Implementing io_engine::getevents and io_engine::events

[lukerip]

Hello,

I am currently attempting to write a custom async io engine for a distributed file system that I am working on and am having troubles understanding the semantics of io_eingine::getevents and io_engine::events. The file system I am working on exposes some sort of bespoke epollable IO interface where the file system kicks the epoll instance whenever there are some events that it can reap from the filesystem queue. Something like this:

void SomePollingThread() {
  Epoll epoll;
  SomeDistributedFileSystem* fs = BuildMe(); 
  epoll->AddWatchedFds(fs->GetFdsForEpoll());

  SomeIOObjects ios = GetIOsToSubmt(); 
  fs->SubmitIOs(&ios);
  epoll.Wait(some_deadline);

  Completions completions;
  fs->GetCompletionsFromQueue(&completions);
}

What I am having troubles with is understanding how to map this to the io_engine::getevents and io_engine::events functions.

At the moment I understand an io_u object should map to some sort io object in my own file system, and the number of Completions broadly maps to the return value of io_engine::getevents.

However, what's not clear to me is how the parameters min and max of io_engine::getevents interacts with io_engine::events and what exactly io_engine::events indexes into.

My main issue is do we have to buffer completions if io_engine::getevents specifies a max that is smaller than the number events available?

For example, let's say I do the following:

• queue and submit $N$ IOs into my filesystem by implementing io_engine::queue and io_engine::commit
• at some point FIO is going to poll the filesystem with io_engine::getevents, let's say the max argument is some number max < $N$. However, the filesystem in fact already has $N$ IOs available (I believe theoretically it could have up to iodepth events available)
• FIO now thinks there are max available events and indexes the events up to [0,max) with io_engine::events and process them as being complete.
• Now, do I have to buffer the remaining max - $N$ events somewhere? and have them be available for the next io_engine::events to index into, or should they be dropped ?

Best,
lukerip