On Mon, Jul 28, 2025 at 07:49:30PM +0100, Pavel Begunkov wrote: > On 7/28/25 19:39, Mina Almasry wrote: > > On Mon, Jul 28, 2025 at 11:35 AM Pavel Begunkov <asml.silence@xxxxxxxxx> wrote: > > > > > > On 7/28/25 06:27, Byungchul Park wrote: > > > > Changes from v1: > > > > 1. Rebase on linux-next. > > > > > > net-next is closed, looks like until August 11. > > > > > > > 2. Initialize net_iov->pp = NULL when allocating net_iov in > > > > net_devmem_bind_dmabuf() and io_zcrx_create_area(). > > > > 3. Use ->pp for net_iov to identify if it's pp rather than > > > > always consider net_iov as pp. > > > > 4. Add Suggested-by: David Hildenbrand <david@xxxxxxxxxx>. > > > > > > Oops, looks you killed my suggested-by tag now. Since it's still > > > pretty much my diff spliced with David's suggestions, maybe > > > Co-developed-by sounds more appropriate. Even more so goes for > > > the second patch getting rid of __netmem_clear_lsb(). > > > > > > Looks fine, just one comment below. > > > > > > ...> diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c > > > > index 100b75ab1e64..34634552cf74 100644 > > > > --- a/io_uring/zcrx.c > > > > +++ b/io_uring/zcrx.c > > > > @@ -444,6 +444,7 @@ static int io_zcrx_create_area(struct io_zcrx_ifq *ifq, > > > > area->freelist[i] = i; > > > > atomic_set(&area->user_refs[i], 0); > > > > niov->type = NET_IOV_IOURING; > > > > + niov->pp = NULL; > > > > > > It's zero initialised, you don't need it. > > > > > > > This may be my bad since I said we should check if it's 0 initialized. > > > > It looks like on the devmem side as well we kvmalloc_array the niovs, > > and if I'm checking through the helpers right, kvmalloc_array does > > 0-initialize indeed. > > I wouldn't rely on that, it's just for zcrx I do: > > kvmalloc_array(..., GFP_KERNEL | __GFP_ZERO); For net_devmem_bind_dmabuf(), __GFP_ZERO will add bigger overhead than just assignment, 'niov->pp = NULL'. I'd like to ask you if you are still good with __GFP_ZERO overhead before going ahead. Byungchul > > -- > Pavel Begunkov
