On Tue, Feb 27, 2024 at 04:39:41PM +0800, Baochen Qiang wrote: > > > On 2/27/2024 3:15 PM, Manivannan Sadhasivam wrote: > > On Tue, Feb 27, 2024 at 02:36:12PM +0800, Baochen Qiang wrote: > > > MHI devices may not be destroyed during suspend/hibernation, so need > > > to unprepare/prepare MHI channels throughout the transition, this is > > > done by adding suspend/resume callbacks. > > > > > > The suspend callback is called in the late suspend stage, this means > > > MHI channels are still alive at suspend stage, and that makes it > > > possible for an MHI controller driver to communicate with others over > > > those channels at suspend stage. While the resume callback is called > > > in the early resume stage, for a similar reason. > > > > > > Tested-on: WCN6855 hw2.0 PCI WLAN.HSP.1.1-03125-QCAHSPSWPL_V1_V2_SILICONZ_LITE-3.6510.30 > > > > > > Signed-off-by: Baochen Qiang <quic_bqiang@xxxxxxxxxxx> > > > > Reviewed-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx> > > > > Could you please confirm if this patch can be applied separately and won't cause > > regression? > Just searched the kernel and found another two drivers using IPCR channels, > one is pci_epf_mhi_driver in drivers/pci/endpoint/functions/pci-epf-mhi.c > and the other is mhi_pci_driver in drivers/bus/mhi/host/pci_generic.c. > > For pci_epf_mhi_driver, this is not impacted because the devices for those > channels are attached to mhi_ep_bus_type while QRTR MHI driver attached to > mhi_bus_type. > > For mhi_pci_driver, I am afraid there would be regressions caused by this > patch. The regression is because when system suspends, mhi_pci_suspend() is > called and then qcom_mhi_qrtr_pm_suspend_late() called, however the latter > would fail because MHI is moved to M3 state by call mhi_pm_suspend() in > mhi_pci_suspend(). To address this, there might be two options: the first is > to move mhi_pci_suspend() to a late suspend stage which would be called > after qcom_mhi_qrtr_pm_suspend_late(). and the second is to avoid whole QRTR > suspend operation in such cases. I prefer the second one, because if MHI is > going to suspend, instead of power down, it is pointless to unprepare MHI > channels and re-prepare them after resume back. We can achieve this purpose > by adding a status_cb() to QRTR MHI driver which would be called when MHI > goes to low power mode. And then QRTR MHI driver could decide not to > suspend/resume if low power mode is notified. > > Your thoughts? > I'd rather query the MHI state of the device in suspend/resume callback using mhi_get_mhi_state(mhi_dev->mhi_cntrl) and if the device is found to be in M3 (suspend) state, then I'd skip prepare/unprepare part. Because this makes it clear that, if the device is in suspend state, then no need for the client drivers to unprepare/prepare the channels. - Mani -- மணிவண்ணன் சதாசிவம்
