Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sat, 1 Apr 2023 22:35:47 +0800
From:   Ming Lei <ming.lei@redhat.com>
To:     Jens Axboe <axboe@kernel.dk>, io-uring@vger.kernel.org,
        linux-block@vger.kernel.org
Cc:     linux-kernel@vger.kernel.org, Miklos Szeredi <mszeredi@redhat.com>,
        ZiyangZhang <ZiyangZhang@linux.alibaba.com>,
        Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>,
        Bernd Schubert <bschubert@ddn.com>,
        Pavel Begunkov <asml.silence@gmail.com>,
        Stefan Hajnoczi <stefanha@redhat.com>,
        Dan Williams <dan.j.williams@intel.com>, ming.lei@redhat.com
Subject: Re: [PATCH V6 03/17] io_uring: add generic IORING_OP_FUSED_CMD
Message-ID: <ZChBQ3EJ/VWFBnF3@ovpn-8-19.pek2.redhat.com>
References: <20230330113630.1388860-1-ming.lei@redhat.com>
 <20230330113630.1388860-4-ming.lei@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230330113630.1388860-4-ming.lei@redhat.com>
Precedence: bulk

On Thu, Mar 30, 2023 at 07:36:16PM +0800, Ming Lei wrote:
> Multiple requests submitted as one whole request logically, and the 1st one
> is primary command(IORING_OP_FUSED_CMD), and the others are secondary
> requests, which number can be retrieved from primary SQE.
> 
> Primary command is responsible for providing resources and submitting
> secondary requests, which depends on primary command's resources, and
> primary command won't be completed until all secondary requests are done.
> 
> The provided resource has same lifetime with primary command, and it
> lifetime won't cross multiple OPs, and this way provides safe way for
> secondary OPs to use the resource.
> 
> Add generic IORING_OP_FUSED_CMD for modeling this primary/secondary
> relationship among requests.

BTW, this model also solves 1:N dependency problem of io_uring.

Current io_uring can't handle this kind of dependency(1:N) efficiently,
and simply convert it into one linked list:

- N requests(1~n) depends on one request(0), and there isn't dependency among
these N requests

- current io_uring converts the dependency to linked list of (N + 1) requests
(0, 1, ...., n), in which all requests are just issued one by one from 0 to n,
inefficiently

The primary/secondary model solves it by issuing request 0 first, then issues
all other N requests concurrently.


Thanks,
Ming