From 660fd04f9317172ae90f414c68b18a26ae88a829 Mon Sep 17 00:00:00 2001 From: Thomas Gleixner Date: Tue, 12 Nov 2019 01:27:09 +0000 Subject: lib/vdso: Prepare for time namespace support To support time namespaces in the vdso with a minimal impact on regular non time namespace affected tasks, the namespace handling needs to be hidden in a slow path. The most obvious place is vdso_seq_begin(). If a task belongs to a time namespace then the VVAR page which contains the system wide vdso data is replaced with a namespace specific page which has the same layout as the VVAR page. That page has vdso_data->seq set to 1 to enforce the slow path and vdso_data->clock_mode set to VCLOCK_TIMENS to enforce the time namespace handling path. The extra check in the case that vdso_data->seq is odd, e.g. a concurrent update of the vdso data is in progress, is not really affecting regular tasks which are not part of a time namespace as the task is spin waiting for the update to finish and vdso_data->seq to become even again. If a time namespace task hits that code path, it invokes the corresponding time getter function which retrieves the real VVAR page, reads host time and then adds the offset for the requested clock which is stored in the special VVAR page. If VDSO time namespace support is disabled the whole magic is compiled out. Initial testing shows that the disabled case is almost identical to the host case which does not take the slow timens path. With the special timens page installed the performance hit is constant time and in the range of 5-7%. For the vdso functions which are not using the sequence count an unconditional check for vdso_data->clock_mode is added which switches to the real vdso when the clock_mode is VCLOCK_TIMENS. [avagin: Make do_hres_timens() work with raw clocks too: choose vdso_data pointer by CS_RAW offset.] Suggested-by: Andy Lutomirski Signed-off-by: Thomas Gleixner Signed-off-by: Andrei Vagin Signed-off-by: Dmitry Safonov Signed-off-by: Thomas Gleixner Link: https://lore.kernel.org/r/20191112012724.250792-21-dima@arista.com --- lib/vdso/Kconfig | 6 ++ lib/vdso/gettimeofday.c | 142 ++++++++++++++++++++++++++++++++++++++++++++++-- 2 files changed, 144 insertions(+), 4 deletions(-) (limited to 'lib/vdso') diff --git a/lib/vdso/Kconfig b/lib/vdso/Kconfig index 9fe698ff62ec..d883ac299508 100644 --- a/lib/vdso/Kconfig +++ b/lib/vdso/Kconfig @@ -24,4 +24,10 @@ config GENERIC_COMPAT_VDSO help This config option enables the compat VDSO layer. +config GENERIC_VDSO_TIME_NS + bool + help + Selected by architectures which support time namespaces in the + VDSO + endif diff --git a/lib/vdso/gettimeofday.c b/lib/vdso/gettimeofday.c index b453d2469b63..f342ac1fce77 100644 --- a/lib/vdso/gettimeofday.c +++ b/lib/vdso/gettimeofday.c @@ -38,15 +38,89 @@ u64 vdso_calc_delta(u64 cycles, u64 last, u64 mask, u32 mult) } #endif +#ifdef CONFIG_TIME_NS +static int do_hres_timens(const struct vdso_data *vdns, clockid_t clk, + struct __kernel_timespec *ts) +{ + const struct vdso_data *vd = __arch_get_timens_vdso_data(); + const struct timens_offset *offs = &vdns->offset[clk]; + const struct vdso_timestamp *vdso_ts; + u64 cycles, last, ns; + u32 seq; + s64 sec; + + if (clk != CLOCK_MONOTONIC_RAW) + vd = &vd[CS_HRES_COARSE]; + else + vd = &vd[CS_RAW]; + vdso_ts = &vd->basetime[clk]; + + do { + seq = vdso_read_begin(vd); + cycles = __arch_get_hw_counter(vd->clock_mode); + ns = vdso_ts->nsec; + last = vd->cycle_last; + if (unlikely((s64)cycles < 0)) + return -1; + + ns += vdso_calc_delta(cycles, last, vd->mask, vd->mult); + ns >>= vd->shift; + sec = vdso_ts->sec; + } while (unlikely(vdso_read_retry(vd, seq))); + + /* Add the namespace offset */ + sec += offs->sec; + ns += offs->nsec; + + /* + * Do this outside the loop: a race inside the loop could result + * in __iter_div_u64_rem() being extremely slow. + */ + ts->tv_sec = sec + __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns); + ts->tv_nsec = ns; + + return 0; +} +#else +static __always_inline const struct vdso_data *__arch_get_timens_vdso_data(void) +{ + return NULL; +} + +static int do_hres_timens(const struct vdso_data *vdns, clockid_t clk, + struct __kernel_timespec *ts) +{ + return -EINVAL; +} +#endif + static __always_inline int do_hres(const struct vdso_data *vd, clockid_t clk, - struct __kernel_timespec *ts) + struct __kernel_timespec *ts) { const struct vdso_timestamp *vdso_ts = &vd->basetime[clk]; u64 cycles, last, sec, ns; u32 seq; do { - seq = vdso_read_begin(vd); + /* + * Open coded to handle VCLOCK_TIMENS. Time namespace + * enabled tasks have a special VVAR page installed which + * has vd->seq set to 1 and vd->clock_mode set to + * VCLOCK_TIMENS. For non time namespace affected tasks + * this does not affect performance because if vd->seq is + * odd, i.e. a concurrent update is in progress the extra + * check for vd->clock_mode is just a few extra + * instructions while spin waiting for vd->seq to become + * even again. + */ + while (unlikely((seq = READ_ONCE(vd->seq)) & 1)) { + if (IS_ENABLED(CONFIG_TIME_NS) && + vd->clock_mode == VCLOCK_TIMENS) + return do_hres_timens(vd, clk, ts); + cpu_relax(); + } + smp_rmb(); + cycles = __arch_get_hw_counter(vd->clock_mode); ns = vdso_ts->nsec; last = vd->cycle_last; @@ -68,6 +142,43 @@ static __always_inline int do_hres(const struct vdso_data *vd, clockid_t clk, return 0; } +#ifdef CONFIG_TIME_NS +static int do_coarse_timens(const struct vdso_data *vdns, clockid_t clk, + struct __kernel_timespec *ts) +{ + const struct vdso_data *vd = __arch_get_timens_vdso_data(); + const struct vdso_timestamp *vdso_ts = &vd->basetime[clk]; + const struct timens_offset *offs = &vdns->offset[clk]; + u64 nsec; + s64 sec; + s32 seq; + + do { + seq = vdso_read_begin(vd); + sec = vdso_ts->sec; + nsec = vdso_ts->nsec; + } while (unlikely(vdso_read_retry(vd, seq))); + + /* Add the namespace offset */ + sec += offs->sec; + nsec += offs->nsec; + + /* + * Do this outside the loop: a race inside the loop could result + * in __iter_div_u64_rem() being extremely slow. + */ + ts->tv_sec = sec + __iter_div_u64_rem(nsec, NSEC_PER_SEC, &nsec); + ts->tv_nsec = nsec; + return 0; +} +#else +static int do_coarse_timens(const struct vdso_data *vdns, clockid_t clk, + struct __kernel_timespec *ts) +{ + return -1; +} +#endif + static __always_inline int do_coarse(const struct vdso_data *vd, clockid_t clk, struct __kernel_timespec *ts) { @@ -75,7 +186,18 @@ static __always_inline int do_coarse(const struct vdso_data *vd, clockid_t clk, u32 seq; do { - seq = vdso_read_begin(vd); + /* + * Open coded to handle VCLOCK_TIMENS. See comment in + * do_hres(). + */ + while ((seq = READ_ONCE(vd->seq)) & 1) { + if (IS_ENABLED(CONFIG_TIME_NS) && + vd->clock_mode == VCLOCK_TIMENS) + return do_coarse_timens(vd, clk, ts); + cpu_relax(); + } + smp_rmb(); + ts->tv_sec = vdso_ts->sec; ts->tv_nsec = vdso_ts->nsec; } while (unlikely(vdso_read_retry(vd, seq))); @@ -156,6 +278,10 @@ __cvdso_gettimeofday(struct __kernel_old_timeval *tv, struct timezone *tz) } if (unlikely(tz != NULL)) { + if (IS_ENABLED(CONFIG_TIME_NS) && + vd->clock_mode == VCLOCK_TIMENS) + vd = __arch_get_timens_vdso_data(); + tz->tz_minuteswest = vd[CS_HRES_COARSE].tz_minuteswest; tz->tz_dsttime = vd[CS_HRES_COARSE].tz_dsttime; } @@ -167,7 +293,12 @@ __cvdso_gettimeofday(struct __kernel_old_timeval *tv, struct timezone *tz) static __maybe_unused __kernel_old_time_t __cvdso_time(__kernel_old_time_t *time) { const struct vdso_data *vd = __arch_get_vdso_data(); - __kernel_old_time_t t = READ_ONCE(vd[CS_HRES_COARSE].basetime[CLOCK_REALTIME].sec); + __kernel_old_time_t t; + + if (IS_ENABLED(CONFIG_TIME_NS) && vd->clock_mode == VCLOCK_TIMENS) + vd = __arch_get_timens_vdso_data(); + + t = READ_ONCE(vd[CS_HRES_COARSE].basetime[CLOCK_REALTIME].sec); if (time) *time = t; @@ -189,6 +320,9 @@ int __cvdso_clock_getres_common(clockid_t clock, struct __kernel_timespec *res) if (unlikely((u32) clock >= MAX_CLOCKS)) return -1; + if (IS_ENABLED(CONFIG_TIME_NS) && vd->clock_mode == VCLOCK_TIMENS) + vd = __arch_get_timens_vdso_data(); + hrtimer_res = READ_ONCE(vd[CS_HRES_COARSE].hrtimer_res); /* * Convert the clockid to a bitmask and use it to check which -- cgit v1.2.3