With a SOCKMAP/SOCKHASH map and an sk_msg program user can steer messages
sent from one TCP socket (s1) to actually egress from another TCP
socket (s2):
tcp_bpf_sendmsg(s1) // = sk_prot->sendmsg
tcp_bpf_send_verdict(s1) // __SK_REDIRECT case
tcp_bpf_sendmsg_redir(s2)
tcp_bpf_push_locked(s2)
tcp_bpf_push(s2)
tcp_rate_check_app_limited(s2) // expects tcp_sock
tcp_sendmsg_locked(s2) // ditto
There is a hard-coded assumption in the call-chain, that the egress
socket (s2) is a TCP socket.
However in commit 122e6c79efe1 ("sock_map: Update sock type checks for
UDP") we have enabled redirects to non-TCP sockets. This was done for the
sake of BPF sk_skb programs. There was no indention to support sk_msg
send-to-egress use case.
As a result, attempts to send-to-egress through a non-TCP socket lead to a
crash due to invalid downcast from sock to tcp_sock:
BUG: kernel NULL pointer dereference, address: 000000000000002f
...
Call Trace:
<TASK>
? show_regs+0x60/0x70
? __die+0x1f/0x70
? page_fault_oops+0x80/0x160
? do_user_addr_fault+0x2d7/0x800
? rcu_is_watching+0x11/0x50
? exc_page_fault+0x70/0x1c0
? asm_exc_page_fault+0x27/0x30
? tcp_tso_segs+0x14/0xa0
tcp_write_xmit+0x67/0xce0
__tcp_push_pending_frames+0x32/0xf0
tcp_push+0x107/0x140
tcp_sendmsg_locked+0x99f/0xbb0
tcp_bpf_push+0x19d/0x3a0
tcp_bpf_sendmsg_redir+0x55/0xd0
tcp_bpf_send_verdict+0x407/0x550
tcp_bpf_sendmsg+0x1a1/0x390
inet_sendmsg+0x6a/0x70
sock_sendmsg+0x9d/0xc0
? sockfd_lookup_light+0x12/0x80
__sys_sendto+0x10e/0x160
? syscall_enter_from_user_mode+0x20/0x60
? __this_cpu_preempt_check+0x13/0x20
? lockdep_hardirqs_on+0x82/0x110
__x64_sys_sendto+0x1f/0x30
do_syscall_64+0x38/0x90
entry_SYSCALL_64_after_hwframe+0x63/0xcd
Reject selecting a non-TCP sockets as redirect target from a BPF sk_msg
program to prevent the crash. When attempted, user will receive an EACCES
error from send/sendto/sendmsg() syscall.
Fixes: 122e6c79efe1 ("sock_map: Update sock type checks for UDP")
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20230920102055.42662-1-jakub@cloudflare.com
Sockmap and sockhash maps are a collection of psocks that are
objects representing a socket plus a set of metadata needed
to manage the BPF programs associated with the socket. These
maps use the stab->lock to protect from concurrent operations
on the maps, e.g. trying to insert to objects into the array
at the same time in the same slot. Additionally, a sockhash map
has a bucket lock to protect iteration and insert/delete into
the hash entry.
Each psock has a psock->link which is a linked list of all the
maps that a psock is attached to. This allows a psock (socket)
to be included in multiple sockmap and sockhash maps. This
linked list is protected the psock->link_lock.
They _must_ be nested correctly to avoid deadlock:
lock(stab->lock)
: do BPF map operations and psock insert/delete
lock(psock->link_lock)
: add map to psock linked list of maps
unlock(psock->link_lock)
unlock(stab->lock)
For non PREEMPT_RT kernels both raw_spin_lock_t and spin_lock_t
are guaranteed to not sleep. But, with PREEMPT_RT kernels the
spin_lock_t variants may sleep. In the current code we have
many patterns like this:
rcu_critical_section:
raw_spin_lock(stab->lock)
spin_lock(psock->link_lock) <- may sleep ouch
spin_unlock(psock->link_lock)
raw_spin_unlock(stab->lock)
rcu_critical_section
Nesting spin_lock() inside a raw_spin_lock() violates locking
rules for PREEMPT_RT kernels. And additionally we do alloc(GFP_ATOMICS)
inside the stab->lock, but those might sleep on PREEMPT_RT kernels.
The result is splats like this:
./test_progs -t sockmap_basic
[ 33.344330] bpf_testmod: loading out-of-tree module taints kernel.
[ 33.441933]
[ 33.442089] =============================
[ 33.442421] [ BUG: Invalid wait context ]
[ 33.442763] 6.5.0-rc5-01731-gec0ded2e0282 #4958 Tainted: G O
[ 33.443320] -----------------------------
[ 33.443624] test_progs/2073 is trying to lock:
[ 33.443960] ffff888102a1c290 (&psock->link_lock){....}-{3:3}, at: sock_map_update_common+0x2c2/0x3d0
[ 33.444636] other info that might help us debug this:
[ 33.444991] context-{5:5}
[ 33.445183] 3 locks held by test_progs/2073:
[ 33.445498] #0: ffff88811a208d30 (sk_lock-AF_INET){+.+.}-{0:0}, at: sock_map_update_elem_sys+0xff/0x330
[ 33.446159] #1: ffffffff842539e0 (rcu_read_lock){....}-{1:3}, at: sock_map_update_elem_sys+0xf5/0x330
[ 33.446809] #2: ffff88810d687240 (&stab->lock){+...}-{2:2}, at: sock_map_update_common+0x177/0x3d0
[ 33.447445] stack backtrace:
[ 33.447655] CPU: 10 PID
To fix observe we can't readily remove the allocations (for that
we would need to use/create something similar to bpf_map_alloc). So
convert raw_spin_lock_t to spin_lock_t. We note that sock_map_update
that would trigger the allocate and potential sleep is only allowed
through sys_bpf ops and via sock_ops which precludes hw interrupts
and low level atomic sections in RT preempt kernel. On non RT
preempt kernel there are no changes here and spin locks sections
and alloc(GFP_ATOMIC) are still not sleepable.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20230830053517.166611-1-john.fastabend@gmail.com
sock_map_del_link() operates on both SOCKMAP and SOCKHASH, although
both types have member named "progs", the offset of "progs" member in
these two types is different, so "progs" should be accessed with the
real map type.
Fixes: 604326b41a6f ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: Xu Kuohai <xukuohai@huawei.com>
Reviewed-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/r/20230804073740.194770-2-xukuohai@huaweicloud.com
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC
allocation later in sk_psock_init_link on PREEMPT_RT kernels, since
GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist
patchset notes for details).
This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to
BUG (sleeping function called from invalid context) on RT kernels.
preempt_disable was introduced together with lock_sk and rcu_read_lock
in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update
in parallel"), probably to match disabled migration of BPF programs, and
is no longer necessary.
Remove preempt_disable to fix BUG in sock_map_update_common on RT.
Signed-off-by: Tomas Glozar <tglozar@redhat.com>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/all/20200224140131.461979697@linutronix.de/
Fixes: 99ba2b5aba24 ("bpf: sockhash, disallow bpf_tcp_close and update in parallel")
Reviewed-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/r/20230728064411.305576-1-tglozar@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This allows to do more centralized decisions later on, and generally makes it very explicit which maps are privileged and which are not (e.g., LRU_HASH and LRU_PERCPU_HASH, which are privileged HASH variants, as opposed to unprivileged HASH and HASH_PERCPU; now this is explicit and easy to verify). Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Stanislav Fomichev <sdf@google.com> Link: https://lore.kernel.org/bpf/20230613223533.3689589-4-andrii@kernel.org
Sk_buffs are fed into sockmap verdict programs either from a strparser
(when the user might want to decide how framing of skb is done by attaching
another parser program) or directly through tcp_read_sock. The
tcp_read_sock is the preferred method for performance when the BPF logic is
a stream parser.
The flow for Cilium's common use case with a stream parser is,
tcp_read_sock()
sk_psock_verdict_recv
ret = bpf_prog_run_pin_on_cpu()
sk_psock_verdict_apply(sock, skb, ret)
// if system is under memory pressure or app is slow we may
// need to queue skb. Do this queuing through ingress_skb and
// then kick timer to wake up handler
skb_queue_tail(ingress_skb, skb)
schedule_work(work);
The work queue is wired up to sk_psock_backlog(). This will then walk the
ingress_skb skb list that holds our sk_buffs that could not be handled,
but should be OK to run at some later point. However, its possible that
the workqueue doing this work still hits an error when sending the skb.
When this happens the skbuff is requeued on a temporary 'state' struct
kept with the workqueue. This is necessary because its possible to
partially send an skbuff before hitting an error and we need to know how
and where to restart when the workqueue runs next.
Now for the trouble, we don't rekick the workqueue. This can cause a
stall where the skbuff we just cached on the state variable might never
be sent. This happens when its the last packet in a flow and no further
packets come along that would cause the system to kick the workqueue from
that side.
To fix we could do simple schedule_work(), but while under memory pressure
it makes sense to back off some instead of continue to retry repeatedly. So
instead to fix convert schedule_work to schedule_delayed_work and add
backoff logic to reschedule from backlog queue on errors. Its not obvious
though what a good backoff is so use '1'.
To test we observed some flakes whil running NGINX compliance test with
sockmap we attributed these failed test to this bug and subsequent issue.
>From on list discussion. This commit
bec217197b41("skmsg: Schedule psock work if the cached skb exists on the psock")
was intended to address similar race, but had a couple cases it missed.
Most obvious it only accounted for receiving traffic on the local socket
so if redirecting into another socket we could still get an sk_buff stuck
here. Next it missed the case where copied=0 in the recv() handler and
then we wouldn't kick the scheduler. Also its sub-optimal to require
userspace to kick the internal mechanisms of sockmap to wake it up and
copy data to user. It results in an extra syscall and requires the app
to actual handle the EAGAIN correctly.
Fixes: 04919bed948dc ("tcp: Introduce tcp_read_skb()")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: William Findlay <will@isovalent.com>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-3-john.fastabend@gmail.com
syzbot reported a splat and bisected it to recent commit ed17aa92dc56 ("bpf,
sockmap: fix deadlocks in the sockhash and sockmap"):
[...]
WARNING: CPU: 1 PID: 9280 at kernel/softirq.c:376 __local_bh_enable_ip+0xbe/0x130 kernel/softirq.c:376
Modules linked in:
CPU: 1 PID: 9280 Comm: syz-executor.1 Not tainted 6.2.0-syzkaller-13249-gd319f344561d #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 03/30/2023
RIP: 0010:__local_bh_enable_ip+0xbe/0x130 kernel/softirq.c:376
[...]
Call Trace:
<TASK>
spin_unlock_bh include/linux/spinlock.h:395 [inline]
sock_map_del_link+0x2ea/0x510 net/core/sock_map.c:165
sock_map_unref+0xb0/0x1d0 net/core/sock_map.c:184
sock_hash_delete_elem+0x1ec/0x2a0 net/core/sock_map.c:945
map_delete_elem kernel/bpf/syscall.c:1536 [inline]
__sys_bpf+0x2edc/0x53e0 kernel/bpf/syscall.c:5053
__do_sys_bpf kernel/bpf/syscall.c:5166 [inline]
__se_sys_bpf kernel/bpf/syscall.c:5164 [inline]
__x64_sys_bpf+0x79/0xc0 kernel/bpf/syscall.c:5164
do_syscall_x64 arch/x86/entry/common.c:50 [inline]
do_syscall_64+0x39/0xb0 arch/x86/entry/common.c:80
entry_SYSCALL_64_after_hwframe+0x63/0xcd
RIP: 0033:0x7fe8f7c8c169
</TASK>
[...]
Revert for now until we have a proper solution.
Fixes: ed17aa92dc56 ("bpf, sockmap: fix deadlocks in the sockhash and sockmap")
Reported-by: syzbot+49f6cef45247ff249498@syzkaller.appspotmail.com
Cc: Hsin-Wei Hung <hsinweih@uci.edu>
Cc: Xin Liu <liuxin350@huawei.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/000000000000f1db9605f939720e@google.com/
When huang uses sched_switch tracepoint, the tracepoint
does only one thing in the mounted ebpf program, which
deletes the fixed elements in sockhash ([0])
It seems that elements in sockhash are rarely actively
deleted by users or ebpf program. Therefore, we do not
pay much attention to their deletion. Compared with hash
maps, sockhash only provides spin_lock_bh protection.
This causes it to appear to have self-locking behavior
in the interrupt context.
[0]:https://lore.kernel.org/all/CABcoxUayum5oOqFMMqAeWuS8+EzojquSOSyDA3J_2omY=2EeAg@mail.gmail.com/
Reported-by: Hsin-Wei Hung <hsinweih@uci.edu>
Fixes: 604326b41a6f ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: Xin Liu <liuxin350@huawei.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/r/20230406122622.109978-1-liuxin350@huawei.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
This patch changes the return types of bpf_map_ops functions to long, where
previously int was returned. Using long allows for bpf programs to maintain
the sign bit in the absence of sign extension during situations where
inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative
error is returned.
The definitions of the helper funcs are generated from comments in the bpf
uapi header at `include/uapi/linux/bpf.h`. The return type of these
helpers was previously changed from int to long in commit bdb7b79b4ce8. For
any case where one of the map helpers call the bpf_map_ops funcs that are
still returning 32-bit int, a compiler might not include sign extension
instructions to properly convert the 32-bit negative value a 64-bit
negative value.
For example:
bpf assembly excerpt of an inlined helper calling a kernel function and
checking for a specific error:
; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST);
...
46: call 0xffffffffe103291c ; htab_map_update_elem
; if (err && err != -EEXIST) {
4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax
kernel function assembly excerpt of return value from
`htab_map_update_elem` returning 32-bit int:
movl $0xffffffef, %r9d
...
movl %r9d, %eax
...results in the comparison:
cmp $0xffffffffffffffef, $0x00000000ffffffef
Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long")
Tested-by: Eduard Zingerman <eddyz87@gmail.com>
Signed-off-by: JP Kobryn <inwardvessel@gmail.com>
Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
sockmap and sockhash don't have something in common in allocation, so let's
introduce different helpers to calculate their memory usage.
The reuslt as follows,
- before
28: sockmap name count_map flags 0x0
key 4B value 4B max_entries 65536 memlock 524288B
29: sockhash name count_map flags 0x0
key 4B value 4B max_entries 65536 memlock 524288B
- after
28: sockmap name count_map flags 0x0
key 4B value 4B max_entries 65536 memlock 524608B
29: sockhash name count_map flags 0x0 <<<< no updated elements
key 4B value 4B max_entries 65536 memlock 1048896B
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Link: https://lore.kernel.org/r/20230305124615.12358-16-laoar.shao@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>