X-Git-Url: http://git.lttng.org/?a=blobdiff_plain;f=lttng-context-callstack.c;h=432fadb652f349566ae1eee7e646ed29f83b507b;hb=e0407e483deb67b6f8617d7100278c97313a9914;hp=3b7859eb8542b8c6bc073288f2380cf3b29e37fb;hpb=64cc198b453a1fd05fb6f23bda08282ed0902c1c;p=lttng-modules.git

diff --git a/lttng-context-callstack.c b/lttng-context-callstack.c
index 3b7859eb..432fadb6 100644
--- a/lttng-context-callstack.c
+++ b/lttng-context-callstack.c
@@ -20,31 +20,32 @@
  * License along with this library; if not, write to the Free Software
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  *
- * The callstack context can be added to any kernel
- * event. It records either the kernel or the userspace callstack, up to a
- * max depth. The context is a CTF sequence, such that it uses only the space
- * required for the number of callstack entries.
+ * The callstack context can be added to any kernel event. It records
+ * either the kernel or the userspace callstack, up to a max depth. The
+ * context is a CTF sequence, such that it uses only the space required
+ * for the number of callstack entries.
  *
- * It allocates callstack buffers per-CPU up to 4 interrupt nesting. This
- * nesting limit is the same as defined in the ring buffer. It therefore uses a
- * fixed amount of memory, proportional to the number of CPUs:
+ * It allocates callstack buffers per-CPU up to 4 interrupt nesting.
+ * This nesting limit is the same as defined in the ring buffer. It
+ * therefore uses a fixed amount of memory, proportional to the number
+ * of CPUs:
  *
  *   size = cpus * nest * depth * sizeof(unsigned long)
  *
- * Which is about 800 bytes per-CPUs on 64-bit host and a depth of 25. The
- * allocation is done at the initialization to avoid memory allocation
- * overhead while tracing, using a shallow stack.
+ * Which is 4096 bytes per CPU on 64-bit host and a depth of 128.
+ * The allocation is done at the initialization to avoid memory
+ * allocation overhead while tracing, using a shallow stack.
  *
  * The kernel callstack is recovered using save_stack_trace(), and the
  * userspace callstack uses save_stack_trace_user(). They rely on frame
- * pointers. These are usually available for the kernel, but the compiler
- * option -fomit-frame-pointer frequently used in popular Linux distributions
- * may cause the userspace callstack to be unreliable, and is a known
- * limitation of this approach. If frame pointers are not available, it
- * produces no error, but the callstack will be empty. We still provide the
- * feature, because it works well for runtime environments having frame
- * pointers. In the future, unwind support and/or last branch record may
- * provide a solution to this problem.
+ * pointers. These are usually available for the kernel, but the
+ * compiler option -fomit-frame-pointer frequently used in popular Linux
+ * distributions may cause the userspace callstack to be unreliable, and
+ * is a known limitation of this approach. If frame pointers are not
+ * available, it produces no error, but the callstack will be empty. We
+ * still provide the feature, because it works well for runtime
+ * environments having frame pointers. In the future, unwind support
+ * and/or last branch record may provide a solution to this problem.
  *
  * The symbol name resolution is left to the trace reader.
  */
@@ -61,20 +62,26 @@
 #include "wrapper/vmalloc.h"
 #include "lttng-tracer.h"
 
-#define MAX_ENTRIES 25
+#define MAX_ENTRIES 128
 
-struct lttng_cs_nesting {
+enum lttng_cs_ctx_modes {
+	CALLSTACK_KERNEL = 0,
+	CALLSTACK_USER = 1,
+	NR_CALLSTACK_MODES,
+};
+
+struct lttng_cs_dispatch {
 	struct stack_trace stack_trace;
 	unsigned long entries[MAX_ENTRIES];
 };
 
 struct lttng_cs {
-	struct lttng_cs_nesting level[RING_BUFFER_MAX_NESTING];
+	struct lttng_cs_dispatch dispatch[RING_BUFFER_MAX_NESTING];
 };
 
 struct field_data {
 	struct lttng_cs __percpu *cs_percpu;
-	int mode;
+	enum lttng_cs_ctx_modes mode;
 };
 
 struct lttng_cs_type {
@@ -83,11 +90,6 @@ struct lttng_cs_type {
 	void (*save_func)(struct stack_trace *trace);
 };
 
-enum lttng_cs_ctx_modes {
-	CALLSTACK_KERNEL = 0,
-	CALLSTACK_USER = 1,
-};
-
 static struct lttng_cs_type cs_types[] = {
 	{
 		.name		= "callstack_kernel",
@@ -102,7 +104,7 @@ static struct lttng_cs_type cs_types[] = {
 };
 
 static
-int init_type(int mode)
+int init_type(enum lttng_cs_ctx_modes mode)
 {
 	unsigned long func;
 
@@ -118,14 +120,26 @@ int init_type(int mode)
 	return 0;
 }
 
+/* Keep track of nesting inside userspace callstack context code */
+DEFINE_PER_CPU(int, callstack_user_nesting);
+
 static
 struct stack_trace *stack_trace_context(struct lttng_ctx_field *field,
 					struct lib_ring_buffer_ctx *ctx)
 {
-	int nesting;
+	int buffer_nesting, cs_user_nesting;
 	struct lttng_cs *cs;
 	struct field_data *fdata = field->priv;
 
+	/*
+	 * Do not gather the userspace callstack context when the event was
+	 * triggered by the userspace callstack context saving mechanism.
+	 */
+	cs_user_nesting = per_cpu(callstack_user_nesting, ctx->cpu);
+
+	if (fdata->mode == CALLSTACK_USER && cs_user_nesting >= 1)
+		return NULL;
+
 	/*
 	 * get_cpu() is not required, preemption is already
 	 * disabled while event is written.
@@ -134,11 +148,11 @@ struct stack_trace *stack_trace_context(struct lttng_ctx_field *field,
 	 * Check it again as a safety net.
 	 */
 	cs = per_cpu_ptr(fdata->cs_percpu, ctx->cpu);
-	nesting = per_cpu(lib_ring_buffer_nesting, ctx->cpu) - 1;
-	if (nesting >= RING_BUFFER_MAX_NESTING) {
+	buffer_nesting = per_cpu(lib_ring_buffer_nesting, ctx->cpu) - 1;
+	if (buffer_nesting >= RING_BUFFER_MAX_NESTING)
 		return NULL;
-	}
-	return &cs->level[nesting].stack_trace;
+
+	return &cs->dispatch[buffer_nesting].stack_trace;
 }
 
 /*
@@ -150,20 +164,31 @@ size_t lttng_callstack_get_size(size_t offset, struct lttng_ctx_field *field,
 				struct lib_ring_buffer_ctx *ctx,
 				struct lttng_channel *chan)
 {
-	size_t size = 0;
 	struct stack_trace *trace;
 	struct field_data *fdata = field->priv;
+	size_t orig_offset = offset;
 
 	/* do not write data if no space is available */
 	trace = stack_trace_context(field, ctx);
-	if (!trace)
-		return 0;
+	if (unlikely(!trace)) {
+		offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
+		offset += sizeof(unsigned int);
+		offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
+		return offset - orig_offset;
+	}
 
 	/* reset stack trace, no need to clear memory */
 	trace->nr_entries = 0;
 
+	if (fdata->mode == CALLSTACK_USER)
+		++per_cpu(callstack_user_nesting, ctx->cpu);
+
 	/* do the real work and reserve space */
 	cs_types[fdata->mode].save_func(trace);
+
+	if (fdata->mode == CALLSTACK_USER)
+		per_cpu(callstack_user_nesting, ctx->cpu)--;
+
 	/*
 	 * Remove final ULONG_MAX delimiter. If we cannot find it, add
 	 * our own marker to show that the stack is incomplete. This is
@@ -173,14 +198,14 @@ size_t lttng_callstack_get_size(size_t offset, struct lttng_ctx_field *field,
 			&& trace->entries[trace->nr_entries - 1] == ULONG_MAX) {
 		trace->nr_entries--;
 	}
-	size += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
-	size += sizeof(unsigned int);
-	size += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
-	size += sizeof(unsigned long) * trace->nr_entries;
+	offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
+	offset += sizeof(unsigned int);
+	offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
+	offset += sizeof(unsigned long) * trace->nr_entries;
 	/* Add our own ULONG_MAX delimiter to show incomplete stack. */
 	if (trace->nr_entries == trace->max_entries)
-		size += sizeof(unsigned long);
-	return size;
+		offset += sizeof(unsigned long);
+	return offset - orig_offset;
 }
 
 static
@@ -191,8 +216,13 @@ void lttng_callstack_record(struct lttng_ctx_field *field,
 	struct stack_trace *trace = stack_trace_context(field, ctx);
 	unsigned int nr_seq_entries;
 
-	if (!trace)
+	if (unlikely(!trace)) {
+		nr_seq_entries = 0;
+		lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
+		chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
+		lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
 		return;
+	}
 	lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
 	nr_seq_entries = trace->nr_entries;
 	if (trace->nr_entries == trace->max_entries)
@@ -219,7 +249,7 @@ void field_data_free(struct field_data *fdata)
 }
 
 static
-struct field_data __percpu *field_data_create(int type)
+struct field_data __percpu *field_data_create(enum lttng_cs_ctx_modes mode)
 {
 	int cpu, i;
 	struct lttng_cs __percpu *cs_set;
@@ -238,14 +268,14 @@ struct field_data __percpu *field_data_create(int type)
 
 		cs = per_cpu_ptr(cs_set, cpu);
 		for (i = 0; i < RING_BUFFER_MAX_NESTING; i++) {
-			struct lttng_cs_nesting *level;
+			struct lttng_cs_dispatch *dispatch;
 
-			level = &cs->level[i];
-			level->stack_trace.entries = level->entries;
-			level->stack_trace.max_entries = MAX_ENTRIES;
+			dispatch = &cs->dispatch[i];
+			dispatch->stack_trace.entries = dispatch->entries;
+			dispatch->stack_trace.max_entries = MAX_ENTRIES;
 		}
 	}
-	fdata->mode = type;
+	fdata->mode = mode;
 	return fdata;
 
 error_alloc:
@@ -262,7 +292,8 @@ void lttng_callstack_destroy(struct lttng_ctx_field *field)
 }
 
 static
-int __lttng_add_callstack_generic(struct lttng_ctx **ctx, int mode)
+int __lttng_add_callstack_generic(struct lttng_ctx **ctx,
+		enum lttng_cs_ctx_modes mode)
 {
 	const char *ctx_name = cs_types[mode].name;
 	struct lttng_ctx_field *field;
@@ -336,8 +367,10 @@ int lttng_add_callstack_to_ctx(struct lttng_ctx **ctx, int type)
 	switch (type) {
 	case LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL:
 		return __lttng_add_callstack_generic(ctx, CALLSTACK_KERNEL);
+#ifdef CONFIG_X86
 	case LTTNG_KERNEL_CONTEXT_CALLSTACK_USER:
 		return __lttng_add_callstack_generic(ctx, CALLSTACK_USER);
+#endif
 	default:
 		return -EINVAL;
 	}