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
16 #include <afsconfig.h>
17 #include "afs/param.h"
20 #include "afs/sysincludes.h" /* Standard vendor system headers */
21 #include "afsincludes.h" /* Afs-based standard headers */
22 #include "afs/afs_stats.h" /* statistics */
23 #include "afs/afs_cbqueue.h"
24 #include "afs/nfsclient.h"
25 #include "afs/afs_osidnlc.h"
26 #include "afs/unified_afs.h"
28 /* question: does afs_create need to set CDirty in the adp or the avc?
29 * I think we can get away without it, but I'm not sure. Note that
30 * afs_setattr is called in here for truncation.
34 afs_create(OSI_VC_DECL(adp), char *aname, struct vattr *attrs, int flags,
35 int amode, struct vcache **avcp, afs_ucred_t *acred)
36 #else /* AFS_SGI64_ENV */
38 afs_create(OSI_VC_DECL(adp), char *aname, struct vattr *attrs,
39 enum vcexcl aexcl, int amode, struct vcache **avcp,
41 #endif /* AFS_SGI64_ENV */
43 afs_int32 origCBs, origZaps, finalZaps;
45 register afs_int32 code;
46 register struct afs_conn *tc;
47 struct VenusFid newFid;
48 struct AFSStoreStatus InStatus;
49 struct AFSFetchStatus OutFidStatus, OutDirStatus;
50 struct AFSVolSync tsync;
51 struct AFSCallBack CallBack;
54 afs_size_t offset, len;
55 struct server *hostp = 0;
57 struct volume *volp = 0;
58 struct afs_fakestat_state fakestate;
63 AFS_STATCNT(afs_create);
64 if ((code = afs_InitReq(&treq, acred)))
67 afs_Trace3(afs_iclSetp, CM_TRACE_CREATE, ICL_TYPE_POINTER, adp,
68 ICL_TYPE_STRING, aname, ICL_TYPE_INT32, amode);
70 afs_InitFakeStat(&fakestate);
73 /* If avcp is passed not null, it's the old reference to this file.
74 * We can use this to avoid create races. For now, just decrement
75 * the reference count on it.
78 AFS_RELE(AFSTOV(*avcp));
83 if (strlen(aname) > AFSNAMEMAX) {
88 if (!afs_ENameOK(aname)) {
92 switch (attrs->va_type) {
95 #if !defined(AFS_SUN5_ENV)
99 /* We don't support special devices or FIFOs */
107 code = afs_EvalFakeStat(&adp, &fakestate, &treq);
111 code = afs_VerifyVCache(adp, &treq);
115 /** If the volume is read-only, return error without making an RPC to the
118 if (adp->f.states & CRO) {
123 if (AFS_IS_DISCONNECTED && !AFS_IS_DISCON_RW) {
128 tdc = afs_GetDCache(adp, (afs_size_t) 0, &treq, &offset, &len, 1);
129 ObtainWriteLock(&adp->lock, 135);
131 ObtainSharedLock(&tdc->lock, 630);
134 * Make sure that the data in the cache is current. We may have
135 * received a callback while we were waiting for the write lock.
137 if (!(adp->f.states & CStatd)
138 || (tdc && !hsame(adp->f.m.DataVersion, tdc->f.versionNo))) {
139 ReleaseWriteLock(&adp->lock);
141 ReleaseSharedLock(&tdc->lock);
147 /* see if file already exists. If it does, we only set
148 * the size attributes (to handle O_TRUNC) */
149 code = afs_dir_Lookup(tdc, aname, &newFid.Fid); /* use dnlc first xxx */
151 ReleaseSharedLock(&tdc->lock);
153 ReleaseWriteLock(&adp->lock);
157 if (aexcl != NONEXCL) {
159 code = EEXIST; /* file exists in excl mode open */
162 /* found the file, so use it */
163 newFid.Cell = adp->f.fid.Cell;
164 newFid.Fid.Volume = adp->f.fid.Fid.Volume;
166 if (newFid.Fid.Unique == 0) {
167 tvc = afs_LookupVCache(&newFid, &treq, NULL, adp, aname);
169 if (!tvc) /* lookup failed or wasn't called */
170 tvc = afs_GetVCache(&newFid, &treq, NULL, NULL);
173 /* if the thing exists, we need the right access to open it.
174 * we must check that here, since no other checks are
175 * made by the open system call */
176 len = attrs->va_size; /* only do the truncate */
178 * We used to check always for READ access before; the
179 * problem is that we will fail if the existing file
180 * has mode -w-w-w, which is wrong.
183 && !afs_AccessOK(tvc, PRSFS_READ, &treq, CHECK_MODE_BITS)) {
188 #if defined(AFS_DARWIN80_ENV)
189 if ((amode & VWRITE) || VATTR_IS_ACTIVE(attrs, va_data_size))
190 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
191 if ((amode & VWRITE) || (attrs->va_mask & AT_SIZE))
193 if ((amode & VWRITE) || len != 0xffffffff)
196 /* needed for write access check */
197 tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
198 tvc->f.parent.unique = adp->f.fid.Fid.Unique;
199 /* need write mode for these guys */
201 (tvc, PRSFS_WRITE, &treq, CHECK_MODE_BITS)) {
207 #if defined(AFS_DARWIN80_ENV)
208 if (VATTR_IS_ACTIVE(attrs, va_data_size))
209 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
210 if (attrs->va_mask & AT_SIZE)
212 if (len != 0xffffffff)
215 if (vType(tvc) != VREG) {
221 #if defined(AFS_DARWIN80_ENV)
223 VATTR_SET_SUPPORTED(attrs, va_data_size);
224 VATTR_SET_ACTIVE(attrs, va_data_size);
225 #elif defined(UKERNEL)
226 attrs->va_mask = ATTR_SIZE;
227 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
228 attrs->va_mask = AT_SIZE;
232 attrs->va_size = len;
233 ObtainWriteLock(&tvc->lock, 136);
234 tvc->f.states |= CCreating;
235 ReleaseWriteLock(&tvc->lock);
236 #if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
237 #if defined(AFS_SGI64_ENV)
239 afs_setattr(VNODE_TO_FIRST_BHV((vnode_t *) tvc),
242 code = afs_setattr(tvc, attrs, 0, acred);
243 #endif /* AFS_SGI64_ENV */
244 #else /* SUN5 || SGI */
245 code = afs_setattr(tvc, attrs, acred);
246 #endif /* SUN5 || SGI */
247 ObtainWriteLock(&tvc->lock, 137);
248 tvc->f.states &= ~CCreating;
249 ReleaseWriteLock(&tvc->lock);
257 code = ENOENT; /* shouldn't get here */
258 /* make sure vrefCount bumped only if code == 0 */
263 /* if we create the file, we don't do any access checks, since
264 * that's how O_CREAT is supposed to work */
265 if (adp->f.states & CForeign) {
266 origCBs = afs_allCBs;
267 origZaps = afs_allZaps;
269 origCBs = afs_evenCBs; /* if changes, we don't really have a callback */
270 origZaps = afs_evenZaps; /* number of even numbered vnodes discarded */
272 InStatus.Mask = AFS_SETMODTIME | AFS_SETMODE | AFS_SETGROUP;
273 InStatus.ClientModTime = osi_Time();
274 InStatus.Group = (afs_int32) acred->cr_gid;
275 if (AFS_NFSXLATORREQ(acred)) {
277 * XXX The following is mainly used to fix a bug in the HP-UX
278 * nfs client where they create files with mode of 0 without
279 * doing any setattr later on to fix it. * XXX
281 #if defined(AFS_AIX_ENV)
282 if (attrs->va_mode != -1) {
284 #if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
285 if (attrs->va_mask & AT_MODE) {
287 if (attrs->va_mode != ((unsigned short)-1)) {
291 attrs->va_mode = 0x1b6; /* XXX default mode: rw-rw-rw XXX */
295 if (!AFS_IS_DISCONNECTED) {
296 /* If not disconnected, connect to the server.*/
298 InStatus.UnixModeBits = attrs->va_mode & 0xffff; /* only care about protection bits */
300 tc = afs_Conn(&adp->f.fid, &treq, SHARED_LOCK);
302 hostp = tc->srvr->server; /* remember for callback processing */
304 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_CREATEFILE);
307 RXAFS_CreateFile(tc->id, (struct AFSFid *)&adp->f.fid.Fid,
308 aname, &InStatus, (struct AFSFid *)
309 &newFid.Fid, &OutFidStatus, &OutDirStatus,
313 CallBack.ExpirationTime += now;
317 (tc, code, &adp->f.fid, &treq, AFS_STATS_FS_RPCIDX_CREATEFILE,
320 if ((code == EEXIST || code == UAEEXIST) &&
323 #else /* AFS_SGI64_ENV */
327 /* if we get an EEXIST in nonexcl mode, just do a lookup */
329 ReleaseSharedLock(&tdc->lock);
332 ReleaseWriteLock(&adp->lock);
335 #if defined(AFS_SGI64_ENV)
336 code = afs_lookup(VNODE_TO_FIRST_BHV((vnode_t *) adp), aname, avcp,
337 NULL, 0, NULL, acred);
338 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
339 code = afs_lookup(adp, aname, avcp, NULL, 0, NULL, acred);
340 #elif defined(UKERNEL)
341 code = afs_lookup(adp, aname, avcp, acred, 0);
342 #elif !defined(AFS_DARWIN_ENV)
343 code = afs_lookup(adp, aname, avcp, acred);
350 ObtainWriteLock(&afs_xcbhash, 488);
351 afs_DequeueCallback(adp);
352 adp->f.states &= ~CStatd;
353 ReleaseWriteLock(&afs_xcbhash);
354 osi_dnlc_purgedp(adp);
356 ReleaseWriteLock(&adp->lock);
358 ReleaseSharedLock(&tdc->lock);
365 #if defined(AFS_DISCON_ENV)
366 /* Generate a fake FID for disconnected mode. */
367 newFid.Cell = adp->f.fid.Cell;
368 newFid.Fid.Volume = adp->f.fid.Fid.Volume;
369 afs_GenFakeFid(&newFid, VREG, 1);
371 } /* if (!AFS_IS_DISCON_RW) */
373 /* otherwise, we should see if we can make the change to the dir locally */
375 UpgradeSToWLock(&tdc->lock, 631);
376 if (AFS_IS_DISCON_RW || afs_LocalHero(adp, tdc, &OutDirStatus, 1)) {
377 /* we can do it locally */
378 ObtainWriteLock(&afs_xdcache, 291);
379 code = afs_dir_Create(tdc, aname, &newFid.Fid);
380 ReleaseWriteLock(&afs_xdcache);
387 ReleaseWriteLock(&tdc->lock);
390 if (AFS_IS_DISCON_RW)
391 adp->f.m.LinkCount++;
393 newFid.Cell = adp->f.fid.Cell;
394 newFid.Fid.Volume = adp->f.fid.Fid.Volume;
395 ReleaseWriteLock(&adp->lock);
396 volp = afs_FindVolume(&newFid, READ_LOCK);
398 /* New tricky optimistic callback handling algorithm for file creation works
399 * as follows. We create the file essentially with no locks set at all. File
400 * server may thus handle operations from others cache managers as well as from
401 * this very own cache manager that reference the file in question before
402 * we managed to create the cache entry. However, if anyone else changes
403 * any of the status information for a file, we'll see afs_evenCBs increase
404 * (files always have even fids). If someone on this workstation manages
405 * to do something to the file, they'll end up having to create a cache
406 * entry for the new file. Either we'll find it once we've got the afs_xvcache
407 * lock set, or it was also *deleted* the vnode before we got there, in which case
408 * we will find evenZaps has changed, too. Thus, we only assume we have the right
409 * status information if no callbacks or vnode removals have occurred to even
410 * numbered files from the time the call started until the time that we got the xvcache
411 * lock set. Of course, this also assumes that any call that modifies a file first
412 * gets a write lock on the file's vnode, but if that weren't true, the whole cache manager
413 * would fail, since no call would be able to update the local vnode status after modifying
414 * a file on a file server. */
415 ObtainWriteLock(&afs_xvcache, 138);
416 if (adp->f.states & CForeign)
417 finalZaps = afs_allZaps; /* do this before calling newvcache */
419 finalZaps = afs_evenZaps; /* do this before calling newvcache */
420 /* don't need to call RemoveVCB, since only path leaving a callback is the
421 * one where we pass through afs_NewVCache. Can't have queued a VCB unless
422 * we created and freed an entry between file creation time and here, and the
423 * freeing of the vnode will change evenZaps. Don't need to update the VLRU
424 * queue, since the find will only succeed in the event of a create race, and
425 * then the vcache will be at the front of the VLRU queue anyway... */
426 if (!(tvc = afs_FindVCache(&newFid, 0, DO_STATS))) {
427 tvc = afs_NewVCache(&newFid, hostp);
430 ObtainWriteLock(&tvc->lock, 139);
432 ObtainWriteLock(&afs_xcbhash, 489);
433 finalCBs = afs_evenCBs;
434 /* add the callback in */
435 if (adp->f.states & CForeign) {
436 tvc->f.states |= CForeign;
437 finalCBs = afs_allCBs;
439 if (origCBs == finalCBs && origZaps == finalZaps) {
440 tvc->f.states |= CStatd; /* we've fake entire thing, so don't stat */
441 tvc->f.states &= ~CBulkFetching;
442 if (!AFS_IS_DISCON_RW) {
443 tvc->cbExpires = CallBack.ExpirationTime;
444 afs_QueueCallback(tvc, CBHash(CallBack.ExpirationTime), volp);
447 afs_DequeueCallback(tvc);
448 tvc->f.states &= ~(CStatd | CUnique);
450 if (tvc->f.fid.Fid.Vnode & 1 || (vType(tvc) == VDIR))
451 osi_dnlc_purgedp(tvc);
453 ReleaseWriteLock(&afs_xcbhash);
454 if (AFS_IS_DISCON_RW) {
455 #if defined(AFS_DISCON_ENV)
456 afs_DisconAddDirty(tvc, VDisconCreate, 0);
457 afs_GenDisconStatus(adp, tvc, &newFid, attrs, &treq, VREG);
460 afs_ProcessFS(tvc, &OutFidStatus, &treq);
463 tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
464 tvc->f.parent.unique = adp->f.fid.Fid.Unique;
465 ReleaseWriteLock(&tvc->lock);
471 /* otherwise cache entry already exists, someone else must
472 * have created it. Comments used to say: "don't need write
473 * lock to *clear* these flags" but we should do it anyway.
474 * Code used to clear stat bit and callback, but I don't see
475 * the point -- we didn't have a create race, somebody else just
476 * snuck into NewVCache before we got here, probably a racing
482 ReleaseWriteLock(&afs_xvcache);
489 afs_PutVolume(volp, READ_LOCK);
492 afs_AddMarinerName(aname, *avcp);
493 /* return the new status in vattr */
494 afs_CopyOutAttrs(*avcp, attrs);
497 afs_PutFakeStat(&fakestate);
498 code = afs_CheckCode(code, &treq, 20);
506 * Check to see if we can track the change locally: requires that
507 * we have sufficiently recent info in data cache. If so, we
508 * know the new DataVersion number, and place it correctly in both the
509 * data and stat cache entries. This routine returns 1 if we should
510 * do the operation locally, and 0 otherwise.
512 * This routine must be called with the stat cache entry write-locked,
513 * and dcache entry write-locked.
516 afs_LocalHero(register struct vcache *avc, register struct dcache *adc,
517 register AFSFetchStatus * astat, register int aincr)
519 register afs_int32 ok;
522 AFS_STATCNT(afs_LocalHero);
523 hset64(avers, astat->dataVersionHigh, astat->DataVersion);
524 /* this *is* the version number, no matter what */
526 ok = (hsame(avc->f.m.DataVersion, adc->f.versionNo) && avc->callback
527 && (avc->f.states & CStatd) && avc->cbExpires >= osi_Time());
531 #if defined(AFS_SGI_ENV)
532 osi_Assert(avc->v.v_type == VDIR);
534 /* The bulk status code used the length as a sequence number. */
535 /* Don't update the vcache entry unless the stats are current. */
536 if (avc->f.states & CStatd) {
537 hset(avc->f.m.DataVersion, avers);
538 #ifdef AFS_64BIT_CLIENT
539 FillInt64(avc->f.m.Length, astat->Length_hi, astat->Length);
540 #else /* AFS_64BIT_CLIENT */
541 avc->f.m.Length = astat->Length;
542 #endif /* AFS_64BIT_CLIENT */
543 avc->f.m.Date = astat->ClientModTime;
546 /* we've been tracking things correctly */
547 adc->dflags |= DFEntryMod;
548 adc->f.versionNo = avers;
555 if (avc->f.states & CStatd) {
556 osi_dnlc_purgedp(avc);