perf sort: Fix symbol sort output by separating unresolved samples by type
authorAnton Blanchard <anton@samba.org>
Wed, 31 Aug 2011 01:51:45 +0000 (11:51 +1000)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Fri, 23 Sep 2011 17:37:17 +0000 (14:37 -0300)
I took a profile that suggested 60% of total CPU time was in the
hypervisor:

...
    60.20%  [H] 0x33d43c
     4.43%  [k] ._spin_lock_irqsave
     1.07%  [k] ._spin_lock

Using perf stat to get the user/kernel/hypervisor breakdown contradicted
this.

The problem is we merge all unresolved samples into the one unknown
bucket. If add a comparison by sample type to sort__sym_cmp we get the
real picture:

...
    57.11%  [.] 0x80fbf63c
     4.43%  [k] ._spin_lock_irqsave
     1.07%  [k] ._spin_lock
     0.65%  [H] 0x33d43c

So it was almost all userspace, not hypervisor as the initial profile
suggested.

I found another issue while adding this. Symbol sorting sometimes shows
multiple entries for the unknown bucket:

...
    16.65%  [.] 0x6cd3a8
     7.25%  [.] 0x422460
     5.37%  [.] yylex
     4.79%  [.] malloc
     4.78%  [.] _int_malloc
     4.03%  [.] _int_free
     3.95%  [.] hash_source_code_string
     2.82%  [.] 0x532908
     2.64%  [.] 0x36b538
     0.94%  [H] 0x8000000000e132a4
     0.82%  [H] 0x800000000000e8b0

This happens because we aren't consistent with our sorting. On
one hand we check to see if both symbols match and for two unresolved
samples sym is NULL so we match:

        if (left->ms.sym == right->ms.sym)
                return 0;

On the other hand we use sample IP for unresolved samples when
comparing against a symbol:

       ip_l = left->ms.sym ? left->ms.sym->start : left->ip;
       ip_r = right->ms.sym ? right->ms.sym->start : right->ip;

This means unresolved samples end up spread across the rbtree and we
can't merge them all.

If we use cmp_null all unresolved samples will end up in the one bucket
and the output makes more sense:

...
    39.12%  [.] 0x36b538
     5.37%  [.] yylex
     4.79%  [.] malloc
     4.78%  [.] _int_malloc
     4.03%  [.] _int_free
     3.95%  [.] hash_source_code_string
     2.26%  [H] 0x800000000000e8b0

Acked-by: Eric B Munson <emunson@mgebm.net>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ian Munsie <imunsie@au1.ibm.com>
Link: http://lkml.kernel.org/r/20110831115145.4f598ab2@kryten
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/util/sort.c

index 401e220566fdab96d87cb1ae24e5a66dfb92992b..1ee8f1e40f18528db9cb2d766fd572f36736f426 100644 (file)
@@ -151,11 +151,17 @@ sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 {
        u64 ip_l, ip_r;
 
+       if (!left->ms.sym && !right->ms.sym)
+               return right->level - left->level;
+
+       if (!left->ms.sym || !right->ms.sym)
+               return cmp_null(left->ms.sym, right->ms.sym);
+
        if (left->ms.sym == right->ms.sym)
                return 0;
 
-       ip_l = left->ms.sym ? left->ms.sym->start : left->ip;
-       ip_r = right->ms.sym ? right->ms.sym->start : right->ip;
+       ip_l = left->ms.sym->start;
+       ip_r = right->ms.sym->start;
 
        return (int64_t)(ip_r - ip_l);
 }