[PATCHSET,0/3] Improve IRQ driven performance

Message ID	20211215163009.15269-1-axboe@kernel.dk (mailing list archive)
Headers	show Return-Path: <linux-block-owner@kernel.org> From: Jens Axboe <axboe@kernel.dk> To: io-uring@vger.kernel.org, linux-block@vger.kernel.org Subject: [PATCHSET 0/3] Improve IRQ driven performance Date: Wed, 15 Dec 2021 09:30:06 -0700 Message-Id: <20211215163009.15269-1-axboe@kernel.dk> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Precedence: bulk
Series	Improve IRQ driven performance \| expand [PATCHSET,0/3] Improve IRQ driven performance [1/3] block: add completion handler for fast path [2/3] block: use singly linked list for bio cache [3/3] block: enable bio allocation cache for IRQ driven IO

Message ID

20211215163009.15269-1-axboe@kernel.dk (mailing list archive)

Headers

From: Jens Axboe <axboe@kernel.dk>
To: io-uring@vger.kernel.org, linux-block@vger.kernel.org
Subject: [PATCHSET 0/3] Improve IRQ driven performance
Date: Wed, 15 Dec 2021 09:30:06 -0700
Message-Id: <20211215163009.15269-1-axboe@kernel.dk>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Series

Improve IRQ driven performance | expand

Message

Jens Axboe Dec. 15, 2021, 4:30 p.m. UTC

Hi,

1/2 are really just optimizations, but were done with the support for
using cached bio allocations with IRQ driven IO. The bio recycling is an
even bigger win on IRQ driven IO than it was on polling, around a 13%
improvement for me.

A caller that is prepared to get a bio passed back at completion time may
set IOCB_BIO_PASSBACK in the iocb, and then the completion side may notice
this and assign iocb->private and set IOCB_PRIV_IS_BIO and have
->ki_complete() handle the freeing. This works for io_uring as IRQ
completions are processed in task context.