[PATCH V2 08/19] x86/vdso: Make delta calculation overflow safe
From: Adrian Hunter <adrian.hunter@intel.com>
Date: 2024-03-25 06:41:31
Also in:
linux-s390, lkml
Subsystem:
the rest, x86 architecture (32-bit and 64-bit) · Maintainers:
Linus Torvalds, Thomas Gleixner, Ingo Molnar, Borislav Petkov, Dave Hansen
Kernel timekeeping is designed to keep the change in cycles (since the last timer interrupt) below max_cycles, which prevents multiplication overflow when converting cycles to nanoseconds. However, if timer interrupts stop, the calculation will eventually overflow. Add protection against that. Select GENERIC_VDSO_OVERFLOW_PROTECT so that max_cycles is made available in the VDSO data page. Check against max_cycles, falling back to a slower higher precision calculation. Take advantage of the opportunity to move masking and negative motion check into the slow path. The result is a calculation that has similar performance as before. Newer machines showed performance benefit, whereas older Skylake-based hardware such as Intel Kaby Lake was seen <1% worse. Suggested-by: Thomas Gleixner <redacted> Signed-off-by: Adrian Hunter <adrian.hunter@intel.com> --- arch/x86/Kconfig | 1 + arch/x86/include/asm/vdso/gettimeofday.h | 29 +++++++++++++++++------- 2 files changed, 22 insertions(+), 8 deletions(-)
diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 03483b23a009..3a70ebb558e7 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig@@ -168,6 +168,7 @@ config X86 select GENERIC_TIME_VSYSCALL select GENERIC_GETTIMEOFDAY select GENERIC_VDSO_TIME_NS + select GENERIC_VDSO_OVERFLOW_PROTECT select GUP_GET_PXX_LOW_HIGH if X86_PAE select HARDIRQS_SW_RESEND select HARDLOCKUP_CHECK_TIMESTAMP if X86_64
diff --git a/arch/x86/include/asm/vdso/gettimeofday.h b/arch/x86/include/asm/vdso/gettimeofday.h
index 5727dedd3549..0ef36190abe6 100644
--- a/arch/x86/include/asm/vdso/gettimeofday.h
+++ b/arch/x86/include/asm/vdso/gettimeofday.h@@ -319,18 +319,31 @@ static inline bool arch_vdso_cycles_ok(u64 cycles) */ static __always_inline u64 vdso_calc_ns(const struct vdso_data *vd, u64 cycles, u64 base) { + u64 delta = cycles - vd->cycle_last; + /* + * Negative motion and deltas which can cause multiplication + * overflow require special treatment. This check covers both as + * negative motion is guaranteed to be greater than @vd::max_cycles + * due to unsigned comparison. + * * Due to the MSB/Sign-bit being used as invalid marker (see - * arch_vdso_cycles_valid() above), the effective mask is S64_MAX. + * arch_vdso_cycles_valid() above), the effective mask is S64_MAX, + * but that case is also unlikely and will also take the unlikely path + * here. */ - u64 delta = (cycles - vd->cycle_last) & S64_MAX; + if (unlikely(delta > vd->max_cycles)) { + /* + * Due to the above mentioned TSC wobbles, filter out + * negative motion. Per the above masking, the effective + * sign bit is now bit 62. + */ + if (delta & (1ULL << 62)) + return base >> vd->shift; - /* - * Due to the above mentioned TSC wobbles, filter out negative motion. - * Per the above masking, the effective sign bit is now bit 62. - */ - if (unlikely(delta & (1ULL << 62))) - return base >> vd->shift; + /* Handle multiplication overflow gracefully */ + return mul_u64_u32_add_u64_shr(delta & S64_MAX, vd->mult, base, vd->shift); + } return ((delta * vd->mult) + base) >> vd->shift; }
--
2.34.1