mm: make sendfile(2) killable
authorJan Kara <jack@suse.com>
Thu, 22 Oct 2015 20:32:21 +0000 (13:32 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 23 Oct 2015 08:55:10 +0000 (17:55 +0900)
Currently a simple program below issues a sendfile(2) system call which
takes about 62 days to complete in my test KVM instance.

        int fd;
        off_t off = 0;

        fd = open("file", O_RDWR | O_TRUNC | O_SYNC | O_CREAT, 0644);
        ftruncate(fd, 2);
        lseek(fd, 0, SEEK_END);
        sendfile(fd, fd, &off, 0xfffffff);

Now you should not ask kernel to do a stupid stuff like copying 256MB in
2-byte chunks and call fsync(2) after each chunk but if you do, sysadmin
should have a way to stop you.

We actually do have a check for fatal_signal_pending() in
generic_perform_write() which triggers in this path however because we
always succeed in writing something before the check is done, we return
value > 0 from generic_perform_write() and thus the information about
signal gets lost.

Fix the problem by doing the signal check before writing anything.  That
way generic_perform_write() returns -EINTR, the error gets propagated up
and the sendfile loop terminates early.

Signed-off-by: Jan Kara <jack@suse.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/filemap.c

index 1cc5467..327910c 100644 (file)
@@ -2488,6 +2488,11 @@ again:
                        break;
                }
 
+               if (fatal_signal_pending(current)) {
+                       status = -EINTR;
+                       break;
+               }
+
                status = a_ops->write_begin(file, mapping, pos, bytes, flags,
                                                &page, &fsdata);
                if (unlikely(status < 0))
@@ -2525,10 +2530,6 @@ again:
                written += copied;
 
                balance_dirty_pages_ratelimited(mapping);
-               if (fatal_signal_pending(current)) {
-                       status = -EINTR;
-                       break;
-               }
        } while (iov_iter_count(i));
 
        return written ? written : status;