Re: [RESEND PATCH v9] mtd: spi-nor: add hisilicon spi-nor flash controller driver

[Jiancheng Xue]

Hi Boris,

On 2016/4/12 17:44, Boris Brezillon wrote:
> +Russell
> 
> Hi Jiancheng,
> 
> On Tue, 12 Apr 2016 17:32:08 +0800
> Jiancheng Xue <xuejiancheng@xxxxxxxxxx> wrote:
> 
>> Hi Marek,
>>
>> On 2016/4/12 3:21, Marek Vasut wrote:
>>> On 04/11/2016 03:28 AM, Jiancheng Xue wrote:
>>>> Hi,
>>>>
>>>> On 2016/4/8 18:04, Marek Vasut wrote:
>>>>> On 04/08/2016 10:26 AM, Jiancheng Xue wrote:
>>>>>> Hi,
>>>>>>
>>>>>> On 2016/4/7 10:28, Marek Vasut wrote:
>>>>>>> On 04/07/2016 04:10 AM, Jiancheng Xue wrote:
>>>>>>>> Hi Brian,
>>>>>>>>    Thank you very much for your comments. I'll fix these issues in next version.
>>>>>>>> In addition, for easy understanding I'd like to rewrite hisi_spi_nor_write and
>>>>>>>> hisi_spi_nor_read. Your comments on these modifications will be highly appreciated.
>>>>>>>
>>>>>>> Would you please stop top-posting ? It rubs some people the wrong way.
>>>>>>>
>>>>>> I feel very sorry about that. I have read the etiquette and won't make the same mistake again.
>>>>>>
>>>>>>>> static int hisi_spi_nor_read(struct spi_nor *nor, loff_t from, size_t len,
>>>>>>>> 		size_t *retlen, u_char *read_buf)
>>>>>>>> {
>>>>>>>> 	struct hifmc_priv *priv = nor->priv;
>>>>>>>> 	struct hifmc_host *host = priv->host;
>>>>>>>> 	int i;
>>>>>>>>
>>>>>>>> 	/* read all bytes in only one time */
>>>>>>>> 	if (len <= HIFMC_DMA_MAX_LEN) {
>>>>>>>> 		hisi_spi_nor_dma_transfer(nor, from, host->dma_buffer,
>>>>>>>> 				len, FMC_OP_READ);
>>>>>>>> 		memcpy(read_buf, host->buffer, len);
>>>>>>>
>>>>>>> Is all the ad-hoc memcpying necessary? I think you can use
>>>>>>> dma_map_single() and co to obtain DMAble memory for your
>>>>>>> controller's use and then you can probably get rid of most
>>>>>>> of this stuff.
>>>>>>>
>>>>>> Considering read_buf >= high_mem case, I think it is also complicated to use dma_map_*
>>>>>> and the DMA buffer allocated by the driver is still needed. But I am not sure about
>>>>>> this. Please let me know if I am wrong. Thank you!
>>>>>
>>>>> Does your controller/DMA have a limitation where it's buffers must be in
>>>>> the bottom 4GiB range ? The DMA framework should be able to take care of
>>>>> such platform limitations.
>>>>>
>>>> When read_buf is allocated by vmalloc, the underlying physical memory may be not contiguous.
>>>> In this case, dma_map_single can't be used directly. I think inner DMA buffer and memcpy are still
>>>> needed. Am I right?
>>>
>>> Take a look at drivers/spi/spi-mxs.c , look for "vmalloc" , does that
>>> solution help you in any way ?
>>>
>> No. I think this solution just processes the buffer within only one page.
>> I had referred to drivers/mtd/onenand/samsung.c and other files.
>> The corresponding code segment is as follows:
>> static int s5pc110_read_bufferram(struct mtd_info *mtd, int area,
>>                 unsigned char *buffer, int offset, size_t count)
>> {
>> 	void *buf = (void *) buffer;
>>         dma_addr_t dma_src, dma_dst;
>> 	...
>> 	/* Handle vmalloc address */
>>         if (buf >= high_memory) {
>>                 struct page *page;
>>
>>                 if (((size_t) buf & PAGE_MASK) !=
>>                     ((size_t) (buf + count - 1) & PAGE_MASK))
>>                         goto normal;
>>                 page = vmalloc_to_page(buf);
>>                 if (!page)
>>                         goto normal;
>>
>>                 ...
>>         } else {
>> 		...
>> 	}
>>
>> normal:
>>         ...
>>         memcpy(buffer, p, count);
>>
>> 	return 0;
>> }
>> I think memcpy in "normal" clause can't be removed. So I'd like to keep my original
>> implementation if it is also OK. What's your opinion?
> 
> You might want to have a look at this series [1], and particularly at
> Russell's answers regarding DMA operations on non-lowmem memory.
> 
Thank you very much for your supplied reference. Besides safety reasons described by Russell,
the dmaengine embeded in this controller doesn't support scatter-list type buffer directly.
So for this controller, I think now it's better to obtain buffer through dma_alloc_coherent
for dma operation, and then copy data to buffers supplied by mtd user. May I keep using this
implementation now?

Regards,
Jiancheng