Re: [PATCH] hw/nvme: be more careful when deasserting IRQs

["Jakub Jermář" <jakub.jermar@kernkonzept.com>]

On 6/10/21 8:14 PM, Klaus Jensen wrote:
> On Jun 10 13:46, Jakub Jermář wrote:
>> An IRQ vector used by a completion queue cannot be deasserted without
>> first checking if the same vector does not need to stay asserted for
>> some other completion queue.
>>
>> Signed-off-by: Jakub Jermar <jakub.jermar@kernkonzept.com>
>> ---
>> hw/nvme/ctrl.c | 21 +++++++++++++++++++--
>> 1 file changed, 19 insertions(+), 2 deletions(-)
>>
>> diff --git a/hw/nvme/ctrl.c b/hw/nvme/ctrl.c
>> index 0bcaf7192f..c0980929eb 100644
>> --- a/hw/nvme/ctrl.c
>> +++ b/hw/nvme/ctrl.c
>> @@ -473,6 +473,21 @@ static void nvme_irq_deassert(NvmeCtrl *n, 
>> NvmeCQueue *cq)
>>     }
>> }
>>
>> +/*
>> + * Check if the vector used by the cq can be deasserted, i.e. it 
>> needn't be
>> + * asserted for some other cq.
>> + */
>> +static bool nvme_irq_can_deassert(NvmeCtrl *n, NvmeCQueue *cq)
>> +{
>> +    for (unsigned qid = 0; qid < n->params.max_ioqpairs + 1; qid++) {
>> +        NvmeCQueue *q = n->cq[qid];
>> +
>> +        if (q && q->vector == cq->vector && q->head != q->tail)
>> +            return false;  /* some queue needs this to stay asserted */
>> +    }
>> +    return true;
>> +}
>> +
>> static void nvme_req_clear(NvmeRequest *req)
>> {
>>     req->ns = NULL;
>> @@ -4089,7 +4104,9 @@ static uint16_t nvme_del_cq(NvmeCtrl *n, 
>> NvmeRequest *req)
>>         trace_pci_nvme_err_invalid_del_cq_notempty(qid);
>>         return NVME_INVALID_QUEUE_DEL;
>>     }
>> -    nvme_irq_deassert(n, cq);
>> +    if (nvme_irq_can_deassert(n, cq)) {
>> +        nvme_irq_deassert(n, cq);
>> +    }
>>     trace_pci_nvme_del_cq(qid);
>>     nvme_free_cq(cq, n);
>>     return NVME_SUCCESS;
>> @@ -5757,7 +5774,7 @@ static void nvme_process_db(NvmeCtrl *n, hwaddr 
>> addr, int val)
>>             timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) 
>> + 500);
>>         }
>>
>> -        if (cq->tail == cq->head) {
>> +        if (nvme_irq_can_deassert(n, cq)) {
>>             nvme_irq_deassert(n, cq);
>>         }
>>     } else {
>> -- 
>> 2.31.1
>>
> 
> This is actually an artifact of commit ca247d35098d3 ("hw/block/nvme: 
> fix pin-based interrupt behavior") that I did a year ago. Prior to that 
> fix, the completion queue id was used to index the internal IS register 
> (irq_status), which, while wrong spec-wise, had the effect of... 
> actually working.
> 
> Anyway, I agree that the logic is flawed right now, since we should only 
> deassert when all outstanding cqe's have been acknowledged by the host.
> 
> nvme_irq_can_deassert should be guarded with a check on msix_enabled(), 
> but in any case I am not happy about looping over all completion queues 
> on each cq doorbell write. I think this can be ref counted? I.e. 
> decrement when cq->tail == cq->head on the cq doorbell write and 
> increment only when going from empty to non-empty in nvme_post_cqes().

I reworked this to use reference counting in v2.

Jakub