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 | ||
50 | /* | |
51 | * After the first stage of log recovery is done, we know where the head and | |
52 | * tail of the log are. We need this log initialisation done before we can | |
53 | * initialise the first CIL checkpoint context. | |
54 | * | |
55 | * Here we allocate a log ticket to track space usage during a CIL push. This | |
56 | * ticket is passed to xlog_write() directly so that we don't slowly leak log | |
57 | * space by failing to account for space used by log headers and additional | |
58 | * region headers for split regions. | |
59 | */ | |
60 | void | |
61 | xlog_cil_init_post_recovery( | |
f7bdf03a | 62 | struct xlog *log) |
71e330b5 | 63 | { |
71e330b5 DC |
64 | log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log); |
65 | log->l_cilp->xc_ctx->sequence = 1; | |
71e330b5 DC |
66 | } |
67 | ||
b1c5ebb2 DC |
68 | static inline int |
69 | xlog_cil_iovec_space( | |
70 | uint niovecs) | |
71 | { | |
72 | return round_up((sizeof(struct xfs_log_vec) + | |
73 | niovecs * sizeof(struct xfs_log_iovec)), | |
74 | sizeof(uint64_t)); | |
75 | } | |
76 | ||
77 | /* | |
78 | * Allocate or pin log vector buffers for CIL insertion. | |
79 | * | |
80 | * The CIL currently uses disposable buffers for copying a snapshot of the | |
81 | * modified items into the log during a push. The biggest problem with this is | |
82 | * the requirement to allocate the disposable buffer during the commit if: | |
83 | * a) does not exist; or | |
84 | * b) it is too small | |
85 | * | |
86 | * If we do this allocation within xlog_cil_insert_format_items(), it is done | |
87 | * under the xc_ctx_lock, which means that a CIL push cannot occur during | |
88 | * the memory allocation. This means that we have a potential deadlock situation | |
89 | * under low memory conditions when we have lots of dirty metadata pinned in | |
90 | * the CIL and we need a CIL commit to occur to free memory. | |
91 | * | |
92 | * To avoid this, we need to move the memory allocation outside the | |
93 | * xc_ctx_lock, but because the log vector buffers are disposable, that opens | |
94 | * up a TOCTOU race condition w.r.t. the CIL committing and removing the log | |
95 | * vector buffers between the check and the formatting of the item into the | |
96 | * log vector buffer within the xc_ctx_lock. | |
97 | * | |
98 | * Because the log vector buffer needs to be unchanged during the CIL push | |
99 | * process, we cannot share the buffer between the transaction commit (which | |
100 | * modifies the buffer) and the CIL push context that is writing the changes | |
101 | * into the log. This means skipping preallocation of buffer space is | |
102 | * unreliable, but we most definitely do not want to be allocating and freeing | |
103 | * buffers unnecessarily during commits when overwrites can be done safely. | |
104 | * | |
105 | * The simplest solution to this problem is to allocate a shadow buffer when a | |
106 | * log item is committed for the second time, and then to only use this buffer | |
107 | * if necessary. The buffer can remain attached to the log item until such time | |
108 | * it is needed, and this is the buffer that is reallocated to match the size of | |
109 | * the incoming modification. Then during the formatting of the item we can swap | |
110 | * the active buffer with the new one if we can't reuse the existing buffer. We | |
111 | * don't free the old buffer as it may be reused on the next modification if | |
112 | * it's size is right, otherwise we'll free and reallocate it at that point. | |
113 | * | |
114 | * This function builds a vector for the changes in each log item in the | |
115 | * transaction. It then works out the length of the buffer needed for each log | |
116 | * item, allocates them and attaches the vector to the log item in preparation | |
117 | * for the formatting step which occurs under the xc_ctx_lock. | |
118 | * | |
119 | * While this means the memory footprint goes up, it avoids the repeated | |
120 | * alloc/free pattern that repeated modifications of an item would otherwise | |
121 | * cause, and hence minimises the CPU overhead of such behaviour. | |
122 | */ | |
123 | static void | |
124 | xlog_cil_alloc_shadow_bufs( | |
125 | struct xlog *log, | |
126 | struct xfs_trans *tp) | |
127 | { | |
e6631f85 | 128 | struct xfs_log_item *lip; |
b1c5ebb2 | 129 | |
e6631f85 | 130 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
b1c5ebb2 DC |
131 | struct xfs_log_vec *lv; |
132 | int niovecs = 0; | |
133 | int nbytes = 0; | |
134 | int buf_size; | |
135 | bool ordered = false; | |
136 | ||
137 | /* Skip items which aren't dirty in this transaction. */ | |
e6631f85 | 138 | if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) |
b1c5ebb2 DC |
139 | continue; |
140 | ||
141 | /* get number of vecs and size of data to be stored */ | |
142 | lip->li_ops->iop_size(lip, &niovecs, &nbytes); | |
143 | ||
144 | /* | |
145 | * Ordered items need to be tracked but we do not wish to write | |
146 | * them. We need a logvec to track the object, but we do not | |
147 | * need an iovec or buffer to be allocated for copying data. | |
148 | */ | |
149 | if (niovecs == XFS_LOG_VEC_ORDERED) { | |
150 | ordered = true; | |
151 | niovecs = 0; | |
152 | nbytes = 0; | |
153 | } | |
154 | ||
155 | /* | |
156 | * We 64-bit align the length of each iovec so that the start | |
157 | * of the next one is naturally aligned. We'll need to | |
158 | * account for that slack space here. Then round nbytes up | |
159 | * to 64-bit alignment so that the initial buffer alignment is | |
160 | * easy to calculate and verify. | |
161 | */ | |
162 | nbytes += niovecs * sizeof(uint64_t); | |
163 | nbytes = round_up(nbytes, sizeof(uint64_t)); | |
164 | ||
165 | /* | |
166 | * The data buffer needs to start 64-bit aligned, so round up | |
167 | * that space to ensure we can align it appropriately and not | |
168 | * overrun the buffer. | |
169 | */ | |
170 | buf_size = nbytes + xlog_cil_iovec_space(niovecs); | |
171 | ||
172 | /* | |
173 | * if we have no shadow buffer, or it is too small, we need to | |
174 | * reallocate it. | |
175 | */ | |
176 | if (!lip->li_lv_shadow || | |
177 | buf_size > lip->li_lv_shadow->lv_size) { | |
178 | ||
179 | /* | |
180 | * We free and allocate here as a realloc would copy | |
cf085a1b | 181 | * unnecessary data. We don't use kmem_zalloc() for the |
b1c5ebb2 DC |
182 | * same reason - we don't need to zero the data area in |
183 | * the buffer, only the log vector header and the iovec | |
184 | * storage. | |
185 | */ | |
186 | kmem_free(lip->li_lv_shadow); | |
187 | ||
d634525d DC |
188 | /* |
189 | * We are in transaction context, which means this | |
190 | * allocation will pick up GFP_NOFS from the | |
191 | * memalloc_nofs_save/restore context the transaction | |
192 | * holds. This means we can use GFP_KERNEL here so the | |
193 | * generic kvmalloc() code will run vmalloc on | |
194 | * contiguous page allocation failure as we require. | |
195 | */ | |
196 | lv = kvmalloc(buf_size, GFP_KERNEL); | |
b1c5ebb2 DC |
197 | memset(lv, 0, xlog_cil_iovec_space(niovecs)); |
198 | ||
199 | lv->lv_item = lip; | |
200 | lv->lv_size = buf_size; | |
201 | if (ordered) | |
202 | lv->lv_buf_len = XFS_LOG_VEC_ORDERED; | |
203 | else | |
204 | lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1]; | |
205 | lip->li_lv_shadow = lv; | |
206 | } else { | |
207 | /* same or smaller, optimise common overwrite case */ | |
208 | lv = lip->li_lv_shadow; | |
209 | if (ordered) | |
210 | lv->lv_buf_len = XFS_LOG_VEC_ORDERED; | |
211 | else | |
212 | lv->lv_buf_len = 0; | |
213 | lv->lv_bytes = 0; | |
214 | lv->lv_next = NULL; | |
215 | } | |
216 | ||
217 | /* Ensure the lv is set up according to ->iop_size */ | |
218 | lv->lv_niovecs = niovecs; | |
219 | ||
220 | /* The allocated data region lies beyond the iovec region */ | |
221 | lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs); | |
222 | } | |
223 | ||
224 | } | |
225 | ||
991aaf65 DC |
226 | /* |
227 | * Prepare the log item for insertion into the CIL. Calculate the difference in | |
228 | * log space and vectors it will consume, and if it is a new item pin it as | |
229 | * well. | |
230 | */ | |
231 | STATIC void | |
232 | xfs_cil_prepare_item( | |
233 | struct xlog *log, | |
234 | struct xfs_log_vec *lv, | |
235 | struct xfs_log_vec *old_lv, | |
236 | int *diff_len, | |
237 | int *diff_iovecs) | |
238 | { | |
239 | /* Account for the new LV being passed in */ | |
240 | if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) { | |
110dc24a | 241 | *diff_len += lv->lv_bytes; |
991aaf65 DC |
242 | *diff_iovecs += lv->lv_niovecs; |
243 | } | |
244 | ||
245 | /* | |
246 | * If there is no old LV, this is the first time we've seen the item in | |
247 | * this CIL context and so we need to pin it. If we are replacing the | |
b1c5ebb2 DC |
248 | * old_lv, then remove the space it accounts for and make it the shadow |
249 | * buffer for later freeing. In both cases we are now switching to the | |
b63da6c8 | 250 | * shadow buffer, so update the pointer to it appropriately. |
991aaf65 | 251 | */ |
b1c5ebb2 | 252 | if (!old_lv) { |
e8b78db7 CH |
253 | if (lv->lv_item->li_ops->iop_pin) |
254 | lv->lv_item->li_ops->iop_pin(lv->lv_item); | |
b1c5ebb2 DC |
255 | lv->lv_item->li_lv_shadow = NULL; |
256 | } else if (old_lv != lv) { | |
991aaf65 DC |
257 | ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED); |
258 | ||
110dc24a | 259 | *diff_len -= old_lv->lv_bytes; |
991aaf65 | 260 | *diff_iovecs -= old_lv->lv_niovecs; |
b1c5ebb2 | 261 | lv->lv_item->li_lv_shadow = old_lv; |
991aaf65 DC |
262 | } |
263 | ||
264 | /* attach new log vector to log item */ | |
265 | lv->lv_item->li_lv = lv; | |
266 | ||
267 | /* | |
268 | * If this is the first time the item is being committed to the | |
269 | * CIL, store the sequence number on the log item so we can | |
270 | * tell in future commits whether this is the first checkpoint | |
271 | * the item is being committed into. | |
272 | */ | |
273 | if (!lv->lv_item->li_seq) | |
274 | lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence; | |
275 | } | |
276 | ||
71e330b5 DC |
277 | /* |
278 | * Format log item into a flat buffers | |
279 | * | |
280 | * For delayed logging, we need to hold a formatted buffer containing all the | |
281 | * changes on the log item. This enables us to relog the item in memory and | |
282 | * write it out asynchronously without needing to relock the object that was | |
283 | * modified at the time it gets written into the iclog. | |
284 | * | |
b1c5ebb2 DC |
285 | * This function takes the prepared log vectors attached to each log item, and |
286 | * formats the changes into the log vector buffer. The buffer it uses is | |
287 | * dependent on the current state of the vector in the CIL - the shadow lv is | |
288 | * guaranteed to be large enough for the current modification, but we will only | |
289 | * use that if we can't reuse the existing lv. If we can't reuse the existing | |
290 | * lv, then simple swap it out for the shadow lv. We don't free it - that is | |
291 | * done lazily either by th enext modification or the freeing of the log item. | |
71e330b5 DC |
292 | * |
293 | * We don't set up region headers during this process; we simply copy the | |
294 | * regions into the flat buffer. We can do this because we still have to do a | |
295 | * formatting step to write the regions into the iclog buffer. Writing the | |
296 | * ophdrs during the iclog write means that we can support splitting large | |
297 | * regions across iclog boundares without needing a change in the format of the | |
298 | * item/region encapsulation. | |
299 | * | |
300 | * Hence what we need to do now is change the rewrite the vector array to point | |
301 | * to the copied region inside the buffer we just allocated. This allows us to | |
302 | * format the regions into the iclog as though they are being formatted | |
303 | * directly out of the objects themselves. | |
304 | */ | |
991aaf65 DC |
305 | static void |
306 | xlog_cil_insert_format_items( | |
307 | struct xlog *log, | |
308 | struct xfs_trans *tp, | |
309 | int *diff_len, | |
310 | int *diff_iovecs) | |
71e330b5 | 311 | { |
e6631f85 | 312 | struct xfs_log_item *lip; |
71e330b5 | 313 | |
0244b960 CH |
314 | |
315 | /* Bail out if we didn't find a log item. */ | |
316 | if (list_empty(&tp->t_items)) { | |
317 | ASSERT(0); | |
991aaf65 | 318 | return; |
0244b960 CH |
319 | } |
320 | ||
e6631f85 | 321 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
7492c5b4 | 322 | struct xfs_log_vec *lv; |
b1c5ebb2 DC |
323 | struct xfs_log_vec *old_lv = NULL; |
324 | struct xfs_log_vec *shadow; | |
fd63875c | 325 | bool ordered = false; |
71e330b5 | 326 | |
0244b960 | 327 | /* Skip items which aren't dirty in this transaction. */ |
e6631f85 | 328 | if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) |
0244b960 CH |
329 | continue; |
330 | ||
fd63875c | 331 | /* |
b1c5ebb2 DC |
332 | * The formatting size information is already attached to |
333 | * the shadow lv on the log item. | |
fd63875c | 334 | */ |
b1c5ebb2 DC |
335 | shadow = lip->li_lv_shadow; |
336 | if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED) | |
fd63875c | 337 | ordered = true; |
fd63875c | 338 | |
b1c5ebb2 DC |
339 | /* Skip items that do not have any vectors for writing */ |
340 | if (!shadow->lv_niovecs && !ordered) | |
341 | continue; | |
0244b960 | 342 | |
f5baac35 | 343 | /* compare to existing item size */ |
b1c5ebb2 DC |
344 | old_lv = lip->li_lv; |
345 | if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) { | |
f5baac35 DC |
346 | /* same or smaller, optimise common overwrite case */ |
347 | lv = lip->li_lv; | |
348 | lv->lv_next = NULL; | |
349 | ||
350 | if (ordered) | |
351 | goto insert; | |
352 | ||
991aaf65 DC |
353 | /* |
354 | * set the item up as though it is a new insertion so | |
355 | * that the space reservation accounting is correct. | |
356 | */ | |
357 | *diff_iovecs -= lv->lv_niovecs; | |
110dc24a | 358 | *diff_len -= lv->lv_bytes; |
b1c5ebb2 DC |
359 | |
360 | /* Ensure the lv is set up according to ->iop_size */ | |
361 | lv->lv_niovecs = shadow->lv_niovecs; | |
362 | ||
363 | /* reset the lv buffer information for new formatting */ | |
364 | lv->lv_buf_len = 0; | |
365 | lv->lv_bytes = 0; | |
366 | lv->lv_buf = (char *)lv + | |
367 | xlog_cil_iovec_space(lv->lv_niovecs); | |
9597df6b | 368 | } else { |
b1c5ebb2 DC |
369 | /* switch to shadow buffer! */ |
370 | lv = shadow; | |
9597df6b | 371 | lv->lv_item = lip; |
9597df6b CH |
372 | if (ordered) { |
373 | /* track as an ordered logvec */ | |
374 | ASSERT(lip->li_lv == NULL); | |
9597df6b CH |
375 | goto insert; |
376 | } | |
f5baac35 DC |
377 | } |
378 | ||
3895e51f | 379 | ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t))); |
bde7cff6 | 380 | lip->li_ops->iop_format(lip, lv); |
7492c5b4 | 381 | insert: |
991aaf65 | 382 | xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs); |
3b93c7aa | 383 | } |
d1583a38 DC |
384 | } |
385 | ||
386 | /* | |
387 | * Insert the log items into the CIL and calculate the difference in space | |
388 | * consumed by the item. Add the space to the checkpoint ticket and calculate | |
389 | * if the change requires additional log metadata. If it does, take that space | |
42b2aa86 | 390 | * as well. Remove the amount of space we added to the checkpoint ticket from |
d1583a38 DC |
391 | * the current transaction ticket so that the accounting works out correctly. |
392 | */ | |
3b93c7aa DC |
393 | static void |
394 | xlog_cil_insert_items( | |
f7bdf03a | 395 | struct xlog *log, |
991aaf65 | 396 | struct xfs_trans *tp) |
3b93c7aa | 397 | { |
d1583a38 DC |
398 | struct xfs_cil *cil = log->l_cilp; |
399 | struct xfs_cil_ctx *ctx = cil->xc_ctx; | |
e6631f85 | 400 | struct xfs_log_item *lip; |
d1583a38 DC |
401 | int len = 0; |
402 | int diff_iovecs = 0; | |
403 | int iclog_space; | |
e2f23426 | 404 | int iovhdr_res = 0, split_res = 0, ctx_res = 0; |
3b93c7aa | 405 | |
991aaf65 | 406 | ASSERT(tp); |
d1583a38 DC |
407 | |
408 | /* | |
d1583a38 DC |
409 | * We can do this safely because the context can't checkpoint until we |
410 | * are done so it doesn't matter exactly how we update the CIL. | |
411 | */ | |
991aaf65 DC |
412 | xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs); |
413 | ||
d1583a38 | 414 | spin_lock(&cil->xc_cil_lock); |
d1583a38 | 415 | |
fd63875c | 416 | /* account for space used by new iovec headers */ |
e2f23426 BF |
417 | iovhdr_res = diff_iovecs * sizeof(xlog_op_header_t); |
418 | len += iovhdr_res; | |
d1583a38 DC |
419 | ctx->nvecs += diff_iovecs; |
420 | ||
991aaf65 DC |
421 | /* attach the transaction to the CIL if it has any busy extents */ |
422 | if (!list_empty(&tp->t_busy)) | |
423 | list_splice_init(&tp->t_busy, &ctx->busy_extents); | |
424 | ||
d1583a38 DC |
425 | /* |
426 | * Now transfer enough transaction reservation to the context ticket | |
427 | * for the checkpoint. The context ticket is special - the unit | |
428 | * reservation has to grow as well as the current reservation as we | |
429 | * steal from tickets so we can correctly determine the space used | |
430 | * during the transaction commit. | |
431 | */ | |
432 | if (ctx->ticket->t_curr_res == 0) { | |
e2f23426 BF |
433 | ctx_res = ctx->ticket->t_unit_res; |
434 | ctx->ticket->t_curr_res = ctx_res; | |
435 | tp->t_ticket->t_curr_res -= ctx_res; | |
d1583a38 DC |
436 | } |
437 | ||
438 | /* do we need space for more log record headers? */ | |
439 | iclog_space = log->l_iclog_size - log->l_iclog_hsize; | |
440 | if (len > 0 && (ctx->space_used / iclog_space != | |
441 | (ctx->space_used + len) / iclog_space)) { | |
e2f23426 | 442 | split_res = (len + iclog_space - 1) / iclog_space; |
d1583a38 | 443 | /* need to take into account split region headers, too */ |
e2f23426 BF |
444 | split_res *= log->l_iclog_hsize + sizeof(struct xlog_op_header); |
445 | ctx->ticket->t_unit_res += split_res; | |
446 | ctx->ticket->t_curr_res += split_res; | |
447 | tp->t_ticket->t_curr_res -= split_res; | |
991aaf65 | 448 | ASSERT(tp->t_ticket->t_curr_res >= len); |
d1583a38 | 449 | } |
991aaf65 | 450 | tp->t_ticket->t_curr_res -= len; |
d1583a38 DC |
451 | ctx->space_used += len; |
452 | ||
d4ca1d55 BF |
453 | /* |
454 | * If we've overrun the reservation, dump the tx details before we move | |
455 | * the log items. Shutdown is imminent... | |
456 | */ | |
457 | if (WARN_ON(tp->t_ticket->t_curr_res < 0)) { | |
458 | xfs_warn(log->l_mp, "Transaction log reservation overrun:"); | |
459 | xfs_warn(log->l_mp, | |
460 | " log items: %d bytes (iov hdrs: %d bytes)", | |
461 | len, iovhdr_res); | |
462 | xfs_warn(log->l_mp, " split region headers: %d bytes", | |
463 | split_res); | |
464 | xfs_warn(log->l_mp, " ctx ticket: %d bytes", ctx_res); | |
465 | xlog_print_trans(tp); | |
466 | } | |
467 | ||
e2f23426 BF |
468 | /* |
469 | * Now (re-)position everything modified at the tail of the CIL. | |
470 | * We do this here so we only need to take the CIL lock once during | |
471 | * the transaction commit. | |
472 | */ | |
e6631f85 | 473 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
e2f23426 BF |
474 | |
475 | /* Skip items which aren't dirty in this transaction. */ | |
e6631f85 | 476 | if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) |
e2f23426 BF |
477 | continue; |
478 | ||
479 | /* | |
480 | * Only move the item if it isn't already at the tail. This is | |
481 | * to prevent a transient list_empty() state when reinserting | |
482 | * an item that is already the only item in the CIL. | |
483 | */ | |
484 | if (!list_is_last(&lip->li_cil, &cil->xc_cil)) | |
485 | list_move_tail(&lip->li_cil, &cil->xc_cil); | |
486 | } | |
487 | ||
d1583a38 | 488 | spin_unlock(&cil->xc_cil_lock); |
d4ca1d55 BF |
489 | |
490 | if (tp->t_ticket->t_curr_res < 0) | |
491 | xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR); | |
71e330b5 DC |
492 | } |
493 | ||
494 | static void | |
495 | xlog_cil_free_logvec( | |
496 | struct xfs_log_vec *log_vector) | |
497 | { | |
498 | struct xfs_log_vec *lv; | |
499 | ||
500 | for (lv = log_vector; lv; ) { | |
501 | struct xfs_log_vec *next = lv->lv_next; | |
71e330b5 DC |
502 | kmem_free(lv); |
503 | lv = next; | |
504 | } | |
505 | } | |
506 | ||
4560e78f CH |
507 | static void |
508 | xlog_discard_endio_work( | |
509 | struct work_struct *work) | |
510 | { | |
511 | struct xfs_cil_ctx *ctx = | |
512 | container_of(work, struct xfs_cil_ctx, discard_endio_work); | |
513 | struct xfs_mount *mp = ctx->cil->xc_log->l_mp; | |
514 | ||
515 | xfs_extent_busy_clear(mp, &ctx->busy_extents, false); | |
516 | kmem_free(ctx); | |
517 | } | |
518 | ||
519 | /* | |
520 | * Queue up the actual completion to a thread to avoid IRQ-safe locking for | |
521 | * pagb_lock. Note that we need a unbounded workqueue, otherwise we might | |
522 | * get the execution delayed up to 30 seconds for weird reasons. | |
523 | */ | |
524 | static void | |
525 | xlog_discard_endio( | |
526 | struct bio *bio) | |
527 | { | |
528 | struct xfs_cil_ctx *ctx = bio->bi_private; | |
529 | ||
530 | INIT_WORK(&ctx->discard_endio_work, xlog_discard_endio_work); | |
531 | queue_work(xfs_discard_wq, &ctx->discard_endio_work); | |
ea7bd56f | 532 | bio_put(bio); |
4560e78f CH |
533 | } |
534 | ||
535 | static void | |
536 | xlog_discard_busy_extents( | |
537 | struct xfs_mount *mp, | |
538 | struct xfs_cil_ctx *ctx) | |
539 | { | |
540 | struct list_head *list = &ctx->busy_extents; | |
541 | struct xfs_extent_busy *busyp; | |
542 | struct bio *bio = NULL; | |
543 | struct blk_plug plug; | |
544 | int error = 0; | |
545 | ||
546 | ASSERT(mp->m_flags & XFS_MOUNT_DISCARD); | |
547 | ||
548 | blk_start_plug(&plug); | |
549 | list_for_each_entry(busyp, list, list) { | |
550 | trace_xfs_discard_extent(mp, busyp->agno, busyp->bno, | |
551 | busyp->length); | |
552 | ||
553 | error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev, | |
554 | XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno), | |
555 | XFS_FSB_TO_BB(mp, busyp->length), | |
556 | GFP_NOFS, 0, &bio); | |
557 | if (error && error != -EOPNOTSUPP) { | |
558 | xfs_info(mp, | |
559 | "discard failed for extent [0x%llx,%u], error %d", | |
560 | (unsigned long long)busyp->bno, | |
561 | busyp->length, | |
562 | error); | |
563 | break; | |
564 | } | |
565 | } | |
566 | ||
567 | if (bio) { | |
568 | bio->bi_private = ctx; | |
569 | bio->bi_end_io = xlog_discard_endio; | |
570 | submit_bio(bio); | |
571 | } else { | |
572 | xlog_discard_endio_work(&ctx->discard_endio_work); | |
573 | } | |
574 | blk_finish_plug(&plug); | |
575 | } | |
576 | ||
71e330b5 DC |
577 | /* |
578 | * Mark all items committed and clear busy extents. We free the log vector | |
579 | * chains in a separate pass so that we unpin the log items as quickly as | |
580 | * possible. | |
581 | */ | |
582 | static void | |
583 | xlog_cil_committed( | |
12e6a0f4 | 584 | struct xfs_cil_ctx *ctx) |
71e330b5 | 585 | { |
e84661aa | 586 | struct xfs_mount *mp = ctx->cil->xc_log->l_mp; |
2039a272 | 587 | bool abort = xlog_is_shutdown(ctx->cil->xc_log); |
71e330b5 | 588 | |
545aa41f BF |
589 | /* |
590 | * If the I/O failed, we're aborting the commit and already shutdown. | |
591 | * Wake any commit waiters before aborting the log items so we don't | |
592 | * block async log pushers on callbacks. Async log pushers explicitly do | |
593 | * not wait on log force completion because they may be holding locks | |
594 | * required to unpin items. | |
595 | */ | |
596 | if (abort) { | |
597 | spin_lock(&ctx->cil->xc_push_lock); | |
598 | wake_up_all(&ctx->cil->xc_commit_wait); | |
599 | spin_unlock(&ctx->cil->xc_push_lock); | |
600 | } | |
601 | ||
0e57f6a3 DC |
602 | xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain, |
603 | ctx->start_lsn, abort); | |
71e330b5 | 604 | |
4ecbfe63 DC |
605 | xfs_extent_busy_sort(&ctx->busy_extents); |
606 | xfs_extent_busy_clear(mp, &ctx->busy_extents, | |
e84661aa | 607 | (mp->m_flags & XFS_MOUNT_DISCARD) && !abort); |
71e330b5 | 608 | |
4bb928cd | 609 | spin_lock(&ctx->cil->xc_push_lock); |
71e330b5 | 610 | list_del(&ctx->committing); |
4bb928cd | 611 | spin_unlock(&ctx->cil->xc_push_lock); |
71e330b5 DC |
612 | |
613 | xlog_cil_free_logvec(ctx->lv_chain); | |
e84661aa | 614 | |
4560e78f CH |
615 | if (!list_empty(&ctx->busy_extents)) |
616 | xlog_discard_busy_extents(mp, ctx); | |
617 | else | |
618 | kmem_free(ctx); | |
71e330b5 DC |
619 | } |
620 | ||
89ae379d CH |
621 | void |
622 | xlog_cil_process_committed( | |
12e6a0f4 | 623 | struct list_head *list) |
89ae379d CH |
624 | { |
625 | struct xfs_cil_ctx *ctx; | |
626 | ||
627 | while ((ctx = list_first_entry_or_null(list, | |
628 | struct xfs_cil_ctx, iclog_entry))) { | |
629 | list_del(&ctx->iclog_entry); | |
12e6a0f4 | 630 | xlog_cil_committed(ctx); |
89ae379d CH |
631 | } |
632 | } | |
633 | ||
71e330b5 | 634 | /* |
c7cc296d CH |
635 | * Push the Committed Item List to the log. |
636 | * | |
637 | * If the current sequence is the same as xc_push_seq we need to do a flush. If | |
638 | * xc_push_seq is less than the current sequence, then it has already been | |
a44f13ed DC |
639 | * flushed and we don't need to do anything - the caller will wait for it to |
640 | * complete if necessary. | |
641 | * | |
c7cc296d CH |
642 | * xc_push_seq is checked unlocked against the sequence number for a match. |
643 | * Hence we can allow log forces to run racily and not issue pushes for the | |
644 | * same sequence twice. If we get a race between multiple pushes for the same | |
645 | * sequence they will block on the first one and then abort, hence avoiding | |
646 | * needless pushes. | |
71e330b5 | 647 | */ |
c7cc296d CH |
648 | static void |
649 | xlog_cil_push_work( | |
650 | struct work_struct *work) | |
71e330b5 | 651 | { |
c7cc296d CH |
652 | struct xfs_cil *cil = |
653 | container_of(work, struct xfs_cil, xc_push_work); | |
654 | struct xlog *log = cil->xc_log; | |
71e330b5 DC |
655 | struct xfs_log_vec *lv; |
656 | struct xfs_cil_ctx *ctx; | |
657 | struct xfs_cil_ctx *new_ctx; | |
658 | struct xlog_in_core *commit_iclog; | |
659 | struct xlog_ticket *tic; | |
71e330b5 | 660 | int num_iovecs; |
71e330b5 DC |
661 | int error = 0; |
662 | struct xfs_trans_header thdr; | |
663 | struct xfs_log_iovec lhdr; | |
664 | struct xfs_log_vec lvhdr = { NULL }; | |
0dc8f7f1 | 665 | xfs_lsn_t preflush_tail_lsn; |
71e330b5 | 666 | xfs_lsn_t commit_lsn; |
0dc8f7f1 | 667 | xfs_csn_t push_seq; |
bad77c37 DC |
668 | struct bio bio; |
669 | DECLARE_COMPLETION_ONSTACK(bdev_flush); | |
71e330b5 | 670 | |
707e0dda | 671 | new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_NOFS); |
71e330b5 DC |
672 | new_ctx->ticket = xlog_cil_ticket_alloc(log); |
673 | ||
4c2d542f | 674 | down_write(&cil->xc_ctx_lock); |
71e330b5 DC |
675 | ctx = cil->xc_ctx; |
676 | ||
4bb928cd | 677 | spin_lock(&cil->xc_push_lock); |
4c2d542f DC |
678 | push_seq = cil->xc_push_seq; |
679 | ASSERT(push_seq <= ctx->sequence); | |
71e330b5 | 680 | |
0e7ab7ef | 681 | /* |
19f4e7cc DC |
682 | * As we are about to switch to a new, empty CIL context, we no longer |
683 | * need to throttle tasks on CIL space overruns. Wake any waiters that | |
684 | * the hard push throttle may have caught so they can start committing | |
685 | * to the new context. The ctx->xc_push_lock provides the serialisation | |
686 | * necessary for safely using the lockless waitqueue_active() check in | |
687 | * this context. | |
0e7ab7ef | 688 | */ |
19f4e7cc | 689 | if (waitqueue_active(&cil->xc_push_wait)) |
c7f87f39 | 690 | wake_up_all(&cil->xc_push_wait); |
0e7ab7ef | 691 | |
4c2d542f DC |
692 | /* |
693 | * Check if we've anything to push. If there is nothing, then we don't | |
694 | * move on to a new sequence number and so we have to be able to push | |
695 | * this sequence again later. | |
696 | */ | |
697 | if (list_empty(&cil->xc_cil)) { | |
698 | cil->xc_push_seq = 0; | |
4bb928cd | 699 | spin_unlock(&cil->xc_push_lock); |
a44f13ed | 700 | goto out_skip; |
4c2d542f | 701 | } |
4c2d542f | 702 | |
a44f13ed | 703 | |
cf085a1b | 704 | /* check for a previously pushed sequence */ |
8af3dcd3 DC |
705 | if (push_seq < cil->xc_ctx->sequence) { |
706 | spin_unlock(&cil->xc_push_lock); | |
df806158 | 707 | goto out_skip; |
8af3dcd3 DC |
708 | } |
709 | ||
710 | /* | |
711 | * We are now going to push this context, so add it to the committing | |
712 | * list before we do anything else. This ensures that anyone waiting on | |
713 | * this push can easily detect the difference between a "push in | |
714 | * progress" and "CIL is empty, nothing to do". | |
715 | * | |
716 | * IOWs, a wait loop can now check for: | |
717 | * the current sequence not being found on the committing list; | |
718 | * an empty CIL; and | |
719 | * an unchanged sequence number | |
720 | * to detect a push that had nothing to do and therefore does not need | |
721 | * waiting on. If the CIL is not empty, we get put on the committing | |
722 | * list before emptying the CIL and bumping the sequence number. Hence | |
723 | * an empty CIL and an unchanged sequence number means we jumped out | |
724 | * above after doing nothing. | |
725 | * | |
726 | * Hence the waiter will either find the commit sequence on the | |
727 | * committing list or the sequence number will be unchanged and the CIL | |
728 | * still dirty. In that latter case, the push has not yet started, and | |
729 | * so the waiter will have to continue trying to check the CIL | |
730 | * committing list until it is found. In extreme cases of delay, the | |
731 | * sequence may fully commit between the attempts the wait makes to wait | |
732 | * on the commit sequence. | |
733 | */ | |
734 | list_add(&ctx->committing, &cil->xc_committing); | |
735 | spin_unlock(&cil->xc_push_lock); | |
df806158 | 736 | |
71e330b5 | 737 | /* |
bad77c37 DC |
738 | * The CIL is stable at this point - nothing new will be added to it |
739 | * because we hold the flush lock exclusively. Hence we can now issue | |
740 | * a cache flush to ensure all the completed metadata in the journal we | |
741 | * are about to overwrite is on stable storage. | |
0dc8f7f1 DC |
742 | * |
743 | * Because we are issuing this cache flush before we've written the | |
744 | * tail lsn to the iclog, we can have metadata IO completions move the | |
745 | * tail forwards between the completion of this flush and the iclog | |
746 | * being written. In this case, we need to re-issue the cache flush | |
747 | * before the iclog write. To detect whether the log tail moves, sample | |
748 | * the tail LSN *before* we issue the flush. | |
bad77c37 | 749 | */ |
0dc8f7f1 | 750 | preflush_tail_lsn = atomic64_read(&log->l_tail_lsn); |
bad77c37 DC |
751 | xfs_flush_bdev_async(&bio, log->l_mp->m_ddev_targp->bt_bdev, |
752 | &bdev_flush); | |
753 | ||
754 | /* | |
755 | * Pull all the log vectors off the items in the CIL, and remove the | |
756 | * items from the CIL. We don't need the CIL lock here because it's only | |
757 | * needed on the transaction commit side which is currently locked out | |
758 | * by the flush lock. | |
71e330b5 DC |
759 | */ |
760 | lv = NULL; | |
71e330b5 | 761 | num_iovecs = 0; |
71e330b5 DC |
762 | while (!list_empty(&cil->xc_cil)) { |
763 | struct xfs_log_item *item; | |
71e330b5 DC |
764 | |
765 | item = list_first_entry(&cil->xc_cil, | |
766 | struct xfs_log_item, li_cil); | |
767 | list_del_init(&item->li_cil); | |
768 | if (!ctx->lv_chain) | |
769 | ctx->lv_chain = item->li_lv; | |
770 | else | |
771 | lv->lv_next = item->li_lv; | |
772 | lv = item->li_lv; | |
773 | item->li_lv = NULL; | |
71e330b5 | 774 | num_iovecs += lv->lv_niovecs; |
71e330b5 DC |
775 | } |
776 | ||
777 | /* | |
778 | * initialise the new context and attach it to the CIL. Then attach | |
c7f87f39 | 779 | * the current context to the CIL committing list so it can be found |
71e330b5 DC |
780 | * during log forces to extract the commit lsn of the sequence that |
781 | * needs to be forced. | |
782 | */ | |
783 | INIT_LIST_HEAD(&new_ctx->committing); | |
784 | INIT_LIST_HEAD(&new_ctx->busy_extents); | |
785 | new_ctx->sequence = ctx->sequence + 1; | |
786 | new_ctx->cil = cil; | |
787 | cil->xc_ctx = new_ctx; | |
788 | ||
789 | /* | |
790 | * The switch is now done, so we can drop the context lock and move out | |
791 | * of a shared context. We can't just go straight to the commit record, | |
792 | * though - we need to synchronise with previous and future commits so | |
793 | * that the commit records are correctly ordered in the log to ensure | |
794 | * that we process items during log IO completion in the correct order. | |
795 | * | |
796 | * For example, if we get an EFI in one checkpoint and the EFD in the | |
797 | * next (e.g. due to log forces), we do not want the checkpoint with | |
798 | * the EFD to be committed before the checkpoint with the EFI. Hence | |
799 | * we must strictly order the commit records of the checkpoints so | |
800 | * that: a) the checkpoint callbacks are attached to the iclogs in the | |
801 | * correct order; and b) the checkpoints are replayed in correct order | |
802 | * in log recovery. | |
803 | * | |
804 | * Hence we need to add this context to the committing context list so | |
805 | * that higher sequences will wait for us to write out a commit record | |
806 | * before they do. | |
f876e446 | 807 | * |
5f9b4b0d | 808 | * xfs_log_force_seq requires us to mirror the new sequence into the cil |
f876e446 DC |
809 | * structure atomically with the addition of this sequence to the |
810 | * committing list. This also ensures that we can do unlocked checks | |
811 | * against the current sequence in log forces without risking | |
812 | * deferencing a freed context pointer. | |
71e330b5 | 813 | */ |
4bb928cd | 814 | spin_lock(&cil->xc_push_lock); |
f876e446 | 815 | cil->xc_current_sequence = new_ctx->sequence; |
4bb928cd | 816 | spin_unlock(&cil->xc_push_lock); |
71e330b5 DC |
817 | up_write(&cil->xc_ctx_lock); |
818 | ||
819 | /* | |
820 | * Build a checkpoint transaction header and write it to the log to | |
821 | * begin the transaction. We need to account for the space used by the | |
822 | * transaction header here as it is not accounted for in xlog_write(). | |
823 | * | |
824 | * The LSN we need to pass to the log items on transaction commit is | |
825 | * the LSN reported by the first log vector write. If we use the commit | |
826 | * record lsn then we can move the tail beyond the grant write head. | |
827 | */ | |
828 | tic = ctx->ticket; | |
829 | thdr.th_magic = XFS_TRANS_HEADER_MAGIC; | |
830 | thdr.th_type = XFS_TRANS_CHECKPOINT; | |
831 | thdr.th_tid = tic->t_tid; | |
832 | thdr.th_num_items = num_iovecs; | |
4e0d5f92 | 833 | lhdr.i_addr = &thdr; |
71e330b5 DC |
834 | lhdr.i_len = sizeof(xfs_trans_header_t); |
835 | lhdr.i_type = XLOG_REG_TYPE_TRANSHDR; | |
836 | tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t); | |
837 | ||
838 | lvhdr.lv_niovecs = 1; | |
839 | lvhdr.lv_iovecp = &lhdr; | |
840 | lvhdr.lv_next = ctx->lv_chain; | |
841 | ||
bad77c37 DC |
842 | /* |
843 | * Before we format and submit the first iclog, we have to ensure that | |
844 | * the metadata writeback ordering cache flush is complete. | |
845 | */ | |
846 | wait_for_completion(&bdev_flush); | |
847 | ||
3468bb1c DC |
848 | error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, |
849 | XLOG_START_TRANS); | |
71e330b5 | 850 | if (error) |
7db37c5e | 851 | goto out_abort_free_ticket; |
71e330b5 DC |
852 | |
853 | /* | |
854 | * now that we've written the checkpoint into the log, strictly | |
855 | * order the commit records so replay will get them in the right order. | |
856 | */ | |
857 | restart: | |
4bb928cd | 858 | spin_lock(&cil->xc_push_lock); |
71e330b5 | 859 | list_for_each_entry(new_ctx, &cil->xc_committing, committing) { |
ac983517 DC |
860 | /* |
861 | * Avoid getting stuck in this loop because we were woken by the | |
862 | * shutdown, but then went back to sleep once already in the | |
863 | * shutdown state. | |
864 | */ | |
2039a272 | 865 | if (xlog_is_shutdown(log)) { |
ac983517 DC |
866 | spin_unlock(&cil->xc_push_lock); |
867 | goto out_abort_free_ticket; | |
868 | } | |
869 | ||
71e330b5 DC |
870 | /* |
871 | * Higher sequences will wait for this one so skip them. | |
ac983517 | 872 | * Don't wait for our own sequence, either. |
71e330b5 DC |
873 | */ |
874 | if (new_ctx->sequence >= ctx->sequence) | |
875 | continue; | |
876 | if (!new_ctx->commit_lsn) { | |
877 | /* | |
878 | * It is still being pushed! Wait for the push to | |
879 | * complete, then start again from the beginning. | |
880 | */ | |
4bb928cd | 881 | xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock); |
71e330b5 DC |
882 | goto restart; |
883 | } | |
884 | } | |
4bb928cd | 885 | spin_unlock(&cil->xc_push_lock); |
71e330b5 | 886 | |
f10e925d | 887 | error = xlog_commit_record(log, tic, &commit_iclog, &commit_lsn); |
dd401770 DC |
888 | if (error) |
889 | goto out_abort_free_ticket; | |
890 | ||
8b41e3f9 | 891 | xfs_log_ticket_ungrant(log, tic); |
71e330b5 | 892 | |
a1bb8505 | 893 | /* |
502a01fa DC |
894 | * Once we attach the ctx to the iclog, it is effectively owned by the |
895 | * iclog and we can only use it while we still have an active reference | |
896 | * to the iclog. i.e. once we call xlog_state_release_iclog() we can no | |
897 | * longer safely reference the ctx. | |
a1bb8505 DC |
898 | */ |
899 | spin_lock(&log->l_icloglock); | |
5112e206 | 900 | if (xlog_is_shutdown(log)) { |
a1bb8505 | 901 | spin_unlock(&log->l_icloglock); |
71e330b5 | 902 | goto out_abort; |
89ae379d CH |
903 | } |
904 | ASSERT_ALWAYS(commit_iclog->ic_state == XLOG_STATE_ACTIVE || | |
905 | commit_iclog->ic_state == XLOG_STATE_WANT_SYNC); | |
906 | list_add_tail(&ctx->iclog_entry, &commit_iclog->ic_callbacks); | |
71e330b5 DC |
907 | |
908 | /* | |
909 | * now the checkpoint commit is complete and we've attached the | |
910 | * callbacks to the iclog we can assign the commit LSN to the context | |
911 | * and wake up anyone who is waiting for the commit to complete. | |
912 | */ | |
4bb928cd | 913 | spin_lock(&cil->xc_push_lock); |
71e330b5 | 914 | ctx->commit_lsn = commit_lsn; |
eb40a875 | 915 | wake_up_all(&cil->xc_commit_wait); |
4bb928cd | 916 | spin_unlock(&cil->xc_push_lock); |
71e330b5 | 917 | |
a79b28c2 | 918 | /* |
1effb72a DC |
919 | * If the checkpoint spans multiple iclogs, wait for all previous iclogs |
920 | * to complete before we submit the commit_iclog. We can't use state | |
921 | * checks for this - ACTIVE can be either a past completed iclog or a | |
922 | * future iclog being filled, while WANT_SYNC through SYNC_DONE can be a | |
923 | * past or future iclog awaiting IO or ordered IO completion to be run. | |
924 | * In the latter case, if it's a future iclog and we wait on it, the we | |
925 | * will hang because it won't get processed through to ic_force_wait | |
926 | * wakeup until this commit_iclog is written to disk. Hence we use the | |
927 | * iclog header lsn and compare it to the commit lsn to determine if we | |
928 | * need to wait on iclogs or not. | |
a79b28c2 DC |
929 | */ |
930 | if (ctx->start_lsn != commit_lsn) { | |
1effb72a DC |
931 | xfs_lsn_t plsn; |
932 | ||
933 | plsn = be64_to_cpu(commit_iclog->ic_prev->ic_header.h_lsn); | |
934 | if (plsn && XFS_LSN_CMP(plsn, commit_lsn) < 0) { | |
935 | /* | |
936 | * Waiting on ic_force_wait orders the completion of | |
937 | * iclogs older than ic_prev. Hence we only need to wait | |
938 | * on the most recent older iclog here. | |
939 | */ | |
940 | xlog_wait_on_iclog(commit_iclog->ic_prev); | |
941 | spin_lock(&log->l_icloglock); | |
942 | } | |
943 | ||
944 | /* | |
945 | * We need to issue a pre-flush so that the ordering for this | |
946 | * checkpoint is correctly preserved down to stable storage. | |
947 | */ | |
eef983ff | 948 | commit_iclog->ic_flags |= XLOG_ICL_NEED_FLUSH; |
a79b28c2 DC |
949 | } |
950 | ||
eef983ff DC |
951 | /* |
952 | * The commit iclog must be written to stable storage to guarantee | |
953 | * journal IO vs metadata writeback IO is correctly ordered on stable | |
954 | * storage. | |
955 | */ | |
956 | commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA; | |
0dc8f7f1 | 957 | xlog_state_release_iclog(log, commit_iclog, preflush_tail_lsn); |
502a01fa DC |
958 | |
959 | /* Not safe to reference ctx now! */ | |
960 | ||
eef983ff | 961 | spin_unlock(&log->l_icloglock); |
c7cc296d | 962 | return; |
71e330b5 DC |
963 | |
964 | out_skip: | |
965 | up_write(&cil->xc_ctx_lock); | |
966 | xfs_log_ticket_put(new_ctx->ticket); | |
967 | kmem_free(new_ctx); | |
c7cc296d | 968 | return; |
71e330b5 | 969 | |
7db37c5e | 970 | out_abort_free_ticket: |
8b41e3f9 | 971 | xfs_log_ticket_ungrant(log, tic); |
71e330b5 | 972 | out_abort: |
2039a272 | 973 | ASSERT(xlog_is_shutdown(log)); |
12e6a0f4 | 974 | xlog_cil_committed(ctx); |
4c2d542f DC |
975 | } |
976 | ||
977 | /* | |
978 | * We need to push CIL every so often so we don't cache more than we can fit in | |
979 | * the log. The limit really is that a checkpoint can't be more than half the | |
980 | * log (the current checkpoint is not allowed to overwrite the previous | |
981 | * checkpoint), but commit latency and memory usage limit this to a smaller | |
982 | * size. | |
983 | */ | |
984 | static void | |
985 | xlog_cil_push_background( | |
0e7ab7ef | 986 | struct xlog *log) __releases(cil->xc_ctx_lock) |
4c2d542f DC |
987 | { |
988 | struct xfs_cil *cil = log->l_cilp; | |
989 | ||
990 | /* | |
991 | * The cil won't be empty because we are called while holding the | |
992 | * context lock so whatever we added to the CIL will still be there | |
993 | */ | |
994 | ASSERT(!list_empty(&cil->xc_cil)); | |
995 | ||
996 | /* | |
19f4e7cc | 997 | * Don't do a background push if we haven't used up all the |
4c2d542f DC |
998 | * space available yet. |
999 | */ | |
0e7ab7ef DC |
1000 | if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log)) { |
1001 | up_read(&cil->xc_ctx_lock); | |
4c2d542f | 1002 | return; |
0e7ab7ef | 1003 | } |
4c2d542f | 1004 | |
4bb928cd | 1005 | spin_lock(&cil->xc_push_lock); |
4c2d542f DC |
1006 | if (cil->xc_push_seq < cil->xc_current_sequence) { |
1007 | cil->xc_push_seq = cil->xc_current_sequence; | |
1008 | queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work); | |
1009 | } | |
0e7ab7ef DC |
1010 | |
1011 | /* | |
1012 | * Drop the context lock now, we can't hold that if we need to sleep | |
1013 | * because we are over the blocking threshold. The push_lock is still | |
1014 | * held, so blocking threshold sleep/wakeup is still correctly | |
1015 | * serialised here. | |
1016 | */ | |
1017 | up_read(&cil->xc_ctx_lock); | |
1018 | ||
1019 | /* | |
1020 | * If we are well over the space limit, throttle the work that is being | |
19f4e7cc DC |
1021 | * done until the push work on this context has begun. Enforce the hard |
1022 | * throttle on all transaction commits once it has been activated, even | |
1023 | * if the committing transactions have resulted in the space usage | |
1024 | * dipping back down under the hard limit. | |
1025 | * | |
1026 | * The ctx->xc_push_lock provides the serialisation necessary for safely | |
1027 | * using the lockless waitqueue_active() check in this context. | |
0e7ab7ef | 1028 | */ |
19f4e7cc DC |
1029 | if (cil->xc_ctx->space_used >= XLOG_CIL_BLOCKING_SPACE_LIMIT(log) || |
1030 | waitqueue_active(&cil->xc_push_wait)) { | |
0e7ab7ef DC |
1031 | trace_xfs_log_cil_wait(log, cil->xc_ctx->ticket); |
1032 | ASSERT(cil->xc_ctx->space_used < log->l_logsize); | |
c7f87f39 | 1033 | xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock); |
0e7ab7ef DC |
1034 | return; |
1035 | } | |
1036 | ||
4bb928cd | 1037 | spin_unlock(&cil->xc_push_lock); |
4c2d542f DC |
1038 | |
1039 | } | |
1040 | ||
f876e446 DC |
1041 | /* |
1042 | * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence | |
1043 | * number that is passed. When it returns, the work will be queued for | |
1044 | * @push_seq, but it won't be completed. The caller is expected to do any | |
1045 | * waiting for push_seq to complete if it is required. | |
1046 | */ | |
4c2d542f | 1047 | static void |
f876e446 | 1048 | xlog_cil_push_now( |
f7bdf03a | 1049 | struct xlog *log, |
4c2d542f DC |
1050 | xfs_lsn_t push_seq) |
1051 | { | |
1052 | struct xfs_cil *cil = log->l_cilp; | |
1053 | ||
1054 | if (!cil) | |
1055 | return; | |
1056 | ||
1057 | ASSERT(push_seq && push_seq <= cil->xc_current_sequence); | |
1058 | ||
1059 | /* start on any pending background push to minimise wait time on it */ | |
1060 | flush_work(&cil->xc_push_work); | |
1061 | ||
1062 | /* | |
1063 | * If the CIL is empty or we've already pushed the sequence then | |
1064 | * there's no work we need to do. | |
1065 | */ | |
4bb928cd | 1066 | spin_lock(&cil->xc_push_lock); |
4c2d542f | 1067 | if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) { |
4bb928cd | 1068 | spin_unlock(&cil->xc_push_lock); |
4c2d542f DC |
1069 | return; |
1070 | } | |
1071 | ||
1072 | cil->xc_push_seq = push_seq; | |
f876e446 | 1073 | queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work); |
4bb928cd | 1074 | spin_unlock(&cil->xc_push_lock); |
4c2d542f DC |
1075 | } |
1076 | ||
2c6e24ce DC |
1077 | bool |
1078 | xlog_cil_empty( | |
1079 | struct xlog *log) | |
1080 | { | |
1081 | struct xfs_cil *cil = log->l_cilp; | |
1082 | bool empty = false; | |
1083 | ||
1084 | spin_lock(&cil->xc_push_lock); | |
1085 | if (list_empty(&cil->xc_cil)) | |
1086 | empty = true; | |
1087 | spin_unlock(&cil->xc_push_lock); | |
1088 | return empty; | |
1089 | } | |
1090 | ||
a44f13ed DC |
1091 | /* |
1092 | * Commit a transaction with the given vector to the Committed Item List. | |
1093 | * | |
1094 | * To do this, we need to format the item, pin it in memory if required and | |
1095 | * account for the space used by the transaction. Once we have done that we | |
1096 | * need to release the unused reservation for the transaction, attach the | |
1097 | * transaction to the checkpoint context so we carry the busy extents through | |
1098 | * to checkpoint completion, and then unlock all the items in the transaction. | |
1099 | * | |
a44f13ed DC |
1100 | * Called with the context lock already held in read mode to lock out |
1101 | * background commit, returns without it held once background commits are | |
1102 | * allowed again. | |
1103 | */ | |
c6f97264 | 1104 | void |
5f9b4b0d DC |
1105 | xlog_cil_commit( |
1106 | struct xlog *log, | |
a44f13ed | 1107 | struct xfs_trans *tp, |
5f9b4b0d | 1108 | xfs_csn_t *commit_seq, |
70393313 | 1109 | bool regrant) |
a44f13ed | 1110 | { |
991aaf65 | 1111 | struct xfs_cil *cil = log->l_cilp; |
195cd83d | 1112 | struct xfs_log_item *lip, *next; |
a44f13ed | 1113 | |
b1c5ebb2 DC |
1114 | /* |
1115 | * Do all necessary memory allocation before we lock the CIL. | |
1116 | * This ensures the allocation does not deadlock with a CIL | |
1117 | * push in memory reclaim (e.g. from kswapd). | |
1118 | */ | |
1119 | xlog_cil_alloc_shadow_bufs(log, tp); | |
1120 | ||
f5baac35 | 1121 | /* lock out background commit */ |
991aaf65 | 1122 | down_read(&cil->xc_ctx_lock); |
f5baac35 | 1123 | |
991aaf65 | 1124 | xlog_cil_insert_items(log, tp); |
a44f13ed | 1125 | |
2039a272 | 1126 | if (regrant && !xlog_is_shutdown(log)) |
8b41e3f9 CH |
1127 | xfs_log_ticket_regrant(log, tp->t_ticket); |
1128 | else | |
1129 | xfs_log_ticket_ungrant(log, tp->t_ticket); | |
ba18781b | 1130 | tp->t_ticket = NULL; |
a44f13ed DC |
1131 | xfs_trans_unreserve_and_mod_sb(tp); |
1132 | ||
1133 | /* | |
1134 | * Once all the items of the transaction have been copied to the CIL, | |
195cd83d | 1135 | * the items can be unlocked and possibly freed. |
a44f13ed DC |
1136 | * |
1137 | * This needs to be done before we drop the CIL context lock because we | |
1138 | * have to update state in the log items and unlock them before they go | |
1139 | * to disk. If we don't, then the CIL checkpoint can race with us and | |
1140 | * we can run checkpoint completion before we've updated and unlocked | |
1141 | * the log items. This affects (at least) processing of stale buffers, | |
1142 | * inodes and EFIs. | |
1143 | */ | |
195cd83d CH |
1144 | trace_xfs_trans_commit_items(tp, _RET_IP_); |
1145 | list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) { | |
1146 | xfs_trans_del_item(lip); | |
1147 | if (lip->li_ops->iop_committing) | |
5f9b4b0d | 1148 | lip->li_ops->iop_committing(lip, cil->xc_ctx->sequence); |
195cd83d | 1149 | } |
5f9b4b0d DC |
1150 | if (commit_seq) |
1151 | *commit_seq = cil->xc_ctx->sequence; | |
a44f13ed | 1152 | |
0e7ab7ef DC |
1153 | /* xlog_cil_push_background() releases cil->xc_ctx_lock */ |
1154 | xlog_cil_push_background(log); | |
a44f13ed DC |
1155 | } |
1156 | ||
71e330b5 DC |
1157 | /* |
1158 | * Conditionally push the CIL based on the sequence passed in. | |
1159 | * | |
1160 | * We only need to push if we haven't already pushed the sequence | |
1161 | * number given. Hence the only time we will trigger a push here is | |
1162 | * if the push sequence is the same as the current context. | |
1163 | * | |
1164 | * We return the current commit lsn to allow the callers to determine if a | |
1165 | * iclog flush is necessary following this call. | |
71e330b5 DC |
1166 | */ |
1167 | xfs_lsn_t | |
5f9b4b0d | 1168 | xlog_cil_force_seq( |
f7bdf03a | 1169 | struct xlog *log, |
5f9b4b0d | 1170 | xfs_csn_t sequence) |
71e330b5 DC |
1171 | { |
1172 | struct xfs_cil *cil = log->l_cilp; | |
1173 | struct xfs_cil_ctx *ctx; | |
1174 | xfs_lsn_t commit_lsn = NULLCOMMITLSN; | |
1175 | ||
a44f13ed DC |
1176 | ASSERT(sequence <= cil->xc_current_sequence); |
1177 | ||
1178 | /* | |
1179 | * check to see if we need to force out the current context. | |
1180 | * xlog_cil_push() handles racing pushes for the same sequence, | |
1181 | * so no need to deal with it here. | |
1182 | */ | |
f876e446 DC |
1183 | restart: |
1184 | xlog_cil_push_now(log, sequence); | |
71e330b5 DC |
1185 | |
1186 | /* | |
1187 | * See if we can find a previous sequence still committing. | |
71e330b5 DC |
1188 | * We need to wait for all previous sequence commits to complete |
1189 | * before allowing the force of push_seq to go ahead. Hence block | |
1190 | * on commits for those as well. | |
1191 | */ | |
4bb928cd | 1192 | spin_lock(&cil->xc_push_lock); |
71e330b5 | 1193 | list_for_each_entry(ctx, &cil->xc_committing, committing) { |
ac983517 DC |
1194 | /* |
1195 | * Avoid getting stuck in this loop because we were woken by the | |
1196 | * shutdown, but then went back to sleep once already in the | |
1197 | * shutdown state. | |
1198 | */ | |
2039a272 | 1199 | if (xlog_is_shutdown(log)) |
ac983517 | 1200 | goto out_shutdown; |
a44f13ed | 1201 | if (ctx->sequence > sequence) |
71e330b5 DC |
1202 | continue; |
1203 | if (!ctx->commit_lsn) { | |
1204 | /* | |
1205 | * It is still being pushed! Wait for the push to | |
1206 | * complete, then start again from the beginning. | |
1207 | */ | |
4bb928cd | 1208 | xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock); |
71e330b5 DC |
1209 | goto restart; |
1210 | } | |
a44f13ed | 1211 | if (ctx->sequence != sequence) |
71e330b5 DC |
1212 | continue; |
1213 | /* found it! */ | |
1214 | commit_lsn = ctx->commit_lsn; | |
1215 | } | |
f876e446 DC |
1216 | |
1217 | /* | |
1218 | * The call to xlog_cil_push_now() executes the push in the background. | |
1219 | * Hence by the time we have got here it our sequence may not have been | |
1220 | * pushed yet. This is true if the current sequence still matches the | |
1221 | * push sequence after the above wait loop and the CIL still contains | |
8af3dcd3 DC |
1222 | * dirty objects. This is guaranteed by the push code first adding the |
1223 | * context to the committing list before emptying the CIL. | |
f876e446 | 1224 | * |
8af3dcd3 DC |
1225 | * Hence if we don't find the context in the committing list and the |
1226 | * current sequence number is unchanged then the CIL contents are | |
1227 | * significant. If the CIL is empty, if means there was nothing to push | |
1228 | * and that means there is nothing to wait for. If the CIL is not empty, | |
1229 | * it means we haven't yet started the push, because if it had started | |
1230 | * we would have found the context on the committing list. | |
f876e446 | 1231 | */ |
f876e446 DC |
1232 | if (sequence == cil->xc_current_sequence && |
1233 | !list_empty(&cil->xc_cil)) { | |
1234 | spin_unlock(&cil->xc_push_lock); | |
1235 | goto restart; | |
1236 | } | |
1237 | ||
4bb928cd | 1238 | spin_unlock(&cil->xc_push_lock); |
71e330b5 | 1239 | return commit_lsn; |
ac983517 DC |
1240 | |
1241 | /* | |
1242 | * We detected a shutdown in progress. We need to trigger the log force | |
1243 | * to pass through it's iclog state machine error handling, even though | |
1244 | * we are already in a shutdown state. Hence we can't return | |
1245 | * NULLCOMMITLSN here as that has special meaning to log forces (i.e. | |
1246 | * LSN is already stable), so we return a zero LSN instead. | |
1247 | */ | |
1248 | out_shutdown: | |
1249 | spin_unlock(&cil->xc_push_lock); | |
1250 | return 0; | |
71e330b5 | 1251 | } |
ccf7c23f DC |
1252 | |
1253 | /* | |
1254 | * Check if the current log item was first committed in this sequence. | |
1255 | * We can't rely on just the log item being in the CIL, we have to check | |
1256 | * the recorded commit sequence number. | |
1257 | * | |
1258 | * Note: for this to be used in a non-racy manner, it has to be called with | |
1259 | * CIL flushing locked out. As a result, it should only be used during the | |
1260 | * transaction commit process when deciding what to format into the item. | |
1261 | */ | |
1262 | bool | |
1263 | xfs_log_item_in_current_chkpt( | |
1264 | struct xfs_log_item *lip) | |
1265 | { | |
5f9b4b0d | 1266 | struct xfs_cil_ctx *ctx = lip->li_mountp->m_log->l_cilp->xc_ctx; |
ccf7c23f | 1267 | |
ccf7c23f DC |
1268 | if (list_empty(&lip->li_cil)) |
1269 | return false; | |
1270 | ||
ccf7c23f DC |
1271 | /* |
1272 | * li_seq is written on the first commit of a log item to record the | |
1273 | * first checkpoint it is written to. Hence if it is different to the | |
1274 | * current sequence, we're in a new checkpoint. | |
1275 | */ | |
5f9b4b0d | 1276 | return lip->li_seq == ctx->sequence; |
ccf7c23f | 1277 | } |
4c2d542f DC |
1278 | |
1279 | /* | |
1280 | * Perform initial CIL structure initialisation. | |
1281 | */ | |
1282 | int | |
1283 | xlog_cil_init( | |
f7bdf03a | 1284 | struct xlog *log) |
4c2d542f DC |
1285 | { |
1286 | struct xfs_cil *cil; | |
1287 | struct xfs_cil_ctx *ctx; | |
1288 | ||
707e0dda | 1289 | cil = kmem_zalloc(sizeof(*cil), KM_MAYFAIL); |
4c2d542f | 1290 | if (!cil) |
2451337d | 1291 | return -ENOMEM; |
4c2d542f | 1292 | |
707e0dda | 1293 | ctx = kmem_zalloc(sizeof(*ctx), KM_MAYFAIL); |
4c2d542f DC |
1294 | if (!ctx) { |
1295 | kmem_free(cil); | |
2451337d | 1296 | return -ENOMEM; |
4c2d542f DC |
1297 | } |
1298 | ||
1299 | INIT_WORK(&cil->xc_push_work, xlog_cil_push_work); | |
1300 | INIT_LIST_HEAD(&cil->xc_cil); | |
1301 | INIT_LIST_HEAD(&cil->xc_committing); | |
1302 | spin_lock_init(&cil->xc_cil_lock); | |
4bb928cd | 1303 | spin_lock_init(&cil->xc_push_lock); |
c7f87f39 | 1304 | init_waitqueue_head(&cil->xc_push_wait); |
4c2d542f DC |
1305 | init_rwsem(&cil->xc_ctx_lock); |
1306 | init_waitqueue_head(&cil->xc_commit_wait); | |
1307 | ||
1308 | INIT_LIST_HEAD(&ctx->committing); | |
1309 | INIT_LIST_HEAD(&ctx->busy_extents); | |
1310 | ctx->sequence = 1; | |
1311 | ctx->cil = cil; | |
1312 | cil->xc_ctx = ctx; | |
1313 | cil->xc_current_sequence = ctx->sequence; | |
1314 | ||
1315 | cil->xc_log = log; | |
1316 | log->l_cilp = cil; | |
1317 | return 0; | |
1318 | } | |
1319 | ||
1320 | void | |
1321 | xlog_cil_destroy( | |
f7bdf03a | 1322 | struct xlog *log) |
4c2d542f DC |
1323 | { |
1324 | if (log->l_cilp->xc_ctx) { | |
1325 | if (log->l_cilp->xc_ctx->ticket) | |
1326 | xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket); | |
1327 | kmem_free(log->l_cilp->xc_ctx); | |
1328 | } | |
1329 | ||
1330 | ASSERT(list_empty(&log->l_cilp->xc_cil)); | |
1331 | kmem_free(log->l_cilp); | |
1332 | } | |
1333 |