[PATCH v6 3/3] qe_common: add qe_muram_ functions to manage muram

[Laura Abbott]

On 08/24/2015 08:03 PM, Zhao Qiang wrote:
>
>> -----Original Message-----
>> From: Laura Abbott [mailto:labbott at redhat.com]
>> Sent: Tuesday, August 25, 2015 7:32 AM
>> To: Zhao Qiang-B45475; Wood Scott-B07421
>> Cc: linux-kernel at vger.kernel.org; linuxppc-dev at lists.ozlabs.org;
>> lauraa at codeaurora.org; Xie Xiaobo-R63061; benh at kernel.crashing.org; Li
>> Yang-Leo-R58472; paulus at samba.org
>> Subject: Re: [PATCH v6 3/3] qe_common: add qe_muram_ functions to manage
>> muram
>>
>> On 08/24/2015 02:31 AM, Zhao Qiang wrote:
>>
>>
>>> +out:
>>> +	of_node_put(np);
>>> +	return ret;
>>> +}
>>> +
>>> +/**
>>> + * qe_muram_alloc - allocate the requested size worth of multi-user
>>> +ram
>>> + * @size: number of bytes to allocate
>>> + * @align: requested alignment, in bytes
>>> + *
>>> + * This function returns an offset into the muram area.
>>> + * Use qe_dpram_addr() to get the virtual address of the area.
>>> + * Use qe_muram_free() to free the allocation.
>>> + */
>>> +unsigned long qe_muram_alloc(unsigned long size, unsigned long align)
>>> +{
>>> +	unsigned long start;
>>> +	unsigned long flags;
>>> +	struct muram_block *entry;
>>> +
>>> +	spin_lock_irqsave(&qe_muram_lock, flags);
>>> +	muram_pool_data.align = align;
>>> +	start = gen_pool_alloc(muram_pool, size);
>>
>> The advantage of creating gen_pool_alloc_data was so that you could pass
>> in the align automatically without having to modify the structure.
>> Is there a reason you aren't using that?
>>
>>> +	memset(qe_muram_addr(start), 0, size);
>>
>> There doesn't seem to be a check for allocation failure from the
>> gen_alloc.
>
> gen_pool_alloc will return 0 if there is error, but if the address returned is
> just 0x0, it can't distinguish it is address or error.
>

Yes, that's a bad limitation of gen_pool. Maybe one day that will get fixed.
In a previous out of tree driver, I worked around this by offsetting the
gen_pool_add by a constant so any return value was non-zero and out of memory
was zero and then subtracting the constant off of the return value. Not sure
if that's better or worse than just fixing gen_alloc.
  
>>
>>> +	entry = kmalloc(sizeof(*entry), GFP_KERNEL);
>>> +	if (!entry)
>>> +		goto out;
>>> +	entry->start = start;
>>> +	entry->size = size;
>>> +	list_add(&entry->head, &muram_block_list);
>>
>> What's the point of keeping the block list anyway? It's used only in this
>> file and it only seems to duplicate what gen_alloc is doing internally.
>> Is there some lookup functionality you still need? Could you use a
>> gen_alloc API to do so?
>
> I need to record the size when allocation, so when free the block, I can get
> The right size for the block, and pass the right size to
> gen_pool_free(struct gen_pool *pool, unsigned long addr, size_t size).
>

Yes, I see now what you are doing.
  
>>
>>
>> Thanks,
>> Laura
> Thanks
> Zhao
>

Thanks,
Laura