Scripts need the map type to know how to interact with a map, e.g.
per-CPU value handling or keyless queue/stack semantics. Add an
info() map method returning type, key/value size and max_entries,
along with BPF_MAP_TYPE constants for comparison.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Add consumer support for BPF_MAP_TYPE_RINGBUF and
BPF_MAP_TYPE_PERF_EVENT_ARRAY maps. ringbuf(callback) and
perf_buffer(callback, pages, lost_callback) map methods return a
buffer resource with fileno, poll and consume methods. The fileno
method exposes the epoll fd and matches the interface uloop.handle()
expects, so a buffer can be passed directly to uloop for event loop
integration.
The ring buffer callback receives the sample data and may return a
negative integer to stop consumption; the perf callbacks additionally
receive the CPU number.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Updates on array/hash-of-maps and prog array maps take the fd of the
inner object as a 4-byte value. Accept bpf.map and bpf.program
resources as map values and convert them to their fd.
Lookups on map-in-map types return the inner map id; add open_map_id
and open_program_id to obtain a usable object from such an id.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Queue and stack maps have zero-sized keys and require a NULL key
pointer; the kernel rejects any non-NULL key with EINVAL, making
these maps unusable. Accept a null key argument on maps without a
key, mapping get() to peek, set() to push and delete() with return
to pop.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Lookup, update and lookup-and-delete on per-CPU maps use buffers of
roundup(value_size, 8) * num_possible_cpus bytes. The previous code
sized buffers for a single value, so get() on a per-CPU map made the
kernel overflow the stack buffer and set() passed undersized data.
Track the map type and size the buffers accordingly. get() and
delete() with return now yield an array with one value per CPU, set()
accepts either such an array or a single value replicated to all
CPUs.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Failures in the detach path returned NULL without recording any error
state, so error() reported a stale or empty error. Attach failures
reported a hardcoded ENOENT regardless of the actual cause. Record
errno for all failures; if_nametoindex, bpf_tc_attach and
bpf_tc_detach all set it.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
bpf_object__next_map does not set errno when it runs out of maps, so
the reported error code was whatever errno happened to hold. With
errno 0 the failure was invisible to error(). Report ENOENT instead.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
The iterator key buffer is a flexible array member placed directly
after a bool, leaving it without natural alignment for the u32/u64
loads in next_int. Copy the key via memcpy instead of dereferencing
a misaligned pointer, which traps on strict-alignment targets.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
uc_bpf_map_arg converted integer arguments through a single static
buffer. When map set was called with both an integer key and an integer
value, the second conversion overwrote the first, so the element was
written at the wrong key. Convert into a caller-provided buffer
instead.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
uc_fn_thisval can return NULL when a method is called with a foreign or
missing this context. pin, foreach and the iterator next functions
dereferenced the result without checking it, crashing the VM.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Forward the OpenWrt CPPFLAGS to the compile process. This fixes fortify
sources support.
Link: https://github.com/openwrt/openwrt/pull/22056
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
The bpf plugin provides functionality for loading and interacting with
eBPF modules.
It allows loading full modules and pinned maps/programs and supports
interacting with maps and attaching programs as tc classifiers.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
