Merge commit 'v2.6.30' into for-2.6.31
[safe/jmp/linux-2.6] / fs / nfs / file.c
index 4e98a56..ec7e27d 100644 (file)
@@ -35,6 +35,7 @@
 #include "delegation.h"
 #include "internal.h"
 #include "iostat.h"
+#include "fscache.h"
 
 #define NFSDBG_FACILITY                NFSDBG_FILE
 
@@ -50,7 +51,7 @@ static ssize_t nfs_file_read(struct kiocb *, const struct iovec *iov,
 static ssize_t nfs_file_write(struct kiocb *, const struct iovec *iov,
                                unsigned long nr_segs, loff_t pos);
 static int  nfs_file_flush(struct file *, fl_owner_t id);
-static int  nfs_fsync(struct file *, struct dentry *dentry, int datasync);
+static int  nfs_file_fsync(struct file *, struct dentry *dentry, int datasync);
 static int nfs_check_flags(int flags);
 static int nfs_lock(struct file *filp, int cmd, struct file_lock *fl);
 static int nfs_flock(struct file *filp, int cmd, struct file_lock *fl);
@@ -64,15 +65,11 @@ const struct file_operations nfs_file_operations = {
        .write          = do_sync_write,
        .aio_read       = nfs_file_read,
        .aio_write      = nfs_file_write,
-#ifdef CONFIG_MMU
        .mmap           = nfs_file_mmap,
-#else
-       .mmap           = generic_file_mmap,
-#endif
        .open           = nfs_file_open,
        .flush          = nfs_file_flush,
        .release        = nfs_file_release,
-       .fsync          = nfs_fsync,
+       .fsync          = nfs_file_fsync,
        .lock           = nfs_lock,
        .flock          = nfs_flock,
        .splice_read    = nfs_file_splice_read,
@@ -119,25 +116,30 @@ nfs_file_open(struct inode *inode, struct file *filp)
 {
        int res;
 
+       dprintk("NFS: open file(%s/%s)\n",
+                       filp->f_path.dentry->d_parent->d_name.name,
+                       filp->f_path.dentry->d_name.name);
+
        res = nfs_check_flags(filp->f_flags);
        if (res)
                return res;
 
        nfs_inc_stats(inode, NFSIOS_VFSOPEN);
-       lock_kernel();
-       res = NFS_PROTO(inode)->file_open(inode, filp);
-       unlock_kernel();
+       res = nfs_open(inode, filp);
        return res;
 }
 
 static int
 nfs_file_release(struct inode *inode, struct file *filp)
 {
-       /* Ensure that dirty pages are flushed out with the right creds */
-       if (filp->f_mode & FMODE_WRITE)
-               nfs_wb_all(filp->f_path.dentry->d_inode);
+       struct dentry *dentry = filp->f_path.dentry;
+
+       dprintk("NFS: release(%s/%s)\n",
+                       dentry->d_parent->d_name.name,
+                       dentry->d_name.name);
+
        nfs_inc_stats(inode, NFSIOS_VFSRELEASE);
-       return NFS_PROTO(inode)->file_release(inode, filp);
+       return nfs_release(inode, filp);
 }
 
 /**
@@ -171,21 +173,30 @@ force_reval:
 static loff_t nfs_file_llseek(struct file *filp, loff_t offset, int origin)
 {
        loff_t loff;
+
+       dprintk("NFS: llseek file(%s/%s, %lld, %d)\n",
+                       filp->f_path.dentry->d_parent->d_name.name,
+                       filp->f_path.dentry->d_name.name,
+                       offset, origin);
+
        /* origin == SEEK_END => we must revalidate the cached file length */
        if (origin == SEEK_END) {
                struct inode *inode = filp->f_mapping->host;
+
                int retval = nfs_revalidate_file_size(inode, filp);
                if (retval < 0)
                        return (loff_t)retval;
-       }
-       lock_kernel();  /* BKL needed? */
-       loff = generic_file_llseek_unlocked(filp, offset, origin);
-       unlock_kernel();
+
+               spin_lock(&inode->i_lock);
+               loff = generic_file_llseek_unlocked(filp, offset, origin);
+               spin_unlock(&inode->i_lock);
+       } else
+               loff = generic_file_llseek_unlocked(filp, offset, origin);
        return loff;
 }
 
 /*
- * Helper for nfs_file_flush() and nfs_fsync()
+ * Helper for nfs_file_flush() and nfs_file_fsync()
  *
  * Notice that it clears the NFS_CONTEXT_ERROR_WRITE before synching to
  * disk, but it retrieves and clears ctx->error after synching, despite
@@ -211,26 +222,24 @@ static int nfs_do_fsync(struct nfs_open_context *ctx, struct inode *inode)
 
 /*
  * Flush all dirty pages, and check for write errors.
- *
  */
 static int
 nfs_file_flush(struct file *file, fl_owner_t id)
 {
        struct nfs_open_context *ctx = nfs_file_open_context(file);
-       struct inode    *inode = file->f_path.dentry->d_inode;
-       int             status;
+       struct dentry   *dentry = file->f_path.dentry;
+       struct inode    *inode = dentry->d_inode;
 
-       dfprintk(VFS, "nfs: flush(%s/%ld)\n", inode->i_sb->s_id, inode->i_ino);
+       dprintk("NFS: flush(%s/%s)\n",
+                       dentry->d_parent->d_name.name,
+                       dentry->d_name.name);
 
        if ((file->f_mode & FMODE_WRITE) == 0)
                return 0;
        nfs_inc_stats(inode, NFSIOS_VFSFLUSH);
 
-       /* Ensure that data+attribute caches are up to date after close() */
-       status = nfs_do_fsync(ctx, inode);
-       if (!status)
-               nfs_revalidate_inode(NFS_SERVER(inode), inode);
-       return status;
+       /* Flush writes to the server and return any errors */
+       return nfs_do_fsync(ctx, inode);
 }
 
 static ssize_t
@@ -245,7 +254,7 @@ nfs_file_read(struct kiocb *iocb, const struct iovec *iov,
        if (iocb->ki_filp->f_flags & O_DIRECT)
                return nfs_file_direct_read(iocb, iov, nr_segs, pos);
 
-       dfprintk(VFS, "nfs: read(%s/%s, %lu@%lu)\n",
+       dprintk("NFS: read(%s/%s, %lu@%lu)\n",
                dentry->d_parent->d_name.name, dentry->d_name.name,
                (unsigned long) count, (unsigned long) pos);
 
@@ -265,7 +274,7 @@ nfs_file_splice_read(struct file *filp, loff_t *ppos,
        struct inode *inode = dentry->d_inode;
        ssize_t res;
 
-       dfprintk(VFS, "nfs: splice_read(%s/%s, %lu@%Lu)\n",
+       dprintk("NFS: splice_read(%s/%s, %lu@%Lu)\n",
                dentry->d_parent->d_name.name, dentry->d_name.name,
                (unsigned long) count, (unsigned long long) *ppos);
 
@@ -282,14 +291,16 @@ nfs_file_mmap(struct file * file, struct vm_area_struct * vma)
        struct inode *inode = dentry->d_inode;
        int     status;
 
-       dfprintk(VFS, "nfs: mmap(%s/%s)\n",
+       dprintk("NFS: mmap(%s/%s)\n",
                dentry->d_parent->d_name.name, dentry->d_name.name);
 
-       status = nfs_revalidate_mapping(inode, file->f_mapping);
+       /* Note: generic_file_mmap() returns ENOSYS on nommu systems
+        *       so we call that before revalidating the mapping
+        */
+       status = generic_file_mmap(file, vma);
        if (!status) {
                vma->vm_ops = &nfs_file_vm_ops;
-               vma->vm_flags |= VM_CAN_NONLINEAR;
-               file_accessed(file);
+               status = nfs_revalidate_mapping(inode, file->f_mapping);
        }
        return status;
 }
@@ -300,12 +311,14 @@ nfs_file_mmap(struct file * file, struct vm_area_struct * vma)
  * whether any write errors occurred for this process.
  */
 static int
-nfs_fsync(struct file *file, struct dentry *dentry, int datasync)
+nfs_file_fsync(struct file *file, struct dentry *dentry, int datasync)
 {
        struct nfs_open_context *ctx = nfs_file_open_context(file);
        struct inode *inode = dentry->d_inode;
 
-       dfprintk(VFS, "nfs: fsync(%s/%ld)\n", inode->i_sb->s_id, inode->i_ino);
+       dprintk("NFS: fsync file(%s/%s) datasync %d\n",
+                       dentry->d_parent->d_name.name, dentry->d_name.name,
+                       datasync);
 
        nfs_inc_stats(inode, NFSIOS_VFSFSYNC);
        return nfs_do_fsync(ctx, inode);
@@ -328,7 +341,21 @@ static int nfs_write_begin(struct file *file, struct address_space *mapping,
        struct page *page;
        index = pos >> PAGE_CACHE_SHIFT;
 
-       page = __grab_cache_page(mapping, index);
+       dfprintk(PAGECACHE, "NFS: write_begin(%s/%s(%ld), %u@%lld)\n",
+               file->f_path.dentry->d_parent->d_name.name,
+               file->f_path.dentry->d_name.name,
+               mapping->host->i_ino, len, (long long) pos);
+
+       /*
+        * Prevent starvation issues if someone is doing a consistency
+        * sync-to-disk
+        */
+       ret = wait_on_bit(&NFS_I(mapping->host)->flags, NFS_INO_FLUSHING,
+                       nfs_wait_bit_killable, TASK_KILLABLE);
+       if (ret)
+               return ret;
+
+       page = grab_cache_page_write_begin(mapping, index, flags);
        if (!page)
                return -ENOMEM;
        *pagep = page;
@@ -348,9 +375,32 @@ static int nfs_write_end(struct file *file, struct address_space *mapping,
        unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
        int status;
 
-       lock_kernel();
+       dfprintk(PAGECACHE, "NFS: write_end(%s/%s(%ld), %u@%lld)\n",
+               file->f_path.dentry->d_parent->d_name.name,
+               file->f_path.dentry->d_name.name,
+               mapping->host->i_ino, len, (long long) pos);
+
+       /*
+        * Zero any uninitialised parts of the page, and then mark the page
+        * as up to date if it turns out that we're extending the file.
+        */
+       if (!PageUptodate(page)) {
+               unsigned pglen = nfs_page_length(page);
+               unsigned end = offset + len;
+
+               if (pglen == 0) {
+                       zero_user_segments(page, 0, offset,
+                                       end, PAGE_CACHE_SIZE);
+                       SetPageUptodate(page);
+               } else if (end >= pglen) {
+                       zero_user_segment(page, end, PAGE_CACHE_SIZE);
+                       if (offset == 0)
+                               SetPageUptodate(page);
+               } else
+                       zero_user_segment(page, pglen, PAGE_CACHE_SIZE);
+       }
+
        status = nfs_updatepage(file, page, offset, copied);
-       unlock_kernel();
 
        unlock_page(page);
        page_cache_release(page);
@@ -360,23 +410,59 @@ static int nfs_write_end(struct file *file, struct address_space *mapping,
        return copied;
 }
 
+/*
+ * Partially or wholly invalidate a page
+ * - Release the private state associated with a page if undergoing complete
+ *   page invalidation
+ * - Called if either PG_private or PG_fscache is set on the page
+ * - Caller holds page lock
+ */
 static void nfs_invalidate_page(struct page *page, unsigned long offset)
 {
+       dfprintk(PAGECACHE, "NFS: invalidate_page(%p, %lu)\n", page, offset);
+
        if (offset != 0)
                return;
        /* Cancel any unstarted writes on this page */
        nfs_wb_page_cancel(page->mapping->host, page);
+
+       nfs_fscache_invalidate_page(page, page->mapping->host);
 }
 
+/*
+ * Attempt to release the private state associated with a page
+ * - Called if either PG_private or PG_fscache is set on the page
+ * - Caller holds page lock
+ * - Return true (may release page) or false (may not)
+ */
 static int nfs_release_page(struct page *page, gfp_t gfp)
 {
+       dfprintk(PAGECACHE, "NFS: release_page(%p)\n", page);
+
        /* If PagePrivate() is set, then the page is not freeable */
-       return 0;
+       if (PagePrivate(page))
+               return 0;
+       return nfs_fscache_release_page(page, gfp);
 }
 
+/*
+ * Attempt to clear the private state associated with a page when an error
+ * occurs that requires the cached contents of an inode to be written back or
+ * destroyed
+ * - Called if either PG_private or fscache is set on the page
+ * - Caller holds page lock
+ * - Return 0 if successful, -error otherwise
+ */
 static int nfs_launder_page(struct page *page)
 {
-       return nfs_wb_page(page->mapping->host, page);
+       struct inode *inode = page->mapping->host;
+       struct nfs_inode *nfsi = NFS_I(inode);
+
+       dfprintk(PAGECACHE, "NFS: launder_page(%ld, %llu)\n",
+               inode->i_ino, (long long)page_offset(page));
+
+       nfs_fscache_wait_on_page_write(nfsi, page);
+       return nfs_wb_page(inode, page);
 }
 
 const struct address_space_operations nfs_file_aops = {
@@ -393,16 +479,31 @@ const struct address_space_operations nfs_file_aops = {
        .launder_page = nfs_launder_page,
 };
 
-static int nfs_vm_page_mkwrite(struct vm_area_struct *vma, struct page *page)
+/*
+ * Notification that a PTE pointing to an NFS page is about to be made
+ * writable, implying that someone is about to modify the page through a
+ * shared-writable mapping
+ */
+static int nfs_vm_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
+       struct page *page = vmf->page;
        struct file *filp = vma->vm_file;
+       struct dentry *dentry = filp->f_path.dentry;
        unsigned pagelen;
        int ret = -EINVAL;
        struct address_space *mapping;
 
+       dfprintk(PAGECACHE, "NFS: vm_page_mkwrite(%s/%s(%ld), offset %lld)\n",
+               dentry->d_parent->d_name.name, dentry->d_name.name,
+               filp->f_mapping->host->i_ino,
+               (long long)page_offset(page));
+
+       /* make sure the cache has finished storing the page */
+       nfs_fscache_wait_on_page_write(NFS_I(dentry->d_inode), page);
+
        lock_page(page);
        mapping = page->mapping;
-       if (mapping != vma->vm_file->f_path.dentry->d_inode->i_mapping)
+       if (mapping != dentry->d_inode->i_mapping)
                goto out_unlock;
 
        ret = 0;
@@ -415,11 +516,11 @@ static int nfs_vm_page_mkwrite(struct vm_area_struct *vma, struct page *page)
                goto out_unlock;
 
        ret = nfs_updatepage(filp, page, 0, pagelen);
-       if (ret == 0)
-               ret = pagelen;
 out_unlock:
+       if (!ret)
+               return VM_FAULT_LOCKED;
        unlock_page(page);
-       return ret;
+       return VM_FAULT_SIGBUS;
 }
 
 static struct vm_operations_struct nfs_file_vm_ops = {
@@ -450,9 +551,9 @@ static ssize_t nfs_file_write(struct kiocb *iocb, const struct iovec *iov,
        if (iocb->ki_filp->f_flags & O_DIRECT)
                return nfs_file_direct_write(iocb, iov, nr_segs, pos);
 
-       dfprintk(VFS, "nfs: write(%s/%s(%ld), %lu@%Ld)\n",
+       dprintk("NFS: write(%s/%s, %lu@%Ld)\n",
                dentry->d_parent->d_name.name, dentry->d_name.name,
-               inode->i_ino, (unsigned long) count, (long long) pos);
+               (unsigned long) count, (long long) pos);
 
        result = -EBUSY;
        if (IS_SWAPFILE(inode))
@@ -586,7 +687,8 @@ static int do_setlk(struct file *filp, int cmd, struct file_lock *fl)
         * This makes locking act as a cache coherency point.
         */
        nfs_sync_mapping(filp->f_mapping);
-       nfs_zap_caches(inode);
+       if (!nfs_have_delegation(inode, FMODE_READ))
+               nfs_zap_caches(inode);
 out:
        return status;
 }
@@ -596,23 +698,35 @@ out:
  */
 static int nfs_lock(struct file *filp, int cmd, struct file_lock *fl)
 {
-       struct inode * inode = filp->f_mapping->host;
+       struct inode *inode = filp->f_mapping->host;
+       int ret = -ENOLCK;
 
-       dprintk("NFS: nfs_lock(f=%s/%ld, t=%x, fl=%x, r=%Ld:%Ld)\n",
-                       inode->i_sb->s_id, inode->i_ino,
+       dprintk("NFS: lock(%s/%s, t=%x, fl=%x, r=%lld:%lld)\n",
+                       filp->f_path.dentry->d_parent->d_name.name,
+                       filp->f_path.dentry->d_name.name,
                        fl->fl_type, fl->fl_flags,
                        (long long)fl->fl_start, (long long)fl->fl_end);
+
        nfs_inc_stats(inode, NFSIOS_VFSLOCK);
 
        /* No mandatory locks over NFS */
        if (__mandatory_lock(inode) && fl->fl_type != F_UNLCK)
-               return -ENOLCK;
+               goto out_err;
+
+       if (NFS_PROTO(inode)->lock_check_bounds != NULL) {
+               ret = NFS_PROTO(inode)->lock_check_bounds(fl);
+               if (ret < 0)
+                       goto out_err;
+       }
 
        if (IS_GETLK(cmd))
-               return do_getlk(filp, cmd, fl);
-       if (fl->fl_type == F_UNLCK)
-               return do_unlk(filp, cmd, fl);
-       return do_setlk(filp, cmd, fl);
+               ret = do_getlk(filp, cmd, fl);
+       else if (fl->fl_type == F_UNLCK)
+               ret = do_unlk(filp, cmd, fl);
+       else
+               ret = do_setlk(filp, cmd, fl);
+out_err:
+       return ret;
 }
 
 /*
@@ -620,18 +734,11 @@ static int nfs_lock(struct file *filp, int cmd, struct file_lock *fl)
  */
 static int nfs_flock(struct file *filp, int cmd, struct file_lock *fl)
 {
-       dprintk("NFS: nfs_flock(f=%s/%ld, t=%x, fl=%x)\n",
-                       filp->f_path.dentry->d_inode->i_sb->s_id,
-                       filp->f_path.dentry->d_inode->i_ino,
+       dprintk("NFS: flock(%s/%s, t=%x, fl=%x)\n",
+                       filp->f_path.dentry->d_parent->d_name.name,
+                       filp->f_path.dentry->d_name.name,
                        fl->fl_type, fl->fl_flags);
 
-       /*
-        * No BSD flocks over NFS allowed.
-        * Note: we could try to fake a POSIX lock request here by
-        * using ((u32) filp | 0x80000000) or some such as the pid.
-        * Not sure whether that would be unique, though, or whether
-        * that would break in other places.
-        */
        if (!(fl->fl_flags & FL_FLOCK))
                return -ENOLCK;
 
@@ -645,12 +752,15 @@ static int nfs_flock(struct file *filp, int cmd, struct file_lock *fl)
        return do_setlk(filp, cmd, fl);
 }
 
+/*
+ * There is no protocol support for leases, so we have no way to implement
+ * them correctly in the face of opens by other clients.
+ */
 static int nfs_setlease(struct file *file, long arg, struct file_lock **fl)
 {
-       /*
-        * There is no protocol support for leases, so we have no way
-        * to implement them correctly in the face of opens by other
-        * clients.
-        */
+       dprintk("NFS: setlease(%s/%s, arg=%ld)\n",
+                       file->f_path.dentry->d_parent->d_name.name,
+                       file->f_path.dentry->d_name.name, arg);
+
        return -EINVAL;
 }