[PATCH V9 1/2] PCI: handle CRS returned by device after FLR

[Bjorn Helgaas]

On Tue, Aug 08, 2017 at 08:57:24PM -0400, Sinan Kaya wrote:
> Sporadic reset issues have been observed with Intel 750 NVMe drive by
> writing to the reset file in sysfs in a loop. The sequence of events
> observed is as follows:
> 
> - perform a Function Level Reset (FLR)
> - sleep up to 1000ms total
> - read ~0 from PCI_COMMAND
> - warn that the device didn't return from FLR
> - touch the device before it's ready

What's the driver-visible or user-visible effect of touching the
device before it's ready?  This sequence is sort of like the joke
without the punch line :)

> An endpoint is allowed to issue Configuration Request Retry Status (CRS)
> following a FLR request to indicate that it is not ready to accept new
> requests. CRS is defined in PCIe r3.1, sec 2.3.1. Request Handling Rules
> and CRS usage in FLR context is mentioned in PCIe r3.1a, sec 6.6.2.
> Function-Level Reset.

You reference both PCIe r3.1 and r3.1a.  Is there something new in
r3.1a in this area?  If not, just reference r3.1 for both cases.

> A CRS indication will only be given if the address to be read is vendor ID
> register. pci_bus_read_dev_vendor_id() knows how to deal with CRS returned
> 0xFFFF0001 value and will continue polling until a value other than
> 0xFFFF0001 is returned within a given timeout.
> 
> Try to discover device presence via CRS first. If device is not found, fall
> through to old behavior.
>
> Signed-off-by: Sinan Kaya <okaya at codeaurora.org>
> ---
>  drivers/pci/pci.c | 26 ++++++++++++++++++++++++--
>  1 file changed, 24 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index af0cc34..4366299 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -3821,17 +3821,39 @@ static void pci_flr_wait(struct pci_dev *dev)
>  {
>  	int i = 0;
>  	u32 id;
> +	bool ret;
> +
> +	/*
> +	 * Don't touch the HW before waiting 100ms. HW has to finish
> +	 * within 100ms according to PCI Express Base Specification
> +	 * Revision 3.1 Section 6.6.2: Function-Level Reset (FLR).
> +	 */
> +	msleep(100);
> +
> +	/*
> +	 * See if we can find a device via CRS first. Physical functions
> +	 * return from here if found, Virtual functions fall through as
> +	 * they return ~0 on vendor id read once CRS is completed.
> +	 */
> +	ret = pci_bus_read_dev_vendor_id(dev->bus, dev->devfn, &id,
> +					 60000);
> +	if (ret)
> +		return;

Alex was concerned that pci_bus_read_dev_vendor_id() could return
false ("no device here") with an ID value of ~0 for a functional VF
[1].

I think that is indeed possible:

  - a VF that is ready will return 0xffff as both Vendor ID and Device
    ID (SR-IOV r1.1, sec 3.4.1.1, .2), so the very first read in
    pci_bus_read_dev_vendor_id() would see 0xffffffff and return
    "false".

  - a VF that needs more time will return CRS and we'll loop in
    pci_bus_read_dev_vendor_id() until it becomes ready, and we'll
    return "true".

Falling into the code below for the "false" case probably will work,
but it's a little bit ugly because (a) we're using two mechanisms to
figure out when the device is ready for config requests, and (b) we
have to worry about VFs both in pci_bus_read_dev_vendor_id() and here
in the caller.

Here's an idea to make pci_bus_read_dev_vendor_id() work for both VFs
and PFs.  It can't distinguish the 0xffffffff from a VF vs one from a
non-existent device, but the caller might be able to pass that
information in, e.g., when we're enumerating and don't know whether
the device exists, we don't have a pci_dev and would use this:

  pci_bus_read_dev_vendor_id(bus, devfn, &l, 60000, 0)

While places where we do have a pci_dev and expect the device to exist
(e.g., waiting after a reset), would do this:

  pci_bus_read_dev_vendor_id(bus, devfn, &l, 60000, dev->is_virtfn)

And we would skip the 0xffffffff check for VFs, e.g.,

  bool pci_bus_read_dev_vendor_id(struct pci_bus *bus, int devfn, u32 *l,
                                  int crs_timeout, int is_vf)
  {
    if (pci_bus_read_config_dword(bus, devfn, PCI_VENDOR_ID, l))
      return false;

    if (!is_vf &&
        (*l == 0xffffffff || *l == 0x00000000 ||
         *l == 0x0000ffff || *l == 0xffff0000))
        return false;

    while ((*l & 0xffff) == 0x0001) {
      ...
      if (pci_bus_read_config_dword(bus, devfn, PCI_VENDOR_ID, l))
	return false;
    }

    return true;
  }

Would that work?

I don't know if this would be the best solution.  This is a messy
area.  We're relying on the host bridge to fabricate the 0xffff data
for non-existent devices, and most (maybe all) do that, but I don't
think that's actually in the spec.

> +	pci_read_config_dword(dev, PCI_COMMAND, &id);
> +	if (id != ~0)
> +		return;
>  
>  	do {
>  		msleep(100);
>  		pci_read_config_dword(dev, PCI_COMMAND, &id);
> -	} while (i++ < 10 && id == ~0);
> +	} while (i++ < 9 && id == ~0);
>
>  	if (id == ~0)
>  		dev_warn(&dev->dev, "Failed to return from FLR\n");
>  	else if (i > 1)
>  		dev_info(&dev->dev, "Required additional %dms to return from FLR\n",
> -			 (i - 1) * 100);
> +			 i * 100);
>  }
>  
>  /**

[1] http://lkml.kernel.org/r/20170221135138.791ba4e2@t450s.home