On Sat, Jul 28, 2012 at 12:53:11AM +0200, Rafael J. Wysocki wrote:
> Unfortunately, your commit
>
> commit ca5481c68e9fbcea62bb3c78ae6cccf99ca8fb73
> Author: Paul Mundt <[email protected]>
> Date: Tue Jul 10 12:08:14 2012 +0900
>
> sh: pfc: Rudimentary pinctrl-backed GPIO support.
>
> breaks system suspend on the Mackerel board (.config attached). The system
> simply doesn't suspend and instead it hangs somewhere while suspending
> devices (apparently before running the "late" callbacks).
>
> If the above commit is reverted, system suspend works normally.
On Tue, Jul 31, 2012 at 08:57:02PM -0700, [email protected] wrote:
> gpio: sh7724_pfc handling gpio 0 -> 486
> core: sh7724_pfc support registered
> HW Breakpoints: SH-4A UBC support registered
> autorequest GPIO-53
> ------------[ cut here ]------------
> WARNING: at /opt/usr/src/WORK/morimoto/gitlinux/linux-2.6/drivers/gpio/gpiolib.3
> Modules linked in:
>
> Pid : 1, Comm: swapper
> CPU : 0 Not tainted (3.5.0-rc6+ #1407)
>
> PC is at gpio_ensure_requested+0x30/0x78
> PR is at gpio_ensure_requested+0x30/0x78
Morimoto-san's logs off-list made it clear what happened. Both of these
platforms are going gpio_request() calls at arch_initcall() time which
completely screwed up the ordering of the pfc core. We seem to -ENODEV
out in one place due to missing a pfc pointer initialization elsewhere
resulting in -EPROBE_DEFER from gpiolib.
Turns out we can just collapse the probe/init stuff anyways, so this
ought to fix it. I've verified that it fixes Morimoto-san's issue, my
expectation is that the mackerel case is likewise getting tripped up but
no one bothered implementing any error detecting logic for gpio_request()
failing, so it doesn't fail gracefully.
I'll be pushing this out to Linus shortly:
---
commit 1e32dfe323d156d5d7b25b9feffe015d19713db2
Author: Paul Mundt <[email protected]>
Date: Wed Aug 1 16:27:38 2012 +0900
sh: pfc: Fix up init ordering mess.
Commit ca5481c68e9fbcea62bb3c78ae6cccf99ca8fb73 ("sh: pfc: Rudimentary
pinctrl-backed GPIO support.") introduced a regression for platforms that
were doing early GPIO API calls (from arch_initcall() or earlier),
leading to a situation where our two-stage registration logic would trip
itself up and we'd -ENODEV out of the pinctrl registration path,
resulting in endless -EPROBE_DEFER errors. Further lack of checking any
sort of errors from gpio_request() resulted in boot time warnings,
tripping on the FLAG_REQUESTED test-and-set in gpio_ensure_requested().
As it turns out there's no particular need to bother with the two-stage
registration, as the platform bus is already available at the point that
we have to start caring. As such, it's easiest to simply fold these
together in to a single init path, the ordering of which is ensured
through the platform's mux registration, as usual.
Reported-by: Rafael J. Wysocki <[email protected]>
Reported-by: Kuninori Morimoto <[email protected]>
Signed-off-by: Paul Mundt <[email protected]>
diff --git a/drivers/sh/pfc/pinctrl.c b/drivers/sh/pfc/pinctrl.c
index 814b292..2804eaa 100644
--- a/drivers/sh/pfc/pinctrl.c
+++ b/drivers/sh/pfc/pinctrl.c
@@ -325,20 +325,6 @@ static struct pinctrl_desc sh_pfc_pinctrl_desc = {
.confops = &sh_pfc_pinconf_ops,
};
-int sh_pfc_register_pinctrl(struct sh_pfc *pfc)
-{
- sh_pfc_pmx = kzalloc(sizeof(struct sh_pfc_pinctrl), GFP_KERNEL);
- if (unlikely(!sh_pfc_pmx))
- return -ENOMEM;
-
- spin_lock_init(&sh_pfc_pmx->lock);
-
- sh_pfc_pmx->pfc = pfc;
-
- return 0;
-}
-EXPORT_SYMBOL_GPL(sh_pfc_register_pinctrl);
-
static inline void __devinit sh_pfc_map_one_gpio(struct sh_pfc *pfc,
struct sh_pfc_pinctrl *pmx,
struct pinmux_gpio *gpio,
@@ -505,7 +491,7 @@ static struct platform_device sh_pfc_pinctrl_device = {
.id = -1,
};
-static int __init sh_pfc_pinctrl_init(void)
+static int sh_pfc_pinctrl_init(void)
{
int rc;
@@ -519,10 +505,22 @@ static int __init sh_pfc_pinctrl_init(void)
return rc;
}
+int sh_pfc_register_pinctrl(struct sh_pfc *pfc)
+{
+ sh_pfc_pmx = kzalloc(sizeof(struct sh_pfc_pinctrl), GFP_KERNEL);
+ if (unlikely(!sh_pfc_pmx))
+ return -ENOMEM;
+
+ spin_lock_init(&sh_pfc_pmx->lock);
+
+ sh_pfc_pmx->pfc = pfc;
+
+ return sh_pfc_pinctrl_init();
+}
+EXPORT_SYMBOL_GPL(sh_pfc_register_pinctrl);
+
static void __exit sh_pfc_pinctrl_exit(void)
{
platform_driver_unregister(&sh_pfc_pinctrl_driver);
}
-
-subsys_initcall(sh_pfc_pinctrl_init);
module_exit(sh_pfc_pinctrl_exit);
On Wednesday, August 01, 2012, Paul Mundt wrote:
> On Sat, Jul 28, 2012 at 12:53:11AM +0200, Rafael J. Wysocki wrote:
> > Unfortunately, your commit
> >
> > commit ca5481c68e9fbcea62bb3c78ae6cccf99ca8fb73
> > Author: Paul Mundt <[email protected]>
> > Date: Tue Jul 10 12:08:14 2012 +0900
> >
> > sh: pfc: Rudimentary pinctrl-backed GPIO support.
> >
> > breaks system suspend on the Mackerel board (.config attached). The system
> > simply doesn't suspend and instead it hangs somewhere while suspending
> > devices (apparently before running the "late" callbacks).
> >
> > If the above commit is reverted, system suspend works normally.
>
> On Tue, Jul 31, 2012 at 08:57:02PM -0700, [email protected] wrote:
> > gpio: sh7724_pfc handling gpio 0 -> 486
> > core: sh7724_pfc support registered
> > HW Breakpoints: SH-4A UBC support registered
> > autorequest GPIO-53
> > ------------[ cut here ]------------
> > WARNING: at /opt/usr/src/WORK/morimoto/gitlinux/linux-2.6/drivers/gpio/gpiolib.3
> > Modules linked in:
> >
> > Pid : 1, Comm: swapper
> > CPU : 0 Not tainted (3.5.0-rc6+ #1407)
> >
> > PC is at gpio_ensure_requested+0x30/0x78
> > PR is at gpio_ensure_requested+0x30/0x78
>
> Morimoto-san's logs off-list made it clear what happened. Both of these
> platforms are going gpio_request() calls at arch_initcall() time which
> completely screwed up the ordering of the pfc core. We seem to -ENODEV
> out in one place due to missing a pfc pointer initialization elsewhere
> resulting in -EPROBE_DEFER from gpiolib.
>
> Turns out we can just collapse the probe/init stuff anyways, so this
> ought to fix it. I've verified that it fixes Morimoto-san's issue, my
> expectation is that the mackerel case is likewise getting tripped up but
> no one bothered implementing any error detecting logic for gpio_request()
> failing, so it doesn't fail gracefully.
>
> I'll be pushing this out to Linus shortly:
Thanks, this helped. Resume works correctly on Mackerel with 3.6-rc1.
However, I'm now seeing a different problem related to system suspend on that,
board which is that sh7372_enter_a3sm_common() returns immediately, as though
at least one of the wakeup signals was permanently asserted. This hadn't
happened before your last pull request was merged, so I suspect that one
of the irqdomain patches might introduce this behavior.
Thanks,
Rafael
On Sun, Aug 05, 2012 at 12:02:43AM +0200, Rafael J. Wysocki wrote:
> On Wednesday, August 01, 2012, Paul Mundt wrote:
> > Turns out we can just collapse the probe/init stuff anyways, so this
> > ought to fix it. I've verified that it fixes Morimoto-san's issue, my
> > expectation is that the mackerel case is likewise getting tripped up but
> > no one bothered implementing any error detecting logic for gpio_request()
> > failing, so it doesn't fail gracefully.
> >
> > I'll be pushing this out to Linus shortly:
>
> Thanks, this helped. Resume works correctly on Mackerel with 3.6-rc1.
>
> However, I'm now seeing a different problem related to system suspend on that,
> board which is that sh7372_enter_a3sm_common() returns immediately, as though
> at least one of the wakeup signals was permanently asserted. This hadn't
> happened before your last pull request was merged, so I suspect that one
> of the irqdomain patches might introduce this behavior.
>
Ok, I'll back off the irqdomain selection until we have a chance to test
it more on the other platforms.
On Tuesday, August 07, 2012, Paul Mundt wrote:
> On Sun, Aug 05, 2012 at 12:02:43AM +0200, Rafael J. Wysocki wrote:
> > On Wednesday, August 01, 2012, Paul Mundt wrote:
> > > Turns out we can just collapse the probe/init stuff anyways, so this
> > > ought to fix it. I've verified that it fixes Morimoto-san's issue, my
> > > expectation is that the mackerel case is likewise getting tripped up but
> > > no one bothered implementing any error detecting logic for gpio_request()
> > > failing, so it doesn't fail gracefully.
> > >
> > > I'll be pushing this out to Linus shortly:
> >
> > Thanks, this helped. Resume works correctly on Mackerel with 3.6-rc1.
> >
> > However, I'm now seeing a different problem related to system suspend on that,
> > board which is that sh7372_enter_a3sm_common() returns immediately, as though
> > at least one of the wakeup signals was permanently asserted. This hadn't
> > happened before your last pull request was merged, so I suspect that one
> > of the irqdomain patches might introduce this behavior.
> >
> Ok, I'll back off the irqdomain selection until we have a chance to test
> it more on the other platforms.
If you mean removing the select from drivers/sh/intc/Kconfig, that alone doesn't
help. Reverting commit 1d6a21b0a672fb29b01ccf397d478e0541e17716
(sh: intc: initial irqdomain support.) helps, though.
Thanks,
Rafael
On Wed, Aug 08, 2012 at 11:23:04AM +0200, Rafael J. Wysocki wrote:
> On Tuesday, August 07, 2012, Paul Mundt wrote:
> > On Sun, Aug 05, 2012 at 12:02:43AM +0200, Rafael J. Wysocki wrote:
> > > On Wednesday, August 01, 2012, Paul Mundt wrote:
> > > > Turns out we can just collapse the probe/init stuff anyways, so this
> > > > ought to fix it. I've verified that it fixes Morimoto-san's issue, my
> > > > expectation is that the mackerel case is likewise getting tripped up but
> > > > no one bothered implementing any error detecting logic for gpio_request()
> > > > failing, so it doesn't fail gracefully.
> > > >
> > > > I'll be pushing this out to Linus shortly:
> > >
> > > Thanks, this helped. Resume works correctly on Mackerel with 3.6-rc1.
> > >
> > > However, I'm now seeing a different problem related to system suspend on that,
> > > board which is that sh7372_enter_a3sm_common() returns immediately, as though
> > > at least one of the wakeup signals was permanently asserted. This hadn't
> > > happened before your last pull request was merged, so I suspect that one
> > > of the irqdomain patches might introduce this behavior.
> > >
> > Ok, I'll back off the irqdomain selection until we have a chance to test
> > it more on the other platforms.
>
> If you mean removing the select from drivers/sh/intc/Kconfig, that alone doesn't
> help. Reverting commit 1d6a21b0a672fb29b01ccf397d478e0541e17716
> (sh: intc: initial irqdomain support.) helps, though.
>
Yeah, I see what happened now. I inadvertently trampled the -EEXIST
behaviour. I've taken care of it now, so hopefully you'll have better
luck with -rc2!
On Thursday, August 09, 2012, Paul Mundt wrote:
> On Wed, Aug 08, 2012 at 11:23:04AM +0200, Rafael J. Wysocki wrote:
> > On Tuesday, August 07, 2012, Paul Mundt wrote:
> > > On Sun, Aug 05, 2012 at 12:02:43AM +0200, Rafael J. Wysocki wrote:
> > > > On Wednesday, August 01, 2012, Paul Mundt wrote:
> > > > > Turns out we can just collapse the probe/init stuff anyways, so this
> > > > > ought to fix it. I've verified that it fixes Morimoto-san's issue, my
> > > > > expectation is that the mackerel case is likewise getting tripped up but
> > > > > no one bothered implementing any error detecting logic for gpio_request()
> > > > > failing, so it doesn't fail gracefully.
> > > > >
> > > > > I'll be pushing this out to Linus shortly:
> > > >
> > > > Thanks, this helped. Resume works correctly on Mackerel with 3.6-rc1.
> > > >
> > > > However, I'm now seeing a different problem related to system suspend on that,
> > > > board which is that sh7372_enter_a3sm_common() returns immediately, as though
> > > > at least one of the wakeup signals was permanently asserted. This hadn't
> > > > happened before your last pull request was merged, so I suspect that one
> > > > of the irqdomain patches might introduce this behavior.
> > > >
> > > Ok, I'll back off the irqdomain selection until we have a chance to test
> > > it more on the other platforms.
> >
> > If you mean removing the select from drivers/sh/intc/Kconfig, that alone doesn't
> > help. Reverting commit 1d6a21b0a672fb29b01ccf397d478e0541e17716
> > (sh: intc: initial irqdomain support.) helps, though.
> >
> Yeah, I see what happened now. I inadvertently trampled the -EEXIST
> behaviour. I've taken care of it now, so hopefully you'll have better
> luck with -rc2!
I tested your "sh: intc: Handle domain association for sparseirq pre-allocated
vectors" patch in the meantime and it does fix the issue.
Thanks,
Rafael