| Message ID | 1445192687-24112-23-git-send-email-mathieu.poirier@linaro.org |
|---|---|
| State | New |
| Headers | show |
Mathieu Poirier <mathieu.poirier@linaro.org> writes: > +static int etm_event_pmu_start(struct perf_event *event) > +{ > + int cpu, ret; > + cpumask_t mask; > + struct coresight_device *csdev; > + > + cpumask_clear(&mask); > + if (event->cpu != -1) > + cpumask_set_cpu(event->cpu, &mask); > + else > + cpumask_copy(&mask, cpu_online_mask); > + > + for_each_cpu(cpu, &mask) { > + csdev = per_cpu(csdev_src, cpu); > + > + if (!source_ops(csdev)->perf_start) > + continue; > + > + ret = source_ops(csdev)->perf_start(csdev); > + if (ret) > + goto err; So long as "perf_start" and "perf_stop" here mean "pm_runtime_get()/put()", this can work, but in that case maybe a better name should be used, because no real starting or stopping of anything takes place here. Since pmu::event_init and event::destroy happen in allocation/deallocation paths and at event scheduling, it's not a good idea to actually start anything here. Regards, -- Alex -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Mathieu Poirier <mathieu.poirier@linaro.org> writes: > +static void *etm_setup_aux(struct perf_event *event, void **pages, > + int nr_pages, bool overwrite) > +{ > + int cpu; > + cpumask_t *mask; > + struct etm_event_data *event_data = NULL; > + struct coresight_device *csdev; > + > + event_data = alloc_event_data(event->cpu); > + if (!event_data) > + return NULL; > + > + mask = &event_data->mask; > + > + if (event->cpu != -1) > + cpumask_set_cpu(event->cpu, mask); > + else > + cpumask_copy(mask, cpu_online_mask); > + > + for_each_cpu(cpu, mask) { > + struct coresight_device *sink; > + > + csdev = per_cpu(csdev_src, cpu); > + if (!csdev) > + goto err; > + > + /* Get the tracer's config from perf */ > + if (!source_ops(csdev)->perf_get_config) > + goto err; > + > + event_data->source_config[cpu] = > + source_ops(csdev)->perf_get_config(csdev, event); > + > + if (!event_data->source_config[cpu]) > + goto err; > + > + /* > + * Get a handle on the sink buffer associated > + * with this tracer. > + */ > + event_data->sink[cpu] = (void *)etm_event_build_path(cpu, true); There are several problems here. What is created/allocated during setup_aux(), has to be undone in free_aux(), however, the effect of build_path() will only be undone in the event::destroy() path. So if the user unmaps the aux buffer and then maps it again, we'll go ahead and try to build the path again. (Btw, coresight_build_paths() and other non-static functions and especially exported ones are really lacking documentation at the moment). It really looks like this has to be done in pmu::add(), so that the source<=>sink connection exists only while the event is scheduled and otherwise other events are free to connect their sources to these sinks. And at pmu::del() the connection has to be torn down. This way we can have a sensible multisession support. That is, provided my understanding of the coresight driver architecture is correct. Also, you won't have to configure things on multiple cpus for cpu==-1 if you keep the source<=>sink connection only between pmu::add() and pmu::del(), as an event can only be scheduled on one cpu at a time, which should make things simpler. > + > + if (!event_data->sink[cpu]) > + goto err; > + > + sink = event_data->sink[cpu]; > + > + if (!sink_ops(sink)->setup_aux) > + goto err; > + > + /* Finally get the AUX specific data from the sink buffer */ > + event_data->sink_config[cpu] = > + sink_ops(sink)->setup_aux(sink, cpu, pages, > + nr_pages, overwrite); Now this is a sensible thing to do. I understand that you'll have to know which sink you're using so that you can pick the right sink_ops and build an appropriate configuration, but perhaps it also makes sense to release it once you got the sink_config. > +static void etm_event_stop(struct perf_event *event, int mode) > +{ > + int cpu = smp_processor_id(); > + struct coresight_device *csdev = per_cpu(csdev_src, cpu); > + > + if (event->hw.state == PERF_HES_STOPPED) > + return; > + > + if (!csdev) > + return; > + > + /* stop tracer */ > + if (!source_ops(csdev)->perf_disable) > + return; This really shouldn't happen. It makes sense to make sure that we have all the callbacks that we rely on in pmu::event_init() or pmu::add() and refuse to start if we don't, but at this point we really shouldn't end up in a situation where we suddenly don't have one of the callbacks. > + if (source_ops(csdev)->perf_disable(csdev)) > + return; This has a similar problem. I'd say that this callback should not be able to fail and return anything other than success. > + /* tell the core */ > + event->hw.state = PERF_HES_STOPPED; > + > + > + if (mode & PERF_EF_UPDATE) { > + struct coresight_device *sink; > + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); > + struct etm_event_data *event_data = perf_get_aux(handle); > + > + if (WARN_ON_ONCE(handle->event != event)) > + return; > + > + if (WARN_ON_ONCE(!event_data)) > + return; > + > + sink = event_data->sink[cpu]; > + if (WARN_ON_ONCE(!sink)) > + return; > + > + /* update trace information */ > + if (!sink_ops(sink)->update_buffer) > + return; > + > + sink_ops(sink)->update_buffer(sink, handle, > + event_data->sink_config[cpu]); > + } > +} > + > +static void etm_event_start(struct perf_event *event, int flags) > +{ > + int cpu = smp_processor_id(); > + struct coresight_device *csdev = per_cpu(csdev_src, cpu); > + > + if (!csdev) > + goto fail; > + > + /* tell the perf core the event is alive */ > + event->hw.state = 0; > + > + if (!source_ops(csdev)->perf_enable) > + goto fail; Same here. > + > + if (source_ops(csdev)->perf_enable(csdev)) > + goto fail; This may fail, I suppose. > + > + return; > + > +fail: > + event->hw.state = PERF_HES_STOPPED; > +} > + > +static void etm_event_del(struct perf_event *event, int mode) > +{ > + int cpu = smp_processor_id(); > + struct coresight_device *sink; > + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); > + struct etm_event_data *event_data = perf_get_aux(handle); > + > + if (WARN_ON_ONCE(!event_data)) > + return; > + > + sink = event_data->sink[cpu]; > + if (!sink) > + return; This also shouldn't be able to prevent us from stopping the event. > + > + etm_event_stop(event, PERF_EF_UPDATE); > + > + if (!sink_ops(sink)->reset_buffer) > + return; > + > + sink_ops(sink)->reset_buffer(sink, handle, > + event_data->sink_config[cpu]); > +} > + > +static int etm_event_add(struct perf_event *event, int mode) > +{ > + > + int ret = -EBUSY, cpu = smp_processor_id(); > + struct etm_event_data *event_data; > + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); > + struct hw_perf_event *hwc = &event->hw; > + struct coresight_device *csdev = per_cpu(csdev_src, cpu); > + struct coresight_device *sink; > + > + if (handle->event) > + goto out; > + > + event_data = perf_aux_output_begin(handle, event); > + ret = -EINVAL; > + if (WARN_ON_ONCE(!event_data)) > + goto fail_stop; > + > + sink = event_data->sink[cpu]; So if you're able to fetch the sink right here and release it in _del(). Of course, this being a hot path and an atomic context needs to be taken into account. Regards, -- Alex -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
On 19 October 2015 at 09:37, Alexander Shishkin <alexander.shishkin@linux.intel.com> wrote: > Mathieu Poirier <mathieu.poirier@linaro.org> writes: > >> +static int etm_event_pmu_start(struct perf_event *event) >> +{ >> + int cpu, ret; >> + cpumask_t mask; >> + struct coresight_device *csdev; >> + >> + cpumask_clear(&mask); >> + if (event->cpu != -1) >> + cpumask_set_cpu(event->cpu, &mask); >> + else >> + cpumask_copy(&mask, cpu_online_mask); >> + >> + for_each_cpu(cpu, &mask) { >> + csdev = per_cpu(csdev_src, cpu); >> + >> + if (!source_ops(csdev)->perf_start) >> + continue; >> + >> + ret = source_ops(csdev)->perf_start(csdev); >> + if (ret) >> + goto err; > > So long as "perf_start" and "perf_stop" here mean > "pm_runtime_get()/put()", this can work, but in that case maybe a better > name should be used, because no real starting or stopping of anything > takes place here. You're correct, nothing else than pm_runtime operations should be happening in there. I will revise the naming convention. > Since pmu::event_init and event::destroy happen in > allocation/deallocation paths and at event scheduling, it's not a good > idea to actually start anything here. > > Regards, > -- > Alex -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
On 20 October 2015 at 03:34, Alexander Shishkin <alexander.shishkin@linux.intel.com> wrote: > Mathieu Poirier <mathieu.poirier@linaro.org> writes: > >> +static void *etm_setup_aux(struct perf_event *event, void **pages, >> + int nr_pages, bool overwrite) >> +{ >> + int cpu; >> + cpumask_t *mask; >> + struct etm_event_data *event_data = NULL; >> + struct coresight_device *csdev; >> + >> + event_data = alloc_event_data(event->cpu); >> + if (!event_data) >> + return NULL; >> + >> + mask = &event_data->mask; >> + >> + if (event->cpu != -1) >> + cpumask_set_cpu(event->cpu, mask); >> + else >> + cpumask_copy(mask, cpu_online_mask); >> + >> + for_each_cpu(cpu, mask) { >> + struct coresight_device *sink; >> + >> + csdev = per_cpu(csdev_src, cpu); >> + if (!csdev) >> + goto err; >> + >> + /* Get the tracer's config from perf */ >> + if (!source_ops(csdev)->perf_get_config) >> + goto err; >> + >> + event_data->source_config[cpu] = >> + source_ops(csdev)->perf_get_config(csdev, event); >> + >> + if (!event_data->source_config[cpu]) >> + goto err; >> + >> + /* >> + * Get a handle on the sink buffer associated >> + * with this tracer. >> + */ >> + event_data->sink[cpu] = (void *)etm_event_build_path(cpu, true); > > There are several problems here. What is created/allocated during > setup_aux(), has to be undone in free_aux(), however, the effect of > build_path() will only be undone in the event::destroy() path. So if the > user unmaps the aux buffer and then maps it again, we'll go ahead and > try to build the path again. (Btw, coresight_build_paths() and other > non-static functions and especially exported ones are really lacking > documentation at the moment). > > It really looks like this has to be done in pmu::add(), so that the > source<=>sink connection exists only while the event is scheduled and > otherwise other events are free to connect their sources to these > sinks. And at pmu::del() the connection has to be torn down. This way we > can have a sensible multisession support. That is, provided my > understanding of the coresight driver architecture is correct. > > Also, you won't have to configure things on multiple cpus for cpu==-1 if > you keep the source<=>sink connection only between pmu::add() and > pmu::del(), as an event can only be scheduled on one cpu at a time, > which should make things simpler. I am well aware of all this... Currently the process of building a path is too heavy to be done at context switch time. To be efficient the components of a path would have to be kept in a linked list that is then enabled/disabled when the time comes. I've been meaning to do something better for a while now. This might be the perfect time to address the problem. Thanks for reviewing the patch set, Mathieu > >> + >> + if (!event_data->sink[cpu]) >> + goto err; >> + >> + sink = event_data->sink[cpu]; >> + >> + if (!sink_ops(sink)->setup_aux) >> + goto err; >> + >> + /* Finally get the AUX specific data from the sink buffer */ >> + event_data->sink_config[cpu] = >> + sink_ops(sink)->setup_aux(sink, cpu, pages, >> + nr_pages, overwrite); > > Now this is a sensible thing to do. I understand that you'll have to > know which sink you're using so that you can pick the right sink_ops and > build an appropriate configuration, but perhaps it also makes sense to > release it once you got the sink_config. > >> +static void etm_event_stop(struct perf_event *event, int mode) >> +{ >> + int cpu = smp_processor_id(); >> + struct coresight_device *csdev = per_cpu(csdev_src, cpu); >> + >> + if (event->hw.state == PERF_HES_STOPPED) >> + return; >> + >> + if (!csdev) >> + return; >> + >> + /* stop tracer */ >> + if (!source_ops(csdev)->perf_disable) >> + return; > > This really shouldn't happen. It makes sense to make sure that we have > all the callbacks that we rely on in pmu::event_init() or pmu::add() and > refuse to start if we don't, but at this point we really shouldn't end > up in a situation where we suddenly don't have one of the callbacks. > >> + if (source_ops(csdev)->perf_disable(csdev)) >> + return; > > This has a similar problem. I'd say that this callback should not be > able to fail and return anything other than success. > >> + /* tell the core */ >> + event->hw.state = PERF_HES_STOPPED; >> + >> + >> + if (mode & PERF_EF_UPDATE) { >> + struct coresight_device *sink; >> + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); >> + struct etm_event_data *event_data = perf_get_aux(handle); >> + >> + if (WARN_ON_ONCE(handle->event != event)) >> + return; >> + >> + if (WARN_ON_ONCE(!event_data)) >> + return; >> + >> + sink = event_data->sink[cpu]; >> + if (WARN_ON_ONCE(!sink)) >> + return; >> + >> + /* update trace information */ >> + if (!sink_ops(sink)->update_buffer) >> + return; >> + >> + sink_ops(sink)->update_buffer(sink, handle, >> + event_data->sink_config[cpu]); >> + } >> +} >> + >> +static void etm_event_start(struct perf_event *event, int flags) >> +{ >> + int cpu = smp_processor_id(); >> + struct coresight_device *csdev = per_cpu(csdev_src, cpu); >> + >> + if (!csdev) >> + goto fail; >> + >> + /* tell the perf core the event is alive */ >> + event->hw.state = 0; >> + >> + if (!source_ops(csdev)->perf_enable) >> + goto fail; > > Same here. > >> + >> + if (source_ops(csdev)->perf_enable(csdev)) >> + goto fail; > > This may fail, I suppose. > >> + >> + return; >> + >> +fail: >> + event->hw.state = PERF_HES_STOPPED; >> +} >> + >> +static void etm_event_del(struct perf_event *event, int mode) >> +{ >> + int cpu = smp_processor_id(); >> + struct coresight_device *sink; >> + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); >> + struct etm_event_data *event_data = perf_get_aux(handle); >> + >> + if (WARN_ON_ONCE(!event_data)) >> + return; >> + >> + sink = event_data->sink[cpu]; >> + if (!sink) >> + return; > > This also shouldn't be able to prevent us from stopping the event. > >> + >> + etm_event_stop(event, PERF_EF_UPDATE); >> + >> + if (!sink_ops(sink)->reset_buffer) >> + return; >> + >> + sink_ops(sink)->reset_buffer(sink, handle, >> + event_data->sink_config[cpu]); >> +} >> + >> +static int etm_event_add(struct perf_event *event, int mode) >> +{ >> + >> + int ret = -EBUSY, cpu = smp_processor_id(); >> + struct etm_event_data *event_data; >> + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); >> + struct hw_perf_event *hwc = &event->hw; >> + struct coresight_device *csdev = per_cpu(csdev_src, cpu); >> + struct coresight_device *sink; >> + >> + if (handle->event) >> + goto out; >> + >> + event_data = perf_aux_output_begin(handle, event); >> + ret = -EINVAL; >> + if (WARN_ON_ONCE(!event_data)) >> + goto fail_stop; >> + >> + sink = event_data->sink[cpu]; > > So if you're able to fetch the sink right here and release it in > _del(). Of course, this being a hot path and an atomic context needs to > be taken into account. > > Regards, > -- > Alex -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
diff --git a/drivers/hwtracing/coresight/Makefile b/drivers/hwtracing/coresight/Makefile index 233d66cf22d3..cf8c6d689747 100644 --- a/drivers/hwtracing/coresight/Makefile +++ b/drivers/hwtracing/coresight/Makefile @@ -9,6 +9,7 @@ obj-$(CONFIG_CORESIGHT_SINK_ETBV10) += coresight-etb10.o obj-$(CONFIG_CORESIGHT_LINKS_AND_SINKS) += coresight-funnel.o \ coresight-replicator.o obj-$(CONFIG_CORESIGHT_SOURCE_ETM3X) += coresight-etm3x.o coresight-etm-cp14.o \ - coresight-etm3x-sysfs.o + coresight-etm3x-sysfs.o \ + coresight-etm-perf.o obj-$(CONFIG_CORESIGHT_SOURCE_ETM4X) += coresight-etm4x.o obj-$(CONFIG_CORESIGHT_QCOM_REPLICATOR) += coresight-replicator-qcom.o diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c new file mode 100644 index 000000000000..dbd02277fcda --- /dev/null +++ b/drivers/hwtracing/coresight/coresight-etm-perf.c @@ -0,0 +1,533 @@ +/* + * Copyright(C) 2015 Linaro Limited. All rights reserved. + * Author: Mathieu Poirier <mathieu.poirier@linaro.org> + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 as published by + * the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for + * more details. + * + * You should have received a copy of the GNU General Public License along with + * this program. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <linux/coresight.h> +#include <linux/cpumask.h> +#include <linux/device.h> +#include <linux/list.h> +#include <linux/mm.h> +#include <linux/module.h> +#include <linux/perf_event.h> +#include <linux/slab.h> +#include <linux/types.h> + +#include "coresight-priv.h" + +#define CORESIGHT_ETM_PMU_NAME "cs_etm" + +static struct pmu etm_pmu; +static bool etm_perf_up; + +/** + * struct etm_event_data - Coresight specifics associated to an event + * @mask: hold the CPU(s) this event was set for. + * @source_config: per CPU tracer configuration associated to a + * trace session. + * @sink_config: per CPU AUX configuration associated to a + * trace session. + * @sink: sink associated to a CPU. + */ +struct etm_event_data { + cpumask_t mask; + void **source_config; + void **sink_config; + void **sink; +}; + +static DEFINE_PER_CPU(struct perf_output_handle, ctx_handle); +static DEFINE_PER_CPU(struct coresight_device *, csdev_src); + +/* ETMCR is 'config' */ +PMU_FORMAT_ATTR(cycacc, "config:12"); +PMU_FORMAT_ATTR(timestamp, "config:28"); + +static struct attribute *etm_config_formats_attr[] = { + &format_attr_cycacc.attr, + &format_attr_timestamp.attr, + NULL, +}; + +static struct attribute_group etm_pmu_format_group = { + .name = "format", + .attrs = etm_config_formats_attr, +}; + +static const struct attribute_group *etm_pmu_attr_groups[] = { + &etm_pmu_format_group, + NULL, +}; + +static void etm_event_read(struct perf_event *event) {} + +/** + * etm_event_build_path() - setup a path between source and sink + * @cpu: The CPU the tracer is associated to. + * @build: Whether the path should be setup or thorned down. + * + * Return: The _first_ sink buffer discovered during the walkthrough. + */ +static struct coresight_device *etm_event_build_path(int cpu, bool build) +{ + int ret = 0; + LIST_HEAD(path); + LIST_HEAD(sinks); + struct coresight_device *csdev_source; + struct coresight_device *csdev_sink = NULL; + + csdev_source = per_cpu(csdev_src, cpu); + + if (!csdev_source) + return ERR_PTR(-EINVAL); + + if (csdev_source->type != CORESIGHT_DEV_TYPE_SOURCE) + return ERR_PTR(-EINVAL); + + if (build) { + ret = coresight_build_paths(csdev_source, &path, &sinks, build); + if (ret) { + dev_dbg(&csdev_source->dev, + "creating path(s) failed\n"); + goto out; + } + + /* Everything is good, record first enabled sink buffer */ + csdev_sink = list_first_entry(&sinks, + struct coresight_device, sinks); + } else { + ret = coresight_build_paths(csdev_source, &path, NULL, build); + if (ret) + dev_dbg(&csdev_source->dev, + "releasing path(s) failed\n"); + } + +out: + return csdev_sink; +} + +static int etm_event_pmu_start(struct perf_event *event) +{ + int cpu, ret; + cpumask_t mask; + struct coresight_device *csdev; + + cpumask_clear(&mask); + if (event->cpu != -1) + cpumask_set_cpu(event->cpu, &mask); + else + cpumask_copy(&mask, cpu_online_mask); + + for_each_cpu(cpu, &mask) { + csdev = per_cpu(csdev_src, cpu); + + if (!source_ops(csdev)->perf_start) + continue; + + ret = source_ops(csdev)->perf_start(csdev); + if (ret) + goto err; + } + +out: + return ret; +err: + for_each_cpu(cpu, &mask) { + csdev = per_cpu(csdev_src, cpu); + + if (!source_ops(csdev)->perf_stop) + continue; + source_ops(csdev)->perf_stop(csdev); + } + + goto out; +} + +static void etm_event_destroy(struct perf_event *event) +{ + int cpu; + cpumask_t mask; + struct coresight_device *csdev; + + cpumask_clear(&mask); + if (event->cpu != -1) + cpumask_set_cpu(event->cpu, &mask); + else + cpumask_copy(&mask, cpu_online_mask); + + for_each_cpu(cpu, &mask) { + csdev = per_cpu(csdev_src, cpu); + etm_event_build_path(cpu, false); + + if (!source_ops(csdev)->perf_stop) + continue; + source_ops(csdev)->perf_stop(csdev); + } +} + +static int etm_event_init(struct perf_event *event) +{ + int ret; + + if (event->attr.type != etm_pmu.type) + return -ENOENT; + + if (event->cpu >= nr_cpu_ids) + return -EINVAL; + + ret = etm_event_pmu_start(event); + if (ret) + return ret; + + event->destroy = etm_event_destroy; + + return 0; +} + +static void *alloc_event_data(int cpu) +{ + int size; + struct etm_event_data *event_data; + void *source_config, *sink_config, *sink; + + event_data = kzalloc(sizeof(struct etm_event_data), GFP_KERNEL); + if (!event_data) + return NULL; + + if (cpu != -1) + size = 1; + else + size = num_online_cpus(); + + source_config = kcalloc(size, sizeof(void *), GFP_KERNEL); + if (!source_config) + goto source_config_err; + + sink_config = kcalloc(size, sizeof(void *), GFP_KERNEL); + if (!sink_config) + goto sink_config_err; + + sink = kcalloc(size, sizeof(void *), GFP_KERNEL); + if (!sink) + goto sink_err; + + cpumask_clear(&event_data->mask); + event_data->source_config = source_config; + event_data->sink_config = sink_config; + event_data->sink = sink; + +out: + return event_data; + +sink_err: + kfree(sink_config); +sink_config_err: + kfree(source_config); +source_config_err: + kfree(event_data); + event_data = NULL; + goto out; +} + +static void free_event_data(struct etm_event_data *event_data) +{ + int cpu; + cpumask_t *mask = &event_data->mask; + + for_each_cpu(cpu, mask) { + kfree(event_data->source_config[cpu]); + kfree(event_data->sink_config[cpu]); + kfree(event_data->sink[cpu]); + } + + kfree(event_data->source_config); + kfree(event_data->sink_config); + kfree(event_data->sink); + kfree(event_data); +} + +static void *etm_setup_aux(struct perf_event *event, void **pages, + int nr_pages, bool overwrite) +{ + int cpu; + cpumask_t *mask; + struct etm_event_data *event_data = NULL; + struct coresight_device *csdev; + + event_data = alloc_event_data(event->cpu); + if (!event_data) + return NULL; + + mask = &event_data->mask; + + if (event->cpu != -1) + cpumask_set_cpu(event->cpu, mask); + else + cpumask_copy(mask, cpu_online_mask); + + for_each_cpu(cpu, mask) { + struct coresight_device *sink; + + csdev = per_cpu(csdev_src, cpu); + if (!csdev) + goto err; + + /* Get the tracer's config from perf */ + if (!source_ops(csdev)->perf_get_config) + goto err; + + event_data->source_config[cpu] = + source_ops(csdev)->perf_get_config(csdev, event); + + if (!event_data->source_config[cpu]) + goto err; + + /* + * Get a handle on the sink buffer associated + * with this tracer. + */ + event_data->sink[cpu] = (void *)etm_event_build_path(cpu, true); + + if (!event_data->sink[cpu]) + goto err; + + sink = event_data->sink[cpu]; + + if (!sink_ops(sink)->setup_aux) + goto err; + + /* Finally get the AUX specific data from the sink buffer */ + event_data->sink_config[cpu] = + sink_ops(sink)->setup_aux(sink, cpu, pages, + nr_pages, overwrite); + if (!event_data->sink_config[cpu]) + goto err; + } + +out: + return event_data; + +err: + for_each_cpu(cpu, mask) { + etm_event_build_path(cpu, false); + } + + free_event_data(event_data); + event_data = NULL; + goto out; +} + +static void etm_free_aux(void *data) +{ + free_event_data(data); +} + +static void etm_event_stop(struct perf_event *event, int mode) +{ + int cpu = smp_processor_id(); + struct coresight_device *csdev = per_cpu(csdev_src, cpu); + + if (event->hw.state == PERF_HES_STOPPED) + return; + + if (!csdev) + return; + + /* stop tracer */ + if (!source_ops(csdev)->perf_disable) + return; + + if (source_ops(csdev)->perf_disable(csdev)) + return; + + /* tell the core */ + event->hw.state = PERF_HES_STOPPED; + + + if (mode & PERF_EF_UPDATE) { + struct coresight_device *sink; + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); + struct etm_event_data *event_data = perf_get_aux(handle); + + if (WARN_ON_ONCE(handle->event != event)) + return; + + if (WARN_ON_ONCE(!event_data)) + return; + + sink = event_data->sink[cpu]; + if (WARN_ON_ONCE(!sink)) + return; + + /* update trace information */ + if (!sink_ops(sink)->update_buffer) + return; + + sink_ops(sink)->update_buffer(sink, handle, + event_data->sink_config[cpu]); + } +} + +static void etm_event_start(struct perf_event *event, int flags) +{ + int cpu = smp_processor_id(); + struct coresight_device *csdev = per_cpu(csdev_src, cpu); + + if (!csdev) + goto fail; + + /* tell the perf core the event is alive */ + event->hw.state = 0; + + if (!source_ops(csdev)->perf_enable) + goto fail; + + if (source_ops(csdev)->perf_enable(csdev)) + goto fail; + + return; + +fail: + event->hw.state = PERF_HES_STOPPED; +} + +static void etm_event_del(struct perf_event *event, int mode) +{ + int cpu = smp_processor_id(); + struct coresight_device *sink; + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); + struct etm_event_data *event_data = perf_get_aux(handle); + + if (WARN_ON_ONCE(!event_data)) + return; + + sink = event_data->sink[cpu]; + if (!sink) + return; + + etm_event_stop(event, PERF_EF_UPDATE); + + if (!sink_ops(sink)->reset_buffer) + return; + + sink_ops(sink)->reset_buffer(sink, handle, + event_data->sink_config[cpu]); +} + +static int etm_event_add(struct perf_event *event, int mode) +{ + + int ret = -EBUSY, cpu = smp_processor_id(); + struct etm_event_data *event_data; + struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle); + struct hw_perf_event *hwc = &event->hw; + struct coresight_device *csdev = per_cpu(csdev_src, cpu); + struct coresight_device *sink; + + if (handle->event) + goto out; + + event_data = perf_aux_output_begin(handle, event); + ret = -EINVAL; + if (WARN_ON_ONCE(!event_data)) + goto fail_stop; + + sink = event_data->sink[cpu]; + if (!sink) + goto fail_end_stop; + + if (!sink_ops(sink)->set_buffer) + goto fail_end_stop; + + ret = sink_ops(sink)->set_buffer(sink, handle, + event_data->sink_config[cpu]); + if (ret) + goto fail_end_stop; + + if (!source_ops(csdev)->perf_set_config) { + ret = -EINVAL; + goto fail_end_stop; + } + + source_ops(csdev)->perf_set_config(csdev, + event_data->source_config[cpu]); + + if (mode & PERF_EF_START) { + etm_event_start(event, 0); + if (hwc->state & PERF_HES_STOPPED) { + etm_event_del(event, 0); + return -EBUSY; + } + } + +out: + return ret; + +fail_end_stop: + perf_aux_output_end(handle, 0, true); +fail_stop: + hwc->state = PERF_HES_STOPPED; + goto out; +} + +int etm_perf_symlink(struct coresight_device *csdev, bool link) +{ + char entry[sizeof("cpu9999999")]; + int ret = 0, cpu = source_ops(csdev)->cpu_id(csdev); + struct device *pmu_dev = etm_pmu.dev; + struct device *cs_dev = &csdev->dev; + + sprintf(entry, "cpu%d", cpu); + + if (!etm_perf_up) + return -EPROBE_DEFER; + + if (link) { + ret = sysfs_create_link(&pmu_dev->kobj, &cs_dev->kobj, entry); + if (ret) + return ret; + per_cpu(csdev_src, cpu) = csdev; + } else { + sysfs_remove_link(&pmu_dev->kobj, entry); + per_cpu(csdev_src, cpu) = NULL; + } + + return 0; +} + +static int __init etm_perf_init(void) +{ + int ret; + + etm_pmu.capabilities = PERF_PMU_CAP_EXCLUSIVE; + + etm_pmu.attr_groups = etm_pmu_attr_groups; + etm_pmu.task_ctx_nr = perf_sw_context; + etm_pmu.read = etm_event_read; + etm_pmu.event_init = etm_event_init; + etm_pmu.setup_aux = etm_setup_aux; + etm_pmu.free_aux = etm_free_aux; + etm_pmu.stop = etm_event_stop; + etm_pmu.start = etm_event_start; + etm_pmu.del = etm_event_del; + etm_pmu.add = etm_event_add; + + ret = perf_pmu_register(&etm_pmu, CORESIGHT_ETM_PMU_NAME, -1); + if (ret == 0) + etm_perf_up = true; + + return ret; +} +module_init(etm_perf_init); diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.h b/drivers/hwtracing/coresight/coresight-etm-perf.h new file mode 100644 index 000000000000..4dd900f2362a --- /dev/null +++ b/drivers/hwtracing/coresight/coresight-etm-perf.h @@ -0,0 +1,27 @@ +/* Copyright (c) 2014-2015, The Linux Foundation. All rights reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 and + * only version 2 as published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + */ + +#ifndef _CORESIGHT_ETM_PERF_H +#define _CORESIGHT_ETM_PERF_H + +struct coresight_device; + +#ifdef CONFIG_CORESIGHT +int etm_perf_symlink(struct coresight_device *csdev, bool link); + +#else +static inline int etm_perf_symlink(struct coresight_device *csdev, bool link) +{ return -EINVAL; } + +#endif /* CONFIG_CORESIGHT */ + +#endif diff --git a/drivers/hwtracing/coresight/coresight-etm3x.c b/drivers/hwtracing/coresight/coresight-etm3x.c index 9b4c0359ca29..7407c7ecf668 100644 --- a/drivers/hwtracing/coresight/coresight-etm3x.c +++ b/drivers/hwtracing/coresight/coresight-etm3x.c @@ -35,6 +35,7 @@ #include <asm/sections.h> #include "coresight-etm.h" +#include "coresight-etm-perf.h" static int boot_enable; module_param_named(boot_enable, boot_enable, int, S_IRUGO); @@ -850,6 +851,12 @@ static int etm_probe(struct amba_device *adev, const struct amba_id *id) goto err_arch_supported; } + ret = etm_perf_symlink(drvdata->csdev, true); + if (ret) { + coresight_unregister(drvdata->csdev); + goto err_arch_supported; + } + pm_runtime_put(&adev->dev); dev_info(dev, "%s initialized\n", (char *)id->data);
Perf is a well known and used tool for performance monitoring and much more. A such it is an ideal condaditate for integration with coresight based HW tracing. This patch introduces a PMU that represent a coresight tracer to the Perf core. Signed-off-by: Mathieu Poirier <mathieu.poirier@linaro.org> --- drivers/hwtracing/coresight/Makefile | 3 +- drivers/hwtracing/coresight/coresight-etm-perf.c | 533 +++++++++++++++++++++++ drivers/hwtracing/coresight/coresight-etm-perf.h | 27 ++ drivers/hwtracing/coresight/coresight-etm3x.c | 7 + 4 files changed, 569 insertions(+), 1 deletion(-) create mode 100644 drivers/hwtracing/coresight/coresight-etm-perf.c create mode 100644 drivers/hwtracing/coresight/coresight-etm-perf.h