netdev - Re: [PATCH net-next v2 13/14] ixgbe: preserve RSS indirection table across admin down/up

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20251022034051.28052-2-enjuk@amazon.com>
Date: Wed, 22 Oct 2025 12:40:45 +0900
From: Kohei Enju <enjuk@...zon.com>
To: <kuba@...nel.org>
CC: <aleksander.lobakin@...el.com>, <andrew+netdev@...n.ch>,
	<anthony.l.nguyen@...el.com>, <corbet@....net>, <davem@...emloft.net>,
	<edumazet@...gle.com>, <enjuk@...zon.com>, <horms@...nel.org>,
	<jacob.e.keller@...el.com>, <jiri@...nulli.us>, <linux-doc@...r.kernel.org>,
	<linux-kernel@...r.kernel.org>, <netdev@...r.kernel.org>,
	<pabeni@...hat.com>, <przemyslaw.kitszel@...el.com>, <sx.rinitha@...el.com>
Subject: Re: [PATCH net-next v2 13/14] ixgbe: preserve RSS indirection table across admin down/up

On Tue, 21 Oct 2025 16:10:06 -0700, Jakub Kicinski wrote:

>On Tue, 21 Oct 2025 12:59:34 +0900 Kohei Enju wrote:
>> For example, consider a scenario where the queue count is 8 with user
>> configuration containing values from 0 to 7. When queue count changes
>> from 8 to 4 and we skip the reinitialization in this scenario, entries
>> pointing to queues 4-7 become invalid. The same issue applies when the
>> RETA table size changes.
>
>Core should reject this. See ethtool_check_max_channel()

Indeed, you're right that the situation above will be rejected. I missed
it.

BTW, I think reinitializing the RETA table when queue count changes or
RETA table size changes is reasonable for predictability and safety.
Does this approach make sense to you?

>
>> Furthermore, IIUC, adding netif_is_rxfh_configured() to the current
>> condition wouldn't provide additional benefit. When parameters remain
>> unchanged, regardless of netif_is_rxfh_configured(), we already preserve
>> the RETA entries which might be user-configured or default values, 
>
>User may decide to "isolate" (take out of RSS) a lower queue,
>to configure it for AF_XDP or other form of zero-copy. Install
>explicit rules to direct traffic to that queue. If you reset
>the RSS table random traffic will get stranded in the ZC queue
>(== dropped).
>

You're correct about the ZC queue scenario. The original implementation
(before this patch) would indeed cause this problem by unconditionally
reinitializing.

I believe this patch addresses that issue - it preserves the user
configuration since neither queue count nor RETA table size changes in
that case. If I'm misunderstanding your scenario, please let me know.

I could update the logic to explicitly check netif_is_rxfh_configured()
as in [1], though the actual behavior would be the same as [2] since
the default RETA table is a deterministic function of (rss_indices,
reta_entries):

[1] Check user configuration explicitly:
    if (!netif_is_rxfh_configured(adapter->netdev) ||
        adapter->last_rss_indices != rss_i ||
        adapter->last_reta_entries != reta_entries) {
        // reinitialize
    }

[2] Current patch:
    if (adapter->last_rss_indices != rss_i ||
        adapter->last_reta_entries != reta_entries) {
        // reinitialize
    }

Do you have any preference between these approaches, or would you
recommend a different solution?