On Sun, Dec 28, 2025 at 02:49:03PM -0800, Cong Wang wrote:
> On Sun, Dec 28, 2025 at 02:31:36PM -0500, Michael S. Tsirkin wrote:
> > On Sat, Dec 27, 2025 at 05:54:51PM -0800, Cong Wang wrote:
> > > From: Cong Wang <[email protected]>
> > > 
> > > The virtio-vsock driver triggers a DMA debug warning during probe:
> > > 
> [...]
> > > This occurs because event_list[8] contains 8 struct virtio_vsock_event
> > > entries, each only 4 bytes (__le32 id). When virtio_vsock_event_fill()
> > > creates DMA mappings for all 8 events via virtqueue_add_inbuf(), these
> > > 32 bytes all fit within a single 64-byte cacheline.
> > > 
> > > The DMA debug subsystem warns about this because multiple DMA_FROM_DEVICE
> > > mappings within the same cacheline can cause data corruption: if the CPU
> > > writes to one event while the device is writing another event in the same
> > > cacheline, the CPU cache writeback could overwrite device data.
> > 
> > But the CPU never writes into one of these, or did I miss anything?
> > 
> > The real issue is other data in the same cache line?
> 
> You are right, it is misleading.
> 
> The CPU never writes to the event buffers themselves, it only reads them
> after the device writes. The problem is other struct fields in the same
> cacheline.
> 
> I will update the commit message.
> 
> > 
> > You want virtqueue_map_alloc_coherent/virtqueue_map_free_coherent
> > methinks.
> > 
> > Then you can use normal inbuf/outbut and not muck around with premapped.
> > 
> > 
> > I prefer keeping fancy premapped APIs for perf sensitive code,
> > let virtio manage DMA API otherwise.
> 
> Yes, I was not aware of these API's, they are indeed better than using
> DMA API's directly.
> 
> Thanks!
> Cong

BTW I sent an RFC fixing these bugs in all drivers. Review/testing would
be appreciated.


Reply via email to