On Thu, Sep 20, 2018 at 11:16:46PM +0530, Devesh Sharma wrote:
> On Thu, Sep 20, 2018 at 10:43 PM Jason Gunthorpe <jgg@xxxxxxxxxxxx> wrote:
> >
> > On Thu, Sep 20, 2018 at 01:08:32PM -0400, Devesh Sharma wrote:
> > > Remove the lock around the doorbell. It isn't necessary
> > > and was a bottleneck for multithreaded clients. Performance of
> > > spdk improves by 40% with it removed.
> > >
> > > Signed-off-by: JD Zheng <jiandong.zheng@xxxxxxxxxxxx>
> > > Signed-off-by: Jonathan Richardson <jonathan.richardson@xxxxxxxxxxxx>
> > > Signed-off-by: Devesh Sharma <devesh.sharma@xxxxxxxxxxxx>
> > > providers/bnxt_re/db.c | 2 --
> > > providers/bnxt_re/main.c | 1 -
> > > providers/bnxt_re/main.h | 1 -
> > > providers/bnxt_re/verbs.c | 2 --
> > > 4 files changed, 6 deletions(-)
> > >
> > > diff --git a/providers/bnxt_re/db.c b/providers/bnxt_re/db.c
> > > index a79f871..2896ceb 100644
> > > +++ b/providers/bnxt_re/db.c
> > > @@ -44,11 +44,9 @@ static void bnxt_re_ring_db(struct bnxt_re_dpi *dpi,
> > > {
> > > __le64 *dbval;
> > >
> > > - pthread_spin_lock(&dpi->db_lock);
> > > dbval = (__le64 *)&hdr->indx;
> > > udma_to_device_barrier();
> > > iowrite64(dpi->dbpage, dbval);
> > > - pthread_spin_unlock(&dpi->db_lock);
> > > }
> >
> > Nope, you have to convert this provider to use the mmio accessor
> > helpers if you want to do this.
> >
> > Ie get rid of iowrite64/iowrite32 and use the macros in util/mmio.h
> >
> > They provide the necessary guarentee for atomicity to allow the lock
> > to be removed.
> >
> > Jason
>
> Okay, this can be done. While removing the lock I was thinking what
> will happen on the 32 bit platform. would those new apis handle 32bit
> platforms too?
Yes
Jason
