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"
16 #include <sys/sleep.h>
19 #include "afs/sysincludes.h" /* Standard vendor system headers */
20 #include "afsincludes.h" /* Afs-based standard headers */
21 #include "afs/afs_stats.h" /* statistics gathering code */
22 #include "afs/afs_cbqueue.h"
24 #include <sys/adspace.h> /* for vm_att(), vm_det() */
27 #if defined(AFS_CACHE_BYPASS)
28 #include "afs/afs_bypasscache.h"
29 #endif /* AFS_CACHE_BYPASS */
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 #ifndef DEFAULT_PROBE_INTERVAL
50 #define DEFAULT_PROBE_INTERVAL 30 /* default to 3 min */
52 afs_int32 afs_probe_interval = DEFAULT_PROBE_INTERVAL;
53 afs_int32 afs_probe_all_interval = 600;
54 afs_int32 afs_nat_probe_interval = 60;
55 afs_int32 afs_preCache = 0;
57 #define PROBE_WAIT() (1000 * (afs_probe_interval - ((afs_random() & 0x7fffffff) \
58 % (afs_probe_interval/2))))
61 afs_SetCheckServerNATmode(int isnat)
63 static afs_int32 old_intvl, old_all_intvl;
66 if (isnat && !wasnat) {
67 old_intvl = afs_probe_interval;
68 old_all_intvl = afs_probe_all_interval;
69 afs_probe_interval = afs_nat_probe_interval;
70 afs_probe_all_interval = afs_nat_probe_interval;
71 afs_osi_CancelWait(&AFS_CSWaitHandler);
72 } else if (!isnat && wasnat) {
73 afs_probe_interval = old_intvl;
74 afs_probe_all_interval = old_all_intvl;
80 afs_CheckServerDaemon(void)
82 afs_int32 now, delay, lastCheck, last10MinCheck;
84 afs_CheckServerDaemonStarted = 1;
86 while (afs_initState < 101)
87 afs_osi_Sleep(&afs_initState);
88 afs_osi_Wait(PROBE_WAIT(), &AFS_CSWaitHandler, 0);
90 last10MinCheck = lastCheck = osi_Time();
92 if (afs_termState == AFSOP_STOP_CS) {
93 afs_termState = AFSOP_STOP_TRUNCDAEMON;
94 afs_osi_Wakeup(&afs_termState);
99 if (afs_probe_interval + lastCheck <= now) {
100 afs_CheckServers(1, NULL); /* check down servers */
101 lastCheck = now = osi_Time();
104 if (afs_probe_all_interval + last10MinCheck <= now) {
105 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, afs_probe_all_interval);
106 afs_CheckServers(0, NULL);
107 last10MinCheck = now = osi_Time();
109 /* shutdown check. */
110 if (afs_termState == AFSOP_STOP_CS) {
111 afs_termState = AFSOP_STOP_TRUNCDAEMON;
112 afs_osi_Wakeup(&afs_termState);
116 /* Compute time to next probe. */
117 delay = afs_probe_interval + lastCheck;
118 if (delay > afs_probe_all_interval + last10MinCheck)
119 delay = afs_probe_all_interval + last10MinCheck;
123 afs_osi_Wait(delay * 1000, &AFS_CSWaitHandler, 0);
125 afs_CheckServerDaemonStarted = 0;
128 extern int vfs_context_ref;
130 /* This function always holds the GLOCK whilst it is running. The caller
131 * gets the GLOCK before invoking it, and afs_osi_Sleep drops the GLOCK
132 * whilst we are sleeping, and regains it when we're woken up.
138 struct afs_exporter *exporter;
140 afs_int32 last3MinCheck, last10MinCheck, last60MinCheck, lastNMinCheck;
141 afs_int32 last1MinCheck, last5MinCheck;
142 afs_uint32 lastCBSlotBump;
145 AFS_STATCNT(afs_Daemon);
147 afs_rootFid.Fid.Volume = 0;
148 while (afs_initState < 101)
149 afs_osi_Sleep(&afs_initState);
151 #ifdef AFS_DARWIN80_ENV
152 if (afs_osi_ctxtp_initialized)
153 osi_Panic("vfs context already initialized");
154 while (afs_osi_ctxtp && vfs_context_ref)
155 afs_osi_Sleep(&afs_osi_ctxtp);
156 if (afs_osi_ctxtp && !vfs_context_ref)
157 vfs_context_rele(afs_osi_ctxtp);
158 afs_osi_ctxtp = vfs_context_create(NULL);
159 afs_osi_ctxtp_initialized = 1;
162 lastCBSlotBump = now;
164 /* when a lot of clients are booted simultaneously, they develop
165 * annoying synchronous VL server bashing behaviors. So we stagger them.
167 last1MinCheck = now + ((afs_random() & 0x7fffffff) % 60); /* an extra 30 */
168 last3MinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
169 last60MinCheck = now - 1800 + ((afs_random() & 0x7fffffff) % 3600);
170 last10MinCheck = now - 300 + ((afs_random() & 0x7fffffff) % 600);
171 last5MinCheck = now - 150 + ((afs_random() & 0x7fffffff) % 300);
172 lastNMinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
174 /* start off with afs_initState >= 101 (basic init done) */
176 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
178 /* things to do every 20 seconds or less - required by protocol spec */
180 afs_FlushActiveVcaches(0); /* flush NFS writes */
181 afs_FlushVCBs(1); /* flush queued callbacks */
183 afs_MaybeWakeupTruncateDaemon(); /* free cache space if have too */
184 rx_CheckPackets(); /* Does RX need more packets? */
187 if (lastCBSlotBump + CBHTSLOTLEN < now) { /* pretty time-dependant */
188 lastCBSlotBump = now;
189 if (afs_BumpBase()) {
190 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
194 if (last1MinCheck + 60 < now) {
195 /* things to do every minute */
196 DFlush(); /* write out dir buffers */
197 afs_WriteThroughDSlots(); /* write through cacheinfo entries */
198 ObtainWriteLock(&afs_xvcache, 736);
199 afs_FlushReclaimedVcaches();
200 ReleaseWriteLock(&afs_xvcache);
201 afs_FlushActiveVcaches(1); /* keep flocks held & flush nfs writes */
203 afs_StoreDirtyVcaches();
209 if (last3MinCheck + 180 < now) {
210 afs_CheckTokenCache(); /* check for access cache resets due to expired
215 if (afsd_dynamic_vcaches && (last5MinCheck + 300 < now)) {
216 /* start with trying to drop us back to our base usage */
217 int anumber = VCACHE_FREE + (afs_vcount - afs_cacheStats);
220 ObtainWriteLock(&afs_xvcache, 734);
221 afs_ShakeLooseVCaches(anumber);
222 ReleaseWriteLock(&afs_xvcache);
227 if (!afs_CheckServerDaemonStarted) {
228 /* Do the check here if the correct afsd is not installed. */
231 afs_warn("Please install afsd with check server daemon.\n");
233 if (lastNMinCheck + afs_probe_interval < now) {
234 /* only check down servers */
235 afs_CheckServers(1, NULL);
239 if (last10MinCheck + 600 < now) {
240 #ifdef AFS_USERSPACE_IP_ADDR
241 extern int rxi_GetcbiInfo(void);
243 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, 600);
244 #ifdef AFS_USERSPACE_IP_ADDR
245 if (rxi_GetcbiInfo()) { /* addresses changed from last time */
248 #else /* AFS_USERSPACE_IP_ADDR */
249 if (rxi_GetIFInfo()) { /* addresses changed from last time */
252 #endif /* else AFS_USERSPACE_IP_ADDR */
253 if (!afs_CheckServerDaemonStarted)
254 afs_CheckServers(0, NULL);
255 afs_GCUserData(0); /* gc old conns */
256 /* This is probably the wrong way of doing GC for the various exporters but it will suffice for a while */
257 for (exporter = root_exported; exporter;
258 exporter = exporter->exp_next) {
259 (void)EXP_GC(exporter, 0); /* Generalize params */
264 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
268 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
273 last10MinCheck = now;
275 if (last60MinCheck + 3600 < now) {
276 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEVOLUME, ICL_TYPE_INT32,
278 afs_CheckRootVolume();
280 if (afs_gcpags == AFS_GCPAGS_OK) {
285 last60MinCheck = now;
287 if (afs_initState < 300) { /* while things ain't rosy */
288 code = afs_CheckRootVolume();
290 afs_initState = 300; /* succeeded */
291 if (afs_initState < 200)
292 afs_initState = 200; /* tried once */
293 afs_osi_Wakeup(&afs_initState);
296 /* 18285 is because we're trying to divide evenly into 128, that is,
297 * CBSlotLen, while staying just under 20 seconds. If CBSlotLen
298 * changes, should probably change this interval, too.
299 * Some of the preceding actions may take quite some time, so we
300 * might not want to wait the entire interval */
301 now = 18285 - (osi_Time() - now);
303 afs_osi_Wait(now, &AFS_WaitHandler, 0);
306 if (afs_termState == AFSOP_STOP_AFS) {
307 if (afs_CheckServerDaemonStarted)
308 afs_termState = AFSOP_STOP_CS;
310 afs_termState = AFSOP_STOP_TRUNCDAEMON;
311 afs_osi_Wakeup(&afs_termState);
318 afs_CheckRootVolume(void)
320 char rootVolName[32];
321 struct volume *tvp = NULL;
322 int usingDynroot = afs_GetDynrootEnable();
325 AFS_STATCNT(afs_CheckRootVolume);
326 if (*afs_rootVolumeName == 0) {
327 strcpy(rootVolName, "root.afs");
329 strcpy(rootVolName, afs_rootVolumeName);
333 afs_GetDynrootFid(&afs_rootFid);
334 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
336 struct cell *lc = afs_GetPrimaryCell(READ_LOCK);
340 localcell = lc->cellNum;
341 afs_PutCell(lc, READ_LOCK);
342 tvp = afs_GetVolumeByName(rootVolName, localcell, 1, NULL, READ_LOCK);
345 int len = strlen(rootVolName);
347 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
348 strcpy(buf, rootVolName);
349 afs_strcat(buf, ".readonly");
350 tvp = afs_GetVolumeByName(buf, localcell, 1, NULL, READ_LOCK);
354 int volid = (tvp->roVol ? tvp->roVol : tvp->volume);
355 afs_rootFid.Cell = localcell;
356 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
358 /* If we had a root fid before and it changed location we reset
359 * the afs_globalVp so that it will be reevaluated.
360 * Just decrement the reference count. This only occurs during
361 * initial cell setup and can panic the machine if we set the
362 * count to zero and fs checkv is executed when the current
365 #ifdef AFS_LINUX20_ENV
367 struct vrequest treq;
373 afs_rootFid.Fid.Volume = volid;
374 afs_rootFid.Fid.Vnode = 1;
375 afs_rootFid.Fid.Unique = 1;
378 if (afs_InitReq(&treq, credp))
380 vcp = afs_GetVCache(&afs_rootFid, &treq, NULL, NULL);
383 afs_getattr(vcp, &vattr, credp);
384 afs_fill_inode(AFSTOV(vcp), &vattr);
386 dp = d_find_alias(AFSTOV(afs_globalVp));
388 #if defined(AFS_LINUX24_ENV)
389 #if defined(HAVE_DCACHE_LOCK)
390 spin_lock(&dcache_lock);
392 spin_lock(&AFSTOV(vcp)->i_lock);
394 #if defined(AFS_LINUX26_ENV)
395 spin_lock(&dp->d_lock);
398 #if defined(D_ALIAS_IS_HLIST)
399 hlist_del_init(&dp->d_alias);
400 hlist_add_head(&dp->d_alias, &(AFSTOV(vcp)->i_dentry));
402 list_del_init(&dp->d_alias);
403 list_add(&dp->d_alias, &(AFSTOV(vcp)->i_dentry));
405 dp->d_inode = AFSTOV(vcp);
406 #if defined(AFS_LINUX24_ENV)
407 #if defined(AFS_LINUX26_ENV)
408 spin_unlock(&dp->d_lock);
410 #if defined(HAVE_DCACHE_LOCK)
411 spin_unlock(&dcache_lock);
413 spin_unlock(&AFSTOV(vcp)->i_lock);
418 AFS_FAST_RELE(afs_globalVp);
424 #ifdef AFS_DARWIN80_ENV
425 afs_PutVCache(afs_globalVp);
427 AFS_FAST_RELE(afs_globalVp);
432 afs_rootFid.Fid.Volume = volid;
433 afs_rootFid.Fid.Vnode = 1;
434 afs_rootFid.Fid.Unique = 1;
438 afs_initState = 300; /* won */
439 afs_osi_Wakeup(&afs_initState);
440 afs_PutVolume(tvp, READ_LOCK);
442 if (afs_rootFid.Fid.Volume)
448 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
450 BPath(struct brequest *ab)
452 struct dcache *tdc = NULL;
453 struct vcache *tvc = NULL;
454 struct vnode *tvn = NULL;
455 #ifdef AFS_LINUX22_ENV
456 struct dentry *dp = NULL;
458 afs_size_t offset, len;
459 struct vrequest treq;
463 if ((code = afs_InitReq(&treq, ab->cred)))
466 #ifdef AFS_LINUX22_ENV
467 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &dp);
469 tvn = (struct vnode *)dp->d_inode;
471 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &tvn);
474 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
477 /* now path may not have been in afs, so check that before calling our cache manager */
478 if (!tvn || !IsAfsVnode(tvn)) {
479 /* release it and give up */
481 #ifdef AFS_LINUX22_ENV
490 /* here we know its an afs vnode, so we can get the data for the chunk */
491 tdc = afs_GetDCache(tvc, ab->size_parm[0], &treq, &offset, &len, 1);
495 #ifdef AFS_LINUX22_ENV
502 /* size_parm 0 to the fetch is the chunk number,
503 * ptr_parm 0 is the dcache entry to wakeup,
504 * size_parm 1 is true iff we should release the dcache entry here.
507 BPrefetch(struct brequest *ab)
511 afs_size_t offset, len, abyte, totallen = 0;
512 struct vrequest treq;
514 AFS_STATCNT(BPrefetch);
515 if ((len = afs_InitReq(&treq, ab->cred)))
517 abyte = ab->size_parm[0];
520 tdc = afs_GetDCache(tvc, abyte, &treq, &offset, &len, 1);
526 } while ((totallen < afs_preCache) && tdc && (len > 0));
527 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
528 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
529 * be waiting for our wakeup anyway.
531 tdc = (struct dcache *)(ab->ptr_parm[0]);
532 ObtainSharedLock(&tdc->lock, 640);
533 if (tdc->mflags & DFFetchReq) {
534 UpgradeSToWLock(&tdc->lock, 641);
535 tdc->mflags &= ~DFFetchReq;
536 ReleaseWriteLock(&tdc->lock);
538 ReleaseSharedLock(&tdc->lock);
540 afs_osi_Wakeup(&tdc->validPos);
541 if (ab->size_parm[1]) {
542 afs_PutDCache(tdc); /* put this one back, too */
546 #if defined(AFS_CACHE_BYPASS)
548 BPrefetchNoCache(struct brequest *ab)
550 struct vrequest treq;
553 if ((len = afs_InitReq(&treq, ab->cred)))
557 /* OS-specific prefetch routine */
558 afs_PrefetchNoCache(ab->vc, ab->cred, (struct nocache_read_request *) ab->ptr_parm[0]);
564 BStore(struct brequest *ab)
568 struct vrequest treq;
569 #if defined(AFS_SGI_ENV)
570 struct cred *tmpcred;
574 if ((code = afs_InitReq(&treq, ab->cred)))
577 #if defined(AFS_SGI_ENV)
579 * Since StoreOnLastReference can end up calling osi_SyncVM which
580 * calls into VM code that assumes that u.u_cred has the
581 * correct credentials, we set our to theirs for this xaction
583 tmpcred = OSI_GET_CURRENT_CRED();
584 OSI_SET_CURRENT_CRED(ab->cred);
587 * To avoid recursion since the WriteLock may be released during VM
588 * operations, we hold the VOP_RWLOCK across this transaction as
589 * do the other callers of StoreOnLastReference
591 AFS_RWLOCK((vnode_t *) tvc, 1);
593 ObtainWriteLock(&tvc->lock, 209);
594 code = afs_StoreOnLastReference(tvc, &treq);
595 ReleaseWriteLock(&tvc->lock);
596 #if defined(AFS_SGI_ENV)
597 OSI_SET_CURRENT_CRED(tmpcred);
598 AFS_RWUNLOCK((vnode_t *) tvc, 1);
600 /* now set final return code, and wakeup anyone waiting */
601 if ((ab->flags & BUVALID) == 0) {
603 /* To explain code_raw/code_checkcode:
604 * Anyone that's waiting won't have our treq, so they won't be able to
605 * call afs_CheckCode themselves on the return code we provide here.
606 * But if we give back only the afs_CheckCode value, they won't know
607 * what the "raw" value was. So give back both values, so the waiter
608 * can know the "raw" value for interpreting the value internally, as
609 * well as the afs_CheckCode value to give to the OS. */
611 ab->code_checkcode = afs_CheckCode(code, &treq, 43);
613 ab->flags |= BUVALID;
614 if (ab->flags & BUWAIT) {
615 ab->flags &= ~BUWAIT;
622 BPartialStore(struct brequest *ab)
626 struct vrequest treq;
627 int locked, shared_locked = 0;
630 if ((code = afs_InitReq(&treq, ab->cred)))
633 locked = tvc->lock.excl_locked? 1:0;
635 ObtainWriteLock(&tvc->lock, 1209);
636 else if (!(tvc->lock.excl_locked & WRITE_LOCK)) {
638 ConvertSToRLock(&tvc->lock);
640 code = afs_StoreAllSegments(tvc, &treq, AFS_ASYNC);
642 ReleaseWriteLock(&tvc->lock);
643 else if (shared_locked)
644 ConvertSToRLock(&tvc->lock);
645 /* now set final return code, and wakeup anyone waiting */
646 if ((ab->flags & BUVALID) == 0) {
647 /* set final code, since treq doesn't go across processes */
649 ab->code_checkcode = afs_CheckCode(code, &treq, 43);
650 ab->flags |= BUVALID;
651 if (ab->flags & BUWAIT) {
652 ab->flags &= ~BUWAIT;
658 /* release a held request buffer */
660 afs_BRelease(struct brequest *ab)
663 AFS_STATCNT(afs_BRelease);
664 ObtainWriteLock(&afs_xbrs, 294);
665 if (--ab->refCount <= 0) {
669 afs_osi_Wakeup(&afs_brsWaiters);
670 ReleaseWriteLock(&afs_xbrs);
673 /* return true if bkg fetch daemons are all busy */
677 AFS_STATCNT(afs_BBusy);
678 if (afs_brsDaemons > 0)
684 afs_BQueue(short aopcode, struct vcache *avc,
685 afs_int32 dontwait, afs_int32 ause, afs_ucred_t *acred,
686 afs_size_t asparm0, afs_size_t asparm1, void *apparm0,
687 void *apparm1, void *apparm2)
692 AFS_STATCNT(afs_BQueue);
693 ObtainWriteLock(&afs_xbrs, 296);
696 for (i = 0; i < NBRS; i++, tb++) {
697 if (tb->refCount == 0)
702 tb->opcode = aopcode;
709 tb->refCount = ause + 1;
710 tb->size_parm[0] = asparm0;
711 tb->size_parm[1] = asparm1;
712 tb->ptr_parm[0] = apparm0;
713 tb->ptr_parm[1] = apparm1;
714 tb->ptr_parm[2] = apparm2;
716 tb->code_raw = tb->code_checkcode = 0;
717 tb->ts = afs_brs_count++;
718 /* if daemons are waiting for work, wake them up */
719 if (afs_brsDaemons > 0) {
720 afs_osi_Wakeup(&afs_brsDaemons);
722 ReleaseWriteLock(&afs_xbrs);
726 ReleaseWriteLock(&afs_xbrs);
729 /* no free buffers, sleep a while */
731 ReleaseWriteLock(&afs_xbrs);
732 afs_osi_Sleep(&afs_brsWaiters);
733 ObtainWriteLock(&afs_xbrs, 301);
739 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
740 * The modifications here will work for either a UP or MP machine.
742 struct buf *afs_asyncbuf = (struct buf *)0;
743 tid_t afs_asyncbuf_cv = EVENT_NULL;
744 afs_int32 afs_biodcnt = 0;
746 /* in implementing this, I assumed that all external linked lists were
749 * Several places in this code traverse a linked list. The algorithm
750 * used here is probably unfamiliar to most people. Careful examination
751 * will show that it eliminates an assignment inside the loop, as compared
752 * to the standard algorithm, at the cost of occasionally using an extra
758 * This function obtains, and returns, a pointer to a buffer for
759 * processing by a daemon. It sleeps until such a buffer is available.
760 * The source of buffers for it is the list afs_asyncbuf (see also
761 * afs_gn_strategy). This function may be invoked concurrently by
762 * several processes, that is, several instances of the same daemon.
763 * afs_gn_strategy, which adds buffers to the list, runs at interrupt
764 * level, while get_bioreq runs at process level.
766 * Since AIX 4.1 can wake just one process at a time, the separate sleep
767 * addresses have been removed.
768 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
769 * The afs_asyncbuf_lock is primarily used to serialize access between
770 * process and interrupts.
772 Simple_lock afs_asyncbuf_lock;
776 struct buf *bp = NULL;
778 struct buf **bestlbpP, **lbpP;
780 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
783 struct afs_bioqueue *s;
785 /* ??? Does the forward pointer of the returned buffer need to be NULL?
788 /* Disable interrupts from the strategy function, and save the
789 * prior priority level and lock access to the afs_asyncbuf.
792 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock);
796 /* look for oldest buffer */
797 bp = bestbp = afs_asyncbuf;
798 bestage = (long)bestbp->av_back;
799 bestlbpP = &afs_asyncbuf;
805 if ((long)bp->av_back - bestage < 0) {
808 bestage = (long)bp->av_back;
812 *bestlbpP = bp->av_forw;
815 /* If afs_asyncbuf is null, it is necessary to go to sleep.
816 * e_wakeup_one() ensures that only one thread wakes.
819 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
820 * lock on an MP machine.
823 e_sleep_thread(&afs_asyncbuf_cv, &afs_asyncbuf_lock,
824 LOCK_HANDLER | INTERRUPTIBLE);
825 if (interrupted == THREAD_INTERRUPTED) {
826 /* re-enable interrupts from strategy */
827 unlock_enable(oldPriority, &afs_asyncbuf_lock);
831 } /* end of "else asyncbuf is empty" */
832 } /* end of "inner loop" */
836 unlock_enable(oldPriority, &afs_asyncbuf_lock);
839 /* For the convenience of other code, replace the gnodes in
840 * the b_vp field of bp and the other buffers on the b_work
841 * chain with the corresponding vnodes.
843 * ??? what happens to the gnodes? They're not just cut loose,
847 t2P = (struct buf *)t1P->b_work;
848 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
852 t1P = (struct buf *)t2P->b_work;
853 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
858 /* If the buffer does not specify I/O, it may immediately
859 * be returned to the caller. This condition is detected
860 * by examining the buffer's flags (the b_flags field). If
861 * the B_PFPROT bit is set, the buffer represents a protection
862 * violation, rather than a request for I/O. The remainder
863 * of the outer loop handles the case where the B_PFPROT bit is clear.
865 if (bp->b_flags & B_PFPROT) {
870 } /* end of function get_bioreq() */
875 * This function is the daemon. It is called from the syscall
876 * interface. Ordinarily, a script or an administrator will run a
877 * daemon startup utility, specifying the number of I/O daemons to
878 * run. The utility will fork off that number of processes,
879 * each making the appropriate syscall, which will cause this
880 * function to be invoked.
882 static int afs_initbiod = 0; /* this is self-initializing code */
885 afs_BioDaemon(afs_int32 nbiods)
887 afs_int32 code, s, pflg = 0;
889 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
897 /* pin lock, since we'll be using it in an interrupt. */
898 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
899 simple_lock_init(&afs_asyncbuf_lock);
900 pin(&afs_asyncbuf, sizeof(struct buf *));
901 pin(&afs_asyncbuf_cv, sizeof(afs_int32));
904 /* Ignore HUP signals... */
906 sigset_t sigbits, osigbits;
908 * add SIGHUP to the set of already masked signals
910 SIGFILLSET(sigbits); /* allow all signals */
911 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
912 limit_sigs(&sigbits, &osigbits); /* and already masked */
914 /* Main body starts here -- this is an intentional infinite loop, and
917 * Now, the loop will exit if get_bioreq() returns NULL, indicating
918 * that we've been interrupted.
921 bp = afs_get_bioreq();
923 break; /* we were interrupted */
924 if (code = setjmpx(&jmpbuf)) {
925 /* This should not have happend, maybe a lack of resources */
927 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
928 for (bp1 = bp; bp; bp = bp1) {
930 bp1 = (struct buf *)bp1->b_work;
933 bp->b_flags |= B_ERROR;
936 unlock_enable(s, &afs_asyncbuf_lock);
940 vcp = VTOAFS(bp->b_vp);
941 if (bp->b_flags & B_PFSTORE) { /* XXXX */
942 ObtainWriteLock(&vcp->lock, 404);
943 if (vcp->v.v_gnode->gn_mwrcnt) {
944 afs_offs_t newlength =
945 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
946 if (vcp->f.m.Length < newlength) {
947 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
948 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
949 __LINE__, ICL_TYPE_OFFSET,
950 ICL_HANDLE_OFFSET(vcp->f.m.Length),
951 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
952 vcp->f.m.Length = newlength;
955 ReleaseWriteLock(&vcp->lock);
957 /* If the buffer represents a protection violation, rather than
958 * an actual request for I/O, no special action need be taken.
960 if (bp->b_flags & B_PFPROT) {
961 iodone(bp); /* Notify all users of the buffer that we're done */
966 ObtainWriteLock(&vcp->pvmlock, 211);
968 * First map its data area to a region in the current address space
969 * by calling vm_att with the subspace identifier, and a pointer to
970 * the data area. vm_att returns a new data area pointer, but we
971 * also want to hang onto the old one.
973 tmpaddr = bp->b_baddr;
974 bp->b_baddr = (caddr_t) vm_att(bp->b_xmemd.subspace_id, tmpaddr);
975 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
976 if (tmperr) { /* in non-error case */
977 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
978 bp->b_error = tmperr;
981 /* Unmap the buffer's data area by calling vm_det. Reset data area
982 * to the value that we saved above.
985 bp->b_baddr = tmpaddr;
988 * buffer may be linked with other buffers via the b_work field.
989 * See also afs_gn_strategy. For each buffer in the chain (including
990 * bp) notify all users of the buffer that the daemon is finished
991 * using it by calling iodone.
992 * assumes iodone can modify the b_work field.
995 tbp2 = (struct buf *)tbp1->b_work;
1000 tbp1 = (struct buf *)tbp2->b_work;
1006 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
1008 } /* infinite loop (unless we're interrupted) */
1009 } /* end of afs_BioDaemon() */
1011 #endif /* AFS_AIX41_ENV */
1016 afs_BackgroundDaemon_once(void)
1018 LOCK_INIT(&afs_xbrs, "afs_xbrs");
1019 memset(afs_brs, 0, sizeof(afs_brs));
1021 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1023 * steal the first daemon for doing delayed DSlot flushing
1024 * (see afs_GetDownDSlot)
1028 exit(CLD_EXITED, 0);
1033 brequest_release(struct brequest *tb)
1036 AFS_RELE(AFSTOV(tb->vc)); /* MUST call vnode layer or could lose vnodes */
1041 tb->cred = (afs_ucred_t *)0;
1043 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1048 afs_BackgroundDaemon(struct afs_uspc_param *uspc, void *param1, void *param2)
1051 afs_BackgroundDaemon(void)
1054 struct brequest *tb;
1057 AFS_STATCNT(afs_BackgroundDaemon);
1058 /* initialize subsystem */
1060 /* Irix with "short stack" exits */
1061 afs_BackgroundDaemon_once();
1064 /* If it's a re-entering syscall, complete the request and release */
1065 if (uspc->ts > -1) {
1067 for (i = 0; i < NBRS; i++, tb++) {
1068 if (tb->ts == uspc->ts) {
1069 /* copy the userspace status back in */
1070 ((struct afs_uspc_param *) tb->ptr_parm[0])->retval =
1072 /* mark it valid and notify our caller */
1073 tb->flags |= BUVALID;
1074 if (tb->flags & BUWAIT) {
1075 tb->flags &= ~BUWAIT;
1078 brequest_release(tb);
1083 afs_osi_MaskUserLoop();
1085 /* Otherwise it's a new one */
1091 ObtainWriteLock(&afs_xbrs, 302);
1094 struct brequest *min_tb = NULL;
1096 if (afs_termState == AFSOP_STOP_BKG) {
1097 if (--afs_nbrs <= 0)
1098 afs_termState = AFSOP_STOP_RXCALLBACK;
1099 ReleaseWriteLock(&afs_xbrs);
1100 afs_osi_Wakeup(&afs_termState);
1108 /* find a request */
1111 for (i = 0; i < NBRS; i++, tb++) {
1112 /* look for request with smallest ts */
1113 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1114 /* new request, not yet picked up */
1115 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1121 if ((tb = min_tb)) {
1122 /* claim and process this request */
1123 tb->flags |= BSTARTED;
1124 ReleaseWriteLock(&afs_xbrs);
1126 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1, ICL_TYPE_INT32,
1128 if (tb->opcode == BOP_FETCH)
1130 #if defined(AFS_CACHE_BYPASS)
1131 else if (tb->opcode == BOP_FETCH_NOCACHE)
1132 BPrefetchNoCache(tb);
1134 else if (tb->opcode == BOP_STORE)
1136 else if (tb->opcode == BOP_PATH)
1138 #ifdef AFS_DARWIN80_ENV
1139 else if (tb->opcode == BOP_MOVE) {
1140 memcpy(uspc, (struct afs_uspc_param *) tb->ptr_parm[0],
1141 sizeof(struct afs_uspc_param));
1143 /* string lengths capped in move vop; copy NUL tho */
1144 memcpy(param1, (char *)tb->ptr_parm[1],
1145 strlen(tb->ptr_parm[1])+1);
1146 memcpy(param2, (char *)tb->ptr_parm[2],
1147 strlen(tb->ptr_parm[2])+1);
1151 else if (tb->opcode == BOP_PARTIAL_STORE)
1154 panic("background bop");
1155 brequest_release(tb);
1156 ObtainWriteLock(&afs_xbrs, 305);
1159 /* wait for new request */
1161 ReleaseWriteLock(&afs_xbrs);
1162 afs_osi_Sleep(&afs_brsDaemons);
1163 ObtainWriteLock(&afs_xbrs, 307);
1174 shutdown_daemons(void)
1176 AFS_STATCNT(shutdown_daemons);
1177 if (afs_cold_shutdown) {
1178 afs_brsDaemons = brsInit = 0;
1179 rxepoch_checked = afs_nbrs = 0;
1180 memset(afs_brs, 0, sizeof(afs_brs));
1181 memset(&afs_xbrs, 0, sizeof(afs_lock_t));
1183 #ifdef AFS_AIX41_ENV
1184 lock_free(&afs_asyncbuf_lock);
1185 unpin(&afs_asyncbuf, sizeof(struct buf *));
1186 unpin(&afs_asyncbuf_cv, sizeof(afs_int32));
1192 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1194 * sgi - daemon - handles certain operations that otherwise
1195 * would use up too much kernel stack space
1197 * This all assumes that since the caller must have the xdcache lock
1198 * exclusively that the list will never be more than one long
1199 * and noone else can attempt to add anything until we're done.
1201 SV_TYPE afs_sgibksync;
1202 SV_TYPE afs_sgibkwait;
1203 lock_t afs_sgibklock;
1204 struct dcache *afs_sgibklist;
1212 if (afs_sgibklock == NULL) {
1213 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1214 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1215 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1217 s = SPLOCK(afs_sgibklock);
1219 /* wait for something to do */
1220 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1221 osi_Assert(afs_sgibklist);
1223 /* XX will probably need to generalize to real list someday */
1224 s = SPLOCK(afs_sgibklock);
1225 while (afs_sgibklist) {
1226 tdc = afs_sgibklist;
1227 afs_sgibklist = NULL;
1228 SPUNLOCK(afs_sgibklock, s);
1230 tdc->dflags &= ~DFEntryMod;
1231 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
1233 s = SPLOCK(afs_sgibklock);
1236 /* done all the work - wake everyone up */
1237 while (SV_SIGNAL(&afs_sgibkwait));