Re: [PATCH] ublk_drv: don't call task_work_add for queueing io commands

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hello Ziyang,

On Mon, Oct 24, 2022 at 05:48:51PM +0800, Ziyang Zhang wrote:
> On 2022/10/23 17:38, Ming Lei wrote:
> > task_work_add() is used for waking ubq daemon task with one batch
> > of io requests/commands queued. However, task_work_add() isn't
> > exported for module code, and it is still debatable if the symbol
> > should be exported.
> > 
> > Fortunately we still have io_uring_cmd_complete_in_task() which just
> > can't handle batched wakeup for us.
> > 
> > Add one one llist into ublk_queue and call io_uring_cmd_complete_in_task()
> > via current command for running them via task work.
> > 
> > This way cleans up current code a lot, meantime allow us to wakeup
> > ubq daemon task after queueing batched requests/io commands.
> > 
> 
> 
> Hi, Ming
> 
> This patch works and I have run some tests to compare current version(ucmd)
> with your patch(ucmd-batch).
> 
> iodepth=128 numjobs=1 direct=1 bs=4k
> 
> --------------------------------------------
> ublk loop target, the backend is a file.
> IOPS(k)
> 
> type		ucmd		ucmd-batch
> seq-read	54.7		54.2	
> rand-read	52.8		52.0
> 
> --------------------------------------------
> ublk null target
> IOPS(k)
> 
> type		ucmd		ucmd-batch
> seq-read	257		257
> rand-read	252		253
> 
> 
> I find that io_req_task_work_add() puts task_work node into a llist
> first, then it may call task_work_add() to run batched task_works. So do we really
> need such llist in ublk_drv? I think io_uring has already considered task_work batch
> optimization.
> 
> BTW, task_work_add() in ublk_drv achieves
> higher IOPS(about 5-10% on my machine) than io_uring_cmd_complete_in_task()
> in ublk_drv.

Yeah, that is same with my observation, and motivation of this patch is
to get same performance with task_work_add by building ublk_drv as
module. One win of task_work_add() is that we get exact batching info
meantime only send TWA_SIGNAL_NO_IPI for whole batch, that is basically
what the patch is doing, but needs help of the following ublksrv patch:

https://github.com/ming1/ubdsrv/commit/dce6d1d222023c1641292713b311ced01e6dc548

which sets IORING_SETUP_COOP_TASKRUN for ublksrv's uring, then
io_uring_cmd_complete_in_task will notify via TWA_SIGNAL_NO_IPI, and 5+%
IOPS boost is observed on loop/001 by putting image on SSD in my test
VM.

Thanks, 
Ming




[Index of Archives]     [Linux RAID]     [Linux SCSI]     [Linux ATA RAID]     [IDE]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Device Mapper]

  Powered by Linux