Maciej Fijalkowski wrote:
> ice_qp_dis() intends to stop a given queue pair that is a target of xsk
> pool attach/detach. One of the steps is to disable interrupts on these
> queues. It currently is broken in a way that txq irq is turned off
> *after* HW flush which in turn takes no effect.
>
> ice_qp_dis():
> -> ice_qvec_dis_irq()
> --> disable rxq irq
> --> flush hw
> -> ice_vsi_stop_tx_ring()
> -->disable txq irq
>
> Below splat can be triggered by following steps:
> - start xdpsock WITHOUT loading xdp prog
> - run xdp_rxq_info with XDP_TX action on this interface
> - start traffic
> - terminate xdpsock
>
> [ 256.312485] BUG: kernel NULL pointer dereference, address: 0000000000000018
> [ 256.319560] #PF: supervisor read access in kernel mode
> [ 256.324775] #PF: error_code(0x0000) - not-present page
> [ 256.329994] PGD 0 P4D 0
> [ 256.332574] Oops: 0000 [#1] PREEMPT SMP NOPTI
> [ 256.337006] CPU: 3 PID: 32 Comm: ksoftirqd/3 Tainted: G OE 6.2.0-rc5+ #51
> [ 256.345218] Hardware name: Intel Corporation S2600WFT/S2600WFT, BIOS SE5C620.86B.02.01.0008.031920191559 03/19/2019
> [ 256.355807] RIP: 0010:ice_clean_rx_irq_zc+0x9c/0x7d0 [ice]
> [ 256.361423] Code: b7 8f 8a 00 00 00 66 39 ca 0f 84 f1 04 00 00 49 8b 47 40 4c 8b 24 d0 41 0f b7 45 04 66 25 ff 3f 66 89 04 24 0f 84 85 02 00 00 <49> 8b 44 24 18 0f b7 14 24 48 05 00 01 00 00 49 89 04 24 49 89 44
> [ 256.380463] RSP: 0018:ffffc900088bfd20 EFLAGS: 00010206
> [ 256.385765] RAX: 000000000000003c RBX: 0000000000000035 RCX: 000000000000067f
> [ 256.393012] RDX: 0000000000000775 RSI: 0000000000000000 RDI: ffff8881deb3ac80
> [ 256.400256] RBP: 000000000000003c R08: ffff889847982710 R09: 0000000000010000
> [ 256.407500] R10: ffffffff82c060c0 R11: 0000000000000004 R12: 0000000000000000
> [ 256.414746] R13: ffff88811165eea0 R14: ffffc9000d255000 R15: ffff888119b37600
> [ 256.421990] FS: 0000000000000000(0000) GS:ffff8897e0cc0000(0000) knlGS:0000000000000000
> [ 256.430207] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [ 256.436036] CR2: 0000000000000018 CR3: 0000000005c0a006 CR4: 00000000007706e0
> [ 256.443283] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [ 256.450527] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [ 256.457770] PKRU: 55555554
> [ 256.460529] Call Trace:
> [ 256.463015] <TASK>
> [ 256.465157] ? ice_xmit_zc+0x6e/0x150 [ice]
> [ 256.469437] ice_napi_poll+0x46d/0x680 [ice]
> [ 256.473815] ? _raw_spin_unlock_irqrestore+0x1b/0x40
> [ 256.478863] __napi_poll+0x29/0x160
> [ 256.482409] net_rx_action+0x136/0x260
> [ 256.486222] __do_softirq+0xe8/0x2e5
> [ 256.489853] ? smpboot_thread_fn+0x2c/0x270
> [ 256.494108] run_ksoftirqd+0x2a/0x50
> [ 256.497747] smpboot_thread_fn+0x1c1/0x270
> [ 256.501907] ? __pfx_smpboot_thread_fn+0x10/0x10
> [ 256.506594] kthread+0xea/0x120
> [ 256.509785] ? __pfx_kthread+0x10/0x10
> [ 256.513597] ret_from_fork+0x29/0x50
> [ 256.517238] </TASK>
>
> In fact, irqs were not disabled and napi managed to be scheduled and run
> while xsk_pool pointer was still valid, but SW ring of xdp_buff pointers
> was already freed.
>
> To fix this, call ice_qvec_dis_irq() after ice_vsi_stop_tx_ring(). Also
> while at it, remove redundant ice_clean_rx_ring() call - this is handled
> in ice_qp_clean_rings().
>
> Fixes: 2d4238f55697 ("ice: Add support for AF_XDP")
> Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx>
Looks reasonable to me for what its worth.
Acked-by: John Fastabend <john.fastabend@xxxxxxxxx>
