Kernel API wishlist

List of linux kernel work that would help rr.

In-kernel buffering of syscall outparam data

Syscallbuf+vsyscall hooking works OK in practice, but the kernel can do so much better, so much more easily. Essentially it "just" has to record copy_to_user() within syscall handlers, plus a bit of metadata. Perhaps could be exposed through perf_event.

Expose Branch Trace Store to userspace

Probably through perf_event. There was a previous attempt to do this, but I think it fizzled out.

Efficient way for ptrace tracers to wait on a tracee's private futex

rr wants to use this for efficiently waiting on CLONE_CLEAR_CHILDTID futexes.

Bugfix: pwrite64/pread64 modify registers in a way visible to userspace

Counter to the syscall ABI. Bug that we shouldn't let slip through the cracks. Ubuntu bug isn't moving, need to test on kernel HEAD and upstream to kernel folks.

Bugfix: agree a kernel/userspace contract for ptrace traps of PR_TSC_SIGSEGV

See #692: the kernel munges sigstate to enable SIGSEGV traps when SIGSEGV is blocked, but apparently doesn't un-munge the state.

Bugfix: raise PTRACE_O_SYSGOOD traps when PTRACE_SINGLESTEP'ing into syscalls?

See #212. Alternatively, create a new ptrace request analogous to PTRACE_SINGLESTEP_SYSEMU, perhaps PTACE_SINGLESTEP_SYSCALL.

Bugfix: are drivers really allowed to use userspace stack as scratch memory?

See #876. That's generally not a good idea because it can let sensitive data leak into userspace.

Bugfix?: ability to open perf_event_open() fds through /proc/TID/fd/x

See #603. We fall back on using SCM_RIGHTS, but it's considerably more complicated than needs to be.

Bugfix?: wait status 0x80007f seems invalid

See #1198. This wait_status (in response to a PTRACE_INTERRUPT) seems to contradict documentation. It's received when we interrupt a task blocked in an untraced system call.