Re: [PATCH 07/15] mm: Add ___GFP_NOTRACE

[Tom Zanussi <tom.zanussi@linux.intel.com>]

On Mon, 2015-03-02 at 10:12 -0800, Alexei Starovoitov wrote:
> On Mon, Mar 2, 2015 at 10:03 AM, Tom Zanussi
> <tom.zanussi@linux.intel.com> wrote:
> > On Mon, 2015-03-02 at 09:58 -0800, Alexei Starovoitov wrote:
> >> On Mon, Mar 2, 2015 at 8:46 AM, Tom Zanussi <tom.zanussi@linux.intel.com> wrote:
> >> > On Mon, 2015-03-02 at 11:37 -0500, Steven Rostedt wrote:
> >> >> On Mon,  2 Mar 2015 10:01:00 -0600
> >> >> Tom Zanussi <tom.zanussi@linux.intel.com> wrote:
> >> >>
> >> >> > Add a gfp flag that allows kmalloc() et al to be used in tracing
> >> >> > functions.
> >> >> >
> >> >> > The problem with using kmalloc for tracing is that the tracing
> >> >> > subsystem should be able to trace kmalloc itself, which it can't do
> >> >> > directly because of paths like kmalloc()->trace_kmalloc()->kmalloc()
> >> >> > or kmalloc()->trace_mm_page_alloc()->kmalloc().
> >> >>
> >> >> This part I don't like at all. Why can't the memory be preallocated
> >> >> when the hist is created (the echo 'hist:...')?
> >> >>
> >> >
> >> > Yeah, I didn't like it either.  My original version did exactly what you
> >> > suggest and preallocated an array of entries to 'allocate' from in order
> >> > to avoid the problem.
> >> >
> >> > But I wanted to attempt to use the bpf_map directly, which already uses
> >> > kmalloc internally.  My fallback in case this wouldn't fly, which it
> >> > obviously won't, would be to add an option to have the bpf_map code
> >> > preallocate a maximum number of entries or pass in a client-owned array
> >> > for the purpose.  I'll do that if I don't hear any better ideas..
> >>
> >> Tom, I'm still reading through the patch set.
> >> Quick comment for the above.
> >> Currently there are two map types: array and hash.
> >> array type is pre-allocating all memory at map creation time.
> >> hash is allocating on demand.
> >
> > OK, so would it make sense to do the same for the hash type, or at least
> > add an option that does that?
> 
> I'm not sure what would be the meaning of hash map that has all
> elements pre-allocated...

The idea would be that instead of getting your individually kmalloc'ed
elements on-demand from kmalloc while in the handler, you'd get them
from a pool you've pre-allocated when you set up the table.  This could
be from a list of individual entries you've already kmalloc'ed ahead of
time, or from an array of n * sizeof(entry).

This would also allow you to avoid GFP_ATOMIC for those.

> As I'm reading your cover letter, I agree, we need to find a way
> to call kmalloc_notrace-like from tracepoints.
> Not sure that patch 8 style of duplicating the functions is clean.

No, it's horrible, but it does the job without changing the normal path
at all.

> Can we keep kmalloc/kfree as-is and do something like
> if (in_tracepoint()) check inside ftrace_raw_kmalloc*  ?

Yeah, that's essentially what TP_CONDITION() in patch 8 (Make kmem
memory allocation tracepoints conditional) does.

Tom

> so that kmalloc will be traced but calls to kmalloc from inside
> tracepoints will be automatically suppressed ?