On Tue 24-04-18 14:35:36, Theodore Ts'o wrote:
> On Tue, Apr 24, 2018 at 10:27:12AM -0600, Michal Hocko wrote:
> > fs/ext4/xattr.c
> >
> > What to do about this? Well, there are two things. Firstly, it would be
> > really great to double check whether the GFP_NOFS is really needed. I
> > cannot judge that because I am not familiar with the code.
>
> *Most* of the time it's not needed, but there are times when it is.
> We could be more smart about sending down GFP_NOFS only when it is
> needed.
Well, the primary idea is that you do not have to. All you care about is
to use the scope api where it matters + a comment describing the
reclaim recursion context (e.g. this lock will be held in the reclaim
path here and there).
> If we are sending too many GFP_NOFS's allocations such that
> it's causing heartburn, we could fix this. (xattr commands are rare
> enough that I dind't think it was worth it to modulate the GFP flags
> for this particular case, but we could make it be smarter if it would
> help.)
Well, the vmalloc is actually a correctness issue rather than a
heartburn...
> > If the use is really valid then we have a way to do the vmalloc
> > allocation properly. We have memalloc_nofs_{save,restore} scope api. How
> > does that work? You simply call memalloc_nofs_save when the reclaim
> > recursion critical section starts (e.g. when you take a lock which is
> > then used in the reclaim path - e.g. shrinker) and memalloc_nofs_restore
> > when the critical section ends. _All_ allocations within that scope
> > will get GFP_NOFS semantic automagically. If you are not sure about the
> > scope itself then the easiest workaround is to wrap the vmalloc itself
> > with a big fat comment that this should be revisited.
>
> This is something we could do in ext4. It hadn't been high priority,
> because we've been rather overloaded.
Well, ext/jbd already has scopes defined for the transaction context so
anything down that road can be converted to GFP_KERNEL (well, unless the
same code path is shared outside of the transaction context and still
requires a protection). It would be really great to identify other
contexts and slowly move away from the explicit GFP_NOFS. Are you aware
of other contexts?
> As a suggestion, could you take
> documentation about how to convert to the memalloc_nofs_{save,restore}
> scope api (which I think you've written about e-mails at length
> before), and put that into a file in Documentation/core-api?
I can.
> The question I was trying to figure out which triggered the above
> request is how/whether to gradually convert to that scope API. Is it
> safe to add the memalloc_nofs_{save,restore} to code and keep the
> GFP_NOFS flags until we're sure we got it all right, for all of the
> code paths, and then drop the GFP_NOFS?
The first stage is to define and document those scopes. I have provided
a debugging patch [1] in the past that would dump_stack when seeing an
explicit GFP_NOFS from a scope which could help to eliminate existing
users.
[1] http://lkml.kernel.org/r/20170106141845.24362-1-mhocko@xxxxxxxxxx
--
Michal Hocko
SUSE Labs
