Re: [ptrace] please review follow fork/exec changes

On Mon, Jan 30, 2012 at 10:26:25AM -0800, Dmitry Mikulin wrote:
> 
> 
> On 01/28/2012 11:48 PM, Kostik Belousov wrote:
> >On Fri, Jan 27, 2012 at 10:12:13AM -0800, Dmitry Mikulin wrote:
> >>Attached are 4 separate patches for each somewhat independent  portion of
> >>the kernel work related to the follow-fork implementation.
> >Ok, as I said, I think that vfork-fork.patch is just wrong.
> 
> Gdb needs to be able to read/write process memory between the time the 
> child is forked and exec is called (in the vfork case). Without the change 
> it causes a kernel panic when gdb tries to read/write process memory. Since 
> my understanding of the kernel is a bit limited, it was the best I could do 
> at the time. I will send more details about the panic once I get a working 
> fbsd system again. Maybe there's a better way to deal with the panic.
Please provide more details, I am looking forward for the panic
message and backtrace.

> 
> >Lets postpone discussion of the orphan.patch for later.
> 
> OK.
> 
> >
> >The follow-fork.patch and follow-exec.patch make me wonder, and I
> >already expressed my doubts. IMO, all features, except one bug, are
> >already presented in the stock src.
> >
> >More, suggested follow-{fork,exec} patches break the SCE/SCX tracers
> >notification of fork and exec events, since TDB_FORK and TBD_EXEC flags
> >are consumed before syscall returns (I also said this previously).
> 
> 
> 
> 
> >
> >Namely, if the process is being debugged, the successfull [f]execve()
> >causes unconditional stop even. This makes PT_FOLLOW_EXEC unneccessary.
> >
> >Existing PT_FOLLOW_FORK implementation indeed has a bug, which was not
> >revealed by my testing during the development, because I only tested
> >SCE/SCX scenario. Namely, if PT_FOLLOW_FORK is requested, but the next
> >stop is not SCX, then follow-fork notification is not sent. After this
> >nit is fixed, PT_FOLLOW_FORK caller gets stops for the child creation.
> >Child is put into stop state as needed to not loose it.
> 
> I think this will fix only a part of the problem, the one that relates to 
> PT_CONTINUE.
> 
> I still need the change that forces a stop in both child and parent on 
> fork(). Without my changes the notification is generated in the child but 
> not in the parent. I need to be able to have both processes stopped in gdb 
> in order to clean up and detach from the parent, and initialize and attach 
> to the child. The main reason I need both processes stopped is that gdb has 
> to be able to read/write into both processes address space and registers.  
> Ideally I would like to have a single event generated for fork() at a point 
> where both child and parent are stopped and available for ptrace read/write 
> requests.  Do you think it's possible?
The lack of the notification for parent is exactly what the patch I
posted fixes. More exactly, it is the lack of notification for parent
with PT_CONTINUE loop. I will commit this today.

Regarding a single notification. Currently, parent arrives at the
syscall return code only after the child is attached to the debugger.
See the cv_wait() in kern_fork.c:739. In other words, if you get the
PL_FLAG_FORK, the child is already attached (at last it shall be). My
scescx.c code illustrates this well, IMO.

You still get a separate stop from the child, but I do not see how is it
harmful.