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;
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;
59 struct rx_connection *rxconn;
64 AFS_STATCNT(afs_create);
65 if ((code = afs_InitReq(&treq, acred)))
68 afs_Trace3(afs_iclSetp, CM_TRACE_CREATE, ICL_TYPE_POINTER, adp,
69 ICL_TYPE_STRING, aname, ICL_TYPE_INT32, amode);
71 afs_InitFakeStat(&fakestate);
74 /* If avcp is passed not null, it's the old reference to this file.
75 * We can use this to avoid create races. For now, just decrement
76 * the reference count on it.
79 AFS_RELE(AFSTOV(*avcp));
84 if (strlen(aname) > AFSNAMEMAX) {
89 if (!afs_ENameOK(aname)) {
93 switch (attrs->va_type) {
96 #if !defined(AFS_SUN5_ENV)
100 /* We don't support special devices or FIFOs */
108 code = afs_EvalFakeStat(&adp, &fakestate, &treq);
112 code = afs_VerifyVCache(adp, &treq);
116 /** If the volume is read-only, return error without making an RPC to the
119 if (adp->f.states & CRO) {
124 if (AFS_IS_DISCONNECTED && !AFS_IS_DISCON_RW) {
129 tdc = afs_GetDCache(adp, (afs_size_t) 0, &treq, &offset, &len, 1);
130 ObtainWriteLock(&adp->lock, 135);
132 ObtainSharedLock(&tdc->lock, 630);
135 * Make sure that the data in the cache is current. We may have
136 * received a callback while we were waiting for the write lock.
138 if (!(adp->f.states & CStatd)
139 || (tdc && !hsame(adp->f.m.DataVersion, tdc->f.versionNo))) {
140 ReleaseWriteLock(&adp->lock);
142 ReleaseSharedLock(&tdc->lock);
148 /* see if file already exists. If it does, we only set
149 * the size attributes (to handle O_TRUNC) */
150 code = afs_dir_Lookup(tdc, aname, &newFid.Fid); /* use dnlc first xxx */
152 ReleaseSharedLock(&tdc->lock);
154 ReleaseWriteLock(&adp->lock);
158 if (aexcl != NONEXCL) {
160 code = EEXIST; /* file exists in excl mode open */
163 /* found the file, so use it */
164 newFid.Cell = adp->f.fid.Cell;
165 newFid.Fid.Volume = adp->f.fid.Fid.Volume;
167 if (newFid.Fid.Unique == 0) {
168 tvc = afs_LookupVCache(&newFid, &treq, NULL, adp, aname);
170 if (!tvc) /* lookup failed or wasn't called */
171 tvc = afs_GetVCache(&newFid, &treq, NULL, NULL);
174 /* if the thing exists, we need the right access to open it.
175 * we must check that here, since no other checks are
176 * made by the open system call */
177 len = attrs->va_size; /* only do the truncate */
179 * We used to check always for READ access before; the
180 * problem is that we will fail if the existing file
181 * has mode -w-w-w, which is wrong.
184 && !afs_AccessOK(tvc, PRSFS_READ, &treq, CHECK_MODE_BITS)) {
189 #if defined(AFS_DARWIN80_ENV)
190 if ((amode & VWRITE) || VATTR_IS_ACTIVE(attrs, va_data_size))
191 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
192 if ((amode & VWRITE) || (attrs->va_mask & AT_SIZE))
194 if ((amode & VWRITE) || len != 0xffffffff)
197 /* needed for write access check */
198 tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
199 tvc->f.parent.unique = adp->f.fid.Fid.Unique;
200 /* need write mode for these guys */
202 (tvc, PRSFS_WRITE, &treq, CHECK_MODE_BITS)) {
208 #if defined(AFS_DARWIN80_ENV)
209 if (VATTR_IS_ACTIVE(attrs, va_data_size))
210 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
211 if (attrs->va_mask & AT_SIZE)
213 if (len != 0xffffffff)
216 if (vType(tvc) != VREG) {
222 #if defined(AFS_DARWIN80_ENV)
224 VATTR_SET_SUPPORTED(attrs, va_data_size);
225 VATTR_SET_ACTIVE(attrs, va_data_size);
226 #elif defined(UKERNEL)
227 attrs->va_mask = ATTR_SIZE;
228 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
229 attrs->va_mask = AT_SIZE;
233 attrs->va_size = len;
234 ObtainWriteLock(&tvc->lock, 136);
235 tvc->f.states |= CCreating;
236 ReleaseWriteLock(&tvc->lock);
237 #if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
238 #if defined(AFS_SGI64_ENV)
240 afs_setattr(VNODE_TO_FIRST_BHV((vnode_t *) tvc),
243 code = afs_setattr(tvc, attrs, 0, acred);
244 #endif /* AFS_SGI64_ENV */
245 #else /* SUN5 || SGI */
246 code = afs_setattr(tvc, attrs, acred);
247 #endif /* SUN5 || SGI */
248 ObtainWriteLock(&tvc->lock, 137);
249 tvc->f.states &= ~CCreating;
250 ReleaseWriteLock(&tvc->lock);
258 code = ENOENT; /* shouldn't get here */
259 /* make sure vrefCount bumped only if code == 0 */
264 /* if we create the file, we don't do any access checks, since
265 * that's how O_CREAT is supposed to work */
266 if (adp->f.states & CForeign) {
267 origCBs = afs_allCBs;
268 origZaps = afs_allZaps;
270 origCBs = afs_evenCBs; /* if changes, we don't really have a callback */
271 origZaps = afs_evenZaps; /* number of even numbered vnodes discarded */
273 InStatus.Mask = AFS_SETMODTIME | AFS_SETMODE | AFS_SETGROUP;
274 InStatus.ClientModTime = osi_Time();
275 InStatus.Group = (afs_int32) afs_cr_gid(acred);
276 if (AFS_NFSXLATORREQ(acred)) {
278 * XXX The following is mainly used to fix a bug in the HP-UX
279 * nfs client where they create files with mode of 0 without
280 * doing any setattr later on to fix it. * XXX
282 #if defined(AFS_AIX_ENV)
283 if (attrs->va_mode != -1) {
285 #if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
286 if (attrs->va_mask & AT_MODE) {
288 if (attrs->va_mode != ((unsigned short)-1)) {
292 attrs->va_mode = 0x1b6; /* XXX default mode: rw-rw-rw XXX */
296 if (!AFS_IS_DISCONNECTED) {
297 /* If not disconnected, connect to the server.*/
299 InStatus.UnixModeBits = attrs->va_mode & 0xffff; /* only care about protection bits */
301 tc = afs_Conn(&adp->f.fid, &treq, SHARED_LOCK, &rxconn);
303 hostp = tc->parent->srvr->server; /* remember for callback processing */
305 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_CREATEFILE);
308 RXAFS_CreateFile(rxconn, (struct AFSFid *)&adp->f.fid.Fid,
309 aname, &InStatus, (struct AFSFid *)
310 &newFid.Fid, &OutFidStatus, &OutDirStatus,
314 CallBack.ExpirationTime += now;
318 (tc, rxconn, code, &adp->f.fid, &treq, AFS_STATS_FS_RPCIDX_CREATEFILE,
321 if ((code == EEXIST || code == UAEEXIST) &&
324 #else /* AFS_SGI64_ENV */
328 /* if we get an EEXIST in nonexcl mode, just do a lookup */
330 ReleaseSharedLock(&tdc->lock);
333 ReleaseWriteLock(&adp->lock);
336 #if defined(AFS_SGI64_ENV)
337 code = afs_lookup(VNODE_TO_FIRST_BHV((vnode_t *) adp), aname, avcp,
338 NULL, 0, NULL, acred);
339 #elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
340 code = afs_lookup(adp, aname, avcp, NULL, 0, NULL, acred);
341 #elif defined(UKERNEL)
342 code = afs_lookup(adp, aname, avcp, acred, 0);
343 #elif !defined(AFS_DARWIN_ENV)
344 code = afs_lookup(adp, aname, avcp, acred);
351 ObtainWriteLock(&afs_xcbhash, 488);
352 afs_DequeueCallback(adp);
353 adp->f.states &= ~CStatd;
354 ReleaseWriteLock(&afs_xcbhash);
355 osi_dnlc_purgedp(adp);
357 ReleaseWriteLock(&adp->lock);
359 ReleaseSharedLock(&tdc->lock);
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);
370 } /* if (!AFS_IS_DISCON_RW) */
372 /* otherwise, we should see if we can make the change to the dir locally */
374 UpgradeSToWLock(&tdc->lock, 631);
375 if (AFS_IS_DISCON_RW || afs_LocalHero(adp, tdc, &OutDirStatus, 1)) {
376 /* we can do it locally */
377 ObtainWriteLock(&afs_xdcache, 291);
378 code = afs_dir_Create(tdc, aname, &newFid.Fid);
379 ReleaseWriteLock(&afs_xdcache);
386 ReleaseWriteLock(&tdc->lock);
389 if (AFS_IS_DISCON_RW)
390 adp->f.m.LinkCount++;
392 newFid.Cell = adp->f.fid.Cell;
393 newFid.Fid.Volume = adp->f.fid.Fid.Volume;
394 ReleaseWriteLock(&adp->lock);
395 volp = afs_FindVolume(&newFid, READ_LOCK);
397 /* New tricky optimistic callback handling algorithm for file creation works
398 * as follows. We create the file essentially with no locks set at all. File
399 * server may thus handle operations from others cache managers as well as from
400 * this very own cache manager that reference the file in question before
401 * we managed to create the cache entry. However, if anyone else changes
402 * any of the status information for a file, we'll see afs_evenCBs increase
403 * (files always have even fids). If someone on this workstation manages
404 * to do something to the file, they'll end up having to create a cache
405 * entry for the new file. Either we'll find it once we've got the afs_xvcache
406 * lock set, or it was also *deleted* the vnode before we got there, in which case
407 * we will find evenZaps has changed, too. Thus, we only assume we have the right
408 * status information if no callbacks or vnode removals have occurred to even
409 * numbered files from the time the call started until the time that we got the xvcache
410 * lock set. Of course, this also assumes that any call that modifies a file first
411 * gets a write lock on the file's vnode, but if that weren't true, the whole cache manager
412 * would fail, since no call would be able to update the local vnode status after modifying
413 * a file on a file server. */
414 ObtainWriteLock(&afs_xvcache, 138);
415 if (adp->f.states & CForeign)
416 finalZaps = afs_allZaps; /* do this before calling newvcache */
418 finalZaps = afs_evenZaps; /* do this before calling newvcache */
419 /* don't need to call RemoveVCB, since only path leaving a callback is the
420 * one where we pass through afs_NewVCache. Can't have queued a VCB unless
421 * we created and freed an entry between file creation time and here, and the
422 * freeing of the vnode will change evenZaps. Don't need to update the VLRU
423 * queue, since the find will only succeed in the event of a create race, and
424 * then the vcache will be at the front of the VLRU queue anyway... */
425 if (!(tvc = afs_FindVCache(&newFid, 0, DO_STATS))) {
426 tvc = afs_NewVCache(&newFid, hostp);
429 ObtainWriteLock(&tvc->lock, 139);
431 ObtainWriteLock(&afs_xcbhash, 489);
432 finalCBs = afs_evenCBs;
433 /* add the callback in */
434 if (adp->f.states & CForeign) {
435 tvc->f.states |= CForeign;
436 finalCBs = afs_allCBs;
438 if (origCBs == finalCBs && origZaps == finalZaps) {
439 tvc->f.states |= CStatd; /* we've fake entire thing, so don't stat */
440 tvc->f.states &= ~CBulkFetching;
441 if (!AFS_IS_DISCON_RW) {
442 tvc->cbExpires = CallBack.ExpirationTime;
443 afs_QueueCallback(tvc, CBHash(CallBack.ExpirationTime), volp);
446 afs_DequeueCallback(tvc);
447 tvc->f.states &= ~(CStatd | CUnique);
449 if (tvc->f.fid.Fid.Vnode & 1 || (vType(tvc) == VDIR))
450 osi_dnlc_purgedp(tvc);
452 ReleaseWriteLock(&afs_xcbhash);
453 if (AFS_IS_DISCON_RW) {
454 afs_DisconAddDirty(tvc, VDisconCreate, 0);
455 afs_GenDisconStatus(adp, tvc, &newFid, attrs, &treq, VREG);
457 afs_ProcessFS(tvc, &OutFidStatus, &treq);
460 tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
461 tvc->f.parent.unique = adp->f.fid.Fid.Unique;
462 ReleaseWriteLock(&tvc->lock);
468 /* otherwise cache entry already exists, someone else must
469 * have created it. Comments used to say: "don't need write
470 * lock to *clear* these flags" but we should do it anyway.
471 * Code used to clear stat bit and callback, but I don't see
472 * the point -- we didn't have a create race, somebody else just
473 * snuck into NewVCache before we got here, probably a racing
479 ReleaseWriteLock(&afs_xvcache);
486 afs_PutVolume(volp, READ_LOCK);
489 afs_AddMarinerName(aname, *avcp);
490 /* return the new status in vattr */
491 afs_CopyOutAttrs(*avcp, attrs);
492 afs_MarinerLog("store$Creating", *avcp);
495 afs_PutFakeStat(&fakestate);
496 code = afs_CheckCode(code, &treq, 20);
504 * Check to see if we can track the change locally: requires that
505 * we have sufficiently recent info in data cache. If so, we
506 * know the new DataVersion number, and place it correctly in both the
507 * data and stat cache entries. This routine returns 1 if we should
508 * do the operation locally, and 0 otherwise.
510 * This routine must be called with the stat cache entry write-locked,
511 * and dcache entry write-locked.
514 afs_LocalHero(struct vcache *avc, struct dcache *adc,
515 AFSFetchStatus * astat, int aincr)
520 AFS_STATCNT(afs_LocalHero);
521 hset64(avers, astat->dataVersionHigh, astat->DataVersion);
522 /* this *is* the version number, no matter what */
524 ok = (hsame(avc->f.m.DataVersion, adc->f.versionNo) && avc->callback
525 && (avc->f.states & CStatd) && avc->cbExpires >= osi_Time());
529 #if defined(AFS_SGI_ENV)
530 osi_Assert(avc->v.v_type == VDIR);
532 /* The bulk status code used the length as a sequence number. */
533 /* Don't update the vcache entry unless the stats are current. */
534 if (avc->f.states & CStatd) {
535 hset(avc->f.m.DataVersion, avers);
536 #ifdef AFS_64BIT_CLIENT
537 FillInt64(avc->f.m.Length, astat->Length_hi, astat->Length);
538 #else /* AFS_64BIT_CLIENT */
539 avc->f.m.Length = astat->Length;
540 #endif /* AFS_64BIT_CLIENT */
541 avc->f.m.Date = astat->ClientModTime;
544 /* we've been tracking things correctly */
545 adc->dflags |= DFEntryMod;
546 adc->f.versionNo = avers;
553 if (avc->f.states & CStatd) {
554 osi_dnlc_purgedp(avc);