On 3/8/2019 10:05 AM, Christoph Hellwig wrote:
On Sun, Mar 03, 2019 at 06:44:21PM +0200, Max Gurtovoy wrote:
This is interesting idea, but it's a big change to be done for each ULP to
use sg_table (only NVMeoF use it AFAIK). I think this transition should be
done separately (and well tested !!). We can do it before or we can do it
after this patchset.
Nothing uses ib_scatterlist either, as you just introduce it. I'd
rather move to a common and more useful structure directly.
We've started exploring the sg_table option and found that there is no
dma_nents arg there.
So I guess you've meant that we should add functionality there (and just
use the sg_table struct "as is" without using
sg_alloc_table_chained/sg_init_table/sg_free_table_chained).
how is this pseudo code in your opinion :
sg_table->sgl = sgl;
sg_table->nents = nents; //get from blk_rq_nr_phys_segments or scsi_sg_count
sg_table->dma_nents = dma_nents; //returned from ib_dma_map_sg
for NON-PI:
ib_map_mr_sg(mr, sg_table, offset_p, SZ_4K);
for PI (set meta_sg_table as above):
ib_map_mr_sg_pi(mr, sg_table, offset_p, meta_sg_table, meta_offset_p,
SZ_4K);
