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
11 * Linux specific vnodeops. Also includes the glue routines required to call
14 * So far the only truly scary part is that Linux relies on the inode cache
15 * to be up to date. Don't you dare break a callback and expect an fstat
16 * to give you meaningful information. This appears to be fixed in the 2.1
17 * development kernels. As it is we can fix this now by intercepting the
21 #include <afsconfig.h>
22 #include "afs/param.h"
25 #include "afs/sysincludes.h"
26 #include "afsincludes.h"
27 #include "afs/afs_stats.h"
29 #ifdef HAVE_MM_INLINE_H
30 #include <linux/mm_inline.h>
32 #include <linux/pagemap.h>
33 #include <linux/writeback.h>
34 #include <linux/pagevec.h>
35 #include <linux/aio.h>
37 #include "afs/afs_bypasscache.h"
39 #include "osi_compat.h"
40 #include "osi_pagecopy.h"
42 #ifndef HAVE_LINUX_PAGEVEC_LRU_ADD_FILE
43 #define __pagevec_lru_add_file __pagevec_lru_add
47 #define MAX_ERRNO 1000L
50 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,34)
51 /* Enable our workaround for a race with d_splice_alias. The race was fixed in
52 * 2.6.34, so don't do it after that point. */
53 # define D_SPLICE_ALIAS_RACE
56 /* Workaround for RH 7.5 which introduced file operation iterate() but requires
57 * each file->f_mode to be marked with FMODE_KABI_ITERATE. Instead OpenAFS will
58 * continue to use file opearation readdir() in this case.
60 #if defined(STRUCT_FILE_OPERATIONS_HAS_ITERATE) && !defined(FMODE_KABI_ITERATE)
61 #define USE_FOP_ITERATE 1
63 #undef USE_FOP_ITERATE
66 int cachefs_noreadpage = 0;
68 extern struct backing_dev_info *afs_backing_dev_info;
70 extern struct vcache *afs_globalVp;
72 /* This function converts a positive error code from AFS into a negative
73 * code suitable for passing into the Linux VFS layer. It checks that the
74 * error code is within the permissable bounds for the ERR_PTR mechanism.
76 * _All_ error codes which come from the AFS layer should be passed through
77 * this function before being returned to the kernel.
81 afs_convert_code(int code) {
82 if ((code >= 0) && (code <= MAX_ERRNO))
88 /* Linux doesn't require a credp for many functions, and crref is an expensive
89 * operation. This helper function avoids obtaining it for VerifyVCache calls
93 afs_linux_VerifyVCache(struct vcache *avc, cred_t **retcred) {
95 struct vrequest *treq = NULL;
98 if (avc->f.states & CStatd) {
106 code = afs_CreateReq(&treq, credp);
108 code = afs_VerifyVCache2(avc, treq);
109 afs_DestroyReq(treq);
117 return afs_convert_code(code);
120 #if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER) || defined(HAVE_LINUX_GENERIC_FILE_AIO_READ)
121 # if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER)
123 afs_linux_read_iter(struct kiocb *iocb, struct iov_iter *iter)
124 # elif defined(LINUX_HAS_NONVECTOR_AIO)
126 afs_linux_aio_read(struct kiocb *iocb, char __user *buf, size_t bufsize,
130 afs_linux_aio_read(struct kiocb *iocb, const struct iovec *buf,
131 unsigned long bufsize, loff_t pos)
134 struct file *fp = iocb->ki_filp;
136 struct vcache *vcp = VTOAFS(fp->f_dentry->d_inode);
137 # if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER)
138 loff_t pos = iocb->ki_pos;
139 unsigned long bufsize = iter->nr_segs;
144 afs_Trace4(afs_iclSetp, CM_TRACE_AIOREADOP, ICL_TYPE_POINTER, vcp,
145 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(pos), ICL_TYPE_INT32,
146 (afs_int32)bufsize, ICL_TYPE_INT32, 99999);
147 code = afs_linux_VerifyVCache(vcp, NULL);
150 /* Linux's FlushPages implementation doesn't ever use credp,
151 * so we optimise by not using it */
152 osi_FlushPages(vcp, NULL); /* ensure stale pages are gone */
154 # if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER)
155 code = generic_file_read_iter(iocb, iter);
157 code = generic_file_aio_read(iocb, buf, bufsize, pos);
162 afs_Trace4(afs_iclSetp, CM_TRACE_AIOREADOP, ICL_TYPE_POINTER, vcp,
163 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(pos), ICL_TYPE_INT32,
164 (afs_int32)bufsize, ICL_TYPE_INT32, code);
170 afs_linux_read(struct file *fp, char *buf, size_t count, loff_t * offp)
173 struct vcache *vcp = VTOAFS(fp->f_dentry->d_inode);
176 afs_Trace4(afs_iclSetp, CM_TRACE_READOP, ICL_TYPE_POINTER, vcp,
177 ICL_TYPE_OFFSET, offp, ICL_TYPE_INT32, count, ICL_TYPE_INT32,
179 code = afs_linux_VerifyVCache(vcp, NULL);
182 /* Linux's FlushPages implementation doesn't ever use credp,
183 * so we optimise by not using it */
184 osi_FlushPages(vcp, NULL); /* ensure stale pages are gone */
186 code = do_sync_read(fp, buf, count, offp);
190 afs_Trace4(afs_iclSetp, CM_TRACE_READOP, ICL_TYPE_POINTER, vcp,
191 ICL_TYPE_OFFSET, offp, ICL_TYPE_INT32, count, ICL_TYPE_INT32,
199 /* Now we have integrated VM for writes as well as reads. the generic write operations
200 * also take care of re-positioning the pointer if file is open in append
201 * mode. Call fake open/close to ensure we do writes of core dumps.
203 #if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER) || defined(HAVE_LINUX_GENERIC_FILE_AIO_READ)
204 # if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER)
206 afs_linux_write_iter(struct kiocb *iocb, struct iov_iter *iter)
207 # elif defined(LINUX_HAS_NONVECTOR_AIO)
209 afs_linux_aio_write(struct kiocb *iocb, const char __user *buf, size_t bufsize,
213 afs_linux_aio_write(struct kiocb *iocb, const struct iovec *buf,
214 unsigned long bufsize, loff_t pos)
218 struct vcache *vcp = VTOAFS(iocb->ki_filp->f_dentry->d_inode);
220 # if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER)
221 loff_t pos = iocb->ki_pos;
222 unsigned long bufsize = iter->nr_segs;
227 afs_Trace4(afs_iclSetp, CM_TRACE_AIOWRITEOP, ICL_TYPE_POINTER, vcp,
228 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(pos), ICL_TYPE_INT32,
229 (afs_int32)bufsize, ICL_TYPE_INT32,
230 (iocb->ki_filp->f_flags & O_APPEND) ? 99998 : 99999);
232 code = afs_linux_VerifyVCache(vcp, &credp);
234 ObtainWriteLock(&vcp->lock, 529);
236 ReleaseWriteLock(&vcp->lock);
239 # if defined(STRUCT_FILE_OPERATIONS_HAS_READ_ITER)
240 code = generic_file_write_iter(iocb, iter);
242 code = generic_file_aio_write(iocb, buf, bufsize, pos);
247 ObtainWriteLock(&vcp->lock, 530);
249 if (vcp->execsOrWriters == 1 && !credp)
252 afs_FakeClose(vcp, credp);
253 ReleaseWriteLock(&vcp->lock);
255 afs_Trace4(afs_iclSetp, CM_TRACE_AIOWRITEOP, ICL_TYPE_POINTER, vcp,
256 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(pos), ICL_TYPE_INT32,
257 (afs_int32)bufsize, ICL_TYPE_INT32, code);
266 afs_linux_write(struct file *fp, const char *buf, size_t count, loff_t * offp)
269 struct vcache *vcp = VTOAFS(fp->f_dentry->d_inode);
274 afs_Trace4(afs_iclSetp, CM_TRACE_WRITEOP, ICL_TYPE_POINTER, vcp,
275 ICL_TYPE_OFFSET, offp, ICL_TYPE_INT32, count, ICL_TYPE_INT32,
276 (fp->f_flags & O_APPEND) ? 99998 : 99999);
278 code = afs_linux_VerifyVCache(vcp, &credp);
280 ObtainWriteLock(&vcp->lock, 529);
282 ReleaseWriteLock(&vcp->lock);
285 code = do_sync_write(fp, buf, count, offp);
289 ObtainWriteLock(&vcp->lock, 530);
291 if (vcp->execsOrWriters == 1 && !credp)
294 afs_FakeClose(vcp, credp);
295 ReleaseWriteLock(&vcp->lock);
297 afs_Trace4(afs_iclSetp, CM_TRACE_WRITEOP, ICL_TYPE_POINTER, vcp,
298 ICL_TYPE_OFFSET, offp, ICL_TYPE_INT32, count, ICL_TYPE_INT32,
308 extern int BlobScan(struct dcache * afile, afs_int32 ablob, afs_int32 *ablobOut);
310 /* This is a complete rewrite of afs_readdir, since we can make use of
311 * filldir instead of afs_readdir_move. Note that changes to vcache/dcache
312 * handling and use of bulkstats will need to be reflected here as well.
315 #if defined(USE_FOP_ITERATE)
316 afs_linux_readdir(struct file *fp, struct dir_context *ctx)
318 afs_linux_readdir(struct file *fp, void *dirbuf, filldir_t filldir)
321 struct vcache *avc = VTOAFS(FILE_INODE(fp));
322 struct vrequest *treq = NULL;
328 struct DirBuffer entry;
331 afs_size_t origOffset, tlen;
332 cred_t *credp = crref();
333 struct afs_fakestat_state fakestat;
336 AFS_STATCNT(afs_readdir);
338 code = afs_convert_code(afs_CreateReq(&treq, credp));
343 afs_InitFakeStat(&fakestat);
344 code = afs_convert_code(afs_EvalFakeStat(&avc, &fakestat, treq));
348 /* update the cache entry */
350 code = afs_convert_code(afs_VerifyVCache2(avc, treq));
354 /* get a reference to the entire directory */
355 tdc = afs_GetDCache(avc, (afs_size_t) 0, treq, &origOffset, &tlen, 1);
361 ObtainWriteLock(&avc->lock, 811);
362 ObtainReadLock(&tdc->lock);
364 * Make sure that the data in the cache is current. There are two
365 * cases we need to worry about:
366 * 1. The cache data is being fetched by another process.
367 * 2. The cache data is no longer valid
369 while ((avc->f.states & CStatd)
370 && (tdc->dflags & DFFetching)
371 && hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
372 ReleaseReadLock(&tdc->lock);
373 ReleaseWriteLock(&avc->lock);
374 afs_osi_Sleep(&tdc->validPos);
375 ObtainWriteLock(&avc->lock, 812);
376 ObtainReadLock(&tdc->lock);
378 if (!(avc->f.states & CStatd)
379 || !hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
380 ReleaseReadLock(&tdc->lock);
381 ReleaseWriteLock(&avc->lock);
386 /* Set the readdir-in-progress flag, and downgrade the lock
387 * to shared so others will be able to acquire a read lock.
389 avc->f.states |= CReadDir;
390 avc->dcreaddir = tdc;
391 avc->readdir_pid = MyPidxx2Pid(MyPidxx);
392 ConvertWToSLock(&avc->lock);
394 /* Fill in until we get an error or we're done. This implementation
395 * takes an offset in units of blobs, rather than bytes.
398 #if defined(USE_FOP_ITERATE)
401 offset = (int) fp->f_pos;
405 code = BlobScan(tdc, offset, &dirpos);
406 if (code == 0 && dirpos == 0) {
407 /* We've reached EOF of the dir blob, so we can stop looking for
413 code = afs_dir_GetVerifiedBlob(tdc, dirpos, &entry);
416 if (!(avc->f.states & CCorrupt)) {
417 struct cell *tc = afs_GetCellStale(avc->f.fid.Cell, READ_LOCK);
418 afs_warn("afs: Corrupt directory (%d.%d.%d.%d [%s] @%lx, pos %d)\n",
419 avc->f.fid.Cell, avc->f.fid.Fid.Volume,
420 avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique,
421 tc ? tc->cellName : "",
422 (unsigned long)&tdc->f.inode, dirpos);
424 afs_PutCell(tc, READ_LOCK);
425 UpgradeSToWLock(&avc->lock, 814);
426 avc->f.states |= CCorrupt;
432 de = (struct DirEntry *)entry.data;
433 ino = afs_calc_inum (avc->f.fid.Cell, avc->f.fid.Fid.Volume,
434 ntohl(de->fid.vnode));
435 len = strlen(de->name);
437 /* filldir returns -EINVAL when the buffer is full. */
439 unsigned int type = DT_UNKNOWN;
440 struct VenusFid afid;
443 afid.Cell = avc->f.fid.Cell;
444 afid.Fid.Volume = avc->f.fid.Fid.Volume;
445 afid.Fid.Vnode = ntohl(de->fid.vnode);
446 afid.Fid.Unique = ntohl(de->fid.vunique);
447 if ((avc->f.states & CForeign) == 0 && (ntohl(de->fid.vnode) & 1)) {
449 } else if ((tvc = afs_FindVCache(&afid, 0, 0))) {
450 if (tvc->mvstat != AFS_MVSTAT_FILE) {
452 } else if (((tvc->f.states) & (CStatd | CTruth))) {
453 /* CTruth will be set if the object has
458 else if (vtype == VREG)
460 /* Don't do this until we're sure it can't be a mtpt */
461 /* else if (vtype == VLNK)
463 /* what other types does AFS support? */
465 /* clean up from afs_FindVCache */
469 * If this is NFS readdirplus, then the filler is going to
470 * call getattr on this inode, which will deadlock if we're
474 #if defined(USE_FOP_ITERATE)
475 /* dir_emit returns a bool - true when it succeeds.
476 * Inverse the result to fit with how we check "code" */
477 code = !dir_emit(ctx, de->name, len, ino, type);
479 code = (*filldir) (dirbuf, de->name, len, offset, ino, type);
486 offset = dirpos + 1 + ((len + 16) >> 5);
488 /* If filldir didn't fill in the last one this is still pointing to that
494 #if defined(USE_FOP_ITERATE)
495 ctx->pos = (loff_t) offset;
497 fp->f_pos = (loff_t) offset;
499 ReleaseReadLock(&tdc->lock);
501 UpgradeSToWLock(&avc->lock, 813);
502 avc->f.states &= ~CReadDir;
504 avc->readdir_pid = 0;
505 ReleaseSharedLock(&avc->lock);
508 afs_PutFakeStat(&fakestat);
509 afs_DestroyReq(treq);
516 /* in afs_pioctl.c */
517 extern int afs_xioctl(struct inode *ip, struct file *fp, unsigned int com,
520 #if defined(HAVE_UNLOCKED_IOCTL) || defined(HAVE_COMPAT_IOCTL)
521 static long afs_unlocked_xioctl(struct file *fp, unsigned int com,
523 return afs_xioctl(FILE_INODE(fp), fp, com, arg);
530 afs_linux_mmap(struct file *fp, struct vm_area_struct *vmap)
532 struct vcache *vcp = VTOAFS(FILE_INODE(fp));
536 afs_Trace3(afs_iclSetp, CM_TRACE_GMAP, ICL_TYPE_POINTER, vcp,
537 ICL_TYPE_POINTER, vmap->vm_start, ICL_TYPE_INT32,
538 vmap->vm_end - vmap->vm_start);
540 /* get a validated vcache entry */
541 code = afs_linux_VerifyVCache(vcp, NULL);
544 /* Linux's Flushpage implementation doesn't use credp, so optimise
545 * our code to not need to crref() it */
546 osi_FlushPages(vcp, NULL); /* ensure stale pages are gone */
548 code = generic_file_mmap(fp, vmap);
551 vcp->f.states |= CMAPPED;
559 afs_linux_open(struct inode *ip, struct file *fp)
561 struct vcache *vcp = VTOAFS(ip);
562 cred_t *credp = crref();
566 code = afs_open(&vcp, fp->f_flags, credp);
570 return afs_convert_code(code);
574 afs_linux_release(struct inode *ip, struct file *fp)
576 struct vcache *vcp = VTOAFS(ip);
577 cred_t *credp = crref();
581 code = afs_close(vcp, fp->f_flags, credp);
582 ObtainWriteLock(&vcp->lock, 807);
587 ReleaseWriteLock(&vcp->lock);
591 return afs_convert_code(code);
595 #if defined(FOP_FSYNC_TAKES_DENTRY)
596 afs_linux_fsync(struct file *fp, struct dentry *dp, int datasync)
597 #elif defined(FOP_FSYNC_TAKES_RANGE)
598 afs_linux_fsync(struct file *fp, loff_t start, loff_t end, int datasync)
600 afs_linux_fsync(struct file *fp, int datasync)
604 struct inode *ip = FILE_INODE(fp);
605 cred_t *credp = crref();
607 #if defined(FOP_FSYNC_TAKES_RANGE)
608 afs_linux_lock_inode(ip);
611 code = afs_fsync(VTOAFS(ip), credp);
613 #if defined(FOP_FSYNC_TAKES_RANGE)
614 afs_linux_unlock_inode(ip);
617 return afs_convert_code(code);
623 afs_linux_lock(struct file *fp, int cmd, struct file_lock *flp)
626 struct vcache *vcp = VTOAFS(FILE_INODE(fp));
627 cred_t *credp = crref();
628 struct AFS_FLOCK flock;
630 /* Convert to a lock format afs_lockctl understands. */
631 memset(&flock, 0, sizeof(flock));
632 flock.l_type = flp->fl_type;
633 flock.l_pid = flp->fl_pid;
635 flock.l_start = flp->fl_start;
636 if (flp->fl_end == OFFSET_MAX)
637 flock.l_len = 0; /* Lock to end of file */
639 flock.l_len = flp->fl_end - flp->fl_start + 1;
641 /* Safe because there are no large files, yet */
642 #if defined(F_GETLK64) && (F_GETLK != F_GETLK64)
643 if (cmd == F_GETLK64)
645 else if (cmd == F_SETLK64)
647 else if (cmd == F_SETLKW64)
649 #endif /* F_GETLK64 && F_GETLK != F_GETLK64 */
652 code = afs_convert_code(afs_lockctl(vcp, &flock, cmd, credp));
655 if ((code == 0 || flp->fl_type == F_UNLCK) &&
656 (cmd == F_SETLK || cmd == F_SETLKW)) {
657 code = afs_posix_lock_file(fp, flp);
658 if (code && flp->fl_type != F_UNLCK) {
659 struct AFS_FLOCK flock2;
661 flock2.l_type = F_UNLCK;
663 afs_lockctl(vcp, &flock2, F_SETLK, credp);
667 /* If lockctl says there are no conflicting locks, then also check with the
668 * kernel, as lockctl knows nothing about byte range locks
670 if (code == 0 && cmd == F_GETLK && flock.l_type == F_UNLCK) {
671 afs_posix_test_lock(fp, flp);
672 /* If we found a lock in the kernel's structure, return it */
673 if (flp->fl_type != F_UNLCK) {
679 /* Convert flock back to Linux's file_lock */
680 flp->fl_type = flock.l_type;
681 flp->fl_pid = flock.l_pid;
682 flp->fl_start = flock.l_start;
683 if (flock.l_len == 0)
684 flp->fl_end = OFFSET_MAX; /* Lock to end of file */
686 flp->fl_end = flock.l_start + flock.l_len - 1;
692 #ifdef STRUCT_FILE_OPERATIONS_HAS_FLOCK
694 afs_linux_flock(struct file *fp, int cmd, struct file_lock *flp) {
696 struct vcache *vcp = VTOAFS(FILE_INODE(fp));
697 cred_t *credp = crref();
698 struct AFS_FLOCK flock;
699 /* Convert to a lock format afs_lockctl understands. */
700 memset(&flock, 0, sizeof(flock));
701 flock.l_type = flp->fl_type;
702 flock.l_pid = flp->fl_pid;
707 /* Safe because there are no large files, yet */
708 #if defined(F_GETLK64) && (F_GETLK != F_GETLK64)
709 if (cmd == F_GETLK64)
711 else if (cmd == F_SETLK64)
713 else if (cmd == F_SETLKW64)
715 #endif /* F_GETLK64 && F_GETLK != F_GETLK64 */
718 code = afs_convert_code(afs_lockctl(vcp, &flock, cmd, credp));
721 if ((code == 0 || flp->fl_type == F_UNLCK) &&
722 (cmd == F_SETLK || cmd == F_SETLKW)) {
723 flp->fl_flags &=~ FL_SLEEP;
724 code = flock_lock_file_wait(fp, flp);
725 if (code && flp->fl_type != F_UNLCK) {
726 struct AFS_FLOCK flock2;
728 flock2.l_type = F_UNLCK;
730 afs_lockctl(vcp, &flock2, F_SETLK, credp);
734 /* Convert flock back to Linux's file_lock */
735 flp->fl_type = flock.l_type;
736 flp->fl_pid = flock.l_pid;
744 * essentially the same as afs_fsync() but we need to get the return
745 * code for the sys_close() here, not afs_linux_release(), so call
746 * afs_StoreAllSegments() with AFS_LASTSTORE
749 #if defined(FOP_FLUSH_TAKES_FL_OWNER_T)
750 afs_linux_flush(struct file *fp, fl_owner_t id)
752 afs_linux_flush(struct file *fp)
755 struct vrequest *treq = NULL;
763 if ((fp->f_flags & O_ACCMODE) == O_RDONLY) { /* readers dont flush */
771 vcp = VTOAFS(FILE_INODE(fp));
773 code = afs_CreateReq(&treq, credp);
776 /* If caching is bypassed for this file, or globally, just return 0 */
777 if (cache_bypass_strategy == ALWAYS_BYPASS_CACHE)
780 ObtainReadLock(&vcp->lock);
781 if (vcp->cachingStates & FCSBypass)
783 ReleaseReadLock(&vcp->lock);
786 /* future proof: don't rely on 0 return from afs_InitReq */
791 ObtainSharedLock(&vcp->lock, 535);
792 if ((vcp->execsOrWriters > 0) && (file_count(fp) == 1)) {
793 UpgradeSToWLock(&vcp->lock, 536);
794 if (!AFS_IS_DISCONNECTED) {
795 code = afs_StoreAllSegments(vcp,
797 AFS_SYNC | AFS_LASTSTORE);
799 afs_DisconAddDirty(vcp, VDisconWriteOsiFlush, 1);
801 ConvertWToSLock(&vcp->lock);
803 code = afs_CheckCode(code, treq, 54);
804 ReleaseSharedLock(&vcp->lock);
807 afs_DestroyReq(treq);
812 return afs_convert_code(code);
815 struct file_operations afs_dir_fops = {
816 .read = generic_read_dir,
817 #if defined(USE_FOP_ITERATE)
818 .iterate = afs_linux_readdir,
820 .readdir = afs_linux_readdir,
822 #ifdef HAVE_UNLOCKED_IOCTL
823 .unlocked_ioctl = afs_unlocked_xioctl,
827 #ifdef HAVE_COMPAT_IOCTL
828 .compat_ioctl = afs_unlocked_xioctl,
830 .open = afs_linux_open,
831 .release = afs_linux_release,
832 .llseek = default_llseek,
833 #ifdef HAVE_LINUX_NOOP_FSYNC
836 .fsync = simple_sync_file,
840 struct file_operations afs_file_fops = {
841 #ifdef STRUCT_FILE_OPERATIONS_HAS_READ_ITER
842 .read_iter = afs_linux_read_iter,
843 .write_iter = afs_linux_write_iter,
844 # if !defined(HAVE_LINUX___VFS_WRITE) && !defined(HAVE_LINUX_KERNEL_WRITE)
845 .read = new_sync_read,
846 .write = new_sync_write,
848 #elif defined(HAVE_LINUX_GENERIC_FILE_AIO_READ)
849 .aio_read = afs_linux_aio_read,
850 .aio_write = afs_linux_aio_write,
851 .read = do_sync_read,
852 .write = do_sync_write,
854 .read = afs_linux_read,
855 .write = afs_linux_write,
857 #ifdef HAVE_UNLOCKED_IOCTL
858 .unlocked_ioctl = afs_unlocked_xioctl,
862 #ifdef HAVE_COMPAT_IOCTL
863 .compat_ioctl = afs_unlocked_xioctl,
865 .mmap = afs_linux_mmap,
866 .open = afs_linux_open,
867 .flush = afs_linux_flush,
868 #if defined(STRUCT_FILE_OPERATIONS_HAS_SENDFILE)
869 .sendfile = generic_file_sendfile,
871 #if defined(STRUCT_FILE_OPERATIONS_HAS_SPLICE) && !defined(HAVE_LINUX_DEFAULT_FILE_SPLICE_READ)
872 # if defined(HAVE_LINUX_ITER_FILE_SPLICE_WRITE)
873 .splice_write = iter_file_splice_write,
875 .splice_write = generic_file_splice_write,
877 .splice_read = generic_file_splice_read,
879 .release = afs_linux_release,
880 .fsync = afs_linux_fsync,
881 .lock = afs_linux_lock,
882 #ifdef STRUCT_FILE_OPERATIONS_HAS_FLOCK
883 .flock = afs_linux_flock,
885 .llseek = default_llseek,
888 static struct dentry *
889 canonical_dentry(struct inode *ip)
891 struct vcache *vcp = VTOAFS(ip);
892 struct dentry *first = NULL, *ret = NULL, *cur;
893 #if defined(D_ALIAS_IS_HLIST) && !defined(HLIST_ITERATOR_NO_NODE)
894 struct hlist_node *p;
898 * if vcp->target_link is set, and can be found in ip->i_dentry, use that.
899 * otherwise, use the first dentry in ip->i_dentry.
900 * if ip->i_dentry is empty, use the 'dentry' argument we were given.
902 /* note that vcp->target_link specifies which dentry to use, but we have
903 * no reference held on that dentry. so, we cannot use or dereference
904 * vcp->target_link itself, since it may have been freed. instead, we only
905 * use it to compare to pointers in the ip->i_dentry list. */
909 afs_d_alias_lock(ip);
911 #if defined(D_ALIAS_IS_HLIST)
912 # if defined(HLIST_ITERATOR_NO_NODE)
913 hlist_for_each_entry(cur, &ip->i_dentry, d_alias) {
915 hlist_for_each_entry(cur, p, &ip->i_dentry, d_alias) {
918 list_for_each_entry_reverse(cur, &ip->i_dentry, d_alias) {
921 if (!vcp->target_link || cur == vcp->target_link) {
934 vcp->target_link = ret;
939 afs_d_alias_unlock(ip);
944 /**********************************************************************
945 * AFS Linux dentry operations
946 **********************************************************************/
948 /* afs_linux_revalidate
949 * Ensure vcache is stat'd before use. Return 0 if entry is valid.
952 afs_linux_revalidate(struct dentry *dp)
954 struct vattr *vattr = NULL;
955 struct vcache *vcp = VTOAFS(dp->d_inode);
959 if (afs_shuttingdown != AFS_RUNNING)
964 code = afs_CreateAttr(&vattr);
969 /* This avoids the crref when we don't have to do it. Watch for
970 * changes in afs_getattr that don't get replicated here!
972 if (vcp->f.states & CStatd &&
973 (!afs_fakestat_enable || vcp->mvstat != AFS_MVSTAT_MTPT) &&
975 (vType(vcp) == VDIR || vType(vcp) == VLNK)) {
976 code = afs_CopyOutAttrs(vcp, vattr);
979 code = afs_getattr(vcp, vattr, credp);
984 afs_fill_inode(AFSTOV(vcp), vattr);
986 afs_DestroyAttr(vattr);
991 return afs_convert_code(code);
995 * Set iattr data into vattr. Assume vattr cleared before call.
998 iattr2vattr(struct vattr *vattrp, struct iattr *iattrp)
1000 vattrp->va_mask = iattrp->ia_valid;
1001 if (iattrp->ia_valid & ATTR_MODE)
1002 vattrp->va_mode = iattrp->ia_mode;
1003 if (iattrp->ia_valid & ATTR_UID)
1004 vattrp->va_uid = afs_from_kuid(iattrp->ia_uid);
1005 if (iattrp->ia_valid & ATTR_GID)
1006 vattrp->va_gid = afs_from_kgid(iattrp->ia_gid);
1007 if (iattrp->ia_valid & ATTR_SIZE)
1008 vattrp->va_size = iattrp->ia_size;
1009 if (iattrp->ia_valid & ATTR_ATIME) {
1010 vattrp->va_atime.tv_sec = iattrp->ia_atime.tv_sec;
1011 vattrp->va_atime.tv_usec = 0;
1013 if (iattrp->ia_valid & ATTR_MTIME) {
1014 vattrp->va_mtime.tv_sec = iattrp->ia_mtime.tv_sec;
1015 vattrp->va_mtime.tv_usec = 0;
1017 if (iattrp->ia_valid & ATTR_CTIME) {
1018 vattrp->va_ctime.tv_sec = iattrp->ia_ctime.tv_sec;
1019 vattrp->va_ctime.tv_usec = 0;
1024 * Rewrite the inode cache from the attr. Assumes all vattr fields are valid.
1027 vattr2inode(struct inode *ip, struct vattr *vp)
1029 ip->i_ino = vp->va_nodeid;
1030 #ifdef HAVE_LINUX_SET_NLINK
1031 set_nlink(ip, vp->va_nlink);
1033 ip->i_nlink = vp->va_nlink;
1035 ip->i_blocks = vp->va_blocks;
1036 #ifdef STRUCT_INODE_HAS_I_BLKBITS
1037 ip->i_blkbits = AFS_BLKBITS;
1039 #ifdef STRUCT_INODE_HAS_I_BLKSIZE
1040 ip->i_blksize = vp->va_blocksize;
1042 ip->i_rdev = vp->va_rdev;
1043 ip->i_mode = vp->va_mode;
1044 ip->i_uid = afs_make_kuid(vp->va_uid);
1045 ip->i_gid = afs_make_kgid(vp->va_gid);
1046 i_size_write(ip, vp->va_size);
1047 ip->i_atime.tv_sec = vp->va_atime.tv_sec;
1048 ip->i_atime.tv_nsec = 0;
1049 ip->i_mtime.tv_sec = vp->va_mtime.tv_sec;
1050 /* Set the mtime nanoseconds to the sysname generation number.
1051 * This convinces NFS clients that all directories have changed
1052 * any time the sysname list changes.
1054 ip->i_mtime.tv_nsec = afs_sysnamegen;
1055 ip->i_ctime.tv_sec = vp->va_ctime.tv_sec;
1056 ip->i_ctime.tv_nsec = 0;
1059 /* afs_notify_change
1060 * Linux version of setattr call. What to change is in the iattr struct.
1061 * We need to set bits in both the Linux inode as well as the vcache.
1064 afs_notify_change(struct dentry *dp, struct iattr *iattrp)
1066 struct vattr *vattr = NULL;
1067 cred_t *credp = crref();
1068 struct inode *ip = dp->d_inode;
1072 code = afs_CreateAttr(&vattr);
1077 iattr2vattr(vattr, iattrp); /* Convert for AFS vnodeops call. */
1079 code = afs_setattr(VTOAFS(ip), vattr, credp);
1081 afs_getattr(VTOAFS(ip), vattr, credp);
1082 vattr2inode(ip, vattr);
1084 afs_DestroyAttr(vattr);
1089 return afs_convert_code(code);
1092 #if defined(IOP_GETATTR_TAKES_PATH_STRUCT)
1094 afs_linux_getattr(const struct path *path, struct kstat *stat, u32 request_mask, unsigned int sync_mode)
1096 int err = afs_linux_revalidate(path->dentry);
1098 generic_fillattr(path->dentry->d_inode, stat);
1104 afs_linux_getattr(struct vfsmount *mnt, struct dentry *dentry, struct kstat *stat)
1106 int err = afs_linux_revalidate(dentry);
1108 generic_fillattr(dentry->d_inode, stat);
1115 parent_vcache_dv(struct inode *inode, cred_t *credp)
1118 struct vcache *pvcp;
1121 * If parent is a mount point and we are using fakestat, we may need
1122 * to look at the fake vcache entry instead of what the vfs is giving
1123 * us. The fake entry is the one with the useful DataVersion.
1125 pvcp = VTOAFS(inode);
1126 if (pvcp->mvstat == AFS_MVSTAT_MTPT && afs_fakestat_enable) {
1127 struct vrequest treq;
1128 struct afs_fakestat_state fakestate;
1134 afs_InitReq(&treq, credp);
1135 afs_InitFakeStat(&fakestate);
1136 afs_TryEvalFakeStat(&pvcp, &fakestate, &treq);
1139 afs_PutFakeStat(&fakestate);
1141 return hgetlo(pvcp->f.m.DataVersion);
1144 #ifndef D_SPLICE_ALIAS_RACE
1146 static inline void dentry_race_lock(void) {}
1147 static inline void dentry_race_unlock(void) {}
1151 # if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,16)
1152 static DEFINE_MUTEX(dentry_race_sem);
1154 static DECLARE_MUTEX(dentry_race_sem);
1158 dentry_race_lock(void)
1160 mutex_lock(&dentry_race_sem);
1163 dentry_race_unlock(void)
1165 mutex_unlock(&dentry_race_sem);
1168 /* Leave some trace that this code is enabled; otherwise it's pretty hard to
1170 static __attribute__((used)) const char dentry_race_marker[] = "d_splice_alias race workaround enabled";
1173 check_dentry_race(struct dentry *dp)
1177 /* In Linux, before commit 4919c5e45a91b5db5a41695fe0357fbdff0d5767,
1178 * d_splice_alias can momentarily hash a dentry before it's fully
1179 * populated. This only happens for a moment, since it's unhashed again
1180 * right after (in d_move), but this can make the dentry be found by
1181 * __d_lookup, and then given to us.
1183 * So check if the dentry is unhashed; if it is, then the dentry is not
1184 * valid. We lock dentry_race_lock() to ensure that d_splice_alias is
1185 * no longer running. Locking d_lock is required to check the dentry's
1186 * flags, so lock that, too.
1189 spin_lock(&dp->d_lock);
1190 if (d_unhashed(dp)) {
1193 spin_unlock(&dp->d_lock);
1194 dentry_race_unlock();
1198 #endif /* D_SPLICE_ALIAS_RACE */
1200 /* Validate a dentry. Return 1 if unchanged, 0 if VFS layer should re-evaluate.
1201 * In kernels 2.2.10 and above, we are passed an additional flags var which
1202 * may have either the LOOKUP_FOLLOW OR LOOKUP_DIRECTORY set in which case
1203 * we are advised to follow the entry if it is a link or to make sure that
1204 * it is a directory. But since the kernel itself checks these possibilities
1205 * later on, we shouldn't have to do it until later. Perhaps in the future..
1207 * The code here assumes that on entry the global lock is not held
1210 #if defined(DOP_REVALIDATE_TAKES_UNSIGNED)
1211 afs_linux_dentry_revalidate(struct dentry *dp, unsigned int flags)
1212 #elif defined(DOP_REVALIDATE_TAKES_NAMEIDATA)
1213 afs_linux_dentry_revalidate(struct dentry *dp, struct nameidata *nd)
1215 afs_linux_dentry_revalidate(struct dentry *dp, int flags)
1218 cred_t *credp = NULL;
1219 struct vcache *vcp, *pvcp, *tvc = NULL;
1220 struct dentry *parent;
1222 struct afs_fakestat_state fakestate;
1224 afs_uint32 parent_dv;
1227 /* We don't support RCU path walking */
1228 # if defined(DOP_REVALIDATE_TAKES_UNSIGNED)
1229 if (flags & LOOKUP_RCU)
1231 if (nd->flags & LOOKUP_RCU)
1236 #ifdef D_SPLICE_ALIAS_RACE
1237 if (check_dentry_race(dp)) {
1244 afs_InitFakeStat(&fakestate);
1247 vcp = VTOAFS(dp->d_inode);
1249 if (vcp == afs_globalVp)
1252 if (vcp->mvstat == AFS_MVSTAT_MTPT) {
1253 if (vcp->mvid.target_root && (vcp->f.states & CMValid)) {
1254 int tryEvalOnly = 0;
1256 struct vrequest *treq = NULL;
1260 code = afs_CreateReq(&treq, credp);
1264 if ((strcmp(dp->d_name.name, ".directory") == 0)) {
1268 code = afs_TryEvalFakeStat(&vcp, &fakestate, treq);
1270 code = afs_EvalFakeStat(&vcp, &fakestate, treq);
1271 afs_DestroyReq(treq);
1272 if ((tryEvalOnly && vcp->mvstat == AFS_MVSTAT_MTPT) || code) {
1273 /* a mount point, not yet replaced by its directory */
1277 } else if (vcp->mvstat == AFS_MVSTAT_ROOT && *dp->d_name.name != '/') {
1278 osi_Assert(vcp->mvid.parent != NULL);
1282 /* If the last looker changes, we should make sure the current
1283 * looker still has permission to examine this file. This would
1284 * always require a crref() which would be "slow".
1286 if (vcp->last_looker != treq.uid) {
1287 if (!afs_AccessOK(vcp, (vType(vcp) == VREG) ? PRSFS_READ : PRSFS_LOOKUP, &treq, CHECK_MODE_BITS)) {
1291 vcp->last_looker = treq.uid;
1295 parent = dget_parent(dp);
1296 pvcp = VTOAFS(parent->d_inode);
1297 parent_dv = parent_vcache_dv(parent->d_inode, credp);
1299 /* If the parent's DataVersion has changed or the vnode
1300 * is longer valid, we need to do a full lookup. VerifyVCache
1301 * isn't enough since the vnode may have been renamed.
1304 if (parent_dv > dp->d_time || !(vcp->f.states & CStatd)) {
1305 struct vattr *vattr = NULL;
1309 if (credp == NULL) {
1312 code = afs_lookup(pvcp, (char *)dp->d_name.name, &tvc, credp);
1315 /* We couldn't perform the lookup, so we're not okay. */
1318 } else if (tvc == vcp) {
1319 /* We got back the same vcache, so we're good. */
1322 } else if (tvc == VTOAFS(dp->d_inode)) {
1323 /* We got back the same vcache, so we're good. This is
1324 * different from the above case, because sometimes 'vcp' is
1325 * not the same as the vcache for dp->d_inode, if 'vcp' was a
1326 * mtpt and we evaluated it to a root dir. In rare cases,
1327 * afs_lookup might not evalute the mtpt when we do, or vice
1328 * versa, so the previous case will not succeed. But this is
1329 * still 'correct', so make sure not to mark the dentry as
1330 * invalid; it still points to the same thing! */
1334 /* We got back a different file, so we're definitely not
1341 /* Force unhash; the name doesn't point to this file
1344 if (code && code != ENOENT) {
1345 /* ...except if we couldn't perform the actual lookup,
1346 * we don't know if the name points to this file or not. */
1352 code = afs_CreateAttr(&vattr);
1358 if (afs_getattr(vcp, vattr, credp)) {
1360 afs_DestroyAttr(vattr);
1364 vattr2inode(AFSTOV(vcp), vattr);
1365 dp->d_time = parent_dv;
1367 afs_DestroyAttr(vattr);
1370 /* should we always update the attributes at this point? */
1371 /* unlikely--the vcache entry hasn't changed */
1377 /* 'dp' represents a cached negative lookup. */
1379 parent = dget_parent(dp);
1380 pvcp = VTOAFS(parent->d_inode);
1381 parent_dv = parent_vcache_dv(parent->d_inode, credp);
1383 if (parent_dv > dp->d_time || !(pvcp->f.states & CStatd)
1384 || afs_IsDynroot(pvcp)) {
1398 #ifndef D_INVALIDATE_IS_VOID
1399 /* When (v3.18) d_invalidate was converted to void, it also started
1400 * being called automatically from revalidate, and automatically
1402 * - shrink_dcache_parent
1403 * - automatic detach of submounts
1405 * Therefore, after that point, OpenAFS revalidate logic no longer needs
1406 * to do any of those things itself for invalid dentry structs. We only need
1407 * to tell VFS it's invalid (by returning 0), and VFS will handle the rest.
1409 if (have_submounts(dp))
1417 afs_PutFakeStat(&fakestate);
1422 #ifndef D_INVALIDATE_IS_VOID
1425 * If we had a negative lookup for the name we want to forcibly
1426 * unhash the dentry.
1427 * Otherwise use d_invalidate which will not unhash it if still in use.
1430 shrink_dcache_parent(dp);
1441 afs_dentry_iput(struct dentry *dp, struct inode *ip)
1443 struct vcache *vcp = VTOAFS(ip);
1446 if (!AFS_IS_DISCONNECTED || (vcp->f.states & CUnlinked)) {
1447 (void) afs_InactiveVCache(vcp, NULL);
1450 afs_linux_clear_nfsfs_renamed(dp);
1456 #if defined(DOP_D_DELETE_TAKES_CONST)
1457 afs_dentry_delete(const struct dentry *dp)
1459 afs_dentry_delete(struct dentry *dp)
1462 if (dp->d_inode && (VTOAFS(dp->d_inode)->f.states & CUnlinked))
1463 return 1; /* bad inode? */
1468 #ifdef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
1469 static struct vfsmount *
1470 afs_dentry_automount(afs_linux_path_t *path)
1472 struct dentry *target;
1475 * Avoid symlink resolution limits when resolving; we cannot contribute to
1476 * an infinite symlink loop.
1478 * On newer kernels the field has moved to the private nameidata structure
1479 * so we can't adjust it here. This may cause ELOOP when using a path with
1480 * 40 or more directories that are not already in the dentry cache.
1482 #if defined(STRUCT_TASK_STRUCT_HAS_TOTAL_LINK_COUNT)
1483 current->total_link_count--;
1486 target = canonical_dentry(path->dentry->d_inode);
1488 if (target == path->dentry) {
1495 path->dentry = target;
1498 spin_lock(&path->dentry->d_lock);
1499 path->dentry->d_flags &= ~DCACHE_NEED_AUTOMOUNT;
1500 spin_unlock(&path->dentry->d_lock);
1505 #endif /* STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT */
1507 struct dentry_operations afs_dentry_operations = {
1508 .d_revalidate = afs_linux_dentry_revalidate,
1509 .d_delete = afs_dentry_delete,
1510 .d_iput = afs_dentry_iput,
1511 #ifdef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
1512 .d_automount = afs_dentry_automount,
1513 #endif /* STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT */
1516 /**********************************************************************
1517 * AFS Linux inode operations
1518 **********************************************************************/
1522 * Merely need to set enough of vattr to get us through the create. Note
1523 * that the higher level code (open_namei) will take care of any tuncation
1524 * explicitly. Exclusive open is also taken care of in open_namei.
1526 * name is in kernel space at this point.
1529 #if defined(IOP_CREATE_TAKES_BOOL)
1530 afs_linux_create(struct inode *dip, struct dentry *dp, umode_t mode,
1532 #elif defined(IOP_CREATE_TAKES_UMODE_T)
1533 afs_linux_create(struct inode *dip, struct dentry *dp, umode_t mode,
1534 struct nameidata *nd)
1535 #elif defined(IOP_CREATE_TAKES_NAMEIDATA)
1536 afs_linux_create(struct inode *dip, struct dentry *dp, int mode,
1537 struct nameidata *nd)
1539 afs_linux_create(struct inode *dip, struct dentry *dp, int mode)
1542 struct vattr *vattr = NULL;
1543 cred_t *credp = crref();
1544 const char *name = dp->d_name.name;
1550 code = afs_CreateAttr(&vattr);
1554 vattr->va_mode = mode;
1555 vattr->va_type = mode & S_IFMT;
1557 code = afs_create(VTOAFS(dip), (char *)name, vattr, NONEXCL, mode,
1561 struct inode *ip = AFSTOV(vcp);
1563 afs_getattr(vcp, vattr, credp);
1564 afs_fill_inode(ip, vattr);
1565 insert_inode_hash(ip);
1566 #if !defined(STRUCT_SUPER_BLOCK_HAS_S_D_OP)
1567 dp->d_op = &afs_dentry_operations;
1569 dp->d_time = parent_vcache_dv(dip, credp);
1570 d_instantiate(dp, ip);
1573 afs_DestroyAttr(vattr);
1579 return afs_convert_code(code);
1582 /* afs_linux_lookup */
1583 static struct dentry *
1584 #if defined(IOP_LOOKUP_TAKES_UNSIGNED)
1585 afs_linux_lookup(struct inode *dip, struct dentry *dp,
1587 #elif defined(IOP_LOOKUP_TAKES_NAMEIDATA)
1588 afs_linux_lookup(struct inode *dip, struct dentry *dp,
1589 struct nameidata *nd)
1591 afs_linux_lookup(struct inode *dip, struct dentry *dp)
1594 cred_t *credp = crref();
1595 struct vcache *vcp = NULL;
1596 const char *comp = dp->d_name.name;
1597 struct inode *ip = NULL;
1598 struct dentry *newdp = NULL;
1603 code = afs_lookup(VTOAFS(dip), (char *)comp, &vcp, credp);
1604 if (code == ENOENT) {
1605 /* It's ok for the file to not be found. That's noted by the caller by
1606 * seeing that the dp->d_inode field is NULL (set by d_splice_alias or
1609 osi_Assert(vcp == NULL);
1617 struct vattr *vattr = NULL;
1618 struct vcache *parent_vc = VTOAFS(dip);
1620 if (parent_vc == vcp) {
1621 /* This is possible if the parent dir is a mountpoint to a volume,
1622 * and the dir entry we looked up is a mountpoint to the same
1623 * volume. Linux cannot cope with this, so return an error instead
1624 * of risking a deadlock or panic. */
1631 code = afs_CreateAttr(&vattr);
1639 afs_getattr(vcp, vattr, credp);
1640 afs_fill_inode(ip, vattr);
1641 if (hlist_unhashed(&ip->i_hash))
1642 insert_inode_hash(ip);
1644 afs_DestroyAttr(vattr);
1646 #if !defined(STRUCT_SUPER_BLOCK_HAS_S_D_OP)
1647 dp->d_op = &afs_dentry_operations;
1649 dp->d_time = parent_vcache_dv(dip, credp);
1653 if (ip && S_ISDIR(ip->i_mode)) {
1654 d_prune_aliases(ip);
1656 #ifdef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
1657 /* Only needed if this is a volume root */
1658 if (vcp->mvstat == 2)
1659 ip->i_flags |= S_AUTOMOUNT;
1663 * Take an extra reference so the inode doesn't go away if
1664 * d_splice_alias drops our reference on error.
1667 #ifdef HAVE_LINUX_IHOLD
1674 newdp = d_splice_alias(ip, dp);
1675 dentry_race_unlock();
1680 if (IS_ERR(newdp)) {
1681 /* d_splice_alias can return an error (EIO) if there is an existing
1682 * connected directory alias for this dentry. Add our dentry manually
1683 * ourselves if this happens. */
1686 #if defined(D_SPLICE_ALIAS_LEAK_ON_ERROR)
1687 /* Depending on the kernel version, d_splice_alias may or may not drop
1688 * the inode reference on error. If it didn't, do it here. */
1697 return ERR_PTR(afs_convert_code(code));
1705 afs_linux_link(struct dentry *olddp, struct inode *dip, struct dentry *newdp)
1708 cred_t *credp = crref();
1709 const char *name = newdp->d_name.name;
1710 struct inode *oldip = olddp->d_inode;
1712 /* If afs_link returned the vnode, we could instantiate the
1713 * dentry. Since it's not, we drop this one and do a new lookup.
1718 code = afs_link(VTOAFS(oldip), VTOAFS(dip), (char *)name, credp);
1722 return afs_convert_code(code);
1725 /* We have to have a Linux specific sillyrename function, because we
1726 * also have to keep the dcache up to date when we're doing a silly
1727 * rename - so we don't want the generic vnodeops doing this behind our
1732 afs_linux_sillyrename(struct inode *dir, struct dentry *dentry,
1735 struct vcache *tvc = VTOAFS(dentry->d_inode);
1736 struct dentry *__dp = NULL;
1737 char *__name = NULL;
1740 if (afs_linux_nfsfs_renamed(dentry))
1748 osi_FreeSmallSpace(__name);
1749 __name = afs_newname();
1752 __dp = lookup_one_len(__name, dentry->d_parent, strlen(__name));
1755 osi_FreeSmallSpace(__name);
1758 } while (__dp->d_inode != NULL);
1761 code = afs_rename(VTOAFS(dir), (char *)dentry->d_name.name,
1762 VTOAFS(dir), (char *)__dp->d_name.name,
1765 tvc->mvid.silly_name = __name;
1768 crfree(tvc->uncred);
1770 tvc->uncred = credp;
1771 tvc->f.states |= CUnlinked;
1772 afs_linux_set_nfsfs_renamed(dentry);
1774 __dp->d_time = 0; /* force to revalidate */
1775 d_move(dentry, __dp);
1777 osi_FreeSmallSpace(__name);
1788 afs_linux_unlink(struct inode *dip, struct dentry *dp)
1791 cred_t *credp = crref();
1792 const char *name = dp->d_name.name;
1793 struct vcache *tvc = VTOAFS(dp->d_inode);
1795 if (VREFCOUNT(tvc) > 1 && tvc->opens > 0
1796 && !(tvc->f.states & CUnlinked)) {
1798 code = afs_linux_sillyrename(dip, dp, credp);
1801 code = afs_remove(VTOAFS(dip), (char *)name, credp);
1808 return afs_convert_code(code);
1813 afs_linux_symlink(struct inode *dip, struct dentry *dp, const char *target)
1816 cred_t *credp = crref();
1817 struct vattr *vattr = NULL;
1818 const char *name = dp->d_name.name;
1820 /* If afs_symlink returned the vnode, we could instantiate the
1821 * dentry. Since it's not, we drop this one and do a new lookup.
1826 code = afs_CreateAttr(&vattr);
1831 code = afs_symlink(VTOAFS(dip), (char *)name, vattr, (char *)target, NULL,
1833 afs_DestroyAttr(vattr);
1838 return afs_convert_code(code);
1842 #if defined(IOP_MKDIR_TAKES_UMODE_T)
1843 afs_linux_mkdir(struct inode *dip, struct dentry *dp, umode_t mode)
1845 afs_linux_mkdir(struct inode *dip, struct dentry *dp, int mode)
1849 cred_t *credp = crref();
1850 struct vcache *tvcp = NULL;
1851 struct vattr *vattr = NULL;
1852 const char *name = dp->d_name.name;
1855 code = afs_CreateAttr(&vattr);
1860 vattr->va_mask = ATTR_MODE;
1861 vattr->va_mode = mode;
1863 code = afs_mkdir(VTOAFS(dip), (char *)name, vattr, &tvcp, credp);
1866 struct inode *ip = AFSTOV(tvcp);
1868 afs_getattr(tvcp, vattr, credp);
1869 afs_fill_inode(ip, vattr);
1871 #if !defined(STRUCT_SUPER_BLOCK_HAS_S_D_OP)
1872 dp->d_op = &afs_dentry_operations;
1874 dp->d_time = parent_vcache_dv(dip, credp);
1875 d_instantiate(dp, ip);
1877 afs_DestroyAttr(vattr);
1883 return afs_convert_code(code);
1887 afs_linux_rmdir(struct inode *dip, struct dentry *dp)
1890 cred_t *credp = crref();
1891 const char *name = dp->d_name.name;
1893 /* locking kernel conflicts with glock? */
1896 code = afs_rmdir(VTOAFS(dip), (char *)name, credp);
1899 /* Linux likes to see ENOTEMPTY returned from an rmdir() syscall
1900 * that failed because a directory is not empty. So, we map
1901 * EEXIST to ENOTEMPTY on linux.
1903 if (code == EEXIST) {
1912 return afs_convert_code(code);
1917 afs_linux_rename(struct inode *oldip, struct dentry *olddp,
1918 struct inode *newip, struct dentry *newdp
1919 #ifdef HAVE_LINUX_INODE_OPERATIONS_RENAME_TAKES_FLAGS
1920 , unsigned int flags
1925 cred_t *credp = crref();
1926 const char *oldname = olddp->d_name.name;
1927 const char *newname = newdp->d_name.name;
1928 struct dentry *rehash = NULL;
1930 #ifdef HAVE_LINUX_INODE_OPERATIONS_RENAME_TAKES_FLAGS
1932 return -EINVAL; /* no support for new flags yet */
1935 /* Prevent any new references during rename operation. */
1937 if (!d_unhashed(newdp)) {
1942 afs_maybe_shrink_dcache(olddp);
1945 code = afs_rename(VTOAFS(oldip), (char *)oldname, VTOAFS(newip), (char *)newname, credp);
1949 olddp->d_time = 0; /* force to revalidate */
1955 return afs_convert_code(code);
1959 /* afs_linux_ireadlink
1960 * Internal readlink which can return link contents to user or kernel space.
1961 * Note that the buffer is NOT supposed to be null-terminated.
1964 afs_linux_ireadlink(struct inode *ip, char *target, int maxlen, uio_seg_t seg)
1967 cred_t *credp = crref();
1971 memset(&tuio, 0, sizeof(tuio));
1972 memset(&iov, 0, sizeof(iov));
1974 setup_uio(&tuio, &iov, target, (afs_offs_t) 0, maxlen, UIO_READ, seg);
1975 code = afs_readlink(VTOAFS(ip), &tuio, credp);
1979 return maxlen - tuio.uio_resid;
1981 return afs_convert_code(code);
1984 #if !defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
1985 /* afs_linux_readlink
1986 * Fill target (which is in user space) with contents of symlink.
1989 afs_linux_readlink(struct dentry *dp, char *target, int maxlen)
1992 struct inode *ip = dp->d_inode;
1995 code = afs_linux_ireadlink(ip, target, maxlen, AFS_UIOUSER);
2001 /* afs_linux_follow_link
2002 * a file system dependent link following routine.
2004 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2005 static const char *afs_linux_follow_link(struct dentry *dentry, void **link_data)
2007 static int afs_linux_follow_link(struct dentry *dentry, struct nameidata *nd)
2013 name = kmalloc(PATH_MAX, GFP_NOFS);
2015 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2016 return ERR_PTR(-EIO);
2023 code = afs_linux_ireadlink(dentry->d_inode, name, PATH_MAX - 1, AFS_UIOSYS);
2027 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2028 return ERR_PTR(code);
2035 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2036 return *link_data = name;
2038 nd_set_link(nd, name);
2043 #if defined(HAVE_LINUX_INODE_OPERATIONS_PUT_LINK_NO_NAMEIDATA)
2045 afs_linux_put_link(struct inode *inode, void *link_data)
2047 char *name = link_data;
2049 if (name && !IS_ERR(name))
2054 afs_linux_put_link(struct dentry *dentry, struct nameidata *nd)
2056 char *name = nd_get_link(nd);
2058 if (name && !IS_ERR(name))
2061 #endif /* HAVE_LINUX_INODE_OPERATIONS_PUT_LINK_NO_NAMEIDATA */
2063 #endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
2065 /* Populate a page by filling it from the cache file pointed at by cachefp
2066 * (which contains indicated chunk)
2067 * If task is NULL, the page copy occurs syncronously, and the routine
2068 * returns with page still locked. If task is non-NULL, then page copies
2069 * may occur in the background, and the page will be unlocked when it is
2073 afs_linux_read_cache(struct file *cachefp, struct page *page,
2074 int chunk, struct pagevec *lrupv,
2075 struct afs_pagecopy_task *task) {
2076 loff_t offset = page_offset(page);
2077 struct inode *cacheinode = cachefp->f_dentry->d_inode;
2078 struct page *newpage, *cachepage;
2079 struct address_space *cachemapping;
2083 cachemapping = cacheinode->i_mapping;
2087 /* If we're trying to read a page that's past the end of the disk
2088 * cache file, then just return a zeroed page */
2089 if (AFS_CHUNKOFFSET(offset) >= i_size_read(cacheinode)) {
2090 zero_user_segment(page, 0, PAGE_SIZE);
2091 SetPageUptodate(page);
2097 /* From our offset, we now need to work out which page in the disk
2098 * file it corresponds to. This will be fun ... */
2099 pageindex = (offset - AFS_CHUNKTOBASE(chunk)) >> PAGE_SHIFT;
2101 while (cachepage == NULL) {
2102 cachepage = find_get_page(cachemapping, pageindex);
2105 newpage = page_cache_alloc(cachemapping);
2111 code = add_to_page_cache(newpage, cachemapping,
2112 pageindex, GFP_KERNEL);
2114 cachepage = newpage;
2117 get_page(cachepage);
2118 if (!pagevec_add(lrupv, cachepage))
2119 __pagevec_lru_add_file(lrupv);
2124 if (code != -EEXIST)
2128 lock_page(cachepage);
2132 if (!PageUptodate(cachepage)) {
2133 ClearPageError(cachepage);
2134 code = cachemapping->a_ops->readpage(NULL, cachepage);
2135 if (!code && !task) {
2136 wait_on_page_locked(cachepage);
2139 unlock_page(cachepage);
2143 if (PageUptodate(cachepage)) {
2144 copy_highpage(page, cachepage);
2145 flush_dcache_page(page);
2146 SetPageUptodate(page);
2151 afs_pagecopy_queue_page(task, cachepage, page);
2163 put_page(cachepage);
2169 afs_linux_readpage_fastpath(struct file *fp, struct page *pp, int *codep)
2171 loff_t offset = page_offset(pp);
2172 struct inode *ip = FILE_INODE(fp);
2173 struct vcache *avc = VTOAFS(ip);
2175 struct file *cacheFp = NULL;
2178 struct pagevec lrupv;
2180 /* Not a UFS cache, don't do anything */
2181 if (cacheDiskType != AFS_FCACHE_TYPE_UFS)
2184 /* No readpage (ex: tmpfs) , skip */
2185 if (cachefs_noreadpage)
2188 /* Can't do anything if the vcache isn't statd , or if the read
2189 * crosses a chunk boundary.
2191 if (!(avc->f.states & CStatd) ||
2192 AFS_CHUNK(offset) != AFS_CHUNK(offset + PAGE_SIZE)) {
2196 ObtainWriteLock(&avc->lock, 911);
2198 /* XXX - See if hinting actually makes things faster !!! */
2200 /* See if we have a suitable entry already cached */
2204 /* We need to lock xdcache, then dcache, to handle situations where
2205 * the hint is on the free list. However, we can't safely do this
2206 * according to the locking hierarchy. So, use a non blocking lock.
2208 ObtainReadLock(&afs_xdcache);
2209 dcLocked = ( 0 == NBObtainReadLock(&tdc->lock));
2211 if (dcLocked && (tdc->index != NULLIDX)
2212 && !FidCmp(&tdc->f.fid, &avc->f.fid)
2213 && tdc->f.chunk == AFS_CHUNK(offset)
2214 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
2215 /* Bonus - the hint was correct */
2218 /* Only destroy the hint if its actually invalid, not if there's
2219 * just been a locking failure */
2221 ReleaseReadLock(&tdc->lock);
2228 ReleaseReadLock(&afs_xdcache);
2231 /* No hint, or hint is no longer valid - see if we can get something
2232 * directly from the dcache
2235 tdc = afs_FindDCache(avc, offset);
2238 ReleaseWriteLock(&avc->lock);
2243 ObtainReadLock(&tdc->lock);
2245 /* Is the dcache we've been given currently up to date */
2246 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) ||
2247 (tdc->dflags & DFFetching))
2250 /* Update our hint for future abuse */
2253 /* Okay, so we've now got a cache file that is up to date */
2255 /* XXX - I suspect we should be locking the inodes before we use them! */
2257 cacheFp = afs_linux_raw_open(&tdc->f.inode);
2258 osi_Assert(cacheFp);
2259 if (!cacheFp->f_dentry->d_inode->i_mapping->a_ops->readpage) {
2260 cachefs_noreadpage = 1;
2264 #if defined(PAGEVEC_INIT_COLD_ARG)
2265 pagevec_init(&lrupv, 0);
2267 pagevec_init(&lrupv);
2270 code = afs_linux_read_cache(cacheFp, pp, tdc->f.chunk, &lrupv, NULL);
2272 if (pagevec_count(&lrupv))
2273 __pagevec_lru_add_file(&lrupv);
2275 filp_close(cacheFp, NULL);
2278 ReleaseReadLock(&tdc->lock);
2279 ReleaseWriteLock(&avc->lock);
2286 ReleaseWriteLock(&avc->lock);
2287 ReleaseReadLock(&tdc->lock);
2292 /* afs_linux_readpage
2294 * This function is split into two, because prepare_write/begin_write
2295 * require a readpage call which doesn't unlock the resulting page upon
2299 afs_linux_fillpage(struct file *fp, struct page *pp)
2304 struct iovec *iovecp;
2305 struct inode *ip = FILE_INODE(fp);
2306 afs_int32 cnt = page_count(pp);
2307 struct vcache *avc = VTOAFS(ip);
2308 afs_offs_t offset = page_offset(pp);
2312 if (afs_linux_readpage_fastpath(fp, pp, &code)) {
2322 auio = kmalloc(sizeof(struct uio), GFP_NOFS);
2323 iovecp = kmalloc(sizeof(struct iovec), GFP_NOFS);
2325 setup_uio(auio, iovecp, (char *)address, offset, PAGE_SIZE, UIO_READ,
2330 afs_Trace4(afs_iclSetp, CM_TRACE_READPAGE, ICL_TYPE_POINTER, ip,
2331 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, cnt, ICL_TYPE_INT32,
2332 99999); /* not a possible code value */
2334 code = afs_rdwr(avc, auio, UIO_READ, 0, credp);
2336 afs_Trace4(afs_iclSetp, CM_TRACE_READPAGE, ICL_TYPE_POINTER, ip,
2337 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, cnt, ICL_TYPE_INT32,
2339 AFS_DISCON_UNLOCK();
2342 /* XXX valid for no-cache also? Check last bits of files... :)
2343 * Cognate code goes in afs_NoCacheFetchProc. */
2344 if (auio->uio_resid) /* zero remainder of page */
2345 memset((void *)(address + (PAGE_SIZE - auio->uio_resid)), 0,
2348 flush_dcache_page(pp);
2349 SetPageUptodate(pp);
2358 return afs_convert_code(code);
2362 afs_linux_prefetch(struct file *fp, struct page *pp)
2365 struct vcache *avc = VTOAFS(FILE_INODE(fp));
2366 afs_offs_t offset = page_offset(pp);
2368 if (AFS_CHUNKOFFSET(offset) == 0) {
2370 struct vrequest *treq = NULL;
2375 code = afs_CreateReq(&treq, credp);
2376 if (!code && !NBObtainWriteLock(&avc->lock, 534)) {
2377 tdc = afs_FindDCache(avc, offset);
2379 if (!(tdc->mflags & DFNextStarted))
2380 afs_PrefetchChunk(avc, tdc, credp, treq);
2383 ReleaseWriteLock(&avc->lock);
2385 afs_DestroyReq(treq);
2389 return afs_convert_code(code);
2394 afs_linux_bypass_readpages(struct file *fp, struct address_space *mapping,
2395 struct list_head *page_list, unsigned num_pages)
2400 struct iovec* iovecp;
2401 struct nocache_read_request *ancr;
2403 struct pagevec lrupv;
2407 struct inode *ip = FILE_INODE(fp);
2408 struct vcache *avc = VTOAFS(ip);
2409 afs_int32 base_index = 0;
2410 afs_int32 page_count = 0;
2413 /* background thread must free: iovecp, auio, ancr */
2414 iovecp = osi_Alloc(num_pages * sizeof(struct iovec));
2416 auio = osi_Alloc(sizeof(struct uio));
2417 auio->uio_iov = iovecp;
2418 auio->uio_iovcnt = num_pages;
2419 auio->uio_flag = UIO_READ;
2420 auio->uio_seg = AFS_UIOSYS;
2421 auio->uio_resid = num_pages * PAGE_SIZE;
2423 ancr = osi_Alloc(sizeof(struct nocache_read_request));
2425 ancr->offset = auio->uio_offset;
2426 ancr->length = auio->uio_resid;
2428 #if defined(PAGEVEC_INIT_COLD_ARG)
2429 pagevec_init(&lrupv, 0);
2431 pagevec_init(&lrupv);
2434 for(page_ix = 0; page_ix < num_pages; ++page_ix) {
2436 if(list_empty(page_list))
2439 pp = list_entry(page_list->prev, struct page, lru);
2440 /* If we allocate a page and don't remove it from page_list,
2441 * the page cache gets upset. */
2443 isize = (i_size_read(fp->f_mapping->host) - 1) >> PAGE_SHIFT;
2444 if(pp->index > isize) {
2451 offset = page_offset(pp);
2452 ancr->offset = auio->uio_offset = offset;
2453 base_index = pp->index;
2455 iovecp[page_ix].iov_len = PAGE_SIZE;
2456 code = add_to_page_cache(pp, mapping, pp->index, GFP_KERNEL);
2457 if(base_index != pp->index) {
2461 iovecp[page_ix].iov_base = (void *) 0;
2463 ancr->length -= PAGE_SIZE;
2471 iovecp[page_ix].iov_base = (void *) 0;
2474 if(!PageLocked(pp)) {
2478 /* increment page refcount--our original design assumed
2479 * that locking it would effectively pin it; protect
2480 * ourselves from the possiblity that this assumption is
2481 * is faulty, at low cost (provided we do not fail to
2482 * do the corresponding decref on the other side) */
2485 /* save the page for background map */
2486 iovecp[page_ix].iov_base = (void*) pp;
2488 /* and put it on the LRU cache */
2489 if (!pagevec_add(&lrupv, pp))
2490 __pagevec_lru_add_file(&lrupv);
2494 /* If there were useful pages in the page list, make sure all pages
2495 * are in the LRU cache, then schedule the read */
2497 if (pagevec_count(&lrupv))
2498 __pagevec_lru_add_file(&lrupv);
2500 code = afs_ReadNoCache(avc, ancr, credp);
2503 /* If there is nothing for the background thread to handle,
2504 * it won't be freeing the things that we never gave it */
2505 osi_Free(iovecp, num_pages * sizeof(struct iovec));
2506 osi_Free(auio, sizeof(struct uio));
2507 osi_Free(ancr, sizeof(struct nocache_read_request));
2509 /* we do not flush, release, or unmap pages--that will be
2510 * done for us by the background thread as each page comes in
2511 * from the fileserver */
2512 return afs_convert_code(code);
2517 afs_linux_bypass_readpage(struct file *fp, struct page *pp)
2519 cred_t *credp = NULL;
2521 struct iovec *iovecp;
2522 struct nocache_read_request *ancr;
2526 * Special case: if page is at or past end of file, just zero it and set
2529 if (page_offset(pp) >= i_size_read(fp->f_mapping->host)) {
2530 zero_user_segment(pp, 0, PAGE_SIZE);
2531 SetPageUptodate(pp);
2538 /* receiver frees */
2539 auio = osi_Alloc(sizeof(struct uio));
2540 iovecp = osi_Alloc(sizeof(struct iovec));
2542 /* address can be NULL, because we overwrite it with 'pp', below */
2543 setup_uio(auio, iovecp, NULL, page_offset(pp),
2544 PAGE_SIZE, UIO_READ, AFS_UIOSYS);
2546 /* save the page for background map */
2547 get_page(pp); /* see above */
2548 auio->uio_iov->iov_base = (void*) pp;
2549 /* the background thread will free this */
2550 ancr = osi_Alloc(sizeof(struct nocache_read_request));
2552 ancr->offset = page_offset(pp);
2553 ancr->length = PAGE_SIZE;
2556 code = afs_ReadNoCache(VTOAFS(FILE_INODE(fp)), ancr, credp);
2559 return afs_convert_code(code);
2563 afs_linux_can_bypass(struct inode *ip) {
2565 switch(cache_bypass_strategy) {
2566 case NEVER_BYPASS_CACHE:
2568 case ALWAYS_BYPASS_CACHE:
2570 case LARGE_FILES_BYPASS_CACHE:
2571 if (i_size_read(ip) > cache_bypass_threshold)
2578 /* Check if a file is permitted to bypass the cache by policy, and modify
2579 * the cache bypass state recorded for that file */
2582 afs_linux_bypass_check(struct inode *ip) {
2585 int bypass = afs_linux_can_bypass(ip);
2588 trydo_cache_transition(VTOAFS(ip), credp, bypass);
2596 afs_linux_readpage(struct file *fp, struct page *pp)
2600 if (afs_linux_bypass_check(FILE_INODE(fp))) {
2601 code = afs_linux_bypass_readpage(fp, pp);
2603 code = afs_linux_fillpage(fp, pp);
2605 code = afs_linux_prefetch(fp, pp);
2612 /* Readpages reads a number of pages for a particular file. We use
2613 * this to optimise the reading, by limiting the number of times upon which
2614 * we have to lookup, lock and open vcaches and dcaches
2618 afs_linux_readpages(struct file *fp, struct address_space *mapping,
2619 struct list_head *page_list, unsigned int num_pages)
2621 struct inode *inode = mapping->host;
2622 struct vcache *avc = VTOAFS(inode);
2624 struct file *cacheFp = NULL;
2626 unsigned int page_idx;
2628 struct pagevec lrupv;
2629 struct afs_pagecopy_task *task;
2631 if (afs_linux_bypass_check(inode))
2632 return afs_linux_bypass_readpages(fp, mapping, page_list, num_pages);
2634 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2637 /* No readpage (ex: tmpfs) , skip */
2638 if (cachefs_noreadpage)
2642 if ((code = afs_linux_VerifyVCache(avc, NULL))) {
2647 ObtainWriteLock(&avc->lock, 912);
2650 task = afs_pagecopy_init_task();
2653 #if defined(PAGEVEC_INIT_COLD_ARG)
2654 pagevec_init(&lrupv, 0);
2656 pagevec_init(&lrupv);
2658 for (page_idx = 0; page_idx < num_pages; page_idx++) {
2659 struct page *page = list_entry(page_list->prev, struct page, lru);
2660 list_del(&page->lru);
2661 offset = page_offset(page);
2663 if (tdc && tdc->f.chunk != AFS_CHUNK(offset)) {
2665 ReleaseReadLock(&tdc->lock);
2670 filp_close(cacheFp, NULL);
2675 if ((tdc = afs_FindDCache(avc, offset))) {
2676 ObtainReadLock(&tdc->lock);
2677 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) ||
2678 (tdc->dflags & DFFetching)) {
2679 ReleaseReadLock(&tdc->lock);
2686 cacheFp = afs_linux_raw_open(&tdc->f.inode);
2687 osi_Assert(cacheFp);
2688 if (!cacheFp->f_dentry->d_inode->i_mapping->a_ops->readpage) {
2689 cachefs_noreadpage = 1;
2695 if (tdc && !add_to_page_cache(page, mapping, page->index,
2698 if (!pagevec_add(&lrupv, page))
2699 __pagevec_lru_add_file(&lrupv);
2701 afs_linux_read_cache(cacheFp, page, tdc->f.chunk, &lrupv, task);
2705 if (pagevec_count(&lrupv))
2706 __pagevec_lru_add_file(&lrupv);
2710 filp_close(cacheFp, NULL);
2712 afs_pagecopy_put_task(task);
2716 ReleaseReadLock(&tdc->lock);
2720 ReleaseWriteLock(&avc->lock);
2725 /* Prepare an AFS vcache for writeback. Should be called with the vcache
2728 afs_linux_prepare_writeback(struct vcache *avc) {
2730 struct pagewriter *pw;
2732 pid = MyPidxx2Pid(MyPidxx);
2733 /* Prevent recursion into the writeback code */
2734 spin_lock(&avc->pagewriter_lock);
2735 list_for_each_entry(pw, &avc->pagewriters, link) {
2736 if (pw->writer == pid) {
2737 spin_unlock(&avc->pagewriter_lock);
2738 return AOP_WRITEPAGE_ACTIVATE;
2741 spin_unlock(&avc->pagewriter_lock);
2743 /* Add ourselves to writer list */
2744 pw = osi_Alloc(sizeof(struct pagewriter));
2746 spin_lock(&avc->pagewriter_lock);
2747 list_add_tail(&pw->link, &avc->pagewriters);
2748 spin_unlock(&avc->pagewriter_lock);
2754 afs_linux_dopartialwrite(struct vcache *avc, cred_t *credp) {
2755 struct vrequest *treq = NULL;
2758 if (!afs_CreateReq(&treq, credp)) {
2759 code = afs_DoPartialWrite(avc, treq);
2760 afs_DestroyReq(treq);
2763 return afs_convert_code(code);
2767 afs_linux_complete_writeback(struct vcache *avc) {
2768 struct pagewriter *pw, *store;
2770 struct list_head tofree;
2772 INIT_LIST_HEAD(&tofree);
2773 pid = MyPidxx2Pid(MyPidxx);
2774 /* Remove ourselves from writer list */
2775 spin_lock(&avc->pagewriter_lock);
2776 list_for_each_entry_safe(pw, store, &avc->pagewriters, link) {
2777 if (pw->writer == pid) {
2778 list_del(&pw->link);
2779 /* osi_Free may sleep so we need to defer it */
2780 list_add_tail(&pw->link, &tofree);
2783 spin_unlock(&avc->pagewriter_lock);
2784 list_for_each_entry_safe(pw, store, &tofree, link) {
2785 list_del(&pw->link);
2786 osi_Free(pw, sizeof(struct pagewriter));
2790 /* Writeback a given page syncronously. Called with no AFS locks held */
2792 afs_linux_page_writeback(struct inode *ip, struct page *pp,
2793 unsigned long offset, unsigned int count,
2796 struct vcache *vcp = VTOAFS(ip);
2804 memset(&tuio, 0, sizeof(tuio));
2805 memset(&iovec, 0, sizeof(iovec));
2807 buffer = kmap(pp) + offset;
2808 base = page_offset(pp) + offset;
2811 afs_Trace4(afs_iclSetp, CM_TRACE_UPDATEPAGE, ICL_TYPE_POINTER, vcp,
2812 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, page_count(pp),
2813 ICL_TYPE_INT32, 99999);
2815 setup_uio(&tuio, &iovec, buffer, base, count, UIO_WRITE, AFS_UIOSYS);
2817 code = afs_write(vcp, &tuio, f_flags, credp, 0);
2819 i_size_write(ip, vcp->f.m.Length);
2820 ip->i_blocks = ((vcp->f.m.Length + 1023) >> 10) << 1;
2822 code = code ? afs_convert_code(code) : count - tuio.uio_resid;
2824 afs_Trace4(afs_iclSetp, CM_TRACE_UPDATEPAGE, ICL_TYPE_POINTER, vcp,
2825 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, page_count(pp),
2826 ICL_TYPE_INT32, code);
2835 afs_linux_writepage_sync(struct inode *ip, struct page *pp,
2836 unsigned long offset, unsigned int count)
2840 struct vcache *vcp = VTOAFS(ip);
2843 /* Catch recursive writeback. This occurs if the kernel decides
2844 * writeback is required whilst we are writing to the cache, or
2845 * flushing to the server. When we're running syncronously (as
2846 * opposed to from writepage) we can't actually do anything about
2847 * this case - as we can't return AOP_WRITEPAGE_ACTIVATE to write()
2850 ObtainWriteLock(&vcp->lock, 532);
2851 afs_linux_prepare_writeback(vcp);
2852 ReleaseWriteLock(&vcp->lock);
2856 code = afs_linux_page_writeback(ip, pp, offset, count, credp);
2859 ObtainWriteLock(&vcp->lock, 533);
2861 code1 = afs_linux_dopartialwrite(vcp, credp);
2862 afs_linux_complete_writeback(vcp);
2863 ReleaseWriteLock(&vcp->lock);
2874 #ifdef AOP_WRITEPAGE_TAKES_WRITEBACK_CONTROL
2875 afs_linux_writepage(struct page *pp, struct writeback_control *wbc)
2877 afs_linux_writepage(struct page *pp)
2880 struct address_space *mapping = pp->mapping;
2881 struct inode *inode;
2884 unsigned int to = PAGE_SIZE;
2891 inode = mapping->host;
2892 vcp = VTOAFS(inode);
2893 isize = i_size_read(inode);
2895 /* Don't defeat an earlier truncate */
2896 if (page_offset(pp) > isize) {
2897 set_page_writeback(pp);
2903 ObtainWriteLock(&vcp->lock, 537);
2904 code = afs_linux_prepare_writeback(vcp);
2905 if (code == AOP_WRITEPAGE_ACTIVATE) {
2906 /* WRITEPAGE_ACTIVATE is the only return value that permits us
2907 * to return with the page still locked */
2908 ReleaseWriteLock(&vcp->lock);
2913 /* Grab the creds structure currently held in the vnode, and
2914 * get a reference to it, in case it goes away ... */
2920 ReleaseWriteLock(&vcp->lock);
2923 set_page_writeback(pp);
2925 SetPageUptodate(pp);
2927 /* We can unlock the page here, because it's protected by the
2928 * page_writeback flag. This should make us less vulnerable to
2929 * deadlocking in afs_write and afs_DoPartialWrite
2933 /* If this is the final page, then just write the number of bytes that
2934 * are actually in it */
2935 if ((isize - page_offset(pp)) < to )
2936 to = isize - page_offset(pp);
2938 code = afs_linux_page_writeback(inode, pp, 0, to, credp);
2941 ObtainWriteLock(&vcp->lock, 538);
2943 /* As much as we might like to ignore a file server error here,
2944 * and just try again when we close(), unfortunately StoreAllSegments
2945 * will invalidate our chunks if the server returns a permanent error,
2946 * so we need to at least try and get that error back to the user
2949 code1 = afs_linux_dopartialwrite(vcp, credp);
2951 afs_linux_complete_writeback(vcp);
2952 ReleaseWriteLock(&vcp->lock);
2957 end_page_writeback(pp);
2969 /* afs_linux_permission
2970 * Check access rights - returns error if can't check or permission denied.
2973 #if defined(IOP_PERMISSION_TAKES_FLAGS)
2974 afs_linux_permission(struct inode *ip, int mode, unsigned int flags)
2975 #elif defined(IOP_PERMISSION_TAKES_NAMEIDATA)
2976 afs_linux_permission(struct inode *ip, int mode, struct nameidata *nd)
2978 afs_linux_permission(struct inode *ip, int mode)
2985 /* Check for RCU path walking */
2986 #if defined(IOP_PERMISSION_TAKES_FLAGS)
2987 if (flags & IPERM_FLAG_RCU)
2989 #elif defined(MAY_NOT_BLOCK)
2990 if (mode & MAY_NOT_BLOCK)
2996 if (mode & MAY_EXEC)
2998 if (mode & MAY_READ)
3000 if (mode & MAY_WRITE)
3002 code = afs_access(VTOAFS(ip), tmp, credp);
3006 return afs_convert_code(code);
3010 afs_linux_commit_write(struct file *file, struct page *page, unsigned offset,
3014 struct inode *inode = FILE_INODE(file);
3015 loff_t pagebase = page_offset(page);
3017 if (i_size_read(inode) < (pagebase + offset))
3018 i_size_write(inode, pagebase + offset);
3020 if (PageChecked(page)) {
3021 SetPageUptodate(page);
3022 ClearPageChecked(page);
3025 code = afs_linux_writepage_sync(inode, page, offset, to - offset);
3031 afs_linux_prepare_write(struct file *file, struct page *page, unsigned from,
3035 /* http://kerneltrap.org/node/4941 details the expected behaviour of
3036 * prepare_write. Essentially, if the page exists within the file,
3037 * and is not being fully written, then we should populate it.
3040 if (!PageUptodate(page)) {
3041 loff_t pagebase = page_offset(page);
3042 loff_t isize = i_size_read(page->mapping->host);
3044 /* Is the location we are writing to beyond the end of the file? */
3045 if (pagebase >= isize ||
3046 ((from == 0) && (pagebase + to) >= isize)) {
3047 zero_user_segments(page, 0, from, to, PAGE_SIZE);
3048 SetPageChecked(page);
3049 /* Are we we writing a full page */
3050 } else if (from == 0 && to == PAGE_SIZE) {
3051 SetPageChecked(page);
3052 /* Is the page readable, if it's wronly, we don't care, because we're
3053 * not actually going to read from it ... */
3054 } else if ((file->f_flags && O_ACCMODE) != O_WRONLY) {
3055 /* We don't care if fillpage fails, because if it does the page
3056 * won't be marked as up to date
3058 afs_linux_fillpage(file, page);
3064 #if defined(STRUCT_ADDRESS_SPACE_OPERATIONS_HAS_WRITE_BEGIN)
3066 afs_linux_write_end(struct file *file, struct address_space *mapping,
3067 loff_t pos, unsigned len, unsigned copied,
3068 struct page *page, void *fsdata)
3071 unsigned int from = pos & (PAGE_SIZE - 1);
3073 code = afs_linux_commit_write(file, page, from, from + copied);
3081 afs_linux_write_begin(struct file *file, struct address_space *mapping,
3082 loff_t pos, unsigned len, unsigned flags,
3083 struct page **pagep, void **fsdata)
3086 pgoff_t index = pos >> PAGE_SHIFT;
3087 unsigned int from = pos & (PAGE_SIZE - 1);
3090 page = grab_cache_page_write_begin(mapping, index, flags);
3097 code = afs_linux_prepare_write(file, page, from, from + len);
3107 #ifndef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
3109 afs_linux_dir_follow_link(struct dentry *dentry, struct nameidata *nd)
3111 struct dentry **dpp;
3112 struct dentry *target;
3114 if (current->total_link_count > 0) {
3115 /* avoid symlink resolution limits when resolving; we cannot contribute to
3116 * an infinite symlink loop */
3117 /* only do this for follow_link when total_link_count is positive to be
3118 * on the safe side; there is at least one code path in the Linux
3119 * kernel where it seems like it may be possible to get here without
3120 * total_link_count getting incremented. it is not clear on how that
3121 * path is actually reached, but guard against it just to be safe */
3122 current->total_link_count--;
3125 target = canonical_dentry(dentry->d_inode);
3127 # ifdef STRUCT_NAMEIDATA_HAS_PATH
3128 dpp = &nd->path.dentry;
3138 *dpp = dget(dentry);
3141 nd->last_type = LAST_BIND;
3145 #endif /* !STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT */
3148 static struct inode_operations afs_file_iops = {
3149 .permission = afs_linux_permission,
3150 .getattr = afs_linux_getattr,
3151 .setattr = afs_notify_change,
3154 static struct address_space_operations afs_file_aops = {
3155 .readpage = afs_linux_readpage,
3156 .readpages = afs_linux_readpages,
3157 .writepage = afs_linux_writepage,
3158 #if defined (STRUCT_ADDRESS_SPACE_OPERATIONS_HAS_WRITE_BEGIN)
3159 .write_begin = afs_linux_write_begin,
3160 .write_end = afs_linux_write_end,
3162 .commit_write = afs_linux_commit_write,
3163 .prepare_write = afs_linux_prepare_write,
3168 /* Separate ops vector for directories. Linux 2.2 tests type of inode
3169 * by what sort of operation is allowed.....
3172 static struct inode_operations afs_dir_iops = {
3173 .setattr = afs_notify_change,
3174 .create = afs_linux_create,
3175 .lookup = afs_linux_lookup,
3176 .link = afs_linux_link,
3177 .unlink = afs_linux_unlink,
3178 .symlink = afs_linux_symlink,
3179 .mkdir = afs_linux_mkdir,
3180 .rmdir = afs_linux_rmdir,
3181 .rename = afs_linux_rename,
3182 .getattr = afs_linux_getattr,
3183 .permission = afs_linux_permission,
3184 #ifndef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
3185 .follow_link = afs_linux_dir_follow_link,
3189 /* We really need a separate symlink set of ops, since do_follow_link()
3190 * determines if it _is_ a link by checking if the follow_link op is set.
3192 #if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
3194 afs_symlink_filler(struct file *file, struct page *page)
3196 struct inode *ip = (struct inode *)page->mapping->host;
3197 char *p = (char *)kmap(page);
3201 code = afs_linux_ireadlink(ip, p, PAGE_SIZE, AFS_UIOSYS);
3206 p[code] = '\0'; /* null terminate? */
3208 SetPageUptodate(page);
3220 static struct address_space_operations afs_symlink_aops = {
3221 .readpage = afs_symlink_filler
3223 #endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
3225 static struct inode_operations afs_symlink_iops = {
3226 #if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
3227 .readlink = page_readlink,
3228 # if defined(HAVE_LINUX_PAGE_GET_LINK)
3229 .get_link = page_get_link,
3230 # elif defined(HAVE_LINUX_PAGE_FOLLOW_LINK)
3231 .follow_link = page_follow_link,
3233 .follow_link = page_follow_link_light,
3234 .put_link = page_put_link,
3236 #else /* !defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE) */
3237 .readlink = afs_linux_readlink,
3238 .follow_link = afs_linux_follow_link,
3239 .put_link = afs_linux_put_link,
3240 #endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
3241 .setattr = afs_notify_change,
3245 afs_fill_inode(struct inode *ip, struct vattr *vattr)
3248 vattr2inode(ip, vattr);
3250 #ifdef STRUCT_ADDRESS_SPACE_HAS_BACKING_DEV_INFO
3251 ip->i_mapping->backing_dev_info = afs_backing_dev_info;
3253 /* Reset ops if symlink or directory. */
3254 if (S_ISREG(ip->i_mode)) {
3255 ip->i_op = &afs_file_iops;
3256 ip->i_fop = &afs_file_fops;
3257 ip->i_data.a_ops = &afs_file_aops;
3259 } else if (S_ISDIR(ip->i_mode)) {
3260 ip->i_op = &afs_dir_iops;
3261 ip->i_fop = &afs_dir_fops;
3263 } else if (S_ISLNK(ip->i_mode)) {
3264 ip->i_op = &afs_symlink_iops;
3265 #if defined(HAVE_LINUX_INODE_NOHIGHMEM)
3266 inode_nohighmem(ip);
3268 #if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
3269 ip->i_data.a_ops = &afs_symlink_aops;
3270 ip->i_mapping = &ip->i_data;