On 10/1/21 12:49 AM, Hugh Dickins wrote:
On Thu, 30 Sep 2021, Matthew Wilcox wrote:
On Wed, Sep 29, 2021 at 10:24:44PM -0700, Hugh Dickins wrote:
Aside from the above page->index mischeck in find_lock_entries(),
I now think this bug needs nothing more than simply removing the
VM_BUG_ON_PAGE(PageTail(page), page) from truncate_inode_page().
I don't think that's right. This bug was also observed when calling
truncate(), so there's clearly a situation where two concurrent calls
to truncate_pagecache() leaves a THP in the cache.
I assume you're thinking of one of the fuzzer blkdev ones:
https://lore.kernel.org/linux-mm/CACkBjsbtF_peC7N_4mRfHML_BeiPe+O9DahTfr84puSG_J9rcQ@xxxxxxxxxxxxxx/
or
https://lore.kernel.org/lkml/CACkBjsYwLYLRmX8GpsDpMthagWOjWWrNxqY6ZLNQVr6yx+f5vA@xxxxxxxxxxxxxx/
I haven't started on those ones yet: yes, I imagine one or both of those
will need a further fix (S_ISREG() check somewhere if we're lucky; but
could well be nastier); but for the bug in this thread, I expect
removing the VM_BUG_ON_PAGE(PageTail) to be enough.
Thanks for your advices!
I'm so sorry that delay the test due to my recent vacation. I plan to
start further test tomorrow. I think removing the
VM_BUG_ON_PAGE(PageTail) is a good idea, but also think using
filemap_invalidate_lock is safer and necessary. And of course, this is
just my own view! So, now, I tend to use filemap_invalidate_lock and
either check page mapping again or remove VM_BUG_ON_PAGE(PageTail).
Anyway, I will run more tests and then give feedback.
Thanks!
If you're thinking of something else, please send a link if you can - thanks.
Hugh
