Reorganisation of the source tree
[ghc-hetmet.git] / rts / win32 / AsyncIO.c
diff --git a/rts/win32/AsyncIO.c b/rts/win32/AsyncIO.c
new file mode 100644 (file)
index 0000000..7bcf571
--- /dev/null
@@ -0,0 +1,345 @@
+/* AsyncIO.c
+ *
+ * Integrating Win32 asynchronous I/O with the GHC RTS.
+ *
+ * (c) sof, 2002-2003.
+ */
+#include "Rts.h"
+#include "RtsUtils.h"
+#include <windows.h>
+#include <stdio.h>
+#include "Schedule.h"
+#include "RtsFlags.h"
+#include "Capability.h"
+#include "win32/AsyncIO.h"
+#include "win32/IOManager.h"
+
+/*
+ * Overview:
+ *
+ * Haskell code issue asynchronous I/O requests via the 
+ * async{Read,Write,DoOp}# primops. These cause addIORequest()
+ * to be invoked, which forwards the request to the underlying
+ * asynchronous I/O subsystem. Each request is tagged with a unique
+ * ID.
+ *
+ * addIORequest() returns this ID, so that when the blocked CH
+ * thread is added onto blocked_queue, its TSO is annotated with
+ * it. Upon completion of an I/O request, the async I/O handling
+ * code makes a back-call to signal its completion; the local
+ * onIOComplete() routine. It adds the IO request ID (along with
+ * its result data) to a queue of completed requests before returning. 
+ *
+ * The queue of completed IO request is read by the thread operating
+ * the RTS scheduler. It de-queues the CH threads corresponding
+ * to the request IDs, making them runnable again.
+ *
+ */
+
+typedef struct CompletedReq {
+    unsigned int   reqID;
+    int            len;
+    int            errCode;
+} CompletedReq;
+
+#define MAX_REQUESTS 200
+
+static CRITICAL_SECTION queue_lock;
+static HANDLE           completed_req_event;
+static HANDLE           abandon_req_wait;
+static HANDLE           wait_handles[2];
+static CompletedReq     completedTable[MAX_REQUESTS];
+static int              completed_hw;
+static HANDLE           completed_table_sema;
+static int              issued_reqs;
+
+static void
+onIOComplete(unsigned int reqID,
+            int   fd STG_UNUSED,
+            int   len,
+            void* buf STG_UNUSED,
+            int   errCode)
+{
+    DWORD dwRes;
+    /* Deposit result of request in queue/table..when there's room. */
+    dwRes = WaitForSingleObject(completed_table_sema, INFINITE);
+    switch (dwRes) {
+    case WAIT_OBJECT_0:
+       break;
+    default:
+       /* Not likely */
+       fprintf(stderr, "onIOComplete: failed to grab table semaphore, dropping request 0x%x\n", reqID);
+       fflush(stderr);
+       return;
+    }
+    EnterCriticalSection(&queue_lock);
+    if (completed_hw == MAX_REQUESTS) {
+       /* Shouldn't happen */
+       fprintf(stderr, "onIOComplete: ERROR -- Request table overflow (%d); dropping.\n", reqID);
+       fflush(stderr);
+    } else {
+#if 0
+       fprintf(stderr, "onCompl: %d %d %d %d %d\n", 
+               reqID, len, errCode, issued_reqs, completed_hw); 
+       fflush(stderr);
+#endif
+       completedTable[completed_hw].reqID   = reqID;
+       completedTable[completed_hw].len     = len;
+       completedTable[completed_hw].errCode = errCode;
+       completed_hw++;
+       issued_reqs--;
+       if (completed_hw == 1) {
+           /* The event is used to wake up the scheduler thread should it
+            * be blocked waiting for requests to complete. The event resets once
+            * that thread has cleared out the request queue/table.
+            */
+           SetEvent(completed_req_event);
+       }
+    }
+    LeaveCriticalSection(&queue_lock);
+}
+
+unsigned int
+addIORequest(int   fd,
+            int   forWriting,
+            int   isSock,
+            int   len,
+            char* buf)
+{
+    EnterCriticalSection(&queue_lock);
+    issued_reqs++;
+    LeaveCriticalSection(&queue_lock);
+#if 0
+    fprintf(stderr, "addIOReq: %d %d %d\n", fd, forWriting, len); fflush(stderr);
+#endif
+    return AddIORequest(fd,forWriting,isSock,len,buf,onIOComplete);
+}
+
+unsigned int
+addDelayRequest(int msecs)
+{
+    EnterCriticalSection(&queue_lock);
+    issued_reqs++;
+    LeaveCriticalSection(&queue_lock);
+#if 0
+    fprintf(stderr, "addDelayReq: %d\n", msecs); fflush(stderr);
+#endif
+    return AddDelayRequest(msecs,onIOComplete);
+}
+
+unsigned int
+addDoProcRequest(void* proc, void* param)
+{
+    EnterCriticalSection(&queue_lock);
+    issued_reqs++;
+    LeaveCriticalSection(&queue_lock);
+#if 0
+    fprintf(stderr, "addProcReq: %p %p\n", proc, param); fflush(stderr);
+#endif
+    return AddProcRequest(proc,param,onIOComplete);
+}
+
+
+int
+startupAsyncIO()
+{
+    if (!StartIOManager()) {
+       return 0;
+    }
+    InitializeCriticalSection(&queue_lock);
+    /* Create a pair of events:
+     *
+     *    - completed_req_event  -- signals the deposit of request result; manual reset.
+     *    - abandon_req_wait     -- external OS thread tells current RTS/Scheduler
+     *                              thread to abandon wait for IO request completion.
+     *                              Auto reset.
+     */
+    completed_req_event = CreateEvent (NULL, TRUE,  FALSE, NULL);
+    abandon_req_wait    = CreateEvent (NULL, FALSE, FALSE, NULL);
+    wait_handles[0] = completed_req_event;
+    wait_handles[1] = abandon_req_wait;
+    completed_hw = 0;
+    if ( !(completed_table_sema = CreateSemaphore (NULL, MAX_REQUESTS, MAX_REQUESTS, NULL)) ) {
+       DWORD rc = GetLastError();
+       fprintf(stderr, "startupAsyncIO: CreateSemaphore failed 0x%x\n", rc);
+       fflush(stderr);
+    }
+
+    return ( completed_req_event  != INVALID_HANDLE_VALUE &&
+            abandon_req_wait     != INVALID_HANDLE_VALUE &&
+            completed_table_sema != NULL );
+}
+
+void
+shutdownAsyncIO()
+{
+    CloseHandle(completed_req_event);
+    ShutdownIOManager();
+}
+
+/*
+ * Function: awaitRequests(wait)
+ *
+ * Check for the completion of external IO work requests. Worker
+ * threads signal completion of IO requests by depositing them
+ * in a table (completedTable). awaitRequests() matches up 
+ * requests in that table with threads on the blocked_queue, 
+ * making the threads whose IO requests have completed runnable
+ * again.
+ * 
+ * awaitRequests() is called by the scheduler periodically _or_ if
+ * it is out of work, and need to wait for the completion of IO
+ * requests to make further progress. In the latter scenario, 
+ * awaitRequests() will simply block waiting for worker threads 
+ * to complete if the 'completedTable' is empty.
+ */
+int
+awaitRequests(rtsBool wait)
+{
+#ifndef THREADED_RTS
+  // none of this is actually used in the threaded RTS
+
+start:
+#if 0
+    fprintf(stderr, "awaitRequests(): %d %d %d\n", issued_reqs, completed_hw, wait);
+    fflush(stderr);
+#endif
+    EnterCriticalSection(&queue_lock);
+    /* Nothing immediately available & we won't wait */
+    if ((!wait && completed_hw == 0)
+#if 0
+       // If we just return when wait==rtsFalse, we'll go into a busy
+       // wait loop, so I disabled this condition --SDM 18/12/2003
+       (issued_reqs == 0 && completed_hw == 0)
+#endif
+       ) {
+       LeaveCriticalSection(&queue_lock);
+       return 0;
+    }
+    if (completed_hw == 0) {
+       /* empty table, drop lock and wait */
+       LeaveCriticalSection(&queue_lock);
+       if ( wait && sched_state == SCHED_RUNNING ) {
+           DWORD dwRes = WaitForMultipleObjects(2, wait_handles, FALSE, INFINITE);
+           switch (dwRes) {
+           case WAIT_OBJECT_0:
+               /* a request was completed */
+               break;
+           case WAIT_OBJECT_0 + 1:
+           case WAIT_TIMEOUT:
+               /* timeout (unlikely) or told to abandon waiting */
+               return 0;
+           case WAIT_FAILED: {
+               DWORD dw = GetLastError();
+               fprintf(stderr, "awaitRequests: wait failed -- error code: %lu\n", dw); fflush(stderr);
+               return 0;
+           }
+           default:
+               fprintf(stderr, "awaitRequests: unexpected wait return code %lu\n", dwRes); fflush(stderr);
+               return 0;
+           }
+       } else {
+           return 0;
+       }
+       goto start;
+    } else {
+       int i;
+       StgTSO *tso, *prev;
+       
+       for (i=0; i < completed_hw; i++) {
+           /* For each of the completed requests, match up their Ids
+            * with those of the threads on the blocked_queue. If the
+            * thread that made the IO request has been subsequently
+            * killed (and removed from blocked_queue), no match will
+            * be found for that request Id. 
+            *
+            * i.e., killing a Haskell thread doesn't attempt to cancel
+            * the IO request it is blocked on.
+            *
+            */
+           unsigned int rID = completedTable[i].reqID;
+           
+           prev = NULL;
+           for(tso = blocked_queue_hd ; tso != END_TSO_QUEUE; prev = tso, tso = tso->link) {
+       
+               switch(tso->why_blocked) {
+               case BlockedOnRead:
+               case BlockedOnWrite:
+               case BlockedOnDoProc:
+                   if (tso->block_info.async_result->reqID == rID) {
+                       /* Found the thread blocked waiting on request; stodgily fill 
+                        * in its result block. 
+                        */
+                       tso->block_info.async_result->len = completedTable[i].len;
+                       tso->block_info.async_result->errCode = completedTable[i].errCode;
+                       
+                       /* Drop the matched TSO from blocked_queue */
+                       if (prev) {
+                           prev->link = tso->link;
+                       } else {
+                           blocked_queue_hd = tso->link;
+                       }
+                       if (blocked_queue_tl == tso) {
+                           blocked_queue_tl = prev ? prev : END_TSO_QUEUE;
+                       }
+                   
+                       /* Terminates the run queue + this inner for-loop. */
+                       tso->link = END_TSO_QUEUE;
+                       tso->why_blocked = NotBlocked;
+                       pushOnRunQueue(&MainCapability, tso);
+                       break;
+                   }
+                   break;
+               default:
+                   if (tso->why_blocked != NotBlocked) {
+                       barf("awaitRequests: odd thread state");
+                   }
+                   break;
+               }
+           }
+           /* Signal that there's completed table slots available */
+           if ( !ReleaseSemaphore(completed_table_sema, 1, NULL) ) {
+               DWORD dw = GetLastError();
+               fprintf(stderr, "awaitRequests: failed to signal semaphore (error code=0x%x)\n", dw);
+               fflush(stderr);
+           }
+       }
+       completed_hw = 0;
+       ResetEvent(completed_req_event);
+       LeaveCriticalSection(&queue_lock);
+       return 1;
+    }
+#endif /* !THREADED_RTS */
+}
+
+/*
+ * Function: abandonRequestWait()
+ *
+ * Wake up a thread that's blocked waiting for new IO requests
+ * to complete (via awaitRequests().)
+ */
+void
+abandonRequestWait( void )
+{
+    /* the event is auto-reset, but in case there's no thread
+     * already waiting on the event, we want to return it to
+     * a non-signalled state.
+     *
+     * Careful!  There is no synchronisation between
+     * abandonRequestWait and awaitRequest, which means that
+     * abandonRequestWait might be called just before a thread
+     * goes into a wait, and we miss the abandon signal.  So we
+     * must SetEvent() here rather than PulseEvent() to ensure
+     * that the event isn't lost.  We can re-optimise by resetting
+     * the event somewhere safe if we know the event has been
+     * properly serviced (see resetAbandon() below).  --SDM 18/12/2003
+     */
+    SetEvent(abandon_req_wait);
+}
+
+void
+resetAbandonRequestWait( void )
+{
+    ResetEvent(abandon_req_wait);
+}
+