[PATCH v2] hv: account for packet descriptor in maximum packet size

[Wei Liu]

On Mon, Jan 10, 2022 at 01:44:19AM +0100, Andrea Parri wrote:
> (Extending Cc: list,)
> 
> On Sun, Jan 09, 2022 at 05:55:16PM +0800, Yanming Liu wrote:
> > Commit adae1e931acd ("Drivers: hv: vmbus: Copy packets sent by Hyper-V
> > out of the ring buffer") introduced a notion of maximum packet size in
> > vmbus channel and used that size to initialize a buffer holding all
> > incoming packet along with their vmbus packet header. Currently, some
> > vmbus drivers set max_pkt_size to the size of their receive buffer
> > passed to vmbus_recvpacket, however vmbus_open expects this size to also
> > include vmbus packet header. This leads to corruption of the ring buffer
> > state when receiving a maximum sized packet.
> > 
> > Specifically, in hv_balloon I have observed of a dm_unballoon_request
> > message of 4096 bytes being truncated to 4080 bytes. When the driver
> > tries to read next packet it starts from a wrong read_index, receives
> > garbage and prints a lot of "Unhandled message: type: <garbage>" in
> > dmesg.
> > 
> > The same mismatch also happens in hv_fcopy, hv_kvp, hv_snapshot,
> > hv_util, hyperv_drm and hyperv_fb, though bad cases are not observed
> > yet.
> > 
> > Allocate the buffer with HV_HYP_PAGE_SIZE more bytes to make room for
> > the descriptor, assuming the vmbus packet header will never be larger
> > than HV_HYP_PAGE_SIZE. This is essentially free compared to just adding
> > 'sizeof(struct vmpacket_descriptor)' because these buffers are all more
> > than HV_HYP_PAGE_SIZE bytes so kmalloc rounds them up anyway.
> > 
> > Fixes: adae1e931acd ("Drivers: hv: vmbus: Copy packets sent by Hyper-V out of the ring buffer")
> > Suggested-by: Andrea Parri (Microsoft) <parri.andrea at gmail.com>
> > Signed-off-by: Yanming Liu <yanminglr at gmail.com>
> 
> Thanks for sorting this out; the patch looks good to me:
> 
> Reviewed-by: Andrea Parri (Microsoft) <parri.andrea at gmail.com>
> 

Thanks. I will pick this up after 5.17-rc1 is out.

Wei.