Re: [PATCH net-next 6/6] mlxsw: pci: Add support for new reset flow

From
Alex Williamson <hidden>
Date
2023-03-30 18:50:49
Also in
linux-pci
Reply to
Ido Schimmel, 2023-03-30
On Thu, 30 Mar 2023 11:26:30 +0300
Ido Schimmel [off-list ref] wrote:
On Wed, Mar 29, 2023 at 11:10:01AM -0600, Alex Williamson wrote:
quoted
I think we don't have it because it's unclear how it's actually
different from a secondary bus reset from the bridge control register,
which is what "bus" would do when selected for the example above.  Per
the spec, both must cause a hot reset.  It seems this device needs a
significantly longer delay though.  
Assuming you are referring to the 2ms sleep in
pci_reset_secondary_bus(), then yes. In our case, after disabling the
link on the downstream port we need to wait for 500ms before enabling
it.
quoted
Note that hot resets can be generated by a userspace driver with
ownership of the device and will make use of the pci-core reset
mechanisms.  Therefore if there is not a device specific reset, we'll
use the standard delays and the device ought not to get itself wedged
if the link becomes active at an unexpected point relative to a
firmware update.  This might be a point in favor of a device specific
reset solution in pci-core.  Thanks,  
I assume you referring to something like this:

# echo 1 > /sys/class/pci_bus/0000:03/device/0000:03:00.0/reset

Doesn't seem to have any effect (network ports remain up, at least).
Anyway, this device is completely managed by the kernel, not a user
space driver. I'm not aware of anyone using this method to reset the
device.
The pci-sysfs reset attribute is only meant to reset the linked device,
so if this is a single function device then it might be accessing bus
reset, but it also might be using FLR or PM reset.  There's a
reset_method attribute to determine and select.

In any case, if the device is unaffected, that suggests we're dealing
with a device that doesn't comply with PCIe reset standards, which
might suggests it needs a device specific reset or to flag broken reset
methods regardless.

Note that QEMU is a vfio-pci userspace driver, so assigning the device
to a VM, where kernel drivers in the guest are managing the device is a
use case of userspace drivers which should have a functional reset
mechanism to avoid data leakage between userspace sessions.
 
If I understand Bjorn and you correctly, we have two options:

1. Keep the current implementation inside the driver.

2. Call __pci_reset_function_locked() from the driver and move the link
toggling to drivers/pci/quirks.c as a "device_specific" method.

Personally, I don't see any benefit in 2, but we can try to implement
it, see if it even works and then decide.
The second option enables use cases like above, where the PCI-core can
perform an effective reset of the device rather than embedding that
into a specific driver.  Even if not intended as a primary use case,
it's a more complete solution and avoids potentially unhappy users that
assume such use cases are available.  Thanks,

Alex
Thread (25 messages) 25 messages, 7 authors, last reply 2023-04-13
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help