1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2004 Poul-Henning Kamp
5 * Copyright (c) 1994,1997 John S. Dyson
6 * Copyright (c) 2013 The FreeBSD Foundation
7 * All rights reserved.
8 *
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 */
33
34 /*
35 * this file contains a new buffer I/O scheme implementing a coherent
36 * VM object and buffer cache scheme. Pains have been taken to make
37 * sure that the performance degradation associated with schemes such
38 * as this is not realized.
39 *
40 * Author: John S. Dyson
41 * Significant help during the development and debugging phases
42 * had been provided by David Greenman, also of the FreeBSD core team.
43 *
44 * see man buf(9) for more info.
45 */
46
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/asan.h>
50 #include <sys/bio.h>
51 #include <sys/bitset.h>
52 #include <sys/boottrace.h>
53 #include <sys/buf.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/devicestat.h>
57 #include <sys/eventhandler.h>
58 #include <sys/fail.h>
59 #include <sys/ktr.h>
60 #include <sys/limits.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/memdesc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/pctrie.h>
69 #include <sys/proc.h>
70 #include <sys/racct.h>
71 #include <sys/refcount.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
74 #include <sys/sched.h>
75 #include <sys/smp.h>
76 #include <sys/sysctl.h>
77 #include <sys/syscallsubr.h>
78 #include <sys/vmem.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 #include <geom/geom.h>
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <vm/vm_kern.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/vm_extern.h>
91 #include <vm/vm_map.h>
92 #include <vm/swap_pager.h>
93
94 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
95
96 struct bio_ops bioops; /* I/O operation notification */
97
98 struct buf_ops buf_ops_bio = {
99 .bop_name = "buf_ops_bio",
100 .bop_write = bufwrite,
101 .bop_strategy = bufstrategy,
102 .bop_sync = bufsync,
103 .bop_bdflush = bufbdflush,
104 };
105
106 struct bufqueue {
107 struct mtx_padalign bq_lock;
108 TAILQ_HEAD(, buf) bq_queue;
109 uint8_t bq_index;
110 uint16_t bq_subqueue;
111 int bq_len;
112 } __aligned(CACHE_LINE_SIZE);
113
114 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
115 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
116 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
117 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
118
119 struct bufdomain {
120 struct bufqueue *bd_subq;
121 struct bufqueue bd_dirtyq;
122 struct bufqueue *bd_cleanq;
123 struct mtx_padalign bd_run_lock;
124 /* Constants */
125 long bd_maxbufspace;
126 long bd_hibufspace;
127 long bd_lobufspace;
128 long bd_bufspacethresh;
129 int bd_hifreebuffers;
130 int bd_lofreebuffers;
131 int bd_hidirtybuffers;
132 int bd_lodirtybuffers;
133 int bd_dirtybufthresh;
134 int bd_lim;
135 /* atomics */
136 int bd_wanted;
137 bool bd_shutdown;
138 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
139 int __aligned(CACHE_LINE_SIZE) bd_running;
140 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
141 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
142 } __aligned(CACHE_LINE_SIZE);
143
144 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
145 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
146 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
147 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
148 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
149 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
150 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
151 #define BD_DOMAIN(bd) (bd - bdomain)
152
153 static char *buf; /* buffer header pool */
154 static struct buf *
nbufp(unsigned i)155 nbufp(unsigned i)
156 {
157 return ((struct buf *)(buf + (sizeof(struct buf) +
158 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
159 }
160
161 caddr_t __read_mostly unmapped_buf;
162 #ifdef INVARIANTS
163 caddr_t poisoned_buf = (void *)-1;
164 #endif
165
166 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
167 struct proc *bufdaemonproc;
168
169 static void vm_hold_free_pages(struct buf *bp, int newbsize);
170 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
171 vm_offset_t to);
172 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
173 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
174 vm_page_t m);
175 static void vfs_clean_pages_dirty_buf(struct buf *bp);
176 static void vfs_setdirty_range(struct buf *bp);
177 static void vfs_vmio_invalidate(struct buf *bp);
178 static void vfs_vmio_truncate(struct buf *bp, int npages);
179 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
180 static int vfs_bio_clcheck(struct vnode *vp, int size,
181 daddr_t lblkno, daddr_t blkno);
182 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
183 void (*)(struct buf *));
184 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
185 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
186 static void buf_daemon(void);
187 static __inline void bd_wakeup(void);
188 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
189 static void bufkva_reclaim(vmem_t *, int);
190 static void bufkva_free(struct buf *);
191 static int buf_import(void *, void **, int, int, int);
192 static void buf_release(void *, void **, int);
193 static void maxbcachebuf_adjust(void);
194 static inline struct bufdomain *bufdomain(struct buf *);
195 static void bq_remove(struct bufqueue *bq, struct buf *bp);
196 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
197 static int buf_recycle(struct bufdomain *, bool kva);
198 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
199 const char *lockname);
200 static void bd_init(struct bufdomain *bd);
201 static int bd_flushall(struct bufdomain *bd);
202 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
203 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
204
205 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
206 int vmiodirenable = TRUE;
207 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
208 "Use the VM system for directory writes");
209 long runningbufspace;
210 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
211 "Amount of presently outstanding async buffer io");
212 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
213 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
214 static counter_u64_t bufkvaspace;
215 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
216 "Kernel virtual memory used for buffers");
217 static long maxbufspace;
218 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
219 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
220 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
221 "Maximum allowed value of bufspace (including metadata)");
222 static long bufmallocspace;
223 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
224 "Amount of malloced memory for buffers");
225 static long maxbufmallocspace;
226 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
227 0, "Maximum amount of malloced memory for buffers");
228 static long lobufspace;
229 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
230 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
231 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
232 "Minimum amount of buffers we want to have");
233 long hibufspace;
234 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
235 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
236 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
237 "Maximum allowed value of bufspace (excluding metadata)");
238 long bufspacethresh;
239 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
240 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
241 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
242 "Bufspace consumed before waking the daemon to free some");
243 static counter_u64_t buffreekvacnt;
244 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
245 "Number of times we have freed the KVA space from some buffer");
246 static counter_u64_t bufdefragcnt;
247 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
248 "Number of times we have had to repeat buffer allocation to defragment");
249 static long lorunningspace;
250 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
251 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
252 "Minimum preferred space used for in-progress I/O");
253 static long hirunningspace;
254 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
255 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
256 "Maximum amount of space to use for in-progress I/O");
257 int dirtybufferflushes;
258 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
259 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
260 int bdwriteskip;
261 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
262 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
263 int altbufferflushes;
264 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
265 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
266 static int recursiveflushes;
267 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
268 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
269 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
270 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
271 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
272 "Number of buffers that are dirty (has unwritten changes) at the moment");
273 static int lodirtybuffers;
274 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
275 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
276 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
277 "How many buffers we want to have free before bufdaemon can sleep");
278 static int hidirtybuffers;
279 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
280 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
281 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
282 "When the number of dirty buffers is considered severe");
283 int dirtybufthresh;
284 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
285 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
286 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
287 "Number of bdwrite to bawrite conversions to clear dirty buffers");
288 static int numfreebuffers;
289 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
290 "Number of free buffers");
291 static int lofreebuffers;
292 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
293 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
294 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
295 "Target number of free buffers");
296 static int hifreebuffers;
297 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
298 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
299 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
300 "Threshold for clean buffer recycling");
301 static counter_u64_t getnewbufcalls;
302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
303 &getnewbufcalls, "Number of calls to getnewbuf");
304 static counter_u64_t getnewbufrestarts;
305 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
306 &getnewbufrestarts,
307 "Number of times getnewbuf has had to restart a buffer acquisition");
308 static counter_u64_t mappingrestarts;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
310 &mappingrestarts,
311 "Number of times getblk has had to restart a buffer mapping for "
312 "unmapped buffer");
313 static counter_u64_t numbufallocfails;
314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
315 &numbufallocfails, "Number of times buffer allocations failed");
316 static int flushbufqtarget = 100;
317 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
318 "Amount of work to do in flushbufqueues when helping bufdaemon");
319 static counter_u64_t notbufdflushes;
320 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
321 "Number of dirty buffer flushes done by the bufdaemon helpers");
322 static long barrierwrites;
323 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
324 &barrierwrites, 0, "Number of barrier writes");
325 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed,
326 CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
327 &unmapped_buf_allowed, 0,
328 "Permit the use of the unmapped i/o");
329 int maxbcachebuf = MAXBCACHEBUF;
330 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
331 "Maximum size of a buffer cache block");
332
333 /*
334 * This lock synchronizes access to bd_request.
335 */
336 static struct mtx_padalign __exclusive_cache_line bdlock;
337
338 /*
339 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
340 * waitrunningbufspace().
341 */
342 static struct mtx_padalign __exclusive_cache_line rbreqlock;
343
344 /*
345 * Lock that protects bdirtywait.
346 */
347 static struct mtx_padalign __exclusive_cache_line bdirtylock;
348
349 /*
350 * bufdaemon shutdown request and sleep channel.
351 */
352 static bool bd_shutdown;
353
354 /*
355 * Wakeup point for bufdaemon, as well as indicator of whether it is already
356 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
357 * is idling.
358 */
359 static int bd_request;
360
361 /*
362 * Request for the buf daemon to write more buffers than is indicated by
363 * lodirtybuf. This may be necessary to push out excess dependencies or
364 * defragment the address space where a simple count of the number of dirty
365 * buffers is insufficient to characterize the demand for flushing them.
366 */
367 static int bd_speedupreq;
368
369 /*
370 * Synchronization (sleep/wakeup) variable for active buffer space requests.
371 * Set when wait starts, cleared prior to wakeup().
372 * Used in runningbufwakeup() and waitrunningbufspace().
373 */
374 static int runningbufreq;
375
376 /*
377 * Synchronization for bwillwrite() waiters.
378 */
379 static int bdirtywait;
380
381 /*
382 * Definitions for the buffer free lists.
383 */
384 #define QUEUE_NONE 0 /* on no queue */
385 #define QUEUE_EMPTY 1 /* empty buffer headers */
386 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
387 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
388 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
389
390 /* Maximum number of buffer domains. */
391 #define BUF_DOMAINS 8
392
393 struct bufdomainset bdlodirty; /* Domains > lodirty */
394 struct bufdomainset bdhidirty; /* Domains > hidirty */
395
396 /* Configured number of clean queues. */
397 static int __read_mostly buf_domains;
398
399 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
400 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
401 struct bufqueue __exclusive_cache_line bqempty;
402
403 /*
404 * per-cpu empty buffer cache.
405 */
406 uma_zone_t buf_zone;
407
408 static int
sysctl_runningspace(SYSCTL_HANDLER_ARGS)409 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
410 {
411 long value;
412 int error;
413
414 value = *(long *)arg1;
415 error = sysctl_handle_long(oidp, &value, 0, req);
416 if (error != 0 || req->newptr == NULL)
417 return (error);
418 mtx_lock(&rbreqlock);
419 if (arg1 == &hirunningspace) {
420 if (value < lorunningspace)
421 error = EINVAL;
422 else
423 hirunningspace = value;
424 } else {
425 KASSERT(arg1 == &lorunningspace,
426 ("%s: unknown arg1", __func__));
427 if (value > hirunningspace)
428 error = EINVAL;
429 else
430 lorunningspace = value;
431 }
432 mtx_unlock(&rbreqlock);
433 return (error);
434 }
435
436 static int
sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)437 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
438 {
439 int error;
440 int value;
441 int i;
442
443 value = *(int *)arg1;
444 error = sysctl_handle_int(oidp, &value, 0, req);
445 if (error != 0 || req->newptr == NULL)
446 return (error);
447 *(int *)arg1 = value;
448 for (i = 0; i < buf_domains; i++)
449 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
450 value / buf_domains;
451
452 return (error);
453 }
454
455 static int
sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)456 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
457 {
458 long value;
459 int error;
460 int i;
461
462 value = *(long *)arg1;
463 error = sysctl_handle_long(oidp, &value, 0, req);
464 if (error != 0 || req->newptr == NULL)
465 return (error);
466 *(long *)arg1 = value;
467 for (i = 0; i < buf_domains; i++)
468 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
469 value / buf_domains;
470
471 return (error);
472 }
473
474 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
475 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
476 static int
sysctl_bufspace(SYSCTL_HANDLER_ARGS)477 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
478 {
479 long lvalue;
480 int ivalue;
481 int i;
482
483 lvalue = 0;
484 for (i = 0; i < buf_domains; i++)
485 lvalue += bdomain[i].bd_bufspace;
486 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
487 return (sysctl_handle_long(oidp, &lvalue, 0, req));
488 if (lvalue > INT_MAX)
489 /* On overflow, still write out a long to trigger ENOMEM. */
490 return (sysctl_handle_long(oidp, &lvalue, 0, req));
491 ivalue = lvalue;
492 return (sysctl_handle_int(oidp, &ivalue, 0, req));
493 }
494 #else
495 static int
sysctl_bufspace(SYSCTL_HANDLER_ARGS)496 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
497 {
498 long lvalue;
499 int i;
500
501 lvalue = 0;
502 for (i = 0; i < buf_domains; i++)
503 lvalue += bdomain[i].bd_bufspace;
504 return (sysctl_handle_long(oidp, &lvalue, 0, req));
505 }
506 #endif
507
508 static int
sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)509 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
510 {
511 int value;
512 int i;
513
514 value = 0;
515 for (i = 0; i < buf_domains; i++)
516 value += bdomain[i].bd_numdirtybuffers;
517 return (sysctl_handle_int(oidp, &value, 0, req));
518 }
519
520 /*
521 * bdirtywakeup:
522 *
523 * Wakeup any bwillwrite() waiters.
524 */
525 static void
bdirtywakeup(void)526 bdirtywakeup(void)
527 {
528 mtx_lock(&bdirtylock);
529 if (bdirtywait) {
530 bdirtywait = 0;
531 wakeup(&bdirtywait);
532 }
533 mtx_unlock(&bdirtylock);
534 }
535
536 /*
537 * bd_clear:
538 *
539 * Clear a domain from the appropriate bitsets when dirtybuffers
540 * is decremented.
541 */
542 static void
bd_clear(struct bufdomain * bd)543 bd_clear(struct bufdomain *bd)
544 {
545
546 mtx_lock(&bdirtylock);
547 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
548 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
549 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
550 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
551 mtx_unlock(&bdirtylock);
552 }
553
554 /*
555 * bd_set:
556 *
557 * Set a domain in the appropriate bitsets when dirtybuffers
558 * is incremented.
559 */
560 static void
bd_set(struct bufdomain * bd)561 bd_set(struct bufdomain *bd)
562 {
563
564 mtx_lock(&bdirtylock);
565 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
566 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
567 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
568 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
569 mtx_unlock(&bdirtylock);
570 }
571
572 /*
573 * bdirtysub:
574 *
575 * Decrement the numdirtybuffers count by one and wakeup any
576 * threads blocked in bwillwrite().
577 */
578 static void
bdirtysub(struct buf * bp)579 bdirtysub(struct buf *bp)
580 {
581 struct bufdomain *bd;
582 int num;
583
584 bd = bufdomain(bp);
585 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
586 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
587 bdirtywakeup();
588 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
589 bd_clear(bd);
590 }
591
592 /*
593 * bdirtyadd:
594 *
595 * Increment the numdirtybuffers count by one and wakeup the buf
596 * daemon if needed.
597 */
598 static void
bdirtyadd(struct buf * bp)599 bdirtyadd(struct buf *bp)
600 {
601 struct bufdomain *bd;
602 int num;
603
604 /*
605 * Only do the wakeup once as we cross the boundary. The
606 * buf daemon will keep running until the condition clears.
607 */
608 bd = bufdomain(bp);
609 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
610 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
611 bd_wakeup();
612 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
613 bd_set(bd);
614 }
615
616 /*
617 * bufspace_daemon_wakeup:
618 *
619 * Wakeup the daemons responsible for freeing clean bufs.
620 */
621 static void
bufspace_daemon_wakeup(struct bufdomain * bd)622 bufspace_daemon_wakeup(struct bufdomain *bd)
623 {
624
625 /*
626 * avoid the lock if the daemon is running.
627 */
628 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
629 BD_RUN_LOCK(bd);
630 atomic_store_int(&bd->bd_running, 1);
631 wakeup(&bd->bd_running);
632 BD_RUN_UNLOCK(bd);
633 }
634 }
635
636 /*
637 * bufspace_adjust:
638 *
639 * Adjust the reported bufspace for a KVA managed buffer, possibly
640 * waking any waiters.
641 */
642 static void
bufspace_adjust(struct buf * bp,int bufsize)643 bufspace_adjust(struct buf *bp, int bufsize)
644 {
645 struct bufdomain *bd;
646 long space;
647 int diff;
648
649 KASSERT((bp->b_flags & B_MALLOC) == 0,
650 ("bufspace_adjust: malloc buf %p", bp));
651 bd = bufdomain(bp);
652 diff = bufsize - bp->b_bufsize;
653 if (diff < 0) {
654 atomic_subtract_long(&bd->bd_bufspace, -diff);
655 } else if (diff > 0) {
656 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
657 /* Wake up the daemon on the transition. */
658 if (space < bd->bd_bufspacethresh &&
659 space + diff >= bd->bd_bufspacethresh)
660 bufspace_daemon_wakeup(bd);
661 }
662 bp->b_bufsize = bufsize;
663 }
664
665 /*
666 * bufspace_reserve:
667 *
668 * Reserve bufspace before calling allocbuf(). metadata has a
669 * different space limit than data.
670 */
671 static int
bufspace_reserve(struct bufdomain * bd,int size,bool metadata)672 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
673 {
674 long limit, new;
675 long space;
676
677 if (metadata)
678 limit = bd->bd_maxbufspace;
679 else
680 limit = bd->bd_hibufspace;
681 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
682 new = space + size;
683 if (new > limit) {
684 atomic_subtract_long(&bd->bd_bufspace, size);
685 return (ENOSPC);
686 }
687
688 /* Wake up the daemon on the transition. */
689 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
690 bufspace_daemon_wakeup(bd);
691
692 return (0);
693 }
694
695 /*
696 * bufspace_release:
697 *
698 * Release reserved bufspace after bufspace_adjust() has consumed it.
699 */
700 static void
bufspace_release(struct bufdomain * bd,int size)701 bufspace_release(struct bufdomain *bd, int size)
702 {
703
704 atomic_subtract_long(&bd->bd_bufspace, size);
705 }
706
707 /*
708 * bufspace_wait:
709 *
710 * Wait for bufspace, acting as the buf daemon if a locked vnode is
711 * supplied. bd_wanted must be set prior to polling for space. The
712 * operation must be re-tried on return.
713 */
714 static void
bufspace_wait(struct bufdomain * bd,struct vnode * vp,int gbflags,int slpflag,int slptimeo)715 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
716 int slpflag, int slptimeo)
717 {
718 struct thread *td;
719 int error, fl, norunbuf;
720
721 if ((gbflags & GB_NOWAIT_BD) != 0)
722 return;
723
724 td = curthread;
725 BD_LOCK(bd);
726 while (bd->bd_wanted) {
727 if (vp != NULL && vp->v_type != VCHR &&
728 (td->td_pflags & TDP_BUFNEED) == 0) {
729 BD_UNLOCK(bd);
730 /*
731 * getblk() is called with a vnode locked, and
732 * some majority of the dirty buffers may as
733 * well belong to the vnode. Flushing the
734 * buffers there would make a progress that
735 * cannot be achieved by the buf_daemon, that
736 * cannot lock the vnode.
737 */
738 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
739 (td->td_pflags & TDP_NORUNNINGBUF);
740
741 /*
742 * Play bufdaemon. The getnewbuf() function
743 * may be called while the thread owns lock
744 * for another dirty buffer for the same
745 * vnode, which makes it impossible to use
746 * VOP_FSYNC() there, due to the buffer lock
747 * recursion.
748 */
749 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
750 fl = buf_flush(vp, bd, flushbufqtarget);
751 td->td_pflags &= norunbuf;
752 BD_LOCK(bd);
753 if (fl != 0)
754 continue;
755 if (bd->bd_wanted == 0)
756 break;
757 }
758 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
759 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
760 if (error != 0)
761 break;
762 }
763 BD_UNLOCK(bd);
764 }
765
766 static void
bufspace_daemon_shutdown(void * arg,int howto __unused)767 bufspace_daemon_shutdown(void *arg, int howto __unused)
768 {
769 struct bufdomain *bd = arg;
770 int error;
771
772 if (KERNEL_PANICKED())
773 return;
774
775 BD_RUN_LOCK(bd);
776 bd->bd_shutdown = true;
777 wakeup(&bd->bd_running);
778 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
779 "bufspace_shutdown", 60 * hz);
780 BD_RUN_UNLOCK(bd);
781 if (error != 0)
782 printf("bufspacedaemon wait error: %d\n", error);
783 }
784
785 /*
786 * bufspace_daemon:
787 *
788 * buffer space management daemon. Tries to maintain some marginal
789 * amount of free buffer space so that requesting processes neither
790 * block nor work to reclaim buffers.
791 */
792 static void
bufspace_daemon(void * arg)793 bufspace_daemon(void *arg)
794 {
795 struct bufdomain *bd = arg;
796
797 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
798 SHUTDOWN_PRI_LAST + 100);
799
800 BD_RUN_LOCK(bd);
801 while (!bd->bd_shutdown) {
802 BD_RUN_UNLOCK(bd);
803
804 /*
805 * Free buffers from the clean queue until we meet our
806 * targets.
807 *
808 * Theory of operation: The buffer cache is most efficient
809 * when some free buffer headers and space are always
810 * available to getnewbuf(). This daemon attempts to prevent
811 * the excessive blocking and synchronization associated
812 * with shortfall. It goes through three phases according
813 * demand:
814 *
815 * 1) The daemon wakes up voluntarily once per-second
816 * during idle periods when the counters are below
817 * the wakeup thresholds (bufspacethresh, lofreebuffers).
818 *
819 * 2) The daemon wakes up as we cross the thresholds
820 * ahead of any potential blocking. This may bounce
821 * slightly according to the rate of consumption and
822 * release.
823 *
824 * 3) The daemon and consumers are starved for working
825 * clean buffers. This is the 'bufspace' sleep below
826 * which will inefficiently trade bufs with bqrelse
827 * until we return to condition 2.
828 */
829 while (bd->bd_bufspace > bd->bd_lobufspace ||
830 bd->bd_freebuffers < bd->bd_hifreebuffers) {
831 if (buf_recycle(bd, false) != 0) {
832 if (bd_flushall(bd))
833 continue;
834 /*
835 * Speedup dirty if we've run out of clean
836 * buffers. This is possible in particular
837 * because softdep may held many bufs locked
838 * pending writes to other bufs which are
839 * marked for delayed write, exhausting
840 * clean space until they are written.
841 */
842 bd_speedup();
843 BD_LOCK(bd);
844 if (bd->bd_wanted) {
845 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
846 PRIBIO|PDROP, "bufspace", hz/10);
847 } else
848 BD_UNLOCK(bd);
849 }
850 maybe_yield();
851 }
852
853 /*
854 * Re-check our limits and sleep. bd_running must be
855 * cleared prior to checking the limits to avoid missed
856 * wakeups. The waker will adjust one of bufspace or
857 * freebuffers prior to checking bd_running.
858 */
859 BD_RUN_LOCK(bd);
860 if (bd->bd_shutdown)
861 break;
862 atomic_store_int(&bd->bd_running, 0);
863 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
864 bd->bd_freebuffers > bd->bd_lofreebuffers) {
865 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
866 PRIBIO, "-", hz);
867 } else {
868 /* Avoid spurious wakeups while running. */
869 atomic_store_int(&bd->bd_running, 1);
870 }
871 }
872 wakeup(&bd->bd_shutdown);
873 BD_RUN_UNLOCK(bd);
874 kthread_exit();
875 }
876
877 /*
878 * bufmallocadjust:
879 *
880 * Adjust the reported bufspace for a malloc managed buffer, possibly
881 * waking any waiters.
882 */
883 static void
bufmallocadjust(struct buf * bp,int bufsize)884 bufmallocadjust(struct buf *bp, int bufsize)
885 {
886 int diff;
887
888 KASSERT((bp->b_flags & B_MALLOC) != 0,
889 ("bufmallocadjust: non-malloc buf %p", bp));
890 diff = bufsize - bp->b_bufsize;
891 if (diff < 0)
892 atomic_subtract_long(&bufmallocspace, -diff);
893 else
894 atomic_add_long(&bufmallocspace, diff);
895 bp->b_bufsize = bufsize;
896 }
897
898 /*
899 * runningwakeup:
900 *
901 * Wake up processes that are waiting on asynchronous writes to fall
902 * below lorunningspace.
903 */
904 static void
runningwakeup(void)905 runningwakeup(void)
906 {
907
908 mtx_lock(&rbreqlock);
909 if (runningbufreq) {
910 runningbufreq = 0;
911 wakeup(&runningbufreq);
912 }
913 mtx_unlock(&rbreqlock);
914 }
915
916 /*
917 * runningbufwakeup:
918 *
919 * Decrement the outstanding write count according.
920 */
921 void
runningbufwakeup(struct buf * bp)922 runningbufwakeup(struct buf *bp)
923 {
924 long space, bspace;
925
926 bspace = bp->b_runningbufspace;
927 if (bspace == 0)
928 return;
929 space = atomic_fetchadd_long(&runningbufspace, -bspace);
930 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
931 space, bspace));
932 bp->b_runningbufspace = 0;
933 /*
934 * Only acquire the lock and wakeup on the transition from exceeding
935 * the threshold to falling below it.
936 */
937 if (space < lorunningspace)
938 return;
939 if (space - bspace > lorunningspace)
940 return;
941 runningwakeup();
942 }
943
944 /*
945 * waitrunningbufspace()
946 *
947 * runningbufspace is a measure of the amount of I/O currently
948 * running. This routine is used in async-write situations to
949 * prevent creating huge backups of pending writes to a device.
950 * Only asynchronous writes are governed by this function.
951 *
952 * This does NOT turn an async write into a sync write. It waits
953 * for earlier writes to complete and generally returns before the
954 * caller's write has reached the device.
955 */
956 void
waitrunningbufspace(void)957 waitrunningbufspace(void)
958 {
959
960 mtx_lock(&rbreqlock);
961 while (runningbufspace > hirunningspace) {
962 runningbufreq = 1;
963 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
964 }
965 mtx_unlock(&rbreqlock);
966 }
967
968 /*
969 * vfs_buf_test_cache:
970 *
971 * Called when a buffer is extended. This function clears the B_CACHE
972 * bit if the newly extended portion of the buffer does not contain
973 * valid data.
974 */
975 static __inline void
vfs_buf_test_cache(struct buf * bp,vm_ooffset_t foff,vm_offset_t off,vm_offset_t size,vm_page_t m)976 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
977 vm_offset_t size, vm_page_t m)
978 {
979
980 /*
981 * This function and its results are protected by higher level
982 * synchronization requiring vnode and buf locks to page in and
983 * validate pages.
984 */
985 if (bp->b_flags & B_CACHE) {
986 int base = (foff + off) & PAGE_MASK;
987 if (vm_page_is_valid(m, base, size) == 0)
988 bp->b_flags &= ~B_CACHE;
989 }
990 }
991
992 /* Wake up the buffer daemon if necessary */
993 static void
bd_wakeup(void)994 bd_wakeup(void)
995 {
996
997 mtx_lock(&bdlock);
998 if (bd_request == 0) {
999 bd_request = 1;
1000 wakeup(&bd_request);
1001 }
1002 mtx_unlock(&bdlock);
1003 }
1004
1005 /*
1006 * Adjust the maxbcachbuf tunable.
1007 */
1008 static void
maxbcachebuf_adjust(void)1009 maxbcachebuf_adjust(void)
1010 {
1011 int i;
1012
1013 /*
1014 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1015 */
1016 i = 2;
1017 while (i * 2 <= maxbcachebuf)
1018 i *= 2;
1019 maxbcachebuf = i;
1020 if (maxbcachebuf < MAXBSIZE)
1021 maxbcachebuf = MAXBSIZE;
1022 if (maxbcachebuf > maxphys)
1023 maxbcachebuf = maxphys;
1024 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1025 printf("maxbcachebuf=%d\n", maxbcachebuf);
1026 }
1027
1028 /*
1029 * bd_speedup - speedup the buffer cache flushing code
1030 */
1031 void
bd_speedup(void)1032 bd_speedup(void)
1033 {
1034 int needwake;
1035
1036 mtx_lock(&bdlock);
1037 needwake = 0;
1038 if (bd_speedupreq == 0 || bd_request == 0)
1039 needwake = 1;
1040 bd_speedupreq = 1;
1041 bd_request = 1;
1042 if (needwake)
1043 wakeup(&bd_request);
1044 mtx_unlock(&bdlock);
1045 }
1046
1047 #ifdef __i386__
1048 #define TRANSIENT_DENOM 5
1049 #else
1050 #define TRANSIENT_DENOM 10
1051 #endif
1052
1053 /*
1054 * Calculating buffer cache scaling values and reserve space for buffer
1055 * headers. This is called during low level kernel initialization and
1056 * may be called more then once. We CANNOT write to the memory area
1057 * being reserved at this time.
1058 */
1059 caddr_t
kern_vfs_bio_buffer_alloc(caddr_t v,long physmem_est)1060 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1061 {
1062 int tuned_nbuf;
1063 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1064
1065 /*
1066 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for
1067 * this when sizing maps based on the amount of physical memory
1068 * available.
1069 */
1070 #if defined(KASAN)
1071 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1072 (KASAN_SHADOW_SCALE + 1);
1073 #elif defined(KMSAN)
1074 physmem_est /= 3;
1075
1076 /*
1077 * KMSAN cannot reliably determine whether buffer data is initialized
1078 * unless it is updated through a KVA mapping.
1079 */
1080 unmapped_buf_allowed = 0;
1081 #endif
1082
1083 /*
1084 * physmem_est is in pages. Convert it to kilobytes (assumes
1085 * PAGE_SIZE is >= 1K)
1086 */
1087 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1088
1089 maxbcachebuf_adjust();
1090 /*
1091 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1092 * For the first 64MB of ram nominally allocate sufficient buffers to
1093 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1094 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1095 * the buffer cache we limit the eventual kva reservation to
1096 * maxbcache bytes.
1097 *
1098 * factor represents the 1/4 x ram conversion.
1099 */
1100 if (nbuf == 0) {
1101 int factor = 4 * BKVASIZE / 1024;
1102
1103 nbuf = 50;
1104 if (physmem_est > 4096)
1105 nbuf += min((physmem_est - 4096) / factor,
1106 65536 / factor);
1107 if (physmem_est > 65536)
1108 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1109 32 * 1024 * 1024 / (factor * 5));
1110
1111 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1112 nbuf = maxbcache / BKVASIZE;
1113 tuned_nbuf = 1;
1114 } else
1115 tuned_nbuf = 0;
1116
1117 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1118 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1119 if (nbuf > maxbuf) {
1120 if (!tuned_nbuf)
1121 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1122 maxbuf);
1123 nbuf = maxbuf;
1124 }
1125
1126 /*
1127 * Ideal allocation size for the transient bio submap is 10%
1128 * of the maximal space buffer map. This roughly corresponds
1129 * to the amount of the buffer mapped for typical UFS load.
1130 *
1131 * Clip the buffer map to reserve space for the transient
1132 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1133 * maximum buffer map extent on the platform.
1134 *
1135 * The fall-back to the maxbuf in case of maxbcache unset,
1136 * allows to not trim the buffer KVA for the architectures
1137 * with ample KVA space.
1138 */
1139 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1140 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1141 buf_sz = (long)nbuf * BKVASIZE;
1142 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1143 (TRANSIENT_DENOM - 1)) {
1144 /*
1145 * There is more KVA than memory. Do not
1146 * adjust buffer map size, and assign the rest
1147 * of maxbuf to transient map.
1148 */
1149 biotmap_sz = maxbuf_sz - buf_sz;
1150 } else {
1151 /*
1152 * Buffer map spans all KVA we could afford on
1153 * this platform. Give 10% (20% on i386) of
1154 * the buffer map to the transient bio map.
1155 */
1156 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1157 buf_sz -= biotmap_sz;
1158 }
1159 if (biotmap_sz / INT_MAX > maxphys)
1160 bio_transient_maxcnt = INT_MAX;
1161 else
1162 bio_transient_maxcnt = biotmap_sz / maxphys;
1163 /*
1164 * Artificially limit to 1024 simultaneous in-flight I/Os
1165 * using the transient mapping.
1166 */
1167 if (bio_transient_maxcnt > 1024)
1168 bio_transient_maxcnt = 1024;
1169 if (tuned_nbuf)
1170 nbuf = buf_sz / BKVASIZE;
1171 }
1172
1173 if (nswbuf == 0) {
1174 /*
1175 * Pager buffers are allocated for short periods, so scale the
1176 * number of reserved buffers based on the number of CPUs rather
1177 * than amount of memory.
1178 */
1179 nswbuf = min(nbuf / 4, 32 * mp_ncpus);
1180 if (nswbuf < NSWBUF_MIN)
1181 nswbuf = NSWBUF_MIN;
1182 }
1183
1184 /*
1185 * Reserve space for the buffer cache buffers
1186 */
1187 buf = (char *)v;
1188 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1189 atop(maxbcachebuf)) * nbuf;
1190
1191 return (v);
1192 }
1193
1194 /*
1195 * Single global constant for BUF_WMESG, to avoid getting multiple
1196 * references.
1197 */
1198 static const char buf_wmesg[] = "bufwait";
1199
1200 /* Initialize the buffer subsystem. Called before use of any buffers. */
1201 void
bufinit(void)1202 bufinit(void)
1203 {
1204 struct buf *bp;
1205 int i;
1206
1207 TSENTER();
1208 KASSERT(maxbcachebuf >= MAXBSIZE,
1209 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1210 MAXBSIZE));
1211 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1212 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1213 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1214 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1215
1216 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1217 #ifdef INVARIANTS
1218 poisoned_buf = unmapped_buf;
1219 #endif
1220
1221 /* finally, initialize each buffer header and stick on empty q */
1222 for (i = 0; i < nbuf; i++) {
1223 bp = nbufp(i);
1224 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1225 bp->b_flags = B_INVAL;
1226 bp->b_rcred = NOCRED;
1227 bp->b_wcred = NOCRED;
1228 bp->b_qindex = QUEUE_NONE;
1229 bp->b_domain = -1;
1230 bp->b_subqueue = mp_maxid + 1;
1231 bp->b_xflags = 0;
1232 bp->b_data = bp->b_kvabase = unmapped_buf;
1233 LIST_INIT(&bp->b_dep);
1234 BUF_LOCKINIT(bp, buf_wmesg);
1235 bq_insert(&bqempty, bp, false);
1236 }
1237
1238 /*
1239 * maxbufspace is the absolute maximum amount of buffer space we are
1240 * allowed to reserve in KVM and in real terms. The absolute maximum
1241 * is nominally used by metadata. hibufspace is the nominal maximum
1242 * used by most other requests. The differential is required to
1243 * ensure that metadata deadlocks don't occur.
1244 *
1245 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1246 * this may result in KVM fragmentation which is not handled optimally
1247 * by the system. XXX This is less true with vmem. We could use
1248 * PAGE_SIZE.
1249 */
1250 maxbufspace = (long)nbuf * BKVASIZE;
1251 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1252 lobufspace = (hibufspace / 20) * 19; /* 95% */
1253 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1254
1255 /*
1256 * Note: The upper limit for hirunningspace was chosen arbitrarily and
1257 * may need further tuning. It corresponds to 128 outstanding write IO
1258 * requests, which fits with many RAID controllers' tagged queuing
1259 * limits.
1260 *
1261 * The lower 1 MiB limit is the historical upper limit for
1262 * hirunningspace.
1263 */
1264 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1265 128 * maxphys), 1024 * 1024);
1266 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1267
1268 /*
1269 * Limit the amount of malloc memory since it is wired permanently into
1270 * the kernel space. Even though this is accounted for in the buffer
1271 * allocation, we don't want the malloced region to grow uncontrolled.
1272 * The malloc scheme improves memory utilization significantly on
1273 * average (small) directories.
1274 */
1275 maxbufmallocspace = hibufspace / 20;
1276
1277 /*
1278 * Reduce the chance of a deadlock occurring by limiting the number
1279 * of delayed-write dirty buffers we allow to stack up.
1280 */
1281 hidirtybuffers = nbuf / 4 + 20;
1282 dirtybufthresh = hidirtybuffers * 9 / 10;
1283 /*
1284 * To support extreme low-memory systems, make sure hidirtybuffers
1285 * cannot eat up all available buffer space. This occurs when our
1286 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1287 * buffer space assuming BKVASIZE'd buffers.
1288 */
1289 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1290 hidirtybuffers >>= 1;
1291 }
1292 lodirtybuffers = hidirtybuffers / 2;
1293
1294 /*
1295 * lofreebuffers should be sufficient to avoid stalling waiting on
1296 * buf headers under heavy utilization. The bufs in per-cpu caches
1297 * are counted as free but will be unavailable to threads executing
1298 * on other cpus.
1299 *
1300 * hifreebuffers is the free target for the bufspace daemon. This
1301 * should be set appropriately to limit work per-iteration.
1302 */
1303 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1304 hifreebuffers = (3 * lofreebuffers) / 2;
1305 numfreebuffers = nbuf;
1306
1307 /* Setup the kva and free list allocators. */
1308 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1309 buf_zone = uma_zcache_create("buf free cache",
1310 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1311 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1312
1313 /*
1314 * Size the clean queue according to the amount of buffer space.
1315 * One queue per-256mb up to the max. More queues gives better
1316 * concurrency but less accurate LRU.
1317 */
1318 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1319 for (i = 0 ; i < buf_domains; i++) {
1320 struct bufdomain *bd;
1321
1322 bd = &bdomain[i];
1323 bd_init(bd);
1324 bd->bd_freebuffers = nbuf / buf_domains;
1325 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1326 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1327 bd->bd_bufspace = 0;
1328 bd->bd_maxbufspace = maxbufspace / buf_domains;
1329 bd->bd_hibufspace = hibufspace / buf_domains;
1330 bd->bd_lobufspace = lobufspace / buf_domains;
1331 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1332 bd->bd_numdirtybuffers = 0;
1333 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1334 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1335 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1336 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1337 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1338 }
1339 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1340 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1341 mappingrestarts = counter_u64_alloc(M_WAITOK);
1342 numbufallocfails = counter_u64_alloc(M_WAITOK);
1343 notbufdflushes = counter_u64_alloc(M_WAITOK);
1344 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1345 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1346 bufkvaspace = counter_u64_alloc(M_WAITOK);
1347 TSEXIT();
1348 }
1349
1350 #ifdef INVARIANTS
1351 static inline void
vfs_buf_check_mapped(struct buf * bp)1352 vfs_buf_check_mapped(struct buf *bp)
1353 {
1354
1355 KASSERT(bp->b_kvabase != unmapped_buf,
1356 ("mapped buf: b_kvabase was not updated %p", bp));
1357 KASSERT(bp->b_data != unmapped_buf,
1358 ("mapped buf: b_data was not updated %p", bp));
1359 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1360 maxphys, ("b_data + b_offset unmapped %p", bp));
1361 }
1362
1363 static inline void
vfs_buf_check_unmapped(struct buf * bp)1364 vfs_buf_check_unmapped(struct buf *bp)
1365 {
1366
1367 KASSERT(bp->b_data == unmapped_buf,
1368 ("unmapped buf: corrupted b_data %p", bp));
1369 }
1370
1371 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1372 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1373 #else
1374 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1375 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1376 #endif
1377
1378 static int
isbufbusy(struct buf * bp)1379 isbufbusy(struct buf *bp)
1380 {
1381 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1382 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1383 return (1);
1384 return (0);
1385 }
1386
1387 /*
1388 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1389 */
1390 void
bufshutdown(int show_busybufs)1391 bufshutdown(int show_busybufs)
1392 {
1393 static int first_buf_printf = 1;
1394 struct buf *bp;
1395 int i, iter, nbusy, pbusy;
1396 #ifndef PREEMPTION
1397 int subiter;
1398 #endif
1399
1400 /*
1401 * Sync filesystems for shutdown
1402 */
1403 wdog_kern_pat(WD_LASTVAL);
1404 kern_sync(curthread);
1405
1406 /*
1407 * With soft updates, some buffers that are
1408 * written will be remarked as dirty until other
1409 * buffers are written.
1410 */
1411 for (iter = pbusy = 0; iter < 20; iter++) {
1412 nbusy = 0;
1413 for (i = nbuf - 1; i >= 0; i--) {
1414 bp = nbufp(i);
1415 if (isbufbusy(bp))
1416 nbusy++;
1417 }
1418 if (nbusy == 0) {
1419 if (first_buf_printf)
1420 printf("All buffers synced.");
1421 break;
1422 }
1423 if (first_buf_printf) {
1424 printf("Syncing disks, buffers remaining... ");
1425 first_buf_printf = 0;
1426 }
1427 printf("%d ", nbusy);
1428 if (nbusy < pbusy)
1429 iter = 0;
1430 pbusy = nbusy;
1431
1432 wdog_kern_pat(WD_LASTVAL);
1433 kern_sync(curthread);
1434
1435 #ifdef PREEMPTION
1436 /*
1437 * Spin for a while to allow interrupt threads to run.
1438 */
1439 DELAY(50000 * iter);
1440 #else
1441 /*
1442 * Context switch several times to allow interrupt
1443 * threads to run.
1444 */
1445 for (subiter = 0; subiter < 50 * iter; subiter++) {
1446 sched_relinquish(curthread);
1447 DELAY(1000);
1448 }
1449 #endif
1450 }
1451 printf("\n");
1452 /*
1453 * Count only busy local buffers to prevent forcing
1454 * a fsck if we're just a client of a wedged NFS server
1455 */
1456 nbusy = 0;
1457 for (i = nbuf - 1; i >= 0; i--) {
1458 bp = nbufp(i);
1459 if (isbufbusy(bp)) {
1460 #if 0
1461 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1462 if (bp->b_dev == NULL) {
1463 TAILQ_REMOVE(&mountlist,
1464 bp->b_vp->v_mount, mnt_list);
1465 continue;
1466 }
1467 #endif
1468 nbusy++;
1469 if (show_busybufs > 0) {
1470 printf(
1471 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1472 nbusy, bp, bp->b_vp, bp->b_flags,
1473 (intmax_t)bp->b_blkno,
1474 (intmax_t)bp->b_lblkno);
1475 BUF_LOCKPRINTINFO(bp);
1476 if (show_busybufs > 1)
1477 vn_printf(bp->b_vp,
1478 "vnode content: ");
1479 }
1480 }
1481 }
1482 if (nbusy) {
1483 /*
1484 * Failed to sync all blocks. Indicate this and don't
1485 * unmount filesystems (thus forcing an fsck on reboot).
1486 */
1487 BOOTTRACE("shutdown failed to sync buffers");
1488 printf("Giving up on %d buffers\n", nbusy);
1489 DELAY(5000000); /* 5 seconds */
1490 swapoff_all();
1491 } else {
1492 BOOTTRACE("shutdown sync complete");
1493 if (!first_buf_printf)
1494 printf("Final sync complete\n");
1495
1496 /*
1497 * Unmount filesystems and perform swapoff, to quiesce
1498 * the system as much as possible. In particular, no
1499 * I/O should be initiated from top levels since it
1500 * might be abruptly terminated by reset, or otherwise
1501 * erronously handled because other parts of the
1502 * system are disabled.
1503 *
1504 * Swapoff before unmount, because file-backed swap is
1505 * non-operational after unmount of the underlying
1506 * filesystem.
1507 */
1508 if (!KERNEL_PANICKED()) {
1509 swapoff_all();
1510 vfs_unmountall();
1511 }
1512 BOOTTRACE("shutdown unmounted all filesystems");
1513 }
1514 DELAY(100000); /* wait for console output to finish */
1515 }
1516
1517 static void
bpmap_qenter(struct buf * bp)1518 bpmap_qenter(struct buf *bp)
1519 {
1520
1521 BUF_CHECK_MAPPED(bp);
1522
1523 /*
1524 * bp->b_data is relative to bp->b_offset, but
1525 * bp->b_offset may be offset into the first page.
1526 */
1527 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1528 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1529 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1530 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1531 }
1532
1533 static inline struct bufdomain *
bufdomain(struct buf * bp)1534 bufdomain(struct buf *bp)
1535 {
1536
1537 return (&bdomain[bp->b_domain]);
1538 }
1539
1540 static struct bufqueue *
bufqueue(struct buf * bp)1541 bufqueue(struct buf *bp)
1542 {
1543
1544 switch (bp->b_qindex) {
1545 case QUEUE_NONE:
1546 /* FALLTHROUGH */
1547 case QUEUE_SENTINEL:
1548 return (NULL);
1549 case QUEUE_EMPTY:
1550 return (&bqempty);
1551 case QUEUE_DIRTY:
1552 return (&bufdomain(bp)->bd_dirtyq);
1553 case QUEUE_CLEAN:
1554 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1555 default:
1556 break;
1557 }
1558 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1559 }
1560
1561 /*
1562 * Return the locked bufqueue that bp is a member of.
1563 */
1564 static struct bufqueue *
bufqueue_acquire(struct buf * bp)1565 bufqueue_acquire(struct buf *bp)
1566 {
1567 struct bufqueue *bq, *nbq;
1568
1569 /*
1570 * bp can be pushed from a per-cpu queue to the
1571 * cleanq while we're waiting on the lock. Retry
1572 * if the queues don't match.
1573 */
1574 bq = bufqueue(bp);
1575 BQ_LOCK(bq);
1576 for (;;) {
1577 nbq = bufqueue(bp);
1578 if (bq == nbq)
1579 break;
1580 BQ_UNLOCK(bq);
1581 BQ_LOCK(nbq);
1582 bq = nbq;
1583 }
1584 return (bq);
1585 }
1586
1587 /*
1588 * binsfree:
1589 *
1590 * Insert the buffer into the appropriate free list. Requires a
1591 * locked buffer on entry and buffer is unlocked before return.
1592 */
1593 static void
binsfree(struct buf * bp,int qindex)1594 binsfree(struct buf *bp, int qindex)
1595 {
1596 struct bufdomain *bd;
1597 struct bufqueue *bq;
1598
1599 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1600 ("binsfree: Invalid qindex %d", qindex));
1601 BUF_ASSERT_XLOCKED(bp);
1602
1603 /*
1604 * Handle delayed bremfree() processing.
1605 */
1606 if (bp->b_flags & B_REMFREE) {
1607 if (bp->b_qindex == qindex) {
1608 bp->b_flags |= B_REUSE;
1609 bp->b_flags &= ~B_REMFREE;
1610 BUF_UNLOCK(bp);
1611 return;
1612 }
1613 bq = bufqueue_acquire(bp);
1614 bq_remove(bq, bp);
1615 BQ_UNLOCK(bq);
1616 }
1617 bd = bufdomain(bp);
1618 if (qindex == QUEUE_CLEAN) {
1619 if (bd->bd_lim != 0)
1620 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1621 else
1622 bq = bd->bd_cleanq;
1623 } else
1624 bq = &bd->bd_dirtyq;
1625 bq_insert(bq, bp, true);
1626 }
1627
1628 /*
1629 * buf_free:
1630 *
1631 * Free a buffer to the buf zone once it no longer has valid contents.
1632 */
1633 static void
buf_free(struct buf * bp)1634 buf_free(struct buf *bp)
1635 {
1636
1637 if (bp->b_flags & B_REMFREE)
1638 bremfreef(bp);
1639 if (bp->b_vflags & BV_BKGRDINPROG)
1640 panic("losing buffer 1");
1641 if (bp->b_rcred != NOCRED) {
1642 crfree(bp->b_rcred);
1643 bp->b_rcred = NOCRED;
1644 }
1645 if (bp->b_wcred != NOCRED) {
1646 crfree(bp->b_wcred);
1647 bp->b_wcred = NOCRED;
1648 }
1649 if (!LIST_EMPTY(&bp->b_dep))
1650 buf_deallocate(bp);
1651 bufkva_free(bp);
1652 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1653 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1654 BUF_UNLOCK(bp);
1655 uma_zfree(buf_zone, bp);
1656 }
1657
1658 /*
1659 * buf_import:
1660 *
1661 * Import bufs into the uma cache from the buf list. The system still
1662 * expects a static array of bufs and much of the synchronization
1663 * around bufs assumes type stable storage. As a result, UMA is used
1664 * only as a per-cpu cache of bufs still maintained on a global list.
1665 */
1666 static int
buf_import(void * arg,void ** store,int cnt,int domain,int flags)1667 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1668 {
1669 struct buf *bp;
1670 int i;
1671
1672 BQ_LOCK(&bqempty);
1673 for (i = 0; i < cnt; i++) {
1674 bp = TAILQ_FIRST(&bqempty.bq_queue);
1675 if (bp == NULL)
1676 break;
1677 bq_remove(&bqempty, bp);
1678 store[i] = bp;
1679 }
1680 BQ_UNLOCK(&bqempty);
1681
1682 return (i);
1683 }
1684
1685 /*
1686 * buf_release:
1687 *
1688 * Release bufs from the uma cache back to the buffer queues.
1689 */
1690 static void
buf_release(void * arg,void ** store,int cnt)1691 buf_release(void *arg, void **store, int cnt)
1692 {
1693 struct bufqueue *bq;
1694 struct buf *bp;
1695 int i;
1696
1697 bq = &bqempty;
1698 BQ_LOCK(bq);
1699 for (i = 0; i < cnt; i++) {
1700 bp = store[i];
1701 /* Inline bq_insert() to batch locking. */
1702 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1703 bp->b_flags &= ~(B_AGE | B_REUSE);
1704 bq->bq_len++;
1705 bp->b_qindex = bq->bq_index;
1706 }
1707 BQ_UNLOCK(bq);
1708 }
1709
1710 /*
1711 * buf_alloc:
1712 *
1713 * Allocate an empty buffer header.
1714 */
1715 static struct buf *
buf_alloc(struct bufdomain * bd)1716 buf_alloc(struct bufdomain *bd)
1717 {
1718 struct buf *bp;
1719 int freebufs, error;
1720
1721 /*
1722 * We can only run out of bufs in the buf zone if the average buf
1723 * is less than BKVASIZE. In this case the actual wait/block will
1724 * come from buf_reycle() failing to flush one of these small bufs.
1725 */
1726 bp = NULL;
1727 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1728 if (freebufs > 0)
1729 bp = uma_zalloc(buf_zone, M_NOWAIT);
1730 if (bp == NULL) {
1731 atomic_add_int(&bd->bd_freebuffers, 1);
1732 bufspace_daemon_wakeup(bd);
1733 counter_u64_add(numbufallocfails, 1);
1734 return (NULL);
1735 }
1736 /*
1737 * Wake-up the bufspace daemon on transition below threshold.
1738 */
1739 if (freebufs == bd->bd_lofreebuffers)
1740 bufspace_daemon_wakeup(bd);
1741
1742 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1743 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1744 error));
1745 (void)error;
1746
1747 KASSERT(bp->b_vp == NULL,
1748 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1749 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1750 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1751 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1752 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1753 KASSERT(bp->b_npages == 0,
1754 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1755 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1756 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1757 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1758
1759 bp->b_domain = BD_DOMAIN(bd);
1760 bp->b_flags = 0;
1761 bp->b_ioflags = 0;
1762 bp->b_xflags = 0;
1763 bp->b_vflags = 0;
1764 bp->b_vp = NULL;
1765 bp->b_blkno = bp->b_lblkno = 0;
1766 bp->b_offset = NOOFFSET;
1767 bp->b_iodone = 0;
1768 bp->b_error = 0;
1769 bp->b_resid = 0;
1770 bp->b_bcount = 0;
1771 bp->b_npages = 0;
1772 bp->b_dirtyoff = bp->b_dirtyend = 0;
1773 bp->b_bufobj = NULL;
1774 bp->b_data = bp->b_kvabase = unmapped_buf;
1775 bp->b_fsprivate1 = NULL;
1776 bp->b_fsprivate2 = NULL;
1777 bp->b_fsprivate3 = NULL;
1778 LIST_INIT(&bp->b_dep);
1779
1780 return (bp);
1781 }
1782
1783 /*
1784 * buf_recycle:
1785 *
1786 * Free a buffer from the given bufqueue. kva controls whether the
1787 * freed buf must own some kva resources. This is used for
1788 * defragmenting.
1789 */
1790 static int
buf_recycle(struct bufdomain * bd,bool kva)1791 buf_recycle(struct bufdomain *bd, bool kva)
1792 {
1793 struct bufqueue *bq;
1794 struct buf *bp, *nbp;
1795
1796 if (kva)
1797 counter_u64_add(bufdefragcnt, 1);
1798 nbp = NULL;
1799 bq = bd->bd_cleanq;
1800 BQ_LOCK(bq);
1801 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1802 ("buf_recycle: Locks don't match"));
1803 nbp = TAILQ_FIRST(&bq->bq_queue);
1804
1805 /*
1806 * Run scan, possibly freeing data and/or kva mappings on the fly
1807 * depending.
1808 */
1809 while ((bp = nbp) != NULL) {
1810 /*
1811 * Calculate next bp (we can only use it if we do not
1812 * release the bqlock).
1813 */
1814 nbp = TAILQ_NEXT(bp, b_freelist);
1815
1816 /*
1817 * If we are defragging then we need a buffer with
1818 * some kva to reclaim.
1819 */
1820 if (kva && bp->b_kvasize == 0)
1821 continue;
1822
1823 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1824 continue;
1825
1826 /*
1827 * Implement a second chance algorithm for frequently
1828 * accessed buffers.
1829 */
1830 if ((bp->b_flags & B_REUSE) != 0) {
1831 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1832 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1833 bp->b_flags &= ~B_REUSE;
1834 BUF_UNLOCK(bp);
1835 continue;
1836 }
1837
1838 /*
1839 * Skip buffers with background writes in progress.
1840 */
1841 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1842 BUF_UNLOCK(bp);
1843 continue;
1844 }
1845
1846 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1847 ("buf_recycle: inconsistent queue %d bp %p",
1848 bp->b_qindex, bp));
1849 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1850 ("getnewbuf: queue domain %d doesn't match request %d",
1851 bp->b_domain, (int)BD_DOMAIN(bd)));
1852 /*
1853 * NOTE: nbp is now entirely invalid. We can only restart
1854 * the scan from this point on.
1855 */
1856 bq_remove(bq, bp);
1857 BQ_UNLOCK(bq);
1858
1859 /*
1860 * Requeue the background write buffer with error and
1861 * restart the scan.
1862 */
1863 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1864 bqrelse(bp);
1865 BQ_LOCK(bq);
1866 nbp = TAILQ_FIRST(&bq->bq_queue);
1867 continue;
1868 }
1869 bp->b_flags |= B_INVAL;
1870 brelse(bp);
1871 return (0);
1872 }
1873 bd->bd_wanted = 1;
1874 BQ_UNLOCK(bq);
1875
1876 return (ENOBUFS);
1877 }
1878
1879 /*
1880 * bremfree:
1881 *
1882 * Mark the buffer for removal from the appropriate free list.
1883 *
1884 */
1885 void
bremfree(struct buf * bp)1886 bremfree(struct buf *bp)
1887 {
1888
1889 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1890 KASSERT((bp->b_flags & B_REMFREE) == 0,
1891 ("bremfree: buffer %p already marked for delayed removal.", bp));
1892 KASSERT(bp->b_qindex != QUEUE_NONE,
1893 ("bremfree: buffer %p not on a queue.", bp));
1894 BUF_ASSERT_XLOCKED(bp);
1895
1896 bp->b_flags |= B_REMFREE;
1897 }
1898
1899 /*
1900 * bremfreef:
1901 *
1902 * Force an immediate removal from a free list. Used only in nfs when
1903 * it abuses the b_freelist pointer.
1904 */
1905 void
bremfreef(struct buf * bp)1906 bremfreef(struct buf *bp)
1907 {
1908 struct bufqueue *bq;
1909
1910 bq = bufqueue_acquire(bp);
1911 bq_remove(bq, bp);
1912 BQ_UNLOCK(bq);
1913 }
1914
1915 static void
bq_init(struct bufqueue * bq,int qindex,int subqueue,const char * lockname)1916 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1917 {
1918
1919 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1920 TAILQ_INIT(&bq->bq_queue);
1921 bq->bq_len = 0;
1922 bq->bq_index = qindex;
1923 bq->bq_subqueue = subqueue;
1924 }
1925
1926 static void
bd_init(struct bufdomain * bd)1927 bd_init(struct bufdomain *bd)
1928 {
1929 int i;
1930
1931 /* Per-CPU clean buf queues, plus one global queue. */
1932 bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue),
1933 M_BIOBUF, M_WAITOK | M_ZERO);
1934 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1935 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1936 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1937 for (i = 0; i <= mp_maxid; i++)
1938 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1939 "bufq clean subqueue lock");
1940 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1941 }
1942
1943 /*
1944 * bq_remove:
1945 *
1946 * Removes a buffer from the free list, must be called with the
1947 * correct qlock held.
1948 */
1949 static void
bq_remove(struct bufqueue * bq,struct buf * bp)1950 bq_remove(struct bufqueue *bq, struct buf *bp)
1951 {
1952
1953 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1954 bp, bp->b_vp, bp->b_flags);
1955 KASSERT(bp->b_qindex != QUEUE_NONE,
1956 ("bq_remove: buffer %p not on a queue.", bp));
1957 KASSERT(bufqueue(bp) == bq,
1958 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1959
1960 BQ_ASSERT_LOCKED(bq);
1961 if (bp->b_qindex != QUEUE_EMPTY) {
1962 BUF_ASSERT_XLOCKED(bp);
1963 }
1964 KASSERT(bq->bq_len >= 1,
1965 ("queue %d underflow", bp->b_qindex));
1966 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1967 bq->bq_len--;
1968 bp->b_qindex = QUEUE_NONE;
1969 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1970 }
1971
1972 static void
bd_flush(struct bufdomain * bd,struct bufqueue * bq)1973 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1974 {
1975 struct buf *bp;
1976
1977 BQ_ASSERT_LOCKED(bq);
1978 if (bq != bd->bd_cleanq) {
1979 BD_LOCK(bd);
1980 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1981 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1982 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1983 b_freelist);
1984 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1985 }
1986 bd->bd_cleanq->bq_len += bq->bq_len;
1987 bq->bq_len = 0;
1988 }
1989 if (bd->bd_wanted) {
1990 bd->bd_wanted = 0;
1991 wakeup(&bd->bd_wanted);
1992 }
1993 if (bq != bd->bd_cleanq)
1994 BD_UNLOCK(bd);
1995 }
1996
1997 static int
bd_flushall(struct bufdomain * bd)1998 bd_flushall(struct bufdomain *bd)
1999 {
2000 struct bufqueue *bq;
2001 int flushed;
2002 int i;
2003
2004 if (bd->bd_lim == 0)
2005 return (0);
2006 flushed = 0;
2007 for (i = 0; i <= mp_maxid; i++) {
2008 bq = &bd->bd_subq[i];
2009 if (bq->bq_len == 0)
2010 continue;
2011 BQ_LOCK(bq);
2012 bd_flush(bd, bq);
2013 BQ_UNLOCK(bq);
2014 flushed++;
2015 }
2016
2017 return (flushed);
2018 }
2019
2020 static void
bq_insert(struct bufqueue * bq,struct buf * bp,bool unlock)2021 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
2022 {
2023 struct bufdomain *bd;
2024
2025 if (bp->b_qindex != QUEUE_NONE)
2026 panic("bq_insert: free buffer %p onto another queue?", bp);
2027
2028 bd = bufdomain(bp);
2029 if (bp->b_flags & B_AGE) {
2030 /* Place this buf directly on the real queue. */
2031 if (bq->bq_index == QUEUE_CLEAN)
2032 bq = bd->bd_cleanq;
2033 BQ_LOCK(bq);
2034 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2035 } else {
2036 BQ_LOCK(bq);
2037 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2038 }
2039 bp->b_flags &= ~(B_AGE | B_REUSE);
2040 bq->bq_len++;
2041 bp->b_qindex = bq->bq_index;
2042 bp->b_subqueue = bq->bq_subqueue;
2043
2044 /*
2045 * Unlock before we notify so that we don't wakeup a waiter that
2046 * fails a trylock on the buf and sleeps again.
2047 */
2048 if (unlock)
2049 BUF_UNLOCK(bp);
2050
2051 if (bp->b_qindex == QUEUE_CLEAN) {
2052 /*
2053 * Flush the per-cpu queue and notify any waiters.
2054 */
2055 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2056 bq->bq_len >= bd->bd_lim))
2057 bd_flush(bd, bq);
2058 }
2059 BQ_UNLOCK(bq);
2060 }
2061
2062 /*
2063 * bufkva_free:
2064 *
2065 * Free the kva allocation for a buffer.
2066 *
2067 */
2068 static void
bufkva_free(struct buf * bp)2069 bufkva_free(struct buf *bp)
2070 {
2071
2072 #ifdef INVARIANTS
2073 if (bp->b_kvasize == 0) {
2074 KASSERT(bp->b_kvabase == unmapped_buf &&
2075 bp->b_data == unmapped_buf,
2076 ("Leaked KVA space on %p", bp));
2077 } else if (buf_mapped(bp))
2078 BUF_CHECK_MAPPED(bp);
2079 else
2080 BUF_CHECK_UNMAPPED(bp);
2081 #endif
2082 if (bp->b_kvasize == 0)
2083 return;
2084
2085 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2086 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2087 counter_u64_add(buffreekvacnt, 1);
2088 bp->b_data = bp->b_kvabase = unmapped_buf;
2089 bp->b_kvasize = 0;
2090 }
2091
2092 /*
2093 * bufkva_alloc:
2094 *
2095 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2096 */
2097 static int
bufkva_alloc(struct buf * bp,int maxsize,int gbflags)2098 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2099 {
2100 vm_offset_t addr;
2101 int error;
2102
2103 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2104 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2105 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2106 KASSERT(maxsize <= maxbcachebuf,
2107 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2108
2109 bufkva_free(bp);
2110
2111 addr = 0;
2112 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2113 if (error != 0) {
2114 /*
2115 * Buffer map is too fragmented. Request the caller
2116 * to defragment the map.
2117 */
2118 return (error);
2119 }
2120 bp->b_kvabase = (caddr_t)addr;
2121 bp->b_kvasize = maxsize;
2122 counter_u64_add(bufkvaspace, bp->b_kvasize);
2123 if ((gbflags & GB_UNMAPPED) != 0) {
2124 bp->b_data = unmapped_buf;
2125 BUF_CHECK_UNMAPPED(bp);
2126 } else {
2127 bp->b_data = bp->b_kvabase;
2128 BUF_CHECK_MAPPED(bp);
2129 }
2130 return (0);
2131 }
2132
2133 /*
2134 * bufkva_reclaim:
2135 *
2136 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2137 * callback that fires to avoid returning failure.
2138 */
2139 static void
bufkva_reclaim(vmem_t * vmem,int flags)2140 bufkva_reclaim(vmem_t *vmem, int flags)
2141 {
2142 bool done;
2143 int q;
2144 int i;
2145
2146 done = false;
2147 for (i = 0; i < 5; i++) {
2148 for (q = 0; q < buf_domains; q++)
2149 if (buf_recycle(&bdomain[q], true) != 0)
2150 done = true;
2151 if (done)
2152 break;
2153 }
2154 return;
2155 }
2156
2157 /*
2158 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2159 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2160 * the buffer is valid and we do not have to do anything.
2161 */
2162 static void
breada(struct vnode * vp,daddr_t * rablkno,int * rabsize,int cnt,struct ucred * cred,int flags,void (* ckhashfunc)(struct buf *))2163 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2164 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2165 {
2166 struct buf *rabp;
2167 struct thread *td;
2168 int i;
2169
2170 td = curthread;
2171
2172 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2173 if (inmem(vp, *rablkno))
2174 continue;
2175 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2176 if ((rabp->b_flags & B_CACHE) != 0) {
2177 brelse(rabp);
2178 continue;
2179 }
2180 #ifdef RACCT
2181 if (racct_enable) {
2182 PROC_LOCK(curproc);
2183 racct_add_buf(curproc, rabp, 0);
2184 PROC_UNLOCK(curproc);
2185 }
2186 #endif /* RACCT */
2187 td->td_ru.ru_inblock++;
2188 rabp->b_flags |= B_ASYNC;
2189 rabp->b_flags &= ~B_INVAL;
2190 if ((flags & GB_CKHASH) != 0) {
2191 rabp->b_flags |= B_CKHASH;
2192 rabp->b_ckhashcalc = ckhashfunc;
2193 }
2194 rabp->b_ioflags &= ~BIO_ERROR;
2195 rabp->b_iocmd = BIO_READ;
2196 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2197 rabp->b_rcred = crhold(cred);
2198 vfs_busy_pages(rabp, 0);
2199 BUF_KERNPROC(rabp);
2200 rabp->b_iooffset = dbtob(rabp->b_blkno);
2201 bstrategy(rabp);
2202 }
2203 }
2204
2205 /*
2206 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2207 *
2208 * Get a buffer with the specified data. Look in the cache first. We
2209 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2210 * is set, the buffer is valid and we do not have to do anything, see
2211 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2212 *
2213 * Always return a NULL buffer pointer (in bpp) when returning an error.
2214 *
2215 * The blkno parameter is the logical block being requested. Normally
2216 * the mapping of logical block number to disk block address is done
2217 * by calling VOP_BMAP(). However, if the mapping is already known, the
2218 * disk block address can be passed using the dblkno parameter. If the
2219 * disk block address is not known, then the same value should be passed
2220 * for blkno and dblkno.
2221 */
2222 int
breadn_flags(struct vnode * vp,daddr_t blkno,daddr_t dblkno,int size,daddr_t * rablkno,int * rabsize,int cnt,struct ucred * cred,int flags,void (* ckhashfunc)(struct buf *),struct buf ** bpp)2223 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2224 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2225 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2226 {
2227 struct buf *bp;
2228 struct thread *td;
2229 int error, readwait, rv;
2230
2231 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2232 td = curthread;
2233 /*
2234 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2235 * are specified.
2236 */
2237 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2238 if (error != 0) {
2239 *bpp = NULL;
2240 return (error);
2241 }
2242 KASSERT(blkno == bp->b_lblkno,
2243 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2244 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2245 flags &= ~GB_NOSPARSE;
2246 *bpp = bp;
2247
2248 /*
2249 * If not found in cache, do some I/O
2250 */
2251 readwait = 0;
2252 if ((bp->b_flags & B_CACHE) == 0) {
2253 #ifdef RACCT
2254 if (racct_enable) {
2255 PROC_LOCK(td->td_proc);
2256 racct_add_buf(td->td_proc, bp, 0);
2257 PROC_UNLOCK(td->td_proc);
2258 }
2259 #endif /* RACCT */
2260 td->td_ru.ru_inblock++;
2261 bp->b_iocmd = BIO_READ;
2262 bp->b_flags &= ~B_INVAL;
2263 if ((flags & GB_CKHASH) != 0) {
2264 bp->b_flags |= B_CKHASH;
2265 bp->b_ckhashcalc = ckhashfunc;
2266 }
2267 if ((flags & GB_CVTENXIO) != 0)
2268 bp->b_xflags |= BX_CVTENXIO;
2269 bp->b_ioflags &= ~BIO_ERROR;
2270 if (bp->b_rcred == NOCRED && cred != NOCRED)
2271 bp->b_rcred = crhold(cred);
2272 vfs_busy_pages(bp, 0);
2273 bp->b_iooffset = dbtob(bp->b_blkno);
2274 bstrategy(bp);
2275 ++readwait;
2276 }
2277
2278 /*
2279 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2280 */
2281 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2282
2283 rv = 0;
2284 if (readwait) {
2285 rv = bufwait(bp);
2286 if (rv != 0) {
2287 brelse(bp);
2288 *bpp = NULL;
2289 }
2290 }
2291 return (rv);
2292 }
2293
2294 /*
2295 * Write, release buffer on completion. (Done by iodone
2296 * if async). Do not bother writing anything if the buffer
2297 * is invalid.
2298 *
2299 * Note that we set B_CACHE here, indicating that buffer is
2300 * fully valid and thus cacheable. This is true even of NFS
2301 * now so we set it generally. This could be set either here
2302 * or in biodone() since the I/O is synchronous. We put it
2303 * here.
2304 */
2305 int
bufwrite(struct buf * bp)2306 bufwrite(struct buf *bp)
2307 {
2308 struct vnode *vp;
2309 long space;
2310 int oldflags, retval;
2311 bool vp_md;
2312
2313 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2314 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2315 bp->b_flags |= B_INVAL | B_RELBUF;
2316 bp->b_flags &= ~B_CACHE;
2317 brelse(bp);
2318 return (ENXIO);
2319 }
2320 if ((bp->b_flags & B_INVAL) != 0) {
2321 brelse(bp);
2322 return (0);
2323 }
2324
2325 if ((bp->b_flags & B_BARRIER) != 0)
2326 atomic_add_long(&barrierwrites, 1);
2327
2328 oldflags = bp->b_flags;
2329
2330 KASSERT((bp->b_vflags & BV_BKGRDINPROG) == 0,
2331 ("FFS background buffer should not get here %p", bp));
2332
2333 vp = bp->b_vp;
2334 vp_md = vp != NULL && (vp->v_vflag & VV_MD) != 0;
2335
2336 /*
2337 * Mark the buffer clean. Increment the bufobj write count
2338 * before bundirty() call, to prevent other thread from seeing
2339 * empty dirty list and zero counter for writes in progress,
2340 * falsely indicating that the bufobj is clean.
2341 */
2342 bufobj_wref(bp->b_bufobj);
2343 bundirty(bp);
2344
2345 bp->b_flags &= ~B_DONE;
2346 bp->b_ioflags &= ~BIO_ERROR;
2347 bp->b_flags |= B_CACHE;
2348 bp->b_iocmd = BIO_WRITE;
2349
2350 vfs_busy_pages(bp, 1);
2351
2352 /*
2353 * Normal bwrites pipeline writes
2354 */
2355 bp->b_runningbufspace = bp->b_bufsize;
2356 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2357
2358 #ifdef RACCT
2359 if (racct_enable) {
2360 PROC_LOCK(curproc);
2361 racct_add_buf(curproc, bp, 1);
2362 PROC_UNLOCK(curproc);
2363 }
2364 #endif /* RACCT */
2365 curthread->td_ru.ru_oublock++;
2366 if ((oldflags & B_ASYNC) != 0)
2367 BUF_KERNPROC(bp);
2368 bp->b_iooffset = dbtob(bp->b_blkno);
2369 buf_track(bp, __func__);
2370 bstrategy(bp);
2371
2372 if ((oldflags & B_ASYNC) == 0) {
2373 retval = bufwait(bp);
2374 brelse(bp);
2375 return (retval);
2376 } else if (space > hirunningspace) {
2377 /*
2378 * Don't allow the async write to saturate the I/O
2379 * system. We do not block here if it is the update
2380 * or syncer daemon trying to clean up as that can
2381 * lead to deadlock.
2382 */
2383 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2384 waitrunningbufspace();
2385 }
2386
2387 return (0);
2388 }
2389
2390 void
bufbdflush(struct bufobj * bo,struct buf * bp)2391 bufbdflush(struct bufobj *bo, struct buf *bp)
2392 {
2393 struct buf *nbp;
2394 struct bufdomain *bd;
2395
2396 bd = &bdomain[bo->bo_domain];
2397 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2398 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2399 altbufferflushes++;
2400 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2401 BO_LOCK(bo);
2402 /*
2403 * Try to find a buffer to flush.
2404 */
2405 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2406 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2407 BUF_LOCK(nbp,
2408 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2409 continue;
2410 if (bp == nbp)
2411 panic("bdwrite: found ourselves");
2412 BO_UNLOCK(bo);
2413 /* Don't countdeps with the bo lock held. */
2414 if (buf_countdeps(nbp, 0)) {
2415 BO_LOCK(bo);
2416 BUF_UNLOCK(nbp);
2417 continue;
2418 }
2419 if (nbp->b_flags & B_CLUSTEROK) {
2420 vfs_bio_awrite(nbp);
2421 } else {
2422 bremfree(nbp);
2423 bawrite(nbp);
2424 }
2425 dirtybufferflushes++;
2426 break;
2427 }
2428 if (nbp == NULL)
2429 BO_UNLOCK(bo);
2430 }
2431 }
2432
2433 /*
2434 * Delayed write. (Buffer is marked dirty). Do not bother writing
2435 * anything if the buffer is marked invalid.
2436 *
2437 * Note that since the buffer must be completely valid, we can safely
2438 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2439 * biodone() in order to prevent getblk from writing the buffer
2440 * out synchronously.
2441 */
2442 void
bdwrite(struct buf * bp)2443 bdwrite(struct buf *bp)
2444 {
2445 struct thread *td = curthread;
2446 struct vnode *vp;
2447 struct bufobj *bo;
2448
2449 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2450 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2451 KASSERT((bp->b_flags & B_BARRIER) == 0,
2452 ("Barrier request in delayed write %p", bp));
2453
2454 if (bp->b_flags & B_INVAL) {
2455 brelse(bp);
2456 return;
2457 }
2458
2459 /*
2460 * If we have too many dirty buffers, don't create any more.
2461 * If we are wildly over our limit, then force a complete
2462 * cleanup. Otherwise, just keep the situation from getting
2463 * out of control. Note that we have to avoid a recursive
2464 * disaster and not try to clean up after our own cleanup!
2465 */
2466 vp = bp->b_vp;
2467 bo = bp->b_bufobj;
2468 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2469 td->td_pflags |= TDP_INBDFLUSH;
2470 BO_BDFLUSH(bo, bp);
2471 td->td_pflags &= ~TDP_INBDFLUSH;
2472 } else
2473 recursiveflushes++;
2474
2475 bdirty(bp);
2476 /*
2477 * Set B_CACHE, indicating that the buffer is fully valid. This is
2478 * true even of NFS now.
2479 */
2480 bp->b_flags |= B_CACHE;
2481
2482 /*
2483 * This bmap keeps the system from needing to do the bmap later,
2484 * perhaps when the system is attempting to do a sync. Since it
2485 * is likely that the indirect block -- or whatever other datastructure
2486 * that the filesystem needs is still in memory now, it is a good
2487 * thing to do this. Note also, that if the pageout daemon is
2488 * requesting a sync -- there might not be enough memory to do
2489 * the bmap then... So, this is important to do.
2490 */
2491 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2492 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2493 }
2494
2495 buf_track(bp, __func__);
2496
2497 /*
2498 * Set the *dirty* buffer range based upon the VM system dirty
2499 * pages.
2500 *
2501 * Mark the buffer pages as clean. We need to do this here to
2502 * satisfy the vnode_pager and the pageout daemon, so that it
2503 * thinks that the pages have been "cleaned". Note that since
2504 * the pages are in a delayed write buffer -- the VFS layer
2505 * "will" see that the pages get written out on the next sync,
2506 * or perhaps the cluster will be completed.
2507 */
2508 vfs_clean_pages_dirty_buf(bp);
2509 bqrelse(bp);
2510
2511 /*
2512 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2513 * due to the softdep code.
2514 */
2515 }
2516
2517 /*
2518 * bdirty:
2519 *
2520 * Turn buffer into delayed write request. We must clear BIO_READ and
2521 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2522 * itself to properly update it in the dirty/clean lists. We mark it
2523 * B_DONE to ensure that any asynchronization of the buffer properly
2524 * clears B_DONE ( else a panic will occur later ).
2525 *
2526 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2527 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2528 * should only be called if the buffer is known-good.
2529 *
2530 * Since the buffer is not on a queue, we do not update the numfreebuffers
2531 * count.
2532 *
2533 * The buffer must be on QUEUE_NONE.
2534 */
2535 void
bdirty(struct buf * bp)2536 bdirty(struct buf *bp)
2537 {
2538
2539 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2540 bp, bp->b_vp, bp->b_flags);
2541 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2542 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2543 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2544 bp->b_flags &= ~(B_RELBUF);
2545 bp->b_iocmd = BIO_WRITE;
2546
2547 if ((bp->b_flags & B_DELWRI) == 0) {
2548 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2549 reassignbuf(bp);
2550 bdirtyadd(bp);
2551 }
2552 }
2553
2554 /*
2555 * bundirty:
2556 *
2557 * Clear B_DELWRI for buffer.
2558 *
2559 * Since the buffer is not on a queue, we do not update the numfreebuffers
2560 * count.
2561 *
2562 * The buffer must be on QUEUE_NONE.
2563 */
2564
2565 void
bundirty(struct buf * bp)2566 bundirty(struct buf *bp)
2567 {
2568
2569 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2570 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2571 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2572 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2573
2574 if (bp->b_flags & B_DELWRI) {
2575 bp->b_flags &= ~B_DELWRI;
2576 reassignbuf(bp);
2577 bdirtysub(bp);
2578 }
2579 /*
2580 * Since it is now being written, we can clear its deferred write flag.
2581 */
2582 bp->b_flags &= ~B_DEFERRED;
2583 }
2584
2585 /*
2586 * bawrite:
2587 *
2588 * Asynchronous write. Start output on a buffer, but do not wait for
2589 * it to complete. The buffer is released when the output completes.
2590 *
2591 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2592 * B_INVAL buffers. Not us.
2593 */
2594 void
bawrite(struct buf * bp)2595 bawrite(struct buf *bp)
2596 {
2597
2598 bp->b_flags |= B_ASYNC;
2599 (void) bwrite(bp);
2600 }
2601
2602 /*
2603 * babarrierwrite:
2604 *
2605 * Asynchronous barrier write. Start output on a buffer, but do not
2606 * wait for it to complete. Place a write barrier after this write so
2607 * that this buffer and all buffers written before it are committed to
2608 * the disk before any buffers written after this write are committed
2609 * to the disk. The buffer is released when the output completes.
2610 */
2611 void
babarrierwrite(struct buf * bp)2612 babarrierwrite(struct buf *bp)
2613 {
2614
2615 bp->b_flags |= B_ASYNC | B_BARRIER;
2616 (void) bwrite(bp);
2617 }
2618
2619 /*
2620 * bbarrierwrite:
2621 *
2622 * Synchronous barrier write. Start output on a buffer and wait for
2623 * it to complete. Place a write barrier after this write so that
2624 * this buffer and all buffers written before it are committed to
2625 * the disk before any buffers written after this write are committed
2626 * to the disk. The buffer is released when the output completes.
2627 */
2628 int
bbarrierwrite(struct buf * bp)2629 bbarrierwrite(struct buf *bp)
2630 {
2631
2632 bp->b_flags |= B_BARRIER;
2633 return (bwrite(bp));
2634 }
2635
2636 /*
2637 * bwillwrite:
2638 *
2639 * Called prior to the locking of any vnodes when we are expecting to
2640 * write. We do not want to starve the buffer cache with too many
2641 * dirty buffers so we block here. By blocking prior to the locking
2642 * of any vnodes we attempt to avoid the situation where a locked vnode
2643 * prevents the various system daemons from flushing related buffers.
2644 */
2645 void
bwillwrite(void)2646 bwillwrite(void)
2647 {
2648
2649 if (buf_dirty_count_severe()) {
2650 mtx_lock(&bdirtylock);
2651 while (buf_dirty_count_severe()) {
2652 bdirtywait = 1;
2653 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2654 "flswai", 0);
2655 }
2656 mtx_unlock(&bdirtylock);
2657 }
2658 }
2659
2660 /*
2661 * Return true if we have too many dirty buffers.
2662 */
2663 int
buf_dirty_count_severe(void)2664 buf_dirty_count_severe(void)
2665 {
2666
2667 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2668 }
2669
2670 /*
2671 * brelse:
2672 *
2673 * Release a busy buffer and, if requested, free its resources. The
2674 * buffer will be stashed in the appropriate bufqueue[] allowing it
2675 * to be accessed later as a cache entity or reused for other purposes.
2676 */
2677 void
brelse(struct buf * bp)2678 brelse(struct buf *bp)
2679 {
2680 struct mount *v_mnt;
2681 int qindex;
2682
2683 /*
2684 * Many functions erroneously call brelse with a NULL bp under rare
2685 * error conditions. Simply return when called with a NULL bp.
2686 */
2687 if (bp == NULL)
2688 return;
2689 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2690 bp, bp->b_vp, bp->b_flags);
2691 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2692 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2693 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2694 ("brelse: non-VMIO buffer marked NOREUSE"));
2695
2696 if (BUF_LOCKRECURSED(bp)) {
2697 /*
2698 * Do not process, in particular, do not handle the
2699 * B_INVAL/B_RELBUF and do not release to free list.
2700 */
2701 BUF_UNLOCK(bp);
2702 return;
2703 }
2704
2705 if (bp->b_flags & B_MANAGED) {
2706 bqrelse(bp);
2707 return;
2708 }
2709
2710 if (LIST_EMPTY(&bp->b_dep)) {
2711 bp->b_flags &= ~B_IOSTARTED;
2712 } else {
2713 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2714 ("brelse: SU io not finished bp %p", bp));
2715 }
2716
2717 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2718 BO_LOCK(bp->b_bufobj);
2719 bp->b_vflags &= ~BV_BKGRDERR;
2720 BO_UNLOCK(bp->b_bufobj);
2721 bdirty(bp);
2722 }
2723
2724 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2725 (bp->b_flags & B_INVALONERR)) {
2726 /*
2727 * Forced invalidation of dirty buffer contents, to be used
2728 * after a failed write in the rare case that the loss of the
2729 * contents is acceptable. The buffer is invalidated and
2730 * freed.
2731 */
2732 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2733 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2734 }
2735
2736 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2737 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2738 !(bp->b_flags & B_INVAL)) {
2739 /*
2740 * Failed write, redirty. All errors except ENXIO (which
2741 * means the device is gone) are treated as being
2742 * transient.
2743 *
2744 * XXX Treating EIO as transient is not correct; the
2745 * contract with the local storage device drivers is that
2746 * they will only return EIO once the I/O is no longer
2747 * retriable. Network I/O also respects this through the
2748 * guarantees of TCP and/or the internal retries of NFS.
2749 * ENOMEM might be transient, but we also have no way of
2750 * knowing when its ok to retry/reschedule. In general,
2751 * this entire case should be made obsolete through better
2752 * error handling/recovery and resource scheduling.
2753 *
2754 * Do this also for buffers that failed with ENXIO, but have
2755 * non-empty dependencies - the soft updates code might need
2756 * to access the buffer to untangle them.
2757 *
2758 * Must clear BIO_ERROR to prevent pages from being scrapped.
2759 */
2760 bp->b_ioflags &= ~BIO_ERROR;
2761 bdirty(bp);
2762 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2763 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2764 /*
2765 * Either a failed read I/O, or we were asked to free or not
2766 * cache the buffer, or we failed to write to a device that's
2767 * no longer present.
2768 */
2769 bp->b_flags |= B_INVAL;
2770 if (!LIST_EMPTY(&bp->b_dep))
2771 buf_deallocate(bp);
2772 if (bp->b_flags & B_DELWRI)
2773 bdirtysub(bp);
2774 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2775 if ((bp->b_flags & B_VMIO) == 0) {
2776 allocbuf(bp, 0);
2777 if (bp->b_vp)
2778 brelvp(bp);
2779 }
2780 }
2781
2782 /*
2783 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2784 * is called with B_DELWRI set, the underlying pages may wind up
2785 * getting freed causing a previous write (bdwrite()) to get 'lost'
2786 * because pages associated with a B_DELWRI bp are marked clean.
2787 *
2788 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2789 * if B_DELWRI is set.
2790 */
2791 if (bp->b_flags & B_DELWRI)
2792 bp->b_flags &= ~B_RELBUF;
2793
2794 /*
2795 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2796 * constituted, not even NFS buffers now. Two flags effect this. If
2797 * B_INVAL, the struct buf is invalidated but the VM object is kept
2798 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2799 *
2800 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2801 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2802 * buffer is also B_INVAL because it hits the re-dirtying code above.
2803 *
2804 * Normally we can do this whether a buffer is B_DELWRI or not. If
2805 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2806 * the commit state and we cannot afford to lose the buffer. If the
2807 * buffer has a background write in progress, we need to keep it
2808 * around to prevent it from being reconstituted and starting a second
2809 * background write.
2810 */
2811
2812 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2813
2814 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2815 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2816 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2817 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2818 vfs_vmio_invalidate(bp);
2819 allocbuf(bp, 0);
2820 }
2821
2822 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2823 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2824 allocbuf(bp, 0);
2825 bp->b_flags &= ~B_NOREUSE;
2826 if (bp->b_vp != NULL)
2827 brelvp(bp);
2828 }
2829
2830 /*
2831 * If the buffer has junk contents signal it and eventually
2832 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2833 * doesn't find it.
2834 */
2835 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2836 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2837 bp->b_flags |= B_INVAL;
2838 if (bp->b_flags & B_INVAL) {
2839 if (bp->b_flags & B_DELWRI)
2840 bundirty(bp);
2841 if (bp->b_vp)
2842 brelvp(bp);
2843 }
2844
2845 buf_track(bp, __func__);
2846
2847 /* buffers with no memory */
2848 if (bp->b_bufsize == 0) {
2849 buf_free(bp);
2850 return;
2851 }
2852 /* buffers with junk contents */
2853 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2854 (bp->b_ioflags & BIO_ERROR)) {
2855 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2856 if (bp->b_vflags & BV_BKGRDINPROG)
2857 panic("losing buffer 2");
2858 qindex = QUEUE_CLEAN;
2859 bp->b_flags |= B_AGE;
2860 /* remaining buffers */
2861 } else if (bp->b_flags & B_DELWRI)
2862 qindex = QUEUE_DIRTY;
2863 else
2864 qindex = QUEUE_CLEAN;
2865
2866 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2867 panic("brelse: not dirty");
2868
2869 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2870 bp->b_xflags &= ~(BX_CVTENXIO);
2871 /* binsfree unlocks bp. */
2872 binsfree(bp, qindex);
2873 }
2874
2875 /*
2876 * Release a buffer back to the appropriate queue but do not try to free
2877 * it. The buffer is expected to be used again soon.
2878 *
2879 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2880 * biodone() to requeue an async I/O on completion. It is also used when
2881 * known good buffers need to be requeued but we think we may need the data
2882 * again soon.
2883 *
2884 * XXX we should be able to leave the B_RELBUF hint set on completion.
2885 */
2886 void
bqrelse(struct buf * bp)2887 bqrelse(struct buf *bp)
2888 {
2889 int qindex;
2890
2891 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2892 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2893 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2894
2895 qindex = QUEUE_NONE;
2896 if (BUF_LOCKRECURSED(bp)) {
2897 /* do not release to free list */
2898 BUF_UNLOCK(bp);
2899 return;
2900 }
2901 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2902 bp->b_xflags &= ~(BX_CVTENXIO);
2903
2904 if (LIST_EMPTY(&bp->b_dep)) {
2905 bp->b_flags &= ~B_IOSTARTED;
2906 } else {
2907 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2908 ("bqrelse: SU io not finished bp %p", bp));
2909 }
2910
2911 if (bp->b_flags & B_MANAGED) {
2912 if (bp->b_flags & B_REMFREE)
2913 bremfreef(bp);
2914 goto out;
2915 }
2916
2917 /* buffers with stale but valid contents */
2918 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2919 BV_BKGRDERR)) == BV_BKGRDERR) {
2920 BO_LOCK(bp->b_bufobj);
2921 bp->b_vflags &= ~BV_BKGRDERR;
2922 BO_UNLOCK(bp->b_bufobj);
2923 qindex = QUEUE_DIRTY;
2924 } else {
2925 if ((bp->b_flags & B_DELWRI) == 0 &&
2926 (bp->b_xflags & BX_VNDIRTY))
2927 panic("bqrelse: not dirty");
2928 if ((bp->b_flags & B_NOREUSE) != 0) {
2929 brelse(bp);
2930 return;
2931 }
2932 qindex = QUEUE_CLEAN;
2933 }
2934 buf_track(bp, __func__);
2935 /* binsfree unlocks bp. */
2936 binsfree(bp, qindex);
2937 return;
2938
2939 out:
2940 buf_track(bp, __func__);
2941 /* unlock */
2942 BUF_UNLOCK(bp);
2943 }
2944
2945 /*
2946 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2947 * restore bogus pages.
2948 */
2949 static void
vfs_vmio_iodone(struct buf * bp)2950 vfs_vmio_iodone(struct buf *bp)
2951 {
2952 vm_ooffset_t foff;
2953 vm_page_t m;
2954 vm_object_t obj;
2955 struct vnode *vp __unused;
2956 int i, iosize, resid;
2957 bool bogus;
2958
2959 obj = bp->b_bufobj->bo_object;
2960 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2961 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2962 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2963
2964 vp = bp->b_vp;
2965 VNPASS(vp->v_holdcnt > 0, vp);
2966 VNPASS(vp->v_object != NULL, vp);
2967
2968 foff = bp->b_offset;
2969 KASSERT(bp->b_offset != NOOFFSET,
2970 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2971
2972 bogus = false;
2973 iosize = bp->b_bcount - bp->b_resid;
2974 for (i = 0; i < bp->b_npages; i++) {
2975 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2976 if (resid > iosize)
2977 resid = iosize;
2978
2979 /*
2980 * cleanup bogus pages, restoring the originals
2981 */
2982 m = bp->b_pages[i];
2983 if (m == bogus_page) {
2984 bogus = true;
2985 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2986 if (m == NULL)
2987 panic("biodone: page disappeared!");
2988 bp->b_pages[i] = m;
2989 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2990 /*
2991 * In the write case, the valid and clean bits are
2992 * already changed correctly ( see bdwrite() ), so we
2993 * only need to do this here in the read case.
2994 */
2995 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2996 resid)) == 0, ("vfs_vmio_iodone: page %p "
2997 "has unexpected dirty bits", m));
2998 vfs_page_set_valid(bp, foff, m);
2999 }
3000 KASSERT(OFF_TO_IDX(foff) == m->pindex,
3001 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
3002 (intmax_t)foff, (uintmax_t)m->pindex));
3003
3004 vm_page_sunbusy(m);
3005 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3006 iosize -= resid;
3007 }
3008 vm_object_pip_wakeupn(obj, bp->b_npages);
3009 if (bogus && buf_mapped(bp)) {
3010 BUF_CHECK_MAPPED(bp);
3011 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3012 bp->b_pages, bp->b_npages);
3013 }
3014 }
3015
3016 /*
3017 * Perform page invalidation when a buffer is released. The fully invalid
3018 * pages will be reclaimed later in vfs_vmio_truncate().
3019 */
3020 static void
vfs_vmio_invalidate(struct buf * bp)3021 vfs_vmio_invalidate(struct buf *bp)
3022 {
3023 vm_object_t obj;
3024 vm_page_t m;
3025 int flags, i, resid, poffset, presid;
3026
3027 if (buf_mapped(bp)) {
3028 BUF_CHECK_MAPPED(bp);
3029 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3030 } else
3031 BUF_CHECK_UNMAPPED(bp);
3032 /*
3033 * Get the base offset and length of the buffer. Note that
3034 * in the VMIO case if the buffer block size is not
3035 * page-aligned then b_data pointer may not be page-aligned.
3036 * But our b_pages[] array *IS* page aligned.
3037 *
3038 * block sizes less then DEV_BSIZE (usually 512) are not
3039 * supported due to the page granularity bits (m->valid,
3040 * m->dirty, etc...).
3041 *
3042 * See man buf(9) for more information
3043 */
3044 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3045 obj = bp->b_bufobj->bo_object;
3046 resid = bp->b_bufsize;
3047 poffset = bp->b_offset & PAGE_MASK;
3048 VM_OBJECT_WLOCK(obj);
3049 for (i = 0; i < bp->b_npages; i++) {
3050 m = bp->b_pages[i];
3051 if (m == bogus_page)
3052 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3053 bp->b_pages[i] = NULL;
3054
3055 presid = resid > (PAGE_SIZE - poffset) ?
3056 (PAGE_SIZE - poffset) : resid;
3057 KASSERT(presid >= 0, ("brelse: extra page"));
3058 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3059 if (pmap_page_wired_mappings(m) == 0)
3060 vm_page_set_invalid(m, poffset, presid);
3061 vm_page_sunbusy(m);
3062 vm_page_release_locked(m, flags);
3063 resid -= presid;
3064 poffset = 0;
3065 }
3066 VM_OBJECT_WUNLOCK(obj);
3067 bp->b_npages = 0;
3068 }
3069
3070 /*
3071 * Page-granular truncation of an existing VMIO buffer.
3072 */
3073 static void
vfs_vmio_truncate(struct buf * bp,int desiredpages)3074 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3075 {
3076 vm_object_t obj;
3077 vm_page_t m;
3078 int flags, i;
3079
3080 if (bp->b_npages == desiredpages)
3081 return;
3082
3083 if (buf_mapped(bp)) {
3084 BUF_CHECK_MAPPED(bp);
3085 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3086 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3087 } else
3088 BUF_CHECK_UNMAPPED(bp);
3089
3090 /*
3091 * The object lock is needed only if we will attempt to free pages.
3092 */
3093 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3094 if ((bp->b_flags & B_DIRECT) != 0) {
3095 flags |= VPR_TRYFREE;
3096 obj = bp->b_bufobj->bo_object;
3097 VM_OBJECT_WLOCK(obj);
3098 } else {
3099 obj = NULL;
3100 }
3101 for (i = desiredpages; i < bp->b_npages; i++) {
3102 m = bp->b_pages[i];
3103 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3104 bp->b_pages[i] = NULL;
3105 if (obj != NULL)
3106 vm_page_release_locked(m, flags);
3107 else
3108 vm_page_release(m, flags);
3109 }
3110 if (obj != NULL)
3111 VM_OBJECT_WUNLOCK(obj);
3112 bp->b_npages = desiredpages;
3113 }
3114
3115 /*
3116 * Byte granular extension of VMIO buffers.
3117 */
3118 static void
vfs_vmio_extend(struct buf * bp,int desiredpages,int size)3119 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3120 {
3121 /*
3122 * We are growing the buffer, possibly in a
3123 * byte-granular fashion.
3124 */
3125 vm_object_t obj;
3126 vm_offset_t toff;
3127 vm_offset_t tinc;
3128 vm_page_t m;
3129
3130 /*
3131 * Step 1, bring in the VM pages from the object, allocating
3132 * them if necessary. We must clear B_CACHE if these pages
3133 * are not valid for the range covered by the buffer.
3134 */
3135 obj = bp->b_bufobj->bo_object;
3136 if (bp->b_npages < desiredpages) {
3137 KASSERT(desiredpages <= atop(maxbcachebuf),
3138 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3139 bp, desiredpages, maxbcachebuf));
3140
3141 /*
3142 * We must allocate system pages since blocking
3143 * here could interfere with paging I/O, no
3144 * matter which process we are.
3145 *
3146 * Only exclusive busy can be tested here.
3147 * Blocking on shared busy might lead to
3148 * deadlocks once allocbuf() is called after
3149 * pages are vfs_busy_pages().
3150 */
3151 (void)vm_page_grab_pages_unlocked(obj,
3152 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3153 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3154 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3155 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3156 bp->b_npages = desiredpages;
3157 }
3158
3159 /*
3160 * Step 2. We've loaded the pages into the buffer,
3161 * we have to figure out if we can still have B_CACHE
3162 * set. Note that B_CACHE is set according to the
3163 * byte-granular range ( bcount and size ), not the
3164 * aligned range ( newbsize ).
3165 *
3166 * The VM test is against m->valid, which is DEV_BSIZE
3167 * aligned. Needless to say, the validity of the data
3168 * needs to also be DEV_BSIZE aligned. Note that this
3169 * fails with NFS if the server or some other client
3170 * extends the file's EOF. If our buffer is resized,
3171 * B_CACHE may remain set! XXX
3172 */
3173 toff = bp->b_bcount;
3174 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3175 while ((bp->b_flags & B_CACHE) && toff < size) {
3176 vm_pindex_t pi;
3177
3178 if (tinc > (size - toff))
3179 tinc = size - toff;
3180 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3181 m = bp->b_pages[pi];
3182 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3183 toff += tinc;
3184 tinc = PAGE_SIZE;
3185 }
3186
3187 /*
3188 * Step 3, fixup the KVA pmap.
3189 */
3190 if (buf_mapped(bp))
3191 bpmap_qenter(bp);
3192 else
3193 BUF_CHECK_UNMAPPED(bp);
3194 }
3195
3196 /*
3197 * Check to see if a block at a particular lbn is available for a clustered
3198 * write.
3199 */
3200 static int
vfs_bio_clcheck(struct vnode * vp,int size,daddr_t lblkno,daddr_t blkno)3201 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3202 {
3203 struct buf *bpa;
3204 int match;
3205
3206 match = 0;
3207
3208 /* If the buf isn't in core skip it */
3209 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3210 return (0);
3211
3212 /* If the buf is busy we don't want to wait for it */
3213 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3214 return (0);
3215
3216 /* Only cluster with valid clusterable delayed write buffers */
3217 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3218 (B_DELWRI | B_CLUSTEROK))
3219 goto done;
3220
3221 if (bpa->b_bufsize != size)
3222 goto done;
3223
3224 /*
3225 * Check to see if it is in the expected place on disk and that the
3226 * block has been mapped.
3227 */
3228 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3229 match = 1;
3230 done:
3231 BUF_UNLOCK(bpa);
3232 return (match);
3233 }
3234
3235 /*
3236 * vfs_bio_awrite:
3237 *
3238 * Implement clustered async writes for clearing out B_DELWRI buffers.
3239 * This is much better then the old way of writing only one buffer at
3240 * a time. Note that we may not be presented with the buffers in the
3241 * correct order, so we search for the cluster in both directions.
3242 */
3243 int
vfs_bio_awrite(struct buf * bp)3244 vfs_bio_awrite(struct buf *bp)
3245 {
3246 struct bufobj *bo;
3247 int i;
3248 int j;
3249 daddr_t lblkno = bp->b_lblkno;
3250 struct vnode *vp = bp->b_vp;
3251 int ncl;
3252 int nwritten;
3253 int size;
3254 int maxcl;
3255 int gbflags;
3256
3257 bo = &vp->v_bufobj;
3258 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3259 /*
3260 * right now we support clustered writing only to regular files. If
3261 * we find a clusterable block we could be in the middle of a cluster
3262 * rather then at the beginning.
3263 */
3264 if ((vp->v_type == VREG) &&
3265 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3266 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3267 size = vp->v_mount->mnt_stat.f_iosize;
3268 maxcl = maxphys / size;
3269
3270 BO_RLOCK(bo);
3271 for (i = 1; i < maxcl; i++)
3272 if (vfs_bio_clcheck(vp, size, lblkno + i,
3273 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3274 break;
3275
3276 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3277 if (vfs_bio_clcheck(vp, size, lblkno - j,
3278 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3279 break;
3280 BO_RUNLOCK(bo);
3281 --j;
3282 ncl = i + j;
3283 /*
3284 * this is a possible cluster write
3285 */
3286 if (ncl != 1) {
3287 BUF_UNLOCK(bp);
3288 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3289 gbflags);
3290 return (nwritten);
3291 }
3292 }
3293 bremfree(bp);
3294 bp->b_flags |= B_ASYNC;
3295 /*
3296 * default (old) behavior, writing out only one block
3297 *
3298 * XXX returns b_bufsize instead of b_bcount for nwritten?
3299 */
3300 nwritten = bp->b_bufsize;
3301 (void) bwrite(bp);
3302
3303 return (nwritten);
3304 }
3305
3306 /*
3307 * getnewbuf_kva:
3308 *
3309 * Allocate KVA for an empty buf header according to gbflags.
3310 */
3311 static int
getnewbuf_kva(struct buf * bp,int gbflags,int maxsize)3312 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3313 {
3314
3315 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3316 /*
3317 * In order to keep fragmentation sane we only allocate kva
3318 * in BKVASIZE chunks. XXX with vmem we can do page size.
3319 */
3320 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3321
3322 if (maxsize != bp->b_kvasize &&
3323 bufkva_alloc(bp, maxsize, gbflags))
3324 return (ENOSPC);
3325 }
3326 return (0);
3327 }
3328
3329 /*
3330 * getnewbuf:
3331 *
3332 * Find and initialize a new buffer header, freeing up existing buffers
3333 * in the bufqueues as necessary. The new buffer is returned locked.
3334 *
3335 * We block if:
3336 * We have insufficient buffer headers
3337 * We have insufficient buffer space
3338 * buffer_arena is too fragmented ( space reservation fails )
3339 * If we have to flush dirty buffers ( but we try to avoid this )
3340 *
3341 * The caller is responsible for releasing the reserved bufspace after
3342 * allocbuf() is called.
3343 */
3344 static struct buf *
getnewbuf(struct vnode * vp,int slpflag,int slptimeo,int maxsize,int gbflags)3345 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3346 {
3347 struct bufdomain *bd;
3348 struct buf *bp;
3349 bool metadata, reserved;
3350
3351 bp = NULL;
3352 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3353 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3354 if (!unmapped_buf_allowed)
3355 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3356
3357 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3358 vp->v_type == VCHR)
3359 metadata = true;
3360 else
3361 metadata = false;
3362 if (vp == NULL)
3363 bd = &bdomain[0];
3364 else
3365 bd = &bdomain[vp->v_bufobj.bo_domain];
3366
3367 counter_u64_add(getnewbufcalls, 1);
3368 reserved = false;
3369 do {
3370 if (reserved == false &&
3371 bufspace_reserve(bd, maxsize, metadata) != 0) {
3372 counter_u64_add(getnewbufrestarts, 1);
3373 continue;
3374 }
3375 reserved = true;
3376 if ((bp = buf_alloc(bd)) == NULL) {
3377 counter_u64_add(getnewbufrestarts, 1);
3378 continue;
3379 }
3380 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3381 return (bp);
3382 break;
3383 } while (buf_recycle(bd, false) == 0);
3384
3385 if (reserved)
3386 bufspace_release(bd, maxsize);
3387 if (bp != NULL) {
3388 bp->b_flags |= B_INVAL;
3389 brelse(bp);
3390 }
3391 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3392
3393 return (NULL);
3394 }
3395
3396 /*
3397 * buf_daemon:
3398 *
3399 * buffer flushing daemon. Buffers are normally flushed by the
3400 * update daemon but if it cannot keep up this process starts to
3401 * take the load in an attempt to prevent getnewbuf() from blocking.
3402 */
3403 static struct kproc_desc buf_kp = {
3404 "bufdaemon",
3405 buf_daemon,
3406 &bufdaemonproc
3407 };
3408 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3409
3410 static int
buf_flush(struct vnode * vp,struct bufdomain * bd,int target)3411 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3412 {
3413 int flushed;
3414
3415 flushed = flushbufqueues(vp, bd, target, 0);
3416 if (flushed == 0) {
3417 /*
3418 * Could not find any buffers without rollback
3419 * dependencies, so just write the first one
3420 * in the hopes of eventually making progress.
3421 */
3422 if (vp != NULL && target > 2)
3423 target /= 2;
3424 flushbufqueues(vp, bd, target, 1);
3425 }
3426 return (flushed);
3427 }
3428
3429 static void
buf_daemon_shutdown(void * arg __unused,int howto __unused)3430 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3431 {
3432 int error;
3433
3434 if (KERNEL_PANICKED())
3435 return;
3436
3437 mtx_lock(&bdlock);
3438 bd_shutdown = true;
3439 wakeup(&bd_request);
3440 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3441 60 * hz);
3442 mtx_unlock(&bdlock);
3443 if (error != 0)
3444 printf("bufdaemon wait error: %d\n", error);
3445 }
3446
3447 static void
buf_daemon(void)3448 buf_daemon(void)
3449 {
3450 struct bufdomain *bd;
3451 int speedupreq;
3452 int lodirty;
3453 int i;
3454
3455 /*
3456 * This process needs to be suspended prior to shutdown sync.
3457 */
3458 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3459 SHUTDOWN_PRI_LAST + 100);
3460
3461 /*
3462 * Start the buf clean daemons as children threads.
3463 */
3464 for (i = 0 ; i < buf_domains; i++) {
3465 int error;
3466
3467 error = kthread_add((void (*)(void *))bufspace_daemon,
3468 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3469 if (error)
3470 panic("error %d spawning bufspace daemon", error);
3471 }
3472
3473 /*
3474 * This process is allowed to take the buffer cache to the limit
3475 */
3476 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3477 mtx_lock(&bdlock);
3478 while (!bd_shutdown) {
3479 bd_request = 0;
3480 mtx_unlock(&bdlock);
3481
3482 /*
3483 * Save speedupreq for this pass and reset to capture new
3484 * requests.
3485 */
3486 speedupreq = bd_speedupreq;
3487 bd_speedupreq = 0;
3488
3489 /*
3490 * Flush each domain sequentially according to its level and
3491 * the speedup request.
3492 */
3493 for (i = 0; i < buf_domains; i++) {
3494 bd = &bdomain[i];
3495 if (speedupreq)
3496 lodirty = bd->bd_numdirtybuffers / 2;
3497 else
3498 lodirty = bd->bd_lodirtybuffers;
3499 while (bd->bd_numdirtybuffers > lodirty) {
3500 if (buf_flush(NULL, bd,
3501 bd->bd_numdirtybuffers - lodirty) == 0)
3502 break;
3503 kern_yield(PRI_USER);
3504 }
3505 }
3506
3507 /*
3508 * Only clear bd_request if we have reached our low water
3509 * mark. The buf_daemon normally waits 1 second and
3510 * then incrementally flushes any dirty buffers that have
3511 * built up, within reason.
3512 *
3513 * If we were unable to hit our low water mark and couldn't
3514 * find any flushable buffers, we sleep for a short period
3515 * to avoid endless loops on unlockable buffers.
3516 */
3517 mtx_lock(&bdlock);
3518 if (bd_shutdown)
3519 break;
3520 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3521 /*
3522 * We reached our low water mark, reset the
3523 * request and sleep until we are needed again.
3524 * The sleep is just so the suspend code works.
3525 */
3526 bd_request = 0;
3527 /*
3528 * Do an extra wakeup in case dirty threshold
3529 * changed via sysctl and the explicit transition
3530 * out of shortfall was missed.
3531 */
3532 bdirtywakeup();
3533 if (runningbufspace <= lorunningspace)
3534 runningwakeup();
3535 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3536 } else {
3537 /*
3538 * We couldn't find any flushable dirty buffers but
3539 * still have too many dirty buffers, we
3540 * have to sleep and try again. (rare)
3541 */
3542 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3543 }
3544 }
3545 wakeup(&bd_shutdown);
3546 mtx_unlock(&bdlock);
3547 kthread_exit();
3548 }
3549
3550 /*
3551 * flushbufqueues:
3552 *
3553 * Try to flush a buffer in the dirty queue. We must be careful to
3554 * free up B_INVAL buffers instead of write them, which NFS is
3555 * particularly sensitive to.
3556 */
3557 static int flushwithdeps = 0;
3558 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3559 &flushwithdeps, 0,
3560 "Number of buffers flushed with dependencies that require rollbacks");
3561
3562 static int
flushbufqueues(struct vnode * lvp,struct bufdomain * bd,int target,int flushdeps)3563 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3564 int flushdeps)
3565 {
3566 struct bufqueue *bq;
3567 struct buf *sentinel;
3568 struct vnode *vp;
3569 struct mount *mp;
3570 struct buf *bp;
3571 int hasdeps;
3572 int flushed;
3573 int error;
3574 bool unlock;
3575
3576 flushed = 0;
3577 bq = &bd->bd_dirtyq;
3578 bp = NULL;
3579 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3580 sentinel->b_qindex = QUEUE_SENTINEL;
3581 BQ_LOCK(bq);
3582 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3583 BQ_UNLOCK(bq);
3584 while (flushed != target) {
3585 maybe_yield();
3586 BQ_LOCK(bq);
3587 bp = TAILQ_NEXT(sentinel, b_freelist);
3588 if (bp != NULL) {
3589 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3590 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3591 b_freelist);
3592 } else {
3593 BQ_UNLOCK(bq);
3594 break;
3595 }
3596 /*
3597 * Skip sentinels inserted by other invocations of the
3598 * flushbufqueues(), taking care to not reorder them.
3599 *
3600 * Only flush the buffers that belong to the
3601 * vnode locked by the curthread.
3602 */
3603 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3604 bp->b_vp != lvp)) {
3605 BQ_UNLOCK(bq);
3606 continue;
3607 }
3608 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3609 BQ_UNLOCK(bq);
3610 if (error != 0)
3611 continue;
3612
3613 /*
3614 * BKGRDINPROG can only be set with the buf and bufobj
3615 * locks both held. We tolerate a race to clear it here.
3616 */
3617 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3618 (bp->b_flags & B_DELWRI) == 0) {
3619 BUF_UNLOCK(bp);
3620 continue;
3621 }
3622 if (bp->b_flags & B_INVAL) {
3623 bremfreef(bp);
3624 brelse(bp);
3625 flushed++;
3626 continue;
3627 }
3628
3629 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3630 if (flushdeps == 0) {
3631 BUF_UNLOCK(bp);
3632 continue;
3633 }
3634 hasdeps = 1;
3635 } else
3636 hasdeps = 0;
3637 /*
3638 * We must hold the lock on a vnode before writing
3639 * one of its buffers. Otherwise we may confuse, or
3640 * in the case of a snapshot vnode, deadlock the
3641 * system.
3642 *
3643 * The lock order here is the reverse of the normal
3644 * of vnode followed by buf lock. This is ok because
3645 * the NOWAIT will prevent deadlock.
3646 */
3647 vp = bp->b_vp;
3648 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3649 BUF_UNLOCK(bp);
3650 continue;
3651 }
3652 if (lvp == NULL) {
3653 unlock = true;
3654 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3655 } else {
3656 ASSERT_VOP_LOCKED(vp, "getbuf");
3657 unlock = false;
3658 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3659 vn_lock(vp, LK_TRYUPGRADE);
3660 }
3661 if (error == 0) {
3662 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3663 bp, bp->b_vp, bp->b_flags);
3664 if (curproc == bufdaemonproc) {
3665 vfs_bio_awrite(bp);
3666 } else {
3667 bremfree(bp);
3668 bwrite(bp);
3669 counter_u64_add(notbufdflushes, 1);
3670 }
3671 vn_finished_write(mp);
3672 if (unlock)
3673 VOP_UNLOCK(vp);
3674 flushwithdeps += hasdeps;
3675 flushed++;
3676
3677 /*
3678 * Sleeping on runningbufspace while holding
3679 * vnode lock leads to deadlock.
3680 */
3681 if (curproc == bufdaemonproc &&
3682 runningbufspace > hirunningspace)
3683 waitrunningbufspace();
3684 continue;
3685 }
3686 vn_finished_write(mp);
3687 BUF_UNLOCK(bp);
3688 }
3689 BQ_LOCK(bq);
3690 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3691 BQ_UNLOCK(bq);
3692 free(sentinel, M_TEMP);
3693 return (flushed);
3694 }
3695
3696 /*
3697 * Check to see if a block is currently memory resident.
3698 */
3699 struct buf *
incore(struct bufobj * bo,daddr_t blkno)3700 incore(struct bufobj *bo, daddr_t blkno)
3701 {
3702 return (gbincore_unlocked(bo, blkno));
3703 }
3704
3705 /*
3706 * Returns true if no I/O is needed to access the
3707 * associated VM object. This is like incore except
3708 * it also hunts around in the VM system for the data.
3709 */
3710 bool
inmem(struct vnode * vp,daddr_t blkno)3711 inmem(struct vnode * vp, daddr_t blkno)
3712 {
3713 vm_object_t obj;
3714 vm_offset_t toff, tinc, size;
3715 vm_page_t m, n;
3716 vm_ooffset_t off;
3717 int valid;
3718
3719 ASSERT_VOP_LOCKED(vp, "inmem");
3720
3721 if (incore(&vp->v_bufobj, blkno))
3722 return (true);
3723 if (vp->v_mount == NULL)
3724 return (false);
3725 obj = vp->v_object;
3726 if (obj == NULL)
3727 return (false);
3728
3729 size = PAGE_SIZE;
3730 if (size > vp->v_mount->mnt_stat.f_iosize)
3731 size = vp->v_mount->mnt_stat.f_iosize;
3732 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3733
3734 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3735 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3736 recheck:
3737 if (m == NULL)
3738 return (false);
3739
3740 tinc = size;
3741 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3742 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3743 /*
3744 * Consider page validity only if page mapping didn't change
3745 * during the check.
3746 */
3747 valid = vm_page_is_valid(m,
3748 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3749 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3750 if (m != n) {
3751 m = n;
3752 goto recheck;
3753 }
3754 if (!valid)
3755 return (false);
3756 }
3757 return (true);
3758 }
3759
3760 /*
3761 * Set the dirty range for a buffer based on the status of the dirty
3762 * bits in the pages comprising the buffer. The range is limited
3763 * to the size of the buffer.
3764 *
3765 * Tell the VM system that the pages associated with this buffer
3766 * are clean. This is used for delayed writes where the data is
3767 * going to go to disk eventually without additional VM intevention.
3768 *
3769 * Note that while we only really need to clean through to b_bcount, we
3770 * just go ahead and clean through to b_bufsize.
3771 */
3772 static void
vfs_clean_pages_dirty_buf(struct buf * bp)3773 vfs_clean_pages_dirty_buf(struct buf *bp)
3774 {
3775 vm_ooffset_t foff, noff, eoff;
3776 vm_page_t m;
3777 int i;
3778
3779 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3780 return;
3781
3782 foff = bp->b_offset;
3783 KASSERT(bp->b_offset != NOOFFSET,
3784 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3785
3786 vfs_busy_pages_acquire(bp);
3787 vfs_setdirty_range(bp);
3788 for (i = 0; i < bp->b_npages; i++) {
3789 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3790 eoff = noff;
3791 if (eoff > bp->b_offset + bp->b_bufsize)
3792 eoff = bp->b_offset + bp->b_bufsize;
3793 m = bp->b_pages[i];
3794 vfs_page_set_validclean(bp, foff, m);
3795 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3796 foff = noff;
3797 }
3798 vfs_busy_pages_release(bp);
3799 }
3800
3801 static void
vfs_setdirty_range(struct buf * bp)3802 vfs_setdirty_range(struct buf *bp)
3803 {
3804 vm_offset_t boffset;
3805 vm_offset_t eoffset;
3806 int i;
3807
3808 /*
3809 * test the pages to see if they have been modified directly
3810 * by users through the VM system.
3811 */
3812 for (i = 0; i < bp->b_npages; i++)
3813 vm_page_test_dirty(bp->b_pages[i]);
3814
3815 /*
3816 * Calculate the encompassing dirty range, boffset and eoffset,
3817 * (eoffset - boffset) bytes.
3818 */
3819
3820 for (i = 0; i < bp->b_npages; i++) {
3821 if (bp->b_pages[i]->dirty)
3822 break;
3823 }
3824 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3825
3826 for (i = bp->b_npages - 1; i >= 0; --i) {
3827 if (bp->b_pages[i]->dirty) {
3828 break;
3829 }
3830 }
3831 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3832
3833 /*
3834 * Fit it to the buffer.
3835 */
3836
3837 if (eoffset > bp->b_bcount)
3838 eoffset = bp->b_bcount;
3839
3840 /*
3841 * If we have a good dirty range, merge with the existing
3842 * dirty range.
3843 */
3844
3845 if (boffset < eoffset) {
3846 if (bp->b_dirtyoff > boffset)
3847 bp->b_dirtyoff = boffset;
3848 if (bp->b_dirtyend < eoffset)
3849 bp->b_dirtyend = eoffset;
3850 }
3851 }
3852
3853 /*
3854 * Allocate the KVA mapping for an existing buffer.
3855 * If an unmapped buffer is provided but a mapped buffer is requested, take
3856 * also care to properly setup mappings between pages and KVA.
3857 */
3858 static void
bp_unmapped_get_kva(struct buf * bp,daddr_t blkno,int size,int gbflags)3859 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3860 {
3861 int bsize, maxsize, need_mapping, need_kva;
3862 off_t offset;
3863
3864 need_mapping = bp->b_data == unmapped_buf &&
3865 (gbflags & GB_UNMAPPED) == 0;
3866 need_kva = bp->b_kvabase == unmapped_buf &&
3867 bp->b_data == unmapped_buf &&
3868 (gbflags & GB_KVAALLOC) != 0;
3869 if (!need_mapping && !need_kva)
3870 return;
3871
3872 BUF_CHECK_UNMAPPED(bp);
3873
3874 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3875 /*
3876 * Buffer is not mapped, but the KVA was already
3877 * reserved at the time of the instantiation. Use the
3878 * allocated space.
3879 */
3880 goto has_addr;
3881 }
3882
3883 /*
3884 * Calculate the amount of the address space we would reserve
3885 * if the buffer was mapped.
3886 */
3887 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3888 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3889 offset = blkno * bsize;
3890 maxsize = size + (offset & PAGE_MASK);
3891 maxsize = imax(maxsize, bsize);
3892
3893 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3894 if ((gbflags & GB_NOWAIT_BD) != 0) {
3895 /*
3896 * XXXKIB: defragmentation cannot
3897 * succeed, not sure what else to do.
3898 */
3899 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3900 }
3901 counter_u64_add(mappingrestarts, 1);
3902 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3903 }
3904 has_addr:
3905 if (need_mapping) {
3906 /* b_offset is handled by bpmap_qenter. */
3907 bp->b_data = bp->b_kvabase;
3908 BUF_CHECK_MAPPED(bp);
3909 bpmap_qenter(bp);
3910 }
3911 }
3912
3913 struct buf *
getblk(struct vnode * vp,daddr_t blkno,int size,int slpflag,int slptimeo,int flags)3914 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3915 int flags)
3916 {
3917 struct buf *bp;
3918 int error;
3919
3920 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3921 if (error != 0)
3922 return (NULL);
3923 return (bp);
3924 }
3925
3926 /*
3927 * getblkx:
3928 *
3929 * Get a block given a specified block and offset into a file/device.
3930 * The buffers B_DONE bit will be cleared on return, making it almost
3931 * ready for an I/O initiation. B_INVAL may or may not be set on
3932 * return. The caller should clear B_INVAL prior to initiating a
3933 * READ.
3934 *
3935 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3936 * an existing buffer.
3937 *
3938 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3939 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3940 * and then cleared based on the backing VM. If the previous buffer is
3941 * non-0-sized but invalid, B_CACHE will be cleared.
3942 *
3943 * If getblk() must create a new buffer, the new buffer is returned with
3944 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3945 * case it is returned with B_INVAL clear and B_CACHE set based on the
3946 * backing VM.
3947 *
3948 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3949 * B_CACHE bit is clear.
3950 *
3951 * What this means, basically, is that the caller should use B_CACHE to
3952 * determine whether the buffer is fully valid or not and should clear
3953 * B_INVAL prior to issuing a read. If the caller intends to validate
3954 * the buffer by loading its data area with something, the caller needs
3955 * to clear B_INVAL. If the caller does this without issuing an I/O,
3956 * the caller should set B_CACHE ( as an optimization ), else the caller
3957 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3958 * a write attempt or if it was a successful read. If the caller
3959 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3960 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3961 *
3962 * The blkno parameter is the logical block being requested. Normally
3963 * the mapping of logical block number to disk block address is done
3964 * by calling VOP_BMAP(). However, if the mapping is already known, the
3965 * disk block address can be passed using the dblkno parameter. If the
3966 * disk block address is not known, then the same value should be passed
3967 * for blkno and dblkno.
3968 */
3969 int
getblkx(struct vnode * vp,daddr_t blkno,daddr_t dblkno,int size,int slpflag,int slptimeo,int flags,struct buf ** bpp)3970 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3971 int slptimeo, int flags, struct buf **bpp)
3972 {
3973 struct buf *bp;
3974 struct bufobj *bo;
3975 daddr_t d_blkno;
3976 int bsize, error, maxsize, vmio;
3977 off_t offset;
3978
3979 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3980 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3981 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3982 if (vp->v_type != VCHR)
3983 ASSERT_VOP_LOCKED(vp, "getblk");
3984 if (size > maxbcachebuf) {
3985 printf("getblkx: size(%d) > maxbcachebuf(%d)\n", size,
3986 maxbcachebuf);
3987 return (EIO);
3988 }
3989 if (!unmapped_buf_allowed)
3990 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3991
3992 bo = &vp->v_bufobj;
3993 d_blkno = dblkno;
3994
3995 /* Attempt lockless lookup first. */
3996 bp = gbincore_unlocked(bo, blkno);
3997 if (bp == NULL) {
3998 /*
3999 * With GB_NOCREAT we must be sure about not finding the buffer
4000 * as it may have been reassigned during unlocked lookup.
4001 * If BO_NONSTERILE is still unset, no reassign has occurred.
4002 */
4003 if ((flags & GB_NOCREAT) != 0) {
4004 /* Ensure bo_flag is loaded after gbincore_unlocked. */
4005 atomic_thread_fence_acq();
4006 if ((bo->bo_flag & BO_NONSTERILE) == 0)
4007 return (EEXIST);
4008 goto loop;
4009 }
4010 goto newbuf_unlocked;
4011 }
4012
4013 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
4014 0);
4015 if (error != 0) {
4016 KASSERT(error == EBUSY,
4017 ("getblk: unexpected error %d from buf try-lock", error));
4018 /*
4019 * We failed a buf try-lock.
4020 *
4021 * With GB_LOCK_NOWAIT, just return, rather than taking the
4022 * bufobj interlock and trying again, since we would probably
4023 * fail again anyway. This is okay even if the buf's identity
4024 * changed and we contended on the wrong lock, as changing
4025 * identity itself requires the buf lock, and we could have
4026 * contended on the right lock.
4027 */
4028 if ((flags & GB_LOCK_NOWAIT) != 0)
4029 return (error);
4030 goto loop;
4031 }
4032
4033 /* Verify buf identify has not changed since lookup. */
4034 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
4035 goto foundbuf_fastpath;
4036
4037 /* It changed, fallback to locked lookup. */
4038 BUF_UNLOCK_RAW(bp);
4039
4040 /* As above, with GB_LOCK_NOWAIT, just return. */
4041 if ((flags & GB_LOCK_NOWAIT) != 0)
4042 return (EBUSY);
4043
4044 loop:
4045 BO_RLOCK(bo);
4046 bp = gbincore(bo, blkno);
4047 if (bp != NULL) {
4048 int lockflags;
4049
4050 /*
4051 * Buffer is in-core. If the buffer is not busy nor managed,
4052 * it must be on a queue.
4053 */
4054 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4055 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4056 #ifdef WITNESS
4057 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4058 #endif
4059
4060 error = BUF_TIMELOCK(bp, lockflags,
4061 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4062
4063 /*
4064 * If we slept and got the lock we have to restart in case
4065 * the buffer changed identities.
4066 */
4067 if (error == ENOLCK)
4068 goto loop;
4069 /* We timed out or were interrupted. */
4070 else if (error != 0)
4071 return (error);
4072
4073 foundbuf_fastpath:
4074 /* If recursed, assume caller knows the rules. */
4075 if (BUF_LOCKRECURSED(bp))
4076 goto end;
4077
4078 /*
4079 * The buffer is locked. B_CACHE is cleared if the buffer is
4080 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4081 * and for a VMIO buffer B_CACHE is adjusted according to the
4082 * backing VM cache.
4083 */
4084 if (bp->b_flags & B_INVAL)
4085 bp->b_flags &= ~B_CACHE;
4086 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4087 bp->b_flags |= B_CACHE;
4088 if (bp->b_flags & B_MANAGED)
4089 MPASS(bp->b_qindex == QUEUE_NONE);
4090 else
4091 bremfree(bp);
4092
4093 /*
4094 * check for size inconsistencies for non-VMIO case.
4095 */
4096 if (bp->b_bcount != size) {
4097 if ((bp->b_flags & B_VMIO) == 0 ||
4098 (size > bp->b_kvasize)) {
4099 if (bp->b_flags & B_DELWRI) {
4100 bp->b_flags |= B_NOCACHE;
4101 bwrite(bp);
4102 } else {
4103 if (LIST_EMPTY(&bp->b_dep)) {
4104 bp->b_flags |= B_RELBUF;
4105 brelse(bp);
4106 } else {
4107 bp->b_flags |= B_NOCACHE;
4108 bwrite(bp);
4109 }
4110 }
4111 goto loop;
4112 }
4113 }
4114
4115 /*
4116 * Handle the case of unmapped buffer which should
4117 * become mapped, or the buffer for which KVA
4118 * reservation is requested.
4119 */
4120 bp_unmapped_get_kva(bp, blkno, size, flags);
4121
4122 /*
4123 * If the size is inconsistent in the VMIO case, we can resize
4124 * the buffer. This might lead to B_CACHE getting set or
4125 * cleared. If the size has not changed, B_CACHE remains
4126 * unchanged from its previous state.
4127 */
4128 allocbuf(bp, size);
4129
4130 KASSERT(bp->b_offset != NOOFFSET,
4131 ("getblk: no buffer offset"));
4132
4133 /*
4134 * A buffer with B_DELWRI set and B_CACHE clear must
4135 * be committed before we can return the buffer in
4136 * order to prevent the caller from issuing a read
4137 * ( due to B_CACHE not being set ) and overwriting
4138 * it.
4139 *
4140 * Most callers, including NFS and FFS, need this to
4141 * operate properly either because they assume they
4142 * can issue a read if B_CACHE is not set, or because
4143 * ( for example ) an uncached B_DELWRI might loop due
4144 * to softupdates re-dirtying the buffer. In the latter
4145 * case, B_CACHE is set after the first write completes,
4146 * preventing further loops.
4147 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4148 * above while extending the buffer, we cannot allow the
4149 * buffer to remain with B_CACHE set after the write
4150 * completes or it will represent a corrupt state. To
4151 * deal with this we set B_NOCACHE to scrap the buffer
4152 * after the write.
4153 *
4154 * We might be able to do something fancy, like setting
4155 * B_CACHE in bwrite() except if B_DELWRI is already set,
4156 * so the below call doesn't set B_CACHE, but that gets real
4157 * confusing. This is much easier.
4158 */
4159
4160 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4161 bp->b_flags |= B_NOCACHE;
4162 bwrite(bp);
4163 goto loop;
4164 }
4165 bp->b_flags &= ~B_DONE;
4166 } else {
4167 /*
4168 * Buffer is not in-core, create new buffer. The buffer
4169 * returned by getnewbuf() is locked. Note that the returned
4170 * buffer is also considered valid (not marked B_INVAL).
4171 */
4172 BO_RUNLOCK(bo);
4173 newbuf_unlocked:
4174 /*
4175 * If the user does not want us to create the buffer, bail out
4176 * here.
4177 */
4178 if (flags & GB_NOCREAT)
4179 return (EEXIST);
4180
4181 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4182 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4183 offset = blkno * bsize;
4184 vmio = vp->v_object != NULL;
4185 if (vmio) {
4186 maxsize = size + (offset & PAGE_MASK);
4187 if (maxsize > maxbcachebuf) {
4188 printf(
4189 "getblkx: maxsize(%d) > maxbcachebuf(%d)\n",
4190 maxsize, maxbcachebuf);
4191 return (EIO);
4192 }
4193 } else {
4194 maxsize = size;
4195 /* Do not allow non-VMIO notmapped buffers. */
4196 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4197 }
4198 maxsize = imax(maxsize, bsize);
4199 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4200 !vn_isdisk(vp)) {
4201 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4202 KASSERT(error != EOPNOTSUPP,
4203 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4204 vp));
4205 if (error != 0)
4206 return (error);
4207 if (d_blkno == -1)
4208 return (EJUSTRETURN);
4209 }
4210
4211 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4212 if (bp == NULL) {
4213 if (slpflag || slptimeo)
4214 return (ETIMEDOUT);
4215 /*
4216 * XXX This is here until the sleep path is diagnosed
4217 * enough to work under very low memory conditions.
4218 *
4219 * There's an issue on low memory, 4BSD+non-preempt
4220 * systems (eg MIPS routers with 32MB RAM) where buffer
4221 * exhaustion occurs without sleeping for buffer
4222 * reclaimation. This just sticks in a loop and
4223 * constantly attempts to allocate a buffer, which
4224 * hits exhaustion and tries to wakeup bufdaemon.
4225 * This never happens because we never yield.
4226 *
4227 * The real solution is to identify and fix these cases
4228 * so we aren't effectively busy-waiting in a loop
4229 * until the reclaimation path has cycles to run.
4230 */
4231 kern_yield(PRI_USER);
4232 goto loop;
4233 }
4234
4235 /*
4236 *
4237 * Insert the buffer into the hash, so that it can
4238 * be found by incore.
4239 *
4240 * We don't hold the bufobj interlock while allocating the new
4241 * buffer. Consequently, we can race on buffer creation. This
4242 * can be a problem whether the vnode is locked or not. If the
4243 * buffer is created out from under us, we have to throw away
4244 * the one we just created.
4245 */
4246 bp->b_lblkno = blkno;
4247 bp->b_blkno = d_blkno;
4248 bp->b_offset = offset;
4249 error = bgetvp(vp, bp);
4250 if (error != 0) {
4251 KASSERT(error == EEXIST,
4252 ("getblk: unexpected error %d from bgetvp",
4253 error));
4254 bp->b_flags |= B_INVAL;
4255 bufspace_release(bufdomain(bp), maxsize);
4256 brelse(bp);
4257 goto loop;
4258 }
4259
4260 /*
4261 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4262 * buffer size starts out as 0, B_CACHE will be set by
4263 * allocbuf() for the VMIO case prior to it testing the
4264 * backing store for validity.
4265 */
4266
4267 if (vmio) {
4268 bp->b_flags |= B_VMIO;
4269 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4270 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4271 bp, vp->v_object, bp->b_bufobj->bo_object));
4272 } else {
4273 bp->b_flags &= ~B_VMIO;
4274 KASSERT(bp->b_bufobj->bo_object == NULL,
4275 ("ARGH! has b_bufobj->bo_object %p %p\n",
4276 bp, bp->b_bufobj->bo_object));
4277 BUF_CHECK_MAPPED(bp);
4278 }
4279
4280 allocbuf(bp, size);
4281 bufspace_release(bufdomain(bp), maxsize);
4282 bp->b_flags &= ~B_DONE;
4283 }
4284 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4285 end:
4286 buf_track(bp, __func__);
4287 KASSERT(bp->b_bufobj == bo,
4288 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4289 *bpp = bp;
4290 return (0);
4291 }
4292
4293 /*
4294 * Get an empty, disassociated buffer of given size. The buffer is initially
4295 * set to B_INVAL.
4296 */
4297 struct buf *
geteblk(int size,int flags)4298 geteblk(int size, int flags)
4299 {
4300 struct buf *bp;
4301 int maxsize;
4302
4303 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4304 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4305 if ((flags & GB_NOWAIT_BD) &&
4306 (curthread->td_pflags & TDP_BUFNEED) != 0)
4307 return (NULL);
4308 }
4309 allocbuf(bp, size);
4310 bufspace_release(bufdomain(bp), maxsize);
4311 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4312 return (bp);
4313 }
4314
4315 /*
4316 * Truncate the backing store for a non-vmio buffer.
4317 */
4318 static void
vfs_nonvmio_truncate(struct buf * bp,int newbsize)4319 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4320 {
4321
4322 if (bp->b_flags & B_MALLOC) {
4323 /*
4324 * malloced buffers are not shrunk
4325 */
4326 if (newbsize == 0) {
4327 bufmallocadjust(bp, 0);
4328 free(bp->b_data, M_BIOBUF);
4329 bp->b_data = bp->b_kvabase;
4330 bp->b_flags &= ~B_MALLOC;
4331 }
4332 return;
4333 }
4334 vm_hold_free_pages(bp, newbsize);
4335 bufspace_adjust(bp, newbsize);
4336 }
4337
4338 /*
4339 * Extend the backing for a non-VMIO buffer.
4340 */
4341 static void
vfs_nonvmio_extend(struct buf * bp,int newbsize)4342 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4343 {
4344 caddr_t origbuf;
4345 int origbufsize;
4346
4347 /*
4348 * We only use malloced memory on the first allocation.
4349 * and revert to page-allocated memory when the buffer
4350 * grows.
4351 *
4352 * There is a potential smp race here that could lead
4353 * to bufmallocspace slightly passing the max. It
4354 * is probably extremely rare and not worth worrying
4355 * over.
4356 */
4357 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4358 bufmallocspace < maxbufmallocspace) {
4359 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4360 bp->b_flags |= B_MALLOC;
4361 bufmallocadjust(bp, newbsize);
4362 return;
4363 }
4364
4365 /*
4366 * If the buffer is growing on its other-than-first
4367 * allocation then we revert to the page-allocation
4368 * scheme.
4369 */
4370 origbuf = NULL;
4371 origbufsize = 0;
4372 if (bp->b_flags & B_MALLOC) {
4373 origbuf = bp->b_data;
4374 origbufsize = bp->b_bufsize;
4375 bp->b_data = bp->b_kvabase;
4376 bufmallocadjust(bp, 0);
4377 bp->b_flags &= ~B_MALLOC;
4378 newbsize = round_page(newbsize);
4379 }
4380 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4381 (vm_offset_t) bp->b_data + newbsize);
4382 if (origbuf != NULL) {
4383 bcopy(origbuf, bp->b_data, origbufsize);
4384 free(origbuf, M_BIOBUF);
4385 }
4386 bufspace_adjust(bp, newbsize);
4387 }
4388
4389 /*
4390 * This code constitutes the buffer memory from either anonymous system
4391 * memory (in the case of non-VMIO operations) or from an associated
4392 * VM object (in the case of VMIO operations). This code is able to
4393 * resize a buffer up or down.
4394 *
4395 * Note that this code is tricky, and has many complications to resolve
4396 * deadlock or inconsistent data situations. Tread lightly!!!
4397 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4398 * the caller. Calling this code willy nilly can result in the loss of data.
4399 *
4400 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4401 * B_CACHE for the non-VMIO case.
4402 */
4403 int
allocbuf(struct buf * bp,int size)4404 allocbuf(struct buf *bp, int size)
4405 {
4406 int newbsize;
4407
4408 if (bp->b_bcount == size)
4409 return (1);
4410
4411 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4412 ("allocbuf: buffer too small %p %#x %#x",
4413 bp, bp->b_kvasize, size));
4414
4415 newbsize = roundup2(size, DEV_BSIZE);
4416 if ((bp->b_flags & B_VMIO) == 0) {
4417 if ((bp->b_flags & B_MALLOC) == 0)
4418 newbsize = round_page(newbsize);
4419 /*
4420 * Just get anonymous memory from the kernel. Don't
4421 * mess with B_CACHE.
4422 */
4423 if (newbsize < bp->b_bufsize)
4424 vfs_nonvmio_truncate(bp, newbsize);
4425 else if (newbsize > bp->b_bufsize)
4426 vfs_nonvmio_extend(bp, newbsize);
4427 } else {
4428 int desiredpages;
4429
4430 desiredpages = size == 0 ? 0 :
4431 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4432
4433 KASSERT((bp->b_flags & B_MALLOC) == 0,
4434 ("allocbuf: VMIO buffer can't be malloced %p", bp));
4435
4436 /*
4437 * Set B_CACHE initially if buffer is 0 length or will become
4438 * 0-length.
4439 */
4440 if (size == 0 || bp->b_bufsize == 0)
4441 bp->b_flags |= B_CACHE;
4442
4443 if (newbsize < bp->b_bufsize)
4444 vfs_vmio_truncate(bp, desiredpages);
4445 /* XXX This looks as if it should be newbsize > b_bufsize */
4446 else if (size > bp->b_bcount)
4447 vfs_vmio_extend(bp, desiredpages, size);
4448 bufspace_adjust(bp, newbsize);
4449 }
4450 bp->b_bcount = size; /* requested buffer size. */
4451 return (1);
4452 }
4453
4454 extern int inflight_transient_maps;
4455
4456 static struct bio_queue nondump_bios;
4457
4458 void
biodone(struct bio * bp)4459 biodone(struct bio *bp)
4460 {
4461 struct mtx *mtxp;
4462 void (*done)(struct bio *);
4463 vm_offset_t start, end;
4464
4465 biotrack(bp, __func__);
4466
4467 /*
4468 * Avoid completing I/O when dumping after a panic since that may
4469 * result in a deadlock in the filesystem or pager code. Note that
4470 * this doesn't affect dumps that were started manually since we aim
4471 * to keep the system usable after it has been resumed.
4472 */
4473 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4474 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4475 return;
4476 }
4477 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4478 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4479 bp->bio_flags |= BIO_UNMAPPED;
4480 start = trunc_page((vm_offset_t)bp->bio_data);
4481 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4482 bp->bio_data = unmapped_buf;
4483 pmap_qremove(start, atop(end - start));
4484 vmem_free(transient_arena, start, end - start);
4485 atomic_add_int(&inflight_transient_maps, -1);
4486 }
4487 done = bp->bio_done;
4488 /*
4489 * The check for done == biodone is to allow biodone to be
4490 * used as a bio_done routine.
4491 */
4492 if (done == NULL || done == biodone) {
4493 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4494 mtx_lock(mtxp);
4495 bp->bio_flags |= BIO_DONE;
4496 wakeup(bp);
4497 mtx_unlock(mtxp);
4498 } else
4499 done(bp);
4500 }
4501
4502 /*
4503 * Wait for a BIO to finish.
4504 */
4505 int
biowait(struct bio * bp,const char * wmesg)4506 biowait(struct bio *bp, const char *wmesg)
4507 {
4508 struct mtx *mtxp;
4509
4510 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4511 mtx_lock(mtxp);
4512 while ((bp->bio_flags & BIO_DONE) == 0)
4513 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4514 mtx_unlock(mtxp);
4515 if (bp->bio_error != 0)
4516 return (bp->bio_error);
4517 if (!(bp->bio_flags & BIO_ERROR))
4518 return (0);
4519 return (EIO);
4520 }
4521
4522 void
biofinish(struct bio * bp,struct devstat * stat,int error)4523 biofinish(struct bio *bp, struct devstat *stat, int error)
4524 {
4525
4526 if (error) {
4527 bp->bio_error = error;
4528 bp->bio_flags |= BIO_ERROR;
4529 }
4530 if (stat != NULL)
4531 devstat_end_transaction_bio(stat, bp);
4532 biodone(bp);
4533 }
4534
4535 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4536 void
biotrack_buf(struct bio * bp,const char * location)4537 biotrack_buf(struct bio *bp, const char *location)
4538 {
4539
4540 buf_track(bp->bio_track_bp, location);
4541 }
4542 #endif
4543
4544 /*
4545 * bufwait:
4546 *
4547 * Wait for buffer I/O completion, returning error status. The buffer
4548 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4549 * error and cleared.
4550 */
4551 int
bufwait(struct buf * bp)4552 bufwait(struct buf *bp)
4553 {
4554 if (bp->b_iocmd == BIO_READ)
4555 bwait(bp, PRIBIO, "biord");
4556 else
4557 bwait(bp, PRIBIO, "biowr");
4558 if (bp->b_flags & B_EINTR) {
4559 bp->b_flags &= ~B_EINTR;
4560 return (EINTR);
4561 }
4562 if (bp->b_ioflags & BIO_ERROR) {
4563 return (bp->b_error ? bp->b_error : EIO);
4564 } else {
4565 return (0);
4566 }
4567 }
4568
4569 /*
4570 * bufdone:
4571 *
4572 * Finish I/O on a buffer, optionally calling a completion function.
4573 * This is usually called from an interrupt so process blocking is
4574 * not allowed.
4575 *
4576 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4577 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4578 * assuming B_INVAL is clear.
4579 *
4580 * For the VMIO case, we set B_CACHE if the op was a read and no
4581 * read error occurred, or if the op was a write. B_CACHE is never
4582 * set if the buffer is invalid or otherwise uncacheable.
4583 *
4584 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4585 * initiator to leave B_INVAL set to brelse the buffer out of existence
4586 * in the biodone routine.
4587 */
4588 void
bufdone(struct buf * bp)4589 bufdone(struct buf *bp)
4590 {
4591 struct bufobj *dropobj;
4592 void (*biodone)(struct buf *);
4593
4594 buf_track(bp, __func__);
4595 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4596 dropobj = NULL;
4597
4598 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4599
4600 runningbufwakeup(bp);
4601 if (bp->b_iocmd == BIO_WRITE)
4602 dropobj = bp->b_bufobj;
4603 /* call optional completion function if requested */
4604 if (bp->b_iodone != NULL) {
4605 biodone = bp->b_iodone;
4606 bp->b_iodone = NULL;
4607 (*biodone) (bp);
4608 if (dropobj)
4609 bufobj_wdrop(dropobj);
4610 return;
4611 }
4612 if (bp->b_flags & B_VMIO) {
4613 /*
4614 * Set B_CACHE if the op was a normal read and no error
4615 * occurred. B_CACHE is set for writes in the b*write()
4616 * routines.
4617 */
4618 if (bp->b_iocmd == BIO_READ &&
4619 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4620 !(bp->b_ioflags & BIO_ERROR))
4621 bp->b_flags |= B_CACHE;
4622 vfs_vmio_iodone(bp);
4623 }
4624 if (!LIST_EMPTY(&bp->b_dep))
4625 buf_complete(bp);
4626 if ((bp->b_flags & B_CKHASH) != 0) {
4627 KASSERT(bp->b_iocmd == BIO_READ,
4628 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4629 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4630 (*bp->b_ckhashcalc)(bp);
4631 }
4632 /*
4633 * For asynchronous completions, release the buffer now. The brelse
4634 * will do a wakeup there if necessary - so no need to do a wakeup
4635 * here in the async case. The sync case always needs to do a wakeup.
4636 */
4637 if (bp->b_flags & B_ASYNC) {
4638 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4639 (bp->b_ioflags & BIO_ERROR))
4640 brelse(bp);
4641 else
4642 bqrelse(bp);
4643 } else
4644 bdone(bp);
4645 if (dropobj)
4646 bufobj_wdrop(dropobj);
4647 }
4648
4649 /*
4650 * This routine is called in lieu of iodone in the case of
4651 * incomplete I/O. This keeps the busy status for pages
4652 * consistent.
4653 */
4654 void
vfs_unbusy_pages(struct buf * bp)4655 vfs_unbusy_pages(struct buf *bp)
4656 {
4657 int i;
4658 vm_object_t obj;
4659 vm_page_t m;
4660
4661 runningbufwakeup(bp);
4662 if (!(bp->b_flags & B_VMIO))
4663 return;
4664
4665 obj = bp->b_bufobj->bo_object;
4666 for (i = 0; i < bp->b_npages; i++) {
4667 m = bp->b_pages[i];
4668 if (m == bogus_page) {
4669 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4670 if (!m)
4671 panic("vfs_unbusy_pages: page missing\n");
4672 bp->b_pages[i] = m;
4673 if (buf_mapped(bp)) {
4674 BUF_CHECK_MAPPED(bp);
4675 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4676 bp->b_pages, bp->b_npages);
4677 } else
4678 BUF_CHECK_UNMAPPED(bp);
4679 }
4680 vm_page_sunbusy(m);
4681 }
4682 vm_object_pip_wakeupn(obj, bp->b_npages);
4683 }
4684
4685 /*
4686 * vfs_page_set_valid:
4687 *
4688 * Set the valid bits in a page based on the supplied offset. The
4689 * range is restricted to the buffer's size.
4690 *
4691 * This routine is typically called after a read completes.
4692 */
4693 static void
vfs_page_set_valid(struct buf * bp,vm_ooffset_t off,vm_page_t m)4694 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4695 {
4696 vm_ooffset_t eoff;
4697
4698 /*
4699 * Compute the end offset, eoff, such that [off, eoff) does not span a
4700 * page boundary and eoff is not greater than the end of the buffer.
4701 * The end of the buffer, in this case, is our file EOF, not the
4702 * allocation size of the buffer.
4703 */
4704 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4705 if (eoff > bp->b_offset + bp->b_bcount)
4706 eoff = bp->b_offset + bp->b_bcount;
4707
4708 /*
4709 * Set valid range. This is typically the entire buffer and thus the
4710 * entire page.
4711 */
4712 if (eoff > off)
4713 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4714 }
4715
4716 /*
4717 * vfs_page_set_validclean:
4718 *
4719 * Set the valid bits and clear the dirty bits in a page based on the
4720 * supplied offset. The range is restricted to the buffer's size.
4721 */
4722 static void
vfs_page_set_validclean(struct buf * bp,vm_ooffset_t off,vm_page_t m)4723 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4724 {
4725 vm_ooffset_t soff, eoff;
4726
4727 /*
4728 * Start and end offsets in buffer. eoff - soff may not cross a
4729 * page boundary or cross the end of the buffer. The end of the
4730 * buffer, in this case, is our file EOF, not the allocation size
4731 * of the buffer.
4732 */
4733 soff = off;
4734 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4735 if (eoff > bp->b_offset + bp->b_bcount)
4736 eoff = bp->b_offset + bp->b_bcount;
4737
4738 /*
4739 * Set valid range. This is typically the entire buffer and thus the
4740 * entire page.
4741 */
4742 if (eoff > soff) {
4743 vm_page_set_validclean(
4744 m,
4745 (vm_offset_t) (soff & PAGE_MASK),
4746 (vm_offset_t) (eoff - soff)
4747 );
4748 }
4749 }
4750
4751 /*
4752 * Acquire a shared busy on all pages in the buf.
4753 */
4754 void
vfs_busy_pages_acquire(struct buf * bp)4755 vfs_busy_pages_acquire(struct buf *bp)
4756 {
4757 int i;
4758
4759 for (i = 0; i < bp->b_npages; i++)
4760 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4761 }
4762
4763 void
vfs_busy_pages_release(struct buf * bp)4764 vfs_busy_pages_release(struct buf *bp)
4765 {
4766 int i;
4767
4768 for (i = 0; i < bp->b_npages; i++)
4769 vm_page_sunbusy(bp->b_pages[i]);
4770 }
4771
4772 /*
4773 * This routine is called before a device strategy routine.
4774 * It is used to tell the VM system that paging I/O is in
4775 * progress, and treat the pages associated with the buffer
4776 * almost as being exclusive busy. Also the object paging_in_progress
4777 * flag is handled to make sure that the object doesn't become
4778 * inconsistent.
4779 *
4780 * Since I/O has not been initiated yet, certain buffer flags
4781 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4782 * and should be ignored.
4783 */
4784 void
vfs_busy_pages(struct buf * bp,int clear_modify)4785 vfs_busy_pages(struct buf *bp, int clear_modify)
4786 {
4787 vm_object_t obj;
4788 vm_ooffset_t foff;
4789 vm_page_t m;
4790 int i;
4791 bool bogus;
4792
4793 if (!(bp->b_flags & B_VMIO))
4794 return;
4795
4796 obj = bp->b_bufobj->bo_object;
4797 foff = bp->b_offset;
4798 KASSERT(bp->b_offset != NOOFFSET,
4799 ("vfs_busy_pages: no buffer offset"));
4800 if ((bp->b_flags & B_CLUSTER) == 0) {
4801 vm_object_pip_add(obj, bp->b_npages);
4802 vfs_busy_pages_acquire(bp);
4803 }
4804 if (bp->b_bufsize != 0)
4805 vfs_setdirty_range(bp);
4806 bogus = false;
4807 for (i = 0; i < bp->b_npages; i++) {
4808 m = bp->b_pages[i];
4809 vm_page_assert_sbusied(m);
4810
4811 /*
4812 * When readying a buffer for a read ( i.e
4813 * clear_modify == 0 ), it is important to do
4814 * bogus_page replacement for valid pages in
4815 * partially instantiated buffers. Partially
4816 * instantiated buffers can, in turn, occur when
4817 * reconstituting a buffer from its VM backing store
4818 * base. We only have to do this if B_CACHE is
4819 * clear ( which causes the I/O to occur in the
4820 * first place ). The replacement prevents the read
4821 * I/O from overwriting potentially dirty VM-backed
4822 * pages. XXX bogus page replacement is, uh, bogus.
4823 * It may not work properly with small-block devices.
4824 * We need to find a better way.
4825 */
4826 if (clear_modify) {
4827 pmap_remove_write(m);
4828 vfs_page_set_validclean(bp, foff, m);
4829 } else if (vm_page_all_valid(m) &&
4830 (bp->b_flags & B_CACHE) == 0) {
4831 bp->b_pages[i] = bogus_page;
4832 bogus = true;
4833 }
4834 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4835 }
4836 if (bogus && buf_mapped(bp)) {
4837 BUF_CHECK_MAPPED(bp);
4838 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4839 bp->b_pages, bp->b_npages);
4840 }
4841 }
4842
4843 /*
4844 * vfs_bio_set_valid:
4845 *
4846 * Set the range within the buffer to valid. The range is
4847 * relative to the beginning of the buffer, b_offset. Note that
4848 * b_offset itself may be offset from the beginning of the first
4849 * page.
4850 */
4851 void
vfs_bio_set_valid(struct buf * bp,int base,int size)4852 vfs_bio_set_valid(struct buf *bp, int base, int size)
4853 {
4854 int i, n;
4855 vm_page_t m;
4856
4857 if (!(bp->b_flags & B_VMIO))
4858 return;
4859
4860 /*
4861 * Fixup base to be relative to beginning of first page.
4862 * Set initial n to be the maximum number of bytes in the
4863 * first page that can be validated.
4864 */
4865 base += (bp->b_offset & PAGE_MASK);
4866 n = PAGE_SIZE - (base & PAGE_MASK);
4867
4868 /*
4869 * Busy may not be strictly necessary here because the pages are
4870 * unlikely to be fully valid and the vnode lock will synchronize
4871 * their access via getpages. It is grabbed for consistency with
4872 * other page validation.
4873 */
4874 vfs_busy_pages_acquire(bp);
4875 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4876 m = bp->b_pages[i];
4877 if (n > size)
4878 n = size;
4879 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4880 base += n;
4881 size -= n;
4882 n = PAGE_SIZE;
4883 }
4884 vfs_busy_pages_release(bp);
4885 }
4886
4887 /*
4888 * vfs_bio_clrbuf:
4889 *
4890 * If the specified buffer is a non-VMIO buffer, clear the entire
4891 * buffer. If the specified buffer is a VMIO buffer, clear and
4892 * validate only the previously invalid portions of the buffer.
4893 * This routine essentially fakes an I/O, so we need to clear
4894 * BIO_ERROR and B_INVAL.
4895 *
4896 * Note that while we only theoretically need to clear through b_bcount,
4897 * we go ahead and clear through b_bufsize.
4898 */
4899 void
vfs_bio_clrbuf(struct buf * bp)4900 vfs_bio_clrbuf(struct buf *bp)
4901 {
4902 int i, j, sa, ea, slide, zbits;
4903 vm_page_bits_t mask;
4904
4905 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4906 clrbuf(bp);
4907 return;
4908 }
4909 bp->b_flags &= ~B_INVAL;
4910 bp->b_ioflags &= ~BIO_ERROR;
4911 vfs_busy_pages_acquire(bp);
4912 sa = bp->b_offset & PAGE_MASK;
4913 slide = 0;
4914 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4915 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4916 ea = slide & PAGE_MASK;
4917 if (ea == 0)
4918 ea = PAGE_SIZE;
4919 if (bp->b_pages[i] == bogus_page)
4920 continue;
4921 j = sa / DEV_BSIZE;
4922 zbits = (sizeof(vm_page_bits_t) * NBBY) -
4923 (ea - sa) / DEV_BSIZE;
4924 mask = (VM_PAGE_BITS_ALL >> zbits) << j;
4925 if ((bp->b_pages[i]->valid & mask) == mask)
4926 continue;
4927 if ((bp->b_pages[i]->valid & mask) == 0)
4928 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4929 else {
4930 for (; sa < ea; sa += DEV_BSIZE, j++) {
4931 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4932 pmap_zero_page_area(bp->b_pages[i],
4933 sa, DEV_BSIZE);
4934 }
4935 }
4936 }
4937 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4938 roundup2(ea - sa, DEV_BSIZE));
4939 }
4940 vfs_busy_pages_release(bp);
4941 bp->b_resid = 0;
4942 }
4943
4944 void
vfs_bio_bzero_buf(struct buf * bp,int base,int size)4945 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4946 {
4947 vm_page_t m;
4948 int i, n;
4949
4950 if (buf_mapped(bp)) {
4951 BUF_CHECK_MAPPED(bp);
4952 bzero(bp->b_data + base, size);
4953 } else {
4954 BUF_CHECK_UNMAPPED(bp);
4955 n = PAGE_SIZE - (base & PAGE_MASK);
4956 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4957 m = bp->b_pages[i];
4958 if (n > size)
4959 n = size;
4960 pmap_zero_page_area(m, base & PAGE_MASK, n);
4961 base += n;
4962 size -= n;
4963 n = PAGE_SIZE;
4964 }
4965 }
4966 }
4967
4968 /*
4969 * Update buffer flags based on I/O request parameters, optionally releasing the
4970 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4971 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4972 * I/O). Otherwise the buffer is released to the cache.
4973 */
4974 static void
b_io_dismiss(struct buf * bp,int ioflag,bool release)4975 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4976 {
4977
4978 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4979 ("buf %p non-VMIO noreuse", bp));
4980
4981 if ((ioflag & IO_DIRECT) != 0)
4982 bp->b_flags |= B_DIRECT;
4983 if ((ioflag & IO_EXT) != 0)
4984 bp->b_xflags |= BX_ALTDATA;
4985 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4986 bp->b_flags |= B_RELBUF;
4987 if ((ioflag & IO_NOREUSE) != 0)
4988 bp->b_flags |= B_NOREUSE;
4989 if (release)
4990 brelse(bp);
4991 } else if (release)
4992 bqrelse(bp);
4993 }
4994
4995 void
vfs_bio_brelse(struct buf * bp,int ioflag)4996 vfs_bio_brelse(struct buf *bp, int ioflag)
4997 {
4998
4999 b_io_dismiss(bp, ioflag, true);
5000 }
5001
5002 void
vfs_bio_set_flags(struct buf * bp,int ioflag)5003 vfs_bio_set_flags(struct buf *bp, int ioflag)
5004 {
5005
5006 b_io_dismiss(bp, ioflag, false);
5007 }
5008
5009 /*
5010 * vm_hold_load_pages and vm_hold_free_pages get pages into
5011 * a buffers address space. The pages are anonymous and are
5012 * not associated with a file object.
5013 */
5014 static void
vm_hold_load_pages(struct buf * bp,vm_offset_t from,vm_offset_t to)5015 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
5016 {
5017 vm_offset_t pg;
5018 vm_page_t p;
5019 int index;
5020
5021 BUF_CHECK_MAPPED(bp);
5022
5023 to = round_page(to);
5024 from = round_page(from);
5025 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5026 MPASS((bp->b_flags & B_MAXPHYS) == 0);
5027 KASSERT(to - from <= maxbcachebuf,
5028 ("vm_hold_load_pages too large %p %#jx %#jx %u",
5029 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
5030
5031 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
5032 /*
5033 * note: must allocate system pages since blocking here
5034 * could interfere with paging I/O, no matter which
5035 * process we are.
5036 */
5037 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
5038 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
5039 pmap_qenter(pg, &p, 1);
5040 bp->b_pages[index] = p;
5041 }
5042 bp->b_npages = index;
5043 }
5044
5045 /* Return pages associated with this buf to the vm system */
5046 static void
vm_hold_free_pages(struct buf * bp,int newbsize)5047 vm_hold_free_pages(struct buf *bp, int newbsize)
5048 {
5049 vm_offset_t from;
5050 vm_page_t p;
5051 int index, newnpages;
5052
5053 BUF_CHECK_MAPPED(bp);
5054
5055 from = round_page((vm_offset_t)bp->b_data + newbsize);
5056 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5057 if (bp->b_npages > newnpages)
5058 pmap_qremove(from, bp->b_npages - newnpages);
5059 for (index = newnpages; index < bp->b_npages; index++) {
5060 p = bp->b_pages[index];
5061 bp->b_pages[index] = NULL;
5062 vm_page_unwire_noq(p);
5063 vm_page_free(p);
5064 }
5065 bp->b_npages = newnpages;
5066 }
5067
5068 /*
5069 * Map an IO request into kernel virtual address space.
5070 *
5071 * All requests are (re)mapped into kernel VA space.
5072 * Notice that we use b_bufsize for the size of the buffer
5073 * to be mapped. b_bcount might be modified by the driver.
5074 *
5075 * Note that even if the caller determines that the address space should
5076 * be valid, a race or a smaller-file mapped into a larger space may
5077 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5078 * check the return value.
5079 *
5080 * This function only works with pager buffers.
5081 */
5082 int
vmapbuf(struct buf * bp,void * uaddr,size_t len,int mapbuf)5083 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5084 {
5085 vm_prot_t prot;
5086 int pidx;
5087
5088 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5089 prot = VM_PROT_READ;
5090 if (bp->b_iocmd == BIO_READ)
5091 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5092 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5093 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5094 if (pidx < 0)
5095 return (-1);
5096 bp->b_bufsize = len;
5097 bp->b_npages = pidx;
5098 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5099 if (mapbuf || !unmapped_buf_allowed) {
5100 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5101 bp->b_data = bp->b_kvabase + bp->b_offset;
5102 } else
5103 bp->b_data = unmapped_buf;
5104 return (0);
5105 }
5106
5107 /*
5108 * Free the io map PTEs associated with this IO operation.
5109 * We also invalidate the TLB entries and restore the original b_addr.
5110 *
5111 * This function only works with pager buffers.
5112 */
5113 void
vunmapbuf(struct buf * bp)5114 vunmapbuf(struct buf *bp)
5115 {
5116 int npages;
5117
5118 npages = bp->b_npages;
5119 if (buf_mapped(bp))
5120 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5121 vm_page_unhold_pages(bp->b_pages, npages);
5122
5123 bp->b_data = unmapped_buf;
5124 }
5125
5126 void
bdone(struct buf * bp)5127 bdone(struct buf *bp)
5128 {
5129 struct mtx *mtxp;
5130
5131 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5132 mtx_lock(mtxp);
5133 bp->b_flags |= B_DONE;
5134 wakeup(bp);
5135 mtx_unlock(mtxp);
5136 }
5137
5138 void
bwait(struct buf * bp,u_char pri,const char * wchan)5139 bwait(struct buf *bp, u_char pri, const char *wchan)
5140 {
5141 struct mtx *mtxp;
5142
5143 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5144 mtx_lock(mtxp);
5145 while ((bp->b_flags & B_DONE) == 0)
5146 msleep(bp, mtxp, pri, wchan, 0);
5147 mtx_unlock(mtxp);
5148 }
5149
5150 int
bufsync(struct bufobj * bo,int waitfor)5151 bufsync(struct bufobj *bo, int waitfor)
5152 {
5153
5154 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5155 }
5156
5157 void
bufstrategy(struct bufobj * bo,struct buf * bp)5158 bufstrategy(struct bufobj *bo, struct buf *bp)
5159 {
5160 int i __unused;
5161 struct vnode *vp;
5162
5163 vp = bp->b_vp;
5164 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5165 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5166 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5167 i = VOP_STRATEGY(vp, bp);
5168 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5169 }
5170
5171 /*
5172 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5173 */
5174 void
bufobj_init(struct bufobj * bo,void * private)5175 bufobj_init(struct bufobj *bo, void *private)
5176 {
5177 static volatile int bufobj_cleanq;
5178
5179 bo->bo_domain =
5180 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5181 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5182 bo->bo_private = private;
5183 TAILQ_INIT(&bo->bo_clean.bv_hd);
5184 pctrie_init(&bo->bo_clean.bv_root);
5185 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5186 pctrie_init(&bo->bo_dirty.bv_root);
5187 }
5188
5189 void
bufobj_wrefl(struct bufobj * bo)5190 bufobj_wrefl(struct bufobj *bo)
5191 {
5192
5193 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5194 ASSERT_BO_WLOCKED(bo);
5195 bo->bo_numoutput++;
5196 }
5197
5198 void
bufobj_wref(struct bufobj * bo)5199 bufobj_wref(struct bufobj *bo)
5200 {
5201
5202 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5203 BO_LOCK(bo);
5204 bo->bo_numoutput++;
5205 BO_UNLOCK(bo);
5206 }
5207
5208 void
bufobj_wdrop(struct bufobj * bo)5209 bufobj_wdrop(struct bufobj *bo)
5210 {
5211
5212 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5213 BO_LOCK(bo);
5214 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5215 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5216 bo->bo_flag &= ~BO_WWAIT;
5217 wakeup(&bo->bo_numoutput);
5218 }
5219 BO_UNLOCK(bo);
5220 }
5221
5222 int
bufobj_wwait(struct bufobj * bo,int slpflag,int timeo)5223 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5224 {
5225 int error;
5226
5227 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5228 ASSERT_BO_WLOCKED(bo);
5229 error = 0;
5230 while (bo->bo_numoutput) {
5231 bo->bo_flag |= BO_WWAIT;
5232 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5233 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5234 if (error)
5235 break;
5236 }
5237 return (error);
5238 }
5239
5240 /*
5241 * Set bio_data or bio_ma for struct bio from the struct buf.
5242 */
5243 void
bdata2bio(struct buf * bp,struct bio * bip)5244 bdata2bio(struct buf *bp, struct bio *bip)
5245 {
5246
5247 if (!buf_mapped(bp)) {
5248 KASSERT(unmapped_buf_allowed, ("unmapped"));
5249 bip->bio_ma = bp->b_pages;
5250 bip->bio_ma_n = bp->b_npages;
5251 bip->bio_data = unmapped_buf;
5252 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5253 bip->bio_flags |= BIO_UNMAPPED;
5254 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5255 PAGE_SIZE == bp->b_npages,
5256 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5257 (long long)bip->bio_length, bip->bio_ma_n));
5258 } else {
5259 bip->bio_data = bp->b_data;
5260 bip->bio_ma = NULL;
5261 }
5262 }
5263
5264 struct memdesc
memdesc_bio(struct bio * bio)5265 memdesc_bio(struct bio *bio)
5266 {
5267 if ((bio->bio_flags & BIO_VLIST) != 0)
5268 return (memdesc_vlist((struct bus_dma_segment *)bio->bio_data,
5269 bio->bio_ma_n));
5270
5271 if ((bio->bio_flags & BIO_UNMAPPED) != 0)
5272 return (memdesc_vmpages(bio->bio_ma, bio->bio_bcount,
5273 bio->bio_ma_offset));
5274
5275 return (memdesc_vaddr(bio->bio_data, bio->bio_bcount));
5276 }
5277
5278 static int buf_pager_relbuf;
5279 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5280 &buf_pager_relbuf, 0,
5281 "Make buffer pager release buffers after reading");
5282
5283 /*
5284 * The buffer pager. It uses buffer reads to validate pages.
5285 *
5286 * In contrast to the generic local pager from vm/vnode_pager.c, this
5287 * pager correctly and easily handles volumes where the underlying
5288 * device block size is greater than the machine page size. The
5289 * buffer cache transparently extends the requested page run to be
5290 * aligned at the block boundary, and does the necessary bogus page
5291 * replacements in the addends to avoid obliterating already valid
5292 * pages.
5293 *
5294 * The only non-trivial issue is that the exclusive busy state for
5295 * pages, which is assumed by the vm_pager_getpages() interface, is
5296 * incompatible with the VMIO buffer cache's desire to share-busy the
5297 * pages. This function performs a trivial downgrade of the pages'
5298 * state before reading buffers, and a less trivial upgrade from the
5299 * shared-busy to excl-busy state after the read.
5300 */
5301 int
vfs_bio_getpages(struct vnode * vp,vm_page_t * ma,int count,int * rbehind,int * rahead,vbg_get_lblkno_t get_lblkno,vbg_get_blksize_t get_blksize)5302 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5303 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5304 vbg_get_blksize_t get_blksize)
5305 {
5306 vm_page_t m;
5307 vm_object_t object;
5308 struct buf *bp;
5309 struct mount *mp;
5310 daddr_t lbn, lbnp;
5311 vm_ooffset_t la, lb, poff, poffe;
5312 long bo_bs, bsize;
5313 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5314 bool redo, lpart;
5315
5316 object = vp->v_object;
5317 mp = vp->v_mount;
5318 error = 0;
5319 la = IDX_TO_OFF(ma[count - 1]->pindex);
5320 if (la >= object->un_pager.vnp.vnp_size)
5321 return (VM_PAGER_BAD);
5322
5323 /*
5324 * Change the meaning of la from where the last requested page starts
5325 * to where it ends, because that's the end of the requested region
5326 * and the start of the potential read-ahead region.
5327 */
5328 la += PAGE_SIZE;
5329 lpart = la > object->un_pager.vnp.vnp_size;
5330 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5331 &bo_bs);
5332 if (error != 0)
5333 return (VM_PAGER_ERROR);
5334
5335 /*
5336 * Calculate read-ahead, behind and total pages.
5337 */
5338 pgsin = count;
5339 lb = IDX_TO_OFF(ma[0]->pindex);
5340 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5341 pgsin += pgsin_b;
5342 if (rbehind != NULL)
5343 *rbehind = pgsin_b;
5344 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5345 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5346 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5347 PAGE_SIZE) - la);
5348 pgsin += pgsin_a;
5349 if (rahead != NULL)
5350 *rahead = pgsin_a;
5351 VM_CNT_INC(v_vnodein);
5352 VM_CNT_ADD(v_vnodepgsin, pgsin);
5353
5354 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5355 != 0) ? GB_UNMAPPED : 0;
5356 again:
5357 for (i = 0; i < count; i++) {
5358 if (ma[i] != bogus_page)
5359 vm_page_busy_downgrade(ma[i]);
5360 }
5361
5362 lbnp = -1;
5363 for (i = 0; i < count; i++) {
5364 m = ma[i];
5365 if (m == bogus_page)
5366 continue;
5367
5368 /*
5369 * Pages are shared busy and the object lock is not
5370 * owned, which together allow for the pages'
5371 * invalidation. The racy test for validity avoids
5372 * useless creation of the buffer for the most typical
5373 * case when invalidation is not used in redo or for
5374 * parallel read. The shared->excl upgrade loop at
5375 * the end of the function catches the race in a
5376 * reliable way (protected by the object lock).
5377 */
5378 if (vm_page_all_valid(m))
5379 continue;
5380
5381 poff = IDX_TO_OFF(m->pindex);
5382 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5383 for (; poff < poffe; poff += bsize) {
5384 lbn = get_lblkno(vp, poff);
5385 if (lbn == lbnp)
5386 goto next_page;
5387 lbnp = lbn;
5388
5389 error = get_blksize(vp, lbn, &bsize);
5390 if (error == 0)
5391 error = bread_gb(vp, lbn, bsize,
5392 curthread->td_ucred, br_flags, &bp);
5393 if (error != 0)
5394 goto end_pages;
5395 if (bp->b_rcred == curthread->td_ucred) {
5396 crfree(bp->b_rcred);
5397 bp->b_rcred = NOCRED;
5398 }
5399 if (LIST_EMPTY(&bp->b_dep)) {
5400 /*
5401 * Invalidation clears m->valid, but
5402 * may leave B_CACHE flag if the
5403 * buffer existed at the invalidation
5404 * time. In this case, recycle the
5405 * buffer to do real read on next
5406 * bread() after redo.
5407 *
5408 * Otherwise B_RELBUF is not strictly
5409 * necessary, enable to reduce buf
5410 * cache pressure.
5411 */
5412 if (buf_pager_relbuf ||
5413 !vm_page_all_valid(m))
5414 bp->b_flags |= B_RELBUF;
5415
5416 bp->b_flags &= ~B_NOCACHE;
5417 brelse(bp);
5418 } else {
5419 bqrelse(bp);
5420 }
5421 }
5422 KASSERT(1 /* racy, enable for debugging */ ||
5423 vm_page_all_valid(m) || i == count - 1,
5424 ("buf %d %p invalid", i, m));
5425 if (i == count - 1 && lpart) {
5426 if (!vm_page_none_valid(m) &&
5427 !vm_page_all_valid(m))
5428 vm_page_zero_invalid(m, TRUE);
5429 }
5430 next_page:;
5431 }
5432 end_pages:
5433
5434 redo = false;
5435 for (i = 0; i < count; i++) {
5436 if (ma[i] == bogus_page)
5437 continue;
5438 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5439 vm_page_sunbusy(ma[i]);
5440 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5441 VM_ALLOC_NORMAL);
5442 }
5443
5444 /*
5445 * Since the pages were only sbusy while neither the
5446 * buffer nor the object lock was held by us, or
5447 * reallocated while vm_page_grab() slept for busy
5448 * relinguish, they could have been invalidated.
5449 * Recheck the valid bits and re-read as needed.
5450 *
5451 * Note that the last page is made fully valid in the
5452 * read loop, and partial validity for the page at
5453 * index count - 1 could mean that the page was
5454 * invalidated or removed, so we must restart for
5455 * safety as well.
5456 */
5457 if (!vm_page_all_valid(ma[i]))
5458 redo = true;
5459 }
5460 if (redo && error == 0)
5461 goto again;
5462 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5463 }
5464
5465 #include "opt_ddb.h"
5466 #ifdef DDB
5467 #include <ddb/ddb.h>
5468
5469 /* DDB command to show buffer data */
DB_SHOW_COMMAND(buffer,db_show_buffer)5470 DB_SHOW_COMMAND(buffer, db_show_buffer)
5471 {
5472 /* get args */
5473 struct buf *bp = (struct buf *)addr;
5474 #ifdef FULL_BUF_TRACKING
5475 uint32_t i, j;
5476 #endif
5477
5478 if (!have_addr) {
5479 db_printf("usage: show buffer <addr>\n");
5480 return;
5481 }
5482
5483 db_printf("buf at %p\n", bp);
5484 db_printf("b_flags = 0x%b, b_xflags = 0x%b\n",
5485 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5486 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5487 db_printf("b_vflags = 0x%b, b_ioflags = 0x%b\n",
5488 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5489 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5490 db_printf(
5491 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5492 "b_bufobj = %p, b_data = %p\n"
5493 "b_blkno = %jd, b_lblkno = %jd, b_vp = %p, b_dep = %p\n",
5494 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5495 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5496 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5497 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5498 bp->b_kvabase, bp->b_kvasize);
5499 if (bp->b_npages) {
5500 int i;
5501 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5502 for (i = 0; i < bp->b_npages; i++) {
5503 vm_page_t m;
5504 m = bp->b_pages[i];
5505 if (m != NULL)
5506 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5507 (u_long)m->pindex,
5508 (u_long)VM_PAGE_TO_PHYS(m));
5509 else
5510 db_printf("( ??? )");
5511 if ((i + 1) < bp->b_npages)
5512 db_printf(",");
5513 }
5514 db_printf("\n");
5515 }
5516 BUF_LOCKPRINTINFO(bp);
5517 #if defined(FULL_BUF_TRACKING)
5518 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5519
5520 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5521 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5522 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5523 continue;
5524 db_printf(" %2u: %s\n", j,
5525 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5526 }
5527 #elif defined(BUF_TRACKING)
5528 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5529 #endif
5530 }
5531
DB_SHOW_COMMAND_FLAGS(bufqueues,bufqueues,DB_CMD_MEMSAFE)5532 DB_SHOW_COMMAND_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE)
5533 {
5534 struct bufdomain *bd;
5535 struct buf *bp;
5536 long total;
5537 int i, j, cnt;
5538
5539 db_printf("bqempty: %d\n", bqempty.bq_len);
5540
5541 for (i = 0; i < buf_domains; i++) {
5542 bd = &bdomain[i];
5543 db_printf("Buf domain %d\n", i);
5544 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5545 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5546 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5547 db_printf("\n");
5548 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5549 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5550 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5551 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5552 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5553 db_printf("\n");
5554 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5555 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5556 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5557 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5558 db_printf("\n");
5559 total = 0;
5560 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5561 total += bp->b_bufsize;
5562 db_printf("\tcleanq count\t%d (%ld)\n",
5563 bd->bd_cleanq->bq_len, total);
5564 total = 0;
5565 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5566 total += bp->b_bufsize;
5567 db_printf("\tdirtyq count\t%d (%ld)\n",
5568 bd->bd_dirtyq.bq_len, total);
5569 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5570 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5571 db_printf("\tCPU ");
5572 for (j = 0; j <= mp_maxid; j++)
5573 db_printf("%d, ", bd->bd_subq[j].bq_len);
5574 db_printf("\n");
5575 cnt = 0;
5576 total = 0;
5577 for (j = 0; j < nbuf; j++) {
5578 bp = nbufp(j);
5579 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5580 cnt++;
5581 total += bp->b_bufsize;
5582 }
5583 }
5584 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5585 cnt = 0;
5586 total = 0;
5587 for (j = 0; j < nbuf; j++) {
5588 bp = nbufp(j);
5589 if (bp->b_domain == i) {
5590 cnt++;
5591 total += bp->b_bufsize;
5592 }
5593 }
5594 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5595 }
5596 }
5597
DB_SHOW_COMMAND_FLAGS(lockedbufs,lockedbufs,DB_CMD_MEMSAFE)5598 DB_SHOW_COMMAND_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE)
5599 {
5600 struct buf *bp;
5601 int i;
5602
5603 for (i = 0; i < nbuf; i++) {
5604 bp = nbufp(i);
5605 if (BUF_ISLOCKED(bp)) {
5606 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5607 db_printf("\n");
5608 if (db_pager_quit)
5609 break;
5610 }
5611 }
5612 }
5613
DB_SHOW_COMMAND(vnodebufs,db_show_vnodebufs)5614 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5615 {
5616 struct vnode *vp;
5617 struct buf *bp;
5618
5619 if (!have_addr) {
5620 db_printf("usage: show vnodebufs <addr>\n");
5621 return;
5622 }
5623 vp = (struct vnode *)addr;
5624 db_printf("Clean buffers:\n");
5625 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5626 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5627 db_printf("\n");
5628 }
5629 db_printf("Dirty buffers:\n");
5630 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5631 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5632 db_printf("\n");
5633 }
5634 }
5635
DB_COMMAND_FLAGS(countfreebufs,db_coundfreebufs,DB_CMD_MEMSAFE)5636 DB_COMMAND_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE)
5637 {
5638 struct buf *bp;
5639 int i, used = 0, nfree = 0;
5640
5641 if (have_addr) {
5642 db_printf("usage: countfreebufs\n");
5643 return;
5644 }
5645
5646 for (i = 0; i < nbuf; i++) {
5647 bp = nbufp(i);
5648 if (bp->b_qindex == QUEUE_EMPTY)
5649 nfree++;
5650 else
5651 used++;
5652 }
5653
5654 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5655 nfree + used);
5656 db_printf("numfreebuffers is %d\n", numfreebuffers);
5657 }
5658 #endif /* DDB */
5659