Re: [Qemu-devel] [PATCH] mttcg: Set jmp_env to handle exit from tb_gen_code

[Pranith Kumar]

On Mon, Feb 20, 2017 at 7:35 PM, Alex Bennée <address@hidden> wrote:
>
> Pranith Kumar <address@hidden> writes:
>
>> Alex Bennée writes:
>>
>>> Pranith Kumar <address@hidden> writes:
>>>
>>>> tb_gen_code() can exit execution using cpu_exit_loop() when it cannot
>>>> allocate new tb's. To handle this, we need to properly set the jmp_env
>>>> pointer ahead of calling tb_gen_code().
>>>>
>>>> CC:Alex Bennée <address@hidden>
>>>> CC: Richard Henderson <address@hidden>
>>>> Signed-off-by: Pranith Kumar <address@hidden>
>>>> ---
>>>>  cpu-exec.c | 23 +++++++++++------------
>>>>  1 file changed, 11 insertions(+), 12 deletions(-)
>>>>
>>>> diff --git a/cpu-exec.c b/cpu-exec.c
>>>> index 97d79612d9..4b70988b24 100644
>>>> --- a/cpu-exec.c
>>>> +++ b/cpu-exec.c
>>>> @@ -236,23 +236,22 @@ static void cpu_exec_step(CPUState *cpu)
>>>>
>>>>      cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
>>>>      tb_lock();
>>>> -    tb = tb_gen_code(cpu, pc, cs_base, flags,
>>>> -                     1 | CF_NOCACHE | CF_IGNORE_ICOUNT);
>>>> -    tb->orig_tb = NULL;
>>>> -    tb_unlock();
>>>> -
>>>> -    cc->cpu_exec_enter(cpu);
>>>> -
>>>
>>> It occurs to me we are also diverging in our locking pattern from
>>> tb_find which takes mmap_lock first. This is a NOP for system emulation
>>> but needed for user-emulation (for which we can do cpu_exec_step but not
>>> cpu_exec_nocache).
>>
>> Right. So we have to take the mmap_lock() before calling
>> tb_gen_code(). However, this lock is released in the error path before 
>> calling
>> cpu_loop_exit() if allocation of a new tb fails. The following is what I have
>> after merging with the previous EXCP_ATOMIC handling patch.
>
> Hmm we are a start/end_exclusive() though so what could we be racing
> against that also wants the mmap_lock()? Could it be held by anything at
> this point as every user needs to be woken up?
>

No, I don't think any other thread holds the mmap/tb lock at this
point. However, we still need to take this lock since the functions we
are calling expect us to hold it and assert otherwise.


-- 
Pranith