The tclEpollNotfy PlatformEventsControl function panics if the TclOSfstat call returns -1, which occurs when using a websocket to a browser and the browser page is refreshed. 

Jan suggested changing the Panic call to LIST_REMOVE(filePtr, readyNode) to remove the filePtr from the hash table as that is more appropriate during the deletion of the handle, but that changes the behavior from a panic to a SIGSEGV.

Tested on Ubuntu 22.04.4 LTS.   Doesn't affect 8.6 since that doesn't have ePoll support on Linux.

Fixed [e3ffca215ef8cf8a|here]. No problems with it reported.

If this change causes newly found problems (which is not expected), please re-open this ticket

Comments added, link fixed

Dear Steve, dear Donal, thanks for this great action. Is the current source code comment sufficiant? Would it feasable to mention this ticket in a source code comment?

I am wondering, why this ticket is shown as "new wiki page" in fossil timeline. Is it me?

Thanks for all, Harald

I've confirmed that just returning from PlatformEventsControl rather than a panic does avoid the crash.  So that's a workaround albeit at the cost of a memory leak.  Is that sufficient or should we keep looking for a better solution?

At the moment, it seems that the fstat call isn't doing anything useful. In particular, the contents of the statbuf aren't examined afterwards on success; at best, it changes what message we panic with.