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 /* PAG garbage collection */
40 /* We induce a compile error if param.h does not define AFS_GCPAGS */
41 afs_int32 afs_gcpags = AFS_GCPAGS;
42 afs_int32 afs_gcpags_procsize = 0;
44 afs_int32 afs_CheckServerDaemonStarted = 0;
45 #ifndef DEFAULT_PROBE_INTERVAL
46 #define DEFAULT_PROBE_INTERVAL 30 /* default to 3 min */
48 afs_int32 afs_probe_interval = DEFAULT_PROBE_INTERVAL;
49 afs_int32 afs_probe_all_interval = 600;
50 afs_int32 afs_nat_probe_interval = 60;
51 afs_int32 afs_preCache = 0;
53 #define PROBE_WAIT() (1000 * (afs_probe_interval - ((afs_random() & 0x7fffffff) \
54 % (afs_probe_interval/2))))
57 afs_SetCheckServerNATmode(int isnat)
59 static afs_int32 old_intvl, old_all_intvl;
62 if (isnat && !wasnat) {
63 old_intvl = afs_probe_interval;
64 old_all_intvl = afs_probe_all_interval;
65 afs_probe_interval = afs_nat_probe_interval;
66 afs_probe_all_interval = afs_nat_probe_interval;
67 afs_osi_CancelWait(&AFS_CSWaitHandler);
68 } else if (!isnat && wasnat) {
69 afs_probe_interval = old_intvl;
70 afs_probe_all_interval = old_all_intvl;
76 afs_CheckServerDaemon(void)
78 afs_int32 now, delay, lastCheck, last10MinCheck;
80 afs_CheckServerDaemonStarted = 1;
82 while (afs_initState < 101)
83 afs_osi_Sleep(&afs_initState);
84 afs_osi_Wait(PROBE_WAIT(), &AFS_CSWaitHandler, 0);
86 last10MinCheck = lastCheck = osi_Time();
88 if (afs_termState == AFSOP_STOP_CS) {
89 afs_termState = AFSOP_STOP_TRUNCDAEMON;
90 afs_osi_Wakeup(&afs_termState);
95 if (afs_probe_interval + lastCheck <= now) {
96 afs_CheckServers(1, NULL); /* check down servers */
97 lastCheck = now = osi_Time();
100 if (afs_probe_all_interval + last10MinCheck <= now) {
101 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, afs_probe_all_interval);
102 afs_CheckServers(0, NULL);
103 last10MinCheck = now = osi_Time();
105 /* shutdown check. */
106 if (afs_termState == AFSOP_STOP_CS) {
107 afs_termState = AFSOP_STOP_TRUNCDAEMON;
108 afs_osi_Wakeup(&afs_termState);
112 /* Compute time to next probe. */
113 delay = afs_probe_interval + lastCheck;
114 if (delay > afs_probe_all_interval + last10MinCheck)
115 delay = afs_probe_all_interval + last10MinCheck;
119 afs_osi_Wait(delay * 1000, &AFS_CSWaitHandler, 0);
121 afs_CheckServerDaemonStarted = 0;
124 extern int vfs_context_ref;
126 /* This function always holds the GLOCK whilst it is running. The caller
127 * gets the GLOCK before invoking it, and afs_osi_Sleep drops the GLOCK
128 * whilst we are sleeping, and regains it when we're woken up.
134 struct afs_exporter *exporter;
136 afs_int32 last3MinCheck, last10MinCheck, last60MinCheck, lastNMinCheck;
137 afs_int32 last1MinCheck, last5MinCheck;
138 afs_uint32 lastCBSlotBump;
140 AFS_STATCNT(afs_Daemon);
142 afs_rootFid.Fid.Volume = 0;
143 while (afs_initState < 101)
144 afs_osi_Sleep(&afs_initState);
146 #ifdef AFS_DARWIN80_ENV
147 if (afs_osi_ctxtp_initialized)
148 osi_Panic("vfs context already initialized");
149 while (afs_osi_ctxtp && vfs_context_ref)
150 afs_osi_Sleep(&afs_osi_ctxtp);
151 if (afs_osi_ctxtp && !vfs_context_ref)
152 vfs_context_rele(afs_osi_ctxtp);
153 afs_osi_ctxtp = vfs_context_create(NULL);
154 afs_osi_ctxtp_initialized = 1;
157 lastCBSlotBump = now;
159 /* when a lot of clients are booted simultaneously, they develop
160 * annoying synchronous VL server bashing behaviors. So we stagger them.
162 last1MinCheck = now + ((afs_random() & 0x7fffffff) % 60); /* an extra 30 */
163 last3MinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
164 last60MinCheck = now - 1800 + ((afs_random() & 0x7fffffff) % 3600);
165 last10MinCheck = now - 300 + ((afs_random() & 0x7fffffff) % 600);
166 last5MinCheck = now - 150 + ((afs_random() & 0x7fffffff) % 300);
167 lastNMinCheck = now - 90 + ((afs_random() & 0x7fffffff) % 180);
169 /* start off with afs_initState >= 101 (basic init done) */
171 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
173 /* things to do every 20 seconds or less - required by protocol spec */
175 afs_FlushActiveVcaches(0); /* flush NFS writes */
176 afs_FlushVCBs(1); /* flush queued callbacks */
178 afs_MaybeWakeupTruncateDaemon(); /* free cache space if have too */
179 rx_CheckPackets(); /* Does RX need more packets? */
182 if (lastCBSlotBump + CBHTSLOTLEN < now) { /* pretty time-dependant */
183 lastCBSlotBump = now;
184 if (afs_BumpBase()) {
185 afs_CheckCallbacks(20); /* unstat anything which will expire soon */
189 if (last1MinCheck + 60 < now) {
190 /* things to do every minute */
191 DFlush(); /* write out dir buffers */
192 (void)afs_WriteThroughDSlots(); /* write through cacheinfo entries */
193 ObtainWriteLock(&afs_xvcache, 736);
194 afs_FlushReclaimedVcaches();
195 ReleaseWriteLock(&afs_xvcache);
196 afs_FlushActiveVcaches(1); /* keep flocks held & flush nfs writes */
200 if (last3MinCheck + 180 < now) {
201 afs_CheckTokenCache(); /* check for access cache resets due to expired
206 if (afsd_dynamic_vcaches && (last5MinCheck + 300 < now)) {
207 /* start with trying to drop us back to our base usage */
208 int anumber = VCACHE_FREE + (afs_vcount - afs_cacheStats);
211 ObtainWriteLock(&afs_xvcache, 734);
212 afs_ShakeLooseVCaches(anumber);
213 ReleaseWriteLock(&afs_xvcache);
218 if (!afs_CheckServerDaemonStarted) {
219 if (lastNMinCheck + afs_probe_interval < now) {
220 /* only check down servers */
221 afs_CheckServers(1, NULL);
225 if (last10MinCheck + 600 < now) {
226 #ifdef AFS_USERSPACE_IP_ADDR
227 extern int rxi_GetcbiInfo(void);
229 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEUP, ICL_TYPE_INT32, 600);
230 #ifdef AFS_USERSPACE_IP_ADDR
231 if (rxi_GetcbiInfo()) { /* addresses changed from last time */
234 #else /* AFS_USERSPACE_IP_ADDR */
235 if (rxi_GetIFInfo()) { /* addresses changed from last time */
238 #endif /* else AFS_USERSPACE_IP_ADDR */
239 if (!afs_CheckServerDaemonStarted)
240 afs_CheckServers(0, NULL);
241 afs_GCUserData(); /* gc old conns */
242 /* This is probably the wrong way of doing GC for the various exporters but it will suffice for a while */
243 for (exporter = root_exported; exporter;
244 exporter = exporter->exp_next) {
245 (void)EXP_GC(exporter, 0); /* Generalize params */
250 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
254 afs_CheckVolumeNames(AFS_VOLCHECK_EXPIRED |
259 last10MinCheck = now;
261 if (last60MinCheck + 3600 < now) {
262 afs_Trace1(afs_iclSetp, CM_TRACE_PROBEVOLUME, ICL_TYPE_INT32,
264 afs_CheckRootVolume();
266 if (afs_gcpags == AFS_GCPAGS_OK) {
271 last60MinCheck = now;
273 if (afs_initState < 300) { /* while things ain't rosy */
274 code = afs_CheckRootVolume();
276 afs_initState = 300; /* succeeded */
277 if (afs_initState < 200)
278 afs_initState = 200; /* tried once */
279 afs_osi_Wakeup(&afs_initState);
282 /* 18285 is because we're trying to divide evenly into 128, that is,
283 * CBSlotLen, while staying just under 20 seconds. If CBSlotLen
284 * changes, should probably change this interval, too.
285 * Some of the preceding actions may take quite some time, so we
286 * might not want to wait the entire interval */
287 now = 18285 - (osi_Time() - now);
289 afs_osi_Wait(now, &AFS_WaitHandler, 0);
292 if (afs_termState == AFSOP_STOP_AFS) {
293 if (afs_CheckServerDaemonStarted)
294 afs_termState = AFSOP_STOP_CS;
296 afs_termState = AFSOP_STOP_TRUNCDAEMON;
297 afs_osi_Wakeup(&afs_termState);
304 afs_CheckRootVolume(void)
306 char rootVolName[MAXROOTVOLNAMELEN];
307 struct volume *tvp = NULL;
308 int usingDynroot = afs_GetDynrootEnable();
311 AFS_STATCNT(afs_CheckRootVolume);
312 if (*afs_rootVolumeName == 0) {
313 strcpy(rootVolName, "root.afs");
315 strcpy(rootVolName, afs_rootVolumeName);
319 afs_GetDynrootFid(&afs_rootFid);
320 tvp = afs_GetVolume(&afs_rootFid, NULL, READ_LOCK);
322 struct cell *lc = afs_GetPrimaryCell(READ_LOCK);
326 localcell = lc->cellNum;
327 afs_PutCell(lc, READ_LOCK);
328 tvp = afs_GetVolumeByName(rootVolName, localcell, 1, NULL, READ_LOCK);
331 int len = strlen(rootVolName);
333 if ((len < 9) || strcmp(&rootVolName[len - 9], ".readonly")) {
334 strcpy(buf, rootVolName);
335 afs_strcat(buf, ".readonly");
336 tvp = afs_GetVolumeByName(buf, localcell, 1, NULL, READ_LOCK);
340 int volid = (tvp->roVol ? tvp->roVol : tvp->volume);
341 afs_rootFid.Cell = localcell;
342 if (afs_rootFid.Fid.Volume && afs_rootFid.Fid.Volume != volid
344 /* If we had a root fid before and it changed location we reset
345 * the afs_globalVp so that it will be reevaluated.
346 * Just decrement the reference count. This only occurs during
347 * initial cell setup and can panic the machine if we set the
348 * count to zero and fs checkv is executed when the current
351 #ifdef AFS_LINUX22_ENV
352 osi_ResetRootVCache(volid);
354 # ifdef AFS_DARWIN80_ENV
355 afs_PutVCache(afs_globalVp);
357 AFS_FAST_RELE(afs_globalVp);
362 afs_rootFid.Fid.Volume = volid;
363 afs_rootFid.Fid.Vnode = 1;
364 afs_rootFid.Fid.Unique = 1;
368 afs_initState = 300; /* won */
369 afs_osi_Wakeup(&afs_initState);
370 afs_PutVolume(tvp, READ_LOCK);
372 if (afs_rootFid.Fid.Volume)
378 /* ptr_parm 0 is the pathname, size_parm 0 to the fetch is the chunk number */
380 BPath(struct brequest *ab)
382 struct dcache *tdc = NULL;
383 struct vcache *tvc = NULL;
384 struct vnode *tvn = NULL;
385 #ifdef AFS_LINUX22_ENV
386 struct dentry *dp = NULL;
388 afs_size_t offset, len;
389 struct vrequest *treq = NULL;
393 if ((code = afs_CreateReq(&treq, ab->cred))) {
397 #ifdef AFS_LINUX22_ENV
398 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &dp);
400 tvn = (struct vnode *)dp->d_inode;
402 code = gop_lookupname((char *)ab->ptr_parm[0], AFS_UIOSYS, 1, &tvn);
405 osi_FreeLargeSpace((char *)ab->ptr_parm[0]); /* free path name buffer here */
407 afs_DestroyReq(treq);
410 /* now path may not have been in afs, so check that before calling our cache manager */
411 if (!tvn || !IsAfsVnode(tvn)) {
412 /* release it and give up */
414 #ifdef AFS_LINUX22_ENV
420 afs_DestroyReq(treq);
424 /* here we know its an afs vnode, so we can get the data for the chunk */
425 tdc = afs_GetDCache(tvc, ab->size_parm[0], treq, &offset, &len, 1);
429 #ifdef AFS_LINUX22_ENV
434 afs_DestroyReq(treq);
437 /* size_parm 0 to the fetch is the chunk number,
438 * ptr_parm 0 is the dcache entry to wakeup,
439 * size_parm 1 is true iff we should release the dcache entry here.
442 BPrefetch(struct brequest *ab)
446 afs_size_t offset, len, abyte, totallen = 0;
447 struct vrequest *treq = NULL;
450 AFS_STATCNT(BPrefetch);
451 if ((code = afs_CreateReq(&treq, ab->cred)))
453 abyte = ab->size_parm[0];
456 tdc = afs_GetDCache(tvc, abyte, treq, &offset, &len, 1);
462 } while ((totallen < afs_preCache) && tdc && (len > 0));
463 /* now, dude may be waiting for us to clear DFFetchReq bit; do so. Can't
464 * use tdc from GetDCache since afs_GetDCache may fail, but someone may
465 * be waiting for our wakeup anyway.
467 tdc = (struct dcache *)(ab->ptr_parm[0]);
468 ObtainSharedLock(&tdc->lock, 640);
469 if (tdc->mflags & DFFetchReq) {
470 UpgradeSToWLock(&tdc->lock, 641);
471 tdc->mflags &= ~DFFetchReq;
472 ReleaseWriteLock(&tdc->lock);
474 ReleaseSharedLock(&tdc->lock);
476 afs_osi_Wakeup(&tdc->validPos);
477 if (ab->size_parm[1]) {
478 afs_PutDCache(tdc); /* put this one back, too */
480 afs_DestroyReq(treq);
483 #if defined(AFS_CACHE_BYPASS)
485 BPrefetchNoCache(struct brequest *ab)
487 struct vrequest *treq = NULL;
490 if ((code = afs_CreateReq(&treq, ab->cred)))
494 /* OS-specific prefetch routine */
495 afs_PrefetchNoCache(ab->vc, ab->cred, (struct nocache_read_request *) ab->ptr_parm[0]);
497 afs_DestroyReq(treq);
502 BStore(struct brequest *ab)
506 struct vrequest *treq = NULL;
507 #if defined(AFS_SGI_ENV)
508 struct cred *tmpcred;
512 if ((code = afs_CreateReq(&treq, ab->cred)))
515 #if defined(AFS_SGI_ENV)
517 * Since StoreOnLastReference can end up calling osi_SyncVM which
518 * calls into VM code that assumes that u.u_cred has the
519 * correct credentials, we set our to theirs for this xaction
521 tmpcred = OSI_GET_CURRENT_CRED();
522 OSI_SET_CURRENT_CRED(ab->cred);
525 * To avoid recursion since the WriteLock may be released during VM
526 * operations, we hold the VOP_RWLOCK across this transaction as
527 * do the other callers of StoreOnLastReference
529 AFS_RWLOCK((vnode_t *) tvc, 1);
531 ObtainWriteLock(&tvc->lock, 209);
532 code = afs_StoreOnLastReference(tvc, treq);
533 ReleaseWriteLock(&tvc->lock);
534 #if defined(AFS_SGI_ENV)
535 OSI_SET_CURRENT_CRED(tmpcred);
536 AFS_RWUNLOCK((vnode_t *) tvc, 1);
538 /* now set final return code, and wakeup anyone waiting */
539 if ((ab->flags & BUVALID) == 0) {
541 /* To explain code_raw/code_checkcode:
542 * Anyone that's waiting won't have our treq, so they won't be able to
543 * call afs_CheckCode themselves on the return code we provide here.
544 * But if we give back only the afs_CheckCode value, they won't know
545 * what the "raw" value was. So give back both values, so the waiter
546 * can know the "raw" value for interpreting the value internally, as
547 * well as the afs_CheckCode value to give to the OS. */
549 ab->code_checkcode = afs_CheckCode(code, treq, 430);
551 ab->flags |= BUVALID;
552 if (ab->flags & BUWAIT) {
553 ab->flags &= ~BUWAIT;
557 afs_DestroyReq(treq);
561 BPartialStore(struct brequest *ab)
565 struct vrequest *treq = NULL;
566 int locked, shared_locked = 0;
569 if ((code = afs_CreateReq(&treq, ab->cred)))
572 locked = tvc->lock.excl_locked? 1:0;
574 ObtainWriteLock(&tvc->lock, 1209);
575 else if (!(tvc->lock.excl_locked & WRITE_LOCK)) {
577 ConvertSToRLock(&tvc->lock);
579 code = afs_StoreAllSegments(tvc, treq, AFS_ASYNC);
581 ReleaseWriteLock(&tvc->lock);
582 else if (shared_locked)
583 ConvertSToRLock(&tvc->lock);
584 /* now set final return code, and wakeup anyone waiting */
585 if ((ab->flags & BUVALID) == 0) {
586 /* set final code, since treq doesn't go across processes */
588 ab->code_checkcode = afs_CheckCode(code, treq, 43);
589 ab->flags |= BUVALID;
590 if (ab->flags & BUWAIT) {
591 ab->flags &= ~BUWAIT;
595 afs_DestroyReq(treq);
599 BInvalidateSegments(struct brequest *ab)
602 struct vcache *tvc = ab->vc;
604 code = afs_InvalidateAllSegments_once(tvc);
606 /* Set return code, and wakeup anyone waiting. */
607 if ((ab->flags & BUVALID) == 0) {
608 ab->code_raw = ab->code_checkcode = code;
609 ab->flags |= BUVALID;
610 if ((ab->flags & BUWAIT)) {
611 ab->flags &= ~BUWAIT;
617 /* release a held request buffer */
619 afs_BRelease(struct brequest *ab)
622 AFS_STATCNT(afs_BRelease);
623 ObtainWriteLock(&afs_xbrs, 294);
624 if (--ab->refCount <= 0) {
628 afs_osi_Wakeup(&afs_brsWaiters);
629 ReleaseWriteLock(&afs_xbrs);
632 /* return true if bkg fetch daemons are all busy */
636 AFS_STATCNT(afs_BBusy);
637 if (afs_brsDaemons > 0)
643 afs_BQueue(short aopcode, struct vcache *avc,
644 afs_int32 dontwait, afs_int32 ause, afs_ucred_t *acred,
645 afs_size_t asparm0, afs_size_t asparm1, void *apparm0,
646 void *apparm1, void *apparm2)
651 AFS_STATCNT(afs_BQueue);
652 ObtainWriteLock(&afs_xbrs, 296);
655 for (i = 0; i < NBRS; i++, tb++) {
656 if (tb->refCount == 0)
661 tb->opcode = aopcode;
668 osi_Assert(osi_vnhold(avc) == 0);
670 tb->refCount = ause + 1;
671 tb->size_parm[0] = asparm0;
672 tb->size_parm[1] = asparm1;
673 tb->ptr_parm[0] = apparm0;
674 tb->ptr_parm[1] = apparm1;
675 tb->ptr_parm[2] = apparm2;
677 tb->code_raw = tb->code_checkcode = 0;
678 tb->ts = afs_brs_count++;
679 /* if daemons are waiting for work, wake them up */
680 if (afs_brsDaemons > 0) {
681 afs_osi_Wakeup(&afs_brsDaemons);
683 ReleaseWriteLock(&afs_xbrs);
687 ReleaseWriteLock(&afs_xbrs);
690 /* no free buffers, sleep a while */
692 ReleaseWriteLock(&afs_xbrs);
693 afs_osi_Sleep(&afs_brsWaiters);
694 ObtainWriteLock(&afs_xbrs, 301);
700 /* AIX 4.1 has a much different sleep/wakeup mechanism available for use.
701 * The modifications here will work for either a UP or MP machine.
703 struct buf *afs_asyncbuf = (struct buf *)0;
704 tid_t afs_asyncbuf_cv = EVENT_NULL;
705 afs_int32 afs_biodcnt = 0;
707 /* in implementing this, I assumed that all external linked lists were
710 * Several places in this code traverse a linked list. The algorithm
711 * used here is probably unfamiliar to most people. Careful examination
712 * will show that it eliminates an assignment inside the loop, as compared
713 * to the standard algorithm, at the cost of occasionally using an extra
719 * This function obtains, and returns, a pointer to a buffer for
720 * processing by a daemon. It sleeps until such a buffer is available.
721 * The source of buffers for it is the list afs_asyncbuf (see also
722 * afs_gn_strategy). This function may be invoked concurrently by
723 * several processes, that is, several instances of the same daemon.
724 * afs_gn_strategy, which adds buffers to the list, runs at interrupt
725 * level, while get_bioreq runs at process level.
727 * Since AIX 4.1 can wake just one process at a time, the separate sleep
728 * addresses have been removed.
729 * Note that the kernel_lock is held until the e_sleep_thread() occurs.
730 * The afs_asyncbuf_lock is primarily used to serialize access between
731 * process and interrupts.
733 Simple_lock afs_asyncbuf_lock;
737 struct buf *bp = NULL;
739 struct buf **bestlbpP, **lbpP;
741 struct buf *t1P, *t2P; /* temp pointers for list manipulation */
744 struct afs_bioqueue *s;
746 /* ??? Does the forward pointer of the returned buffer need to be NULL?
749 /* Disable interrupts from the strategy function, and save the
750 * prior priority level and lock access to the afs_asyncbuf.
753 oldPriority = disable_lock(INTMAX, &afs_asyncbuf_lock);
757 /* look for oldest buffer */
758 bp = bestbp = afs_asyncbuf;
759 bestage = (long)bestbp->av_back;
760 bestlbpP = &afs_asyncbuf;
766 if ((long)bp->av_back - bestage < 0) {
769 bestage = (long)bp->av_back;
773 *bestlbpP = bp->av_forw;
776 /* If afs_asyncbuf is null, it is necessary to go to sleep.
777 * e_wakeup_one() ensures that only one thread wakes.
780 /* The LOCK_HANDLER indicates to e_sleep_thread to only drop the
781 * lock on an MP machine.
784 e_sleep_thread(&afs_asyncbuf_cv, &afs_asyncbuf_lock,
785 LOCK_HANDLER | INTERRUPTIBLE);
786 if (interrupted == THREAD_INTERRUPTED) {
787 /* re-enable interrupts from strategy */
788 unlock_enable(oldPriority, &afs_asyncbuf_lock);
792 } /* end of "else asyncbuf is empty" */
793 } /* end of "inner loop" */
797 unlock_enable(oldPriority, &afs_asyncbuf_lock);
800 /* For the convenience of other code, replace the gnodes in
801 * the b_vp field of bp and the other buffers on the b_work
802 * chain with the corresponding vnodes.
804 * ??? what happens to the gnodes? They're not just cut loose,
808 t2P = (struct buf *)t1P->b_work;
809 t1P->b_vp = ((struct gnode *)t1P->b_vp)->gn_vnode;
813 t1P = (struct buf *)t2P->b_work;
814 t2P->b_vp = ((struct gnode *)t2P->b_vp)->gn_vnode;
819 /* If the buffer does not specify I/O, it may immediately
820 * be returned to the caller. This condition is detected
821 * by examining the buffer's flags (the b_flags field). If
822 * the B_PFPROT bit is set, the buffer represents a protection
823 * violation, rather than a request for I/O. The remainder
824 * of the outer loop handles the case where the B_PFPROT bit is clear.
826 if (bp->b_flags & B_PFPROT) {
831 } /* end of function get_bioreq() */
836 * This function is the daemon. It is called from the syscall
837 * interface. Ordinarily, a script or an administrator will run a
838 * daemon startup utility, specifying the number of I/O daemons to
839 * run. The utility will fork off that number of processes,
840 * each making the appropriate syscall, which will cause this
841 * function to be invoked.
843 static int afs_initbiod = 0; /* this is self-initializing code */
846 afs_BioDaemon(afs_int32 nbiods)
848 afs_int32 code, s, pflg = 0;
850 struct buf *bp, *bp1, *tbp1, *tbp2; /* temp pointers only */
858 /* pin lock, since we'll be using it in an interrupt. */
859 lock_alloc(&afs_asyncbuf_lock, LOCK_ALLOC_PIN, 2, 1);
860 simple_lock_init(&afs_asyncbuf_lock);
861 pin(&afs_asyncbuf, sizeof(struct buf *));
862 pin(&afs_asyncbuf_cv, sizeof(afs_int32));
865 /* Ignore HUP signals... */
867 sigset_t sigbits, osigbits;
869 * add SIGHUP to the set of already masked signals
871 SIGFILLSET(sigbits); /* allow all signals */
872 SIGDELSET(sigbits, SIGHUP); /* except SIGHUP */
873 limit_sigs(&sigbits, &osigbits); /* and already masked */
875 /* Main body starts here -- this is an intentional infinite loop, and
878 * Now, the loop will exit if get_bioreq() returns NULL, indicating
879 * that we've been interrupted.
882 bp = afs_get_bioreq();
884 break; /* we were interrupted */
885 if (code = setjmpx(&jmpbuf)) {
886 /* This should not have happend, maybe a lack of resources */
888 s = disable_lock(INTMAX, &afs_asyncbuf_lock);
889 for (bp1 = bp; bp; bp = bp1) {
891 bp1 = (struct buf *)bp1->b_work;
894 bp->b_flags |= B_ERROR;
897 unlock_enable(s, &afs_asyncbuf_lock);
901 vcp = VTOAFS(bp->b_vp);
902 if (bp->b_flags & B_PFSTORE) { /* XXXX */
903 ObtainWriteLock(&vcp->lock, 404);
904 if (vcp->v.v_gnode->gn_mwrcnt) {
905 afs_offs_t newlength =
906 (afs_offs_t) dbtob(bp->b_blkno) + bp->b_bcount;
907 if (vcp->f.m.Length < newlength) {
908 afs_Trace4(afs_iclSetp, CM_TRACE_SETLENGTH,
909 ICL_TYPE_STRING, __FILE__, ICL_TYPE_LONG,
910 __LINE__, ICL_TYPE_OFFSET,
911 ICL_HANDLE_OFFSET(vcp->f.m.Length),
912 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(newlength));
913 vcp->f.m.Length = newlength;
916 ReleaseWriteLock(&vcp->lock);
918 /* If the buffer represents a protection violation, rather than
919 * an actual request for I/O, no special action need be taken.
921 if (bp->b_flags & B_PFPROT) {
922 iodone(bp); /* Notify all users of the buffer that we're done */
927 ObtainWriteLock(&vcp->pvmlock, 211);
929 * First map its data area to a region in the current address space
930 * by calling vm_att with the subspace identifier, and a pointer to
931 * the data area. vm_att returns a new data area pointer, but we
932 * also want to hang onto the old one.
934 tmpaddr = bp->b_baddr;
935 bp->b_baddr = (caddr_t) vm_att(bp->b_xmemd.subspace_id, tmpaddr);
936 tmperr = afs_ustrategy(bp); /* temp variable saves offset calculation */
937 if (tmperr) { /* in non-error case */
938 bp->b_flags |= B_ERROR; /* should other flags remain set ??? */
939 bp->b_error = tmperr;
942 /* Unmap the buffer's data area by calling vm_det. Reset data area
943 * to the value that we saved above.
946 bp->b_baddr = tmpaddr;
949 * buffer may be linked with other buffers via the b_work field.
950 * See also afs_gn_strategy. For each buffer in the chain (including
951 * bp) notify all users of the buffer that the daemon is finished
952 * using it by calling iodone.
953 * assumes iodone can modify the b_work field.
956 tbp2 = (struct buf *)tbp1->b_work;
961 tbp1 = (struct buf *)tbp2->b_work;
967 ReleaseWriteLock(&vcp->pvmlock); /* Unlock the vnode. */
969 } /* infinite loop (unless we're interrupted) */
970 } /* end of afs_BioDaemon() */
972 #endif /* AFS_AIX41_ENV */
977 afs_BackgroundDaemon_once(void)
979 LOCK_INIT(&afs_xbrs, "afs_xbrs");
980 memset(afs_brs, 0, sizeof(afs_brs));
982 #if defined (AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
984 * steal the first daemon for doing delayed DSlot flushing
985 * (see afs_GetDownDSlot)
994 brequest_release(struct brequest *tb)
997 AFS_RELE(AFSTOV(tb->vc)); /* MUST call vnode layer or could lose vnodes */
1002 tb->cred = (afs_ucred_t *)0;
1004 afs_BRelease(tb); /* this grabs and releases afs_xbrs lock */
1009 should_do_noop(int foundAny, int n_processed)
1011 if (!foundAny && n_processed > 0) {
1012 /* If there aren't any requests right now, and we've processed
1013 * at least one request, do a noop. */
1015 } else if (foundAny && n_processed > 100) {
1016 /* If we've processed over 100 requests in a row, do a noop. */
1024 * Entry point for background daemon processes.
1026 * For old-style background daemons (non-AFS_NEW_BKG), a background daemon afsd
1027 * process will end up in this function, and it will stay in here forever
1028 * processing in-kernel bkg requests until the client shuts down.
1030 * For new-style background daemons (AFS_NEW_BKG), we can pass data back to
1031 * afsd to perform some background operations in userspace, by populating
1032 * 'uspc' with the operation to perform and then returning. When the afsd
1033 * process enters this function again, the return code for that operation is
1034 * also provided in 'uspc'.
1036 * @param[inout] uspc Userspace operation data. If uspc->ts is non-negative
1037 * on entry, the related background request has finished,
1038 * and we're providing the return code. On return,
1039 * contains the userspace operation to perform.
1040 * @param[inout] param1 Operation-specific pointer.
1041 * @param[inout] param2 Operation-specific pointer.
1043 * @return Always returns 0.
1047 afs_BackgroundDaemon(struct afs_uspc_param *uspc, void *param1, void *param2)
1050 afs_BackgroundDaemon(void)
1053 struct brequest *tb;
1055 int n_processed = 0;
1057 AFS_STATCNT(afs_BackgroundDaemon);
1058 /* initialize subsystem */
1060 /* Irix with "short stack" exits */
1061 afs_BackgroundDaemon_once();
1064 if (uspc->reqtype == AFS_USPC_NOOP) {
1065 /* The daemon is re-entering from a noop, not actually returning data;
1066 * don't look for an existing request. */
1069 } else if (uspc->ts > -1) {
1070 /* If it's a re-entering syscall, complete the request and release */
1072 for (i = 0; i < NBRS; i++, tb++) {
1073 if (tb->ts == uspc->ts) {
1074 /* copy the userspace status back in */
1075 tb->code_raw = tb->code_checkcode = uspc->retval;
1076 /* mark it valid and notify our caller */
1077 tb->flags |= BUVALID;
1078 if (tb->flags & BUWAIT) {
1079 tb->flags &= ~BUWAIT;
1082 brequest_release(tb);
1087 afs_osi_MaskUserLoop();
1089 /* Otherwise it's a new one */
1095 ObtainWriteLock(&afs_xbrs, 302);
1098 struct brequest *min_tb = NULL;
1100 if (afs_termState == AFSOP_STOP_BKG) {
1101 if (--afs_nbrs <= 0)
1102 afs_termState = AFSOP_STOP_RXCALLBACK;
1103 ReleaseWriteLock(&afs_xbrs);
1104 afs_osi_Wakeup(&afs_termState);
1106 memset(uspc, 0, sizeof(*uspc));
1107 uspc->reqtype = AFS_USPC_SHUTDOWN;
1114 /* find a request */
1117 for (i = 0; i < NBRS; i++, tb++) {
1118 /* look for request with smallest ts */
1119 if ((tb->refCount > 0) && !(tb->flags & BSTARTED)) {
1120 /* new request, not yet picked up */
1121 if ((min_tb && (min_ts - tb->ts > 0)) || !min_tb) {
1127 if ((tb = min_tb)) {
1128 /* claim and process this request */
1129 tb->flags |= BSTARTED;
1130 ReleaseWriteLock(&afs_xbrs);
1133 afs_Trace1(afs_iclSetp, CM_TRACE_BKG1, ICL_TYPE_INT32,
1135 if (tb->opcode == BOP_FETCH)
1137 #if defined(AFS_CACHE_BYPASS)
1138 else if (tb->opcode == BOP_FETCH_NOCACHE)
1139 BPrefetchNoCache(tb);
1141 else if (tb->opcode == BOP_STORE)
1143 else if (tb->opcode == BOP_PATH)
1145 #ifdef AFS_DARWIN80_ENV
1146 else if (tb->opcode == BOP_MOVE) {
1147 memcpy(uspc, (struct afs_uspc_param *) tb->ptr_parm[0],
1148 sizeof(struct afs_uspc_param));
1150 /* string lengths capped in move vop; copy NUL tho */
1151 memcpy(param1, (char *)tb->ptr_parm[1],
1152 strlen(tb->ptr_parm[1])+1);
1153 memcpy(param2, (char *)tb->ptr_parm[2],
1154 strlen(tb->ptr_parm[2])+1);
1158 else if (tb->opcode == BOP_PARTIAL_STORE)
1160 else if (tb->opcode == BOP_INVALIDATE_SEGMENTS)
1161 BInvalidateSegments(tb);
1163 panic("background bop");
1164 brequest_release(tb);
1165 ObtainWriteLock(&afs_xbrs, 305);
1169 if (should_do_noop(foundAny, n_processed)) {
1170 ReleaseWriteLock(&afs_xbrs);
1171 memset(uspc, 0, sizeof(*uspc));
1172 uspc->reqtype = AFS_USPC_NOOP;
1178 /* wait for new request */
1180 ReleaseWriteLock(&afs_xbrs);
1181 afs_osi_Sleep(&afs_brsDaemons);
1182 ObtainWriteLock(&afs_xbrs, 307);
1193 shutdown_daemons(void)
1195 AFS_STATCNT(shutdown_daemons);
1196 if (afs_cold_shutdown) {
1197 afs_brsDaemons = brsInit = 0;
1199 memset(afs_brs, 0, sizeof(afs_brs));
1200 memset(&afs_xbrs, 0, sizeof(afs_lock_t));
1202 #ifdef AFS_AIX41_ENV
1203 lock_free(&afs_asyncbuf_lock);
1204 unpin(&afs_asyncbuf, sizeof(struct buf *));
1205 unpin(&afs_asyncbuf_cv, sizeof(afs_int32));
1211 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1213 * sgi - daemon - handles certain operations that otherwise
1214 * would use up too much kernel stack space
1216 * This all assumes that since the caller must have the xdcache lock
1217 * exclusively that the list will never be more than one long
1218 * and noone else can attempt to add anything until we're done.
1220 SV_TYPE afs_sgibksync;
1221 SV_TYPE afs_sgibkwait;
1222 lock_t afs_sgibklock;
1223 struct dcache *afs_sgibklist;
1231 if (afs_sgibklock == NULL) {
1232 SV_INIT(&afs_sgibksync, "bksync", 0, 0);
1233 SV_INIT(&afs_sgibkwait, "bkwait", 0, 0);
1234 SPINLOCK_INIT(&afs_sgibklock, "bklock");
1236 s = SPLOCK(afs_sgibklock);
1238 /* wait for something to do */
1239 SP_WAIT(afs_sgibklock, s, &afs_sgibksync, PINOD);
1240 osi_Assert(afs_sgibklist);
1242 /* XX will probably need to generalize to real list someday */
1243 s = SPLOCK(afs_sgibklock);
1244 while (afs_sgibklist) {
1245 tdc = afs_sgibklist;
1246 afs_sgibklist = NULL;
1247 SPUNLOCK(afs_sgibklock, s);
1249 tdc->dflags &= ~DFEntryMod;
1250 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
1252 s = SPLOCK(afs_sgibklock);
1255 /* done all the work - wake everyone up */
1256 while (SV_SIGNAL(&afs_sgibkwait));