Re: [PATCH v2] fuse: use newer inode info when writeback cache is enabled

[Fengnan Chang <changfengnan@vivo.com>]

Hi Miklos:

FYI, I run xfstest on fuse, with linux 5.4.61 + patch, no new failure case.

On 2021/6/24 15:42, Fengnan Chang wrote:
> Hi Miklos:
> 
> Thank you for the information, I have been able to reproduce the problem.
> 
> The new version of the patch as below. Previous fsx test is pass now. 
> Need do more test, Can you help to test new patch? or send me your test 
> case, I will test this.
> 
> Here is my test case, and is the problem this patch is trying to solve.
> Case A:
> mkdir /tmp/test
> passthrough_ll -ocache=always,writeback /mnt/test/
> echo "11111" > /tmp/test/fsx
> ls -l /mnt/test/tmp/test/
> echo "2222" >> /tmp/test/fsx
> ls -l /mnt/test/tmp/test/
> 
> Case B:
> mkdir /tmp/test
> passthrough_ll -ocache=always,writeback /mnt/test/
> passthrough_ll -ocache=always,writeback /mnt/test2/
> echo "11111" > /tmp/test/fsx
> ls -l /mnt/test/tmp/test/
> ls -l /mnt/test2/tmp/test/
> echo "222" >> /mnt/test/tmp/test/fsx
> ls -l /mnt/test/tmp/test/
> ls -l /mnt/test2/tmp/test/
> 
> 
> diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c
> index b9beb39a4a18..8e22a31b55c4 100644
> --- a/fs/fuse/inode.c
> +++ b/fs/fuse/inode.c
> @@ -60,6 +60,10 @@ MODULE_PARM_DESC(max_user_congthresh,
>   /** Congestion starts at 75% of maximum */
>   #define FUSE_DEFAULT_CONGESTION_THRESHOLD (FUSE_DEFAULT_MAX_BACKGROUND 
> * 3 / 4)
> 
> +static inline bool attr_newer_than_local(struct fuse_attr *attr, struct 
> inode *inode) {
> +    return (attr->mtime > inode->i_mtime.tv_sec)
> +               || ((attr->mtime == inode->i_mtime.tv_sec) && 
> (attr->mtimensec > inode->i_mtime.tv_nsec));
> +}
>   #ifdef CONFIG_BLOCK
>   static struct file_system_type fuseblk_fs_type;
>   #endif
> @@ -241,8 +245,10 @@ void fuse_change_attributes(struct inode *inode, 
> struct fuse_attr *attr,
>           * extend local i_size without keeping userspace server in 
> sync. So,
>           * attr->size coming from server can be stale. We cannot trust it.
>           */
> -       if (!is_wb || !S_ISREG(inode->i_mode))
> +       if (!is_wb || !S_ISREG(inode->i_mode)
> +               || (attr_newer_than_local(attr, inode) && 
> !inode_is_open_for_write(inode))) {
>                  i_size_write(inode, attr->size);
> +       }
>          spin_unlock(&fi->lock);
> 
>          if (!is_wb && S_ISREG(inode->i_mode)) {
> 
> On 2021/6/22 23:19, Miklos Szeredi wrote:
>> On Tue, 22 Jun 2021 at 14:25, Fengnan Chang <changfengnan@vivo.com> 
>> wrote:
>>>
>>> Unh, it seems i_writecount not work.
>>> If we modify file through lowerfs, i_writecount won't change, but the
>>> size already changed.
>>> For example:
>>> echo "111" > /lowerfs/test
>>> ls -l /upper/test
>>> echo "2222" >> /lowerfs/test
>>> ls -l /upper/test
>>>
>>> So, can you describe your test enviroment? including kernel version and
>>> fsx parameters, I will check it.
>>
>> linux-5.13-rc5 + patch
>> mkdir /tmp/test
>> libfuse/example/passthrough_ll -ocache=always,writeback /mnt/fuse/
>> fsx-linux -N 1000000 /mnt/fuse/tmp/test/fsx
>>
>> Thanks,
>> Miklos
>>