]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/commitdiff
nfs: don't ignore return value from nfs_pageio_add_request
authorFred Isaman <iisaman@citi.umich.edu>
Wed, 19 Mar 2008 15:24:39 +0000 (11:24 -0400)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Wed, 19 Mar 2008 21:59:02 +0000 (17:59 -0400)
Ignoring the return value from nfs_pageio_add_request can cause deadlocks.

In read path:
  call nfs_pageio_add_request from readpage_async_filler
  assume at this point that there are requests already in desc, that
    can't be merged with the current request.
  so nfs_pageio_doio is fired up to clear out desc.
  assume something goes wrong in setting up the io, so desc->pg_error is set.
  This causes nfs_pageio_add_request to return 0, *WITHOUT* adding the original
    request.
  BUT, since return code is ignored, readpage_async_filler assumes it has
    been added, and does nothing further, leaving page locked.
  do_generic_mapping_read will eventually call lock_page, resulting in deadlock

In write path:
  page is marked dirty by generic_perform_write
  nfs_writepages is called
  call nfs_pageio_add_request from nfs_page_async_flush
  assume at this point that there are requests already in desc, that
    can't be merged with the current request.
  so nfs_pageio_doio is fired up to clear out desc.
  assume something goes wrong in setting up the io, so desc->pg_error is set.
  This causes nfs_page_async_flush to return 0, *WITHOUT* adding the original
    request, yet marking the request as locked (PG_BUSY) and in writeback,
    clearing dirty marks.
  The next time a write is done to the page, deadlock will result as
    nfs_write_end calls nfs_update_request

Signed-off-by: Fred Isaman <iisaman@citi.umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
fs/nfs/read.c
fs/nfs/write.c

index 3d7d9631e1254f7ad898863e66519740af1e34e9..5a70be589bbe2025eee4c6fc64aaba1c0b5958c3 100644 (file)
@@ -533,7 +533,10 @@ readpage_async_filler(void *data, struct page *page)
 
        if (len < PAGE_CACHE_SIZE)
                zero_user_segment(page, len, PAGE_CACHE_SIZE);
-       nfs_pageio_add_request(desc->pgio, new);
+       if (!nfs_pageio_add_request(desc->pgio, new)) {
+               error = desc->pgio->pg_error;
+               goto out_unlock;
+       }
        return 0;
 out_error:
        error = PTR_ERR(new);
index 80c61fdb2720c4b05f86ff7c590cd0af566e2b8d..bed63416a55b3c51ab539666c4f3f5f5ca8321b0 100644 (file)
@@ -39,6 +39,7 @@ static struct nfs_page * nfs_update_request(struct nfs_open_context*,
                                            unsigned int, unsigned int);
 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
                                  struct inode *inode, int ioflags);
+static void nfs_redirty_request(struct nfs_page *req);
 static const struct rpc_call_ops nfs_write_partial_ops;
 static const struct rpc_call_ops nfs_write_full_ops;
 static const struct rpc_call_ops nfs_commit_ops;
@@ -288,7 +289,12 @@ static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
                BUG();
        }
        spin_unlock(&inode->i_lock);
-       nfs_pageio_add_request(pgio, req);
+       if (!nfs_pageio_add_request(pgio, req)) {
+               nfs_redirty_request(req);
+               nfs_end_page_writeback(page);
+               nfs_clear_page_tag_locked(req);
+               return pgio->pg_error;
+       }
        return 0;
 }