Re: [f2fs-dev] [PATCH 2/2] f2fs: avoid infinite GC loop due to stale atomic files

[Jaegeuk Kim]

On 09/09, Jaegeuk Kim wrote:
> On 09/09, Chao Yu wrote:
> > On 2019/9/9 16:21, Jaegeuk Kim wrote:
> > > On 09/09, Chao Yu wrote:
> > >> On 2019/9/9 16:01, Jaegeuk Kim wrote:
> > >>> On 09/09, Chao Yu wrote:
> > >>>> On 2019/9/9 15:30, Jaegeuk Kim wrote:
> > >>>>> On 09/09, Chao Yu wrote:
> > >>>>>> On 2019/9/9 9:25, Jaegeuk Kim wrote:
> > >>>>>>> If committing atomic pages is failed when doing f2fs_do_sync_file(), we can
> > >>>>>>> get commited pages but atomic_file being still set like:
> > >>>>>>>
> > >>>>>>> - inmem:    0, atomic IO:    4 (Max.   10), volatile IO:    0 (Max.    0)
> > >>>>>>>
> > >>>>>>> If GC selects this block, we can get an infinite loop like this:
> > >>>>>>>
> > >>>>>>> f2fs_submit_page_bio: dev = (253,7), ino = 2, page_index = 0x2359a8, oldaddr = 0x2359a8, newaddr = 0x2359a8, rw = READ(), type = COLD_DATA
> > >>>>>>> f2fs_submit_read_bio: dev = (253,7)/(253,7), rw = READ(), DATA, sector = 18533696, size = 4096
> > >>>>>>> f2fs_get_victim: dev = (253,7), type = No TYPE, policy = (Foreground GC, LFS-mode, Greedy), victim = 4355, cost = 1, ofs_unit = 1, pre_victim_secno = 4355, prefree = 0, free = 234
> > >>>>>>> f2fs_iget: dev = (253,7), ino = 6247, pino = 5845, i_mode = 0x81b0, i_size = 319488, i_nlink = 1, i_blocks = 624, i_advise = 0x2c
> > >>>>>>> f2fs_submit_page_bio: dev = (253,7), ino = 2, page_index = 0x2359a8, oldaddr = 0x2359a8, newaddr = 0x2359a8, rw = READ(), type = COLD_DATA
> > >>>>>>> f2fs_submit_read_bio: dev = (253,7)/(253,7), rw = READ(), DATA, sector = 18533696, size = 4096
> > >>>>>>> f2fs_get_victim: dev = (253,7), type = No TYPE, policy = (Foreground GC, LFS-mode, Greedy), victim = 4355, cost = 1, ofs_unit = 1, pre_victim_secno = 4355, prefree = 0, free = 234
> > >>>>>>> f2fs_iget: dev = (253,7), ino = 6247, pino = 5845, i_mode = 0x81b0, i_size = 319488, i_nlink = 1, i_blocks = 624, i_advise = 0x2c
> > >>>>>>>
> > >>>>>>> In that moment, we can observe:
> > >>>>>>>
> > >>>>>>> [Before]
> > >>>>>>> Try to move 5084219 blocks (BG: 384508)
> > >>>>>>>   - data blocks : 4962373 (274483)
> > >>>>>>>   - node blocks : 121846 (110025)
> > >>>>>>> Skipped : atomic write 4534686 (10)
> > >>>>>>>
> > >>>>>>> [After]
> > >>>>>>> Try to move 5088973 blocks (BG: 384508)
> > >>>>>>>   - data blocks : 4967127 (274483)
> > >>>>>>>   - node blocks : 121846 (110025)
> > >>>>>>> Skipped : atomic write 4539440 (10)
> > >>>>>>>
> > >>>>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > >>>>>>> ---
> > >>>>>>>  fs/f2fs/file.c | 10 +++++-----
> > >>>>>>>  1 file changed, 5 insertions(+), 5 deletions(-)
> > >>>>>>>
> > >>>>>>> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > >>>>>>> index 7ae2f3bd8c2f..68b6da734e5f 100644
> > >>>>>>> --- a/fs/f2fs/file.c
> > >>>>>>> +++ b/fs/f2fs/file.c
> > >>>>>>> @@ -1997,11 +1997,11 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
> > >>>>>>>  			goto err_out;
> > >>>>>>>  
> > >>>>>>>  		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
> > >>>>>>> -		if (!ret) {
> > >>>>>>> -			clear_inode_flag(inode, FI_ATOMIC_FILE);
> > >>>>>>> -			F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
> > >>>>>>> -			stat_dec_atomic_write(inode);
> > >>>>>>> -		}
> > >>>>>>> +
> > >>>>>>> +		/* doesn't need to check error */
> > >>>>>>> +		clear_inode_flag(inode, FI_ATOMIC_FILE);
> > >>>>>>> +		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
> > >>>>>>> +		stat_dec_atomic_write(inode);
> > >>>>>>
> > >>>>>> If there are still valid atomic write pages linked in .inmem_pages, it may cause
> > >>>>>> memory leak when we just clear FI_ATOMIC_FILE flag.
> > >>>>>
> > >>>>> f2fs_commit_inmem_pages() should have flushed them.
> > >>>>
> > >>>> Oh, we failed to flush its nodes.
> > >>>>
> > >>>> However we won't clear such info if we failed to flush inmen pages, it looks
> > >>>> inconsistent.
> > >>>>
> > >>>> Any interface needed to drop inmem pages or clear ATOMIC_FILE flag in that two
> > >>>> error path? I'm not very clear how sqlite handle such error.
> > >>>
> > >>> f2fs_drop_inmem_pages() did that, but not in this case.
> > >>
> > >> What I mean is, for any error returned from atomic_commit() interface, should
> > >> userspace application handle it with consistent way, like trigger
> > >> f2fs_drop_inmem_pages(), so we don't need to handle it inside atomic_commit().
> > > 
> > > f2fs_ioc_abort_volatile_write() will be triggered.
> > 
> > If userspace can do this, we can get rid of this patch, or am I missing sth?
> 
> We don't know when that will come. And, other threads are waiting for GC here.
> 

Actually, we can call this.

---
 fs/f2fs/file.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
index 7ae2f3bd8c2f..98e2f58467d3 100644
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -1997,11 +1997,7 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
 			goto err_out;
 
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
-		if (!ret) {
-			clear_inode_flag(inode, FI_ATOMIC_FILE);
-			F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
-			stat_dec_atomic_write(inode);
-		}
+		f2fs_drop_inmem_pages(inode);
 	} else {
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 1, false);
 	}
-- 
2.19.0.605.g01d371f741-goog