On 02/24/2015 03:56 PM, Mark Brown wrote:
> On Tue, Feb 24, 2015 at 03:00:03PM +0200, Stanimir Varbanov wrote:
>
>> +static void spi_qup_dma_done(void *data)
>> +{
>> + struct spi_qup *qup = data;
>> +
>> + if (atomic_dec_and_test(&qup->dma_outstanding))
>> + complete(&qup->done);
>> +}
>
> I'm finding it hard to be thrilled about the use of atomics for
> synchronization (they're just generally hard to work with) and...
>
>> + cookie = dmaengine_submit(desc);
>> + ret = dma_submit_error(cookie);
>> + if (ret)
>> + return ret;
>
>> + atomic_inc(&qup->dma_outstanding);
>
> ..don't we have two potential races here: one if somehow the DMA manages
> to complete prior to the atomic_inc() (unlikely but that's what race
> conditions are all about really) and one if we are issuing multiple DMAs
> and the early ones complete before the later ones are issued?
>
yes, there is a potential race between atomic_inc and dma callback. I
reordered these calls to save few checks, and now it returns to me.
I imagine few options here:
- reorder the dmaengine calls and atomic operations, i.e.
call atomic_inc for rx and tx channels before corresponding
dmaengine_submit and dmaengine_issue_pending.
- have two different dma callbacks and two completions and waiting for
the two.
- manage to receive only one dma callback, i.e. the last transfer in
case of presence of the rx_buf and tx_buf at the same time.
- let me see for better solution.
Thanks for the comments.
regards,
Stan
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
