Re: [PATCH 0/3] Btrfs: stop abusing current->journal_info for direct I/O

[Omar Sandoval <osandov@osandov.com>]

On Fri, May 11, 2018 at 12:53:36PM +0300, Nikolay Borisov wrote:
> 
> 
> On 11.05.2018 09:30, Omar Sandoval wrote:
> > From: Omar Sandoval <osandov@fb.com>
> > 
> > Hi, everyone,
> > 
> > Btrfs currently abuses current->journal_info in btrfs_direct_IO() in
> > order to pass around some state to get_block() and submit_io(). This
> > hack is ugly and unnecessary because the data we pass around is only
> > used in one call frame. Robbie Ko also pointed out [1] that it could
> > potentially cause a crash if we happen to end up in start_transaction()
> > (e.g., from memory reclaim calling into btrfs_evict_inode(), which can
> > start a transaction). I'm not convinced that Robbie's case can happen in
> > practice since we are using GFP_NOFS for allocations during direct I/O,
> > but either way it's fragile and nasty.
> 
> When I worked initially on btrfs-over-swap I managed to hit a case where
> ext4 stacked on top of btrfs would crash since btrfs will overwrite
> journal_info which was set by ext4. So this change is indeed welcome :)

Yup, that's what I originally made these patches for. Although my latest
idea for swap is to do something along the lines of Darrick's
iomap_swap_activate(): https://patchwork.kernel.org/patch/10376435/,
I'll be getting back to that soon.