Hello, This patchset implements stats estimation in kthread context. It replaces the code that runs on single CPU in timer context every 2 seconds and causing latency splats as shown in reports [1], [2], [3]. The solution targets setups with thousands of IPVS services, destinations and multi-CPU boxes. Spread the estimation on multiple (configured) CPUs and multiple time slots (timer ticks) by using multiple chains organized under RCU rules. When stats are not needed, it is recommended to use run_estimation=0 as already implemented before this change. Overview of the basic concepts. More in the commit messages... RCU Locking: - As stats are now RCU-locked, tot_stats, svc and dest which hold estimator structures are now always freed from RCU callback. This ensures RCU grace period after the ip_vs_stop_estimator() call. Kthread data: - every kthread works over its own data structure and all such structures are attached to array. For now we limit kthreads depending on the number of CPUs. - even while there can be a kthread structure, its task may not be running, eg. before first service is added or while the sysctl var is set to an empty cpulist or when run_estimation is set to 0 to disable the estimation. - the allocated kthread context may grow from 1 to 50 allocated structures for timer ticks which saves memory for setups with small number of estimators - a task and its structure may be released if all estimators are unlinked from its chains, leaving the slot in the array empty - every kthread data structure allows limited number of estimators. Kthread 0 is also used to initially calculate the max number of estimators to allow in every chain considering a sub-100 microsecond cond_resched rate. This number can be from 1 to hundreds. - kthread 0 has an additional job of optimizing the adding of estimators: they are first added in temp list (est_temp_list) and later kthread 0 distributes them to other kthreads. The optimization is based on the fact that newly added estimator should be estimated after 2 seconds, so we have the time to offload the adding to chain from controlling process to kthread 0. - to add new estimators we use the last added kthread context (est_add_ktid). The new estimators are linked to the chains just before the estimated one, based on add_row. This ensures their estimation will start after 2 seconds. If estimators are added in bursts, common case if all services and dests are initially configured, we may spread the estimators to more chains and as result, reducing the initial delay below 2 seconds. Many thanks to Jiri Wiesner for his valuable comments and for spending a lot of time reviewing and testing the changes on different platforms with 48-256 CPUs and 1-8 NUMA nodes under different cpufreq governors. Changes in v7: Patch 5: * use the IPVS_EST_CPU_KTHREADS value from patch 4 to limit kthreads Patch 4: * add IPVS_EST_LOAD_DIVISOR=8 to specify the max CPU load per kthread * add IPVS_EST_CPU_KTHREADS to limit load from kthreads per CPU * add build checks for the values of ktrow and ktcid Changes in v6: Patch 4 (was 3 in v5): * while ests are in est_temp_list use est->ktrow as delay (offset from est_row) to apply on enqueue, initially or later when estimators are relinked. By this way we try to preserve the current timer delay on relinking. If there is no space in the desired time slot, the initial delay can be reduced, while on relinking the delay can be increased. Still, this is far from perfect and the relinking continues to apply wrong delay for some of the ests while its goal is still to fill kthreads data one by one. * ip_vs_est_calc_limits() now relinks the entries trying to preserve the delay, ests with lowest delay will get higher chance to preserve it * add_row is now not updated (it is wrong on relinking) but will continue to be used as starting position for the bit searching, so logic is preserved * while relinking try to keep tot_stats in kt0, we do not want to keep some kthread 1..N just for it * ip_vs_est_calc_limits() now performs 12 tests but sleeps for 20ms on every 4 tests to make sure CPU speed is increased Patch 3: * new patch that converts per-cpu counters to u64_stats_t Changes in v5: Patch 4 (was 3 in v4): * use ip_vs_est_max_threads() helper Patch 3 (was 2 in v4): * kthread 0 now disables BH instead of using SCHED_FIFO mode, this should work because now we perform less number of repeated test over pre-allocated kd->calc_stats structure. Use cache_factor = 4 to approximate the time non-cached (due to large number) per-cpu stats are estimated compared to the cached data we estimate in calc phase. This needs to be tested with different number of CPUs and NUMA nodes. * limit max threads using the formula 4 * Number of CPUs, not on rlimit * remove _len suffix from vars like chain_max_len, tick_max_len, est_chain_max_len Patch 2: * new patch that adds functions for stats allocations Changes in v4: Patch 2: * kthread 0 can start with calculation phase in SCHED_FIFO mode to determine chain_max_len suitable for 100us cond_resched rate and 12% of 40ms CPU usage in a tick. Current value of IPVS_EST_TICK_CHAINS=48 determines tick time of 4.8ms (i.e. in units of 100us) which is 12% of max tick time of 40ms. The question is how reliable will be such calculation test. * est_calc_phase indicates a mode where we dequeue estimators from kthreads, apply new chain_max_len and enqueue again all estimators to kthreads, done by kthread 0 * est->ktid now can be -1 to indicate est is in est_temp_list ready to be distributed to kthread by kt 0, done in ip_vs_est_drain_temp_list(). kthread 0 data is now released only after the data for others kthreads * ip_vs_start_estimator was not setting ret = 0 * READ_ONCE not needed for volatile jiffies Patch 3: * restrict cpulist based on the cpus_allowed of process that assigns cpulist, not on cpu_possible_mask * change of cpulist will trigger calc phase Patch 5: * print message every minute, not 2 seconds Changes in v3: Patch 2: * calculate chain_max_len (was IPVS_EST_CHAIN_DEPTH) but it needs further tuning based on real estimation test * est_max_threads set from rlimit(RLIMIT_NPROC). I don't see analog to get_ucounts_value() to get the max value. * the atomic bitop for td->present is not needed, remove it * start filling based on est_row after 2 ticks are fully allocated. As 2/50 is 4% this can be increased more. Changes in v2: Patch 2: * kd->mutex is gone, cond_resched rate determined by IPVS_EST_CHAIN_DEPTH * IPVS_EST_MAX_COUNT is a hard limit now * kthread data is now 1-50 allocated tick structures, each containing heads for limited chains. Bitmaps should allow faster access. We avoid large allocations for structs. * as the td->present bitmap is shared, use atomic bitops * ip_vs_start_estimator now returns error code * _bh locking removed from stats->lock * bump arg is gone from ip_vs_est_reload_start * prepare for upcoming changes that remove _irq from u64_stats_fetch_begin_irq/u64_stats_fetch_retry_irq * est_add_ktid is now always valid Patch 3: * use .. in est_nice docs [1] Report from Yunhong Jiang: https://lore.kernel.org/netdev/D25792C1-1B89-45DE-9F10-EC350DC04ADC@xxxxxxxxx/ [2] https://marc.info/?l=linux-virtual-server&m=159679809118027&w=2 [3] Report from Dust: https://archive.linuxvirtualserver.org/html/lvs-devel/2020-12/msg00000.html Julian Anastasov (6): ipvs: add rcu protection to stats ipvs: use common functions for stats allocation ipvs: use u64_stats_t for the per-cpu counters ipvs: use kthreads for stats estimation ipvs: add est_cpulist and est_nice sysctl vars ipvs: run_estimation should control the kthread tasks Documentation/networking/ipvs-sysctl.rst | 24 +- include/net/ip_vs.h | 171 ++++- net/netfilter/ipvs/ip_vs_core.c | 40 +- net/netfilter/ipvs/ip_vs_ctl.c | 448 +++++++++--- net/netfilter/ipvs/ip_vs_est.c | 882 +++++++++++++++++++++-- 5 files changed, 1380 insertions(+), 185 deletions(-) -- 2.38.1
