CVE-2026-98299

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

tcp: do not let tcp rmem be set below 4096

We can hit a division by zero crash in tcp rcvbuf grow() and tcp rcv space adjust():

divide error: 0000 [#1] PREEMPT SMP RIP: 0010:tcp rcvbuf grow+0x187/0x450 net/ipv4/tcp input.c:939 ... grow = div u64(((u64)rcvwin << 1) * (newval - oldval), oldval);

The division uses oldval = tp->rcvq space.space as divisor. When tp->rcvq space.space is zero, this leads to a divide-by-zero exception.

tp->rcvq space.space is initialized in tcp init buffer space(): tp->rcvq space.space = min3(tp->rcv ssthresh, tp->rcv wnd, (u32)TCP INIT CWND * tp->advmss);

If tcp rmem[1] is configured to very small values (such as 1), sk->sk rcvbuf is initialized to 1. Then tcp full space(sk), which computes (sk->sk rcvbuf * scaling ratio) >> 8, truncates to 0. This sets tp->window clamp = 0, tp->rcv ssthresh = 0, and tp->rcvq space.space = 0. Later, when data arrives and DRS is invoked, tcp rcvbuf grow() divides by oldval == 0.

Back in 2015, commit b1cb59cf2efe ("net: sysctl net core: check SNDBUF and RCVBUF for min length") ensured that net.core.rmem default and net.core.rmem max cannot be set below SOCK MIN RCVBUF. Similarly, SO RCVBUF setsockopt enforces max t(int, val * 2, SOCK MIN RCVBUF).

However, net.ipv4.tcp rmem still had .extra1 = SYSCTL ONE, allowing arbitrarily small values.

Because SOCK MIN RCVBUF depends on sizeof(struct sk buff) and cacheline alignment, its value varies across architectures and configuration options. Using a fixed constant of 4096 ensures a predictable, architecture- independent lower bound that is safely above SOCK MIN RCVBUF everywhere and matches the documented 4K default.

Fix this by setting tcp rmem.extra1 to 4096 and updating the documentation.