[CRIU] [PATCH] ptrace: Skip GETREGS on exits from syscalls when possible

[Pavel Emelyanov]

On 09/16/2014 09:19 PM, Andrew Vagin wrote:
> On Tue, Sep 16, 2014 at 08:56:38PM +0400, Pavel Emelyanov wrote:
>> On 09/16/2014 03:46 PM, Andrew Vagin wrote:
>>> On Tue, Sep 16, 2014 at 01:26:15PM +0400, Andrew Vagin wrote:
>>>> On Tue, Sep 16, 2014 at 12:27:07AM +0400, Pavel Emelyanov wrote:
>>>>> From: Pavel Emelyanov <xemul at parallels.com>
>>>>>
>>>>> The PTRACE_SYSCALL traps task twice -- first on enter into
>>>>> and then on exit from syscall. If we trace a single task (and
>>>>> we do it on dump two times per task) we may skip half of all
>>>>> getregs calls -- on exit we don't need them.
>>>>
>>>> Acked-by: Andrew Vagin <avagin at parallels.com>
>>>>
>>>> Can we use breakpoints to avoid step-by-step tracing?
>>>
>>> No, We can't. If sigreturn restarts syscalls, we doesn't have a place
>>> for breakpoint.
>>
>> How to treat the "[PATCH 3/3] x86: use breakpoints instead of tracing syscalls"
>> sent 3 hours (and 15 minutes) after this?
> 
> I explain there, that sigreturn() doesn't restart system calls. They are
> restarted by other code, which adjusts eip to call __NR_restart_syscall.

Can we put breakpoint one instruction before sigreturn? E.g. put artificial
sys_getpid() and catch the task in it?

However, not sure how useful it is, really. After we get to step-by-step
tasks they do 4 syscalls -- receive message, close log, close control
socket and the sigreturn itself. Maybe it makes more sense to cure tasks
from parasite in batch mode? And to stop_on_syscall() for may tasks,
like we do on restore?

> Look at get_task_regs() for more details.
> 
>>
>>> Sorry for talking with myself.
>>
>> You're always welcome!
>>
>>>>>
>>>>> Signed-off-by: Pavel Emelyanov <xemul at parallels.com>
>>>>> ---
>>>>>  parasite-syscall.c | 22 +++++++++++++++++++++-
>>>>>  1 file changed, 21 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/parasite-syscall.c b/parasite-syscall.c
>>>>> index 6f56617..266ef0b 100644
>>>>> --- a/parasite-syscall.c
>>>>> +++ b/parasite-syscall.c
>>>>> @@ -886,6 +886,10 @@ static int parasite_fini_seized(struct parasite_ctl *ctl)
>>>>>  	return 0;
>>>>>  }
>>>>>  
>>>>> +#define TRACE_ALL	1
>>>>> +#define TRACE_ENTER	2
>>>>> +#define TRACE_EXIT	3
>>>>> +
>>>>>  /*
>>>>>   * Trap tasks on the exit from the specified syscall
>>>>>   *
>>>>> @@ -897,6 +901,13 @@ int parasite_stop_on_syscall(int tasks, const int sys_nr)
>>>>>  	user_regs_struct_t regs;
>>>>>  	int status, ret;
>>>>>  	pid_t pid;
>>>>> +	/*
>>>>> +	 * The PTRACE_SYSCALL will trap task twice -- on
>>>>> +	 * enter into and on exit from syscall. If we trace
>>>>> +	 * a single task, we may skip half of all getregs
>>>>> +	 * calls -- on exit we don't need them.
>>>>> +	 */
>>>>> +	int trace = (tasks == 1 ? TRACE_ENTER : TRACE_ALL);
>>>>>  
>>>>>  	/* Stop all threads on the enter point in sys_rt_sigreturn */
>>>>>  	while (tasks) {
>>>>> @@ -916,6 +927,15 @@ int parasite_stop_on_syscall(int tasks, const int sys_nr)
>>>>>  			pr_err("%d\n", status);
>>>>>  			return -1;
>>>>>  		}
>>>>> +
>>>>> +		if (trace == TRACE_EXIT) {
>>>>> +			trace = TRACE_ENTER;
>>>>> +			pr_debug("`- Expecting exit\n");
>>>>> +			goto goon;
>>>>> +		}
>>>>> +		if (trace == TRACE_ENTER)
>>>>> +			trace = TRACE_EXIT;
>>>>> +
>>>>>  		ret = ptrace_get_regs(pid, &regs);
>>>>>  		if (ret) {
>>>>>  			pr_perror("ptrace");
>>>>> @@ -949,7 +969,7 @@ int parasite_stop_on_syscall(int tasks, const int sys_nr)
>>>>>  			tasks--;
>>>>>  			continue;
>>>>>  		}
>>>>> -
>>>>> +goon:
>>>>>  		ret = ptrace(PTRACE_SYSCALL, pid, NULL, NULL);
>>>>>  		if (ret) {
>>>>>  			pr_perror("ptrace");
>>>>> -- 
>>>>> 1.8.4.2
>>>>>
>>>>> _______________________________________________
>>>>> CRIU mailing list
>>>>> CRIU at openvz.org
>>>>> https://lists.openvz.org/mailman/listinfo/criu
>>>> _______________________________________________
>>>> CRIU mailing list
>>>> CRIU at openvz.org
>>>> https://lists.openvz.org/mailman/listinfo/criu
>>> _______________________________________________
>>> CRIU mailing list
>>> CRIU at openvz.org
>>> https://lists.openvz.org/mailman/listinfo/criu
>>> .
>>>
>>
> .
>