lttng-context-callstack.c

   1 /*
   2  * lttng-context-callstack.c
   3  *
   4  * LTTng callstack event context.
   5  *
   6  * Copyright (C) 2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  * Copyright (C) 2014 Francis Giraldeau <francis.giraldeau@gmail.com>
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; only
  12  * version 2.1 of the License.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  *
  23  * The callstack context can be added to any kernel
  24  * event. It records either the kernel or the userspace callstack, up to a
  25  * max depth. The context is a CTF sequence, such that it uses only the space
  26  * required for the number of callstack entries.
  27  *
  28  * It allocates callstack buffers per-CPU up to 4 interrupt nesting. This
  29  * nesting limit is the same as defined in the ring buffer. It therefore uses a
  30  * fixed amount of memory, proportional to the number of CPUs:
  31  *
  32  *   size = cpus * nest * depth * sizeof(unsigned long)
  33  *
  34  * Which is about 800 bytes per-CPUs on 64-bit host and a depth of 25. The
  35  * allocation is done at the initialization to avoid memory allocation
  36  * overhead while tracing, using a shallow stack.
  37  *
  38  * The kernel callstack is recovered using save_stack_trace(), and the
  39  * userspace callstack uses save_stack_trace_user(). They rely on frame
  40  * pointers. These are usually available for the kernel, but the compiler
  41  * option -fomit-frame-pointer frequently used in popular Linux distributions
  42  * may cause the userspace callstack to be unreliable, and is a known
  43  * limitation of this approach. If frame pointers are not available, it
  44  * produces no error, but the callstack will be empty. We still provide the
  45  * feature, because it works well for runtime environments having frame
  46  * pointers. In the future, unwind support and/or last branch record may
  47  * provide a solution to this problem.
  48  *
  49  * The symbol name resolution is left to the trace reader.
  50  */
  51
  52 #include <linux/module.h>
  53 #include <linux/slab.h>
  54 #include <linux/sched.h>
  55 #include <linux/utsname.h>
  56 #include <linux/stacktrace.h>
  57 #include <linux/spinlock.h>
  58 #include "lttng-events.h"
  59 #include "wrapper/ringbuffer/backend.h"
  60 #include "wrapper/ringbuffer/frontend.h"
  61 #include "wrapper/vmalloc.h"
  62 #include "lttng-tracer.h"
  63
  64 #define MAX_ENTRIES 25
  65
  66 struct lttng_cs_nesting {
  67         struct stack_trace stack_trace;
  68         unsigned long entries[MAX_ENTRIES];
  69 };
  70
  71 struct lttng_cs {
  72         struct lttng_cs_nesting level[RING_BUFFER_MAX_NESTING];
  73 };
  74
  75 struct field_data {
  76         struct lttng_cs __percpu *cs_percpu;
  77         int mode;
  78 };
  79
  80 struct lttng_cs_type {
  81         const char *name;
  82         const char *save_func_name;
  83         void (*save_func)(struct stack_trace *trace);
  84 };
  85
  86 enum lttng_cs_ctx_modes {
  87         CALLSTACK_KERNEL = 0,
  88         CALLSTACK_USER = 1,
  89 };
  90
  91 static struct lttng_cs_type cs_types[] = {
  92         {
  93                 .name           = "callstack_kernel",
  94                 .save_func_name = "save_stack_trace",
  95                 .save_func      = NULL,
  96         },
  97         {
  98                 .name           = "callstack_user",
  99                 .save_func_name = "save_stack_trace_user",
 100                 .save_func      = NULL,
 101         },
 102 };
 103
 104 static
 105 int init_type(int mode)
 106 {
 107         unsigned long func;
 108
 109         if (cs_types[mode].save_func)
 110                 return 0;
 111         func = kallsyms_lookup_funcptr(cs_types[mode].save_func_name);
 112         if (!func) {
 113                 printk(KERN_WARNING "LTTng: symbol lookup failed: %s\n",
 114                                 cs_types[mode].save_func_name);
 115                 return -EINVAL;
 116         }
 117         cs_types[mode].save_func = (void *) func;
 118         return 0;
 119 }
 120
 121 static
 122 struct stack_trace *stack_trace_context(struct lttng_ctx_field *field,
 123                                         struct lib_ring_buffer_ctx *ctx)
 124 {
 125         int nesting;
 126         struct lttng_cs *cs;
 127         struct field_data *fdata = field->priv;
 128
 129         /*
 130          * get_cpu() is not required, preemption is already
 131          * disabled while event is written.
 132          *
 133          * max nesting is checked in lib_ring_buffer_get_cpu().
 134          * Check it again as a safety net.
 135          */
 136         cs = per_cpu_ptr(fdata->cs_percpu, ctx->cpu);
 137         nesting = per_cpu(lib_ring_buffer_nesting, ctx->cpu) - 1;
 138         if (nesting >= RING_BUFFER_MAX_NESTING) {
 139                 return NULL;
 140         }
 141         return &cs->level[nesting].stack_trace;
 142 }
 143
 144 /*
 145  * In order to reserve the correct size, the callstack is computed. The
 146  * resulting callstack is saved to be accessed in the record step.
 147  */
 148 static
 149 size_t lttng_callstack_get_size(size_t offset, struct lttng_ctx_field *field,
 150                                 struct lib_ring_buffer_ctx *ctx,
 151                                 struct lttng_channel *chan)
 152 {
 153         size_t size = 0;
 154         struct stack_trace *trace;
 155         struct field_data *fdata = field->priv;
 156
 157         /* do not write data if no space is available */
 158         trace = stack_trace_context(field, ctx);
 159         if (!trace)
 160                 return 0;
 161
 162         /* reset stack trace, no need to clear memory */
 163         trace->nr_entries = 0;
 164
 165         /* do the real work and reserve space */
 166         cs_types[fdata->mode].save_func(trace);
 167         /*
 168          * Remove final ULONG_MAX delimiter. If we cannot find it, add
 169          * our own marker to show that the stack is incomplete. This is
 170          * more compact for a trace.
 171          */
 172         if (trace->nr_entries > 0
 173                         && trace->entries[trace->nr_entries - 1] == ULONG_MAX) {
 174                 trace->nr_entries--;
 175         }
 176         size += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
 177         size += sizeof(unsigned int);
 178         size += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
 179         size += sizeof(unsigned long) * trace->nr_entries;
 180         /* Add our own ULONG_MAX delimiter to show incomplete stack. */
 181         if (trace->nr_entries == trace->max_entries)
 182                 size += sizeof(unsigned long);
 183         return size;
 184 }
 185
 186 static
 187 void lttng_callstack_record(struct lttng_ctx_field *field,
 188                         struct lib_ring_buffer_ctx *ctx,
 189                         struct lttng_channel *chan)
 190 {
 191         struct stack_trace *trace = stack_trace_context(field, ctx);
 192         unsigned int nr_seq_entries;
 193
 194         if (!trace)
 195                 return;
 196         lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
 197         nr_seq_entries = trace->nr_entries;
 198         if (trace->nr_entries == trace->max_entries)
 199                 nr_seq_entries++;
 200         chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
 201         lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
 202         chan->ops->event_write(ctx, trace->entries,
 203                         sizeof(unsigned long) * trace->nr_entries);
 204         /* Add our own ULONG_MAX delimiter to show incomplete stack. */
 205         if (trace->nr_entries == trace->max_entries) {
 206                 unsigned long delim = ULONG_MAX;
 207
 208                 chan->ops->event_write(ctx, &delim, sizeof(unsigned long));
 209         }
 210 }
 211
 212 static
 213 void field_data_free(struct field_data *fdata)
 214 {
 215         if (!fdata)
 216                 return;
 217         free_percpu(fdata->cs_percpu);
 218         kfree(fdata);
 219 }
 220
 221 static
 222 struct field_data __percpu *field_data_create(int type)
 223 {
 224         int cpu, i;
 225         struct lttng_cs __percpu *cs_set;
 226         struct field_data *fdata;
 227
 228         fdata = kzalloc(sizeof(*fdata), GFP_KERNEL);
 229         if (!fdata)
 230                 return NULL;
 231         cs_set = alloc_percpu(struct lttng_cs);
 232         if (!cs_set)
 233                 goto error_alloc;
 234
 235         fdata->cs_percpu = cs_set;
 236         for_each_possible_cpu(cpu) {
 237                 struct lttng_cs *cs;
 238
 239                 cs = per_cpu_ptr(cs_set, cpu);
 240                 for (i = 0; i < RING_BUFFER_MAX_NESTING; i++) {
 241                         struct lttng_cs_nesting *level;
 242
 243                         level = &cs->level[i];
 244                         level->stack_trace.entries = level->entries;
 245                         level->stack_trace.max_entries = MAX_ENTRIES;
 246                 }
 247         }
 248         fdata->mode = type;
 249         return fdata;
 250
 251 error_alloc:
 252         field_data_free(fdata);
 253         return NULL;
 254 }
 255
 256 static
 257 void lttng_callstack_destroy(struct lttng_ctx_field *field)
 258 {
 259         struct field_data *fdata = field->priv;
 260
 261         field_data_free(fdata);
 262 }
 263
 264 static
 265 int __lttng_add_callstack_generic(struct lttng_ctx **ctx, int mode)
 266 {
 267         const char *ctx_name = cs_types[mode].name;
 268         struct lttng_ctx_field *field;
 269         struct field_data *fdata;
 270         int ret;
 271
 272         ret = init_type(mode);
 273         if (ret)
 274                 return ret;
 275         field = lttng_append_context(ctx);
 276         if (!field)
 277                 return -ENOMEM;
 278         if (lttng_find_context(*ctx, ctx_name)) {
 279                 ret = -EEXIST;
 280                 goto error_find;
 281         }
 282         fdata = field_data_create(mode);
 283         if (!fdata) {
 284                 ret = -ENOMEM;
 285                 goto error_create;
 286         }
 287
 288         field->event_field.name = ctx_name;
 289         field->event_field.type.atype = atype_sequence;
 290         field->event_field.type.u.sequence.elem_type.atype = atype_integer;
 291         field->event_field.type.u.sequence.elem_type.u.basic.integer.size = sizeof(unsigned long) * CHAR_BIT;
 292         field->event_field.type.u.sequence.elem_type.u.basic.integer.alignment = lttng_alignof(long) * CHAR_BIT;
 293         field->event_field.type.u.sequence.elem_type.u.basic.integer.signedness = lttng_is_signed_type(unsigned long);
 294         field->event_field.type.u.sequence.elem_type.u.basic.integer.reverse_byte_order = 0;
 295         field->event_field.type.u.sequence.elem_type.u.basic.integer.base = 16;
 296         field->event_field.type.u.sequence.elem_type.u.basic.integer.encoding = lttng_encode_none;
 297
 298         field->event_field.type.u.sequence.length_type.atype = atype_integer;
 299         field->event_field.type.u.sequence.length_type.u.basic.integer.size = sizeof(unsigned int) * CHAR_BIT;
 300         field->event_field.type.u.sequence.length_type.u.basic.integer.alignment = lttng_alignof(unsigned int) * CHAR_BIT;
 301         field->event_field.type.u.sequence.length_type.u.basic.integer.signedness = lttng_is_signed_type(unsigned int);
 302         field->event_field.type.u.sequence.length_type.u.basic.integer.reverse_byte_order = 0;
 303         field->event_field.type.u.sequence.length_type.u.basic.integer.base = 10;
 304         field->event_field.type.u.sequence.length_type.u.basic.integer.encoding = lttng_encode_none;
 305
 306         field->get_size_arg = lttng_callstack_get_size;
 307         field->record = lttng_callstack_record;
 308         field->priv = fdata;
 309         field->destroy = lttng_callstack_destroy;
 310         wrapper_vmalloc_sync_all();
 311         return 0;
 312
 313 error_create:
 314         field_data_free(fdata);
 315 error_find:
 316         lttng_remove_context_field(ctx, field);
 317         return ret;
 318 }
 319
 320 /**
 321  *      lttng_add_callstack_to_ctx - add callstack event context
 322  *
 323  *      @ctx: the lttng_ctx pointer to initialize
 324  *      @type: the context type
 325  *
 326  *      Supported callstack type supported:
 327  *      LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL
 328  *              Records the callstack of the kernel
 329  *      LTTNG_KERNEL_CONTEXT_CALLSTACK_USER
 330  *              Records the callstack of the userspace program (from the kernel)
 331  *
 332  * Return 0 for success, or error code.
 333  */
 334 int lttng_add_callstack_to_ctx(struct lttng_ctx **ctx, int type)
 335 {
 336         switch (type) {
 337         case LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL:
 338                 return __lttng_add_callstack_generic(ctx, CALLSTACK_KERNEL);
 339         case LTTNG_KERNEL_CONTEXT_CALLSTACK_USER:
 340                 return __lttng_add_callstack_generic(ctx, CALLSTACK_USER);
 341         default:
 342                 return -EINVAL;
 343         }
 344 }
 345 EXPORT_SYMBOL_GPL(lttng_add_callstack_to_ctx);
 346
 347 MODULE_LICENSE("GPL and additional rights");
 348 MODULE_AUTHOR("Francis Giraldeau");
 349 MODULE_DESCRIPTION("Linux Trace Toolkit Callstack Support");