On Thu, Aug 04, 2022 at 01:05:04PM -0400, Jim Quinlan wrote: > On Mon, Aug 1, 2022 at 6:19 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > > On Tue, Jul 26, 2022 at 04:41:09PM -0700, Florian Fainelli wrote: > > > On 7/26/22 15:03, Bjorn Helgaas wrote: > > > > On Mon, Jul 25, 2022 at 11:12:49AM -0400, Jim Quinlan wrote: > > > >> ... > > > >> Jim Quinlan (7): > > > >> PCI: brcmstb: Remove unnecessary forward declarations > > > >> PCI: brcmstb: Split brcm_pcie_setup() into two funcs > > > >> PCI: brcmstb: Gate config space access on link status > > > >> PCI: brcmstb: Add mechanism to turn on subdev regulators > > > >> PCI: brcmstb: Add control of subdevice voltage regulators > > > >> PCI: brcmstb: Do not turn off WOL regulators on suspend > > > >> PCI: brcmstb: Have .map_bus function names end with 'map_bus' > > > >> > > > >> drivers/pci/controller/pcie-brcmstb.c | 476 ++++++++++++++++++-------- > > > >> 1 file changed, 341 insertions(+), 135 deletions(-) > > > > > > > > I reworked these and put them on pci/ctrl/brcm for v5.20. This is a > > > > proposal, not something set in stone. But time is of the essence to > > > > figure out how we want to proceed. > > > > > > > > I changed a lot of stuff and it's likely I broke something in the > > > > process, so please take a look and test this out. Here's an outline > > > > of what I changed: > > > > > > > > - Moved the config access "link up" check earlier because it's not > > > > related to the power regulator patches. > > > > > > > > - Changed config access "link up" checks to use PCIE_ECAM_REG() > > > > instead of hard-coding 0xfff masks. The 32-bit accessors already > > > > mask out the low two bits, so we don't need to do that here. > > > > > > > > - Squashed pci_subdev_regulators_add_bus() directly into > > > > brcm_pcie_add_bus() for readability. Similarly for > > > > pci_subdev_regulators_remove_bus(). > > > > > > > > - This makes a clear split between: > > > > > > > > * A patch that adds get/enable of regulators, and starting the > > > > link after enabling regulators, and > > > > > > > > * A patch that disables/enables regulators for suspend/resume. > > > > > > > > - Since we only support one set of subregulator info (for one Root > > > > Port, and brcm_pcie_suspend_noirq() depends on this since it uses > > > > the pcie->sr pointer), use pcie->sr always instead of > > > > dev->driver_data. > > > > > > > > - Squashed wakeup device checking into the suspend/resume patch so > > > > there's not a time when suspend might turn off power to a wakeup > > > > device. > > > > > > > > - Renamed brcm_pcie_map_bus32() to brcm7425_pcie_map_bus() so it > > > > ends in "_map_bus()" like other drivers. Also, > > > > brcm7425_pcie_map_bus() doesn't actually depend on the 32-bitness. > > > > > > Attached is the diff between Jim's and your branch just so it is easier to see what moved around. > > > > > > Initial testing on an ARCH_BRCMSTB system with PCIe appears to be good, we don't have any regulator on that board so the dummy ones get picked up which is expected. Same thing with a Raspberry Pi 4B system. > > > > > > I could unbind and bind again and there were no reference count leaks on the regulators, so this looks good to me. > > > > > > Tested-by: Florian Fainelli <f.fainelli@xxxxxxxxx> > > > > > > of course, we should have Jim's test results as well as Cyril's ideally to make sure there are no regressions on the CM4 board. > > > > Cyril, any chance you could test this to be sure it fixes the problem > > you reported? This is in -next and hopefully headed for v5.20/v6.0 > > soon. > > Cyril sent me an email about a week ago saying that he probably > wouldn't have the bandwidth to test this. > I immediately ordered an overpriced CM4 via Ebay and it recently > arrived. I'm happy to say that this > patchset tests successfully, w/ or w/o a device in the slot. Great, thanks a lot for testing this! > That being said, there is an old device, when paired with the CM4, > works with RPi Linux but not with upstream Linux. It is unrelated > to this patchset; i.e. it fails w/ or w/o this patchset applied. I > know the reason for this failure: the current driver > assumes clkreq# asserted, which is true for all STB boards. I can add > a patch for this now or in the next release > cycle, your choice. It's too late for v5.20-rc1, but if this would fix a regression or otherwise exceptional bug, Lorenzo might still consider it for the v5.20. Bjorn
