Re: [RESEND PATCH 2/3] PCI: aardvark: Fix checking for PIO status

[Pali Rohár]

On Monday 19 July 2021 18:12:27 Bjorn Helgaas wrote:
> On Fri, Jun 25, 2021 at 12:04:29PM +0100, Lorenzo Pieralisi wrote:
> > On Thu, Jun 24, 2021 at 11:33:44PM +0200, Pali Rohár wrote:
> >
> > [...]
> > 
> > > -static void advk_pcie_check_pio_status(struct advk_pcie *pcie)
> > > +static int advk_pcie_check_pio_status(struct advk_pcie *pcie, u32 *val)
> > >  {
> > >  	struct device *dev = &pcie->pdev->dev;
> > >  	u32 reg;
> > > @@ -472,15 +476,50 @@ static void advk_pcie_check_pio_status(struct advk_pcie *pcie)
> > >  	status = (reg & PIO_COMPLETION_STATUS_MASK) >>
> > >  		PIO_COMPLETION_STATUS_SHIFT;
> > >  
> > > -	if (!status)
> > > -		return;
> > > -
> > > +	/*
> > > +	 * According to HW spec, the PIO status check sequence as below:
> > > +	 * 1) even if COMPLETION_STATUS(bit9:7) indicates successful,
> > > +	 *    it still needs to check Error Status(bit11), only when this bit
> > > +	 *    indicates no error happen, the operation is successful.
> > > +	 * 2) value Unsupported Request(1) of COMPLETION_STATUS(bit9:7) only
> > > +	 *    means a PIO write error, and for PIO read it is successful with
> > > +	 *    a read value of 0xFFFFFFFF.
> > > +	 * 3) value Completion Retry Status(CRS) of COMPLETION_STATUS(bit9:7)
> > > +	 *    only means a PIO write error, and for PIO read it is successful
> > > +	 *    with a read value of 0xFFFF0001.
> > > +	 * 4) value Completer Abort (CA) of COMPLETION_STATUS(bit9:7) means
> > > +	 *    error for both PIO read and PIO write operation.
> > > +	 * 5) other errors are indicated as 'unknown'.
> > > +	 */
> > >  	switch (status) {
> > > +	case PIO_COMPLETION_STATUS_OK:
> > > +		if (reg & PIO_ERR_STATUS) {
> > > +			strcomp_status = "COMP_ERR";
> > > +			break;
> > > +		}
> > > +		/* Get the read result */
> > > +		if (val)
> > > +			*val = advk_readl(pcie, PIO_RD_DATA);
> > > +		/* No error */
> > > +		strcomp_status = NULL;
> > > +		break;
> > >  	case PIO_COMPLETION_STATUS_UR:
> > > -		strcomp_status = "UR";
> > > +		if (val) {
> > > +			/* For reading, UR is not an error status */
> > > +			*val = CFG_RD_UR_VAL;
> 
> I think the comment is incorrect.  Unsupported Request *is* an error
> status.  But most platforms log it and fabricate ~0 data
> (CFG_RD_UR_VAL) to return to the CPU, and I think that's what you're
> doing here.  So I think the code is fine, but the "not an error
> status" comment is wrong.

Ok, and what we should driver set as return value for pci_ops.read
callback in this case?

> Per the flowchart in PCIe r5.0, sec 6.2.5., fig 6-2, I think the
> hardware should be setting the "Unsupported Request Detected" bit in
> the Device Status register when this occurs.

Yes there is register in kernel's emulated PCIe bridge which at bit 19
has: Unsupported Request Detected - The core sets this bit to 1 when an
unsupported request is received. Write this bit to 1 to clear.

> > > +			strcomp_status = NULL;
> > > +		} else {
> > > +			strcomp_status = "UR";
> > > +		}
> > >  		break;
> > >  	case PIO_COMPLETION_STATUS_CRS:
> > > -		strcomp_status = "CRS";
> > > +		if (val) {
> > > +			/* For reading, CRS is not an error status */
> > > +			*val = CFG_RD_CRS_VAL;
> > 
> > Need Bjorn's input on this. I don't think this is what is expected from
> > from a root complex according to the PCI specifications (depending on
> > whether CSR software visibility is supported or not).
> > 
> > Here we are fabricating a CRS completion value for all PCI config read
> > transactions that are hitting a CRS completion status (and that's not
> > the expected behaviour according to the PCI specifications and I don't
> > think that's correct).
> 
> Right.  I think any config access (read or write) can be completed
> with a CRS completion (sec 2.3.1).
> 
> Per sec 2.3.2, when CRS SV (in Root Control register, sec 7.5.3.12) is
> enabled and a config read that includes both bytes of the Vendor ID
> receives a CRS completion, we must return 0x0001 for the Vendor ID and
> 0xff for any additional bytes.  Note that a config read of only the
> two Vendor ID bytes is legal and should receive 0x0001 data.
> 
> But if CRS SV is disabled, I think config reads that receive CRS
> completions should fail the normal way, i.e., fabricate ~0 data.

In PCIe base 2.0 is:

For other Configuration Requests, or when CRS Software Visibility is not
enabled, the Root Complex will generally re-issue the Configuration
Request until it completes with a status other than CRS as described in
Section 2.3.2.

So what should pci-aardvark driver in this case do? Return ~0 or re-send
this config read request (and how many times)?

Also this relates to previous discussion about PCI_EXP_RTCTL_CRSSVE:
https://lore.kernel.org/linux-pci/20210507152542.sd54lk7bk56qapf3@pali/

> > > +			strcomp_status = NULL;
> > > +		} else {
> > > +			strcomp_status = "CRS";
> > > +		}
> > >  		break;
> > >  	case PIO_COMPLETION_STATUS_CA:
> > >  		strcomp_status = "CA";
> > > @@ -490,6 +529,9 @@ static void advk_pcie_check_pio_status(struct advk_pcie *pcie)
> > >  		break;
> > >  	}
> > >  
> > > +	if (!strcomp_status)
> > > +		return 0;
> > > +
> > >  	if (reg & PIO_NON_POSTED_REQ)
> > >  		str_posted = "Non-posted";
> > >  	else
> > > @@ -497,6 +539,8 @@ static void advk_pcie_check_pio_status(struct advk_pcie *pcie)
> > >  
> > >  	dev_err(dev, "%s PIO Response Status: %s, %#x @ %#x\n",
> > >  		str_posted, strcomp_status, reg, advk_readl(pcie, PIO_ADDR_LS));
> > > +
> > > +	return -EFAULT;
> > >  }
> > >  
> > >  static int advk_pcie_wait_pio(struct advk_pcie *pcie)
> > > @@ -703,8 +747,17 @@ static int advk_pcie_rd_conf(struct pci_bus *bus, u32 devfn,
> > >  						 size, val);
> > >  
> > >  	if (advk_pcie_pio_is_running(pcie)) {
> > > -		*val = 0xffffffff;
> > > -		return PCIBIOS_SET_FAILED;
> > > +		/*
> > > +		 * For PCI_VENDOR_ID register, return Completion Retry Status
> > > +		 * so caller tries to issue the request again insted of failing
> > > +		 */
> > > +		if (where == PCI_VENDOR_ID) {
> > > +			*val = CFG_RD_CRS_VAL;
> > > +			return PCIBIOS_SUCCESSFUL;
> > 
> > Mmmm..here we are faking a CRS completion value to coerce the kernel
> > into believing a CRS completion was received (which is not necessarily
> > true) ?
> > 
> > if advk_pcie_pio_is_running(pcie) == true, is that an HW error ?
> > 
> > Lorenzo
> > 
> > > +		} else {
> > > +			*val = 0xffffffff;
> > > +			return PCIBIOS_SET_FAILED;
> > > +		}
> > >  	}
> > >  
> > >  	/* Program the control register */
> > > @@ -729,15 +782,27 @@ static int advk_pcie_rd_conf(struct pci_bus *bus, u32 devfn,
> > >  	advk_writel(pcie, 1, PIO_START);
> > >  
> > >  	ret = advk_pcie_wait_pio(pcie);
> > > +	if (ret < 0) {
> > > +		/*
> > > +		 * For PCI_VENDOR_ID register, return Completion Retry Status
> > > +		 * so caller tries to issue the request again instead of failing
> > > +		 */
> > > +		if (where == PCI_VENDOR_ID) {
> > > +			*val = CFG_RD_CRS_VAL;
> > > +			return PCIBIOS_SUCCESSFUL;
> > > +		} else {
> > > +			*val = 0xffffffff;
> > > +			return PCIBIOS_SET_FAILED;
> > > +		}
> > > +	}
> > > +
> > > +	/* Check PIO status and get the read result */
> > > +	ret = advk_pcie_check_pio_status(pcie, val);
> > >  	if (ret < 0) {
> > >  		*val = 0xffffffff;
> > >  		return PCIBIOS_SET_FAILED;
> > >  	}
> > >  
> > > -	advk_pcie_check_pio_status(pcie);
> > > -
> > > -	/* Get the read result */
> > > -	*val = advk_readl(pcie, PIO_RD_DATA);
> > >  	if (size == 1)
> > >  		*val = (*val >> (8 * (where & 3))) & 0xff;
> > >  	else if (size == 2)
> > > @@ -801,7 +866,9 @@ static int advk_pcie_wr_conf(struct pci_bus *bus, u32 devfn,
> > >  	if (ret < 0)
> > >  		return PCIBIOS_SET_FAILED;
> > >  
> > > -	advk_pcie_check_pio_status(pcie);
> > > +	ret = advk_pcie_check_pio_status(pcie, NULL);
> > > +	if (ret < 0)
> > > +		return PCIBIOS_SET_FAILED;
> > >  
> > >  	return PCIBIOS_SUCCESSFUL;
> > >  }
> > > -- 
> > > 2.20.1
> > >