fs: fix nobh error handling
authorNick Piggin <npiggin@suse.de>
Tue, 16 Oct 2007 08:24:48 +0000 (01:24 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 16 Oct 2007 16:42:54 +0000 (09:42 -0700)
nobh mode error handling is not just pretty slack, it's wrong.

One cannot zero out the whole page to ensure new blocks are zeroed, because
it just brings the whole page "uptodate" with zeroes even if that may not
be the correct uptodate data.  Also, other parts of the page may already
contain dirty data which would get lost by zeroing it out.  Thirdly, the
writeback of zeroes to the new blocks will also erase existing blocks.  All
these conditions are pagecache and/or filesystem corruption.

The problem comes about because we didn't keep track of which buffers
actually are new or old.  However it is not enough just to keep only this
state, because at the point we start dirtying parts of the page (new
blocks, with zeroes), the handling of IO errors becomes impossible without
buffers because the page may only be partially uptodate, in which case the
page flags allone cannot capture the state of the parts of the page.

So allocate all buffers for the page upfront, but leave them unattached so
that they don't pick up any other references and can be freed when we're
done.  If the error path is hit, then zero the new buffers as the regular
buffer path does, then attach the buffers to the page so that it can
actually be written out correctly and be subject to the normal IO error
handling paths.

As an upshot, we save 1K of kernel stack on ia64 or powerpc 64K page
systems.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/buffer.c

index b144fc367b8b9c32d0edda2aaa96fc44448c1e8a..09bb80c479d8b84a02d6defc7f70fef831bb6eb8 100644 (file)
@@ -2274,51 +2274,64 @@ int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
        struct inode *inode = page->mapping->host;
        const unsigned blkbits = inode->i_blkbits;
        const unsigned blocksize = 1 << blkbits;
-       struct buffer_head map_bh;
-       struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
+       struct buffer_head *head, *bh;
        unsigned block_in_page;
-       unsigned block_start;
+       unsigned block_start, block_end;
        sector_t block_in_file;
        char *kaddr;
        int nr_reads = 0;
-       int i;
        int ret = 0;
        int is_mapped_to_disk = 1;
 
+       if (page_has_buffers(page))
+               return block_prepare_write(page, from, to, get_block);
+
        if (PageMappedToDisk(page))
                return 0;
 
+       /*
+        * Allocate buffers so that we can keep track of state, and potentially
+        * attach them to the page if an error occurs. In the common case of
+        * no error, they will just be freed again without ever being attached
+        * to the page (which is all OK, because we're under the page lock).
+        *
+        * Be careful: the buffer linked list is a NULL terminated one, rather
+        * than the circular one we're used to.
+        */
+       head = alloc_page_buffers(page, blocksize, 0);
+       if (!head)
+               return -ENOMEM;
+
        block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
-       map_bh.b_page = page;
 
        /*
         * We loop across all blocks in the page, whether or not they are
         * part of the affected region.  This is so we can discover if the
         * page is fully mapped-to-disk.
         */
-       for (block_start = 0, block_in_page = 0;
+       for (block_start = 0, block_in_page = 0, bh = head;
                  block_start < PAGE_CACHE_SIZE;
-                 block_in_page++, block_start += blocksize) {
-               unsigned block_end = block_start + blocksize;
+                 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
                int create;
 
-               map_bh.b_state = 0;
+               block_end = block_start + blocksize;
+               bh->b_state = 0;
                create = 1;
                if (block_start >= to)
                        create = 0;
-               map_bh.b_size = blocksize;
                ret = get_block(inode, block_in_file + block_in_page,
-                                       &map_bh, create);
+                                       bh, create);
                if (ret)
                        goto failed;
-               if (!buffer_mapped(&map_bh))
+               if (!buffer_mapped(bh))
                        is_mapped_to_disk = 0;
-               if (buffer_new(&map_bh))
-                       unmap_underlying_metadata(map_bh.b_bdev,
-                                                       map_bh.b_blocknr);
-               if (PageUptodate(page))
+               if (buffer_new(bh))
+                       unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+               if (PageUptodate(page)) {
+                       set_buffer_uptodate(bh);
                        continue;
-               if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
+               }
+               if (buffer_new(bh) || !buffer_mapped(bh)) {
                        kaddr = kmap_atomic(page, KM_USER0);
                        if (block_start < from)
                                memset(kaddr+block_start, 0, from-block_start);
@@ -2328,49 +2341,26 @@ int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
                        kunmap_atomic(kaddr, KM_USER0);
                        continue;
                }
-               if (buffer_uptodate(&map_bh))
+               if (buffer_uptodate(bh))
                        continue;       /* reiserfs does this */
                if (block_start < from || block_end > to) {
-                       struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
-
-                       if (!bh) {
-                               ret = -ENOMEM;
-                               goto failed;
-                       }
-                       bh->b_state = map_bh.b_state;
-                       atomic_set(&bh->b_count, 0);
-                       bh->b_this_page = NULL;
-                       bh->b_page = page;
-                       bh->b_blocknr = map_bh.b_blocknr;
-                       bh->b_size = blocksize;
-                       bh->b_data = (char *)(long)block_start;
-                       bh->b_bdev = map_bh.b_bdev;
-                       bh->b_private = NULL;
-                       read_bh[nr_reads++] = bh;
+                       lock_buffer(bh);
+                       bh->b_end_io = end_buffer_read_nobh;
+                       submit_bh(READ, bh);
+                       nr_reads++;
                }
        }
 
        if (nr_reads) {
-               struct buffer_head *bh;
-
                /*
                 * The page is locked, so these buffers are protected from
                 * any VM or truncate activity.  Hence we don't need to care
                 * for the buffer_head refcounts.
                 */
-               for (i = 0; i < nr_reads; i++) {
-                       bh = read_bh[i];
-                       lock_buffer(bh);
-                       bh->b_end_io = end_buffer_read_nobh;
-                       submit_bh(READ, bh);
-               }
-               for (i = 0; i < nr_reads; i++) {
-                       bh = read_bh[i];
+               for (bh = head; bh; bh = bh->b_this_page) {
                        wait_on_buffer(bh);
                        if (!buffer_uptodate(bh))
                                ret = -EIO;
-                       free_buffer_head(bh);
-                       read_bh[i] = NULL;
                }
                if (ret)
                        goto failed;
@@ -2379,21 +2369,54 @@ int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
        if (is_mapped_to_disk)
                SetPageMappedToDisk(page);
 
+       do {
+               bh = head;
+               head = head->b_this_page;
+               free_buffer_head(bh);
+       } while (head);
+
        return 0;
 
 failed:
-       for (i = 0; i < nr_reads; i++) {
-               if (read_bh[i])
-                       free_buffer_head(read_bh[i]);
-       }
-
        /*
-        * Error recovery is pretty slack.  Clear the page and mark it dirty
-        * so we'll later zero out any blocks which _were_ allocated.
+        * Error recovery is a bit difficult. We need to zero out blocks that
+        * were newly allocated, and dirty them to ensure they get written out.
+        * Buffers need to be attached to the page at this point, otherwise
+        * the handling of potential IO errors during writeout would be hard
+        * (could try doing synchronous writeout, but what if that fails too?)
         */
-       zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
-       SetPageUptodate(page);
-       set_page_dirty(page);
+       spin_lock(&page->mapping->private_lock);
+       bh = head;
+       block_start = 0;
+       do {
+               if (PageUptodate(page))
+                       set_buffer_uptodate(bh);
+               if (PageDirty(page))
+                       set_buffer_dirty(bh);
+
+               block_end = block_start+blocksize;
+               if (block_end <= from)
+                       goto next;
+               if (block_start >= to)
+                       goto next;
+
+               if (buffer_new(bh)) {
+                       clear_buffer_new(bh);
+                       if (!buffer_uptodate(bh)) {
+                               zero_user_page(page, block_start, bh->b_size, KM_USER0);
+                               set_buffer_uptodate(bh);
+                       }
+                       mark_buffer_dirty(bh);
+               }
+next:
+               block_start = block_end;
+               if (!bh->b_this_page)
+                       bh->b_this_page = head;
+               bh = bh->b_this_page;
+       } while (bh != head);
+       attach_page_buffers(page, head);
+       spin_unlock(&page->mapping->private_lock);
+
        return ret;
 }
 EXPORT_SYMBOL(nobh_prepare_write);
@@ -2408,6 +2431,9 @@ int nobh_commit_write(struct file *file, struct page *page,
        struct inode *inode = page->mapping->host;
        loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
 
+       if (page_has_buffers(page))
+               return generic_commit_write(file, page, from, to);
+
        SetPageUptodate(page);
        set_page_dirty(page);
        if (pos > inode->i_size) {