Messages in this thread | | | From | Andy Lutomirski <> | Subject | [PATCH v4 0/6] Micro-optimize vclock_gettime | Date | Mon, 16 May 2011 12:00:57 -0400 |
| |
This series speeds up vclock_gettime(CLOCK_MONOTONIC) on by almost 30% (tested on Sandy Bridge). These patches are intended for 2.6.40, and if I'm feeling really ambitious I'll try to shave a few more ns off for 2.6.41. (There are lots more optimization opportunities in there.)
x86-64: Clean up vdso/kernel shared variables
Because vsyscall_gtod_data's address isn't known until load time, the code contains unnecessary address calculations. The code is also rather complicated. Clean it up and use addresses that are known at compile time.
x86-64: Remove unnecessary barrier in vread_tsc
A fair amount of testing on lots of machines has failed to find a single example in which the barrier *after* rdtsc is needed. So remove it. (The manuals give no real justification for it, and rdtsc has no dependencies so there's no sensible reason for a CPU to delay it.)
x86-64: Don't generate cmov in vread_tsc
GCC likes to generate a cmov on a branch that's almost completely predictable. Force it to generate a real branch instead.
x86-64: vclock_gettime(CLOCK_MONOTONIC) can't ever see nsec < 0
vset_normalize_timespec was more general than necessary. Open-code the appropriate normalization loops. This is a big win for CLOCK_MONOTONIC_COARSE.
x86-64: Move vread_tsc into a new file with sensible options
This way vread_tsc doesn't have a frame pointer, with saves about 0.3ns. I guess that the CPU's stack frame optimizations aren't quite as good as I thought.
x86-64: Turn off -pg and turn on -foptimize-sibling-calls for vDSO
We're building the vDSO with optimizations disabled that were meant for kernel code. Override that, except for -fno-omit-frame-pointers, which might make userspace debugging harder.
Changes from v3: - Put jiffies and vgetcpu_mode into the same cacheline. I folded it into the vsyscall cleanup patch because it's literally just changing a number. (In theory one more cacheline could be saved by putting jiffies and vgetcpu_mode at the end of gtod_data, but that would be annoying to maintain and would, I think, have little benefit. - Don't turn off frame pointers in vDSO code.
Changes from v2: - Just remove the second barrier instead of hacking it. Tests still pass.
Changes from v1: - Redo the vsyscall_gtod_data address patch to make the code cleaner instead of uglier and to make it work for all the vsyscall variables. - Improve the comments for clarity and formatting. - Fix up the changelog for the nsec < 0 tweak (the normalization code can't be inline because the two callers are different). - Move vread_tsc into its own file, removing a GCC version dependence and making it more maintainable.
Andy Lutomirski (6): x86-64: Clean up vdso/kernel shared variables x86-64: Remove unnecessary barrier in vread_tsc x86-64: Don't generate cmov in vread_tsc x86-64: vclock_gettime(CLOCK_MONOTONIC) can't ever see nsec < 0 x86-64: Move vread_tsc into a new file with sensible options x86-64: Turn off -pg and turn on -foptimize-sibling-calls for vDSO
arch/x86/include/asm/tsc.h | 4 +++ arch/x86/include/asm/vdso.h | 14 ---------- arch/x86/include/asm/vgtod.h | 2 - arch/x86/include/asm/vsyscall.h | 12 +------- arch/x86/include/asm/vvar.h | 52 +++++++++++++++++++++++++++++++++++++++ arch/x86/kernel/Makefile | 8 +++-- arch/x86/kernel/time.c | 2 +- arch/x86/kernel/tsc.c | 19 -------------- arch/x86/kernel/vmlinux.lds.S | 34 ++++++++----------------- arch/x86/kernel/vread_tsc_64.c | 36 +++++++++++++++++++++++++++ arch/x86/kernel/vsyscall_64.c | 46 +++++++++++++++------------------- arch/x86/vdso/Makefile | 17 +++++++++++- arch/x86/vdso/vclock_gettime.c | 43 +++++++++++++++++--------------- arch/x86/vdso/vdso.lds.S | 7 ----- arch/x86/vdso/vextern.h | 16 ------------ arch/x86/vdso/vgetcpu.c | 3 +- arch/x86/vdso/vma.c | 27 -------------------- arch/x86/vdso/vvar.c | 12 --------- 18 files changed, 170 insertions(+), 184 deletions(-) create mode 100644 arch/x86/include/asm/vvar.h create mode 100644 arch/x86/kernel/vread_tsc_64.c delete mode 100644 arch/x86/vdso/vextern.h delete mode 100644 arch/x86/vdso/vvar.c
-- 1.7.5.1
| |