perf tools: Fix mmap limitations on 32-bit

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@davemloft.net>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20121110.141219.582924082787523608.davem@davemloft.net
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
index d5fb60760bacd6e9b5b6a18ccecb86962e4623f4..aa5e58255cba8bd5daddf5f44511f8a4ee1813a8 100644 (file)
--- 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);
@@ -1386,6 +1377,18 @@ 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)
@@ -1393,7 +1396,7 @@ 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  mmap_size;
-       char *buf, *mmaps[8];
+       char *buf, *mmaps[NUM_MMAPS];
        union perf_event *event;
        uint32_t size;
 
@@ -1408,7 +1411,7 @@ 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 cea133a6bdf1fd54061a3be27927fc32d3996184..c18fabdff19b8a12955cd4ca74cb52107bea528a 100644 (file)
--- a/tools/perf/util/session.h
+++ b/tools/perf/util/session.h
@@ -30,7 +30,6 @@ 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;