Nested EPOLL FD

Question

I find very limited information online about the behaviour of nested FDs.

Lets say EPOLL FD1 is shared between EPOLL FD2 and FD3. So what would happen if epoll FD1 get an event

1. Is it going to raise an event to both FD2 and Fd3 if it is added without epoll_exclusive flag.
2. After getting an event, do i need to do epoll_wait on FD1 to clear the event.

And what is the use case of nested epoll FD.

Answer 1

For the level-triggered case, it is pretty straightforward:

1. Yes, you can combine the answers to Q2 and Q3 from the Q&A part in epoll(7) manpage:
   A2: If the same file descriptor is added to two epoll instances, events are reported to both of them.
   A3: If an epoll file descriptor has events waiting, then it will indicate as being readable.
2. No, since clearing the event is only relevant for edge-triggered epoll.

For the edge-triggered case, we would expect the answer to 1 to stay the same, since regardless of nesting, if the same FD is added with EPOLLET to two epoll instances they will both report every event once.

I don't know what the answer to the second question SHOULD be, but it doesn't matter. There is a bug when adding epoll FDs in edge-triggerd mode, they just always behave like level-triggered, FD2 and FD3 will continue to show that FD1 is readable as long as it has pending events. You can see this thread about an attempt to fix it which I don't think made to mainland yet: fs/epoll: fix the edge-triggered mode for nested epoll
Which links to this GitHub repo which tests a lot of different scenarios, that you might find interesting.