Re: [f2fs-dev] [PATCH v2] f2fs: introduce periodic iostat io latency traces

From: Chao Yu
Date: Thu Aug 12 2021 - 21:14:09 EST


On 2021/8/13 4:52, Jaegeuk Kim wrote:
On 08/11, Chao Yu wrote:
Hi Daeho,

On 2021/8/4 6:55, Daeho Jeong wrote:
From: Daeho Jeong <daehojeong@xxxxxxxxxx>

Whenever we notice some sluggish issues on our machines, we are always
curious about how well all types of I/O in the f2fs filesystem are
handled. But, it's hard to get this kind of real data. First of all,
we need to reproduce the issue while turning on the profiling tool like
blktrace, but the issue doesn't happen again easily. Second, with the
intervention of any tools, the overall timing of the issue will be
slightly changed and it sometimes makes us hard to figure it out.

So, I added F2FS_IOSTAT_IO_LATENCY config option to support printing out
IO latency statistics tracepoint events which are minimal things to
understand filesystem's I/O related behaviors. With "iostat_enable" sysfs
node on, we can get this statistics info in a periodic way and it
would cause the least overhead.

[samples]
f2fs_ckpt-254:1-507 [003] .... 2842.439683: f2fs_iostat_latency:
dev = (254,11), iotype [peak lat.(ms)/avg lat.(ms)/count],
rd_data [136/1/801], rd_node [136/1/1704], rd_meta [4/2/4],
wr_sync_data [164/16/3331], wr_sync_node [152/3/648],
wr_sync_meta [160/2/4243], wr_async_data [24/13/15],
wr_async_node [0/0/0], wr_async_meta [0/0/0]

f2fs_ckpt-254:1-507 [002] .... 2845.450514: f2fs_iostat_latency:
dev = (254,11), iotype [peak lat.(ms)/avg lat.(ms)/count],
rd_data [60/3/456], rd_node [60/3/1258], rd_meta [0/0/1],
wr_sync_data [120/12/2285], wr_sync_node [88/5/428],
wr_sync_meta [52/6/2990], wr_async_data [4/1/3],
wr_async_node [0/0/0], wr_async_meta [0/0/0]

Signed-off-by: Daeho Jeong <daehojeong@xxxxxxxxxx>

---
v2: clean up with wrappers and fix a build breakage reported by
kernel test robot <lkp@xxxxxxxxx>
---
fs/f2fs/Kconfig | 9 +++

I try to apply this patch in my local dev branch, but it failed due to
conflicting with below commit, it needs to rebase this patch to last dev
branch.

I applied this in dev branch. Could you please check?

Yeah, I see.

+config F2FS_IOSTAT_IO_LATENCY
+ bool "F2FS IO statistics IO latency information"
+ depends on F2FS_FS
+ default n
+ help
+ Support printing out periodic IO latency statistics tracepoint
+ events. With this, you have to turn on "iostat_enable" sysfs
+ node to print this out.

This functionality looks independent, how about introuducing iostat.h
and iostat.c (not sure, maybe trace.[hc])to include newly added structure
and functions for dispersive codes cleanup.

Thoughts? this also can avoid using CONFIG_F2FS_IOSTAT_IO_LATENCY in many places.

Thanks,