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? */
150 if (lastCBSlotBump + CBHTSLOTLEN < now) { /* pretty time-dependant */
151 lastCBSlotBump = now;
152 if (afs_BumpBase()) {
153 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
157 if (last1MinCheck + 60 < now) {
158 /* things to do every minute */
159 DFlush(); /* write out dir buffers */
160 afs_WriteThroughDSlots(); /* write through cacheinfo entries */
161 afs_FlushActiveVcaches(1); /* keep flocks held & flush nfs writes */
162 #ifdef AFS_DISCON_ENV
163 afs_StoreDirtyVcaches();
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, ICL_TYPE_INT32, 600);
191 #ifdef AFS_USERSPACE_IP_ADDR
192 if (rxi_GetcbiInfo()) { /* addresses changed from last time */
195 #else /* AFS_USERSPACE_IP_ADDR */
196 if (rxi_GetIFInfo()) { /* addresses changed from last time */
199 #endif /* else AFS_USERSPACE_IP_ADDR */
200 if (!afs_CheckServerDaemonStarted)
201 afs_CheckServers(0, NULL);
202 afs_GCUserData(0); /* gc old conns */
203 /* This is probably the wrong way of doing GC for the various exporters but it will suffice for a while */
204 for (exporter = root_exported; exporter;
205 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, ICL_TYPE_INT32,
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();
237 afs_initState = 300; /* succeeded */
238 if (afs_initState < 200)
239 afs_initState = 200; /* tried once */
240 afs_osi_Wakeup(&afs_initState);
243 /* 18285 is because we're trying to divide evenly into 128, that is,
244 * CBSlotLen, while staying just under 20 seconds. If CBSlotLen
245 * changes, should probably change this interval, too.
246 * Some of the preceding actions may take quite some time, so we
247 * might not want to wait the entire interval */
248 now = 18285 - (osi_Time() - now);
250 afs_osi_Wait(now, &AFS_WaitHandler, 0);
253 if (afs_termState == AFSOP_STOP_AFS) {
254 if (afs_CheckServerDaemonStarted)
255 afs_termState = AFSOP_STOP_CS;
257 afs_termState = AFSOP_STOP_BKG;
258 afs_osi_Wakeup(&afs_termState);
265 afs_CheckRootVolume(void)
267 char rootVolName[32];
268 struct volume *tvp = NULL;
269 int usingDynroot = afs_GetDynrootEnable();
272 AFS_STATCNT(afs_CheckRootVolume);
273 if (*afs_rootVolumeName == 0) {
274 strcpy(rootVolName, "root.afs");
276 strcpy(rootVolName, afs_rootVolumeName);
280 afs_GetDynrootFid(&afs_rootFid);
281 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
283 struct cell *lc = afs_GetPrimaryCell(READ_LOCK);
287 localcell = lc->cellNum;
288 afs_PutCell(lc, READ_LOCK);
289 tvp = afs_GetVolumeByName(rootVolName, localcell, 1, NULL, READ_LOCK);
292 int len = strlen(rootVolName);
294 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
295 strcpy(buf, rootVolName);
296 afs_strcat(buf, ".readonly");
297 tvp = afs_GetVolumeByName(buf, localcell, 1, NULL, READ_LOCK);
301 int volid = (tvp->roVol ? tvp->roVol : tvp->volume);
302 afs_rootFid.Cell = localcell;
303 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
305 /* If we had a root fid before and it changed location we reset
306 * the afs_globalVp so that it will be reevaluated.
307 * Just decrement the reference count. This only occurs during
308 * initial cell setup and can panic the machine if we set the
309 * count to zero and fs checkv is executed when the current
312 AFS_FAST_RELE(afs_globalVp);
315 afs_rootFid.Fid.Volume = volid;
316 afs_rootFid.Fid.Vnode = 1;
317 afs_rootFid.Fid.Unique = 1;
321 afs_initState = 300; /* won */
322 afs_osi_Wakeup(&afs_initState);
323 afs_PutVolume(tvp, READ_LOCK);
325 if (afs_rootFid.Fid.Volume)
331 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
333 BPath(register struct brequest *ab)
335 register struct dcache *tdc = NULL;
336 struct vcache *tvc = NULL;
337 struct vnode *tvn = NULL;
338 #ifdef AFS_LINUX22_ENV
339 struct dentry *dp = NULL;
341 afs_size_t offset, len;
342 struct vrequest treq;
346 if ((code = afs_InitReq(&treq, ab->cred)))
349 #ifdef AFS_LINUX22_ENV
350 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &dp);
352 tvn = (struct vnode *)dp->d_inode;
354 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &tvn);
357 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
360 /* now path may not have been in afs, so check that before calling our cache manager */
361 if (!tvn || !IsAfsVnode(tvn)) {
362 /* release it and give up */
364 #ifdef AFS_LINUX22_ENV
373 /* here we know its an afs vnode, so we can get the data for the chunk */
374 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
378 #ifdef AFS_LINUX22_ENV
385 /* size_parm 0 to the fetch is the chunk number,
386 * ptr_parm 0 is the dcache entry to wakeup,
387 * size_parm 1 is true iff we should release the dcache entry here.
390 BPrefetch(register struct brequest *ab)
392 register struct dcache *tdc;
393 register struct vcache *tvc;
394 afs_size_t offset, len;
395 struct vrequest treq;
397 AFS_STATCNT(BPrefetch);
398 if ((len = afs_InitReq(&treq, ab->cred)))
401 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
405 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
406 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
407 * be waiting for our wakeup anyway.
409 tdc = (struct dcache *)(ab->ptr_parm[0]);
410 ObtainSharedLock(&tdc->lock, 640);
411 if (tdc->mflags & DFFetchReq) {
412 UpgradeSToWLock(&tdc->lock, 641);
413 tdc->mflags &= ~DFFetchReq;
414 ReleaseWriteLock(&tdc->lock);
416 ReleaseSharedLock(&tdc->lock);
418 afs_osi_Wakeup(&tdc->validPos);
419 if (ab->size_parm[1]) {
420 afs_PutDCache(tdc); /* put this one back, too */
426 BStore(register struct brequest *ab)
428 register struct vcache *tvc;
429 register afs_int32 code;
430 struct vrequest treq;
431 #if defined(AFS_SGI_ENV)
432 struct cred *tmpcred;
436 if ((code = afs_InitReq(&treq, ab->cred)))
440 #if defined(AFS_SGI_ENV)
442 * Since StoreOnLastReference can end up calling osi_SyncVM which
443 * calls into VM code that assumes that u.u_cred has the
444 * correct credentials, we set our to theirs for this xaction
446 tmpcred = OSI_GET_CURRENT_CRED();
447 OSI_SET_CURRENT_CRED(ab->cred);
450 * To avoid recursion since the WriteLock may be released during VM
451 * operations, we hold the VOP_RWLOCK across this transaction as
452 * do the other callers of StoreOnLastReference
454 AFS_RWLOCK((vnode_t *) tvc, 1);
456 ObtainWriteLock(&tvc->lock, 209);
457 code = afs_StoreOnLastReference(tvc, &treq);
458 ReleaseWriteLock(&tvc->lock);
459 #if defined(AFS_SGI_ENV)
460 OSI_SET_CURRENT_CRED(tmpcred);
461 AFS_RWUNLOCK((vnode_t *) tvc, 1);
463 /* now set final return code, and wakeup anyone waiting */
464 if ((ab->flags & BUVALID) == 0) {
465 ab->code = afs_CheckCode(code, &treq, 43); /* set final code, since treq doesn't go across processes */
466 ab->flags |= BUVALID;
467 if (ab->flags & BUWAIT) {
468 ab->flags &= ~BUWAIT;
474 /* release a held request buffer */
476 afs_BRelease(register struct brequest *ab)
479 AFS_STATCNT(afs_BRelease);
480 MObtainWriteLock(&afs_xbrs, 294);
481 if (--ab->refCount <= 0) {
485 afs_osi_Wakeup(&afs_brsWaiters);
486 MReleaseWriteLock(&afs_xbrs);
489 /* return true if bkg fetch daemons are all busy */
493 AFS_STATCNT(afs_BBusy);
494 if (afs_brsDaemons > 0)
500 afs_BQueue(register short aopcode, register struct vcache *avc,
501 afs_int32 dontwait, afs_int32 ause, struct AFS_UCRED *acred,
502 afs_size_t asparm0, afs_size_t asparm1, void *apparm0)
505 register struct brequest *tb;
507 AFS_STATCNT(afs_BQueue);
508 MObtainWriteLock(&afs_xbrs, 296);
511 for (i = 0; i < NBRS; i++, tb++) {
512 if (tb->refCount == 0)
517 tb->opcode = aopcode;
522 VN_HOLD(AFSTOV(avc));
524 tb->refCount = ause + 1;
525 tb->size_parm[0] = asparm0;
526 tb->size_parm[1] = asparm1;
527 tb->ptr_parm[0] = apparm0;
530 tb->ts = afs_brs_count++;
531 /* if daemons are waiting for work, wake them up */
532 if (afs_brsDaemons > 0) {
533 afs_osi_Wakeup(&afs_brsDaemons);
535 MReleaseWriteLock(&afs_xbrs);
539 MReleaseWriteLock(&afs_xbrs);
542 /* no free buffers, sleep a while */
544 MReleaseWriteLock(&afs_xbrs);
545 afs_osi_Sleep(&afs_brsWaiters);
546 MObtainWriteLock(&afs_xbrs, 301);
553 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
554 * The modifications here will work for either a UP or MP machine.
556 struct buf *afs_asyncbuf = (struct buf *)0;
557 tid_t afs_asyncbuf_cv = EVENT_NULL;
558 afs_int32 afs_biodcnt = 0;
560 /* in implementing this, I assumed that all external linked lists were
563 * Several places in this code traverse a linked list. The algorithm
564 * used here is probably unfamiliar to most people. Careful examination
565 * will show that it eliminates an assignment inside the loop, as compared
566 * to the standard algorithm, at the cost of occasionally using an extra
572 * This function obtains, and returns, a pointer to a buffer for
573 * processing by a daemon. It sleeps until such a buffer is available.
574 * The source of buffers for it is the list afs_asyncbuf (see also
575 * naix_vm_strategy). This function may be invoked concurrently by
576 * several processes, that is, several instances of the same daemon.
577 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
578 * level, while get_bioreq runs at process level.
580 * Since AIX 4.1 can wake just one process at a time, the separate sleep
581 * addresses have been removed.
582 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
583 * The afs_asyncbuf_lock is primarily used to serialize access between
584 * process and interrupts.
586 Simple_lock afs_asyncbuf_lock;
587 /*static*/ struct buf *
590 struct buf *bp = NULL;
592 struct buf **bestlbpP, **lbpP;
594 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
597 struct afs_bioqueue *s;
599 /* ??? Does the forward pointer of the returned buffer need to be NULL?
602 /* Disable interrupts from the strategy function, and save the
603 * prior priority level and lock access to the afs_asyncbuf.
606 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock);
610 /* look for oldest buffer */
611 bp = bestbp = afs_asyncbuf;
612 bestage = (long)bestbp->av_back;
613 bestlbpP = &afs_asyncbuf;
619 if ((long)bp->av_back - bestage < 0) {
622 bestage = (long)bp->av_back;
626 *bestlbpP = bp->av_forw;
629 /* If afs_asyncbuf is null, it is necessary to go to sleep.
630 * e_wakeup_one() ensures that only one thread wakes.
633 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
634 * lock on an MP machine.
637 e_sleep_thread(&afs_asyncbuf_cv, &afs_asyncbuf_lock,
638 LOCK_HANDLER | INTERRUPTIBLE);
639 if (interrupted == THREAD_INTERRUPTED) {
640 /* re-enable interrupts from strategy */
641 unlock_enable(oldPriority, &afs_asyncbuf_lock);
645 } /* end of "else asyncbuf is empty" */
646 } /* end of "inner loop" */
650 unlock_enable(oldPriority, &afs_asyncbuf_lock);
653 /* For the convenience of other code, replace the gnodes in
654 * the b_vp field of bp and the other buffers on the b_work
655 * chain with the corresponding vnodes.
657 * ??? what happens to the gnodes? They're not just cut loose,
661 t2P = (struct buf *)t1P->b_work;
662 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
666 t1P = (struct buf *)t2P->b_work;
667 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
672 /* If the buffer does not specify I/O, it may immediately
673 * be returned to the caller. This condition is detected
674 * by examining the buffer's flags (the b_flags field). If
675 * the B_PFPROT bit is set, the buffer represents a protection
676 * violation, rather than a request for I/O. The remainder
677 * of the outer loop handles the case where the B_PFPROT bit is clear.
679 if (bp->b_flags & B_PFPROT) {
684 } /* end of function get_bioreq() */
689 * This function is the daemon. It is called from the syscall
690 * interface. Ordinarily, a script or an administrator will run a
691 * daemon startup utility, specifying the number of I/O daemons to
692 * run. The utility will fork off that number of processes,
693 * each making the appropriate syscall, which will cause this
694 * function to be invoked.
696 static int afs_initbiod = 0; /* this is self-initializing code */
699 afs_BioDaemon(afs_int32 nbiods)
701 afs_int32 code, s, pflg = 0;
703 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
711 /* pin lock, since we'll be using it in an interrupt. */
712 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
713 simple_lock_init(&afs_asyncbuf_lock);
714 pin(&afs_asyncbuf, sizeof(struct buf *));
715 pin(&afs_asyncbuf_cv, sizeof(afs_int32));
718 /* Ignore HUP signals... */
720 sigset_t sigbits, osigbits;
722 * add SIGHUP to the set of already masked signals
724 SIGFILLSET(sigbits); /* allow all signals */
725 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
726 limit_sigs(&sigbits, &osigbits); /* and already masked */
728 /* Main body starts here -- this is an intentional infinite loop, and
731 * Now, the loop will exit if get_bioreq() returns NULL, indicating
732 * that we've been interrupted.
735 bp = afs_get_bioreq();
737 break; /* we were interrupted */
738 if (code = setjmpx(&jmpbuf)) {
739 /* This should not have happend, maybe a lack of resources */
741 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
742 for (bp1 = bp; bp; bp = bp1) {
744 bp1 = (struct buf *)bp1->b_work;
747 bp->b_flags |= B_ERROR;
750 unlock_enable(s, &afs_asyncbuf_lock);
754 vcp = VTOAFS(bp->b_vp);
755 if (bp->b_flags & B_PFSTORE) { /* XXXX */
756 ObtainWriteLock(&vcp->lock, 404);
757 if (vcp->v.v_gnode->gn_mwrcnt) {
758 afs_offs_t newlength =
759 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
760 if (vcp->m.Length < newlength) {
761 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
762 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
763 __LINE__, ICL_TYPE_OFFSET,
764 ICL_HANDLE_OFFSET(vcp->m.Length),
765 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
766 vcp->m.Length = newlength;
769 ReleaseWriteLock(&vcp->lock);
771 /* If the buffer represents a protection violation, rather than
772 * an actual request for I/O, no special action need be taken.
774 if (bp->b_flags & B_PFPROT) {
775 iodone(bp); /* Notify all users of the buffer that we're done */
780 ObtainWriteLock(&vcp->pvmlock, 211);
782 * First map its data area to a region in the current address space
783 * by calling vm_att with the subspace identifier, and a pointer to
784 * the data area. vm_att returns a new data area pointer, but we
785 * also want to hang onto the old one.
787 tmpaddr = bp->b_baddr;
788 bp->b_baddr = (caddr_t) vm_att(bp->b_xmemd.subspace_id, tmpaddr);
789 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
790 if (tmperr) { /* in non-error case */
791 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
792 bp->b_error = tmperr;
795 /* Unmap the buffer's data area by calling vm_det. Reset data area
796 * to the value that we saved above.
799 bp->b_baddr = tmpaddr;
802 * buffer may be linked with other buffers via the b_work field.
803 * See also naix_vm_strategy. For each buffer in the chain (including
804 * bp) notify all users of the buffer that the daemon is finished
805 * using it by calling iodone.
806 * assumes iodone can modify the b_work field.
809 tbp2 = (struct buf *)tbp1->b_work;
814 tbp1 = (struct buf *)tbp2->b_work;
820 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
822 } /* infinite loop (unless we're interrupted) */
823 } /* end of afs_BioDaemon() */
825 #else /* AFS_AIX41_ENV */
829 struct afs_bioqueue {
834 struct afs_bioqueue afs_bioqueue;
835 struct buf *afs_busyq = NULL;
836 struct buf *afs_asyncbuf;
837 afs_int32 afs_biodcnt = 0;
839 /* in implementing this, I assumed that all external linked lists were
842 * Several places in this code traverse a linked list. The algorithm
843 * used here is probably unfamiliar to most people. Careful examination
844 * will show that it eliminates an assignment inside the loop, as compared
845 * to the standard algorithm, at the cost of occasionally using an extra
851 * This function obtains, and returns, a pointer to a buffer for
852 * processing by a daemon. It sleeps until such a buffer is available.
853 * The source of buffers for it is the list afs_asyncbuf (see also
854 * naix_vm_strategy). This function may be invoked concurrently by
855 * several processes, that is, several instances of the same daemon.
856 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
857 * level, while get_bioreq runs at process level.
859 * The common kernel paradigm of sleeping and waking up, in which all the
860 * competing processes sleep waiting for wakeups on one address, is not
861 * followed here. Instead, the following paradigm is used: when a daemon
862 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
863 * it sleeps on the address of variable afs_asyncbuf. But if there is
864 * already a daemon sleeping on that address, it threads its own unique
865 * address onto a list, and sleeps on that address. This way, every
866 * sleeper is sleeping on a different address, and every wakeup wakes up
867 * exactly one daemon. This prevents a whole bunch of daemons from waking
868 * up and then immediately having to go back to sleep. This provides a
869 * performance gain and makes the I/O scheduling a bit more deterministic.
870 * The list of sleepers is variable afs_bioqueue. The unique address
871 * on which to sleep is passed to get_bioreq as its parameter.
873 /*static*/ struct buf *
875 struct afs_bioqueue *self; /* address on which to sleep */
878 struct buf *bp = NULL;
880 struct buf **bestlbpP, **lbpP;
882 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
885 struct afs_bioqueue *s;
887 /* ??? Does the forward pointer of the returned buffer need to be NULL?
890 /* Disable interrupts from the strategy function, and save the
891 * prior priority level
893 oldPriority = i_disable(INTMAX);
895 /* Each iteration of following loop either pulls
896 * a buffer off afs_asyncbuf, or sleeps.
898 while (1) { /* inner loop */
900 /* look for oldest buffer */
901 bp = bestbp = afs_asyncbuf;
902 bestage = (int)bestbp->av_back;
903 bestlbpP = &afs_asyncbuf;
909 if ((int)bp->av_back - bestage < 0) {
912 bestage = (int)bp->av_back;
916 *bestlbpP = bp->av_forw;
921 /* If afs_asyncbuf is null, it is necessary to go to sleep.
922 * There are two possibilities: either there is already a
923 * daemon that is sleeping on the address of afs_asyncbuf,
926 if (afs_bioqueue.sleeper) {
928 QAdd(&(afs_bioqueue.lruq), &(self->lruq));
929 interrupted = sleep((caddr_t) self, PCATCH | (PZERO + 1));
930 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
931 QRemove(&(self->lruq)); /* dequeue */
934 afs_bioqueue.sleeper = FALSE;
936 /* re-enable interrupts from strategy */
937 i_enable(oldPriority);
942 afs_bioqueue.sleeper = TRUE;
944 sleep((caddr_t) & afs_asyncbuf, PCATCH | (PZERO + 1));
945 afs_bioqueue.sleeper = FALSE;
948 * We need to wakeup another daemon if present
949 * since we were waiting on afs_asyncbuf.
951 #ifdef notdef /* The following doesn't work as advertised */
952 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq) {
953 struct squeue *bq = afs_bioqueue.lruq.next;
958 /* re-enable interrupts from strategy */
959 i_enable(oldPriority);
965 } /* end of "else asyncbuf is empty" */
966 } /* end of "inner loop" */
970 i_enable(oldPriority); /* re-enable interrupts from strategy */
972 /* For the convenience of other code, replace the gnodes in
973 * the b_vp field of bp and the other buffers on the b_work
974 * chain with the corresponding vnodes.
976 * ??? what happens to the gnodes? They're not just cut loose,
980 t2P = (struct buf *)t1P->b_work;
981 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
985 t1P = (struct buf *)t2P->b_work;
986 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
991 /* If the buffer does not specify I/O, it may immediately
992 * be returned to the caller. This condition is detected
993 * by examining the buffer's flags (the b_flags field). If
994 * the B_PFPROT bit is set, the buffer represents a protection
995 * violation, rather than a request for I/O. The remainder
996 * of the outer loop handles the case where the B_PFPROT bit is clear.
998 if (bp->b_flags & B_PFPROT) {
1002 /* wake up another process to handle the next buffer, and return
1005 oldPriority = i_disable(INTMAX);
1007 /* determine where to find the sleeping process.
1008 * There are two cases: either it is sleeping on
1009 * afs_asyncbuf, or it is sleeping on its own unique
1010 * address. These cases are distinguished by examining
1011 * the sleeper field of afs_bioqueue.
1013 if (afs_bioqueue.sleeper) {
1014 wakeup(&afs_asyncbuf);
1016 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1017 /* queue is empty, what now? ??? */
1018 /* Should this be impossible, or does */
1019 /* it just mean that nobody is sleeping? */ ;
1021 struct squeue *bq = afs_bioqueue.lruq.next;
1025 afs_bioqueue.sleeper = TRUE;
1028 i_enable(oldPriority); /* re-enable interrupts from strategy */
1031 } /* end of function get_bioreq() */
1036 * This function is the daemon. It is called from the syscall
1037 * interface. Ordinarily, a script or an administrator will run a
1038 * daemon startup utility, specifying the number of I/O daemons to
1039 * run. The utility will fork off that number of processes,
1040 * each making the appropriate syscall, which will cause this
1041 * function to be invoked.
1043 static int afs_initbiod = 0; /* this is self-initializing code */
1045 afs_BioDaemon(nbiods)
1048 struct afs_bioqueue *self;
1049 afs_int32 code, s, pflg = 0;
1051 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1056 if (!afs_initbiod) {
1059 /* Initialize the queue of waiting processes, afs_bioqueue. */
1060 QInit(&(afs_bioqueue.lruq));
1063 /* establish ourself as a kernel process so shutdown won't kill us */
1064 /* u.u_procp->p_flag |= SKPROC;*/
1066 /* Initialize a token (self) to use in the queue of sleeping processes. */
1067 self = (struct afs_bioqueue *)afs_osi_Alloc(sizeof(struct afs_bioqueue));
1068 pin(self, sizeof(struct afs_bioqueue)); /* fix in memory */
1069 memset(self, 0, sizeof(*self));
1070 QInit(&(self->lruq)); /* initialize queue entry pointers */
1073 /* Ignore HUP signals... */
1074 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1075 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1076 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1077 /* Main body starts here -- this is an intentional infinite loop, and
1080 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1081 * that we've been interrupted.
1084 bp = afs_get_bioreq(self);
1086 break; /* we were interrupted */
1087 if (code = setjmpx(&jmpbuf)) {
1088 /* This should not have happend, maybe a lack of resources */
1090 for (bp1 = bp; bp; bp = bp1) {
1095 bp->b_flags |= B_ERROR;
1101 vcp = VTOAFS(bp->b_vp);
1102 if (bp->b_flags & B_PFSTORE) {
1103 ObtainWriteLock(&vcp->lock, 210);
1104 if (vcp->v.v_gnode->gn_mwrcnt) {
1105 afs_offs_t newlength =
1106 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
1107 if (vcp->m.Length < newlength) {
1108 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
1109 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
1110 __LINE__, ICL_TYPE_OFFSET,
1111 ICL_HANDLE_OFFSET(vcp->m.Length),
1112 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
1113 vcp->m.Length = newlength;
1116 ReleaseWriteLock(&vcp->lock);
1118 /* If the buffer represents a protection violation, rather than
1119 * an actual request for I/O, no special action need be taken.
1121 if (bp->b_flags & B_PFPROT) {
1122 iodone(bp); /* Notify all users of the buffer that we're done */
1126 ObtainWriteLock(&vcp->pvmlock, 558);
1128 * First map its data area to a region in the current address space
1129 * by calling vm_att with the subspace identifier, and a pointer to
1130 * the data area. vm_att returns a new data area pointer, but we
1131 * also want to hang onto the old one.
1133 tmpaddr = bp->b_baddr;
1134 bp->b_baddr = vm_att(bp->b_xmemd.subspace_id, tmpaddr);
1135 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1136 if (tmperr) { /* in non-error case */
1137 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1138 bp->b_error = tmperr;
1141 /* Unmap the buffer's data area by calling vm_det. Reset data area
1142 * to the value that we saved above.
1144 vm_det(bp->b_un.b_addr);
1145 bp->b_baddr = tmpaddr;
1148 * buffer may be linked with other buffers via the b_work field.
1149 * See also naix_vm_strategy. For each buffer in the chain (including
1150 * bp) notify all users of the buffer that the daemon is finished
1151 * using it by calling iodone.
1152 * assumes iodone can modify the b_work field.
1155 tbp2 = (struct buf *)tbp1->b_work;
1160 tbp1 = (struct buf *)tbp2->b_work;
1166 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1168 } /* infinite loop (unless we're interrupted) */
1169 unpin(self, sizeof(struct afs_bioqueue));
1170 afs_osi_Free(self, sizeof(struct afs_bioqueue));
1171 } /* end of afs_BioDaemon() */
1172 #endif /* AFS_AIX41_ENV */
1173 #endif /* AFS_AIX32_ENV */
1178 afs_BackgroundDaemon(void)
1180 struct brequest *tb;
1183 AFS_STATCNT(afs_BackgroundDaemon);
1184 /* initialize subsystem */
1186 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1187 memset((char *)afs_brs, 0, sizeof(afs_brs));
1189 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1191 * steal the first daemon for doing delayed DSlot flushing
1192 * (see afs_GetDownDSlot)
1201 MObtainWriteLock(&afs_xbrs, 302);
1204 struct brequest *min_tb = NULL;
1206 if (afs_termState == AFSOP_STOP_BKG) {
1207 if (--afs_nbrs <= 0)
1208 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1209 MReleaseWriteLock(&afs_xbrs);
1210 afs_osi_Wakeup(&afs_termState);
1214 /* find a request */
1217 for (i = 0; i < NBRS; i++, tb++) {
1218 /* look for request with smallest ts */
1219 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1220 /* new request, not yet picked up */
1221 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1227 if ((tb = min_tb)) {
1228 /* claim and process this request */
1229 tb->flags |= BSTARTED;
1230 MReleaseWriteLock(&afs_xbrs);
1232 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1, ICL_TYPE_INT32,
1234 if (tb->opcode == BOP_FETCH)
1236 else if (tb->opcode == BOP_STORE)
1238 else if (tb->opcode == BOP_PATH)
1241 panic("background bop");
1243 AFS_RELE(AFSTOV(tb->vc)); /* MUST call vnode layer or could lose vnodes */
1248 tb->cred = (struct AFS_UCRED *)0;
1250 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1251 MObtainWriteLock(&afs_xbrs, 305);
1254 /* wait for new request */
1256 MReleaseWriteLock(&afs_xbrs);
1257 afs_osi_Sleep(&afs_brsDaemons);
1258 MObtainWriteLock(&afs_xbrs, 307);
1266 shutdown_daemons(void)
1268 AFS_STATCNT(shutdown_daemons);
1269 if (afs_cold_shutdown) {
1270 afs_brsDaemons = brsInit = 0;
1271 rxepoch_checked = afs_nbrs = 0;
1272 memset((char *)afs_brs, 0, sizeof(afs_brs));
1273 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1275 #ifdef AFS_AIX32_ENV
1276 #ifdef AFS_AIX41_ENV
1277 lock_free(&afs_asyncbuf_lock);
1278 unpin(&afs_asyncbuf, sizeof(struct buf *));
1279 unpin(&afs_asyncbuf_cv, sizeof(afs_int32));
1280 #else /* AFS_AIX41_ENV */
1283 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1290 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1292 * sgi - daemon - handles certain operations that otherwise
1293 * would use up too much kernel stack space
1295 * This all assumes that since the caller must have the xdcache lock
1296 * exclusively that the list will never be more than one long
1297 * and noone else can attempt to add anything until we're done.
1299 SV_TYPE afs_sgibksync;
1300 SV_TYPE afs_sgibkwait;
1301 lock_t afs_sgibklock;
1302 struct dcache *afs_sgibklist;
1310 if (afs_sgibklock == NULL) {
1311 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1312 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1313 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1315 s = SPLOCK(afs_sgibklock);
1317 /* wait for something to do */
1318 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1319 osi_Assert(afs_sgibklist);
1321 /* XX will probably need to generalize to real list someday */
1322 s = SPLOCK(afs_sgibklock);
1323 while (afs_sgibklist) {
1324 tdc = afs_sgibklist;
1325 afs_sgibklist = NULL;
1326 SPUNLOCK(afs_sgibklock, s);
1328 tdc->dflags &= ~DFEntryMod;
1329 afs_WriteDCache(tdc, 1);
1331 s = SPLOCK(afs_sgibklock);
1334 /* done all the work - wake everyone up */
1335 while (SV_SIGNAL(&afs_sgibkwait));