Re: [PATCH v7 05/10] coresight: syscfg: Add API to activate and enable configurations

HI Suzuki,

On Fri, 25 Jun 2021 at 16:09, Suzuki K Poulose suzuki.poulose@arm.com wrote:

On 25/06/2021 15:37, Mike Leach wrote:

...

HI Branislav, Mathieu,

I have started to look into this as a top priority.

However, when I enable the CONFIG_DEBUG_ option my board fails to boot at all. (Dragonboard DB410c), so the resolution make take a little time.

At this point I am assuming that the mutex will need to be replaced by a spinlock - so a re-design + v8 set may well be required.

Or may be a refcount on the "config" instance may work too ? i.e, when the config is loaded onto a device, a refcount could be taken on the config and dropped on unload.

Whoever drops it to zero could release it.

Config activate / deactivate is protected by a refcount - and when dynamic loading we prevent unload if active.

The issue here is dertermining which config instance on the device to program up when enabling the device. perf will activate the config when it goes through setup_aux (at the same time as all the coresight paths / sinks are activated). the coresight device (ETM /ETE) will enable the config on a specific ETM instance when that ETM is enabled for the perf event. To do this it checks the event data to determine which config is active for the event and then calls this API to enable it. conversely when the perf event causes a disable ETM - then the config will be disabled - which may mean saving some registers. perf will deactivate the config in the free_event_data, once the perf session disappears.

So, as the event is scheduled on and off etms, the config is programmed or deprogrammed on the etm - so it is not really a case of ref counting, as it is either on or off. This allows for the event moving when --per-thread - only devices in use by perf are programmed with the config - and also ensures that this scales to support multiple perf sessions with differing configs (when we support multiple perf sessions).

The requirement here is to protect the list of loaded configs in the instance while we determine which one matches the active config for the perf session.

Moving from a cscfg global mutex, to a per csdev instance list splnlock makes sense in this context.

Regards

Mike

Thoughts ?

Suzuki

...

Regards

Mike

On Mon, 14 Jun 2021 at 18:03, Mathieu Poirier mathieu.poirier@linaro.org wrote:

...

Hi Branislav,

On Thu, 10 Jun 2021 at 05:05, Branislav Rankov branislav.rankov@arm.com wrote:

...

On 4/14/21 8:18 PM, Mike Leach wrote:

...

Configurations are first activated, then when any coresight device is enabled, the active configurations are checked and any matching one is enabled.

This patch provides the activation / enable API.

Signed-off-by: Mike Leach mike.leach@linaro.org Reviewed-by: Mathieu Poirier mathieu.poirier@linaro.org

---

.../hwtracing/coresight/coresight-config.h | 2 + .../hwtracing/coresight/coresight-syscfg.c | 172 ++++++++++++++++++ .../hwtracing/coresight/coresight-syscfg.h | 8 + include/linux/coresight.h | 2 + 4 files changed, 184 insertions(+)

diff --git a/drivers/hwtracing/coresight/coresight-config.h b/drivers/hwtracing/coresight/coresight-config.h index 0667581822c1..25eb6c632692 100644 --- a/drivers/hwtracing/coresight/coresight-config.h +++ b/drivers/hwtracing/coresight/coresight-config.h @@ -127,6 +127,7 @@ struct cscfg_feature_desc { * @nr_total_params: Sum of all parameters declared by used features * @presets: Array of preset values. * @event_ea: Extended attribute for perf event value

• • @active_cnt: ref count for activate on this configuration.
    */
  struct cscfg_config_desc {

@@ -139,6 +140,7 @@ struct cscfg_config_desc { int nr_total_params; const u64 *presets; /* nr_presets * nr_total_params */ struct dev_ext_attribute *event_ea;

• atomic_t active_cnt;
  

  };

  /**

diff --git a/drivers/hwtracing/coresight/coresight-syscfg.c b/drivers/hwtracing/coresight/coresight-syscfg.c index e35f8c0ac2f8..b234e45c153f 100644 --- a/drivers/hwtracing/coresight/coresight-syscfg.c +++ b/drivers/hwtracing/coresight/coresight-syscfg.c @@ -283,6 +283,7 @@ static int cscfg_load_config(struct cscfg_config_desc *config_desc) return err;

   list_add(&config_desc->item, &cscfg_mgr->config_desc_list);

• atomic_set(&config_desc->active_cnt, 0);
   return 0;
  

  }

@@ -468,6 +469,176 @@ void cscfg_unregister_csdev(struct coresight_device *csdev) } EXPORT_SYMBOL_GPL(cscfg_unregister_csdev);

+/**

• • cscfg_csdev_reset_feats - reset features for a CoreSight device.
• • Resets all parameters and register values for any features loaded
• • into @csdev to their default values.
• • @csdev: The CoreSight device.
• */

+void cscfg_csdev_reset_feats(struct coresight_device *csdev) +{

• struct cscfg_feature_csdev *feat_csdev;
  

• mutex_lock(&cscfg_csdev_mutex);
  

• if (list_empty(&csdev->feature_csdev_list))
  

•         goto unlock_exit;
  

• list_for_each_entry(feat_csdev, &csdev->feature_csdev_list, node)
  

•         cscfg_reset_feat(feat_csdev);
  

+unlock_exit:

• mutex_unlock(&cscfg_csdev_mutex);
  

+} +EXPORT_SYMBOL_GPL(cscfg_csdev_reset_feats);

+/**

• • cscfg_activate_config - Mark a configuration descriptor as active.
• • This will be seen when csdev devices are enabled in the system.
• • Only activated configurations can be enabled on individual devices.
• • Activation protects the configuration from alteration or removal while
• • active.
• • Selection by hash value - generated from the configuration name when it
• • was loaded and added to the cs_etm/configurations file system for selection
• • by perf.
• • Increments the configuration descriptor active count and the global active
• • count.
• • @cfg_hash: Hash value of the selected configuration name.
• */

+int cscfg_activate_config(unsigned long cfg_hash) +{

• struct cscfg_config_desc *config_desc;
  

• int err = -EINVAL;
  

• mutex_lock(&cscfg_mutex);
  

• list_for_each_entry(config_desc, &cscfg_mgr->config_desc_list, item) {
  

•         if ((unsigned long)config_desc->event_ea->var == cfg_hash) {
  

•                 /*
  

•                  * increment the global active count - control changes to
  

•                  * active configurations
  

•                  */
  

•                 atomic_inc(&cscfg_mgr->sys_active_cnt);
  

•                 /*
  

•                  * mark the descriptor as active so enable config on a
  

•                  * device instance will use it
  

•                  */
  

•                 atomic_inc(&config_desc->active_cnt);
  

•                 err = 0;
  

•                 dev_dbg(cscfg_device(), "Activate config %s.\n", config_desc->name);
  

•                 break;
  

•         }
  

• }
  

• mutex_unlock(&cscfg_mutex);
  

• return err;
  

+} +EXPORT_SYMBOL_GPL(cscfg_activate_config);

+/**

• • cscfg_deactivate_config - Mark a config descriptor as inactive.
• • Decrement the configuration and global active counts.
• • @cfg_hash: Hash value of the selected configuration name.
• */

+void cscfg_deactivate_config(unsigned long cfg_hash) +{

• struct cscfg_config_desc *config_desc;
  

• mutex_lock(&cscfg_mutex);
  

• list_for_each_entry(config_desc, &cscfg_mgr->config_desc_list, item) {
  

•         if ((unsigned long)config_desc->event_ea->var == cfg_hash) {
  

•                 atomic_dec(&config_desc->active_cnt);
  

•                 atomic_dec(&cscfg_mgr->sys_active_cnt);
  

•                 dev_dbg(cscfg_device(), "Deactivate config %s.\n", config_desc->name);
  

•                 break;
  

•         }
  

• }
  

• mutex_unlock(&cscfg_mutex);
  

+} +EXPORT_SYMBOL_GPL(cscfg_deactivate_config);

+/**

• • cscfg_csdev_enable_active_config - Enable matching active configuration for device.
• • Enables the configuration selected by @cfg_hash if the configuration is supported
• • on the device and has been activated.
• • If active and supported the CoreSight device @csdev will be programmed with the
• • configuration, using @preset parameters.
• • Should be called before driver hardware enable for the requested device, prior to
• • programming and enabling the physical hardware.
• • @csdev: CoreSight device to program.
• • @cfg_hash: Selector for the configuration.
• • @preset: Preset parameter values to use, 0 for current / default values.
• */

+int cscfg_csdev_enable_active_config(struct coresight_device *csdev,

•                              unsigned long cfg_hash, int preset)
  

+{

• struct cscfg_config_csdev *config_csdev_active = NULL, *config_csdev_item;
  

• const struct cscfg_config_desc *config_desc;
  

• int err = 0;
  

• /* quickly check global count */
  

• if (!atomic_read(&cscfg_mgr->sys_active_cnt))
  

•         return 0;
  

• mutex_lock(&cscfg_csdev_mutex);
  

• list_for_each_entry(config_csdev_item, &csdev->config_csdev_list, node) {
  

•         config_desc = config_csdev_item->config_desc;
  

•         if ((atomic_read(&config_desc->active_cnt)) &&
  

•             ((unsigned long)config_desc->event_ea->var == cfg_hash)) {
  

•                 config_csdev_active = config_csdev_item;
  

•                 break;
  

•         }
  

• }
  

• if (config_csdev_active) {
  

•         err = cscfg_csdev_enable_config(config_csdev_active, preset);
  

•         if (!err)
  

•                 csdev->active_cscfg_ctxt = (void *)config_csdev_active;
  

• }
  

• mutex_unlock(&cscfg_csdev_mutex);
  

• return err;
  

+} +EXPORT_SYMBOL_GPL(cscfg_csdev_enable_active_config);

+/**

• • cscfg_csdev_disable_active_config - disable an active config on the device.
• • Disables the active configuration on the CoreSight device @csdev.
• • Disable will save the values of any registers marked in the configurations
• • as save on disable.
• • Should be called after driver hardware disable for the requested device,
• • after disabling the physical hardware and reading back registers.
• • @csdev: The CoreSight device.
• */

+void cscfg_csdev_disable_active_config(struct coresight_device *csdev) +{

• struct cscfg_config_csdev *config_csdev;
  

• mutex_lock(&cscfg_csdev_mutex);
  

This line seems to cause a bug report when the kernel is compiled with CONFIG_DEBUG_ATOMIC_SLEEP=y

I have tested this by applying the series to 5.10 kernel on Dragonboard 845c.

It only happens when strobing is used.

Thanks for reporting this. Unfortunately Mike is away and likely won't be able to look at this in time for the merge window. As such I backed out the complex configuration feature from the next branch but that code is still available on next-complex-configuration.

Thanks, Mathieu

...

The report is this:

[13431.885395] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:283 [13431.893919] in_atomic(): 1, irqs_disabled(): 128, non_block: 0, pid: 3525, name: perf [13431.901882] CPU: 0 PID: 3525 Comm: perf Tainted: G W 5.10.0-rc5-17282-g923b456a55fb #124 [13431.911436] Hardware name: Thundercomm Dragonboard 845c (DT) [13431.917450] Call trace: [13431.919938] dump_backtrace+0x0/0x1a0 [13431.923644] show_stack+0x18/0x68 [13431.927008] dump_stack+0xd0/0x12c [13431.930444] ___might_sleep+0xf0/0x130 [13431.934228] __might_sleep+0x54/0x90 [13431.937945] mutex_lock+0x28/0x80 [13431.941468] cscfg_csdev_disable_active_config+0x24/0x50 [13431.946952] etm4_disable+0xf0/0x100 [13431.950571] etm_event_stop+0xb8/0x130 [13431.954380] etm_event_del+0x14/0x20 [13431.958008] event_sched_out.isra.0+0x84/0x1c8 [13431.962496] group_sched_out.part.0+0x44/0xc8 [13431.967010] __perf_event_disable+0xe4/0x198 [13431.971549] event_function+0x8c/0xe8 [13431.975256] remote_function+0x64/0x78 [13431.979161] generic_exec_single+0xa0/0x100 [13431.983876] smp_call_function_single+0x158/0x1d8 [13431.988629] event_function_call+0x128/0x138 [13431.992942] _perf_event_disable+0x48/0x70 [13431.997082] perf_event_for_each_child+0x3c/0x90 [13432.001751] _perf_ioctl+0x198/0x4a8 [13432.005359] perf_ioctl+0x4c/0x78 [13432.008715] __arm64_sys_ioctl+0xa8/0xf0 [13432.012835] el0_svc_common.constprop.0+0x78/0x1a0 [13432.017767] do_el0_svc+0x24/0x90 [13432.021616] el0_sync_handler+0x160/0x168 [13432.025670] el0_sync+0x174/0x180

...

• config_csdev = (struct cscfg_config_csdev *)csdev->active_cscfg_ctxt;
  

• if (config_csdev) {
  

•         cscfg_csdev_disable_config(config_csdev);
  

•         csdev->active_cscfg_ctxt = NULL;
  

• }
  

• mutex_unlock(&cscfg_csdev_mutex);
  

+} +EXPORT_SYMBOL_GPL(cscfg_csdev_disable_active_config);

• /* Initialise system configuration management device. */

  struct device *cscfg_device(void)

@@ -536,6 +707,7 @@ int __init cscfg_init(void) INIT_LIST_HEAD(&cscfg_mgr->csdev_desc_list); INIT_LIST_HEAD(&cscfg_mgr->feat_desc_list); INIT_LIST_HEAD(&cscfg_mgr->config_desc_list);

• atomic_set(&cscfg_mgr->sys_active_cnt, 0);
  
   dev_info(cscfg_device(), "CoreSight Configuration manager initialised");
   return 0;
  

diff --git a/drivers/hwtracing/coresight/coresight-syscfg.h b/drivers/hwtracing/coresight/coresight-syscfg.h index 5bcae3b374c6..a52775890670 100644 --- a/drivers/hwtracing/coresight/coresight-syscfg.h +++ b/drivers/hwtracing/coresight/coresight-syscfg.h @@ -24,12 +24,14 @@ * @csdev_desc_list: List of coresight devices registered with the configuration manager. * @feat_desc_list: List of feature descriptors to load into registered devices. * @config_desc_list: List of system configuration descriptors to load into registered devices.

• • @sys_active_cnt: Total number of active config descriptor references. */

  struct cscfg_manager { struct device dev; struct list_head csdev_desc_list; struct list_head feat_desc_list; struct list_head config_desc_list;

• atomic_t sys_active_cnt;
  

  };

  /* get reference to dev in cscfg_manager */

@@ -61,5 +63,11 @@ int cscfg_load_config_sets(struct cscfg_config_desc **cfg_descs, int cscfg_register_csdev(struct coresight_device *csdev, u32 match_flags, struct cscfg_csdev_feat_ops *ops); void cscfg_unregister_csdev(struct coresight_device *csdev); +int cscfg_activate_config(unsigned long cfg_hash); +void cscfg_deactivate_config(unsigned long cfg_hash); +void cscfg_csdev_reset_feats(struct coresight_device *csdev); +int cscfg_csdev_enable_active_config(struct coresight_device *csdev,

•                              unsigned long cfg_hash, int preset);
  

+void cscfg_csdev_disable_active_config(struct coresight_device *csdev);

#endif /* CORESIGHT_SYSCFG_H */ diff --git a/include/linux/coresight.h b/include/linux/coresight.h index 6fb516e1b22e..a348049ee08b 100644 --- a/include/linux/coresight.h +++ b/include/linux/coresight.h @@ -222,6 +222,7 @@ struct coresight_sysfs_link { * @has_conns_grp: Have added a "connections" group for sysfs links. * @feature_csdev_list: List of complex feature programming added to the device. * @config_csdev_list: List of system configurations added to the device.

• • @active_cscfg_ctxt: Context information for current active system configuration. */
  struct coresight_device { struct coresight_platform_data *pdata;

@@ -246,6 +247,7 @@ struct coresight_device { /* system configuration and feature lists */ struct list_head feature_csdev_list; struct list_head config_csdev_list;

• void *active_cscfg_ctxt;
  

  };

  /*

-- Mike Leach Principal Engineer, ARM Ltd. Manchester Design Centre. UK