I have tried running seastar's tests (http://seastar.io/) with valgrind, but some of the tests fail when they try to use io_pgetevents (syscall number 333 on x86_64) which is not implemented yet.
I have hit the same bug on Ubuntu 24.04 running valgrind version 3.22.0, and did not see a fix in git. Steps to reproduce: missing-syscall.c ```C #include <assert.h> #include <fcntl.h> #include <libaio.h> #include <unistd.h> int main() { const char *msg = "hello world\n"; struct iocb iocb = {}; struct io_event event; io_context_t ctx = 0; struct iocb *iocbp; int rc, fd; rc = io_setup(1, &ctx); assert(rc == 0); fd = open("test.txt", O_CREAT | O_RDWR, 0666); assert(fd >= 0); io_prep_pwrite(&iocb, fd, (void *)msg, 12, 0); iocbp = &iocb; rc = io_submit(ctx, 1, &iocbp); assert(rc == 1); rc = io_getevents(ctx, 1, 1, &event, NULL); assert(rc = 1); close(fd); io_destroy(ctx); } ``` ``` gcc -g -Og missing-syscall.c -o test-valgrind -laio valgrind ./test-valgrind ``` output: ``` ==39192== Memcheck, a memory error detector ==39192== Copyright (C) 2002-2022, and GNU GPL'd, by Julian Seward et al. ==39192== Using Valgrind-3.22.0 and LibVEX; rerun with -h for copyright info ==39192== Command: ./test-valgrind ==39192== --39192-- WARNING: unhandled amd64-linux syscall: 333 --39192-- You may be able to write your own handler. --39192-- Read the file README_MISSING_SYSCALL_OR_IOCTL. --39192-- Nevertheless we consider this a bug. Please report --39192-- it at http://valgrind.org/support/bug_reports.html. ==39192== ==39192== HEAP SUMMARY: ==39192== in use at exit: 0 bytes in 0 blocks ==39192== total heap usage: 0 allocs, 0 frees, 0 bytes allocated ==39192== ==39192== All heap blocks were freed -- no leaks are possible ==39192== ==39192== For lists of detected and suppressed errors, rerun with: -s ==39192== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0) ```
(In reply to David Sloan from comment #1) > I have hit the same bug on Ubuntu 24.04 running valgrind version 3.22.0, and > did not see a fix in git. It works for me on Fedora 41. But that's using sys_io_getevents rather than sys_io_pgetevents SYSCALL[5989,1](208) sys_io_getevents ( 139871270969344, 1, 1, 0x1ffefff850, 0x0 ) --> [async] ... SYSCALL[5989,1](208) ... [async] --> Success(0x1) It looks like the same interface but with an extra const struct __aio_sigset* usig argument.
Created attachment 179731 [details] initial patch I haven't tested it yet properly, but could you try this patch?
(In reply to Paul Floyd from comment #3) > Created attachment 179731 [details] > initial patch > > I haven't tested it yet properly, but could you try this patch? This works on my system, thanks! Tested on valgrind: commit f697142e15c44c2d3e3910ae436a6597e8782aea (HEAD -> master, origin/master, origin/HEAD)
I've written a test for this and cleaned up the error messages a bit - there were a few errors in the untested code. Buildbot says I need to improve the configure.ac test the the syscall. There is only the amd64 wrapper for the moment.
I couldn't see the syscall number for ppc. The value was wrong for arm64, now corrected. Added x86 s390 mips32, mips64 and riscv64 as well as amd64. commit 19c00de89e0fa21c62a5c0c0d1b1cc081f4ce00a (HEAD -> master, origin/master, origin/HEAD, bug420682) Author: Paul Floyd <pjfloyd@wanadoo.fr> Date: Fri Mar 28 21:45:15 2025 +0100 Bug 420682 - io_pgetevents is not supported