Hi there,
ext4_aops and ext4_da_aops both have a migratepage callback, whereas
ext4_journalled_aops lacks such a callback. Why is this so? I’m asking
this due to the following: when a page containing EXT4 journal buffer
heads ends up being migrated, fallback_migrate_page() is used, which
eventually calls try_to_free_buffers(), which will call drop_buffers().
Drop buffers() can fail for a page if that page is on the LRU list (see
https://elixir.bootlin.com/linux/v5.8.14/source/fs/buffer.c#L3225).
Now, if buffer_migrate_page() was supplied as the migratepage callback
for the journaled aops, this wouldn’t be problem since we ignore the LRU
lists altogether.
Resolving this issue will benefit CMA allocations, which might have to
migrate movable pages that were allocated from a CMA region (the
assumption is that these pages can be migrated once the memory backing
these pages is needed).
Thanks,
Chris.
--
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum,
a Linux Foundation Collaborative Project
tOn 2020-10-08 23:42, Chris Goldsworthy wrote:
> Hi there,
>
> ext4_aops and ext4_da_aops both have a migratepage callback, whereas
> ext4_journalled_aops lacks such a callback. Why is this so? I’m
> asking this due to the following: when a page containing EXT4 journal
> buffer heads ends up being migrated, fallback_migrate_page() is used,
> which eventually calls try_to_free_buffers(), which will call
> drop_buffers(). Drop buffers() can fail for a page if that page is on
> the LRU list (see
> https://elixir.bootlin.com/linux/v5.8.14/source/fs/buffer.c#L3225).
> Now, if buffer_migrate_page() was supplied as the migratepage callback
> for the journaled aops, this wouldn’t be problem since we ignore the
> LRU lists altogether.
>
> Resolving this issue will benefit CMA allocations, which might have to
> migrate movable pages that were allocated from a CMA region (the
> assumption is that these pages can be migrated once the memory backing
> these pages is needed).
>
> Thanks,
>
> Chris.
An alternative solution to to placing in a migratepage callback is to
actually remove the buffer heads from an LRU list in drop_buffers() - we
tried upstreaming such a patch some time ago:
https://lore.kernel.org/patchwork/patch/325406/ . This would solve the
problem, and would avoid having to get a migratepage callback for any
other files systems that have aops without the callback. This didn't
get off the ground however, I'm picking up the line of questioning left
in the previous thread, to determine why EXT4 doesn't have a migratepage
callback.
Thanks,
Chris.
--
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum,
a Linux Foundation Collaborative Project