On Wed, May 30, 2018 at 06:04:05PM -0700, Sarah Newman wrote:
[ Not relevant upstream, therefore no upstream commit. ]
To fix, unmap the page as soon as possible.
When swiotlb is in use, calling dma_unmap_page means that the original page mapped with dma_map_page must still be valid, as swiotlb will copy data from its internal cache back to the originally requested DMA location.
When GRO is enabled, before this patch all references to the original frag may be put and the page freed before dma_unmap_page in mlx4_en_free_frag is called.
It is possible there is a path where the use-after-free occurs even with GRO disabled, but this has not been observed so far.
The bug can be trivially detected by doing the following:
- Compile the kernel with DEBUG_PAGEALLOC
- Run the kernel as a Xen Dom0
- Leave GRO enabled on the interface
- Run a 10 second or more test with iperf over the interface.
This bug was likely introduced in commit 4cce66cdd14a ("mlx4_en: map entire pages to increase throughput"), first part of u3.6.
It was incidentally fixed in commit 34db548bfb95 ("mlx4: add page recycling in receive path"), first part of v4.12.
Why not just apply this patch instead?
This version applies to the v4.9 series.
What about 4.4? Why not just use 4.14 for this hardware?
thanks,
greg k-h