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"
18 #include <sys/sleep.h>
21 #include "afs/sysincludes.h" /* Standard vendor system headers */
22 #include "afsincludes.h" /* Afs-based standard headers */
23 #include "afs/afs_stats.h" /* statistics gathering code */
24 #include "afs/afs_cbqueue.h"
26 #include <sys/adspace.h> /* for vm_att(), vm_det() */
30 /* background request queue size */
31 afs_lock_t afs_xbrs; /* lock for brs */
32 static int brsInit = 0;
33 short afs_brsWaiters = 0; /* number of users waiting for brs buffers */
34 short afs_brsDaemons = 0; /* number of daemons waiting for brs requests */
35 struct brequest afs_brs[NBRS]; /* request structures */
36 struct afs_osi_WaitHandle AFS_WaitHandler, AFS_CSWaitHandler;
37 static int afs_brs_count = 0; /* request counter, to service reqs in order */
39 static int rxepoch_checked = 0;
40 #define afs_CheckRXEpoch() {if (rxepoch_checked == 0 && rxkad_EpochWasSet) { \
41 rxepoch_checked = 1; afs_GCUserData(/* force flag */ 1); } }
43 /* PAG garbage collection */
44 /* We induce a compile error if param.h does not define AFS_GCPAGS */
45 afs_int32 afs_gcpags = AFS_GCPAGS;
46 afs_int32 afs_gcpags_procsize = 0;
48 afs_int32 afs_CheckServerDaemonStarted = 0;
49 #ifdef DEFAULT_PROBE_INTERVAL
50 afs_int32 PROBE_INTERVAL = DEFAULT_PROBE_INTERVAL; /* overridding during compile */
52 afs_int32 PROBE_INTERVAL = 180; /* default to 3 min */
55 #define PROBE_WAIT() (1000 * (PROBE_INTERVAL - ((afs_random() & 0x7fffffff) \
56 % (PROBE_INTERVAL/2))))
59 afs_CheckServerDaemon(void)
61 afs_int32 now, delay, lastCheck, last10MinCheck;
63 afs_CheckServerDaemonStarted = 1;
65 while (afs_initState < 101)
66 afs_osi_Sleep(&afs_initState);
67 afs_osi_Wait(PROBE_WAIT(), &AFS_CSWaitHandler, 0);
69 last10MinCheck = lastCheck = osi_Time();
71 if (afs_termState == AFSOP_STOP_CS) {
72 afs_termState = AFSOP_STOP_BKG;
73 afs_osi_Wakeup(&afs_termState);
78 if (PROBE_INTERVAL + lastCheck <= now) {
79 afs_CheckServers(1, NULL); /* check down servers */
80 lastCheck = now = osi_Time();
83 if (600 + last10MinCheck <= now) {
84 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, 600);
85 afs_CheckServers(0, NULL);
86 last10MinCheck = now = osi_Time();
89 if (afs_termState == AFSOP_STOP_CS) {
90 afs_termState = AFSOP_STOP_BKG;
91 afs_osi_Wakeup(&afs_termState);
95 /* Compute time to next probe. */
96 delay = PROBE_INTERVAL + lastCheck;
97 if (delay > 600 + last10MinCheck)
98 delay = 600 + last10MinCheck;
102 afs_osi_Wait(delay * 1000, &AFS_CSWaitHandler, 0);
104 afs_CheckServerDaemonStarted = 0;
111 struct afs_exporter *exporter;
113 afs_int32 last3MinCheck, last10MinCheck, last60MinCheck, lastNMinCheck;
114 afs_int32 last1MinCheck;
115 afs_uint32 lastCBSlotBump;
118 AFS_STATCNT(afs_Daemon);
119 last1MinCheck = last3MinCheck = last60MinCheck = last10MinCheck =
122 afs_rootFid.Fid.Volume = 0;
123 while (afs_initState < 101)
124 afs_osi_Sleep(&afs_initState);
127 lastCBSlotBump = now;
129 /* when a lot of clients are booted simultaneously, they develop
130 * annoying synchronous VL server bashing behaviors. So we stagger them.
132 last1MinCheck = now + ((afs_random() & 0x7fffffff) % 60); /* an extra 30 */
133 last3MinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
134 last60MinCheck = now - 1800 + ((afs_random() & 0x7fffffff) % 3600);
135 last10MinCheck = now - 300 + ((afs_random() & 0x7fffffff) % 600);
136 lastNMinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
138 /* start off with afs_initState >= 101 (basic init done) */
140 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
142 /* things to do every 20 seconds or less - required by protocol spec */
144 afs_FlushActiveVcaches(0); /* flush NFS writes */
145 afs_FlushVCBs(1); /* flush queued callbacks */
146 afs_MaybeWakeupTruncateDaemon(); /* free cache space if have too */
147 rx_CheckPackets(); /* Does RX need more packets? */
148 #if defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)
150 * Hack: We always want to make sure there are plenty free
151 * entries in the small free pool so that we don't have to
152 * worry about rx (with disabled interrupts) to have to call
153 * malloc). So we do the dummy call below...
155 if (((afs_stats_cmperf.SmallBlocksAlloced -
156 afs_stats_cmperf.SmallBlocksActive)
157 <= AFS_SALLOC_LOW_WATER))
158 osi_FreeSmallSpace(osi_AllocSmallSpace(AFS_SMALLOCSIZ));
159 if (((afs_stats_cmperf.MediumBlocksAlloced -
160 afs_stats_cmperf.MediumBlocksActive)
161 <= AFS_MALLOC_LOW_WATER + 50))
162 osi_AllocMoreMSpace(AFS_MALLOC_LOW_WATER * 2);
166 if (lastCBSlotBump + CBHTSLOTLEN < now) { /* pretty time-dependant */
167 lastCBSlotBump = now;
168 if (afs_BumpBase()) {
169 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
173 if (last1MinCheck + 60 < now) {
174 /* things to do every minute */
175 DFlush(); /* write out dir buffers */
176 afs_WriteThroughDSlots(); /* write through cacheinfo entries */
177 afs_FlushActiveVcaches(1); /* keep flocks held & flush nfs writes */
178 #ifdef AFS_DISCON_ENV
179 afs_StoreDirtyVcaches();
185 if (last3MinCheck + 180 < now) {
186 afs_CheckTokenCache(); /* check for access cache resets due to expired
190 if (!afs_CheckServerDaemonStarted) {
191 /* Do the check here if the correct afsd is not installed. */
194 printf("Please install afsd with check server daemon.\n");
196 if (lastNMinCheck + PROBE_INTERVAL < now) {
197 /* only check down servers */
198 afs_CheckServers(1, NULL);
202 if (last10MinCheck + 600 < now) {
203 #ifdef AFS_USERSPACE_IP_ADDR
204 extern int rxi_GetcbiInfo(void);
206 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, 600);
207 #ifdef AFS_USERSPACE_IP_ADDR
208 if (rxi_GetcbiInfo()) { /* addresses changed from last time */
211 #else /* AFS_USERSPACE_IP_ADDR */
212 if (rxi_GetIFInfo()) { /* addresses changed from last time */
215 #endif /* else AFS_USERSPACE_IP_ADDR */
216 if (!afs_CheckServerDaemonStarted)
217 afs_CheckServers(0, NULL);
218 afs_GCUserData(0); /* gc old conns */
219 /* This is probably the wrong way of doing GC for the various exporters but it will suffice for a while */
220 for (exporter = root_exported; exporter;
221 exporter = exporter->exp_next) {
222 (void)EXP_GC(exporter, 0); /* Generalize params */
227 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
231 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
236 last10MinCheck = now;
238 if (last60MinCheck + 3600 < now) {
239 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEVOLUME, ICL_TYPE_INT32,
241 afs_CheckRootVolume();
243 if (afs_gcpags == AFS_GCPAGS_OK) {
248 last60MinCheck = now;
250 if (afs_initState < 300) { /* while things ain't rosy */
251 code = afs_CheckRootVolume();
253 afs_initState = 300; /* succeeded */
254 if (afs_initState < 200)
255 afs_initState = 200; /* tried once */
256 afs_osi_Wakeup(&afs_initState);
259 /* 18285 is because we're trying to divide evenly into 128, that is,
260 * CBSlotLen, while staying just under 20 seconds. If CBSlotLen
261 * changes, should probably change this interval, too.
262 * Some of the preceding actions may take quite some time, so we
263 * might not want to wait the entire interval */
264 now = 18285 - (osi_Time() - now);
266 afs_osi_Wait(now, &AFS_WaitHandler, 0);
269 if (afs_termState == AFSOP_STOP_AFS) {
270 if (afs_CheckServerDaemonStarted)
271 afs_termState = AFSOP_STOP_CS;
273 afs_termState = AFSOP_STOP_BKG;
274 afs_osi_Wakeup(&afs_termState);
281 afs_CheckRootVolume(void)
283 char rootVolName[32];
284 struct volume *tvp = NULL;
285 int usingDynroot = afs_GetDynrootEnable();
288 AFS_STATCNT(afs_CheckRootVolume);
289 if (*afs_rootVolumeName == 0) {
290 strcpy(rootVolName, "root.afs");
292 strcpy(rootVolName, afs_rootVolumeName);
296 afs_GetDynrootFid(&afs_rootFid);
297 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
299 struct cell *lc = afs_GetPrimaryCell(READ_LOCK);
303 localcell = lc->cellNum;
304 afs_PutCell(lc, READ_LOCK);
305 tvp = afs_GetVolumeByName(rootVolName, localcell, 1, NULL, READ_LOCK);
308 int len = strlen(rootVolName);
310 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
311 strcpy(buf, rootVolName);
312 afs_strcat(buf, ".readonly");
313 tvp = afs_GetVolumeByName(buf, localcell, 1, NULL, READ_LOCK);
317 int volid = (tvp->roVol ? tvp->roVol : tvp->volume);
318 afs_rootFid.Cell = localcell;
319 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
321 /* If we had a root fid before and it changed location we reset
322 * the afs_globalVp so that it will be reevaluated.
323 * Just decrement the reference count. This only occurs during
324 * initial cell setup and can panic the machine if we set the
325 * count to zero and fs checkv is executed when the current
328 AFS_FAST_RELE(afs_globalVp);
331 afs_rootFid.Fid.Volume = volid;
332 afs_rootFid.Fid.Vnode = 1;
333 afs_rootFid.Fid.Unique = 1;
337 afs_initState = 300; /* won */
338 afs_osi_Wakeup(&afs_initState);
339 afs_PutVolume(tvp, READ_LOCK);
341 if (afs_rootFid.Fid.Volume)
347 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
349 BPath(register struct brequest *ab)
351 register struct dcache *tdc = NULL;
352 struct vcache *tvc = NULL;
353 struct vnode *tvn = NULL;
354 #ifdef AFS_LINUX22_ENV
355 struct dentry *dp = NULL;
357 afs_size_t offset, len;
358 struct vrequest treq;
362 if ((code = afs_InitReq(&treq, ab->cred)))
365 #ifdef AFS_LINUX22_ENV
366 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &dp);
368 tvn = (struct vnode *)dp->d_inode;
370 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &tvn);
373 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(tvn)) {
378 /* release it and give up */
380 #ifdef AFS_LINUX22_ENV
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);
394 #ifdef AFS_LINUX22_ENV
401 /* size_parm 0 to the fetch is the chunk number,
402 * ptr_parm 0 is the dcache entry to wakeup,
403 * size_parm 1 is true iff we should release the dcache entry here.
406 BPrefetch(register struct brequest *ab)
408 register struct dcache *tdc;
409 register struct vcache *tvc;
410 afs_size_t offset, len;
411 struct vrequest treq;
413 AFS_STATCNT(BPrefetch);
414 if ((len = afs_InitReq(&treq, ab->cred)))
417 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
421 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
422 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
423 * be waiting for our wakeup anyway.
425 tdc = (struct dcache *)(ab->ptr_parm[0]);
426 ObtainSharedLock(&tdc->lock, 640);
427 if (tdc->mflags & DFFetchReq) {
428 UpgradeSToWLock(&tdc->lock, 641);
429 tdc->mflags &= ~DFFetchReq;
430 ReleaseWriteLock(&tdc->lock);
432 ReleaseSharedLock(&tdc->lock);
434 afs_osi_Wakeup(&tdc->validPos);
435 if (ab->size_parm[1]) {
436 afs_PutDCache(tdc); /* put this one back, too */
442 BStore(register struct brequest *ab)
444 register struct vcache *tvc;
445 register afs_int32 code;
446 struct vrequest treq;
447 #if defined(AFS_SGI_ENV)
448 struct cred *tmpcred;
452 if ((code = afs_InitReq(&treq, ab->cred)))
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 */
492 afs_BRelease(register struct brequest *ab)
495 AFS_STATCNT(afs_BRelease);
496 MObtainWriteLock(&afs_xbrs, 294);
497 if (--ab->refCount <= 0) {
501 afs_osi_Wakeup(&afs_brsWaiters);
502 MReleaseWriteLock(&afs_xbrs);
505 /* return true if bkg fetch daemons are all busy */
509 AFS_STATCNT(afs_BBusy);
510 if (afs_brsDaemons > 0)
516 afs_BQueue(register short aopcode, register struct vcache *avc,
517 afs_int32 dontwait, afs_int32 ause, struct AFS_UCRED *acred,
518 afs_size_t asparm0, afs_size_t asparm1, void *apparm0)
521 register struct brequest *tb;
523 AFS_STATCNT(afs_BQueue);
524 MObtainWriteLock(&afs_xbrs, 296);
527 for (i = 0; i < NBRS; i++, tb++) {
528 if (tb->refCount == 0)
533 tb->opcode = aopcode;
538 VN_HOLD(AFSTOV(avc));
540 tb->refCount = ause + 1;
541 tb->size_parm[0] = asparm0;
542 tb->size_parm[1] = asparm1;
543 tb->ptr_parm[0] = apparm0;
546 tb->ts = afs_brs_count++;
547 /* if daemons are waiting for work, wake them up */
548 if (afs_brsDaemons > 0) {
549 afs_osi_Wakeup(&afs_brsDaemons);
551 MReleaseWriteLock(&afs_xbrs);
555 MReleaseWriteLock(&afs_xbrs);
558 /* no free buffers, sleep a while */
560 MReleaseWriteLock(&afs_xbrs);
561 afs_osi_Sleep(&afs_brsWaiters);
562 MObtainWriteLock(&afs_xbrs, 301);
569 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
570 * The modifications here will work for either a UP or MP machine.
572 struct buf *afs_asyncbuf = (struct buf *)0;
573 tid_t afs_asyncbuf_cv = EVENT_NULL;
574 afs_int32 afs_biodcnt = 0;
576 /* in implementing this, I assumed that all external linked lists were
579 * Several places in this code traverse a linked list. The algorithm
580 * used here is probably unfamiliar to most people. Careful examination
581 * will show that it eliminates an assignment inside the loop, as compared
582 * to the standard algorithm, at the cost of occasionally using an extra
588 * This function obtains, and returns, a pointer to a buffer for
589 * processing by a daemon. It sleeps until such a buffer is available.
590 * The source of buffers for it is the list afs_asyncbuf (see also
591 * naix_vm_strategy). This function may be invoked concurrently by
592 * several processes, that is, several instances of the same daemon.
593 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
594 * level, while get_bioreq runs at process level.
596 * Since AIX 4.1 can wake just one process at a time, the separate sleep
597 * addresses have been removed.
598 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
599 * The afs_asyncbuf_lock is primarily used to serialize access between
600 * process and interrupts.
602 Simple_lock afs_asyncbuf_lock;
603 /*static*/ struct buf *
606 struct buf *bp = NULL;
608 struct buf **bestlbpP, **lbpP;
610 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
613 struct afs_bioqueue *s;
615 /* ??? Does the forward pointer of the returned buffer need to be NULL?
618 /* Disable interrupts from the strategy function, and save the
619 * prior priority level and lock access to the afs_asyncbuf.
622 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock);
626 /* look for oldest buffer */
627 bp = bestbp = afs_asyncbuf;
628 bestage = (long)bestbp->av_back;
629 bestlbpP = &afs_asyncbuf;
635 if ((long)bp->av_back - bestage < 0) {
638 bestage = (long)bp->av_back;
642 *bestlbpP = bp->av_forw;
645 /* If afs_asyncbuf is null, it is necessary to go to sleep.
646 * e_wakeup_one() ensures that only one thread wakes.
649 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
650 * lock on an MP machine.
653 e_sleep_thread(&afs_asyncbuf_cv, &afs_asyncbuf_lock,
654 LOCK_HANDLER | INTERRUPTIBLE);
655 if (interrupted == THREAD_INTERRUPTED) {
656 /* re-enable interrupts from strategy */
657 unlock_enable(oldPriority, &afs_asyncbuf_lock);
661 } /* end of "else asyncbuf is empty" */
662 } /* end of "inner loop" */
666 unlock_enable(oldPriority, &afs_asyncbuf_lock);
669 /* For the convenience of other code, replace the gnodes in
670 * the b_vp field of bp and the other buffers on the b_work
671 * chain with the corresponding vnodes.
673 * ??? what happens to the gnodes? They're not just cut loose,
677 t2P = (struct buf *)t1P->b_work;
678 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
682 t1P = (struct buf *)t2P->b_work;
683 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
688 /* If the buffer does not specify I/O, it may immediately
689 * be returned to the caller. This condition is detected
690 * by examining the buffer's flags (the b_flags field). If
691 * the B_PFPROT bit is set, the buffer represents a protection
692 * violation, rather than a request for I/O. The remainder
693 * of the outer loop handles the case where the B_PFPROT bit is clear.
695 if (bp->b_flags & B_PFPROT) {
700 } /* end of function get_bioreq() */
705 * This function is the daemon. It is called from the syscall
706 * interface. Ordinarily, a script or an administrator will run a
707 * daemon startup utility, specifying the number of I/O daemons to
708 * run. The utility will fork off that number of processes,
709 * each making the appropriate syscall, which will cause this
710 * function to be invoked.
712 static int afs_initbiod = 0; /* this is self-initializing code */
715 afs_BioDaemon(afs_int32 nbiods)
717 afs_int32 code, s, pflg = 0;
719 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
727 /* pin lock, since we'll be using it in an interrupt. */
728 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
729 simple_lock_init(&afs_asyncbuf_lock);
730 pin(&afs_asyncbuf, sizeof(struct buf *));
731 pin(&afs_asyncbuf_cv, sizeof(afs_int32));
734 /* Ignore HUP signals... */
736 sigset_t sigbits, osigbits;
738 * add SIGHUP to the set of already masked signals
740 SIGFILLSET(sigbits); /* allow all signals */
741 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
742 limit_sigs(&sigbits, &osigbits); /* and already masked */
744 /* Main body starts here -- this is an intentional infinite loop, and
747 * Now, the loop will exit if get_bioreq() returns NULL, indicating
748 * that we've been interrupted.
751 bp = afs_get_bioreq();
753 break; /* we were interrupted */
754 if (code = setjmpx(&jmpbuf)) {
755 /* This should not have happend, maybe a lack of resources */
757 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
758 for (bp1 = bp; bp; bp = bp1) {
760 bp1 = (struct buf *)bp1->b_work;
763 bp->b_flags |= B_ERROR;
766 unlock_enable(s, &afs_asyncbuf_lock);
770 vcp = VTOAFS(bp->b_vp);
771 if (bp->b_flags & B_PFSTORE) { /* XXXX */
772 ObtainWriteLock(&vcp->lock, 404);
773 if (vcp->v.v_gnode->gn_mwrcnt) {
774 afs_offs_t newlength =
775 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
776 if (vcp->m.Length < newlength) {
777 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
778 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
779 __LINE__, ICL_TYPE_OFFSET,
780 ICL_HANDLE_OFFSET(vcp->m.Length),
781 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
782 vcp->m.Length = newlength;
785 ReleaseWriteLock(&vcp->lock);
787 /* If the buffer represents a protection violation, rather than
788 * an actual request for I/O, no special action need be taken.
790 if (bp->b_flags & B_PFPROT) {
791 iodone(bp); /* Notify all users of the buffer that we're done */
796 ObtainWriteLock(&vcp->pvmlock, 211);
798 * First map its data area to a region in the current address space
799 * by calling vm_att with the subspace identifier, and a pointer to
800 * the data area. vm_att returns a new data area pointer, but we
801 * also want to hang onto the old one.
803 tmpaddr = bp->b_baddr;
804 bp->b_baddr = (caddr_t) vm_att(bp->b_xmemd.subspace_id, tmpaddr);
805 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
806 if (tmperr) { /* in non-error case */
807 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
808 bp->b_error = tmperr;
811 /* Unmap the buffer's data area by calling vm_det. Reset data area
812 * to the value that we saved above.
815 bp->b_baddr = tmpaddr;
818 * buffer may be linked with other buffers via the b_work field.
819 * See also naix_vm_strategy. For each buffer in the chain (including
820 * bp) notify all users of the buffer that the daemon is finished
821 * using it by calling iodone.
822 * assumes iodone can modify the b_work field.
825 tbp2 = (struct buf *)tbp1->b_work;
830 tbp1 = (struct buf *)tbp2->b_work;
836 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
838 } /* infinite loop (unless we're interrupted) */
839 } /* end of afs_BioDaemon() */
841 #else /* AFS_AIX41_ENV */
845 struct afs_bioqueue {
850 struct afs_bioqueue afs_bioqueue;
851 struct buf *afs_busyq = NULL;
852 struct buf *afs_asyncbuf;
853 afs_int32 afs_biodcnt = 0;
855 /* in implementing this, I assumed that all external linked lists were
858 * Several places in this code traverse a linked list. The algorithm
859 * used here is probably unfamiliar to most people. Careful examination
860 * will show that it eliminates an assignment inside the loop, as compared
861 * to the standard algorithm, at the cost of occasionally using an extra
867 * This function obtains, and returns, a pointer to a buffer for
868 * processing by a daemon. It sleeps until such a buffer is available.
869 * The source of buffers for it is the list afs_asyncbuf (see also
870 * naix_vm_strategy). This function may be invoked concurrently by
871 * several processes, that is, several instances of the same daemon.
872 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
873 * level, while get_bioreq runs at process level.
875 * The common kernel paradigm of sleeping and waking up, in which all the
876 * competing processes sleep waiting for wakeups on one address, is not
877 * followed here. Instead, the following paradigm is used: when a daemon
878 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
879 * it sleeps on the address of variable afs_asyncbuf. But if there is
880 * already a daemon sleeping on that address, it threads its own unique
881 * address onto a list, and sleeps on that address. This way, every
882 * sleeper is sleeping on a different address, and every wakeup wakes up
883 * exactly one daemon. This prevents a whole bunch of daemons from waking
884 * up and then immediately having to go back to sleep. This provides a
885 * performance gain and makes the I/O scheduling a bit more deterministic.
886 * The list of sleepers is variable afs_bioqueue. The unique address
887 * on which to sleep is passed to get_bioreq as its parameter.
889 /*static*/ struct buf *
891 struct afs_bioqueue *self; /* address on which to sleep */
894 struct buf *bp = NULL;
896 struct buf **bestlbpP, **lbpP;
898 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
901 struct afs_bioqueue *s;
903 /* ??? Does the forward pointer of the returned buffer need to be NULL?
906 /* Disable interrupts from the strategy function, and save the
907 * prior priority level
909 oldPriority = i_disable(INTMAX);
911 /* Each iteration of following loop either pulls
912 * a buffer off afs_asyncbuf, or sleeps.
914 while (1) { /* inner loop */
916 /* look for oldest buffer */
917 bp = bestbp = afs_asyncbuf;
918 bestage = (int)bestbp->av_back;
919 bestlbpP = &afs_asyncbuf;
925 if ((int)bp->av_back - bestage < 0) {
928 bestage = (int)bp->av_back;
932 *bestlbpP = bp->av_forw;
937 /* If afs_asyncbuf is null, it is necessary to go to sleep.
938 * There are two possibilities: either there is already a
939 * daemon that is sleeping on the address of afs_asyncbuf,
942 if (afs_bioqueue.sleeper) {
944 QAdd(&(afs_bioqueue.lruq), &(self->lruq));
945 interrupted = sleep((caddr_t) self, PCATCH | (PZERO + 1));
946 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
947 QRemove(&(self->lruq)); /* dequeue */
950 afs_bioqueue.sleeper = FALSE;
952 /* re-enable interrupts from strategy */
953 i_enable(oldPriority);
958 afs_bioqueue.sleeper = TRUE;
960 sleep((caddr_t) & afs_asyncbuf, PCATCH | (PZERO + 1));
961 afs_bioqueue.sleeper = FALSE;
964 * We need to wakeup another daemon if present
965 * since we were waiting on afs_asyncbuf.
967 #ifdef notdef /* The following doesn't work as advertised */
968 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq) {
969 struct squeue *bq = afs_bioqueue.lruq.next;
974 /* re-enable interrupts from strategy */
975 i_enable(oldPriority);
981 } /* end of "else asyncbuf is empty" */
982 } /* end of "inner loop" */
986 i_enable(oldPriority); /* re-enable interrupts from strategy */
988 /* For the convenience of other code, replace the gnodes in
989 * the b_vp field of bp and the other buffers on the b_work
990 * chain with the corresponding vnodes.
992 * ??? what happens to the gnodes? They're not just cut loose,
996 t2P = (struct buf *)t1P->b_work;
997 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
1001 t1P = (struct buf *)t2P->b_work;
1002 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
1007 /* If the buffer does not specify I/O, it may immediately
1008 * be returned to the caller. This condition is detected
1009 * by examining the buffer's flags (the b_flags field). If
1010 * the B_PFPROT bit is set, the buffer represents a protection
1011 * violation, rather than a request for I/O. The remainder
1012 * of the outer loop handles the case where the B_PFPROT bit is clear.
1014 if (bp->b_flags & B_PFPROT) {
1018 /* wake up another process to handle the next buffer, and return
1021 oldPriority = i_disable(INTMAX);
1023 /* determine where to find the sleeping process.
1024 * There are two cases: either it is sleeping on
1025 * afs_asyncbuf, or it is sleeping on its own unique
1026 * address. These cases are distinguished by examining
1027 * the sleeper field of afs_bioqueue.
1029 if (afs_bioqueue.sleeper) {
1030 wakeup(&afs_asyncbuf);
1032 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1033 /* queue is empty, what now? ??? */
1034 /* Should this be impossible, or does */
1035 /* it just mean that nobody is sleeping? */ ;
1037 struct squeue *bq = afs_bioqueue.lruq.next;
1041 afs_bioqueue.sleeper = TRUE;
1044 i_enable(oldPriority); /* re-enable interrupts from strategy */
1047 } /* end of function get_bioreq() */
1052 * This function is the daemon. It is called from the syscall
1053 * interface. Ordinarily, a script or an administrator will run a
1054 * daemon startup utility, specifying the number of I/O daemons to
1055 * run. The utility will fork off that number of processes,
1056 * each making the appropriate syscall, which will cause this
1057 * function to be invoked.
1059 static int afs_initbiod = 0; /* this is self-initializing code */
1061 afs_BioDaemon(nbiods)
1064 struct afs_bioqueue *self;
1065 afs_int32 code, s, pflg = 0;
1067 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1072 if (!afs_initbiod) {
1075 /* Initialize the queue of waiting processes, afs_bioqueue. */
1076 QInit(&(afs_bioqueue.lruq));
1079 /* establish ourself as a kernel process so shutdown won't kill us */
1080 /* u.u_procp->p_flag |= SKPROC;*/
1082 /* Initialize a token (self) to use in the queue of sleeping processes. */
1083 self = (struct afs_bioqueue *)afs_osi_Alloc(sizeof(struct afs_bioqueue));
1084 pin(self, sizeof(struct afs_bioqueue)); /* fix in memory */
1085 memset(self, 0, sizeof(*self));
1086 QInit(&(self->lruq)); /* initialize queue entry pointers */
1089 /* Ignore HUP signals... */
1090 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1091 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1092 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1093 /* Main body starts here -- this is an intentional infinite loop, and
1096 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1097 * that we've been interrupted.
1100 bp = afs_get_bioreq(self);
1102 break; /* we were interrupted */
1103 if (code = setjmpx(&jmpbuf)) {
1104 /* This should not have happend, maybe a lack of resources */
1106 for (bp1 = bp; bp; bp = bp1) {
1111 bp->b_flags |= B_ERROR;
1117 vcp = VTOAFS(bp->b_vp);
1118 if (bp->b_flags & B_PFSTORE) {
1119 ObtainWriteLock(&vcp->lock, 210);
1120 if (vcp->v.v_gnode->gn_mwrcnt) {
1121 afs_offs_t newlength =
1122 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
1123 if (vcp->m.Length < newlength) {
1124 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
1125 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
1126 __LINE__, ICL_TYPE_OFFSET,
1127 ICL_HANDLE_OFFSET(vcp->m.Length),
1128 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
1129 vcp->m.Length = newlength;
1132 ReleaseWriteLock(&vcp->lock);
1134 /* If the buffer represents a protection violation, rather than
1135 * an actual request for I/O, no special action need be taken.
1137 if (bp->b_flags & B_PFPROT) {
1138 iodone(bp); /* Notify all users of the buffer that we're done */
1142 ObtainWriteLock(&vcp->pvmlock, 558);
1144 * First map its data area to a region in the current address space
1145 * by calling vm_att with the subspace identifier, and a pointer to
1146 * the data area. vm_att returns a new data area pointer, but we
1147 * also want to hang onto the old one.
1149 tmpaddr = bp->b_baddr;
1150 bp->b_baddr = vm_att(bp->b_xmemd.subspace_id, tmpaddr);
1151 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1152 if (tmperr) { /* in non-error case */
1153 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1154 bp->b_error = tmperr;
1157 /* Unmap the buffer's data area by calling vm_det. Reset data area
1158 * to the value that we saved above.
1160 vm_det(bp->b_un.b_addr);
1161 bp->b_baddr = tmpaddr;
1164 * buffer may be linked with other buffers via the b_work field.
1165 * See also naix_vm_strategy. For each buffer in the chain (including
1166 * bp) notify all users of the buffer that the daemon is finished
1167 * using it by calling iodone.
1168 * assumes iodone can modify the b_work field.
1171 tbp2 = (struct buf *)tbp1->b_work;
1176 tbp1 = (struct buf *)tbp2->b_work;
1182 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1184 } /* infinite loop (unless we're interrupted) */
1185 unpin(self, sizeof(struct afs_bioqueue));
1186 afs_osi_Free(self, sizeof(struct afs_bioqueue));
1187 } /* end of afs_BioDaemon() */
1188 #endif /* AFS_AIX41_ENV */
1189 #endif /* AFS_AIX32_ENV */
1194 afs_BackgroundDaemon(void)
1196 struct brequest *tb;
1199 AFS_STATCNT(afs_BackgroundDaemon);
1200 /* initialize subsystem */
1202 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1203 memset((char *)afs_brs, 0, sizeof(afs_brs));
1205 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1207 * steal the first daemon for doing delayed DSlot flushing
1208 * (see afs_GetDownDSlot)
1217 MObtainWriteLock(&afs_xbrs, 302);
1220 struct brequest *min_tb = NULL;
1222 if (afs_termState == AFSOP_STOP_BKG) {
1223 if (--afs_nbrs <= 0)
1224 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1225 MReleaseWriteLock(&afs_xbrs);
1226 afs_osi_Wakeup(&afs_termState);
1230 /* find a request */
1233 for (i = 0; i < NBRS; i++, tb++) {
1234 /* look for request with smallest ts */
1235 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1236 /* new request, not yet picked up */
1237 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1243 if ((tb = min_tb)) {
1244 /* claim and process this request */
1245 tb->flags |= BSTARTED;
1246 MReleaseWriteLock(&afs_xbrs);
1248 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1, ICL_TYPE_INT32,
1250 if (tb->opcode == BOP_FETCH)
1252 else if (tb->opcode == BOP_STORE)
1254 else if (tb->opcode == BOP_PATH)
1257 panic("background bop");
1259 AFS_RELE(AFSTOV(tb->vc)); /* MUST call vnode layer or could lose vnodes */
1264 tb->cred = (struct AFS_UCRED *)0;
1266 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1267 MObtainWriteLock(&afs_xbrs, 305);
1270 /* wait for new request */
1272 MReleaseWriteLock(&afs_xbrs);
1273 afs_osi_Sleep(&afs_brsDaemons);
1274 MObtainWriteLock(&afs_xbrs, 307);
1282 shutdown_daemons(void)
1284 AFS_STATCNT(shutdown_daemons);
1285 if (afs_cold_shutdown) {
1286 afs_brsDaemons = brsInit = 0;
1287 rxepoch_checked = afs_nbrs = 0;
1288 memset((char *)afs_brs, 0, sizeof(afs_brs));
1289 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1291 #ifdef AFS_AIX32_ENV
1292 #ifdef AFS_AIX41_ENV
1293 lock_free(&afs_asyncbuf_lock);
1294 unpin(&afs_asyncbuf, sizeof(struct buf *));
1295 unpin(&afs_asyncbuf_cv, sizeof(afs_int32));
1296 #else /* AFS_AIX41_ENV */
1299 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1306 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1308 * sgi - daemon - handles certain operations that otherwise
1309 * would use up too much kernel stack space
1311 * This all assumes that since the caller must have the xdcache lock
1312 * exclusively that the list will never be more than one long
1313 * and noone else can attempt to add anything until we're done.
1315 SV_TYPE afs_sgibksync;
1316 SV_TYPE afs_sgibkwait;
1317 lock_t afs_sgibklock;
1318 struct dcache *afs_sgibklist;
1326 if (afs_sgibklock == NULL) {
1327 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1328 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1329 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1331 s = SPLOCK(afs_sgibklock);
1333 /* wait for something to do */
1334 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1335 osi_Assert(afs_sgibklist);
1337 /* XX will probably need to generalize to real list someday */
1338 s = SPLOCK(afs_sgibklock);
1339 while (afs_sgibklist) {
1340 tdc = afs_sgibklist;
1341 afs_sgibklist = NULL;
1342 SPUNLOCK(afs_sgibklock, s);
1344 tdc->dflags &= ~DFEntryMod;
1345 afs_WriteDCache(tdc, 1);
1347 s = SPLOCK(afs_sgibklock);
1350 /* done all the work - wake everyone up */
1351 while (SV_SIGNAL(&afs_sgibkwait));