2 * Copyright (c) 2004 Poul-Henning Kamp
3 * Copyright (c) 1994,1997 John S. Dyson
4 * Copyright (c) 2013 The FreeBSD Foundation
7 * Portions of this software were developed by Konstantin Belousov
8 * under sponsorship from the FreeBSD Foundation.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * this file contains a new buffer I/O scheme implementing a coherent
34 * VM object and buffer cache scheme. Pains have been taken to make
35 * sure that the performance degradation associated with schemes such
36 * as this is not realized.
38 * Author: John S. Dyson
39 * Significant help during the development and debugging phases
40 * had been provided by David Greenman, also of the FreeBSD core team.
42 * see man buf(9) for more info.
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include <sys/param.h>
49 #include <sys/systm.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
56 #include <sys/limits.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
64 #include <sys/racct.h>
65 #include <sys/resourcevar.h>
66 #include <sys/rwlock.h>
68 #include <sys/sysctl.h>
69 #include <sys/sysproto.h>
71 #include <sys/vmmeter.h>
72 #include <sys/vnode.h>
73 #include <sys/watchdog.h>
74 #include <geom/geom.h>
76 #include <vm/vm_param.h>
77 #include <vm/vm_kern.h>
78 #include <vm/vm_pageout.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_extern.h>
82 #include <vm/vm_map.h>
83 #include <vm/swap_pager.h>
84 #include "opt_compat.h"
87 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
89 struct bio_ops bioops; /* I/O operation notification */
91 struct buf_ops buf_ops_bio = {
92 .bop_name = "buf_ops_bio",
93 .bop_write = bufwrite,
94 .bop_strategy = bufstrategy,
96 .bop_bdflush = bufbdflush,
99 static struct buf *buf; /* buffer header pool */
100 extern struct buf *swbuf; /* Swap buffer header pool. */
101 caddr_t unmapped_buf;
103 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
104 struct proc *bufdaemonproc;
105 struct proc *bufspacedaemonproc;
107 static int inmem(struct vnode *vp, daddr_t blkno);
108 static void vm_hold_free_pages(struct buf *bp, int newbsize);
109 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
111 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
112 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
114 static void vfs_clean_pages_dirty_buf(struct buf *bp);
115 static void vfs_setdirty_locked_object(struct buf *bp);
116 static void vfs_vmio_invalidate(struct buf *bp);
117 static void vfs_vmio_truncate(struct buf *bp, int npages);
118 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
119 static int vfs_bio_clcheck(struct vnode *vp, int size,
120 daddr_t lblkno, daddr_t blkno);
121 static int buf_flush(struct vnode *vp, int);
122 static int buf_recycle(bool);
123 static int buf_scan(bool);
124 static int flushbufqueues(struct vnode *, int, int);
125 static void buf_daemon(void);
126 static void bremfreel(struct buf *bp);
127 static __inline void bd_wakeup(void);
128 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
129 static void bufkva_reclaim(vmem_t *, int);
130 static void bufkva_free(struct buf *);
131 static int buf_import(void *, void **, int, int);
132 static void buf_release(void *, void **, int);
134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
135 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
136 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
139 int vmiodirenable = TRUE;
140 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
141 "Use the VM system for directory writes");
142 long runningbufspace;
143 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
144 "Amount of presently outstanding async buffer io");
145 static long bufspace;
146 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
147 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
148 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
149 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
151 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
152 "Physical memory used for buffers");
154 static long bufkvaspace;
155 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0,
156 "Kernel virtual memory used for buffers");
157 static long maxbufspace;
158 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, &maxbufspace, 0,
159 "Maximum allowed value of bufspace (including metadata)");
160 static long bufmallocspace;
161 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
162 "Amount of malloced memory for buffers");
163 static long maxbufmallocspace;
164 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
165 0, "Maximum amount of malloced memory for buffers");
166 static long lobufspace;
167 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RW, &lobufspace, 0,
168 "Minimum amount of buffers we want to have");
170 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RW, &hibufspace, 0,
171 "Maximum allowed value of bufspace (excluding metadata)");
173 SYSCTL_LONG(_vfs, OID_AUTO, bufspacethresh, CTLFLAG_RW, &bufspacethresh,
174 0, "Bufspace consumed before waking the daemon to free some");
175 static int buffreekvacnt;
176 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
177 "Number of times we have freed the KVA space from some buffer");
178 static int bufdefragcnt;
179 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
180 "Number of times we have had to repeat buffer allocation to defragment");
181 static long lorunningspace;
182 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
183 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
184 "Minimum preferred space used for in-progress I/O");
185 static long hirunningspace;
186 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
187 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
188 "Maximum amount of space to use for in-progress I/O");
189 int dirtybufferflushes;
190 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
191 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
193 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
194 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
195 int altbufferflushes;
196 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
197 0, "Number of fsync flushes to limit dirty buffers");
198 static int recursiveflushes;
199 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
200 0, "Number of flushes skipped due to being recursive");
201 static int numdirtybuffers;
202 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
203 "Number of buffers that are dirty (has unwritten changes) at the moment");
204 static int lodirtybuffers;
205 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
206 "How many buffers we want to have free before bufdaemon can sleep");
207 static int hidirtybuffers;
208 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
209 "When the number of dirty buffers is considered severe");
211 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
212 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
213 static int numfreebuffers;
214 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
215 "Number of free buffers");
216 static int lofreebuffers;
217 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
218 "Target number of free buffers");
219 static int hifreebuffers;
220 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
221 "Threshold for clean buffer recycling");
222 static int getnewbufcalls;
223 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
224 "Number of calls to getnewbuf");
225 static int getnewbufrestarts;
226 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
227 "Number of times getnewbuf has had to restart a buffer acquisition");
228 static int mappingrestarts;
229 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
230 "Number of times getblk has had to restart a buffer mapping for "
232 static int numbufallocfails;
233 SYSCTL_INT(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, 0,
234 "Number of times buffer allocations failed");
235 static int flushbufqtarget = 100;
236 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
237 "Amount of work to do in flushbufqueues when helping bufdaemon");
238 static long notbufdflushes;
239 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0,
240 "Number of dirty buffer flushes done by the bufdaemon helpers");
241 static long barrierwrites;
242 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
243 "Number of barrier writes");
244 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
245 &unmapped_buf_allowed, 0,
246 "Permit the use of the unmapped i/o");
249 * This lock synchronizes access to bd_request.
251 static struct mtx_padalign bdlock;
254 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
255 * waitrunningbufspace().
257 static struct mtx_padalign rbreqlock;
260 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
262 static struct rwlock_padalign nblock;
265 * Lock that protects bdirtywait.
267 static struct mtx_padalign bdirtylock;
270 * Wakeup point for bufdaemon, as well as indicator of whether it is already
271 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
274 static int bd_request;
277 * Request/wakeup point for the bufspace daemon.
279 static int bufspace_request;
282 * Request for the buf daemon to write more buffers than is indicated by
283 * lodirtybuf. This may be necessary to push out excess dependencies or
284 * defragment the address space where a simple count of the number of dirty
285 * buffers is insufficient to characterize the demand for flushing them.
287 static int bd_speedupreq;
290 * bogus page -- for I/O to/from partially complete buffers
291 * this is a temporary solution to the problem, but it is not
292 * really that bad. it would be better to split the buffer
293 * for input in the case of buffers partially already in memory,
294 * but the code is intricate enough already.
296 vm_page_t bogus_page;
299 * Synchronization (sleep/wakeup) variable for active buffer space requests.
300 * Set when wait starts, cleared prior to wakeup().
301 * Used in runningbufwakeup() and waitrunningbufspace().
303 static int runningbufreq;
306 * Synchronization (sleep/wakeup) variable for buffer requests.
307 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
309 * Used in numdirtywakeup(), bufspace_wakeup(), bwillwrite(),
310 * getnewbuf(), and getblk().
312 static volatile int needsbuffer;
315 * Synchronization for bwillwrite() waiters.
317 static int bdirtywait;
320 * Definitions for the buffer free lists.
322 #define QUEUE_NONE 0 /* on no queue */
323 #define QUEUE_EMPTY 1 /* empty buffer headers */
324 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
325 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
326 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
328 /* Maximum number of clean buffer queues. */
329 #define CLEAN_QUEUES 16
331 /* Configured number of clean queues. */
332 static int clean_queues;
334 /* Maximum number of buffer queues. */
335 #define BUFFER_QUEUES (QUEUE_CLEAN + CLEAN_QUEUES)
337 /* Queues for free buffers with various properties */
338 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
340 static int bq_len[BUFFER_QUEUES];
344 * Lock for each bufqueue
346 static struct mtx_padalign bqlocks[BUFFER_QUEUES];
349 * per-cpu empty buffer cache.
354 * Single global constant for BUF_WMESG, to avoid getting multiple references.
355 * buf_wmesg is referred from macros.
357 const char *buf_wmesg = BUF_WMESG;
360 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
365 value = *(long *)arg1;
366 error = sysctl_handle_long(oidp, &value, 0, req);
367 if (error != 0 || req->newptr == NULL)
369 mtx_lock(&rbreqlock);
370 if (arg1 == &hirunningspace) {
371 if (value < lorunningspace)
374 hirunningspace = value;
376 KASSERT(arg1 == &lorunningspace,
377 ("%s: unknown arg1", __func__));
378 if (value > hirunningspace)
381 lorunningspace = value;
383 mtx_unlock(&rbreqlock);
387 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
388 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
390 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
395 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
396 return (sysctl_handle_long(oidp, arg1, arg2, req));
397 lvalue = *(long *)arg1;
398 if (lvalue > INT_MAX)
399 /* On overflow, still write out a long to trigger ENOMEM. */
400 return (sysctl_handle_long(oidp, &lvalue, 0, req));
402 return (sysctl_handle_int(oidp, &ivalue, 0, req));
411 return ((atomic_fetchadd_int(&nextq, 1) % clean_queues) + QUEUE_CLEAN);
415 bqisclean(int qindex)
418 return (qindex >= QUEUE_CLEAN && qindex < QUEUE_CLEAN + CLEAN_QUEUES);
424 * Return the appropriate queue lock based on the index.
426 static inline struct mtx *
430 return (struct mtx *)&bqlocks[qindex];
436 * Wakeup any bwillwrite() waiters.
441 mtx_lock(&bdirtylock);
446 mtx_unlock(&bdirtylock);
452 * Decrement the numdirtybuffers count by one and wakeup any
453 * threads blocked in bwillwrite().
459 if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
460 (lodirtybuffers + hidirtybuffers) / 2)
467 * Increment the numdirtybuffers count by one and wakeup the buf
475 * Only do the wakeup once as we cross the boundary. The
476 * buf daemon will keep running until the condition clears.
478 if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
479 (lodirtybuffers + hidirtybuffers) / 2)
486 * Called when buffer space is potentially available for recovery.
487 * getnewbuf() will block on this flag when it is unable to free
488 * sufficient buffer space. Buffer space becomes recoverable when
489 * bp's get placed back in the queues.
492 bufspace_wakeup(void)
496 * If someone is waiting for bufspace, wake them up.
498 * Since needsbuffer is set prior to doing an additional queue
499 * scan it is safe to check for the flag prior to acquiring the
500 * lock. The thread that is preparing to scan again before
501 * blocking would discover the buf we released.
505 if (atomic_cmpset_int(&needsbuffer, 1, 0) == 1)
506 wakeup(__DEVOLATILE(void *, &needsbuffer));
512 * bufspace_daemonwakeup:
514 * Wakeup the daemon responsible for freeing clean bufs.
517 bufspace_daemonwakeup(void)
520 if (bufspace_request == 0) {
521 bufspace_request = 1;
522 wakeup(&bufspace_request);
530 * Adjust the reported bufspace for a KVA managed buffer, possibly
531 * waking any waiters.
534 bufspace_adjust(struct buf *bp, int bufsize)
539 KASSERT((bp->b_flags & B_MALLOC) == 0,
540 ("bufspace_adjust: malloc buf %p", bp));
541 diff = bufsize - bp->b_bufsize;
543 atomic_subtract_long(&bufspace, -diff);
546 space = atomic_fetchadd_long(&bufspace, diff);
547 /* Wake up the daemon on the transition. */
548 if (space < bufspacethresh && space + diff >= bufspacethresh)
549 bufspace_daemonwakeup();
551 bp->b_bufsize = bufsize;
557 * Reserve bufspace before calling allocbuf(). metadata has a
558 * different space limit than data.
561 bufspace_reserve(int size, bool metadata)
572 if (space + size > limit)
574 } while (atomic_cmpset_long(&bufspace, space, space + size) == 0);
576 /* Wake up the daemon on the transition. */
577 if (space < bufspacethresh && space + size >= bufspacethresh)
578 bufspace_daemonwakeup();
586 * Release reserved bufspace after bufspace_adjust() has consumed it.
589 bufspace_release(int size)
591 atomic_subtract_long(&bufspace, size);
598 * Wait for bufspace, acting as the buf daemon if a locked vnode is
599 * supplied. needsbuffer must be set in a safe fashion prior to
600 * polling for space. The operation must be re-tried on return.
603 bufspace_wait(struct vnode *vp, int gbflags, int slpflag, int slptimeo)
606 int error, fl, norunbuf;
608 if ((gbflags & GB_NOWAIT_BD) != 0)
613 while (needsbuffer != 0) {
614 if (vp != NULL && vp->v_type != VCHR &&
615 (td->td_pflags & TDP_BUFNEED) == 0) {
618 * getblk() is called with a vnode locked, and
619 * some majority of the dirty buffers may as
620 * well belong to the vnode. Flushing the
621 * buffers there would make a progress that
622 * cannot be achieved by the buf_daemon, that
623 * cannot lock the vnode.
625 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
626 (td->td_pflags & TDP_NORUNNINGBUF);
629 * Play bufdaemon. The getnewbuf() function
630 * may be called while the thread owns lock
631 * for another dirty buffer for the same
632 * vnode, which makes it impossible to use
633 * VOP_FSYNC() there, due to the buffer lock
636 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
637 fl = buf_flush(vp, flushbufqtarget);
638 td->td_pflags &= norunbuf;
642 if (needsbuffer == 0)
645 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
646 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
657 * buffer space management daemon. Tries to maintain some marginal
658 * amount of free buffer space so that requesting processes neither
659 * block nor work to reclaim buffers.
662 bufspace_daemon(void)
665 kproc_suspend_check(bufspacedaemonproc);
668 * Free buffers from the clean queue until we meet our
671 * Theory of operation: The buffer cache is most efficient
672 * when some free buffer headers and space are always
673 * available to getnewbuf(). This daemon attempts to prevent
674 * the excessive blocking and synchronization associated
675 * with shortfall. It goes through three phases according
678 * 1) The daemon wakes up voluntarily once per-second
679 * during idle periods when the counters are below
680 * the wakeup thresholds (bufspacethresh, lofreebuffers).
682 * 2) The daemon wakes up as we cross the thresholds
683 * ahead of any potential blocking. This may bounce
684 * slightly according to the rate of consumption and
687 * 3) The daemon and consumers are starved for working
688 * clean buffers. This is the 'bufspace' sleep below
689 * which will inefficiently trade bufs with bqrelse
690 * until we return to condition 2.
692 while (bufspace > lobufspace ||
693 numfreebuffers < hifreebuffers) {
694 if (buf_recycle(false) != 0) {
695 atomic_set_int(&needsbuffer, 1);
696 if (buf_recycle(false) != 0) {
699 rw_sleep(__DEVOLATILE(void *,
700 &needsbuffer), &nblock,
701 PRIBIO|PDROP, "bufspace",
711 * Re-check our limits under the exclusive nblock.
714 if (bufspace < bufspacethresh &&
715 numfreebuffers > lofreebuffers) {
716 bufspace_request = 0;
717 rw_sleep(&bufspace_request, &nblock, PRIBIO|PDROP,
724 static struct kproc_desc bufspace_kp = {
729 SYSINIT(bufspacedaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start,
735 * Adjust the reported bufspace for a malloc managed buffer, possibly
736 * waking any waiters.
739 bufmallocadjust(struct buf *bp, int bufsize)
743 KASSERT((bp->b_flags & B_MALLOC) != 0,
744 ("bufmallocadjust: non-malloc buf %p", bp));
745 diff = bufsize - bp->b_bufsize;
747 atomic_subtract_long(&bufmallocspace, -diff);
749 atomic_add_long(&bufmallocspace, diff);
750 bp->b_bufsize = bufsize;
756 * Wake up processes that are waiting on asynchronous writes to fall
757 * below lorunningspace.
763 mtx_lock(&rbreqlock);
766 wakeup(&runningbufreq);
768 mtx_unlock(&rbreqlock);
774 * Decrement the outstanding write count according.
777 runningbufwakeup(struct buf *bp)
781 bspace = bp->b_runningbufspace;
784 space = atomic_fetchadd_long(&runningbufspace, -bspace);
785 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
787 bp->b_runningbufspace = 0;
789 * Only acquire the lock and wakeup on the transition from exceeding
790 * the threshold to falling below it.
792 if (space < lorunningspace)
794 if (space - bspace > lorunningspace)
800 * waitrunningbufspace()
802 * runningbufspace is a measure of the amount of I/O currently
803 * running. This routine is used in async-write situations to
804 * prevent creating huge backups of pending writes to a device.
805 * Only asynchronous writes are governed by this function.
807 * This does NOT turn an async write into a sync write. It waits
808 * for earlier writes to complete and generally returns before the
809 * caller's write has reached the device.
812 waitrunningbufspace(void)
815 mtx_lock(&rbreqlock);
816 while (runningbufspace > hirunningspace) {
818 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
820 mtx_unlock(&rbreqlock);
825 * vfs_buf_test_cache:
827 * Called when a buffer is extended. This function clears the B_CACHE
828 * bit if the newly extended portion of the buffer does not contain
832 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
833 vm_offset_t size, vm_page_t m)
836 VM_OBJECT_ASSERT_LOCKED(m->object);
837 if (bp->b_flags & B_CACHE) {
838 int base = (foff + off) & PAGE_MASK;
839 if (vm_page_is_valid(m, base, size) == 0)
840 bp->b_flags &= ~B_CACHE;
844 /* Wake up the buffer daemon if necessary */
850 if (bd_request == 0) {
858 * bd_speedup - speedup the buffer cache flushing code
867 if (bd_speedupreq == 0 || bd_request == 0)
877 #define NSWBUF_MIN 16
881 #define TRANSIENT_DENOM 5
883 #define TRANSIENT_DENOM 10
887 * Calculating buffer cache scaling values and reserve space for buffer
888 * headers. This is called during low level kernel initialization and
889 * may be called more then once. We CANNOT write to the memory area
890 * being reserved at this time.
893 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
896 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
899 * physmem_est is in pages. Convert it to kilobytes (assumes
900 * PAGE_SIZE is >= 1K)
902 physmem_est = physmem_est * (PAGE_SIZE / 1024);
905 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
906 * For the first 64MB of ram nominally allocate sufficient buffers to
907 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
908 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
909 * the buffer cache we limit the eventual kva reservation to
912 * factor represents the 1/4 x ram conversion.
915 int factor = 4 * BKVASIZE / 1024;
918 if (physmem_est > 4096)
919 nbuf += min((physmem_est - 4096) / factor,
921 if (physmem_est > 65536)
922 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
923 32 * 1024 * 1024 / (factor * 5));
925 if (maxbcache && nbuf > maxbcache / BKVASIZE)
926 nbuf = maxbcache / BKVASIZE;
931 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
932 maxbuf = (LONG_MAX / 3) / BKVASIZE;
935 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
941 * Ideal allocation size for the transient bio submap is 10%
942 * of the maximal space buffer map. This roughly corresponds
943 * to the amount of the buffer mapped for typical UFS load.
945 * Clip the buffer map to reserve space for the transient
946 * BIOs, if its extent is bigger than 90% (80% on i386) of the
947 * maximum buffer map extent on the platform.
949 * The fall-back to the maxbuf in case of maxbcache unset,
950 * allows to not trim the buffer KVA for the architectures
951 * with ample KVA space.
953 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
954 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
955 buf_sz = (long)nbuf * BKVASIZE;
956 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
957 (TRANSIENT_DENOM - 1)) {
959 * There is more KVA than memory. Do not
960 * adjust buffer map size, and assign the rest
961 * of maxbuf to transient map.
963 biotmap_sz = maxbuf_sz - buf_sz;
966 * Buffer map spans all KVA we could afford on
967 * this platform. Give 10% (20% on i386) of
968 * the buffer map to the transient bio map.
970 biotmap_sz = buf_sz / TRANSIENT_DENOM;
971 buf_sz -= biotmap_sz;
973 if (biotmap_sz / INT_MAX > MAXPHYS)
974 bio_transient_maxcnt = INT_MAX;
976 bio_transient_maxcnt = biotmap_sz / MAXPHYS;
978 * Artificially limit to 1024 simultaneous in-flight I/Os
979 * using the transient mapping.
981 if (bio_transient_maxcnt > 1024)
982 bio_transient_maxcnt = 1024;
984 nbuf = buf_sz / BKVASIZE;
988 * swbufs are used as temporary holders for I/O, such as paging I/O.
989 * We have no less then 16 and no more then 256.
991 nswbuf = min(nbuf / 4, 256);
992 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
993 if (nswbuf < NSWBUF_MIN)
997 * Reserve space for the buffer cache buffers
1000 v = (caddr_t)(swbuf + nswbuf);
1002 v = (caddr_t)(buf + nbuf);
1007 /* Initialize the buffer subsystem. Called before use of any buffers. */
1014 CTASSERT(MAXBCACHEBUF >= MAXBSIZE);
1015 mtx_init(&bqlocks[QUEUE_DIRTY], "bufq dirty lock", NULL, MTX_DEF);
1016 mtx_init(&bqlocks[QUEUE_EMPTY], "bufq empty lock", NULL, MTX_DEF);
1017 for (i = QUEUE_CLEAN; i < QUEUE_CLEAN + CLEAN_QUEUES; i++)
1018 mtx_init(&bqlocks[i], "bufq clean lock", NULL, MTX_DEF);
1019 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1020 rw_init(&nblock, "needsbuffer lock");
1021 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1022 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1024 /* next, make a null set of free lists */
1025 for (i = 0; i < BUFFER_QUEUES; i++)
1026 TAILQ_INIT(&bufqueues[i]);
1028 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1030 /* finally, initialize each buffer header and stick on empty q */
1031 for (i = 0; i < nbuf; i++) {
1033 bzero(bp, sizeof *bp);
1034 bp->b_flags = B_INVAL;
1035 bp->b_rcred = NOCRED;
1036 bp->b_wcred = NOCRED;
1037 bp->b_qindex = QUEUE_EMPTY;
1039 bp->b_data = bp->b_kvabase = unmapped_buf;
1040 LIST_INIT(&bp->b_dep);
1042 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
1044 bq_len[QUEUE_EMPTY]++;
1049 * maxbufspace is the absolute maximum amount of buffer space we are
1050 * allowed to reserve in KVM and in real terms. The absolute maximum
1051 * is nominally used by metadata. hibufspace is the nominal maximum
1052 * used by most other requests. The differential is required to
1053 * ensure that metadata deadlocks don't occur.
1055 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1056 * this may result in KVM fragmentation which is not handled optimally
1057 * by the system. XXX This is less true with vmem. We could use
1060 maxbufspace = (long)nbuf * BKVASIZE;
1061 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10);
1062 lobufspace = (hibufspace / 20) * 19; /* 95% */
1063 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1066 * Note: The 16 MiB upper limit for hirunningspace was chosen
1067 * arbitrarily and may need further tuning. It corresponds to
1068 * 128 outstanding write IO requests (if IO size is 128 KiB),
1069 * which fits with many RAID controllers' tagged queuing limits.
1070 * The lower 1 MiB limit is the historical upper limit for
1073 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF),
1074 16 * 1024 * 1024), 1024 * 1024);
1075 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF);
1078 * Limit the amount of malloc memory since it is wired permanently into
1079 * the kernel space. Even though this is accounted for in the buffer
1080 * allocation, we don't want the malloced region to grow uncontrolled.
1081 * The malloc scheme improves memory utilization significantly on
1082 * average (small) directories.
1084 maxbufmallocspace = hibufspace / 20;
1087 * Reduce the chance of a deadlock occurring by limiting the number
1088 * of delayed-write dirty buffers we allow to stack up.
1090 hidirtybuffers = nbuf / 4 + 20;
1091 dirtybufthresh = hidirtybuffers * 9 / 10;
1092 numdirtybuffers = 0;
1094 * To support extreme low-memory systems, make sure hidirtybuffers
1095 * cannot eat up all available buffer space. This occurs when our
1096 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1097 * buffer space assuming BKVASIZE'd buffers.
1099 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1100 hidirtybuffers >>= 1;
1102 lodirtybuffers = hidirtybuffers / 2;
1105 * lofreebuffers should be sufficient to avoid stalling waiting on
1106 * buf headers under heavy utilization. The bufs in per-cpu caches
1107 * are counted as free but will be unavailable to threads executing
1110 * hifreebuffers is the free target for the bufspace daemon. This
1111 * should be set appropriately to limit work per-iteration.
1113 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1114 hifreebuffers = (3 * lofreebuffers) / 2;
1115 numfreebuffers = nbuf;
1117 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
1118 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
1120 /* Setup the kva and free list allocators. */
1121 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1122 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1123 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1126 * Size the clean queue according to the amount of buffer space.
1127 * One queue per-256mb up to the max. More queues gives better
1128 * concurrency but less accurate LRU.
1130 clean_queues = MIN(howmany(maxbufspace, 256*1024*1024), CLEAN_QUEUES);
1136 vfs_buf_check_mapped(struct buf *bp)
1139 KASSERT(bp->b_kvabase != unmapped_buf,
1140 ("mapped buf: b_kvabase was not updated %p", bp));
1141 KASSERT(bp->b_data != unmapped_buf,
1142 ("mapped buf: b_data was not updated %p", bp));
1143 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1144 MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1148 vfs_buf_check_unmapped(struct buf *bp)
1151 KASSERT(bp->b_data == unmapped_buf,
1152 ("unmapped buf: corrupted b_data %p", bp));
1155 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1156 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1158 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1159 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1163 isbufbusy(struct buf *bp)
1165 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1166 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1172 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1175 bufshutdown(int show_busybufs)
1177 static int first_buf_printf = 1;
1179 int iter, nbusy, pbusy;
1185 * Sync filesystems for shutdown
1187 wdog_kern_pat(WD_LASTVAL);
1188 sys_sync(curthread, NULL);
1191 * With soft updates, some buffers that are
1192 * written will be remarked as dirty until other
1193 * buffers are written.
1195 for (iter = pbusy = 0; iter < 20; iter++) {
1197 for (bp = &buf[nbuf]; --bp >= buf; )
1201 if (first_buf_printf)
1202 printf("All buffers synced.");
1205 if (first_buf_printf) {
1206 printf("Syncing disks, buffers remaining... ");
1207 first_buf_printf = 0;
1209 printf("%d ", nbusy);
1214 wdog_kern_pat(WD_LASTVAL);
1215 sys_sync(curthread, NULL);
1219 * Drop Giant and spin for a while to allow
1220 * interrupt threads to run.
1223 DELAY(50000 * iter);
1227 * Drop Giant and context switch several times to
1228 * allow interrupt threads to run.
1231 for (subiter = 0; subiter < 50 * iter; subiter++) {
1232 thread_lock(curthread);
1233 mi_switch(SW_VOL, NULL);
1234 thread_unlock(curthread);
1242 * Count only busy local buffers to prevent forcing
1243 * a fsck if we're just a client of a wedged NFS server
1246 for (bp = &buf[nbuf]; --bp >= buf; ) {
1247 if (isbufbusy(bp)) {
1249 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1250 if (bp->b_dev == NULL) {
1251 TAILQ_REMOVE(&mountlist,
1252 bp->b_vp->v_mount, mnt_list);
1257 if (show_busybufs > 0) {
1259 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1260 nbusy, bp, bp->b_vp, bp->b_flags,
1261 (intmax_t)bp->b_blkno,
1262 (intmax_t)bp->b_lblkno);
1263 BUF_LOCKPRINTINFO(bp);
1264 if (show_busybufs > 1)
1272 * Failed to sync all blocks. Indicate this and don't
1273 * unmount filesystems (thus forcing an fsck on reboot).
1275 printf("Giving up on %d buffers\n", nbusy);
1276 DELAY(5000000); /* 5 seconds */
1278 if (!first_buf_printf)
1279 printf("Final sync complete\n");
1281 * Unmount filesystems
1283 if (panicstr == NULL)
1287 DELAY(100000); /* wait for console output to finish */
1291 bpmap_qenter(struct buf *bp)
1294 BUF_CHECK_MAPPED(bp);
1297 * bp->b_data is relative to bp->b_offset, but
1298 * bp->b_offset may be offset into the first page.
1300 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1301 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1302 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1303 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1309 * Insert the buffer into the appropriate free list.
1312 binsfree(struct buf *bp, int qindex)
1314 struct mtx *olock, *nlock;
1316 if (qindex != QUEUE_EMPTY) {
1317 BUF_ASSERT_XLOCKED(bp);
1321 * Stick to the same clean queue for the lifetime of the buf to
1322 * limit locking below. Otherwise pick ont sequentially.
1324 if (qindex == QUEUE_CLEAN) {
1325 if (bqisclean(bp->b_qindex))
1326 qindex = bp->b_qindex;
1328 qindex = bqcleanq();
1332 * Handle delayed bremfree() processing.
1334 nlock = bqlock(qindex);
1335 if (bp->b_flags & B_REMFREE) {
1336 olock = bqlock(bp->b_qindex);
1339 if (olock != nlock) {
1346 if (bp->b_qindex != QUEUE_NONE)
1347 panic("binsfree: free buffer onto another queue???");
1349 bp->b_qindex = qindex;
1350 if (bp->b_flags & B_AGE)
1351 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1353 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1355 bq_len[bp->b_qindex]++;
1363 * Free a buffer to the buf zone once it no longer has valid contents.
1366 buf_free(struct buf *bp)
1369 if (bp->b_flags & B_REMFREE)
1371 if (bp->b_vflags & BV_BKGRDINPROG)
1372 panic("losing buffer 1");
1373 if (bp->b_rcred != NOCRED) {
1374 crfree(bp->b_rcred);
1375 bp->b_rcred = NOCRED;
1377 if (bp->b_wcred != NOCRED) {
1378 crfree(bp->b_wcred);
1379 bp->b_wcred = NOCRED;
1381 if (!LIST_EMPTY(&bp->b_dep))
1385 uma_zfree(buf_zone, bp);
1386 atomic_add_int(&numfreebuffers, 1);
1393 * Import bufs into the uma cache from the buf list. The system still
1394 * expects a static array of bufs and much of the synchronization
1395 * around bufs assumes type stable storage. As a result, UMA is used
1396 * only as a per-cpu cache of bufs still maintained on a global list.
1399 buf_import(void *arg, void **store, int cnt, int flags)
1404 mtx_lock(&bqlocks[QUEUE_EMPTY]);
1405 for (i = 0; i < cnt; i++) {
1406 bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1412 mtx_unlock(&bqlocks[QUEUE_EMPTY]);
1420 * Release bufs from the uma cache back to the buffer queues.
1423 buf_release(void *arg, void **store, int cnt)
1427 for (i = 0; i < cnt; i++)
1428 binsfree(store[i], QUEUE_EMPTY);
1434 * Allocate an empty buffer header.
1441 bp = uma_zalloc(buf_zone, M_NOWAIT);
1443 bufspace_daemonwakeup();
1444 atomic_add_int(&numbufallocfails, 1);
1449 * Wake-up the bufspace daemon on transition.
1451 if (atomic_fetchadd_int(&numfreebuffers, -1) == lofreebuffers)
1452 bufspace_daemonwakeup();
1454 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1455 panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1457 KASSERT(bp->b_vp == NULL,
1458 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1459 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1460 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1461 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1462 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1463 KASSERT(bp->b_npages == 0,
1464 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1465 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1466 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1473 bp->b_blkno = bp->b_lblkno = 0;
1474 bp->b_offset = NOOFFSET;
1480 bp->b_dirtyoff = bp->b_dirtyend = 0;
1481 bp->b_bufobj = NULL;
1482 bp->b_data = bp->b_kvabase = unmapped_buf;
1483 bp->b_fsprivate1 = NULL;
1484 bp->b_fsprivate2 = NULL;
1485 bp->b_fsprivate3 = NULL;
1486 LIST_INIT(&bp->b_dep);
1494 * Free a buffer from the given bufqueue. kva controls whether the
1495 * freed buf must own some kva resources. This is used for
1499 buf_qrecycle(int qindex, bool kva)
1501 struct buf *bp, *nbp;
1504 atomic_add_int(&bufdefragcnt, 1);
1506 mtx_lock(&bqlocks[qindex]);
1507 nbp = TAILQ_FIRST(&bufqueues[qindex]);
1510 * Run scan, possibly freeing data and/or kva mappings on the fly
1513 while ((bp = nbp) != NULL) {
1515 * Calculate next bp (we can only use it if we do not
1516 * release the bqlock).
1518 nbp = TAILQ_NEXT(bp, b_freelist);
1521 * If we are defragging then we need a buffer with
1522 * some kva to reclaim.
1524 if (kva && bp->b_kvasize == 0)
1527 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1531 * Skip buffers with background writes in progress.
1533 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1538 KASSERT(bp->b_qindex == qindex,
1539 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1541 * NOTE: nbp is now entirely invalid. We can only restart
1542 * the scan from this point on.
1545 mtx_unlock(&bqlocks[qindex]);
1548 * Requeue the background write buffer with error and
1551 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1553 mtx_lock(&bqlocks[qindex]);
1554 nbp = TAILQ_FIRST(&bufqueues[qindex]);
1557 bp->b_flags |= B_INVAL;
1561 mtx_unlock(&bqlocks[qindex]);
1569 * Iterate through all clean queues until we find a buf to recycle or
1570 * exhaust the search.
1573 buf_recycle(bool kva)
1575 int qindex, first_qindex;
1577 qindex = first_qindex = bqcleanq();
1579 if (buf_qrecycle(qindex, kva) == 0)
1581 if (++qindex == QUEUE_CLEAN + clean_queues)
1582 qindex = QUEUE_CLEAN;
1583 } while (qindex != first_qindex);
1591 * Scan the clean queues looking for a buffer to recycle. needsbuffer
1592 * is set on failure so that the caller may optionally bufspace_wait()
1593 * in a race-free fashion.
1596 buf_scan(bool defrag)
1601 * To avoid heavy synchronization and wakeup races we set
1602 * needsbuffer and re-poll before failing. This ensures that
1603 * no frees can be missed between an unsuccessful poll and
1604 * going to sleep in a synchronized fashion.
1606 if ((error = buf_recycle(defrag)) != 0) {
1607 atomic_set_int(&needsbuffer, 1);
1608 bufspace_daemonwakeup();
1609 error = buf_recycle(defrag);
1612 atomic_add_int(&getnewbufrestarts, 1);
1619 * Mark the buffer for removal from the appropriate free list.
1623 bremfree(struct buf *bp)
1626 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1627 KASSERT((bp->b_flags & B_REMFREE) == 0,
1628 ("bremfree: buffer %p already marked for delayed removal.", bp));
1629 KASSERT(bp->b_qindex != QUEUE_NONE,
1630 ("bremfree: buffer %p not on a queue.", bp));
1631 BUF_ASSERT_XLOCKED(bp);
1633 bp->b_flags |= B_REMFREE;
1639 * Force an immediate removal from a free list. Used only in nfs when
1640 * it abuses the b_freelist pointer.
1643 bremfreef(struct buf *bp)
1647 qlock = bqlock(bp->b_qindex);
1656 * Removes a buffer from the free list, must be called with the
1657 * correct qlock held.
1660 bremfreel(struct buf *bp)
1663 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1664 bp, bp->b_vp, bp->b_flags);
1665 KASSERT(bp->b_qindex != QUEUE_NONE,
1666 ("bremfreel: buffer %p not on a queue.", bp));
1667 if (bp->b_qindex != QUEUE_EMPTY) {
1668 BUF_ASSERT_XLOCKED(bp);
1670 mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1672 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1674 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1676 bq_len[bp->b_qindex]--;
1678 bp->b_qindex = QUEUE_NONE;
1679 bp->b_flags &= ~B_REMFREE;
1685 * Free the kva allocation for a buffer.
1689 bufkva_free(struct buf *bp)
1693 if (bp->b_kvasize == 0) {
1694 KASSERT(bp->b_kvabase == unmapped_buf &&
1695 bp->b_data == unmapped_buf,
1696 ("Leaked KVA space on %p", bp));
1697 } else if (buf_mapped(bp))
1698 BUF_CHECK_MAPPED(bp);
1700 BUF_CHECK_UNMAPPED(bp);
1702 if (bp->b_kvasize == 0)
1705 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1706 atomic_subtract_long(&bufkvaspace, bp->b_kvasize);
1707 atomic_add_int(&buffreekvacnt, 1);
1708 bp->b_data = bp->b_kvabase = unmapped_buf;
1715 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
1718 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
1723 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
1724 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
1729 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
1732 * Buffer map is too fragmented. Request the caller
1733 * to defragment the map.
1737 bp->b_kvabase = (caddr_t)addr;
1738 bp->b_kvasize = maxsize;
1739 atomic_add_long(&bufkvaspace, bp->b_kvasize);
1740 if ((gbflags & GB_UNMAPPED) != 0) {
1741 bp->b_data = unmapped_buf;
1742 BUF_CHECK_UNMAPPED(bp);
1744 bp->b_data = bp->b_kvabase;
1745 BUF_CHECK_MAPPED(bp);
1753 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
1754 * callback that fires to avoid returning failure.
1757 bufkva_reclaim(vmem_t *vmem, int flags)
1761 for (i = 0; i < 5; i++)
1762 if (buf_scan(true) != 0)
1769 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
1770 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1771 * the buffer is valid and we do not have to do anything.
1774 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1775 int cnt, struct ucred * cred)
1780 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1781 if (inmem(vp, *rablkno))
1783 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1785 if ((rabp->b_flags & B_CACHE) == 0) {
1786 if (!TD_IS_IDLETHREAD(curthread)) {
1790 racct_add_buf(curproc, rabp, 0);
1791 PROC_UNLOCK(curproc);
1794 curthread->td_ru.ru_inblock++;
1796 rabp->b_flags |= B_ASYNC;
1797 rabp->b_flags &= ~B_INVAL;
1798 rabp->b_ioflags &= ~BIO_ERROR;
1799 rabp->b_iocmd = BIO_READ;
1800 if (rabp->b_rcred == NOCRED && cred != NOCRED)
1801 rabp->b_rcred = crhold(cred);
1802 vfs_busy_pages(rabp, 0);
1804 rabp->b_iooffset = dbtob(rabp->b_blkno);
1813 * Entry point for bread() and breadn() via #defines in sys/buf.h.
1815 * Get a buffer with the specified data. Look in the cache first. We
1816 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
1817 * is set, the buffer is valid and we do not have to do anything, see
1818 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1820 * Always return a NULL buffer pointer (in bpp) when returning an error.
1823 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1824 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1827 int rv = 0, readwait = 0;
1829 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1831 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1833 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1837 /* if not found in cache, do some I/O */
1838 if ((bp->b_flags & B_CACHE) == 0) {
1839 if (!TD_IS_IDLETHREAD(curthread)) {
1843 racct_add_buf(curproc, bp, 0);
1844 PROC_UNLOCK(curproc);
1847 curthread->td_ru.ru_inblock++;
1849 bp->b_iocmd = BIO_READ;
1850 bp->b_flags &= ~B_INVAL;
1851 bp->b_ioflags &= ~BIO_ERROR;
1852 if (bp->b_rcred == NOCRED && cred != NOCRED)
1853 bp->b_rcred = crhold(cred);
1854 vfs_busy_pages(bp, 0);
1855 bp->b_iooffset = dbtob(bp->b_blkno);
1860 breada(vp, rablkno, rabsize, cnt, cred);
1873 * Write, release buffer on completion. (Done by iodone
1874 * if async). Do not bother writing anything if the buffer
1877 * Note that we set B_CACHE here, indicating that buffer is
1878 * fully valid and thus cacheable. This is true even of NFS
1879 * now so we set it generally. This could be set either here
1880 * or in biodone() since the I/O is synchronous. We put it
1884 bufwrite(struct buf *bp)
1891 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1892 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
1893 bp->b_flags |= B_INVAL | B_RELBUF;
1894 bp->b_flags &= ~B_CACHE;
1898 if (bp->b_flags & B_INVAL) {
1903 if (bp->b_flags & B_BARRIER)
1906 oldflags = bp->b_flags;
1908 BUF_ASSERT_HELD(bp);
1910 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1911 ("FFS background buffer should not get here %p", bp));
1915 vp_md = vp->v_vflag & VV_MD;
1920 * Mark the buffer clean. Increment the bufobj write count
1921 * before bundirty() call, to prevent other thread from seeing
1922 * empty dirty list and zero counter for writes in progress,
1923 * falsely indicating that the bufobj is clean.
1925 bufobj_wref(bp->b_bufobj);
1928 bp->b_flags &= ~B_DONE;
1929 bp->b_ioflags &= ~BIO_ERROR;
1930 bp->b_flags |= B_CACHE;
1931 bp->b_iocmd = BIO_WRITE;
1933 vfs_busy_pages(bp, 1);
1936 * Normal bwrites pipeline writes
1938 bp->b_runningbufspace = bp->b_bufsize;
1939 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1941 if (!TD_IS_IDLETHREAD(curthread)) {
1945 racct_add_buf(curproc, bp, 1);
1946 PROC_UNLOCK(curproc);
1949 curthread->td_ru.ru_oublock++;
1951 if (oldflags & B_ASYNC)
1953 bp->b_iooffset = dbtob(bp->b_blkno);
1956 if ((oldflags & B_ASYNC) == 0) {
1957 int rtval = bufwait(bp);
1960 } else if (space > hirunningspace) {
1962 * don't allow the async write to saturate the I/O
1963 * system. We will not deadlock here because
1964 * we are blocking waiting for I/O that is already in-progress
1965 * to complete. We do not block here if it is the update
1966 * or syncer daemon trying to clean up as that can lead
1969 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1970 waitrunningbufspace();
1977 bufbdflush(struct bufobj *bo, struct buf *bp)
1981 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1982 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1984 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1987 * Try to find a buffer to flush.
1989 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1990 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1992 LK_EXCLUSIVE | LK_NOWAIT, NULL))
1995 panic("bdwrite: found ourselves");
1997 /* Don't countdeps with the bo lock held. */
1998 if (buf_countdeps(nbp, 0)) {
2003 if (nbp->b_flags & B_CLUSTEROK) {
2004 vfs_bio_awrite(nbp);
2009 dirtybufferflushes++;
2018 * Delayed write. (Buffer is marked dirty). Do not bother writing
2019 * anything if the buffer is marked invalid.
2021 * Note that since the buffer must be completely valid, we can safely
2022 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2023 * biodone() in order to prevent getblk from writing the buffer
2024 * out synchronously.
2027 bdwrite(struct buf *bp)
2029 struct thread *td = curthread;
2033 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2034 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2035 KASSERT((bp->b_flags & B_BARRIER) == 0,
2036 ("Barrier request in delayed write %p", bp));
2037 BUF_ASSERT_HELD(bp);
2039 if (bp->b_flags & B_INVAL) {
2045 * If we have too many dirty buffers, don't create any more.
2046 * If we are wildly over our limit, then force a complete
2047 * cleanup. Otherwise, just keep the situation from getting
2048 * out of control. Note that we have to avoid a recursive
2049 * disaster and not try to clean up after our own cleanup!
2053 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2054 td->td_pflags |= TDP_INBDFLUSH;
2056 td->td_pflags &= ~TDP_INBDFLUSH;
2062 * Set B_CACHE, indicating that the buffer is fully valid. This is
2063 * true even of NFS now.
2065 bp->b_flags |= B_CACHE;
2068 * This bmap keeps the system from needing to do the bmap later,
2069 * perhaps when the system is attempting to do a sync. Since it
2070 * is likely that the indirect block -- or whatever other datastructure
2071 * that the filesystem needs is still in memory now, it is a good
2072 * thing to do this. Note also, that if the pageout daemon is
2073 * requesting a sync -- there might not be enough memory to do
2074 * the bmap then... So, this is important to do.
2076 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2077 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2081 * Set the *dirty* buffer range based upon the VM system dirty
2084 * Mark the buffer pages as clean. We need to do this here to
2085 * satisfy the vnode_pager and the pageout daemon, so that it
2086 * thinks that the pages have been "cleaned". Note that since
2087 * the pages are in a delayed write buffer -- the VFS layer
2088 * "will" see that the pages get written out on the next sync,
2089 * or perhaps the cluster will be completed.
2091 vfs_clean_pages_dirty_buf(bp);
2095 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2096 * due to the softdep code.
2103 * Turn buffer into delayed write request. We must clear BIO_READ and
2104 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2105 * itself to properly update it in the dirty/clean lists. We mark it
2106 * B_DONE to ensure that any asynchronization of the buffer properly
2107 * clears B_DONE ( else a panic will occur later ).
2109 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2110 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2111 * should only be called if the buffer is known-good.
2113 * Since the buffer is not on a queue, we do not update the numfreebuffers
2116 * The buffer must be on QUEUE_NONE.
2119 bdirty(struct buf *bp)
2122 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2123 bp, bp->b_vp, bp->b_flags);
2124 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2125 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2126 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2127 BUF_ASSERT_HELD(bp);
2128 bp->b_flags &= ~(B_RELBUF);
2129 bp->b_iocmd = BIO_WRITE;
2131 if ((bp->b_flags & B_DELWRI) == 0) {
2132 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2141 * Clear B_DELWRI for buffer.
2143 * Since the buffer is not on a queue, we do not update the numfreebuffers
2146 * The buffer must be on QUEUE_NONE.
2150 bundirty(struct buf *bp)
2153 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2154 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2155 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2156 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2157 BUF_ASSERT_HELD(bp);
2159 if (bp->b_flags & B_DELWRI) {
2160 bp->b_flags &= ~B_DELWRI;
2165 * Since it is now being written, we can clear its deferred write flag.
2167 bp->b_flags &= ~B_DEFERRED;
2173 * Asynchronous write. Start output on a buffer, but do not wait for
2174 * it to complete. The buffer is released when the output completes.
2176 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2177 * B_INVAL buffers. Not us.
2180 bawrite(struct buf *bp)
2183 bp->b_flags |= B_ASYNC;
2190 * Asynchronous barrier write. Start output on a buffer, but do not
2191 * wait for it to complete. Place a write barrier after this write so
2192 * that this buffer and all buffers written before it are committed to
2193 * the disk before any buffers written after this write are committed
2194 * to the disk. The buffer is released when the output completes.
2197 babarrierwrite(struct buf *bp)
2200 bp->b_flags |= B_ASYNC | B_BARRIER;
2207 * Synchronous barrier write. Start output on a buffer and wait for
2208 * it to complete. Place a write barrier after this write so that
2209 * this buffer and all buffers written before it are committed to
2210 * the disk before any buffers written after this write are committed
2211 * to the disk. The buffer is released when the output completes.
2214 bbarrierwrite(struct buf *bp)
2217 bp->b_flags |= B_BARRIER;
2218 return (bwrite(bp));
2224 * Called prior to the locking of any vnodes when we are expecting to
2225 * write. We do not want to starve the buffer cache with too many
2226 * dirty buffers so we block here. By blocking prior to the locking
2227 * of any vnodes we attempt to avoid the situation where a locked vnode
2228 * prevents the various system daemons from flushing related buffers.
2234 if (numdirtybuffers >= hidirtybuffers) {
2235 mtx_lock(&bdirtylock);
2236 while (numdirtybuffers >= hidirtybuffers) {
2238 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2241 mtx_unlock(&bdirtylock);
2246 * Return true if we have too many dirty buffers.
2249 buf_dirty_count_severe(void)
2252 return(numdirtybuffers >= hidirtybuffers);
2258 * Release a busy buffer and, if requested, free its resources. The
2259 * buffer will be stashed in the appropriate bufqueue[] allowing it
2260 * to be accessed later as a cache entity or reused for other purposes.
2263 brelse(struct buf *bp)
2268 * Many functions erroneously call brelse with a NULL bp under rare
2269 * error conditions. Simply return when called with a NULL bp.
2273 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2274 bp, bp->b_vp, bp->b_flags);
2275 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2276 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2277 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2278 ("brelse: non-VMIO buffer marked NOREUSE"));
2280 if (BUF_LOCKRECURSED(bp)) {
2282 * Do not process, in particular, do not handle the
2283 * B_INVAL/B_RELBUF and do not release to free list.
2289 if (bp->b_flags & B_MANAGED) {
2294 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2295 BO_LOCK(bp->b_bufobj);
2296 bp->b_vflags &= ~BV_BKGRDERR;
2297 BO_UNLOCK(bp->b_bufobj);
2300 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2301 !(bp->b_flags & B_INVAL)) {
2303 * Failed write, redirty. Must clear BIO_ERROR to prevent
2304 * pages from being scrapped.
2306 bp->b_ioflags &= ~BIO_ERROR;
2308 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2309 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2311 * Either a failed read I/O or we were asked to free or not
2314 bp->b_flags |= B_INVAL;
2315 if (!LIST_EMPTY(&bp->b_dep))
2317 if (bp->b_flags & B_DELWRI)
2319 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2320 if ((bp->b_flags & B_VMIO) == 0) {
2328 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2329 * is called with B_DELWRI set, the underlying pages may wind up
2330 * getting freed causing a previous write (bdwrite()) to get 'lost'
2331 * because pages associated with a B_DELWRI bp are marked clean.
2333 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2334 * if B_DELWRI is set.
2336 if (bp->b_flags & B_DELWRI)
2337 bp->b_flags &= ~B_RELBUF;
2340 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2341 * constituted, not even NFS buffers now. Two flags effect this. If
2342 * B_INVAL, the struct buf is invalidated but the VM object is kept
2343 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2345 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2346 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2347 * buffer is also B_INVAL because it hits the re-dirtying code above.
2349 * Normally we can do this whether a buffer is B_DELWRI or not. If
2350 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2351 * the commit state and we cannot afford to lose the buffer. If the
2352 * buffer has a background write in progress, we need to keep it
2353 * around to prevent it from being reconstituted and starting a second
2356 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2357 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2358 !(bp->b_vp->v_mount != NULL &&
2359 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2360 !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI))) {
2361 vfs_vmio_invalidate(bp);
2365 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2366 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2368 bp->b_flags &= ~B_NOREUSE;
2369 if (bp->b_vp != NULL)
2374 * If the buffer has junk contents signal it and eventually
2375 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2378 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2379 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2380 bp->b_flags |= B_INVAL;
2381 if (bp->b_flags & B_INVAL) {
2382 if (bp->b_flags & B_DELWRI)
2388 /* buffers with no memory */
2389 if (bp->b_bufsize == 0) {
2393 /* buffers with junk contents */
2394 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2395 (bp->b_ioflags & BIO_ERROR)) {
2396 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2397 if (bp->b_vflags & BV_BKGRDINPROG)
2398 panic("losing buffer 2");
2399 qindex = QUEUE_CLEAN;
2400 bp->b_flags |= B_AGE;
2401 /* remaining buffers */
2402 } else if (bp->b_flags & B_DELWRI)
2403 qindex = QUEUE_DIRTY;
2405 qindex = QUEUE_CLEAN;
2407 binsfree(bp, qindex);
2409 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
2410 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2411 panic("brelse: not dirty");
2414 if (qindex == QUEUE_CLEAN)
2419 * Release a buffer back to the appropriate queue but do not try to free
2420 * it. The buffer is expected to be used again soon.
2422 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2423 * biodone() to requeue an async I/O on completion. It is also used when
2424 * known good buffers need to be requeued but we think we may need the data
2427 * XXX we should be able to leave the B_RELBUF hint set on completion.
2430 bqrelse(struct buf *bp)
2434 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2435 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2436 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2438 qindex = QUEUE_NONE;
2439 if (BUF_LOCKRECURSED(bp)) {
2440 /* do not release to free list */
2444 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2446 if (bp->b_flags & B_MANAGED) {
2447 if (bp->b_flags & B_REMFREE)
2452 /* buffers with stale but valid contents */
2453 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2454 BV_BKGRDERR)) == BV_BKGRDERR) {
2455 BO_LOCK(bp->b_bufobj);
2456 bp->b_vflags &= ~BV_BKGRDERR;
2457 BO_UNLOCK(bp->b_bufobj);
2458 qindex = QUEUE_DIRTY;
2460 if ((bp->b_flags & B_DELWRI) == 0 &&
2461 (bp->b_xflags & BX_VNDIRTY))
2462 panic("bqrelse: not dirty");
2463 if ((bp->b_flags & B_NOREUSE) != 0) {
2467 qindex = QUEUE_CLEAN;
2469 binsfree(bp, qindex);
2474 if (qindex == QUEUE_CLEAN)
2479 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2480 * restore bogus pages.
2483 vfs_vmio_iodone(struct buf *bp)
2489 int bogus, i, iosize;
2491 obj = bp->b_bufobj->bo_object;
2492 KASSERT(obj->paging_in_progress >= bp->b_npages,
2493 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2494 obj->paging_in_progress, bp->b_npages));
2497 KASSERT(vp->v_holdcnt > 0,
2498 ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2499 KASSERT(vp->v_object != NULL,
2500 ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2502 foff = bp->b_offset;
2503 KASSERT(bp->b_offset != NOOFFSET,
2504 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2507 iosize = bp->b_bcount - bp->b_resid;
2508 VM_OBJECT_WLOCK(obj);
2509 for (i = 0; i < bp->b_npages; i++) {
2512 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2517 * cleanup bogus pages, restoring the originals
2520 if (m == bogus_page) {
2522 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2524 panic("biodone: page disappeared!");
2526 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2528 * In the write case, the valid and clean bits are
2529 * already changed correctly ( see bdwrite() ), so we
2530 * only need to do this here in the read case.
2532 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2533 resid)) == 0, ("vfs_vmio_iodone: page %p "
2534 "has unexpected dirty bits", m));
2535 vfs_page_set_valid(bp, foff, m);
2537 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2538 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2539 (intmax_t)foff, (uintmax_t)m->pindex));
2542 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2545 vm_object_pip_wakeupn(obj, bp->b_npages);
2546 VM_OBJECT_WUNLOCK(obj);
2547 if (bogus && buf_mapped(bp)) {
2548 BUF_CHECK_MAPPED(bp);
2549 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2550 bp->b_pages, bp->b_npages);
2555 * Unwire a page held by a buf and place it on the appropriate vm queue.
2558 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2563 if (vm_page_unwire(m, PQ_NONE)) {
2565 * Determine if the page should be freed before adding
2566 * it to the inactive queue.
2568 if (m->valid == 0) {
2569 freed = !vm_page_busied(m);
2572 } else if ((bp->b_flags & B_DIRECT) != 0)
2573 freed = vm_page_try_to_free(m);
2578 * If the page is unlikely to be reused, let the
2579 * VM know. Otherwise, maintain LRU page
2580 * ordering and put the page at the tail of the
2583 if ((bp->b_flags & B_NOREUSE) != 0)
2584 vm_page_deactivate_noreuse(m);
2586 vm_page_deactivate(m);
2593 * Perform page invalidation when a buffer is released. The fully invalid
2594 * pages will be reclaimed later in vfs_vmio_truncate().
2597 vfs_vmio_invalidate(struct buf *bp)
2601 int i, resid, poffset, presid;
2603 if (buf_mapped(bp)) {
2604 BUF_CHECK_MAPPED(bp);
2605 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2607 BUF_CHECK_UNMAPPED(bp);
2609 * Get the base offset and length of the buffer. Note that
2610 * in the VMIO case if the buffer block size is not
2611 * page-aligned then b_data pointer may not be page-aligned.
2612 * But our b_pages[] array *IS* page aligned.
2614 * block sizes less then DEV_BSIZE (usually 512) are not
2615 * supported due to the page granularity bits (m->valid,
2616 * m->dirty, etc...).
2618 * See man buf(9) for more information
2620 obj = bp->b_bufobj->bo_object;
2621 resid = bp->b_bufsize;
2622 poffset = bp->b_offset & PAGE_MASK;
2623 VM_OBJECT_WLOCK(obj);
2624 for (i = 0; i < bp->b_npages; i++) {
2626 if (m == bogus_page)
2627 panic("vfs_vmio_invalidate: Unexpected bogus page.");
2628 bp->b_pages[i] = NULL;
2630 presid = resid > (PAGE_SIZE - poffset) ?
2631 (PAGE_SIZE - poffset) : resid;
2632 KASSERT(presid >= 0, ("brelse: extra page"));
2633 while (vm_page_xbusied(m)) {
2635 VM_OBJECT_WUNLOCK(obj);
2636 vm_page_busy_sleep(m, "mbncsh");
2637 VM_OBJECT_WLOCK(obj);
2639 if (pmap_page_wired_mappings(m) == 0)
2640 vm_page_set_invalid(m, poffset, presid);
2641 vfs_vmio_unwire(bp, m);
2645 VM_OBJECT_WUNLOCK(obj);
2650 * Page-granular truncation of an existing VMIO buffer.
2653 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2659 if (bp->b_npages == desiredpages)
2662 if (buf_mapped(bp)) {
2663 BUF_CHECK_MAPPED(bp);
2664 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2665 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2667 BUF_CHECK_UNMAPPED(bp);
2668 obj = bp->b_bufobj->bo_object;
2670 VM_OBJECT_WLOCK(obj);
2671 for (i = desiredpages; i < bp->b_npages; i++) {
2673 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
2674 bp->b_pages[i] = NULL;
2675 vfs_vmio_unwire(bp, m);
2678 VM_OBJECT_WUNLOCK(obj);
2679 bp->b_npages = desiredpages;
2683 * Byte granular extension of VMIO buffers.
2686 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
2689 * We are growing the buffer, possibly in a
2690 * byte-granular fashion.
2698 * Step 1, bring in the VM pages from the object, allocating
2699 * them if necessary. We must clear B_CACHE if these pages
2700 * are not valid for the range covered by the buffer.
2702 obj = bp->b_bufobj->bo_object;
2703 VM_OBJECT_WLOCK(obj);
2704 while (bp->b_npages < desiredpages) {
2706 * We must allocate system pages since blocking
2707 * here could interfere with paging I/O, no
2708 * matter which process we are.
2710 * Only exclusive busy can be tested here.
2711 * Blocking on shared busy might lead to
2712 * deadlocks once allocbuf() is called after
2713 * pages are vfs_busy_pages().
2715 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages,
2716 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2717 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY |
2718 VM_ALLOC_COUNT(desiredpages - bp->b_npages));
2720 bp->b_flags &= ~B_CACHE;
2721 bp->b_pages[bp->b_npages] = m;
2726 * Step 2. We've loaded the pages into the buffer,
2727 * we have to figure out if we can still have B_CACHE
2728 * set. Note that B_CACHE is set according to the
2729 * byte-granular range ( bcount and size ), not the
2730 * aligned range ( newbsize ).
2732 * The VM test is against m->valid, which is DEV_BSIZE
2733 * aligned. Needless to say, the validity of the data
2734 * needs to also be DEV_BSIZE aligned. Note that this
2735 * fails with NFS if the server or some other client
2736 * extends the file's EOF. If our buffer is resized,
2737 * B_CACHE may remain set! XXX
2739 toff = bp->b_bcount;
2740 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2741 while ((bp->b_flags & B_CACHE) && toff < size) {
2744 if (tinc > (size - toff))
2746 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
2747 m = bp->b_pages[pi];
2748 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
2752 VM_OBJECT_WUNLOCK(obj);
2755 * Step 3, fixup the KVA pmap.
2760 BUF_CHECK_UNMAPPED(bp);
2764 * Check to see if a block at a particular lbn is available for a clustered
2768 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
2775 /* If the buf isn't in core skip it */
2776 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
2779 /* If the buf is busy we don't want to wait for it */
2780 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2783 /* Only cluster with valid clusterable delayed write buffers */
2784 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
2785 (B_DELWRI | B_CLUSTEROK))
2788 if (bpa->b_bufsize != size)
2792 * Check to see if it is in the expected place on disk and that the
2793 * block has been mapped.
2795 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
2805 * Implement clustered async writes for clearing out B_DELWRI buffers.
2806 * This is much better then the old way of writing only one buffer at
2807 * a time. Note that we may not be presented with the buffers in the
2808 * correct order, so we search for the cluster in both directions.
2811 vfs_bio_awrite(struct buf *bp)
2816 daddr_t lblkno = bp->b_lblkno;
2817 struct vnode *vp = bp->b_vp;
2825 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
2827 * right now we support clustered writing only to regular files. If
2828 * we find a clusterable block we could be in the middle of a cluster
2829 * rather then at the beginning.
2831 if ((vp->v_type == VREG) &&
2832 (vp->v_mount != 0) && /* Only on nodes that have the size info */
2833 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
2835 size = vp->v_mount->mnt_stat.f_iosize;
2836 maxcl = MAXPHYS / size;
2839 for (i = 1; i < maxcl; i++)
2840 if (vfs_bio_clcheck(vp, size, lblkno + i,
2841 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2844 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2845 if (vfs_bio_clcheck(vp, size, lblkno - j,
2846 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2852 * this is a possible cluster write
2856 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2862 bp->b_flags |= B_ASYNC;
2864 * default (old) behavior, writing out only one block
2866 * XXX returns b_bufsize instead of b_bcount for nwritten?
2868 nwritten = bp->b_bufsize;
2877 * Allocate KVA for an empty buf header according to gbflags.
2880 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
2883 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
2885 * In order to keep fragmentation sane we only allocate kva
2886 * in BKVASIZE chunks. XXX with vmem we can do page size.
2888 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2890 if (maxsize != bp->b_kvasize &&
2891 bufkva_alloc(bp, maxsize, gbflags))
2900 * Find and initialize a new buffer header, freeing up existing buffers
2901 * in the bufqueues as necessary. The new buffer is returned locked.
2904 * We have insufficient buffer headers
2905 * We have insufficient buffer space
2906 * buffer_arena is too fragmented ( space reservation fails )
2907 * If we have to flush dirty buffers ( but we try to avoid this )
2909 * The caller is responsible for releasing the reserved bufspace after
2910 * allocbuf() is called.
2913 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
2916 bool metadata, reserved;
2919 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2920 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2921 if (!unmapped_buf_allowed)
2922 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2924 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2929 atomic_add_int(&getnewbufcalls, 1);
2932 if (reserved == false &&
2933 bufspace_reserve(maxsize, metadata) != 0)
2936 if ((bp = buf_alloc()) == NULL)
2938 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
2941 } while(buf_scan(false) == 0);
2944 atomic_subtract_long(&bufspace, maxsize);
2946 bp->b_flags |= B_INVAL;
2949 bufspace_wait(vp, gbflags, slpflag, slptimeo);
2957 * buffer flushing daemon. Buffers are normally flushed by the
2958 * update daemon but if it cannot keep up this process starts to
2959 * take the load in an attempt to prevent getnewbuf() from blocking.
2961 static struct kproc_desc buf_kp = {
2966 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2969 buf_flush(struct vnode *vp, int target)
2973 flushed = flushbufqueues(vp, target, 0);
2976 * Could not find any buffers without rollback
2977 * dependencies, so just write the first one
2978 * in the hopes of eventually making progress.
2980 if (vp != NULL && target > 2)
2982 flushbufqueues(vp, target, 1);
2993 * This process needs to be suspended prior to shutdown sync.
2995 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2999 * This process is allowed to take the buffer cache to the limit
3001 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3005 mtx_unlock(&bdlock);
3007 kproc_suspend_check(bufdaemonproc);
3008 lodirty = lodirtybuffers;
3009 if (bd_speedupreq) {
3010 lodirty = numdirtybuffers / 2;
3014 * Do the flush. Limit the amount of in-transit I/O we
3015 * allow to build up, otherwise we would completely saturate
3018 while (numdirtybuffers > lodirty) {
3019 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
3021 kern_yield(PRI_USER);
3025 * Only clear bd_request if we have reached our low water
3026 * mark. The buf_daemon normally waits 1 second and
3027 * then incrementally flushes any dirty buffers that have
3028 * built up, within reason.
3030 * If we were unable to hit our low water mark and couldn't
3031 * find any flushable buffers, we sleep for a short period
3032 * to avoid endless loops on unlockable buffers.
3035 if (numdirtybuffers <= lodirtybuffers) {
3037 * We reached our low water mark, reset the
3038 * request and sleep until we are needed again.
3039 * The sleep is just so the suspend code works.
3043 * Do an extra wakeup in case dirty threshold
3044 * changed via sysctl and the explicit transition
3045 * out of shortfall was missed.
3048 if (runningbufspace <= lorunningspace)
3050 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3053 * We couldn't find any flushable dirty buffers but
3054 * still have too many dirty buffers, we
3055 * have to sleep and try again. (rare)
3057 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3065 * Try to flush a buffer in the dirty queue. We must be careful to
3066 * free up B_INVAL buffers instead of write them, which NFS is
3067 * particularly sensitive to.
3069 static int flushwithdeps = 0;
3070 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3071 0, "Number of buffers flushed with dependecies that require rollbacks");
3074 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
3076 struct buf *sentinel;
3087 queue = QUEUE_DIRTY;
3089 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3090 sentinel->b_qindex = QUEUE_SENTINEL;
3091 mtx_lock(&bqlocks[queue]);
3092 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
3093 mtx_unlock(&bqlocks[queue]);
3094 while (flushed != target) {
3096 mtx_lock(&bqlocks[queue]);
3097 bp = TAILQ_NEXT(sentinel, b_freelist);
3099 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3100 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
3103 mtx_unlock(&bqlocks[queue]);
3107 * Skip sentinels inserted by other invocations of the
3108 * flushbufqueues(), taking care to not reorder them.
3110 * Only flush the buffers that belong to the
3111 * vnode locked by the curthread.
3113 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3115 mtx_unlock(&bqlocks[queue]);
3118 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3119 mtx_unlock(&bqlocks[queue]);
3124 * BKGRDINPROG can only be set with the buf and bufobj
3125 * locks both held. We tolerate a race to clear it here.
3127 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3128 (bp->b_flags & B_DELWRI) == 0) {
3132 if (bp->b_flags & B_INVAL) {
3139 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3140 if (flushdeps == 0) {
3148 * We must hold the lock on a vnode before writing
3149 * one of its buffers. Otherwise we may confuse, or
3150 * in the case of a snapshot vnode, deadlock the
3153 * The lock order here is the reverse of the normal
3154 * of vnode followed by buf lock. This is ok because
3155 * the NOWAIT will prevent deadlock.
3158 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3164 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3166 ASSERT_VOP_LOCKED(vp, "getbuf");
3168 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3169 vn_lock(vp, LK_TRYUPGRADE);
3172 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3173 bp, bp->b_vp, bp->b_flags);
3174 if (curproc == bufdaemonproc) {
3181 vn_finished_write(mp);
3184 flushwithdeps += hasdeps;
3188 * Sleeping on runningbufspace while holding
3189 * vnode lock leads to deadlock.
3191 if (curproc == bufdaemonproc &&
3192 runningbufspace > hirunningspace)
3193 waitrunningbufspace();
3196 vn_finished_write(mp);
3199 mtx_lock(&bqlocks[queue]);
3200 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3201 mtx_unlock(&bqlocks[queue]);
3202 free(sentinel, M_TEMP);
3207 * Check to see if a block is currently memory resident.
3210 incore(struct bufobj *bo, daddr_t blkno)
3215 bp = gbincore(bo, blkno);
3221 * Returns true if no I/O is needed to access the
3222 * associated VM object. This is like incore except
3223 * it also hunts around in the VM system for the data.
3227 inmem(struct vnode * vp, daddr_t blkno)
3230 vm_offset_t toff, tinc, size;
3234 ASSERT_VOP_LOCKED(vp, "inmem");
3236 if (incore(&vp->v_bufobj, blkno))
3238 if (vp->v_mount == NULL)
3245 if (size > vp->v_mount->mnt_stat.f_iosize)
3246 size = vp->v_mount->mnt_stat.f_iosize;
3247 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3249 VM_OBJECT_RLOCK(obj);
3250 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3251 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3255 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3256 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3257 if (vm_page_is_valid(m,
3258 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3261 VM_OBJECT_RUNLOCK(obj);
3265 VM_OBJECT_RUNLOCK(obj);
3270 * Set the dirty range for a buffer based on the status of the dirty
3271 * bits in the pages comprising the buffer. The range is limited
3272 * to the size of the buffer.
3274 * Tell the VM system that the pages associated with this buffer
3275 * are clean. This is used for delayed writes where the data is
3276 * going to go to disk eventually without additional VM intevention.
3278 * Note that while we only really need to clean through to b_bcount, we
3279 * just go ahead and clean through to b_bufsize.
3282 vfs_clean_pages_dirty_buf(struct buf *bp)
3284 vm_ooffset_t foff, noff, eoff;
3288 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3291 foff = bp->b_offset;
3292 KASSERT(bp->b_offset != NOOFFSET,
3293 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3295 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3296 vfs_drain_busy_pages(bp);
3297 vfs_setdirty_locked_object(bp);
3298 for (i = 0; i < bp->b_npages; i++) {
3299 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3301 if (eoff > bp->b_offset + bp->b_bufsize)
3302 eoff = bp->b_offset + bp->b_bufsize;
3304 vfs_page_set_validclean(bp, foff, m);
3305 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3308 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3312 vfs_setdirty_locked_object(struct buf *bp)
3317 object = bp->b_bufobj->bo_object;
3318 VM_OBJECT_ASSERT_WLOCKED(object);
3321 * We qualify the scan for modified pages on whether the
3322 * object has been flushed yet.
3324 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3325 vm_offset_t boffset;
3326 vm_offset_t eoffset;
3329 * test the pages to see if they have been modified directly
3330 * by users through the VM system.
3332 for (i = 0; i < bp->b_npages; i++)
3333 vm_page_test_dirty(bp->b_pages[i]);
3336 * Calculate the encompassing dirty range, boffset and eoffset,
3337 * (eoffset - boffset) bytes.
3340 for (i = 0; i < bp->b_npages; i++) {
3341 if (bp->b_pages[i]->dirty)
3344 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3346 for (i = bp->b_npages - 1; i >= 0; --i) {
3347 if (bp->b_pages[i]->dirty) {
3351 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3354 * Fit it to the buffer.
3357 if (eoffset > bp->b_bcount)
3358 eoffset = bp->b_bcount;
3361 * If we have a good dirty range, merge with the existing
3365 if (boffset < eoffset) {
3366 if (bp->b_dirtyoff > boffset)
3367 bp->b_dirtyoff = boffset;
3368 if (bp->b_dirtyend < eoffset)
3369 bp->b_dirtyend = eoffset;
3375 * Allocate the KVA mapping for an existing buffer.
3376 * If an unmapped buffer is provided but a mapped buffer is requested, take
3377 * also care to properly setup mappings between pages and KVA.
3380 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3382 int bsize, maxsize, need_mapping, need_kva;
3385 need_mapping = bp->b_data == unmapped_buf &&
3386 (gbflags & GB_UNMAPPED) == 0;
3387 need_kva = bp->b_kvabase == unmapped_buf &&
3388 bp->b_data == unmapped_buf &&
3389 (gbflags & GB_KVAALLOC) != 0;
3390 if (!need_mapping && !need_kva)
3393 BUF_CHECK_UNMAPPED(bp);
3395 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3397 * Buffer is not mapped, but the KVA was already
3398 * reserved at the time of the instantiation. Use the
3405 * Calculate the amount of the address space we would reserve
3406 * if the buffer was mapped.
3408 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3409 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3410 offset = blkno * bsize;
3411 maxsize = size + (offset & PAGE_MASK);
3412 maxsize = imax(maxsize, bsize);
3414 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3415 if ((gbflags & GB_NOWAIT_BD) != 0) {
3417 * XXXKIB: defragmentation cannot
3418 * succeed, not sure what else to do.
3420 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3422 atomic_add_int(&mappingrestarts, 1);
3423 bufspace_wait(bp->b_vp, gbflags, 0, 0);
3427 /* b_offset is handled by bpmap_qenter. */
3428 bp->b_data = bp->b_kvabase;
3429 BUF_CHECK_MAPPED(bp);
3437 * Get a block given a specified block and offset into a file/device.
3438 * The buffers B_DONE bit will be cleared on return, making it almost
3439 * ready for an I/O initiation. B_INVAL may or may not be set on
3440 * return. The caller should clear B_INVAL prior to initiating a
3443 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3444 * an existing buffer.
3446 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3447 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3448 * and then cleared based on the backing VM. If the previous buffer is
3449 * non-0-sized but invalid, B_CACHE will be cleared.
3451 * If getblk() must create a new buffer, the new buffer is returned with
3452 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3453 * case it is returned with B_INVAL clear and B_CACHE set based on the
3456 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3457 * B_CACHE bit is clear.
3459 * What this means, basically, is that the caller should use B_CACHE to
3460 * determine whether the buffer is fully valid or not and should clear
3461 * B_INVAL prior to issuing a read. If the caller intends to validate
3462 * the buffer by loading its data area with something, the caller needs
3463 * to clear B_INVAL. If the caller does this without issuing an I/O,
3464 * the caller should set B_CACHE ( as an optimization ), else the caller
3465 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3466 * a write attempt or if it was a successful read. If the caller
3467 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3468 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3471 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3476 int bsize, error, maxsize, vmio;
3479 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3480 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3481 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3482 ASSERT_VOP_LOCKED(vp, "getblk");
3483 if (size > MAXBCACHEBUF)
3484 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size,
3486 if (!unmapped_buf_allowed)
3487 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3492 bp = gbincore(bo, blkno);
3496 * Buffer is in-core. If the buffer is not busy nor managed,
3497 * it must be on a queue.
3499 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3501 if (flags & GB_LOCK_NOWAIT)
3502 lockflags |= LK_NOWAIT;
3504 error = BUF_TIMELOCK(bp, lockflags,
3505 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3508 * If we slept and got the lock we have to restart in case
3509 * the buffer changed identities.
3511 if (error == ENOLCK)
3513 /* We timed out or were interrupted. */
3516 /* If recursed, assume caller knows the rules. */
3517 else if (BUF_LOCKRECURSED(bp))
3521 * The buffer is locked. B_CACHE is cleared if the buffer is
3522 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3523 * and for a VMIO buffer B_CACHE is adjusted according to the
3526 if (bp->b_flags & B_INVAL)
3527 bp->b_flags &= ~B_CACHE;
3528 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3529 bp->b_flags |= B_CACHE;
3530 if (bp->b_flags & B_MANAGED)
3531 MPASS(bp->b_qindex == QUEUE_NONE);
3536 * check for size inconsistencies for non-VMIO case.
3538 if (bp->b_bcount != size) {
3539 if ((bp->b_flags & B_VMIO) == 0 ||
3540 (size > bp->b_kvasize)) {
3541 if (bp->b_flags & B_DELWRI) {
3542 bp->b_flags |= B_NOCACHE;
3545 if (LIST_EMPTY(&bp->b_dep)) {
3546 bp->b_flags |= B_RELBUF;
3549 bp->b_flags |= B_NOCACHE;
3558 * Handle the case of unmapped buffer which should
3559 * become mapped, or the buffer for which KVA
3560 * reservation is requested.
3562 bp_unmapped_get_kva(bp, blkno, size, flags);
3565 * If the size is inconsistent in the VMIO case, we can resize
3566 * the buffer. This might lead to B_CACHE getting set or
3567 * cleared. If the size has not changed, B_CACHE remains
3568 * unchanged from its previous state.
3572 KASSERT(bp->b_offset != NOOFFSET,
3573 ("getblk: no buffer offset"));
3576 * A buffer with B_DELWRI set and B_CACHE clear must
3577 * be committed before we can return the buffer in
3578 * order to prevent the caller from issuing a read
3579 * ( due to B_CACHE not being set ) and overwriting
3582 * Most callers, including NFS and FFS, need this to
3583 * operate properly either because they assume they
3584 * can issue a read if B_CACHE is not set, or because
3585 * ( for example ) an uncached B_DELWRI might loop due
3586 * to softupdates re-dirtying the buffer. In the latter
3587 * case, B_CACHE is set after the first write completes,
3588 * preventing further loops.
3589 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3590 * above while extending the buffer, we cannot allow the
3591 * buffer to remain with B_CACHE set after the write
3592 * completes or it will represent a corrupt state. To
3593 * deal with this we set B_NOCACHE to scrap the buffer
3596 * We might be able to do something fancy, like setting
3597 * B_CACHE in bwrite() except if B_DELWRI is already set,
3598 * so the below call doesn't set B_CACHE, but that gets real
3599 * confusing. This is much easier.
3602 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3603 bp->b_flags |= B_NOCACHE;
3607 bp->b_flags &= ~B_DONE;
3610 * Buffer is not in-core, create new buffer. The buffer
3611 * returned by getnewbuf() is locked. Note that the returned
3612 * buffer is also considered valid (not marked B_INVAL).
3616 * If the user does not want us to create the buffer, bail out
3619 if (flags & GB_NOCREAT)
3621 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3624 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3625 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3626 offset = blkno * bsize;
3627 vmio = vp->v_object != NULL;
3629 maxsize = size + (offset & PAGE_MASK);
3632 /* Do not allow non-VMIO notmapped buffers. */
3633 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3635 maxsize = imax(maxsize, bsize);
3637 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
3639 if (slpflag || slptimeo)
3642 * XXX This is here until the sleep path is diagnosed
3643 * enough to work under very low memory conditions.
3645 * There's an issue on low memory, 4BSD+non-preempt
3646 * systems (eg MIPS routers with 32MB RAM) where buffer
3647 * exhaustion occurs without sleeping for buffer
3648 * reclaimation. This just sticks in a loop and
3649 * constantly attempts to allocate a buffer, which
3650 * hits exhaustion and tries to wakeup bufdaemon.
3651 * This never happens because we never yield.
3653 * The real solution is to identify and fix these cases
3654 * so we aren't effectively busy-waiting in a loop
3655 * until the reclaimation path has cycles to run.
3657 kern_yield(PRI_USER);
3662 * This code is used to make sure that a buffer is not
3663 * created while the getnewbuf routine is blocked.
3664 * This can be a problem whether the vnode is locked or not.
3665 * If the buffer is created out from under us, we have to
3666 * throw away the one we just created.
3668 * Note: this must occur before we associate the buffer
3669 * with the vp especially considering limitations in
3670 * the splay tree implementation when dealing with duplicate
3674 if (gbincore(bo, blkno)) {
3676 bp->b_flags |= B_INVAL;
3678 bufspace_release(maxsize);
3683 * Insert the buffer into the hash, so that it can
3684 * be found by incore.
3686 bp->b_blkno = bp->b_lblkno = blkno;
3687 bp->b_offset = offset;
3692 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
3693 * buffer size starts out as 0, B_CACHE will be set by
3694 * allocbuf() for the VMIO case prior to it testing the
3695 * backing store for validity.
3699 bp->b_flags |= B_VMIO;
3700 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3701 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3702 bp, vp->v_object, bp->b_bufobj->bo_object));
3704 bp->b_flags &= ~B_VMIO;
3705 KASSERT(bp->b_bufobj->bo_object == NULL,
3706 ("ARGH! has b_bufobj->bo_object %p %p\n",
3707 bp, bp->b_bufobj->bo_object));
3708 BUF_CHECK_MAPPED(bp);
3712 bufspace_release(maxsize);
3713 bp->b_flags &= ~B_DONE;
3715 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3716 BUF_ASSERT_HELD(bp);
3718 KASSERT(bp->b_bufobj == bo,
3719 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3724 * Get an empty, disassociated buffer of given size. The buffer is initially
3728 geteblk(int size, int flags)
3733 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3734 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
3735 if ((flags & GB_NOWAIT_BD) &&
3736 (curthread->td_pflags & TDP_BUFNEED) != 0)
3740 bufspace_release(maxsize);
3741 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3742 BUF_ASSERT_HELD(bp);
3747 * Truncate the backing store for a non-vmio buffer.
3750 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
3753 if (bp->b_flags & B_MALLOC) {
3755 * malloced buffers are not shrunk
3757 if (newbsize == 0) {
3758 bufmallocadjust(bp, 0);
3759 free(bp->b_data, M_BIOBUF);
3760 bp->b_data = bp->b_kvabase;
3761 bp->b_flags &= ~B_MALLOC;
3765 vm_hold_free_pages(bp, newbsize);
3766 bufspace_adjust(bp, newbsize);
3770 * Extend the backing for a non-VMIO buffer.
3773 vfs_nonvmio_extend(struct buf *bp, int newbsize)
3779 * We only use malloced memory on the first allocation.
3780 * and revert to page-allocated memory when the buffer
3783 * There is a potential smp race here that could lead
3784 * to bufmallocspace slightly passing the max. It
3785 * is probably extremely rare and not worth worrying
3788 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
3789 bufmallocspace < maxbufmallocspace) {
3790 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
3791 bp->b_flags |= B_MALLOC;
3792 bufmallocadjust(bp, newbsize);
3797 * If the buffer is growing on its other-than-first
3798 * allocation then we revert to the page-allocation
3803 if (bp->b_flags & B_MALLOC) {
3804 origbuf = bp->b_data;
3805 origbufsize = bp->b_bufsize;
3806 bp->b_data = bp->b_kvabase;
3807 bufmallocadjust(bp, 0);
3808 bp->b_flags &= ~B_MALLOC;
3809 newbsize = round_page(newbsize);
3811 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
3812 (vm_offset_t) bp->b_data + newbsize);
3813 if (origbuf != NULL) {
3814 bcopy(origbuf, bp->b_data, origbufsize);
3815 free(origbuf, M_BIOBUF);
3817 bufspace_adjust(bp, newbsize);
3821 * This code constitutes the buffer memory from either anonymous system
3822 * memory (in the case of non-VMIO operations) or from an associated
3823 * VM object (in the case of VMIO operations). This code is able to
3824 * resize a buffer up or down.
3826 * Note that this code is tricky, and has many complications to resolve
3827 * deadlock or inconsistent data situations. Tread lightly!!!
3828 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3829 * the caller. Calling this code willy nilly can result in the loss of data.
3831 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3832 * B_CACHE for the non-VMIO case.
3835 allocbuf(struct buf *bp, int size)
3839 BUF_ASSERT_HELD(bp);
3841 if (bp->b_bcount == size)
3844 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
3845 panic("allocbuf: buffer too small");
3847 newbsize = roundup2(size, DEV_BSIZE);
3848 if ((bp->b_flags & B_VMIO) == 0) {
3849 if ((bp->b_flags & B_MALLOC) == 0)
3850 newbsize = round_page(newbsize);
3852 * Just get anonymous memory from the kernel. Don't
3853 * mess with B_CACHE.
3855 if (newbsize < bp->b_bufsize)
3856 vfs_nonvmio_truncate(bp, newbsize);
3857 else if (newbsize > bp->b_bufsize)
3858 vfs_nonvmio_extend(bp, newbsize);
3862 desiredpages = (size == 0) ? 0 :
3863 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3865 if (bp->b_flags & B_MALLOC)
3866 panic("allocbuf: VMIO buffer can't be malloced");
3868 * Set B_CACHE initially if buffer is 0 length or will become
3871 if (size == 0 || bp->b_bufsize == 0)
3872 bp->b_flags |= B_CACHE;
3874 if (newbsize < bp->b_bufsize)
3875 vfs_vmio_truncate(bp, desiredpages);
3876 /* XXX This looks as if it should be newbsize > b_bufsize */
3877 else if (size > bp->b_bcount)
3878 vfs_vmio_extend(bp, desiredpages, size);
3879 bufspace_adjust(bp, newbsize);
3881 bp->b_bcount = size; /* requested buffer size. */
3885 extern int inflight_transient_maps;
3888 biodone(struct bio *bp)
3891 void (*done)(struct bio *);
3892 vm_offset_t start, end;
3894 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3895 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3896 bp->bio_flags |= BIO_UNMAPPED;
3897 start = trunc_page((vm_offset_t)bp->bio_data);
3898 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3899 bp->bio_data = unmapped_buf;
3900 pmap_qremove(start, OFF_TO_IDX(end - start));
3901 vmem_free(transient_arena, start, end - start);
3902 atomic_add_int(&inflight_transient_maps, -1);
3904 done = bp->bio_done;
3906 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3908 bp->bio_flags |= BIO_DONE;
3912 bp->bio_flags |= BIO_DONE;
3918 * Wait for a BIO to finish.
3921 biowait(struct bio *bp, const char *wchan)
3925 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3927 while ((bp->bio_flags & BIO_DONE) == 0)
3928 msleep(bp, mtxp, PRIBIO, wchan, 0);
3930 if (bp->bio_error != 0)
3931 return (bp->bio_error);
3932 if (!(bp->bio_flags & BIO_ERROR))
3938 biofinish(struct bio *bp, struct devstat *stat, int error)
3942 bp->bio_error = error;
3943 bp->bio_flags |= BIO_ERROR;
3946 devstat_end_transaction_bio(stat, bp);
3953 * Wait for buffer I/O completion, returning error status. The buffer
3954 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3955 * error and cleared.
3958 bufwait(struct buf *bp)
3960 if (bp->b_iocmd == BIO_READ)
3961 bwait(bp, PRIBIO, "biord");
3963 bwait(bp, PRIBIO, "biowr");
3964 if (bp->b_flags & B_EINTR) {
3965 bp->b_flags &= ~B_EINTR;
3968 if (bp->b_ioflags & BIO_ERROR) {
3969 return (bp->b_error ? bp->b_error : EIO);
3978 * Finish I/O on a buffer, optionally calling a completion function.
3979 * This is usually called from an interrupt so process blocking is
3982 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3983 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3984 * assuming B_INVAL is clear.
3986 * For the VMIO case, we set B_CACHE if the op was a read and no
3987 * read error occurred, or if the op was a write. B_CACHE is never
3988 * set if the buffer is invalid or otherwise uncacheable.
3990 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3991 * initiator to leave B_INVAL set to brelse the buffer out of existence
3992 * in the biodone routine.
3995 bufdone(struct buf *bp)
3997 struct bufobj *dropobj;
3998 void (*biodone)(struct buf *);
4000 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4003 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4004 BUF_ASSERT_HELD(bp);
4006 runningbufwakeup(bp);
4007 if (bp->b_iocmd == BIO_WRITE)
4008 dropobj = bp->b_bufobj;
4009 /* call optional completion function if requested */
4010 if (bp->b_iodone != NULL) {
4011 biodone = bp->b_iodone;
4012 bp->b_iodone = NULL;
4015 bufobj_wdrop(dropobj);
4022 bufobj_wdrop(dropobj);
4026 bufdone_finish(struct buf *bp)
4028 BUF_ASSERT_HELD(bp);
4030 if (!LIST_EMPTY(&bp->b_dep))
4033 if (bp->b_flags & B_VMIO) {
4035 * Set B_CACHE if the op was a normal read and no error
4036 * occurred. B_CACHE is set for writes in the b*write()
4039 if (bp->b_iocmd == BIO_READ &&
4040 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4041 !(bp->b_ioflags & BIO_ERROR))
4042 bp->b_flags |= B_CACHE;
4043 vfs_vmio_iodone(bp);
4047 * For asynchronous completions, release the buffer now. The brelse
4048 * will do a wakeup there if necessary - so no need to do a wakeup
4049 * here in the async case. The sync case always needs to do a wakeup.
4051 if (bp->b_flags & B_ASYNC) {
4052 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4053 (bp->b_ioflags & BIO_ERROR))
4062 * This routine is called in lieu of iodone in the case of
4063 * incomplete I/O. This keeps the busy status for pages
4067 vfs_unbusy_pages(struct buf *bp)
4073 runningbufwakeup(bp);
4074 if (!(bp->b_flags & B_VMIO))
4077 obj = bp->b_bufobj->bo_object;
4078 VM_OBJECT_WLOCK(obj);
4079 for (i = 0; i < bp->b_npages; i++) {
4081 if (m == bogus_page) {
4082 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4084 panic("vfs_unbusy_pages: page missing\n");
4086 if (buf_mapped(bp)) {
4087 BUF_CHECK_MAPPED(bp);
4088 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4089 bp->b_pages, bp->b_npages);
4091 BUF_CHECK_UNMAPPED(bp);
4095 vm_object_pip_wakeupn(obj, bp->b_npages);
4096 VM_OBJECT_WUNLOCK(obj);
4100 * vfs_page_set_valid:
4102 * Set the valid bits in a page based on the supplied offset. The
4103 * range is restricted to the buffer's size.
4105 * This routine is typically called after a read completes.
4108 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4113 * Compute the end offset, eoff, such that [off, eoff) does not span a
4114 * page boundary and eoff is not greater than the end of the buffer.
4115 * The end of the buffer, in this case, is our file EOF, not the
4116 * allocation size of the buffer.
4118 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4119 if (eoff > bp->b_offset + bp->b_bcount)
4120 eoff = bp->b_offset + bp->b_bcount;
4123 * Set valid range. This is typically the entire buffer and thus the
4127 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4131 * vfs_page_set_validclean:
4133 * Set the valid bits and clear the dirty bits in a page based on the
4134 * supplied offset. The range is restricted to the buffer's size.
4137 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4139 vm_ooffset_t soff, eoff;
4142 * Start and end offsets in buffer. eoff - soff may not cross a
4143 * page boundary or cross the end of the buffer. The end of the
4144 * buffer, in this case, is our file EOF, not the allocation size
4148 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4149 if (eoff > bp->b_offset + bp->b_bcount)
4150 eoff = bp->b_offset + bp->b_bcount;
4153 * Set valid range. This is typically the entire buffer and thus the
4157 vm_page_set_validclean(
4159 (vm_offset_t) (soff & PAGE_MASK),
4160 (vm_offset_t) (eoff - soff)
4166 * Ensure that all buffer pages are not exclusive busied. If any page is
4167 * exclusive busy, drain it.
4170 vfs_drain_busy_pages(struct buf *bp)
4175 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4177 for (i = 0; i < bp->b_npages; i++) {
4179 if (vm_page_xbusied(m)) {
4180 for (; last_busied < i; last_busied++)
4181 vm_page_sbusy(bp->b_pages[last_busied]);
4182 while (vm_page_xbusied(m)) {
4184 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4185 vm_page_busy_sleep(m, "vbpage");
4186 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4190 for (i = 0; i < last_busied; i++)
4191 vm_page_sunbusy(bp->b_pages[i]);
4195 * This routine is called before a device strategy routine.
4196 * It is used to tell the VM system that paging I/O is in
4197 * progress, and treat the pages associated with the buffer
4198 * almost as being exclusive busy. Also the object paging_in_progress
4199 * flag is handled to make sure that the object doesn't become
4202 * Since I/O has not been initiated yet, certain buffer flags
4203 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4204 * and should be ignored.
4207 vfs_busy_pages(struct buf *bp, int clear_modify)
4214 if (!(bp->b_flags & B_VMIO))
4217 obj = bp->b_bufobj->bo_object;
4218 foff = bp->b_offset;
4219 KASSERT(bp->b_offset != NOOFFSET,
4220 ("vfs_busy_pages: no buffer offset"));
4221 VM_OBJECT_WLOCK(obj);
4222 vfs_drain_busy_pages(bp);
4223 if (bp->b_bufsize != 0)
4224 vfs_setdirty_locked_object(bp);
4226 for (i = 0; i < bp->b_npages; i++) {
4229 if ((bp->b_flags & B_CLUSTER) == 0) {
4230 vm_object_pip_add(obj, 1);
4234 * When readying a buffer for a read ( i.e
4235 * clear_modify == 0 ), it is important to do
4236 * bogus_page replacement for valid pages in
4237 * partially instantiated buffers. Partially
4238 * instantiated buffers can, in turn, occur when
4239 * reconstituting a buffer from its VM backing store
4240 * base. We only have to do this if B_CACHE is
4241 * clear ( which causes the I/O to occur in the
4242 * first place ). The replacement prevents the read
4243 * I/O from overwriting potentially dirty VM-backed
4244 * pages. XXX bogus page replacement is, uh, bogus.
4245 * It may not work properly with small-block devices.
4246 * We need to find a better way.
4249 pmap_remove_write(m);
4250 vfs_page_set_validclean(bp, foff, m);
4251 } else if (m->valid == VM_PAGE_BITS_ALL &&
4252 (bp->b_flags & B_CACHE) == 0) {
4253 bp->b_pages[i] = bogus_page;
4256 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4258 VM_OBJECT_WUNLOCK(obj);
4259 if (bogus && buf_mapped(bp)) {
4260 BUF_CHECK_MAPPED(bp);
4261 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4262 bp->b_pages, bp->b_npages);
4267 * vfs_bio_set_valid:
4269 * Set the range within the buffer to valid. The range is
4270 * relative to the beginning of the buffer, b_offset. Note that
4271 * b_offset itself may be offset from the beginning of the first
4275 vfs_bio_set_valid(struct buf *bp, int base, int size)
4280 if (!(bp->b_flags & B_VMIO))
4284 * Fixup base to be relative to beginning of first page.
4285 * Set initial n to be the maximum number of bytes in the
4286 * first page that can be validated.
4288 base += (bp->b_offset & PAGE_MASK);
4289 n = PAGE_SIZE - (base & PAGE_MASK);
4291 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4292 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4296 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4301 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4307 * If the specified buffer is a non-VMIO buffer, clear the entire
4308 * buffer. If the specified buffer is a VMIO buffer, clear and
4309 * validate only the previously invalid portions of the buffer.
4310 * This routine essentially fakes an I/O, so we need to clear
4311 * BIO_ERROR and B_INVAL.
4313 * Note that while we only theoretically need to clear through b_bcount,
4314 * we go ahead and clear through b_bufsize.
4317 vfs_bio_clrbuf(struct buf *bp)
4319 int i, j, mask, sa, ea, slide;
4321 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4325 bp->b_flags &= ~B_INVAL;
4326 bp->b_ioflags &= ~BIO_ERROR;
4327 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4328 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4329 (bp->b_offset & PAGE_MASK) == 0) {
4330 if (bp->b_pages[0] == bogus_page)
4332 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4333 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4334 if ((bp->b_pages[0]->valid & mask) == mask)
4336 if ((bp->b_pages[0]->valid & mask) == 0) {
4337 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4338 bp->b_pages[0]->valid |= mask;
4342 sa = bp->b_offset & PAGE_MASK;
4344 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4345 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4346 ea = slide & PAGE_MASK;
4349 if (bp->b_pages[i] == bogus_page)
4352 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4353 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4354 if ((bp->b_pages[i]->valid & mask) == mask)
4356 if ((bp->b_pages[i]->valid & mask) == 0)
4357 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4359 for (; sa < ea; sa += DEV_BSIZE, j++) {
4360 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4361 pmap_zero_page_area(bp->b_pages[i],
4366 bp->b_pages[i]->valid |= mask;
4369 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4374 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4379 if (buf_mapped(bp)) {
4380 BUF_CHECK_MAPPED(bp);
4381 bzero(bp->b_data + base, size);
4383 BUF_CHECK_UNMAPPED(bp);
4384 n = PAGE_SIZE - (base & PAGE_MASK);
4385 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4389 pmap_zero_page_area(m, base & PAGE_MASK, n);
4398 * vm_hold_load_pages and vm_hold_free_pages get pages into
4399 * a buffers address space. The pages are anonymous and are
4400 * not associated with a file object.
4403 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4409 BUF_CHECK_MAPPED(bp);
4411 to = round_page(to);
4412 from = round_page(from);
4413 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4415 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4418 * note: must allocate system pages since blocking here
4419 * could interfere with paging I/O, no matter which
4422 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4423 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4428 pmap_qenter(pg, &p, 1);
4429 bp->b_pages[index] = p;
4431 bp->b_npages = index;
4434 /* Return pages associated with this buf to the vm system */
4436 vm_hold_free_pages(struct buf *bp, int newbsize)
4440 int index, newnpages;
4442 BUF_CHECK_MAPPED(bp);
4444 from = round_page((vm_offset_t)bp->b_data + newbsize);
4445 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4446 if (bp->b_npages > newnpages)
4447 pmap_qremove(from, bp->b_npages - newnpages);
4448 for (index = newnpages; index < bp->b_npages; index++) {
4449 p = bp->b_pages[index];
4450 bp->b_pages[index] = NULL;
4451 if (vm_page_sbusied(p))
4452 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4453 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4456 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4458 bp->b_npages = newnpages;
4462 * Map an IO request into kernel virtual address space.
4464 * All requests are (re)mapped into kernel VA space.
4465 * Notice that we use b_bufsize for the size of the buffer
4466 * to be mapped. b_bcount might be modified by the driver.
4468 * Note that even if the caller determines that the address space should
4469 * be valid, a race or a smaller-file mapped into a larger space may
4470 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4471 * check the return value.
4473 * This function only works with pager buffers.
4476 vmapbuf(struct buf *bp, int mapbuf)
4481 if (bp->b_bufsize < 0)
4483 prot = VM_PROT_READ;
4484 if (bp->b_iocmd == BIO_READ)
4485 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4486 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4487 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4488 btoc(MAXPHYS))) < 0)
4490 bp->b_npages = pidx;
4491 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4492 if (mapbuf || !unmapped_buf_allowed) {
4493 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4494 bp->b_data = bp->b_kvabase + bp->b_offset;
4496 bp->b_data = unmapped_buf;
4501 * Free the io map PTEs associated with this IO operation.
4502 * We also invalidate the TLB entries and restore the original b_addr.
4504 * This function only works with pager buffers.
4507 vunmapbuf(struct buf *bp)
4511 npages = bp->b_npages;
4513 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4514 vm_page_unhold_pages(bp->b_pages, npages);
4516 bp->b_data = unmapped_buf;
4520 bdone(struct buf *bp)
4524 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4526 bp->b_flags |= B_DONE;
4532 bwait(struct buf *bp, u_char pri, const char *wchan)
4536 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4538 while ((bp->b_flags & B_DONE) == 0)
4539 msleep(bp, mtxp, pri, wchan, 0);
4544 bufsync(struct bufobj *bo, int waitfor)
4547 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4551 bufstrategy(struct bufobj *bo, struct buf *bp)
4557 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4558 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4559 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4560 i = VOP_STRATEGY(vp, bp);
4561 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4565 bufobj_wrefl(struct bufobj *bo)
4568 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4569 ASSERT_BO_WLOCKED(bo);
4574 bufobj_wref(struct bufobj *bo)
4577 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4584 bufobj_wdrop(struct bufobj *bo)
4587 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4589 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4590 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4591 bo->bo_flag &= ~BO_WWAIT;
4592 wakeup(&bo->bo_numoutput);
4598 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4602 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4603 ASSERT_BO_WLOCKED(bo);
4605 while (bo->bo_numoutput) {
4606 bo->bo_flag |= BO_WWAIT;
4607 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4608 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4616 * Set bio_data or bio_ma for struct bio from the struct buf.
4619 bdata2bio(struct buf *bp, struct bio *bip)
4622 if (!buf_mapped(bp)) {
4623 KASSERT(unmapped_buf_allowed, ("unmapped"));
4624 bip->bio_ma = bp->b_pages;
4625 bip->bio_ma_n = bp->b_npages;
4626 bip->bio_data = unmapped_buf;
4627 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4628 bip->bio_flags |= BIO_UNMAPPED;
4629 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4630 PAGE_SIZE == bp->b_npages,
4631 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4632 (long long)bip->bio_length, bip->bio_ma_n));
4634 bip->bio_data = bp->b_data;
4639 #include "opt_ddb.h"
4641 #include <ddb/ddb.h>
4643 /* DDB command to show buffer data */
4644 DB_SHOW_COMMAND(buffer, db_show_buffer)
4647 struct buf *bp = (struct buf *)addr;
4650 db_printf("usage: show buffer <addr>\n");
4654 db_printf("buf at %p\n", bp);
4655 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4656 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4657 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4659 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4660 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4662 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4663 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4664 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4665 db_printf("b_kvabase = %p, b_kvasize = %d\n",
4666 bp->b_kvabase, bp->b_kvasize);
4669 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4670 for (i = 0; i < bp->b_npages; i++) {
4674 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
4676 (u_long)VM_PAGE_TO_PHYS(m));
4678 db_printf("( ??? )");
4679 if ((i + 1) < bp->b_npages)
4685 BUF_LOCKPRINTINFO(bp);
4688 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4693 for (i = 0; i < nbuf; i++) {
4695 if (BUF_ISLOCKED(bp)) {
4696 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4702 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4708 db_printf("usage: show vnodebufs <addr>\n");
4711 vp = (struct vnode *)addr;
4712 db_printf("Clean buffers:\n");
4713 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4714 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4717 db_printf("Dirty buffers:\n");
4718 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4719 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4724 DB_COMMAND(countfreebufs, db_coundfreebufs)
4727 int i, used = 0, nfree = 0;
4730 db_printf("usage: countfreebufs\n");
4734 for (i = 0; i < nbuf; i++) {
4736 if (bp->b_qindex == QUEUE_EMPTY)
4742 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4744 db_printf("numfreebuffers is %d\n", numfreebuffers);