Re: [vfs] 8bb3c61baf: vm-scalability.median -23.7% regression

[Hugh Dickins]

On Mon, 9 Sep 2019, Al Viro wrote:
> 
> Anyway, see vfs.git#uncertain.shmem for what I've got with those folded in.
> Do you see any problems with that one?  That's the last 5 commits in there...

It's mostly fine, I've no problem with going your way instead of what
we had in mmotm; but I have seen some problems with it, and had been
intending to send you a fixup patch tonight (shmem_reconfigure() missing
unlock on error is the main problem, but there are other fixes needed).

But I'm growing tired. I've a feeling my "swap" of the mpols, instead
of immediate mpol_put(), was necessary to protect against a race with
shmem_get_sbmpol(), but I'm not clear-headed enough to trust myself on
that now.  And I've a mystery to solve, that shmem_reconfigure() gets
stuck into showing the wrong error message.

Tomorrow....

Oh, and my first attempt to build and boot that series over 5.3-rc5
wouldn't boot. Luckily there was a tell-tale "i915" in the stacktrace,
which reminded me of the drivers/gpu/drm/i915/gem/i915_gemfs.c fix
we discussed earlier in the cycle.  That is of course in linux-next
by now, but I wonder if your branch ought to contain a duplicate of
that fix, so that people with i915 doing bisections on 5.4-rc do not
fall into an unbootable hole between vfs and gpu merges.

Hugh