[PATCH v2 bpf-next 7/7] docs/bpf: add BPF ring buffer design notes

[Andrii Nakryiko <andriin@fb.com>]

Add commit description from patch #1 as a stand-alone documentation under
Documentation/bpf, as it might be more convenient format, in long term
perspective.

Suggested-by: Stanislav Fomichev <sdf@google.com>
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
---
 Documentation/bpf/ringbuf.txt | 191 ++++++++++++++++++++++++++++++++++
 1 file changed, 191 insertions(+)
 create mode 100644 Documentation/bpf/ringbuf.txt

diff --git a/Documentation/bpf/ringbuf.txt b/Documentation/bpf/ringbuf.txt
new file mode 100644
index 000000000000..72f0a2480db3
--- /dev/null
+++ b/Documentation/bpf/ringbuf.txt
@@ -0,0 +1,191 @@
+BPF ring buffer
+==============
+
+Motivation
+----------
+There are two distinctive motivators for this work, which are not satisfied by
+existing perf buffer, which prompted creation of a new ring buffer
+implementation.
+  - more efficient memory utilization by sharing ring buffer across CPUs;
+  - preserving ordering of events that happen sequentially in time, even
+  across multiple CPUs (e.g., fork/exec/exit events for a task).
+
+These two problems are independent, but perf buffer fails to satisfy both.
+Both are a result of a choice to have per-CPU perf ring buffer.  Both can be
+also solved by having an MPSC implementation of ring buffer. The ordering
+problem could technically be solved for perf buffer with some in-kernel
+counting, but given the first one requires an MPSC buffer, the same solution
+would solve the second problem automatically.
+
+Semantics and APIs
+------------------
+Single ring buffer is presented to BPF programs as an instance of BPF map of
+type BPF_MAP_TYPE_RINGBUF. Two other alternatives considered, but ultimately
+rejected.
+
+One way would be to, similar to BPF_MAP_TYPE_PERF_EVENT_ARRAY, make
+BPF_MAP_TYPE_RINGBUF could represent an array of ring buffers, but not enforce
+"same CPU only" rule. This would be more familiar interface compatible with
+existing perf buffer use in BPF, but would fail if application needed more
+advanced logic to lookup ring buffer by arbitrary key. HASH_OF_MAPS addresses
+this with current approach. Additionally, given the performance of BPF
+ringbuf, many use cases would just opt into a simple single ring buffer shared
+among all CPUs, for which current approach would be an overkill.
+
+Another approach could introduce a new concept, alongside BPF map, to
+represent generic "container" object, which doesn't necessarily have key/value
+interface with lookup/update/delete operations. This approach would add a lot
+of extra infrastructure that has to be built for observability and verifier
+support. It would also add another concept that BPF developers would have to
+familiarize themselves with, new syntax in libbpf, etc. But then would really
+provide no additional benefits over the approach of using a map.
+BPF_MAP_TYPE_RINGBUF doesn't support lookup/update/delete operations, but so
+doesn't few other map types (e.g., queue and stack; array doesn't support
+delete, etc).
+
+The approach chosen has an advantage of re-using existing BPF map
+infrastructure (introspection APIs in kernel, libbpf support, etc), being
+familiar concept (no need to teach users a new type of object in BPF program),
+and utilizing existing tooling (bpftool). For common scenario of using
+a single ring buffer for all CPUs, it's as simple and straightforward, as
+would be with a dedicated "container" object. On the other hand, by being
+a map, it can be combined with ARRAY_OF_MAPS and HASH_OF_MAPS map-in-maps to
+implement a wide variety of topologies, from one ring buffer for each CPU
+(e.g., as a replacement for perf buffer use cases), to a complicated
+application hashing/sharding of ring buffers (e.g., having a small pool of
+ring buffers with hashed task's tgid being a look up key to preserve order,
+but reduce contention).
+
+Key and value sizes are enforced to be zero. max_entries is used to specify
+the size of ring buffer and has to be a power of 2 value.
+
+There are a bunch of similarities between perf buffer
+(BPF_MAP_TYPE_PERF_EVENT_ARRAY) and new BPF ring buffer semantics:
+  - variable-length records;
+  - if there is no more space left in ring buffer, reservation fails, no
+    blocking;
+  - memory-mappable data area for user-space applications for ease of
+    consumption and high performance;
+  - epoll notifications for new incoming data;
+  - but still the ability to do busy polling for new data to achieve the
+    lowest latency, if necessary.
+
+BPF ringbuf provides two sets of APIs to BPF programs:
+  - bpf_ringbuf_output() allows to *copy* data from one place to a ring
+    buffer, similarly to bpf_perf_event_output();
+  - bpf_ringbuf_reserve()/bpf_ringbuf_commit()/bpf_ringbuf_discard() APIs
+    split the whole process into two steps. First, a fixed amount of space is
+    reserved. If successful, a pointer to a data inside ring buffer data area
+    is returned, which BPF programs can use similarly to a data inside
+    array/hash maps. Once ready, this piece of memory is either committed or
+    discarded. Discard is similar to commit, but makes consumer ignore the
+    record.
+
+bpf_ringbuf_output() has disadvantage of incurring extra memory copy, because
+record has to be prepared in some other place first. But it allows to submit
+records of the length that's not known to verifier beforehand. It also closely
+matches bpf_perf_event_output(), so will simplify migration significantly.
+
+bpf_ringbuf_reserve() avoids the extra copy of memory by providing a memory
+pointer directly to ring buffer memory. In a lot of cases records are larger
+than BPF stack space allows, so many programs have use extra per-CPU array as
+a temporary heap for preparing sample. bpf_ringbuf_reserve() avoid this needs
+completely. But in exchange, it only allows a known constant size of memory to
+be reserved, such that verifier can verify that BPF program can't access
+memory outside its reserved record space. bpf_ringbuf_output(), while slightly
+slower due to extra memory copy, covers some use cases that are not suitable
+for bpf_ringbuf_reserve().
+
+The difference between commit and discard is very small. Discard just marks
+a record as discarded, and such records are supposed to be ignored by consumer
+code. Discard is useful for some advanced use-cases, such as ensuring
+all-or-nothing multi-record submission, or emulating temporary malloc()/free()
+within single BPF program invocation.
+
+Each reserved record is tracked by verifier through existing
+reference-tracking logic, similar to socket ref-tracking. It is thus
+impossible to reserve a record, but forget to submit (or discard) it.
+
+bpf_ringbuf_query() helper allows to query various properties of ring buffer.
+Currently 4 are supported:
+  - BPF_RB_AVAIL_DATA returns amount of unconsumed data in ring buffer;
+  - BPF_RB_RING_SIZE returns the size of ring buffer;
+  - BPF_RB_CONS_POS/BPF_RB_PROD_POS returns current logical possition of
+    consumer/producer, respectively.
+Returned values are momentarily snapshots of ring buffer state and could be
+off by the time helper returns, so this should be used only for
+debugging/reporting reasons or for implementing various heuristics, that take
+into account highly-changeable nature of some of those characteristics.
+
+One such heuristic might involve more fine-grained control over poll/epoll
+notifications about new data availability in ring buffer. Together with
+BPF_RB_NO_WAKEUP/BPF_RB_FORCE_WAKEUP flags for output/commit/discard helpers,
+it allows BPF program a high degree of control and, e.g., more efficient
+batched notifications. Default self-balancing strategy, though, should be
+adequate for most applications and will work reliable and efficiently already.
+
+Design and implementation
+-------------------------
+This reserve/commit schema allows a natural way for multiple producers, either
+on different CPUs or even on the same CPU/in the same BPF program, to reserve
+independent records and work with them without blocking other producers. This
+means that if BPF program was interruped by another BPF program sharing the
+same ring buffer, they will both get a record reserved (provided there is
+enough space left) and can work with it and submit it independently. This
+applies to NMI context as well, except that due to using a spinlock during
+reservation, in NMI context, bpf_ringbuf_reserve() might fail to get a lock,
+in which case reservation will fail even if ring buffer is not full.
+
+The ring buffer itself internally is implemented as a power-of-2 sized
+circular buffer, with two logical and ever-increasing counters (which might
+wrap around on 32-bit architectures, that's not a problem):
+  - consumer counter shows up to which logical position consumer consumed the
+    data;
+  - producer counter denotes amount of data reserved by all producers.
+
+Each time a record is reserved, producer that "owns" the record will
+successfully advance producer counter. At that point, data is still not yet
+ready to be consumed, though. Each record has 8 byte header, which contains
+the length of reserved record, as well as two extra bits: busy bit to denote
+that record is still being worked on, and discard bit, which might be set at
+commit time if record is discarded. In the latter case, consumer is supposed
+to skip the record and move on to the next one. Record header also encodes
+record's relative offset from the beginning of ring buffer data area (in
+pages). This allows bpf_ringbuf_commit()/bpf_ringbuf_discard() to accept only
+the pointer to the record itself, without requiring also the pointer to ring
+buffer itself. Ring buffer memory location will be restored from record
+metadata header. This significantly simplifies verifier, as well as improving
+API usability.
+
+Producer counter increments are serialized under spinlock, so there is
+a strict ordering between reservations. Commits, on the other hand, are
+completely lockless and independent. All records become available to consumer
+in the order of reservations, but only after all previous records where
+already committed. It is thus possible for slow producers to temporarily hold
+off submitted records, that were reserved later.
+
+Reservation/commit/consumer protocol is verified by litmus tests in
+tools/memory-model/litmus-tests, see mpsc-rb*.litmus files.
+
+One interesting implementation bit, that significantly simplifies (and thus
+speeds up as well) implementation of both producers and consumers is how data
+area is mapped twice contiguously back-to-back in the virtual memory. This
+allows to not take any special measures for samples that have to wrap around
+at the end of the circular buffer data area, because the next page after the
+last data page would be first data page again, and thus the sample will still
+appear completely contiguous in virtual memory. See comment and a simple ASCII
+diagram showing this visually in bpf_ringbuf_area_alloc().
+
+Another feature that distinguishes BPF ringbuf from perf ring buffer is
+a self-pacing notifications of new data being availability.
+bpf_ringbuf_commit() implementation will send a notification of new record
+being available after commit only if consumer has already caught up right up
+to the record being committed. If not, consumer still has to catch up and thus
+will see new data anyways without needing an extra poll notification.
+Benchmarks (see tools/testing/selftests/bpf/benchs/bench_ringbuf.c) show that
+this allows to achieve a very high throughput without having to resort to
+tricks like "notify only every Nth sample", which are necessary with perf
+buffer. For extreme cases, when BPF program wants more manual control of
+notifications, commit/discard/output helpers accept BPF_RB_NO_WAKEUP and
+BPF_RB_FORCE_WAKEUP flags, which give full control over notifications of data
+availability, but require extra caution and diligence in using this API.
-- 
2.24.1