Skip to content

L
linux
Switch branch/tag
  1. 23 May, 2023 10 commits
    • John Fastabend's avatar
      bpf, sockmap: Build helper to create connected socket pair · 298970c8
      John Fastabend authored 
      A common operation for testing is to spin up a pair of sockets that are
connected. Then we can use these to run specific tests that need to
send data, check BPF programs and so on.

The sockmap_listen programs already have this logic lets move it into
the new sockmap_helpers header file for general use.
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-11-john.fastabend@gmail.com
      298970c8
    • John Fastabend's avatar
      bpf, sockmap: Pull socket helpers out of listen test for general use · 4e02588d
      John Fastabend authored 
      No functional change here we merely pull the helpers in sockmap_listen.c
into a header file so we can use these in other programs. The tests we
are about to add aren't really _listen tests so doesn't make sense
to add them here.
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-10-john.fastabend@gmail.com
      4e02588d
    • John Fastabend's avatar
      bpf, sockmap: Incorrectly handling copied_seq · e5c6de5f
      John Fastabend authored 
      The read_skb() logic is incrementing the tcp->copied_seq which is used for
among other things calculating how many outstanding bytes can be read by
the application. This results in application errors, if the application
does an ioctl(FIONREAD) we return zero because this is calculated from
the copied_seq value.

To fix this we move tcp->copied_seq accounting into the recv handler so
that we update these when the recvmsg() hook is called and data is in
fact copied into user buffers. This gives an accurate FIONREAD value
as expected and improves ACK handling. Before we were calling the
tcp_rcv_space_adjust() which would update 'number of bytes copied to
user in last RTT' which is wrong for programs returning SK_PASS. The
bytes are only copied to the user when recvmsg is handled.

Doing the fix for recvmsg is straightforward, but fixing redirect and
SK_DROP pkts is a bit tricker. Build a tcp_psock_eat() helper and then
call this from skmsg handlers. This fixes another issue where a broken
socket with a BPF program doing a resubmit could hang the receiver. This
happened because although read_skb() consumed the skb through sock_drop()
it did not update the copied_seq. Now if a single reccv socket is
redirecting to many sockets (for example for lb) the receiver sk will be
hung even though we might expect it to continue. The hang comes from
not updating the copied_seq numbers and memory pressure resulting from
that.

We have a slight layer problem of calling tcp_eat_skb even if its not
a TCP socket. To fix we could refactor and create per type receiver
handlers. I decided this is more work than we want in the fix and we
already have some small tweaks depending on caller that use the
helper skb_bpf_strparser(). So we extend that a bit and always set
the strparser bit when it is in use and then we can gate the
seq_copied updates on this.

Fixes: 04919bed ("tcp: Introduce tcp_read_skb()")
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-9-john.fastabend@gmail.com
      e5c6de5f
    • John Fastabend's avatar
      bpf, sockmap: Wake up polling after data copy · 6df7f764
      John Fastabend authored 
      When TCP stack has data ready to read sk_data_ready() is called. Sockmap
overwrites this with its own handler to call into BPF verdict program.
But, the original TCP socket had sock_def_readable that would additionally
wake up any user space waiters with sk_wake_async().

Sockmap saved the callback when the socket was created so call the saved
data ready callback and then we can wake up any epoll() logic waiting
on the read.

Note we call on 'copied >= 0' to account for returning 0 when a FIN is
received because we need to wake up user for this as well so they
can do the recvmsg() -> 0 and detect the shutdown.

Fixes: 04919bed ("tcp: Introduce tcp_read_skb()")
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-8-john.fastabend@gmail.com
      6df7f764
    • John Fastabend's avatar
      bpf, sockmap: TCP data stall on recv before accept · ea444185
      John Fastabend authored 
      A common mechanism to put a TCP socket into the sockmap is to hook the
BPF_SOCK_OPS_{ACTIVE_PASSIVE}_ESTABLISHED_CB event with a BPF program
that can map the socket info to the correct BPF verdict parser. When
the user adds the socket to the map the psock is created and the new
ops are assigned to ensure the verdict program will 'see' the sk_buffs
as they arrive.

Part of this process hooks the sk_data_ready op with a BPF specific
handler to wake up the BPF verdict program when data is ready to read.
The logic is simple enough (posted here for easy reading)

 static void sk_psock_verdict_data_ready(struct sock *sk)
 {
	struct socket *sock = sk->sk_socket;

	if (unlikely(!sock || !sock->ops || !sock->ops->read_skb))
		return;
	sock->ops->read_skb(sk, sk_psock_verdict_recv);
 }

The oversight here is sk->sk_socket is not assigned until the application
accepts() the new socket. However, its entirely ok for the peer application
to do a connect() followed immediately by sends. The socket on the receiver
is sitting on the backlog queue of the listening socket until its accepted
and the data is queued up. If the peer never accepts the socket or is slow
it will eventually hit data limits and rate limit the session. But,
important for BPF sockmap hooks when this data is received TCP stack does
the sk_data_ready() call but the read_skb() for this data is never called
because sk_socket is missing. The data sits on the sk_receive_queue.

Then once the socket is accepted if we never receive more data from the
peer there will be no further sk_data_ready calls and all the data
is still on the sk_receive_queue(). Then user calls recvmsg after accept()
and for TCP sockets in sockmap we use the tcp_bpf_recvmsg_parser() handler.
The handler checks for data in the sk_msg ingress queue expecting that
the BPF program has already run from the sk_data_ready hook and enqueued
the data as needed. So we are stuck.

To fix do an unlikely check in recvmsg handler for data on the
sk_receive_queue and if it exists wake up data_ready. We have the sock
locked in both read_skb and recvmsg so should avoid having multiple
runners.

Fixes: 04919bed ("tcp: Introduce tcp_read_skb()")
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-7-john.fastabend@gmail.com
      ea444185
    • John Fastabend's avatar
      bpf, sockmap: Handle fin correctly · 901546fd
      John Fastabend authored 
      The sockmap code is returning EAGAIN after a FIN packet is received and no
more data is on the receive queue. Correct behavior is to return 0 to the
user and the user can then close the socket. The EAGAIN causes many apps
to retry which masks the problem. Eventually the socket is evicted from
the sockmap because its released from sockmap sock free handling. The
issue creates a delay and can cause some errors on application side.

To fix this check on sk_msg_recvmsg side if length is zero and FIN flag
is set then set return to zero. A selftest will be added to check this
condition.

Fixes: 04919bed ("tcp: Introduce tcp_read_skb()")
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Tested-by: default avatarWilliam Findlay <will@isovalent.com>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-6-john.fastabend@gmail.com
      901546fd
    • John Fastabend's avatar
      bpf, sockmap: Improved check for empty queue · 405df89d
      John Fastabend authored 
      We noticed some rare sk_buffs were stepping past the queue when system was
under memory pressure. The general theory is to skip enqueueing
sk_buffs when its not necessary which is the normal case with a system
that is properly provisioned for the task, no memory pressure and enough
cpu assigned.

But, if we can't allocate memory due to an ENOMEM error when enqueueing
the sk_buff into the sockmap receive queue we push it onto a delayed
workqueue to retry later. When a new sk_buff is received we then check
if that queue is empty. However, there is a problem with simply checking
the queue length. When a sk_buff is being processed from the ingress queue
but not yet on the sockmap msg receive queue its possible to also recv
a sk_buff through normal path. It will check the ingress queue which is
zero and then skip ahead of the pkt being processed.

Previously we used sock lock from both contexts which made the problem
harder to hit, but not impossible.

To fix instead of popping the skb from the queue entirely we peek the
skb from the queue and do the copy there. This ensures checks to the
queue length are non-zero while skb is being processed. Then finally
when the entire skb has been copied to user space queue or another
socket we pop it off the queue. This way the queue length check allows
bypassing the queue only after the list has been completely processed.

To reproduce issue we run NGINX compliance test with sockmap running and
observe some flakes in our testing that we attributed to this issue.

Fixes: 04919bed ("tcp: Introduce tcp_read_skb()")
Suggested-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Tested-by: default avatarWilliam Findlay <will@isovalent.com>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-5-john.fastabend@gmail.com
      405df89d
    • John Fastabend's avatar
      bpf, sockmap: Reschedule is now done through backlog · bce22552
      John Fastabend authored 
      Now that the backlog manages the reschedule() logic correctly we can drop
the partial fix to reschedule from recvmsg hook.

Rescheduling on recvmsg hook was added to address a corner case where we
still had data in the backlog state but had nothing to kick it and
reschedule the backlog worker to run and finish copying data out of the
state. This had a couple limitations, first it required user space to
kick it introducing an unnecessary EBUSY and retry. Second it only
handled the ingress case and egress redirects would still be hung.

With the correct fix, pushing the reschedule logic down to where the
enomem error occurs we can drop this fix.

Fixes: bec21719 ("skmsg: Schedule psock work if the cached skb exists on the psock")
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-4-john.fastabend@gmail.com
      bce22552
    • John Fastabend's avatar
      bpf, sockmap: Convert schedule_work into delayed_work · 29173d07
      John Fastabend authored 
      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

 bec21719("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: 04919bed ("tcp: Introduce tcp_read_skb()")
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Tested-by: default avatarWilliam Findlay <will@isovalent.com>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-3-john.fastabend@gmail.com
      29173d07
    • John Fastabend's avatar
      bpf, sockmap: Pass skb ownership through read_skb · 78fa0d61
      John Fastabend authored 
      The read_skb hook calls consume_skb() now, but this means that if the
recv_actor program wants to use the skb it needs to inc the ref cnt
so that the consume_skb() doesn't kfree the sk_buff.

This is problematic because in some error cases under memory pressure
we may need to linearize the sk_buff from sk_psock_skb_ingress_enqueue().
Then we get this,

 skb_linearize()
   __pskb_pull_tail()
     pskb_expand_head()
       BUG_ON(skb_shared(skb))

Because we incremented users refcnt from sk_psock_verdict_recv() we
hit the bug on with refcnt > 1 and trip it.

To fix lets simply pass ownership of the sk_buff through the skb_read
call. Then we can drop the consume from read_skb handlers and assume
the verdict recv does any required kfree.

Bug found while testing in our CI which runs in VMs that hit memory
constraints rather regularly. William tested TCP read_skb handlers.

[  106.536188] ------------[ cut here ]------------
[  106.536197] kernel BUG at net/core/skbuff.c:1693!
[  106.536479] invalid opcode: 0000 [#1] PREEMPT SMP PTI
[  106.536726] CPU: 3 PID: 1495 Comm: curl Not tainted 5.19.0-rc5 #1
[  106.537023] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ArchLinux 1.16.0-1 04/01/2014
[  106.537467] RIP: 0010:pskb_expand_head+0x269/0x330
[  106.538585] RSP: 0018:ffffc90000138b68 EFLAGS: 00010202
[  106.538839] RAX: 000000000000003f RBX: ffff8881048940e8 RCX: 0000000000000a20
[  106.539186] RDX: 0000000000000002 RSI: 0000000000000000 RDI: ffff8881048940e8
[  106.539529] RBP: ffffc90000138be8 R08: 00000000e161fd1a R09: 0000000000000000
[  106.539877] R10: 0000000000000018 R11: 0000000000000000 R12: ffff8881048940e8
[  106.540222] R13: 0000000000000003 R14: 0000000000000000 R15: ffff8881048940e8
[  106.540568] FS:  00007f277dde9f00(0000) GS:ffff88813bd80000(0000) knlGS:0000000000000000
[  106.540954] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  106.541227] CR2: 00007f277eeede64 CR3: 000000000ad3e000 CR4: 00000000000006e0
[  106.541569] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  106.541915] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[  106.542255] Call Trace:
[  106.542383]  <IRQ>
[  106.542487]  __pskb_pull_tail+0x4b/0x3e0
[  106.542681]  skb_ensure_writable+0x85/0xa0
[  106.542882]  sk_skb_pull_data+0x18/0x20
[  106.543084]  bpf_prog_b517a65a242018b0_bpf_skskb_http_verdict+0x3a9/0x4aa9
[  106.543536]  ? migrate_disable+0x66/0x80
[  106.543871]  sk_psock_verdict_recv+0xe2/0x310
[  106.544258]  ? sk_psock_write_space+0x1f0/0x1f0
[  106.544561]  tcp_read_skb+0x7b/0x120
[  106.544740]  tcp_data_queue+0x904/0xee0
[  106.544931]  tcp_rcv_established+0x212/0x7c0
[  106.545142]  tcp_v4_do_rcv+0x174/0x2a0
[  106.545326]  tcp_v4_rcv+0xe70/0xf60
[  106.545500]  ip_protocol_deliver_rcu+0x48/0x290
[  106.545744]  ip_local_deliver_finish+0xa7/0x150

Fixes: 04919bed ("tcp: Introduce tcp_read_skb()")
Reported-by: default avatarWilliam Findlay <will@isovalent.com>
Signed-off-by: default avatarJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Tested-by: default avatarWilliam Findlay <will@isovalent.com>
Reviewed-by: default avatarJakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20230523025618.113937-2-john.fastabend@gmail.com
      78fa0d61
  3. 22 May, 2023 1 commit
  5. 19 May, 2023 1 commit
    • Will Deacon's avatar
      bpf: Fix mask generation for 32-bit narrow loads of 64-bit fields · 0613d8ca
      Will Deacon authored 
      A narrow load from a 64-bit context field results in a 64-bit load
followed potentially by a 64-bit right-shift and then a bitwise AND
operation to extract the relevant data.

In the case of a 32-bit access, an immediate mask of 0xffffffff is used
to construct a 64-bit BPP_AND operation which then sign-extends the mask
value and effectively acts as a glorified no-op. For example:

0:	61 10 00 00 00 00 00 00	r0 = *(u32 *)(r1 + 0)

results in the following code generation for a 64-bit field:

	ldr	x7, [x7]	// 64-bit load
	mov	x10, #0xffffffffffffffff
	and	x7, x7, x10

Fix the mask generation so that narrow loads always perform a 32-bit AND
operation:

	ldr	x7, [x7]	// 64-bit load
	mov	w10, #0xffffffff
	and	w7, w7, w10

Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: Krzesimir Nowak <krzesimir@kinvolk.io>
Cc: Andrey Ignatov <rdna@fb.com>
Acked-by: default avatarYonghong Song <yhs@fb.com>
Fixes: 31fd8581 ("bpf: permits narrower load from bpf program context fields")
Signed-off-by: default avatarWill Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20230518102528.1341-1-will@kernel.orgSigned-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
      0613d8ca
  7. 16 May, 2023 1 commit
  9. 15 May, 2023 1 commit
  11. 14 May, 2023 1 commit
  13. 13 May, 2023 10 commits
  15. 12 May, 2023 15 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7