From: Rafael J. Wysocki <[email protected]>
Because acpi_install_fixed_event_handler() enables the event
automatically on success, it is incorrect to call it before the
handler routine passed to it is ready to handle events.
Unfortunately, the rtc-cmos driver does exactly the incorrect thing
by calling cmos_wake_setup(), which passes rtc_handler() to
acpi_install_fixed_event_handler(), before cmos_do_probe(), because
rtc_handler() uses dev_get_drvdata() to get to the cmos object
pointer and the driver data pointer is only populated in
cmos_do_probe().
This leads to a NULL pointer dereference in rtc_handler() on boot
if the RTC fixed event happens to be active at the init time.
To address this issue, change the initialization ordering of the
driver so that cmos_wake_setup() is always called after a successful
cmos_do_probe() call.
While at it, change cmos_pnp_probe() to call cmos_do_probe() after
the initial if () statement used for computing the IRQ argument to
be passed to cmos_do_probe() which is cleaner than calling it in
each branch of that if () (local variable "irq" can be of type int,
because it is passed to that function as an argument of type int).
Note that commit 6492fed7d8c9 ("rtc: rtc-cmos: Do not check
ACPI_FADT_LOW_POWER_S0") caused this issue to affect a larger number
of systems, because previously it only affected systems with
ACPI_FADT_LOW_POWER_S0 set, but it is present regardless of that
commit.
Fixes: 6492fed7d8c9 ("rtc: rtc-cmos: Do not check ACPI_FADT_LOW_POWER_S0")
Fixes: a474aaedac99 ("rtc-cmos: move wake setup from ACPI glue into RTC driver")
Link: https://lore.kernel.org/linux-acpi/[email protected]/
Reported-by: Mel Gorman <[email protected]>
Signed-off-by: Rafael J. Wysocki <[email protected]>
---
drivers/rtc/rtc-cmos.c | 29 +++++++++++++++++++----------
1 file changed, 19 insertions(+), 10 deletions(-)
Index: linux-pm/drivers/rtc/rtc-cmos.c
===================================================================
--- linux-pm.orig/drivers/rtc/rtc-cmos.c
+++ linux-pm/drivers/rtc/rtc-cmos.c
@@ -1352,10 +1352,10 @@ static void cmos_check_acpi_rtc_status(s
static int cmos_pnp_probe(struct pnp_dev *pnp, const struct pnp_device_id *id)
{
- cmos_wake_setup(&pnp->dev);
+ int irq, ret;
if (pnp_port_start(pnp, 0) == 0x70 && !pnp_irq_valid(pnp, 0)) {
- unsigned int irq = 0;
+ irq = 0;
#ifdef CONFIG_X86
/* Some machines contain a PNP entry for the RTC, but
* don't define the IRQ. It should always be safe to
@@ -1364,13 +1364,17 @@ static int cmos_pnp_probe(struct pnp_dev
if (nr_legacy_irqs())
irq = RTC_IRQ;
#endif
- return cmos_do_probe(&pnp->dev,
- pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
} else {
- return cmos_do_probe(&pnp->dev,
- pnp_get_resource(pnp, IORESOURCE_IO, 0),
- pnp_irq(pnp, 0));
+ irq = pnp_irq(pnp, 0);
}
+
+ ret = cmos_do_probe(&pnp->dev, pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
+ if (ret)
+ return ret;
+
+ cmos_wake_setup(&pnp->dev);
+
+ return 0;
}
static void cmos_pnp_remove(struct pnp_dev *pnp)
@@ -1454,10 +1458,9 @@ static inline void cmos_of_init(struct p
static int __init cmos_platform_probe(struct platform_device *pdev)
{
struct resource *resource;
- int irq;
+ int irq, ret;
cmos_of_init(pdev);
- cmos_wake_setup(&pdev->dev);
if (RTC_IOMAPPED)
resource = platform_get_resource(pdev, IORESOURCE_IO, 0);
@@ -1467,7 +1470,13 @@ static int __init cmos_platform_probe(st
if (irq < 0)
irq = -1;
- return cmos_do_probe(&pdev->dev, resource, irq);
+ ret = cmos_do_probe(&pdev->dev, resource, irq);
+ if (ret)
+ return ret;
+
+ cmos_wake_setup(&pdev->dev);
+
+ return 0;
}
static int cmos_platform_remove(struct platform_device *pdev)
On Wed, Oct 12, 2022 at 08:07:01PM +0200, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <[email protected]>
>
> Because acpi_install_fixed_event_handler() enables the event
> automatically on success, it is incorrect to call it before the
> handler routine passed to it is ready to handle events.
>
> Unfortunately, the rtc-cmos driver does exactly the incorrect thing
> by calling cmos_wake_setup(), which passes rtc_handler() to
> acpi_install_fixed_event_handler(), before cmos_do_probe(), because
> rtc_handler() uses dev_get_drvdata() to get to the cmos object
> pointer and the driver data pointer is only populated in
> cmos_do_probe().
>
> This leads to a NULL pointer dereference in rtc_handler() on boot
> if the RTC fixed event happens to be active at the init time.
>
> To address this issue, change the initialization ordering of the
> driver so that cmos_wake_setup() is always called after a successful
> cmos_do_probe() call.
>
> While at it, change cmos_pnp_probe() to call cmos_do_probe() after
> the initial if () statement used for computing the IRQ argument to
> be passed to cmos_do_probe() which is cleaner than calling it in
> each branch of that if () (local variable "irq" can be of type int,
> because it is passed to that function as an argument of type int).
>
> Note that commit 6492fed7d8c9 ("rtc: rtc-cmos: Do not check
> ACPI_FADT_LOW_POWER_S0") caused this issue to affect a larger number
> of systems, because previously it only affected systems with
> ACPI_FADT_LOW_POWER_S0 set, but it is present regardless of that
> commit.
>
> Fixes: 6492fed7d8c9 ("rtc: rtc-cmos: Do not check ACPI_FADT_LOW_POWER_S0")
> Fixes: a474aaedac99 ("rtc-cmos: move wake setup from ACPI glue into RTC driver")
> Link: https://lore.kernel.org/linux-acpi/[email protected]/
> Reported-by: Mel Gorman <[email protected]>
> Signed-off-by: Rafael J. Wysocki <[email protected]>
Works for me so;
Tested-by: Mel Gorman <[email protected]>
--
Mel Gorman
SUSE Labs
On Wed, Oct 12, 2022 at 08:07:01PM +0200, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <[email protected]>
>
> Because acpi_install_fixed_event_handler() enables the event
> automatically on success, it is incorrect to call it before the
> handler routine passed to it is ready to handle events.
>
> Unfortunately, the rtc-cmos driver does exactly the incorrect thing
> by calling cmos_wake_setup(), which passes rtc_handler() to
> acpi_install_fixed_event_handler(), before cmos_do_probe(), because
> rtc_handler() uses dev_get_drvdata() to get to the cmos object
> pointer and the driver data pointer is only populated in
> cmos_do_probe().
>
> This leads to a NULL pointer dereference in rtc_handler() on boot
> if the RTC fixed event happens to be active at the init time.
>
> To address this issue, change the initialization ordering of the
> driver so that cmos_wake_setup() is always called after a successful
> cmos_do_probe() call.
>
> While at it, change cmos_pnp_probe() to call cmos_do_probe() after
> the initial if () statement used for computing the IRQ argument to
> be passed to cmos_do_probe() which is cleaner than calling it in
> each branch of that if () (local variable "irq" can be of type int,
> because it is passed to that function as an argument of type int).
>
> Note that commit 6492fed7d8c9 ("rtc: rtc-cmos: Do not check
> ACPI_FADT_LOW_POWER_S0") caused this issue to affect a larger number
> of systems, because previously it only affected systems with
> ACPI_FADT_LOW_POWER_S0 set, but it is present regardless of that
> commit.
>
> Fixes: 6492fed7d8c9 ("rtc: rtc-cmos: Do not check ACPI_FADT_LOW_POWER_S0")
> Fixes: a474aaedac99 ("rtc-cmos: move wake setup from ACPI glue into RTC driver")
> Link: https://lore.kernel.org/linux-acpi/[email protected]/
> Reported-by: Mel Gorman <[email protected]>
> Signed-off-by: Rafael J. Wysocki <[email protected]>
Reviewed-by: Bjorn Helgaas <[email protected]>
Yep, I blew it with a474aaedac99, sorry about that.
Possibly could call cmos_wake_setup() from cmos_do_probe() instead of
from cmos_pnp_probe() and cmos_platform_probe()? Then there would be
a single call site and it would be closer to the actual dependency on
dev_set_drvdata(). Either way is fine with me.
Unrelated, but I happened to notice that pnp_irq() returns -1 for
failure, and this note suggests that possibly returning 0 would be
better:
https://lore.kernel.org/r/CAHk-=wg2Pkb9kbfbstbB91AJA2SF6cySbsgHG-iQMq56j3VTcA@mail.gmail.com
> ---
> drivers/rtc/rtc-cmos.c | 29 +++++++++++++++++++----------
> 1 file changed, 19 insertions(+), 10 deletions(-)
>
> Index: linux-pm/drivers/rtc/rtc-cmos.c
> ===================================================================
> --- linux-pm.orig/drivers/rtc/rtc-cmos.c
> +++ linux-pm/drivers/rtc/rtc-cmos.c
> @@ -1352,10 +1352,10 @@ static void cmos_check_acpi_rtc_status(s
>
> static int cmos_pnp_probe(struct pnp_dev *pnp, const struct pnp_device_id *id)
> {
> - cmos_wake_setup(&pnp->dev);
> + int irq, ret;
>
> if (pnp_port_start(pnp, 0) == 0x70 && !pnp_irq_valid(pnp, 0)) {
> - unsigned int irq = 0;
> + irq = 0;
> #ifdef CONFIG_X86
> /* Some machines contain a PNP entry for the RTC, but
> * don't define the IRQ. It should always be safe to
> @@ -1364,13 +1364,17 @@ static int cmos_pnp_probe(struct pnp_dev
> if (nr_legacy_irqs())
> irq = RTC_IRQ;
> #endif
> - return cmos_do_probe(&pnp->dev,
> - pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
> } else {
> - return cmos_do_probe(&pnp->dev,
> - pnp_get_resource(pnp, IORESOURCE_IO, 0),
> - pnp_irq(pnp, 0));
> + irq = pnp_irq(pnp, 0);
> }
> +
> + ret = cmos_do_probe(&pnp->dev, pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
> + if (ret)
> + return ret;
> +
> + cmos_wake_setup(&pnp->dev);
> +
> + return 0;
> }
>
> static void cmos_pnp_remove(struct pnp_dev *pnp)
> @@ -1454,10 +1458,9 @@ static inline void cmos_of_init(struct p
> static int __init cmos_platform_probe(struct platform_device *pdev)
> {
> struct resource *resource;
> - int irq;
> + int irq, ret;
>
> cmos_of_init(pdev);
> - cmos_wake_setup(&pdev->dev);
>
> if (RTC_IOMAPPED)
> resource = platform_get_resource(pdev, IORESOURCE_IO, 0);
> @@ -1467,7 +1470,13 @@ static int __init cmos_platform_probe(st
> if (irq < 0)
> irq = -1;
>
> - return cmos_do_probe(&pdev->dev, resource, irq);
> + ret = cmos_do_probe(&pdev->dev, resource, irq);
> + if (ret)
> + return ret;
> +
> + cmos_wake_setup(&pdev->dev);
> +
> + return 0;
> }
>
> static int cmos_platform_remove(struct platform_device *pdev)
>
>
>
On Wed, Oct 12, 2022 at 11:00 PM Bjorn Helgaas <[email protected]> wrote:
>
> On Wed, Oct 12, 2022 at 08:07:01PM +0200, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <[email protected]>
> >
> > Because acpi_install_fixed_event_handler() enables the event
> > automatically on success, it is incorrect to call it before the
> > handler routine passed to it is ready to handle events.
> >
> > Unfortunately, the rtc-cmos driver does exactly the incorrect thing
> > by calling cmos_wake_setup(), which passes rtc_handler() to
> > acpi_install_fixed_event_handler(), before cmos_do_probe(), because
> > rtc_handler() uses dev_get_drvdata() to get to the cmos object
> > pointer and the driver data pointer is only populated in
> > cmos_do_probe().
> >
> > This leads to a NULL pointer dereference in rtc_handler() on boot
> > if the RTC fixed event happens to be active at the init time.
> >
> > To address this issue, change the initialization ordering of the
> > driver so that cmos_wake_setup() is always called after a successful
> > cmos_do_probe() call.
> >
> > While at it, change cmos_pnp_probe() to call cmos_do_probe() after
> > the initial if () statement used for computing the IRQ argument to
> > be passed to cmos_do_probe() which is cleaner than calling it in
> > each branch of that if () (local variable "irq" can be of type int,
> > because it is passed to that function as an argument of type int).
> >
> > Note that commit 6492fed7d8c9 ("rtc: rtc-cmos: Do not check
> > ACPI_FADT_LOW_POWER_S0") caused this issue to affect a larger number
> > of systems, because previously it only affected systems with
> > ACPI_FADT_LOW_POWER_S0 set, but it is present regardless of that
> > commit.
> >
> > Fixes: 6492fed7d8c9 ("rtc: rtc-cmos: Do not check ACPI_FADT_LOW_POWER_S0")
> > Fixes: a474aaedac99 ("rtc-cmos: move wake setup from ACPI glue into RTC driver")
> > Link: https://lore.kernel.org/linux-acpi/[email protected]/
> > Reported-by: Mel Gorman <[email protected]>
> > Signed-off-by: Rafael J. Wysocki <[email protected]>
>
> Reviewed-by: Bjorn Helgaas <[email protected]>
>
> Yep, I blew it with a474aaedac99, sorry about that.
>
> Possibly could call cmos_wake_setup() from cmos_do_probe() instead of
> from cmos_pnp_probe() and cmos_platform_probe()?
Sounds good.
I would prefer to send a separate patch for this on top of the
$subject one, unless Alexandre wants me to do it all in one go.
Alexandre, what's your preference here? Or would you prefer if I
pushed this forward?
> Then there would be a single call site and it would be closer to the actual dependency on
> dev_set_drvdata(). Either way is fine with me.
OK
> Unrelated, but I happened to notice that pnp_irq() returns -1 for
> failure, and this note suggests that possibly returning 0 would be
> better:
>
> https://lore.kernel.org/r/CAHk-=wg2Pkb9kbfbstbB91AJA2SF6cySbsgHG-iQMq56j3VTcA@mail.gmail.com
Probably.
In that case, though, it would be prudent to also explicitly discard
IRQ resources where start is equal to 0.
>
> > ---
> > drivers/rtc/rtc-cmos.c | 29 +++++++++++++++++++----------
> > 1 file changed, 19 insertions(+), 10 deletions(-)
> >
> > Index: linux-pm/drivers/rtc/rtc-cmos.c
> > ===================================================================
> > --- linux-pm.orig/drivers/rtc/rtc-cmos.c
> > +++ linux-pm/drivers/rtc/rtc-cmos.c
> > @@ -1352,10 +1352,10 @@ static void cmos_check_acpi_rtc_status(s
> >
> > static int cmos_pnp_probe(struct pnp_dev *pnp, const struct pnp_device_id *id)
> > {
> > - cmos_wake_setup(&pnp->dev);
> > + int irq, ret;
> >
> > if (pnp_port_start(pnp, 0) == 0x70 && !pnp_irq_valid(pnp, 0)) {
> > - unsigned int irq = 0;
> > + irq = 0;
> > #ifdef CONFIG_X86
> > /* Some machines contain a PNP entry for the RTC, but
> > * don't define the IRQ. It should always be safe to
> > @@ -1364,13 +1364,17 @@ static int cmos_pnp_probe(struct pnp_dev
> > if (nr_legacy_irqs())
> > irq = RTC_IRQ;
> > #endif
> > - return cmos_do_probe(&pnp->dev,
> > - pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
> > } else {
> > - return cmos_do_probe(&pnp->dev,
> > - pnp_get_resource(pnp, IORESOURCE_IO, 0),
> > - pnp_irq(pnp, 0));
> > + irq = pnp_irq(pnp, 0);
> > }
> > +
> > + ret = cmos_do_probe(&pnp->dev, pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
> > + if (ret)
> > + return ret;
> > +
> > + cmos_wake_setup(&pnp->dev);
> > +
> > + return 0;
> > }
> >
> > static void cmos_pnp_remove(struct pnp_dev *pnp)
> > @@ -1454,10 +1458,9 @@ static inline void cmos_of_init(struct p
> > static int __init cmos_platform_probe(struct platform_device *pdev)
> > {
> > struct resource *resource;
> > - int irq;
> > + int irq, ret;
> >
> > cmos_of_init(pdev);
> > - cmos_wake_setup(&pdev->dev);
> >
> > if (RTC_IOMAPPED)
> > resource = platform_get_resource(pdev, IORESOURCE_IO, 0);
> > @@ -1467,7 +1470,13 @@ static int __init cmos_platform_probe(st
> > if (irq < 0)
> > irq = -1;
> >
> > - return cmos_do_probe(&pdev->dev, resource, irq);
> > + ret = cmos_do_probe(&pdev->dev, resource, irq);
> > + if (ret)
> > + return ret;
> > +
> > + cmos_wake_setup(&pdev->dev);
> > +
> > + return 0;
> > }
> >
> > static int cmos_platform_remove(struct platform_device *pdev)
> >
> >
> >
On 13/10/2022 13:38:31+0200, Rafael J. Wysocki wrote:
> On Wed, Oct 12, 2022 at 11:00 PM Bjorn Helgaas <[email protected]> wrote:
> >
> > On Wed, Oct 12, 2022 at 08:07:01PM +0200, Rafael J. Wysocki wrote:
> > > From: Rafael J. Wysocki <[email protected]>
> > >
> > > Because acpi_install_fixed_event_handler() enables the event
> > > automatically on success, it is incorrect to call it before the
> > > handler routine passed to it is ready to handle events.
> > >
> > > Unfortunately, the rtc-cmos driver does exactly the incorrect thing
> > > by calling cmos_wake_setup(), which passes rtc_handler() to
> > > acpi_install_fixed_event_handler(), before cmos_do_probe(), because
> > > rtc_handler() uses dev_get_drvdata() to get to the cmos object
> > > pointer and the driver data pointer is only populated in
> > > cmos_do_probe().
> > >
> > > This leads to a NULL pointer dereference in rtc_handler() on boot
> > > if the RTC fixed event happens to be active at the init time.
> > >
> > > To address this issue, change the initialization ordering of the
> > > driver so that cmos_wake_setup() is always called after a successful
> > > cmos_do_probe() call.
> > >
> > > While at it, change cmos_pnp_probe() to call cmos_do_probe() after
> > > the initial if () statement used for computing the IRQ argument to
> > > be passed to cmos_do_probe() which is cleaner than calling it in
> > > each branch of that if () (local variable "irq" can be of type int,
> > > because it is passed to that function as an argument of type int).
> > >
> > > Note that commit 6492fed7d8c9 ("rtc: rtc-cmos: Do not check
> > > ACPI_FADT_LOW_POWER_S0") caused this issue to affect a larger number
> > > of systems, because previously it only affected systems with
> > > ACPI_FADT_LOW_POWER_S0 set, but it is present regardless of that
> > > commit.
> > >
> > > Fixes: 6492fed7d8c9 ("rtc: rtc-cmos: Do not check ACPI_FADT_LOW_POWER_S0")
> > > Fixes: a474aaedac99 ("rtc-cmos: move wake setup from ACPI glue into RTC driver")
> > > Link: https://lore.kernel.org/linux-acpi/[email protected]/
> > > Reported-by: Mel Gorman <[email protected]>
> > > Signed-off-by: Rafael J. Wysocki <[email protected]>
> >
> > Reviewed-by: Bjorn Helgaas <[email protected]>
> >
> > Yep, I blew it with a474aaedac99, sorry about that.
> >
> > Possibly could call cmos_wake_setup() from cmos_do_probe() instead of
> > from cmos_pnp_probe() and cmos_platform_probe()?
>
> Sounds good.
>
> I would prefer to send a separate patch for this on top of the
> $subject one, unless Alexandre wants me to do it all in one go.
>
> Alexandre, what's your preference here? Or would you prefer if I
> pushed this forward?
>
I applied your patch, feel free to improve on top of that ;)
> > Then there would be a single call site and it would be closer to the actual dependency on
> > dev_set_drvdata(). Either way is fine with me.
>
> OK
>
> > Unrelated, but I happened to notice that pnp_irq() returns -1 for
> > failure, and this note suggests that possibly returning 0 would be
> > better:
> >
> > https://lore.kernel.org/r/CAHk-=wg2Pkb9kbfbstbB91AJA2SF6cySbsgHG-iQMq56j3VTcA@mail.gmail.com
>
> Probably.
>
> In that case, though, it would be prudent to also explicitly discard
> IRQ resources where start is equal to 0.
>
> >
> > > ---
> > > drivers/rtc/rtc-cmos.c | 29 +++++++++++++++++++----------
> > > 1 file changed, 19 insertions(+), 10 deletions(-)
> > >
> > > Index: linux-pm/drivers/rtc/rtc-cmos.c
> > > ===================================================================
> > > --- linux-pm.orig/drivers/rtc/rtc-cmos.c
> > > +++ linux-pm/drivers/rtc/rtc-cmos.c
> > > @@ -1352,10 +1352,10 @@ static void cmos_check_acpi_rtc_status(s
> > >
> > > static int cmos_pnp_probe(struct pnp_dev *pnp, const struct pnp_device_id *id)
> > > {
> > > - cmos_wake_setup(&pnp->dev);
> > > + int irq, ret;
> > >
> > > if (pnp_port_start(pnp, 0) == 0x70 && !pnp_irq_valid(pnp, 0)) {
> > > - unsigned int irq = 0;
> > > + irq = 0;
> > > #ifdef CONFIG_X86
> > > /* Some machines contain a PNP entry for the RTC, but
> > > * don't define the IRQ. It should always be safe to
> > > @@ -1364,13 +1364,17 @@ static int cmos_pnp_probe(struct pnp_dev
> > > if (nr_legacy_irqs())
> > > irq = RTC_IRQ;
> > > #endif
> > > - return cmos_do_probe(&pnp->dev,
> > > - pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
> > > } else {
> > > - return cmos_do_probe(&pnp->dev,
> > > - pnp_get_resource(pnp, IORESOURCE_IO, 0),
> > > - pnp_irq(pnp, 0));
> > > + irq = pnp_irq(pnp, 0);
> > > }
> > > +
> > > + ret = cmos_do_probe(&pnp->dev, pnp_get_resource(pnp, IORESOURCE_IO, 0), irq);
> > > + if (ret)
> > > + return ret;
> > > +
> > > + cmos_wake_setup(&pnp->dev);
> > > +
> > > + return 0;
> > > }
> > >
> > > static void cmos_pnp_remove(struct pnp_dev *pnp)
> > > @@ -1454,10 +1458,9 @@ static inline void cmos_of_init(struct p
> > > static int __init cmos_platform_probe(struct platform_device *pdev)
> > > {
> > > struct resource *resource;
> > > - int irq;
> > > + int irq, ret;
> > >
> > > cmos_of_init(pdev);
> > > - cmos_wake_setup(&pdev->dev);
> > >
> > > if (RTC_IOMAPPED)
> > > resource = platform_get_resource(pdev, IORESOURCE_IO, 0);
> > > @@ -1467,7 +1470,13 @@ static int __init cmos_platform_probe(st
> > > if (irq < 0)
> > > irq = -1;
> > >
> > > - return cmos_do_probe(&pdev->dev, resource, irq);
> > > + ret = cmos_do_probe(&pdev->dev, resource, irq);
> > > + if (ret)
> > > + return ret;
> > > +
> > > + cmos_wake_setup(&pdev->dev);
> > > +
> > > + return 0;
> > > }
> > >
> > > static int cmos_platform_remove(struct platform_device *pdev)
> > >
> > >
> > >
--
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
On Wed, 12 Oct 2022 20:07:01 +0200, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <[email protected]>
>
> Because acpi_install_fixed_event_handler() enables the event
> automatically on success, it is incorrect to call it before the
> handler routine passed to it is ready to handle events.
>
> Unfortunately, the rtc-cmos driver does exactly the incorrect thing
> by calling cmos_wake_setup(), which passes rtc_handler() to
> acpi_install_fixed_event_handler(), before cmos_do_probe(), because
> rtc_handler() uses dev_get_drvdata() to get to the cmos object
> pointer and the driver data pointer is only populated in
> cmos_do_probe().
>
> [...]
Applied, thanks!
[1/1] rtc: rtc-cmos: Fix event handler registration ordering issue
commit: 4919d3eb2ec0ee364f7e3cf2d99646c1b224fae8
Best regards,
--
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
On Thu, Oct 13, 2022 at 11:34 PM Alexandre Belloni
<[email protected]> wrote:
>
> On 13/10/2022 13:38:31+0200, Rafael J. Wysocki wrote:
> > On Wed, Oct 12, 2022 at 11:00 PM Bjorn Helgaas <[email protected]> wrote:
> > >
> > > On Wed, Oct 12, 2022 at 08:07:01PM +0200, Rafael J. Wysocki wrote:
> > > > From: Rafael J. Wysocki <[email protected]>
> > > >
> > > > Because acpi_install_fixed_event_handler() enables the event
> > > > automatically on success, it is incorrect to call it before the
> > > > handler routine passed to it is ready to handle events.
> > > >
> > > > Unfortunately, the rtc-cmos driver does exactly the incorrect thing
> > > > by calling cmos_wake_setup(), which passes rtc_handler() to
> > > > acpi_install_fixed_event_handler(), before cmos_do_probe(), because
> > > > rtc_handler() uses dev_get_drvdata() to get to the cmos object
> > > > pointer and the driver data pointer is only populated in
> > > > cmos_do_probe().
> > > >
> > > > This leads to a NULL pointer dereference in rtc_handler() on boot
> > > > if the RTC fixed event happens to be active at the init time.
> > > >
> > > > To address this issue, change the initialization ordering of the
> > > > driver so that cmos_wake_setup() is always called after a successful
> > > > cmos_do_probe() call.
> > > >
> > > > While at it, change cmos_pnp_probe() to call cmos_do_probe() after
> > > > the initial if () statement used for computing the IRQ argument to
> > > > be passed to cmos_do_probe() which is cleaner than calling it in
> > > > each branch of that if () (local variable "irq" can be of type int,
> > > > because it is passed to that function as an argument of type int).
> > > >
> > > > Note that commit 6492fed7d8c9 ("rtc: rtc-cmos: Do not check
> > > > ACPI_FADT_LOW_POWER_S0") caused this issue to affect a larger number
> > > > of systems, because previously it only affected systems with
> > > > ACPI_FADT_LOW_POWER_S0 set, but it is present regardless of that
> > > > commit.
> > > >
> > > > Fixes: 6492fed7d8c9 ("rtc: rtc-cmos: Do not check ACPI_FADT_LOW_POWER_S0")
> > > > Fixes: a474aaedac99 ("rtc-cmos: move wake setup from ACPI glue into RTC driver")
> > > > Link: https://lore.kernel.org/linux-acpi/[email protected]/
> > > > Reported-by: Mel Gorman <[email protected]>
> > > > Signed-off-by: Rafael J. Wysocki <[email protected]>
> > >
> > > Reviewed-by: Bjorn Helgaas <[email protected]>
> > >
> > > Yep, I blew it with a474aaedac99, sorry about that.
> > >
> > > Possibly could call cmos_wake_setup() from cmos_do_probe() instead of
> > > from cmos_pnp_probe() and cmos_platform_probe()?
> >
> > Sounds good.
> >
> > I would prefer to send a separate patch for this on top of the
> > $subject one, unless Alexandre wants me to do it all in one go.
> >
> > Alexandre, what's your preference here? Or would you prefer if I
> > pushed this forward?
> >
>
> I applied your patch, feel free to improve on top of that ;)
Thank you!
Unfortunately, I broke the wake alarm with this change and a fix has
just been posted:
https://lore.kernel.org/linux-acpi/5887691.lOV4Wx5bFT@kreacher/