[PATCH 1/1] ARM: OMAP2+: hwmod: Print an error message if no match exists for a hwmod class

If no match is found for a hwmod class, omap_hwmod_for_each_by_class()
currently does not spit out any error messages. To make debugging
issues related to non-existent hwmod classes simpler make this function
print an error message in case no match is found for the supplied hwmod
class.

Signed-off-by: Vaibhav Bedia <redacted>
---
 arch/arm/mach-omap2/omap_hwmod.c |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)

diff --git a/arch/arm/mach-omap2/omap_hwmod.c b/arch/arm/mach-omap2/omap_hwmod.c
index 6ca8e51..90e2306 100644
--- a/arch/arm/mach-omap2/omap_hwmod.c
+++ b/arch/arm/mach-omap2/omap_hwmod.c

@@ -3480,7 +3480,8 @@ int omap_hwmod_read_hardreset(struct omap_hwmod *oh, const char *name)
  * If the callback function returns something other than
  * zero, the iterator is terminated, and the callback function's return
  * value is passed back to the caller.  Returns 0 upon success, -EINVAL
- * if @classname or @fn are NULL, or passes back the error code from @fn.
+ * if @classname or @fn are NULL or no match is found for @classname,
+ * or passes back the error code from @fn.
  */
 int omap_hwmod_for_each_by_class(const char *classname,
 				 int (*fn)(struct omap_hwmod *oh,

@@ -3489,6 +3490,7 @@ int omap_hwmod_for_each_by_class(const char *classname,
 {
 	struct omap_hwmod *temp_oh;
 	int ret = 0;
+	bool found = false;
 
 	if (!classname || !fn)
 		return -EINVAL;

@@ -3498,6 +3500,7 @@ int omap_hwmod_for_each_by_class(const char *classname,
 
 	list_for_each_entry(temp_oh, &omap_hwmod_list, node) {
 		if (!strcmp(temp_oh->class->name, classname)) {
+			found = true;
 			pr_debug("omap_hwmod: %s: %s: calling callback fn\n",
 				 __func__, temp_oh->name);
 			ret = (*fn)(temp_oh, user);

@@ -3506,6 +3509,12 @@ int omap_hwmod_for_each_by_class(const char *classname,
 		}
 	}
 
+	if (!found) {
+		pr_err("omap_hwmod: %s: no match for class %s\n",
+			__func__, classname);
+		return -EINVAL;
+	}
+
 	if (ret)
 		pr_debug("omap_hwmod: %s: iterator terminated early: %d\n",
 			 __func__, ret);

-- 
1.7.0.4

Hello Vaibhav,

On Wed, 8 Aug 2012, Vaibhav Bedia wrote:

If no match is found for a hwmod class, omap_hwmod_for_each_by_class()
currently does not spit out any error messages. To make debugging
issues related to non-existent hwmod classes simpler make this function
print an error message in case no match is found for the supplied hwmod
class.

Signed-off-by: Vaibhav Bedia <redacted>

It doesn't seem to me that this would necessarily always be an error.  
Suppose some init code to create cpsw devices is running on an OMAP3430.  
Then a omap_hwmod_for_each_by_class() that doesn't locate anything 
wouldn't be an error, right?


- Paul

Hi Paul,

On Wed, Aug 08, 2012 at 19:35:50, Paul Walmsley wrote:
[...]

It doesn't seem to me that this would necessarily always be an error.  
Suppose some init code to create cpsw devices is running on an OMAP3430.  
Then a omap_hwmod_for_each_by_class() that doesn't locate anything 
wouldn't be an error, right?

Well, I think a failed lookup operation will be a bit expensive since it
iterates over the whole list of registered hwmods and hence put in the
pr_err(). But I don't see how the SoC specific example that you pointed out
could be handled without putting in cpu_is_* checks in the registration
process itself :(

Would a pr_debug() be ok?

Regards,
Vaibhav B.

Hi Vaibhav

On Wed, 8 Aug 2012, Bedia, Vaibhav wrote:

On Wed, Aug 08, 2012 at 19:35:50, Paul Walmsley wrote:

quoted

It doesn't seem to me that this would necessarily always be an error.  
Suppose some init code to create cpsw devices is running on an OMAP3430.  
Then a omap_hwmod_for_each_by_class() that doesn't locate anything 
wouldn't be an error, right?

Well, I think a failed lookup operation will be a bit expensive since it
iterates over the whole list of registered hwmods and hence put in the
pr_err().

If performance is a serious problem there, we can maintain a list of hwmod 
pointers inside each struct omap_hwmod_class, updated when the hwmods are 
registered.  But as far as I know, performance hasn't been an issue here 
yet.  If performance is the real motivation, I'd rather see an 
implementation of per-class lists as described above.

But I don't see how the SoC specific example that you pointed out
could be handled without putting in cpu_is_* checks in the registration
process itself :(

Hmm, why not?

To continue the earlier example, the init code in the platform code simply 
calls omap_hwmod_for_each_by_class(), and creates the appropriate device 
if anything was found.  If no cpsw IP blocks are present on a given SoC, 
nothing would happen.  No cpu_is_*() would be needed.


- Paul

Hi Paul,

On Thu, Aug 09, 2012 at 01:49:34, Paul Walmsley wrote:

Hi Vaibhav

On Wed, 8 Aug 2012, Bedia, Vaibhav wrote:

quoted

On Wed, Aug 08, 2012 at 19:35:50, Paul Walmsley wrote:

quoted

It doesn't seem to me that this would necessarily always be an error.  
Suppose some init code to create cpsw devices is running on an OMAP3430.  
Then a omap_hwmod_for_each_by_class() that doesn't locate anything 
wouldn't be an error, right?

Well, I think a failed lookup operation will be a bit expensive since it
iterates over the whole list of registered hwmods and hence put in the
pr_err().

If performance is a serious problem there, we can maintain a list of hwmod 
pointers inside each struct omap_hwmod_class, updated when the hwmods are 
registered.  But as far as I know, performance hasn't been an issue here 
yet.  If performance is the real motivation, I'd rather see an 
implementation of per-class lists as described above.

Ok. I'll try to do some profiling using a timer to see how much time we
spend over here. On AM335x there are 2 such lookups, one for the dma
and one for the mcbsp. I could get this done today but will get to it
tomorrow. Based on how this goes I'll try out the omap_hwmod_class list
like you suggested.

quoted

But I don't see how the SoC specific example that you pointed out
could be handled without putting in cpu_is_* checks in the registration
process itself :(

Hmm, why not?

To continue the earlier example, the init code in the platform code simply 
calls omap_hwmod_for_each_by_class(), and creates the appropriate device 
if anything was found.  If no cpsw IP blocks are present on a given SoC, 
nothing would happen.  No cpu_is_*() would be needed.

I was thinking of avoiding the lookup process itself and hence mentioned
the need for cpu checks surrounding the platform init code. The current
API does help avoid all this in an excellent manner.

Regards,
Vaibhav B. 

Hi Paul,

[...]

quoted

quoted

On Wed, Aug 08, 2012 at 19:35:50, Paul Walmsley wrote:

quoted

It doesn't seem to me that this would necessarily always be an error.  
Suppose some init code to create cpsw devices is running on an OMAP3430.  
Then a omap_hwmod_for_each_by_class() that doesn't locate anything 
wouldn't be an error, right?

Well, I think a failed lookup operation will be a bit expensive since it
iterates over the whole list of registered hwmods and hence put in the
pr_err().

If performance is a serious problem there, we can maintain a list of hwmod 
pointers inside each struct omap_hwmod_class, updated when the hwmods are 
registered.  But as far as I know, performance hasn't been an issue here 
yet.  If performance is the real motivation, I'd rather see an 
implementation of per-class lists as described above.

Ok. I'll try to do some profiling using a timer to see how much time we
spend over here. On AM335x there are 2 such lookups, one for the dma
and one for the mcbsp. I could get this done today but will get to it
tomorrow. Based on how this goes I'll try out the omap_hwmod_class list
like you suggested.

I used a dmtimer clocked at 24MHz (41ns resolution) to get the time
spent in the lookup. There are two failed lookups that happen in AM335x
and following are the times taken

Run1
Lookup for dma - 435 cycles -> 17.835 usecs
Lookup for mcbsp - 83 cycles -> 3.4 usecs

Run2
Lookup for dma - 503 cycles -> 20.623 usecs
Lookup for mcbsp - 90 cycles -> 3.69 usescs

Run3
Lookup for dma - 375 cycles -> 15.375 usecs
Lookup for mcbsp - 91 cycles -> 3.731 usecs

The 2nd lookup happens immediately after the 1st one so probably
the difference is coming due to the data sitting in the cache?

The average times don't look too high to me but I guess you already
knew that ;)

If you are ok with a pr_debug() instead of a pr_err() I'll resubmit
the patch with this change.

Regards,
Vaibhav B.

Hello Vaibhav

On Fri, 10 Aug 2012, Bedia, Vaibhav wrote:

quoted

Ok. I'll try to do some profiling using a timer to see how much time we
spend over here. On AM335x there are 2 such lookups, one for the dma
and one for the mcbsp. I could get this done today but will get to it
tomorrow. Based on how this goes I'll try out the omap_hwmod_class list
like you suggested.

I used a dmtimer clocked at 24MHz (41ns resolution) to get the time
spent in the lookup. There are two failed lookups that happen in AM335x
and following are the times taken

Run1
Lookup for dma - 435 cycles -> 17.835 usecs
Lookup for mcbsp - 83 cycles -> 3.4 usecs

Run2
Lookup for dma - 503 cycles -> 20.623 usecs
Lookup for mcbsp - 90 cycles -> 3.69 usescs

Run3
Lookup for dma - 375 cycles -> 15.375 usecs
Lookup for mcbsp - 91 cycles -> 3.731 usecs

The 2nd lookup happens immediately after the 1st one so probably
the difference is coming due to the data sitting in the cache?

The average times don't look too high to me but I guess you already
knew that ;)

If you are ok with a pr_debug() instead of a pr_err() I'll resubmit
the patch with this change.

Well I'd much rather see a patch that maintains a list of hwmods 
pertaining to each class.  That would solve the performance issue at 
fairly minimal cost (especially if those lists were freed after kernel 
init). 

It sounds like the basic difference in opinion is about whether an 
omap_hwmod_for_each_by_class() that doesn't find anything is an error 
condition or not.  My sense is that it is not.


- Paul