On Thu, 2024-06-20 at 14:06 +0200, Niklas Schnelle wrote: > On Wed, 2024-06-19 at 21:09 -0700, Christoph Hellwig wrote: > > On Wed, Jun 19, 2024 at 12:56:47PM +0200, Niklas Schnelle wrote: > > > In short, the ISM BAR 0 is stupidly large but this is > > > intentional. It not fitting in the VMAP is simply the least crazy > > > filter I could come up with to keep the ISM device from causing > > > trouble for use of vfio-pci mmap() for other, normal, PCI > > > devices. > > > > Then maybe add a PCI quirk to prevent mapping it. This would also > > affect the sysfs resource0 file unless I'm missing something. > > > > The resource<N> files are globally disabled on s390x due to lack of > HAVE_PCI_MMAP/ARCH_GENERIC_PCI_MMAP_RESOURCE at the moment. I might > change that in the future with a analogous argument as for > VFIO_PCI_MMAP but for its not there. Once we add it we of course need > the same special ISM treatment there too. > > Looking at the existing PCI quirks I don't see anything that would > fit so I'm guessing you would want to add a new quirk? As I > understand it we would then have to export something like a > is_pci_mmap_broken(pdev) function while currently the only quirk > function that seems to be exported is pxi_fixup_device(). But then > what happens if CONFIG_PCI_QUIRKS=n? Also the header comment in > drivers/pci/quirks.c says that platform specific devices shouldn't go > in there and ISM is platform specific. I took a cursory look at that file as well and arrived at a similar conclusion that drivers/pci/quirks.c (as is) does not sound like a good match for the required functionality > Instead of exporting IOREMAP_START/IOREMAP_END maybe there is another > reasonable maximum resource length? Or maybe we could create a size_t > ioremap_area_size() helper similar to is_ioremap_addr() but not > inlined. The latter already uses IOREMAP_START/IOREMAP_END so not > sure how that works when IOREMAP_END is not exported? But seeing how the PCI quirks filter against Vendor and Device ID, I just had the idea if it would make sense to create a similar infrastructure in the form of VFIO PCI quirks - and simply reject mmap on ISM devices regardless of the size of iomap'able address spaces? Sound rather coarse-grained, though... > Thanks, > Niklas Just a thought, Gerd
