Re: [PATCH v6 0/6] [PATCH v6 0/6] Add reclaim to the dmem cgroup controller

[Thadeu Lima de Souza Cascardo]

On Thu, Jun 25, 2026 at 11:19:19AM +0200, Michal Koutný wrote:
> On Mon, Jun 22, 2026 at 12:26:54PM -0300, Thadeu Lima de Souza Cascardo <cascardo@xxxxxxxxxx> wrote:
> > As far as I understood the patchset, it doesn't fail the write if it fails
> > to reclaim. It sets the new max, then, if the write is blocking, starts
> > reclaim and eventually returns after multiple attempts. But it still
> > returns success.
> > 
> > So I believe this is behaving as you would expect.
> 
> I was alarmed by the EBUSY mention similarly to Tejun but then I
> couldn't find it in pre-patch (840ef6c78e6a2) nor in patched (v5) code.
> Please make sure the EBUSY return behavior is not introduced
> (essentially match memory.max behavior) and that the accompanying
> message refers up to date code ;-)
> 
> Michal

I think this is a reference to the fact that right now writing to dmem.max
calls page_counter_set_max, which will return -EBUSY and fail to change the
max value. It is just that we are not returning that error today to
userspace (a fix I have submitted back in April).

So, this patchset by Thomas is setting max, and, then, if it is a blocking
write, tries to evict on a best-effort basis, and returns success after a
few attempts, still setting max to the written value.

Cascardo.