Re: [PATCH 10/17] prmem: documentation

From: Igor Stoppa
Date: Wed Oct 31 2018 - 15:38:33 EST

On 31/10/2018 11:08, Igor Stoppa wrote:
Adding SELinux folks and the SELinux ml

I think it's better if they participate in this discussion.

On 31/10/2018 06:41, Andy Lutomirski wrote:
On Tue, Oct 30, 2018 at 2:36 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:

On Tue, Oct 30, 2018 at 10:43:14PM +0200, Igor Stoppa wrote:
On 30/10/2018 21:20, Matthew Wilcox wrote:
So the API might look something like this:

ÂÂÂÂÂÂÂÂ void *p = rare_alloc(...);ÂÂÂÂÂ /* writable pointer */
ÂÂÂÂÂÂÂÂ p->a = x;
ÂÂÂÂÂÂÂÂ q = rare_protect(p);ÂÂÂÂÂÂÂÂÂÂÂ /* read-only pointer */

With pools and memory allocated from vmap_areas, I was able to say


and that would do a swipe on all the pages currently in use.
In the SELinux policyDB, for example, one doesn't really want to
individually protect each allocation.

The loading phase happens usually at boot, when the system can be assumed to
be sane (one might even preload a bare-bone set of rules from initramfs and
then replace it later on, with the full blown set).

There is no need to process each of these tens of thousands allocations and
initialization as write-rare.

Would it be possible to do the same here?

What Andy is proposing effectively puts all rare allocations into
one pool. Although I suppose it could be generalised to multiple pools
... one mm_struct per pool. Andy, what do you think to doing that?

Hmm. Let's see.

To clarify some of this thread, I think that the fact that rare_write
uses an mm_struct and alias mappings under the hood should be
completely invisible to users of the API.

I agree.

No one should ever be
handed a writable pointer to rare_write memory (except perhaps during
bootup or when initializing a large complex data structure that will
be rare_write but isn't yet, e.g. the policy db).

The policy db doesn't need to be write rare.
Actually, it really shouldn't be write rare.

Maybe it's just a matter of wording, but effectively the policyDB can be trated with this sequence:

1) allocate various data structures in writable form

2) initialize them

3) go back to 1 as needed

4) lock down everything that has been allocated, as Read-Only
The reason why I stress ReadOnly is that differentiating what is really ReadOnly from what is WriteRare provides an extra edge against attacks, because attempts to alter ReadOnly data through a WriteRare API could be detected

5) read any part of the policyDB during regular operations

6) in case of update, create a temporary new version, using steps 1..3

7) if update successful, use the new one and destroy the old one

8) if the update failed, destroy the new one

The destruction at points 7 and 8 is not so much a write operation, as it is a release of the memory.

So, we might have a bit different interpretation of what write-rare means wrt destroying the memory and its content.

To clarify: I've been using write-rare to indicate primarily small operations that one would otherwise achieve with "=", memcpy or memset or more complex variants, like atomic ops, rcu pointer assignment, etc.

Tearing down an entire set of allocations like the policyDB doesn't fit very well with that model.

The only part which _needs_ to be write rare, in the policyDB, is the set of pointers which are used to access all the dynamically allocated data set.

These pointers must be updated when a new policyDB is allocated.

For example, there could easily be architectures where having a
writable alias is problematic. On such architectures, an entirely
different mechanism might work better. And, if a tool like KNOX ever
becomes a *part* of the Linux kernel (hint hint!)

Something related, albeit not identical is going on here [1]
Eventually, it could be expanded to deal also with write rare.

If you have multiple pools and one mm_struct per pool, you'll need a
way to find the mm_struct from a given allocation.

Indeed. In my patchset, based on vmas, I do the following:
* a private field from the page struct points to the vma using that page
* inside the vma there is alist_head used only during deletion
 - one pointer is used to chain vmas fro mthe same pool
 - one pointer points to the pool struct
* the pool struct has the property to use for all the associated allocations: is it write-rare, read-only, does it auto protect, etc.

Regardless of how
the mm_structs are set up, changing rare_write memory to normal memory
or vice versa will require a global TLB flush (all ASIDs and global
pages) on all CPUs, so having extra mm_structs doesn't seem to buy

1) it supports differnt levels of protection:
 temporarily unprotected vs read-only vs write-rare

2) the change of write permission should be possible only toward more restrictive rules (writable -> write-rare -> read-only) and only to the point that was specified while creating the pool, to avoid DOS attacks, where a write-rare is flipped into read-only and further updates fail
(ex: prevent IMA from registering modifications to a file, by not letting it store new information - I'm not 100% sure this would work, but it gives the idea, I think)

3) being able to track all the allocations related to a pool would allow to perform mass operations, like reducing the writabilty or destroying all the allocations.

(It's just possible that changing rare_write back to normal might be
able to avoid the flush if the spurious faults can be handled

I do not see the need for such a case of degrading the write permissions of an allocation, unless it refers to the release of a pool of allocations (see updating the SELinux policy DB)

A few more thoughts about pools. Not sure if they are all correct.
Note: I stick to "pool", instead of mm_struct, because what I'll say is mostly independent from the implementation.

- As Peter Zijlstra wrote earlier, protecting a target moves the focus of the attack to something else. In this case, probably, the "something else" would be the metadata of the pool(s).

- The amount of pools needed should be known at compilation time, so the metadata used for pools could be statically allocated and any change to it could be treated as write-rare.

- only certain fields of a pool structure would be writable, even as write-rare, after the pool is initialized.
In case the pool is a mm_struct or a superset (containing also additional properties, like the type of writability: RO or WR), the field

struct vm_area_struct *mmap;

is an example of what could be protected. It should be alterable only when creating/destroying the pool and making the first initialization.

- to speed up and also improve the validation of the target of a write-rare operation, it would be really desirable if the target had some intrinsic property which clearly differentiates it from non-write-rare memory. Its address, for example. The amount of write rare data needed by a full kernel should not exceed a few tens of megabytes. On a 64 bit system it shouldn't be so bad to reserve an address range maybe one order of magnitude lager than that.
It could even become a parameter for the creation of a pool.
SELinux, for example, should fit within 10-20MB. Or it could be a command line parameter.

- even if an hypervisor was present, it would be preferable to use it exclusively as extra protection, which triggers an exception only when something abnormal happens. The hypervisor should not become aware of the actual meaning of kernel (meta)data. Ideally, it would be mostly used for trapping unexpected writes to pages which are not supposed to be modified.

- one more reason for using pools is that, if each pool was also acting as memory cache for its users, attacks relying on use-after-free would not have access to possible vulnerability, because the memory and addresses associated with a pool would stay with it.