[PATCH] gpio/omap: fix invalid context restore of gpio bank-0

Currently the gpio _runtime_resume/suspend functions are calling the
get_context_loss_count() platform function if the function is populated for
a gpio bank. This function is used to determine if the gpio bank logic state
needs to be restored due to a power transition. This function will be populated
for all banks, but it should only be called for banks that have the
"loses_context" variable set. It is pointless to call this if loses_context is
false as we know the context will never be lost and will not need restoring.

For all OMAP2+ devices gpio bank-0 is in an always-on power domain and so will
never lose context. We found that the get_context_loss_count() was being called
for bank-0 during the probe and returning 1 instead of 0 indicating that the
context had been lost. This was causing the context restore function to be
called at probe time for this bank and because the context had never been saved,
was restoring an invalid state. This ultimately resulted in a crash [1].

There are multiple bugs here that need to be addressed ...

1. Why the always-on power domain returns a context loss count of 1? This needs
   to be fixed in the power domain code. However, the gpio driver should not
   assume the loss count is 0 to begin with.
2. The omap gpio driver should never be calling get_context_loss_count for a
   gpio bank in a always-on domain. This is pointless and adds unneccessary
   overhead.
3. The OMAP gpio driver assumes that the initial power domain context loss count
   will be 0 at the time the gpio driver is probed. However, it could be
   possible that this is not the case and an invalid context restore could be
   performed during the probe. To avoid this otherwise only populated the
   get_context_loss_count() function pointer after the initial call to
   pm_runtime_get() has occurred. This will ensure that the first
   pm_runtime_put() initialised the loss count correctly.

This patch addresses issues 2 and 3 above.

[1] http://marc.info/?l=linux-omap&m=134065775323775&w=2

Cc: Grant Likely <redacted>
Cc: Linus Walleij <redacted>
Cc: Kevin Hilman <redacted>
Cc: Tarun Kanti DebBarma <redacted>
Cc: Franky Lin <redacted>

Reported-by: Franky Lin <redacted>
Signed-off-by: Jon Hunter <redacted>
---
 drivers/gpio/gpio-omap.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/gpio/gpio-omap.c b/drivers/gpio/gpio-omap.c
index c4ed172..f13fc9c 100644
--- a/drivers/gpio/gpio-omap.c
+++ b/drivers/gpio/gpio-omap.c

@@ -1081,7 +1081,6 @@ static int __devinit omap_gpio_probe(struct platform_device *pdev)
 	bank->is_mpuio = pdata->is_mpuio;
 	bank->non_wakeup_gpios = pdata->non_wakeup_gpios;
 	bank->loses_context = pdata->loses_context;
-	bank->get_context_loss_count = pdata->get_context_loss_count;
 	bank->regs = pdata->regs;
 #ifdef CONFIG_OF_GPIO
 	bank->chip.of_node = of_node_get(node);

@@ -1135,6 +1134,9 @@ static int __devinit omap_gpio_probe(struct platform_device *pdev)
 	omap_gpio_chip_init(bank);
 	omap_gpio_show_rev(bank);
 
+	if (bank->loses_context)
+		bank->get_context_loss_count = pdata->get_context_loss_count;
+
 	pm_runtime_put(bank->dev);
 
 	list_add_tail(&bank->node, &omap_gpio_list);

-- 
1.7.9.5

On 06/29/2012 10:22 AM, Jon Hunter wrote:

Currently the gpio _runtime_resume/suspend functions are calling the
get_context_loss_count() platform function if the function is populated for
a gpio bank. This function is used to determine if the gpio bank logic state
needs to be restored due to a power transition. This function will be populated
for all banks, but it should only be called for banks that have the
"loses_context" variable set. It is pointless to call this if loses_context is
false as we know the context will never be lost and will not need restoring.

For all OMAP2+ devices gpio bank-0 is in an always-on power domain and so will
never lose context. We found that the get_context_loss_count() was being called
for bank-0 during the probe and returning 1 instead of 0 indicating that the
context had been lost. This was causing the context restore function to be
called at probe time for this bank and because the context had never been saved,
was restoring an invalid state. This ultimately resulted in a crash [1].

There are multiple bugs here that need to be addressed ...

1. Why the always-on power domain returns a context loss count of 1? This needs
    to be fixed in the power domain code. However, the gpio driver should not
    assume the loss count is 0 to begin with.
2. The omap gpio driver should never be calling get_context_loss_count for a
    gpio bank in a always-on domain. This is pointless and adds unneccessary
    overhead.
3. The OMAP gpio driver assumes that the initial power domain context loss count
    will be 0 at the time the gpio driver is probed. However, it could be
    possible that this is not the case and an invalid context restore could be
    performed during the probe. To avoid this otherwise only populated the
    get_context_loss_count() function pointer after the initial call to
    pm_runtime_get() has occurred. This will ensure that the first
    pm_runtime_put() initialised the loss count correctly.

This patch addresses issues 2 and 3 above.

[1] http://marc.info/?l=linux-omap&m=134065775323775&w=2

Cc: Grant Likely <redacted>
Cc: Linus Walleij <redacted>
Cc: Kevin Hilman <redacted>
Cc: Tarun Kanti DebBarma <redacted>
Cc: Franky Lin <redacted>

Reported-by: Franky Lin <redacted>
Signed-off-by: Jon Hunter <redacted>
---

Tested-by: Franky Lin <redacted>

On Fri, Jun 29, 2012 at 10:52 PM, Jon Hunter [off-list ref] wrote:

Currently the gpio _runtime_resume/suspend functions are calling the
get_context_loss_count() platform function if the function is populated for
a gpio bank. This function is used to determine if the gpio bank logic state
needs to be restored due to a power transition. This function will be populated
for all banks, but it should only be called for banks that have the
"loses_context" variable set. It is pointless to call this if loses_context is
false as we know the context will never be lost and will not need restoring.

For all OMAP2+ devices gpio bank-0 is in an always-on power domain and so will
never lose context. We found that the get_context_loss_count() was being called
for bank-0 during the probe and returning 1 instead of 0 indicating that the
context had been lost. This was causing the context restore function to be
called at probe time for this bank and because the context had never been saved,
was restoring an invalid state. This ultimately resulted in a crash [1].

There are multiple bugs here that need to be addressed ...

1. Why the always-on power domain returns a context loss count of 1? This needs
? to be fixed in the power domain code. However, the gpio driver should not
? assume the loss count is 0 to begin with.

Indeed. GPIO driver should not assume the value.

2. The omap gpio driver should never be calling get_context_loss_count for a
? gpio bank in a always-on domain. This is pointless and adds unneccessary
? overhead.

Make sense too.

3. The OMAP gpio driver assumes that the initial power domain context loss count
? will be 0 at the time the gpio driver is probed. However, it could be
? possible that this is not the case and an invalid context restore could be
? performed during the probe. To avoid this otherwise only populated the
? get_context_loss_count() function pointer after the initial call to
? pm_runtime_get() has occurred. This will ensure that the first
? pm_runtime_put() initialised the loss count correctly.

This patch addresses issues 2 and 3 above.

[1] http://marc.info/?l=linux-omap&m=134065775323775&w=2

Cc: Grant Likely <redacted>
Cc: Linus Walleij <redacted>
Cc: Kevin Hilman <redacted>
Cc: Tarun Kanti DebBarma <redacted>
Cc: Franky Lin <redacted>

Reported-by: Franky Lin <redacted>
Signed-off-by: Jon Hunter <redacted>
---

Thanks Jon for sorting this out. Patch looks good to me.

Reviewed-by: Santosh Shilimkar <redacted>

* Shilimkar, Santosh [off-list ref] [120629 21:23]:

On Fri, Jun 29, 2012 at 10:52 PM, Jon Hunter [off-list ref] wrote:

quoted

Currently the gpio _runtime_resume/suspend functions are calling the
get_context_loss_count() platform function if the function is populated for
a gpio bank. This function is used to determine if the gpio bank logic state
needs to be restored due to a power transition. This function will be populated
for all banks, but it should only be called for banks that have the
"loses_context" variable set. It is pointless to call this if loses_context is
false as we know the context will never be lost and will not need restoring.

For all OMAP2+ devices gpio bank-0 is in an always-on power domain and so will
never lose context. We found that the get_context_loss_count() was being called
for bank-0 during the probe and returning 1 instead of 0 indicating that the
context had been lost. This was causing the context restore function to be
called at probe time for this bank and because the context had never been saved,
was restoring an invalid state. This ultimately resulted in a crash [1].

There are multiple bugs here that need to be addressed ...

1. Why the always-on power domain returns a context loss count of 1? This needs
? to be fixed in the power domain code. However, the gpio driver should not
? assume the loss count is 0 to begin with.

Indeed. GPIO driver should not assume the value.

quoted

2. The omap gpio driver should never be calling get_context_loss_count for a
? gpio bank in a always-on domain. This is pointless and adds unneccessary
? overhead.

Make sense too.

quoted

3. The OMAP gpio driver assumes that the initial power domain context loss count
? will be 0 at the time the gpio driver is probed. However, it could be
? possible that this is not the case and an invalid context restore could be
? performed during the probe. To avoid this otherwise only populated the
? get_context_loss_count() function pointer after the initial call to
? pm_runtime_get() has occurred. This will ensure that the first
? pm_runtime_put() initialised the loss count correctly.

This patch addresses issues 2 and 3 above.

Should this one be Cc: stable? If this is a regression, then the regression
causing commit should be mentioned.

Tony

On 07/01/2012 03:45 AM, Tony Lindgren wrote:

* Shilimkar, Santosh [off-list ref] [120629 21:23]:

quoted

On Fri, Jun 29, 2012 at 10:52 PM, Jon Hunter [off-list ref] wrote:

quoted

Currently the gpio _runtime_resume/suspend functions are calling the
get_context_loss_count() platform function if the function is populated for
a gpio bank. This function is used to determine if the gpio bank logic state
needs to be restored due to a power transition. This function will be populated
for all banks, but it should only be called for banks that have the
"loses_context" variable set. It is pointless to call this if loses_context is
false as we know the context will never be lost and will not need restoring.

For all OMAP2+ devices gpio bank-0 is in an always-on power domain and so will
never lose context. We found that the get_context_loss_count() was being called
for bank-0 during the probe and returning 1 instead of 0 indicating that the
context had been lost. This was causing the context restore function to be
called at probe time for this bank and because the context had never been saved,
was restoring an invalid state. This ultimately resulted in a crash [1].

There are multiple bugs here that need to be addressed ...

1. Why the always-on power domain returns a context loss count of 1? This needs
  to be fixed in the power domain code. However, the gpio driver should not
  assume the loss count is 0 to begin with.

Indeed. GPIO driver should not assume the value.

quoted

2. The omap gpio driver should never be calling get_context_loss_count for a
  gpio bank in a always-on domain. This is pointless and adds unneccessary
  overhead.

Make sense too.

quoted

3. The OMAP gpio driver assumes that the initial power domain context loss count
  will be 0 at the time the gpio driver is probed. However, it could be
  possible that this is not the case and an invalid context restore could be
  performed during the probe. To avoid this otherwise only populated the
  get_context_loss_count() function pointer after the initial call to
  pm_runtime_get() has occurred. This will ensure that the first
  pm_runtime_put() initialised the loss count correctly.

This patch addresses issues 2 and 3 above.

Should this one be Cc: stable? If this is a regression, then the regression
causing commit should be mentioned.

So that raises a good point. Looking at the stable branch (3.4.4) it is
missing 3 other fixes too [1][2][3]. So this particular problem would
not have been exposed, however, I am wondering if there are other
problems lingering there.

This is a regression is exposed by [2]. I should add that to the changelog.

Cheers
Jon

[1]
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b3c64bc30af67ed328a8d919e41160942b870451
[2]
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=1b1287032df3a69d3ef9a486b444f4ffcca50d01
[3]
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=22770de11cb13e7120f973bca6c800de371a6717