[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20250906032227.70729-1-enjuk@amazon.com>
Date: Sat, 6 Sep 2025 12:22:24 +0900
From: Kohei Enju <enjuk@...zon.com>
To: <aleksandr.loktionov@...el.com>
CC: <andrew+netdev@...n.ch>, <anthony.l.nguyen@...el.com>,
<davem@...emloft.net>, <edumazet@...gle.com>, <enjuk@...zon.com>,
<intel-wired-lan@...ts.osuosl.org>, <kohei.enju@...il.com>,
<kuba@...nel.org>, <netdev@...r.kernel.org>, <pabeni@...hat.com>,
<przemyslaw.kitszel@...el.com>
Subject: Re: [Intel-wired-lan] [PATCH iwl-next v3] ixgbe: preserve RSS indirection table across admin down/up
On Wed, 3 Sep 2025 06:21:05 +0000, Loktionov, Aleksandr wrote:
> [...]
>> On Wed, 3 Sep 2025 06:04:43 +0900, Kohei Enju wrote:
>>
>> >On Tue, 2 Sep 2025 13:25:56 +0000, Loktionov, Aleksandr wrote:
>> >
>> >> [...]
>> >>> -
>> >>> - for (i = 0, j = 0; i < reta_entries; i++, j++) {
>> >>> - if (j == rss_i)
>> >>> - j = 0;
>> >>> + /* Update redirection table in memory on first init, queue
>> >>> count change,
>> >>> + * or reta entries change, otherwise preserve user
>> >>> configurations. Then
>> >>> + * always write to hardware.
>> >>> + */
>> >>> + if (adapter->last_rss_indices != rss_i ||
>> >>> + adapter->last_reta_entries != reta_entries) {
>> >>> + for (i = 0; i < reta_entries; i++)
>> >>> + adapter->rss_indir_tbl[i] = i % rss_i;
>> >>Are you sure rss_i never ever can be a 0?
>> >>This is the only thing I'm worrying about.
>> >
>> >Oops, you're exactly right. Good catch!
>> >
>> >I see the original code assigns 0 to rss_indir_tbl[i] when rss_i is
>> 0,
>> >like:
>> > adapter->rss_indir_tbl[i] = 0;
>>
>> Ahh, that's not true, my brain was not working... Sorry for messing
>> up.
>> Anyway, in a situation where rss_i == 0, we should handle it somehow
>> to avoid zero-divisor.
>>
>> >
>> >To handle this with keeping the behavior when rss_i == 0, I'm
>> >considering Option 1:
>> > adapter->rss_indir_tbl[i] = rss_i ? i % rss_i : 0;
>> >
>> >Option 2:
>> > if (rss_i)
>> > for (i = 0; i < reta_entries; i++)
>> > adapter->rss_indir_tbl[i] = i % rss_i;
>> > else
>> > memset(adapter->rss_indir_tbl, 0, reta_entries);
>> >
>> >Since this is not in the data path, the overhead of checking rss_i in
>> >each iteration might be acceptable. Therefore I'd like to adopt the
>> >option 1 for simplicity.
>> >
>> >Do you have any preference or other suggestions?
>
>I lean toward option 2, as the explicit if (rss_i) guard makes the logic clearer and easier to follow.
>
>Handling the simplified case first with:
>if (unlikely(!rss_i))
> memset(adapter->rss_indir_tbl, 0, reta_entries);
>else
> for (i = 0; i < reta_entries; i++)
> adapter->rss_indir_tbl[i] = i % rss_i;
>
>Improves readability and separates the edge case from the main logic.
>
>While it's possible to use a ternary expression like adapter->rss_indir_tbl[i] = rss_i ? i % rss_i : 0;,
>I find the conditional block more maintainable, especially if this logic evolves later.
Okay, I got it.
>
>Regarding unlikely(), unless there's profiling data showing a performance benefit,
>I'd avoid it here - this isn't in the fast path, and clarity should take precedence.
Yes, I agree on that this isn't in the fast path therefore unlikely is
not always necessary.
Thank you again for reviewing and suggesting!
>With the best regards Alex
Powered by blists - more mailing lists