X-Git-Url: http://git.lttng.org/?a=blobdiff_plain;f=lttng-context-callstack.c;h=4dd984be40fee4e3ef6ba9bbbc2d153244f50ee1;hb=0bb47c89c36280339e560a15a3f66d64293b5304;hp=d95cbc2eef3f7117f0fcb38508ff173106f9fadf;hpb=3c1a57e8ce82366ce96774e4ac1d7481c561cc4c;p=lttng-modules.git diff --git a/lttng-context-callstack.c b/lttng-context-callstack.c index d95cbc2e..4dd984be 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 about 800 bytes per CPU 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. * * 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. */ @@ -61,15 +62,26 @@ #include "wrapper/vmalloc.h" #include "lttng-tracer.h" -#define MAX_ENTRIES 25 /* BUG: saving more than 30 entries causes trace corruption */ +#define MAX_ENTRIES 25 + +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 { @@ -78,26 +90,21 @@ 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; @@ -133,7 +140,7 @@ struct stack_trace *stack_trace_context(struct lttng_ctx_field *field, if (nesting >= RING_BUFFER_MAX_NESTING) { return NULL; } - return &cs->items[nesting]; + return &cs->dispatch[nesting].stack_trace; } /* @@ -151,65 +158,85 @@ size_t lttng_callstack_get_size(size_t offset, struct lttng_ctx_field *field, /* do not write data if no space is available */ trace = stack_trace_context(field, ctx); - if (!trace) - return 0; + if (unlikely(!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)); + return size; + } /* reset stack trace, no need to clear memory */ trace->nr_entries = 0; /* do the real work and reserve space */ cs_types[fdata->mode].save_func(trace); + /* + * 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--; + } 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; + /* Add our own ULONG_MAX delimiter to show incomplete stack. */ + if (trace->nr_entries == trace->max_entries) + size += sizeof(unsigned long); return size; } 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); @@ -218,17 +245,18 @@ struct field_data __percpu *field_data_create(unsigned int entries, int type) 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: @@ -245,7 +273,8 @@ 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; @@ -262,7 +291,7 @@ int __lttng_add_callstack_generic(struct lttng_ctx **ctx, int mode) ret = -EEXIST; goto error_find; } - fdata = field_data_create(MAX_ENTRIES, mode); + fdata = field_data_create(mode); if (!fdata) { ret = -ENOMEM; goto error_create;