Messages in this thread | | | Date | Fri, 9 Oct 2015 15:06:54 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH v2 00/36] x86: Rewrite all syscall entries except native 64-bit |
| |
* Andy Lutomirski <luto@kernel.org> wrote:
> The first two patches are optimizations that I'm surprised we didn't > already have. I noticed them when I was looking at the generated > asm. > > The next two patches are tests and some old stuff. There's a test > that validates the vDSO AT_SYSINFO annotations. There's also a test > that exercises some assumptions that signal handling and ptracers > make about syscalls that currently do *not* hold on 64-bit AMD using > 32-bit AT_SYSINFO. > > The next three patches are NT cleanups and a lockdep cleanup. > > It may pay to apply the beginning of the series (at most through > "x86/entry/64/compat: After SYSENTER, move STI after the NT fixup") > without waiting for everyone to wrap their heads around the rest. > > The rest is basically a rewrite of syscalls for all cases except > 64-bit native. With these patches applied, there is a single 32-bit > vDSO and it uses SYSCALL, SYSENTER, and INT80 almost interchangeably > via alternatives. The semantics of SYSENTER and SYSCALL are defined > as: > > 1. If SYSCALL, ESP = ECX > 2. ECX = *ESP > 3. IP = INT80 landing pad > 4. Opportunistic SYSRET/SYSEXIT is enabled on return > > The vDSO is rearranged so that these semantics work. Anything that > backs IP up by 2 ends up pointing at a bona fide int $0x80 > instruction with the expected regs. > > In the process, the vDSO CFI annotations (which are actually used) > get rewritten using normal CFI directives. > > Opportunistic SYSRET/SYSEXIT only happens on return when CS and SS > are as expected, IP points to the INT80 landing pad, and flags are > in good shape. (There is no longer any assumption that full > fast-path 32-bit syscalls don't muck with the registers that matter > for fast exits -- I played with maintaining an optimization like > that with poor results. I may try again if it saves a few cycles.) > > Other than that, the system call entries are simplified to the bare > minimum prologue and a call to a C function. Amusingly, SYSENTER > and SYSCALL32 use the same C function. > > To make that work, I had to remove all the 32-bit syscall stubs > except the clone argument hack. This is because, for C code to call > through the system call table, the system call table entries need to > be real function pointers with C-compatible ABIs. > > There is nothing at all anymore that requires that x86_32 syscalls > be asmlinkage. That could be removed in a subsequent patch. > > The upshot appears to be a ~16 cycle performance hit on 32-bit fast > path syscalls. (On my system, my little prctl test takes 172 cycles > before and 188 cycles with these patches applied.) > > The slow path is probably faster under most circumstances and, if > the exit slow path gets hit, it'll be much faster because (as we > already do in the 64-bit native case) we can still use > SYSEXIT/SYSRET. > > The patchset is structured as a removal of the old fast syscall > code, then the change that makes syscalls into real functions, then > a clean re-implementation of fast syscalls. > > If we want some of the 25 cycles back, we could consider open-coding > a new C fast path. > > Changes from v1: > - The unwind_vdso_32 test now warns on broken Debian installations > instead of failing. The problem is now fully understood, will > be fixed by Debian and possibly also fixed by upstream glibc. > - execve was rather broken in v1. > - It's quite a bit faster now (the optimizations at the end are mostly new). > - int80 on 64-bit no longer clobbers extra regs (thanks Denys!). > - The uaccess stuff is new. > - Lots of other things that I forgot, I'm sure. > > Andy Lutomirski (36): > x86/uaccess: Tell the compiler that uaccess is unlikely to fault > x86/uaccess: __chk_range_not_ok is unlikely to return true > selftests/x86: Add a test for vDSO unwinding > selftests/x86: Add a test for syscall restart and arg modification > x86/entry/64/compat: Fix SYSENTER's NT flag before user memory access > x86/entry: Move lockdep_sys_exit to prepare_exit_to_usermode > x86/entry/64/compat: After SYSENTER, move STI after the NT fixup > x86/vdso: Remove runtime 32-bit vDSO selection > x86/asm: Re-add manual CFI infrastructure > x86/vdso: Define BUILD_VDSO while building and emit .eh_frame in asm > x86/vdso: Replace hex int80 CFI annotations with gas directives > x86/elf/64: Clear more registers in elf_common_init > x86/vdso/32: Save extra registers in the INT80 vsyscall path > x86/entry/64/compat: Disable SYSENTER and SYSCALL32 entries > x86/entry/64/compat: Remove audit optimizations > x86/entry/64/compat: Remove most of the fast system call machinery > x86/entry/64/compat: Set up full pt_regs for all compat syscalls > x86/entry/syscalls: Move syscall table declarations into > asm/syscalls.h > x86/syscalls: Give sys_call_ptr_t a useful type > x86/entry: Add do_syscall_32, a C function to do 32-bit syscalls > x86/entry/64/compat: Migrate the body of the syscall entry to C > x86/entry: Add C code for fast system call entries > x86/vdso/compat: Wire up SYSENTER and SYSCSALL for compat userspace > x86/entry/compat: Implement opportunistic SYSRETL for compat syscalls > x86/entry/32: Open-code return tracking from fork and kthreads > x86/entry/32: Switch INT80 to the new C syscall path > x86/entry/32: Re-implement SYSENTER using the new C path > x86/asm: Remove thread_info.sysenter_return > x86/entry: Remove unnecessary IRQ twiddling in fast 32-bit syscalls > x86/entry: Make irqs_disabled checks in exit code depend on lockdep > x86/entry: Force inlining of 32-bit syscall code > x86/entry: Micro-optimize compat fast syscall arg fetch > x86/entry: Hide two syscall entry assertions behind CONFIG_DEBUG_ENTRY > x86/entry: Use pt_regs_to_thread_info() in syscall entry tracing > x86/entry: Split and inline prepare_exit_to_usermode > x86/entry: Split and inline syscall_return_slowpath > > arch/x86/Makefile | 10 +- > arch/x86/entry/common.c | 255 ++++++++-- > arch/x86/entry/entry_32.S | 184 +++---- > arch/x86/entry/entry_64.S | 9 +- > arch/x86/entry/entry_64_compat.S | 541 +++++---------------- > arch/x86/entry/syscall_32.c | 9 +- > arch/x86/entry/syscall_64.c | 4 +- > arch/x86/entry/syscalls/syscall_32.tbl | 12 +- > arch/x86/entry/vdso/Makefile | 39 +- > arch/x86/entry/vdso/vdso2c.c | 2 +- > arch/x86/entry/vdso/vdso32-setup.c | 28 +- > arch/x86/entry/vdso/vdso32/int80.S | 56 --- > arch/x86/entry/vdso/vdso32/syscall.S | 75 --- > arch/x86/entry/vdso/vdso32/sysenter.S | 116 ----- > arch/x86/entry/vdso/vdso32/system_call.S | 57 +++ > arch/x86/entry/vdso/vma.c | 13 +- > arch/x86/ia32/ia32_signal.c | 4 +- > arch/x86/include/asm/dwarf2.h | 177 +++++++ > arch/x86/include/asm/elf.h | 10 +- > arch/x86/include/asm/syscall.h | 14 +- > arch/x86/include/asm/thread_info.h | 1 - > arch/x86/include/asm/uaccess.h | 14 +- > arch/x86/include/asm/vdso.h | 10 +- > arch/x86/kernel/asm-offsets.c | 3 - > arch/x86/kernel/signal.c | 4 +- > arch/x86/um/sys_call_table_32.c | 7 +- > arch/x86/um/sys_call_table_64.c | 7 +- > arch/x86/xen/setup.c | 13 +- > tools/testing/selftests/x86/Makefile | 5 +- > tools/testing/selftests/x86/ptrace_syscall.c | 294 +++++++++++ > .../testing/selftests/x86/raw_syscall_helper_32.S | 46 ++ > tools/testing/selftests/x86/unwind_vdso.c | 209 ++++++++ > 32 files changed, 1258 insertions(+), 970 deletions(-) > delete mode 100644 arch/x86/entry/vdso/vdso32/int80.S > delete mode 100644 arch/x86/entry/vdso/vdso32/syscall.S > delete mode 100644 arch/x86/entry/vdso/vdso32/sysenter.S > create mode 100644 arch/x86/entry/vdso/vdso32/system_call.S > create mode 100644 arch/x86/include/asm/dwarf2.h > create mode 100644 tools/testing/selftests/x86/ptrace_syscall.c > create mode 100644 tools/testing/selftests/x86/raw_syscall_helper_32.S > create mode 100644 tools/testing/selftests/x86/unwind_vdso.c
Ok, so I applied all of them to tip:x86/asm, in two phases, with small (stylistic) edits - it all seems to work fine for me so far, so I pushed it all out to -tip and linux-next.
Thanks,
Ingo
| |