Thomas Shackell wrote:
The first thing to note is that concurrency is implemented in Yhc by the interpretter running some instructions for one process, then some more instructions for the next process and so on.
This is rather than using one OS thread per Haskell process because all processes can access the global heap, so if using OS threads it would be necessary to either lock every heap access (incredibly slow) or use some of the truely nasty tricks that concurrent GHC uses.
I resent that :-) There are other choices too: one OS thread per Haskell thread works fine if there's a Giant Lock around the whole runtime (this is the approach that OCaml uses, for instance). Your approach is clearly the right one in the context of YHC, though. I think it's fantastic that you guys have got concurrency working, BTW. Can you really context switch between two arbitrary instructions? Is the stack in a GC'able state at all times?
The first problem, as you note, is that originally we only had one stack but now every process needs its own stack. After considering possible options I decided the easiest way was: - process stacks become normal nodes in the heap, and are garbage collected like other heap nodes. This preserves the property that we can't run out of heap but not stack (and vice versa).
- a process is initially allocated a small stack space in the heap and if that runs out we simply allocate a new stack space (twice as large as before) and let the old stack be garbage collected.
This is exactly what GHC does. I found we had to be very careful with pointers to the old version of the stack, which occur from time to time. eg. ThreadIds are just pointers to the TSO (== stack), they have to be indirected to the new stack when accessed.
Now going back, the approach of running some instructions from one thread then some more from another is fine providing every instruction runs quickly. They all do, with one exception, PRIMITIVE. This is because PRIMITIVE is used for IO actions/FFI calls.
The solution? Well the basic idea is "just spawn an OS thread to do the IO action while the main thread continues running Haskell".
In detail, when PRIMITIVE is called for an IO action/FFI call:
- an OS thread is taken from a pool of threads
- the main thread tells the OS thread to run the IO action/FFI call
- the main thread then reschedules so another haskell process can run
- eventually the OS thread complete it's action and is put back in the pool. It adds the process to the list of processes that have completed their IO actions and are waiting to be rescheduled.
- when the main thread next reschedules it checks the list of completed IO actions and makes the corresponding blocked processes ready to run again.
What happens if the FFI call invokes a callback? (a foreign export, or foreign import wrapper)? Can callbacks be invoked by multiple OS threads? Are you planning to implement bound threads? Cheers, Simmon