vfs_bio.c revision 137193
1139749Simp/*- 275353Smjacob * Copyright (c) 2004 Poul-Henning Kamp 375353Smjacob * Copyright (c) 1994,1997 John S. Dyson 475353Smjacob * All rights reserved. 575353Smjacob * 675353Smjacob * Redistribution and use in source and binary forms, with or without 775353Smjacob * modification, are permitted provided that the following conditions 875353Smjacob * are met: 975353Smjacob * 1. Redistributions of source code must retain the above copyright 1075353Smjacob * notice, this list of conditions and the following disclaimer. 1175353Smjacob * 2. Redistributions in binary form must reproduce the above copyright 1275353Smjacob * notice, this list of conditions and the following disclaimer in the 1375353Smjacob * documentation and/or other materials provided with the distribution. 1475353Smjacob * 1575353Smjacob * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 1675353Smjacob * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1775353Smjacob * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1875353Smjacob * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 1975353Smjacob * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2075353Smjacob * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2175353Smjacob * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2275353Smjacob * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2375353Smjacob * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2475353Smjacob * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2575353Smjacob * SUCH DAMAGE. 2675353Smjacob */ 2775353Smjacob 28220938Smarius/* 2975353Smjacob * this file contains a new buffer I/O scheme implementing a coherent 3075353Smjacob * VM object and buffer cache scheme. Pains have been taken to make 3175353Smjacob * sure that the performance degradation associated with schemes such 32129843Smarius * as this is not realized. 33129843Smarius * 34129843Smarius * Author: John S. Dyson 3575353Smjacob * Significant help during the development and debugging phases 3675353Smjacob * had been provided by David Greenman, also of the FreeBSD core team. 3775353Smjacob * 3875353Smjacob * see man buf(9) for more info. 39120281Swilko */ 40120281Swilko 41120281Swilko#include <sys/cdefs.h> 42120281Swilko__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 137193 2004-11-04 09:06:54Z phk $"); 43120281Swilko 44120281Swilko#include <sys/param.h> 45120281Swilko#include <sys/systm.h> 4675353Smjacob#include <sys/bio.h> 4775353Smjacob#include <sys/conf.h> 4875353Smjacob#include <sys/buf.h> 49129876Sphk#include <sys/devicestat.h> 5075353Smjacob#include <sys/eventhandler.h> 5175353Smjacob#include <sys/lock.h> 5275353Smjacob#include <sys/malloc.h> 5375353Smjacob#include <sys/mount.h> 5475353Smjacob#include <sys/mutex.h> 5575353Smjacob#include <sys/kernel.h> 5675353Smjacob#include <sys/kthread.h> 5775353Smjacob#include <sys/proc.h> 5875353Smjacob#include <sys/resourcevar.h> 59109514Sobrien#include <sys/sysctl.h> 6075353Smjacob#include <sys/vmmeter.h> 6175353Smjacob#include <sys/vnode.h> 6275353Smjacob#include <geom/geom.h> 6375353Smjacob#include <vm/vm.h> 6475353Smjacob#include <vm/vm_param.h> 65165095Syongari#include <vm/vm_kern.h> 66165095Syongari#include <vm/vm_pageout.h> 6775353Smjacob#include <vm/vm_page.h> 6875353Smjacob#include <vm/vm_object.h> 6975353Smjacob#include <vm/vm_extern.h> 7075353Smjacob#include <vm/vm_map.h> 7175353Smjacob#include "opt_directio.h" 7295722Sphk#include "opt_swap.h" 7375353Smjacob 74227908Smariusstatic MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 7575353Smjacob 7675353Smjacobstruct bio_ops bioops; /* I/O operation notification */ 7775353Smjacob 7875353Smjacobstruct buf_ops buf_ops_bio = { 79165099Syongari .bop_name = "buf_ops_bio", 80165099Syongari .bop_write = bufwrite, 81221407Smarius .bop_strategy = bufstrategy, 8275353Smjacob}; 83165099Syongari 8475353Smjacob/* 8575353Smjacob * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 8684145Sjlemon * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 8784145Sjlemon */ 8884145Sjlemonstruct buf *buf; /* buffer header pool */ 89221407Smarius 9075353Smjacobstatic struct proc *bufdaemonproc; 91165099Syongari 92165099Syongaristatic int inmem(struct vnode *vp, daddr_t blkno); 93165099Syongaristatic void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 94165099Syongari vm_offset_t to); 95165099Syongaristatic void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 96165099Syongari vm_offset_t to); 97165099Syongaristatic void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 98165099Syongari int pageno, vm_page_t m); 99165099Syongaristatic void vfs_clean_pages(struct buf *bp); 100221407Smariusstatic void vfs_setdirty(struct buf *bp); 101165099Syongaristatic void vfs_vmio_release(struct buf *bp); 102221407Smariusstatic void vfs_backgroundwritedone(struct buf *bp); 103221407Smariusstatic int vfs_bio_clcheck(struct vnode *vp, int size, 104221407Smarius daddr_t lblkno, daddr_t blkno); 105221407Smariusstatic int flushbufqueues(int flushdeps); 106165099Syongaristatic void buf_daemon(void); 107221407Smariusvoid bremfreel(struct buf *bp); 108221407Smarius 109238874Shrsint vmiodirenable = TRUE; 110221407SmariusSYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 111242272Sjmallett "Use the VM system for directory writes"); 112223688Simpint runningbufspace; 113221407SmariusSYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 114221407Smarius "Amount of presently outstanding async buffer io"); 115165099Syongaristatic int bufspace; 116165099SyongariSYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 11775353Smjacob "KVA memory used for bufs"); 118221407Smariusstatic int maxbufspace; 119221407SmariusSYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 120221407Smarius "Maximum allowed value of bufspace (including buf_daemon)"); 121221407Smariusstatic int bufmallocspace; 122221407SmariusSYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 123221407Smarius "Amount of malloced memory for buffers"); 12475353Smjacobstatic int maxbufmallocspace; 12575353SmjacobSYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 12675353Smjacob "Maximum amount of malloced memory for buffers"); 12775353Smjacobstatic int lobufspace; 128165099SyongariSYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 12975353Smjacob "Minimum amount of buffers we want to have"); 13075353Smjacobstatic int hibufspace; 13175353SmjacobSYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 13275353Smjacob "Maximum allowed value of bufspace (excluding buf_daemon)"); 13375353Smjacobstatic int bufreusecnt; 13475353SmjacobSYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 135197590Syongari "Number of times we have reused a buffer"); 13675353Smjacobstatic int buffreekvacnt; 137221407SmariusSYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 13875353Smjacob "Number of times we have freed the KVA space from some buffer"); 139221407Smariusstatic int bufdefragcnt; 14075353SmjacobSYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 141197590Syongari "Number of times we have had to repeat buffer allocation to defragment"); 142213893Smariusstatic int lorunningspace; 143213893SmariusSYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 144213893Smarius "Minimum preferred space used for in-progress I/O"); 145197590Syongaristatic int hirunningspace; 146221407SmariusSYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 147221407Smarius "Maximum amount of space to use for in-progress I/O"); 148221407Smariusstatic int dirtybufferflushes; 149165099SyongariSYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 150165099Syongari 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 151165099Syongaristatic int altbufferflushes; 152221407SmariusSYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 153223688Simp 0, "Number of fsync flushes to limit dirty buffers"); 154183966Syongaristatic int recursiveflushes; 155183966SyongariSYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 156183966Syongari 0, "Number of flushes skipped due to being recursive"); 157183966Syongaristatic int numdirtybuffers; 158183966SyongariSYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 159183966Syongari "Number of buffers that are dirty (has unwritten changes) at the moment"); 160183966Syongaristatic int lodirtybuffers; 161183966SyongariSYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 162183966Syongari "How many buffers we want to have free before bufdaemon can sleep"); 163183966Syongaristatic int hidirtybuffers; 164183966SyongariSYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 165183966Syongari "When the number of dirty buffers is considered severe"); 166183966Syongaristatic int dirtybufthresh; 167165099SyongariSYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 168165099Syongari 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 169221407Smariusstatic int numfreebuffers; 17075353SmjacobSYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 171221407Smarius "Number of free buffers"); 172192708Syongaristatic int lofreebuffers; 173192708SyongariSYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 17484144Sjlemon "XXX Unused"); 175192708Syongaristatic int hifreebuffers; 176192708SyongariSYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 17784144Sjlemon "XXX Complicatedly unused"); 17875353Smjacobstatic int getnewbufcalls; 179165095SyongariSYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 18075353Smjacob "Number of calls to getnewbuf"); 18175353Smjacobstatic int getnewbufrestarts; 18275353SmjacobSYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 18375353Smjacob "Number of times getnewbuf has had to restart a buffer aquisition"); 18475353Smjacobstatic int dobkgrdwrite = 1; 185183493SyongariSYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 18675353Smjacob "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 18775353Smjacob 188165099Syongari/* 189165099Syongari * Wakeup point for bufdaemon, as well as indicator of whether it is already 190165099Syongari * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 191221407Smarius * is idling. 192165099Syongari */ 193183493Syongaristatic int bd_request; 194165099Syongari 195165099Syongari/* 196165099Syongari * This lock synchronizes access to bd_request. 197165099Syongari */ 198165099Syongaristatic struct mtx bdlock; 199214566Smarius 200197590Syongari/* 201197590Syongari * bogus page -- for I/O to/from partially complete buffers 202197590Syongari * this is a temporary solution to the problem, but it is not 203197590Syongari * really that bad. it would be better to split the buffer 204197590Syongari * for input in the case of buffers partially already in memory, 205197590Syongari * but the code is intricate enough already. 206183493Syongari */ 207165099Syongarivm_page_t bogus_page; 208165099Syongari 209221407Smarius/* 210221407Smarius * Synchronization (sleep/wakeup) variable for active buffer space requests. 211221407Smarius * Set when wait starts, cleared prior to wakeup(). 212221407Smarius * Used in runningbufwakeup() and waitrunningbufspace(). 213238874Shrs */ 214221407Smariusstatic int runningbufreq; 215221407Smarius 216223688Simp/* 217221407Smarius * This lock protects the runningbufreq and synchronizes runningbufwakeup and 218165099Syongari * waitrunningbufspace(). 219165099Syongari */ 220165099Syongaristatic struct mtx rbreqlock; 221238874Shrs 222238874Shrs/* 223173133Syongari * Synchronization (sleep/wakeup) variable for buffer requests. 224192713Syongari * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 225165099Syongari * by and/or. 226221407Smarius * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 227165099Syongari * getnewbuf(), and getblk(). 228192713Syongari */ 229165099Syongaristatic int needsbuffer; 230221407Smarius 231192713Syongari/* 232192713Syongari * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 233192713Syongari */ 234192713Syongaristatic struct mtx nblock; 235192713Syongari 236192713Syongari/* 237192713Syongari * Lock that protects against bwait()/bdone()/B_DONE races. 238165099Syongari */ 239165099Syongari 240192713Syongaristatic struct mtx bdonelock; 241165099Syongari 242165099Syongari/* 243221407Smarius * Definitions for the buffer free lists. 244192713Syongari */ 245192713Syongari#define BUFFER_QUEUES 5 /* number of free buffer queues */ 246192713Syongari 247165099Syongari#define QUEUE_NONE 0 /* on no queue */ 248173133Syongari#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 249221407Smarius#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 250238874Shrs#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 251223688Simp#define QUEUE_EMPTY 4 /* empty buffer headers */ 252223688Simp 253173133Syongari/* Queues for free buffers with various properties */ 254173133Syongaristatic TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 255173133Syongari 256173133Syongari/* Lock for the bufqueues */ 257196366Syongaristatic struct mtx bqlock; 258173133Syongari 259165099Syongari/* 26075353Smjacob * Single global constant for BUF_WMESG, to avoid getting multiple references. 261221407Smarius * buf_wmesg is referred from macros. 262221407Smarius */ 263221407Smariusconst char *buf_wmesg = BUF_WMESG; 264221407Smarius 265193291Syongari#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 266221407Smarius#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 267238874Shrs#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 268193291Syongari#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 269193291Syongari 270193291Syongari#ifdef DIRECTIO 271193291Syongariextern void ffs_rawread_setup(void); 272193291Syongari#endif /* DIRECTIO */ 273193291Syongari/* 274193291Syongari * numdirtywakeup: 275193291Syongari * 276193291Syongari * If someone is blocked due to there being too many dirty buffers, 277193291Syongari * and numdirtybuffers is now reasonable, wake them up. 278193291Syongari */ 279193291Syongari 280165099Syongaristatic __inline void 281221407Smariusnumdirtywakeup(int level) 282192713Syongari{ 283192713Syongari 284192713Syongari if (numdirtybuffers <= level) { 285192713Syongari mtx_lock(&nblock); 286192713Syongari if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 287192713Syongari needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 288165099Syongari wakeup(&needsbuffer); 289165099Syongari } 290165099Syongari mtx_unlock(&nblock); 291165099Syongari } 292165099Syongari} 293165099Syongari 294165099Syongari/* 295165099Syongari * bufspacewakeup: 296165099Syongari * 29775353Smjacob * Called when buffer space is potentially available for recovery. 29875353Smjacob * getnewbuf() will block on this flag when it is unable to free 29975353Smjacob * sufficient buffer space. Buffer space becomes recoverable when 30075353Smjacob * bp's get placed back in the queues. 30175353Smjacob */ 30284145Sjlemon 30375353Smjacobstatic __inline void 30475353Smjacobbufspacewakeup(void) 30575353Smjacob{ 306165099Syongari 30775353Smjacob /* 30875353Smjacob * If someone is waiting for BUF space, wake them up. Even 30975353Smjacob * though we haven't freed the kva space yet, the waiting 31075353Smjacob * process will be able to now. 31175353Smjacob */ 31275353Smjacob mtx_lock(&nblock); 31375353Smjacob if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 31475353Smjacob needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 31575353Smjacob wakeup(&needsbuffer); 31675353Smjacob } 317165095Syongari mtx_unlock(&nblock); 31875353Smjacob} 31975353Smjacob 320165099Syongari/* 321221407Smarius * runningbufwakeup() - in-progress I/O accounting. 32275353Smjacob * 323165099Syongari */ 32475353Smjacobstatic __inline void 325165099Syongarirunningbufwakeup(struct buf *bp) 326165099Syongari{ 32795673Sphk 328192708Syongari if (bp->b_runningbufspace) { 329192708Syongari atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 330165099Syongari bp->b_runningbufspace = 0; 331165099Syongari mtx_lock(&rbreqlock); 33275353Smjacob if (runningbufreq && runningbufspace <= lorunningspace) { 333120281Swilko runningbufreq = 0; 334192708Syongari wakeup(&runningbufreq); 335192708Syongari } 336165099Syongari mtx_unlock(&rbreqlock); 337165099Syongari } 338120281Swilko} 33975353Smjacob 340165099Syongari/* 34175353Smjacob * bufcountwakeup: 34275353Smjacob * 343165099Syongari * Called when a buffer has been added to one of the free queues to 34475353Smjacob * account for the buffer and to wakeup anyone waiting for free buffers. 345165099Syongari * This typically occurs when large amounts of metadata are being handled 346165099Syongari * by the buffer cache ( else buffer space runs out first, usually ). 347165099Syongari */ 348165099Syongari 349165099Syongaristatic __inline void 35075353Smjacobbufcountwakeup(void) 35175353Smjacob{ 35275353Smjacob 35375353Smjacob atomic_add_int(&numfreebuffers, 1); 354217412Smarius mtx_lock(&nblock); 355165099Syongari if (needsbuffer) { 356165099Syongari needsbuffer &= ~VFS_BIO_NEED_ANY; 357165099Syongari if (numfreebuffers >= hifreebuffers) 358165099Syongari needsbuffer &= ~VFS_BIO_NEED_FREE; 359165099Syongari wakeup(&needsbuffer); 360165099Syongari } 361165099Syongari mtx_unlock(&nblock); 362165099Syongari} 363165099Syongari 364215297Smarius/* 365215297Smarius * waitrunningbufspace() 366215297Smarius * 367215297Smarius * runningbufspace is a measure of the amount of I/O currently 368215297Smarius * running. This routine is used in async-write situations to 369215297Smarius * prevent creating huge backups of pending writes to a device. 370217412Smarius * Only asynchronous writes are governed by this function. 371217412Smarius * 372165099Syongari * Reads will adjust runningbufspace, but will not block based on it. 373165099Syongari * The read load has a side effect of reducing the allowed write load. 374165099Syongari * 37575353Smjacob * This does NOT turn an async write into a sync write. It waits 37675353Smjacob * for earlier writes to complete and generally returns before the 37775353Smjacob * caller's write has reached the device. 37884145Sjlemon */ 37975353Smjacobstatic __inline void 38084145Sjlemonwaitrunningbufspace(void) 38175353Smjacob{ 38275353Smjacob 38375353Smjacob mtx_lock(&rbreqlock); 38484145Sjlemon while (runningbufspace > hirunningspace) { 38575353Smjacob ++runningbufreq; 386173667Syongari msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 387173667Syongari } 38884145Sjlemon mtx_unlock(&rbreqlock); 389173667Syongari} 39075353Smjacob 39175353Smjacob 39284145Sjlemon/* 39384145Sjlemon * vfs_buf_test_cache: 39475353Smjacob * 39584145Sjlemon * Called when a buffer is extended. This function clears the B_CACHE 396165099Syongari * bit if the newly extended portion of the buffer does not contain 397165099Syongari * valid data. 39884145Sjlemon */ 399165099Syongaristatic __inline 40075353Smjacobvoid 401165099Syongarivfs_buf_test_cache(struct buf *bp, 402173667Syongari vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 403173667Syongari vm_page_t m) 404165099Syongari{ 405192709Syongari 40675353Smjacob GIANT_REQUIRED; 40784145Sjlemon 408221407Smarius VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 409221407Smarius if (bp->b_flags & B_CACHE) { 410165099Syongari int base = (foff + off) & PAGE_MASK; 41175353Smjacob if (vm_page_is_valid(m, base, size) == 0) 41275353Smjacob bp->b_flags &= ~B_CACHE; 41375353Smjacob } 414221407Smarius} 41575353Smjacob 41675353Smjacob/* Wake up the buffer deamon if necessary */ 41784145Sjlemonstatic __inline 41875353Smjacobvoid 41975353Smjacobbd_wakeup(int dirtybuflevel) 42075353Smjacob{ 42184145Sjlemon 42275353Smjacob mtx_lock(&bdlock); 42375353Smjacob if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 42475353Smjacob bd_request = 1; 425215297Smarius wakeup(&bd_request); 42675353Smjacob } 42775353Smjacob mtx_unlock(&bdlock); 42875353Smjacob} 42975353Smjacob 43075353Smjacob/* 43175353Smjacob * bd_speedup - speedup the buffer cache flushing code 43275353Smjacob */ 43375353Smjacob 43475353Smjacobstatic __inline 43575353Smjacobvoid 43675353Smjacobbd_speedup(void) 43775353Smjacob{ 43875353Smjacob 43975353Smjacob bd_wakeup(1); 440192710Syongari} 441192710Syongari 44275353Smjacob/* 44375353Smjacob * Calculating buffer cache scaling values and reserve space for buffer 44475353Smjacob * headers. This is called during low level kernel initialization and 44575353Smjacob * may be called more then once. We CANNOT write to the memory area 44675353Smjacob * being reserved at this time. 447120281Swilko */ 448192710Syongaricaddr_t 449192710Syongarikern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 450120281Swilko{ 451192710Syongari 452192710Syongari /* 453120281Swilko * physmem_est is in pages. Convert it to kilobytes (assumes 454192710Syongari * PAGE_SIZE is >= 1K) 455192710Syongari */ 456120281Swilko physmem_est = physmem_est * (PAGE_SIZE / 1024); 457192710Syongari 458192710Syongari /* 459192710Syongari * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 460192710Syongari * For the first 64MB of ram nominally allocate sufficient buffers to 461192710Syongari * cover 1/4 of our ram. Beyond the first 64MB allocate additional 462120281Swilko * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 463197588Syongari * the buffer cache we limit the eventual kva reservation to 464197588Syongari * maxbcache bytes. 465197588Syongari * 466197588Syongari * factor represents the 1/4 x ram conversion. 467197588Syongari */ 468120281Swilko if (nbuf == 0) { 46975353Smjacob int factor = 4 * BKVASIZE / 1024; 470215297Smarius 47175353Smjacob nbuf = 50; 472215297Smarius if (physmem_est > 4096) 473215297Smarius nbuf += min((physmem_est - 4096) / factor, 474215297Smarius 65536 / factor); 47575353Smjacob if (physmem_est > 65536) 47675353Smjacob nbuf += (physmem_est - 65536) * 2 / (factor * 5); 477215297Smarius 478215297Smarius if (maxbcache && nbuf > maxbcache / BKVASIZE) 479215297Smarius nbuf = maxbcache / BKVASIZE; 480215297Smarius } 48175353Smjacob 48275353Smjacob#if 0 48375353Smjacob /* 48475353Smjacob * Do not allow the buffer_map to be more then 1/2 the size of the 485221407Smarius * kernel_map. 48675353Smjacob */ 487192711Syongari if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 48875353Smjacob (BKVASIZE * 2)) { 489192711Syongari nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 490192711Syongari (BKVASIZE * 2); 491215923Smarius printf("Warning: nbufs capped at %d\n", nbuf); 492192711Syongari } 493215297Smarius#endif 494215297Smarius 495215297Smarius /* 496215297Smarius * swbufs are used as temporary holders for I/O, such as paging I/O. 497192711Syongari * We have no less then 16 and no more then 256. 498192711Syongari */ 499215297Smarius nswbuf = max(min(nbuf/4, 256), 16); 500192708Syongari#ifdef NSWBUF_MIN 501165099Syongari if (nswbuf < NSWBUF_MIN) 502165099Syongari nswbuf = NSWBUF_MIN; 503165099Syongari#endif 504165099Syongari#ifdef DIRECTIO 50575353Smjacob ffs_rawread_setup(); 50675353Smjacob#endif 50775353Smjacob 508 /* 509 * Reserve space for the buffer cache buffers 510 */ 511 swbuf = (void *)v; 512 v = (caddr_t)(swbuf + nswbuf); 513 buf = (void *)v; 514 v = (caddr_t)(buf + nbuf); 515 516 return(v); 517} 518 519/* Initialize the buffer subsystem. Called before use of any buffers. */ 520void 521bufinit(void) 522{ 523 struct buf *bp; 524 int i; 525 526 GIANT_REQUIRED; 527 528 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 529 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 530 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 531 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 532 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 533 534 /* next, make a null set of free lists */ 535 for (i = 0; i < BUFFER_QUEUES; i++) 536 TAILQ_INIT(&bufqueues[i]); 537 538 /* finally, initialize each buffer header and stick on empty q */ 539 for (i = 0; i < nbuf; i++) { 540 bp = &buf[i]; 541 bzero(bp, sizeof *bp); 542 bp->b_flags = B_INVAL; /* we're just an empty header */ 543 bp->b_rcred = NOCRED; 544 bp->b_wcred = NOCRED; 545 bp->b_qindex = QUEUE_EMPTY; 546 bp->b_vflags = 0; 547 bp->b_xflags = 0; 548 LIST_INIT(&bp->b_dep); 549 BUF_LOCKINIT(bp); 550 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 551 } 552 553 /* 554 * maxbufspace is the absolute maximum amount of buffer space we are 555 * allowed to reserve in KVM and in real terms. The absolute maximum 556 * is nominally used by buf_daemon. hibufspace is the nominal maximum 557 * used by most other processes. The differential is required to 558 * ensure that buf_daemon is able to run when other processes might 559 * be blocked waiting for buffer space. 560 * 561 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 562 * this may result in KVM fragmentation which is not handled optimally 563 * by the system. 564 */ 565 maxbufspace = nbuf * BKVASIZE; 566 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 567 lobufspace = hibufspace - MAXBSIZE; 568 569 lorunningspace = 512 * 1024; 570 hirunningspace = 1024 * 1024; 571 572/* 573 * Limit the amount of malloc memory since it is wired permanently into 574 * the kernel space. Even though this is accounted for in the buffer 575 * allocation, we don't want the malloced region to grow uncontrolled. 576 * The malloc scheme improves memory utilization significantly on average 577 * (small) directories. 578 */ 579 maxbufmallocspace = hibufspace / 20; 580 581/* 582 * Reduce the chance of a deadlock occuring by limiting the number 583 * of delayed-write dirty buffers we allow to stack up. 584 */ 585 hidirtybuffers = nbuf / 4 + 20; 586 dirtybufthresh = hidirtybuffers * 9 / 10; 587 numdirtybuffers = 0; 588/* 589 * To support extreme low-memory systems, make sure hidirtybuffers cannot 590 * eat up all available buffer space. This occurs when our minimum cannot 591 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 592 * BKVASIZE'd (8K) buffers. 593 */ 594 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 595 hidirtybuffers >>= 1; 596 } 597 lodirtybuffers = hidirtybuffers / 2; 598 599/* 600 * Try to keep the number of free buffers in the specified range, 601 * and give special processes (e.g. like buf_daemon) access to an 602 * emergency reserve. 603 */ 604 lofreebuffers = nbuf / 18 + 5; 605 hifreebuffers = 2 * lofreebuffers; 606 numfreebuffers = nbuf; 607 608/* 609 * Maximum number of async ops initiated per buf_daemon loop. This is 610 * somewhat of a hack at the moment, we really need to limit ourselves 611 * based on the number of bytes of I/O in-transit that were initiated 612 * from buf_daemon. 613 */ 614 615 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 616 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 617} 618 619/* 620 * bfreekva() - free the kva allocation for a buffer. 621 * 622 * Must be called at splbio() or higher as this is the only locking for 623 * buffer_map. 624 * 625 * Since this call frees up buffer space, we call bufspacewakeup(). 626 */ 627static void 628bfreekva(struct buf *bp) 629{ 630 631 GIANT_REQUIRED; 632 633 if (bp->b_kvasize) { 634 atomic_add_int(&buffreekvacnt, 1); 635 atomic_subtract_int(&bufspace, bp->b_kvasize); 636 vm_map_delete(buffer_map, 637 (vm_offset_t) bp->b_kvabase, 638 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 639 ); 640 bp->b_kvasize = 0; 641 bufspacewakeup(); 642 } 643} 644 645/* 646 * bremfree: 647 * 648 * Remove the buffer from the appropriate free list. 649 */ 650void 651bremfree(struct buf *bp) 652{ 653 654 mtx_lock(&bqlock); 655 bremfreel(bp); 656 mtx_unlock(&bqlock); 657} 658 659void 660bremfreel(struct buf *bp) 661{ 662 int s = splbio(); 663 int old_qindex = bp->b_qindex; 664 665 GIANT_REQUIRED; 666 667 if (bp->b_qindex != QUEUE_NONE) { 668 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 669 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 670 bp->b_qindex = QUEUE_NONE; 671 } else { 672 if (BUF_REFCNT(bp) <= 1) 673 panic("bremfree: removing a buffer not on a queue"); 674 } 675 676 /* 677 * Fixup numfreebuffers count. If the buffer is invalid or not 678 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 679 * the buffer was free and we must decrement numfreebuffers. 680 */ 681 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 682 switch(old_qindex) { 683 case QUEUE_DIRTY: 684 case QUEUE_CLEAN: 685 case QUEUE_EMPTY: 686 case QUEUE_EMPTYKVA: 687 atomic_subtract_int(&numfreebuffers, 1); 688 break; 689 default: 690 break; 691 } 692 } 693 splx(s); 694} 695 696 697/* 698 * Get a buffer with the specified data. Look in the cache first. We 699 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 700 * is set, the buffer is valid and we do not have to do anything ( see 701 * getblk() ). This is really just a special case of breadn(). 702 */ 703int 704bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 705 struct buf **bpp) 706{ 707 708 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 709} 710 711/* 712 * Operates like bread, but also starts asynchronous I/O on 713 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 714 * to initiating I/O . If B_CACHE is set, the buffer is valid 715 * and we do not have to do anything. 716 */ 717int 718breadn(struct vnode * vp, daddr_t blkno, int size, 719 daddr_t * rablkno, int *rabsize, 720 int cnt, struct ucred * cred, struct buf **bpp) 721{ 722 struct buf *bp, *rabp; 723 int i; 724 int rv = 0, readwait = 0; 725 726 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 727 728 /* if not found in cache, do some I/O */ 729 if ((bp->b_flags & B_CACHE) == 0) { 730 if (curthread != PCPU_GET(idlethread)) 731 curthread->td_proc->p_stats->p_ru.ru_inblock++; 732 bp->b_iocmd = BIO_READ; 733 bp->b_flags &= ~B_INVAL; 734 bp->b_ioflags &= ~BIO_ERROR; 735 if (bp->b_rcred == NOCRED && cred != NOCRED) 736 bp->b_rcred = crhold(cred); 737 vfs_busy_pages(bp, 0); 738 bp->b_iooffset = dbtob(bp->b_blkno); 739 bstrategy(bp); 740 ++readwait; 741 } 742 743 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 744 if (inmem(vp, *rablkno)) 745 continue; 746 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 747 748 if ((rabp->b_flags & B_CACHE) == 0) { 749 if (curthread != PCPU_GET(idlethread)) 750 curthread->td_proc->p_stats->p_ru.ru_inblock++; 751 rabp->b_flags |= B_ASYNC; 752 rabp->b_flags &= ~B_INVAL; 753 rabp->b_ioflags &= ~BIO_ERROR; 754 rabp->b_iocmd = BIO_READ; 755 if (rabp->b_rcred == NOCRED && cred != NOCRED) 756 rabp->b_rcred = crhold(cred); 757 vfs_busy_pages(rabp, 0); 758 BUF_KERNPROC(rabp); 759 rabp->b_iooffset = dbtob(rabp->b_blkno); 760 bstrategy(rabp); 761 } else { 762 brelse(rabp); 763 } 764 } 765 766 if (readwait) { 767 rv = bufwait(bp); 768 } 769 return (rv); 770} 771 772/* 773 * Write, release buffer on completion. (Done by iodone 774 * if async). Do not bother writing anything if the buffer 775 * is invalid. 776 * 777 * Note that we set B_CACHE here, indicating that buffer is 778 * fully valid and thus cacheable. This is true even of NFS 779 * now so we set it generally. This could be set either here 780 * or in biodone() since the I/O is synchronous. We put it 781 * here. 782 */ 783int 784bufwrite(struct buf *bp) 785{ 786 int oldflags, s; 787 struct buf *newbp; 788 789 if (bp->b_flags & B_INVAL) { 790 brelse(bp); 791 return (0); 792 } 793 794 oldflags = bp->b_flags; 795 796 if (BUF_REFCNT(bp) == 0) 797 panic("bufwrite: buffer is not busy???"); 798 s = splbio(); 799 /* 800 * If a background write is already in progress, delay 801 * writing this block if it is asynchronous. Otherwise 802 * wait for the background write to complete. 803 */ 804 BO_LOCK(bp->b_bufobj); 805 if (bp->b_vflags & BV_BKGRDINPROG) { 806 if (bp->b_flags & B_ASYNC) { 807 BO_UNLOCK(bp->b_bufobj); 808 splx(s); 809 bdwrite(bp); 810 return (0); 811 } 812 bp->b_vflags |= BV_BKGRDWAIT; 813 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO, "bwrbg", 0); 814 if (bp->b_vflags & BV_BKGRDINPROG) 815 panic("bufwrite: still writing"); 816 } 817 BO_UNLOCK(bp->b_bufobj); 818 819 /* Mark the buffer clean */ 820 bundirty(bp); 821 822 /* 823 * If this buffer is marked for background writing and we 824 * do not have to wait for it, make a copy and write the 825 * copy so as to leave this buffer ready for further use. 826 * 827 * This optimization eats a lot of memory. If we have a page 828 * or buffer shortfall we can't do it. 829 */ 830 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 831 (bp->b_flags & B_ASYNC) && 832 !vm_page_count_severe() && 833 !buf_dirty_count_severe()) { 834 KASSERT(bp->b_iodone == NULL, 835 ("bufwrite: needs chained iodone (%p)", bp->b_iodone)); 836 837 /* get a new block */ 838 newbp = geteblk(bp->b_bufsize); 839 840 /* 841 * set it to be identical to the old block. We have to 842 * set b_lblkno and BKGRDMARKER before calling bgetvp() 843 * to avoid confusing the splay tree and gbincore(). 844 */ 845 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 846 newbp->b_lblkno = bp->b_lblkno; 847 newbp->b_xflags |= BX_BKGRDMARKER; 848 BO_LOCK(bp->b_bufobj); 849 bp->b_vflags |= BV_BKGRDINPROG; 850 bgetvp(bp->b_vp, newbp); 851 BO_UNLOCK(bp->b_bufobj); 852 newbp->b_bufobj = &bp->b_vp->v_bufobj; 853 newbp->b_blkno = bp->b_blkno; 854 newbp->b_offset = bp->b_offset; 855 newbp->b_iodone = vfs_backgroundwritedone; 856 newbp->b_flags |= B_ASYNC; 857 newbp->b_flags &= ~B_INVAL; 858 859 /* move over the dependencies */ 860 if (LIST_FIRST(&bp->b_dep) != NULL) 861 buf_movedeps(bp, newbp); 862 863 /* 864 * Initiate write on the copy, release the original to 865 * the B_LOCKED queue so that it cannot go away until 866 * the background write completes. If not locked it could go 867 * away and then be reconstituted while it was being written. 868 * If the reconstituted buffer were written, we could end up 869 * with two background copies being written at the same time. 870 */ 871 bqrelse(bp); 872 bp = newbp; 873 } 874 875 bp->b_flags &= ~B_DONE; 876 bp->b_ioflags &= ~BIO_ERROR; 877 bp->b_flags |= B_CACHE; 878 bp->b_iocmd = BIO_WRITE; 879 880 bufobj_wref(bp->b_bufobj); 881 vfs_busy_pages(bp, 1); 882 883 /* 884 * Normal bwrites pipeline writes 885 */ 886 bp->b_runningbufspace = bp->b_bufsize; 887 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 888 889 if (curthread != PCPU_GET(idlethread)) 890 curthread->td_proc->p_stats->p_ru.ru_oublock++; 891 splx(s); 892 if (oldflags & B_ASYNC) 893 BUF_KERNPROC(bp); 894 bp->b_iooffset = dbtob(bp->b_blkno); 895 bstrategy(bp); 896 897 if ((oldflags & B_ASYNC) == 0) { 898 int rtval = bufwait(bp); 899 brelse(bp); 900 return (rtval); 901 } else { 902 /* 903 * don't allow the async write to saturate the I/O 904 * system. We will not deadlock here because 905 * we are blocking waiting for I/O that is already in-progress 906 * to complete. We do not block here if it is the update 907 * or syncer daemon trying to clean up as that can lead 908 * to deadlock. 909 */ 910 if (curthread->td_proc != bufdaemonproc && 911 curthread->td_proc != updateproc) 912 waitrunningbufspace(); 913 } 914 915 return (0); 916} 917 918/* 919 * Complete a background write started from bwrite. 920 */ 921static void 922vfs_backgroundwritedone(struct buf *bp) 923{ 924 struct buf *origbp; 925 926 /* 927 * Find the original buffer that we are writing. 928 */ 929 BO_LOCK(bp->b_bufobj); 930 if ((origbp = gbincore(bp->b_bufobj, bp->b_lblkno)) == NULL) 931 panic("backgroundwritedone: lost buffer"); 932 933 /* 934 * Clear the BV_BKGRDINPROG flag in the original buffer 935 * and awaken it if it is waiting for the write to complete. 936 * If BV_BKGRDINPROG is not set in the original buffer it must 937 * have been released and re-instantiated - which is not legal. 938 */ 939 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 940 ("backgroundwritedone: lost buffer2")); 941 origbp->b_vflags &= ~BV_BKGRDINPROG; 942 if (origbp->b_vflags & BV_BKGRDWAIT) { 943 origbp->b_vflags &= ~BV_BKGRDWAIT; 944 wakeup(&origbp->b_xflags); 945 } 946 BO_UNLOCK(bp->b_bufobj); 947 /* 948 * Process dependencies then return any unfinished ones. 949 */ 950 if (LIST_FIRST(&bp->b_dep) != NULL) 951 buf_complete(bp); 952 if (LIST_FIRST(&bp->b_dep) != NULL) 953 buf_movedeps(bp, origbp); 954 955 /* 956 * This buffer is marked B_NOCACHE, so when it is released 957 * by biodone, it will be tossed. We mark it with BIO_READ 958 * to avoid biodone doing a second bufobj_wdrop. 959 */ 960 bp->b_flags |= B_NOCACHE; 961 bp->b_iocmd = BIO_READ; 962 bp->b_flags &= ~(B_CACHE | B_DONE); 963 bp->b_iodone = 0; 964 bufdone(bp); 965} 966 967/* 968 * Delayed write. (Buffer is marked dirty). Do not bother writing 969 * anything if the buffer is marked invalid. 970 * 971 * Note that since the buffer must be completely valid, we can safely 972 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 973 * biodone() in order to prevent getblk from writing the buffer 974 * out synchronously. 975 */ 976void 977bdwrite(struct buf *bp) 978{ 979 struct thread *td = curthread; 980 struct vnode *vp; 981 struct buf *nbp; 982 struct bufobj *bo; 983 984 GIANT_REQUIRED; 985 986 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 987 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 988 989 if (bp->b_flags & B_INVAL) { 990 brelse(bp); 991 return; 992 } 993 994 /* 995 * If we have too many dirty buffers, don't create any more. 996 * If we are wildly over our limit, then force a complete 997 * cleanup. Otherwise, just keep the situation from getting 998 * out of control. Note that we have to avoid a recursive 999 * disaster and not try to clean up after our own cleanup! 1000 */ 1001 vp = bp->b_vp; 1002 bo = bp->b_bufobj; 1003 BO_LOCK(bo); 1004 if (td->td_pflags & TDP_COWINPROGRESS) { 1005 recursiveflushes++; 1006 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1007 BO_UNLOCK(bo); 1008 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 1009 BO_LOCK(bo); 1010 altbufferflushes++; 1011 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1012 /* 1013 * Try to find a buffer to flush. 1014 */ 1015 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1016 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1017 buf_countdeps(nbp, 0) || 1018 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1019 continue; 1020 if (bp == nbp) 1021 panic("bdwrite: found ourselves"); 1022 BO_UNLOCK(bo); 1023 if (nbp->b_flags & B_CLUSTEROK) { 1024 vfs_bio_awrite(nbp); 1025 } else { 1026 bremfree(nbp); 1027 bawrite(nbp); 1028 } 1029 BO_LOCK(bo); 1030 dirtybufferflushes++; 1031 break; 1032 } 1033 } 1034 BO_UNLOCK(bo); 1035 1036 bdirty(bp); 1037 /* 1038 * Set B_CACHE, indicating that the buffer is fully valid. This is 1039 * true even of NFS now. 1040 */ 1041 bp->b_flags |= B_CACHE; 1042 1043 /* 1044 * This bmap keeps the system from needing to do the bmap later, 1045 * perhaps when the system is attempting to do a sync. Since it 1046 * is likely that the indirect block -- or whatever other datastructure 1047 * that the filesystem needs is still in memory now, it is a good 1048 * thing to do this. Note also, that if the pageout daemon is 1049 * requesting a sync -- there might not be enough memory to do 1050 * the bmap then... So, this is important to do. 1051 */ 1052 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1053 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1054 } 1055 1056 /* 1057 * Set the *dirty* buffer range based upon the VM system dirty pages. 1058 */ 1059 vfs_setdirty(bp); 1060 1061 /* 1062 * We need to do this here to satisfy the vnode_pager and the 1063 * pageout daemon, so that it thinks that the pages have been 1064 * "cleaned". Note that since the pages are in a delayed write 1065 * buffer -- the VFS layer "will" see that the pages get written 1066 * out on the next sync, or perhaps the cluster will be completed. 1067 */ 1068 vfs_clean_pages(bp); 1069 bqrelse(bp); 1070 1071 /* 1072 * Wakeup the buffer flushing daemon if we have a lot of dirty 1073 * buffers (midpoint between our recovery point and our stall 1074 * point). 1075 */ 1076 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1077 1078 /* 1079 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1080 * due to the softdep code. 1081 */ 1082} 1083 1084/* 1085 * bdirty: 1086 * 1087 * Turn buffer into delayed write request. We must clear BIO_READ and 1088 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1089 * itself to properly update it in the dirty/clean lists. We mark it 1090 * B_DONE to ensure that any asynchronization of the buffer properly 1091 * clears B_DONE ( else a panic will occur later ). 1092 * 1093 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1094 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1095 * should only be called if the buffer is known-good. 1096 * 1097 * Since the buffer is not on a queue, we do not update the numfreebuffers 1098 * count. 1099 * 1100 * Must be called at splbio(). 1101 * The buffer must be on QUEUE_NONE. 1102 */ 1103void 1104bdirty(struct buf *bp) 1105{ 1106 1107 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1108 KASSERT(bp->b_qindex == QUEUE_NONE, 1109 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1110 bp->b_flags &= ~(B_RELBUF); 1111 bp->b_iocmd = BIO_WRITE; 1112 1113 if ((bp->b_flags & B_DELWRI) == 0) { 1114 bp->b_flags |= B_DONE | B_DELWRI; 1115 reassignbuf(bp); 1116 atomic_add_int(&numdirtybuffers, 1); 1117 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1118 } 1119} 1120 1121/* 1122 * bundirty: 1123 * 1124 * Clear B_DELWRI for buffer. 1125 * 1126 * Since the buffer is not on a queue, we do not update the numfreebuffers 1127 * count. 1128 * 1129 * Must be called at splbio(). 1130 * The buffer must be on QUEUE_NONE. 1131 */ 1132 1133void 1134bundirty(struct buf *bp) 1135{ 1136 1137 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1138 KASSERT(bp->b_qindex == QUEUE_NONE, 1139 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1140 1141 if (bp->b_flags & B_DELWRI) { 1142 bp->b_flags &= ~B_DELWRI; 1143 reassignbuf(bp); 1144 atomic_subtract_int(&numdirtybuffers, 1); 1145 numdirtywakeup(lodirtybuffers); 1146 } 1147 /* 1148 * Since it is now being written, we can clear its deferred write flag. 1149 */ 1150 bp->b_flags &= ~B_DEFERRED; 1151} 1152 1153/* 1154 * bawrite: 1155 * 1156 * Asynchronous write. Start output on a buffer, but do not wait for 1157 * it to complete. The buffer is released when the output completes. 1158 * 1159 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1160 * B_INVAL buffers. Not us. 1161 */ 1162void 1163bawrite(struct buf *bp) 1164{ 1165 1166 bp->b_flags |= B_ASYNC; 1167 (void) bwrite(bp); 1168} 1169 1170/* 1171 * bwillwrite: 1172 * 1173 * Called prior to the locking of any vnodes when we are expecting to 1174 * write. We do not want to starve the buffer cache with too many 1175 * dirty buffers so we block here. By blocking prior to the locking 1176 * of any vnodes we attempt to avoid the situation where a locked vnode 1177 * prevents the various system daemons from flushing related buffers. 1178 */ 1179 1180void 1181bwillwrite(void) 1182{ 1183 1184 if (numdirtybuffers >= hidirtybuffers) { 1185 int s; 1186 1187 mtx_lock(&Giant); 1188 s = splbio(); 1189 mtx_lock(&nblock); 1190 while (numdirtybuffers >= hidirtybuffers) { 1191 bd_wakeup(1); 1192 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1193 msleep(&needsbuffer, &nblock, 1194 (PRIBIO + 4), "flswai", 0); 1195 } 1196 splx(s); 1197 mtx_unlock(&nblock); 1198 mtx_unlock(&Giant); 1199 } 1200} 1201 1202/* 1203 * Return true if we have too many dirty buffers. 1204 */ 1205int 1206buf_dirty_count_severe(void) 1207{ 1208 1209 return(numdirtybuffers >= hidirtybuffers); 1210} 1211 1212/* 1213 * brelse: 1214 * 1215 * Release a busy buffer and, if requested, free its resources. The 1216 * buffer will be stashed in the appropriate bufqueue[] allowing it 1217 * to be accessed later as a cache entity or reused for other purposes. 1218 */ 1219void 1220brelse(struct buf *bp) 1221{ 1222 int s; 1223 1224 GIANT_REQUIRED; 1225 1226 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1227 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1228 1229 s = splbio(); 1230 1231 if (bp->b_iocmd == BIO_WRITE && 1232 (bp->b_ioflags & BIO_ERROR) && 1233 !(bp->b_flags & B_INVAL)) { 1234 /* 1235 * Failed write, redirty. Must clear BIO_ERROR to prevent 1236 * pages from being scrapped. If B_INVAL is set then 1237 * this case is not run and the next case is run to 1238 * destroy the buffer. B_INVAL can occur if the buffer 1239 * is outside the range supported by the underlying device. 1240 */ 1241 bp->b_ioflags &= ~BIO_ERROR; 1242 bdirty(bp); 1243 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1244 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1245 /* 1246 * Either a failed I/O or we were asked to free or not 1247 * cache the buffer. 1248 */ 1249 bp->b_flags |= B_INVAL; 1250 if (LIST_FIRST(&bp->b_dep) != NULL) 1251 buf_deallocate(bp); 1252 if (bp->b_flags & B_DELWRI) { 1253 atomic_subtract_int(&numdirtybuffers, 1); 1254 numdirtywakeup(lodirtybuffers); 1255 } 1256 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1257 if ((bp->b_flags & B_VMIO) == 0) { 1258 if (bp->b_bufsize) 1259 allocbuf(bp, 0); 1260 if (bp->b_vp) 1261 brelvp(bp); 1262 } 1263 } 1264 1265 /* 1266 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1267 * is called with B_DELWRI set, the underlying pages may wind up 1268 * getting freed causing a previous write (bdwrite()) to get 'lost' 1269 * because pages associated with a B_DELWRI bp are marked clean. 1270 * 1271 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1272 * if B_DELWRI is set. 1273 * 1274 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1275 * on pages to return pages to the VM page queues. 1276 */ 1277 if (bp->b_flags & B_DELWRI) 1278 bp->b_flags &= ~B_RELBUF; 1279 else if (vm_page_count_severe()) { 1280 /* 1281 * XXX This lock may not be necessary since BKGRDINPROG 1282 * cannot be set while we hold the buf lock, it can only be 1283 * cleared if it is already pending. 1284 */ 1285 if (bp->b_vp) { 1286 BO_LOCK(bp->b_bufobj); 1287 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1288 bp->b_flags |= B_RELBUF; 1289 BO_UNLOCK(bp->b_bufobj); 1290 } else 1291 bp->b_flags |= B_RELBUF; 1292 } 1293 1294 /* 1295 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1296 * constituted, not even NFS buffers now. Two flags effect this. If 1297 * B_INVAL, the struct buf is invalidated but the VM object is kept 1298 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1299 * 1300 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1301 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1302 * buffer is also B_INVAL because it hits the re-dirtying code above. 1303 * 1304 * Normally we can do this whether a buffer is B_DELWRI or not. If 1305 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1306 * the commit state and we cannot afford to lose the buffer. If the 1307 * buffer has a background write in progress, we need to keep it 1308 * around to prevent it from being reconstituted and starting a second 1309 * background write. 1310 */ 1311 if ((bp->b_flags & B_VMIO) 1312 && !(bp->b_vp->v_mount != NULL && 1313 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1314 !vn_isdisk(bp->b_vp, NULL) && 1315 (bp->b_flags & B_DELWRI)) 1316 ) { 1317 1318 int i, j, resid; 1319 vm_page_t m; 1320 off_t foff; 1321 vm_pindex_t poff; 1322 vm_object_t obj; 1323 1324 obj = bp->b_bufobj->bo_object; 1325 1326 /* 1327 * Get the base offset and length of the buffer. Note that 1328 * in the VMIO case if the buffer block size is not 1329 * page-aligned then b_data pointer may not be page-aligned. 1330 * But our b_pages[] array *IS* page aligned. 1331 * 1332 * block sizes less then DEV_BSIZE (usually 512) are not 1333 * supported due to the page granularity bits (m->valid, 1334 * m->dirty, etc...). 1335 * 1336 * See man buf(9) for more information 1337 */ 1338 resid = bp->b_bufsize; 1339 foff = bp->b_offset; 1340 VM_OBJECT_LOCK(obj); 1341 for (i = 0; i < bp->b_npages; i++) { 1342 int had_bogus = 0; 1343 1344 m = bp->b_pages[i]; 1345 1346 /* 1347 * If we hit a bogus page, fixup *all* the bogus pages 1348 * now. 1349 */ 1350 if (m == bogus_page) { 1351 poff = OFF_TO_IDX(bp->b_offset); 1352 had_bogus = 1; 1353 1354 for (j = i; j < bp->b_npages; j++) { 1355 vm_page_t mtmp; 1356 mtmp = bp->b_pages[j]; 1357 if (mtmp == bogus_page) { 1358 mtmp = vm_page_lookup(obj, poff + j); 1359 if (!mtmp) { 1360 panic("brelse: page missing\n"); 1361 } 1362 bp->b_pages[j] = mtmp; 1363 } 1364 } 1365 1366 if ((bp->b_flags & B_INVAL) == 0) { 1367 pmap_qenter( 1368 trunc_page((vm_offset_t)bp->b_data), 1369 bp->b_pages, bp->b_npages); 1370 } 1371 m = bp->b_pages[i]; 1372 } 1373 if ((bp->b_flags & B_NOCACHE) || 1374 (bp->b_ioflags & BIO_ERROR)) { 1375 int poffset = foff & PAGE_MASK; 1376 int presid = resid > (PAGE_SIZE - poffset) ? 1377 (PAGE_SIZE - poffset) : resid; 1378 1379 KASSERT(presid >= 0, ("brelse: extra page")); 1380 vm_page_lock_queues(); 1381 vm_page_set_invalid(m, poffset, presid); 1382 vm_page_unlock_queues(); 1383 if (had_bogus) 1384 printf("avoided corruption bug in bogus_page/brelse code\n"); 1385 } 1386 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1387 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1388 } 1389 VM_OBJECT_UNLOCK(obj); 1390 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1391 vfs_vmio_release(bp); 1392 1393 } else if (bp->b_flags & B_VMIO) { 1394 1395 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1396 vfs_vmio_release(bp); 1397 } 1398 1399 } 1400 1401 if (bp->b_qindex != QUEUE_NONE) 1402 panic("brelse: free buffer onto another queue???"); 1403 if (BUF_REFCNT(bp) > 1) { 1404 /* do not release to free list */ 1405 BUF_UNLOCK(bp); 1406 splx(s); 1407 return; 1408 } 1409 1410 /* enqueue */ 1411 mtx_lock(&bqlock); 1412 1413 /* buffers with no memory */ 1414 if (bp->b_bufsize == 0) { 1415 bp->b_flags |= B_INVAL; 1416 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1417 if (bp->b_vflags & BV_BKGRDINPROG) 1418 panic("losing buffer 1"); 1419 if (bp->b_kvasize) { 1420 bp->b_qindex = QUEUE_EMPTYKVA; 1421 } else { 1422 bp->b_qindex = QUEUE_EMPTY; 1423 } 1424 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1425 /* buffers with junk contents */ 1426 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1427 (bp->b_ioflags & BIO_ERROR)) { 1428 bp->b_flags |= B_INVAL; 1429 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1430 if (bp->b_vflags & BV_BKGRDINPROG) 1431 panic("losing buffer 2"); 1432 bp->b_qindex = QUEUE_CLEAN; 1433 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1434 /* remaining buffers */ 1435 } else { 1436 if (bp->b_flags & B_DELWRI) 1437 bp->b_qindex = QUEUE_DIRTY; 1438 else 1439 bp->b_qindex = QUEUE_CLEAN; 1440 if (bp->b_flags & B_AGE) 1441 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1442 else 1443 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1444 } 1445 mtx_unlock(&bqlock); 1446 1447 /* 1448 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1449 * placed the buffer on the correct queue. We must also disassociate 1450 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1451 * find it. 1452 */ 1453 if (bp->b_flags & B_INVAL) { 1454 if (bp->b_flags & B_DELWRI) 1455 bundirty(bp); 1456 if (bp->b_vp) 1457 brelvp(bp); 1458 } 1459 1460 /* 1461 * Fixup numfreebuffers count. The bp is on an appropriate queue 1462 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1463 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1464 * if B_INVAL is set ). 1465 */ 1466 1467 if (!(bp->b_flags & B_DELWRI)) 1468 bufcountwakeup(); 1469 1470 /* 1471 * Something we can maybe free or reuse 1472 */ 1473 if (bp->b_bufsize || bp->b_kvasize) 1474 bufspacewakeup(); 1475 1476 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1477 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1478 panic("brelse: not dirty"); 1479 /* unlock */ 1480 BUF_UNLOCK(bp); 1481 splx(s); 1482} 1483 1484/* 1485 * Release a buffer back to the appropriate queue but do not try to free 1486 * it. The buffer is expected to be used again soon. 1487 * 1488 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1489 * biodone() to requeue an async I/O on completion. It is also used when 1490 * known good buffers need to be requeued but we think we may need the data 1491 * again soon. 1492 * 1493 * XXX we should be able to leave the B_RELBUF hint set on completion. 1494 */ 1495void 1496bqrelse(struct buf *bp) 1497{ 1498 int s; 1499 1500 s = splbio(); 1501 1502 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1503 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1504 1505 if (bp->b_qindex != QUEUE_NONE) 1506 panic("bqrelse: free buffer onto another queue???"); 1507 if (BUF_REFCNT(bp) > 1) { 1508 /* do not release to free list */ 1509 BUF_UNLOCK(bp); 1510 splx(s); 1511 return; 1512 } 1513 mtx_lock(&bqlock); 1514 /* buffers with stale but valid contents */ 1515 if (bp->b_flags & B_DELWRI) { 1516 bp->b_qindex = QUEUE_DIRTY; 1517 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1518 } else { 1519 /* 1520 * XXX This lock may not be necessary since BKGRDINPROG 1521 * cannot be set while we hold the buf lock, it can only be 1522 * cleared if it is already pending. 1523 */ 1524 BO_LOCK(bp->b_bufobj); 1525 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1526 BO_UNLOCK(bp->b_bufobj); 1527 bp->b_qindex = QUEUE_CLEAN; 1528 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1529 b_freelist); 1530 } else { 1531 /* 1532 * We are too low on memory, we have to try to free 1533 * the buffer (most importantly: the wired pages 1534 * making up its backing store) *now*. 1535 */ 1536 BO_UNLOCK(bp->b_bufobj); 1537 mtx_unlock(&bqlock); 1538 splx(s); 1539 brelse(bp); 1540 return; 1541 } 1542 } 1543 mtx_unlock(&bqlock); 1544 1545 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1546 bufcountwakeup(); 1547 1548 /* 1549 * Something we can maybe free or reuse. 1550 */ 1551 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1552 bufspacewakeup(); 1553 1554 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1555 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1556 panic("bqrelse: not dirty"); 1557 /* unlock */ 1558 BUF_UNLOCK(bp); 1559 splx(s); 1560} 1561 1562/* Give pages used by the bp back to the VM system (where possible) */ 1563static void 1564vfs_vmio_release(struct buf *bp) 1565{ 1566 int i; 1567 vm_page_t m; 1568 1569 GIANT_REQUIRED; 1570 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1571 vm_page_lock_queues(); 1572 for (i = 0; i < bp->b_npages; i++) { 1573 m = bp->b_pages[i]; 1574 bp->b_pages[i] = NULL; 1575 /* 1576 * In order to keep page LRU ordering consistent, put 1577 * everything on the inactive queue. 1578 */ 1579 vm_page_unwire(m, 0); 1580 /* 1581 * We don't mess with busy pages, it is 1582 * the responsibility of the process that 1583 * busied the pages to deal with them. 1584 */ 1585 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1586 continue; 1587 1588 if (m->wire_count == 0) { 1589 /* 1590 * Might as well free the page if we can and it has 1591 * no valid data. We also free the page if the 1592 * buffer was used for direct I/O 1593 */ 1594 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1595 m->hold_count == 0) { 1596 pmap_remove_all(m); 1597 vm_page_free(m); 1598 } else if (bp->b_flags & B_DIRECT) { 1599 vm_page_try_to_free(m); 1600 } else if (vm_page_count_severe()) { 1601 vm_page_try_to_cache(m); 1602 } 1603 } 1604 } 1605 vm_page_unlock_queues(); 1606 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1607 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1608 1609 if (bp->b_bufsize) { 1610 bufspacewakeup(); 1611 bp->b_bufsize = 0; 1612 } 1613 bp->b_npages = 0; 1614 bp->b_flags &= ~B_VMIO; 1615 if (bp->b_vp) 1616 brelvp(bp); 1617} 1618 1619/* 1620 * Check to see if a block at a particular lbn is available for a clustered 1621 * write. 1622 */ 1623static int 1624vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1625{ 1626 struct buf *bpa; 1627 int match; 1628 1629 match = 0; 1630 1631 /* If the buf isn't in core skip it */ 1632 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1633 return (0); 1634 1635 /* If the buf is busy we don't want to wait for it */ 1636 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1637 return (0); 1638 1639 /* Only cluster with valid clusterable delayed write buffers */ 1640 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1641 (B_DELWRI | B_CLUSTEROK)) 1642 goto done; 1643 1644 if (bpa->b_bufsize != size) 1645 goto done; 1646 1647 /* 1648 * Check to see if it is in the expected place on disk and that the 1649 * block has been mapped. 1650 */ 1651 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1652 match = 1; 1653done: 1654 BUF_UNLOCK(bpa); 1655 return (match); 1656} 1657 1658/* 1659 * vfs_bio_awrite: 1660 * 1661 * Implement clustered async writes for clearing out B_DELWRI buffers. 1662 * This is much better then the old way of writing only one buffer at 1663 * a time. Note that we may not be presented with the buffers in the 1664 * correct order, so we search for the cluster in both directions. 1665 */ 1666int 1667vfs_bio_awrite(struct buf *bp) 1668{ 1669 int i; 1670 int j; 1671 daddr_t lblkno = bp->b_lblkno; 1672 struct vnode *vp = bp->b_vp; 1673 int s; 1674 int ncl; 1675 int nwritten; 1676 int size; 1677 int maxcl; 1678 1679 s = splbio(); 1680 /* 1681 * right now we support clustered writing only to regular files. If 1682 * we find a clusterable block we could be in the middle of a cluster 1683 * rather then at the beginning. 1684 */ 1685 if ((vp->v_type == VREG) && 1686 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1687 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1688 1689 size = vp->v_mount->mnt_stat.f_iosize; 1690 maxcl = MAXPHYS / size; 1691 1692 VI_LOCK(vp); 1693 for (i = 1; i < maxcl; i++) 1694 if (vfs_bio_clcheck(vp, size, lblkno + i, 1695 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1696 break; 1697 1698 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1699 if (vfs_bio_clcheck(vp, size, lblkno - j, 1700 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1701 break; 1702 1703 VI_UNLOCK(vp); 1704 --j; 1705 ncl = i + j; 1706 /* 1707 * this is a possible cluster write 1708 */ 1709 if (ncl != 1) { 1710 BUF_UNLOCK(bp); 1711 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1712 splx(s); 1713 return nwritten; 1714 } 1715 } 1716 1717 bremfree(bp); 1718 bp->b_flags |= B_ASYNC; 1719 1720 splx(s); 1721 /* 1722 * default (old) behavior, writing out only one block 1723 * 1724 * XXX returns b_bufsize instead of b_bcount for nwritten? 1725 */ 1726 nwritten = bp->b_bufsize; 1727 (void) bwrite(bp); 1728 1729 return nwritten; 1730} 1731 1732/* 1733 * getnewbuf: 1734 * 1735 * Find and initialize a new buffer header, freeing up existing buffers 1736 * in the bufqueues as necessary. The new buffer is returned locked. 1737 * 1738 * Important: B_INVAL is not set. If the caller wishes to throw the 1739 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1740 * 1741 * We block if: 1742 * We have insufficient buffer headers 1743 * We have insufficient buffer space 1744 * buffer_map is too fragmented ( space reservation fails ) 1745 * If we have to flush dirty buffers ( but we try to avoid this ) 1746 * 1747 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1748 * Instead we ask the buf daemon to do it for us. We attempt to 1749 * avoid piecemeal wakeups of the pageout daemon. 1750 */ 1751 1752static struct buf * 1753getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1754{ 1755 struct buf *bp; 1756 struct buf *nbp; 1757 int defrag = 0; 1758 int nqindex; 1759 static int flushingbufs; 1760 1761 GIANT_REQUIRED; 1762 1763 /* 1764 * We can't afford to block since we might be holding a vnode lock, 1765 * which may prevent system daemons from running. We deal with 1766 * low-memory situations by proactively returning memory and running 1767 * async I/O rather then sync I/O. 1768 */ 1769 1770 atomic_add_int(&getnewbufcalls, 1); 1771 atomic_subtract_int(&getnewbufrestarts, 1); 1772restart: 1773 atomic_add_int(&getnewbufrestarts, 1); 1774 1775 /* 1776 * Setup for scan. If we do not have enough free buffers, 1777 * we setup a degenerate case that immediately fails. Note 1778 * that if we are specially marked process, we are allowed to 1779 * dip into our reserves. 1780 * 1781 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1782 * 1783 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1784 * However, there are a number of cases (defragging, reusing, ...) 1785 * where we cannot backup. 1786 */ 1787 mtx_lock(&bqlock); 1788 nqindex = QUEUE_EMPTYKVA; 1789 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1790 1791 if (nbp == NULL) { 1792 /* 1793 * If no EMPTYKVA buffers and we are either 1794 * defragging or reusing, locate a CLEAN buffer 1795 * to free or reuse. If bufspace useage is low 1796 * skip this step so we can allocate a new buffer. 1797 */ 1798 if (defrag || bufspace >= lobufspace) { 1799 nqindex = QUEUE_CLEAN; 1800 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1801 } 1802 1803 /* 1804 * If we could not find or were not allowed to reuse a 1805 * CLEAN buffer, check to see if it is ok to use an EMPTY 1806 * buffer. We can only use an EMPTY buffer if allocating 1807 * its KVA would not otherwise run us out of buffer space. 1808 */ 1809 if (nbp == NULL && defrag == 0 && 1810 bufspace + maxsize < hibufspace) { 1811 nqindex = QUEUE_EMPTY; 1812 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1813 } 1814 } 1815 1816 /* 1817 * Run scan, possibly freeing data and/or kva mappings on the fly 1818 * depending. 1819 */ 1820 1821 while ((bp = nbp) != NULL) { 1822 int qindex = nqindex; 1823 1824 /* 1825 * Calculate next bp ( we can only use it if we do not block 1826 * or do other fancy things ). 1827 */ 1828 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1829 switch(qindex) { 1830 case QUEUE_EMPTY: 1831 nqindex = QUEUE_EMPTYKVA; 1832 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1833 break; 1834 /* FALLTHROUGH */ 1835 case QUEUE_EMPTYKVA: 1836 nqindex = QUEUE_CLEAN; 1837 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1838 break; 1839 /* FALLTHROUGH */ 1840 case QUEUE_CLEAN: 1841 /* 1842 * nbp is NULL. 1843 */ 1844 break; 1845 } 1846 } 1847 if (bp->b_vp) { 1848 BO_LOCK(bp->b_bufobj); 1849 if (bp->b_vflags & BV_BKGRDINPROG) { 1850 BO_UNLOCK(bp->b_bufobj); 1851 continue; 1852 } 1853 BO_UNLOCK(bp->b_bufobj); 1854 } 1855 1856 /* 1857 * Sanity Checks 1858 */ 1859 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1860 1861 /* 1862 * Note: we no longer distinguish between VMIO and non-VMIO 1863 * buffers. 1864 */ 1865 1866 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1867 1868 /* 1869 * If we are defragging then we need a buffer with 1870 * b_kvasize != 0. XXX this situation should no longer 1871 * occur, if defrag is non-zero the buffer's b_kvasize 1872 * should also be non-zero at this point. XXX 1873 */ 1874 if (defrag && bp->b_kvasize == 0) { 1875 printf("Warning: defrag empty buffer %p\n", bp); 1876 continue; 1877 } 1878 1879 /* 1880 * Start freeing the bp. This is somewhat involved. nbp 1881 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1882 */ 1883 1884 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1885 panic("getnewbuf: locked buf"); 1886 bremfreel(bp); 1887 mtx_unlock(&bqlock); 1888 1889 if (qindex == QUEUE_CLEAN) { 1890 if (bp->b_flags & B_VMIO) { 1891 bp->b_flags &= ~B_ASYNC; 1892 vfs_vmio_release(bp); 1893 } 1894 if (bp->b_vp) 1895 brelvp(bp); 1896 } 1897 1898 /* 1899 * NOTE: nbp is now entirely invalid. We can only restart 1900 * the scan from this point on. 1901 * 1902 * Get the rest of the buffer freed up. b_kva* is still 1903 * valid after this operation. 1904 */ 1905 1906 if (bp->b_rcred != NOCRED) { 1907 crfree(bp->b_rcred); 1908 bp->b_rcred = NOCRED; 1909 } 1910 if (bp->b_wcred != NOCRED) { 1911 crfree(bp->b_wcred); 1912 bp->b_wcred = NOCRED; 1913 } 1914 if (LIST_FIRST(&bp->b_dep) != NULL) 1915 buf_deallocate(bp); 1916 if (bp->b_vflags & BV_BKGRDINPROG) 1917 panic("losing buffer 3"); 1918 1919 if (bp->b_bufsize) 1920 allocbuf(bp, 0); 1921 1922 bp->b_flags = 0; 1923 bp->b_ioflags = 0; 1924 bp->b_xflags = 0; 1925 bp->b_vflags = 0; 1926 bp->b_vp = NULL; 1927 bp->b_blkno = bp->b_lblkno = 0; 1928 bp->b_offset = NOOFFSET; 1929 bp->b_iodone = 0; 1930 bp->b_error = 0; 1931 bp->b_resid = 0; 1932 bp->b_bcount = 0; 1933 bp->b_npages = 0; 1934 bp->b_dirtyoff = bp->b_dirtyend = 0; 1935 bp->b_magic = B_MAGIC_BIO; 1936 bp->b_bufobj = NULL; 1937 1938 LIST_INIT(&bp->b_dep); 1939 1940 /* 1941 * If we are defragging then free the buffer. 1942 */ 1943 if (defrag) { 1944 bp->b_flags |= B_INVAL; 1945 bfreekva(bp); 1946 brelse(bp); 1947 defrag = 0; 1948 goto restart; 1949 } 1950 1951 /* 1952 * If we are overcomitted then recover the buffer and its 1953 * KVM space. This occurs in rare situations when multiple 1954 * processes are blocked in getnewbuf() or allocbuf(). 1955 */ 1956 if (bufspace >= hibufspace) 1957 flushingbufs = 1; 1958 if (flushingbufs && bp->b_kvasize != 0) { 1959 bp->b_flags |= B_INVAL; 1960 bfreekva(bp); 1961 brelse(bp); 1962 goto restart; 1963 } 1964 if (bufspace < lobufspace) 1965 flushingbufs = 0; 1966 break; 1967 } 1968 1969 /* 1970 * If we exhausted our list, sleep as appropriate. We may have to 1971 * wakeup various daemons and write out some dirty buffers. 1972 * 1973 * Generally we are sleeping due to insufficient buffer space. 1974 */ 1975 1976 if (bp == NULL) { 1977 int flags; 1978 char *waitmsg; 1979 1980 mtx_unlock(&bqlock); 1981 if (defrag) { 1982 flags = VFS_BIO_NEED_BUFSPACE; 1983 waitmsg = "nbufkv"; 1984 } else if (bufspace >= hibufspace) { 1985 waitmsg = "nbufbs"; 1986 flags = VFS_BIO_NEED_BUFSPACE; 1987 } else { 1988 waitmsg = "newbuf"; 1989 flags = VFS_BIO_NEED_ANY; 1990 } 1991 1992 bd_speedup(); /* heeeelp */ 1993 1994 mtx_lock(&nblock); 1995 needsbuffer |= flags; 1996 while (needsbuffer & flags) { 1997 if (msleep(&needsbuffer, &nblock, 1998 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1999 mtx_unlock(&nblock); 2000 return (NULL); 2001 } 2002 } 2003 mtx_unlock(&nblock); 2004 } else { 2005 /* 2006 * We finally have a valid bp. We aren't quite out of the 2007 * woods, we still have to reserve kva space. In order 2008 * to keep fragmentation sane we only allocate kva in 2009 * BKVASIZE chunks. 2010 */ 2011 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2012 2013 if (maxsize != bp->b_kvasize) { 2014 vm_offset_t addr = 0; 2015 2016 bfreekva(bp); 2017 2018 if (vm_map_findspace(buffer_map, 2019 vm_map_min(buffer_map), maxsize, &addr)) { 2020 /* 2021 * Uh oh. Buffer map is to fragmented. We 2022 * must defragment the map. 2023 */ 2024 atomic_add_int(&bufdefragcnt, 1); 2025 defrag = 1; 2026 bp->b_flags |= B_INVAL; 2027 brelse(bp); 2028 goto restart; 2029 } 2030 if (addr) { 2031 vm_map_insert(buffer_map, NULL, 0, 2032 addr, addr + maxsize, 2033 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2034 2035 bp->b_kvabase = (caddr_t) addr; 2036 bp->b_kvasize = maxsize; 2037 atomic_add_int(&bufspace, bp->b_kvasize); 2038 atomic_add_int(&bufreusecnt, 1); 2039 } 2040 } 2041 bp->b_saveaddr = bp->b_kvabase; 2042 bp->b_data = bp->b_saveaddr; 2043 } 2044 return(bp); 2045} 2046 2047/* 2048 * buf_daemon: 2049 * 2050 * buffer flushing daemon. Buffers are normally flushed by the 2051 * update daemon but if it cannot keep up this process starts to 2052 * take the load in an attempt to prevent getnewbuf() from blocking. 2053 */ 2054 2055static struct kproc_desc buf_kp = { 2056 "bufdaemon", 2057 buf_daemon, 2058 &bufdaemonproc 2059}; 2060SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2061 2062static void 2063buf_daemon() 2064{ 2065 int s; 2066 2067 mtx_lock(&Giant); 2068 2069 /* 2070 * This process needs to be suspended prior to shutdown sync. 2071 */ 2072 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2073 SHUTDOWN_PRI_LAST); 2074 2075 /* 2076 * This process is allowed to take the buffer cache to the limit 2077 */ 2078 s = splbio(); 2079 mtx_lock(&bdlock); 2080 2081 for (;;) { 2082 bd_request = 0; 2083 mtx_unlock(&bdlock); 2084 2085 kthread_suspend_check(bufdaemonproc); 2086 2087 /* 2088 * Do the flush. Limit the amount of in-transit I/O we 2089 * allow to build up, otherwise we would completely saturate 2090 * the I/O system. Wakeup any waiting processes before we 2091 * normally would so they can run in parallel with our drain. 2092 */ 2093 while (numdirtybuffers > lodirtybuffers) { 2094 if (flushbufqueues(0) == 0) { 2095 /* 2096 * Could not find any buffers without rollback 2097 * dependencies, so just write the first one 2098 * in the hopes of eventually making progress. 2099 */ 2100 flushbufqueues(1); 2101 break; 2102 } 2103 waitrunningbufspace(); 2104 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2105 } 2106 2107 /* 2108 * Only clear bd_request if we have reached our low water 2109 * mark. The buf_daemon normally waits 1 second and 2110 * then incrementally flushes any dirty buffers that have 2111 * built up, within reason. 2112 * 2113 * If we were unable to hit our low water mark and couldn't 2114 * find any flushable buffers, we sleep half a second. 2115 * Otherwise we loop immediately. 2116 */ 2117 mtx_lock(&bdlock); 2118 if (numdirtybuffers <= lodirtybuffers) { 2119 /* 2120 * We reached our low water mark, reset the 2121 * request and sleep until we are needed again. 2122 * The sleep is just so the suspend code works. 2123 */ 2124 bd_request = 0; 2125 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2126 } else { 2127 /* 2128 * We couldn't find any flushable dirty buffers but 2129 * still have too many dirty buffers, we 2130 * have to sleep and try again. (rare) 2131 */ 2132 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2133 } 2134 } 2135} 2136 2137/* 2138 * flushbufqueues: 2139 * 2140 * Try to flush a buffer in the dirty queue. We must be careful to 2141 * free up B_INVAL buffers instead of write them, which NFS is 2142 * particularly sensitive to. 2143 */ 2144int flushwithdeps = 0; 2145SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2146 0, "Number of buffers flushed with dependecies that require rollbacks"); 2147 2148static int 2149flushbufqueues(int flushdeps) 2150{ 2151 struct thread *td = curthread; 2152 struct vnode *vp; 2153 struct mount *mp; 2154 struct buf *bp; 2155 int hasdeps; 2156 2157 mtx_lock(&bqlock); 2158 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2159 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2160 continue; 2161 KASSERT((bp->b_flags & B_DELWRI), 2162 ("unexpected clean buffer %p", bp)); 2163 BO_LOCK(bp->b_bufobj); 2164 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2165 BO_UNLOCK(bp->b_bufobj); 2166 BUF_UNLOCK(bp); 2167 continue; 2168 } 2169 BO_UNLOCK(bp->b_bufobj); 2170 if (bp->b_flags & B_INVAL) { 2171 bremfreel(bp); 2172 mtx_unlock(&bqlock); 2173 brelse(bp); 2174 return (1); 2175 } 2176 2177 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2178 if (flushdeps == 0) { 2179 BUF_UNLOCK(bp); 2180 continue; 2181 } 2182 hasdeps = 1; 2183 } else 2184 hasdeps = 0; 2185 /* 2186 * We must hold the lock on a vnode before writing 2187 * one of its buffers. Otherwise we may confuse, or 2188 * in the case of a snapshot vnode, deadlock the 2189 * system. 2190 * 2191 * The lock order here is the reverse of the normal 2192 * of vnode followed by buf lock. This is ok because 2193 * the NOWAIT will prevent deadlock. 2194 */ 2195 vp = bp->b_vp; 2196 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2197 BUF_UNLOCK(bp); 2198 continue; 2199 } 2200 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2201 mtx_unlock(&bqlock); 2202 vfs_bio_awrite(bp); 2203 vn_finished_write(mp); 2204 VOP_UNLOCK(vp, 0, td); 2205 flushwithdeps += hasdeps; 2206 return (1); 2207 } 2208 vn_finished_write(mp); 2209 BUF_UNLOCK(bp); 2210 } 2211 mtx_unlock(&bqlock); 2212 return (0); 2213} 2214 2215/* 2216 * Check to see if a block is currently memory resident. 2217 */ 2218struct buf * 2219incore(struct bufobj *bo, daddr_t blkno) 2220{ 2221 struct buf *bp; 2222 2223 int s = splbio(); 2224 BO_LOCK(bo); 2225 bp = gbincore(bo, blkno); 2226 BO_UNLOCK(bo); 2227 splx(s); 2228 return (bp); 2229} 2230 2231/* 2232 * Returns true if no I/O is needed to access the 2233 * associated VM object. This is like incore except 2234 * it also hunts around in the VM system for the data. 2235 */ 2236 2237static int 2238inmem(struct vnode * vp, daddr_t blkno) 2239{ 2240 vm_object_t obj; 2241 vm_offset_t toff, tinc, size; 2242 vm_page_t m; 2243 vm_ooffset_t off; 2244 2245 GIANT_REQUIRED; 2246 ASSERT_VOP_LOCKED(vp, "inmem"); 2247 2248 if (incore(&vp->v_bufobj, blkno)) 2249 return 1; 2250 if (vp->v_mount == NULL) 2251 return 0; 2252 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2253 return 0; 2254 2255 size = PAGE_SIZE; 2256 if (size > vp->v_mount->mnt_stat.f_iosize) 2257 size = vp->v_mount->mnt_stat.f_iosize; 2258 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2259 2260 VM_OBJECT_LOCK(obj); 2261 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2262 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2263 if (!m) 2264 goto notinmem; 2265 tinc = size; 2266 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2267 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2268 if (vm_page_is_valid(m, 2269 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2270 goto notinmem; 2271 } 2272 VM_OBJECT_UNLOCK(obj); 2273 return 1; 2274 2275notinmem: 2276 VM_OBJECT_UNLOCK(obj); 2277 return (0); 2278} 2279 2280/* 2281 * vfs_setdirty: 2282 * 2283 * Sets the dirty range for a buffer based on the status of the dirty 2284 * bits in the pages comprising the buffer. 2285 * 2286 * The range is limited to the size of the buffer. 2287 * 2288 * This routine is primarily used by NFS, but is generalized for the 2289 * B_VMIO case. 2290 */ 2291static void 2292vfs_setdirty(struct buf *bp) 2293{ 2294 int i; 2295 vm_object_t object; 2296 2297 GIANT_REQUIRED; 2298 /* 2299 * Degenerate case - empty buffer 2300 */ 2301 2302 if (bp->b_bufsize == 0) 2303 return; 2304 2305 /* 2306 * We qualify the scan for modified pages on whether the 2307 * object has been flushed yet. The OBJ_WRITEABLE flag 2308 * is not cleared simply by protecting pages off. 2309 */ 2310 2311 if ((bp->b_flags & B_VMIO) == 0) 2312 return; 2313 2314 object = bp->b_pages[0]->object; 2315 VM_OBJECT_LOCK(object); 2316 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2317 printf("Warning: object %p writeable but not mightbedirty\n", object); 2318 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2319 printf("Warning: object %p mightbedirty but not writeable\n", object); 2320 2321 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2322 vm_offset_t boffset; 2323 vm_offset_t eoffset; 2324 2325 vm_page_lock_queues(); 2326 /* 2327 * test the pages to see if they have been modified directly 2328 * by users through the VM system. 2329 */ 2330 for (i = 0; i < bp->b_npages; i++) 2331 vm_page_test_dirty(bp->b_pages[i]); 2332 2333 /* 2334 * Calculate the encompassing dirty range, boffset and eoffset, 2335 * (eoffset - boffset) bytes. 2336 */ 2337 2338 for (i = 0; i < bp->b_npages; i++) { 2339 if (bp->b_pages[i]->dirty) 2340 break; 2341 } 2342 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2343 2344 for (i = bp->b_npages - 1; i >= 0; --i) { 2345 if (bp->b_pages[i]->dirty) { 2346 break; 2347 } 2348 } 2349 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2350 2351 vm_page_unlock_queues(); 2352 /* 2353 * Fit it to the buffer. 2354 */ 2355 2356 if (eoffset > bp->b_bcount) 2357 eoffset = bp->b_bcount; 2358 2359 /* 2360 * If we have a good dirty range, merge with the existing 2361 * dirty range. 2362 */ 2363 2364 if (boffset < eoffset) { 2365 if (bp->b_dirtyoff > boffset) 2366 bp->b_dirtyoff = boffset; 2367 if (bp->b_dirtyend < eoffset) 2368 bp->b_dirtyend = eoffset; 2369 } 2370 } 2371 VM_OBJECT_UNLOCK(object); 2372} 2373 2374/* 2375 * getblk: 2376 * 2377 * Get a block given a specified block and offset into a file/device. 2378 * The buffers B_DONE bit will be cleared on return, making it almost 2379 * ready for an I/O initiation. B_INVAL may or may not be set on 2380 * return. The caller should clear B_INVAL prior to initiating a 2381 * READ. 2382 * 2383 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2384 * an existing buffer. 2385 * 2386 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2387 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2388 * and then cleared based on the backing VM. If the previous buffer is 2389 * non-0-sized but invalid, B_CACHE will be cleared. 2390 * 2391 * If getblk() must create a new buffer, the new buffer is returned with 2392 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2393 * case it is returned with B_INVAL clear and B_CACHE set based on the 2394 * backing VM. 2395 * 2396 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2397 * B_CACHE bit is clear. 2398 * 2399 * What this means, basically, is that the caller should use B_CACHE to 2400 * determine whether the buffer is fully valid or not and should clear 2401 * B_INVAL prior to issuing a read. If the caller intends to validate 2402 * the buffer by loading its data area with something, the caller needs 2403 * to clear B_INVAL. If the caller does this without issuing an I/O, 2404 * the caller should set B_CACHE ( as an optimization ), else the caller 2405 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2406 * a write attempt or if it was a successfull read. If the caller 2407 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2408 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2409 */ 2410struct buf * 2411getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2412 int flags) 2413{ 2414 struct buf *bp; 2415 struct bufobj *bo; 2416 int s; 2417 int error; 2418 ASSERT_VOP_LOCKED(vp, "getblk"); 2419 struct vm_object *vmo; 2420 2421 if (size > MAXBSIZE) 2422 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2423 2424 bo = &vp->v_bufobj; 2425 s = splbio(); 2426loop: 2427 /* 2428 * Block if we are low on buffers. Certain processes are allowed 2429 * to completely exhaust the buffer cache. 2430 * 2431 * If this check ever becomes a bottleneck it may be better to 2432 * move it into the else, when gbincore() fails. At the moment 2433 * it isn't a problem. 2434 * 2435 * XXX remove if 0 sections (clean this up after its proven) 2436 */ 2437 if (numfreebuffers == 0) { 2438 if (curthread == PCPU_GET(idlethread)) 2439 return NULL; 2440 mtx_lock(&nblock); 2441 needsbuffer |= VFS_BIO_NEED_ANY; 2442 mtx_unlock(&nblock); 2443 } 2444 2445 VI_LOCK(vp); 2446 bp = gbincore(bo, blkno); 2447 if (bp != NULL) { 2448 int lockflags; 2449 /* 2450 * Buffer is in-core. If the buffer is not busy, it must 2451 * be on a queue. 2452 */ 2453 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2454 2455 if (flags & GB_LOCK_NOWAIT) 2456 lockflags |= LK_NOWAIT; 2457 2458 error = BUF_TIMELOCK(bp, lockflags, 2459 VI_MTX(vp), "getblk", slpflag, slptimeo); 2460 2461 /* 2462 * If we slept and got the lock we have to restart in case 2463 * the buffer changed identities. 2464 */ 2465 if (error == ENOLCK) 2466 goto loop; 2467 /* We timed out or were interrupted. */ 2468 else if (error) 2469 return (NULL); 2470 2471 /* 2472 * The buffer is locked. B_CACHE is cleared if the buffer is 2473 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2474 * and for a VMIO buffer B_CACHE is adjusted according to the 2475 * backing VM cache. 2476 */ 2477 if (bp->b_flags & B_INVAL) 2478 bp->b_flags &= ~B_CACHE; 2479 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2480 bp->b_flags |= B_CACHE; 2481 bremfree(bp); 2482 2483 /* 2484 * check for size inconsistancies for non-VMIO case. 2485 */ 2486 2487 if (bp->b_bcount != size) { 2488 if ((bp->b_flags & B_VMIO) == 0 || 2489 (size > bp->b_kvasize)) { 2490 if (bp->b_flags & B_DELWRI) { 2491 bp->b_flags |= B_NOCACHE; 2492 bwrite(bp); 2493 } else { 2494 if ((bp->b_flags & B_VMIO) && 2495 (LIST_FIRST(&bp->b_dep) == NULL)) { 2496 bp->b_flags |= B_RELBUF; 2497 brelse(bp); 2498 } else { 2499 bp->b_flags |= B_NOCACHE; 2500 bwrite(bp); 2501 } 2502 } 2503 goto loop; 2504 } 2505 } 2506 2507 /* 2508 * If the size is inconsistant in the VMIO case, we can resize 2509 * the buffer. This might lead to B_CACHE getting set or 2510 * cleared. If the size has not changed, B_CACHE remains 2511 * unchanged from its previous state. 2512 */ 2513 2514 if (bp->b_bcount != size) 2515 allocbuf(bp, size); 2516 2517 KASSERT(bp->b_offset != NOOFFSET, 2518 ("getblk: no buffer offset")); 2519 2520 /* 2521 * A buffer with B_DELWRI set and B_CACHE clear must 2522 * be committed before we can return the buffer in 2523 * order to prevent the caller from issuing a read 2524 * ( due to B_CACHE not being set ) and overwriting 2525 * it. 2526 * 2527 * Most callers, including NFS and FFS, need this to 2528 * operate properly either because they assume they 2529 * can issue a read if B_CACHE is not set, or because 2530 * ( for example ) an uncached B_DELWRI might loop due 2531 * to softupdates re-dirtying the buffer. In the latter 2532 * case, B_CACHE is set after the first write completes, 2533 * preventing further loops. 2534 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2535 * above while extending the buffer, we cannot allow the 2536 * buffer to remain with B_CACHE set after the write 2537 * completes or it will represent a corrupt state. To 2538 * deal with this we set B_NOCACHE to scrap the buffer 2539 * after the write. 2540 * 2541 * We might be able to do something fancy, like setting 2542 * B_CACHE in bwrite() except if B_DELWRI is already set, 2543 * so the below call doesn't set B_CACHE, but that gets real 2544 * confusing. This is much easier. 2545 */ 2546 2547 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2548 bp->b_flags |= B_NOCACHE; 2549 bwrite(bp); 2550 goto loop; 2551 } 2552 2553 splx(s); 2554 bp->b_flags &= ~B_DONE; 2555 } else { 2556 int bsize, maxsize, vmio; 2557 off_t offset; 2558 2559 /* 2560 * Buffer is not in-core, create new buffer. The buffer 2561 * returned by getnewbuf() is locked. Note that the returned 2562 * buffer is also considered valid (not marked B_INVAL). 2563 */ 2564 VI_UNLOCK(vp); 2565 /* 2566 * If the user does not want us to create the buffer, bail out 2567 * here. 2568 */ 2569 if (flags & GB_NOCREAT) { 2570 splx(s); 2571 return NULL; 2572 } 2573 2574 bsize = bo->bo_bsize; 2575 offset = blkno * bsize; 2576 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2577 (vp->v_vflag & VV_OBJBUF); 2578 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2579 maxsize = imax(maxsize, bsize); 2580 2581 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2582 if (bp == NULL) { 2583 if (slpflag || slptimeo) { 2584 splx(s); 2585 return NULL; 2586 } 2587 goto loop; 2588 } 2589 2590 /* 2591 * This code is used to make sure that a buffer is not 2592 * created while the getnewbuf routine is blocked. 2593 * This can be a problem whether the vnode is locked or not. 2594 * If the buffer is created out from under us, we have to 2595 * throw away the one we just created. There is now window 2596 * race because we are safely running at splbio() from the 2597 * point of the duplicate buffer creation through to here, 2598 * and we've locked the buffer. 2599 * 2600 * Note: this must occur before we associate the buffer 2601 * with the vp especially considering limitations in 2602 * the splay tree implementation when dealing with duplicate 2603 * lblkno's. 2604 */ 2605 BO_LOCK(bo); 2606 if (gbincore(bo, blkno)) { 2607 BO_UNLOCK(bo); 2608 bp->b_flags |= B_INVAL; 2609 brelse(bp); 2610 goto loop; 2611 } 2612 2613 /* 2614 * Insert the buffer into the hash, so that it can 2615 * be found by incore. 2616 */ 2617 bp->b_blkno = bp->b_lblkno = blkno; 2618 bp->b_offset = offset; 2619 2620 bgetvp(vp, bp); 2621 BO_UNLOCK(bo); 2622 2623 /* 2624 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2625 * buffer size starts out as 0, B_CACHE will be set by 2626 * allocbuf() for the VMIO case prior to it testing the 2627 * backing store for validity. 2628 */ 2629 2630 if (vmio) { 2631 bp->b_flags |= B_VMIO; 2632#if defined(VFS_BIO_DEBUG) 2633 if (vn_canvmio(vp) != TRUE) 2634 printf("getblk: VMIO on vnode type %d\n", 2635 vp->v_type); 2636#endif 2637 VOP_GETVOBJECT(vp, &vmo); 2638 KASSERT(vmo == bp->b_bufobj->bo_object, 2639 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2640 bp, vmo, bp->b_bufobj->bo_object)); 2641 } else { 2642 bp->b_flags &= ~B_VMIO; 2643 KASSERT(bp->b_bufobj->bo_object == NULL, 2644 ("ARGH! has b_bufobj->bo_object %p %p\n", 2645 bp, bp->b_bufobj->bo_object)); 2646 } 2647 2648 allocbuf(bp, size); 2649 2650 splx(s); 2651 bp->b_flags &= ~B_DONE; 2652 } 2653 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2654 KASSERT(bp->b_bufobj == bo, 2655 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2656 return (bp); 2657} 2658 2659/* 2660 * Get an empty, disassociated buffer of given size. The buffer is initially 2661 * set to B_INVAL. 2662 */ 2663struct buf * 2664geteblk(int size) 2665{ 2666 struct buf *bp; 2667 int s; 2668 int maxsize; 2669 2670 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2671 2672 s = splbio(); 2673 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2674 continue; 2675 splx(s); 2676 allocbuf(bp, size); 2677 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2678 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2679 return (bp); 2680} 2681 2682 2683/* 2684 * This code constitutes the buffer memory from either anonymous system 2685 * memory (in the case of non-VMIO operations) or from an associated 2686 * VM object (in the case of VMIO operations). This code is able to 2687 * resize a buffer up or down. 2688 * 2689 * Note that this code is tricky, and has many complications to resolve 2690 * deadlock or inconsistant data situations. Tread lightly!!! 2691 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2692 * the caller. Calling this code willy nilly can result in the loss of data. 2693 * 2694 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2695 * B_CACHE for the non-VMIO case. 2696 */ 2697 2698int 2699allocbuf(struct buf *bp, int size) 2700{ 2701 int newbsize, mbsize; 2702 int i; 2703 2704 GIANT_REQUIRED; 2705 2706 if (BUF_REFCNT(bp) == 0) 2707 panic("allocbuf: buffer not busy"); 2708 2709 if (bp->b_kvasize < size) 2710 panic("allocbuf: buffer too small"); 2711 2712 if ((bp->b_flags & B_VMIO) == 0) { 2713 caddr_t origbuf; 2714 int origbufsize; 2715 /* 2716 * Just get anonymous memory from the kernel. Don't 2717 * mess with B_CACHE. 2718 */ 2719 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2720 if (bp->b_flags & B_MALLOC) 2721 newbsize = mbsize; 2722 else 2723 newbsize = round_page(size); 2724 2725 if (newbsize < bp->b_bufsize) { 2726 /* 2727 * malloced buffers are not shrunk 2728 */ 2729 if (bp->b_flags & B_MALLOC) { 2730 if (newbsize) { 2731 bp->b_bcount = size; 2732 } else { 2733 free(bp->b_data, M_BIOBUF); 2734 if (bp->b_bufsize) { 2735 atomic_subtract_int( 2736 &bufmallocspace, 2737 bp->b_bufsize); 2738 bufspacewakeup(); 2739 bp->b_bufsize = 0; 2740 } 2741 bp->b_saveaddr = bp->b_kvabase; 2742 bp->b_data = bp->b_saveaddr; 2743 bp->b_bcount = 0; 2744 bp->b_flags &= ~B_MALLOC; 2745 } 2746 return 1; 2747 } 2748 vm_hold_free_pages( 2749 bp, 2750 (vm_offset_t) bp->b_data + newbsize, 2751 (vm_offset_t) bp->b_data + bp->b_bufsize); 2752 } else if (newbsize > bp->b_bufsize) { 2753 /* 2754 * We only use malloced memory on the first allocation. 2755 * and revert to page-allocated memory when the buffer 2756 * grows. 2757 */ 2758 /* 2759 * There is a potential smp race here that could lead 2760 * to bufmallocspace slightly passing the max. It 2761 * is probably extremely rare and not worth worrying 2762 * over. 2763 */ 2764 if ( (bufmallocspace < maxbufmallocspace) && 2765 (bp->b_bufsize == 0) && 2766 (mbsize <= PAGE_SIZE/2)) { 2767 2768 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2769 bp->b_bufsize = mbsize; 2770 bp->b_bcount = size; 2771 bp->b_flags |= B_MALLOC; 2772 atomic_add_int(&bufmallocspace, mbsize); 2773 return 1; 2774 } 2775 origbuf = NULL; 2776 origbufsize = 0; 2777 /* 2778 * If the buffer is growing on its other-than-first allocation, 2779 * then we revert to the page-allocation scheme. 2780 */ 2781 if (bp->b_flags & B_MALLOC) { 2782 origbuf = bp->b_data; 2783 origbufsize = bp->b_bufsize; 2784 bp->b_data = bp->b_kvabase; 2785 if (bp->b_bufsize) { 2786 atomic_subtract_int(&bufmallocspace, 2787 bp->b_bufsize); 2788 bufspacewakeup(); 2789 bp->b_bufsize = 0; 2790 } 2791 bp->b_flags &= ~B_MALLOC; 2792 newbsize = round_page(newbsize); 2793 } 2794 vm_hold_load_pages( 2795 bp, 2796 (vm_offset_t) bp->b_data + bp->b_bufsize, 2797 (vm_offset_t) bp->b_data + newbsize); 2798 if (origbuf) { 2799 bcopy(origbuf, bp->b_data, origbufsize); 2800 free(origbuf, M_BIOBUF); 2801 } 2802 } 2803 } else { 2804 int desiredpages; 2805 2806 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2807 desiredpages = (size == 0) ? 0 : 2808 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2809 2810 if (bp->b_flags & B_MALLOC) 2811 panic("allocbuf: VMIO buffer can't be malloced"); 2812 /* 2813 * Set B_CACHE initially if buffer is 0 length or will become 2814 * 0-length. 2815 */ 2816 if (size == 0 || bp->b_bufsize == 0) 2817 bp->b_flags |= B_CACHE; 2818 2819 if (newbsize < bp->b_bufsize) { 2820 /* 2821 * DEV_BSIZE aligned new buffer size is less then the 2822 * DEV_BSIZE aligned existing buffer size. Figure out 2823 * if we have to remove any pages. 2824 */ 2825 if (desiredpages < bp->b_npages) { 2826 vm_page_t m; 2827 2828 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2829 vm_page_lock_queues(); 2830 for (i = desiredpages; i < bp->b_npages; i++) { 2831 /* 2832 * the page is not freed here -- it 2833 * is the responsibility of 2834 * vnode_pager_setsize 2835 */ 2836 m = bp->b_pages[i]; 2837 KASSERT(m != bogus_page, 2838 ("allocbuf: bogus page found")); 2839 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2840 vm_page_lock_queues(); 2841 2842 bp->b_pages[i] = NULL; 2843 vm_page_unwire(m, 0); 2844 } 2845 vm_page_unlock_queues(); 2846 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2847 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2848 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2849 bp->b_npages = desiredpages; 2850 } 2851 } else if (size > bp->b_bcount) { 2852 /* 2853 * We are growing the buffer, possibly in a 2854 * byte-granular fashion. 2855 */ 2856 struct vnode *vp; 2857 vm_object_t obj; 2858 vm_offset_t toff; 2859 vm_offset_t tinc; 2860 2861 /* 2862 * Step 1, bring in the VM pages from the object, 2863 * allocating them if necessary. We must clear 2864 * B_CACHE if these pages are not valid for the 2865 * range covered by the buffer. 2866 */ 2867 2868 vp = bp->b_vp; 2869 obj = bp->b_bufobj->bo_object; 2870 2871 VM_OBJECT_LOCK(obj); 2872 while (bp->b_npages < desiredpages) { 2873 vm_page_t m; 2874 vm_pindex_t pi; 2875 2876 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2877 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2878 /* 2879 * note: must allocate system pages 2880 * since blocking here could intefere 2881 * with paging I/O, no matter which 2882 * process we are. 2883 */ 2884 m = vm_page_alloc(obj, pi, 2885 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2886 VM_ALLOC_WIRED); 2887 if (m == NULL) { 2888 atomic_add_int(&vm_pageout_deficit, 2889 desiredpages - bp->b_npages); 2890 VM_OBJECT_UNLOCK(obj); 2891 VM_WAIT; 2892 VM_OBJECT_LOCK(obj); 2893 } else { 2894 bp->b_flags &= ~B_CACHE; 2895 bp->b_pages[bp->b_npages] = m; 2896 ++bp->b_npages; 2897 } 2898 continue; 2899 } 2900 2901 /* 2902 * We found a page. If we have to sleep on it, 2903 * retry because it might have gotten freed out 2904 * from under us. 2905 * 2906 * We can only test PG_BUSY here. Blocking on 2907 * m->busy might lead to a deadlock: 2908 * 2909 * vm_fault->getpages->cluster_read->allocbuf 2910 * 2911 */ 2912 vm_page_lock_queues(); 2913 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2914 continue; 2915 2916 /* 2917 * We have a good page. Should we wakeup the 2918 * page daemon? 2919 */ 2920 if ((curproc != pageproc) && 2921 ((m->queue - m->pc) == PQ_CACHE) && 2922 ((cnt.v_free_count + cnt.v_cache_count) < 2923 (cnt.v_free_min + cnt.v_cache_min))) { 2924 pagedaemon_wakeup(); 2925 } 2926 vm_page_wire(m); 2927 vm_page_unlock_queues(); 2928 bp->b_pages[bp->b_npages] = m; 2929 ++bp->b_npages; 2930 } 2931 2932 /* 2933 * Step 2. We've loaded the pages into the buffer, 2934 * we have to figure out if we can still have B_CACHE 2935 * set. Note that B_CACHE is set according to the 2936 * byte-granular range ( bcount and size ), new the 2937 * aligned range ( newbsize ). 2938 * 2939 * The VM test is against m->valid, which is DEV_BSIZE 2940 * aligned. Needless to say, the validity of the data 2941 * needs to also be DEV_BSIZE aligned. Note that this 2942 * fails with NFS if the server or some other client 2943 * extends the file's EOF. If our buffer is resized, 2944 * B_CACHE may remain set! XXX 2945 */ 2946 2947 toff = bp->b_bcount; 2948 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2949 2950 while ((bp->b_flags & B_CACHE) && toff < size) { 2951 vm_pindex_t pi; 2952 2953 if (tinc > (size - toff)) 2954 tinc = size - toff; 2955 2956 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2957 PAGE_SHIFT; 2958 2959 vfs_buf_test_cache( 2960 bp, 2961 bp->b_offset, 2962 toff, 2963 tinc, 2964 bp->b_pages[pi] 2965 ); 2966 toff += tinc; 2967 tinc = PAGE_SIZE; 2968 } 2969 VM_OBJECT_UNLOCK(obj); 2970 2971 /* 2972 * Step 3, fixup the KVM pmap. Remember that 2973 * bp->b_data is relative to bp->b_offset, but 2974 * bp->b_offset may be offset into the first page. 2975 */ 2976 2977 bp->b_data = (caddr_t) 2978 trunc_page((vm_offset_t)bp->b_data); 2979 pmap_qenter( 2980 (vm_offset_t)bp->b_data, 2981 bp->b_pages, 2982 bp->b_npages 2983 ); 2984 2985 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2986 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2987 } 2988 } 2989 if (newbsize < bp->b_bufsize) 2990 bufspacewakeup(); 2991 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2992 bp->b_bcount = size; /* requested buffer size */ 2993 return 1; 2994} 2995 2996void 2997biodone(struct bio *bp) 2998{ 2999 3000 mtx_lock(&bdonelock); 3001 bp->bio_flags |= BIO_DONE; 3002 if (bp->bio_done == NULL) 3003 wakeup(bp); 3004 mtx_unlock(&bdonelock); 3005 if (bp->bio_done != NULL) 3006 bp->bio_done(bp); 3007} 3008 3009/* 3010 * Wait for a BIO to finish. 3011 * 3012 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3013 * case is not yet clear. 3014 */ 3015int 3016biowait(struct bio *bp, const char *wchan) 3017{ 3018 3019 mtx_lock(&bdonelock); 3020 while ((bp->bio_flags & BIO_DONE) == 0) 3021 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3022 mtx_unlock(&bdonelock); 3023 if (bp->bio_error != 0) 3024 return (bp->bio_error); 3025 if (!(bp->bio_flags & BIO_ERROR)) 3026 return (0); 3027 return (EIO); 3028} 3029 3030void 3031biofinish(struct bio *bp, struct devstat *stat, int error) 3032{ 3033 3034 if (error) { 3035 bp->bio_error = error; 3036 bp->bio_flags |= BIO_ERROR; 3037 } 3038 if (stat != NULL) 3039 devstat_end_transaction_bio(stat, bp); 3040 biodone(bp); 3041} 3042 3043/* 3044 * bufwait: 3045 * 3046 * Wait for buffer I/O completion, returning error status. The buffer 3047 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3048 * error and cleared. 3049 */ 3050int 3051bufwait(struct buf *bp) 3052{ 3053 int s; 3054 3055 s = splbio(); 3056 if (bp->b_iocmd == BIO_READ) 3057 bwait(bp, PRIBIO, "biord"); 3058 else 3059 bwait(bp, PRIBIO, "biowr"); 3060 splx(s); 3061 if (bp->b_flags & B_EINTR) { 3062 bp->b_flags &= ~B_EINTR; 3063 return (EINTR); 3064 } 3065 if (bp->b_ioflags & BIO_ERROR) { 3066 return (bp->b_error ? bp->b_error : EIO); 3067 } else { 3068 return (0); 3069 } 3070} 3071 3072 /* 3073 * Call back function from struct bio back up to struct buf. 3074 */ 3075static void 3076bufdonebio(struct bio *bip) 3077{ 3078 struct buf *bp; 3079 3080 /* Device drivers may or may not hold giant, hold it here. */ 3081 mtx_lock(&Giant); 3082 bp = bip->bio_caller2; 3083 bp->b_resid = bp->b_bcount - bip->bio_completed; 3084 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3085 bp->b_ioflags = bip->bio_flags; 3086 bp->b_error = bip->bio_error; 3087 if (bp->b_error) 3088 bp->b_ioflags |= BIO_ERROR; 3089 bufdone(bp); 3090 mtx_unlock(&Giant); 3091 g_destroy_bio(bip); 3092} 3093 3094void 3095dev_strategy(struct cdev *dev, struct buf *bp) 3096{ 3097 struct cdevsw *csw; 3098 struct bio *bip; 3099 3100 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3101 panic("b_iocmd botch"); 3102 for (;;) { 3103 bip = g_new_bio(); 3104 if (bip != NULL) 3105 break; 3106 /* Try again later */ 3107 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3108 } 3109 bip->bio_cmd = bp->b_iocmd; 3110 bip->bio_offset = bp->b_iooffset; 3111 bip->bio_length = bp->b_bcount; 3112 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3113 bip->bio_data = bp->b_data; 3114 bip->bio_done = bufdonebio; 3115 bip->bio_caller2 = bp; 3116 bip->bio_dev = dev; 3117 KASSERT(dev->si_refcount > 0, 3118 ("dev_strategy on un-referenced struct cdev *(%s)", 3119 devtoname(dev))); 3120 csw = dev_refthread(dev); 3121 if (csw == NULL) { 3122 bp->b_error = ENXIO; 3123 bp->b_ioflags = BIO_ERROR; 3124 mtx_lock(&Giant); /* XXX: too defensive ? */ 3125 bufdone(bp); 3126 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3127 return; 3128 } 3129 (*csw->d_strategy)(bip); 3130 dev_relthread(dev); 3131} 3132 3133/* 3134 * bufdone: 3135 * 3136 * Finish I/O on a buffer, optionally calling a completion function. 3137 * This is usually called from an interrupt so process blocking is 3138 * not allowed. 3139 * 3140 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3141 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3142 * assuming B_INVAL is clear. 3143 * 3144 * For the VMIO case, we set B_CACHE if the op was a read and no 3145 * read error occured, or if the op was a write. B_CACHE is never 3146 * set if the buffer is invalid or otherwise uncacheable. 3147 * 3148 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3149 * initiator to leave B_INVAL set to brelse the buffer out of existance 3150 * in the biodone routine. 3151 */ 3152void 3153bufdone(struct buf *bp) 3154{ 3155 int s; 3156 void (*biodone)(struct buf *); 3157 3158 3159 s = splbio(); 3160 3161 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3162 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3163 3164 bp->b_flags |= B_DONE; 3165 runningbufwakeup(bp); 3166 3167 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3168 bufobj_wdrop(bp->b_bufobj); 3169 3170 /* call optional completion function if requested */ 3171 if (bp->b_iodone != NULL) { 3172 biodone = bp->b_iodone; 3173 bp->b_iodone = NULL; 3174 (*biodone) (bp); 3175 splx(s); 3176 return; 3177 } 3178 if (LIST_FIRST(&bp->b_dep) != NULL) 3179 buf_complete(bp); 3180 3181 if (bp->b_flags & B_VMIO) { 3182 int i; 3183 vm_ooffset_t foff; 3184 vm_page_t m; 3185 vm_object_t obj; 3186 int iosize; 3187 struct vnode *vp = bp->b_vp; 3188 3189 obj = bp->b_bufobj->bo_object; 3190 3191#if defined(VFS_BIO_DEBUG) 3192 mp_fixme("usecount and vflag accessed without locks."); 3193 if (vp->v_usecount == 0) { 3194 panic("biodone: zero vnode ref count"); 3195 } 3196 3197 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3198 panic("biodone: vnode is not setup for merged cache"); 3199 } 3200#endif 3201 3202 foff = bp->b_offset; 3203 KASSERT(bp->b_offset != NOOFFSET, 3204 ("biodone: no buffer offset")); 3205 3206 VM_OBJECT_LOCK(obj); 3207#if defined(VFS_BIO_DEBUG) 3208 if (obj->paging_in_progress < bp->b_npages) { 3209 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3210 obj->paging_in_progress, bp->b_npages); 3211 } 3212#endif 3213 3214 /* 3215 * Set B_CACHE if the op was a normal read and no error 3216 * occured. B_CACHE is set for writes in the b*write() 3217 * routines. 3218 */ 3219 iosize = bp->b_bcount - bp->b_resid; 3220 if (bp->b_iocmd == BIO_READ && 3221 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3222 !(bp->b_ioflags & BIO_ERROR)) { 3223 bp->b_flags |= B_CACHE; 3224 } 3225 vm_page_lock_queues(); 3226 for (i = 0; i < bp->b_npages; i++) { 3227 int bogusflag = 0; 3228 int resid; 3229 3230 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3231 if (resid > iosize) 3232 resid = iosize; 3233 3234 /* 3235 * cleanup bogus pages, restoring the originals 3236 */ 3237 m = bp->b_pages[i]; 3238 if (m == bogus_page) { 3239 bogusflag = 1; 3240 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3241 if (m == NULL) 3242 panic("biodone: page disappeared!"); 3243 bp->b_pages[i] = m; 3244 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3245 } 3246#if defined(VFS_BIO_DEBUG) 3247 if (OFF_TO_IDX(foff) != m->pindex) { 3248 printf( 3249"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3250 (intmax_t)foff, (uintmax_t)m->pindex); 3251 } 3252#endif 3253 3254 /* 3255 * In the write case, the valid and clean bits are 3256 * already changed correctly ( see bdwrite() ), so we 3257 * only need to do this here in the read case. 3258 */ 3259 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3260 vfs_page_set_valid(bp, foff, i, m); 3261 } 3262 3263 /* 3264 * when debugging new filesystems or buffer I/O methods, this 3265 * is the most common error that pops up. if you see this, you 3266 * have not set the page busy flag correctly!!! 3267 */ 3268 if (m->busy == 0) { 3269 printf("biodone: page busy < 0, " 3270 "pindex: %d, foff: 0x(%x,%x), " 3271 "resid: %d, index: %d\n", 3272 (int) m->pindex, (int)(foff >> 32), 3273 (int) foff & 0xffffffff, resid, i); 3274 if (!vn_isdisk(vp, NULL)) 3275 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3276 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3277 (intmax_t) bp->b_lblkno, 3278 bp->b_flags, bp->b_npages); 3279 else 3280 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3281 (intmax_t) bp->b_lblkno, 3282 bp->b_flags, bp->b_npages); 3283 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3284 (u_long)m->valid, (u_long)m->dirty, 3285 m->wire_count); 3286 panic("biodone: page busy < 0\n"); 3287 } 3288 vm_page_io_finish(m); 3289 vm_object_pip_subtract(obj, 1); 3290 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3291 iosize -= resid; 3292 } 3293 vm_page_unlock_queues(); 3294 vm_object_pip_wakeupn(obj, 0); 3295 VM_OBJECT_UNLOCK(obj); 3296 } 3297 3298 /* 3299 * For asynchronous completions, release the buffer now. The brelse 3300 * will do a wakeup there if necessary - so no need to do a wakeup 3301 * here in the async case. The sync case always needs to do a wakeup. 3302 */ 3303 3304 if (bp->b_flags & B_ASYNC) { 3305 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3306 brelse(bp); 3307 else 3308 bqrelse(bp); 3309 } else { 3310 bdone(bp); 3311 } 3312 splx(s); 3313} 3314 3315/* 3316 * This routine is called in lieu of iodone in the case of 3317 * incomplete I/O. This keeps the busy status for pages 3318 * consistant. 3319 */ 3320void 3321vfs_unbusy_pages(struct buf *bp) 3322{ 3323 int i; 3324 vm_object_t obj; 3325 vm_page_t m; 3326 3327 runningbufwakeup(bp); 3328 if (!(bp->b_flags & B_VMIO)) 3329 return; 3330 3331 obj = bp->b_bufobj->bo_object; 3332 VM_OBJECT_LOCK(obj); 3333 vm_page_lock_queues(); 3334 for (i = 0; i < bp->b_npages; i++) { 3335 m = bp->b_pages[i]; 3336 if (m == bogus_page) { 3337 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3338 if (!m) 3339 panic("vfs_unbusy_pages: page missing\n"); 3340 bp->b_pages[i] = m; 3341 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3342 bp->b_pages, bp->b_npages); 3343 } 3344 vm_object_pip_subtract(obj, 1); 3345 vm_page_io_finish(m); 3346 } 3347 vm_page_unlock_queues(); 3348 vm_object_pip_wakeupn(obj, 0); 3349 VM_OBJECT_UNLOCK(obj); 3350} 3351 3352/* 3353 * vfs_page_set_valid: 3354 * 3355 * Set the valid bits in a page based on the supplied offset. The 3356 * range is restricted to the buffer's size. 3357 * 3358 * This routine is typically called after a read completes. 3359 */ 3360static void 3361vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3362{ 3363 vm_ooffset_t soff, eoff; 3364 3365 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3366 /* 3367 * Start and end offsets in buffer. eoff - soff may not cross a 3368 * page boundry or cross the end of the buffer. The end of the 3369 * buffer, in this case, is our file EOF, not the allocation size 3370 * of the buffer. 3371 */ 3372 soff = off; 3373 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3374 if (eoff > bp->b_offset + bp->b_bcount) 3375 eoff = bp->b_offset + bp->b_bcount; 3376 3377 /* 3378 * Set valid range. This is typically the entire buffer and thus the 3379 * entire page. 3380 */ 3381 if (eoff > soff) { 3382 vm_page_set_validclean( 3383 m, 3384 (vm_offset_t) (soff & PAGE_MASK), 3385 (vm_offset_t) (eoff - soff) 3386 ); 3387 } 3388} 3389 3390/* 3391 * This routine is called before a device strategy routine. 3392 * It is used to tell the VM system that paging I/O is in 3393 * progress, and treat the pages associated with the buffer 3394 * almost as being PG_BUSY. Also the object paging_in_progress 3395 * flag is handled to make sure that the object doesn't become 3396 * inconsistant. 3397 * 3398 * Since I/O has not been initiated yet, certain buffer flags 3399 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3400 * and should be ignored. 3401 */ 3402void 3403vfs_busy_pages(struct buf *bp, int clear_modify) 3404{ 3405 int i, bogus; 3406 vm_object_t obj; 3407 vm_ooffset_t foff; 3408 vm_page_t m; 3409 3410 if (!(bp->b_flags & B_VMIO)) 3411 return; 3412 3413 obj = bp->b_bufobj->bo_object; 3414 foff = bp->b_offset; 3415 KASSERT(bp->b_offset != NOOFFSET, 3416 ("vfs_busy_pages: no buffer offset")); 3417 vfs_setdirty(bp); 3418 VM_OBJECT_LOCK(obj); 3419retry: 3420 vm_page_lock_queues(); 3421 for (i = 0; i < bp->b_npages; i++) { 3422 m = bp->b_pages[i]; 3423 3424 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3425 goto retry; 3426 } 3427 bogus = 0; 3428 for (i = 0; i < bp->b_npages; i++) { 3429 m = bp->b_pages[i]; 3430 3431 if ((bp->b_flags & B_CLUSTER) == 0) { 3432 vm_object_pip_add(obj, 1); 3433 vm_page_io_start(m); 3434 } 3435 /* 3436 * When readying a buffer for a read ( i.e 3437 * clear_modify == 0 ), it is important to do 3438 * bogus_page replacement for valid pages in 3439 * partially instantiated buffers. Partially 3440 * instantiated buffers can, in turn, occur when 3441 * reconstituting a buffer from its VM backing store 3442 * base. We only have to do this if B_CACHE is 3443 * clear ( which causes the I/O to occur in the 3444 * first place ). The replacement prevents the read 3445 * I/O from overwriting potentially dirty VM-backed 3446 * pages. XXX bogus page replacement is, uh, bogus. 3447 * It may not work properly with small-block devices. 3448 * We need to find a better way. 3449 */ 3450 pmap_remove_all(m); 3451 if (clear_modify) 3452 vfs_page_set_valid(bp, foff, i, m); 3453 else if (m->valid == VM_PAGE_BITS_ALL && 3454 (bp->b_flags & B_CACHE) == 0) { 3455 bp->b_pages[i] = bogus_page; 3456 bogus++; 3457 } 3458 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3459 } 3460 vm_page_unlock_queues(); 3461 VM_OBJECT_UNLOCK(obj); 3462 if (bogus) 3463 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3464 bp->b_pages, bp->b_npages); 3465} 3466 3467/* 3468 * Tell the VM system that the pages associated with this buffer 3469 * are clean. This is used for delayed writes where the data is 3470 * going to go to disk eventually without additional VM intevention. 3471 * 3472 * Note that while we only really need to clean through to b_bcount, we 3473 * just go ahead and clean through to b_bufsize. 3474 */ 3475static void 3476vfs_clean_pages(struct buf *bp) 3477{ 3478 int i; 3479 vm_ooffset_t foff, noff, eoff; 3480 vm_page_t m; 3481 3482 if (!(bp->b_flags & B_VMIO)) 3483 return; 3484 3485 foff = bp->b_offset; 3486 KASSERT(bp->b_offset != NOOFFSET, 3487 ("vfs_clean_pages: no buffer offset")); 3488 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3489 vm_page_lock_queues(); 3490 for (i = 0; i < bp->b_npages; i++) { 3491 m = bp->b_pages[i]; 3492 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3493 eoff = noff; 3494 3495 if (eoff > bp->b_offset + bp->b_bufsize) 3496 eoff = bp->b_offset + bp->b_bufsize; 3497 vfs_page_set_valid(bp, foff, i, m); 3498 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3499 foff = noff; 3500 } 3501 vm_page_unlock_queues(); 3502 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3503} 3504 3505/* 3506 * vfs_bio_set_validclean: 3507 * 3508 * Set the range within the buffer to valid and clean. The range is 3509 * relative to the beginning of the buffer, b_offset. Note that b_offset 3510 * itself may be offset from the beginning of the first page. 3511 * 3512 */ 3513 3514void 3515vfs_bio_set_validclean(struct buf *bp, int base, int size) 3516{ 3517 int i, n; 3518 vm_page_t m; 3519 3520 if (!(bp->b_flags & B_VMIO)) 3521 return; 3522 /* 3523 * Fixup base to be relative to beginning of first page. 3524 * Set initial n to be the maximum number of bytes in the 3525 * first page that can be validated. 3526 */ 3527 3528 base += (bp->b_offset & PAGE_MASK); 3529 n = PAGE_SIZE - (base & PAGE_MASK); 3530 3531 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3532 vm_page_lock_queues(); 3533 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3534 m = bp->b_pages[i]; 3535 if (n > size) 3536 n = size; 3537 vm_page_set_validclean(m, base & PAGE_MASK, n); 3538 base += n; 3539 size -= n; 3540 n = PAGE_SIZE; 3541 } 3542 vm_page_unlock_queues(); 3543 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3544} 3545 3546/* 3547 * vfs_bio_clrbuf: 3548 * 3549 * clear a buffer. This routine essentially fakes an I/O, so we need 3550 * to clear BIO_ERROR and B_INVAL. 3551 * 3552 * Note that while we only theoretically need to clear through b_bcount, 3553 * we go ahead and clear through b_bufsize. 3554 */ 3555 3556void 3557vfs_bio_clrbuf(struct buf *bp) 3558{ 3559 int i, j, mask = 0; 3560 caddr_t sa, ea; 3561 3562 GIANT_REQUIRED; 3563 3564 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3565 clrbuf(bp); 3566 return; 3567 } 3568 3569 bp->b_flags &= ~B_INVAL; 3570 bp->b_ioflags &= ~BIO_ERROR; 3571 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3572 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3573 (bp->b_offset & PAGE_MASK) == 0) { 3574 if (bp->b_pages[0] == bogus_page) 3575 goto unlock; 3576 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3577 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3578 if ((bp->b_pages[0]->valid & mask) == mask) 3579 goto unlock; 3580 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3581 ((bp->b_pages[0]->valid & mask) == 0)) { 3582 bzero(bp->b_data, bp->b_bufsize); 3583 bp->b_pages[0]->valid |= mask; 3584 goto unlock; 3585 } 3586 } 3587 ea = sa = bp->b_data; 3588 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3589 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3590 ea = (caddr_t)(vm_offset_t)ulmin( 3591 (u_long)(vm_offset_t)ea, 3592 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3593 if (bp->b_pages[i] == bogus_page) 3594 continue; 3595 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3596 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3597 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3598 if ((bp->b_pages[i]->valid & mask) == mask) 3599 continue; 3600 if ((bp->b_pages[i]->valid & mask) == 0) { 3601 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3602 bzero(sa, ea - sa); 3603 } else { 3604 for (; sa < ea; sa += DEV_BSIZE, j++) { 3605 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3606 (bp->b_pages[i]->valid & (1 << j)) == 0) 3607 bzero(sa, DEV_BSIZE); 3608 } 3609 } 3610 bp->b_pages[i]->valid |= mask; 3611 } 3612unlock: 3613 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3614 bp->b_resid = 0; 3615} 3616 3617/* 3618 * vm_hold_load_pages and vm_hold_free_pages get pages into 3619 * a buffers address space. The pages are anonymous and are 3620 * not associated with a file object. 3621 */ 3622static void 3623vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3624{ 3625 vm_offset_t pg; 3626 vm_page_t p; 3627 int index; 3628 3629 to = round_page(to); 3630 from = round_page(from); 3631 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3632 3633 VM_OBJECT_LOCK(kernel_object); 3634 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3635tryagain: 3636 /* 3637 * note: must allocate system pages since blocking here 3638 * could intefere with paging I/O, no matter which 3639 * process we are. 3640 */ 3641 p = vm_page_alloc(kernel_object, 3642 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3643 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3644 if (!p) { 3645 atomic_add_int(&vm_pageout_deficit, 3646 (to - pg) >> PAGE_SHIFT); 3647 VM_OBJECT_UNLOCK(kernel_object); 3648 VM_WAIT; 3649 VM_OBJECT_LOCK(kernel_object); 3650 goto tryagain; 3651 } 3652 p->valid = VM_PAGE_BITS_ALL; 3653 pmap_qenter(pg, &p, 1); 3654 bp->b_pages[index] = p; 3655 } 3656 VM_OBJECT_UNLOCK(kernel_object); 3657 bp->b_npages = index; 3658} 3659 3660/* Return pages associated with this buf to the vm system */ 3661static void 3662vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3663{ 3664 vm_offset_t pg; 3665 vm_page_t p; 3666 int index, newnpages; 3667 3668 from = round_page(from); 3669 to = round_page(to); 3670 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3671 3672 VM_OBJECT_LOCK(kernel_object); 3673 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3674 p = bp->b_pages[index]; 3675 if (p && (index < bp->b_npages)) { 3676 if (p->busy) { 3677 printf( 3678 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3679 (intmax_t)bp->b_blkno, 3680 (intmax_t)bp->b_lblkno); 3681 } 3682 bp->b_pages[index] = NULL; 3683 pmap_qremove(pg, 1); 3684 vm_page_lock_queues(); 3685 vm_page_unwire(p, 0); 3686 vm_page_free(p); 3687 vm_page_unlock_queues(); 3688 } 3689 } 3690 VM_OBJECT_UNLOCK(kernel_object); 3691 bp->b_npages = newnpages; 3692} 3693 3694/* 3695 * Map an IO request into kernel virtual address space. 3696 * 3697 * All requests are (re)mapped into kernel VA space. 3698 * Notice that we use b_bufsize for the size of the buffer 3699 * to be mapped. b_bcount might be modified by the driver. 3700 * 3701 * Note that even if the caller determines that the address space should 3702 * be valid, a race or a smaller-file mapped into a larger space may 3703 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3704 * check the return value. 3705 */ 3706int 3707vmapbuf(struct buf *bp) 3708{ 3709 caddr_t addr, kva; 3710 vm_prot_t prot; 3711 int pidx, i; 3712 struct vm_page *m; 3713 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3714 3715 if (bp->b_bufsize < 0) 3716 return (-1); 3717 prot = VM_PROT_READ; 3718 if (bp->b_iocmd == BIO_READ) 3719 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3720 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3721 addr < bp->b_data + bp->b_bufsize; 3722 addr += PAGE_SIZE, pidx++) { 3723 /* 3724 * Do the vm_fault if needed; do the copy-on-write thing 3725 * when reading stuff off device into memory. 3726 * 3727 * NOTE! Must use pmap_extract() because addr may be in 3728 * the userland address space, and kextract is only guarenteed 3729 * to work for the kernland address space (see: sparc64 port). 3730 */ 3731retry: 3732 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3733 prot) < 0) { 3734 vm_page_lock_queues(); 3735 for (i = 0; i < pidx; ++i) { 3736 vm_page_unhold(bp->b_pages[i]); 3737 bp->b_pages[i] = NULL; 3738 } 3739 vm_page_unlock_queues(); 3740 return(-1); 3741 } 3742 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3743 if (m == NULL) 3744 goto retry; 3745 bp->b_pages[pidx] = m; 3746 } 3747 if (pidx > btoc(MAXPHYS)) 3748 panic("vmapbuf: mapped more than MAXPHYS"); 3749 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3750 3751 kva = bp->b_saveaddr; 3752 bp->b_npages = pidx; 3753 bp->b_saveaddr = bp->b_data; 3754 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3755 return(0); 3756} 3757 3758/* 3759 * Free the io map PTEs associated with this IO operation. 3760 * We also invalidate the TLB entries and restore the original b_addr. 3761 */ 3762void 3763vunmapbuf(struct buf *bp) 3764{ 3765 int pidx; 3766 int npages; 3767 3768 npages = bp->b_npages; 3769 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3770 vm_page_lock_queues(); 3771 for (pidx = 0; pidx < npages; pidx++) 3772 vm_page_unhold(bp->b_pages[pidx]); 3773 vm_page_unlock_queues(); 3774 3775 bp->b_data = bp->b_saveaddr; 3776} 3777 3778void 3779bdone(struct buf *bp) 3780{ 3781 3782 mtx_lock(&bdonelock); 3783 bp->b_flags |= B_DONE; 3784 wakeup(bp); 3785 mtx_unlock(&bdonelock); 3786} 3787 3788void 3789bwait(struct buf *bp, u_char pri, const char *wchan) 3790{ 3791 3792 mtx_lock(&bdonelock); 3793 while ((bp->b_flags & B_DONE) == 0) 3794 msleep(bp, &bdonelock, pri, wchan, 0); 3795 mtx_unlock(&bdonelock); 3796} 3797 3798void 3799bufstrategy(struct bufobj *bo, struct buf *bp) 3800{ 3801 int i = 0; 3802 struct vnode *vp; 3803 3804 vp = bp->b_vp; 3805 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3806 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3807 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3808 i = VOP_STRATEGY(vp, bp); 3809 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3810} 3811 3812void 3813bufobj_wref(struct bufobj *bo) 3814{ 3815 3816 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3817 BO_LOCK(bo); 3818 bo->bo_numoutput++; 3819 BO_UNLOCK(bo); 3820} 3821 3822void 3823bufobj_wdrop(struct bufobj *bo) 3824{ 3825 3826 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3827 BO_LOCK(bo); 3828 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3829 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3830 bo->bo_flag &= ~BO_WWAIT; 3831 wakeup(&bo->bo_numoutput); 3832 } 3833 BO_UNLOCK(bo); 3834} 3835 3836int 3837bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3838{ 3839 int error; 3840 3841 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3842 ASSERT_BO_LOCKED(bo); 3843 error = 0; 3844 while (bo->bo_numoutput) { 3845 bo->bo_flag |= BO_WWAIT; 3846 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3847 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3848 if (error) 3849 break; 3850 } 3851 return (error); 3852} 3853 3854#include "opt_ddb.h" 3855#ifdef DDB 3856#include <ddb/ddb.h> 3857 3858/* DDB command to show buffer data */ 3859DB_SHOW_COMMAND(buffer, db_show_buffer) 3860{ 3861 /* get args */ 3862 struct buf *bp = (struct buf *)addr; 3863 3864 if (!have_addr) { 3865 db_printf("usage: show buffer <addr>\n"); 3866 return; 3867 } 3868 3869 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3870 db_printf( 3871 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3872 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3873 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3874 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3875 if (bp->b_npages) { 3876 int i; 3877 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3878 for (i = 0; i < bp->b_npages; i++) { 3879 vm_page_t m; 3880 m = bp->b_pages[i]; 3881 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3882 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3883 if ((i + 1) < bp->b_npages) 3884 db_printf(","); 3885 } 3886 db_printf("\n"); 3887 } 3888} 3889#endif /* DDB */ 3890