xref: /dragonfly/sys/kern/vfs_bio.c (revision cb0f6a61)
1 /*
2  * Copyright (c) 1994,1997 John S. Dyson
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice immediately at the beginning of the file, without modification,
10  *    this list of conditions, and the following disclaimer.
11  * 2. Absolutely no warranty of function or purpose is made by the author
12  *		John S. Dyson.
13  *
14  * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15  */
16 
17 /*
18  * this file contains a new buffer I/O scheme implementing a coherent
19  * VM object and buffer cache scheme.  Pains have been taken to make
20  * sure that the performance degradation associated with schemes such
21  * as this is not realized.
22  *
23  * Author:  John S. Dyson
24  * Significant help during the development and debugging phases
25  * had been provided by David Greenman, also of the FreeBSD core team.
26  *
27  * see man buf(9) for more info.  Note that man buf(9) doesn't reflect
28  * the actual buf/bio implementation in DragonFly.
29  */
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/buf.h>
34 #include <sys/conf.h>
35 #include <sys/devicestat.h>
36 #include <sys/eventhandler.h>
37 #include <sys/lock.h>
38 #include <sys/malloc.h>
39 #include <sys/mount.h>
40 #include <sys/kernel.h>
41 #include <sys/kthread.h>
42 #include <sys/proc.h>
43 #include <sys/reboot.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sysctl.h>
46 #include <sys/vmmeter.h>
47 #include <sys/vnode.h>
48 #include <sys/dsched.h>
49 #include <vm/vm.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_pageout.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/swap_pager.h>
59 
60 #include <sys/buf2.h>
61 #include <sys/spinlock2.h>
62 #include <vm/vm_page2.h>
63 
64 #include "opt_ddb.h"
65 #ifdef DDB
66 #include <ddb/ddb.h>
67 #endif
68 
69 /*
70  * Buffer queues.
71  */
72 enum bufq_type {
73 	BQUEUE_NONE,    	/* not on any queue */
74 	BQUEUE_LOCKED,  	/* locked buffers */
75 	BQUEUE_CLEAN,   	/* non-B_DELWRI buffers */
76 	BQUEUE_DIRTY,   	/* B_DELWRI buffers */
77 	BQUEUE_DIRTY_HW,   	/* B_DELWRI buffers - heavy weight */
78 	BQUEUE_EMPTY,    	/* empty buffer headers */
79 
80 	BUFFER_QUEUES		/* number of buffer queues */
81 };
82 
83 typedef enum bufq_type bufq_type_t;
84 
85 #define BD_WAKE_SIZE	16384
86 #define BD_WAKE_MASK	(BD_WAKE_SIZE - 1)
87 
88 TAILQ_HEAD(bqueues, buf);
89 
90 struct bufpcpu {
91 	struct spinlock spin;
92 	struct bqueues bufqueues[BUFFER_QUEUES];
93 } __cachealign;
94 
95 struct bufpcpu bufpcpu[MAXCPU];
96 
97 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
98 
99 struct buf *buf;		/* buffer header pool */
100 
101 static void vfs_clean_pages(struct buf *bp);
102 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
103 #if 0
104 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
105 #endif
106 static void vfs_vmio_release(struct buf *bp);
107 static int flushbufqueues(struct buf *marker, bufq_type_t q);
108 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj,
109 				vm_pindex_t pg, int deficit);
110 
111 static void bd_signal(long totalspace);
112 static void buf_daemon(void);
113 static void buf_daemon_hw(void);
114 
115 /*
116  * bogus page -- for I/O to/from partially complete buffers
117  * this is a temporary solution to the problem, but it is not
118  * really that bad.  it would be better to split the buffer
119  * for input in the case of buffers partially already in memory,
120  * but the code is intricate enough already.
121  */
122 vm_page_t bogus_page;
123 
124 /*
125  * These are all static, but make the ones we export globals so we do
126  * not need to use compiler magic.
127  */
128 long bufspace;			/* atomic ops */
129 long maxbufspace;
130 long lobufspace, hibufspace;
131 static long lorunningspace;
132 static long hirunningspace;
133 static long dirtykvaspace;		/* atomic */
134 long dirtybufspace;			/* atomic (global for systat) */
135 static long dirtybufcount;		/* atomic */
136 static long dirtybufspacehw;		/* atomic */
137 static long dirtybufcounthw;		/* atomic */
138 static long runningbufspace;		/* atomic */
139 static long runningbufcount;		/* atomic */
140 long lodirtybufspace;
141 long hidirtybufspace;
142 static int getnewbufcalls;
143 static int needsbuffer;			/* atomic */
144 static int runningbufreq;		/* atomic */
145 static int bd_request;			/* atomic */
146 static int bd_request_hw;		/* atomic */
147 static u_int bd_wake_ary[BD_WAKE_SIZE];
148 static u_int bd_wake_index;
149 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
150 static int debug_commit;
151 static int debug_bufbio;
152 static int debug_kvabio;
153 static long bufcache_bw = 200 * 1024 * 1024;
154 
155 static struct thread *bufdaemon_td;
156 static struct thread *bufdaemonhw_td;
157 static u_int lowmempgallocs;
158 static u_int flushperqueue = 1024;
159 
160 /*
161  * Sysctls for operational control of the buffer cache.
162  */
163 SYSCTL_UINT(_vfs, OID_AUTO, flushperqueue, CTLFLAG_RW, &flushperqueue, 0,
164 	"Number of buffers to flush from each per-cpu queue");
165 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
166 	"Number of dirty buffers to flush before bufdaemon becomes inactive");
167 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
168 	"High watermark used to trigger explicit flushing of dirty buffers");
169 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
170 	"Minimum amount of buffer space required for active I/O");
171 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
172 	"Maximum amount of buffer space to usable for active I/O");
173 SYSCTL_LONG(_vfs, OID_AUTO, bufcache_bw, CTLFLAG_RW, &bufcache_bw, 0,
174 	"Buffer-cache -> VM page cache transfer bandwidth");
175 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
176 	"Page allocations done during periods of very low free memory");
177 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
178 	"Recycle pages to active or inactive queue transition pt 0-64");
179 /*
180  * Sysctls determining current state of the buffer cache.
181  */
182 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
183 	"Total number of buffers in buffer cache");
184 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0,
185 	"KVA reserved by dirty buffers (all)");
186 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
187 	"Pending bytes of dirty buffers (all)");
188 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
189 	"Pending bytes of dirty buffers (heavy weight)");
190 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
191 	"Pending number of dirty buffers");
192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
193 	"Pending number of dirty buffers (heavy weight)");
194 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
195 	"I/O bytes currently in progress due to asynchronous writes");
196 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
197 	"I/O buffers currently in progress due to asynchronous writes");
198 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
199 	"Hard limit on maximum amount of memory usable for buffer space");
200 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
201 	"Soft limit on maximum amount of memory usable for buffer space");
202 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
203 	"Minimum amount of memory to reserve for system buffer space");
204 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
205 	"Amount of memory available for buffers");
206 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
207 	"New buffer header acquisition requests");
208 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
209 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, "");
210 SYSCTL_INT(_vfs, OID_AUTO, debug_kvabio, CTLFLAG_RW, &debug_kvabio, 0, "");
211 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
212 	"sizeof(struct buf)");
213 
214 char *buf_wmesg = BUF_WMESG;
215 
216 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
217 #define VFS_BIO_NEED_UNUSED02	0x02
218 #define VFS_BIO_NEED_UNUSED04	0x04
219 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
220 
221 /*
222  * Called when buffer space is potentially available for recovery.
223  * getnewbuf() will block on this flag when it is unable to free
224  * sufficient buffer space.  Buffer space becomes recoverable when
225  * bp's get placed back in the queues.
226  */
227 static __inline void
bufspacewakeup(void)228 bufspacewakeup(void)
229 {
230 	/*
231 	 * If someone is waiting for BUF space, wake them up.  Even
232 	 * though we haven't freed the kva space yet, the waiting
233 	 * process will be able to now.
234 	 */
235 	for (;;) {
236 		int flags = needsbuffer;
237 		cpu_ccfence();
238 		if ((flags & VFS_BIO_NEED_BUFSPACE) == 0)
239 			break;
240 		if (atomic_cmpset_int(&needsbuffer, flags,
241 				      flags & ~VFS_BIO_NEED_BUFSPACE)) {
242 			wakeup(&needsbuffer);
243 			break;
244 		}
245 		/* retry */
246 	}
247 }
248 
249 /*
250  * runningbufwakeup:
251  *
252  *	Accounting for I/O in progress.
253  *
254  */
255 static __inline void
runningbufwakeup(struct buf * bp)256 runningbufwakeup(struct buf *bp)
257 {
258 	long totalspace;
259 	long flags;
260 
261 	if ((totalspace = bp->b_runningbufspace) != 0) {
262 		atomic_add_long(&runningbufspace, -totalspace);
263 		atomic_add_long(&runningbufcount, -1);
264 		bp->b_runningbufspace = 0;
265 
266 		/*
267 		 * see waitrunningbufspace() for limit test.
268 		 */
269 		for (;;) {
270 			flags = runningbufreq;
271 			cpu_ccfence();
272 			if (flags == 0)
273 				break;
274 			if (atomic_cmpset_int(&runningbufreq, flags, 0)) {
275 				wakeup(&runningbufreq);
276 				break;
277 			}
278 			/* retry */
279 		}
280 		bd_signal(totalspace);
281 	}
282 }
283 
284 /*
285  * bufcountwakeup:
286  *
287  *	Called when a buffer has been added to one of the free queues to
288  *	account for the buffer and to wakeup anyone waiting for free buffers.
289  *	This typically occurs when large amounts of metadata are being handled
290  *	by the buffer cache ( else buffer space runs out first, usually ).
291  */
292 static __inline void
bufcountwakeup(void)293 bufcountwakeup(void)
294 {
295 	long flags;
296 
297 	for (;;) {
298 		flags = needsbuffer;
299 		if (flags == 0)
300 			break;
301 		if (atomic_cmpset_int(&needsbuffer, flags,
302 				      (flags & ~VFS_BIO_NEED_ANY))) {
303 			wakeup(&needsbuffer);
304 			break;
305 		}
306 		/* retry */
307 	}
308 }
309 
310 /*
311  * waitrunningbufspace()
312  *
313  * If runningbufspace exceeds 4/6 hirunningspace we block until
314  * runningbufspace drops to 3/6 hirunningspace.  We also block if another
315  * thread blocked here in order to be fair, even if runningbufspace
316  * is now lower than the limit.
317  *
318  * The caller may be using this function to block in a tight loop, we
319  * must block while runningbufspace is greater than at least
320  * hirunningspace * 3 / 6.
321  */
322 void
waitrunningbufspace(void)323 waitrunningbufspace(void)
324 {
325 	long limit = hirunningspace * 4 / 6;
326 	long flags;
327 
328 	while (runningbufspace > limit || runningbufreq) {
329 		tsleep_interlock(&runningbufreq, 0);
330 		flags = atomic_fetchadd_int(&runningbufreq, 1);
331 		if (runningbufspace > limit || flags)
332 			tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz);
333 	}
334 }
335 
336 /*
337  * buf_dirty_count_severe:
338  *
339  *	Return true if we have too many dirty buffers.
340  */
341 int
buf_dirty_count_severe(void)342 buf_dirty_count_severe(void)
343 {
344 	return (runningbufspace + dirtykvaspace >= hidirtybufspace ||
345 	        dirtybufcount >= nbuf / 2);
346 }
347 
348 /*
349  * Return true if the amount of running I/O is severe and BIOQ should
350  * start bursting.
351  */
352 int
buf_runningbufspace_severe(void)353 buf_runningbufspace_severe(void)
354 {
355 	return (runningbufspace >= hirunningspace * 4 / 6);
356 }
357 
358 /*
359  * vfs_buf_test_cache:
360  *
361  * Called when a buffer is extended.  This function clears the B_CACHE
362  * bit if the newly extended portion of the buffer does not contain
363  * valid data.
364  *
365  * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
366  * cache buffers.  The VM pages remain dirty, as someone had mmap()'d
367  * them while a clean buffer was present.
368  */
369 static __inline__
370 void
vfs_buf_test_cache(struct buf * bp,vm_ooffset_t foff,vm_offset_t off,vm_offset_t size,vm_page_t m)371 vfs_buf_test_cache(struct buf *bp,
372 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
373 		  vm_page_t m)
374 {
375 	if (bp->b_flags & B_CACHE) {
376 		int base = (foff + off) & PAGE_MASK;
377 		if (vm_page_is_valid(m, base, size) == 0)
378 			bp->b_flags &= ~B_CACHE;
379 	}
380 }
381 
382 /*
383  * bd_speedup()
384  *
385  * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
386  * low water mark.
387  */
388 static __inline__
389 void
bd_speedup(void)390 bd_speedup(void)
391 {
392 	if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2)
393 		return;
394 
395 	if (bd_request == 0 &&
396 	    (dirtykvaspace > lodirtybufspace / 2 ||
397 	     dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
398 		if (atomic_fetchadd_int(&bd_request, 1) == 0)
399 			wakeup(&bd_request);
400 	}
401 	if (bd_request_hw == 0 &&
402 	    (dirtykvaspace > lodirtybufspace / 2 ||
403 	     dirtybufcounthw >= nbuf / 2)) {
404 		if (atomic_fetchadd_int(&bd_request_hw, 1) == 0)
405 			wakeup(&bd_request_hw);
406 	}
407 }
408 
409 /*
410  * bd_heatup()
411  *
412  *	Get the buf_daemon heated up when the number of running and dirty
413  *	buffers exceeds the mid-point.
414  *
415  *	Return the total number of dirty bytes past the second mid point
416  *	as a measure of how much excess dirty data there is in the system.
417  */
418 long
bd_heatup(void)419 bd_heatup(void)
420 {
421 	long mid1;
422 	long mid2;
423 	long totalspace;
424 
425 	mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
426 
427 	totalspace = runningbufspace + dirtykvaspace;
428 	if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
429 		bd_speedup();
430 		mid2 = mid1 + (hidirtybufspace - mid1) / 2;
431 		if (totalspace >= mid2)
432 			return(totalspace - mid2);
433 	}
434 	return(0);
435 }
436 
437 /*
438  * bd_wait()
439  *
440  *	Wait for the buffer cache to flush (totalspace) bytes worth of
441  *	buffers, then return.
442  *
443  *	Regardless this function blocks while the number of dirty buffers
444  *	exceeds hidirtybufspace.
445  */
446 void
bd_wait(long totalspace)447 bd_wait(long totalspace)
448 {
449 	u_int i;
450 	u_int j;
451 	u_int mi;
452 	int count;
453 
454 	if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
455 		return;
456 
457 	while (totalspace > 0) {
458 		bd_heatup();
459 
460 		/*
461 		 * Order is important.  Suppliers adjust bd_wake_index after
462 		 * updating runningbufspace/dirtykvaspace.  We want to fetch
463 		 * bd_wake_index before accessing.  Any error should thus
464 		 * be in our favor.
465 		 */
466 		i = atomic_fetchadd_int(&bd_wake_index, 0);
467 		if (totalspace > runningbufspace + dirtykvaspace)
468 			totalspace = runningbufspace + dirtykvaspace;
469 		count = totalspace / MAXBSIZE;
470 		if (count >= BD_WAKE_SIZE / 2)
471 			count = BD_WAKE_SIZE / 2;
472 		i = i + count;
473 		mi = i & BD_WAKE_MASK;
474 
475 		/*
476 		 * This is not a strict interlock, so we play a bit loose
477 		 * with locking access to dirtybufspace*.  We have to re-check
478 		 * bd_wake_index to ensure that it hasn't passed us.
479 		 */
480 		tsleep_interlock(&bd_wake_ary[mi], 0);
481 		atomic_add_int(&bd_wake_ary[mi], 1);
482 		j = atomic_fetchadd_int(&bd_wake_index, 0);
483 		if ((int)(i - j) >= 0)
484 			tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz);
485 
486 		totalspace = runningbufspace + dirtykvaspace - hidirtybufspace;
487 	}
488 }
489 
490 /*
491  * bd_signal()
492  *
493  *	This function is called whenever runningbufspace or dirtykvaspace
494  *	is reduced.  Track threads waiting for run+dirty buffer I/O
495  *	complete.
496  */
497 static void
bd_signal(long totalspace)498 bd_signal(long totalspace)
499 {
500 	u_int i;
501 
502 	if (totalspace > 0) {
503 		if (totalspace > MAXBSIZE * BD_WAKE_SIZE)
504 			totalspace = MAXBSIZE * BD_WAKE_SIZE;
505 		while (totalspace > 0) {
506 			i = atomic_fetchadd_int(&bd_wake_index, 1);
507 			i &= BD_WAKE_MASK;
508 			if (atomic_readandclear_int(&bd_wake_ary[i]))
509 				wakeup(&bd_wake_ary[i]);
510 			totalspace -= MAXBSIZE;
511 		}
512 	}
513 }
514 
515 /*
516  * BIO tracking support routines.
517  *
518  * Release a ref on a bio_track.  Wakeup requests are atomically released
519  * along with the last reference so bk_active will never wind up set to
520  * only 0x80000000.
521  */
522 static
523 void
bio_track_rel(struct bio_track * track)524 bio_track_rel(struct bio_track *track)
525 {
526 	int	active;
527 	int	desired;
528 
529 	/*
530 	 * Shortcut
531 	 */
532 	active = track->bk_active;
533 	if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
534 		return;
535 
536 	/*
537 	 * Full-on.  Note that the wait flag is only atomically released on
538 	 * the 1->0 count transition.
539 	 *
540 	 * We check for a negative count transition using bit 30 since bit 31
541 	 * has a different meaning.
542 	 */
543 	for (;;) {
544 		desired = (active & 0x7FFFFFFF) - 1;
545 		if (desired)
546 			desired |= active & 0x80000000;
547 		if (atomic_cmpset_int(&track->bk_active, active, desired)) {
548 			if (desired & 0x40000000)
549 				panic("bio_track_rel: bad count: %p", track);
550 			if (active & 0x80000000)
551 				wakeup(track);
552 			break;
553 		}
554 		active = track->bk_active;
555 	}
556 }
557 
558 /*
559  * Wait for the tracking count to reach 0.
560  *
561  * Use atomic ops such that the wait flag is only set atomically when
562  * bk_active is non-zero.
563  */
564 int
bio_track_wait(struct bio_track * track,int slp_flags,int slp_timo)565 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
566 {
567 	int	active;
568 	int	desired;
569 	int	error;
570 
571 	/*
572 	 * Shortcut
573 	 */
574 	if (track->bk_active == 0)
575 		return(0);
576 
577 	/*
578 	 * Full-on.  Note that the wait flag may only be atomically set if
579 	 * the active count is non-zero.
580 	 *
581 	 * NOTE: We cannot optimize active == desired since a wakeup could
582 	 *	 clear active prior to our tsleep_interlock().
583 	 */
584 	error = 0;
585 	while ((active = track->bk_active) != 0) {
586 		cpu_ccfence();
587 		desired = active | 0x80000000;
588 		tsleep_interlock(track, slp_flags);
589 		if (atomic_cmpset_int(&track->bk_active, active, desired)) {
590 			error = tsleep(track, slp_flags | PINTERLOCKED,
591 				       "trwait", slp_timo);
592 			if (error)
593 				break;
594 		}
595 	}
596 	return (error);
597 }
598 
599 /*
600  * bufinit:
601  *
602  *	Load time initialisation of the buffer cache, called from machine
603  *	dependant initialization code.
604  */
605 static
606 void
bufinit(void * dummy __unused)607 bufinit(void *dummy __unused)
608 {
609 	struct bufpcpu *pcpu;
610 	struct buf *bp;
611 	vm_offset_t bogus_offset;
612 	int i;
613 	int j;
614 	long n;
615 
616 	/* next, make a null set of free lists */
617 	for (i = 0; i < ncpus; ++i) {
618 		pcpu = &bufpcpu[i];
619 		spin_init(&pcpu->spin, "bufinit");
620 		for (j = 0; j < BUFFER_QUEUES; j++)
621 			TAILQ_INIT(&pcpu->bufqueues[j]);
622 	}
623 
624 	/*
625 	 * Finally, initialize each buffer header and stick on empty q.
626 	 * Each buffer gets its own KVA reservation.
627 	 */
628 	i = 0;
629 	pcpu = &bufpcpu[i];
630 
631 	for (n = 0; n < nbuf; n++) {
632 		bp = &buf[n];
633 		bzero(bp, sizeof *bp);
634 		bp->b_flags = B_INVAL;	/* we're just an empty header */
635 		bp->b_cmd = BUF_CMD_DONE;
636 		bp->b_qindex = BQUEUE_EMPTY;
637 		bp->b_qcpu = i;
638 		bp->b_kvabase = (void *)(vm_map_min(buffer_map) +
639 					 MAXBSIZE * n);
640 		bp->b_kvasize = MAXBSIZE;
641 		initbufbio(bp);
642 		xio_init(&bp->b_xio);
643 		buf_dep_init(bp);
644 		TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
645 				  bp, b_freelist);
646 
647 		i = (i + 1) % ncpus;
648 		pcpu = &bufpcpu[i];
649 	}
650 
651 	/*
652 	 * maxbufspace is the absolute maximum amount of buffer space we are
653 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
654 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
655 	 * used by most other processes.  The differential is required to
656 	 * ensure that buf_daemon is able to run when other processes might
657 	 * be blocked waiting for buffer space.
658 	 *
659 	 * Calculate hysteresis (lobufspace, hibufspace).  Don't make it
660 	 * too large or we might lockup a cpu for too long a period of
661 	 * time in our tight loop.
662 	 */
663 	maxbufspace = nbuf * NBUFCALCSIZE;
664 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
665 	lobufspace = hibufspace * 7 / 8;
666 	if (hibufspace - lobufspace > 64 * 1024 * 1024)
667 		lobufspace = hibufspace - 64 * 1024 * 1024;
668 	if (lobufspace > hibufspace - MAXBSIZE)
669 		lobufspace = hibufspace - MAXBSIZE;
670 
671 	lorunningspace = 512 * 1024;
672 	/* hirunningspace -- see below */
673 
674 	/*
675 	 * Reduce the chance of a deadlock occuring by limiting the number
676 	 * of delayed-write dirty buffers we allow to stack up.
677 	 *
678 	 * We don't want too much actually queued to the device at once
679 	 * (XXX this needs to be per-mount!), because the buffers will
680 	 * wind up locked for a very long period of time while the I/O
681 	 * drains.
682 	 */
683 	hidirtybufspace = hibufspace / 2;	/* dirty + running */
684 	hirunningspace = hibufspace / 16;	/* locked & queued to device */
685 	if (hirunningspace < 1024 * 1024)
686 		hirunningspace = 1024 * 1024;
687 
688 	dirtykvaspace = 0;
689 	dirtybufspace = 0;
690 	dirtybufspacehw = 0;
691 
692 	lodirtybufspace = hidirtybufspace / 2;
693 
694 	/*
695 	 * Maximum number of async ops initiated per buf_daemon loop.  This is
696 	 * somewhat of a hack at the moment, we really need to limit ourselves
697 	 * based on the number of bytes of I/O in-transit that were initiated
698 	 * from buf_daemon.
699 	 */
700 
701 	bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE,
702 					   VM_SUBSYS_BOGUS);
703 	vm_object_hold(kernel_object);
704 	bogus_page = vm_page_alloc(kernel_object,
705 				   (bogus_offset >> PAGE_SHIFT),
706 				   VM_ALLOC_NORMAL);
707 	vm_object_drop(kernel_object);
708 	vmstats.v_wire_count++;
709 
710 }
711 
712 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL);
713 
714 /*
715  * Initialize the embedded bio structures, typically used by
716  * deprecated code which tries to allocate its own struct bufs.
717  */
718 void
initbufbio(struct buf * bp)719 initbufbio(struct buf *bp)
720 {
721 	bp->b_bio1.bio_buf = bp;
722 	bp->b_bio1.bio_prev = NULL;
723 	bp->b_bio1.bio_offset = NOOFFSET;
724 	bp->b_bio1.bio_next = &bp->b_bio2;
725 	bp->b_bio1.bio_done = NULL;
726 	bp->b_bio1.bio_flags = 0;
727 
728 	bp->b_bio2.bio_buf = bp;
729 	bp->b_bio2.bio_prev = &bp->b_bio1;
730 	bp->b_bio2.bio_offset = NOOFFSET;
731 	bp->b_bio2.bio_next = NULL;
732 	bp->b_bio2.bio_done = NULL;
733 	bp->b_bio2.bio_flags = 0;
734 
735 	BUF_LOCKINIT(bp);
736 }
737 
738 /*
739  * Reinitialize the embedded bio structures as well as any additional
740  * translation cache layers.
741  */
742 void
reinitbufbio(struct buf * bp)743 reinitbufbio(struct buf *bp)
744 {
745 	struct bio *bio;
746 
747 	for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
748 		bio->bio_done = NULL;
749 		bio->bio_offset = NOOFFSET;
750 	}
751 }
752 
753 /*
754  * Undo the effects of an initbufbio().
755  */
756 void
uninitbufbio(struct buf * bp)757 uninitbufbio(struct buf *bp)
758 {
759 	dsched_buf_exit(bp);
760 	BUF_LOCKFREE(bp);
761 }
762 
763 /*
764  * Push another BIO layer onto an existing BIO and return it.  The new
765  * BIO layer may already exist, holding cached translation data.
766  */
767 struct bio *
push_bio(struct bio * bio)768 push_bio(struct bio *bio)
769 {
770 	struct bio *nbio;
771 
772 	if ((nbio = bio->bio_next) == NULL) {
773 		int index = bio - &bio->bio_buf->b_bio_array[0];
774 		if (index >= NBUF_BIO - 1) {
775 			panic("push_bio: too many layers %d for bp %p",
776 				index, bio->bio_buf);
777 		}
778 		nbio = &bio->bio_buf->b_bio_array[index + 1];
779 		bio->bio_next = nbio;
780 		nbio->bio_prev = bio;
781 		nbio->bio_buf = bio->bio_buf;
782 		nbio->bio_offset = NOOFFSET;
783 		nbio->bio_done = NULL;
784 		nbio->bio_next = NULL;
785 	}
786 	KKASSERT(nbio->bio_done == NULL);
787 	return(nbio);
788 }
789 
790 /*
791  * Pop a BIO translation layer, returning the previous layer.  The
792  * must have been previously pushed.
793  */
794 struct bio *
pop_bio(struct bio * bio)795 pop_bio(struct bio *bio)
796 {
797 	return(bio->bio_prev);
798 }
799 
800 void
clearbiocache(struct bio * bio)801 clearbiocache(struct bio *bio)
802 {
803 	while (bio) {
804 		bio->bio_offset = NOOFFSET;
805 		bio = bio->bio_next;
806 	}
807 }
808 
809 /*
810  * Remove the buffer from the appropriate free list.
811  * (caller must be locked)
812  */
813 static __inline void
_bremfree(struct buf * bp)814 _bremfree(struct buf *bp)
815 {
816 	struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
817 
818 	if (bp->b_qindex != BQUEUE_NONE) {
819 		KASSERT(BUF_LOCKINUSE(bp), ("bremfree: bp %p not locked", bp));
820 		TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist);
821 		bp->b_qindex = BQUEUE_NONE;
822 	} else {
823 		if (!BUF_LOCKINUSE(bp))
824 			panic("bremfree: removing a buffer not on a queue");
825 	}
826 }
827 
828 /*
829  * bremfree() - must be called with a locked buffer
830  */
831 void
bremfree(struct buf * bp)832 bremfree(struct buf *bp)
833 {
834 	struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
835 
836 	spin_lock(&pcpu->spin);
837 	_bremfree(bp);
838 	spin_unlock(&pcpu->spin);
839 }
840 
841 /*
842  * bremfree_locked - must be called with pcpu->spin locked
843  */
844 static void
bremfree_locked(struct buf * bp)845 bremfree_locked(struct buf *bp)
846 {
847 	_bremfree(bp);
848 }
849 
850 /*
851  * This version of bread issues any required I/O asyncnronously and
852  * makes a callback on completion.
853  *
854  * The callback must check whether BIO_DONE is set in the bio and issue
855  * the bpdone(bp, 0) if it isn't.  The callback is responsible for clearing
856  * BIO_DONE and disposing of the I/O (bqrelse()ing it).
857  */
858 void
breadcb(struct vnode * vp,off_t loffset,int size,int bflags,void (* func)(struct bio *),void * arg)859 breadcb(struct vnode *vp, off_t loffset, int size, int bflags,
860 	void (*func)(struct bio *), void *arg)
861 {
862 	struct buf *bp;
863 
864 	bp = getblk(vp, loffset, size, 0, 0);
865 
866 	/* if not found in cache, do some I/O */
867 	if ((bp->b_flags & B_CACHE) == 0) {
868 		bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA);
869 		bp->b_flags |= bflags;
870 		bp->b_cmd = BUF_CMD_READ;
871 		bp->b_bio1.bio_done = func;
872 		bp->b_bio1.bio_caller_info1.ptr = arg;
873 		vfs_busy_pages(vp, bp);
874 		BUF_KERNPROC(bp);
875 		vn_strategy(vp, &bp->b_bio1);
876 	} else if (func) {
877 		/*
878 		 * Since we are issuing the callback synchronously it cannot
879 		 * race the BIO_DONE, so no need for atomic ops here.
880 		 */
881 		/*bp->b_bio1.bio_done = func;*/
882 		bp->b_bio1.bio_caller_info1.ptr = arg;
883 		bp->b_bio1.bio_flags |= BIO_DONE;
884 		func(&bp->b_bio1);
885 	} else {
886 		bqrelse(bp);
887 	}
888 }
889 
890 /*
891  * breadnx() - Terminal function for bread() and breadn().
892  *
893  * This function will start asynchronous I/O on read-ahead blocks as well
894  * as satisfy the primary request.
895  *
896  * We must clear B_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE is
897  * set, the buffer is valid and we do not have to do anything.
898  */
899 int
breadnx(struct vnode * vp,off_t loffset,int size,int bflags,off_t * raoffset,int * rabsize,int cnt,struct buf ** bpp)900 breadnx(struct vnode *vp, off_t loffset, int size, int bflags,
901 	off_t *raoffset, int *rabsize,
902 	int cnt, struct buf **bpp)
903 {
904 	struct buf *bp, *rabp;
905 	int i;
906 	int rv = 0, readwait = 0;
907 	int blkflags = (bflags & B_KVABIO) ? GETBLK_KVABIO : 0;
908 
909 	if (*bpp)
910 		bp = *bpp;
911 	else
912 		*bpp = bp = getblk(vp, loffset, size, blkflags, 0);
913 
914 	/* if not found in cache, do some I/O */
915 	if ((bp->b_flags & B_CACHE) == 0) {
916 		bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA);
917 		bp->b_flags |= bflags;
918 		bp->b_cmd = BUF_CMD_READ;
919 		bp->b_bio1.bio_done = biodone_sync;
920 		bp->b_bio1.bio_flags |= BIO_SYNC;
921 		vfs_busy_pages(vp, bp);
922 		vn_strategy(vp, &bp->b_bio1);
923 		++readwait;
924 	}
925 
926 	for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
927 		if (inmem(vp, *raoffset))
928 			continue;
929 		rabp = getblk(vp, *raoffset, *rabsize, GETBLK_KVABIO, 0);
930 
931 		if ((rabp->b_flags & B_CACHE) == 0) {
932 			rabp->b_flags &= ~(B_ERROR | B_EINTR |
933 					   B_INVAL | B_NOTMETA);
934 			rabp->b_flags |= (bflags & ~B_KVABIO);
935 			rabp->b_cmd = BUF_CMD_READ;
936 			vfs_busy_pages(vp, rabp);
937 			BUF_KERNPROC(rabp);
938 			vn_strategy(vp, &rabp->b_bio1);
939 		} else {
940 			brelse(rabp);
941 		}
942 	}
943 	if (readwait)
944 		rv = biowait(&bp->b_bio1, "biord");
945 	return (rv);
946 }
947 
948 /*
949  * bwrite:
950  *
951  *	Synchronous write, waits for completion.
952  *
953  *	Write, release buffer on completion.  (Done by iodone
954  *	if async).  Do not bother writing anything if the buffer
955  *	is invalid.
956  *
957  *	Note that we set B_CACHE here, indicating that buffer is
958  *	fully valid and thus cacheable.  This is true even of NFS
959  *	now so we set it generally.  This could be set either here
960  *	or in biodone() since the I/O is synchronous.  We put it
961  *	here.
962  */
963 int
bwrite(struct buf * bp)964 bwrite(struct buf *bp)
965 {
966 	int error;
967 
968 	if (bp->b_flags & B_INVAL) {
969 		brelse(bp);
970 		return (0);
971 	}
972 	if (BUF_LOCKINUSE(bp) == 0)
973 		panic("bwrite: buffer is not busy???");
974 
975 	/*
976 	 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
977 	 *	 call because it will remove the buffer from the vnode's
978 	 *	 dirty buffer list prematurely and possibly cause filesystem
979 	 *	 checks to race buffer flushes.  This is now handled in
980 	 *	 bpdone().
981 	 *
982 	 *	 bundirty(bp); REMOVED
983 	 */
984 
985 	bp->b_flags &= ~(B_ERROR | B_EINTR);
986 	bp->b_flags |= B_CACHE;
987 	bp->b_cmd = BUF_CMD_WRITE;
988 	bp->b_error = 0;
989 	bp->b_bio1.bio_done = biodone_sync;
990 	bp->b_bio1.bio_flags |= BIO_SYNC;
991 	vfs_busy_pages(bp->b_vp, bp);
992 
993 	/*
994 	 * Normal bwrites pipeline writes.  NOTE: b_bufsize is only
995 	 * valid for vnode-backed buffers.
996 	 */
997 	bsetrunningbufspace(bp, bp->b_bufsize);
998 	vn_strategy(bp->b_vp, &bp->b_bio1);
999 	error = biowait(&bp->b_bio1, "biows");
1000 	brelse(bp);
1001 
1002 	return (error);
1003 }
1004 
1005 /*
1006  * bawrite:
1007  *
1008  *	Asynchronous write.  Start output on a buffer, but do not wait for
1009  *	it to complete.  The buffer is released when the output completes.
1010  *
1011  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1012  *	B_INVAL buffers.  Not us.
1013  */
1014 void
bawrite(struct buf * bp)1015 bawrite(struct buf *bp)
1016 {
1017 	if (bp->b_flags & B_INVAL) {
1018 		brelse(bp);
1019 		return;
1020 	}
1021 	if (BUF_LOCKINUSE(bp) == 0)
1022 		panic("bawrite: buffer is not busy???");
1023 
1024 	/*
1025 	 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
1026 	 *	 call because it will remove the buffer from the vnode's
1027 	 *	 dirty buffer list prematurely and possibly cause filesystem
1028 	 *	 checks to race buffer flushes.  This is now handled in
1029 	 *	 bpdone().
1030 	 *
1031 	 *	 bundirty(bp); REMOVED
1032 	 */
1033 	bp->b_flags &= ~(B_ERROR | B_EINTR);
1034 	bp->b_flags |= B_CACHE;
1035 	bp->b_cmd = BUF_CMD_WRITE;
1036 	bp->b_error = 0;
1037 	KKASSERT(bp->b_bio1.bio_done == NULL);
1038 	vfs_busy_pages(bp->b_vp, bp);
1039 
1040 	/*
1041 	 * Normal bwrites pipeline writes.  NOTE: b_bufsize is only
1042 	 * valid for vnode-backed buffers.
1043 	 */
1044 	bsetrunningbufspace(bp, bp->b_bufsize);
1045 	BUF_KERNPROC(bp);
1046 	vn_strategy(bp->b_vp, &bp->b_bio1);
1047 }
1048 
1049 /*
1050  * bdwrite:
1051  *
1052  *	Delayed write. (Buffer is marked dirty).  Do not bother writing
1053  *	anything if the buffer is marked invalid.
1054  *
1055  *	Note that since the buffer must be completely valid, we can safely
1056  *	set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1057  *	biodone() in order to prevent getblk from writing the buffer
1058  *	out synchronously.
1059  */
1060 void
bdwrite(struct buf * bp)1061 bdwrite(struct buf *bp)
1062 {
1063 	if (BUF_LOCKINUSE(bp) == 0)
1064 		panic("bdwrite: buffer is not busy");
1065 
1066 	if (bp->b_flags & B_INVAL) {
1067 		brelse(bp);
1068 		return;
1069 	}
1070 	bdirty(bp);
1071 
1072 	dsched_buf_enter(bp);	/* might stack */
1073 
1074 	/*
1075 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1076 	 * true even of NFS now.
1077 	 */
1078 	bp->b_flags |= B_CACHE;
1079 
1080 	/*
1081 	 * This bmap keeps the system from needing to do the bmap later,
1082 	 * perhaps when the system is attempting to do a sync.  Since it
1083 	 * is likely that the indirect block -- or whatever other datastructure
1084 	 * that the filesystem needs is still in memory now, it is a good
1085 	 * thing to do this.  Note also, that if the pageout daemon is
1086 	 * requesting a sync -- there might not be enough memory to do
1087 	 * the bmap then...  So, this is important to do.
1088 	 */
1089 	if (bp->b_bio2.bio_offset == NOOFFSET) {
1090 		VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1091 			 NULL, NULL, BUF_CMD_WRITE);
1092 	}
1093 
1094 	/*
1095 	 * Because the underlying pages may still be mapped and
1096 	 * writable trying to set the dirty buffer (b_dirtyoff/end)
1097 	 * range here will be inaccurate.
1098 	 *
1099 	 * However, we must still clean the pages to satisfy the
1100 	 * vnode_pager and pageout daemon, so they think the pages
1101 	 * have been "cleaned".  What has really occured is that
1102 	 * they've been earmarked for later writing by the buffer
1103 	 * cache.
1104 	 *
1105 	 * So we get the b_dirtyoff/end update but will not actually
1106 	 * depend on it (NFS that is) until the pages are busied for
1107 	 * writing later on.
1108 	 */
1109 	vfs_clean_pages(bp);
1110 	bqrelse(bp);
1111 
1112 	/*
1113 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1114 	 * due to the softdep code.
1115 	 */
1116 }
1117 
1118 /*
1119  * Fake write - return pages to VM system as dirty, leave the buffer clean.
1120  * This is used by tmpfs.
1121  *
1122  * It is important for any VFS using this routine to NOT use it for
1123  * IO_SYNC or IO_ASYNC operations which occur when the system really
1124  * wants to flush VM pages to backing store.
1125  */
1126 void
buwrite(struct buf * bp)1127 buwrite(struct buf *bp)
1128 {
1129 	vm_page_t m;
1130 	int i;
1131 
1132 	/*
1133 	 * Only works for VMIO buffers.  If the buffer is already
1134 	 * marked for delayed-write we can't avoid the bdwrite().
1135 	 */
1136 	if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1137 		bdwrite(bp);
1138 		return;
1139 	}
1140 
1141 	/*
1142 	 * Mark as needing a commit.
1143 	 */
1144 	for (i = 0; i < bp->b_xio.xio_npages; i++) {
1145 		m = bp->b_xio.xio_pages[i];
1146 		vm_page_need_commit(m);
1147 	}
1148 	bqrelse(bp);
1149 }
1150 
1151 /*
1152  * bdirty:
1153  *
1154  *	Turn buffer into delayed write request by marking it B_DELWRI.
1155  *	B_RELBUF and B_NOCACHE must be cleared.
1156  *
1157  *	We reassign the buffer to itself to properly update it in the
1158  *	dirty/clean lists.
1159  *
1160  *	Must be called from a critical section.
1161  *	The buffer must be on BQUEUE_NONE.
1162  */
1163 void
bdirty(struct buf * bp)1164 bdirty(struct buf *bp)
1165 {
1166 	KASSERT(bp->b_qindex == BQUEUE_NONE,
1167 		("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1168 	if (bp->b_flags & B_NOCACHE) {
1169 		kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1170 		bp->b_flags &= ~B_NOCACHE;
1171 	}
1172 	if (bp->b_flags & B_INVAL) {
1173 		kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1174 	}
1175 	bp->b_flags &= ~B_RELBUF;
1176 
1177 	if ((bp->b_flags & B_DELWRI) == 0) {
1178 		lwkt_gettoken(&bp->b_vp->v_token);
1179 		bp->b_flags |= B_DELWRI;
1180 		reassignbuf(bp);
1181 		lwkt_reltoken(&bp->b_vp->v_token);
1182 
1183 		atomic_add_long(&dirtybufcount, 1);
1184 		atomic_add_long(&dirtykvaspace, bp->b_kvasize);
1185 		atomic_add_long(&dirtybufspace, bp->b_bufsize);
1186 		if (bp->b_flags & B_HEAVY) {
1187 			atomic_add_long(&dirtybufcounthw, 1);
1188 			atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1189 		}
1190 		bd_heatup();
1191 	}
1192 }
1193 
1194 /*
1195  * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1196  * needs to be flushed with a different buf_daemon thread to avoid
1197  * deadlocks.  B_HEAVY also imposes restrictions in getnewbuf().
1198  */
1199 void
bheavy(struct buf * bp)1200 bheavy(struct buf *bp)
1201 {
1202 	if ((bp->b_flags & B_HEAVY) == 0) {
1203 		bp->b_flags |= B_HEAVY;
1204 		if (bp->b_flags & B_DELWRI) {
1205 			atomic_add_long(&dirtybufcounthw, 1);
1206 			atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1207 		}
1208 	}
1209 }
1210 
1211 /*
1212  * bundirty:
1213  *
1214  *	Clear B_DELWRI for buffer.
1215  *
1216  *	Must be called from a critical section.
1217  *
1218  *	The buffer is typically on BQUEUE_NONE but there is one case in
1219  *	brelse() that calls this function after placing the buffer on
1220  *	a different queue.
1221  */
1222 void
bundirty(struct buf * bp)1223 bundirty(struct buf *bp)
1224 {
1225 	if (bp->b_flags & B_DELWRI) {
1226 		lwkt_gettoken(&bp->b_vp->v_token);
1227 		bp->b_flags &= ~B_DELWRI;
1228 		reassignbuf(bp);
1229 		lwkt_reltoken(&bp->b_vp->v_token);
1230 
1231 		atomic_add_long(&dirtybufcount, -1);
1232 		atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1233 		atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1234 		if (bp->b_flags & B_HEAVY) {
1235 			atomic_add_long(&dirtybufcounthw, -1);
1236 			atomic_add_long(&dirtybufspacehw, -bp->b_bufsize);
1237 		}
1238 		bd_signal(bp->b_bufsize);
1239 	}
1240 	/*
1241 	 * Since it is now being written, we can clear its deferred write flag.
1242 	 */
1243 	bp->b_flags &= ~B_DEFERRED;
1244 }
1245 
1246 /*
1247  * Set the b_runningbufspace field, used to track how much I/O is
1248  * in progress at any given moment.
1249  */
1250 void
bsetrunningbufspace(struct buf * bp,int bytes)1251 bsetrunningbufspace(struct buf *bp, int bytes)
1252 {
1253 	bp->b_runningbufspace = bytes;
1254 	if (bytes) {
1255 		atomic_add_long(&runningbufspace, bytes);
1256 		atomic_add_long(&runningbufcount, 1);
1257 	}
1258 }
1259 
1260 /*
1261  * brelse:
1262  *
1263  *	Release a busy buffer and, if requested, free its resources.  The
1264  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1265  *	to be accessed later as a cache entity or reused for other purposes.
1266  */
1267 void
brelse(struct buf * bp)1268 brelse(struct buf *bp)
1269 {
1270 	struct bufpcpu *pcpu;
1271 #ifdef INVARIANTS
1272 	int saved_flags = bp->b_flags;
1273 #endif
1274 
1275 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1276 		("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1277 
1278 	/*
1279 	 * If B_NOCACHE is set we are being asked to destroy the buffer and
1280 	 * its backing store.  Clear B_DELWRI.
1281 	 *
1282 	 * B_NOCACHE is set in two cases: (1) when the caller really wants
1283 	 * to destroy the buffer and backing store and (2) when the caller
1284 	 * wants to destroy the buffer and backing store after a write
1285 	 * completes.
1286 	 */
1287 	if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1288 		bundirty(bp);
1289 	}
1290 
1291 	if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1292 		/*
1293 		 * A re-dirtied buffer is only subject to destruction
1294 		 * by B_INVAL.  B_ERROR and B_NOCACHE are ignored.
1295 		 */
1296 		/* leave buffer intact */
1297 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1298 		   (bp->b_bufsize <= 0)) {
1299 		/*
1300 		 * Either a failed read or we were asked to free or not
1301 		 * cache the buffer.  This path is reached with B_DELWRI
1302 		 * set only if B_INVAL is already set.  B_NOCACHE governs
1303 		 * backing store destruction.
1304 		 *
1305 		 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1306 		 * buffer cannot be immediately freed.
1307 		 */
1308 		bp->b_flags |= B_INVAL;
1309 		if (LIST_FIRST(&bp->b_dep) != NULL)
1310 			buf_deallocate(bp);
1311 		if (bp->b_flags & B_DELWRI) {
1312 			atomic_add_long(&dirtybufcount, -1);
1313 			atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1314 			atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1315 			if (bp->b_flags & B_HEAVY) {
1316 				atomic_add_long(&dirtybufcounthw, -1);
1317 				atomic_add_long(&dirtybufspacehw,
1318 						-bp->b_bufsize);
1319 			}
1320 			bd_signal(bp->b_bufsize);
1321 		}
1322 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1323 	}
1324 
1325 	/*
1326 	 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1327 	 * or if b_refs is non-zero.
1328 	 *
1329 	 * If vfs_vmio_release() is called with either bit set, the
1330 	 * underlying pages may wind up getting freed causing a previous
1331 	 * write (bdwrite()) to get 'lost' because pages associated with
1332 	 * a B_DELWRI bp are marked clean.  Pages associated with a
1333 	 * B_LOCKED buffer may be mapped by the filesystem.
1334 	 *
1335 	 * If we want to release the buffer ourselves (rather then the
1336 	 * originator asking us to release it), give the originator a
1337 	 * chance to countermand the release by setting B_LOCKED.
1338 	 *
1339 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1340 	 * if B_DELWRI is set.
1341 	 *
1342 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1343 	 * on pages to return pages to the VM page queues.
1344 	 */
1345 	if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1346 		bp->b_flags &= ~B_RELBUF;
1347 	} else if (vm_paging_min()) {
1348 		if (LIST_FIRST(&bp->b_dep) != NULL)
1349 			buf_deallocate(bp);		/* can set B_LOCKED */
1350 		if (bp->b_flags & (B_DELWRI | B_LOCKED))
1351 			bp->b_flags &= ~B_RELBUF;
1352 		else
1353 			bp->b_flags |= B_RELBUF;
1354 	}
1355 
1356 	/*
1357 	 * Make sure b_cmd is clear.  It may have already been cleared by
1358 	 * biodone().
1359 	 *
1360 	 * At this point destroying the buffer is governed by the B_INVAL
1361 	 * or B_RELBUF flags.
1362 	 */
1363 	bp->b_cmd = BUF_CMD_DONE;
1364 	dsched_buf_exit(bp);
1365 
1366 	/*
1367 	 * VMIO buffer rundown.  Make sure the VM page array is restored
1368 	 * after an I/O may have replaces some of the pages with bogus pages
1369 	 * in order to not destroy dirty pages in a fill-in read.
1370 	 *
1371 	 * Note that due to the code above, if a buffer is marked B_DELWRI
1372 	 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1373 	 * B_INVAL may still be set, however.
1374 	 *
1375 	 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1376 	 * but not the backing store.   B_NOCACHE will destroy the backing
1377 	 * store.
1378 	 *
1379 	 * Note that dirty NFS buffers contain byte-granular write ranges
1380 	 * and should not be destroyed w/ B_INVAL even if the backing store
1381 	 * is left intact.
1382 	 */
1383 	if (bp->b_flags & B_VMIO) {
1384 		/*
1385 		 * Rundown for VMIO buffers which are not dirty NFS buffers.
1386 		 */
1387 		int i, j, resid;
1388 		vm_page_t m;
1389 		off_t foff;
1390 		vm_pindex_t poff;
1391 		vm_object_t obj;
1392 		struct vnode *vp;
1393 
1394 		vp = bp->b_vp;
1395 
1396 		/*
1397 		 * Get the base offset and length of the buffer.  Note that
1398 		 * in the VMIO case if the buffer block size is not
1399 		 * page-aligned then b_data pointer may not be page-aligned.
1400 		 * But our b_xio.xio_pages array *IS* page aligned.
1401 		 *
1402 		 * block sizes less then DEV_BSIZE (usually 512) are not
1403 		 * supported due to the page granularity bits (m->valid,
1404 		 * m->dirty, etc...).
1405 		 *
1406 		 * See man buf(9) for more information
1407 		 */
1408 
1409 		resid = bp->b_bufsize;
1410 		foff = bp->b_loffset;
1411 
1412 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
1413 			m = bp->b_xio.xio_pages[i];
1414 
1415 			/*
1416 			 * If we hit a bogus page, fixup *all* of them
1417 			 * now.  Note that we left these pages wired
1418 			 * when we removed them so they had better exist,
1419 			 * and they cannot be ripped out from under us so
1420 			 * no critical section protection is necessary.
1421 			 */
1422 			if (m == bogus_page) {
1423 				obj = vp->v_object;
1424 				poff = OFF_TO_IDX(bp->b_loffset);
1425 
1426 				vm_object_hold(obj);
1427 				for (j = i; j < bp->b_xio.xio_npages; j++) {
1428 					vm_page_t mtmp;
1429 
1430 					mtmp = bp->b_xio.xio_pages[j];
1431 					if (mtmp == bogus_page) {
1432 						if ((bp->b_flags & B_HASBOGUS) == 0)
1433 							panic("brelse: bp %p corrupt bogus", bp);
1434 						mtmp = vm_page_lookup(obj, poff + j);
1435 						if (!mtmp)
1436 							panic("brelse: bp %p page %d missing", bp, j);
1437 						bp->b_xio.xio_pages[j] = mtmp;
1438 					}
1439 				}
1440 				vm_object_drop(obj);
1441 
1442 				if ((bp->b_flags & B_HASBOGUS) ||
1443 				    (bp->b_flags & B_INVAL) == 0) {
1444 					pmap_qenter_noinval(
1445 					    trunc_page((vm_offset_t)bp->b_data),
1446 					    bp->b_xio.xio_pages,
1447 					    bp->b_xio.xio_npages);
1448 					bp->b_flags &= ~B_HASBOGUS;
1449 					bp->b_flags |= B_KVABIO;
1450 					bkvareset(bp);
1451 				}
1452 				m = bp->b_xio.xio_pages[i];
1453 			}
1454 
1455 			/*
1456 			 * Invalidate the backing store if B_NOCACHE is set
1457 			 * (e.g. used with vinvalbuf()).  If this is NFS
1458 			 * we impose a requirement that the block size be
1459 			 * a multiple of PAGE_SIZE and create a temporary
1460 			 * hack to basically invalidate the whole page.  The
1461 			 * problem is that NFS uses really odd buffer sizes
1462 			 * especially when tracking piecemeal writes and
1463 			 * it also vinvalbuf()'s a lot, which would result
1464 			 * in only partial page validation and invalidation
1465 			 * here.  If the file page is mmap()'d, however,
1466 			 * all the valid bits get set so after we invalidate
1467 			 * here we would end up with weird m->valid values
1468 			 * like 0xfc.  nfs_getpages() can't handle this so
1469 			 * we clear all the valid bits for the NFS case
1470 			 * instead of just some of them.
1471 			 *
1472 			 * The real bug is the VM system having to set m->valid
1473 			 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1474 			 * itself is an artifact of the whole 512-byte
1475 			 * granular mess that exists to support odd block
1476 			 * sizes and UFS meta-data block sizes (e.g. 6144).
1477 			 * A complete rewrite is required.
1478 			 *
1479 			 * XXX
1480 			 */
1481 			if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1482 				int poffset = foff & PAGE_MASK;
1483 				int presid;
1484 
1485 				presid = PAGE_SIZE - poffset;
1486 				if (bp->b_vp->v_tag == VT_NFS &&
1487 				    bp->b_vp->v_type == VREG) {
1488 					; /* entire page */
1489 				} else if (presid > resid) {
1490 					presid = resid;
1491 				}
1492 				KASSERT(presid >= 0, ("brelse: extra page"));
1493 				vm_page_set_invalid(m, poffset, presid);
1494 
1495 				/*
1496 				 * Also make sure any swap cache is removed
1497 				 * as it is now stale (HAMMER in particular
1498 				 * uses B_NOCACHE to deal with buffer
1499 				 * aliasing).
1500 				 */
1501 				swap_pager_unswapped(m);
1502 			}
1503 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1504 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1505 		}
1506 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1507 			vfs_vmio_release(bp);
1508 	} else {
1509 		/*
1510 		 * Rundown for non-VMIO buffers.
1511 		 *
1512 		 * XXX With B_MALLOC buffers removed, there should no longer
1513 		 * be any situation where brelse() is called on a non B_VMIO
1514 		 * buffer.  Recommend assertion here.  XXX
1515 		 */
1516 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1517 			if (bp->b_bufsize)
1518 				allocbuf(bp, 0);
1519 			KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1520 			if (bp->b_vp)
1521 				brelvp(bp);
1522 		}
1523 	}
1524 
1525 	if (bp->b_qindex != BQUEUE_NONE)
1526 		panic("brelse: free buffer onto another queue???");
1527 
1528 	/*
1529 	 * Figure out the correct queue to place the cleaned up buffer on.
1530 	 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1531 	 * disassociated from their vnode.
1532 	 *
1533 	 * Return the buffer to its original pcpu area
1534 	 */
1535 	pcpu = &bufpcpu[bp->b_qcpu];
1536 	spin_lock(&pcpu->spin);
1537 
1538 	if (bp->b_flags & B_LOCKED) {
1539 		/*
1540 		 * Buffers that are locked are placed in the locked queue
1541 		 * immediately, regardless of their state.
1542 		 */
1543 		bp->b_qindex = BQUEUE_LOCKED;
1544 		TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1545 				  bp, b_freelist);
1546 	} else if (bp->b_bufsize == 0) {
1547 		/*
1548 		 * Buffers with no memory.  Due to conditionals near the top
1549 		 * of brelse() such buffers should probably already be
1550 		 * marked B_INVAL and disassociated from their vnode.
1551 		 */
1552 		bp->b_flags |= B_INVAL;
1553 		KASSERT(bp->b_vp == NULL,
1554 			("bp1 %p flags %08x/%08x vnode %p "
1555 			 "unexpectededly still associated!",
1556 			bp, saved_flags, bp->b_flags, bp->b_vp));
1557 		KKASSERT((bp->b_flags & B_HASHED) == 0);
1558 		bp->b_qindex = BQUEUE_EMPTY;
1559 		TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1560 				  bp, b_freelist);
1561 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1562 		/*
1563 		 * Buffers with junk contents.   Again these buffers had better
1564 		 * already be disassociated from their vnode.
1565 		 */
1566 		KASSERT(bp->b_vp == NULL,
1567 			("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1568 			 "still associated!",
1569 			bp, saved_flags, bp->b_flags, bp->b_vp));
1570 		KKASSERT((bp->b_flags & B_HASHED) == 0);
1571 		bp->b_flags |= B_INVAL;
1572 		bp->b_qindex = BQUEUE_CLEAN;
1573 		TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1574 				  bp, b_freelist);
1575 	} else {
1576 		/*
1577 		 * Remaining buffers.  These buffers are still associated with
1578 		 * their vnode.
1579 		 */
1580 		switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1581 		case B_DELWRI:
1582 			bp->b_qindex = BQUEUE_DIRTY;
1583 			TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1584 					  bp, b_freelist);
1585 			break;
1586 		case B_DELWRI | B_HEAVY:
1587 			bp->b_qindex = BQUEUE_DIRTY_HW;
1588 			TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1589 					  bp, b_freelist);
1590 			break;
1591 		default:
1592 			/*
1593 			 * NOTE: Buffers are always placed at the end of the
1594 			 * queue.  If B_AGE is not set the buffer will cycle
1595 			 * through the queue twice.
1596 			 */
1597 			bp->b_qindex = BQUEUE_CLEAN;
1598 			TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1599 					  bp, b_freelist);
1600 			break;
1601 		}
1602 	}
1603 	spin_unlock(&pcpu->spin);
1604 
1605 	/*
1606 	 * If B_INVAL, clear B_DELWRI.  We've already placed the buffer
1607 	 * on the correct queue but we have not yet unlocked it.
1608 	 */
1609 	if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1610 		bundirty(bp);
1611 
1612 	/*
1613 	 * The bp is on an appropriate queue unless locked.  If it is not
1614 	 * locked or dirty we can wakeup threads waiting for buffer space.
1615 	 *
1616 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1617 	 * if B_INVAL is set ).
1618 	 */
1619 	if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1620 		bufcountwakeup();
1621 
1622 	/*
1623 	 * Something we can maybe free or reuse
1624 	 */
1625 	if (bp->b_bufsize || bp->b_kvasize)
1626 		bufspacewakeup();
1627 
1628 	/*
1629 	 * Clean up temporary flags and unlock the buffer.
1630 	 */
1631 	bp->b_flags &= ~(B_NOCACHE | B_RELBUF | B_DIRECT);
1632 	BUF_UNLOCK(bp);
1633 }
1634 
1635 /*
1636  * bqrelse:
1637  *
1638  *	Release a buffer back to the appropriate queue but do not try to free
1639  *	it.  The buffer is expected to be used again soon.
1640  *
1641  *	bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1642  *	biodone() to requeue an async I/O on completion.  It is also used when
1643  *	known good buffers need to be requeued but we think we may need the data
1644  *	again soon.
1645  *
1646  *	XXX we should be able to leave the B_RELBUF hint set on completion.
1647  */
1648 void
bqrelse(struct buf * bp)1649 bqrelse(struct buf *bp)
1650 {
1651 	struct bufpcpu *pcpu;
1652 
1653 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1654 		("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1655 
1656 	if (bp->b_qindex != BQUEUE_NONE)
1657 		panic("bqrelse: free buffer onto another queue???");
1658 
1659 	buf_act_advance(bp);
1660 
1661 	pcpu = &bufpcpu[bp->b_qcpu];
1662 	spin_lock(&pcpu->spin);
1663 
1664 	if (bp->b_flags & B_LOCKED) {
1665 		/*
1666 		 * Locked buffers are released to the locked queue.  However,
1667 		 * if the buffer is dirty it will first go into the dirty
1668 		 * queue and later on after the I/O completes successfully it
1669 		 * will be released to the locked queue.
1670 		 */
1671 		bp->b_qindex = BQUEUE_LOCKED;
1672 		TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1673 				  bp, b_freelist);
1674 	} else if (bp->b_flags & B_DELWRI) {
1675 		bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1676 			       BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1677 		TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1678 				  bp, b_freelist);
1679 	} else if (vm_paging_min()) {
1680 		/*
1681 		 * We are too low on memory, we have to try to free the
1682 		 * buffer (most importantly: the wired pages making up its
1683 		 * backing store) *now*.
1684 		 */
1685 		spin_unlock(&pcpu->spin);
1686 		brelse(bp);
1687 		return;
1688 	} else {
1689 		bp->b_qindex = BQUEUE_CLEAN;
1690 		TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1691 				  bp, b_freelist);
1692 	}
1693 	spin_unlock(&pcpu->spin);
1694 
1695 	/*
1696 	 * We have now placed the buffer on the proper queue, but have yet
1697 	 * to unlock it.
1698 	 */
1699 	if ((bp->b_flags & B_LOCKED) == 0 &&
1700 	    ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1701 		bufcountwakeup();
1702 	}
1703 
1704 	/*
1705 	 * Something we can maybe free or reuse.
1706 	 */
1707 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1708 		bufspacewakeup();
1709 
1710 	/*
1711 	 * Final cleanup and unlock.  Clear bits that are only used while a
1712 	 * buffer is actively locked.
1713 	 */
1714 	bp->b_flags &= ~(B_NOCACHE | B_RELBUF);
1715 	dsched_buf_exit(bp);
1716 	BUF_UNLOCK(bp);
1717 }
1718 
1719 /*
1720  * Hold a buffer, preventing it from being reused.  This will prevent
1721  * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1722  * operations.  If a B_INVAL operation occurs the buffer will remain held
1723  * but the underlying pages may get ripped out.
1724  *
1725  * These functions are typically used in VOP_READ/VOP_WRITE functions
1726  * to hold a buffer during a copyin or copyout, preventing deadlocks
1727  * or recursive lock panics when read()/write() is used over mmap()'d
1728  * space.
1729  *
1730  * NOTE: bqhold() requires that the buffer be locked at the time of the
1731  *	 hold.  bqdrop() has no requirements other than the buffer having
1732  *	 previously been held.
1733  */
1734 void
bqhold(struct buf * bp)1735 bqhold(struct buf *bp)
1736 {
1737 	atomic_add_int(&bp->b_refs, 1);
1738 }
1739 
1740 void
bqdrop(struct buf * bp)1741 bqdrop(struct buf *bp)
1742 {
1743 	KKASSERT(bp->b_refs > 0);
1744 	atomic_add_int(&bp->b_refs, -1);
1745 }
1746 
1747 /*
1748  * Return backing pages held by the buffer 'bp' back to the VM system.
1749  * This routine is called when the bp is invalidated, released, or
1750  * reused.
1751  *
1752  * The KVA mapping (b_data) for the underlying pages is removed by
1753  * this function.
1754  *
1755  * WARNING! This routine is integral to the low memory critical path
1756  *          when a buffer is B_RELBUF'd.  If the system has a severe page
1757  *          deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1758  *          queues so they can be reused in the current pageout daemon
1759  *          pass.
1760  */
1761 static void
vfs_vmio_release(struct buf * bp)1762 vfs_vmio_release(struct buf *bp)
1763 {
1764 	int i;
1765 	vm_page_t m;
1766 
1767 	for (i = 0; i < bp->b_xio.xio_npages; i++) {
1768 		m = bp->b_xio.xio_pages[i];
1769 		bp->b_xio.xio_pages[i] = NULL;
1770 
1771 		/*
1772 		 * We need to own the page in order to safely unwire it.
1773 		 */
1774 		vm_page_busy_wait(m, FALSE, "vmiopg");
1775 
1776 		/*
1777 		 * The VFS is telling us this is not a meta-data buffer
1778 		 * even if it is backed by a block device.
1779 		 */
1780 		if (bp->b_flags & B_NOTMETA)
1781 			vm_page_flag_set(m, PG_NOTMETA);
1782 
1783 		/*
1784 		 * This is a very important bit of code.  We try to track
1785 		 * VM page use whether the pages are wired into the buffer
1786 		 * cache or not.  While wired into the buffer cache the
1787 		 * bp tracks the act_count.
1788 		 *
1789 		 * We can choose to place unwired pages on the inactive
1790 		 * queue (0) or active queue (1).  If we place too many
1791 		 * on the active queue the queue will cycle the act_count
1792 		 * on pages we'd like to keep, just from single-use pages
1793 		 * (such as when doing a tar-up or file scan).
1794 		 */
1795 		if (bp->b_act_count < vm_cycle_point)
1796 			vm_page_unwire(m, 0);
1797 		else
1798 			vm_page_unwire(m, 1);
1799 
1800 		/*
1801 		 * If the wire_count has dropped to 0 we may need to take
1802 		 * further action before unbusying the page.
1803 		 *
1804 		 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1805 		 */
1806 		if (m->wire_count == 0) {
1807 			if (bp->b_flags & B_DIRECT) {
1808 				/*
1809 				 * Attempt to free the page if B_DIRECT is
1810 				 * set, the caller does not desire the page
1811 				 * to be cached.
1812 				 */
1813 				vm_page_wakeup(m);
1814 				vm_page_try_to_free(m);
1815 			} else if ((bp->b_flags & (B_NOTMETA | B_TTC)) ||
1816 				   vm_paging_min()) {
1817 				/*
1818 				 * Attempt to move the page to PQ_CACHE
1819 				 * if B_NOTMETA is set.  This flag is set
1820 				 * by HAMMER to remove one of the two pages
1821 				 * present when double buffering is enabled.
1822 				 *
1823 				 * Attempt to move the page to PQ_CACHE
1824 				 * If we have a severe page deficit.  This
1825 				 * will cause buffer cache operations related
1826 				 * to pageouts to recycle the related pages
1827 				 * in order to avoid a low memory deadlock.
1828 				 */
1829 				m->act_count = bp->b_act_count;
1830 				vm_page_try_to_cache(m);
1831 			} else {
1832 				/*
1833 				 * Nominal case, leave the page on the
1834 				 * queue the original unwiring placed it on
1835 				 * (active or inactive).
1836 				 */
1837 				m->act_count = bp->b_act_count;
1838 				vm_page_wakeup(m);
1839 			}
1840 		} else {
1841 			vm_page_wakeup(m);
1842 		}
1843 	}
1844 
1845 	/*
1846 	 * Zero out the pmap pte's for the mapping, but don't bother
1847 	 * invalidating the TLB.  The range will be properly invalidating
1848 	 * when new pages are entered into the mapping.
1849 	 *
1850 	 * This in particular reduces tmpfs tear-down overhead and reduces
1851 	 * buffer cache re-use overhead (one invalidation sequence instead
1852 	 * of two per re-use).
1853 	 */
1854 	pmap_qremove_noinval(trunc_page((vm_offset_t) bp->b_data),
1855 			     bp->b_xio.xio_npages);
1856 	CPUMASK_ASSZERO(bp->b_cpumask);
1857 	if (bp->b_bufsize) {
1858 		atomic_add_long(&bufspace, -bp->b_bufsize);
1859 		bp->b_bufsize = 0;
1860 		bufspacewakeup();
1861 	}
1862 	bp->b_xio.xio_npages = 0;
1863 	bp->b_flags &= ~(B_VMIO | B_TTC);
1864 	KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1865 	if (bp->b_vp)
1866 		brelvp(bp);
1867 }
1868 
1869 /*
1870  * Find and initialize a new buffer header, freeing up existing buffers
1871  * in the bufqueues as necessary.  The new buffer is returned locked.
1872  *
1873  * Important:  B_INVAL is not set.  If the caller wishes to throw the
1874  * buffer away, the caller must set B_INVAL prior to calling brelse().
1875  *
1876  * We block if:
1877  *	We have insufficient buffer headers
1878  *	We have insufficient buffer space
1879  *
1880  * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1881  * Instead we ask the buf daemon to do it for us.  We attempt to
1882  * avoid piecemeal wakeups of the pageout daemon.
1883  */
1884 struct buf *
getnewbuf(int blkflags,int slptimeo,int size,int maxsize)1885 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1886 {
1887 	struct bufpcpu *pcpu;
1888 	struct buf *bp;
1889 	struct buf *nbp;
1890 	int nqindex;
1891 	int nqcpu;
1892 	int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1893 	int maxloops = 200000;
1894 	int restart_reason = 0;
1895 	struct buf *restart_bp = NULL;
1896 	static char flushingbufs[MAXCPU];
1897 	char *flushingp;
1898 
1899 	/*
1900 	 * We can't afford to block since we might be holding a vnode lock,
1901 	 * which may prevent system daemons from running.  We deal with
1902 	 * low-memory situations by proactively returning memory and running
1903 	 * async I/O rather then sync I/O.
1904 	 */
1905 
1906 	++getnewbufcalls;
1907 	nqcpu = mycpu->gd_cpuid;
1908 	flushingp = &flushingbufs[nqcpu];
1909 restart:
1910 	if (bufspace < lobufspace)
1911 		*flushingp = 0;
1912 
1913 	if (debug_bufbio && --maxloops == 0)
1914 		panic("getnewbuf, excessive loops on cpu %d restart %d (%p)",
1915 			mycpu->gd_cpuid, restart_reason, restart_bp);
1916 
1917 	/*
1918 	 * Setup for scan.  If we do not have enough free buffers,
1919 	 * we setup a degenerate case that immediately fails.  Note
1920 	 * that if we are specially marked process, we are allowed to
1921 	 * dip into our reserves.
1922 	 *
1923 	 * The scanning sequence is nominally:  EMPTY->CLEAN
1924 	 */
1925 	pcpu = &bufpcpu[nqcpu];
1926 	spin_lock(&pcpu->spin);
1927 
1928 	/*
1929 	 * Prime the scan for this cpu.  Locate the first buffer to
1930 	 * check.  If we are flushing buffers we must skip the
1931 	 * EMPTY queue.
1932 	 */
1933 	nqindex = BQUEUE_EMPTY;
1934 	nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]);
1935 	if (nbp == NULL || *flushingp) {
1936 		nqindex = BQUEUE_CLEAN;
1937 		nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]);
1938 	}
1939 
1940 	/*
1941 	 * Run scan, possibly freeing data and/or kva mappings on the fly,
1942 	 * depending.
1943 	 *
1944 	 * WARNING! spin is held!
1945 	 */
1946 	while ((bp = nbp) != NULL) {
1947 		int qindex = nqindex;
1948 
1949 		nbp = TAILQ_NEXT(bp, b_freelist);
1950 
1951 		/*
1952 		 * BQUEUE_CLEAN - B_AGE special case.  If not set the bp
1953 		 * cycles through the queue twice before being selected.
1954 		 */
1955 		if (qindex == BQUEUE_CLEAN &&
1956 		    (bp->b_flags & B_AGE) == 0 && nbp) {
1957 			bp->b_flags |= B_AGE;
1958 			TAILQ_REMOVE(&pcpu->bufqueues[qindex],
1959 				     bp, b_freelist);
1960 			TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex],
1961 					  bp, b_freelist);
1962 			continue;
1963 		}
1964 
1965 		/*
1966 		 * Calculate next bp ( we can only use it if we do not block
1967 		 * or do other fancy things ).
1968 		 */
1969 		if (nbp == NULL) {
1970 			switch(qindex) {
1971 			case BQUEUE_EMPTY:
1972 				nqindex = BQUEUE_CLEAN;
1973 				if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN])))
1974 					break;
1975 				/* fall through */
1976 			case BQUEUE_CLEAN:
1977 				/*
1978 				 * nbp is NULL.
1979 				 */
1980 				break;
1981 			}
1982 		}
1983 
1984 		/*
1985 		 * Sanity Checks
1986 		 */
1987 		KASSERT(bp->b_qindex == qindex,
1988 			("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1989 
1990 		/*
1991 		 * Note: we no longer distinguish between VMIO and non-VMIO
1992 		 * buffers.
1993 		 */
1994 		KASSERT((bp->b_flags & B_DELWRI) == 0,
1995 			("delwri buffer %p found in queue %d", bp, qindex));
1996 
1997 		/*
1998 		 * Do not try to reuse a buffer with a non-zero b_refs.
1999 		 * This is an unsynchronized test.  A synchronized test
2000 		 * is also performed after we lock the buffer.
2001 		 */
2002 		if (bp->b_refs)
2003 			continue;
2004 
2005 		/*
2006 		 * Start freeing the bp.  This is somewhat involved.  nbp
2007 		 * remains valid only for BQUEUE_EMPTY bp's.  Buffers
2008 		 * on the clean list must be disassociated from their
2009 		 * current vnode.  Buffers on the empty lists have
2010 		 * already been disassociated.
2011 		 *
2012 		 * b_refs is checked after locking along with queue changes.
2013 		 * We must check here to deal with zero->nonzero transitions
2014 		 * made by the owner of the buffer lock, which is used by
2015 		 * VFS's to hold the buffer while issuing an unlocked
2016 		 * uiomove()s.  We cannot invalidate the buffer's pages
2017 		 * for this case.  Once we successfully lock a buffer the
2018 		 * only 0->1 transitions of b_refs will occur via findblk().
2019 		 *
2020 		 * We must also check for queue changes after successful
2021 		 * locking as the current lock holder may dispose of the
2022 		 * buffer and change its queue.
2023 		 */
2024 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2025 			spin_unlock(&pcpu->spin);
2026 			tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2027 			restart_reason = 1;
2028 			restart_bp = bp;
2029 			goto restart;
2030 		}
2031 		if (bp->b_qindex != qindex || bp->b_refs) {
2032 			spin_unlock(&pcpu->spin);
2033 			BUF_UNLOCK(bp);
2034 			restart_reason = 2;
2035 			restart_bp = bp;
2036 			goto restart;
2037 		}
2038 		bremfree_locked(bp);
2039 		spin_unlock(&pcpu->spin);
2040 
2041 		/*
2042 		 * Dependancies must be handled before we disassociate the
2043 		 * vnode.
2044 		 *
2045 		 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2046 		 * be immediately disassociated.  HAMMER then becomes
2047 		 * responsible for releasing the buffer.
2048 		 *
2049 		 * NOTE: spin is UNLOCKED now.
2050 		 */
2051 		if (LIST_FIRST(&bp->b_dep) != NULL) {
2052 			buf_deallocate(bp);
2053 			if (bp->b_flags & B_LOCKED) {
2054 				bqrelse(bp);
2055 				restart_reason = 3;
2056 				restart_bp = bp;
2057 				goto restart;
2058 			}
2059 			KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2060 		}
2061 
2062 		/*
2063 		 * CLEAN buffers have content or associations that must be
2064 		 * cleaned out if not repurposing.
2065 		 */
2066 		if (qindex == BQUEUE_CLEAN) {
2067 			if (bp->b_flags & B_VMIO)
2068 				vfs_vmio_release(bp);
2069 			if (bp->b_vp)
2070 				brelvp(bp);
2071 		}
2072 
2073 		/*
2074 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2075 		 * the scan from this point on.
2076 		 *
2077 		 * Get the rest of the buffer freed up.  b_kva* is still
2078 		 * valid after this operation.
2079 		 */
2080 		KASSERT(bp->b_vp == NULL,
2081 			("bp3 %p flags %08x vnode %p qindex %d "
2082 			 "unexpectededly still associated!",
2083 			 bp, bp->b_flags, bp->b_vp, qindex));
2084 		KKASSERT((bp->b_flags & B_HASHED) == 0);
2085 
2086 		if (bp->b_bufsize)
2087 			allocbuf(bp, 0);
2088 
2089                 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) {
2090 			kprintf("getnewbuf: caught bug vp queue "
2091 				"%p/%08x qidx %d\n",
2092 				bp, bp->b_flags, qindex);
2093 			brelvp(bp);
2094 		}
2095 		bp->b_flags = B_BNOCLIP;
2096 		bp->b_cmd = BUF_CMD_DONE;
2097 		bp->b_vp = NULL;
2098 		bp->b_error = 0;
2099 		bp->b_resid = 0;
2100 		bp->b_bcount = 0;
2101 		bp->b_xio.xio_npages = 0;
2102 		bp->b_dirtyoff = bp->b_dirtyend = 0;
2103 		bp->b_act_count = ACT_INIT;
2104 		reinitbufbio(bp);
2105 		KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2106 		buf_dep_init(bp);
2107 		if (blkflags & GETBLK_BHEAVY)
2108 			bp->b_flags |= B_HEAVY;
2109 
2110 		if (bufspace >= hibufspace)
2111 			*flushingp = 1;
2112 		if (bufspace < lobufspace)
2113 			*flushingp = 0;
2114 		if (*flushingp) {
2115 			bp->b_flags |= B_INVAL;
2116 			brelse(bp);
2117 			restart_reason = 5;
2118 			restart_bp = bp;
2119 			goto restart;
2120 		}
2121 
2122 		/*
2123 		 * b_refs can transition to a non-zero value while we hold
2124 		 * the buffer locked due to a findblk().  Our brelvp() above
2125 		 * interlocked any future possible transitions due to
2126 		 * findblk()s.
2127 		 *
2128 		 * If we find b_refs to be non-zero we can destroy the
2129 		 * buffer's contents but we cannot yet reuse the buffer.
2130 		 */
2131 		if (bp->b_refs) {
2132 			bp->b_flags |= B_INVAL;
2133 			brelse(bp);
2134 			restart_reason = 6;
2135 			restart_bp = bp;
2136 
2137 			goto restart;
2138 		}
2139 
2140 		/*
2141 		 * We found our buffer!
2142 		 */
2143 		break;
2144 	}
2145 
2146 	/*
2147 	 * If we exhausted our list, iterate other cpus.  If that fails,
2148 	 * sleep as appropriate.  We may have to wakeup various daemons
2149 	 * and write out some dirty buffers.
2150 	 *
2151 	 * Generally we are sleeping due to insufficient buffer space.
2152 	 *
2153 	 * NOTE: spin is held if bp is NULL, else it is not held.
2154 	 */
2155 	if (bp == NULL) {
2156 		int flags;
2157 		char *waitmsg;
2158 
2159 		spin_unlock(&pcpu->spin);
2160 
2161 		nqcpu = (nqcpu + 1) % ncpus;
2162 		if (nqcpu != mycpu->gd_cpuid) {
2163 			restart_reason = 7;
2164 			restart_bp = bp;
2165 			goto restart;
2166 		}
2167 
2168 		if (bufspace >= hibufspace) {
2169 			waitmsg = "bufspc";
2170 			flags = VFS_BIO_NEED_BUFSPACE;
2171 		} else {
2172 			waitmsg = "newbuf";
2173 			flags = VFS_BIO_NEED_ANY;
2174 		}
2175 
2176 		bd_speedup();	/* heeeelp */
2177 		atomic_set_int(&needsbuffer, flags);
2178 		while (needsbuffer & flags) {
2179 			int value;
2180 
2181 			tsleep_interlock(&needsbuffer, 0);
2182 			value = atomic_fetchadd_int(&needsbuffer, 0);
2183 			if (value & flags) {
2184 				if (tsleep(&needsbuffer, PINTERLOCKED|slpflags,
2185 					   waitmsg, slptimeo)) {
2186 					return (NULL);
2187 				}
2188 			}
2189 		}
2190 	} else {
2191 		/*
2192 		 * We finally have a valid bp.  Reset b_data.
2193 		 *
2194 		 * (spin is not held)
2195 		 */
2196 		bp->b_data = bp->b_kvabase;
2197 	}
2198 	return(bp);
2199 }
2200 
2201 /*
2202  * buf_daemon:
2203  *
2204  *	Buffer flushing daemon.  Buffers are normally flushed by the
2205  *	update daemon but if it cannot keep up this process starts to
2206  *	take the load in an attempt to prevent getnewbuf() from blocking.
2207  *
2208  *	Once a flush is initiated it does not stop until the number
2209  *	of buffers falls below lodirtybuffers, but we will wake up anyone
2210  *	waiting at the mid-point.
2211  */
2212 static struct kproc_desc buf_kp = {
2213 	"bufdaemon",
2214 	buf_daemon,
2215 	&bufdaemon_td
2216 };
2217 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2218 	kproc_start, &buf_kp);
2219 
2220 static struct kproc_desc bufhw_kp = {
2221 	"bufdaemon_hw",
2222 	buf_daemon_hw,
2223 	&bufdaemonhw_td
2224 };
2225 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2226 	kproc_start, &bufhw_kp);
2227 
2228 static void
buf_daemon1(struct thread * td,int queue,int (* buf_limit_fn)(long),int * bd_req)2229 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long),
2230 	    int *bd_req)
2231 {
2232 	long limit;
2233 	struct buf *marker;
2234 
2235 	marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO);
2236 	marker->b_flags |= B_MARKER;
2237 	marker->b_qindex = BQUEUE_NONE;
2238 	marker->b_qcpu = 0;
2239 
2240 	/*
2241 	 * This process needs to be suspended prior to shutdown sync.
2242 	 */
2243 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2244 			      td, SHUTDOWN_PRI_LAST);
2245 	curthread->td_flags |= TDF_SYSTHREAD;
2246 
2247 	/*
2248 	 * This process is allowed to take the buffer cache to the limit
2249 	 */
2250 	for (;;) {
2251 		kproc_suspend_loop();
2252 
2253 		/*
2254 		 * Do the flush as long as the number of dirty buffers
2255 		 * (including those running) exceeds lodirtybufspace.
2256 		 *
2257 		 * When flushing limit running I/O to hirunningspace
2258 		 * Do the flush.  Limit the amount of in-transit I/O we
2259 		 * allow to build up, otherwise we would completely saturate
2260 		 * the I/O system.  Wakeup any waiting processes before we
2261 		 * normally would so they can run in parallel with our drain.
2262 		 *
2263 		 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2264 		 * but because we split the operation into two threads we
2265 		 * have to cut it in half for each thread.
2266 		 */
2267 		waitrunningbufspace();
2268 		limit = lodirtybufspace / 2;
2269 		while (buf_limit_fn(limit)) {
2270 			if (flushbufqueues(marker, queue) == 0)
2271 				break;
2272 			if (runningbufspace < hirunningspace)
2273 				continue;
2274 			waitrunningbufspace();
2275 		}
2276 
2277 		/*
2278 		 * We reached our low water mark, reset the
2279 		 * request and sleep until we are needed again.
2280 		 * The sleep is just so the suspend code works.
2281 		 */
2282 		tsleep_interlock(bd_req, 0);
2283 		if (atomic_swap_int(bd_req, 0) == 0)
2284 			tsleep(bd_req, PINTERLOCKED, "psleep", hz);
2285 	}
2286 	/* NOT REACHED */
2287 	/*kfree(marker, M_BIOBUF);*/
2288 }
2289 
2290 static int
buf_daemon_limit(long limit)2291 buf_daemon_limit(long limit)
2292 {
2293 	return (runningbufspace + dirtykvaspace > limit ||
2294 		dirtybufcount - dirtybufcounthw >= nbuf / 2);
2295 }
2296 
2297 static int
buf_daemon_hw_limit(long limit)2298 buf_daemon_hw_limit(long limit)
2299 {
2300 	return (runningbufspace + dirtykvaspace > limit ||
2301 		dirtybufcounthw >= nbuf / 2);
2302 }
2303 
2304 static void
buf_daemon(void)2305 buf_daemon(void)
2306 {
2307 	buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit,
2308 		    &bd_request);
2309 }
2310 
2311 static void
buf_daemon_hw(void)2312 buf_daemon_hw(void)
2313 {
2314 	buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit,
2315 		    &bd_request_hw);
2316 }
2317 
2318 /*
2319  * Flush up to (flushperqueue) buffers in the dirty queue.  Each cpu has a
2320  * localized version of the queue.  Each call made to this function iterates
2321  * to another cpu.  It is desireable to flush several buffers from the same
2322  * cpu's queue at once, as these are likely going to be linear.
2323  *
2324  * We must be careful to free up B_INVAL buffers instead of write them, which
2325  * NFS is particularly sensitive to.
2326  *
2327  * B_RELBUF may only be set by VFSs.  We do set B_AGE to indicate that we
2328  * really want to try to get the buffer out and reuse it due to the write
2329  * load on the machine.
2330  *
2331  * We must lock the buffer in order to check its validity before we can mess
2332  * with its contents.  spin isn't enough.
2333  */
2334 static int
flushbufqueues(struct buf * marker,bufq_type_t q)2335 flushbufqueues(struct buf *marker, bufq_type_t q)
2336 {
2337 	struct bufpcpu *pcpu;
2338 	struct buf *bp;
2339 	int r = 0;
2340 	u_int loops = flushperqueue;
2341 	int lcpu = marker->b_qcpu;
2342 
2343 	KKASSERT(marker->b_qindex == BQUEUE_NONE);
2344 	KKASSERT(marker->b_flags & B_MARKER);
2345 
2346 again:
2347 	/*
2348 	 * Spinlock needed to perform operations on the queue and may be
2349 	 * held through a non-blocking BUF_LOCK(), but cannot be held when
2350 	 * BUF_UNLOCK()ing or through any other major operation.
2351 	 */
2352 	pcpu = &bufpcpu[marker->b_qcpu];
2353 	spin_lock(&pcpu->spin);
2354 	marker->b_qindex = q;
2355 	TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist);
2356 	bp = marker;
2357 
2358 	while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) {
2359 		/*
2360 		 * NOTE: spinlock is always held at the top of the loop
2361 		 */
2362 		if (bp->b_flags & B_MARKER)
2363 			continue;
2364 		if ((bp->b_flags & B_DELWRI) == 0) {
2365 			kprintf("Unexpected clean buffer %p\n", bp);
2366 			continue;
2367 		}
2368 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
2369 			continue;
2370 		KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q);
2371 
2372 		/*
2373 		 * Once the buffer is locked we will have no choice but to
2374 		 * unlock the spinlock around a later BUF_UNLOCK and re-set
2375 		 * bp = marker when looping.  Move the marker now to make
2376 		 * things easier.
2377 		 */
2378 		TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2379 		TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist);
2380 
2381 		/*
2382 		 * Must recheck B_DELWRI after successfully locking
2383 		 * the buffer.
2384 		 */
2385 		if ((bp->b_flags & B_DELWRI) == 0) {
2386 			spin_unlock(&pcpu->spin);
2387 			BUF_UNLOCK(bp);
2388 			spin_lock(&pcpu->spin);
2389 			bp = marker;
2390 			continue;
2391 		}
2392 
2393 		/*
2394 		 * Remove the buffer from its queue.  We still own the
2395 		 * spinlock here.
2396 		 */
2397 		_bremfree(bp);
2398 
2399 		/*
2400 		 * Disposing of an invalid buffer counts as a flush op
2401 		 */
2402 		if (bp->b_flags & B_INVAL) {
2403 			spin_unlock(&pcpu->spin);
2404 			brelse(bp);
2405 			goto doloop;
2406 		}
2407 
2408 		/*
2409 		 * Release the spinlock for the more complex ops we
2410 		 * are now going to do.
2411 		 */
2412 		spin_unlock(&pcpu->spin);
2413 		lwkt_yield();
2414 
2415 		/*
2416 		 * This is a bit messy
2417 		 */
2418 		if (LIST_FIRST(&bp->b_dep) != NULL &&
2419 		    (bp->b_flags & B_DEFERRED) == 0 &&
2420 		    buf_countdeps(bp, 0)) {
2421 			spin_lock(&pcpu->spin);
2422 			TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist);
2423 			bp->b_qindex = q;
2424 			bp->b_flags |= B_DEFERRED;
2425 			spin_unlock(&pcpu->spin);
2426 			BUF_UNLOCK(bp);
2427 			spin_lock(&pcpu->spin);
2428 			bp = marker;
2429 			continue;
2430 		}
2431 
2432 		/*
2433 		 * spinlock not held here.
2434 		 *
2435 		 * If the buffer has a dependancy, buf_checkwrite() must
2436 		 * also return 0 for us to be able to initate the write.
2437 		 *
2438 		 * If the buffer is flagged B_ERROR it may be requeued
2439 		 * over and over again, we try to avoid a live lock.
2440 		 */
2441 		if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2442 			brelse(bp);
2443 		} else if (bp->b_flags & B_ERROR) {
2444 			tsleep(bp, 0, "bioer", 1);
2445 			bp->b_flags &= ~B_AGE;
2446 			cluster_awrite(bp);
2447 		} else {
2448 			bp->b_flags |= B_AGE | B_KVABIO;
2449 			cluster_awrite(bp);
2450 		}
2451 		/* bp invalid but needs to be NULL-tested if we break out */
2452 doloop:
2453 		spin_lock(&pcpu->spin);
2454 		++r;
2455 		if (--loops == 0)
2456 			break;
2457 		bp = marker;
2458 	}
2459 	/* bp is invalid here but can be NULL-tested to advance */
2460 
2461 	TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2462 	marker->b_qindex = BQUEUE_NONE;
2463 	spin_unlock(&pcpu->spin);
2464 
2465 	/*
2466 	 * Advance the marker to be fair.
2467 	 */
2468 	marker->b_qcpu = (marker->b_qcpu + 1) % ncpus;
2469 	if (bp == NULL) {
2470 		if (marker->b_qcpu != lcpu)
2471 			goto again;
2472 	}
2473 
2474 	return (r);
2475 }
2476 
2477 /*
2478  * inmem:
2479  *
2480  *	Returns true if no I/O is needed to access the associated VM object.
2481  *	This is like findblk except it also hunts around in the VM system for
2482  *	the data.
2483  *
2484  *	Note that we ignore vm_page_free() races from interrupts against our
2485  *	lookup, since if the caller is not protected our return value will not
2486  *	be any more valid then otherwise once we exit the critical section.
2487  */
2488 int
inmem(struct vnode * vp,off_t loffset)2489 inmem(struct vnode *vp, off_t loffset)
2490 {
2491 	vm_object_t obj;
2492 	vm_offset_t toff, tinc, size;
2493 	vm_page_t m;
2494 	int res = 1;
2495 
2496 	if (findblk(vp, loffset, FINDBLK_TEST))
2497 		return 1;
2498 	if (vp->v_mount == NULL)
2499 		return 0;
2500 	if ((obj = vp->v_object) == NULL)
2501 		return 0;
2502 
2503 	size = PAGE_SIZE;
2504 	if (size > vp->v_mount->mnt_stat.f_iosize)
2505 		size = vp->v_mount->mnt_stat.f_iosize;
2506 
2507 	vm_object_hold(obj);
2508 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2509 		m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2510 		if (m == NULL) {
2511 			res = 0;
2512 			break;
2513 		}
2514 		tinc = size;
2515 		if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2516 			tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2517 		if (vm_page_is_valid(m,
2518 		    (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2519 			res = 0;
2520 			break;
2521 		}
2522 	}
2523 	vm_object_drop(obj);
2524 	return (res);
2525 }
2526 
2527 /*
2528  * findblk:
2529  *
2530  *	Locate and return the specified buffer.  Unless flagged otherwise,
2531  *	a locked buffer will be returned if it exists or NULL if it does not.
2532  *
2533  *	findblk()'d buffers are still on the bufqueues and if you intend
2534  *	to use your (locked NON-TEST) buffer you need to bremfree(bp)
2535  *	and possibly do other stuff to it.
2536  *
2537  *	FINDBLK_TEST	- Do not lock the buffer.  The caller is responsible
2538  *			  for locking the buffer and ensuring that it remains
2539  *			  the desired buffer after locking.
2540  *
2541  *	FINDBLK_NBLOCK	- Lock the buffer non-blocking.  If we are unable
2542  *			  to acquire the lock we return NULL, even if the
2543  *			  buffer exists.
2544  *
2545  *	FINDBLK_REF	- Returns the buffer ref'd, which prevents normal
2546  *			  reuse by getnewbuf() but does not prevent
2547  *			  disassociation (B_INVAL).  Used to avoid deadlocks
2548  *			  against random (vp,loffset)s due to reassignment.
2549  *
2550  *	FINDBLK_KVABIO	- Only applicable when returning a locked buffer.
2551  *			  Indicates that the caller supports B_KVABIO.
2552  *
2553  *	(0)		- Lock the buffer blocking.
2554  */
2555 struct buf *
findblk(struct vnode * vp,off_t loffset,int flags)2556 findblk(struct vnode *vp, off_t loffset, int flags)
2557 {
2558 	struct buf *bp;
2559 	int lkflags;
2560 
2561 	lkflags = LK_EXCLUSIVE;
2562 	if (flags & FINDBLK_NBLOCK)
2563 		lkflags |= LK_NOWAIT;
2564 
2565 	for (;;) {
2566 		/*
2567 		 * Lookup.  Ref the buf while holding v_token to prevent
2568 		 * reuse (but does not prevent diassociation).
2569 		 */
2570 		lwkt_gettoken_shared(&vp->v_token);
2571 		bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2572 		if (bp == NULL) {
2573 			lwkt_reltoken(&vp->v_token);
2574 			return(NULL);
2575 		}
2576 		bqhold(bp);
2577 		lwkt_reltoken(&vp->v_token);
2578 
2579 		/*
2580 		 * If testing only break and return bp, do not lock.
2581 		 */
2582 		if (flags & FINDBLK_TEST)
2583 			break;
2584 
2585 		/*
2586 		 * Lock the buffer, return an error if the lock fails.
2587 		 * (only FINDBLK_NBLOCK can cause the lock to fail).
2588 		 */
2589 		if (BUF_LOCK(bp, lkflags)) {
2590 			atomic_subtract_int(&bp->b_refs, 1);
2591 			/* bp = NULL; not needed */
2592 			return(NULL);
2593 		}
2594 
2595 		/*
2596 		 * Revalidate the locked buf before allowing it to be
2597 		 * returned.
2598 		 *
2599 		 * B_KVABIO is only set/cleared when locking.  When
2600 		 * clearing B_KVABIO, we must ensure that the buffer
2601 		 * is synchronized to all cpus.
2602 		 */
2603 		if (bp->b_vp == vp && bp->b_loffset == loffset) {
2604 			if (flags & FINDBLK_KVABIO)
2605 				bp->b_flags |= B_KVABIO;
2606 			else
2607 				bkvasync_all(bp);
2608 			break;
2609 		}
2610 		atomic_subtract_int(&bp->b_refs, 1);
2611 		BUF_UNLOCK(bp);
2612 	}
2613 
2614 	/*
2615 	 * Success
2616 	 */
2617 	if ((flags & FINDBLK_REF) == 0)
2618 		atomic_subtract_int(&bp->b_refs, 1);
2619 	return(bp);
2620 }
2621 
2622 /*
2623  * getcacheblk:
2624  *
2625  *	Similar to getblk() except only returns the buffer if it is
2626  *	B_CACHE and requires no other manipulation.  Otherwise NULL
2627  *	is returned.  NULL is also returned if GETBLK_NOWAIT is set
2628  *	and the getblk() would block.
2629  *
2630  *	If B_RAM is set the buffer might be just fine, but we return
2631  *	NULL anyway because we want the code to fall through to the
2632  *	cluster read to issue more read-aheads.  Otherwise read-ahead breaks.
2633  *
2634  *	If blksize is 0 the buffer cache buffer must already be fully
2635  *	cached.
2636  *
2637  *	If blksize is non-zero getblk() will be used, allowing a buffer
2638  *	to be reinstantiated from its VM backing store.  The buffer must
2639  *	still be fully cached after reinstantiation to be returned.
2640  */
2641 struct buf *
getcacheblk(struct vnode * vp,off_t loffset,int blksize,int blkflags)2642 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags)
2643 {
2644 	struct buf *bp;
2645 	int fndflags = 0;
2646 
2647 	if (blkflags & GETBLK_NOWAIT)
2648 		fndflags |= FINDBLK_NBLOCK;
2649 	if (blkflags & GETBLK_KVABIO)
2650 		fndflags |= FINDBLK_KVABIO;
2651 
2652 	if (blksize) {
2653 		bp = getblk(vp, loffset, blksize, blkflags, 0);
2654 		if (bp) {
2655 			if ((bp->b_flags & (B_INVAL | B_CACHE)) == B_CACHE) {
2656 				bp->b_flags &= ~B_AGE;
2657 				if (bp->b_flags & B_RAM) {
2658 					bqrelse(bp);
2659 					bp = NULL;
2660 				}
2661 			} else {
2662 				brelse(bp);
2663 				bp = NULL;
2664 			}
2665 		}
2666 	} else {
2667 		bp = findblk(vp, loffset, fndflags);
2668 		if (bp) {
2669 			if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2670 			    B_CACHE) {
2671 				bp->b_flags &= ~B_AGE;
2672 				bremfree(bp);
2673 			} else {
2674 				BUF_UNLOCK(bp);
2675 				bp = NULL;
2676 			}
2677 		}
2678 	}
2679 	return (bp);
2680 }
2681 
2682 /*
2683  * getblk:
2684  *
2685  *	Get a block given a specified block and offset into a file/device.
2686  * 	B_INVAL may or may not be set on return.  The caller should clear
2687  *	B_INVAL prior to initiating a READ.
2688  *
2689  *	IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2690  *	IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2691  *	OR SET B_INVAL BEFORE RETIRING IT.  If you retire a getblk'd buffer
2692  *	without doing any of those things the system will likely believe
2693  *	the buffer to be valid (especially if it is not B_VMIO), and the
2694  *	next getblk() will return the buffer with B_CACHE set.
2695  *
2696  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2697  *	an existing buffer.
2698  *
2699  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2700  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2701  *	and then cleared based on the backing VM.  If the previous buffer is
2702  *	non-0-sized but invalid, B_CACHE will be cleared.
2703  *
2704  *	If getblk() must create a new buffer, the new buffer is returned with
2705  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2706  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2707  *	backing VM.
2708  *
2709  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2710  *	B_CACHE bit is clear.
2711  *
2712  *	What this means, basically, is that the caller should use B_CACHE to
2713  *	determine whether the buffer is fully valid or not and should clear
2714  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2715  *	the buffer by loading its data area with something, the caller needs
2716  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2717  *	the caller should set B_CACHE ( as an optimization ), else the caller
2718  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2719  *	a write attempt or if it was a successfull read.  If the caller
2720  *	intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2721  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2722  *
2723  *	getblk flags:
2724  *
2725  *	GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2726  *	GETBLK_BHEAVY - heavy-weight buffer cache buffer
2727  */
2728 struct buf *
getblk(struct vnode * vp,off_t loffset,int size,int blkflags,int slptimeo)2729 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2730 {
2731 	struct buf *bp;
2732 	int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2733 	int error;
2734 	int lkflags;
2735 
2736 	if (size > MAXBSIZE)
2737 		panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2738 	if (vp->v_object == NULL)
2739 		panic("getblk: vnode %p has no object!", vp);
2740 
2741 	/*
2742 	 * NOTE: findblk does not try to resolve KVABIO in REF-only mode.
2743 	 *	 we still have to handle that ourselves.
2744 	 */
2745 loop:
2746 	if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2747 		/*
2748 		 * The buffer was found in the cache, but we need to lock it.
2749 		 * We must acquire a ref on the bp to prevent reuse, but
2750 		 * this will not prevent disassociation (brelvp()) so we
2751 		 * must recheck (vp,loffset) after acquiring the lock.
2752 		 *
2753 		 * Without the ref the buffer could potentially be reused
2754 		 * before we acquire the lock and create a deadlock
2755 		 * situation between the thread trying to reuse the buffer
2756 		 * and us due to the fact that we would wind up blocking
2757 		 * on a random (vp,loffset).
2758 		 */
2759 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2760 			if (blkflags & GETBLK_NOWAIT) {
2761 				bqdrop(bp);
2762 				return(NULL);
2763 			}
2764 			lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2765 			if (blkflags & GETBLK_PCATCH)
2766 				lkflags |= LK_PCATCH;
2767 			error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2768 			if (error) {
2769 				bqdrop(bp);
2770 				if (error == ENOLCK)
2771 					goto loop;
2772 				return (NULL);
2773 			}
2774 			/* buffer may have changed on us */
2775 		}
2776 		bqdrop(bp);
2777 
2778 		/*
2779 		 * Once the buffer has been locked, make sure we didn't race
2780 		 * a buffer recyclement.  Buffers that are no longer hashed
2781 		 * will have b_vp == NULL, so this takes care of that check
2782 		 * as well.
2783 		 */
2784 		if (bp->b_vp != vp || bp->b_loffset != loffset) {
2785 #if 0
2786 			kprintf("Warning buffer %p (vp %p loffset %lld) "
2787 				"was recycled\n",
2788 				bp, vp, (long long)loffset);
2789 #endif
2790 			BUF_UNLOCK(bp);
2791 			goto loop;
2792 		}
2793 
2794 		/*
2795 		 * If SZMATCH any pre-existing buffer must be of the requested
2796 		 * size or NULL is returned.  The caller absolutely does not
2797 		 * want getblk() to bwrite() the buffer on a size mismatch.
2798 		 */
2799 		if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2800 			BUF_UNLOCK(bp);
2801 			return(NULL);
2802 		}
2803 
2804 		/*
2805 		 * All vnode-based buffers must be backed by a VM object.
2806 		 *
2807 		 * Set B_KVABIO for any incidental work, we will fix it
2808 		 * up later.
2809 		 */
2810 		KKASSERT(bp->b_flags & B_VMIO);
2811 		KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2812 		bp->b_flags &= ~B_AGE;
2813 		bp->b_flags |= B_KVABIO;
2814 
2815 		/*
2816 		 * Make sure that B_INVAL buffers do not have a cached
2817 		 * block number translation.
2818 		 */
2819 		if ((bp->b_flags & B_INVAL) &&
2820 		    (bp->b_bio2.bio_offset != NOOFFSET)) {
2821 			kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2822 				" did not have cleared bio_offset cache\n",
2823 				bp, vp, (long long)loffset);
2824 			clearbiocache(&bp->b_bio2);
2825 		}
2826 
2827 		/*
2828 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2829 		 * invalid.
2830 		 *
2831 		 * After the bremfree(), disposals must use b[q]relse().
2832 		 */
2833 		if (bp->b_flags & B_INVAL)
2834 			bp->b_flags &= ~B_CACHE;
2835 		bremfree(bp);
2836 
2837 		/*
2838 		 * Any size inconsistancy with a dirty buffer or a buffer
2839 		 * with a softupdates dependancy must be resolved.  Resizing
2840 		 * the buffer in such circumstances can lead to problems.
2841 		 *
2842 		 * Dirty or dependant buffers are written synchronously.
2843 		 * Other types of buffers are simply released and
2844 		 * reconstituted as they may be backed by valid, dirty VM
2845 		 * pages (but not marked B_DELWRI).
2846 		 *
2847 		 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2848 		 * and may be left over from a prior truncation (and thus
2849 		 * no longer represent the actual EOF point), so we
2850 		 * definitely do not want to B_NOCACHE the backing store.
2851 		 */
2852 		if (size != bp->b_bcount) {
2853 			if (bp->b_flags & B_DELWRI) {
2854 				bp->b_flags |= B_RELBUF;
2855 				bwrite(bp);
2856 			} else if (LIST_FIRST(&bp->b_dep)) {
2857 				bp->b_flags |= B_RELBUF;
2858 				bwrite(bp);
2859 			} else {
2860 				bp->b_flags |= B_RELBUF;
2861 				brelse(bp);
2862 			}
2863 			goto loop;
2864 		}
2865 		KKASSERT(size <= bp->b_kvasize);
2866 		KASSERT(bp->b_loffset != NOOFFSET,
2867 			("getblk: no buffer offset"));
2868 
2869 		/*
2870 		 * A buffer with B_DELWRI set and B_CACHE clear must
2871 		 * be committed before we can return the buffer in
2872 		 * order to prevent the caller from issuing a read
2873 		 * ( due to B_CACHE not being set ) and overwriting
2874 		 * it.
2875 		 *
2876 		 * Most callers, including NFS and FFS, need this to
2877 		 * operate properly either because they assume they
2878 		 * can issue a read if B_CACHE is not set, or because
2879 		 * ( for example ) an uncached B_DELWRI might loop due
2880 		 * to softupdates re-dirtying the buffer.  In the latter
2881 		 * case, B_CACHE is set after the first write completes,
2882 		 * preventing further loops.
2883 		 *
2884 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2885 		 * above while extending the buffer, we cannot allow the
2886 		 * buffer to remain with B_CACHE set after the write
2887 		 * completes or it will represent a corrupt state.  To
2888 		 * deal with this we set B_NOCACHE to scrap the buffer
2889 		 * after the write.
2890 		 *
2891 		 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2892 		 *     I'm not even sure this state is still possible
2893 		 *     now that getblk() writes out any dirty buffers
2894 		 *     on size changes.
2895 		 *
2896 		 * We might be able to do something fancy, like setting
2897 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2898 		 * so the below call doesn't set B_CACHE, but that gets real
2899 		 * confusing.  This is much easier.
2900 		 */
2901 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2902 			kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2903 				"and CACHE clear, b_flags %08x\n",
2904 				bp, (uintmax_t)bp->b_loffset, bp->b_flags);
2905 			bp->b_flags |= B_NOCACHE;
2906 			bwrite(bp);
2907 			goto loop;
2908 		}
2909 	} else {
2910 		/*
2911 		 * Buffer is not in-core, create new buffer.  The buffer
2912 		 * returned by getnewbuf() is locked.  Note that the returned
2913 		 * buffer is also considered valid (not marked B_INVAL).
2914 		 *
2915 		 * Calculating the offset for the I/O requires figuring out
2916 		 * the block size.  We use DEV_BSIZE for VBLK or VCHR and
2917 		 * the mount's f_iosize otherwise.  If the vnode does not
2918 		 * have an associated mount we assume that the passed size is
2919 		 * the block size.
2920 		 *
2921 		 * Note that vn_isdisk() cannot be used here since it may
2922 		 * return a failure for numerous reasons.   Note that the
2923 		 * buffer size may be larger then the block size (the caller
2924 		 * will use block numbers with the proper multiple).  Beware
2925 		 * of using any v_* fields which are part of unions.  In
2926 		 * particular, in DragonFly the mount point overloading
2927 		 * mechanism uses the namecache only and the underlying
2928 		 * directory vnode is not a special case.
2929 		 */
2930 		int bsize, maxsize;
2931 
2932 		if (vp->v_type == VBLK || vp->v_type == VCHR)
2933 			bsize = DEV_BSIZE;
2934 		else if (vp->v_mount)
2935 			bsize = vp->v_mount->mnt_stat.f_iosize;
2936 		else
2937 			bsize = size;
2938 
2939 		maxsize = size + (loffset & PAGE_MASK);
2940 		maxsize = imax(maxsize, bsize);
2941 
2942 		bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2943 		if (bp == NULL) {
2944 			if (slpflags || slptimeo)
2945 				return NULL;
2946 			goto loop;
2947 		}
2948 
2949 		/*
2950 		 * Atomically insert the buffer into the hash, so that it can
2951 		 * be found by findblk().
2952 		 *
2953 		 * If bgetvp() returns non-zero a collision occured, and the
2954 		 * bp will not be associated with the vnode.
2955 		 *
2956 		 * Make sure the translation layer has been cleared.
2957 		 */
2958 		bp->b_loffset = loffset;
2959 		bp->b_bio2.bio_offset = NOOFFSET;
2960 		/* bp->b_bio2.bio_next = NULL; */
2961 
2962 		if (bgetvp(vp, bp, size)) {
2963 			bp->b_flags |= B_INVAL;
2964 			brelse(bp);
2965 			goto loop;
2966 		}
2967 
2968 		/*
2969 		 * All vnode-based buffers must be backed by a VM object.
2970 		 *
2971 		 * Set B_KVABIO for incidental work
2972 		 */
2973 		KKASSERT(vp->v_object != NULL);
2974 		bp->b_flags |= B_VMIO | B_KVABIO;
2975 		KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2976 
2977 		allocbuf(bp, size);
2978 	}
2979 
2980 	/*
2981 	 * Do the nasty smp broadcast (if the buffer needs it) when KVABIO
2982 	 * is not supported.
2983 	 */
2984 	if (bp && (blkflags & GETBLK_KVABIO) == 0) {
2985 		bkvasync_all(bp);
2986 	}
2987 	return (bp);
2988 }
2989 
2990 /*
2991  * regetblk(bp)
2992  *
2993  * Reacquire a buffer that was previously released to the locked queue,
2994  * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2995  * set B_LOCKED (which handles the acquisition race).
2996  *
2997  * To this end, either B_LOCKED must be set or the dependancy list must be
2998  * non-empty.
2999  */
3000 void
regetblk(struct buf * bp)3001 regetblk(struct buf *bp)
3002 {
3003 	KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3004 	BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3005 	bremfree(bp);
3006 }
3007 
3008 /*
3009  * allocbuf:
3010  *
3011  *	This code constitutes the buffer memory from either anonymous system
3012  *	memory (in the case of non-VMIO operations) or from an associated
3013  *	VM object (in the case of VMIO operations).  This code is able to
3014  *	resize a buffer up or down.
3015  *
3016  *	Note that this code is tricky, and has many complications to resolve
3017  *	deadlock or inconsistant data situations.  Tread lightly!!!
3018  *	There are B_CACHE and B_DELWRI interactions that must be dealt with by
3019  *	the caller.  Calling this code willy nilly can result in the loss of
3020  *	data.
3021  *
3022  *	allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3023  *	B_CACHE for the non-VMIO case.
3024  *
3025  *	This routine does not need to be called from a critical section but you
3026  *	must own the buffer.
3027  */
3028 void
allocbuf(struct buf * bp,int size)3029 allocbuf(struct buf *bp, int size)
3030 {
3031 	vm_page_t m;
3032 	int newbsize;
3033 	int desiredpages;
3034 	int i;
3035 
3036 	if (BUF_LOCKINUSE(bp) == 0)
3037 		panic("allocbuf: buffer not busy");
3038 
3039 	if (bp->b_kvasize < size)
3040 		panic("allocbuf: buffer too small");
3041 
3042 	KKASSERT(bp->b_flags & B_VMIO);
3043 
3044 	newbsize = roundup2(size, DEV_BSIZE);
3045 	desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3046 			newbsize + PAGE_MASK) >> PAGE_SHIFT;
3047 	KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3048 
3049 	/*
3050 	 * Set B_CACHE initially if buffer is 0 length or will become
3051 	 * 0-length.
3052 	 */
3053 	if (size == 0 || bp->b_bufsize == 0)
3054 		bp->b_flags |= B_CACHE;
3055 
3056 	if (newbsize < bp->b_bufsize) {
3057 		/*
3058 		 * DEV_BSIZE aligned new buffer size is less then the
3059 		 * DEV_BSIZE aligned existing buffer size.  Figure out
3060 		 * if we have to remove any pages.
3061 		 */
3062 		if (desiredpages < bp->b_xio.xio_npages) {
3063 			for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3064 				/*
3065 				 * the page is not freed here -- it
3066 				 * is the responsibility of
3067 				 * vnode_pager_setsize
3068 				 */
3069 				m = bp->b_xio.xio_pages[i];
3070 				KASSERT(m != bogus_page,
3071 				    ("allocbuf: bogus page found"));
3072 				vm_page_busy_wait(m, TRUE, "biodep");
3073 				bp->b_xio.xio_pages[i] = NULL;
3074 				vm_page_unwire(m, 0);
3075 				vm_page_wakeup(m);
3076 			}
3077 			pmap_qremove_noinval((vm_offset_t)
3078 				      trunc_page((vm_offset_t)bp->b_data) +
3079 				      (desiredpages << PAGE_SHIFT),
3080 				     (bp->b_xio.xio_npages - desiredpages));
3081 			bp->b_xio.xio_npages = desiredpages;
3082 
3083 			/*
3084 			 * Don't bother invalidating the pmap changes
3085 			 * (which wastes global SMP invalidation IPIs)
3086 			 * when setting the size to 0.  This case occurs
3087 			 * when called via getnewbuf() during buffer
3088 			 * recyclement.
3089 			 */
3090 			if (desiredpages == 0) {
3091 				CPUMASK_ASSZERO(bp->b_cpumask);
3092 			} else {
3093 				bkvareset(bp);
3094 			}
3095 		}
3096 	} else if (size > bp->b_bcount) {
3097 		/*
3098 		 * We are growing the buffer, possibly in a
3099 		 * byte-granular fashion.
3100 		 */
3101 		struct vnode *vp;
3102 		vm_object_t obj;
3103 		vm_offset_t toff;
3104 		vm_offset_t tinc;
3105 
3106 		/*
3107 		 * Step 1, bring in the VM pages from the object,
3108 		 * allocating them if necessary.  We must clear
3109 		 * B_CACHE if these pages are not valid for the
3110 		 * range covered by the buffer.
3111 		 */
3112 		vp = bp->b_vp;
3113 		obj = vp->v_object;
3114 
3115 		vm_object_hold(obj);
3116 		while (bp->b_xio.xio_npages < desiredpages) {
3117 			vm_page_t m;
3118 			vm_pindex_t pi;
3119 			int error;
3120 
3121 			pi = OFF_TO_IDX(bp->b_loffset) +
3122 			     bp->b_xio.xio_npages;
3123 
3124 			/*
3125 			 * Blocking on m->busy_count might lead to a
3126 			 * deadlock:
3127 			 *
3128 			 *  vm_fault->getpages->cluster_read->allocbuf
3129 			 */
3130 			m = vm_page_lookup_busy_try(obj, pi, FALSE,
3131 						    &error);
3132 			if (error) {
3133 				vm_page_sleep_busy(m, FALSE, "pgtblk");
3134 				continue;
3135 			}
3136 			if (m == NULL) {
3137 				/*
3138 				 * note: must allocate system pages
3139 				 * since blocking here could intefere
3140 				 * with paging I/O, no matter which
3141 				 * process we are.
3142 				 */
3143 				m = bio_page_alloc(bp, obj, pi,
3144 						   desiredpages -
3145 						    bp->b_xio.xio_npages);
3146 				if (m) {
3147 					vm_page_wire(m);
3148 					vm_page_wakeup(m);
3149 					bp->b_flags &= ~B_CACHE;
3150 					bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3151 					++bp->b_xio.xio_npages;
3152 				}
3153 				continue;
3154 			}
3155 
3156 			/*
3157 			 * We found a page and were able to busy it.
3158 			 */
3159 			vm_page_wire(m);
3160 			vm_page_wakeup(m);
3161 			bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3162 			++bp->b_xio.xio_npages;
3163 			if (bp->b_act_count < m->act_count)
3164 				bp->b_act_count = m->act_count;
3165 		}
3166 		vm_object_drop(obj);
3167 
3168 		/*
3169 		 * Step 2.  We've loaded the pages into the buffer,
3170 		 * we have to figure out if we can still have B_CACHE
3171 		 * set.  Note that B_CACHE is set according to the
3172 		 * byte-granular range ( bcount and size ), not the
3173 		 * aligned range ( newbsize ).
3174 		 *
3175 		 * The VM test is against m->valid, which is DEV_BSIZE
3176 		 * aligned.  Needless to say, the validity of the data
3177 		 * needs to also be DEV_BSIZE aligned.  Note that this
3178 		 * fails with NFS if the server or some other client
3179 		 * extends the file's EOF.  If our buffer is resized,
3180 		 * B_CACHE may remain set! XXX
3181 		 */
3182 
3183 		toff = bp->b_bcount;
3184 		tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3185 
3186 		while ((bp->b_flags & B_CACHE) && toff < size) {
3187 			vm_pindex_t pi;
3188 
3189 			if (tinc > (size - toff))
3190 				tinc = size - toff;
3191 
3192 			pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3193 			    PAGE_SHIFT;
3194 
3195 			vfs_buf_test_cache(
3196 			    bp,
3197 			    bp->b_loffset,
3198 			    toff,
3199 			    tinc,
3200 			    bp->b_xio.xio_pages[pi]
3201 			);
3202 			toff += tinc;
3203 			tinc = PAGE_SIZE;
3204 		}
3205 
3206 		/*
3207 		 * Step 3, fixup the KVM pmap.  Remember that
3208 		 * bp->b_data is relative to bp->b_loffset, but
3209 		 * bp->b_loffset may be offset into the first page.
3210 		 */
3211 		bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
3212 		pmap_qenter_noinval((vm_offset_t)bp->b_data,
3213 			    bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3214 		bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3215 				      (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3216 		bkvareset(bp);
3217 	}
3218 	atomic_add_long(&bufspace, newbsize - bp->b_bufsize);
3219 
3220 	/* adjust space use on already-dirty buffer */
3221 	if (bp->b_flags & B_DELWRI) {
3222 		/* dirtykvaspace unchanged */
3223 		atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize);
3224 		if (bp->b_flags & B_HEAVY) {
3225 			atomic_add_long(&dirtybufspacehw,
3226 					newbsize - bp->b_bufsize);
3227 		}
3228 	}
3229 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3230 	bp->b_bcount = size;		/* requested buffer size	*/
3231 	bufspacewakeup();
3232 }
3233 
3234 /*
3235  * biowait:
3236  *
3237  *	Wait for buffer I/O completion, returning error status. B_EINTR
3238  *	is converted into an EINTR error but not cleared (since a chain
3239  *	of biowait() calls may occur).
3240  *
3241  *	On return bpdone() will have been called but the buffer will remain
3242  *	locked and will not have been brelse()'d.
3243  *
3244  *	NOTE!  If a timeout is specified and ETIMEDOUT occurs the I/O is
3245  *	likely still in progress on return.
3246  *
3247  *	NOTE!  This operation is on a BIO, not a BUF.
3248  *
3249  *	NOTE!  BIO_DONE is cleared by vn_strategy()
3250  */
3251 static __inline int
_biowait(struct bio * bio,const char * wmesg,int to)3252 _biowait(struct bio *bio, const char *wmesg, int to)
3253 {
3254 	struct buf *bp = bio->bio_buf;
3255 	u_int32_t flags;
3256 	u_int32_t nflags;
3257 	int error;
3258 
3259 	KKASSERT(bio == &bp->b_bio1);
3260 	for (;;) {
3261 		flags = bio->bio_flags;
3262 		if (flags & BIO_DONE)
3263 			break;
3264 		nflags = flags | BIO_WANT;
3265 		tsleep_interlock(bio, 0);
3266 		if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3267 			if (wmesg)
3268 				error = tsleep(bio, PINTERLOCKED, wmesg, to);
3269 			else if (bp->b_cmd == BUF_CMD_READ)
3270 				error = tsleep(bio, PINTERLOCKED, "biord", to);
3271 			else
3272 				error = tsleep(bio, PINTERLOCKED, "biowr", to);
3273 			if (error) {
3274 				kprintf("tsleep error biowait %d\n", error);
3275 				return (error);
3276 			}
3277 		}
3278 	}
3279 
3280 	/*
3281 	 * Finish up.
3282 	 */
3283 	KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3284 	bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3285 	if (bp->b_flags & B_EINTR)
3286 		return (EINTR);
3287 	if (bp->b_flags & B_ERROR)
3288 		return (bp->b_error ? bp->b_error : EIO);
3289 	return (0);
3290 }
3291 
3292 int
biowait(struct bio * bio,const char * wmesg)3293 biowait(struct bio *bio, const char *wmesg)
3294 {
3295 	return(_biowait(bio, wmesg, 0));
3296 }
3297 
3298 int
biowait_timeout(struct bio * bio,const char * wmesg,int to)3299 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3300 {
3301 	return(_biowait(bio, wmesg, to));
3302 }
3303 
3304 /*
3305  * This associates a tracking count with an I/O.  vn_strategy() and
3306  * dev_dstrategy() do this automatically but there are a few cases
3307  * where a vnode or device layer is bypassed when a block translation
3308  * is cached.  In such cases bio_start_transaction() may be called on
3309  * the bypassed layers so the system gets an I/O in progress indication
3310  * for those higher layers.
3311  */
3312 void
bio_start_transaction(struct bio * bio,struct bio_track * track)3313 bio_start_transaction(struct bio *bio, struct bio_track *track)
3314 {
3315 	bio->bio_track = track;
3316 	bio_track_ref(track);
3317 	dsched_buf_enter(bio->bio_buf);	/* might stack */
3318 }
3319 
3320 /*
3321  * Initiate I/O on a vnode.
3322  *
3323  * SWAPCACHE OPERATION:
3324  *
3325  *	Real buffer cache buffers have a non-NULL bp->b_vp.  Unfortunately
3326  *	devfs also uses b_vp for fake buffers so we also have to check
3327  *	that B_PAGING is 0.  In this case the passed 'vp' is probably the
3328  *	underlying block device.  The swap assignments are related to the
3329  *	buffer cache buffer's b_vp, not the passed vp.
3330  *
3331  *	The passed vp == bp->b_vp only in the case where the strategy call
3332  *	is made on the vp itself for its own buffers (a regular file or
3333  *	block device vp).  The filesystem usually then re-calls vn_strategy()
3334  *	after translating the request to an underlying device.
3335  *
3336  *	Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3337  *	underlying buffer cache buffers.
3338  *
3339  *	We can only deal with page-aligned buffers at the moment, because
3340  *	we can't tell what the real dirty state for pages straddling a buffer
3341  *	are.
3342  *
3343  *	In order to call swap_pager_strategy() we must provide the VM object
3344  *	and base offset for the underlying buffer cache pages so it can find
3345  *	the swap blocks.
3346  */
3347 void
vn_strategy(struct vnode * vp,struct bio * bio)3348 vn_strategy(struct vnode *vp, struct bio *bio)
3349 {
3350 	struct bio_track *track;
3351 	struct buf *bp = bio->bio_buf;
3352 
3353 	KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3354 
3355 	/*
3356 	 * Set when an I/O is issued on the bp.  Cleared by consumers
3357 	 * (aka HAMMER), allowing the consumer to determine if I/O had
3358 	 * actually occurred.
3359 	 */
3360 	bp->b_flags |= B_IOISSUED;
3361 
3362 	/*
3363 	 * Handle the swapcache intercept.
3364 	 *
3365 	 * NOTE: The swapcache itself always supports KVABIO and will
3366 	 *	 do the right thing if its underlying devices do not.
3367 	 */
3368 	if (vn_cache_strategy(vp, bio))
3369 		return;
3370 
3371 	/*
3372 	 * If the vnode does not support KVABIO and the buffer is using
3373 	 * KVABIO, we must synchronize b_data to all cpus before dispatching.
3374 	 */
3375 	if ((vp->v_flag & VKVABIO) == 0 && (bp->b_flags & B_KVABIO))
3376 		bkvasync_all(bp);
3377 
3378 	/*
3379 	 * Otherwise do the operation through the filesystem
3380 	 */
3381         if (bp->b_cmd == BUF_CMD_READ)
3382                 track = &vp->v_track_read;
3383         else
3384                 track = &vp->v_track_write;
3385 	KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3386 	bio->bio_track = track;
3387 	bio_track_ref(track);
3388 	dsched_buf_enter(bp);	/* might stack */
3389         vop_strategy(*vp->v_ops, vp, bio);
3390 }
3391 
3392 /*
3393  * vn_cache_strategy()
3394  *
3395  * Returns 1 if the interrupt was successful, 0 if not.
3396  *
3397  * NOTE: This function supports the KVABIO API wherein b_data might not
3398  *	 be synchronized to the current cpu.
3399  */
3400 static void vn_cache_strategy_callback(struct bio *bio);
3401 
3402 int
vn_cache_strategy(struct vnode * vp,struct bio * bio)3403 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3404 {
3405 	struct buf *bp = bio->bio_buf;
3406 	struct bio *nbio;
3407 	vm_object_t object;
3408 	vm_page_t m;
3409 	int i;
3410 
3411 	/*
3412 	 * Stop using swapcache if paniced, dumping, or dumped
3413 	 */
3414 	if (panicstr || dumping)
3415 		return(0);
3416 
3417 	/*
3418 	 * Is this buffer cache buffer suitable for reading from
3419 	 * the swap cache?
3420 	 */
3421 	if (vm_swapcache_read_enable == 0 ||
3422 	    bp->b_cmd != BUF_CMD_READ ||
3423 	    ((bp->b_flags & B_CLUSTER) == 0 &&
3424 	     (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3425 	    ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3426 	    (bp->b_bcount & PAGE_MASK) != 0) {
3427 		return(0);
3428 	}
3429 
3430 	/*
3431 	 * Figure out the original VM object (it will match the underlying
3432 	 * VM pages).  Note that swap cached data uses page indices relative
3433 	 * to that object, not relative to bio->bio_offset.
3434 	 */
3435 	if (bp->b_flags & B_CLUSTER)
3436 		object = vp->v_object;
3437 	else
3438 		object = bp->b_vp->v_object;
3439 
3440 	/*
3441 	 * In order to be able to use the swap cache all underlying VM
3442 	 * pages must be marked as such, and we can't have any bogus pages.
3443 	 */
3444 	for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3445 		m = bp->b_xio.xio_pages[i];
3446 		if ((m->flags & PG_SWAPPED) == 0)
3447 			break;
3448 		if (m == bogus_page)
3449 			break;
3450 	}
3451 
3452 	/*
3453 	 * If we are good then issue the I/O using swap_pager_strategy().
3454 	 *
3455 	 * We can only do this if the buffer actually supports object-backed
3456 	 * I/O.  If it doesn't npages will be 0.
3457 	 */
3458 	if (i && i == bp->b_xio.xio_npages) {
3459 		m = bp->b_xio.xio_pages[0];
3460 		nbio = push_bio(bio);
3461 		nbio->bio_done = vn_cache_strategy_callback;
3462 		nbio->bio_offset = ptoa(m->pindex);
3463 		KKASSERT(m->object == object);
3464 		swap_pager_strategy(object, nbio);
3465 		return(1);
3466 	}
3467 	return(0);
3468 }
3469 
3470 /*
3471  * This is a bit of a hack but since the vn_cache_strategy() function can
3472  * override a VFS's strategy function we must make sure that the bio, which
3473  * is probably bio2, doesn't leak an unexpected offset value back to the
3474  * filesystem.  The filesystem (e.g. UFS) might otherwise assume that the
3475  * bio went through its own file strategy function and the the bio2 offset
3476  * is a cached disk offset when, in fact, it isn't.
3477  */
3478 static void
vn_cache_strategy_callback(struct bio * bio)3479 vn_cache_strategy_callback(struct bio *bio)
3480 {
3481 	bio->bio_offset = NOOFFSET;
3482 	biodone(pop_bio(bio));
3483 }
3484 
3485 /*
3486  * bpdone:
3487  *
3488  *	Finish I/O on a buffer after all BIOs have been processed.
3489  *	Called when the bio chain is exhausted or by biowait.  If called
3490  *	by biowait, elseit is typically 0.
3491  *
3492  *	bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3493  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3494  *	assuming B_INVAL is clear.
3495  *
3496  *	For the VMIO case, we set B_CACHE if the op was a read and no
3497  *	read error occured, or if the op was a write.  B_CACHE is never
3498  *	set if the buffer is invalid or otherwise uncacheable.
3499  *
3500  *	bpdone does not mess with B_INVAL, allowing the I/O routine or the
3501  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3502  *	in the biodone routine.
3503  *
3504  *	bpdone is responsible for calling bundirty() on the buffer after a
3505  *	successful write.  We previously did this prior to initiating the
3506  *	write under the assumption that the buffer might be dirtied again
3507  *	while the write was in progress, however doing it before-hand creates
3508  *	a race condition prior to the call to vn_strategy() where the
3509  *	filesystem may not be aware that a dirty buffer is present.
3510  *	It should not be possible for the buffer or its underlying pages to
3511  *	be redirtied prior to bpdone()'s unbusying of the underlying VM
3512  *	pages.
3513  */
3514 void
bpdone(struct buf * bp,int elseit)3515 bpdone(struct buf *bp, int elseit)
3516 {
3517 	buf_cmd_t cmd;
3518 
3519 	KASSERT(BUF_LOCKINUSE(bp), ("bpdone: bp %p not busy", bp));
3520 	KASSERT(bp->b_cmd != BUF_CMD_DONE,
3521 		("bpdone: bp %p already done!", bp));
3522 
3523 	/*
3524 	 * No more BIOs are left.  All completion functions have been dealt
3525 	 * with, now we clean up the buffer.
3526 	 */
3527 	cmd = bp->b_cmd;
3528 	bp->b_cmd = BUF_CMD_DONE;
3529 
3530 	/*
3531 	 * Only reads and writes are processed past this point.
3532 	 */
3533 	if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3534 		if (cmd == BUF_CMD_FREEBLKS)
3535 			bp->b_flags |= B_NOCACHE;
3536 		if (elseit)
3537 			brelse(bp);
3538 		return;
3539 	}
3540 
3541 	/*
3542 	 * A failed write must re-dirty the buffer unless B_INVAL
3543 	 * was set.
3544 	 *
3545 	 * A successful write must clear the dirty flag.  This is done after
3546 	 * the write to ensure that the buffer remains on the vnode's dirty
3547 	 * list for filesystem interlocks / checks until the write is actually
3548 	 * complete.  HAMMER2 is sensitive to this issue.
3549 	 *
3550 	 * Only applicable to normal buffers (with VPs).  vinum buffers may
3551 	 * not have a vp.
3552 	 *
3553 	 * Must be done prior to calling buf_complete() as the callback might
3554 	 * re-dirty the buffer.
3555 	 */
3556 	if (cmd == BUF_CMD_WRITE) {
3557 		if ((bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3558 			bp->b_flags &= ~B_NOCACHE;
3559 			if (bp->b_vp)
3560 				bdirty(bp);
3561 		} else {
3562 			if (bp->b_vp)
3563 				bundirty(bp);
3564 		}
3565 	}
3566 
3567 	/*
3568 	 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3569 	 * a lot worse.  XXX - move this above the clearing of b_cmd
3570 	 */
3571 	if (LIST_FIRST(&bp->b_dep) != NULL)
3572 		buf_complete(bp);
3573 
3574 	if (bp->b_flags & B_VMIO) {
3575 		int i;
3576 		vm_ooffset_t foff;
3577 		vm_page_t m;
3578 		vm_object_t obj;
3579 		int iosize;
3580 		struct vnode *vp = bp->b_vp;
3581 
3582 		obj = vp->v_object;
3583 
3584 #if defined(VFS_BIO_DEBUG)
3585 		if (vp->v_auxrefs == 0)
3586 			panic("bpdone: zero vnode hold count");
3587 		if ((vp->v_flag & VOBJBUF) == 0)
3588 			panic("bpdone: vnode is not setup for merged cache");
3589 #endif
3590 
3591 		foff = bp->b_loffset;
3592 		KASSERT(foff != NOOFFSET, ("bpdone: no buffer offset"));
3593 		KASSERT(obj != NULL, ("bpdone: missing VM object"));
3594 
3595 #if defined(VFS_BIO_DEBUG)
3596 		if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3597 			kprintf("bpdone: paging in progress(%d) < "
3598 				"bp->b_xio.xio_npages(%d)\n",
3599 				obj->paging_in_progress,
3600 				bp->b_xio.xio_npages);
3601 		}
3602 #endif
3603 
3604 		/*
3605 		 * Set B_CACHE if the op was a normal read and no error
3606 		 * occured.  B_CACHE is set for writes in the b*write()
3607 		 * routines.
3608 		 */
3609 		iosize = bp->b_bcount - bp->b_resid;
3610 		if (cmd == BUF_CMD_READ &&
3611 		    (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3612 			bp->b_flags |= B_CACHE;
3613 		}
3614 
3615 		vm_object_hold(obj);
3616 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3617 			int resid;
3618 			int isbogus;
3619 
3620 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3621 			if (resid > iosize)
3622 				resid = iosize;
3623 
3624 			/*
3625 			 * cleanup bogus pages, restoring the originals.  Since
3626 			 * the originals should still be wired, we don't have
3627 			 * to worry about interrupt/freeing races destroying
3628 			 * the VM object association.
3629 			 */
3630 			m = bp->b_xio.xio_pages[i];
3631 			if (m == bogus_page) {
3632 				if ((bp->b_flags & B_HASBOGUS) == 0)
3633 					panic("bpdone: bp %p corrupt bogus", bp);
3634 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3635 				if (m == NULL)
3636 					panic("bpdone: page disappeared");
3637 				bp->b_xio.xio_pages[i] = m;
3638 				isbogus = 1;
3639 			} else {
3640 				isbogus = 0;
3641 			}
3642 #if defined(VFS_BIO_DEBUG)
3643 			if (OFF_TO_IDX(foff) != m->pindex) {
3644 				kprintf("bpdone: foff(%lu)/m->pindex(%ld) "
3645 					"mismatch\n",
3646 					(unsigned long)foff, (long)m->pindex);
3647 			}
3648 #endif
3649 
3650 			/*
3651 			 * In the write case, the valid and clean bits are
3652 			 * already changed correctly (see bdwrite()), so we
3653 			 * only need to do this here in the read case.
3654 			 */
3655 			vm_page_busy_wait(m, FALSE, "bpdpgw");
3656 			if (cmd == BUF_CMD_READ && isbogus == 0 && resid > 0)
3657 				vfs_clean_one_page(bp, i, m);
3658 
3659 			/*
3660 			 * when debugging new filesystems or buffer I/O
3661 			 * methods, this is the most common error that pops
3662 			 * up.  if you see this, you have not set the page
3663 			 * busy flag correctly!!!
3664 			 */
3665 			if ((m->busy_count & PBUSY_MASK) == 0) {
3666 				kprintf("bpdone: page busy < 0, "
3667 				    "pindex: %d, foff: 0x(%x,%x), "
3668 				    "resid: %d, index: %d\n",
3669 				    (int) m->pindex, (int)(foff >> 32),
3670 						(int) foff & 0xffffffff, resid, i);
3671 				if (!vn_isdisk(vp, NULL))
3672 					kprintf(" iosize: %ld, loffset: %lld, "
3673 						"flags: 0x%08x, npages: %d\n",
3674 					    bp->b_vp->v_mount->mnt_stat.f_iosize,
3675 					    (long long)bp->b_loffset,
3676 					    bp->b_flags, bp->b_xio.xio_npages);
3677 				else
3678 					kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3679 					    (long long)bp->b_loffset,
3680 					    bp->b_flags, bp->b_xio.xio_npages);
3681 				kprintf(" valid: 0x%x, dirty: 0x%x, "
3682 					"wired: %d\n",
3683 					m->valid, m->dirty,
3684 					m->wire_count);
3685 				panic("bpdone: page busy < 0");
3686 			}
3687 			vm_page_io_finish(m);
3688 			vm_page_wakeup(m);
3689 			vm_object_pip_wakeup(obj);
3690 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3691 			iosize -= resid;
3692 		}
3693 		if (bp->b_flags & B_HASBOGUS) {
3694 			pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
3695 					    bp->b_xio.xio_pages,
3696 					    bp->b_xio.xio_npages);
3697 			bp->b_flags &= ~B_HASBOGUS;
3698 			bkvareset(bp);
3699 		}
3700 		vm_object_drop(obj);
3701 	}
3702 
3703 	/*
3704 	 * Finish up by releasing the buffer.  There are no more synchronous
3705 	 * or asynchronous completions, those were handled by bio_done
3706 	 * callbacks.
3707 	 */
3708 	if (elseit) {
3709 		if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3710 			brelse(bp);
3711 		else
3712 			bqrelse(bp);
3713 	}
3714 }
3715 
3716 /*
3717  * Normal biodone.
3718  */
3719 void
biodone(struct bio * bio)3720 biodone(struct bio *bio)
3721 {
3722 	struct buf *bp = bio->bio_buf;
3723 
3724 	runningbufwakeup(bp);
3725 
3726 	/*
3727 	 * Run up the chain of BIO's.   Leave b_cmd intact for the duration.
3728 	 */
3729 	while (bio) {
3730 		biodone_t *done_func;
3731 		struct bio_track *track;
3732 
3733 		/*
3734 		 * BIO tracking.  Most but not all BIOs are tracked.
3735 		 */
3736 		if ((track = bio->bio_track) != NULL) {
3737 			bio_track_rel(track);
3738 			bio->bio_track = NULL;
3739 		}
3740 
3741 		/*
3742 		 * A bio_done function terminates the loop.  The function
3743 		 * will be responsible for any further chaining and/or
3744 		 * buffer management.
3745 		 *
3746 		 * WARNING!  The done function can deallocate the buffer!
3747 		 */
3748 		if ((done_func = bio->bio_done) != NULL) {
3749 			bio->bio_done = NULL;
3750 			done_func(bio);
3751 			return;
3752 		}
3753 		bio = bio->bio_prev;
3754 	}
3755 
3756 	/*
3757 	 * If we've run out of bio's do normal [a]synchronous completion.
3758 	 */
3759 	bpdone(bp, 1);
3760 }
3761 
3762 /*
3763  * Synchronous biodone - this terminates a synchronous BIO.
3764  *
3765  * bpdone() is called with elseit=FALSE, leaving the buffer completed
3766  * but still locked.  The caller must brelse() the buffer after waiting
3767  * for completion.
3768  */
3769 void
biodone_sync(struct bio * bio)3770 biodone_sync(struct bio *bio)
3771 {
3772 	struct buf *bp = bio->bio_buf;
3773 	int flags;
3774 	int nflags;
3775 
3776 	KKASSERT(bio == &bp->b_bio1);
3777 	bpdone(bp, 0);
3778 
3779 	for (;;) {
3780 		flags = bio->bio_flags;
3781 		nflags = (flags | BIO_DONE) & ~BIO_WANT;
3782 
3783 		if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3784 			if (flags & BIO_WANT)
3785 				wakeup(bio);
3786 			break;
3787 		}
3788 	}
3789 }
3790 
3791 /*
3792  * vfs_unbusy_pages:
3793  *
3794  *	This routine is called in lieu of iodone in the case of
3795  *	incomplete I/O.  This keeps the busy status for pages
3796  *	consistant.
3797  */
3798 void
vfs_unbusy_pages(struct buf * bp)3799 vfs_unbusy_pages(struct buf *bp)
3800 {
3801 	int i;
3802 
3803 	runningbufwakeup(bp);
3804 
3805 	if (bp->b_flags & B_VMIO) {
3806 		struct vnode *vp = bp->b_vp;
3807 		vm_object_t obj;
3808 
3809 		obj = vp->v_object;
3810 		vm_object_hold(obj);
3811 
3812 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3813 			vm_page_t m = bp->b_xio.xio_pages[i];
3814 
3815 			/*
3816 			 * When restoring bogus changes the original pages
3817 			 * should still be wired, so we are in no danger of
3818 			 * losing the object association and do not need
3819 			 * critical section protection particularly.
3820 			 */
3821 			if (m == bogus_page) {
3822 				m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3823 				if (!m) {
3824 					panic("vfs_unbusy_pages: page missing");
3825 				}
3826 				bp->b_xio.xio_pages[i] = m;
3827 			}
3828 			vm_page_busy_wait(m, FALSE, "bpdpgw");
3829 			vm_page_io_finish(m);
3830 			vm_page_wakeup(m);
3831 			vm_object_pip_wakeup(obj);
3832 		}
3833 		if (bp->b_flags & B_HASBOGUS) {
3834 			pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
3835 					    bp->b_xio.xio_pages,
3836 					    bp->b_xio.xio_npages);
3837 			bp->b_flags &= ~B_HASBOGUS;
3838 			bkvareset(bp);
3839 		}
3840 		vm_object_drop(obj);
3841 	}
3842 }
3843 
3844 /*
3845  * vfs_busy_pages:
3846  *
3847  *	This routine is called before a device strategy routine.
3848  *	It is used to tell the VM system that paging I/O is in
3849  *	progress, and treat the pages associated with the buffer
3850  *	almost as being PBUSY_LOCKED.  Also the object 'paging_in_progress'
3851  *	flag is handled to make sure that the object doesn't become
3852  *	inconsistant.
3853  *
3854  *	Since I/O has not been initiated yet, certain buffer flags
3855  *	such as B_ERROR or B_INVAL may be in an inconsistant state
3856  *	and should be ignored.
3857  */
3858 void
vfs_busy_pages(struct vnode * vp,struct buf * bp)3859 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3860 {
3861 	int i, bogus;
3862 	struct lwp *lp = curthread->td_lwp;
3863 
3864 	/*
3865 	 * The buffer's I/O command must already be set.  If reading,
3866 	 * B_CACHE must be 0 (double check against callers only doing
3867 	 * I/O when B_CACHE is 0).
3868 	 */
3869 	KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3870 	KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3871 
3872 	if (bp->b_flags & B_VMIO) {
3873 		vm_object_t obj;
3874 
3875 		obj = vp->v_object;
3876 		KASSERT(bp->b_loffset != NOOFFSET,
3877 			("vfs_busy_pages: no buffer offset"));
3878 
3879 		/*
3880 		 * Busy all the pages.  We have to busy them all at once
3881 		 * to avoid deadlocks.
3882 		 */
3883 retry:
3884 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3885 			vm_page_t m = bp->b_xio.xio_pages[i];
3886 
3887 			if (vm_page_busy_try(m, FALSE)) {
3888 				vm_page_sleep_busy(m, FALSE, "vbpage");
3889 				while (--i >= 0)
3890 					vm_page_wakeup(bp->b_xio.xio_pages[i]);
3891 				goto retry;
3892 			}
3893 		}
3894 
3895 		/*
3896 		 * Setup for I/O, soft-busy the page right now because
3897 		 * the next loop may block.
3898 		 */
3899 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3900 			vm_page_t m = bp->b_xio.xio_pages[i];
3901 
3902 			if ((bp->b_flags & B_CLUSTER) == 0) {
3903 				vm_object_pip_add(obj, 1);
3904 				vm_page_io_start(m);
3905 			}
3906 		}
3907 
3908 		/*
3909 		 * Adjust protections for I/O and do bogus-page mapping.
3910 		 * Assume that vm_page_protect() can block (it can block
3911 		 * if VM_PROT_NONE, don't take any chances regardless).
3912 		 *
3913 		 * In particular note that for writes we must incorporate
3914 		 * page dirtyness from the VM system into the buffer's
3915 		 * dirty range.
3916 		 *
3917 		 * For reads we theoretically must incorporate page dirtyness
3918 		 * from the VM system to determine if the page needs bogus
3919 		 * replacement, but we shortcut the test by simply checking
3920 		 * that all m->valid bits are set, indicating that the page
3921 		 * is fully valid and does not need to be re-read.  For any
3922 		 * VM system dirtyness the page will also be fully valid
3923 		 * since it was mapped at one point.
3924 		 */
3925 		bogus = 0;
3926 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3927 			vm_page_t m = bp->b_xio.xio_pages[i];
3928 
3929 			if (bp->b_cmd == BUF_CMD_WRITE) {
3930 				/*
3931 				 * When readying a vnode-backed buffer for
3932 				 * a write we must zero-fill any invalid
3933 				 * portions of the backing VM pages, mark
3934 				 * it valid and clear related dirty bits.
3935 				 *
3936 				 * vfs_clean_one_page() incorporates any
3937 				 * VM dirtyness and updates the b_dirtyoff
3938 				 * range (after we've made the page RO).
3939 				 *
3940 				 * It is also expected that the pmap modified
3941 				 * bit has already been cleared by the
3942 				 * vm_page_protect().  We may not be able
3943 				 * to clear all dirty bits for a page if it
3944 				 * was also memory mapped (NFS).
3945 				 *
3946 				 * Finally be sure to unassign any swap-cache
3947 				 * backing store as it is now stale.
3948 				 */
3949 				vm_page_protect(m, VM_PROT_READ);
3950 				vfs_clean_one_page(bp, i, m);
3951 				swap_pager_unswapped(m);
3952 			} else if (m->valid == VM_PAGE_BITS_ALL) {
3953 				/*
3954 				 * When readying a vnode-backed buffer for
3955 				 * read we must replace any dirty pages with
3956 				 * a bogus page so dirty data is not destroyed
3957 				 * when filling gaps.
3958 				 *
3959 				 * To avoid testing whether the page is
3960 				 * dirty we instead test that the page was
3961 				 * at some point mapped (m->valid fully
3962 				 * valid) with the understanding that
3963 				 * this also covers the dirty case.
3964 				 */
3965 				bp->b_xio.xio_pages[i] = bogus_page;
3966 				bp->b_flags |= B_HASBOGUS;
3967 				bogus++;
3968 			} else if (m->valid & m->dirty) {
3969 				/*
3970 				 * This case should not occur as partial
3971 				 * dirtyment can only happen if the buffer
3972 				 * is B_CACHE, and this code is not entered
3973 				 * if the buffer is B_CACHE.
3974 				 */
3975 				kprintf("Warning: vfs_busy_pages - page not "
3976 					"fully valid! loff=%jx bpf=%08x "
3977 					"idx=%d val=%02x dir=%02x\n",
3978 					(uintmax_t)bp->b_loffset, bp->b_flags,
3979 					i, m->valid, m->dirty);
3980 				vm_page_protect(m, VM_PROT_NONE);
3981 			} else {
3982 				/*
3983 				 * The page is not valid and can be made
3984 				 * part of the read.
3985 				 */
3986 				vm_page_protect(m, VM_PROT_NONE);
3987 			}
3988 			vm_page_wakeup(m);
3989 		}
3990 		if (bogus) {
3991 			pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
3992 					    bp->b_xio.xio_pages,
3993 					    bp->b_xio.xio_npages);
3994 			bkvareset(bp);
3995 		}
3996 	}
3997 
3998 	/*
3999 	 * This is the easiest place to put the process accounting for the I/O
4000 	 * for now.
4001 	 */
4002 	if (lp != NULL) {
4003 		if (bp->b_cmd == BUF_CMD_READ)
4004 			lp->lwp_ru.ru_inblock++;
4005 		else
4006 			lp->lwp_ru.ru_oublock++;
4007 	}
4008 }
4009 
4010 /*
4011  * Tell the VM system that the pages associated with this buffer
4012  * are clean.  This is used for delayed writes where the data is
4013  * going to go to disk eventually without additional VM intevention.
4014  *
4015  * NOTE: While we only really need to clean through to b_bcount, we
4016  *	 just go ahead and clean through to b_bufsize.
4017  */
4018 static void
vfs_clean_pages(struct buf * bp)4019 vfs_clean_pages(struct buf *bp)
4020 {
4021 	vm_page_t m;
4022 	int i;
4023 
4024 	if ((bp->b_flags & B_VMIO) == 0)
4025 		return;
4026 
4027 	KASSERT(bp->b_loffset != NOOFFSET,
4028 		("vfs_clean_pages: no buffer offset"));
4029 
4030 	for (i = 0; i < bp->b_xio.xio_npages; i++) {
4031 		m = bp->b_xio.xio_pages[i];
4032 		vfs_clean_one_page(bp, i, m);
4033 	}
4034 }
4035 
4036 /*
4037  * vfs_clean_one_page:
4038  *
4039  *	Set the valid bits and clear the dirty bits in a page within a
4040  *	buffer.  The range is restricted to the buffer's size and the
4041  *	buffer's logical offset might index into the first page.
4042  *
4043  *	The caller has busied or soft-busied the page and it is not mapped,
4044  *	test and incorporate the dirty bits into b_dirtyoff/end before
4045  *	clearing them.  Note that we need to clear the pmap modified bits
4046  *	after determining the the page was dirty, vm_page_set_validclean()
4047  *	does not do it for us.
4048  *
4049  *	This routine is typically called after a read completes (dirty should
4050  *	be zero in that case as we are not called on bogus-replace pages),
4051  *	or before a write is initiated.
4052  */
4053 static void
vfs_clean_one_page(struct buf * bp,int pageno,vm_page_t m)4054 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4055 {
4056 	int bcount;
4057 	int xoff;
4058 	int soff;
4059 	int eoff;
4060 
4061 	/*
4062 	 * Calculate offset range within the page but relative to buffer's
4063 	 * loffset.  loffset might be offset into the first page.
4064 	 */
4065 	xoff = (int)bp->b_loffset & PAGE_MASK;	/* loffset offset into pg 0 */
4066 	bcount = bp->b_bcount + xoff;		/* offset adjusted */
4067 
4068 	if (pageno == 0) {
4069 		soff = xoff;
4070 		eoff = PAGE_SIZE;
4071 	} else {
4072 		soff = (pageno << PAGE_SHIFT);
4073 		eoff = soff + PAGE_SIZE;
4074 	}
4075 	if (eoff > bcount)
4076 		eoff = bcount;
4077 	if (soff >= eoff)
4078 		return;
4079 
4080 	/*
4081 	 * Test dirty bits and adjust b_dirtyoff/end.
4082 	 *
4083 	 * If dirty pages are incorporated into the bp any prior
4084 	 * B_NEEDCOMMIT state (NFS) must be cleared because the
4085 	 * caller has not taken into account the new dirty data.
4086 	 *
4087 	 * If the page was memory mapped the dirty bits might go beyond the
4088 	 * end of the buffer, but we can't really make the assumption that
4089 	 * a file EOF straddles the buffer (even though this is the case for
4090 	 * NFS if B_NEEDCOMMIT is also set).  So for the purposes of clearing
4091 	 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4092 	 * This also saves some console spam.
4093 	 *
4094 	 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4095 	 * NFS can handle huge commits but not huge writes.
4096 	 */
4097 	vm_page_test_dirty(m);
4098 	if (m->dirty) {
4099 		if ((bp->b_flags & B_NEEDCOMMIT) &&
4100 		    (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4101 			if (debug_commit)
4102 				kprintf("Warning: vfs_clean_one_page: bp %p "
4103 				    "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4104 				    " cmd %d vd %02x/%02x x/s/e %d %d %d "
4105 				    "doff/end %d %d\n",
4106 				    bp, (uintmax_t)bp->b_loffset, bp->b_bcount,
4107 				    bp->b_flags, bp->b_cmd,
4108 				    m->valid, m->dirty, xoff, soff, eoff,
4109 				    bp->b_dirtyoff, bp->b_dirtyend);
4110 			bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4111 			if (debug_commit)
4112 				print_backtrace(-1);
4113 		}
4114 		/*
4115 		 * Only clear the pmap modified bits if ALL the dirty bits
4116 		 * are set, otherwise the system might mis-clear portions
4117 		 * of a page.
4118 		 */
4119 		if (m->dirty == VM_PAGE_BITS_ALL &&
4120 		    (bp->b_flags & B_NEEDCOMMIT) == 0) {
4121 			pmap_clear_modify(m);
4122 		}
4123 		if (bp->b_dirtyoff > soff - xoff)
4124 			bp->b_dirtyoff = soff - xoff;
4125 		if (bp->b_dirtyend < eoff - xoff)
4126 			bp->b_dirtyend = eoff - xoff;
4127 	}
4128 
4129 	/*
4130 	 * Set related valid bits, clear related dirty bits.
4131 	 * Does not mess with the pmap modified bit.
4132 	 *
4133 	 * WARNING!  We cannot just clear all of m->dirty here as the
4134 	 *	     buffer cache buffers may use a DEV_BSIZE'd aligned
4135 	 *	     block size, or have an odd size (e.g. NFS at file EOF).
4136 	 *	     The putpages code can clear m->dirty to 0.
4137 	 *
4138 	 *	     If a VOP_WRITE generates a buffer cache buffer which
4139 	 *	     covers the same space as mapped writable pages the
4140 	 *	     buffer flush might not be able to clear all the dirty
4141 	 *	     bits and still require a putpages from the VM system
4142 	 *	     to finish it off.
4143 	 *
4144 	 * WARNING!  vm_page_set_validclean() currently assumes vm_token
4145 	 *	     is held.  The page might not be busied (bdwrite() case).
4146 	 *	     XXX remove this comment once we've validated that this
4147 	 *	     is no longer an issue.
4148 	 */
4149 	vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4150 }
4151 
4152 #if 0
4153 /*
4154  * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4155  * The page data is assumed to be valid (there is no zeroing here).
4156  */
4157 static void
4158 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4159 {
4160 	int bcount;
4161 	int xoff;
4162 	int soff;
4163 	int eoff;
4164 
4165 	/*
4166 	 * Calculate offset range within the page but relative to buffer's
4167 	 * loffset.  loffset might be offset into the first page.
4168 	 */
4169 	xoff = (int)bp->b_loffset & PAGE_MASK;	/* loffset offset into pg 0 */
4170 	bcount = bp->b_bcount + xoff;		/* offset adjusted */
4171 
4172 	if (pageno == 0) {
4173 		soff = xoff;
4174 		eoff = PAGE_SIZE;
4175 	} else {
4176 		soff = (pageno << PAGE_SHIFT);
4177 		eoff = soff + PAGE_SIZE;
4178 	}
4179 	if (eoff > bcount)
4180 		eoff = bcount;
4181 	if (soff >= eoff)
4182 		return;
4183 	vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4184 }
4185 #endif
4186 
4187 /*
4188  * vfs_bio_clrbuf:
4189  *
4190  *	Clear a buffer.  This routine essentially fakes an I/O, so we need
4191  *	to clear B_ERROR and B_INVAL.
4192  *
4193  *	Note that while we only theoretically need to clear through b_bcount,
4194  *	we go ahead and clear through b_bufsize.
4195  */
4196 void
vfs_bio_clrbuf(struct buf * bp)4197 vfs_bio_clrbuf(struct buf *bp)
4198 {
4199 	int i, mask = 0;
4200 	caddr_t sa, ea;
4201 	KKASSERT(bp->b_flags & B_VMIO);
4202 
4203 	bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4204 	bkvasync(bp);
4205 
4206 	if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4207 	    (bp->b_loffset & PAGE_MASK) == 0) {
4208 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4209 		if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4210 			bp->b_resid = 0;
4211 			return;
4212 		}
4213 		if ((bp->b_xio.xio_pages[0]->valid & mask) == 0) {
4214 			bzero(bp->b_data, bp->b_bufsize);
4215 			bp->b_xio.xio_pages[0]->valid |= mask;
4216 			bp->b_resid = 0;
4217 			return;
4218 		}
4219 	}
4220 	sa = bp->b_data;
4221 	for(i = 0; i < bp->b_xio.xio_npages; i++, sa=ea) {
4222 		int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4223 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4224 		ea = (caddr_t)(vm_offset_t)ulmin(
4225 			    (u_long)(vm_offset_t)ea,
4226 			    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4227 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4228 		if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4229 			continue;
4230 		if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4231 			bzero(sa, ea - sa);
4232 		} else {
4233 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4234 				if ((bp->b_xio.xio_pages[i]->valid &
4235 				    (1<<j)) == 0) {
4236 					bzero(sa, DEV_BSIZE);
4237 				}
4238 			}
4239 		}
4240 		bp->b_xio.xio_pages[i]->valid |= mask;
4241 	}
4242 	bp->b_resid = 0;
4243 }
4244 
4245 /*
4246  * Allocate a page for a buffer cache buffer.
4247  *
4248  * If NULL is returned the caller is expected to retry (typically check if
4249  * the page already exists on retry before trying to allocate one).
4250  *
4251  * NOTE! Low-memory handling is dealt with in b[q]relse(), not here.  This
4252  *	 function will use the system reserve with the hope that the page
4253  *	 allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4254  *	 is done with the buffer.
4255  *
4256  * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4257  *	 to TMPFS doesn't clean the page.  For TMPFS, only the pagedaemon
4258  *	 is capable of retiring pages (to swap).  For TMPFS we don't dig
4259  *	 into the system reserve because doing so could stall out pretty
4260  *	 much every process running on the system.
4261  */
4262 static
4263 vm_page_t
bio_page_alloc(struct buf * bp,vm_object_t obj,vm_pindex_t pg,int deficit)4264 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit)
4265 {
4266 	int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK;
4267 	vm_page_t p;
4268 
4269 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4270 
4271 	/*
4272 	 * Avoid localized page-queue exhaustion by rotating the effective
4273 	 * cpu-base for the BIO page allocation.  Remember we are trying to
4274 	 * avoid contention, so we want all the cpus to be in lockstep with
4275 	 * different cpuids.  Really serious contention in the kernel page
4276 	 * allocator can occur without this.
4277 	 *
4278 	 * This is kinda anti-NUMA, but localizing file data is a really hard
4279 	 * call.  It works great in some situations (temporary files in tmpfs),
4280 	 * and horribly in other situations.
4281 	 *
4282 	 * XXX add some NUMA relocalization (2 zones or 4 zones).
4283 	 */
4284 	vmflags |= VM_ALLOC_CPU((mycpu->gd_cpuid + (u_short)ticks) % ncpus);
4285 
4286 	/*
4287 	 * Try a normal allocation first.
4288 	 */
4289 	p = vm_page_alloc(obj, pg, vmflags);
4290 	if (p)
4291 		return(p);
4292 	if (vm_page_lookup(obj, pg))
4293 		return(NULL);
4294 	vm_pageout_deficit += deficit;
4295 
4296 	/*
4297 	 * Try again, digging into the system reserve.
4298 	 *
4299 	 * Trying to recover pages from the buffer cache here can deadlock
4300 	 * against other threads trying to busy underlying pages so we
4301 	 * depend on the code in brelse() and bqrelse() to free/cache the
4302 	 * underlying buffer cache pages when memory is low.
4303 	 */
4304 	if (curthread->td_flags & TDF_SYSTHREAD)
4305 		vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
4306 	else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS)
4307 		vmflags |= 0;
4308 	else
4309 		vmflags |= VM_ALLOC_SYSTEM;
4310 
4311 	/*recoverbufpages();*/
4312 	p = vm_page_alloc(obj, pg, vmflags);
4313 	if (p)
4314 		return(p);
4315 	if (vm_page_lookup(obj, pg))
4316 		return(NULL);
4317 
4318 	/*
4319 	 * Wait for memory to free up and try again
4320 	 */
4321 	if (vm_paging_severe())
4322 		++lowmempgallocs;
4323 	vm_wait(hz / 20 + 1);
4324 
4325 	p = vm_page_alloc(obj, pg, vmflags);
4326 	if (p)
4327 		return(p);
4328 	if (vm_page_lookup(obj, pg))
4329 		return(NULL);
4330 
4331 	/*
4332 	 * Ok, now we are really in trouble.
4333 	 */
4334 	if (bootverbose) {
4335 		static struct krate biokrate = { .freq = 1 };
4336 		krateprintf(&biokrate,
4337 			    "Warning: bio_page_alloc: memory exhausted "
4338 			    "during buffer cache page allocation from %s\n",
4339 			    curthread->td_comm);
4340 	}
4341 	if (curthread->td_flags & TDF_SYSTHREAD)
4342 		vm_wait(hz / 20 + 1);
4343 	else
4344 		vm_wait(hz / 2 + 1);
4345 	return (NULL);
4346 }
4347 
4348 /*
4349  * The buffer's mapping has changed.  Adjust the buffer's memory
4350  * synchronization.  The caller is the exclusive holder of the buffer
4351  * and has set or cleared B_KVABIO according to preference.
4352  *
4353  * WARNING! If the caller is using B_KVABIO mode, this function will
4354  *	    not map the data to the current cpu.  The caller must also
4355  *	    call bkvasync(bp).
4356  */
4357 void
bkvareset(struct buf * bp)4358 bkvareset(struct buf *bp)
4359 {
4360 	if (bp->b_flags & B_KVABIO) {
4361 		CPUMASK_ASSZERO(bp->b_cpumask);
4362 	} else {
4363 		CPUMASK_ORMASK(bp->b_cpumask, smp_active_mask);
4364 		smp_invltlb();
4365 		cpu_invltlb();
4366 	}
4367 }
4368 
4369 /*
4370  * The buffer will be used by the caller on the caller's cpu, synchronize
4371  * its data to the current cpu.  Caller must control the buffer by holding
4372  * its lock, but calling cpu does not necessarily have to be the owner of
4373  * the lock (i.e. HAMMER2's concurrent I/O accessors).
4374  *
4375  * If B_KVABIO is not set, the buffer is already fully synchronized.
4376  */
4377 void
bkvasync(struct buf * bp)4378 bkvasync(struct buf *bp)
4379 {
4380 	int cpuid = mycpu->gd_cpuid;
4381 	char *bdata;
4382 
4383 	if ((bp->b_flags & B_KVABIO) &&
4384 	    CPUMASK_TESTBIT(bp->b_cpumask, cpuid) == 0) {
4385 		bdata = bp->b_data;
4386 		while (bdata < bp->b_data + bp->b_bufsize) {
4387 			cpu_invlpg(bdata);
4388 			bdata += PAGE_SIZE -
4389 				 ((intptr_t)bdata & PAGE_MASK);
4390 		}
4391 		ATOMIC_CPUMASK_ORBIT(bp->b_cpumask, cpuid);
4392 	}
4393 }
4394 
4395 /*
4396  * The buffer will be used by a subsystem that does not understand
4397  * the KVABIO API.  Make sure its data is synchronized to all cpus.
4398  *
4399  * If B_KVABIO is not set, the buffer is already fully synchronized.
4400  *
4401  * NOTE! This is the only safe way to clear B_KVABIO on a buffer.
4402  */
4403 void
bkvasync_all(struct buf * bp)4404 bkvasync_all(struct buf *bp)
4405 {
4406 	if (debug_kvabio > 0) {
4407 		--debug_kvabio;
4408 		print_backtrace(10);
4409 	}
4410 
4411 	if ((bp->b_flags & B_KVABIO) &&
4412 	    CPUMASK_CMPMASKNEQ(bp->b_cpumask, smp_active_mask)) {
4413 		smp_invltlb();
4414 		cpu_invltlb();
4415 		ATOMIC_CPUMASK_ORMASK(bp->b_cpumask, smp_active_mask);
4416 	}
4417 	bp->b_flags &= ~B_KVABIO;
4418 }
4419 
4420 /*
4421  * Scan all buffers in the system and issue the callback.
4422  */
4423 int
scan_all_buffers(int (* callback)(struct buf *,void *),void * info)4424 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4425 {
4426 	int count = 0;
4427 	int error;
4428 	long n;
4429 
4430 	for (n = 0; n < nbuf; ++n) {
4431 		if ((error = callback(&buf[n], info)) < 0) {
4432 			count = error;
4433 			break;
4434 		}
4435 		count += error;
4436 	}
4437 	return (count);
4438 }
4439 
4440 /*
4441  * nestiobuf_iodone: biodone callback for nested buffers and propagate
4442  * completion to the master buffer.
4443  */
4444 static void
nestiobuf_iodone(struct bio * bio)4445 nestiobuf_iodone(struct bio *bio)
4446 {
4447 	struct bio *mbio;
4448 	struct buf *mbp, *bp;
4449 	struct devstat *stats;
4450 	int error;
4451 	int donebytes;
4452 
4453 	bp = bio->bio_buf;
4454 	mbio = bio->bio_caller_info1.ptr;
4455 	stats = bio->bio_caller_info2.ptr;
4456 	mbp = mbio->bio_buf;
4457 
4458 	KKASSERT(bp->b_bcount <= bp->b_bufsize);
4459 	KKASSERT(mbp != bp);
4460 
4461 	error = bp->b_error;
4462 	if (bp->b_error == 0 &&
4463 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4464 		/*
4465 		 * Not all got transfered, raise an error. We have no way to
4466 		 * propagate these conditions to mbp.
4467 		 */
4468 		error = EIO;
4469 	}
4470 
4471 	donebytes = bp->b_bufsize;
4472 
4473 	relpbuf(bp, NULL);
4474 
4475 	nestiobuf_done(mbio, donebytes, error, stats);
4476 }
4477 
4478 void
nestiobuf_done(struct bio * mbio,int donebytes,int error,struct devstat * stats)4479 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4480 {
4481 	struct buf *mbp;
4482 
4483 	mbp = mbio->bio_buf;
4484 
4485 	KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4486 
4487 	/*
4488 	 * If an error occured, propagate it to the master buffer.
4489 	 *
4490 	 * Several biodone()s may wind up running concurrently so
4491 	 * use an atomic op to adjust b_flags.
4492 	 */
4493 	if (error) {
4494 		mbp->b_error = error;
4495 		atomic_set_int(&mbp->b_flags, B_ERROR);
4496 	}
4497 
4498 	/*
4499 	 * Decrement the operations in progress counter and terminate the
4500 	 * I/O if this was the last bit.
4501 	 */
4502 	if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4503 		mbp->b_resid = 0;
4504 		if (stats)
4505 			devstat_end_transaction_buf(stats, mbp);
4506 		biodone(mbio);
4507 	}
4508 }
4509 
4510 /*
4511  * Initialize a nestiobuf for use.  Set an initial count of 1 to prevent
4512  * the mbio from being biodone()'d while we are still adding sub-bios to
4513  * it.
4514  */
4515 void
nestiobuf_init(struct bio * bio)4516 nestiobuf_init(struct bio *bio)
4517 {
4518 	bio->bio_driver_info = (void *)1;
4519 }
4520 
4521 /*
4522  * The BIOs added to the nestedio have already been started, remove the
4523  * count that placeheld our mbio and biodone() it if the count would
4524  * transition to 0.
4525  */
4526 void
nestiobuf_start(struct bio * mbio)4527 nestiobuf_start(struct bio *mbio)
4528 {
4529 	struct buf *mbp = mbio->bio_buf;
4530 
4531 	/*
4532 	 * Decrement the operations in progress counter and terminate the
4533 	 * I/O if this was the last bit.
4534 	 */
4535 	if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4536 		if (mbp->b_flags & B_ERROR)
4537 			mbp->b_resid = mbp->b_bcount;
4538 		else
4539 			mbp->b_resid = 0;
4540 		biodone(mbio);
4541 	}
4542 }
4543 
4544 /*
4545  * Set an intermediate error prior to calling nestiobuf_start()
4546  */
4547 void
nestiobuf_error(struct bio * mbio,int error)4548 nestiobuf_error(struct bio *mbio, int error)
4549 {
4550 	struct buf *mbp = mbio->bio_buf;
4551 
4552 	if (error) {
4553 		mbp->b_error = error;
4554 		atomic_set_int(&mbp->b_flags, B_ERROR);
4555 	}
4556 }
4557 
4558 /*
4559  * nestiobuf_add: setup a "nested" buffer.
4560  *
4561  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4562  * => 'bp' should be a buffer allocated by getiobuf.
4563  * => 'offset' is a byte offset in the master buffer.
4564  * => 'size' is a size in bytes of this nested buffer.
4565  */
4566 void
nestiobuf_add(struct bio * mbio,struct buf * bp,int offset,size_t size,struct devstat * stats)4567 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4568 {
4569 	struct buf *mbp = mbio->bio_buf;
4570 	struct vnode *vp = mbp->b_vp;
4571 
4572 	KKASSERT(mbp->b_bcount >= offset + size);
4573 
4574 	atomic_add_int((int *)&mbio->bio_driver_info, 1);
4575 
4576 	/* kernel needs to own the lock for it to be released in biodone */
4577 	BUF_KERNPROC(bp);
4578 	bp->b_vp = vp;
4579 	bp->b_cmd = mbp->b_cmd;
4580 	bp->b_bio1.bio_done = nestiobuf_iodone;
4581 	bp->b_data = (char *)mbp->b_data + offset;
4582 	bp->b_resid = bp->b_bcount = size;
4583 	bp->b_bufsize = bp->b_bcount;
4584 
4585 	bp->b_bio1.bio_track = NULL;
4586 	bp->b_bio1.bio_caller_info1.ptr = mbio;
4587 	bp->b_bio1.bio_caller_info2.ptr = stats;
4588 }
4589 
4590 const char *
buf_cmd_name(struct buf * bp)4591 buf_cmd_name(struct buf *bp)
4592 {
4593 	const char *name;
4594 
4595 	switch(bp->b_cmd) {
4596 	case BUF_CMD_DONE:
4597 		name = "(DONE)";
4598 		break;
4599 	case BUF_CMD_READ:
4600 		name = "READ";
4601 		break;
4602 	case BUF_CMD_WRITE:
4603 		name = "WRITE";
4604 		break;
4605 	case BUF_CMD_FREEBLKS:
4606 		name = "FREEBLKS";
4607 		break;
4608 	case BUF_CMD_FORMAT:
4609 		name = "FORMAT";
4610 		break;
4611 	case BUF_CMD_FLUSH:
4612 		name = "FLUSH";
4613 		break;
4614 	default:
4615 		name = "(UNKNOWN)";
4616 		break;
4617 	}
4618 	return name;
4619 }
4620 
4621 
4622 #ifdef DDB
4623 
DB_SHOW_COMMAND(buffer,db_show_buffer)4624 DB_SHOW_COMMAND(buffer, db_show_buffer)
4625 {
4626 	/* get args */
4627 	struct buf *bp = (struct buf *)addr;
4628 
4629 	if (!have_addr) {
4630 		db_printf("usage: show buffer <addr>\n");
4631 		return;
4632 	}
4633 
4634 	db_printf("b_flags = 0x%pb%i\n", PRINT_BUF_FLAGS, bp->b_flags);
4635 	db_printf("b_cmd = %d\n", bp->b_cmd);
4636 	db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4637 		  "b_resid = %d\n, b_data = %p, "
4638 		  "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4639 		  bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4640 		  bp->b_data,
4641 		  (long long)bp->b_bio2.bio_offset,
4642 		  (long long)(bp->b_bio2.bio_next ?
4643 				bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4644 	if (bp->b_xio.xio_npages) {
4645 		int i;
4646 		db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4647 			bp->b_xio.xio_npages);
4648 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
4649 			vm_page_t m;
4650 			m = bp->b_xio.xio_pages[i];
4651 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4652 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4653 			if ((i + 1) < bp->b_xio.xio_npages)
4654 				db_printf(",");
4655 		}
4656 		db_printf("\n");
4657 	}
4658 }
4659 #endif /* DDB */
4660