Live › vulnerability › CVE-2026-99164
Go to dashboardThe i.MX SPI driver allocates its DMA bounce buffers smaller than the length it maps, so DMA operations run past the end of the allocation. On a reported i.MX8MP system this corrupted kernel memory and eventually caused crashes; the sources do not describe remote code execution.
Summary written from the sources below. Check the originals before acting.
First seen by InfraStack 2026-10-28 11:32 UTC. All times on this page are UTC.
Systems running a Linux kernel with the i.MX ECSPI driver in DMA mode, such as i.MX8MP boards with SPI peripherals under heavy DMA or RPMSG load. The sources describe a custom platform with an mcp251xfd CAN controller; the sources do not say whether other boards are affected in practice. Systems without i.MX SPI hardware can skip it.
No product ranges from the sources yet.
uname -r and compare against the fixed kernel release, which the sources do not namelsmod | grep spi_imx and whether DMA mode is used (unknown from sources)Each row is the first time that source carried this signal. The analyst’s rows are its own findings.
In the Linux kernel, the following vulnerability has been resolved:
spi: imx: fix slab corruption from undersized DMA bounce buffers
Since commit faa8e404ad8e ("spi: imx: support dynamic burst length for ECSPI DMA mode"), the driver allocates its private DMA TX/RX buffers with kzalloc(dma_len), where dma_len is only rounded up to 4 bytes, but maps, syncs and unmaps them with DMA_CACHE_ALIGNED_LEN(dma_len), i.e. rounded up to the cache line size. As a result, the DMA API operates on memory beyond the end of the allocation.
This was observed on a custom i.MX8MP platform with an mcp251xfd CAN controller under heavy RPMSG load, as a series of refcount_t warnings triggered from rpmsg_recv_done() in virtio_rpmsg_bus:
refcount_t: addition on 0; use-after-free. refcount_t: underflow; use-after-free. refcount_t: saturated; leaking memory.
Leading finally to an "Unable to handle kernel paging request" oops. Reproducing this problem took between 30 minutes and roughly 30 hours under the same workload.
Hardware watchpoints on the refcount of a struct rpmsg_endpoint showed it was overwritten from 2 to 1 by __pi_memcpy_generic() in the irq/155-spi2.0 thread. Additionally it was found that a "normal" rpmsg_recv_done() took 11µs +/-2µs from its kref_get to kref_put. In the error case the same sequence took about 215µs +/-10µs. Based on this, an LLM was used to identify the size mismatch fixed here.
Allocate the buffers with DMA_CACHE_ALIGNED_LEN(dma_len) so they cover the full mapped length, and use the same length for dma_sync_single_for_cpu() to stay consistent with map and unmap.
With this patch applied, the same workload ran for more than 200 hours without the problem reproducing.
This page as .md · .json · in your agent: get_signal("CVE-2026-99164")
NVD data is provided by NIST and is public domain. Posts quoted belong to their authors.