Re: [PATCH] Remove GFP_DMA and GFP_DMA32 from flags before passing it into kmalloc.

Re: [PATCH] Remove GFP_DMA and GFP_DMA32 from flags before passing it into kmalloc.

[Catalin Marinas]

On 27 Sep 2014, at 16:09, Miles MH Chen <orca.chen@gmail.com> wrote:
> Signed-off-by: Min-Hua Chen <orca.chen@gmail.com>
> ---
>  arch/arm64/mm/dma-mapping.c |    3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
> index 4164c5a..273cf6d 100644
> --- a/arch/arm64/mm/dma-mapping.c
> +++ b/arch/arm64/mm/dma-mapping.c
> @@ -103,7 +103,8 @@ static void *__dma_alloc_noncoherent(struct device *dev, size_t size,
>      ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
>      if (!ptr)
>          goto no_mem;
> -    map = kmalloc(sizeof(struct page *) << order, flags & ~GFP_DMA);
> +    map = kmalloc(sizeof(struct page *) << order,
> +                    flags & ~(GFP_DMA | GFP_DMA32));

Do you have an explanation on why this is needed (and such explanation
should also be included in the commit log)? We don’t use ZONE_DMA32 on
arm64 (we did initially but it was for the wrong reasons).

Thanks,

Catalin

Re: [PATCH] Remove GFP_DMA and GFP_DMA32 from flags before passing it into kmalloc.

[Miles MH Chen]

On Sat, Sep 27, 2014 at 11:13 PM, Catalin Marinas
<catalin.marinas@arm.com> wrote:
> On 27 Sep 2014, at 16:09, Miles MH Chen <orca.chen@gmail.com> wrote:
>> Signed-off-by: Min-Hua Chen <orca.chen@gmail.com>
>> ---
>>  arch/arm64/mm/dma-mapping.c |    3 ++-
>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
>> index 4164c5a..273cf6d 100644
>> --- a/arch/arm64/mm/dma-mapping.c
>> +++ b/arch/arm64/mm/dma-mapping.c
>> @@ -103,7 +103,8 @@ static void *__dma_alloc_noncoherent(struct device *dev, size_t size,
>>      ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
>>      if (!ptr)
>>          goto no_mem;
>> -    map = kmalloc(sizeof(struct page *) << order, flags & ~GFP_DMA);
>> +    map = kmalloc(sizeof(struct page *) << order,
>> +                    flags & ~(GFP_DMA | GFP_DMA32));
>
> Do you have an explanation on why this is needed (and such explanation
> should also be included in the commit log)? We don’t use ZONE_DMA32 on
> arm64 (we did initially but it was for the wrong reasons).

If GFP_DMA32 is passed to dma_alloc_coherent, the flag will go to kmalloc and
trigger a BUG_ON check in slab allocator:

__dma_alloc_noncoherent
kmalloc
new_slab  BUG_ON(flags & GFP_SLAB_BUG_MASK);

It can be avoided this by removing GFP_DMA32 before passing it to kmalloc
or we should block GFP_DMA32 flag earlier in arch/arm64/dma-mapping.c if
GFP_DMA32 is not allowed in arch/arm64/dma-mapping.c.

Thanks,
MH

>
> Thanks,
>
> Catalin

Re: [PATCH] Remove GFP_DMA and GFP_DMA32 from flags before passing it into kmalloc.

[Catalin Marinas]

On Sun, Sep 28, 2014 at 01:44:15AM +0100, Miles MH Chen wrote:
> On Sat, Sep 27, 2014 at 11:13 PM, Catalin Marinas
> <catalin.marinas@arm.com> wrote:
> > On 27 Sep 2014, at 16:09, Miles MH Chen <orca.chen@gmail.com> wrote:
> >> Signed-off-by: Min-Hua Chen <orca.chen@gmail.com>
> >> ---
> >>  arch/arm64/mm/dma-mapping.c |    3 ++-
> >>  1 file changed, 2 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
> >> index 4164c5a..273cf6d 100644
> >> --- a/arch/arm64/mm/dma-mapping.c
> >> +++ b/arch/arm64/mm/dma-mapping.c
> >> @@ -103,7 +103,8 @@ static void *__dma_alloc_noncoherent(struct device *dev, size_t size,
> >>      ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
> >>      if (!ptr)
> >>          goto no_mem;
> >> -    map = kmalloc(sizeof(struct page *) << order, flags & ~GFP_DMA);
> >> +    map = kmalloc(sizeof(struct page *) << order,
> >> +                    flags & ~(GFP_DMA | GFP_DMA32));
> >
> > Do you have an explanation on why this is needed (and such explanation
> > should also be included in the commit log)? We don’t use ZONE_DMA32 on
> > arm64 (we did initially but it was for the wrong reasons).
> 
> If GFP_DMA32 is passed to dma_alloc_coherent, the flag will go to kmalloc and
> trigger a BUG_ON check in slab allocator:
> 
> __dma_alloc_noncoherent
> kmalloc
> new_slab  BUG_ON(flags & GFP_SLAB_BUG_MASK);
> 
> It can be avoided this by removing GFP_DMA32 before passing it to kmalloc
> or we should block GFP_DMA32 flag earlier in arch/arm64/dma-mapping.c if
> GFP_DMA32 is not allowed in arch/arm64/dma-mapping.c.

With commit d4932f9e81ae (arm64: add atomic pool for non-coherent and CMA
allocations), this fix is no longer needed. Prior to this we could
simply use GFP_KERNEL rather than masking out bits as we don't care
about atomic contexts prior to the commit I mentioned.

-- 
Catalin