Re: mm/percpu.c: use smarter memory allocation for struct pcpu_alloc_info (crisv32 hang)

[Guenter Roeck]

Hi,

On Tue, Oct 03, 2017 at 06:29:49PM -0400, Nicolas Pitre wrote:
> On Tue, 3 Oct 2017, Tejun Heo wrote:
> 
> > On Tue, Oct 03, 2017 at 04:57:44PM -0400, Nicolas Pitre wrote:
> > > This can be much smaller than a page on very small memory systems. 
> > > Always rounding up the size to a page is wasteful in that case, and 
> > > required alignment is smaller than the memblock default. Let's round 
> > > things up to a page size only when the actual size is >= page size, and 
> > > then it makes sense to page-align for a nicer allocation pattern.
> > 
> > Isn't that a temporary area which gets freed later during boot?
> 
> Hmmm...
> 
> It may get freed through 3 different paths where 2 of them are error 
> paths. What looks like a non-error path is in pcpu_embed_first_chunk() 
> called from setup_per_cpu_areas(). But there are two versions of 
> setup_per_cpu_areas(): one for SMP and one for !SMP. And the !SMP case 
> never calls pcpu_free_alloc_info() currently.
> 
> I'm not sure i understand that code fully, but maybe the following patch 
> could be a better fit:
> 
> ----- >8
> Subject: [PATCH] percpu: don't forget to free the temporary struct pcpu_alloc_info
> 
> Unlike the SMP case, the !SMP case does not free the memory for struct 
> pcpu_alloc_info allocated in setup_per_cpu_areas(). And to give it a 
> chance of being reused by the page allocator later, align it to a page 
> boundary just like its size.
> 
> Signed-off-by: Nicolas Pitre <nico@xxxxxxxxxx>

This patch causes my crisv32 qemu emulation to hang with no console output.

> 
> diff --git a/mm/percpu.c b/mm/percpu.c
> index 434844415d..caab63375b 100644
> --- a/mm/percpu.c
> +++ b/mm/percpu.c
> @@ -1416,7 +1416,7 @@ struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
>  			  __alignof__(ai->groups[0].cpu_map[0]));
>  	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
>  
> -	ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
> +	ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), PAGE_SIZE);
>  	if (!ptr)
>  		return NULL;
>  	ai = ptr;
> @@ -2295,6 +2295,7 @@ void __init setup_per_cpu_areas(void)
>  
>  	if (pcpu_setup_first_chunk(ai, fc) < 0)
>  		panic("Failed to initialize percpu areas.");
> +	pcpu_free_alloc_info(ai);

This is the culprit. Everything works fine if I remove this line.

No idea if the problem is here or in the cris core.
Copying cris maintainers for input.

Guenter