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 /* 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;
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 int 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 register struct volume *tvp;
266 int usingDynroot = afs_GetDynrootEnable();
268 AFS_STATCNT(afs_CheckRootVolume);
269 if (*afs_rootVolumeName == 0) {
270 strcpy(rootVolName, "root.afs");
273 strcpy(rootVolName, afs_rootVolumeName);
276 afs_GetDynrootFid(&afs_rootFid);
277 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
279 tvp = afs_GetVolumeByName(rootVolName, LOCALCELL, 1, NULL, READ_LOCK);
283 int len = strlen(rootVolName);
285 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
286 strcpy(buf, rootVolName);
287 afs_strcat(buf, ".readonly");
288 tvp = afs_GetVolumeByName(buf, LOCALCELL, 1, NULL, READ_LOCK);
293 int volid = (tvp->roVol? tvp->roVol : tvp->volume);
294 afs_rootFid.Cell = LOCALCELL;
295 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
297 /* If we had a root fid before and it changed location we reset
298 * the afs_globalVp so that it will be reevaluated.
299 * Just decrement the reference count. This only occurs during
300 * initial cell setup and can panic the machine if we set the
301 * count to zero and fs checkv is executed when the current
304 AFS_FAST_RELE(afs_globalVp);
307 afs_rootFid.Fid.Volume = volid;
308 afs_rootFid.Fid.Vnode = 1;
309 afs_rootFid.Fid.Unique = 1;
311 afs_initState = 300; /* won */
312 afs_osi_Wakeup(&afs_initState);
313 afs_PutVolume(tvp, READ_LOCK);
316 /* This is to make sure that we update the root gnode */
317 /* every time root volume gets released */
319 struct gnode *rootgp;
323 /* Only do this if afs_globalVFS is properly set due to race conditions
324 this routine could be called before the gfs_mount is performed!
325 Furthermore, afs_root (called below) *waits* until
326 initState >= 200, so we don't try this until we've gotten
328 if (afs_globalVFS && afs_initState >= 200) {
329 if (code = afs_root(afs_globalVFS, &rootgp))
331 mp = (struct mount *) afs_globalVFS->vfs_data ;
332 mp->m_rootgp = gget(mp, 0, 0, (char *)rootgp);
333 afs_unlock(mp->m_rootgp); /* unlock basic gnode */
334 afs_vrele(VTOAFS(rootgp)); /* zap afs_root's vnode hold */
338 if (afs_rootFid.Fid.Volume) return 0;
342 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
343 static void BPath(register struct brequest *ab)
345 register struct dcache *tdc = NULL;
346 struct vcache *tvc = NULL;
347 struct vnode *tvn = NULL;
348 #ifdef AFS_LINUX22_ENV
349 struct dentry *dp = NULL;
351 afs_size_t offset, len;
352 struct vrequest treq;
356 if ((code = afs_InitReq(&treq, ab->cred))) return;
358 #ifdef AFS_LINUX22_ENV
359 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, NULL, &dp);
361 tvn = (struct vnode*)dp->d_inode;
363 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, NULL, (struct vnode **)&tvn);
366 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
368 /* now path may not have been in afs, so check that before calling our cache manager */
369 if (!tvn || !IsAfsVnode((struct vnode *) tvn)) {
370 /* release it and give up */
375 #ifdef AFS_LINUX22_ENV
378 AFS_RELE((struct vnode *) tvn);
385 tvc = VTOAFS(afs_gntovn(tvn));
389 /* here we know its an afs vnode, so we can get the data for the chunk */
390 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
397 #ifdef AFS_LINUX22_ENV
400 AFS_RELE((struct vnode *) tvn);
405 /* size_parm 0 to the fetch is the chunk number,
406 * ptr_parm 0 is the dcache entry to wakeup,
407 * size_parm 1 is true iff we should release the dcache entry here.
409 static void BPrefetch(register struct brequest *ab)
411 register struct dcache *tdc;
412 register struct vcache *tvc;
413 afs_size_t offset, len;
414 struct vrequest treq;
416 AFS_STATCNT(BPrefetch);
417 if ((len = afs_InitReq(&treq, ab->cred))) return;
419 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
423 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
424 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
425 * be waiting for our wakeup anyway.
427 tdc = (struct dcache *) (ab->ptr_parm[0]);
428 ObtainSharedLock(&tdc->lock, 640);
429 if (tdc->mflags & DFFetchReq) {
430 UpgradeSToWLock(&tdc->lock, 641);
431 tdc->mflags &= ~DFFetchReq;
432 ReleaseWriteLock(&tdc->lock);
434 ReleaseSharedLock(&tdc->lock);
436 afs_osi_Wakeup(&tdc->validPos);
437 if (ab->size_parm[1]) {
438 afs_PutDCache(tdc); /* put this one back, too */
443 static void BStore(register struct brequest *ab)
445 register struct vcache *tvc;
446 register afs_int32 code;
447 struct vrequest treq;
448 #if defined(AFS_SGI_ENV)
449 struct cred *tmpcred;
453 if ((code = afs_InitReq(&treq, ab->cred))) return;
456 #if defined(AFS_SGI_ENV)
458 * Since StoreOnLastReference can end up calling osi_SyncVM which
459 * calls into VM code that assumes that u.u_cred has the
460 * correct credentials, we set our to theirs for this xaction
462 tmpcred = OSI_GET_CURRENT_CRED();
463 OSI_SET_CURRENT_CRED(ab->cred);
466 * To avoid recursion since the WriteLock may be released during VM
467 * operations, we hold the VOP_RWLOCK across this transaction as
468 * do the other callers of StoreOnLastReference
470 AFS_RWLOCK((vnode_t *)tvc, 1);
472 ObtainWriteLock(&tvc->lock,209);
473 code = afs_StoreOnLastReference(tvc, &treq);
474 ReleaseWriteLock(&tvc->lock);
475 #if defined(AFS_SGI_ENV)
476 OSI_SET_CURRENT_CRED(tmpcred);
477 AFS_RWUNLOCK((vnode_t *)tvc, 1);
479 /* now set final return code, and wakeup anyone waiting */
480 if ((ab->flags & BUVALID) == 0) {
481 ab->code = afs_CheckCode(code, &treq, 43); /* set final code, since treq doesn't go across processes */
482 ab->flags |= BUVALID;
483 if (ab->flags & BUWAIT) {
484 ab->flags &= ~BUWAIT;
490 /* release a held request buffer */
491 void afs_BRelease(register struct brequest *ab)
494 AFS_STATCNT(afs_BRelease);
495 MObtainWriteLock(&afs_xbrs,294);
496 if (--ab->refCount <= 0) {
499 if (afs_brsWaiters) afs_osi_Wakeup(&afs_brsWaiters);
500 MReleaseWriteLock(&afs_xbrs);
503 /* return true if bkg fetch daemons are all busy */
506 AFS_STATCNT(afs_BBusy);
507 if (afs_brsDaemons > 0) return 0;
511 struct brequest *afs_BQueue(register short aopcode, register struct vcache *avc,
512 afs_int32 dontwait, afs_int32 ause, struct AFS_UCRED *acred,
513 afs_size_t asparm0, afs_size_t asparm1, void *apparm0)
516 register struct brequest *tb;
518 AFS_STATCNT(afs_BQueue);
519 MObtainWriteLock(&afs_xbrs,296);
522 for(i=0;i<NBRS;i++,tb++) {
523 if (tb->refCount == 0) break;
527 tb->opcode = aopcode;
535 VN_HOLD(AFSTOV(avc));
538 tb->refCount = ause+1;
539 tb->size_parm[0] = asparm0;
540 tb->size_parm[1] = asparm1;
541 tb->ptr_parm[0] = apparm0;
544 tb->ts = afs_brs_count++;
545 /* if daemons are waiting for work, wake them up */
546 if (afs_brsDaemons > 0) {
547 afs_osi_Wakeup(&afs_brsDaemons);
549 MReleaseWriteLock(&afs_xbrs);
553 MReleaseWriteLock(&afs_xbrs);
556 /* no free buffers, sleep a while */
558 MReleaseWriteLock(&afs_xbrs);
559 afs_osi_Sleep(&afs_brsWaiters);
560 MObtainWriteLock(&afs_xbrs,301);
567 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
568 * The modifications here will work for either a UP or MP machine.
570 struct buf *afs_asyncbuf = (struct buf*)0;
571 afs_int32 afs_asyncbuf_cv = EVENT_NULL;
572 afs_int32 afs_biodcnt = 0;
574 /* in implementing this, I assumed that all external linked lists were
577 * Several places in this code traverse a linked list. The algorithm
578 * used here is probably unfamiliar to most people. Careful examination
579 * will show that it eliminates an assignment inside the loop, as compared
580 * to the standard algorithm, at the cost of occasionally using an extra
586 * This function obtains, and returns, a pointer to a buffer for
587 * processing by a daemon. It sleeps until such a buffer is available.
588 * The source of buffers for it is the list afs_asyncbuf (see also
589 * naix_vm_strategy). This function may be invoked concurrently by
590 * several processes, that is, several instances of the same daemon.
591 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
592 * level, while get_bioreq runs at process level.
594 * Since AIX 4.1 can wake just one process at a time, the separate sleep
595 * addresses have been removed.
596 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
597 * The afs_asyncbuf_lock is primarily used to serialize access between
598 * process and interrupts.
600 Simple_lock afs_asyncbuf_lock;
601 /*static*/ struct buf *afs_get_bioreq()
603 struct buf *bp = NULL;
605 struct buf **bestlbpP, **lbpP;
607 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
610 struct afs_bioqueue *s;
612 /* ??? Does the forward pointer of the returned buffer need to be NULL?
615 /* Disable interrupts from the strategy function, and save the
616 * prior priority level and lock access to the afs_asyncbuf.
619 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock) ;
623 /* look for oldest buffer */
624 bp = bestbp = afs_asyncbuf;
625 bestage = (int) bestbp->av_back;
626 bestlbpP = &afs_asyncbuf;
631 if ((int) bp->av_back - bestage < 0) {
634 bestage = (int) bp->av_back;
638 *bestlbpP = bp->av_forw;
642 /* If afs_asyncbuf is null, it is necessary to go to sleep.
643 * e_wakeup_one() ensures that only one thread wakes.
646 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
647 * lock on an MP machine.
649 interrupted = e_sleep_thread(&afs_asyncbuf_cv,
651 LOCK_HANDLER|INTERRUPTIBLE);
652 if (interrupted==THREAD_INTERRUPTED) {
653 /* re-enable interrupts from strategy */
654 unlock_enable(oldPriority, &afs_asyncbuf_lock);
658 } /* end of "else asyncbuf is empty" */
659 } /* end of "inner loop" */
663 unlock_enable(oldPriority, &afs_asyncbuf_lock);
666 /* For the convenience of other code, replace the gnodes in
667 * the b_vp field of bp and the other buffers on the b_work
668 * chain with the corresponding vnodes.
670 * ??? what happens to the gnodes? They're not just cut loose,
674 t2P = (struct buf *) t1P->b_work;
675 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
679 t1P = (struct buf *) t2P->b_work;
680 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
685 /* If the buffer does not specify I/O, it may immediately
686 * be returned to the caller. This condition is detected
687 * by examining the buffer's flags (the b_flags field). If
688 * the B_PFPROT bit is set, the buffer represents a protection
689 * violation, rather than a request for I/O. The remainder
690 * of the outer loop handles the case where the B_PFPROT bit is clear.
692 if (bp->b_flags & B_PFPROT) {
697 } /* end of function get_bioreq() */
702 * This function is the daemon. It is called from the syscall
703 * interface. Ordinarily, a script or an administrator will run a
704 * daemon startup utility, specifying the number of I/O daemons to
705 * run. The utility will fork off that number of processes,
706 * each making the appropriate syscall, which will cause this
707 * function to be invoked.
709 static int afs_initbiod = 0; /* this is self-initializing code */
711 int afs_BioDaemon (afs_int32 nbiods)
713 afs_int32 code, s, pflg = 0;
715 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
723 /* pin lock, since we'll be using it in an interrupt. */
724 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
725 simple_lock_init(&afs_asyncbuf_lock);
726 pin (&afs_asyncbuf, sizeof(struct buf*));
727 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
730 /* Ignore HUP signals... */
732 sigset_t sigbits, osigbits;
734 * add SIGHUP to the set of already masked signals
736 SIGFILLSET(sigbits); /* allow all signals */
737 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
738 limit_sigs(&sigbits, &osigbits); /* and already masked */
740 /* Main body starts here -- this is an intentional infinite loop, and
743 * Now, the loop will exit if get_bioreq() returns NULL, indicating
744 * that we've been interrupted.
747 bp = afs_get_bioreq();
749 break; /* we were interrupted */
750 if (code = setjmpx(&jmpbuf)) {
751 /* This should not have happend, maybe a lack of resources */
753 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
754 for (bp1 = bp; bp ; bp = bp1) {
756 bp1 = (struct buf *) bp1->b_work;
759 bp->b_flags |= B_ERROR;
762 unlock_enable(s, &afs_asyncbuf_lock);
766 vcp = VTOAFS(bp->b_vp);
767 if (bp->b_flags & B_PFSTORE) { /* XXXX */
768 ObtainWriteLock(&vcp->lock,404);
769 if (vcp->v.v_gnode->gn_mwrcnt) {
770 afs_offs_t newlength =
771 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
772 if (vcp->m.Length < newlength) {
773 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
774 ICL_TYPE_STRING, __FILE__,
775 ICL_TYPE_LONG, __LINE__,
776 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(vcp->m.Length),
777 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
778 vcp->m.Length = newlength;
781 ReleaseWriteLock(&vcp->lock);
783 /* If the buffer represents a protection violation, rather than
784 * an actual request for I/O, no special action need be taken.
786 if ( bp->b_flags & B_PFPROT ) {
787 iodone (bp); /* Notify all users of the buffer that we're done */
792 ObtainWriteLock(&vcp->pvmlock,211);
794 * First map its data area to a region in the current address space
795 * by calling vm_att with the subspace identifier, and a pointer to
796 * the data area. vm_att returns a new data area pointer, but we
797 * also want to hang onto the old one.
799 tmpaddr = bp->b_baddr;
800 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
801 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
802 if (tmperr) { /* in non-error case */
803 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
804 bp->b_error = tmperr;
807 /* Unmap the buffer's data area by calling vm_det. Reset data area
808 * to the value that we saved above.
810 vm_det(bp->b_un.b_addr);
811 bp->b_baddr = tmpaddr;
814 * buffer may be linked with other buffers via the b_work field.
815 * See also naix_vm_strategy. For each buffer in the chain (including
816 * bp) notify all users of the buffer that the daemon is finished
817 * using it by calling iodone.
818 * assumes iodone can modify the b_work field.
821 tbp2 = (struct buf *) tbp1->b_work;
826 tbp1 = (struct buf *) tbp2->b_work;
832 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
834 } /* infinite loop (unless we're interrupted) */
835 } /* end of afs_BioDaemon() */
837 #else /* AFS_AIX41_ENV */
841 struct afs_bioqueue {
846 struct afs_bioqueue afs_bioqueue;
847 struct buf *afs_busyq = NULL;
848 struct buf *afs_asyncbuf;
849 afs_int32 afs_biodcnt = 0;
851 /* in implementing this, I assumed that all external linked lists were
854 * Several places in this code traverse a linked list. The algorithm
855 * used here is probably unfamiliar to most people. Careful examination
856 * will show that it eliminates an assignment inside the loop, as compared
857 * to the standard algorithm, at the cost of occasionally using an extra
863 * This function obtains, and returns, a pointer to a buffer for
864 * processing by a daemon. It sleeps until such a buffer is available.
865 * The source of buffers for it is the list afs_asyncbuf (see also
866 * naix_vm_strategy). This function may be invoked concurrently by
867 * several processes, that is, several instances of the same daemon.
868 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
869 * level, while get_bioreq runs at process level.
871 * The common kernel paradigm of sleeping and waking up, in which all the
872 * competing processes sleep waiting for wakeups on one address, is not
873 * followed here. Instead, the following paradigm is used: when a daemon
874 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
875 * it sleeps on the address of variable afs_asyncbuf. But if there is
876 * already a daemon sleeping on that address, it threads its own unique
877 * address onto a list, and sleeps on that address. This way, every
878 * sleeper is sleeping on a different address, and every wakeup wakes up
879 * exactly one daemon. This prevents a whole bunch of daemons from waking
880 * up and then immediately having to go back to sleep. This provides a
881 * performance gain and makes the I/O scheduling a bit more deterministic.
882 * The list of sleepers is variable afs_bioqueue. The unique address
883 * on which to sleep is passed to get_bioreq as its parameter.
885 /*static*/ struct buf *afs_get_bioreq(self)
886 struct afs_bioqueue *self; /* address on which to sleep */
889 struct buf *bp = NULL;
891 struct buf **bestlbpP, **lbpP;
893 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
896 struct afs_bioqueue *s;
898 /* ??? Does the forward pointer of the returned buffer need to be NULL?
901 /* Disable interrupts from the strategy function, and save the
902 * prior priority level
904 oldPriority = i_disable ( INTMAX ) ;
906 /* Each iteration of following loop either pulls
907 * a buffer off afs_asyncbuf, or sleeps.
909 while (1) { /* inner loop */
911 /* look for oldest buffer */
912 bp = bestbp = afs_asyncbuf;
913 bestage = (int) bestbp->av_back;
914 bestlbpP = &afs_asyncbuf;
919 if ((int) bp->av_back - bestage < 0) {
922 bestage = (int) bp->av_back;
926 *bestlbpP = bp->av_forw;
932 /* If afs_asyncbuf is null, it is necessary to go to sleep.
933 * There are two possibilities: either there is already a
934 * daemon that is sleeping on the address of afs_asyncbuf,
937 if (afs_bioqueue.sleeper) {
939 QAdd (&(afs_bioqueue.lruq), &(self->lruq));
940 interrupted = sleep ((caddr_t) self, PCATCH|(PZERO + 1));
941 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
942 QRemove (&(self->lruq)); /* dequeue */
945 afs_bioqueue.sleeper = FALSE;
947 /* re-enable interrupts from strategy */
948 i_enable (oldPriority);
953 afs_bioqueue.sleeper = TRUE;
954 interrupted = sleep ((caddr_t) &afs_asyncbuf, PCATCH|(PZERO + 1));
955 afs_bioqueue.sleeper = FALSE;
959 * We need to wakeup another daemon if present
960 * since we were waiting on afs_asyncbuf.
962 #ifdef notdef /* The following doesn't work as advertised */
963 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq)
965 struct squeue *bq = afs_bioqueue.lruq.next;
970 /* re-enable interrupts from strategy */
971 i_enable (oldPriority);
977 } /* end of "else asyncbuf is empty" */
978 } /* end of "inner loop" */
982 i_enable (oldPriority); /* re-enable interrupts from strategy */
984 /* For the convenience of other code, replace the gnodes in
985 * the b_vp field of bp and the other buffers on the b_work
986 * chain with the corresponding vnodes.
988 * ??? what happens to the gnodes? They're not just cut loose,
992 t2P = (struct buf *) t1P->b_work;
993 t1P->b_vp = ((struct gnode *) t1P->b_vp)->gn_vnode;
997 t1P = (struct buf *) t2P->b_work;
998 t2P->b_vp = ((struct gnode *) t2P->b_vp)->gn_vnode;
1003 /* If the buffer does not specify I/O, it may immediately
1004 * be returned to the caller. This condition is detected
1005 * by examining the buffer's flags (the b_flags field). If
1006 * the B_PFPROT bit is set, the buffer represents a protection
1007 * violation, rather than a request for I/O. The remainder
1008 * of the outer loop handles the case where the B_PFPROT bit is clear.
1010 if (bp->b_flags & B_PFPROT) {
1014 /* wake up another process to handle the next buffer, and return
1017 oldPriority = i_disable ( INTMAX ) ;
1019 /* determine where to find the sleeping process.
1020 * There are two cases: either it is sleeping on
1021 * afs_asyncbuf, or it is sleeping on its own unique
1022 * address. These cases are distinguished by examining
1023 * the sleeper field of afs_bioqueue.
1025 if (afs_bioqueue.sleeper) {
1026 wakeup (&afs_asyncbuf);
1029 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1030 /* queue is empty, what now? ???*/
1031 /* Should this be impossible, or does */
1032 /* it just mean that nobody is sleeping? */;
1035 struct squeue *bq = afs_bioqueue.lruq.next;
1039 afs_bioqueue.sleeper = TRUE;
1042 i_enable (oldPriority); /* re-enable interrupts from strategy */
1045 } /* end of function get_bioreq() */
1050 * This function is the daemon. It is called from the syscall
1051 * interface. Ordinarily, a script or an administrator will run a
1052 * daemon startup utility, specifying the number of I/O daemons to
1053 * run. The utility will fork off that number of processes,
1054 * each making the appropriate syscall, which will cause this
1055 * function to be invoked.
1057 static int afs_initbiod = 0; /* this is self-initializing code */
1059 afs_BioDaemon (nbiods)
1062 struct afs_bioqueue *self;
1063 afs_int32 code, s, pflg = 0;
1065 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1070 if (!afs_initbiod) {
1073 /* Initialize the queue of waiting processes, afs_bioqueue. */
1074 QInit (&(afs_bioqueue.lruq));
1077 /* establish ourself as a kernel process so shutdown won't kill us */
1078 /* u.u_procp->p_flag |= SKPROC;*/
1080 /* Initialize a token (self) to use in the queue of sleeping processes. */
1081 self = (struct afs_bioqueue *) afs_osi_Alloc (sizeof (struct afs_bioqueue));
1082 pin (self, sizeof (struct afs_bioqueue)); /* fix in memory */
1083 memset(self, 0, sizeof(*self));
1084 QInit (&(self->lruq)); /* initialize queue entry pointers */
1087 /* Ignore HUP signals... */
1088 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1089 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1090 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1091 /* Main body starts here -- this is an intentional infinite loop, and
1094 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1095 * that we've been interrupted.
1098 bp = afs_get_bioreq(self);
1100 break; /* we were interrupted */
1101 if (code = setjmpx(&jmpbuf)) {
1102 /* This should not have happend, maybe a lack of resources */
1104 for (bp1 = bp; bp ; bp = bp1) {
1109 bp->b_flags |= B_ERROR;
1115 vcp = VTOAFS(bp->b_vp);
1116 if (bp->b_flags & B_PFSTORE) {
1117 ObtainWriteLock(&vcp->lock,210);
1118 if (vcp->v.v_gnode->gn_mwrcnt) {
1119 if (vcp->m.Length < bp->b_bcount + (u_int)dbtob(bp->b_blkno))
1120 vcp->m.Length = bp->b_bcount + (u_int)dbtob(bp->b_blkno);
1122 ReleaseWriteLock(&vcp->lock);
1124 /* If the buffer represents a protection violation, rather than
1125 * an actual request for I/O, no special action need be taken.
1127 if ( bp->b_flags & B_PFPROT ) {
1128 iodone (bp); /* Notify all users of the buffer that we're done */
1132 ObtainWriteLock(&vcp->pvmlock,558);
1134 * First map its data area to a region in the current address space
1135 * by calling vm_att with the subspace identifier, and a pointer to
1136 * the data area. vm_att returns a new data area pointer, but we
1137 * also want to hang onto the old one.
1139 tmpaddr = bp->b_baddr;
1140 bp->b_baddr = vm_att (bp->b_xmemd.subspace_id, tmpaddr);
1141 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1142 if (tmperr) { /* in non-error case */
1143 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1144 bp->b_error = tmperr;
1147 /* Unmap the buffer's data area by calling vm_det. Reset data area
1148 * to the value that we saved above.
1150 vm_det(bp->b_un.b_addr);
1151 bp->b_baddr = tmpaddr;
1154 * buffer may be linked with other buffers via the b_work field.
1155 * See also naix_vm_strategy. For each buffer in the chain (including
1156 * bp) notify all users of the buffer that the daemon is finished
1157 * using it by calling iodone.
1158 * assumes iodone can modify the b_work field.
1161 tbp2 = (struct buf *) tbp1->b_work;
1166 tbp1 = (struct buf *) tbp2->b_work;
1172 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1174 } /* infinite loop (unless we're interrupted) */
1175 unpin (self, sizeof (struct afs_bioqueue));
1176 afs_osi_Free (self, sizeof (struct afs_bioqueue));
1177 } /* end of afs_BioDaemon() */
1178 #endif /* AFS_AIX41_ENV */
1179 #endif /* AFS_AIX32_ENV */
1183 void afs_BackgroundDaemon(void)
1185 struct brequest *tb;
1188 AFS_STATCNT(afs_BackgroundDaemon);
1189 /* initialize subsystem */
1191 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1192 memset((char *)afs_brs, 0, sizeof(afs_brs));
1194 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1196 * steal the first daemon for doing delayed DSlot flushing
1197 * (see afs_GetDownDSlot)
1206 MObtainWriteLock(&afs_xbrs,302);
1209 struct brequest *min_tb;
1211 if (afs_termState == AFSOP_STOP_BKG) {
1212 if (--afs_nbrs <= 0)
1213 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1214 MReleaseWriteLock(&afs_xbrs);
1215 afs_osi_Wakeup(&afs_termState);
1219 /* find a request */
1223 for(i=0; i<NBRS; i++, tb++) {
1224 /* look for request with smallest ts */
1225 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1226 /* new request, not yet picked up */
1227 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1233 if ((tb = min_tb)) {
1234 /* claim and process this request */
1235 tb->flags |= BSTARTED;
1236 MReleaseWriteLock(&afs_xbrs);
1238 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1,
1239 ICL_TYPE_INT32, tb->opcode);
1240 if (tb->opcode == BOP_FETCH)
1242 else if (tb->opcode == BOP_STORE)
1244 else if (tb->opcode == BOP_PATH)
1246 else panic("background bop");
1249 tb->vnode->vrefCount--; /* fix up reference count */
1251 AFS_RELE((struct vnode *)(tb->vnode)); /* MUST call vnode layer or could lose vnodes */
1257 tb->cred = (struct AFS_UCRED *) 0;
1259 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1260 MObtainWriteLock(&afs_xbrs,305);
1263 /* wait for new request */
1265 MReleaseWriteLock(&afs_xbrs);
1266 afs_osi_Sleep(&afs_brsDaemons);
1267 MObtainWriteLock(&afs_xbrs,307);
1274 void shutdown_daemons(void)
1276 AFS_STATCNT(shutdown_daemons);
1277 if (afs_cold_shutdown) {
1278 afs_brsDaemons = brsInit = 0;
1279 rxepoch_checked = afs_nbrs = 0;
1280 memset((char *)afs_brs, 0, sizeof(afs_brs));
1281 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1283 #ifdef AFS_AIX32_ENV
1284 #ifdef AFS_AIX41_ENV
1285 lock_free(&afs_asyncbuf_lock);
1286 unpin(&afs_asyncbuf, sizeof(struct buf*));
1287 pin (&afs_asyncbuf_cv, sizeof(afs_int32));
1288 #else /* AFS_AIX41_ENV */
1291 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1298 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1300 * sgi - daemon - handles certain operations that otherwise
1301 * would use up too much kernel stack space
1303 * This all assumes that since the caller must have the xdcache lock
1304 * exclusively that the list will never be more than one long
1305 * and noone else can attempt to add anything until we're done.
1307 SV_TYPE afs_sgibksync;
1308 SV_TYPE afs_sgibkwait;
1309 lock_t afs_sgibklock;
1310 struct dcache *afs_sgibklist;
1312 int afs_sgidaemon(void)
1317 if (afs_sgibklock == NULL) {
1318 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1319 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1320 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1322 s = SPLOCK(afs_sgibklock);
1324 /* wait for something to do */
1325 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1326 osi_Assert(afs_sgibklist);
1328 /* XX will probably need to generalize to real list someday */
1329 s = SPLOCK(afs_sgibklock);
1330 while (afs_sgibklist) {
1331 tdc = afs_sgibklist;
1332 afs_sgibklist = NULL;
1333 SPUNLOCK(afs_sgibklock, s);
1335 tdc->dflags &= ~DFEntryMod;
1336 afs_WriteDCache(tdc, 1);
1338 s = SPLOCK(afs_sgibklock);
1341 /* done all the work - wake everyone up */
1342 while (SV_SIGNAL(&afs_sgibkwait))