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 struct cell *lc = afs_GetPrimaryCell(READ_LOCK);
300 localcell = lc->cellNum;
301 afs_PutCell(lc, READ_LOCK);
305 afs_GetDynrootFid(&afs_rootFid);
306 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
308 tvp = afs_GetVolumeByName(rootVolName, localcell, 1, NULL, READ_LOCK);
310 if (!tvp && !usingDynroot) {
312 int len = strlen(rootVolName);
314 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
315 strcpy(buf, rootVolName);
316 afs_strcat(buf, ".readonly");
317 tvp = afs_GetVolumeByName(buf, localcell, 1, NULL, READ_LOCK);
322 int volid = (tvp->roVol ? tvp->roVol : tvp->volume);
323 afs_rootFid.Cell = localcell;
324 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
326 /* If we had a root fid before and it changed location we reset
327 * the afs_globalVp so that it will be reevaluated.
328 * Just decrement the reference count. This only occurs during
329 * initial cell setup and can panic the machine if we set the
330 * count to zero and fs checkv is executed when the current
333 AFS_FAST_RELE(afs_globalVp);
336 afs_rootFid.Fid.Volume = volid;
337 afs_rootFid.Fid.Vnode = 1;
338 afs_rootFid.Fid.Unique = 1;
340 afs_initState = 300; /* won */
341 afs_osi_Wakeup(&afs_initState);
342 afs_PutVolume(tvp, READ_LOCK);
345 /* This is to make sure that we update the root gnode */
346 /* every time root volume gets released */
348 struct gnode *rootgp;
352 /* Only do this if afs_globalVFS is properly set due to race conditions
353 * this routine could be called before the gfs_mount is performed!
354 * Furthermore, afs_root (called below) *waits* until
355 * initState >= 200, so we don't try this until we've gotten
356 * at least that far */
357 if (afs_globalVFS && afs_initState >= 200) {
358 if (code = afs_root(afs_globalVFS, &rootgp))
360 mp = (struct mount *)afs_globalVFS->vfs_data;
361 mp->m_rootgp = gget(mp, 0, 0, (char *)rootgp);
362 afs_unlock(mp->m_rootgp); /* unlock basic gnode */
363 afs_vrele(VTOAFS(rootgp)); /* zap afs_root's vnode hold */
367 if (afs_rootFid.Fid.Volume)
373 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
375 BPath(register struct brequest *ab)
377 register struct dcache *tdc = NULL;
378 struct vcache *tvc = NULL;
379 struct vnode *tvn = NULL;
380 #ifdef AFS_LINUX22_ENV
381 struct dentry *dp = NULL;
383 afs_size_t offset, len;
384 struct vrequest treq;
388 if ((code = afs_InitReq(&treq, ab->cred)))
391 #ifdef AFS_LINUX22_ENV
392 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, NULL, &dp);
394 tvn = (struct vnode *)dp->d_inode;
396 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, NULL, &tvn);
399 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
402 /* now path may not have been in afs, so check that before calling our cache manager */
403 if (!tvn || !IsAfsVnode(tvn)) {
404 /* release it and give up */
409 #ifdef AFS_LINUX22_ENV
419 tvc = VTOAFS(afs_gntovn(tvn));
423 /* here we know its an afs vnode, so we can get the data for the chunk */
424 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
431 #ifdef AFS_LINUX22_ENV
439 /* size_parm 0 to the fetch is the chunk number,
440 * ptr_parm 0 is the dcache entry to wakeup,
441 * size_parm 1 is true iff we should release the dcache entry here.
444 BPrefetch(register struct brequest *ab)
446 register struct dcache *tdc;
447 register struct vcache *tvc;
448 afs_size_t offset, len;
449 struct vrequest treq;
451 AFS_STATCNT(BPrefetch);
452 if ((len = afs_InitReq(&treq, ab->cred)))
455 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
459 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
460 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
461 * be waiting for our wakeup anyway.
463 tdc = (struct dcache *)(ab->ptr_parm[0]);
464 ObtainSharedLock(&tdc->lock, 640);
465 if (tdc->mflags & DFFetchReq) {
466 UpgradeSToWLock(&tdc->lock, 641);
467 tdc->mflags &= ~DFFetchReq;
468 ReleaseWriteLock(&tdc->lock);
470 ReleaseSharedLock(&tdc->lock);
472 afs_osi_Wakeup(&tdc->validPos);
473 if (ab->size_parm[1]) {
474 afs_PutDCache(tdc); /* put this one back, too */
480 BStore(register struct brequest *ab)
482 register struct vcache *tvc;
483 register afs_int32 code;
484 struct vrequest treq;
485 #if defined(AFS_SGI_ENV)
486 struct cred *tmpcred;
490 if ((code = afs_InitReq(&treq, ab->cred)))
494 #if defined(AFS_SGI_ENV)
496 * Since StoreOnLastReference can end up calling osi_SyncVM which
497 * calls into VM code that assumes that u.u_cred has the
498 * correct credentials, we set our to theirs for this xaction
500 tmpcred = OSI_GET_CURRENT_CRED();
501 OSI_SET_CURRENT_CRED(ab->cred);
504 * To avoid recursion since the WriteLock may be released during VM
505 * operations, we hold the VOP_RWLOCK across this transaction as
506 * do the other callers of StoreOnLastReference
508 AFS_RWLOCK((vnode_t *) tvc, 1);
510 ObtainWriteLock(&tvc->lock, 209);
511 code = afs_StoreOnLastReference(tvc, &treq);
512 ReleaseWriteLock(&tvc->lock);
513 #if defined(AFS_SGI_ENV)
514 OSI_SET_CURRENT_CRED(tmpcred);
515 AFS_RWUNLOCK((vnode_t *) tvc, 1);
517 /* now set final return code, and wakeup anyone waiting */
518 if ((ab->flags & BUVALID) == 0) {
519 ab->code = afs_CheckCode(code, &treq, 43); /* set final code, since treq doesn't go across processes */
520 ab->flags |= BUVALID;
521 if (ab->flags & BUWAIT) {
522 ab->flags &= ~BUWAIT;
528 /* release a held request buffer */
530 afs_BRelease(register struct brequest *ab)
533 AFS_STATCNT(afs_BRelease);
534 MObtainWriteLock(&afs_xbrs, 294);
535 if (--ab->refCount <= 0) {
539 afs_osi_Wakeup(&afs_brsWaiters);
540 MReleaseWriteLock(&afs_xbrs);
543 /* return true if bkg fetch daemons are all busy */
547 AFS_STATCNT(afs_BBusy);
548 if (afs_brsDaemons > 0)
554 afs_BQueue(register short aopcode, register struct vcache *avc,
555 afs_int32 dontwait, afs_int32 ause, struct AFS_UCRED *acred,
556 afs_size_t asparm0, afs_size_t asparm1, void *apparm0)
559 register struct brequest *tb;
561 AFS_STATCNT(afs_BQueue);
562 MObtainWriteLock(&afs_xbrs, 296);
565 for (i = 0; i < NBRS; i++, tb++) {
566 if (tb->refCount == 0)
571 tb->opcode = aopcode;
579 VN_HOLD(AFSTOV(avc));
582 tb->refCount = ause + 1;
583 tb->size_parm[0] = asparm0;
584 tb->size_parm[1] = asparm1;
585 tb->ptr_parm[0] = apparm0;
588 tb->ts = afs_brs_count++;
589 /* if daemons are waiting for work, wake them up */
590 if (afs_brsDaemons > 0) {
591 afs_osi_Wakeup(&afs_brsDaemons);
593 MReleaseWriteLock(&afs_xbrs);
597 MReleaseWriteLock(&afs_xbrs);
600 /* no free buffers, sleep a while */
602 MReleaseWriteLock(&afs_xbrs);
603 afs_osi_Sleep(&afs_brsWaiters);
604 MObtainWriteLock(&afs_xbrs, 301);
611 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
612 * The modifications here will work for either a UP or MP machine.
614 struct buf *afs_asyncbuf = (struct buf *)0;
615 tid_t afs_asyncbuf_cv = EVENT_NULL;
616 afs_int32 afs_biodcnt = 0;
618 /* in implementing this, I assumed that all external linked lists were
621 * Several places in this code traverse a linked list. The algorithm
622 * used here is probably unfamiliar to most people. Careful examination
623 * will show that it eliminates an assignment inside the loop, as compared
624 * to the standard algorithm, at the cost of occasionally using an extra
630 * This function obtains, and returns, a pointer to a buffer for
631 * processing by a daemon. It sleeps until such a buffer is available.
632 * The source of buffers for it is the list afs_asyncbuf (see also
633 * naix_vm_strategy). This function may be invoked concurrently by
634 * several processes, that is, several instances of the same daemon.
635 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
636 * level, while get_bioreq runs at process level.
638 * Since AIX 4.1 can wake just one process at a time, the separate sleep
639 * addresses have been removed.
640 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
641 * The afs_asyncbuf_lock is primarily used to serialize access between
642 * process and interrupts.
644 Simple_lock afs_asyncbuf_lock;
645 /*static*/ struct buf *
648 struct buf *bp = NULL;
650 struct buf **bestlbpP, **lbpP;
652 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
655 struct afs_bioqueue *s;
657 /* ??? Does the forward pointer of the returned buffer need to be NULL?
660 /* Disable interrupts from the strategy function, and save the
661 * prior priority level and lock access to the afs_asyncbuf.
664 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock);
668 /* look for oldest buffer */
669 bp = bestbp = afs_asyncbuf;
670 bestage = (long)bestbp->av_back;
671 bestlbpP = &afs_asyncbuf;
677 if ((long)bp->av_back - bestage < 0) {
680 bestage = (long)bp->av_back;
684 *bestlbpP = bp->av_forw;
687 /* If afs_asyncbuf is null, it is necessary to go to sleep.
688 * e_wakeup_one() ensures that only one thread wakes.
691 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
692 * lock on an MP machine.
695 e_sleep_thread(&afs_asyncbuf_cv, &afs_asyncbuf_lock,
696 LOCK_HANDLER | INTERRUPTIBLE);
697 if (interrupted == THREAD_INTERRUPTED) {
698 /* re-enable interrupts from strategy */
699 unlock_enable(oldPriority, &afs_asyncbuf_lock);
703 } /* end of "else asyncbuf is empty" */
704 } /* end of "inner loop" */
708 unlock_enable(oldPriority, &afs_asyncbuf_lock);
711 /* For the convenience of other code, replace the gnodes in
712 * the b_vp field of bp and the other buffers on the b_work
713 * chain with the corresponding vnodes.
715 * ??? what happens to the gnodes? They're not just cut loose,
719 t2P = (struct buf *)t1P->b_work;
720 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
724 t1P = (struct buf *)t2P->b_work;
725 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
730 /* If the buffer does not specify I/O, it may immediately
731 * be returned to the caller. This condition is detected
732 * by examining the buffer's flags (the b_flags field). If
733 * the B_PFPROT bit is set, the buffer represents a protection
734 * violation, rather than a request for I/O. The remainder
735 * of the outer loop handles the case where the B_PFPROT bit is clear.
737 if (bp->b_flags & B_PFPROT) {
742 } /* end of function get_bioreq() */
747 * This function is the daemon. It is called from the syscall
748 * interface. Ordinarily, a script or an administrator will run a
749 * daemon startup utility, specifying the number of I/O daemons to
750 * run. The utility will fork off that number of processes,
751 * each making the appropriate syscall, which will cause this
752 * function to be invoked.
754 static int afs_initbiod = 0; /* this is self-initializing code */
757 afs_BioDaemon(afs_int32 nbiods)
759 afs_int32 code, s, pflg = 0;
761 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
769 /* pin lock, since we'll be using it in an interrupt. */
770 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
771 simple_lock_init(&afs_asyncbuf_lock);
772 pin(&afs_asyncbuf, sizeof(struct buf *));
773 pin(&afs_asyncbuf_cv, sizeof(afs_int32));
776 /* Ignore HUP signals... */
778 sigset_t sigbits, osigbits;
780 * add SIGHUP to the set of already masked signals
782 SIGFILLSET(sigbits); /* allow all signals */
783 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
784 limit_sigs(&sigbits, &osigbits); /* and already masked */
786 /* Main body starts here -- this is an intentional infinite loop, and
789 * Now, the loop will exit if get_bioreq() returns NULL, indicating
790 * that we've been interrupted.
793 bp = afs_get_bioreq();
795 break; /* we were interrupted */
796 if (code = setjmpx(&jmpbuf)) {
797 /* This should not have happend, maybe a lack of resources */
799 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
800 for (bp1 = bp; bp; bp = bp1) {
802 bp1 = (struct buf *)bp1->b_work;
805 bp->b_flags |= B_ERROR;
808 unlock_enable(s, &afs_asyncbuf_lock);
812 vcp = VTOAFS(bp->b_vp);
813 if (bp->b_flags & B_PFSTORE) { /* XXXX */
814 ObtainWriteLock(&vcp->lock, 404);
815 if (vcp->v.v_gnode->gn_mwrcnt) {
816 afs_offs_t newlength =
817 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
818 if (vcp->m.Length < newlength) {
819 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
820 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
821 __LINE__, ICL_TYPE_OFFSET,
822 ICL_HANDLE_OFFSET(vcp->m.Length),
823 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
824 vcp->m.Length = newlength;
827 ReleaseWriteLock(&vcp->lock);
829 /* If the buffer represents a protection violation, rather than
830 * an actual request for I/O, no special action need be taken.
832 if (bp->b_flags & B_PFPROT) {
833 iodone(bp); /* Notify all users of the buffer that we're done */
838 ObtainWriteLock(&vcp->pvmlock, 211);
840 * First map its data area to a region in the current address space
841 * by calling vm_att with the subspace identifier, and a pointer to
842 * the data area. vm_att returns a new data area pointer, but we
843 * also want to hang onto the old one.
845 tmpaddr = bp->b_baddr;
846 bp->b_baddr = (caddr_t) vm_att(bp->b_xmemd.subspace_id, tmpaddr);
847 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
848 if (tmperr) { /* in non-error case */
849 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
850 bp->b_error = tmperr;
853 /* Unmap the buffer's data area by calling vm_det. Reset data area
854 * to the value that we saved above.
857 bp->b_baddr = tmpaddr;
860 * buffer may be linked with other buffers via the b_work field.
861 * See also naix_vm_strategy. For each buffer in the chain (including
862 * bp) notify all users of the buffer that the daemon is finished
863 * using it by calling iodone.
864 * assumes iodone can modify the b_work field.
867 tbp2 = (struct buf *)tbp1->b_work;
872 tbp1 = (struct buf *)tbp2->b_work;
878 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
880 } /* infinite loop (unless we're interrupted) */
881 } /* end of afs_BioDaemon() */
883 #else /* AFS_AIX41_ENV */
887 struct afs_bioqueue {
892 struct afs_bioqueue afs_bioqueue;
893 struct buf *afs_busyq = NULL;
894 struct buf *afs_asyncbuf;
895 afs_int32 afs_biodcnt = 0;
897 /* in implementing this, I assumed that all external linked lists were
900 * Several places in this code traverse a linked list. The algorithm
901 * used here is probably unfamiliar to most people. Careful examination
902 * will show that it eliminates an assignment inside the loop, as compared
903 * to the standard algorithm, at the cost of occasionally using an extra
909 * This function obtains, and returns, a pointer to a buffer for
910 * processing by a daemon. It sleeps until such a buffer is available.
911 * The source of buffers for it is the list afs_asyncbuf (see also
912 * naix_vm_strategy). This function may be invoked concurrently by
913 * several processes, that is, several instances of the same daemon.
914 * naix_vm_strategy, which adds buffers to the list, runs at interrupt
915 * level, while get_bioreq runs at process level.
917 * The common kernel paradigm of sleeping and waking up, in which all the
918 * competing processes sleep waiting for wakeups on one address, is not
919 * followed here. Instead, the following paradigm is used: when a daemon
920 * goes to sleep, it checks for other sleeping daemons. If there aren't any,
921 * it sleeps on the address of variable afs_asyncbuf. But if there is
922 * already a daemon sleeping on that address, it threads its own unique
923 * address onto a list, and sleeps on that address. This way, every
924 * sleeper is sleeping on a different address, and every wakeup wakes up
925 * exactly one daemon. This prevents a whole bunch of daemons from waking
926 * up and then immediately having to go back to sleep. This provides a
927 * performance gain and makes the I/O scheduling a bit more deterministic.
928 * The list of sleepers is variable afs_bioqueue. The unique address
929 * on which to sleep is passed to get_bioreq as its parameter.
931 /*static*/ struct buf *
933 struct afs_bioqueue *self; /* address on which to sleep */
936 struct buf *bp = NULL;
938 struct buf **bestlbpP, **lbpP;
940 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
943 struct afs_bioqueue *s;
945 /* ??? Does the forward pointer of the returned buffer need to be NULL?
948 /* Disable interrupts from the strategy function, and save the
949 * prior priority level
951 oldPriority = i_disable(INTMAX);
953 /* Each iteration of following loop either pulls
954 * a buffer off afs_asyncbuf, or sleeps.
956 while (1) { /* inner loop */
958 /* look for oldest buffer */
959 bp = bestbp = afs_asyncbuf;
960 bestage = (int)bestbp->av_back;
961 bestlbpP = &afs_asyncbuf;
967 if ((int)bp->av_back - bestage < 0) {
970 bestage = (int)bp->av_back;
974 *bestlbpP = bp->av_forw;
979 /* If afs_asyncbuf is null, it is necessary to go to sleep.
980 * There are two possibilities: either there is already a
981 * daemon that is sleeping on the address of afs_asyncbuf,
984 if (afs_bioqueue.sleeper) {
986 QAdd(&(afs_bioqueue.lruq), &(self->lruq));
987 interrupted = sleep((caddr_t) self, PCATCH | (PZERO + 1));
988 if (self->lruq.next != &self->lruq) { /* XXX ##3 XXX */
989 QRemove(&(self->lruq)); /* dequeue */
992 afs_bioqueue.sleeper = FALSE;
994 /* re-enable interrupts from strategy */
995 i_enable(oldPriority);
1000 afs_bioqueue.sleeper = TRUE;
1002 sleep((caddr_t) & afs_asyncbuf, PCATCH | (PZERO + 1));
1003 afs_bioqueue.sleeper = FALSE;
1006 * We need to wakeup another daemon if present
1007 * since we were waiting on afs_asyncbuf.
1009 #ifdef notdef /* The following doesn't work as advertised */
1010 if (afs_bioqueue.lruq.next != &afs_bioqueue.lruq) {
1011 struct squeue *bq = afs_bioqueue.lruq.next;
1016 /* re-enable interrupts from strategy */
1017 i_enable(oldPriority);
1023 } /* end of "else asyncbuf is empty" */
1024 } /* end of "inner loop" */
1028 i_enable(oldPriority); /* re-enable interrupts from strategy */
1030 /* For the convenience of other code, replace the gnodes in
1031 * the b_vp field of bp and the other buffers on the b_work
1032 * chain with the corresponding vnodes.
1034 * ??? what happens to the gnodes? They're not just cut loose,
1038 t2P = (struct buf *)t1P->b_work;
1039 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
1043 t1P = (struct buf *)t2P->b_work;
1044 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
1049 /* If the buffer does not specify I/O, it may immediately
1050 * be returned to the caller. This condition is detected
1051 * by examining the buffer's flags (the b_flags field). If
1052 * the B_PFPROT bit is set, the buffer represents a protection
1053 * violation, rather than a request for I/O. The remainder
1054 * of the outer loop handles the case where the B_PFPROT bit is clear.
1056 if (bp->b_flags & B_PFPROT) {
1060 /* wake up another process to handle the next buffer, and return
1063 oldPriority = i_disable(INTMAX);
1065 /* determine where to find the sleeping process.
1066 * There are two cases: either it is sleeping on
1067 * afs_asyncbuf, or it is sleeping on its own unique
1068 * address. These cases are distinguished by examining
1069 * the sleeper field of afs_bioqueue.
1071 if (afs_bioqueue.sleeper) {
1072 wakeup(&afs_asyncbuf);
1074 if (afs_bioqueue.lruq.next == &afs_bioqueue.lruq) {
1075 /* queue is empty, what now? ??? */
1076 /* Should this be impossible, or does */
1077 /* it just mean that nobody is sleeping? */ ;
1079 struct squeue *bq = afs_bioqueue.lruq.next;
1083 afs_bioqueue.sleeper = TRUE;
1086 i_enable(oldPriority); /* re-enable interrupts from strategy */
1089 } /* end of function get_bioreq() */
1094 * This function is the daemon. It is called from the syscall
1095 * interface. Ordinarily, a script or an administrator will run a
1096 * daemon startup utility, specifying the number of I/O daemons to
1097 * run. The utility will fork off that number of processes,
1098 * each making the appropriate syscall, which will cause this
1099 * function to be invoked.
1101 static int afs_initbiod = 0; /* this is self-initializing code */
1103 afs_BioDaemon(nbiods)
1106 struct afs_bioqueue *self;
1107 afs_int32 code, s, pflg = 0;
1109 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
1114 if (!afs_initbiod) {
1117 /* Initialize the queue of waiting processes, afs_bioqueue. */
1118 QInit(&(afs_bioqueue.lruq));
1121 /* establish ourself as a kernel process so shutdown won't kill us */
1122 /* u.u_procp->p_flag |= SKPROC;*/
1124 /* Initialize a token (self) to use in the queue of sleeping processes. */
1125 self = (struct afs_bioqueue *)afs_osi_Alloc(sizeof(struct afs_bioqueue));
1126 pin(self, sizeof(struct afs_bioqueue)); /* fix in memory */
1127 memset(self, 0, sizeof(*self));
1128 QInit(&(self->lruq)); /* initialize queue entry pointers */
1131 /* Ignore HUP signals... */
1132 SIGDELSET(u.u_procp->p_sig, SIGHUP);
1133 SIGADDSET(u.u_procp->p_sigignore, SIGHUP);
1134 SIGDELSET(u.u_procp->p_sigcatch, SIGHUP);
1135 /* Main body starts here -- this is an intentional infinite loop, and
1138 * Now, the loop will exit if get_bioreq() returns NULL, indicating
1139 * that we've been interrupted.
1142 bp = afs_get_bioreq(self);
1144 break; /* we were interrupted */
1145 if (code = setjmpx(&jmpbuf)) {
1146 /* This should not have happend, maybe a lack of resources */
1148 for (bp1 = bp; bp; bp = bp1) {
1153 bp->b_flags |= B_ERROR;
1159 vcp = VTOAFS(bp->b_vp);
1160 if (bp->b_flags & B_PFSTORE) {
1161 ObtainWriteLock(&vcp->lock, 210);
1162 if (vcp->v.v_gnode->gn_mwrcnt) {
1163 afs_offs_t newlength =
1164 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
1165 if (vcp->m.Length < newlength) {
1166 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
1167 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
1168 __LINE__, ICL_TYPE_OFFSET,
1169 ICL_HANDLE_OFFSET(vcp->m.Length),
1170 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
1171 vcp->m.Length = newlength;
1174 ReleaseWriteLock(&vcp->lock);
1176 /* If the buffer represents a protection violation, rather than
1177 * an actual request for I/O, no special action need be taken.
1179 if (bp->b_flags & B_PFPROT) {
1180 iodone(bp); /* Notify all users of the buffer that we're done */
1184 ObtainWriteLock(&vcp->pvmlock, 558);
1186 * First map its data area to a region in the current address space
1187 * by calling vm_att with the subspace identifier, and a pointer to
1188 * the data area. vm_att returns a new data area pointer, but we
1189 * also want to hang onto the old one.
1191 tmpaddr = bp->b_baddr;
1192 bp->b_baddr = vm_att(bp->b_xmemd.subspace_id, tmpaddr);
1193 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
1194 if (tmperr) { /* in non-error case */
1195 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
1196 bp->b_error = tmperr;
1199 /* Unmap the buffer's data area by calling vm_det. Reset data area
1200 * to the value that we saved above.
1202 vm_det(bp->b_un.b_addr);
1203 bp->b_baddr = tmpaddr;
1206 * buffer may be linked with other buffers via the b_work field.
1207 * See also naix_vm_strategy. For each buffer in the chain (including
1208 * bp) notify all users of the buffer that the daemon is finished
1209 * using it by calling iodone.
1210 * assumes iodone can modify the b_work field.
1213 tbp2 = (struct buf *)tbp1->b_work;
1218 tbp1 = (struct buf *)tbp2->b_work;
1224 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1226 } /* infinite loop (unless we're interrupted) */
1227 unpin(self, sizeof(struct afs_bioqueue));
1228 afs_osi_Free(self, sizeof(struct afs_bioqueue));
1229 } /* end of afs_BioDaemon() */
1230 #endif /* AFS_AIX41_ENV */
1231 #endif /* AFS_AIX32_ENV */
1236 afs_BackgroundDaemon(void)
1238 struct brequest *tb;
1241 AFS_STATCNT(afs_BackgroundDaemon);
1242 /* initialize subsystem */
1244 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1245 memset((char *)afs_brs, 0, sizeof(afs_brs));
1247 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1249 * steal the first daemon for doing delayed DSlot flushing
1250 * (see afs_GetDownDSlot)
1259 MObtainWriteLock(&afs_xbrs, 302);
1262 struct brequest *min_tb = NULL;
1264 if (afs_termState == AFSOP_STOP_BKG) {
1265 if (--afs_nbrs <= 0)
1266 afs_termState = AFSOP_STOP_TRUNCDAEMON;
1267 MReleaseWriteLock(&afs_xbrs);
1268 afs_osi_Wakeup(&afs_termState);
1272 /* find a request */
1275 for (i = 0; i < NBRS; i++, tb++) {
1276 /* look for request with smallest ts */
1277 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1278 /* new request, not yet picked up */
1279 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1285 if ((tb = min_tb)) {
1286 /* claim and process this request */
1287 tb->flags |= BSTARTED;
1288 MReleaseWriteLock(&afs_xbrs);
1290 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1, ICL_TYPE_INT32,
1292 if (tb->opcode == BOP_FETCH)
1294 else if (tb->opcode == BOP_STORE)
1296 else if (tb->opcode == BOP_PATH)
1299 panic("background bop");
1302 tb->vc->vrefCount--; /* fix up reference count */
1304 AFS_RELE(AFSTOV(tb->vc)); /* MUST call vnode layer or could lose vnodes */
1310 tb->cred = (struct AFS_UCRED *)0;
1312 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1313 MObtainWriteLock(&afs_xbrs, 305);
1316 /* wait for new request */
1318 MReleaseWriteLock(&afs_xbrs);
1319 afs_osi_Sleep(&afs_brsDaemons);
1320 MObtainWriteLock(&afs_xbrs, 307);
1328 shutdown_daemons(void)
1330 AFS_STATCNT(shutdown_daemons);
1331 if (afs_cold_shutdown) {
1332 afs_brsDaemons = brsInit = 0;
1333 rxepoch_checked = afs_nbrs = 0;
1334 memset((char *)afs_brs, 0, sizeof(afs_brs));
1335 memset((char *)&afs_xbrs, 0, sizeof(afs_lock_t));
1337 #ifdef AFS_AIX32_ENV
1338 #ifdef AFS_AIX41_ENV
1339 lock_free(&afs_asyncbuf_lock);
1340 unpin(&afs_asyncbuf, sizeof(struct buf *));
1341 pin(&afs_asyncbuf_cv, sizeof(afs_int32));
1342 #else /* AFS_AIX41_ENV */
1345 memset((char *)&afs_bioqueue, 0, sizeof(struct afs_bioqueue));
1352 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1354 * sgi - daemon - handles certain operations that otherwise
1355 * would use up too much kernel stack space
1357 * This all assumes that since the caller must have the xdcache lock
1358 * exclusively that the list will never be more than one long
1359 * and noone else can attempt to add anything until we're done.
1361 SV_TYPE afs_sgibksync;
1362 SV_TYPE afs_sgibkwait;
1363 lock_t afs_sgibklock;
1364 struct dcache *afs_sgibklist;
1372 if (afs_sgibklock == NULL) {
1373 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1374 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1375 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1377 s = SPLOCK(afs_sgibklock);
1379 /* wait for something to do */
1380 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1381 osi_Assert(afs_sgibklist);
1383 /* XX will probably need to generalize to real list someday */
1384 s = SPLOCK(afs_sgibklock);
1385 while (afs_sgibklist) {
1386 tdc = afs_sgibklist;
1387 afs_sgibklist = NULL;
1388 SPUNLOCK(afs_sgibklock, s);
1390 tdc->dflags &= ~DFEntryMod;
1391 afs_WriteDCache(tdc, 1);
1393 s = SPLOCK(afs_sgibklock);
1396 /* done all the work - wake everyone up */
1397 while (SV_SIGNAL(&afs_sgibkwait));