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
18 #include <afsconfig.h>
19 #include "afs/param.h"
22 #include "afs/sysincludes.h" /* Standard vendor system headers */
23 #include "afsincludes.h" /* Afs-based standard headers */
24 #include "afs/afs_stats.h" /* statistics */
25 #include "afs/afs_cbqueue.h"
26 #include "afs/nfsclient.h"
27 #include "afs/afs_osidnlc.h"
28 #include "afs/afs_osi.h"
31 extern char afs_zeros[AFS_ZEROS];
33 /* Imported variables */
34 extern afs_rwlock_t afs_xdcache;
35 extern unsigned char *afs_indexFlags;
36 extern afs_hyper_t *afs_indexTimes; /* Dcache entry Access times */
37 extern afs_hyper_t afs_indexCounter; /* Fake time for marking index */
40 /* Forward declarations */
41 void afs_PrefetchChunk(struct vcache *avc, struct dcache *adc,
42 afs_ucred_t *acred, struct vrequest *areq);
45 afs_read(struct vcache *avc, struct uio *auio, afs_ucred_t *acred,
48 afs_size_t totalLength;
49 afs_size_t transferLength;
51 afs_size_t offset, tlen;
54 struct dcache *tdc = 0;
55 afs_int32 error, trybusy = 1;
56 struct uio *tuiop = NULL;
60 AFS_STATCNT(afs_read);
67 /* check that we have the latest status info in the vnode cache */
68 if ((code = afs_InitReq(&treq, acred)))
73 osi_Panic("null avc in afs_GenericRead");
75 code = afs_VerifyVCache(avc, &treq);
77 code = afs_CheckCode(code, &treq, 8); /* failed to get it */
81 #ifndef AFS_VM_RDWR_ENV
82 if (AFS_NFSXLATORREQ(acred)) {
84 (avc, PRSFS_READ, &treq,
85 CHECK_MODE_BITS | CMB_ALLOW_EXEC_AS_READ)) {
86 code = afs_CheckCode(EACCES, &treq, 9);
92 totalLength = AFS_UIO_RESID(auio);
93 filePos = AFS_UIO_OFFSET(auio);
94 afs_Trace4(afs_iclSetp, CM_TRACE_READ, ICL_TYPE_POINTER, avc,
95 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_INT32,
96 totalLength, ICL_TYPE_OFFSET,
97 ICL_HANDLE_OFFSET(avc->f.m.Length));
101 ObtainReadLock(&avc->lock);
102 #if defined(AFS_TEXT_ENV) && !defined(AFS_VM_RDWR_ENV)
103 if (avc->flushDV.high == AFS_MAXDV && avc->flushDV.low == AFS_MAXDV) {
104 hset(avc->flushDV, avc->f.m.DataVersion);
113 /* This bit is bogus. We're checking to see if the read goes past the
114 * end of the file. If so, we should be zeroing out all of the buffers
115 * that the client has passed into us (there is a danger that we may leak
116 * kernel memory if we do not). However, this behaviour is disabled by
117 * not setting len before this segment runs, and by setting len to 0
118 * immediately we enter it. In addition, we also need to check for a read
119 * which partially goes off the end of the file in the while loop below.
122 if (filePos >= avc->f.m.Length) {
124 len = sizeof(afs_zeros); /* and in 0 buffer */
127 tuiop = afsio_partialcopy(auio, trimlen);
128 AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code);
131 while (avc->f.m.Length > 0 && totalLength > 0) {
132 /* read all of the cached info */
133 if (filePos >= avc->f.m.Length)
134 break; /* all done */
137 ReleaseReadLock(&tdc->lock);
140 tdc = afs_FindDCache(avc, filePos);
142 ObtainReadLock(&tdc->lock);
143 offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk);
144 len = tdc->validPos - filePos;
147 /* a tricky question: does the presence of the DFFetching flag
148 * mean that we're fetching the latest version of the file? No.
149 * The server could update the file as soon as the fetch responsible
150 * for the setting of the DFFetching flag completes.
152 * However, the presence of the DFFetching flag (visible under
153 * a dcache read lock since it is set and cleared only under a
154 * dcache write lock) means that we're fetching as good a version
155 * as was known to this client at the time of the last call to
156 * afs_VerifyVCache, since the latter updates the stat cache's
157 * m.DataVersion field under a vcache write lock, and from the
158 * time that the DFFetching flag goes on in afs_GetDCache (before
159 * the fetch starts), to the time it goes off (after the fetch
160 * completes), afs_GetDCache keeps at least a read lock on the
163 * This means that if the DFFetching flag is set, we can use that
164 * data for any reads that must come from the current version of
165 * the file (current == m.DataVersion).
167 * Another way of looking at this same point is this: if we're
168 * fetching some data and then try do an afs_VerifyVCache, the
169 * VerifyVCache operation will not complete until after the
170 * DFFetching flag is turned off and the dcache entry's f.versionNo
173 * Note, by the way, that if DFFetching is set,
174 * m.DataVersion > f.versionNo (the latter is not updated until
175 * after the fetch completes).
178 ReleaseReadLock(&tdc->lock);
179 afs_PutDCache(tdc); /* before reusing tdc */
181 tdc = afs_GetDCache(avc, filePos, &treq, &offset, &len, 2);
187 ObtainReadLock(&tdc->lock);
188 /* now, first try to start transfer, if we'll need the data. If
189 * data already coming, we don't need to do this, obviously. Type
190 * 2 requests never return a null dcache entry, btw.
192 if (!(tdc->dflags & DFFetching)
193 && !hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
194 /* have cache entry, it is not coming in now,
195 * and we'll need new data */
197 if (trybusy && !afs_BBusy()) {
199 /* daemon is not busy */
200 ObtainSharedLock(&tdc->mflock, 665);
201 if (!(tdc->mflags & DFFetchReq)) {
202 /* start the daemon (may already be running, however) */
203 UpgradeSToWLock(&tdc->mflock, 666);
204 tdc->mflags |= DFFetchReq;
205 bp = afs_BQueue(BOP_FETCH, avc, B_DONTWAIT, 0, acred,
206 (afs_size_t) filePos, (afs_size_t) 0,
209 /* Bkg table full; retry deadlocks */
210 tdc->mflags &= ~DFFetchReq;
211 trybusy = 0; /* Avoid bkg daemon since they're too busy */
212 ReleaseWriteLock(&tdc->mflock);
215 ConvertWToSLock(&tdc->mflock);
216 /* don't use bp pointer! */
219 ConvertSToRLock(&tdc->mflock);
220 while (!code && tdc->mflags & DFFetchReq) {
221 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT,
222 ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32,
223 __LINE__, ICL_TYPE_POINTER, tdc,
224 ICL_TYPE_INT32, tdc->dflags);
225 /* don't need waiting flag on this one */
226 ReleaseReadLock(&tdc->mflock);
227 ReleaseReadLock(&tdc->lock);
228 ReleaseReadLock(&avc->lock);
229 code = afs_osi_SleepSig(&tdc->validPos);
230 ObtainReadLock(&avc->lock);
231 ObtainReadLock(&tdc->lock);
232 ObtainReadLock(&tdc->mflock);
234 ReleaseReadLock(&tdc->mflock);
241 /* now data may have started flowing in (if DFFetching is on). If
242 * data is now streaming in, then wait for some interesting stuff.
245 while (!code && (tdc->dflags & DFFetching)
246 && tdc->validPos <= filePos) {
247 /* too early: wait for DFFetching flag to vanish,
248 * or data to appear */
249 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING,
250 __FILE__, ICL_TYPE_INT32, __LINE__,
251 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
253 ReleaseReadLock(&tdc->lock);
254 ReleaseReadLock(&avc->lock);
255 code = afs_osi_SleepSig(&tdc->validPos);
256 ObtainReadLock(&avc->lock);
257 ObtainReadLock(&tdc->lock);
263 /* fetching flag gone, data is here, or we never tried
264 * (BBusy for instance) */
265 if (tdc->dflags & DFFetching) {
266 /* still fetching, some new data is here:
267 * compute length and offset */
268 offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk);
269 len = tdc->validPos - filePos;
271 /* no longer fetching, verify data version
272 * (avoid new GetDCache call) */
273 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)
274 && ((len = tdc->validPos - filePos) > 0)) {
275 offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk);
277 /* don't have current data, so get it below */
278 afs_Trace3(afs_iclSetp, CM_TRACE_VERSIONNO,
279 ICL_TYPE_INT64, ICL_HANDLE_OFFSET(filePos),
280 ICL_TYPE_HYPER, &avc->f.m.DataVersion,
281 ICL_TYPE_HYPER, &tdc->f.versionNo);
282 ReleaseReadLock(&tdc->lock);
289 /* If we get, it was not possible to start the
290 * background daemon. With flag == 1 afs_GetDCache
291 * does the FetchData rpc synchronously.
293 ReleaseReadLock(&avc->lock);
294 tdc = afs_GetDCache(avc, filePos, &treq, &offset, &len, 1);
295 ObtainReadLock(&avc->lock);
297 ObtainReadLock(&tdc->lock);
301 afs_Trace3(afs_iclSetp, CM_TRACE_VNODEREAD, ICL_TYPE_POINTER, tdc,
302 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(offset),
303 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(len));
315 if (len > totalLength)
316 len = totalLength; /* will read len bytes */
317 if (len <= 0) { /* shouldn't get here if DFFetching is on */
318 /* read past the end of a chunk, may not be at next chunk yet, and yet
319 * also not at eof, so may have to supply fake zeros */
320 len = AFS_CHUNKTOSIZE(tdc->f.chunk) - offset; /* bytes left in chunk addr space */
321 if (len > totalLength)
322 len = totalLength; /* and still within xfr request */
323 tlen = avc->f.m.Length - offset; /* and still within file */
327 len = sizeof(afs_zeros); /* and in 0 buffer */
329 tuiop = afsio_partialcopy(auio, trimlen);
330 AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code);
336 /* get the data from the cache */
338 /* mung uio structure to be right for this transfer */
340 tuiop = afsio_partialcopy(auio, trimlen);
341 AFS_UIO_SETOFFSET(tuiop, offset);
343 code = (*(afs_cacheType->vreadUIO))(&tdc->f.inode, tuiop);
350 /* otherwise we've read some, fixup length, etc and continue with next seg */
351 len = len - AFS_UIO_RESID(tuiop); /* compute amount really transferred */
353 afsio_skip(auio, trimlen); /* update input uio structure */
355 transferLength += len;
359 break; /* surprise eof */
364 } /* the whole while loop */
369 * tdc->lock(R) if tdc
372 /* if we make it here with tdc non-zero, then it is the last chunk we
373 * dealt with, and we have to release it when we're done. We hold on
374 * to it in case we need to do a prefetch.
377 ReleaseReadLock(&tdc->lock);
378 #if !defined(AFS_VM_RDWR_ENV)
379 /* try to queue prefetch, if needed */
381 if (!(tdc->mflags &DFNextStarted))
382 afs_PrefetchChunk(avc, tdc, acred, &treq);
388 ReleaseReadLock(&avc->lock);
390 code = afs_CheckCode(error, &treq, 10);
400 /* called with the dcache entry triggering the fetch, the vcache entry involved,
401 * and a vrequest for the read call. Marks the dcache entry as having already
402 * triggered a prefetch, starts the prefetch going and sets the DFFetchReq
403 * flag in the prefetched block, so that the next call to read knows to wait
404 * for the daemon to start doing things.
406 * This function must be called with the vnode at least read-locked, and
407 * no locks on the dcache, because it plays around with dcache entries.
410 afs_PrefetchChunk(struct vcache *avc, struct dcache *adc,
411 afs_ucred_t *acred, struct vrequest *areq)
415 afs_size_t j1, j2; /* junk vbls for GetDCache to trash */
417 offset = adc->f.chunk + 1; /* next chunk we'll need */
418 offset = AFS_CHUNKTOBASE(offset); /* base of next chunk */
419 ObtainReadLock(&adc->lock);
420 ObtainSharedLock(&adc->mflock, 662);
421 if (offset < avc->f.m.Length && !(adc->mflags & DFNextStarted)
425 UpgradeSToWLock(&adc->mflock, 663);
426 adc->mflags |= DFNextStarted; /* we've tried to prefetch for this guy */
427 ReleaseWriteLock(&adc->mflock);
428 ReleaseReadLock(&adc->lock);
430 tdc = afs_GetDCache(avc, offset, areq, &j1, &j2, 2); /* type 2 never returns 0 */
432 * In disconnected mode, type 2 can return 0 because it doesn't
433 * make any sense to allocate a dcache we can never fill
438 ObtainSharedLock(&tdc->mflock, 651);
439 if (!(tdc->mflags & DFFetchReq)) {
440 /* ask the daemon to do the work */
441 UpgradeSToWLock(&tdc->mflock, 652);
442 tdc->mflags |= DFFetchReq; /* guaranteed to be cleared by BKG or GetDCache */
443 /* last parm (1) tells bkg daemon to do an afs_PutDCache when it is done,
444 * since we don't want to wait for it to finish before doing so ourselves.
446 bp = afs_BQueue(BOP_FETCH, avc, B_DONTWAIT, 0, acred,
447 (afs_size_t) offset, (afs_size_t) 1, tdc,
448 (void *)0, (void *)0);
450 /* Bkg table full; just abort non-important prefetching to avoid deadlocks */
451 tdc->mflags &= ~DFFetchReq;
452 ReleaseWriteLock(&tdc->mflock);
456 * DCLOCKXXX: This is a little sketchy, since someone else
457 * could have already started a prefetch.. In practice,
458 * this probably doesn't matter; at most it would cause an
459 * extra slot in the BKG table to be used up when someone
460 * prefetches this for the second time.
462 ObtainReadLock(&adc->lock);
463 ObtainWriteLock(&adc->mflock, 664);
464 adc->mflags &= ~DFNextStarted;
465 ReleaseWriteLock(&adc->mflock);
466 ReleaseReadLock(&adc->lock);
468 ReleaseWriteLock(&tdc->mflock);
471 ReleaseSharedLock(&tdc->mflock);
475 ReleaseSharedLock(&adc->mflock);
476 ReleaseReadLock(&adc->lock);
481 afs_UFSReadUIO(afs_dcache_id_t *cacheId, struct uio *tuiop)
484 struct osi_file *tfile;
486 tfile = (struct osi_file *) osi_UFSOpen(cacheId);
488 #if defined(AFS_AIX41_ENV)
491 VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, tuiop, NULL, NULL,
492 NULL, afs_osi_credp);
494 #elif defined(AFS_AIX32_ENV)
496 VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, tuiop, NULL, NULL);
497 /* Flush all JFS pages now for big performance gain in big file cases
498 * If we do something like this, must check to be sure that AFS file
499 * isn't mmapped... see afs_gn_map() for why.
502 if (tfile->vnode->v_gnode && tfile->vnode->v_gnode->gn_seg) {
503 any different ways to do similar things:
504 so far, the best performing one is #2, but #1 might match it if we
505 straighten out the confusion regarding which pages to flush. It
507 1. vm_flushp(tfile->vnode->v_gnode->gn_seg, 0, len/PAGESIZE - 1);
508 2. vm_releasep(tfile->vnode->v_gnode->gn_seg, offset/PAGESIZE,
509 (len + PAGESIZE-1)/PAGESIZE);
510 3. vms_inactive(tfile->vnode->v_gnode->gn_seg) Doesn't work correctly
511 4. vms_delete(tfile->vnode->v_gnode->gn_seg) probably also fails
512 tfile->vnode->v_gnode->gn_seg = NULL;
516 Unfortunately, this seems to cause frequent "cache corruption" episodes.
517 vm_releasep(tfile->vnode->v_gnode->gn_seg, offset/PAGESIZE,
518 (len + PAGESIZE-1)/PAGESIZE);
521 #elif defined(AFS_AIX_ENV)
523 VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, (off_t) & offset,
524 tuiop, NULL, NULL, -1);
525 #elif defined(AFS_SUN5_ENV)
527 #ifdef AFS_SUN510_ENV
530 VOP_RWLOCK(tfile->vnode, 0, &ct);
531 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp, &ct);
532 VOP_RWUNLOCK(tfile->vnode, 0, &ct);
535 VOP_RWLOCK(tfile->vnode, 0);
536 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp);
537 VOP_RWUNLOCK(tfile->vnode, 0);
540 #elif defined(AFS_SGI_ENV)
542 AFS_VOP_RWLOCK(tfile->vnode, VRWLOCK_READ);
543 AFS_VOP_READ(tfile->vnode, tuiop, IO_ISLOCKED, afs_osi_credp,
545 AFS_VOP_RWUNLOCK(tfile->vnode, VRWLOCK_READ);
547 #elif defined(AFS_HPUX100_ENV)
549 code = VOP_RDWR(tfile->vnode, tuiop, UIO_READ, 0, afs_osi_credp);
551 #elif defined(AFS_LINUX20_ENV)
553 code = osi_rdwr(tfile, tuiop, UIO_READ);
555 #elif defined(AFS_DARWIN80_ENV)
557 code = VNOP_READ(tfile->vnode, tuiop, 0, afs_osi_ctxtp);
559 #elif defined(AFS_DARWIN_ENV)
561 VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, current_proc());
562 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp);
563 VOP_UNLOCK(tfile->vnode, 0, current_proc());
565 #elif defined(AFS_FBSD80_ENV)
567 VOP_LOCK(tfile->vnode, LK_EXCLUSIVE);
568 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp);
569 VOP_UNLOCK(tfile->vnode, 0);
571 #elif defined(AFS_FBSD_ENV)
573 VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, curthread);
574 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp);
575 VOP_UNLOCK(tfile->vnode, 0, curthread);
577 #elif defined(AFS_NBSD_ENV)
578 tuiop->uio_rw = UIO_READ;
580 VOP_LOCK(tfile->vnode, LK_EXCLUSIVE);
581 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp);
582 # if defined(AFS_NBSD60_ENV)
583 VOP_UNLOCK(tfile->vnode);
585 VOP_UNLOCK(tfile->vnode, 0);
588 #elif defined(AFS_XBSD_ENV)
590 VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, curproc);
591 code = VOP_READ(tfile->vnode, tuiop, 0, afs_osi_credp);
592 VOP_UNLOCK(tfile->vnode, 0, curproc);
595 code = VOP_RDWR(tfile->vnode, tuiop, UIO_READ, 0, afs_osi_credp);