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 event. It records
  24  * either the kernel or the userspace callstack, up to a max depth. The
  25  * context is a CTF sequence, such that it uses only the space required
  26  * for the number of callstack entries.
  27  *
  28  * It allocates callstack buffers per-CPU up to 4 interrupt nesting.
  29  * This nesting limit is the same as defined in the ring buffer. It
  30  * therefore uses a fixed amount of memory, proportional to the number
  31  * of CPUs:
  32  *
  33  *   size = cpus * nest * depth * sizeof(unsigned long)
  34  *
  35  * Which is 4096 bytes per CPU on 64-bit host and a depth of 128.
  36  * The allocation is done at the initialization to avoid memory
  37  * allocation overhead while tracing, using a shallow stack.
  38  *
  39  * The kernel callstack is recovered using save_stack_trace(), and the
  40  * userspace callstack uses save_stack_trace_user(). They rely on frame
  41  * pointers. These are usually available for the kernel, but the
  42  * compiler option -fomit-frame-pointer frequently used in popular Linux
  43  * distributions may cause the userspace callstack to be unreliable, and
  44  * is a known limitation of this approach. If frame pointers are not
  45  * available, it produces no error, but the callstack will be empty. We
  46  * still provide the feature, because it works well for runtime
  47  * environments having frame pointers. In the future, unwind support
  48  * and/or last branch record may provide a solution to this problem.
  49  *
  50  * The symbol name resolution is left to the trace reader.
  51  */
  52
  53 #include <linux/module.h>
  54 #include <linux/slab.h>
  55 #include <linux/sched.h>
  56 #include <linux/utsname.h>
  57 #include <linux/stacktrace.h>
  58 #include <linux/spinlock.h>
  59 #include "lttng-events.h"
  60 #include "wrapper/ringbuffer/backend.h"
  61 #include "wrapper/ringbuffer/frontend.h"
  62 #include "wrapper/vmalloc.h"
  63 #include "lttng-tracer.h"
  64
  65 #define MAX_ENTRIES 128
  66
  67 enum lttng_cs_ctx_modes {
  68         CALLSTACK_KERNEL = 0,
  69         CALLSTACK_USER = 1,
  70         NR_CALLSTACK_MODES,
  71 };
  72
  73 struct lttng_cs_dispatch {
  74         struct stack_trace stack_trace;
  75         unsigned long entries[MAX_ENTRIES];
  76 };
  77
  78 struct lttng_cs {
  79         struct lttng_cs_dispatch dispatch[RING_BUFFER_MAX_NESTING];
  80 };
  81
  82 struct field_data {
  83         struct lttng_cs __percpu *cs_percpu;
  84         enum lttng_cs_ctx_modes mode;
  85 };
  86
  87 struct lttng_cs_type {
  88         const char *name;
  89         const char *save_func_name;
  90         void (*save_func)(struct stack_trace *trace);
  91 };
  92
  93 static struct lttng_cs_type cs_types[] = {
  94         {
  95                 .name           = "callstack_kernel",
  96                 .save_func_name = "save_stack_trace",
  97                 .save_func      = NULL,
  98         },
  99         {
 100                 .name           = "callstack_user",
 101                 .save_func_name = "save_stack_trace_user",
 102                 .save_func      = NULL,
 103         },
 104 };
 105
 106 static
 107 int init_type(enum lttng_cs_ctx_modes mode)
 108 {
 109         unsigned long func;
 110
 111         if (cs_types[mode].save_func)
 112                 return 0;
 113         func = kallsyms_lookup_funcptr(cs_types[mode].save_func_name);
 114         if (!func) {
 115                 printk(KERN_WARNING "LTTng: symbol lookup failed: %s\n",
 116                                 cs_types[mode].save_func_name);
 117                 return -EINVAL;
 118         }
 119         cs_types[mode].save_func = (void *) func;
 120         return 0;
 121 }
 122
 123 /* Keep track of nesting inside userspace callstack context code */
 124 DEFINE_PER_CPU(int, callstack_user_nesting);
 125
 126 static
 127 struct stack_trace *stack_trace_context(struct lttng_ctx_field *field,
 128                                         struct lib_ring_buffer_ctx *ctx)
 129 {
 130         int buffer_nesting, cs_user_nesting;
 131         struct lttng_cs *cs;
 132         struct field_data *fdata = field->priv;
 133
 134         /*
 135          * Do not gather the userspace callstack context when the event was
 136          * triggered by the userspace callstack context saving mechanism.
 137          */
 138         cs_user_nesting = per_cpu(callstack_user_nesting, ctx->cpu);
 139
 140         if (fdata->mode == CALLSTACK_USER && cs_user_nesting >= 1)
 141                 return NULL;
 142
 143         /*
 144          * get_cpu() is not required, preemption is already
 145          * disabled while event is written.
 146          *
 147          * max nesting is checked in lib_ring_buffer_get_cpu().
 148          * Check it again as a safety net.
 149          */
 150         cs = per_cpu_ptr(fdata->cs_percpu, ctx->cpu);
 151         buffer_nesting = per_cpu(lib_ring_buffer_nesting, ctx->cpu) - 1;
 152         if (buffer_nesting >= RING_BUFFER_MAX_NESTING)
 153                 return NULL;
 154
 155         return &cs->dispatch[buffer_nesting].stack_trace;
 156 }
 157
 158 /*
 159  * In order to reserve the correct size, the callstack is computed. The
 160  * resulting callstack is saved to be accessed in the record step.
 161  */
 162 static
 163 size_t lttng_callstack_get_size(size_t offset, struct lttng_ctx_field *field,
 164                                 struct lib_ring_buffer_ctx *ctx,
 165                                 struct lttng_channel *chan)
 166 {
 167         struct stack_trace *trace;
 168         struct field_data *fdata = field->priv;
 169         size_t orig_offset = offset;
 170
 171         /* do not write data if no space is available */
 172         trace = stack_trace_context(field, ctx);
 173         if (unlikely(!trace)) {
 174                 offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
 175                 offset += sizeof(unsigned int);
 176                 offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
 177                 return offset - orig_offset;
 178         }
 179
 180         /* reset stack trace, no need to clear memory */
 181         trace->nr_entries = 0;
 182
 183         if (fdata->mode == CALLSTACK_USER)
 184                 ++per_cpu(callstack_user_nesting, ctx->cpu);
 185
 186         /* do the real work and reserve space */
 187         cs_types[fdata->mode].save_func(trace);
 188
 189         if (fdata->mode == CALLSTACK_USER)
 190                 per_cpu(callstack_user_nesting, ctx->cpu)--;
 191
 192         /*
 193          * Remove final ULONG_MAX delimiter. If we cannot find it, add
 194          * our own marker to show that the stack is incomplete. This is
 195          * more compact for a trace.
 196          */
 197         if (trace->nr_entries > 0
 198                         && trace->entries[trace->nr_entries - 1] == ULONG_MAX) {
 199                 trace->nr_entries--;
 200         }
 201         offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
 202         offset += sizeof(unsigned int);
 203         offset += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
 204         offset += sizeof(unsigned long) * trace->nr_entries;
 205         /* Add our own ULONG_MAX delimiter to show incomplete stack. */
 206         if (trace->nr_entries == trace->max_entries)
 207                 offset += sizeof(unsigned long);
 208         return offset - orig_offset;
 209 }
 210
 211 static
 212 void lttng_callstack_record(struct lttng_ctx_field *field,
 213                         struct lib_ring_buffer_ctx *ctx,
 214                         struct lttng_channel *chan)
 215 {
 216         struct stack_trace *trace = stack_trace_context(field, ctx);
 217         unsigned int nr_seq_entries;
 218
 219         if (unlikely(!trace)) {
 220                 nr_seq_entries = 0;
 221                 lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
 222                 chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
 223                 lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
 224                 return;
 225         }
 226         lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
 227         nr_seq_entries = trace->nr_entries;
 228         if (trace->nr_entries == trace->max_entries)
 229                 nr_seq_entries++;
 230         chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
 231         lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
 232         chan->ops->event_write(ctx, trace->entries,
 233                         sizeof(unsigned long) * trace->nr_entries);
 234         /* Add our own ULONG_MAX delimiter to show incomplete stack. */
 235         if (trace->nr_entries == trace->max_entries) {
 236                 unsigned long delim = ULONG_MAX;
 237
 238                 chan->ops->event_write(ctx, &delim, sizeof(unsigned long));
 239         }
 240 }
 241
 242 static
 243 void field_data_free(struct field_data *fdata)
 244 {
 245         if (!fdata)
 246                 return;
 247         free_percpu(fdata->cs_percpu);
 248         kfree(fdata);
 249 }
 250
 251 static
 252 struct field_data __percpu *field_data_create(enum lttng_cs_ctx_modes mode)
 253 {
 254         int cpu, i;
 255         struct lttng_cs __percpu *cs_set;
 256         struct field_data *fdata;
 257
 258         fdata = kzalloc(sizeof(*fdata), GFP_KERNEL);
 259         if (!fdata)
 260                 return NULL;
 261         cs_set = alloc_percpu(struct lttng_cs);
 262         if (!cs_set)
 263                 goto error_alloc;
 264
 265         fdata->cs_percpu = cs_set;
 266         for_each_possible_cpu(cpu) {
 267                 struct lttng_cs *cs;
 268
 269                 cs = per_cpu_ptr(cs_set, cpu);
 270                 for (i = 0; i < RING_BUFFER_MAX_NESTING; i++) {
 271                         struct lttng_cs_dispatch *dispatch;
 272
 273                         dispatch = &cs->dispatch[i];
 274                         dispatch->stack_trace.entries = dispatch->entries;
 275                         dispatch->stack_trace.max_entries = MAX_ENTRIES;
 276                 }
 277         }
 278         fdata->mode = mode;
 279         return fdata;
 280
 281 error_alloc:
 282         field_data_free(fdata);
 283         return NULL;
 284 }
 285
 286 static
 287 void lttng_callstack_destroy(struct lttng_ctx_field *field)
 288 {
 289         struct field_data *fdata = field->priv;
 290
 291         field_data_free(fdata);
 292 }
 293
 294 static
 295 int __lttng_add_callstack_generic(struct lttng_ctx **ctx,
 296                 enum lttng_cs_ctx_modes mode)
 297 {
 298         const char *ctx_name = cs_types[mode].name;
 299         struct lttng_ctx_field *field;
 300         struct field_data *fdata;
 301         int ret;
 302
 303         ret = init_type(mode);
 304         if (ret)
 305                 return ret;
 306         field = lttng_append_context(ctx);
 307         if (!field)
 308                 return -ENOMEM;
 309         if (lttng_find_context(*ctx, ctx_name)) {
 310                 ret = -EEXIST;
 311                 goto error_find;
 312         }
 313         fdata = field_data_create(mode);
 314         if (!fdata) {
 315                 ret = -ENOMEM;
 316                 goto error_create;
 317         }
 318
 319         field->event_field.name = ctx_name;
 320         field->event_field.type.atype = atype_sequence;
 321         field->event_field.type.u.sequence.elem_type.atype = atype_integer;
 322         field->event_field.type.u.sequence.elem_type.u.basic.integer.size = sizeof(unsigned long) * CHAR_BIT;
 323         field->event_field.type.u.sequence.elem_type.u.basic.integer.alignment = lttng_alignof(long) * CHAR_BIT;
 324         field->event_field.type.u.sequence.elem_type.u.basic.integer.signedness = lttng_is_signed_type(unsigned long);
 325         field->event_field.type.u.sequence.elem_type.u.basic.integer.reverse_byte_order = 0;
 326         field->event_field.type.u.sequence.elem_type.u.basic.integer.base = 16;
 327         field->event_field.type.u.sequence.elem_type.u.basic.integer.encoding = lttng_encode_none;
 328
 329         field->event_field.type.u.sequence.length_type.atype = atype_integer;
 330         field->event_field.type.u.sequence.length_type.u.basic.integer.size = sizeof(unsigned int) * CHAR_BIT;
 331         field->event_field.type.u.sequence.length_type.u.basic.integer.alignment = lttng_alignof(unsigned int) * CHAR_BIT;
 332         field->event_field.type.u.sequence.length_type.u.basic.integer.signedness = lttng_is_signed_type(unsigned int);
 333         field->event_field.type.u.sequence.length_type.u.basic.integer.reverse_byte_order = 0;
 334         field->event_field.type.u.sequence.length_type.u.basic.integer.base = 10;
 335         field->event_field.type.u.sequence.length_type.u.basic.integer.encoding = lttng_encode_none;
 336
 337         field->get_size_arg = lttng_callstack_get_size;
 338         field->record = lttng_callstack_record;
 339         field->priv = fdata;
 340         field->destroy = lttng_callstack_destroy;
 341         wrapper_vmalloc_sync_all();
 342         return 0;
 343
 344 error_create:
 345         field_data_free(fdata);
 346 error_find:
 347         lttng_remove_context_field(ctx, field);
 348         return ret;
 349 }
 350
 351 /**
 352  *      lttng_add_callstack_to_ctx - add callstack event context
 353  *
 354  *      @ctx: the lttng_ctx pointer to initialize
 355  *      @type: the context type
 356  *
 357  *      Supported callstack type supported:
 358  *      LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL
 359  *              Records the callstack of the kernel
 360  *      LTTNG_KERNEL_CONTEXT_CALLSTACK_USER
 361  *              Records the callstack of the userspace program (from the kernel)
 362  *
 363  * Return 0 for success, or error code.
 364  */
 365 int lttng_add_callstack_to_ctx(struct lttng_ctx **ctx, int type)
 366 {
 367         switch (type) {
 368         case LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL:
 369                 return __lttng_add_callstack_generic(ctx, CALLSTACK_KERNEL);
 370 #ifdef CONFIG_X86
 371         case LTTNG_KERNEL_CONTEXT_CALLSTACK_USER:
 372                 return __lttng_add_callstack_generic(ctx, CALLSTACK_USER);
 373 #endif
 374         default:
 375                 return -EINVAL;
 376         }
 377 }
 378 EXPORT_SYMBOL_GPL(lttng_add_callstack_to_ctx);
 379
 380 MODULE_LICENSE("GPL and additional rights");
 381 MODULE_AUTHOR("Francis Giraldeau");
 382 MODULE_DESCRIPTION("Linux Trace Toolkit Callstack Support");