On Tue, Jun 25, 2019 at 9:39 AM Florian Weimer <fweimer@xxxxxxxxxx> wrote: > > * Thomas Gleixner: > > > On Tue, 25 Jun 2019, Florian Weimer wrote: > >> We're trying to create portable binaries which use VSYSCALL on older > >> kernels (to avoid performance regressions), but gracefully degrade to > >> full system calls on kernels which do not have VSYSCALL support compiled > >> in (or disabled at boot). > >> > >> For technical reasons, we cannot use vDSO fallback. Trying vDSO first > >> and only then use VSYSCALL is the way this has been tackled in the past, > >> which is why this userspace ABI breakage goes generally unnoticed. But > >> we don't have a dynamic linker in our scenario. > > > > I'm not following. On newer kernels which usually have vsyscall disabled > > you need to use real syscalls anyway, so why are you so worried about > > performance on older kernels. That doesn't make sense. > > We want binaries that run fast on VSYSCALL kernels, but can fall back to > full system calls on kernels that do not have them (instead of > crashing). Define "VSYSCALL kernels." On any remotely recent kernel (*all* new kernels and all kernels for the last several years that haven't specifically requested vsyscall=native), using vsyscalls is much, much slower than just doing syscalls. I know a way you can tell whether vsyscalls are fast, but it's unreliable, and I'm disinclined to suggest it. There are also at least two pending patch series that will interfere. > > We could parse the vDSO and prefer the functions found there, but this > is for the statically linked case. We currently do not have a (minimal) > dynamic loader there in that version of the code base, so that doesn't > really work for us. Is anything preventing you from adding a vDSO parser? I wrote one just for this type of use: $ wc -l tools/testing/selftests/vDSO/parse_vdso.c 269 tools/testing/selftests/vDSO/parse_vdso.c (289 lines includes quite a bit of comment.) $ head -n8 tools/testing/selftests/vDSO/parse_vdso.c /* * parse_vdso.c: Linux reference vDSO parser * Written by Andrew Lutomirski, 2011-2014. * * This code is meant to be linked in to various programs that run on Linux. * As such, it is available with as few restrictions as possible. This file * is licensed under the Creative Commons Zero License, version 1.0, * available at http://creativecommons.org/publicdomain/zero/1.0/legalcode If this license is too restrictive for you, I could probably be convinced to relicense it, I'd be surprised :) In hindsight, I kind of wish I'd used MIT instead, since the Go runtime took advantage of the CC0 license to import it without attribution *and* break it quite badly in the process. IMO the correct solution is to parse the vDSO and, if that fails, to use plain syscalls as a fallback. You should not ship anything that uses a vsyscall under any circumstances, unless you need the last ounce of performance on that one ancient version of OpenSuSE that crashes if the vDSO is enabled.
