A tiny tool to inspect your go process's goroutine info, without single line change in your code. Inspired by py-spy, learned a lot from delve
sudo ./gospy summary --pid 1234
aggregate goroutines by functions: sudo ./gospy top --pid 1234
dump a global variable value by name: sudo ./gospy var --pid 1234 --name runtime.ncpu
type: int32, value: 6
Support types:
- int(8, 16, 32, 64)
- uint(8, 16, 32, 64)
- bool
- string
dump runtime heap info: sudo ./gospy heap --pid 1234
PagesInUse: 918, PagesSwept: 977
Large Object(>32KB) Stats: AllocNum: 21, AllocRamSize: 4.52MB, FreeNum: 16, FreedRamSize: 3.38MB
SweepDone: 1, Sweepers: 0, Sweepgen: 82
P0, FlushGen:82:
Tiny size object(<16B): AllocNum: 9, BytesUsage: 11/16
Large size object freed(>32KB): FreeNum: 0, FreedRamSize: 0B
Small size object(<32KB):
0B~8B: npages: 1, allocCount: 114
8B~16B: npages: 1, allocCount: 259
16B~32B: npages: 1, allocCount: 146
32B~48B: npages: 1, allocCount: 21
48B~64B: npages: 1, allocCount: 40
64B~80B: npages: 1, allocCount: 88
80B~96B: npages: 1, allocCount: 76
96B~112B: npages: 1, allocCount: 17
112B~128B: npages: 1, allocCount: 48
144B~160B: npages: 1, allocCount: 26
160B~176B: npages: 1, allocCount: 13
176B~192B: npages: 1, allocCount: 14
208B~224B: npages: 1, allocCount: 24
240B~256B: npages: 1, allocCount: 19
256B~288B: npages: 1, allocCount: 15
288B~320B: npages: 1, allocCount: 16
352B~384B: npages: 1, allocCount: 19
480B~512B: npages: 1, allocCount: 14
512B~576B: npages: 1, allocCount: 1
576B~640B: npages: 1, allocCount: 5
704B~768B: npages: 1, allocCount: 6
768B~896B: npages: 1, allocCount: 7
896B~1.00KB: npages: 1, allocCount: 2
1.00KB~1.12KB: npages: 1, allocCount: 4
1.12KB~1.25KB: npages: 1, allocCount: 3
1.25KB~1.38KB: npages: 2, allocCount: 8
2.00KB~2.25KB: npages: 2, allocCount: 5
4.75KB~5.25KB: npages: 2, allocCount: 3
P1, FlushGen: 82:
...
Works on target go version >= 1.21 tested:
- 1.21.x
Minium go version to build: 1.13.0
git clone https://github.com/monsterxx03/gospy.git
gospy is based on go module, please don't put repo under GOPATH, it won't work.
cd gospy && make
binary will be created under gospy/bin/
- x86_64 linux only
- Don't work with binaries without debug info, if you build with linker flag
-w -s
, gospy won't be able to figure out function/variable names.
If binary is build with -buildmode=pie
, only support go >= 1.15. Since .gopclntab is missing in PIE mode before 1.15.
Read DWARF ino from ELF binary(embed by go compiler), parse some basic global variables'(runtime.allgs, runtime.allglen...) virtual memory address, then read target process's memory space to recreate runtime structs(runtime.g, runtime.p, runtime.m, runtime.sched...)
There're three ways:
PTRACE_PEEKDATA
, it can only read a long word(8 bytes) at a time. If we want to read a continuous memory space, need to call it multi times.process_vm_readv
, available after kernel 3.2, can read a continuous block of process memory space, not exposed directly in go, maybe need cgo to call?- Read
/proc/{pid}/mem
directly, it's the easiest way on linux. Also more efficient than a syscall in go.
gospy takes the third way(/proc/{pid}/mem
). Bad side is sudo privilege is required.
Yes. By default, gospy use PTRACE_ATTACH
to suspend target process to get a consistent memory view, after reading, PTRACE_DEATCH
to resume target process.
If --non-blocking
option is provided, gospy will do memory reading directly, won't suspend target process. If target process is creating/destorying goroutines actively, it may fail during reading memory.
You can compile a binary with debug info and specify the --bin
option. Ensure compile with same code revision, same go version.
Yes :)
- kubectl cp bin/gospy pod-xxx:/tmp/gospy
- kubectl exec -it pod-xxx -- /tmp/gospy top --pid 1
go get github.com/monsterxx03/gospy
import (
"fmt"
"github.com/monsterxx03/gospy/pkg/proc"
)
func main() {
p, _ := proc.New(14791, "")
ver, _ := p.GoVersion()
fmt.Println(ver) // print target process go version
sum, _ := p.Summary(true)
fmt.Println(sum)
for _, g := range sum.Gs {
fmt.Println(g)
}
}
- Support dump more variable types
- Optimize performance