On 07/10, Chao Yu wrote: > On 2020/7/10 11:26, Jaegeuk Kim wrote: > > On 07/10, Chao Yu wrote: > >> On 2020/7/10 3:05, Jaegeuk Kim wrote: > >>> On 07/09, Chao Yu wrote: > >>>> On 2020/7/9 13:30, Jaegeuk Kim wrote: > >>>>> It doesn't need to bypass flushing quota data in background. > >>>> > >>>> The condition is used to flush quota data in batch to avoid random > >>>> small-sized udpate, did you hit any problem here? > >>> > >>> I suspect this causes fault injection test being stuck by waiting for > >>> inode > >>> writeback completion. With this patch, it has been running w/o any issue > >>> so far. > >>> I keep an eye on this. > >> > >> Hmmm.. so that this patch may not fix the root cause, and it may hiding the > >> issue deeper. > >> > >> How about just keeping this patch in our private branch to let fault > >> injection > >> test not be stuck? until we find the root cause in upstream codes. > > > > Well, I don't think this hides something. When the issue happens, I saw > > inodes > > being stuck due to writeback while only quota has some dirty data. At that > > time, > > there was no dirty data page from other inodes. > > Okay, > > > > > More specifically, I suspect __writeback_inodes_sb_nr() gives WB_SYNC_NONE > > and > > waits for wb_wait_for_completion(). > > Did you record any callstack after the issue happened?
I found this. [213389.297642] __schedule+0x2dd/0x780^M [213389.299224] schedule+0x55/0xc0^M [213389.300745] wb_wait_for_completion+0x56/0x90^M [213389.302469] ? wait_woken+0x80/0x80^M [213389.303997] __writeback_inodes_sb_nr+0xa8/0xd0^M [213389.305760] writeback_inodes_sb+0x4b/0x60^M [213389.307439] sync_filesystem+0x2e/0xa0^M [213389.308999] generic_shutdown_super+0x27/0x110^M [213389.310738] kill_block_super+0x27/0x50^M [213389.312327] kill_f2fs_super+0x76/0xe0 [f2fs]^M [213389.314014] deactivate_locked_super+0x3b/0x80^M [213389.315692] deactivate_super+0x3e/0x50^M [213389.317226] cleanup_mnt+0x109/0x160^M [213389.318718] __cleanup_mnt+0x12/0x20^M [213389.320177] task_work_run+0x70/0xb0^M [213389.321609] exit_to_usermode_loop+0x131/0x160^M [213389.323306] do_syscall_64+0x170/0x1b0^M [213389.324762] entry_SYSCALL_64_after_hwframe+0x44/0xa9^M [213389.326477] RIP: 0033:0x7fc4b5e6a35b^M > > Still I'm confused that why directory's data written could be skipped, but > quota's data couldn't, what's the difference? I suspect different blocking timing from cp_error between quota and dentry. e.g., we block dir operations right after cp_error, while quota can make dirty pages in more fine granularity. > > > > >> > >> Thanks, > >> > >>> > >>> Thanks, > >>> > >>>> > >>>> Thanks, > >>>> > >>>>> > >>>>> Signed-off-by: Jaegeuk Kim <[email protected]> > >>>>> --- > >>>>> fs/f2fs/data.c | 2 +- > >>>>> 1 file changed, 1 insertion(+), 1 deletion(-) > >>>>> > >>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c > >>>>> index 44645f4f914b6..72e8b50e588c1 100644 > >>>>> --- a/fs/f2fs/data.c > >>>>> +++ b/fs/f2fs/data.c > >>>>> @@ -3148,7 +3148,7 @@ static int __f2fs_write_data_pages(struct > >>>>> address_space *mapping, > >>>>> if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING))) > >>>>> goto skip_write; > >>>>> > >>>>> - if ((S_ISDIR(inode->i_mode) || IS_NOQUOTA(inode)) && > >>>>> + if (S_ISDIR(inode->i_mode) && > >>>>> wbc->sync_mode == WB_SYNC_NONE && > >>>>> get_dirty_pages(inode) < nr_pages_to_skip(sbi, > >>>>> DATA) && > >>>>> f2fs_available_free_memory(sbi, DIRTY_DENTS)) > >>>>> > >>> . > >>> > > . > >
