From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id 9E73AC433EF for ; Fri, 11 Mar 2022 07:26:04 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1345665AbiCKH1F (ORCPT ); Fri, 11 Mar 2022 02:27:05 -0500 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:45760 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1347230AbiCKH07 (ORCPT ); Fri, 11 Mar 2022 02:26:59 -0500 Received: from mail-pl1-x644.google.com (mail-pl1-x644.google.com [IPv6:2607:f8b0:4864:20::644]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 559DB1B7551; Thu, 10 Mar 2022 23:25:55 -0800 (PST) Received: by mail-pl1-x644.google.com with SMTP id e2so7000954pls.10; Thu, 10 Mar 2022 23:25:55 -0800 (PST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20210112; h=date:from:to:cc:subject:message-id:references:mime-version :content-disposition:content-transfer-encoding:in-reply-to; bh=tywLyN8StSWR/Rxw81cSTsVeOZKqkeuokSX2+f8M3Xk=; b=N8rr+fRBx99vOKKNJxuZfMTnIxccg8CDp5K4KQbRKdaX1TS/oUu2mrUvdH+oB6C5QP TcpvMfpLVxMQBgCYJoRopN5AH5jiJlF4ZNuq0pMKwn/huaKZMnv7Lb0DPRe3gsgHq1IP rhB3Rht54BuPm23FUO9JZI8EdHtuDmKbd+vmS+Lj18KhKpb6U9iO6350qvad/BG5yr1r OmGU7j2UvY4If7Ni0eAoskLUu93gHF9Ke9UWfjcWi1fEPnE03u/oETJuSJgmpyZfO0sa xqOlJ4RSXBcUSb8fmHbUqB84iN1c3+RH12Aqs5ZYo0VYh3gJrMOlInFKlYqUspQDjgp7 mDFQ== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=x-gm-message-state:date:from:to:cc:subject:message-id:references :mime-version:content-disposition:content-transfer-encoding :in-reply-to; bh=tywLyN8StSWR/Rxw81cSTsVeOZKqkeuokSX2+f8M3Xk=; b=w7dEPRMqnm5qhYA7hupV8ilVShG1+Z20zcSkbTaXidj7j1IWpgaiOEurs78XCqw3xF L4iXk/hPSfkdhXjGCoL1Hmnb1jOjlzxew7/hCD9RwaNUFqlcYoJbw9NQaXzmh5om303l PXcXsAhAtfc/t3J8QbEAYG20jDGnu1Uj9i0szxccgeS9OEAyzo5oi36dFddtH4+fhmxm 0D6CIjp/jJKX4aCVxYzVSm3Uo/xq/jikqniT7sP21NboutfV51tdpgWOilixbwIPuZcB MYLife3C1BxAZEpslo/PqpRBKC09IOor2mzJuX/6QFUjsay9hWmP2x5FjJ69xhADdrgv P5zA== X-Gm-Message-State: AOAM531CWdg2uMfG44a6UNM+voTRhlUNPp0zfqF7at+d/WS4vGtNPTHO /TPL8Gz0SVcUTsoRzefd9Ko= X-Google-Smtp-Source: ABdhPJxRTM2sGc9O6tLKMktrgrXnRitaqhRTB9rPYN5udtphtS47UedIhk3KZt4Vh0tGGkmW+q7Fjw== X-Received: by 2002:a17:90a:a385:b0:1b9:cfb8:de07 with SMTP id x5-20020a17090aa38500b001b9cfb8de07mr9207794pjp.162.1646983554715; Thu, 10 Mar 2022 23:25:54 -0800 (PST) Received: from localhost ([112.79.142.206]) by smtp.gmail.com with ESMTPSA id s14-20020a056a0008ce00b004f66dcd4f1csm10402788pfu.32.2022.03.10.23.25.53 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Thu, 10 Mar 2022 23:25:54 -0800 (PST) Date: Fri, 11 Mar 2022 12:55:45 +0530 From: Kumar Kartikeya Dwivedi To: Alexei Starovoitov Cc: Toke =?utf-8?Q?H=C3=B8iland-J=C3=B8rgensen?= , Andrii Nakryiko , bpf , Alexei Starovoitov , Andrii Nakryiko , Daniel Borkmann , Martin KaFai Lau , Jesper Dangaard Brouer , Lorenzo Bianconi , John Fastabend , Jakub Kicinski , Lorenz Bauer , Networking Subject: Re: [PATCH bpf-next v1 0/5] Introduce bpf_packet_pointer helper Message-ID: <20220311072545.deeifraq4u74dagb@apollo> References: <20220306234311.452206-1-memxor@gmail.com> <20220308070828.4tjiuvvyqwmhru6a@apollo.legion> <87lexky33s.fsf@toke.dk> <87bkydxu59.fsf@toke.dk> MIME-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Disposition: inline Content-Transfer-Encoding: 8bit In-Reply-To: Precedence: bulk List-ID: X-Mailing-List: netdev@vger.kernel.org On Fri, Mar 11, 2022 at 05:05:24AM IST, Alexei Starovoitov wrote: > On Thu, Mar 10, 2022 at 3:30 PM Toke Høiland-Jørgensen wrote: > > > > Andrii Nakryiko writes: > > > > > On Tue, Mar 8, 2022 at 5:40 AM Toke Høiland-Jørgensen wrote: > > >> > > >> Kumar Kartikeya Dwivedi writes: > > >> > > >> > On Tue, Mar 08, 2022 at 11:18:52AM IST, Andrii Nakryiko wrote: > > >> >> On Sun, Mar 6, 2022 at 3:43 PM Kumar Kartikeya Dwivedi wrote: > > >> >> > > > >> >> > Expose existing 'bpf_xdp_pointer' as a BPF helper named 'bpf_packet_pointer' > > >> >> > returning a packet pointer with a fixed immutable range. This can be useful to > > >> >> > enable DPA without having to use memcpy (currently the case in > > >> >> > bpf_xdp_load_bytes and bpf_xdp_store_bytes). > > >> >> > > > >> >> > The intended usage to read and write data for multi-buff XDP is: > > >> >> > > > >> >> > int err = 0; > > >> >> > char buf[N]; > > >> >> > > > >> >> > off &= 0xffff; > > >> >> > ptr = bpf_packet_pointer(ctx, off, sizeof(buf), &err); > > >> >> > if (unlikely(!ptr)) { > > >> >> > if (err < 0) > > >> >> > return XDP_ABORTED; > > >> >> > err = bpf_xdp_load_bytes(ctx, off, buf, sizeof(buf)); > > >> >> > if (err < 0) > > >> >> > return XDP_ABORTED; > > >> >> > ptr = buf; > > >> >> > } > > >> >> > ... > > >> >> > // Do some stores and loads in [ptr, ptr + N) region > > >> >> > ... > > >> >> > if (unlikely(ptr == buf)) { > > >> >> > err = bpf_xdp_store_bytes(ctx, off, buf, sizeof(buf)); > > >> >> > if (err < 0) > > >> >> > return XDP_ABORTED; > > >> >> > } > > >> >> > > > >> >> > Note that bpf_packet_pointer returns a PTR_TO_PACKET, not PTR_TO_MEM, because > > >> >> > these pointers need to be invalidated on clear_all_pkt_pointers invocation, and > > >> >> > it is also more meaningful to the user to see return value as R0=pkt. > > >> >> > > > >> >> > This series is meant to collect feedback on the approach, next version can > > >> >> > include a bpf_skb_pointer and exposing it as bpf_packet_pointer helper for TC > > >> >> > hooks, and explore not resetting range to zero on r0 += rX, instead check access > > >> >> > like check_mem_region_access (var_off + off < range), since there would be no > > >> >> > data_end to compare against and obtain a new range. > > >> >> > > > >> >> > The common name and func_id is supposed to allow writing generic code using > > >> >> > bpf_packet_pointer that works for both XDP and TC programs. > > >> >> > > > >> >> > Please see the individual patches for implementation details. > > >> >> > > > >> >> > > >> >> Joanne is working on a "bpf_dynptr" framework that will support > > >> >> exactly this feature, in addition to working with dynamically > > >> >> allocated memory, working with memory of statically unknown size (but > > >> >> safe and checked at runtime), etc. And all that within a generic > > >> >> common feature implemented uniformly within the verifier. E.g., it > > >> >> won't need any of the custom bits of logic added in patch #2 and #3. > > >> >> So I'm thinking that instead of custom-implementing a partial case of > > >> >> bpf_dynptr just for skb and xdp packets, let's maybe wait for dynptr > > >> >> and do it only once there? > > >> >> > > >> > > > >> > Interesting stuff, looking forward to it. > > >> > > > >> >> See also my ARG_CONSTANT comment. It seems like a pretty common thing > > >> >> where input constant is used to characterize some pointer returned > > >> >> from the helper (e.g., bpf_ringbuf_reserve() case), and we'll need > > >> >> that for bpf_dynptr for exactly this "give me direct access of N > > >> >> bytes, if possible" case. So improving/generalizing it now before > > >> >> dynptr lands makes a lot of sense, outside of bpf_packet_pointer() > > >> >> feature itself. > > >> > > > >> > No worries, we can continue the discussion in patch 1, I'll split out the arg > > >> > changes into a separate patch, and wait for dynptr to be posted before reworking > > >> > this. > > >> > > >> This does raise the question of what we do in the meantime, though? Your > > >> patch includes a change to bpf_xdp_{load,store}_bytes() which, if we're > > >> making it, really has to go in before those hit a release and become > > >> UAPI. > > >> > > >> One option would be to still make the change to those other helpers; > > >> they'd become a bit slower, but if we have a solution for that coming, > > >> that may be OK for a single release? WDYT? > > > > > > I must have missed important changes to bpf_xdp_{load,store}_bytes(). > > > Does anything change about its behavior? If there are some fixes > > > specific to those helpers, we should fix them as well as a separate > > > patch. My main objection is adding a bpf_packet_pointer() special case > > > when we have a generic mechanism in the works that will come this use > > > case (among other use cases). > > > > Well it's not a functional change per se, but Kartikeya's patch is > > removing an optimisation from bpf_xdp_{load_store}_bytes() (i.e., the > > use of the bpf_xdp_pointer()) in favour of making it available directly > > to BPF. So if we don't do that change before those helpers are > > finalised, we will end up either introducing a performance regression > > for code using those helpers, or being stuck with the bpf_xdp_pointer() > > use inside them even though it makes more sense to move it out to BPF. > > > > So the "safe" thing to do would do the change to the store/load helpers > > now, and get rid of the bpf_xdp_pointer() entirely until it can be > > introduced as a BPF helper in a generic way. Of course this depends on > > whether you consider performance regressions to be something to avoid, > > but this being XDP IMO we should :) > > I don't follow this logic. > Would you mean by "get rid of the bpf_xdp_pointer" ? > It's just an internal static function. > > Also I don't believe that this patch set and exposing > bpf_xdp_pointer as a helper actually gives measurable performance wins. > It looks quirky to me and hard to use. This is actually inspired from your idea to avoid memcpy when reading and writing to multi-buff XDP [0]. But instead of passing in the stack or mem pointer (as discussed in that thread), I let the user set it and detect it themselves, which makes the implementation simpler. I am sure accessing a few bytes directly is going to be faster than first memcpy'ing it to a local buffer, reading, and then possibly writing things back again using a memcpy, but I will be happy to come with some numbers when I respin this later, when Joanne posts the dynptr series. Ofcourse, we could just make return value PTR_TO_MEM even for the 'pass buf pointer' idea, but then we have to conservatively invalidate the pointer even if it points to stack buffer on clear_all_pkt_pointers. The current approach looked better to me. [0]: https://lore.kernel.org/bpf/CAADnVQKbrkOxfNoixUx-RLJEWULJLyhqjZ=M_X2cFG_APwNyCg@mail.gmail.com -- Kartikeya