Re: [syzbot] [netfs?] kernel BUG in folio_unlock (3)

[Dominique Martinet]

Matthew Wilcox wrote on Thu, Feb 20, 2025 at 04:15:53PM +0000:
> On Thu, Feb 20, 2025 at 08:00:24AM -0800, syzbot wrote:
> > ------------[ cut here ]------------
> > kernel BUG at mm/filemap.c:1499!
> 
> Tried to unlock a folio that wasn't locked.
> 
> The entire log is interesting:
> 
> https://syzkaller.appspot.com/x/log.txt?x=12af2fdf980000
> 
> It injects a failure which hits p9_tag_alloc() (so adding the 9p people
> to the cc)

9p is calling iov_iter_revert() in p9_client_write() on failure, but at
this point of the failure copy_from_iter_full (which advanced the iter)
wasn't called yet because the format processing happens after
allocation...

This was changed by Al Viro in 2015 so it's a "fairly old" bug, but it's
a bug on 9p side alright - thanks for the cc

Now to figure out how to decide if we want to revert or not... I
honestly don't have any bright idea, but I don't know the iov API well
at all -- perhaps it's possible to copy without advancing and only
advance the iov if IO worked?

-- 
Dominique Martinet | Asmadeus