Re: [PATCH v2] Revert "vrf: Remove unnecessary RCU-bh critical section"

Ben Greear wrote:

On 9/26/24 02:03, Willem de Bruijn wrote:

quoted

greearb@ wrote:

quoted

From: Ben Greear <redacted>

This reverts commit 504fc6f4f7f681d2a03aa5f68aad549d90eab853.

dev_queue_xmit_nit needs to run with bh locking, otherwise
it conflicts with packets coming in from a nic in softirq
context and packets being transmitted from user context.

================================
WARNING: inconsistent lock state
6.11.0 #1 Tainted: G        W
--------------------------------
inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage.
btserver/134819 [HC0[0]:SC0[0]:HE1:SE1] takes:
ffff8882da30c118 (rlock-AF_PACKET){+.?.}-{2:2}, at: tpacket_rcv+0x863/0x3b30
{IN-SOFTIRQ-W} state was registered at:
   lock_acquire+0x19a/0x4f0
   _raw_spin_lock+0x27/0x40
   packet_rcv+0xa33/0x1320
   __netif_receive_skb_core.constprop.0+0xcb0/0x3a90
   __netif_receive_skb_list_core+0x2c9/0x890
   netif_receive_skb_list_internal+0x610/0xcc0
   napi_complete_done+0x1c0/0x7c0
   igb_poll+0x1dbb/0x57e0 [igb]
   __napi_poll.constprop.0+0x99/0x430
   net_rx_action+0x8e7/0xe10
   handle_softirqs+0x1b7/0x800
   __irq_exit_rcu+0x91/0xc0
   irq_exit_rcu+0x5/0x10
   [snip]

other info that might help us debug this:
  Possible unsafe locking scenario:

        CPU0
        ----
   lock(rlock-AF_PACKET);
   <Interrupt>
     lock(rlock-AF_PACKET);

  *** DEADLOCK ***

Call Trace:
  <TASK>
  dump_stack_lvl+0x73/0xa0
  mark_lock+0x102e/0x16b0
  __lock_acquire+0x9ae/0x6170
  lock_acquire+0x19a/0x4f0
  _raw_spin_lock+0x27/0x40
  tpacket_rcv+0x863/0x3b30
  dev_queue_xmit_nit+0x709/0xa40
  vrf_finish_direct+0x26e/0x340 [vrf]
  vrf_l3_out+0x5f4/0xe80 [vrf]
  __ip_local_out+0x51e/0x7a0
[snip]

Fixes: 504fc6f4f7f6 ("vrf: Remove unnecessary RCU-bh critical section")
Link: https://lore.kernel.org/netdev/05765015-f727-2f30-58da-2ad6fa7ea99f@candelatech.com/T/ (local)

Signed-off-by: Ben Greear <redacted>

Please Cc: all previous reviewers and folks who participated in the
discussion. I entirely missed this. No need to add as Cc tags, just
--cc in git send-email will do.

quoted

---

v2:  Edit patch description.

  drivers/net/vrf.c | 2 ++
  net/core/dev.c    | 1 +
  2 files changed, 3 insertions(+)

diff --git a/drivers/net/vrf.c b/drivers/net/vrf.c
index 4d8ccaf9a2b4..4087f72f0d2b 100644
--- a/drivers/net/vrf.c
+++ b/drivers/net/vrf.c

@@ -608,7 +608,9 @@ static void vrf_finish_direct(struct sk_buff *skb)
  		eth_zero_addr(eth->h_dest);
  		eth->h_proto = skb->protocol;
  
+		rcu_read_lock_bh();
  		dev_queue_xmit_nit(skb, vrf_dev);
+		rcu_read_unlock_bh();
  
  		skb_pull(skb, ETH_HLEN);
  	}

diff --git a/net/core/dev.c b/net/core/dev.c
index cd479f5f22f6..566e69a38eed 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -2285,6 +2285,7 @@ EXPORT_SYMBOL_GPL(dev_nit_active);
  /*
   *	Support routine. Sends outgoing frames to any network
   *	taps currently in use.
+ *	BH must be disabled before calling this.

Unnecessary. Especially patches for stable should be minimal to
reduce the chance of conflicts.

I was asked to add this, and also asked to CC stable.

Conflicting feedback is annoying, but I disagree with the other
comment.

Not only does it potentially complicate backporting, it also is no
longer a strict revert. In which case it must not be labeled as such.

Easier is to keep the revert unmodified, and add the comment to the
commit message.

Most important is the Link to our earlier thread that explains the
history.

If the process appears byzantine, if you prefer I can also send it.

Thanks,

  Willem