On Wed, Sep 29, 2021 at 5:02 PM Matthew Wilcox [off-list ref] wrote:

On Wed, Sep 29, 2021 at 04:41:48PM -0700, Song Liu wrote:

quoted

The issue is NOT caused by concurrent khugepaged:collapse_file() and
truncate_pagecache(inode, 0). With some printks, we can see a clear
time gap (>2 second )  between collapse_file() finishes, and
truncate_pagecache() (which crashes soon). Therefore, my earlier
suggestion that adds deny_write_access() to collapse_file() does NOT
work.

The crash is actually caused by concurrent truncate_pagecache(inode, 0).
If I change the number of write thread in stress_madvise_dso.c to one,
(IOW, one thread_read and one thread_write), I cannot reproduce the
crash anymore.

I think this means we cannot fix this issue in collapse_file(), because it
finishes long before the crash.

Ah!  So are we missing one or more of these locks:

        inode_lock(inode);
        filemap_invalidate_lock(mapping);

in the open path?

The following fixes the crash in my test. But I am not sure whether this is the
best fix.

Rongwei, could you please run more tests on it?

Thanks,
Song

diff --git i/fs/open.c w/fs/open.c
index daa324606a41f..d13c4668b2e53 100644
--- i/fs/open.c
+++ w/fs/open.c

@@ -856,8 +856,11 @@ static int do_dentry_open(struct file *f,
                 * of THPs into the page cache will fail.
                 */
                smp_mb();
-               if (filemap_nr_thps(inode->i_mapping))
+               if (filemap_nr_thps(inode->i_mapping)) {
+                       filemap_invalidate_lock(inode->i_mapping);
                        truncate_pagecache(inode, 0);
+                       filemap_invalidate_unlock(inode->i_mapping);
+               }
        }

        return 0;