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/racct.h> 65 #include <sys/resourcevar.h> 66 #include <sys/rwlock.h> 67 #include <sys/smp.h> 68 #include <sys/sysctl.h> 69 #include <sys/sysproto.h> 70 #include <sys/vmem.h> 71 #include <sys/vmmeter.h> 72 #include <sys/vnode.h> 73 #include <sys/watchdog.h> 74 #include <geom/geom.h> 75 #include <vm/vm.h> 76 #include <vm/vm_param.h> 77 #include <vm/vm_kern.h> 78 #include <vm/vm_object.h> 79 #include <vm/vm_page.h> 80 #include <vm/vm_pageout.h> 81 #include <vm/vm_pager.h> 82 #include <vm/vm_extern.h> 83 #include <vm/vm_map.h> 84 #include <vm/swap_pager.h> 85 #include "opt_compat.h" 86 #include "opt_swap.h" 87 88 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 89 90 struct bio_ops bioops; /* I/O operation notification */ 91 92 struct buf_ops buf_ops_bio = { 93 .bop_name = "buf_ops_bio", 94 .bop_write = bufwrite, 95 .bop_strategy = bufstrategy, 96 .bop_sync = bufsync, 97 .bop_bdflush = bufbdflush, 98 }; 99 100 static struct buf *buf; /* buffer header pool */ 101 extern struct buf *swbuf; /* Swap buffer header pool. */ 102 caddr_t unmapped_buf; 103 104 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 105 struct proc *bufdaemonproc; 106 struct proc *bufspacedaemonproc; 107 108 static int inmem(struct vnode *vp, daddr_t blkno); 109 static void vm_hold_free_pages(struct buf *bp, int newbsize); 110 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 111 vm_offset_t to); 112 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 113 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 114 vm_page_t m); 115 static void vfs_clean_pages_dirty_buf(struct buf *bp); 116 static void vfs_setdirty_locked_object(struct buf *bp); 117 static void vfs_vmio_invalidate(struct buf *bp); 118 static void vfs_vmio_truncate(struct buf *bp, int npages); 119 static void vfs_vmio_extend(struct buf *bp, int npages, int size); 120 static int vfs_bio_clcheck(struct vnode *vp, int size, 121 daddr_t lblkno, daddr_t blkno); 122 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, 123 void (*)(struct buf *)); 124 static int buf_flush(struct vnode *vp, int); 125 static int buf_recycle(bool); 126 static int buf_scan(bool); 127 static int flushbufqueues(struct vnode *, int, int); 128 static void buf_daemon(void); 129 static void bremfreel(struct buf *bp); 130 static __inline void bd_wakeup(void); 131 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 132 static void bufkva_reclaim(vmem_t *, int); 133 static void bufkva_free(struct buf *); 134 static int buf_import(void *, void **, int, int); 135 static void buf_release(void *, void **, int); 136 static void maxbcachebuf_adjust(void); 137 138 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 139 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 140 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 141 #endif 142 143 int vmiodirenable = TRUE; 144 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 145 "Use the VM system for directory writes"); 146 long runningbufspace; 147 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 148 "Amount of presently outstanding async buffer io"); 149 static long bufspace; 150 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 151 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 152 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 153 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 154 #else 155 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 156 "Physical memory used for buffers"); 157 #endif 158 static long bufkvaspace; 159 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0, 160 "Kernel virtual memory used for buffers"); 161 static long maxbufspace; 162 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, &maxbufspace, 0, 163 "Maximum allowed value of bufspace (including metadata)"); 164 static long bufmallocspace; 165 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 166 "Amount of malloced memory for buffers"); 167 static long maxbufmallocspace; 168 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 169 0, "Maximum amount of malloced memory for buffers"); 170 static long lobufspace; 171 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RW, &lobufspace, 0, 172 "Minimum amount of buffers we want to have"); 173 long hibufspace; 174 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RW, &hibufspace, 0, 175 "Maximum allowed value of bufspace (excluding metadata)"); 176 long bufspacethresh; 177 SYSCTL_LONG(_vfs, OID_AUTO, bufspacethresh, CTLFLAG_RW, &bufspacethresh, 178 0, "Bufspace consumed before waking the daemon to free some"); 179 static int buffreekvacnt; 180 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 181 "Number of times we have freed the KVA space from some buffer"); 182 static int bufdefragcnt; 183 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 184 "Number of times we have had to repeat buffer allocation to defragment"); 185 static long lorunningspace; 186 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 187 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 188 "Minimum preferred space used for in-progress I/O"); 189 static long hirunningspace; 190 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 191 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 192 "Maximum amount of space to use for in-progress I/O"); 193 int dirtybufferflushes; 194 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 195 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 196 int bdwriteskip; 197 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 198 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 199 int altbufferflushes; 200 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 201 0, "Number of fsync flushes to limit dirty buffers"); 202 static int recursiveflushes; 203 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 204 0, "Number of flushes skipped due to being recursive"); 205 static int numdirtybuffers; 206 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 207 "Number of buffers that are dirty (has unwritten changes) at the moment"); 208 static int lodirtybuffers; 209 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 210 "How many buffers we want to have free before bufdaemon can sleep"); 211 static int hidirtybuffers; 212 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 213 "When the number of dirty buffers is considered severe"); 214 int dirtybufthresh; 215 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 216 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 217 static int numfreebuffers; 218 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 219 "Number of free buffers"); 220 static int lofreebuffers; 221 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 222 "Target number of free buffers"); 223 static int hifreebuffers; 224 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 225 "Threshold for clean buffer recycling"); 226 static int getnewbufcalls; 227 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 228 "Number of calls to getnewbuf"); 229 static int getnewbufrestarts; 230 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 231 "Number of times getnewbuf has had to restart a buffer acquisition"); 232 static int mappingrestarts; 233 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 234 "Number of times getblk has had to restart a buffer mapping for " 235 "unmapped buffer"); 236 static int numbufallocfails; 237 SYSCTL_INT(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, 0, 238 "Number of times buffer allocations failed"); 239 static int flushbufqtarget = 100; 240 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 241 "Amount of work to do in flushbufqueues when helping bufdaemon"); 242 static long notbufdflushes; 243 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0, 244 "Number of dirty buffer flushes done by the bufdaemon helpers"); 245 static long barrierwrites; 246 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 247 "Number of barrier writes"); 248 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 249 &unmapped_buf_allowed, 0, 250 "Permit the use of the unmapped i/o"); 251 int maxbcachebuf = MAXBCACHEBUF; 252 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, 253 "Maximum size of a buffer cache block"); 254 255 /* 256 * This lock synchronizes access to bd_request. 257 */ 258 static struct mtx_padalign __exclusive_cache_line bdlock; 259 260 /* 261 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 262 * waitrunningbufspace(). 263 */ 264 static struct mtx_padalign __exclusive_cache_line rbreqlock; 265 266 /* 267 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 268 */ 269 static struct rwlock_padalign __exclusive_cache_line nblock; 270 271 /* 272 * Lock that protects bdirtywait. 273 */ 274 static struct mtx_padalign __exclusive_cache_line bdirtylock; 275 276 /* 277 * Wakeup point for bufdaemon, as well as indicator of whether it is already 278 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 279 * is idling. 280 */ 281 static int bd_request; 282 283 /* 284 * Request/wakeup point for the bufspace daemon. 285 */ 286 static int bufspace_request; 287 288 /* 289 * Request for the buf daemon to write more buffers than is indicated by 290 * lodirtybuf. This may be necessary to push out excess dependencies or 291 * defragment the address space where a simple count of the number of dirty 292 * buffers is insufficient to characterize the demand for flushing them. 293 */ 294 static int bd_speedupreq; 295 296 /* 297 * Synchronization (sleep/wakeup) variable for active buffer space requests. 298 * Set when wait starts, cleared prior to wakeup(). 299 * Used in runningbufwakeup() and waitrunningbufspace(). 300 */ 301 static int runningbufreq; 302 303 /* 304 * Synchronization (sleep/wakeup) variable for buffer requests. 305 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 306 * by and/or. 307 * Used in numdirtywakeup(), bufspace_wakeup(), bwillwrite(), 308 * getnewbuf(), and getblk(). 309 */ 310 static volatile int needsbuffer; 311 312 /* 313 * Synchronization for bwillwrite() waiters. 314 */ 315 static int bdirtywait; 316 317 /* 318 * Definitions for the buffer free lists. 319 */ 320 #define QUEUE_NONE 0 /* on no queue */ 321 #define QUEUE_EMPTY 1 /* empty buffer headers */ 322 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 323 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ 324 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 325 326 /* Maximum number of clean buffer queues. */ 327 #define CLEAN_QUEUES 16 328 329 /* Configured number of clean queues. */ 330 static int clean_queues; 331 332 /* Maximum number of buffer queues. */ 333 #define BUFFER_QUEUES (QUEUE_CLEAN + CLEAN_QUEUES) 334 335 /* Queues for free buffers with various properties */ 336 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 337 #ifdef INVARIANTS 338 static int bq_len[BUFFER_QUEUES]; 339 #endif 340 341 /* 342 * Lock for each bufqueue 343 */ 344 static struct mtx_padalign __exclusive_cache_line bqlocks[BUFFER_QUEUES]; 345 346 /* 347 * per-cpu empty buffer cache. 348 */ 349 uma_zone_t buf_zone; 350 351 /* 352 * Single global constant for BUF_WMESG, to avoid getting multiple references. 353 * buf_wmesg is referred from macros. 354 */ 355 const char *buf_wmesg = BUF_WMESG; 356 357 static int 358 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 359 { 360 long value; 361 int error; 362 363 value = *(long *)arg1; 364 error = sysctl_handle_long(oidp, &value, 0, req); 365 if (error != 0 || req->newptr == NULL) 366 return (error); 367 mtx_lock(&rbreqlock); 368 if (arg1 == &hirunningspace) { 369 if (value < lorunningspace) 370 error = EINVAL; 371 else 372 hirunningspace = value; 373 } else { 374 KASSERT(arg1 == &lorunningspace, 375 ("%s: unknown arg1", __func__)); 376 if (value > hirunningspace) 377 error = EINVAL; 378 else 379 lorunningspace = value; 380 } 381 mtx_unlock(&rbreqlock); 382 return (error); 383 } 384 385 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 386 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 387 static int 388 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 389 { 390 long lvalue; 391 int ivalue; 392 393 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 394 return (sysctl_handle_long(oidp, arg1, arg2, req)); 395 lvalue = *(long *)arg1; 396 if (lvalue > INT_MAX) 397 /* On overflow, still write out a long to trigger ENOMEM. */ 398 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 399 ivalue = lvalue; 400 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 401 } 402 #endif 403 404 static int 405 bqcleanq(void) 406 { 407 static int nextq; 408 409 return ((atomic_fetchadd_int(&nextq, 1) % clean_queues) + QUEUE_CLEAN); 410 } 411 412 static int 413 bqisclean(int qindex) 414 { 415 416 return (qindex >= QUEUE_CLEAN && qindex < QUEUE_CLEAN + CLEAN_QUEUES); 417 } 418 419 /* 420 * bqlock: 421 * 422 * Return the appropriate queue lock based on the index. 423 */ 424 static inline struct mtx * 425 bqlock(int qindex) 426 { 427 428 return (struct mtx *)&bqlocks[qindex]; 429 } 430 431 /* 432 * bdirtywakeup: 433 * 434 * Wakeup any bwillwrite() waiters. 435 */ 436 static void 437 bdirtywakeup(void) 438 { 439 mtx_lock(&bdirtylock); 440 if (bdirtywait) { 441 bdirtywait = 0; 442 wakeup(&bdirtywait); 443 } 444 mtx_unlock(&bdirtylock); 445 } 446 447 /* 448 * bdirtysub: 449 * 450 * Decrement the numdirtybuffers count by one and wakeup any 451 * threads blocked in bwillwrite(). 452 */ 453 static void 454 bdirtysub(void) 455 { 456 457 if (atomic_fetchadd_int(&numdirtybuffers, -1) == 458 (lodirtybuffers + hidirtybuffers) / 2) 459 bdirtywakeup(); 460 } 461 462 /* 463 * bdirtyadd: 464 * 465 * Increment the numdirtybuffers count by one and wakeup the buf 466 * daemon if needed. 467 */ 468 static void 469 bdirtyadd(void) 470 { 471 472 /* 473 * Only do the wakeup once as we cross the boundary. The 474 * buf daemon will keep running until the condition clears. 475 */ 476 if (atomic_fetchadd_int(&numdirtybuffers, 1) == 477 (lodirtybuffers + hidirtybuffers) / 2) 478 bd_wakeup(); 479 } 480 481 /* 482 * bufspace_wakeup: 483 * 484 * Called when buffer space is potentially available for recovery. 485 * getnewbuf() will block on this flag when it is unable to free 486 * sufficient buffer space. Buffer space becomes recoverable when 487 * bp's get placed back in the queues. 488 */ 489 static void 490 bufspace_wakeup(void) 491 { 492 493 /* 494 * If someone is waiting for bufspace, wake them up. 495 * 496 * Since needsbuffer is set prior to doing an additional queue 497 * scan it is safe to check for the flag prior to acquiring the 498 * lock. The thread that is preparing to scan again before 499 * blocking would discover the buf we released. 500 */ 501 if (needsbuffer) { 502 rw_rlock(&nblock); 503 if (atomic_cmpset_int(&needsbuffer, 1, 0) == 1) 504 wakeup(__DEVOLATILE(void *, &needsbuffer)); 505 rw_runlock(&nblock); 506 } 507 } 508 509 /* 510 * bufspace_daemonwakeup: 511 * 512 * Wakeup the daemon responsible for freeing clean bufs. 513 */ 514 static void 515 bufspace_daemonwakeup(void) 516 { 517 rw_rlock(&nblock); 518 if (bufspace_request == 0) { 519 bufspace_request = 1; 520 wakeup(&bufspace_request); 521 } 522 rw_runlock(&nblock); 523 } 524 525 /* 526 * bufspace_adjust: 527 * 528 * Adjust the reported bufspace for a KVA managed buffer, possibly 529 * waking any waiters. 530 */ 531 static void 532 bufspace_adjust(struct buf *bp, int bufsize) 533 { 534 long space; 535 int diff; 536 537 KASSERT((bp->b_flags & B_MALLOC) == 0, 538 ("bufspace_adjust: malloc buf %p", bp)); 539 diff = bufsize - bp->b_bufsize; 540 if (diff < 0) { 541 atomic_subtract_long(&bufspace, -diff); 542 bufspace_wakeup(); 543 } else { 544 space = atomic_fetchadd_long(&bufspace, diff); 545 /* Wake up the daemon on the transition. */ 546 if (space < bufspacethresh && space + diff >= bufspacethresh) 547 bufspace_daemonwakeup(); 548 } 549 bp->b_bufsize = bufsize; 550 } 551 552 /* 553 * bufspace_reserve: 554 * 555 * Reserve bufspace before calling allocbuf(). metadata has a 556 * different space limit than data. 557 */ 558 static int 559 bufspace_reserve(int size, bool metadata) 560 { 561 long limit; 562 long space; 563 564 if (metadata) 565 limit = maxbufspace; 566 else 567 limit = hibufspace; 568 do { 569 space = bufspace; 570 if (space + size > limit) 571 return (ENOSPC); 572 } while (atomic_cmpset_long(&bufspace, space, space + size) == 0); 573 574 /* Wake up the daemon on the transition. */ 575 if (space < bufspacethresh && space + size >= bufspacethresh) 576 bufspace_daemonwakeup(); 577 578 return (0); 579 } 580 581 /* 582 * bufspace_release: 583 * 584 * Release reserved bufspace after bufspace_adjust() has consumed it. 585 */ 586 static void 587 bufspace_release(int size) 588 { 589 atomic_subtract_long(&bufspace, size); 590 bufspace_wakeup(); 591 } 592 593 /* 594 * bufspace_wait: 595 * 596 * Wait for bufspace, acting as the buf daemon if a locked vnode is 597 * supplied. needsbuffer must be set in a safe fashion prior to 598 * polling for space. The operation must be re-tried on return. 599 */ 600 static void 601 bufspace_wait(struct vnode *vp, int gbflags, int slpflag, int slptimeo) 602 { 603 struct thread *td; 604 int error, fl, norunbuf; 605 606 if ((gbflags & GB_NOWAIT_BD) != 0) 607 return; 608 609 td = curthread; 610 rw_wlock(&nblock); 611 while (needsbuffer != 0) { 612 if (vp != NULL && vp->v_type != VCHR && 613 (td->td_pflags & TDP_BUFNEED) == 0) { 614 rw_wunlock(&nblock); 615 /* 616 * getblk() is called with a vnode locked, and 617 * some majority of the dirty buffers may as 618 * well belong to the vnode. Flushing the 619 * buffers there would make a progress that 620 * cannot be achieved by the buf_daemon, that 621 * cannot lock the vnode. 622 */ 623 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 624 (td->td_pflags & TDP_NORUNNINGBUF); 625 626 /* 627 * Play bufdaemon. The getnewbuf() function 628 * may be called while the thread owns lock 629 * for another dirty buffer for the same 630 * vnode, which makes it impossible to use 631 * VOP_FSYNC() there, due to the buffer lock 632 * recursion. 633 */ 634 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 635 fl = buf_flush(vp, flushbufqtarget); 636 td->td_pflags &= norunbuf; 637 rw_wlock(&nblock); 638 if (fl != 0) 639 continue; 640 if (needsbuffer == 0) 641 break; 642 } 643 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 644 (PRIBIO + 4) | slpflag, "newbuf", slptimeo); 645 if (error != 0) 646 break; 647 } 648 rw_wunlock(&nblock); 649 } 650 651 652 /* 653 * bufspace_daemon: 654 * 655 * buffer space management daemon. Tries to maintain some marginal 656 * amount of free buffer space so that requesting processes neither 657 * block nor work to reclaim buffers. 658 */ 659 static void 660 bufspace_daemon(void) 661 { 662 for (;;) { 663 kproc_suspend_check(bufspacedaemonproc); 664 665 /* 666 * Free buffers from the clean queue until we meet our 667 * targets. 668 * 669 * Theory of operation: The buffer cache is most efficient 670 * when some free buffer headers and space are always 671 * available to getnewbuf(). This daemon attempts to prevent 672 * the excessive blocking and synchronization associated 673 * with shortfall. It goes through three phases according 674 * demand: 675 * 676 * 1) The daemon wakes up voluntarily once per-second 677 * during idle periods when the counters are below 678 * the wakeup thresholds (bufspacethresh, lofreebuffers). 679 * 680 * 2) The daemon wakes up as we cross the thresholds 681 * ahead of any potential blocking. This may bounce 682 * slightly according to the rate of consumption and 683 * release. 684 * 685 * 3) The daemon and consumers are starved for working 686 * clean buffers. This is the 'bufspace' sleep below 687 * which will inefficiently trade bufs with bqrelse 688 * until we return to condition 2. 689 */ 690 while (bufspace > lobufspace || 691 numfreebuffers < hifreebuffers) { 692 if (buf_recycle(false) != 0) { 693 atomic_set_int(&needsbuffer, 1); 694 if (buf_recycle(false) != 0) { 695 rw_wlock(&nblock); 696 if (needsbuffer) 697 rw_sleep(__DEVOLATILE(void *, 698 &needsbuffer), &nblock, 699 PRIBIO|PDROP, "bufspace", 700 hz/10); 701 else 702 rw_wunlock(&nblock); 703 } 704 } 705 maybe_yield(); 706 } 707 708 /* 709 * Re-check our limits under the exclusive nblock. 710 */ 711 rw_wlock(&nblock); 712 if (bufspace < bufspacethresh && 713 numfreebuffers > lofreebuffers) { 714 bufspace_request = 0; 715 rw_sleep(&bufspace_request, &nblock, PRIBIO|PDROP, 716 "-", hz); 717 } else 718 rw_wunlock(&nblock); 719 } 720 } 721 722 static struct kproc_desc bufspace_kp = { 723 "bufspacedaemon", 724 bufspace_daemon, 725 &bufspacedaemonproc 726 }; 727 SYSINIT(bufspacedaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, 728 &bufspace_kp); 729 730 /* 731 * bufmallocadjust: 732 * 733 * Adjust the reported bufspace for a malloc managed buffer, possibly 734 * waking any waiters. 735 */ 736 static void 737 bufmallocadjust(struct buf *bp, int bufsize) 738 { 739 int diff; 740 741 KASSERT((bp->b_flags & B_MALLOC) != 0, 742 ("bufmallocadjust: non-malloc buf %p", bp)); 743 diff = bufsize - bp->b_bufsize; 744 if (diff < 0) 745 atomic_subtract_long(&bufmallocspace, -diff); 746 else 747 atomic_add_long(&bufmallocspace, diff); 748 bp->b_bufsize = bufsize; 749 } 750 751 /* 752 * runningwakeup: 753 * 754 * Wake up processes that are waiting on asynchronous writes to fall 755 * below lorunningspace. 756 */ 757 static void 758 runningwakeup(void) 759 { 760 761 mtx_lock(&rbreqlock); 762 if (runningbufreq) { 763 runningbufreq = 0; 764 wakeup(&runningbufreq); 765 } 766 mtx_unlock(&rbreqlock); 767 } 768 769 /* 770 * runningbufwakeup: 771 * 772 * Decrement the outstanding write count according. 773 */ 774 void 775 runningbufwakeup(struct buf *bp) 776 { 777 long space, bspace; 778 779 bspace = bp->b_runningbufspace; 780 if (bspace == 0) 781 return; 782 space = atomic_fetchadd_long(&runningbufspace, -bspace); 783 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 784 space, bspace)); 785 bp->b_runningbufspace = 0; 786 /* 787 * Only acquire the lock and wakeup on the transition from exceeding 788 * the threshold to falling below it. 789 */ 790 if (space < lorunningspace) 791 return; 792 if (space - bspace > lorunningspace) 793 return; 794 runningwakeup(); 795 } 796 797 /* 798 * waitrunningbufspace() 799 * 800 * runningbufspace is a measure of the amount of I/O currently 801 * running. This routine is used in async-write situations to 802 * prevent creating huge backups of pending writes to a device. 803 * Only asynchronous writes are governed by this function. 804 * 805 * This does NOT turn an async write into a sync write. It waits 806 * for earlier writes to complete and generally returns before the 807 * caller's write has reached the device. 808 */ 809 void 810 waitrunningbufspace(void) 811 { 812 813 mtx_lock(&rbreqlock); 814 while (runningbufspace > hirunningspace) { 815 runningbufreq = 1; 816 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 817 } 818 mtx_unlock(&rbreqlock); 819 } 820 821 822 /* 823 * vfs_buf_test_cache: 824 * 825 * Called when a buffer is extended. This function clears the B_CACHE 826 * bit if the newly extended portion of the buffer does not contain 827 * valid data. 828 */ 829 static __inline void 830 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 831 vm_offset_t size, vm_page_t m) 832 { 833 834 VM_OBJECT_ASSERT_LOCKED(m->object); 835 if (bp->b_flags & B_CACHE) { 836 int base = (foff + off) & PAGE_MASK; 837 if (vm_page_is_valid(m, base, size) == 0) 838 bp->b_flags &= ~B_CACHE; 839 } 840 } 841 842 /* Wake up the buffer daemon if necessary */ 843 static __inline void 844 bd_wakeup(void) 845 { 846 847 mtx_lock(&bdlock); 848 if (bd_request == 0) { 849 bd_request = 1; 850 wakeup(&bd_request); 851 } 852 mtx_unlock(&bdlock); 853 } 854 855 /* 856 * Adjust the maxbcachbuf tunable. 857 */ 858 static void 859 maxbcachebuf_adjust(void) 860 { 861 int i; 862 863 /* 864 * maxbcachebuf must be a power of 2 >= MAXBSIZE. 865 */ 866 i = 2; 867 while (i * 2 <= maxbcachebuf) 868 i *= 2; 869 maxbcachebuf = i; 870 if (maxbcachebuf < MAXBSIZE) 871 maxbcachebuf = MAXBSIZE; 872 if (maxbcachebuf > MAXPHYS) 873 maxbcachebuf = MAXPHYS; 874 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) 875 printf("maxbcachebuf=%d\n", maxbcachebuf); 876 } 877 878 /* 879 * bd_speedup - speedup the buffer cache flushing code 880 */ 881 void 882 bd_speedup(void) 883 { 884 int needwake; 885 886 mtx_lock(&bdlock); 887 needwake = 0; 888 if (bd_speedupreq == 0 || bd_request == 0) 889 needwake = 1; 890 bd_speedupreq = 1; 891 bd_request = 1; 892 if (needwake) 893 wakeup(&bd_request); 894 mtx_unlock(&bdlock); 895 } 896 897 #ifndef NSWBUF_MIN 898 #define NSWBUF_MIN 16 899 #endif 900 901 #ifdef __i386__ 902 #define TRANSIENT_DENOM 5 903 #else 904 #define TRANSIENT_DENOM 10 905 #endif 906 907 /* 908 * Calculating buffer cache scaling values and reserve space for buffer 909 * headers. This is called during low level kernel initialization and 910 * may be called more then once. We CANNOT write to the memory area 911 * being reserved at this time. 912 */ 913 caddr_t 914 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 915 { 916 int tuned_nbuf; 917 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 918 919 /* 920 * physmem_est is in pages. Convert it to kilobytes (assumes 921 * PAGE_SIZE is >= 1K) 922 */ 923 physmem_est = physmem_est * (PAGE_SIZE / 1024); 924 925 maxbcachebuf_adjust(); 926 /* 927 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 928 * For the first 64MB of ram nominally allocate sufficient buffers to 929 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 930 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 931 * the buffer cache we limit the eventual kva reservation to 932 * maxbcache bytes. 933 * 934 * factor represents the 1/4 x ram conversion. 935 */ 936 if (nbuf == 0) { 937 int factor = 4 * BKVASIZE / 1024; 938 939 nbuf = 50; 940 if (physmem_est > 4096) 941 nbuf += min((physmem_est - 4096) / factor, 942 65536 / factor); 943 if (physmem_est > 65536) 944 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 945 32 * 1024 * 1024 / (factor * 5)); 946 947 if (maxbcache && nbuf > maxbcache / BKVASIZE) 948 nbuf = maxbcache / BKVASIZE; 949 tuned_nbuf = 1; 950 } else 951 tuned_nbuf = 0; 952 953 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 954 maxbuf = (LONG_MAX / 3) / BKVASIZE; 955 if (nbuf > maxbuf) { 956 if (!tuned_nbuf) 957 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 958 maxbuf); 959 nbuf = maxbuf; 960 } 961 962 /* 963 * Ideal allocation size for the transient bio submap is 10% 964 * of the maximal space buffer map. This roughly corresponds 965 * to the amount of the buffer mapped for typical UFS load. 966 * 967 * Clip the buffer map to reserve space for the transient 968 * BIOs, if its extent is bigger than 90% (80% on i386) of the 969 * maximum buffer map extent on the platform. 970 * 971 * The fall-back to the maxbuf in case of maxbcache unset, 972 * allows to not trim the buffer KVA for the architectures 973 * with ample KVA space. 974 */ 975 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 976 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 977 buf_sz = (long)nbuf * BKVASIZE; 978 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 979 (TRANSIENT_DENOM - 1)) { 980 /* 981 * There is more KVA than memory. Do not 982 * adjust buffer map size, and assign the rest 983 * of maxbuf to transient map. 984 */ 985 biotmap_sz = maxbuf_sz - buf_sz; 986 } else { 987 /* 988 * Buffer map spans all KVA we could afford on 989 * this platform. Give 10% (20% on i386) of 990 * the buffer map to the transient bio map. 991 */ 992 biotmap_sz = buf_sz / TRANSIENT_DENOM; 993 buf_sz -= biotmap_sz; 994 } 995 if (biotmap_sz / INT_MAX > MAXPHYS) 996 bio_transient_maxcnt = INT_MAX; 997 else 998 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 999 /* 1000 * Artificially limit to 1024 simultaneous in-flight I/Os 1001 * using the transient mapping. 1002 */ 1003 if (bio_transient_maxcnt > 1024) 1004 bio_transient_maxcnt = 1024; 1005 if (tuned_nbuf) 1006 nbuf = buf_sz / BKVASIZE; 1007 } 1008 1009 /* 1010 * swbufs are used as temporary holders for I/O, such as paging I/O. 1011 * We have no less then 16 and no more then 256. 1012 */ 1013 nswbuf = min(nbuf / 4, 256); 1014 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf); 1015 if (nswbuf < NSWBUF_MIN) 1016 nswbuf = NSWBUF_MIN; 1017 1018 /* 1019 * Reserve space for the buffer cache buffers 1020 */ 1021 swbuf = (void *)v; 1022 v = (caddr_t)(swbuf + nswbuf); 1023 buf = (void *)v; 1024 v = (caddr_t)(buf + nbuf); 1025 1026 return(v); 1027 } 1028 1029 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1030 void 1031 bufinit(void) 1032 { 1033 struct buf *bp; 1034 int i; 1035 1036 KASSERT(maxbcachebuf >= MAXBSIZE, 1037 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1038 MAXBSIZE)); 1039 mtx_init(&bqlocks[QUEUE_DIRTY], "bufq dirty lock", NULL, MTX_DEF); 1040 mtx_init(&bqlocks[QUEUE_EMPTY], "bufq empty lock", NULL, MTX_DEF); 1041 for (i = QUEUE_CLEAN; i < QUEUE_CLEAN + CLEAN_QUEUES; i++) 1042 mtx_init(&bqlocks[i], "bufq clean lock", NULL, MTX_DEF); 1043 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1044 rw_init(&nblock, "needsbuffer lock"); 1045 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1046 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1047 1048 /* next, make a null set of free lists */ 1049 for (i = 0; i < BUFFER_QUEUES; i++) 1050 TAILQ_INIT(&bufqueues[i]); 1051 1052 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 1053 1054 /* finally, initialize each buffer header and stick on empty q */ 1055 for (i = 0; i < nbuf; i++) { 1056 bp = &buf[i]; 1057 bzero(bp, sizeof *bp); 1058 bp->b_flags = B_INVAL; 1059 bp->b_rcred = NOCRED; 1060 bp->b_wcred = NOCRED; 1061 bp->b_qindex = QUEUE_EMPTY; 1062 bp->b_xflags = 0; 1063 bp->b_data = bp->b_kvabase = unmapped_buf; 1064 LIST_INIT(&bp->b_dep); 1065 BUF_LOCKINIT(bp); 1066 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 1067 #ifdef INVARIANTS 1068 bq_len[QUEUE_EMPTY]++; 1069 #endif 1070 } 1071 1072 /* 1073 * maxbufspace is the absolute maximum amount of buffer space we are 1074 * allowed to reserve in KVM and in real terms. The absolute maximum 1075 * is nominally used by metadata. hibufspace is the nominal maximum 1076 * used by most other requests. The differential is required to 1077 * ensure that metadata deadlocks don't occur. 1078 * 1079 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1080 * this may result in KVM fragmentation which is not handled optimally 1081 * by the system. XXX This is less true with vmem. We could use 1082 * PAGE_SIZE. 1083 */ 1084 maxbufspace = (long)nbuf * BKVASIZE; 1085 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1086 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1087 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1088 1089 /* 1090 * Note: The 16 MiB upper limit for hirunningspace was chosen 1091 * arbitrarily and may need further tuning. It corresponds to 1092 * 128 outstanding write IO requests (if IO size is 128 KiB), 1093 * which fits with many RAID controllers' tagged queuing limits. 1094 * The lower 1 MiB limit is the historical upper limit for 1095 * hirunningspace. 1096 */ 1097 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1098 16 * 1024 * 1024), 1024 * 1024); 1099 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1100 1101 /* 1102 * Limit the amount of malloc memory since it is wired permanently into 1103 * the kernel space. Even though this is accounted for in the buffer 1104 * allocation, we don't want the malloced region to grow uncontrolled. 1105 * The malloc scheme improves memory utilization significantly on 1106 * average (small) directories. 1107 */ 1108 maxbufmallocspace = hibufspace / 20; 1109 1110 /* 1111 * Reduce the chance of a deadlock occurring by limiting the number 1112 * of delayed-write dirty buffers we allow to stack up. 1113 */ 1114 hidirtybuffers = nbuf / 4 + 20; 1115 dirtybufthresh = hidirtybuffers * 9 / 10; 1116 numdirtybuffers = 0; 1117 /* 1118 * To support extreme low-memory systems, make sure hidirtybuffers 1119 * cannot eat up all available buffer space. This occurs when our 1120 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1121 * buffer space assuming BKVASIZE'd buffers. 1122 */ 1123 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1124 hidirtybuffers >>= 1; 1125 } 1126 lodirtybuffers = hidirtybuffers / 2; 1127 1128 /* 1129 * lofreebuffers should be sufficient to avoid stalling waiting on 1130 * buf headers under heavy utilization. The bufs in per-cpu caches 1131 * are counted as free but will be unavailable to threads executing 1132 * on other cpus. 1133 * 1134 * hifreebuffers is the free target for the bufspace daemon. This 1135 * should be set appropriately to limit work per-iteration. 1136 */ 1137 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1138 hifreebuffers = (3 * lofreebuffers) / 2; 1139 numfreebuffers = nbuf; 1140 1141 /* Setup the kva and free list allocators. */ 1142 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1143 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), 1144 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1145 1146 /* 1147 * Size the clean queue according to the amount of buffer space. 1148 * One queue per-256mb up to the max. More queues gives better 1149 * concurrency but less accurate LRU. 1150 */ 1151 clean_queues = MIN(howmany(maxbufspace, 256*1024*1024), CLEAN_QUEUES); 1152 1153 } 1154 1155 #ifdef INVARIANTS 1156 static inline void 1157 vfs_buf_check_mapped(struct buf *bp) 1158 { 1159 1160 KASSERT(bp->b_kvabase != unmapped_buf, 1161 ("mapped buf: b_kvabase was not updated %p", bp)); 1162 KASSERT(bp->b_data != unmapped_buf, 1163 ("mapped buf: b_data was not updated %p", bp)); 1164 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1165 MAXPHYS, ("b_data + b_offset unmapped %p", bp)); 1166 } 1167 1168 static inline void 1169 vfs_buf_check_unmapped(struct buf *bp) 1170 { 1171 1172 KASSERT(bp->b_data == unmapped_buf, 1173 ("unmapped buf: corrupted b_data %p", bp)); 1174 } 1175 1176 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1177 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1178 #else 1179 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1180 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1181 #endif 1182 1183 static int 1184 isbufbusy(struct buf *bp) 1185 { 1186 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1187 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1188 return (1); 1189 return (0); 1190 } 1191 1192 /* 1193 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1194 */ 1195 void 1196 bufshutdown(int show_busybufs) 1197 { 1198 static int first_buf_printf = 1; 1199 struct buf *bp; 1200 int iter, nbusy, pbusy; 1201 #ifndef PREEMPTION 1202 int subiter; 1203 #endif 1204 1205 /* 1206 * Sync filesystems for shutdown 1207 */ 1208 wdog_kern_pat(WD_LASTVAL); 1209 sys_sync(curthread, NULL); 1210 1211 /* 1212 * With soft updates, some buffers that are 1213 * written will be remarked as dirty until other 1214 * buffers are written. 1215 */ 1216 for (iter = pbusy = 0; iter < 20; iter++) { 1217 nbusy = 0; 1218 for (bp = &buf[nbuf]; --bp >= buf; ) 1219 if (isbufbusy(bp)) 1220 nbusy++; 1221 if (nbusy == 0) { 1222 if (first_buf_printf) 1223 printf("All buffers synced."); 1224 break; 1225 } 1226 if (first_buf_printf) { 1227 printf("Syncing disks, buffers remaining... "); 1228 first_buf_printf = 0; 1229 } 1230 printf("%d ", nbusy); 1231 if (nbusy < pbusy) 1232 iter = 0; 1233 pbusy = nbusy; 1234 1235 wdog_kern_pat(WD_LASTVAL); 1236 sys_sync(curthread, NULL); 1237 1238 #ifdef PREEMPTION 1239 /* 1240 * Drop Giant and spin for a while to allow 1241 * interrupt threads to run. 1242 */ 1243 DROP_GIANT(); 1244 DELAY(50000 * iter); 1245 PICKUP_GIANT(); 1246 #else 1247 /* 1248 * Drop Giant and context switch several times to 1249 * allow interrupt threads to run. 1250 */ 1251 DROP_GIANT(); 1252 for (subiter = 0; subiter < 50 * iter; subiter++) { 1253 thread_lock(curthread); 1254 mi_switch(SW_VOL, NULL); 1255 thread_unlock(curthread); 1256 DELAY(1000); 1257 } 1258 PICKUP_GIANT(); 1259 #endif 1260 } 1261 printf("\n"); 1262 /* 1263 * Count only busy local buffers to prevent forcing 1264 * a fsck if we're just a client of a wedged NFS server 1265 */ 1266 nbusy = 0; 1267 for (bp = &buf[nbuf]; --bp >= buf; ) { 1268 if (isbufbusy(bp)) { 1269 #if 0 1270 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1271 if (bp->b_dev == NULL) { 1272 TAILQ_REMOVE(&mountlist, 1273 bp->b_vp->v_mount, mnt_list); 1274 continue; 1275 } 1276 #endif 1277 nbusy++; 1278 if (show_busybufs > 0) { 1279 printf( 1280 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1281 nbusy, bp, bp->b_vp, bp->b_flags, 1282 (intmax_t)bp->b_blkno, 1283 (intmax_t)bp->b_lblkno); 1284 BUF_LOCKPRINTINFO(bp); 1285 if (show_busybufs > 1) 1286 vn_printf(bp->b_vp, 1287 "vnode content: "); 1288 } 1289 } 1290 } 1291 if (nbusy) { 1292 /* 1293 * Failed to sync all blocks. Indicate this and don't 1294 * unmount filesystems (thus forcing an fsck on reboot). 1295 */ 1296 printf("Giving up on %d buffers\n", nbusy); 1297 DELAY(5000000); /* 5 seconds */ 1298 } else { 1299 if (!first_buf_printf) 1300 printf("Final sync complete\n"); 1301 /* 1302 * Unmount filesystems 1303 */ 1304 if (panicstr == NULL) 1305 vfs_unmountall(); 1306 } 1307 swapoff_all(); 1308 DELAY(100000); /* wait for console output to finish */ 1309 } 1310 1311 static void 1312 bpmap_qenter(struct buf *bp) 1313 { 1314 1315 BUF_CHECK_MAPPED(bp); 1316 1317 /* 1318 * bp->b_data is relative to bp->b_offset, but 1319 * bp->b_offset may be offset into the first page. 1320 */ 1321 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1322 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1323 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1324 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1325 } 1326 1327 /* 1328 * binsfree: 1329 * 1330 * Insert the buffer into the appropriate free list. 1331 */ 1332 static void 1333 binsfree(struct buf *bp, int qindex) 1334 { 1335 struct mtx *olock, *nlock; 1336 1337 if (qindex != QUEUE_EMPTY) { 1338 BUF_ASSERT_XLOCKED(bp); 1339 } 1340 1341 /* 1342 * Stick to the same clean queue for the lifetime of the buf to 1343 * limit locking below. Otherwise pick ont sequentially. 1344 */ 1345 if (qindex == QUEUE_CLEAN) { 1346 if (bqisclean(bp->b_qindex)) 1347 qindex = bp->b_qindex; 1348 else 1349 qindex = bqcleanq(); 1350 } 1351 1352 /* 1353 * Handle delayed bremfree() processing. 1354 */ 1355 nlock = bqlock(qindex); 1356 if (bp->b_flags & B_REMFREE) { 1357 olock = bqlock(bp->b_qindex); 1358 mtx_lock(olock); 1359 bremfreel(bp); 1360 if (olock != nlock) { 1361 mtx_unlock(olock); 1362 mtx_lock(nlock); 1363 } 1364 } else 1365 mtx_lock(nlock); 1366 1367 if (bp->b_qindex != QUEUE_NONE) 1368 panic("binsfree: free buffer onto another queue???"); 1369 1370 bp->b_qindex = qindex; 1371 if (bp->b_flags & B_AGE) 1372 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1373 else 1374 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1375 #ifdef INVARIANTS 1376 bq_len[bp->b_qindex]++; 1377 #endif 1378 mtx_unlock(nlock); 1379 } 1380 1381 /* 1382 * buf_free: 1383 * 1384 * Free a buffer to the buf zone once it no longer has valid contents. 1385 */ 1386 static void 1387 buf_free(struct buf *bp) 1388 { 1389 1390 if (bp->b_flags & B_REMFREE) 1391 bremfreef(bp); 1392 if (bp->b_vflags & BV_BKGRDINPROG) 1393 panic("losing buffer 1"); 1394 if (bp->b_rcred != NOCRED) { 1395 crfree(bp->b_rcred); 1396 bp->b_rcred = NOCRED; 1397 } 1398 if (bp->b_wcred != NOCRED) { 1399 crfree(bp->b_wcred); 1400 bp->b_wcred = NOCRED; 1401 } 1402 if (!LIST_EMPTY(&bp->b_dep)) 1403 buf_deallocate(bp); 1404 bufkva_free(bp); 1405 BUF_UNLOCK(bp); 1406 uma_zfree(buf_zone, bp); 1407 atomic_add_int(&numfreebuffers, 1); 1408 bufspace_wakeup(); 1409 } 1410 1411 /* 1412 * buf_import: 1413 * 1414 * Import bufs into the uma cache from the buf list. The system still 1415 * expects a static array of bufs and much of the synchronization 1416 * around bufs assumes type stable storage. As a result, UMA is used 1417 * only as a per-cpu cache of bufs still maintained on a global list. 1418 */ 1419 static int 1420 buf_import(void *arg, void **store, int cnt, int flags) 1421 { 1422 struct buf *bp; 1423 int i; 1424 1425 mtx_lock(&bqlocks[QUEUE_EMPTY]); 1426 for (i = 0; i < cnt; i++) { 1427 bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1428 if (bp == NULL) 1429 break; 1430 bremfreel(bp); 1431 store[i] = bp; 1432 } 1433 mtx_unlock(&bqlocks[QUEUE_EMPTY]); 1434 1435 return (i); 1436 } 1437 1438 /* 1439 * buf_release: 1440 * 1441 * Release bufs from the uma cache back to the buffer queues. 1442 */ 1443 static void 1444 buf_release(void *arg, void **store, int cnt) 1445 { 1446 int i; 1447 1448 for (i = 0; i < cnt; i++) 1449 binsfree(store[i], QUEUE_EMPTY); 1450 } 1451 1452 /* 1453 * buf_alloc: 1454 * 1455 * Allocate an empty buffer header. 1456 */ 1457 static struct buf * 1458 buf_alloc(void) 1459 { 1460 struct buf *bp; 1461 1462 bp = uma_zalloc(buf_zone, M_NOWAIT); 1463 if (bp == NULL) { 1464 bufspace_daemonwakeup(); 1465 atomic_add_int(&numbufallocfails, 1); 1466 return (NULL); 1467 } 1468 1469 /* 1470 * Wake-up the bufspace daemon on transition. 1471 */ 1472 if (atomic_fetchadd_int(&numfreebuffers, -1) == lofreebuffers) 1473 bufspace_daemonwakeup(); 1474 1475 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1476 panic("getnewbuf_empty: Locked buf %p on free queue.", bp); 1477 1478 KASSERT(bp->b_vp == NULL, 1479 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1480 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1481 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1482 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1483 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1484 KASSERT(bp->b_npages == 0, 1485 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1486 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1487 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1488 1489 bp->b_flags = 0; 1490 bp->b_ioflags = 0; 1491 bp->b_xflags = 0; 1492 bp->b_vflags = 0; 1493 bp->b_vp = NULL; 1494 bp->b_blkno = bp->b_lblkno = 0; 1495 bp->b_offset = NOOFFSET; 1496 bp->b_iodone = 0; 1497 bp->b_error = 0; 1498 bp->b_resid = 0; 1499 bp->b_bcount = 0; 1500 bp->b_npages = 0; 1501 bp->b_dirtyoff = bp->b_dirtyend = 0; 1502 bp->b_bufobj = NULL; 1503 bp->b_data = bp->b_kvabase = unmapped_buf; 1504 bp->b_fsprivate1 = NULL; 1505 bp->b_fsprivate2 = NULL; 1506 bp->b_fsprivate3 = NULL; 1507 LIST_INIT(&bp->b_dep); 1508 1509 return (bp); 1510 } 1511 1512 /* 1513 * buf_qrecycle: 1514 * 1515 * Free a buffer from the given bufqueue. kva controls whether the 1516 * freed buf must own some kva resources. This is used for 1517 * defragmenting. 1518 */ 1519 static int 1520 buf_qrecycle(int qindex, bool kva) 1521 { 1522 struct buf *bp, *nbp; 1523 1524 if (kva) 1525 atomic_add_int(&bufdefragcnt, 1); 1526 nbp = NULL; 1527 mtx_lock(&bqlocks[qindex]); 1528 nbp = TAILQ_FIRST(&bufqueues[qindex]); 1529 1530 /* 1531 * Run scan, possibly freeing data and/or kva mappings on the fly 1532 * depending. 1533 */ 1534 while ((bp = nbp) != NULL) { 1535 /* 1536 * Calculate next bp (we can only use it if we do not 1537 * release the bqlock). 1538 */ 1539 nbp = TAILQ_NEXT(bp, b_freelist); 1540 1541 /* 1542 * If we are defragging then we need a buffer with 1543 * some kva to reclaim. 1544 */ 1545 if (kva && bp->b_kvasize == 0) 1546 continue; 1547 1548 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1549 continue; 1550 1551 /* 1552 * Skip buffers with background writes in progress. 1553 */ 1554 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1555 BUF_UNLOCK(bp); 1556 continue; 1557 } 1558 1559 KASSERT(bp->b_qindex == qindex, 1560 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 1561 /* 1562 * NOTE: nbp is now entirely invalid. We can only restart 1563 * the scan from this point on. 1564 */ 1565 bremfreel(bp); 1566 mtx_unlock(&bqlocks[qindex]); 1567 1568 /* 1569 * Requeue the background write buffer with error and 1570 * restart the scan. 1571 */ 1572 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1573 bqrelse(bp); 1574 mtx_lock(&bqlocks[qindex]); 1575 nbp = TAILQ_FIRST(&bufqueues[qindex]); 1576 continue; 1577 } 1578 bp->b_flags |= B_INVAL; 1579 brelse(bp); 1580 return (0); 1581 } 1582 mtx_unlock(&bqlocks[qindex]); 1583 1584 return (ENOBUFS); 1585 } 1586 1587 /* 1588 * buf_recycle: 1589 * 1590 * Iterate through all clean queues until we find a buf to recycle or 1591 * exhaust the search. 1592 */ 1593 static int 1594 buf_recycle(bool kva) 1595 { 1596 int qindex, first_qindex; 1597 1598 qindex = first_qindex = bqcleanq(); 1599 do { 1600 if (buf_qrecycle(qindex, kva) == 0) 1601 return (0); 1602 if (++qindex == QUEUE_CLEAN + clean_queues) 1603 qindex = QUEUE_CLEAN; 1604 } while (qindex != first_qindex); 1605 1606 return (ENOBUFS); 1607 } 1608 1609 /* 1610 * buf_scan: 1611 * 1612 * Scan the clean queues looking for a buffer to recycle. needsbuffer 1613 * is set on failure so that the caller may optionally bufspace_wait() 1614 * in a race-free fashion. 1615 */ 1616 static int 1617 buf_scan(bool defrag) 1618 { 1619 int error; 1620 1621 /* 1622 * To avoid heavy synchronization and wakeup races we set 1623 * needsbuffer and re-poll before failing. This ensures that 1624 * no frees can be missed between an unsuccessful poll and 1625 * going to sleep in a synchronized fashion. 1626 */ 1627 if ((error = buf_recycle(defrag)) != 0) { 1628 atomic_set_int(&needsbuffer, 1); 1629 bufspace_daemonwakeup(); 1630 error = buf_recycle(defrag); 1631 } 1632 if (error == 0) 1633 atomic_add_int(&getnewbufrestarts, 1); 1634 return (error); 1635 } 1636 1637 /* 1638 * bremfree: 1639 * 1640 * Mark the buffer for removal from the appropriate free list. 1641 * 1642 */ 1643 void 1644 bremfree(struct buf *bp) 1645 { 1646 1647 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1648 KASSERT((bp->b_flags & B_REMFREE) == 0, 1649 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1650 KASSERT(bp->b_qindex != QUEUE_NONE, 1651 ("bremfree: buffer %p not on a queue.", bp)); 1652 BUF_ASSERT_XLOCKED(bp); 1653 1654 bp->b_flags |= B_REMFREE; 1655 } 1656 1657 /* 1658 * bremfreef: 1659 * 1660 * Force an immediate removal from a free list. Used only in nfs when 1661 * it abuses the b_freelist pointer. 1662 */ 1663 void 1664 bremfreef(struct buf *bp) 1665 { 1666 struct mtx *qlock; 1667 1668 qlock = bqlock(bp->b_qindex); 1669 mtx_lock(qlock); 1670 bremfreel(bp); 1671 mtx_unlock(qlock); 1672 } 1673 1674 /* 1675 * bremfreel: 1676 * 1677 * Removes a buffer from the free list, must be called with the 1678 * correct qlock held. 1679 */ 1680 static void 1681 bremfreel(struct buf *bp) 1682 { 1683 1684 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1685 bp, bp->b_vp, bp->b_flags); 1686 KASSERT(bp->b_qindex != QUEUE_NONE, 1687 ("bremfreel: buffer %p not on a queue.", bp)); 1688 if (bp->b_qindex != QUEUE_EMPTY) { 1689 BUF_ASSERT_XLOCKED(bp); 1690 } 1691 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1692 1693 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1694 #ifdef INVARIANTS 1695 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1696 bp->b_qindex)); 1697 bq_len[bp->b_qindex]--; 1698 #endif 1699 bp->b_qindex = QUEUE_NONE; 1700 bp->b_flags &= ~B_REMFREE; 1701 } 1702 1703 /* 1704 * bufkva_free: 1705 * 1706 * Free the kva allocation for a buffer. 1707 * 1708 */ 1709 static void 1710 bufkva_free(struct buf *bp) 1711 { 1712 1713 #ifdef INVARIANTS 1714 if (bp->b_kvasize == 0) { 1715 KASSERT(bp->b_kvabase == unmapped_buf && 1716 bp->b_data == unmapped_buf, 1717 ("Leaked KVA space on %p", bp)); 1718 } else if (buf_mapped(bp)) 1719 BUF_CHECK_MAPPED(bp); 1720 else 1721 BUF_CHECK_UNMAPPED(bp); 1722 #endif 1723 if (bp->b_kvasize == 0) 1724 return; 1725 1726 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 1727 atomic_subtract_long(&bufkvaspace, bp->b_kvasize); 1728 atomic_add_int(&buffreekvacnt, 1); 1729 bp->b_data = bp->b_kvabase = unmapped_buf; 1730 bp->b_kvasize = 0; 1731 } 1732 1733 /* 1734 * bufkva_alloc: 1735 * 1736 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 1737 */ 1738 static int 1739 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 1740 { 1741 vm_offset_t addr; 1742 int error; 1743 1744 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 1745 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 1746 1747 bufkva_free(bp); 1748 1749 addr = 0; 1750 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 1751 if (error != 0) { 1752 /* 1753 * Buffer map is too fragmented. Request the caller 1754 * to defragment the map. 1755 */ 1756 return (error); 1757 } 1758 bp->b_kvabase = (caddr_t)addr; 1759 bp->b_kvasize = maxsize; 1760 atomic_add_long(&bufkvaspace, bp->b_kvasize); 1761 if ((gbflags & GB_UNMAPPED) != 0) { 1762 bp->b_data = unmapped_buf; 1763 BUF_CHECK_UNMAPPED(bp); 1764 } else { 1765 bp->b_data = bp->b_kvabase; 1766 BUF_CHECK_MAPPED(bp); 1767 } 1768 return (0); 1769 } 1770 1771 /* 1772 * bufkva_reclaim: 1773 * 1774 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 1775 * callback that fires to avoid returning failure. 1776 */ 1777 static void 1778 bufkva_reclaim(vmem_t *vmem, int flags) 1779 { 1780 int i; 1781 1782 for (i = 0; i < 5; i++) 1783 if (buf_scan(true) != 0) 1784 break; 1785 return; 1786 } 1787 1788 /* 1789 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1790 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1791 * the buffer is valid and we do not have to do anything. 1792 */ 1793 static void 1794 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 1795 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 1796 { 1797 struct buf *rabp; 1798 int i; 1799 1800 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1801 if (inmem(vp, *rablkno)) 1802 continue; 1803 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1804 if ((rabp->b_flags & B_CACHE) != 0) { 1805 brelse(rabp); 1806 continue; 1807 } 1808 if (!TD_IS_IDLETHREAD(curthread)) { 1809 #ifdef RACCT 1810 if (racct_enable) { 1811 PROC_LOCK(curproc); 1812 racct_add_buf(curproc, rabp, 0); 1813 PROC_UNLOCK(curproc); 1814 } 1815 #endif /* RACCT */ 1816 curthread->td_ru.ru_inblock++; 1817 } 1818 rabp->b_flags |= B_ASYNC; 1819 rabp->b_flags &= ~B_INVAL; 1820 if ((flags & GB_CKHASH) != 0) { 1821 rabp->b_flags |= B_CKHASH; 1822 rabp->b_ckhashcalc = ckhashfunc; 1823 } 1824 rabp->b_ioflags &= ~BIO_ERROR; 1825 rabp->b_iocmd = BIO_READ; 1826 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1827 rabp->b_rcred = crhold(cred); 1828 vfs_busy_pages(rabp, 0); 1829 BUF_KERNPROC(rabp); 1830 rabp->b_iooffset = dbtob(rabp->b_blkno); 1831 bstrategy(rabp); 1832 } 1833 } 1834 1835 /* 1836 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1837 * 1838 * Get a buffer with the specified data. Look in the cache first. We 1839 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1840 * is set, the buffer is valid and we do not have to do anything, see 1841 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1842 * 1843 * Always return a NULL buffer pointer (in bpp) when returning an error. 1844 */ 1845 int 1846 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1847 int *rabsize, int cnt, struct ucred *cred, int flags, 1848 void (*ckhashfunc)(struct buf *), struct buf **bpp) 1849 { 1850 struct buf *bp; 1851 int readwait, rv; 1852 1853 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1854 /* 1855 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1856 */ 1857 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1858 if (bp == NULL) 1859 return (EBUSY); 1860 1861 /* 1862 * If not found in cache, do some I/O 1863 */ 1864 readwait = 0; 1865 if ((bp->b_flags & B_CACHE) == 0) { 1866 if (!TD_IS_IDLETHREAD(curthread)) { 1867 #ifdef RACCT 1868 if (racct_enable) { 1869 PROC_LOCK(curproc); 1870 racct_add_buf(curproc, bp, 0); 1871 PROC_UNLOCK(curproc); 1872 } 1873 #endif /* RACCT */ 1874 curthread->td_ru.ru_inblock++; 1875 } 1876 bp->b_iocmd = BIO_READ; 1877 bp->b_flags &= ~B_INVAL; 1878 if ((flags & GB_CKHASH) != 0) { 1879 bp->b_flags |= B_CKHASH; 1880 bp->b_ckhashcalc = ckhashfunc; 1881 } 1882 bp->b_ioflags &= ~BIO_ERROR; 1883 if (bp->b_rcred == NOCRED && cred != NOCRED) 1884 bp->b_rcred = crhold(cred); 1885 vfs_busy_pages(bp, 0); 1886 bp->b_iooffset = dbtob(bp->b_blkno); 1887 bstrategy(bp); 1888 ++readwait; 1889 } 1890 1891 /* 1892 * Attempt to initiate asynchronous I/O on read-ahead blocks. 1893 */ 1894 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); 1895 1896 rv = 0; 1897 if (readwait) { 1898 rv = bufwait(bp); 1899 if (rv != 0) { 1900 brelse(bp); 1901 *bpp = NULL; 1902 } 1903 } 1904 return (rv); 1905 } 1906 1907 /* 1908 * Write, release buffer on completion. (Done by iodone 1909 * if async). Do not bother writing anything if the buffer 1910 * is invalid. 1911 * 1912 * Note that we set B_CACHE here, indicating that buffer is 1913 * fully valid and thus cacheable. This is true even of NFS 1914 * now so we set it generally. This could be set either here 1915 * or in biodone() since the I/O is synchronous. We put it 1916 * here. 1917 */ 1918 int 1919 bufwrite(struct buf *bp) 1920 { 1921 int oldflags; 1922 struct vnode *vp; 1923 long space; 1924 int vp_md; 1925 1926 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1927 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 1928 bp->b_flags |= B_INVAL | B_RELBUF; 1929 bp->b_flags &= ~B_CACHE; 1930 brelse(bp); 1931 return (ENXIO); 1932 } 1933 if (bp->b_flags & B_INVAL) { 1934 brelse(bp); 1935 return (0); 1936 } 1937 1938 if (bp->b_flags & B_BARRIER) 1939 barrierwrites++; 1940 1941 oldflags = bp->b_flags; 1942 1943 BUF_ASSERT_HELD(bp); 1944 1945 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1946 ("FFS background buffer should not get here %p", bp)); 1947 1948 vp = bp->b_vp; 1949 if (vp) 1950 vp_md = vp->v_vflag & VV_MD; 1951 else 1952 vp_md = 0; 1953 1954 /* 1955 * Mark the buffer clean. Increment the bufobj write count 1956 * before bundirty() call, to prevent other thread from seeing 1957 * empty dirty list and zero counter for writes in progress, 1958 * falsely indicating that the bufobj is clean. 1959 */ 1960 bufobj_wref(bp->b_bufobj); 1961 bundirty(bp); 1962 1963 bp->b_flags &= ~B_DONE; 1964 bp->b_ioflags &= ~BIO_ERROR; 1965 bp->b_flags |= B_CACHE; 1966 bp->b_iocmd = BIO_WRITE; 1967 1968 vfs_busy_pages(bp, 1); 1969 1970 /* 1971 * Normal bwrites pipeline writes 1972 */ 1973 bp->b_runningbufspace = bp->b_bufsize; 1974 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1975 1976 if (!TD_IS_IDLETHREAD(curthread)) { 1977 #ifdef RACCT 1978 if (racct_enable) { 1979 PROC_LOCK(curproc); 1980 racct_add_buf(curproc, bp, 1); 1981 PROC_UNLOCK(curproc); 1982 } 1983 #endif /* RACCT */ 1984 curthread->td_ru.ru_oublock++; 1985 } 1986 if (oldflags & B_ASYNC) 1987 BUF_KERNPROC(bp); 1988 bp->b_iooffset = dbtob(bp->b_blkno); 1989 buf_track(bp, __func__); 1990 bstrategy(bp); 1991 1992 if ((oldflags & B_ASYNC) == 0) { 1993 int rtval = bufwait(bp); 1994 brelse(bp); 1995 return (rtval); 1996 } else if (space > hirunningspace) { 1997 /* 1998 * don't allow the async write to saturate the I/O 1999 * system. We will not deadlock here because 2000 * we are blocking waiting for I/O that is already in-progress 2001 * to complete. We do not block here if it is the update 2002 * or syncer daemon trying to clean up as that can lead 2003 * to deadlock. 2004 */ 2005 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 2006 waitrunningbufspace(); 2007 } 2008 2009 return (0); 2010 } 2011 2012 void 2013 bufbdflush(struct bufobj *bo, struct buf *bp) 2014 { 2015 struct buf *nbp; 2016 2017 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 2018 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 2019 altbufferflushes++; 2020 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 2021 BO_LOCK(bo); 2022 /* 2023 * Try to find a buffer to flush. 2024 */ 2025 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2026 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2027 BUF_LOCK(nbp, 2028 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2029 continue; 2030 if (bp == nbp) 2031 panic("bdwrite: found ourselves"); 2032 BO_UNLOCK(bo); 2033 /* Don't countdeps with the bo lock held. */ 2034 if (buf_countdeps(nbp, 0)) { 2035 BO_LOCK(bo); 2036 BUF_UNLOCK(nbp); 2037 continue; 2038 } 2039 if (nbp->b_flags & B_CLUSTEROK) { 2040 vfs_bio_awrite(nbp); 2041 } else { 2042 bremfree(nbp); 2043 bawrite(nbp); 2044 } 2045 dirtybufferflushes++; 2046 break; 2047 } 2048 if (nbp == NULL) 2049 BO_UNLOCK(bo); 2050 } 2051 } 2052 2053 /* 2054 * Delayed write. (Buffer is marked dirty). Do not bother writing 2055 * anything if the buffer is marked invalid. 2056 * 2057 * Note that since the buffer must be completely valid, we can safely 2058 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 2059 * biodone() in order to prevent getblk from writing the buffer 2060 * out synchronously. 2061 */ 2062 void 2063 bdwrite(struct buf *bp) 2064 { 2065 struct thread *td = curthread; 2066 struct vnode *vp; 2067 struct bufobj *bo; 2068 2069 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2070 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2071 KASSERT((bp->b_flags & B_BARRIER) == 0, 2072 ("Barrier request in delayed write %p", bp)); 2073 BUF_ASSERT_HELD(bp); 2074 2075 if (bp->b_flags & B_INVAL) { 2076 brelse(bp); 2077 return; 2078 } 2079 2080 /* 2081 * If we have too many dirty buffers, don't create any more. 2082 * If we are wildly over our limit, then force a complete 2083 * cleanup. Otherwise, just keep the situation from getting 2084 * out of control. Note that we have to avoid a recursive 2085 * disaster and not try to clean up after our own cleanup! 2086 */ 2087 vp = bp->b_vp; 2088 bo = bp->b_bufobj; 2089 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2090 td->td_pflags |= TDP_INBDFLUSH; 2091 BO_BDFLUSH(bo, bp); 2092 td->td_pflags &= ~TDP_INBDFLUSH; 2093 } else 2094 recursiveflushes++; 2095 2096 bdirty(bp); 2097 /* 2098 * Set B_CACHE, indicating that the buffer is fully valid. This is 2099 * true even of NFS now. 2100 */ 2101 bp->b_flags |= B_CACHE; 2102 2103 /* 2104 * This bmap keeps the system from needing to do the bmap later, 2105 * perhaps when the system is attempting to do a sync. Since it 2106 * is likely that the indirect block -- or whatever other datastructure 2107 * that the filesystem needs is still in memory now, it is a good 2108 * thing to do this. Note also, that if the pageout daemon is 2109 * requesting a sync -- there might not be enough memory to do 2110 * the bmap then... So, this is important to do. 2111 */ 2112 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2113 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2114 } 2115 2116 buf_track(bp, __func__); 2117 2118 /* 2119 * Set the *dirty* buffer range based upon the VM system dirty 2120 * pages. 2121 * 2122 * Mark the buffer pages as clean. We need to do this here to 2123 * satisfy the vnode_pager and the pageout daemon, so that it 2124 * thinks that the pages have been "cleaned". Note that since 2125 * the pages are in a delayed write buffer -- the VFS layer 2126 * "will" see that the pages get written out on the next sync, 2127 * or perhaps the cluster will be completed. 2128 */ 2129 vfs_clean_pages_dirty_buf(bp); 2130 bqrelse(bp); 2131 2132 /* 2133 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2134 * due to the softdep code. 2135 */ 2136 } 2137 2138 /* 2139 * bdirty: 2140 * 2141 * Turn buffer into delayed write request. We must clear BIO_READ and 2142 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2143 * itself to properly update it in the dirty/clean lists. We mark it 2144 * B_DONE to ensure that any asynchronization of the buffer properly 2145 * clears B_DONE ( else a panic will occur later ). 2146 * 2147 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2148 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2149 * should only be called if the buffer is known-good. 2150 * 2151 * Since the buffer is not on a queue, we do not update the numfreebuffers 2152 * count. 2153 * 2154 * The buffer must be on QUEUE_NONE. 2155 */ 2156 void 2157 bdirty(struct buf *bp) 2158 { 2159 2160 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2161 bp, bp->b_vp, bp->b_flags); 2162 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2163 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2164 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2165 BUF_ASSERT_HELD(bp); 2166 bp->b_flags &= ~(B_RELBUF); 2167 bp->b_iocmd = BIO_WRITE; 2168 2169 if ((bp->b_flags & B_DELWRI) == 0) { 2170 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2171 reassignbuf(bp); 2172 bdirtyadd(); 2173 } 2174 } 2175 2176 /* 2177 * bundirty: 2178 * 2179 * Clear B_DELWRI for buffer. 2180 * 2181 * Since the buffer is not on a queue, we do not update the numfreebuffers 2182 * count. 2183 * 2184 * The buffer must be on QUEUE_NONE. 2185 */ 2186 2187 void 2188 bundirty(struct buf *bp) 2189 { 2190 2191 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2192 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2193 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2194 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2195 BUF_ASSERT_HELD(bp); 2196 2197 if (bp->b_flags & B_DELWRI) { 2198 bp->b_flags &= ~B_DELWRI; 2199 reassignbuf(bp); 2200 bdirtysub(); 2201 } 2202 /* 2203 * Since it is now being written, we can clear its deferred write flag. 2204 */ 2205 bp->b_flags &= ~B_DEFERRED; 2206 } 2207 2208 /* 2209 * bawrite: 2210 * 2211 * Asynchronous write. Start output on a buffer, but do not wait for 2212 * it to complete. The buffer is released when the output completes. 2213 * 2214 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2215 * B_INVAL buffers. Not us. 2216 */ 2217 void 2218 bawrite(struct buf *bp) 2219 { 2220 2221 bp->b_flags |= B_ASYNC; 2222 (void) bwrite(bp); 2223 } 2224 2225 /* 2226 * babarrierwrite: 2227 * 2228 * Asynchronous barrier write. Start output on a buffer, but do not 2229 * wait for it to complete. Place a write barrier after this write so 2230 * that this buffer and all buffers written before it are committed to 2231 * the disk before any buffers written after this write are committed 2232 * to the disk. The buffer is released when the output completes. 2233 */ 2234 void 2235 babarrierwrite(struct buf *bp) 2236 { 2237 2238 bp->b_flags |= B_ASYNC | B_BARRIER; 2239 (void) bwrite(bp); 2240 } 2241 2242 /* 2243 * bbarrierwrite: 2244 * 2245 * Synchronous barrier write. Start output on a buffer and wait for 2246 * it to complete. Place a write barrier after this write so that 2247 * this buffer and all buffers written before it are committed to 2248 * the disk before any buffers written after this write are committed 2249 * to the disk. The buffer is released when the output completes. 2250 */ 2251 int 2252 bbarrierwrite(struct buf *bp) 2253 { 2254 2255 bp->b_flags |= B_BARRIER; 2256 return (bwrite(bp)); 2257 } 2258 2259 /* 2260 * bwillwrite: 2261 * 2262 * Called prior to the locking of any vnodes when we are expecting to 2263 * write. We do not want to starve the buffer cache with too many 2264 * dirty buffers so we block here. By blocking prior to the locking 2265 * of any vnodes we attempt to avoid the situation where a locked vnode 2266 * prevents the various system daemons from flushing related buffers. 2267 */ 2268 void 2269 bwillwrite(void) 2270 { 2271 2272 if (numdirtybuffers >= hidirtybuffers) { 2273 mtx_lock(&bdirtylock); 2274 while (numdirtybuffers >= hidirtybuffers) { 2275 bdirtywait = 1; 2276 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2277 "flswai", 0); 2278 } 2279 mtx_unlock(&bdirtylock); 2280 } 2281 } 2282 2283 /* 2284 * Return true if we have too many dirty buffers. 2285 */ 2286 int 2287 buf_dirty_count_severe(void) 2288 { 2289 2290 return(numdirtybuffers >= hidirtybuffers); 2291 } 2292 2293 /* 2294 * brelse: 2295 * 2296 * Release a busy buffer and, if requested, free its resources. The 2297 * buffer will be stashed in the appropriate bufqueue[] allowing it 2298 * to be accessed later as a cache entity or reused for other purposes. 2299 */ 2300 void 2301 brelse(struct buf *bp) 2302 { 2303 int qindex; 2304 2305 /* 2306 * Many functions erroneously call brelse with a NULL bp under rare 2307 * error conditions. Simply return when called with a NULL bp. 2308 */ 2309 if (bp == NULL) 2310 return; 2311 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2312 bp, bp->b_vp, bp->b_flags); 2313 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2314 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2315 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2316 ("brelse: non-VMIO buffer marked NOREUSE")); 2317 2318 if (BUF_LOCKRECURSED(bp)) { 2319 /* 2320 * Do not process, in particular, do not handle the 2321 * B_INVAL/B_RELBUF and do not release to free list. 2322 */ 2323 BUF_UNLOCK(bp); 2324 return; 2325 } 2326 2327 if (bp->b_flags & B_MANAGED) { 2328 bqrelse(bp); 2329 return; 2330 } 2331 2332 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2333 BO_LOCK(bp->b_bufobj); 2334 bp->b_vflags &= ~BV_BKGRDERR; 2335 BO_UNLOCK(bp->b_bufobj); 2336 bdirty(bp); 2337 } 2338 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2339 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && 2340 !(bp->b_flags & B_INVAL)) { 2341 /* 2342 * Failed write, redirty. All errors except ENXIO (which 2343 * means the device is gone) are treated as being 2344 * transient. 2345 * 2346 * XXX Treating EIO as transient is not correct; the 2347 * contract with the local storage device drivers is that 2348 * they will only return EIO once the I/O is no longer 2349 * retriable. Network I/O also respects this through the 2350 * guarantees of TCP and/or the internal retries of NFS. 2351 * ENOMEM might be transient, but we also have no way of 2352 * knowing when its ok to retry/reschedule. In general, 2353 * this entire case should be made obsolete through better 2354 * error handling/recovery and resource scheduling. 2355 * 2356 * Do this also for buffers that failed with ENXIO, but have 2357 * non-empty dependencies - the soft updates code might need 2358 * to access the buffer to untangle them. 2359 * 2360 * Must clear BIO_ERROR to prevent pages from being scrapped. 2361 */ 2362 bp->b_ioflags &= ~BIO_ERROR; 2363 bdirty(bp); 2364 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2365 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2366 /* 2367 * Either a failed read I/O, or we were asked to free or not 2368 * cache the buffer, or we failed to write to a device that's 2369 * no longer present. 2370 */ 2371 bp->b_flags |= B_INVAL; 2372 if (!LIST_EMPTY(&bp->b_dep)) 2373 buf_deallocate(bp); 2374 if (bp->b_flags & B_DELWRI) 2375 bdirtysub(); 2376 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2377 if ((bp->b_flags & B_VMIO) == 0) { 2378 allocbuf(bp, 0); 2379 if (bp->b_vp) 2380 brelvp(bp); 2381 } 2382 } 2383 2384 /* 2385 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2386 * is called with B_DELWRI set, the underlying pages may wind up 2387 * getting freed causing a previous write (bdwrite()) to get 'lost' 2388 * because pages associated with a B_DELWRI bp are marked clean. 2389 * 2390 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2391 * if B_DELWRI is set. 2392 */ 2393 if (bp->b_flags & B_DELWRI) 2394 bp->b_flags &= ~B_RELBUF; 2395 2396 /* 2397 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2398 * constituted, not even NFS buffers now. Two flags effect this. If 2399 * B_INVAL, the struct buf is invalidated but the VM object is kept 2400 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2401 * 2402 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2403 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2404 * buffer is also B_INVAL because it hits the re-dirtying code above. 2405 * 2406 * Normally we can do this whether a buffer is B_DELWRI or not. If 2407 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2408 * the commit state and we cannot afford to lose the buffer. If the 2409 * buffer has a background write in progress, we need to keep it 2410 * around to prevent it from being reconstituted and starting a second 2411 * background write. 2412 */ 2413 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2414 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2415 !(bp->b_vp->v_mount != NULL && 2416 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2417 !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI))) { 2418 vfs_vmio_invalidate(bp); 2419 allocbuf(bp, 0); 2420 } 2421 2422 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2423 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2424 allocbuf(bp, 0); 2425 bp->b_flags &= ~B_NOREUSE; 2426 if (bp->b_vp != NULL) 2427 brelvp(bp); 2428 } 2429 2430 /* 2431 * If the buffer has junk contents signal it and eventually 2432 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2433 * doesn't find it. 2434 */ 2435 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2436 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2437 bp->b_flags |= B_INVAL; 2438 if (bp->b_flags & B_INVAL) { 2439 if (bp->b_flags & B_DELWRI) 2440 bundirty(bp); 2441 if (bp->b_vp) 2442 brelvp(bp); 2443 } 2444 2445 buf_track(bp, __func__); 2446 2447 /* buffers with no memory */ 2448 if (bp->b_bufsize == 0) { 2449 buf_free(bp); 2450 return; 2451 } 2452 /* buffers with junk contents */ 2453 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2454 (bp->b_ioflags & BIO_ERROR)) { 2455 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2456 if (bp->b_vflags & BV_BKGRDINPROG) 2457 panic("losing buffer 2"); 2458 qindex = QUEUE_CLEAN; 2459 bp->b_flags |= B_AGE; 2460 /* remaining buffers */ 2461 } else if (bp->b_flags & B_DELWRI) 2462 qindex = QUEUE_DIRTY; 2463 else 2464 qindex = QUEUE_CLEAN; 2465 2466 binsfree(bp, qindex); 2467 2468 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 2469 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2470 panic("brelse: not dirty"); 2471 /* unlock */ 2472 BUF_UNLOCK(bp); 2473 if (qindex == QUEUE_CLEAN) 2474 bufspace_wakeup(); 2475 } 2476 2477 /* 2478 * Release a buffer back to the appropriate queue but do not try to free 2479 * it. The buffer is expected to be used again soon. 2480 * 2481 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2482 * biodone() to requeue an async I/O on completion. It is also used when 2483 * known good buffers need to be requeued but we think we may need the data 2484 * again soon. 2485 * 2486 * XXX we should be able to leave the B_RELBUF hint set on completion. 2487 */ 2488 void 2489 bqrelse(struct buf *bp) 2490 { 2491 int qindex; 2492 2493 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2494 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2495 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2496 2497 qindex = QUEUE_NONE; 2498 if (BUF_LOCKRECURSED(bp)) { 2499 /* do not release to free list */ 2500 BUF_UNLOCK(bp); 2501 return; 2502 } 2503 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2504 2505 if (bp->b_flags & B_MANAGED) { 2506 if (bp->b_flags & B_REMFREE) 2507 bremfreef(bp); 2508 goto out; 2509 } 2510 2511 /* buffers with stale but valid contents */ 2512 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2513 BV_BKGRDERR)) == BV_BKGRDERR) { 2514 BO_LOCK(bp->b_bufobj); 2515 bp->b_vflags &= ~BV_BKGRDERR; 2516 BO_UNLOCK(bp->b_bufobj); 2517 qindex = QUEUE_DIRTY; 2518 } else { 2519 if ((bp->b_flags & B_DELWRI) == 0 && 2520 (bp->b_xflags & BX_VNDIRTY)) 2521 panic("bqrelse: not dirty"); 2522 if ((bp->b_flags & B_NOREUSE) != 0) { 2523 brelse(bp); 2524 return; 2525 } 2526 qindex = QUEUE_CLEAN; 2527 } 2528 binsfree(bp, qindex); 2529 2530 out: 2531 buf_track(bp, __func__); 2532 /* unlock */ 2533 BUF_UNLOCK(bp); 2534 if (qindex == QUEUE_CLEAN) 2535 bufspace_wakeup(); 2536 } 2537 2538 /* 2539 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2540 * restore bogus pages. 2541 */ 2542 static void 2543 vfs_vmio_iodone(struct buf *bp) 2544 { 2545 vm_ooffset_t foff; 2546 vm_page_t m; 2547 vm_object_t obj; 2548 struct vnode *vp; 2549 int i, iosize, resid; 2550 bool bogus; 2551 2552 obj = bp->b_bufobj->bo_object; 2553 KASSERT(obj->paging_in_progress >= bp->b_npages, 2554 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2555 obj->paging_in_progress, bp->b_npages)); 2556 2557 vp = bp->b_vp; 2558 KASSERT(vp->v_holdcnt > 0, 2559 ("vfs_vmio_iodone: vnode %p has zero hold count", vp)); 2560 KASSERT(vp->v_object != NULL, 2561 ("vfs_vmio_iodone: vnode %p has no vm_object", vp)); 2562 2563 foff = bp->b_offset; 2564 KASSERT(bp->b_offset != NOOFFSET, 2565 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2566 2567 bogus = false; 2568 iosize = bp->b_bcount - bp->b_resid; 2569 VM_OBJECT_WLOCK(obj); 2570 for (i = 0; i < bp->b_npages; i++) { 2571 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2572 if (resid > iosize) 2573 resid = iosize; 2574 2575 /* 2576 * cleanup bogus pages, restoring the originals 2577 */ 2578 m = bp->b_pages[i]; 2579 if (m == bogus_page) { 2580 bogus = true; 2581 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2582 if (m == NULL) 2583 panic("biodone: page disappeared!"); 2584 bp->b_pages[i] = m; 2585 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2586 /* 2587 * In the write case, the valid and clean bits are 2588 * already changed correctly ( see bdwrite() ), so we 2589 * only need to do this here in the read case. 2590 */ 2591 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2592 resid)) == 0, ("vfs_vmio_iodone: page %p " 2593 "has unexpected dirty bits", m)); 2594 vfs_page_set_valid(bp, foff, m); 2595 } 2596 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2597 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2598 (intmax_t)foff, (uintmax_t)m->pindex)); 2599 2600 vm_page_sunbusy(m); 2601 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2602 iosize -= resid; 2603 } 2604 vm_object_pip_wakeupn(obj, bp->b_npages); 2605 VM_OBJECT_WUNLOCK(obj); 2606 if (bogus && buf_mapped(bp)) { 2607 BUF_CHECK_MAPPED(bp); 2608 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2609 bp->b_pages, bp->b_npages); 2610 } 2611 } 2612 2613 /* 2614 * Unwire a page held by a buf and place it on the appropriate vm queue. 2615 */ 2616 static void 2617 vfs_vmio_unwire(struct buf *bp, vm_page_t m) 2618 { 2619 bool freed; 2620 2621 vm_page_lock(m); 2622 if (vm_page_unwire(m, PQ_NONE)) { 2623 /* 2624 * Determine if the page should be freed before adding 2625 * it to the inactive queue. 2626 */ 2627 if (m->valid == 0) { 2628 freed = !vm_page_busied(m); 2629 if (freed) 2630 vm_page_free(m); 2631 } else if ((bp->b_flags & B_DIRECT) != 0) 2632 freed = vm_page_try_to_free(m); 2633 else 2634 freed = false; 2635 if (!freed) { 2636 /* 2637 * If the page is unlikely to be reused, let the 2638 * VM know. Otherwise, maintain LRU page 2639 * ordering and put the page at the tail of the 2640 * inactive queue. 2641 */ 2642 if ((bp->b_flags & B_NOREUSE) != 0) 2643 vm_page_deactivate_noreuse(m); 2644 else 2645 vm_page_deactivate(m); 2646 } 2647 } 2648 vm_page_unlock(m); 2649 } 2650 2651 /* 2652 * Perform page invalidation when a buffer is released. The fully invalid 2653 * pages will be reclaimed later in vfs_vmio_truncate(). 2654 */ 2655 static void 2656 vfs_vmio_invalidate(struct buf *bp) 2657 { 2658 vm_object_t obj; 2659 vm_page_t m; 2660 int i, resid, poffset, presid; 2661 2662 if (buf_mapped(bp)) { 2663 BUF_CHECK_MAPPED(bp); 2664 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 2665 } else 2666 BUF_CHECK_UNMAPPED(bp); 2667 /* 2668 * Get the base offset and length of the buffer. Note that 2669 * in the VMIO case if the buffer block size is not 2670 * page-aligned then b_data pointer may not be page-aligned. 2671 * But our b_pages[] array *IS* page aligned. 2672 * 2673 * block sizes less then DEV_BSIZE (usually 512) are not 2674 * supported due to the page granularity bits (m->valid, 2675 * m->dirty, etc...). 2676 * 2677 * See man buf(9) for more information 2678 */ 2679 obj = bp->b_bufobj->bo_object; 2680 resid = bp->b_bufsize; 2681 poffset = bp->b_offset & PAGE_MASK; 2682 VM_OBJECT_WLOCK(obj); 2683 for (i = 0; i < bp->b_npages; i++) { 2684 m = bp->b_pages[i]; 2685 if (m == bogus_page) 2686 panic("vfs_vmio_invalidate: Unexpected bogus page."); 2687 bp->b_pages[i] = NULL; 2688 2689 presid = resid > (PAGE_SIZE - poffset) ? 2690 (PAGE_SIZE - poffset) : resid; 2691 KASSERT(presid >= 0, ("brelse: extra page")); 2692 while (vm_page_xbusied(m)) { 2693 vm_page_lock(m); 2694 VM_OBJECT_WUNLOCK(obj); 2695 vm_page_busy_sleep(m, "mbncsh", true); 2696 VM_OBJECT_WLOCK(obj); 2697 } 2698 if (pmap_page_wired_mappings(m) == 0) 2699 vm_page_set_invalid(m, poffset, presid); 2700 vfs_vmio_unwire(bp, m); 2701 resid -= presid; 2702 poffset = 0; 2703 } 2704 VM_OBJECT_WUNLOCK(obj); 2705 bp->b_npages = 0; 2706 } 2707 2708 /* 2709 * Page-granular truncation of an existing VMIO buffer. 2710 */ 2711 static void 2712 vfs_vmio_truncate(struct buf *bp, int desiredpages) 2713 { 2714 vm_object_t obj; 2715 vm_page_t m; 2716 int i; 2717 2718 if (bp->b_npages == desiredpages) 2719 return; 2720 2721 if (buf_mapped(bp)) { 2722 BUF_CHECK_MAPPED(bp); 2723 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 2724 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 2725 } else 2726 BUF_CHECK_UNMAPPED(bp); 2727 obj = bp->b_bufobj->bo_object; 2728 if (obj != NULL) 2729 VM_OBJECT_WLOCK(obj); 2730 for (i = desiredpages; i < bp->b_npages; i++) { 2731 m = bp->b_pages[i]; 2732 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 2733 bp->b_pages[i] = NULL; 2734 vfs_vmio_unwire(bp, m); 2735 } 2736 if (obj != NULL) 2737 VM_OBJECT_WUNLOCK(obj); 2738 bp->b_npages = desiredpages; 2739 } 2740 2741 /* 2742 * Byte granular extension of VMIO buffers. 2743 */ 2744 static void 2745 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 2746 { 2747 /* 2748 * We are growing the buffer, possibly in a 2749 * byte-granular fashion. 2750 */ 2751 vm_object_t obj; 2752 vm_offset_t toff; 2753 vm_offset_t tinc; 2754 vm_page_t m; 2755 2756 /* 2757 * Step 1, bring in the VM pages from the object, allocating 2758 * them if necessary. We must clear B_CACHE if these pages 2759 * are not valid for the range covered by the buffer. 2760 */ 2761 obj = bp->b_bufobj->bo_object; 2762 VM_OBJECT_WLOCK(obj); 2763 if (bp->b_npages < desiredpages) { 2764 /* 2765 * We must allocate system pages since blocking 2766 * here could interfere with paging I/O, no 2767 * matter which process we are. 2768 * 2769 * Only exclusive busy can be tested here. 2770 * Blocking on shared busy might lead to 2771 * deadlocks once allocbuf() is called after 2772 * pages are vfs_busy_pages(). 2773 */ 2774 (void)vm_page_grab_pages(obj, 2775 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 2776 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 2777 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 2778 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 2779 bp->b_npages = desiredpages; 2780 } 2781 2782 /* 2783 * Step 2. We've loaded the pages into the buffer, 2784 * we have to figure out if we can still have B_CACHE 2785 * set. Note that B_CACHE is set according to the 2786 * byte-granular range ( bcount and size ), not the 2787 * aligned range ( newbsize ). 2788 * 2789 * The VM test is against m->valid, which is DEV_BSIZE 2790 * aligned. Needless to say, the validity of the data 2791 * needs to also be DEV_BSIZE aligned. Note that this 2792 * fails with NFS if the server or some other client 2793 * extends the file's EOF. If our buffer is resized, 2794 * B_CACHE may remain set! XXX 2795 */ 2796 toff = bp->b_bcount; 2797 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2798 while ((bp->b_flags & B_CACHE) && toff < size) { 2799 vm_pindex_t pi; 2800 2801 if (tinc > (size - toff)) 2802 tinc = size - toff; 2803 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 2804 m = bp->b_pages[pi]; 2805 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 2806 toff += tinc; 2807 tinc = PAGE_SIZE; 2808 } 2809 VM_OBJECT_WUNLOCK(obj); 2810 2811 /* 2812 * Step 3, fixup the KVA pmap. 2813 */ 2814 if (buf_mapped(bp)) 2815 bpmap_qenter(bp); 2816 else 2817 BUF_CHECK_UNMAPPED(bp); 2818 } 2819 2820 /* 2821 * Check to see if a block at a particular lbn is available for a clustered 2822 * write. 2823 */ 2824 static int 2825 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 2826 { 2827 struct buf *bpa; 2828 int match; 2829 2830 match = 0; 2831 2832 /* If the buf isn't in core skip it */ 2833 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 2834 return (0); 2835 2836 /* If the buf is busy we don't want to wait for it */ 2837 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2838 return (0); 2839 2840 /* Only cluster with valid clusterable delayed write buffers */ 2841 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 2842 (B_DELWRI | B_CLUSTEROK)) 2843 goto done; 2844 2845 if (bpa->b_bufsize != size) 2846 goto done; 2847 2848 /* 2849 * Check to see if it is in the expected place on disk and that the 2850 * block has been mapped. 2851 */ 2852 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 2853 match = 1; 2854 done: 2855 BUF_UNLOCK(bpa); 2856 return (match); 2857 } 2858 2859 /* 2860 * vfs_bio_awrite: 2861 * 2862 * Implement clustered async writes for clearing out B_DELWRI buffers. 2863 * This is much better then the old way of writing only one buffer at 2864 * a time. Note that we may not be presented with the buffers in the 2865 * correct order, so we search for the cluster in both directions. 2866 */ 2867 int 2868 vfs_bio_awrite(struct buf *bp) 2869 { 2870 struct bufobj *bo; 2871 int i; 2872 int j; 2873 daddr_t lblkno = bp->b_lblkno; 2874 struct vnode *vp = bp->b_vp; 2875 int ncl; 2876 int nwritten; 2877 int size; 2878 int maxcl; 2879 int gbflags; 2880 2881 bo = &vp->v_bufobj; 2882 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 2883 /* 2884 * right now we support clustered writing only to regular files. If 2885 * we find a clusterable block we could be in the middle of a cluster 2886 * rather then at the beginning. 2887 */ 2888 if ((vp->v_type == VREG) && 2889 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 2890 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 2891 2892 size = vp->v_mount->mnt_stat.f_iosize; 2893 maxcl = MAXPHYS / size; 2894 2895 BO_RLOCK(bo); 2896 for (i = 1; i < maxcl; i++) 2897 if (vfs_bio_clcheck(vp, size, lblkno + i, 2898 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 2899 break; 2900 2901 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 2902 if (vfs_bio_clcheck(vp, size, lblkno - j, 2903 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 2904 break; 2905 BO_RUNLOCK(bo); 2906 --j; 2907 ncl = i + j; 2908 /* 2909 * this is a possible cluster write 2910 */ 2911 if (ncl != 1) { 2912 BUF_UNLOCK(bp); 2913 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 2914 gbflags); 2915 return (nwritten); 2916 } 2917 } 2918 bremfree(bp); 2919 bp->b_flags |= B_ASYNC; 2920 /* 2921 * default (old) behavior, writing out only one block 2922 * 2923 * XXX returns b_bufsize instead of b_bcount for nwritten? 2924 */ 2925 nwritten = bp->b_bufsize; 2926 (void) bwrite(bp); 2927 2928 return (nwritten); 2929 } 2930 2931 /* 2932 * getnewbuf_kva: 2933 * 2934 * Allocate KVA for an empty buf header according to gbflags. 2935 */ 2936 static int 2937 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 2938 { 2939 2940 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 2941 /* 2942 * In order to keep fragmentation sane we only allocate kva 2943 * in BKVASIZE chunks. XXX with vmem we can do page size. 2944 */ 2945 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2946 2947 if (maxsize != bp->b_kvasize && 2948 bufkva_alloc(bp, maxsize, gbflags)) 2949 return (ENOSPC); 2950 } 2951 return (0); 2952 } 2953 2954 /* 2955 * getnewbuf: 2956 * 2957 * Find and initialize a new buffer header, freeing up existing buffers 2958 * in the bufqueues as necessary. The new buffer is returned locked. 2959 * 2960 * We block if: 2961 * We have insufficient buffer headers 2962 * We have insufficient buffer space 2963 * buffer_arena is too fragmented ( space reservation fails ) 2964 * If we have to flush dirty buffers ( but we try to avoid this ) 2965 * 2966 * The caller is responsible for releasing the reserved bufspace after 2967 * allocbuf() is called. 2968 */ 2969 static struct buf * 2970 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 2971 { 2972 struct buf *bp; 2973 bool metadata, reserved; 2974 2975 bp = NULL; 2976 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2977 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2978 if (!unmapped_buf_allowed) 2979 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2980 2981 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2982 vp->v_type == VCHR) 2983 metadata = true; 2984 else 2985 metadata = false; 2986 atomic_add_int(&getnewbufcalls, 1); 2987 reserved = false; 2988 do { 2989 if (reserved == false && 2990 bufspace_reserve(maxsize, metadata) != 0) 2991 continue; 2992 reserved = true; 2993 if ((bp = buf_alloc()) == NULL) 2994 continue; 2995 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 2996 return (bp); 2997 break; 2998 } while(buf_scan(false) == 0); 2999 3000 if (reserved) 3001 atomic_subtract_long(&bufspace, maxsize); 3002 if (bp != NULL) { 3003 bp->b_flags |= B_INVAL; 3004 brelse(bp); 3005 } 3006 bufspace_wait(vp, gbflags, slpflag, slptimeo); 3007 3008 return (NULL); 3009 } 3010 3011 /* 3012 * buf_daemon: 3013 * 3014 * buffer flushing daemon. Buffers are normally flushed by the 3015 * update daemon but if it cannot keep up this process starts to 3016 * take the load in an attempt to prevent getnewbuf() from blocking. 3017 */ 3018 static struct kproc_desc buf_kp = { 3019 "bufdaemon", 3020 buf_daemon, 3021 &bufdaemonproc 3022 }; 3023 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3024 3025 static int 3026 buf_flush(struct vnode *vp, int target) 3027 { 3028 int flushed; 3029 3030 flushed = flushbufqueues(vp, target, 0); 3031 if (flushed == 0) { 3032 /* 3033 * Could not find any buffers without rollback 3034 * dependencies, so just write the first one 3035 * in the hopes of eventually making progress. 3036 */ 3037 if (vp != NULL && target > 2) 3038 target /= 2; 3039 flushbufqueues(vp, target, 1); 3040 } 3041 return (flushed); 3042 } 3043 3044 static void 3045 buf_daemon() 3046 { 3047 int lodirty; 3048 3049 /* 3050 * This process needs to be suspended prior to shutdown sync. 3051 */ 3052 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 3053 SHUTDOWN_PRI_LAST); 3054 3055 /* 3056 * This process is allowed to take the buffer cache to the limit 3057 */ 3058 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3059 mtx_lock(&bdlock); 3060 for (;;) { 3061 bd_request = 0; 3062 mtx_unlock(&bdlock); 3063 3064 kproc_suspend_check(bufdaemonproc); 3065 lodirty = lodirtybuffers; 3066 if (bd_speedupreq) { 3067 lodirty = numdirtybuffers / 2; 3068 bd_speedupreq = 0; 3069 } 3070 /* 3071 * Do the flush. Limit the amount of in-transit I/O we 3072 * allow to build up, otherwise we would completely saturate 3073 * the I/O system. 3074 */ 3075 while (numdirtybuffers > lodirty) { 3076 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0) 3077 break; 3078 kern_yield(PRI_USER); 3079 } 3080 3081 /* 3082 * Only clear bd_request if we have reached our low water 3083 * mark. The buf_daemon normally waits 1 second and 3084 * then incrementally flushes any dirty buffers that have 3085 * built up, within reason. 3086 * 3087 * If we were unable to hit our low water mark and couldn't 3088 * find any flushable buffers, we sleep for a short period 3089 * to avoid endless loops on unlockable buffers. 3090 */ 3091 mtx_lock(&bdlock); 3092 if (numdirtybuffers <= lodirtybuffers) { 3093 /* 3094 * We reached our low water mark, reset the 3095 * request and sleep until we are needed again. 3096 * The sleep is just so the suspend code works. 3097 */ 3098 bd_request = 0; 3099 /* 3100 * Do an extra wakeup in case dirty threshold 3101 * changed via sysctl and the explicit transition 3102 * out of shortfall was missed. 3103 */ 3104 bdirtywakeup(); 3105 if (runningbufspace <= lorunningspace) 3106 runningwakeup(); 3107 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3108 } else { 3109 /* 3110 * We couldn't find any flushable dirty buffers but 3111 * still have too many dirty buffers, we 3112 * have to sleep and try again. (rare) 3113 */ 3114 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3115 } 3116 } 3117 } 3118 3119 /* 3120 * flushbufqueues: 3121 * 3122 * Try to flush a buffer in the dirty queue. We must be careful to 3123 * free up B_INVAL buffers instead of write them, which NFS is 3124 * particularly sensitive to. 3125 */ 3126 static int flushwithdeps = 0; 3127 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 3128 0, "Number of buffers flushed with dependecies that require rollbacks"); 3129 3130 static int 3131 flushbufqueues(struct vnode *lvp, int target, int flushdeps) 3132 { 3133 struct buf *sentinel; 3134 struct vnode *vp; 3135 struct mount *mp; 3136 struct buf *bp; 3137 int hasdeps; 3138 int flushed; 3139 int queue; 3140 int error; 3141 bool unlock; 3142 3143 flushed = 0; 3144 queue = QUEUE_DIRTY; 3145 bp = NULL; 3146 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3147 sentinel->b_qindex = QUEUE_SENTINEL; 3148 mtx_lock(&bqlocks[queue]); 3149 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 3150 mtx_unlock(&bqlocks[queue]); 3151 while (flushed != target) { 3152 maybe_yield(); 3153 mtx_lock(&bqlocks[queue]); 3154 bp = TAILQ_NEXT(sentinel, b_freelist); 3155 if (bp != NULL) { 3156 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 3157 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 3158 b_freelist); 3159 } else { 3160 mtx_unlock(&bqlocks[queue]); 3161 break; 3162 } 3163 /* 3164 * Skip sentinels inserted by other invocations of the 3165 * flushbufqueues(), taking care to not reorder them. 3166 * 3167 * Only flush the buffers that belong to the 3168 * vnode locked by the curthread. 3169 */ 3170 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3171 bp->b_vp != lvp)) { 3172 mtx_unlock(&bqlocks[queue]); 3173 continue; 3174 } 3175 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3176 mtx_unlock(&bqlocks[queue]); 3177 if (error != 0) 3178 continue; 3179 3180 /* 3181 * BKGRDINPROG can only be set with the buf and bufobj 3182 * locks both held. We tolerate a race to clear it here. 3183 */ 3184 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3185 (bp->b_flags & B_DELWRI) == 0) { 3186 BUF_UNLOCK(bp); 3187 continue; 3188 } 3189 if (bp->b_flags & B_INVAL) { 3190 bremfreef(bp); 3191 brelse(bp); 3192 flushed++; 3193 continue; 3194 } 3195 3196 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3197 if (flushdeps == 0) { 3198 BUF_UNLOCK(bp); 3199 continue; 3200 } 3201 hasdeps = 1; 3202 } else 3203 hasdeps = 0; 3204 /* 3205 * We must hold the lock on a vnode before writing 3206 * one of its buffers. Otherwise we may confuse, or 3207 * in the case of a snapshot vnode, deadlock the 3208 * system. 3209 * 3210 * The lock order here is the reverse of the normal 3211 * of vnode followed by buf lock. This is ok because 3212 * the NOWAIT will prevent deadlock. 3213 */ 3214 vp = bp->b_vp; 3215 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3216 BUF_UNLOCK(bp); 3217 continue; 3218 } 3219 if (lvp == NULL) { 3220 unlock = true; 3221 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3222 } else { 3223 ASSERT_VOP_LOCKED(vp, "getbuf"); 3224 unlock = false; 3225 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3226 vn_lock(vp, LK_TRYUPGRADE); 3227 } 3228 if (error == 0) { 3229 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3230 bp, bp->b_vp, bp->b_flags); 3231 if (curproc == bufdaemonproc) { 3232 vfs_bio_awrite(bp); 3233 } else { 3234 bremfree(bp); 3235 bwrite(bp); 3236 notbufdflushes++; 3237 } 3238 vn_finished_write(mp); 3239 if (unlock) 3240 VOP_UNLOCK(vp, 0); 3241 flushwithdeps += hasdeps; 3242 flushed++; 3243 3244 /* 3245 * Sleeping on runningbufspace while holding 3246 * vnode lock leads to deadlock. 3247 */ 3248 if (curproc == bufdaemonproc && 3249 runningbufspace > hirunningspace) 3250 waitrunningbufspace(); 3251 continue; 3252 } 3253 vn_finished_write(mp); 3254 BUF_UNLOCK(bp); 3255 } 3256 mtx_lock(&bqlocks[queue]); 3257 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 3258 mtx_unlock(&bqlocks[queue]); 3259 free(sentinel, M_TEMP); 3260 return (flushed); 3261 } 3262 3263 /* 3264 * Check to see if a block is currently memory resident. 3265 */ 3266 struct buf * 3267 incore(struct bufobj *bo, daddr_t blkno) 3268 { 3269 struct buf *bp; 3270 3271 BO_RLOCK(bo); 3272 bp = gbincore(bo, blkno); 3273 BO_RUNLOCK(bo); 3274 return (bp); 3275 } 3276 3277 /* 3278 * Returns true if no I/O is needed to access the 3279 * associated VM object. This is like incore except 3280 * it also hunts around in the VM system for the data. 3281 */ 3282 3283 static int 3284 inmem(struct vnode * vp, daddr_t blkno) 3285 { 3286 vm_object_t obj; 3287 vm_offset_t toff, tinc, size; 3288 vm_page_t m; 3289 vm_ooffset_t off; 3290 3291 ASSERT_VOP_LOCKED(vp, "inmem"); 3292 3293 if (incore(&vp->v_bufobj, blkno)) 3294 return 1; 3295 if (vp->v_mount == NULL) 3296 return 0; 3297 obj = vp->v_object; 3298 if (obj == NULL) 3299 return (0); 3300 3301 size = PAGE_SIZE; 3302 if (size > vp->v_mount->mnt_stat.f_iosize) 3303 size = vp->v_mount->mnt_stat.f_iosize; 3304 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3305 3306 VM_OBJECT_RLOCK(obj); 3307 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3308 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 3309 if (!m) 3310 goto notinmem; 3311 tinc = size; 3312 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3313 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3314 if (vm_page_is_valid(m, 3315 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 3316 goto notinmem; 3317 } 3318 VM_OBJECT_RUNLOCK(obj); 3319 return 1; 3320 3321 notinmem: 3322 VM_OBJECT_RUNLOCK(obj); 3323 return (0); 3324 } 3325 3326 /* 3327 * Set the dirty range for a buffer based on the status of the dirty 3328 * bits in the pages comprising the buffer. The range is limited 3329 * to the size of the buffer. 3330 * 3331 * Tell the VM system that the pages associated with this buffer 3332 * are clean. This is used for delayed writes where the data is 3333 * going to go to disk eventually without additional VM intevention. 3334 * 3335 * Note that while we only really need to clean through to b_bcount, we 3336 * just go ahead and clean through to b_bufsize. 3337 */ 3338 static void 3339 vfs_clean_pages_dirty_buf(struct buf *bp) 3340 { 3341 vm_ooffset_t foff, noff, eoff; 3342 vm_page_t m; 3343 int i; 3344 3345 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3346 return; 3347 3348 foff = bp->b_offset; 3349 KASSERT(bp->b_offset != NOOFFSET, 3350 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3351 3352 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3353 vfs_drain_busy_pages(bp); 3354 vfs_setdirty_locked_object(bp); 3355 for (i = 0; i < bp->b_npages; i++) { 3356 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3357 eoff = noff; 3358 if (eoff > bp->b_offset + bp->b_bufsize) 3359 eoff = bp->b_offset + bp->b_bufsize; 3360 m = bp->b_pages[i]; 3361 vfs_page_set_validclean(bp, foff, m); 3362 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3363 foff = noff; 3364 } 3365 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3366 } 3367 3368 static void 3369 vfs_setdirty_locked_object(struct buf *bp) 3370 { 3371 vm_object_t object; 3372 int i; 3373 3374 object = bp->b_bufobj->bo_object; 3375 VM_OBJECT_ASSERT_WLOCKED(object); 3376 3377 /* 3378 * We qualify the scan for modified pages on whether the 3379 * object has been flushed yet. 3380 */ 3381 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 3382 vm_offset_t boffset; 3383 vm_offset_t eoffset; 3384 3385 /* 3386 * test the pages to see if they have been modified directly 3387 * by users through the VM system. 3388 */ 3389 for (i = 0; i < bp->b_npages; i++) 3390 vm_page_test_dirty(bp->b_pages[i]); 3391 3392 /* 3393 * Calculate the encompassing dirty range, boffset and eoffset, 3394 * (eoffset - boffset) bytes. 3395 */ 3396 3397 for (i = 0; i < bp->b_npages; i++) { 3398 if (bp->b_pages[i]->dirty) 3399 break; 3400 } 3401 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3402 3403 for (i = bp->b_npages - 1; i >= 0; --i) { 3404 if (bp->b_pages[i]->dirty) { 3405 break; 3406 } 3407 } 3408 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3409 3410 /* 3411 * Fit it to the buffer. 3412 */ 3413 3414 if (eoffset > bp->b_bcount) 3415 eoffset = bp->b_bcount; 3416 3417 /* 3418 * If we have a good dirty range, merge with the existing 3419 * dirty range. 3420 */ 3421 3422 if (boffset < eoffset) { 3423 if (bp->b_dirtyoff > boffset) 3424 bp->b_dirtyoff = boffset; 3425 if (bp->b_dirtyend < eoffset) 3426 bp->b_dirtyend = eoffset; 3427 } 3428 } 3429 } 3430 3431 /* 3432 * Allocate the KVA mapping for an existing buffer. 3433 * If an unmapped buffer is provided but a mapped buffer is requested, take 3434 * also care to properly setup mappings between pages and KVA. 3435 */ 3436 static void 3437 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3438 { 3439 int bsize, maxsize, need_mapping, need_kva; 3440 off_t offset; 3441 3442 need_mapping = bp->b_data == unmapped_buf && 3443 (gbflags & GB_UNMAPPED) == 0; 3444 need_kva = bp->b_kvabase == unmapped_buf && 3445 bp->b_data == unmapped_buf && 3446 (gbflags & GB_KVAALLOC) != 0; 3447 if (!need_mapping && !need_kva) 3448 return; 3449 3450 BUF_CHECK_UNMAPPED(bp); 3451 3452 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3453 /* 3454 * Buffer is not mapped, but the KVA was already 3455 * reserved at the time of the instantiation. Use the 3456 * allocated space. 3457 */ 3458 goto has_addr; 3459 } 3460 3461 /* 3462 * Calculate the amount of the address space we would reserve 3463 * if the buffer was mapped. 3464 */ 3465 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3466 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3467 offset = blkno * bsize; 3468 maxsize = size + (offset & PAGE_MASK); 3469 maxsize = imax(maxsize, bsize); 3470 3471 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3472 if ((gbflags & GB_NOWAIT_BD) != 0) { 3473 /* 3474 * XXXKIB: defragmentation cannot 3475 * succeed, not sure what else to do. 3476 */ 3477 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3478 } 3479 atomic_add_int(&mappingrestarts, 1); 3480 bufspace_wait(bp->b_vp, gbflags, 0, 0); 3481 } 3482 has_addr: 3483 if (need_mapping) { 3484 /* b_offset is handled by bpmap_qenter. */ 3485 bp->b_data = bp->b_kvabase; 3486 BUF_CHECK_MAPPED(bp); 3487 bpmap_qenter(bp); 3488 } 3489 } 3490 3491 /* 3492 * getblk: 3493 * 3494 * Get a block given a specified block and offset into a file/device. 3495 * The buffers B_DONE bit will be cleared on return, making it almost 3496 * ready for an I/O initiation. B_INVAL may or may not be set on 3497 * return. The caller should clear B_INVAL prior to initiating a 3498 * READ. 3499 * 3500 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3501 * an existing buffer. 3502 * 3503 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3504 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3505 * and then cleared based on the backing VM. If the previous buffer is 3506 * non-0-sized but invalid, B_CACHE will be cleared. 3507 * 3508 * If getblk() must create a new buffer, the new buffer is returned with 3509 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3510 * case it is returned with B_INVAL clear and B_CACHE set based on the 3511 * backing VM. 3512 * 3513 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3514 * B_CACHE bit is clear. 3515 * 3516 * What this means, basically, is that the caller should use B_CACHE to 3517 * determine whether the buffer is fully valid or not and should clear 3518 * B_INVAL prior to issuing a read. If the caller intends to validate 3519 * the buffer by loading its data area with something, the caller needs 3520 * to clear B_INVAL. If the caller does this without issuing an I/O, 3521 * the caller should set B_CACHE ( as an optimization ), else the caller 3522 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3523 * a write attempt or if it was a successful read. If the caller 3524 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3525 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3526 */ 3527 struct buf * 3528 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3529 int flags) 3530 { 3531 struct buf *bp; 3532 struct bufobj *bo; 3533 int bsize, error, maxsize, vmio; 3534 off_t offset; 3535 3536 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3537 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3538 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3539 ASSERT_VOP_LOCKED(vp, "getblk"); 3540 if (size > maxbcachebuf) 3541 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, 3542 maxbcachebuf); 3543 if (!unmapped_buf_allowed) 3544 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3545 3546 bo = &vp->v_bufobj; 3547 loop: 3548 BO_RLOCK(bo); 3549 bp = gbincore(bo, blkno); 3550 if (bp != NULL) { 3551 int lockflags; 3552 /* 3553 * Buffer is in-core. If the buffer is not busy nor managed, 3554 * it must be on a queue. 3555 */ 3556 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3557 3558 if (flags & GB_LOCK_NOWAIT) 3559 lockflags |= LK_NOWAIT; 3560 3561 error = BUF_TIMELOCK(bp, lockflags, 3562 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3563 3564 /* 3565 * If we slept and got the lock we have to restart in case 3566 * the buffer changed identities. 3567 */ 3568 if (error == ENOLCK) 3569 goto loop; 3570 /* We timed out or were interrupted. */ 3571 else if (error) 3572 return (NULL); 3573 /* If recursed, assume caller knows the rules. */ 3574 else if (BUF_LOCKRECURSED(bp)) 3575 goto end; 3576 3577 /* 3578 * The buffer is locked. B_CACHE is cleared if the buffer is 3579 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3580 * and for a VMIO buffer B_CACHE is adjusted according to the 3581 * backing VM cache. 3582 */ 3583 if (bp->b_flags & B_INVAL) 3584 bp->b_flags &= ~B_CACHE; 3585 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3586 bp->b_flags |= B_CACHE; 3587 if (bp->b_flags & B_MANAGED) 3588 MPASS(bp->b_qindex == QUEUE_NONE); 3589 else 3590 bremfree(bp); 3591 3592 /* 3593 * check for size inconsistencies for non-VMIO case. 3594 */ 3595 if (bp->b_bcount != size) { 3596 if ((bp->b_flags & B_VMIO) == 0 || 3597 (size > bp->b_kvasize)) { 3598 if (bp->b_flags & B_DELWRI) { 3599 bp->b_flags |= B_NOCACHE; 3600 bwrite(bp); 3601 } else { 3602 if (LIST_EMPTY(&bp->b_dep)) { 3603 bp->b_flags |= B_RELBUF; 3604 brelse(bp); 3605 } else { 3606 bp->b_flags |= B_NOCACHE; 3607 bwrite(bp); 3608 } 3609 } 3610 goto loop; 3611 } 3612 } 3613 3614 /* 3615 * Handle the case of unmapped buffer which should 3616 * become mapped, or the buffer for which KVA 3617 * reservation is requested. 3618 */ 3619 bp_unmapped_get_kva(bp, blkno, size, flags); 3620 3621 /* 3622 * If the size is inconsistent in the VMIO case, we can resize 3623 * the buffer. This might lead to B_CACHE getting set or 3624 * cleared. If the size has not changed, B_CACHE remains 3625 * unchanged from its previous state. 3626 */ 3627 allocbuf(bp, size); 3628 3629 KASSERT(bp->b_offset != NOOFFSET, 3630 ("getblk: no buffer offset")); 3631 3632 /* 3633 * A buffer with B_DELWRI set and B_CACHE clear must 3634 * be committed before we can return the buffer in 3635 * order to prevent the caller from issuing a read 3636 * ( due to B_CACHE not being set ) and overwriting 3637 * it. 3638 * 3639 * Most callers, including NFS and FFS, need this to 3640 * operate properly either because they assume they 3641 * can issue a read if B_CACHE is not set, or because 3642 * ( for example ) an uncached B_DELWRI might loop due 3643 * to softupdates re-dirtying the buffer. In the latter 3644 * case, B_CACHE is set after the first write completes, 3645 * preventing further loops. 3646 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3647 * above while extending the buffer, we cannot allow the 3648 * buffer to remain with B_CACHE set after the write 3649 * completes or it will represent a corrupt state. To 3650 * deal with this we set B_NOCACHE to scrap the buffer 3651 * after the write. 3652 * 3653 * We might be able to do something fancy, like setting 3654 * B_CACHE in bwrite() except if B_DELWRI is already set, 3655 * so the below call doesn't set B_CACHE, but that gets real 3656 * confusing. This is much easier. 3657 */ 3658 3659 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3660 bp->b_flags |= B_NOCACHE; 3661 bwrite(bp); 3662 goto loop; 3663 } 3664 bp->b_flags &= ~B_DONE; 3665 } else { 3666 /* 3667 * Buffer is not in-core, create new buffer. The buffer 3668 * returned by getnewbuf() is locked. Note that the returned 3669 * buffer is also considered valid (not marked B_INVAL). 3670 */ 3671 BO_RUNLOCK(bo); 3672 /* 3673 * If the user does not want us to create the buffer, bail out 3674 * here. 3675 */ 3676 if (flags & GB_NOCREAT) 3677 return NULL; 3678 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3679 return NULL; 3680 3681 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3682 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3683 offset = blkno * bsize; 3684 vmio = vp->v_object != NULL; 3685 if (vmio) { 3686 maxsize = size + (offset & PAGE_MASK); 3687 } else { 3688 maxsize = size; 3689 /* Do not allow non-VMIO notmapped buffers. */ 3690 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3691 } 3692 maxsize = imax(maxsize, bsize); 3693 3694 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 3695 if (bp == NULL) { 3696 if (slpflag || slptimeo) 3697 return NULL; 3698 /* 3699 * XXX This is here until the sleep path is diagnosed 3700 * enough to work under very low memory conditions. 3701 * 3702 * There's an issue on low memory, 4BSD+non-preempt 3703 * systems (eg MIPS routers with 32MB RAM) where buffer 3704 * exhaustion occurs without sleeping for buffer 3705 * reclaimation. This just sticks in a loop and 3706 * constantly attempts to allocate a buffer, which 3707 * hits exhaustion and tries to wakeup bufdaemon. 3708 * This never happens because we never yield. 3709 * 3710 * The real solution is to identify and fix these cases 3711 * so we aren't effectively busy-waiting in a loop 3712 * until the reclaimation path has cycles to run. 3713 */ 3714 kern_yield(PRI_USER); 3715 goto loop; 3716 } 3717 3718 /* 3719 * This code is used to make sure that a buffer is not 3720 * created while the getnewbuf routine is blocked. 3721 * This can be a problem whether the vnode is locked or not. 3722 * If the buffer is created out from under us, we have to 3723 * throw away the one we just created. 3724 * 3725 * Note: this must occur before we associate the buffer 3726 * with the vp especially considering limitations in 3727 * the splay tree implementation when dealing with duplicate 3728 * lblkno's. 3729 */ 3730 BO_LOCK(bo); 3731 if (gbincore(bo, blkno)) { 3732 BO_UNLOCK(bo); 3733 bp->b_flags |= B_INVAL; 3734 brelse(bp); 3735 bufspace_release(maxsize); 3736 goto loop; 3737 } 3738 3739 /* 3740 * Insert the buffer into the hash, so that it can 3741 * be found by incore. 3742 */ 3743 bp->b_blkno = bp->b_lblkno = blkno; 3744 bp->b_offset = offset; 3745 bgetvp(vp, bp); 3746 BO_UNLOCK(bo); 3747 3748 /* 3749 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3750 * buffer size starts out as 0, B_CACHE will be set by 3751 * allocbuf() for the VMIO case prior to it testing the 3752 * backing store for validity. 3753 */ 3754 3755 if (vmio) { 3756 bp->b_flags |= B_VMIO; 3757 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3758 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3759 bp, vp->v_object, bp->b_bufobj->bo_object)); 3760 } else { 3761 bp->b_flags &= ~B_VMIO; 3762 KASSERT(bp->b_bufobj->bo_object == NULL, 3763 ("ARGH! has b_bufobj->bo_object %p %p\n", 3764 bp, bp->b_bufobj->bo_object)); 3765 BUF_CHECK_MAPPED(bp); 3766 } 3767 3768 allocbuf(bp, size); 3769 bufspace_release(maxsize); 3770 bp->b_flags &= ~B_DONE; 3771 } 3772 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3773 BUF_ASSERT_HELD(bp); 3774 end: 3775 buf_track(bp, __func__); 3776 KASSERT(bp->b_bufobj == bo, 3777 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3778 return (bp); 3779 } 3780 3781 /* 3782 * Get an empty, disassociated buffer of given size. The buffer is initially 3783 * set to B_INVAL. 3784 */ 3785 struct buf * 3786 geteblk(int size, int flags) 3787 { 3788 struct buf *bp; 3789 int maxsize; 3790 3791 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3792 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 3793 if ((flags & GB_NOWAIT_BD) && 3794 (curthread->td_pflags & TDP_BUFNEED) != 0) 3795 return (NULL); 3796 } 3797 allocbuf(bp, size); 3798 bufspace_release(maxsize); 3799 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3800 BUF_ASSERT_HELD(bp); 3801 return (bp); 3802 } 3803 3804 /* 3805 * Truncate the backing store for a non-vmio buffer. 3806 */ 3807 static void 3808 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 3809 { 3810 3811 if (bp->b_flags & B_MALLOC) { 3812 /* 3813 * malloced buffers are not shrunk 3814 */ 3815 if (newbsize == 0) { 3816 bufmallocadjust(bp, 0); 3817 free(bp->b_data, M_BIOBUF); 3818 bp->b_data = bp->b_kvabase; 3819 bp->b_flags &= ~B_MALLOC; 3820 } 3821 return; 3822 } 3823 vm_hold_free_pages(bp, newbsize); 3824 bufspace_adjust(bp, newbsize); 3825 } 3826 3827 /* 3828 * Extend the backing for a non-VMIO buffer. 3829 */ 3830 static void 3831 vfs_nonvmio_extend(struct buf *bp, int newbsize) 3832 { 3833 caddr_t origbuf; 3834 int origbufsize; 3835 3836 /* 3837 * We only use malloced memory on the first allocation. 3838 * and revert to page-allocated memory when the buffer 3839 * grows. 3840 * 3841 * There is a potential smp race here that could lead 3842 * to bufmallocspace slightly passing the max. It 3843 * is probably extremely rare and not worth worrying 3844 * over. 3845 */ 3846 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 3847 bufmallocspace < maxbufmallocspace) { 3848 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 3849 bp->b_flags |= B_MALLOC; 3850 bufmallocadjust(bp, newbsize); 3851 return; 3852 } 3853 3854 /* 3855 * If the buffer is growing on its other-than-first 3856 * allocation then we revert to the page-allocation 3857 * scheme. 3858 */ 3859 origbuf = NULL; 3860 origbufsize = 0; 3861 if (bp->b_flags & B_MALLOC) { 3862 origbuf = bp->b_data; 3863 origbufsize = bp->b_bufsize; 3864 bp->b_data = bp->b_kvabase; 3865 bufmallocadjust(bp, 0); 3866 bp->b_flags &= ~B_MALLOC; 3867 newbsize = round_page(newbsize); 3868 } 3869 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 3870 (vm_offset_t) bp->b_data + newbsize); 3871 if (origbuf != NULL) { 3872 bcopy(origbuf, bp->b_data, origbufsize); 3873 free(origbuf, M_BIOBUF); 3874 } 3875 bufspace_adjust(bp, newbsize); 3876 } 3877 3878 /* 3879 * This code constitutes the buffer memory from either anonymous system 3880 * memory (in the case of non-VMIO operations) or from an associated 3881 * VM object (in the case of VMIO operations). This code is able to 3882 * resize a buffer up or down. 3883 * 3884 * Note that this code is tricky, and has many complications to resolve 3885 * deadlock or inconsistent data situations. Tread lightly!!! 3886 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3887 * the caller. Calling this code willy nilly can result in the loss of data. 3888 * 3889 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3890 * B_CACHE for the non-VMIO case. 3891 */ 3892 int 3893 allocbuf(struct buf *bp, int size) 3894 { 3895 int newbsize; 3896 3897 BUF_ASSERT_HELD(bp); 3898 3899 if (bp->b_bcount == size) 3900 return (1); 3901 3902 if (bp->b_kvasize != 0 && bp->b_kvasize < size) 3903 panic("allocbuf: buffer too small"); 3904 3905 newbsize = roundup2(size, DEV_BSIZE); 3906 if ((bp->b_flags & B_VMIO) == 0) { 3907 if ((bp->b_flags & B_MALLOC) == 0) 3908 newbsize = round_page(newbsize); 3909 /* 3910 * Just get anonymous memory from the kernel. Don't 3911 * mess with B_CACHE. 3912 */ 3913 if (newbsize < bp->b_bufsize) 3914 vfs_nonvmio_truncate(bp, newbsize); 3915 else if (newbsize > bp->b_bufsize) 3916 vfs_nonvmio_extend(bp, newbsize); 3917 } else { 3918 int desiredpages; 3919 3920 desiredpages = (size == 0) ? 0 : 3921 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3922 3923 if (bp->b_flags & B_MALLOC) 3924 panic("allocbuf: VMIO buffer can't be malloced"); 3925 /* 3926 * Set B_CACHE initially if buffer is 0 length or will become 3927 * 0-length. 3928 */ 3929 if (size == 0 || bp->b_bufsize == 0) 3930 bp->b_flags |= B_CACHE; 3931 3932 if (newbsize < bp->b_bufsize) 3933 vfs_vmio_truncate(bp, desiredpages); 3934 /* XXX This looks as if it should be newbsize > b_bufsize */ 3935 else if (size > bp->b_bcount) 3936 vfs_vmio_extend(bp, desiredpages, size); 3937 bufspace_adjust(bp, newbsize); 3938 } 3939 bp->b_bcount = size; /* requested buffer size. */ 3940 return (1); 3941 } 3942 3943 extern int inflight_transient_maps; 3944 3945 void 3946 biodone(struct bio *bp) 3947 { 3948 struct mtx *mtxp; 3949 void (*done)(struct bio *); 3950 vm_offset_t start, end; 3951 3952 biotrack(bp, __func__); 3953 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3954 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3955 bp->bio_flags |= BIO_UNMAPPED; 3956 start = trunc_page((vm_offset_t)bp->bio_data); 3957 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3958 bp->bio_data = unmapped_buf; 3959 pmap_qremove(start, atop(end - start)); 3960 vmem_free(transient_arena, start, end - start); 3961 atomic_add_int(&inflight_transient_maps, -1); 3962 } 3963 done = bp->bio_done; 3964 if (done == NULL) { 3965 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3966 mtx_lock(mtxp); 3967 bp->bio_flags |= BIO_DONE; 3968 wakeup(bp); 3969 mtx_unlock(mtxp); 3970 } else 3971 done(bp); 3972 } 3973 3974 /* 3975 * Wait for a BIO to finish. 3976 */ 3977 int 3978 biowait(struct bio *bp, const char *wchan) 3979 { 3980 struct mtx *mtxp; 3981 3982 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3983 mtx_lock(mtxp); 3984 while ((bp->bio_flags & BIO_DONE) == 0) 3985 msleep(bp, mtxp, PRIBIO, wchan, 0); 3986 mtx_unlock(mtxp); 3987 if (bp->bio_error != 0) 3988 return (bp->bio_error); 3989 if (!(bp->bio_flags & BIO_ERROR)) 3990 return (0); 3991 return (EIO); 3992 } 3993 3994 void 3995 biofinish(struct bio *bp, struct devstat *stat, int error) 3996 { 3997 3998 if (error) { 3999 bp->bio_error = error; 4000 bp->bio_flags |= BIO_ERROR; 4001 } 4002 if (stat != NULL) 4003 devstat_end_transaction_bio(stat, bp); 4004 biodone(bp); 4005 } 4006 4007 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4008 void 4009 biotrack_buf(struct bio *bp, const char *location) 4010 { 4011 4012 buf_track(bp->bio_track_bp, location); 4013 } 4014 #endif 4015 4016 /* 4017 * bufwait: 4018 * 4019 * Wait for buffer I/O completion, returning error status. The buffer 4020 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4021 * error and cleared. 4022 */ 4023 int 4024 bufwait(struct buf *bp) 4025 { 4026 if (bp->b_iocmd == BIO_READ) 4027 bwait(bp, PRIBIO, "biord"); 4028 else 4029 bwait(bp, PRIBIO, "biowr"); 4030 if (bp->b_flags & B_EINTR) { 4031 bp->b_flags &= ~B_EINTR; 4032 return (EINTR); 4033 } 4034 if (bp->b_ioflags & BIO_ERROR) { 4035 return (bp->b_error ? bp->b_error : EIO); 4036 } else { 4037 return (0); 4038 } 4039 } 4040 4041 /* 4042 * bufdone: 4043 * 4044 * Finish I/O on a buffer, optionally calling a completion function. 4045 * This is usually called from an interrupt so process blocking is 4046 * not allowed. 4047 * 4048 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4049 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4050 * assuming B_INVAL is clear. 4051 * 4052 * For the VMIO case, we set B_CACHE if the op was a read and no 4053 * read error occurred, or if the op was a write. B_CACHE is never 4054 * set if the buffer is invalid or otherwise uncacheable. 4055 * 4056 * biodone does not mess with B_INVAL, allowing the I/O routine or the 4057 * initiator to leave B_INVAL set to brelse the buffer out of existence 4058 * in the biodone routine. 4059 */ 4060 void 4061 bufdone(struct buf *bp) 4062 { 4063 struct bufobj *dropobj; 4064 void (*biodone)(struct buf *); 4065 4066 buf_track(bp, __func__); 4067 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4068 dropobj = NULL; 4069 4070 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4071 BUF_ASSERT_HELD(bp); 4072 4073 runningbufwakeup(bp); 4074 if (bp->b_iocmd == BIO_WRITE) 4075 dropobj = bp->b_bufobj; 4076 else if ((bp->b_flags & B_CKHASH) != 0) { 4077 KASSERT(buf_mapped(bp), ("biodone: bp %p not mapped", bp)); 4078 (*bp->b_ckhashcalc)(bp); 4079 } 4080 /* call optional completion function if requested */ 4081 if (bp->b_iodone != NULL) { 4082 biodone = bp->b_iodone; 4083 bp->b_iodone = NULL; 4084 (*biodone) (bp); 4085 if (dropobj) 4086 bufobj_wdrop(dropobj); 4087 return; 4088 } 4089 4090 bufdone_finish(bp); 4091 4092 if (dropobj) 4093 bufobj_wdrop(dropobj); 4094 } 4095 4096 void 4097 bufdone_finish(struct buf *bp) 4098 { 4099 BUF_ASSERT_HELD(bp); 4100 4101 if (!LIST_EMPTY(&bp->b_dep)) 4102 buf_complete(bp); 4103 4104 if (bp->b_flags & B_VMIO) { 4105 /* 4106 * Set B_CACHE if the op was a normal read and no error 4107 * occurred. B_CACHE is set for writes in the b*write() 4108 * routines. 4109 */ 4110 if (bp->b_iocmd == BIO_READ && 4111 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4112 !(bp->b_ioflags & BIO_ERROR)) 4113 bp->b_flags |= B_CACHE; 4114 vfs_vmio_iodone(bp); 4115 } 4116 4117 /* 4118 * For asynchronous completions, release the buffer now. The brelse 4119 * will do a wakeup there if necessary - so no need to do a wakeup 4120 * here in the async case. The sync case always needs to do a wakeup. 4121 */ 4122 if (bp->b_flags & B_ASYNC) { 4123 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4124 (bp->b_ioflags & BIO_ERROR)) 4125 brelse(bp); 4126 else 4127 bqrelse(bp); 4128 } else 4129 bdone(bp); 4130 } 4131 4132 /* 4133 * This routine is called in lieu of iodone in the case of 4134 * incomplete I/O. This keeps the busy status for pages 4135 * consistent. 4136 */ 4137 void 4138 vfs_unbusy_pages(struct buf *bp) 4139 { 4140 int i; 4141 vm_object_t obj; 4142 vm_page_t m; 4143 4144 runningbufwakeup(bp); 4145 if (!(bp->b_flags & B_VMIO)) 4146 return; 4147 4148 obj = bp->b_bufobj->bo_object; 4149 VM_OBJECT_WLOCK(obj); 4150 for (i = 0; i < bp->b_npages; i++) { 4151 m = bp->b_pages[i]; 4152 if (m == bogus_page) { 4153 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4154 if (!m) 4155 panic("vfs_unbusy_pages: page missing\n"); 4156 bp->b_pages[i] = m; 4157 if (buf_mapped(bp)) { 4158 BUF_CHECK_MAPPED(bp); 4159 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4160 bp->b_pages, bp->b_npages); 4161 } else 4162 BUF_CHECK_UNMAPPED(bp); 4163 } 4164 vm_page_sunbusy(m); 4165 } 4166 vm_object_pip_wakeupn(obj, bp->b_npages); 4167 VM_OBJECT_WUNLOCK(obj); 4168 } 4169 4170 /* 4171 * vfs_page_set_valid: 4172 * 4173 * Set the valid bits in a page based on the supplied offset. The 4174 * range is restricted to the buffer's size. 4175 * 4176 * This routine is typically called after a read completes. 4177 */ 4178 static void 4179 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4180 { 4181 vm_ooffset_t eoff; 4182 4183 /* 4184 * Compute the end offset, eoff, such that [off, eoff) does not span a 4185 * page boundary and eoff is not greater than the end of the buffer. 4186 * The end of the buffer, in this case, is our file EOF, not the 4187 * allocation size of the buffer. 4188 */ 4189 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4190 if (eoff > bp->b_offset + bp->b_bcount) 4191 eoff = bp->b_offset + bp->b_bcount; 4192 4193 /* 4194 * Set valid range. This is typically the entire buffer and thus the 4195 * entire page. 4196 */ 4197 if (eoff > off) 4198 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4199 } 4200 4201 /* 4202 * vfs_page_set_validclean: 4203 * 4204 * Set the valid bits and clear the dirty bits in a page based on the 4205 * supplied offset. The range is restricted to the buffer's size. 4206 */ 4207 static void 4208 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4209 { 4210 vm_ooffset_t soff, eoff; 4211 4212 /* 4213 * Start and end offsets in buffer. eoff - soff may not cross a 4214 * page boundary or cross the end of the buffer. The end of the 4215 * buffer, in this case, is our file EOF, not the allocation size 4216 * of the buffer. 4217 */ 4218 soff = off; 4219 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4220 if (eoff > bp->b_offset + bp->b_bcount) 4221 eoff = bp->b_offset + bp->b_bcount; 4222 4223 /* 4224 * Set valid range. This is typically the entire buffer and thus the 4225 * entire page. 4226 */ 4227 if (eoff > soff) { 4228 vm_page_set_validclean( 4229 m, 4230 (vm_offset_t) (soff & PAGE_MASK), 4231 (vm_offset_t) (eoff - soff) 4232 ); 4233 } 4234 } 4235 4236 /* 4237 * Ensure that all buffer pages are not exclusive busied. If any page is 4238 * exclusive busy, drain it. 4239 */ 4240 void 4241 vfs_drain_busy_pages(struct buf *bp) 4242 { 4243 vm_page_t m; 4244 int i, last_busied; 4245 4246 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4247 last_busied = 0; 4248 for (i = 0; i < bp->b_npages; i++) { 4249 m = bp->b_pages[i]; 4250 if (vm_page_xbusied(m)) { 4251 for (; last_busied < i; last_busied++) 4252 vm_page_sbusy(bp->b_pages[last_busied]); 4253 while (vm_page_xbusied(m)) { 4254 vm_page_lock(m); 4255 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4256 vm_page_busy_sleep(m, "vbpage", true); 4257 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4258 } 4259 } 4260 } 4261 for (i = 0; i < last_busied; i++) 4262 vm_page_sunbusy(bp->b_pages[i]); 4263 } 4264 4265 /* 4266 * This routine is called before a device strategy routine. 4267 * It is used to tell the VM system that paging I/O is in 4268 * progress, and treat the pages associated with the buffer 4269 * almost as being exclusive busy. Also the object paging_in_progress 4270 * flag is handled to make sure that the object doesn't become 4271 * inconsistent. 4272 * 4273 * Since I/O has not been initiated yet, certain buffer flags 4274 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4275 * and should be ignored. 4276 */ 4277 void 4278 vfs_busy_pages(struct buf *bp, int clear_modify) 4279 { 4280 vm_object_t obj; 4281 vm_ooffset_t foff; 4282 vm_page_t m; 4283 int i; 4284 bool bogus; 4285 4286 if (!(bp->b_flags & B_VMIO)) 4287 return; 4288 4289 obj = bp->b_bufobj->bo_object; 4290 foff = bp->b_offset; 4291 KASSERT(bp->b_offset != NOOFFSET, 4292 ("vfs_busy_pages: no buffer offset")); 4293 VM_OBJECT_WLOCK(obj); 4294 vfs_drain_busy_pages(bp); 4295 if (bp->b_bufsize != 0) 4296 vfs_setdirty_locked_object(bp); 4297 bogus = false; 4298 for (i = 0; i < bp->b_npages; i++) { 4299 m = bp->b_pages[i]; 4300 4301 if ((bp->b_flags & B_CLUSTER) == 0) { 4302 vm_object_pip_add(obj, 1); 4303 vm_page_sbusy(m); 4304 } 4305 /* 4306 * When readying a buffer for a read ( i.e 4307 * clear_modify == 0 ), it is important to do 4308 * bogus_page replacement for valid pages in 4309 * partially instantiated buffers. Partially 4310 * instantiated buffers can, in turn, occur when 4311 * reconstituting a buffer from its VM backing store 4312 * base. We only have to do this if B_CACHE is 4313 * clear ( which causes the I/O to occur in the 4314 * first place ). The replacement prevents the read 4315 * I/O from overwriting potentially dirty VM-backed 4316 * pages. XXX bogus page replacement is, uh, bogus. 4317 * It may not work properly with small-block devices. 4318 * We need to find a better way. 4319 */ 4320 if (clear_modify) { 4321 pmap_remove_write(m); 4322 vfs_page_set_validclean(bp, foff, m); 4323 } else if (m->valid == VM_PAGE_BITS_ALL && 4324 (bp->b_flags & B_CACHE) == 0) { 4325 bp->b_pages[i] = bogus_page; 4326 bogus = true; 4327 } 4328 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4329 } 4330 VM_OBJECT_WUNLOCK(obj); 4331 if (bogus && buf_mapped(bp)) { 4332 BUF_CHECK_MAPPED(bp); 4333 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4334 bp->b_pages, bp->b_npages); 4335 } 4336 } 4337 4338 /* 4339 * vfs_bio_set_valid: 4340 * 4341 * Set the range within the buffer to valid. The range is 4342 * relative to the beginning of the buffer, b_offset. Note that 4343 * b_offset itself may be offset from the beginning of the first 4344 * page. 4345 */ 4346 void 4347 vfs_bio_set_valid(struct buf *bp, int base, int size) 4348 { 4349 int i, n; 4350 vm_page_t m; 4351 4352 if (!(bp->b_flags & B_VMIO)) 4353 return; 4354 4355 /* 4356 * Fixup base to be relative to beginning of first page. 4357 * Set initial n to be the maximum number of bytes in the 4358 * first page that can be validated. 4359 */ 4360 base += (bp->b_offset & PAGE_MASK); 4361 n = PAGE_SIZE - (base & PAGE_MASK); 4362 4363 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4364 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4365 m = bp->b_pages[i]; 4366 if (n > size) 4367 n = size; 4368 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4369 base += n; 4370 size -= n; 4371 n = PAGE_SIZE; 4372 } 4373 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4374 } 4375 4376 /* 4377 * vfs_bio_clrbuf: 4378 * 4379 * If the specified buffer is a non-VMIO buffer, clear the entire 4380 * buffer. If the specified buffer is a VMIO buffer, clear and 4381 * validate only the previously invalid portions of the buffer. 4382 * This routine essentially fakes an I/O, so we need to clear 4383 * BIO_ERROR and B_INVAL. 4384 * 4385 * Note that while we only theoretically need to clear through b_bcount, 4386 * we go ahead and clear through b_bufsize. 4387 */ 4388 void 4389 vfs_bio_clrbuf(struct buf *bp) 4390 { 4391 int i, j, mask, sa, ea, slide; 4392 4393 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4394 clrbuf(bp); 4395 return; 4396 } 4397 bp->b_flags &= ~B_INVAL; 4398 bp->b_ioflags &= ~BIO_ERROR; 4399 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4400 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4401 (bp->b_offset & PAGE_MASK) == 0) { 4402 if (bp->b_pages[0] == bogus_page) 4403 goto unlock; 4404 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4405 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4406 if ((bp->b_pages[0]->valid & mask) == mask) 4407 goto unlock; 4408 if ((bp->b_pages[0]->valid & mask) == 0) { 4409 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4410 bp->b_pages[0]->valid |= mask; 4411 goto unlock; 4412 } 4413 } 4414 sa = bp->b_offset & PAGE_MASK; 4415 slide = 0; 4416 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4417 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4418 ea = slide & PAGE_MASK; 4419 if (ea == 0) 4420 ea = PAGE_SIZE; 4421 if (bp->b_pages[i] == bogus_page) 4422 continue; 4423 j = sa / DEV_BSIZE; 4424 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4425 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4426 if ((bp->b_pages[i]->valid & mask) == mask) 4427 continue; 4428 if ((bp->b_pages[i]->valid & mask) == 0) 4429 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4430 else { 4431 for (; sa < ea; sa += DEV_BSIZE, j++) { 4432 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4433 pmap_zero_page_area(bp->b_pages[i], 4434 sa, DEV_BSIZE); 4435 } 4436 } 4437 } 4438 bp->b_pages[i]->valid |= mask; 4439 } 4440 unlock: 4441 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4442 bp->b_resid = 0; 4443 } 4444 4445 void 4446 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4447 { 4448 vm_page_t m; 4449 int i, n; 4450 4451 if (buf_mapped(bp)) { 4452 BUF_CHECK_MAPPED(bp); 4453 bzero(bp->b_data + base, size); 4454 } else { 4455 BUF_CHECK_UNMAPPED(bp); 4456 n = PAGE_SIZE - (base & PAGE_MASK); 4457 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4458 m = bp->b_pages[i]; 4459 if (n > size) 4460 n = size; 4461 pmap_zero_page_area(m, base & PAGE_MASK, n); 4462 base += n; 4463 size -= n; 4464 n = PAGE_SIZE; 4465 } 4466 } 4467 } 4468 4469 /* 4470 * Update buffer flags based on I/O request parameters, optionally releasing the 4471 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4472 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4473 * I/O). Otherwise the buffer is released to the cache. 4474 */ 4475 static void 4476 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4477 { 4478 4479 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4480 ("buf %p non-VMIO noreuse", bp)); 4481 4482 if ((ioflag & IO_DIRECT) != 0) 4483 bp->b_flags |= B_DIRECT; 4484 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4485 bp->b_flags |= B_RELBUF; 4486 if ((ioflag & IO_NOREUSE) != 0) 4487 bp->b_flags |= B_NOREUSE; 4488 if (release) 4489 brelse(bp); 4490 } else if (release) 4491 bqrelse(bp); 4492 } 4493 4494 void 4495 vfs_bio_brelse(struct buf *bp, int ioflag) 4496 { 4497 4498 b_io_dismiss(bp, ioflag, true); 4499 } 4500 4501 void 4502 vfs_bio_set_flags(struct buf *bp, int ioflag) 4503 { 4504 4505 b_io_dismiss(bp, ioflag, false); 4506 } 4507 4508 /* 4509 * vm_hold_load_pages and vm_hold_free_pages get pages into 4510 * a buffers address space. The pages are anonymous and are 4511 * not associated with a file object. 4512 */ 4513 static void 4514 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4515 { 4516 vm_offset_t pg; 4517 vm_page_t p; 4518 int index; 4519 4520 BUF_CHECK_MAPPED(bp); 4521 4522 to = round_page(to); 4523 from = round_page(from); 4524 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4525 4526 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4527 /* 4528 * note: must allocate system pages since blocking here 4529 * could interfere with paging I/O, no matter which 4530 * process we are. 4531 */ 4532 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4533 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | 4534 VM_ALLOC_WAITOK); 4535 pmap_qenter(pg, &p, 1); 4536 bp->b_pages[index] = p; 4537 } 4538 bp->b_npages = index; 4539 } 4540 4541 /* Return pages associated with this buf to the vm system */ 4542 static void 4543 vm_hold_free_pages(struct buf *bp, int newbsize) 4544 { 4545 vm_offset_t from; 4546 vm_page_t p; 4547 int index, newnpages; 4548 4549 BUF_CHECK_MAPPED(bp); 4550 4551 from = round_page((vm_offset_t)bp->b_data + newbsize); 4552 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4553 if (bp->b_npages > newnpages) 4554 pmap_qremove(from, bp->b_npages - newnpages); 4555 for (index = newnpages; index < bp->b_npages; index++) { 4556 p = bp->b_pages[index]; 4557 bp->b_pages[index] = NULL; 4558 p->wire_count--; 4559 vm_page_free(p); 4560 } 4561 atomic_subtract_int(&vm_cnt.v_wire_count, bp->b_npages - newnpages); 4562 bp->b_npages = newnpages; 4563 } 4564 4565 /* 4566 * Map an IO request into kernel virtual address space. 4567 * 4568 * All requests are (re)mapped into kernel VA space. 4569 * Notice that we use b_bufsize for the size of the buffer 4570 * to be mapped. b_bcount might be modified by the driver. 4571 * 4572 * Note that even if the caller determines that the address space should 4573 * be valid, a race or a smaller-file mapped into a larger space may 4574 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4575 * check the return value. 4576 * 4577 * This function only works with pager buffers. 4578 */ 4579 int 4580 vmapbuf(struct buf *bp, int mapbuf) 4581 { 4582 vm_prot_t prot; 4583 int pidx; 4584 4585 if (bp->b_bufsize < 0) 4586 return (-1); 4587 prot = VM_PROT_READ; 4588 if (bp->b_iocmd == BIO_READ) 4589 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4590 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4591 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4592 btoc(MAXPHYS))) < 0) 4593 return (-1); 4594 bp->b_npages = pidx; 4595 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4596 if (mapbuf || !unmapped_buf_allowed) { 4597 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 4598 bp->b_data = bp->b_kvabase + bp->b_offset; 4599 } else 4600 bp->b_data = unmapped_buf; 4601 return(0); 4602 } 4603 4604 /* 4605 * Free the io map PTEs associated with this IO operation. 4606 * We also invalidate the TLB entries and restore the original b_addr. 4607 * 4608 * This function only works with pager buffers. 4609 */ 4610 void 4611 vunmapbuf(struct buf *bp) 4612 { 4613 int npages; 4614 4615 npages = bp->b_npages; 4616 if (buf_mapped(bp)) 4617 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4618 vm_page_unhold_pages(bp->b_pages, npages); 4619 4620 bp->b_data = unmapped_buf; 4621 } 4622 4623 void 4624 bdone(struct buf *bp) 4625 { 4626 struct mtx *mtxp; 4627 4628 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4629 mtx_lock(mtxp); 4630 bp->b_flags |= B_DONE; 4631 wakeup(bp); 4632 mtx_unlock(mtxp); 4633 } 4634 4635 void 4636 bwait(struct buf *bp, u_char pri, const char *wchan) 4637 { 4638 struct mtx *mtxp; 4639 4640 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4641 mtx_lock(mtxp); 4642 while ((bp->b_flags & B_DONE) == 0) 4643 msleep(bp, mtxp, pri, wchan, 0); 4644 mtx_unlock(mtxp); 4645 } 4646 4647 int 4648 bufsync(struct bufobj *bo, int waitfor) 4649 { 4650 4651 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 4652 } 4653 4654 void 4655 bufstrategy(struct bufobj *bo, struct buf *bp) 4656 { 4657 int i = 0; 4658 struct vnode *vp; 4659 4660 vp = bp->b_vp; 4661 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4662 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4663 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4664 i = VOP_STRATEGY(vp, bp); 4665 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4666 } 4667 4668 void 4669 bufobj_wrefl(struct bufobj *bo) 4670 { 4671 4672 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4673 ASSERT_BO_WLOCKED(bo); 4674 bo->bo_numoutput++; 4675 } 4676 4677 void 4678 bufobj_wref(struct bufobj *bo) 4679 { 4680 4681 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4682 BO_LOCK(bo); 4683 bo->bo_numoutput++; 4684 BO_UNLOCK(bo); 4685 } 4686 4687 void 4688 bufobj_wdrop(struct bufobj *bo) 4689 { 4690 4691 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4692 BO_LOCK(bo); 4693 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4694 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4695 bo->bo_flag &= ~BO_WWAIT; 4696 wakeup(&bo->bo_numoutput); 4697 } 4698 BO_UNLOCK(bo); 4699 } 4700 4701 int 4702 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4703 { 4704 int error; 4705 4706 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4707 ASSERT_BO_WLOCKED(bo); 4708 error = 0; 4709 while (bo->bo_numoutput) { 4710 bo->bo_flag |= BO_WWAIT; 4711 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4712 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4713 if (error) 4714 break; 4715 } 4716 return (error); 4717 } 4718 4719 /* 4720 * Set bio_data or bio_ma for struct bio from the struct buf. 4721 */ 4722 void 4723 bdata2bio(struct buf *bp, struct bio *bip) 4724 { 4725 4726 if (!buf_mapped(bp)) { 4727 KASSERT(unmapped_buf_allowed, ("unmapped")); 4728 bip->bio_ma = bp->b_pages; 4729 bip->bio_ma_n = bp->b_npages; 4730 bip->bio_data = unmapped_buf; 4731 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4732 bip->bio_flags |= BIO_UNMAPPED; 4733 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4734 PAGE_SIZE == bp->b_npages, 4735 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4736 (long long)bip->bio_length, bip->bio_ma_n)); 4737 } else { 4738 bip->bio_data = bp->b_data; 4739 bip->bio_ma = NULL; 4740 } 4741 } 4742 4743 /* 4744 * The MIPS pmap code currently doesn't handle aliased pages. 4745 * The VIPT caches may not handle page aliasing themselves, leading 4746 * to data corruption. 4747 * 4748 * As such, this code makes a system extremely unhappy if said 4749 * system doesn't support unaliasing the above situation in hardware. 4750 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable 4751 * this feature at build time, so it has to be handled in software. 4752 * 4753 * Once the MIPS pmap/cache code grows to support this function on 4754 * earlier chips, it should be flipped back off. 4755 */ 4756 #ifdef __mips__ 4757 static int buf_pager_relbuf = 1; 4758 #else 4759 static int buf_pager_relbuf = 0; 4760 #endif 4761 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 4762 &buf_pager_relbuf, 0, 4763 "Make buffer pager release buffers after reading"); 4764 4765 /* 4766 * The buffer pager. It uses buffer reads to validate pages. 4767 * 4768 * In contrast to the generic local pager from vm/vnode_pager.c, this 4769 * pager correctly and easily handles volumes where the underlying 4770 * device block size is greater than the machine page size. The 4771 * buffer cache transparently extends the requested page run to be 4772 * aligned at the block boundary, and does the necessary bogus page 4773 * replacements in the addends to avoid obliterating already valid 4774 * pages. 4775 * 4776 * The only non-trivial issue is that the exclusive busy state for 4777 * pages, which is assumed by the vm_pager_getpages() interface, is 4778 * incompatible with the VMIO buffer cache's desire to share-busy the 4779 * pages. This function performs a trivial downgrade of the pages' 4780 * state before reading buffers, and a less trivial upgrade from the 4781 * shared-busy to excl-busy state after the read. 4782 */ 4783 int 4784 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 4785 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 4786 vbg_get_blksize_t get_blksize) 4787 { 4788 vm_page_t m; 4789 vm_object_t object; 4790 struct buf *bp; 4791 struct mount *mp; 4792 daddr_t lbn, lbnp; 4793 vm_ooffset_t la, lb, poff, poffe; 4794 long bsize; 4795 int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b; 4796 bool redo, lpart; 4797 4798 object = vp->v_object; 4799 mp = vp->v_mount; 4800 la = IDX_TO_OFF(ma[count - 1]->pindex); 4801 if (la >= object->un_pager.vnp.vnp_size) 4802 return (VM_PAGER_BAD); 4803 lpart = la + PAGE_SIZE > object->un_pager.vnp.vnp_size; 4804 bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex))); 4805 4806 /* 4807 * Calculate read-ahead, behind and total pages. 4808 */ 4809 pgsin = count; 4810 lb = IDX_TO_OFF(ma[0]->pindex); 4811 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 4812 pgsin += pgsin_b; 4813 if (rbehind != NULL) 4814 *rbehind = pgsin_b; 4815 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 4816 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 4817 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 4818 PAGE_SIZE) - la); 4819 pgsin += pgsin_a; 4820 if (rahead != NULL) 4821 *rahead = pgsin_a; 4822 VM_CNT_INC(v_vnodein); 4823 VM_CNT_ADD(v_vnodepgsin, pgsin); 4824 4825 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 4826 != 0) ? GB_UNMAPPED : 0; 4827 VM_OBJECT_WLOCK(object); 4828 again: 4829 for (i = 0; i < count; i++) 4830 vm_page_busy_downgrade(ma[i]); 4831 VM_OBJECT_WUNLOCK(object); 4832 4833 lbnp = -1; 4834 for (i = 0; i < count; i++) { 4835 m = ma[i]; 4836 4837 /* 4838 * Pages are shared busy and the object lock is not 4839 * owned, which together allow for the pages' 4840 * invalidation. The racy test for validity avoids 4841 * useless creation of the buffer for the most typical 4842 * case when invalidation is not used in redo or for 4843 * parallel read. The shared->excl upgrade loop at 4844 * the end of the function catches the race in a 4845 * reliable way (protected by the object lock). 4846 */ 4847 if (m->valid == VM_PAGE_BITS_ALL) 4848 continue; 4849 4850 poff = IDX_TO_OFF(m->pindex); 4851 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 4852 for (; poff < poffe; poff += bsize) { 4853 lbn = get_lblkno(vp, poff); 4854 if (lbn == lbnp) 4855 goto next_page; 4856 lbnp = lbn; 4857 4858 bsize = get_blksize(vp, lbn); 4859 error = bread_gb(vp, lbn, bsize, curthread->td_ucred, 4860 br_flags, &bp); 4861 if (error != 0) 4862 goto end_pages; 4863 if (LIST_EMPTY(&bp->b_dep)) { 4864 /* 4865 * Invalidation clears m->valid, but 4866 * may leave B_CACHE flag if the 4867 * buffer existed at the invalidation 4868 * time. In this case, recycle the 4869 * buffer to do real read on next 4870 * bread() after redo. 4871 * 4872 * Otherwise B_RELBUF is not strictly 4873 * necessary, enable to reduce buf 4874 * cache pressure. 4875 */ 4876 if (buf_pager_relbuf || 4877 m->valid != VM_PAGE_BITS_ALL) 4878 bp->b_flags |= B_RELBUF; 4879 4880 bp->b_flags &= ~B_NOCACHE; 4881 brelse(bp); 4882 } else { 4883 bqrelse(bp); 4884 } 4885 } 4886 KASSERT(1 /* racy, enable for debugging */ || 4887 m->valid == VM_PAGE_BITS_ALL || i == count - 1, 4888 ("buf %d %p invalid", i, m)); 4889 if (i == count - 1 && lpart) { 4890 VM_OBJECT_WLOCK(object); 4891 if (m->valid != 0 && 4892 m->valid != VM_PAGE_BITS_ALL) 4893 vm_page_zero_invalid(m, TRUE); 4894 VM_OBJECT_WUNLOCK(object); 4895 } 4896 next_page:; 4897 } 4898 end_pages: 4899 4900 VM_OBJECT_WLOCK(object); 4901 redo = false; 4902 for (i = 0; i < count; i++) { 4903 vm_page_sunbusy(ma[i]); 4904 ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL); 4905 4906 /* 4907 * Since the pages were only sbusy while neither the 4908 * buffer nor the object lock was held by us, or 4909 * reallocated while vm_page_grab() slept for busy 4910 * relinguish, they could have been invalidated. 4911 * Recheck the valid bits and re-read as needed. 4912 * 4913 * Note that the last page is made fully valid in the 4914 * read loop, and partial validity for the page at 4915 * index count - 1 could mean that the page was 4916 * invalidated or removed, so we must restart for 4917 * safety as well. 4918 */ 4919 if (ma[i]->valid != VM_PAGE_BITS_ALL) 4920 redo = true; 4921 } 4922 if (redo && error == 0) 4923 goto again; 4924 VM_OBJECT_WUNLOCK(object); 4925 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 4926 } 4927 4928 #include "opt_ddb.h" 4929 #ifdef DDB 4930 #include <ddb/ddb.h> 4931 4932 /* DDB command to show buffer data */ 4933 DB_SHOW_COMMAND(buffer, db_show_buffer) 4934 { 4935 /* get args */ 4936 struct buf *bp = (struct buf *)addr; 4937 #ifdef FULL_BUF_TRACKING 4938 uint32_t i, j; 4939 #endif 4940 4941 if (!have_addr) { 4942 db_printf("usage: show buffer <addr>\n"); 4943 return; 4944 } 4945 4946 db_printf("buf at %p\n", bp); 4947 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4948 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4949 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4950 db_printf( 4951 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4952 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4953 "b_dep = %p\n", 4954 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4955 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4956 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4957 db_printf("b_kvabase = %p, b_kvasize = %d\n", 4958 bp->b_kvabase, bp->b_kvasize); 4959 if (bp->b_npages) { 4960 int i; 4961 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4962 for (i = 0; i < bp->b_npages; i++) { 4963 vm_page_t m; 4964 m = bp->b_pages[i]; 4965 if (m != NULL) 4966 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 4967 (u_long)m->pindex, 4968 (u_long)VM_PAGE_TO_PHYS(m)); 4969 else 4970 db_printf("( ??? )"); 4971 if ((i + 1) < bp->b_npages) 4972 db_printf(","); 4973 } 4974 db_printf("\n"); 4975 } 4976 #if defined(FULL_BUF_TRACKING) 4977 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 4978 4979 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 4980 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 4981 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 4982 continue; 4983 db_printf(" %2u: %s\n", j, 4984 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 4985 } 4986 #elif defined(BUF_TRACKING) 4987 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 4988 #endif 4989 db_printf(" "); 4990 BUF_LOCKPRINTINFO(bp); 4991 } 4992 4993 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4994 { 4995 struct buf *bp; 4996 int i; 4997 4998 for (i = 0; i < nbuf; i++) { 4999 bp = &buf[i]; 5000 if (BUF_ISLOCKED(bp)) { 5001 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5002 db_printf("\n"); 5003 if (db_pager_quit) 5004 break; 5005 } 5006 } 5007 } 5008 5009 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5010 { 5011 struct vnode *vp; 5012 struct buf *bp; 5013 5014 if (!have_addr) { 5015 db_printf("usage: show vnodebufs <addr>\n"); 5016 return; 5017 } 5018 vp = (struct vnode *)addr; 5019 db_printf("Clean buffers:\n"); 5020 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5021 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5022 db_printf("\n"); 5023 } 5024 db_printf("Dirty buffers:\n"); 5025 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5026 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5027 db_printf("\n"); 5028 } 5029 } 5030 5031 DB_COMMAND(countfreebufs, db_coundfreebufs) 5032 { 5033 struct buf *bp; 5034 int i, used = 0, nfree = 0; 5035 5036 if (have_addr) { 5037 db_printf("usage: countfreebufs\n"); 5038 return; 5039 } 5040 5041 for (i = 0; i < nbuf; i++) { 5042 bp = &buf[i]; 5043 if (bp->b_qindex == QUEUE_EMPTY) 5044 nfree++; 5045 else 5046 used++; 5047 } 5048 5049 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5050 nfree + used); 5051 db_printf("numfreebuffers is %d\n", numfreebuffers); 5052 } 5053 #endif /* DDB */ 5054