On Wed, Apr 24, 2019 at 8:01 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > > On Wed, Apr 24, 2019 at 05:58:19PM +0300, Mika Westerberg wrote: > > On Wed, Apr 24, 2019 at 09:11:48AM -0500, Bjorn Helgaas wrote: > > > - Maybe the PCI sysfs accessors (pci_mmap_resource(), etc) should > > > turn off runtime PM? If we allow mmap of a BAR and then put the > > > device in D3hot, that seems like a bug that could affect lots of > > > things. But maybe that's already done magically elsewhere? > > > > IIRC there is no PM magic happening for MMIO userspace accesses. > > > > What you suggest above sounds like a good way to fix it. We already do > > similar for config space access from userspace (if the device is in > > D3cold) so definitely makes sense to do the same for MMIO. However, I > > don't think we need to disable runtime PM - it should be enough to > > increase the reference count (pm_runtime_get_sync() and friends) during > > the time the MMIO resource is mmapped. > > OK, so if I understand correctly this would be basically adding > pm_runtime_get_sync(dev) in pci_mmap_resource(). I don't know what > the unmap path is, but there would have to be a matching > pm_runtime_put() somewhere. Right. > And a similar change in the read/write path for /sys/.../resource<N>; > I think this must be related to the sysfs_create_bin_file() call in > pci_create_attr(), but I don't see the path where the actual > read/write to the device is done. > > And probably something similar should be done in pci_resource_io(), > pci_map_rom(), and pci_unmap_rom(). In general, every path in which there is a memory or IO address space access requires pm_runtime_get_sync()/pm_runtime_put() around it as these accesses are only guaranteed to work in D0. Cheers, Rafael
