1 /*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dmapi.h"
28 #include "xfs_mount.h"
29 #include "xfs_buf_item.h"
30 #include "xfs_trans_priv.h"
31 #include "xfs_error.h"
32
33
34 kmem_zone_t *xfs_buf_item_zone;
35
36 #ifdef XFS_TRANS_DEBUG
37 /*
38 * This function uses an alternate strategy for tracking the bytes
39 * that the user requests to be logged. This can then be used
40 * in conjunction with the bli_orig array in the buf log item to
41 * catch bugs in our callers' code.
42 *
43 * We also double check the bits set in xfs_buf_item_log using a
44 * simple algorithm to check that every byte is accounted for.
45 */
46 STATIC void
xfs_buf_item_log_debug(xfs_buf_log_item_t * bip,uint first,uint last)47 xfs_buf_item_log_debug(
48 xfs_buf_log_item_t *bip,
49 uint first,
50 uint last)
51 {
52 uint x;
53 uint byte;
54 uint nbytes;
55 uint chunk_num;
56 uint word_num;
57 uint bit_num;
58 uint bit_set;
59 uint *wordp;
60
61 ASSERT(bip->bli_logged != NULL);
62 byte = first;
63 nbytes = last - first + 1;
64 bfset(bip->bli_logged, first, nbytes);
65 for (x = 0; x < nbytes; x++) {
66 chunk_num = byte >> XFS_BLI_SHIFT;
67 word_num = chunk_num >> BIT_TO_WORD_SHIFT;
68 bit_num = chunk_num & (NBWORD - 1);
69 wordp = &(bip->bli_format.blf_data_map[word_num]);
70 bit_set = *wordp & (1 << bit_num);
71 ASSERT(bit_set);
72 byte++;
73 }
74 }
75
76 /*
77 * This function is called when we flush something into a buffer without
78 * logging it. This happens for things like inodes which are logged
79 * separately from the buffer.
80 */
81 void
xfs_buf_item_flush_log_debug(xfs_buf_t * bp,uint first,uint last)82 xfs_buf_item_flush_log_debug(
83 xfs_buf_t *bp,
84 uint first,
85 uint last)
86 {
87 xfs_buf_log_item_t *bip;
88 uint nbytes;
89
90 bip = XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t*);
91 if ((bip == NULL) || (bip->bli_item.li_type != XFS_LI_BUF)) {
92 return;
93 }
94
95 ASSERT(bip->bli_logged != NULL);
96 nbytes = last - first + 1;
97 bfset(bip->bli_logged, first, nbytes);
98 }
99
100 /*
101 * This function is called to verify that our callers have logged
102 * all the bytes that they changed.
103 *
104 * It does this by comparing the original copy of the buffer stored in
105 * the buf log item's bli_orig array to the current copy of the buffer
106 * and ensuring that all bytes which mismatch are set in the bli_logged
107 * array of the buf log item.
108 */
109 STATIC void
xfs_buf_item_log_check(xfs_buf_log_item_t * bip)110 xfs_buf_item_log_check(
111 xfs_buf_log_item_t *bip)
112 {
113 char *orig;
114 char *buffer;
115 int x;
116 xfs_buf_t *bp;
117
118 ASSERT(bip->bli_orig != NULL);
119 ASSERT(bip->bli_logged != NULL);
120
121 bp = bip->bli_buf;
122 ASSERT(XFS_BUF_COUNT(bp) > 0);
123 ASSERT(XFS_BUF_PTR(bp) != NULL);
124 orig = bip->bli_orig;
125 buffer = XFS_BUF_PTR(bp);
126 for (x = 0; x < XFS_BUF_COUNT(bp); x++) {
127 if (orig[x] != buffer[x] && !btst(bip->bli_logged, x))
128 cmn_err(CE_PANIC,
129 "xfs_buf_item_log_check bip %x buffer %x orig %x index %d",
130 bip, bp, orig, x);
131 }
132 }
133 #else
134 #define xfs_buf_item_log_debug(x,y,z)
135 #define xfs_buf_item_log_check(x)
136 #endif
137
138 STATIC void xfs_buf_error_relse(xfs_buf_t *bp);
139 STATIC void xfs_buf_do_callbacks(xfs_buf_t *bp, xfs_log_item_t *lip);
140
141 /*
142 * This returns the number of log iovecs needed to log the
143 * given buf log item.
144 *
145 * It calculates this as 1 iovec for the buf log format structure
146 * and 1 for each stretch of non-contiguous chunks to be logged.
147 * Contiguous chunks are logged in a single iovec.
148 *
149 * If the XFS_BLI_STALE flag has been set, then log nothing.
150 */
151 STATIC uint
xfs_buf_item_size(xfs_buf_log_item_t * bip)152 xfs_buf_item_size(
153 xfs_buf_log_item_t *bip)
154 {
155 uint nvecs;
156 int next_bit;
157 int last_bit;
158 xfs_buf_t *bp;
159
160 ASSERT(atomic_read(&bip->bli_refcount) > 0);
161 if (bip->bli_flags & XFS_BLI_STALE) {
162 /*
163 * The buffer is stale, so all we need to log
164 * is the buf log format structure with the
165 * cancel flag in it.
166 */
167 xfs_buf_item_trace("SIZE STALE", bip);
168 ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL);
169 return 1;
170 }
171
172 bp = bip->bli_buf;
173 ASSERT(bip->bli_flags & XFS_BLI_LOGGED);
174 nvecs = 1;
175 last_bit = xfs_next_bit(bip->bli_format.blf_data_map,
176 bip->bli_format.blf_map_size, 0);
177 ASSERT(last_bit != -1);
178 nvecs++;
179 while (last_bit != -1) {
180 /*
181 * This takes the bit number to start looking from and
182 * returns the next set bit from there. It returns -1
183 * if there are no more bits set or the start bit is
184 * beyond the end of the bitmap.
185 */
186 next_bit = xfs_next_bit(bip->bli_format.blf_data_map,
187 bip->bli_format.blf_map_size,
188 last_bit + 1);
189 /*
190 * If we run out of bits, leave the loop,
191 * else if we find a new set of bits bump the number of vecs,
192 * else keep scanning the current set of bits.
193 */
194 if (next_bit == -1) {
195 last_bit = -1;
196 } else if (next_bit != last_bit + 1) {
197 last_bit = next_bit;
198 nvecs++;
199 } else if (xfs_buf_offset(bp, next_bit * XFS_BLI_CHUNK) !=
200 (xfs_buf_offset(bp, last_bit * XFS_BLI_CHUNK) +
201 XFS_BLI_CHUNK)) {
202 last_bit = next_bit;
203 nvecs++;
204 } else {
205 last_bit++;
206 }
207 }
208
209 xfs_buf_item_trace("SIZE NORM", bip);
210 return nvecs;
211 }
212
213 /*
214 * This is called to fill in the vector of log iovecs for the
215 * given log buf item. It fills the first entry with a buf log
216 * format structure, and the rest point to contiguous chunks
217 * within the buffer.
218 */
219 STATIC void
xfs_buf_item_format(xfs_buf_log_item_t * bip,xfs_log_iovec_t * log_vector)220 xfs_buf_item_format(
221 xfs_buf_log_item_t *bip,
222 xfs_log_iovec_t *log_vector)
223 {
224 uint base_size;
225 uint nvecs;
226 xfs_log_iovec_t *vecp;
227 xfs_buf_t *bp;
228 int first_bit;
229 int last_bit;
230 int next_bit;
231 uint nbits;
232 uint buffer_offset;
233
234 ASSERT(atomic_read(&bip->bli_refcount) > 0);
235 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) ||
236 (bip->bli_flags & XFS_BLI_STALE));
237 bp = bip->bli_buf;
238 vecp = log_vector;
239
240 /*
241 * The size of the base structure is the size of the
242 * declared structure plus the space for the extra words
243 * of the bitmap. We subtract one from the map size, because
244 * the first element of the bitmap is accounted for in the
245 * size of the base structure.
246 */
247 base_size =
248 (uint)(sizeof(xfs_buf_log_format_t) +
249 ((bip->bli_format.blf_map_size - 1) * sizeof(uint)));
250 vecp->i_addr = (xfs_caddr_t)&bip->bli_format;
251 vecp->i_len = base_size;
252 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BFORMAT);
253 vecp++;
254 nvecs = 1;
255
256 if (bip->bli_flags & XFS_BLI_STALE) {
257 /*
258 * The buffer is stale, so all we need to log
259 * is the buf log format structure with the
260 * cancel flag in it.
261 */
262 xfs_buf_item_trace("FORMAT STALE", bip);
263 ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL);
264 bip->bli_format.blf_size = nvecs;
265 return;
266 }
267
268 /*
269 * Fill in an iovec for each set of contiguous chunks.
270 */
271 first_bit = xfs_next_bit(bip->bli_format.blf_data_map,
272 bip->bli_format.blf_map_size, 0);
273 ASSERT(first_bit != -1);
274 last_bit = first_bit;
275 nbits = 1;
276 for (;;) {
277 /*
278 * This takes the bit number to start looking from and
279 * returns the next set bit from there. It returns -1
280 * if there are no more bits set or the start bit is
281 * beyond the end of the bitmap.
282 */
283 next_bit = xfs_next_bit(bip->bli_format.blf_data_map,
284 bip->bli_format.blf_map_size,
285 (uint)last_bit + 1);
286 /*
287 * If we run out of bits fill in the last iovec and get
288 * out of the loop.
289 * Else if we start a new set of bits then fill in the
290 * iovec for the series we were looking at and start
291 * counting the bits in the new one.
292 * Else we're still in the same set of bits so just
293 * keep counting and scanning.
294 */
295 if (next_bit == -1) {
296 buffer_offset = first_bit * XFS_BLI_CHUNK;
297 vecp->i_addr = xfs_buf_offset(bp, buffer_offset);
298 vecp->i_len = nbits * XFS_BLI_CHUNK;
299 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BCHUNK);
300 nvecs++;
301 break;
302 } else if (next_bit != last_bit + 1) {
303 buffer_offset = first_bit * XFS_BLI_CHUNK;
304 vecp->i_addr = xfs_buf_offset(bp, buffer_offset);
305 vecp->i_len = nbits * XFS_BLI_CHUNK;
306 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BCHUNK);
307 nvecs++;
308 vecp++;
309 first_bit = next_bit;
310 last_bit = next_bit;
311 nbits = 1;
312 } else if (xfs_buf_offset(bp, next_bit << XFS_BLI_SHIFT) !=
313 (xfs_buf_offset(bp, last_bit << XFS_BLI_SHIFT) +
314 XFS_BLI_CHUNK)) {
315 buffer_offset = first_bit * XFS_BLI_CHUNK;
316 vecp->i_addr = xfs_buf_offset(bp, buffer_offset);
317 vecp->i_len = nbits * XFS_BLI_CHUNK;
318 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BCHUNK);
319 /* You would think we need to bump the nvecs here too, but we do not
320 * this number is used by recovery, and it gets confused by the boundary
321 * split here
322 * nvecs++;
323 */
324 vecp++;
325 first_bit = next_bit;
326 last_bit = next_bit;
327 nbits = 1;
328 } else {
329 last_bit++;
330 nbits++;
331 }
332 }
333 bip->bli_format.blf_size = nvecs;
334
335 /*
336 * Check to make sure everything is consistent.
337 */
338 xfs_buf_item_trace("FORMAT NORM", bip);
339 xfs_buf_item_log_check(bip);
340 }
341
342 /*
343 * This is called to pin the buffer associated with the buf log
344 * item in memory so it cannot be written out. Simply call bpin()
345 * on the buffer to do this.
346 */
347 STATIC void
xfs_buf_item_pin(xfs_buf_log_item_t * bip)348 xfs_buf_item_pin(
349 xfs_buf_log_item_t *bip)
350 {
351 xfs_buf_t *bp;
352
353 bp = bip->bli_buf;
354 ASSERT(XFS_BUF_ISBUSY(bp));
355 ASSERT(atomic_read(&bip->bli_refcount) > 0);
356 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) ||
357 (bip->bli_flags & XFS_BLI_STALE));
358 xfs_buf_item_trace("PIN", bip);
359 xfs_buftrace("XFS_PIN", bp);
360 xfs_bpin(bp);
361 }
362
363
364 /*
365 * This is called to unpin the buffer associated with the buf log
366 * item which was previously pinned with a call to xfs_buf_item_pin().
367 * Just call bunpin() on the buffer to do this.
368 *
369 * Also drop the reference to the buf item for the current transaction.
370 * If the XFS_BLI_STALE flag is set and we are the last reference,
371 * then free up the buf log item and unlock the buffer.
372 */
373 STATIC void
xfs_buf_item_unpin(xfs_buf_log_item_t * bip,int stale)374 xfs_buf_item_unpin(
375 xfs_buf_log_item_t *bip,
376 int stale)
377 {
378 struct xfs_ail *ailp;
379 xfs_buf_t *bp;
380 int freed;
381
382 bp = bip->bli_buf;
383 ASSERT(bp != NULL);
384 ASSERT(XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t *) == bip);
385 ASSERT(atomic_read(&bip->bli_refcount) > 0);
386 xfs_buf_item_trace("UNPIN", bip);
387 xfs_buftrace("XFS_UNPIN", bp);
388
389 freed = atomic_dec_and_test(&bip->bli_refcount);
390 ailp = bip->bli_item.li_ailp;
391 xfs_bunpin(bp);
392 if (freed && stale) {
393 ASSERT(bip->bli_flags & XFS_BLI_STALE);
394 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
395 ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
396 ASSERT(XFS_BUF_ISSTALE(bp));
397 ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL);
398 xfs_buf_item_trace("UNPIN STALE", bip);
399 xfs_buftrace("XFS_UNPIN STALE", bp);
400 /*
401 * If we get called here because of an IO error, we may
402 * or may not have the item on the AIL. xfs_trans_ail_delete()
403 * will take care of that situation.
404 * xfs_trans_ail_delete() drops the AIL lock.
405 */
406 if (bip->bli_flags & XFS_BLI_STALE_INODE) {
407 xfs_buf_do_callbacks(bp, (xfs_log_item_t *)bip);
408 XFS_BUF_SET_FSPRIVATE(bp, NULL);
409 XFS_BUF_CLR_IODONE_FUNC(bp);
410 } else {
411 spin_lock(&ailp->xa_lock);
412 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)bip);
413 xfs_buf_item_relse(bp);
414 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL);
415 }
416 xfs_buf_relse(bp);
417 }
418 }
419
420 /*
421 * this is called from uncommit in the forced-shutdown path.
422 * we need to check to see if the reference count on the log item
423 * is going to drop to zero. If so, unpin will free the log item
424 * so we need to free the item's descriptor (that points to the item)
425 * in the transaction.
426 */
427 STATIC void
xfs_buf_item_unpin_remove(xfs_buf_log_item_t * bip,xfs_trans_t * tp)428 xfs_buf_item_unpin_remove(
429 xfs_buf_log_item_t *bip,
430 xfs_trans_t *tp)
431 {
432 xfs_buf_t *bp;
433 xfs_log_item_desc_t *lidp;
434 int stale = 0;
435
436 bp = bip->bli_buf;
437 /*
438 * will xfs_buf_item_unpin() call xfs_buf_item_relse()?
439 */
440 if ((atomic_read(&bip->bli_refcount) == 1) &&
441 (bip->bli_flags & XFS_BLI_STALE)) {
442 ASSERT(XFS_BUF_VALUSEMA(bip->bli_buf) <= 0);
443 xfs_buf_item_trace("UNPIN REMOVE", bip);
444 xfs_buftrace("XFS_UNPIN_REMOVE", bp);
445 /*
446 * yes -- clear the xaction descriptor in-use flag
447 * and free the chunk if required. We can safely
448 * do some work here and then call buf_item_unpin
449 * to do the rest because if the if is true, then
450 * we are holding the buffer locked so no one else
451 * will be able to bump up the refcount.
452 */
453 lidp = xfs_trans_find_item(tp, (xfs_log_item_t *) bip);
454 stale = lidp->lid_flags & XFS_LID_BUF_STALE;
455 xfs_trans_free_item(tp, lidp);
456 /*
457 * Since the transaction no longer refers to the buffer,
458 * the buffer should no longer refer to the transaction.
459 */
460 XFS_BUF_SET_FSPRIVATE2(bp, NULL);
461 }
462
463 xfs_buf_item_unpin(bip, stale);
464
465 return;
466 }
467
468 /*
469 * This is called to attempt to lock the buffer associated with this
470 * buf log item. Don't sleep on the buffer lock. If we can't get
471 * the lock right away, return 0. If we can get the lock, pull the
472 * buffer from the free list, mark it busy, and return 1.
473 */
474 STATIC uint
xfs_buf_item_trylock(xfs_buf_log_item_t * bip)475 xfs_buf_item_trylock(
476 xfs_buf_log_item_t *bip)
477 {
478 xfs_buf_t *bp;
479
480 bp = bip->bli_buf;
481
482 if (XFS_BUF_ISPINNED(bp)) {
483 return XFS_ITEM_PINNED;
484 }
485
486 if (!XFS_BUF_CPSEMA(bp)) {
487 return XFS_ITEM_LOCKED;
488 }
489
490 /*
491 * Remove the buffer from the free list. Only do this
492 * if it's on the free list. Private buffers like the
493 * superblock buffer are not.
494 */
495 XFS_BUF_HOLD(bp);
496
497 ASSERT(!(bip->bli_flags & XFS_BLI_STALE));
498 xfs_buf_item_trace("TRYLOCK SUCCESS", bip);
499 return XFS_ITEM_SUCCESS;
500 }
501
502 /*
503 * Release the buffer associated with the buf log item.
504 * If there is no dirty logged data associated with the
505 * buffer recorded in the buf log item, then free the
506 * buf log item and remove the reference to it in the
507 * buffer.
508 *
509 * This call ignores the recursion count. It is only called
510 * when the buffer should REALLY be unlocked, regardless
511 * of the recursion count.
512 *
513 * If the XFS_BLI_HOLD flag is set in the buf log item, then
514 * free the log item if necessary but do not unlock the buffer.
515 * This is for support of xfs_trans_bhold(). Make sure the
516 * XFS_BLI_HOLD field is cleared if we don't free the item.
517 */
518 STATIC void
xfs_buf_item_unlock(xfs_buf_log_item_t * bip)519 xfs_buf_item_unlock(
520 xfs_buf_log_item_t *bip)
521 {
522 int aborted;
523 xfs_buf_t *bp;
524 uint hold;
525
526 bp = bip->bli_buf;
527 xfs_buftrace("XFS_UNLOCK", bp);
528
529 /*
530 * Clear the buffer's association with this transaction.
531 */
532 XFS_BUF_SET_FSPRIVATE2(bp, NULL);
533
534 /*
535 * If this is a transaction abort, don't return early.
536 * Instead, allow the brelse to happen.
537 * Normally it would be done for stale (cancelled) buffers
538 * at unpin time, but we'll never go through the pin/unpin
539 * cycle if we abort inside commit.
540 */
541 aborted = (bip->bli_item.li_flags & XFS_LI_ABORTED) != 0;
542
543 /*
544 * If the buf item is marked stale, then don't do anything.
545 * We'll unlock the buffer and free the buf item when the
546 * buffer is unpinned for the last time.
547 */
548 if (bip->bli_flags & XFS_BLI_STALE) {
549 bip->bli_flags &= ~XFS_BLI_LOGGED;
550 xfs_buf_item_trace("UNLOCK STALE", bip);
551 ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL);
552 if (!aborted)
553 return;
554 }
555
556 /*
557 * Drop the transaction's reference to the log item if
558 * it was not logged as part of the transaction. Otherwise
559 * we'll drop the reference in xfs_buf_item_unpin() when
560 * the transaction is really through with the buffer.
561 */
562 if (!(bip->bli_flags & XFS_BLI_LOGGED)) {
563 atomic_dec(&bip->bli_refcount);
564 } else {
565 /*
566 * Clear the logged flag since this is per
567 * transaction state.
568 */
569 bip->bli_flags &= ~XFS_BLI_LOGGED;
570 }
571
572 /*
573 * Before possibly freeing the buf item, determine if we should
574 * release the buffer at the end of this routine.
575 */
576 hold = bip->bli_flags & XFS_BLI_HOLD;
577 xfs_buf_item_trace("UNLOCK", bip);
578
579 /*
580 * If the buf item isn't tracking any data, free it.
581 * Otherwise, if XFS_BLI_HOLD is set clear it.
582 */
583 if (xfs_bitmap_empty(bip->bli_format.blf_data_map,
584 bip->bli_format.blf_map_size)) {
585 xfs_buf_item_relse(bp);
586 } else if (hold) {
587 bip->bli_flags &= ~XFS_BLI_HOLD;
588 }
589
590 /*
591 * Release the buffer if XFS_BLI_HOLD was not set.
592 */
593 if (!hold) {
594 xfs_buf_relse(bp);
595 }
596 }
597
598 /*
599 * This is called to find out where the oldest active copy of the
600 * buf log item in the on disk log resides now that the last log
601 * write of it completed at the given lsn.
602 * We always re-log all the dirty data in a buffer, so usually the
603 * latest copy in the on disk log is the only one that matters. For
604 * those cases we simply return the given lsn.
605 *
606 * The one exception to this is for buffers full of newly allocated
607 * inodes. These buffers are only relogged with the XFS_BLI_INODE_BUF
608 * flag set, indicating that only the di_next_unlinked fields from the
609 * inodes in the buffers will be replayed during recovery. If the
610 * original newly allocated inode images have not yet been flushed
611 * when the buffer is so relogged, then we need to make sure that we
612 * keep the old images in the 'active' portion of the log. We do this
613 * by returning the original lsn of that transaction here rather than
614 * the current one.
615 */
616 STATIC xfs_lsn_t
xfs_buf_item_committed(xfs_buf_log_item_t * bip,xfs_lsn_t lsn)617 xfs_buf_item_committed(
618 xfs_buf_log_item_t *bip,
619 xfs_lsn_t lsn)
620 {
621 xfs_buf_item_trace("COMMITTED", bip);
622 if ((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) &&
623 (bip->bli_item.li_lsn != 0)) {
624 return bip->bli_item.li_lsn;
625 }
626 return (lsn);
627 }
628
629 /*
630 * This is called to asynchronously write the buffer associated with this
631 * buf log item out to disk. The buffer will already have been locked by
632 * a successful call to xfs_buf_item_trylock(). If the buffer still has
633 * B_DELWRI set, then get it going out to disk with a call to bawrite().
634 * If not, then just release the buffer.
635 */
636 STATIC void
xfs_buf_item_push(xfs_buf_log_item_t * bip)637 xfs_buf_item_push(
638 xfs_buf_log_item_t *bip)
639 {
640 xfs_buf_t *bp;
641
642 ASSERT(!(bip->bli_flags & XFS_BLI_STALE));
643 xfs_buf_item_trace("PUSH", bip);
644
645 bp = bip->bli_buf;
646
647 if (XFS_BUF_ISDELAYWRITE(bp)) {
648 int error;
649 error = xfs_bawrite(bip->bli_item.li_mountp, bp);
650 if (error)
651 xfs_fs_cmn_err(CE_WARN, bip->bli_item.li_mountp,
652 "xfs_buf_item_push: pushbuf error %d on bip %p, bp %p",
653 error, bip, bp);
654 } else {
655 xfs_buf_relse(bp);
656 }
657 }
658
659 /* ARGSUSED */
660 STATIC void
xfs_buf_item_committing(xfs_buf_log_item_t * bip,xfs_lsn_t commit_lsn)661 xfs_buf_item_committing(xfs_buf_log_item_t *bip, xfs_lsn_t commit_lsn)
662 {
663 }
664
665 /*
666 * This is the ops vector shared by all buf log items.
667 */
668 static struct xfs_item_ops xfs_buf_item_ops = {
669 .iop_size = (uint(*)(xfs_log_item_t*))xfs_buf_item_size,
670 .iop_format = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
671 xfs_buf_item_format,
672 .iop_pin = (void(*)(xfs_log_item_t*))xfs_buf_item_pin,
673 .iop_unpin = (void(*)(xfs_log_item_t*, int))xfs_buf_item_unpin,
674 .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t *))
675 xfs_buf_item_unpin_remove,
676 .iop_trylock = (uint(*)(xfs_log_item_t*))xfs_buf_item_trylock,
677 .iop_unlock = (void(*)(xfs_log_item_t*))xfs_buf_item_unlock,
678 .iop_committed = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
679 xfs_buf_item_committed,
680 .iop_push = (void(*)(xfs_log_item_t*))xfs_buf_item_push,
681 .iop_pushbuf = NULL,
682 .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
683 xfs_buf_item_committing
684 };
685
686
687 /*
688 * Allocate a new buf log item to go with the given buffer.
689 * Set the buffer's b_fsprivate field to point to the new
690 * buf log item. If there are other item's attached to the
691 * buffer (see xfs_buf_attach_iodone() below), then put the
692 * buf log item at the front.
693 */
694 void
xfs_buf_item_init(xfs_buf_t * bp,xfs_mount_t * mp)695 xfs_buf_item_init(
696 xfs_buf_t *bp,
697 xfs_mount_t *mp)
698 {
699 xfs_log_item_t *lip;
700 xfs_buf_log_item_t *bip;
701 int chunks;
702 int map_size;
703
704 /*
705 * Check to see if there is already a buf log item for
706 * this buffer. If there is, it is guaranteed to be
707 * the first. If we do already have one, there is
708 * nothing to do here so return.
709 */
710 if (bp->b_mount != mp)
711 bp->b_mount = mp;
712 XFS_BUF_SET_BDSTRAT_FUNC(bp, xfs_bdstrat_cb);
713 if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) {
714 lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
715 if (lip->li_type == XFS_LI_BUF) {
716 return;
717 }
718 }
719
720 /*
721 * chunks is the number of XFS_BLI_CHUNK size pieces
722 * the buffer can be divided into. Make sure not to
723 * truncate any pieces. map_size is the size of the
724 * bitmap needed to describe the chunks of the buffer.
725 */
726 chunks = (int)((XFS_BUF_COUNT(bp) + (XFS_BLI_CHUNK - 1)) >> XFS_BLI_SHIFT);
727 map_size = (int)((chunks + NBWORD) >> BIT_TO_WORD_SHIFT);
728
729 bip = (xfs_buf_log_item_t*)kmem_zone_zalloc(xfs_buf_item_zone,
730 KM_SLEEP);
731 bip->bli_item.li_type = XFS_LI_BUF;
732 bip->bli_item.li_ops = &xfs_buf_item_ops;
733 bip->bli_item.li_mountp = mp;
734 bip->bli_item.li_ailp = mp->m_ail;
735 bip->bli_buf = bp;
736 xfs_buf_hold(bp);
737 bip->bli_format.blf_type = XFS_LI_BUF;
738 bip->bli_format.blf_blkno = (__int64_t)XFS_BUF_ADDR(bp);
739 bip->bli_format.blf_len = (ushort)BTOBB(XFS_BUF_COUNT(bp));
740 bip->bli_format.blf_map_size = map_size;
741 #ifdef XFS_BLI_TRACE
742 bip->bli_trace = ktrace_alloc(XFS_BLI_TRACE_SIZE, KM_NOFS);
743 #endif
744
745 #ifdef XFS_TRANS_DEBUG
746 /*
747 * Allocate the arrays for tracking what needs to be logged
748 * and what our callers request to be logged. bli_orig
749 * holds a copy of the original, clean buffer for comparison
750 * against, and bli_logged keeps a 1 bit flag per byte in
751 * the buffer to indicate which bytes the callers have asked
752 * to have logged.
753 */
754 bip->bli_orig = (char *)kmem_alloc(XFS_BUF_COUNT(bp), KM_SLEEP);
755 memcpy(bip->bli_orig, XFS_BUF_PTR(bp), XFS_BUF_COUNT(bp));
756 bip->bli_logged = (char *)kmem_zalloc(XFS_BUF_COUNT(bp) / NBBY, KM_SLEEP);
757 #endif
758
759 /*
760 * Put the buf item into the list of items attached to the
761 * buffer at the front.
762 */
763 if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) {
764 bip->bli_item.li_bio_list =
765 XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
766 }
767 XFS_BUF_SET_FSPRIVATE(bp, bip);
768 }
769
770
771 /*
772 * Mark bytes first through last inclusive as dirty in the buf
773 * item's bitmap.
774 */
775 void
xfs_buf_item_log(xfs_buf_log_item_t * bip,uint first,uint last)776 xfs_buf_item_log(
777 xfs_buf_log_item_t *bip,
778 uint first,
779 uint last)
780 {
781 uint first_bit;
782 uint last_bit;
783 uint bits_to_set;
784 uint bits_set;
785 uint word_num;
786 uint *wordp;
787 uint bit;
788 uint end_bit;
789 uint mask;
790
791 /*
792 * Mark the item as having some dirty data for
793 * quick reference in xfs_buf_item_dirty.
794 */
795 bip->bli_flags |= XFS_BLI_DIRTY;
796
797 /*
798 * Convert byte offsets to bit numbers.
799 */
800 first_bit = first >> XFS_BLI_SHIFT;
801 last_bit = last >> XFS_BLI_SHIFT;
802
803 /*
804 * Calculate the total number of bits to be set.
805 */
806 bits_to_set = last_bit - first_bit + 1;
807
808 /*
809 * Get a pointer to the first word in the bitmap
810 * to set a bit in.
811 */
812 word_num = first_bit >> BIT_TO_WORD_SHIFT;
813 wordp = &(bip->bli_format.blf_data_map[word_num]);
814
815 /*
816 * Calculate the starting bit in the first word.
817 */
818 bit = first_bit & (uint)(NBWORD - 1);
819
820 /*
821 * First set any bits in the first word of our range.
822 * If it starts at bit 0 of the word, it will be
823 * set below rather than here. That is what the variable
824 * bit tells us. The variable bits_set tracks the number
825 * of bits that have been set so far. End_bit is the number
826 * of the last bit to be set in this word plus one.
827 */
828 if (bit) {
829 end_bit = MIN(bit + bits_to_set, (uint)NBWORD);
830 mask = ((1 << (end_bit - bit)) - 1) << bit;
831 *wordp |= mask;
832 wordp++;
833 bits_set = end_bit - bit;
834 } else {
835 bits_set = 0;
836 }
837
838 /*
839 * Now set bits a whole word at a time that are between
840 * first_bit and last_bit.
841 */
842 while ((bits_to_set - bits_set) >= NBWORD) {
843 *wordp |= 0xffffffff;
844 bits_set += NBWORD;
845 wordp++;
846 }
847
848 /*
849 * Finally, set any bits left to be set in one last partial word.
850 */
851 end_bit = bits_to_set - bits_set;
852 if (end_bit) {
853 mask = (1 << end_bit) - 1;
854 *wordp |= mask;
855 }
856
857 xfs_buf_item_log_debug(bip, first, last);
858 }
859
860
861 /*
862 * Return 1 if the buffer has some data that has been logged (at any
863 * point, not just the current transaction) and 0 if not.
864 */
865 uint
xfs_buf_item_dirty(xfs_buf_log_item_t * bip)866 xfs_buf_item_dirty(
867 xfs_buf_log_item_t *bip)
868 {
869 return (bip->bli_flags & XFS_BLI_DIRTY);
870 }
871
872 STATIC void
xfs_buf_item_free(xfs_buf_log_item_t * bip)873 xfs_buf_item_free(
874 xfs_buf_log_item_t *bip)
875 {
876 #ifdef XFS_TRANS_DEBUG
877 kmem_free(bip->bli_orig);
878 kmem_free(bip->bli_logged);
879 #endif /* XFS_TRANS_DEBUG */
880
881 #ifdef XFS_BLI_TRACE
882 ktrace_free(bip->bli_trace);
883 #endif
884 kmem_zone_free(xfs_buf_item_zone, bip);
885 }
886
887 /*
888 * This is called when the buf log item is no longer needed. It should
889 * free the buf log item associated with the given buffer and clear
890 * the buffer's pointer to the buf log item. If there are no more
891 * items in the list, clear the b_iodone field of the buffer (see
892 * xfs_buf_attach_iodone() below).
893 */
894 void
xfs_buf_item_relse(xfs_buf_t * bp)895 xfs_buf_item_relse(
896 xfs_buf_t *bp)
897 {
898 xfs_buf_log_item_t *bip;
899
900 xfs_buftrace("XFS_RELSE", bp);
901 bip = XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t*);
902 XFS_BUF_SET_FSPRIVATE(bp, bip->bli_item.li_bio_list);
903 if ((XFS_BUF_FSPRIVATE(bp, void *) == NULL) &&
904 (XFS_BUF_IODONE_FUNC(bp) != NULL)) {
905 XFS_BUF_CLR_IODONE_FUNC(bp);
906 }
907 xfs_buf_rele(bp);
908 xfs_buf_item_free(bip);
909 }
910
911
912 /*
913 * Add the given log item with its callback to the list of callbacks
914 * to be called when the buffer's I/O completes. If it is not set
915 * already, set the buffer's b_iodone() routine to be
916 * xfs_buf_iodone_callbacks() and link the log item into the list of
917 * items rooted at b_fsprivate. Items are always added as the second
918 * entry in the list if there is a first, because the buf item code
919 * assumes that the buf log item is first.
920 */
921 void
xfs_buf_attach_iodone(xfs_buf_t * bp,void (* cb)(xfs_buf_t *,xfs_log_item_t *),xfs_log_item_t * lip)922 xfs_buf_attach_iodone(
923 xfs_buf_t *bp,
924 void (*cb)(xfs_buf_t *, xfs_log_item_t *),
925 xfs_log_item_t *lip)
926 {
927 xfs_log_item_t *head_lip;
928
929 ASSERT(XFS_BUF_ISBUSY(bp));
930 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
931
932 lip->li_cb = cb;
933 if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) {
934 head_lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
935 lip->li_bio_list = head_lip->li_bio_list;
936 head_lip->li_bio_list = lip;
937 } else {
938 XFS_BUF_SET_FSPRIVATE(bp, lip);
939 }
940
941 ASSERT((XFS_BUF_IODONE_FUNC(bp) == xfs_buf_iodone_callbacks) ||
942 (XFS_BUF_IODONE_FUNC(bp) == NULL));
943 XFS_BUF_SET_IODONE_FUNC(bp, xfs_buf_iodone_callbacks);
944 }
945
946 STATIC void
xfs_buf_do_callbacks(xfs_buf_t * bp,xfs_log_item_t * lip)947 xfs_buf_do_callbacks(
948 xfs_buf_t *bp,
949 xfs_log_item_t *lip)
950 {
951 xfs_log_item_t *nlip;
952
953 while (lip != NULL) {
954 nlip = lip->li_bio_list;
955 ASSERT(lip->li_cb != NULL);
956 /*
957 * Clear the next pointer so we don't have any
958 * confusion if the item is added to another buf.
959 * Don't touch the log item after calling its
960 * callback, because it could have freed itself.
961 */
962 lip->li_bio_list = NULL;
963 lip->li_cb(bp, lip);
964 lip = nlip;
965 }
966 }
967
968 /*
969 * This is the iodone() function for buffers which have had callbacks
970 * attached to them by xfs_buf_attach_iodone(). It should remove each
971 * log item from the buffer's list and call the callback of each in turn.
972 * When done, the buffer's fsprivate field is set to NULL and the buffer
973 * is unlocked with a call to iodone().
974 */
975 void
xfs_buf_iodone_callbacks(xfs_buf_t * bp)976 xfs_buf_iodone_callbacks(
977 xfs_buf_t *bp)
978 {
979 xfs_log_item_t *lip;
980 static ulong lasttime;
981 static xfs_buftarg_t *lasttarg;
982 xfs_mount_t *mp;
983
984 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) != NULL);
985 lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
986
987 if (XFS_BUF_GETERROR(bp) != 0) {
988 /*
989 * If we've already decided to shutdown the filesystem
990 * because of IO errors, there's no point in giving this
991 * a retry.
992 */
993 mp = lip->li_mountp;
994 if (XFS_FORCED_SHUTDOWN(mp)) {
995 ASSERT(XFS_BUF_TARGET(bp) == mp->m_ddev_targp);
996 XFS_BUF_SUPER_STALE(bp);
997 xfs_buftrace("BUF_IODONE_CB", bp);
998 xfs_buf_do_callbacks(bp, lip);
999 XFS_BUF_SET_FSPRIVATE(bp, NULL);
1000 XFS_BUF_CLR_IODONE_FUNC(bp);
1001 xfs_biodone(bp);
1002 return;
1003 }
1004
1005 if ((XFS_BUF_TARGET(bp) != lasttarg) ||
1006 (time_after(jiffies, (lasttime + 5*HZ)))) {
1007 lasttime = jiffies;
1008 cmn_err(CE_ALERT, "Device %s, XFS metadata write error"
1009 " block 0x%llx in %s",
1010 XFS_BUFTARG_NAME(XFS_BUF_TARGET(bp)),
1011 (__uint64_t)XFS_BUF_ADDR(bp), mp->m_fsname);
1012 }
1013 lasttarg = XFS_BUF_TARGET(bp);
1014
1015 if (XFS_BUF_ISASYNC(bp)) {
1016 /*
1017 * If the write was asynchronous then noone will be
1018 * looking for the error. Clear the error state
1019 * and write the buffer out again delayed write.
1020 *
1021 * XXXsup This is OK, so long as we catch these
1022 * before we start the umount; we don't want these
1023 * DELWRI metadata bufs to be hanging around.
1024 */
1025 XFS_BUF_ERROR(bp,0); /* errno of 0 unsets the flag */
1026
1027 if (!(XFS_BUF_ISSTALE(bp))) {
1028 XFS_BUF_DELAYWRITE(bp);
1029 XFS_BUF_DONE(bp);
1030 XFS_BUF_SET_START(bp);
1031 }
1032 ASSERT(XFS_BUF_IODONE_FUNC(bp));
1033 xfs_buftrace("BUF_IODONE ASYNC", bp);
1034 xfs_buf_relse(bp);
1035 } else {
1036 /*
1037 * If the write of the buffer was not asynchronous,
1038 * then we want to make sure to return the error
1039 * to the caller of bwrite(). Because of this we
1040 * cannot clear the B_ERROR state at this point.
1041 * Instead we install a callback function that
1042 * will be called when the buffer is released, and
1043 * that routine will clear the error state and
1044 * set the buffer to be written out again after
1045 * some delay.
1046 */
1047 /* We actually overwrite the existing b-relse
1048 function at times, but we're gonna be shutting down
1049 anyway. */
1050 XFS_BUF_SET_BRELSE_FUNC(bp,xfs_buf_error_relse);
1051 XFS_BUF_DONE(bp);
1052 XFS_BUF_FINISH_IOWAIT(bp);
1053 }
1054 return;
1055 }
1056 #ifdef XFSERRORDEBUG
1057 xfs_buftrace("XFS BUFCB NOERR", bp);
1058 #endif
1059 xfs_buf_do_callbacks(bp, lip);
1060 XFS_BUF_SET_FSPRIVATE(bp, NULL);
1061 XFS_BUF_CLR_IODONE_FUNC(bp);
1062 xfs_biodone(bp);
1063 }
1064
1065 /*
1066 * This is a callback routine attached to a buffer which gets an error
1067 * when being written out synchronously.
1068 */
1069 STATIC void
xfs_buf_error_relse(xfs_buf_t * bp)1070 xfs_buf_error_relse(
1071 xfs_buf_t *bp)
1072 {
1073 xfs_log_item_t *lip;
1074 xfs_mount_t *mp;
1075
1076 lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
1077 mp = (xfs_mount_t *)lip->li_mountp;
1078 ASSERT(XFS_BUF_TARGET(bp) == mp->m_ddev_targp);
1079
1080 XFS_BUF_STALE(bp);
1081 XFS_BUF_DONE(bp);
1082 XFS_BUF_UNDELAYWRITE(bp);
1083 XFS_BUF_ERROR(bp,0);
1084 xfs_buftrace("BUF_ERROR_RELSE", bp);
1085 if (! XFS_FORCED_SHUTDOWN(mp))
1086 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1087 /*
1088 * We have to unpin the pinned buffers so do the
1089 * callbacks.
1090 */
1091 xfs_buf_do_callbacks(bp, lip);
1092 XFS_BUF_SET_FSPRIVATE(bp, NULL);
1093 XFS_BUF_CLR_IODONE_FUNC(bp);
1094 XFS_BUF_SET_BRELSE_FUNC(bp,NULL);
1095 xfs_buf_relse(bp);
1096 }
1097
1098
1099 /*
1100 * This is the iodone() function for buffers which have been
1101 * logged. It is called when they are eventually flushed out.
1102 * It should remove the buf item from the AIL, and free the buf item.
1103 * It is called by xfs_buf_iodone_callbacks() above which will take
1104 * care of cleaning up the buffer itself.
1105 */
1106 /* ARGSUSED */
1107 void
xfs_buf_iodone(xfs_buf_t * bp,xfs_buf_log_item_t * bip)1108 xfs_buf_iodone(
1109 xfs_buf_t *bp,
1110 xfs_buf_log_item_t *bip)
1111 {
1112 struct xfs_ail *ailp = bip->bli_item.li_ailp;
1113
1114 ASSERT(bip->bli_buf == bp);
1115
1116 xfs_buf_rele(bp);
1117
1118 /*
1119 * If we are forcibly shutting down, this may well be
1120 * off the AIL already. That's because we simulate the
1121 * log-committed callbacks to unpin these buffers. Or we may never
1122 * have put this item on AIL because of the transaction was
1123 * aborted forcibly. xfs_trans_ail_delete() takes care of these.
1124 *
1125 * Either way, AIL is useless if we're forcing a shutdown.
1126 */
1127 spin_lock(&ailp->xa_lock);
1128 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)bip);
1129 xfs_buf_item_free(bip);
1130 }
1131
1132 #if defined(XFS_BLI_TRACE)
1133 void
xfs_buf_item_trace(char * id,xfs_buf_log_item_t * bip)1134 xfs_buf_item_trace(
1135 char *id,
1136 xfs_buf_log_item_t *bip)
1137 {
1138 xfs_buf_t *bp;
1139 ASSERT(bip->bli_trace != NULL);
1140
1141 bp = bip->bli_buf;
1142 ktrace_enter(bip->bli_trace,
1143 (void *)id,
1144 (void *)bip->bli_buf,
1145 (void *)((unsigned long)bip->bli_flags),
1146 (void *)((unsigned long)bip->bli_recur),
1147 (void *)((unsigned long)atomic_read(&bip->bli_refcount)),
1148 (void *)((unsigned long)
1149 (0xFFFFFFFF & XFS_BUF_ADDR(bp) >> 32)),
1150 (void *)((unsigned long)(0xFFFFFFFF & XFS_BUF_ADDR(bp))),
1151 (void *)((unsigned long)XFS_BUF_COUNT(bp)),
1152 (void *)((unsigned long)XFS_BUF_BFLAGS(bp)),
1153 XFS_BUF_FSPRIVATE(bp, void *),
1154 XFS_BUF_FSPRIVATE2(bp, void *),
1155 (void *)(unsigned long)XFS_BUF_ISPINNED(bp),
1156 (void *)XFS_BUF_IODONE_FUNC(bp),
1157 (void *)((unsigned long)(XFS_BUF_VALUSEMA(bp))),
1158 (void *)bip->bli_item.li_desc,
1159 (void *)((unsigned long)bip->bli_item.li_flags));
1160 }
1161 #endif /* XFS_BLI_TRACE */
1162