Work around 32 bits epoll_wait() bug

[bnoordhuis]

A bug in kernels < 2.6.37 makes timeouts larger than LONG_MAX /
CONFIG_HZ (approx. 30 minutes with CONFIG_HZ=1200) effectively
infinite on 32 bits architectures. The magic number is the same
constant used by libuv, see commit libuv/libuv@d1b5008e76.

If you don't care about kernels that old, I'm a-okay with closing this. No hard feelings.

[Thomasdezeeuw]

Officially we don't support kernels that old. That said since the change is rather small I would be OK with it it. @carllerche, @kleimkuhler what are your thoughts?

@bnoordhuis to fix the CI you can use the following patch. It's complaining about cfg! not being constant.

diff --git a/src/sys/unix/selector/epoll.rs b/src/sys/unix/selector/epoll.rs
index eb7cbe3..6955848 100644
--- a/src/sys/unix/selector/epoll.rs
+++ b/src/sys/unix/selector/epoll.rs
@@ -49,11 +49,11 @@ impl Selector {
         // A bug in kernels < 2.6.37 makes timeouts larger than LONG_MAX / CONFIG_HZ
         // (approx. 30 minutes with CONFIG_HZ=1200) effectively infinite on 32 bits
         // architectures. The magic number is the same constant used by libuv.
-        const MAX_SAFE_TIMEOUT: u128 = if cfg!(target_pointer_width = "32") {
-            1789569
-        } else {
-            libc::c_int::max_value() as u128
-        };
+        #[cfg(target_pointer_width = "32")]
+        const MAX_SAFE_TIMEOUT: u128 = 1789569;
+        #[cfg(not(target_pointer_width = "32"))]
+        const MAX_SAFE_TIMEOUT: u128 = libc::c_int::max_value() as u128;
 
         let timeout = timeout
             .map(|to| cmp::min(to.as_millis(), MAX_SAFE_TIMEOUT) as libc::c_int)

[carllerche]

👍

[Thomasdezeeuw]

Thanks @bnoordhuis

[Thomasdezeeuw]

@carllerche is this something we want to backport to v0.6?