[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <72658bca-c2b2-b3cb-64a0-35540b247a11@intel.com>
Date: Mon, 3 Jul 2023 15:50:55 +0200
From: Alexander Lobakin <aleksander.lobakin@...el.com>
To: Jakub Kicinski <kuba@...nel.org>
CC: "David S. Miller" <davem@...emloft.net>,
Eric Dumazet <edumazet@...gle.com>,
Paolo Abeni <pabeni@...hat.com>,
Maciej Fijalkowski <maciej.fijalkowski@...el.com>,
Larysa Zaremba <larysa.zaremba@...el.com>,
Yunsheng Lin <linyunsheng@...wei.com>,
Alexander Duyck <alexanderduyck@...com>,
Jesper Dangaard Brouer <hawk@...nel.org>,
"Ilias Apalodimas" <ilias.apalodimas@...aro.org>,
<netdev@...r.kernel.org>, <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH RFC net-next 0/4] net: page_pool: a couple assorted
optimizations
From: Jakub Kicinski <kuba@...nel.org>
Date: Sat, 1 Jul 2023 17:01:55 -0700
> On Thu, 29 Jun 2023 17:23:01 +0200 Alexander Lobakin wrote:
>> #3: new, prereq to #4. Add NAPI state flag, which would indicate
>> napi->poll() is running right now, so that napi->list_owner would
>> point to the CPU where it's being run, not just scheduled;
>> #4: new. In addition to recycling skb PP pages directly when @napi_safe
>> is set, check for the flag from #3, which will mean the same if
>> ->list_owner is pointing to us. This allows to use direct recycling
>> anytime we're inside a NAPI polling loop or GRO stuff going right
>> after it, covering way more cases than is right now.
>
> You know NAPI pretty well so I'm worried I'm missing something.
I wouldn't say I know it well :D
> I don't think the new flag adds any value. NAPI does not have to
> be running, you can drop patch 3 and use in_softirq() instead of
> the new flag, AFAIU.
That's most likely true for the patch 4 case, but I wanted to add some
flag for wider usage.
For example, busy polling relies on whether ->poll() returned whole
budget to decide whether interrupts were reenabled to avoid possible
concurrent access, but I wouldn't say it's precise enough.
napi_complete_done() doesn't always return true.
OTOH, the new flag or, more precisely, flag + list_owner combo would
tell for sure.
>
> The reason I did not do that is that I wasn't sure if there is no
> weird (netcons?) case where skb gets freed from an IRQ :(
Shouldn't they use dev_kfree_skb_any() or _irq()? Usage of plain
kfree_skb() is not allowed in the TH :s
Anyway, if the flag really makes no sense, I can replace it with
in_softirq(), it's my hobby to break weird drivers :D
Thanks,
Olek
Powered by blists - more mailing lists