[PATCH v8 24/24] nvme-pci: optimize single-segment handling

[Leon Romanovsky]

On Fri, Apr 18, 2025 at 05:02:38PM +0900, Damien Le Moal wrote:
> On 4/18/25 15:47, Leon Romanovsky wrote:
> > From: Kanchan Joshi <joshi.k at samsung.com>
> > 
> > blk_rq_dma_map API is costly for single-segment requests.
> > Avoid using it and map the bio_vec directly.
> > 
> > Signed-off-by: Kanchan Joshi <joshi.k at samsung.com>
> > Signed-off-by: Nitesh Shetty <nj.shetty at samsung.com>
> > Signed-off-by: Leon Romanovsky <leonro at nvidia.com>
> > ---
> >  drivers/nvme/host/pci.c | 65 +++++++++++++++++++++++++++++++++++++----
> >  1 file changed, 60 insertions(+), 5 deletions(-)

<...>

> >  static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
> >  {
> >  	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
> > +	unsigned int nr_segments = blk_rq_nr_phys_segments(req);
> > +	dma_addr_t dma_addr;
> > +
> > +	if (nr_segments == 1 && (iod->flags & IOD_SINGLE_SEGMENT)) {
> 
> nvme_pci_setup_prps() calls nvme_try_setup_prp_simple() which sets
> IOD_SINGLE_SEGMENT if and only if the req has a single phys segment. So why do
> you need to count the segments again here ? Looking at the flag only should be
> enough, no ?

Yes, you are right. There is no need in extra check of nr_segments and
it is enough to rely on IOD_SINGLE_SEGMENT.

Thanks