[PATCH 2/3] nvme: Fix missing address translation for PCIe inbound access

[Neil Armstrong]

On 5/12/26 11:30, Torsten Duwe wrote:
> On Mon, 11 May 2026 16:50:48 +0200
> Neil Armstrong <neil.armstrong at linaro.org> wrote:
> 
>> On 5/8/26 17:42, Torsten Duwe wrote:
> [...]
>>> +#define DEV_ADDR(a)		dev_phys_to_bus(dev->udev, (a))
>>
>> This doesn't look very clean, I would do something closer to linux by precalculating
>> the DMA addresses after allocation like:
>> ====================><================================
>> diff --git a/drivers/nvme/nvme.c b/drivers/nvme/nvme.c
>> index 2b14437f69c..1ce06402a90 100644
>> --- a/drivers/nvme/nvme.c
>> +++ b/drivers/nvme/nvme.c
>> @@ -241,6 +241,7 @@ static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev,
>>    	nvmeq->sq_cmds = (void *)memalign(4096, NVME_SQ_SIZE(depth));
>>    	if (!nvmeq->sq_cmds)
>>    		goto free_queue;
>> +	nvmeq->sq_dma_addr = dev_phys_to_bus(dev->udev, nvmeq->sq_cmds);
>>    	memset((void *)nvmeq->sq_cmds, 0, NVME_SQ_SIZE(depth));
>>
>>    	nvmeq->dev = dev;
>> @@ -393,7 +394,7 @@ static int nvme_configure_admin_queue(struct nvme_dev *dev)
>>    	dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
>>
>>    	writel(aqa, &dev->bar->aqa);
>> -	nvme_writeq((ulong)nvmeq->sq_cmds, &dev->bar->asq);
>> +	nvme_writeq((ulong)nvmeq->sq_dma_addr, &dev->bar->asq);
>>    	nvme_writeq((ulong)nvmeq->cqes, &dev->bar->acq);
>>
>>    	result = nvme_enable_ctrl(dev);
> 
> Hm, time / space tradeoff. I don't really have a strong preference
> regarding this, I'd worry most about U-Boot code size (for which this
> doesn't make much of a difference). However, I was under the impression
> that during regular NVMe business a lot of those buffers and addresses
> get allocated and the translations are required often. That's probably
> only true once a full-blown operating system is running.
> 
> I'll check the allocations and assignments. Any other opinions?

Other that that, I think it's fine, Linux is basically doing the same.

Thanks,
Neil

> 
> 	Torsten