On 2021/3/15 下午11:42, David Sterba wrote:

On Mon, Mar 15, 2021 at 01:39:13PM +0800, Qu Wenruo wrote:

quoted

During the fstests run for btrfs subpage read-write support, generic/475
crashes the system with a very high chance.

It turns out the cause is also affecting btrfs subpage read-only mount
so it's worthy a quick fix.

Also the crash call site shows a new rabbit hole of hard coded
PAGE_SHIFT in readahead.

There's still a lot of PAGE_SHIFT use, not all of them are wrong but I
think we'll need to do an audit and categorize the valid uses, otherwise
it'll be a whack-a-mole.

Already did that.

The current valid use case for PAGE_SHIFT are:
- Grab page
   Including:
   * compression
   * raid56
   * relocation
   * buffered write in file.c
   * sb cross page check in volumes.c
   * send
   * zoned
   * sb rw in disk-io.c
   * tree csum in disk-io.c
   * free space cache v1
- Some legacy code still runs in full page mode
   Including:
   * defrag

- Verification code
   That part has way more hardcoded part to be addressed.
   Will be addressed in the final part, along with selftest enhancement.

Although there can be something missing, I believe it shouldn't be that
hard to hit during fstests then.

Thanks,
Qu