CVE-2026-89602

Quoted from the kernel CNA's own scoring file, word for word. A highlighted metric carries the most severe value CVSS v3.1 defines for it, which is what pushes the score up. source

AV:LAttack vector, local

z_erofs_gbuf_growsize() runs only during a local EROFS mount (erofs_fc_get_tree → erofs_fc_fill_super → erofs_read_superblock → z_erofs_parse_cfgs → z_erofs_load_lz4_config) of a block, loop, or file-backed image. The LZ4 max_pclusterblks field is not parsed from a network protocol, so the vector is Local.

AC:LAttack complexity, low

The attacker picks max_pclusterblks on two images (a large grow, then a strictly smaller size). After GFP_KERNEL alloc_pages_bulk()/vmap fails mid-loop, some per-CPU gbufs stay enlarged while z_erofs_gbuf_nrpages does not; the next intermediate mount then copies gbuf->nrpages pointers into a smaller tmp_pages array with no race. A max-size image plus memory pressure makes that partial failure attacker-inducible.

PR:NPrivileges required, none

Crafting and supplying the malicious EROFS images needs no account or capability on the target. erofs_fs_type lacks FS_USERNS_MOUNT, so the CAP_SYS_ADMIN mount is done by a victim, automounter, or container runtime (OCI/Nydus/Android image handling), matching other crafted-image EROFS CVEs.

UI:RUser interaction, required

mount_capable() requires CAP_SYS_ADMIN in the initial user namespace because EROFS is not FS_USERNS_MOUNT. A separate user, automounter, or container runtime must mount the attacker-supplied images before z_erofs_gbuf_growsize() runs.

S:UScope, unchanged

The heap overflow corrupts kmalloc objects inside the host kernel. That is standard kernel memory corruption within one security authority, not a VM escape, IOMMU bypass, or other cross-boundary impact.

C:HConfidentiality impact, high

The copy loop writes a nearly page-count-sized run of struct page * values past the kmalloc tmp_pages array (up to hundreds of pointers when max_pclusterblks is 256). That slab corruption can be turned into an arbitrary-read primitive against adjacent objects, so confidentiality impact is High.

I:HIntegrity impact, high

This is a kernel heap out-of-bounds write whose length is chosen by the two image sizes, smashing adjacent slab objects with kernel page pointers. Out-of-bounds writes are exploitable for arbitrary write and control-flow hijacking, so integrity impact is High.

A:HAvailability impact, high

Overwriting adjacent kmalloc metadata and objects causes a kernel oops or panic. Independently, retrying the same size after a partial grow treats a full tmp_pages array as -ENOMEM and leaves the pool stuck, so later EROFS mounts fail until reboot.