-/*
+/* SPDX-License-Identifier: (GPL-2.0 or LGPL-2.1)
+ *
* lttng-context-callstack.c
*
* LTTng callstack event context.
* Copyright (C) 2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
* Copyright (C) 2014 Francis Giraldeau <francis.giraldeau@gmail.com>
*
- * This library is free software; you can redistribute it and/or
- * modify it under the terms of the GNU Lesser General Public
- * License as published by the Free Software Foundation; only
- * version 2.1 of the License.
- *
- * This library is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * Lesser General Public License for more details.
+ * 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.
*
- * You should have received a copy of the GNU Lesser General Public
- * 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.
- *
- * 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.
*/
#include "wrapper/vmalloc.h"
#include "lttng-tracer.h"
-#define MAX_ENTRIES 25 /* BUG: saving more than 30 entries causes trace corruption */
+#define MAX_ENTRIES 128
+
+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 stack_trace items[RING_BUFFER_MAX_NESTING];
+ struct lttng_cs_dispatch dispatch[RING_BUFFER_MAX_NESTING];
};
struct field_data {
- int mode;
struct lttng_cs __percpu *cs_percpu;
+ enum lttng_cs_ctx_modes mode;
};
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",
- .save_func_name = "save_stack_trace",
- .save_func = NULL,
+ .name = "callstack_kernel",
+ .save_func_name = "save_stack_trace",
+ .save_func = NULL,
},
{
- .name = "callstack_user",
- .save_func_name = "save_stack_trace_user",
- .save_func = NULL,
+ .name = "callstack_user",
+ .save_func_name = "save_stack_trace_user",
+ .save_func = NULL,
},
};
static
-int init_type(int mode)
+int init_type(enum lttng_cs_ctx_modes mode)
{
unsigned long func;
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.
* 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->items[nesting];
+
+ return &cs->dispatch[buffer_nesting].stack_trace;
}
/*
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);
- 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;
- return size;
+
+ 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
+ * more compact for a trace.
+ */
+ if (trace->nr_entries > 0
+ && trace->entries[trace->nr_entries - 1] == ULONG_MAX) {
+ 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)
+ offset += sizeof(unsigned long);
+ return offset - orig_offset;
}
static
void lttng_callstack_record(struct lttng_ctx_field *field,
- struct lib_ring_buffer_ctx *ctx,
- struct lttng_channel *chan)
+ struct lib_ring_buffer_ctx *ctx,
+ struct lttng_channel *chan)
{
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));
- chan->ops->event_write(ctx, &trace->nr_entries, sizeof(unsigned int));
+ nr_seq_entries = trace->nr_entries;
+ if (trace->nr_entries == trace->max_entries)
+ nr_seq_entries++;
+ chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
chan->ops->event_write(ctx, trace->entries,
sizeof(unsigned long) * trace->nr_entries);
+ /* Add our own ULONG_MAX delimiter to show incomplete stack. */
+ if (trace->nr_entries == trace->max_entries) {
+ unsigned long delim = ULONG_MAX;
+
+ chan->ops->event_write(ctx, &delim, sizeof(unsigned long));
+ }
}
static
void field_data_free(struct field_data *fdata)
{
- int cpu, i;
- struct lttng_cs *cs;
-
if (!fdata)
return;
- for_each_possible_cpu(cpu) {
- cs = per_cpu_ptr(fdata->cs_percpu, cpu);
- for (i = 0; i < RING_BUFFER_MAX_NESTING; i++) {
- kfree(cs->items[i].entries);
- }
- }
free_percpu(fdata->cs_percpu);
kfree(fdata);
}
static
-struct field_data __percpu *field_data_create(unsigned int entries, int type)
+struct field_data __percpu *field_data_create(enum lttng_cs_ctx_modes mode)
{
int cpu, i;
- struct stack_trace *item;
- struct lttng_cs *cs;
struct lttng_cs __percpu *cs_set;
- struct field_data* fdata;
+ struct field_data *fdata;
- fdata = kzalloc(sizeof(unsigned long) * entries, GFP_KERNEL);
+ fdata = kzalloc(sizeof(*fdata), GFP_KERNEL);
if (!fdata)
return NULL;
cs_set = alloc_percpu(struct lttng_cs);
fdata->cs_percpu = cs_set;
for_each_possible_cpu(cpu) {
+ struct lttng_cs *cs;
+
cs = per_cpu_ptr(cs_set, cpu);
for (i = 0; i < RING_BUFFER_MAX_NESTING; i++) {
- item = &cs->items[i];
- item->entries = kzalloc(sizeof(unsigned long) * entries, GFP_KERNEL);
- if (!item->entries) {
- goto error_alloc;
- }
- item->max_entries = entries;
+ struct lttng_cs_dispatch *dispatch;
+
+ 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:
}
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;
ret = -EEXIST;
goto error_find;
}
- fdata = field_data_create(MAX_ENTRIES, mode);
+ fdata = field_data_create(mode);
if (!fdata) {
ret = -ENOMEM;
goto error_create;
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;
}
}
EXPORT_SYMBOL_GPL(lttng_add_callstack_to_ctx);
-
-MODULE_LICENSE("GPL and additional rights");
-MODULE_AUTHOR("Francis Giraldeau");
-MODULE_DESCRIPTION("Linux Trace Toolkit Callstack Support");
projects / lttng-modules.git / blobdiff