[PATCH 2/2] media: rkisp1: Fix IRQ handling due to shared interrupts
Tomi Valkeinen
tomi.valkeinen at ideasonboard.com
Tue Dec 19 05:51:09 PST 2023
On 19/12/2023 15:08, Laurent Pinchart wrote:
> Hi Tomi,
>
> CC'ing Sakari
>
> On Tue, Dec 19, 2023 at 10:50:05AM +0200, Tomi Valkeinen wrote:
>> On 18/12/2023 11:22, Laurent Pinchart wrote:
>>> On Mon, Dec 18, 2023 at 09:54:01AM +0200, Tomi Valkeinen wrote:
>>>> The driver requests the interrupts as IRQF_SHARED, so the interrupt
>>>> handlers can be called at any time. If such a call happens while the ISP
>>>> is powered down, the SoC will hang as the driver tries to access the
>>>> ISP registers.
>>>>
>>>> This can be reproduced even without the platform sharing the IRQ line:
>>>> Enable CONFIG_DEBUG_SHIRQ and unload the driver, and the board will
>>>> hang.
>>>>
>>>> Fix this by adding a new field, 'irqs_enabled', which is used to bail
>>>> out from the interrupt handler when the ISP is not operational.
>>>>
>>>> Signed-off-by: Tomi Valkeinen <tomi.valkeinen at ideasonboard.com>
>>>> ---
>>>> .../platform/rockchip/rkisp1/rkisp1-capture.c | 3 +++
>>>> .../media/platform/rockchip/rkisp1/rkisp1-common.h | 2 ++
>>>> .../media/platform/rockchip/rkisp1/rkisp1-csi.c | 3 +++
>>>> .../media/platform/rockchip/rkisp1/rkisp1-dev.c | 22 ++++++++++++++++++++++
>>>> .../media/platform/rockchip/rkisp1/rkisp1-isp.c | 3 +++
>>>> 5 files changed, 33 insertions(+)
>>>>
>>>> diff --git a/drivers/media/platform/rockchip/rkisp1/rkisp1-capture.c b/drivers/media/platform/rockchip/rkisp1/rkisp1-capture.c
>>>> index aebd3c12020b..c381c22135a2 100644
>>>> --- a/drivers/media/platform/rockchip/rkisp1/rkisp1-capture.c
>>>> +++ b/drivers/media/platform/rockchip/rkisp1/rkisp1-capture.c
>>>> @@ -725,6 +725,9 @@ irqreturn_t rkisp1_capture_isr(int irq, void *ctx)
>>>> unsigned int i;
>>>> u32 status;
>>>>
>>>> + if (!rkisp1->irqs_enabled)
>>>> + return IRQ_NONE;
>>>
>>> Given that this is something all drivers that use shared IRQs have to
>>> do, would it make sense to use a standard helper here, such as
>>> pm_runtime_suspended() for instance ? I haven't looked at which one
>>> would be the most appropriate (if any), there's also
>>> pm_runtime_active() and pm_runtime_status_suspended(). That would
>>> simplify this patch.
>>
>> I did consider that when writing the patch. But I just wasn't very
>> comfortable using the runtime PM here, even if it would make sense, as
>> I'm just not quite sure how it works.
>>
>> For example, pm_runtime_suspended() checks if the device is in
>> RPM_SUSPENDED state, and the device will be in RPM_SUSPENDED after the
>> driver's suspend callback has finished. This makes sense.
>>
>> However, _while_ suspending (not after we have suspended), we want to
>> make sure that 1) no new irq handling will start, 2) we'll wait until
>> any currently running irq handler has finished. For 1), we can't use
>> pm_runtime_suspended() in the irq handler, as the status is not
>> RPM_SUSPENDED. We could probably check for:
>>
>> spin_lock(&dev->power.lock);
>> off = dev->power.runtime_status == RPM_SUSPENDED ||
>> dev->power.runtime_status == RPM_SUSPENDING;
>> spin_unlock(&dev->power.lock);
>> if (off)
>> return IRQ_NONE;
>>
>> That would not work if we would depend on the irq handling while in the
>> suspend callback (e.g. waiting for an irq which signals that the device
>> has finished processing). But we don't do that at the moment, and that
>> kind of this probably can usually be done before calling runtime_put().
>>
>> When we take into account the resume part, I think we could just check
>> for RPM_ACTIVE in the irq handler, which would then rule out
>> RPM_RESUMING, RPM_SUSPENDED and RPM_SUSPENDING.
>>
>> But we can't use pm_runtime_active(), as that also checks for
>> dev->power.disable_depth. In other words, when we disable the PM for our
>> device (e.g. when unloading the driver), the PM framework says our
>> device became active.
>>
>> Soo... I think this should work in the irq handler:
>>
>> spin_lock(&dev->power.lock);
>> active = dev->power.runtime_status == RPM_ACTIVE;
>
> It would be nice to use pm_runtime_active() instead. This would however
> require unregistering the IRQ handler before disabling runtime PM in the
> remove path. I think that should be done nonetheless though, as relying
> on devm to unregister the IRQ handler means it will happen after
> .remove() returns, which could cause all sort of issues (I'm thinking
> about the calls to dev_get_drvdata() in the IRQ handlers for instance).
> What do you think ?
I agree, devm with irqs sounds scary.
What I'd like to do is reserve an irq handler without activating it.
That call could return an error, if, e.g. there's no such irq. Then
later I would enable it (could be in the resume callback, but as well in
the start-stream call), which would never return an error.
Having those both combined in a single call is not nice, as we have to
deal with irq handler calls even when the driver knows it doesn't want them.
>> spin_unlock(&dev->power.lock);
>> if (!active)
>> return IRQ_NONE;
>>
>> I think the driver depends on runtime PM, but if no-PM was an option, I
>> guess we'd need to ifdef the above away, and trust that the device is
>> always powered on.
>>
>> So, as I said in the beginning, "I just wasn't very comfortable using
>> the runtime PM here". And that's still the case =). The runtime PM is
>> horribly complex. If you think the above is clearer, and you think it's
>> correct, I can make the change.
>
> It sounds it may require some more work, and we should land this fix in
> v6.8, with the revert, right ? If so, I'm fine merging this patch, and
> moving to runtime PM checks on top if we decide to do so.
Yes. I think this fix should work as it is now. That said, I haven't
heard anyone else reporting this issue, so maybe just applying the
revert (to fix the driver) would be enough. And we could then figure out
later how exactly to handle the shared interrupts.
> Reviewed-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
>
> By the way, I wonder if it would make sense to handle this in the driver
> core. The prospect of copying this code in all drivers doesn't make me
> happy.
Indeed. But I wonder if it's always like this. E.g. a "wakeup" irq which
is supposed to happen while the device is off.
Tomi
>>>> +
>>>> status = rkisp1_read(rkisp1, RKISP1_CIF_MI_MIS);
>>>> if (!status)
>>>> return IRQ_NONE;
>>>> diff --git a/drivers/media/platform/rockchip/rkisp1/rkisp1-common.h b/drivers/media/platform/rockchip/rkisp1/rkisp1-common.h
>>>> index 4b6b28c05b89..b757f75edecf 100644
>>>> --- a/drivers/media/platform/rockchip/rkisp1/rkisp1-common.h
>>>> +++ b/drivers/media/platform/rockchip/rkisp1/rkisp1-common.h
>>>> @@ -450,6 +450,7 @@ struct rkisp1_debug {
>>>> * @debug: debug params to be exposed on debugfs
>>>> * @info: version-specific ISP information
>>>> * @irqs: IRQ line numbers
>>>> + * @irqs_enabled: the hardware is enabled and can cause interrupts
>>>> */
>>>> struct rkisp1_device {
>>>> void __iomem *base_addr;
>>>> @@ -471,6 +472,7 @@ struct rkisp1_device {
>>>> struct rkisp1_debug debug;
>>>> const struct rkisp1_info *info;
>>>> int irqs[RKISP1_NUM_IRQS];
>>>> + bool irqs_enabled;
>>>> };
>>>>
>>>> /*
>>>> diff --git a/drivers/media/platform/rockchip/rkisp1/rkisp1-csi.c b/drivers/media/platform/rockchip/rkisp1/rkisp1-csi.c
>>>> index b6e47e2f1b94..4202642e0523 100644
>>>> --- a/drivers/media/platform/rockchip/rkisp1/rkisp1-csi.c
>>>> +++ b/drivers/media/platform/rockchip/rkisp1/rkisp1-csi.c
>>>> @@ -196,6 +196,9 @@ irqreturn_t rkisp1_csi_isr(int irq, void *ctx)
>>>> struct rkisp1_device *rkisp1 = dev_get_drvdata(dev);
>>>> u32 val, status;
>>>>
>>>> + if (!rkisp1->irqs_enabled)
>>>> + return IRQ_NONE;
>>>> +
>>>> status = rkisp1_read(rkisp1, RKISP1_CIF_MIPI_MIS);
>>>> if (!status)
>>>> return IRQ_NONE;
>>>> diff --git a/drivers/media/platform/rockchip/rkisp1/rkisp1-dev.c b/drivers/media/platform/rockchip/rkisp1/rkisp1-dev.c
>>>> index acc559652d6e..73cf08a74011 100644
>>>> --- a/drivers/media/platform/rockchip/rkisp1/rkisp1-dev.c
>>>> +++ b/drivers/media/platform/rockchip/rkisp1/rkisp1-dev.c
>>>> @@ -305,6 +305,24 @@ static int __maybe_unused rkisp1_runtime_suspend(struct device *dev)
>>>> {
>>>> struct rkisp1_device *rkisp1 = dev_get_drvdata(dev);
>>>>
>>>> + rkisp1->irqs_enabled = false;
>>>> + /* Make sure the IRQ handler will see the above */
>>>> + mb();
>>>> +
>>>> + /*
>>>> + * Wait until any running IRQ handler has returned. The IRQ handler
>>>> + * may get called even after this (as it's a shared interrupt line)
>>>> + * but the 'irqs_enabled' flag will make the handler return immediately.
>>>> + */
>>>> + for (unsigned int il = 0; il < ARRAY_SIZE(rkisp1->irqs); ++il) {
>>>> + if (rkisp1->irqs[il] == -1)
>>>> + continue;
>>>> +
>>>> + /* Skip if the irq line is the same as previous */
>>>> + if (il == 0 || rkisp1->irqs[il - 1] != rkisp1->irqs[il])
>>>> + synchronize_irq(rkisp1->irqs[il]);
>>>> + }
>>>> +
>>>> clk_bulk_disable_unprepare(rkisp1->clk_size, rkisp1->clks);
>>>> return pinctrl_pm_select_sleep_state(dev);
>>>> }
>>>> @@ -321,6 +339,10 @@ static int __maybe_unused rkisp1_runtime_resume(struct device *dev)
>>>> if (ret)
>>>> return ret;
>>>>
>>>> + rkisp1->irqs_enabled = true;
>>>> + /* Make sure the IRQ handler will see the above */
>>>> + mb();
>>>> +
>>>> return 0;
>>>> }
>>>>
>>>> diff --git a/drivers/media/platform/rockchip/rkisp1/rkisp1-isp.c b/drivers/media/platform/rockchip/rkisp1/rkisp1-isp.c
>>>> index f00873d31c42..78a1f7a1499b 100644
>>>> --- a/drivers/media/platform/rockchip/rkisp1/rkisp1-isp.c
>>>> +++ b/drivers/media/platform/rockchip/rkisp1/rkisp1-isp.c
>>>> @@ -976,6 +976,9 @@ irqreturn_t rkisp1_isp_isr(int irq, void *ctx)
>>>> struct rkisp1_device *rkisp1 = dev_get_drvdata(dev);
>>>> u32 status, isp_err;
>>>>
>>>> + if (!rkisp1->irqs_enabled)
>>>> + return IRQ_NONE;
>>>> +
>>>> status = rkisp1_read(rkisp1, RKISP1_CIF_ISP_MIS);
>>>> if (!status)
>>>> return IRQ_NONE;
>
More information about the Linux-rockchip
mailing list