BPF: RCU use-after-reallocation of hash table elements?

BPF: RCU use-after-reallocation of hash table elements?

[Jann Horn]

Hi!

Note: I haven't tested any of this; feel free to tell me that I've
completely misunderstood how all this works.

The BPF manpage, at the moment, states about BPF hash tables:

       BPF_MAP_TYPE_HASH
              Hash-table maps have the following characteristics:

              *  Maps are created and destroyed by user-space programs.
                 Both user-space and eBPF programs can perform lookup,
                 update, and delete operations.

              *  The kernel takes care of allocating and freeing key/value
                 pairs.

              *  The map_update_elem() helper will fail to insert new ele‐
                 ment when the max_entries limit is reached.  (This ensures
                 that eBPF programs cannot exhaust memory.)

              *  map_update_elem() replaces existing elements atomically.

              Hash-table maps are optimized for speed of lookup.

This documentation claims that elements are replaced "atomically", and
that the kernel "takes care of allocating and freeing key/value
pairs". But as far as I can tell, that's not quite the whole story
least since commit 6c90598174322b8888029e40dd84a4eb01f56afe (first in
4.6).

Unless a BPF hash table is created with the (undocumented) flag
BPF_F_NO_PREALLOC, the kernel now actually pre-allocates the hash
table elements. Hash table elements can be freed and reused for new
allocations (!) without waiting for an RCU grace period: Freed
elements are immediately pushed on the percpu freelist, and can be
immediately reused from there. The most obvious consequence of this is
that if a BPF program looks up a hash table entry and then reads the
value, the value can be replaced with a new value in between. A more
subtle consequence is that BPF map lookups can return false-positive
results: If the first half of the lookup key matches the old key, and
the second half of the lookup key matches the new key, then a BPF map
lookup can return a false-positive result, as far as I can tell.

If what I'm saying is correct, I'm not sure what the best fix is.

Add a grace period when freeing hash map entries, and add a new -EBUSY
return value for attempts to create hash map entries when all free
entries are waiting for the end of an RCU grace period?

Add a grace period when freeing hash map entries, and use
rcu_synchronize() when inserting BPF hashmap entries from userspace
and all free entries are waiting for RCU? But that still leaves the
bpf_map_update_elem_proto helper that can be called from BPF.
Deprecate that helper for access to hash maps?

Document the race, and advise people who use BPF for
non-performance-tracing purposes (where occasional false positives
actually matter) to use BPF_F_NO_PREALLOC?

Add some sort of sequence lock to BPF (yuck)?

Re: BPF: RCU use-after-reallocation of hash table elements?

[Alexei Starovoitov]

On Wed, Oct 03, 2018 at 06:15:08PM +0200, Jann Horn wrote:
> Hi!
> 
> Note: I haven't tested any of this; feel free to tell me that I've
> completely misunderstood how all this works.
> 
> The BPF manpage, at the moment, states about BPF hash tables:

unfortunately man pages have been obsolete for long time.
the only doc that is mostly up to date is
http://docs.cilium.io/en/latest/bpf/

> Unless a BPF hash table is created with the (undocumented) flag
> BPF_F_NO_PREALLOC, the kernel now actually pre-allocates the hash
> table elements. Hash table elements can be freed and reused for new
> allocations (!) without waiting for an RCU grace period: Freed
> elements are immediately pushed on the percpu freelist, and can be
> immediately reused from there. The most obvious consequence of this is
> that if a BPF program looks up a hash table entry and then reads the
> value, the value can be replaced with a new value in between. A more
> subtle consequence is that BPF map lookups can return false-positive
> results: If the first half of the lookup key matches the old key, and
> the second half of the lookup key matches the new key, then a BPF map
> lookup can return a false-positive result, as far as I can tell.

Correct. That is the case when one cpu deletes the pre-allocated element
and immediately reuses the slot in update_elem while another cpu is still
looking into it.

> If what I'm saying is correct, I'm not sure what the best fix is.
>
> Add a grace period when freeing hash map entries, and add a new -EBUSY
> return value for attempts to create hash map entries when all free
> entries are waiting for the end of an RCU grace period?
> 
> Add a grace period when freeing hash map entries, and use
> rcu_synchronize() when inserting BPF hashmap entries from userspace
> and all free entries are waiting for RCU? But that still leaves the
> bpf_map_update_elem_proto helper that can be called from BPF.
> Deprecate that helper for access to hash maps?
> 
> Document the race, and advise people who use BPF for
> non-performance-tracing purposes (where occasional false positives
> actually matter) to use BPF_F_NO_PREALLOC?

Definitely our fault that it's not documented clearly.
There are subtle issues with kmalloc style as well.
See free_htab_elem():
  atomic_dec(&htab->count);
  l->htab = htab;
  call_rcu(&l->rcu, htab_elem_free_rcu);
which means that for short time the total memory
consumption of the no_prealloc hash map can be higher
than max_entries limit while rcu callbacks are working their way through.
For bpf prog that is invoked million times a second and constantly
doing delete/update the rcu callback latency can be an issue.
Hence the use of no_prealloc needs to be carefully considered as well.

we tried to atomic_dec in rcu callback, but that caused even more
issues, since already written bpf progs expected max_entries limit
to be accurate.

> Add some sort of sequence lock to BPF (yuck)?

why yuck reaction? ;)
Locks are mandatory primitives in concurrent programming and
the lock support is coming to bpf as well.
bpf progs themselves will be able to do lock/unlock of objects
to maintain atomicity of multi field updates.
Right now updating two counters atomically is pretty much impossible.
hash map update can replace the element with a copy, but that's not
usable for counters.
Soon the program will be able to do:
struct value_t {
  u32 var1, var2;
  bpf_lock_t lock;
};
struct value_t * value = bpf_map_lookup_elem(map, &key);
bpf_lock(&value->lock);
value->var1++;
value->var2++;
bpf_unlock(&value->lock);
and verifier will be smart enough to make sure locks don't escape.

Also alloc/free will be available to bpf progs and new program
types won't have to run as one giant rcu_lock+preempt_disable section.
Progs will be able to do rcu_lock/unlock and preempt_enable/disable
from inside the program.

Re: BPF: RCU use-after-reallocation of hash table elements?

[Jann Horn]

On Wed, Oct 3, 2018 at 7:05 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
> On Wed, Oct 03, 2018 at 06:15:08PM +0200, Jann Horn wrote:
> > Hi!
> >
> > Note: I haven't tested any of this; feel free to tell me that I've
> > completely misunderstood how all this works.
> >
> > The BPF manpage, at the moment, states about BPF hash tables:
>
> unfortunately man pages have been obsolete for long time.
> the only doc that is mostly up to date is
> http://docs.cilium.io/en/latest/bpf/

Oh. If you're not updating the manpage, could you at least put a
reference to that other documentation in the manpage?