Hi,
On 2019/1/23 21:50, Dominique Martinet wrote:
> Hou Tao wrote on Wed, Jan 23, 2019:
>>> write_end() has a comment that i_size cannot change under it because it
>>> has the i_mutex, but it's obviously not sufficient given the stat2inode
>>> code does not have it, so it needs to do the same dance as write_iter.
>>
>> OK, will do that in v3
After checking the code, i think v9fs_write_begin() truly doesn't need i_lock.
Because it is used for LOOSE or FSCACHE cache mode and under these two
modes the i_size will not updated at all (neither in v9fs_refresh_inode()
nor v9fs_vfs_getattr()). And for v9fs_file_write_iter(), the i_lock is needed.
>>
>> How about adding a helper as shown in the following lines ?
>>
>> static inline void v9fs_i_size_write(struct inode *inode, loff_t i_size)
>> {
>> spin_lock(&inode->i_lock);
>> i_size_write(inode, i_size);
>> spin_unlock(&inode->i_lock);
>> }
>
> Sure. I'm actually surprise no other part of the kernel has such helper,
> cifs seems to be using that pattern a lot too.
> Actually, looking a bit deeper fs/stack.c has this code:
> /*
> * If CONFIG_SMP or CONFIG_PREEMPT on 32-bit, it's vital for
> * fsstack_copy_inode_size() to hold some lock around
> * i_size_write(), otherwise i_size_read() may spin forever (see
> * include/linux/fs.h). We don't necessarily hold i_mutex when this
> * is called, so take i_lock for that case.
> *
> * And if CONFIG_LBDAF (on 32-bit), continue our effort to keep the
> * two halves of i_blocks in sync despite SMP or PREEMPT: use i_lock
> * for that case too, and do both at once by combining the tests.
> *
> * There is none of this locking overhead in the 64-bit case.
> */
> if (sizeof(i_size) > sizeof(long) || sizeof(i_blocks) > sizeof(long))
> spin_lock(&dst->i_lock);
> i_size_write(dst, i_size);
> dst->i_blocks = i_blocks;
> if (sizeof(i_size) > sizeof(long) || sizeof(i_blocks) > sizeof(long))
> spin_unlock(&dst->i_lock);
>
> It might make sense to do the same in our little helper ?
OK. And there will be no performance loss in 64-bit case.
>
> (it looks like i_blocks has the same problem? speaking of which we
> probably do not want to update i_blocks either in the KEEP_SIZE
> case...?)
>
Yes, i_blocks may be corrupted under 32-bit case, and maybe it's appropriate
to fix it in a separated patch.
>>> As a nitpick I don't really like foo() vs foo_flags() as
>>> foo-that-takes-extra-flags.
>>> There are a few such examples in the kernel already but I think it does
>>> not really convery information; it's better to have the base function
>>> take flags and just use it, or if you want wrappers then just never
>>> expose the flags but make a static _v9fs_stat2inode take flags,
>>> v9fs_stat2inode behave as the old one and a new
>>> v9fs_stat2inode_keepisize for the update with cache.
>>> I'd personally go with the former are there only are four call sites.
>>
>> I agree with you. I will add a new flags parameter to v9fs_stat2inode() and use
>> it directly instead of creating inline wrappers around it.
>
> Thanks.
>
