Messages in this thread | | | Subject | Re: [f2fs-dev] [PATCH 2/2] f2fs: avoid infinite GC loop due to stale atomic files | From | Chao Yu <> | Date | Mon, 9 Sep 2019 19:26:17 +0800 |
| |
On 2019/9/9 16:38, 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@kernel.org> >>>>>>>>> --- >>>>>>>>> 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.
Yes, however, even atomic_write won't be called sometimes... that's why we add handling logic in f2fs_gc().
> >> >> - f2fs_ioc_abort_volatile_write >> - f2fs_drop_inmem_pages >> - clear_inode_flag(inode, FI_ATOMIC_FILE); >> - fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0; >> - stat_dec_atomic_write(inode); >> >>> >>>> >>>>> >>>>>> >>>>>> Thanks, >>>>>> >>>>>>> >>>>>>>> >>>>>>>> So my question is why below logic didn't handle such condition well? >>>>>>>> >>>>>>>> f2fs_gc() >>>>>>>> >>>>>>>> if (has_not_enough_free_secs(sbi, sec_freed, 0)) { >>>>>>>> if (skipped_round <= MAX_SKIP_GC_COUNT || >>>>>>>> skipped_round * 2 < round) { >>>>>>>> segno = NULL_SEGNO; >>>>>>>> goto gc_more; >>>>>>>> } >>>>>>>> >>>>>>>> if (first_skipped < last_skipped && >>>>>>>> (last_skipped - first_skipped) > >>>>>>>> sbi->skipped_gc_rwsem) { >>>>>>>> f2fs_drop_inmem_pages_all(sbi, true); >>>>>>> >>>>>>> This is doing nothing, since f2fs_commit_inmem_pages() removed the inode >>>>>>> from inmem list.
One possible solution is keeping inode in sbi->inode_lock[ATOMIC_FILE] list until f2fs_do_sync_file() return, then f2fs_drop_inmem_pages_all() can find and drop the inode.
Thanks,
>>>>>>> >>>>>>>> segno = NULL_SEGNO; >>>>>>>> goto gc_more; >>>>>>>> } >>>>>>>> if (gc_type == FG_GC && !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) >>>>>>>> ret = f2fs_write_checkpoint(sbi, &cpc); >>>>>>>> } >>>>>>>> >>>>>>>>> } else { >>>>>>>>> ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 1, false); >>>>>>>>> } >>>>>>>>> >>>>>>> . >>>>>>> >>>>> . >>>>> >>> . >>> > . >
| |