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 about 800 bytes per CPU on 64-bit host and a depth of 25.
  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 25
  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 static
 124 struct stack_trace *stack_trace_context(struct lttng_ctx_field *field,
 125                                         struct lib_ring_buffer_ctx *ctx)
 126 {
 127         int nesting;
 128         struct lttng_cs *cs;
 129         struct field_data *fdata = field->priv;
 130
 131         /*
 132          * get_cpu() is not required, preemption is already
 133          * disabled while event is written.
 134          *
 135          * max nesting is checked in lib_ring_buffer_get_cpu().
 136          * Check it again as a safety net.
 137          */
 138         cs = per_cpu_ptr(fdata->cs_percpu, ctx->cpu);
 139         nesting = per_cpu(lib_ring_buffer_nesting, ctx->cpu) - 1;
 140         if (nesting >= RING_BUFFER_MAX_NESTING) {
 141                 return NULL;
 142         }
 143         return &cs->dispatch[nesting].stack_trace;
 144 }
 145
 146 /*
 147  * In order to reserve the correct size, the callstack is computed. The
 148  * resulting callstack is saved to be accessed in the record step.
 149  */
 150 static
 151 size_t lttng_callstack_get_size(size_t offset, struct lttng_ctx_field *field,
 152                                 struct lib_ring_buffer_ctx *ctx,
 153                                 struct lttng_channel *chan)
 154 {
 155         size_t size = 0;
 156         struct stack_trace *trace;
 157         struct field_data *fdata = field->priv;
 158
 159         /* do not write data if no space is available */
 160         trace = stack_trace_context(field, ctx);
 161         if (unlikely(!trace)) {
 162                 size += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
 163                 size += sizeof(unsigned int);
 164                 size += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
 165                 return size;
 166         }
 167
 168         /* reset stack trace, no need to clear memory */
 169         trace->nr_entries = 0;
 170
 171         /* do the real work and reserve space */
 172         cs_types[fdata->mode].save_func(trace);
 173         /*
 174          * Remove final ULONG_MAX delimiter. If we cannot find it, add
 175          * our own marker to show that the stack is incomplete. This is
 176          * more compact for a trace.
 177          */
 178         if (trace->nr_entries > 0
 179                         && trace->entries[trace->nr_entries - 1] == ULONG_MAX) {
 180                 trace->nr_entries--;
 181         }
 182         size += lib_ring_buffer_align(offset, lttng_alignof(unsigned int));
 183         size += sizeof(unsigned int);
 184         size += lib_ring_buffer_align(offset, lttng_alignof(unsigned long));
 185         size += sizeof(unsigned long) * trace->nr_entries;
 186         /* Add our own ULONG_MAX delimiter to show incomplete stack. */
 187         if (trace->nr_entries == trace->max_entries)
 188                 size += sizeof(unsigned long);
 189         return size;
 190 }
 191
 192 static
 193 void lttng_callstack_record(struct lttng_ctx_field *field,
 194                         struct lib_ring_buffer_ctx *ctx,
 195                         struct lttng_channel *chan)
 196 {
 197         struct stack_trace *trace = stack_trace_context(field, ctx);
 198         unsigned int nr_seq_entries;
 199
 200         if (unlikely(!trace)) {
 201                 nr_seq_entries = 0;
 202                 lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
 203                 chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
 204                 lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
 205                 return;
 206         }
 207         lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned int));
 208         nr_seq_entries = trace->nr_entries;
 209         if (trace->nr_entries == trace->max_entries)
 210                 nr_seq_entries++;
 211         chan->ops->event_write(ctx, &nr_seq_entries, sizeof(unsigned int));
 212         lib_ring_buffer_align_ctx(ctx, lttng_alignof(unsigned long));
 213         chan->ops->event_write(ctx, trace->entries,
 214                         sizeof(unsigned long) * trace->nr_entries);
 215         /* Add our own ULONG_MAX delimiter to show incomplete stack. */
 216         if (trace->nr_entries == trace->max_entries) {
 217                 unsigned long delim = ULONG_MAX;
 218
 219                 chan->ops->event_write(ctx, &delim, sizeof(unsigned long));
 220         }
 221 }
 222
 223 static
 224 void field_data_free(struct field_data *fdata)
 225 {
 226         if (!fdata)
 227                 return;
 228         free_percpu(fdata->cs_percpu);
 229         kfree(fdata);
 230 }
 231
 232 static
 233 struct field_data __percpu *field_data_create(enum lttng_cs_ctx_modes mode)
 234 {
 235         int cpu, i;
 236         struct lttng_cs __percpu *cs_set;
 237         struct field_data *fdata;
 238
 239         fdata = kzalloc(sizeof(*fdata), GFP_KERNEL);
 240         if (!fdata)
 241                 return NULL;
 242         cs_set = alloc_percpu(struct lttng_cs);
 243         if (!cs_set)
 244                 goto error_alloc;
 245
 246         fdata->cs_percpu = cs_set;
 247         for_each_possible_cpu(cpu) {
 248                 struct lttng_cs *cs;
 249
 250                 cs = per_cpu_ptr(cs_set, cpu);
 251                 for (i = 0; i < RING_BUFFER_MAX_NESTING; i++) {
 252                         struct lttng_cs_dispatch *dispatch;
 253
 254                         dispatch = &cs->dispatch[i];
 255                         dispatch->stack_trace.entries = dispatch->entries;
 256                         dispatch->stack_trace.max_entries = MAX_ENTRIES;
 257                 }
 258         }
 259         fdata->mode = mode;
 260         return fdata;
 261
 262 error_alloc:
 263         field_data_free(fdata);
 264         return NULL;
 265 }
 266
 267 static
 268 void lttng_callstack_destroy(struct lttng_ctx_field *field)
 269 {
 270         struct field_data *fdata = field->priv;
 271
 272         field_data_free(fdata);
 273 }
 274
 275 static
 276 int __lttng_add_callstack_generic(struct lttng_ctx **ctx,
 277                 enum lttng_cs_ctx_modes mode)
 278 {
 279         const char *ctx_name = cs_types[mode].name;
 280         struct lttng_ctx_field *field;
 281         struct field_data *fdata;
 282         int ret;
 283
 284         ret = init_type(mode);
 285         if (ret)
 286                 return ret;
 287         field = lttng_append_context(ctx);
 288         if (!field)
 289                 return -ENOMEM;
 290         if (lttng_find_context(*ctx, ctx_name)) {
 291                 ret = -EEXIST;
 292                 goto error_find;
 293         }
 294         fdata = field_data_create(mode);
 295         if (!fdata) {
 296                 ret = -ENOMEM;
 297                 goto error_create;
 298         }
 299
 300         field->event_field.name = ctx_name;
 301         field->event_field.type.atype = atype_sequence;
 302         field->event_field.type.u.sequence.elem_type.atype = atype_integer;
 303         field->event_field.type.u.sequence.elem_type.u.basic.integer.size = sizeof(unsigned long) * CHAR_BIT;
 304         field->event_field.type.u.sequence.elem_type.u.basic.integer.alignment = lttng_alignof(long) * CHAR_BIT;
 305         field->event_field.type.u.sequence.elem_type.u.basic.integer.signedness = lttng_is_signed_type(unsigned long);
 306         field->event_field.type.u.sequence.elem_type.u.basic.integer.reverse_byte_order = 0;
 307         field->event_field.type.u.sequence.elem_type.u.basic.integer.base = 16;
 308         field->event_field.type.u.sequence.elem_type.u.basic.integer.encoding = lttng_encode_none;
 309
 310         field->event_field.type.u.sequence.length_type.atype = atype_integer;
 311         field->event_field.type.u.sequence.length_type.u.basic.integer.size = sizeof(unsigned int) * CHAR_BIT;
 312         field->event_field.type.u.sequence.length_type.u.basic.integer.alignment = lttng_alignof(unsigned int) * CHAR_BIT;
 313         field->event_field.type.u.sequence.length_type.u.basic.integer.signedness = lttng_is_signed_type(unsigned int);
 314         field->event_field.type.u.sequence.length_type.u.basic.integer.reverse_byte_order = 0;
 315         field->event_field.type.u.sequence.length_type.u.basic.integer.base = 10;
 316         field->event_field.type.u.sequence.length_type.u.basic.integer.encoding = lttng_encode_none;
 317
 318         field->get_size_arg = lttng_callstack_get_size;
 319         field->record = lttng_callstack_record;
 320         field->priv = fdata;
 321         field->destroy = lttng_callstack_destroy;
 322         wrapper_vmalloc_sync_all();
 323         return 0;
 324
 325 error_create:
 326         field_data_free(fdata);
 327 error_find:
 328         lttng_remove_context_field(ctx, field);
 329         return ret;
 330 }
 331
 332 /**
 333  *      lttng_add_callstack_to_ctx - add callstack event context
 334  *
 335  *      @ctx: the lttng_ctx pointer to initialize
 336  *      @type: the context type
 337  *
 338  *      Supported callstack type supported:
 339  *      LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL
 340  *              Records the callstack of the kernel
 341  *      LTTNG_KERNEL_CONTEXT_CALLSTACK_USER
 342  *              Records the callstack of the userspace program (from the kernel)
 343  *
 344  * Return 0 for success, or error code.
 345  */
 346 int lttng_add_callstack_to_ctx(struct lttng_ctx **ctx, int type)
 347 {
 348         switch (type) {
 349         case LTTNG_KERNEL_CONTEXT_CALLSTACK_KERNEL:
 350                 return __lttng_add_callstack_generic(ctx, CALLSTACK_KERNEL);
 351         case LTTNG_KERNEL_CONTEXT_CALLSTACK_USER:
 352                 return __lttng_add_callstack_generic(ctx, CALLSTACK_USER);
 353         default:
 354                 return -EINVAL;
 355         }
 356 }
 357 EXPORT_SYMBOL_GPL(lttng_add_callstack_to_ctx);
 358
 359 MODULE_LICENSE("GPL and additional rights");
 360 MODULE_AUTHOR("Francis Giraldeau");
 361 MODULE_DESCRIPTION("Linux Trace Toolkit Callstack Support");