tegra-profiler: record offset of the stack pointer

Tegra Profiler: when collecting backtraces, record
remaining data stack size.

Bug 1584533
Bug 1598009

Change-Id: I608ab73f8e1b7da84221a17a782080fdf5598111
Signed-off-by: Igor Nabirushkin <inabirushkin@nvidia.com>
Reviewed-on: http://git-master/r/658431
(cherry picked from commit 803c1a41e9b7ebcbf297bd7ce5caeed1d1f800ce)
Reviewed-on: http://git-master/r/672039
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Venkat Moganty <vmoganty@nvidia.com>

5 files changed, 57 insertions, 7 deletions

diff --git a/drivers/misc/tegra-profiler/hrt.c b/drivers/misc/tegra-profiler/hrt.c
index 1850f3f8003c..b7bd00c795ac 100644
--- a/drivers/misc/tegra-profiler/hrt.c
+++ b/drivers/misc/tegra-profiler/hrt.c
@@ -186,6 +186,9 @@ static void put_header(void)
 	if (hrt.use_arch_timer)
 		hdr->reserved |= QUADD_HDR_USE_ARCH_TIMER;
 
+	if (hrt.get_stack_offset)
+		hdr->reserved |= QUADD_HDR_STACK_OFFSET;
+
 	if (pmu)
 		nr_events += pmu->get_current_events(events, max_events);
 
@@ -295,6 +298,28 @@ static int read_source(struct quadd_event_source_interface *source,
 	return nr_events;
 }
 
+static long
+get_stack_offset(struct task_struct *task,
+		 struct pt_regs *regs,
+		 struct quadd_callchain *cc)
+{
+	unsigned long sp;
+	struct vm_area_struct *vma;
+	struct mm_struct *mm = task->mm;
+
+	if (!regs || !mm)
+		return -ENOMEM;
+
+	sp = cc->nr > 0 ? cc->curr_sp :
+		quadd_user_stack_pointer(regs);
+
+	vma = find_vma(mm, sp);
+	if (!vma)
+		return -ENOMEM;
+
+	return vma->vm_end - sp;
+}
+
 static void
 read_all_sources(struct pt_regs *regs, struct task_struct *task)
 {
@@ -355,6 +380,11 @@ read_all_sources(struct pt_regs *regs, struct task_struct *task)
 
 	s->reserved = 0;
 
+	cc->nr = 0;
+	cc->curr_sp = 0;
+	cc->curr_fp = 0;
+	cc->curr_pc = 0;
+
 	if (ctx->param.backtrace) {
 		cc->unw_method = hrt.unw_method;
 		bt_size = quadd_get_user_callchain(user_regs, cc, ctx, task);
@@ -394,6 +424,15 @@ read_all_sources(struct pt_regs *regs, struct task_struct *task)
 	}
 	s->callchain_nr = bt_size;
 
+	if (hrt.get_stack_offset) {
+		long offset = get_stack_offset(task, user_regs, cc);
+		if (offset > 0) {
+			u32 off = offset >> 2;
+			off = min_t(u32, off, 0xffff);
+			extra_data |= off << QUADD_SED_STACK_OFFSET_SHIFT;
+		}
+	}
+
 	record_data.record_type = QUADD_RECORD_TYPE_SAMPLE;
 
 	s->events_flags = 0;
@@ -627,6 +666,9 @@ int quadd_hrt_start(void)
 
 	pr_info("timer: %s\n", hrt.use_arch_timer ? "arch" : "monotonic clock");
 
+	hrt.get_stack_offset =
+		(extra & QUADD_PARAM_EXTRA_STACK_OFFSET) ? 1 : 0;
+
 	put_header();
 
 	if (extra & QUADD_PARAM_EXTRA_GET_MMAP) {
diff --git a/drivers/misc/tegra-profiler/hrt.h b/drivers/misc/tegra-profiler/hrt.h
index 47662fbf7682..3361c61e6f3f 100644
--- a/drivers/misc/tegra-profiler/hrt.h
+++ b/drivers/misc/tegra-profiler/hrt.h
@@ -65,6 +65,7 @@ struct quadd_hrt_ctx {
 	int use_arch_timer;
 
 	unsigned int unw_method;
+	int get_stack_offset;
 };
 
 #define QUADD_HRT_MIN_FREQ	100
diff --git a/drivers/misc/tegra-profiler/main.c b/drivers/misc/tegra-profiler/main.c
index 1e967a419bef..f0116af49035 100644
--- a/drivers/misc/tegra-profiler/main.c
+++ b/drivers/misc/tegra-profiler/main.c
@@ -518,11 +518,11 @@ static int __init quadd_module_init(void)
 							  QUADD_MAX_COUNTERS);
 		ctx.pmu_info.nr_supported_events = nr_events;
 
-		pr_info("PMU: amount of events: %d\n", nr_events);
+		pr_debug("PMU: amount of events: %d\n", nr_events);
 
 		for (i = 0; i < nr_events; i++)
-			pr_info("PMU event: %s\n",
-				quadd_get_event_str(events[i]));
+			pr_debug("PMU event: %s\n",
+				 quadd_get_event_str(events[i]));
 	}
 
 #ifdef CONFIG_CACHE_L2X0
@@ -543,7 +543,7 @@ static int __init quadd_module_init(void)
 			pr_info("pl310 event: %s\n",
 				quadd_get_event_str(events[i]));
 	} else {
-		pr_info("PL310 not found\n");
+		pr_debug("PL310 not found\n");
 	}
 
 	ctx.hrt = quadd_hrt_init(&ctx);
diff --git a/drivers/misc/tegra-profiler/version.h b/drivers/misc/tegra-profiler/version.h
index 88210df23026..3836f6683f55 100644
--- a/drivers/misc/tegra-profiler/version.h
+++ b/drivers/misc/tegra-profiler/version.h
@@ -18,7 +18,7 @@
 #ifndef __QUADD_VERSION_H
 #define __QUADD_VERSION_H
 
-#define QUADD_MODULE_VERSION		"1.84"
+#define QUADD_MODULE_VERSION		"1.85"
 #define QUADD_MODULE_BRANCH		"Dev"
 
 #endif	/* __QUADD_VERSION_H */
diff --git a/include/linux/tegra_profiler.h b/include/linux/tegra_profiler.h
index ad62d9a8e918..4225e7e9274e 100644
--- a/include/linux/tegra_profiler.h
+++ b/include/linux/tegra_profiler.h
@@ -19,8 +19,8 @@
 
 #include <linux/ioctl.h>
 
-#define QUADD_SAMPLES_VERSION	30
-#define QUADD_IO_VERSION	15
+#define QUADD_SAMPLES_VERSION	31
+#define QUADD_IO_VERSION	16
 
 #define QUADD_IO_VERSION_DYNAMIC_RB		5
 #define QUADD_IO_VERSION_RB_MAX_FILL_COUNT	6
@@ -33,6 +33,7 @@
 #define QUADD_IO_VERSION_ARCH_TIMER_OPT		13
 #define QUADD_IO_VERSION_DATA_MMAP		14
 #define QUADD_IO_VERSION_BT_LOWER_BOUND		15
+#define QUADD_IO_VERSION_STACK_OFFSET		16
 
 #define QUADD_SAMPLE_VERSION_THUMB_MODE_FLAG	17
 #define QUADD_SAMPLE_VERSION_GROUP_SAMPLES	18
@@ -46,6 +47,7 @@
 #define QUADD_SAMPLE_VERSION_SCHED_SAMPLES	28
 #define QUADD_SAMPLE_VERSION_HDR_UNW_METHOD	29
 #define QUADD_SAMPLE_VERSION_HDR_ARCH_TIMER	30
+#define QUADD_SAMPLE_VERSION_STACK_OFFSET	31
 
 #define QUADD_MMAP_HEADER_VERSION		1
 
@@ -192,6 +194,9 @@ enum {
 #define QUADD_SED_UNW_METHOD_SHIFT	1
 #define QUADD_SED_UNW_METHOD_MASK	(0x07 << QUADD_SED_UNW_METHOD_SHIFT)
 
+#define QUADD_SED_STACK_OFFSET_SHIFT	4
+#define QUADD_SED_STACK_OFFSET_MASK	(0xffff << QUADD_SED_STACK_OFFSET_SHIFT)
+
 enum {
 	QUADD_UNW_TYPE_FP = 0,
 	QUADD_UNW_TYPE_UT,
@@ -302,6 +307,7 @@ struct quadd_debug_data {
 #define QUADD_HDR_UNW_METHOD_MASK	(0x07 << QUADD_HDR_UNW_METHOD_SHIFT)
 
 #define QUADD_HDR_USE_ARCH_TIMER	(1 << 3)
+#define QUADD_HDR_STACK_OFFSET		(1 << 4)
 
 struct quadd_header_data {
 	u16 magic;
@@ -354,6 +360,7 @@ enum {
 #define QUADD_PARAM_EXTRA_BT_UNWIND_TABLES	(1 << 2)
 #define QUADD_PARAM_EXTRA_BT_MIXED		(1 << 3)
 #define QUADD_PARAM_EXTRA_USE_ARCH_TIMER	(1 << 4)
+#define QUADD_PARAM_EXTRA_STACK_OFFSET		(1 << 5)
 
 struct quadd_parameters {
 	u32 freq;