On 3/29/23 17:14, Christoph Hellwig wrote:
> On Wed, Mar 29, 2023 at 02:58:23PM +0900, Damien Le Moal wrote:
>> + /*
>> + * If the inode block size (sector size) is smaller than the
>> + * page size, we may be appending data belonging to an already
>> + * cached last page of the inode. So make sure to invalidate that
>> + * last cached page. This will always be a no-op for the case where
>> + * the block size is equal to the page size.
>> + */
>> + ret = invalidate_inode_pages2_range(inode->i_mapping,
>> + iocb->ki_pos >> PAGE_SHIFT, -1);
>> + if (ret)
>> + return ret;
>
> The missing truncate here obviously is a bug and needs fixing.
>
> But why does this not follow the logic in __iomap_dio_rw to to return
> -ENOTBLK for any error so that the write falls back to buffered I/O.
This is a write to sequential zones so we cannot use buffered writes. We have to
do a direct write to ensure ordering between writes.
Note that this is the special blocking write case where we issue a zone append.
For async regular writes, we use iomap so this bug does not exist. But then I
now realize that __iomap_dio_rw() falling back to buffered IOs could also create
an issue with write ordering.
> Also as far as I can tell from reading the code, -1 is not a valid
> end special case for invalidate_inode_pages2_range, so you'll actually
> have to pass a valid end here.
I wondered about that but then saw:
int invalidate_inode_pages2(struct address_space *mapping)
{
return invalidate_inode_pages2_range(mapping, 0, -1);
}
EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
which tend to indicate that "-1" is fine. The end is passed to
find_get_entries() -> find_get_entry() where it becomes a "max" pgoff_t, so
using -1 seems fine.
--
Damien Le Moal
Western Digital Research
