#include "afs/sysincludes.h"
#include "afsincludes.h"
#include "afs/afs_stats.h"
-#include "h/mm.h"
+#include <linux/mm.h>
#ifdef HAVE_MM_INLINE_H
-#include "h/mm_inline.h"
+#include <linux/mm_inline.h>
#endif
-#include "h/pagemap.h"
-#include "h/smp_lock.h"
-#include "h/writeback.h"
-#include "h/pagevec.h"
+#include <linux/pagemap.h>
+#include <linux/smp_lock.h>
+#include <linux/writeback.h>
+#include <linux/pagevec.h>
#if defined(AFS_CACHE_BYPASS)
#include "afs/lock.h"
#include "afs/afs_bypasscache.h"
#include "osi_compat.h"
#include "osi_pagecopy.h"
-#ifdef pgoff2loff
-#define pageoff(pp) pgoff2loff((pp)->index)
-#else
-#define pageoff(pp) pp->offset
-#endif
-
-#ifndef HAVE_PAGEVEC_LRU_ADD_FILE
+#ifndef HAVE_LINUX_PAGEVEC_LRU_ADD_FILE
#define __pagevec_lru_add_file __pagevec_lru_add
#endif
#define MAX_ERRNO 1000L
#endif
-#define LockPage(pp) lock_page(pp)
-#define UnlockPage(pp) unlock_page(pp)
-extern struct backing_dev_info afs_backing_dev_info;
+extern struct backing_dev_info *afs_backing_dev_info;
extern struct vcache *afs_globalVp;
extern int afs_notify_change(struct dentry *dp, struct iattr *iattrp);
-/* Some uses of BKL are perhaps not needed for bypass or memcache--
- * why don't we try it out? */
-extern struct afs_cacheOps afs_UfsCacheOps;
-
-static inline void
-afs_maybe_lock_kernel(void) {
- if(afs_cacheType == &afs_UfsCacheOps)
- lock_kernel();
-}
-
-static inline void
-afs_maybe_unlock_kernel(void) {
- if(afs_cacheType == &afs_UfsCacheOps)
- unlock_kernel();
-}
/* This function converts a positive error code from AFS into a negative
* code suitable for passing into the Linux VFS layer. It checks that the
{
struct vcache *avc = VTOAFS(FILE_INODE(fp));
struct vrequest treq;
- register struct dcache *tdc;
+ struct dcache *tdc;
int code;
int offset;
int dirpos;
cred_t *credp = crref();
struct afs_fakestat_state fakestat;
- afs_maybe_lock_kernel();
AFS_GLOCK();
AFS_STATCNT(afs_readdir);
afs_PutFakeStat(&fakestat);
out1:
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
return code;
}
/* get a validated vcache entry */
code = afs_linux_VerifyVCache(vcp, NULL);
- /* Linux's Flushpage implementation doesn't use credp, so optimise
- * our code to not need to crref() it */
- osi_FlushPages(vcp, NULL); /* ensure stale pages are gone */
+ if (code == 0) {
+ /* Linux's Flushpage implementation doesn't use credp, so optimise
+ * our code to not need to crref() it */
+ osi_FlushPages(vcp, NULL); /* ensure stale pages are gone */
+ AFS_GUNLOCK();
+ code = generic_file_mmap(fp, vmap);
+ AFS_GLOCK();
+ if (!code)
+ vcp->f.states |= CMAPPED;
+ }
AFS_GUNLOCK();
- code = generic_file_mmap(fp, vmap);
- AFS_GLOCK();
- if (!code)
- vcp->f.states |= CMAPPED;
- AFS_GUNLOCK();
return code;
}
cred_t *credp = crref();
int code;
- afs_maybe_lock_kernel();
AFS_GLOCK();
code = afs_open(&vcp, fp->f_flags, credp);
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
crfree(credp);
return afs_convert_code(code);
cred_t *credp = crref();
int code = 0;
- afs_maybe_lock_kernel();
AFS_GLOCK();
code = afs_close(vcp, fp->f_flags, credp);
+ ObtainWriteLock(&vcp->lock, 807);
+ if (vcp->cred) {
+ crfree(vcp->cred);
+ vcp->cred = NULL;
+ }
+ ReleaseWriteLock(&vcp->lock);
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
crfree(credp);
return afs_convert_code(code);
}
static int
+#if defined(FOP_FSYNC_TAKES_DENTRY)
afs_linux_fsync(struct file *fp, struct dentry *dp, int datasync)
+#else
+afs_linux_fsync(struct file *fp, int datasync)
+#endif
{
int code;
struct inode *ip = FILE_INODE(fp);
cred_t *credp = crref();
- afs_maybe_lock_kernel();
AFS_GLOCK();
code = afs_fsync(VTOAFS(ip), credp);
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
crfree(credp);
return afs_convert_code(code);
struct AFS_FLOCK flock;
/* Convert to a lock format afs_lockctl understands. */
- memset((char *)&flock, 0, sizeof(flock));
+ memset(&flock, 0, sizeof(flock));
flock.l_type = flp->fl_type;
flock.l_pid = flp->fl_pid;
flock.l_whence = 0;
flock.l_start = flp->fl_start;
- flock.l_len = flp->fl_end - flp->fl_start + 1;
+ if (flp->fl_end == OFFSET_MAX)
+ flock.l_len = 0; /* Lock to end of file */
+ else
+ flock.l_len = flp->fl_end - flp->fl_start + 1;
/* Safe because there are no large files, yet */
#if defined(F_GETLK64) && (F_GETLK != F_GETLK64)
#endif /* F_GETLK64 && F_GETLK != F_GETLK64 */
AFS_GLOCK();
- code = afs_lockctl(vcp, &flock, cmd, credp);
+ code = afs_convert_code(afs_lockctl(vcp, &flock, cmd, credp));
AFS_GUNLOCK();
if ((code == 0 || flp->fl_type == F_UNLCK) &&
flp->fl_type = flock.l_type;
flp->fl_pid = flock.l_pid;
flp->fl_start = flock.l_start;
- flp->fl_end = flock.l_start + flock.l_len - 1;
+ if (flock.l_len == 0)
+ flp->fl_end = OFFSET_MAX; /* Lock to end of file */
+ else
+ flp->fl_end = flock.l_start + flock.l_len - 1;
crfree(credp);
- return afs_convert_code(code);
+ return code;
}
#ifdef STRUCT_FILE_OPERATIONS_HAS_FLOCK
cred_t *credp = crref();
struct AFS_FLOCK flock;
/* Convert to a lock format afs_lockctl understands. */
- memset((char *)&flock, 0, sizeof(flock));
+ memset(&flock, 0, sizeof(flock));
flock.l_type = flp->fl_type;
flock.l_pid = flp->fl_pid;
flock.l_whence = 0;
flock.l_start = 0;
- flock.l_len = OFFSET_MAX;
+ flock.l_len = 0;
/* Safe because there are no large files, yet */
#if defined(F_GETLK64) && (F_GETLK != F_GETLK64)
#endif /* F_GETLK64 && F_GETLK != F_GETLK64 */
AFS_GLOCK();
- code = afs_lockctl(vcp, &flock, cmd, credp);
+ code = afs_convert_code(afs_lockctl(vcp, &flock, cmd, credp));
AFS_GUNLOCK();
if ((code == 0 || flp->fl_type == F_UNLCK) &&
flp->fl_pid = flock.l_pid;
crfree(credp);
- return afs_convert_code(code);
+ return code;
}
#endif
struct file_operations afs_file_fops = {
.read = afs_linux_read,
.write = afs_linux_write,
-#ifdef GENERIC_FILE_AIO_READ
+#ifdef HAVE_LINUX_GENERIC_FILE_AIO_READ
.aio_read = generic_file_aio_read,
.aio_write = generic_file_aio_write,
#endif
check_bad_parent(struct dentry *dp)
{
cred_t *credp;
- struct vcache *vcp = VTOAFS(dp->d_inode), *avc = NULL;
- struct vcache *pvc = VTOAFS(dp->d_parent->d_inode);
+ struct dentry *parent;
+ struct vcache *vcp, *pvc, *avc = NULL;
+
+ vcp = VTOAFS(dp->d_inode);
+ parent = dget_parent(dp);
+ pvc = VTOAFS(parent->d_inode);
if (vcp->mvid->Fid.Volume != pvc->f.fid.Fid.Volume) { /* bad parent */
credp = crref();
crfree(credp);
}
+ dput(parent);
+
return;
}
if (afs_shuttingdown)
return EIO;
- afs_maybe_lock_kernel();
AFS_GLOCK();
#ifdef notyet
check_bad_parent(dp); /* check and correct mvid */
AFS_GUNLOCK();
- unlock_kernel();
return 0;
}
#endif
*/
if (vcp->f.states & CStatd &&
(!afs_fakestat_enable || vcp->mvstat != 1) &&
- !afs_nfsexporter) {
+ !afs_nfsexporter &&
+ (vType(vcp) == VDIR || vType(vcp) == VLNK)) {
code = afs_CopyOutAttrs(vcp, &vattr);
} else {
credp = crref();
afs_fill_inode(AFSTOV(vcp), &vattr);
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
return afs_convert_code(code);
}
struct vattr vattr;
cred_t *credp = NULL;
struct vcache *vcp, *pvcp, *tvc = NULL;
+ struct dentry *parent;
int valid;
struct afs_fakestat_state fakestate;
- afs_maybe_lock_kernel();
AFS_GLOCK();
afs_InitFakeStat(&fakestate);
if (dp->d_inode) {
-
vcp = VTOAFS(dp->d_inode);
- pvcp = VTOAFS(dp->d_parent->d_inode); /* dget_parent()? */
if (vcp == afs_globalVp)
goto good_dentry;
}
#endif
+ parent = dget_parent(dp);
+ pvcp = VTOAFS(parent->d_inode);
+
/* If the parent's DataVersion has changed or the vnode
* is longer valid, we need to do a full lookup. VerifyVCache
* isn't enough since the vnode may have been renamed.
credp = crref();
afs_lookup(pvcp, (char *)dp->d_name.name, &tvc, credp);
- if (!tvc || tvc != vcp)
+ if (!tvc || tvc != vcp) {
+ dput(parent);
goto bad_dentry;
+ }
- if (afs_getattr(vcp, &vattr, credp))
+ if (afs_getattr(vcp, &vattr, credp)) {
+ dput(parent);
goto bad_dentry;
+ }
vattr2inode(AFSTOV(vcp), &vattr);
dp->d_time = hgetlo(pvcp->f.m.DataVersion);
/* should we always update the attributes at this point? */
/* unlikely--the vcache entry hasn't changed */
+ dput(parent);
} else {
#ifdef notyet
- pvcp = VTOAFS(dp->d_parent->d_inode); /* dget_parent()? */
+ /* If this code is ever enabled, we should use dget_parent to handle
+ * getting the parent, and dput() to dispose of it. See above for an
+ * example ... */
+ pvcp = VTOAFS(dp->d_parent->d_inode);
if (hgetlo(pvcp->f.m.DataVersion) > dp->d_time)
goto bad_dentry;
#endif
shrink_dcache_parent(dp);
d_drop(dp);
}
- afs_maybe_unlock_kernel();
return valid;
bad_dentry:
vattr.va_mode = mode;
vattr.va_type = mode & S_IFMT;
- afs_maybe_lock_kernel();
AFS_GLOCK();
code = afs_create(VTOAFS(dip), (char *)name, &vattr, NONEXCL, mode,
&vcp, credp);
}
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
crfree(credp);
return afs_convert_code(code);
}
struct dentry *newdp = NULL;
int code;
- afs_maybe_lock_kernel();
AFS_GLOCK();
code = afs_lookup(VTOAFS(dip), (char *)comp, &vcp, credp);
dput(alias);
} else {
iput(ip);
- unlock_kernel();
crfree(credp);
return alias;
}
}
newdp = d_splice_alias(ip, dp);
- afs_maybe_unlock_kernel();
crfree(credp);
/* It's ok for the file to not be found. That's noted by the caller by
return afs_convert_code(code);
}
+/* We have to have a Linux specific sillyrename function, because we
+ * also have to keep the dcache up to date when we're doing a silly
+ * rename - so we don't want the generic vnodeops doing this behind our
+ * back.
+ */
+
+static int
+afs_linux_sillyrename(struct inode *dir, struct dentry *dentry,
+ cred_t *credp)
+{
+ struct vcache *tvc = VTOAFS(dentry->d_inode);
+ struct dentry *__dp = NULL;
+ char *__name = NULL;
+ int code;
+
+ if (afs_linux_nfsfs_renamed(dentry))
+ return EBUSY;
+
+ do {
+ dput(__dp);
+
+ AFS_GLOCK();
+ if (__name)
+ osi_FreeSmallSpace(__name);
+ __name = afs_newname();
+ AFS_GUNLOCK();
+
+ __dp = lookup_one_len(__name, dentry->d_parent, strlen(__name));
+
+ if (IS_ERR(__dp)) {
+ osi_FreeSmallSpace(__name);
+ return EBUSY;
+ }
+ } while (__dp->d_inode != NULL);
+
+ AFS_GLOCK();
+ code = afs_rename(VTOAFS(dir), (char *)dentry->d_name.name,
+ VTOAFS(dir), (char *)__dp->d_name.name,
+ credp);
+ if (!code) {
+ tvc->mvid = (void *) __name;
+ crhold(credp);
+ if (tvc->uncred) {
+ crfree(tvc->uncred);
+ }
+ tvc->uncred = credp;
+ tvc->f.states |= CUnlinked;
+ afs_linux_set_nfsfs_renamed(dentry);
+ } else {
+ osi_FreeSmallSpace(__name);
+ }
+ AFS_GUNLOCK();
+
+ if (!code) {
+ __dp->d_time = hgetlo(VTOAFS(dir)->f.m.DataVersion);
+ d_move(dentry, __dp);
+ }
+ dput(__dp);
+
+ return code;
+}
+
+
static int
afs_linux_unlink(struct inode *dip, struct dentry *dp)
{
const char *name = dp->d_name.name;
struct vcache *tvc = VTOAFS(dp->d_inode);
- afs_maybe_lock_kernel();
if (VREFCOUNT(tvc) > 1 && tvc->opens > 0
&& !(tvc->f.states & CUnlinked)) {
- struct dentry *__dp;
- char *__name;
-
- __dp = NULL;
- __name = NULL;
- do {
- dput(__dp);
-
- AFS_GLOCK();
- if (__name)
- osi_FreeSmallSpace(__name);
- __name = afs_newname();
- AFS_GUNLOCK();
-
- __dp = lookup_one_len(__name, dp->d_parent, strlen(__name));
-
- if (IS_ERR(__dp))
- goto out;
- } while (__dp->d_inode != NULL);
+ code = afs_linux_sillyrename(dip, dp, credp);
+ } else {
AFS_GLOCK();
- code = afs_rename(VTOAFS(dip), (char *)dp->d_name.name, VTOAFS(dip), (char *)__dp->d_name.name, credp);
- if (!code) {
- tvc->mvid = (void *) __name;
- crhold(credp);
- if (tvc->uncred) {
- crfree(tvc->uncred);
- }
- tvc->uncred = credp;
- tvc->f.states |= CUnlinked;
- afs_linux_set_nfsfs_renamed(dp);
- } else {
- osi_FreeSmallSpace(__name);
- }
+ code = afs_remove(VTOAFS(dip), (char *)name, credp);
AFS_GUNLOCK();
-
- if (!code) {
- __dp->d_time = hgetlo(VTOAFS(dip)->f.m.DataVersion);
- d_move(dp, __dp);
- }
- dput(__dp);
-
- goto out;
+ if (!code)
+ d_drop(dp);
}
- AFS_GLOCK();
- code = afs_remove(VTOAFS(dip), (char *)name, credp);
- AFS_GUNLOCK();
- if (!code)
- d_drop(dp);
-out:
- afs_maybe_unlock_kernel();
crfree(credp);
return afs_convert_code(code);
}
struct vattr vattr;
const char *name = dp->d_name.name;
- afs_maybe_lock_kernel();
VATTR_NULL(&vattr);
vattr.va_mask = ATTR_MODE;
vattr.va_mode = mode;
}
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
crfree(credp);
return afs_convert_code(code);
}
struct dentry *rehash = NULL;
/* Prevent any new references during rename operation. */
- afs_maybe_lock_kernel();
if (!d_unhashed(newdp)) {
d_drop(newdp);
if (rehash)
d_rehash(rehash);
- afs_maybe_unlock_kernel();
-
crfree(credp);
return afs_convert_code(code);
}
#endif /* USABLE_KERNEL_PAGE_SYMLINK_CACHE */
#if defined(AFS_CACHE_BYPASS)
-static inline int
-afs_linux_can_bypass(struct inode *ip) {
- switch(cache_bypass_strategy) {
- case NEVER_BYPASS_CACHE:
- return 0;
- case ALWAYS_BYPASS_CACHE:
- return 1;
- case LARGE_FILES_BYPASS_CACHE:
- if(i_size_read(ip) > cache_bypass_threshold)
- return 1;
- default:
- }
- return 0;
-}
-
-static int
-afs_linux_cache_bypass_read(struct file *fp, struct address_space *mapping,
- struct list_head *page_list, unsigned num_pages)
-{
- afs_int32 page_ix;
- uio_t *auio;
- afs_offs_t offset;
- struct iovec* iovecp;
- struct nocache_read_request *ancr;
- struct page *pp, *ppt;
- struct pagevec lrupv;
- afs_int32 code = 0;
-
- cred_t *credp;
- struct inode *ip = FILE_INODE(fp);
- struct vcache *avc = VTOAFS(ip);
- afs_int32 bypasscache = 0; /* bypass for this read */
- afs_int32 base_index = 0;
- afs_int32 page_count = 0;
- afs_int32 isize;
-
- /* background thread must free: iovecp, auio, ancr */
- iovecp = osi_Alloc(num_pages * sizeof(struct iovec));
-
- auio = osi_Alloc(sizeof(uio_t));
- auio->uio_iov = iovecp;
- auio->uio_iovcnt = num_pages;
- auio->uio_flag = UIO_READ;
- auio->uio_seg = AFS_UIOSYS;
- auio->uio_resid = num_pages * PAGE_SIZE;
-
- ancr = osi_Alloc(sizeof(struct nocache_read_request));
- ancr->auio = auio;
- ancr->offset = auio->uio_offset;
- ancr->length = auio->uio_resid;
-
- pagevec_init(&lrupv, 0);
-
- for(page_ix = 0; page_ix < num_pages; ++page_ix) {
-
- if(list_empty(page_list))
- break;
-
- pp = list_entry(page_list->prev, struct page, lru);
- /* If we allocate a page and don't remove it from page_list,
- * the page cache gets upset. */
- list_del(&pp->lru);
- isize = (i_size_read(fp->f_mapping->host) - 1) >> PAGE_CACHE_SHIFT;
- if(pp->index > isize) {
- if(PageLocked(pp))
- UnlockPage(pp);
- continue;
- }
-
- if(page_ix == 0) {
- offset = ((loff_t) pp->index) << PAGE_CACHE_SHIFT;
- auio->uio_offset = offset;
- base_index = pp->index;
- }
- iovecp[page_ix].iov_len = PAGE_SIZE;
- code = add_to_page_cache(pp, mapping, pp->index, GFP_KERNEL);
- if(base_index != pp->index) {
- if(PageLocked(pp))
- UnlockPage(pp);
- page_cache_release(pp);
- iovecp[page_ix].iov_base = (void *) 0;
- base_index++;
- continue;
- }
- base_index++;
- if(code) {
- if(PageLocked(pp))
- UnlockPage(pp);
- page_cache_release(pp);
- iovecp[page_ix].iov_base = (void *) 0;
- } else {
- page_count++;
- if(!PageLocked(pp)) {
- LockPage(pp);
- }
-
- /* save the page for background map */
- iovecp[page_ix].iov_base = (void*) pp;
-
- /* and put it on the LRU cache */
- if (!pagevec_add(&lrupv, pp))
- __pagevec_lru_add(&lrupv);
- }
- }
-
- /* If there were useful pages in the page list, make sure all pages
- * are in the LRU cache, then schedule the read */
- if(page_count) {
- pagevec_lru_add(&lrupv);
- credp = crref();
- code = afs_ReadNoCache(avc, ancr, credp);
- crfree(credp);
- } else {
- /* If there is nothing for the background thread to handle,
- * it won't be freeing the things that we never gave it */
- osi_Free(iovecp, num_pages * sizeof(struct iovec));
- osi_Free(auio, sizeof(uio_t));
- osi_Free(ancr, sizeof(struct nocache_read_request));
- }
- /* we do not flush, release, or unmap pages--that will be
- * done for us by the background thread as each page comes in
- * from the fileserver */
-out:
- return afs_convert_code(code);
-}
-
#endif /* defined(AFS_CACHE_BYPASS */
+/* Populate a page by filling it from the cache file pointed at by cachefp
+ * (which contains indicated chunk)
+ * If task is NULL, the page copy occurs syncronously, and the routine
+ * returns with page still locked. If task is non-NULL, then page copies
+ * may occur in the background, and the page will be unlocked when it is
+ * ready for use.
+ */
static int
afs_linux_read_cache(struct file *cachefp, struct page *page,
int chunk, struct pagevec *lrupv,
copy_highpage(page, cachepage);
flush_dcache_page(page);
SetPageUptodate(page);
- UnlockPage(page);
+
+ if (task)
+ unlock_page(page);
} else if (task) {
afs_pagecopy_queue_page(task, cachepage, page);
} else {
}
}
- if (code) {
- UnlockPage(page);
+ if (code && task) {
+ unlock_page(page);
}
out:
/* XXX - I suspect we should be locking the inodes before we use them! */
AFS_GUNLOCK();
- cacheFp = afs_linux_raw_open(&tdc->f.inode, NULL);
+ cacheFp = afs_linux_raw_open(&tdc->f.inode);
pagevec_init(&lrupv, 0);
code = afs_linux_read_cache(cacheFp, pp, tdc->f.chunk, &lrupv, NULL);
}
/* afs_linux_readpage
- * all reads come through here. A strategy-like read call.
+ *
+ * This function is split into two, because prepare_write/begin_write
+ * require a readpage call which doesn't unlock the resulting page upon
+ * success.
*/
static int
-afs_linux_readpage(struct file *fp, struct page *pp)
+afs_linux_fillpage(struct file *fp, struct page *pp)
{
afs_int32 code;
char *address;
- afs_offs_t offset = ((loff_t) pp->index) << PAGE_CACHE_SHIFT;
-#if defined(AFS_CACHE_BYPASS)
- afs_int32 bypasscache = 0; /* bypass for this read */
- struct nocache_read_request *ancr;
- afs_int32 isize;
-#endif
uio_t *auio;
struct iovec *iovecp;
struct inode *ip = FILE_INODE(fp);
afs_int32 cnt = page_count(pp);
struct vcache *avc = VTOAFS(ip);
+ afs_offs_t offset = page_offset(pp);
cred_t *credp;
AFS_GLOCK();
address = kmap(pp);
ClearPageError(pp);
- /* if bypasscache, receiver frees, else we do */
auio = osi_Alloc(sizeof(uio_t));
iovecp = osi_Alloc(sizeof(struct iovec));
setup_uio(auio, iovecp, (char *)address, offset, PAGE_SIZE, UIO_READ,
- AFS_UIOSYS);
-
-#if defined(AFS_CACHE_BYPASS)
- bypasscache = afs_linux_can_bypass(ip);
+ AFS_UIOSYS);
- /* In the new incarnation of selective caching, a file's caching policy
- * can change, eg because file size exceeds threshold, etc. */
- trydo_cache_transition(avc, credp, bypasscache);
-
- if(bypasscache) {
- if(address)
- kunmap(pp);
- /* save the page for background map */
- auio->uio_iov->iov_base = (void*) pp;
- /* the background thread will free this */
- ancr = osi_Alloc(sizeof(struct nocache_read_request));
- ancr->auio = auio;
- ancr->offset = offset;
- ancr->length = PAGE_SIZE;
-
- afs_maybe_lock_kernel();
- code = afs_ReadNoCache(avc, ancr, credp);
- afs_maybe_unlock_kernel();
-
- goto done; /* skips release page, doing it in bg thread */
- }
-#endif
-
- afs_maybe_lock_kernel();
AFS_GLOCK();
AFS_DISCON_LOCK();
afs_Trace4(afs_iclSetp, CM_TRACE_READPAGE, ICL_TYPE_POINTER, ip,
code);
AFS_DISCON_UNLOCK();
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
if (!code) {
/* XXX valid for no-cache also? Check last bits of files... :)
* Cognate code goes in afs_NoCacheFetchProc. */
} /* !code */
kunmap(pp);
- UnlockPage(pp);
-
-#if defined(AFS_CACHE_BYPASS)
- /* do not call afs_GetDCache if cache is bypassed */
- if(bypasscache)
- goto done;
-#endif
- /* free if not bypassing cache */
osi_Free(auio, sizeof(uio_t));
osi_Free(iovecp, sizeof(struct iovec));
- if (!code && AFS_CHUNKOFFSET(offset) == 0) {
+ crfree(credp);
+ return afs_convert_code(code);
+}
+
+static int
+afs_linux_prefetch(struct file *fp, struct page *pp)
+{
+ int code = 0;
+ struct vcache *avc = VTOAFS(FILE_INODE(fp));
+ afs_offs_t offset = page_offset(pp);
+
+ if (AFS_CHUNKOFFSET(offset) == 0) {
struct dcache *tdc;
struct vrequest treq;
+ cred_t *credp;
+ credp = crref();
AFS_GLOCK();
code = afs_InitReq(&treq, credp);
if (!code && !NBObtainWriteLock(&avc->lock, 534)) {
ReleaseWriteLock(&avc->lock);
}
AFS_GUNLOCK();
+ crfree(credp);
}
+ return afs_convert_code(code);
+
+}
#if defined(AFS_CACHE_BYPASS)
-done:
-#endif
+
+static int
+afs_linux_bypass_readpages(struct file *fp, struct address_space *mapping,
+ struct list_head *page_list, unsigned num_pages)
+{
+ afs_int32 page_ix;
+ uio_t *auio;
+ afs_offs_t offset;
+ struct iovec* iovecp;
+ struct nocache_read_request *ancr;
+ struct page *pp;
+ struct pagevec lrupv;
+ afs_int32 code = 0;
+
+ cred_t *credp;
+ struct inode *ip = FILE_INODE(fp);
+ struct vcache *avc = VTOAFS(ip);
+ afs_int32 base_index = 0;
+ afs_int32 page_count = 0;
+ afs_int32 isize;
+
+ /* background thread must free: iovecp, auio, ancr */
+ iovecp = osi_Alloc(num_pages * sizeof(struct iovec));
+
+ auio = osi_Alloc(sizeof(uio_t));
+ auio->uio_iov = iovecp;
+ auio->uio_iovcnt = num_pages;
+ auio->uio_flag = UIO_READ;
+ auio->uio_seg = AFS_UIOSYS;
+ auio->uio_resid = num_pages * PAGE_SIZE;
+
+ ancr = osi_Alloc(sizeof(struct nocache_read_request));
+ ancr->auio = auio;
+ ancr->offset = auio->uio_offset;
+ ancr->length = auio->uio_resid;
+
+ pagevec_init(&lrupv, 0);
+
+ for(page_ix = 0; page_ix < num_pages; ++page_ix) {
+
+ if(list_empty(page_list))
+ break;
+
+ pp = list_entry(page_list->prev, struct page, lru);
+ /* If we allocate a page and don't remove it from page_list,
+ * the page cache gets upset. */
+ list_del(&pp->lru);
+ isize = (i_size_read(fp->f_mapping->host) - 1) >> PAGE_CACHE_SHIFT;
+ if(pp->index > isize) {
+ if(PageLocked(pp))
+ unlock_page(pp);
+ continue;
+ }
+
+ if(page_ix == 0) {
+ offset = page_offset(pp);
+ auio->uio_offset = offset;
+ base_index = pp->index;
+ }
+ iovecp[page_ix].iov_len = PAGE_SIZE;
+ code = add_to_page_cache(pp, mapping, pp->index, GFP_KERNEL);
+ if(base_index != pp->index) {
+ if(PageLocked(pp))
+ unlock_page(pp);
+ page_cache_release(pp);
+ iovecp[page_ix].iov_base = (void *) 0;
+ base_index++;
+ continue;
+ }
+ base_index++;
+ if(code) {
+ if(PageLocked(pp))
+ unlock_page(pp);
+ page_cache_release(pp);
+ iovecp[page_ix].iov_base = (void *) 0;
+ } else {
+ page_count++;
+ if(!PageLocked(pp)) {
+ lock_page(pp);
+ }
+
+ /* increment page refcount--our original design assumed
+ * that locking it would effectively pin it; protect
+ * ourselves from the possiblity that this assumption is
+ * is faulty, at low cost (provided we do not fail to
+ * do the corresponding decref on the other side) */
+ get_page(pp);
+
+ /* save the page for background map */
+ iovecp[page_ix].iov_base = (void*) pp;
+
+ /* and put it on the LRU cache */
+ if (!pagevec_add(&lrupv, pp))
+ __pagevec_lru_add_file(&lrupv);
+ }
+ }
+
+ /* If there were useful pages in the page list, make sure all pages
+ * are in the LRU cache, then schedule the read */
+ if(page_count) {
+ if (pagevec_count(&lrupv))
+ __pagevec_lru_add_file(&lrupv);
+ credp = crref();
+ code = afs_ReadNoCache(avc, ancr, credp);
+ crfree(credp);
+ } else {
+ /* If there is nothing for the background thread to handle,
+ * it won't be freeing the things that we never gave it */
+ osi_Free(iovecp, num_pages * sizeof(struct iovec));
+ osi_Free(auio, sizeof(uio_t));
+ osi_Free(ancr, sizeof(struct nocache_read_request));
+ }
+ /* we do not flush, release, or unmap pages--that will be
+ * done for us by the background thread as each page comes in
+ * from the fileserver */
+ return afs_convert_code(code);
+}
+
+
+static int
+afs_linux_bypass_readpage(struct file *fp, struct page *pp)
+{
+ cred_t *credp = NULL;
+ uio_t *auio;
+ struct iovec *iovecp;
+ struct nocache_read_request *ancr;
+ int code;
+
+ ClearPageError(pp);
+
+ /* receiver frees */
+ auio = osi_Alloc(sizeof(uio_t));
+ iovecp = osi_Alloc(sizeof(struct iovec));
+
+ /* address can be NULL, because we overwrite it with 'pp', below */
+ setup_uio(auio, iovecp, NULL, page_offset(pp),
+ PAGE_SIZE, UIO_READ, AFS_UIOSYS);
+
+ /* save the page for background map */
+ get_page(pp); /* see above */
+ auio->uio_iov->iov_base = (void*) pp;
+ /* the background thread will free this */
+ ancr = osi_Alloc(sizeof(struct nocache_read_request));
+ ancr->auio = auio;
+ ancr->offset = page_offset(pp);
+ ancr->length = PAGE_SIZE;
+
+ credp = crref();
+ code = afs_ReadNoCache(VTOAFS(FILE_INODE(fp)), ancr, credp);
crfree(credp);
+
return afs_convert_code(code);
}
+static inline int
+afs_linux_can_bypass(struct inode *ip) {
+ switch(cache_bypass_strategy) {
+ case NEVER_BYPASS_CACHE:
+ return 0;
+ case ALWAYS_BYPASS_CACHE:
+ return 1;
+ case LARGE_FILES_BYPASS_CACHE:
+ if(i_size_read(ip) > cache_bypass_threshold)
+ return 1;
+ default:
+ return 0;
+ }
+}
+
+/* Check if a file is permitted to bypass the cache by policy, and modify
+ * the cache bypass state recorded for that file */
+
+static inline int
+afs_linux_bypass_check(struct inode *ip) {
+ cred_t* credp;
+
+ int bypass = afs_linux_can_bypass(ip);
+
+ credp = crref();
+ trydo_cache_transition(VTOAFS(ip), credp, bypass);
+ crfree(credp);
+
+ return bypass;
+}
+
+#else
+static inline int
+afs_linux_bypass_check(struct inode *ip) {
+ return 0;
+}
+static inline int
+afs_linux_bypass_readpage(struct file *fp, struct page *pp) {
+ return 0;
+}
+static inline int
+afs_linux_bypass_readpages(struct file *fp, struct address_space *mapping,
+ struct list_head *page_list, unsigned int num_pages) {
+ return 0;
+}
+#endif
+
+static int
+afs_linux_readpage(struct file *fp, struct page *pp)
+{
+ int code;
+
+ if (afs_linux_bypass_check(FILE_INODE(fp))) {
+ code = afs_linux_bypass_readpage(fp, pp);
+ } else {
+ code = afs_linux_fillpage(fp, pp);
+ if (!code)
+ code = afs_linux_prefetch(fp, pp);
+ unlock_page(pp);
+ }
+
+ return code;
+}
+
/* Readpages reads a number of pages for a particular file. We use
* this to optimise the reading, by limiting the number of times upon which
* we have to lookup, lock and open vcaches and dcaches
struct pagevec lrupv;
struct afs_pagecopy_task *task;
-#if defined(AFS_CACHE_BYPASS)
- bypasscache = afs_linux_can_bypass(ip);
-
- /* In the new incarnation of selective caching, a file's caching policy
- * can change, eg because file size exceeds threshold, etc. */
- trydo_cache_transition(avc, credp, bypasscache);
+ if (afs_linux_bypass_check(inode))
+ return afs_linux_bypass_readpages(fp, mapping, page_list, num_pages);
- if (bypasscache)
- return afs_linux_cache_bypass_read(ip, mapping, page_list, num_pages);
-#endif
+ if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
+ return 0;
AFS_GLOCK();
if ((code = afs_linux_VerifyVCache(avc, NULL))) {
}
AFS_GUNLOCK();
if (tdc)
- cacheFp = afs_linux_raw_open(&tdc->f.inode, NULL);
+ cacheFp = afs_linux_raw_open(&tdc->f.inode);
}
if (tdc && !add_to_page_cache(page, mapping, page->index,
return 0;
}
+/* Prepare an AFS vcache for writeback. Should be called with the vcache
+ * locked */
+static inline int
+afs_linux_prepare_writeback(struct vcache *avc) {
+ if (avc->f.states & CPageWrite) {
+ return AOP_WRITEPAGE_ACTIVATE;
+ }
+ avc->f.states |= CPageWrite;
+ return 0;
+}
+
+static inline int
+afs_linux_dopartialwrite(struct vcache *avc, cred_t *credp) {
+ struct vrequest treq;
+ int code = 0;
+
+ if (!afs_InitReq(&treq, credp))
+ code = afs_DoPartialWrite(avc, &treq);
+
+ return afs_convert_code(code);
+}
+
+static inline void
+afs_linux_complete_writeback(struct vcache *avc) {
+ avc->f.states &= ~CPageWrite;
+}
+
+/* Writeback a given page syncronously. Called with no AFS locks held */
static int
-afs_linux_writepage_sync(struct inode *ip, struct page *pp,
- unsigned long offset, unsigned int count)
+afs_linux_page_writeback(struct inode *ip, struct page *pp,
+ unsigned long offset, unsigned int count,
+ cred_t *credp)
{
struct vcache *vcp = VTOAFS(ip);
char *buffer;
afs_offs_t base;
int code = 0;
- cred_t *credp;
uio_t tuio;
struct iovec iovec;
int f_flags = 0;
buffer = kmap(pp) + offset;
- base = (((loff_t) pp->index) << PAGE_CACHE_SHIFT) + offset;
+ base = page_offset(pp) + offset;
- credp = crref();
- afs_maybe_lock_kernel();
AFS_GLOCK();
afs_Trace4(afs_iclSetp, CM_TRACE_UPDATEPAGE, ICL_TYPE_POINTER, vcp,
ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, page_count(pp),
ICL_TYPE_INT32, 99999);
- ObtainWriteLock(&vcp->lock, 532);
- if (vcp->f.states & CPageWrite) {
- ReleaseWriteLock(&vcp->lock);
- AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
- crfree(credp);
- kunmap(pp);
- return AOP_WRITEPAGE_ACTIVATE;
- }
- vcp->f.states |= CPageWrite;
- ReleaseWriteLock(&vcp->lock);
-
setup_uio(&tuio, &iovec, buffer, base, count, UIO_WRITE, AFS_UIOSYS);
code = afs_write(vcp, &tuio, f_flags, credp, 0);
i_size_write(ip, vcp->f.m.Length);
ip->i_blocks = ((vcp->f.m.Length + 1023) >> 10) << 1;
- ObtainWriteLock(&vcp->lock, 533);
- if (!code) {
- struct vrequest treq;
-
- if (!afs_InitReq(&treq, credp))
- code = afs_DoPartialWrite(vcp, &treq);
- }
code = code ? afs_convert_code(code) : count - tuio.uio_resid;
- vcp->f.states &= ~CPageWrite;
- ReleaseWriteLock(&vcp->lock);
-
afs_Trace4(afs_iclSetp, CM_TRACE_UPDATEPAGE, ICL_TYPE_POINTER, vcp,
ICL_TYPE_POINTER, pp, ICL_TYPE_INT32, page_count(pp),
ICL_TYPE_INT32, code);
AFS_GUNLOCK();
- afs_maybe_unlock_kernel();
- crfree(credp);
kunmap(pp);
return code;
}
+static int
+afs_linux_writepage_sync(struct inode *ip, struct page *pp,
+ unsigned long offset, unsigned int count)
+{
+ int code;
+ int code1 = 0;
+ struct vcache *vcp = VTOAFS(ip);
+ cred_t *credp;
+
+ /* Catch recursive writeback. This occurs if the kernel decides
+ * writeback is required whilst we are writing to the cache, or
+ * flushing to the server. When we're running syncronously (as
+ * opposed to from writepage) we can't actually do anything about
+ * this case - as we can't return AOP_WRITEPAGE_ACTIVATE to write()
+ */
+ AFS_GLOCK();
+ ObtainWriteLock(&vcp->lock, 532);
+ afs_linux_prepare_writeback(vcp);
+ ReleaseWriteLock(&vcp->lock);
+ AFS_GUNLOCK();
+
+ credp = crref();
+ code = afs_linux_page_writeback(ip, pp, offset, count, credp);
+
+ AFS_GLOCK();
+ ObtainWriteLock(&vcp->lock, 533);
+ if (code > 0)
+ code1 = afs_linux_dopartialwrite(vcp, credp);
+ afs_linux_complete_writeback(vcp);
+ ReleaseWriteLock(&vcp->lock);
+ AFS_GUNLOCK();
+ crfree(credp);
+
+ if (code1)
+ return code1;
+
+ return code;
+}
static int
#ifdef AOP_WRITEPAGE_TAKES_WRITEBACK_CONTROL
{
struct address_space *mapping = pp->mapping;
struct inode *inode;
- unsigned long end_index;
- unsigned offset = PAGE_CACHE_SIZE;
- long status;
+ struct vcache *vcp;
+ cred_t *credp;
+ unsigned int to = PAGE_CACHE_SIZE;
+ loff_t isize;
+ int code = 0;
+ int code1 = 0;
if (PageReclaim(pp)) {
return AOP_WRITEPAGE_ACTIVATE;
+ /* XXX - Do we need to redirty the page here? */
}
- inode = (struct inode *)mapping->host;
- end_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
+ page_cache_get(pp);
+
+ inode = mapping->host;
+ vcp = VTOAFS(inode);
+ isize = i_size_read(inode);
+
+ /* Don't defeat an earlier truncate */
+ if (page_offset(pp) > isize) {
+ set_page_writeback(pp);
+ unlock_page(pp);
+ goto done;
+ }
+
+ AFS_GLOCK();
+ ObtainWriteLock(&vcp->lock, 537);
+ code = afs_linux_prepare_writeback(vcp);
+ if (code == AOP_WRITEPAGE_ACTIVATE) {
+ /* WRITEPAGE_ACTIVATE is the only return value that permits us
+ * to return with the page still locked */
+ ReleaseWriteLock(&vcp->lock);
+ AFS_GUNLOCK();
+ return code;
+ }
+
+ /* Grab the creds structure currently held in the vnode, and
+ * get a reference to it, in case it goes away ... */
+ credp = vcp->cred;
+ if (credp)
+ crhold(credp);
+ else
+ credp = crref();
+ ReleaseWriteLock(&vcp->lock);
+ AFS_GUNLOCK();
+
+ set_page_writeback(pp);
- /* easy case */
- if (pp->index < end_index)
- goto do_it;
- /* things got complicated... */
- offset = i_size_read(inode) & (PAGE_CACHE_SIZE - 1);
- /* OK, are we completely out? */
- if (pp->index >= end_index + 1 || !offset)
- return -EIO;
- do_it:
- status = afs_linux_writepage_sync(inode, pp, 0, offset);
SetPageUptodate(pp);
- if ( status != AOP_WRITEPAGE_ACTIVATE )
- UnlockPage(pp);
- if (status == offset)
+
+ /* We can unlock the page here, because it's protected by the
+ * page_writeback flag. This should make us less vulnerable to
+ * deadlocking in afs_write and afs_DoPartialWrite
+ */
+ unlock_page(pp);
+
+ /* If this is the final page, then just write the number of bytes that
+ * are actually in it */
+ if ((isize - page_offset(pp)) < to )
+ to = isize - page_offset(pp);
+
+ code = afs_linux_page_writeback(inode, pp, 0, to, credp);
+
+ AFS_GLOCK();
+ ObtainWriteLock(&vcp->lock, 538);
+
+ /* As much as we might like to ignore a file server error here,
+ * and just try again when we close(), unfortunately StoreAllSegments
+ * will invalidate our chunks if the server returns a permanent error,
+ * so we need to at least try and get that error back to the user
+ */
+ if (code == to)
+ code1 = afs_linux_dopartialwrite(vcp, credp);
+
+ afs_linux_complete_writeback(vcp);
+ ReleaseWriteLock(&vcp->lock);
+ crfree(credp);
+ AFS_GUNLOCK();
+
+done:
+ end_page_writeback(pp);
+ page_cache_release(pp);
+
+ if (code1)
+ return code1;
+
+ if (code == to)
return 0;
- else
- return status;
+
+ return code;
}
/* afs_linux_permission
return afs_convert_code(code);
}
-#if !defined(HAVE_WRITE_BEGIN)
static int
afs_linux_commit_write(struct file *file, struct page *page, unsigned offset,
unsigned to)
{
int code;
+ struct inode *inode = FILE_INODE(file);
+ loff_t pagebase = page_offset(page);
+
+ if (i_size_read(inode) < (pagebase + offset))
+ i_size_write(inode, pagebase + offset);
+
+ if (PageChecked(page)) {
+ SetPageUptodate(page);
+ ClearPageChecked(page);
+ }
- code = afs_linux_writepage_sync(file->f_dentry->d_inode, page,
- offset, to - offset);
+ code = afs_linux_writepage_sync(inode, page, offset, to - offset);
return code;
}
afs_linux_prepare_write(struct file *file, struct page *page, unsigned from,
unsigned to)
{
+
+ /* http://kerneltrap.org/node/4941 details the expected behaviour of
+ * prepare_write. Essentially, if the page exists within the file,
+ * and is not being fully written, then we should populate it.
+ */
+
+ if (!PageUptodate(page)) {
+ loff_t pagebase = page_offset(page);
+ loff_t isize = i_size_read(page->mapping->host);
+
+ /* Is the location we are writing to beyond the end of the file? */
+ if (pagebase >= isize ||
+ ((from == 0) && (pagebase + to) >= isize)) {
+ zero_user_segments(page, 0, from, to, PAGE_CACHE_SIZE);
+ SetPageChecked(page);
+ /* Are we we writing a full page */
+ } else if (from == 0 && to == PAGE_CACHE_SIZE) {
+ SetPageChecked(page);
+ /* Is the page readable, if it's wronly, we don't care, because we're
+ * not actually going to read from it ... */
+ } else if ((file->f_flags && O_ACCMODE) != O_WRONLY) {
+ /* We don't care if fillpage fails, because if it does the page
+ * won't be marked as up to date
+ */
+ afs_linux_fillpage(file, page);
+ }
+ }
return 0;
}
-#else
+#if defined(STRUCT_ADDRESS_SPACE_OPERATIONS_HAS_WRITE_BEGIN)
static int
afs_linux_write_end(struct file *file, struct address_space *mapping,
loff_t pos, unsigned len, unsigned copied,
struct page *page, void *fsdata)
{
int code;
- unsigned from = pos & (PAGE_CACHE_SIZE - 1);
+ unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
+
+ code = afs_linux_commit_write(file, page, from, from + len);
- code = afs_linux_writepage_sync(file->f_dentry->d_inode, page,
- from, copied);
unlock_page(page);
page_cache_release(page);
return code;
{
struct page *page;
pgoff_t index = pos >> PAGE_CACHE_SHIFT;
+ unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
+ int code;
+
page = grab_cache_page_write_begin(mapping, index, flags);
*pagep = page;
- return 0;
+ code = afs_linux_prepare_write(file, page, from, from + len);
+ if (code) {
+ unlock_page(page);
+ page_cache_release(page);
+ }
+
+ return code;
}
#endif
.readpage = afs_linux_readpage,
.readpages = afs_linux_readpages,
.writepage = afs_linux_writepage,
-#if defined (HAVE_WRITE_BEGIN)
+#if defined (STRUCT_ADDRESS_SPACE_OPERATIONS_HAS_WRITE_BEGIN)
.write_begin = afs_linux_write_begin,
.write_end = afs_linux_write_end,
#else
char *p = (char *)kmap(page);
int code;
- afs_maybe_lock_kernel();
AFS_GLOCK();
code = afs_linux_ireadlink(ip, p, PAGE_SIZE, AFS_UIOSYS);
AFS_GUNLOCK();
if (code < 0)
goto fail;
p[code] = '\0'; /* null terminate? */
- afs_maybe_unlock_kernel();
SetPageUptodate(page);
kunmap(page);
- UnlockPage(page);
+ unlock_page(page);
return 0;
fail:
- afs_maybe_unlock_kernel();
-
SetPageError(page);
kunmap(page);
- UnlockPage(page);
+ unlock_page(page);
return code;
}
static struct inode_operations afs_symlink_iops = {
#if defined(USABLE_KERNEL_PAGE_SYMLINK_CACHE)
.readlink = page_readlink,
-# if defined(HAVE_KERNEL_PAGE_FOLLOW_LINK)
+# if defined(HAVE_LINUX_PAGE_FOLLOW_LINK)
.follow_link = page_follow_link,
# else
.follow_link = page_follow_link_light,
if (vattr)
vattr2inode(ip, vattr);
- ip->i_mapping->backing_dev_info = &afs_backing_dev_info;
+ ip->i_mapping->backing_dev_info = afs_backing_dev_info;
/* Reset ops if symlink or directory. */
if (S_ISREG(ip->i_mode)) {
ip->i_op = &afs_file_iops;
git://git.openafs.org / openafs.git / blobdiff