tegra-profiler: backtraces from the kernel context

Tegra Profiler: collect backtraces from the kernel context

Bug 1394804

Change-Id: I1db7435c9e1fc753dd8c02252076287572f5f5af
Signed-off-by: Igor Nabirushkin <inabirushkin@nvidia.com>
Reviewed-on: http://git-master/r/324091
(cherry picked from commit f7213f6e694089f866ccfcbca02c1e61648048db)
Reviewed-on: http://git-master/r/340016
Reviewed-by: Harry Hong <hhong@nvidia.com>
Tested-by: Harry Hong <hhong@nvidia.com>

6 files changed, 55 insertions, 25 deletions

diff --git a/drivers/misc/tegra-profiler/backtrace.c b/drivers/misc/tegra-profiler/backtrace.c
index 340f4faa1405..3191def82ce0 100644
--- a/drivers/misc/tegra-profiler/backtrace.c
+++ b/drivers/misc/tegra-profiler/backtrace.c
@@ -17,7 +17,9 @@
 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 
 #include <linux/module.h>
-#include <asm-generic/uaccess.h>
+#include <linux/uaccess.h>
+#include <linux/sched.h>
+#include <linux/mm.h>
 
 #include <linux/tegra_profiler.h>
 
@@ -111,7 +113,7 @@ quadd_get_user_callchain(struct pt_regs *regs,
 
 	callchain_data->nr = 0;
 
-	if (!regs || !user_mode(regs) || !mm)
+	if (!regs || !mm)
 		return 0;
 
 	if (thumb_mode(regs))
@@ -126,12 +128,18 @@ quadd_get_user_callchain(struct pt_regs *regs,
 		return 0;
 
 	vma = find_vma(mm, sp);
+	if (!vma)
+		return 0;
+
 	if (check_vma_address(fp, vma))
 		return 0;
 
-	if (__copy_from_user_inatomic(&reg, (unsigned long __user *)fp,
-				      sizeof(unsigned long)))
+	if (probe_kernel_address(fp, reg)) {
+		pr_warn_once("frame error: sp/fp: %#lx/%#lx, pc/lr: %#lx/%#lx, vma: %#lx-%#lx\n",
+			     sp, fp, regs->ARM_pc, regs->ARM_lr,
+			     vma->vm_start, vma->vm_end);
 		return 0;
+	}
 
 	if (thumb_mode(regs)) {
 		if (reg <= fp || check_vma_address(reg, vma))
diff --git a/drivers/misc/tegra-profiler/hrt.c b/drivers/misc/tegra-profiler/hrt.c
index 0066f3af5520..8bc30727d9bb 100644
--- a/drivers/misc/tegra-profiler/hrt.c
+++ b/drivers/misc/tegra-profiler/hrt.c
@@ -25,6 +25,7 @@
 #include <linux/cpu.h>
 #include <linux/ratelimit.h>
 #include <asm/irq_regs.h>
+#include <linux/ptrace.h>
 
 #include <linux/tegra_profiler.h>
 
@@ -241,6 +242,7 @@ static void read_source(struct quadd_event_source_interface *source,
 	struct quadd_cpu_context *cpu_ctx = this_cpu_ptr(hrt.cpu_ctx);
 	struct quadd_callchain *callchain_data = &cpu_ctx->callchain_data;
 	struct quadd_ctx *quadd_ctx = hrt.quadd_ctx;
+	struct pt_regs *user_regs;
 
 	if (!source)
 		return;
@@ -256,8 +258,14 @@ static void read_source(struct quadd_event_source_interface *source,
 	if (atomic_read(&cpu_ctx->nr_active) == 0)
 		return;
 
-	if (user_mode(regs) && hrt.quadd_ctx->param.backtrace) {
-		callchain_nr = quadd_get_user_callchain(regs, callchain_data);
+	if (user_mode(regs))
+		user_regs = regs;
+	else
+		user_regs = task_pt_regs(current);
+
+	if (hrt.quadd_ctx->param.backtrace) {
+		callchain_nr =
+			quadd_get_user_callchain(user_regs, callchain_data);
 		if (callchain_nr > 0) {
 			extra_data = (char *)cpu_ctx->callchain_data.callchain;
 			extra_length = callchain_nr * sizeof(u32);
diff --git a/drivers/misc/tegra-profiler/main.c b/drivers/misc/tegra-profiler/main.c
index e8856d302d3b..f1ed75cf5833 100644
--- a/drivers/misc/tegra-profiler/main.c
+++ b/drivers/misc/tegra-profiler/main.c
@@ -246,6 +246,7 @@ static int set_parameters(struct quadd_parameters *param, uid_t *debug_app_uid)
 static void get_capabilities(struct quadd_comm_cap *cap)
 {
 	int i, event;
+	unsigned int extra = 0;
 	struct quadd_events_cap *events_cap = &cap->events_cap;
 
 	cap->pmu = ctx.pmu ? 1 : 0;
@@ -357,6 +358,9 @@ static void get_capabilities(struct quadd_comm_cap *cap)
 	cap->tegra_lp_cluster = quadd_is_cpu_with_lp_cluster();
 	cap->power_rate = 1;
 	cap->blocked_read = 1;
+
+	extra |= QUADD_COMM_CAP_EXTRA_BT_KERNEL_CTX;
+	cap->reserved[QUADD_COMM_CAP_IDX_EXTRA] = extra;
 }
 
 void quadd_get_state(struct quadd_module_state *state)
diff --git a/drivers/misc/tegra-profiler/quadd_proc.c b/drivers/misc/tegra-profiler/quadd_proc.c
index dbde7041fa04..980a810f4e40 100644
--- a/drivers/misc/tegra-profiler/quadd_proc.c
+++ b/drivers/misc/tegra-profiler/quadd_proc.c
@@ -56,47 +56,50 @@ static int show_capabilities(struct seq_file *f, void *offset)
 {
 	struct quadd_comm_cap *cap = &ctx->cap;
 	struct quadd_events_cap *event = &cap->events_cap;
+	unsigned int extra = cap->reserved[QUADD_COMM_CAP_IDX_EXTRA];
 
-	seq_printf(f, "pmu:                    %s\n",
+	seq_printf(f, "pmu:                           %s\n",
 		   YES_NO(cap->pmu));
-	seq_printf(f, "tegra 3 LP cluster:     %s\n",
+	seq_printf(f, "tegra 3 LP cluster:            %s\n",
 		   YES_NO(cap->tegra_lp_cluster));
-	seq_printf(f, "power rate samples:     %s\n",
+	seq_printf(f, "power rate samples:            %s\n",
 		   YES_NO(cap->power_rate));
 
-	seq_printf(f, "l2 cache:               %s\n",
+	seq_printf(f, "l2 cache:                      %s\n",
 		   YES_NO(cap->l2_cache));
 	if (cap->l2_cache) {
-		seq_printf(f, "Multiple l2 events:     %s\n",
+		seq_printf(f, "multiple l2 events:            %s\n",
 			   YES_NO(cap->l2_multiple_events));
 	}
 
-	seq_printf(f, "support polling mode:   %s\n",
+	seq_printf(f, "support polling mode:          %s\n",
 		   YES_NO(cap->blocked_read));
+	seq_printf(f, "backtrace from the kernel ctx: %s\n",
+		   YES_NO(extra & QUADD_COMM_CAP_EXTRA_BT_KERNEL_CTX));
 
 	seq_puts(f, "\n");
 	seq_puts(f, "Supported events:\n");
-	seq_printf(f, "cpu_cycles:             %s\n",
+	seq_printf(f, "cpu_cycles:                    %s\n",
 		   YES_NO(event->cpu_cycles));
-	seq_printf(f, "instructions:           %s\n",
+	seq_printf(f, "instructions:                  %s\n",
 		   YES_NO(event->instructions));
-	seq_printf(f, "branch_instructions:    %s\n",
+	seq_printf(f, "branch_instructions:           %s\n",
 		   YES_NO(event->branch_instructions));
-	seq_printf(f, "branch_misses:          %s\n",
+	seq_printf(f, "branch_misses:                 %s\n",
 		   YES_NO(event->branch_misses));
-	seq_printf(f, "bus_cycles:             %s\n",
+	seq_printf(f, "bus_cycles:                    %s\n",
 		   YES_NO(event->bus_cycles));
-	seq_printf(f, "l1_dcache_read_misses:  %s\n",
+	seq_printf(f, "l1_dcache_read_misses:         %s\n",
 		   YES_NO(event->l1_dcache_read_misses));
-	seq_printf(f, "l1_dcache_write_misses: %s\n",
+	seq_printf(f, "l1_dcache_write_misses:        %s\n",
 		   YES_NO(event->l1_dcache_write_misses));
-	seq_printf(f, "l1_icache_misses:       %s\n",
+	seq_printf(f, "l1_icache_misses:              %s\n",
 		   YES_NO(event->l1_icache_misses));
-	seq_printf(f, "l2_dcache_read_misses:  %s\n",
+	seq_printf(f, "l2_dcache_read_misses:         %s\n",
 		   YES_NO(event->l2_dcache_read_misses));
-	seq_printf(f, "l2_dcache_write_misses: %s\n",
+	seq_printf(f, "l2_dcache_write_misses:        %s\n",
 		   YES_NO(event->l2_dcache_write_misses));
-	seq_printf(f, "l2_icache_misses:       %s\n",
+	seq_printf(f, "l2_icache_misses:              %s\n",
 		   YES_NO(event->l2_icache_misses));
 
 	return 0;
diff --git a/drivers/misc/tegra-profiler/version.h b/drivers/misc/tegra-profiler/version.h
index 1b2ecc045aca..32b074609292 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		"1.36"
+#define QUADD_MODULE_VERSION		"1.37"
 #define QUADD_MODULE_BRANCH		"Dev"
 
 #endif	/* __QUADD_VERSION_H */
diff --git a/include/linux/tegra_profiler.h b/include/linux/tegra_profiler.h
index 1fda9594ce94..ed21f75e9556 100644
--- a/include/linux/tegra_profiler.h
+++ b/include/linux/tegra_profiler.h
@@ -20,11 +20,12 @@
 #include <linux/ioctl.h>
 
 #define QUADD_SAMPLES_VERSION	17
-#define QUADD_IO_VERSION	7
+#define QUADD_IO_VERSION	8
 
 #define QUADD_IO_VERSION_DYNAMIC_RB		5
 #define QUADD_IO_VERSION_RB_MAX_FILL_COUNT	6
 #define QUADD_IO_VERSION_MOD_STATE_STATUS_FIELD	7
+#define QUADD_IO_VERSION_BT_KERNEL_CTX		8
 
 #define QUADD_SAMPLE_VERSION_THUMB_MODE_FLAG	17
 
@@ -275,6 +276,12 @@ struct quadd_events_cap {
 		l2_icache_misses:1;
 };
 
+enum {
+	QUADD_COMM_CAP_IDX_EXTRA = 0,
+};
+
+#define QUADD_COMM_CAP_EXTRA_BT_KERNEL_CTX	(1 << 0)
+
 struct quadd_comm_cap {
 	u32	pmu:1,
 		power_rate:1,