Re: [PATCH 2/5] perf kmem: Analyze page allocator events also
From: Ingo Molnar
Date: Tue Mar 24 2015 - 03:08:14 EST
* Joonsoo Kim <js1304@xxxxxxxxx> wrote:
> 2015-03-24 9:18 GMT+09:00 Namhyung Kim <namhyung@xxxxxxxxxx>:
> > On Tue, Mar 24, 2015 at 02:32:17AM +0900, Joonsoo Kim wrote:
> >> 2015-03-23 15:30 GMT+09:00 Namhyung Kim <namhyung@xxxxxxxxxx>:
> >> > The perf kmem command records and analyze kernel memory allocation
> >> > only for SLAB objects. This patch implement a simple page allocator
> >> > analyzer using kmem:mm_page_alloc and kmem:mm_page_free events.
> >> >
> >> > It adds two new options of --slab and --page. The --slab option is
> >> > for analyzing SLAB allocator and that's what perf kmem currently does.
> >> >
> >> > The new --page option enables page allocator events and analyze kernel
> >> > memory usage in page unit. Currently, 'stat --alloc' subcommand is
> >> > implemented only.
> >> >
> >> > If none of these --slab nor --page is specified, --slab is implied.
> >> >
> >> > # perf kmem stat --page --alloc --line 10
> >> >
> >> > -------------------------------------------------------------------------------------
> >> > Page | Total alloc (KB) | Hits | Order | Migration type | GFP flags
> >> > -------------------------------------------------------------------------------------
> >> > ffffea0015e48e00 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea0015e47400 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea001440f600 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea001440cc00 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea00140c6300 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea00140c5c00 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea00140c5000 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea00140c4f00 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea00140c4e00 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ffffea00140c4d00 | 16 | 1 | 2 | RECLAIMABLE | 00285250
> >> > ... | ... | ... | ... | ... | ...
> >> > -------------------------------------------------------------------------------------
> >>
> >> Tracepoint on mm_page_alloc print out pfn as well as pointer of struct page.
> >> How about printing pfn rather than pointer of struct page?
> >
> > I'd really like to have pfn rather than struct page. But I don't know
> > how to convert page pointer to pfn in userspace.
> >
> > The output of tracepoint via $debugfs/tracing/trace file is generated
> > from kernel-side, so it can easily have pfn from page pointer. But
> > tracepoint itself only saves page pointer and we need to convert/print
> > it in userspace.
>
> Ah...I didn't realize that perf don't use output of $debugfs/tracing/trace
> file. So, perf just uses raw trace buffer directly? If pfn is saved to
> the trace buffer, perf can print pfn rather than pointer of struct page?
>
> > Yes, perf script (or libtraceevent) shows pfn when printing those
> > events. But that's bogus since it cannot determine the size of the
> > struct page so the pointer arithmetic in open-coded page_to_pfn()
> > which is saved in the print_fmt of the tracepoint will end up with an
> > normal integer arithmatic.
>
> How about following change and making 'perf kmem' print pfn?
> If we store pfn on the trace buffer, we can print $debugfs/tracing/trace
> as is and 'perf kmem' can also print pfn.
>
> Thanks.
>
> diff --git a/include/trace/events/kmem.h b/include/trace/events/kmem.h
> index 4ad10ba..9dcfd0b 100644
> --- a/include/trace/events/kmem.h
> +++ b/include/trace/events/kmem.h
> @@ -199,22 +199,22 @@ TRACE_EVENT(mm_page_alloc,
> TP_ARGS(page, order, gfp_flags, migratetype),
>
> TP_STRUCT__entry(
> - __field( struct page *, page )
> + __field( unsigned long, pfn )
> __field( unsigned int, order )
> __field( gfp_t, gfp_flags )
> __field( int, migratetype )
> ),
>
> TP_fast_assign(
> - __entry->page = page;
> + __entry->pfn = page ? page_to_pfn(page) : -1;
> __entry->order = order;
> __entry->gfp_flags = gfp_flags;
> __entry->migratetype = migratetype;
> ),
>
> TP_printk("page=%p pfn=%lu order=%d migratetype=%d gfp_flags=%s",
> - __entry->page,
> - __entry->page ? page_to_pfn(__entry->page) : 0,
> + __entry->pfn != -1 ? pfn_to_page(__entry->pfn) : NULL,
> + __entry->pfn != -1 ? __entry->pfn : 0,
> __entry->order,
> __entry->migratetype,
> show_gfp_flags(__entry->gfp_flags))
Acked-by: Ingo Molnar <mingo@xxxxxxxxxx>
It would be very nice to make all the other page granular tracepoints
output pfn (which is a physical address that can be resolved to 'node'
and other properties), not 'struct page *' (which is a kernel resource
with little meaning to user-space tooling).
I.e. the following tracepoints:
triton:~/tip> git grep -E '__field.*struct page *' include/trace/
include/trace/events/filemap.h: __field(struct page *, page)
include/trace/events/kmem.h: __field( struct page *, page )
include/trace/events/kmem.h: __field( struct page *, page )
include/trace/events/kmem.h: __field( struct page *, page )
include/trace/events/kmem.h: __field( struct page *, page )
include/trace/events/kmem.h: __field( struct page *, page )
include/trace/events/pagemap.h: __field(struct page *, page )
include/trace/events/pagemap.h: __field(struct page *, page )
include/trace/events/vmscan.h: __field(struct page *, page)
there's very little breakage I can imagine: they have traced pointers
to 'struct page', which is a pretty opaque page identifier to
user-space, and they'll trace pfn's in the future, which still serves
as a page identifier.
One thing would be important: to do all these changes at once, to make
sure that the various page identifiers can be compared.
Also, we might keep the 'page' field name if anything relies on that -
but 'pfn' is even better.
Thanks,
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/