From: John Fastabend <john.fastabend@gmail.com> Date: 2021-03-24 21:00:30
This addresses an issue found while reviewing latest round of sock
map patches and an issue reported from CI via Andrii.
The CI discovered issue was introduced by over correcting our
previously broken memory accounting. After the fix, "bpf, sockmap:
Avoid returning unneeded EAGAIN when redirecting to self" we fixed
a dropped packet and a missing fwd_alloc calculations, but pushed
it too far back into the packet pipeline creating an issue in the
unlikely case socket tear down happens with an enqueued skb. See
patch for details.
Tested with usual suspects: test_sockmap, test_maps, test_progs
and test_progs-no_alu32.
---
John Fastabend (2):
bpf, sockmap: fix sk->prot unhash op reset
bpf, sockmap: fix incorrect fwd_alloc accounting
include/linux/skmsg.h | 1 -
net/core/skmsg.c | 13 ++++++-------
net/tls/tls_main.c | 6 ++++++
3 files changed, 12 insertions(+), 8 deletions(-)
--
Signature
From: John Fastabend <john.fastabend@gmail.com> Date: 2021-03-24 21:00:29
In '4da6a196f93b1' we fixed a potential unhash loop caused when
a TLS socket in a sockmap was removed from the sockmap. This
happened because the unhash operation on the TLS ctx continued
to point at the sockmap implementation of unhash even though the
psock has already been removed. The sockmap unhash handler when a
psock is removed does the following,
void sock_map_unhash(struct sock *sk)
{
void (*saved_unhash)(struct sock *sk);
struct sk_psock *psock;
rcu_read_lock();
psock = sk_psock(sk);
if (unlikely(!psock)) {
rcu_read_unlock();
if (sk->sk_prot->unhash)
sk->sk_prot->unhash(sk);
return;
}
[...]
}
The unlikely() case is there to handle the case where psock is detached
but the proto ops have not been updated yet. But, in the above case
with TLS and removed psock we never fixed sk_prot->unhash() and unhash()
points back to sock_map_unhash resulting in a loop. To fix this we added
this bit of code,
static inline void sk_psock_restore_proto(struct sock *sk,
struct sk_psock *psock)
{
sk->sk_prot->unhash = psock->saved_unhash;
This will set the sk_prot->unhash back to its saved value. This is the
correct callback for a TLS socket that has been removed from the sock_map.
Unfortunately, this also overwrites the unhash pointer for all psocks.
We effectively break sockmap unhash handling for any future socks.
Omitting the unhash operation will leave stale entries in the map if
a socket transition through unhash, but does not do close() op.
To fix handle similar to write_space and rewrite it in the TLS update
hook. This way the TLS enabled socket will point to the saved unhash()
handler.
Fixes: 4da6a196f93b1 ("bpf: Sockmap/tls, during free we may call tcp_bpf_unhash() in loop")
Reported-by: Cong Wang <redacted>
Reported-by: Lorenz Bauer <redacted>
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
---
include/linux/skmsg.h | 1 -
net/tls/tls_main.c | 6 ++++++
2 files changed, 6 insertions(+), 1 deletion(-)
@@ -754,6 +754,12 @@ static void tls_update(struct sock *sk, struct proto *p,ctx=tls_get_ctx(sk);if(likely(ctx)){+/* TLS does not have an unhash proto in SW cases, but we need+*toensurewestopusingthesock_mapunhashroutinebecause+*theassociatedpsockisbeingremoved.Sousetheoriginal+*unhashhandler.+*/+WRITE_ONCE(sk->sk_prot->unhash,p->unhash);ctx->sk_write_space=write_space;ctx->sk_proto=p;}else{
From: John Fastabend <john.fastabend@gmail.com> Date: 2021-03-24 21:01:01
Incorrect accounting fwd_alloc can result in a warning when the socket
is torn down,
[18455.319240] WARNING: CPU: 0 PID: 24075 at net/core/stream.c:208 sk_stream_kill_queues+0x21f/0x230
[...]
[18455.319543] Call Trace:
[18455.319556] inet_csk_destroy_sock+0xba/0x1f0
[18455.319577] tcp_rcv_state_process+0x1b4e/0x2380
[18455.319593] ? lock_downgrade+0x3a0/0x3a0
[18455.319617] ? tcp_finish_connect+0x1e0/0x1e0
[18455.319631] ? sk_reset_timer+0x15/0x70
[18455.319646] ? tcp_schedule_loss_probe+0x1b2/0x240
[18455.319663] ? lock_release+0xb2/0x3f0
[18455.319676] ? __release_sock+0x8a/0x1b0
[18455.319690] ? lock_downgrade+0x3a0/0x3a0
[18455.319704] ? lock_release+0x3f0/0x3f0
[18455.319717] ? __tcp_close+0x2c6/0x790
[18455.319736] ? tcp_v4_do_rcv+0x168/0x370
[18455.319750] tcp_v4_do_rcv+0x168/0x370
[18455.319767] __release_sock+0xbc/0x1b0
[18455.319785] __tcp_close+0x2ee/0x790
[18455.319805] tcp_close+0x20/0x80
This currently happens because on redirect case we do skb_set_owner_r()
with the original sock. This increments the fwd_alloc memory accounting
on the original sock. Then on redirect we may push this into the queue
of the psock we are redirecting to. When the skb is flushed from the
queue we give the memory back to the original sock. The problem is if
the original sock is destroyed/closed with skbs on another psocks queue
then the original sock will not have a way to reclaim the memory before
being destroyed. Then above warning will be thrown
sockA sockB
sk_psock_strp_read()
sk_psock_verdict_apply()
-- SK_REDIRECT --
sk_psock_skb_redirect()
skb_queue_tail(psock_other->ingress_skb..)
sk_close()
sock_map_unref()
sk_psock_put()
sk_psock_drop()
sk_psock_zap_ingress()
At this point we have torn down our own psock, but have the outstanding
skb in psock_other. Note that SK_PASS doesn't have this problem because
the sk_psock_drop() logic releases the skb, its still associated with
our psock.
To resolve lets only account for sockets on the ingress queue that are
still associated with the current socket. On the redirect case we will
check memory limits per 6fa9201a89898, but will omit fwd_alloc accounting
until skb is actually enqueued. When the skb is sent via skb_send_sock_locked
or received with sk_psock_skb_ingress memory will be claimed on psock_other.
Reported-by: Andrii Nakryiko <andrii@kernel.org>
Fixes: 6fa9201a89898 ("bpf, sockmap: Avoid returning unneeded EAGAIN when redirecting to self")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
---
net/core/skmsg.c | 13 ++++++-------
1 file changed, 6 insertions(+), 7 deletions(-)
@@ -808,10 +808,6 @@ int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)rcu_read_lock();prog=READ_ONCE(psock->progs.skb_verdict);if(likely(prog)){-/* We skip full set_owner_r here because if we do a SK_PASS-*orSK_DROPwecanskipskbmemoryaccountingandusethe-*TLScontext.-*/skb->sk=psock->sk;tcp_skb_bpf_redirect_clear(skb);ret=sk_psock_bpf_run(psock,prog,skb);
@@ -754,6 +754,12 @@ static void tls_update(struct sock *sk, struct proto *p,ctx=tls_get_ctx(sk);if(likely(ctx)){+/* TLS does not have an unhash proto in SW cases, but we need+*toensurewestopusingthesock_mapunhashroutinebecause+*theassociatedpsockisbeingremoved.Sousetheoriginal+*unhashhandler.+*/+WRITE_ONCE(sk->sk_prot->unhash,p->unhash);ctx->sk_write_space=write_space;ctx->sk_proto=p;
It looks awkward to update sk->sk_proto inside tls_update(),
at least when ctx!=NULL.
What is wrong with updating it in sk_psock_restore_proto()
when inet_csk_has_ulp() is true? It looks better to me.
From: Cong Wang <hidden> Date: 2021-03-25 00:45:39
On Wed, Mar 24, 2021 at 2:00 PM John Fastabend [off-list ref] wrote:
Incorrect accounting fwd_alloc can result in a warning when the socket
is torn down,
[18455.319240] WARNING: CPU: 0 PID: 24075 at net/core/stream.c:208 sk_stream_kill_queues+0x21f/0x230
[...]
[18455.319543] Call Trace:
[18455.319556] inet_csk_destroy_sock+0xba/0x1f0
[18455.319577] tcp_rcv_state_process+0x1b4e/0x2380
[18455.319593] ? lock_downgrade+0x3a0/0x3a0
[18455.319617] ? tcp_finish_connect+0x1e0/0x1e0
[18455.319631] ? sk_reset_timer+0x15/0x70
[18455.319646] ? tcp_schedule_loss_probe+0x1b2/0x240
[18455.319663] ? lock_release+0xb2/0x3f0
[18455.319676] ? __release_sock+0x8a/0x1b0
[18455.319690] ? lock_downgrade+0x3a0/0x3a0
[18455.319704] ? lock_release+0x3f0/0x3f0
[18455.319717] ? __tcp_close+0x2c6/0x790
[18455.319736] ? tcp_v4_do_rcv+0x168/0x370
[18455.319750] tcp_v4_do_rcv+0x168/0x370
[18455.319767] __release_sock+0xbc/0x1b0
[18455.319785] __tcp_close+0x2ee/0x790
[18455.319805] tcp_close+0x20/0x80
This currently happens because on redirect case we do skb_set_owner_r()
with the original sock. This increments the fwd_alloc memory accounting
on the original sock. Then on redirect we may push this into the queue
of the psock we are redirecting to. When the skb is flushed from the
queue we give the memory back to the original sock. The problem is if
the original sock is destroyed/closed with skbs on another psocks queue
then the original sock will not have a way to reclaim the memory before
being destroyed. Then above warning will be thrown
sockA sockB
sk_psock_strp_read()
sk_psock_verdict_apply()
-- SK_REDIRECT --
sk_psock_skb_redirect()
skb_queue_tail(psock_other->ingress_skb..)
sk_close()
sock_map_unref()
sk_psock_put()
sk_psock_drop()
sk_psock_zap_ingress()
At this point we have torn down our own psock, but have the outstanding
skb in psock_other. Note that SK_PASS doesn't have this problem because
the sk_psock_drop() logic releases the skb, its still associated with
our psock.
To resolve lets only account for sockets on the ingress queue that are
still associated with the current socket. On the redirect case we will
check memory limits per 6fa9201a89898, but will omit fwd_alloc accounting
until skb is actually enqueued. When the skb is sent via skb_send_sock_locked
or received with sk_psock_skb_ingress memory will be claimed on psock_other.
@@ -808,10 +808,6 @@ int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)rcu_read_lock();prog=READ_ONCE(psock->progs.skb_verdict);if(likely(prog)){-/* We skip full set_owner_r here because if we do a SK_PASS-*orSK_DROPwecanskipskbmemoryaccountingandusethe-*TLScontext.-*/skb->sk=psock->sk;tcp_skb_bpf_redirect_clear(skb);ret=sk_psock_bpf_run(psock,prog,skb);
@@ -754,6 +754,12 @@ static void tls_update(struct sock *sk, struct proto *p,ctx=tls_get_ctx(sk);if(likely(ctx)){+/* TLS does not have an unhash proto in SW cases, but we need+*toensurewestopusingthesock_mapunhashroutinebecause+*theassociatedpsockisbeingremoved.Sousetheoriginal+*unhashhandler.+*/+WRITE_ONCE(sk->sk_prot->unhash,p->unhash);ctx->sk_write_space=write_space;ctx->sk_proto=p;
It looks awkward to update sk->sk_proto inside tls_update(),
at least when ctx!=NULL.
hmm. It doesn't strike me as paticularly awkward but OK.
What is wrong with updating it in sk_psock_restore_proto()
when inet_csk_has_ulp() is true? It looks better to me.
It could be wrong if inet_csk_has_ulp has an unhash callback
already assigned. But, because we know inet_csk_has_ulp()
really means is_tls_attached() it would be fine.
From: John Fastabend <john.fastabend@gmail.com> Date: 2021-03-25 02:47:02
Cong Wang wrote:
On Wed, Mar 24, 2021 at 2:00 PM John Fastabend [off-list ref] wrote:
quoted
Incorrect accounting fwd_alloc can result in a warning when the socket
is torn down,
[...]
quoted
To resolve lets only account for sockets on the ingress queue that are
still associated with the current socket. On the redirect case we will
check memory limits per 6fa9201a89898, but will omit fwd_alloc accounting
until skb is actually enqueued. When the skb is sent via skb_send_sock_locked
or received with sk_psock_skb_ingress memory will be claimed on psock_other.
These come from strparser which do not have skb->sk set.
Nit: You can just use 'sk' here, so "skb->sk = sk".
Sure that is a bit nicer, will respin with this.
quoted
tcp_skb_bpf_redirect_clear(skb);
ret = sk_psock_bpf_run(psock, prog, skb);
ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
+ skb->sk = NULL;
Why do you want to set it to NULL here?
So we don't cause the stack to throw other errors later if we
were to call skb_orphan for example. Various places in the skb
helpers expect both skb->sk and skb->destructor to be set together
and here we are just using it as a mechanism to feed the sk into
the BPF program side. The above skb_set_owner_r for example
would likely BUG().
@@ -754,6 +754,12 @@ static void tls_update(struct sock *sk, struct proto *p,ctx=tls_get_ctx(sk);if(likely(ctx)){+/* TLS does not have an unhash proto in SW cases, but we need+*toensurewestopusingthesock_mapunhashroutinebecause+*theassociatedpsockisbeingremoved.Sousetheoriginal+*unhashhandler.+*/+WRITE_ONCE(sk->sk_prot->unhash,p->unhash);ctx->sk_write_space=write_space;ctx->sk_proto=p;
It looks awkward to update sk->sk_proto inside tls_update(),
at least when ctx!=NULL.
hmm. It doesn't strike me as paticularly awkward but OK.
I read tls_update() as "updating ctx when it is initialized", with your
patch, we are updating sk->sk_prot->unhash too when updating ctx,
pretty much like a piggyback, hence it reads odd to me.
Thanks.
From: Cong Wang <hidden> Date: 2021-03-25 19:28:26
On Wed, Mar 24, 2021 at 7:46 PM John Fastabend [off-list ref] wrote:
Cong Wang wrote:
quoted
On Wed, Mar 24, 2021 at 2:00 PM John Fastabend [off-list ref] wrote:
quoted
Incorrect accounting fwd_alloc can result in a warning when the socket
is torn down,
[...]
quoted
quoted
To resolve lets only account for sockets on the ingress queue that are
still associated with the current socket. On the redirect case we will
check memory limits per 6fa9201a89898, but will omit fwd_alloc accounting
until skb is actually enqueued. When the skb is sent via skb_send_sock_locked
or received with sk_psock_skb_ingress memory will be claimed on psock_other.
^^^^^^^^^^^^^^^^^^^^
quoted
You mean sk_psock_skb_ingress(), right?
Yes.
skb_send_sock_locked() actually allocates its own skb when sending, hence
it uses a different skb for memory accounting.
These come from strparser which do not have skb->sk set.
Hmm, but sk_psock_verdict_recv() passes a clone too, like
strparser, so either we need it for both, or not at all. Clones
do not have skb->sk, so I think you can remove the one in
sk_psock_verdict_recv() too.
quoted
Nit: You can just use 'sk' here, so "skb->sk = sk".
Sure that is a bit nicer, will respin with this.
quoted
quoted
tcp_skb_bpf_redirect_clear(skb);
ret = sk_psock_bpf_run(psock, prog, skb);
ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
+ skb->sk = NULL;
Why do you want to set it to NULL here?
So we don't cause the stack to throw other errors later if we
were to call skb_orphan for example. Various places in the skb
helpers expect both skb->sk and skb->destructor to be set together
and here we are just using it as a mechanism to feed the sk into
the BPF program side. The above skb_set_owner_r for example
would likely BUG().
@@ -754,6 +754,12 @@ static void tls_update(struct sock *sk, struct proto *p,ctx=tls_get_ctx(sk);if(likely(ctx)){+/* TLS does not have an unhash proto in SW cases, but we need+*toensurewestopusingthesock_mapunhashroutinebecause+*theassociatedpsockisbeingremoved.Sousetheoriginal+*unhashhandler.+*/+WRITE_ONCE(sk->sk_prot->unhash,p->unhash);ctx->sk_write_space=write_space;ctx->sk_proto=p;
It looks awkward to update sk->sk_proto inside tls_update(),
at least when ctx!=NULL.
hmm. It doesn't strike me as paticularly awkward but OK.
I read tls_update() as "updating ctx when it is initialized", with your
patch, we are updating sk->sk_prot->unhash too when updating ctx,
pretty much like a piggyback, hence it reads odd to me.
Thanks.
From: John Fastabend <john.fastabend@gmail.com> Date: 2021-03-26 01:00:10
Cong Wang wrote:
On Wed, Mar 24, 2021 at 7:46 PM John Fastabend [off-list ref] wrote:
quoted
Cong Wang wrote:
quoted
On Wed, Mar 24, 2021 at 2:00 PM John Fastabend [off-list ref] wrote:
quoted
Incorrect accounting fwd_alloc can result in a warning when the socket
is torn down,
[...]
quoted
quoted
To resolve lets only account for sockets on the ingress queue that are
still associated with the current socket. On the redirect case we will
check memory limits per 6fa9201a89898, but will omit fwd_alloc accounting
until skb is actually enqueued. When the skb is sent via skb_send_sock_locked
or received with sk_psock_skb_ingress memory will be claimed on psock_other.
^^^^^^^^^^^^^^^^^^^^
quoted
You mean sk_psock_skb_ingress(), right?
Yes.
skb_send_sock_locked() actually allocates its own skb when sending, hence
it uses a different skb for memory accounting.
These come from strparser which do not have skb->sk set.
Hmm, but sk_psock_verdict_recv() passes a clone too, like
strparser, so either we need it for both, or not at all. Clones
do not have skb->sk, so I think you can remove the one in
sk_psock_verdict_recv() too.
Agree skb_orphan can just be removed, I was being overly
paranoid.
Thanks.