Linux Kernel CVE-2025-38398

In the Linux kernel, the following vulnerability has been resolved:

spi: spi-qpic-snand: reallocate BAM transactions

Using the mtd_nandbiterrs module for testing the driver occasionally results in weird things like below.

1. swiotlb mapping fails with the following message:

[ 85.926216] qcom_snand 79b0000.spi: swiotlb buffer is full (sz: 4294967294 bytes), total 512 (slots), used 0 (slots) [ 85.932937] qcom_snand 79b0000.spi: failure in mapping desc [ 87.999314] qcom_snand 79b0000.spi: failure to write raw page [ 87.999352] mtd_nandbiterrs: error: write_oob failed (-110)

Rebooting the board after this causes a panic due to a NULL pointer dereference.

1. If the swiotlb mapping does not fail, rebooting the board may result

in a different panic due to a bad spinlock magic:

[ 256.104459] BUG: spinlock bad magic on CPU#3, procd/2241 [ 256.104488] Unable to handle kernel paging request at virtual address ffffffff0000049b ...

Investigating the issue revealed that these symptoms are results of memory corruption which is caused by out of bounds access within the driver.

The driver uses a dynamically allocated structure for BAM transactions, which structure must have enough space for all possible variations of different flash operations initiated by the driver. The required space heavily depends on the actual number of 'codewords' which is calculated from the pagesize of the actual NAND chip.

Although the qcom_nandc_alloc() function allocates memory for the BAM transactions during probe, but since the actual number of 'codewords' is not yet know the allocation is done for one 'codeword' only.

Because of this, whenever the driver does a flash operation, and the number of the required transactions exceeds the size of the allocated arrays the driver accesses memory out of the allocated range.

To avoid this, change the code to free the initially allocated BAM transactions memory, and allocate a new one once the actual number of 'codewords' required for a given NAND chip is known.