Netbsd forever #4793
Netbsd forever #4793
Conversation
On NetBSD this caused the libuserland compilation to fail. Closes ziglang#3719
* Use the correct versioned libc calls to avoid nasty surprises
* NetBSD is stricter than other OSs and doesn't allow mprotect to mark a non-accessible region as RW * Fix mprotect call over the whole stack, oops
The guard page size shouldn't be taken into account, pthread is only interested in the usable area.
On NetBSD this is needed to avoid crashes in pthread_join as the default value for the guard page size is not ignored even though a custom stack address is specified.
lib/std/os.zig
Outdated
| @@ -2542,6 +2542,17 @@ pub fn fstat(fd: fd_t) FStatError!Stat { | |||
| } | |||
| } | |||
|
|
|||
| if (std.Target.current.os.tag == .netbsd) { | |||
| switch (errno(system.__fstat50(fd, &stat))) { | |||
There was a problem hiding this comment.
lets share the error code paths with the generic posix ones below?
switch (errno(switch(std.Target.current.os.tag) {
.netbsd => system.__fstat50(fd, &stat),
else => system.fstat(fd, &stat),
})) {
....
}There was a problem hiding this comment.
I was thinking of something like this to avoid excessive nesting:
const fstat_symbol = switch (std.Target.current.os.tag) {
.netbsd => "__fstat50",
else => "fstat",
};
switch (errno(@field(system, fstat_symbol))) {
...
}
|
looking good so far! netbsd 9.0: |
|
Yeah that's expected, the event loop code expects evfilt_user to be available but that's not the case for NetBSD (nor OpenBSD afair). But again, even on FreeBSD that test compiles and hangs, so it's safe to say that only Linux is well supported at the moment. |
|
|
||
| break :mem mmap_slice; | ||
| } else { | ||
| // NetBSD mprotect is very strict and doesn't allow to "upgrade" |
There was a problem hiding this comment.
Is there any reason the NetBSD code path isn't used on every OS? Basically, you're dropping privileges instead of acquiring them later.
This should also be correct for grsec and similar hardened kernels, AFAIK.
There was a problem hiding this comment.
On Linux we can tell the kernel that the guard page is effectively empty and can avoid allocating memory for that. The savings are small but scale linearly with the number of threads we spawn.
This should also be correct for grsec and similar hardened kernels, AFAIK.
Hmm, that may be a problem...
Let me show some love for a really nice OS, NetBSD is the only BSD that manages to build Zig using LLVM from
pkg!What this patchset does:
Speaking of the event code, on NetBSD we have kqueue but without
EVFILT_USERso cross-thread signaling has to be done like it's done for Linux using a pipe (or a socketpair) per-thread.