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 "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 /* PAG garbage collection */
38 /* We induce a compile error if param.h does not define AFS_GCPAGS */
39 afs_int32 afs_gcpags=AFS_GCPAGS;
40 afs_int32 afs_gcpags_procsize = 0;
42 afs_int32 afs_CheckServerDaemonStarted = 0;
43 #ifdef DEFAULT_PROBE_INTERVAL
44 afs_int32 PROBE_INTERVAL=DEFAULT_PROBE_INTERVAL; /* overridding during compile */
46 afs_int32 PROBE_INTERVAL=180; /* default to 3 min */
49 #define PROBE_WAIT() (1000 * (PROBE_INTERVAL - ((afs_random() & 0x7fffffff) \
50 % (PROBE_INTERVAL/2))))
52 void afs_CheckServerDaemon(void)
54 afs_int32 now, delay, lastCheck, last10MinCheck;
56 afs_CheckServerDaemonStarted = 1;
58 while (afs_initState < 101) afs_osi_Sleep(&afs_initState);
59 afs_osi_Wait(PROBE_WAIT(), &AFS_CSWaitHandler, 0);
61 last10MinCheck = lastCheck = osi_Time();
63 if (afs_termState == AFSOP_STOP_CS) {
64 afs_termState = AFSOP_STOP_BKG;
65 afs_osi_Wakeup(&afs_termState);
70 if (PROBE_INTERVAL + lastCheck <= now) {
71 afs_CheckServers(1, NULL); /* check down servers */
72 lastCheck = now = osi_Time();
75 if (600 + last10MinCheck <= now) {
76 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, 600);
77 afs_CheckServers(0, NULL);
78 last10MinCheck = now = osi_Time();
81 if (afs_termState == AFSOP_STOP_CS) {
82 afs_termState = AFSOP_STOP_BKG;
83 afs_osi_Wakeup(&afs_termState);
87 /* Compute time to next probe. */
88 delay = PROBE_INTERVAL + lastCheck;
89 if (delay > 600 + last10MinCheck)
90 delay = 600 + last10MinCheck;
94 afs_osi_Wait(delay * 1000, &AFS_CSWaitHandler, 0);
96 afs_CheckServerDaemonStarted = 0;
102 struct afs_exporter *exporter;
104 afs_int32 last3MinCheck, last10MinCheck, last60MinCheck, lastNMinCheck;
105 afs_int32 last1MinCheck;
106 afs_uint32 lastCBSlotBump;
109 AFS_STATCNT(afs_Daemon);
110 last1MinCheck = last3MinCheck = last60MinCheck = last10MinCheck = lastNMinCheck = 0;
112 afs_rootFid.Fid.Volume = 0;
113 while (afs_initState < 101) afs_osi_Sleep(&afs_initState);
116 lastCBSlotBump = now;
118 /* when a lot of clients are booted simultaneously, they develop
119 * annoying synchronous VL server bashing behaviors. So we stagger them.
121 last1MinCheck = now + ((afs_random() & 0x7fffffff) % 60); /* an extra 30 */
122 last3MinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
123 last60MinCheck = now - 1800 + ((afs_random() & 0x7fffffff) % 3600);
124 last10MinCheck = now - 300 + ((afs_random() & 0x7fffffff) % 600);
125 lastNMinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
127 /* start off with afs_initState >= 101 (basic init done) */
129 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
131 /* things to do every 20 seconds or less - required by protocol spec */
133 afs_FlushActiveVcaches(0); /* flush NFS writes */
134 afs_FlushVCBs(1); /* flush queued callbacks */
135 afs_MaybeWakeupTruncateDaemon(); /* free cache space if have too */
136 rx_CheckPackets(); /* Does RX need more packets? */
137 #if defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)
139 * Hack: We always want to make sure there are plenty free
140 * entries in the small free pool so that we don't have to
141 * worry about rx (with disabled interrupts) to have to call
142 * malloc). So we do the dummy call below...
144 if (((afs_stats_cmperf.SmallBlocksAlloced - afs_stats_cmperf.SmallBlocksActive)
145 <= AFS_SALLOC_LOW_WATER))
146 osi_FreeSmallSpace(osi_AllocSmallSpace(AFS_SMALLOCSIZ));
147 if (((afs_stats_cmperf.MediumBlocksAlloced - afs_stats_cmperf.MediumBlocksActive)
148 <= AFS_MALLOC_LOW_WATER+50))
149 osi_AllocMoreMSpace(AFS_MALLOC_LOW_WATER * 2);
153 if (lastCBSlotBump + CBHTSLOTLEN < now) { /* pretty time-dependant */
154 lastCBSlotBump = now;
155 if (afs_BumpBase()) {
156 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
160 if (last1MinCheck + 60 < now) {
161 /* things to do every minute */
162 DFlush(); /* write out dir buffers */
163 afs_WriteThroughDSlots(); /* write through cacheinfo entries */
164 afs_FlushActiveVcaches(1);/* keep flocks held & flush nfs writes */
169 if (last3MinCheck + 180 < now) {
170 afs_CheckTokenCache(); /* check for access cache resets due to expired
174 if (!afs_CheckServerDaemonStarted) {
175 /* Do the check here if the correct afsd is not installed. */
178 printf("Please install afsd with check server daemon.\n");
180 if (lastNMinCheck + PROBE_INTERVAL < now) {
181 /* only check down servers */
182 afs_CheckServers(1, NULL);
186 if (last10MinCheck + 600 < now) {
187 #ifdef AFS_USERSPACE_IP_ADDR
188 extern int rxi_GetcbiInfo(void);
190 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP,
191 ICL_TYPE_INT32, 600);
192 #ifdef AFS_USERSPACE_IP_ADDR
193 if (rxi_GetcbiInfo()) { /* addresses changed from last time */
196 #else /* AFS_USERSPACE_IP_ADDR */
197 if (rxi_GetIFInfo()) { /* addresses changed from last time */
200 #endif /* else AFS_USERSPACE_IP_ADDR */
201 if (!afs_CheckServerDaemonStarted)
202 afs_CheckServers(0, NULL);
203 afs_GCUserData(0); /* gc old conns */
204 /* This is probably the wrong way of doing GC for the various exporters but it will suffice for a while */
205 for (exporter = root_exported; exporter; exporter = exporter->exp_next) {
206 (void) EXP_GC(exporter, 0); /* Generalize params */
211 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
215 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
220 last10MinCheck = now;
222 if (last60MinCheck + 3600 < now) {
223 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEVOLUME,
224 ICL_TYPE_INT32, 3600);
225 afs_CheckRootVolume();
227 if (afs_gcpags == AFS_GCPAGS_OK) {
232 last60MinCheck = now;
234 if (afs_initState < 300) { /* while things ain't rosy */
235 code = afs_CheckRootVolume();
236 if (code == 0) afs_initState = 300; /* succeeded */
237 if (afs_initState < 200) afs_initState = 200; /* tried once */
238 afs_osi_Wakeup(&afs_initState);
241 /* 18285 is because we're trying to divide evenly into 128, that is,
242 * CBSlotLen, while staying just under 20 seconds. If CBSlotLen
243 * changes, should probably change this interval, too.
244 * Some of the preceding actions may take quite some time, so we
245 * might not want to wait the entire interval */
246 now = 18285 - (osi_Time() - now);
248 afs_osi_Wait(now, &AFS_WaitHandler, 0);
251 if (afs_termState == AFSOP_STOP_AFS) {
252 if (afs_CheckServerDaemonStarted)
253 afs_termState = AFSOP_STOP_CS;
255 afs_termState = AFSOP_STOP_BKG;
256 afs_osi_Wakeup(&afs_termState);
262 int afs_CheckRootVolume (void)
264 char rootVolName[32];
265 struct volume *tvp = NULL;
266 int usingDynroot = afs_GetDynrootEnable();
269 AFS_STATCNT(afs_CheckRootVolume);
270 if (*afs_rootVolumeName == 0) {
271 strcpy(rootVolName, "root.afs");
273 strcpy(rootVolName, afs_rootVolumeName);
277 struct cell *lc = afs_GetPrimaryCell(READ_LOCK);
281 localcell = lc->cellNum;
282 afs_PutCell(lc, READ_LOCK);
286 afs_GetDynrootFid(&afs_rootFid);
287 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
289 tvp = afs_GetVolumeByName(rootVolName, localcell, 1, NULL, READ_LOCK);
291 if (!tvp && !usingDynroot) {
293 int len = strlen(rootVolName);
295 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
296 strcpy(buf, rootVolName);
297 afs_strcat(buf, ".readonly");
298 tvp = afs_GetVolumeByName(buf, localcell, 1, NULL, READ_LOCK);
303 int volid = (tvp->roVol? tvp->roVol : tvp->volume);
304 afs_rootFid.Cell = localcell;
305 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
307 /* If we had a root fid before and it changed location we reset
308 * the afs_globalVp so that it will be reevaluated.
309 * Just decrement the reference count. This only occurs during
310 * initial cell setup and can panic the machine if we set the
311 * count to zero and fs checkv is executed when the current
314 AFS_FAST_RELE(afs_globalVp);
317 afs_rootFid.Fid.Volume = volid;
318 afs_rootFid.Fid.Vnode = 1;
319 afs_rootFid.Fid.Unique = 1;
321 afs_initState = 300; /* won */
322 afs_osi_Wakeup(&afs_initState);
323 afs_PutVolume(tvp, READ_LOCK);
326 /* This is to make sure that we update the root gnode */
327 /* every time root volume gets released */
329 struct gnode *rootgp;
333 /* Only do this if afs_globalVFS is properly set due to race conditions
334 this routine could be called before the gfs_mount is performed!
335 Furthermore, afs_root (called below) *waits* until
336 initState >= 200, so we don't try this until we've gotten
338 if (afs_globalVFS && afs_initState >= 200) {
339 if (code = afs_root(afs_globalVFS, &rootgp))
341 mp = (struct mount *) afs_globalVFS->vfs_data ;
342 mp->m_rootgp = gget(mp, 0, 0, (char *)rootgp);
343 afs_unlock(mp->m_rootgp); /* unlock basic gnode */
344 afs_vrele(VTOAFS(rootgp)); /* zap afs_root's vnode hold */
348 if (afs_rootFid.Fid.Volume) return 0;
352 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
353 static void BPath(register struct brequest *ab)
355 register struct dcache *tdc = NULL;
356 struct vcache *tvc = NULL;
357 struct vnode *tvn = NULL;
358 #ifdef AFS_LINUX22_ENV
359 struct dentry *dp = NULL;
361 afs_size_t offset, len;
362 struct vrequest treq;
366 if ((code = afs_InitReq(&treq, ab->cred))) return;
368 #ifdef AFS_LINUX22_ENV
369 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, NULL, &dp);
371 tvn = (struct vnode*)dp->d_inode;
373 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, NULL, (struct vnode **)&tvn);
376 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
378 /* now path may not have been in afs, so check that before calling our cache manager */
379 if (!tvn || !IsAfsVnode((struct vnode *) tvn)) {
380 /* release it and give up */
385 #ifdef AFS_LINUX22_ENV
388 AFS_RELE((struct vnode *) tvn);
395 tvc = VTOAFS(afs_gntovn(tvn));
399 /* here we know its an afs vnode, so we can get the data for the chunk */
400 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
407 #ifdef AFS_LINUX22_ENV
410 AFS_RELE((struct vnode *) tvn);
415 /* size_parm 0 to the fetch is the chunk number,
416 * ptr_parm 0 is the dcache entry to wakeup,
417 * size_parm 1 is true iff we should release the dcache entry here.
419 static void BPrefetch(register struct brequest *ab)
421 register struct dcache *tdc;
422 register struct vcache *tvc;
423 afs_size_t offset, len;
424 struct vrequest treq;
426 AFS_STATCNT(BPrefetch);
427 if ((len = afs_InitReq(&treq, ab->cred))) return;
429 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
433 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
434 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
435 * be waiting for our wakeup anyway.
437 tdc = (struct dcache *) (ab->ptr_parm[0]);
438 ObtainSharedLock(&tdc->lock, 640);
439 if (tdc->mflags & DFFetchReq) {
440 UpgradeSToWLock(&tdc->lock, 641);
441 tdc->mflags &= ~DFFetchReq;
442 ReleaseWriteLock(&tdc->lock);
444 ReleaseSharedLock(&tdc->lock);
446 afs_osi_Wakeup(&tdc->validPos);
447 if (ab->size_parm[1]) {
448 afs_PutDCache(tdc); /* put this one back, too */
453 static void BStore(register struct brequest *ab)
455 register struct vcache *tvc;
456 register afs_int32 code;
457 struct vrequest treq;
458 #if defined(AFS_SGI_ENV)
459 struct cred *tmpcred;
463 if ((code = afs_InitReq(&treq, ab->cred))) return;
466 #if defined(AFS_SGI_ENV)
468 * Since StoreOnLastReference can end up calling osi_SyncVM which
469 * calls into VM code that assumes that u.u_cred has the
470 * correct credentials, we set our to theirs for this xaction
472 tmpcred = OSI_GET_CURRENT_CRED();
473 OSI_SET_CURRENT_CRED(ab->cred);
476 * To avoid recursion since the WriteLock may be released during VM
477 * operations, we hold the VOP_RWLOCK across this transaction as
478 * do the other callers of StoreOnLastReference
480 AFS_RWLOCK((vnode_t *)tvc, 1);
482 ObtainWriteLock(&tvc->lock,209);
483 code = afs_StoreOnLastReference(tvc, &treq);
484 ReleaseWriteLock(&tvc->lock);
485 #if defined(AFS_SGI_ENV)
486 OSI_SET_CURRENT_CRED(tmpcred);
487 AFS_RWUNLOCK((vnode_t *)tvc, 1);
489 /* now set final return code, and wakeup anyone waiting */
490 if ((ab->flags & BUVALID) == 0) {
491 ab->code = afs_CheckCode(code, &treq, 43); /* set final code, since treq doesn't go across processes */
492 ab->flags |= BUVALID;
493 if (ab->flags & BUWAIT) {
494 ab->flags &= ~BUWAIT;
500 /* release a held request buffer */
501 void afs_BRelease(register struct brequest *ab)
504 AFS_STATCNT(afs_BRelease);
505 MObtainWriteLock(&afs_xbrs,294);
506 if (--ab->refCount <= 0) {
509 if (afs_brsWaiters) afs_osi_Wakeup(&afs_brsWaiters);
510 MReleaseWriteLock(&afs_xbrs);
513 /* return true if bkg fetch daemons are all busy */
516 AFS_STATCNT(afs_BBusy);
517 if (afs_brsDaemons > 0) return 0;
521 struct brequest *afs_BQueue(register short aopcode, register struct vcache *avc,
522 afs_int32 dontwait, afs_int32 ause, struct AFS_UCRED *acred,
523 afs_size_t asparm0, afs_size_t asparm1, void *apparm0)
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 #if defined(AFS_NETBSD_ENV) || defined(AFS_OBSD_ENV)
546 AFS_HOLD(AFSTOV(avc));
548 VN_HOLD(AFSTOV(avc));
552 tb->refCount = ause+1;
553 tb->size_parm[0] = asparm0;
554 tb->size_parm[1] = asparm1;
555 tb->ptr_parm[0] = apparm0;
558 tb->ts = afs_brs_count++;
559 /* if daemons are waiting for work, wake them up */
560 if (afs_brsDaemons > 0) {
561 afs_osi_Wakeup(&afs_brsDaemons);
563 MReleaseWriteLock(&afs_xbrs);
567 MReleaseWriteLock(&afs_xbrs);
570 /* no free buffers, sleep a while */
572 MReleaseWriteLock(&afs_xbrs);
573 afs_osi_Sleep(&afs_brsWaiters);
574 MObtainWriteLock(&afs_xbrs,301);
581 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
582 * The modifications here will work for either a UP or MP machine.
584 struct buf *afs_asyncbuf = (struct buf*)0;
585 afs_int32 afs_asyncbuf_cv = EVENT_NULL;
586 afs_int32 afs_biodcnt = 0;
588 /* in implementing this, I assumed that all external linked lists were
591 * Several places in this code traverse a linked list. The algorithm
592 * used here is probably unfamiliar to most people. Careful examination
593 * will show that it eliminates an assignment inside the loop, as compared
594 * to the standard algorithm, at the cost of occasionally using an extra
600 * This function obtains, and returns, a pointer to a buffer for
601 * processing by a daemon. It sleeps until such a buffer is available.
602 * The source of buffers for it is the list afs_asyncbuf (see also
603 * naix_vm_strategy). This function may be invoked concurrently by
604 * several processes, that is, several instances of the same daemon.
605 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
606 * level, while get_bioreq runs at process level.
608 * Since AIX 4.1 can wake just one process at a time, the separate sleep
609 * addresses have been removed.
610 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
611 * The afs_asyncbuf_lock is primarily used to serialize access between
612 * process and interrupts.
614 Simple_lock afs_asyncbuf_lock;
615 /*static*/ struct buf *afs_get_bioreq()
617 struct buf *bp = NULL;
619 struct buf **bestlbpP, **lbpP;
621 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
624 struct afs_bioqueue *s;
626 /* ??? Does the forward pointer of the returned buffer need to be NULL?
629 /* Disable interrupts from the strategy function, and save the
630 * prior priority level and lock access to the afs_asyncbuf.
633 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock) ;
637 /* look for oldest buffer */
638 bp = bestbp = afs_asyncbuf;
639 bestage = (int) bestbp->av_back;
640 bestlbpP = &afs_asyncbuf;
645 if ((int) bp->av_back - bestage < 0) {
648 bestage = (int) bp->av_back;
652 *bestlbpP = bp->av_forw;
656 /* If afs_asyncbuf is null, it is necessary to go to sleep.
657 * e_wakeup_one() ensures that only one thread wakes.
660 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
661 * lock on an MP machine.
663 interrupted = e_sleep_thread(&afs_asyncbuf_cv,
665 LOCK_HANDLER|INTERRUPTIBLE);
666 if (interrupted==THREAD_INTERRUPTED) {
667 /* re-enable interrupts from strategy */
668 unlock_enable(oldPriority, &afs_asyncbuf_lock);
672 } /* end of "else asyncbuf is empty" */
673 } /* end of "inner loop" */
677 unlock_enable(oldPriority, &afs_asyncbuf_lock);
680 /* For the convenience of other code, replace the gnodes in
681 * the b_vp field of bp and the other buffers on the b_work
682 * chain with the corresponding vnodes.
684 * ??? what happens to the gnodes? They're not just cut loose,
688 t2P = (struct buf *) t1P->b_work;
689 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
693 t1P = (struct buf *) t2P->b_work;
694 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
699 /* If the buffer does not specify I/O, it may immediately
700 * be returned to the caller. This condition is detected
701 * by examining the buffer's flags (the b_flags field). If
702 * the B_PFPROT bit is set, the buffer represents a protection
703 * violation, rather than a request for I/O. The remainder
704 * of the outer loop handles the case where the B_PFPROT bit is clear.
706 if (bp->b_flags & B_PFPROT) {
711 } /* end of function get_bioreq() */
716 * This function is the daemon. It is called from the syscall
717 * interface. Ordinarily, a script or an administrator will run a
718 * daemon startup utility, specifying the number of I/O daemons to
719 * run. The utility will fork off that number of processes,
720 * each making the appropriate syscall, which will cause this
721 * function to be invoked.
723 static int afs_initbiod = 0; /* this is self-initializing code */
725 int afs_BioDaemon (afs_int32 nbiods)
727 afs_int32 code, s, pflg = 0;
729 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
737 /* pin lock, since we'll be using it in an interrupt. */
738 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
739 simple_lock_init(&afs_asyncbuf_lock);
740 pin (&afs_asyncbuf, sizeof(struct buf*));
741 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
744 /* Ignore HUP signals... */
746 sigset_t sigbits, osigbits;
748 * add SIGHUP to the set of already masked signals
750 SIGFILLSET(sigbits); /* allow all signals */
751 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
752 limit_sigs(&sigbits, &osigbits); /* and already masked */
754 /* Main body starts here -- this is an intentional infinite loop, and
757 * Now, the loop will exit if get_bioreq() returns NULL, indicating
758 * that we've been interrupted.
761 bp = afs_get_bioreq();
763 break; /* we were interrupted */
764 if (code = setjmpx(&jmpbuf)) {
765 /* This should not have happend, maybe a lack of resources */
767 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
768 for (bp1 = bp; bp ; bp = bp1) {
770 bp1 = (struct buf *) bp1->b_work;
773 bp->b_flags |= B_ERROR;
776 unlock_enable(s, &afs_asyncbuf_lock);
780 vcp = VTOAFS(bp->b_vp);
781 if (bp->b_flags & B_PFSTORE) { /* XXXX */
782 ObtainWriteLock(&vcp->lock,404);
783 if (vcp->v.v_gnode->gn_mwrcnt) {
784 afs_offs_t newlength =
785 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
786 if (vcp->m.Length < newlength) {
787 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
788 ICL_TYPE_STRING, __FILE__,
789 ICL_TYPE_LONG, __LINE__,
790 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(vcp->m.Length),
791 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
792 vcp->m.Length = newlength;
795 ReleaseWriteLock(&vcp->lock);
797 /* If the buffer represents a protection violation, rather than
798 * an actual request for I/O, no special action need be taken.
800 if ( bp->b_flags & B_PFPROT ) {
801 iodone (bp); /* Notify all users of the buffer that we're done */
806 ObtainWriteLock(&vcp->pvmlock,211);
808 * First map its data area to a region in the current address space
809 * by calling vm_att with the subspace identifier, and a pointer to
810 * the data area. vm_att returns a new data area pointer, but we
811 * also want to hang onto the old one.
813 tmpaddr = bp->b_baddr;
814 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
815 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
816 if (tmperr) { /* in non-error case */
817 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
818 bp->b_error = tmperr;
821 /* Unmap the buffer's data area by calling vm_det. Reset data area
822 * to the value that we saved above.
824 vm_det(bp->b_un.b_addr);
825 bp->b_baddr = tmpaddr;
828 * buffer may be linked with other buffers via the b_work field.
829 * See also naix_vm_strategy. For each buffer in the chain (including
830 * bp) notify all users of the buffer that the daemon is finished
831 * using it by calling iodone.
832 * assumes iodone can modify the b_work field.
835 tbp2 = (struct buf *) tbp1->b_work;
840 tbp1 = (struct buf *) tbp2->b_work;
846 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
848 } /* infinite loop (unless we're interrupted) */
849 } /* end of afs_BioDaemon() */
851 #else /* AFS_AIX41_ENV */
855 struct afs_bioqueue {
860 struct afs_bioqueue afs_bioqueue;
861 struct buf *afs_busyq = NULL;
862 struct buf *afs_asyncbuf;
863 afs_int32 afs_biodcnt = 0;
865 /* in implementing this, I assumed that all external linked lists were
868 * Several places in this code traverse a linked list. The algorithm
869 * used here is probably unfamiliar to most people. Careful examination
870 * will show that it eliminates an assignment inside the loop, as compared
871 * to the standard algorithm, at the cost of occasionally using an extra
877 * This function obtains, and returns, a pointer to a buffer for
878 * processing by a daemon. It sleeps until such a buffer is available.
879 * The source of buffers for it is the list afs_asyncbuf (see also
880 * naix_vm_strategy). This function may be invoked concurrently by
881 * several processes, that is, several instances of the same daemon.
882 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
883 * level, while get_bioreq runs at process level.
885 * The common kernel paradigm of sleeping and waking up, in which all the
886 * competing processes sleep waiting for wakeups on one address, is not
887 * followed here. Instead, the following paradigm is used: when a daemon
888 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
889 * it sleeps on the address of variable afs_asyncbuf. But if there is
890 * already a daemon sleeping on that address, it threads its own unique
891 * address onto a list, and sleeps on that address. This way, every
892 * sleeper is sleeping on a different address, and every wakeup wakes up
893 * exactly one daemon. This prevents a whole bunch of daemons from waking
894 * up and then immediately having to go back to sleep. This provides a
895 * performance gain and makes the I/O scheduling a bit more deterministic.
896 * The list of sleepers is variable afs_bioqueue. The unique address
897 * on which to sleep is passed to get_bioreq as its parameter.
899 /*static*/ struct buf *afs_get_bioreq(self)
900 struct afs_bioqueue *self; /* address on which to sleep */
903 struct buf *bp = NULL;
905 struct buf **bestlbpP, **lbpP;
907 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
910 struct afs_bioqueue *s;
912 /* ??? Does the forward pointer of the returned buffer need to be NULL?
915 /* Disable interrupts from the strategy function, and save the
916 * prior priority level
918 oldPriority = i_disable ( INTMAX ) ;
920 /* Each iteration of following loop either pulls
921 * a buffer off afs_asyncbuf, or sleeps.
923 while (1) { /* inner loop */
925 /* look for oldest buffer */
926 bp = bestbp = afs_asyncbuf;
927 bestage = (int) bestbp->av_back;
928 bestlbpP = &afs_asyncbuf;
933 if ((int) bp->av_back - bestage < 0) {
936 bestage = (int) bp->av_back;
940 *bestlbpP = bp->av_forw;
946 /* If afs_asyncbuf is null, it is necessary to go to sleep.
947 * There are two possibilities: either there is already a
948 * daemon that is sleeping on the address of afs_asyncbuf,
951 if (afs_bioqueue.sleeper) {
953 QAdd (&(afs_bioqueue.lruq), &(self->lruq));
954 interrupted = sleep ((caddr_t) self, PCATCH|(PZERO + 1));
955 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
956 QRemove (&(self->lruq)); /* dequeue */
959 afs_bioqueue.sleeper = FALSE;
961 /* re-enable interrupts from strategy */
962 i_enable (oldPriority);
967 afs_bioqueue.sleeper = TRUE;
968 interrupted = sleep ((caddr_t) &afs_asyncbuf, PCATCH|(PZERO + 1));
969 afs_bioqueue.sleeper = FALSE;
973 * We need to wakeup another daemon if present
974 * since we were waiting on afs_asyncbuf.
976 #ifdef notdef /* The following doesn't work as advertised */
977 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq)
979 struct squeue *bq = afs_bioqueue.lruq.next;
984 /* re-enable interrupts from strategy */
985 i_enable (oldPriority);
991 } /* end of "else asyncbuf is empty" */
992 } /* end of "inner loop" */
996 i_enable (oldPriority); /* re-enable interrupts from strategy */
998 /* For the convenience of other code, replace the gnodes in
999 * the b_vp field of bp and the other buffers on the b_work
1000 * chain with the corresponding vnodes.
1002 * ??? what happens to the gnodes? They're not just cut loose,
1006 t2P = (struct buf *) t1P->b_work;
1007 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
1011 t1P = (struct buf *) t2P->b_work;
1012 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
1017 /* If the buffer does not specify I/O, it may immediately
1018 * be returned to the caller. This condition is detected
1019 * by examining the buffer's flags (the b_flags field). If
1020 * the B_PFPROT bit is set, the buffer represents a protection
1021 * violation, rather than a request for I/O. The remainder
1022 * of the outer loop handles the case where the B_PFPROT bit is clear.
1024 if (bp->b_flags & B_PFPROT) {
1028 /* wake up another process to handle the next buffer, and return
1031 oldPriority = i_disable ( INTMAX ) ;
1033 /* determine where to find the sleeping process.
1034 * There are two cases: either it is sleeping on
1035 * afs_asyncbuf, or it is sleeping on its own unique
1036 * address. These cases are distinguished by examining
1037 * the sleeper field of afs_bioqueue.
1039 if (afs_bioqueue.sleeper) {
1040 wakeup (&afs_asyncbuf);
1043 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1044 /* queue is empty, what now? ???*/
1045 /* Should this be impossible, or does */
1046 /* it just mean that nobody is sleeping? */;
1049 struct squeue *bq = afs_bioqueue.lruq.next;
1053 afs_bioqueue.sleeper = TRUE;
1056 i_enable (oldPriority); /* re-enable interrupts from strategy */
1059 } /* end of function get_bioreq() */
1064 * This function is the daemon. It is called from the syscall
1065 * interface. Ordinarily, a script or an administrator will run a
1066 * daemon startup utility, specifying the number of I/O daemons to
1067 * run. The utility will fork off that number of processes,
1068 * each making the appropriate syscall, which will cause this
1069 * function to be invoked.
1071 static int afs_initbiod = 0; /* this is self-initializing code */
1073 afs_BioDaemon (nbiods)
1076 struct afs_bioqueue *self;
1077 afs_int32 code, s, pflg = 0;
1079 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1084 if (!afs_initbiod) {
1087 /* Initialize the queue of waiting processes, afs_bioqueue. */
1088 QInit (&(afs_bioqueue.lruq));
1091 /* establish ourself as a kernel process so shutdown won't kill us */
1092 /* u.u_procp->p_flag |= SKPROC;*/
1094 /* Initialize a token (self) to use in the queue of sleeping processes. */
1095 self = (struct afs_bioqueue *) afs_osi_Alloc (sizeof (struct afs_bioqueue));
1096 pin (self, sizeof (struct afs_bioqueue)); /* fix in memory */
1097 memset(self, 0, sizeof(*self));
1098 QInit (&(self->lruq)); /* initialize queue entry pointers */
1101 /* Ignore HUP signals... */
1102 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1103 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1104 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1105 /* Main body starts here -- this is an intentional infinite loop, and
1108 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1109 * that we've been interrupted.
1112 bp = afs_get_bioreq(self);
1114 break; /* we were interrupted */
1115 if (code = setjmpx(&jmpbuf)) {
1116 /* This should not have happend, maybe a lack of resources */
1118 for (bp1 = bp; bp ; bp = bp1) {
1123 bp->b_flags |= B_ERROR;
1129 vcp = VTOAFS(bp->b_vp);
1130 if (bp->b_flags & B_PFSTORE) {
1131 ObtainWriteLock(&vcp->lock,210);
1132 if (vcp->v.v_gnode->gn_mwrcnt) {
1133 afs_offs_t newlength =
1134 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
1135 if (vcp->m.Length < newlength) {
1136 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
1137 ICL_TYPE_STRING, __FILE__,
1138 ICL_TYPE_LONG, __LINE__,
1139 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(vcp->m.Length),
1140 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
1141 vcp->m.Length = newlength;
1144 ReleaseWriteLock(&vcp->lock);
1146 /* If the buffer represents a protection violation, rather than
1147 * an actual request for I/O, no special action need be taken.
1149 if ( bp->b_flags & B_PFPROT ) {
1150 iodone (bp); /* Notify all users of the buffer that we're done */
1154 ObtainWriteLock(&vcp->pvmlock,558);
1156 * First map its data area to a region in the current address space
1157 * by calling vm_att with the subspace identifier, and a pointer to
1158 * the data area. vm_att returns a new data area pointer, but we
1159 * also want to hang onto the old one.
1161 tmpaddr = bp->b_baddr;
1162 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
1163 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1164 if (tmperr) { /* in non-error case */
1165 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1166 bp->b_error = tmperr;
1169 /* Unmap the buffer's data area by calling vm_det. Reset data area
1170 * to the value that we saved above.
1172 vm_det(bp->b_un.b_addr);
1173 bp->b_baddr = tmpaddr;
1176 * buffer may be linked with other buffers via the b_work field.
1177 * See also naix_vm_strategy. For each buffer in the chain (including
1178 * bp) notify all users of the buffer that the daemon is finished
1179 * using it by calling iodone.
1180 * assumes iodone can modify the b_work field.
1183 tbp2 = (struct buf *) tbp1->b_work;
1188 tbp1 = (struct buf *) tbp2->b_work;
1194 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1196 } /* infinite loop (unless we're interrupted) */
1197 unpin (self, sizeof (struct afs_bioqueue));
1198 afs_osi_Free (self, sizeof (struct afs_bioqueue));
1199 } /* end of afs_BioDaemon() */
1200 #endif /* AFS_AIX41_ENV */
1201 #endif /* AFS_AIX32_ENV */
1205 void afs_BackgroundDaemon(void)
1207 struct brequest *tb;
1210 AFS_STATCNT(afs_BackgroundDaemon);
1211 /* initialize subsystem */
1213 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1214 memset((char *)afs_brs, 0, sizeof(afs_brs));
1216 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1218 * steal the first daemon for doing delayed DSlot flushing
1219 * (see afs_GetDownDSlot)
1228 MObtainWriteLock(&afs_xbrs,302);
1231 struct brequest *min_tb = NULL;
1233 if (afs_termState == AFSOP_STOP_BKG) {
1234 if (--afs_nbrs <= 0)
1235 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1236 MReleaseWriteLock(&afs_xbrs);
1237 afs_osi_Wakeup(&afs_termState);
1241 /* find a request */
1244 for(i=0; i<NBRS; i++, tb++) {
1245 /* look for request with smallest ts */
1246 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1247 /* new request, not yet picked up */
1248 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1254 if ((tb = min_tb)) {
1255 /* claim and process this request */
1256 tb->flags |= BSTARTED;
1257 MReleaseWriteLock(&afs_xbrs);
1259 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1,
1260 ICL_TYPE_INT32, tb->opcode);
1261 if (tb->opcode == BOP_FETCH)
1263 else if (tb->opcode == BOP_STORE)
1265 else if (tb->opcode == BOP_PATH)
1267 else panic("background bop");
1270 tb->vnode->vrefCount--; /* fix up reference count */
1272 AFS_RELE((struct vnode *)(tb->vnode)); /* MUST call vnode layer or could lose vnodes */
1278 tb->cred = (struct AFS_UCRED *) 0;
1280 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1281 MObtainWriteLock(&afs_xbrs,305);
1284 /* wait for new request */
1286 MReleaseWriteLock(&afs_xbrs);
1287 afs_osi_Sleep(&afs_brsDaemons);
1288 MObtainWriteLock(&afs_xbrs,307);
1295 void shutdown_daemons(void)
1297 AFS_STATCNT(shutdown_daemons);
1298 if (afs_cold_shutdown) {
1299 afs_brsDaemons = brsInit = 0;
1300 rxepoch_checked = afs_nbrs = 0;
1301 memset((char *)afs_brs, 0, sizeof(afs_brs));
1302 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1304 #ifdef AFS_AIX32_ENV
1305 #ifdef AFS_AIX41_ENV
1306 lock_free(&afs_asyncbuf_lock);
1307 unpin(&afs_asyncbuf, sizeof(struct buf*));
1308 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
1309 #else /* AFS_AIX41_ENV */
1312 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1319 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1321 * sgi - daemon - handles certain operations that otherwise
1322 * would use up too much kernel stack space
1324 * This all assumes that since the caller must have the xdcache lock
1325 * exclusively that the list will never be more than one long
1326 * and noone else can attempt to add anything until we're done.
1328 SV_TYPE afs_sgibksync;
1329 SV_TYPE afs_sgibkwait;
1330 lock_t afs_sgibklock;
1331 struct dcache *afs_sgibklist;
1333 int afs_sgidaemon(void)
1338 if (afs_sgibklock == NULL) {
1339 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1340 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1341 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1343 s = SPLOCK(afs_sgibklock);
1345 /* wait for something to do */
1346 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1347 osi_Assert(afs_sgibklist);
1349 /* XX will probably need to generalize to real list someday */
1350 s = SPLOCK(afs_sgibklock);
1351 while (afs_sgibklist) {
1352 tdc = afs_sgibklist;
1353 afs_sgibklist = NULL;
1354 SPUNLOCK(afs_sgibklock, s);
1356 tdc->dflags &= ~DFEntryMod;
1357 afs_WriteDCache(tdc, 1);
1359 s = SPLOCK(afs_sgibklock);
1362 /* done all the work - wake everyone up */
1363 while (SV_SIGNAL(&afs_sgibkwait))