Re: [PATCH V2] PCI: rcar: Add the initialization of PCIe link in resume_noirq

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Friday, March 8, 2019 10:35:49 PM CET Marek Vasut wrote:
> On 3/8/19 6:17 PM, Bjorn Helgaas wrote:
> > [+cc linux-pm, Rafael for SET_NOIRQ_SYSTEM_SLEEP_PM_OPS question at the end]
> > 
> > On Thu, Mar 07, 2019 at 11:49:34PM +0100, Marek Vasut wrote:
> >> On 3/7/19 9:50 PM, Bjorn Helgaas wrote:
> >>> On Sun, Feb 17, 2019 at 02:24:41PM +0100, marek.vasut@gmail.com wrote:
> >>>> From: Kazufumi Ikeda <kaz-ikeda@xc.jp.nec.com>
> >>>>
> >>>> Reestablish the PCIe link very early in the resume process in case it
> >>>> went down to prevent PCI accesses from hanging the bus. Such accesses
> >>>> can happen early in the PCI resume process, in the resume_noirq, thus
> >>>> the link must be reestablished in the resume_noirq callback of the
> >>>> driver.
> >>>>
> >>>> Signed-off-by: Kazufumi Ikeda <kaz-ikeda@xc.jp.nec.com>
> >>>> Signed-off-by: Gaku Inami <gaku.inami.xw@bp.renesas.com>
> >>>> Signed-off-by: Marek Vasut <marek.vasut+renesas@gmail.com>
> >>>> Cc: Geert Uytterhoeven <geert+renesas@glider.be>
> >>>> Cc: Phil Edworthy <phil.edworthy@renesas.com>
> >>>> Cc: Simon Horman <horms+renesas@verge.net.au>
> >>>> Cc: Wolfram Sang <wsa@the-dreams.de>
> >>>> Cc: linux-renesas-soc@vger.kernel.org
> >>>> ---
> >>>> V2: - Use BIT() macro for (1 << n)
> >>>>     - Since polling in rcar_pcie_wait_for_dl() uses udelay(), do not
> >>>>       add extra changes to this function anymore
> >>>>     - Make resume_noirq return early and clean up parenthesis therein
> >>>> ---
> >>>>  drivers/pci/controller/pcie-rcar.c | 20 ++++++++++++++++++++
> >>>>  1 file changed, 20 insertions(+)
> >>>>
> >>>> diff --git a/drivers/pci/controller/pcie-rcar.c b/drivers/pci/controller/pcie-rcar.c
> >>>> index c8febb009454..b8f8fb3bc640 100644
> >>>> --- a/drivers/pci/controller/pcie-rcar.c
> >>>> +++ b/drivers/pci/controller/pcie-rcar.c
> >>>> @@ -46,6 +46,7 @@
> >>>>  
> >>>>  /* Transfer control */
> >>>>  #define PCIETCTLR		0x02000
> >>>> +#define  DL_DOWN		BIT(3)
> >>>>  #define  CFINIT			1
> >>>
> >>> I saw discussion after the V1 patch about using BIT() and making
> >>> similar constants also use BIT() for consistency.  That makes sense to
> >>> me, and I think the best way would be:
> >>>
> >>>   1) in *this* patch, use "#define DL_DOWN 8"
> >>>   2) in a followup patch, convert them all to BIT()
> >>>
> >>> That way each revision of pcie-rcar.c is self-consistent.
> >>
> >> But the BIT() macros are already cleaned , see commit
> >> 0ee40820989b330e24926d82953ffb9e1c7a8425
> >>
> >>     PCI: rcar: Clean up the macros
> > 
> > Hmmm.  Maybe I'm missing something, but it looks like 0ee40820989b
> > didn't touch CFINIT, DATA_LINK_ACTIVE, or MSIFE.  Arguably,
> > LINK_SPEED_2_5GTS and LINK_SPEED_5_0GTS could use BIT() also.
> > 
> > I guess I'm just old-school, but in my personal opinion, BIT() is more
> > trouble than it's worth.  I'd rather see a complete bitmask because I
> > can easily match it with the typical pictures in a spec, multi-bit
> > fields are easy (you don't have to mix BIT() and GENMASK()), it gives
> > a hint about the register width, it's easy to match with a hexdump,
> > etc, e.g.,
> > 
> >   #define  DL_DOWN        0x00000008
> >   #define  CFINIT         0x00000001
> > 
> > But I'm not suggesting that you get rid of BIT() in this driver.  I'm
> > fine with it as long as it's used consistently.
> > 
> > BTW, while we're looking at it, I think rcar_pci_read_reg() and
> > rcar_pci_write_reg() really should use "u32" instead of "unsigned
> > long", since they deal with hardware registers that are explicitly
> > 32 bits wide.
> 
> OK, I can send those as separate patches.
> 
> >>>>  #define PCIETSTR		0x02004
> >>>>  #define  DATA_LINK_ACTIVE	1
> >>>> @@ -1130,6 +1131,7 @@ static int rcar_pcie_probe(struct platform_device *pdev)
> >>>>  	pcie = pci_host_bridge_priv(bridge);
> >>>>  
> >>>>  	pcie->dev = dev;
> >>>> +	platform_set_drvdata(pdev, pcie);
> >>>>  
> >>>>  	err = pci_parse_request_of_pci_ranges(dev, &pcie->resources, NULL);
> >>>>  	if (err)
> >>>> @@ -1221,10 +1223,28 @@ static int rcar_pcie_probe(struct platform_device *pdev)
> >>>>  	return err;
> >>>>  }
> >>>>  
> >>>> +static int rcar_pcie_resume_noirq(struct device *dev)
> >>>> +{
> >>>> +	struct rcar_pcie *pcie = dev_get_drvdata(dev);
> >>>> +
> >>>> +	if (rcar_pci_read_reg(pcie, PMSR) &&
> >>>> +	    !(rcar_pci_read_reg(pcie, PCIETCTLR) & DL_DOWN))
> >>>> +		return 0;
> >>>> +
> >>>> +	/* Re-establish the PCIe link */
> >>>> +	rcar_pci_write_reg(pcie, CFINIT, PCIETCTLR);
> >>>> +	return rcar_pcie_wait_for_dl(pcie);
> >>>> +}
> >>>> +
> >>>> +static const struct dev_pm_ops rcar_pcie_pm_ops = {
> >>>> +	.resume_noirq = rcar_pcie_resume_noirq,
> >>>> +};
> >>>
> >>> I think there's the beginning of a convention to use #ifdef
> >>> CONFIG_PM_SLEEP around the ops themselves [1].  Otherwise I think
> >>> we'll get a warning about unused code when CONFIG_PM_SLEEP is unset.
> >>
> >> Only if I used SET_NOIRQ_SYSTEM_SLEEP_PM_OPS() , but I set the
> >> resume_noirq directly.
> > 
> > Fair enough.  I guess in this case if CONFIG_PM_SLEEP is unset, you
> > set the pointer, which avoids the "unused function" warning, but we
> > just never use that function pointer.
> > 
> > My intent is to avoid needless differences when possible, so when I
> > review things like this I look at how other drivers do things.  It
> > looks like all the other controllers use
> > SET_NOIRQ_SYSTEM_SLEEP_PM_OPS() or similar:
> > 
> >   git grep -A3 "static.*dev_pm_ops" drivers/pci/controller
> > 
> > In the rcar case you only need the resume_fn, not the suspend_fn, so
> > SET_NOIRQ_SYSTEM_SLEEP_PM_OPS() does a little more than you need and
> > you'd have to pass NULL for suspend_fn.  I didn't check them all
> > (suspend_noirq, freeze_noirq, poweroff_noirq), but at least for
> > suspend_noirq, all users check for NULL before calling through the
> > .suspend_noirq() function pointer, so I think that should be safe.

Yes, this should be safe from the NULL pointer checks perspective.

> > This *does* raise the question of whether you should be setting
> > .thaw_noirq and .restore_noirq in addition to .resume_noirq, as
> > SET_NOIRQ_SYSTEM_SLEEP_PM_OPS() would.  I'm not a PM person, but maybe
> > Rafael or others will chime in.

That depends on what happens during hibernation/restore.

The driver may not need to do anything before the memory snapshot is taken or
before the restore kernel jumps back to the image kernel.  In that case you
don't need freeze/thaw/restore_noirq at all.

However, chances are that it needs to do something and in that case doing
the same thing as during system-wide suspend/resume generally works, even
though it may not be particularly energy-efficient.  That's why it very
ofther is done.