On 03/20/2012 05:13 PM, Turquette, Mike wrote:

On Tue, Mar 20, 2012 at 12:46 AM, Saravana Kannan
[off-list ref]  wrote:

quoted

On Tue, March 20, 2012 12:19 am, Sascha Hauer wrote:

quoted

On Mon, Mar 19, 2012 at 08:38:25PM -0700, Saravana Kannan wrote:

quoted

If memory allocation for the parents array or the parent string fails,
then
fail the registration immediately instead of calling clk_register and
hoping it fails there.

Return -ENOMEM on failure.

Signed-off-by: Saravana Kannan<redacted>
Cc: Mike Turquette<redacted>
Cc: Andrew Lunn<andrew@lunn.ch>
Cc: Rob Herring<redacted>
Cc: Russell King<redacted>
Cc: Jeremy Kerr<redacted>
Cc: Thomas Gleixner<redacted>
Cc: Arnd Bergman<redacted>
Cc: Paul Walmsley<paul@pwsan.com>
Cc: Shawn Guo<redacted>
Cc: Sascha Hauer<s.hauer@pengutronix.de>
Cc: Jamie Iles<redacted>
Cc: Richard Zhao<redacted>
Cc: Saravana Kannan<redacted>
Cc: Magnus Damm<magnus.damm@gmail.com>
Cc: Mark Brown<redacted>
Cc: Linus Walleij<redacted>
Cc: Stephen Boyd<redacted>
Cc: Amit Kucheria<redacted>
Cc: Deepak Saxena<redacted>
Cc: Grant Likely<redacted>
---
There are still some memory free issues when clk_register() fails, but I
will
fix it when I fixed the other register() fns to return ENOMEM of alloc
failure instead of a NULL.

  drivers/clk/clk-fixed-rate.c |   10 +++++++---
  1 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/drivers/clk/clk-fixed-rate.c b/drivers/clk/clk-fixed-rate.c
index 90c79fb..6423ae9 100644
--- a/drivers/clk/clk-fixed-rate.c
+++ b/drivers/clk/clk-fixed-rate.c

@@ -61,22 +61,26 @@ struct clk *clk_register_fixed_rate(struct device

*dev, const char *name,
              parent_names = kmalloc(sizeof(char *), GFP_KERNEL);

              if (! parent_names)
-                    goto out;
+                    goto fail_ptr;

              len = sizeof(char) * strlen(parent_name);

              parent_names[0] = kmalloc(len, GFP_KERNEL);

              if (!parent_names[0])
-                    goto out;
+                    goto fail_str;

              strncpy(parent_names[0], parent_name, len);
      }

It's easier to add a char *parent to struct clk_fixed and pass it to
clk_register with&fixed->parent. This saves you a kmalloc call and
makes the error path simpler. It's the same way already done in the
divider.

I thought I had done this for v7... hmm looks like one got left out.
I'll line up a patch to get it in sync with the others as part of my
fixes.

quoted

I thought about that since I saw the same was done for gated and divider
(I think). Here is my guess at Mike's reasoning for this:

Gated and divider clocks have to have a parent. There's nothing to gate
otherwise. But fixed rate clocks might not have a parent. It could be XO's
or PLLs running off of always on XOs not controlled by the SoC. So, it's
arguable to not have a parent. I don't have a strong opinion on this --
since Mike took the time to write it, it left it to his subjective
preference.

I appreciate the thoughtfulness.  Re-using the same type of mechanism
as the divider and gate clocks will still allow the fixed-rate clock
to be parentless, and it makes for cleaner code, one less allocation
and lines up with how the other single-parent basic clocks are done,
so I'll take that method in instead of your patch.

No problem, go for it.

quoted

I sent this patch first since it was around the place I was cleaning up. I
didn't want to actually just shuffle around a bug. As I mentioned, this
patch still leaves a bug open -- what if clk_register() fails. I plan to
fix that once my two patches are picked up (hopefully).

Do you still find it useful to return -ENOMEM from the registration
function instead of a NULL clock?  I'm always worried that people
don't check for error codes on pointers in their platform code and
only check for NULL...

The last discussion I remember, NULL was considered a valid clock. So, I 
think on error, we shouldn't ever return NULL when the return type is 
struct clk *.

Thanks,
Saravana

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.