[tip:perfcounters/core] perf_counter, x86: Fix kernel-space call-chains

From: tip-bot for Ingo Molnar
Date: Mon Jun 15 2009 - 04:05:37 EST


Commit-ID: 1f3d4ecee6fe08ca7ed1ef131a8a8903d613d730
Gitweb: http://git.kernel.org/tip/1f3d4ecee6fe08ca7ed1ef131a8a8903d613d730
Author: Ingo Molnar <mingo@xxxxxxx>
AuthorDate: Mon, 15 Jun 2009 09:57:59 +0200
Committer: Ingo Molnar <mingo@xxxxxxx>
CommitDate: Mon, 15 Jun 2009 09:58:00 +0200

perf_counter, x86: Fix kernel-space call-chains

Kernel-space call-chains were trimmed at the first entry because
we never processed anything beyond the first stack context.

Allow the backtrace to jump from NMI to IRQ stack then to task stack
and finally user-space stack.

Also calculate the stack and bp variables correctly so that the
stack walker does not exit early.

We can get deep traces as a result, visible in perf report -D output:

0x32af0 [0xe0]: PERF_EVENT (IP, 5): 15134: 0xffffffff815225fd period: 1
... chain: u:2, k:22, nr:24
..... 0: 0xffffffff815225fd
..... 1: 0xffffffff810ac51c
..... 2: 0xffffffff81018e29
..... 3: 0xffffffff81523939
..... 4: 0xffffffff81524b8f
..... 5: 0xffffffff81524bd9
..... 6: 0xffffffff8105e498
..... 7: 0xffffffff8152315a
..... 8: 0xffffffff81522c3a
..... 9: 0xffffffff810d9b74
..... 10: 0xffffffff810dbeec
..... 11: 0xffffffff810dc3fb

This is a 22-entries kernel-space chain.

(We still only record reliable stack entries.)

Cc: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
Cc: Mike Galbraith <efault@xxxxxx>
Cc: Paul Mackerras <paulus@xxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@xxxxxxx>


---
arch/x86/kernel/cpu/perf_counter.c | 22 +++++++++-------------
1 files changed, 9 insertions(+), 13 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 09d8cb6..6d5e7cf 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -1575,8 +1575,8 @@ static void backtrace_warning(void *data, char *msg)

static int backtrace_stack(void *data, char *name)
{
- /* Don't bother with IRQ stacks for now */
- return -1;
+ /* Process all stacks: */
+ return 0;
}

static void backtrace_address(void *data, unsigned long addr, int reliable)
@@ -1594,6 +1594,8 @@ static const struct stacktrace_ops backtrace_ops = {
.address = backtrace_address,
};

+#include "../dumpstack.h"
+
static void
perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
{
@@ -1601,26 +1603,20 @@ perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
char *stack;
int nr = entry->nr;

- callchain_store(entry, instruction_pointer(regs));
+ callchain_store(entry, regs->ip);

stack = ((char *)regs + sizeof(struct pt_regs));
#ifdef CONFIG_FRAME_POINTER
- bp = frame_pointer(regs);
+ get_bp(bp);
#else
bp = 0;
#endif

- dump_trace(NULL, regs, (void *)stack, bp, &backtrace_ops, entry);
+ dump_trace(NULL, regs, (void *)&stack, bp, &backtrace_ops, entry);

entry->kernel = entry->nr - nr;
}

-
-struct stack_frame {
- const void __user *next_fp;
- unsigned long return_address;
-};
-
static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
{
int ret;
@@ -1652,7 +1648,7 @@ perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
callchain_store(entry, regs->ip);

while (entry->nr < MAX_STACK_DEPTH) {
- frame.next_fp = NULL;
+ frame.next_frame = NULL;
frame.return_address = 0;

if (!copy_stack_frame(fp, &frame))
@@ -1662,7 +1658,7 @@ perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
break;

callchain_store(entry, frame.return_address);
- fp = frame.next_fp;
+ fp = frame.next_frame;
}

entry->user = entry->nr - nr;
--
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/