On 11/10/2019 09:01, Peter Zijlstra wrote:
> On Fri, Oct 04, 2019 at 10:10:47AM +0200, Daniel Bristot de Oliveira wrote:
>> Currently, ftrace_rec entries are ordered inside the group of functions, but
>> "groups of function" are not ordered. So, the current int3 handler does a 
>> (*):
> We can insert a sort() of the vector right before doing
> text_poke_bp_batch() of course...

I agree!

What I tried to do earlier this week was to order the ftrace_pages in the
insertion [1], and so, while sequentially reading the pages with
do_for_each_ftrace_rec() we would already see the "ip"s in order.

As ftrace_pages are inserted only at boot and during a load of a module, this
would push the ordering for a very very slow path.

It works! But under the assumption that the address of functions in a module
does not intersect with the address of other modules/kernel, e.g.:

kernel:          module A:      module B:
[ 1, 2, 3, 4 ]   [ 7, 8, 9 ]    [ 15, 16, 19 ]

But this does not happen in practice, as I saw things like:

kernel:          module A:      module B:
[ 1, 2, 3, 4 ]   [ 7, 8, 18 ]   [ 15, 16, 19 ]
                         ^^ <--- greater than the first of the next

Is this expected?

At this point, I stopped working on it to give a time for my brain o think about
a better solution, also because Steve is reworking ftrace_pages to save some
space. So, it was better to wait.

But, yes, we will need [ as an optimization ] to sort the address right before
inserting them in the batch. Still, having the ftrace_pages ordered seems to be
a good thing, as in many cases, the ftrace_pages are disjoint sets.

[1] see ftrace_pages_insert() and ftrace_pages_start.

-- Daniel

Reply via email to