Re: [PATCH 01/24] xfs: cow unwritten conversion uses uninitialized dfops

[Brian Foster <bfoster@redhat.com>]

On Tue, Jul 03, 2018 at 08:21:53AM -0700, Darrick J. Wong wrote:
> On Tue, Jul 03, 2018 at 11:10:26AM -0400, Brian Foster wrote:
> > On Tue, Jul 03, 2018 at 07:59:38AM -0700, Darrick J. Wong wrote:
> > > On Mon, Jul 02, 2018 at 01:32:41PM -0400, Brian Foster wrote:
> > > > On Mon, Jul 02, 2018 at 06:43:04AM -0700, Christoph Hellwig wrote:
> > > > > On Thu, Jun 28, 2018 at 12:36:13PM -0400, Brian Foster wrote:
> > > > > > A couple COW fork unwritten extent conversion helpers pass an
> > > > > > uninitialized dfops pointer to xfs_bmapi_write(). This does not
> > > > > > cause problems because conversion does not use a transaction or the
> > > > > > dfops structure for the COW fork.  Drop the uninitialized usage of
> > > > > > dfops in these codepaths and pass NULL along to xfs_bmapi_write()
> > > > > > instead.
> > > > > 
> > > > > Looks good.
> > > > > 
> > > > > Is this something we should maybe queue up for 4.18?
> > > > > 
> > > > 
> > > > That might make sense because of all the refactoring, but otherwise I
> > > > don't have a strong opinion. Let's see what Darrick wants to do...
> > > 
> > > AFAICT this only eliminates the passing around of an unus{ed,able} dfops
> > > parameter, right?  We're not fixing a regression or some other breakage,
> > > just eliminating cpu cycle waste, so I think this can soak (along with
> > > everything else) until 4.19.
> > > 
> > 
> > Works for me. This does have the side effect of enabling the deferred
> > AGFL block free behavior wherever dfops is used, but that is a not a
> > critical change/fix. The problem that inspired the behavior in the first
> > place was resolved by the more targeted changes in the first series.
> > The goals of this (and the series or two to follow) are primarily follow
> > up refactoring and to provide more consistent behavior fs-wide.
> 
> Follow-up question, then: are there situations where we defered agfl
> freeing is /not/ desirable?  I couldn't think of any, but my head (and
> shop vac) are full of sawdust. :)
> 

I don't think so. My presumption since the original series is that this
is essentially an analogus situation to freeing file blocks. We defer it
unconditionally to provide some semblence of consistent/predictable
transaction reservation consumption. I haven't audited to the point of
seeing if we need to add dfops in certain paths purely for deferred AGFL
frees (i.e., adding dfops to paths that don't already have a dfops), but
the bit I mentioned in my last mail would effectively cover that last
case by embedding ->t_dfops in the transaction. With that, any
transaction is capable of deferring operations to be finished
automatically at commit time (and the boilerplate trans_alloc() ->
defer_init() -> defer_finish() -> trans_commit() sequence can be
factored away). One obvious tradeoff is that the transaction becomes
bigger, but I'm not sure how much that matters given that we have a
dfops in most cases anyways and we're using it more progressively.

BTW, I don't think any of this precludes us from not deferring AGFL
frees in a particular case if that becomes necessary for some reason in
the future. We'd just need to decide out how to communicate that to the
allocator (alloc_arg flag, tx flag, etc.).

Brian

> --D
> 
> > Brian
> > 
> > > --D
> > > 
> > > > Brian
> > > > 
> > > > > Reviewed-by: Christoph Hellwig <hch@lst.de>
> > > > > --
> > > > > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> > > > > the body of a message to majordomo@vger.kernel.org
> > > > > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > > > --
> > > > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> > > > the body of a message to majordomo@vger.kernel.org
> > > > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > > --
> > > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> > > the body of a message to majordomo@vger.kernel.org
> > > More majordomo info at  http://vger.kernel.org/majordomo-info.html