cygrunsrv + sshd + rsync = 20 times too slow -- throttled?

[Jon Turney]

On 18/09/2021 12:35, Takashi Yano wrote:
> On Sat, 18 Sep 2021 07:12:56 -0400
> Ken Brown wrote:
>>> It seems that SIGPIPE is handled differently. I guess
>>> SIGPIPE is handled by SIG_IGN if the terminal started
>>> from xwin-xdg-menu, otherwize, it is handled by SIG_DFL.
>>
>> Yes, I remember now that the same issue came up a couple years ago:
>>
>>     https://cygwin.com/pipermail/cygwin/2019-August/242060.html

Oops.  Somehow I missed the conclusion in that thread.

>>> Ah, the following patch for xwin-xdg-menu may be the
>>> right thing.
>>>
>>> --- execute.c.orig	2021-06-23 23:59:37.000000000 +0900
>>> +++ execute.c	2021-09-18 16:53:52.144248600 +0900
>>> @@ -76,6 +76,7 @@
>>>        {
>>>            struct rlimit rl;
>>>            unsigned int fd;
>>> +        int sig;
>>>    
>>>            /* dup write end of pipes onto stderr and stdout */
>>>            close(STDOUT_FILENO);
>>> @@ -89,6 +90,10 @@
>>>            for (fd = STDERR_FILENO + 1; fd < rl.rlim_cur; fd++)
>>>                close(fd);
>>>    
>>> +        /* Set all signal handlers to SIG_DFL. */
>>> +        for (sig = 1; sig < NSIG; sig++)
>>> +            signal(sig, SIG_DFL);
>>> +
>>>            /* Disassociate any TTYs */
>>>            setsid();
>>

Thanks for the patch!

>> This may be unnecessarily drastic, given the analysis in my message cited above.
> 
> It seems resetting handler only for SIGPIPE is enough according to
> the message you posted two years ago. Thanks.

I've applied Takashi-san's patch as is, and rebuilt the xwin-xdg-menu 
package.

I think NSIG is sufficiently small that it's worth doing this for all 
signals, just in case we ever ignore some other signal.

I probably need to audit that code more carefully for anything else it's 
leaking into the child that it shouldn't be.