Re: [PATCH 1/2] f2fs: fix to commit bio cache after flushing node pages

[Jaegeuk Kim]

On Thu, Sep 29, 2016 at 06:45:03PM +0800, Chao Yu wrote:
> On 2016/9/29 4:19, Jaegeuk Kim wrote:
> > On Tue, Sep 27, 2016 at 10:09:03AM +0800, Chao Yu wrote:
> >> On 2016/9/27 9:39, Jaegeuk Kim wrote:
> >>> On Tue, Sep 27, 2016 at 08:57:41AM +0800, Chao Yu wrote:
> >>>> Hi Jaegeuk,
> >>>>
> >>>> On 2016/9/27 2:33, Jaegeuk Kim wrote:
> >>>>> Hi Chao,
> >>>>>
> >>>>> On Tue, Sep 27, 2016 at 12:09:52AM +0800, Chao Yu wrote:
> >>>>>> From: Chao Yu <yuchao0@xxxxxxxxxx>
> >>>>>>
> >>>>>> In sync_node_pages, we won't check and commit last merged pages in private
> >>>>>> bio cache of f2fs, as these pages were taged as writeback, someone who is
> >>>>>> waiting for writebacking of the page will be blocked until the cache was
> >>>>>> committed by someone else.
> >>>>>>
> >>>>>> We need to commit node type bio cache to avoid potential deadlock or long
> >>>>>> delay of waiting writeback.
> >>>>>>
> >>>>>> Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
> >>>>>> ---
> >>>>>>  fs/f2fs/node.c | 11 +++++++++--
> >>>>>>  1 file changed, 9 insertions(+), 2 deletions(-)
> >>>>>>
> >>>>>> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
> >>>>>> index 9faddcd..f73f774 100644
> >>>>>> --- a/fs/f2fs/node.c
> >>>>>> +++ b/fs/f2fs/node.c
> >>>>>> @@ -1416,6 +1416,7 @@ int sync_node_pages(struct f2fs_sb_info *sbi, struct writeback_control *wbc)
> >>>>>>  	struct pagevec pvec;
> >>>>>>  	int step = 0;
> >>>>>>  	int nwritten = 0;
> >>>>>> +	int ret = 0;
> >>>>>>  
> >>>>>>  	pagevec_init(&pvec, 0);
> >>>>>>  
> >>>>>> @@ -1436,7 +1437,8 @@ next_step:
> >>>>>>  
> >>>>>>  			if (unlikely(f2fs_cp_error(sbi))) {
> >>>>>>  				pagevec_release(&pvec);
> >>>>>> -				return -EIO;
> >>>>>> +				ret = -EIO;
> >>>>>> +				goto out;
> >>>>>>  			}
> >>>>>>  
> >>>>>>  			/*
> >>>>>> @@ -1487,6 +1489,8 @@ continue_unlock:
> >>>>>>  
> >>>>>>  			if (NODE_MAPPING(sbi)->a_ops->writepage(page, wbc))
> >>>>>>  				unlock_page(page);
> >>>>>> +			else
> >>>>>> +				nwritten++;
> >>>>>>  
> >>>>>>  			if (--wbc->nr_to_write == 0)
> >>>>>>  				break;
> >>>>>> @@ -1504,7 +1508,10 @@ continue_unlock:
> >>>>>>  		step++;
> >>>>>>  		goto next_step;
> >>>>>>  	}
> >>>>>> -	return nwritten;
> >>>>>> +out:
> >>>>>> +	if (nwritten)
> >>>>>> +		f2fs_submit_merged_bio(sbi, NODE, WRITE);
> >>>>>
> >>>>> IIRC, we don't need to flush this, since f2fs_submit_merged_bio_cond() would
> >>>>> handle this in f2fs_wait_on_page_writeback().
> >>>>
> >>>> Yes, it covers all the cases in f2fs private codes, but there are still some
> >>>> codes in mm or fs directory, and they didn't use f2fs_wait_on_page_writeback
> >>>> when waiting page writeback. Such as do_writepages && filemap_fdatawait in
> >>>> __writeback_single_inode...
> >>>
> >>> The do_writepages() is okay, which will call f2fs_write_node_pages().
> >>> The __writeback_single_inode() won't do filemap_fdatawait() with WB_SYNC_ALL.
> >>> We don't need to take care of truncation as well.
> >>>
> >>> Any missing one?
> >>
> >> Another is: while testing with first version of checkpoint error injection, I
> >> encounter below dump stack:
> >>
> >> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >> mount           D ffff8801c1bf7960     0 97685  97397 0x00080000
> >>  ffff8801c1bf7960 ffff8801c1bf7930 ffff880175900000 ffff8801c1bf7980
> >>  ffff8801c1bf8000 0000000000000000 7fffffffffffffff ffff88021f7be340
> >>  ffffffff817c8880 ffff8801c1bf7978 ffffffff817c80a5 ffff880214f58fc0
> >> Call Trace:
> >>  [<ffffffff817c8880>] ? bit_wait+0x50/0x50
> >>  [<ffffffff817c80a5>] schedule+0x35/0x80
> >>  [<ffffffff817cb152>] schedule_timeout+0x292/0x3d0
> >>  [<ffffffff81022ab5>] ? xen_clocksource_get_cycles+0x15/0x20
> >>  [<ffffffff810eeb5c>] ? ktime_get+0x3c/0xb0
> >>  [<ffffffff817c8880>] ? bit_wait+0x50/0x50
> >>  [<ffffffff817c7906>] io_schedule_timeout+0xa6/0x110
> >>  [<ffffffff817c889b>] bit_wait_io+0x1b/0x60
> >>  [<ffffffff817c84e4>] __wait_on_bit+0x64/0x90
> >>  [<ffffffff8117dcd4>] wait_on_page_bit+0xc4/0xd0
> >>  [<ffffffff810bc4d0>] ? autoremove_wake_function+0x40/0x40
> >>  [<ffffffff81190a29>] truncate_inode_pages_range+0x409/0x840
> >>  [<ffffffff811a406d>] ? pcpu_free_area+0x13d/0x1a0
> >>  [<ffffffff810bc025>] ? wake_up_bit+0x25/0x30
> >>  [<ffffffff81190ecc>] truncate_inode_pages_final+0x4c/0x60
> >>  [<ffffffffa025e9e8>] f2fs_evict_inode+0x48/0x390 [f2fs]
> >>  [<ffffffff812212f7>] evict+0xc7/0x1a0
> >>  [<ffffffff81221f77>] iput+0x197/0x200
> >>  [<ffffffffa0268242>] f2fs_fill_super+0xab2/0x1130 [f2fs]
> >>  [<ffffffff81209454>] mount_bdev+0x184/0x1c0
> >>  [<ffffffffa0267790>] ? f2fs_commit_super+0x100/0x100 [f2fs]
> >>  [<ffffffffa02646a5>] f2fs_mount+0x15/0x20 [f2fs]
> >>  [<ffffffff81209e19>] mount_fs+0x39/0x160
> >>  [<ffffffff81225e47>] vfs_kern_mount+0x67/0x110
> >>  [<ffffffff812283bb>] do_mount+0x1bb/0xc80
> >>  [<ffffffff81229163>] SyS_mount+0x83/0xd0
> >>  [<ffffffff8100391e>] do_syscall_64+0x6e/0x170
> >>  [<ffffffff817cc325>] entry_SYSCALL64_slow_path+0x25/0x25
> >>
> >> Any thoughts?
> > 
> > I think this should not happen normally, since f2fs_stop_checkpoint() calls
> > f2fs_flush_merged_bios().
> 
> In write_end_io, f2fs_stop_checkpoint will not call f2fs_flush_merged_bios.

So, in write_node_page(), we call f2fs_submit_merged_bio() if f2fs_cp_error()
is set.

> One other problem here is it can cause latency during waiting writeback:
> 
> In fsync()
> 1. fsync_node_pages a/b/c pages is submitted, and d/e/f pages is still in bio
> cache as we didn't commit the bio in the end.
> 2. wait_on_node_pages_writeback will wait a/b/c pages for writebacking, then
> submit bio with d/e/f pages and the wait until they have writebacked to device.
> 
> Here we need to submit cached bios at the end of
> fsync_node_pages/sync_node_pages to let them being merged in block layer as much
> as possible, and also to avoid suffering more delay time due to double submit&wait.

I think this is more reasonable to me. ;)

Thanks,