Thread (92 messages) flat view 92 messages, 6 authors, 2017-07-13

Re: [RFC v5 00/38] powerpc: Memory Protection Keys

From: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Date: 2017-07-12 22:54:39
Also in: linux-arch, linux-mm, lkml

On Wed, 2017-07-12 at 09:23 +0200, Michal Hocko wrote:
quoted
Ideally the MMU looks at the PTE for keys, in order to enforce
protection. This is the case with x86 and is the case with power9 Radix
page table. Hence the keys have to be programmed into the PTE.
But x86 doesn't update ptes for PKEYs, that would be just too expensive.
You could use standard mprotect to do the same...
What do you mean ? x86 ends up in mprotect_fixup -> change_protection()
which will update the PTEs just the same as we do.

Changing the key for a page is a form mprotect. Changing the access
permissions for keys is different, for us it's a special register
(AMR).

I don't understand why you think we are doing any differently than x86
here.
quoted
However with HPT on power, these keys do not necessarily have to be
programmed into the PTE. We could bypass the Linux Page Table Entry(PTE)
and instead just program them into the Hash Page Table(HPTE), since
the MMU does not refer the PTE but refers the HPTE. The last version
of the page attempted to do that.   It worked as follows:

a) when a address range is requested to be associated with a key; by the
   application through key_mprotect() system call, the kernel
   stores that key in the vmas corresponding to that address
   range.

b) Whenever there is a hash page fault for that address, the fault
   handler reads the key from the VMA and programs the key into the
   HPTE. __hash_page() is the function that does that.
What causes the fault here?
The hardware. With the hash MMU, the HW walks a hash table which is
effectively a large in-memory TLB extension. When a page isn't found
there, a  "hash fault" is generated allowing Linux to populate that
hash table with the content of the corresponding PTE. 
quoted
c) Once the hpte is programmed, the MMU can sense key violations and
   generate key-faults.

The problem is with step (b).  This step is really a very critical
path which is performance sensitive. We dont want to add any delays.
However if we want to access the key from the vma, we will have to
hold the vma semaphore, and that is a big NO-NO. As a result, this
design had to be dropped.



I reverted back to the old design i.e the design in v4 version. In this
version we do the following:

a) when a address range is requested to be associated with a key; by the
   application through key_mprotect() system call, the kernel
   stores that key in the vmas corresponding to that address
   range. Also the kernel programs the key into Linux PTE coresponding to all the
   pages associated with the address range.
OK, so how is this any different from the regular mprotect then?
It takes the key argument. This is nothing new. This was done for x86
already, we are just re-using the infrastructure. Look at
do_mprotect_pkey() in mm/mprotect.c today. It's all the same code,
pkey_mprotect() is just mprotect with an added key argument.
quoted
b) Whenever there is a hash page fault for that address, the fault
   handler reads the key from the Linux PTE and programs the key into 
   the HPTE.

c) Once the HPTE is programmed, the MMU can sense key violations and
   generate key-faults.


Since step (b) in this case has easy access to the Linux PTE, and hence
to the key, it is fast to access it and program the HPTE. Thus we avoid
taking any performance hit on this critical path.

Hope this explains the rationale,


As promised here is the high level design:
I will read through that later
[...]
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help