[REGRESSION] Commit 5745392e0c2b ("PCI: Apply the new generic I/O management on PCI IO hosts") breaks PCI for legacy virtio devices with kvmtool on arm64

[Will Deacon]

Hi all,

I'm seeing a regression in Linux guests since 4.17 under kvmtool, where
legacy virtio devices using the PCI transport fail to probe. Legacy virtio
PCI devices must be accessed via "I/O space" (e.g. BAR0, which is
IORESOURCE_IO) and kvmtool assigns this to the guest physical range
0x0 - 0x10000.

On arm64, when the virtio legacy PCI driver calls pci_iomap() for this BAR,
it expands to ioport_map():

  static inline void __iomem *ioport_map(unsigned long port, unsigned int nr)
  {
  	return PCI_IOBASE + (port & MMIO_UPPER_LIMIT);
  }

Since the indirect PIO changes, MMIO_UPPER_LIMIT is defined as:

  /*
   * We reserve 0x4000 bytes for Indirect IO as so far this library is only
   * used by the HiSilicon LPC Host. If needed, we can reserve a wider IO
   * area by redefining the macro below.
   */
  #define PIO_INDIRECT_SIZE 0x4000
  #define MMIO_UPPER_LIMIT (IO_SPACE_LIMIT - PIO_INDIRECT_SIZE)

which corrupts the BAR address. For example, kvmtool has the BAR pointing
at 0x6200 on my system, but pci_iomap() actually maps offset 0x2200.
Changing PIO_INDIRECT_SIZE to 0 gets things working again.

Since this stuff doesn't revert nicely, I'm not sure how to proceed. Any
thoughts? Generally, having a per-platform magic constant hardcoded in
the PCI mapping code makes me feel slightly ill...

Cheers,

Will