2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 #include <afsconfig.h>
11 #include "../afs/param.h"
15 #include "../afs/sysincludes.h" /* Standard vendor system headers */
16 #include "../afs/afsincludes.h" /* Afs-based standard headers */
17 #include "../afs/afs_stats.h" /* statistics gathering code */
18 #include "../afs/afs_cbqueue.h"
20 #include <sys/adspace.h> /* for vm_att(), vm_det() */
24 /* background request queue size */
25 afs_lock_t afs_xbrs; /* lock for brs */
26 static int brsInit = 0;
27 short afs_brsWaiters = 0; /* number of users waiting for brs buffers */
28 short afs_brsDaemons = 0; /* number of daemons waiting for brs requests */
29 struct brequest afs_brs[NBRS]; /* request structures */
30 struct afs_osi_WaitHandle AFS_WaitHandler, AFS_CSWaitHandler;
31 static int afs_brs_count = 0; /* request counter, to service reqs in order */
33 static int rxepoch_checked=0;
34 #define afs_CheckRXEpoch() {if (rxepoch_checked == 0 && rxkad_EpochWasSet) { \
35 rxepoch_checked = 1; afs_GCUserData(/* force flag */ 1); } }
37 extern char afs_rootVolumeName[];
38 extern struct vcache *afs_globalVp;
39 extern struct VenusFid afs_rootFid;
40 extern struct osi_dev cacheDev;
41 extern char *afs_indexFlags;
42 extern afs_rwlock_t afs_xvcache;
43 extern struct afs_exporter *afs_nfsexporter;
44 extern int cacheDiskType;
45 extern int afs_BumpBase();
46 extern void afs_CheckCallbacks();
48 /* PAG garbage collection */
49 /* We induce a compile error if param.h does not define AFS_GCPAGS */
50 afs_int32 afs_gcpags=AFS_GCPAGS;
51 afs_int32 afs_gcpags_procsize;
53 afs_int32 afs_CheckServerDaemonStarted = 0;
54 afs_int32 PROBE_INTERVAL=180; /* default to 3 min */
56 #define PROBE_WAIT() (1000 * (PROBE_INTERVAL - ((afs_random() & 0x7fffffff) \
57 % (PROBE_INTERVAL/2))))
59 afs_CheckServerDaemon()
61 afs_int32 now, delay, lastCheck, last10MinCheck;
63 afs_CheckServerDaemonStarted = 1;
65 while (afs_initState < 101) afs_osi_Sleep(&afs_initState);
66 afs_osi_Wait(PROBE_WAIT(), &AFS_CSWaitHandler, 0);
68 last10MinCheck = lastCheck = osi_Time();
70 if (afs_termState == AFSOP_STOP_CS) {
71 afs_termState = AFSOP_STOP_BKG;
72 afs_osi_Wakeup(&afs_termState);
77 if (PROBE_INTERVAL + lastCheck <= now) {
78 afs_CheckServers(1, (struct cell *) 0); /* check down servers */
79 lastCheck = now = osi_Time();
82 if (600 + last10MinCheck <= now) {
83 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, 600);
84 afs_CheckServers(0, (struct cell *) 0);
85 last10MinCheck = now = osi_Time();
88 if (afs_termState == AFSOP_STOP_CS) {
89 afs_termState = AFSOP_STOP_BKG;
90 afs_osi_Wakeup(&afs_termState);
94 /* Compute time to next probe. */
95 delay = PROBE_INTERVAL + lastCheck;
96 if (delay > 600 + last10MinCheck)
97 delay = 600 + last10MinCheck;
101 afs_osi_Wait(delay * 1000, &AFS_CSWaitHandler, 0);
103 afs_CheckServerDaemonStarted = 0;
108 extern struct afs_exporter *root_exported;
109 struct afs_exporter *exporter;
111 afs_int32 last3MinCheck, last10MinCheck, last60MinCheck, lastNMinCheck;
112 afs_int32 last1MinCheck;
113 afs_uint32 lastCBSlotBump;
116 AFS_STATCNT(afs_Daemon);
117 last1MinCheck = last3MinCheck = last60MinCheck = last10MinCheck = lastNMinCheck = 0;
119 afs_rootFid.Fid.Volume = 0;
120 while (afs_initState < 101) afs_osi_Sleep(&afs_initState);
123 lastCBSlotBump = now;
125 /* when a lot of clients are booted simultaneously, they develop
126 * annoying synchronous VL server bashing behaviors. So we stagger them.
128 last1MinCheck = now + ((afs_random() & 0x7fffffff) % 60); /* an extra 30 */
129 last3MinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
130 last60MinCheck = now - 1800 + ((afs_random() & 0x7fffffff) % 3600);
131 last10MinCheck = now - 300 + ((afs_random() & 0x7fffffff) % 600);
132 lastNMinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
134 /* start off with afs_initState >= 101 (basic init done) */
136 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
138 /* things to do every 20 seconds or less - required by protocol spec */
140 afs_FlushActiveVcaches(0); /* flush NFS writes */
141 afs_FlushVCBs(1); /* flush queued callbacks */
142 afs_MaybeWakeupTruncateDaemon(); /* free cache space if have too */
143 rx_CheckPackets(); /* Does RX need more packets? */
144 #if defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)
146 * Hack: We always want to make sure there are plenty free
147 * entries in the small free pool so that we don't have to
148 * worry about rx (with disabled interrupts) to have to call
149 * malloc). So we do the dummy call below...
151 if (((afs_stats_cmperf.SmallBlocksAlloced - afs_stats_cmperf.SmallBlocksActive)
152 <= AFS_SALLOC_LOW_WATER))
153 osi_FreeSmallSpace(osi_AllocSmallSpace(AFS_SMALLOCSIZ));
154 if (((afs_stats_cmperf.MediumBlocksAlloced - afs_stats_cmperf.MediumBlocksActive)
155 <= AFS_MALLOC_LOW_WATER+50))
156 osi_AllocMoreMSpace(AFS_MALLOC_LOW_WATER * 2);
160 if (lastCBSlotBump + CBHTSLOTLEN < now) { /* pretty time-dependant */
161 lastCBSlotBump = now;
162 if (afs_BumpBase()) {
163 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
167 if (last1MinCheck + 60 < now) {
168 /* things to do every minute */
169 DFlush(); /* write out dir buffers */
170 afs_WriteThroughDSlots(); /* write through cacheinfo entries */
171 afs_FlushActiveVcaches(1);/* keep flocks held & flush nfs writes */
176 if (last3MinCheck + 180 < now) {
177 afs_CheckTokenCache(); /* check for access cache resets due to expired
181 if (!afs_CheckServerDaemonStarted) {
182 /* Do the check here if the correct afsd is not installed. */
185 printf("Please install afsd with check server daemon.\n");
187 if (lastNMinCheck + PROBE_INTERVAL < now) {
188 /* only check down servers */
189 afs_CheckServers(1, (struct cell *) 0);
193 if (last10MinCheck + 600 < now) {
194 #ifdef AFS_USERSPACE_IP_ADDR
195 extern int rxi_GetcbiInfo(void);
197 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP,
198 ICL_TYPE_INT32, 600);
199 #ifdef AFS_USERSPACE_IP_ADDR
200 if (rxi_GetcbiInfo()) { /* addresses changed from last time */
203 #else /* AFS_USERSPACE_IP_ADDR */
204 if (rxi_GetIFInfo()) { /* addresses changed from last time */
207 #endif /* else AFS_USERSPACE_IP_ADDR */
208 if (!afs_CheckServerDaemonStarted)
209 afs_CheckServers(0, (struct cell *) 0);
210 afs_GCUserData(0); /* gc old conns */
211 /* This is probably the wrong way of doing GC for the various exporters but it will suffice for a while */
212 for (exporter = root_exported; exporter; exporter = exporter->exp_next) {
213 (void) EXP_GC(exporter, 0); /* Generalize params */
218 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
222 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
227 last10MinCheck = now;
229 if (last60MinCheck + 3600 < now) {
230 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEVOLUME,
231 ICL_TYPE_INT32, 3600);
232 afs_CheckRootVolume();
234 if (afs_gcpags == AFS_GCPAGS_OK) {
239 last60MinCheck = now;
241 if (afs_initState < 300) { /* while things ain't rosy */
242 code = afs_CheckRootVolume();
243 if (code == 0) afs_initState = 300; /* succeeded */
244 if (afs_initState < 200) afs_initState = 200; /* tried once */
245 afs_osi_Wakeup(&afs_initState);
248 /* 18285 is because we're trying to divide evenly into 128, that is,
249 * CBSlotLen, while staying just under 20 seconds. If CBSlotLen
250 * changes, should probably change this interval, too.
251 * Some of the preceding actions may take quite some time, so we
252 * might not want to wait the entire interval */
253 now = 18285 - (osi_Time() - now);
255 afs_osi_Wait(now, &AFS_WaitHandler, 0);
258 if (afs_termState == AFSOP_STOP_AFS) {
259 if (afs_CheckServerDaemonStarted)
260 afs_termState = AFSOP_STOP_CS;
262 afs_termState = AFSOP_STOP_BKG;
263 afs_osi_Wakeup(&afs_termState);
269 afs_CheckRootVolume () {
270 char rootVolName[32];
271 register struct volume *tvp;
272 int usingDynroot = afs_GetDynrootEnable();
274 AFS_STATCNT(afs_CheckRootVolume);
275 if (*afs_rootVolumeName == 0) {
276 strcpy(rootVolName, "root.afs");
279 strcpy(rootVolName, afs_rootVolumeName);
282 afs_GetDynrootFid(&afs_rootFid);
283 tvp = afs_GetVolume(&afs_rootFid, (struct vrequest *) 0, READ_LOCK);
285 tvp = afs_GetVolumeByName(rootVolName, LOCALCELL, 1, (struct vrequest *) 0, READ_LOCK);
289 int len = strlen(rootVolName);
291 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
292 strcpy(buf, rootVolName);
293 afs_strcat(buf, ".readonly");
294 tvp = afs_GetVolumeByName(buf, LOCALCELL, 1, (struct vrequest *) 0, READ_LOCK);
299 int volid = (tvp->roVol? tvp->roVol : tvp->volume);
300 afs_rootFid.Cell = LOCALCELL;
301 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
303 /* If we had a root fid before and it changed location we reset
304 * the afs_globalVp so that it will be reevaluated.
305 * Just decrement the reference count. This only occurs during
306 * initial cell setup and can panic the machine if we set the
307 * count to zero and fs checkv is executed when the current
310 AFS_FAST_RELE(afs_globalVp);
313 afs_rootFid.Fid.Volume = volid;
314 afs_rootFid.Fid.Vnode = 1;
315 afs_rootFid.Fid.Unique = 1;
317 afs_initState = 300; /* won */
318 afs_osi_Wakeup(&afs_initState);
319 afs_PutVolume(tvp, READ_LOCK);
322 /* This is to make sure that we update the root gnode */
323 /* every time root volume gets released */
325 extern struct vfs *afs_globalVFS;
326 extern int afs_root();
327 struct gnode *rootgp;
331 /* Only do this if afs_globalVFS is properly set due to race conditions
332 this routine could be called before the gfs_mount is performed!
333 Furthermore, afs_root (called below) *waits* until
334 initState >= 200, so we don't try this until we've gotten
336 if (afs_globalVFS && afs_initState >= 200) {
337 if (code = afs_root(afs_globalVFS, &rootgp))
339 mp = (struct mount *) afs_globalVFS->vfs_data ;
340 mp->m_rootgp = gget(mp, 0, 0, (char *)rootgp);
341 afs_unlock(mp->m_rootgp); /* unlock basic gnode */
342 afs_vrele((struct vcache *) rootgp); /* zap afs_root's vnode hold */
346 if (afs_rootFid.Fid.Volume) return 0;
350 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
352 register struct brequest *ab; {
353 register struct dcache *tdc;
356 #ifdef AFS_LINUX22_ENV
357 struct dentry *dp = NULL;
359 afs_size_t offset, len;
360 struct vrequest treq;
364 if (code = afs_InitReq(&treq, ab->cred)) return;
366 #ifdef AFS_LINUX22_ENV
367 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, (struct vnode **) 0, &dp);
369 tvn = (struct vnode*)dp->d_inode;
371 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, (struct vnode **) 0, (struct vnode **)&tvn);
374 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
376 /* now path may not have been in afs, so check that before calling our cache manager */
377 if (!tvn || !IsAfsVnode((struct vnode *) tvn)) {
378 /* release it and give up */
383 #ifdef AFS_LINUX22_ENV
386 AFS_RELE((struct vnode *) tvn);
393 tvc = (struct vcache *) afs_gntovn(tvn);
395 tvc = (struct vcache *) tvn;
397 /* here we know its an afs vnode, so we can get the data for the chunk */
398 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
405 #ifdef AFS_LINUX22_ENV
408 AFS_RELE((struct vnode *) tvn);
413 /* size_parm 0 to the fetch is the chunk number,
414 * ptr_parm 0 is the dcache entry to wakeup,
415 * size_parm 1 is true iff we should release the dcache entry here.
418 register struct brequest *ab; {
419 register struct dcache *tdc;
420 register struct vcache *tvc;
421 afs_size_t offset, len;
422 struct vrequest treq;
424 AFS_STATCNT(BPrefetch);
425 if (len = afs_InitReq(&treq, ab->cred)) return;
427 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
431 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
432 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
433 * be waiting for our wakeup anyway.
435 tdc = (struct dcache *) (ab->ptr_parm[0]);
436 tdc->flags &= ~DFFetchReq;
437 afs_osi_Wakeup(&tdc->validPos);
438 if (ab->size_parm[1]) {
440 mutex_enter(&tdc->lock);
442 mutex_exit(&tdc->lock);
444 afs_PutDCache(tdc); /* put this one back, too */
451 register struct brequest *ab; {
452 register struct vcache *tvc;
453 register afs_int32 code;
454 struct vrequest treq;
455 #if defined(AFS_SGI_ENV)
456 struct cred *tmpcred;
460 if (code = afs_InitReq(&treq, ab->cred)) return;
463 #if defined(AFS_SGI_ENV)
465 * Since StoreOnLastReference can end up calling osi_SyncVM which
466 * calls into VM code that assumes that u.u_cred has the
467 * correct credentials, we set our to theirs for this xaction
469 tmpcred = OSI_GET_CURRENT_CRED();
470 OSI_SET_CURRENT_CRED(ab->cred);
473 * To avoid recursion since the WriteLock may be released during VM
474 * operations, we hold the VOP_RWLOCK across this transaction as
475 * do the other callers of StoreOnLastReference
477 AFS_RWLOCK((vnode_t *)tvc, 1);
479 ObtainWriteLock(&tvc->lock,209);
480 code = afs_StoreOnLastReference(tvc, &treq);
481 ReleaseWriteLock(&tvc->lock);
482 #if defined(AFS_SGI_ENV)
483 OSI_SET_CURRENT_CRED(tmpcred);
484 AFS_RWUNLOCK((vnode_t *)tvc, 1);
486 /* now set final return code, and wakeup anyone waiting */
487 if ((ab->flags & BUVALID) == 0) {
488 ab->code = afs_CheckCode(code, &treq, 43); /* set final code, since treq doesn't go across processes */
489 ab->flags |= BUVALID;
490 if (ab->flags & BUWAIT) {
491 ab->flags &= ~BUWAIT;
497 /* release a held request buffer */
498 void afs_BRelease(ab)
499 register struct brequest *ab; {
501 AFS_STATCNT(afs_BRelease);
502 MObtainWriteLock(&afs_xbrs,294);
503 if (--ab->refCount <= 0) {
506 if (afs_brsWaiters) afs_osi_Wakeup(&afs_brsWaiters);
507 MReleaseWriteLock(&afs_xbrs);
510 /* return true if bkg fetch daemons are all busy */
512 AFS_STATCNT(afs_BBusy);
513 if (afs_brsDaemons > 0) return 0;
517 struct brequest *afs_BQueue(aopcode, avc, dontwait, ause, acred, asparm0, asparm1, apparm0)
518 register short aopcode;
519 afs_int32 ause, dontwait;
520 register struct vcache *avc;
521 struct AFS_UCRED *acred;
522 afs_size_t asparm0, asparm1;
526 register struct brequest *tb;
528 AFS_STATCNT(afs_BQueue);
529 MObtainWriteLock(&afs_xbrs,296);
532 for(i=0;i<NBRS;i++,tb++) {
533 if (tb->refCount == 0) break;
537 tb->opcode = aopcode;
545 VN_HOLD((struct vnode *)avc);
548 tb->refCount = ause+1;
549 tb->size_parm[0] = asparm0;
550 tb->size_parm[1] = asparm1;
551 tb->ptr_parm[0] = apparm0;
554 tb->ts = afs_brs_count++;
555 /* if daemons are waiting for work, wake them up */
556 if (afs_brsDaemons > 0) {
557 afs_osi_Wakeup(&afs_brsDaemons);
559 MReleaseWriteLock(&afs_xbrs);
563 MReleaseWriteLock(&afs_xbrs);
564 return (struct brequest *)0;
566 /* no free buffers, sleep a while */
568 MReleaseWriteLock(&afs_xbrs);
569 afs_osi_Sleep(&afs_brsWaiters);
570 MObtainWriteLock(&afs_xbrs,301);
577 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
578 * The modifications here will work for either a UP or MP machine.
580 struct buf *afs_asyncbuf = (struct buf*)0;
581 afs_int32 afs_asyncbuf_cv = EVENT_NULL;
582 afs_int32 afs_biodcnt = 0;
584 /* in implementing this, I assumed that all external linked lists were
587 * Several places in this code traverse a linked list. The algorithm
588 * used here is probably unfamiliar to most people. Careful examination
589 * will show that it eliminates an assignment inside the loop, as compared
590 * to the standard algorithm, at the cost of occasionally using an extra
596 * This function obtains, and returns, a pointer to a buffer for
597 * processing by a daemon. It sleeps until such a buffer is available.
598 * The source of buffers for it is the list afs_asyncbuf (see also
599 * naix_vm_strategy). This function may be invoked concurrently by
600 * several processes, that is, several instances of the same daemon.
601 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
602 * level, while get_bioreq runs at process level.
604 * Since AIX 4.1 can wake just one process at a time, the separate sleep
605 * addresses have been removed.
606 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
607 * The afs_asyncbuf_lock is primarily used to serialize access between
608 * process and interrupts.
610 Simple_lock afs_asyncbuf_lock;
611 /*static*/ struct buf *afs_get_bioreq()
613 struct buf *bp = (struct buf *) 0;
615 struct buf **bestlbpP, **lbpP;
617 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
620 struct afs_bioqueue *s;
622 /* ??? Does the forward pointer of the returned buffer need to be NULL?
625 /* Disable interrupts from the strategy function, and save the
626 * prior priority level and lock access to the afs_asyncbuf.
629 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock) ;
633 /* look for oldest buffer */
634 bp = bestbp = afs_asyncbuf;
635 bestage = (int) bestbp->av_back;
636 bestlbpP = &afs_asyncbuf;
641 if ((int) bp->av_back - bestage < 0) {
644 bestage = (int) bp->av_back;
648 *bestlbpP = bp->av_forw;
652 /* If afs_asyncbuf is null, it is necessary to go to sleep.
653 * e_wakeup_one() ensures that only one thread wakes.
656 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
657 * lock on an MP machine.
659 interrupted = e_sleep_thread(&afs_asyncbuf_cv,
661 LOCK_HANDLER|INTERRUPTIBLE);
662 if (interrupted==THREAD_INTERRUPTED) {
663 /* re-enable interrupts from strategy */
664 unlock_enable(oldPriority, &afs_asyncbuf_lock);
668 } /* end of "else asyncbuf is empty" */
669 } /* end of "inner loop" */
673 unlock_enable(oldPriority, &afs_asyncbuf_lock);
676 /* For the convenience of other code, replace the gnodes in
677 * the b_vp field of bp and the other buffers on the b_work
678 * chain with the corresponding vnodes.
680 * ??? what happens to the gnodes? They're not just cut loose,
684 t2P = (struct buf *) t1P->b_work;
685 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
689 t1P = (struct buf *) t2P->b_work;
690 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
695 /* If the buffer does not specify I/O, it may immediately
696 * be returned to the caller. This condition is detected
697 * by examining the buffer's flags (the b_flags field). If
698 * the B_PFPROT bit is set, the buffer represents a protection
699 * violation, rather than a request for I/O. The remainder
700 * of the outer loop handles the case where the B_PFPROT bit is clear.
702 if (bp->b_flags & B_PFPROT) {
707 } /* end of function get_bioreq() */
712 * This function is the daemon. It is called from the syscall
713 * interface. Ordinarily, a script or an administrator will run a
714 * daemon startup utility, specifying the number of I/O daemons to
715 * run. The utility will fork off that number of processes,
716 * each making the appropriate syscall, which will cause this
717 * function to be invoked.
719 static int afs_initbiod = 0; /* this is self-initializing code */
721 afs_BioDaemon (nbiods)
724 afs_int32 code, s, pflg = 0;
726 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
734 /* pin lock, since we'll be using it in an interrupt. */
735 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
736 simple_lock_init(&afs_asyncbuf_lock);
737 pin (&afs_asyncbuf, sizeof(struct buf*));
738 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
741 /* Ignore HUP signals... */
743 sigset_t sigbits, osigbits;
745 * add SIGHUP to the set of already masked signals
747 SIGFILLSET(sigbits); /* allow all signals */
748 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
749 limit_sigs(&sigbits, &osigbits); /* and already masked */
751 /* Main body starts here -- this is an intentional infinite loop, and
754 * Now, the loop will exit if get_bioreq() returns NULL, indicating
755 * that we've been interrupted.
758 bp = afs_get_bioreq();
760 break; /* we were interrupted */
761 if (code = setjmpx(&jmpbuf)) {
762 /* This should not have happend, maybe a lack of resources */
764 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
765 for (bp1 = bp; bp ; bp = bp1) {
767 bp1 = (struct buf *) bp1->b_work;
770 bp->b_flags |= B_ERROR;
773 unlock_enable(s, &afs_asyncbuf_lock);
777 vcp = (struct vcache *)bp->b_vp;
778 if (bp->b_flags & B_PFSTORE) { /* XXXX */
779 ObtainWriteLock(&vcp->lock,404);
780 if (vcp->v.v_gnode->gn_mwrcnt) {
781 afs_offs_t newlength =
782 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
783 if (vcp->m.Length < newlength) {
784 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
785 ICL_TYPE_STRING, __FILE__,
786 ICL_TYPE_LONG, __LINE__,
787 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(vcp->m.Length),
788 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
789 vcp->m.Length = newlength;
792 ReleaseWriteLock(&vcp->lock);
794 /* If the buffer represents a protection violation, rather than
795 * an actual request for I/O, no special action need be taken.
797 if ( bp->b_flags & B_PFPROT ) {
798 iodone (bp); /* Notify all users of the buffer that we're done */
803 ObtainWriteLock(&vcp->pvmlock,211);
805 * First map its data area to a region in the current address space
806 * by calling vm_att with the subspace identifier, and a pointer to
807 * the data area. vm_att returns a new data area pointer, but we
808 * also want to hang onto the old one.
810 tmpaddr = bp->b_baddr;
811 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
812 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
813 if (tmperr) { /* in non-error case */
814 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
815 bp->b_error = tmperr;
818 /* Unmap the buffer's data area by calling vm_det. Reset data area
819 * to the value that we saved above.
821 vm_det(bp->b_un.b_addr);
822 bp->b_baddr = tmpaddr;
825 * buffer may be linked with other buffers via the b_work field.
826 * See also naix_vm_strategy. For each buffer in the chain (including
827 * bp) notify all users of the buffer that the daemon is finished
828 * using it by calling iodone.
829 * assumes iodone can modify the b_work field.
832 tbp2 = (struct buf *) tbp1->b_work;
837 tbp1 = (struct buf *) tbp2->b_work;
843 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
845 } /* infinite loop (unless we're interrupted) */
846 } /* end of afs_BioDaemon() */
848 #else /* AFS_AIX41_ENV */
852 struct afs_bioqueue {
857 struct afs_bioqueue afs_bioqueue;
858 struct buf *afs_busyq = NULL;
859 struct buf *afs_asyncbuf;
860 afs_int32 afs_biodcnt = 0;
862 /* in implementing this, I assumed that all external linked lists were
865 * Several places in this code traverse a linked list. The algorithm
866 * used here is probably unfamiliar to most people. Careful examination
867 * will show that it eliminates an assignment inside the loop, as compared
868 * to the standard algorithm, at the cost of occasionally using an extra
874 * This function obtains, and returns, a pointer to a buffer for
875 * processing by a daemon. It sleeps until such a buffer is available.
876 * The source of buffers for it is the list afs_asyncbuf (see also
877 * naix_vm_strategy). This function may be invoked concurrently by
878 * several processes, that is, several instances of the same daemon.
879 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
880 * level, while get_bioreq runs at process level.
882 * The common kernel paradigm of sleeping and waking up, in which all the
883 * competing processes sleep waiting for wakeups on one address, is not
884 * followed here. Instead, the following paradigm is used: when a daemon
885 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
886 * it sleeps on the address of variable afs_asyncbuf. But if there is
887 * already a daemon sleeping on that address, it threads its own unique
888 * address onto a list, and sleeps on that address. This way, every
889 * sleeper is sleeping on a different address, and every wakeup wakes up
890 * exactly one daemon. This prevents a whole bunch of daemons from waking
891 * up and then immediately having to go back to sleep. This provides a
892 * performance gain and makes the I/O scheduling a bit more deterministic.
893 * The list of sleepers is variable afs_bioqueue. The unique address
894 * on which to sleep is passed to get_bioreq as its parameter.
896 /*static*/ struct buf *afs_get_bioreq(self)
897 struct afs_bioqueue *self; /* address on which to sleep */
900 struct buf *bp = (struct buf *) 0;
902 struct buf **bestlbpP, **lbpP;
904 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
907 struct afs_bioqueue *s;
909 /* ??? Does the forward pointer of the returned buffer need to be NULL?
912 /* Disable interrupts from the strategy function, and save the
913 * prior priority level
915 oldPriority = i_disable ( INTMAX ) ;
917 /* Each iteration of following loop either pulls
918 * a buffer off afs_asyncbuf, or sleeps.
920 while (1) { /* inner loop */
922 /* look for oldest buffer */
923 bp = bestbp = afs_asyncbuf;
924 bestage = (int) bestbp->av_back;
925 bestlbpP = &afs_asyncbuf;
930 if ((int) bp->av_back - bestage < 0) {
933 bestage = (int) bp->av_back;
937 *bestlbpP = bp->av_forw;
943 /* If afs_asyncbuf is null, it is necessary to go to sleep.
944 * There are two possibilities: either there is already a
945 * daemon that is sleeping on the address of afs_asyncbuf,
948 if (afs_bioqueue.sleeper) {
950 QAdd (&(afs_bioqueue.lruq), &(self->lruq));
951 interrupted = sleep ((caddr_t) self, PCATCH|(PZERO + 1));
952 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
953 QRemove (&(self->lruq)); /* dequeue */
956 afs_bioqueue.sleeper = FALSE;
958 /* re-enable interrupts from strategy */
959 i_enable (oldPriority);
964 afs_bioqueue.sleeper = TRUE;
965 interrupted = sleep ((caddr_t) &afs_asyncbuf, PCATCH|(PZERO + 1));
966 afs_bioqueue.sleeper = FALSE;
970 * We need to wakeup another daemon if present
971 * since we were waiting on afs_asyncbuf.
973 #ifdef notdef /* The following doesn't work as advertised */
974 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq)
976 struct squeue *bq = afs_bioqueue.lruq.next;
981 /* re-enable interrupts from strategy */
982 i_enable (oldPriority);
988 } /* end of "else asyncbuf is empty" */
989 } /* end of "inner loop" */
993 i_enable (oldPriority); /* re-enable interrupts from strategy */
995 /* For the convenience of other code, replace the gnodes in
996 * the b_vp field of bp and the other buffers on the b_work
997 * chain with the corresponding vnodes.
999 * ??? what happens to the gnodes? They're not just cut loose,
1003 t2P = (struct buf *) t1P->b_work;
1004 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
1008 t1P = (struct buf *) t2P->b_work;
1009 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
1014 /* If the buffer does not specify I/O, it may immediately
1015 * be returned to the caller. This condition is detected
1016 * by examining the buffer's flags (the b_flags field). If
1017 * the B_PFPROT bit is set, the buffer represents a protection
1018 * violation, rather than a request for I/O. The remainder
1019 * of the outer loop handles the case where the B_PFPROT bit is clear.
1021 if (bp->b_flags & B_PFPROT) {
1025 /* wake up another process to handle the next buffer, and return
1028 oldPriority = i_disable ( INTMAX ) ;
1030 /* determine where to find the sleeping process.
1031 * There are two cases: either it is sleeping on
1032 * afs_asyncbuf, or it is sleeping on its own unique
1033 * address. These cases are distinguished by examining
1034 * the sleeper field of afs_bioqueue.
1036 if (afs_bioqueue.sleeper) {
1037 wakeup (&afs_asyncbuf);
1040 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1041 /* queue is empty, what now? ???*/
1042 /* Should this be impossible, or does */
1043 /* it just mean that nobody is sleeping? */;
1046 struct squeue *bq = afs_bioqueue.lruq.next;
1050 afs_bioqueue.sleeper = TRUE;
1053 i_enable (oldPriority); /* re-enable interrupts from strategy */
1056 } /* end of function get_bioreq() */
1061 * This function is the daemon. It is called from the syscall
1062 * interface. Ordinarily, a script or an administrator will run a
1063 * daemon startup utility, specifying the number of I/O daemons to
1064 * run. The utility will fork off that number of processes,
1065 * each making the appropriate syscall, which will cause this
1066 * function to be invoked.
1068 static int afs_initbiod = 0; /* this is self-initializing code */
1070 afs_BioDaemon (nbiods)
1073 struct afs_bioqueue *self;
1074 afs_int32 code, s, pflg = 0;
1076 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1081 if (!afs_initbiod) {
1084 /* Initialize the queue of waiting processes, afs_bioqueue. */
1085 QInit (&(afs_bioqueue.lruq));
1088 /* establish ourself as a kernel process so shutdown won't kill us */
1089 /* u.u_procp->p_flag |= SKPROC;*/
1091 /* Initialize a token (self) to use in the queue of sleeping processes. */
1092 self = (struct afs_bioqueue *) afs_osi_Alloc (sizeof (struct afs_bioqueue));
1093 pin (self, sizeof (struct afs_bioqueue)); /* fix in memory */
1094 memset(self, 0, sizeof(*self));
1095 QInit (&(self->lruq)); /* initialize queue entry pointers */
1098 /* Ignore HUP signals... */
1099 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1100 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1101 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1102 /* Main body starts here -- this is an intentional infinite loop, and
1105 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1106 * that we've been interrupted.
1109 bp = afs_get_bioreq(self);
1111 break; /* we were interrupted */
1112 if (code = setjmpx(&jmpbuf)) {
1113 /* This should not have happend, maybe a lack of resources */
1115 for (bp1 = bp; bp ; bp = bp1) {
1120 bp->b_flags |= B_ERROR;
1126 vcp = (struct vcache *)bp->b_vp;
1127 if (bp->b_flags & B_PFSTORE) {
1128 ObtainWriteLock(&vcp->lock,210);
1129 if (vcp->v.v_gnode->gn_mwrcnt) {
1130 if (vcp->m.Length < bp->b_bcount + (u_int)dbtob(bp->b_blkno))
1131 vcp->m.Length = bp->b_bcount + (u_int)dbtob(bp->b_blkno);
1133 ReleaseWriteLock(&vcp->lock);
1135 /* If the buffer represents a protection violation, rather than
1136 * an actual request for I/O, no special action need be taken.
1138 if ( bp->b_flags & B_PFPROT ) {
1139 iodone (bp); /* Notify all users of the buffer that we're done */
1143 ObtainWriteLock(&vcp->pvmlock,558);
1145 * First map its data area to a region in the current address space
1146 * by calling vm_att with the subspace identifier, and a pointer to
1147 * the data area. vm_att returns a new data area pointer, but we
1148 * also want to hang onto the old one.
1150 tmpaddr = bp->b_baddr;
1151 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
1152 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1153 if (tmperr) { /* in non-error case */
1154 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1155 bp->b_error = tmperr;
1158 /* Unmap the buffer's data area by calling vm_det. Reset data area
1159 * to the value that we saved above.
1161 vm_det(bp->b_un.b_addr);
1162 bp->b_baddr = tmpaddr;
1165 * buffer may be linked with other buffers via the b_work field.
1166 * See also naix_vm_strategy. For each buffer in the chain (including
1167 * bp) notify all users of the buffer that the daemon is finished
1168 * using it by calling iodone.
1169 * assumes iodone can modify the b_work field.
1172 tbp2 = (struct buf *) tbp1->b_work;
1177 tbp1 = (struct buf *) tbp2->b_work;
1183 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1185 } /* infinite loop (unless we're interrupted) */
1186 unpin (self, sizeof (struct afs_bioqueue));
1187 afs_osi_Free (self, sizeof (struct afs_bioqueue));
1188 } /* end of afs_BioDaemon() */
1189 #endif /* AFS_AIX41_ENV */
1190 #endif /* AFS_AIX32_ENV */
1194 void afs_BackgroundDaemon() {
1195 struct brequest *tb;
1199 AFS_STATCNT(afs_BackgroundDaemon);
1200 /* initialize subsystem */
1202 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1203 memset((char *)afs_brs, 0, sizeof(afs_brs));
1205 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1207 * steal the first daemon for doing delayed DSlot flushing
1208 * (see afs_GetDownDSlot)
1217 MObtainWriteLock(&afs_xbrs,302);
1220 struct brequest *min_tb;
1222 if (afs_termState == AFSOP_STOP_BKG) {
1223 if (--afs_nbrs <= 0)
1224 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1225 MReleaseWriteLock(&afs_xbrs);
1226 afs_osi_Wakeup(&afs_termState);
1230 /* find a request */
1234 for(i=0; i<NBRS; i++, tb++) {
1235 /* look for request with smallest ts */
1236 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1237 /* new request, not yet picked up */
1238 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1245 /* claim and process this request */
1246 tb->flags |= BSTARTED;
1247 MReleaseWriteLock(&afs_xbrs);
1249 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1,
1250 ICL_TYPE_INT32, tb->opcode);
1251 if (tb->opcode == BOP_FETCH)
1253 else if (tb->opcode == BOP_STORE)
1255 else if (tb->opcode == BOP_PATH)
1257 else panic("background bop");
1260 tb->vnode->vrefCount--; /* fix up reference count */
1262 AFS_RELE((struct vnode *)(tb->vnode)); /* MUST call vnode layer or could lose vnodes */
1264 tb->vnode = (struct vcache *) 0;
1268 tb->cred = (struct AFS_UCRED *) 0;
1270 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1271 MObtainWriteLock(&afs_xbrs,305);
1274 /* wait for new request */
1276 MReleaseWriteLock(&afs_xbrs);
1277 afs_osi_Sleep(&afs_brsDaemons);
1278 MObtainWriteLock(&afs_xbrs,307);
1285 void shutdown_daemons()
1287 extern int afs_cold_shutdown;
1289 register struct brequest *tb;
1291 AFS_STATCNT(shutdown_daemons);
1292 if (afs_cold_shutdown) {
1293 afs_brsDaemons = brsInit = 0;
1294 rxepoch_checked = afs_nbrs = 0;
1295 memset((char *)afs_brs, 0, sizeof(afs_brs));
1296 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1298 #ifdef AFS_AIX32_ENV
1299 #ifdef AFS_AIX41_ENV
1300 lock_free(&afs_asyncbuf_lock);
1301 unpin(&afs_asyncbuf, sizeof(struct buf*));
1302 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
1303 #else /* AFS_AIX41_ENV */
1306 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1313 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1315 * sgi - daemon - handles certain operations that otherwise
1316 * would use up too much kernel stack space
1318 * This all assumes that since the caller must have the xdcache lock
1319 * exclusively that the list will never be more than one long
1320 * and noone else can attempt to add anything until we're done.
1322 SV_TYPE afs_sgibksync;
1323 SV_TYPE afs_sgibkwait;
1324 lock_t afs_sgibklock;
1325 struct dcache *afs_sgibklist;
1333 if (afs_sgibklock == NULL) {
1334 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1335 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1336 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1338 s = SPLOCK(afs_sgibklock);
1340 /* wait for something to do */
1341 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1342 osi_Assert(afs_sgibklist);
1344 /* XX will probably need to generalize to real list someday */
1345 s = SPLOCK(afs_sgibklock);
1346 while (afs_sgibklist) {
1347 tdc = afs_sgibklist;
1348 afs_sgibklist = NULL;
1349 SPUNLOCK(afs_sgibklock, s);
1351 tdc->flags &= ~DFEntryMod;
1352 afs_WriteDCache(tdc, 1);
1354 s = SPLOCK(afs_sgibklock);
1357 /* done all the work - wake everyone up */
1358 while (SV_SIGNAL(&afs_sgibkwait))