On 2020/12/3 4:18, Bjorn Helgaas wrote:
On Wed, Dec 02, 2020 at 05:18:12PM +0800, Chiqijun wrote:
On 2020/11/30 23:46, Alex Williamson wrote:
On Sat, 28 Nov 2020 17:29:19 -0600
Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
On Sat, Nov 28, 2020 at 02:18:25PM +0800, Chiqijun wrote:
When multiple VFs do FLR at the same time, the firmware is
processed serially, resulting in some VF FLRs being delayed more
than 100ms, when the virtual machine restarts and the device
driver is loaded, the firmware is doing the corresponding VF
FLR, causing the driver to fail to load.
To solve this problem, add host and firmware status synchronization
during FLR.
Is this because the Huawei Intelligent NIC isn't following the spec,
or is it because Linux isn't correctly waiting for the FLR to
complete?
Seems like a spec compliance issue, I don't recall anything in the spec
about coordinating FLR between VFs.
The spec stipulates that the FLR time of a single VF does not exceed 100ms,
but when multiple VMs are reset concurrently in Linux, there will be
multiple VF parallel FLRs, VF of Huawei Intelligent NIC
FLR will exceed 100ms in this case.
Can you somehow just serialize Huawei Intelligent NIC FLR and
otherwise use the normal FLR path instead of the iomap, PCI_COMMAND
fiddling, and huge timeout below?
FLR triggering is not controlled by Huawei Intelligent NIC. For example,
multiple VFs are assigned to different VMs, one VM corresponds to a qemu
process, multiple VMs perform life cycle operations (such as reset) at
the same time, and FLR triggered by different Qemu processes will be
parallel.
The FLR of a single VF of Huawei Intelligent NIC is within 100ms (for
example, 50ms), while the regular FLR process will sleep for 100ms. If
the VF FLR is forced to be serialized (for example, adding a lock), the
overall FLR time will become longer and the VF FLR between different
cards also become serial, the actual effect will be worse than setting
huge timeout.
If this is a Huawei Intelligent NIC defect, is there documentation
somewhere (errata) that you can reference? Will it be fixed in future
designs, so we don't have to add future Device IDs to the quirk?
Signed-off-by: Chiqijun <chiqijun@xxxxxxxxxx>
---
drivers/pci/quirks.c | 67 ++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 67 insertions(+)
diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
index f70692ac79c5..bd6236ea9064 100644
--- a/drivers/pci/quirks.c
+++ b/drivers/pci/quirks.c
@@ -3912,6 +3912,71 @@ static int delay_250ms_after_flr(struct pci_dev *dev, int probe)
return 0;
}
+#define PCI_DEVICE_ID_HINIC_VF 0x375E
+#define HINIC_VF_FLR_TYPE 0x1000
+#define HINIC_VF_OP 0xE80
+#define HINIC_OPERATION_TIMEOUT 15000
+
+/* Device-specific reset method for Huawei Intelligent NIC virtual functions */
+static int reset_hinic_vf_dev(struct pci_dev *pdev, int probe)
+{
+ unsigned long timeout;
+ void __iomem *bar;
+ u16 old_command;
+ u32 val;
+
+ if (probe)
+ return 0;
+
+ bar = pci_iomap(pdev, 0, 0);
+ if (!bar)
+ return -ENOTTY;
+
+ pci_read_config_word(pdev, PCI_COMMAND, &old_command);
+
+ /*
+ * FLR cap bit bit30, FLR ACK bit: bit18, to avoid big-endian conversion
+ * the big-endian bit6, bit10 is directly operated here
+ */
+ val = readl(bar + HINIC_VF_FLR_TYPE);
+ if (!(val & (1UL << 6))) {
+ pci_iounmap(pdev, bar);
+ return -ENOTTY;
+ }
I don't know exactly what this is testing, but it seems like a
feature/capability test that can fail, why is it not done as part of
the probe? Can we define bit 6 with a macro? Same for bit 10 in the
VF op register below.
The firmware of Huawei Intelligent NIC does not support this feature in the
old version. here is the reading ability to determine whether the firmware
supports it.
In the next patch, I will add a comment here and replace bit 6 and bit 10
with macro definitions.
+
+ val = readl(bar + HINIC_VF_OP);
+ val = val | (1UL << 10);
+ writel(val, bar + HINIC_VF_OP);
+
+ /* Perform the actual device function reset */
+ pcie_flr(pdev);
+
+ pci_write_config_word(pdev, PCI_COMMAND,
+ old_command | PCI_COMMAND_MEMORY);
+
+ /* Waiting for device reset complete */
+ timeout = jiffies + msecs_to_jiffies(HINIC_OPERATION_TIMEOUT);
Yikes, 15s timeout!
Huawei Intelligent NIC supports a maximum of 496 VFs, so the total timeout
period is set to 15s, which will not reach the timeout time under normal
circumstances.
+ do {
+ val = readl(bar + HINIC_VF_OP);
+ if (!(val & (1UL << 10)))
+ goto reset_complete;
+ msleep(20);
+ } while (time_before(jiffies, timeout));
+
+ val = readl(bar + HINIC_VF_OP);
+ if (!(val & (1UL << 10)))
+ goto reset_complete;
+
+ pci_warn(pdev, "Reset dev timeout, flr ack reg: %x\n",
+ be32_to_cpu(val));
+
+reset_complete:
+ pci_write_config_word(pdev, PCI_COMMAND, old_command);
+ pci_iounmap(pdev, bar);
+
+ return 0;
+}
+
static const struct pci_dev_reset_methods pci_dev_reset_methods[] = {
{ PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_82599_SFP_VF,
reset_intel_82599_sfp_virtfn },
@@ -3923,6 +3988,8 @@ static const struct pci_dev_reset_methods pci_dev_reset_methods[] = {
{ PCI_VENDOR_ID_INTEL, 0x0953, delay_250ms_after_flr },
{ PCI_VENDOR_ID_CHELSIO, PCI_ANY_ID,
reset_chelsio_generic_dev },
+ { PCI_VENDOR_ID_HUAWEI, PCI_DEVICE_ID_HINIC_VF,
+ reset_hinic_vf_dev },
{ 0 }
};
--
2.17.1
.
.
