2BPF ring buffer
   5This document describes BPF ring buffer design, API, and implementation details.
   7.. contents::
   8    :local:
   9    :depth: 2
  14There are two distinctive motivators for this work, which are not satisfied by
  15existing perf buffer, which prompted creation of a new ring buffer
  18- more efficient memory utilization by sharing ring buffer across CPUs;
  19- preserving ordering of events that happen sequentially in time, even across
  20  multiple CPUs (e.g., fork/exec/exit events for a task).
  22These two problems are independent, but perf buffer fails to satisfy both.
  23Both are a result of a choice to have per-CPU perf ring buffer.  Both can be
  24also solved by having an MPSC implementation of ring buffer. The ordering
  25problem could technically be solved for perf buffer with some in-kernel
  26counting, but given the first one requires an MPSC buffer, the same solution
  27would solve the second problem automatically.
  29Semantics and APIs
  32Single ring buffer is presented to BPF programs as an instance of BPF map of
  33type ``BPF_MAP_TYPE_RINGBUF``. Two other alternatives considered, but
  34ultimately rejected.
  36One way would be to, similar to ``BPF_MAP_TYPE_PERF_EVENT_ARRAY``, make
  37``BPF_MAP_TYPE_RINGBUF`` could represent an array of ring buffers, but not
  38enforce "same CPU only" rule. This would be more familiar interface compatible
  39with existing perf buffer use in BPF, but would fail if application needed more
  40advanced logic to lookup ring buffer by arbitrary key.
  41``BPF_MAP_TYPE_HASH_OF_MAPS`` addresses this with current approach.
  42Additionally, given the performance of BPF ringbuf, many use cases would just
  43opt into a simple single ring buffer shared among all CPUs, for which current
  44approach would be an overkill.
  46Another approach could introduce a new concept, alongside BPF map, to represent
  47generic "container" object, which doesn't necessarily have key/value interface
  48with lookup/update/delete operations. This approach would add a lot of extra
  49infrastructure that has to be built for observability and verifier support. It
  50would also add another concept that BPF developers would have to familiarize
  51themselves with, new syntax in libbpf, etc. But then would really provide no
  52additional benefits over the approach of using a map.  ``BPF_MAP_TYPE_RINGBUF``
  53doesn't support lookup/update/delete operations, but so doesn't few other map
  54types (e.g., queue and stack; array doesn't support delete, etc).
  56The approach chosen has an advantage of re-using existing BPF map
  57infrastructure (introspection APIs in kernel, libbpf support, etc), being
  58familiar concept (no need to teach users a new type of object in BPF program),
  59and utilizing existing tooling (bpftool). For common scenario of using a single
  60ring buffer for all CPUs, it's as simple and straightforward, as would be with
  61a dedicated "container" object. On the other hand, by being a map, it can be
  62combined with ``ARRAY_OF_MAPS`` and ``HASH_OF_MAPS`` map-in-maps to implement
  63a wide variety of topologies, from one ring buffer for each CPU (e.g., as
  64a replacement for perf buffer use cases), to a complicated application
  65hashing/sharding of ring buffers (e.g., having a small pool of ring buffers
  66with hashed task's tgid being a look up key to preserve order, but reduce
  69Key and value sizes are enforced to be zero. ``max_entries`` is used to specify
  70the size of ring buffer and has to be a power of 2 value.
  72There are a bunch of similarities between perf buffer
  73(``BPF_MAP_TYPE_PERF_EVENT_ARRAY``) and new BPF ring buffer semantics:
  75- variable-length records;
  76- if there is no more space left in ring buffer, reservation fails, no
  77  blocking;
  78- memory-mappable data area for user-space applications for ease of
  79  consumption and high performance;
  80- epoll notifications for new incoming data;
  81- but still the ability to do busy polling for new data to achieve the
  82  lowest latency, if necessary.
  84BPF ringbuf provides two sets of APIs to BPF programs:
  86- ``bpf_ringbuf_output()`` allows to *copy* data from one place to a ring
  87  buffer, similarly to ``bpf_perf_event_output()``;
  88- ``bpf_ringbuf_reserve()``/``bpf_ringbuf_commit()``/``bpf_ringbuf_discard()``
  89  APIs split the whole process into two steps. First, a fixed amount of space
  90  is reserved. If successful, a pointer to a data inside ring buffer data
  91  area is returned, which BPF programs can use similarly to a data inside
  92  array/hash maps. Once ready, this piece of memory is either committed or
  93  discarded. Discard is similar to commit, but makes consumer ignore the
  94  record.
  96``bpf_ringbuf_output()`` has disadvantage of incurring extra memory copy,
  97because record has to be prepared in some other place first. But it allows to
  98submit records of the length that's not known to verifier beforehand. It also
  99closely matches ``bpf_perf_event_output()``, so will simplify migration
 102``bpf_ringbuf_reserve()`` avoids the extra copy of memory by providing a memory
 103pointer directly to ring buffer memory. In a lot of cases records are larger
 104than BPF stack space allows, so many programs have use extra per-CPU array as
 105a temporary heap for preparing sample. bpf_ringbuf_reserve() avoid this needs
 106completely. But in exchange, it only allows a known constant size of memory to
 107be reserved, such that verifier can verify that BPF program can't access memory
 108outside its reserved record space. bpf_ringbuf_output(), while slightly slower
 109due to extra memory copy, covers some use cases that are not suitable for
 112The difference between commit and discard is very small. Discard just marks
 113a record as discarded, and such records are supposed to be ignored by consumer
 114code. Discard is useful for some advanced use-cases, such as ensuring
 115all-or-nothing multi-record submission, or emulating temporary
 116``malloc()``/``free()`` within single BPF program invocation.
 118Each reserved record is tracked by verifier through existing
 119reference-tracking logic, similar to socket ref-tracking. It is thus
 120impossible to reserve a record, but forget to submit (or discard) it.
 122``bpf_ringbuf_query()`` helper allows to query various properties of ring
 123buffer.  Currently 4 are supported:
 125- ``BPF_RB_AVAIL_DATA`` returns amount of unconsumed data in ring buffer;
 126- ``BPF_RB_RING_SIZE`` returns the size of ring buffer;
 127- ``BPF_RB_CONS_POS``/``BPF_RB_PROD_POS`` returns current logical possition
 128  of consumer/producer, respectively.
 130Returned values are momentarily snapshots of ring buffer state and could be
 131off by the time helper returns, so this should be used only for
 132debugging/reporting reasons or for implementing various heuristics, that take
 133into account highly-changeable nature of some of those characteristics.
 135One such heuristic might involve more fine-grained control over poll/epoll
 136notifications about new data availability in ring buffer. Together with
 137``BPF_RB_NO_WAKEUP``/``BPF_RB_FORCE_WAKEUP`` flags for output/commit/discard
 138helpers, it allows BPF program a high degree of control and, e.g., more
 139efficient batched notifications. Default self-balancing strategy, though,
 140should be adequate for most applications and will work reliable and efficiently
 143Design and Implementation
 146This reserve/commit schema allows a natural way for multiple producers, either
 147on different CPUs or even on the same CPU/in the same BPF program, to reserve
 148independent records and work with them without blocking other producers. This
 149means that if BPF program was interruped by another BPF program sharing the
 150same ring buffer, they will both get a record reserved (provided there is
 151enough space left) and can work with it and submit it independently. This
 152applies to NMI context as well, except that due to using a spinlock during
 153reservation, in NMI context, ``bpf_ringbuf_reserve()`` might fail to get
 154a lock, in which case reservation will fail even if ring buffer is not full.
 156The ring buffer itself internally is implemented as a power-of-2 sized
 157circular buffer, with two logical and ever-increasing counters (which might
 158wrap around on 32-bit architectures, that's not a problem):
 160- consumer counter shows up to which logical position consumer consumed the
 161  data;
 162- producer counter denotes amount of data reserved by all producers.
 164Each time a record is reserved, producer that "owns" the record will
 165successfully advance producer counter. At that point, data is still not yet
 166ready to be consumed, though. Each record has 8 byte header, which contains the
 167length of reserved record, as well as two extra bits: busy bit to denote that
 168record is still being worked on, and discard bit, which might be set at commit
 169time if record is discarded. In the latter case, consumer is supposed to skip
 170the record and move on to the next one. Record header also encodes record's
 171relative offset from the beginning of ring buffer data area (in pages). This
 172allows ``bpf_ringbuf_commit()``/``bpf_ringbuf_discard()`` to accept only the
 173pointer to the record itself, without requiring also the pointer to ring buffer
 174itself. Ring buffer memory location will be restored from record metadata
 175header. This significantly simplifies verifier, as well as improving API
 178Producer counter increments are serialized under spinlock, so there is
 179a strict ordering between reservations. Commits, on the other hand, are
 180completely lockless and independent. All records become available to consumer
 181in the order of reservations, but only after all previous records where
 182already committed. It is thus possible for slow producers to temporarily hold
 183off submitted records, that were reserved later.
 185One interesting implementation bit, that significantly simplifies (and thus
 186speeds up as well) implementation of both producers and consumers is how data
 187area is mapped twice contiguously back-to-back in the virtual memory. This
 188allows to not take any special measures for samples that have to wrap around
 189at the end of the circular buffer data area, because the next page after the
 190last data page would be first data page again, and thus the sample will still
 191appear completely contiguous in virtual memory. See comment and a simple ASCII
 192diagram showing this visually in ``bpf_ringbuf_area_alloc()``.
 194Another feature that distinguishes BPF ringbuf from perf ring buffer is
 195a self-pacing notifications of new data being availability.
 196``bpf_ringbuf_commit()`` implementation will send a notification of new record
 197being available after commit only if consumer has already caught up right up to
 198the record being committed. If not, consumer still has to catch up and thus
 199will see new data anyways without needing an extra poll notification.
 200Benchmarks (see tools/testing/selftests/bpf/benchs/bench_ringbufs.c) show that
 201this allows to achieve a very high throughput without having to resort to
 202tricks like "notify only every Nth sample", which are necessary with perf
 203buffer. For extreme cases, when BPF program wants more manual control of
 204notifications, commit/discard/output helpers accept ``BPF_RB_NO_WAKEUP`` and
 205``BPF_RB_FORCE_WAKEUP`` flags, which give full control over notifications of
 206data availability, but require extra caution and diligence in using this API.