Errors on MMIO read access on VM suspend / resume operations

[Stefan Berger <stefanb@linux.vnet.ibm.com>]

Hi!

   I am currently doing some long-term testing of a device model using 
memory mapped IO (TPM TIS) and am seeing some strange errors when the 
suspend occurs in the middle of a read operation in the Linux TPM TIS 
device driver where the driver reads the result packet from the mmio 
location.

  Short background: The TPM response packet is read in 2 chunks. First 
the first 10 bytes are read containing the response's header. 
Subsequently the rest of the packet is read using knowledge of the total 
size of the response packet from the header (bytes 2-5 in big endian 
format). The corresponding code reading the data from the hardware 
interface is here:


http://lxr.linux.no/#linux+v2.6.37/drivers/char/tpm/tpm_tis.c#L228


The test I am running is setup as follows:

- inside the VM keys are permanently generated by sending commands to 
the TPM; packets read from the interface are dumped to the screen

- on the host a script suspends the VM every 6 seconds and resumes it 
immediately afterwards (using libvirt)


As it happens, sometimes the VM is suspended in the middle of a read 
operation on the TPM TIS interface -- see above code reference. I see 
that because I do dump the state of the TPM TIS when suspending and see 
that the read offset is pointing to a location somewhere in the middle 
of the packet - so the TPM TIS Linux driver is in the above loop 
currently reading the data. I am observing two types of results if this 
happens:


- either the result read by the Linux TPM TIS driver is ok, so no 
problem here

- or the problematic case where the TPM TIS driver reads a packet with a 
byte missing and then at the end gets a zero byte from the TPM TIS 
interface indicating that it read beyond the available data. If the 
suspend happened while reading the first chunk of data (header), the TPM 
TIS driver will also complain that the available data for the 2nd  chunk 
(burst size) is less than what's expected  -- it's an off-by-one error


So, I then modified the TPM TIS device model to decrement the read 
offset pointer by '1' in case it was detected that the suspend happened 
in the middle of the read operation -- in Qemu I do this in the 
post-load 'method'. This then leads to the following types of results:


- the problematic(!) case where the read packet was ok

- the expected case where the TPM TIS driver reads the packet and ends 
up having two same bytes in the result in consecutive array locations; 
besides that the TPM TIS driver will in this case complain that it has 
left-over data


So my conclusion from the above tests are:

- for some reason the memory read to the MMIO location happens as the 
last instruction executed on suspend and again as the very first on 
executed on resume. This explains to me that the TPM TIS model internal 
pointer into the packet was advanced by '1' (the packet is read by 
subsequently reading from the same memory location) and the above 
problematic cases make sense

- the other instruction in the Linux TPM TIS drivers that for example 
advance the buffer location do not execute twice, i.e., size++ in the 
buf[size++] = ... in the Linux driver.


What puzzles me is that the read operation may be run twice but others 
don't.


If you have insights as why the above may be occurring, please let me 
know. A simple solution to work around this may be to introduce a 
register holding the index into the result packet where to read the next 
byte from (rather than advancing an internal pointer to the next byte), 
though this would deviate the driver from the standard interface the 
model currently implements.


   Regards,

      Stefan