Don't crash on failing DRIVER_FindFromHDrvr

[Uwe Bonnes]

Hallo,

I sent a patch to prevent wine from crashing last friday. It happend, when
DRIVER_FindFromHDrvr returned NULL. No I found out, that I had sent that
patch before last year. Eric answered
> this doesn't seem right to me: either the loading fails and we shouldn't 
> get a valid handle, or it succeeds and we should be able to find it
> there's must be something wrong elsewhere
> could you please send a -debugmsg +driver,+module,+winmm to see what 
> happens ?

and like before, I can't reproduce the error after reverting the patch:
> the error no longer exists. I guess some registry entry was corrupt and a
> successfull  run corrected it later.
>
> However I see a lot of check after  DRIVER_FindFromHDrvr() in other parts of
> the code...

If one dose "grep -3 DRIVER_FindFromHDrvr wine/dlls/winmm/*", every other
usage of DRIVER_FindFromHDrvr is like if 
((lpDrv = DRIVER_FindFromHDrvr(hDrvr)) != NULL)

Why should expecting a NULL return be wrong? It  happens, as the the crashes
show.

Bye

[Eric Pouech]

> Why should expecting a NULL return be wrong? It  happens, as the the crashes
> show.

no. the cases are different. The ones you refer to (mainly in driver.c) 
use a hDrvr passed by the called. So, we have to catch erroneous hDrvr 
values.
The case in lolvldrv.c is different. The hDrvr used is the one returned 
by the call to OpenDriver in the same function. That's why it shouldn't 
fail (unless someone closes the driver between the call to OpenDriver 
and the one to DRIVER_FindFromHDrvr - and yes, winmm functions should be 
made reentrant, why they are not today)
A+

[Uwe Bonnes]

> > > > > "Eric" == Eric Pouech  writes:

>> Why should expecting a NULL return be wrong? It happens, as the the
    >> crashes show.
    Eric> no. the cases are different. The ones you refer to (mainly in
    Eric> driver.c) use a hDrvr passed by the called. So, we have to catch
    Eric> erroneous hDrvr values.  The case in lolvldrv.c is different. The
    Eric> hDrvr used is the one returned by the call to OpenDriver in the
    Eric> same function. That's why it shouldn't fail (unless someone closes
    Eric> the driver between the call to OpenDriver and the one to
    Eric> DRIVER_FindFromHDrvr - and yes, winmm functions should be made
    Eric> reentrant, why they are not today) A+

Eric,

we discussed the problem some time ago, and came to no conclusion. You
requested a debug log, but I didn't hit that error in the meantime. Now I
hit it again.

I changed in winmm/driver.c
LPWINE_DRIVER	DRIVER_TryOpenDriver32(LPCSTR fn, LPARAM lParam2)
{
    LPWINE_DRIVER 	lpDrv = NULL;
    HMODULE		hModule = 0;
    LPSTR		ptr;
    LPCSTR		cause = 0;

TRACE("(%s, %08lX);\n", debugstr_a(fn), lParam2);

if ((ptr = strchr(fn, ' ')) != NULL) {
    *ptr++ = '\0';
    while (*ptr == ' ') ptr++;
    if (*ptr == '\0') ptr = NULL;
    }

FIXME("BON: lpDrv before HeapAlloc is %p\n",lpDrv);
    lpDrv = HeapAlloc(GetProcessHeap(), 0, sizeof(WINE_DRIVER));
    FIXME("BON: after HeapAlloc %p\n",lpDrv);
    if(!HeapValidate(GetProcessHeap(), 0, lpDrv))
    {
        FIXME("BON:HeapValidate failed\n");
    }

A +relay,+heap,+winmm,+driver log looks like:

trace:winmm:MMDRV_Install ('wineoss.drv', 'wineoss.drv', mapper=N);
trace:driver:OpenDriverA ("wineoss.drv", (null), 0x00000000);
0009:Call kernel32.lstrcpynA(406ef9b8,406efb28 "wineoss.drv",00000080) ret=41285668
0009:Ret  kernel32.lstrcpynA() retval=406ef9b8 ret=41285668
trace:driver:DRIVER_TryOpenDriver32 ("wineoss.drv", 00000000);
fixme:driver:DRIVER_TryOpenDriver32 BON: HeapAlloc is (nil)
0009:Call kernel32.GetCurrentThreadId() ret=0a908a76
0009:Ret  kernel32.GetCurrentThreadId() retval=00000009 ret=0a908a76
fixme:driver:DRIVER_TryOpenDriver32 BON: HeapAlloc got 0x41c401c0
0009:Call kernel32.HeapValidate(403a0000,00000000,41c401c0) ret=4128527e
0009:Call ntdll.RtlValidateHeap(403a0000,00000000,41c401c0) ret=4050a731
warn:heap:HEAP_IsRealArena Heap 0x403a0000: block 0x41c401c0 is not inside heap
0009:Ret  ntdll.RtlValidateHeap() retval=00000000 ret=4050a731
0009:Ret  kernel32.HeapValidate() retval=00000000 ret=4128527e
fixme:driver:DRIVER_TryOpenDriver32 BON:HeapValidate failed

Strange: we don't see the RtlHeapAllocate function. But then I remember that
the application (_impact.exe from the Xilinx ISE suite) uses the shsmp library
for memory allocation (http://www.microquill.com/smartheap/sh_tspec.htm).

So our HeapAlloc get patched by the Microquill code, but probably
HeapValidate dose not get patched and things diverge.

Eric, can we do the memory validation perhaps in some other way then by
HeapValidate ?

Bye

[Eric Pouech]

> Strange: we don't see the RtlHeapAllocate function. But then I remember that
> the application (_impact.exe from the Xilinx ISE suite) uses the shsmp library
> for memory allocation (http://www.microquill.com/smartheap/sh_tspec.htm).
>
> So our HeapAlloc get patched by the Microquill code, but probably
> HeapValidate dose not get patched and things diverge.
>
> Eric, can we do the memory validation perhaps in some other way then by
> HeapValidate ?

several questions here:
- if your theory (about missing HeapValidate func) is right: then I 
think that this memory manager is rather broken and should be fixed and 
not Wine
- but on the other hand, how could you explain that it did work at some 
point (ie didn't crash) and did it afterwards ? This doesn't seem too 
logical to me. Perhaps the smartheap API interception mechanism isn't 
that stable on Wine.
IMO, you should start looking at the interception code and see if it 
puts something in place for HeapValidate, then we'll see what we could 
do (one quickie is to use RtlValidateHeap instead of HeapValidate, but 
that's *really* ugly).
A+

[Uwe Bonnes]

> > > > > "Eric" == Eric Pouech  writes:

>> Strange: we don't see the RtlHeapAllocate function. But then I
    >> remember that the application (_impact.exe from the Xilinx ISE suite)
    >> uses the shsmp library for memory allocation
    >> (http://www.microquill.com/smartheap/sh_tspec.htm).
    >> 
    >> So our HeapAlloc get patched by the Microquill code, but probably
    >> HeapValidate dose not get patched and things diverge.
    >> 
    >> Eric, can we do the memory validation perhaps in some other way then
    >> by HeapValidate ?
    Eric> several questions here: - if your theory (about missing
    Eric> HeapValidate func) is right: then I think that this memory manager
    Eric> is rather broken and should be fixed and not Wine - but on the
    Eric> other hand, how could you explain that it did work at some point
    Eric> (ie didn't crash) and did it afterwards ? This doesn't seem too
    Eric> logical to me. Perhaps the smartheap API interception mechanism
    Eric> isn't that stable on Wine.  IMO, you should start looking at the
    Eric> interception code and see if it puts something in place for
    Eric> HeapValidate, then we'll see what we could do (one quickie is to
    Eric> use RtlValidateHeap instead of HeapValidate, but that's *really*
    Eric> ugly).  A+

Eric> - if your theory (about missing HeapValidate func) is right: 
    Eric>   then I think that this memory manager is rather broken and 
    Eric>   should be fixed and not Wine

We have to deal with a lot of brokenness...

I will try to get in contact with Microquill. But even if they fix it, it
will take some time until there fix propagates into those projects that use
this library and then there are all these old programs out in the wild. And
I think a some big project use this library..

And HeapValidate is not a much used function. For example in our wine tree
only uses winmm/driver.c and winmm/lolvldrv.c use HeapValidate outside the
heap handling itself.

An probably any user of the Microquill libray who had problems with
HeapValidate circumvented the problem in the meantime...

Eric> - but on the other hand, how could you explain that it did 
    Eric>   work at some point (ie didn't crash) and did it afterwards ? 
    Eric>   This doesn't seem too logical to me. Perhaps the smartheap 
    Eric>   API interception mechanism isn't that stable on Wine.

Well, impact is the end of the tool chain. There were a lot of other
problems, so I didn't look at the impact problem. And it only happens with
wine running as win95/98. This is probably the reason why Xilinx ISE
requests win2k or winxp.

Eric> IMO, you should start looking at the interception code and 
    Eric> see if it puts something in place for HeapValidate, then we'll 
    Eric> see what we could do (one quickie is to use RtlValidateHeap 
    Eric> instead of HeapValidate, but that's *really* ugly).

I tried the other way. Using RtlAllocateHeap also showed the problem. But
when I did a GetProcAddres on RtlAllocateHeap, things worked as expected
(the trace showed a RtlAllocateHeap and the following Heapvalidate reported
success).

Is using GetProcAddress on HeapAllocate an acceptable solution for that
problem?

Bye