Re: [PATCH RFC net] linux/dim: Fix divide 0 in RDMA DIM.

[Tao Liu <thomas.liu@ucloud.cn>]

On Mon, Jun 27, 2022 at 10:42:02AM +0300, Leon Romanovsky wrote:
> On Thu, Jun 23, 2022 at 04:58:58PM +0800, Tao Liu wrote:
> > We hit a divide 0 error in ofed 5.1.2.3.7.1. But dim.c and
> > rdma_dim.c seem same as upstream.
> > 
> > CallTrace:
> >   Hardware name: H3C R4900 G3/RS33M2C9S, BIOS 2.00.37P21 03/12/2020
> >   task: ffff880194b78000 task.stack: ffffc90006714000
> >   RIP: 0010:backport_rdma_dim+0x10e/0x240 [mlx_compat]
> >   RSP: 0018:ffff880c10e83ec0 EFLAGS: 00010202
> >   RAX: 0000000000002710 RBX: ffff88096cd7f780 RCX: 0000000000000064
> >   RDX: 0000000000000000 RSI: 0000000000000002 RDI: 0000000000000001
> >   RBP: 0000000000000001 R08: 0000000000000000 R09: 0000000000000000
> >   R10: 0000000000000000 R11: 0000000000000000 R12: 000000001d7c6c09
> >   R13: ffff88096cd7f780 R14: ffff880b174fe800 R15: 0000000000000000
> >   FS:  0000000000000000(0000) GS:ffff880c10e80000(0000)
> >   knlGS:0000000000000000
> >   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >   CR2: 00000000a0965b00 CR3: 000000000200a003 CR4: 00000000007606e0
> >   DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> >   DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> >   PKRU: 55555554
> >   Call Trace:
> >    <IRQ>
> >    ib_poll_handler+0x43/0x80 [ib_core]
> >    irq_poll_softirq+0xae/0x110
> >    __do_softirq+0xd1/0x28c
> >    irq_exit+0xde/0xf0
> >    do_IRQ+0x54/0xe0
> >    common_interrupt+0x8f/0x8f
> >    </IRQ>
> >    ? cpuidle_enter_state+0xd9/0x2a0
> >    ? cpuidle_enter_state+0xc7/0x2a0
> >    ? do_idle+0x170/0x1d0
> >    ? cpu_startup_entry+0x6f/0x80
> >    ? start_secondary+0x1b9/0x210
> >    ? secondary_startup_64+0xa5/0xb0
> >   Code: 0f 87 e1 00 00 00 8b 4c 24 14 44 8b 43 14 89 c8 4d 63 c8 44 29 c0 99 31 d0 29 d0 31 d2 48 98 48 8d 04 80 48 8d 04 80 48 c1 e0 02 <49> f7 f1 48 83 f8 0a 0f 86 c1 00 00 00 44 39 c1 7f 10 48 89 df
> >   RIP: backport_rdma_dim+0x10e/0x240 [mlx_compat] RSP: ffff880c10e83ec0
> > 
> > crash> struct dim ffff88096cd7f780
> > struct dim {
> >   state = 1 '\001',
> >   prev_stats = {
> >       ppms = 2142150514,
> >       bpms = 391112768,
> >       epms = -30709,
> >       cpms = 1,
> >       cpe_ratio = 0
> >     },
> >   start_sample = {
> >       time = 51174507127193614,
> >       pkt_ctr = 0,
> >       byte_ctr = 0,
> >       event_ctr = 1,
> >       comp_ctr = 494693321
> >     },
> >   measuring_sample = {
> >       time = 51174507266581985,
> >       pkt_ctr = 0,
> >       byte_ctr = 0,
> >       event_ctr = 65,
> >       comp_ctr = 494693385
> >     },
> >   work = {
> >       data = {
> >             counter = 128
> >           },
> >       entry = {
> >             next = 0xffff88096cd7f7d0,
> >             prev = 0xffff88096cd7f7d0
> >           },
> >       func = 0xffffffffa02b9f80
> >     },
> >   priv = 0xffff880b174fe800,
> >   profile_ix = 1 '\001',
> >   mode = 0 '\000',
> >   tune_state = 2 '\002',
> >   steps_right = 1 '\001',
> >   steps_left = 1 '\001',
> >   tired = 0 '\000'
> > }
> > 
> > Fixes: f4915455dcf0 ("linux/dim: Implement RDMA adaptive moderation (DIM)")
> > Signed-off-by: Tao Liu <thomas.liu@ucloud.cn>
> > ---
> >  lib/dim/rdma_dim.c | 3 +++
> >  1 file changed, 3 insertions(+)
> 
> I think that this change will be better as it won't change
> decision order in rdma_dim_stats_compare()
> 
> diff --git a/include/linux/dim.h b/include/linux/dim.h
> index b698266d0035..69ae238ec2dc 100644
> --- a/include/linux/dim.h
> +++ b/include/linux/dim.h
> @@ -21,7 +21,7 @@
>   * We consider 10% difference as significant.
>   */
>  #define IS_SIGNIFICANT_DIFF(val, ref) \
> -       (((100UL * abs((val) - (ref))) / (ref)) > 10)
> +       (ref && (((100UL * abs((val) - (ref))) / (ref)) > 10))
>  
>  /*
>   * Calculate the gap between two values.
> 
> 
Reviewed code in net_dim_stats_compare() and rdma_dim_stats_compare(), the
crash point is the only place not covered 0 condition. So it maybe not
need to change the macro.

But I am not familiar with the algorithm, and not sure what is the right
return value.
> > 
> > diff --git a/lib/dim/rdma_dim.c b/lib/dim/rdma_dim.c
> > index 15462d54758d..a657b106343c 100644
> > --- a/lib/dim/rdma_dim.c
> > +++ b/lib/dim/rdma_dim.c
> > @@ -34,6 +34,9 @@ static int rdma_dim_stats_compare(struct dim_stats *curr,
> >  		return (curr->cpms > prev->cpms) ? DIM_STATS_BETTER :
> >  						DIM_STATS_WORSE;
> >  
> > +	if (!prev->cpe_ratio)
> > +		return DIM_STATS_SAME;
> > +
> >  	if (IS_SIGNIFICANT_DIFF(curr->cpe_ratio, prev->cpe_ratio))
> >  		return (curr->cpe_ratio > prev->cpe_ratio) ? DIM_STATS_BETTER :
> >  						DIM_STATS_WORSE;
> > -- 
> > 2.30.1 (Apple Git-130)
> > 
>