Non-threaded WinIO is broken

It appears that WinIO is in a pretty sad state:

• it is no longer easily possible to bootstrap a compiler which uses WinIO by default since 85b0f87a
• once you figure out how to do so you quickly find that the result segfaults upon any IO

added Pnormal Tbug Windows winio labels

assigned to @bgamari

FWIW, I used the following to select WinIO by default:

diff --git a/rts/configure.ac b/rts/configure.ac
index e9181aaf55..b85873e383 100644
--- a/rts/configure.ac
+++ b/rts/configure.ac
@@ -401,10 +401,10 @@ dnl Now we establish a default I/O manager for the threaded and non-threaded
 dnl RTS. We select the default based on which I/O managers are enabled. They
 dnl are checked in reverse order of priority, the last enabled one wins:
 if test "$HostOS" = "mingw32"; then
-  GHC_IOMANAGER_DEFAULT_SELECT([IOManagerNonThreadedDefault], [winio], [EnableIOManagerWinIO],)
   GHC_IOMANAGER_DEFAULT_SELECT([IOManagerNonThreadedDefault], [win32-legacy], [EnableIOManagerWin32Legacy])
-  GHC_IOMANAGER_DEFAULT_SELECT([IOManagerThreadedDefault], [winio], [EnableIOManagerWinIO])
+  GHC_IOMANAGER_DEFAULT_SELECT([IOManagerNonThreadedDefault], [winio], [EnableIOManagerWinIO],)
   GHC_IOMANAGER_DEFAULT_SELECT([IOManagerThreadedDefault], [mio], [EnableIOManagerMIO])
+  GHC_IOMANAGER_DEFAULT_SELECT([IOManagerThreadedDefault], [winio], [EnableIOManagerWinIO])
 else
   GHC_IOMANAGER_DEFAULT_SELECT([IOManagerNonThreadedDefault], [select], [EnableIOManagerSelect])
   GHC_IOMANAGER_DEFAULT_SELECT([IOManagerThreadedDefault], [mio], [EnableIOManagerMIO])

The problem appears to be that ioMan (of type IOManagerState*) is NULL:

Thread 1 received signal SIGSEGV, Segmentation fault.
0x00007ff668abc690 in AddIORequest (fd=fd@entry=2, forWriting=true, isSocket=false, len=len@entry=1, buffer=0x7ff4fd7f7010 "g",
    onCompletion=0x7ff668931cb0 <onIOComplete>) at rts\win32/MIOManager.c:446
446         unsigned int reqID = ioMan->requestID++;
(gdb) print ioMan
$1 = (IOManagerState *) 0x0
(gdb) bt
#0  0x00007ff668abc690 in AddIORequest (fd=fd@entry=2, forWriting=true, isSocket=false, len=len@entry=1, buffer=0x7ff4fd7f7010 "g",
    onCompletion=0x7ff668931cb0 <onIOComplete>) at rts\win32/MIOManager.c:446
#1  0x00007ff668931c9c in addIORequest (fd=2, forWriting=true, isSock=false, len=1, buf=0x7ff4fd7f7010 "g")
    at rts\win32/AsyncMIO.c:125
#2  0x00007ff66803be8c in stg_asyncWritezh ()
#3  0x0000000000000002 in ?? ()
#4  0x00007ff667f11e88 in stg_marked_upd_frame_info ()
#5  0x00007ff4fda04b08 in ?? ()
#6  0x00007ff668d60501 in ghczminternal_GHCziInternalziMaybe_Nothing_closure ()
#7  0x00007ff4fd7f7010 in ?? ()
#8  0x006e00690000005c in ?? ()
#9  0x0000000000000214 in ?? ()
#10 0x000000000014db18 in ?? ()
#11 0x0000000000120010 in ?? ()
#12 0x000000000014db50 in ?? ()
#13 0x0000000000000000 in ?? ()
(gdb)

Indeed it appears that StartIOManager, which appears to initialize ioMan, is not called.

Actually, the real question here is why we are ending up in MIOManager.c at all; we should be in the WinIO path.

The stage2 ghc and indeed executables built with -threaded work fine so I suspect that something in !9676 (closed) may have broken WinIO in the non-threaded configuration.

The problem is

(gdb) print rts_IOManagerIsWin32Native
$1 = false

The problem is this: !9676 (diffs).

Namely rts_IOManagerIsWin32Native isn't properly set when the IO_MNGR_FLAG_AUTO path is taken.

Fixed by:

diff --git a/rts/IOManager.c b/rts/IOManager.c
index d70713fa34..375d2adb73 100644
--- a/rts/IOManager.c
+++ b/rts/IOManager.c
@@ -233,7 +233,6 @@ void selectIOManager(void)
 #if defined(IOMGR_ENABLED_WINIO)
         case IO_MNGR_FLAG_WINIO:
             iomgr_type = IO_MANAGER_WINIO;
-            rts_IOManagerIsWin32Native = true;
             break;
 #endif
@@ -246,6 +245,8 @@ void selectIOManager(void)
         default:
           barf("selectIOManager: %d", RtsFlags.MiscFlags.ioManager);
     }
+
+    rts_IOManagerIsWin32Native = iomgr_type == IO_MANAGER_WINIO;
 }

changed title from WinIO is broken to Non-thraeded WinIO is broken

changed title from Non-thraeded WinIO is broken to Non-threaded WinIO is broken

mentioned in commit 3bc8cfdc

mentioned in merge request !14088 (closed)

Fixed in !14088 (closed) although next it appears that there are further issues since Haddock crashes with:

haddock: internal error: copyFile:atomicCopyFileContents:withReplacementFile:CloseHandle: invalid argument (The handle is invalid.)

mentioned in commit 9d680679

mentioned in issue #25111

mentioned in commit 5ccaa6ad

mentioned in commit bad34019

mentioned in commit cf96bfaf

closed with commit 5b94f99f

mentioned in commit 5b94f99f