1 /*-
2 * Copyright (c) 1993
3 * The Regents of the University of California. All rights reserved.
4 * Modifications/enhancements:
5 * Copyright (c) 1995 John S. Dyson. All rights reserved.
6 * Copyright (c) 2012-2013 Matthew Dillon. All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
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 the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 */
32
33 #include "opt_debug_cluster.h"
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/proc.h>
39 #include <sys/buf.h>
40 #include <sys/vnode.h>
41 #include <sys/malloc.h>
42 #include <sys/mount.h>
43 #include <sys/resourcevar.h>
44 #include <sys/vmmeter.h>
45 #include <vm/vm.h>
46 #include <vm/vm_object.h>
47 #include <vm/vm_page.h>
48 #include <sys/sysctl.h>
49
50 #include <sys/buf2.h>
51 #include <vm/vm_page2.h>
52
53 #include <machine/limits.h>
54
55 /*
56 * Cluster tracking cache - replaces the original vnode v_* fields which had
57 * limited utility and were not MP safe.
58 *
59 * The cluster tracking cache is a simple 4-way set-associative non-chained
60 * cache. It is capable of tracking up to four zones separated by 1MB or
61 * more per vnode.
62 *
63 * NOTE: We want this structure to be cache-line friendly so the iterator
64 * is embedded rather than in a separate array.
65 *
66 * NOTE: A cluster cache entry can become stale when a vnode is recycled.
67 * For now we treat the values as heuristical but also self-consistent.
68 * i.e. the values cannot be completely random and cannot be SMP unsafe
69 * or the cluster code might end-up clustering non-contiguous buffers
70 * at the wrong offsets.
71 */
72 struct cluster_cache {
73 struct vnode *vp;
74 u_int locked;
75 off_t v_lastw; /* last write (end) (write cluster) */
76 off_t v_cstart; /* start block (beg) of cluster */
77 off_t v_lasta; /* last allocation (end) */
78 u_int v_clen; /* length of current cluster */
79 u_int iterator;
80 } __cachealign;
81
82 typedef struct cluster_cache cluster_cache_t;
83
84 #define CLUSTER_CACHE_SIZE 512
85 #define CLUSTER_CACHE_MASK (CLUSTER_CACHE_SIZE - 1)
86
87 #define CLUSTER_ZONE ((off_t)(1024 * 1024))
88
89 cluster_cache_t cluster_array[CLUSTER_CACHE_SIZE];
90
91 #if defined(CLUSTERDEBUG)
92 static int rcluster= 0;
93 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
94 #endif
95
96 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
97
98 static struct cluster_save *
99 cluster_collectbufs (cluster_cache_t *cc, struct vnode *vp,
100 struct buf *last_bp, int blksize);
101 static struct buf *
102 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
103 off_t doffset, int blksize, int run,
104 struct buf *fbp, int *srp);
105 static void cluster_callback (struct bio *);
106 static void cluster_setram (struct buf *);
107 static void cluster_clrram (struct buf *);
108 static int cluster_wbuild(struct vnode *vp, struct buf **bpp, int blksize,
109 off_t start_loffset, int bytes);
110
111 static int write_behind = 1;
112 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
113 "Cluster write-behind setting");
114 static quad_t write_behind_minfilesize = 10 * 1024 * 1024;
115 SYSCTL_QUAD(_vfs, OID_AUTO, write_behind_minfilesize, CTLFLAG_RW,
116 &write_behind_minfilesize, 0, "Cluster write-behind setting");
117 static int max_readahead = 2 * 1024 * 1024;
118 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
119 "Limit in bytes for desired cluster read-ahead");
120
121 extern vm_page_t bogus_page;
122
123 /*
124 * nblks is our cluster_rbuild request size. The approximate number of
125 * physical read-ahead requests is maxra / nblks. The physical request
126 * size is limited by the device (maxrbuild). We also do not want to make
127 * the request size too big or it will mess up the B_RAM streaming.
128 */
129 static __inline
130 int
calc_rbuild_reqsize(int maxra,int maxrbuild)131 calc_rbuild_reqsize(int maxra, int maxrbuild)
132 {
133 int nblks;
134
135 if ((nblks = maxra / 4) > maxrbuild)
136 nblks = maxrbuild;
137 if (nblks < 1)
138 nblks = maxra;
139 return nblks;
140 }
141
142 /*
143 * Acquire/release cluster cache (can return dummy entry)
144 */
145 static
146 cluster_cache_t *
cluster_getcache(cluster_cache_t * dummy,struct vnode * vp,off_t loffset)147 cluster_getcache(cluster_cache_t *dummy, struct vnode *vp, off_t loffset)
148 {
149 cluster_cache_t *cc;
150 size_t hv;
151 int i;
152 int xact;
153
154 hv = (size_t)(intptr_t)vp ^ (size_t)(intptr_t)vp / sizeof(*vp);
155 hv &= CLUSTER_CACHE_MASK & ~3;
156 cc = &cluster_array[hv];
157
158 xact = -1;
159 for (i = 0; i < 4; ++i) {
160 if (cc[i].vp != vp)
161 continue;
162 if (rounddown2(cc[i].v_cstart ^ loffset, CLUSTER_ZONE) == 0) {
163 xact = i;
164 break;
165 }
166 }
167 if (xact >= 0 && atomic_swap_int(&cc[xact].locked, 1) == 0) {
168 if (cc[xact].vp == vp &&
169 rounddown2(cc[i].v_cstart ^ loffset, CLUSTER_ZONE) == 0) {
170 return(&cc[xact]);
171 }
172 atomic_swap_int(&cc[xact].locked, 0);
173 }
174
175 /*
176 * New entry. If we can't acquire the cache line then use the
177 * passed-in dummy element and reset all fields.
178 *
179 * When we are able to acquire the cache line we only clear the
180 * fields if the vp does not match. This allows us to multi-zone
181 * a vp and for excessive zones / partial clusters to be retired.
182 */
183 i = cc->iterator++ & 3;
184 cc += i;
185 if (atomic_swap_int(&cc->locked, 1) != 0) {
186 cc = dummy;
187 cc->locked = 1;
188 cc->vp = NULL;
189 }
190 if (cc->vp != vp) {
191 cc->vp = vp;
192 cc->v_lasta = 0;
193 cc->v_clen = 0;
194 cc->v_cstart = 0;
195 cc->v_lastw = 0;
196 }
197 return(cc);
198 }
199
200 static
201 void
cluster_putcache(cluster_cache_t * cc)202 cluster_putcache(cluster_cache_t *cc)
203 {
204 atomic_swap_int(&cc->locked, 0);
205 }
206
207 /*
208 * This replaces bread(), providing a synchronous read of the requested
209 * buffer plus asynchronous read-ahead within the specified bounds.
210 *
211 * The caller may pre-populate *bpp if it already has the requested buffer
212 * in-hand, else must set *bpp to NULL. Note that the cluster_read() inline
213 * sets *bpp to NULL and then calls cluster_readx() for compatibility.
214 *
215 * filesize - read-ahead @ blksize will not cross this boundary
216 * loffset - loffset for returned *bpp
217 * blksize - blocksize for returned *bpp and read-ahead bps
218 * minreq - minimum (not a hard minimum) in bytes, typically reflects
219 * a higher level uio resid.
220 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB)
221 * bpp - return buffer (*bpp) for (loffset,blksize)
222 */
223 int
cluster_readx(struct vnode * vp,off_t filesize,off_t loffset,int blksize,int bflags,size_t minreq,size_t maxreq,struct buf ** bpp)224 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset, int blksize,
225 int bflags, size_t minreq, size_t maxreq,
226 struct buf **bpp)
227 {
228 struct buf *bp, *rbp, *reqbp;
229 off_t origoffset;
230 off_t doffset;
231 int error;
232 int i;
233 int maxra;
234 int maxrbuild;
235 int sr;
236 int blkflags = (bflags & B_KVABIO) ? GETBLK_KVABIO : 0;
237
238 sr = 0;
239
240 /*
241 * Calculate the desired read-ahead in blksize'd blocks (maxra).
242 * To do this we calculate maxreq.
243 *
244 * maxreq typically starts out as a sequential heuristic. If the
245 * high level uio/resid is bigger (minreq), we pop maxreq up to
246 * minreq. This represents the case where random I/O is being
247 * performed by the userland is issuing big read()'s.
248 *
249 * Then we limit maxreq to max_readahead to ensure it is a reasonable
250 * value.
251 *
252 * Finally we must ensure that (loffset + maxreq) does not cross the
253 * boundary (filesize) for the current blocksize. If we allowed it
254 * to cross we could end up with buffers past the boundary with the
255 * wrong block size (HAMMER large-data areas use mixed block sizes).
256 * minreq is also absolutely limited to filesize.
257 */
258 if (maxreq < minreq)
259 maxreq = minreq;
260 /* minreq not used beyond this point */
261
262 if (maxreq > max_readahead) {
263 maxreq = max_readahead;
264 if (maxreq > 16 * 1024 * 1024)
265 maxreq = 16 * 1024 * 1024;
266 }
267 if (maxreq < blksize)
268 maxreq = blksize;
269 if (loffset + maxreq > filesize) {
270 if (loffset > filesize)
271 maxreq = 0;
272 else
273 maxreq = filesize - loffset;
274 }
275
276 maxra = (int)(maxreq / blksize);
277
278 /*
279 * Get the requested block.
280 */
281 if (*bpp)
282 reqbp = bp = *bpp;
283 else
284 *bpp = reqbp = bp = getblk(vp, loffset, blksize, blkflags, 0);
285 origoffset = loffset;
286
287 /*
288 * Calculate the maximum cluster size for a single I/O, used
289 * by cluster_rbuild().
290 */
291 maxrbuild = vmaxiosize(vp) / blksize;
292
293 /*
294 * If it is in the cache, then check to see if the reads have been
295 * sequential. If they have, then try some read-ahead, otherwise
296 * back-off on prospective read-aheads.
297 */
298 if (bp->b_flags & B_CACHE) {
299 /*
300 * Not sequential, do not do any read-ahead
301 */
302 if (maxra <= 1)
303 return 0;
304
305 /*
306 * No read-ahead mark, do not do any read-ahead
307 * yet.
308 */
309 if ((bp->b_flags & B_RAM) == 0)
310 return 0;
311
312 /*
313 * We hit a read-ahead-mark, figure out how much read-ahead
314 * to do (maxra) and where to start (loffset).
315 *
316 * Typically the way this works is that B_RAM is set in the
317 * middle of the cluster and triggers an overlapping
318 * read-ahead of 1/2 a cluster more blocks. This ensures
319 * that the cluster read-ahead scales with the read-ahead
320 * count and is thus better-able to absorb the caller's
321 * latency.
322 *
323 * Estimate where the next unread block will be by assuming
324 * that the B_RAM's are placed at the half-way point.
325 */
326 bp->b_flags &= ~B_RAM;
327
328 i = maxra / 2;
329 rbp = findblk(vp, loffset + i * blksize, FINDBLK_TEST);
330 if (rbp == NULL || (rbp->b_flags & B_CACHE) == 0) {
331 while (i) {
332 --i;
333 rbp = findblk(vp, loffset + i * blksize,
334 FINDBLK_TEST);
335 if (rbp) {
336 ++i;
337 break;
338 }
339 }
340 } else {
341 while (i < maxra) {
342 rbp = findblk(vp, loffset + i * blksize,
343 FINDBLK_TEST);
344 if (rbp == NULL)
345 break;
346 ++i;
347 }
348 }
349
350 /*
351 * We got everything or everything is in the cache, no
352 * point continuing.
353 */
354 if (i >= maxra)
355 return 0;
356
357 /*
358 * Calculate where to start the read-ahead and how much
359 * to do. Generally speaking we want to read-ahead by
360 * (maxra) when we've found a read-ahead mark. We do
361 * not want to reduce maxra here as it will cause
362 * successive read-ahead I/O's to be smaller and smaller.
363 *
364 * However, we have to make sure we don't break the
365 * filesize limitation for the clustered operation.
366 */
367 loffset += i * blksize;
368 reqbp = bp = NULL;
369
370 if (loffset >= filesize)
371 return 0;
372 if (loffset + maxra * blksize > filesize) {
373 maxreq = filesize - loffset;
374 maxra = (int)(maxreq / blksize);
375 }
376
377 /*
378 * Set RAM on first read-ahead block since we still have
379 * approximate maxra/2 blocks ahead of us that are already
380 * cached or in-progress.
381 */
382 sr = 1;
383 } else {
384 /*
385 * Start block is not valid, we will want to do a
386 * full read-ahead.
387 */
388 __debugvar off_t firstread = bp->b_loffset;
389 int nblks;
390
391 /*
392 * Set-up synchronous read for bp.
393 */
394 bp->b_cmd = BUF_CMD_READ;
395 bp->b_bio1.bio_done = biodone_sync;
396 bp->b_bio1.bio_flags |= BIO_SYNC;
397
398 KASSERT(firstread != NOOFFSET,
399 ("cluster_read: no buffer offset"));
400
401 nblks = calc_rbuild_reqsize(maxra, maxrbuild);
402
403 /*
404 * Set RAM half-way through the full-cluster.
405 */
406 sr = (maxra + 1) / 2;
407
408 if (nblks > 1) {
409 int burstbytes;
410
411 error = VOP_BMAP(vp, loffset, &doffset,
412 &burstbytes, NULL, BUF_CMD_READ);
413 if (error)
414 goto single_block_read;
415 if (nblks > burstbytes / blksize)
416 nblks = burstbytes / blksize;
417 if (doffset == NOOFFSET)
418 goto single_block_read;
419 if (nblks <= 1)
420 goto single_block_read;
421
422 bp = cluster_rbuild(vp, filesize, loffset,
423 doffset, blksize, nblks, bp, &sr);
424 loffset += bp->b_bufsize;
425 maxra -= bp->b_bufsize / blksize;
426 } else {
427 single_block_read:
428 /*
429 * If it isn't in the cache, then get a chunk from
430 * disk if sequential, otherwise just get the block.
431 */
432 loffset += blksize;
433 --maxra;
434 }
435 }
436
437 /*
438 * If B_CACHE was not set issue bp. bp will either be an
439 * asynchronous cluster buf or a synchronous single-buf.
440 * If it is a single buf it will be the same as reqbp.
441 *
442 * NOTE: Once an async cluster buf is issued bp becomes invalid.
443 */
444 if (bp) {
445 #if defined(CLUSTERDEBUG)
446 if (rcluster)
447 kprintf("S(%012jx,%d,%d)\n",
448 (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
449 #endif
450 if ((bp->b_flags & B_CLUSTER) == 0)
451 vfs_busy_pages(vp, bp);
452 bp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA);
453 bp->b_flags |= bflags;
454 vn_strategy(vp, &bp->b_bio1);
455 /* bp invalid now */
456 bp = NULL;
457 }
458
459 #if defined(CLUSTERDEBUG)
460 if (rcluster)
461 kprintf("cluster_rd %016jx/%d maxra=%d sr=%d\n",
462 loffset, blksize, maxra, sr);
463 #endif
464
465 /*
466 * If we have been doing sequential I/O, then do some read-ahead.
467 * The code above us should have positioned us at the next likely
468 * offset.
469 *
470 * Only mess with buffers which we can immediately lock. HAMMER
471 * will do device-readahead irrespective of what the blocks
472 * represent.
473 *
474 * Set B_RAM on the first buffer (the next likely offset needing
475 * read-ahead), under the assumption that there are still
476 * approximately maxra/2 blocks good ahead of us.
477 */
478 while (maxra > 0) {
479 int burstbytes;
480 int nblks;
481
482 rbp = getblk(vp, loffset, blksize,
483 GETBLK_SZMATCH | GETBLK_NOWAIT | GETBLK_KVABIO,
484 0);
485 #if defined(CLUSTERDEBUG)
486 if (rcluster) {
487 kprintf("read-ahead %016jx rbp=%p ",
488 loffset, rbp);
489 }
490 #endif
491 if (rbp == NULL)
492 goto no_read_ahead;
493 if ((rbp->b_flags & B_CACHE)) {
494 bqrelse(rbp);
495 goto no_read_ahead;
496 }
497
498 /*
499 * If BMAP is not supported or has an issue, we still do
500 * (maxra) read-ahead, but we do not try to use rbuild.
501 */
502 error = VOP_BMAP(vp, loffset, &doffset,
503 &burstbytes, NULL, BUF_CMD_READ);
504 if (error || doffset == NOOFFSET) {
505 nblks = 1;
506 doffset = NOOFFSET;
507 } else {
508 nblks = calc_rbuild_reqsize(maxra, maxrbuild);
509 if (nblks > burstbytes / blksize)
510 nblks = burstbytes / blksize;
511 }
512 rbp->b_cmd = BUF_CMD_READ;
513
514 if (nblks > 1) {
515 rbp = cluster_rbuild(vp, filesize, loffset,
516 doffset, blksize,
517 nblks, rbp, &sr);
518 } else {
519 rbp->b_bio2.bio_offset = doffset;
520 if (--sr == 0)
521 cluster_setram(rbp);
522 }
523
524 rbp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA);
525 rbp->b_flags |= bflags;
526
527 if ((rbp->b_flags & B_CLUSTER) == 0)
528 vfs_busy_pages(vp, rbp);
529 BUF_KERNPROC(rbp);
530 loffset += rbp->b_bufsize;
531 maxra -= rbp->b_bufsize / blksize;
532 vn_strategy(vp, &rbp->b_bio1);
533 /* rbp invalid now */
534 }
535
536 /*
537 * Wait for our original buffer to complete its I/O. reqbp will
538 * be NULL if the original buffer was B_CACHE. We are returning
539 * (*bpp) which is the same as reqbp when reqbp != NULL.
540 */
541 no_read_ahead:
542 if (reqbp) {
543 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
544 error = biowait(&reqbp->b_bio1, "clurd");
545 } else {
546 error = 0;
547 }
548 return (error);
549 }
550
551 /*
552 * This replaces breadcb(), providing an asynchronous read of the requested
553 * buffer with a callback, plus an asynchronous read-ahead within the
554 * specified bounds.
555 *
556 * The callback must check whether BIO_DONE is set in the bio and issue
557 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
558 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
559 *
560 * filesize - read-ahead @ blksize will not cross this boundary
561 * loffset - loffset for returned *bpp
562 * blksize - blocksize for returned *bpp and read-ahead bps
563 * minreq - minimum (not a hard minimum) in bytes, typically reflects
564 * a higher level uio resid.
565 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB)
566 * bpp - return buffer (*bpp) for (loffset,blksize)
567 */
568 void
cluster_readcb(struct vnode * vp,off_t filesize,off_t loffset,int blksize,int bflags,size_t minreq,size_t maxreq,void (* func)(struct bio *),void * arg)569 cluster_readcb(struct vnode *vp, off_t filesize, off_t loffset, int blksize,
570 int bflags, size_t minreq, size_t maxreq,
571 void (*func)(struct bio *), void *arg)
572 {
573 struct buf *bp, *rbp, *reqbp;
574 off_t origoffset;
575 off_t doffset;
576 int i;
577 int maxra;
578 int maxrbuild;
579 int sr;
580 int blkflags = (bflags & B_KVABIO) ? GETBLK_KVABIO : 0;
581
582 sr = 0;
583
584 /*
585 * Calculate the desired read-ahead in blksize'd blocks (maxra).
586 * To do this we calculate maxreq.
587 *
588 * maxreq typically starts out as a sequential heuristic. If the
589 * high level uio/resid is bigger (minreq), we pop maxreq up to
590 * minreq. This represents the case where random I/O is being
591 * performed by the userland is issuing big read()'s.
592 *
593 * Then we limit maxreq to max_readahead to ensure it is a reasonable
594 * value.
595 *
596 * Finally we must ensure that (loffset + maxreq) does not cross the
597 * boundary (filesize) for the current blocksize. If we allowed it
598 * to cross we could end up with buffers past the boundary with the
599 * wrong block size (HAMMER large-data areas use mixed block sizes).
600 * minreq is also absolutely limited to filesize.
601 */
602 if (maxreq < minreq)
603 maxreq = minreq;
604 /* minreq not used beyond this point */
605
606 if (maxreq > max_readahead) {
607 maxreq = max_readahead;
608 if (maxreq > 16 * 1024 * 1024)
609 maxreq = 16 * 1024 * 1024;
610 }
611 if (maxreq < blksize)
612 maxreq = blksize;
613 if (loffset + maxreq > filesize) {
614 if (loffset > filesize)
615 maxreq = 0;
616 else
617 maxreq = filesize - loffset;
618 }
619
620 maxra = (int)(maxreq / blksize);
621
622 /*
623 * Get the requested block.
624 */
625 reqbp = bp = getblk(vp, loffset, blksize, blkflags, 0);
626 origoffset = loffset;
627
628 /*
629 * Calculate the maximum cluster size for a single I/O, used
630 * by cluster_rbuild().
631 */
632 maxrbuild = vmaxiosize(vp) / blksize;
633
634 /*
635 * if it is in the cache, then check to see if the reads have been
636 * sequential. If they have, then try some read-ahead, otherwise
637 * back-off on prospective read-aheads.
638 */
639 if (bp->b_flags & B_CACHE) {
640 /*
641 * Setup for func() call whether we do read-ahead or not.
642 */
643 bp->b_bio1.bio_caller_info1.ptr = arg;
644 bp->b_bio1.bio_flags |= BIO_DONE;
645
646 /*
647 * Not sequential, do not do any read-ahead
648 */
649 if (maxra <= 1)
650 goto no_read_ahead;
651
652 /*
653 * No read-ahead mark, do not do any read-ahead
654 * yet.
655 */
656 if ((bp->b_flags & B_RAM) == 0)
657 goto no_read_ahead;
658 bp->b_flags &= ~B_RAM;
659
660 /*
661 * We hit a read-ahead-mark, figure out how much read-ahead
662 * to do (maxra) and where to start (loffset).
663 *
664 * Shortcut the scan. Typically the way this works is that
665 * we've built up all the blocks inbetween except for the
666 * last in previous iterations, so if the second-to-last
667 * block is present we just skip ahead to it.
668 *
669 * This algorithm has O(1) cpu in the steady state no
670 * matter how large maxra is.
671 */
672 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
673 i = maxra - 1;
674 else
675 i = 1;
676 while (i < maxra) {
677 if (findblk(vp, loffset + i * blksize,
678 FINDBLK_TEST) == NULL) {
679 break;
680 }
681 ++i;
682 }
683
684 /*
685 * We got everything or everything is in the cache, no
686 * point continuing.
687 */
688 if (i >= maxra)
689 goto no_read_ahead;
690
691 /*
692 * Calculate where to start the read-ahead and how much
693 * to do. Generally speaking we want to read-ahead by
694 * (maxra) when we've found a read-ahead mark. We do
695 * not want to reduce maxra here as it will cause
696 * successive read-ahead I/O's to be smaller and smaller.
697 *
698 * However, we have to make sure we don't break the
699 * filesize limitation for the clustered operation.
700 */
701 loffset += i * blksize;
702 bp = NULL;
703 /* leave reqbp intact to force function callback */
704
705 if (loffset >= filesize)
706 goto no_read_ahead;
707 if (loffset + maxra * blksize > filesize) {
708 maxreq = filesize - loffset;
709 maxra = (int)(maxreq / blksize);
710 }
711 sr = 1;
712 } else {
713 /*
714 * bp is not valid, no prior cluster in progress so get a
715 * full cluster read-ahead going.
716 */
717 __debugvar off_t firstread = bp->b_loffset;
718 int nblks;
719 int error;
720
721 /*
722 * Set-up synchronous read for bp.
723 */
724 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA);
725 bp->b_flags |= bflags;
726 bp->b_cmd = BUF_CMD_READ;
727 bp->b_bio1.bio_done = func;
728 bp->b_bio1.bio_caller_info1.ptr = arg;
729 BUF_KERNPROC(bp);
730 reqbp = NULL; /* don't func() reqbp, it's running async */
731
732 KASSERT(firstread != NOOFFSET,
733 ("cluster_read: no buffer offset"));
734
735 /*
736 * nblks is our cluster_rbuild request size, limited
737 * primarily by the device.
738 */
739 nblks = calc_rbuild_reqsize(maxra, maxrbuild);
740
741 /*
742 * Set RAM half-way through the full-cluster.
743 */
744 sr = (maxra + 1) / 2;
745
746 if (nblks > 1) {
747 int burstbytes;
748
749 error = VOP_BMAP(vp, loffset, &doffset,
750 &burstbytes, NULL, BUF_CMD_READ);
751 if (error)
752 goto single_block_read;
753 if (nblks > burstbytes / blksize)
754 nblks = burstbytes / blksize;
755 if (doffset == NOOFFSET)
756 goto single_block_read;
757 if (nblks <= 1)
758 goto single_block_read;
759
760 bp = cluster_rbuild(vp, filesize, loffset,
761 doffset, blksize, nblks, bp, &sr);
762 loffset += bp->b_bufsize;
763 maxra -= bp->b_bufsize / blksize;
764 } else {
765 single_block_read:
766 /*
767 * If it isn't in the cache, then get a chunk from
768 * disk if sequential, otherwise just get the block.
769 */
770 loffset += blksize;
771 --maxra;
772 }
773 }
774
775 /*
776 * If bp != NULL then B_CACHE was *NOT* set and bp must be issued.
777 * bp will either be an asynchronous cluster buf or an asynchronous
778 * single-buf.
779 *
780 * NOTE: Once an async cluster buf is issued bp becomes invalid.
781 */
782 if (bp) {
783 #if defined(CLUSTERDEBUG)
784 if (rcluster)
785 kprintf("S(%012jx,%d,%d)\n",
786 (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
787 #endif
788 if ((bp->b_flags & B_CLUSTER) == 0)
789 vfs_busy_pages(vp, bp);
790 bp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA);
791 bp->b_flags |= bflags;
792 vn_strategy(vp, &bp->b_bio1);
793 /* bp invalid now */
794 bp = NULL;
795 }
796
797 #if defined(CLUSTERDEBUG)
798 if (rcluster)
799 kprintf("cluster_rd %016jx/%d maxra=%d sr=%d\n",
800 loffset, blksize, maxra, sr);
801 #endif
802
803 /*
804 * If we have been doing sequential I/O, then do some read-ahead.
805 * The code above us should have positioned us at the next likely
806 * offset.
807 *
808 * Only mess with buffers which we can immediately lock. HAMMER
809 * will do device-readahead irrespective of what the blocks
810 * represent.
811 */
812 while (maxra > 0) {
813 int burstbytes;
814 int error;
815 int nblks;
816
817 rbp = getblk(vp, loffset, blksize,
818 GETBLK_SZMATCH | GETBLK_NOWAIT | GETBLK_KVABIO,
819 0);
820 if (rbp == NULL)
821 goto no_read_ahead;
822 if ((rbp->b_flags & B_CACHE)) {
823 bqrelse(rbp);
824 goto no_read_ahead;
825 }
826
827 /*
828 * If BMAP is not supported or has an issue, we still do
829 * (maxra) read-ahead, but we do not try to use rbuild.
830 */
831 error = VOP_BMAP(vp, loffset, &doffset,
832 &burstbytes, NULL, BUF_CMD_READ);
833 if (error || doffset == NOOFFSET) {
834 nblks = 1;
835 doffset = NOOFFSET;
836 } else {
837 nblks = calc_rbuild_reqsize(maxra, maxrbuild);
838 if (nblks > burstbytes / blksize)
839 nblks = burstbytes / blksize;
840 }
841 rbp->b_cmd = BUF_CMD_READ;
842
843 if (nblks > 1) {
844 rbp = cluster_rbuild(vp, filesize, loffset,
845 doffset, blksize,
846 nblks, rbp, &sr);
847 } else {
848 rbp->b_bio2.bio_offset = doffset;
849 if (--sr == 0)
850 cluster_setram(rbp);
851 }
852
853 rbp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA);
854 rbp->b_flags |= bflags;
855
856 if ((rbp->b_flags & B_CLUSTER) == 0)
857 vfs_busy_pages(vp, rbp);
858 BUF_KERNPROC(rbp);
859 loffset += rbp->b_bufsize;
860 maxra -= rbp->b_bufsize / blksize;
861 vn_strategy(vp, &rbp->b_bio1);
862 /* rbp invalid now */
863 }
864
865 /*
866 * If reqbp is non-NULL it had B_CACHE set and we issue the
867 * function callback synchronously.
868 *
869 * Note that we may start additional asynchronous I/O before doing
870 * the func() callback for the B_CACHE case
871 */
872 no_read_ahead:
873 if (reqbp)
874 func(&reqbp->b_bio1);
875 }
876
877 /*
878 * If blocks are contiguous on disk, use this to provide clustered
879 * read ahead. We will read as many blocks as possible sequentially
880 * and then parcel them up into logical blocks in the buffer hash table.
881 *
882 * This function either returns a cluster buf or it returns fbp. fbp is
883 * already expected to be set up as a synchronous or asynchronous request.
884 *
885 * If a cluster buf is returned it will always be async.
886 *
887 * (*srp) counts down original blocks to determine where B_RAM should be set.
888 * Set B_RAM when *srp drops to 0. If (*srp) starts at 0, B_RAM will not be
889 * set on any buffer. Make sure B_RAM is cleared on any other buffers to
890 * prevent degenerate read-aheads from being generated.
891 */
892 static struct buf *
cluster_rbuild(struct vnode * vp,off_t filesize,off_t loffset,off_t doffset,int blksize,int run,struct buf * fbp,int * srp)893 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
894 int blksize, int run, struct buf *fbp, int *srp)
895 {
896 struct buf *bp, *tbp;
897 off_t boffset;
898 int i, j;
899 int maxiosize = vmaxiosize(vp);
900
901 /*
902 * avoid a division
903 */
904 while (loffset + run * blksize > filesize) {
905 --run;
906 }
907
908 tbp = fbp;
909 tbp->b_bio2.bio_offset = doffset;
910 if (((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
911 if (--*srp == 0)
912 cluster_setram(tbp);
913 else
914 cluster_clrram(tbp);
915 return tbp;
916 }
917
918 /*
919 * Get a pbuf, limit cluster I/O on a per-device basis. If
920 * doing cluster I/O for a file, limit cluster I/O on a
921 * per-mount basis.
922 */
923 if (vp->v_type == VCHR || vp->v_type == VBLK)
924 bp = trypbuf_kva(&vp->v_pbuf_count);
925 else
926 bp = trypbuf_kva(&vp->v_mount->mnt_pbuf_count);
927
928 if (bp == NULL)
929 return tbp;
930
931 /*
932 * We are synthesizing a buffer out of vm_page_t's, but
933 * if the block size is not page aligned then the starting
934 * address may not be either. Inherit the b_data offset
935 * from the original buffer.
936 */
937 bp->b_vp = vp;
938 bp->b_data = (char *)((vm_offset_t)bp->b_data |
939 ((vm_offset_t)tbp->b_data & PAGE_MASK));
940 bp->b_flags |= B_CLUSTER | B_VMIO | B_KVABIO;
941 bp->b_cmd = BUF_CMD_READ;
942 bp->b_bio1.bio_done = cluster_callback; /* default to async */
943 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
944 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
945 bp->b_loffset = loffset;
946 bp->b_bio2.bio_offset = doffset;
947 KASSERT(bp->b_loffset != NOOFFSET,
948 ("cluster_rbuild: no buffer offset"));
949
950 bp->b_bcount = 0;
951 bp->b_bufsize = 0;
952 bp->b_xio.xio_npages = 0;
953
954 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
955 if (i) {
956 if ((bp->b_xio.xio_npages * PAGE_SIZE) +
957 round_page(blksize) > maxiosize) {
958 break;
959 }
960
961 /*
962 * Shortcut some checks and try to avoid buffers that
963 * would block in the lock. The same checks have to
964 * be made again after we officially get the buffer.
965 */
966 tbp = getblk(vp, loffset + i * blksize, blksize,
967 GETBLK_SZMATCH |
968 GETBLK_NOWAIT |
969 GETBLK_KVABIO,
970 0);
971 if (tbp == NULL)
972 break;
973 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
974 if (tbp->b_xio.xio_pages[j]->valid)
975 break;
976 }
977 if (j != tbp->b_xio.xio_npages) {
978 bqrelse(tbp);
979 break;
980 }
981
982 /*
983 * Stop scanning if the buffer is fuly valid
984 * (marked B_CACHE), or locked (may be doing a
985 * background write), or if the buffer is not
986 * VMIO backed. The clustering code can only deal
987 * with VMIO-backed buffers.
988 */
989 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
990 (tbp->b_flags & B_VMIO) == 0 ||
991 (LIST_FIRST(&tbp->b_dep) != NULL &&
992 buf_checkread(tbp))
993 ) {
994 bqrelse(tbp);
995 break;
996 }
997
998 /*
999 * The buffer must be completely invalid in order to
1000 * take part in the cluster. If it is partially valid
1001 * then we stop.
1002 */
1003 for (j = 0;j < tbp->b_xio.xio_npages; j++) {
1004 if (tbp->b_xio.xio_pages[j]->valid)
1005 break;
1006 }
1007 if (j != tbp->b_xio.xio_npages) {
1008 bqrelse(tbp);
1009 break;
1010 }
1011
1012 /*
1013 * Depress the priority of buffers not explicitly
1014 * requested.
1015 */
1016 /* tbp->b_flags |= B_AGE; */
1017
1018 /*
1019 * Set the block number if it isn't set, otherwise
1020 * if it is make sure it matches the block number we
1021 * expect.
1022 */
1023 if (tbp->b_bio2.bio_offset == NOOFFSET) {
1024 tbp->b_bio2.bio_offset = boffset;
1025 } else if (tbp->b_bio2.bio_offset != boffset) {
1026 brelse(tbp);
1027 break;
1028 }
1029 }
1030
1031 /*
1032 * Set B_RAM if (*srp) is 1. B_RAM is only set on one buffer
1033 * in the cluster, including potentially the first buffer
1034 * once we start streaming the read-aheads.
1035 */
1036 if (--*srp == 0)
1037 cluster_setram(tbp);
1038 else
1039 cluster_clrram(tbp);
1040
1041 /*
1042 * The passed-in tbp (i == 0) will already be set up for
1043 * async or sync operation. All other tbp's acquire in
1044 * our loop are set up for async operation.
1045 */
1046 tbp->b_cmd = BUF_CMD_READ;
1047 BUF_KERNPROC(tbp);
1048 cluster_append(&bp->b_bio1, tbp);
1049 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
1050 vm_page_t m;
1051
1052 m = tbp->b_xio.xio_pages[j];
1053 vm_page_busy_wait(m, FALSE, "clurpg");
1054 vm_page_io_start(m);
1055 vm_page_wakeup(m);
1056 vm_object_pip_add(m->object, 1);
1057 if ((bp->b_xio.xio_npages == 0) ||
1058 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
1059 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
1060 bp->b_xio.xio_npages++;
1061 }
1062 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) {
1063 tbp->b_xio.xio_pages[j] = bogus_page;
1064 tbp->b_flags |= B_HASBOGUS;
1065 }
1066 }
1067 /*
1068 * XXX shouldn't this be += size for both, like in
1069 * cluster_wbuild()?
1070 *
1071 * Don't inherit tbp->b_bufsize as it may be larger due to
1072 * a non-page-aligned size. Instead just aggregate using
1073 * 'size'.
1074 */
1075 if (tbp->b_bcount != blksize)
1076 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
1077 if (tbp->b_bufsize != blksize)
1078 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
1079 bp->b_bcount += blksize;
1080 bp->b_bufsize += blksize;
1081 }
1082
1083 /*
1084 * Fully valid pages in the cluster are already good and do not need
1085 * to be re-read from disk. Replace the page with bogus_page
1086 */
1087 for (j = 0; j < bp->b_xio.xio_npages; j++) {
1088 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
1089 VM_PAGE_BITS_ALL) {
1090 bp->b_xio.xio_pages[j] = bogus_page;
1091 bp->b_flags |= B_HASBOGUS;
1092 }
1093 }
1094 if (bp->b_bufsize > bp->b_kvasize) {
1095 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
1096 bp->b_bufsize, bp->b_kvasize);
1097 }
1098 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
1099 (vm_page_t *)bp->b_xio.xio_pages,
1100 bp->b_xio.xio_npages);
1101 BUF_KERNPROC(bp);
1102 return (bp);
1103 }
1104
1105 /*
1106 * Cleanup after a clustered read or write.
1107 * This is complicated by the fact that any of the buffers might have
1108 * extra memory (if there were no empty buffer headers at allocbuf time)
1109 * that we will need to shift around.
1110 *
1111 * The returned bio is &bp->b_bio1
1112 */
1113 static void
cluster_callback(struct bio * bio)1114 cluster_callback(struct bio *bio)
1115 {
1116 struct buf *bp = bio->bio_buf;
1117 struct buf *tbp;
1118 struct buf *next;
1119 struct vnode *vp;
1120 int error = 0;
1121 int bpflags;
1122
1123 /*
1124 * Must propogate errors to all the components. A short read (EOF)
1125 * is a critical error.
1126 */
1127 if (bp->b_flags & B_ERROR) {
1128 error = bp->b_error;
1129 } else if (bp->b_bcount != bp->b_bufsize) {
1130 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
1131 }
1132
1133 pmap_qremove_noinval(trunc_page((vm_offset_t) bp->b_data),
1134 bp->b_xio.xio_npages);
1135
1136 /*
1137 * Retrieve the cluster head and dispose of the cluster buffer.
1138 * the vp is only valid while we hold one or more cluster elements,
1139 * so we have to do this before disposing of them.
1140 */
1141 tbp = bio->bio_caller_info1.cluster_head;
1142 bio->bio_caller_info1.cluster_head = NULL;
1143 bpflags = bp->b_flags;
1144 vp = bp->b_vp;
1145 bp->b_vp = NULL;
1146
1147 if (vp->v_type == VCHR || vp->v_type == VBLK)
1148 relpbuf(bp, &vp->v_pbuf_count);
1149 else
1150 relpbuf(bp, &vp->v_mount->mnt_pbuf_count);
1151 bp = NULL; /* SAFETY */
1152
1153 /*
1154 * Move memory from the large cluster buffer into the component
1155 * buffers and mark IO as done on these. Since the memory map
1156 * is the same, no actual copying is required.
1157 *
1158 * (And we already disposed of the larger cluster buffer)
1159 */
1160 while (tbp) {
1161 next = tbp->b_cluster_next;
1162 if (error) {
1163 tbp->b_flags |= B_ERROR | B_IOISSUED;
1164 tbp->b_error = error;
1165 } else {
1166 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
1167 tbp->b_flags &= ~(B_ERROR | B_INVAL);
1168 if (tbp->b_cmd == BUF_CMD_READ) {
1169 tbp->b_flags = (tbp->b_flags & ~B_NOTMETA) |
1170 (bpflags & B_NOTMETA);
1171 }
1172 tbp->b_flags |= B_IOISSUED;
1173
1174 /*
1175 * XXX the bdwrite()/bqrelse() issued during
1176 * cluster building clears B_RELBUF (see bqrelse()
1177 * comment). If direct I/O was specified, we have
1178 * to restore it here to allow the buffer and VM
1179 * to be freed.
1180 */
1181 if (tbp->b_flags & B_DIRECT)
1182 tbp->b_flags |= B_RELBUF;
1183
1184 /*
1185 * XXX I think biodone() below will do this, but do
1186 * it here anyway for consistency.
1187 */
1188 if (tbp->b_cmd == BUF_CMD_WRITE)
1189 bundirty(tbp);
1190 }
1191 biodone(&tbp->b_bio1);
1192 tbp = next;
1193 }
1194 }
1195
1196 /*
1197 * Implement modified write build for cluster.
1198 *
1199 * write_behind = 0 write behind disabled
1200 * write_behind = 1 write behind normal (default)
1201 * write_behind = 2 write behind backed-off
1202 *
1203 * In addition, write_behind is only activated for files that have
1204 * grown past a certain size (default 10MB). Otherwise temporary files
1205 * wind up generating a lot of unnecessary disk I/O.
1206 */
1207 static __inline int
cluster_wbuild_wb(struct vnode * vp,int blksize,off_t start_loffset,int len)1208 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
1209 {
1210 int r = 0;
1211
1212 switch(write_behind) {
1213 case 2:
1214 if (start_loffset < len)
1215 break;
1216 start_loffset -= len;
1217 /* fall through */
1218 case 1:
1219 if (vp->v_filesize >= write_behind_minfilesize) {
1220 r = cluster_wbuild(vp, NULL, blksize,
1221 start_loffset, len);
1222 }
1223 /* fall through */
1224 default:
1225 /* fall through */
1226 break;
1227 }
1228 return(r);
1229 }
1230
1231 /*
1232 * Do clustered write for FFS.
1233 *
1234 * Three cases:
1235 * 1. Write is not sequential (write asynchronously)
1236 * Write is sequential:
1237 * 2. beginning of cluster - begin cluster
1238 * 3. middle of a cluster - add to cluster
1239 * 4. end of a cluster - asynchronously write cluster
1240 *
1241 * WARNING! vnode fields are not locked and must ONLY be used heuristically.
1242 */
1243 void
cluster_write(struct buf * bp,off_t filesize,int blksize,int seqcount)1244 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
1245 {
1246 struct vnode *vp;
1247 off_t loffset;
1248 int maxclen, cursize;
1249 int async;
1250 cluster_cache_t dummy;
1251 cluster_cache_t *cc;
1252
1253 vp = bp->b_vp;
1254 if (vp->v_type == VREG)
1255 async = vp->v_mount->mnt_flag & MNT_ASYNC;
1256 else
1257 async = 0;
1258 loffset = bp->b_loffset;
1259 KASSERT(bp->b_loffset != NOOFFSET,
1260 ("cluster_write: no buffer offset"));
1261
1262 cc = cluster_getcache(&dummy, vp, loffset);
1263
1264 /*
1265 * Initialize vnode to beginning of file.
1266 */
1267 if (loffset == 0)
1268 cc->v_lasta = cc->v_clen = cc->v_cstart = cc->v_lastw = 0;
1269
1270 if (cc->v_clen == 0 || loffset != cc->v_lastw ||
1271 (bp->b_bio2.bio_offset != NOOFFSET &&
1272 (bp->b_bio2.bio_offset != cc->v_lasta))) {
1273 /*
1274 * Next block is not logically sequential, or, if physical
1275 * block offsets are available, not physically sequential.
1276 *
1277 * If physical block offsets are not available we only
1278 * get here if we weren't logically sequential.
1279 */
1280 maxclen = vmaxiosize(vp);
1281 if (cc->v_clen != 0) {
1282 /*
1283 * Next block is not sequential.
1284 *
1285 * If we are not writing at end of file, the process
1286 * seeked to another point in the file since its last
1287 * write, or we have reached our maximum cluster size,
1288 * then push the previous cluster. Otherwise try
1289 * reallocating to make it sequential.
1290 *
1291 * Change to algorithm: only push previous cluster if
1292 * it was sequential from the point of view of the
1293 * seqcount heuristic, otherwise leave the buffer
1294 * intact so we can potentially optimize the I/O
1295 * later on in the buf_daemon or update daemon
1296 * flush.
1297 */
1298 cursize = cc->v_lastw - cc->v_cstart;
1299 if (bp->b_loffset + blksize < filesize ||
1300 loffset != cc->v_lastw ||
1301 cc->v_clen <= cursize) {
1302 if (!async && seqcount > 0) {
1303 cluster_wbuild_wb(vp, blksize,
1304 cc->v_cstart, cursize);
1305 }
1306 } else {
1307 struct buf **bpp, **endbp;
1308 struct cluster_save *buflist;
1309
1310 buflist = cluster_collectbufs(cc, vp,
1311 bp, blksize);
1312 endbp = &buflist->bs_children
1313 [buflist->bs_nchildren - 1];
1314 if (VOP_REALLOCBLKS(vp, buflist)) {
1315 /*
1316 * Failed, push the previous cluster
1317 * if *really* writing sequentially
1318 * in the logical file (seqcount > 1),
1319 * otherwise delay it in the hopes that
1320 * the low level disk driver can
1321 * optimize the write ordering.
1322 *
1323 * NOTE: We do not brelse the last
1324 * element which is bp, and we
1325 * do not return here.
1326 */
1327 for (bpp = buflist->bs_children;
1328 bpp < endbp; bpp++)
1329 brelse(*bpp);
1330 kfree(buflist, M_SEGMENT);
1331 if (seqcount > 1) {
1332 cluster_wbuild_wb(vp,
1333 blksize, cc->v_cstart,
1334 cursize);
1335 }
1336 } else {
1337 /*
1338 * Succeeded, keep building cluster.
1339 */
1340 for (bpp = buflist->bs_children;
1341 bpp <= endbp; bpp++)
1342 bdwrite(*bpp);
1343 kfree(buflist, M_SEGMENT);
1344 cc->v_lastw = loffset + blksize;
1345 cc->v_lasta = bp->b_bio2.bio_offset +
1346 blksize;
1347 cluster_putcache(cc);
1348 return;
1349 }
1350 }
1351 }
1352
1353 /*
1354 * Consider beginning a cluster. If at end of file, make
1355 * cluster as large as possible, otherwise find size of
1356 * existing cluster.
1357 */
1358 if ((vp->v_type == VREG) &&
1359 bp->b_loffset + blksize < filesize &&
1360 (bp->b_bio2.bio_offset == NOOFFSET) &&
1361 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
1362 bp->b_bio2.bio_offset == NOOFFSET)) {
1363 bdwrite(bp);
1364 cc->v_clen = 0;
1365 cc->v_lasta = bp->b_bio2.bio_offset + blksize;
1366 cc->v_cstart = loffset;
1367 cc->v_lastw = loffset + blksize;
1368 cluster_putcache(cc);
1369 return;
1370 }
1371 if (maxclen > blksize)
1372 cc->v_clen = maxclen;
1373 else
1374 cc->v_clen = blksize;
1375 if (!async && cc->v_clen == 0) { /* I/O not contiguous */
1376 cc->v_cstart = loffset;
1377 bdwrite(bp);
1378 } else { /* Wait for rest of cluster */
1379 cc->v_cstart = loffset;
1380 bdwrite(bp);
1381 }
1382 } else if (loffset == cc->v_cstart + cc->v_clen) {
1383 /*
1384 * At end of cluster, write it out if seqcount tells us we
1385 * are operating sequentially, otherwise let the buf or
1386 * update daemon handle it.
1387 */
1388 bdwrite(bp);
1389 if (seqcount > 1)
1390 cluster_wbuild_wb(vp, blksize, cc->v_cstart,
1391 cc->v_clen + blksize);
1392 cc->v_clen = 0;
1393 cc->v_cstart = loffset;
1394 } else if (vm_paging_severe() &&
1395 bp->b_loffset + blksize < filesize) {
1396 /*
1397 * We are low on memory, get it going NOW. However, do not
1398 * try to push out a partial block at the end of the file
1399 * as this could lead to extremely non-optimal write activity.
1400 */
1401 bawrite(bp);
1402 } else {
1403 /*
1404 * In the middle of a cluster, so just delay the I/O for now.
1405 */
1406 bdwrite(bp);
1407 }
1408 cc->v_lastw = loffset + blksize;
1409 cc->v_lasta = bp->b_bio2.bio_offset + blksize;
1410 cluster_putcache(cc);
1411 }
1412
1413 /*
1414 * This is the clustered version of bawrite(). It works similarly to
1415 * cluster_write() except I/O on the buffer is guaranteed to occur.
1416 */
1417 int
cluster_awrite(struct buf * bp)1418 cluster_awrite(struct buf *bp)
1419 {
1420 int total;
1421
1422 /*
1423 * Don't bother if it isn't clusterable.
1424 */
1425 if ((bp->b_flags & B_CLUSTEROK) == 0 ||
1426 bp->b_vp == NULL ||
1427 (bp->b_vp->v_flag & VOBJBUF) == 0) {
1428 total = bp->b_bufsize;
1429 bawrite(bp);
1430 return (total);
1431 }
1432
1433 total = cluster_wbuild(bp->b_vp, &bp, bp->b_bufsize,
1434 bp->b_loffset, vmaxiosize(bp->b_vp));
1435
1436 /*
1437 * If bp is still non-NULL then cluster_wbuild() did not initiate
1438 * I/O on it and we must do so here to provide the API guarantee.
1439 */
1440 if (bp)
1441 bawrite(bp);
1442
1443 return total;
1444 }
1445
1446 /*
1447 * This is an awful lot like cluster_rbuild...wish they could be combined.
1448 * The last lbn argument is the current block on which I/O is being
1449 * performed. Check to see that it doesn't fall in the middle of
1450 * the current block (if last_bp == NULL).
1451 *
1452 * cluster_wbuild() normally does not guarantee anything. If bpp is
1453 * non-NULL and cluster_wbuild() is able to incorporate it into the
1454 * I/O it will set *bpp to NULL, otherwise it will leave it alone and
1455 * the caller must dispose of *bpp.
1456 */
1457 static int
cluster_wbuild(struct vnode * vp,struct buf ** bpp,int blksize,off_t start_loffset,int bytes)1458 cluster_wbuild(struct vnode *vp, struct buf **bpp,
1459 int blksize, off_t start_loffset, int bytes)
1460 {
1461 struct buf *bp, *tbp;
1462 int i, j;
1463 int totalwritten = 0;
1464 int must_initiate;
1465 int maxiosize = vmaxiosize(vp);
1466
1467 while (bytes > 0) {
1468 /*
1469 * If the buffer matches the passed locked & removed buffer
1470 * we used the passed buffer (which might not be B_DELWRI).
1471 *
1472 * Otherwise locate the buffer and determine if it is
1473 * compatible.
1474 */
1475 if (bpp && (*bpp)->b_loffset == start_loffset) {
1476 tbp = *bpp;
1477 *bpp = NULL;
1478 bpp = NULL;
1479 } else {
1480 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK |
1481 FINDBLK_KVABIO);
1482 if (tbp == NULL ||
1483 (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) !=
1484 B_DELWRI ||
1485 (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
1486 if (tbp)
1487 BUF_UNLOCK(tbp);
1488 start_loffset += blksize;
1489 bytes -= blksize;
1490 continue;
1491 }
1492 bremfree(tbp);
1493 }
1494 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
1495
1496 /*
1497 * Extra memory in the buffer, punt on this buffer.
1498 * XXX we could handle this in most cases, but we would
1499 * have to push the extra memory down to after our max
1500 * possible cluster size and then potentially pull it back
1501 * up if the cluster was terminated prematurely--too much
1502 * hassle.
1503 */
1504 if ((tbp->b_flags & B_CLUSTEROK) == 0 ||
1505 tbp->b_bcount != tbp->b_bufsize ||
1506 tbp->b_bcount != blksize ||
1507 bytes == blksize) {
1508 totalwritten += tbp->b_bufsize;
1509 bawrite(tbp);
1510 start_loffset += blksize;
1511 bytes -= blksize;
1512 continue;
1513 }
1514
1515 /*
1516 * Get a pbuf, limit cluster I/O on a per-device basis. If
1517 * doing cluster I/O for a file, limit cluster I/O on a
1518 * per-mount basis.
1519 *
1520 * HAMMER and other filesystems may attempt to queue a massive
1521 * amount of write I/O, using trypbuf() here easily results in
1522 * situation where the I/O stream becomes non-clustered.
1523 */
1524 if (vp->v_type == VCHR || vp->v_type == VBLK)
1525 bp = getpbuf_kva(&vp->v_pbuf_count);
1526 else
1527 bp = getpbuf_kva(&vp->v_mount->mnt_pbuf_count);
1528
1529 /*
1530 * Set up the pbuf. Track our append point with b_bcount
1531 * and b_bufsize. b_bufsize is not used by the device but
1532 * our caller uses it to loop clusters and we use it to
1533 * detect a premature EOF on the block device.
1534 */
1535 bp->b_bcount = 0;
1536 bp->b_bufsize = 0;
1537 bp->b_xio.xio_npages = 0;
1538 bp->b_loffset = tbp->b_loffset;
1539 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
1540 bp->b_vp = vp;
1541
1542 /*
1543 * We are synthesizing a buffer out of vm_page_t's, but
1544 * if the block size is not page aligned then the starting
1545 * address may not be either. Inherit the b_data offset
1546 * from the original buffer.
1547 */
1548 bp->b_data = (char *)((vm_offset_t)bp->b_data |
1549 ((vm_offset_t)tbp->b_data & PAGE_MASK));
1550 bp->b_flags &= ~(B_ERROR | B_NOTMETA);
1551 bp->b_flags |= B_CLUSTER | B_BNOCLIP | B_KVABIO |
1552 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT |
1553 B_NOTMETA));
1554 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
1555 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
1556
1557 /*
1558 * From this location in the file, scan forward to see
1559 * if there are buffers with adjacent data that need to
1560 * be written as well.
1561 *
1562 * IO *must* be initiated on index 0 at this point
1563 * (particularly when called from cluster_awrite()).
1564 */
1565 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
1566 if (i == 0) {
1567 must_initiate = 1;
1568 } else {
1569 /*
1570 * Not first buffer.
1571 */
1572 must_initiate = 0;
1573 tbp = findblk(vp, start_loffset,
1574 FINDBLK_NBLOCK | FINDBLK_KVABIO);
1575 /*
1576 * Buffer not found or could not be locked
1577 * non-blocking.
1578 */
1579 if (tbp == NULL)
1580 break;
1581
1582 /*
1583 * If it IS in core, but has different
1584 * characteristics, then don't cluster
1585 * with it.
1586 */
1587 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
1588 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
1589 != (B_DELWRI | B_CLUSTEROK |
1590 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
1591 (tbp->b_flags & B_LOCKED)
1592 ) {
1593 BUF_UNLOCK(tbp);
1594 break;
1595 }
1596
1597 /*
1598 * Check that the combined cluster
1599 * would make sense with regard to pages
1600 * and would not be too large
1601 *
1602 * WARNING! buf_checkwrite() must be the last
1603 * check made. If it returns 0 then
1604 * we must initiate the I/O.
1605 */
1606 if ((tbp->b_bcount != blksize) ||
1607 ((bp->b_bio2.bio_offset + i) !=
1608 tbp->b_bio2.bio_offset) ||
1609 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
1610 (maxiosize / PAGE_SIZE)) ||
1611 (LIST_FIRST(&tbp->b_dep) &&
1612 buf_checkwrite(tbp))
1613 ) {
1614 BUF_UNLOCK(tbp);
1615 break;
1616 }
1617 if (LIST_FIRST(&tbp->b_dep))
1618 must_initiate = 1;
1619 /*
1620 * Ok, it's passed all the tests,
1621 * so remove it from the free list
1622 * and mark it busy. We will use it.
1623 */
1624 bremfree(tbp);
1625 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
1626 }
1627
1628 /*
1629 * If the IO is via the VM then we do some
1630 * special VM hackery (yuck). Since the buffer's
1631 * block size may not be page-aligned it is possible
1632 * for a page to be shared between two buffers. We
1633 * have to get rid of the duplication when building
1634 * the cluster.
1635 */
1636 if (tbp->b_flags & B_VMIO) {
1637 vm_page_t m;
1638
1639 /*
1640 * Try to avoid deadlocks with the VM system.
1641 * However, we cannot abort the I/O if
1642 * must_initiate is non-zero.
1643 */
1644 if (must_initiate == 0) {
1645 for (j = 0;
1646 j < tbp->b_xio.xio_npages;
1647 ++j) {
1648 m = tbp->b_xio.xio_pages[j];
1649 if (m->busy_count &
1650 PBUSY_LOCKED) {
1651 bqrelse(tbp);
1652 goto finishcluster;
1653 }
1654 }
1655 }
1656
1657 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
1658 m = tbp->b_xio.xio_pages[j];
1659 vm_page_busy_wait(m, FALSE, "clurpg");
1660 vm_page_io_start(m);
1661 vm_page_wakeup(m);
1662 vm_object_pip_add(m->object, 1);
1663 if ((bp->b_xio.xio_npages == 0) ||
1664 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
1665 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
1666 bp->b_xio.xio_npages++;
1667 }
1668 }
1669 }
1670 bp->b_bcount += blksize;
1671 bp->b_bufsize += blksize;
1672
1673 /*
1674 * NOTE: see bwrite/bawrite code for why we no longer
1675 * undirty tbp here.
1676 *
1677 * bundirty(tbp); REMOVED
1678 */
1679 tbp->b_flags &= ~B_ERROR;
1680 tbp->b_cmd = BUF_CMD_WRITE;
1681 BUF_KERNPROC(tbp);
1682 cluster_append(&bp->b_bio1, tbp);
1683
1684 /*
1685 * check for latent dependencies to be handled
1686 */
1687 if (LIST_FIRST(&tbp->b_dep) != NULL)
1688 buf_start(tbp);
1689 }
1690 finishcluster:
1691 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
1692 (vm_page_t *)bp->b_xio.xio_pages,
1693 bp->b_xio.xio_npages);
1694 if (bp->b_bufsize > bp->b_kvasize) {
1695 panic("cluster_wbuild: b_bufsize(%d) "
1696 "> b_kvasize(%d)\n",
1697 bp->b_bufsize, bp->b_kvasize);
1698 }
1699 totalwritten += bp->b_bufsize;
1700 bp->b_dirtyoff = 0;
1701 bp->b_dirtyend = bp->b_bufsize;
1702 bp->b_bio1.bio_done = cluster_callback;
1703 bp->b_cmd = BUF_CMD_WRITE;
1704
1705 vfs_busy_pages(vp, bp);
1706 bsetrunningbufspace(bp, bp->b_bufsize);
1707 BUF_KERNPROC(bp);
1708 vn_strategy(vp, &bp->b_bio1);
1709
1710 bytes -= i;
1711 }
1712 return totalwritten;
1713 }
1714
1715 /*
1716 * Collect together all the buffers in a cluster, plus add one
1717 * additional buffer passed-in.
1718 *
1719 * Only pre-existing buffers whos block size matches blksize are collected.
1720 * (this is primarily because HAMMER1 uses varying block sizes and we don't
1721 * want to override its choices).
1722 *
1723 * This code will not try to collect buffers that it cannot lock, otherwise
1724 * it might deadlock against SMP-friendly filesystems.
1725 */
1726 static struct cluster_save *
cluster_collectbufs(cluster_cache_t * cc,struct vnode * vp,struct buf * last_bp,int blksize)1727 cluster_collectbufs(cluster_cache_t *cc, struct vnode *vp,
1728 struct buf *last_bp, int blksize)
1729 {
1730 struct cluster_save *buflist;
1731 struct buf *bp;
1732 off_t loffset;
1733 int i, len;
1734 int j;
1735 int k;
1736
1737 len = (int)(cc->v_lastw - cc->v_cstart) / blksize;
1738 KKASSERT(len > 0);
1739 buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1740 M_SEGMENT, M_WAITOK);
1741 buflist->bs_nchildren = 0;
1742 buflist->bs_children = (struct buf **) (buflist + 1);
1743 for (loffset = cc->v_cstart, i = 0, j = 0;
1744 i < len;
1745 (loffset += blksize), i++) {
1746 bp = getcacheblk(vp, loffset,
1747 last_bp->b_bcount, GETBLK_SZMATCH |
1748 GETBLK_NOWAIT);
1749 buflist->bs_children[i] = bp;
1750 if (bp == NULL) {
1751 j = i + 1;
1752 } else if (bp->b_bio2.bio_offset == NOOFFSET) {
1753 VOP_BMAP(bp->b_vp, bp->b_loffset,
1754 &bp->b_bio2.bio_offset,
1755 NULL, NULL, BUF_CMD_WRITE);
1756 }
1757 }
1758
1759 /*
1760 * Get rid of gaps
1761 */
1762 for (k = 0; k < j; ++k) {
1763 if (buflist->bs_children[k]) {
1764 bqrelse(buflist->bs_children[k]);
1765 buflist->bs_children[k] = NULL;
1766 }
1767 }
1768 if (j != 0) {
1769 if (j != i) {
1770 bcopy(buflist->bs_children + j,
1771 buflist->bs_children + 0,
1772 sizeof(buflist->bs_children[0]) * (i - j));
1773 }
1774 i -= j;
1775 }
1776 buflist->bs_children[i] = bp = last_bp;
1777 if (bp->b_bio2.bio_offset == NOOFFSET) {
1778 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1779 NULL, NULL, BUF_CMD_WRITE);
1780 }
1781 buflist->bs_nchildren = i + 1;
1782 return (buflist);
1783 }
1784
1785 void
cluster_append(struct bio * bio,struct buf * tbp)1786 cluster_append(struct bio *bio, struct buf *tbp)
1787 {
1788 tbp->b_cluster_next = NULL;
1789 if (bio->bio_caller_info1.cluster_head == NULL) {
1790 bio->bio_caller_info1.cluster_head = tbp;
1791 bio->bio_caller_info2.cluster_tail = tbp;
1792 } else {
1793 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1794 bio->bio_caller_info2.cluster_tail = tbp;
1795 }
1796 }
1797
1798 static
1799 void
cluster_setram(struct buf * bp)1800 cluster_setram(struct buf *bp)
1801 {
1802 bp->b_flags |= B_RAM;
1803 if (bp->b_xio.xio_npages)
1804 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1805 }
1806
1807 static
1808 void
cluster_clrram(struct buf * bp)1809 cluster_clrram(struct buf *bp)
1810 {
1811 bp->b_flags &= ~B_RAM;
1812 if (bp->b_xio.xio_npages)
1813 vm_page_flag_clear(bp->b_xio.xio_pages[0], PG_RAM);
1814 }
1815