[0/2] vmxnet3: Fix inconsistent DMA accesses

Message

Brian Johannesmeyer Nov. 13, 2024, 7:59 p.m. UTC

Hello,

We found hundreds of inconsistent DMA accesses in the VMXNET3 driver. This
patch series aims to fix them. (For a nice summary of the rules around
accessing streaming DMA --- which, if violated, result in inconsistent
accesses --- see Figure 4a of this paper [0]).

The inconsistent accesses occur because the `adapter` object is mapped into
streaming DMA. However, when it is mapped into streaming DMA, it is then
"owned" by the device. Hence, any access to `adapter` thereafter, if not
preceded by a CPU-synchronization operation (e.g.,
`dma_sync_single_for_cpu()`), may cause unexpected hardware behaviors.

This patch series consists of two patches:
- Patch 1 adds synchronization operations into `vmxnet3_probe_device()`, to
  mitigate the inconsistent accesses when `adapter` is initialized.
However, this unfortunately does not mitigate all inconsistent accesses to
it, because `adapter` is accessed elsewhere in the driver without proper
synchronization.
- Patch 2 removes `adapter` from streaming DMA, which entirely mitigates
  the inconsistent accesses to it. It is not clear to me why `adapter` was
mapped into DMA in the first place (in [1]), because it seems that before
[1], it was not mapped into DMA. (However, I am not very familiar with the
VMXNET3 internals, so someone is welcome to correct me here). Alternatively
--- if `adapter` should indeed remain mapped in DMA --- then
synchronization operations should be added throughout the driver code (as
Patch 1 begins to do).

[0] Link: https://www.usenix.org/system/files/sec21-bai.pdf
[1] commit b0eb57cb97e7837ebb746404c2c58c6f536f23fa ("VMXNET3: Add support
for virtual IOMMU")

Brian Johannesmeyer (2):
  vmxnet3: Fix inconsistent DMA accesses in vmxnet3_probe_device()
  vmxnet3: Remove adapter from DMA region

 drivers/net/vmxnet3/vmxnet3_drv.c | 17 ++---------------
 drivers/net/vmxnet3/vmxnet3_int.h |  1 -
 2 files changed, 2 insertions(+), 16 deletions(-)