pci, vmm: Identify a moving PCI BAR by index, not by address - #8695

The device manager will use the constant of ivshmem BAR2 as data BAR.
Rename IVSHMEM_BAR2_IDX to IVSHMEM_DATA_BAR_IDX to give a meaningful
name for later use.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>

As renamed IVSHMEM_BAR2_IDX to IVSHMEM_DATA_BAR_IDX, update the code
to use data_bar instead of bar2 for consistency.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>

For code readability rename VIRTIO_COMMON_BAR_INDEX to
VIRTIO_CONFIG_BAR_INDEX as CONFIG makes more sense than COMMON.  At
the same time, change the type from u8 to usize.  PCI BAR index is
typed as usize in common code.  u8 is used only within pci_device.rs.
So change the type from u8 to usize for consistency.  The trade-off is
the cast to u8 for VirtioPciCap/MsixCap.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>

Setting up the virtio ioeventfds and the ivshmem RAM region used to
read the BAR base address from the device.  Instead, read BAR base
address with the BAR index from the BAR list returned by
allocate_pci_bars().

The change set of c7cabc8 ("vmm: Conditionally update ioeventfds
for virtio PCI device") first introduced config_bar_addr() to
determine which BAR was updated by its BAR base address to avoid
unnecessarily updating the underlying ioeventfd resources.  Later the
other devices adopted the idiom, and introduced data_bar_addr().
I guess that because move_bar() doesn't receive PCI BAR index, but
old_base and new_base, using base address was easier to implement
than adding bar index to move_bar(). but the commit message and the
diff doesn't tell why bar index wasn't used.

allocate_pci_bars() sets up PCI BARs in pci_device.configuration
internally and returns the BAR list including the base address.
config_bar_addr()/data_bar_addr() return the base address of a
specified BAR from pci_device.configuration.  It's redundant to query
the PCI device for those base addresses again because the caller of
allocate_pci_bars() already obtained BAR information including the
base address.  It is straightforward and uniform to use the BARs
returned by allocate_pci_bars().  This untangles the unnecessary calls
from the device manager to pci devices.

Introduce MissingPciBar error to cover the case when the BAR index is
invalid.  As long as a valid BAR index is passed, it won't be
triggered.

This is a step towards dropping the BAR base address comparison in the
BAR movement code.  Because there are still callers of
config_bar_addr()/data_bar_addr(), this patch does not remove them
yet.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>
Assisted-by: Claude:Fable-5

Identify PCI BAR by (pci device, BAR index) in the PCI BAR relocation
code instead of BAR base address.  In theory, a guest can program two
or more BARs to the same base address, so an address comparison
can match the wrong BAR.  Well-behaved software won't do this, though.

Pass around the BAR index through DeviceRelocation::move_bar() and
PciDevice::move_bar() and match on it instead of base address. VFIO
and vfio-user derive the old address from their own region records.
move_bar() keeps old_base on the DeviceRelocation side: it is still
needed to unregister the ioeventfds and to unmap the shared memory
region at their current addresses.

BarReprogrammingParams is passed around on the snapshot state as part
of PciConfigurationState::pending_bar_reprogram, so bar_idx is
optional: an entry restored from a snapshot that predates it cannot
identify its BAR.  Such an entry is rolled back rather than applied to
a guessed slot, which is what already happens when a move fails.

Device eject therefore takes the settings BAR's mapped address from
the device-tree resource instead of config space.  The device is off
the device tree by the time eject gets there, so a resource that
cannot be found is reported and skipped rather than failing the eject
half way through.

For the virtio shared-memory BAR the address-equality guard is
replaced by the index check.  A move of any other BAR index is now a
no-op, which changes nothing for virtio because only the config and
shared-memory BARs exist and the existing code assumes a non-config
bar means the shared-memory BAR.  Now invalid bar index results in
real error instead of being silently ignored.  In practice that
execution path won't be hit because such a BAR cannot be touched via
pci_config_write().

There is no caller of config_bar_addr() and data_bar_addr() so the
following patch removes them.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>
Assisted-by: Claude:Fable-5

There is no caller for IvshmemDevice::config_bar_addr(),
data_bar_addr(), PvPanicDevice::config_bar_addr(), and
VirtioPciDevice::config_bar_addr().  Because they are pub, they are
not reported as dead code.  Remove them.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>
Assisted-by: Claude:Opus-5

Cover the two mechanisms the previous patches introduced.

For detect_bar_reprogramming(), each of the three ways a BAR move is
reported now asserts the slot it names: a plain 32-bit BAR, the
high-dword write of a 64-bit BAR (which must name the low slot, the one
devices record their BAR under), and the expansion ROM BAR.  One test
programs two BARs to the same base first, so the reported index is the
only thing that distinguishes them - the case an address comparison
gets wrong.

Two tests pin the snapshot behavior: a pending entry restored without
a bar_idx stays None rather than defaulting to slot 0, and the wire
form of BarReprogrammingParams still decodes when the field is absent,
which is what a snapshot taken before this series looks like.
apply_bar_reprogramming() is then checked to move the named BAR, and
to roll an unnamed one back instead of moving a guessed slot.
MockDeviceRelocation grows a record of the BAR indexes it was asked to
move, so that those two can tell the cases apart.

For bar_addr_of_idx(), the lookup is checked against a BAR list that is
not ordered by index, and for the error raised when the index is
absent.

Signed-off-by: Isaku Yamahata <iyamahata@crusoe.ai>
Assisted-by: Claude:Opus-5