Ok, so here's my current patch (work in progress). I have not run tests yet, and finding a way to add a test case is virtually impossible though I expect we will find ways of using gdb to confirm local fix later.
After careful code review, I settled on the following approach which seemed to be the least intrusive.
1. Change the retry logic so that it does not retry of QueryCancelPending or ProcDiePending are set. In these cases EINTR is passed back to the caller
2. Change the error handling logic of the caller so that EINTR is handled by the next CHECK_FOR_INTERRUPTS() after cleanup.
This means that the file descriptor is now already closed and that we handle this the same way we would with a ENOSPC. The reason for this is that there are many places in the code where it is not clear what degree of safety is present for throwing errors using ereport, and the patch needs to be as unobtrusive as possible to facilitate back porting.
At this point the patch is for discussion. I have not even compiled it or tested it but maybe there is something wrong with my approach so figured I would send it out early.
The major assumptions are:
1. close() will never take longer than the interrupt interval that caused the loop to break.
2. close() does not get interrupted in a way that will not cause cleanup problems later.
3. We will reach the next interrupt check at a reasonable interval and safe spot.
Any initial arguments against?
--
Best Regards,
Chris Travers
Head of Database
Saarbrücker Straße 37a, 10405 Berlin