[GHC] #12038: Shutdown interacts badly with requestSync()

[GHC]

#12038: Shutdown interacts badly with requestSync()
-------------------------------------+-------------------------------------
        Reporter:  simonmar          |                Owner:
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  8.2.1
       Component:  Runtime System    |              Version:  7.10.3
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------
Description changed by simonmar:

@@ -0,0 +1,3 @@
+ I've been investigating #10860, and the problem goes pretty deep,
+ so I'm going to record what I know here and come back to fix it
+ properly later.
@@ -2,0 +5,19 @@
+ We have this mechanism `requestSync()` for operations that need
+ to seize control of the whole runtime to do something.  It is used by
+
+ * `scheduleDoGC()`
+ * `setNumCapabilities()`
+ * `forkProcess()`
+
+ `requestSync()` ensures that only one of these operations wins,
+ the others will `yieldCapability()` to the winner, before
+ continuing with their own sync.
+
+ The problem is that this interacts badly with shutdown.  Shutdown
+ might start at any time (initiated by `exitScheduler()`).  If it
+ starts during a sync, then a deadlock is likely: some
+ capabilities will be already shut down, and cannot be acquired by
+ `acquireAllCapabilities()`.  This happens in #10860.
+
+ Really, shutdown should play the `requestSync()` game too, but
+ that requires a lot of thought.

New description:

 I've been investigating #10860, and the problem goes pretty deep,
 so I'm going to record what I know here and come back to fix it
 properly later.

 We have this mechanism `requestSync()` for operations that need
 to seize control of the whole runtime to do something.  It is used by

 * `scheduleDoGC()`
 * `setNumCapabilities()`
 * `forkProcess()`

 `requestSync()` ensures that only one of these operations wins,
 the others will `yieldCapability()` to the winner, before
 continuing with their own sync.

 The problem is that this interacts badly with shutdown.  Shutdown
 might start at any time (initiated by `exitScheduler()`).  If it
 starts during a sync, then a deadlock is likely: some
 capabilities will be already shut down, and cannot be acquired by
 `acquireAllCapabilities()`.  This happens in #10860.

 Really, shutdown should play the `requestSync()` game too, but
 that requires a lot of thought.

--

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12038#comment:1>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler