Performance degradation for WinXP boot time after b55f54bc

added accel: TCG kindBug target: i386 labels

@rth7680 - Can you triage further?

On my Mac I was able to use the Instruments application to profile what qemu-system-i386 was doing when Windows XP was starting up. Here is a picture of the results. Hope it gives us some idea what is going on.

In the file accel/tcg/translate-all.c, I added code to the tb_flush() function to tell us how many times it is called. It looks like this:

void tb_flush(CPUState *cpu)
{
    // count how many times this function is called
    static int counter = 0;
    if (counter % 10 == 0) {
        printf("tb_flush count: %d\n", counter);
    }
    counter++;


    if (tcg_enabled()) {
        unsigned tb_flush_count = qatomic_mb_read(&tb_ctx.tb_flush_count);

        if (cpu_in_exclusive_context(cpu)) {
            do_tb_flush(cpu, RUN_ON_CPU_HOST_INT(tb_flush_count));
        } else {
            async_safe_run_on_cpu(cpu, do_tb_flush,
                                  RUN_ON_CPU_HOST_INT(tb_flush_count));
        }
    }
}

After compiling the changes I saw that Windows XP calls this function over 263,830 times! I tried Windows 7 and it didn't call the function.

Good news, I found a way to speed up Windows XP's boot time. See the patch below:

From 4554aaf5bfbddf37620eb1d5bdff16a672ab562b Mon Sep 17 00:00:00 2001
From: John Arbuckle <programmingkidx@gmail.com>
Date: Wed, 9 Jun 2021 21:30:27 -0400
Subject: [PATCH] Speeds up Windows XP boot time from 3 minutes to 20 seconds

---
 cpu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cpu.c b/cpu.c
index bfbe5a66f9..297c2e4281 100644
--- a/cpu.c
+++ b/cpu.c
@@ -253,7 +253,7 @@ static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
      * Flush the whole TB cache to force re-translation of such TBs.
      * This is heavyweight, but we're debugging anyway.
      */
-    tb_flush(cpu);
+    /* tb_flush(cpu); */
 }
 #endif
 
-- 
2.24.3 (Apple Git-128)

I does need more testing but so far it seems good. I was able to play Pinball without problem.

It looks like it's a too simple implementation. At least it should be reverted to the state before b55f54bc, because now debugging can become broken.

Hi Максим, you say debugging is broken. Could you send us a test someone could use to reproduce the problem? I'm thinking some x86 assembly program that test debugging by writing to the debug registers is what we need to see.

Using the /nodebug option in the boot.ini file did not have any effect for me.

marked this issue as related to #404 (closed)

I dunno if gitlab forwards comments from linked issues, but

https://patchew.org/QEMU/20210712154004.1410832-1-richard.henderson@linaro.org/

Richard Henderson commented:

I dunno if gitlab forwards comments from linked issues, but

https://patchew.org/QEMU/20210712154004.1410832-1-richard.henderson@linaro.org/

I did not receive an email. I did apply and test out your patches. They work! Windows XP boots so much faster now.