On Thu, Apr 30, 2020 at 11:43:20PM +0100, Filipe Manana wrote:
> On Thu, Apr 30, 2020 at 11:23 PM Dennis Zhou <dennis@xxxxxxxxxx> wrote:
> >
> > On Thu, Apr 30, 2020 at 02:40:18PM -0700, Andrew Morton wrote:
> > > On Thu, 30 Apr 2020 17:43:56 +0100 fdmanana@xxxxxxxxxx wrote:
> > >
> > > > From: Filipe Manana <fdmanana@xxxxxxxx>
> > > >
> > > > Since 5.7-rc1, on btrfs we have a percpu counter initialization for which
> > > > we always pass a GFP_KERNEL gfp_t argument (this happens since commit
> > > > 2992df73268f78 ("btrfs: Implement DREW lock")). That is safe in some
> > > > contextes but not on others where allowing fs reclaim could lead to a
> > > > deadlock because we are either holding some btrfs lock needed for a
> > > > transaction commit or holding a btrfs transaction handle open. Because
> > > > of that we surround the call to the function that initializes the percpu
> > > > counter with a NOFS context using memalloc_nofs_save() (this is done at
> > > > btrfs_init_fs_root()).
> > > >
> > > > However it turns out that this is not enough to prevent a possible
> > > > deadlock because percpu_alloc() determines if it is in an atomic context
> > > > by looking exclusively at the gfp flags passed to it (GFP_KERNEL in this
> > > > case) and it is not aware that a NOFS context is set. Because it thinks
> > > > it is in a non atomic context it locks the pcpu_alloc_mutex, which can
> > > > result in a btrfs deadlock when pcpu_balance_workfn() is running, has
> > > > acquired that mutex and is waiting for reclaim, while the btrfs task that
> > > > called percpu_counter_init() (and therefore percpu_alloc()) is holding
> > > > either the btrfs commit_root semaphore or a transaction handle (done at
> > > > fs/btrfs/backref.c:iterate_extent_inodes()), which prevents reclaim from
> > > > finishing as an attempt to commit the current btrfs transaction will
> > > > deadlock.
> > > >
> > >
> > > Patch looks good and seems sensible, thanks.
> > >
> >
> > Acked-by: Dennis Zhou <dennis@xxxxxxxxxx>
> >
> > > But why did btrfs use memalloc_nofs_save()/restore() rather than
> > > s/GFP_KERNEL/GFP_NOFS/?
> >
> > I would also like to know.
>
> For 2 reasons:
>
> 1) It's the preferred way to do it since
> memalloc_nofs_save()/restore() was added (according to
> Documentation/core-api/gfp_mask-from-fs-io.rst);
>
Thanks. I didn't realize it completely superceded GFP_NOFS.
> 2) According to Documentation/core-api/gfp_mask-from-fs-io.rst,
> passing GFP_NOFS to __vmalloc() doesn't work, so one has to use the
> memalloc_nofs_save()/restore() API for that. And pcpu_alloc() calls
> helpers that end up calling __vmalloc() (through pcpu_mem_zalloc()).
>
> And that's it.
>
I'm starting to remember a bit more. I guess it's not great how
percpu manages GFP_ATOMIC as !GFP_KERNEL for the possible vmalloc()
calls. At the time I believe the whitelist was the only way to deal with
the recursive case. If I get a chance I'll look at the flags again and
see if we can't do something better/ more aligned today.
>
> >
> > Thanks,
> > Dennis
