This is the mail archive of the
mailing list for the GDB project.
Re: [Discussion/Prec] The record speed up plan (Make speed like without prec)
I have some doubts, which I would like to get clarified.
1) In this approach, we fork and create child the moment user does 'record' right ?
2) Now we parent/chile and let other freely run, right ?
3) when user wants to do reverse we record parent/child insns right ?
4) how are we planning to handle events like (sending signals, closing files, freeing memory etc...), and some special instructions which changes some hw configuration may be...
will not all these happen twice ?
----- Original Message ----
From: Hui Zhu <email@example.com>
Cc: Michael Snyder <firstname.lastname@example.org>; paawan oza <email@example.com>; Eli Zaretskii <firstname.lastname@example.org>
Sent: Fri, May 21, 2010 1:12:22 PM
Subject: Re: [Discussion/Prec] The record speed up plan (Make speed like without prec)
On Wed, May 19, 2010 at 15:18, Hui Zhu <email@example.com> wrote:
> This is a demo.
> Still not support segment register, system call and some others.
Before I send RFA for these patches, I think there are 12 parts need complete:
1. Find a better name.
Maybe it can be change to checkpoint record or something.
2. The code that check the record_ptid and inferior_ptid is exec to
the same place need to more works.
Now, it check the stop_pc, all the registers and the frame.
But if the code like:
for (i = 1; i< 99; i++)
If the inferior stop in the go() and replay. The stop_pc , reg and
frame will be same.
So I think we need add the check for memory change. The inferior exec
to the same place, the memory that changed must be same. So Check the
each record_mem addr of record_ptid and inferior_ptid. If all of them
are same. I think it is really same.
3. System call support.
Like what I said in the plan. We need support system call with catch point.
4. handle special insns that we will not know what will happen after it exec.
Like what I said in the plan. Give them up.
5. Signal support.
Need do special record (maybe fork), when record_ptid inferior exec to
this point, record this signal.
6. User change the memory when record.
It need special record like signal.
7. In demo, record.c call ptrace to kill the record_ptid and call
linux_nat_switch_fork to switch ptid.
I think both of them need a target interface.
8. When gdb get record_list from record_ptid. Ctrl-c can stop it.
9. Let user can use goto command to the begin and end of record_list
without get each record_list from record_ptid. It will make reverse
debug more speed up.
10. record_end have a insn_num to help goto. I am not sure we can
support it now. Maybe we can close it when record pre is opened.
11. record_insn_max, I need make it works OK with pre prec or close it
when record pre is opened.
12. Make record save works OK with record pre.
> On Fri, Apr 30, 2010 at 21:23, Hui Zhu <firstname.lastname@example.org> wrote:
>> I think the record speed is the biggest trouble of prec.
>> After I did a long think and a lot of test around with it. I got a
>> idea. Actually, I have began the code work.
>> I found that the big trouble is prec let the inferior just step. It
>> make inferior speed very low. Because the setp need a lot of context
>> So I think let inferior continue can make it speed up. But How to
>> record the change of each step?
>> Some physicists said all the things in the world have execution rules.
>> So use the current stat of this thing, we will get what will happen
>> in the future. Looks most of rules are still not found. :)
>> But lucky for us that insns exec rules we know. So most of insns
>> (There a some special, I will talk it later), if we have the a
>> inferior value(memory and reg), we can get the each value of next
>> So if we can record the all the value of a inferior A(or all the value
>> that will be change, but to get it will need parse the insns that will
>> be exec, this is not easy.) , we can let the inferior exec without
>> step. If the user want reverse exec, get the each step value from A.
>> Then the record speed will very faster than before.
>> But this way have a 2 question.
>> 1. How to record all the status of a inferiorï For the linux,
>> checkpoint already use fork to record the inferior. So prec will use
>> it too.
>> And when we want get the old status of inferior step by step, we can
>> let the forked process step by step. That will easy by parse the insn
>> and know what will happen.
>> 2. How to handle special insns that we will not know what will happen
>> after it exec?
>> The first type of this insns is system call. Linux have catchpoint
>> that make inferior stop before and after syscall. Then we can record
>> the change after the system call.
>> The other insn is like rdtsc, I don't know howto get the feature value
>> of this type. My current idea with them is give up all the record
>> after this insn.
>> If user need, insert special breakpoint for it. Next time, inferior
>> will stop on this insn, then prec can record the value after it exec.
>> BTW, I call this new function pre_record, I agree with you if you said
>> this name is ugly. :)
>> Please tell me your opinions about my idea. That will help me a lot. Thanks.