[libvirt] [PATCH] vl: allow "cont" from panicked state

[Paolo Bonzini]

Il 22/08/2013 12:34, Laszlo Ersek ha scritto:
> <academic>

Actually it's fine to clarify these things!  Hence the longish
digression below.

> I think before priority comes into the picture, the access size would
> matter first, no?
> 
> (I think I'm recalling this from the 0xCF9 reset control register, which
> falls into the [0xCF8..0xCFA] range.)

The base address is what matters.  A 2- or 4-byte access to x will
always go to the region that includes address x, even if there are other
regions between x and respectively x+1 or x+3.  So an access to 0xCF8
will go to the PCI address register, while an access to 0xCF9 will
always go to the reset control register.

This happens in address_space_translate_internal:

    diff = int128_sub(section->mr->size, int128_make64(addr));

For a write to 0xCF8, addr would be 0 (it is relative to the base of the
MemoryRegion).  section->size would be 1 because the next section starts
at 0xCF9.  However, section->mr->size would be 4 as specified e.g. in
i440fx_pcihost_initfn:

    memory_region_init_io(&s->conf_mem, obj, &pci_host_conf_le_ops, s,
                          "pci-conf-idx", 4);

Using section->size would be wrong---it would attempt a 1-byte write to
0xCF8, another 1-byte write to 0xCF9, and a 2-byte write to 0xCFA.
section->mr->size instead does a single write to 0xCF8, the same as on
real hardware.

BTW, the behavior changed slightly in QEMU 1.6 for 8-byte accesses. All
such accesses were split to two 4-byte accesses before; now the maximum
size of a "direct" MMIO operation (the data bus size, effectively) is 64
bits, so a 64-bit write will always address a single MemoryRegion.

For example, say you had the PCI address and data registers occupying
two separate 4-byte MemoryRegions in 8 consecutive bytes of memory.  In
1.5 you could write both of them with a single 64-bit write.  In 1.6,
this would only write four bytes to the first MemoryRegion.  This
matches hardware more closely, and is really unlikely to be a problem: a
target with 32-bit data bus probably would not have 64-bit CPU registers
to begin with.  If it did, it would resemble the architecture of the
80386SX or 8088 processors.

> Unless ioport 0 is accessed with width 1 for dma-chan purposes, I think
> such an access would be unique to pvpanic, and always dispatched to pvpanic.

It is:

static const MemoryRegionOps channel_io_ops = {
    .read = read_chan,
    .write = write_chan,
    .endianness = DEVICE_NATIVE_ENDIAN,
    .impl = {
        .min_access_size = 1,
        .max_access_size = 1,
    },
};

>> Channel 0 is (was) used for DRAM refresh, so
>> it should not have any visible effect.  However, it may not be entirely
>> disabling pvpanic, just making it mostly invisible.
> 
> That's good enough for the guest to reach kexec :)

Yes, I cannot deny that. :)

Paolo