[RFC PATCH 1/3] PCI: rockchip: Add support for pcie wake irq

Hi Brian,

i've tried dev_pm_set_dedicated_wake_irq, it doesn't work well on 
upstream kernel(for level irq).


it looks like we would delay real set trigger type to request irq after 
this commit:
1e2a7d78499e irqdomain: Don't set type when mapping an IRQ

so calling irq_set_status_flags before request irq would lose trigger 
type setting.


things work well after reverted that commit, so i will send my patch 
soon, and ask irq people about it :)


On 08/17/2017 01:49 AM, Brian Norris wrote:

 Hi,

On Wed, Aug 16, 2017 at 03:52:22PM +0800, Jeffy Chen wrote:

quoted

Add support for PCIE_WAKE pin in rockchip pcie driver.

Signed-off-by: Jeffy Chen <redacted>
---

  drivers/pci/host/pcie-rockchip.c | 58 ++++++++++++++++++++++++++++++++++++++++
  1 file changed, 58 insertions(+)

diff --git a/drivers/pci/host/pcie-rockchip.c b/drivers/pci/host/pcie-rockchip.c
index 7bb9870f6d8c..f969a6d3cd85 100644
--- a/drivers/pci/host/pcie-rockchip.c
+++ b/drivers/pci/host/pcie-rockchip.c

@@ -38,6 +38,7 @@
  #include <linux/platform_device.h>
  #include <linux/reset.h>
  #include <linux/regmap.h>
+#include <linux/suspend.h>

  /*
   * The upper 16 bits of PCIE_CLIENT_CONFIG are a write mask for the lower 16

@@ -226,6 +227,8 @@ struct rockchip_pcie {
  	struct	regulator *vpcie1v8; /* 1.8V power supply */
  	struct	regulator *vpcie0v9; /* 0.9V power supply */
  	struct	gpio_desc *ep_gpio;
+	int	wake_irq;
+	bool	wake_by_pcie;
  	u32	lanes;
  	u8	root_bus_nr;
  	int	link_gen;

@@ -853,6 +856,20 @@ static void rockchip_pcie_legacy_int_handler(struct irq_desc *desc)
  	chained_irq_exit(chip, desc);
  }

+static irqreturn_t rockchip_pcie_wake_irq_handler(int irq, void *arg)
+{
+	struct rockchip_pcie *rockchip = arg;
+
+	rockchip->wake_by_pcie = true;
+
+	disable_irq_nosync(rockchip->wake_irq);
+	disable_irq_wake(rockchip->wake_irq);
+
+	pm_wakeup_event(rockchip->dev, 0);
+	pm_system_wakeup();
+
+	return IRQ_HANDLED;
+}

  /**
   * rockchip_pcie_parse_dt - Parse Device Tree

@@ -868,6 +885,7 @@ static int rockchip_pcie_parse_dt(struct rockchip_pcie *rockchip)
  	struct resource *regs;
  	int irq;
  	int err;
+	bool wakeup = 0;

'0' should be 'false'.

right

quoted

  	regs = platform_get_resource_byname(pdev,
  					    IORESOURCE_MEM,

@@ -1018,6 +1036,21 @@ static int rockchip_pcie_parse_dt(struct rockchip_pcie *rockchip)
  		return err;
  	}

+	rockchip->wake_irq = platform_get_irq_byname(pdev, "wake");
+	if (rockchip->wake_irq >= 0) {
+		err = devm_request_irq(dev, rockchip->wake_irq,
+				       rockchip_pcie_wake_irq_handler,
+				       0, "pcie-wake", rockchip);
+		if (err) {
+			dev_err(dev, "failed to request PCIe wake IRQ\n");
+			return err;
+		}
+
+		disable_irq(rockchip->wake_irq);

If you're worried about keeping this disabled at first, you can just use
this nifty trick (since this isn't a shared interrupt) -- call this
before requesting the IRQ:

	irq_set_status_flags(rockchip->wake_irq, IRQ_NOAUTOEN);

You could also consider using dev_pm_set_dedicated_wake_irq() to handle
this -- but beware, it still might not quite handle level-triggered
interrupt properly. I'm pretty sure Tony Lindgren would be happy to get
testing or patches for that though :) He already sent me something a
while back but I didn't have time to test it out.

ok, that seems very handy

quoted

+		wakeup = device_property_read_bool(dev, "wakeup-source");
+	}
+	device_init_wakeup(dev, wakeup);

Shouldn't you call 'device_init_wakeup(dev, false)' on remove()?

ok

quoted

+
  	rockchip->vpcie3v3 = devm_regulator_get_optional(dev, "vpcie3v3");
  	if (IS_ERR(rockchip->vpcie3v3)) {
  		if (PTR_ERR(rockchip->vpcie3v3) == -EPROBE_DEFER)

@@ -1270,6 +1303,30 @@ static int rockchip_pcie_wait_l2(struct rockchip_pcie *rockchip)
  	return 0;
  }

+static int __maybe_unused rockchip_pcie_suspend(struct device *dev)

Why do this in suspend() instead of suspend_noirq()? You shouldn't
really need a separate method here.

Note that this should be a level-triggered interrupt which remains
asserted, so there should be no chance of "missing" it if you don't
enable it in time.

And on a related note: if you try the dedicated wake irq approach, this
will only occur just before the noirq phase anyway, since
device_wakeup_arm_wake_irqs() is called in dpm_suspend_noirq().

right

quoted

+{
+	struct rockchip_pcie *rockchip = dev_get_drvdata(dev);
+
+	rockchip->wake_by_pcie = false;
+
+	if (device_may_wakeup(dev)) {
+		enable_irq_wake(rockchip->wake_irq);
+		enable_irq(rockchip->wake_irq);
+	}
+	return 0;
+}
+
+static int __maybe_unused rockchip_pcie_resume(struct device *dev)
+{
+	struct rockchip_pcie *rockchip = dev_get_drvdata(dev);
+
+	if (device_may_wakeup(dev) && !rockchip->wake_by_pcie) {

The use of 'wake_by_pcie' is racy; an interrupt could be in flight (but
not completed), and so it could set 'wake_by_pcie' just after you're
reading this. Then, you'll get a double-disable.

I believe the safe way to handle this would be to use an atomic
test-and-set / test-and-clear approach (either atomic_cmpxchg(), or use
a spinlock).

right, and we don't need these if using dev_pm_set_dedicated_wake_irq

quoted

+		disable_irq(rockchip->wake_irq);
+		disable_irq_wake(rockchip->wake_irq);
+	}
+	return 0;
+}
+
  static int __maybe_unused rockchip_pcie_suspend_noirq(struct device *dev)
  {
  	struct rockchip_pcie *rockchip = dev_get_drvdata(dev);

@@ -1548,6 +1605,7 @@ static int rockchip_pcie_remove(struct platform_device *pdev)
  }

  static const struct dev_pm_ops rockchip_pcie_pm_ops = {
+	SET_SYSTEM_SLEEP_PM_OPS(rockchip_pcie_suspend, rockchip_pcie_resume)
  	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(rockchip_pcie_suspend_noirq,
  				      rockchip_pcie_resume_noirq)
  };

Brian