Re: [ros-svn] [ion] 16635: - Fix nasty APC delivery bug (in case a Kernel-Mode Special APC still returned with a Normal Routine, the Normal Routine was called with incorrect values (Special Routines take PVOID* arguments, while Normal Routines do not!))

[Hartmut Birr]

[e-mail] wrote:
> Fix nasty APC delivery bug (in case a Kernel-Mode Special APC still returned with a Normal Routine, the Normal Routine was called with incorrect values (Special Routines take PVOID* arguments, while Normal Routines do not!))
> Remove APC from list before setting it to non-inserted.
> Do proper thread termination piggybacking; terminate threads in user-mode as Hartmut correctly fixed.
>
> Updated files:
> trunk/reactos/ntoskrnl/ke/apc.c
> trunk/reactos/ntoskrnl/ps/kill.c

Hi,

I do not agree to this changes. For a non system thread, we use a kernel
mode apc first, which uses a second user mode apc. This overhead isn't
necessary. For system threads it isn't possible to use a kernel mode apc
which calls PspExitThread. A system thread must terminate it self. It is
only possible to help by marking a flag, setting an event or unwait the
thread.

- Hartmut

[Alex Ionescu]

> Hi,
>
> I do not agree to this changes. For a non system thread, we use a kernel
> mode apc first, which uses a second user mode apc. This overhead isn't
> necessary. For system threads it isn't possible to use a kernel mode apc
> which calls PspExitThread. A system thread must terminate it self. It is
> only possible to help by marking a flag, setting an event or unwait the
> thread.
> Hartmut

Hi,

I based these changes on some APC tracing on my system. First, a kernel 
APC is queded up without the special flag and its normal function is 
called PspExitNormalApc. This function then kills the thread immediately 
if it's a system thread. I believe this is an optimization and allows us 
to kill the thread directly from the first APC...there's no use in going 
down to user-mode for this. If this is a user-thread, then a second APC 
is queued, this time usermode, to kill the thread. I don't really see 
any overhead...this is faster then queuing a user-mode APC to deal with 
everything, since system threads will stay in kernel and we won't need 
the overhead of a user APC (much larger then the overhead of killing a 
user thread by doing two apcs).

Best regards,
Alex Ionescu