *busyp; | |
594 | struct bio *bio = NULL; | |
595 | struct blk_plug plug; | |
596 | int error = 0; | |
597 | ||
0560f31a | 598 | ASSERT(xfs_has_discard(mp)); |
4560e78f CH |
599 | |
600 | blk_start_plug(&plug); | |
601 | list_for_each_entry(busyp, list, list) { | |
602 | trace_xfs_discard_extent(mp, busyp->agno, busyp->bno, | |
603 | busyp->length); | |
604 | ||
605 | error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev, | |
606 | XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno), | |
607 | XFS_FSB_TO_BB(mp, busyp->length), | |
608 | GFP_NOFS, 0, &bio); | |
609 | if (error && error != -EOPNOTSUPP) { | |
610 | xfs_info(mp, | |
611 | "discard failed for extent [0x%llx,%u], error %d", | |
612 | (unsigned long long)busyp->bno, | |
613 | busyp->length, | |
614 | error); | |
615 | break; | |
616 | } | |
617 | } | |
618 | ||
619 | if (bio) { | |
620 | bio->bi_private = ctx; | |
621 | bio->bi_end_io = xlog_discard_endio; | |
622 | submit_bio(bio); | |
623 | } else { | |
624 | xlog_discard_endio_work(&ctx->discard_endio_work); | |
625 | } | |
626 | blk_finish_plug(&plug); | |
627 | } | |
628 | ||
71e330b5 DC |
629 | /* |
630 | * Mark all items committed and clear busy extents. We free the log vector | |
631 | * chains in a separate pass so that we unpin the log items as quickly as | |
632 | * possible. | |
633 | */ | |
634 | static void | |
635 | xlog_cil_committed( | |
12e6a0f4 | 636 | struct xfs_cil_ctx *ctx) |
71e330b5 | 637 | { |
e84661aa | 638 | struct xfs_mount *mp = ctx->cil->xc_log->l_mp; |
2039a272 | 639 | bool abort = xlog_is_shutdown(ctx->cil->xc_log); |
71e330b5 | 640 | |
545aa41f BF |
641 | /* |
642 | * If the I/O failed, we're aborting the commit and already shutdown. | |
643 | * Wake any commit waiters before aborting the log items so we don't | |
644 | * block async log pushers on callbacks. Async log pushers explicitly do | |
645 | * not wait on log force completion because they may be holding locks | |
646 | * required to unpin items. | |
647 | */ | |
648 | if (abort) { | |
649 | spin_lock(&ctx->cil->xc_push_lock); | |
68a74dca | 650 | wake_up_all(&ctx->cil->xc_start_wait); |
545aa41f BF |
651 | wake_up_all(&ctx->cil->xc_commit_wait); |
652 | spin_unlock(&ctx->cil->xc_push_lock); | |
653 | } | |
654 | ||
0e57f6a3 DC |
655 | xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain, |
656 | ctx->start_lsn, abort); | |
71e330b5 | 657 | |
4ecbfe63 DC |
658 | xfs_extent_busy_sort(&ctx->busy_extents); |
659 | xfs_extent_busy_clear(mp, &ctx->busy_extents, | |
0560f31a | 660 | xfs_has_discard(mp) && !abort); |
71e330b5 | 661 | |
4bb928cd | 662 | spin_lock(&ctx->cil->xc_push_lock); |
71e330b5 | 663 | list_del(&ctx->committing); |
4bb928cd | 664 | spin_unlock(&ctx->cil->xc_push_lock); |
71e330b5 DC |
665 | |
666 | xlog_cil_free_logvec(ctx->lv_chain); | |
e84661aa | 667 | |
4560e78f CH |
668 | if (!list_empty(&ctx->busy_extents)) |
669 | xlog_discard_busy_extents(mp, ctx); | |
670 | else | |
671 | kmem_free(ctx); | |
71e330b5 DC |
672 | } |
673 | ||
89ae379d CH |
674 | void |
675 | xlog_cil_process_committed( | |
12e6a0f4 | 676 | struct list_head *list) |
89ae379d CH |
677 | { |
678 | struct xfs_cil_ctx *ctx; | |
679 | ||
680 | while ((ctx = list_first_entry_or_null(list, | |
681 | struct xfs_cil_ctx, iclog_entry))) { | |
682 | list_del(&ctx->iclog_entry); | |
12e6a0f4 | 683 | xlog_cil_committed(ctx); |
89ae379d CH |
684 | } |
685 | } | |
686 | ||
c45aba40 DC |
687 | /* |
688 | * Record the LSN of the iclog we were just granted space to start writing into. | |
689 | * If the context doesn't have a start_lsn recorded, then this iclog will | |
690 | * contain the start record for the checkpoint. Otherwise this write contains | |
691 | * the commit record for the checkpoint. | |
692 | */ | |
693 | void | |
694 | xlog_cil_set_ctx_write_state( | |
695 | struct xfs_cil_ctx *ctx, | |
696 | struct xlog_in_core *iclog) | |
697 | { | |
698 | struct xfs_cil *cil = ctx->cil; | |
699 | xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn); | |
700 | ||
701 | ASSERT(!ctx->commit_lsn); | |
caa80090 DC |
702 | if (!ctx->start_lsn) { |
703 | spin_lock(&cil->xc_push_lock); | |
68a74dca DC |
704 | /* |
705 | * The LSN we need to pass to the log items on transaction | |
706 | * commit is the LSN reported by the first log vector write, not | |
707 | * the commit lsn. If we use the commit record lsn then we can | |
919edbad DC |
708 | * move the grant write head beyond the tail LSN and overwrite |
709 | * it. | |
68a74dca | 710 | */ |
c45aba40 | 711 | ctx->start_lsn = lsn; |
68a74dca | 712 | wake_up_all(&cil->xc_start_wait); |
caa80090 | 713 | spin_unlock(&cil->xc_push_lock); |
919edbad DC |
714 | |
715 | /* | |
716 | * Make sure the metadata we are about to overwrite in the log | |
717 | * has been flushed to stable storage before this iclog is | |
718 | * issued. | |
719 | */ | |
720 | spin_lock(&cil->xc_log->l_icloglock); | |
721 | iclog->ic_flags |= XLOG_ICL_NEED_FLUSH; | |
722 | spin_unlock(&cil->xc_log->l_icloglock); | |
caa80090 DC |
723 | return; |
724 | } | |
725 | ||
726 | /* | |
727 | * Take a reference to the iclog for the context so that we still hold | |
728 | * it when xlog_write is done and has released it. This means the | |
729 | * context controls when the iclog is released for IO. | |
730 | */ | |
731 | atomic_inc(&iclog->ic_refcnt); | |
732 | ||
733 | /* | |
734 | * xlog_state_get_iclog_space() guarantees there is enough space in the | |
735 | * iclog for an entire commit record, so we can attach the context | |
736 | * callbacks now. This needs to be done before we make the commit_lsn | |
737 | * visible to waiters so that checkpoints with commit records in the | |
738 | * same iclog order their IO completion callbacks in the same order that | |
739 | * the commit records appear in the iclog. | |
740 | */ | |
741 | spin_lock(&cil->xc_log->l_icloglock); | |
742 | list_add_tail(&ctx->iclog_entry, &iclog->ic_callbacks); | |
743 | spin_unlock(&cil->xc_log->l_icloglock); | |
744 | ||
745 | /* | |
746 | * Now we can record the commit LSN and wake anyone waiting for this | |
747 | * sequence to have the ordered commit record assigned to a physical | |
748 | * location in the log. | |
749 | */ | |
750 | spin_lock(&cil->xc_push_lock); | |
751 | ctx->commit_iclog = iclog; | |
752 | ctx->commit_lsn = lsn; | |
753 | wake_up_all(&cil->xc_commit_wait); | |
c45aba40 DC |
754 | spin_unlock(&cil->xc_push_lock); |
755 | } | |
756 | ||
757 | ||
2ce82b72 | 758 | /* |
bf034bc8 DC |
759 | * Ensure that the order of log writes follows checkpoint sequence order. This |
760 | * relies on the context LSN being zero until the log write has guaranteed the | |
761 | * LSN that the log write will start at via xlog_state_get_iclog_space(). | |
762 | */ | |
68a74dca DC |
763 | enum _record_type { |
764 | _START_RECORD, | |
765 | _COMMIT_RECORD, | |
766 | }; | |
767 | ||
bf034bc8 DC |
768 | static int |
769 | xlog_cil_order_write( | |
770 | struct xfs_cil *cil, | |
68a74dca DC |
771 | xfs_csn_t sequence, |
772 | enum _record_type record) | |
bf034bc8 DC |
773 | { |
774 | struct xfs_cil_ctx *ctx; | |
775 | ||
776 | restart: | |
777 | spin_lock(&cil->xc_push_lock); | |
778 | list_for_each_entry(ctx, &cil->xc_committing, committing) { | |
779 | /* | |
780 | * Avoid getting stuck in this loop because we were woken by the | |
781 | * shutdown, but then went back to sleep once already in the | |
782 | * shutdown state. | |
783 | */ | |
784 | if (xlog_is_shutdown(cil->xc_log)) { | |
785 | spin_unlock(&cil->xc_push_lock); | |
786 | return -EIO; | |
787 | } | |
788 | ||
789 | /* | |
790 | * Higher sequences will wait for this one so skip them. | |
791 | * Don't wait for our own sequence, either. | |
792 | */ | |
793 | if (ctx->sequence >= sequence) | |
794 | continue; | |
68a74dca DC |
795 | |
796 | /* Wait until the LSN for the record has been recorded. */ | |
797 | switch (record) { | |
798 | case _START_RECORD: | |
799 | if (!ctx->start_lsn) { | |
800 | xlog_wait(&cil->xc_start_wait, &cil->xc_push_lock); | |
801 | goto restart; | |
802 | } | |
803 | break; | |
804 | case _COMMIT_RECORD: | |
805 | if (!ctx->commit_lsn) { | |
806 | xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock); | |
807 | goto restart; | |
808 | } | |
809 | break; | |
bf034bc8 DC |
810 | } |
811 | } | |
812 | spin_unlock(&cil->xc_push_lock); | |
813 | return 0; | |
814 | } | |
815 | ||
68a74dca DC |
816 | /* |
817 | * Write out the log vector change now attached to the CIL context. This will | |
818 | * write a start record that needs to be strictly ordered in ascending CIL | |
819 | * sequence order so that log recovery will always use in-order start LSNs when | |
820 | * replaying checkpoints. | |
821 | */ | |
822 | static int | |
823 | xlog_cil_write_chain( | |
824 | struct xfs_cil_ctx *ctx, | |
825 | struct xfs_log_vec *chain) | |
826 | { | |
827 | struct xlog *log = ctx->cil->xc_log; | |
828 | int error; | |
829 | ||
830 | error = xlog_cil_order_write(ctx->cil, ctx->sequence, _START_RECORD); | |
831 | if (error) | |
832 | return error; | |
833 | return xlog_write(log, ctx, chain, ctx->ticket, XLOG_START_TRANS); | |
834 | } | |
835 | ||
bf034bc8 DC |
836 | /* |
837 | * Write out the commit record of a checkpoint transaction to close off a | |
838 | * running log write. These commit records are strictly ordered in ascending CIL | |
839 | * sequence order so that log recovery will always replay the checkpoints in the | |
840 | * correct order. | |
2ce82b72 DC |
841 | */ |
842 | static int | |
843 | xlog_cil_write_commit_record( | |
caa80090 | 844 | struct xfs_cil_ctx *ctx) |
2ce82b72 | 845 | { |
c45aba40 | 846 | struct xlog *log = ctx->cil->xc_log; |
54021b62 DC |
847 | struct xlog_op_header ophdr = { |
848 | .oh_clientid = XFS_TRANSACTION, | |
849 | .oh_tid = cpu_to_be32(ctx->ticket->t_tid), | |
850 | .oh_flags = XLOG_COMMIT_TRANS, | |
851 | }; | |
c45aba40 | 852 | struct xfs_log_iovec reg = { |
54021b62 DC |
853 | .i_addr = &ophdr, |
854 | .i_len = sizeof(struct xlog_op_header), | |
2ce82b72 DC |
855 | .i_type = XLOG_REG_TYPE_COMMIT, |
856 | }; | |
c45aba40 | 857 | struct xfs_log_vec vec = { |
2ce82b72 DC |
858 | .lv_niovecs = 1, |
859 | .lv_iovecp = ®, | |
860 | }; | |
c45aba40 | 861 | int error; |
2ce82b72 DC |
862 | |
863 | if (xlog_is_shutdown(log)) | |
864 | return -EIO; | |
865 | ||
68a74dca DC |
866 | error = xlog_cil_order_write(ctx->cil, ctx->sequence, _COMMIT_RECORD); |
867 | if (error) | |
868 | return error; | |
869 | ||
54021b62 DC |
870 | /* account for space used by record data */ |
871 | ctx->ticket->t_curr_res -= reg.i_len; | |
caa80090 | 872 | error = xlog_write(log, ctx, &vec, ctx->ticket, XLOG_COMMIT_TRANS); |
2ce82b72 | 873 | if (error) |
b5f17bec | 874 | xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); |
2ce82b72 DC |
875 | return error; |
876 | } | |
877 | ||
735fbf67 | 878 | struct xlog_cil_trans_hdr { |
6eaed95e | 879 | struct xlog_op_header oph[2]; |
735fbf67 | 880 | struct xfs_trans_header thdr; |
6eaed95e | 881 | struct xfs_log_iovec lhdr[2]; |
735fbf67 DC |
882 | }; |
883 | ||
884 | /* | |
885 | * Build a checkpoint transaction header to begin the journal transaction. We | |
886 | * need to account for the space used by the transaction header here as it is | |
887 | * not accounted for in xlog_write(). | |
6eaed95e DC |
888 | * |
889 | * This is the only place we write a transaction header, so we also build the | |
890 | * log opheaders that indicate the start of a log transaction and wrap the | |
891 | * transaction header. We keep the start record in it's own log vector rather | |
892 | * than compacting them into a single region as this ends up making the logic | |
893 | * in xlog_write() for handling empty opheaders for start, commit and unmount | |
894 | * records much simpler. | |
735fbf67 DC |
895 | */ |
896 | static void | |
897 | xlog_cil_build_trans_hdr( | |
898 | struct xfs_cil_ctx *ctx, | |
899 | struct xlog_cil_trans_hdr *hdr, | |
900 | struct xfs_log_vec *lvhdr, | |
901 | int num_iovecs) | |
902 | { | |
903 | struct xlog_ticket *tic = ctx->ticket; | |
6eaed95e | 904 | __be32 tid = cpu_to_be32(tic->t_tid); |
735fbf67 DC |
905 | |
906 | memset(hdr, 0, sizeof(*hdr)); | |
907 | ||
6eaed95e DC |
908 | /* Log start record */ |
909 | hdr->oph[0].oh_tid = tid; | |
910 | hdr->oph[0].oh_clientid = XFS_TRANSACTION; | |
911 | hdr->oph[0].oh_flags = XLOG_START_TRANS; | |
912 | ||
913 | /* log iovec region pointer */ | |
914 | hdr->lhdr[0].i_addr = &hdr->oph[0]; | |
915 | hdr->lhdr[0].i_len = sizeof(struct xlog_op_header); | |
916 | hdr->lhdr[0].i_type = XLOG_REG_TYPE_LRHEADER; | |
917 | ||
918 | /* log opheader */ | |
919 | hdr->oph[1].oh_tid = tid; | |
920 | hdr->oph[1].oh_clientid = XFS_TRANSACTION; | |
921 | hdr->oph[1].oh_len = cpu_to_be32(sizeof(struct xfs_trans_header)); | |
922 | ||
923 | /* transaction header in host byte order format */ | |
735fbf67 DC |
924 | hdr->thdr.th_magic = XFS_TRANS_HEADER_MAGIC; |
925 | hdr->thdr.th_type = XFS_TRANS_CHECKPOINT; | |
926 | hdr->thdr.th_tid = tic->t_tid; | |
927 | hdr->thdr.th_num_items = num_iovecs; | |
735fbf67 | 928 | |
6eaed95e DC |
929 | /* log iovec region pointer */ |
930 | hdr->lhdr[1].i_addr = &hdr->oph[1]; | |
931 | hdr->lhdr[1].i_len = sizeof(struct xlog_op_header) + | |
932 | sizeof(struct xfs_trans_header); | |
933 | hdr->lhdr[1].i_type = XLOG_REG_TYPE_TRANSHDR; | |
934 | ||
935 | tic->t_curr_res -= hdr->lhdr[0].i_len + hdr->lhdr[1].i_len; | |
936 | ||
937 | lvhdr->lv_niovecs = 2; | |
938 | lvhdr->lv_iovecp = &hdr->lhdr[0]; | |
735fbf67 DC |
939 | lvhdr->lv_next = ctx->lv_chain; |
940 | } | |
941 | ||
71e330b5 | 942 | /* |
c7cc296d CH |
943 | * Push the Committed Item List to the log. |
944 | * | |
945 | * If the current sequence is the same as xc_push_seq we need to do a flush. If | |
946 | * xc_push_seq is less than the current sequence, then it has already been | |
a44f13ed DC |
947 | * flushed and we don't need to do anything - the caller will wait for it to |
948 | * complete if necessary. | |
949 | * | |
c7cc296d CH |
950 | * xc_push_seq is checked unlocked against the sequence number for a match. |
951 | * Hence we can allow log forces to run racily and not issue pushes for the | |
952 | * same sequence twice. If we get a race between multiple pushes for the same | |
953 | * sequence they will block on the first one and then abort, hence avoiding | |
954 | * needless pushes. | |
71e330b5 | 955 | */ |
c7cc296d CH |
956 | static void |
957 | xlog_cil_push_work( | |
958 | struct work_struct *work) | |
71e330b5 | 959 | { |
39823d0f DC |
960 | struct xfs_cil_ctx *ctx = |
961 | container_of(work, struct xfs_cil_ctx, push_work); | |
962 | struct xfs_cil *cil = ctx->cil; | |
c7cc296d | 963 | struct xlog *log = cil->xc_log; |
71e330b5 | 964 | struct xfs_log_vec *lv; |
71e330b5 | 965 | struct xfs_cil_ctx *new_ctx; |
71e330b5 | 966 | int num_iovecs; |
71e330b5 | 967 | int error = 0; |
735fbf67 | 968 | struct xlog_cil_trans_hdr thdr; |
71e330b5 | 969 | struct xfs_log_vec lvhdr = { NULL }; |
0dc8f7f1 | 970 | xfs_csn_t push_seq; |
0020a190 | 971 | bool push_commit_stable; |
71e330b5 | 972 | |
39823d0f | 973 | new_ctx = xlog_cil_ctx_alloc(); |
71e330b5 DC |
974 | new_ctx->ticket = xlog_cil_ticket_alloc(log); |
975 | ||
4c2d542f | 976 | down_write(&cil->xc_ctx_lock); |
71e330b5 | 977 | |
4bb928cd | 978 | spin_lock(&cil->xc_push_lock); |
4c2d542f DC |
979 | push_seq = cil->xc_push_seq; |
980 | ASSERT(push_seq <= ctx->sequence); | |
0020a190 DC |
981 | push_commit_stable = cil->xc_push_commit_stable; |
982 | cil->xc_push_commit_stable = false; | |
71e330b5 | 983 | |
0e7ab7ef | 984 | /* |
19f4e7cc DC |
985 | * As we are about to switch to a new, empty CIL context, we no longer |
986 | * need to throttle tasks on CIL space overruns. Wake any waiters that | |
987 | * the hard push throttle may have caught so they can start committing | |
988 | * to the new context. The ctx->xc_push_lock provides the serialisation | |
989 | * necessary for safely using the lockless waitqueue_active() check in | |
990 | * this context. | |
0e7ab7ef | 991 | */ |
19f4e7cc | 992 | if (waitqueue_active(&cil->xc_push_wait)) |
c7f87f39 | 993 | wake_up_all(&cil->xc_push_wait); |
0e7ab7ef | 994 | |
4c2d542f DC |
995 | /* |
996 | * Check if we've anything to push. If there is nothing, then we don't | |
997 | * move on to a new sequence number and so we have to be able to push | |
998 | * this sequence again later. | |
999 | */ | |
1000 | if (list_empty(&cil->xc_cil)) { | |
1001 | cil->xc_push_seq = 0; | |
4bb928cd | 1002 | spin_unlock(&cil->xc_push_lock); |
a44f13ed | 1003 | goto out_skip; |
4c2d542f | 1004 | } |
4c2d542f | 1005 | |
a44f13ed | 1006 | |
cf085a1b | 1007 | /* check for a previously pushed sequence */ |
39823d0f | 1008 | if (push_seq < ctx->sequence) { |
8af3dcd3 | 1009 | spin_unlock(&cil->xc_push_lock); |
df806158 | 1010 | goto out_skip; |
8af3dcd3 DC |
1011 | } |
1012 | ||
1013 | /* | |
1014 | * We are now going to push this context, so add it to the committing | |
1015 | * list before we do anything else. This ensures that anyone waiting on | |
1016 | * this push can easily detect the difference between a "push in | |
1017 | * progress" and "CIL is empty, nothing to do". | |
1018 | * | |
1019 | * IOWs, a wait loop can now check for: | |
1020 | * the current sequence not being found on the committing list; | |
1021 | * an empty CIL; and | |
1022 | * an unchanged sequence number | |
1023 | * to detect a push that had nothing to do and therefore does not need | |
1024 | * waiting on. If the CIL is not empty, we get put on the committing | |
1025 | * list before emptying the CIL and bumping the sequence number. Hence | |
1026 | * an empty CIL and an unchanged sequence number means we jumped out | |
1027 | * above after doing nothing. | |
1028 | * | |
1029 | * Hence the waiter will either find the commit sequence on the | |
1030 | * committing list or the sequence number will be unchanged and the CIL | |
1031 | * still dirty. In that latter case, the push has not yet started, and | |
1032 | * so the waiter will have to continue trying to check the CIL | |
1033 | * committing list until it is found. In extreme cases of delay, the | |
1034 | * sequence may fully commit between the attempts the wait makes to wait | |
1035 | * on the commit sequence. | |
1036 | */ | |
1037 | list_add(&ctx->committing, &cil->xc_committing); | |
1038 | spin_unlock(&cil->xc_push_lock); | |
df806158 | 1039 | |
bad77c37 DC |
1040 | /* |
1041 | * Pull all the log vectors off the items in the CIL, and remove the | |
1042 | * items from the CIL. We don't need the CIL lock here because it's only | |
1043 | * needed on the transaction commit side which is currently locked out | |
1044 | * by the flush lock. | |
71e330b5 DC |
1045 | */ |
1046 | lv = NULL; | |
71e330b5 | 1047 | num_iovecs = 0; |
71e330b5 DC |
1048 | while (!list_empty(&cil->xc_cil)) { |
1049 | struct xfs_log_item *item; | |
71e330b5 DC |
1050 | |
1051 | item = list_first_entry(&cil->xc_cil, | |
1052 | struct xfs_log_item, li_cil); | |
1053 | list_del_init(&item->li_cil); | |
1054 | if (!ctx->lv_chain) | |
1055 | ctx->lv_chain = item->li_lv; | |
1056 | else | |
1057 | lv->lv_next = item->li_lv; | |
1058 | lv = item->li_lv; | |
1059 | item->li_lv = NULL; | |
71e330b5 | 1060 | num_iovecs += lv->lv_niovecs; |
71e330b5 DC |
1061 | } |
1062 | ||
1063 | /* | |
39823d0f | 1064 | * Switch the contexts so we can drop the context lock and move out |
71e330b5 DC |
1065 | * of a shared context. We can't just go straight to the commit record, |
1066 | * though - we need to synchronise with previous and future commits so | |
1067 | * that the commit records are correctly ordered in the log to ensure | |
1068 | * that we process items during log IO completion in the correct order. | |
1069 | * | |
1070 | * For example, if we get an EFI in one checkpoint and the EFD in the | |
1071 | * next (e.g. due to log forces), we do not want the checkpoint with | |
1072 | * the EFD to be committed before the checkpoint with the EFI. Hence | |
1073 | * we must strictly order the commit records of the checkpoints so | |
1074 | * that: a) the checkpoint callbacks are attached to the iclogs in the | |
1075 | * correct order; and b) the checkpoints are replayed in correct order | |
1076 | * in log recovery. | |
1077 | * | |
1078 | * Hence we need to add this context to the committing context list so | |
1079 | * that higher sequences will wait for us to write out a commit record | |
1080 | * before they do. | |
f876e446 | 1081 | * |
5f9b4b0d | 1082 | * xfs_log_force_seq requires us to mirror the new sequence into the cil |
f876e446 DC |
1083 | * structure atomically with the addition of this sequence to the |
1084 | * committing list. This also ensures that we can do unlocked checks | |
1085 | * against the current sequence in log forces without risking | |
1086 | * deferencing a freed context pointer. | |
71e330b5 | 1087 | */ |
4bb928cd | 1088 | spin_lock(&cil->xc_push_lock); |
39823d0f | 1089 | xlog_cil_ctx_switch(cil, new_ctx); |
4bb928cd | 1090 | spin_unlock(&cil->xc_push_lock); |
71e330b5 DC |
1091 | up_write(&cil->xc_ctx_lock); |
1092 | ||
1093 | /* | |
1094 | * Build a checkpoint transaction header and write it to the log to | |
1095 | * begin the transaction. We need to account for the space used by the | |
1096 | * transaction header here as it is not accounted for in xlog_write(). | |
71e330b5 | 1097 | */ |
735fbf67 | 1098 | xlog_cil_build_trans_hdr(ctx, &thdr, &lvhdr, num_iovecs); |
71e330b5 | 1099 | |
68a74dca | 1100 | error = xlog_cil_write_chain(ctx, &lvhdr); |
bf034bc8 DC |
1101 | if (error) |
1102 | goto out_abort_free_ticket; | |
71e330b5 | 1103 | |
caa80090 | 1104 | error = xlog_cil_write_commit_record(ctx); |
dd401770 DC |
1105 | if (error) |
1106 | goto out_abort_free_ticket; | |
1107 | ||
735fbf67 | 1108 | xfs_log_ticket_ungrant(log, ctx->ticket); |
71e330b5 | 1109 | |
a79b28c2 | 1110 | /* |
1effb72a DC |
1111 | * If the checkpoint spans multiple iclogs, wait for all previous iclogs |
1112 | * to complete before we submit the commit_iclog. We can't use state | |
1113 | * checks for this - ACTIVE can be either a past completed iclog or a | |
1114 | * future iclog being filled, while WANT_SYNC through SYNC_DONE can be a | |
1115 | * past or future iclog awaiting IO or ordered IO completion to be run. | |
1116 | * In the latter case, if it's a future iclog and we wait on it, the we | |
1117 | * will hang because it won't get processed through to ic_force_wait | |
1118 | * wakeup until this commit_iclog is written to disk. Hence we use the | |
1119 | * iclog header lsn and compare it to the commit lsn to determine if we | |
1120 | * need to wait on iclogs or not. | |
a79b28c2 | 1121 | */ |
caa80090 | 1122 | spin_lock(&log->l_icloglock); |
c45aba40 | 1123 | if (ctx->start_lsn != ctx->commit_lsn) { |
1effb72a DC |
1124 | xfs_lsn_t plsn; |
1125 | ||
caa80090 | 1126 | plsn = be64_to_cpu(ctx->commit_iclog->ic_prev->ic_header.h_lsn); |
c45aba40 | 1127 | if (plsn && XFS_LSN_CMP(plsn, ctx->commit_lsn) < 0) { |
1effb72a DC |
1128 | /* |
1129 | * Waiting on ic_force_wait orders the completion of | |
1130 | * iclogs older than ic_prev. Hence we only need to wait | |
1131 | * on the most recent older iclog here. | |
1132 | */ | |
caa80090 | 1133 | xlog_wait_on_iclog(ctx->commit_iclog->ic_prev); |
1effb72a DC |
1134 | spin_lock(&log->l_icloglock); |
1135 | } | |
1136 | ||
1137 | /* | |
1138 | * We need to issue a pre-flush so that the ordering for this | |
1139 | * checkpoint is correctly preserved down to stable storage. | |
1140 | */ | |
caa80090 | 1141 | ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FLUSH; |
a79b28c2 DC |
1142 | } |
1143 | ||
eef983ff DC |
1144 | /* |
1145 | * The commit iclog must be written to stable storage to guarantee | |
1146 | * journal IO vs metadata writeback IO is correctly ordered on stable | |
1147 | * storage. | |
0020a190 DC |
1148 | * |
1149 | * If the push caller needs the commit to be immediately stable and the | |
1150 | * commit_iclog is not yet marked as XLOG_STATE_WANT_SYNC to indicate it | |
1151 | * will be written when released, switch it's state to WANT_SYNC right | |
1152 | * now. | |
eef983ff | 1153 | */ |
caa80090 | 1154 | ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA; |
0020a190 DC |
1155 | if (push_commit_stable && |
1156 | ctx->commit_iclog->ic_state == XLOG_STATE_ACTIVE) | |
1157 | xlog_state_switch_iclogs(log, ctx->commit_iclog, 0); | |
919edbad | 1158 | xlog_state_release_iclog(log, ctx->commit_iclog); |
502a01fa DC |
1159 | |
1160 | /* Not safe to reference ctx now! */ | |
1161 | ||
eef983ff | 1162 | spin_unlock(&log->l_icloglock); |
c7cc296d | 1163 | return; |
71e330b5 DC |
1164 | |
1165 | out_skip: | |
1166 | up_write(&cil->xc_ctx_lock); | |
1167 | xfs_log_ticket_put(new_ctx->ticket); | |
1168 | kmem_free(new_ctx); | |
c7cc296d | 1169 | return; |
71e330b5 | 1170 | |
7db37c5e | 1171 | out_abort_free_ticket: |
735fbf67 | 1172 | xfs_log_ticket_ungrant(log, ctx->ticket); |
2039a272 | 1173 | ASSERT(xlog_is_shutdown(log)); |
caa80090 DC |
1174 | if (!ctx->commit_iclog) { |
1175 | xlog_cil_committed(ctx); | |
1176 | return; | |
1177 | } | |
1178 | spin_lock(&log->l_icloglock); | |
919edbad | 1179 | xlog_state_release_iclog(log, ctx->commit_iclog); |
caa80090 DC |
1180 | /* Not safe to reference ctx now! */ |
1181 | spin_unlock(&log->l_icloglock); | |
4c2d542f DC |
1182 | } |
1183 | ||
1184 | /* | |
1185 | * We need to push CIL every so often so we don't cache more than we can fit in | |
1186 | * the log. The limit really is that a checkpoint can't be more than half the | |
1187 | * log (the current checkpoint is not allowed to overwrite the previous | |
1188 | * checkpoint), but commit latency and memory usage limit this to a smaller | |
1189 | * size. | |
1190 | */ | |
1191 | static void | |
1192 | xlog_cil_push_background( | |
0e7ab7ef | 1193 | struct xlog *log) __releases(cil->xc_ctx_lock) |
4c2d542f DC |
1194 | { |
1195 | struct xfs_cil *cil = log->l_cilp; | |
1196 | ||
1197 | /* | |
1198 | * The cil won't be empty because we are called while holding the | |
1199 | * context lock so whatever we added to the CIL will still be there | |
1200 | */ | |
1201 | ASSERT(!list_empty(&cil->xc_cil)); | |
1202 | ||
1203 | /* | |
19f4e7cc | 1204 | * Don't do a background push if we haven't used up all the |
4c2d542f DC |
1205 | * space available yet. |
1206 | */ | |
0e7ab7ef DC |
1207 | if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log)) { |
1208 | up_read(&cil->xc_ctx_lock); | |
4c2d542f | 1209 | return; |
0e7ab7ef | 1210 | } |
4c2d542f | 1211 | |
4bb928cd | 1212 | spin_lock(&cil->xc_push_lock); |
4c2d542f DC |
1213 | if (cil->xc_push_seq < cil->xc_current_sequence) { |
1214 | cil->xc_push_seq = cil->xc_current_sequence; | |
33c0dd78 | 1215 | queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work); |
4c2d542f | 1216 | } |
0e7ab7ef DC |
1217 | |
1218 | /* | |
1219 | * Drop the context lock now, we can't hold that if we need to sleep | |
1220 | * because we are over the blocking threshold. The push_lock is still | |
1221 | * held, so blocking threshold sleep/wakeup is still correctly | |
1222 | * serialised here. | |
1223 | */ | |
1224 | up_read(&cil->xc_ctx_lock); | |
1225 | ||
1226 | /* | |
1227 | * If we are well over the space limit, throttle the work that is being | |
19f4e7cc DC |
1228 | * done until the push work on this context has begun. Enforce the hard |
1229 | * throttle on all transaction commits once it has been activated, even | |
1230 | * if the committing transactions have resulted in the space usage | |
1231 | * dipping back down under the hard limit. | |
1232 | * | |
1233 | * The ctx->xc_push_lock provides the serialisation necessary for safely | |
1234 | * using the lockless waitqueue_active() check in this context. | |
0e7ab7ef | 1235 | */ |
19f4e7cc DC |
1236 | if (cil->xc_ctx->space_used >= XLOG_CIL_BLOCKING_SPACE_LIMIT(log) || |
1237 | waitqueue_active(&cil->xc_push_wait)) { | |
0e7ab7ef DC |
1238 | trace_xfs_log_cil_wait(log, cil->xc_ctx->ticket); |
1239 | ASSERT(cil->xc_ctx->space_used < log->l_logsize); | |
c7f87f39 | 1240 | xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock); |
0e7ab7ef DC |
1241 | return; |
1242 | } | |
1243 | ||
4bb928cd | 1244 | spin_unlock(&cil->xc_push_lock); |
4c2d542f DC |
1245 | |
1246 | } | |
1247 | ||
f876e446 DC |
1248 | /* |
1249 | * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence | |
1250 | * number that is passed. When it returns, the work will be queued for | |
0020a190 DC |
1251 | * @push_seq, but it won't be completed. |
1252 | * | |
1253 | * If the caller is performing a synchronous force, we will flush the workqueue | |
1254 | * to get previously queued work moving to minimise the wait time they will | |
1255 | * undergo waiting for all outstanding pushes to complete. The caller is | |
1256 | * expected to do the required waiting for push_seq to complete. | |
1257 | * | |
1258 | * If the caller is performing an async push, we need to ensure that the | |
1259 | * checkpoint is fully flushed out of the iclogs when we finish the push. If we | |
1260 | * don't do this, then the commit record may remain sitting in memory in an | |
1261 | * ACTIVE iclog. This then requires another full log force to push to disk, | |
1262 | * which defeats the purpose of having an async, non-blocking CIL force | |
1263 | * mechanism. Hence in this case we need to pass a flag to the push work to | |
1264 | * indicate it needs to flush the commit record itself. | |
f876e446 | 1265 | */ |
4c2d542f | 1266 | static void |
f876e446 | 1267 | xlog_cil_push_now( |
f7bdf03a | 1268 | struct xlog *log, |
0020a190 DC |
1269 | xfs_lsn_t push_seq, |
1270 | bool async) | |
4c2d542f DC |
1271 | { |
1272 | struct xfs_cil *cil = log->l_cilp; | |
1273 | ||
1274 | if (!cil) | |
1275 | return; | |
1276 | ||
1277 | ASSERT(push_seq && push_seq <= cil->xc_current_sequence); | |
1278 | ||
1279 | /* start on any pending background push to minimise wait time on it */ | |
0020a190 | 1280 | if (!async) |
33c0dd78 | 1281 | flush_workqueue(cil->xc_push_wq); |
4c2d542f | 1282 | |
70447e0a DC |
1283 | spin_lock(&cil->xc_push_lock); |
1284 | ||
1285 | /* | |
1286 | * If this is an async flush request, we always need to set the | |
1287 | * xc_push_commit_stable flag even if something else has already queued | |
1288 | * a push. The flush caller is asking for the CIL to be on stable | |
1289 | * storage when the next push completes, so regardless of who has queued | |
1290 | * the push, the flush requires stable semantics from it. | |
1291 | */ | |
1292 | cil->xc_push_commit_stable = async; | |
1293 | ||
4c2d542f DC |
1294 | /* |
1295 | * If the CIL is empty or we've already pushed the sequence then | |
70447e0a | 1296 | * there's no more work that we need to do. |
4c2d542f | 1297 | */ |
4c2d542f | 1298 | if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) { |
4bb928cd | 1299 | spin_unlock(&cil->xc_push_lock); |
4c2d542f DC |
1300 | return; |
1301 | } | |
1302 | ||
1303 | cil->xc_push_seq = push_seq; | |
33c0dd78 | 1304 | queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work); |
4bb928cd | 1305 | spin_unlock(&cil->xc_push_lock); |
4c2d542f DC |
1306 | } |
1307 | ||
2c6e24ce DC |
1308 | bool |
1309 | xlog_cil_empty( | |
1310 | struct xlog *log) | |
1311 | { | |
1312 | struct xfs_cil *cil = log->l_cilp; | |
1313 | bool empty = false; | |
1314 | ||
1315 | spin_lock(&cil->xc_push_lock); | |
1316 | if (list_empty(&cil->xc_cil)) | |
1317 | empty = true; | |
1318 | spin_unlock(&cil->xc_push_lock); | |
1319 | return empty; | |
1320 | } | |
1321 | ||
a44f13ed DC |
1322 | /* |
1323 | * Commit a transaction with the given vector to the Committed Item List. | |
1324 | * | |
1325 | * To do this, we need to format the item, pin it in memory if required and | |
1326 | * account for the space used by the transaction. Once we have done that we | |
1327 | * need to release the unused reservation for the transaction, attach the | |
1328 | * transaction to the checkpoint context so we carry the busy extents through | |
1329 | * to checkpoint completion, and then unlock all the items in the transaction. | |
1330 | * | |
a44f13ed DC |
1331 | * Called with the context lock already held in read mode to lock out |
1332 | * background commit, returns without it held once background commits are | |
1333 | * allowed again. | |
1334 | */ | |
c6f97264 | 1335 | void |
5f9b4b0d DC |
1336 | xlog_cil_commit( |
1337 | struct xlog *log, | |
a44f13ed | 1338 | struct xfs_trans *tp, |
5f9b4b0d | 1339 | xfs_csn_t *commit_seq, |
70393313 | 1340 | bool regrant) |
a44f13ed | 1341 | { |
991aaf65 | 1342 | struct xfs_cil *cil = log->l_cilp; |
195cd83d | 1343 | struct xfs_log_item *lip, *next; |
a44f13ed | 1344 | |
b1c5ebb2 DC |
1345 | /* |
1346 | * Do all necessary memory allocation before we lock the CIL. | |
1347 | * This ensures the allocation does not deadlock with a CIL | |
1348 | * push in memory reclaim (e.g. from kswapd). | |
1349 | */ | |
1350 | xlog_cil_alloc_shadow_bufs(log, tp); | |
1351 | ||
f5baac35 | 1352 | /* lock out background commit */ |
991aaf65 | 1353 | down_read(&cil->xc_ctx_lock); |
f5baac35 | 1354 | |
991aaf65 | 1355 | xlog_cil_insert_items(log, tp); |
a44f13ed | 1356 | |
2039a272 | 1357 | if (regrant && !xlog_is_shutdown(log)) |
8b41e3f9 CH |
1358 | xfs_log_ticket_regrant(log, tp->t_ticket); |
1359 | else | |
1360 | xfs_log_ticket_ungrant(log, tp->t_ticket); | |
ba18781b | 1361 | tp->t_ticket = NULL; |
a44f13ed DC |
1362 | xfs_trans_unreserve_and_mod_sb(tp); |
1363 | ||
1364 | /* | |
1365 | * Once all the items of the transaction have been copied to the CIL, | |
195cd83d | 1366 | * the items can be unlocked and possibly freed. |
a44f13ed DC |
1367 | * |
1368 | * This needs to be done before we drop the CIL context lock because we | |
1369 | * have to update state in the log items and unlock them before they go | |
1370 | * to disk. If we don't, then the CIL checkpoint can race with us and | |
1371 | * we can run checkpoint completion before we've updated and unlocked | |
1372 | * the log items. This affects (at least) processing of stale buffers, | |
1373 | * inodes and EFIs. | |
1374 | */ | |
195cd83d CH |
1375 | trace_xfs_trans_commit_items(tp, _RET_IP_); |
1376 | list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) { | |
1377 | xfs_trans_del_item(lip); | |
1378 | if (lip->li_ops->iop_committing) | |
5f9b4b0d | 1379 | lip->li_ops->iop_committing(lip, cil->xc_ctx->sequence); |
195cd83d | 1380 | } |
5f9b4b0d DC |
1381 | if (commit_seq) |
1382 | *commit_seq = cil->xc_ctx->sequence; | |
a44f13ed | 1383 | |
0e7ab7ef DC |
1384 | /* xlog_cil_push_background() releases cil->xc_ctx_lock */ |
1385 | xlog_cil_push_background(log); | |
a44f13ed DC |
1386 | } |
1387 | ||
0020a190 DC |
1388 | /* |
1389 | * Flush the CIL to stable storage but don't wait for it to complete. This | |
1390 | * requires the CIL push to ensure the commit record for the push hits the disk, | |
1391 | * but otherwise is no different to a push done from a log force. | |
1392 | */ | |
1393 | void | |
1394 | xlog_cil_flush( | |
1395 | struct xlog *log) | |
1396 | { | |
1397 | xfs_csn_t seq = log->l_cilp->xc_current_sequence; | |
1398 | ||
1399 | trace_xfs_log_force(log->l_mp, seq, _RET_IP_); | |
1400 | xlog_cil_push_now(log, seq, true); | |
70447e0a DC |
1401 | |
1402 | /* | |
1403 | * If the CIL is empty, make sure that any previous checkpoint that may | |
1404 | * still be in an active iclog is pushed to stable storage. | |
1405 | */ | |
1406 | if (list_empty(&log->l_cilp->xc_cil)) | |
1407 | xfs_log_force(log->l_mp, 0); | |
0020a190 DC |
1408 | } |
1409 | ||
71e330b5 DC |
1410 | /* |
1411 | * Conditionally push the CIL based on the sequence passed in. | |
1412 | * | |
0020a190 DC |
1413 | * We only need to push if we haven't already pushed the sequence number given. |
1414 | * Hence the only time we will trigger a push here is if the push sequence is | |
1415 | * the same as the current context. | |
71e330b5 DC |
1416 | * |
1417 | * We return the current commit lsn to allow the callers to determine if a | |
1418 | * iclog flush is necessary following this call. | |
71e330b5 DC |
1419 | */ |
1420 | xfs_lsn_t | |
5f9b4b0d | 1421 | xlog_cil_force_seq( |
f7bdf03a | 1422 | struct xlog *log, |
5f9b4b0d | 1423 | xfs_csn_t sequence) |
71e330b5 DC |
1424 | { |
1425 | struct xfs_cil *cil = log->l_cilp; | |
1426 | struct xfs_cil_ctx *ctx; | |
1427 | xfs_lsn_t commit_lsn = NULLCOMMITLSN; | |
1428 | ||
a44f13ed DC |
1429 | ASSERT(sequence <= cil->xc_current_sequence); |
1430 | ||
0020a190 DC |
1431 | if (!sequence) |
1432 | sequence = cil->xc_current_sequence; | |
1433 | trace_xfs_log_force(log->l_mp, sequence, _RET_IP_); | |
1434 | ||
a44f13ed DC |
1435 | /* |
1436 | * check to see if we need to force out the current context. | |
1437 | * xlog_cil_push() handles racing pushes for the same sequence, | |
1438 | * so no need to deal with it here. | |
1439 | */ | |
f876e446 | 1440 | restart: |
0020a190 | 1441 | xlog_cil_push_now(log, sequence, false); |
71e330b5 DC |
1442 | |
1443 | /* | |
1444 | * See if we can find a previous sequence still committing. | |
71e330b5 DC |
1445 | * We need to wait for all previous sequence commits to complete |
1446 | * before allowing the force of push_seq to go ahead. Hence block | |
1447 | * on commits for those as well. | |
1448 | */ | |
4bb928cd | 1449 | spin_lock(&cil->xc_push_lock); |
71e330b5 | 1450 | list_for_each_entry(ctx, &cil->xc_committing, committing) { |
ac983517 DC |
1451 | /* |
1452 | * Avoid getting stuck in this loop because we were woken by the | |
1453 | * shutdown, but then went back to sleep once already in the | |
1454 | * shutdown state. | |
1455 | */ | |
2039a272 | 1456 | if (xlog_is_shutdown(log)) |
ac983517 | 1457 | goto out_shutdown; |
a44f13ed | 1458 | if (ctx->sequence > sequence) |
71e330b5 DC |
1459 | continue; |
1460 | if (!ctx->commit_lsn) { | |
1461 | /* | |
1462 | * It is still being pushed! Wait for the push to | |
1463 | * complete, then start again from the beginning. | |
1464 | */ | |
0020a190 | 1465 | XFS_STATS_INC(log->l_mp, xs_log_force_sleep); |
4bb928cd | 1466 | xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock); |
71e330b5 DC |
1467 | goto restart; |
1468 | } | |
a44f13ed | 1469 | if (ctx->sequence != sequence) |
71e330b5 DC |
1470 | continue; |
1471 | /* found it! */ | |
1472 | commit_lsn = ctx->commit_lsn; | |
1473 | } | |
f876e446 DC |
1474 | |
1475 | /* | |
1476 | * The call to xlog_cil_push_now() executes the push in the background. | |
1477 | * Hence by the time we have got here it our sequence may not have been | |
1478 | * pushed yet. This is true if the current sequence still matches the | |
1479 | * push sequence after the above wait loop and the CIL still contains | |
8af3dcd3 DC |
1480 | * dirty objects. This is guaranteed by the push code first adding the |
1481 | * context to the committing list before emptying the CIL. | |
f876e446 | 1482 | * |
8af3dcd3 DC |
1483 | * Hence if we don't find the context in the committing list and the |
1484 | * current sequence number is unchanged then the CIL contents are | |
1485 | * significant. If the CIL is empty, if means there was nothing to push | |
1486 | * and that means there is nothing to wait for. If the CIL is not empty, | |
1487 | * it means we haven't yet started the push, because if it had started | |
1488 | * we would have found the context on the committing list. | |
f876e446 | 1489 | */ |
f876e446 DC |
1490 | if (sequence == cil->xc_current_sequence && |
1491 | !list_empty(&cil->xc_cil)) { | |
1492 | spin_unlock(&cil->xc_push_lock); | |
1493 | goto restart; | |
1494 | } | |
1495 | ||
4bb928cd | 1496 | spin_unlock(&cil->xc_push_lock); |
71e330b5 | 1497 | return commit_lsn; |
ac983517 DC |
1498 | |
1499 | /* | |
1500 | * We detected a shutdown in progress. We need to trigger the log force | |
1501 | * to pass through it's iclog state machine error handling, even though | |
1502 | * we are already in a shutdown state. Hence we can't return | |
1503 | * NULLCOMMITLSN here as that has special meaning to log forces (i.e. | |
1504 | * LSN is already stable), so we return a zero LSN instead. | |
1505 | */ | |
1506 | out_shutdown: | |
1507 | spin_unlock(&cil->xc_push_lock); | |
1508 | return 0; | |
71e330b5 | 1509 | } |
ccf7c23f DC |
1510 | |
1511 | /* | |
1512 | * Check if the current log item was first committed in this sequence. | |
1513 | * We can't rely on just the log item being in the CIL, we have to check | |
1514 | * the recorded commit sequence number. | |
1515 | * | |
1516 | * Note: for this to be used in a non-racy manner, it has to be called with | |
1517 | * CIL flushing locked out. As a result, it should only be used during the | |
1518 | * transaction commit process when deciding what to format into the item. | |
1519 | */ | |
1520 | bool | |
1521 | xfs_log_item_in_current_chkpt( | |
f8d92a66 | 1522 | struct xfs_log_item *lip) |
ccf7c23f | 1523 | { |
d86142dd | 1524 | struct xfs_cil *cil = lip->li_log->l_cilp; |
ccf7c23f | 1525 | |
ccf7c23f DC |
1526 | if (list_empty(&lip->li_cil)) |
1527 | return false; | |
1528 | ||
ccf7c23f DC |
1529 | /* |
1530 | * li_seq is written on the first commit of a log item to record the | |
1531 | * first checkpoint it is written to. Hence if it is different to the | |
1532 | * current sequence, we're in a new checkpoint. | |
1533 | */ | |
f8d92a66 | 1534 | return lip->li_seq == READ_ONCE(cil->xc_current_sequence); |
ccf7c23f | 1535 | } |
4c2d542f DC |
1536 | |
1537 | /* | |
1538 | * Perform initial CIL structure initialisation. | |
1539 | */ | |
1540 | int | |
1541 | xlog_cil_init( | |
f7bdf03a | 1542 | struct xlog *log) |
4c2d542f DC |
1543 | { |
1544 | struct xfs_cil *cil; | |
1545 | struct xfs_cil_ctx *ctx; | |
1546 | ||
707e0dda | 1547 | cil = kmem_zalloc(sizeof(*cil), KM_MAYFAIL); |
4c2d542f | 1548 | if (!cil) |
2451337d | 1549 | return -ENOMEM; |
33c0dd78 DC |
1550 | /* |
1551 | * Limit the CIL pipeline depth to 4 concurrent works to bound the | |
1552 | * concurrency the log spinlocks will be exposed to. | |
1553 | */ | |
1554 | cil->xc_push_wq = alloc_workqueue("xfs-cil/%s", | |
1555 | XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | WQ_UNBOUND), | |
1556 | 4, log->l_mp->m_super->s_id); | |
1557 | if (!cil->xc_push_wq) | |
1558 | goto out_destroy_cil; | |
4c2d542f | 1559 | |
4c2d542f DC |
1560 | INIT_LIST_HEAD(&cil->xc_cil); |
1561 | INIT_LIST_HEAD(&cil->xc_committing); | |
1562 | spin_lock_init(&cil->xc_cil_lock); | |
4bb928cd | 1563 | spin_lock_init(&cil->xc_push_lock); |
c7f87f39 | 1564 | init_waitqueue_head(&cil->xc_push_wait); |
4c2d542f | 1565 | init_rwsem(&cil->xc_ctx_lock); |
68a74dca | 1566 | init_waitqueue_head(&cil->xc_start_wait); |
4c2d542f | 1567 | init_waitqueue_head(&cil->xc_commit_wait); |
4c2d542f DC |
1568 | cil->xc_log = log; |
1569 | log->l_cilp = cil; | |
39823d0f DC |
1570 | |
1571 | ctx = xlog_cil_ctx_alloc(); | |
1572 | xlog_cil_ctx_switch(cil, ctx); | |
1573 | ||
4c2d542f | 1574 | return 0; |
33c0dd78 DC |
1575 | |
1576 | out_destroy_cil: | |
1577 | kmem_free(cil); | |
1578 | return -ENOMEM; | |
4c2d542f DC |
1579 | } |
1580 | ||
1581 | void | |
1582 | xlog_cil_destroy( | |
f7bdf03a | 1583 | struct xlog *log) |
4c2d542f DC |
1584 | { |
1585 | if (log->l_cilp->xc_ctx) { | |
1586 | if (log->l_cilp->xc_ctx->ticket) | |
1587 | xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket); | |
1588 | kmem_free(log->l_cilp->xc_ctx); | |
1589 | } | |
1590 | ||
1591 | ASSERT(list_empty(&log->l_cilp->xc_cil)); | |
33c0dd78 | 1592 | destroy_workqueue(log->l_cilp->xc_push_wq); |
4c2d542f DC |
1593 | kmem_free(log->l_cilp); |
1594 | } | |
1595 |