LXR linux/include/linux/psi

   1#ifndef _LINUX_PSI_TYPES_H
   2#define _LINUX_PSI_TYPES_H
   3
   4#include <linux/kthread.h>
   5#include <linux/seqlock.h>
   6#include <linux/types.h>
   7#include <linux/kref.h>
   8#include <linux/wait.h>
   9
  10#ifdef CONFIG_PSI
  11
  12/* Tracked task states */
  13enum psi_task_count {
  14        NR_IOWAIT,
  15        NR_MEMSTALL,
  16        NR_RUNNING,
  17        NR_PSI_TASK_COUNTS = 3,
  18};
  19
  20/* Task state bitmasks */
  21#define TSK_IOWAIT      (1 << NR_IOWAIT)
  22#define TSK_MEMSTALL    (1 << NR_MEMSTALL)
  23#define TSK_RUNNING     (1 << NR_RUNNING)
  24
  25/* Resources that workloads could be stalled on */
  26enum psi_res {
  27        PSI_IO,
  28        PSI_MEM,
  29        PSI_CPU,
  30        NR_PSI_RESOURCES = 3,
  31};
  32
  33/*
  34 * Pressure states for each resource:
  35 *
  36 * SOME: Stalled tasks & working tasks
  37 * FULL: Stalled tasks & no working tasks
  38 */
  39enum psi_states {
  40        PSI_IO_SOME,
  41        PSI_IO_FULL,
  42        PSI_MEM_SOME,
  43        PSI_MEM_FULL,
  44        PSI_CPU_SOME,
  45        /* Only per-CPU, to weigh the CPU in the global average: */
  46        PSI_NONIDLE,
  47        NR_PSI_STATES = 6,
  48};
  49
  50enum psi_aggregators {
  51        PSI_AVGS = 0,
  52        PSI_POLL,
  53        NR_PSI_AGGREGATORS,
  54};
  55
  56struct psi_group_cpu {
  57        /* 1st cacheline updated by the scheduler */
  58
  59        /* Aggregator needs to know of concurrent changes */
  60        seqcount_t seq ____cacheline_aligned_in_smp;
  61
  62        /* States of the tasks belonging to this group */
  63        unsigned int tasks[NR_PSI_TASK_COUNTS];
  64
  65        /* Aggregate pressure state derived from the tasks */
  66        u32 state_mask;
  67
  68        /* Period time sampling buckets for each state of interest (ns) */
  69        u32 times[NR_PSI_STATES];
  70
  71        /* Time of last task change in this group (rq_clock) */
  72        u64 state_start;
  73
  74        /* 2nd cacheline updated by the aggregator */
  75
  76        /* Delta detection against the sampling buckets */
  77        u32 times_prev[NR_PSI_AGGREGATORS][NR_PSI_STATES]
  78                        ____cacheline_aligned_in_smp;
  79};
  80
  81/* PSI growth tracking window */
  82struct psi_window {
  83        /* Window size in ns */
  84        u64 size;
  85
  86        /* Start time of the current window in ns */
  87        u64 start_time;
  88
  89        /* Value at the start of the window */
  90        u64 start_value;
  91
  92        /* Value growth in the previous window */
  93        u64 prev_growth;
  94};
  95
  96struct psi_trigger {
  97        /* PSI state being monitored by the trigger */
  98        enum psi_states state;
  99
 100        /* User-spacified threshold in ns */
 101        u64 threshold;
 102
 103        /* List node inside triggers list */
 104        struct list_head node;
 105
 106        /* Backpointer needed during trigger destruction */
 107        struct psi_group *group;
 108
 109        /* Wait queue for polling */
 110        wait_queue_head_t event_wait;
 111
 112        /* Pending event flag */
 113        int event;
 114
 115        /* Tracking window */
 116        struct psi_window win;
 117
 118        /*
 119         * Time last event was generated. Used for rate-limiting
 120         * events to one per window
 121         */
 122        u64 last_event_time;
 123
 124        /* Refcounting to prevent premature destruction */
 125        struct kref refcount;
 126};
 127
 128struct psi_group {
 129        /* Protects data used by the aggregator */
 130        struct mutex avgs_lock;
 131
 132        /* Per-cpu task state & time tracking */
 133        struct psi_group_cpu __percpu *pcpu;
 134
 135        /* Running pressure averages */
 136        u64 avg_total[NR_PSI_STATES - 1];
 137        u64 avg_last_update;
 138        u64 avg_next_update;
 139
 140        /* Aggregator work control */
 141        struct delayed_work avgs_work;
 142
 143        /* Total stall times and sampled pressure averages */
 144        u64 total[NR_PSI_AGGREGATORS][NR_PSI_STATES - 1];
 145        unsigned long avg[NR_PSI_STATES - 1][3];
 146
 147        /* Monitor work control */
 148        atomic_t poll_scheduled;
 149        struct kthread_worker __rcu *poll_kworker;
 150        struct kthread_delayed_work poll_work;
 151
 152        /* Protects data used by the monitor */
 153        struct mutex trigger_lock;
 154
 155        /* Configured polling triggers */
 156        struct list_head triggers;
 157        u32 nr_triggers[NR_PSI_STATES - 1];
 158        u32 poll_states;
 159        u64 poll_min_period;
 160
 161        /* Total stall times at the start of monitor activation */
 162        u64 polling_total[NR_PSI_STATES - 1];
 163        u64 polling_next_update;
 164        u64 polling_until;
 165};
 166
 167#else /* CONFIG_PSI */
 168
 169struct psi_group { };
 170
 171#endif /* CONFIG_PSI */
 172
 173#endif /* _LINUX_PSI_TYPES_H */
 174