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(VTOAFS(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 = VTOAFS(afs_gntovn(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 ObtainSharedLock(&tdc->lock, 640);
437 if (tdc->mflags & DFFetchReq) {
438 UpgradeSToWLock(&tdc->lock, 641);
439 tdc->mflags &= ~DFFetchReq;
440 ReleaseWriteLock(&tdc->lock);
442 ReleaseSharedLock(&tdc->lock);
444 afs_osi_Wakeup(&tdc->validPos);
445 if (ab->size_parm[1]) {
446 afs_PutDCache(tdc); /* put this one back, too */
452 register struct brequest *ab; {
453 register struct vcache *tvc;
454 register afs_int32 code;
455 struct vrequest treq;
456 #if defined(AFS_SGI_ENV)
457 struct cred *tmpcred;
461 if (code = afs_InitReq(&treq, ab->cred)) return;
464 #if defined(AFS_SGI_ENV)
466 * Since StoreOnLastReference can end up calling osi_SyncVM which
467 * calls into VM code that assumes that u.u_cred has the
468 * correct credentials, we set our to theirs for this xaction
470 tmpcred = OSI_GET_CURRENT_CRED();
471 OSI_SET_CURRENT_CRED(ab->cred);
474 * To avoid recursion since the WriteLock may be released during VM
475 * operations, we hold the VOP_RWLOCK across this transaction as
476 * do the other callers of StoreOnLastReference
478 AFS_RWLOCK((vnode_t *)tvc, 1);
480 ObtainWriteLock(&tvc->lock,209);
481 code = afs_StoreOnLastReference(tvc, &treq);
482 ReleaseWriteLock(&tvc->lock);
483 #if defined(AFS_SGI_ENV)
484 OSI_SET_CURRENT_CRED(tmpcred);
485 AFS_RWUNLOCK((vnode_t *)tvc, 1);
487 /* now set final return code, and wakeup anyone waiting */
488 if ((ab->flags & BUVALID) == 0) {
489 ab->code = afs_CheckCode(code, &treq, 43); /* set final code, since treq doesn't go across processes */
490 ab->flags |= BUVALID;
491 if (ab->flags & BUWAIT) {
492 ab->flags &= ~BUWAIT;
498 /* release a held request buffer */
499 void afs_BRelease(ab)
500 register struct brequest *ab; {
502 AFS_STATCNT(afs_BRelease);
503 MObtainWriteLock(&afs_xbrs,294);
504 if (--ab->refCount <= 0) {
507 if (afs_brsWaiters) afs_osi_Wakeup(&afs_brsWaiters);
508 MReleaseWriteLock(&afs_xbrs);
511 /* return true if bkg fetch daemons are all busy */
513 AFS_STATCNT(afs_BBusy);
514 if (afs_brsDaemons > 0) return 0;
518 struct brequest *afs_BQueue(aopcode, avc, dontwait, ause, acred, asparm0, asparm1, apparm0)
519 register short aopcode;
520 afs_int32 ause, dontwait;
521 register struct vcache *avc;
522 struct AFS_UCRED *acred;
523 afs_size_t asparm0, asparm1;
527 register struct brequest *tb;
529 AFS_STATCNT(afs_BQueue);
530 MObtainWriteLock(&afs_xbrs,296);
533 for(i=0;i<NBRS;i++,tb++) {
534 if (tb->refCount == 0) break;
538 tb->opcode = aopcode;
546 VN_HOLD(AFSTOV(avc));
549 tb->refCount = ause+1;
550 tb->size_parm[0] = asparm0;
551 tb->size_parm[1] = asparm1;
552 tb->ptr_parm[0] = apparm0;
555 tb->ts = afs_brs_count++;
556 /* if daemons are waiting for work, wake them up */
557 if (afs_brsDaemons > 0) {
558 afs_osi_Wakeup(&afs_brsDaemons);
560 MReleaseWriteLock(&afs_xbrs);
564 MReleaseWriteLock(&afs_xbrs);
565 return (struct brequest *)0;
567 /* no free buffers, sleep a while */
569 MReleaseWriteLock(&afs_xbrs);
570 afs_osi_Sleep(&afs_brsWaiters);
571 MObtainWriteLock(&afs_xbrs,301);
578 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
579 * The modifications here will work for either a UP or MP machine.
581 struct buf *afs_asyncbuf = (struct buf*)0;
582 afs_int32 afs_asyncbuf_cv = EVENT_NULL;
583 afs_int32 afs_biodcnt = 0;
585 /* in implementing this, I assumed that all external linked lists were
588 * Several places in this code traverse a linked list. The algorithm
589 * used here is probably unfamiliar to most people. Careful examination
590 * will show that it eliminates an assignment inside the loop, as compared
591 * to the standard algorithm, at the cost of occasionally using an extra
597 * This function obtains, and returns, a pointer to a buffer for
598 * processing by a daemon. It sleeps until such a buffer is available.
599 * The source of buffers for it is the list afs_asyncbuf (see also
600 * naix_vm_strategy). This function may be invoked concurrently by
601 * several processes, that is, several instances of the same daemon.
602 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
603 * level, while get_bioreq runs at process level.
605 * Since AIX 4.1 can wake just one process at a time, the separate sleep
606 * addresses have been removed.
607 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
608 * The afs_asyncbuf_lock is primarily used to serialize access between
609 * process and interrupts.
611 Simple_lock afs_asyncbuf_lock;
612 /*static*/ struct buf *afs_get_bioreq()
614 struct buf *bp = (struct buf *) 0;
616 struct buf **bestlbpP, **lbpP;
618 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
621 struct afs_bioqueue *s;
623 /* ??? Does the forward pointer of the returned buffer need to be NULL?
626 /* Disable interrupts from the strategy function, and save the
627 * prior priority level and lock access to the afs_asyncbuf.
630 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock) ;
634 /* look for oldest buffer */
635 bp = bestbp = afs_asyncbuf;
636 bestage = (int) bestbp->av_back;
637 bestlbpP = &afs_asyncbuf;
642 if ((int) bp->av_back - bestage < 0) {
645 bestage = (int) bp->av_back;
649 *bestlbpP = bp->av_forw;
653 /* If afs_asyncbuf is null, it is necessary to go to sleep.
654 * e_wakeup_one() ensures that only one thread wakes.
657 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
658 * lock on an MP machine.
660 interrupted = e_sleep_thread(&afs_asyncbuf_cv,
662 LOCK_HANDLER|INTERRUPTIBLE);
663 if (interrupted==THREAD_INTERRUPTED) {
664 /* re-enable interrupts from strategy */
665 unlock_enable(oldPriority, &afs_asyncbuf_lock);
669 } /* end of "else asyncbuf is empty" */
670 } /* end of "inner loop" */
674 unlock_enable(oldPriority, &afs_asyncbuf_lock);
677 /* For the convenience of other code, replace the gnodes in
678 * the b_vp field of bp and the other buffers on the b_work
679 * chain with the corresponding vnodes.
681 * ??? what happens to the gnodes? They're not just cut loose,
685 t2P = (struct buf *) t1P->b_work;
686 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
690 t1P = (struct buf *) t2P->b_work;
691 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
696 /* If the buffer does not specify I/O, it may immediately
697 * be returned to the caller. This condition is detected
698 * by examining the buffer's flags (the b_flags field). If
699 * the B_PFPROT bit is set, the buffer represents a protection
700 * violation, rather than a request for I/O. The remainder
701 * of the outer loop handles the case where the B_PFPROT bit is clear.
703 if (bp->b_flags & B_PFPROT) {
708 } /* end of function get_bioreq() */
713 * This function is the daemon. It is called from the syscall
714 * interface. Ordinarily, a script or an administrator will run a
715 * daemon startup utility, specifying the number of I/O daemons to
716 * run. The utility will fork off that number of processes,
717 * each making the appropriate syscall, which will cause this
718 * function to be invoked.
720 static int afs_initbiod = 0; /* this is self-initializing code */
722 afs_BioDaemon (nbiods)
725 afs_int32 code, s, pflg = 0;
727 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
735 /* pin lock, since we'll be using it in an interrupt. */
736 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
737 simple_lock_init(&afs_asyncbuf_lock);
738 pin (&afs_asyncbuf, sizeof(struct buf*));
739 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
742 /* Ignore HUP signals... */
744 sigset_t sigbits, osigbits;
746 * add SIGHUP to the set of already masked signals
748 SIGFILLSET(sigbits); /* allow all signals */
749 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
750 limit_sigs(&sigbits, &osigbits); /* and already masked */
752 /* Main body starts here -- this is an intentional infinite loop, and
755 * Now, the loop will exit if get_bioreq() returns NULL, indicating
756 * that we've been interrupted.
759 bp = afs_get_bioreq();
761 break; /* we were interrupted */
762 if (code = setjmpx(&jmpbuf)) {
763 /* This should not have happend, maybe a lack of resources */
765 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
766 for (bp1 = bp; bp ; bp = bp1) {
768 bp1 = (struct buf *) bp1->b_work;
771 bp->b_flags |= B_ERROR;
774 unlock_enable(s, &afs_asyncbuf_lock);
778 vcp = VTOAFS(bp->b_vp);
779 if (bp->b_flags & B_PFSTORE) { /* XXXX */
780 ObtainWriteLock(&vcp->lock,404);
781 if (vcp->v.v_gnode->gn_mwrcnt) {
782 afs_offs_t newlength =
783 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
784 if (vcp->m.Length < newlength) {
785 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
786 ICL_TYPE_STRING, __FILE__,
787 ICL_TYPE_LONG, __LINE__,
788 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(vcp->m.Length),
789 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
790 vcp->m.Length = newlength;
793 ReleaseWriteLock(&vcp->lock);
795 /* If the buffer represents a protection violation, rather than
796 * an actual request for I/O, no special action need be taken.
798 if ( bp->b_flags & B_PFPROT ) {
799 iodone (bp); /* Notify all users of the buffer that we're done */
804 ObtainWriteLock(&vcp->pvmlock,211);
806 * First map its data area to a region in the current address space
807 * by calling vm_att with the subspace identifier, and a pointer to
808 * the data area. vm_att returns a new data area pointer, but we
809 * also want to hang onto the old one.
811 tmpaddr = bp->b_baddr;
812 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
813 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
814 if (tmperr) { /* in non-error case */
815 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
816 bp->b_error = tmperr;
819 /* Unmap the buffer's data area by calling vm_det. Reset data area
820 * to the value that we saved above.
822 vm_det(bp->b_un.b_addr);
823 bp->b_baddr = tmpaddr;
826 * buffer may be linked with other buffers via the b_work field.
827 * See also naix_vm_strategy. For each buffer in the chain (including
828 * bp) notify all users of the buffer that the daemon is finished
829 * using it by calling iodone.
830 * assumes iodone can modify the b_work field.
833 tbp2 = (struct buf *) tbp1->b_work;
838 tbp1 = (struct buf *) tbp2->b_work;
844 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
846 } /* infinite loop (unless we're interrupted) */
847 } /* end of afs_BioDaemon() */
849 #else /* AFS_AIX41_ENV */
853 struct afs_bioqueue {
858 struct afs_bioqueue afs_bioqueue;
859 struct buf *afs_busyq = NULL;
860 struct buf *afs_asyncbuf;
861 afs_int32 afs_biodcnt = 0;
863 /* in implementing this, I assumed that all external linked lists were
866 * Several places in this code traverse a linked list. The algorithm
867 * used here is probably unfamiliar to most people. Careful examination
868 * will show that it eliminates an assignment inside the loop, as compared
869 * to the standard algorithm, at the cost of occasionally using an extra
875 * This function obtains, and returns, a pointer to a buffer for
876 * processing by a daemon. It sleeps until such a buffer is available.
877 * The source of buffers for it is the list afs_asyncbuf (see also
878 * naix_vm_strategy). This function may be invoked concurrently by
879 * several processes, that is, several instances of the same daemon.
880 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
881 * level, while get_bioreq runs at process level.
883 * The common kernel paradigm of sleeping and waking up, in which all the
884 * competing processes sleep waiting for wakeups on one address, is not
885 * followed here. Instead, the following paradigm is used: when a daemon
886 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
887 * it sleeps on the address of variable afs_asyncbuf. But if there is
888 * already a daemon sleeping on that address, it threads its own unique
889 * address onto a list, and sleeps on that address. This way, every
890 * sleeper is sleeping on a different address, and every wakeup wakes up
891 * exactly one daemon. This prevents a whole bunch of daemons from waking
892 * up and then immediately having to go back to sleep. This provides a
893 * performance gain and makes the I/O scheduling a bit more deterministic.
894 * The list of sleepers is variable afs_bioqueue. The unique address
895 * on which to sleep is passed to get_bioreq as its parameter.
897 /*static*/ struct buf *afs_get_bioreq(self)
898 struct afs_bioqueue *self; /* address on which to sleep */
901 struct buf *bp = (struct buf *) 0;
903 struct buf **bestlbpP, **lbpP;
905 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
908 struct afs_bioqueue *s;
910 /* ??? Does the forward pointer of the returned buffer need to be NULL?
913 /* Disable interrupts from the strategy function, and save the
914 * prior priority level
916 oldPriority = i_disable ( INTMAX ) ;
918 /* Each iteration of following loop either pulls
919 * a buffer off afs_asyncbuf, or sleeps.
921 while (1) { /* inner loop */
923 /* look for oldest buffer */
924 bp = bestbp = afs_asyncbuf;
925 bestage = (int) bestbp->av_back;
926 bestlbpP = &afs_asyncbuf;
931 if ((int) bp->av_back - bestage < 0) {
934 bestage = (int) bp->av_back;
938 *bestlbpP = bp->av_forw;
944 /* If afs_asyncbuf is null, it is necessary to go to sleep.
945 * There are two possibilities: either there is already a
946 * daemon that is sleeping on the address of afs_asyncbuf,
949 if (afs_bioqueue.sleeper) {
951 QAdd (&(afs_bioqueue.lruq), &(self->lruq));
952 interrupted = sleep ((caddr_t) self, PCATCH|(PZERO + 1));
953 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
954 QRemove (&(self->lruq)); /* dequeue */
957 afs_bioqueue.sleeper = FALSE;
959 /* re-enable interrupts from strategy */
960 i_enable (oldPriority);
965 afs_bioqueue.sleeper = TRUE;
966 interrupted = sleep ((caddr_t) &afs_asyncbuf, PCATCH|(PZERO + 1));
967 afs_bioqueue.sleeper = FALSE;
971 * We need to wakeup another daemon if present
972 * since we were waiting on afs_asyncbuf.
974 #ifdef notdef /* The following doesn't work as advertised */
975 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq)
977 struct squeue *bq = afs_bioqueue.lruq.next;
982 /* re-enable interrupts from strategy */
983 i_enable (oldPriority);
989 } /* end of "else asyncbuf is empty" */
990 } /* end of "inner loop" */
994 i_enable (oldPriority); /* re-enable interrupts from strategy */
996 /* For the convenience of other code, replace the gnodes in
997 * the b_vp field of bp and the other buffers on the b_work
998 * chain with the corresponding vnodes.
1000 * ??? what happens to the gnodes? They're not just cut loose,
1004 t2P = (struct buf *) t1P->b_work;
1005 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
1009 t1P = (struct buf *) t2P->b_work;
1010 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
1015 /* If the buffer does not specify I/O, it may immediately
1016 * be returned to the caller. This condition is detected
1017 * by examining the buffer's flags (the b_flags field). If
1018 * the B_PFPROT bit is set, the buffer represents a protection
1019 * violation, rather than a request for I/O. The remainder
1020 * of the outer loop handles the case where the B_PFPROT bit is clear.
1022 if (bp->b_flags & B_PFPROT) {
1026 /* wake up another process to handle the next buffer, and return
1029 oldPriority = i_disable ( INTMAX ) ;
1031 /* determine where to find the sleeping process.
1032 * There are two cases: either it is sleeping on
1033 * afs_asyncbuf, or it is sleeping on its own unique
1034 * address. These cases are distinguished by examining
1035 * the sleeper field of afs_bioqueue.
1037 if (afs_bioqueue.sleeper) {
1038 wakeup (&afs_asyncbuf);
1041 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1042 /* queue is empty, what now? ???*/
1043 /* Should this be impossible, or does */
1044 /* it just mean that nobody is sleeping? */;
1047 struct squeue *bq = afs_bioqueue.lruq.next;
1051 afs_bioqueue.sleeper = TRUE;
1054 i_enable (oldPriority); /* re-enable interrupts from strategy */
1057 } /* end of function get_bioreq() */
1062 * This function is the daemon. It is called from the syscall
1063 * interface. Ordinarily, a script or an administrator will run a
1064 * daemon startup utility, specifying the number of I/O daemons to
1065 * run. The utility will fork off that number of processes,
1066 * each making the appropriate syscall, which will cause this
1067 * function to be invoked.
1069 static int afs_initbiod = 0; /* this is self-initializing code */
1071 afs_BioDaemon (nbiods)
1074 struct afs_bioqueue *self;
1075 afs_int32 code, s, pflg = 0;
1077 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1082 if (!afs_initbiod) {
1085 /* Initialize the queue of waiting processes, afs_bioqueue. */
1086 QInit (&(afs_bioqueue.lruq));
1089 /* establish ourself as a kernel process so shutdown won't kill us */
1090 /* u.u_procp->p_flag |= SKPROC;*/
1092 /* Initialize a token (self) to use in the queue of sleeping processes. */
1093 self = (struct afs_bioqueue *) afs_osi_Alloc (sizeof (struct afs_bioqueue));
1094 pin (self, sizeof (struct afs_bioqueue)); /* fix in memory */
1095 memset(self, 0, sizeof(*self));
1096 QInit (&(self->lruq)); /* initialize queue entry pointers */
1099 /* Ignore HUP signals... */
1100 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1101 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1102 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1103 /* Main body starts here -- this is an intentional infinite loop, and
1106 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1107 * that we've been interrupted.
1110 bp = afs_get_bioreq(self);
1112 break; /* we were interrupted */
1113 if (code = setjmpx(&jmpbuf)) {
1114 /* This should not have happend, maybe a lack of resources */
1116 for (bp1 = bp; bp ; bp = bp1) {
1121 bp->b_flags |= B_ERROR;
1127 vcp = VTOAFS(bp->b_vp);
1128 if (bp->b_flags & B_PFSTORE) {
1129 ObtainWriteLock(&vcp->lock,210);
1130 if (vcp->v.v_gnode->gn_mwrcnt) {
1131 if (vcp->m.Length < bp->b_bcount + (u_int)dbtob(bp->b_blkno))
1132 vcp->m.Length = bp->b_bcount + (u_int)dbtob(bp->b_blkno);
1134 ReleaseWriteLock(&vcp->lock);
1136 /* If the buffer represents a protection violation, rather than
1137 * an actual request for I/O, no special action need be taken.
1139 if ( bp->b_flags & B_PFPROT ) {
1140 iodone (bp); /* Notify all users of the buffer that we're done */
1144 ObtainWriteLock(&vcp->pvmlock,558);
1146 * First map its data area to a region in the current address space
1147 * by calling vm_att with the subspace identifier, and a pointer to
1148 * the data area. vm_att returns a new data area pointer, but we
1149 * also want to hang onto the old one.
1151 tmpaddr = bp->b_baddr;
1152 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
1153 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1154 if (tmperr) { /* in non-error case */
1155 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1156 bp->b_error = tmperr;
1159 /* Unmap the buffer's data area by calling vm_det. Reset data area
1160 * to the value that we saved above.
1162 vm_det(bp->b_un.b_addr);
1163 bp->b_baddr = tmpaddr;
1166 * buffer may be linked with other buffers via the b_work field.
1167 * See also naix_vm_strategy. For each buffer in the chain (including
1168 * bp) notify all users of the buffer that the daemon is finished
1169 * using it by calling iodone.
1170 * assumes iodone can modify the b_work field.
1173 tbp2 = (struct buf *) tbp1->b_work;
1178 tbp1 = (struct buf *) tbp2->b_work;
1184 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1186 } /* infinite loop (unless we're interrupted) */
1187 unpin (self, sizeof (struct afs_bioqueue));
1188 afs_osi_Free (self, sizeof (struct afs_bioqueue));
1189 } /* end of afs_BioDaemon() */
1190 #endif /* AFS_AIX41_ENV */
1191 #endif /* AFS_AIX32_ENV */
1195 void afs_BackgroundDaemon() {
1196 struct brequest *tb;
1200 AFS_STATCNT(afs_BackgroundDaemon);
1201 /* initialize subsystem */
1203 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1204 memset((char *)afs_brs, 0, sizeof(afs_brs));
1206 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1208 * steal the first daemon for doing delayed DSlot flushing
1209 * (see afs_GetDownDSlot)
1218 MObtainWriteLock(&afs_xbrs,302);
1221 struct brequest *min_tb;
1223 if (afs_termState == AFSOP_STOP_BKG) {
1224 if (--afs_nbrs <= 0)
1225 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1226 MReleaseWriteLock(&afs_xbrs);
1227 afs_osi_Wakeup(&afs_termState);
1231 /* find a request */
1235 for(i=0; i<NBRS; i++, tb++) {
1236 /* look for request with smallest ts */
1237 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1238 /* new request, not yet picked up */
1239 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1246 /* claim and process this request */
1247 tb->flags |= BSTARTED;
1248 MReleaseWriteLock(&afs_xbrs);
1250 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1,
1251 ICL_TYPE_INT32, tb->opcode);
1252 if (tb->opcode == BOP_FETCH)
1254 else if (tb->opcode == BOP_STORE)
1256 else if (tb->opcode == BOP_PATH)
1258 else panic("background bop");
1261 tb->vnode->vrefCount--; /* fix up reference count */
1263 AFS_RELE((struct vnode *)(tb->vnode)); /* MUST call vnode layer or could lose vnodes */
1265 tb->vnode = (struct vcache *) 0;
1269 tb->cred = (struct AFS_UCRED *) 0;
1271 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1272 MObtainWriteLock(&afs_xbrs,305);
1275 /* wait for new request */
1277 MReleaseWriteLock(&afs_xbrs);
1278 afs_osi_Sleep(&afs_brsDaemons);
1279 MObtainWriteLock(&afs_xbrs,307);
1286 void shutdown_daemons()
1288 extern int afs_cold_shutdown;
1290 register struct brequest *tb;
1292 AFS_STATCNT(shutdown_daemons);
1293 if (afs_cold_shutdown) {
1294 afs_brsDaemons = brsInit = 0;
1295 rxepoch_checked = afs_nbrs = 0;
1296 memset((char *)afs_brs, 0, sizeof(afs_brs));
1297 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1299 #ifdef AFS_AIX32_ENV
1300 #ifdef AFS_AIX41_ENV
1301 lock_free(&afs_asyncbuf_lock);
1302 unpin(&afs_asyncbuf, sizeof(struct buf*));
1303 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
1304 #else /* AFS_AIX41_ENV */
1307 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1314 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1316 * sgi - daemon - handles certain operations that otherwise
1317 * would use up too much kernel stack space
1319 * This all assumes that since the caller must have the xdcache lock
1320 * exclusively that the list will never be more than one long
1321 * and noone else can attempt to add anything until we're done.
1323 SV_TYPE afs_sgibksync;
1324 SV_TYPE afs_sgibkwait;
1325 lock_t afs_sgibklock;
1326 struct dcache *afs_sgibklist;
1334 if (afs_sgibklock == NULL) {
1335 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1336 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1337 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1339 s = SPLOCK(afs_sgibklock);
1341 /* wait for something to do */
1342 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1343 osi_Assert(afs_sgibklist);
1345 /* XX will probably need to generalize to real list someday */
1346 s = SPLOCK(afs_sgibklock);
1347 while (afs_sgibklist) {
1348 tdc = afs_sgibklist;
1349 afs_sgibklist = NULL;
1350 SPUNLOCK(afs_sgibklock, s);
1352 tdc->dflags &= ~DFEntryMod;
1353 afs_WriteDCache(tdc, 1);
1355 s = SPLOCK(afs_sgibklock);
1358 /* done all the work - wake everyone up */
1359 while (SV_SIGNAL(&afs_sgibkwait))