[nginx] Fix valgrind errors in mail tests

[Maxim Dounin]

Hello!

On Mon, May 13, 2024 at 03:43:58PM +1000, Robert Mueller wrote:

> # HG changeset patch
> # User Rob Mueller <robm at fastmailteam.com>
> Fix valgrind errors in mail tests
> 
> Applying this fix allows nginx to be run under valgrind and all
> mail_* tests to pass with no valgrand issues detected
> ---
>  src/event/modules/ngx_epoll_module.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/src/event/modules/ngx_epoll_module.c b/src/event/modules/ngx_epoll_module.c
> index 98e3ce7c8..c89a56d95 100644
> --- a/src/event/modules/ngx_epoll_module.c
> +++ b/src/event/modules/ngx_epoll_module.c
> @@ -474,6 +474,7 @@ ngx_epoll_test_rdhup(ngx_cycle_t *cycle)
>      }
>  
>      ee.events = EPOLLET|EPOLLIN|EPOLLRDHUP;
> +    ee.data.u64 = 0;
>  
>      if (epoll_ctl(ep, EPOLL_CTL_ADD, s[0], &ee) == -1) {
>          ngx_log_error(NGX_LOG_ALERT, cycle->log, ngx_errno,

There were reports about similar complaints from Valgrind on 
32-bit platforms a while ago, but for normal epoll_ctl() calls, 
where only data.ptr is set:

https://mailman.nginx.org/pipermail/nginx-devel/2013-July/003892.html

Given this is a hot path, and there are multiple such calls, 
suggested approach was to do additional unneeded initialization 
only if NGX_VALGRIND is defined (which is already used to ignore 
sigaction() results under Valgrind, see 5244:593d344999f5).  This 
was never implemented though.

Still, I cannot reproduce it with recent Valgrind versions. It 
looks like it is fixed in the Valgrind itself, and now it only 
checks ee.events, but not ee.data and padding:

https://bugs.kde.org/show_bug.cgi?id=422623

The fix is available starting with Valgrind 4.17.0, which is not 
exactly new.  In particular, the only supported Ubuntu version 
without the fix is 20.04 LTS, while Ubuntu 22.04 LTS packages 
contain Valgrind 3.18.0, with the fix.

OTOH, I'm not really seeing this on Ubuntu 20.04 LTS / amd64 with 
Valgrind 3.15.0 either, not sure why, probably due to some code 
generation nuances.  But on Ubuntu 20.04 LTS I'm seeing another 
Valgrind bug ("Source and destination overlap in memcpy_chk()" for 
memmove(), https://bugs.kde.org/show_bug.cgi?id=402833).

I was only able to reproduce this on Ubuntu 18.04 / i386 (with 
Valgrind 3.13.0, along with complaints about normal epoll_ctl() 
calls.

Overall, this patch is probably no longer needed, as the proper 
fix is already in Valgrind, and there are other bugs in old 
Valgrind versions anyway.

-- 
Maxim Dounin
http://mdounin.ru/