linux: epoll_pwait() syscall wrapper is probably wrong

[bnoordhuis]

It currently passes sizeof(sigset_t) to the kernel but that's wrong on at least x86_64. A patch:

diff --git a/src/unix/linux-syscalls.c b/src/unix/linux-syscalls.c
index 06cc594..ee90b84 100644
--- a/src/unix/linux-syscalls.c
+++ b/src/unix/linux-syscalls.c
@@ -21,6 +21,7 @@

 #include "linux-syscalls.h"
 #include <unistd.h>
+#include <signal.h>
 #include <sys/syscall.h>
 #include <sys/types.h>
 #include <errno.h>
@@ -298,7 +299,7 @@ int uv__epoll_pwait(int epfd,
                  nevents,
                  timeout,
                  sigmask,
-                 sizeof(*sigmask));
+                 _NSIG / 8);
 #else
   return errno = ENOSYS, -1;
 #endif

_NSIG / 8 is 8, whereas sizeof(*sigmask) is 128 and makes the system call fail with EINVAL. I'm reasonably sure that the above is correct (it's how glibc implements epoll_wait) but it perhaps needs some additional verification.

[saghul]

IIRC this was tested both by the reporter and by @lucab on Debian's AArch64 build: joyent/libuv#1443 any chance things should be handled differently depending on the architecture here?

[saghul]

Ah, wait, in libuv we always pass NULL to the call uv__epoll_pwait, that's why it never surfaced. LGTM if you are sure it works. But otherwise we can fix it later too :-)

[saghul]

musl libc also does it, so land it at will.

[bnoordhuis]

Cheers, landed in v0.10 in b705b53. What's the current strategy for up-merges? At will or after a release?

[saghul]

Usually after a release, but if you need it feel free to up-merge.

[bnoordhuis]

No rush, I'll wait for the next release.