On 2020/6/11 0:27, Jan Kara wrote:
> On Wed 10-06-20 11:45:43, Theodore Y. Ts'o wrote:
>> On Wed, Jun 10, 2020 at 11:57:39AM +0200, Jan Kara wrote:
>>>> So I guess it may still lead to inconsistency. How about add this checking
>>>> into ext4_journal_get_write_access() ?
>>>
>>> Yes, this also occured to me later. Adding the check to
>>> ext4_journal_get_write_access() should be safer.
>>
>> There's another thing which we could do. One of the issues is that we
>> allow buffered writeback for block devices once the change to the
>> block has been committed. What if we add a change to block device
>> writeback code and in fs/buffer.c so that optionally, the file system
>> can specify a callback function can get called when an I/O error has
>> been reflected back up from the block layer?
>>
>> It seems unfortunate that currently, we can immediately report the I/O
>> error for buffered writes to *files*, but for metadata blocks, we
>> would only be able to report the problem when we next try to modify
>> it.
>>
>> Making changes to fs/buffer.c might be controversial, but I think it
>> might be result in a better solution.
>
> Yeah, what you propose certainly makes sence could be relatively easily
> done by blkdev_writepage() using __block_write_full_page() with appropriate
> endio handler which calls fs callback. I'm just not sure how propagate the
> callback function from the fs to the blkdev...
>
I have thought about this solution, we could add a hook in 'struct super_operations'
and call it in blkdev_writepage() like blkdev_releasepage() does, and pick out a
wrapper from block_write_full_page() to pass our endio handler in, something like
this.
static const struct super_operations ext4_sops = {
...
.bdev_write_page = ext4_bdev_write_page,
...
};
static int blkdev_writepage(struct page *page, struct writeback_control *wbc)
{
struct super_block *super = BDEV_I(page->mapping->host)->bdev.bd_super;
if (super && super->s_op->bdev_write_page)
return super->s_op->bdev_write_page(page, blkdev_get_block, wbc);
return block_write_full_page(page, blkdev_get_block, wbc);
}
But I'm not sure it's a optimal ieda. So I continue to realize the "wb_err"
solution now ?
Thanks,
Yi.
