]> git.proxmox.com Git - qemu.git/commit
PIIX3: reset the VM when the Reset Control Register's RCPU bit gets set
authorLaszlo Ersek <lersek@redhat.com>
Thu, 24 Jan 2013 09:31:20 +0000 (10:31 +0100)
committerMichael S. Tsirkin <mst@redhat.com>
Tue, 29 Jan 2013 23:31:09 +0000 (01:31 +0200)
commit1ec4ba741630699665a6334f3959271da3effec7
treefbde5b99a2ce9b06bc8aedf1517ade216c4a2b2b
parent91c3f2f00810a9ba5e4404c9611197efd8f694c8
PIIX3: reset the VM when the Reset Control Register's RCPU bit gets set

  Traditional PCI config space access is achieved by writing a 32 bit
  value to io port 0xcf8 to identify the bus, device, function and config
  register. Port 0xcfc then contains the register in question. But if you
  write the appropriate pair of magic values to 0xcf9, the machine will
  reboot. Spectacular! And not standardised in any way (certainly not part
  of the PCI spec), so different chipsets may have different requirements.
  Booo.

In the PIIX3 spec, IO port 0xcf9 is specified as the Reset Control
Register. Bit 1 (System Reset, SRST) would normally differentiate between
soft reset and hard reset, but we ignore the difference beyond allowing
the guest to read it back.

RHBZ reference: 890459

This patch introduces the following overlap between the preexistent
"pci-conf-idx" region and the "piix3-reset-control" region just being
added. Partial output from "info mtree":

  I/O
  0000000000000000-000000000000ffff (prio 0, RW): io
    0000000000000cf8-0000000000000cfb (prio 0, RW): pci-conf-idx
    0000000000000cf9-0000000000000cf9 (prio 1, RW): piix3-reset-control

I sanity-checked the patch by booting a RHEL-6.3 guest and found no
problems. I summoned gdb and set a breakpoint on rcr_write() in order to
gather a bit more confidence. Relevant frames of the stack:

  kvm_handle_io (port=3321, data=0x7f3f5f3de000, direction=1, size=1,
                 count=1)                                 [kvm-all.c:1422]
    cpu_outb (addr=3321, val=6 '\006')                      [ioport.c:289]
      ioport_write (index=0, address=3321, data=6)           [ioport.c:83]
        ioport_writeb_thunk (opaque=0x7f3f622c4680, addr=3321, data=6)
                                                            [ioport.c:212]
          memory_region_iorange_write (iorange=0x7f3f622c4680, offset=0,
                                       width=1, data=6)     [memory.c:439]
            access_with_adjusted_size (addr=0, value=0x7f3f531fbac0,
                                       size=1, access_size_min=1,
                                       access_size_max=4,
                                       access=0x7f3f5f6e0f90
                                           <memory_region_write_accessor>,
                                       opaque=0x7f3f6227b668)
                                                            [memory.c:364]
              memory_region_write_accessor (opaque=0x7f3f6227b668, addr=0,
                                            value=0x7f3f531fbac0, size=1,
                                            shift=0, mask=255)
                                                            [memory.c:334]
                rcr_write (opaque=0x7f3f6227afb0, addr=0, val=6, len=1)
                                                       [hw/piix_pci.c:498]

The dispatch happens in ioport_write(); "index=0" means byte-wide access:

    static void ioport_write(int index, uint32_t address, uint32_t data)
    {
        static IOPortWriteFunc * const default_func[3] = {
            default_ioport_writeb,
            default_ioport_writew,
            default_ioport_writel
        };
        IOPortWriteFunc *func = ioport_write_table[index][address];
        if (!func)
            func = default_func[index];
        func(ioport_opaque[address], address, data);
    }

The "ioport_write_table" and "ioport_opaque" arrays describe the flattened
IO port space. The first array is less interesting (it selects a thunk
function). The "ioport_opaque" array is interesting because it decides how
writing to the port is implemented ultimately.

4-byte wide access to 0xcf8 (pci-conf-idx):

  (gdb) print ioport_write_table[2][0xcf8]
  $1 = (IOPortWriteFunc *) 0x7f3f5f6d99ba <ioport_writel_thunk>

  (gdb) print \
        ((struct MemoryRegionIORange*)ioport_opaque[0xcf8])->mr->ops.write
  $2 = (void (*)(void *, hwaddr, uint64_t, unsigned int))
       0x7f3f5f5575cb <pci_host_config_write>

1-byte wide access to 0xcf9 (piix3-reset-control):

  (gdb) print ioport_write_table[0][0xcf9]
  $3 = (IOPortWriteFunc *) 0x7f3f5f6d98d0 <ioport_writeb_thunk>

  (gdb) print \
        ((struct MemoryRegionIORange*)ioport_opaque[0xcf9])->mr->ops.write
  $4 = (void (*)(void *, hwaddr, uint64_t, unsigned int))
       0x7f3f5f6b42f1 <rcr_write>

The higher priority of "piix3-reset-control" ensures that the 0xcf9
entries in ioport_write_table / ioport_opaque will always belong to it,
independently of its relative registration order versus "pci-conf-idx".

Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
hw/piix_pci.c