Re: [PATCH] mm/page_alloc: restore 0-handling to zone_set_pageset_high_and_batch

[Vlastimil Babka]

On 12/17/25 14:02, Joshua Hahn wrote:
> On Wed, 17 Dec 2025 12:23:58 +0100 Vlastimil Babka <vbabka@xxxxxxx> wrote:
> 
>> On 12/17/25 12:12, Vlastimil Babka wrote:
>> > On 12/17/25 07:05, Joshua Hahn wrote:
>> >> Commit 2783088ef24e ("mm/page_alloc: prevent reporting pcp->batch = 0")
>> >> moved the error handling (0-handling) of zone_batchsize from its
>> >> callers to inside the function. However, the commit left out the error
>> >> handling for the NOMMU case, leading to deadlocks on NOMMU systems.
>> >> 
>> >> Since in the NOMMU case the reported-to-user batchsize should still be 0,
>> > 
>> > Should it? The value is effectively set to 1 despite what zone_batchsize()
>> > returns, because of that adjustment this patch reinstates. Also does anyone
>> > care, really?
>> > 
>> >> we would only like the error handling to exist in the callsites that
>> >> set the internal value for the zone (i.e. zone_set_pageset_high_and_batch).
>> >> 
>> >> Restore max(1, zone_batchsize(zone)) to the callsite to prevent errors
>> >> on NOMMU systems.
>> > 
>> > I would rather make zone_batchsize() for !CONFIG_MMU return 1 instead of 0.
>> 
>> Ah looks like you considered it too, initially:
>> https://lore.kernel.org/all/20251211225947.822866-1-joshua.hahnjy@xxxxxxxxx/
>> 
>> It makes more sense to me than doing effectively two fixups in the MMU case.
> 
> Hi Vlastimil,
> 
> Thank you for your review as always.
> 
> Yes, I had also considered returning 1 for the !MMU case, since I think it
> would make it a lot simpler as well (It would also make my original patch
> function as intended).
> 
> However, I was unsure if changing this user-facing behavior for one line of
> simplification would be worth it. I am not a NOMMU user, so I have very
> little experience here, but I imagine that there is someone out there who
> looks at zone_batchsize() returning 0 for NOMMU and interpreting it as
> "there is no batching" as opposed to "there is batching, and it processes
> 1 page at a time" (which, actually isn't even true anyways because of the
> bitshift). Maybe an option is to just make batchsize not visible
> in the NOMMU case in addition to always returning 1 to avoid confusion.
> 
> Anyways, back to your original question of "does anyone care". . .
> 
> I am not sure : -)

It's a pr_debug(), it's not even being printed by default, nothing can
possibly break by changing 0 to 1 there. So I really wouldn't overthink this...

> For me, both solutions work, and in fact I prefer the original solution of
> always reurning 1 for !NOMMU. Maybe some NOMMU users like Daniel and Guenter
> can comment on whether this change really matters?

I would be surprised if they were aware of that pr_debug() in the first place :)

> Thank you again for your review and follow-up. I hope you have a great day!
> Joshua