1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1993
5 * The Regents of the University of California. All rights reserved.
6 * Modifications/enhancements:
7 * Copyright (c) 1995 John S. Dyson. All rights reserved.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 */
33
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/proc.h>
38 #include <sys/bio.h>
39 #include <sys/buf.h>
40 #include <sys/vnode.h>
41 #include <sys/malloc.h>
42 #include <sys/mount.h>
43 #include <sys/racct.h>
44 #include <sys/resourcevar.h>
45 #include <sys/rwlock.h>
46 #include <sys/vmmeter.h>
47 #include <vm/vm.h>
48 #include <vm/vm_object.h>
49 #include <vm/vm_page.h>
50 #include <sys/sysctl.h>
51
52 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
53 static uma_zone_t cluster_pbuf_zone;
54
55 static void cluster_init(void *);
56 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
57 struct vn_clusterw *vnc, struct buf *last_bp, int gbflags);
58 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
59 daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
60 struct buf *fbp);
61 static void cluster_callback(struct buf *);
62
63 static int write_behind = 1;
64 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
65 "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
66
67 static int read_max = 64;
68 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
69 "Cluster read-ahead max block count");
70
71 static int read_min = 1;
72 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
73 "Cluster read min block count");
74
75 SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL);
76
77 static void
cluster_init(void * dummy)78 cluster_init(void *dummy)
79 {
80
81 cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2);
82 }
83
84 /*
85 * Read data to a buf, including read-ahead if we find this to be beneficial.
86 * cluster_read replaces bread.
87 */
88 int
cluster_read(struct vnode * vp,u_quad_t filesize,daddr_t lblkno,long size,struct ucred * cred,long totread,int seqcount,int gbflags,struct buf ** bpp)89 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
90 struct ucred *cred, long totread, int seqcount, int gbflags,
91 struct buf **bpp)
92 {
93 struct buf *bp, *rbp, *reqbp;
94 struct bufobj *bo;
95 struct thread *td;
96 daddr_t blkno, origblkno;
97 int maxra, racluster;
98 int error, ncontig;
99 int i;
100
101 error = 0;
102 td = curthread;
103 bo = &vp->v_bufobj;
104 if (!unmapped_buf_allowed)
105 gbflags &= ~GB_UNMAPPED;
106
107 /*
108 * Try to limit the amount of read-ahead by a few
109 * ad-hoc parameters. This needs work!!!
110 */
111 racluster = vp->v_mount->mnt_iosize_max / size;
112 maxra = seqcount;
113 maxra = min(read_max, maxra);
114 maxra = min(nbuf/8, maxra);
115 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
116 maxra = (filesize / size) - lblkno;
117
118 /*
119 * get the requested block
120 */
121 error = getblkx(vp, lblkno, lblkno, size, 0, 0, gbflags, &bp);
122 if (error != 0) {
123 *bpp = NULL;
124 return (error);
125 }
126 gbflags &= ~GB_NOSPARSE;
127 origblkno = lblkno;
128 *bpp = reqbp = bp;
129
130 /*
131 * if it is in the cache, then check to see if the reads have been
132 * sequential. If they have, then try some read-ahead, otherwise
133 * back-off on prospective read-aheads.
134 */
135 if (bp->b_flags & B_CACHE) {
136 if (!seqcount) {
137 return 0;
138 } else if ((bp->b_flags & B_RAM) == 0) {
139 return 0;
140 } else {
141 bp->b_flags &= ~B_RAM;
142 BO_RLOCK(bo);
143 for (i = 1; i < maxra; i++) {
144 /*
145 * Stop if the buffer does not exist or it
146 * is invalid (about to go away?)
147 */
148 rbp = gbincore(&vp->v_bufobj, lblkno+i);
149 if (rbp == NULL || (rbp->b_flags & B_INVAL))
150 break;
151
152 /*
153 * Set another read-ahead mark so we know
154 * to check again. (If we can lock the
155 * buffer without waiting)
156 */
157 if ((((i % racluster) == (racluster - 1)) ||
158 (i == (maxra - 1)))
159 && (0 == BUF_LOCK(rbp,
160 LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
161 rbp->b_flags |= B_RAM;
162 BUF_UNLOCK(rbp);
163 }
164 }
165 BO_RUNLOCK(bo);
166 if (i >= maxra) {
167 return 0;
168 }
169 lblkno += i;
170 }
171 reqbp = bp = NULL;
172 /*
173 * If it isn't in the cache, then get a chunk from
174 * disk if sequential, otherwise just get the block.
175 */
176 } else {
177 off_t firstread = bp->b_offset;
178 int nblks;
179 long minread;
180
181 KASSERT(bp->b_offset != NOOFFSET,
182 ("cluster_read: no buffer offset"));
183
184 ncontig = 0;
185
186 /*
187 * Adjust totread if needed
188 */
189 minread = read_min * size;
190 if (minread > totread)
191 totread = minread;
192
193 /*
194 * Compute the total number of blocks that we should read
195 * synchronously.
196 */
197 if (firstread + totread > filesize)
198 totread = filesize - firstread;
199 nblks = howmany(totread, size);
200 if (nblks > racluster)
201 nblks = racluster;
202
203 /*
204 * Now compute the number of contiguous blocks.
205 */
206 if (nblks > 1) {
207 error = VOP_BMAP(vp, lblkno, NULL,
208 &blkno, &ncontig, NULL);
209 /*
210 * If this failed to map just do the original block.
211 */
212 if (error || blkno == -1)
213 ncontig = 0;
214 }
215
216 /*
217 * If we have contiguous data available do a cluster
218 * otherwise just read the requested block.
219 */
220 if (ncontig) {
221 /* Account for our first block. */
222 ncontig = min(ncontig + 1, nblks);
223 if (ncontig < nblks)
224 nblks = ncontig;
225 bp = cluster_rbuild(vp, filesize, lblkno,
226 blkno, size, nblks, gbflags, bp);
227 lblkno += (bp->b_bufsize / size);
228 } else {
229 bp->b_flags |= B_RAM;
230 bp->b_iocmd = BIO_READ;
231 lblkno += 1;
232 }
233 }
234
235 /*
236 * handle the synchronous read so that it is available ASAP.
237 */
238 if (bp) {
239 if ((bp->b_flags & B_CLUSTER) == 0) {
240 vfs_busy_pages(bp, 0);
241 }
242 bp->b_flags &= ~B_INVAL;
243 bp->b_ioflags &= ~BIO_ERROR;
244 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
245 BUF_KERNPROC(bp);
246 bp->b_iooffset = dbtob(bp->b_blkno);
247 bstrategy(bp);
248 #ifdef RACCT
249 if (racct_enable) {
250 PROC_LOCK(td->td_proc);
251 racct_add_buf(td->td_proc, bp, 0);
252 PROC_UNLOCK(td->td_proc);
253 }
254 #endif /* RACCT */
255 td->td_ru.ru_inblock++;
256 }
257
258 /*
259 * If we have been doing sequential I/O, then do some read-ahead.
260 */
261 while (lblkno < (origblkno + maxra)) {
262 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
263 if (error) {
264 error = 0;
265 break;
266 }
267
268 if (blkno == -1)
269 break;
270
271 /*
272 * We could throttle ncontig here by maxra but we might as
273 * well read the data if it is contiguous. We're throttled
274 * by racluster anyway.
275 */
276 if (ncontig) {
277 ncontig = min(ncontig + 1, racluster);
278 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
279 size, ncontig, gbflags, NULL);
280 lblkno += (rbp->b_bufsize / size);
281 if (rbp->b_flags & B_DELWRI) {
282 bqrelse(rbp);
283 continue;
284 }
285 } else {
286 rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
287 lblkno += 1;
288 if (rbp->b_flags & B_DELWRI) {
289 bqrelse(rbp);
290 continue;
291 }
292 rbp->b_flags |= B_ASYNC | B_RAM;
293 rbp->b_iocmd = BIO_READ;
294 rbp->b_blkno = blkno;
295 }
296 if (rbp->b_flags & B_CACHE) {
297 rbp->b_flags &= ~B_ASYNC;
298 bqrelse(rbp);
299 continue;
300 }
301 if ((rbp->b_flags & B_CLUSTER) == 0) {
302 vfs_busy_pages(rbp, 0);
303 }
304 rbp->b_flags &= ~B_INVAL;
305 rbp->b_ioflags &= ~BIO_ERROR;
306 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
307 BUF_KERNPROC(rbp);
308 rbp->b_iooffset = dbtob(rbp->b_blkno);
309 bstrategy(rbp);
310 #ifdef RACCT
311 if (racct_enable) {
312 PROC_LOCK(td->td_proc);
313 racct_add_buf(td->td_proc, rbp, 0);
314 PROC_UNLOCK(td->td_proc);
315 }
316 #endif /* RACCT */
317 td->td_ru.ru_inblock++;
318 }
319
320 if (reqbp) {
321 /*
322 * Like bread, always brelse() the buffer when
323 * returning an error.
324 */
325 error = bufwait(reqbp);
326 if (error != 0) {
327 brelse(reqbp);
328 *bpp = NULL;
329 }
330 }
331 return (error);
332 }
333
334 /*
335 * If blocks are contiguous on disk, use this to provide clustered
336 * read ahead. We will read as many blocks as possible sequentially
337 * and then parcel them up into logical blocks in the buffer hash table.
338 */
339 static struct buf *
cluster_rbuild(struct vnode * vp,u_quad_t filesize,daddr_t lbn,daddr_t blkno,long size,int run,int gbflags,struct buf * fbp)340 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
341 daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
342 {
343 struct buf *bp, *tbp;
344 daddr_t bn;
345 off_t off;
346 long tinc, tsize;
347 int i, inc, j, k, toff;
348
349 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
350 ("cluster_rbuild: size %ld != f_iosize %jd\n",
351 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
352
353 /*
354 * avoid a division
355 */
356 while ((u_quad_t) size * (lbn + run) > filesize) {
357 --run;
358 }
359
360 if (fbp) {
361 tbp = fbp;
362 tbp->b_iocmd = BIO_READ;
363 } else {
364 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
365 if (tbp->b_flags & B_CACHE)
366 return tbp;
367 tbp->b_flags |= B_ASYNC | B_RAM;
368 tbp->b_iocmd = BIO_READ;
369 }
370 tbp->b_blkno = blkno;
371 if ( (tbp->b_flags & B_MALLOC) ||
372 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
373 return tbp;
374
375 bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT);
376 if (bp == NULL)
377 return tbp;
378 MPASS((bp->b_flags & B_MAXPHYS) != 0);
379
380 /*
381 * We are synthesizing a buffer out of vm_page_t's, but
382 * if the block size is not page aligned then the starting
383 * address may not be either. Inherit the b_data offset
384 * from the original buffer.
385 */
386 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
387 if ((gbflags & GB_UNMAPPED) != 0) {
388 bp->b_data = unmapped_buf;
389 } else {
390 bp->b_data = (char *)((vm_offset_t)bp->b_data |
391 ((vm_offset_t)tbp->b_data & PAGE_MASK));
392 }
393 bp->b_iocmd = BIO_READ;
394 bp->b_iodone = cluster_callback;
395 bp->b_blkno = blkno;
396 bp->b_lblkno = lbn;
397 bp->b_offset = tbp->b_offset;
398 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
399 pbgetvp(vp, bp);
400
401 TAILQ_INIT(&bp->b_cluster.cluster_head);
402
403 bp->b_bcount = 0;
404 bp->b_bufsize = 0;
405 bp->b_npages = 0;
406
407 inc = btodb(size);
408 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
409 if (i == 0) {
410 vm_object_pip_add(tbp->b_bufobj->bo_object,
411 tbp->b_npages);
412 vfs_busy_pages_acquire(tbp);
413 } else {
414 if ((bp->b_npages * PAGE_SIZE) +
415 round_page(size) > vp->v_mount->mnt_iosize_max) {
416 break;
417 }
418
419 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
420 (gbflags & GB_UNMAPPED));
421
422 /* Don't wait around for locked bufs. */
423 if (tbp == NULL)
424 break;
425
426 /*
427 * Stop scanning if the buffer is fully valid
428 * (marked B_CACHE), or locked (may be doing a
429 * background write), or if the buffer is not
430 * VMIO backed. The clustering code can only deal
431 * with VMIO-backed buffers. The bo lock is not
432 * required for the BKGRDINPROG check since it
433 * can not be set without the buf lock.
434 */
435 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
436 (tbp->b_flags & B_CACHE) ||
437 (tbp->b_flags & B_VMIO) == 0) {
438 bqrelse(tbp);
439 break;
440 }
441
442 /*
443 * The buffer must be completely invalid in order to
444 * take part in the cluster. If it is partially valid
445 * then we stop.
446 */
447 off = tbp->b_offset;
448 tsize = size;
449 for (j = 0; tsize > 0; j++) {
450 toff = off & PAGE_MASK;
451 tinc = tsize;
452 if (toff + tinc > PAGE_SIZE)
453 tinc = PAGE_SIZE - toff;
454 if (vm_page_trysbusy(tbp->b_pages[j]) == 0)
455 break;
456 if ((tbp->b_pages[j]->valid &
457 vm_page_bits(toff, tinc)) != 0) {
458 vm_page_sunbusy(tbp->b_pages[j]);
459 break;
460 }
461 vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
462 off += tinc;
463 tsize -= tinc;
464 }
465 if (tsize > 0) {
466 clean_sbusy:
467 vm_object_pip_wakeupn(tbp->b_bufobj->bo_object,
468 j);
469 for (k = 0; k < j; k++)
470 vm_page_sunbusy(tbp->b_pages[k]);
471 bqrelse(tbp);
472 break;
473 }
474
475 /*
476 * Set a read-ahead mark as appropriate
477 */
478 if ((fbp && (i == 1)) || (i == (run - 1)))
479 tbp->b_flags |= B_RAM;
480
481 /*
482 * Set the buffer up for an async read (XXX should
483 * we do this only if we do not wind up brelse()ing?).
484 * Set the block number if it isn't set, otherwise
485 * if it is make sure it matches the block number we
486 * expect.
487 */
488 tbp->b_flags |= B_ASYNC;
489 tbp->b_iocmd = BIO_READ;
490 if (tbp->b_blkno == tbp->b_lblkno) {
491 tbp->b_blkno = bn;
492 } else if (tbp->b_blkno != bn) {
493 goto clean_sbusy;
494 }
495 }
496 /*
497 * XXX fbp from caller may not be B_ASYNC, but we are going
498 * to biodone() it in cluster_callback() anyway
499 */
500 BUF_KERNPROC(tbp);
501 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
502 tbp, b_cluster.cluster_entry);
503 for (j = 0; j < tbp->b_npages; j += 1) {
504 vm_page_t m;
505
506 m = tbp->b_pages[j];
507 if ((bp->b_npages == 0) ||
508 (bp->b_pages[bp->b_npages-1] != m)) {
509 bp->b_pages[bp->b_npages] = m;
510 bp->b_npages++;
511 }
512 if (vm_page_all_valid(m))
513 tbp->b_pages[j] = bogus_page;
514 }
515
516 /*
517 * Don't inherit tbp->b_bufsize as it may be larger due to
518 * a non-page-aligned size. Instead just aggregate using
519 * 'size'.
520 */
521 if (tbp->b_bcount != size)
522 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
523 if (tbp->b_bufsize != size)
524 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
525 bp->b_bcount += size;
526 bp->b_bufsize += size;
527 }
528
529 /*
530 * Fully valid pages in the cluster are already good and do not need
531 * to be re-read from disk. Replace the page with bogus_page
532 */
533 for (j = 0; j < bp->b_npages; j++) {
534 if (vm_page_all_valid(bp->b_pages[j]))
535 bp->b_pages[j] = bogus_page;
536 }
537 if (bp->b_bufsize > bp->b_kvasize)
538 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
539 bp->b_bufsize, bp->b_kvasize);
540
541 if (buf_mapped(bp)) {
542 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
543 (vm_page_t *)bp->b_pages, bp->b_npages);
544 }
545 return (bp);
546 }
547
548 /*
549 * Cleanup after a clustered read or write.
550 * This is complicated by the fact that any of the buffers might have
551 * extra memory (if there were no empty buffer headers at allocbuf time)
552 * that we will need to shift around.
553 */
554 static void
cluster_callback(struct buf * bp)555 cluster_callback(struct buf *bp)
556 {
557 struct buf *nbp, *tbp;
558 int error = 0;
559
560 /*
561 * Must propagate errors to all the components.
562 */
563 if (bp->b_ioflags & BIO_ERROR)
564 error = bp->b_error;
565
566 if (buf_mapped(bp)) {
567 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
568 bp->b_npages);
569 }
570 /*
571 * Move memory from the large cluster buffer into the component
572 * buffers and mark IO as done on these.
573 */
574 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
575 tbp; tbp = nbp) {
576 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
577 if (error) {
578 tbp->b_ioflags |= BIO_ERROR;
579 tbp->b_error = error;
580 } else {
581 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
582 tbp->b_flags &= ~B_INVAL;
583 tbp->b_ioflags &= ~BIO_ERROR;
584 /*
585 * XXX the bdwrite()/bqrelse() issued during
586 * cluster building clears B_RELBUF (see bqrelse()
587 * comment). If direct I/O was specified, we have
588 * to restore it here to allow the buffer and VM
589 * to be freed.
590 */
591 if (tbp->b_flags & B_DIRECT)
592 tbp->b_flags |= B_RELBUF;
593 }
594 bufdone(tbp);
595 }
596 pbrelvp(bp);
597 uma_zfree(cluster_pbuf_zone, bp);
598 }
599
600 /*
601 * cluster_wbuild_wb:
602 *
603 * Implement modified write build for cluster.
604 *
605 * write_behind = 0 write behind disabled
606 * write_behind = 1 write behind normal (default)
607 * write_behind = 2 write behind backed-off
608 */
609
610 static __inline int
cluster_wbuild_wb(struct vnode * vp,long size,daddr_t start_lbn,int len,int gbflags)611 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
612 int gbflags)
613 {
614 int r = 0;
615
616 switch (write_behind) {
617 case 2:
618 if (start_lbn < len)
619 break;
620 start_lbn -= len;
621 /* FALLTHROUGH */
622 case 1:
623 r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
624 /* FALLTHROUGH */
625 default:
626 /* FALLTHROUGH */
627 break;
628 }
629 return(r);
630 }
631
632 /*
633 * Do clustered write for FFS.
634 *
635 * Three cases:
636 * 1. Write is not sequential (write asynchronously)
637 * Write is sequential:
638 * 2. beginning of cluster - begin cluster
639 * 3. middle of a cluster - add to cluster
640 * 4. end of a cluster - asynchronously write cluster
641 */
642 void
cluster_write(struct vnode * vp,struct vn_clusterw * vnc,struct buf * bp,u_quad_t filesize,int seqcount,int gbflags)643 cluster_write(struct vnode *vp, struct vn_clusterw *vnc, struct buf *bp,
644 u_quad_t filesize, int seqcount, int gbflags)
645 {
646 daddr_t lbn, pbn;
647 int maxclen, cursize;
648 int lblocksize;
649 int async;
650
651 if (!unmapped_buf_allowed)
652 gbflags &= ~GB_UNMAPPED;
653
654 if (vp->v_type == VREG) {
655 async = DOINGASYNC(vp);
656 lblocksize = vp->v_mount->mnt_stat.f_iosize;
657 } else {
658 async = 0;
659 lblocksize = bp->b_bufsize;
660 }
661 lbn = bp->b_lblkno;
662 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
663
664 /* Initialize vnode to beginning of file. */
665 if (lbn == 0)
666 vnc->v_lasta = vnc->v_clen = vnc->v_cstart = vnc->v_lastw = 0;
667
668 if (vnc->v_clen == 0 || lbn != vnc->v_lastw + 1 ||
669 (bp->b_blkno != vnc->v_lasta + btodb(lblocksize))) {
670 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
671 if (vnc->v_clen != 0) {
672 /*
673 * Next block is not sequential.
674 *
675 * If we are not writing at end of file, the process
676 * seeked to another point in the file since its last
677 * write, or we have reached our maximum cluster size,
678 * then push the previous cluster. Otherwise try
679 * reallocating to make it sequential.
680 *
681 * Change to algorithm: only push previous cluster if
682 * it was sequential from the point of view of the
683 * seqcount heuristic, otherwise leave the buffer
684 * intact so we can potentially optimize the I/O
685 * later on in the buf_daemon or update daemon
686 * flush.
687 */
688 cursize = vnc->v_lastw - vnc->v_cstart + 1;
689 if ((u_quad_t)bp->b_offset + lblocksize != filesize ||
690 lbn != vnc->v_lastw + 1 || vnc->v_clen <= cursize) {
691 if (!async && seqcount > 0) {
692 cluster_wbuild_wb(vp, lblocksize,
693 vnc->v_cstart, cursize, gbflags);
694 }
695 } else {
696 struct buf **bpp, **endbp;
697 struct cluster_save *buflist;
698
699 buflist = cluster_collectbufs(vp, vnc, bp,
700 gbflags);
701 if (buflist == NULL) {
702 /*
703 * Cluster build failed so just write
704 * it now.
705 */
706 bawrite(bp);
707 return;
708 }
709 endbp = &buflist->bs_children
710 [buflist->bs_nchildren - 1];
711 if (VOP_REALLOCBLKS(vp, buflist)) {
712 /*
713 * Failed, push the previous cluster
714 * if *really* writing sequentially
715 * in the logical file (seqcount > 1),
716 * otherwise delay it in the hopes that
717 * the low level disk driver can
718 * optimize the write ordering.
719 */
720 for (bpp = buflist->bs_children;
721 bpp < endbp; bpp++)
722 brelse(*bpp);
723 free(buflist, M_SEGMENT);
724 if (seqcount > 1) {
725 cluster_wbuild_wb(vp,
726 lblocksize, vnc->v_cstart,
727 cursize, gbflags);
728 }
729 } else {
730 /*
731 * Succeeded, keep building cluster.
732 */
733 for (bpp = buflist->bs_children;
734 bpp <= endbp; bpp++)
735 bdwrite(*bpp);
736 free(buflist, M_SEGMENT);
737 vnc->v_lastw = lbn;
738 vnc->v_lasta = bp->b_blkno;
739 return;
740 }
741 }
742 }
743 /*
744 * Consider beginning a cluster. If at end of file, make
745 * cluster as large as possible, otherwise find size of
746 * existing cluster.
747 */
748 if (vp->v_type == VREG &&
749 (u_quad_t) bp->b_offset + lblocksize != filesize &&
750 bp->b_blkno == bp->b_lblkno &&
751 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen,
752 NULL) != 0 || bp->b_blkno == -1)) {
753 pbn = bp->b_blkno;
754 bawrite(bp);
755 vnc->v_clen = 0;
756 vnc->v_lasta = pbn;
757 vnc->v_cstart = lbn + 1;
758 vnc->v_lastw = lbn;
759 return;
760 }
761 vnc->v_clen = maxclen;
762 pbn = bp->b_blkno;
763 if (!async && maxclen == 0) { /* I/O not contiguous */
764 vnc->v_cstart = lbn + 1;
765 bawrite(bp);
766 } else { /* Wait for rest of cluster */
767 vnc->v_cstart = lbn;
768 bdwrite(bp);
769 }
770 } else if (lbn == vnc->v_cstart + vnc->v_clen) {
771 /*
772 * At end of cluster, write it out if seqcount tells us we
773 * are operating sequentially, otherwise let the buf or
774 * update daemon handle it.
775 */
776 pbn = bp->b_blkno;
777 bdwrite(bp);
778 if (seqcount > 1) {
779 cluster_wbuild_wb(vp, lblocksize, vnc->v_cstart,
780 vnc->v_clen + 1, gbflags);
781 }
782 vnc->v_clen = 0;
783 vnc->v_cstart = lbn + 1;
784 } else if (vm_page_count_severe()) {
785 /*
786 * We are low on memory, get it going NOW
787 */
788 pbn = bp->b_blkno;
789 bawrite(bp);
790 } else {
791 /*
792 * In the middle of a cluster, so just delay the I/O for now.
793 */
794 pbn = bp->b_blkno;
795 bdwrite(bp);
796 }
797 vnc->v_lastw = lbn;
798 vnc->v_lasta = pbn;
799 }
800
801 /*
802 * This is an awful lot like cluster_rbuild...wish they could be combined.
803 * The last lbn argument is the current block on which I/O is being
804 * performed. Check to see that it doesn't fall in the middle of
805 * the current block (if last_bp == NULL).
806 */
807 int
cluster_wbuild(struct vnode * vp,long size,daddr_t start_lbn,int len,int gbflags)808 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
809 int gbflags)
810 {
811 struct buf *bp, *tbp;
812 struct bufobj *bo;
813 int i, j;
814 int totalwritten = 0;
815 int dbsize = btodb(size);
816
817 if (!unmapped_buf_allowed)
818 gbflags &= ~GB_UNMAPPED;
819
820 bo = &vp->v_bufobj;
821 while (len > 0) {
822 /*
823 * If the buffer is not delayed-write (i.e. dirty), or it
824 * is delayed-write but either locked or inval, it cannot
825 * partake in the clustered write.
826 */
827 BO_LOCK(bo);
828 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
829 (tbp->b_vflags & BV_BKGRDINPROG)) {
830 BO_UNLOCK(bo);
831 ++start_lbn;
832 --len;
833 continue;
834 }
835 if (BUF_LOCK(tbp,
836 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
837 ++start_lbn;
838 --len;
839 continue;
840 }
841 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
842 BUF_UNLOCK(tbp);
843 ++start_lbn;
844 --len;
845 continue;
846 }
847 bremfree(tbp);
848 tbp->b_flags &= ~B_DONE;
849
850 /*
851 * Extra memory in the buffer, punt on this buffer.
852 * XXX we could handle this in most cases, but we would
853 * have to push the extra memory down to after our max
854 * possible cluster size and then potentially pull it back
855 * up if the cluster was terminated prematurely--too much
856 * hassle.
857 */
858 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
859 (B_CLUSTEROK | B_VMIO)) ||
860 (tbp->b_bcount != tbp->b_bufsize) ||
861 (tbp->b_bcount != size) ||
862 (len == 1) ||
863 ((bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT)) == NULL)) {
864 totalwritten += tbp->b_bufsize;
865 bawrite(tbp);
866 ++start_lbn;
867 --len;
868 continue;
869 }
870 MPASS((bp->b_flags & B_MAXPHYS) != 0);
871
872 /*
873 * We got a pbuf to make the cluster in.
874 * so initialise it.
875 */
876 TAILQ_INIT(&bp->b_cluster.cluster_head);
877 bp->b_bcount = 0;
878 bp->b_bufsize = 0;
879 bp->b_npages = 0;
880 if (tbp->b_wcred != NOCRED)
881 bp->b_wcred = crhold(tbp->b_wcred);
882
883 bp->b_blkno = tbp->b_blkno;
884 bp->b_lblkno = tbp->b_lblkno;
885 bp->b_offset = tbp->b_offset;
886
887 /*
888 * We are synthesizing a buffer out of vm_page_t's, but
889 * if the block size is not page aligned then the starting
890 * address may not be either. Inherit the b_data offset
891 * from the original buffer.
892 */
893 if ((gbflags & GB_UNMAPPED) == 0 ||
894 (tbp->b_flags & B_VMIO) == 0) {
895 bp->b_data = (char *)((vm_offset_t)bp->b_data |
896 ((vm_offset_t)tbp->b_data & PAGE_MASK));
897 } else {
898 bp->b_data = unmapped_buf;
899 }
900 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
901 B_NEEDCOMMIT));
902 bp->b_iodone = cluster_callback;
903 pbgetvp(vp, bp);
904 /*
905 * From this location in the file, scan forward to see
906 * if there are buffers with adjacent data that need to
907 * be written as well.
908 */
909 for (i = 0; i < len; ++i, ++start_lbn) {
910 if (i != 0) { /* If not the first buffer */
911 /*
912 * If the adjacent data is not even in core it
913 * can't need to be written.
914 */
915 BO_LOCK(bo);
916 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
917 (tbp->b_vflags & BV_BKGRDINPROG)) {
918 BO_UNLOCK(bo);
919 break;
920 }
921
922 /*
923 * If it IS in core, but has different
924 * characteristics, or is locked (which
925 * means it could be undergoing a background
926 * I/O or be in a weird state), then don't
927 * cluster with it.
928 */
929 if (BUF_LOCK(tbp,
930 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
931 BO_LOCKPTR(bo)))
932 break;
933
934 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
935 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
936 != (B_DELWRI | B_CLUSTEROK |
937 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
938 tbp->b_wcred != bp->b_wcred) {
939 BUF_UNLOCK(tbp);
940 break;
941 }
942
943 /*
944 * Check that the combined cluster
945 * would make sense with regard to pages
946 * and would not be too large
947 */
948 if ((tbp->b_bcount != size) ||
949 ((bp->b_blkno + (dbsize * i)) !=
950 tbp->b_blkno) ||
951 ((tbp->b_npages + bp->b_npages) >
952 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
953 BUF_UNLOCK(tbp);
954 break;
955 }
956
957 /*
958 * Ok, it's passed all the tests,
959 * so remove it from the free list
960 * and mark it busy. We will use it.
961 */
962 bremfree(tbp);
963 tbp->b_flags &= ~B_DONE;
964 } /* end of code for non-first buffers only */
965 /*
966 * If the IO is via the VM then we do some
967 * special VM hackery (yuck). Since the buffer's
968 * block size may not be page-aligned it is possible
969 * for a page to be shared between two buffers. We
970 * have to get rid of the duplication when building
971 * the cluster.
972 */
973 if (tbp->b_flags & B_VMIO) {
974 vm_page_t m;
975
976 if (i == 0) {
977 vfs_busy_pages_acquire(tbp);
978 } else { /* if not first buffer */
979 for (j = 0; j < tbp->b_npages; j += 1) {
980 m = tbp->b_pages[j];
981 if (vm_page_trysbusy(m) == 0) {
982 for (j--; j >= 0; j--)
983 vm_page_sunbusy(
984 tbp->b_pages[j]);
985 bqrelse(tbp);
986 goto finishcluster;
987 }
988 }
989 }
990 vm_object_pip_add(tbp->b_bufobj->bo_object,
991 tbp->b_npages);
992 for (j = 0; j < tbp->b_npages; j += 1) {
993 m = tbp->b_pages[j];
994 if ((bp->b_npages == 0) ||
995 (bp->b_pages[bp->b_npages - 1] != m)) {
996 bp->b_pages[bp->b_npages] = m;
997 bp->b_npages++;
998 }
999 }
1000 }
1001 bp->b_bcount += size;
1002 bp->b_bufsize += size;
1003 /*
1004 * If any of the clustered buffers have their
1005 * B_BARRIER flag set, transfer that request to
1006 * the cluster.
1007 */
1008 bp->b_flags |= (tbp->b_flags & B_BARRIER);
1009 tbp->b_flags &= ~(B_DONE | B_BARRIER);
1010 tbp->b_flags |= B_ASYNC;
1011 tbp->b_ioflags &= ~BIO_ERROR;
1012 tbp->b_iocmd = BIO_WRITE;
1013 bundirty(tbp);
1014 reassignbuf(tbp); /* put on clean list */
1015 bufobj_wref(tbp->b_bufobj);
1016 BUF_KERNPROC(tbp);
1017 buf_track(tbp, __func__);
1018 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1019 tbp, b_cluster.cluster_entry);
1020 }
1021 finishcluster:
1022 if (buf_mapped(bp)) {
1023 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1024 (vm_page_t *)bp->b_pages, bp->b_npages);
1025 }
1026 if (bp->b_bufsize > bp->b_kvasize)
1027 panic(
1028 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1029 bp->b_bufsize, bp->b_kvasize);
1030 totalwritten += bp->b_bufsize;
1031 bp->b_dirtyoff = 0;
1032 bp->b_dirtyend = bp->b_bufsize;
1033 bawrite(bp);
1034
1035 len -= i;
1036 }
1037 return totalwritten;
1038 }
1039
1040 /*
1041 * Collect together all the buffers in a cluster.
1042 * Plus add one additional buffer.
1043 */
1044 static struct cluster_save *
cluster_collectbufs(struct vnode * vp,struct vn_clusterw * vnc,struct buf * last_bp,int gbflags)1045 cluster_collectbufs(struct vnode *vp, struct vn_clusterw *vnc,
1046 struct buf *last_bp, int gbflags)
1047 {
1048 struct cluster_save *buflist;
1049 struct buf *bp;
1050 daddr_t lbn;
1051 int i, j, len, error;
1052
1053 len = vnc->v_lastw - vnc->v_cstart + 1;
1054 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1055 M_SEGMENT, M_WAITOK);
1056 buflist->bs_nchildren = 0;
1057 buflist->bs_children = (struct buf **) (buflist + 1);
1058 for (lbn = vnc->v_cstart, i = 0; i < len; lbn++, i++) {
1059 error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1060 gbflags, &bp);
1061 if (error != 0) {
1062 /*
1063 * If read fails, release collected buffers
1064 * and return failure.
1065 */
1066 for (j = 0; j < i; j++)
1067 brelse(buflist->bs_children[j]);
1068 free(buflist, M_SEGMENT);
1069 return (NULL);
1070 }
1071 buflist->bs_children[i] = bp;
1072 if (bp->b_blkno == bp->b_lblkno)
1073 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1074 NULL, NULL);
1075 }
1076 buflist->bs_children[i] = bp = last_bp;
1077 if (bp->b_blkno == bp->b_lblkno)
1078 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1079 buflist->bs_nchildren = i + 1;
1080 return (buflist);
1081 }
1082
1083 void
cluster_init_vn(struct vn_clusterw * vnc)1084 cluster_init_vn(struct vn_clusterw *vnc)
1085 {
1086 vnc->v_lasta = 0;
1087 vnc->v_clen = 0;
1088 vnc->v_cstart = 0;
1089 vnc->v_lastw = 0;
1090 }
1091