Re: [GIT PULL] uprobes: tmpfs support

[Hugh Dickins <hughd@google.com>]

On Mon, 2 Jun 2014, Christoph Hellwig wrote:
> On Mon, Jun 02, 2014 at 04:14:06PM +0200, Oleg Nesterov wrote:
> > Ingo, please pull from
> > 
> >   git://git.kernel.org/pub/scm/linux/kernel/git/oleg/misc uprobes/core
> > 
> > Based on tip:perf/uprobes
> 
> Eww, adding tmpfs-specific code to uprobes screams layering violation.
> 
> Hugh, what is the problem with implementing ->readpage for tmpfs again?

The problem is that ->readpage invites the caller to allocate a page
of their choice for pagecache, and then pass it down to the filesystem
to fill and use thereafter.

There are several ways in which that does not suit tmpfs, 3 spring to mind:
1. the page may already be in memory, but currently in swapcache not in
filecache: tmpfs knows how to manage that, the ->readpage caller does not
2. there may be a NUMA mempolicy applied to that file, which would choose
to allocate the page differently: tmpfs knows about that, caller does not
3. (handy side-effect) it happens to disable use of tmpfs file as swapfile

It was a great relief when tmpfs could finally jettison its ->readpage
back in v3.1 (though if you press me, I could admit to some remaining
embarrassments).  I certainly do not want it back.

Just think of tmpfs as a layering violation itself (memory as backing!
no wonder it has peculiar demands on the allocation of its backing)
and we're all good - there's a variety of ways in which the generic
code already happens to accommodate it (many PageSwapBacked tests,
or the mapping_cap_account_dirty/writeback tests, for example).

IIRC, you were in on the discussion of shmem_read_mapping_page() when we
introduced it: Oleg is simply adding a call to it to fix a uprobes bug.  
That the name explicitly mentions shmem instead of concealing it,
is not necessarily a bad thing.

Hugh