misc: tegra-profiler: add unwind reason codes

Unwinding: store individual URC codes for each method.

Bug 1624134

Change-Id: I3b2045f9c9147354f3440e326fd3aeccb5e0458d
Signed-off-by: Igor Nabirushkin <inabirushkin@nvidia.com>
Reviewed-on: http://git-master/r/717848
(cherry picked from commit e5ceff53d63e668a19d36196823f6c185ce48e88)
Reviewed-on: http://git-master/r/748086
GVS: Gerrit_Virtual_Submit
Reviewed-by: Andrey Trachenko <atrachenko@nvidia.com>
Reviewed-by: Winnie Hsu <whsu@nvidia.com>

11 files changed, 142 insertions, 152 deletions

diff --git a/drivers/misc/tegra-profiler/backtrace.c b/drivers/misc/tegra-profiler/backtrace.c
index 2dc39b1d018c..2d947615978d 100644
--- a/drivers/misc/tegra-profiler/backtrace.c
+++ b/drivers/misc/tegra-profiler/backtrace.c
@@ -29,6 +29,12 @@
 #include "hrt.h"
 #include "tegra.h"
 
+static inline int
+is_table_unwinding(struct quadd_callchain *cc)
+{
+	return cc->um.ut || cc->um.dwarf;
+}
+
 unsigned long
 quadd_user_stack_pointer(struct pt_regs *regs)
 {
@@ -87,12 +93,12 @@ quadd_callchain_store(struct quadd_callchain *cc,
 	unsigned long low_addr = cc->hrt->low_addr;
 
 	if (ip < low_addr || !validate_pc_addr(ip, sizeof(unsigned long))) {
-		cc->unw_rc = QUADD_URC_PC_INCORRECT;
+		cc->urc_fp = QUADD_URC_PC_INCORRECT;
 		return 0;
 	}
 
 	if (cc->nr >= QUADD_MAX_STACK_DEPTH) {
-		cc->unw_rc = QUADD_URC_LEVEL_TOO_DEEP;
+		cc->urc_fp = QUADD_URC_LEVEL_TOO_DEEP;
 		return 0;
 	}
 
@@ -130,7 +136,7 @@ user_backtrace(struct pt_regs *regs,
 		return NULL;
 
 	if (__copy_from_user_inatomic(&value, tail, sizeof(unsigned long))) {
-		cc->unw_rc = QUADD_URC_EACCESS;
+		cc->urc_fp = QUADD_URC_EACCESS;
 		return NULL;
 	}
 
@@ -144,7 +150,7 @@ user_backtrace(struct pt_regs *regs,
 
 		if (__copy_from_user_inatomic(&value_lr, tail + 1,
 					      sizeof(value_lr))) {
-			cc->unw_rc = QUADD_URC_EACCESS;
+			cc->urc_fp = QUADD_URC_EACCESS;
 			return NULL;
 		}
 
@@ -155,7 +161,7 @@ user_backtrace(struct pt_regs *regs,
 		/* gcc arm frame */
 		if (__copy_from_user_inatomic(&value_fp, tail - 1,
 					      sizeof(value_fp))) {
-			cc->unw_rc = QUADD_URC_EACCESS;
+			cc->urc_fp = QUADD_URC_EACCESS;
 			return NULL;
 		}
 
@@ -175,7 +181,7 @@ user_backtrace(struct pt_regs *regs,
 	if (nr_added == 0)
 		return NULL;
 
-	if (cc->unw_method == QUADD_UNW_METHOD_MIXED &&
+	if (is_table_unwinding(cc) &&
 	    is_ex_entry_exist(regs, value_lr, task))
 		return NULL;
 
@@ -193,10 +199,10 @@ get_user_callchain_fp(struct pt_regs *regs,
 	struct mm_struct *mm = task->mm;
 
 	cc->nr = 0;
-	cc->unw_rc = QUADD_URC_FP_INCORRECT;
+	cc->urc_fp = QUADD_URC_FP_INCORRECT;
 
 	if (!regs || !mm) {
-		cc->unw_rc = QUADD_URC_FAILURE;
+		cc->urc_fp = QUADD_URC_FAILURE;
 		return 0;
 	}
 
@@ -209,7 +215,7 @@ get_user_callchain_fp(struct pt_regs *regs,
 
 	vma = find_vma(mm, sp);
 	if (!vma) {
-		cc->unw_rc = QUADD_URC_SP_INCORRECT;
+		cc->urc_fp = QUADD_URC_SP_INCORRECT;
 		return 0;
 	}
 
@@ -218,7 +224,7 @@ get_user_callchain_fp(struct pt_regs *regs,
 
 	if (probe_kernel_address(fp, reg)) {
 		pr_warn_once("%s: failed for address: %#lx\n", __func__, fp);
-		cc->unw_rc = QUADD_URC_EACCESS;
+		cc->urc_fp = QUADD_URC_EACCESS;
 		return 0;
 	}
 
@@ -239,7 +245,7 @@ get_user_callchain_fp(struct pt_regs *regs,
 					&value,
 					(unsigned long __user *)fp + 1,
 					sizeof(unsigned long))) {
-				cc->unw_rc = QUADD_URC_EACCESS;
+				cc->urc_fp = QUADD_URC_EACCESS;
 				return 0;
 			}
 
@@ -279,16 +285,16 @@ __user_backtrace(struct pt_regs *regs,
 	struct vm_area_struct *vma;
 	unsigned long __user *tail;
 
-	cc->unw_rc = QUADD_URC_FP_INCORRECT;
+	cc->urc_fp = QUADD_URC_FP_INCORRECT;
 
 	if (!mm) {
-		cc->unw_rc = QUADD_URC_FAILURE;
+		cc->urc_fp = QUADD_URC_FAILURE;
 		return cc->nr;
 	}
 
 	vma = find_vma(mm, cc->curr_sp);
 	if (!vma) {
-		cc->unw_rc = QUADD_URC_SP_INCORRECT;
+		cc->urc_fp = QUADD_URC_SP_INCORRECT;
 		return cc->nr;
 	}
 
@@ -316,7 +322,7 @@ user_backtrace_compat(struct pt_regs *regs,
 		return NULL;
 
 	if (__copy_from_user_inatomic(&value, tail, sizeof(value))) {
-		cc->unw_rc = QUADD_URC_EACCESS;
+		cc->urc_fp = QUADD_URC_EACCESS;
 		return NULL;
 	}
 
@@ -330,7 +336,7 @@ user_backtrace_compat(struct pt_regs *regs,
 
 		if (__copy_from_user_inatomic(&value_lr, tail + 1,
 					      sizeof(value_lr))) {
-			cc->unw_rc = QUADD_URC_EACCESS;
+			cc->urc_fp = QUADD_URC_EACCESS;
 			return NULL;
 		}
 
@@ -341,7 +347,7 @@ user_backtrace_compat(struct pt_regs *regs,
 		/* gcc arm frame */
 		if (__copy_from_user_inatomic(&value_fp, tail - 1,
 					      sizeof(value_fp))) {
-			cc->unw_rc = QUADD_URC_EACCESS;
+			cc->urc_fp = QUADD_URC_EACCESS;
 			return NULL;
 		}
 
@@ -361,7 +367,7 @@ user_backtrace_compat(struct pt_regs *regs,
 	if (nr_added == 0)
 		return NULL;
 
-	if (cc->unw_method == QUADD_UNW_METHOD_MIXED &&
+	if (is_table_unwinding(cc) &&
 	    is_ex_entry_exist(regs, value_lr, task))
 		return NULL;
 
@@ -379,10 +385,10 @@ get_user_callchain_fp_compat(struct pt_regs *regs,
 	struct mm_struct *mm = task->mm;
 
 	cc->nr = 0;
-	cc->unw_rc = QUADD_URC_FP_INCORRECT;
+	cc->urc_fp = QUADD_URC_FP_INCORRECT;
 
 	if (!regs || !mm) {
-		cc->unw_rc = QUADD_URC_FAILURE;
+		cc->urc_fp = QUADD_URC_FAILURE;
 		return 0;
 	}
 
@@ -395,7 +401,7 @@ get_user_callchain_fp_compat(struct pt_regs *regs,
 
 	vma = find_vma(mm, sp);
 	if (!vma) {
-		cc->unw_rc = QUADD_URC_SP_INCORRECT;
+		cc->urc_fp = QUADD_URC_SP_INCORRECT;
 		return 0;
 	}
 
@@ -404,7 +410,7 @@ get_user_callchain_fp_compat(struct pt_regs *regs,
 
 	if (probe_kernel_address((unsigned long)fp, reg)) {
 		pr_warn_once("%s: failed for address: %#x\n", __func__, fp);
-		cc->unw_rc = QUADD_URC_EACCESS;
+		cc->urc_fp = QUADD_URC_EACCESS;
 		return 0;
 	}
 
@@ -425,7 +431,7 @@ get_user_callchain_fp_compat(struct pt_regs *regs,
 					&value,
 					(u32 __user *)(fp + sizeof(u32)),
 					sizeof(value))) {
-				cc->unw_rc = QUADD_URC_EACCESS;
+				cc->urc_fp = QUADD_URC_EACCESS;
 				return 0;
 			}
 
@@ -465,16 +471,16 @@ __user_backtrace_compat(struct pt_regs *regs,
 	struct vm_area_struct *vma;
 	u32 __user *tail;
 
-	cc->unw_rc = QUADD_URC_FP_INCORRECT;
+	cc->urc_fp = QUADD_URC_FP_INCORRECT;
 
 	if (!mm) {
-		cc->unw_rc = QUADD_URC_FAILURE;
+		cc->urc_fp = QUADD_URC_FAILURE;
 		return cc->nr;
 	}
 
 	vma = find_vma(mm, cc->curr_sp);
 	if (!vma) {
-		cc->unw_rc = QUADD_URC_SP_INCORRECT;
+		cc->urc_fp = QUADD_URC_SP_INCORRECT;
 		return cc->nr;
 	}
 
@@ -494,7 +500,7 @@ __get_user_callchain_fp(struct pt_regs *regs,
 		      struct task_struct *task)
 {
 	if (cc->nr > 0) {
-		if (cc->unw_rc == QUADD_URC_LEVEL_TOO_DEEP)
+		if (cc->urc_fp == QUADD_URC_LEVEL_TOO_DEEP)
 			return cc->nr;
 
 #ifdef CONFIG_ARM64
@@ -518,44 +524,22 @@ __get_user_callchain_fp(struct pt_regs *regs,
 }
 
 static unsigned int
-get_user_callchain_ut(struct pt_regs *regs,
-		      struct quadd_callchain *cc,
-		      struct task_struct *task)
-{
-	int nr_prev;
-	unsigned long prev_sp;
-
-	do {
-		nr_prev = cc->nr;
-		prev_sp = cc->curr_sp;
-
-		quadd_get_user_cc_dwarf(regs, cc, task);
-		if (nr_prev > 0 && cc->nr == nr_prev)
-			break;
-
-		nr_prev = cc->nr;
-
-		quadd_get_user_cc_arm32_ehabi(regs, cc, task);
-	} while (nr_prev != cc->nr &&
-		 (cc->nr <= 1 || cc->curr_sp > prev_sp));
-
-	return cc->nr;
-}
-
-static unsigned int
 get_user_callchain_mixed(struct pt_regs *regs,
 		      struct quadd_callchain *cc,
 		      struct task_struct *task)
 {
 	int nr_prev;
 	unsigned long prev_sp;
+	struct quadd_unw_methods *um = &cc->um;
 
 	do {
 		nr_prev = cc->nr;
 		prev_sp = cc->curr_sp;
 
-		quadd_get_user_cc_dwarf(regs, cc, task);
-		quadd_get_user_cc_arm32_ehabi(regs, cc, task);
+		if (um->dwarf)
+			quadd_get_user_cc_dwarf(regs, cc, task);
+		if (um->ut)
+			quadd_get_user_cc_arm32_ehabi(regs, cc, task);
 
 		if (nr_prev != cc->nr) {
 			if (cc->nr > 1 &&
@@ -565,7 +549,8 @@ get_user_callchain_mixed(struct pt_regs *regs,
 			continue;
 		}
 
-		__get_user_callchain_fp(regs, cc, task);
+		if (um->fp)
+			__get_user_callchain_fp(regs, cc, task);
 	} while (nr_prev != cc->nr &&
 		 (cc->nr <= 1 || cc->curr_sp > prev_sp));
 
@@ -578,12 +563,12 @@ quadd_get_user_callchain(struct pt_regs *regs,
 			 struct quadd_ctx *ctx,
 			 struct task_struct *task)
 {
-	unsigned int method = cc->unw_method;
-
 	cc->nr = 0;
 
 	if (!regs) {
-		cc->unw_rc = QUADD_URC_FAILURE;
+		cc->urc_fp = QUADD_URC_FAILURE;
+		cc->urc_ut = QUADD_URC_FAILURE;
+		cc->urc_dwarf = QUADD_URC_FAILURE;
 		return 0;
 	}
 
@@ -598,25 +583,11 @@ quadd_get_user_callchain(struct pt_regs *regs,
 	cc->cs_64 = 0;
 #endif
 
-	cc->unw_rc = 0;
-
-	switch (method) {
-	case QUADD_UNW_METHOD_FP:
-		__get_user_callchain_fp(regs, cc, task);
-		break;
+	cc->urc_fp = QUADD_URC_NONE;
+	cc->urc_ut = QUADD_URC_NONE;
+	cc->urc_dwarf = QUADD_URC_NONE;
 
-	case QUADD_UNW_METHOD_EHT:
-		get_user_callchain_ut(regs, cc, task);
-		break;
-
-	case QUADD_UNW_METHOD_MIXED:
-		get_user_callchain_mixed(regs, cc, task);
-		break;
-
-	case QUADD_UNW_METHOD_NONE:
-	default:
-		break;
-	}
+	get_user_callchain_mixed(regs, cc, task);
 
 	return cc->nr;
 }
diff --git a/drivers/misc/tegra-profiler/backtrace.h b/drivers/misc/tegra-profiler/backtrace.h
index 5ee544cdd496..724c614a9408 100644
--- a/drivers/misc/tegra-profiler/backtrace.h
+++ b/drivers/misc/tegra-profiler/backtrace.h
@@ -27,6 +27,14 @@
 
 struct quadd_hrt_ctx;
 
+struct quadd_unw_methods {
+	unsigned int
+		fp:1,
+		ut:1,
+		ut_ce:1,
+		dwarf:1;
+};
+
 struct quadd_callchain {
 	int nr;
 
@@ -39,8 +47,11 @@ struct quadd_callchain {
 
 	int cs_64;
 
-	unsigned int unw_method;
-	unsigned int unw_rc;
+	struct quadd_unw_methods um;
+
+	unsigned int urc_fp;
+	unsigned int urc_ut;
+	unsigned int urc_dwarf;
 
 	unsigned long curr_sp;
 	unsigned long curr_fp;
diff --git a/drivers/misc/tegra-profiler/comm.c b/drivers/misc/tegra-profiler/comm.c
index 7c4b6fcb1a7a..2d70f6318d3e 100644
--- a/drivers/misc/tegra-profiler/comm.c
+++ b/drivers/misc/tegra-profiler/comm.c
@@ -370,6 +370,7 @@ init_mmap_hdr(struct quadd_mmap_rb_info *mmap_rb,
 	mmap_hdr->magic = QUADD_MMAP_HEADER_MAGIC;
 	mmap_hdr->version = QUADD_MMAP_HEADER_VERSION;
 	mmap_hdr->cpu_id = cpu_id;
+	mmap_hdr->samples_version = QUADD_SAMPLES_VERSION;
 
 	rb_hdr = (struct quadd_ring_buffer_hdr *)(mmap_hdr + 1);
 	rb->rb_hdr = rb_hdr;
diff --git a/drivers/misc/tegra-profiler/dwarf_unwind.c b/drivers/misc/tegra-profiler/dwarf_unwind.c
index ef5c4ad420eb..2e624352c4b7 100644
--- a/drivers/misc/tegra-profiler/dwarf_unwind.c
+++ b/drivers/misc/tegra-profiler/dwarf_unwind.c
@@ -1905,7 +1905,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 	unsigned int unw_type;
 	int is_eh = 1, mode = sf->mode;
 
-	cc->unw_rc = QUADD_URC_FAILURE;
+	cc->urc_dwarf = QUADD_URC_FAILURE;
 	user_reg_size = get_user_reg_size(mode);
 
 	while (1) {
@@ -1922,7 +1922,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 		sp = sf->vregs[regnum_sp(mode)];
 
 		if (!validate_stack_addr(sp, vma_sp, user_reg_size)) {
-			cc->unw_rc = -QUADD_URC_SP_INCORRECT;
+			cc->urc_dwarf = QUADD_URC_SP_INCORRECT;
 			break;
 		}
 
@@ -1935,7 +1935,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 		if (!is_vma_addr(addr, vma_pc, user_reg_size)) {
 			err = quadd_get_dw_frames(vma_pc->vm_start, &ri_new);
 			if (err) {
-				cc->unw_rc = QUADD_URC_TBL_NOT_EXIST;
+				cc->urc_dwarf = QUADD_URC_TBL_NOT_EXIST;
 				break;
 			}
 
@@ -1947,7 +1947,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 
 		if (!is_fde_entry_exist(ri, sf->pc, &__is_eh, &__is_debug)) {
 			pr_debug("eh/debug fde entries are not existed\n");
-			cc->unw_rc = QUADD_URC_IDX_NOT_FOUND;
+			cc->urc_dwarf = QUADD_URC_IDX_NOT_FOUND;
 			break;
 		}
 		pr_debug("is_eh: %d, is_debug: %d\n", __is_eh, __is_debug);
@@ -1967,11 +1967,11 @@ unwind_backtrace(struct quadd_callchain *cc,
 
 				err = unwind_frame(ri, sf, vma_sp, is_eh);
 				if (err < 0) {
-					cc->unw_rc = -err;
+					cc->urc_dwarf = -err;
 					break;
 				}
 			} else {
-				cc->unw_rc = -err;
+				cc->urc_dwarf = -err;
 				break;
 			}
 		}
@@ -2038,10 +2038,10 @@ quadd_get_user_cc_dwarf(struct pt_regs *regs,
 	if (!regs || !mm)
 		return 0;
 
-	if (cc->unw_rc == QUADD_URC_LEVEL_TOO_DEEP)
+	if (cc->urc_dwarf == QUADD_URC_LEVEL_TOO_DEEP)
 		return nr_prev;
 
-	cc->unw_rc = QUADD_URC_FAILURE;
+	cc->urc_dwarf = QUADD_URC_FAILURE;
 
 	if (nr_prev > 0) {
 		ip = cc->curr_pc;
@@ -2109,7 +2109,7 @@ quadd_get_user_cc_dwarf(struct pt_regs *regs,
 
 	err = quadd_get_dw_frames(vma->vm_start, &ri);
 	if (err) {
-		cc->unw_rc = QUADD_URC_TBL_NOT_EXIST;
+		cc->urc_dwarf = QUADD_URC_TBL_NOT_EXIST;
 		return 0;
 	}
 
diff --git a/drivers/misc/tegra-profiler/eh_unwind.c b/drivers/misc/tegra-profiler/eh_unwind.c
index bfbcd144822f..b48765678aee 100644
--- a/drivers/misc/tegra-profiler/eh_unwind.c
+++ b/drivers/misc/tegra-profiler/eh_unwind.c
@@ -1060,7 +1060,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 {
 	struct ex_region_info ri_new;
 
-	cc->unw_rc = QUADD_URC_FAILURE;
+	cc->urc_ut = QUADD_URC_FAILURE;
 
 	pr_debug("fp_arm: %#lx, fp_thumb: %#lx, sp: %#lx, lr: %#lx, pc: %#lx\n",
 		 frame->fp_arm, frame->fp_thumb,
@@ -1081,7 +1081,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 			break;
 
 		if (!validate_stack_addr(frame->sp, vma_sp, sizeof(u32))) {
-			cc->unw_rc = -QUADD_URC_SP_INCORRECT;
+			cc->urc_ut = QUADD_URC_SP_INCORRECT;
 			break;
 		}
 
@@ -1094,7 +1094,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 		if (!is_vma_addr(ti->addr, vma_pc, sizeof(u32))) {
 			err = get_extabs_ehabi(vma_pc->vm_start, &ri_new);
 			if (err) {
-				cc->unw_rc = QUADD_URC_TBL_NOT_EXIST;
+				cc->urc_ut = QUADD_URC_TBL_NOT_EXIST;
 				break;
 			}
 
@@ -1104,7 +1104,7 @@ unwind_backtrace(struct quadd_callchain *cc,
 		err = unwind_frame(ri, frame, vma_sp);
 		if (err < 0) {
 			pr_debug("end unwind, urc: %ld\n", err);
-			cc->unw_rc = -err;
+			cc->urc_ut = -err;
 			break;
 		}
 
@@ -1144,10 +1144,10 @@ quadd_get_user_cc_arm32_ehabi(struct pt_regs *regs,
 		return 0;
 #endif
 
-	if (cc->unw_rc == QUADD_URC_LEVEL_TOO_DEEP)
+	if (cc->urc_ut == QUADD_URC_LEVEL_TOO_DEEP)
 		return nr_prev;
 
-	cc->unw_rc = QUADD_URC_FAILURE;
+	cc->urc_ut = QUADD_URC_FAILURE;
 
 	if (nr_prev > 0) {
 		ip = cc->curr_pc;
@@ -1188,7 +1188,7 @@ quadd_get_user_cc_arm32_ehabi(struct pt_regs *regs,
 
 	err = get_extabs_ehabi(vma->vm_start, &ri);
 	if (err) {
-		cc->unw_rc = QUADD_URC_TBL_NOT_EXIST;
+		cc->urc_ut = QUADD_URC_TBL_NOT_EXIST;
 		return 0;
 	}
 
diff --git a/drivers/misc/tegra-profiler/hrt.c b/drivers/misc/tegra-profiler/hrt.c
index 88a0d967e096..214b79c56aeb 100644
--- a/drivers/misc/tegra-profiler/hrt.c
+++ b/drivers/misc/tegra-profiler/hrt.c
@@ -62,7 +62,7 @@ static enum hrtimer_restart hrtimer_handler(struct hrtimer *hrtimer)
 
 	regs = get_irq_regs();
 
-	if (!hrt.active)
+	if (!atomic_read(&hrt.active))
 		return HRTIMER_NORESTART;
 
 	qm_debug_handler_sample(regs);
@@ -193,7 +193,14 @@ static void put_header(void)
 	hdr->reserved = 0;
 	hdr->extra_length = 0;
 
-	hdr->reserved |= hrt.unw_method << QUADD_HDR_UNW_METHOD_SHIFT;
+	if (hdr->backtrace) {
+		struct quadd_unw_methods *um = &hrt.um;
+
+		hdr->reserved |= um->fp ? QUADD_HDR_BT_FP : 0;
+		hdr->reserved |= um->ut ? QUADD_HDR_BT_UT : 0;
+		hdr->reserved |= um->ut_ce ? QUADD_HDR_BT_UT_CE : 0;
+		hdr->reserved |= um->dwarf ? QUADD_HDR_BT_DWARF : 0;
+	}
 
 	if (hrt.use_arch_timer)
 		hdr->reserved |= QUADD_HDR_USE_ARCH_TIMER;
@@ -335,11 +342,11 @@ get_stack_offset(struct task_struct *task,
 static void
 read_all_sources(struct pt_regs *regs, struct task_struct *task)
 {
-	u32 state, extra_data = 0;
+	u32 state, extra_data = 0, urcs = 0;
 	int i, vec_idx = 0, bt_size = 0;
 	int nr_events = 0, nr_positive_events = 0;
 	struct pt_regs *user_regs;
-	struct quadd_iovec vec[5];
+	struct quadd_iovec vec[6];
 	struct hrt_event_value events[QUADD_MAX_COUNTERS];
 	u32 events_extra[QUADD_MAX_COUNTERS];
 
@@ -401,7 +408,8 @@ read_all_sources(struct pt_regs *regs, struct task_struct *task)
 	cc->curr_pc = 0;
 
 	if (ctx->param.backtrace) {
-		cc->unw_method = hrt.unw_method;
+		cc->um = hrt.um;
+
 		bt_size = quadd_get_user_callchain(user_regs, cc, ctx, task);
 
 		if (!bt_size && !user_mode(regs)) {
@@ -434,8 +442,18 @@ read_all_sources(struct pt_regs *regs, struct task_struct *task)
 				extra_data |= QUADD_SED_IP64;
 		}
 
-		extra_data |= cc->unw_method << QUADD_SED_UNW_METHOD_SHIFT;
-		s->reserved |= cc->unw_rc << QUADD_SAMPLE_URC_SHIFT;
+		urcs |= (cc->urc_fp & QUADD_SAMPLE_URC_MASK) <<
+			QUADD_SAMPLE_URC_SHIFT_FP;
+		urcs |= (cc->urc_ut & QUADD_SAMPLE_URC_MASK) <<
+			QUADD_SAMPLE_URC_SHIFT_UT;
+		urcs |= (cc->urc_dwarf & QUADD_SAMPLE_URC_MASK) <<
+			QUADD_SAMPLE_URC_SHIFT_DWARF;
+
+		s->reserved |= QUADD_SAMPLE_RES_URCS_ENABLED;
+
+		vec[vec_idx].base = &urcs;
+		vec[vec_idx].len = sizeof(urcs);
+		vec_idx++;
 	}
 	s->callchain_nr = bt_size;
 
@@ -540,7 +558,7 @@ void __quadd_task_sched_in(struct task_struct *prev,
 	struct event_data events[QUADD_MAX_COUNTERS];
 	/* static DEFINE_RATELIMIT_STATE(ratelimit_state, 5 * HZ, 2); */
 
-	if (likely(!hrt.active))
+	if (likely(!atomic_read(&hrt.active)))
 		return;
 /*
 	if (__ratelimit(&ratelimit_state))
@@ -578,7 +596,7 @@ void __quadd_task_sched_out(struct task_struct *prev,
 	struct quadd_ctx *ctx = hrt.quadd_ctx;
 	/* static DEFINE_RATELIMIT_STATE(ratelimit_state, 5 * HZ, 2); */
 
-	if (likely(!hrt.active))
+	if (likely(!atomic_read(&hrt.active)))
 		return;
 /*
 	if (__ratelimit(&ratelimit_state))
@@ -612,7 +630,7 @@ void __quadd_event_mmap(struct vm_area_struct *vma)
 {
 	struct quadd_parameters *param;
 
-	if (likely(!hrt.active))
+	if (likely(!atomic_read(&hrt.active)))
 		return;
 
 	if (!is_profile_process(current))
@@ -665,14 +683,17 @@ int quadd_hrt_start(void)
 
 	extra = param->reserved[QUADD_PARAM_IDX_EXTRA];
 
-	if (extra & QUADD_PARAM_EXTRA_BT_MIXED)
-		hrt.unw_method = QUADD_UNW_METHOD_MIXED;
-	else if (extra & QUADD_PARAM_EXTRA_BT_UNWIND_TABLES)
-		hrt.unw_method = QUADD_UNW_METHOD_EHT;
-	else if (extra & QUADD_PARAM_EXTRA_BT_FP)
-		hrt.unw_method = QUADD_UNW_METHOD_FP;
-	else
-		hrt.unw_method = QUADD_UNW_METHOD_NONE;
+	if (param->backtrace) {
+		struct quadd_unw_methods *um = &hrt.um;
+
+		um->fp = extra & QUADD_PARAM_EXTRA_BT_FP ? 1 : 0;
+		um->ut = extra & QUADD_PARAM_EXTRA_BT_UT ? 1 : 0;
+		um->ut_ce = extra & QUADD_PARAM_EXTRA_BT_UT_CE ? 1 : 0;
+		um->dwarf = extra & QUADD_PARAM_EXTRA_BT_DWARF ? 1 : 0;
+
+		pr_info("unw methods: fp/ut/ut_ce/dwarf: %u/%u/%u/%u\n",
+			um->fp, um->ut, um->ut_ce, um->dwarf);
+	}
 
 	if (hrt.tc && (extra & QUADD_PARAM_EXTRA_USE_ARCH_TIMER))
 		hrt.use_arch_timer = 1;
@@ -699,7 +720,7 @@ int quadd_hrt_start(void)
 
 	quadd_ma_start(&hrt);
 
-	hrt.active = 1;
+	atomic_set(&hrt.active, 1);
 
 	pr_info("Start hrt: freq/period: %ld/%llu\n", freq, period);
 	return 0;
@@ -718,7 +739,7 @@ void quadd_hrt_stop(void)
 
 	quadd_ma_stop(&hrt);
 
-	hrt.active = 0;
+	atomic_set(&hrt.active, 0);
 
 	atomic64_set(&hrt.counter_samples, 0);
 	atomic64_set(&hrt.skipped_samples, 0);
@@ -728,7 +749,7 @@ void quadd_hrt_stop(void)
 
 void quadd_hrt_deinit(void)
 {
-	if (hrt.active)
+	if (atomic_read(&hrt.active))
 		quadd_hrt_stop();
 
 	free_percpu(hrt.cpu_ctx);
@@ -758,7 +779,7 @@ struct quadd_hrt_ctx *quadd_hrt_init(struct quadd_ctx *ctx)
 	struct quadd_cpu_context *cpu_ctx;
 
 	hrt.quadd_ctx = ctx;
-	hrt.active = 0;
+	atomic_set(&hrt.active, 0);
 
 	freq = ctx->param.freq;
 	freq = max_t(long, QUADD_HRT_MIN_FREQ, freq);
diff --git a/drivers/misc/tegra-profiler/hrt.h b/drivers/misc/tegra-profiler/hrt.h
index 172e2f1719dd..2b5254a52fa6 100644
--- a/drivers/misc/tegra-profiler/hrt.h
+++ b/drivers/misc/tegra-profiler/hrt.h
@@ -49,7 +49,7 @@ struct quadd_hrt_ctx {
 
 	struct quadd_ctx *quadd_ctx;
 
-	int active;
+	atomic_t active;
 	atomic_t nr_active_all_core;
 
 	atomic64_t counter_samples;
@@ -64,7 +64,7 @@ struct quadd_hrt_ctx {
 	struct timecounter *tc;
 	int use_arch_timer;
 
-	unsigned int unw_method;
+	struct quadd_unw_methods um;
 	int get_stack_offset;
 };
 
diff --git a/drivers/misc/tegra-profiler/ma.c b/drivers/misc/tegra-profiler/ma.c
index c7ccd82ac1b0..83ae8195f80e 100644
--- a/drivers/misc/tegra-profiler/ma.c
+++ b/drivers/misc/tegra-profiler/ma.c
@@ -85,7 +85,7 @@ static void timer_interrupt(unsigned long data)
 	struct quadd_hrt_ctx *hrt_ctx = (struct quadd_hrt_ctx *)data;
 	struct timer_list *timer = &hrt_ctx->ma_timer;
 
-	if (hrt_ctx->active == 0)
+	if (!atomic_read(&hrt_ctx->active))
 		return;
 
 	check_ma(hrt_ctx);
diff --git a/drivers/misc/tegra-profiler/main.c b/drivers/misc/tegra-profiler/main.c
index f9d7d4bcdd9b..d3236970e27a 100644
--- a/drivers/misc/tegra-profiler/main.c
+++ b/drivers/misc/tegra-profiler/main.c
@@ -185,7 +185,6 @@ set_parameters(struct quadd_parameters *p, uid_t *debug_app_uid)
 	int pl310_events_id;
 	int nr_pmu = 0, nr_pl310 = 0;
 	struct task_struct *task;
-	unsigned int extra;
 	u64 *low_addr_p;
 
 	if (!validate_freq(p->freq)) {
@@ -293,17 +292,6 @@ set_parameters(struct quadd_parameters *p, uid_t *debug_app_uid)
 		}
 	}
 
-	extra = p->reserved[QUADD_PARAM_IDX_EXTRA];
-
-	if (extra & QUADD_PARAM_EXTRA_BT_UNWIND_TABLES)
-		pr_info("unwinding: exception-handling tables\n");
-
-	if (extra & QUADD_PARAM_EXTRA_BT_FP)
-		pr_info("unwinding: frame pointers\n");
-
-	if (extra & QUADD_PARAM_EXTRA_BT_MIXED)
-		pr_info("unwinding: mixed mode\n");
-
 	low_addr_p = (u64 *)&p->reserved[QUADD_PARAM_IDX_BT_LOWER_BOUND];
 	ctx.hrt->low_addr = (unsigned long)*low_addr_p;
 	pr_info("bt lower bound: %#lx\n", ctx.hrt->low_addr);
diff --git a/drivers/misc/tegra-profiler/version.h b/drivers/misc/tegra-profiler/version.h
index f956c1d40861..d3254f99c063 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		"1.96"
+#define QUADD_MODULE_VERSION		"1.97"
 #define QUADD_MODULE_BRANCH		"Dev"
 
 #endif	/* __QUADD_VERSION_H */
diff --git a/include/linux/tegra_profiler.h b/include/linux/tegra_profiler.h
index 66d5bf240c74..fffa74a7e7df 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	32
-#define QUADD_IO_VERSION	17
+#define QUADD_SAMPLES_VERSION	33
+#define QUADD_IO_VERSION	18
 
 #define QUADD_IO_VERSION_DYNAMIC_RB		5
 #define QUADD_IO_VERSION_RB_MAX_FILL_COUNT	6
@@ -35,6 +35,7 @@
 #define QUADD_IO_VERSION_BT_LOWER_BOUND		15
 #define QUADD_IO_VERSION_STACK_OFFSET		16
 #define QUADD_IO_VERSION_SECTIONS_INFO		17
+#define QUADD_IO_VERSION_UNW_METHODS_OPT	18
 
 #define QUADD_SAMPLE_VERSION_THUMB_MODE_FLAG	17
 #define QUADD_SAMPLE_VERSION_GROUP_SAMPLES	18
@@ -50,6 +51,7 @@
 #define QUADD_SAMPLE_VERSION_HDR_ARCH_TIMER	30
 #define QUADD_SAMPLE_VERSION_STACK_OFFSET	31
 #define QUADD_SAMPLE_VERSION_SCHED_TASK_STATE	32
+#define QUADD_SAMPLE_VERSION_URCS		33
 
 #define QUADD_MMAP_HEADER_VERSION		1
 
@@ -165,18 +167,13 @@ enum quadd_cpu_mode {
 
 #pragma pack(push, 1)
 
-#define QUADD_SAMPLE_UNW_METHOD_SHIFT	0
-#define QUADD_SAMPLE_UNW_METHOD_MASK	(1 << QUADD_SAMPLE_UNW_METHOD_SHIFT)
+#define QUADD_SAMPLE_RES_URCS_ENABLED	(1 << 0)
 
-enum {
-	QUADD_UNW_METHOD_FP = 0,
-	QUADD_UNW_METHOD_EHT,
-	QUADD_UNW_METHOD_MIXED,
-	QUADD_UNW_METHOD_NONE,
-};
+#define QUADD_SAMPLE_URC_MASK		0xff
 
-#define QUADD_SAMPLE_URC_SHIFT		1
-#define QUADD_SAMPLE_URC_MASK		(0x0f << QUADD_SAMPLE_URC_SHIFT)
+#define QUADD_SAMPLE_URC_SHIFT_FP	0
+#define QUADD_SAMPLE_URC_SHIFT_UT	(1 * 8)
+#define QUADD_SAMPLE_URC_SHIFT_DWARF	(2 * 8)
 
 enum {
 	QUADD_URC_SUCCESS = 0,
@@ -194,15 +191,13 @@ enum {
 	QUADD_URC_PC_INCORRECT,
 	QUADD_URC_LEVEL_TOO_DEEP,
 	QUADD_URC_FP_INCORRECT,
+	QUADD_URC_NONE,
 	QUADD_URC_MAX,
 };
 
 #define QUADD_SED_IP64			(1 << 0)
 
-#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_SHIFT	1
 #define QUADD_SED_STACK_OFFSET_MASK	(0xffff << QUADD_SED_STACK_OFFSET_SHIFT)
 
 enum {
@@ -318,11 +313,12 @@ struct quadd_debug_data {
 
 #define QUADD_HEADER_MAGIC	0x1122
 
-#define QUADD_HDR_UNW_METHOD_SHIFT	0
-#define QUADD_HDR_UNW_METHOD_MASK	(0x07 << QUADD_HDR_UNW_METHOD_SHIFT)
-
+#define QUADD_HDR_BT_FP			(1 << 0)
+#define QUADD_HDR_BT_UT			(1 << 1)
+#define QUADD_HDR_BT_UT_CE		(1 << 2)
 #define QUADD_HDR_USE_ARCH_TIMER	(1 << 3)
 #define QUADD_HDR_STACK_OFFSET		(1 << 4)
+#define QUADD_HDR_BT_DWARF		(1 << 5)
 
 struct quadd_header_data {
 	u16 magic;
@@ -372,10 +368,12 @@ enum {
 
 #define QUADD_PARAM_EXTRA_GET_MMAP		(1 << 0)
 #define QUADD_PARAM_EXTRA_BT_FP			(1 << 1)
-#define QUADD_PARAM_EXTRA_BT_UNWIND_TABLES	(1 << 2)
+#define QUADD_PARAM_EXTRA_BT_UT			(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)
+#define QUADD_PARAM_EXTRA_BT_UT_CE		(1 << 6)
+#define QUADD_PARAM_EXTRA_BT_DWARF		(1 << 7)
 
 struct quadd_parameters {
 	u32 freq;