Message-ID: <20151210201817.GB367@localhost>
Date:	Thu, 10 Dec 2015 14:18:17 -0600
From:	Bjorn Helgaas <helgaas@...nel.org>
To:	Bharat Kumar Gogada <bharat.kumar.gogada@...inx.com>
Cc:	"robh+dt@...nel.org" <robh+dt@...nel.org>,
	"pawel.moll@....com" <pawel.moll@....com>,
	"mark.rutland@....com" <mark.rutland@....com>,
	"ijc+devicetree@...lion.org.uk" <ijc+devicetree@...lion.org.uk>,
	"galak@...eaurora.org" <galak@...eaurora.org>,
	Michal Simek <michals@...inx.com>,
	Soren Brinkmann <sorenb@...inx.com>,
	"bhelgaas@...gle.com" <bhelgaas@...gle.com>,
	"arnd@...db.de" <arnd@...db.de>,
	"tinamdar@....com" <tinamdar@....com>,
	"treding@...dia.com" <treding@...dia.com>,
	"rjui@...adcom.com" <rjui@...adcom.com>,
	"Minghuan.Lian@...escale.com" <Minghuan.Lian@...escale.com>,
	"m-karicheri2@...com" <m-karicheri2@...com>,
	"hauke@...ke-m.de" <hauke@...ke-m.de>,
	"marc.zyngier@....com" <marc.zyngier@....com>,
	"dhdang@....com" <dhdang@....com>,
	"sbranden@...adcom.com" <sbranden@...adcom.com>,
	"devicetree@...r.kernel.org" <devicetree@...r.kernel.org>,
	"linux-arm-kernel@...ts.infradead.org" 
	<linux-arm-kernel@...ts.infradead.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"linux-pci@...r.kernel.org" <linux-pci@...r.kernel.org>,
	Ravikiran Gummaluri <rgummal@...inx.com>,
	Paul Burton <paul.burton@...tec.com>,
	Thierry Reding <thierry.reding@...il.com>,
	Stephen Warren <swarren@...dotorg.org>,
	Alexandre Courbot <gnurou@...il.com>
Subject: Re: [PATCH v11] PCI: Xilinx-NWL-PCIe: Added support for Xilinx NWL
 PCIe Host Controller

[+cc Marc for irq_dispose_mapping() question]

On Thu, Dec 10, 2015 at 02:10:34PM +0000, Bharat Kumar Gogada wrote:
> > Subject: Re: [PATCH v11] PCI: Xilinx-NWL-PCIe: Added support for Xilinx NWL
> > > +static bool nwl_pcie_valid_device(struct pci_bus *bus, unsigned int
> > > +devfn) {
> > > +	struct nwl_pcie *pcie = bus->sysdata;
> > > +
> > > +	/* Check link,before accessing downstream ports */
> > > +	if (bus->number != pcie->root_busno) {
> > > +		if (!nwl_pcie_link_up(pcie, PCIE_USER_LINKUP))
> > > +			return false;
> > > +	}
> > > +
> > > +	/* Only one device down on each root port */
> > > +	if (bus->number == pcie->root_busno && devfn > 0)
> > > +		return false;
> > > +
> > > +	/*
> > > +	 * Do not read more than one device on the bus directly attached
> > > +	 * to root port.
> > > +	 */
> > > +	if (bus->primary == pcie->root_busno && devfn > 0)
> > 
> > Why is this?  Do you have some restriction on the PCIe topology below the
> > Root Port?  If there's a spec-compliant link coming from the Root Port, it
> > seems like any legal PCIe device could be attached, including a multifunction
> > device.
> > 
> Primary refers to root port itself, this check is not for devices downstream on link. 

I'm trying to figure out what the difference is between these two
checks and why you have both of them:

> +	if (bus->number == pcie->root_busno && devfn > 0)
> +	if (bus->primary == pcie->root_busno && devfn > 0)

If I understand correctly, pcie->root_busno is the bus number of the
Root Port device (likely 00).  I think the "bus->number ==
pcie->root_busno && devfn > 0" check means that the Root Port, e.g.,
00:00.0, is the only device allowed on bus 00.  Often a Root Complex
contains several Root Ports and other integrated devices that
typically are on bus 00.  But in your case, I think you're saying
there is only the single Root Port and no other devices.

I think that first check takes care of everything on bus 00, so I'm
trying to figure out what the second check is for.  Assume your Root
Port is device 00:00.0 and it is a bridge to [bus 01-ff].  Then we
have two pci_bus structs with these values:

  bus->number = 00
  bus->primary = 00
  bus->busn_res = [bus 00-ff]

  bus->number = 01
  bus->primary = 00
  bus->busn_res = [bus 01-ff]

Because of the first check, 00:00.0 is the only possible device on bus
00, and because of the second check, 01:00.0 is the only possible
device on bus 01.  Therefore, you don't support a multifunction device
connected to the Root Port.  Right?

> > > +		return false;
> > > +
> > > +	return true;
> > > +}
> > > + * nwl_setup_sspl - Set Slot Power limit
> > > + *
> > > + * @pcie: PCIe port information
> > > + */
> > > +static int nwl_setup_sspl(struct nwl_pcie *pcie)
> > 
> > This function needs a little explanation because it doesn't really make any
> > sense when compared with the PCIe spec.  The spec says the Slot Power
> > Limit Value is HwInit, which means it's supposed to be initialized by firmware
> > or hardware, and it should be read-only by the time we get here.
> > I think the value there should reflect parameters of the hardware design; we
> > wouldn't want a user to write an arbitrary limit that would tell a card it can
> > pull more power than the slot can supply.
> > 
> > But I saw another driver recently (pcie-snpsdev.c) that's doing something
> > similar, so there must be something else going on here.  In any case, a
> > comment here about exactly *what* is going on would be much appreciated.
> 
> The spec says the following:
[This is all from PCIe spec r3.0, sec 2.2.8.5]

> The Set_Slot_Power_Limit Message includes a one DW data payload. The
> data payload is copied from the Slot Capabilities register of the
> Downstream Port and is written into the Device Capabilities register
> of the Upstream Port on the other side of the Link. Bits 9:8 of the
> data payload map to the Slot Power Limit Scale field and bits 7:0
> map to the Slot Power Limit Value field. Bits 31:10 of the data
> payload must be set to all 0's by the Transmitter and ignored by the
> Receiver.

> This Message is sent automatically by the Downstream Port (of a Root
> Complex or a Switch) when one of the following events occurs:
> -> On a Configuration Write to the Slot Capabilities register (see
> Section 7.8.9) when the Data Link Layer reports DL_Up status.

I interpret this as meaning "the *hardware* automatically sends a
Set_Slot_Power_Limit Message."  There's no mention of software doing
anything other than the configuration write.

If your hardware doesn't do that, I think it's a defect.  It's fine to
work around it, but we should have a comment to that effect so people
don't copy the code to new drivers that don't need it.

It's a little strange that 7.8.9 talks about writing to this register
when all of its fields are HwInit and supposedly read-only.  I had
assumed devices would use strapping or implementation-specific
registers to set the Slot Power values, but maybe some devices use
direct writes to Slot Capabilities instead.

BTW, I noticed a related lspci bug: it didn't decode the Capture Slot
Power Limit in Device Capabilities of Endpoints.  I posted a fix for
that separately.

The Slot Power Limit (in Slot Capabilities) indicates how much power
the slot can supply to a downstream device.  That's a function of the
platform design, so it seems like this is something you want to set
via DT or some other mechanism that knows about the platform.
Intercepting all config writes and updating it with whatever the
caller supplies doesn't sound wise.  The value might be coming from
setpci or some other source with no knowledge of the platform.

> > > +			status = nwl_bridge_readl(pcie, TX_PCIE_MSG)
> > > +						  & MSG_DONE_BIT;
> > > +			if (status) {
> > > +				status = nwl_bridge_readl(pcie,
> > TX_PCIE_MSG)
> > > +						  & MSG_DONE_STATUS_BIT;

> > It's not clear to me whether you need to re-read TX_PCIE_MSG here.
> 
> MSG_DONE_BIT qualifies MSG_DONE_STATUS_BIT; so value of
> MSG_DONE_STATUS_BIT is valid only when MSG_DONE_BIT = 1

That doesn't answer the question of whether another read is required.
In fact, I would argue that if MSG_DONE_STATUS_BIT is only valid when
MSG_DONE_BIT = 1, you *should* only do one read, because you want to
capture both bits simultaneously so you know they're consistent, e.g.,

  status = nwl_bridge_readl(pcie, TX_PCIE_MSG);
  if (status & MSG_DONE_BIT) {
    if (status & MSG_DONE_STATUS_BIT)
      ...
  }

If you read the register twice, you always have to worry about what
changes MSG_DONE_BIT, and how you guarantee that the second read
happens before MSG_DONE_BIT changes.

> > > +		}
> > > +	} while (status);
> > > +
> > > +	return retval;
> > > 
> > > +	for (i = 0; i < 4; i++) {
> > > +		irq = irq_find_mapping(pcie->legacy_irq_domain, i + 1);
> > > +		if (irq > 0)
> > > +			irq_dispose_mapping(irq);
> > > +	}
> > > +	if (pcie->legacy_irq_domain)
> > > +		irq_domain_remove(pcie->legacy_irq_domain);
> > 
> > Something seems wrong here.  I don't know when irq_dispose_mapping() is
> > required, but it's not used consistently in drivers/pci, and it should be.
> > Currently, only pci-tegra.c, pcie-xilinx.c, and this new driver use it.  Tegra uses
> > it only for MSIs, and Xilinx seems to use it for both MSIs and INTx.  What's
> > right?
> Its not related to MSI or INTx, its related to domain, for freeing irq descriptor associated with irq.

So are you saying that other drivers in drivers/pci/host should be
using irq_dispose_mapping(), but they aren't?

Marc, can you chime in here?

> > > +	pcie = devm_kzalloc(&pdev->dev, sizeof(*pcie), GFP_KERNEL);
> > > +	if (!pcie)
> > > +		return -ENOMEM;
> > > +
> > > +	err = nwl_pcie_parse_dt(pcie, pdev);
> > > +	if (err) {
> > > +		dev_err(pcie->dev, "Parsing DT failed\n");
> > > +		return err;
> > 
> > This leaks the nwl_pcie struct you just allocated.  And all the subsequent
> > error paths do too, of course.
> > 
> Its resource-managed kzalloc(), kernel will take care of it.

Heh, devm_*() always trips me up, sorry about that.

Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists