acc333c40b
When GTS is running in a container runtime which has configured CPU or memory limits or under an init system that uses cgroups to impose CPU and memory limits the values the Go runtime sees for GOMAXPROCS and GOMEMLIMIT are still based on the host resources, not the cgroup. At least for the throttling middlewares which use GOMAXPROCS to configure their queue size, this can result in GTS running with values too big compared to the resources that will actuall be available to it. This introduces 2 dependencies which can pick up resource contraints from the current cgroup and tune the Go runtime accordingly. This should result in the different queues being appropriately sized and in general more predictable performance. These dependencies are a no-op on non-Linux systems or if running in a cgroup that doesn't set a limit on CPU or memory. The automatic tuning of GOMEMLIMIT can be disabled by either explicitly setting GOMEMLIMIT yourself or by setting AUTOMEMLIMIT=off. The automatic tuning of GOMAXPROCS can similarly be counteracted by setting GOMAXPROCS yourself.
3.5 KiB
Architecture of the library
ELF -> Specifications -> Objects -> Links
ELF
BPF is usually produced by using Clang to compile a subset of C. Clang outputs an ELF file which contains program byte code (aka BPF), but also metadata for maps used by the program. The metadata follows the conventions set by libbpf shipped with the kernel. Certain ELF sections have special meaning and contain structures defined by libbpf. Newer versions of clang emit additional metadata in BPF Type Format (aka BTF).
The library aims to be compatible with libbpf so that moving from a C toolchain to a Go one creates little friction. To that end, the ELF reader is tested against the Linux selftests and avoids introducing custom behaviour if possible.
The output of the ELF reader is a CollectionSpec which encodes
all of the information contained in the ELF in a form that is easy to work with
in Go.
BTF
The BPF Type Format describes more than just the types used by a BPF program. It includes debug aids like which source line corresponds to which instructions and what global variables are used.
BTF parsing lives in a separate internal package since exposing
it would mean an additional maintenance burden, and because the API still
has sharp corners. The most important concept is the btf.Type interface, which
also describes things that aren't really types like .rodata or .bss sections.
btf.Types can form cyclical graphs, which can easily lead to infinite loops if
one is not careful. Hopefully a safe pattern to work with btf.Type emerges as
we write more code that deals with it.
Specifications
CollectionSpec, ProgramSpec and MapSpec are blueprints for in-kernel
objects and contain everything necessary to execute the relevant bpf(2)
syscalls. Since the ELF reader outputs a CollectionSpec it's possible to
modify clang-compiled BPF code, for example to rewrite constants. At the same
time the asm package provides an assembler that can be used to generate
ProgramSpec on the fly.
Creating a spec should never require any privileges or be restricted in any way, for example by only allowing programs in native endianness. This ensures that the library stays flexible.
Objects
Program and Map are the result of loading specs into the kernel. Sometimes
loading a spec will fail because the kernel is too old, or a feature is not
enabled. There are multiple ways the library deals with that:
-
Fallback: older kernels don't allowing naming programs and maps. The library automatically detects support for names, and omits them during load if necessary. This works since name is primarily a debug aid.
-
Sentinel error: sometimes it's possible to detect that a feature isn't available. In that case the library will return an error wrapping
ErrNotSupported. This is also useful to skip tests that can't run on the current kernel.
Once program and map objects are loaded they expose the kernel's low-level API,
e.g. NextKey. Often this API is awkward to use in Go, so there are safer
wrappers on top of the low-level API, like MapIterator. The low-level API is
useful as an out when our higher-level API doesn't support a particular use case.
Links
BPF can be attached to many different points in the kernel and newer BPF hooks tend to use bpf_link to do so. Older hooks unfortunately use a combination of syscalls, netlink messages, etc. Adding support for a new link type should not pull in large dependencies like netlink, so XDP programs or tracepoints are out of scope.