1 /*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * this file contains a new buffer I/O scheme implementing a coherent 30 * VM object and buffer cache scheme. Pains have been taken to make 31 * sure that the performance degradation associated with schemes such 32 * as this is not realized. 33 * 34 * Author: John S. Dyson 35 * Significant help during the development and debugging phases 36 * had been provided by David Greenman, also of the FreeBSD core team. 37 * 38 * see man buf(9) for more info. 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/bio.h> 47 #include <sys/conf.h> 48 #include <sys/buf.h> 49 #include <sys/devicestat.h> 50 #include <sys/eventhandler.h> 51 #include <sys/limits.h> 52 #include <sys/lock.h> 53 #include <sys/malloc.h> 54 #include <sys/mount.h> 55 #include <sys/mutex.h> 56 #include <sys/kernel.h> 57 #include <sys/kthread.h> 58 #include <sys/proc.h> 59 #include <sys/resourcevar.h> 60 #include <sys/sysctl.h> 61 #include <sys/vmmeter.h> 62 #include <sys/vnode.h> 63 #include <geom/geom.h> 64 #include <vm/vm.h> 65 #include <vm/vm_param.h> 66 #include <vm/vm_kern.h> 67 #include <vm/vm_pageout.h> 68 #include <vm/vm_page.h> 69 #include <vm/vm_object.h> 70 #include <vm/vm_extern.h> 71 #include <vm/vm_map.h> 72 #include "opt_compat.h" 73 #include "opt_directio.h" 74 #include "opt_swap.h" 75 76 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 77 78 struct bio_ops bioops; /* I/O operation notification */ 79 80 struct buf_ops buf_ops_bio = { 81 .bop_name = "buf_ops_bio", 82 .bop_write = bufwrite, 83 .bop_strategy = bufstrategy, 84 .bop_sync = bufsync, 85 .bop_bdflush = bufbdflush, 86 }; 87 88 /* 89 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 90 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 91 */ 92 struct buf *buf; /* buffer header pool */ 93 94 static struct proc *bufdaemonproc; 95 96 static int inmem(struct vnode *vp, daddr_t blkno); 97 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 98 vm_offset_t to); 99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 100 vm_offset_t to); 101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 103 vm_page_t m); 104 static void vfs_clean_pages(struct buf *bp); 105 static void vfs_setdirty(struct buf *bp); 106 static void vfs_setdirty_locked_object(struct buf *bp); 107 static void vfs_vmio_release(struct buf *bp); 108 static int vfs_bio_clcheck(struct vnode *vp, int size, 109 daddr_t lblkno, daddr_t blkno); 110 static int buf_do_flush(struct vnode *vp); 111 static int flushbufqueues(struct vnode *, int, int); 112 static void buf_daemon(void); 113 static void bremfreel(struct buf *bp); 114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 117 #endif 118 119 int vmiodirenable = TRUE; 120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 121 "Use the VM system for directory writes"); 122 long runningbufspace; 123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 124 "Amount of presently outstanding async buffer io"); 125 static long bufspace; 126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 130 #else 131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 132 "Virtual memory used for buffers"); 133 #endif 134 static long maxbufspace; 135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 136 "Maximum allowed value of bufspace (including buf_daemon)"); 137 static long bufmallocspace; 138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 139 "Amount of malloced memory for buffers"); 140 static long maxbufmallocspace; 141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 142 "Maximum amount of malloced memory for buffers"); 143 static long lobufspace; 144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 145 "Minimum amount of buffers we want to have"); 146 long hibufspace; 147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 148 "Maximum allowed value of bufspace (excluding buf_daemon)"); 149 static int bufreusecnt; 150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 151 "Number of times we have reused a buffer"); 152 static int buffreekvacnt; 153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 154 "Number of times we have freed the KVA space from some buffer"); 155 static int bufdefragcnt; 156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 157 "Number of times we have had to repeat buffer allocation to defragment"); 158 static long lorunningspace; 159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 160 "Minimum preferred space used for in-progress I/O"); 161 static long hirunningspace; 162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 163 "Maximum amount of space to use for in-progress I/O"); 164 int dirtybufferflushes; 165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 167 int bdwriteskip; 168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 170 int altbufferflushes; 171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 172 0, "Number of fsync flushes to limit dirty buffers"); 173 static int recursiveflushes; 174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 175 0, "Number of flushes skipped due to being recursive"); 176 static int numdirtybuffers; 177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 178 "Number of buffers that are dirty (has unwritten changes) at the moment"); 179 static int lodirtybuffers; 180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 181 "How many buffers we want to have free before bufdaemon can sleep"); 182 static int hidirtybuffers; 183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 184 "When the number of dirty buffers is considered severe"); 185 int dirtybufthresh; 186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 188 static int numfreebuffers; 189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 190 "Number of free buffers"); 191 static int lofreebuffers; 192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 193 "XXX Unused"); 194 static int hifreebuffers; 195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 196 "XXX Complicatedly unused"); 197 static int getnewbufcalls; 198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 199 "Number of calls to getnewbuf"); 200 static int getnewbufrestarts; 201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 202 "Number of times getnewbuf has had to restart a buffer aquisition"); 203 static int flushbufqtarget = 100; 204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 205 "Amount of work to do in flushbufqueues when helping bufdaemon"); 206 static long notbufdflashes; 207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 208 "Number of dirty buffer flushes done by the bufdaemon helpers"); 209 210 /* 211 * Wakeup point for bufdaemon, as well as indicator of whether it is already 212 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 213 * is idling. 214 */ 215 static int bd_request; 216 217 /* 218 * This lock synchronizes access to bd_request. 219 */ 220 static struct mtx bdlock; 221 222 /* 223 * bogus page -- for I/O to/from partially complete buffers 224 * this is a temporary solution to the problem, but it is not 225 * really that bad. it would be better to split the buffer 226 * for input in the case of buffers partially already in memory, 227 * but the code is intricate enough already. 228 */ 229 vm_page_t bogus_page; 230 231 /* 232 * Synchronization (sleep/wakeup) variable for active buffer space requests. 233 * Set when wait starts, cleared prior to wakeup(). 234 * Used in runningbufwakeup() and waitrunningbufspace(). 235 */ 236 static int runningbufreq; 237 238 /* 239 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 240 * waitrunningbufspace(). 241 */ 242 static struct mtx rbreqlock; 243 244 /* 245 * Synchronization (sleep/wakeup) variable for buffer requests. 246 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 247 * by and/or. 248 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 249 * getnewbuf(), and getblk(). 250 */ 251 static int needsbuffer; 252 253 /* 254 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 255 */ 256 static struct mtx nblock; 257 258 /* 259 * Definitions for the buffer free lists. 260 */ 261 #define BUFFER_QUEUES 6 /* number of free buffer queues */ 262 263 #define QUEUE_NONE 0 /* on no queue */ 264 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 265 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 266 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 267 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 268 #define QUEUE_EMPTY 5 /* empty buffer headers */ 269 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 270 271 /* Queues for free buffers with various properties */ 272 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 273 274 /* Lock for the bufqueues */ 275 static struct mtx bqlock; 276 277 /* 278 * Single global constant for BUF_WMESG, to avoid getting multiple references. 279 * buf_wmesg is referred from macros. 280 */ 281 const char *buf_wmesg = BUF_WMESG; 282 283 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 284 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 285 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 286 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 287 288 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 289 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 290 static int 291 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 292 { 293 long lvalue; 294 int ivalue; 295 296 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 297 return (sysctl_handle_long(oidp, arg1, arg2, req)); 298 lvalue = *(long *)arg1; 299 if (lvalue > INT_MAX) 300 /* On overflow, still write out a long to trigger ENOMEM. */ 301 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 302 ivalue = lvalue; 303 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 304 } 305 #endif 306 307 #ifdef DIRECTIO 308 extern void ffs_rawread_setup(void); 309 #endif /* DIRECTIO */ 310 /* 311 * numdirtywakeup: 312 * 313 * If someone is blocked due to there being too many dirty buffers, 314 * and numdirtybuffers is now reasonable, wake them up. 315 */ 316 317 static __inline void 318 numdirtywakeup(int level) 319 { 320 321 if (numdirtybuffers <= level) { 322 mtx_lock(&nblock); 323 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 324 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 325 wakeup(&needsbuffer); 326 } 327 mtx_unlock(&nblock); 328 } 329 } 330 331 /* 332 * bufspacewakeup: 333 * 334 * Called when buffer space is potentially available for recovery. 335 * getnewbuf() will block on this flag when it is unable to free 336 * sufficient buffer space. Buffer space becomes recoverable when 337 * bp's get placed back in the queues. 338 */ 339 340 static __inline void 341 bufspacewakeup(void) 342 { 343 344 /* 345 * If someone is waiting for BUF space, wake them up. Even 346 * though we haven't freed the kva space yet, the waiting 347 * process will be able to now. 348 */ 349 mtx_lock(&nblock); 350 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 351 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 352 wakeup(&needsbuffer); 353 } 354 mtx_unlock(&nblock); 355 } 356 357 /* 358 * runningbufwakeup() - in-progress I/O accounting. 359 * 360 */ 361 void 362 runningbufwakeup(struct buf *bp) 363 { 364 365 if (bp->b_runningbufspace) { 366 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 367 bp->b_runningbufspace = 0; 368 mtx_lock(&rbreqlock); 369 if (runningbufreq && runningbufspace <= lorunningspace) { 370 runningbufreq = 0; 371 wakeup(&runningbufreq); 372 } 373 mtx_unlock(&rbreqlock); 374 } 375 } 376 377 /* 378 * bufcountwakeup: 379 * 380 * Called when a buffer has been added to one of the free queues to 381 * account for the buffer and to wakeup anyone waiting for free buffers. 382 * This typically occurs when large amounts of metadata are being handled 383 * by the buffer cache ( else buffer space runs out first, usually ). 384 */ 385 386 static __inline void 387 bufcountwakeup(void) 388 { 389 390 atomic_add_int(&numfreebuffers, 1); 391 mtx_lock(&nblock); 392 if (needsbuffer) { 393 needsbuffer &= ~VFS_BIO_NEED_ANY; 394 if (numfreebuffers >= hifreebuffers) 395 needsbuffer &= ~VFS_BIO_NEED_FREE; 396 wakeup(&needsbuffer); 397 } 398 mtx_unlock(&nblock); 399 } 400 401 /* 402 * waitrunningbufspace() 403 * 404 * runningbufspace is a measure of the amount of I/O currently 405 * running. This routine is used in async-write situations to 406 * prevent creating huge backups of pending writes to a device. 407 * Only asynchronous writes are governed by this function. 408 * 409 * Reads will adjust runningbufspace, but will not block based on it. 410 * The read load has a side effect of reducing the allowed write load. 411 * 412 * This does NOT turn an async write into a sync write. It waits 413 * for earlier writes to complete and generally returns before the 414 * caller's write has reached the device. 415 */ 416 void 417 waitrunningbufspace(void) 418 { 419 420 mtx_lock(&rbreqlock); 421 while (runningbufspace > hirunningspace) { 422 ++runningbufreq; 423 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 424 } 425 mtx_unlock(&rbreqlock); 426 } 427 428 429 /* 430 * vfs_buf_test_cache: 431 * 432 * Called when a buffer is extended. This function clears the B_CACHE 433 * bit if the newly extended portion of the buffer does not contain 434 * valid data. 435 */ 436 static __inline 437 void 438 vfs_buf_test_cache(struct buf *bp, 439 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 440 vm_page_t m) 441 { 442 443 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 444 if (bp->b_flags & B_CACHE) { 445 int base = (foff + off) & PAGE_MASK; 446 if (vm_page_is_valid(m, base, size) == 0) 447 bp->b_flags &= ~B_CACHE; 448 } 449 } 450 451 /* Wake up the buffer daemon if necessary */ 452 static __inline 453 void 454 bd_wakeup(int dirtybuflevel) 455 { 456 457 mtx_lock(&bdlock); 458 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 459 bd_request = 1; 460 wakeup(&bd_request); 461 } 462 mtx_unlock(&bdlock); 463 } 464 465 /* 466 * bd_speedup - speedup the buffer cache flushing code 467 */ 468 469 static __inline 470 void 471 bd_speedup(void) 472 { 473 474 bd_wakeup(1); 475 } 476 477 /* 478 * Calculating buffer cache scaling values and reserve space for buffer 479 * headers. This is called during low level kernel initialization and 480 * may be called more then once. We CANNOT write to the memory area 481 * being reserved at this time. 482 */ 483 caddr_t 484 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 485 { 486 int tuned_nbuf; 487 long maxbuf; 488 489 /* 490 * physmem_est is in pages. Convert it to kilobytes (assumes 491 * PAGE_SIZE is >= 1K) 492 */ 493 physmem_est = physmem_est * (PAGE_SIZE / 1024); 494 495 /* 496 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 497 * For the first 64MB of ram nominally allocate sufficient buffers to 498 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 499 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 500 * the buffer cache we limit the eventual kva reservation to 501 * maxbcache bytes. 502 * 503 * factor represents the 1/4 x ram conversion. 504 */ 505 if (nbuf == 0) { 506 int factor = 4 * BKVASIZE / 1024; 507 508 nbuf = 50; 509 if (physmem_est > 4096) 510 nbuf += min((physmem_est - 4096) / factor, 511 65536 / factor); 512 if (physmem_est > 65536) 513 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 514 515 if (maxbcache && nbuf > maxbcache / BKVASIZE) 516 nbuf = maxbcache / BKVASIZE; 517 tuned_nbuf = 1; 518 } else 519 tuned_nbuf = 0; 520 521 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 522 maxbuf = (LONG_MAX / 3) / BKVASIZE; 523 if (nbuf > maxbuf) { 524 if (!tuned_nbuf) 525 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 526 maxbuf); 527 nbuf = maxbuf; 528 } 529 530 /* 531 * swbufs are used as temporary holders for I/O, such as paging I/O. 532 * We have no less then 16 and no more then 256. 533 */ 534 nswbuf = max(min(nbuf/4, 256), 16); 535 #ifdef NSWBUF_MIN 536 if (nswbuf < NSWBUF_MIN) 537 nswbuf = NSWBUF_MIN; 538 #endif 539 #ifdef DIRECTIO 540 ffs_rawread_setup(); 541 #endif 542 543 /* 544 * Reserve space for the buffer cache buffers 545 */ 546 swbuf = (void *)v; 547 v = (caddr_t)(swbuf + nswbuf); 548 buf = (void *)v; 549 v = (caddr_t)(buf + nbuf); 550 551 return(v); 552 } 553 554 /* Initialize the buffer subsystem. Called before use of any buffers. */ 555 void 556 bufinit(void) 557 { 558 struct buf *bp; 559 int i; 560 561 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 562 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 563 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 564 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 565 566 /* next, make a null set of free lists */ 567 for (i = 0; i < BUFFER_QUEUES; i++) 568 TAILQ_INIT(&bufqueues[i]); 569 570 /* finally, initialize each buffer header and stick on empty q */ 571 for (i = 0; i < nbuf; i++) { 572 bp = &buf[i]; 573 bzero(bp, sizeof *bp); 574 bp->b_flags = B_INVAL; /* we're just an empty header */ 575 bp->b_rcred = NOCRED; 576 bp->b_wcred = NOCRED; 577 bp->b_qindex = QUEUE_EMPTY; 578 bp->b_vflags = 0; 579 bp->b_xflags = 0; 580 LIST_INIT(&bp->b_dep); 581 BUF_LOCKINIT(bp); 582 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 583 } 584 585 /* 586 * maxbufspace is the absolute maximum amount of buffer space we are 587 * allowed to reserve in KVM and in real terms. The absolute maximum 588 * is nominally used by buf_daemon. hibufspace is the nominal maximum 589 * used by most other processes. The differential is required to 590 * ensure that buf_daemon is able to run when other processes might 591 * be blocked waiting for buffer space. 592 * 593 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 594 * this may result in KVM fragmentation which is not handled optimally 595 * by the system. 596 */ 597 maxbufspace = (long)nbuf * BKVASIZE; 598 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 599 lobufspace = hibufspace - MAXBSIZE; 600 601 lorunningspace = 512 * 1024; 602 hirunningspace = 1024 * 1024; 603 604 /* 605 * Limit the amount of malloc memory since it is wired permanently into 606 * the kernel space. Even though this is accounted for in the buffer 607 * allocation, we don't want the malloced region to grow uncontrolled. 608 * The malloc scheme improves memory utilization significantly on average 609 * (small) directories. 610 */ 611 maxbufmallocspace = hibufspace / 20; 612 613 /* 614 * Reduce the chance of a deadlock occuring by limiting the number 615 * of delayed-write dirty buffers we allow to stack up. 616 */ 617 hidirtybuffers = nbuf / 4 + 20; 618 dirtybufthresh = hidirtybuffers * 9 / 10; 619 numdirtybuffers = 0; 620 /* 621 * To support extreme low-memory systems, make sure hidirtybuffers cannot 622 * eat up all available buffer space. This occurs when our minimum cannot 623 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 624 * BKVASIZE'd (8K) buffers. 625 */ 626 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 627 hidirtybuffers >>= 1; 628 } 629 lodirtybuffers = hidirtybuffers / 2; 630 631 /* 632 * Try to keep the number of free buffers in the specified range, 633 * and give special processes (e.g. like buf_daemon) access to an 634 * emergency reserve. 635 */ 636 lofreebuffers = nbuf / 18 + 5; 637 hifreebuffers = 2 * lofreebuffers; 638 numfreebuffers = nbuf; 639 640 /* 641 * Maximum number of async ops initiated per buf_daemon loop. This is 642 * somewhat of a hack at the moment, we really need to limit ourselves 643 * based on the number of bytes of I/O in-transit that were initiated 644 * from buf_daemon. 645 */ 646 647 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 648 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 649 } 650 651 /* 652 * bfreekva() - free the kva allocation for a buffer. 653 * 654 * Since this call frees up buffer space, we call bufspacewakeup(). 655 */ 656 static void 657 bfreekva(struct buf *bp) 658 { 659 660 if (bp->b_kvasize) { 661 atomic_add_int(&buffreekvacnt, 1); 662 atomic_subtract_long(&bufspace, bp->b_kvasize); 663 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 664 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 665 bp->b_kvasize = 0; 666 bufspacewakeup(); 667 } 668 } 669 670 /* 671 * bremfree: 672 * 673 * Mark the buffer for removal from the appropriate free list in brelse. 674 * 675 */ 676 void 677 bremfree(struct buf *bp) 678 { 679 680 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 681 KASSERT((bp->b_flags & B_REMFREE) == 0, 682 ("bremfree: buffer %p already marked for delayed removal.", bp)); 683 KASSERT(bp->b_qindex != QUEUE_NONE, 684 ("bremfree: buffer %p not on a queue.", bp)); 685 BUF_ASSERT_HELD(bp); 686 687 bp->b_flags |= B_REMFREE; 688 /* Fixup numfreebuffers count. */ 689 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 690 atomic_subtract_int(&numfreebuffers, 1); 691 } 692 693 /* 694 * bremfreef: 695 * 696 * Force an immediate removal from a free list. Used only in nfs when 697 * it abuses the b_freelist pointer. 698 */ 699 void 700 bremfreef(struct buf *bp) 701 { 702 mtx_lock(&bqlock); 703 bremfreel(bp); 704 mtx_unlock(&bqlock); 705 } 706 707 /* 708 * bremfreel: 709 * 710 * Removes a buffer from the free list, must be called with the 711 * bqlock held. 712 */ 713 static void 714 bremfreel(struct buf *bp) 715 { 716 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 717 bp, bp->b_vp, bp->b_flags); 718 KASSERT(bp->b_qindex != QUEUE_NONE, 719 ("bremfreel: buffer %p not on a queue.", bp)); 720 BUF_ASSERT_HELD(bp); 721 mtx_assert(&bqlock, MA_OWNED); 722 723 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 724 bp->b_qindex = QUEUE_NONE; 725 /* 726 * If this was a delayed bremfree() we only need to remove the buffer 727 * from the queue and return the stats are already done. 728 */ 729 if (bp->b_flags & B_REMFREE) { 730 bp->b_flags &= ~B_REMFREE; 731 return; 732 } 733 /* 734 * Fixup numfreebuffers count. If the buffer is invalid or not 735 * delayed-write, the buffer was free and we must decrement 736 * numfreebuffers. 737 */ 738 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 739 atomic_subtract_int(&numfreebuffers, 1); 740 } 741 742 743 /* 744 * Get a buffer with the specified data. Look in the cache first. We 745 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 746 * is set, the buffer is valid and we do not have to do anything ( see 747 * getblk() ). This is really just a special case of breadn(). 748 */ 749 int 750 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 751 struct buf **bpp) 752 { 753 754 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 755 } 756 757 /* 758 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 759 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 760 * the buffer is valid and we do not have to do anything. 761 */ 762 void 763 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 764 int cnt, struct ucred * cred) 765 { 766 struct buf *rabp; 767 int i; 768 769 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 770 if (inmem(vp, *rablkno)) 771 continue; 772 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 773 774 if ((rabp->b_flags & B_CACHE) == 0) { 775 if (!TD_IS_IDLETHREAD(curthread)) 776 curthread->td_ru.ru_inblock++; 777 rabp->b_flags |= B_ASYNC; 778 rabp->b_flags &= ~B_INVAL; 779 rabp->b_ioflags &= ~BIO_ERROR; 780 rabp->b_iocmd = BIO_READ; 781 if (rabp->b_rcred == NOCRED && cred != NOCRED) 782 rabp->b_rcred = crhold(cred); 783 vfs_busy_pages(rabp, 0); 784 BUF_KERNPROC(rabp); 785 rabp->b_iooffset = dbtob(rabp->b_blkno); 786 bstrategy(rabp); 787 } else { 788 brelse(rabp); 789 } 790 } 791 } 792 793 /* 794 * Operates like bread, but also starts asynchronous I/O on 795 * read-ahead blocks. 796 */ 797 int 798 breadn(struct vnode * vp, daddr_t blkno, int size, 799 daddr_t * rablkno, int *rabsize, 800 int cnt, struct ucred * cred, struct buf **bpp) 801 { 802 struct buf *bp; 803 int rv = 0, readwait = 0; 804 805 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 806 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 807 808 /* if not found in cache, do some I/O */ 809 if ((bp->b_flags & B_CACHE) == 0) { 810 if (!TD_IS_IDLETHREAD(curthread)) 811 curthread->td_ru.ru_inblock++; 812 bp->b_iocmd = BIO_READ; 813 bp->b_flags &= ~B_INVAL; 814 bp->b_ioflags &= ~BIO_ERROR; 815 if (bp->b_rcred == NOCRED && cred != NOCRED) 816 bp->b_rcred = crhold(cred); 817 vfs_busy_pages(bp, 0); 818 bp->b_iooffset = dbtob(bp->b_blkno); 819 bstrategy(bp); 820 ++readwait; 821 } 822 823 breada(vp, rablkno, rabsize, cnt, cred); 824 825 if (readwait) { 826 rv = bufwait(bp); 827 } 828 return (rv); 829 } 830 831 /* 832 * Write, release buffer on completion. (Done by iodone 833 * if async). Do not bother writing anything if the buffer 834 * is invalid. 835 * 836 * Note that we set B_CACHE here, indicating that buffer is 837 * fully valid and thus cacheable. This is true even of NFS 838 * now so we set it generally. This could be set either here 839 * or in biodone() since the I/O is synchronous. We put it 840 * here. 841 */ 842 int 843 bufwrite(struct buf *bp) 844 { 845 int oldflags; 846 struct vnode *vp; 847 int vp_md; 848 849 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 850 if (bp->b_flags & B_INVAL) { 851 brelse(bp); 852 return (0); 853 } 854 855 oldflags = bp->b_flags; 856 857 BUF_ASSERT_HELD(bp); 858 859 if (bp->b_pin_count > 0) 860 bunpin_wait(bp); 861 862 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 863 ("FFS background buffer should not get here %p", bp)); 864 865 vp = bp->b_vp; 866 if (vp) 867 vp_md = vp->v_vflag & VV_MD; 868 else 869 vp_md = 0; 870 871 /* Mark the buffer clean */ 872 bundirty(bp); 873 874 bp->b_flags &= ~B_DONE; 875 bp->b_ioflags &= ~BIO_ERROR; 876 bp->b_flags |= B_CACHE; 877 bp->b_iocmd = BIO_WRITE; 878 879 bufobj_wref(bp->b_bufobj); 880 vfs_busy_pages(bp, 1); 881 882 /* 883 * Normal bwrites pipeline writes 884 */ 885 bp->b_runningbufspace = bp->b_bufsize; 886 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 887 888 if (!TD_IS_IDLETHREAD(curthread)) 889 curthread->td_ru.ru_oublock++; 890 if (oldflags & B_ASYNC) 891 BUF_KERNPROC(bp); 892 bp->b_iooffset = dbtob(bp->b_blkno); 893 bstrategy(bp); 894 895 if ((oldflags & B_ASYNC) == 0) { 896 int rtval = bufwait(bp); 897 brelse(bp); 898 return (rtval); 899 } else { 900 /* 901 * don't allow the async write to saturate the I/O 902 * system. We will not deadlock here because 903 * we are blocking waiting for I/O that is already in-progress 904 * to complete. We do not block here if it is the update 905 * or syncer daemon trying to clean up as that can lead 906 * to deadlock. 907 */ 908 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 909 waitrunningbufspace(); 910 } 911 912 return (0); 913 } 914 915 void 916 bufbdflush(struct bufobj *bo, struct buf *bp) 917 { 918 struct buf *nbp; 919 920 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 921 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 922 altbufferflushes++; 923 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 924 BO_LOCK(bo); 925 /* 926 * Try to find a buffer to flush. 927 */ 928 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 929 if ((nbp->b_vflags & BV_BKGRDINPROG) || 930 BUF_LOCK(nbp, 931 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 932 continue; 933 if (bp == nbp) 934 panic("bdwrite: found ourselves"); 935 BO_UNLOCK(bo); 936 /* Don't countdeps with the bo lock held. */ 937 if (buf_countdeps(nbp, 0)) { 938 BO_LOCK(bo); 939 BUF_UNLOCK(nbp); 940 continue; 941 } 942 if (nbp->b_flags & B_CLUSTEROK) { 943 vfs_bio_awrite(nbp); 944 } else { 945 bremfree(nbp); 946 bawrite(nbp); 947 } 948 dirtybufferflushes++; 949 break; 950 } 951 if (nbp == NULL) 952 BO_UNLOCK(bo); 953 } 954 } 955 956 /* 957 * Delayed write. (Buffer is marked dirty). Do not bother writing 958 * anything if the buffer is marked invalid. 959 * 960 * Note that since the buffer must be completely valid, we can safely 961 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 962 * biodone() in order to prevent getblk from writing the buffer 963 * out synchronously. 964 */ 965 void 966 bdwrite(struct buf *bp) 967 { 968 struct thread *td = curthread; 969 struct vnode *vp; 970 struct bufobj *bo; 971 972 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 973 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 974 BUF_ASSERT_HELD(bp); 975 976 if (bp->b_flags & B_INVAL) { 977 brelse(bp); 978 return; 979 } 980 981 /* 982 * If we have too many dirty buffers, don't create any more. 983 * If we are wildly over our limit, then force a complete 984 * cleanup. Otherwise, just keep the situation from getting 985 * out of control. Note that we have to avoid a recursive 986 * disaster and not try to clean up after our own cleanup! 987 */ 988 vp = bp->b_vp; 989 bo = bp->b_bufobj; 990 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 991 td->td_pflags |= TDP_INBDFLUSH; 992 BO_BDFLUSH(bo, bp); 993 td->td_pflags &= ~TDP_INBDFLUSH; 994 } else 995 recursiveflushes++; 996 997 bdirty(bp); 998 /* 999 * Set B_CACHE, indicating that the buffer is fully valid. This is 1000 * true even of NFS now. 1001 */ 1002 bp->b_flags |= B_CACHE; 1003 1004 /* 1005 * This bmap keeps the system from needing to do the bmap later, 1006 * perhaps when the system is attempting to do a sync. Since it 1007 * is likely that the indirect block -- or whatever other datastructure 1008 * that the filesystem needs is still in memory now, it is a good 1009 * thing to do this. Note also, that if the pageout daemon is 1010 * requesting a sync -- there might not be enough memory to do 1011 * the bmap then... So, this is important to do. 1012 */ 1013 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1014 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1015 } 1016 1017 /* 1018 * Set the *dirty* buffer range based upon the VM system dirty pages. 1019 */ 1020 vfs_setdirty(bp); 1021 1022 /* 1023 * We need to do this here to satisfy the vnode_pager and the 1024 * pageout daemon, so that it thinks that the pages have been 1025 * "cleaned". Note that since the pages are in a delayed write 1026 * buffer -- the VFS layer "will" see that the pages get written 1027 * out on the next sync, or perhaps the cluster will be completed. 1028 */ 1029 vfs_clean_pages(bp); 1030 bqrelse(bp); 1031 1032 /* 1033 * Wakeup the buffer flushing daemon if we have a lot of dirty 1034 * buffers (midpoint between our recovery point and our stall 1035 * point). 1036 */ 1037 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1038 1039 /* 1040 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1041 * due to the softdep code. 1042 */ 1043 } 1044 1045 /* 1046 * bdirty: 1047 * 1048 * Turn buffer into delayed write request. We must clear BIO_READ and 1049 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1050 * itself to properly update it in the dirty/clean lists. We mark it 1051 * B_DONE to ensure that any asynchronization of the buffer properly 1052 * clears B_DONE ( else a panic will occur later ). 1053 * 1054 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1055 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1056 * should only be called if the buffer is known-good. 1057 * 1058 * Since the buffer is not on a queue, we do not update the numfreebuffers 1059 * count. 1060 * 1061 * The buffer must be on QUEUE_NONE. 1062 */ 1063 void 1064 bdirty(struct buf *bp) 1065 { 1066 1067 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1068 bp, bp->b_vp, bp->b_flags); 1069 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1070 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1071 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1072 BUF_ASSERT_HELD(bp); 1073 bp->b_flags &= ~(B_RELBUF); 1074 bp->b_iocmd = BIO_WRITE; 1075 1076 if ((bp->b_flags & B_DELWRI) == 0) { 1077 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1078 reassignbuf(bp); 1079 atomic_add_int(&numdirtybuffers, 1); 1080 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1081 } 1082 } 1083 1084 /* 1085 * bundirty: 1086 * 1087 * Clear B_DELWRI for buffer. 1088 * 1089 * Since the buffer is not on a queue, we do not update the numfreebuffers 1090 * count. 1091 * 1092 * The buffer must be on QUEUE_NONE. 1093 */ 1094 1095 void 1096 bundirty(struct buf *bp) 1097 { 1098 1099 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1100 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1101 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1102 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1103 BUF_ASSERT_HELD(bp); 1104 1105 if (bp->b_flags & B_DELWRI) { 1106 bp->b_flags &= ~B_DELWRI; 1107 reassignbuf(bp); 1108 atomic_subtract_int(&numdirtybuffers, 1); 1109 numdirtywakeup(lodirtybuffers); 1110 } 1111 /* 1112 * Since it is now being written, we can clear its deferred write flag. 1113 */ 1114 bp->b_flags &= ~B_DEFERRED; 1115 } 1116 1117 /* 1118 * bawrite: 1119 * 1120 * Asynchronous write. Start output on a buffer, but do not wait for 1121 * it to complete. The buffer is released when the output completes. 1122 * 1123 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1124 * B_INVAL buffers. Not us. 1125 */ 1126 void 1127 bawrite(struct buf *bp) 1128 { 1129 1130 bp->b_flags |= B_ASYNC; 1131 (void) bwrite(bp); 1132 } 1133 1134 /* 1135 * bwillwrite: 1136 * 1137 * Called prior to the locking of any vnodes when we are expecting to 1138 * write. We do not want to starve the buffer cache with too many 1139 * dirty buffers so we block here. By blocking prior to the locking 1140 * of any vnodes we attempt to avoid the situation where a locked vnode 1141 * prevents the various system daemons from flushing related buffers. 1142 */ 1143 1144 void 1145 bwillwrite(void) 1146 { 1147 1148 if (numdirtybuffers >= hidirtybuffers) { 1149 mtx_lock(&nblock); 1150 while (numdirtybuffers >= hidirtybuffers) { 1151 bd_wakeup(1); 1152 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1153 msleep(&needsbuffer, &nblock, 1154 (PRIBIO + 4), "flswai", 0); 1155 } 1156 mtx_unlock(&nblock); 1157 } 1158 } 1159 1160 /* 1161 * Return true if we have too many dirty buffers. 1162 */ 1163 int 1164 buf_dirty_count_severe(void) 1165 { 1166 1167 return(numdirtybuffers >= hidirtybuffers); 1168 } 1169 1170 /* 1171 * brelse: 1172 * 1173 * Release a busy buffer and, if requested, free its resources. The 1174 * buffer will be stashed in the appropriate bufqueue[] allowing it 1175 * to be accessed later as a cache entity or reused for other purposes. 1176 */ 1177 void 1178 brelse(struct buf *bp) 1179 { 1180 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1181 bp, bp->b_vp, bp->b_flags); 1182 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1183 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1184 1185 if (bp->b_flags & B_MANAGED) { 1186 bqrelse(bp); 1187 return; 1188 } 1189 1190 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1191 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1192 /* 1193 * Failed write, redirty. Must clear BIO_ERROR to prevent 1194 * pages from being scrapped. If the error is anything 1195 * other than an I/O error (EIO), assume that retryingi 1196 * is futile. 1197 */ 1198 bp->b_ioflags &= ~BIO_ERROR; 1199 bdirty(bp); 1200 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1201 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1202 /* 1203 * Either a failed I/O or we were asked to free or not 1204 * cache the buffer. 1205 */ 1206 bp->b_flags |= B_INVAL; 1207 if (!LIST_EMPTY(&bp->b_dep)) 1208 buf_deallocate(bp); 1209 if (bp->b_flags & B_DELWRI) { 1210 atomic_subtract_int(&numdirtybuffers, 1); 1211 numdirtywakeup(lodirtybuffers); 1212 } 1213 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1214 if ((bp->b_flags & B_VMIO) == 0) { 1215 if (bp->b_bufsize) 1216 allocbuf(bp, 0); 1217 if (bp->b_vp) 1218 brelvp(bp); 1219 } 1220 } 1221 1222 /* 1223 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1224 * is called with B_DELWRI set, the underlying pages may wind up 1225 * getting freed causing a previous write (bdwrite()) to get 'lost' 1226 * because pages associated with a B_DELWRI bp are marked clean. 1227 * 1228 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1229 * if B_DELWRI is set. 1230 * 1231 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1232 * on pages to return pages to the VM page queues. 1233 */ 1234 if (bp->b_flags & B_DELWRI) 1235 bp->b_flags &= ~B_RELBUF; 1236 else if (vm_page_count_severe()) { 1237 /* 1238 * The locking of the BO_LOCK is not necessary since 1239 * BKGRDINPROG cannot be set while we hold the buf 1240 * lock, it can only be cleared if it is already 1241 * pending. 1242 */ 1243 if (bp->b_vp) { 1244 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1245 bp->b_flags |= B_RELBUF; 1246 } else 1247 bp->b_flags |= B_RELBUF; 1248 } 1249 1250 /* 1251 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1252 * constituted, not even NFS buffers now. Two flags effect this. If 1253 * B_INVAL, the struct buf is invalidated but the VM object is kept 1254 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1255 * 1256 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1257 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1258 * buffer is also B_INVAL because it hits the re-dirtying code above. 1259 * 1260 * Normally we can do this whether a buffer is B_DELWRI or not. If 1261 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1262 * the commit state and we cannot afford to lose the buffer. If the 1263 * buffer has a background write in progress, we need to keep it 1264 * around to prevent it from being reconstituted and starting a second 1265 * background write. 1266 */ 1267 if ((bp->b_flags & B_VMIO) 1268 && !(bp->b_vp->v_mount != NULL && 1269 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1270 !vn_isdisk(bp->b_vp, NULL) && 1271 (bp->b_flags & B_DELWRI)) 1272 ) { 1273 1274 int i, j, resid; 1275 vm_page_t m; 1276 off_t foff; 1277 vm_pindex_t poff; 1278 vm_object_t obj; 1279 1280 obj = bp->b_bufobj->bo_object; 1281 1282 /* 1283 * Get the base offset and length of the buffer. Note that 1284 * in the VMIO case if the buffer block size is not 1285 * page-aligned then b_data pointer may not be page-aligned. 1286 * But our b_pages[] array *IS* page aligned. 1287 * 1288 * block sizes less then DEV_BSIZE (usually 512) are not 1289 * supported due to the page granularity bits (m->valid, 1290 * m->dirty, etc...). 1291 * 1292 * See man buf(9) for more information 1293 */ 1294 resid = bp->b_bufsize; 1295 foff = bp->b_offset; 1296 VM_OBJECT_LOCK(obj); 1297 for (i = 0; i < bp->b_npages; i++) { 1298 int had_bogus = 0; 1299 1300 m = bp->b_pages[i]; 1301 1302 /* 1303 * If we hit a bogus page, fixup *all* the bogus pages 1304 * now. 1305 */ 1306 if (m == bogus_page) { 1307 poff = OFF_TO_IDX(bp->b_offset); 1308 had_bogus = 1; 1309 1310 for (j = i; j < bp->b_npages; j++) { 1311 vm_page_t mtmp; 1312 mtmp = bp->b_pages[j]; 1313 if (mtmp == bogus_page) { 1314 mtmp = vm_page_lookup(obj, poff + j); 1315 if (!mtmp) { 1316 panic("brelse: page missing\n"); 1317 } 1318 bp->b_pages[j] = mtmp; 1319 } 1320 } 1321 1322 if ((bp->b_flags & B_INVAL) == 0) { 1323 pmap_qenter( 1324 trunc_page((vm_offset_t)bp->b_data), 1325 bp->b_pages, bp->b_npages); 1326 } 1327 m = bp->b_pages[i]; 1328 } 1329 if ((bp->b_flags & B_NOCACHE) || 1330 (bp->b_ioflags & BIO_ERROR)) { 1331 int poffset = foff & PAGE_MASK; 1332 int presid = resid > (PAGE_SIZE - poffset) ? 1333 (PAGE_SIZE - poffset) : resid; 1334 1335 KASSERT(presid >= 0, ("brelse: extra page")); 1336 vm_page_lock_queues(); 1337 vm_page_set_invalid(m, poffset, presid); 1338 vm_page_unlock_queues(); 1339 if (had_bogus) 1340 printf("avoided corruption bug in bogus_page/brelse code\n"); 1341 } 1342 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1343 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1344 } 1345 VM_OBJECT_UNLOCK(obj); 1346 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1347 vfs_vmio_release(bp); 1348 1349 } else if (bp->b_flags & B_VMIO) { 1350 1351 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1352 vfs_vmio_release(bp); 1353 } 1354 1355 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1356 if (bp->b_bufsize != 0) 1357 allocbuf(bp, 0); 1358 if (bp->b_vp != NULL) 1359 brelvp(bp); 1360 } 1361 1362 if (BUF_LOCKRECURSED(bp)) { 1363 /* do not release to free list */ 1364 BUF_UNLOCK(bp); 1365 return; 1366 } 1367 1368 /* enqueue */ 1369 mtx_lock(&bqlock); 1370 /* Handle delayed bremfree() processing. */ 1371 if (bp->b_flags & B_REMFREE) 1372 bremfreel(bp); 1373 if (bp->b_qindex != QUEUE_NONE) 1374 panic("brelse: free buffer onto another queue???"); 1375 1376 /* 1377 * If the buffer has junk contents signal it and eventually 1378 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1379 * doesn't find it. 1380 */ 1381 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1382 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1383 bp->b_flags |= B_INVAL; 1384 if (bp->b_flags & B_INVAL) { 1385 if (bp->b_flags & B_DELWRI) 1386 bundirty(bp); 1387 if (bp->b_vp) 1388 brelvp(bp); 1389 } 1390 1391 /* buffers with no memory */ 1392 if (bp->b_bufsize == 0) { 1393 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1394 if (bp->b_vflags & BV_BKGRDINPROG) 1395 panic("losing buffer 1"); 1396 if (bp->b_kvasize) { 1397 bp->b_qindex = QUEUE_EMPTYKVA; 1398 } else { 1399 bp->b_qindex = QUEUE_EMPTY; 1400 } 1401 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1402 /* buffers with junk contents */ 1403 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1404 (bp->b_ioflags & BIO_ERROR)) { 1405 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1406 if (bp->b_vflags & BV_BKGRDINPROG) 1407 panic("losing buffer 2"); 1408 bp->b_qindex = QUEUE_CLEAN; 1409 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1410 /* remaining buffers */ 1411 } else { 1412 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1413 (B_DELWRI|B_NEEDSGIANT)) 1414 bp->b_qindex = QUEUE_DIRTY_GIANT; 1415 else if (bp->b_flags & B_DELWRI) 1416 bp->b_qindex = QUEUE_DIRTY; 1417 else 1418 bp->b_qindex = QUEUE_CLEAN; 1419 if (bp->b_flags & B_AGE) 1420 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1421 else 1422 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1423 } 1424 mtx_unlock(&bqlock); 1425 1426 /* 1427 * Fixup numfreebuffers count. The bp is on an appropriate queue 1428 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1429 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1430 * if B_INVAL is set ). 1431 */ 1432 1433 if (!(bp->b_flags & B_DELWRI)) 1434 bufcountwakeup(); 1435 1436 /* 1437 * Something we can maybe free or reuse 1438 */ 1439 if (bp->b_bufsize || bp->b_kvasize) 1440 bufspacewakeup(); 1441 1442 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1443 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1444 panic("brelse: not dirty"); 1445 /* unlock */ 1446 BUF_UNLOCK(bp); 1447 } 1448 1449 /* 1450 * Release a buffer back to the appropriate queue but do not try to free 1451 * it. The buffer is expected to be used again soon. 1452 * 1453 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1454 * biodone() to requeue an async I/O on completion. It is also used when 1455 * known good buffers need to be requeued but we think we may need the data 1456 * again soon. 1457 * 1458 * XXX we should be able to leave the B_RELBUF hint set on completion. 1459 */ 1460 void 1461 bqrelse(struct buf *bp) 1462 { 1463 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1464 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1465 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1466 1467 if (BUF_LOCKRECURSED(bp)) { 1468 /* do not release to free list */ 1469 BUF_UNLOCK(bp); 1470 return; 1471 } 1472 1473 if (bp->b_flags & B_MANAGED) { 1474 if (bp->b_flags & B_REMFREE) { 1475 mtx_lock(&bqlock); 1476 bremfreel(bp); 1477 mtx_unlock(&bqlock); 1478 } 1479 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1480 BUF_UNLOCK(bp); 1481 return; 1482 } 1483 1484 mtx_lock(&bqlock); 1485 /* Handle delayed bremfree() processing. */ 1486 if (bp->b_flags & B_REMFREE) 1487 bremfreel(bp); 1488 if (bp->b_qindex != QUEUE_NONE) 1489 panic("bqrelse: free buffer onto another queue???"); 1490 /* buffers with stale but valid contents */ 1491 if (bp->b_flags & B_DELWRI) { 1492 if (bp->b_flags & B_NEEDSGIANT) 1493 bp->b_qindex = QUEUE_DIRTY_GIANT; 1494 else 1495 bp->b_qindex = QUEUE_DIRTY; 1496 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1497 } else { 1498 /* 1499 * The locking of the BO_LOCK for checking of the 1500 * BV_BKGRDINPROG is not necessary since the 1501 * BV_BKGRDINPROG cannot be set while we hold the buf 1502 * lock, it can only be cleared if it is already 1503 * pending. 1504 */ 1505 if (!vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1506 bp->b_qindex = QUEUE_CLEAN; 1507 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1508 b_freelist); 1509 } else { 1510 /* 1511 * We are too low on memory, we have to try to free 1512 * the buffer (most importantly: the wired pages 1513 * making up its backing store) *now*. 1514 */ 1515 mtx_unlock(&bqlock); 1516 brelse(bp); 1517 return; 1518 } 1519 } 1520 mtx_unlock(&bqlock); 1521 1522 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1523 bufcountwakeup(); 1524 1525 /* 1526 * Something we can maybe free or reuse. 1527 */ 1528 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1529 bufspacewakeup(); 1530 1531 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1532 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1533 panic("bqrelse: not dirty"); 1534 /* unlock */ 1535 BUF_UNLOCK(bp); 1536 } 1537 1538 /* Give pages used by the bp back to the VM system (where possible) */ 1539 static void 1540 vfs_vmio_release(struct buf *bp) 1541 { 1542 int i; 1543 vm_page_t m; 1544 1545 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1546 vm_page_lock_queues(); 1547 for (i = 0; i < bp->b_npages; i++) { 1548 m = bp->b_pages[i]; 1549 bp->b_pages[i] = NULL; 1550 /* 1551 * In order to keep page LRU ordering consistent, put 1552 * everything on the inactive queue. 1553 */ 1554 vm_page_unwire(m, 0); 1555 /* 1556 * We don't mess with busy pages, it is 1557 * the responsibility of the process that 1558 * busied the pages to deal with them. 1559 */ 1560 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1561 continue; 1562 1563 if (m->wire_count == 0) { 1564 /* 1565 * Might as well free the page if we can and it has 1566 * no valid data. We also free the page if the 1567 * buffer was used for direct I/O 1568 */ 1569 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1570 m->hold_count == 0) { 1571 vm_page_free(m); 1572 } else if (bp->b_flags & B_DIRECT) { 1573 vm_page_try_to_free(m); 1574 } else if (vm_page_count_severe()) { 1575 vm_page_try_to_cache(m); 1576 } 1577 } 1578 } 1579 vm_page_unlock_queues(); 1580 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1581 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1582 1583 if (bp->b_bufsize) { 1584 bufspacewakeup(); 1585 bp->b_bufsize = 0; 1586 } 1587 bp->b_npages = 0; 1588 bp->b_flags &= ~B_VMIO; 1589 if (bp->b_vp) 1590 brelvp(bp); 1591 } 1592 1593 /* 1594 * Check to see if a block at a particular lbn is available for a clustered 1595 * write. 1596 */ 1597 static int 1598 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1599 { 1600 struct buf *bpa; 1601 int match; 1602 1603 match = 0; 1604 1605 /* If the buf isn't in core skip it */ 1606 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1607 return (0); 1608 1609 /* If the buf is busy we don't want to wait for it */ 1610 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1611 return (0); 1612 1613 /* Only cluster with valid clusterable delayed write buffers */ 1614 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1615 (B_DELWRI | B_CLUSTEROK)) 1616 goto done; 1617 1618 if (bpa->b_bufsize != size) 1619 goto done; 1620 1621 /* 1622 * Check to see if it is in the expected place on disk and that the 1623 * block has been mapped. 1624 */ 1625 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1626 match = 1; 1627 done: 1628 BUF_UNLOCK(bpa); 1629 return (match); 1630 } 1631 1632 /* 1633 * vfs_bio_awrite: 1634 * 1635 * Implement clustered async writes for clearing out B_DELWRI buffers. 1636 * This is much better then the old way of writing only one buffer at 1637 * a time. Note that we may not be presented with the buffers in the 1638 * correct order, so we search for the cluster in both directions. 1639 */ 1640 int 1641 vfs_bio_awrite(struct buf *bp) 1642 { 1643 struct bufobj *bo; 1644 int i; 1645 int j; 1646 daddr_t lblkno = bp->b_lblkno; 1647 struct vnode *vp = bp->b_vp; 1648 int ncl; 1649 int nwritten; 1650 int size; 1651 int maxcl; 1652 1653 bo = &vp->v_bufobj; 1654 /* 1655 * right now we support clustered writing only to regular files. If 1656 * we find a clusterable block we could be in the middle of a cluster 1657 * rather then at the beginning. 1658 */ 1659 if ((vp->v_type == VREG) && 1660 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1661 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1662 1663 size = vp->v_mount->mnt_stat.f_iosize; 1664 maxcl = MAXPHYS / size; 1665 1666 BO_LOCK(bo); 1667 for (i = 1; i < maxcl; i++) 1668 if (vfs_bio_clcheck(vp, size, lblkno + i, 1669 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1670 break; 1671 1672 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1673 if (vfs_bio_clcheck(vp, size, lblkno - j, 1674 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1675 break; 1676 BO_UNLOCK(bo); 1677 --j; 1678 ncl = i + j; 1679 /* 1680 * this is a possible cluster write 1681 */ 1682 if (ncl != 1) { 1683 BUF_UNLOCK(bp); 1684 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1685 return nwritten; 1686 } 1687 } 1688 bremfree(bp); 1689 bp->b_flags |= B_ASYNC; 1690 /* 1691 * default (old) behavior, writing out only one block 1692 * 1693 * XXX returns b_bufsize instead of b_bcount for nwritten? 1694 */ 1695 nwritten = bp->b_bufsize; 1696 (void) bwrite(bp); 1697 1698 return nwritten; 1699 } 1700 1701 /* 1702 * getnewbuf: 1703 * 1704 * Find and initialize a new buffer header, freeing up existing buffers 1705 * in the bufqueues as necessary. The new buffer is returned locked. 1706 * 1707 * Important: B_INVAL is not set. If the caller wishes to throw the 1708 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1709 * 1710 * We block if: 1711 * We have insufficient buffer headers 1712 * We have insufficient buffer space 1713 * buffer_map is too fragmented ( space reservation fails ) 1714 * If we have to flush dirty buffers ( but we try to avoid this ) 1715 * 1716 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1717 * Instead we ask the buf daemon to do it for us. We attempt to 1718 * avoid piecemeal wakeups of the pageout daemon. 1719 */ 1720 1721 static struct buf * 1722 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1723 int gbflags) 1724 { 1725 struct thread *td; 1726 struct buf *bp; 1727 struct buf *nbp; 1728 int defrag = 0; 1729 int nqindex; 1730 static int flushingbufs; 1731 1732 td = curthread; 1733 /* 1734 * We can't afford to block since we might be holding a vnode lock, 1735 * which may prevent system daemons from running. We deal with 1736 * low-memory situations by proactively returning memory and running 1737 * async I/O rather then sync I/O. 1738 */ 1739 atomic_add_int(&getnewbufcalls, 1); 1740 atomic_subtract_int(&getnewbufrestarts, 1); 1741 restart: 1742 atomic_add_int(&getnewbufrestarts, 1); 1743 1744 /* 1745 * Setup for scan. If we do not have enough free buffers, 1746 * we setup a degenerate case that immediately fails. Note 1747 * that if we are specially marked process, we are allowed to 1748 * dip into our reserves. 1749 * 1750 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1751 * 1752 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1753 * However, there are a number of cases (defragging, reusing, ...) 1754 * where we cannot backup. 1755 */ 1756 mtx_lock(&bqlock); 1757 nqindex = QUEUE_EMPTYKVA; 1758 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1759 1760 if (nbp == NULL) { 1761 /* 1762 * If no EMPTYKVA buffers and we are either 1763 * defragging or reusing, locate a CLEAN buffer 1764 * to free or reuse. If bufspace useage is low 1765 * skip this step so we can allocate a new buffer. 1766 */ 1767 if (defrag || bufspace >= lobufspace) { 1768 nqindex = QUEUE_CLEAN; 1769 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1770 } 1771 1772 /* 1773 * If we could not find or were not allowed to reuse a 1774 * CLEAN buffer, check to see if it is ok to use an EMPTY 1775 * buffer. We can only use an EMPTY buffer if allocating 1776 * its KVA would not otherwise run us out of buffer space. 1777 */ 1778 if (nbp == NULL && defrag == 0 && 1779 bufspace + maxsize < hibufspace) { 1780 nqindex = QUEUE_EMPTY; 1781 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1782 } 1783 } 1784 1785 /* 1786 * Run scan, possibly freeing data and/or kva mappings on the fly 1787 * depending. 1788 */ 1789 1790 while ((bp = nbp) != NULL) { 1791 int qindex = nqindex; 1792 1793 /* 1794 * Calculate next bp ( we can only use it if we do not block 1795 * or do other fancy things ). 1796 */ 1797 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1798 switch(qindex) { 1799 case QUEUE_EMPTY: 1800 nqindex = QUEUE_EMPTYKVA; 1801 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1802 break; 1803 /* FALLTHROUGH */ 1804 case QUEUE_EMPTYKVA: 1805 nqindex = QUEUE_CLEAN; 1806 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1807 break; 1808 /* FALLTHROUGH */ 1809 case QUEUE_CLEAN: 1810 /* 1811 * nbp is NULL. 1812 */ 1813 break; 1814 } 1815 } 1816 /* 1817 * If we are defragging then we need a buffer with 1818 * b_kvasize != 0. XXX this situation should no longer 1819 * occur, if defrag is non-zero the buffer's b_kvasize 1820 * should also be non-zero at this point. XXX 1821 */ 1822 if (defrag && bp->b_kvasize == 0) { 1823 printf("Warning: defrag empty buffer %p\n", bp); 1824 continue; 1825 } 1826 1827 /* 1828 * Start freeing the bp. This is somewhat involved. nbp 1829 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1830 */ 1831 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1832 continue; 1833 if (bp->b_vp) { 1834 BO_LOCK(bp->b_bufobj); 1835 if (bp->b_vflags & BV_BKGRDINPROG) { 1836 BO_UNLOCK(bp->b_bufobj); 1837 BUF_UNLOCK(bp); 1838 continue; 1839 } 1840 BO_UNLOCK(bp->b_bufobj); 1841 } 1842 CTR6(KTR_BUF, 1843 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1844 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1845 bp->b_kvasize, bp->b_bufsize, qindex); 1846 1847 /* 1848 * Sanity Checks 1849 */ 1850 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1851 1852 /* 1853 * Note: we no longer distinguish between VMIO and non-VMIO 1854 * buffers. 1855 */ 1856 1857 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1858 1859 bremfreel(bp); 1860 mtx_unlock(&bqlock); 1861 1862 if (qindex == QUEUE_CLEAN) { 1863 if (bp->b_flags & B_VMIO) { 1864 bp->b_flags &= ~B_ASYNC; 1865 vfs_vmio_release(bp); 1866 } 1867 if (bp->b_vp) 1868 brelvp(bp); 1869 } 1870 1871 /* 1872 * NOTE: nbp is now entirely invalid. We can only restart 1873 * the scan from this point on. 1874 * 1875 * Get the rest of the buffer freed up. b_kva* is still 1876 * valid after this operation. 1877 */ 1878 1879 if (bp->b_rcred != NOCRED) { 1880 crfree(bp->b_rcred); 1881 bp->b_rcred = NOCRED; 1882 } 1883 if (bp->b_wcred != NOCRED) { 1884 crfree(bp->b_wcred); 1885 bp->b_wcred = NOCRED; 1886 } 1887 if (!LIST_EMPTY(&bp->b_dep)) 1888 buf_deallocate(bp); 1889 if (bp->b_vflags & BV_BKGRDINPROG) 1890 panic("losing buffer 3"); 1891 KASSERT(bp->b_vp == NULL, 1892 ("bp: %p still has vnode %p. qindex: %d", 1893 bp, bp->b_vp, qindex)); 1894 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1895 ("bp: %p still on a buffer list. xflags %X", 1896 bp, bp->b_xflags)); 1897 1898 if (bp->b_bufsize) 1899 allocbuf(bp, 0); 1900 1901 bp->b_flags = 0; 1902 bp->b_ioflags = 0; 1903 bp->b_xflags = 0; 1904 bp->b_vflags = 0; 1905 bp->b_vp = NULL; 1906 bp->b_blkno = bp->b_lblkno = 0; 1907 bp->b_offset = NOOFFSET; 1908 bp->b_iodone = 0; 1909 bp->b_error = 0; 1910 bp->b_resid = 0; 1911 bp->b_bcount = 0; 1912 bp->b_npages = 0; 1913 bp->b_dirtyoff = bp->b_dirtyend = 0; 1914 bp->b_bufobj = NULL; 1915 bp->b_pin_count = 0; 1916 bp->b_fsprivate1 = NULL; 1917 bp->b_fsprivate2 = NULL; 1918 bp->b_fsprivate3 = NULL; 1919 1920 LIST_INIT(&bp->b_dep); 1921 1922 /* 1923 * If we are defragging then free the buffer. 1924 */ 1925 if (defrag) { 1926 bp->b_flags |= B_INVAL; 1927 bfreekva(bp); 1928 brelse(bp); 1929 defrag = 0; 1930 goto restart; 1931 } 1932 1933 /* 1934 * Notify any waiters for the buffer lock about 1935 * identity change by freeing the buffer. 1936 */ 1937 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1938 bp->b_flags |= B_INVAL; 1939 bfreekva(bp); 1940 brelse(bp); 1941 goto restart; 1942 } 1943 1944 /* 1945 * If we are overcomitted then recover the buffer and its 1946 * KVM space. This occurs in rare situations when multiple 1947 * processes are blocked in getnewbuf() or allocbuf(). 1948 */ 1949 if (bufspace >= hibufspace) 1950 flushingbufs = 1; 1951 if (flushingbufs && bp->b_kvasize != 0) { 1952 bp->b_flags |= B_INVAL; 1953 bfreekva(bp); 1954 brelse(bp); 1955 goto restart; 1956 } 1957 if (bufspace < lobufspace) 1958 flushingbufs = 0; 1959 break; 1960 } 1961 1962 /* 1963 * If we exhausted our list, sleep as appropriate. We may have to 1964 * wakeup various daemons and write out some dirty buffers. 1965 * 1966 * Generally we are sleeping due to insufficient buffer space. 1967 */ 1968 1969 if (bp == NULL) { 1970 int flags, norunbuf; 1971 char *waitmsg; 1972 int fl; 1973 1974 if (defrag) { 1975 flags = VFS_BIO_NEED_BUFSPACE; 1976 waitmsg = "nbufkv"; 1977 } else if (bufspace >= hibufspace) { 1978 waitmsg = "nbufbs"; 1979 flags = VFS_BIO_NEED_BUFSPACE; 1980 } else { 1981 waitmsg = "newbuf"; 1982 flags = VFS_BIO_NEED_ANY; 1983 } 1984 mtx_lock(&nblock); 1985 needsbuffer |= flags; 1986 mtx_unlock(&nblock); 1987 mtx_unlock(&bqlock); 1988 1989 bd_speedup(); /* heeeelp */ 1990 if (gbflags & GB_NOWAIT_BD) 1991 return (NULL); 1992 1993 mtx_lock(&nblock); 1994 while (needsbuffer & flags) { 1995 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 1996 mtx_unlock(&nblock); 1997 /* 1998 * getblk() is called with a vnode 1999 * locked, and some majority of the 2000 * dirty buffers may as well belong to 2001 * the vnode. Flushing the buffers 2002 * there would make a progress that 2003 * cannot be achieved by the 2004 * buf_daemon, that cannot lock the 2005 * vnode. 2006 */ 2007 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2008 (td->td_pflags & TDP_NORUNNINGBUF); 2009 /* play bufdaemon */ 2010 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2011 fl = buf_do_flush(vp); 2012 td->td_pflags &= norunbuf; 2013 mtx_lock(&nblock); 2014 if (fl != 0) 2015 continue; 2016 if ((needsbuffer & flags) == 0) 2017 break; 2018 } 2019 if (msleep(&needsbuffer, &nblock, 2020 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2021 mtx_unlock(&nblock); 2022 return (NULL); 2023 } 2024 } 2025 mtx_unlock(&nblock); 2026 } else { 2027 /* 2028 * We finally have a valid bp. We aren't quite out of the 2029 * woods, we still have to reserve kva space. In order 2030 * to keep fragmentation sane we only allocate kva in 2031 * BKVASIZE chunks. 2032 */ 2033 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2034 2035 if (maxsize != bp->b_kvasize) { 2036 vm_offset_t addr = 0; 2037 2038 bfreekva(bp); 2039 2040 vm_map_lock(buffer_map); 2041 if (vm_map_findspace(buffer_map, 2042 vm_map_min(buffer_map), maxsize, &addr)) { 2043 /* 2044 * Uh oh. Buffer map is to fragmented. We 2045 * must defragment the map. 2046 */ 2047 atomic_add_int(&bufdefragcnt, 1); 2048 vm_map_unlock(buffer_map); 2049 defrag = 1; 2050 bp->b_flags |= B_INVAL; 2051 brelse(bp); 2052 goto restart; 2053 } 2054 if (addr) { 2055 vm_map_insert(buffer_map, NULL, 0, 2056 addr, addr + maxsize, 2057 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2058 2059 bp->b_kvabase = (caddr_t) addr; 2060 bp->b_kvasize = maxsize; 2061 atomic_add_long(&bufspace, bp->b_kvasize); 2062 atomic_add_int(&bufreusecnt, 1); 2063 } 2064 vm_map_unlock(buffer_map); 2065 } 2066 bp->b_saveaddr = bp->b_kvabase; 2067 bp->b_data = bp->b_saveaddr; 2068 } 2069 return(bp); 2070 } 2071 2072 /* 2073 * buf_daemon: 2074 * 2075 * buffer flushing daemon. Buffers are normally flushed by the 2076 * update daemon but if it cannot keep up this process starts to 2077 * take the load in an attempt to prevent getnewbuf() from blocking. 2078 */ 2079 2080 static struct kproc_desc buf_kp = { 2081 "bufdaemon", 2082 buf_daemon, 2083 &bufdaemonproc 2084 }; 2085 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2086 2087 static int 2088 buf_do_flush(struct vnode *vp) 2089 { 2090 int flushed; 2091 2092 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2093 /* The list empty check here is slightly racy */ 2094 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2095 mtx_lock(&Giant); 2096 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2097 mtx_unlock(&Giant); 2098 } 2099 if (flushed == 0) { 2100 /* 2101 * Could not find any buffers without rollback 2102 * dependencies, so just write the first one 2103 * in the hopes of eventually making progress. 2104 */ 2105 flushbufqueues(vp, QUEUE_DIRTY, 1); 2106 if (!TAILQ_EMPTY( 2107 &bufqueues[QUEUE_DIRTY_GIANT])) { 2108 mtx_lock(&Giant); 2109 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2110 mtx_unlock(&Giant); 2111 } 2112 } 2113 return (flushed); 2114 } 2115 2116 static void 2117 buf_daemon() 2118 { 2119 2120 /* 2121 * This process needs to be suspended prior to shutdown sync. 2122 */ 2123 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2124 SHUTDOWN_PRI_LAST); 2125 2126 /* 2127 * This process is allowed to take the buffer cache to the limit 2128 */ 2129 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2130 mtx_lock(&bdlock); 2131 for (;;) { 2132 bd_request = 0; 2133 mtx_unlock(&bdlock); 2134 2135 kproc_suspend_check(bufdaemonproc); 2136 2137 /* 2138 * Do the flush. Limit the amount of in-transit I/O we 2139 * allow to build up, otherwise we would completely saturate 2140 * the I/O system. Wakeup any waiting processes before we 2141 * normally would so they can run in parallel with our drain. 2142 */ 2143 while (numdirtybuffers > lodirtybuffers) { 2144 if (buf_do_flush(NULL) == 0) 2145 break; 2146 uio_yield(); 2147 } 2148 2149 /* 2150 * Only clear bd_request if we have reached our low water 2151 * mark. The buf_daemon normally waits 1 second and 2152 * then incrementally flushes any dirty buffers that have 2153 * built up, within reason. 2154 * 2155 * If we were unable to hit our low water mark and couldn't 2156 * find any flushable buffers, we sleep half a second. 2157 * Otherwise we loop immediately. 2158 */ 2159 mtx_lock(&bdlock); 2160 if (numdirtybuffers <= lodirtybuffers) { 2161 /* 2162 * We reached our low water mark, reset the 2163 * request and sleep until we are needed again. 2164 * The sleep is just so the suspend code works. 2165 */ 2166 bd_request = 0; 2167 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2168 } else { 2169 /* 2170 * We couldn't find any flushable dirty buffers but 2171 * still have too many dirty buffers, we 2172 * have to sleep and try again. (rare) 2173 */ 2174 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2175 } 2176 } 2177 } 2178 2179 /* 2180 * flushbufqueues: 2181 * 2182 * Try to flush a buffer in the dirty queue. We must be careful to 2183 * free up B_INVAL buffers instead of write them, which NFS is 2184 * particularly sensitive to. 2185 */ 2186 static int flushwithdeps = 0; 2187 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2188 0, "Number of buffers flushed with dependecies that require rollbacks"); 2189 2190 static int 2191 flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2192 { 2193 struct buf *sentinel; 2194 struct vnode *vp; 2195 struct mount *mp; 2196 struct buf *bp; 2197 int hasdeps; 2198 int flushed; 2199 int target; 2200 2201 if (lvp == NULL) { 2202 target = numdirtybuffers - lodirtybuffers; 2203 if (flushdeps && target > 2) 2204 target /= 2; 2205 } else 2206 target = flushbufqtarget; 2207 flushed = 0; 2208 bp = NULL; 2209 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2210 sentinel->b_qindex = QUEUE_SENTINEL; 2211 mtx_lock(&bqlock); 2212 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2213 while (flushed != target) { 2214 bp = TAILQ_NEXT(sentinel, b_freelist); 2215 if (bp != NULL) { 2216 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2217 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2218 b_freelist); 2219 } else 2220 break; 2221 /* 2222 * Skip sentinels inserted by other invocations of the 2223 * flushbufqueues(), taking care to not reorder them. 2224 */ 2225 if (bp->b_qindex == QUEUE_SENTINEL) 2226 continue; 2227 /* 2228 * Only flush the buffers that belong to the 2229 * vnode locked by the curthread. 2230 */ 2231 if (lvp != NULL && bp->b_vp != lvp) 2232 continue; 2233 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2234 continue; 2235 if (bp->b_pin_count > 0) { 2236 BUF_UNLOCK(bp); 2237 continue; 2238 } 2239 BO_LOCK(bp->b_bufobj); 2240 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2241 (bp->b_flags & B_DELWRI) == 0) { 2242 BO_UNLOCK(bp->b_bufobj); 2243 BUF_UNLOCK(bp); 2244 continue; 2245 } 2246 BO_UNLOCK(bp->b_bufobj); 2247 if (bp->b_flags & B_INVAL) { 2248 bremfreel(bp); 2249 mtx_unlock(&bqlock); 2250 brelse(bp); 2251 flushed++; 2252 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2253 mtx_lock(&bqlock); 2254 continue; 2255 } 2256 2257 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2258 if (flushdeps == 0) { 2259 BUF_UNLOCK(bp); 2260 continue; 2261 } 2262 hasdeps = 1; 2263 } else 2264 hasdeps = 0; 2265 /* 2266 * We must hold the lock on a vnode before writing 2267 * one of its buffers. Otherwise we may confuse, or 2268 * in the case of a snapshot vnode, deadlock the 2269 * system. 2270 * 2271 * The lock order here is the reverse of the normal 2272 * of vnode followed by buf lock. This is ok because 2273 * the NOWAIT will prevent deadlock. 2274 */ 2275 vp = bp->b_vp; 2276 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2277 BUF_UNLOCK(bp); 2278 continue; 2279 } 2280 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2281 mtx_unlock(&bqlock); 2282 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2283 bp, bp->b_vp, bp->b_flags); 2284 if (curproc == bufdaemonproc) 2285 vfs_bio_awrite(bp); 2286 else { 2287 bremfree(bp); 2288 bwrite(bp); 2289 notbufdflashes++; 2290 } 2291 vn_finished_write(mp); 2292 VOP_UNLOCK(vp, 0); 2293 flushwithdeps += hasdeps; 2294 flushed++; 2295 2296 /* 2297 * Sleeping on runningbufspace while holding 2298 * vnode lock leads to deadlock. 2299 */ 2300 if (curproc == bufdaemonproc) 2301 waitrunningbufspace(); 2302 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2303 mtx_lock(&bqlock); 2304 continue; 2305 } 2306 vn_finished_write(mp); 2307 BUF_UNLOCK(bp); 2308 } 2309 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2310 mtx_unlock(&bqlock); 2311 free(sentinel, M_TEMP); 2312 return (flushed); 2313 } 2314 2315 /* 2316 * Check to see if a block is currently memory resident. 2317 */ 2318 struct buf * 2319 incore(struct bufobj *bo, daddr_t blkno) 2320 { 2321 struct buf *bp; 2322 2323 BO_LOCK(bo); 2324 bp = gbincore(bo, blkno); 2325 BO_UNLOCK(bo); 2326 return (bp); 2327 } 2328 2329 /* 2330 * Returns true if no I/O is needed to access the 2331 * associated VM object. This is like incore except 2332 * it also hunts around in the VM system for the data. 2333 */ 2334 2335 static int 2336 inmem(struct vnode * vp, daddr_t blkno) 2337 { 2338 vm_object_t obj; 2339 vm_offset_t toff, tinc, size; 2340 vm_page_t m; 2341 vm_ooffset_t off; 2342 2343 ASSERT_VOP_LOCKED(vp, "inmem"); 2344 2345 if (incore(&vp->v_bufobj, blkno)) 2346 return 1; 2347 if (vp->v_mount == NULL) 2348 return 0; 2349 obj = vp->v_object; 2350 if (obj == NULL) 2351 return (0); 2352 2353 size = PAGE_SIZE; 2354 if (size > vp->v_mount->mnt_stat.f_iosize) 2355 size = vp->v_mount->mnt_stat.f_iosize; 2356 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2357 2358 VM_OBJECT_LOCK(obj); 2359 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2360 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2361 if (!m) 2362 goto notinmem; 2363 tinc = size; 2364 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2365 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2366 if (vm_page_is_valid(m, 2367 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2368 goto notinmem; 2369 } 2370 VM_OBJECT_UNLOCK(obj); 2371 return 1; 2372 2373 notinmem: 2374 VM_OBJECT_UNLOCK(obj); 2375 return (0); 2376 } 2377 2378 /* 2379 * vfs_setdirty: 2380 * 2381 * Sets the dirty range for a buffer based on the status of the dirty 2382 * bits in the pages comprising the buffer. 2383 * 2384 * The range is limited to the size of the buffer. 2385 * 2386 * This routine is primarily used by NFS, but is generalized for the 2387 * B_VMIO case. 2388 */ 2389 static void 2390 vfs_setdirty(struct buf *bp) 2391 { 2392 2393 /* 2394 * Degenerate case - empty buffer 2395 */ 2396 2397 if (bp->b_bufsize == 0) 2398 return; 2399 2400 /* 2401 * We qualify the scan for modified pages on whether the 2402 * object has been flushed yet. 2403 */ 2404 2405 if ((bp->b_flags & B_VMIO) == 0) 2406 return; 2407 2408 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2409 vfs_setdirty_locked_object(bp); 2410 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2411 } 2412 2413 static void 2414 vfs_setdirty_locked_object(struct buf *bp) 2415 { 2416 vm_object_t object; 2417 int i; 2418 2419 object = bp->b_bufobj->bo_object; 2420 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2421 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2422 vm_offset_t boffset; 2423 vm_offset_t eoffset; 2424 2425 vm_page_lock_queues(); 2426 /* 2427 * test the pages to see if they have been modified directly 2428 * by users through the VM system. 2429 */ 2430 for (i = 0; i < bp->b_npages; i++) 2431 vm_page_test_dirty(bp->b_pages[i]); 2432 2433 /* 2434 * Calculate the encompassing dirty range, boffset and eoffset, 2435 * (eoffset - boffset) bytes. 2436 */ 2437 2438 for (i = 0; i < bp->b_npages; i++) { 2439 if (bp->b_pages[i]->dirty) 2440 break; 2441 } 2442 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2443 2444 for (i = bp->b_npages - 1; i >= 0; --i) { 2445 if (bp->b_pages[i]->dirty) { 2446 break; 2447 } 2448 } 2449 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2450 2451 vm_page_unlock_queues(); 2452 /* 2453 * Fit it to the buffer. 2454 */ 2455 2456 if (eoffset > bp->b_bcount) 2457 eoffset = bp->b_bcount; 2458 2459 /* 2460 * If we have a good dirty range, merge with the existing 2461 * dirty range. 2462 */ 2463 2464 if (boffset < eoffset) { 2465 if (bp->b_dirtyoff > boffset) 2466 bp->b_dirtyoff = boffset; 2467 if (bp->b_dirtyend < eoffset) 2468 bp->b_dirtyend = eoffset; 2469 } 2470 } 2471 } 2472 2473 /* 2474 * getblk: 2475 * 2476 * Get a block given a specified block and offset into a file/device. 2477 * The buffers B_DONE bit will be cleared on return, making it almost 2478 * ready for an I/O initiation. B_INVAL may or may not be set on 2479 * return. The caller should clear B_INVAL prior to initiating a 2480 * READ. 2481 * 2482 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2483 * an existing buffer. 2484 * 2485 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2486 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2487 * and then cleared based on the backing VM. If the previous buffer is 2488 * non-0-sized but invalid, B_CACHE will be cleared. 2489 * 2490 * If getblk() must create a new buffer, the new buffer is returned with 2491 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2492 * case it is returned with B_INVAL clear and B_CACHE set based on the 2493 * backing VM. 2494 * 2495 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2496 * B_CACHE bit is clear. 2497 * 2498 * What this means, basically, is that the caller should use B_CACHE to 2499 * determine whether the buffer is fully valid or not and should clear 2500 * B_INVAL prior to issuing a read. If the caller intends to validate 2501 * the buffer by loading its data area with something, the caller needs 2502 * to clear B_INVAL. If the caller does this without issuing an I/O, 2503 * the caller should set B_CACHE ( as an optimization ), else the caller 2504 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2505 * a write attempt or if it was a successfull read. If the caller 2506 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2507 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2508 */ 2509 struct buf * 2510 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2511 int flags) 2512 { 2513 struct buf *bp; 2514 struct bufobj *bo; 2515 int error; 2516 2517 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2518 ASSERT_VOP_LOCKED(vp, "getblk"); 2519 if (size > MAXBSIZE) 2520 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2521 2522 bo = &vp->v_bufobj; 2523 loop: 2524 /* 2525 * Block if we are low on buffers. Certain processes are allowed 2526 * to completely exhaust the buffer cache. 2527 * 2528 * If this check ever becomes a bottleneck it may be better to 2529 * move it into the else, when gbincore() fails. At the moment 2530 * it isn't a problem. 2531 * 2532 * XXX remove if 0 sections (clean this up after its proven) 2533 */ 2534 if (numfreebuffers == 0) { 2535 if (TD_IS_IDLETHREAD(curthread)) 2536 return NULL; 2537 mtx_lock(&nblock); 2538 needsbuffer |= VFS_BIO_NEED_ANY; 2539 mtx_unlock(&nblock); 2540 } 2541 2542 BO_LOCK(bo); 2543 bp = gbincore(bo, blkno); 2544 if (bp != NULL) { 2545 int lockflags; 2546 /* 2547 * Buffer is in-core. If the buffer is not busy, it must 2548 * be on a queue. 2549 */ 2550 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2551 2552 if (flags & GB_LOCK_NOWAIT) 2553 lockflags |= LK_NOWAIT; 2554 2555 error = BUF_TIMELOCK(bp, lockflags, 2556 BO_MTX(bo), "getblk", slpflag, slptimeo); 2557 2558 /* 2559 * If we slept and got the lock we have to restart in case 2560 * the buffer changed identities. 2561 */ 2562 if (error == ENOLCK) 2563 goto loop; 2564 /* We timed out or were interrupted. */ 2565 else if (error) 2566 return (NULL); 2567 2568 /* 2569 * The buffer is locked. B_CACHE is cleared if the buffer is 2570 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2571 * and for a VMIO buffer B_CACHE is adjusted according to the 2572 * backing VM cache. 2573 */ 2574 if (bp->b_flags & B_INVAL) 2575 bp->b_flags &= ~B_CACHE; 2576 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2577 bp->b_flags |= B_CACHE; 2578 bremfree(bp); 2579 2580 /* 2581 * check for size inconsistancies for non-VMIO case. 2582 */ 2583 2584 if (bp->b_bcount != size) { 2585 if ((bp->b_flags & B_VMIO) == 0 || 2586 (size > bp->b_kvasize)) { 2587 if (bp->b_flags & B_DELWRI) { 2588 /* 2589 * If buffer is pinned and caller does 2590 * not want sleep waiting for it to be 2591 * unpinned, bail out 2592 * */ 2593 if (bp->b_pin_count > 0) { 2594 if (flags & GB_LOCK_NOWAIT) { 2595 bqrelse(bp); 2596 return (NULL); 2597 } else { 2598 bunpin_wait(bp); 2599 } 2600 } 2601 bp->b_flags |= B_NOCACHE; 2602 bwrite(bp); 2603 } else { 2604 if (LIST_EMPTY(&bp->b_dep)) { 2605 bp->b_flags |= B_RELBUF; 2606 brelse(bp); 2607 } else { 2608 bp->b_flags |= B_NOCACHE; 2609 bwrite(bp); 2610 } 2611 } 2612 goto loop; 2613 } 2614 } 2615 2616 /* 2617 * If the size is inconsistant in the VMIO case, we can resize 2618 * the buffer. This might lead to B_CACHE getting set or 2619 * cleared. If the size has not changed, B_CACHE remains 2620 * unchanged from its previous state. 2621 */ 2622 2623 if (bp->b_bcount != size) 2624 allocbuf(bp, size); 2625 2626 KASSERT(bp->b_offset != NOOFFSET, 2627 ("getblk: no buffer offset")); 2628 2629 /* 2630 * A buffer with B_DELWRI set and B_CACHE clear must 2631 * be committed before we can return the buffer in 2632 * order to prevent the caller from issuing a read 2633 * ( due to B_CACHE not being set ) and overwriting 2634 * it. 2635 * 2636 * Most callers, including NFS and FFS, need this to 2637 * operate properly either because they assume they 2638 * can issue a read if B_CACHE is not set, or because 2639 * ( for example ) an uncached B_DELWRI might loop due 2640 * to softupdates re-dirtying the buffer. In the latter 2641 * case, B_CACHE is set after the first write completes, 2642 * preventing further loops. 2643 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2644 * above while extending the buffer, we cannot allow the 2645 * buffer to remain with B_CACHE set after the write 2646 * completes or it will represent a corrupt state. To 2647 * deal with this we set B_NOCACHE to scrap the buffer 2648 * after the write. 2649 * 2650 * We might be able to do something fancy, like setting 2651 * B_CACHE in bwrite() except if B_DELWRI is already set, 2652 * so the below call doesn't set B_CACHE, but that gets real 2653 * confusing. This is much easier. 2654 */ 2655 2656 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2657 bp->b_flags |= B_NOCACHE; 2658 bwrite(bp); 2659 goto loop; 2660 } 2661 bp->b_flags &= ~B_DONE; 2662 } else { 2663 int bsize, maxsize, vmio; 2664 off_t offset; 2665 2666 /* 2667 * Buffer is not in-core, create new buffer. The buffer 2668 * returned by getnewbuf() is locked. Note that the returned 2669 * buffer is also considered valid (not marked B_INVAL). 2670 */ 2671 BO_UNLOCK(bo); 2672 /* 2673 * If the user does not want us to create the buffer, bail out 2674 * here. 2675 */ 2676 if (flags & GB_NOCREAT) 2677 return NULL; 2678 bsize = bo->bo_bsize; 2679 offset = blkno * bsize; 2680 vmio = vp->v_object != NULL; 2681 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2682 maxsize = imax(maxsize, bsize); 2683 2684 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2685 if (bp == NULL) { 2686 if (slpflag || slptimeo) 2687 return NULL; 2688 goto loop; 2689 } 2690 2691 /* 2692 * This code is used to make sure that a buffer is not 2693 * created while the getnewbuf routine is blocked. 2694 * This can be a problem whether the vnode is locked or not. 2695 * If the buffer is created out from under us, we have to 2696 * throw away the one we just created. 2697 * 2698 * Note: this must occur before we associate the buffer 2699 * with the vp especially considering limitations in 2700 * the splay tree implementation when dealing with duplicate 2701 * lblkno's. 2702 */ 2703 BO_LOCK(bo); 2704 if (gbincore(bo, blkno)) { 2705 BO_UNLOCK(bo); 2706 bp->b_flags |= B_INVAL; 2707 brelse(bp); 2708 goto loop; 2709 } 2710 2711 /* 2712 * Insert the buffer into the hash, so that it can 2713 * be found by incore. 2714 */ 2715 bp->b_blkno = bp->b_lblkno = blkno; 2716 bp->b_offset = offset; 2717 bgetvp(vp, bp); 2718 BO_UNLOCK(bo); 2719 2720 /* 2721 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2722 * buffer size starts out as 0, B_CACHE will be set by 2723 * allocbuf() for the VMIO case prior to it testing the 2724 * backing store for validity. 2725 */ 2726 2727 if (vmio) { 2728 bp->b_flags |= B_VMIO; 2729 #if defined(VFS_BIO_DEBUG) 2730 if (vn_canvmio(vp) != TRUE) 2731 printf("getblk: VMIO on vnode type %d\n", 2732 vp->v_type); 2733 #endif 2734 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2735 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2736 bp, vp->v_object, bp->b_bufobj->bo_object)); 2737 } else { 2738 bp->b_flags &= ~B_VMIO; 2739 KASSERT(bp->b_bufobj->bo_object == NULL, 2740 ("ARGH! has b_bufobj->bo_object %p %p\n", 2741 bp, bp->b_bufobj->bo_object)); 2742 } 2743 2744 allocbuf(bp, size); 2745 bp->b_flags &= ~B_DONE; 2746 } 2747 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2748 BUF_ASSERT_HELD(bp); 2749 KASSERT(bp->b_bufobj == bo, 2750 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2751 return (bp); 2752 } 2753 2754 /* 2755 * Get an empty, disassociated buffer of given size. The buffer is initially 2756 * set to B_INVAL. 2757 */ 2758 struct buf * 2759 geteblk(int size, int flags) 2760 { 2761 struct buf *bp; 2762 int maxsize; 2763 2764 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2765 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2766 if ((flags & GB_NOWAIT_BD) && 2767 (curthread->td_pflags & TDP_BUFNEED) != 0) 2768 return (NULL); 2769 } 2770 allocbuf(bp, size); 2771 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2772 BUF_ASSERT_HELD(bp); 2773 return (bp); 2774 } 2775 2776 2777 /* 2778 * This code constitutes the buffer memory from either anonymous system 2779 * memory (in the case of non-VMIO operations) or from an associated 2780 * VM object (in the case of VMIO operations). This code is able to 2781 * resize a buffer up or down. 2782 * 2783 * Note that this code is tricky, and has many complications to resolve 2784 * deadlock or inconsistant data situations. Tread lightly!!! 2785 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2786 * the caller. Calling this code willy nilly can result in the loss of data. 2787 * 2788 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2789 * B_CACHE for the non-VMIO case. 2790 */ 2791 2792 int 2793 allocbuf(struct buf *bp, int size) 2794 { 2795 int newbsize, mbsize; 2796 int i; 2797 2798 BUF_ASSERT_HELD(bp); 2799 2800 if (bp->b_kvasize < size) 2801 panic("allocbuf: buffer too small"); 2802 2803 if ((bp->b_flags & B_VMIO) == 0) { 2804 caddr_t origbuf; 2805 int origbufsize; 2806 /* 2807 * Just get anonymous memory from the kernel. Don't 2808 * mess with B_CACHE. 2809 */ 2810 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2811 if (bp->b_flags & B_MALLOC) 2812 newbsize = mbsize; 2813 else 2814 newbsize = round_page(size); 2815 2816 if (newbsize < bp->b_bufsize) { 2817 /* 2818 * malloced buffers are not shrunk 2819 */ 2820 if (bp->b_flags & B_MALLOC) { 2821 if (newbsize) { 2822 bp->b_bcount = size; 2823 } else { 2824 free(bp->b_data, M_BIOBUF); 2825 if (bp->b_bufsize) { 2826 atomic_subtract_long( 2827 &bufmallocspace, 2828 bp->b_bufsize); 2829 bufspacewakeup(); 2830 bp->b_bufsize = 0; 2831 } 2832 bp->b_saveaddr = bp->b_kvabase; 2833 bp->b_data = bp->b_saveaddr; 2834 bp->b_bcount = 0; 2835 bp->b_flags &= ~B_MALLOC; 2836 } 2837 return 1; 2838 } 2839 vm_hold_free_pages( 2840 bp, 2841 (vm_offset_t) bp->b_data + newbsize, 2842 (vm_offset_t) bp->b_data + bp->b_bufsize); 2843 } else if (newbsize > bp->b_bufsize) { 2844 /* 2845 * We only use malloced memory on the first allocation. 2846 * and revert to page-allocated memory when the buffer 2847 * grows. 2848 */ 2849 /* 2850 * There is a potential smp race here that could lead 2851 * to bufmallocspace slightly passing the max. It 2852 * is probably extremely rare and not worth worrying 2853 * over. 2854 */ 2855 if ( (bufmallocspace < maxbufmallocspace) && 2856 (bp->b_bufsize == 0) && 2857 (mbsize <= PAGE_SIZE/2)) { 2858 2859 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2860 bp->b_bufsize = mbsize; 2861 bp->b_bcount = size; 2862 bp->b_flags |= B_MALLOC; 2863 atomic_add_long(&bufmallocspace, mbsize); 2864 return 1; 2865 } 2866 origbuf = NULL; 2867 origbufsize = 0; 2868 /* 2869 * If the buffer is growing on its other-than-first allocation, 2870 * then we revert to the page-allocation scheme. 2871 */ 2872 if (bp->b_flags & B_MALLOC) { 2873 origbuf = bp->b_data; 2874 origbufsize = bp->b_bufsize; 2875 bp->b_data = bp->b_kvabase; 2876 if (bp->b_bufsize) { 2877 atomic_subtract_long(&bufmallocspace, 2878 bp->b_bufsize); 2879 bufspacewakeup(); 2880 bp->b_bufsize = 0; 2881 } 2882 bp->b_flags &= ~B_MALLOC; 2883 newbsize = round_page(newbsize); 2884 } 2885 vm_hold_load_pages( 2886 bp, 2887 (vm_offset_t) bp->b_data + bp->b_bufsize, 2888 (vm_offset_t) bp->b_data + newbsize); 2889 if (origbuf) { 2890 bcopy(origbuf, bp->b_data, origbufsize); 2891 free(origbuf, M_BIOBUF); 2892 } 2893 } 2894 } else { 2895 int desiredpages; 2896 2897 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2898 desiredpages = (size == 0) ? 0 : 2899 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2900 2901 if (bp->b_flags & B_MALLOC) 2902 panic("allocbuf: VMIO buffer can't be malloced"); 2903 /* 2904 * Set B_CACHE initially if buffer is 0 length or will become 2905 * 0-length. 2906 */ 2907 if (size == 0 || bp->b_bufsize == 0) 2908 bp->b_flags |= B_CACHE; 2909 2910 if (newbsize < bp->b_bufsize) { 2911 /* 2912 * DEV_BSIZE aligned new buffer size is less then the 2913 * DEV_BSIZE aligned existing buffer size. Figure out 2914 * if we have to remove any pages. 2915 */ 2916 if (desiredpages < bp->b_npages) { 2917 vm_page_t m; 2918 2919 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2920 vm_page_lock_queues(); 2921 for (i = desiredpages; i < bp->b_npages; i++) { 2922 /* 2923 * the page is not freed here -- it 2924 * is the responsibility of 2925 * vnode_pager_setsize 2926 */ 2927 m = bp->b_pages[i]; 2928 KASSERT(m != bogus_page, 2929 ("allocbuf: bogus page found")); 2930 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2931 vm_page_lock_queues(); 2932 2933 bp->b_pages[i] = NULL; 2934 vm_page_unwire(m, 0); 2935 } 2936 vm_page_unlock_queues(); 2937 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2938 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2939 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2940 bp->b_npages = desiredpages; 2941 } 2942 } else if (size > bp->b_bcount) { 2943 /* 2944 * We are growing the buffer, possibly in a 2945 * byte-granular fashion. 2946 */ 2947 struct vnode *vp; 2948 vm_object_t obj; 2949 vm_offset_t toff; 2950 vm_offset_t tinc; 2951 2952 /* 2953 * Step 1, bring in the VM pages from the object, 2954 * allocating them if necessary. We must clear 2955 * B_CACHE if these pages are not valid for the 2956 * range covered by the buffer. 2957 */ 2958 2959 vp = bp->b_vp; 2960 obj = bp->b_bufobj->bo_object; 2961 2962 VM_OBJECT_LOCK(obj); 2963 while (bp->b_npages < desiredpages) { 2964 vm_page_t m; 2965 vm_pindex_t pi; 2966 2967 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2968 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2969 /* 2970 * note: must allocate system pages 2971 * since blocking here could intefere 2972 * with paging I/O, no matter which 2973 * process we are. 2974 */ 2975 m = vm_page_alloc(obj, pi, 2976 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2977 VM_ALLOC_WIRED); 2978 if (m == NULL) { 2979 atomic_add_int(&vm_pageout_deficit, 2980 desiredpages - bp->b_npages); 2981 VM_OBJECT_UNLOCK(obj); 2982 VM_WAIT; 2983 VM_OBJECT_LOCK(obj); 2984 } else { 2985 if (m->valid == 0) 2986 bp->b_flags &= ~B_CACHE; 2987 bp->b_pages[bp->b_npages] = m; 2988 ++bp->b_npages; 2989 } 2990 continue; 2991 } 2992 2993 /* 2994 * We found a page. If we have to sleep on it, 2995 * retry because it might have gotten freed out 2996 * from under us. 2997 * 2998 * We can only test VPO_BUSY here. Blocking on 2999 * m->busy might lead to a deadlock: 3000 * 3001 * vm_fault->getpages->cluster_read->allocbuf 3002 * 3003 */ 3004 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 3005 continue; 3006 3007 /* 3008 * We have a good page. 3009 */ 3010 vm_page_lock_queues(); 3011 vm_page_wire(m); 3012 vm_page_unlock_queues(); 3013 bp->b_pages[bp->b_npages] = m; 3014 ++bp->b_npages; 3015 } 3016 3017 /* 3018 * Step 2. We've loaded the pages into the buffer, 3019 * we have to figure out if we can still have B_CACHE 3020 * set. Note that B_CACHE is set according to the 3021 * byte-granular range ( bcount and size ), new the 3022 * aligned range ( newbsize ). 3023 * 3024 * The VM test is against m->valid, which is DEV_BSIZE 3025 * aligned. Needless to say, the validity of the data 3026 * needs to also be DEV_BSIZE aligned. Note that this 3027 * fails with NFS if the server or some other client 3028 * extends the file's EOF. If our buffer is resized, 3029 * B_CACHE may remain set! XXX 3030 */ 3031 3032 toff = bp->b_bcount; 3033 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3034 3035 while ((bp->b_flags & B_CACHE) && toff < size) { 3036 vm_pindex_t pi; 3037 3038 if (tinc > (size - toff)) 3039 tinc = size - toff; 3040 3041 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3042 PAGE_SHIFT; 3043 3044 vfs_buf_test_cache( 3045 bp, 3046 bp->b_offset, 3047 toff, 3048 tinc, 3049 bp->b_pages[pi] 3050 ); 3051 toff += tinc; 3052 tinc = PAGE_SIZE; 3053 } 3054 VM_OBJECT_UNLOCK(obj); 3055 3056 /* 3057 * Step 3, fixup the KVM pmap. Remember that 3058 * bp->b_data is relative to bp->b_offset, but 3059 * bp->b_offset may be offset into the first page. 3060 */ 3061 3062 bp->b_data = (caddr_t) 3063 trunc_page((vm_offset_t)bp->b_data); 3064 pmap_qenter( 3065 (vm_offset_t)bp->b_data, 3066 bp->b_pages, 3067 bp->b_npages 3068 ); 3069 3070 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3071 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3072 } 3073 } 3074 if (newbsize < bp->b_bufsize) 3075 bufspacewakeup(); 3076 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3077 bp->b_bcount = size; /* requested buffer size */ 3078 return 1; 3079 } 3080 3081 void 3082 biodone(struct bio *bp) 3083 { 3084 struct mtx *mtxp; 3085 void (*done)(struct bio *); 3086 3087 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3088 mtx_lock(mtxp); 3089 bp->bio_flags |= BIO_DONE; 3090 done = bp->bio_done; 3091 if (done == NULL) 3092 wakeup(bp); 3093 mtx_unlock(mtxp); 3094 if (done != NULL) 3095 done(bp); 3096 } 3097 3098 /* 3099 * Wait for a BIO to finish. 3100 * 3101 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3102 * case is not yet clear. 3103 */ 3104 int 3105 biowait(struct bio *bp, const char *wchan) 3106 { 3107 struct mtx *mtxp; 3108 3109 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3110 mtx_lock(mtxp); 3111 while ((bp->bio_flags & BIO_DONE) == 0) 3112 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3113 mtx_unlock(mtxp); 3114 if (bp->bio_error != 0) 3115 return (bp->bio_error); 3116 if (!(bp->bio_flags & BIO_ERROR)) 3117 return (0); 3118 return (EIO); 3119 } 3120 3121 void 3122 biofinish(struct bio *bp, struct devstat *stat, int error) 3123 { 3124 3125 if (error) { 3126 bp->bio_error = error; 3127 bp->bio_flags |= BIO_ERROR; 3128 } 3129 if (stat != NULL) 3130 devstat_end_transaction_bio(stat, bp); 3131 biodone(bp); 3132 } 3133 3134 /* 3135 * bufwait: 3136 * 3137 * Wait for buffer I/O completion, returning error status. The buffer 3138 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3139 * error and cleared. 3140 */ 3141 int 3142 bufwait(struct buf *bp) 3143 { 3144 if (bp->b_iocmd == BIO_READ) 3145 bwait(bp, PRIBIO, "biord"); 3146 else 3147 bwait(bp, PRIBIO, "biowr"); 3148 if (bp->b_flags & B_EINTR) { 3149 bp->b_flags &= ~B_EINTR; 3150 return (EINTR); 3151 } 3152 if (bp->b_ioflags & BIO_ERROR) { 3153 return (bp->b_error ? bp->b_error : EIO); 3154 } else { 3155 return (0); 3156 } 3157 } 3158 3159 /* 3160 * Call back function from struct bio back up to struct buf. 3161 */ 3162 static void 3163 bufdonebio(struct bio *bip) 3164 { 3165 struct buf *bp; 3166 3167 bp = bip->bio_caller2; 3168 bp->b_resid = bp->b_bcount - bip->bio_completed; 3169 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3170 bp->b_ioflags = bip->bio_flags; 3171 bp->b_error = bip->bio_error; 3172 if (bp->b_error) 3173 bp->b_ioflags |= BIO_ERROR; 3174 bufdone(bp); 3175 g_destroy_bio(bip); 3176 } 3177 3178 void 3179 dev_strategy(struct cdev *dev, struct buf *bp) 3180 { 3181 struct cdevsw *csw; 3182 struct bio *bip; 3183 3184 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3185 panic("b_iocmd botch"); 3186 for (;;) { 3187 bip = g_new_bio(); 3188 if (bip != NULL) 3189 break; 3190 /* Try again later */ 3191 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3192 } 3193 bip->bio_cmd = bp->b_iocmd; 3194 bip->bio_offset = bp->b_iooffset; 3195 bip->bio_length = bp->b_bcount; 3196 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3197 bip->bio_data = bp->b_data; 3198 bip->bio_done = bufdonebio; 3199 bip->bio_caller2 = bp; 3200 bip->bio_dev = dev; 3201 KASSERT(dev->si_refcount > 0, 3202 ("dev_strategy on un-referenced struct cdev *(%s)", 3203 devtoname(dev))); 3204 csw = dev_refthread(dev); 3205 if (csw == NULL) { 3206 g_destroy_bio(bip); 3207 bp->b_error = ENXIO; 3208 bp->b_ioflags = BIO_ERROR; 3209 bufdone(bp); 3210 return; 3211 } 3212 (*csw->d_strategy)(bip); 3213 dev_relthread(dev); 3214 } 3215 3216 /* 3217 * bufdone: 3218 * 3219 * Finish I/O on a buffer, optionally calling a completion function. 3220 * This is usually called from an interrupt so process blocking is 3221 * not allowed. 3222 * 3223 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3224 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3225 * assuming B_INVAL is clear. 3226 * 3227 * For the VMIO case, we set B_CACHE if the op was a read and no 3228 * read error occured, or if the op was a write. B_CACHE is never 3229 * set if the buffer is invalid or otherwise uncacheable. 3230 * 3231 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3232 * initiator to leave B_INVAL set to brelse the buffer out of existance 3233 * in the biodone routine. 3234 */ 3235 void 3236 bufdone(struct buf *bp) 3237 { 3238 struct bufobj *dropobj; 3239 void (*biodone)(struct buf *); 3240 3241 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3242 dropobj = NULL; 3243 3244 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3245 BUF_ASSERT_HELD(bp); 3246 3247 runningbufwakeup(bp); 3248 if (bp->b_iocmd == BIO_WRITE) 3249 dropobj = bp->b_bufobj; 3250 /* call optional completion function if requested */ 3251 if (bp->b_iodone != NULL) { 3252 biodone = bp->b_iodone; 3253 bp->b_iodone = NULL; 3254 (*biodone) (bp); 3255 if (dropobj) 3256 bufobj_wdrop(dropobj); 3257 return; 3258 } 3259 3260 bufdone_finish(bp); 3261 3262 if (dropobj) 3263 bufobj_wdrop(dropobj); 3264 } 3265 3266 void 3267 bufdone_finish(struct buf *bp) 3268 { 3269 BUF_ASSERT_HELD(bp); 3270 3271 if (!LIST_EMPTY(&bp->b_dep)) 3272 buf_complete(bp); 3273 3274 if (bp->b_flags & B_VMIO) { 3275 int i; 3276 vm_ooffset_t foff; 3277 vm_page_t m; 3278 vm_object_t obj; 3279 int iosize; 3280 struct vnode *vp = bp->b_vp; 3281 3282 obj = bp->b_bufobj->bo_object; 3283 3284 #if defined(VFS_BIO_DEBUG) 3285 mp_fixme("usecount and vflag accessed without locks."); 3286 if (vp->v_usecount == 0) { 3287 panic("biodone: zero vnode ref count"); 3288 } 3289 3290 KASSERT(vp->v_object != NULL, 3291 ("biodone: vnode %p has no vm_object", vp)); 3292 #endif 3293 3294 foff = bp->b_offset; 3295 KASSERT(bp->b_offset != NOOFFSET, 3296 ("biodone: no buffer offset")); 3297 3298 VM_OBJECT_LOCK(obj); 3299 #if defined(VFS_BIO_DEBUG) 3300 if (obj->paging_in_progress < bp->b_npages) { 3301 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3302 obj->paging_in_progress, bp->b_npages); 3303 } 3304 #endif 3305 3306 /* 3307 * Set B_CACHE if the op was a normal read and no error 3308 * occured. B_CACHE is set for writes in the b*write() 3309 * routines. 3310 */ 3311 iosize = bp->b_bcount - bp->b_resid; 3312 if (bp->b_iocmd == BIO_READ && 3313 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3314 !(bp->b_ioflags & BIO_ERROR)) { 3315 bp->b_flags |= B_CACHE; 3316 } 3317 for (i = 0; i < bp->b_npages; i++) { 3318 int bogusflag = 0; 3319 int resid; 3320 3321 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3322 if (resid > iosize) 3323 resid = iosize; 3324 3325 /* 3326 * cleanup bogus pages, restoring the originals 3327 */ 3328 m = bp->b_pages[i]; 3329 if (m == bogus_page) { 3330 bogusflag = 1; 3331 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3332 if (m == NULL) 3333 panic("biodone: page disappeared!"); 3334 bp->b_pages[i] = m; 3335 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3336 bp->b_pages, bp->b_npages); 3337 } 3338 #if defined(VFS_BIO_DEBUG) 3339 if (OFF_TO_IDX(foff) != m->pindex) { 3340 printf( 3341 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3342 (intmax_t)foff, (uintmax_t)m->pindex); 3343 } 3344 #endif 3345 3346 /* 3347 * In the write case, the valid and clean bits are 3348 * already changed correctly ( see bdwrite() ), so we 3349 * only need to do this here in the read case. 3350 */ 3351 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3352 KASSERT((m->dirty & vm_page_bits(foff & 3353 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3354 " page %p has unexpected dirty bits", m)); 3355 vfs_page_set_valid(bp, foff, m); 3356 } 3357 3358 /* 3359 * when debugging new filesystems or buffer I/O methods, this 3360 * is the most common error that pops up. if you see this, you 3361 * have not set the page busy flag correctly!!! 3362 */ 3363 if (m->busy == 0) { 3364 printf("biodone: page busy < 0, " 3365 "pindex: %d, foff: 0x(%x,%x), " 3366 "resid: %d, index: %d\n", 3367 (int) m->pindex, (int)(foff >> 32), 3368 (int) foff & 0xffffffff, resid, i); 3369 if (!vn_isdisk(vp, NULL)) 3370 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3371 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3372 (intmax_t) bp->b_lblkno, 3373 bp->b_flags, bp->b_npages); 3374 else 3375 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3376 (intmax_t) bp->b_lblkno, 3377 bp->b_flags, bp->b_npages); 3378 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3379 (u_long)m->valid, (u_long)m->dirty, 3380 m->wire_count); 3381 panic("biodone: page busy < 0\n"); 3382 } 3383 vm_page_io_finish(m); 3384 vm_object_pip_subtract(obj, 1); 3385 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3386 iosize -= resid; 3387 } 3388 vm_object_pip_wakeupn(obj, 0); 3389 VM_OBJECT_UNLOCK(obj); 3390 } 3391 3392 /* 3393 * For asynchronous completions, release the buffer now. The brelse 3394 * will do a wakeup there if necessary - so no need to do a wakeup 3395 * here in the async case. The sync case always needs to do a wakeup. 3396 */ 3397 3398 if (bp->b_flags & B_ASYNC) { 3399 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3400 brelse(bp); 3401 else 3402 bqrelse(bp); 3403 } else 3404 bdone(bp); 3405 } 3406 3407 /* 3408 * This routine is called in lieu of iodone in the case of 3409 * incomplete I/O. This keeps the busy status for pages 3410 * consistant. 3411 */ 3412 void 3413 vfs_unbusy_pages(struct buf *bp) 3414 { 3415 int i; 3416 vm_object_t obj; 3417 vm_page_t m; 3418 3419 runningbufwakeup(bp); 3420 if (!(bp->b_flags & B_VMIO)) 3421 return; 3422 3423 obj = bp->b_bufobj->bo_object; 3424 VM_OBJECT_LOCK(obj); 3425 for (i = 0; i < bp->b_npages; i++) { 3426 m = bp->b_pages[i]; 3427 if (m == bogus_page) { 3428 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3429 if (!m) 3430 panic("vfs_unbusy_pages: page missing\n"); 3431 bp->b_pages[i] = m; 3432 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3433 bp->b_pages, bp->b_npages); 3434 } 3435 vm_object_pip_subtract(obj, 1); 3436 vm_page_io_finish(m); 3437 } 3438 vm_object_pip_wakeupn(obj, 0); 3439 VM_OBJECT_UNLOCK(obj); 3440 } 3441 3442 /* 3443 * vfs_page_set_valid: 3444 * 3445 * Set the valid bits in a page based on the supplied offset. The 3446 * range is restricted to the buffer's size. 3447 * 3448 * This routine is typically called after a read completes. 3449 */ 3450 static void 3451 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3452 { 3453 vm_ooffset_t eoff; 3454 3455 /* 3456 * Compute the end offset, eoff, such that [off, eoff) does not span a 3457 * page boundary and eoff is not greater than the end of the buffer. 3458 * The end of the buffer, in this case, is our file EOF, not the 3459 * allocation size of the buffer. 3460 */ 3461 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3462 if (eoff > bp->b_offset + bp->b_bcount) 3463 eoff = bp->b_offset + bp->b_bcount; 3464 3465 /* 3466 * Set valid range. This is typically the entire buffer and thus the 3467 * entire page. 3468 */ 3469 if (eoff > off) 3470 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3471 } 3472 3473 /* 3474 * vfs_page_set_validclean: 3475 * 3476 * Set the valid bits and clear the dirty bits in a page based on the 3477 * supplied offset. The range is restricted to the buffer's size. 3478 */ 3479 static void 3480 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3481 { 3482 vm_ooffset_t soff, eoff; 3483 3484 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3485 /* 3486 * Start and end offsets in buffer. eoff - soff may not cross a 3487 * page boundry or cross the end of the buffer. The end of the 3488 * buffer, in this case, is our file EOF, not the allocation size 3489 * of the buffer. 3490 */ 3491 soff = off; 3492 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3493 if (eoff > bp->b_offset + bp->b_bcount) 3494 eoff = bp->b_offset + bp->b_bcount; 3495 3496 /* 3497 * Set valid range. This is typically the entire buffer and thus the 3498 * entire page. 3499 */ 3500 if (eoff > soff) { 3501 vm_page_set_validclean( 3502 m, 3503 (vm_offset_t) (soff & PAGE_MASK), 3504 (vm_offset_t) (eoff - soff) 3505 ); 3506 } 3507 } 3508 3509 /* 3510 * This routine is called before a device strategy routine. 3511 * It is used to tell the VM system that paging I/O is in 3512 * progress, and treat the pages associated with the buffer 3513 * almost as being VPO_BUSY. Also the object paging_in_progress 3514 * flag is handled to make sure that the object doesn't become 3515 * inconsistant. 3516 * 3517 * Since I/O has not been initiated yet, certain buffer flags 3518 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3519 * and should be ignored. 3520 */ 3521 void 3522 vfs_busy_pages(struct buf *bp, int clear_modify) 3523 { 3524 int i, bogus; 3525 vm_object_t obj; 3526 vm_ooffset_t foff; 3527 vm_page_t m; 3528 3529 if (!(bp->b_flags & B_VMIO)) 3530 return; 3531 3532 obj = bp->b_bufobj->bo_object; 3533 foff = bp->b_offset; 3534 KASSERT(bp->b_offset != NOOFFSET, 3535 ("vfs_busy_pages: no buffer offset")); 3536 VM_OBJECT_LOCK(obj); 3537 if (bp->b_bufsize != 0) 3538 vfs_setdirty_locked_object(bp); 3539 retry: 3540 for (i = 0; i < bp->b_npages; i++) { 3541 m = bp->b_pages[i]; 3542 3543 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3544 goto retry; 3545 } 3546 bogus = 0; 3547 if (clear_modify) 3548 vm_page_lock_queues(); 3549 for (i = 0; i < bp->b_npages; i++) { 3550 m = bp->b_pages[i]; 3551 3552 if ((bp->b_flags & B_CLUSTER) == 0) { 3553 vm_object_pip_add(obj, 1); 3554 vm_page_io_start(m); 3555 } 3556 /* 3557 * When readying a buffer for a read ( i.e 3558 * clear_modify == 0 ), it is important to do 3559 * bogus_page replacement for valid pages in 3560 * partially instantiated buffers. Partially 3561 * instantiated buffers can, in turn, occur when 3562 * reconstituting a buffer from its VM backing store 3563 * base. We only have to do this if B_CACHE is 3564 * clear ( which causes the I/O to occur in the 3565 * first place ). The replacement prevents the read 3566 * I/O from overwriting potentially dirty VM-backed 3567 * pages. XXX bogus page replacement is, uh, bogus. 3568 * It may not work properly with small-block devices. 3569 * We need to find a better way. 3570 */ 3571 if (clear_modify) { 3572 pmap_remove_write(m); 3573 vfs_page_set_validclean(bp, foff, m); 3574 } else if (m->valid == VM_PAGE_BITS_ALL && 3575 (bp->b_flags & B_CACHE) == 0) { 3576 bp->b_pages[i] = bogus_page; 3577 bogus++; 3578 } 3579 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3580 } 3581 if (clear_modify) 3582 vm_page_unlock_queues(); 3583 VM_OBJECT_UNLOCK(obj); 3584 if (bogus) 3585 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3586 bp->b_pages, bp->b_npages); 3587 } 3588 3589 /* 3590 * Tell the VM system that the pages associated with this buffer 3591 * are clean. This is used for delayed writes where the data is 3592 * going to go to disk eventually without additional VM intevention. 3593 * 3594 * Note that while we only really need to clean through to b_bcount, we 3595 * just go ahead and clean through to b_bufsize. 3596 */ 3597 static void 3598 vfs_clean_pages(struct buf *bp) 3599 { 3600 int i; 3601 vm_ooffset_t foff, noff, eoff; 3602 vm_page_t m; 3603 3604 if (!(bp->b_flags & B_VMIO)) 3605 return; 3606 3607 foff = bp->b_offset; 3608 KASSERT(bp->b_offset != NOOFFSET, 3609 ("vfs_clean_pages: no buffer offset")); 3610 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3611 vm_page_lock_queues(); 3612 for (i = 0; i < bp->b_npages; i++) { 3613 m = bp->b_pages[i]; 3614 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3615 eoff = noff; 3616 3617 if (eoff > bp->b_offset + bp->b_bufsize) 3618 eoff = bp->b_offset + bp->b_bufsize; 3619 vfs_page_set_validclean(bp, foff, m); 3620 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3621 foff = noff; 3622 } 3623 vm_page_unlock_queues(); 3624 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3625 } 3626 3627 /* 3628 * vfs_bio_set_valid: 3629 * 3630 * Set the range within the buffer to valid. The range is 3631 * relative to the beginning of the buffer, b_offset. Note that 3632 * b_offset itself may be offset from the beginning of the first 3633 * page. 3634 */ 3635 void 3636 vfs_bio_set_valid(struct buf *bp, int base, int size) 3637 { 3638 int i, n; 3639 vm_page_t m; 3640 3641 if (!(bp->b_flags & B_VMIO)) 3642 return; 3643 3644 /* 3645 * Fixup base to be relative to beginning of first page. 3646 * Set initial n to be the maximum number of bytes in the 3647 * first page that can be validated. 3648 */ 3649 base += (bp->b_offset & PAGE_MASK); 3650 n = PAGE_SIZE - (base & PAGE_MASK); 3651 3652 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3653 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3654 m = bp->b_pages[i]; 3655 if (n > size) 3656 n = size; 3657 vm_page_set_valid(m, base & PAGE_MASK, n); 3658 base += n; 3659 size -= n; 3660 n = PAGE_SIZE; 3661 } 3662 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3663 } 3664 3665 /* 3666 * vfs_bio_set_validclean: 3667 * 3668 * Set the range within the buffer to valid and clean. The range is 3669 * relative to the beginning of the buffer, b_offset. Note that b_offset 3670 * itself may be offset from the beginning of the first page. 3671 * 3672 */ 3673 3674 void 3675 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3676 { 3677 int i, n; 3678 vm_page_t m; 3679 3680 if (!(bp->b_flags & B_VMIO)) 3681 return; 3682 /* 3683 * Fixup base to be relative to beginning of first page. 3684 * Set initial n to be the maximum number of bytes in the 3685 * first page that can be validated. 3686 */ 3687 3688 base += (bp->b_offset & PAGE_MASK); 3689 n = PAGE_SIZE - (base & PAGE_MASK); 3690 3691 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3692 vm_page_lock_queues(); 3693 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3694 m = bp->b_pages[i]; 3695 if (n > size) 3696 n = size; 3697 vm_page_set_validclean(m, base & PAGE_MASK, n); 3698 base += n; 3699 size -= n; 3700 n = PAGE_SIZE; 3701 } 3702 vm_page_unlock_queues(); 3703 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3704 } 3705 3706 /* 3707 * vfs_bio_clrbuf: 3708 * 3709 * If the specified buffer is a non-VMIO buffer, clear the entire 3710 * buffer. If the specified buffer is a VMIO buffer, clear and 3711 * validate only the previously invalid portions of the buffer. 3712 * This routine essentially fakes an I/O, so we need to clear 3713 * BIO_ERROR and B_INVAL. 3714 * 3715 * Note that while we only theoretically need to clear through b_bcount, 3716 * we go ahead and clear through b_bufsize. 3717 */ 3718 void 3719 vfs_bio_clrbuf(struct buf *bp) 3720 { 3721 int i, j, mask; 3722 caddr_t sa, ea; 3723 3724 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3725 clrbuf(bp); 3726 return; 3727 } 3728 bp->b_flags &= ~B_INVAL; 3729 bp->b_ioflags &= ~BIO_ERROR; 3730 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3731 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3732 (bp->b_offset & PAGE_MASK) == 0) { 3733 if (bp->b_pages[0] == bogus_page) 3734 goto unlock; 3735 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3736 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3737 if ((bp->b_pages[0]->valid & mask) == mask) 3738 goto unlock; 3739 if ((bp->b_pages[0]->valid & mask) == 0) { 3740 bzero(bp->b_data, bp->b_bufsize); 3741 bp->b_pages[0]->valid |= mask; 3742 goto unlock; 3743 } 3744 } 3745 ea = sa = bp->b_data; 3746 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3747 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3748 ea = (caddr_t)(vm_offset_t)ulmin( 3749 (u_long)(vm_offset_t)ea, 3750 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3751 if (bp->b_pages[i] == bogus_page) 3752 continue; 3753 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3754 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3755 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3756 if ((bp->b_pages[i]->valid & mask) == mask) 3757 continue; 3758 if ((bp->b_pages[i]->valid & mask) == 0) 3759 bzero(sa, ea - sa); 3760 else { 3761 for (; sa < ea; sa += DEV_BSIZE, j++) { 3762 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3763 bzero(sa, DEV_BSIZE); 3764 } 3765 } 3766 bp->b_pages[i]->valid |= mask; 3767 } 3768 unlock: 3769 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3770 bp->b_resid = 0; 3771 } 3772 3773 /* 3774 * vm_hold_load_pages and vm_hold_free_pages get pages into 3775 * a buffers address space. The pages are anonymous and are 3776 * not associated with a file object. 3777 */ 3778 static void 3779 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3780 { 3781 vm_offset_t pg; 3782 vm_page_t p; 3783 int index; 3784 3785 to = round_page(to); 3786 from = round_page(from); 3787 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3788 3789 VM_OBJECT_LOCK(kernel_object); 3790 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3791 tryagain: 3792 /* 3793 * note: must allocate system pages since blocking here 3794 * could intefere with paging I/O, no matter which 3795 * process we are. 3796 */ 3797 p = vm_page_alloc(kernel_object, 3798 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3799 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3800 if (!p) { 3801 atomic_add_int(&vm_pageout_deficit, 3802 (to - pg) >> PAGE_SHIFT); 3803 VM_OBJECT_UNLOCK(kernel_object); 3804 VM_WAIT; 3805 VM_OBJECT_LOCK(kernel_object); 3806 goto tryagain; 3807 } 3808 p->valid = VM_PAGE_BITS_ALL; 3809 pmap_qenter(pg, &p, 1); 3810 bp->b_pages[index] = p; 3811 } 3812 VM_OBJECT_UNLOCK(kernel_object); 3813 bp->b_npages = index; 3814 } 3815 3816 /* Return pages associated with this buf to the vm system */ 3817 static void 3818 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3819 { 3820 vm_offset_t pg; 3821 vm_page_t p; 3822 int index, newnpages; 3823 3824 from = round_page(from); 3825 to = round_page(to); 3826 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3827 3828 VM_OBJECT_LOCK(kernel_object); 3829 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3830 p = bp->b_pages[index]; 3831 if (p && (index < bp->b_npages)) { 3832 if (p->busy) { 3833 printf( 3834 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3835 (intmax_t)bp->b_blkno, 3836 (intmax_t)bp->b_lblkno); 3837 } 3838 bp->b_pages[index] = NULL; 3839 pmap_qremove(pg, 1); 3840 vm_page_lock_queues(); 3841 vm_page_unwire(p, 0); 3842 vm_page_free(p); 3843 vm_page_unlock_queues(); 3844 } 3845 } 3846 VM_OBJECT_UNLOCK(kernel_object); 3847 bp->b_npages = newnpages; 3848 } 3849 3850 /* 3851 * Map an IO request into kernel virtual address space. 3852 * 3853 * All requests are (re)mapped into kernel VA space. 3854 * Notice that we use b_bufsize for the size of the buffer 3855 * to be mapped. b_bcount might be modified by the driver. 3856 * 3857 * Note that even if the caller determines that the address space should 3858 * be valid, a race or a smaller-file mapped into a larger space may 3859 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3860 * check the return value. 3861 */ 3862 int 3863 vmapbuf(struct buf *bp) 3864 { 3865 caddr_t addr, kva; 3866 vm_prot_t prot; 3867 int pidx, i; 3868 struct vm_page *m; 3869 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3870 3871 if (bp->b_bufsize < 0) 3872 return (-1); 3873 prot = VM_PROT_READ; 3874 if (bp->b_iocmd == BIO_READ) 3875 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3876 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3877 addr < bp->b_data + bp->b_bufsize; 3878 addr += PAGE_SIZE, pidx++) { 3879 /* 3880 * Do the vm_fault if needed; do the copy-on-write thing 3881 * when reading stuff off device into memory. 3882 * 3883 * NOTE! Must use pmap_extract() because addr may be in 3884 * the userland address space, and kextract is only guarenteed 3885 * to work for the kernland address space (see: sparc64 port). 3886 */ 3887 retry: 3888 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3889 prot) < 0) { 3890 vm_page_lock_queues(); 3891 for (i = 0; i < pidx; ++i) { 3892 vm_page_unhold(bp->b_pages[i]); 3893 bp->b_pages[i] = NULL; 3894 } 3895 vm_page_unlock_queues(); 3896 return(-1); 3897 } 3898 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3899 if (m == NULL) 3900 goto retry; 3901 bp->b_pages[pidx] = m; 3902 } 3903 if (pidx > btoc(MAXPHYS)) 3904 panic("vmapbuf: mapped more than MAXPHYS"); 3905 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3906 3907 kva = bp->b_saveaddr; 3908 bp->b_npages = pidx; 3909 bp->b_saveaddr = bp->b_data; 3910 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3911 return(0); 3912 } 3913 3914 /* 3915 * Free the io map PTEs associated with this IO operation. 3916 * We also invalidate the TLB entries and restore the original b_addr. 3917 */ 3918 void 3919 vunmapbuf(struct buf *bp) 3920 { 3921 int pidx; 3922 int npages; 3923 3924 npages = bp->b_npages; 3925 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3926 vm_page_lock_queues(); 3927 for (pidx = 0; pidx < npages; pidx++) 3928 vm_page_unhold(bp->b_pages[pidx]); 3929 vm_page_unlock_queues(); 3930 3931 bp->b_data = bp->b_saveaddr; 3932 } 3933 3934 void 3935 bdone(struct buf *bp) 3936 { 3937 struct mtx *mtxp; 3938 3939 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3940 mtx_lock(mtxp); 3941 bp->b_flags |= B_DONE; 3942 wakeup(bp); 3943 mtx_unlock(mtxp); 3944 } 3945 3946 void 3947 bwait(struct buf *bp, u_char pri, const char *wchan) 3948 { 3949 struct mtx *mtxp; 3950 3951 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3952 mtx_lock(mtxp); 3953 while ((bp->b_flags & B_DONE) == 0) 3954 msleep(bp, mtxp, pri, wchan, 0); 3955 mtx_unlock(mtxp); 3956 } 3957 3958 int 3959 bufsync(struct bufobj *bo, int waitfor) 3960 { 3961 3962 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3963 } 3964 3965 void 3966 bufstrategy(struct bufobj *bo, struct buf *bp) 3967 { 3968 int i = 0; 3969 struct vnode *vp; 3970 3971 vp = bp->b_vp; 3972 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3973 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3974 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3975 i = VOP_STRATEGY(vp, bp); 3976 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3977 } 3978 3979 void 3980 bufobj_wrefl(struct bufobj *bo) 3981 { 3982 3983 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3984 ASSERT_BO_LOCKED(bo); 3985 bo->bo_numoutput++; 3986 } 3987 3988 void 3989 bufobj_wref(struct bufobj *bo) 3990 { 3991 3992 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3993 BO_LOCK(bo); 3994 bo->bo_numoutput++; 3995 BO_UNLOCK(bo); 3996 } 3997 3998 void 3999 bufobj_wdrop(struct bufobj *bo) 4000 { 4001 4002 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4003 BO_LOCK(bo); 4004 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4005 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4006 bo->bo_flag &= ~BO_WWAIT; 4007 wakeup(&bo->bo_numoutput); 4008 } 4009 BO_UNLOCK(bo); 4010 } 4011 4012 int 4013 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4014 { 4015 int error; 4016 4017 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4018 ASSERT_BO_LOCKED(bo); 4019 error = 0; 4020 while (bo->bo_numoutput) { 4021 bo->bo_flag |= BO_WWAIT; 4022 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 4023 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4024 if (error) 4025 break; 4026 } 4027 return (error); 4028 } 4029 4030 void 4031 bpin(struct buf *bp) 4032 { 4033 struct mtx *mtxp; 4034 4035 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4036 mtx_lock(mtxp); 4037 bp->b_pin_count++; 4038 mtx_unlock(mtxp); 4039 } 4040 4041 void 4042 bunpin(struct buf *bp) 4043 { 4044 struct mtx *mtxp; 4045 4046 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4047 mtx_lock(mtxp); 4048 if (--bp->b_pin_count == 0) 4049 wakeup(bp); 4050 mtx_unlock(mtxp); 4051 } 4052 4053 void 4054 bunpin_wait(struct buf *bp) 4055 { 4056 struct mtx *mtxp; 4057 4058 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4059 mtx_lock(mtxp); 4060 while (bp->b_pin_count > 0) 4061 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4062 mtx_unlock(mtxp); 4063 } 4064 4065 #include "opt_ddb.h" 4066 #ifdef DDB 4067 #include <ddb/ddb.h> 4068 4069 /* DDB command to show buffer data */ 4070 DB_SHOW_COMMAND(buffer, db_show_buffer) 4071 { 4072 /* get args */ 4073 struct buf *bp = (struct buf *)addr; 4074 4075 if (!have_addr) { 4076 db_printf("usage: show buffer <addr>\n"); 4077 return; 4078 } 4079 4080 db_printf("buf at %p\n", bp); 4081 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4082 db_printf( 4083 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4084 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4085 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4086 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4087 bp->b_dep.lh_first); 4088 if (bp->b_npages) { 4089 int i; 4090 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4091 for (i = 0; i < bp->b_npages; i++) { 4092 vm_page_t m; 4093 m = bp->b_pages[i]; 4094 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4095 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4096 if ((i + 1) < bp->b_npages) 4097 db_printf(","); 4098 } 4099 db_printf("\n"); 4100 } 4101 db_printf(" "); 4102 lockmgr_printinfo(&bp->b_lock); 4103 } 4104 4105 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4106 { 4107 struct buf *bp; 4108 int i; 4109 4110 for (i = 0; i < nbuf; i++) { 4111 bp = &buf[i]; 4112 if (BUF_ISLOCKED(bp)) { 4113 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4114 db_printf("\n"); 4115 } 4116 } 4117 } 4118 4119 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4120 { 4121 struct vnode *vp; 4122 struct buf *bp; 4123 4124 if (!have_addr) { 4125 db_printf("usage: show vnodebufs <addr>\n"); 4126 return; 4127 } 4128 vp = (struct vnode *)addr; 4129 db_printf("Clean buffers:\n"); 4130 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4131 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4132 db_printf("\n"); 4133 } 4134 db_printf("Dirty buffers:\n"); 4135 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4136 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4137 db_printf("\n"); 4138 } 4139 } 4140 #endif /* DDB */ 4141