|
This commits updates the following functions to allow reads from
uninitialized stack locations when env->allow_uninit_stack option is
enabled:
- check_stack_read_fixed_off()
- check_stack_range_initialized(), called from:
- check_stack_read_var_off()
- check_helper_mem_access()
Such change allows to relax logic in stacksafe() to treat STACK_MISC
and STACK_INVALID in a same way and make the following stack slot
configurations equivalent:
| Cached state | Current state |
| stack slot | stack slot |
|------------------+------------------|
| STACK_INVALID or | STACK_INVALID or |
| STACK_MISC | STACK_SPILL or |
| | STACK_MISC or |
| | STACK_ZERO or |
| | STACK_DYNPTR |
This leads to significant verification speed gains (see below).
The idea was suggested by Andrii Nakryiko [1] and initial patch was
created by Alexei Starovoitov [2].
Currently the env->allow_uninit_stack is allowed for programs loaded
by users with CAP_PERFMON or CAP_SYS_ADMIN capabilities.
A number of test cases from verifier/*.c were expecting uninitialized
stack access to be an error. These test cases were updated to execute
in unprivileged mode (thus preserving the tests).
The test progs/test_global_func10.c expected "invalid indirect read
from stack" error message because of the access to uninitialized
memory region. This error is no longer possible in privileged mode.
The test is updated to provoke an error "invalid indirect access to
stack" because of access to invalid stack address (such error is not
verified by progs/test_global_func*.c series of tests).
The following tests had to be removed because these can't be made
unprivileged:
- verifier/sock.c:
- "sk_storage_get(map, skb->sk, &stack_value, 1): partially init
stack_value"
BPF_PROG_TYPE_SCHED_CLS programs are not executed in unprivileged mode.
- verifier/var_off.c:
- "indirect variable-offset stack access, max_off+size > max_initialized"
- "indirect variable-offset stack access, uninitialized"
These tests verify that access to uninitialized stack values is
detected when stack offset is not a constant. However, variable
stack access is prohibited in unprivileged mode, thus these tests
are no longer valid.
* * *
Here is veristat log comparing this patch with current master on a
set of selftest binaries listed in tools/testing/selftests/bpf/veristat.cfg
and cilium BPF binaries (see [3]):
$ ./veristat -e file,prog,states -C -f 'states_pct<-30' master.log current.log
File Program States (A) States (B) States (DIFF)
-------------------------- -------------------------- ---------- ---------- ----------------
bpf_host.o tail_handle_ipv6_from_host 349 244 -105 (-30.09%)
bpf_host.o tail_handle_nat_fwd_ipv4 1320 895 -425 (-32.20%)
bpf_lxc.o tail_handle_nat_fwd_ipv4 1320 895 -425 (-32.20%)
bpf_sock.o cil_sock4_connect 70 48 -22 (-31.43%)
bpf_sock.o cil_sock4_sendmsg 68 46 -22 (-32.35%)
bpf_xdp.o tail_handle_nat_fwd_ipv4 1554 803 -751 (-48.33%)
bpf_xdp.o tail_lb_ipv4 6457 2473 -3984 (-61.70%)
bpf_xdp.o tail_lb_ipv6 7249 3908 -3341 (-46.09%)
pyperf600_bpf_loop.bpf.o on_event 287 145 -142 (-49.48%)
strobemeta.bpf.o on_event 15915 4772 -11143 (-70.02%)
strobemeta_nounroll2.bpf.o on_event 17087 3820 -13267 (-77.64%)
xdp_synproxy_kern.bpf.o syncookie_tc 21271 6635 -14636 (-68.81%)
xdp_synproxy_kern.bpf.o syncookie_xdp 23122 6024 -17098 (-73.95%)
-------------------------- -------------------------- ---------- ---------- ----------------
Note: I limited selection by states_pct<-30%.
Inspection of differences in pyperf600_bpf_loop behavior shows that
the following patch for the test removes almost all differences:
- a/tools/testing/selftests/bpf/progs/pyperf.h
+ b/tools/testing/selftests/bpf/progs/pyperf.h
@ -266,8 +266,8 @ int __on_event(struct bpf_raw_tracepoint_args *ctx)
}
if (event->pthread_match || !pidData->use_tls) {
- void* frame_ptr;
- FrameData frame;
+ void* frame_ptr = 0;
+ FrameData frame = {};
Symbol sym = {};
int cur_cpu = bpf_get_smp_processor_id();
W/o this patch the difference comes from the following pattern
(for different variables):
static bool get_frame_data(... FrameData *frame ...)
{
...
bpf_probe_read_user(&frame->f_code, ...);
if (!frame->f_code)
return false;
...
bpf_probe_read_user(&frame->co_name, ...);
if (frame->co_name)
...;
}
int __on_event(struct bpf_raw_tracepoint_args *ctx)
{
FrameData frame;
...
get_frame_data(... &frame ...) // indirectly via a bpf_loop & callback
...
}
SEC("raw_tracepoint/kfree_skb")
int on_event(struct bpf_raw_tracepoint_args* ctx)
{
...
ret |= __on_event(ctx);
ret |= __on_event(ctx);
...
}
With regards to value `frame->co_name` the following is important:
- Because of the conditional `if (!frame->f_code)` each call to
__on_event() produces two states, one with `frame->co_name` marked
as STACK_MISC, another with it as is (and marked STACK_INVALID on a
first call).
- The call to bpf_probe_read_user() does not mark stack slots
corresponding to `&frame->co_name` as REG_LIVE_WRITTEN but it marks
these slots as BPF_MISC, this happens because of the following loop
in the check_helper_call():
for (i = 0; i < meta.access_size; i++) {
err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
BPF_WRITE, -1, false);
if (err)
return err;
}
Note the size of the write, it is a one byte write for each byte
touched by a helper. The BPF_B write does not lead to write marks
for the target stack slot.
- Which means that w/o this patch when second __on_event() call is
verified `if (frame->co_name)` will propagate read marks first to a
stack slot with STACK_MISC marks and second to a stack slot with
STACK_INVALID marks and these states would be considered different.
[1] https://lore.kernel.org/bpf/CAEf4BzY3e+ZuC6HUa8dCiUovQRg2SzEk7M-dSkqNZyn=xEmnPA@mail.gmail.com/
[2] https://lore.kernel.org/bpf/CAADnVQKs2i1iuZ5SUGuJtxWVfGYR9kDgYKhq3rNV+kBLQCu7rA@mail.gmail.com/
[3] git@github.com:anakryiko/cilium.git
Suggested-by: Andrii Nakryiko <andrii@kernel.org>
Co-developed-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20230219200427.606541-2-eddyz87@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
|
|
This commit adds a new MPSC ring buffer implementation into BPF ecosystem,
which allows multiple CPUs to submit data to a single shared ring buffer. On
the consumption side, only single consumer is assumed.
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
Documentation/litmus-test/bpf-rb.
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.
Comparison to alternatives
--------------------------
Before considering implementing BPF ring buffer from scratch existing
alternatives in kernel were evaluated, but didn't seem to meet the needs. They
largely fell into few categores:
- per-CPU buffers (perf, ftrace, etc), which don't satisfy two motivations
outlined above (ordering and memory consumption);
- linked list-based implementations; while some were multi-producer designs,
consuming these from user-space would be very complicated and most
probably not performant; memory-mapping contiguous piece of memory is
simpler and more performant for user-space consumers;
- io_uring is SPSC, but also requires fixed-sized elements. Naively turning
SPSC queue into MPSC w/ lock would have subpar performance compared to
locked reserve + lockless commit, as with BPF ring buffer. Fixed sized
elements would be too limiting for BPF programs, given existing BPF
programs heavily rely on variable-sized perf buffer already;
- specialized implementations (like a new printk ring buffer, [0]) with lots
of printk-specific limitations and implications, that didn't seem to fit
well for intended use with BPF programs.
[0] https://lwn.net/Articles/779550/
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20200529075424.3139988-2-andriin@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
|