Re: [RFC][PATCH 11/14] function_graph: Convert ret_stack to a series of longs

[Joel Fernandes]

On Mon, Nov 26, 2018 at 11:26:03AM -0500, Steven Rostedt wrote:
> On Tue, 27 Nov 2018 01:07:55 +0900
> Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> 
> > > > --- a/include/linux/sched.h
> > > > +++ b/include/linux/sched.h
> > > > @@ -1119,7 +1119,7 @@ struct task_struct {
> > > >  	int				curr_ret_depth;
> > > >  
> > > >  	/* Stack of return addresses for return function tracing: */
> > > > -	struct ftrace_ret_stack		*ret_stack;
> > > > +	unsigned long			*ret_stack;
> > > >  
> > > >  	/* Timestamp for last schedule: */
> > > >  	unsigned long long		ftrace_timestamp;
> > > > diff --git a/kernel/trace/fgraph.c b/kernel/trace/fgraph.c
> > > > index 9b85638ecded..1389fe39f64c 100644
> > > > --- a/kernel/trace/fgraph.c
> > > > +++ b/kernel/trace/fgraph.c
> > > > @@ -23,6 +23,17 @@
> > > >  #define ASSIGN_OPS_HASH(opsname, val)
> > > >  #endif
> > > >  
> > > > +#define FGRAPH_RET_SIZE (sizeof(struct ftrace_ret_stack))
> > > > +#define FGRAPH_RET_INDEX (ALIGN(FGRAPH_RET_SIZE, sizeof(long)) / sizeof(long))
> > > > +#define SHADOW_STACK_SIZE (FTRACE_RETFUNC_DEPTH * FGRAPH_RET_SIZE)
> > > > +#define SHADOW_STACK_INDEX			\
> > > > +	(ALIGN(SHADOW_STACK_SIZE, sizeof(long)) / sizeof(long))
> > > > +#define SHADOW_STACK_MAX_INDEX (SHADOW_STACK_INDEX - FGRAPH_RET_INDEX)
> > > > +
> > > > +#define RET_STACK(t, index) ((struct ftrace_ret_stack *)(&(t)->ret_stack[index]))
> > > > +#define RET_STACK_INC(c) ({ c += FGRAPH_RET_INDEX; })
> > > > +#define RET_STACK_DEC(c) ({ c -= FGRAPH_RET_INDEX; })
> > > > +  
> > > [...]  
> > > > @@ -514,7 +531,7 @@ void ftrace_graph_init_task(struct task_struct *t)
> > > >  
> > > >  void ftrace_graph_exit_task(struct task_struct *t)
> > > >  {
> > > > -	struct ftrace_ret_stack	*ret_stack = t->ret_stack;
> > > > +	unsigned long *ret_stack = t->ret_stack;
> > > >  
> > > >  	t->ret_stack = NULL;
> > > >  	/* NULL must become visible to IRQs before we free it: */
> > > > @@ -526,12 +543,10 @@ void ftrace_graph_exit_task(struct task_struct *t)
> > > >  /* Allocate a return stack for each task */
> > > >  static int start_graph_tracing(void)
> > > >  {
> > > > -	struct ftrace_ret_stack **ret_stack_list;
> > > > +	unsigned long **ret_stack_list;
> > > >  	int ret, cpu;
> > > >  
> > > > -	ret_stack_list = kmalloc_array(FTRACE_RETSTACK_ALLOC_SIZE,
> > > > -				       sizeof(struct ftrace_ret_stack *),
> > > > -				       GFP_KERNEL);
> > > > +	ret_stack_list = kmalloc(SHADOW_STACK_SIZE, GFP_KERNEL);
> > > >    
> > > 
> > > I had dumped the fgraph size related macros to understand the patch better, I
> > > got:
> > > [    0.909528] val of FGRAPH_RET_SIZE is 40
> > > [    0.910250] val of FGRAPH_RET_INDEX is 5
> > > [    0.910866] val of FGRAPH_ARRAY_SIZE is 16
> > > [    0.911488] val of FGRAPH_ARRAY_MASK is 255
> > > [    0.912134] val of FGRAPH_MAX_INDEX is 16
> > > [    0.912751] val of FGRAPH_INDEX_SHIFT is 8
> > > [    0.913382] val of FGRAPH_FRAME_SIZE is 168
> > > [    0.914033] val of FGRAPH_FRAME_INDEX is 21
> > >                       FTRACE_RETFUNC_DEPTH is 50
> > > [    0.914686] val of SHADOW_STACK_SIZE is 8400
> > > 
> > > I had a concern about memory overhead per-task. It seems the total memory
> > > needed per task for the stack is 8400 bytes (with my configuration with
> > > FUNCTION_PROFILE
> > > turned off).
> > > 
> > > Where as before it would be 32 * 40 = 1280 bytes. That looks like ~7 times
> > > more than before.  
> > 
> > Hmm, this seems too big... I thought the shadow-stack size should be
> > smaller than 1 page (4kB). Steve, can we give a 4k page for shadow stack
> > and define FTRACE_RETFUNC_DEPTH = 4096 / FGRAPH_RET_SIZE ?
> 
> For the first pass, I decided not to worry about the size. It made the
> code less complex :-)
> 
> Yes, I plan on working on making the size of the stack smaller, but
> that will probably be added on patches to do so.

Cool, sounds good.

> > > On my system with ~4000 threads, that becomes ~32MB which seems a bit
> > > wasteful especially if there was only one or 2 function graph callbacks
> > > registered and most of the callback array in the stack isn't used.
> 
> Note, all 4000 threads could be doing those trace backs, and if you are
> doing full function graph tracing, it will use a lot.

But I think each of the threads will only use N entries in the callback array
where N is the number of function graph callback users who registered, right?
So the remaining total-N allocated callback array entries per thread will not
be used.

> > > Could we make the array size configurable at compile time and start it with a
> > > small number like 4 or 6?  
> > 
> > Or, we can introduce online setting :)
> 
> Yes, that can easily be added. I didn't try to make this into the
> perfect solution, I wanted a solid one first, and then massage it into
> something that is more efficient, both with memory consumption and
> performance.
> 
> Joel and Masami, thanks for the feedback.

I agree the first step is good so far. Looking forward to the patches, thanks
a lot,

 - Joel