Re: [PATCH 2/2] pci: Don't set RCB bit in LNKCTL if the upstream bridge hasn't

[Bjorn Helgaas]

Hi Johannes,

On Wed, Nov 02, 2016 at 04:35:52PM -0600, Johannes Thumshirn wrote:
> The Read Completion Boundary (RCB) bit must only be set on a device or
> endpoint if it is set on the root complex.

I propose the following slightly modified patch.  The interesting
difference is that your patch only touches the _HPX "OR" mask, so it
refrains from *setting* RCB in some cases, but it never actually
*clears* it.  The only time we clear RCB is when the _HPX "AND" mask
has RCB == 0.

My intent below is that we completely ignore the _HPX RCB bits, and we
set an Endpoint's RCB if and only if the Root Port's RCB is set.

I made an ugly ASCII table to think about the cases:

      Root   EP    _HPX  _HPX     Final Endpoint RCB state
      Port  (init)  AND   OR     (curr)  (yours)  (mine)
  0)   0     0      0    0          0       0       0
  1)   0     0      0    1          1       0       0
  2)   0     0      1    0          0       0       0
  3)   0     0      1    1          1       0       0
  4)   0     1      0    0          0       0       0
  5)   0     1      0    1          1       0       0
  6)   0     1      1    0          1       1       0
  7)   0     1      1    1          1       1       0
  8)   1     0      0    0          0       0       1
  9)   1     0      0    1          1       1       1
  A)   1     0      1    0          0       0       1
  B)   1     0      1    1          1       1       1
  C)   1     1      0    0          0       0       1
  D)   1     1      0    1          1       1       1
  E)   1     1      1    0          1       1       1
  F)   1     1      1    1          1       1       1

Cases 0-7 should all result in the Endpoint RCB being zero because the
Root Port RCB is zero.  Case 1 is the bug you're fixing.  Cases 3 & 5
are similar hypothetical bugs your patch also fixes.

Cases 6 & 7, where firmware left the Endpoint RCB set and _HPX didn't
tell us to clear it, are hypothetical firmware bugs that your patch
wouldn't fix.

In cases 8, A, and C, we currently leave the Endpoint RCB cleared,
either because firmware left it clear and _HPX didn't tell us to set
it (8 and A), or because firmware set it but _HPX told us to clear it
(C).

One could argue that 8, A, and C should stay as they currently are, as
a way for _HPX to work around hardware bugs, e.g., a Root Port that
advertises a 128-byte RCB but doesn't actually support it.  I didn't
bother with that and set the Endpoint's RCB to 128 in all cases when
the Root Port claims to support it.

It'd be great if you could test this and comment.

If you get a chance, collect the /proc/iomem contents, too.  That's
not for this bug; it's because I'm curious about the

  ERST: Can not request [mem 0xb928b000-0xb928cbff] for ERST
  
problem in your dmesg log.


commit da567a6c67cb5c5a55db7b26b46e9b7e9af69299
Author: Johannes Thumshirn <jthumshirn@xxxxxxx>
Date:   Wed Nov 2 16:35:52 2016 -0600

    PCI: Set Read Completion Boundary to 128 iff Root Port supports it
    
    Per PCIe spec r3.0, sec 2.3.1.1, the Read Completion Boundary (RCB)
    determines the naturally aligned address boundaries on which a Read Request
    may be serviced with multiple Completions:
    
      - For a Root Complex, RCB is 64 bytes or 128 bytes
        This value is reported in the Link Control Register
    
        Note: Bridges and Endpoints may implement a corresponding command bit
        which may be set by system software to indicate the RCB value for the
        Root Complex, allowing the Bridge/Endpoint to optimize its behavior
        when the Root Complexâs RCB is 128 bytes.
    
      - For all other system elements, RCB is 128 bytes
    
    Per sec 7.8.7, if a Root Port only supports a 64-byte RCB, the RCB of all
    downstream devices must be clear, indicating an RCB of 64 bytes.  If the
    Root Port supports a 128-byte RCB, we may optionally set the RCB of
    downstream devices so they know they can generate larger Completions.
    
    Some BIOSes supply an _HPX that tells us to set RCB, even though the Root
    Port doesn't have RCB set, which may lead to Malformed TLP errors if the
    Endpoint generates completions larger than the Root Port can handle.
    
    The IBM x3850 X6 with BIOS version -[A8E120CUS-1.30]- 08/22/2016 supplies
    such an _HPX and a Mellanox MT27500 ConnectX-3 device fails to initialize:
    
      mlx4_core 0000:41:00.0: command 0xfff timed out (go bit not cleared)
      mlx4_core 0000:41:00.0: device is going to be reset
      mlx4_core 0000:41:00.0: Failed to obtain HW semaphore, aborting
      mlx4_core 0000:41:00.0: Fail to reset HCA
      ------------[ cut here ]------------
      kernel BUG at drivers/net/ethernet/mellanox/mlx4/catas.c:193!
    
    After 6cd33649fa83 ("PCI: Add pci_configure_device() during enumeration")
    and 7a1562d4f2d0 ("PCI: Apply _HPX Link Control settings to all devices
    with a link"), we apply _HPX settings to *all* devices, not just those
    hot-added after boot.
    
    Before 7a1562d4f2d0, we didn't touch the Mellanox RCB, and the device
    worked.  After 7a1562d4f2d0, we set its RCB to 128, and it failed.
    
    Set the RCB to 128 iff the Root Port supports a 128-byte RCB.  Otherwise,
    set RCB to 64 bytes.  This effectively ignores what _HPX tells us about
    RCB.
    
    [bhelgaas: changelog, clear RCB if not set for Root Port]
    Fixes: 6cd33649fa83 ("PCI: Add pci_configure_device() during enumeration")
    Fixes: 7a1562d4f2d0 ("PCI: Apply _HPX Link Control settings to all devices with a link")
    Link: https://bugzilla.kernel.org/show_bug.cgi?id=187781
    Signed-off-by: Johannes Thumshirn <jthumshirn@xxxxxxx>
    Signed-off-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
    CC: stable@xxxxxxxxxxxxxxx	# v3.18+

diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index ab00267..8854161 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -1439,6 +1439,21 @@ static void program_hpp_type1(struct pci_dev *dev, struct hpp_type1 *hpp)
 		dev_warn(&dev->dev, "PCI-X settings not supported\n");
 }
 
+static bool pcie_root_rcb_set(struct pci_dev *dev)
+{
+	struct pci_dev *rp = pcie_find_root_port(dev);
+	u16 lnkctl;
+
+	if (!rp)
+		return false;
+
+	pcie_capability_read_word(rp, PCI_EXP_LNKCTL, &lnkctl);
+	if (lnkctl & PCI_EXP_LNKCTL_RCB)
+		return true;
+
+	return false;
+}
+
 static void program_hpp_type2(struct pci_dev *dev, struct hpp_type2 *hpp)
 {
 	int pos;
@@ -1468,9 +1483,19 @@ static void program_hpp_type2(struct pci_dev *dev, struct hpp_type2 *hpp)
 			~hpp->pci_exp_devctl_and, hpp->pci_exp_devctl_or);
 
 	/* Initialize Link Control Register */
-	if (pcie_cap_has_lnkctl(dev))
+	if (pcie_cap_has_lnkctl(dev)) {
+
+		/*
+		 * If the Root Port supports Read Completion Boundary of
+		 * 128, set RCB to 128.  Otherwise, clear it.
+		 */
+		hpp->pci_exp_lnkctl_and |= PCI_EXP_LNKCTL_RCB;
+		if (pcie_root_rcb_set(dev))
+			hpp->pci_exp_lnkctl_or |= PCI_EXP_LNKCTL_RCB;
+
 		pcie_capability_clear_and_set_word(dev, PCI_EXP_LNKCTL,
 			~hpp->pci_exp_lnkctl_and, hpp->pci_exp_lnkctl_or);
+	}
 
 	/* Find Advanced Error Reporting Enhanced Capability */
 	pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ERR);