Re: [Algorithms] General purpose task parallel threading approach

[Jon W. <jw...@gm...>]

asy...@gm... wrote:
> Where do you get a deadlock from ?

If a specific operation is not cancellable, and waits for completion no 
matter what, then that means that cancelling the fiber will have to wait 
for that operation to complete. If the operation in turn needs another 
task to complete, which may be later in the work queue, or require 
locks/resources currently held by the fiber doing the cancellation, then 
you have a deadlock.

In general, non-cancellable operations are bad (similar to 
non-signal-breakable syscalls in UNIX). The right thing to do is to make 
any operation cancellable, and we've found that the best way to make 
sure nobody does the wrong thing is to simply throw an exception out of 
the yield call when an operation is cancelled. Proper use of smart 
pointers and RAII will then take care of the unwind semantics.

That being said, blocking is still rare in game loops; we use all of 
this only for the transactional application server back-end.

Sincerely,

jw