CURLOPT_NOSIGNAL does not work everywhere

[martinus]

I did this

I use the curl_easy interfaces in 2 threads, and immediately call curl_easy_setopt(curl, CURLOPT_NOSIGNAL);. Eventually, in curl_easy_cleanup the instances are destroyed, where unfortunately sigaction(SIGPIPE, &action, NULL); is called and in our case due to race condition this keeps the SIG_IGN installed if both destructions happen at almost the same time.

Here is the relevant stacktrace:

#0  0xfffffd7fff082824 in sigaction () from /lib/64/libc.so.1
#1  0x000000000095837a in sigpipe_ignore (ig=ig@entry=0xfffffd7ffe6001e0, data=<optimized out>) at sigpipe.h:51
#2  0x0000000000958ee9 in sigpipe_ignore (data=<optimized out>, ig=0xfffffd7ffe6001e0) at conncache.c:604
#3  Curl_conncache_close_all_connections (connc=connc@entry=0x12d0738) at conncache.c:597
#4  0x0000000000952220 in curl_multi_cleanup (multi=0x12d0640) at multi.c:2219
#5  0x000000000096ff8d in Curl_close (data=data@entry=0x12c6f70) at url.c:326
#6  0x000000000094e797 in curl_easy_cleanup (data=0x12c6f70) at easy.c:832

I expected the following

sigaction()should never be called when we use curl_easy_setopt(curl, CURLOPT_NOSIGNAL);

curl/libcurl version

curl-7.61.1

operating system

Solaris, Linux

[martinus]

I have a small patch that seems to work for us:

--- lib/easy.c	2018-10-15 10:43:12.715335746 +0200
+++ lib/easy.c	2018-10-15 10:43:22.039433726 +0200
@@ -761,6 +761,8 @@
   /* Copy the MAXCONNECTS option to the multi handle */
   curl_multi_setopt(multi, CURLMOPT_MAXCONNECTS, data->set.maxconnects);
 
+  multi->conn_cache.closure_handle->set.no_signal = data->set.no_signal;
+
   mcode = curl_multi_add_handle(multi, data);
   if(mcode) {
     curl_multi_cleanup(multi);

I'm not sure though if that is the best place to do this.

[bagder]

Thanks for the report, I think you're entirely correct and I'll give your suggested a fix a closer look...

[bagder]

@martinus: I just put the assignment in another place, next to two other options that also get "inherited" by the closure handle. I'll appreciate if you can just verify that this also works for you. See #3147!

[martinus]

I've just tried this your fix and it works for us. Thanks!

[bagder]

Excellent, thanks for confirming!