11541Srgrimes/*- 21541Srgrimes * Copyright (c) 1993 31541Srgrimes * The Regents of the University of California. All rights reserved. 45455Sdg * Modifications/enhancements: 55455Sdg * Copyright (c) 1995 John S. Dyson. All rights reserved. 61541Srgrimes * 71541Srgrimes * Redistribution and use in source and binary forms, with or without 81541Srgrimes * modification, are permitted provided that the following conditions 91541Srgrimes * are met: 101541Srgrimes * 1. Redistributions of source code must retain the above copyright 111541Srgrimes * notice, this list of conditions and the following disclaimer. 121541Srgrimes * 2. Redistributions in binary form must reproduce the above copyright 131541Srgrimes * notice, this list of conditions and the following disclaimer in the 141541Srgrimes * documentation and/or other materials provided with the distribution. 151541Srgrimes * 4. Neither the name of the University nor the names of its contributors 161541Srgrimes * may be used to endorse or promote products derived from this software 171541Srgrimes * without specific prior written permission. 181541Srgrimes * 191541Srgrimes * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 201541Srgrimes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 211541Srgrimes * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 221541Srgrimes * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 231541Srgrimes * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 241541Srgrimes * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 251541Srgrimes * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 261541Srgrimes * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 271541Srgrimes * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 281541Srgrimes * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 291541Srgrimes * SUCH DAMAGE. 301541Srgrimes * 311541Srgrimes * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 321541Srgrimes */ 331541Srgrimes 34116182Sobrien#include <sys/cdefs.h> 35116182Sobrien__FBSDID("$FreeBSD: stable/11/sys/kern/vfs_cluster.c 352947 2019-10-01 23:28:22Z mckusick $"); 36116182Sobrien 3732929Seivind#include "opt_debug_cluster.h" 3832929Seivind 391541Srgrimes#include <sys/param.h> 401549Srgrimes#include <sys/systm.h> 4141168Sbde#include <sys/kernel.h> 421541Srgrimes#include <sys/proc.h> 4360041Sphk#include <sys/bio.h> 441541Srgrimes#include <sys/buf.h> 451541Srgrimes#include <sys/vnode.h> 4641124Sdg#include <sys/malloc.h> 471541Srgrimes#include <sys/mount.h> 48297633Strasz#include <sys/racct.h> 491541Srgrimes#include <sys/resourcevar.h> 50248084Sattilio#include <sys/rwlock.h> 5168885Sdillon#include <sys/vmmeter.h> 526621Sdg#include <vm/vm.h> 5310541Sdyson#include <vm/vm_object.h> 5410541Sdyson#include <vm/vm_page.h> 5548545Smckusick#include <sys/sysctl.h> 561541Srgrimes 5721002Sdyson#if defined(CLUSTERDEBUG) 5821002Sdysonstatic int rcluster= 0; 5991690SeivindSYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, 6091690Seivind "Debug VFS clustering code"); 6121002Sdyson#endif 6221002Sdyson 63167214Swkoszekstatic MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer"); 6441124Sdg 65248508Skibstatic struct cluster_save *cluster_collectbufs(struct vnode *vp, 66248508Skib struct buf *last_bp, int gbflags); 67248508Skibstatic struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize, 68248508Skib daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, 69248508Skib struct buf *fbp); 70141628Sphkstatic void cluster_callback(struct buf *); 711541Srgrimes 7248545Smckusickstatic int write_behind = 1; 7391690SeivindSYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, 7491690Seivind "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); 7548545Smckusick 76219699Sivorasstatic int read_max = 64; 77112080SjeffSYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, 78112080Sjeff "Cluster read-ahead max block count"); 79112080Sjeff 80250327Sscottlstatic int read_min = 1; 81250327SscottlSYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0, 82250327Sscottl "Cluster read min block count"); 83250327Sscottl 8491690Seivind/* Page expended to mark partially backed buffers */ 8512973Sbdeextern vm_page_t bogus_page; 865455Sdg 8791690Seivind/* 8891690Seivind * Read data to a buf, including read-ahead if we find this to be beneficial. 8991690Seivind * cluster_read replaces bread. 9010541Sdyson */ 911549Srgrimesint 92248282Skibcluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size, 93248282Skib struct ucred *cred, long totread, int seqcount, int gbflags, 94248282Skib struct buf **bpp) 951541Srgrimes{ 9621002Sdyson struct buf *bp, *rbp, *reqbp; 97177493Sjeff struct bufobj *bo; 9896572Sphk daddr_t blkno, origblkno; 99112080Sjeff int maxra, racluster; 100112080Sjeff int error, ncontig; 10110541Sdyson int i; 1021541Srgrimes 1031541Srgrimes error = 0; 104177493Sjeff bo = &vp->v_bufobj; 105248508Skib if (!unmapped_buf_allowed) 106248508Skib gbflags &= ~GB_UNMAPPED; 10721002Sdyson 1085455Sdg /* 10921002Sdyson * Try to limit the amount of read-ahead by a few 11021002Sdyson * ad-hoc parameters. This needs work!!! 11121002Sdyson */ 11251797Sphk racluster = vp->v_mount->mnt_iosize_max / size; 113112080Sjeff maxra = seqcount; 114112080Sjeff maxra = min(read_max, maxra); 115112080Sjeff maxra = min(nbuf/8, maxra); 116112080Sjeff if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) 117112080Sjeff maxra = (filesize / size) - lblkno; 11821002Sdyson 11921002Sdyson /* 1205455Sdg * get the requested block 1215455Sdg */ 122248508Skib *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags); 123294954Smckusick if (bp == NULL) 124294954Smckusick return (EBUSY); 12521002Sdyson origblkno = lblkno; 12612767Sdyson 1275455Sdg /* 1285455Sdg * if it is in the cache, then check to see if the reads have been 1295455Sdg * sequential. If they have, then try some read-ahead, otherwise 1305455Sdg * back-off on prospective read-aheads. 1315455Sdg */ 1321541Srgrimes if (bp->b_flags & B_CACHE) { 13321002Sdyson if (!seqcount) { 1345455Sdg return 0; 13521002Sdyson } else if ((bp->b_flags & B_RAM) == 0) { 13621002Sdyson return 0; 13721002Sdyson } else { 13821002Sdyson bp->b_flags &= ~B_RAM; 139251171Sjeff BO_RLOCK(bo); 14048225Smckusick for (i = 1; i < maxra; i++) { 14199737Sdillon /* 14299737Sdillon * Stop if the buffer does not exist or it 14399737Sdillon * is invalid (about to go away?) 14499737Sdillon */ 145136767Sphk rbp = gbincore(&vp->v_bufobj, lblkno+i); 146112080Sjeff if (rbp == NULL || (rbp->b_flags & B_INVAL)) 14721002Sdyson break; 14821002Sdyson 14921002Sdyson /* 15048677Smckusick * Set another read-ahead mark so we know 151151621Sups * to check again. (If we can lock the 152151621Sups * buffer without waiting) 15321002Sdyson */ 154151621Sups if ((((i % racluster) == (racluster - 1)) || 155151621Sups (i == (maxra - 1))) 156151621Sups && (0 == BUF_LOCK(rbp, 157151621Sups LK_EXCLUSIVE | LK_NOWAIT, NULL))) { 158112080Sjeff rbp->b_flags |= B_RAM; 159151621Sups BUF_UNLOCK(rbp); 160151621Sups } 16121002Sdyson } 162251171Sjeff BO_RUNLOCK(bo); 16321002Sdyson if (i >= maxra) { 1645839Sdg return 0; 16510541Sdyson } 16621002Sdyson lblkno += i; 16721002Sdyson } 16821002Sdyson reqbp = bp = NULL; 169111886Sjeff /* 170111886Sjeff * If it isn't in the cache, then get a chunk from 171111886Sjeff * disk if sequential, otherwise just get the block. 172111886Sjeff */ 17321002Sdyson } else { 17442453Seivind off_t firstread = bp->b_offset; 175111886Sjeff int nblks; 176250327Sscottl long minread; 17742453Seivind 17842408Seivind KASSERT(bp->b_offset != NOOFFSET, 17942453Seivind ("cluster_read: no buffer offset")); 180111886Sjeff 181112080Sjeff ncontig = 0; 182111886Sjeff 183111886Sjeff /* 184250327Sscottl * Adjust totread if needed 185250327Sscottl */ 186250327Sscottl minread = read_min * size; 187250327Sscottl if (minread > totread) 188250327Sscottl totread = minread; 189250327Sscottl 190250327Sscottl /* 191111886Sjeff * Compute the total number of blocks that we should read 192111886Sjeff * synchronously. 193111886Sjeff */ 19421002Sdyson if (firstread + totread > filesize) 19521002Sdyson totread = filesize - firstread; 196111886Sjeff nblks = howmany(totread, size); 197111886Sjeff if (nblks > racluster) 198111886Sjeff nblks = racluster; 19921002Sdyson 200111886Sjeff /* 201111886Sjeff * Now compute the number of contiguous blocks. 202111886Sjeff */ 203111886Sjeff if (nblks > 1) { 20421002Sdyson error = VOP_BMAP(vp, lblkno, NULL, 205112080Sjeff &blkno, &ncontig, NULL); 206111886Sjeff /* 207111886Sjeff * If this failed to map just do the original block. 208111886Sjeff */ 209111886Sjeff if (error || blkno == -1) 210112080Sjeff ncontig = 0; 211111886Sjeff } 21221002Sdyson 213111886Sjeff /* 214111886Sjeff * If we have contiguous data available do a cluster 215111886Sjeff * otherwise just read the requested block. 216111886Sjeff */ 217112080Sjeff if (ncontig) { 218111886Sjeff /* Account for our first block. */ 219112080Sjeff ncontig = min(ncontig + 1, nblks); 220112080Sjeff if (ncontig < nblks) 221112080Sjeff nblks = ncontig; 22221002Sdyson bp = cluster_rbuild(vp, filesize, lblkno, 223248508Skib blkno, size, nblks, gbflags, bp); 22434694Sdyson lblkno += (bp->b_bufsize / size); 22510541Sdyson } else { 22658345Sphk bp->b_flags |= B_RAM; 22758345Sphk bp->b_iocmd = BIO_READ; 22810541Sdyson lblkno += 1; 2298876Srgrimes } 2301541Srgrimes } 2315455Sdg 2325455Sdg /* 233112080Sjeff * handle the synchronous read so that it is available ASAP. 2345455Sdg */ 2355455Sdg if (bp) { 23670374Sdillon if ((bp->b_flags & B_CLUSTER) == 0) { 23736275Sdyson vfs_busy_pages(bp, 0); 23870374Sdillon } 23958934Sphk bp->b_flags &= ~B_INVAL; 24058934Sphk bp->b_ioflags &= ~BIO_ERROR; 24158345Sphk if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) 24248333Speter BUF_KERNPROC(bp); 243121205Sphk bp->b_iooffset = dbtob(bp->b_blkno); 244136927Sphk bstrategy(bp); 245297633Strasz#ifdef RACCT 246297633Strasz if (racct_enable) { 247297633Strasz PROC_LOCK(curproc); 248297633Strasz racct_add_buf(curproc, bp, 0); 249297633Strasz PROC_UNLOCK(curproc); 250297633Strasz } 251297633Strasz#endif /* RACCT */ 252170174Sjeff curthread->td_ru.ru_inblock++; 2535455Sdg } 25434611Sdyson 2555455Sdg /* 256112080Sjeff * If we have been doing sequential I/O, then do some read-ahead. 2575455Sdg */ 258112080Sjeff while (lblkno < (origblkno + maxra)) { 259112080Sjeff error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); 260112080Sjeff if (error) 261112080Sjeff break; 262112080Sjeff 263112080Sjeff if (blkno == -1) 264112080Sjeff break; 265112080Sjeff 266112080Sjeff /* 267112080Sjeff * We could throttle ncontig here by maxra but we might as 268112080Sjeff * well read the data if it is contiguous. We're throttled 269112080Sjeff * by racluster anyway. 270112080Sjeff */ 271112080Sjeff if (ncontig) { 272112080Sjeff ncontig = min(ncontig + 1, racluster); 273112080Sjeff rbp = cluster_rbuild(vp, filesize, lblkno, blkno, 274248508Skib size, ncontig, gbflags, NULL); 275112080Sjeff lblkno += (rbp->b_bufsize / size); 276112838Sjeff if (rbp->b_flags & B_DELWRI) { 277112838Sjeff bqrelse(rbp); 278112838Sjeff continue; 279112838Sjeff } 280112080Sjeff } else { 281248508Skib rbp = getblk(vp, lblkno, size, 0, 0, gbflags); 282112838Sjeff lblkno += 1; 283112838Sjeff if (rbp->b_flags & B_DELWRI) { 284112838Sjeff bqrelse(rbp); 285112838Sjeff continue; 286112838Sjeff } 287112080Sjeff rbp->b_flags |= B_ASYNC | B_RAM; 288112080Sjeff rbp->b_iocmd = BIO_READ; 289112080Sjeff rbp->b_blkno = blkno; 290112080Sjeff } 291112080Sjeff if (rbp->b_flags & B_CACHE) { 29258345Sphk rbp->b_flags &= ~B_ASYNC; 29313490Sdyson bqrelse(rbp); 294112080Sjeff continue; 2955455Sdg } 296112080Sjeff if ((rbp->b_flags & B_CLUSTER) == 0) { 297112080Sjeff vfs_busy_pages(rbp, 0); 298112080Sjeff } 299112080Sjeff rbp->b_flags &= ~B_INVAL; 300112080Sjeff rbp->b_ioflags &= ~BIO_ERROR; 301112080Sjeff if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) 302112080Sjeff BUF_KERNPROC(rbp); 303121205Sphk rbp->b_iooffset = dbtob(rbp->b_blkno); 304136927Sphk bstrategy(rbp); 305297633Strasz#ifdef RACCT 306297633Strasz if (racct_enable) { 307297633Strasz PROC_LOCK(curproc); 308297633Strasz racct_add_buf(curproc, rbp, 0); 309297633Strasz PROC_UNLOCK(curproc); 310297633Strasz } 311297633Strasz#endif /* RACCT */ 312170174Sjeff curthread->td_ru.ru_inblock++; 3135455Sdg } 314112080Sjeff 315294954Smckusick if (reqbp) { 316294954Smckusick /* 317294954Smckusick * Like bread, always brelse() the buffer when 318294954Smckusick * returning an error. 319294954Smckusick */ 320294954Smckusick error = bufwait(reqbp); 321294954Smckusick if (error != 0) { 322294954Smckusick brelse(reqbp); 323294954Smckusick *bpp = NULL; 324294954Smckusick } 325294954Smckusick } 326294954Smckusick return (error); 3271541Srgrimes} 3281541Srgrimes 3291541Srgrimes/* 3301541Srgrimes * If blocks are contiguous on disk, use this to provide clustered 3311541Srgrimes * read ahead. We will read as many blocks as possible sequentially 3321541Srgrimes * and then parcel them up into logical blocks in the buffer hash table. 3331541Srgrimes */ 33410541Sdysonstatic struct buf * 335248508Skibcluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, 336248508Skib daddr_t blkno, long size, int run, int gbflags, struct buf *fbp) 3371541Srgrimes{ 33810541Sdyson struct buf *bp, *tbp; 3391541Srgrimes daddr_t bn; 340195122Salc off_t off; 341195122Salc long tinc, tsize; 342254668Skib int i, inc, j, k, toff; 3431541Srgrimes 34442408Seivind KASSERT(size == vp->v_mount->mnt_stat.f_iosize, 345239315Salc ("cluster_rbuild: size %ld != f_iosize %jd\n", 346122537Smckusick size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); 34742453Seivind 34812767Sdyson /* 34912767Sdyson * avoid a division 35012767Sdyson */ 35112767Sdyson while ((u_quad_t) size * (lbn + run) > filesize) { 3521541Srgrimes --run; 35312767Sdyson } 35410541Sdyson 35521002Sdyson if (fbp) { 35621002Sdyson tbp = fbp; 35758345Sphk tbp->b_iocmd = BIO_READ; 35821002Sdyson } else { 359248508Skib tbp = getblk(vp, lbn, size, 0, 0, gbflags); 36021002Sdyson if (tbp->b_flags & B_CACHE) 36121002Sdyson return tbp; 36258345Sphk tbp->b_flags |= B_ASYNC | B_RAM; 36358345Sphk tbp->b_iocmd = BIO_READ; 36421002Sdyson } 36510541Sdyson tbp->b_blkno = blkno; 36616086Sdyson if( (tbp->b_flags & B_MALLOC) || 36716086Sdyson ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) 36810541Sdyson return tbp; 36910541Sdyson 37042957Sdillon bp = trypbuf(&cluster_pbuf_freecnt); 371298069Spfg if (bp == NULL) 37210541Sdyson return tbp; 37310541Sdyson 37485272Sdillon /* 37585272Sdillon * We are synthesizing a buffer out of vm_page_t's, but 37685272Sdillon * if the block size is not page aligned then the starting 37785272Sdillon * address may not be either. Inherit the b_data offset 37885272Sdillon * from the original buffer. 37985272Sdillon */ 38058345Sphk bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; 381248508Skib if ((gbflags & GB_UNMAPPED) != 0) { 382248508Skib bp->b_data = unmapped_buf; 383248508Skib } else { 384248508Skib bp->b_data = (char *)((vm_offset_t)bp->b_data | 385248508Skib ((vm_offset_t)tbp->b_data & PAGE_MASK)); 386248508Skib } 38758345Sphk bp->b_iocmd = BIO_READ; 3885455Sdg bp->b_iodone = cluster_callback; 3895455Sdg bp->b_blkno = blkno; 3905455Sdg bp->b_lblkno = lbn; 39134611Sdyson bp->b_offset = tbp->b_offset; 39242453Seivind KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); 3935455Sdg pbgetvp(vp, bp); 3941541Srgrimes 39512404Sdyson TAILQ_INIT(&bp->b_cluster.cluster_head); 3961541Srgrimes 3975455Sdg bp->b_bcount = 0; 3985455Sdg bp->b_bufsize = 0; 3995455Sdg bp->b_npages = 0; 4005455Sdg 4011541Srgrimes inc = btodb(size); 40210541Sdyson for (bn = blkno, i = 0; i < run; ++i, bn += inc) { 403254668Skib if (i == 0) { 404254668Skib VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 405254668Skib vfs_drain_busy_pages(tbp); 406254668Skib vm_object_pip_add(tbp->b_bufobj->bo_object, 407254668Skib tbp->b_npages); 408254668Skib for (k = 0; k < tbp->b_npages; k++) 409254668Skib vm_page_sbusy(tbp->b_pages[k]); 410254668Skib VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 411254668Skib } else { 41212767Sdyson if ((bp->b_npages * PAGE_SIZE) + 41385272Sdillon round_page(size) > vp->v_mount->mnt_iosize_max) { 41410541Sdyson break; 41585272Sdillon } 41610978Sdyson 417248508Skib tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT | 418248508Skib (gbflags & GB_UNMAPPED)); 41912767Sdyson 420111886Sjeff /* Don't wait around for locked bufs. */ 421111886Sjeff if (tbp == NULL) 422111886Sjeff break; 42334611Sdyson 42471230Sdillon /* 42585272Sdillon * Stop scanning if the buffer is fully valid 42685272Sdillon * (marked B_CACHE), or locked (may be doing a 42785272Sdillon * background write), or if the buffer is not 42885272Sdillon * VMIO backed. The clustering code can only deal 429251171Sjeff * with VMIO-backed buffers. The bo lock is not 430251171Sjeff * required for the BKGRDINPROG check since it 431251171Sjeff * can not be set without the buf lock. 43271230Sdillon */ 433119521Sjeff if ((tbp->b_vflags & BV_BKGRDINPROG) || 434119521Sjeff (tbp->b_flags & B_CACHE) || 435119521Sjeff (tbp->b_flags & B_VMIO) == 0) { 43613490Sdyson bqrelse(tbp); 4375455Sdg break; 4385455Sdg } 43910541Sdyson 44085272Sdillon /* 44185272Sdillon * The buffer must be completely invalid in order to 44285272Sdillon * take part in the cluster. If it is partially valid 44385272Sdillon * then we stop. 44485272Sdillon */ 445195122Salc off = tbp->b_offset; 446195122Salc tsize = size; 447248084Sattilio VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 448195122Salc for (j = 0; tsize > 0; j++) { 449195122Salc toff = off & PAGE_MASK; 450195122Salc tinc = tsize; 451195122Salc if (toff + tinc > PAGE_SIZE) 452195122Salc tinc = PAGE_SIZE - toff; 453248084Sattilio VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object); 454195122Salc if ((tbp->b_pages[j]->valid & 455195122Salc vm_page_bits(toff, tinc)) != 0) 45610541Sdyson break; 457254668Skib if (vm_page_xbusied(tbp->b_pages[j])) 458254668Skib break; 459254668Skib vm_object_pip_add(tbp->b_bufobj->bo_object, 1); 460254668Skib vm_page_sbusy(tbp->b_pages[j]); 461195122Salc off += tinc; 462195122Salc tsize -= tinc; 46371230Sdillon } 464195122Salc if (tsize > 0) { 465254668Skibclean_sbusy: 466254668Skib vm_object_pip_add(tbp->b_bufobj->bo_object, -j); 467254668Skib for (k = 0; k < j; k++) 468254668Skib vm_page_sunbusy(tbp->b_pages[k]); 469254668Skib VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 47034611Sdyson bqrelse(tbp); 47110541Sdyson break; 47210541Sdyson } 473254668Skib VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 47410541Sdyson 47585272Sdillon /* 47685272Sdillon * Set a read-ahead mark as appropriate 47785272Sdillon */ 47821002Sdyson if ((fbp && (i == 1)) || (i == (run - 1))) 47921002Sdyson tbp->b_flags |= B_RAM; 48085272Sdillon 48185272Sdillon /* 48285272Sdillon * Set the buffer up for an async read (XXX should 48385272Sdillon * we do this only if we do not wind up brelse()ing?). 48485272Sdillon * Set the block number if it isn't set, otherwise 48585272Sdillon * if it is make sure it matches the block number we 48685272Sdillon * expect. 48785272Sdillon */ 48858345Sphk tbp->b_flags |= B_ASYNC; 48958345Sphk tbp->b_iocmd = BIO_READ; 49012767Sdyson if (tbp->b_blkno == tbp->b_lblkno) { 49110541Sdyson tbp->b_blkno = bn; 49210541Sdyson } else if (tbp->b_blkno != bn) { 493254668Skib VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 494254668Skib goto clean_sbusy; 49510541Sdyson } 4961541Srgrimes } 49748333Speter /* 49848333Speter * XXX fbp from caller may not be B_ASYNC, but we are going 49948333Speter * to biodone() it in cluster_callback() anyway 50048333Speter */ 50148333Speter BUF_KERNPROC(tbp); 50212404Sdyson TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 50312404Sdyson tbp, b_cluster.cluster_entry); 504248084Sattilio VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 5055455Sdg for (j = 0; j < tbp->b_npages; j += 1) { 50610541Sdyson vm_page_t m; 50710541Sdyson m = tbp->b_pages[j]; 50810541Sdyson if ((bp->b_npages == 0) || 509254668Skib (bp->b_pages[bp->b_npages-1] != m)) { 51010541Sdyson bp->b_pages[bp->b_npages] = m; 51110541Sdyson bp->b_npages++; 51210541Sdyson } 513193643Salc if (m->valid == VM_PAGE_BITS_ALL) 51418737Sdyson tbp->b_pages[j] = bogus_page; 5151541Srgrimes } 516248084Sattilio VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 51785511Sdillon /* 51885511Sdillon * Don't inherit tbp->b_bufsize as it may be larger due to 51985511Sdillon * a non-page-aligned size. Instead just aggregate using 52085511Sdillon * 'size'. 52185511Sdillon */ 52285511Sdillon if (tbp->b_bcount != size) 52385511Sdillon printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); 52485511Sdillon if (tbp->b_bufsize != size) 52585511Sdillon printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); 52685511Sdillon bp->b_bcount += size; 52785511Sdillon bp->b_bufsize += size; 5281541Srgrimes } 52918737Sdyson 53085272Sdillon /* 53185272Sdillon * Fully valid pages in the cluster are already good and do not need 53285272Sdillon * to be re-read from disk. Replace the page with bogus_page 53385272Sdillon */ 534248084Sattilio VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 53585272Sdillon for (j = 0; j < bp->b_npages; j++) { 536248084Sattilio VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object); 537193643Salc if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL) 53818737Sdyson bp->b_pages[j] = bogus_page; 53918737Sdyson } 540248084Sattilio VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 54120054Sdyson if (bp->b_bufsize > bp->b_kvasize) 54237559Sbde panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 54337559Sbde bp->b_bufsize, bp->b_kvasize); 54418737Sdyson 545285819Sjeff if (buf_mapped(bp)) { 546248508Skib pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 547248508Skib (vm_page_t *)bp->b_pages, bp->b_npages); 548248508Skib } 5495455Sdg return (bp); 5501541Srgrimes} 5511541Srgrimes 5521541Srgrimes/* 5531541Srgrimes * Cleanup after a clustered read or write. 5541541Srgrimes * This is complicated by the fact that any of the buffers might have 5551541Srgrimes * extra memory (if there were no empty buffer headers at allocbuf time) 5561541Srgrimes * that we will need to shift around. 5571541Srgrimes */ 558141628Sphkstatic void 5591541Srgrimescluster_callback(bp) 5601541Srgrimes struct buf *bp; 5611541Srgrimes{ 56212404Sdyson struct buf *nbp, *tbp; 5631541Srgrimes int error = 0; 5641541Srgrimes 5651541Srgrimes /* 566298819Spfg * Must propagate errors to all the components. 5671541Srgrimes */ 56858934Sphk if (bp->b_ioflags & BIO_ERROR) 5691541Srgrimes error = bp->b_error; 5701541Srgrimes 571285819Sjeff if (buf_mapped(bp)) { 572248508Skib pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 573248508Skib bp->b_npages); 574248508Skib } 5751541Srgrimes /* 5761541Srgrimes * Move memory from the large cluster buffer into the component 5771541Srgrimes * buffers and mark IO as done on these. 5781541Srgrimes */ 57921002Sdyson for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 58012404Sdyson tbp; tbp = nbp) { 58121002Sdyson nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 5821541Srgrimes if (error) { 58358934Sphk tbp->b_ioflags |= BIO_ERROR; 5841541Srgrimes tbp->b_error = error; 58546349Salc } else { 58646349Salc tbp->b_dirtyoff = tbp->b_dirtyend = 0; 58758934Sphk tbp->b_flags &= ~B_INVAL; 58858934Sphk tbp->b_ioflags &= ~BIO_ERROR; 58977115Sdillon /* 59077115Sdillon * XXX the bdwrite()/bqrelse() issued during 59177115Sdillon * cluster building clears B_RELBUF (see bqrelse() 59277115Sdillon * comment). If direct I/O was specified, we have 59377115Sdillon * to restore it here to allow the buffer and VM 59477115Sdillon * to be freed. 59577115Sdillon */ 59677115Sdillon if (tbp->b_flags & B_DIRECT) 59777115Sdillon tbp->b_flags |= B_RELBUF; 59846349Salc } 59959249Sphk bufdone(tbp); 6001541Srgrimes } 601137719Sphk pbrelvp(bp); 60242957Sdillon relpbuf(bp, &cluster_pbuf_freecnt); 6031541Srgrimes} 6041541Srgrimes 6051541Srgrimes/* 60648545Smckusick * cluster_wbuild_wb: 60748545Smckusick * 60848545Smckusick * Implement modified write build for cluster. 60948545Smckusick * 61048545Smckusick * write_behind = 0 write behind disabled 61148545Smckusick * write_behind = 1 write behind normal (default) 61248545Smckusick * write_behind = 2 write behind backed-off 61348545Smckusick */ 61448545Smckusick 61548545Smckusickstatic __inline int 616248508Skibcluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len, 617248508Skib int gbflags) 61848545Smckusick{ 61948545Smckusick int r = 0; 62048545Smckusick 621248283Skib switch (write_behind) { 62248545Smckusick case 2: 62348545Smckusick if (start_lbn < len) 62448545Smckusick break; 62548545Smckusick start_lbn -= len; 626102412Scharnier /* FALLTHROUGH */ 62748545Smckusick case 1: 628248508Skib r = cluster_wbuild(vp, size, start_lbn, len, gbflags); 629102412Scharnier /* FALLTHROUGH */ 63048545Smckusick default: 631102412Scharnier /* FALLTHROUGH */ 63248545Smckusick break; 63348545Smckusick } 63448545Smckusick return(r); 63548545Smckusick} 63648545Smckusick 63748545Smckusick/* 6381541Srgrimes * Do clustered write for FFS. 6391541Srgrimes * 6401541Srgrimes * Three cases: 6411541Srgrimes * 1. Write is not sequential (write asynchronously) 6421541Srgrimes * Write is sequential: 6431541Srgrimes * 2. beginning of cluster - begin cluster 6441541Srgrimes * 3. middle of a cluster - add to cluster 6451541Srgrimes * 4. end of a cluster - asynchronously write cluster 6461541Srgrimes */ 6471541Srgrimesvoid 648248282Skibcluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount, 649248282Skib int gbflags) 6501541Srgrimes{ 6515455Sdg daddr_t lbn; 6525455Sdg int maxclen, cursize; 6535455Sdg int lblocksize; 65412404Sdyson int async; 6551541Srgrimes 656248508Skib if (!unmapped_buf_allowed) 657248508Skib gbflags &= ~GB_UNMAPPED; 658248508Skib 65932286Sdyson if (vp->v_type == VREG) { 660231204Skib async = DOINGASYNC(vp); 66132286Sdyson lblocksize = vp->v_mount->mnt_stat.f_iosize; 66232286Sdyson } else { 66332286Sdyson async = 0; 66432286Sdyson lblocksize = bp->b_bufsize; 66532286Sdyson } 6665455Sdg lbn = bp->b_lblkno; 66742408Seivind KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 66834694Sdyson 6691541Srgrimes /* Initialize vnode to beginning of file. */ 6701541Srgrimes if (lbn == 0) 6711541Srgrimes vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 6721541Srgrimes 6735455Sdg if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 6745455Sdg (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 67551797Sphk maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; 6761541Srgrimes if (vp->v_clen != 0) { 6771541Srgrimes /* 6781541Srgrimes * Next block is not sequential. 6798876Srgrimes * 6801541Srgrimes * If we are not writing at end of file, the process 6815455Sdg * seeked to another point in the file since its last 6825455Sdg * write, or we have reached our maximum cluster size, 6835455Sdg * then push the previous cluster. Otherwise try 6845455Sdg * reallocating to make it sequential. 68558909Sdillon * 68658909Sdillon * Change to algorithm: only push previous cluster if 68758909Sdillon * it was sequential from the point of view of the 68858909Sdillon * seqcount heuristic, otherwise leave the buffer 68958909Sdillon * intact so we can potentially optimize the I/O 69058909Sdillon * later on in the buf_daemon or update daemon 69158909Sdillon * flush. 6921541Srgrimes */ 6931541Srgrimes cursize = vp->v_lastw - vp->v_cstart + 1; 69434611Sdyson if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 69510541Sdyson lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 69658909Sdillon if (!async && seqcount > 0) { 69748677Smckusick cluster_wbuild_wb(vp, lblocksize, 698248508Skib vp->v_cstart, cursize, gbflags); 69958909Sdillon } 70010541Sdyson } else { 70110541Sdyson struct buf **bpp, **endbp; 70210541Sdyson struct cluster_save *buflist; 70310541Sdyson 704248508Skib buflist = cluster_collectbufs(vp, bp, gbflags); 705352947Smckusick if (buflist == NULL) { 706352947Smckusick /* 707352947Smckusick * Cluster build failed so just write 708352947Smckusick * it now. 709352947Smckusick */ 710352947Smckusick bawrite(bp); 711352947Smckusick return; 712352947Smckusick } 71310541Sdyson endbp = &buflist->bs_children 71410541Sdyson [buflist->bs_nchildren - 1]; 71510541Sdyson if (VOP_REALLOCBLKS(vp, buflist)) { 71610541Sdyson /* 71758909Sdillon * Failed, push the previous cluster 71858909Sdillon * if *really* writing sequentially 71958909Sdillon * in the logical file (seqcount > 1), 72058909Sdillon * otherwise delay it in the hopes that 72158909Sdillon * the low level disk driver can 72258909Sdillon * optimize the write ordering. 72310541Sdyson */ 72410541Sdyson for (bpp = buflist->bs_children; 72510541Sdyson bpp < endbp; bpp++) 72610541Sdyson brelse(*bpp); 72710541Sdyson free(buflist, M_SEGMENT); 72858909Sdillon if (seqcount > 1) { 72958909Sdillon cluster_wbuild_wb(vp, 73058909Sdillon lblocksize, vp->v_cstart, 731248508Skib cursize, gbflags); 73258909Sdillon } 73310541Sdyson } else { 73410541Sdyson /* 73510541Sdyson * Succeeded, keep building cluster. 73610541Sdyson */ 73710541Sdyson for (bpp = buflist->bs_children; 73810541Sdyson bpp <= endbp; bpp++) 73910541Sdyson bdwrite(*bpp); 74010541Sdyson free(buflist, M_SEGMENT); 74110541Sdyson vp->v_lastw = lbn; 74210541Sdyson vp->v_lasta = bp->b_blkno; 74310541Sdyson return; 74410541Sdyson } 74510541Sdyson } 7461541Srgrimes } 7471541Srgrimes /* 7485455Sdg * Consider beginning a cluster. If at end of file, make 7495455Sdg * cluster as large as possible, otherwise find size of 7505455Sdg * existing cluster. 7511541Srgrimes */ 75232286Sdyson if ((vp->v_type == VREG) && 75334611Sdyson ((u_quad_t) bp->b_offset + lblocksize) != filesize && 7547613Sdg (bp->b_blkno == bp->b_lblkno) && 75510551Sdyson (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 75610541Sdyson bp->b_blkno == -1)) { 7571541Srgrimes bawrite(bp); 7581541Srgrimes vp->v_clen = 0; 7591541Srgrimes vp->v_lasta = bp->b_blkno; 7601541Srgrimes vp->v_cstart = lbn + 1; 7611541Srgrimes vp->v_lastw = lbn; 7621541Srgrimes return; 7631541Srgrimes } 7645455Sdg vp->v_clen = maxclen; 76512404Sdyson if (!async && maxclen == 0) { /* I/O not contiguous */ 7661541Srgrimes vp->v_cstart = lbn + 1; 76713490Sdyson bawrite(bp); 7685455Sdg } else { /* Wait for rest of cluster */ 7691541Srgrimes vp->v_cstart = lbn; 7705455Sdg bdwrite(bp); 7711541Srgrimes } 7721541Srgrimes } else if (lbn == vp->v_cstart + vp->v_clen) { 7731541Srgrimes /* 77458909Sdillon * At end of cluster, write it out if seqcount tells us we 77558909Sdillon * are operating sequentially, otherwise let the buf or 77658909Sdillon * update daemon handle it. 7771541Srgrimes */ 77812404Sdyson bdwrite(bp); 779248508Skib if (seqcount > 1) { 780248508Skib cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, 781248508Skib vp->v_clen + 1, gbflags); 782248508Skib } 7831541Srgrimes vp->v_clen = 0; 7841541Srgrimes vp->v_cstart = lbn + 1; 78568885Sdillon } else if (vm_page_count_severe()) { 78668885Sdillon /* 78768885Sdillon * We are low on memory, get it going NOW 78868885Sdillon */ 78968885Sdillon bawrite(bp); 79058909Sdillon } else { 7911541Srgrimes /* 7925455Sdg * In the middle of a cluster, so just delay the I/O for now. 7931541Srgrimes */ 7941541Srgrimes bdwrite(bp); 79558909Sdillon } 7961541Srgrimes vp->v_lastw = lbn; 7971541Srgrimes vp->v_lasta = bp->b_blkno; 7981541Srgrimes} 7991541Srgrimes 8001541Srgrimes 8011541Srgrimes/* 8021541Srgrimes * This is an awful lot like cluster_rbuild...wish they could be combined. 8031541Srgrimes * The last lbn argument is the current block on which I/O is being 8041541Srgrimes * performed. Check to see that it doesn't fall in the middle of 8051541Srgrimes * the current block (if last_bp == NULL). 8061541Srgrimes */ 80712767Sdysonint 808248282Skibcluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len, 809248282Skib int gbflags) 8101541Srgrimes{ 81112404Sdyson struct buf *bp, *tbp; 812177493Sjeff struct bufobj *bo; 813145734Sjeff int i, j; 81412767Sdyson int totalwritten = 0; 81512404Sdyson int dbsize = btodb(size); 81635595Sbde 817248508Skib if (!unmapped_buf_allowed) 818248508Skib gbflags &= ~GB_UNMAPPED; 819248508Skib 820177493Sjeff bo = &vp->v_bufobj; 82112767Sdyson while (len > 0) { 82271230Sdillon /* 82371230Sdillon * If the buffer is not delayed-write (i.e. dirty), or it 82471230Sdillon * is delayed-write but either locked or inval, it cannot 82572080Sasmodai * partake in the clustered write. 82671230Sdillon */ 827177493Sjeff BO_LOCK(bo); 828136767Sphk if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL || 829119521Sjeff (tbp->b_vflags & BV_BKGRDINPROG)) { 830177493Sjeff BO_UNLOCK(bo); 83112767Sdyson ++start_lbn; 83212767Sdyson --len; 83312767Sdyson continue; 83412767Sdyson } 835111886Sjeff if (BUF_LOCK(tbp, 836251171Sjeff LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) { 837111886Sjeff ++start_lbn; 838111886Sjeff --len; 839111886Sjeff continue; 840111886Sjeff } 841119521Sjeff if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { 842111886Sjeff BUF_UNLOCK(tbp); 843111886Sjeff ++start_lbn; 844111886Sjeff --len; 845111886Sjeff continue; 846111886Sjeff } 847153192Srodrigc if (tbp->b_pin_count > 0) { 848153192Srodrigc BUF_UNLOCK(tbp); 849153192Srodrigc ++start_lbn; 850153192Srodrigc --len; 851153192Srodrigc continue; 852153192Srodrigc } 85312767Sdyson bremfree(tbp); 85412767Sdyson tbp->b_flags &= ~B_DONE; 8551541Srgrimes 85647967Sjulian /* 85747967Sjulian * Extra memory in the buffer, punt on this buffer. 85847967Sjulian * XXX we could handle this in most cases, but we would 85947967Sjulian * have to push the extra memory down to after our max 86047967Sjulian * possible cluster size and then potentially pull it back 86147967Sjulian * up if the cluster was terminated prematurely--too much 86247967Sjulian * hassle. 86347967Sjulian */ 86468868Stegge if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != 86568868Stegge (B_CLUSTEROK | B_VMIO)) || 86634630Sjulian (tbp->b_bcount != tbp->b_bufsize) || 86734630Sjulian (tbp->b_bcount != size) || 86834630Sjulian (len == 1) || 869254945Skib ((bp = (vp->v_vflag & VV_MD) != 0 ? 870254945Skib trypbuf(&cluster_pbuf_freecnt) : 871254945Skib getpbuf(&cluster_pbuf_freecnt)) == NULL)) { 87212767Sdyson totalwritten += tbp->b_bufsize; 87312767Sdyson bawrite(tbp); 87412767Sdyson ++start_lbn; 87512767Sdyson --len; 87612767Sdyson continue; 87712767Sdyson } 87812404Sdyson 87934630Sjulian /* 88034630Sjulian * We got a pbuf to make the cluster in. 88134630Sjulian * so initialise it. 88234630Sjulian */ 88312767Sdyson TAILQ_INIT(&bp->b_cluster.cluster_head); 88412767Sdyson bp->b_bcount = 0; 88512767Sdyson bp->b_bufsize = 0; 88612767Sdyson bp->b_npages = 0; 88784827Sjhb if (tbp->b_wcred != NOCRED) 88884827Sjhb bp->b_wcred = crhold(tbp->b_wcred); 8891541Srgrimes 89012767Sdyson bp->b_blkno = tbp->b_blkno; 89112767Sdyson bp->b_lblkno = tbp->b_lblkno; 89234611Sdyson bp->b_offset = tbp->b_offset; 89385272Sdillon 89485272Sdillon /* 89585272Sdillon * We are synthesizing a buffer out of vm_page_t's, but 89685272Sdillon * if the block size is not page aligned then the starting 89785272Sdillon * address may not be either. Inherit the b_data offset 89885272Sdillon * from the original buffer. 89985272Sdillon */ 900248508Skib if ((gbflags & GB_UNMAPPED) == 0 || 901248508Skib (tbp->b_flags & B_VMIO) == 0) { 902248508Skib bp->b_data = (char *)((vm_offset_t)bp->b_data | 903248508Skib ((vm_offset_t)tbp->b_data & PAGE_MASK)); 904248508Skib } else { 905248508Skib bp->b_data = unmapped_buf; 906248508Skib } 907248508Skib bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO | 908248508Skib B_NEEDCOMMIT)); 90912767Sdyson bp->b_iodone = cluster_callback; 91012767Sdyson pbgetvp(vp, bp); 91134630Sjulian /* 91234630Sjulian * From this location in the file, scan forward to see 91334630Sjulian * if there are buffers with adjacent data that need to 91434630Sjulian * be written as well. 91534630Sjulian */ 91612767Sdyson for (i = 0; i < len; ++i, ++start_lbn) { 91734630Sjulian if (i != 0) { /* If not the first buffer */ 91834630Sjulian /* 91934630Sjulian * If the adjacent data is not even in core it 92034630Sjulian * can't need to be written. 92134630Sjulian */ 922177493Sjeff BO_LOCK(bo); 923177493Sjeff if ((tbp = gbincore(bo, start_lbn)) == NULL || 924119521Sjeff (tbp->b_vflags & BV_BKGRDINPROG)) { 925177493Sjeff BO_UNLOCK(bo); 92612767Sdyson break; 92712767Sdyson } 9281541Srgrimes 92934630Sjulian /* 93034630Sjulian * If it IS in core, but has different 93171230Sdillon * characteristics, or is locked (which 93271230Sdillon * means it could be undergoing a background 93371230Sdillon * I/O or be in a weird state), then don't 93471230Sdillon * cluster with it. 93534630Sjulian */ 936111886Sjeff if (BUF_LOCK(tbp, 937111886Sjeff LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, 938251171Sjeff BO_LOCKPTR(bo))) 939111886Sjeff break; 940111886Sjeff 94148225Smckusick if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 94248225Smckusick B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 943111886Sjeff != (B_DELWRI | B_CLUSTEROK | 94448225Smckusick (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 945111886Sjeff tbp->b_wcred != bp->b_wcred) { 946112347Sjeff BUF_UNLOCK(tbp); 94712767Sdyson break; 94812767Sdyson } 94912767Sdyson 95034630Sjulian /* 95134630Sjulian * Check that the combined cluster 95234630Sjulian * would make sense with regard to pages 95334630Sjulian * and would not be too large 95434630Sjulian */ 95512767Sdyson if ((tbp->b_bcount != size) || 95634630Sjulian ((bp->b_blkno + (dbsize * i)) != 95734694Sdyson tbp->b_blkno) || 95834630Sjulian ((tbp->b_npages + bp->b_npages) > 95951797Sphk (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { 96048225Smckusick BUF_UNLOCK(tbp); 96112767Sdyson break; 96212767Sdyson } 963153192Srodrigc 96434630Sjulian /* 965153192Srodrigc * Do not pull in pinned buffers. 966153192Srodrigc */ 967153192Srodrigc if (tbp->b_pin_count > 0) { 968153192Srodrigc BUF_UNLOCK(tbp); 969153192Srodrigc break; 970153192Srodrigc } 971153192Srodrigc 972153192Srodrigc /* 97334630Sjulian * Ok, it's passed all the tests, 97434630Sjulian * so remove it from the free list 97534630Sjulian * and mark it busy. We will use it. 97634630Sjulian */ 97712767Sdyson bremfree(tbp); 97812767Sdyson tbp->b_flags &= ~B_DONE; 97934630Sjulian } /* end of code for non-first buffers only */ 98034630Sjulian /* 98134630Sjulian * If the IO is via the VM then we do some 98285272Sdillon * special VM hackery (yuck). Since the buffer's 98385272Sdillon * block size may not be page-aligned it is possible 98485272Sdillon * for a page to be shared between two buffers. We 98585272Sdillon * have to get rid of the duplication when building 98685272Sdillon * the cluster. 98734630Sjulian */ 98813490Sdyson if (tbp->b_flags & B_VMIO) { 98932937Sdyson vm_page_t m; 99032937Sdyson 991248084Sattilio VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 992254668Skib if (i == 0) { 993254668Skib vfs_drain_busy_pages(tbp); 994254717Sjkim } else { /* if not first buffer */ 99532937Sdyson for (j = 0; j < tbp->b_npages; j += 1) { 99632937Sdyson m = tbp->b_pages[j]; 997254138Sattilio if (vm_page_xbusied(m)) { 998248084Sattilio VM_OBJECT_WUNLOCK( 999136985Salc tbp->b_object); 100050701Stegge bqrelse(tbp); 100132937Sdyson goto finishcluster; 100250701Stegge } 100332937Sdyson } 100432937Sdyson } 100513490Sdyson for (j = 0; j < tbp->b_npages; j += 1) { 100613490Sdyson m = tbp->b_pages[j]; 1007254138Sattilio vm_page_sbusy(m); 100838517Sdfr vm_object_pip_add(m->object, 1); 100913490Sdyson if ((bp->b_npages == 0) || 101034630Sjulian (bp->b_pages[bp->b_npages - 1] != m)) { 101113490Sdyson bp->b_pages[bp->b_npages] = m; 101213490Sdyson bp->b_npages++; 101313490Sdyson } 101412767Sdyson } 1015248084Sattilio VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 101612767Sdyson } 101712767Sdyson bp->b_bcount += size; 101812767Sdyson bp->b_bufsize += size; 1019246876Smckusick /* 1020246876Smckusick * If any of the clustered buffers have their 1021246876Smckusick * B_BARRIER flag set, transfer that request to 1022246876Smckusick * the cluster. 1023246876Smckusick */ 1024246876Smckusick bp->b_flags |= (tbp->b_flags & B_BARRIER); 1025246876Smckusick tbp->b_flags &= ~(B_DONE | B_BARRIER); 1026246876Smckusick tbp->b_flags |= B_ASYNC; 102758934Sphk tbp->b_ioflags &= ~BIO_ERROR; 102858345Sphk tbp->b_iocmd = BIO_WRITE; 1029246876Smckusick bundirty(tbp); 1030132640Sphk reassignbuf(tbp); /* put on clean list */ 1031136767Sphk bufobj_wref(tbp->b_bufobj); 103248333Speter BUF_KERNPROC(tbp); 103312767Sdyson TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 103412767Sdyson tbp, b_cluster.cluster_entry); 10351541Srgrimes } 103632937Sdyson finishcluster: 1037285819Sjeff if (buf_mapped(bp)) { 1038248508Skib pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 1039248508Skib (vm_page_t *)bp->b_pages, bp->b_npages); 1040248508Skib } 104120054Sdyson if (bp->b_bufsize > bp->b_kvasize) 104237559Sbde panic( 104337559Sbde "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 104437559Sbde bp->b_bufsize, bp->b_kvasize); 104512767Sdyson totalwritten += bp->b_bufsize; 104617304Sdyson bp->b_dirtyoff = 0; 104717304Sdyson bp->b_dirtyend = bp->b_bufsize; 104812767Sdyson bawrite(bp); 10491541Srgrimes 105012767Sdyson len -= i; 10511541Srgrimes } 105212767Sdyson return totalwritten; 10531541Srgrimes} 10541541Srgrimes 10551541Srgrimes/* 10561541Srgrimes * Collect together all the buffers in a cluster. 10571541Srgrimes * Plus add one additional buffer. 10581541Srgrimes */ 105912973Sbdestatic struct cluster_save * 1060248508Skibcluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags) 10611541Srgrimes{ 10621541Srgrimes struct cluster_save *buflist; 106341205Smckusick struct buf *bp; 10645455Sdg daddr_t lbn; 1065352947Smckusick int i, j, len, error; 10661541Srgrimes 10671541Srgrimes len = vp->v_lastw - vp->v_cstart + 1; 10681541Srgrimes buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1069111119Simp M_SEGMENT, M_WAITOK); 10701541Srgrimes buflist->bs_nchildren = 0; 10715455Sdg buflist->bs_children = (struct buf **) (buflist + 1); 107241205Smckusick for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 1073352947Smckusick error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED, 1074248508Skib gbflags, &bp); 1075352947Smckusick if (error != 0) { 1076352947Smckusick /* 1077352947Smckusick * If read fails, release collected buffers 1078352947Smckusick * and return failure. 1079352947Smckusick */ 1080352947Smckusick for (j = 0; j < i; j++) 1081352947Smckusick brelse(buflist->bs_children[j]); 1082352947Smckusick free(buflist, M_SEGMENT); 1083352947Smckusick return (NULL); 1084352947Smckusick } 108541205Smckusick buflist->bs_children[i] = bp; 108641205Smckusick if (bp->b_blkno == bp->b_lblkno) 1087136989Sphk VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, 108841205Smckusick NULL, NULL); 108941205Smckusick } 109041529Smckusick buflist->bs_children[i] = bp = last_bp; 109141529Smckusick if (bp->b_blkno == bp->b_lblkno) 1092136989Sphk VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 10931541Srgrimes buflist->bs_nchildren = i + 1; 10941541Srgrimes return (buflist); 10951541Srgrimes} 1096