Re: [PATCH] Fix perf mmap limitations on 32-bit.

From: David Miller
Date: Wed Dec 05 2012 - 13:33:46 EST


From: David Miller <davem@xxxxxxxxxxxxx>
Date: Sat, 10 Nov 2012 14:12:19 -0500 (EST)


Ping?

> This is a suggested patch to fix the bug I reported at:
>
> http://marc.info/?l=linux-kernel&m=135033028924652&w=2
>
> Essentially, there is a hard requirement that when perf analyzes a
> trace, it must have the entire thing mmap()'d.
>
> Therefore the scheme used on 32-bit where we have a fixed (8) number
> of 32MB mmaps, and cycle through them, simply does not work.
>
> One of the reasons this requirement exists is because the iterators
> maintain references to perf entry objects and those references don't
> just simply go away when this mmap code decides to cycle an old mmap
> area out and reuse it. At this point, those entry pointers now point
> to garbage resulting in unpredictable behavior and crashes.
>
> It is better to try to mmap() as much as we can and if we do actually
> run into address space limitations, the failure of the mmap() call
> will indicate that and stop processing.
>
> I noticed that perf_session->mmap_window is set to a constant in one
> location, and only used in one other location. So I got rid of it
> altogether.
>
> So we adjust the size of the mmaps[] array to the maximum we could
> need. On 64-bit we only need one slot. On 32-bit we could need
> up to 128 (128 * 32MB == 4GB).
>
> I've verified that this allows a large (~600MB) perf.data file to
> be analyzed properly with a 32-bit perf binary, which previously
> was not possible.
>
> Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
>
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index 8cdd232..2cd3cc3 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -128,15 +128,6 @@ struct perf_session *perf_session__new(const char *filename, int mode,
> goto out;
>
> memcpy(self->filename, filename, len);
> - /*
> - * On 64bit we can mmap the data file in one go. No need for tiny mmap
> - * slices. On 32bit we use 32MB.
> - */
> -#if BITS_PER_LONG == 64
> - self->mmap_window = ULLONG_MAX;
> -#else
> - self->mmap_window = 32 * 1024 * 1024ULL;
> -#endif
> self->machines = RB_ROOT;
> self->repipe = repipe;
> INIT_LIST_HEAD(&self->ordered_samples.samples);
> @@ -1369,6 +1360,18 @@ fetch_mmaped_event(struct perf_session *session,
> return event;
> }
>
> +/*
> + * On 64bit we can mmap the data file in one go. No need for tiny mmap
> + * slices. On 32bit we use 32MB.
> + */
> +#if BITS_PER_LONG == 64
> +#define MMAP_SIZE ULLONG_MAX
> +#define NUM_MMAPS 1
> +#else
> +#define MMAP_SIZE (32 * 1024 * 1024ULL)
> +#define NUM_MMAPS 128
> +#endif
> +
> int __perf_session__process_events(struct perf_session *session,
> u64 data_offset, u64 data_size,
> u64 file_size, struct perf_tool *tool)
> @@ -1376,7 +1379,7 @@ int __perf_session__process_events(struct perf_session *session,
> u64 head, page_offset, file_offset, file_pos, progress_next;
> int err, mmap_prot, mmap_flags, map_idx = 0;
> size_t page_size, mmap_size;
> - char *buf, *mmaps[8];
> + char *buf, *mmaps[NUM_MMAPS];
> union perf_event *event;
> uint32_t size;
>
> @@ -1393,7 +1396,7 @@ int __perf_session__process_events(struct perf_session *session,
>
> progress_next = file_size / 16;
>
> - mmap_size = session->mmap_window;
> + mmap_size = MMAP_SIZE;
> if (mmap_size > file_size)
> mmap_size = file_size;
>
> diff --git a/tools/perf/util/session.h b/tools/perf/util/session.h
> index dd64261..903966b 100644
> --- a/tools/perf/util/session.h
> +++ b/tools/perf/util/session.h
> @@ -29,7 +29,6 @@ struct ordered_samples {
> struct perf_session {
> struct perf_header header;
> unsigned long size;
> - unsigned long mmap_window;
> struct machine host_machine;
> struct rb_root machines;
> struct perf_evlist *evlist;
--
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/