On 2019/12/12 0:54, Vinod Koul wrote:
> On 11-12-19, 19:43, Zhou Wang wrote:
>> On 2019/12/11 18:52, Vinod Koul wrote:
>>> On 10-12-19, 10:13, Zhou Wang wrote:
>
>>>> +static int hisi_dma_terminate_all(struct dma_chan *c)
>>>> +{
>>>> + struct hisi_dma_chan *chan = to_hisi_dma_chan(c);
>>>> + unsigned long flags;
>>>> + LIST_HEAD(head);
>>>> +
>>>> + spin_lock_irqsave(&chan->vc.lock, flags);
>>>> +
>>>> + hisi_dma_pause_dma(chan->hdma_dev, chan->qp_num, true);
>>>> + if (chan->desc) {
>>>> + vchan_terminate_vdesc(&chan->desc->vd);
>>>> + chan->desc = NULL;
>>>> + }
>>>> +
>>>> + vchan_get_all_descriptors(&chan->vc, &head);
>>>> +
>>>> + spin_unlock_irqrestore(&chan->vc.lock, flags);
>>>> +
>>>> + vchan_dma_desc_free_list(&chan->vc, &head);
>>>> + hisi_dma_pause_dma(chan->hdma_dev, chan->qp_num, false);
>>>
>>> pause on terminate? Not DISABLE?
>>
>> here this function just aborts transfers on specific channel.
>
> yeah and I would expect the channel to go into disable state right!
>
>>>> +static struct pci_driver hisi_dma_pci_driver = {
>>>> + .name = "hisi_dma",
>>>> + .id_table = hisi_dma_pci_tbl,
>>>> + .probe = hisi_dma_probe,
>>>
>>> no .remove and kconfig has a tristate option!
>>
>> Use devres APIs in probe, so seems nothing should be done in remove :)
>
> who will de-register from dmaengine, you have dangiling chan_tasklet
> which needs to be killed and you have isr which is still enabled, yeah
> what could go wrong!
I missed to do tasklet_kill when disabling qps. And also missed to do
pci_free_irq_vectors, which seems can be done by adding a devres callback
by devm_add_action_or_reset.
Here we use devres API to register to dmaengine, so no need to de-register
from it.
Thanks!
Zhou
>
> Please, deregister from dmaengine, kill the vchan tasklet and make sur
> irq is disabled and tasklets killed
>
