Re: tb_flush() calls causing long Windows XP boot times

["Alex Bennée" <alex.bennee@linaro.org>]

Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> writes:

> On 11/06/2021 19:22, Alex Bennée wrote:
>
> (added Gitlab on CC)
>
>> Paolo Bonzini <pbonzini@redhat.com> writes:
>> 
>>> On 11/06/21 17:01, Programmingkid wrote:
>>>> Hello Alex,
>>>> The good news is the source code to Windows XP is available
>>>> online:https://github.com/cryptoAlgorithm/nt5src
>>>
>>> It's leaked, so I doubt anybody who's paid to work on Linux or QEMU
>>> would touch that with a ten-foot pole.
>> Indeed.
>> Anyway what the OP could do is run QEMU with gdb and -d nochain and
>> stick a breakpoint (sic) in breakpoint_invalidate. Then each time it
>> hits you can examine the backtrace to cpu_loop_exec_tb and collect the
>> data from tb->pc. Then you will have a bunch of addresses in Windows
>> that keep triggering the behaviour. You can then re-run with -dfilter
>> and -d in_asm,cpu to get some sort of idea of what Windows is up to.
>
> I have been able to recreate this locally using my WinXP and it looks
> like during boot WinXP goes into a tight loop where it writes and
> clears a set of breakpoints via writes to DB7 which is what causes the
> very slow boot time.
>
> Once boot proceeds further into the login screen, the same code seems
> to called periodically once every second or so which has less of a
> performance impact.
>
>
> This gives a repeated set of outputs like this:
>
> ##### bpi @ 0x90
> ### dp7 add bp inst @ 0x8053cab8, index 1
> ##### bpi @ 0xa4
> ### dp7 add bp inst @ 0x8053cab8, index 2
> ##### bpi @ 0xff
> ### dp7 add bp inst @ 0x8053cab8, index 3
> ##### bpi @ 0xf

That's weird - maybe this is a misunderstanding of the x86 debug
registers but it looks like it's setting each one to all the same value. 

> ### dp7 remove bp inst @ 0x8053f58a, index 0
> ##### bpi @ 0x90
> ### dp7 remove bp inst @ 0x8053f58a, index 1
> ##### bpi @ 0xa4
> ### dp7 remove bp inst @ 0x8053f58a, index 2
> ##### bpi @ 0xff
> ### dp7 remove bp inst @ 0x8053f58a, index 3
> ...
> ...
> ### dp7 add bp inst @ 0x8053c960, index 0
> ##### bpi @ 0x90
> ### dp7 add bp inst @ 0x8053c960, index 1
> ##### bpi @ 0xa4
> ### dp7 add bp inst @ 0x8053c960, index 2
> ##### bpi @ 0xff
> ### dp7 add bp inst @ 0x8053c960, index 3
> ##### bpi @ 0xf
> ### dp7 remove bp inst @ 0x8053c730, index 0
> ##### bpi @ 0x90
> ### dp7 remove bp inst @ 0x8053c730, index 1
> ##### bpi @ 0xa4
> ### dp7 remove bp inst @ 0x8053c730, index 2
> ##### bpi @ 0xff
> ### dp7 remove bp inst @ 0x8053c730, index 3
> ...
> ...

I wonder if this is Windows check pointing itself by observing when it
gets to a particular place in the boot sequence. I guess we don't have
any symbols for the addresses it's setting?

>
> From a vanilla XP install the 2 main sections of code which alter the
> breakpoint registers are at 0x8053cab8 (enable) and 0x8053f58a
> (disable):

Ahh I misread - so those are the addresses of the routines and not where
it's sticking the breakpoint?

I notice from a bit of googling that there is a boot debugger. I wonder
if /nodebug in boot.ini stops this behaviour?

  https://docs.microsoft.com/en-us/troubleshoot/windows-server/performance/switch-options-for-boot-files

-- 
Alex Bennée