[PATCH v9] PCI: Add device-specific reset for Qualcomm devices

[Manivannan Sadhasivam]

On Mon, Jun 22, 2026 at 04:08:22PM -0600, Alex Williamson wrote:
> On Mon, 22 Jun 2026 18:22:39 +0200
> Manivannan Sadhasivam <mani at kernel.org> wrote:
> 
> > On Thu, Jun 18, 2026 at 08:33:08AM +0200, Jose Ignacio Tornos Martinez wrote:
> > > Hi Mani,
> > > 
> > > Let me clarify the exact scenario and where the reset is necessary:
> > > 
> > > * For the commented WiFi devices (WCN6855/WCN7850):
> > > 
> > > Standard VFIO passthrough flow (this works fine):
> > >   1. Unbind native driver (ath11k/ath12k/MHI)
> > >   2. Bind vfio-pci driver
> > >   3. Assign device to VM
> > >   4. VM boots, loads its own driver → device works perfectly
> > >   5. VM shuts down cleanly → device can be reassigned → works fine
> > > 
> > > The problem occurs with unclean VM termination:
> > >   1. VM crashes or is force-terminated
> > >   2. VFIO tries to reset the device before reassignment
> > >   3. Without a working PCI reset method, reset fails
> > >   4. Device stuck in undefined state → cannot be reassigned to another VM
> > >   
> > >      Unbinding the driver again doesn't help because the device hardware
> > >      itself is in a bad state. From hypervisor:
> > >      $ lspci -vvv -s 0000:03:00.0
> > >         03:00.0 Network controller: Qualcomm Technologies, Inc (rev ff) (prog-if ff)
> > >             !!! Unknown header type 7f
> > >      And a full host power-cycle is necessary to recover.
> > >        
> > 
> > Can you try the global reset available in the WLAN device BAR space?
> > 
> > WCN6855: https://git.kernel.org/pub/scm/linux/kernel/git/pci/pci.git/tree/drivers/net/wireless/ath/ath11k/pci.c#n193
> > WCN7850: https://git.kernel.org/pub/scm/linux/kernel/git/pci/pci.git/tree/drivers/net/wireless/ath/ath12k/pci.c#n182
> > 
> > > * For the commented modem devices (SDX62/SDX65): 
> > > 
> > > Even worse because it fails during the first VM boot without proper reset
> > > capability, standard VFIO passthrough flow:
> > >   1. Unbind native driver (MHI)
> > >   2. Bind vfio-pci driver
> > >   3. Assign device to VM
> > >   4. VM boots, loads its own driver and crashes:
> > >      [   24.024165] mhi mhi0: Device failed to enter MHI Ready
> > >      [   24.024168] mhi mhi0: MHI did not enter READY state
> > >      
> > >      Unbind/rebind attempts fail:
> > >      [  352.643601] mhi mhi0: Requested to power ON
> > >      [  352.643611] mhi mhi0: Power on setup success
> > >      [  373.442954] mhi mhi0: Device failed to clear MHI Reset
> > >      [  373.442970] mhi mhi0: MHI did not enter READY state
> > >      And requires a full host power cycle to recover,
> > >      even outside of VFIO scenarios.
> > > 
> > > * MHI Host driver's remove callback may handle clean software state
> > > teardown, but it doesn't provide a PCI reset capability that VFIO can
> > > invoke. VFIO needs a reset method registered in the PCI reset hierarchy
> > > (device_specific, pm, flr, bus, etc.). VFIO invokes this reset both during
> > > initial device binding (before the VM starts) and when reassigning the
> > > device between VMs - without a working reset method, the device cannot
> > > reach a clean state for initialization.
> > >   
> > 
> > Likewise, there is a SoC reset available in the modem BAR space. You can try it:
> > https://git.kernel.org/pub/scm/linux/kernel/git/pci/pci.git/tree/drivers/bus/mhi/host/main.c#n178
> > 
> > If these works, then you can hook these in the device_specific reset callback.
> 
> 
> These look promising as simple flows to implement in a device specific
> reset: save command register, set memory enable, ioremap BAR space,
> match read/write/delay sequences of reset function and caller, iounmap,
> restore command.
> 
> Note the delay in this latter reset is in the caller.  It's also
> surprising that none of these implement a read to flush the posted
> write that initiates the reset.  I wonder if that contributes to the 2s
> delay in the latter example.
> 

While I agree that the absence of read-back after the posted write is a problem,
especially with a delay afterwards, the delay itself is required to let the
device reset itself and do a warm boot.

I'll submit fixes to do a read-back after triggering the reset to ensure that
the write gets flushed to the device before the delay.

- Mani

-- 
மணிவண்ணன் சதாசிவம்