The mutex of NetAccept::action_->continuation is optional when the EVENT_ERROR event is called back.

[oknet]

In general, NetAccept::action_->continuation is a type of ProtocolProbeSessionAccept object.

The mutex of ProtocolProbeSessionAccept is NULL to allow parallel accepts.

Resolve issue #4726.

[scw00]

How to process EVENT_ERROR in ProtocolProbeSessionAccept.

It still crash .

[oknet]

@scw00

If errno returns from accept(2) neither EAGAIN nor 'EINTR`, it means an error that cannot be recovered.

In the situation, we must abort the ATS process, and output the errno into log file to help system admin find the cause of the problem.

[oknet]

@scw00 @masaori335

I guess the value of errno might be:

• EMFILE (24)
  • The per-process limit of open file descriptors has been reached.
• ENFILE (23)
  • The system limit on the total number of open files has been reached.
• ENOMEM (12)
  • Not enough free memory. This often means that the memory allocation is limited by the socket buffer limits, not by the system memory.
• EBADF (9)
  • The descriptor is invalid.

[scw00]

+1

[scw00]

Thanks for your explain.

[zwoop]

So, do we think this still fixes the crashes in #4726 ?

[scw00]

Yep, we need to catch the errno to figure out what happen during crashing. Some code can not be recover like ENFILE 、 EMFILE and ENOMEM because it is memory leaking.

[oknet]

@zwoop With this PR, ATS still crash (ink_abort) on L190 of ProtocolProbeSessionAccept.cc.

190   ink_abort("Protocol probe received a fatal error: errno = %d", -((int)(intptr_t)data));

We will get errno from the abort message and find out the cause of the problem.

[bryancall]

Cherry picked to 8.1.0