Merge 9606690a73 into 09815d3c0c

aya/src/programs/tc.rs

@@ -14,8 +14,8 @@ use crate::{
     },
     programs::{define_link_wrapper, load_program, Link, ProgramData, ProgramError},
     sys::{
-        netlink_find_filter_with_name, netlink_qdisc_add_clsact, netlink_qdisc_attach,
+        netlink_clsact_qdisc_exists, netlink_find_filter_with_name, netlink_qdisc_add_clsact,
-        netlink_qdisc_detach,
+        netlink_qdisc_attach, netlink_qdisc_detach,
     },
     util::{ifindex_from_ifname, tc_handler_make},
     VerifierLogLevel,
@@ -390,3 +390,9 @@ fn qdisc_detach_program_fast(
 
     Ok(())
 }
+
+/// Check if the `clasct` qdisc exists on the given interface.
+pub fn clsact_qdisc_exists(if_name: &str) -> Result<bool, io::Error> {
+    let if_index = ifindex_from_ifname(if_name)?;
+    unsafe { netlink_clsact_qdisc_exists(if_index as i32) }
+}

aya/src/sys/netlink.rs

@@ -10,8 +10,8 @@ use libc::{
     getsockname, nlattr, nlmsgerr, nlmsghdr, recv, send, setsockopt, sockaddr_nl, socket,
     AF_NETLINK, AF_UNSPEC, ETH_P_ALL, IFF_UP, IFLA_XDP, NETLINK_EXT_ACK, NETLINK_ROUTE,
     NLA_ALIGNTO, NLA_F_NESTED, NLA_TYPE_MASK, NLMSG_DONE, NLMSG_ERROR, NLM_F_ACK, NLM_F_CREATE,
-    NLM_F_DUMP, NLM_F_ECHO, NLM_F_EXCL, NLM_F_MULTI, NLM_F_REQUEST, RTM_DELTFILTER, RTM_GETTFILTER,
+    NLM_F_DUMP, NLM_F_ECHO, NLM_F_EXCL, NLM_F_MULTI, NLM_F_REQUEST, RTM_DELTFILTER, RTM_GETQDISC,
-    RTM_NEWQDISC, RTM_NEWTFILTER, RTM_SETLINK, SOCK_RAW, SOL_NETLINK,
+    RTM_GETTFILTER, RTM_NEWQDISC, RTM_NEWTFILTER, RTM_SETLINK, SOCK_RAW, SOL_NETLINK,
 };
 use thiserror::Error;
 
@@ -80,6 +80,46 @@ pub(crate) unsafe fn netlink_set_xdp_fd(
     Ok(())
 }
 
+pub(crate) unsafe fn netlink_clsact_qdisc_exists(if_index: i32) -> Result<bool, io::Error> {
+    let sock = NetlinkSocket::open()?;
+
+    let mut req = mem::zeroed::<TcRequest>();
+
+    let nlmsg_len = mem::size_of::<nlmsghdr>() + mem::size_of::<tcmsg>();
+
+    req.header = nlmsghdr {
+        nlmsg_len: nlmsg_len as u32,
+        nlmsg_flags: (NLM_F_REQUEST | NLM_F_ACK | NLM_F_DUMP) as u16,
+        nlmsg_type: RTM_GETQDISC,
+        nlmsg_pid: 0,
+        nlmsg_seq: 1,
+    };
+    req.tc_info.tcm_family = AF_UNSPEC as u8;
+
+    sock.send(&bytes_of(&req)[..req.header.nlmsg_len as usize])?;
+
+    for msg in sock.recv()? {
+        if msg.header.nlmsg_type != RTM_NEWQDISC {
+            continue;
+        }
+
+        let tc_msg = ptr::read_unaligned(msg.data.as_ptr() as *const tcmsg);
+        if tc_msg.tcm_ifindex != if_index {
+            continue;
+        }
+
+        let attrs = parse_attrs(&msg.data[mem::size_of::<tcmsg>()..])?;
+
+        if let Some(opts) = attrs.get(&(TCA_KIND as u16)) {
+            if opts.data == b"clsact\0" {
+                return Ok(true);
+            }
+        }
+    }
+
+    Ok(false)
+}
+
 pub(crate) unsafe fn netlink_qdisc_add_clsact(if_index: i32) -> Result<(), io::Error> {
     let sock = NetlinkSocket::open()?;
 
@@ -303,7 +343,13 @@ struct Request {
 struct TcRequest {
     header: nlmsghdr,
     tc_info: tcmsg,
-    attrs: [u8; 64],
+    // The buffer for attributes should be sized to hold at least 256 bytes,
+    // based on `CLS_BPF_NAME_LEN = 256` from the kernel:
+    // https://github.com/torvalds/linux/blob/02aee814/net/sched/cls_bpf.c#L28
+    // We currently use around ~30 bytes of attributes in addition to name.
+    // Rather than picking a "right sized buffer" for the payload (which is of
+    // varying length anyway) we use the next largest power of 2.
+    attrs: [u8; 512],
 }
 
 struct NetlinkSocket {

test/integration-ebpf/Cargo.toml

@@ -59,3 +59,11 @@ path = "src/xdp_sec.rs"
 [[bin]]
 name = "ring_buf"
 path = "src/ring_buf.rs"
+
+[[bin]]
+name = "tc_name_limit"
+path = "src/tc_name_limit.rs"
+
+[[bin]]
+name = "tc_name_limit_exceeded"
+path = "src/tc_name_limit_exceeded.rs"

test/integration-ebpf/src/tc_name_limit.rs

@@ -0,0 +1,22 @@
+#![no_std]
+#![no_main]
+
+use aya_bpf::{macros::classifier, programs::TcContext};
+
+// A function with a 256-byte-long name (all 'a's) to be used as the name of
+// the ebpf program. This name must match the name passed to userspace side
+// of the program (i.e. test/integration-test/src/tests/load.rs).
+// 256 is the maximum length allowed by the kernel, so this test should pass.
+// https://github.com/torvalds/linux/blob/02aee814/net/sched/cls_bpf.c#L28
+#[classifier]
+pub fn aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa(
+    _ctx: TcContext,
+) -> i32 {
+    0
+}
+
+#[cfg(not(test))]
+#[panic_handler]
+fn panic(_info: &core::panic::PanicInfo) -> ! {
+    loop {}
+}

test/integration-ebpf/src/tc_name_limit_exceeded.rs

@@ -0,0 +1,22 @@
+#![no_std]
+#![no_main]
+
+use aya_bpf::{macros::classifier, programs::TcContext};
+
+// A function with a 257-byte-long name (all 'a's) to be used as the name of
+// the ebpf program. This name must match the name passed to userspace side
+// of the program (i.e. test/integration-test/src/tests/load.rs).
+// 256 is the maximum length allowed by the kernel, so this test should fail.
+// https://github.com/torvalds/linux/blob/02aee814/net/sched/cls_bpf.c#L28
+#[classifier]
+pub fn aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa(
+    _ctx: TcContext,
+) -> i32 {
+    0
+}
+
+#[cfg(not(test))]
+#[panic_handler]
+fn panic(_info: &core::panic::PanicInfo) -> ! {
+    loop {}
+}

test/integration-test/src/lib.rs

@@ -22,6 +22,10 @@ pub const BPF_PROBE_READ: &[u8] =
 pub const REDIRECT: &[u8] = include_bytes_aligned!(concat!(env!("OUT_DIR"), "/redirect"));
 pub const XDP_SEC: &[u8] = include_bytes_aligned!(concat!(env!("OUT_DIR"), "/xdp_sec"));
 pub const RING_BUF: &[u8] = include_bytes_aligned!(concat!(env!("OUT_DIR"), "/ring_buf"));
+pub const TC_NAME_LIMIT_TEST: &[u8] =
+    include_bytes_aligned!(concat!(env!("OUT_DIR"), "/tc_name_limit"));
+pub const TC_NAME_LIMIT_EXCEEDED_TEST: &[u8] =
+    include_bytes_aligned!(concat!(env!("OUT_DIR"), "/tc_name_limit_exceeded"));
 
 #[cfg(test)]
 mod tests;

test/integration-test/src/tests/load.rs

@@ -6,11 +6,15 @@ use std::{
     time::{Duration, SystemTime},
 };
 
+use assert_matches::assert_matches;
 use aya::{
     maps::Array,
     programs::{
         links::{FdLink, PinnedLink},
-        loaded_links, loaded_programs, KProbe, TracePoint, UProbe, Xdp, XdpFlags,
+        loaded_links, loaded_programs,
+        tc::{clsact_qdisc_exists, qdisc_add_clsact},
+        KProbe, ProgramError, SchedClassifier, TcAttachType, TcError, TracePoint, UProbe, Xdp,
+        XdpFlags,
     },
     util::KernelVersion,
     Ebpf,
@@ -608,3 +612,51 @@ fn pin_lifecycle_uprobe() {
     // Make sure the function isn't optimized out.
     uprobe_function();
 }
+
+#[test]
+fn tc_name_limit() {
+    let clsact_exists = clsact_qdisc_exists("lo").unwrap();
+    if !clsact_exists {
+        qdisc_add_clsact("lo").unwrap();
+    }
+
+    let mut bpf = Bpf::load(crate::TC_NAME_LIMIT_TEST).unwrap();
+
+    let long_name = "a".repeat(256);
+
+    let program: &mut SchedClassifier = bpf
+        .program_mut(long_name.as_str())
+        .unwrap()
+        .try_into()
+        .unwrap();
+    program.load().unwrap();
+    program.attach("lo", TcAttachType::Ingress).unwrap();
+}
+
+#[test]
+fn tc_name_limit_exceeded() {
+    let clsact_exists = clsact_qdisc_exists("lo").unwrap();
+    if !clsact_exists {
+        qdisc_add_clsact("lo").unwrap();
+    }
+
+    let mut bpf = Bpf::load(crate::TC_NAME_LIMIT_EXCEEDED_TEST).unwrap();
+
+    let long_name = "a".repeat(257);
+
+    let program: &mut SchedClassifier = bpf
+        .program_mut(long_name.as_str())
+        .unwrap()
+        .try_into()
+        .unwrap();
+    program.load().unwrap();
+
+    assert_matches!(
+      program.attach("lo", TcAttachType::Ingress),
+      Err(ProgramError::TcError(TcError::NetlinkError { io_error })) => {
+        // An invalid argument error (EINVAL) with code 22 should occur.
+        // The invalid argument is the tc program name which is too long.
+        assert_eq!(io_error.raw_os_error(), Some(22))
+      }
+    );
+}

xtask/public-api/aya.txt

@@ -5673,6 +5673,7 @@ impl<T> core::borrow::BorrowMut<T> for aya::programs::tc::TcOptions where T: cor
 pub fn aya::programs::tc::TcOptions::borrow_mut(&mut self) -> &mut T
 impl<T> core::convert::From<T> for aya::programs::tc::TcOptions
 pub fn aya::programs::tc::TcOptions::from(t: T) -> T
+pub fn aya::programs::tc::clsact_qdisc_exists(if_name: &str) -> core::result::Result<bool, std::io::error::Error>
 pub fn aya::programs::tc::qdisc_add_clsact(if_name: &str) -> core::result::Result<(), std::io::error::Error>
 pub fn aya::programs::tc::qdisc_detach_program(if_name: &str, attach_type: aya::programs::tc::TcAttachType, name: &str) -> core::result::Result<(), std::io::error::Error>
 pub mod aya::programs::tp_btf