Re: Filling holes in ext2

[Adrian Cox <adrian@humboldt.co.uk>]

Andrew Morton wrote:

> generic_file_write() will mark the page not up-to-date in this case.
> I wonder what's actually going on?  Perhaps the fact that we've
> instantiated a block in ext2 outside i_size?

The problem is that the on-disk metadata now says that that disk block 
is part of the file. So as the page is not up-to-date, the next read 
operation will go to the disk and fetch that block of garbage into the 
page cache.

> If you change the error path in -ac's generic_file_write() thusly:
> 
> -	goto fail_write;
> +	status = -EFAULT;
> +	goto sync_failure;
> 
> does it fix it?

No difference, as the write is to a hole in the middle of the file, and 
doesn't change i_size.

> Can you send the code you're using to demonstrate this?

The heart of it is this:
fd = open(file, O_CREAT | O_TRUNC | O_RDWR, S_IRUSR | S_IWUSR);
lseek(fd, 16385, SEEK_SET);
write(fd, "x", 1);
lseek(fd, 12345, SEEK_SET);
write(fd, "y", 1);
	
During the final write, I intervene with gdb to force the 
__copy_from_user() to fail. It should be possible to do something 
similar with a multi-threaded application where another thread modifies 
the VM, but I haven't actually built a real example.

Then I just take a look at the file with od.

[usermode:~] od -x /mnt/adrian/test
0000000 0000 0000 0000 0000 0000 0000 0000 0000
*
0030000 0a79 0a79 0a79 0a79 0a79 0a79 0a79 0a79
*
0040000 7800
0040002


-- 
Adrian Cox   http://www.humboldt.co.uk/