rtnl_fill_vf() emits one IFLA_VF_INFO per VF into the IFLA_VFINFO_LIST
nest and closes it with nla_nest_end(), which stores the accumulated
length into nla_len. That field is a u16, so a nest larger than 65535
bytes is written truncated modulo 65536. The list dates back to commit
c02db8c629 ("rtnetlink: make SR-IOV VF interface symmetric") in 2010
and has never been able to describe an arbitrary number of VFs; nothing
regressed, the encoding simply cannot represent it.
Nothing catches it on the way. if_nlmsg_size() adds rtnl_vfinfo_size()
for every VF, so the skb really is large enough and none of the nla_put()
calls fails. Userspace then walks the message with RTA_NEXT(), which
advances by the stored length, so parsing resumes inside VF payload and
the attributes after the nest are read out of VF data: IFLA_VF_PORTS,
IFLA_XDP, IFLA_LINKINFO, IFLA_PERM_ADDRESS, IFLA_AF_SPEC. iproute2
prints "!!!Deficit" and strictly validating parsers reject the message.
On CONFIG_DEBUG_NET kernels nla_nest_end() also splats, via the
DEBUG_NET_WARN_ON_ONCE() added in commit ff205bf8c5 ("netlink: add one
debug check in nla_nest_end()").
Where the wrap falls depends on what was asked for and on the host. A VF
costs 196 bytes, 296 with statistics, 236 with GUIDs and 336 with both,
and on a kernel without CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the
statistics carry a padding attribute each and cost 32 bytes more, making
those two 328 and 368. The nest therefore overflows somewhere between 179
and 335 VFs, and ice allows 256 per PF (ICE_MAX_SRIOV_VFS), which reaches
it. Statistics are included unless the request sets
RTEXT_FILTER_SKIP_STATS, so the common case is the one that wraps first.
A limit that moves with the requested attribute set and with the host's
alignment requirements is not something userspace can be told, so use
fixed numbers instead and document them as what the interface supports:
256 VFs, or 128 when statistics are included. Both stay well inside
U16_MAX even in the largest per-VF encoding, at 60416 and 47104 bytes
respectively. rtnl_vfinfo_cap() applies the cap in both places, so
rtnl_vfinfo_size() does not size the skb for VFs that will not be
emitted.
A device with more VFs than the limit reports a shorter
IFLA_VFINFO_LIST. IFLA_NUM_VF keeps carrying the real count, and
everything after the nest stays parsable, which is the part that is
broken today. An empty nest is already emitted for a PF with no VFs, so a
list shorter than IFLA_NUM_VF is not a new encoding.
Returning -EMSGSIZE instead, which is what nla_nest_end_safe() would
give, is not an option here: a nest that does not fit in a u16 will not
fit in a retried skb either, so it would turn a link dump on such a
device into a hard failure.
The other large nests in rtnl_fill_ifinfo() were audited and cannot
overflow. IFLA_AF_SPEC is bounded by a handful of address families at
about a kilobyte each, and IFLA_VF_PORTS would need more than 560 VFs,
which no in-tree driver allows.
Reported-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/netdev/16b289f6-b025-5dd3-443d-92d4c167e79c@intel.com/
Assisted-by: Claude:claude-fable-5
Signed-off-by: Artem Lytkin <iprintercanon@gmail.com>
Link: https://patch.msgid.link/20260801114944.115272-1-iprintercanon@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
51e15308c6