netdev - net DIM (Dynamic interrupt moderation) queries

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>] [day] [month] [year] [list]

Date:   Tue, 10 Nov 2020 12:33:40 +0000
From:   Manish Chopra <manishc@...vell.com>
To:     "talgi@...lanox.com" <talgi@...lanox.com>
CC:     "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Igor Russkikh <irusskikh@...vell.com>
Subject: net DIM (Dynamic interrupt moderation) queries

Hello Tal,

I looked through the documentation but some of the things were not cleared, so I am looking for some fine details on this topic if you could help please -

Q1- 	In general, can this feature be problematic or not useful if device only support time based moderation but does not support packets count based moderation ?
	IOW, does DIM strictly demands for devices to support both time based and packets count based moderation ?

Q2- 	What is DIM mode actually (DIM_CQ_PERIOD_MODE_START_FROM_EQE and DIM_CQ_PERIOD_MODE_START_FROM_CQE) ?
	Document mentions that one of them is used for timer reset, but unsure what exactly it is ? By looking at the code, it seems it just differs from profile values.
	How do I get to know that which one should be used or could be benefitted in general by the drivers and in what circumstances ?

Q3- 	I started with using DIM_CQ_PERIOD_MODE_START_FROM_EQE with our drivers (drivers/net/ethernet/qlogic/qede) and our device only supports
	time based moderation but does not support packets based moderation, our driver simply updating data-path interrupts as an event counts, rcv_bytes and
	rcv_pkts through the DIM api in NAPI poll routine, I was experiencing some strange behaviors - (unsure if they are expected)

	a. When simply running constant PING flood from peer (ping <ip> -f), I was seeing that DIM causing the driver to change the
	    usec moderation values (1,8,..) quite frequently, it did not make stable moderation configuration although it was steady PING flood,
	    same was observed with single TCP connection too (with TCP it wasn't that frequent but it kept on changing over the time, still not stable configuration on const TCP stream)

	b. With TCP, I could see that DIM causing driver to use higher usec values (compared to PING flood), but strangely when stopping TCP stream and running PING
	    back, it did not reduce the moderation values back (I would have expected that PING after TCP would reduce the moderation value)

# uname -r
5.5.0-rc2net-next

Regards,
Manish