1 /*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * Copyright (c) 2013 The FreeBSD Foundation 5 * All rights reserved. 6 * 7 * Portions of this software were developed by Konstantin Belousov 8 * under sponsorship from the FreeBSD Foundation. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 */ 31 32 /* 33 * this file contains a new buffer I/O scheme implementing a coherent 34 * VM object and buffer cache scheme. Pains have been taken to make 35 * sure that the performance degradation associated with schemes such 36 * as this is not realized. 37 * 38 * Author: John S. Dyson 39 * Significant help during the development and debugging phases 40 * had been provided by David Greenman, also of the FreeBSD core team. 41 * 42 * see man buf(9) for more info. 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/bio.h> 51 #include <sys/conf.h> 52 #include <sys/buf.h> 53 #include <sys/devicestat.h> 54 #include <sys/eventhandler.h> 55 #include <sys/fail.h> 56 #include <sys/limits.h> 57 #include <sys/lock.h> 58 #include <sys/malloc.h> 59 #include <sys/mount.h> 60 #include <sys/mutex.h> 61 #include <sys/kernel.h> 62 #include <sys/kthread.h> 63 #include <sys/proc.h> 64 #include <sys/resourcevar.h> 65 #include <sys/rwlock.h> 66 #include <sys/smp.h> 67 #include <sys/sysctl.h> 68 #include <sys/sysproto.h> 69 #include <sys/vmem.h> 70 #include <sys/vmmeter.h> 71 #include <sys/vnode.h> 72 #include <sys/watchdog.h> 73 #include <geom/geom.h> 74 #include <vm/vm.h> 75 #include <vm/vm_param.h> 76 #include <vm/vm_kern.h> 77 #include <vm/vm_pageout.h> 78 #include <vm/vm_page.h> 79 #include <vm/vm_object.h> 80 #include <vm/vm_extern.h> 81 #include <vm/vm_map.h> 82 #include <vm/swap_pager.h> 83 #include "opt_compat.h" 84 #include "opt_swap.h" 85 86 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 87 88 struct bio_ops bioops; /* I/O operation notification */ 89 90 struct buf_ops buf_ops_bio = { 91 .bop_name = "buf_ops_bio", 92 .bop_write = bufwrite, 93 .bop_strategy = bufstrategy, 94 .bop_sync = bufsync, 95 .bop_bdflush = bufbdflush, 96 }; 97 98 static struct buf *buf; /* buffer header pool */ 99 extern struct buf *swbuf; /* Swap buffer header pool. */ 100 caddr_t unmapped_buf; 101 102 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 103 struct proc *bufdaemonproc; 104 struct proc *bufspacedaemonproc; 105 106 static int inmem(struct vnode *vp, daddr_t blkno); 107 static void vm_hold_free_pages(struct buf *bp, int newbsize); 108 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 109 vm_offset_t to); 110 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 111 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 112 vm_page_t m); 113 static void vfs_clean_pages_dirty_buf(struct buf *bp); 114 static void vfs_setdirty_locked_object(struct buf *bp); 115 static void vfs_vmio_invalidate(struct buf *bp); 116 static void vfs_vmio_truncate(struct buf *bp, int npages); 117 static void vfs_vmio_extend(struct buf *bp, int npages, int size); 118 static int vfs_bio_clcheck(struct vnode *vp, int size, 119 daddr_t lblkno, daddr_t blkno); 120 static int buf_flush(struct vnode *vp, int); 121 static int buf_recycle(bool); 122 static int buf_scan(bool); 123 static int flushbufqueues(struct vnode *, int, int); 124 static void buf_daemon(void); 125 static void bremfreel(struct buf *bp); 126 static __inline void bd_wakeup(void); 127 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 128 static void bufkva_reclaim(vmem_t *, int); 129 static void bufkva_free(struct buf *); 130 static int buf_import(void *, void **, int, int); 131 static void buf_release(void *, void **, int); 132 133 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 134 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 135 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 136 #endif 137 138 int vmiodirenable = TRUE; 139 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 140 "Use the VM system for directory writes"); 141 long runningbufspace; 142 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 143 "Amount of presently outstanding async buffer io"); 144 static long bufspace; 145 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 146 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 147 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 148 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 149 #else 150 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 151 "Physical memory used for buffers"); 152 #endif 153 static long bufkvaspace; 154 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0, 155 "Kernel virtual memory used for buffers"); 156 static long maxbufspace; 157 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, &maxbufspace, 0, 158 "Maximum allowed value of bufspace (including metadata)"); 159 static long bufmallocspace; 160 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 161 "Amount of malloced memory for buffers"); 162 static long maxbufmallocspace; 163 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 164 0, "Maximum amount of malloced memory for buffers"); 165 static long lobufspace; 166 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RW, &lobufspace, 0, 167 "Minimum amount of buffers we want to have"); 168 long hibufspace; 169 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RW, &hibufspace, 0, 170 "Maximum allowed value of bufspace (excluding metadata)"); 171 long bufspacethresh; 172 SYSCTL_LONG(_vfs, OID_AUTO, bufspacethresh, CTLFLAG_RW, &bufspacethresh, 173 0, "Bufspace consumed before waking the daemon to free some"); 174 static int buffreekvacnt; 175 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 176 "Number of times we have freed the KVA space from some buffer"); 177 static int bufdefragcnt; 178 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 179 "Number of times we have had to repeat buffer allocation to defragment"); 180 static long lorunningspace; 181 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 182 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 183 "Minimum preferred space used for in-progress I/O"); 184 static long hirunningspace; 185 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 186 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 187 "Maximum amount of space to use for in-progress I/O"); 188 int dirtybufferflushes; 189 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 190 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 191 int bdwriteskip; 192 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 193 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 194 int altbufferflushes; 195 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 196 0, "Number of fsync flushes to limit dirty buffers"); 197 static int recursiveflushes; 198 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 199 0, "Number of flushes skipped due to being recursive"); 200 static int numdirtybuffers; 201 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 202 "Number of buffers that are dirty (has unwritten changes) at the moment"); 203 static int lodirtybuffers; 204 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 205 "How many buffers we want to have free before bufdaemon can sleep"); 206 static int hidirtybuffers; 207 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 208 "When the number of dirty buffers is considered severe"); 209 int dirtybufthresh; 210 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 211 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 212 static int numfreebuffers; 213 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 214 "Number of free buffers"); 215 static int lofreebuffers; 216 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 217 "Target number of free buffers"); 218 static int hifreebuffers; 219 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 220 "Threshold for clean buffer recycling"); 221 static int getnewbufcalls; 222 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 223 "Number of calls to getnewbuf"); 224 static int getnewbufrestarts; 225 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 226 "Number of times getnewbuf has had to restart a buffer aquisition"); 227 static int mappingrestarts; 228 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 229 "Number of times getblk has had to restart a buffer mapping for " 230 "unmapped buffer"); 231 static int numbufallocfails; 232 SYSCTL_INT(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, 0, 233 "Number of times buffer allocations failed"); 234 static int flushbufqtarget = 100; 235 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 236 "Amount of work to do in flushbufqueues when helping bufdaemon"); 237 static long notbufdflushes; 238 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0, 239 "Number of dirty buffer flushes done by the bufdaemon helpers"); 240 static long barrierwrites; 241 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 242 "Number of barrier writes"); 243 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 244 &unmapped_buf_allowed, 0, 245 "Permit the use of the unmapped i/o"); 246 247 /* 248 * This lock synchronizes access to bd_request. 249 */ 250 static struct mtx_padalign bdlock; 251 252 /* 253 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 254 * waitrunningbufspace(). 255 */ 256 static struct mtx_padalign rbreqlock; 257 258 /* 259 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 260 */ 261 static struct rwlock_padalign nblock; 262 263 /* 264 * Lock that protects bdirtywait. 265 */ 266 static struct mtx_padalign bdirtylock; 267 268 /* 269 * Wakeup point for bufdaemon, as well as indicator of whether it is already 270 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 271 * is idling. 272 */ 273 static int bd_request; 274 275 /* 276 * Request/wakeup point for the bufspace daemon. 277 */ 278 static int bufspace_request; 279 280 /* 281 * Request for the buf daemon to write more buffers than is indicated by 282 * lodirtybuf. This may be necessary to push out excess dependencies or 283 * defragment the address space where a simple count of the number of dirty 284 * buffers is insufficient to characterize the demand for flushing them. 285 */ 286 static int bd_speedupreq; 287 288 /* 289 * bogus page -- for I/O to/from partially complete buffers 290 * this is a temporary solution to the problem, but it is not 291 * really that bad. it would be better to split the buffer 292 * for input in the case of buffers partially already in memory, 293 * but the code is intricate enough already. 294 */ 295 vm_page_t bogus_page; 296 297 /* 298 * Synchronization (sleep/wakeup) variable for active buffer space requests. 299 * Set when wait starts, cleared prior to wakeup(). 300 * Used in runningbufwakeup() and waitrunningbufspace(). 301 */ 302 static int runningbufreq; 303 304 /* 305 * Synchronization (sleep/wakeup) variable for buffer requests. 306 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 307 * by and/or. 308 * Used in numdirtywakeup(), bufspace_wakeup(), bwillwrite(), 309 * getnewbuf(), and getblk(). 310 */ 311 static volatile int needsbuffer; 312 313 /* 314 * Synchronization for bwillwrite() waiters. 315 */ 316 static int bdirtywait; 317 318 /* 319 * Definitions for the buffer free lists. 320 */ 321 #define QUEUE_NONE 0 /* on no queue */ 322 #define QUEUE_EMPTY 1 /* empty buffer headers */ 323 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 324 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ 325 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 326 327 /* Maximum number of clean buffer queues. */ 328 #define CLEAN_QUEUES 16 329 330 /* Configured number of clean queues. */ 331 static int clean_queues; 332 333 /* Maximum number of buffer queues. */ 334 #define BUFFER_QUEUES (QUEUE_CLEAN + CLEAN_QUEUES) 335 336 /* Queues for free buffers with various properties */ 337 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 338 #ifdef INVARIANTS 339 static int bq_len[BUFFER_QUEUES]; 340 #endif 341 342 /* 343 * Lock for each bufqueue 344 */ 345 static struct mtx_padalign bqlocks[BUFFER_QUEUES]; 346 347 /* 348 * per-cpu empty buffer cache. 349 */ 350 uma_zone_t buf_zone; 351 352 /* 353 * Single global constant for BUF_WMESG, to avoid getting multiple references. 354 * buf_wmesg is referred from macros. 355 */ 356 const char *buf_wmesg = BUF_WMESG; 357 358 static int 359 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 360 { 361 long value; 362 int error; 363 364 value = *(long *)arg1; 365 error = sysctl_handle_long(oidp, &value, 0, req); 366 if (error != 0 || req->newptr == NULL) 367 return (error); 368 mtx_lock(&rbreqlock); 369 if (arg1 == &hirunningspace) { 370 if (value < lorunningspace) 371 error = EINVAL; 372 else 373 hirunningspace = value; 374 } else { 375 KASSERT(arg1 == &lorunningspace, 376 ("%s: unknown arg1", __func__)); 377 if (value > hirunningspace) 378 error = EINVAL; 379 else 380 lorunningspace = value; 381 } 382 mtx_unlock(&rbreqlock); 383 return (error); 384 } 385 386 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 387 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 388 static int 389 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 390 { 391 long lvalue; 392 int ivalue; 393 394 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 395 return (sysctl_handle_long(oidp, arg1, arg2, req)); 396 lvalue = *(long *)arg1; 397 if (lvalue > INT_MAX) 398 /* On overflow, still write out a long to trigger ENOMEM. */ 399 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 400 ivalue = lvalue; 401 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 402 } 403 #endif 404 405 static int 406 bqcleanq(void) 407 { 408 static int nextq; 409 410 return ((atomic_fetchadd_int(&nextq, 1) % clean_queues) + QUEUE_CLEAN); 411 } 412 413 static int 414 bqisclean(int qindex) 415 { 416 417 return (qindex >= QUEUE_CLEAN && qindex < QUEUE_CLEAN + CLEAN_QUEUES); 418 } 419 420 /* 421 * bqlock: 422 * 423 * Return the appropriate queue lock based on the index. 424 */ 425 static inline struct mtx * 426 bqlock(int qindex) 427 { 428 429 return (struct mtx *)&bqlocks[qindex]; 430 } 431 432 /* 433 * bdirtywakeup: 434 * 435 * Wakeup any bwillwrite() waiters. 436 */ 437 static void 438 bdirtywakeup(void) 439 { 440 mtx_lock(&bdirtylock); 441 if (bdirtywait) { 442 bdirtywait = 0; 443 wakeup(&bdirtywait); 444 } 445 mtx_unlock(&bdirtylock); 446 } 447 448 /* 449 * bdirtysub: 450 * 451 * Decrement the numdirtybuffers count by one and wakeup any 452 * threads blocked in bwillwrite(). 453 */ 454 static void 455 bdirtysub(void) 456 { 457 458 if (atomic_fetchadd_int(&numdirtybuffers, -1) == 459 (lodirtybuffers + hidirtybuffers) / 2) 460 bdirtywakeup(); 461 } 462 463 /* 464 * bdirtyadd: 465 * 466 * Increment the numdirtybuffers count by one and wakeup the buf 467 * daemon if needed. 468 */ 469 static void 470 bdirtyadd(void) 471 { 472 473 /* 474 * Only do the wakeup once as we cross the boundary. The 475 * buf daemon will keep running until the condition clears. 476 */ 477 if (atomic_fetchadd_int(&numdirtybuffers, 1) == 478 (lodirtybuffers + hidirtybuffers) / 2) 479 bd_wakeup(); 480 } 481 482 /* 483 * bufspace_wakeup: 484 * 485 * Called when buffer space is potentially available for recovery. 486 * getnewbuf() will block on this flag when it is unable to free 487 * sufficient buffer space. Buffer space becomes recoverable when 488 * bp's get placed back in the queues. 489 */ 490 static void 491 bufspace_wakeup(void) 492 { 493 494 /* 495 * If someone is waiting for bufspace, wake them up. 496 * 497 * Since needsbuffer is set prior to doing an additional queue 498 * scan it is safe to check for the flag prior to acquiring the 499 * lock. The thread that is preparing to scan again before 500 * blocking would discover the buf we released. 501 */ 502 if (needsbuffer) { 503 rw_rlock(&nblock); 504 if (atomic_cmpset_int(&needsbuffer, 1, 0) == 1) 505 wakeup(__DEVOLATILE(void *, &needsbuffer)); 506 rw_runlock(&nblock); 507 } 508 } 509 510 /* 511 * bufspace_daemonwakeup: 512 * 513 * Wakeup the daemon responsible for freeing clean bufs. 514 */ 515 static void 516 bufspace_daemonwakeup(void) 517 { 518 rw_rlock(&nblock); 519 if (bufspace_request == 0) { 520 bufspace_request = 1; 521 wakeup(&bufspace_request); 522 } 523 rw_runlock(&nblock); 524 } 525 526 /* 527 * bufspace_adjust: 528 * 529 * Adjust the reported bufspace for a KVA managed buffer, possibly 530 * waking any waiters. 531 */ 532 static void 533 bufspace_adjust(struct buf *bp, int bufsize) 534 { 535 long space; 536 int diff; 537 538 KASSERT((bp->b_flags & B_MALLOC) == 0, 539 ("bufspace_adjust: malloc buf %p", bp)); 540 diff = bufsize - bp->b_bufsize; 541 if (diff < 0) { 542 atomic_subtract_long(&bufspace, -diff); 543 bufspace_wakeup(); 544 } else { 545 space = atomic_fetchadd_long(&bufspace, diff); 546 /* Wake up the daemon on the transition. */ 547 if (space < bufspacethresh && space + diff >= bufspacethresh) 548 bufspace_daemonwakeup(); 549 } 550 bp->b_bufsize = bufsize; 551 } 552 553 /* 554 * bufspace_reserve: 555 * 556 * Reserve bufspace before calling allocbuf(). metadata has a 557 * different space limit than data. 558 */ 559 static int 560 bufspace_reserve(int size, bool metadata) 561 { 562 long limit; 563 long space; 564 565 if (metadata) 566 limit = maxbufspace; 567 else 568 limit = hibufspace; 569 do { 570 space = bufspace; 571 if (space + size > limit) 572 return (ENOSPC); 573 } while (atomic_cmpset_long(&bufspace, space, space + size) == 0); 574 575 /* Wake up the daemon on the transition. */ 576 if (space < bufspacethresh && space + size >= bufspacethresh) 577 bufspace_daemonwakeup(); 578 579 return (0); 580 } 581 582 /* 583 * bufspace_release: 584 * 585 * Release reserved bufspace after bufspace_adjust() has consumed it. 586 */ 587 static void 588 bufspace_release(int size) 589 { 590 atomic_subtract_long(&bufspace, size); 591 bufspace_wakeup(); 592 } 593 594 /* 595 * bufspace_wait: 596 * 597 * Wait for bufspace, acting as the buf daemon if a locked vnode is 598 * supplied. needsbuffer must be set in a safe fashion prior to 599 * polling for space. The operation must be re-tried on return. 600 */ 601 static void 602 bufspace_wait(struct vnode *vp, int gbflags, int slpflag, int slptimeo) 603 { 604 struct thread *td; 605 int error, fl, norunbuf; 606 607 if ((gbflags & GB_NOWAIT_BD) != 0) 608 return; 609 610 td = curthread; 611 rw_wlock(&nblock); 612 while (needsbuffer != 0) { 613 if (vp != NULL && vp->v_type != VCHR && 614 (td->td_pflags & TDP_BUFNEED) == 0) { 615 rw_wunlock(&nblock); 616 /* 617 * getblk() is called with a vnode locked, and 618 * some majority of the dirty buffers may as 619 * well belong to the vnode. Flushing the 620 * buffers there would make a progress that 621 * cannot be achieved by the buf_daemon, that 622 * cannot lock the vnode. 623 */ 624 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 625 (td->td_pflags & TDP_NORUNNINGBUF); 626 627 /* 628 * Play bufdaemon. The getnewbuf() function 629 * may be called while the thread owns lock 630 * for another dirty buffer for the same 631 * vnode, which makes it impossible to use 632 * VOP_FSYNC() there, due to the buffer lock 633 * recursion. 634 */ 635 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 636 fl = buf_flush(vp, flushbufqtarget); 637 td->td_pflags &= norunbuf; 638 rw_wlock(&nblock); 639 if (fl != 0) 640 continue; 641 if (needsbuffer == 0) 642 break; 643 } 644 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 645 (PRIBIO + 4) | slpflag, "newbuf", slptimeo); 646 if (error != 0) 647 break; 648 } 649 rw_wunlock(&nblock); 650 } 651 652 653 /* 654 * bufspace_daemon: 655 * 656 * buffer space management daemon. Tries to maintain some marginal 657 * amount of free buffer space so that requesting processes neither 658 * block nor work to reclaim buffers. 659 */ 660 static void 661 bufspace_daemon(void) 662 { 663 for (;;) { 664 kproc_suspend_check(bufspacedaemonproc); 665 666 /* 667 * Free buffers from the clean queue until we meet our 668 * targets. 669 * 670 * Theory of operation: The buffer cache is most efficient 671 * when some free buffer headers and space are always 672 * available to getnewbuf(). This daemon attempts to prevent 673 * the excessive blocking and synchronization associated 674 * with shortfall. It goes through three phases according 675 * demand: 676 * 677 * 1) The daemon wakes up voluntarily once per-second 678 * during idle periods when the counters are below 679 * the wakeup thresholds (bufspacethresh, lofreebuffers). 680 * 681 * 2) The daemon wakes up as we cross the thresholds 682 * ahead of any potential blocking. This may bounce 683 * slightly according to the rate of consumption and 684 * release. 685 * 686 * 3) The daemon and consumers are starved for working 687 * clean buffers. This is the 'bufspace' sleep below 688 * which will inefficiently trade bufs with bqrelse 689 * until we return to condition 2. 690 */ 691 while (bufspace > lobufspace || 692 numfreebuffers < hifreebuffers) { 693 if (buf_recycle(false) != 0) { 694 atomic_set_int(&needsbuffer, 1); 695 if (buf_recycle(false) != 0) { 696 rw_wlock(&nblock); 697 if (needsbuffer) 698 rw_sleep(__DEVOLATILE(void *, 699 &needsbuffer), &nblock, 700 PRIBIO|PDROP, "bufspace", 701 hz/10); 702 else 703 rw_wunlock(&nblock); 704 } 705 } 706 maybe_yield(); 707 } 708 709 /* 710 * Re-check our limits under the exclusive nblock. 711 */ 712 rw_wlock(&nblock); 713 if (bufspace < bufspacethresh && 714 numfreebuffers > lofreebuffers) { 715 bufspace_request = 0; 716 rw_sleep(&bufspace_request, &nblock, PRIBIO|PDROP, 717 "-", hz); 718 } else 719 rw_wunlock(&nblock); 720 } 721 } 722 723 static struct kproc_desc bufspace_kp = { 724 "bufspacedaemon", 725 bufspace_daemon, 726 &bufspacedaemonproc 727 }; 728 SYSINIT(bufspacedaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, 729 &bufspace_kp); 730 731 /* 732 * bufmallocadjust: 733 * 734 * Adjust the reported bufspace for a malloc managed buffer, possibly 735 * waking any waiters. 736 */ 737 static void 738 bufmallocadjust(struct buf *bp, int bufsize) 739 { 740 int diff; 741 742 KASSERT((bp->b_flags & B_MALLOC) != 0, 743 ("bufmallocadjust: non-malloc buf %p", bp)); 744 diff = bufsize - bp->b_bufsize; 745 if (diff < 0) 746 atomic_subtract_long(&bufmallocspace, -diff); 747 else 748 atomic_add_long(&bufmallocspace, diff); 749 bp->b_bufsize = bufsize; 750 } 751 752 /* 753 * runningwakeup: 754 * 755 * Wake up processes that are waiting on asynchronous writes to fall 756 * below lorunningspace. 757 */ 758 static void 759 runningwakeup(void) 760 { 761 762 mtx_lock(&rbreqlock); 763 if (runningbufreq) { 764 runningbufreq = 0; 765 wakeup(&runningbufreq); 766 } 767 mtx_unlock(&rbreqlock); 768 } 769 770 /* 771 * runningbufwakeup: 772 * 773 * Decrement the outstanding write count according. 774 */ 775 void 776 runningbufwakeup(struct buf *bp) 777 { 778 long space, bspace; 779 780 bspace = bp->b_runningbufspace; 781 if (bspace == 0) 782 return; 783 space = atomic_fetchadd_long(&runningbufspace, -bspace); 784 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 785 space, bspace)); 786 bp->b_runningbufspace = 0; 787 /* 788 * Only acquire the lock and wakeup on the transition from exceeding 789 * the threshold to falling below it. 790 */ 791 if (space < lorunningspace) 792 return; 793 if (space - bspace > lorunningspace) 794 return; 795 runningwakeup(); 796 } 797 798 /* 799 * waitrunningbufspace() 800 * 801 * runningbufspace is a measure of the amount of I/O currently 802 * running. This routine is used in async-write situations to 803 * prevent creating huge backups of pending writes to a device. 804 * Only asynchronous writes are governed by this function. 805 * 806 * This does NOT turn an async write into a sync write. It waits 807 * for earlier writes to complete and generally returns before the 808 * caller's write has reached the device. 809 */ 810 void 811 waitrunningbufspace(void) 812 { 813 814 mtx_lock(&rbreqlock); 815 while (runningbufspace > hirunningspace) { 816 runningbufreq = 1; 817 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 818 } 819 mtx_unlock(&rbreqlock); 820 } 821 822 823 /* 824 * vfs_buf_test_cache: 825 * 826 * Called when a buffer is extended. This function clears the B_CACHE 827 * bit if the newly extended portion of the buffer does not contain 828 * valid data. 829 */ 830 static __inline void 831 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 832 vm_offset_t size, vm_page_t m) 833 { 834 835 VM_OBJECT_ASSERT_LOCKED(m->object); 836 if (bp->b_flags & B_CACHE) { 837 int base = (foff + off) & PAGE_MASK; 838 if (vm_page_is_valid(m, base, size) == 0) 839 bp->b_flags &= ~B_CACHE; 840 } 841 } 842 843 /* Wake up the buffer daemon if necessary */ 844 static __inline void 845 bd_wakeup(void) 846 { 847 848 mtx_lock(&bdlock); 849 if (bd_request == 0) { 850 bd_request = 1; 851 wakeup(&bd_request); 852 } 853 mtx_unlock(&bdlock); 854 } 855 856 /* 857 * bd_speedup - speedup the buffer cache flushing code 858 */ 859 void 860 bd_speedup(void) 861 { 862 int needwake; 863 864 mtx_lock(&bdlock); 865 needwake = 0; 866 if (bd_speedupreq == 0 || bd_request == 0) 867 needwake = 1; 868 bd_speedupreq = 1; 869 bd_request = 1; 870 if (needwake) 871 wakeup(&bd_request); 872 mtx_unlock(&bdlock); 873 } 874 875 #ifndef NSWBUF_MIN 876 #define NSWBUF_MIN 16 877 #endif 878 879 #ifdef __i386__ 880 #define TRANSIENT_DENOM 5 881 #else 882 #define TRANSIENT_DENOM 10 883 #endif 884 885 /* 886 * Calculating buffer cache scaling values and reserve space for buffer 887 * headers. This is called during low level kernel initialization and 888 * may be called more then once. We CANNOT write to the memory area 889 * being reserved at this time. 890 */ 891 caddr_t 892 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 893 { 894 int tuned_nbuf; 895 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 896 897 /* 898 * physmem_est is in pages. Convert it to kilobytes (assumes 899 * PAGE_SIZE is >= 1K) 900 */ 901 physmem_est = physmem_est * (PAGE_SIZE / 1024); 902 903 /* 904 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 905 * For the first 64MB of ram nominally allocate sufficient buffers to 906 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 907 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 908 * the buffer cache we limit the eventual kva reservation to 909 * maxbcache bytes. 910 * 911 * factor represents the 1/4 x ram conversion. 912 */ 913 if (nbuf == 0) { 914 int factor = 4 * BKVASIZE / 1024; 915 916 nbuf = 50; 917 if (physmem_est > 4096) 918 nbuf += min((physmem_est - 4096) / factor, 919 65536 / factor); 920 if (physmem_est > 65536) 921 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 922 32 * 1024 * 1024 / (factor * 5)); 923 924 if (maxbcache && nbuf > maxbcache / BKVASIZE) 925 nbuf = maxbcache / BKVASIZE; 926 tuned_nbuf = 1; 927 } else 928 tuned_nbuf = 0; 929 930 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 931 maxbuf = (LONG_MAX / 3) / BKVASIZE; 932 if (nbuf > maxbuf) { 933 if (!tuned_nbuf) 934 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 935 maxbuf); 936 nbuf = maxbuf; 937 } 938 939 /* 940 * Ideal allocation size for the transient bio submap is 10% 941 * of the maximal space buffer map. This roughly corresponds 942 * to the amount of the buffer mapped for typical UFS load. 943 * 944 * Clip the buffer map to reserve space for the transient 945 * BIOs, if its extent is bigger than 90% (80% on i386) of the 946 * maximum buffer map extent on the platform. 947 * 948 * The fall-back to the maxbuf in case of maxbcache unset, 949 * allows to not trim the buffer KVA for the architectures 950 * with ample KVA space. 951 */ 952 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 953 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 954 buf_sz = (long)nbuf * BKVASIZE; 955 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 956 (TRANSIENT_DENOM - 1)) { 957 /* 958 * There is more KVA than memory. Do not 959 * adjust buffer map size, and assign the rest 960 * of maxbuf to transient map. 961 */ 962 biotmap_sz = maxbuf_sz - buf_sz; 963 } else { 964 /* 965 * Buffer map spans all KVA we could afford on 966 * this platform. Give 10% (20% on i386) of 967 * the buffer map to the transient bio map. 968 */ 969 biotmap_sz = buf_sz / TRANSIENT_DENOM; 970 buf_sz -= biotmap_sz; 971 } 972 if (biotmap_sz / INT_MAX > MAXPHYS) 973 bio_transient_maxcnt = INT_MAX; 974 else 975 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 976 /* 977 * Artifically limit to 1024 simultaneous in-flight I/Os 978 * using the transient mapping. 979 */ 980 if (bio_transient_maxcnt > 1024) 981 bio_transient_maxcnt = 1024; 982 if (tuned_nbuf) 983 nbuf = buf_sz / BKVASIZE; 984 } 985 986 /* 987 * swbufs are used as temporary holders for I/O, such as paging I/O. 988 * We have no less then 16 and no more then 256. 989 */ 990 nswbuf = min(nbuf / 4, 256); 991 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf); 992 if (nswbuf < NSWBUF_MIN) 993 nswbuf = NSWBUF_MIN; 994 995 /* 996 * Reserve space for the buffer cache buffers 997 */ 998 swbuf = (void *)v; 999 v = (caddr_t)(swbuf + nswbuf); 1000 buf = (void *)v; 1001 v = (caddr_t)(buf + nbuf); 1002 1003 return(v); 1004 } 1005 1006 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1007 void 1008 bufinit(void) 1009 { 1010 struct buf *bp; 1011 int i; 1012 1013 CTASSERT(MAXBCACHEBUF >= MAXBSIZE); 1014 mtx_init(&bqlocks[QUEUE_DIRTY], "bufq dirty lock", NULL, MTX_DEF); 1015 mtx_init(&bqlocks[QUEUE_EMPTY], "bufq empty lock", NULL, MTX_DEF); 1016 for (i = QUEUE_CLEAN; i < QUEUE_CLEAN + CLEAN_QUEUES; i++) 1017 mtx_init(&bqlocks[i], "bufq clean lock", NULL, MTX_DEF); 1018 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1019 rw_init(&nblock, "needsbuffer lock"); 1020 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1021 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1022 1023 /* next, make a null set of free lists */ 1024 for (i = 0; i < BUFFER_QUEUES; i++) 1025 TAILQ_INIT(&bufqueues[i]); 1026 1027 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 1028 1029 /* finally, initialize each buffer header and stick on empty q */ 1030 for (i = 0; i < nbuf; i++) { 1031 bp = &buf[i]; 1032 bzero(bp, sizeof *bp); 1033 bp->b_flags = B_INVAL; 1034 bp->b_rcred = NOCRED; 1035 bp->b_wcred = NOCRED; 1036 bp->b_qindex = QUEUE_EMPTY; 1037 bp->b_xflags = 0; 1038 bp->b_data = bp->b_kvabase = unmapped_buf; 1039 LIST_INIT(&bp->b_dep); 1040 BUF_LOCKINIT(bp); 1041 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 1042 #ifdef INVARIANTS 1043 bq_len[QUEUE_EMPTY]++; 1044 #endif 1045 } 1046 1047 /* 1048 * maxbufspace is the absolute maximum amount of buffer space we are 1049 * allowed to reserve in KVM and in real terms. The absolute maximum 1050 * is nominally used by metadata. hibufspace is the nominal maximum 1051 * used by most other requests. The differential is required to 1052 * ensure that metadata deadlocks don't occur. 1053 * 1054 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1055 * this may result in KVM fragmentation which is not handled optimally 1056 * by the system. XXX This is less true with vmem. We could use 1057 * PAGE_SIZE. 1058 */ 1059 maxbufspace = (long)nbuf * BKVASIZE; 1060 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10); 1061 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1062 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1063 1064 /* 1065 * Note: The 16 MiB upper limit for hirunningspace was chosen 1066 * arbitrarily and may need further tuning. It corresponds to 1067 * 128 outstanding write IO requests (if IO size is 128 KiB), 1068 * which fits with many RAID controllers' tagged queuing limits. 1069 * The lower 1 MiB limit is the historical upper limit for 1070 * hirunningspace. 1071 */ 1072 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF), 1073 16 * 1024 * 1024), 1024 * 1024); 1074 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF); 1075 1076 /* 1077 * Limit the amount of malloc memory since it is wired permanently into 1078 * the kernel space. Even though this is accounted for in the buffer 1079 * allocation, we don't want the malloced region to grow uncontrolled. 1080 * The malloc scheme improves memory utilization significantly on 1081 * average (small) directories. 1082 */ 1083 maxbufmallocspace = hibufspace / 20; 1084 1085 /* 1086 * Reduce the chance of a deadlock occuring by limiting the number 1087 * of delayed-write dirty buffers we allow to stack up. 1088 */ 1089 hidirtybuffers = nbuf / 4 + 20; 1090 dirtybufthresh = hidirtybuffers * 9 / 10; 1091 numdirtybuffers = 0; 1092 /* 1093 * To support extreme low-memory systems, make sure hidirtybuffers 1094 * cannot eat up all available buffer space. This occurs when our 1095 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1096 * buffer space assuming BKVASIZE'd buffers. 1097 */ 1098 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1099 hidirtybuffers >>= 1; 1100 } 1101 lodirtybuffers = hidirtybuffers / 2; 1102 1103 /* 1104 * lofreebuffers should be sufficient to avoid stalling waiting on 1105 * buf headers under heavy utilization. The bufs in per-cpu caches 1106 * are counted as free but will be unavailable to threads executing 1107 * on other cpus. 1108 * 1109 * hifreebuffers is the free target for the bufspace daemon. This 1110 * should be set appropriately to limit work per-iteration. 1111 */ 1112 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1113 hifreebuffers = (3 * lofreebuffers) / 2; 1114 numfreebuffers = nbuf; 1115 1116 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 1117 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 1118 1119 /* Setup the kva and free list allocators. */ 1120 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1121 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), 1122 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1123 1124 /* 1125 * Size the clean queue according to the amount of buffer space. 1126 * One queue per-256mb up to the max. More queues gives better 1127 * concurrency but less accurate LRU. 1128 */ 1129 clean_queues = MIN(howmany(maxbufspace, 256*1024*1024), CLEAN_QUEUES); 1130 1131 } 1132 1133 #ifdef INVARIANTS 1134 static inline void 1135 vfs_buf_check_mapped(struct buf *bp) 1136 { 1137 1138 KASSERT(bp->b_kvabase != unmapped_buf, 1139 ("mapped buf: b_kvabase was not updated %p", bp)); 1140 KASSERT(bp->b_data != unmapped_buf, 1141 ("mapped buf: b_data was not updated %p", bp)); 1142 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1143 MAXPHYS, ("b_data + b_offset unmapped %p", bp)); 1144 } 1145 1146 static inline void 1147 vfs_buf_check_unmapped(struct buf *bp) 1148 { 1149 1150 KASSERT(bp->b_data == unmapped_buf, 1151 ("unmapped buf: corrupted b_data %p", bp)); 1152 } 1153 1154 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1155 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1156 #else 1157 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1158 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1159 #endif 1160 1161 static int 1162 isbufbusy(struct buf *bp) 1163 { 1164 if (((bp->b_flags & (B_INVAL | B_PERSISTENT)) == 0 && 1165 BUF_ISLOCKED(bp)) || 1166 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1167 return (1); 1168 return (0); 1169 } 1170 1171 /* 1172 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1173 */ 1174 void 1175 bufshutdown(int show_busybufs) 1176 { 1177 static int first_buf_printf = 1; 1178 struct buf *bp; 1179 int iter, nbusy, pbusy; 1180 #ifndef PREEMPTION 1181 int subiter; 1182 #endif 1183 1184 /* 1185 * Sync filesystems for shutdown 1186 */ 1187 wdog_kern_pat(WD_LASTVAL); 1188 sys_sync(curthread, NULL); 1189 1190 /* 1191 * With soft updates, some buffers that are 1192 * written will be remarked as dirty until other 1193 * buffers are written. 1194 */ 1195 for (iter = pbusy = 0; iter < 20; iter++) { 1196 nbusy = 0; 1197 for (bp = &buf[nbuf]; --bp >= buf; ) 1198 if (isbufbusy(bp)) 1199 nbusy++; 1200 if (nbusy == 0) { 1201 if (first_buf_printf) 1202 printf("All buffers synced."); 1203 break; 1204 } 1205 if (first_buf_printf) { 1206 printf("Syncing disks, buffers remaining... "); 1207 first_buf_printf = 0; 1208 } 1209 printf("%d ", nbusy); 1210 if (nbusy < pbusy) 1211 iter = 0; 1212 pbusy = nbusy; 1213 1214 wdog_kern_pat(WD_LASTVAL); 1215 sys_sync(curthread, NULL); 1216 1217 #ifdef PREEMPTION 1218 /* 1219 * Drop Giant and spin for a while to allow 1220 * interrupt threads to run. 1221 */ 1222 DROP_GIANT(); 1223 DELAY(50000 * iter); 1224 PICKUP_GIANT(); 1225 #else 1226 /* 1227 * Drop Giant and context switch several times to 1228 * allow interrupt threads to run. 1229 */ 1230 DROP_GIANT(); 1231 for (subiter = 0; subiter < 50 * iter; subiter++) { 1232 thread_lock(curthread); 1233 mi_switch(SW_VOL, NULL); 1234 thread_unlock(curthread); 1235 DELAY(1000); 1236 } 1237 PICKUP_GIANT(); 1238 #endif 1239 } 1240 printf("\n"); 1241 /* 1242 * Count only busy local buffers to prevent forcing 1243 * a fsck if we're just a client of a wedged NFS server 1244 */ 1245 nbusy = 0; 1246 for (bp = &buf[nbuf]; --bp >= buf; ) { 1247 if (isbufbusy(bp)) { 1248 #if 0 1249 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1250 if (bp->b_dev == NULL) { 1251 TAILQ_REMOVE(&mountlist, 1252 bp->b_vp->v_mount, mnt_list); 1253 continue; 1254 } 1255 #endif 1256 nbusy++; 1257 if (show_busybufs > 0) { 1258 printf( 1259 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1260 nbusy, bp, bp->b_vp, bp->b_flags, 1261 (intmax_t)bp->b_blkno, 1262 (intmax_t)bp->b_lblkno); 1263 BUF_LOCKPRINTINFO(bp); 1264 if (show_busybufs > 1) 1265 vn_printf(bp->b_vp, 1266 "vnode content: "); 1267 } 1268 } 1269 } 1270 if (nbusy) { 1271 /* 1272 * Failed to sync all blocks. Indicate this and don't 1273 * unmount filesystems (thus forcing an fsck on reboot). 1274 */ 1275 printf("Giving up on %d buffers\n", nbusy); 1276 DELAY(5000000); /* 5 seconds */ 1277 } else { 1278 if (!first_buf_printf) 1279 printf("Final sync complete\n"); 1280 /* 1281 * Unmount filesystems 1282 */ 1283 if (panicstr == NULL) 1284 vfs_unmountall(); 1285 } 1286 swapoff_all(); 1287 DELAY(100000); /* wait for console output to finish */ 1288 } 1289 1290 static void 1291 bpmap_qenter(struct buf *bp) 1292 { 1293 1294 BUF_CHECK_MAPPED(bp); 1295 1296 /* 1297 * bp->b_data is relative to bp->b_offset, but 1298 * bp->b_offset may be offset into the first page. 1299 */ 1300 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1301 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1302 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1303 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1304 } 1305 1306 /* 1307 * binsfree: 1308 * 1309 * Insert the buffer into the appropriate free list. 1310 */ 1311 static void 1312 binsfree(struct buf *bp, int qindex) 1313 { 1314 struct mtx *olock, *nlock; 1315 1316 if (qindex != QUEUE_EMPTY) { 1317 BUF_ASSERT_XLOCKED(bp); 1318 } 1319 1320 /* 1321 * Stick to the same clean queue for the lifetime of the buf to 1322 * limit locking below. Otherwise pick ont sequentially. 1323 */ 1324 if (qindex == QUEUE_CLEAN) { 1325 if (bqisclean(bp->b_qindex)) 1326 qindex = bp->b_qindex; 1327 else 1328 qindex = bqcleanq(); 1329 } 1330 1331 /* 1332 * Handle delayed bremfree() processing. 1333 */ 1334 nlock = bqlock(qindex); 1335 if (bp->b_flags & B_REMFREE) { 1336 olock = bqlock(bp->b_qindex); 1337 mtx_lock(olock); 1338 bremfreel(bp); 1339 if (olock != nlock) { 1340 mtx_unlock(olock); 1341 mtx_lock(nlock); 1342 } 1343 } else 1344 mtx_lock(nlock); 1345 1346 if (bp->b_qindex != QUEUE_NONE) 1347 panic("binsfree: free buffer onto another queue???"); 1348 1349 bp->b_qindex = qindex; 1350 if (bp->b_flags & B_AGE) 1351 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1352 else 1353 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1354 #ifdef INVARIANTS 1355 bq_len[bp->b_qindex]++; 1356 #endif 1357 mtx_unlock(nlock); 1358 } 1359 1360 /* 1361 * buf_free: 1362 * 1363 * Free a buffer to the buf zone once it no longer has valid contents. 1364 */ 1365 static void 1366 buf_free(struct buf *bp) 1367 { 1368 1369 if (bp->b_flags & B_REMFREE) 1370 bremfreef(bp); 1371 if (bp->b_vflags & BV_BKGRDINPROG) 1372 panic("losing buffer 1"); 1373 if (bp->b_rcred != NOCRED) { 1374 crfree(bp->b_rcred); 1375 bp->b_rcred = NOCRED; 1376 } 1377 if (bp->b_wcred != NOCRED) { 1378 crfree(bp->b_wcred); 1379 bp->b_wcred = NOCRED; 1380 } 1381 if (!LIST_EMPTY(&bp->b_dep)) 1382 buf_deallocate(bp); 1383 bufkva_free(bp); 1384 BUF_UNLOCK(bp); 1385 uma_zfree(buf_zone, bp); 1386 atomic_add_int(&numfreebuffers, 1); 1387 bufspace_wakeup(); 1388 } 1389 1390 /* 1391 * buf_import: 1392 * 1393 * Import bufs into the uma cache from the buf list. The system still 1394 * expects a static array of bufs and much of the synchronization 1395 * around bufs assumes type stable storage. As a result, UMA is used 1396 * only as a per-cpu cache of bufs still maintained on a global list. 1397 */ 1398 static int 1399 buf_import(void *arg, void **store, int cnt, int flags) 1400 { 1401 struct buf *bp; 1402 int i; 1403 1404 mtx_lock(&bqlocks[QUEUE_EMPTY]); 1405 for (i = 0; i < cnt; i++) { 1406 bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1407 if (bp == NULL) 1408 break; 1409 bremfreel(bp); 1410 store[i] = bp; 1411 } 1412 mtx_unlock(&bqlocks[QUEUE_EMPTY]); 1413 1414 return (i); 1415 } 1416 1417 /* 1418 * buf_release: 1419 * 1420 * Release bufs from the uma cache back to the buffer queues. 1421 */ 1422 static void 1423 buf_release(void *arg, void **store, int cnt) 1424 { 1425 int i; 1426 1427 for (i = 0; i < cnt; i++) 1428 binsfree(store[i], QUEUE_EMPTY); 1429 } 1430 1431 /* 1432 * buf_alloc: 1433 * 1434 * Allocate an empty buffer header. 1435 */ 1436 static struct buf * 1437 buf_alloc(void) 1438 { 1439 struct buf *bp; 1440 1441 bp = uma_zalloc(buf_zone, M_NOWAIT); 1442 if (bp == NULL) { 1443 bufspace_daemonwakeup(); 1444 atomic_add_int(&numbufallocfails, 1); 1445 return (NULL); 1446 } 1447 1448 /* 1449 * Wake-up the bufspace daemon on transition. 1450 */ 1451 if (atomic_fetchadd_int(&numfreebuffers, -1) == lofreebuffers) 1452 bufspace_daemonwakeup(); 1453 1454 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1455 panic("getnewbuf_empty: Locked buf %p on free queue.", bp); 1456 1457 KASSERT(bp->b_vp == NULL, 1458 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1459 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1460 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1461 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1462 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1463 KASSERT(bp->b_npages == 0, 1464 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1465 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1466 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1467 1468 bp->b_flags = 0; 1469 bp->b_ioflags = 0; 1470 bp->b_xflags = 0; 1471 bp->b_vflags = 0; 1472 bp->b_vp = NULL; 1473 bp->b_blkno = bp->b_lblkno = 0; 1474 bp->b_offset = NOOFFSET; 1475 bp->b_iodone = 0; 1476 bp->b_error = 0; 1477 bp->b_resid = 0; 1478 bp->b_bcount = 0; 1479 bp->b_npages = 0; 1480 bp->b_dirtyoff = bp->b_dirtyend = 0; 1481 bp->b_bufobj = NULL; 1482 bp->b_pin_count = 0; 1483 bp->b_data = bp->b_kvabase = unmapped_buf; 1484 bp->b_fsprivate1 = NULL; 1485 bp->b_fsprivate2 = NULL; 1486 bp->b_fsprivate3 = NULL; 1487 LIST_INIT(&bp->b_dep); 1488 1489 return (bp); 1490 } 1491 1492 /* 1493 * buf_qrecycle: 1494 * 1495 * Free a buffer from the given bufqueue. kva controls whether the 1496 * freed buf must own some kva resources. This is used for 1497 * defragmenting. 1498 */ 1499 static int 1500 buf_qrecycle(int qindex, bool kva) 1501 { 1502 struct buf *bp, *nbp; 1503 1504 if (kva) 1505 atomic_add_int(&bufdefragcnt, 1); 1506 nbp = NULL; 1507 mtx_lock(&bqlocks[qindex]); 1508 nbp = TAILQ_FIRST(&bufqueues[qindex]); 1509 1510 /* 1511 * Run scan, possibly freeing data and/or kva mappings on the fly 1512 * depending. 1513 */ 1514 while ((bp = nbp) != NULL) { 1515 /* 1516 * Calculate next bp (we can only use it if we do not 1517 * release the bqlock). 1518 */ 1519 nbp = TAILQ_NEXT(bp, b_freelist); 1520 1521 /* 1522 * If we are defragging then we need a buffer with 1523 * some kva to reclaim. 1524 */ 1525 if (kva && bp->b_kvasize == 0) 1526 continue; 1527 1528 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1529 continue; 1530 1531 /* 1532 * Skip buffers with background writes in progress. 1533 */ 1534 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1535 BUF_UNLOCK(bp); 1536 continue; 1537 } 1538 1539 KASSERT(bp->b_qindex == qindex, 1540 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 1541 /* 1542 * NOTE: nbp is now entirely invalid. We can only restart 1543 * the scan from this point on. 1544 */ 1545 bremfreel(bp); 1546 mtx_unlock(&bqlocks[qindex]); 1547 1548 /* 1549 * Requeue the background write buffer with error and 1550 * restart the scan. 1551 */ 1552 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1553 bqrelse(bp); 1554 mtx_lock(&bqlocks[qindex]); 1555 nbp = TAILQ_FIRST(&bufqueues[qindex]); 1556 continue; 1557 } 1558 bp->b_flags |= B_INVAL; 1559 brelse(bp); 1560 return (0); 1561 } 1562 mtx_unlock(&bqlocks[qindex]); 1563 1564 return (ENOBUFS); 1565 } 1566 1567 /* 1568 * buf_recycle: 1569 * 1570 * Iterate through all clean queues until we find a buf to recycle or 1571 * exhaust the search. 1572 */ 1573 static int 1574 buf_recycle(bool kva) 1575 { 1576 int qindex, first_qindex; 1577 1578 qindex = first_qindex = bqcleanq(); 1579 do { 1580 if (buf_qrecycle(qindex, kva) == 0) 1581 return (0); 1582 if (++qindex == QUEUE_CLEAN + clean_queues) 1583 qindex = QUEUE_CLEAN; 1584 } while (qindex != first_qindex); 1585 1586 return (ENOBUFS); 1587 } 1588 1589 /* 1590 * buf_scan: 1591 * 1592 * Scan the clean queues looking for a buffer to recycle. needsbuffer 1593 * is set on failure so that the caller may optionally bufspace_wait() 1594 * in a race-free fashion. 1595 */ 1596 static int 1597 buf_scan(bool defrag) 1598 { 1599 int error; 1600 1601 /* 1602 * To avoid heavy synchronization and wakeup races we set 1603 * needsbuffer and re-poll before failing. This ensures that 1604 * no frees can be missed between an unsuccessful poll and 1605 * going to sleep in a synchronized fashion. 1606 */ 1607 if ((error = buf_recycle(defrag)) != 0) { 1608 atomic_set_int(&needsbuffer, 1); 1609 bufspace_daemonwakeup(); 1610 error = buf_recycle(defrag); 1611 } 1612 if (error == 0) 1613 atomic_add_int(&getnewbufrestarts, 1); 1614 return (error); 1615 } 1616 1617 /* 1618 * bremfree: 1619 * 1620 * Mark the buffer for removal from the appropriate free list. 1621 * 1622 */ 1623 void 1624 bremfree(struct buf *bp) 1625 { 1626 1627 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1628 KASSERT((bp->b_flags & B_REMFREE) == 0, 1629 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1630 KASSERT(bp->b_qindex != QUEUE_NONE, 1631 ("bremfree: buffer %p not on a queue.", bp)); 1632 BUF_ASSERT_XLOCKED(bp); 1633 1634 bp->b_flags |= B_REMFREE; 1635 } 1636 1637 /* 1638 * bremfreef: 1639 * 1640 * Force an immediate removal from a free list. Used only in nfs when 1641 * it abuses the b_freelist pointer. 1642 */ 1643 void 1644 bremfreef(struct buf *bp) 1645 { 1646 struct mtx *qlock; 1647 1648 qlock = bqlock(bp->b_qindex); 1649 mtx_lock(qlock); 1650 bremfreel(bp); 1651 mtx_unlock(qlock); 1652 } 1653 1654 /* 1655 * bremfreel: 1656 * 1657 * Removes a buffer from the free list, must be called with the 1658 * correct qlock held. 1659 */ 1660 static void 1661 bremfreel(struct buf *bp) 1662 { 1663 1664 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1665 bp, bp->b_vp, bp->b_flags); 1666 KASSERT(bp->b_qindex != QUEUE_NONE, 1667 ("bremfreel: buffer %p not on a queue.", bp)); 1668 if (bp->b_qindex != QUEUE_EMPTY) { 1669 BUF_ASSERT_XLOCKED(bp); 1670 } 1671 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1672 1673 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1674 #ifdef INVARIANTS 1675 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1676 bp->b_qindex)); 1677 bq_len[bp->b_qindex]--; 1678 #endif 1679 bp->b_qindex = QUEUE_NONE; 1680 bp->b_flags &= ~B_REMFREE; 1681 } 1682 1683 /* 1684 * bufkva_free: 1685 * 1686 * Free the kva allocation for a buffer. 1687 * 1688 */ 1689 static void 1690 bufkva_free(struct buf *bp) 1691 { 1692 1693 #ifdef INVARIANTS 1694 if (bp->b_kvasize == 0) { 1695 KASSERT(bp->b_kvabase == unmapped_buf && 1696 bp->b_data == unmapped_buf, 1697 ("Leaked KVA space on %p", bp)); 1698 } else if (buf_mapped(bp)) 1699 BUF_CHECK_MAPPED(bp); 1700 else 1701 BUF_CHECK_UNMAPPED(bp); 1702 #endif 1703 if (bp->b_kvasize == 0) 1704 return; 1705 1706 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 1707 atomic_subtract_long(&bufkvaspace, bp->b_kvasize); 1708 atomic_add_int(&buffreekvacnt, 1); 1709 bp->b_data = bp->b_kvabase = unmapped_buf; 1710 bp->b_kvasize = 0; 1711 } 1712 1713 /* 1714 * bufkva_alloc: 1715 * 1716 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 1717 */ 1718 static int 1719 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 1720 { 1721 vm_offset_t addr; 1722 int error; 1723 1724 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 1725 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 1726 1727 bufkva_free(bp); 1728 1729 addr = 0; 1730 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 1731 if (error != 0) { 1732 /* 1733 * Buffer map is too fragmented. Request the caller 1734 * to defragment the map. 1735 */ 1736 return (error); 1737 } 1738 bp->b_kvabase = (caddr_t)addr; 1739 bp->b_kvasize = maxsize; 1740 atomic_add_long(&bufkvaspace, bp->b_kvasize); 1741 if ((gbflags & GB_UNMAPPED) != 0) { 1742 bp->b_data = unmapped_buf; 1743 BUF_CHECK_UNMAPPED(bp); 1744 } else { 1745 bp->b_data = bp->b_kvabase; 1746 BUF_CHECK_MAPPED(bp); 1747 } 1748 return (0); 1749 } 1750 1751 /* 1752 * bufkva_reclaim: 1753 * 1754 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 1755 * callback that fires to avoid returning failure. 1756 */ 1757 static void 1758 bufkva_reclaim(vmem_t *vmem, int flags) 1759 { 1760 int i; 1761 1762 for (i = 0; i < 5; i++) 1763 if (buf_scan(true) != 0) 1764 break; 1765 return; 1766 } 1767 1768 1769 /* 1770 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1771 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1772 * the buffer is valid and we do not have to do anything. 1773 */ 1774 void 1775 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 1776 int cnt, struct ucred * cred) 1777 { 1778 struct buf *rabp; 1779 int i; 1780 1781 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1782 if (inmem(vp, *rablkno)) 1783 continue; 1784 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1785 1786 if ((rabp->b_flags & B_CACHE) == 0) { 1787 if (!TD_IS_IDLETHREAD(curthread)) 1788 curthread->td_ru.ru_inblock++; 1789 rabp->b_flags |= B_ASYNC; 1790 rabp->b_flags &= ~B_INVAL; 1791 rabp->b_ioflags &= ~BIO_ERROR; 1792 rabp->b_iocmd = BIO_READ; 1793 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1794 rabp->b_rcred = crhold(cred); 1795 vfs_busy_pages(rabp, 0); 1796 BUF_KERNPROC(rabp); 1797 rabp->b_iooffset = dbtob(rabp->b_blkno); 1798 bstrategy(rabp); 1799 } else { 1800 brelse(rabp); 1801 } 1802 } 1803 } 1804 1805 /* 1806 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1807 * 1808 * Get a buffer with the specified data. Look in the cache first. We 1809 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1810 * is set, the buffer is valid and we do not have to do anything, see 1811 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1812 */ 1813 int 1814 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1815 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 1816 { 1817 struct buf *bp; 1818 int rv = 0, readwait = 0; 1819 1820 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1821 /* 1822 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1823 */ 1824 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1825 if (bp == NULL) 1826 return (EBUSY); 1827 1828 /* if not found in cache, do some I/O */ 1829 if ((bp->b_flags & B_CACHE) == 0) { 1830 if (!TD_IS_IDLETHREAD(curthread)) 1831 curthread->td_ru.ru_inblock++; 1832 bp->b_iocmd = BIO_READ; 1833 bp->b_flags &= ~B_INVAL; 1834 bp->b_ioflags &= ~BIO_ERROR; 1835 if (bp->b_rcred == NOCRED && cred != NOCRED) 1836 bp->b_rcred = crhold(cred); 1837 vfs_busy_pages(bp, 0); 1838 bp->b_iooffset = dbtob(bp->b_blkno); 1839 bstrategy(bp); 1840 ++readwait; 1841 } 1842 1843 breada(vp, rablkno, rabsize, cnt, cred); 1844 1845 if (readwait) { 1846 rv = bufwait(bp); 1847 } 1848 return (rv); 1849 } 1850 1851 /* 1852 * Write, release buffer on completion. (Done by iodone 1853 * if async). Do not bother writing anything if the buffer 1854 * is invalid. 1855 * 1856 * Note that we set B_CACHE here, indicating that buffer is 1857 * fully valid and thus cacheable. This is true even of NFS 1858 * now so we set it generally. This could be set either here 1859 * or in biodone() since the I/O is synchronous. We put it 1860 * here. 1861 */ 1862 int 1863 bufwrite(struct buf *bp) 1864 { 1865 int oldflags; 1866 struct vnode *vp; 1867 long space; 1868 int vp_md; 1869 1870 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1871 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 1872 bp->b_flags |= B_INVAL | B_RELBUF; 1873 bp->b_flags &= ~B_CACHE; 1874 brelse(bp); 1875 return (ENXIO); 1876 } 1877 if (bp->b_flags & B_INVAL) { 1878 brelse(bp); 1879 return (0); 1880 } 1881 1882 if (bp->b_flags & B_BARRIER) 1883 barrierwrites++; 1884 1885 oldflags = bp->b_flags; 1886 1887 BUF_ASSERT_HELD(bp); 1888 1889 if (bp->b_pin_count > 0) 1890 bunpin_wait(bp); 1891 1892 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1893 ("FFS background buffer should not get here %p", bp)); 1894 1895 vp = bp->b_vp; 1896 if (vp) 1897 vp_md = vp->v_vflag & VV_MD; 1898 else 1899 vp_md = 0; 1900 1901 /* 1902 * Mark the buffer clean. Increment the bufobj write count 1903 * before bundirty() call, to prevent other thread from seeing 1904 * empty dirty list and zero counter for writes in progress, 1905 * falsely indicating that the bufobj is clean. 1906 */ 1907 bufobj_wref(bp->b_bufobj); 1908 bundirty(bp); 1909 1910 bp->b_flags &= ~B_DONE; 1911 bp->b_ioflags &= ~BIO_ERROR; 1912 bp->b_flags |= B_CACHE; 1913 bp->b_iocmd = BIO_WRITE; 1914 1915 vfs_busy_pages(bp, 1); 1916 1917 /* 1918 * Normal bwrites pipeline writes 1919 */ 1920 bp->b_runningbufspace = bp->b_bufsize; 1921 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1922 1923 if (!TD_IS_IDLETHREAD(curthread)) 1924 curthread->td_ru.ru_oublock++; 1925 if (oldflags & B_ASYNC) 1926 BUF_KERNPROC(bp); 1927 bp->b_iooffset = dbtob(bp->b_blkno); 1928 bstrategy(bp); 1929 1930 if ((oldflags & B_ASYNC) == 0) { 1931 int rtval = bufwait(bp); 1932 brelse(bp); 1933 return (rtval); 1934 } else if (space > hirunningspace) { 1935 /* 1936 * don't allow the async write to saturate the I/O 1937 * system. We will not deadlock here because 1938 * we are blocking waiting for I/O that is already in-progress 1939 * to complete. We do not block here if it is the update 1940 * or syncer daemon trying to clean up as that can lead 1941 * to deadlock. 1942 */ 1943 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1944 waitrunningbufspace(); 1945 } 1946 1947 return (0); 1948 } 1949 1950 void 1951 bufbdflush(struct bufobj *bo, struct buf *bp) 1952 { 1953 struct buf *nbp; 1954 1955 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1956 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1957 altbufferflushes++; 1958 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1959 BO_LOCK(bo); 1960 /* 1961 * Try to find a buffer to flush. 1962 */ 1963 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1964 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1965 BUF_LOCK(nbp, 1966 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1967 continue; 1968 if (bp == nbp) 1969 panic("bdwrite: found ourselves"); 1970 BO_UNLOCK(bo); 1971 /* Don't countdeps with the bo lock held. */ 1972 if (buf_countdeps(nbp, 0)) { 1973 BO_LOCK(bo); 1974 BUF_UNLOCK(nbp); 1975 continue; 1976 } 1977 if (nbp->b_flags & B_CLUSTEROK) { 1978 vfs_bio_awrite(nbp); 1979 } else { 1980 bremfree(nbp); 1981 bawrite(nbp); 1982 } 1983 dirtybufferflushes++; 1984 break; 1985 } 1986 if (nbp == NULL) 1987 BO_UNLOCK(bo); 1988 } 1989 } 1990 1991 /* 1992 * Delayed write. (Buffer is marked dirty). Do not bother writing 1993 * anything if the buffer is marked invalid. 1994 * 1995 * Note that since the buffer must be completely valid, we can safely 1996 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1997 * biodone() in order to prevent getblk from writing the buffer 1998 * out synchronously. 1999 */ 2000 void 2001 bdwrite(struct buf *bp) 2002 { 2003 struct thread *td = curthread; 2004 struct vnode *vp; 2005 struct bufobj *bo; 2006 2007 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2008 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2009 KASSERT((bp->b_flags & B_BARRIER) == 0, 2010 ("Barrier request in delayed write %p", bp)); 2011 BUF_ASSERT_HELD(bp); 2012 2013 if (bp->b_flags & B_INVAL) { 2014 brelse(bp); 2015 return; 2016 } 2017 2018 /* 2019 * If we have too many dirty buffers, don't create any more. 2020 * If we are wildly over our limit, then force a complete 2021 * cleanup. Otherwise, just keep the situation from getting 2022 * out of control. Note that we have to avoid a recursive 2023 * disaster and not try to clean up after our own cleanup! 2024 */ 2025 vp = bp->b_vp; 2026 bo = bp->b_bufobj; 2027 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2028 td->td_pflags |= TDP_INBDFLUSH; 2029 BO_BDFLUSH(bo, bp); 2030 td->td_pflags &= ~TDP_INBDFLUSH; 2031 } else 2032 recursiveflushes++; 2033 2034 bdirty(bp); 2035 /* 2036 * Set B_CACHE, indicating that the buffer is fully valid. This is 2037 * true even of NFS now. 2038 */ 2039 bp->b_flags |= B_CACHE; 2040 2041 /* 2042 * This bmap keeps the system from needing to do the bmap later, 2043 * perhaps when the system is attempting to do a sync. Since it 2044 * is likely that the indirect block -- or whatever other datastructure 2045 * that the filesystem needs is still in memory now, it is a good 2046 * thing to do this. Note also, that if the pageout daemon is 2047 * requesting a sync -- there might not be enough memory to do 2048 * the bmap then... So, this is important to do. 2049 */ 2050 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2051 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2052 } 2053 2054 /* 2055 * Set the *dirty* buffer range based upon the VM system dirty 2056 * pages. 2057 * 2058 * Mark the buffer pages as clean. We need to do this here to 2059 * satisfy the vnode_pager and the pageout daemon, so that it 2060 * thinks that the pages have been "cleaned". Note that since 2061 * the pages are in a delayed write buffer -- the VFS layer 2062 * "will" see that the pages get written out on the next sync, 2063 * or perhaps the cluster will be completed. 2064 */ 2065 vfs_clean_pages_dirty_buf(bp); 2066 bqrelse(bp); 2067 2068 /* 2069 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2070 * due to the softdep code. 2071 */ 2072 } 2073 2074 /* 2075 * bdirty: 2076 * 2077 * Turn buffer into delayed write request. We must clear BIO_READ and 2078 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2079 * itself to properly update it in the dirty/clean lists. We mark it 2080 * B_DONE to ensure that any asynchronization of the buffer properly 2081 * clears B_DONE ( else a panic will occur later ). 2082 * 2083 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2084 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2085 * should only be called if the buffer is known-good. 2086 * 2087 * Since the buffer is not on a queue, we do not update the numfreebuffers 2088 * count. 2089 * 2090 * The buffer must be on QUEUE_NONE. 2091 */ 2092 void 2093 bdirty(struct buf *bp) 2094 { 2095 2096 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2097 bp, bp->b_vp, bp->b_flags); 2098 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2099 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2100 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2101 BUF_ASSERT_HELD(bp); 2102 bp->b_flags &= ~(B_RELBUF); 2103 bp->b_iocmd = BIO_WRITE; 2104 2105 if ((bp->b_flags & B_DELWRI) == 0) { 2106 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2107 reassignbuf(bp); 2108 bdirtyadd(); 2109 } 2110 } 2111 2112 /* 2113 * bundirty: 2114 * 2115 * Clear B_DELWRI for buffer. 2116 * 2117 * Since the buffer is not on a queue, we do not update the numfreebuffers 2118 * count. 2119 * 2120 * The buffer must be on QUEUE_NONE. 2121 */ 2122 2123 void 2124 bundirty(struct buf *bp) 2125 { 2126 2127 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2128 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2129 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2130 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2131 BUF_ASSERT_HELD(bp); 2132 2133 if (bp->b_flags & B_DELWRI) { 2134 bp->b_flags &= ~B_DELWRI; 2135 reassignbuf(bp); 2136 bdirtysub(); 2137 } 2138 /* 2139 * Since it is now being written, we can clear its deferred write flag. 2140 */ 2141 bp->b_flags &= ~B_DEFERRED; 2142 } 2143 2144 /* 2145 * bawrite: 2146 * 2147 * Asynchronous write. Start output on a buffer, but do not wait for 2148 * it to complete. The buffer is released when the output completes. 2149 * 2150 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2151 * B_INVAL buffers. Not us. 2152 */ 2153 void 2154 bawrite(struct buf *bp) 2155 { 2156 2157 bp->b_flags |= B_ASYNC; 2158 (void) bwrite(bp); 2159 } 2160 2161 /* 2162 * babarrierwrite: 2163 * 2164 * Asynchronous barrier write. Start output on a buffer, but do not 2165 * wait for it to complete. Place a write barrier after this write so 2166 * that this buffer and all buffers written before it are committed to 2167 * the disk before any buffers written after this write are committed 2168 * to the disk. The buffer is released when the output completes. 2169 */ 2170 void 2171 babarrierwrite(struct buf *bp) 2172 { 2173 2174 bp->b_flags |= B_ASYNC | B_BARRIER; 2175 (void) bwrite(bp); 2176 } 2177 2178 /* 2179 * bbarrierwrite: 2180 * 2181 * Synchronous barrier write. Start output on a buffer and wait for 2182 * it to complete. Place a write barrier after this write so that 2183 * this buffer and all buffers written before it are committed to 2184 * the disk before any buffers written after this write are committed 2185 * to the disk. The buffer is released when the output completes. 2186 */ 2187 int 2188 bbarrierwrite(struct buf *bp) 2189 { 2190 2191 bp->b_flags |= B_BARRIER; 2192 return (bwrite(bp)); 2193 } 2194 2195 /* 2196 * bwillwrite: 2197 * 2198 * Called prior to the locking of any vnodes when we are expecting to 2199 * write. We do not want to starve the buffer cache with too many 2200 * dirty buffers so we block here. By blocking prior to the locking 2201 * of any vnodes we attempt to avoid the situation where a locked vnode 2202 * prevents the various system daemons from flushing related buffers. 2203 */ 2204 void 2205 bwillwrite(void) 2206 { 2207 2208 if (numdirtybuffers >= hidirtybuffers) { 2209 mtx_lock(&bdirtylock); 2210 while (numdirtybuffers >= hidirtybuffers) { 2211 bdirtywait = 1; 2212 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2213 "flswai", 0); 2214 } 2215 mtx_unlock(&bdirtylock); 2216 } 2217 } 2218 2219 /* 2220 * Return true if we have too many dirty buffers. 2221 */ 2222 int 2223 buf_dirty_count_severe(void) 2224 { 2225 2226 return(numdirtybuffers >= hidirtybuffers); 2227 } 2228 2229 /* 2230 * brelse: 2231 * 2232 * Release a busy buffer and, if requested, free its resources. The 2233 * buffer will be stashed in the appropriate bufqueue[] allowing it 2234 * to be accessed later as a cache entity or reused for other purposes. 2235 */ 2236 void 2237 brelse(struct buf *bp) 2238 { 2239 int qindex; 2240 2241 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2242 bp, bp->b_vp, bp->b_flags); 2243 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2244 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2245 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2246 ("brelse: non-VMIO buffer marked NOREUSE")); 2247 2248 if (BUF_LOCKRECURSED(bp)) { 2249 /* 2250 * Do not process, in particular, do not handle the 2251 * B_INVAL/B_RELBUF and do not release to free list. 2252 */ 2253 BUF_UNLOCK(bp); 2254 return; 2255 } 2256 2257 if (bp->b_flags & B_MANAGED) { 2258 bqrelse(bp); 2259 return; 2260 } 2261 2262 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2263 BO_LOCK(bp->b_bufobj); 2264 bp->b_vflags &= ~BV_BKGRDERR; 2265 BO_UNLOCK(bp->b_bufobj); 2266 bdirty(bp); 2267 } 2268 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2269 !(bp->b_flags & B_INVAL)) { 2270 /* 2271 * Failed write, redirty. Must clear BIO_ERROR to prevent 2272 * pages from being scrapped. 2273 */ 2274 bp->b_ioflags &= ~BIO_ERROR; 2275 bdirty(bp); 2276 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2277 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2278 /* 2279 * Either a failed read I/O or we were asked to free or not 2280 * cache the buffer. 2281 */ 2282 bp->b_flags |= B_INVAL; 2283 if (!LIST_EMPTY(&bp->b_dep)) 2284 buf_deallocate(bp); 2285 if (bp->b_flags & B_DELWRI) 2286 bdirtysub(); 2287 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2288 if ((bp->b_flags & B_VMIO) == 0) { 2289 allocbuf(bp, 0); 2290 if (bp->b_vp) 2291 brelvp(bp); 2292 } 2293 } 2294 2295 /* 2296 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2297 * is called with B_DELWRI set, the underlying pages may wind up 2298 * getting freed causing a previous write (bdwrite()) to get 'lost' 2299 * because pages associated with a B_DELWRI bp are marked clean. 2300 * 2301 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2302 * if B_DELWRI is set. 2303 */ 2304 if (bp->b_flags & B_DELWRI) 2305 bp->b_flags &= ~B_RELBUF; 2306 2307 /* 2308 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2309 * constituted, not even NFS buffers now. Two flags effect this. If 2310 * B_INVAL, the struct buf is invalidated but the VM object is kept 2311 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2312 * 2313 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2314 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2315 * buffer is also B_INVAL because it hits the re-dirtying code above. 2316 * 2317 * Normally we can do this whether a buffer is B_DELWRI or not. If 2318 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2319 * the commit state and we cannot afford to lose the buffer. If the 2320 * buffer has a background write in progress, we need to keep it 2321 * around to prevent it from being reconstituted and starting a second 2322 * background write. 2323 */ 2324 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2325 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2326 !(bp->b_vp->v_mount != NULL && 2327 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2328 !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI))) { 2329 vfs_vmio_invalidate(bp); 2330 allocbuf(bp, 0); 2331 } 2332 2333 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2334 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2335 allocbuf(bp, 0); 2336 bp->b_flags &= ~B_NOREUSE; 2337 if (bp->b_vp != NULL) 2338 brelvp(bp); 2339 } 2340 2341 /* 2342 * If the buffer has junk contents signal it and eventually 2343 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2344 * doesn't find it. 2345 */ 2346 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2347 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2348 bp->b_flags |= B_INVAL; 2349 if (bp->b_flags & B_INVAL) { 2350 if (bp->b_flags & B_DELWRI) 2351 bundirty(bp); 2352 if (bp->b_vp) 2353 brelvp(bp); 2354 } 2355 2356 /* buffers with no memory */ 2357 if (bp->b_bufsize == 0) { 2358 buf_free(bp); 2359 return; 2360 } 2361 /* buffers with junk contents */ 2362 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2363 (bp->b_ioflags & BIO_ERROR)) { 2364 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2365 if (bp->b_vflags & BV_BKGRDINPROG) 2366 panic("losing buffer 2"); 2367 qindex = QUEUE_CLEAN; 2368 bp->b_flags |= B_AGE; 2369 /* remaining buffers */ 2370 } else if (bp->b_flags & B_DELWRI) 2371 qindex = QUEUE_DIRTY; 2372 else 2373 qindex = QUEUE_CLEAN; 2374 2375 binsfree(bp, qindex); 2376 2377 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 2378 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2379 panic("brelse: not dirty"); 2380 /* unlock */ 2381 BUF_UNLOCK(bp); 2382 if (qindex == QUEUE_CLEAN) 2383 bufspace_wakeup(); 2384 } 2385 2386 /* 2387 * Release a buffer back to the appropriate queue but do not try to free 2388 * it. The buffer is expected to be used again soon. 2389 * 2390 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2391 * biodone() to requeue an async I/O on completion. It is also used when 2392 * known good buffers need to be requeued but we think we may need the data 2393 * again soon. 2394 * 2395 * XXX we should be able to leave the B_RELBUF hint set on completion. 2396 */ 2397 void 2398 bqrelse(struct buf *bp) 2399 { 2400 int qindex; 2401 2402 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2403 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2404 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2405 2406 qindex = QUEUE_NONE; 2407 if (BUF_LOCKRECURSED(bp)) { 2408 /* do not release to free list */ 2409 BUF_UNLOCK(bp); 2410 return; 2411 } 2412 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2413 2414 if (bp->b_flags & B_MANAGED) { 2415 if (bp->b_flags & B_REMFREE) 2416 bremfreef(bp); 2417 goto out; 2418 } 2419 2420 /* buffers with stale but valid contents */ 2421 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2422 BV_BKGRDERR)) == BV_BKGRDERR) { 2423 BO_LOCK(bp->b_bufobj); 2424 bp->b_vflags &= ~BV_BKGRDERR; 2425 BO_UNLOCK(bp->b_bufobj); 2426 qindex = QUEUE_DIRTY; 2427 } else { 2428 if ((bp->b_flags & B_DELWRI) == 0 && 2429 (bp->b_xflags & BX_VNDIRTY)) 2430 panic("bqrelse: not dirty"); 2431 if ((bp->b_flags & B_NOREUSE) != 0) { 2432 brelse(bp); 2433 return; 2434 } 2435 qindex = QUEUE_CLEAN; 2436 } 2437 binsfree(bp, qindex); 2438 2439 out: 2440 /* unlock */ 2441 BUF_UNLOCK(bp); 2442 if (qindex == QUEUE_CLEAN) 2443 bufspace_wakeup(); 2444 } 2445 2446 /* 2447 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2448 * restore bogus pages. 2449 */ 2450 static void 2451 vfs_vmio_iodone(struct buf *bp) 2452 { 2453 vm_ooffset_t foff; 2454 vm_page_t m; 2455 vm_object_t obj; 2456 struct vnode *vp; 2457 int bogus, i, iosize; 2458 2459 obj = bp->b_bufobj->bo_object; 2460 KASSERT(obj->paging_in_progress >= bp->b_npages, 2461 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2462 obj->paging_in_progress, bp->b_npages)); 2463 2464 vp = bp->b_vp; 2465 KASSERT(vp->v_holdcnt > 0, 2466 ("vfs_vmio_iodone: vnode %p has zero hold count", vp)); 2467 KASSERT(vp->v_object != NULL, 2468 ("vfs_vmio_iodone: vnode %p has no vm_object", vp)); 2469 2470 foff = bp->b_offset; 2471 KASSERT(bp->b_offset != NOOFFSET, 2472 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2473 2474 bogus = 0; 2475 iosize = bp->b_bcount - bp->b_resid; 2476 VM_OBJECT_WLOCK(obj); 2477 for (i = 0; i < bp->b_npages; i++) { 2478 int resid; 2479 2480 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2481 if (resid > iosize) 2482 resid = iosize; 2483 2484 /* 2485 * cleanup bogus pages, restoring the originals 2486 */ 2487 m = bp->b_pages[i]; 2488 if (m == bogus_page) { 2489 bogus = 1; 2490 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2491 if (m == NULL) 2492 panic("biodone: page disappeared!"); 2493 bp->b_pages[i] = m; 2494 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2495 /* 2496 * In the write case, the valid and clean bits are 2497 * already changed correctly ( see bdwrite() ), so we 2498 * only need to do this here in the read case. 2499 */ 2500 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2501 resid)) == 0, ("vfs_vmio_iodone: page %p " 2502 "has unexpected dirty bits", m)); 2503 vfs_page_set_valid(bp, foff, m); 2504 } 2505 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2506 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2507 (intmax_t)foff, (uintmax_t)m->pindex)); 2508 2509 vm_page_sunbusy(m); 2510 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2511 iosize -= resid; 2512 } 2513 vm_object_pip_wakeupn(obj, bp->b_npages); 2514 VM_OBJECT_WUNLOCK(obj); 2515 if (bogus && buf_mapped(bp)) { 2516 BUF_CHECK_MAPPED(bp); 2517 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2518 bp->b_pages, bp->b_npages); 2519 } 2520 } 2521 2522 /* 2523 * Unwire a page held by a buf and place it on the appropriate vm queue. 2524 */ 2525 static void 2526 vfs_vmio_unwire(struct buf *bp, vm_page_t m) 2527 { 2528 bool freed; 2529 2530 vm_page_lock(m); 2531 if (vm_page_unwire(m, PQ_NONE)) { 2532 /* 2533 * Determine if the page should be freed before adding 2534 * it to the inactive queue. 2535 */ 2536 if (m->valid == 0) { 2537 freed = !vm_page_busied(m); 2538 if (freed) 2539 vm_page_free(m); 2540 } else if ((bp->b_flags & B_DIRECT) != 0) 2541 freed = vm_page_try_to_free(m); 2542 else 2543 freed = false; 2544 if (!freed) { 2545 /* 2546 * If the page is unlikely to be reused, let the 2547 * VM know. Otherwise, maintain LRU page 2548 * ordering and put the page at the tail of the 2549 * inactive queue. 2550 */ 2551 if ((bp->b_flags & B_NOREUSE) != 0) 2552 vm_page_deactivate_noreuse(m); 2553 else 2554 vm_page_deactivate(m); 2555 } 2556 } 2557 vm_page_unlock(m); 2558 } 2559 2560 /* 2561 * Perform page invalidation when a buffer is released. The fully invalid 2562 * pages will be reclaimed later in vfs_vmio_truncate(). 2563 */ 2564 static void 2565 vfs_vmio_invalidate(struct buf *bp) 2566 { 2567 vm_object_t obj; 2568 vm_page_t m; 2569 int i, resid, poffset, presid; 2570 2571 if (buf_mapped(bp)) { 2572 BUF_CHECK_MAPPED(bp); 2573 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 2574 } else 2575 BUF_CHECK_UNMAPPED(bp); 2576 /* 2577 * Get the base offset and length of the buffer. Note that 2578 * in the VMIO case if the buffer block size is not 2579 * page-aligned then b_data pointer may not be page-aligned. 2580 * But our b_pages[] array *IS* page aligned. 2581 * 2582 * block sizes less then DEV_BSIZE (usually 512) are not 2583 * supported due to the page granularity bits (m->valid, 2584 * m->dirty, etc...). 2585 * 2586 * See man buf(9) for more information 2587 */ 2588 obj = bp->b_bufobj->bo_object; 2589 resid = bp->b_bufsize; 2590 poffset = bp->b_offset & PAGE_MASK; 2591 VM_OBJECT_WLOCK(obj); 2592 for (i = 0; i < bp->b_npages; i++) { 2593 m = bp->b_pages[i]; 2594 if (m == bogus_page) 2595 panic("vfs_vmio_invalidate: Unexpected bogus page."); 2596 bp->b_pages[i] = NULL; 2597 2598 presid = resid > (PAGE_SIZE - poffset) ? 2599 (PAGE_SIZE - poffset) : resid; 2600 KASSERT(presid >= 0, ("brelse: extra page")); 2601 while (vm_page_xbusied(m)) { 2602 vm_page_lock(m); 2603 VM_OBJECT_WUNLOCK(obj); 2604 vm_page_busy_sleep(m, "mbncsh"); 2605 VM_OBJECT_WLOCK(obj); 2606 } 2607 if (pmap_page_wired_mappings(m) == 0) 2608 vm_page_set_invalid(m, poffset, presid); 2609 vfs_vmio_unwire(bp, m); 2610 resid -= presid; 2611 poffset = 0; 2612 } 2613 VM_OBJECT_WUNLOCK(obj); 2614 bp->b_npages = 0; 2615 } 2616 2617 /* 2618 * Page-granular truncation of an existing VMIO buffer. 2619 */ 2620 static void 2621 vfs_vmio_truncate(struct buf *bp, int desiredpages) 2622 { 2623 vm_object_t obj; 2624 vm_page_t m; 2625 int i; 2626 2627 if (bp->b_npages == desiredpages) 2628 return; 2629 2630 if (buf_mapped(bp)) { 2631 BUF_CHECK_MAPPED(bp); 2632 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 2633 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 2634 } else 2635 BUF_CHECK_UNMAPPED(bp); 2636 obj = bp->b_bufobj->bo_object; 2637 if (obj != NULL) 2638 VM_OBJECT_WLOCK(obj); 2639 for (i = desiredpages; i < bp->b_npages; i++) { 2640 m = bp->b_pages[i]; 2641 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 2642 bp->b_pages[i] = NULL; 2643 vfs_vmio_unwire(bp, m); 2644 } 2645 if (obj != NULL) 2646 VM_OBJECT_WUNLOCK(obj); 2647 bp->b_npages = desiredpages; 2648 } 2649 2650 /* 2651 * Byte granular extension of VMIO buffers. 2652 */ 2653 static void 2654 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 2655 { 2656 /* 2657 * We are growing the buffer, possibly in a 2658 * byte-granular fashion. 2659 */ 2660 vm_object_t obj; 2661 vm_offset_t toff; 2662 vm_offset_t tinc; 2663 vm_page_t m; 2664 2665 /* 2666 * Step 1, bring in the VM pages from the object, allocating 2667 * them if necessary. We must clear B_CACHE if these pages 2668 * are not valid for the range covered by the buffer. 2669 */ 2670 obj = bp->b_bufobj->bo_object; 2671 VM_OBJECT_WLOCK(obj); 2672 while (bp->b_npages < desiredpages) { 2673 /* 2674 * We must allocate system pages since blocking 2675 * here could interfere with paging I/O, no 2676 * matter which process we are. 2677 * 2678 * Only exclusive busy can be tested here. 2679 * Blocking on shared busy might lead to 2680 * deadlocks once allocbuf() is called after 2681 * pages are vfs_busy_pages(). 2682 */ 2683 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages, 2684 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2685 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY | 2686 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 2687 if (m->valid == 0) 2688 bp->b_flags &= ~B_CACHE; 2689 bp->b_pages[bp->b_npages] = m; 2690 ++bp->b_npages; 2691 } 2692 2693 /* 2694 * Step 2. We've loaded the pages into the buffer, 2695 * we have to figure out if we can still have B_CACHE 2696 * set. Note that B_CACHE is set according to the 2697 * byte-granular range ( bcount and size ), not the 2698 * aligned range ( newbsize ). 2699 * 2700 * The VM test is against m->valid, which is DEV_BSIZE 2701 * aligned. Needless to say, the validity of the data 2702 * needs to also be DEV_BSIZE aligned. Note that this 2703 * fails with NFS if the server or some other client 2704 * extends the file's EOF. If our buffer is resized, 2705 * B_CACHE may remain set! XXX 2706 */ 2707 toff = bp->b_bcount; 2708 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2709 while ((bp->b_flags & B_CACHE) && toff < size) { 2710 vm_pindex_t pi; 2711 2712 if (tinc > (size - toff)) 2713 tinc = size - toff; 2714 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 2715 m = bp->b_pages[pi]; 2716 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 2717 toff += tinc; 2718 tinc = PAGE_SIZE; 2719 } 2720 VM_OBJECT_WUNLOCK(obj); 2721 2722 /* 2723 * Step 3, fixup the KVA pmap. 2724 */ 2725 if (buf_mapped(bp)) 2726 bpmap_qenter(bp); 2727 else 2728 BUF_CHECK_UNMAPPED(bp); 2729 } 2730 2731 /* 2732 * Check to see if a block at a particular lbn is available for a clustered 2733 * write. 2734 */ 2735 static int 2736 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 2737 { 2738 struct buf *bpa; 2739 int match; 2740 2741 match = 0; 2742 2743 /* If the buf isn't in core skip it */ 2744 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 2745 return (0); 2746 2747 /* If the buf is busy we don't want to wait for it */ 2748 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2749 return (0); 2750 2751 /* Only cluster with valid clusterable delayed write buffers */ 2752 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 2753 (B_DELWRI | B_CLUSTEROK)) 2754 goto done; 2755 2756 if (bpa->b_bufsize != size) 2757 goto done; 2758 2759 /* 2760 * Check to see if it is in the expected place on disk and that the 2761 * block has been mapped. 2762 */ 2763 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 2764 match = 1; 2765 done: 2766 BUF_UNLOCK(bpa); 2767 return (match); 2768 } 2769 2770 /* 2771 * vfs_bio_awrite: 2772 * 2773 * Implement clustered async writes for clearing out B_DELWRI buffers. 2774 * This is much better then the old way of writing only one buffer at 2775 * a time. Note that we may not be presented with the buffers in the 2776 * correct order, so we search for the cluster in both directions. 2777 */ 2778 int 2779 vfs_bio_awrite(struct buf *bp) 2780 { 2781 struct bufobj *bo; 2782 int i; 2783 int j; 2784 daddr_t lblkno = bp->b_lblkno; 2785 struct vnode *vp = bp->b_vp; 2786 int ncl; 2787 int nwritten; 2788 int size; 2789 int maxcl; 2790 int gbflags; 2791 2792 bo = &vp->v_bufobj; 2793 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 2794 /* 2795 * right now we support clustered writing only to regular files. If 2796 * we find a clusterable block we could be in the middle of a cluster 2797 * rather then at the beginning. 2798 */ 2799 if ((vp->v_type == VREG) && 2800 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 2801 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 2802 2803 size = vp->v_mount->mnt_stat.f_iosize; 2804 maxcl = MAXPHYS / size; 2805 2806 BO_RLOCK(bo); 2807 for (i = 1; i < maxcl; i++) 2808 if (vfs_bio_clcheck(vp, size, lblkno + i, 2809 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 2810 break; 2811 2812 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 2813 if (vfs_bio_clcheck(vp, size, lblkno - j, 2814 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 2815 break; 2816 BO_RUNLOCK(bo); 2817 --j; 2818 ncl = i + j; 2819 /* 2820 * this is a possible cluster write 2821 */ 2822 if (ncl != 1) { 2823 BUF_UNLOCK(bp); 2824 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 2825 gbflags); 2826 return (nwritten); 2827 } 2828 } 2829 bremfree(bp); 2830 bp->b_flags |= B_ASYNC; 2831 /* 2832 * default (old) behavior, writing out only one block 2833 * 2834 * XXX returns b_bufsize instead of b_bcount for nwritten? 2835 */ 2836 nwritten = bp->b_bufsize; 2837 (void) bwrite(bp); 2838 2839 return (nwritten); 2840 } 2841 2842 /* 2843 * getnewbuf_kva: 2844 * 2845 * Allocate KVA for an empty buf header according to gbflags. 2846 */ 2847 static int 2848 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 2849 { 2850 2851 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 2852 /* 2853 * In order to keep fragmentation sane we only allocate kva 2854 * in BKVASIZE chunks. XXX with vmem we can do page size. 2855 */ 2856 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2857 2858 if (maxsize != bp->b_kvasize && 2859 bufkva_alloc(bp, maxsize, gbflags)) 2860 return (ENOSPC); 2861 } 2862 return (0); 2863 } 2864 2865 /* 2866 * getnewbuf: 2867 * 2868 * Find and initialize a new buffer header, freeing up existing buffers 2869 * in the bufqueues as necessary. The new buffer is returned locked. 2870 * 2871 * We block if: 2872 * We have insufficient buffer headers 2873 * We have insufficient buffer space 2874 * buffer_arena is too fragmented ( space reservation fails ) 2875 * If we have to flush dirty buffers ( but we try to avoid this ) 2876 * 2877 * The caller is responsible for releasing the reserved bufspace after 2878 * allocbuf() is called. 2879 */ 2880 static struct buf * 2881 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 2882 { 2883 struct buf *bp; 2884 bool metadata, reserved; 2885 2886 bp = NULL; 2887 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2888 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2889 if (!unmapped_buf_allowed) 2890 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2891 2892 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2893 vp->v_type == VCHR) 2894 metadata = true; 2895 else 2896 metadata = false; 2897 atomic_add_int(&getnewbufcalls, 1); 2898 reserved = false; 2899 do { 2900 if (reserved == false && 2901 bufspace_reserve(maxsize, metadata) != 0) 2902 continue; 2903 reserved = true; 2904 if ((bp = buf_alloc()) == NULL) 2905 continue; 2906 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 2907 return (bp); 2908 break; 2909 } while(buf_scan(false) == 0); 2910 2911 if (reserved) 2912 atomic_subtract_long(&bufspace, maxsize); 2913 if (bp != NULL) { 2914 bp->b_flags |= B_INVAL; 2915 brelse(bp); 2916 } 2917 bufspace_wait(vp, gbflags, slpflag, slptimeo); 2918 2919 return (NULL); 2920 } 2921 2922 /* 2923 * buf_daemon: 2924 * 2925 * buffer flushing daemon. Buffers are normally flushed by the 2926 * update daemon but if it cannot keep up this process starts to 2927 * take the load in an attempt to prevent getnewbuf() from blocking. 2928 */ 2929 static struct kproc_desc buf_kp = { 2930 "bufdaemon", 2931 buf_daemon, 2932 &bufdaemonproc 2933 }; 2934 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2935 2936 static int 2937 buf_flush(struct vnode *vp, int target) 2938 { 2939 int flushed; 2940 2941 flushed = flushbufqueues(vp, target, 0); 2942 if (flushed == 0) { 2943 /* 2944 * Could not find any buffers without rollback 2945 * dependencies, so just write the first one 2946 * in the hopes of eventually making progress. 2947 */ 2948 if (vp != NULL && target > 2) 2949 target /= 2; 2950 flushbufqueues(vp, target, 1); 2951 } 2952 return (flushed); 2953 } 2954 2955 static void 2956 buf_daemon() 2957 { 2958 int lodirty; 2959 2960 /* 2961 * This process needs to be suspended prior to shutdown sync. 2962 */ 2963 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2964 SHUTDOWN_PRI_LAST); 2965 2966 /* 2967 * This process is allowed to take the buffer cache to the limit 2968 */ 2969 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2970 mtx_lock(&bdlock); 2971 for (;;) { 2972 bd_request = 0; 2973 mtx_unlock(&bdlock); 2974 2975 kproc_suspend_check(bufdaemonproc); 2976 lodirty = lodirtybuffers; 2977 if (bd_speedupreq) { 2978 lodirty = numdirtybuffers / 2; 2979 bd_speedupreq = 0; 2980 } 2981 /* 2982 * Do the flush. Limit the amount of in-transit I/O we 2983 * allow to build up, otherwise we would completely saturate 2984 * the I/O system. 2985 */ 2986 while (numdirtybuffers > lodirty) { 2987 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0) 2988 break; 2989 kern_yield(PRI_USER); 2990 } 2991 2992 /* 2993 * Only clear bd_request if we have reached our low water 2994 * mark. The buf_daemon normally waits 1 second and 2995 * then incrementally flushes any dirty buffers that have 2996 * built up, within reason. 2997 * 2998 * If we were unable to hit our low water mark and couldn't 2999 * find any flushable buffers, we sleep for a short period 3000 * to avoid endless loops on unlockable buffers. 3001 */ 3002 mtx_lock(&bdlock); 3003 if (numdirtybuffers <= lodirtybuffers) { 3004 /* 3005 * We reached our low water mark, reset the 3006 * request and sleep until we are needed again. 3007 * The sleep is just so the suspend code works. 3008 */ 3009 bd_request = 0; 3010 /* 3011 * Do an extra wakeup in case dirty threshold 3012 * changed via sysctl and the explicit transition 3013 * out of shortfall was missed. 3014 */ 3015 bdirtywakeup(); 3016 if (runningbufspace <= lorunningspace) 3017 runningwakeup(); 3018 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3019 } else { 3020 /* 3021 * We couldn't find any flushable dirty buffers but 3022 * still have too many dirty buffers, we 3023 * have to sleep and try again. (rare) 3024 */ 3025 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3026 } 3027 } 3028 } 3029 3030 /* 3031 * flushbufqueues: 3032 * 3033 * Try to flush a buffer in the dirty queue. We must be careful to 3034 * free up B_INVAL buffers instead of write them, which NFS is 3035 * particularly sensitive to. 3036 */ 3037 static int flushwithdeps = 0; 3038 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 3039 0, "Number of buffers flushed with dependecies that require rollbacks"); 3040 3041 static int 3042 flushbufqueues(struct vnode *lvp, int target, int flushdeps) 3043 { 3044 struct buf *sentinel; 3045 struct vnode *vp; 3046 struct mount *mp; 3047 struct buf *bp; 3048 int hasdeps; 3049 int flushed; 3050 int queue; 3051 int error; 3052 bool unlock; 3053 3054 flushed = 0; 3055 queue = QUEUE_DIRTY; 3056 bp = NULL; 3057 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3058 sentinel->b_qindex = QUEUE_SENTINEL; 3059 mtx_lock(&bqlocks[queue]); 3060 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 3061 mtx_unlock(&bqlocks[queue]); 3062 while (flushed != target) { 3063 maybe_yield(); 3064 mtx_lock(&bqlocks[queue]); 3065 bp = TAILQ_NEXT(sentinel, b_freelist); 3066 if (bp != NULL) { 3067 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 3068 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 3069 b_freelist); 3070 } else { 3071 mtx_unlock(&bqlocks[queue]); 3072 break; 3073 } 3074 /* 3075 * Skip sentinels inserted by other invocations of the 3076 * flushbufqueues(), taking care to not reorder them. 3077 * 3078 * Only flush the buffers that belong to the 3079 * vnode locked by the curthread. 3080 */ 3081 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3082 bp->b_vp != lvp)) { 3083 mtx_unlock(&bqlocks[queue]); 3084 continue; 3085 } 3086 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3087 mtx_unlock(&bqlocks[queue]); 3088 if (error != 0) 3089 continue; 3090 if (bp->b_pin_count > 0) { 3091 BUF_UNLOCK(bp); 3092 continue; 3093 } 3094 /* 3095 * BKGRDINPROG can only be set with the buf and bufobj 3096 * locks both held. We tolerate a race to clear it here. 3097 */ 3098 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3099 (bp->b_flags & B_DELWRI) == 0) { 3100 BUF_UNLOCK(bp); 3101 continue; 3102 } 3103 if (bp->b_flags & B_INVAL) { 3104 bremfreef(bp); 3105 brelse(bp); 3106 flushed++; 3107 continue; 3108 } 3109 3110 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3111 if (flushdeps == 0) { 3112 BUF_UNLOCK(bp); 3113 continue; 3114 } 3115 hasdeps = 1; 3116 } else 3117 hasdeps = 0; 3118 /* 3119 * We must hold the lock on a vnode before writing 3120 * one of its buffers. Otherwise we may confuse, or 3121 * in the case of a snapshot vnode, deadlock the 3122 * system. 3123 * 3124 * The lock order here is the reverse of the normal 3125 * of vnode followed by buf lock. This is ok because 3126 * the NOWAIT will prevent deadlock. 3127 */ 3128 vp = bp->b_vp; 3129 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3130 BUF_UNLOCK(bp); 3131 continue; 3132 } 3133 if (lvp == NULL) { 3134 unlock = true; 3135 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3136 } else { 3137 ASSERT_VOP_LOCKED(vp, "getbuf"); 3138 unlock = false; 3139 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3140 vn_lock(vp, LK_TRYUPGRADE); 3141 } 3142 if (error == 0) { 3143 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3144 bp, bp->b_vp, bp->b_flags); 3145 if (curproc == bufdaemonproc) { 3146 vfs_bio_awrite(bp); 3147 } else { 3148 bremfree(bp); 3149 bwrite(bp); 3150 notbufdflushes++; 3151 } 3152 vn_finished_write(mp); 3153 if (unlock) 3154 VOP_UNLOCK(vp, 0); 3155 flushwithdeps += hasdeps; 3156 flushed++; 3157 3158 /* 3159 * Sleeping on runningbufspace while holding 3160 * vnode lock leads to deadlock. 3161 */ 3162 if (curproc == bufdaemonproc && 3163 runningbufspace > hirunningspace) 3164 waitrunningbufspace(); 3165 continue; 3166 } 3167 vn_finished_write(mp); 3168 BUF_UNLOCK(bp); 3169 } 3170 mtx_lock(&bqlocks[queue]); 3171 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 3172 mtx_unlock(&bqlocks[queue]); 3173 free(sentinel, M_TEMP); 3174 return (flushed); 3175 } 3176 3177 /* 3178 * Check to see if a block is currently memory resident. 3179 */ 3180 struct buf * 3181 incore(struct bufobj *bo, daddr_t blkno) 3182 { 3183 struct buf *bp; 3184 3185 BO_RLOCK(bo); 3186 bp = gbincore(bo, blkno); 3187 BO_RUNLOCK(bo); 3188 return (bp); 3189 } 3190 3191 /* 3192 * Returns true if no I/O is needed to access the 3193 * associated VM object. This is like incore except 3194 * it also hunts around in the VM system for the data. 3195 */ 3196 3197 static int 3198 inmem(struct vnode * vp, daddr_t blkno) 3199 { 3200 vm_object_t obj; 3201 vm_offset_t toff, tinc, size; 3202 vm_page_t m; 3203 vm_ooffset_t off; 3204 3205 ASSERT_VOP_LOCKED(vp, "inmem"); 3206 3207 if (incore(&vp->v_bufobj, blkno)) 3208 return 1; 3209 if (vp->v_mount == NULL) 3210 return 0; 3211 obj = vp->v_object; 3212 if (obj == NULL) 3213 return (0); 3214 3215 size = PAGE_SIZE; 3216 if (size > vp->v_mount->mnt_stat.f_iosize) 3217 size = vp->v_mount->mnt_stat.f_iosize; 3218 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3219 3220 VM_OBJECT_RLOCK(obj); 3221 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3222 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 3223 if (!m) 3224 goto notinmem; 3225 tinc = size; 3226 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3227 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3228 if (vm_page_is_valid(m, 3229 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 3230 goto notinmem; 3231 } 3232 VM_OBJECT_RUNLOCK(obj); 3233 return 1; 3234 3235 notinmem: 3236 VM_OBJECT_RUNLOCK(obj); 3237 return (0); 3238 } 3239 3240 /* 3241 * Set the dirty range for a buffer based on the status of the dirty 3242 * bits in the pages comprising the buffer. The range is limited 3243 * to the size of the buffer. 3244 * 3245 * Tell the VM system that the pages associated with this buffer 3246 * are clean. This is used for delayed writes where the data is 3247 * going to go to disk eventually without additional VM intevention. 3248 * 3249 * Note that while we only really need to clean through to b_bcount, we 3250 * just go ahead and clean through to b_bufsize. 3251 */ 3252 static void 3253 vfs_clean_pages_dirty_buf(struct buf *bp) 3254 { 3255 vm_ooffset_t foff, noff, eoff; 3256 vm_page_t m; 3257 int i; 3258 3259 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3260 return; 3261 3262 foff = bp->b_offset; 3263 KASSERT(bp->b_offset != NOOFFSET, 3264 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3265 3266 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3267 vfs_drain_busy_pages(bp); 3268 vfs_setdirty_locked_object(bp); 3269 for (i = 0; i < bp->b_npages; i++) { 3270 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3271 eoff = noff; 3272 if (eoff > bp->b_offset + bp->b_bufsize) 3273 eoff = bp->b_offset + bp->b_bufsize; 3274 m = bp->b_pages[i]; 3275 vfs_page_set_validclean(bp, foff, m); 3276 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3277 foff = noff; 3278 } 3279 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3280 } 3281 3282 static void 3283 vfs_setdirty_locked_object(struct buf *bp) 3284 { 3285 vm_object_t object; 3286 int i; 3287 3288 object = bp->b_bufobj->bo_object; 3289 VM_OBJECT_ASSERT_WLOCKED(object); 3290 3291 /* 3292 * We qualify the scan for modified pages on whether the 3293 * object has been flushed yet. 3294 */ 3295 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 3296 vm_offset_t boffset; 3297 vm_offset_t eoffset; 3298 3299 /* 3300 * test the pages to see if they have been modified directly 3301 * by users through the VM system. 3302 */ 3303 for (i = 0; i < bp->b_npages; i++) 3304 vm_page_test_dirty(bp->b_pages[i]); 3305 3306 /* 3307 * Calculate the encompassing dirty range, boffset and eoffset, 3308 * (eoffset - boffset) bytes. 3309 */ 3310 3311 for (i = 0; i < bp->b_npages; i++) { 3312 if (bp->b_pages[i]->dirty) 3313 break; 3314 } 3315 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3316 3317 for (i = bp->b_npages - 1; i >= 0; --i) { 3318 if (bp->b_pages[i]->dirty) { 3319 break; 3320 } 3321 } 3322 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3323 3324 /* 3325 * Fit it to the buffer. 3326 */ 3327 3328 if (eoffset > bp->b_bcount) 3329 eoffset = bp->b_bcount; 3330 3331 /* 3332 * If we have a good dirty range, merge with the existing 3333 * dirty range. 3334 */ 3335 3336 if (boffset < eoffset) { 3337 if (bp->b_dirtyoff > boffset) 3338 bp->b_dirtyoff = boffset; 3339 if (bp->b_dirtyend < eoffset) 3340 bp->b_dirtyend = eoffset; 3341 } 3342 } 3343 } 3344 3345 /* 3346 * Allocate the KVA mapping for an existing buffer. 3347 * If an unmapped buffer is provided but a mapped buffer is requested, take 3348 * also care to properly setup mappings between pages and KVA. 3349 */ 3350 static void 3351 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3352 { 3353 int bsize, maxsize, need_mapping, need_kva; 3354 off_t offset; 3355 3356 need_mapping = bp->b_data == unmapped_buf && 3357 (gbflags & GB_UNMAPPED) == 0; 3358 need_kva = bp->b_kvabase == unmapped_buf && 3359 bp->b_data == unmapped_buf && 3360 (gbflags & GB_KVAALLOC) != 0; 3361 if (!need_mapping && !need_kva) 3362 return; 3363 3364 BUF_CHECK_UNMAPPED(bp); 3365 3366 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3367 /* 3368 * Buffer is not mapped, but the KVA was already 3369 * reserved at the time of the instantiation. Use the 3370 * allocated space. 3371 */ 3372 goto has_addr; 3373 } 3374 3375 /* 3376 * Calculate the amount of the address space we would reserve 3377 * if the buffer was mapped. 3378 */ 3379 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3380 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3381 offset = blkno * bsize; 3382 maxsize = size + (offset & PAGE_MASK); 3383 maxsize = imax(maxsize, bsize); 3384 3385 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3386 if ((gbflags & GB_NOWAIT_BD) != 0) { 3387 /* 3388 * XXXKIB: defragmentation cannot 3389 * succeed, not sure what else to do. 3390 */ 3391 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3392 } 3393 atomic_add_int(&mappingrestarts, 1); 3394 bufspace_wait(bp->b_vp, gbflags, 0, 0); 3395 } 3396 has_addr: 3397 if (need_mapping) { 3398 /* b_offset is handled by bpmap_qenter. */ 3399 bp->b_data = bp->b_kvabase; 3400 BUF_CHECK_MAPPED(bp); 3401 bpmap_qenter(bp); 3402 } 3403 } 3404 3405 /* 3406 * getblk: 3407 * 3408 * Get a block given a specified block and offset into a file/device. 3409 * The buffers B_DONE bit will be cleared on return, making it almost 3410 * ready for an I/O initiation. B_INVAL may or may not be set on 3411 * return. The caller should clear B_INVAL prior to initiating a 3412 * READ. 3413 * 3414 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3415 * an existing buffer. 3416 * 3417 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3418 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3419 * and then cleared based on the backing VM. If the previous buffer is 3420 * non-0-sized but invalid, B_CACHE will be cleared. 3421 * 3422 * If getblk() must create a new buffer, the new buffer is returned with 3423 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3424 * case it is returned with B_INVAL clear and B_CACHE set based on the 3425 * backing VM. 3426 * 3427 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3428 * B_CACHE bit is clear. 3429 * 3430 * What this means, basically, is that the caller should use B_CACHE to 3431 * determine whether the buffer is fully valid or not and should clear 3432 * B_INVAL prior to issuing a read. If the caller intends to validate 3433 * the buffer by loading its data area with something, the caller needs 3434 * to clear B_INVAL. If the caller does this without issuing an I/O, 3435 * the caller should set B_CACHE ( as an optimization ), else the caller 3436 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3437 * a write attempt or if it was a successfull read. If the caller 3438 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3439 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3440 */ 3441 struct buf * 3442 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3443 int flags) 3444 { 3445 struct buf *bp; 3446 struct bufobj *bo; 3447 int bsize, error, maxsize, vmio; 3448 off_t offset; 3449 3450 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3451 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3452 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3453 ASSERT_VOP_LOCKED(vp, "getblk"); 3454 if (size > MAXBCACHEBUF) 3455 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size, 3456 MAXBCACHEBUF); 3457 if (!unmapped_buf_allowed) 3458 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3459 3460 bo = &vp->v_bufobj; 3461 loop: 3462 BO_RLOCK(bo); 3463 bp = gbincore(bo, blkno); 3464 if (bp != NULL) { 3465 int lockflags; 3466 /* 3467 * Buffer is in-core. If the buffer is not busy nor managed, 3468 * it must be on a queue. 3469 */ 3470 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3471 3472 if (flags & GB_LOCK_NOWAIT) 3473 lockflags |= LK_NOWAIT; 3474 3475 error = BUF_TIMELOCK(bp, lockflags, 3476 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3477 3478 /* 3479 * If we slept and got the lock we have to restart in case 3480 * the buffer changed identities. 3481 */ 3482 if (error == ENOLCK) 3483 goto loop; 3484 /* We timed out or were interrupted. */ 3485 else if (error) 3486 return (NULL); 3487 /* If recursed, assume caller knows the rules. */ 3488 else if (BUF_LOCKRECURSED(bp)) 3489 goto end; 3490 3491 /* 3492 * The buffer is locked. B_CACHE is cleared if the buffer is 3493 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3494 * and for a VMIO buffer B_CACHE is adjusted according to the 3495 * backing VM cache. 3496 */ 3497 if (bp->b_flags & B_INVAL) 3498 bp->b_flags &= ~B_CACHE; 3499 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3500 bp->b_flags |= B_CACHE; 3501 if (bp->b_flags & B_MANAGED) 3502 MPASS(bp->b_qindex == QUEUE_NONE); 3503 else 3504 bremfree(bp); 3505 3506 /* 3507 * check for size inconsistencies for non-VMIO case. 3508 */ 3509 if (bp->b_bcount != size) { 3510 if ((bp->b_flags & B_VMIO) == 0 || 3511 (size > bp->b_kvasize)) { 3512 if (bp->b_flags & B_DELWRI) { 3513 /* 3514 * If buffer is pinned and caller does 3515 * not want sleep waiting for it to be 3516 * unpinned, bail out 3517 * */ 3518 if (bp->b_pin_count > 0) { 3519 if (flags & GB_LOCK_NOWAIT) { 3520 bqrelse(bp); 3521 return (NULL); 3522 } else { 3523 bunpin_wait(bp); 3524 } 3525 } 3526 bp->b_flags |= B_NOCACHE; 3527 bwrite(bp); 3528 } else { 3529 if (LIST_EMPTY(&bp->b_dep)) { 3530 bp->b_flags |= B_RELBUF; 3531 brelse(bp); 3532 } else { 3533 bp->b_flags |= B_NOCACHE; 3534 bwrite(bp); 3535 } 3536 } 3537 goto loop; 3538 } 3539 } 3540 3541 /* 3542 * Handle the case of unmapped buffer which should 3543 * become mapped, or the buffer for which KVA 3544 * reservation is requested. 3545 */ 3546 bp_unmapped_get_kva(bp, blkno, size, flags); 3547 3548 /* 3549 * If the size is inconsistant in the VMIO case, we can resize 3550 * the buffer. This might lead to B_CACHE getting set or 3551 * cleared. If the size has not changed, B_CACHE remains 3552 * unchanged from its previous state. 3553 */ 3554 allocbuf(bp, size); 3555 3556 KASSERT(bp->b_offset != NOOFFSET, 3557 ("getblk: no buffer offset")); 3558 3559 /* 3560 * A buffer with B_DELWRI set and B_CACHE clear must 3561 * be committed before we can return the buffer in 3562 * order to prevent the caller from issuing a read 3563 * ( due to B_CACHE not being set ) and overwriting 3564 * it. 3565 * 3566 * Most callers, including NFS and FFS, need this to 3567 * operate properly either because they assume they 3568 * can issue a read if B_CACHE is not set, or because 3569 * ( for example ) an uncached B_DELWRI might loop due 3570 * to softupdates re-dirtying the buffer. In the latter 3571 * case, B_CACHE is set after the first write completes, 3572 * preventing further loops. 3573 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3574 * above while extending the buffer, we cannot allow the 3575 * buffer to remain with B_CACHE set after the write 3576 * completes or it will represent a corrupt state. To 3577 * deal with this we set B_NOCACHE to scrap the buffer 3578 * after the write. 3579 * 3580 * We might be able to do something fancy, like setting 3581 * B_CACHE in bwrite() except if B_DELWRI is already set, 3582 * so the below call doesn't set B_CACHE, but that gets real 3583 * confusing. This is much easier. 3584 */ 3585 3586 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3587 bp->b_flags |= B_NOCACHE; 3588 bwrite(bp); 3589 goto loop; 3590 } 3591 bp->b_flags &= ~B_DONE; 3592 } else { 3593 /* 3594 * Buffer is not in-core, create new buffer. The buffer 3595 * returned by getnewbuf() is locked. Note that the returned 3596 * buffer is also considered valid (not marked B_INVAL). 3597 */ 3598 BO_RUNLOCK(bo); 3599 /* 3600 * If the user does not want us to create the buffer, bail out 3601 * here. 3602 */ 3603 if (flags & GB_NOCREAT) 3604 return NULL; 3605 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3606 return NULL; 3607 3608 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3609 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3610 offset = blkno * bsize; 3611 vmio = vp->v_object != NULL; 3612 if (vmio) { 3613 maxsize = size + (offset & PAGE_MASK); 3614 } else { 3615 maxsize = size; 3616 /* Do not allow non-VMIO notmapped buffers. */ 3617 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3618 } 3619 maxsize = imax(maxsize, bsize); 3620 3621 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 3622 if (bp == NULL) { 3623 if (slpflag || slptimeo) 3624 return NULL; 3625 /* 3626 * XXX This is here until the sleep path is diagnosed 3627 * enough to work under very low memory conditions. 3628 * 3629 * There's an issue on low memory, 4BSD+non-preempt 3630 * systems (eg MIPS routers with 32MB RAM) where buffer 3631 * exhaustion occurs without sleeping for buffer 3632 * reclaimation. This just sticks in a loop and 3633 * constantly attempts to allocate a buffer, which 3634 * hits exhaustion and tries to wakeup bufdaemon. 3635 * This never happens because we never yield. 3636 * 3637 * The real solution is to identify and fix these cases 3638 * so we aren't effectively busy-waiting in a loop 3639 * until the reclaimation path has cycles to run. 3640 */ 3641 kern_yield(PRI_USER); 3642 goto loop; 3643 } 3644 3645 /* 3646 * This code is used to make sure that a buffer is not 3647 * created while the getnewbuf routine is blocked. 3648 * This can be a problem whether the vnode is locked or not. 3649 * If the buffer is created out from under us, we have to 3650 * throw away the one we just created. 3651 * 3652 * Note: this must occur before we associate the buffer 3653 * with the vp especially considering limitations in 3654 * the splay tree implementation when dealing with duplicate 3655 * lblkno's. 3656 */ 3657 BO_LOCK(bo); 3658 if (gbincore(bo, blkno)) { 3659 BO_UNLOCK(bo); 3660 bp->b_flags |= B_INVAL; 3661 brelse(bp); 3662 bufspace_release(maxsize); 3663 goto loop; 3664 } 3665 3666 /* 3667 * Insert the buffer into the hash, so that it can 3668 * be found by incore. 3669 */ 3670 bp->b_blkno = bp->b_lblkno = blkno; 3671 bp->b_offset = offset; 3672 bgetvp(vp, bp); 3673 BO_UNLOCK(bo); 3674 3675 /* 3676 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3677 * buffer size starts out as 0, B_CACHE will be set by 3678 * allocbuf() for the VMIO case prior to it testing the 3679 * backing store for validity. 3680 */ 3681 3682 if (vmio) { 3683 bp->b_flags |= B_VMIO; 3684 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3685 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3686 bp, vp->v_object, bp->b_bufobj->bo_object)); 3687 } else { 3688 bp->b_flags &= ~B_VMIO; 3689 KASSERT(bp->b_bufobj->bo_object == NULL, 3690 ("ARGH! has b_bufobj->bo_object %p %p\n", 3691 bp, bp->b_bufobj->bo_object)); 3692 BUF_CHECK_MAPPED(bp); 3693 } 3694 3695 allocbuf(bp, size); 3696 bufspace_release(maxsize); 3697 bp->b_flags &= ~B_DONE; 3698 } 3699 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3700 BUF_ASSERT_HELD(bp); 3701 end: 3702 KASSERT(bp->b_bufobj == bo, 3703 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3704 return (bp); 3705 } 3706 3707 /* 3708 * Get an empty, disassociated buffer of given size. The buffer is initially 3709 * set to B_INVAL. 3710 */ 3711 struct buf * 3712 geteblk(int size, int flags) 3713 { 3714 struct buf *bp; 3715 int maxsize; 3716 3717 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3718 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 3719 if ((flags & GB_NOWAIT_BD) && 3720 (curthread->td_pflags & TDP_BUFNEED) != 0) 3721 return (NULL); 3722 } 3723 allocbuf(bp, size); 3724 bufspace_release(maxsize); 3725 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3726 BUF_ASSERT_HELD(bp); 3727 return (bp); 3728 } 3729 3730 /* 3731 * Truncate the backing store for a non-vmio buffer. 3732 */ 3733 static void 3734 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 3735 { 3736 3737 if (bp->b_flags & B_MALLOC) { 3738 /* 3739 * malloced buffers are not shrunk 3740 */ 3741 if (newbsize == 0) { 3742 bufmallocadjust(bp, 0); 3743 free(bp->b_data, M_BIOBUF); 3744 bp->b_data = bp->b_kvabase; 3745 bp->b_flags &= ~B_MALLOC; 3746 } 3747 return; 3748 } 3749 vm_hold_free_pages(bp, newbsize); 3750 bufspace_adjust(bp, newbsize); 3751 } 3752 3753 /* 3754 * Extend the backing for a non-VMIO buffer. 3755 */ 3756 static void 3757 vfs_nonvmio_extend(struct buf *bp, int newbsize) 3758 { 3759 caddr_t origbuf; 3760 int origbufsize; 3761 3762 /* 3763 * We only use malloced memory on the first allocation. 3764 * and revert to page-allocated memory when the buffer 3765 * grows. 3766 * 3767 * There is a potential smp race here that could lead 3768 * to bufmallocspace slightly passing the max. It 3769 * is probably extremely rare and not worth worrying 3770 * over. 3771 */ 3772 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 3773 bufmallocspace < maxbufmallocspace) { 3774 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 3775 bp->b_flags |= B_MALLOC; 3776 bufmallocadjust(bp, newbsize); 3777 return; 3778 } 3779 3780 /* 3781 * If the buffer is growing on its other-than-first 3782 * allocation then we revert to the page-allocation 3783 * scheme. 3784 */ 3785 origbuf = NULL; 3786 origbufsize = 0; 3787 if (bp->b_flags & B_MALLOC) { 3788 origbuf = bp->b_data; 3789 origbufsize = bp->b_bufsize; 3790 bp->b_data = bp->b_kvabase; 3791 bufmallocadjust(bp, 0); 3792 bp->b_flags &= ~B_MALLOC; 3793 newbsize = round_page(newbsize); 3794 } 3795 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 3796 (vm_offset_t) bp->b_data + newbsize); 3797 if (origbuf != NULL) { 3798 bcopy(origbuf, bp->b_data, origbufsize); 3799 free(origbuf, M_BIOBUF); 3800 } 3801 bufspace_adjust(bp, newbsize); 3802 } 3803 3804 /* 3805 * This code constitutes the buffer memory from either anonymous system 3806 * memory (in the case of non-VMIO operations) or from an associated 3807 * VM object (in the case of VMIO operations). This code is able to 3808 * resize a buffer up or down. 3809 * 3810 * Note that this code is tricky, and has many complications to resolve 3811 * deadlock or inconsistant data situations. Tread lightly!!! 3812 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3813 * the caller. Calling this code willy nilly can result in the loss of data. 3814 * 3815 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3816 * B_CACHE for the non-VMIO case. 3817 */ 3818 int 3819 allocbuf(struct buf *bp, int size) 3820 { 3821 int newbsize; 3822 3823 BUF_ASSERT_HELD(bp); 3824 3825 if (bp->b_bcount == size) 3826 return (1); 3827 3828 if (bp->b_kvasize != 0 && bp->b_kvasize < size) 3829 panic("allocbuf: buffer too small"); 3830 3831 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3832 if ((bp->b_flags & B_VMIO) == 0) { 3833 if ((bp->b_flags & B_MALLOC) == 0) 3834 newbsize = round_page(newbsize); 3835 /* 3836 * Just get anonymous memory from the kernel. Don't 3837 * mess with B_CACHE. 3838 */ 3839 if (newbsize < bp->b_bufsize) 3840 vfs_nonvmio_truncate(bp, newbsize); 3841 else if (newbsize > bp->b_bufsize) 3842 vfs_nonvmio_extend(bp, newbsize); 3843 } else { 3844 int desiredpages; 3845 3846 desiredpages = (size == 0) ? 0 : 3847 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3848 3849 if (bp->b_flags & B_MALLOC) 3850 panic("allocbuf: VMIO buffer can't be malloced"); 3851 /* 3852 * Set B_CACHE initially if buffer is 0 length or will become 3853 * 0-length. 3854 */ 3855 if (size == 0 || bp->b_bufsize == 0) 3856 bp->b_flags |= B_CACHE; 3857 3858 if (newbsize < bp->b_bufsize) 3859 vfs_vmio_truncate(bp, desiredpages); 3860 /* XXX This looks as if it should be newbsize > b_bufsize */ 3861 else if (size > bp->b_bcount) 3862 vfs_vmio_extend(bp, desiredpages, size); 3863 bufspace_adjust(bp, newbsize); 3864 } 3865 bp->b_bcount = size; /* requested buffer size. */ 3866 return (1); 3867 } 3868 3869 extern int inflight_transient_maps; 3870 3871 void 3872 biodone(struct bio *bp) 3873 { 3874 struct mtx *mtxp; 3875 void (*done)(struct bio *); 3876 vm_offset_t start, end; 3877 3878 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3879 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3880 bp->bio_flags |= BIO_UNMAPPED; 3881 start = trunc_page((vm_offset_t)bp->bio_data); 3882 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3883 bp->bio_data = unmapped_buf; 3884 pmap_qremove(start, OFF_TO_IDX(end - start)); 3885 vmem_free(transient_arena, start, end - start); 3886 atomic_add_int(&inflight_transient_maps, -1); 3887 } 3888 done = bp->bio_done; 3889 if (done == NULL) { 3890 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3891 mtx_lock(mtxp); 3892 bp->bio_flags |= BIO_DONE; 3893 wakeup(bp); 3894 mtx_unlock(mtxp); 3895 } else { 3896 bp->bio_flags |= BIO_DONE; 3897 done(bp); 3898 } 3899 } 3900 3901 /* 3902 * Wait for a BIO to finish. 3903 */ 3904 int 3905 biowait(struct bio *bp, const char *wchan) 3906 { 3907 struct mtx *mtxp; 3908 3909 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3910 mtx_lock(mtxp); 3911 while ((bp->bio_flags & BIO_DONE) == 0) 3912 msleep(bp, mtxp, PRIBIO, wchan, 0); 3913 mtx_unlock(mtxp); 3914 if (bp->bio_error != 0) 3915 return (bp->bio_error); 3916 if (!(bp->bio_flags & BIO_ERROR)) 3917 return (0); 3918 return (EIO); 3919 } 3920 3921 void 3922 biofinish(struct bio *bp, struct devstat *stat, int error) 3923 { 3924 3925 if (error) { 3926 bp->bio_error = error; 3927 bp->bio_flags |= BIO_ERROR; 3928 } 3929 if (stat != NULL) 3930 devstat_end_transaction_bio(stat, bp); 3931 biodone(bp); 3932 } 3933 3934 /* 3935 * bufwait: 3936 * 3937 * Wait for buffer I/O completion, returning error status. The buffer 3938 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3939 * error and cleared. 3940 */ 3941 int 3942 bufwait(struct buf *bp) 3943 { 3944 if (bp->b_iocmd == BIO_READ) 3945 bwait(bp, PRIBIO, "biord"); 3946 else 3947 bwait(bp, PRIBIO, "biowr"); 3948 if (bp->b_flags & B_EINTR) { 3949 bp->b_flags &= ~B_EINTR; 3950 return (EINTR); 3951 } 3952 if (bp->b_ioflags & BIO_ERROR) { 3953 return (bp->b_error ? bp->b_error : EIO); 3954 } else { 3955 return (0); 3956 } 3957 } 3958 3959 /* 3960 * bufdone: 3961 * 3962 * Finish I/O on a buffer, optionally calling a completion function. 3963 * This is usually called from an interrupt so process blocking is 3964 * not allowed. 3965 * 3966 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3967 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3968 * assuming B_INVAL is clear. 3969 * 3970 * For the VMIO case, we set B_CACHE if the op was a read and no 3971 * read error occured, or if the op was a write. B_CACHE is never 3972 * set if the buffer is invalid or otherwise uncacheable. 3973 * 3974 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3975 * initiator to leave B_INVAL set to brelse the buffer out of existance 3976 * in the biodone routine. 3977 */ 3978 void 3979 bufdone(struct buf *bp) 3980 { 3981 struct bufobj *dropobj; 3982 void (*biodone)(struct buf *); 3983 3984 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3985 dropobj = NULL; 3986 3987 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3988 BUF_ASSERT_HELD(bp); 3989 3990 runningbufwakeup(bp); 3991 if (bp->b_iocmd == BIO_WRITE) 3992 dropobj = bp->b_bufobj; 3993 /* call optional completion function if requested */ 3994 if (bp->b_iodone != NULL) { 3995 biodone = bp->b_iodone; 3996 bp->b_iodone = NULL; 3997 (*biodone) (bp); 3998 if (dropobj) 3999 bufobj_wdrop(dropobj); 4000 return; 4001 } 4002 4003 bufdone_finish(bp); 4004 4005 if (dropobj) 4006 bufobj_wdrop(dropobj); 4007 } 4008 4009 void 4010 bufdone_finish(struct buf *bp) 4011 { 4012 BUF_ASSERT_HELD(bp); 4013 4014 if (!LIST_EMPTY(&bp->b_dep)) 4015 buf_complete(bp); 4016 4017 if (bp->b_flags & B_VMIO) { 4018 /* 4019 * Set B_CACHE if the op was a normal read and no error 4020 * occured. B_CACHE is set for writes in the b*write() 4021 * routines. 4022 */ 4023 if (bp->b_iocmd == BIO_READ && 4024 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4025 !(bp->b_ioflags & BIO_ERROR)) 4026 bp->b_flags |= B_CACHE; 4027 vfs_vmio_iodone(bp); 4028 } 4029 4030 /* 4031 * For asynchronous completions, release the buffer now. The brelse 4032 * will do a wakeup there if necessary - so no need to do a wakeup 4033 * here in the async case. The sync case always needs to do a wakeup. 4034 */ 4035 if (bp->b_flags & B_ASYNC) { 4036 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4037 (bp->b_ioflags & BIO_ERROR)) 4038 brelse(bp); 4039 else 4040 bqrelse(bp); 4041 } else 4042 bdone(bp); 4043 } 4044 4045 /* 4046 * This routine is called in lieu of iodone in the case of 4047 * incomplete I/O. This keeps the busy status for pages 4048 * consistant. 4049 */ 4050 void 4051 vfs_unbusy_pages(struct buf *bp) 4052 { 4053 int i; 4054 vm_object_t obj; 4055 vm_page_t m; 4056 4057 runningbufwakeup(bp); 4058 if (!(bp->b_flags & B_VMIO)) 4059 return; 4060 4061 obj = bp->b_bufobj->bo_object; 4062 VM_OBJECT_WLOCK(obj); 4063 for (i = 0; i < bp->b_npages; i++) { 4064 m = bp->b_pages[i]; 4065 if (m == bogus_page) { 4066 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4067 if (!m) 4068 panic("vfs_unbusy_pages: page missing\n"); 4069 bp->b_pages[i] = m; 4070 if (buf_mapped(bp)) { 4071 BUF_CHECK_MAPPED(bp); 4072 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4073 bp->b_pages, bp->b_npages); 4074 } else 4075 BUF_CHECK_UNMAPPED(bp); 4076 } 4077 vm_page_sunbusy(m); 4078 } 4079 vm_object_pip_wakeupn(obj, bp->b_npages); 4080 VM_OBJECT_WUNLOCK(obj); 4081 } 4082 4083 /* 4084 * vfs_page_set_valid: 4085 * 4086 * Set the valid bits in a page based on the supplied offset. The 4087 * range is restricted to the buffer's size. 4088 * 4089 * This routine is typically called after a read completes. 4090 */ 4091 static void 4092 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4093 { 4094 vm_ooffset_t eoff; 4095 4096 /* 4097 * Compute the end offset, eoff, such that [off, eoff) does not span a 4098 * page boundary and eoff is not greater than the end of the buffer. 4099 * The end of the buffer, in this case, is our file EOF, not the 4100 * allocation size of the buffer. 4101 */ 4102 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4103 if (eoff > bp->b_offset + bp->b_bcount) 4104 eoff = bp->b_offset + bp->b_bcount; 4105 4106 /* 4107 * Set valid range. This is typically the entire buffer and thus the 4108 * entire page. 4109 */ 4110 if (eoff > off) 4111 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4112 } 4113 4114 /* 4115 * vfs_page_set_validclean: 4116 * 4117 * Set the valid bits and clear the dirty bits in a page based on the 4118 * supplied offset. The range is restricted to the buffer's size. 4119 */ 4120 static void 4121 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4122 { 4123 vm_ooffset_t soff, eoff; 4124 4125 /* 4126 * Start and end offsets in buffer. eoff - soff may not cross a 4127 * page boundry or cross the end of the buffer. The end of the 4128 * buffer, in this case, is our file EOF, not the allocation size 4129 * of the buffer. 4130 */ 4131 soff = off; 4132 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4133 if (eoff > bp->b_offset + bp->b_bcount) 4134 eoff = bp->b_offset + bp->b_bcount; 4135 4136 /* 4137 * Set valid range. This is typically the entire buffer and thus the 4138 * entire page. 4139 */ 4140 if (eoff > soff) { 4141 vm_page_set_validclean( 4142 m, 4143 (vm_offset_t) (soff & PAGE_MASK), 4144 (vm_offset_t) (eoff - soff) 4145 ); 4146 } 4147 } 4148 4149 /* 4150 * Ensure that all buffer pages are not exclusive busied. If any page is 4151 * exclusive busy, drain it. 4152 */ 4153 void 4154 vfs_drain_busy_pages(struct buf *bp) 4155 { 4156 vm_page_t m; 4157 int i, last_busied; 4158 4159 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4160 last_busied = 0; 4161 for (i = 0; i < bp->b_npages; i++) { 4162 m = bp->b_pages[i]; 4163 if (vm_page_xbusied(m)) { 4164 for (; last_busied < i; last_busied++) 4165 vm_page_sbusy(bp->b_pages[last_busied]); 4166 while (vm_page_xbusied(m)) { 4167 vm_page_lock(m); 4168 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4169 vm_page_busy_sleep(m, "vbpage"); 4170 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4171 } 4172 } 4173 } 4174 for (i = 0; i < last_busied; i++) 4175 vm_page_sunbusy(bp->b_pages[i]); 4176 } 4177 4178 /* 4179 * This routine is called before a device strategy routine. 4180 * It is used to tell the VM system that paging I/O is in 4181 * progress, and treat the pages associated with the buffer 4182 * almost as being exclusive busy. Also the object paging_in_progress 4183 * flag is handled to make sure that the object doesn't become 4184 * inconsistant. 4185 * 4186 * Since I/O has not been initiated yet, certain buffer flags 4187 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4188 * and should be ignored. 4189 */ 4190 void 4191 vfs_busy_pages(struct buf *bp, int clear_modify) 4192 { 4193 int i, bogus; 4194 vm_object_t obj; 4195 vm_ooffset_t foff; 4196 vm_page_t m; 4197 4198 if (!(bp->b_flags & B_VMIO)) 4199 return; 4200 4201 obj = bp->b_bufobj->bo_object; 4202 foff = bp->b_offset; 4203 KASSERT(bp->b_offset != NOOFFSET, 4204 ("vfs_busy_pages: no buffer offset")); 4205 VM_OBJECT_WLOCK(obj); 4206 vfs_drain_busy_pages(bp); 4207 if (bp->b_bufsize != 0) 4208 vfs_setdirty_locked_object(bp); 4209 bogus = 0; 4210 for (i = 0; i < bp->b_npages; i++) { 4211 m = bp->b_pages[i]; 4212 4213 if ((bp->b_flags & B_CLUSTER) == 0) { 4214 vm_object_pip_add(obj, 1); 4215 vm_page_sbusy(m); 4216 } 4217 /* 4218 * When readying a buffer for a read ( i.e 4219 * clear_modify == 0 ), it is important to do 4220 * bogus_page replacement for valid pages in 4221 * partially instantiated buffers. Partially 4222 * instantiated buffers can, in turn, occur when 4223 * reconstituting a buffer from its VM backing store 4224 * base. We only have to do this if B_CACHE is 4225 * clear ( which causes the I/O to occur in the 4226 * first place ). The replacement prevents the read 4227 * I/O from overwriting potentially dirty VM-backed 4228 * pages. XXX bogus page replacement is, uh, bogus. 4229 * It may not work properly with small-block devices. 4230 * We need to find a better way. 4231 */ 4232 if (clear_modify) { 4233 pmap_remove_write(m); 4234 vfs_page_set_validclean(bp, foff, m); 4235 } else if (m->valid == VM_PAGE_BITS_ALL && 4236 (bp->b_flags & B_CACHE) == 0) { 4237 bp->b_pages[i] = bogus_page; 4238 bogus++; 4239 } 4240 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4241 } 4242 VM_OBJECT_WUNLOCK(obj); 4243 if (bogus && buf_mapped(bp)) { 4244 BUF_CHECK_MAPPED(bp); 4245 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4246 bp->b_pages, bp->b_npages); 4247 } 4248 } 4249 4250 /* 4251 * vfs_bio_set_valid: 4252 * 4253 * Set the range within the buffer to valid. The range is 4254 * relative to the beginning of the buffer, b_offset. Note that 4255 * b_offset itself may be offset from the beginning of the first 4256 * page. 4257 */ 4258 void 4259 vfs_bio_set_valid(struct buf *bp, int base, int size) 4260 { 4261 int i, n; 4262 vm_page_t m; 4263 4264 if (!(bp->b_flags & B_VMIO)) 4265 return; 4266 4267 /* 4268 * Fixup base to be relative to beginning of first page. 4269 * Set initial n to be the maximum number of bytes in the 4270 * first page that can be validated. 4271 */ 4272 base += (bp->b_offset & PAGE_MASK); 4273 n = PAGE_SIZE - (base & PAGE_MASK); 4274 4275 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4276 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4277 m = bp->b_pages[i]; 4278 if (n > size) 4279 n = size; 4280 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4281 base += n; 4282 size -= n; 4283 n = PAGE_SIZE; 4284 } 4285 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4286 } 4287 4288 /* 4289 * vfs_bio_clrbuf: 4290 * 4291 * If the specified buffer is a non-VMIO buffer, clear the entire 4292 * buffer. If the specified buffer is a VMIO buffer, clear and 4293 * validate only the previously invalid portions of the buffer. 4294 * This routine essentially fakes an I/O, so we need to clear 4295 * BIO_ERROR and B_INVAL. 4296 * 4297 * Note that while we only theoretically need to clear through b_bcount, 4298 * we go ahead and clear through b_bufsize. 4299 */ 4300 void 4301 vfs_bio_clrbuf(struct buf *bp) 4302 { 4303 int i, j, mask, sa, ea, slide; 4304 4305 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4306 clrbuf(bp); 4307 return; 4308 } 4309 bp->b_flags &= ~B_INVAL; 4310 bp->b_ioflags &= ~BIO_ERROR; 4311 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4312 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4313 (bp->b_offset & PAGE_MASK) == 0) { 4314 if (bp->b_pages[0] == bogus_page) 4315 goto unlock; 4316 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4317 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4318 if ((bp->b_pages[0]->valid & mask) == mask) 4319 goto unlock; 4320 if ((bp->b_pages[0]->valid & mask) == 0) { 4321 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4322 bp->b_pages[0]->valid |= mask; 4323 goto unlock; 4324 } 4325 } 4326 sa = bp->b_offset & PAGE_MASK; 4327 slide = 0; 4328 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4329 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4330 ea = slide & PAGE_MASK; 4331 if (ea == 0) 4332 ea = PAGE_SIZE; 4333 if (bp->b_pages[i] == bogus_page) 4334 continue; 4335 j = sa / DEV_BSIZE; 4336 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4337 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4338 if ((bp->b_pages[i]->valid & mask) == mask) 4339 continue; 4340 if ((bp->b_pages[i]->valid & mask) == 0) 4341 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4342 else { 4343 for (; sa < ea; sa += DEV_BSIZE, j++) { 4344 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4345 pmap_zero_page_area(bp->b_pages[i], 4346 sa, DEV_BSIZE); 4347 } 4348 } 4349 } 4350 bp->b_pages[i]->valid |= mask; 4351 } 4352 unlock: 4353 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4354 bp->b_resid = 0; 4355 } 4356 4357 void 4358 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4359 { 4360 vm_page_t m; 4361 int i, n; 4362 4363 if (buf_mapped(bp)) { 4364 BUF_CHECK_MAPPED(bp); 4365 bzero(bp->b_data + base, size); 4366 } else { 4367 BUF_CHECK_UNMAPPED(bp); 4368 n = PAGE_SIZE - (base & PAGE_MASK); 4369 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4370 m = bp->b_pages[i]; 4371 if (n > size) 4372 n = size; 4373 pmap_zero_page_area(m, base & PAGE_MASK, n); 4374 base += n; 4375 size -= n; 4376 n = PAGE_SIZE; 4377 } 4378 } 4379 } 4380 4381 /* 4382 * vm_hold_load_pages and vm_hold_free_pages get pages into 4383 * a buffers address space. The pages are anonymous and are 4384 * not associated with a file object. 4385 */ 4386 static void 4387 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4388 { 4389 vm_offset_t pg; 4390 vm_page_t p; 4391 int index; 4392 4393 BUF_CHECK_MAPPED(bp); 4394 4395 to = round_page(to); 4396 from = round_page(from); 4397 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4398 4399 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4400 tryagain: 4401 /* 4402 * note: must allocate system pages since blocking here 4403 * could interfere with paging I/O, no matter which 4404 * process we are. 4405 */ 4406 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4407 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4408 if (p == NULL) { 4409 VM_WAIT; 4410 goto tryagain; 4411 } 4412 pmap_qenter(pg, &p, 1); 4413 bp->b_pages[index] = p; 4414 } 4415 bp->b_npages = index; 4416 } 4417 4418 /* Return pages associated with this buf to the vm system */ 4419 static void 4420 vm_hold_free_pages(struct buf *bp, int newbsize) 4421 { 4422 vm_offset_t from; 4423 vm_page_t p; 4424 int index, newnpages; 4425 4426 BUF_CHECK_MAPPED(bp); 4427 4428 from = round_page((vm_offset_t)bp->b_data + newbsize); 4429 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4430 if (bp->b_npages > newnpages) 4431 pmap_qremove(from, bp->b_npages - newnpages); 4432 for (index = newnpages; index < bp->b_npages; index++) { 4433 p = bp->b_pages[index]; 4434 bp->b_pages[index] = NULL; 4435 if (vm_page_sbusied(p)) 4436 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4437 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4438 p->wire_count--; 4439 vm_page_free(p); 4440 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 4441 } 4442 bp->b_npages = newnpages; 4443 } 4444 4445 /* 4446 * Map an IO request into kernel virtual address space. 4447 * 4448 * All requests are (re)mapped into kernel VA space. 4449 * Notice that we use b_bufsize for the size of the buffer 4450 * to be mapped. b_bcount might be modified by the driver. 4451 * 4452 * Note that even if the caller determines that the address space should 4453 * be valid, a race or a smaller-file mapped into a larger space may 4454 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4455 * check the return value. 4456 * 4457 * This function only works with pager buffers. 4458 */ 4459 int 4460 vmapbuf(struct buf *bp, int mapbuf) 4461 { 4462 vm_prot_t prot; 4463 int pidx; 4464 4465 if (bp->b_bufsize < 0) 4466 return (-1); 4467 prot = VM_PROT_READ; 4468 if (bp->b_iocmd == BIO_READ) 4469 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4470 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4471 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4472 btoc(MAXPHYS))) < 0) 4473 return (-1); 4474 bp->b_npages = pidx; 4475 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4476 if (mapbuf || !unmapped_buf_allowed) { 4477 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 4478 bp->b_data = bp->b_kvabase + bp->b_offset; 4479 } else 4480 bp->b_data = unmapped_buf; 4481 return(0); 4482 } 4483 4484 /* 4485 * Free the io map PTEs associated with this IO operation. 4486 * We also invalidate the TLB entries and restore the original b_addr. 4487 * 4488 * This function only works with pager buffers. 4489 */ 4490 void 4491 vunmapbuf(struct buf *bp) 4492 { 4493 int npages; 4494 4495 npages = bp->b_npages; 4496 if (buf_mapped(bp)) 4497 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4498 vm_page_unhold_pages(bp->b_pages, npages); 4499 4500 bp->b_data = unmapped_buf; 4501 } 4502 4503 void 4504 bdone(struct buf *bp) 4505 { 4506 struct mtx *mtxp; 4507 4508 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4509 mtx_lock(mtxp); 4510 bp->b_flags |= B_DONE; 4511 wakeup(bp); 4512 mtx_unlock(mtxp); 4513 } 4514 4515 void 4516 bwait(struct buf *bp, u_char pri, const char *wchan) 4517 { 4518 struct mtx *mtxp; 4519 4520 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4521 mtx_lock(mtxp); 4522 while ((bp->b_flags & B_DONE) == 0) 4523 msleep(bp, mtxp, pri, wchan, 0); 4524 mtx_unlock(mtxp); 4525 } 4526 4527 int 4528 bufsync(struct bufobj *bo, int waitfor) 4529 { 4530 4531 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4532 } 4533 4534 void 4535 bufstrategy(struct bufobj *bo, struct buf *bp) 4536 { 4537 int i = 0; 4538 struct vnode *vp; 4539 4540 vp = bp->b_vp; 4541 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4542 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4543 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4544 i = VOP_STRATEGY(vp, bp); 4545 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4546 } 4547 4548 void 4549 bufobj_wrefl(struct bufobj *bo) 4550 { 4551 4552 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4553 ASSERT_BO_WLOCKED(bo); 4554 bo->bo_numoutput++; 4555 } 4556 4557 void 4558 bufobj_wref(struct bufobj *bo) 4559 { 4560 4561 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4562 BO_LOCK(bo); 4563 bo->bo_numoutput++; 4564 BO_UNLOCK(bo); 4565 } 4566 4567 void 4568 bufobj_wdrop(struct bufobj *bo) 4569 { 4570 4571 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4572 BO_LOCK(bo); 4573 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4574 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4575 bo->bo_flag &= ~BO_WWAIT; 4576 wakeup(&bo->bo_numoutput); 4577 } 4578 BO_UNLOCK(bo); 4579 } 4580 4581 int 4582 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4583 { 4584 int error; 4585 4586 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4587 ASSERT_BO_WLOCKED(bo); 4588 error = 0; 4589 while (bo->bo_numoutput) { 4590 bo->bo_flag |= BO_WWAIT; 4591 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4592 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4593 if (error) 4594 break; 4595 } 4596 return (error); 4597 } 4598 4599 void 4600 bpin(struct buf *bp) 4601 { 4602 struct mtx *mtxp; 4603 4604 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4605 mtx_lock(mtxp); 4606 bp->b_pin_count++; 4607 mtx_unlock(mtxp); 4608 } 4609 4610 void 4611 bunpin(struct buf *bp) 4612 { 4613 struct mtx *mtxp; 4614 4615 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4616 mtx_lock(mtxp); 4617 if (--bp->b_pin_count == 0) 4618 wakeup(bp); 4619 mtx_unlock(mtxp); 4620 } 4621 4622 void 4623 bunpin_wait(struct buf *bp) 4624 { 4625 struct mtx *mtxp; 4626 4627 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4628 mtx_lock(mtxp); 4629 while (bp->b_pin_count > 0) 4630 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4631 mtx_unlock(mtxp); 4632 } 4633 4634 /* 4635 * Set bio_data or bio_ma for struct bio from the struct buf. 4636 */ 4637 void 4638 bdata2bio(struct buf *bp, struct bio *bip) 4639 { 4640 4641 if (!buf_mapped(bp)) { 4642 KASSERT(unmapped_buf_allowed, ("unmapped")); 4643 bip->bio_ma = bp->b_pages; 4644 bip->bio_ma_n = bp->b_npages; 4645 bip->bio_data = unmapped_buf; 4646 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4647 bip->bio_flags |= BIO_UNMAPPED; 4648 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4649 PAGE_SIZE == bp->b_npages, 4650 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4651 (long long)bip->bio_length, bip->bio_ma_n)); 4652 } else { 4653 bip->bio_data = bp->b_data; 4654 bip->bio_ma = NULL; 4655 } 4656 } 4657 4658 #include "opt_ddb.h" 4659 #ifdef DDB 4660 #include <ddb/ddb.h> 4661 4662 /* DDB command to show buffer data */ 4663 DB_SHOW_COMMAND(buffer, db_show_buffer) 4664 { 4665 /* get args */ 4666 struct buf *bp = (struct buf *)addr; 4667 4668 if (!have_addr) { 4669 db_printf("usage: show buffer <addr>\n"); 4670 return; 4671 } 4672 4673 db_printf("buf at %p\n", bp); 4674 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4675 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4676 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4677 db_printf( 4678 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4679 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4680 "b_dep = %p\n", 4681 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4682 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4683 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4684 db_printf("b_kvabase = %p, b_kvasize = %d\n", 4685 bp->b_kvabase, bp->b_kvasize); 4686 if (bp->b_npages) { 4687 int i; 4688 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4689 for (i = 0; i < bp->b_npages; i++) { 4690 vm_page_t m; 4691 m = bp->b_pages[i]; 4692 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4693 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4694 if ((i + 1) < bp->b_npages) 4695 db_printf(","); 4696 } 4697 db_printf("\n"); 4698 } 4699 db_printf(" "); 4700 BUF_LOCKPRINTINFO(bp); 4701 } 4702 4703 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4704 { 4705 struct buf *bp; 4706 int i; 4707 4708 for (i = 0; i < nbuf; i++) { 4709 bp = &buf[i]; 4710 if (BUF_ISLOCKED(bp)) { 4711 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4712 db_printf("\n"); 4713 } 4714 } 4715 } 4716 4717 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4718 { 4719 struct vnode *vp; 4720 struct buf *bp; 4721 4722 if (!have_addr) { 4723 db_printf("usage: show vnodebufs <addr>\n"); 4724 return; 4725 } 4726 vp = (struct vnode *)addr; 4727 db_printf("Clean buffers:\n"); 4728 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4729 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4730 db_printf("\n"); 4731 } 4732 db_printf("Dirty buffers:\n"); 4733 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4734 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4735 db_printf("\n"); 4736 } 4737 } 4738 4739 DB_COMMAND(countfreebufs, db_coundfreebufs) 4740 { 4741 struct buf *bp; 4742 int i, used = 0, nfree = 0; 4743 4744 if (have_addr) { 4745 db_printf("usage: countfreebufs\n"); 4746 return; 4747 } 4748 4749 for (i = 0; i < nbuf; i++) { 4750 bp = &buf[i]; 4751 if (bp->b_qindex == QUEUE_EMPTY) 4752 nfree++; 4753 else 4754 used++; 4755 } 4756 4757 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4758 nfree + used); 4759 db_printf("numfreebuffers is %d\n", numfreebuffers); 4760 } 4761 #endif /* DDB */ 4762