HAMMER VFS - Version 4 part 1/many - UNDO FIFO layout work.
[dragonfly.git] / sys / vfs / hammer / hammer_io.c
CommitLineData
66325755 1/*
b84de5af 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
66325755
MD
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
e83ca595 34 * $DragonFly: src/sys/vfs/hammer/hammer_io.c,v 1.55 2008/09/15 17:02:49 dillon Exp $
66325755
MD
35 */
36/*
37 * IO Primitives and buffer cache management
38 *
39 * All major data-tracking structures in HAMMER contain a struct hammer_io
40 * which is used to manage their backing store. We use filesystem buffers
41 * for backing store and we leave them passively associated with their
42 * HAMMER structures.
43 *
9f5097dc 44 * If the kernel tries to destroy a passively associated buf which we cannot
66325755
MD
45 * yet let go we set B_LOCKED in the buffer and then actively released it
46 * later when we can.
47 */
48
49#include "hammer.h"
50#include <sys/fcntl.h>
51#include <sys/nlookup.h>
52#include <sys/buf.h>
53#include <sys/buf2.h>
54
10a5d1ba 55static void hammer_io_modify(hammer_io_t io, int count);
055f5ff8 56static void hammer_io_deallocate(struct buf *bp);
1b0ab2c3
MD
57#if 0
58static void hammer_io_direct_read_complete(struct bio *nbio);
59#endif
60static void hammer_io_direct_write_complete(struct bio *nbio);
43c665ae 61static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data);
cdb6e4e6 62static void hammer_io_set_modlist(struct hammer_io *io);
748efb59 63static void hammer_io_flush_mark(hammer_volume_t volume);
748efb59 64
055f5ff8
MD
65
66/*
10a5d1ba
MD
67 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
68 * an existing hammer_io structure which may have switched to another type.
055f5ff8
MD
69 */
70void
748efb59 71hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type)
055f5ff8 72{
748efb59
MD
73 io->volume = volume;
74 io->hmp = volume->io.hmp;
055f5ff8 75 io->type = type;
055f5ff8
MD
76}
77
66325755 78/*
fbc6e32a 79 * Helper routine to disassociate a buffer cache buffer from an I/O
ecca949a 80 * structure. The buffer is unlocked and marked appropriate for reclamation.
055f5ff8
MD
81 *
82 * The io may have 0 or 1 references depending on who called us. The
83 * caller is responsible for dealing with the refs.
84 *
85 * This call can only be made when no action is required on the buffer.
ecca949a
MD
86 *
87 * The caller must own the buffer and the IO must indicate that the
88 * structure no longer owns it (io.released != 0).
66325755
MD
89 */
90static void
ecca949a 91hammer_io_disassociate(hammer_io_structure_t iou)
66325755 92{
055f5ff8 93 struct buf *bp = iou->io.bp;
66325755 94
ecca949a 95 KKASSERT(iou->io.released);
b58c6388 96 KKASSERT(iou->io.modified == 0);
af209b0f 97 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou);
4d75d829 98 buf_dep_init(bp);
055f5ff8 99 iou->io.bp = NULL;
9f5097dc
MD
100
101 /*
102 * If the buffer was locked someone wanted to get rid of it.
103 */
a99b9ea2
MD
104 if (bp->b_flags & B_LOCKED) {
105 --hammer_count_io_locked;
9f5097dc 106 bp->b_flags &= ~B_LOCKED;
a99b9ea2 107 }
ecca949a
MD
108 if (iou->io.reclaim) {
109 bp->b_flags |= B_NOCACHE|B_RELBUF;
110 iou->io.reclaim = 0;
055f5ff8 111 }
66325755 112
055f5ff8 113 switch(iou->io.type) {
66325755 114 case HAMMER_STRUCTURE_VOLUME:
055f5ff8 115 iou->volume.ondisk = NULL;
66325755 116 break;
10a5d1ba
MD
117 case HAMMER_STRUCTURE_DATA_BUFFER:
118 case HAMMER_STRUCTURE_META_BUFFER:
119 case HAMMER_STRUCTURE_UNDO_BUFFER:
055f5ff8 120 iou->buffer.ondisk = NULL;
66325755
MD
121 break;
122 }
fbc6e32a
MD
123}
124
125/*
055f5ff8 126 * Wait for any physical IO to complete
ae8e83e6
MD
127 *
128 * XXX we aren't interlocked against a spinlock or anything so there
129 * is a small window in the interlock / io->running == 0 test.
fbc6e32a 130 */
1b0ab2c3 131void
055f5ff8 132hammer_io_wait(hammer_io_t io)
fbc6e32a 133{
055f5ff8 134 if (io->running) {
055f5ff8 135 for (;;) {
ae8e83e6
MD
136 io->waiting = 1;
137 tsleep_interlock(io, 0);
055f5ff8
MD
138 if (io->running == 0)
139 break;
ae8e83e6 140 tsleep(io, PINTERLOCKED, "hmrflw", hz);
055f5ff8
MD
141 if (io->running == 0)
142 break;
143 }
66325755
MD
144 }
145}
146
af209b0f
MD
147/*
148 * Wait for all hammer_io-initated write I/O's to complete. This is not
149 * supposed to count direct I/O's but some can leak through (for
150 * non-full-sized direct I/Os).
151 */
152void
153hammer_io_wait_all(hammer_mount_t hmp, const char *ident)
154{
748efb59 155 hammer_io_flush_sync(hmp);
af209b0f 156 crit_enter();
f5a07a7a
MD
157 while (hmp->io_running_space)
158 tsleep(&hmp->io_running_space, 0, ident, 0);
af209b0f
MD
159 crit_exit();
160}
161
2f85fa4d
MD
162#define HAMMER_MAXRA 4
163
66325755 164/*
10a5d1ba
MD
165 * Load bp for a HAMMER structure. The io must be exclusively locked by
166 * the caller.
2f85fa4d 167 *
a99b9ea2
MD
168 * This routine is mostly used on meta-data and small-data blocks. Generally
169 * speaking HAMMER assumes some locality of reference and will cluster
170 * a 64K read.
af209b0f
MD
171 *
172 * Note that clustering occurs at the device layer, not the logical layer.
173 * If the buffers do not apply to the current operation they may apply to
174 * some other.
66325755
MD
175 */
176int
2f85fa4d 177hammer_io_read(struct vnode *devvp, struct hammer_io *io, hammer_off_t limit)
66325755
MD
178{
179 struct buf *bp;
2f85fa4d 180 int error;
66325755
MD
181
182 if ((bp = io->bp) == NULL) {
f5a07a7a 183 hammer_count_io_running_read += io->bytes;
ce0138a6
MD
184 if (hammer_cluster_enable) {
185 error = cluster_read(devvp, limit,
186 io->offset, io->bytes,
187 HAMMER_CLUSTER_SIZE,
188 HAMMER_CLUSTER_BUFS, &io->bp);
189 } else {
190 error = bread(devvp, io->offset, io->bytes, &io->bp);
191 }
192 hammer_stats_disk_read += io->bytes;
f5a07a7a 193 hammer_count_io_running_read -= io->bytes;
cdb6e4e6
MD
194
195 /*
196 * The code generally assumes b_ops/b_dep has been set-up,
197 * even if we error out here.
198 */
199 bp = io->bp;
200 bp->b_ops = &hammer_bioops;
201 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
202 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
203 BUF_KERNPROC(bp);
10a5d1ba
MD
204 KKASSERT(io->modified == 0);
205 KKASSERT(io->running == 0);
206 KKASSERT(io->waiting == 0);
66325755
MD
207 io->released = 0; /* we hold an active lock on bp */
208 } else {
209 error = 0;
210 }
211 return(error);
212}
213
214/*
215 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
10a5d1ba
MD
216 * Must be called with the IO exclusively locked.
217 *
66325755 218 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
10a5d1ba
MD
219 * I/O by forcing the buffer to not be in a released state before calling
220 * it.
055f5ff8 221 *
10a5d1ba
MD
222 * This function will also mark the IO as modified but it will not
223 * increment the modify_refs count.
66325755
MD
224 */
225int
226hammer_io_new(struct vnode *devvp, struct hammer_io *io)
227{
228 struct buf *bp;
229
230 if ((bp = io->bp) == NULL) {
4a2796f3 231 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0);
66325755
MD
232 bp = io->bp;
233 bp->b_ops = &hammer_bioops;
af209b0f 234 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
66325755 235 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
055f5ff8 236 io->released = 0;
10a5d1ba 237 KKASSERT(io->running == 0);
055f5ff8 238 io->waiting = 0;
66325755
MD
239 BUF_KERNPROC(bp);
240 } else {
241 if (io->released) {
242 regetblk(bp);
66325755 243 BUF_KERNPROC(bp);
d113fda1 244 io->released = 0;
66325755
MD
245 }
246 }
10a5d1ba 247 hammer_io_modify(io, 0);
66325755
MD
248 vfs_bio_clrbuf(bp);
249 return(0);
250}
251
252/*
0e8bd897
MD
253 * Advance the activity count on the underlying buffer because
254 * HAMMER does not getblk/brelse on every access.
255 */
256void
257hammer_io_advance(struct hammer_io *io)
258{
259 if (io->bp)
260 buf_act_advance(io->bp);
261}
262
263/*
47637bff 264 * Remove potential device level aliases against buffers managed by high level
362ec2dc
MD
265 * vnodes. Aliases can also be created due to mixed buffer sizes or via
266 * direct access to the backing store device.
e469566b
MD
267 *
268 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
269 * does not exist its backing VM pages might, and we have to invalidate
270 * those as well or a getblk() will reinstate them.
362ec2dc
MD
271 *
272 * Buffer cache buffers associated with hammer_buffers cannot be
273 * invalidated.
47637bff 274 */
362ec2dc 275int
47637bff
MD
276hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset)
277{
cebe9493 278 hammer_io_structure_t iou;
47637bff
MD
279 hammer_off_t phys_offset;
280 struct buf *bp;
362ec2dc 281 int error;
47637bff
MD
282
283 phys_offset = volume->ondisk->vol_buf_beg +
284 (zone2_offset & HAMMER_OFF_SHORT_MASK);
4a2796f3 285 crit_enter();
b1c20cfa 286 if ((bp = findblk(volume->devvp, phys_offset, FINDBLK_TEST)) != NULL)
4a2796f3 287 bp = getblk(volume->devvp, phys_offset, bp->b_bufsize, 0, 0);
e469566b
MD
288 else
289 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0);
290 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) {
362ec2dc 291#if 0
5c8d05e2 292 hammer_ref(&iou->io.lock);
e469566b
MD
293 hammer_io_clear_modify(&iou->io, 1);
294 bundirty(bp);
e83ca595
MD
295 iou->io.released = 0;
296 BUF_KERNPROC(bp);
e469566b 297 iou->io.reclaim = 1;
5c8d05e2 298 iou->io.waitdep = 1;
e83ca595 299 KKASSERT(iou->io.lock.refs == 1);
5c8d05e2
MD
300 hammer_rel_buffer(&iou->buffer, 0);
301 /*hammer_io_deallocate(bp);*/
362ec2dc 302#endif
04b04ca6 303 bqrelse(bp);
362ec2dc 304 error = EAGAIN;
e469566b
MD
305 } else {
306 KKASSERT((bp->b_flags & B_LOCKED) == 0);
307 bundirty(bp);
308 bp->b_flags |= B_NOCACHE|B_RELBUF;
e83ca595 309 brelse(bp);
362ec2dc 310 error = 0;
47637bff 311 }
4a2796f3 312 crit_exit();
362ec2dc 313 return(error);
47637bff
MD
314}
315
316/*
b3deaf57 317 * This routine is called on the last reference to a hammer structure.
ecca949a 318 * The io is usually interlocked with io.loading and io.refs must be 1.
b3deaf57 319 *
ecca949a
MD
320 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
321 * simply means the caller finishes decrementing the ref-count on the
322 * IO structure then brelse()'s the bp. The bp may or may not still be
323 * passively associated with the IO.
324 *
325 * The only requirement here is that modified meta-data and volume-header
326 * buffer may NOT be disassociated from the IO structure, and consequently
327 * we also leave such buffers actively associated with the IO if they already
328 * are (since the kernel can't do anything with them anyway). Only the
329 * flusher is allowed to write such buffers out. Modified pure-data and
330 * undo buffers are returned to the kernel but left passively associated
331 * so we can track when the kernel writes the bp out.
66325755 332 */
ecca949a 333struct buf *
09ac686b 334hammer_io_release(struct hammer_io *io, int flush)
66325755 335{
9f5097dc 336 union hammer_io_structure *iou = (void *)io;
66325755
MD
337 struct buf *bp;
338
055f5ff8 339 if ((bp = io->bp) == NULL)
ecca949a 340 return(NULL);
fbc6e32a 341
055f5ff8 342 /*
10a5d1ba
MD
343 * Try to flush a dirty IO to disk if asked to by the
344 * caller or if the kernel tried to flush the buffer in the past.
055f5ff8 345 *
10a5d1ba
MD
346 * Kernel-initiated flushes are only allowed for pure-data buffers.
347 * meta-data and volume buffers can only be flushed explicitly
348 * by HAMMER.
055f5ff8 349 */
10a5d1ba 350 if (io->modified) {
09ac686b 351 if (flush) {
710733a6 352 hammer_io_flush(io, 0);
10a5d1ba
MD
353 } else if (bp->b_flags & B_LOCKED) {
354 switch(io->type) {
355 case HAMMER_STRUCTURE_DATA_BUFFER:
710733a6
MD
356 hammer_io_flush(io, 0);
357 break;
10a5d1ba 358 case HAMMER_STRUCTURE_UNDO_BUFFER:
710733a6 359 hammer_io_flush(io, hammer_undo_reclaim(io));
10a5d1ba
MD
360 break;
361 default:
362 break;
363 }
364 } /* else no explicit request to flush the buffer */
365 }
fbc6e32a 366
055f5ff8 367 /*
5c8d05e2
MD
368 * Wait for the IO to complete if asked to. This occurs when
369 * the buffer must be disposed of definitively during an umount
370 * or buffer invalidation.
055f5ff8 371 */
b58c6388 372 if (io->waitdep && io->running) {
055f5ff8
MD
373 hammer_io_wait(io);
374 }
375
376 /*
10a5d1ba
MD
377 * Return control of the buffer to the kernel (with the provisio
378 * that our bioops can override kernel decisions with regards to
379 * the buffer).
055f5ff8 380 */
cebe9493 381 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) {
10a5d1ba
MD
382 /*
383 * Always disassociate the bp if an explicit flush
384 * was requested and the IO completed with no error
385 * (so unmount can really clean up the structure).
386 */
055f5ff8 387 if (io->released) {
b3deaf57 388 regetblk(bp);
46fe7ae1 389 BUF_KERNPROC(bp);
ecca949a
MD
390 } else {
391 io->released = 1;
055f5ff8 392 }
ecca949a
MD
393 hammer_io_disassociate((hammer_io_structure_t)io);
394 /* return the bp */
055f5ff8 395 } else if (io->modified) {
10a5d1ba 396 /*
ecca949a
MD
397 * Only certain IO types can be released to the kernel if
398 * the buffer has been modified.
399 *
400 * volume and meta-data IO types may only be explicitly
401 * flushed by HAMMER.
10a5d1ba
MD
402 */
403 switch(io->type) {
404 case HAMMER_STRUCTURE_DATA_BUFFER:
405 case HAMMER_STRUCTURE_UNDO_BUFFER:
406 if (io->released == 0) {
407 io->released = 1;
408 bdwrite(bp);
409 }
410 break;
411 default:
412 break;
055f5ff8 413 }
ecca949a 414 bp = NULL; /* bp left associated */
055f5ff8 415 } else if (io->released == 0) {
10a5d1ba
MD
416 /*
417 * Clean buffers can be generally released to the kernel.
418 * We leave the bp passively associated with the HAMMER
419 * structure and use bioops to disconnect it later on
420 * if the kernel wants to discard the buffer.
ecca949a
MD
421 *
422 * We can steal the structure's ownership of the bp.
10a5d1ba 423 */
ecca949a 424 io->released = 1;
9f5097dc 425 if (bp->b_flags & B_LOCKED) {
ecca949a
MD
426 hammer_io_disassociate(iou);
427 /* return the bp */
9f5097dc 428 } else {
cebe9493 429 if (io->reclaim) {
ecca949a
MD
430 hammer_io_disassociate(iou);
431 /* return the bp */
cebe9493 432 } else {
ecca949a 433 /* return the bp (bp passively associated) */
cebe9493 434 }
9f5097dc 435 }
19b97e01
MD
436 } else {
437 /*
af209b0f
MD
438 * A released buffer is passively associate with our
439 * hammer_io structure. The kernel cannot destroy it
440 * without making a bioops call. If the kernel (B_LOCKED)
441 * or we (reclaim) requested that the buffer be destroyed
442 * we destroy it, otherwise we do a quick get/release to
443 * reset its position in the kernel's LRU list.
444 *
445 * Leaving the buffer passively associated allows us to
446 * use the kernel's LRU buffer flushing mechanisms rather
447 * then rolling our own.
cb51be26
MD
448 *
449 * XXX there are two ways of doing this. We can re-acquire
450 * and passively release to reset the LRU, or not.
19b97e01 451 */
af209b0f 452 if (io->running == 0) {
19b97e01 453 regetblk(bp);
cebe9493 454 if ((bp->b_flags & B_LOCKED) || io->reclaim) {
ecca949a
MD
455 hammer_io_disassociate(iou);
456 /* return the bp */
9f5097dc 457 } else {
ecca949a 458 /* return the bp (bp passively associated) */
9f5097dc 459 }
ecca949a
MD
460 } else {
461 /*
462 * bp is left passively associated but we do not
463 * try to reacquire it. Interactions with the io
464 * structure will occur on completion of the bp's
465 * I/O.
466 */
467 bp = NULL;
19b97e01 468 }
66325755 469 }
ecca949a 470 return(bp);
66325755
MD
471}
472
473/*
b33e2cc0
MD
474 * This routine is called with a locked IO when a flush is desired and
475 * no other references to the structure exists other then ours. This
476 * routine is ONLY called when HAMMER believes it is safe to flush a
477 * potentially modified buffer out.
fbc6e32a
MD
478 */
479void
710733a6 480hammer_io_flush(struct hammer_io *io, int reclaim)
fbc6e32a
MD
481{
482 struct buf *bp;
fbc6e32a 483
055f5ff8 484 /*
10a5d1ba 485 * Degenerate case - nothing to flush if nothing is dirty.
055f5ff8 486 */
b58c6388 487 if (io->modified == 0) {
055f5ff8 488 return;
b58c6388 489 }
055f5ff8
MD
490
491 KKASSERT(io->bp);
9f5097dc 492 KKASSERT(io->modify_refs <= 0);
055f5ff8 493
b33e2cc0 494 /*
77062c8a
MD
495 * Acquire ownership of the bp, particularly before we clear our
496 * modified flag.
497 *
498 * We are going to bawrite() this bp. Don't leave a window where
499 * io->released is set, we actually own the bp rather then our
500 * buffer.
501 */
502 bp = io->bp;
503 if (io->released) {
504 regetblk(bp);
505 /* BUF_KERNPROC(io->bp); */
506 /* io->released = 0; */
507 KKASSERT(io->released);
508 KKASSERT(io->bp == bp);
509 }
510 io->released = 1;
511
710733a6
MD
512 if (reclaim) {
513 io->reclaim = 1;
514 if ((bp->b_flags & B_LOCKED) == 0) {
515 bp->b_flags |= B_LOCKED;
516 ++hammer_count_io_locked;
517 }
518 }
519
77062c8a 520 /*
10a5d1ba
MD
521 * Acquire exclusive access to the bp and then clear the modified
522 * state of the buffer prior to issuing I/O to interlock any
523 * modifications made while the I/O is in progress. This shouldn't
524 * happen anyway but losing data would be worse. The modified bit
525 * will be rechecked after the IO completes.
526 *
4a2796f3
MD
527 * NOTE: This call also finalizes the buffer's content (inval == 0).
528 *
b33e2cc0
MD
529 * This is only legal when lock.refs == 1 (otherwise we might clear
530 * the modified bit while there are still users of the cluster
531 * modifying the data).
532 *
b33e2cc0
MD
533 * Do this before potentially blocking so any attempt to modify the
534 * ondisk while we are blocked blocks waiting for us.
535 */
5c8d05e2 536 hammer_ref(&io->lock);
4a2796f3 537 hammer_io_clear_modify(io, 0);
5c8d05e2 538 hammer_unref(&io->lock);
bcac4bbb
MD
539
540 /*
10a5d1ba
MD
541 * Transfer ownership to the kernel and initiate I/O.
542 */
055f5ff8 543 io->running = 1;
f5a07a7a
MD
544 io->hmp->io_running_space += io->bytes;
545 hammer_count_io_running_write += io->bytes;
055f5ff8 546 bawrite(bp);
748efb59 547 hammer_io_flush_mark(io->volume);
fbc6e32a
MD
548}
549
055f5ff8
MD
550/************************************************************************
551 * BUFFER DIRTYING *
552 ************************************************************************
553 *
554 * These routines deal with dependancies created when IO buffers get
555 * modified. The caller must call hammer_modify_*() on a referenced
556 * HAMMER structure prior to modifying its on-disk data.
0b075555 557 *
055f5ff8
MD
558 * Any intent to modify an IO buffer acquires the related bp and imposes
559 * various write ordering dependancies.
0b075555 560 */
055f5ff8
MD
561
562/*
10a5d1ba
MD
563 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
564 * are locked until the flusher can deal with them, pure data buffers
565 * can be written out.
055f5ff8 566 */
10a5d1ba 567static
b58c6388 568void
10a5d1ba 569hammer_io_modify(hammer_io_t io, int count)
0b075555 570{
46fe7ae1 571 /*
9f5097dc
MD
572 * io->modify_refs must be >= 0
573 */
574 while (io->modify_refs < 0) {
575 io->waitmod = 1;
576 tsleep(io, 0, "hmrmod", 0);
577 }
578
579 /*
46fe7ae1
MD
580 * Shortcut if nothing to do.
581 */
0b075555 582 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
10a5d1ba 583 io->modify_refs += count;
b58c6388
MD
584 if (io->modified && io->released == 0)
585 return;
46fe7ae1
MD
586
587 hammer_lock_ex(&io->lock);
10a5d1ba 588 if (io->modified == 0) {
cdb6e4e6 589 hammer_io_set_modlist(io);
10a5d1ba
MD
590 io->modified = 1;
591 }
46fe7ae1
MD
592 if (io->released) {
593 regetblk(io->bp);
594 BUF_KERNPROC(io->bp);
595 io->released = 0;
596 KKASSERT(io->modified != 0);
597 }
46fe7ae1 598 hammer_unlock(&io->lock);
055f5ff8
MD
599}
600
10a5d1ba
MD
601static __inline
602void
603hammer_io_modify_done(hammer_io_t io)
604{
605 KKASSERT(io->modify_refs > 0);
606 --io->modify_refs;
9f5097dc
MD
607 if (io->modify_refs == 0 && io->waitmod) {
608 io->waitmod = 0;
609 wakeup(io);
610 }
611}
612
613void
614hammer_io_write_interlock(hammer_io_t io)
615{
616 while (io->modify_refs != 0) {
617 io->waitmod = 1;
618 tsleep(io, 0, "hmrmod", 0);
619 }
620 io->modify_refs = -1;
621}
622
623void
624hammer_io_done_interlock(hammer_io_t io)
625{
626 KKASSERT(io->modify_refs == -1);
627 io->modify_refs = 0;
628 if (io->waitmod) {
629 io->waitmod = 0;
630 wakeup(io);
631 }
10a5d1ba
MD
632}
633
2f85fa4d
MD
634/*
635 * Caller intends to modify a volume's ondisk structure.
636 *
637 * This is only allowed if we are the flusher or we have a ref on the
638 * sync_lock.
639 */
055f5ff8 640void
36f82b23
MD
641hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
642 void *base, int len)
055f5ff8 643{
2f85fa4d 644 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
055f5ff8 645
2f85fa4d 646 hammer_io_modify(&volume->io, 1);
47197d71
MD
647 if (len) {
648 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
649 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
02428fb6 650 hammer_generate_undo(trans,
47197d71
MD
651 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
652 base, len);
653 }
0b075555
MD
654}
655
055f5ff8 656/*
2f85fa4d
MD
657 * Caller intends to modify a buffer's ondisk structure.
658 *
659 * This is only allowed if we are the flusher or we have a ref on the
660 * sync_lock.
055f5ff8 661 */
0b075555 662void
36f82b23
MD
663hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
664 void *base, int len)
46fe7ae1 665{
2f85fa4d
MD
666 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
667
10a5d1ba 668 hammer_io_modify(&buffer->io, 1);
47197d71
MD
669 if (len) {
670 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
671 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
02428fb6 672 hammer_generate_undo(trans,
34d829f7 673 buffer->zone2_offset + rel_offset,
47197d71
MD
674 base, len);
675 }
46fe7ae1
MD
676}
677
10a5d1ba
MD
678void
679hammer_modify_volume_done(hammer_volume_t volume)
680{
681 hammer_io_modify_done(&volume->io);
682}
683
684void
685hammer_modify_buffer_done(hammer_buffer_t buffer)
686{
687 hammer_io_modify_done(&buffer->io);
688}
689
46fe7ae1 690/*
4a2796f3
MD
691 * Mark an entity as not being dirty any more and finalize any
692 * delayed adjustments to the buffer.
693 *
694 * Delayed adjustments are an important performance enhancement, allowing
695 * us to avoid recalculating B-Tree node CRCs over and over again when
696 * making bulk-modifications to the B-Tree.
697 *
698 * If inval is non-zero delayed adjustments are ignored.
5c8d05e2
MD
699 *
700 * This routine may dereference related btree nodes and cause the
701 * buffer to be dereferenced. The caller must own a reference on io.
61aeeb33
MD
702 */
703void
4a2796f3 704hammer_io_clear_modify(struct hammer_io *io, int inval)
61aeeb33 705{
4a2796f3
MD
706 if (io->modified == 0)
707 return;
708
709 /*
710 * Take us off the mod-list and clear the modified bit.
711 */
712 KKASSERT(io->mod_list != NULL);
713 if (io->mod_list == &io->hmp->volu_list ||
714 io->mod_list == &io->hmp->meta_list) {
f5a07a7a
MD
715 io->hmp->locked_dirty_space -= io->bytes;
716 hammer_count_dirtybufspace -= io->bytes;
4a2796f3
MD
717 }
718 TAILQ_REMOVE(io->mod_list, io, mod_entry);
719 io->mod_list = NULL;
720 io->modified = 0;
721
722 /*
723 * If this bit is not set there are no delayed adjustments.
724 */
725 if (io->gencrc == 0)
726 return;
727 io->gencrc = 0;
728
729 /*
730 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
731 * on the node (& underlying buffer). Release the node after clearing
732 * the flag.
733 */
734 if (io->type == HAMMER_STRUCTURE_META_BUFFER) {
735 hammer_buffer_t buffer = (void *)io;
736 hammer_node_t node;
737
738restart:
739 TAILQ_FOREACH(node, &buffer->clist, entry) {
740 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0)
741 continue;
742 node->flags &= ~HAMMER_NODE_NEEDSCRC;
743 KKASSERT(node->ondisk);
744 if (inval == 0)
745 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
746 hammer_rel_node(node);
747 goto restart;
61aeeb33 748 }
cebe9493 749 }
5c8d05e2
MD
750 /* caller must still have ref on io */
751 KKASSERT(io->lock.refs > 0);
cebe9493
MD
752}
753
754/*
755 * Clear the IO's modify list. Even though the IO is no longer modified
756 * it may still be on the lose_list. This routine is called just before
757 * the governing hammer_buffer is destroyed.
758 */
759void
760hammer_io_clear_modlist(struct hammer_io *io)
761{
4a2796f3 762 KKASSERT(io->modified == 0);
cebe9493 763 if (io->mod_list) {
a99b9ea2 764 crit_enter(); /* biodone race against list */
cebe9493
MD
765 KKASSERT(io->mod_list == &io->hmp->lose_list);
766 TAILQ_REMOVE(io->mod_list, io, mod_entry);
767 io->mod_list = NULL;
a99b9ea2 768 crit_exit();
61aeeb33
MD
769 }
770}
771
cdb6e4e6
MD
772static void
773hammer_io_set_modlist(struct hammer_io *io)
774{
775 struct hammer_mount *hmp = io->hmp;
776
777 KKASSERT(io->mod_list == NULL);
778
779 switch(io->type) {
780 case HAMMER_STRUCTURE_VOLUME:
781 io->mod_list = &hmp->volu_list;
782 hmp->locked_dirty_space += io->bytes;
783 hammer_count_dirtybufspace += io->bytes;
784 break;
785 case HAMMER_STRUCTURE_META_BUFFER:
786 io->mod_list = &hmp->meta_list;
787 hmp->locked_dirty_space += io->bytes;
788 hammer_count_dirtybufspace += io->bytes;
789 break;
790 case HAMMER_STRUCTURE_UNDO_BUFFER:
791 io->mod_list = &hmp->undo_list;
792 break;
793 case HAMMER_STRUCTURE_DATA_BUFFER:
794 io->mod_list = &hmp->data_list;
795 break;
796 }
797 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry);
798}
799
055f5ff8
MD
800/************************************************************************
801 * HAMMER_BIOOPS *
802 ************************************************************************
803 *
66325755
MD
804 */
805
806/*
055f5ff8 807 * Pre-IO initiation kernel callback - cluster build only
66325755 808 */
66325755
MD
809static void
810hammer_io_start(struct buf *bp)
811{
812}
813
055f5ff8 814/*
7bc5b8c2 815 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
b33e2cc0
MD
816 *
817 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
818 * may also be set if we were marking a cluster header open. Only remove
819 * our dependancy if the modified bit is clear.
055f5ff8 820 */
66325755
MD
821static void
822hammer_io_complete(struct buf *bp)
823{
055f5ff8
MD
824 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
825
826 KKASSERT(iou->io.released == 1);
fbc6e32a 827
bf3b416b
MD
828 /*
829 * Deal with people waiting for I/O to drain
830 */
f90dde4c 831 if (iou->io.running) {
cdb6e4e6
MD
832 /*
833 * Deal with critical write errors. Once a critical error
834 * has been flagged in hmp the UNDO FIFO will not be updated.
835 * That way crash recover will give us a consistent
836 * filesystem.
837 *
838 * Because of this we can throw away failed UNDO buffers. If
839 * we throw away META or DATA buffers we risk corrupting
840 * the now read-only version of the filesystem visible to
841 * the user. Clear B_ERROR so the buffer is not re-dirtied
842 * by the kernel and ref the io so it doesn't get thrown
843 * away.
844 */
845 if (bp->b_flags & B_ERROR) {
846 hammer_critical_error(iou->io.hmp, NULL, bp->b_error,
847 "while flushing meta-data");
848 switch(iou->io.type) {
849 case HAMMER_STRUCTURE_UNDO_BUFFER:
850 break;
851 default:
852 if (iou->io.ioerror == 0) {
853 iou->io.ioerror = 1;
854 if (iou->io.lock.refs == 0)
855 ++hammer_count_refedbufs;
856 hammer_ref(&iou->io.lock);
857 }
858 break;
859 }
860 bp->b_flags &= ~B_ERROR;
861 bundirty(bp);
862#if 0
863 hammer_io_set_modlist(&iou->io);
864 iou->io.modified = 1;
865#endif
866 }
ce0138a6 867 hammer_stats_disk_write += iou->io.bytes;
f5a07a7a
MD
868 hammer_count_io_running_write -= iou->io.bytes;
869 iou->io.hmp->io_running_space -= iou->io.bytes;
870 if (iou->io.hmp->io_running_space == 0)
871 wakeup(&iou->io.hmp->io_running_space);
872 KKASSERT(iou->io.hmp->io_running_space >= 0);
f90dde4c 873 iou->io.running = 0;
ce0138a6
MD
874 } else {
875 hammer_stats_disk_read += iou->io.bytes;
f90dde4c
MD
876 }
877
055f5ff8
MD
878 if (iou->io.waiting) {
879 iou->io.waiting = 0;
880 wakeup(iou);
881 }
882
883 /*
bf3b416b
MD
884 * If B_LOCKED is set someone wanted to deallocate the bp at some
885 * point, do it now if refs has become zero.
055f5ff8
MD
886 */
887 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
b33e2cc0 888 KKASSERT(iou->io.modified == 0);
a99b9ea2 889 --hammer_count_io_locked;
d5ef456e 890 bp->b_flags &= ~B_LOCKED;
055f5ff8
MD
891 hammer_io_deallocate(bp);
892 /* structure may be dead now */
893 }
66325755
MD
894}
895
896/*
897 * Callback from kernel when it wishes to deallocate a passively
10a5d1ba
MD
898 * associated structure. This mostly occurs with clean buffers
899 * but it may be possible for a holding structure to be marked dirty
7bc5b8c2 900 * while its buffer is passively associated. The caller owns the bp.
66325755
MD
901 *
902 * If we cannot disassociate we set B_LOCKED to prevent the buffer
903 * from getting reused.
46fe7ae1
MD
904 *
905 * WARNING: Because this can be called directly by getnewbuf we cannot
906 * recurse into the tree. If a bp cannot be immediately disassociated
907 * our only recourse is to set B_LOCKED.
7bc5b8c2
MD
908 *
909 * WARNING: This may be called from an interrupt via hammer_io_complete()
66325755
MD
910 */
911static void
912hammer_io_deallocate(struct buf *bp)
913{
055f5ff8 914 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
66325755 915
055f5ff8 916 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
46fe7ae1 917 if (iou->io.lock.refs > 0 || iou->io.modified) {
10a5d1ba
MD
918 /*
919 * It is not legal to disassociate a modified buffer. This
920 * case really shouldn't ever occur.
921 */
055f5ff8 922 bp->b_flags |= B_LOCKED;
a99b9ea2 923 ++hammer_count_io_locked;
055f5ff8 924 } else {
10a5d1ba
MD
925 /*
926 * Disassociate the BP. If the io has no refs left we
927 * have to add it to the loose list.
928 */
ecca949a
MD
929 hammer_io_disassociate(iou);
930 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) {
931 KKASSERT(iou->io.bp == NULL);
10a5d1ba 932 KKASSERT(iou->io.mod_list == NULL);
a99b9ea2 933 crit_enter(); /* biodone race against list */
10a5d1ba
MD
934 iou->io.mod_list = &iou->io.hmp->lose_list;
935 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry);
a99b9ea2 936 crit_exit();
66325755 937 }
66325755 938 }
66325755
MD
939}
940
941static int
942hammer_io_fsync(struct vnode *vp)
943{
944 return(0);
945}
946
947/*
948 * NOTE: will not be called unless we tell the kernel about the
949 * bioops. Unused... we use the mount's VFS_SYNC instead.
950 */
951static int
952hammer_io_sync(struct mount *mp)
953{
954 return(0);
955}
956
957static void
958hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
959{
960}
961
962/*
963 * I/O pre-check for reading and writing. HAMMER only uses this for
964 * B_CACHE buffers so checkread just shouldn't happen, but if it does
965 * allow it.
966 *
fbc6e32a
MD
967 * Writing is a different case. We don't want the kernel to try to write
968 * out a buffer that HAMMER may be modifying passively or which has a
10a5d1ba
MD
969 * dependancy. In addition, kernel-demanded writes can only proceed for
970 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
971 * buffer types can only be explicitly written by the flusher.
fbc6e32a 972 *
10a5d1ba
MD
973 * checkwrite will only be called for bdwrite()n buffers. If we return
974 * success the kernel is guaranteed to initiate the buffer write.
66325755
MD
975 */
976static int
977hammer_io_checkread(struct buf *bp)
978{
979 return(0);
980}
981
982static int
983hammer_io_checkwrite(struct buf *bp)
984{
10a5d1ba 985 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
66325755 986
77062c8a
MD
987 /*
988 * This shouldn't happen under normal operation.
989 */
990 if (io->type == HAMMER_STRUCTURE_VOLUME ||
991 io->type == HAMMER_STRUCTURE_META_BUFFER) {
992 if (!panicstr)
993 panic("hammer_io_checkwrite: illegal buffer");
a99b9ea2
MD
994 if ((bp->b_flags & B_LOCKED) == 0) {
995 bp->b_flags |= B_LOCKED;
996 ++hammer_count_io_locked;
997 }
77062c8a
MD
998 return(1);
999 }
c9b9e29d 1000
055f5ff8 1001 /*
10a5d1ba
MD
1002 * We can only clear the modified bit if the IO is not currently
1003 * undergoing modification. Otherwise we may miss changes.
5c8d05e2
MD
1004 *
1005 * Only data and undo buffers can reach here. These buffers do
1006 * not have terminal crc functions but we temporarily reference
1007 * the IO anyway, just in case.
b33e2cc0 1008 */
5c8d05e2
MD
1009 if (io->modify_refs == 0 && io->modified) {
1010 hammer_ref(&io->lock);
4a2796f3 1011 hammer_io_clear_modify(io, 0);
5c8d05e2
MD
1012 hammer_unref(&io->lock);
1013 } else if (io->modified) {
1014 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER);
1015 }
f90dde4c
MD
1016
1017 /*
1018 * The kernel is going to start the IO, set io->running.
1019 */
1020 KKASSERT(io->running == 0);
1021 io->running = 1;
f5a07a7a
MD
1022 io->hmp->io_running_space += io->bytes;
1023 hammer_count_io_running_write += io->bytes;
055f5ff8 1024 return(0);
66325755
MD
1025}
1026
8cd0a023 1027/*
66325755
MD
1028 * Return non-zero if we wish to delay the kernel's attempt to flush
1029 * this buffer to disk.
1030 */
1031static int
1032hammer_io_countdeps(struct buf *bp, int n)
1033{
1034 return(0);
1035}
1036
1037struct bio_ops hammer_bioops = {
1038 .io_start = hammer_io_start,
1039 .io_complete = hammer_io_complete,
1040 .io_deallocate = hammer_io_deallocate,
1041 .io_fsync = hammer_io_fsync,
1042 .io_sync = hammer_io_sync,
1043 .io_movedeps = hammer_io_movedeps,
1044 .io_countdeps = hammer_io_countdeps,
1045 .io_checkread = hammer_io_checkread,
1046 .io_checkwrite = hammer_io_checkwrite,
1047};
1048
47637bff
MD
1049/************************************************************************
1050 * DIRECT IO OPS *
1051 ************************************************************************
1052 *
1053 * These functions operate directly on the buffer cache buffer associated
1054 * with a front-end vnode rather then a back-end device vnode.
1055 */
1056
1057/*
1058 * Read a buffer associated with a front-end vnode directly from the
1b0ab2c3
MD
1059 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1060 * we validate the CRC.
a99b9ea2 1061 *
1b0ab2c3
MD
1062 * We must check for the presence of a HAMMER buffer to handle the case
1063 * where the reblocker has rewritten the data (which it does via the HAMMER
1064 * buffer system, not via the high-level vnode buffer cache), but not yet
1065 * committed the buffer to the media.
47637bff
MD
1066 */
1067int
1b0ab2c3
MD
1068hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1069 hammer_btree_leaf_elm_t leaf)
47637bff 1070{
1b0ab2c3 1071 hammer_off_t buf_offset;
47637bff
MD
1072 hammer_off_t zone2_offset;
1073 hammer_volume_t volume;
1074 struct buf *bp;
1075 struct bio *nbio;
1076 int vol_no;
1077 int error;
1078
1b0ab2c3
MD
1079 buf_offset = bio->bio_offset;
1080 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1081 HAMMER_ZONE_LARGE_DATA);
1082
1083 /*
1084 * The buffer cache may have an aliased buffer (the reblocker can
1085 * write them). If it does we have to sync any dirty data before
1086 * we can build our direct-read. This is a non-critical code path.
1087 */
1088 bp = bio->bio_buf;
1089 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
4a2796f3 1090
1b0ab2c3
MD
1091 /*
1092 * Resolve to a zone-2 offset. The conversion just requires
1093 * munging the top 4 bits but we want to abstract it anyway
1094 * so the blockmap code can verify the zone assignment.
1095 */
1096 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1097 if (error)
1098 goto done;
43c665ae
MD
1099 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1100 HAMMER_ZONE_RAW_BUFFER);
1101
1b0ab2c3
MD
1102 /*
1103 * Resolve volume and raw-offset for 3rd level bio. The
1104 * offset will be specific to the volume.
1105 */
43c665ae
MD
1106 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1107 volume = hammer_get_volume(hmp, vol_no, &error);
1108 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1109 error = EIO;
1110
47637bff 1111 if (error == 0) {
e469566b
MD
1112 /*
1113 * 3rd level bio
1114 */
43c665ae
MD
1115 nbio = push_bio(bio);
1116 nbio->bio_offset = volume->ondisk->vol_buf_beg +
e469566b 1117 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1b0ab2c3
MD
1118#if 0
1119 /*
1120 * XXX disabled - our CRC check doesn't work if the OS
1121 * does bogus_page replacement on the direct-read.
1122 */
1123 if (leaf && hammer_verify_data) {
1124 nbio->bio_done = hammer_io_direct_read_complete;
1125 nbio->bio_caller_info1.uvalue32 = leaf->data_crc;
1126 }
1127#endif
ce0138a6 1128 hammer_stats_disk_read += bp->b_bufsize;
43c665ae 1129 vn_strategy(volume->devvp, nbio);
47637bff 1130 }
43c665ae 1131 hammer_rel_volume(volume, 0);
1b0ab2c3 1132done:
47637bff 1133 if (error) {
cebe9493 1134 kprintf("hammer_direct_read: failed @ %016llx\n",
973c11b9 1135 (long long)zone2_offset);
47637bff
MD
1136 bp->b_error = error;
1137 bp->b_flags |= B_ERROR;
1138 biodone(bio);
1139 }
1140 return(error);
1141}
1142
1b0ab2c3
MD
1143#if 0
1144/*
1145 * On completion of the BIO this callback must check the data CRC
1146 * and chain to the previous bio.
1147 */
1148static
1149void
1150hammer_io_direct_read_complete(struct bio *nbio)
1151{
1152 struct bio *obio;
1153 struct buf *bp;
1154 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32;
1155
1156 bp = nbio->bio_buf;
1157 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) {
1158 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1159 nbio->bio_offset, bp->b_bufsize);
1160 if (hammer_debug_debug)
1161 Debugger("");
1162 bp->b_flags |= B_ERROR;
1163 bp->b_error = EIO;
1164 }
1165 obio = pop_bio(nbio);
1166 biodone(obio);
1167}
1168#endif
1169
47637bff
MD
1170/*
1171 * Write a buffer associated with a front-end vnode directly to the
1172 * disk media. The bio may be issued asynchronously.
1b0ab2c3
MD
1173 *
1174 * The BIO is associated with the specified record and RECF_DIRECT_IO
e469566b 1175 * is set. The recorded is added to its object.
47637bff
MD
1176 */
1177int
1b0ab2c3 1178hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
47637bff
MD
1179 struct bio *bio)
1180{
1b0ab2c3 1181 hammer_btree_leaf_elm_t leaf = &record->leaf;
0832c9bb 1182 hammer_off_t buf_offset;
47637bff
MD
1183 hammer_off_t zone2_offset;
1184 hammer_volume_t volume;
0832c9bb 1185 hammer_buffer_t buffer;
47637bff
MD
1186 struct buf *bp;
1187 struct bio *nbio;
0832c9bb 1188 char *ptr;
47637bff
MD
1189 int vol_no;
1190 int error;
1191
0832c9bb
MD
1192 buf_offset = leaf->data_offset;
1193
1194 KKASSERT(buf_offset > HAMMER_ZONE_BTREE);
47637bff
MD
1195 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE);
1196
0832c9bb 1197 if ((buf_offset & HAMMER_BUFMASK) == 0 &&
4a2796f3 1198 leaf->data_len >= HAMMER_BUFSIZE) {
0832c9bb
MD
1199 /*
1200 * We are using the vnode's bio to write directly to the
1201 * media, any hammer_buffer at the same zone-X offset will
1202 * now have stale data.
1203 */
1204 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
47637bff
MD
1205 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1206 volume = hammer_get_volume(hmp, vol_no, &error);
1207
1208 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1209 error = EIO;
1210 if (error == 0) {
0832c9bb 1211 bp = bio->bio_buf;
4a2796f3 1212 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0);
e469566b 1213 /*
4a2796f3
MD
1214 hammer_del_buffers(hmp, buf_offset,
1215 zone2_offset, bp->b_bufsize);
e469566b 1216 */
1b0ab2c3 1217
43c665ae
MD
1218 /*
1219 * Second level bio - cached zone2 offset.
1b0ab2c3
MD
1220 *
1221 * (We can put our bio_done function in either the
1222 * 2nd or 3rd level).
43c665ae 1223 */
47637bff 1224 nbio = push_bio(bio);
43c665ae 1225 nbio->bio_offset = zone2_offset;
1b0ab2c3
MD
1226 nbio->bio_done = hammer_io_direct_write_complete;
1227 nbio->bio_caller_info1.ptr = record;
e469566b
MD
1228 record->zone2_offset = zone2_offset;
1229 record->flags |= HAMMER_RECF_DIRECT_IO |
1230 HAMMER_RECF_DIRECT_INVAL;
43c665ae
MD
1231
1232 /*
1233 * Third level bio - raw offset specific to the
1234 * correct volume.
1235 */
1236 zone2_offset &= HAMMER_OFF_SHORT_MASK;
1237 nbio = push_bio(nbio);
47637bff 1238 nbio->bio_offset = volume->ondisk->vol_buf_beg +
0832c9bb 1239 zone2_offset;
ce0138a6 1240 hammer_stats_disk_write += bp->b_bufsize;
47637bff 1241 vn_strategy(volume->devvp, nbio);
748efb59 1242 hammer_io_flush_mark(volume);
47637bff
MD
1243 }
1244 hammer_rel_volume(volume, 0);
0832c9bb 1245 } else {
1b0ab2c3
MD
1246 /*
1247 * Must fit in a standard HAMMER buffer. In this case all
1248 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1249 * does not need to be set-up.
1250 */
0832c9bb
MD
1251 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0);
1252 buffer = NULL;
1253 ptr = hammer_bread(hmp, buf_offset, &error, &buffer);
1254 if (error == 0) {
0832c9bb 1255 bp = bio->bio_buf;
7bc5b8c2 1256 bp->b_flags |= B_AGE;
0832c9bb
MD
1257 hammer_io_modify(&buffer->io, 1);
1258 bcopy(bp->b_data, ptr, leaf->data_len);
1259 hammer_io_modify_done(&buffer->io);
7bc5b8c2 1260 hammer_rel_buffer(buffer, 0);
0832c9bb
MD
1261 bp->b_resid = 0;
1262 biodone(bio);
1263 }
47637bff 1264 }
e469566b
MD
1265 if (error == 0) {
1266 /*
1267 * The record is all setup now, add it. Potential conflics
1268 * have already been dealt with.
1269 */
1270 error = hammer_mem_add(record);
1271 KKASSERT(error == 0);
1272 } else {
1273 /*
3214ade6
MD
1274 * Major suckage occured. Also note: The record was never added
1275 * to the tree so we do not have to worry about the backend.
e469566b 1276 */
cebe9493 1277 kprintf("hammer_direct_write: failed @ %016llx\n",
973c11b9 1278 (long long)leaf->data_offset);
47637bff
MD
1279 bp = bio->bio_buf;
1280 bp->b_resid = 0;
1281 bp->b_error = EIO;
1282 bp->b_flags |= B_ERROR;
1283 biodone(bio);
e469566b
MD
1284 record->flags |= HAMMER_RECF_DELETED_FE;
1285 hammer_rel_mem_record(record);
47637bff
MD
1286 }
1287 return(error);
1288}
1289
43c665ae 1290/*
1b0ab2c3
MD
1291 * On completion of the BIO this callback must disconnect
1292 * it from the hammer_record and chain to the previous bio.
cdb6e4e6
MD
1293 *
1294 * An I/O error forces the mount to read-only. Data buffers
1295 * are not B_LOCKED like meta-data buffers are, so we have to
1296 * throw the buffer away to prevent the kernel from retrying.
1b0ab2c3
MD
1297 */
1298static
1299void
1300hammer_io_direct_write_complete(struct bio *nbio)
1301{
1302 struct bio *obio;
e469566b 1303 struct buf *bp;
1b0ab2c3
MD
1304 hammer_record_t record = nbio->bio_caller_info1.ptr;
1305
e469566b 1306 bp = nbio->bio_buf;
1b0ab2c3 1307 obio = pop_bio(nbio);
e469566b 1308 if (bp->b_flags & B_ERROR) {
cdb6e4e6 1309 hammer_critical_error(record->ip->hmp, record->ip,
e469566b 1310 bp->b_error,
cdb6e4e6 1311 "while writing bulk data");
e469566b 1312 bp->b_flags |= B_INVAL;
cdb6e4e6 1313 }
1b0ab2c3 1314 biodone(obio);
e469566b
MD
1315
1316 KKASSERT(record != NULL);
1317 KKASSERT(record->flags & HAMMER_RECF_DIRECT_IO);
1b0ab2c3 1318 if (record->flags & HAMMER_RECF_DIRECT_WAIT) {
de996e86
MD
1319 record->flags &= ~(HAMMER_RECF_DIRECT_IO |
1320 HAMMER_RECF_DIRECT_WAIT);
1321 /* record can disappear once DIRECT_IO flag is cleared */
1b0ab2c3 1322 wakeup(&record->flags);
de996e86
MD
1323 } else {
1324 record->flags &= ~HAMMER_RECF_DIRECT_IO;
1325 /* record can disappear once DIRECT_IO flag is cleared */
1b0ab2c3
MD
1326 }
1327}
1328
1329
1330/*
1331 * This is called before a record is either committed to the B-Tree
e469566b 1332 * or destroyed, to resolve any associated direct-IO.
1b0ab2c3 1333 *
e469566b
MD
1334 * (1) We must wait for any direct-IO related to the record to complete.
1335 *
1336 * (2) We must remove any buffer cache aliases for data accessed via
1337 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1338 * (the mirroring and reblocking code) do not see stale data.
1b0ab2c3
MD
1339 */
1340void
1341hammer_io_direct_wait(hammer_record_t record)
1342{
e469566b
MD
1343 /*
1344 * Wait for I/O to complete
1345 */
1346 if (record->flags & HAMMER_RECF_DIRECT_IO) {
1347 crit_enter();
1348 while (record->flags & HAMMER_RECF_DIRECT_IO) {
1349 record->flags |= HAMMER_RECF_DIRECT_WAIT;
1350 tsleep(&record->flags, 0, "hmdiow", 0);
1351 }
1352 crit_exit();
1353 }
1354
1355 /*
362ec2dc
MD
1356 * Invalidate any related buffer cache aliases associated with the
1357 * backing device. This is needed because the buffer cache buffer
1358 * for file data is associated with the file vnode, not the backing
1359 * device vnode.
1360 *
1361 * XXX I do not think this case can occur any more now that
1362 * reservations ensure that all such buffers are removed before
1363 * an area can be reused.
e469566b
MD
1364 */
1365 if (record->flags & HAMMER_RECF_DIRECT_INVAL) {
1366 KKASSERT(record->leaf.data_offset);
362ec2dc
MD
1367 hammer_del_buffers(record->ip->hmp, record->leaf.data_offset,
1368 record->zone2_offset, record->leaf.data_len,
1369 1);
e469566b 1370 record->flags &= ~HAMMER_RECF_DIRECT_INVAL;
1b0ab2c3 1371 }
1b0ab2c3
MD
1372}
1373
1374/*
43c665ae
MD
1375 * This is called to remove the second-level cached zone-2 offset from
1376 * frontend buffer cache buffers, now stale due to a data relocation.
1377 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1378 * by hammer_vop_strategy_read().
1379 *
1380 * This is rather nasty because here we have something like the reblocker
1381 * scanning the raw B-Tree with no held references on anything, really,
1382 * other then a shared lock on the B-Tree node, and we have to access the
1383 * frontend's buffer cache to check for and clean out the association.
1384 * Specifically, if the reblocker is moving data on the disk, these cached
1385 * offsets will become invalid.
1386 *
1387 * Only data record types associated with the large-data zone are subject
1388 * to direct-io and need to be checked.
1389 *
1390 */
1391void
1392hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf)
1393{
1394 struct hammer_inode_info iinfo;
1395 int zone;
1396
1397 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA)
1398 return;
1399 zone = HAMMER_ZONE_DECODE(leaf->data_offset);
1400 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX)
1401 return;
1402 iinfo.obj_id = leaf->base.obj_id;
1403 iinfo.obj_asof = 0; /* unused */
1404 iinfo.obj_localization = leaf->base.localization &
5a930e66 1405 HAMMER_LOCALIZE_PSEUDOFS_MASK;
43c665ae
MD
1406 iinfo.u.leaf = leaf;
1407 hammer_scan_inode_snapshots(hmp, &iinfo,
1408 hammer_io_direct_uncache_callback,
1409 leaf);
1410}
1411
1412static int
1413hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data)
1414{
1415 hammer_inode_info_t iinfo = data;
1416 hammer_off_t data_offset;
1417 hammer_off_t file_offset;
1418 struct vnode *vp;
1419 struct buf *bp;
1420 int blksize;
1421
1422 if (ip->vp == NULL)
1423 return(0);
1424 data_offset = iinfo->u.leaf->data_offset;
1425 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len;
1426 blksize = iinfo->u.leaf->data_len;
1427 KKASSERT((blksize & HAMMER_BUFMASK) == 0);
1428
1429 hammer_ref(&ip->lock);
1430 if (hammer_get_vnode(ip, &vp) == 0) {
b1c20cfa 1431 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL &&
43c665ae
MD
1432 bp->b_bio2.bio_offset != NOOFFSET) {
1433 bp = getblk(ip->vp, file_offset, blksize, 0, 0);
1434 bp->b_bio2.bio_offset = NOOFFSET;
1435 brelse(bp);
1436 }
1437 vput(vp);
1438 }
1439 hammer_rel_inode(ip, 0);
1440 return(0);
1441}
47637bff 1442
748efb59
MD
1443
1444/*
1445 * This function is called when writes may have occured on the volume,
1446 * indicating that the device may be holding cached writes.
1447 */
1448static void
1449hammer_io_flush_mark(hammer_volume_t volume)
1450{
1451 volume->vol_flags |= HAMMER_VOLF_NEEDFLUSH;
1452}
1453
1454/*
1455 * This function ensures that the device has flushed any cached writes out.
1456 */
1457void
1458hammer_io_flush_sync(hammer_mount_t hmp)
1459{
1460 hammer_volume_t volume;
1461 struct buf *bp_base = NULL;
1462 struct buf *bp;
1463
1464 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) {
1465 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) {
1466 volume->vol_flags &= ~HAMMER_VOLF_NEEDFLUSH;
1467 bp = getpbuf(NULL);
1468 bp->b_bio1.bio_offset = 0;
1469 bp->b_bufsize = 0;
1470 bp->b_bcount = 0;
1471 bp->b_cmd = BUF_CMD_FLUSH;
1472 bp->b_bio1.bio_caller_info1.cluster_head = bp_base;
ae8e83e6
MD
1473 bp->b_bio1.bio_done = biodone_sync;
1474 bp->b_bio1.bio_flags |= BIO_SYNC;
748efb59
MD
1475 bp_base = bp;
1476 vn_strategy(volume->devvp, &bp->b_bio1);
1477 }
1478 }
1479 while ((bp = bp_base) != NULL) {
1480 bp_base = bp->b_bio1.bio_caller_info1.cluster_head;
ae8e83e6 1481 biowait(&bp->b_bio1, "hmrFLS");
748efb59
MD
1482 relpbuf(bp, NULL);
1483 }
1484}