> -----Original Message----- > From: papito favaro [mailto:papito.favaro@xxxxxxxxx] > Sent: Monday, February 25, 2019 10:11 PM > To: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx> > Cc: xdp-newbies@xxxxxxxxxxxxxxx; Karlsson, Magnus <magnus.karlsson@xxxxxxxxx>; Björn Töpel <bjorn.topel@xxxxxxxxx>; Maciej Fijałkowski <maciejromanfijalkowski@xxxxxxxxx>; Jesper Dangaard Brouer <brouer@xxxxxxxxxx> > Subject: Re: AF_XDP xdp_umem_reg.headroom XDP_ZEROCOPY > > Yes, would be nice that the documentation would mention this facts (like copy vs zerocopy differences) I also had an issue with the rx-path. In zerocopy is required that the fill ring be "initialized" (with indexes) before the setup of rx ring. > In copy mode this can be done afterwards but in zerocopy must be done before. also not sure minimum (valid) sizes for the rings. > sorry don't have a ixgbe to test this right now. You are supposed to be able to insert entries into the fill ring at any point in time independent on copy mode or zero-copy mode. That it did not work was a bug in both the i40e and ixgbe drivers. Links to patches below. As for documentation, yes it needs to be improved :-). Anything else you would like to see in there? i40e: https://lore.kernel.org/netdev/1548770597-16141-1-git-send-email-magnus.karlsson@xxxxxxxxx/. ixgbe: https://lore.kernel.org/netdev/1548770630-16189-1-git-send-email-magnus.karlsson@xxxxxxxxx/ Thanks: Magnus > anyway thanks for your time > > Il giorno lun 25 feb 2019 alle ore 21:10 Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx> ha scritto: > > > > On Mon, 25 Feb 2019 10:16:15 +0100, papito favaro wrote: > > > Hi all, > > > I woulld like to understand if the behaviour i'm getting is correct. > > > > > > When I configure the XDP_UMEM_REG in zerocopy mode with a headroom > > > == > > > 0 on a i40e, > > > the driver (i guess) actually sets a headroom of 256. > > > if i set a headroom == 256 the driver sets a headroom of 512. > > > > > > If this behaviour is correct, would be nice that the xdp_umem_reg > > > struct after the setsockopt call would contain the actual configured > > > headroom. > > > > Kernel's XDP may require up to 256B (XDP_PACKET_HEADROOM) of headroom > > this is for calls to xdp_adjust_head() which may want to prepend data. > > The AF_XDP headroom is extra, beyond that. > > > > Perhaps we should document that as expected behaviour? (Does it hold > > for ixgbe?)
|