Set initial state of 'data_event' and 'highwater_event' to *not* signaled
Since commit fafed094 (2 months ago), the Synchronet Web Server on my Windows systems has occasionally gone into a state where every HTTP session thread was causing 100% CPU utilization and each new HTTP session thread would just exacerbate the problem eventually leading to complete system instability/unresponsiveness until the sbbs instance was terminated. This was more readily reproducible on a Win7-32 VM, but would occasionally, though much less frequently, happen in a native instance on Win10-64 as well. Using the VisualStudio debugger, I was able to narrow it down to this: Each new HTTP thread (eventually, hundreds of such threads) would get stuck in this loop in http_output_thread(): while(session->socket!=INVALID_SOCKET) { /* Wait for something to output in the RingBuffer */ if((avail=RingBufFull(obuf))==0) { /* empty */ if(WaitForEvent(obuf->data_event, 1000) != WAIT_OBJECT_0) continue; /* Check for spurious sem post... */ if((avail=RingBufFull(obuf))==0) continue; // <--- data_event signaled, but never cleared } There appears to be a race condition where this logic could be executed immediately after the output ringbuf was created, but before writebuf() was ever called (which would have actually placed data in the output buffer), causing a potential high-utilization infinite loop: the data_event is signaled but there is no data and the event is never reset and nothing can ever add data to the ringbuf due to starvation of CPU cycles. Uses of ringbuf's data_event elsewhere in Synchronet don't seem to be subject to this issue since they always call RingBufRead after, which will clear the data_event when the ringbuf is actually empty (no similar loops to this one). The root cause just appears to be a simple copy/paste issue in the code added to RingBufInit(): the preexisting 'empty_event' was initialized with a correct initiate state of 'signaled' (because by default a ringbuf is empty) but the newly added events (data_event and highwater_event) should *not* be initially-signaled because... the ringbuf is empty. So I added some parameter comments to these calls to CreateEvent() to hopefully make that more clear and prevent similar mistakes in the future. Relieved to have this one resolved finally.
parent
2e934268
No related branches found
No related tags found
-
The correct culprit commit was a2aa9631 (the one after fafed...)
Please register or sign in to comment