Re: [PATCH 05/11] ftrace: create memcache for hash entries

[Changbin Du]

On Mon, Aug 26, 2019 at 09:44:37AM +0200, Peter Zijlstra wrote:
> On Sun, Aug 25, 2019 at 09:23:24PM +0800, Changbin Du wrote:
> > When CONFIG_FTRACE_FUNC_PROTOTYPE is enabled, thousands of
> > ftrace_func_entry instances are created. So create a dedicated
> > memcache to enhance performance.
> > 
> > Signed-off-by: Changbin Du <changbin.du@xxxxxxxxx>
> > ---
> >  kernel/trace/ftrace.c | 17 ++++++++++++++++-
> >  1 file changed, 16 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
> > index a314f0768b2c..cfcb8dad93ea 100644
> > --- a/kernel/trace/ftrace.c
> > +++ b/kernel/trace/ftrace.c
> > @@ -94,6 +94,8 @@ struct ftrace_ops *function_trace_op __read_mostly = &ftrace_list_end;
> >  /* What to set function_trace_op to */
> >  static struct ftrace_ops *set_function_trace_op;
> >  
> > +struct kmem_cache *hash_entry_cache;
> > +
> >  static bool ftrace_pids_enabled(struct ftrace_ops *ops)
> >  {
> >  	struct trace_array *tr;
> > @@ -1169,7 +1171,7 @@ static int add_hash_entry(struct ftrace_hash *hash, unsigned long ip,
> >  {
> >  	struct ftrace_func_entry *entry;
> >  
> > -	entry = kmalloc(sizeof(*entry), GFP_KERNEL);
> > +	entry = kmem_cache_alloc(hash_entry_cache, GFP_KERNEL);
> >  	if (!entry)
> >  		return -ENOMEM;
> >  
> > @@ -6153,6 +6155,15 @@ void __init ftrace_init(void)
> >  	if (ret)
> >  		goto failed;
> >  
> > +	hash_entry_cache = kmem_cache_create("ftrace-hash",
> > +					     sizeof(struct ftrace_func_entry),
> > +					     sizeof(struct ftrace_func_entry),
> > +					     0, NULL);
> > +	if (!hash_entry_cache) {
> > +		pr_err("failed to create ftrace hash entry cache\n");
> > +		goto failed;
> > +	}
> 
> Wait what; you already have then in the binary image, now you're
> allocating extra memory for each of them?
>
No, here we only allocate ftrace hash entries. The prototype data is not copied.
The entry->priv points to prototype data in binary.

> Did you look at what ORC does? Is the binary search really not fast
> enough?
For ftrace, binary search is not enough. Just like the hash tables
(ftrace_graph_notrace_hash, ftrace_graph_hash) we already have which is used to
filter traced functions.


-- 
Cheers,
Changbin Du