CVE-2026-72496 PUBLISHED

RDMA/bnxt_re: Proper rollback if the ioremap fails

Assigner: Linux
Reserved: 09.08.2026 Published: 15.08.2026 Updated: 17.08.2026

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

RDMA/bnxt_re: Proper rollback if the ioremap fails

bnxt_qplib_alloc_dpi returns success even if ioremap fails. Add the proper rollback when the ioremap fails and return -ENOMEM status.

Metrics

CVSS Vector: CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:C/C:L/I:H/A:H
CVSS Score: 9.2

AV:L - The flaw is reached only through local RDMA uverbs ioctl/write on /dev/infiniband/uverbs* (ibv_alloc_pd and BNXT_RE_ALLOC_WC_PAGE call bnxt_qplib_alloc_dpi); remote RoCE packet handling and in-kernel RDMA target/server paths never invoke this DPI setup routine. AC:L - An attacker can induce vmalloc/ioremap failure with local memory pressure, then repeatedly invoke alloc_pd or WC-page allocation; each call deterministically leaves a consumed DPI with NULL dbr, and subsequent create_cq/QP/SRQ doorbell rings fault without races or uncontrollable layout. PR:N - ib_uverbs creates per-device character nodes as mode 0666, so any local process can open uverbs, allocate a context, and call alloc_pd or alloc_page without CAP_SYS_ADMIN, init-namespace root, or other elevated capabilities on typical HPC/cloud RDMA hosts. UI:N - Exploitation requires only the attacker's own RDMA verbs sequence (alloc_context, alloc_pd or alloc_page, then create_cq/QP/SRQ and I/O); no cooperative action by another user, administrator, or victim mounting devices or opening files is needed. S:C - On Kubernetes/HPC hosts exposing bnxt_re to tenant containers or unprivileged RDMA workloads, triggering host-kernel NULL doorbell writes or exhausting finite DPI indices from a workload crosses the container/tenant security boundary to the hypervisor/host authority. C:L - Returning success after failed ioremap leaves DPI metadata and a mmapable BAR doorbell offset active while kernel dpi->dbr stays NULL, disclosing doorbell-index/BAR layout to the caller and enabling follow-on corruption when doorbells are later rung. I:H - With dpi->dbr left NULL, bnxt_qplib_ring_db and ring_prod_db issue writeq() to NULL during CQ, SQ, RQ, and SRQ doorbell updates on subsequent verbs activity, providing a kernel write primitive and control-flow disruption beyond a benign error return. A:H - NULL doorbell writeq faults oops or panic the kernel when CQ/QP/SRQ objects are used after the bogus allocation, and each successful-but-broken DPI allocation permanently consumes a finite doorbell index, enabling repeated denial of RDMA service to other tenants.

Product Status

Vendor Linux
Product Linux
Versions Default: unaffected
  • affected from 0ac20faf5d837b59fb4c041ea320932ed47fd67f to 303f6fef95df5e5316970746d861cf6daeeca77f (excl.)
  • affected from 0ac20faf5d837b59fb4c041ea320932ed47fd67f to 87267803a8c824616eb147c5dad7030a5db6f878 (excl.)
Vendor Linux
Product Linux
Versions Default: affected
  • Version 6.5 is affected
  • unaffected from 0 to 6.5 (excl.)
  • unaffected from 7.1.5 to 7.1.* (incl.)
  • unaffected from 7.2 to * (incl.)

References