io_uring

Very quietly I take my leave,
As quietly as I came here;
Gently waving my sleeve,
I am not taking away a single cloud. - Xu Zhimo

---

syscall

1. man syscall or know more: linux-insides SysCall , linux-insides-zh SysCall

2. syscall interface (liburing interface wrapper syscall for cpu arch, man liburing help doc learn more):
   

/*
 * System calls
 */
extern int io_uring_setup(unsigned entries, struct io_uring_params *p);
extern int io_uring_enter(int fd, unsigned to_submit,
	unsigned min_complete, unsigned flags, sigset_t *sig);
extern int io_uring_register(int fd, unsigned int opcode, void *arg,
	unsigned int nr_args);

3. linux kernel v6.3-rc2 interface ( syscall )

   • io_uring_setup -> syscall --(soft interrupt)-->> sys_io_uring_setup -> SYSCALL_DEFINE2() > sysreturn io_uring_setup
   • io_uring_enter -> syscall --(soft interrupt)-->> sys_io_uring_enter -> SYSCALL_DEFINE6(sysreturn io_uring_enter)
   • io_uring_register -> syscall --(soft interrupt)-->> sys_io_uring_register -> SYSCALL_DEFINE4(io_uring_register)

4. see more support io_uring_op

5. io_uring kernel bug (fixed: io_put_kbuf):

   • IORING_OP_PROVIDE_BUFFERS
   • http://www.aqwu.net/blog/index.php/2022/03/17/io_uring-linux/
   • https://starlabs.sg/blog/2022/06-io_uring-new-code-new-bugs-and-a-new-exploit-technique/

liburing

1. use liburing see readme, more detail: https://github.com/axboe/liburing
   use cgo Notice: https://dave.cheney.net/2016/01/18/cgo-is-not-go
   io_uring want learn more see those [updating]:
   • https://kernel.dk/io_uring.pdf
   • https://www.youtube.com/watch?v=-5T4Cjw46ys
   • https://kernel.dk/axboe-kr2022.pdf
   • https://kernel-recipes.org/en/2022/whats-new-with-io_uring/

# see linux os kernel support uring kernel symbols
cat /proc/kallsyms | grep uring
# see syscall table idx no
ausyscall --dump
#sudo bpftrace -l | grep io_uring
#sudo perf list 'io_uring:*'
#sudo perf stat -e io_uring:io_uring_submit_sqe -- timeout 1 {process}
#sudo perf stat -e io_uring:io_uring_submit_sqe --timeout 1000 -p {pid}
#sudo perf stat -a -d -d -d --timeout 1000 -p {pid}

# when use io_uring check iou-***-*** kernel thread {wrk(work), sqp(sq-poll)} or old version {io_uring-sq(sq-poll)}
sudo bpftrace --btf -e 'kretprobe:create_io_thread { @[retval] = count(); } interval:s:1 { print(@); clear(@); } END { clear(@); }' | cat -s
#sudo bpftrace -e 'tracepoint:io_uring:io_uring_submit_sqe {printf("%s(%d)\n", comm, pid);}'

2. if need use golang runtime native support, please see this Note: #31908

3. 3rd io_uring support for golang:

   • https://github.com/hodgesds/iouring-go (old, no maintenance)
   • https://github.com/godzie44/go-uring (use multi iouring impl golang net listener, acceptor,conn interface)
   • https://github.com/Iceber/iouring-go (use epoll add eventfd IN event for io_uring peek cqe, depend epoll)
   • https://github.com/ii64/gouring [√] (one to one liburing cp,test coverage ok,use userData need notice GC)

4. RocksDB MultiGet use IO Uring interface: https://github.com/facebook/rocksdb/wiki/MultiGet-Performance

bench scene (net/storage IO)

1. net IO for netpoll scenes (more unbounded work stream requests, S_IFSOCK type fd)
   • tcp echo server, build & bench:

   make build-echo
   make bench-echo

2. storage IO for data file storage in HDD, NVMe SSD etc hardware scenes (more bounded work requests, eg: S_IFREG, S_ISBLK type fd)

learn more try to change IO

• net IO
  1. redis: redis/redis#9440
• storage IO
  1. badger: https://dgraph.io/blog/post/badger/
  2. pebble: https://www.cockroachlabs.com/blog/pebble-rocksdb-kv-store/

linux kernel for new io_uring feature

compiling linux kernel for develop io_uring-** tag branch

1. kernel_compile
2. the linux kernel archives

upgrade release linux kernel for ubuntu

https://sypalo.com/how-to-upgrade-ubuntu eg: upgrage linux kernel to v6.3-rc2 for Ubuntu 22.04.2 LTS

sudo apt-get upgrade
sudo apt-get update
# mainline: https://kernel.ubuntu.com/~kernel-ppa/mainline/?C=N;O=D ; u can chose latest linux kernel
sudo wget -c https://kernel.ubuntu.com/~kernel-ppa/mainline/v6.3-rc2/amd64/linux-headers-6.3.0-060300rc2-generic_6.3.0-060300rc2.202303122031_amd64.deb
sudo wget -c https://kernel.ubuntu.com/~kernel-ppa/mainline/v6.3-rc2/amd64/linux-headers-6.3.0-060300rc2_6.3.0-060300rc2.202303122031_all.deb
sudo wget -c https://kernel.ubuntu.com/~kernel-ppa/mainline/v6.3-rc2/amd64/linux-image-unsigned-6.3.0-060300rc2-generic_6.3.0-060300rc2.202303122031_amd64.deb
sudo wget -c https://kernel.ubuntu.com/~kernel-ppa/mainline/v6.3-rc2/amd64/linux-modules-6.3.0-060300rc2-generic_6.3.0-060300rc2.202303122031_amd64.deb
sudo apt install ./linux-*.deb
#restart
sudo reboot

Develop Tips:

• use UTM/VirtualBox/VMware install VM to run;
• use docker container to run;
• use vscode ssh remote or devcontainer to develop; recomend dev container with local env;
• bench net IO , server run in VM or physical machine;
• bench storage IO , server run in physical machine mount HDD, NVMe SSD etc hardware device;
• perf programe, cpu, memery, IO cost

Bench Tips:

• Bench net IO, use tcp tools:
  • tcpdump check RST;
  • netstat check es,cw,tw stat and static tcp send recv;
  • ss static tcp stat;
  • dstat report processors related statistics iowait,sys use;,send,recv;
• Bench storage IO, use storage IO tools:
  • fio bench ioengine, check IOPS, bw;
  • vmstat check io bi,bo, swap;
  • iostat check device r/w tps, iowait;

sudo apt install moreutilscis
sudo apt install dstat
sudo apt install sysstat
sudo apt-get install linux-tools-common linux-tools-generic linux-tools-`uname -r`

Perf tools:

• pprof(go): https://github.com/google/pprof
• gperftools(c/c++): https://github.com/gperftools/gperftools
• Linux kernel profiling with perf tutorial
• more perf tools : https://www.brendangregg.com/linuxperf.html

bench result

test VM

Distributor ID: Ubuntu
Description: Ubuntu Lunar Lobster (development branch)
Release: 23.04
Codename: lunar
Linux ubuntu2 6.2.0-18-generic #18-Ubuntu SMP PREEMPT_DYNAMIC Thu Mar 16 00:09:48 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux
physical id 1
processor 4
cpu MHz : 1996.800
4 Intel(R) Core(TM) i5-1038NG7 CPU @ 2.00GHz
MemTotal: 2005084 kB

case1. golang echo server bench

one cpu vcore, set affinity cpu 0 taskset -cp 0 $SRV_PID (c pthread_setaffinity_np, syscall sched_setaffinity for thread pid)

	go-netpoll	go-iouring	go-iouring-sqpoll
c:300 bytes:128	38630	61884	78348
c:500 bytes:128	44235	65096	76976
c:1000 bytes:128	41551	62788	74774
c:2000 bytes:128	40133	64316	78213
c:300 bytes:512	42702	67029	82585
c:500 bytes:512	40298	61839	73856
c:1000 bytes:512	43908	65027	74533
c:2000 bytes:512	41182	63614	71960
c:300 bytes:1000	41143	65582	74441
c:500 bytes:1000	39191	63407	77784
c:1000 bytes:1000	42663	60978	74310
c:2000 bytes:1000	42207	58816	66598

• golang netpoll more workload(cpu), more G:M (init G:M, some M and G per connect);
• golang io_uring (sqp) less workload(cpu), less G:M (init G:M);

reference

1. https://unixism.net/loti/
2. https://unixism.net/2020/04/io-uring-by-example-article-series/
3. windows IORing: https://windows-internals.com/ioring-vs-io_uring-a-comparison-of-windows-and-linux-implementations/
4. Diego Didona - Understanding Modern Storage APIs: A systematic study of libaio, SPDK, and io_uring , video
5. awesome-iouring
6. https://openanolis.cn/sig/high-perf-storage/doc/218455073889779745
7. io_uring-and-networking-in-2023