Re: [Qemu-devel] [PULL 22/26] target/aarch64: optimize indirect branches

[Alex Bennée]

Alex Bennée <address@hidden> writes:

> Alex Bennée <address@hidden> writes:
>
>> Richard Henderson <address@hidden> writes:
>>
>>> From: "Emilio G. Cota" <address@hidden>
>>>
>>> Measurements:
>>>
>>> [Baseline performance is that before applying this and the previous
>>> commit]
>>
>> Sadly this has regressed my qemu-system-aarch64 EL2 run. It was slightly
>> masked by an unrelated assertion breakage which I had to fix. However
>> with this patch my boot hangs spinning all 4 threads. Once reverted
>> things work again.
>>
>> My command line:
>>
>>   timeout -k 1s --foreground 120s ./aarch64-softmmu/qemu-system-aarch64 
>> -machine type=virt -display none -m 4096 -cpu cortex-a57 -serial mon:stdio 
>> -netdev user,id=unet -device virtio-net-device,netdev=unet -drive 
>> file=/home/alex/lsrc/qemu/images/jessie-arm64.qcow2,id=myblock,index=0,if=none
>>  -device virtio-blk-device,drive=myblock -append "console=ttyAMA0 
>> root=/dev/vda1 systemd.unit=benchmark.service" -kernel 
>> /home/alex/lsrc/qemu/images/aarch64-current-linux-kernel-only.img -smp 4 
>> -machine gic-version=3 -machine virtualization=true -name debug-threads=on
>>
>> My tree with fix and revert:
>>
>>   https://github.com/stsquad/qemu/tree/debug/aarch64-hang
>>
>> I'm investigating now.
>
> Well this seems to be a case of hangs with -smp > 1 (which I guess was
> obvious seeing as the TCG threads seem to be spinning against each
> other).

So the minimum fix I've found to get things working again is:

void *HELPER(lookup_tb_ptr)(CPUArchState *env, target_ulong addr)
{
    CPUState *cpu = ENV_GET_CPU(env);
    TranslationBlock *tb;
    target_ulong cs_base, pc;
    uint32_t flags;

    tb = atomic_rcu_read(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(addr)]);
    if (likely(tb)) {
        cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
        if (likely(tb->pc == addr && tb->cs_base == cs_base &&
                   tb->flags == flags)) {
            goto found;
        }
        tb = tb_htable_lookup(cpu, addr, cs_base, flags);
        if (likely(tb)) {
            atomic_set(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(addr)], tb);
            /* goto found; */
        }
    }
    return tcg_ctx.code_gen_epilogue;
 found:
    qemu_log_mask_and_addr(CPU_LOG_EXEC, addr,
                           "Chain %p [%d: " TARGET_FMT_lx "] %s\n",
                           tb->tc_ptr, cpu->cpu_index, addr,
                           lookup_symbol(addr));
    return tb->tc_ptr;
}

Which I find very odd. It would imply that tb_htable_lookup is giving us
a bad result, except I would find it very unlikely what ever funky value
we stored in the jmp cache would never get hit again.

/me is very confused.

--
Alex Bennée