[Qemu-devel] TB monitoring

[Clemens Kolbitsch]

Guys,

I need a Qemu-internals expert to help me out here:

I'm trying to monitor execution of certain (user-land) TBs in a i386-softmmu 
system. For this, the cpu-main loop has been patched:

Before jumping into a TB's generated code, I first check its guest-virtual 
start  address whether it matches any of my monitored EIPs. Obviously, this 
only works for monitoring function starts (or other locations where a new TB 
is guaranteed to start) and with TB-chaining turned off. So far, this has been 
working great (for years even).

Now I tried to improve the performance of my monitoring system by doing the 
following: Each TB is associated with an additional 2 bits (in the "cflags"): 
"seen-before" and "monitored". Whenever I check a TB, I also set its seen-
before flag and update the monitored bit (if the TB contains a monitored start 
EIP).

Since a TB might be shared between two guest processes (the system is 
monitoring a Windows XP guest), I also save the a guest-process unique ID with 
the TB (using some bits from its CR3).

With this, I can skip checking a TB if its seen-before bit is set, but the 
monitored is clear. If the TB's guest-process-ID is different from the 
currently executing process-ID I just reset the seen-before bit (which forces 
the EIP analysis).

I have tried this on various programs, running inside my guest and it seems to 
be fast and working (almost ;-)). However, sometimes (*very* rarely), the 
above approach misses a monitored TB (skips analysis).

Could someone comment on my approach, maybe including some of the following 
thoughts:

1) is it possible that a TB grows ... i.e., it merges with a second TB where 
the second is monitored but the first one isn't?

2) are there any situations how a TB can be created without going through 
tb_alloc (which internally resets the cflags and thus my bits)?

Note that my code is based on Qemu 0.9.1 (if you wonder why I refer to certain 
old-style Qemu-internals).

Any help would be greatly appreciated!! If you need some more information or 
don't fully understand the problem, please don't hesitate to ask.

Thanks!!
Clemens