Re: [spi-devel-general] Re: [PATCH/RFC] SPI: add DMAUNSAFE analog

[Vitaly Wool <vwool@ru.mvista.com>]

David Brownell wrote:

>On Wednesday 14 December 2005 10:47 pm, Vitaly Wool wrote:
>
>  
>
>>One cannot allocate memory in interrupt context, so the way to go is 
>>allocating it on stack, thus the buffer is not DMA-safe.
>>    
>>
>
>kmalloc(..., GFP_ATOMIC) is the way to allocate memory in irq context.
>It's done that way throughout the kernel.
>  
>
It's not applicable within the RT-related changes. kmalloc anyway takes 
mutexes, so allocationg it in interrupt context is buggy.
*Legacy* kernel code does that but why produce a new code with that?

>
>  
>
>>Making it DMA-safe in thread that does the very message processing is a 
>>good way of overcoming this.
>>    
>>
>
>The rest of Linux appears to work fine without needing such mechanisms...
>  
>
The rest of Linux still has a lot of bugs. Noone I guess is ready to 
argue that.

>I really fail to see why you think SPI needs that.  USB isn't the only
>counterexample, but it's particularly relevant since both USB and SPI
>use asynchronous message passing over serial links ... and USB has a
>rather complete driver stack over it.   (None of the USB based WLAN
>drivers need those static buffers you worry about, by the way...)
>  
>
I haven't heard of USB device registers needing to be written in IRQ 
context. I'm not that well familiar with USB, so if you give such an 
example, that'd be fine.

>>Using preallocated buffer is not a good way, since it may well be 
>>already used by another interrupt or not yet processed by the worker 
>>thread (or tasklet, or whatever).
>>    
>>
>
>We would call those "driver bugs" and expect patches.  :)
>
>  
>
Well, I know one such patch ;)

>>One can not predict how many transfers are  
>>gonna be dropped due to "previous trransfer is being processed" problem, 
>>it depends on the system load. And though it's not a problem for 
>>touchscreen, it *will* be a problem if it were MMC, for instance.
>>    
>>
>
>Huh, well I've already seen one nice "mmc/sd over SPI" driver, using
>a slightly earlier version of the framework than is found in mm3.
>It's being used for root filesystems.
>
>So demonstrably there is no problem for MMC/SD, either.
>  
>
No problem... Has the driver been tested in stress conditions? If not, 
then I guess you can't say this for sure :)

Vitaly