Capture backtraces when measuring stack depth

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>

6 files changed, 38 insertions, 17 deletions

diff --git a/asm/asm-offsets.c b/asm/asm-offsets.c
index 4fb2344..74f3124 100644
--- a/asm/asm-offsets.c
+++ b/asm/asm-offsets.c
@@ -37,10 +37,11 @@ int main(void)
 	OFFSET(CPUTHREAD_STATE, cpu_thread, state);
 	OFFSET(CPUTHREAD_CUR_TOKEN, cpu_thread, current_token);
 	DEFINE(CPUTHREAD_GAP, sizeof(struct cpu_thread) + STACK_SAFETY_GAP);
+#ifdef STACK_CHECK_ENABLED
 	OFFSET(CPUTHREAD_STACK_BOT_MARK, cpu_thread, stack_bot_mark);
 	OFFSET(CPUTHREAD_STACK_BOT_PC, cpu_thread, stack_bot_pc);
 	OFFSET(CPUTHREAD_STACK_BOT_TOK, cpu_thread, stack_bot_tok);
-
+#endif
 	OFFSET(STACK_TYPE,	stack_frame, type);
 	OFFSET(STACK_LOCALS,	stack_frame, locals);
 	OFFSET(STACK_GPR0,	stack_frame, gpr[0]);
diff --git a/asm/head.S b/asm/head.S
index 75eb170..c4d5240 100644
--- a/asm/head.S
+++ b/asm/head.S
@@ -311,10 +311,11 @@ boot_entry:
 	/* Get our per-cpu pointer into r13 */
 	GET_CPU()
 
+#ifdef STACK_CHECK_ENABLED
 	/* Initialize stack bottom mark to 0, it will be updated in C code */
 	li	%r0,0
 	std	%r0,CPUTHREAD_STACK_BOT_MARK(%r13)
-
+#endif
 	/* Jump to C */
 	mr	%r3,%r27
 	mr	%r4,%r25
diff --git a/core/cpu.c b/core/cpu.c
index 5e60da2..1e405f8 100644
--- a/core/cpu.c
+++ b/core/cpu.c
@@ -339,7 +339,9 @@ static void init_cpu_thread(struct cpu_thread *t,
 	list_head_init(&t->job_queue);
 	t->state = state;
 	t->pir = pir;
+#ifdef STACK_CHECK_ENABLED
 	t->stack_bot_mark = LONG_MAX;
+#endif
 	assert(pir == container_of(t, struct cpu_stack, cpu) - cpu_stacks);
 }
 
diff --git a/core/stack.c b/core/stack.c
index 0d8a7eb..9858968 100644
--- a/core/stack.c
+++ b/core/stack.c
@@ -110,6 +110,9 @@ void __noreturn __nomcount __stack_chk_fail(void)
 
 #ifdef STACK_CHECK_ENABLED
 
+static int64_t lowest_stack_mark = LONG_MAX;
+static struct lock stack_check_lock = LOCK_UNLOCKED;
+
 void __nomcount __mcount_stack_check(uint64_t sp, uint64_t lr);
 void __nomcount __mcount_stack_check(uint64_t sp, uint64_t lr)
 {
@@ -129,9 +132,14 @@ void __nomcount __mcount_stack_check(uint64_t sp, uint64_t lr)
 
 	/* Capture lowest stack for this thread */
 	if (mark < c->stack_bot_mark) {
+		unsigned int count = CPU_BACKTRACE_SIZE;
+		lock(&stack_check_lock);
 		c->stack_bot_mark = mark;
 		c->stack_bot_pc = lr;
 		c->stack_bot_tok = c->current_token;
+		__backtrace(c->stack_bot_bt, &count);
+		c->stack_bot_bt_count = count;
+		unlock(&stack_check_lock);
 	}
 
 	/* Stack is within bounds ? check for warning and bail */
@@ -152,14 +160,17 @@ void __nomcount __mcount_stack_check(uint64_t sp, uint64_t lr)
 	abort();
 }
 
-static int64_t lowest_stack_mark = LONG_MAX;
-static struct lock stack_check_lock = LOCK_UNLOCKED;
-
 void check_stacks(void)
 {
 	struct cpu_thread *c;
-	uint64_t lmark, lpc, ltok;
-	int found = -1;
+
+	/* We should never call that from mcount */
+	assert(!this_cpu()->in_mcount);
+
+	/* Mark ourselves "in_mcount" to avoid deadlock on stack
+	 * check lock
+	 */
+	this_cpu()->in_mcount = true;
 
 	for_each_cpu(c) {
 		if (!c->stack_bot_mark ||
@@ -170,15 +181,15 @@ void check_stacks(void)
 			unlock(&stack_check_lock);
 			continue;
 		}
-		lmark = lowest_stack_mark = c->stack_bot_mark;
-		lpc = c->stack_bot_pc;
-		ltok = c->stack_bot_tok;
-		found = c->pir;
+		prlog(PR_NOTICE, "CPU %04x lowest stack mark %lld bytes left"
+		      " pc=%08llx token=%lld\n",
+		      c->pir, c->stack_bot_mark, c->stack_bot_pc,
+		      c->stack_bot_tok);
+		__print_backtrace(c->pir, c->stack_bot_bt,
+				  c->stack_bot_bt_count, NULL, NULL);
 		unlock(&stack_check_lock);
 	}
-	if (found >= 0)
-		prlog(PR_NOTICE, "CPU %04x lowest stack mark %lld bytes left"
-		      " pc=%08llx token=%lld\n", found, lmark, lpc, ltok);
-}
 
+	this_cpu()->in_mcount = false;
+}
 #endif /* STACK_CHECK_ENABLED */
diff --git a/include/cpu.h b/include/cpu.h
index 6c8274c..d7acd25 100644
--- a/include/cpu.h
+++ b/include/cpu.h
@@ -22,6 +22,7 @@
 #include <lock.h>
 #include <device.h>
 #include <opal.h>
+#include <stack.h>
 
 /*
  * cpu_thread is our internal structure representing each
@@ -60,9 +61,14 @@ struct cpu_thread {
 	uint32_t			hbrt_spec_wakeup; /* primary only */
 	uint64_t			save_l2_fir_action1;
 	uint64_t			current_token;
+#ifdef STACK_CHECK_ENABLED
 	int64_t				stack_bot_mark;
 	uint64_t			stack_bot_pc;
 	uint64_t			stack_bot_tok;
+#define CPU_BACKTRACE_SIZE	20
+	struct bt_entry			stack_bot_bt[CPU_BACKTRACE_SIZE];
+	unsigned int			stack_bot_bt_count;
+#endif
 	struct lock			job_lock;
 	struct list_head		job_queue;
 };
diff --git a/include/stack.h b/include/stack.h
index d4664dd..d6a0609 100644
--- a/include/stack.h
+++ b/include/stack.h
@@ -47,9 +47,9 @@
 #define STACK_SAFETY_GAP	512
 
 /* Warning threshold, if stack goes below that on mcount, print a
- * warning
+ * warning.
  */
-#define STACK_WARNING_GAP	1024
+#define STACK_WARNING_GAP	2048
 
 #ifndef __ASSEMBLY__