On 2021/2/18 下午11:28, Nikolay Borisov wrote:

On 8.01.21 г. 7:36 ч., Qu Wenruo wrote:

quoted

There are several qgroup flush related bugs fixed recently, all of them
are caused by the fact that we can trigger qgroup metadata space
reservation holding a transaction handle.

Thankfully the only situation to trigger above reservation is
btrfs_dirty_inode().

Currently btrfs_dirty_inode() will try join transactio first, then
update the inode.
If btrfs_update_inode() fails with -ENOSPC, then it retry to start
transaction to reserve metadata space.

This not only forces us to reserve metadata space with a transaction
handle hold, but can't handle other errors like -EDQUOT.

This patch will make btrfs_dirty_inode() to call
btrfs_start_transaction() directly without first try joining then
starting, so that in try_flush_qgroup() we won't hold a trans handle.

This will slow down btrfs_dirty_inode() but my fstests doesn't show too
much different for most test cases, thus it may be worthy to skip such
performance "optimization".

Signed-off-by: Qu Wenruo <redacted>

Ok I actually run 2 tests against this patch. The first one is a 10
second run of  stress-ng's utime test (stress-ng --temp-path
/media/scratch --utime 4 -M -t 10 ; done) to see if I can reproduce
intel's results and here's what I found:


bogo ops/s real (Before-patch)	bogo ops/s real (After Patch)
	35993	                         32968
	35712	                         33146
	35369	                         32996
	35544	                         33159
	35623	                         33000
	35939	                         33016
	35693	                         32829
	35562	                         32685
	35675	                         32815
Std dev	182.161981912585	146.829034703967
HMean	35677.9600871036	32957.1111111111
Diff%:		                -7.626

So there's a 7.6% decrease in the rate of utime() calls we can make,
given that we now start a transaction I'd say that's expected.

The other test was a randwrite with fio as I was mostly worried that
making btrfs_dirty_inode more expensive would hit write performance
since file_update_times is called from the generic iter. But inspecting
the code btrfs uses update_time_for_write which doesn't dirty the inode
per-se as this is deferred to endio completion time.  I also measured
the impact during buffered read time as file_accessed is called a lot of
times but the following bpftrace script:

BEGIN {@execs = 0; }
kprobe:btrfs_dirty_inode
{
	@test[kstack] = count();
	@execs++;
}

kprobe:touch_atime
{
	@test[kstack] = count();
}
END{
	printf("Total btrfs_dirty_inode calls: %llu\n", @execs);
}


confirmed we only ever execute around 8 btrfs_dirty_inode out of 1048773
execution of touch_atimes from generic_file_buffered_read with the
following fio workload:

fio --name=random-readers --thread --ioengine=sync --iodepth=4
--rw=randread --bs=4k --direct=0 --size=1g --numjobs=4
--directory=/media/scratch --filename_format=FioWorkloads.\$jobnum
--new_group --group_reporting=1


So performance-wise I'm inclined to give it a "pass".

Great. Mind me to add such info into the commit message and add you as sob?

Thanks,
Qu