Open
Description
The code used to make system calls from the runtime is a mess and a big pain point when making OS-level changes in the runtime.
Quick summary of some of the pain points:
- Completely independent implementation from package
syscall
. - System call stubs exist in a variety of OS- and arch-specific
sys*
files (57, to be precise, mostly assembly), totaling ~23k lines of code [1], which is 100% hand-written. Adding a new syscall stub often requires adding new assembly to 10+ files. - There is no generic
Syscall
function that can be used to make an arbitrary system call. You must add a new stub to make a new call.
Some things we could do to improve the situation:
- Add generic
Syscall
functions, used by other stubs. And/or use code generation to generate stubs rather than doing so manually. - Move syscall code to a new package (like
runtime/internal/syscall
or perhapsruntime/internal/linux
?) to avoid littering theruntime
package itself with so many files. - Make package
syscall
depend on the runtime'sSyscall
stub so we only have a single implementation.
[1] Some of this code is things other than syscall stubs, like thread entrypoints.
Metadata
Metadata
Assignees
Labels
Type
Projects
Status
In Progress
Milestone
Relationships
Development
No branches or pull requests
Activity
aclements commentedon Feb 8, 2022
All or most of the information we would need to generate the stubs is already in the syscall package. The needs of the syscall package are of course a bit different, but the metadata should be effectively the same.
Orthogonal to code generation, given a generic runtime Syscall function (or a few of them for different numbers of arguments), would there even be any overhead to having the stubs be in Go rather than assembly? In addition to significantly reducing the amount of assembly, I think this would encourage us to do better error handling.
cherrymui commentedon Feb 8, 2022
Just to add another pain point: the stubs handle errors differently. Some returns errno (negated or not?), some just crash hard (writing to an invalid address), which is difficult to debug sometimes.
gopherbot commentedon Feb 8, 2022
Change https://go.dev/cl/383999 mentions this issue:
runtime/internal/syscall: new package for linux
runtime/internal/syscall: new package for linux
ii64 commentedon Feb 23, 2022
Seems this introduce build error, i got:

gofrontend commit 45fd14ab8baf5e86012a808426f8ef52c1d77943
cherrymui commentedon Feb 23, 2022
@ii64 this issue is about a plan for the future, not what has happened. Your comment is unrelated to this issue.
Looks like you are using gollvm. What version of gollvm are you using? Try updating to the latest. If that still fail, please file a new issue. Thanks.
gopherbot commentedon Feb 28, 2022
Change https://go.dev/cl/388094 mentions this issue:
test: rename live_syscall.go to live_uintptrkeepalive.go
gopherbot commentedon Feb 28, 2022
Change https://go.dev/cl/388477 mentions this issue:
syscall: define Syscall in terms of RawSyscall
35 remaining items
gopherbot commentedon Aug 8, 2022
Change https://go.dev/cl/421994 mentions this issue:
runtime: move epoll syscalls to runtime/internal/syscall
runtime: move epoll syscalls to runtime/internal/syscall
gopherbot commentedon Sep 30, 2022
Change https://go.dev/cl/437295 mentions this issue:
Revert "runtime: move epoll syscalls to runtime/internal/syscall"
Revert "runtime: move epoll syscalls to runtime/internal/syscall"
gopherbot commentedon Oct 7, 2022
Change https://go.dev/cl/440115 mentions this issue:
runtime: move epoll syscalls to runtime/internal/syscall
runtime: move epoll syscalls to runtime/internal/syscall
runtime: move epoll syscalls to runtime/internal/syscall
internal/runtime
#65355runtime: use getrandom(2) for readRandom on Linux
gopherbot commentedon Sep 12, 2024
Change https://go.dev/cl/612715 mentions this issue:
runtime: use getrandom(2) for readRandom on Linux