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 && afs_IsDCacheFresh(tdc, avc)) {
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 || !afs_IsDCacheFresh(tdc, avc)) {
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_Trace4(afs_iclSetp, CM_TRACE_GMAP, ICL_TYPE_POINTER, vcp,
537 ICL_TYPE_POINTER, vmap->vm_start, ICL_TYPE_LONG,
538 vmap->vm_end - vmap->vm_start, ICL_TYPE_LONG, 0);
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_nsec = 0;
1013 if (iattrp->ia_valid & ATTR_MTIME) {
1014 vattrp->va_mtime.tv_sec = iattrp->ia_mtime.tv_sec;
1015 vattrp->va_mtime.tv_nsec = 0;
1017 if (iattrp->ia_valid & ATTR_CTIME) {
1018 vattrp->va_ctime.tv_sec = iattrp->ia_ctime.tv_sec;
1019 vattrp->va_ctime.tv_nsec = 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);
1145 filter_enoent(int code)
1147 #ifdef HAVE_LINUX_FATAL_SIGNAL_PENDING
1148 if (code == ENOENT && fatal_signal_pending(current)) {
1155 #ifndef D_SPLICE_ALIAS_RACE
1157 static inline void dentry_race_lock(void) {}
1158 static inline void dentry_race_unlock(void) {}
1162 # if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,16)
1163 static DEFINE_MUTEX(dentry_race_sem);
1165 static DECLARE_MUTEX(dentry_race_sem);
1169 dentry_race_lock(void)
1171 mutex_lock(&dentry_race_sem);
1174 dentry_race_unlock(void)
1176 mutex_unlock(&dentry_race_sem);
1179 /* Leave some trace that this code is enabled; otherwise it's pretty hard to
1181 static __attribute__((used)) const char dentry_race_marker[] = "d_splice_alias race workaround enabled";
1184 check_dentry_race(struct dentry *dp)
1188 /* In Linux, before commit 4919c5e45a91b5db5a41695fe0357fbdff0d5767,
1189 * d_splice_alias can momentarily hash a dentry before it's fully
1190 * populated. This only happens for a moment, since it's unhashed again
1191 * right after (in d_move), but this can make the dentry be found by
1192 * __d_lookup, and then given to us.
1194 * So check if the dentry is unhashed; if it is, then the dentry is not
1195 * valid. We lock dentry_race_lock() to ensure that d_splice_alias is
1196 * no longer running. Locking d_lock is required to check the dentry's
1197 * flags, so lock that, too.
1200 spin_lock(&dp->d_lock);
1201 if (d_unhashed(dp)) {
1204 spin_unlock(&dp->d_lock);
1205 dentry_race_unlock();
1209 #endif /* D_SPLICE_ALIAS_RACE */
1211 /* Validate a dentry. Return 1 if unchanged, 0 if VFS layer should re-evaluate.
1212 * In kernels 2.2.10 and above, we are passed an additional flags var which
1213 * may have either the LOOKUP_FOLLOW OR LOOKUP_DIRECTORY set in which case
1214 * we are advised to follow the entry if it is a link or to make sure that
1215 * it is a directory. But since the kernel itself checks these possibilities
1216 * later on, we shouldn't have to do it until later. Perhaps in the future..
1218 * The code here assumes that on entry the global lock is not held
1221 #if defined(DOP_REVALIDATE_TAKES_UNSIGNED)
1222 afs_linux_dentry_revalidate(struct dentry *dp, unsigned int flags)
1223 #elif defined(DOP_REVALIDATE_TAKES_NAMEIDATA)
1224 afs_linux_dentry_revalidate(struct dentry *dp, struct nameidata *nd)
1226 afs_linux_dentry_revalidate(struct dentry *dp, int flags)
1229 cred_t *credp = NULL;
1230 struct vcache *vcp, *pvcp, *tvc = NULL;
1231 struct dentry *parent;
1233 struct afs_fakestat_state fakestate;
1235 afs_uint32 parent_dv;
1238 /* We don't support RCU path walking */
1239 # if defined(DOP_REVALIDATE_TAKES_UNSIGNED)
1240 if (flags & LOOKUP_RCU)
1242 if (nd->flags & LOOKUP_RCU)
1247 #ifdef D_SPLICE_ALIAS_RACE
1248 if (check_dentry_race(dp)) {
1255 afs_InitFakeStat(&fakestate);
1258 vcp = VTOAFS(dp->d_inode);
1260 if (vcp == afs_globalVp)
1263 if (vcp->mvstat == AFS_MVSTAT_MTPT) {
1264 if (vcp->mvid.target_root && (vcp->f.states & CMValid)) {
1265 int tryEvalOnly = 0;
1267 struct vrequest *treq = NULL;
1271 code = afs_CreateReq(&treq, credp);
1275 if ((strcmp(dp->d_name.name, ".directory") == 0)) {
1279 code = afs_TryEvalFakeStat(&vcp, &fakestate, treq);
1281 code = afs_EvalFakeStat(&vcp, &fakestate, treq);
1282 afs_DestroyReq(treq);
1283 if ((tryEvalOnly && vcp->mvstat == AFS_MVSTAT_MTPT) || code) {
1284 /* a mount point, not yet replaced by its directory */
1288 } else if (vcp->mvstat == AFS_MVSTAT_ROOT && *dp->d_name.name != '/') {
1289 osi_Assert(vcp->mvid.parent != NULL);
1292 parent = dget_parent(dp);
1293 pvcp = VTOAFS(parent->d_inode);
1294 parent_dv = parent_vcache_dv(parent->d_inode, credp);
1296 /* If the parent's DataVersion has changed or the vnode
1297 * is longer valid, we need to do a full lookup. VerifyVCache
1298 * isn't enough since the vnode may have been renamed.
1301 if (parent_dv > dp->d_time || !(vcp->f.states & CStatd)) {
1302 struct vattr *vattr = NULL;
1306 if (credp == NULL) {
1309 code = afs_lookup(pvcp, (char *)dp->d_name.name, &tvc, credp);
1310 code = filter_enoent(code);
1313 /* We couldn't perform the lookup, so we're not okay. */
1316 } else if (tvc == vcp) {
1317 /* We got back the same vcache, so we're good. */
1320 } else if (tvc == VTOAFS(dp->d_inode)) {
1321 /* We got back the same vcache, so we're good. This is
1322 * different from the above case, because sometimes 'vcp' is
1323 * not the same as the vcache for dp->d_inode, if 'vcp' was a
1324 * mtpt and we evaluated it to a root dir. In rare cases,
1325 * afs_lookup might not evalute the mtpt when we do, or vice
1326 * versa, so the previous case will not succeed. But this is
1327 * still 'correct', so make sure not to mark the dentry as
1328 * invalid; it still points to the same thing! */
1332 /* We got back a different file, so we're definitely not
1339 /* Force unhash; the name doesn't point to this file
1342 if (code && code != ENOENT) {
1343 /* ...except if we couldn't perform the actual lookup,
1344 * we don't know if the name points to this file or not. */
1350 code = afs_CreateAttr(&vattr);
1356 if (afs_getattr(vcp, vattr, credp)) {
1358 afs_DestroyAttr(vattr);
1362 vattr2inode(AFSTOV(vcp), vattr);
1363 dp->d_time = parent_dv;
1365 afs_DestroyAttr(vattr);
1368 /* should we always update the attributes at this point? */
1369 /* unlikely--the vcache entry hasn't changed */
1375 /* 'dp' represents a cached negative lookup. */
1377 parent = dget_parent(dp);
1378 pvcp = VTOAFS(parent->d_inode);
1379 parent_dv = parent_vcache_dv(parent->d_inode, credp);
1381 if (parent_dv > dp->d_time || !(pvcp->f.states & CStatd)
1382 || afs_IsDynroot(pvcp)) {
1396 #ifndef D_INVALIDATE_IS_VOID
1397 /* When (v3.18) d_invalidate was converted to void, it also started
1398 * being called automatically from revalidate, and automatically
1400 * - shrink_dcache_parent
1401 * - automatic detach of submounts
1403 * Therefore, after that point, OpenAFS revalidate logic no longer needs
1404 * to do any of those things itself for invalid dentry structs. We only need
1405 * to tell VFS it's invalid (by returning 0), and VFS will handle the rest.
1407 if (have_submounts(dp))
1415 afs_PutFakeStat(&fakestate);
1420 #ifndef D_INVALIDATE_IS_VOID
1423 * If we had a negative lookup for the name we want to forcibly
1424 * unhash the dentry.
1425 * Otherwise use d_invalidate which will not unhash it if still in use.
1428 shrink_dcache_parent(dp);
1439 afs_dentry_iput(struct dentry *dp, struct inode *ip)
1441 struct vcache *vcp = VTOAFS(ip);
1442 int haveGlock = ISAFS_GLOCK();
1448 if (!AFS_IS_DISCONNECTED || (vcp->f.states & CUnlinked)) {
1449 (void) afs_InactiveVCache(vcp, NULL);
1456 afs_linux_clear_nfsfs_renamed(dp);
1462 #if defined(DOP_D_DELETE_TAKES_CONST)
1463 afs_dentry_delete(const struct dentry *dp)
1465 afs_dentry_delete(struct dentry *dp)
1468 if (dp->d_inode && (VTOAFS(dp->d_inode)->f.states & CUnlinked))
1469 return 1; /* bad inode? */
1474 #ifdef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
1475 static struct vfsmount *
1476 afs_dentry_automount(afs_linux_path_t *path)
1478 struct dentry *target;
1481 * Avoid symlink resolution limits when resolving; we cannot contribute to
1482 * an infinite symlink loop.
1484 * On newer kernels the field has moved to the private nameidata structure
1485 * so we can't adjust it here. This may cause ELOOP when using a path with
1486 * 40 or more directories that are not already in the dentry cache.
1488 #if defined(STRUCT_TASK_STRUCT_HAS_TOTAL_LINK_COUNT)
1489 current->total_link_count--;
1492 target = canonical_dentry(path->dentry->d_inode);
1494 if (target == path->dentry) {
1501 path->dentry = target;
1504 spin_lock(&path->dentry->d_lock);
1505 path->dentry->d_flags &= ~DCACHE_NEED_AUTOMOUNT;
1506 spin_unlock(&path->dentry->d_lock);
1511 #endif /* STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT */
1513 struct dentry_operations afs_dentry_operations = {
1514 .d_revalidate = afs_linux_dentry_revalidate,
1515 .d_delete = afs_dentry_delete,
1516 .d_iput = afs_dentry_iput,
1517 #ifdef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
1518 .d_automount = afs_dentry_automount,
1519 #endif /* STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT */
1522 /**********************************************************************
1523 * AFS Linux inode operations
1524 **********************************************************************/
1528 * Merely need to set enough of vattr to get us through the create. Note
1529 * that the higher level code (open_namei) will take care of any tuncation
1530 * explicitly. Exclusive open is also taken care of in open_namei.
1532 * name is in kernel space at this point.
1535 #if defined(IOP_CREATE_TAKES_BOOL)
1536 afs_linux_create(struct inode *dip, struct dentry *dp, umode_t mode,
1538 #elif defined(IOP_CREATE_TAKES_UMODE_T)
1539 afs_linux_create(struct inode *dip, struct dentry *dp, umode_t mode,
1540 struct nameidata *nd)
1541 #elif defined(IOP_CREATE_TAKES_NAMEIDATA)
1542 afs_linux_create(struct inode *dip, struct dentry *dp, int mode,
1543 struct nameidata *nd)
1545 afs_linux_create(struct inode *dip, struct dentry *dp, int mode)
1548 struct vattr *vattr = NULL;
1549 cred_t *credp = crref();
1550 const char *name = dp->d_name.name;
1556 code = afs_CreateAttr(&vattr);
1560 vattr->va_mode = mode;
1561 vattr->va_type = mode & S_IFMT;
1563 code = afs_create(VTOAFS(dip), (char *)name, vattr, NONEXCL, mode,
1567 struct inode *ip = AFSTOV(vcp);
1569 afs_getattr(vcp, vattr, credp);
1570 afs_fill_inode(ip, vattr);
1571 insert_inode_hash(ip);
1572 #if !defined(STRUCT_SUPER_BLOCK_HAS_S_D_OP)
1573 dp->d_op = &afs_dentry_operations;
1575 dp->d_time = parent_vcache_dv(dip, credp);
1576 d_instantiate(dp, ip);
1579 afs_DestroyAttr(vattr);
1585 return afs_convert_code(code);
1588 /* afs_linux_lookup */
1589 static struct dentry *
1590 #if defined(IOP_LOOKUP_TAKES_UNSIGNED)
1591 afs_linux_lookup(struct inode *dip, struct dentry *dp,
1593 #elif defined(IOP_LOOKUP_TAKES_NAMEIDATA)
1594 afs_linux_lookup(struct inode *dip, struct dentry *dp,
1595 struct nameidata *nd)
1597 afs_linux_lookup(struct inode *dip, struct dentry *dp)
1600 cred_t *credp = crref();
1601 struct vcache *vcp = NULL;
1602 const char *comp = dp->d_name.name;
1603 struct inode *ip = NULL;
1604 struct dentry *newdp = NULL;
1609 code = afs_lookup(VTOAFS(dip), (char *)comp, &vcp, credp);
1610 code = filter_enoent(code);
1611 if (code == ENOENT) {
1612 /* It's ok for the file to not be found. That's noted by the caller by
1613 * seeing that the dp->d_inode field is NULL (set by d_splice_alias or
1616 osi_Assert(vcp == NULL);
1624 struct vattr *vattr = NULL;
1625 struct vcache *parent_vc = VTOAFS(dip);
1627 if (parent_vc == vcp) {
1628 /* This is possible if the parent dir is a mountpoint to a volume,
1629 * and the dir entry we looked up is a mountpoint to the same
1630 * volume. Linux cannot cope with this, so return an error instead
1631 * of risking a deadlock or panic. */
1638 code = afs_CreateAttr(&vattr);
1646 afs_getattr(vcp, vattr, credp);
1647 afs_fill_inode(ip, vattr);
1648 if (hlist_unhashed(&ip->i_hash))
1649 insert_inode_hash(ip);
1651 afs_DestroyAttr(vattr);
1653 #if !defined(STRUCT_SUPER_BLOCK_HAS_S_D_OP)
1654 dp->d_op = &afs_dentry_operations;
1656 dp->d_time = parent_vcache_dv(dip, credp);
1660 if (ip && S_ISDIR(ip->i_mode)) {
1661 d_prune_aliases(ip);
1663 #ifdef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
1664 /* Only needed if this is a volume root */
1665 if (vcp->mvstat == 2)
1666 ip->i_flags |= S_AUTOMOUNT;
1670 * Take an extra reference so the inode doesn't go away if
1671 * d_splice_alias drops our reference on error.
1674 #ifdef HAVE_LINUX_IHOLD
1681 newdp = d_splice_alias(ip, dp);
1682 dentry_race_unlock();
1687 if (IS_ERR(newdp)) {
1688 /* d_splice_alias can return an error (EIO) if there is an existing
1689 * connected directory alias for this dentry. Add our dentry manually
1690 * ourselves if this happens. */
1693 #if defined(D_SPLICE_ALIAS_LEAK_ON_ERROR)
1694 /* Depending on the kernel version, d_splice_alias may or may not drop
1695 * the inode reference on error. If it didn't, do it here. */
1704 return ERR_PTR(afs_convert_code(code));
1712 afs_linux_link(struct dentry *olddp, struct inode *dip, struct dentry *newdp)
1715 cred_t *credp = crref();
1716 const char *name = newdp->d_name.name;
1717 struct inode *oldip = olddp->d_inode;
1719 /* If afs_link returned the vnode, we could instantiate the
1720 * dentry. Since it's not, we drop this one and do a new lookup.
1725 code = afs_link(VTOAFS(oldip), VTOAFS(dip), (char *)name, credp);
1729 return afs_convert_code(code);
1732 /* We have to have a Linux specific sillyrename function, because we
1733 * also have to keep the dcache up to date when we're doing a silly
1734 * rename - so we don't want the generic vnodeops doing this behind our
1739 afs_linux_sillyrename(struct inode *dir, struct dentry *dentry,
1742 struct vcache *tvc = VTOAFS(dentry->d_inode);
1743 struct dentry *__dp = NULL;
1744 char *__name = NULL;
1747 if (afs_linux_nfsfs_renamed(dentry))
1755 osi_FreeSmallSpace(__name);
1756 __name = afs_newname();
1759 __dp = lookup_one_len(__name, dentry->d_parent, strlen(__name));
1762 osi_FreeSmallSpace(__name);
1765 } while (__dp->d_inode != NULL);
1768 code = afs_rename(VTOAFS(dir), (char *)dentry->d_name.name,
1769 VTOAFS(dir), (char *)__dp->d_name.name,
1772 tvc->mvid.silly_name = __name;
1775 crfree(tvc->uncred);
1777 tvc->uncred = credp;
1778 tvc->f.states |= CUnlinked;
1779 afs_linux_set_nfsfs_renamed(dentry);
1781 __dp->d_time = 0; /* force to revalidate */
1782 d_move(dentry, __dp);
1784 osi_FreeSmallSpace(__name);
1795 afs_linux_unlink(struct inode *dip, struct dentry *dp)
1798 cred_t *credp = crref();
1799 const char *name = dp->d_name.name;
1800 struct vcache *tvc = VTOAFS(dp->d_inode);
1802 if (VREFCOUNT(tvc) > 1 && tvc->opens > 0
1803 && !(tvc->f.states & CUnlinked)) {
1805 code = afs_linux_sillyrename(dip, dp, credp);
1808 code = afs_remove(VTOAFS(dip), (char *)name, credp);
1815 return afs_convert_code(code);
1820 afs_linux_symlink(struct inode *dip, struct dentry *dp, const char *target)
1823 cred_t *credp = crref();
1824 struct vattr *vattr = NULL;
1825 const char *name = dp->d_name.name;
1827 /* If afs_symlink returned the vnode, we could instantiate the
1828 * dentry. Since it's not, we drop this one and do a new lookup.
1833 code = afs_CreateAttr(&vattr);
1838 code = afs_symlink(VTOAFS(dip), (char *)name, vattr, (char *)target, NULL,
1840 afs_DestroyAttr(vattr);
1845 return afs_convert_code(code);
1849 #if defined(IOP_MKDIR_TAKES_UMODE_T)
1850 afs_linux_mkdir(struct inode *dip, struct dentry *dp, umode_t mode)
1852 afs_linux_mkdir(struct inode *dip, struct dentry *dp, int mode)
1856 cred_t *credp = crref();
1857 struct vcache *tvcp = NULL;
1858 struct vattr *vattr = NULL;
1859 const char *name = dp->d_name.name;
1862 code = afs_CreateAttr(&vattr);
1867 vattr->va_mask = ATTR_MODE;
1868 vattr->va_mode = mode;
1870 code = afs_mkdir(VTOAFS(dip), (char *)name, vattr, &tvcp, credp);
1873 struct inode *ip = AFSTOV(tvcp);
1875 afs_getattr(tvcp, vattr, credp);
1876 afs_fill_inode(ip, vattr);
1878 #if !defined(STRUCT_SUPER_BLOCK_HAS_S_D_OP)
1879 dp->d_op = &afs_dentry_operations;
1881 dp->d_time = parent_vcache_dv(dip, credp);
1882 d_instantiate(dp, ip);
1884 afs_DestroyAttr(vattr);
1890 return afs_convert_code(code);
1894 afs_linux_rmdir(struct inode *dip, struct dentry *dp)
1897 cred_t *credp = crref();
1898 const char *name = dp->d_name.name;
1900 /* locking kernel conflicts with glock? */
1903 code = afs_rmdir(VTOAFS(dip), (char *)name, credp);
1906 /* Linux likes to see ENOTEMPTY returned from an rmdir() syscall
1907 * that failed because a directory is not empty. So, we map
1908 * EEXIST to ENOTEMPTY on linux.
1910 if (code == EEXIST) {
1919 return afs_convert_code(code);
1924 afs_linux_rename(struct inode *oldip, struct dentry *olddp,
1925 struct inode *newip, struct dentry *newdp
1926 #ifdef HAVE_LINUX_INODE_OPERATIONS_RENAME_TAKES_FLAGS
1927 , unsigned int flags
1932 cred_t *credp = crref();
1933 const char *oldname = olddp->d_name.name;
1934 const char *newname = newdp->d_name.name;
1935 struct dentry *rehash = NULL;
1937 #ifdef HAVE_LINUX_INODE_OPERATIONS_RENAME_TAKES_FLAGS
1939 return -EINVAL; /* no support for new flags yet */
1942 /* Prevent any new references during rename operation. */
1944 if (!d_unhashed(newdp)) {
1949 afs_maybe_shrink_dcache(olddp);
1952 code = afs_rename(VTOAFS(oldip), (char *)oldname, VTOAFS(newip), (char *)newname, credp);
1956 olddp->d_time = 0; /* force to revalidate */
1962 return afs_convert_code(code);
1966 /* afs_linux_ireadlink
1967 * Internal readlink which can return link contents to user or kernel space.
1968 * Note that the buffer is NOT supposed to be null-terminated.
1971 afs_linux_ireadlink(struct inode *ip, char *target, int maxlen, uio_seg_t seg)
1974 cred_t *credp = crref();
1978 memset(&tuio, 0, sizeof(tuio));
1979 memset(&iov, 0, sizeof(iov));
1981 setup_uio(&tuio, &iov, target, (afs_offs_t) 0, maxlen, UIO_READ, seg);
1982 code = afs_readlink(VTOAFS(ip), &tuio, credp);
1986 return maxlen - tuio.uio_resid;
1988 return afs_convert_code(code);
1991 #if !defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
1992 /* afs_linux_readlink
1993 * Fill target (which is in user space) with contents of symlink.
1996 afs_linux_readlink(struct dentry *dp, char *target, int maxlen)
1999 struct inode *ip = dp->d_inode;
2002 code = afs_linux_ireadlink(ip, target, maxlen, AFS_UIOUSER);
2008 /* afs_linux_follow_link
2009 * a file system dependent link following routine.
2011 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2012 static const char *afs_linux_follow_link(struct dentry *dentry, void **link_data)
2014 static int afs_linux_follow_link(struct dentry *dentry, struct nameidata *nd)
2020 name = kmalloc(PATH_MAX, GFP_NOFS);
2022 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2023 return ERR_PTR(-EIO);
2030 code = afs_linux_ireadlink(dentry->d_inode, name, PATH_MAX - 1, AFS_UIOSYS);
2034 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2035 return ERR_PTR(code);
2042 #if defined(HAVE_LINUX_INODE_OPERATIONS_FOLLOW_LINK_NO_NAMEIDATA)
2043 return *link_data = name;
2045 nd_set_link(nd, name);
2050 #if defined(HAVE_LINUX_INODE_OPERATIONS_PUT_LINK_NO_NAMEIDATA)
2052 afs_linux_put_link(struct inode *inode, void *link_data)
2054 char *name = link_data;
2056 if (name && !IS_ERR(name))
2061 afs_linux_put_link(struct dentry *dentry, struct nameidata *nd)
2063 char *name = nd_get_link(nd);
2065 if (name && !IS_ERR(name))
2068 #endif /* HAVE_LINUX_INODE_OPERATIONS_PUT_LINK_NO_NAMEIDATA */
2070 #endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
2072 /* Populate a page by filling it from the cache file pointed at by cachefp
2073 * (which contains indicated chunk)
2074 * If task is NULL, the page copy occurs syncronously, and the routine
2075 * returns with page still locked. If task is non-NULL, then page copies
2076 * may occur in the background, and the page will be unlocked when it is
2077 * ready for use. Note that if task is non-NULL and we encounter an error
2078 * before we start the background copy, we MUST unlock 'page' before we return.
2081 afs_linux_read_cache(struct file *cachefp, struct page *page,
2082 int chunk, struct pagevec *lrupv,
2083 struct afs_pagecopy_task *task) {
2084 loff_t offset = page_offset(page);
2085 struct inode *cacheinode = cachefp->f_dentry->d_inode;
2086 struct page *newpage, *cachepage;
2087 struct address_space *cachemapping;
2091 cachemapping = cacheinode->i_mapping;
2095 /* If we're trying to read a page that's past the end of the disk
2096 * cache file, then just return a zeroed page */
2097 if (AFS_CHUNKOFFSET(offset) >= i_size_read(cacheinode)) {
2098 zero_user_segment(page, 0, PAGE_SIZE);
2099 SetPageUptodate(page);
2105 /* From our offset, we now need to work out which page in the disk
2106 * file it corresponds to. This will be fun ... */
2107 pageindex = (offset - AFS_CHUNKTOBASE(chunk)) >> PAGE_SHIFT;
2109 while (cachepage == NULL) {
2110 cachepage = find_get_page(cachemapping, pageindex);
2113 newpage = page_cache_alloc(cachemapping);
2119 code = add_to_page_cache(newpage, cachemapping,
2120 pageindex, GFP_KERNEL);
2122 cachepage = newpage;
2125 get_page(cachepage);
2126 if (!pagevec_add(lrupv, cachepage))
2127 __pagevec_lru_add_file(lrupv);
2132 if (code != -EEXIST)
2136 lock_page(cachepage);
2140 if (!PageUptodate(cachepage)) {
2141 ClearPageError(cachepage);
2142 /* Note that ->readpage always handles unlocking the given page, even
2143 * when an error is returned. */
2144 code = cachemapping->a_ops->readpage(NULL, cachepage);
2145 if (!code && !task) {
2146 wait_on_page_locked(cachepage);
2149 unlock_page(cachepage);
2153 if (PageUptodate(cachepage)) {
2154 copy_highpage(page, cachepage);
2155 flush_dcache_page(page);
2156 SetPageUptodate(page);
2161 afs_pagecopy_queue_page(task, cachepage, page);
2173 put_page(cachepage);
2179 afs_linux_readpage_fastpath(struct file *fp, struct page *pp, int *codep)
2181 loff_t offset = page_offset(pp);
2182 struct inode *ip = FILE_INODE(fp);
2183 struct vcache *avc = VTOAFS(ip);
2185 struct file *cacheFp = NULL;
2188 struct pagevec lrupv;
2190 /* Not a UFS cache, don't do anything */
2191 if (cacheDiskType != AFS_FCACHE_TYPE_UFS)
2194 /* No readpage (ex: tmpfs) , skip */
2195 if (cachefs_noreadpage)
2198 /* Can't do anything if the vcache isn't statd , or if the read
2199 * crosses a chunk boundary.
2201 if (!(avc->f.states & CStatd) ||
2202 AFS_CHUNK(offset) != AFS_CHUNK(offset + PAGE_SIZE)) {
2206 ObtainWriteLock(&avc->lock, 911);
2208 /* XXX - See if hinting actually makes things faster !!! */
2210 /* See if we have a suitable entry already cached */
2214 /* We need to lock xdcache, then dcache, to handle situations where
2215 * the hint is on the free list. However, we can't safely do this
2216 * according to the locking hierarchy. So, use a non blocking lock.
2218 ObtainReadLock(&afs_xdcache);
2219 dcLocked = ( 0 == NBObtainReadLock(&tdc->lock));
2221 if (dcLocked && (tdc->index != NULLIDX)
2222 && !FidCmp(&tdc->f.fid, &avc->f.fid)
2223 && tdc->f.chunk == AFS_CHUNK(offset)
2224 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
2225 /* Bonus - the hint was correct */
2228 /* Only destroy the hint if its actually invalid, not if there's
2229 * just been a locking failure */
2231 ReleaseReadLock(&tdc->lock);
2238 ReleaseReadLock(&afs_xdcache);
2241 /* No hint, or hint is no longer valid - see if we can get something
2242 * directly from the dcache
2245 tdc = afs_FindDCache(avc, offset);
2248 ReleaseWriteLock(&avc->lock);
2253 ObtainReadLock(&tdc->lock);
2255 /* Is the dcache we've been given currently up to date */
2256 if (!afs_IsDCacheFresh(tdc, avc) ||
2257 (tdc->dflags & DFFetching))
2260 /* Update our hint for future abuse */
2263 /* Okay, so we've now got a cache file that is up to date */
2265 /* XXX - I suspect we should be locking the inodes before we use them! */
2267 cacheFp = afs_linux_raw_open(&tdc->f.inode);
2268 osi_Assert(cacheFp);
2269 if (!cacheFp->f_dentry->d_inode->i_mapping->a_ops->readpage) {
2270 cachefs_noreadpage = 1;
2274 #if defined(PAGEVEC_INIT_COLD_ARG)
2275 pagevec_init(&lrupv, 0);
2277 pagevec_init(&lrupv);
2280 code = afs_linux_read_cache(cacheFp, pp, tdc->f.chunk, &lrupv, NULL);
2282 if (pagevec_count(&lrupv))
2283 __pagevec_lru_add_file(&lrupv);
2285 filp_close(cacheFp, NULL);
2288 ReleaseReadLock(&tdc->lock);
2289 ReleaseWriteLock(&avc->lock);
2296 ReleaseWriteLock(&avc->lock);
2297 ReleaseReadLock(&tdc->lock);
2302 /* afs_linux_readpage
2304 * This function is split into two, because prepare_write/begin_write
2305 * require a readpage call which doesn't unlock the resulting page upon
2309 afs_linux_fillpage(struct file *fp, struct page *pp)
2314 struct iovec *iovecp;
2315 struct inode *ip = FILE_INODE(fp);
2316 afs_int32 cnt = page_count(pp);
2317 struct vcache *avc = VTOAFS(ip);
2318 afs_offs_t offset = page_offset(pp);
2322 if (afs_linux_readpage_fastpath(fp, pp, &code)) {
2332 auio = kmalloc(sizeof(struct uio), GFP_NOFS);
2333 iovecp = kmalloc(sizeof(struct iovec), GFP_NOFS);
2335 setup_uio(auio, iovecp, (char *)address, offset, PAGE_SIZE, UIO_READ,
2340 afs_Trace4(afs_iclSetp, CM_TRACE_READPAGE, ICL_TYPE_POINTER, ip,
2341 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, cnt, ICL_TYPE_INT32,
2342 99999); /* not a possible code value */
2344 code = afs_rdwr(avc, auio, UIO_READ, 0, credp);
2346 afs_Trace4(afs_iclSetp, CM_TRACE_READPAGE, ICL_TYPE_POINTER, ip,
2347 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, cnt, ICL_TYPE_INT32,
2349 AFS_DISCON_UNLOCK();
2352 /* XXX valid for no-cache also? Check last bits of files... :)
2353 * Cognate code goes in afs_NoCacheFetchProc. */
2354 if (auio->uio_resid) /* zero remainder of page */
2355 memset((void *)(address + (PAGE_SIZE - auio->uio_resid)), 0,
2358 flush_dcache_page(pp);
2359 SetPageUptodate(pp);
2368 return afs_convert_code(code);
2372 afs_linux_prefetch(struct file *fp, struct page *pp)
2375 struct vcache *avc = VTOAFS(FILE_INODE(fp));
2376 afs_offs_t offset = page_offset(pp);
2378 if (AFS_CHUNKOFFSET(offset) == 0) {
2380 struct vrequest *treq = NULL;
2385 code = afs_CreateReq(&treq, credp);
2386 if (!code && !NBObtainWriteLock(&avc->lock, 534)) {
2387 tdc = afs_FindDCache(avc, offset);
2389 if (!(tdc->mflags & DFNextStarted))
2390 afs_PrefetchChunk(avc, tdc, credp, treq);
2393 ReleaseWriteLock(&avc->lock);
2395 afs_DestroyReq(treq);
2399 return afs_convert_code(code);
2404 afs_linux_bypass_readpages(struct file *fp, struct address_space *mapping,
2405 struct list_head *page_list, unsigned num_pages)
2410 struct iovec* iovecp;
2411 struct nocache_read_request *ancr;
2413 struct pagevec lrupv;
2417 struct inode *ip = FILE_INODE(fp);
2418 struct vcache *avc = VTOAFS(ip);
2419 afs_int32 base_index = 0;
2420 afs_int32 page_count = 0;
2423 /* background thread must free: iovecp, auio, ancr */
2424 iovecp = osi_Alloc(num_pages * sizeof(struct iovec));
2426 auio = osi_Alloc(sizeof(struct uio));
2427 auio->uio_iov = iovecp;
2428 auio->uio_iovcnt = num_pages;
2429 auio->uio_flag = UIO_READ;
2430 auio->uio_seg = AFS_UIOSYS;
2431 auio->uio_resid = num_pages * PAGE_SIZE;
2433 ancr = osi_Alloc(sizeof(struct nocache_read_request));
2435 ancr->offset = auio->uio_offset;
2436 ancr->length = auio->uio_resid;
2438 #if defined(PAGEVEC_INIT_COLD_ARG)
2439 pagevec_init(&lrupv, 0);
2441 pagevec_init(&lrupv);
2444 for(page_ix = 0; page_ix < num_pages; ++page_ix) {
2446 if(list_empty(page_list))
2449 pp = list_entry(page_list->prev, struct page, lru);
2450 /* If we allocate a page and don't remove it from page_list,
2451 * the page cache gets upset. */
2453 isize = (i_size_read(fp->f_mapping->host) - 1) >> PAGE_SHIFT;
2454 if(pp->index > isize) {
2461 offset = page_offset(pp);
2462 ancr->offset = auio->uio_offset = offset;
2463 base_index = pp->index;
2465 iovecp[page_ix].iov_len = PAGE_SIZE;
2466 code = add_to_page_cache(pp, mapping, pp->index, GFP_KERNEL);
2467 if(base_index != pp->index) {
2471 iovecp[page_ix].iov_base = (void *) 0;
2473 ancr->length -= PAGE_SIZE;
2481 iovecp[page_ix].iov_base = (void *) 0;
2484 if(!PageLocked(pp)) {
2488 /* increment page refcount--our original design assumed
2489 * that locking it would effectively pin it; protect
2490 * ourselves from the possiblity that this assumption is
2491 * is faulty, at low cost (provided we do not fail to
2492 * do the corresponding decref on the other side) */
2495 /* save the page for background map */
2496 iovecp[page_ix].iov_base = (void*) pp;
2498 /* and put it on the LRU cache */
2499 if (!pagevec_add(&lrupv, pp))
2500 __pagevec_lru_add_file(&lrupv);
2504 /* If there were useful pages in the page list, make sure all pages
2505 * are in the LRU cache, then schedule the read */
2507 if (pagevec_count(&lrupv))
2508 __pagevec_lru_add_file(&lrupv);
2510 code = afs_ReadNoCache(avc, ancr, credp);
2513 /* If there is nothing for the background thread to handle,
2514 * it won't be freeing the things that we never gave it */
2515 osi_Free(iovecp, num_pages * sizeof(struct iovec));
2516 osi_Free(auio, sizeof(struct uio));
2517 osi_Free(ancr, sizeof(struct nocache_read_request));
2519 /* we do not flush, release, or unmap pages--that will be
2520 * done for us by the background thread as each page comes in
2521 * from the fileserver */
2522 return afs_convert_code(code);
2527 afs_linux_bypass_readpage(struct file *fp, struct page *pp)
2529 cred_t *credp = NULL;
2531 struct iovec *iovecp;
2532 struct nocache_read_request *ancr;
2536 * Special case: if page is at or past end of file, just zero it and set
2539 if (page_offset(pp) >= i_size_read(fp->f_mapping->host)) {
2540 zero_user_segment(pp, 0, PAGE_SIZE);
2541 SetPageUptodate(pp);
2548 /* receiver frees */
2549 auio = osi_Alloc(sizeof(struct uio));
2550 iovecp = osi_Alloc(sizeof(struct iovec));
2552 /* address can be NULL, because we overwrite it with 'pp', below */
2553 setup_uio(auio, iovecp, NULL, page_offset(pp),
2554 PAGE_SIZE, UIO_READ, AFS_UIOSYS);
2556 /* save the page for background map */
2557 get_page(pp); /* see above */
2558 auio->uio_iov->iov_base = (void*) pp;
2559 /* the background thread will free this */
2560 ancr = osi_Alloc(sizeof(struct nocache_read_request));
2562 ancr->offset = page_offset(pp);
2563 ancr->length = PAGE_SIZE;
2566 code = afs_ReadNoCache(VTOAFS(FILE_INODE(fp)), ancr, credp);
2569 return afs_convert_code(code);
2573 afs_linux_can_bypass(struct inode *ip) {
2575 switch(cache_bypass_strategy) {
2576 case NEVER_BYPASS_CACHE:
2578 case ALWAYS_BYPASS_CACHE:
2580 case LARGE_FILES_BYPASS_CACHE:
2581 if (i_size_read(ip) > cache_bypass_threshold)
2589 /* Check if a file is permitted to bypass the cache by policy, and modify
2590 * the cache bypass state recorded for that file */
2593 afs_linux_bypass_check(struct inode *ip) {
2596 int bypass = afs_linux_can_bypass(ip);
2599 trydo_cache_transition(VTOAFS(ip), credp, bypass);
2607 afs_linux_readpage(struct file *fp, struct page *pp)
2611 if (afs_linux_bypass_check(FILE_INODE(fp))) {
2612 code = afs_linux_bypass_readpage(fp, pp);
2614 code = afs_linux_fillpage(fp, pp);
2616 code = afs_linux_prefetch(fp, pp);
2623 /* Readpages reads a number of pages for a particular file. We use
2624 * this to optimise the reading, by limiting the number of times upon which
2625 * we have to lookup, lock and open vcaches and dcaches
2629 afs_linux_readpages(struct file *fp, struct address_space *mapping,
2630 struct list_head *page_list, unsigned int num_pages)
2632 struct inode *inode = mapping->host;
2633 struct vcache *avc = VTOAFS(inode);
2635 struct file *cacheFp = NULL;
2637 unsigned int page_idx;
2639 struct pagevec lrupv;
2640 struct afs_pagecopy_task *task;
2642 if (afs_linux_bypass_check(inode))
2643 return afs_linux_bypass_readpages(fp, mapping, page_list, num_pages);
2645 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2648 /* No readpage (ex: tmpfs) , skip */
2649 if (cachefs_noreadpage)
2653 if ((code = afs_linux_VerifyVCache(avc, NULL))) {
2658 ObtainWriteLock(&avc->lock, 912);
2661 task = afs_pagecopy_init_task();
2664 #if defined(PAGEVEC_INIT_COLD_ARG)
2665 pagevec_init(&lrupv, 0);
2667 pagevec_init(&lrupv);
2669 for (page_idx = 0; page_idx < num_pages; page_idx++) {
2670 struct page *page = list_entry(page_list->prev, struct page, lru);
2671 list_del(&page->lru);
2672 offset = page_offset(page);
2674 if (tdc && tdc->f.chunk != AFS_CHUNK(offset)) {
2676 ReleaseReadLock(&tdc->lock);
2681 filp_close(cacheFp, NULL);
2686 if ((tdc = afs_FindDCache(avc, offset))) {
2687 ObtainReadLock(&tdc->lock);
2688 if (!afs_IsDCacheFresh(tdc, avc) ||
2689 (tdc->dflags & DFFetching)) {
2690 ReleaseReadLock(&tdc->lock);
2697 cacheFp = afs_linux_raw_open(&tdc->f.inode);
2698 osi_Assert(cacheFp);
2699 if (!cacheFp->f_dentry->d_inode->i_mapping->a_ops->readpage) {
2700 cachefs_noreadpage = 1;
2706 if (tdc && !add_to_page_cache(page, mapping, page->index,
2709 if (!pagevec_add(&lrupv, page))
2710 __pagevec_lru_add_file(&lrupv);
2712 /* Note that add_to_page_cache() locked 'page'.
2713 * afs_linux_read_cache() is guaranteed to handle unlocking it. */
2714 afs_linux_read_cache(cacheFp, page, tdc->f.chunk, &lrupv, task);
2718 if (pagevec_count(&lrupv))
2719 __pagevec_lru_add_file(&lrupv);
2723 filp_close(cacheFp, NULL);
2725 afs_pagecopy_put_task(task);
2729 ReleaseReadLock(&tdc->lock);
2733 ReleaseWriteLock(&avc->lock);
2738 /* Prepare an AFS vcache for writeback. Should be called with the vcache
2741 afs_linux_prepare_writeback(struct vcache *avc) {
2743 struct pagewriter *pw;
2745 pid = MyPidxx2Pid(MyPidxx);
2746 /* Prevent recursion into the writeback code */
2747 spin_lock(&avc->pagewriter_lock);
2748 list_for_each_entry(pw, &avc->pagewriters, link) {
2749 if (pw->writer == pid) {
2750 spin_unlock(&avc->pagewriter_lock);
2751 return AOP_WRITEPAGE_ACTIVATE;
2754 spin_unlock(&avc->pagewriter_lock);
2756 /* Add ourselves to writer list */
2757 pw = osi_Alloc(sizeof(struct pagewriter));
2759 spin_lock(&avc->pagewriter_lock);
2760 list_add_tail(&pw->link, &avc->pagewriters);
2761 spin_unlock(&avc->pagewriter_lock);
2767 afs_linux_dopartialwrite(struct vcache *avc, cred_t *credp) {
2768 struct vrequest *treq = NULL;
2771 if (!afs_CreateReq(&treq, credp)) {
2772 code = afs_DoPartialWrite(avc, treq);
2773 afs_DestroyReq(treq);
2776 return afs_convert_code(code);
2780 afs_linux_complete_writeback(struct vcache *avc) {
2781 struct pagewriter *pw, *store;
2783 struct list_head tofree;
2785 INIT_LIST_HEAD(&tofree);
2786 pid = MyPidxx2Pid(MyPidxx);
2787 /* Remove ourselves from writer list */
2788 spin_lock(&avc->pagewriter_lock);
2789 list_for_each_entry_safe(pw, store, &avc->pagewriters, link) {
2790 if (pw->writer == pid) {
2791 list_del(&pw->link);
2792 /* osi_Free may sleep so we need to defer it */
2793 list_add_tail(&pw->link, &tofree);
2796 spin_unlock(&avc->pagewriter_lock);
2797 list_for_each_entry_safe(pw, store, &tofree, link) {
2798 list_del(&pw->link);
2799 osi_Free(pw, sizeof(struct pagewriter));
2803 /* Writeback a given page syncronously. Called with no AFS locks held */
2805 afs_linux_page_writeback(struct inode *ip, struct page *pp,
2806 unsigned long offset, unsigned int count,
2809 struct vcache *vcp = VTOAFS(ip);
2817 memset(&tuio, 0, sizeof(tuio));
2818 memset(&iovec, 0, sizeof(iovec));
2820 buffer = kmap(pp) + offset;
2821 base = page_offset(pp) + offset;
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, 99999);
2828 setup_uio(&tuio, &iovec, buffer, base, count, UIO_WRITE, AFS_UIOSYS);
2830 code = afs_write(vcp, &tuio, f_flags, credp, 0);
2832 i_size_write(ip, vcp->f.m.Length);
2833 ip->i_blocks = ((vcp->f.m.Length + 1023) >> 10) << 1;
2835 code = code ? afs_convert_code(code) : count - tuio.uio_resid;
2837 afs_Trace4(afs_iclSetp, CM_TRACE_UPDATEPAGE, ICL_TYPE_POINTER, vcp,
2838 ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, page_count(pp),
2839 ICL_TYPE_INT32, code);
2848 afs_linux_writepage_sync(struct inode *ip, struct page *pp,
2849 unsigned long offset, unsigned int count)
2853 struct vcache *vcp = VTOAFS(ip);
2856 /* Catch recursive writeback. This occurs if the kernel decides
2857 * writeback is required whilst we are writing to the cache, or
2858 * flushing to the server. When we're running syncronously (as
2859 * opposed to from writepage) we can't actually do anything about
2860 * this case - as we can't return AOP_WRITEPAGE_ACTIVATE to write()
2863 ObtainWriteLock(&vcp->lock, 532);
2864 afs_linux_prepare_writeback(vcp);
2865 ReleaseWriteLock(&vcp->lock);
2869 code = afs_linux_page_writeback(ip, pp, offset, count, credp);
2872 ObtainWriteLock(&vcp->lock, 533);
2874 code1 = afs_linux_dopartialwrite(vcp, credp);
2875 afs_linux_complete_writeback(vcp);
2876 ReleaseWriteLock(&vcp->lock);
2887 #ifdef AOP_WRITEPAGE_TAKES_WRITEBACK_CONTROL
2888 afs_linux_writepage(struct page *pp, struct writeback_control *wbc)
2890 afs_linux_writepage(struct page *pp)
2893 struct address_space *mapping = pp->mapping;
2894 struct inode *inode;
2897 unsigned int to = PAGE_SIZE;
2904 inode = mapping->host;
2905 vcp = VTOAFS(inode);
2906 isize = i_size_read(inode);
2908 /* Don't defeat an earlier truncate */
2909 if (page_offset(pp) > isize) {
2910 set_page_writeback(pp);
2916 ObtainWriteLock(&vcp->lock, 537);
2917 code = afs_linux_prepare_writeback(vcp);
2918 if (code == AOP_WRITEPAGE_ACTIVATE) {
2919 /* WRITEPAGE_ACTIVATE is the only return value that permits us
2920 * to return with the page still locked */
2921 ReleaseWriteLock(&vcp->lock);
2926 /* Grab the creds structure currently held in the vnode, and
2927 * get a reference to it, in case it goes away ... */
2933 ReleaseWriteLock(&vcp->lock);
2936 set_page_writeback(pp);
2938 SetPageUptodate(pp);
2940 /* We can unlock the page here, because it's protected by the
2941 * page_writeback flag. This should make us less vulnerable to
2942 * deadlocking in afs_write and afs_DoPartialWrite
2946 /* If this is the final page, then just write the number of bytes that
2947 * are actually in it */
2948 if ((isize - page_offset(pp)) < to )
2949 to = isize - page_offset(pp);
2951 code = afs_linux_page_writeback(inode, pp, 0, to, credp);
2954 ObtainWriteLock(&vcp->lock, 538);
2956 /* As much as we might like to ignore a file server error here,
2957 * and just try again when we close(), unfortunately StoreAllSegments
2958 * will invalidate our chunks if the server returns a permanent error,
2959 * so we need to at least try and get that error back to the user
2962 code1 = afs_linux_dopartialwrite(vcp, credp);
2964 afs_linux_complete_writeback(vcp);
2965 ReleaseWriteLock(&vcp->lock);
2970 end_page_writeback(pp);
2982 /* afs_linux_permission
2983 * Check access rights - returns error if can't check or permission denied.
2986 #if defined(IOP_PERMISSION_TAKES_FLAGS)
2987 afs_linux_permission(struct inode *ip, int mode, unsigned int flags)
2988 #elif defined(IOP_PERMISSION_TAKES_NAMEIDATA)
2989 afs_linux_permission(struct inode *ip, int mode, struct nameidata *nd)
2991 afs_linux_permission(struct inode *ip, int mode)
2998 /* Check for RCU path walking */
2999 #if defined(IOP_PERMISSION_TAKES_FLAGS)
3000 if (flags & IPERM_FLAG_RCU)
3002 #elif defined(MAY_NOT_BLOCK)
3003 if (mode & MAY_NOT_BLOCK)
3009 if (mode & MAY_EXEC)
3011 if (mode & MAY_READ)
3013 if (mode & MAY_WRITE)
3015 code = afs_access(VTOAFS(ip), tmp, credp);
3019 return afs_convert_code(code);
3023 afs_linux_commit_write(struct file *file, struct page *page, unsigned offset,
3027 struct inode *inode = FILE_INODE(file);
3028 loff_t pagebase = page_offset(page);
3030 if (i_size_read(inode) < (pagebase + offset))
3031 i_size_write(inode, pagebase + offset);
3033 if (PageChecked(page)) {
3034 SetPageUptodate(page);
3035 ClearPageChecked(page);
3038 code = afs_linux_writepage_sync(inode, page, offset, to - offset);
3044 afs_linux_prepare_write(struct file *file, struct page *page, unsigned from,
3048 /* http://kerneltrap.org/node/4941 details the expected behaviour of
3049 * prepare_write. Essentially, if the page exists within the file,
3050 * and is not being fully written, then we should populate it.
3053 if (!PageUptodate(page)) {
3054 loff_t pagebase = page_offset(page);
3055 loff_t isize = i_size_read(page->mapping->host);
3057 /* Is the location we are writing to beyond the end of the file? */
3058 if (pagebase >= isize ||
3059 ((from == 0) && (pagebase + to) >= isize)) {
3060 zero_user_segments(page, 0, from, to, PAGE_SIZE);
3061 SetPageChecked(page);
3062 /* Are we we writing a full page */
3063 } else if (from == 0 && to == PAGE_SIZE) {
3064 SetPageChecked(page);
3065 /* Is the page readable, if it's wronly, we don't care, because we're
3066 * not actually going to read from it ... */
3067 } else if ((file->f_flags && O_ACCMODE) != O_WRONLY) {
3068 /* We don't care if fillpage fails, because if it does the page
3069 * won't be marked as up to date
3071 afs_linux_fillpage(file, page);
3077 #if defined(STRUCT_ADDRESS_SPACE_OPERATIONS_HAS_WRITE_BEGIN)
3079 afs_linux_write_end(struct file *file, struct address_space *mapping,
3080 loff_t pos, unsigned len, unsigned copied,
3081 struct page *page, void *fsdata)
3084 unsigned int from = pos & (PAGE_SIZE - 1);
3086 code = afs_linux_commit_write(file, page, from, from + copied);
3094 afs_linux_write_begin(struct file *file, struct address_space *mapping,
3095 loff_t pos, unsigned len, unsigned flags,
3096 struct page **pagep, void **fsdata)
3099 pgoff_t index = pos >> PAGE_SHIFT;
3100 unsigned int from = pos & (PAGE_SIZE - 1);
3103 page = grab_cache_page_write_begin(mapping, index, flags);
3110 code = afs_linux_prepare_write(file, page, from, from + len);
3120 #ifndef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
3122 afs_linux_dir_follow_link(struct dentry *dentry, struct nameidata *nd)
3124 struct dentry **dpp;
3125 struct dentry *target;
3127 if (current->total_link_count > 0) {
3128 /* avoid symlink resolution limits when resolving; we cannot contribute to
3129 * an infinite symlink loop */
3130 /* only do this for follow_link when total_link_count is positive to be
3131 * on the safe side; there is at least one code path in the Linux
3132 * kernel where it seems like it may be possible to get here without
3133 * total_link_count getting incremented. it is not clear on how that
3134 * path is actually reached, but guard against it just to be safe */
3135 current->total_link_count--;
3138 target = canonical_dentry(dentry->d_inode);
3140 # ifdef STRUCT_NAMEIDATA_HAS_PATH
3141 dpp = &nd->path.dentry;
3151 *dpp = dget(dentry);
3154 nd->last_type = LAST_BIND;
3158 #endif /* !STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT */
3161 static struct inode_operations afs_file_iops = {
3162 .permission = afs_linux_permission,
3163 .getattr = afs_linux_getattr,
3164 .setattr = afs_notify_change,
3167 static struct address_space_operations afs_file_aops = {
3168 .readpage = afs_linux_readpage,
3169 .readpages = afs_linux_readpages,
3170 .writepage = afs_linux_writepage,
3171 #if defined (STRUCT_ADDRESS_SPACE_OPERATIONS_HAS_WRITE_BEGIN)
3172 .write_begin = afs_linux_write_begin,
3173 .write_end = afs_linux_write_end,
3175 .commit_write = afs_linux_commit_write,
3176 .prepare_write = afs_linux_prepare_write,
3181 /* Separate ops vector for directories. Linux 2.2 tests type of inode
3182 * by what sort of operation is allowed.....
3185 static struct inode_operations afs_dir_iops = {
3186 .setattr = afs_notify_change,
3187 .create = afs_linux_create,
3188 .lookup = afs_linux_lookup,
3189 .link = afs_linux_link,
3190 .unlink = afs_linux_unlink,
3191 .symlink = afs_linux_symlink,
3192 .mkdir = afs_linux_mkdir,
3193 .rmdir = afs_linux_rmdir,
3194 .rename = afs_linux_rename,
3195 .getattr = afs_linux_getattr,
3196 .permission = afs_linux_permission,
3197 #ifndef STRUCT_DENTRY_OPERATIONS_HAS_D_AUTOMOUNT
3198 .follow_link = afs_linux_dir_follow_link,
3202 /* We really need a separate symlink set of ops, since do_follow_link()
3203 * determines if it _is_ a link by checking if the follow_link op is set.
3205 #if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
3207 afs_symlink_filler(struct file *file, struct page *page)
3209 struct inode *ip = (struct inode *)page->mapping->host;
3210 char *p = (char *)kmap(page);
3214 code = afs_linux_ireadlink(ip, p, PAGE_SIZE, AFS_UIOSYS);
3219 p[code] = '\0'; /* null terminate? */
3221 SetPageUptodate(page);
3233 static struct address_space_operations afs_symlink_aops = {
3234 .readpage = afs_symlink_filler
3236 #endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
3238 static struct inode_operations afs_symlink_iops = {
3239 #if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
3240 .readlink = page_readlink,
3241 # if defined(HAVE_LINUX_PAGE_GET_LINK)
3242 .get_link = page_get_link,
3243 # elif defined(HAVE_LINUX_PAGE_FOLLOW_LINK)
3244 .follow_link = page_follow_link,
3246 .follow_link = page_follow_link_light,
3247 .put_link = page_put_link,
3249 #else /* !defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE) */
3250 .readlink = afs_linux_readlink,
3251 .follow_link = afs_linux_follow_link,
3252 .put_link = afs_linux_put_link,
3253 #endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
3254 .setattr = afs_notify_change,
3258 afs_fill_inode(struct inode *ip, struct vattr *vattr)
3261 vattr2inode(ip, vattr);
3263 #ifdef STRUCT_ADDRESS_SPACE_HAS_BACKING_DEV_INFO
3264 ip->i_mapping->backing_dev_info = afs_backing_dev_info;
3266 /* Reset ops if symlink or directory. */
3267 if (S_ISREG(ip->i_mode)) {
3268 ip->i_op = &afs_file_iops;
3269 ip->i_fop = &afs_file_fops;
3270 ip->i_data.a_ops = &afs_file_aops;
3272 } else if (S_ISDIR(ip->i_mode)) {
3273 ip->i_op = &afs_dir_iops;
3274 ip->i_fop = &afs_dir_fops;
3276 } else if (S_ISLNK(ip->i_mode)) {
3277 ip->i_op = &afs_symlink_iops;
3278 #if defined(HAVE_LINUX_INODE_NOHIGHMEM)
3279 inode_nohighmem(ip);
3281 #if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
3282 ip->i_data.a_ops = &afs_symlink_aops;
3283 ip->i_mapping = &ip->i_data;