CVE-2026-80789: nvmet-tcp: bound SGL data length before allocating command buffers
In the Linux kernel, the following vulnerability has been resolved:
nvmet-tcp: bound SGL data length before allocating command buffers
nvmettcpmapdata() reads the host-controlled 32-bit sgl->length and, for the in-capsule offset descriptor (type 0x01), checks it against port->inlinedatasize before use. Any other SGL descriptor type -- including the non-inline transport SGL data-block descriptor (type (NVMETRANSPORTSGLDATADESC << 4) | NVMESGLFMTTRANSPORTA, the type a real host uses for out-of-capsule writes) skips that check entirely and falls straight through to:
cmd->req.sg = sglalloc(len, GFPKERNEL, &cmd->req.sgcnt);
with len taken directly from the wire, unbounded up to 4 GiB.
nvmetreqinit() only parses the command and never inspects sgl->length, and nvmetchecktransferlen() -- the only other place transferlen is validated -- runs later, from req->execute(), after the allocation has already happened. For a write command the target responds with an R2T and parks the command waiting for the host to send the data; if the host (or an unauthenticated peer that simply never follows up) never does, the sglalloc() buffer stays resident for the life of the command. NVMe/TCP has no mandatory authentication in the default configuration, so any peer able to reach the target portal and complete a Fabrics connect can drive this with a single crafted command, repeatable across queues and connections for amplification. This is unbounded kernel memory allocation triggered by a remote, effectively unauthenticated peer.
Validate len against the same NVMETTCPMAXH2CDATA ceiling this file already uses to bound per-PDU H2C data, for every SGL descriptor type, before doing any allocation. This closes the gap for the non-inline descriptor while leaving the existing, tighter inlinedatasize check in place for the in-capsule case.
Runtime-verified on a v6.19 KASAN stand: with this bound in place, a crafted write command carrying an oversized non-inline SGL length is rejected before sglalloc() runs, where the same request previously drove an unbounded ~256 MiB kernel allocation (up to 4 GiB) that stayed resident pending an R2T the host never satisfies.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Configuration
Enable mandatory authentication for NVMe/TCP, since the vulnerability described is "NVMe/TCP has no mandatory authentication".
Linux kernel NVMe/TCP (nvmet-tcp) mandatory authentication = enabled - Configuration
In nvmet-tcp, validate host-controlled SGL length (sgl->length/len from the wire) against "NVMET_TCP_MAXH2CDATA" and "port->inline_data_size" before any sgl_alloc() call; ensure the in-capsule offset descriptor (type 0x01) checks occur before allocation as well.
Linux kernel NVMe/TCP (nvmet-tcp) SGL length validation = validated against NVMET_TCP_MAXH2CDATA and port->inline_data_size before allocation
Event History
Frequently Asked Questions
Which systems are exposed to this issue?
Systems acting as NVMe-over-TCP targets are exposed when they accept NVMe/TCP commands from a host or peer. The affected path handles out-of-capsule write requests using a transport SGL data-block descriptor.
What does an attacker need to do to trigger the resource consumption?
An attacker needs to send a write command with a host-controlled SGL length and a descriptor type that bypasses the inline-data-size check. After the target allocates the command buffer and sends R2T, the attacker can leave the command unfinished by not sending the requested data.
Is validation elsewhere sufficient to prevent the allocation?
No. nvmet_req_init() does not inspect the SGL length, and nvmet_check_transfer_len() runs only later from request execution, after sgl_alloc() has already used the wire-supplied length.
What is the immediate impact of leaving commands incomplete?
The allocated sgl_alloc() buffer remains resident for the lifetime of the command. Repeated unfinished write requests can therefore retain large command buffers and consume target memory.