In usb_device_post_load, certain values of dev->setup_len or
dev->setup_index can cause -EINVAL to be returned. One example is when
setup_len exceeds 4096, the hard-coded value of sizeof(dev->data_buf).
This can happen through legitimate guest activity and will cause all
subsequent attempts to migrate the guest to fail in vmstate_load_state.

The values of these variables can be set by USB packets originating in
the guest. There are two ways in which they can be set: in
do_token_setup and in do_parameter in hw/usb/core.c.

It is easy to craft a USB packet in a guest that causes do_token_setup
to set setup_len to a value larger than 4096. When this has been done
once, all subsequent attempts to migrate the VM will fail in
usb_device_post_load until the VM is next power-cycled or a
smaller-sized USB packet is sent to the device.

Sample code for achieving this in a VM started with "-device usb-tablet"
running Linux with CONFIG_HIDRAW=y and HID_MAX_BUFFER_SIZE > 4096:

  #include <sys/types.h>
  #include <sys/stat.h>
  #include <fcntl.h>
  #include <unistd.h>

  int main() {
           char buf[4097];
           int fd = open("/dev/hidraw0", O_RDWR|O_NONBLOCK);

           buf[0] = 0x1;
           write(fd, buf, 4097);

           return 0;
  }

When this code is run in the VM, qemu will output:

  usb_generic_handle_packet: ctrl buffer too small (4097 > 4096)

A subsequent attempt to migrate the VM will fail and output the
following on the destination host:

  qemu-kvm: error while loading state for instance 0x0 of device '0000:00:06.7/1/usb-ptr'
  qemu-kvm: load of migration failed: Invalid argument

The idea behind checking the values of setup_len and setup_index before
they are used is correct, but doing it in usb_device_post_load feels
arbitrary, and will cause unnecessary migration failures. Indeed, none
of the commit messages for c60174e8, 9f8e9895 and 719ffe1f justify why
post_load is the right place to do these checks. They correctly point
out that the important thing to protect is the usb_packet_copy.

Instead, the right place to do the checks is in do_token_setup and
do_parameter. Indeed, there are already some checks here. We can examine
each of the disjuncts currently tested in usb_device_post_load to see
whether any need adding to do_token_setup or do_parameter to improve
safety there:

  * dev->setup_index < 0
     - This test is not needed because setup_index is explicitly set to
0 in do_token_setup and do_parameter.

  * dev->setup_len < 0
     - In both do_token_setup and do_parameter, the value of setup_len
is computed by (s->setup_buf[7] << 8) | s->setup_buf[6]. Since
s->setup_buf is a byte array and setup_len is an int32_t, it's
impossible for this arithmetic to set setup_len's top bit, so it can
never be negative.

  * dev->setup_index > dev->setup_len
     - Since setup_index is 0, this is equivalent to the previous test,
so is redundant.

  * dev->setup_len > sizeof(dev->data_buf)
     - This condition is already explicitly checked in both
do_token_setup and do_parameter.

Hence there is no need to bolster the existing checks in do_token_setup
or do_parameter, and we can safely remove these checks from
usb_device_post_load without reducing safety but allowing migrations to
proceed regardless of what USB packets have been generated by the guest.

Signed-off-by: Jonathan Davies <jonathan.davies@nutanix.com>
Message-Id: <20190107175117.23769-1-jonathan.davies@nutanix.com>
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>