While scanning the [email protected] mailing list, I've learned time(2) may return the wrong value in the first 1 to 2.5 ms of every second. While I'm not sure if the Date: response header matters to anyone, returning the correct time seems prudent.
Link: https://lore.kernel.org/git/[email protected]/ Link: https://inbox.sourceware.org/libc-alpha/[email protected]/T/ Link: https://sourceware.org/bugzilla/show_bug.cgi?id=30200 --- ext/unicorn_http/httpdate.c | 21 ++++++++++++++++----- 1 file changed, 16 insertions(+), 5 deletions(-) diff --git a/ext/unicorn_http/httpdate.c b/ext/unicorn_http/httpdate.c index 3f512dd..27a8f51 100644 --- a/ext/unicorn_http/httpdate.c +++ b/ext/unicorn_http/httpdate.c @@ -1,5 +1,5 @@ #include <ruby.h> -#include <time.h> +#include <sys/time.h> #include <stdio.h> static const size_t buf_capa = sizeof("Thu, 01 Jan 1970 00:00:00 GMT"); @@ -43,13 +43,24 @@ static struct tm * my_gmtime_r(time_t *now, struct tm *tm) static VALUE httpdate(VALUE self) { static time_t last; - time_t now = time(NULL); /* not a syscall on modern 64-bit systems */ + struct timeval now; struct tm tm; - if (last == now) + /* + * Favor gettimeofday(2) over time(2), as the latter can return the + * wrong value in the first 1 .. 2.5 ms of every second(!) + * + * https://lore.kernel.org/git/[email protected]/ + * https://inbox.sourceware.org/libc-alpha/[email protected]/T/ + * https://sourceware.org/bugzilla/show_bug.cgi?id=30200 + */ + if (gettimeofday(&now, NULL)) + rb_sys_fail("gettimeofday"); + + if (last == now.tv_sec) return buf; - last = now; - gmtime_r(&now, &tm); + last = now.tv_sec; + gmtime_r(&now.tv_sec, &tm); /* we can make this thread-safe later if our Ruby loses the GVL */ snprintf(buf_ptr, buf_capa,
