Google Gvisor Gvisor

Add or extend syscall: $target

You are adding or extending Linux syscall support in gvisor. The target is: $target

Phase 1: Understand the current state

1. Check the syscall table — search pkg/sentry/syscalls/linux/linux64.go for the syscall name. Determine:

   • Is it registered? (has a table entry)
   • What is its support level? (Supported, PartiallySupported, ErrorWithEvent, Error)
   • What handler function is it mapped to?

2. Read the existing implementation (if any) — find the handler in pkg/sentry/syscalls/linux/sys_*.go. Look for:

   • Which flags/options are already handled
   • Which flags fall through to default: / return EINVAL / ENOSYS
   • The function signature pattern: func Name(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error)

3. Check existing ABI constants — look in pkg/abi/linux/ for the relevant constants file (e.g., prctl.go, ioctl.go, socket.go)

4. Read the Linux man page / kernel source — use WebSearch or WebFetch to look up the exact Linux behavior for the syscall or flag. This is critical for correctness. Search for the man page (e.g., man 2 prctl) and/or relevant kernel source.

Phase 2: Plan the changes

Present the user a summary before writing code:

• What the target syscall/flag does in Linux
• What changes are needed in gvisor (list files)
• Any kernel subsystems that need new support (e.g., new task fields, VFS operations)
• Whether this can be a simple implementation or needs deeper infrastructure

Wait for user confirmation before proceeding.

Phase 3: Implement

Follow this checklist — not all steps apply to every change:

A. ABI Constants (pkg/abi/linux/)

• Add any missing constants to the appropriate file
• Follow the existing naming convention (e.g., PR_ prefix for prctl, CLONE_ for clone flags)
• Add a one-line comment matching the Linux kernel comment style: // PR_FOO does bar.
• If adding a new struct, use pkg/marshal for user-space copying

B. Syscall Implementation (pkg/sentry/syscalls/linux/sys_*.go)

• New syscall: Create sys_<name>.go with the handler function. Follow the standard signature:

  func SyscallName(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
  

• New flag on existing syscall: Add a case to the existing switch statement

• Use linuxerr package for errors (e.g., linuxerr.EINVAL, linuxerr.EPERM)

• Use t.CopyIn* / t.CopyOut* or primitive.Copy* for userspace memory access

• Use hostarch.Addr for user pointers: args[N].Pointer()

• Use args[N].Int(), args[N].Uint(), args[N].Uint64() for scalar arguments

C. Syscall Table (pkg/sentry/syscalls/linux/linux64.go)

• New syscall: Change from syscalls.ErrorWithEvent(...) or syscalls.Error(...) to syscalls.Supported("name", HandlerFunc) or syscalls.PartiallySupported("name", HandlerFunc, "note about limitations", nil)
• Improved existing: Update PartiallySupported note to reflect newly supported options, or upgrade to Supported if fully implemented
• The table has entries for both AMD64 and ARM64 — update both if the syscall applies to both architectures

D. BUILD files

• Add new .go files to srcs in pkg/sentry/syscalls/linux/BUILD
• Add any new package dependencies to deps
• Add new ABI files to pkg/abi/linux/BUILD if created

E. Kernel/Sentry Support (if needed)

• New task-level state → pkg/sentry/kernel/task.go or related
• New filesystem operations → pkg/sentry/vfs/ or pkg/sentry/fsimpl/
• New network features → pkg/tcpip/

F. Tests (test/syscalls/linux/)

• Add or extend C++ tests in test/syscalls/linux/<syscall>.cc
• Use the gtest framework with gvisor's test utilities:
  • SyscallSucceeds(), SyscallFailsWithErrno(EINVAL), etc. from test/util/test_util.h
  • ASSERT_NO_ERRNO_AND_VALUE() for operations that return PosixErrorOr
• Add test to test/syscalls/linux/BUILD if it's a new file
• Test both the happy path and error cases

Phase 4: Test-driven verification against native Linux

The test MUST pass on native Linux first. This ensures the test itself is correct before running it inside gvisor. Native test targets run the C++ test binary directly on the host kernel.

Step 1: Run the native test

The native test target naming convention is:

bazel test //test/syscalls:<syscall>_test_native

For example:

bazel test //test/syscalls:access_test_native
bazel test //test/syscalls:prctl_test_native
bazel test //test/syscalls:eventfd_test_native

The native target is auto-generated by the syscall_test() macro in test/syscalls/BUILD from the cc_binary in test/syscalls/linux/BUILD. It runs with --platform=native (directly on the host kernel, no gvisor sandbox).

Run the native test FIRST and iterate until it passes. If the native test fails, the test itself is buggy — fix the test before touching the gvisor implementation.

bazel test //test/syscalls:<syscall>_test_native --test_output=errors

To run a specific test case:

bazel test //test/syscalls:<syscall>_test_native --test_output=errors --test_arg=--gtest_filter='TestSuite.TestCase'

Step 2: Build the gvisor implementation

After the native test passes, verify the gvisor code compiles:

bazel build //pkg/sentry/syscalls/linux/...

If ABI constants changed:

bazel build //pkg/abi/linux/...

Step 3: Run the gvisor (runsc) test

After the native test passes and the gvisor implementation compiles, run the test under gvisor to verify the implementation is correct:

bazel test //test/syscalls:<syscall>_test_runsc_ptrace_shared --test_output=errors

This test MUST pass. If it fails, the gvisor implementation has a bug — fix the implementation and re-run until it passes.

Iterative workflow

The expected loop is:

1. Write/update the C++ test in test/syscalls/linux/<syscall>.cc
2. Run bazel test //test/syscalls:<syscall>_test_native --test_output=errors
3. If it fails → fix the test (it's a bug in the test, not Linux)
4. Once native passes → implement/update the gvisor handler
5. Build gvisor: bazel build //pkg/sentry/syscalls/linux/...
6. Run bazel test //test/syscalls:<syscall>_test_runsc_ptrace_shared --test_output=errors
7. If it fails → fix the gvisor implementation and repeat from step 5

Key patterns to follow

Error handling

// Return Linux error codes from linuxerr package
return 0, nil, linuxerr.EINVAL
return 0, nil, linuxerr.EPERM
return 0, nil, linuxerr.ENOSYS

Copying data to/from userspace

// Copy a single int32 to userspace
_, err := primitive.CopyInt32Out(t, args[1].Pointer(), value)

// Copy a struct from userspace
var s SomeStruct
_, err := s.CopyIn(t, args[0].Pointer())

// Copy a string from userspace
name, err := t.CopyInString(addr, maxLen)

Unimplemented options pattern

// For options you deliberately don't implement, emit an event:
t.Kernel().EmitUnimplementedEvent(t, sysno)
return 0, nil, linuxerr.ENOSYS

Credential checks

creds := t.Credentials()
if !creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, creds.UserNamespace) {
    return 0, nil, linuxerr.EPERM
}

Important notes

• Match Linux behavior exactly — check the man page and kernel source for edge cases
• gvisor does NOT have a real kernel, so some things (hardware access, kernel modules) cannot be implemented — stub them with appropriate errors
• Always handle the default: case in switch statements
• Use t.Kernel().EmitUnimplementedEvent(t, sysno) before returning ENOSYS for deliberate non-implementation — this enables tracking of missing features
• Prefer returning errors over panicking
• Consider both amd64 and arm64 when the syscall exists on both architectures

Source: google/gvisor — distributed by TomeVault.