2014-10-25 19:19:36

by Zhihui Zhang

Subject: [PATCH] [percpu] Make the unit size of the first chunk the same as other chunks

Since we have already allocated the full unit size for the first chunk, we might as well use
it so that the unit size are the same for all chunks. The page first chunk allocator already
has this effect because it allocates one page at a time.

Signed-off-by: Zhihui Zhang <[email protected]>
---
mm/percpu.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/mm/percpu.c b/mm/percpu.c
index 014bab6..7242360 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -1960,6 +1960,7 @@ int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
return PTR_ERR(ai);

size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
+ ai->dyn_size += ai->unit_size - size_sum;
areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));

areas = memblock_virt_alloc_nopanic(areas_size, 0);
@@ -2006,9 +2007,8 @@ int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
free_fn(ptr, ai->unit_size);
continue;
}
- /* copy and return the unused part */
+ /* copy static data */
memcpy(ptr, __per_cpu_load, ai->static_size);
- free_fn(ptr + size_sum, ai->unit_size - size_sum);
}
}

@@ -2034,7 +2034,7 @@ int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
}

pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
- PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
+ PFN_DOWN(ai->unit_size), base, ai->static_size, ai->reserved_size,
ai->dyn_size, ai->unit_size);

rc = pcpu_setup_first_chunk(ai, base);
--
1.8.1.2

2014-10-27 14:08:32

by Tejun Heo

Subject: Re: [PATCH] [percpu] Make the unit size of the first chunk the same as other chunks

On Sat, Oct 25, 2014 at 11:05:58AM -0400, Zhihui Zhang wrote:
> Since we have already allocated the full unit size for the first chunk, we might as well use
> it so that the unit size are the same for all chunks. The page first chunk allocator already
> has this effect because it allocates one page at a time.

I'm not following. Where do we allocate the full unit size for the
first chunk?

Thanks.

--
tejun

2014-10-27 23:32:48

by Zhihui Zhang

Subject: Re: [PATCH] [percpu] Make the unit size of the first chunk the same as other chunks

In pcpu_embed_first_chunk(), we allocate full unit size for each CPU
in the first chunk:

1981 /* allocate space for the whole group */
1982 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size,
atom_size);
1983 if (!ptr) {
1984 rc = -ENOMEM;
1985 goto out_free_areas;
1986 }


Later we freed unused part:


2009 /* copy and return the unused part */
2010 memcpy(ptr, __per_cpu_load, ai->static_size);
2011 free_fn(ptr + size_sum, ai->unit_size - size_sum);

I am trying to make each CPU to have a full unit size in the first
chunk, same as in all other chunks. Does this make sense?

-Zhihui

On Mon, Oct 27, 2014 at 10:08 AM, Tejun Heo <[email protected]> wrote:
> On Sat, Oct 25, 2014 at 11:05:58AM -0400, Zhihui Zhang wrote:
>> Since we have already allocated the full unit size for the first chunk, we might as well use
>> it so that the unit size are the same for all chunks. The page first chunk allocator already
>> has this effect because it allocates one page at a time.
>
> I'm not following. Where do we allocate the full unit size for the
> first chunk?
>
> Thanks.
>
> --
> tejun

2014-10-28 03:57:00

by Tejun Heo

Subject: Re: [PATCH] [percpu] Make the unit size of the first chunk the same as other chunks

On Mon, Oct 27, 2014 at 07:32:47PM -0400, Zhihui Zhang wrote:
> I am trying to make each CPU to have a full unit size in the first
> chunk, same as in all other chunks. Does this make sense?

And why are you trying to do that? Unit size may be highly variable.
Why would it make sense to reserve memory area according to a value
which can fluctuate for reasons completely unrelated to actual percpu
memory usage?

Thanks.

--
tejun

2014-10-30 00:07:08

by Zhihui Zhang

Subject: Re: [PATCH] [percpu] Make the unit size of the first chunk the same as other chunks

I see your point. Thanks.

-Zhihui

On Wed, Oct 29, 2014 at 12:18 AM, Tejun Heo <[email protected]> wrote:
> Plesae restore lkml cc when replying.
>
> On Tue, Oct 28, 2014 at 08:12:30PM -0400, Zhihui Zhang wrote:
>> My patch just increases the dynamic area in the first chunk slightly
>> to cover the round up surplus. On my 64-bit laptop, it is 12288 bytes.
>
> As I wrote before, it's 12288 bytes on your laptop but it can be much
> larger on other setups.
>
>> It will mostly likely be used, and in fact, a second chunk will be
>> most likely needed as well. So in theory you are right, but in
>> practice, it probably won't matter.
>
> If the initial dynamic reserve is too small, which could be the case
> given the overall increase in percpu memory, increase
> PERCPU_DYNAMIC_EARLY_SIZE.
>
> --
> tejun