Re: [PATCH v2 00/24] exec: Move unshare_files and guarantee files_struct.count is correct

[Al Viro]

On Sat, Nov 28, 2020 at 05:12:26AM +0000, Al Viro wrote:
> On Fri, Nov 20, 2020 at 04:05:47PM -0800, Linus Torvalds wrote:
> > On Fri, Nov 20, 2020 at 3:11 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
> > >
> > > This set of changes cleanups of the code in exec so hopefully this code
> > > will not regress again.  Then it adds helpers and fixes the users of
> > > files_struct so the reference count is only incremented if COPY_FILES is
> > > passed to clone (or if io_uring takes a reference).  Then it removes
> > > helpers (get_files_struct, __install_fd, __alloc_fd, __close_fd) that
> > > are no longer needed and if used would encourage code that increments
> > > the count of files_struct somewhere besides in clone when COPY_FILES is
> > > passed.
> > 
> > I'm not seeing anything that triggered me going "that looks dodgy". It
> > all looks like nice cleanups.
> > 
> > But that's just from reading the patches (and in some cases going and
> > looking at the context), so I didn't actually _test_ any of it. It all
> > looks sane to me, though, and the fact that it removes a fair number
> > of lines of code is always a good sign.
> > 
> > It would be good for people to review and test (Al? Oleg? others?),
> > but my gut feel is "this is good".
> 
> Will check (sorry, the last couple of weeks had been bloody awful -
> off-net and very short on sleep); I'm digging through the piles of
> email right now.

	TBH, the thing that makes me uneasy about that series is handling
the task_lock().  Holding it for longer periods in general, plus boinking
it a _lot_ at least on /proc/*/fd getdents(2).  There might be other
places like that I'd missed - I'm rereading the series now that I've
noticed that one.

	Other than that I mostly like it.  I would rather reduce the
access to other processes descriptor tables (in particular, slapped
a very clear "don't use those helpers unless you really need it; we
will be watching for new call sites and you *will* have to explain
yourself if you add such" on them), but that's not a new problem -
fcheck_files() et.al. had to be watched anyway.

	One thing I would like to achieve is fewer places grabbing
struct file references from descriptor tables other than current->files,
but that's mostly orthogonal to this series.  I'll need to resurrect
a local branch trying to do that; there are conflicts (unsurprisingly),
but due to the amount of bitrot it has accumulated it'll have to be
redone anyway.