The scheduling logic for sched_ext schedulers is written in eBPF (BPF). For high level documentation the kernel docs should be referenced.
When working on schedulers the following documentation is rather useful as schedulers will use a combination of BPF cpumasks, helper functions, kfuncs and maps for scheduling logic.
The kernel BPF tests are also a useful source of examples of BPF functionality.
The kernel scheduling docs provide a high level overview of the existing scheduler subsystem. The kernel docs cover various topics such as deadline scheduling, realtime scheduling and the interaction of schedulers with other system resources.
When schedulers are written to scale beyond more than a single core eventually
the scheduler needs to implement a load balancing algorithm. Calculating the
load between scheduling domains becomes a difficult problem. sched_ext has a
common crate for calculating weights between scheduling domains. See the
infeasible crate in rust/scx_utils/src for the implementation.
We use cargo fmt to ensure consistency in our Rust code. This runs on PRs in
the CI and will fail with a patch if your code doesn't match. We currently need
a nightly version of Rust to format so have pinned this for consistency. To run
locally (with rustup) run:
$ rustup install nightly-2024-09-10
$ cargo +nightly-2024-09-10 fmt
Perfetto is a profiling and trace visualization
platform. It can be used to view scheduling data, which is useful for
understanding scheduling decisions. The sched_ftrace.py
script can be used to generate a ftrace compatible with Perfetto.
$ sudo ./scripts/sched_ftrace.py > sched.ftrace
The output of the script can then be loaded into the perfetto UI:
The linux perf tool has a subcommand for profiling scheduling perf sched.
The interface is text driven, but is able to provide various timeline views and
aggregations of scheduler events. The following is an example of using perf sched to get a timeline histogram with additional scheduling metrics.
$ perf sched record
$ perf sched timehist -Vw --state
time cpu 0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef0 task name wait time sch delay run time state
[tid/pid] (msec) (msec) (msec)
--------------- ------ --------------------------------------------------------------------------------- ------------------------------ --------- --------- --------- -----
960264.500659 [0000] perf[1635250] awakened: migration/0[19]
960264.500680 [0000] s perf[1635250] 0.000 0.000 0.000 D
960264.500683 [0000] migration/0[19] awakened: perf[1635250]
960264.500809 [0001] perf[1635250] awakened: migration/1[24]
960264.500814 [0001] s perf[1635250] 0.000 0.000 0.000 D
960264.500816 [0001] migration/1[24] awakened: perf[1635250]
960264.500824 [0001] s migration/1[24] 0.000 0.005 0.009 S
960264.502403 [0001] i <idle> 0.000 0.000 1.579 I
960264.502418 [0001] s HTTPSrvExec39[3403538/3403436] 0.000 0.000 0.014 S
960264.506002 [0001] i <idle> 0.014 0.000 3.583 I
960264.506045 [0001] s CfgrIO0[13302/13094] 0.000 0.000 0.043 S
960264.506763 [0001] swapper awakened: chef-client[1629157]
960264.506767 [0001] i <idle> 0.043 0.000 0.721 I
960264.506784 [0001] s chef-client[1629157] 0.000 0.003 0.017 S
960264.507622 [0001] i <idle> 0.017 0.000 0.837 I
960264.507806 [0001] mcrcfg-fci[1635235/1635080] awakened: GlobalCPUThread[1635186/1635080
960264.507937 [0001] mcrcfg-fci[1635235/1635080] awakened: FalconClientThr[1635187/1635080
960264.507996 [0001] mcrcfg-fci[1635235/1635080] awakened: CfgrIO0[1635185/1635080]
960264.508007 [0001] s mcrcfg-fci[1635235/1635080] 0.000 0.000 0.384 S
960264.508079 [0001] i <idle> 0.384 0.000 0.071 I
960264.508100 [0001] ThriftSrv.N2104[1635036/2683498 awakened: IOThreadPool0[2685229/2683498]
960264.508108 [0001] s ThriftSrv.N2104[1635036/2683498 0.000 0.000 0.029 S
960264.508638 [0001] i <idle> 0.029 0.000 0.529 I
960264.508655 [0001] ThriftSrv.N2104[1635036/2683498 awakened: ThriftIO70[2683693/2683498]
bpftool contains many utilities for
interacting with the BPF subsystem and BPF programs. If you need to know
what BPF programs, maps, iterators are loaded on a system bpftool will
provide all this information.
Listing BPF maps:
$ sudo bpftool map list
11: hash_of_maps name cgroup_hash flags 0x0
key 8B value 4B max_entries 2048 memlock 172992B
pids systemd(1)
Listing struct_ops:
$ sudo bpftool struct_ops list
21381: layered sched_ext_ops
retsnoop is a BPF tool for tracing
linux. It is very useful if you are trying to understand the flow of kernel
functions. This can be useful when BPF verification issues are encountered. The
following example shows how the verifier do_check_common function can be
traced.
$ sudo retsnoop -e 'do_check*' -a ':kernel/bpf/*.c' -T
07:55:28.049718 -> 07:55:28.049797 TID/PID 270611/270611 (bpftool/bpftool):
FUNCTION CALL TRACE RESULT DURATION
--------------------------------- --------- --------
→ do_check_common
→ init_func_state
↔ tnum_const [0] 2.084us
← init_func_state [void] 6.648us
↔ tnum_const [0] 2.662us
→ do_check
↔ mark_reg_unknown [void] 2.251us
↔ tnum_const [0] 2.421us
↔ reg_bounds_sanity_check [0] 2.049us
↔ check_reference_leak [0] 2.014us
→ check_return_code
↔ mark_reg_read [0] 2.212us
← check_return_code [0] 6.531us
↔ pop_stack [-ENOENT] 2.099us
← do_check [0] 34.822us
↔ pop_stack [-ENOENT] 2.167us
← do_check_common [0] 76.413us
entry_SYSCALL_64_after_hwframe+0x4b (entry_SYSCALL_64 @ arch/x86/entry/entry_64.S:130:0)
do_syscall_64+0x6a (arch/x86/entry/common.c:0:0)
__x64_sys_bpf+0x18 (kernel/bpf/syscall.c:5792:1)
. __se_sys_bpf (kernel/bpf/syscall.c:5792:1)
. __do_sys_bpf (kernel/bpf/syscall.c:5794:9)
__sys_bpf+0x27e (kernel/bpf/syscall.c:0:9)
bpf_prog_load+0x593 (kernel/bpf/syscall.c:2908:6)
bpf_check+0x1066 (kernel/bpf/verifier.c:21608:8)
. do_check_main (kernel/bpf/verifier.c:20938:8)
76us [0] do_check_common+0x552 (kernel/bpf/verifier.c:20856:9)
! 2us [-ENOENT] pop_stack
bpftrace is a high level tracing
language for BPF. When working with sched_ext bpftrace programs can be used
for understanding scheduler run queue latency as other scheduler internals. See
the scripts dir for examples.
For generating synthetic load on a system
stress-ng can be used.
stress-ng can generate different types of load on the system including cpu
bound, fork heavy, NUMA, cache heavy and more.
veristat is a tool to provide statics
from the BPF verifier for BPF programs. It can also be used to compare
verification stats across runs. This is useful when trying to optimize BPF
programs for their instruction count.
turbostat
is a tool for inspecting CPU frequency as well as power utilization. When
optimizing schedulers for energy performance turbostat can be used to
understand the energy required per operation.