Line data    Source code

       1             : /* The nohz-full stage checks that Firedancer tile CPUs run in full
       2             :    dynticks ("nohz_full") mode.
       3             : 
       4             :    On a nohz_full CPU running exactly one runnable task, the kernel
       5             :    stops the periodic scheduler tick, and (since kernel 4.17) offloads
       6             :    even the residual 1Hz tick to a housekeeping CPU via the global
       7             :    workqueue (see the kworkers stage, which keeps that workqueue off
       8             :    tile CPUs).  This removes ~CONFIG_HZ (typically 250/s) timer
       9             :    interrupts per second per tile CPU, each costing ~2-5us plus cache
      10             :    and branch predictor pollution.  See also the rcu-nocbs stage:
      11             :    nohz_full CPUs need rcu_nocbs to reach full silence.
      12             : 
      13             :    nohz_full is a boot-time kernel parameter: it cannot be configured
      14             :    at runtime, so this stage cannot have an init step.  It is a
      15             :    check-only stage (like hyperthreads) that WARNS with the exact
      16             :    suggested parameter rather than failing, since Firedancer runs
      17             :    correctly (with slightly degraded jitter) without it.
      18             : 
      19             :    The kernel exposes the active set in
      20             :    /sys/devices/system/cpu/nohz_full (cpulist; may contain "(null)" or
      21             :    be absent when CONFIG_NO_HZ_FULL is off).
      22             : 
      23             :    A CPU listed in nohz_full but running multiple tasks just keeps its
      24             :    tick, so listing extra CPUs is harmless.  Hence the suggested
      25             :    parameter covers ALL fixed tile CPUs.  One CPU must remain outside
      26             :    the set for timekeeping; since the topology never covers every host
      27             :    CPU (other stages enforce housekeeping CPUs exist), the tile set is
      28             :    always safe to suggest. */
      29             : 
      30             : #include "configure.h"
      31             : #include "fd_cpu_isolation.h"
      32             : 
      33             : #define NAME "nohz-full"
      34             : 
      35             : #define NOHZ_FULL_PATH "/sys/devices/system/cpu/nohz_full"
      36             : 
      37             : static configure_result_t
      38             : check( config_t const * config,
      39           0 :        int              check_type ) {
      40           0 :   if( !( check_type==FD_CONFIGURE_CHECK_TYPE_CHECK ||
      41           0 :          check_type==FD_CONFIGURE_CHECK_TYPE_RUN ) ) CONFIGURE_OK();
      42             : 
      43             :   /* Jitter tuning only matters for production validators against a
      44             :      live cluster, don't nag in development. */
      45           0 :   if( FD_LIKELY( !config->is_live_cluster || config->is_dev ) ) CONFIGURE_OK();
      46             : 
      47           0 :   FD_CPUSET_DECL( tile_cpus );
      48           0 :   fd_cpu_isolation_tile_cpus( tile_cpus, &config->topo );
      49           0 :   if( FD_UNLIKELY( !fd_cpuset_cnt( tile_cpus ) ) ) CONFIGURE_OK();
      50             : 
      51           0 :   char suggested[ FD_CPU_ISOLATION_LIST_MAX ];
      52           0 :   fd_cpu_isolation_format_list( suggested, sizeof(suggested), tile_cpus );
      53             : 
      54           0 :   FD_CPUSET_DECL( nohz );
      55           0 :   if( FD_UNLIKELY( !fd_cpu_isolation_read_list( NOHZ_FULL_PATH, nohz ) ) ) {
      56           0 :     FD_LOG_WARNING(( "kernel has no nohz_full support %s(missing " NOHZ_FULL_PATH ")%s. Firedancer tiles will be "
      57           0 :                      "interrupted by periodic timer ticks. For lower jitter, use a kernel with CONFIG_NO_HZ_FULL "
      58           0 :                      "and boot with %snohz_full=%s%s.",
      59           0 :                      fd_log_style_dim(), fd_log_style_normal(),
      60           0 :                      fd_log_style_bold(), suggested, fd_log_style_normal() ));
      61           0 :     CONFIGURE_OK();
      62           0 :   }
      63             : 
      64           0 :   FD_CPUSET_DECL( missing );
      65           0 :   fd_cpuset_subtract( missing, tile_cpus, nohz );
      66           0 :   if( FD_UNLIKELY( !fd_cpuset_is_null( missing ) ) ) {
      67           0 :     char missing_str[ FD_CPU_ISOLATION_LIST_MAX ];
      68           0 :     fd_cpu_isolation_format_list( missing_str, sizeof(missing_str), missing );
      69           0 :     FD_LOG_WARNING(( "tile cpus %s are not in the kernel nohz_full= set and will be interrupted by "
      70           0 :                      "periodic timer ticks. For lower jitter, boot with %snohz_full=%s%s.",
      71           0 :                      missing_str, fd_log_style_bold(), suggested, fd_log_style_normal() ));
      72           0 :   }
      73             : 
      74           0 :   CONFIGURE_OK();
      75           0 : }
      76             : 
      77             : configure_stage_t fd_cfg_stage_nohz_full = {
      78             :   .name            = NAME,
      79             :   .always_recreate = 0,
      80             :   .enabled         = NULL,
      81             :   .init_perm       = NULL,
      82             :   .fini_perm       = NULL,
      83             :   .init            = NULL,
      84             :   .fini            = NULL,
      85             :   .check           = check,
      86             : };
      87             : 
      88             : #undef NAME
      89             : #undef NOHZ_FULL_PATH