Den fre 10 aug. 2018 kl 12:18 skrev Jesper Dangaard Brouer <bro...@redhat.com>: > > On Fri, 10 Aug 2018 11:28:02 +0200 > Björn Töpel <bjorn.to...@gmail.com> wrote: > > > From: Björn Töpel <bjorn.to...@intel.com> > > > > This reverts commit 36e0f12bbfd3016f495904b35e41c5711707509f. > > > > The reverted commit adds a WARN to check against NULL entries in the > > mem_id_ht rhashtable. Any kernel path implementing the XDP (generic or > > driver) fast path is required to make a paired > > xdp_rxq_info_reg/xdp_rxq_info_unreg call for proper function. In > > addition, a driver using a different allocation scheme than the > > default MEM_TYPE_PAGE_SHARED is required to additionally call > > xdp_rxq_info_reg_mem_model. > > > > For MEM_TYPE_ZERO_COPY, an xdp_rxq_info_reg_mem_model call ensures > > that the mem_id_ht rhashtable has a properly inserted allocator id. If > > not, this would be a driver bug. A NULL pointer kernel OOPS is > > preferred to the WARN. > > Acked-by: Jesper Dangaard Brouer <bro...@redhat.com> > > As a comment says in the code: /* NB! Only valid from an xdp_buff! */ > Which is (currently) guarded by the return/exit in convert_to_xdp_frame(). > > This means that this code path can only be invoked while the driver is > still running under the RX NAPI process. Thus, there is no chance that > the allocator-id is gone (via calling xdp_rxq_info_unreg) for this code > path. > > But I really hope we at somepoint can convert a MEM_TYPE_ZERO_COPY into > a form of xdp_frame, that can travel further into the redirect-core. > In which case, we likely need to handle the NULL case (but also need > other code to handle what to do with the memory backing the frame) > > (I'm my vision here:) > > I really dislike that the current Zero-Copy mode steal ALL packets, > when ZC is enabled on a RX-queue. This is not better than the existing > bypass solutions, which have ugly ways of re-injecting packet back into > the network stack. With the integration with XDP, we have the > flexibility of selecting frames, that we don't want to be "bypassed" > into AF_XDP, and want the kernel process these. (The most common > use-case is letting the kernel handle the arptable). IHMO this is what > will/would make AF_XDP superior to other bypass solutions. > >
Thanks for putting your visions/ideas here! I agree with both of your last sections, and this is what we're working towards. AF_XDP ZC has to play nicer with XDP. The current (well, the soon-to-be-published [1] ;-)) ZC scheme is just a first step, and should be seen as a starting point so people can start playing using AF_XDP. Jakub also mentioned these issues a couple of threads ago, so there are definitely more people feeling the ZC allocator pains. Mid-term a sophisticated/proper and generic (for inter-driver reuse) ZC allocator is needed; Converting xdp_buffs to xdp_frames cheaply for multi-CPU completion, and hopefully dito for the XDP_PASS/kernel stack path. But let's start with something simple that works, and take it from there. Björn [1] WIP: https://github.com/bjoto/linux/tree/af-xdp-i40e-zc > > Suggested-by: Jesper Dangaard Brouer <bro...@redhat.com> > > Signed-off-by: Björn Töpel <bjorn.to...@intel.com> > > --- > > net/core/xdp.c | 3 +-- > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > diff --git a/net/core/xdp.c b/net/core/xdp.c > > index 6771f1855b96..9d1f22072d5d 100644 > > --- a/net/core/xdp.c > > +++ b/net/core/xdp.c > > @@ -345,8 +345,7 @@ static void __xdp_return(void *data, struct > > xdp_mem_info *mem, bool napi_direct, > > rcu_read_lock(); > > /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() > > */ > > xa = rhashtable_lookup(mem_id_ht, &mem->id, > > mem_id_rht_params); > > - if (!WARN_ON_ONCE(!xa)) > > - xa->zc_alloc->free(xa->zc_alloc, handle); > > + xa->zc_alloc->free(xa->zc_alloc, handle); > > rcu_read_unlock(); > > default: > > /* Not possible, checked in xdp_rxq_info_reg_mem_model() */ > > > > -- > Best regards, > Jesper Dangaard Brouer > MSc.CS, Principal Kernel Engineer at Red Hat > LinkedIn: http://www.linkedin.com/in/brouer
