Re: [PATCH v4 0/3] delay timer_new from init to realize to fix memleaks.

[Mark Cave-Ayland]

On 08/03/2020 13:39, Peter Maydell wrote:

> On Sun, 8 Mar 2020 at 11:58, Mark Cave-Ayland
> <address@hidden> wrote:
>> I just tried this patchset applied on top of git master and it causes 
>> qemu-system-ppc
>> to segfault on startup:
>>
>> $ gdb --args ./qemu-system-ppc
>> ...
>> ...
>> Thread 1 "qemu-system-ppc" received signal SIGSEGV, Segmentation fault.
>> 0x0000555555e7e38c in timer_del (ts=0x0) at util/qemu-timer.c:429
>> 429         QEMUTimerList *timer_list = ts->timer_list;
>> (gdb) bt
>> #0  0x0000555555e7e38c in timer_del (ts=0x0) at util/qemu-timer.c:429
>> #1  0x0000555555b5d2c1 in mos6522_reset (dev=0x555556e0ac50) at 
>> hw/misc/mos6522.c:468
>> #2  0x0000555555b63570 in mos6522_cuda_reset (dev=0x555556e0ac50) at
>> hw/misc/macio/cuda.c:599
> 
> It looks like we haven't caught all the cases of "somebody created a
> MOS6522 (or one of its subclasses) but forgot to realize it". This
> particular one I think is the s->cuda which is inited in macio_oldworld_init()
> but not realized in macio_oldworld_realize(). I think that pmu_init() in
> hw/misc/macio/pmu.c also has this bug. We need to go through and
> audit all the places where we create TYPE_MOS6522 or any of its
> subclasses and make sure they are also realizing the devices they create.
> (The presence of the new 3-phase reset infrastructure in the backtrace
> is a red herring here -- this would have crashed the same way with the
> old code too.)
> 
> We should probably find some generic place in Device code where we
> can stick an assert "are we trying to reset an unrealized device?"
> because I bet we have other instances of this bug which we haven't
> noticed because the reset function happens to not misbehave on
> an inited-but-not-realized device...

Yeah that's probably my fault - I remember struggling quite a bit to get 
everything
to initialise correctly in the right order when I worked on this.

I tested first on cuda and then used the same pattern for pmu and mac_via so 
I'm not
surprised at all that the same problem appears in all three.


ATB,

Mark.