xref: /linux/fs/xfs/xfs_buf_item.c (revision 643d1f7fe3aa12c8bdea6fa5b4ba874ff6dd601d)
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
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
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
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
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
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
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
374 xfs_buf_item_unpin(
375 	xfs_buf_log_item_t	*bip,
376 	int			stale)
377 {
378 	xfs_mount_t	*mp;
379 	xfs_buf_t	*bp;
380 	int		freed;
381 	SPLDECL(s);
382 
383 	bp = bip->bli_buf;
384 	ASSERT(bp != NULL);
385 	ASSERT(XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t *) == bip);
386 	ASSERT(atomic_read(&bip->bli_refcount) > 0);
387 	xfs_buf_item_trace("UNPIN", bip);
388 	xfs_buftrace("XFS_UNPIN", bp);
389 
390 	freed = atomic_dec_and_test(&bip->bli_refcount);
391 	mp = bip->bli_item.li_mountp;
392 	xfs_bunpin(bp);
393 	if (freed && stale) {
394 		ASSERT(bip->bli_flags & XFS_BLI_STALE);
395 		ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
396 		ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
397 		ASSERT(XFS_BUF_ISSTALE(bp));
398 		ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL);
399 		xfs_buf_item_trace("UNPIN STALE", bip);
400 		xfs_buftrace("XFS_UNPIN STALE", bp);
401 		/*
402 		 * If we get called here because of an IO error, we may
403 		 * or may not have the item on the AIL. xfs_trans_delete_ail()
404 		 * will take care of that situation.
405 		 * xfs_trans_delete_ail() drops the AIL lock.
406 		 */
407 		if (bip->bli_flags & XFS_BLI_STALE_INODE) {
408 			xfs_buf_do_callbacks(bp, (xfs_log_item_t *)bip);
409 			XFS_BUF_SET_FSPRIVATE(bp, NULL);
410 			XFS_BUF_CLR_IODONE_FUNC(bp);
411 		} else {
412 			AIL_LOCK(mp,s);
413 			xfs_trans_delete_ail(mp, (xfs_log_item_t *)bip, s);
414 			xfs_buf_item_relse(bp);
415 			ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL);
416 		}
417 		xfs_buf_relse(bp);
418 	}
419 }
420 
421 /*
422  * this is called from uncommit in the forced-shutdown path.
423  * we need to check to see if the reference count on the log item
424  * is going to drop to zero.  If so, unpin will free the log item
425  * so we need to free the item's descriptor (that points to the item)
426  * in the transaction.
427  */
428 STATIC void
429 xfs_buf_item_unpin_remove(
430 	xfs_buf_log_item_t	*bip,
431 	xfs_trans_t		*tp)
432 {
433 	xfs_buf_t		*bp;
434 	xfs_log_item_desc_t	*lidp;
435 	int			stale = 0;
436 
437 	bp = bip->bli_buf;
438 	/*
439 	 * will xfs_buf_item_unpin() call xfs_buf_item_relse()?
440 	 */
441 	if ((atomic_read(&bip->bli_refcount) == 1) &&
442 	    (bip->bli_flags & XFS_BLI_STALE)) {
443 		ASSERT(XFS_BUF_VALUSEMA(bip->bli_buf) <= 0);
444 		xfs_buf_item_trace("UNPIN REMOVE", bip);
445 		xfs_buftrace("XFS_UNPIN_REMOVE", bp);
446 		/*
447 		 * yes -- clear the xaction descriptor in-use flag
448 		 * and free the chunk if required.  We can safely
449 		 * do some work here and then call buf_item_unpin
450 		 * to do the rest because if the if is true, then
451 		 * we are holding the buffer locked so no one else
452 		 * will be able to bump up the refcount.
453 		 */
454 		lidp = xfs_trans_find_item(tp, (xfs_log_item_t *) bip);
455 		stale = lidp->lid_flags & XFS_LID_BUF_STALE;
456 		xfs_trans_free_item(tp, lidp);
457 		/*
458 		 * Since the transaction no longer refers to the buffer,
459 		 * the buffer should no longer refer to the transaction.
460 		 */
461 		XFS_BUF_SET_FSPRIVATE2(bp, NULL);
462 	}
463 
464 	xfs_buf_item_unpin(bip, stale);
465 
466 	return;
467 }
468 
469 /*
470  * This is called to attempt to lock the buffer associated with this
471  * buf log item.  Don't sleep on the buffer lock.  If we can't get
472  * the lock right away, return 0.  If we can get the lock, pull the
473  * buffer from the free list, mark it busy, and return 1.
474  */
475 STATIC uint
476 xfs_buf_item_trylock(
477 	xfs_buf_log_item_t	*bip)
478 {
479 	xfs_buf_t	*bp;
480 
481 	bp = bip->bli_buf;
482 
483 	if (XFS_BUF_ISPINNED(bp)) {
484 		return XFS_ITEM_PINNED;
485 	}
486 
487 	if (!XFS_BUF_CPSEMA(bp)) {
488 		return XFS_ITEM_LOCKED;
489 	}
490 
491 	/*
492 	 * Remove the buffer from the free list.  Only do this
493 	 * if it's on the free list.  Private buffers like the
494 	 * superblock buffer are not.
495 	 */
496 	XFS_BUF_HOLD(bp);
497 
498 	ASSERT(!(bip->bli_flags & XFS_BLI_STALE));
499 	xfs_buf_item_trace("TRYLOCK SUCCESS", bip);
500 	return XFS_ITEM_SUCCESS;
501 }
502 
503 /*
504  * Release the buffer associated with the buf log item.
505  * If there is no dirty logged data associated with the
506  * buffer recorded in the buf log item, then free the
507  * buf log item and remove the reference to it in the
508  * buffer.
509  *
510  * This call ignores the recursion count.  It is only called
511  * when the buffer should REALLY be unlocked, regardless
512  * of the recursion count.
513  *
514  * If the XFS_BLI_HOLD flag is set in the buf log item, then
515  * free the log item if necessary but do not unlock the buffer.
516  * This is for support of xfs_trans_bhold(). Make sure the
517  * XFS_BLI_HOLD field is cleared if we don't free the item.
518  */
519 STATIC void
520 xfs_buf_item_unlock(
521 	xfs_buf_log_item_t	*bip)
522 {
523 	int		aborted;
524 	xfs_buf_t	*bp;
525 	uint		hold;
526 
527 	bp = bip->bli_buf;
528 	xfs_buftrace("XFS_UNLOCK", bp);
529 
530 	/*
531 	 * Clear the buffer's association with this transaction.
532 	 */
533 	XFS_BUF_SET_FSPRIVATE2(bp, NULL);
534 
535 	/*
536 	 * If this is a transaction abort, don't return early.
537 	 * Instead, allow the brelse to happen.
538 	 * Normally it would be done for stale (cancelled) buffers
539 	 * at unpin time, but we'll never go through the pin/unpin
540 	 * cycle if we abort inside commit.
541 	 */
542 	aborted = (bip->bli_item.li_flags & XFS_LI_ABORTED) != 0;
543 
544 	/*
545 	 * If the buf item is marked stale, then don't do anything.
546 	 * We'll unlock the buffer and free the buf item when the
547 	 * buffer is unpinned for the last time.
548 	 */
549 	if (bip->bli_flags & XFS_BLI_STALE) {
550 		bip->bli_flags &= ~XFS_BLI_LOGGED;
551 		xfs_buf_item_trace("UNLOCK STALE", bip);
552 		ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL);
553 		if (!aborted)
554 			return;
555 	}
556 
557 	/*
558 	 * Drop the transaction's reference to the log item if
559 	 * it was not logged as part of the transaction.  Otherwise
560 	 * we'll drop the reference in xfs_buf_item_unpin() when
561 	 * the transaction is really through with the buffer.
562 	 */
563 	if (!(bip->bli_flags & XFS_BLI_LOGGED)) {
564 		atomic_dec(&bip->bli_refcount);
565 	} else {
566 		/*
567 		 * Clear the logged flag since this is per
568 		 * transaction state.
569 		 */
570 		bip->bli_flags &= ~XFS_BLI_LOGGED;
571 	}
572 
573 	/*
574 	 * Before possibly freeing the buf item, determine if we should
575 	 * release the buffer at the end of this routine.
576 	 */
577 	hold = bip->bli_flags & XFS_BLI_HOLD;
578 	xfs_buf_item_trace("UNLOCK", bip);
579 
580 	/*
581 	 * If the buf item isn't tracking any data, free it.
582 	 * Otherwise, if XFS_BLI_HOLD is set clear it.
583 	 */
584 	if (xfs_bitmap_empty(bip->bli_format.blf_data_map,
585 			     bip->bli_format.blf_map_size)) {
586 		xfs_buf_item_relse(bp);
587 	} else if (hold) {
588 		bip->bli_flags &= ~XFS_BLI_HOLD;
589 	}
590 
591 	/*
592 	 * Release the buffer if XFS_BLI_HOLD was not set.
593 	 */
594 	if (!hold) {
595 		xfs_buf_relse(bp);
596 	}
597 }
598 
599 /*
600  * This is called to find out where the oldest active copy of the
601  * buf log item in the on disk log resides now that the last log
602  * write of it completed at the given lsn.
603  * We always re-log all the dirty data in a buffer, so usually the
604  * latest copy in the on disk log is the only one that matters.  For
605  * those cases we simply return the given lsn.
606  *
607  * The one exception to this is for buffers full of newly allocated
608  * inodes.  These buffers are only relogged with the XFS_BLI_INODE_BUF
609  * flag set, indicating that only the di_next_unlinked fields from the
610  * inodes in the buffers will be replayed during recovery.  If the
611  * original newly allocated inode images have not yet been flushed
612  * when the buffer is so relogged, then we need to make sure that we
613  * keep the old images in the 'active' portion of the log.  We do this
614  * by returning the original lsn of that transaction here rather than
615  * the current one.
616  */
617 STATIC xfs_lsn_t
618 xfs_buf_item_committed(
619 	xfs_buf_log_item_t	*bip,
620 	xfs_lsn_t		lsn)
621 {
622 	xfs_buf_item_trace("COMMITTED", bip);
623 	if ((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) &&
624 	    (bip->bli_item.li_lsn != 0)) {
625 		return bip->bli_item.li_lsn;
626 	}
627 	return (lsn);
628 }
629 
630 /*
631  * This is called to asynchronously write the buffer associated with this
632  * buf log item out to disk. The buffer will already have been locked by
633  * a successful call to xfs_buf_item_trylock().  If the buffer still has
634  * B_DELWRI set, then get it going out to disk with a call to bawrite().
635  * If not, then just release the buffer.
636  */
637 STATIC void
638 xfs_buf_item_push(
639 	xfs_buf_log_item_t	*bip)
640 {
641 	xfs_buf_t	*bp;
642 
643 	ASSERT(!(bip->bli_flags & XFS_BLI_STALE));
644 	xfs_buf_item_trace("PUSH", bip);
645 
646 	bp = bip->bli_buf;
647 
648 	if (XFS_BUF_ISDELAYWRITE(bp)) {
649 		xfs_bawrite(bip->bli_item.li_mountp, bp);
650 	} else {
651 		xfs_buf_relse(bp);
652 	}
653 }
654 
655 /* ARGSUSED */
656 STATIC void
657 xfs_buf_item_committing(xfs_buf_log_item_t *bip, xfs_lsn_t commit_lsn)
658 {
659 }
660 
661 /*
662  * This is the ops vector shared by all buf log items.
663  */
664 static struct xfs_item_ops xfs_buf_item_ops = {
665 	.iop_size	= (uint(*)(xfs_log_item_t*))xfs_buf_item_size,
666 	.iop_format	= (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
667 					xfs_buf_item_format,
668 	.iop_pin	= (void(*)(xfs_log_item_t*))xfs_buf_item_pin,
669 	.iop_unpin	= (void(*)(xfs_log_item_t*, int))xfs_buf_item_unpin,
670 	.iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t *))
671 					xfs_buf_item_unpin_remove,
672 	.iop_trylock	= (uint(*)(xfs_log_item_t*))xfs_buf_item_trylock,
673 	.iop_unlock	= (void(*)(xfs_log_item_t*))xfs_buf_item_unlock,
674 	.iop_committed	= (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
675 					xfs_buf_item_committed,
676 	.iop_push	= (void(*)(xfs_log_item_t*))xfs_buf_item_push,
677 	.iop_pushbuf	= NULL,
678 	.iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
679 					xfs_buf_item_committing
680 };
681 
682 
683 /*
684  * Allocate a new buf log item to go with the given buffer.
685  * Set the buffer's b_fsprivate field to point to the new
686  * buf log item.  If there are other item's attached to the
687  * buffer (see xfs_buf_attach_iodone() below), then put the
688  * buf log item at the front.
689  */
690 void
691 xfs_buf_item_init(
692 	xfs_buf_t	*bp,
693 	xfs_mount_t	*mp)
694 {
695 	xfs_log_item_t		*lip;
696 	xfs_buf_log_item_t	*bip;
697 	int			chunks;
698 	int			map_size;
699 
700 	/*
701 	 * Check to see if there is already a buf log item for
702 	 * this buffer.  If there is, it is guaranteed to be
703 	 * the first.  If we do already have one, there is
704 	 * nothing to do here so return.
705 	 */
706 	if (XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *) != mp)
707 		XFS_BUF_SET_FSPRIVATE3(bp, mp);
708 	XFS_BUF_SET_BDSTRAT_FUNC(bp, xfs_bdstrat_cb);
709 	if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) {
710 		lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
711 		if (lip->li_type == XFS_LI_BUF) {
712 			return;
713 		}
714 	}
715 
716 	/*
717 	 * chunks is the number of XFS_BLI_CHUNK size pieces
718 	 * the buffer can be divided into. Make sure not to
719 	 * truncate any pieces.  map_size is the size of the
720 	 * bitmap needed to describe the chunks of the buffer.
721 	 */
722 	chunks = (int)((XFS_BUF_COUNT(bp) + (XFS_BLI_CHUNK - 1)) >> XFS_BLI_SHIFT);
723 	map_size = (int)((chunks + NBWORD) >> BIT_TO_WORD_SHIFT);
724 
725 	bip = (xfs_buf_log_item_t*)kmem_zone_zalloc(xfs_buf_item_zone,
726 						    KM_SLEEP);
727 	bip->bli_item.li_type = XFS_LI_BUF;
728 	bip->bli_item.li_ops = &xfs_buf_item_ops;
729 	bip->bli_item.li_mountp = mp;
730 	bip->bli_buf = bp;
731 	bip->bli_format.blf_type = XFS_LI_BUF;
732 	bip->bli_format.blf_blkno = (__int64_t)XFS_BUF_ADDR(bp);
733 	bip->bli_format.blf_len = (ushort)BTOBB(XFS_BUF_COUNT(bp));
734 	bip->bli_format.blf_map_size = map_size;
735 #ifdef XFS_BLI_TRACE
736 	bip->bli_trace = ktrace_alloc(XFS_BLI_TRACE_SIZE, KM_SLEEP);
737 #endif
738 
739 #ifdef XFS_TRANS_DEBUG
740 	/*
741 	 * Allocate the arrays for tracking what needs to be logged
742 	 * and what our callers request to be logged.  bli_orig
743 	 * holds a copy of the original, clean buffer for comparison
744 	 * against, and bli_logged keeps a 1 bit flag per byte in
745 	 * the buffer to indicate which bytes the callers have asked
746 	 * to have logged.
747 	 */
748 	bip->bli_orig = (char *)kmem_alloc(XFS_BUF_COUNT(bp), KM_SLEEP);
749 	memcpy(bip->bli_orig, XFS_BUF_PTR(bp), XFS_BUF_COUNT(bp));
750 	bip->bli_logged = (char *)kmem_zalloc(XFS_BUF_COUNT(bp) / NBBY, KM_SLEEP);
751 #endif
752 
753 	/*
754 	 * Put the buf item into the list of items attached to the
755 	 * buffer at the front.
756 	 */
757 	if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) {
758 		bip->bli_item.li_bio_list =
759 				XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
760 	}
761 	XFS_BUF_SET_FSPRIVATE(bp, bip);
762 }
763 
764 
765 /*
766  * Mark bytes first through last inclusive as dirty in the buf
767  * item's bitmap.
768  */
769 void
770 xfs_buf_item_log(
771 	xfs_buf_log_item_t	*bip,
772 	uint			first,
773 	uint			last)
774 {
775 	uint		first_bit;
776 	uint		last_bit;
777 	uint		bits_to_set;
778 	uint		bits_set;
779 	uint		word_num;
780 	uint		*wordp;
781 	uint		bit;
782 	uint		end_bit;
783 	uint		mask;
784 
785 	/*
786 	 * Mark the item as having some dirty data for
787 	 * quick reference in xfs_buf_item_dirty.
788 	 */
789 	bip->bli_flags |= XFS_BLI_DIRTY;
790 
791 	/*
792 	 * Convert byte offsets to bit numbers.
793 	 */
794 	first_bit = first >> XFS_BLI_SHIFT;
795 	last_bit = last >> XFS_BLI_SHIFT;
796 
797 	/*
798 	 * Calculate the total number of bits to be set.
799 	 */
800 	bits_to_set = last_bit - first_bit + 1;
801 
802 	/*
803 	 * Get a pointer to the first word in the bitmap
804 	 * to set a bit in.
805 	 */
806 	word_num = first_bit >> BIT_TO_WORD_SHIFT;
807 	wordp = &(bip->bli_format.blf_data_map[word_num]);
808 
809 	/*
810 	 * Calculate the starting bit in the first word.
811 	 */
812 	bit = first_bit & (uint)(NBWORD - 1);
813 
814 	/*
815 	 * First set any bits in the first word of our range.
816 	 * If it starts at bit 0 of the word, it will be
817 	 * set below rather than here.  That is what the variable
818 	 * bit tells us. The variable bits_set tracks the number
819 	 * of bits that have been set so far.  End_bit is the number
820 	 * of the last bit to be set in this word plus one.
821 	 */
822 	if (bit) {
823 		end_bit = MIN(bit + bits_to_set, (uint)NBWORD);
824 		mask = ((1 << (end_bit - bit)) - 1) << bit;
825 		*wordp |= mask;
826 		wordp++;
827 		bits_set = end_bit - bit;
828 	} else {
829 		bits_set = 0;
830 	}
831 
832 	/*
833 	 * Now set bits a whole word at a time that are between
834 	 * first_bit and last_bit.
835 	 */
836 	while ((bits_to_set - bits_set) >= NBWORD) {
837 		*wordp |= 0xffffffff;
838 		bits_set += NBWORD;
839 		wordp++;
840 	}
841 
842 	/*
843 	 * Finally, set any bits left to be set in one last partial word.
844 	 */
845 	end_bit = bits_to_set - bits_set;
846 	if (end_bit) {
847 		mask = (1 << end_bit) - 1;
848 		*wordp |= mask;
849 	}
850 
851 	xfs_buf_item_log_debug(bip, first, last);
852 }
853 
854 
855 /*
856  * Return 1 if the buffer has some data that has been logged (at any
857  * point, not just the current transaction) and 0 if not.
858  */
859 uint
860 xfs_buf_item_dirty(
861 	xfs_buf_log_item_t	*bip)
862 {
863 	return (bip->bli_flags & XFS_BLI_DIRTY);
864 }
865 
866 /*
867  * This is called when the buf log item is no longer needed.  It should
868  * free the buf log item associated with the given buffer and clear
869  * the buffer's pointer to the buf log item.  If there are no more
870  * items in the list, clear the b_iodone field of the buffer (see
871  * xfs_buf_attach_iodone() below).
872  */
873 void
874 xfs_buf_item_relse(
875 	xfs_buf_t	*bp)
876 {
877 	xfs_buf_log_item_t	*bip;
878 
879 	xfs_buftrace("XFS_RELSE", bp);
880 	bip = XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t*);
881 	XFS_BUF_SET_FSPRIVATE(bp, bip->bli_item.li_bio_list);
882 	if ((XFS_BUF_FSPRIVATE(bp, void *) == NULL) &&
883 	    (XFS_BUF_IODONE_FUNC(bp) != NULL)) {
884 		XFS_BUF_CLR_IODONE_FUNC(bp);
885 	}
886 
887 #ifdef XFS_TRANS_DEBUG
888 	kmem_free(bip->bli_orig, XFS_BUF_COUNT(bp));
889 	bip->bli_orig = NULL;
890 	kmem_free(bip->bli_logged, XFS_BUF_COUNT(bp) / NBBY);
891 	bip->bli_logged = NULL;
892 #endif /* XFS_TRANS_DEBUG */
893 
894 #ifdef XFS_BLI_TRACE
895 	ktrace_free(bip->bli_trace);
896 #endif
897 	kmem_zone_free(xfs_buf_item_zone, bip);
898 }
899 
900 
901 /*
902  * Add the given log item with its callback to the list of callbacks
903  * to be called when the buffer's I/O completes.  If it is not set
904  * already, set the buffer's b_iodone() routine to be
905  * xfs_buf_iodone_callbacks() and link the log item into the list of
906  * items rooted at b_fsprivate.  Items are always added as the second
907  * entry in the list if there is a first, because the buf item code
908  * assumes that the buf log item is first.
909  */
910 void
911 xfs_buf_attach_iodone(
912 	xfs_buf_t	*bp,
913 	void		(*cb)(xfs_buf_t *, xfs_log_item_t *),
914 	xfs_log_item_t	*lip)
915 {
916 	xfs_log_item_t	*head_lip;
917 
918 	ASSERT(XFS_BUF_ISBUSY(bp));
919 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
920 
921 	lip->li_cb = cb;
922 	if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) {
923 		head_lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
924 		lip->li_bio_list = head_lip->li_bio_list;
925 		head_lip->li_bio_list = lip;
926 	} else {
927 		XFS_BUF_SET_FSPRIVATE(bp, lip);
928 	}
929 
930 	ASSERT((XFS_BUF_IODONE_FUNC(bp) == xfs_buf_iodone_callbacks) ||
931 	       (XFS_BUF_IODONE_FUNC(bp) == NULL));
932 	XFS_BUF_SET_IODONE_FUNC(bp, xfs_buf_iodone_callbacks);
933 }
934 
935 STATIC void
936 xfs_buf_do_callbacks(
937 	xfs_buf_t	*bp,
938 	xfs_log_item_t	*lip)
939 {
940 	xfs_log_item_t	*nlip;
941 
942 	while (lip != NULL) {
943 		nlip = lip->li_bio_list;
944 		ASSERT(lip->li_cb != NULL);
945 		/*
946 		 * Clear the next pointer so we don't have any
947 		 * confusion if the item is added to another buf.
948 		 * Don't touch the log item after calling its
949 		 * callback, because it could have freed itself.
950 		 */
951 		lip->li_bio_list = NULL;
952 		lip->li_cb(bp, lip);
953 		lip = nlip;
954 	}
955 }
956 
957 /*
958  * This is the iodone() function for buffers which have had callbacks
959  * attached to them by xfs_buf_attach_iodone().  It should remove each
960  * log item from the buffer's list and call the callback of each in turn.
961  * When done, the buffer's fsprivate field is set to NULL and the buffer
962  * is unlocked with a call to iodone().
963  */
964 void
965 xfs_buf_iodone_callbacks(
966 	xfs_buf_t	*bp)
967 {
968 	xfs_log_item_t	*lip;
969 	static ulong	lasttime;
970 	static xfs_buftarg_t *lasttarg;
971 	xfs_mount_t	*mp;
972 
973 	ASSERT(XFS_BUF_FSPRIVATE(bp, void *) != NULL);
974 	lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
975 
976 	if (XFS_BUF_GETERROR(bp) != 0) {
977 		/*
978 		 * If we've already decided to shutdown the filesystem
979 		 * because of IO errors, there's no point in giving this
980 		 * a retry.
981 		 */
982 		mp = lip->li_mountp;
983 		if (XFS_FORCED_SHUTDOWN(mp)) {
984 			ASSERT(XFS_BUF_TARGET(bp) == mp->m_ddev_targp);
985 			XFS_BUF_SUPER_STALE(bp);
986 			xfs_buftrace("BUF_IODONE_CB", bp);
987 			xfs_buf_do_callbacks(bp, lip);
988 			XFS_BUF_SET_FSPRIVATE(bp, NULL);
989 			XFS_BUF_CLR_IODONE_FUNC(bp);
990 
991 			/*
992 			 * XFS_SHUT flag gets set when we go thru the
993 			 * entire buffer cache and deliberately start
994 			 * throwing away delayed write buffers.
995 			 * Since there's no biowait done on those,
996 			 * we should just brelse them.
997 			 */
998 			if (XFS_BUF_ISSHUT(bp)) {
999 			    XFS_BUF_UNSHUT(bp);
1000 				xfs_buf_relse(bp);
1001 			} else {
1002 				xfs_biodone(bp);
1003 			}
1004 
1005 			return;
1006 		}
1007 
1008 		if ((XFS_BUF_TARGET(bp) != lasttarg) ||
1009 		    (time_after(jiffies, (lasttime + 5*HZ)))) {
1010 			lasttime = jiffies;
1011 			cmn_err(CE_ALERT, "Device %s, XFS metadata write error"
1012 					" block 0x%llx in %s",
1013 				XFS_BUFTARG_NAME(XFS_BUF_TARGET(bp)),
1014 			      (__uint64_t)XFS_BUF_ADDR(bp), mp->m_fsname);
1015 		}
1016 		lasttarg = XFS_BUF_TARGET(bp);
1017 
1018 		if (XFS_BUF_ISASYNC(bp)) {
1019 			/*
1020 			 * If the write was asynchronous then noone will be
1021 			 * looking for the error.  Clear the error state
1022 			 * and write the buffer out again delayed write.
1023 			 *
1024 			 * XXXsup This is OK, so long as we catch these
1025 			 * before we start the umount; we don't want these
1026 			 * DELWRI metadata bufs to be hanging around.
1027 			 */
1028 			XFS_BUF_ERROR(bp,0); /* errno of 0 unsets the flag */
1029 
1030 			if (!(XFS_BUF_ISSTALE(bp))) {
1031 				XFS_BUF_DELAYWRITE(bp);
1032 				XFS_BUF_DONE(bp);
1033 				XFS_BUF_SET_START(bp);
1034 			}
1035 			ASSERT(XFS_BUF_IODONE_FUNC(bp));
1036 			xfs_buftrace("BUF_IODONE ASYNC", bp);
1037 			xfs_buf_relse(bp);
1038 		} else {
1039 			/*
1040 			 * If the write of the buffer was not asynchronous,
1041 			 * then we want to make sure to return the error
1042 			 * to the caller of bwrite().  Because of this we
1043 			 * cannot clear the B_ERROR state at this point.
1044 			 * Instead we install a callback function that
1045 			 * will be called when the buffer is released, and
1046 			 * that routine will clear the error state and
1047 			 * set the buffer to be written out again after
1048 			 * some delay.
1049 			 */
1050 			/* We actually overwrite the existing b-relse
1051 			   function at times, but we're gonna be shutting down
1052 			   anyway. */
1053 			XFS_BUF_SET_BRELSE_FUNC(bp,xfs_buf_error_relse);
1054 			XFS_BUF_DONE(bp);
1055 			XFS_BUF_V_IODONESEMA(bp);
1056 		}
1057 		return;
1058 	}
1059 #ifdef XFSERRORDEBUG
1060 	xfs_buftrace("XFS BUFCB NOERR", bp);
1061 #endif
1062 	xfs_buf_do_callbacks(bp, lip);
1063 	XFS_BUF_SET_FSPRIVATE(bp, NULL);
1064 	XFS_BUF_CLR_IODONE_FUNC(bp);
1065 	xfs_biodone(bp);
1066 }
1067 
1068 /*
1069  * This is a callback routine attached to a buffer which gets an error
1070  * when being written out synchronously.
1071  */
1072 STATIC void
1073 xfs_buf_error_relse(
1074 	xfs_buf_t	*bp)
1075 {
1076 	xfs_log_item_t	*lip;
1077 	xfs_mount_t	*mp;
1078 
1079 	lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
1080 	mp = (xfs_mount_t *)lip->li_mountp;
1081 	ASSERT(XFS_BUF_TARGET(bp) == mp->m_ddev_targp);
1082 
1083 	XFS_BUF_STALE(bp);
1084 	XFS_BUF_DONE(bp);
1085 	XFS_BUF_UNDELAYWRITE(bp);
1086 	XFS_BUF_ERROR(bp,0);
1087 	xfs_buftrace("BUF_ERROR_RELSE", bp);
1088 	if (! XFS_FORCED_SHUTDOWN(mp))
1089 		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1090 	/*
1091 	 * We have to unpin the pinned buffers so do the
1092 	 * callbacks.
1093 	 */
1094 	xfs_buf_do_callbacks(bp, lip);
1095 	XFS_BUF_SET_FSPRIVATE(bp, NULL);
1096 	XFS_BUF_CLR_IODONE_FUNC(bp);
1097 	XFS_BUF_SET_BRELSE_FUNC(bp,NULL);
1098 	xfs_buf_relse(bp);
1099 }
1100 
1101 
1102 /*
1103  * This is the iodone() function for buffers which have been
1104  * logged.  It is called when they are eventually flushed out.
1105  * It should remove the buf item from the AIL, and free the buf item.
1106  * It is called by xfs_buf_iodone_callbacks() above which will take
1107  * care of cleaning up the buffer itself.
1108  */
1109 /* ARGSUSED */
1110 void
1111 xfs_buf_iodone(
1112 	xfs_buf_t		*bp,
1113 	xfs_buf_log_item_t	*bip)
1114 {
1115 	struct xfs_mount	*mp;
1116 	SPLDECL(s);
1117 
1118 	ASSERT(bip->bli_buf == bp);
1119 
1120 	mp = bip->bli_item.li_mountp;
1121 
1122 	/*
1123 	 * If we are forcibly shutting down, this may well be
1124 	 * off the AIL already. That's because we simulate the
1125 	 * log-committed callbacks to unpin these buffers. Or we may never
1126 	 * have put this item on AIL because of the transaction was
1127 	 * aborted forcibly. xfs_trans_delete_ail() takes care of these.
1128 	 *
1129 	 * Either way, AIL is useless if we're forcing a shutdown.
1130 	 */
1131 	AIL_LOCK(mp,s);
1132 	/*
1133 	 * xfs_trans_delete_ail() drops the AIL lock.
1134 	 */
1135 	xfs_trans_delete_ail(mp, (xfs_log_item_t *)bip, s);
1136 
1137 #ifdef XFS_TRANS_DEBUG
1138 	kmem_free(bip->bli_orig, XFS_BUF_COUNT(bp));
1139 	bip->bli_orig = NULL;
1140 	kmem_free(bip->bli_logged, XFS_BUF_COUNT(bp) / NBBY);
1141 	bip->bli_logged = NULL;
1142 #endif /* XFS_TRANS_DEBUG */
1143 
1144 #ifdef XFS_BLI_TRACE
1145 	ktrace_free(bip->bli_trace);
1146 #endif
1147 	kmem_zone_free(xfs_buf_item_zone, bip);
1148 }
1149 
1150 #if defined(XFS_BLI_TRACE)
1151 void
1152 xfs_buf_item_trace(
1153 	char			*id,
1154 	xfs_buf_log_item_t	*bip)
1155 {
1156 	xfs_buf_t		*bp;
1157 	ASSERT(bip->bli_trace != NULL);
1158 
1159 	bp = bip->bli_buf;
1160 	ktrace_enter(bip->bli_trace,
1161 		     (void *)id,
1162 		     (void *)bip->bli_buf,
1163 		     (void *)((unsigned long)bip->bli_flags),
1164 		     (void *)((unsigned long)bip->bli_recur),
1165 		     (void *)((unsigned long)atomic_read(&bip->bli_refcount)),
1166 		     (void *)((unsigned long)
1167 				(0xFFFFFFFF & XFS_BUF_ADDR(bp) >> 32)),
1168 		     (void *)((unsigned long)(0xFFFFFFFF & XFS_BUF_ADDR(bp))),
1169 		     (void *)((unsigned long)XFS_BUF_COUNT(bp)),
1170 		     (void *)((unsigned long)XFS_BUF_BFLAGS(bp)),
1171 		     XFS_BUF_FSPRIVATE(bp, void *),
1172 		     XFS_BUF_FSPRIVATE2(bp, void *),
1173 		     (void *)(unsigned long)XFS_BUF_ISPINNED(bp),
1174 		     (void *)XFS_BUF_IODONE_FUNC(bp),
1175 		     (void *)((unsigned long)(XFS_BUF_VALUSEMA(bp))),
1176 		     (void *)bip->bli_item.li_desc,
1177 		     (void *)((unsigned long)bip->bli_item.li_flags));
1178 }
1179 #endif /* XFS_BLI_TRACE */
1180