From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jakub Kicinski Subject: Re: [PATCH bpf] xdp: add NULL pointer check in __xdp_return() Date: Mon, 23 Jul 2018 12:58:43 -0700 Message-ID: <20180723125843.7431eb3a@cakuba.netronome.com> References: <20180720160445.17475-1-ap420073@gmail.com> <20180720171821.ivtpo6obzx2v737c@kafai-mbp.dhcp.thefacebook.com> <20180720130538.10c22c61@cakuba.netronome.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable Cc: Daniel Borkmann , =?UTF-8?B?QmrDtnJuIFTDtnBlbA==?= , Jesper Dangaard Brouer , kafai@fb.com, ap420073@gmail.com, ast@kernel.org, Netdev , "Karlsson, Magnus" To: =?UTF-8?B?QmrDtnJuIFTDtnBlbA==?= Return-path: Received: from mail-qt0-f196.google.com ([209.85.216.196]:34942 "EHLO mail-qt0-f196.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S2387883AbeGWVBf (ORCPT ); Mon, 23 Jul 2018 17:01:35 -0400 Received: by mail-qt0-f196.google.com with SMTP id a5-v6so1906882qtp.2 for ; Mon, 23 Jul 2018 12:58:48 -0700 (PDT) In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: On Mon, 23 Jul 2018 11:39:36 +0200, Bj=C3=B6rn T=C3=B6pel wrote: > Den fre 20 juli 2018 kl 22:08 skrev Jakub Kicinski: > > On Fri, 20 Jul 2018 10:18:21 -0700, Martin KaFai Lau wrote: =20 > > > On Sat, Jul 21, 2018 at 01:04:45AM +0900, Taehee Yoo wrote: =20 > > > > rhashtable_lookup() can return NULL. so that NULL pointer > > > > check routine should be added. > > > > > > > > Fixes: 02b55e5657c3 ("xdp: add MEM_TYPE_ZERO_COPY") > > > > Signed-off-by: Taehee Yoo > > > > --- > > > > net/core/xdp.c | 3 ++- > > > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/net/core/xdp.c b/net/core/xdp.c > > > > index 9d1f220..1c12bc7 100644 > > > > --- a/net/core/xdp.c > > > > +++ b/net/core/xdp.c > > > > @@ -345,7 +345,8 @@ 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_mo= del() */ > > > > xa =3D rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rh= t_params); > > > > - xa->zc_alloc->free(xa->zc_alloc, handle); > > > > + if (xa) > > > > + xa->zc_alloc->free(xa->zc_alloc, handle); =20 > > > hmm...It is not clear to me the "!xa" case don't have to be handled? = =20 > > > > Actually I have a more fundamental question about this interface I've > > been meaning to ask. > > > > IIUC free() can happen on any CPU at any time, when whatever device, > > socket or CPU this got redirected to completed the TX. IOW there may > > be multiple producers. Drivers would need to create spin lock a'la the > > a9744f7ca200 ("xsk: fix potential race in SKB TX completion code") fix? > > =20 >=20 > Jakub, apologies for the slow response. I'm still in > "holiday/hammock&beer mode", but will be back in a week. :-P Ah, sorry to interrupt! :) > The idea with the xdp_return_* functions are that an xdp_buff and > xdp_frame can have custom allocations schemes. The difference beween > struct xdp_buff and struct xdp_frame is lifetime. The xdp_buff > lifetime is within the napi context, whereas xdp_frame can have a > lifetime longer/outside the napi context. E.g. for a XDP_REDIRECT > scenario an xdp_buff is converted to a xdp_frame. The conversion is > done in include/net/xdp.h:convert_to_xdp_frame. >=20 > Currently, the zero-copy MEM_TYPE_ZERO_COPY memtype can *only* be used > for xdp_buff, meaning that the lifetime is constrained to a napi > context. Further, given an xdp_buff with memtype MEM_TYPE_ZERO_COPY, > doing XDP_REDIRECT to a target that is *not* an AF_XDP socket would > mean converting the xdp_buff to an xdp_frame. The xdp_frame can then > be free'd on any CPU. >=20 > Note that the xsk_rcv* functions is always called from an napi > context, and therefore is using the xdp_return_buff calls. >=20 > To answer your question -- no, this fix is *not* needed, because the > xdp_buff napi constrained, and the xdp_buff will only be free'd on one > CPU. Oh, thanks, I missed the check in convert_to_xdp_frame(), so the only frames which can come back via the free path are out of the error path in __xsk_rcv_zc()? That path looks a little surprising too, isn't the expectation that if xdp_do_redirect() returns an error the driver retains the ownership of the buffer?=20 =20 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) { int err =3D xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len); if (err) { xdp_return_buff(xdp); xs->rx_dropped++; } return err; } This seems to call xdp_return_buff() *and* return an error. > > We need some form of internal kernel circulation which would be MPSC. > > I'm currently hacking up the XSK code to tell me whether the frame was > > consumed by the correct XSK, and always clone the frame otherwise > > (claiming to be the "traditional" MEM_TYPE_PAGE_ORDER0). > > > > I feel like I'm missing something about the code. Is redirect of > > ZC/UMEM frame outside the xsk not possible and the only returns we will > > see are from net/xdp/xsk.c? That would work, but I don't see such a > > check. Help would be appreciated. > > =20 >=20 > Right now, this is the case (refer to the TODO in > convert_to_xdp_frame), i.e. you cannot redirect an ZC/UMEM allocated > xdp_buff to a target that is not an xsk. This must, obviously, change > so that an xdp_buff (of MEM_TYPE_ZERO_COPY) can be converted to an > xdp_frame. The xdp_frame must be able to be free'd from multiple CPUs, > so here the a more sophisticated allocation scheme is required. >=20 > > Also the fact that XSK bufs can't be freed, only completed, adds to the > > pain of implementing AF_XDP, we'd certainly need some form of "give > > back the frame, but I may need it later" SPSC mechanism, otherwise > > driver writers will have tough time. Unless, again, I'm missing > > something about the code :) > > =20 >=20 > Yup, moving the recycling scheme from driver to "generic" is a good > idea! I need to finish up those i40e zerocopy patches first though... Interesting, FWIW I wasn't necessarily thinking about full recycling, although that would be the holy grail. Just a generic way of giving up buffers for example when user changes ring sizes or brings the device down. > (...and I'm very excited that you're doing nfp support for AF_XDP!!!) Thanks, I'm still way out in the weeds but it's interesting work :)