Dexuan wrote:

  "[...]  when we disable AccelNet, the host PCI VSP driver sends a
   PCI_EJECT message first, and the channel callback may set
   hpdev->state to hv_pcichild_ejecting on a different CPU.  This can
   cause hv_compose_msi_msg() to exit from the loop and 'return', and
   the on-stack variable 'ctxt' is invalid.  Now, if the response
   message from the host arrives, the channel callback will try to
   access the invalid 'ctxt' variable, and this may cause a crash."

Schematically:

  Hyper-V sends PCI_EJECT msg
    hv_pci_onchannelcallback()
      state = hv_pcichild_ejecting
                                       hv_compose_msi_msg()
                                         alloc and init comp_pkt
                                         state == hv_pcichild_ejecting
  Hyper-V sends VM_PKT_COMP msg
    hv_pci_onchannelcallback()
      retrieve address of comp_pkt
                                         'free' comp_pkt and return
      comp_pkt->completion_func()

Dexuan also showed how the crash can be triggered after introducing
suitable delays in the driver code, thus validating the 'assumption'
that the host can still normally respond to the guest's compose_msi
request after the host has started to eject the PCI device.

Fix the synchronization by leveraging the requestor lock as follows:

  - Before 'return'-ing in hv_compose_msi_msg(), remove the ID (while
    holding the requestor lock) associated to the completion packet.

  - Retrieve the address *and call ->completion_func() within a same
    (requestor) critical section in hv_pci_onchannelcallback().
Reported-by: Wei Hu <weh@microsoft.com>
Reported-by: Dexuan Cui <decui@microsoft.com>
Suggested-by: Michael Kelley <mikelley@microsoft.com>
Signed-off-by: Andrea Parri (Microsoft) <parri.andrea@gmail.com>
Reviewed-by: Michael Kelley <mikelley@microsoft.com>
Link: https://lore.kernel.org/r/20220419122325.10078-7-parri.andrea@gmail.comSigned-off-by: Wei Liu <wei.liu@kernel.org>