mm: direct IO starvation improvement
authorNick Piggin <npiggin@suse.de>
Tue, 6 Jan 2009 22:40:22 +0000 (14:40 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 6 Jan 2009 23:59:09 +0000 (15:59 -0800)
Direct IO can invalidate and sync a lot of pagecache pages in the mapping.
 A 4K direct IO will actually try to sync and/or invalidate the pagecache
of the entire file, for example (which might be many GB or TB large).

Improve this by doing range syncs.  Also, memory no longer has to be
unmapped to catch the dirty bits for syncing, as dirty bits would remain
coherent due to dirty mmap accounting.

This fixes the immediate DM deadlocks when doing direct IO reads to block
device with a mounted filesystem, if only by papering over the problem
somewhat rather than addressing the fsync starvation cases.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/filemap.c

index 9c5e6235cc743e37787c124e9f26a9e287ac8c06..f3555fb806d39f17fa663de470a3a08a765133a3 100644 (file)
@@ -1317,7 +1317,8 @@ generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
                        goto out; /* skip atime */
                size = i_size_read(inode);
                if (pos < size) {
-                       retval = filemap_write_and_wait(mapping);
+                       retval = filemap_write_and_wait_range(mapping, pos,
+                                       pos + iov_length(iov, nr_segs) - 1);
                        if (!retval) {
                                retval = mapping->a_ops->direct_IO(READ, iocb,
                                                        iov, pos, nr_segs);
@@ -2059,18 +2060,10 @@ generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
        if (count != ocount)
                *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
 
-       /*
-        * Unmap all mmappings of the file up-front.
-        *
-        * This will cause any pte dirty bits to be propagated into the
-        * pageframes for the subsequent filemap_write_and_wait().
-        */
        write_len = iov_length(iov, *nr_segs);
        end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
-       if (mapping_mapped(mapping))
-               unmap_mapping_range(mapping, pos, write_len, 0);
 
-       written = filemap_write_and_wait(mapping);
+       written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
        if (written)
                goto out;
 
@@ -2290,7 +2283,8 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
         * the file data here, to try to honour O_DIRECT expectations.
         */
        if (unlikely(file->f_flags & O_DIRECT) && written)
-               status = filemap_write_and_wait(mapping);
+               status = filemap_write_and_wait_range(mapping,
+                                       pos, pos + written - 1);
 
        return written ? written : status;
 }