[PATCH v17 07/10] mm: introduce memfd_secret system call to create "secret" memory areas

[James Bottomley]

On Mon, 2021-02-15 at 10:13 +0100, Michal Hocko wrote:
> On Sun 14-02-21 11:21:02, James Bottomley wrote:
> > On Sun, 2021-02-14 at 10:58 +0100, David Hildenbrand wrote:
> > [...]
> > > > And here we come to the question "what are the differences that
> > > > justify a new system call?" and the answer to this is very
> > > > subjective. And as such we can continue bikeshedding forever.
> > > 
> > > I think this fits into the existing memfd_create() syscall just
> > > fine, and I heard no compelling argument why it shouldn‘t. That‘s
> > > all I can say.
> > 
> > OK, so let's review history.  In the first two incarnations of the
> > patch, it was an extension of memfd_create().  The specific
> > objection by Kirill Shutemov was that it doesn't share any code in
> > common with memfd and so should be a separate system call:
> > 
> > https://lore.kernel.org/linux-api/20200713105812.dnwtdhsuyj3xbh4f@box/
> 
> Thanks for the pointer. But this argument hasn't been challenged at
> all. It hasn't been brought up that the overlap would be considerable
> higher by the hugetlb/sealing support. And so far nobody has claimed
> those combinations as unviable.

Kirill is actually interested in the sealing path for his KVM code so
we took a look.  There might be a two line overlap in memfd_create for
the seal case, but there's no real overlap in memfd_add_seals which is
the bulk of the code.  So the best way would seem to lift the inode ...
-> seals helpers to be non-static so they can be reused and roll our
own add_seals.

I can't see a use case at all for hugetlb support, so it seems to be a
bit of an angels on pin head discussion.  However, if one were to come
along handling it in the same way seems reasonable.

> > The other objection raised offlist is that if we do use
> > memfd_create, then we have to add all the secret memory flags as an
> > additional ioctl, whereas they can be specified on open if we do a
> > separate system call.  The container people violently objected to
> > the ioctl because it can't be properly analysed by seccomp and much
> > preferred the syscall version.
> > 
> > Since we're dumping the uncached variant, the ioctl problem
> > disappears but so does the possibility of ever adding it back if we
> > take on the container peoples' objection.  This argues for a
> > separate syscall because we can add additional features and extend
> > the API with flags without causing anti-ioctl riots.
> 
> I am sorry but I do not understand this argument.

You don't understand why container guarding technology doesn't like
ioctls?  The problem is each ioctl is the multiplexor is specific to
the particular fd implementation, so unlike fcntl you don't have global
ioctl numbers (although we do try to separate the space somewhat with
the _IO macros).  This makes analysis and blocking a hard problem for
container seccomp.

>  What kind of flags are we talking about and why would that be a
> problem with memfd_create interface? Could you be more specific
> please?

You mean what were the ioctl flags in the patch series linked above? 
They were SECRETMEM_EXCLUSIVE and SECRETMEM_UNCACHED in patch 3/5. 
They were eventually dropped after v10, because of problems with
architectural semantics, with the idea that it could be added back
again if a compelling need arose:

https://lore.kernel.org/linux-api/20201123095432.5860-1-rppt@kernel.org/

In theory the extra flags could be multiplexed into the memfd_create
flags like hugetlbfs is but with 32 flags and a lot already taken it
gets messy for expansion.  When we run out of flags the first question
people will ask is "why didn't you do separate system calls?".

James