mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2024-09-12 05:48:40 +00:00
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>
This commit is contained in:
parent
ee8d7787e1
commit
35d48ddfc0
2 changed files with 14 additions and 12 deletions
|
@ -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;
|
||||
|
||||
|
|
|
@ -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;
|
||||
|
|
Loading…
Reference in a new issue