Re: [PATCH] MIPS: Alchemy: fix dbdma ring destruction memory debugcheck.

[David Daney <ddaney@caviumnetworks.com>]

Manuel Lauss wrote:
> David,
> 
> On Tue, Jan 26, 2010 at 6:58 PM, Manuel Lauss
> <manuel.lauss@googlemail.com> wrote:
>> On Tue, Jan 26, 2010 at 6:44 PM, David Daney <ddaney@caviumnetworks.com> wrote:
>>> Manuel Lauss wrote:
>>>> DBDMA descriptors need to be located at 32-byte aligned addresses;
>>>> however kmalloc rarely delivers such addresses.  The dbdma code
>>>> works around that by allocating 63 bytes and re-aligning the
>>>> descriptor base afterwards.  Hoewever when freeing memory it does
>>>> not account for this adjustment and trips the kfree debugcheck:
>>>>
>>> Correct me if I am wrong, but don't kmalloc et al. return blocks aligned
>>>  boundaries of the size rounded up the the next power of two?  So if you
>>> need 32-byte aligned addresses, just use a size value of 32 or greater.  You
>>> wouldn't have to add 63 and do masking and remember the membase value as you
>>> do in the patch.
>> The description is not completely correct (I suck a writing those):
>> It allocates a number
>> of descriptor entries (64 bytes each) specified by the driver in a single block:
>>
>>        desc_base = (u32)kmalloc(entries * sizeof(au1x_ddma_desc_t),
>>                                 GFP_KERNEL|GFP_DMA);
>>        if (desc_base == 0)
>>                return 0;
>>
>>        if (desc_base & 0x1f) {
>>
>> So far the 3 users I have (mmc, spi, audio) always return true on the above
>> check (2 descriptors for audio for instance).
> 
> ... but only if CONFIG_DEBUG_SLAB is enabled.

Oh no!  I make this assumption in my drivers.  If CONFIG_DEBUG_SLAB 
causes it to not be true... Bad news.

David Daney

> You are correct in that
> kmalloc() returns aligned blocks with a non-debug slab allocator and this
> fix becomes superfluous.   I'll try to confirm this with the other
> allocators and
> resend with a fixed description.
> 
> Thank you for the pointer!
>      Manuel Lauss
>