Message ID | 1466623636-24179-1-git-send-email-jdesfossez@efficios.com |
---|---|
State | Accepted, archived |
Headers | show |
On Wed, Jun 22, 2016 at 3:27 PM, Julien Desfossez <jdesfossez at efficios.com> wrote: > Allow enabling perf PMU counters by raw ID in addition to the generic > list already provided. The format for kernel tracing is > "perf:cpu:raw:rNNN:<name>" and "perf:thread:raw:rNNN:<name> for > user-space. The rNNN format is the same as perf-record(1) where NNN is a > hexadecimal event descriptor in the form of umask+eventsel. The <name> > field allows the user to give a more friendly name. to associate a clearer name to the counter. > > Example usage on Intel i7-3520M to get the unhalted reference cycles > (eventsel: 0x13c) count at privilege level 0 (umask: 0x00): > lttng add-context -k -t perf:cpu:raw:r0013c:x86unhalted > > Result in the trace: > sched_switch: { cpu_id = 3 }, { > perf_cpu_raw_r0013c_x86unhalted = 27632578 }, [...] > > Signed-off-by: Julien Desfossez <jdesfossez at efficios.com> > --- > doc/man/lttng-add-context.1.txt | 8 +++ > src/bin/lttng/commands/add_context.c | 94 +++++++++++++++++++++++++++++++++++- > 2 files changed, 101 insertions(+), 1 deletion(-) > > diff --git a/doc/man/lttng-add-context.1.txt b/doc/man/lttng-add-context.1.txt > index f995a7f..c43581a 100644 > --- a/doc/man/lttng-add-context.1.txt > +++ b/doc/man/lttng-add-context.1.txt > @@ -45,6 +45,14 @@ per-thread (`perf:thread:` prefix) counters. Currently, per-CPU counters > can only be used in the Linux kernel tracing domain, while per-thread > counters can only be used in the user space tracing domain. > > +It is also possible to enable PMU counters by raw ID using the > +`perf:cpu:raw:r<N>:<name>` or `perf:thread:raw:r<N>:<name>` format for the > +kernel and user-space respectively. `<N>` is a hexadecimal event descriptor > +which is the same format as perf-record(1): a concatenation of the `Umask > +value` and `Event number` provided by the processors manufacturer. The possible > +values for this field are processor-specific. The `<name>` field is used to > +give a symbolic name to the counter in the trace. ...to associate a... > + > Application-specific context fields can be added to a channel using the > following syntax: > > diff --git a/src/bin/lttng/commands/add_context.c b/src/bin/lttng/commands/add_context.c > index 2f43dc7..711c1e0 100644 > --- a/src/bin/lttng/commands/add_context.c > +++ b/src/bin/lttng/commands/add_context.c > @@ -87,6 +87,7 @@ enum perf_type { > PERF_TYPE_HARDWARE = 0, > PERF_TYPE_SOFTWARE = 1, > PERF_TYPE_HW_CACHE = 3, > + PERF_TYPE_RAW = 4, > }; > > enum perf_count_hard { > @@ -688,9 +689,88 @@ end: > } > > static > +int find_ctx_type_perf_raw(const char *ctx, struct ctx_type *type) > +{ > + char *next; Reduce scope of "next" to the "for" loop. > + int ret; > + int field_pos = 0; > + char *tmp_list; > + > + tmp_list = strdup(ctx); > + if (!tmp_list) { > + PERROR("strdup temp list"); > + ret = -ENOMEM; > + goto error; No need for a separate error label to skip the free() as tmp_list is NULL. > + } > + > + /* Looking for "perf:[cpu|thread]:raw:<mask>:<name>". */ > + for (;;) { > + next = strtok(tmp_list, ":"); > + if (!next) { > + break; > + } > + tmp_list = NULL; tmp_list is leaked. > + switch (field_pos) { > + case 0: > + if (strncmp(next, "perf", 4) != 0) { > + ret = -1; > + goto end; > + } > + break; > + case 1: > + if (strncmp(next, "cpu", 3) == 0) { > + type->opt->ctx_type = CONTEXT_PERF_CPU_COUNTER; > + } else if (strncmp(next, "thread", 4) == 0) { > + type->opt->ctx_type = CONTEXT_PERF_THREAD_COUNTER; > + } else { > + ret = -1; > + goto end; > + } > + break; > + case 2: > + if (strncmp(next, "raw", 3) != 0) { > + ret = -1; > + goto end; > + } > + break; > + case 3: > + if (strlen(next) < 2 || next[0] != 'r') { > + ERR("Wrong perf raw mask format: rNNN"); > + ret = -1; > + goto end; > + } > + type->opt->u.perf.config = strtoll(next + 1, NULL, 16); strtoll() should be checked for error (setting errno to 0 beforehand, and checking after for EINVAL and ERANGE, but also using the endptr). This currently allows, for instance: lttng add-context -u -t perf:thread:raw:rXYZ:oops Such malformed cases should be added to the test suite. > + break; > + case 4: > + /* name */ > + break; > + case 5: > + ERR("Too many ':' in perf raw format"); > + ret = -1; > + goto end; > + }; > + field_pos++; > + } > + > + if (field_pos < 5) { > + ERR("Wrong perf raw format"); "Invalid perf counter specifier, expected a specifier of the form perf:cpu:raw:rNNN:<name> or perf:thread:raw:rNNN:<name>" > + ret = -1; > + goto end; > + } > + > + ret = 0; > + goto end; > + > +end: > + free(tmp_list); > +error: > + return ret; > +} > + > +static > struct ctx_type *get_context_type(const char *ctx) > { > - int opt_index; > + int opt_index, ret; > struct ctx_type *type = NULL; > const char app_ctx_prefix[] = "$app."; > char *provider_name = NULL, *ctx_name = NULL; > @@ -713,6 +793,18 @@ struct ctx_type *get_context_type(const char *ctx) > goto found; > } > > + /* Check if ctx is a raw perf context. */ > + ret = find_ctx_type_perf_raw(ctx, type); > + if (ret == 0) { > + type->opt->u.perf.type = PERF_TYPE_RAW; > + type->opt->symbol = strdup(ctx); > + if (!type->opt->symbol) { > + PERROR("Copy perf field name"); > + goto not_found; > + } > + goto found; > + } > + > /* > * No match found against static contexts; check if it is an app > * context. > -- > 1.9.1 > > _______________________________________________ > lttng-dev mailing list > lttng-dev at lists.lttng.org > https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
diff --git a/doc/man/lttng-add-context.1.txt b/doc/man/lttng-add-context.1.txt index f995a7f..c43581a 100644 --- a/doc/man/lttng-add-context.1.txt +++ b/doc/man/lttng-add-context.1.txt @@ -45,6 +45,14 @@ per-thread (`perf:thread:` prefix) counters. Currently, per-CPU counters can only be used in the Linux kernel tracing domain, while per-thread counters can only be used in the user space tracing domain. +It is also possible to enable PMU counters by raw ID using the +`perf:cpu:raw:r<N>:<name>` or `perf:thread:raw:r<N>:<name>` format for the +kernel and user-space respectively. `<N>` is a hexadecimal event descriptor +which is the same format as perf-record(1): a concatenation of the `Umask +value` and `Event number` provided by the processors manufacturer. The possible +values for this field are processor-specific. The `<name>` field is used to +give a symbolic name to the counter in the trace. + Application-specific context fields can be added to a channel using the following syntax: diff --git a/src/bin/lttng/commands/add_context.c b/src/bin/lttng/commands/add_context.c index 2f43dc7..711c1e0 100644 --- a/src/bin/lttng/commands/add_context.c +++ b/src/bin/lttng/commands/add_context.c @@ -87,6 +87,7 @@ enum perf_type { PERF_TYPE_HARDWARE = 0, PERF_TYPE_SOFTWARE = 1, PERF_TYPE_HW_CACHE = 3, + PERF_TYPE_RAW = 4, }; enum perf_count_hard { @@ -688,9 +689,88 @@ end: } static +int find_ctx_type_perf_raw(const char *ctx, struct ctx_type *type) +{ + char *next; + int ret; + int field_pos = 0; + char *tmp_list; + + tmp_list = strdup(ctx); + if (!tmp_list) { + PERROR("strdup temp list"); + ret = -ENOMEM; + goto error; + } + + /* Looking for "perf:[cpu|thread]:raw:<mask>:<name>". */ + for (;;) { + next = strtok(tmp_list, ":"); + if (!next) { + break; + } + tmp_list = NULL; + switch (field_pos) { + case 0: + if (strncmp(next, "perf", 4) != 0) { + ret = -1; + goto end; + } + break; + case 1: + if (strncmp(next, "cpu", 3) == 0) { + type->opt->ctx_type = CONTEXT_PERF_CPU_COUNTER; + } else if (strncmp(next, "thread", 4) == 0) { + type->opt->ctx_type = CONTEXT_PERF_THREAD_COUNTER; + } else { + ret = -1; + goto end; + } + break; + case 2: + if (strncmp(next, "raw", 3) != 0) { + ret = -1; + goto end; + } + break; + case 3: + if (strlen(next) < 2 || next[0] != 'r') { + ERR("Wrong perf raw mask format: rNNN"); + ret = -1; + goto end; + } + type->opt->u.perf.config = strtoll(next + 1, NULL, 16); + break; + case 4: + /* name */ + break; + case 5: + ERR("Too many ':' in perf raw format"); + ret = -1; + goto end; + }; + field_pos++; + } + + if (field_pos < 5) { + ERR("Wrong perf raw format"); + ret = -1; + goto end; + } + + ret = 0; + goto end; + +end: + free(tmp_list); +error: + return ret; +} + +static struct ctx_type *get_context_type(const char *ctx) { - int opt_index; + int opt_index, ret; struct ctx_type *type = NULL; const char app_ctx_prefix[] = "$app."; char *provider_name = NULL, *ctx_name = NULL; @@ -713,6 +793,18 @@ struct ctx_type *get_context_type(const char *ctx) goto found; } + /* Check if ctx is a raw perf context. */ + ret = find_ctx_type_perf_raw(ctx, type); + if (ret == 0) { + type->opt->u.perf.type = PERF_TYPE_RAW; + type->opt->symbol = strdup(ctx); + if (!type->opt->symbol) { + PERROR("Copy perf field name"); + goto not_found; + } + goto found; + } + /* * No match found against static contexts; check if it is an app * context.
Allow enabling perf PMU counters by raw ID in addition to the generic list already provided. The format for kernel tracing is "perf:cpu:raw:rNNN:<name>" and "perf:thread:raw:rNNN:<name> for user-space. The rNNN format is the same as perf-record(1) where NNN is a hexadecimal event descriptor in the form of umask+eventsel. The <name> field allows the user to give a more friendly name. Example usage on Intel i7-3520M to get the unhalted reference cycles (eventsel: 0x13c) count at privilege level 0 (umask: 0x00): lttng add-context -k -t perf:cpu:raw:r0013c:x86unhalted Result in the trace: sched_switch: { cpu_id = 3 }, { perf_cpu_raw_r0013c_x86unhalted = 27632578 }, [...] Signed-off-by: Julien Desfossez <jdesfossez at efficios.com> --- doc/man/lttng-add-context.1.txt | 8 +++ src/bin/lttng/commands/add_context.c | 94 +++++++++++++++++++++++++++++++++++- 2 files changed, 101 insertions(+), 1 deletion(-)