Merge branch 'splice' of git://brick.kernel.dk/data/git/linux-2.6-block
authorLinus Torvalds <torvalds@g5.osdl.org>
Wed, 19 Apr 2006 16:25:52 +0000 (09:25 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 19 Apr 2006 16:25:52 +0000 (09:25 -0700)
* 'splice' of git://brick.kernel.dk/data/git/linux-2.6-block:
  [PATCH] splice: fixup writeout path after ->map changes
  [PATCH] splice: offset fixes
  [PATCH] tee: link_pipe() must be careful when dropping one of the pipe locks
  [PATCH] splice: cleanup the SPLICE_F_NONBLOCK handling
  [PATCH] splice: close i_size truncate races on read

fs/splice.c

index 8d57e89..22fac87 100644 (file)
@@ -50,7 +50,8 @@ static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
        struct page *page = buf->page;
        struct address_space *mapping = page_mapping(page);
 
-       WARN_ON(!PageLocked(page));
+       lock_page(page);
+
        WARN_ON(!PageUptodate(page));
 
        /*
@@ -65,8 +66,10 @@ static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
        if (PagePrivate(page))
                try_to_release_page(page, mapping_gfp_mask(mapping));
 
-       if (!remove_mapping(mapping, page))
+       if (!remove_mapping(mapping, page)) {
+               unlock_page(page);
                return 1;
+       }
 
        buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
        return 0;
@@ -145,8 +148,8 @@ static struct pipe_buf_operations page_cache_pipe_buf_ops = {
  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
  */
 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
-                           int nr_pages, unsigned long offset,
-                           unsigned long len, unsigned int flags)
+                           int nr_pages, unsigned long len,
+                           unsigned int offset, unsigned int flags)
 {
        int ret, do_wakeup, i;
 
@@ -243,14 +246,16 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                           unsigned int flags)
 {
        struct address_space *mapping = in->f_mapping;
-       unsigned int offset, nr_pages;
+       unsigned int loff, offset, nr_pages;
        struct page *pages[PIPE_BUFFERS];
        struct page *page;
-       pgoff_t index;
+       pgoff_t index, end_index;
+       loff_t isize;
+       size_t bytes;
        int i, error;
 
        index = *ppos >> PAGE_CACHE_SHIFT;
-       offset = *ppos & ~PAGE_CACHE_MASK;
+       loff = offset = *ppos & ~PAGE_CACHE_MASK;
        nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 
        if (nr_pages > PIPE_BUFFERS)
@@ -268,6 +273,7 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
         * Now fill in the holes:
         */
        error = 0;
+       bytes = 0;
        for (i = 0; i < nr_pages; i++, index++) {
 find_page:
                /*
@@ -275,14 +281,6 @@ find_page:
                 */
                page = find_get_page(mapping, index);
                if (!page) {
-                       /*
-                        * If in nonblock mode then dont block on
-                        * readpage (we've kicked readahead so there
-                        * will be asynchronous progress):
-                        */
-                       if (flags & SPLICE_F_NONBLOCK)
-                               break;
-
                        /*
                         * page didn't exist, allocate one
                         */
@@ -304,6 +302,13 @@ find_page:
                 * If the page isn't uptodate, we may need to start io on it
                 */
                if (!PageUptodate(page)) {
+                       /*
+                        * If in nonblock mode then dont block on waiting
+                        * for an in-flight io page
+                        */
+                       if (flags & SPLICE_F_NONBLOCK)
+                               break;
+
                        lock_page(page);
 
                        /*
@@ -336,13 +341,41 @@ readpage:
                                        goto find_page;
                                break;
                        }
+
+                       /*
+                        * i_size must be checked after ->readpage().
+                        */
+                       isize = i_size_read(mapping->host);
+                       end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
+                       if (unlikely(!isize || index > end_index)) {
+                               page_cache_release(page);
+                               break;
+                       }
+
+                       /*
+                        * if this is the last page, see if we need to shrink
+                        * the length and stop
+                        */
+                       if (end_index == index) {
+                               loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
+                               if (bytes + loff > isize) {
+                                       page_cache_release(page);
+                                       break;
+                               }
+                               /*
+                                * force quit after adding this page
+                                */
+                               nr_pages = i;
+                       }
                }
 fill_it:
                pages[i] = page;
+               bytes += PAGE_CACHE_SIZE - loff;
+               loff = 0;
        }
 
        if (i)
-               return move_to_pipe(pipe, pages, i, offset, len, flags);
+               return move_to_pipe(pipe, pages, i, bytes, offset, flags);
 
        return error;
 }
@@ -369,17 +402,20 @@ ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
        while (len) {
                ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 
-               if (ret <= 0)
+               if (ret < 0)
                        break;
+               else if (!ret) {
+                       if (spliced)
+                               break;
+                       if (flags & SPLICE_F_NONBLOCK) {
+                               ret = -EAGAIN;
+                               break;
+                       }
+               }
 
                *ppos += ret;
                len -= ret;
                spliced += ret;
-
-               if (!(flags & SPLICE_F_NONBLOCK))
-                       continue;
-               ret = -EAGAIN;
-               break;
        }
 
        if (spliced)
@@ -474,14 +510,12 @@ static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
        if (sd->flags & SPLICE_F_MOVE) {
                /*
                 * If steal succeeds, buf->page is now pruned from the vm
-                * side (LRU and page cache) and we can reuse it.
+                * side (LRU and page cache) and we can reuse it. The page
+                * will also be looked on successful return.
                 */
                if (buf->ops->steal(info, buf))
                        goto find_page;
 
-               /*
-                * this will also set the page locked
-                */
                page = buf->page;
                if (add_to_page_cache(page, mapping, index, gfp_mask))
                        goto find_page;
@@ -490,15 +524,27 @@ static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
                        lru_cache_add(page);
        } else {
 find_page:
-               ret = -ENOMEM;
-               page = find_or_create_page(mapping, index, gfp_mask);
-               if (!page)
-                       goto out_nomem;
+               page = find_lock_page(mapping, index);
+               if (!page) {
+                       ret = -ENOMEM;
+                       page = page_cache_alloc_cold(mapping);
+                       if (unlikely(!page))
+                               goto out_nomem;
+
+                       /*
+                        * This will also lock the page
+                        */
+                       ret = add_to_page_cache_lru(page, mapping, index,
+                                                   gfp_mask);
+                       if (unlikely(ret))
+                               goto out;
+               }
 
                /*
-                * If the page is uptodate, it is also locked. If it isn't
-                * uptodate, we can mark it uptodate if we are filling the
-                * full page. Otherwise we need to read it in first...
+                * We get here with the page locked. If the page is also
+                * uptodate, we don't need to do more. If it isn't, we
+                * may need to bring it in if we are not going to overwrite
+                * the full page.
                 */
                if (!PageUptodate(page)) {
                        if (sd->len < PAGE_CACHE_SIZE) {
@@ -520,10 +566,8 @@ find_page:
                                        ret = -EIO;
                                        goto out;
                                }
-                       } else {
-                               WARN_ON(!PageLocked(page));
+                       } else
                                SetPageUptodate(page);
-                       }
                }
        }
 
@@ -552,10 +596,10 @@ find_page:
        mark_page_accessed(page);
        balance_dirty_pages_ratelimited(mapping);
 out:
-       if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
+       if (!(buf->flags & PIPE_BUF_FLAG_STOLEN))
                page_cache_release(page);
-               unlock_page(page);
-       }
+
+       unlock_page(page);
 out_nomem:
        buf->ops->unmap(info, buf);
        return ret;
@@ -687,22 +731,26 @@ generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
        ssize_t ret;
 
        ret = move_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
-
-       /*
-        * If file or inode is SYNC and we actually wrote some data, sync it.
-        */
-       if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
-           && ret > 0) {
+       if (ret > 0) {
                struct inode *inode = mapping->host;
-               int err;
 
-               mutex_lock(&inode->i_mutex);
-               err = generic_osync_inode(mapping->host, mapping,
-                                         OSYNC_METADATA|OSYNC_DATA);
-               mutex_unlock(&inode->i_mutex);
+               *ppos += ret;
+
+               /*
+                * If file or inode is SYNC and we actually wrote some data,
+                * sync it.
+                */
+               if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
+                       int err;
+
+                       mutex_lock(&inode->i_mutex);
+                       err = generic_osync_inode(inode, mapping,
+                                                 OSYNC_METADATA|OSYNC_DATA);
+                       mutex_unlock(&inode->i_mutex);
 
-               if (err)
-                       ret = err;
+                       if (err)
+                               ret = err;
+               }
        }
 
        return ret;
@@ -904,6 +952,7 @@ static long do_splice(struct file *in, loff_t __user *off_in,
 {
        struct pipe_inode_info *pipe;
        loff_t offset, *off;
+       long ret;
 
        pipe = in->f_dentry->d_inode->i_pipe;
        if (pipe) {
@@ -918,7 +967,12 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                } else
                        off = &out->f_pos;
 
-               return do_splice_from(pipe, out, off, len, flags);
+               ret = do_splice_from(pipe, out, off, len, flags);
+
+               if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
+                       ret = -EFAULT;
+
+               return ret;
        }
 
        pipe = out->f_dentry->d_inode->i_pipe;
@@ -934,7 +988,12 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                } else
                        off = &in->f_pos;
 
-               return do_splice_to(in, off, pipe, len, flags);
+               ret = do_splice_to(in, off, pipe, len, flags);
+
+               if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
+                       ret = -EFAULT;
+
+               return ret;
        }
 
        return -EINVAL;
@@ -979,7 +1038,9 @@ static int link_pipe(struct pipe_inode_info *ipipe,
                     size_t len, unsigned int flags)
 {
        struct pipe_buffer *ibuf, *obuf;
-       int ret = 0, do_wakeup = 0, i;
+       int ret, do_wakeup, i, ipipe_first;
+
+       ret = do_wakeup = ipipe_first = 0;
 
        /*
         * Potential ABBA deadlock, work around it by ordering lock
@@ -987,6 +1048,7 @@ static int link_pipe(struct pipe_inode_info *ipipe,
         * could deadlock (one doing tee from A -> B, the other from B -> A).
         */
        if (ipipe->inode < opipe->inode) {
+               ipipe_first = 1;
                mutex_lock(&ipipe->inode->i_mutex);
                mutex_lock(&opipe->inode->i_mutex);
        } else {
@@ -1035,9 +1097,11 @@ static int link_pipe(struct pipe_inode_info *ipipe,
 
                        /*
                         * We have input available, but no output room.
-                        * If we already copied data, return that.
+                        * If we already copied data, return that. If we
+                        * need to drop the opipe lock, it must be ordered
+                        * last to avoid deadlocks.
                         */
-                       if (flags & SPLICE_F_NONBLOCK) {
+                       if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
                                if (!ret)
                                        ret = -EAGAIN;
                                break;
@@ -1071,7 +1135,12 @@ static int link_pipe(struct pipe_inode_info *ipipe,
                        if (ret)
                                break;
                }
-               if (flags & SPLICE_F_NONBLOCK) {
+               /*
+                * pipe_wait() drops the ipipe mutex. To avoid deadlocks
+                * with another process, we can only safely do that if
+                * the ipipe lock is ordered last.
+                */
+               if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
                        if (!ret)
                                ret = -EAGAIN;
                        break;