On 11/17/22 20:23, Dennis Zhou wrote:
> On Wed, Nov 16, 2022 at 07:32:03PM +0800, Baoquan He wrote:
>> On 11/15/22 at 12:00pm, Dennis Zhou wrote:
>> > On Tue, Nov 15, 2022 at 05:08:52PM +0800, Baoquan He wrote:
>> > > Hi Dennis,
>> > >
>> > > On 11/14/22 at 08:13pm, Dennis Zhou wrote:
>> > > > Hi Vlastimil & Baoquan,
>> > > >
>> > > > On Mon, Nov 14, 2022 at 06:58:13PM +0100, Vlastimil Babka wrote:
>> > > > > On 11/14/22 08:44, Baoquan He wrote:
>> > > > > > Hi,
>> > > > > >
>> > > > > > I reproduced the build failure according to lkp report and made a patch
>> > > > > > as below to fix it.
>> > > > > >
>> > > > > > From dae7dd9705015ce36db757e88c78802584f949b1 Mon Sep 17 00:00:00 2001
>> > > > > > From: Baoquan He <bhe@xxxxxxxxxx>
>> > > > > > Date: Sun, 13 Nov 2022 18:08:27 +0800
>> > > > > > Subject: [PATCH] percpu: adjust the value of PERCPU_DYNAMIC_EARLY_SIZE
>> > > > > > Content-type: text/plain
>> > > > > >
>> > > > > > LKP reported a build failure as below on the patch "mm/slub, percpu:
>> > > > > > correct the calculation of early percpu allocation size"
>> > > > >
>> > > > > Since I have that patch in slab.git exposed to -next, should I take this fix
>> > > > > too, to make things simpler? Dennis?
>> > > > >
>> > > >
>> > > > I don't have any problems with you running a fix, but I'm not quite sure
>> > > > this is the right fix. Though this might cause a trivial merge conflict
>> > > > with: d667c94962c1 ("mm/percpu: remove unused PERCPU_DYNAMIC_EARLY_SLOTS")
>> > > > in my percpu#for-6.2 branch.
>> > > >
>> > > > If I'm understanding this correctly, slub requires additional percpu
>> > > > memory due to the use of 64k pages. By increasing
>> > > > PERCPU_DYNAMIC_EARLY_SIZE, we solve the problem for 64k page users, but
>> > > > require a few unnecessary pages that can bloat the size of subsequent
>> > > > percpu chunks. Though, I'm not sure if that's an issue today for
>> > > > embedded devices.
>> > >
>> > > Thanks for looking into this.
>> > >
>> > > I guess you are talking about PERCPU_DYNAMIC_EARLY_SIZE will impact the
>> > > first dynamic chunk size of page first chunk, because the embed first
>> > > chunk will take PERCPU_DYNAMIC_RESERVE. And the impact is done in below
>> > > max() invacation.
>> > >
>> > > static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
>> > > size_t reserved_size, size_t dyn_size,
>> > > size_t atom_size,
>> > > pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
>> > > {
>> > > ......
>> > > /* calculate size_sum and ensure dyn_size is enough for early alloc */
>> > > size_sum = PFN_ALIGN(static_size + reserved_size +
>> > > max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
>> > > ......
>> > > }
>> > >
>> > > >
>> > > > I think adding parity to PERCPU_DYNAMIC_EARLY_SIZE with
>> > > > PERCPU_DYNAMIC_RESERVE is defined by BITS_PER_LONG is a safer option
>> > > > here. A small TODO item would be to make PERCPU_DYNAMIC_RESERVE be a +
>> > > > value instead of a max() with PERCPU_DYNAMIC_EARLY_SIZE.
>> > >
>> > > Hmm, the below change may not take power arch into account. Please
>> > > check arch/powerpc/include/asm/page.h, seems the 32bit ppc could have
>> > > 256K pages too. Adding PERCPU_DYNAMIC_EARLY_SIZE to 20K may cost extra
>> > > memory during boot. But th left space of 1st dynamic chunk will join
>> > > the later percpu dynamic allocation, it's not wasted, right?
>> > >
>> > > Not sure if I got your point.
>> > >
>> > >
>> >
>> > Ah, I'm not familiar with all the PAGE_SIZE and word length
>> > combinations.
>> >
>> > The first chunk is smaller in the embedded case with the assumption that
>> > static percpu variables are highly accessed along with the limited
>> > initial allocations. While adding an additional 8KB is not the biggest
>> > deal to the first chunk, this can cause the unit_size for subsequent
>> > chunks to be larger. For example, x86 unit size jumps in powers of 2 due
>> > to alignment and packing against an allocation size of 2MB. So if we're
>> > at say 60KB for the first chunk, subsequent chunks could be 64KB. But
>> > adding 8KB, we'd go from 60KB -> 68KB and a chunk size of 64KB -> 128KB.
>>
>> I could have misunderstanding about the first chunk usage and percpu
>> code. Below is my personal uderstanding about the 1st chunk size and
>> how PERCPU_DYNAMIC_EARLY_SIZE could impact it, please help point out
>> if I am wrong.
>>
>> ~~~
>> Abstract the definition of them here for reference.
>> /*
>> * Percpu allocator can serve percpu allocations before slab is
>> * initialized which allows slab to depend on the percpu allocator.
>> * The following parameter decide how much resource to preallocate
>> * for this. Keep PERCPU_DYNAMIC_RESERVE equal to or larger than
>> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>> * PERCPU_DYNAMIC_EARLY_SIZE.
>> ~~~~~~~~~~~~~~~~~~~~~
>> */
>> #define PERCPU_DYNAMIC_EARLY_SIZE (12 << 10)
>> ......
>> #if BITS_PER_LONG > 32
>> #define PERCPU_DYNAMIC_RESERVE (28 << 10)
>> #else
>> #define PERCPU_DYNAMIC_RESERVE (20 << 10)
>> #endif
>>
>> From above definition, we can see that no matter how big
>> PERCPU_DYNAMIC_RESERVE is , it's >= PERCPU_DYNAMIC_EARLY_SIZE as the
>> code comment says. So the max() in pcpu_build_alloc_info() won't impact
>> the embeded 1st chunk at all.
>>
>> So, PERCPU_DYNAMIC_EARLY_SIZE can only impact the page 1st chunk case,
>> namely when calling pcpu_page_first_chunk() to do that. In
>> pcpu_page_first_chunk(), we don't provide dyn_size, so with the help of
>> max(), it will get final dyn_size as PERCPU_DYNAMIC_EARLY_SIZE. This is
>> the only place where PERCPU_DYNAMIC_EARLY_SIZE takes effect on percpu.
>> However, the atom size of page 1st chunk is PAGE_SIZE, it doesn't have
>> the issue of possible bloating unit_size by the atom size, e.g 2M on
>> x86_64. Since pcpu_page_first_chunk() is the fallback of
>> pcpu_embed_first_chunk(), if we decide to provide PERCPU_DYNAMIC_RESERVE
>> as the current value, why we grudge setting it as the smaller value,
>> 20K, whether it's 32bit or 64bit.
>>
>
> I think I might be overindexing on the out of tree modifications here.
> Currently, I think it's clear how modifying PERCPU_DYNAMIC_RESERVE
> affects the system with the lower bound being dictated by
> PERCPU_DYNAMIC_EARLY_SIZE. If we bump PERCPU_DYNAMIC_EARLY_SIZE, it's
> not inherently obvious you can drop that value lower depending on your
> system config.
>
> Ultimately, it is only a few pages, so is saving it that big of a deal
> today? Likely not, just a bit wasteful to potentially orphan a few extra
> pages unnecessarily.
>
> Let's just fix this now and I can massage this in the future if anything
> comes up. I appreciate you taking the time to have this discussion with
> me.
>
> Vlastimil, can you please pick up this fix.
Sorry, got a bit lost, so do you mean the original uncoditional bump, or the
modification with BITS_PER_LONG > 32 (or PAGE_SHIFT > 12)?
> Acked-by: Dennis Zhou <dennis@xxxxxxxxxx>
>
> Thanks,
> Dennis
>
>>
>> >
>> > If not `BITS_PER_LONG >32`, we could do `PAGE_SHIFT > 12`.
>> >
>> > Thanks,
>> > Dennis
>> >
>> > > >
>> > > > ---
>> > > > diff --git a/include/linux/percpu.h b/include/linux/percpu.h
>> > > > index f1ec5ad1351c..22ce3271eed2 100644
>> > > > --- a/include/linux/percpu.h
>> > > > +++ b/include/linux/percpu.h
>> > > > @@ -42,7 +42,11 @@
>> > > > * larger than PERCPU_DYNAMIC_EARLY_SIZE.
>> > > > */
>> > > > #define PERCPU_DYNAMIC_EARLY_SLOTS 128
>> > > > +#if BITS_PER_LONG > 32
>> > > > +#define PERCPU_DYNAMIC_EARLY_SIZE (20 << 10)
>> > > > +#else
>> > > > #define PERCPU_DYNAMIC_EARLY_SIZE (12 << 10)
>> > > > +#endif
>> > > >
>> > > > /*
>> > > > * PERCPU_DYNAMIC_RESERVE indicates the amount of free area to piggy
>> > > >
>> > >
>> > >
>> >
>>
>>
