Re: [RFC PATCH bpf-next 0/6] bpf: Handle reuse in bpf memory alloc

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Tue, Jan 3, 2023 at 5:40 AM Hou Tao <houtao@huaweicloud.com> wrote:
>
> Hi,
>
> On 1/1/2023 9:26 AM, Alexei Starovoitov wrote:
> > On Fri, Dec 30, 2022 at 12:11:45PM +0800, Hou Tao wrote:
> >> From: Hou Tao <houtao1@huawei.com>
> >>
> >> Hi,
> >>
> >> The patchset tries to fix the problems found when checking how htab map
> >> handles element reuse in bpf memory allocator. The immediate reuse of
> >> freed elements may lead to two problems in htab map:
> >>
> >> (1) reuse will reinitialize special fields (e.g., bpf_spin_lock) in
> >>     htab map value and it may corrupt lookup procedure with BFP_F_LOCK
> >>     flag which acquires bpf-spin-lock during value copying. The
> >>     corruption of bpf-spin-lock may result in hard lock-up.
> >> (2) lookup procedure may get incorrect map value if the found element is
> >>     freed and then reused.
> >>
> >> Because the type of htab map elements are the same, so problem #1 can be
> >> fixed by supporting ctor in bpf memory allocator. The ctor initializes
> >> these special fields in map element only when the map element is newly
> >> allocated. If it is just a reused element, there will be no
> >> reinitialization.
> > Instead of adding the overhead of ctor callback let's just
> > add __GFP_ZERO to flags in __alloc().
> > That will address the issue 1 and will make bpf_mem_alloc behave just
> > like percpu_freelist, so hashmap with BPF_F_NO_PREALLOC and default
> > will behave the same way.
> Use __GPF_ZERO will be simpler, but the overhead of memset() for every allocated
> object may be bigger than ctor callback when the size of allocated object is
> large. And it also introduces unnecessary memory zeroing if there is no special
> field in the map value.

Small memset is often faster than an indirect call.
I doubt that adding GFP_ZERO will have a measurable perf difference
in map_perf_test and other benchmarks.

> >> Problem #2 exists for both non-preallocated and preallocated htab map.
> >> By adding seq in htab element, doing reuse check and retrying the
> >> lookup procedure may be a feasible solution, but it will make the
> >> lookup API being hard to use, because the user needs to check whether
> >> the found element is reused or not and repeat the lookup procedure if it
> >> is reused. A simpler solution would be just disabling freed elements
> >> reuse and freeing these elements after lookup procedure ends.
> > You've proposed this 'solution' twice already in qptrie thread and both
> > times the answer was 'no, we cannot do this' with reasons explained.
> > The 3rd time the answer is still the same.
> This time a workable demo which calls call_rcu_task_trace() in batch is provided
> :) Also because I can not find a better solution for the reuse problem. But you
> are right, although don't reuse the freed element will make the implementation
> of map simpler, the potential OOM problem is hard to solve specially when RCU
> tasks trace grace period is slow. Hope Paul can provide some insight about the
> problem.

OOM is exactly the reason why we cannot do this delaying logic
in the general case. We don't control what progs do and rcu tasks trace
may take a long time.

> > This 'issue 2' existed in hashmap since very beginning for many years.
> > It's a known quirk. There is nothing to fix really.
> Do we need to document the unexpected behavior somewhere, because I really don't
> know nothing about the quirk ?

Yeah. It's not documented in Documentation/bpf/map_hash.rst.
Please send a patch to add it.

> >
> > The graph apis (aka new gen data structs) with link list and rbtree are
> > in active development. Soon bpf progs will be able to implement their own
> > hash maps with explicit bpf_rcu_read_lock. At that time the progs will
> > be making the trade off between performance and lookup/delete race.
> It seems these new gen data struct also need to solve the reuse problem because
> a global bpf memory allocator is used.

Currently the graph api is single owner and kptr_xchg is used to allow
single cpu at a time to observe the object.
In the future we will add bpf_refcount and multi owner.
Then multiple cpus will be able to access the same object concurrently.
They will race to read/write the fields and it will be prog decision
to arbitrate the access.
In other words the bpf prog needs to have mechanisms to deal with reuse.