[Devel,vz7] fs: avoid writeback busy-loop if redirty

Submitted by Maxim Patlasov on Dec. 14, 2016, 4:09 a.m.

Details

Message ID 148168840532.21778.15860368316917000865.stgit@maxim-thinkpad
State New
Series "fs: avoid writeback busy-loop if redirty"
Headers show

Commit Message

Maxim Patlasov Dec. 14, 2016, 4:09 a.m.
Andrey, can you please review this one-liner below. I think it can
be responsible for inexplicable delays and CPU peaks in bugs like
https://jira.sw.ru/browse/PSBM-55919. Thanks, Maxim

wb_writeback() bails out only if either nr_pages comes to zero or
"progress" (as reported by writeback_sb_inodes or __writeback_inodes_wb)
is zero.

If a page is temporarily unavailable (is under fuse-writeback),
fuse_writepages_fill calls redirty_page_for_writepage, unlock the page
and returns OK. This OK results in progress=1. So wb_writeback
loops over writeback_sb_inodes/__writeback_inodes_wb again and again
until all nr_pages are exhausted.

The patch ensures that progress=0 if fuse_writepages_fill did nothing.

Signed-off-by: Maxim Patlasov <mpatlasov@virtuozzo.com>
---
 fs/fs-writeback.c |    3 +++
 1 file changed, 3 insertions(+)

Patch hide | download patch | download mbox

diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index 5c9a082..483bbb9 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -774,6 +774,9 @@  static long writeback_sb_inodes(struct super_block *sb,
 			if (work->nr_pages <= 0)
 				break;
 		}
+
+		BUG_ON(wbc.pages_skipped > write_chunk - wbc.nr_to_write);
+		wrote -= wbc.pages_skipped;
 	}
 	return wrote;
 }

Comments

Vasily Averin Dec. 14, 2016, 9:25 a.m.
Maxim,
did you asked Alexey probably?
Also can you replace BUG_ON on WARN_ON ? 
IMHO it's bad idea to crash kernel in any such cases. 
If it can be triggered somehow from inside container -- it will be pure security issue.

Thank you,
	Vasily Averin

On 12/14/2016 07:09 AM, Maxim Patlasov wrote:
> Andrey, can you please review this one-liner below. I think it can
> be responsible for inexplicable delays and CPU peaks in bugs like
> https://jira.sw.ru/browse/PSBM-55919. Thanks, Maxim
> 
> wb_writeback() bails out only if either nr_pages comes to zero or
> "progress" (as reported by writeback_sb_inodes or __writeback_inodes_wb)
> is zero.
> 
> If a page is temporarily unavailable (is under fuse-writeback),
> fuse_writepages_fill calls redirty_page_for_writepage, unlock the page
> and returns OK. This OK results in progress=1. So wb_writeback
> loops over writeback_sb_inodes/__writeback_inodes_wb again and again
> until all nr_pages are exhausted.
> 
> The patch ensures that progress=0 if fuse_writepages_fill did nothing.
> 
> Signed-off-by: Maxim Patlasov <mpatlasov@virtuozzo.com>
> ---
>  fs/fs-writeback.c |    3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index 5c9a082..483bbb9 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -774,6 +774,9 @@ static long writeback_sb_inodes(struct super_block *sb,
>  			if (work->nr_pages <= 0)
>  				break;
>  		}
> +
> +		BUG_ON(wbc.pages_skipped > write_chunk - wbc.nr_to_write);
> +		wrote -= wbc.pages_skipped;
>  	}
>  	return wrote;
>  }
> 
>
Maxim Patlasov Dec. 14, 2016, 6:47 p.m.
Thank you, Vasily, I'll resend the patch.


On 12/14/2016 01:25 AM, Vasily Averin wrote:
> Maxim,
> did you asked Alexey probably?
> Also can you replace BUG_ON on WARN_ON ?
> IMHO it's bad idea to crash kernel in any such cases.
> If it can be triggered somehow from inside container -- it will be pure security issue.
>
> Thank you,
> 	Vasily Averin
>
> On 12/14/2016 07:09 AM, Maxim Patlasov wrote:
>> Andrey, can you please review this one-liner below. I think it can
>> be responsible for inexplicable delays and CPU peaks in bugs like
>> https://jira.sw.ru/browse/PSBM-55919. Thanks, Maxim
>>
>> wb_writeback() bails out only if either nr_pages comes to zero or
>> "progress" (as reported by writeback_sb_inodes or __writeback_inodes_wb)
>> is zero.
>>
>> If a page is temporarily unavailable (is under fuse-writeback),
>> fuse_writepages_fill calls redirty_page_for_writepage, unlock the page
>> and returns OK. This OK results in progress=1. So wb_writeback
>> loops over writeback_sb_inodes/__writeback_inodes_wb again and again
>> until all nr_pages are exhausted.
>>
>> The patch ensures that progress=0 if fuse_writepages_fill did nothing.
>>
>> Signed-off-by: Maxim Patlasov <mpatlasov@virtuozzo.com>
>> ---
>>   fs/fs-writeback.c |    3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
>> index 5c9a082..483bbb9 100644
>> --- a/fs/fs-writeback.c
>> +++ b/fs/fs-writeback.c
>> @@ -774,6 +774,9 @@ static long writeback_sb_inodes(struct super_block *sb,
>>   			if (work->nr_pages <= 0)
>>   				break;
>>   		}
>> +
>> +		BUG_ON(wbc.pages_skipped > write_chunk - wbc.nr_to_write);
>> +		wrote -= wbc.pages_skipped;
>>   	}
>>   	return wrote;
>>   }
>>
>>