On Sat, Sep 09, 2023 at 10:03:04PM -0400, Kent Overstreet wrote:
> On Wed, Sep 06, 2023 at 03:07:58PM -0400, Brian Foster wrote:
> > On Tue, Sep 05, 2023 at 08:59:02AM -0400, Brian Foster wrote:
> > > On Sun, Sep 03, 2023 at 10:29:28PM -0400, Kent Overstreet wrote:
> > > > On Sun, Sep 03, 2023 at 05:18:12PM -0400, Kent Overstreet wrote:
> > > > > On Thu, Aug 31, 2023 at 07:07:34AM -0400, Brian Foster wrote:
> > > > > > bcachefs freeze testing via fstests generic/390 occasionally
> > > > > > reproduces the following BUG from bch2_fs_read_only():
> > > > > > 
> > > > > >   BUG_ON(atomic_long_read(&c->btree_key_cache.nr_dirty));
> > > > >
> > > > > Your fix makes sense, but I wonder if it might be simpler and better 
> > > > > to
> > > > > fix this in bch2_journal_reclaim_fast().
> > ...
> > > 
> > > FWIW, one thing I disliked about the original patch was the need to
> > > mostly duplicate the buf put helper due to the locking context quirk. I
> > > was trying to avoid that, but I ran out of ideas before I wanted to move
> > > on actually fixing the bug. My preference would be to address the
> > > reference counting issue as is (to preserve design simplicity), and then
> > > maybe think a bit harder about cleaning up the res put implementation if
> > > the primary concern is that we feel like this starts to make things a
> > > bit convoluted..
> > > 
> > 
> > Another thought that comes to mind is to perhaps allow the journal_res
> > to hold a reference to the pin fifo for the associated seq.. The idea
> > would be we could continue to hold a reference during the open/close
> > journal buffer lifecycle, but a res of the same seq would acquire an
> > additional reference as well to keep the tail from popping before a
> > transaction can actually set a pin (i.e. essentially parallel to the
> > buffer reference).
> 
> Yeah that would probably be cleanest - but it would be much too heavy.
> 

I'm curious how so? Additional atomics? 

If so, we could also consider something that allows a pin list ref on
the reservation to simply transfer to a pin set via the transaction.
That might not add any new overhead over the current code, but would
require some plumbing.

> > In a sense the behavior in this patch is kind of an optimization of that
> > approach, but I think implementing it directly would eliminate the need
> > to mostly duplicate the put helper, perhaps making things a bit more
> > natural. We could still fall into bch2_journal_reclaim_fast() from
> > res_put(), but now only for cases where the reservation outlives the
> > active journal buffer (due to a journal flush or whatever). That also
> > still addresses the race by properly using the reference count. I'd
> > probably have to try it to be sure I'm not missing something, but...
> > thoughts?
> 
> We can do your approach if that's what you feel is going to be cleanest,
> but if we go that way here's my two main concerns:
> 

Ok..

FWIW, it's not so much that I think it's the cleanest (I agree that the
variant discussed above probably ends up cleaner code-wise), but rather
that it provides sufficient confidence we won't just have to revisit the
same underlying problem the next time some bit of code assumes that an
outstanding reservation guarantees a usable pin list.

>  - we can't put more slowpath code into the inline fastpath, remember
>    this is all inlined into the main __bch2_trans_commit() path
> 

Sure. I missed that closure_call() was itself inlined. I think that is
what initially confused me wrt to the feedback on patch 2. I'll try to
keep that helper external and perhaps just rename it.

>  - your approach (the optimized version) means there's a new state
>    transition where we do work, i.e. when journal_state_count() hits 0.
> 

Indeed.

>    So we have to make sure that that refcount isn't hitting 0 multiple
>    times. The appropriate assertion already exists in
>    journal_res_get_fast() - for a refcount to hit 0 multiple times it
>    must leave 0 while in use, and we check for that.
> 
>    Let's just add a comment by that EBUG_ON(!journal_state_count(new,
>    new.idx)) line indicating why it's now important.
> 

Sounds reasonable. I'll try to come up with something. Thanks.

Brian

Reply via email to