[CRIU] crash in pb_read_one?

[Tycho Andersen]

On Wed, Sep 17, 2014 at 06:27:14PM +0400, Pavel Emelyanov wrote:
> On 09/17/2014 06:05 PM, Tycho Andersen wrote:
> > Hi Pavel,
> > 
> > On Tue, Sep 16, 2014 at 02:50:09PM -0500, Tycho Andersen wrote:
> >>
> >> Yes, this is very sticky.
> > 
> > I think I've finally gotten it to a state where I understand
> > everything about the synchronization. I am having problems using
> > rst_mem_alloc and friends, though. The top two patches here are the
> > ones that are causing problems:
> > 
> > https://github.com/tych0/criu/commits/missing-pid
> > 
> > The one that segfaults is (I think?) closer to the right way to do
> > things, but the pointer in task_args is bad, so when we call
> > wait_helpers it segfaults.
> 
> You misuse the rst-mem API. The proper usage is:
> 
> In CRIU:
> 
> 	foo_pos = rst_mem_cpos(type);
> 	while (...) {
> 		foo = rst_mem_alloc(type, size);
> 		init_foo(foo);
> 	}
> 
> 	/* after rst_mem_remap() */
> 
> 	task_args->foo = rst_mem_remap_ptr(foo_pos, type);
> 
> In restorer blob:
> 
> 	use_foo(task_args->foo);
> 
> 
> In your patch you call rst_mem_remap_ptr not on _pos, but
> on the allocated mem.
> 
> And one more thing -- the rst_mem_grow_last is not required for
> the usage described, as you pin the beginning of your buffer
> first (with the rst_mem_cpos()), then grow it.

Excellent, thanks for the writeup!

Tycho

> 
> > Any thoughts are much appreciated,
> > 
> > Tycho
> > 
>