[PATCH v2 2/4] spi: Fix cache corruption due to DMA/PIO overlap

[Marek Szyprowski]

Hi,

CCed: Christoph and Robin, as the issue is partially dma-mapping related.

On 27.09.2022 13:21, Vincent Whitchurch wrote:
> The SPI core DMA mapping support performs cache management once for the
> entire message and not between transfers, and this leads to cache
> corruption if a message has two or more RX transfers with both
> transfers targeting the same cache line, and the controller driver
> decides to handle one using DMA and the other using PIO (for example,
> because one is much larger than the other).
>
> Fix it by syncing before/after the actual transfers.  This also means
> that we can skip the sync during the map/unmap of the message.
>
> Fixes: 99adef310f68 ("spi: Provide core support for DMA mapping transfers")
> Signed-off-by: Vincent Whitchurch <vincent.whitchurch at axis.com>
> ---

This patch landed in linux next-20220929 as commit 0c17ba73c08f ("spi: 
Fix cache corruption due to DMA/PIO overlap"). Unfortunately it causes 
kernel oops on one of my test systems:

8<--- cut here ---
Unable to handle kernel NULL pointer dereference at virtual address 0000000c
[0000000c] *pgd=00000000
Internal error: Oops: 5 [#1] PREEMPT SMP ARM
Modules linked in: cmac bnep btsdio hci_uart btbcm s5p_mfc btintel 
brcmfmac bluetooth videobuf2_dma_contig videobuf2_memops videobuf2_v4l2 
videobuf2_common videodev cfg80211 mc ecdh_generic ecc brcmutil
CPU: 0 PID: 12 Comm: kworker/0:1 Not tainted 
6.0.0-rc7-next-20220929-dirty #12903
Hardware name: Samsung Exynos (Flattened Device Tree)
Workqueue: events ax88796c_work
PC is at dma_direct_sync_sg_for_device+0x24/0xb8
LR is at spi_transfer_one_message+0x4c4/0xabc
pc : [<c01cbcf0>]    lr : [<c0739fcc>]    psr: 20000013
...
Process kworker/0:1 (pid: 12, stack limit = 0xca429928)
Stack: (0xe0071d38 to 0xe0072000)
...
  dma_direct_sync_sg_for_device from spi_transfer_one_message+0x4c4/0xabc
  spi_transfer_one_message from __spi_pump_transfer_message+0x300/0x770
  __spi_pump_transfer_message from __spi_sync+0x304/0x3f4
  __spi_sync from spi_sync+0x28/0x40
  spi_sync from axspi_read_rxq+0x98/0xc8
  axspi_read_rxq from ax88796c_work+0x7a8/0xf6c
  ax88796c_work from process_one_work+0x288/0x774
  process_one_work from worker_thread+0x44/0x504
  worker_thread from kthread+0xf0/0x124
  kthread from ret_from_fork+0x14/0x2c
Exception stack(0xe0071fb0 to 0xe0071ff8)
...
---[ end trace 0000000000000000 ]---

This happens because sg_free_table() doesn't clear table->orig_nents nor 
table->nents. If the given spi xfer object is reused without dma-mapped 
buffer, then a NULL pointer de-reference happens at table->sgl 
spi_dma_sync_for_device()/spi_dma_sync_for_cpu(). A possible fix would 
be to zero table->orig_nents in spi_unmap_buf_attrs(). I will send a 
patch for this soon.

However, I think that clearing table->orig_nents and table->nents should 
be added to __sg_free_table() in lib/scatterlist.c to avoid this kind of 
issue in the future. This however will be a significant change that 
might break code somewhere, if it relies on the nents/orig_nents value 
after calling sg_free_table(). Christoph, Robin - what is your opinion?


>   drivers/spi/spi.c | 109 +++++++++++++++++++++++++++++++++++++---------
>   1 file changed, 88 insertions(+), 21 deletions(-)
>
> ...

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland