From mboxrd@z Thu Jan 1 00:00:00 1970 From: Alexei Starovoitov Subject: Re: [PATCH bpf-next 1/3] bpf: add bpf queue map Date: Thu, 9 Aug 2018 09:23:30 -0700 Message-ID: <20180809162328.rnyxihpiim2wz5dk@ast-mbp> References: <153356387977.6981.12236150594041620482.stgit@kernel> <153356390770.6981.4228793745105954649.stgit@kernel> <20180807144226.pmalhe3mjvc3a45y@ast-mbp.dhcp.thefacebook.com> <20180809044806.bbvzyp3sqh6orcnl@ast-mbp> <867e95e1-cb45-54c8-ce66-d3e49161d5e6@polito.it> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Daniel Borkmann , Alexei Starovoitov , netdev@vger.kernel.org To: Mauricio Vasquez Return-path: Received: from mail-pl0-f65.google.com ([209.85.160.65]:36150 "EHLO mail-pl0-f65.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1730419AbeHIStK (ORCPT ); Thu, 9 Aug 2018 14:49:10 -0400 Received: by mail-pl0-f65.google.com with SMTP id e11-v6so2764255plb.3 for ; Thu, 09 Aug 2018 09:23:33 -0700 (PDT) Content-Disposition: inline In-Reply-To: <867e95e1-cb45-54c8-ce66-d3e49161d5e6@polito.it> Sender: netdev-owner@vger.kernel.org List-ID: On Thu, Aug 09, 2018 at 09:51:49AM -0500, Mauricio Vasquez wrote: > > > Agree that existing ops are not the right alias, but deferring to user > > space as inline function also doesn't really seem like a good fit, imho, > > so I'd prefer rather to have something native. (Aside from that, the > > above inline bpf_pop() would also race between CPUs.) > > I think we should have push/pop/peek syscalls as well, having a bpf_pop() > that is race prone would create problems. Users expected maps operations to > be safe, so having one that is not will confuse them. agree the races are not acceptable. How about a mixed solution: - introduce bpf_push/pop/peak helpers that programs will use, so they don't need to pass useless key=NULL - introduce map->ops->lookup_and_delete and map->ops->lookup_or_init that prog-side helpers can use and syscall has 1-1 mapping for Native lookup_or_init() helper for programs and syscall is badly missing. Most of the bcc scripts use it and bcc has a racy workaround. Similarly lookup_and_delete() syscall is 1-1 to pop() for stack/queue and useful for regular hash maps. At the end for stack/queue map the programs will use: int bpf_push(map, value); value_or_null = bpf_pop(map); // guaranteed non-racy for multi-cpu value_or_null = bpf_peak(map); // racy if 2+ cpus doing it from syscall: bpf_map_lookup_elem(map, NULL, &value); // returns top of stack bpf_map_lookup_and_delete_elem(map, NULL, &value); // returns top and deletes top atomically bpf_map_update_elem(map, NULL, &value); // pushes new value into stack atomically Eventually hash and other maps will implement bpf_map_lookup_and_delete() for both bpf progs and syscall. The main point that prog-side api doesn't have to match 1-1 to syscall-side, since they're different enough already. Like lookup_or_init() is badly needed for programs, but unnecessary for syscall. Thoughts?