Den ons 5 sep. 2018 kl 19:14 skrev Jakub Kicinski <jakub.kicin...@netronome.com>: > > On Tue, 4 Sep 2018 20:11:01 +0200, Björn Töpel wrote: > > From: Björn Töpel <bjorn.to...@intel.com> > > > > This series addresses an AF_XDP zero-copy issue that buffers passed > > from userspace to the kernel was leaked when the hardware descriptor > > ring was torn down. > > > > The patches fixes the i40e AF_XDP zero-copy implementation. > > > > Thanks to Jakub Kicinski for pointing this out! > > > > Some background for folks that don't know the details: A zero-copy > > capable driver picks buffers off the fill ring and places them on the > > hardware Rx ring to be completed at a later point when DMA is > > complete. Similar on the Tx side; The driver picks buffers off the Tx > > ring and places them on the Tx hardware ring. > > > > In the typical flow, the Rx buffer will be placed onto an Rx ring > > (completed to the user), and the Tx buffer will be placed on the > > completion ring to notify the user that the transfer is done. > > > > However, if the driver needs to tear down the hardware rings for some > > reason (interface goes down, reconfiguration and such), the userspace > > buffers cannot be leaked. They have to be reused or completed back to > > userspace. > > > > The implementation does the following: > > > > * Outstanding Tx descriptors will be passed to the completion > > ring. The Tx code has back-pressure mechanism in place, so that > > enough empty space in the completion ring is guaranteed. > > > > * Outstanding Rx descriptors are temporarily stored on a stash/reuse > > queue. The reuse queue is based on Jakub's RFC. When/if the HW rings > > comes up again, entries from the stash are used to re-populate the > > ring. > > > > * When AF_XDP ZC is enabled, disallow changing the number of hardware > > descriptors via ethtool. Otherwise, the size of the stash/reuse > > queue can grow unbounded. > > > > Going forward, introducing a "zero-copy allocator" analogous to Jesper > > Brouer's page pool would be a more robust and reuseable solution. > > > > Jakub: I've made a minor checkpatch-fix to your RFC, prior adding it > > into this series. > > Thanks for the fix! :) > > Out of curiosity, did checking the reuse queue have a noticeable impact > in your test (i.e. always using the _rq() helpers)? You seem to be > adding an indirect call, would that not be way worse on a retpoline > kernel?
Do you mean the indirection in __i40e_alloc_rx_buffers_zc (patch #3)? The indirect call is elided by the __always_inline -- without that retpoline took 2.5Mpps worth of Rx. :-( I'm only using the _rq helpers in the configuration/slow path, so the fast-path is unchanged. Björn
