[TS-2187] failed assert `nr == sizeof(uint64_t)` in EventNotify::signal()

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.1.2
• Component/s: Core
• Assignee: Yunkai Zhang
• Reporter: Yunkai Zhang
• Created: 08/Sep/13 15:34
• Updated: 05/Dec/13 23:51
• Resolved: 11/Sep/13 07:20

Description

I have noticed two bug report on "failed assert `nr == sizeof(uint64_t)`" in EventNotify::signal():

[TrafficServer] using root directory '/usr/local/trafficserver-4.1.0'
FATAL: EventNotify.cc:73: failed assert `nr == sizeof(uint64_t)`
/usr/local/trafficserver-4.1.0/bin/traffic_server - STACK TRACE:
/usr/local/trafficserver-4.1.0/lib/libtsutil.so.4(+0x14b57)[0x2b71a2d84b57]
/usr/local/trafficserver-4.1.0/lib/libtsutil.so.4(+0x13d7f)[0x2b71a2d83d7f]
/usr/local/trafficserver-4.1.0/lib/libtsutil.so.4(+0x1c32e)[0x2b71a2d8c32e]
/usr/local/trafficserver-4.1.0/bin/traffic_server(LogObject::_checkout_write(unsigned long*, unsigned long)+0x1f5)[0x5adf25]
/usr/local/trafficserver-4.1.0/bin/traffic_server(LogObjectManager::check_buffer_expiration(long)+0x7b)[0x5afbfb]
/usr/local/trafficserver-4.1.0/bin/traffic_server(Log::periodic_tasks(long)+0xe2)[0x595c92]
/usr/local/trafficserver-4.1.0/bin/traffic_server(Log::flush_thread_main(void*)+0x28e)[0x596cee]
/usr/local/trafficserver-4.1.0/bin/traffic_server[0x59abcd]
/usr/local/trafficserver-4.1.0/bin/traffic_server(EThread::execute()+0x1159)[0x6ae139]
/usr/local/trafficserver-4.1.0/bin/traffic_server[0x6ab93a]
/lib/x86_64-linux-gnu/libpthread.so.0(+0x6b50)[0x2b71a479cb50]
/lib/x86_64-linux-gnu/libc.so.6(clone+0x6d)[0x2b71a5430a7d]
[E. Mgmt] log ==> [TrafficManager] using root directory '/usr/local/trafficserver-4.1.0'
[TrafficServer] using root directory '/usr/local/trafficserver-4.1.0'

In order to fix this issue, I cook a patch:
1) Use nonblock eventfd, so that we can tolerate write() failed with errno EAGAIN – which is acceptable as the signal receiver will be notified eventually in this case.

2) After using nonblock eventfd, read() will not block in wait(). So I use epoll_wait() to implement block behavior, just like timedwait().

3) nonblock eventfd can fix a potential problem: if receiver didn't read() data immediately, senders might block in write().

Please test this patch, any feedback is welcome.