Re: åå: [PATCH] f2fs: fix a race condition between f2fs_write_end_io and f2fs_del_fsync_node_entry

From: Chao Yu
Date: Wed Jun 17 2020 - 22:57:54 EST


On 2020/6/18 10:39, Zac wrote:
>
>> On 2020/6/17 17:04, zhaowuyun@xxxxxxxxxxxx wrote:
>>> From: Wuyun Zhao <zhaowuyun@xxxxxxxxxxxx>
>>>
>>> Under some condition, the __write_node_page will submit a page which is
>> not
>>> f2fs_in_warm_node_list and will not call f2fs_add_fsync_node_entry.
>>> f2fs_gc continue to run to invoke f2fs_iget -> do_read_inode to read the
>> same node page
>>> and set code node, which make f2fs_in_warm_node_list become true,
>>> that will cause f2fs_bug_on in f2fs_del_fsync_node_entry when
>> f2fs_write_end_io called.
>> Could you please add below race condition description into commit
>> message?
>>
>> - f2fs_write_end_io
>> - f2fs_iget
>> - do_read_inode
>> - set_cold_node
>> recover cold node flag
>> - f2fs_in_warm_node_list
>> - is_cold_node
>> if node is cold, assume we have added
>> node to fsync_node_list during writepages()
>> - f2fs_del_fsync_node_entry
>> - f2fs_bug_on() due to node page
>> is not in fsync_node_list
>
> Ok, will add the commit message.
>
>> BTW, I'm curious about why we can lose cold flag for non-dir inode?
>> any clue to reproduce this bug (I mean losing cold flag)?
>
> it's a f2fs image with 25600MB
> flash this image to device
> the device will resize it according to the userdata partition size which is
> about 94GB
> the device mount the f2fs partition
> then hit this f2fs_bug_on
>
> seems that the cold flag is not been set when mkfs

Ah, I guess both mkfs/sload ignores setting cold node flag for non-dir inode,
could you please send another patch to fix this issue?

>
> I think the issue is that
>
> 1. the node page in the storage is without cold bit
> 2. f2fs_disable_checkpoint -> f2fs_gc -> f2fs_get_node_page, this page won't
> be set cold flag
> 3. f2fs_move_node_page -> __write_node_page to write this page
> 4. f2fs_gc -> f2fs_iget -> do_read_inode to read this page and set cold flag

Clear enough, thanks for your explanation. :)

Thanks,

>
>>>
>>> [ 34.966133] Call trace:
>>> [ 34.969902] f2fs_del_fsync_node_entry+0x100/0x108
>>> [ 34.976071] f2fs_write_end_io+0x1e0/0x288
>>> [ 34.981539] bio_endio+0x248/0x270
>>> [ 34.986289] blk_update_request+0x2b0/0x4d8
>>> [ 34.991841] scsi_end_request+0x40/0x440
>>> [ 34.997126] scsi_io_completion+0xa4/0x748
>>> [ 35.002593] scsi_finish_command+0xdc/0x110
>>> [ 35.008143] scsi_softirq_done+0x118/0x150
>>> [ 35.013610] blk_done_softirq+0x8c/0xe8
>>> [ 35.018811] __do_softirq+0x2e8/0x578
>>> [ 35.023828] irq_exit+0xfc/0x120
>>> [ 35.028398] handle_IPI+0x1d8/0x330
>>> [ 35.033233] gic_handle_irq+0x110/0x1d4
>>> [ 35.038433] el1_irq+0xb4/0x130
>>> [ 35.042917] kmem_cache_alloc+0x3f0/0x418
>>> [ 35.048288] radix_tree_node_alloc+0x50/0xf8
>>> [ 35.053933] __radix_tree_create+0xf8/0x188
>>> [ 35.059484] __radix_tree_insert+0x3c/0x128
>>> [ 35.065035] add_gc_inode+0x90/0x118
>>> [ 35.069967] f2fs_gc+0x1b80/0x2d70
>>> [ 35.074718] f2fs_disable_checkpoint+0x94/0x1d0
>>> [ 35.080621] f2fs_fill_super+0x10c4/0x1b88
>>> [ 35.086088] mount_bdev+0x194/0x1e0
>>> [ 35.090923] f2fs_mount+0x40/0x50
>>> [ 35.095589] mount_fs+0xb4/0x190
>>> [ 35.100159] vfs_kern_mount+0x80/0x1d8
>>> [ 35.105260] do_mount+0x478/0xf18
>>> [ 35.109926] ksys_mount+0x90/0xd0
>>> [ 35.114592] __arm64_sys_mount+0x24/0x38
>>>
>>> Signed-off-by: Wuyun Zhao <zhaowuyun@xxxxxxxxxxxx>
>>
>> Reviewed-by: Chao Yu <yuchao0@xxxxxxxxxx>
>>
>> Thanks,
>>
>>> ---
>>> fs/f2fs/inode.c | 1 +
>>> 1 file changed, 1 insertion(+)
>>>
>>> diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c
>>> index be6ac33..0df5c8c 100644
>>> --- a/fs/f2fs/inode.c
>>> +++ b/fs/f2fs/inode.c
>>> @@ -402,6 +402,7 @@ static int do_read_inode(struct inode *inode)
>>>
>>> /* try to recover cold bit for non-dir inode */
>>> if (!S_ISDIR(inode->i_mode) && !is_cold_node(node_page)) {
>>> + f2fs_wait_on_page_writeback(node_page, NODE, true, true);
>>> set_cold_node(node_page, false);
>>> set_page_dirty(node_page);
>>> }
>>>
>
> .
>