Re: [PATCH 2/4] tee: Remove vmalloc page support

[Ira Weiny]

On Fri, Oct 07, 2022 at 10:12:57AM +0200, Jens Wiklander wrote:
> On Thu, Oct 6, 2022 at 8:20 PM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Wed, Oct 5, 2022 at 11:24 PM Sumit Garg <sumit.garg@xxxxxxxxxx> wrote:
> > >
> > > Sorry but you need to get your driver mainline in order to support
> > > vmalloc interface.
> >
> > Actually, I think even then we shouldn't support vmalloc - and
> > register_shm_helper() just needs to be changed to pass in an array of
> > actual page pointers instead.
> 
> register_shm_helper() is an internal function, I suppose it's what's
> passed to tee_shm_register_user_buf() and especially
> tee_shm_register_kernel_buf() in this case.
> 
> So the gain is that in the kernel it becomes the caller's
> responsibility to provide the array of page pointers and the TEE
> subsystem doesn't need to care about what kind of kernel memory it is
> any longer. Yes, that should avoid eventual complexities with
> vmalloc() etc.

I finally spent some time digging into this again.

Overall I'm not opposed to trying to clean up the code more but I feel like the
removal of TEE_SHM_USER_MAPPED is too complex for the main goal; to remove a
caller of kmap_to_page().

Not only is that flag used in release_registered_pages() but it is also used in
tee_shm_fop_mmap().  I'm not following exactly why.  I think this is to allow
mmap of the tee_shm's allocated by kernel users?  Which I _think_ is
orthogonal to the callers of tee_shm_register_kernel_buf()?

> 
> >
> > At that point TEE_SHM_USER_MAPPED should also go away, because then
> > it's the caller that should just do either the user space page
> > pinning, or pass in the kernel page pointer.
> >
> > JensW, is there some reason that wouldn't work?
> 
> We still need to know if it's kernel or user pages in
> release_registered_pages().

Yes.

As I dug into this it seemed ok to define a tee_shm_kernel_free().  Pull out
the allocation of the page array from register_shm_helper() such that it could
be handled by tee_shm_register_kernel_buf() and this new tee_shm_kernel_free().

This seems reasonable because the only callers of tee_shm_register_kernel_buf()
are in trusted_tee.c and they all call tee_shm_free() on the registered memory
prior to returning.

Other callers[*] of tee_shm_free() obtained tee_shm from
tee_shm_alloc_kernel_buf() which AFAICT avoids all this nonsense.

[*] such as .../drivers/firmware/broadcom/tee_bnxt_fw.c.

> 
> The struct tee_shm:s acquired with syscalls from user space are
> reference counted. As are the kernel tee_shm:s, but I believe we could
> separate them to avoid reference counting tee_shm:s used by kernel
> clients if needed. I'll need to look closer at this if we're going to
> use that approach.
> 
> Without reference counting the caller of tee_shm_free() can be certain
> that the secure world is done with the memory so we could delegate the
> kernel pages part of release_registered_pages() to the caller instead.
> 

I'm not sure I follow you here.  Would this be along the lines of creating a
tee_shm_free_kernel() to be used in trusted_tee.c for those specific kernel
data?

Overall I feel like submitting this series again with Christoph and Al's
comments addressed is the best way forward to get rid of kmap_to_page().  I
would really like to get moving on that to avoid any further issues with the
kmap conversions.

But if folks feel strongly enough about removing that flag I can certainly try
to do so.

Ira

> Cheers,
> Jens
> 
> >
> >                  Linus