On 10/13/21 3:09 AM, John Garry wrote:
> On 12/10/2021 19:17, Jens Axboe wrote:
>> Signed-off-by: Jens Axboe<axboe@xxxxxxxxx>
>> ---
>> drivers/nvme/host/pci.c | 69 +++++++++++++++++++++++++++++++++++++----
>> 1 file changed, 63 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
>> index 4ad63bb9f415..4713da708cd4 100644
>> --- a/drivers/nvme/host/pci.c
>> +++ b/drivers/nvme/host/pci.c
>> @@ -959,7 +959,7 @@ static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
>> return ret;
>> }
>>
>> -static void nvme_pci_complete_rq(struct request *req)
>> +static void nvme_pci_unmap_rq(struct request *req)
>> {
>> struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
>> struct nvme_dev *dev = iod->nvmeq->dev;
>> @@ -969,9 +969,34 @@ static void nvme_pci_complete_rq(struct request *req)
>> rq_integrity_vec(req)->bv_len, rq_data_dir(req));
>> if (blk_rq_nr_phys_segments(req))
>> nvme_unmap_data(dev, req);
>> +}
>> +
>> +static void nvme_pci_complete_rq(struct request *req)
>> +{
>> + nvme_pci_unmap_rq(req);
>> nvme_complete_rq(req);
>> }
>>
>> +static void nvme_pci_complete_batch(struct io_batch *ib)
>> +{
>> + struct request *req;
>> +
>> + req = ib->req_list;
>> + while (req) {
>> + nvme_pci_unmap_rq(req);
>
> This will do the DMA SG unmap per request. Often this is a performance
> bottle neck when we have an IOMMU enabled in strict mode. So since we
> complete in batches, could we combine all the SGs in the batch to do one
> big DMA unmap SG, and not one-by-one?
It is indeed, I actually have a patch for persistent maps as well. But even
without that, it would make sense to handle these unmaps a bit smarter. That
requires some iommu work though which I'm not that interested in right now,
could be done on top of this one for someone motivated enough.
--
Jens Axboe
