linux-kernel - Re: [PATCH] fix readahead pipeline break caused by block plug

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>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20120201070247.GA29083@localhost>
Date:	Wed, 1 Feb 2012 15:02:47 +0800
From:	Wu Fengguang <wfg@...ux.intel.com>
To:	Vivek Goyal <vgoyal@...hat.com>
Cc:	Shaohua Li <shaohua.li@...el.com>,
	lkml <linux-kernel@...r.kernel.org>,
	linux-mm <linux-mm@...ck.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Jens Axboe <axboe@...nel.dk>,
	Herbert Poetzl <herbert@...hfloor.at>,
	Eric Dumazet <eric.dumazet@...il.com>
Subject: Re: [PATCH] fix readahead pipeline break caused by block plug

On Tue, Jan 31, 2012 at 05:03:33PM -0500, Vivek Goyal wrote:
> On Tue, Jan 31, 2012 at 03:59:40PM +0800, Shaohua Li wrote:
> > Herbert Poetzl reported a performance regression since 2.6.39. The test
> > is a simple dd read, but with big block size. The reason is:
> > 
> > T1: ra (A, A+128k), (A+128k, A+256k)
> > T2: lock_page for page A, submit the 256k
> > T3: hit page A+128K, ra (A+256k, A+384). the range isn't submitted
> > because of plug and there isn't any lock_page till we hit page A+256k
> > because all pages from A to A+256k is in memory
> > T4: hit page A+256k, ra (A+384, A+ 512). Because of plug, the range isn't
> > submitted again.
> > T5: lock_page A+256k, so (A+256k, A+512k) will be submitted. The task is
> > waitting for (A+256k, A+512k) finish.
> > 
> > There is no request to disk in T3 and T4, so readahead pipeline breaks.
> > 
> > We really don't need block plug for generic_file_aio_read() for buffered
> > I/O. The readahead already has plug and has fine grained control when I/O
> > should be submitted. Deleting plug for buffered I/O fixes the regression.
> > 
> > One side effect is plug makes the request size 256k, the size is 128k
> > without it. This is because default ra size is 128k and not a reason we
> > need plug here.
> 
> For me, this patch helps only so much and does not get back all the
> performance lost in case of raw disk read. It does improve the throughput
> from around 85-90 MB/s to 110-120 MB/s but running the same dd with
> iflag=direct, gets me more than 250MB/s.
> 
> # echo 3 > /proc/sys/vm/drop_caches 
> # dd if=/dev/sdb of=/dev/null bs=1M count=1K
> 1024+0 records in
> 1024+0 records out
> 1073741824 bytes (1.1 GB) copied, 9.03305 s, 119 MB/s
> 
> echo 3 > /proc/sys/vm/drop_caches 
> # dd if=/dev/sdb of=/dev/null bs=1M count=1K iflag=direct
> 1024+0 records in
> 1024+0 records out
> 1073741824 bytes (1.1 GB) copied, 4.07426 s, 264 MB/s
> 
> I think it is happening because in case of raw read we are submitting
> one page at a time to request queue and by the time all the pages
> are submitted and one big merged request is formed it wates lot of time.
> 
> In case of direct IO, we are getting bigger IOs at request queue so
> less cpu overhead, less idling on queue.

Note that "dd bs=1M" will result in 128KB readahead IO. The buffered
dd reads may perform much better if 1MB readahead size is used:

blockdev --setra 2048 /dev/sda

> I created ext4 filesystem on same SSD and did the buffered read and
> that seems to work just fine. Now I am getting bigger requests at
> the request queue. (128K, 256 sectors).
> 
> [root@...lli common]# echo 3 > /proc/sys/vm/drop_caches 
> [root@...lli common]# dd if=zerofile-4G of=/dev/null bs=1M count=1K
> 1024+0 records in
> 1024+0 records out
> 1073741824 bytes (1.1 GB) copied, 4.09186 s, 262 MB/s

So the raw sda reads have some performance problems. What's the exact
blktrace sequence for sda reads? And the block size?

blockdev --getbsz /dev/sda               

> Anyway, remvoing top level plug in case of buffered reads sounds
> reasonable.

Yup.

Thanks,
Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/