This series fixes the coresight generic layer to handle the reference
counting for the STM source properly, to allow multiple applications
to share the STM. Without this series, the STM is disabled when the
first user closes its connection, causing trace data losses.
Suzuki K Poulose (2):
coresight: Disable the path only when the source is disabled
coresight: Fix reference count for software sources
drivers/hwtracing/coresight/coresight.c | 20 +++++++++++++++-----
1 file changed, 15 insertions(+), 5 deletions(-)
--
2.7.4
With a coresight tracing session, the components along the path
from the source to sink are disabled after the source is disabled.
However, if the source was not actually disabled due to active
users, we should not disable the components in the path.
Cc: Mathieu Poirier <[email protected]>
Signed-off-by: Suzuki K Poulose <[email protected]>
---
drivers/hwtracing/coresight/coresight.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
index 0c37356..34cd1ed 100644
--- a/drivers/hwtracing/coresight/coresight.c
+++ b/drivers/hwtracing/coresight/coresight.c
@@ -253,7 +253,8 @@ static int coresight_enable_source(struct coresight_device *csdev, u32 mode)
return 0;
}
-static void coresight_disable_source(struct coresight_device *csdev)
+/* coresight_disable_source: Returns true if the device has been disabled */
+static bool coresight_disable_source(struct coresight_device *csdev)
{
if (atomic_dec_return(csdev->refcnt) == 0) {
if (source_ops(csdev)->disable) {
@@ -261,6 +262,7 @@ static void coresight_disable_source(struct coresight_device *csdev)
csdev->enable = false;
}
}
+ return !csdev->enable;
}
void coresight_disable_path(struct list_head *path)
@@ -629,7 +631,7 @@ void coresight_disable(struct coresight_device *csdev)
if (ret)
goto out;
- if (!csdev->enable)
+ if (!csdev->enable || !coresight_disable_source(csdev))
goto out;
switch (csdev->subtype.source_subtype) {
@@ -647,7 +649,6 @@ void coresight_disable(struct coresight_device *csdev)
break;
}
- coresight_disable_source(csdev);
coresight_disable_path(path);
coresight_release_path(path);
--
2.7.4
For software sources (i.e STM), there could be multiple agents
generating the trace data, unlike the ETMs. So we need to
properly do the accounting for the active number of users
to disable the device when the last user goes away. Right
now, the reference counting is broken for sources as we skip
the actions when we detect that the source is enabled.
This patch fixes the problem by adding the refcounting for
software sources, even when they are enabled.
Cc: Mathieu Poirier <[email protected]>
Reported-by: Robert Walker <[email protected]>
Signed-off-by: Suzuki K Poulose <[email protected]>
---
drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
index 34cd1ed..2da9e39 100644
--- a/drivers/hwtracing/coresight/coresight.c
+++ b/drivers/hwtracing/coresight/coresight.c
@@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
int cpu, ret = 0;
struct coresight_device *sink;
struct list_head *path;
+ enum coresight_dev_subtype_source subtype = csdev->subtype.source_subtype;
mutex_lock(&coresight_mutex);
@@ -559,8 +560,16 @@ int coresight_enable(struct coresight_device *csdev)
if (ret)
goto out;
- if (csdev->enable)
+ if (csdev->enable) {
+ /*
+ * There could be multiple applications driving the software
+ * source. So keep the refcount for each such user when the
+ * source is already enabled.
+ */
+ if (subtype == CORESIGHT_DEV_SUBTYPE_SOURCE_SOFTWARE)
+ atomic_inc(csdev->refcnt);
goto out;
+ }
/*
* Search for a valid sink for this session but don't reset the
@@ -587,7 +596,7 @@ int coresight_enable(struct coresight_device *csdev)
if (ret)
goto err_source;
- switch (csdev->subtype.source_subtype) {
+ switch (subtype) {
case CORESIGHT_DEV_SUBTYPE_SOURCE_PROC:
/*
* When working from sysFS it is important to keep track
--
2.7.4
Hi Suzuki,
On 14 March 2017 at 11:31, Mathieu Poirier <[email protected]> wrote:
> From: "Suzuki K. Poulose" <[email protected]>
>
> With a coresight tracing session, the components along the path
> from the source to sink are disabled after the source is disabled.
> However, if the source was not actually disabled due to active
> users, we should not disable the components in the path.
>
> Cc: Mathieu Poirier <[email protected]>
> Signed-off-by: Suzuki K Poulose <[email protected]>
> ---
> drivers/hwtracing/coresight/coresight.c | 7 ++++---
> 1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
> index 0c37356..34cd1ed 100644
> --- a/drivers/hwtracing/coresight/coresight.c
> +++ b/drivers/hwtracing/coresight/coresight.c
> @@ -253,7 +253,8 @@ static int coresight_enable_source(struct coresight_device *csdev, u32 mode)
> return 0;
> }
>
> -static void coresight_disable_source(struct coresight_device *csdev)
> +/* coresight_disable_source: Returns true if the device has been disabled */
To be coherent with the rest of the file please use:
/*
* coresight_disable_source: Returns true if the device has been disabled
*/
I would have done the modification myself but 2/2 needs tending as well.
Thanks,
Mathieu
> +static bool coresight_disable_source(struct coresight_device *csdev)
> {
> if (atomic_dec_return(csdev->refcnt) == 0) {
> if (source_ops(csdev)->disable) {
> @@ -261,6 +262,7 @@ static void coresight_disable_source(struct coresight_device *csdev)
> csdev->enable = false;
> }
> }
> + return !csdev->enable;
> }
>
> void coresight_disable_path(struct list_head *path)
> @@ -629,7 +631,7 @@ void coresight_disable(struct coresight_device *csdev)
> if (ret)
> goto out;
>
> - if (!csdev->enable)
> + if (!csdev->enable || !coresight_disable_source(csdev))
> goto out;
>
> switch (csdev->subtype.source_subtype) {
> @@ -647,7 +649,6 @@ void coresight_disable(struct coresight_device *csdev)
> break;
> }
>
> - coresight_disable_source(csdev);
> coresight_disable_path(path);
> coresight_release_path(path);
>
On 14/03/17 17:37, Mathieu Poirier wrote:
> Hi Suzuki,
>
> On 14 March 2017 at 11:31, Mathieu Poirier <[email protected]> wrote:
>> From: "Suzuki K. Poulose" <[email protected]>
>>
>> With a coresight tracing session, the components along the path
>> from the source to sink are disabled after the source is disabled.
>> However, if the source was not actually disabled due to active
>> users, we should not disable the components in the path.
>>
>> Cc: Mathieu Poirier <[email protected]>
>> Signed-off-by: Suzuki K Poulose <[email protected]>
>> ---
>> drivers/hwtracing/coresight/coresight.c | 7 ++++---
>> 1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
>> index 0c37356..34cd1ed 100644
>> --- a/drivers/hwtracing/coresight/coresight.c
>> +++ b/drivers/hwtracing/coresight/coresight.c
>> @@ -253,7 +253,8 @@ static int coresight_enable_source(struct coresight_device *csdev, u32 mode)
>> return 0;
>> }
>>
>> -static void coresight_disable_source(struct coresight_device *csdev)
>> +/* coresight_disable_source: Returns true if the device has been disabled */
>
> To be coherent with the rest of the file please use:
>
> /*
> * coresight_disable_source: Returns true if the device has been disabled
> */
>
> I would have done the modification myself but 2/2 needs tending as well.
>
Ah, sorry about that, will resend it. Thanks for spotting.
Suzuki
On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]> wrote:
> From: "Suzuki K. Poulose" <[email protected]>
>
> For software sources (i.e STM), there could be multiple agents
> generating the trace data, unlike the ETMs. So we need to
> properly do the accounting for the active number of users
> to disable the device when the last user goes away. Right
> now, the reference counting is broken for sources as we skip
> the actions when we detect that the source is enabled.
>
> This patch fixes the problem by adding the refcounting for
> software sources, even when they are enabled.
>
> Cc: Mathieu Poirier <[email protected]>
> Reported-by: Robert Walker <[email protected]>
> Signed-off-by: Suzuki K Poulose <[email protected]>
> ---
> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
> 1 file changed, 11 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
> index 34cd1ed..2da9e39 100644
> --- a/drivers/hwtracing/coresight/coresight.c
> +++ b/drivers/hwtracing/coresight/coresight.c
> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
> int cpu, ret = 0;
> struct coresight_device *sink;
> struct list_head *path;
> + enum coresight_dev_subtype_source subtype = csdev->subtype.source_subtype;
Checkpatch.pl complains about a line over 80 characters.
>
> mutex_lock(&coresight_mutex);
>
> @@ -559,8 +560,16 @@ int coresight_enable(struct coresight_device *csdev)
> if (ret)
> goto out;
>
> - if (csdev->enable)
> + if (csdev->enable) {
> + /*
> + * There could be multiple applications driving the software
> + * source. So keep the refcount for each such user when the
> + * source is already enabled.
> + */
> + if (subtype == CORESIGHT_DEV_SUBTYPE_SOURCE_SOFTWARE)
> + atomic_inc(csdev->refcnt);
> goto out;
> + }
>
> /*
> * Search for a valid sink for this session but don't reset the
> @@ -587,7 +596,7 @@ int coresight_enable(struct coresight_device *csdev)
> if (ret)
> goto err_source;
>
> - switch (csdev->subtype.source_subtype) {
> + switch (subtype) {
> case CORESIGHT_DEV_SUBTYPE_SOURCE_PROC:
> /*
> * When working from sysFS it is important to keep track
On 14/03/17 17:40, Mathieu Poirier wrote:
> On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]> wrote:
>> From: "Suzuki K. Poulose" <[email protected]>
>>
>> For software sources (i.e STM), there could be multiple agents
>> generating the trace data, unlike the ETMs. So we need to
>> properly do the accounting for the active number of users
>> to disable the device when the last user goes away. Right
>> now, the reference counting is broken for sources as we skip
>> the actions when we detect that the source is enabled.
>>
>> This patch fixes the problem by adding the refcounting for
>> software sources, even when they are enabled.
>>
>> Cc: Mathieu Poirier <[email protected]>
>> Reported-by: Robert Walker <[email protected]>
>> Signed-off-by: Suzuki K Poulose <[email protected]>
>> ---
>> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
>> 1 file changed, 11 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
>> index 34cd1ed..2da9e39 100644
>> --- a/drivers/hwtracing/coresight/coresight.c
>> +++ b/drivers/hwtracing/coresight/coresight.c
>> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
>> int cpu, ret = 0;
>> struct coresight_device *sink;
>> struct list_head *path;
>> + enum coresight_dev_subtype_source subtype = csdev->subtype.source_subtype;
>
> Checkpatch.pl complains about a line over 80 characters.
Hmm, I think the name of the enum is unusually long. I could clean it up a little bit.
Also, it does look silly to split a simple variable assignment though.
What do you prefer ?
Suzuki
On 14/03/17 17:40, Mathieu Poirier wrote:
> On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]> wrote:
>> From: "Suzuki K. Poulose" <[email protected]>
>>
>> For software sources (i.e STM), there could be multiple agents
>> generating the trace data, unlike the ETMs. So we need to
>> properly do the accounting for the active number of users
>> to disable the device when the last user goes away. Right
>> now, the reference counting is broken for sources as we skip
>> the actions when we detect that the source is enabled.
>>
>> This patch fixes the problem by adding the refcounting for
>> software sources, even when they are enabled.
>>
>> Cc: Mathieu Poirier <[email protected]>
>> Reported-by: Robert Walker <[email protected]>
>> Signed-off-by: Suzuki K Poulose <[email protected]>
>> ---
>> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
>> 1 file changed, 11 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/hwtracing/coresight/coresight.c b/drivers/hwtracing/coresight/coresight.c
>> index 34cd1ed..2da9e39 100644
>> --- a/drivers/hwtracing/coresight/coresight.c
>> +++ b/drivers/hwtracing/coresight/coresight.c
>> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
>> int cpu, ret = 0;
>> struct coresight_device *sink;
>> struct list_head *path;
>> + enum coresight_dev_subtype_source subtype = csdev->subtype.source_subtype;
>
> Checkpatch.pl complains about a line over 80 characters.
>
>>
>> mutex_lock(&coresight_mutex);
>>
>> @@ -559,8 +560,16 @@ int coresight_enable(struct coresight_device *csdev)
>> if (ret)
>> goto out;
>>
>> - if (csdev->enable)
>> + if (csdev->enable) {
>> + /*
>> + * There could be multiple applications driving the software
>> + * source. So keep the refcount for each such user when the
>> + * source is already enabled.
>> + */
>> + if (subtype == CORESIGHT_DEV_SUBTYPE_SOURCE_SOFTWARE)
>> + atomic_inc(csdev->refcnt);
>> goto out;
>> + }
>>
Btw, should we allow the user to turn on the STM from sysfs (echo 1 > $STM/enable_source) ?
All STM users should set their policy via ioctls and that in turn turns the device on. So
it doesn't make sense for enable_source to really enable the hardware unless someone really
opens it.
Chunyan, Mathieu,
thoughts ?
Suzuki
On 14 March 2017 at 11:55, Suzuki K Poulose <[email protected]> wrote:
> On 14/03/17 17:40, Mathieu Poirier wrote:
>>
>> On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]>
>> wrote:
>>>
>>> From: "Suzuki K. Poulose" <[email protected]>
>>>
>>> For software sources (i.e STM), there could be multiple agents
>>> generating the trace data, unlike the ETMs. So we need to
>>> properly do the accounting for the active number of users
>>> to disable the device when the last user goes away. Right
>>> now, the reference counting is broken for sources as we skip
>>> the actions when we detect that the source is enabled.
>>>
>>> This patch fixes the problem by adding the refcounting for
>>> software sources, even when they are enabled.
>>>
>>> Cc: Mathieu Poirier <[email protected]>
>>> Reported-by: Robert Walker <[email protected]>
>>> Signed-off-by: Suzuki K Poulose <[email protected]>
>>> ---
>>> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
>>> 1 file changed, 11 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/hwtracing/coresight/coresight.c
>>> b/drivers/hwtracing/coresight/coresight.c
>>> index 34cd1ed..2da9e39 100644
>>> --- a/drivers/hwtracing/coresight/coresight.c
>>> +++ b/drivers/hwtracing/coresight/coresight.c
>>> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
>>> int cpu, ret = 0;
>>> struct coresight_device *sink;
>>> struct list_head *path;
>>> + enum coresight_dev_subtype_source subtype =
>>> csdev->subtype.source_subtype;
>>
>>
>> Checkpatch.pl complains about a line over 80 characters.
>
>
>
> Hmm, I think the name of the enum is unusually long. I could clean it up a
> little bit.
> Also, it does look silly to split a simple variable assignment though.
>
> What do you prefer ?
enum coresight_dev_subtype_source subtype;
...
...
subtype = csdev->subtype.source_subtype;
That is probably the easiest.
>
> Suzuki
Hi Suzuki,
On 15 March 2017 at 02:06, Suzuki K Poulose <[email protected]> wrote:
> On 14/03/17 17:40, Mathieu Poirier wrote:
>>
>> On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]>
>> wrote:
>>>
>>> From: "Suzuki K. Poulose" <[email protected]>
>>>
>>> For software sources (i.e STM), there could be multiple agents
>>> generating the trace data, unlike the ETMs. So we need to
>>> properly do the accounting for the active number of users
>>> to disable the device when the last user goes away. Right
>>> now, the reference counting is broken for sources as we skip
>>> the actions when we detect that the source is enabled.
>>>
>>> This patch fixes the problem by adding the refcounting for
>>> software sources, even when they are enabled.
>>>
>>> Cc: Mathieu Poirier <[email protected]>
>>> Reported-by: Robert Walker <[email protected]>
>>> Signed-off-by: Suzuki K Poulose <[email protected]>
>>> ---
>>> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
>>> 1 file changed, 11 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/hwtracing/coresight/coresight.c
>>> b/drivers/hwtracing/coresight/coresight.c
>>> index 34cd1ed..2da9e39 100644
>>> --- a/drivers/hwtracing/coresight/coresight.c
>>> +++ b/drivers/hwtracing/coresight/coresight.c
>>> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
>>> int cpu, ret = 0;
>>> struct coresight_device *sink;
>>> struct list_head *path;
>>> + enum coresight_dev_subtype_source subtype =
>>> csdev->subtype.source_subtype;
>>
>>
>> Checkpatch.pl complains about a line over 80 characters.
>>
>>>
>>> mutex_lock(&coresight_mutex);
>>>
>>> @@ -559,8 +560,16 @@ int coresight_enable(struct coresight_device *csdev)
>>> if (ret)
>>> goto out;
>>>
>>> - if (csdev->enable)
>>> + if (csdev->enable) {
>>> + /*
>>> + * There could be multiple applications driving the
>>> software
>>> + * source. So keep the refcount for each such user when
>>> the
>>> + * source is already enabled.
>>> + */
>>> + if (subtype == CORESIGHT_DEV_SUBTYPE_SOURCE_SOFTWARE)
>>> + atomic_inc(csdev->refcnt);
>>> goto out;
>>> + }
>>>
>
> Btw, should we allow the user to turn on the STM from sysfs (echo 1 >
> $STM/enable_source) ?
If enabling STM can not be allowed via sysfs, how should we allow
users to turn on STM when they want to mmap STM to user space, and
write STM device from user space directly? For example this kind of
use case [1].
> All STM users should set their policy via ioctls and that in turn turns the
> device on.
Yes users can set policy via ioctls to request resource of STM (i.e.
which STM channel(s) will be written), but they still need to use
sysfs to enable STM.
> So it doesn't make sense for enable_source to really enable
> the hardware unless someone really opens it.
Right, there're two ways to enable STM currently, e.g.
1) echo <addr>.stm > /sys/class/stm_source/stm_ftrace/stm_source_link
2) echo 1 > $STM/enable_source
That would probably make people confused, I would appreciate any
better solution.
Thanks,
Chunyan
[1] https://github.com/lyrazhang/user-space-tests/tree/mmap-wrapper/mmap-wrapper
>
> Chunyan, Mathieu,
>
> thoughts ?
>
> Suzuki
>
On 15/03/17 03:51, Chunyan Zhang wrote:
> Hi Suzuki,
>
> On 15 March 2017 at 02:06, Suzuki K Poulose <[email protected]> wrote:
>> On 14/03/17 17:40, Mathieu Poirier wrote:
>>>
>>> On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]>
>>> wrote:
>>>>
>>>> From: "Suzuki K. Poulose" <[email protected]>
>>>>
>>>> For software sources (i.e STM), there could be multiple agents
>>>> generating the trace data, unlike the ETMs. So we need to
>>>> properly do the accounting for the active number of users
>>>> to disable the device when the last user goes away. Right
>>>> now, the reference counting is broken for sources as we skip
>>>> the actions when we detect that the source is enabled.
>>>>
>>>> This patch fixes the problem by adding the refcounting for
>>>> software sources, even when they are enabled.
>>>>
>>>> Cc: Mathieu Poirier <[email protected]>
>>>> Reported-by: Robert Walker <[email protected]>
>>>> Signed-off-by: Suzuki K Poulose <[email protected]>
>>>> ---
>>>> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
>>>> 1 file changed, 11 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/drivers/hwtracing/coresight/coresight.c
>>>> b/drivers/hwtracing/coresight/coresight.c
>>>> index 34cd1ed..2da9e39 100644
>>>> --- a/drivers/hwtracing/coresight/coresight.c
>>>> +++ b/drivers/hwtracing/coresight/coresight.c
>>>> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device *csdev)
>>>> int cpu, ret = 0;
>>>> struct coresight_device *sink;
>>>> struct list_head *path;
>>>> + enum coresight_dev_subtype_source subtype =
>>>> csdev->subtype.source_subtype;
>>>
>>>
>>> Checkpatch.pl complains about a line over 80 characters.
>>>
>>>>
>>>> mutex_lock(&coresight_mutex);
>>>>
>>>> @@ -559,8 +560,16 @@ int coresight_enable(struct coresight_device *csdev)
>>>> if (ret)
>>>> goto out;
>>>>
>>>> - if (csdev->enable)
>>>> + if (csdev->enable) {
>>>> + /*
>>>> + * There could be multiple applications driving the
>>>> software
>>>> + * source. So keep the refcount for each such user when
>>>> the
>>>> + * source is already enabled.
>>>> + */
>>>> + if (subtype == CORESIGHT_DEV_SUBTYPE_SOURCE_SOFTWARE)
>>>> + atomic_inc(csdev->refcnt);
>>>> goto out;
>>>> + }
>>>>
>>
>> Btw, should we allow the user to turn on the STM from sysfs (echo 1 >
>> $STM/enable_source) ?
>
> If enabling STM can not be allowed via sysfs, how should we allow
> users to turn on STM when they want to mmap STM to user space, and
> write STM device from user space directly? For example this kind of
> use case [1].
The ioct(, STP_POLICY_ID_SET) indirectly turns on the STM hardware via :
stm_char_policy_set_ioctl()->stm.link (stm_generic_link)-> coresight_enable().
>
>> All STM users should set their policy via ioctls and that in turn turns the
>> device on.
>
> Yes users can set policy via ioctls to request resource of STM (i.e.
> which STM channel(s) will be written), but they still need to use
> sysfs to enable STM.
As mentioned above, it is not necessary.
>
>> So it doesn't make sense for enable_source to really enable
>> the hardware unless someone really opens it.
>
> Right, there're two ways to enable STM currently, e.g.
> 1) echo <addr>.stm > /sys/class/stm_source/stm_ftrace/stm_source_link
I am not familiar with the stm_source class. From a quick glance, it looks like,
writing to stm_source_link triggers :
stm_source_link_store()->stm_source_link_add()->(stm->data->link()).
which is fine for connecting a source (ftrace,console or heartbeat) to STM.
> 2) echo 1 > $STM/enable_source
This is a bit awkward for an application who wants to mmap and stream data,
and is quite unnecessary from my explanation above.
>
> That would probably make people confused, I would appreciate any
> better solution.
Please let me know if you have any outstanding concerns.
Cheers
Suzuki
On 16 March 2017 at 00:00, Suzuki K Poulose <[email protected]> wrote:
> On 15/03/17 03:51, Chunyan Zhang wrote:
>>
>> Hi Suzuki,
>>
>> On 15 March 2017 at 02:06, Suzuki K Poulose <[email protected]>
>> wrote:
>>>
>>> On 14/03/17 17:40, Mathieu Poirier wrote:
>>>>
>>>>
>>>> On 14 March 2017 at 11:32, Mathieu Poirier <[email protected]>
>>>> wrote:
>>>>>
>>>>>
>>>>> From: "Suzuki K. Poulose" <[email protected]>
>>>>>
>>>>> For software sources (i.e STM), there could be multiple agents
>>>>> generating the trace data, unlike the ETMs. So we need to
>>>>> properly do the accounting for the active number of users
>>>>> to disable the device when the last user goes away. Right
>>>>> now, the reference counting is broken for sources as we skip
>>>>> the actions when we detect that the source is enabled.
>>>>>
>>>>> This patch fixes the problem by adding the refcounting for
>>>>> software sources, even when they are enabled.
>>>>>
>>>>> Cc: Mathieu Poirier <[email protected]>
>>>>> Reported-by: Robert Walker <[email protected]>
>>>>> Signed-off-by: Suzuki K Poulose <[email protected]>
>>>>> ---
>>>>> drivers/hwtracing/coresight/coresight.c | 13 +++++++++++--
>>>>> 1 file changed, 11 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/hwtracing/coresight/coresight.c
>>>>> b/drivers/hwtracing/coresight/coresight.c
>>>>> index 34cd1ed..2da9e39 100644
>>>>> --- a/drivers/hwtracing/coresight/coresight.c
>>>>> +++ b/drivers/hwtracing/coresight/coresight.c
>>>>> @@ -552,6 +552,7 @@ int coresight_enable(struct coresight_device
>>>>> *csdev)
>>>>> int cpu, ret = 0;
>>>>> struct coresight_device *sink;
>>>>> struct list_head *path;
>>>>> + enum coresight_dev_subtype_source subtype =
>>>>> csdev->subtype.source_subtype;
>>>>
>>>>
>>>>
>>>> Checkpatch.pl complains about a line over 80 characters.
>>>>
>>>>>
>>>>> mutex_lock(&coresight_mutex);
>>>>>
>>>>> @@ -559,8 +560,16 @@ int coresight_enable(struct coresight_device
>>>>> *csdev)
>>>>> if (ret)
>>>>> goto out;
>>>>>
>>>>> - if (csdev->enable)
>>>>> + if (csdev->enable) {
>>>>> + /*
>>>>> + * There could be multiple applications driving the
>>>>> software
>>>>> + * source. So keep the refcount for each such user when
>>>>> the
>>>>> + * source is already enabled.
>>>>> + */
>>>>> + if (subtype == CORESIGHT_DEV_SUBTYPE_SOURCE_SOFTWARE)
>>>>> + atomic_inc(csdev->refcnt);
>>>>> goto out;
>>>>> + }
>>>>>
>>>
>>> Btw, should we allow the user to turn on the STM from sysfs (echo 1 >
>>> $STM/enable_source) ?
>>
>>
>> If enabling STM can not be allowed via sysfs, how should we allow
>> users to turn on STM when they want to mmap STM to user space, and
>> write STM device from user space directly? For example this kind of
>> use case [1].
>
>
> The ioct(, STP_POLICY_ID_SET) indirectly turns on the STM hardware via :
> stm_char_policy_set_ioctl()->stm.link (stm_generic_link)->
> coresight_enable().
>
Ah, that's right. Especially after your patch [1] merged, it is
indeed not necessary. Thanks for pointing this.
>
>>
>>> All STM users should set their policy via ioctls and that in turn turns
>>> the
>>> device on.
>>
>>
>> Yes users can set policy via ioctls to request resource of STM (i.e.
>> which STM channel(s) will be written), but they still need to use
>> sysfs to enable STM.
>
>
> As mentioned above, it is not necessary.
>
>>
>>> So it doesn't make sense for enable_source to really enable
>>> the hardware unless someone really opens it.
>>
>>
>> Right, there're two ways to enable STM currently, e.g.
>> 1) echo <addr>.stm > /sys/class/stm_source/stm_ftrace/stm_source_link
>
>
> I am not familiar with the stm_source class. From a quick glance, it looks
> like,
> writing to stm_source_link triggers :
> stm_source_link_store()->stm_source_link_add()->(stm->data->link()).
>
> which is fine for connecting a source (ftrace,console or heartbeat) to STM.
>
>> 2) echo 1 > $STM/enable_source
>
>
> This is a bit awkward for an application who wants to mmap and stream data,
> and is quite unnecessary from my explanation above.
>
>>
>> That would probably make people confused, I would appreciate any
>> better solution.
>
>
> Please let me know if you have any outstanding concerns.
I haven't thought out other use case which need this sysfs interface
for CoreSight STM device, I'm ok with hiding it, but I'm not sure if
it's good to remove this ABI from CoreSight STM but leave it to other
CoreSight source components.
Mathieu,
Any thought?
Thanks,
Chunyan
[1] http://lists.infradead.org/pipermail/linux-arm-kernel/2017-January/479893.html
>
> Cheers
> Suzuki
>
On 16/03/17 12:13, Chunyan Zhang wrote:
> On 16 March 2017 at 00:00, Suzuki K Poulose <[email protected]> wrote:
>> On 15/03/17 03:51, Chunyan Zhang wrote:
>>>> Btw, should we allow the user to turn on the STM from sysfs (echo 1 >
>>>> $STM/enable_source) ?
>>>
>>>
>>> If enabling STM can not be allowed via sysfs, how should we allow
>>> users to turn on STM when they want to mmap STM to user space, and
>>> write STM device from user space directly? For example this kind of
>>> use case [1].
>>
>>
>> The ioct(, STP_POLICY_ID_SET) indirectly turns on the STM hardware via :
>> stm_char_policy_set_ioctl()->stm.link (stm_generic_link)->
>> coresight_enable().
>>
>
> Ah, that's right. Especially after your patch [1] merged, it is
> indeed not necessary. Thanks for pointing this.
>
>>
>>>
>>>> All STM users should set their policy via ioctls and that in turn turns
>>>> the
>>>> device on.
>>>
>>>
>>> Yes users can set policy via ioctls to request resource of STM (i.e.
>>> which STM channel(s) will be written), but they still need to use
>>> sysfs to enable STM.
>>
>>
>> As mentioned above, it is not necessary.
>>
>>>
>>>> So it doesn't make sense for enable_source to really enable
>>>> the hardware unless someone really opens it.
>>>
>>>
>>> Right, there're two ways to enable STM currently, e.g.
>>> 1) echo <addr>.stm > /sys/class/stm_source/stm_ftrace/stm_source_link
>>
>>
>> I am not familiar with the stm_source class. From a quick glance, it looks
>> like,
>> writing to stm_source_link triggers :
>> stm_source_link_store()->stm_source_link_add()->(stm->data->link()).
>>
>> which is fine for connecting a source (ftrace,console or heartbeat) to STM.
>>
>>> 2) echo 1 > $STM/enable_source
>>
>>
>> This is a bit awkward for an application who wants to mmap and stream data,
>> and is quite unnecessary from my explanation above.
>>
>>>
>>> That would probably make people confused, I would appreciate any
>>> better solution.
>>
>>
>> Please let me know if you have any outstanding concerns.
>
> I haven't thought out other use case which need this sysfs interface
> for CoreSight STM device, I'm ok with hiding it, but I'm not sure if
> it's good to remove this ABI from CoreSight STM but leave it to other
> CoreSight source components.
I was not planning to remove the ABI, but make it a no-op.
I had a chat with some of our experts in the area and they have one use case
which needs the STM turned on. If the user wants to run two applications
back to back (without overlap in execution), there is no way to make sure that
the STM stays on after the first application exits. So I think that is a valid
use case. So, we cannot make it
The reason I brought up this discussion is that we don't know how many times the
STM is enabled from the sysfs vs. stm_generic_link(). This could possibly cause
issues with sysfs disabling the STM (by unbalanced enable/disable). As the user
doesn't necessarily know how many times he/she enabled the device. May be we
should keep track of the sysfs refcount separately and pass it down to the csdev->refcnt
only once.
I will cook up a patch.
Suzuki
>
> Thanks,
> Chunyan
>
> [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2017-January/479893.html
>
>>
>> Cheers
>> Suzuki
>>