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