Re: [RFC PATCH 31/35] Revert "vfs: add d_real_inode() helper"

[Miklos Szeredi]

On Thu, Apr 19, 2018 at 9:54 PM, Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:
> On Wed, Apr 18, 2018 at 03:49:02PM +0200, Miklos Szeredi wrote:
>> On Wed, Apr 18, 2018 at 3:38 PM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>> > On Wed, 18 Apr 2018 13:42:03 +0200
>> > Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>> >
>> >> On Wed, Apr 18, 2018 at 10:19 AM, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
>> >> > On Thu, Apr 12, 2018 at 6:08 PM, Miklos Szeredi <mszeredi@xxxxxxxxxx> wrote:
>> >> >> This reverts commit a118084432d642eeccb961c7c8cc61525a941fcb.
>> >> >>
>> >> >> No user of d_real_inode() remains, so it can be removed.
>> >> >>
>> >> >
>> >> > FYI, there is a new user in v4.17-rc1 added by commit
>> >> > f0a2aa5a2a40 tracing/uprobe: Add support for overlayfs
>> >> >
>> >> > Seems like this patch got merged without any CC to overlayfs
>> >> > mailing list nor maintainer?
>> >
>> > It appeared to be a small change with lots of reviewers. I didn't think
>> > it was something to notify the overlayfs folks with. But perhaps I was
>> > wrong.
>>
>> The patch is correct.  The code surrounding it isn't, though.
>>
>> >
>> >> >
>> >> > Not sure yet if overlayfs-rorw patches would allow reverting this
>> >> > change.
>> >>
>> >> Not trivial, because uprobe is looking at i_mapping to get a list of
>> >> current memory maps.   We could set i_mapping at overlay inode
>> >> initialization time, but we definitely can't *change* i_mapping at
>> >> copy up.  Which is bound to result in some weird inconsistencies.  So
>> >> likely we'll need to keep d_real_inode() for the time being.
>> >
>> > I just received this patch:
>> >
>> >  http://lkml.kernel.org/r/20180418062907.3210386-1-songliubraving@xxxxxx
>> >
>> > Which removes this code. Can you review it and I'll take it.
>>
>> It shouldn't remove d_real_inode(), because that part is correct and
>> fixes a real bug in handling overlayfs files.
>
> I am wondering what does it practically mean for metdata only copy up
> patches. Given this is uprobe code, I am assuming its modifying some
> executable code dynamically. And for the the case of metadata only
> copy up, it will return inode of lower. That probably means, as long
> as all running instances of that exeutable are using that inode, things
> will work fine.
>
> But if for some reason somebody opens that file for WRITE and triggers
> copy up and new instances of same binary will not see the probe taking
> affect?
>
> Which is probably very similar to what will happen if a lower executable
> is copied up. Having said that, in normal cases there should not be a
> need to copy up a binary in normal circumstances.

The only thing we need to ensure when uprobes interact with copy-ups
is that the kernel doesn't crash and doesn't leak memory.  Other than
that, it's a totally uninteresting corner case and we don't need to
worry about its behavior.

Thanks,
Miklos