On Thu, Oct 8, 2020 at 12:29 AM Dan Williams dan.j.williams@intel.com wrote:
On Wed, Oct 7, 2020 at 3:23 PM Dan Williams dan.j.williams@intel.com wrote:
On Wed, Oct 7, 2020 at 12:49 PM Daniel Vetter daniel.vetter@ffwll.ch wrote:
On Wed, Oct 7, 2020 at 9:33 PM Dan Williams dan.j.williams@intel.com wrote:
On Wed, Oct 7, 2020 at 11:11 AM Daniel Vetter daniel.vetter@ffwll.ch wrote:
Since 3234ac664a87 ("/dev/mem: Revoke mappings when a driver claims the region") /dev/kmem zaps ptes when the kernel requests exclusive acccess to an iomem region. And with CONFIG_IO_STRICT_DEVMEM, this is the default for all driver uses.
Except there's two more ways to access pci bars: sysfs and proc mmap support. Let's plug that hole.
Ooh, yes, lets.
For revoke_devmem() to work we need to link our vma into the same address_space, with consistent vma->vm_pgoff. ->pgoff is already adjusted, because that's how (io_)remap_pfn_range works, but for the mapping we need to adjust vma->vm_file->f_mapping. Usually that's done at ->open time, but that's a bit tricky here with all the entry points and arch code. So instead create a fake file and adjust vma->vm_file.
I don't think you want to share the devmem inode for this, this should be based off the sysfs inode which I believe there is already only one instance per resource. In contrast /dev/mem can have multiple inodes because anyone can just mknod a new character device file, the same problem does not exist for sysfs.
But then I need to find the right one, plus I also need to find the right one for the procfs side. That gets messy, and I already have no idea how to really test this. Shared address_space is the same trick we're using in drm (where we have multiple things all pointing to the same underlying resources, through different files), and it gets the job done. So that's why I figured the shared address_space is the cleaner solution since then unmap_mapping_range takes care of iterating over all vma for us. I guess I could reimplement that logic with our own locking and everything in revoke_devmem, but feels a bit silly. But it would also solve the problem of having mutliple different mknod of /dev/kmem with different address_space behind them. Also because of how remap_pfn_range works, all these vma do use the same pgoff already anyway.
True, remap_pfn_range() makes sure that ->pgoff is an absolute physical address offset for all use cases. So you might be able to just point proc_bus_pci_open() at the shared devmem address space. For sysfs it's messier. I think you would need to somehow get the inode from kernfs_fop_open() to adjust its address space, but only if the bin_file will ultimately be used for PCI memory.
Just read the code a bit more, and for proc it's impossible. There's only a single file, and before you mmap it you have to call a few ioctl to select the right pci resource on that device you want to mmap. Which includes legacy ioport stuff, and at least for now those don't get revoked (maybe they should, but I'm looking at iomem here now). Setting the mapping too early in ->open means that on architectures which can do ioport as mmaps (not many, but powerpc is among them) we'd shoot down these mmaps too.
Looking at the code there's the generic implementation, which consults pci_iobar_pfn. And the only other implementation for sparc looks similar, they separate iomem vs ioport through different pfn. So I think this should indeed work.
To me this seems like a new sysfs_create_bin_file() flavor that registers the file with the common devmem address_space.
Hm I think we could just add a i_mapping member to bin_attributes and let the normal open code set that up for us. That should work. mmapable binary sysfs file is already a similar special case. -Daniel
-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch