On Mon, May 15, 2023 at 03:52:47PM +0200, Maciej Fijalkowski wrote:
> The ice driver caches next_to_clean value at the beginning of
> ice_clean_rx_irq() in order to remember the first buffer that has to be
> freed/recycled after main Rx processing loop. The end boundary is
> indicated by first descriptor of frame that Rx processing loop has ended
> its duties. Note that if mentioned loop ended in the middle of gathering
> multi-buffer frame, next_to_clean would be pointing to the descriptor in
> the middle of the frame BUT freeing/recycling stage will stop at the
> first descriptor. This means that next iteration of ice_clean_rx_irq()
> will miss the (first_desc, next_to_clean - 1) entries.
>
> When running various 9K MTU workloads, such splats were observed:
>
> [ 540.780716] BUG: kernel NULL pointer dereference, address: 0000000000000000
> [ 540.787787] #PF: supervisor read access in kernel mode
> [ 540.793002] #PF: error_code(0x0000) - not-present page
> [ 540.798218] PGD 0 P4D 0
> [ 540.800801] Oops: 0000 [#1] PREEMPT SMP NOPTI
> [ 540.805231] CPU: 18 PID: 3984 Comm: xskxceiver Tainted: G W 6.3.0-rc7+ #96
> [ 540.813619] Hardware name: Intel Corporation S2600WFT/S2600WFT, BIOS SE5C620.86B.02.01.0008.031920191559 03/19/2019
> [ 540.824209] RIP: 0010:ice_clean_rx_irq+0x2b6/0xf00 [ice]
> [ 540.829678] Code: 74 24 10 e9 aa 00 00 00 8b 55 78 41 31 57 10 41 09 c4 4d 85 ff 0f 84 83 00 00 00 49 8b 57 08 41 8b 4f 1c 65 8b 35 1a fa 4b 3f <48> 8b 02 48 c1 e8 3a 39 c6 0f 85 a2 00 00 00 f6 42 08 02 0f 85 98
> [ 540.848717] RSP: 0018:ffffc9000f42fc50 EFLAGS: 00010282
> [ 540.854029] RAX: 0000000000000004 RBX: 0000000000000002 RCX: 000000000000fffe
> [ 540.861272] RDX: 0000000000000000 RSI: 0000000000000001 RDI: 00000000ffffffff
> [ 540.868519] RBP: ffff88984a05ac00 R08: 0000000000000000 R09: dead000000000100
> [ 540.875760] R10: ffff88983fffcd00 R11: 000000000010f2b8 R12: 0000000000000004
> [ 540.883008] R13: 0000000000000003 R14: 0000000000000800 R15: ffff889847a10040
> [ 540.890253] FS: 00007f6ddf7fe640(0000) GS:ffff88afdf800000(0000) knlGS:0000000000000000
> [ 540.898465] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [ 540.904299] CR2: 0000000000000000 CR3: 000000010d3da001 CR4: 00000000007706e0
> [ 540.911542] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [ 540.918789] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [ 540.926032] PKRU: 55555554
> [ 540.928790] Call Trace:
> [ 540.931276] <TASK>
> [ 540.933418] ice_napi_poll+0x4ca/0x6d0 [ice]
> [ 540.937804] ? __pfx_ice_napi_poll+0x10/0x10 [ice]
> [ 540.942716] napi_busy_loop+0xd7/0x320
> [ 540.946537] xsk_recvmsg+0x143/0x170
> [ 540.950178] sock_recvmsg+0x99/0xa0
> [ 540.953729] __sys_recvfrom+0xa8/0x120
> [ 540.957543] ? do_futex+0xbd/0x1d0
> [ 540.961008] ? __x64_sys_futex+0x73/0x1d0
> [ 540.965083] __x64_sys_recvfrom+0x20/0x30
> [ 540.969155] do_syscall_64+0x38/0x90
> [ 540.972796] entry_SYSCALL_64_after_hwframe+0x72/0xdc
> [ 540.977934] RIP: 0033:0x7f6de5f27934
>
> To fix this, set cached_ntc to first_desc so that at the end, when
> freeing/recycling buffers, descriptors from first to ntc are not missed.
>
> Fixes: 2fba7dc5157b ("ice: Add support for XDP multi-buffer on Rx side")
> Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx>
> ---
> v2: set cached_ntc directly to first_desc [Simon]
Thanks.
Reviewed-by: Simon Horman <simon.horman@xxxxxxxxxxxx>
