irqchip: bcm2835: Add FIQ support

Commit Message

Noralf Trønnes June 18, 2015, 1:32 p.m. UTC

Den 18.06.2015 04:26, skrev Stephen Warren:
> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>> Add a duplicate irq range with an offset on the hwirq's so the
>> driver can detect that enable_fiq() is used.
>> Tested with downstream dwc_otg USB controller driver.
> This basically looks OK, but a few comments/thoughts:
>
> a) Should the Kconfig change be a separate patch since it's a separate
> subsystem?

I can separate it out.

> b) Doesn't the driver need to refuse some operation (handler
> registration, IRQ setup, IRQ enable, ...?) for more than 1 IRQ in the
> FIQ range, since the FIQ control register only allows routing 1 IRQ to FIQ.

claim_fiq() protects the FIQ. See d) answer below.

> c) The DT binding needs updating to describe the extra IRQs:
>
>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt

Ok.

> d) I wonder how the FIQ handler actually gets routed to this controller
> and hooked to its handler etc. I assume there's a separate patch for
> that coming?

set_fiq_handler() sets the handler and enable_fiq() enables it:

     if (claim_fiq(&fh))
         ERROR;
     set_fiq_handler(...)
     set_fiq_regs(&regs);
     enable_fiq(irq);
     local_fiq_enable();


Downstream dwc_otg
------------------

FIQ handler:
https://github.com/raspberrypi/linux/blob/rpi-4.0.y/drivers/usb/host/dwc_otg/dwc_otg_fiq_stub.S

FIQ is set up in hcd_init_fiq():
https://github.com/raspberrypi/linux/blob/rpi-4.0.y/drivers/usb/host/dwc_otg/dwc_otg_hcd_linux.c

This patch is also necessary:

DT node:
         usb: usb@7e980000 {
             compatible = "brcm,bcm2708-usb";
             reg = <0x7e980000 0x10000>,
                   <0x7e006000 0x1000>;
             interrupts = <2 0>,
                      <1 9>;
         };

Comments

Noralf Trønnes June 18, 2015, 4:23 p.m. UTC | #1

Den 18.06.2015 15:32, skrev Noralf Trønnes:
>
> Den 18.06.2015 04:26, skrev Stephen Warren:
>> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>>> Add a duplicate irq range with an offset on the hwirq's so the
>>> driver can detect that enable_fiq() is used.
>>> Tested with downstream dwc_otg USB controller driver.
>> This basically looks OK, but a few comments/thoughts:
>>

>> c) The DT binding needs updating to describe the extra IRQs:
>>
>>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt 
>>>
>
> Ok.
>

I have seconds thoughts on this:
This patch does not change the DT bindings so I don't see what update
I should make. This patch only adds support for the Linux way of
handling FIQ's through enable_fiq(). It doesn't change how interrupts
are described in the DT.

Stephen Warren July 11, 2015, 4:09 a.m. UTC | #2

(Sorry for the slow reply; I was on vacation)

On 06/18/2015 07:32 AM, Noralf Trønnes wrote:
> Den 18.06.2015 04:26, skrev Stephen Warren:
>> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>>> Add a duplicate irq range with an offset on the hwirq's so the
>>> driver can detect that enable_fiq() is used.
>>> Tested with downstream dwc_otg USB controller driver.
>> This basically looks OK, but a few comments/thoughts:

>> b) Doesn't the driver need to refuse some operation (handler
>> registration, IRQ setup, IRQ enable, ...?) for more than 1 IRQ in the
>> FIQ range, since the FIQ control register only allows routing 1 IRQ to
>> FIQ.
> 
> claim_fiq() protects the FIQ. See d) answer below.

That assumes the IRQ is "accessed" via the fiq-specific APIs. Since this
patch changes the IRQ domain from having n IRQs to having 2*n IRQs, and
doesn't do anything special to prevent clients from using IRQs n..2n-1
via the existing IRQ APIs, it's quite possible the a buggy client would.

(From another email):
>>> c) The DT binding needs updating to describe the extra IRQs:
>>>
>>>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt
>>
>> Ok.
> 
> I have seconds thoughts on this:
> This patch does not change the DT bindings so I don't see what update
> I should make. This patch only adds support for the Linux way of
> handling FIQ's through enable_fiq(). It doesn't change how interrupts
> are described in the DT.

The intention of the patch may not be to expand the set of IRQs
available via DT, but it does in practice. I think you need to add a
custom of_xlate for the IRQ domain to ensure that only IRQs 0..n-1 can
be translated from DT, and not IRQs n..2n-1. If you do that, then I
agree that no DT binding update should be required.

Even with a custom of_xlate function, some code could hard-code an IRQ
number and hence end up registering a FIQ handler that way. However, I
guess that's a bug that the driver doesn't need to solve. We can just
fix that bug in the kernel code in that case. The same argument doesn't
apply to bad DTs; we need to more aggressively protect against that case.

>> d) I wonder how the FIQ handler actually gets routed to this controller
>> and hooked to its handler etc. I assume there's a separate patch for
>> that coming?
> 
> set_fiq_handler() sets the handler and enable_fiq() enables it:
> 
>     if (claim_fiq(&fh))
>         ERROR;
>     set_fiq_handler(...)
>     set_fiq_regs(&regs);
>     enable_fiq(irq);
>     local_fiq_enable();

Noralf Trønnes July 11, 2015, 3:26 p.m. UTC | #3

Den 11.07.2015 06:09, skrev Stephen Warren:
> (Sorry for the slow reply; I was on vacation)
>
> On 06/18/2015 07:32 AM, Noralf Trønnes wrote:
>> Den 18.06.2015 04:26, skrev Stephen Warren:
>>> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>>>> Add a duplicate irq range with an offset on the hwirq's so the
>>>> driver can detect that enable_fiq() is used.
>>>> Tested with downstream dwc_otg USB controller driver.
>>> This basically looks OK, but a few comments/thoughts:
>>> b) Doesn't the driver need to refuse some operation (handler
>>> registration, IRQ setup, IRQ enable, ...?) for more than 1 IRQ in the
>>> FIQ range, since the FIQ control register only allows routing 1 IRQ to
>>> FIQ.
>> claim_fiq() protects the FIQ. See d) answer below.
> That assumes the IRQ is "accessed" via the fiq-specific APIs. Since this
> patch changes the IRQ domain from having n IRQs to having 2*n IRQs, and
> doesn't do anything special to prevent clients from using IRQs n..2n-1
> via the existing IRQ APIs, it's quite possible the a buggy client would.

Yes, but doesn't this apply to all irq use, using the wrong one doesn't 
work.
If FIQ's where in more common use, we might have seen a FIQ IRQ flag instead
of special FIQ irqs.

> (From another email):
>>>> c) The DT binding needs updating to describe the extra IRQs:
>>>>
>>>>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt
>>> Ok.
>> I have seconds thoughts on this:
>> This patch does not change the DT bindings so I don't see what update
>> I should make. This patch only adds support for the Linux way of
>> handling FIQ's through enable_fiq(). It doesn't change how interrupts
>> are described in the DT.
> The intention of the patch may not be to expand the set of IRQs
> available via DT, but it does in practice. I think you need to add a
> custom of_xlate for the IRQ domain to ensure that only IRQs 0..n-1 can
> be translated from DT, and not IRQs n..2n-1. If you do that, then I
> agree that no DT binding update should be required.

armctrl_xlate() maps to the same hwirqs as before. This patch adds a
new range of hwirqs at the end of the "real" hwirq range.
It's not possible to get to these FIQ shadow hwirqs through DT.

Stephen Warren July 14, 2015, 4:50 a.m. UTC | #4

On 07/11/2015 09:26 AM, Noralf Trønnes wrote:
> 
> Den 11.07.2015 06:09, skrev Stephen Warren:
>> (Sorry for the slow reply; I was on vacation)
>>
>> On 06/18/2015 07:32 AM, Noralf Trønnes wrote:
>>> Den 18.06.2015 04:26, skrev Stephen Warren:
>>>> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>>>>> Add a duplicate irq range with an offset on the hwirq's so the
>>>>> driver can detect that enable_fiq() is used.
>>>>> Tested with downstream dwc_otg USB controller driver.
>>>> This basically looks OK, but a few comments/thoughts:
>>>> b) Doesn't the driver need to refuse some operation (handler
>>>> registration, IRQ setup, IRQ enable, ...?) for more than 1 IRQ in the
>>>> FIQ range, since the FIQ control register only allows routing 1 IRQ to
>>>> FIQ.
>>> claim_fiq() protects the FIQ. See d) answer below.
>> That assumes the IRQ is "accessed" via the fiq-specific APIs. Since this
>> patch changes the IRQ domain from having n IRQs to having 2*n IRQs, and
>> doesn't do anything special to prevent clients from using IRQs n..2n-1
>> via the existing IRQ APIs, it's quite possible the a buggy client would.
> 
> Yes, but doesn't this apply to all irq use, using the wrong one doesn't
> work.
> If FIQ's where in more common use, we might have seen a FIQ IRQ flag
> instead
> of special FIQ irqs.
> 
>> (From another email):
>>>>> c) The DT binding needs updating to describe the extra IRQs:
>>>>>
>>>>>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt
>>>>>>
>>>> Ok.
>>> I have seconds thoughts on this:
>>> This patch does not change the DT bindings so I don't see what update
>>> I should make. This patch only adds support for the Linux way of
>>> handling FIQ's through enable_fiq(). It doesn't change how interrupts
>>> are described in the DT.
>> The intention of the patch may not be to expand the set of IRQs
>> available via DT, but it does in practice. I think you need to add a
>> custom of_xlate for the IRQ domain to ensure that only IRQs 0..n-1 can
>> be translated from DT, and not IRQs n..2n-1. If you do that, then I
>> agree that no DT binding update should be required.
> 
> armctrl_xlate() maps to the same hwirqs as before. This patch adds a
> new range of hwirqs at the end of the "real" hwirq range.
> It's not possible to get to these FIQ shadow hwirqs through DT.

What prevents a DT from (incorrectly) referencing the extra hwirqs?

Noralf Trønnes July 14, 2015, 11:48 a.m. UTC | #5

Den 14.07.2015 06:50, skrev Stephen Warren:
> On 07/11/2015 09:26 AM, Noralf Trønnes wrote:
>> Den 11.07.2015 06:09, skrev Stephen Warren:
>>> (Sorry for the slow reply; I was on vacation)
>>>
>>> On 06/18/2015 07:32 AM, Noralf Trønnes wrote:
>>>> Den 18.06.2015 04:26, skrev Stephen Warren:
>>>>> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>>>>>> Add a duplicate irq range with an offset on the hwirq's so the
>>>>>> driver can detect that enable_fiq() is used.
>>>>>> Tested with downstream dwc_otg USB controller driver.
>>>>> This basically looks OK, but a few comments/thoughts:
>>>>> b) Doesn't the driver need to refuse some operation (handler
>>>>> registration, IRQ setup, IRQ enable, ...?) for more than 1 IRQ in the
>>>>> FIQ range, since the FIQ control register only allows routing 1 IRQ to
>>>>> FIQ.
>>>> claim_fiq() protects the FIQ. See d) answer below.
>>> That assumes the IRQ is "accessed" via the fiq-specific APIs. Since this
>>> patch changes the IRQ domain from having n IRQs to having 2*n IRQs, and
>>> doesn't do anything special to prevent clients from using IRQs n..2n-1
>>> via the existing IRQ APIs, it's quite possible the a buggy client would.
>> Yes, but doesn't this apply to all irq use, using the wrong one doesn't
>> work.
>> If FIQ's where in more common use, we might have seen a FIQ IRQ flag
>> instead
>> of special FIQ irqs.
>>
>>> (From another email):
>>>>>> c) The DT binding needs updating to describe the extra IRQs:
>>>>>>
>>>>>>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt
>>>>>>>
>>>>> Ok.
>>>> I have seconds thoughts on this:
>>>> This patch does not change the DT bindings so I don't see what update
>>>> I should make. This patch only adds support for the Linux way of
>>>> handling FIQ's through enable_fiq(). It doesn't change how interrupts
>>>> are described in the DT.
>>> The intention of the patch may not be to expand the set of IRQs
>>> available via DT, but it does in practice. I think you need to add a
>>> custom of_xlate for the IRQ domain to ensure that only IRQs 0..n-1 can
>>> be translated from DT, and not IRQs n..2n-1. If you do that, then I
>>> agree that no DT binding update should be required.
>> armctrl_xlate() maps to the same hwirqs as before. This patch adds a
>> new range of hwirqs at the end of the "real" hwirq range.
>> It's not possible to get to these FIQ shadow hwirqs through DT.
> What prevents a DT from (incorrectly) referencing the extra hwirqs?

armctrl_xlate() has these limits:

if (WARN_ON(intspec[0] >= NR_BANKS))
if (WARN_ON(intspec[1] >= IRQS_PER_BANK))
if (WARN_ON(intspec[0] == 0 && intspec[1] >= NR_IRQS_BANK0))

Thus the maximum values allowed are:
intspec[0]: (NR_BANKS - 1) = 2
intspec[1]: (IRQS_PER_BANK - 1) = 31

This gives a maximum hwirq:
*out_hwirq = MAKE_HWIRQ(intspec[0], intspec[1]);
*out_hwirq = (2 << 5) | 31 = 95

The FIQ shadow hwirq range starts at 96:
irq = irq_create_mapping(intc.domain, MAKE_HWIRQ(b, i) + NUMBER_IRQS);

NUMBER_IRQS = MAKE_HWIRQ(NR_BANKS, 0) = 96

Stephen Warren July 22, 2015, 1:50 a.m. UTC | #6

On 07/14/2015 05:48 AM, Noralf Trønnes wrote:
> Den 14.07.2015 06:50, skrev Stephen Warren:
>> On 07/11/2015 09:26 AM, Noralf Trønnes wrote:
>>> Den 11.07.2015 06:09, skrev Stephen Warren:
>>>> (Sorry for the slow reply; I was on vacation)
>>>>
>>>> On 06/18/2015 07:32 AM, Noralf Trønnes wrote:
>>>>> Den 18.06.2015 04:26, skrev Stephen Warren:
>>>>>> On 06/12/2015 11:26 AM, Noralf Trønnes wrote:
>>>>>>> Add a duplicate irq range with an offset on the hwirq's so the
>>>>>>> driver can detect that enable_fiq() is used.
>>>>>>> Tested with downstream dwc_otg USB controller driver.
>>>>>> This basically looks OK, but a few comments/thoughts:
>>>>>> b) Doesn't the driver need to refuse some operation (handler
>>>>>> registration, IRQ setup, IRQ enable, ...?) for more than 1 IRQ in the
>>>>>> FIQ range, since the FIQ control register only allows routing 1
>>>>>> IRQ to
>>>>>> FIQ.
>>>>> claim_fiq() protects the FIQ. See d) answer below.
>>>> That assumes the IRQ is "accessed" via the fiq-specific APIs. Since
>>>> this
>>>> patch changes the IRQ domain from having n IRQs to having 2*n IRQs, and
>>>> doesn't do anything special to prevent clients from using IRQs n..2n-1
>>>> via the existing IRQ APIs, it's quite possible the a buggy client
>>>> would.
>>> Yes, but doesn't this apply to all irq use, using the wrong one doesn't
>>> work.
>>> If FIQ's where in more common use, we might have seen a FIQ IRQ flag
>>> instead
>>> of special FIQ irqs.
>>>
>>>> (From another email):
>>>>>>> c) The DT binding needs updating to describe the extra IRQs:
>>>>>>>
>>>>>>>> Documentation/devicetree/bindings/interrupt-controller/brcm,bcm28armctrl-ic.txt
>>>>>>>>
>>>>>>>>
>>>>>> Ok.
>>>>> I have seconds thoughts on this:
>>>>> This patch does not change the DT bindings so I don't see what update
>>>>> I should make. This patch only adds support for the Linux way of
>>>>> handling FIQ's through enable_fiq(). It doesn't change how interrupts
>>>>> are described in the DT.
>>>> The intention of the patch may not be to expand the set of IRQs
>>>> available via DT, but it does in practice. I think you need to add a
>>>> custom of_xlate for the IRQ domain to ensure that only IRQs 0..n-1 can
>>>> be translated from DT, and not IRQs n..2n-1. If you do that, then I
>>>> agree that no DT binding update should be required.
>>> armctrl_xlate() maps to the same hwirqs as before. This patch adds a
>>> new range of hwirqs at the end of the "real" hwirq range.
>>> It's not possible to get to these FIQ shadow hwirqs through DT.
>> What prevents a DT from (incorrectly) referencing the extra hwirqs?
> 
> armctrl_xlate() has these limits:
> 
> if (WARN_ON(intspec[0] >= NR_BANKS))
> if (WARN_ON(intspec[1] >= IRQS_PER_BANK))
> if (WARN_ON(intspec[0] == 0 && intspec[1] >= NR_IRQS_BANK0))
> 
> Thus the maximum values allowed are:
> intspec[0]: (NR_BANKS - 1) = 2
> intspec[1]: (IRQS_PER_BANK - 1) = 31
> 
> This gives a maximum hwirq:
> *out_hwirq = MAKE_HWIRQ(intspec[0], intspec[1]);
> *out_hwirq = (2 << 5) | 31 = 95
> 
> The FIQ shadow hwirq range starts at 96:
> irq = irq_create_mapping(intc.domain, MAKE_HWIRQ(b, i) + NUMBER_IRQS);
> 
> NUMBER_IRQS = MAKE_HWIRQ(NR_BANKS, 0) = 96

Great, thanks for the explanation. That should be fine then.

Patch

diff --git a/drivers/usb/host/dwc_otg/dwc_otg_driver.c 
b/drivers/usb/host/dwc_otg/dwc_otg_driver.c
index 53307f0..95edadf 100644
--- a/drivers/usb/host/dwc_otg/dwc_otg_driver.c
+++ b/drivers/usb/host/dwc_otg/dwc_otg_driver.c
@@ -723,6 +723,7 @@  static int dwc_otg_driver_probe(

      memset(dwc_otg_device, 0, sizeof(*dwc_otg_device));
      dwc_otg_device->os_dep.reg_offset = 0xFFFFFFFF;
+    dwc_otg_device->os_dep.platformdev = _dev;

      /*
       * Map the DWC_otg Core memory into virtual address space.
diff --git a/drivers/usb/host/dwc_otg/dwc_otg_hcd_intr.c 
b/drivers/usb/host/dwc_otg/dwc_otg_hcd_intr.c
index 8a31562..2961985 100644
--- a/drivers/usb/host/dwc_otg/dwc_otg_hcd_intr.c
+++ b/drivers/usb/host/dwc_otg/dwc_otg_hcd_intr.c
@@ -36,10 +36,8 @@ 
  #include "dwc_otg_regs.h"

  #include <linux/jiffies.h>
-#include <mach/hardware.h>
  #include <asm/fiq.h>

-
  extern bool microframe_schedule;

  /** @file
diff --git a/drivers/usb/host/dwc_otg/dwc_otg_hcd_linux.c 
b/drivers/usb/host/dwc_otg/dwc_otg_hcd_linux.c
index 6aad9c4..0440c66 100644
--- a/drivers/usb/host/dwc_otg/dwc_otg_hcd_linux.c
+++ b/drivers/usb/host/dwc_otg/dwc_otg_hcd_linux.c
@@ -445,7 +445,11 @@  static void hcd_init_fiq(void *cookie)
          DWC_WARN("MPHI periph has NOT been enabled");
  #endif
      // Enable FIQ interrupt from USB peripheral
+#ifdef CONFIG_ARCH_BCM2835
+    enable_fiq(platform_get_irq(otg_dev->os_dep.platformdev, 1));
+#else
      enable_fiq(INTERRUPT_VC_USB);
+#endif
      local_fiq_enable();
  }