Skip to main content

scx_layered/
stats.rs

1use std::collections::BTreeMap;
2use std::io::Write;
3use std::sync::atomic::AtomicBool;
4use std::sync::atomic::Ordering;
5use std::sync::Arc;
6use std::thread::current;
7use std::thread::ThreadId;
8use std::time::Duration;
9use std::time::SystemTime;
10use std::time::UNIX_EPOCH;
11
12use anyhow::bail;
13use anyhow::Result;
14use chrono::DateTime;
15use chrono::Local;
16use scx_stats::prelude::*;
17use scx_stats_derive::stat_doc;
18use scx_stats_derive::Stats;
19use scx_utils::Cpumask;
20use scx_utils::Topology;
21use serde::Deserialize;
22use serde::Serialize;
23use tracing::warn;
24
25use crate::bpf_intf;
26use crate::BpfStats;
27use crate::Layer;
28use crate::LayerKind;
29use crate::Stats;
30use crate::LAYER_USAGE_OPEN;
31use crate::LAYER_USAGE_PROTECTED;
32use crate::LAYER_USAGE_PROTECTED_PREEMPT;
33use crate::LAYER_USAGE_SUM_UPTO;
34
35const GSTAT_EXCL_IDLE: usize = bpf_intf::global_stat_id_GSTAT_EXCL_IDLE as usize;
36const GSTAT_EXCL_WAKEUP: usize = bpf_intf::global_stat_id_GSTAT_EXCL_WAKEUP as usize;
37const GSTAT_HI_FB_EVENTS: usize = bpf_intf::global_stat_id_GSTAT_HI_FB_EVENTS as usize;
38const GSTAT_HI_FB_USAGE: usize = bpf_intf::global_stat_id_GSTAT_HI_FB_USAGE as usize;
39const GSTAT_LO_FB_EVENTS: usize = bpf_intf::global_stat_id_GSTAT_LO_FB_EVENTS as usize;
40const GSTAT_LO_FB_USAGE: usize = bpf_intf::global_stat_id_GSTAT_LO_FB_USAGE as usize;
41const GSTAT_FB_CPU_USAGE: usize = bpf_intf::global_stat_id_GSTAT_FB_CPU_USAGE as usize;
42const GSTAT_ANTISTALL: usize = bpf_intf::global_stat_id_GSTAT_ANTISTALL as usize;
43const GSTAT_SKIP_PREEMPT: usize = bpf_intf::global_stat_id_GSTAT_SKIP_PREEMPT as usize;
44const GSTAT_FIXUP_VTIME: usize = bpf_intf::global_stat_id_GSTAT_FIXUP_VTIME as usize;
45const GSTAT_PREEMPTING_MISMATCH: usize =
46    bpf_intf::global_stat_id_GSTAT_PREEMPTING_MISMATCH as usize;
47
48const LSTAT_SEL_LOCAL: usize = bpf_intf::layer_stat_id_LSTAT_SEL_LOCAL as usize;
49const LSTAT_ENQ_LOCAL: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_LOCAL as usize;
50const LSTAT_ENQ_WAKEUP: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_WAKEUP as usize;
51const LSTAT_ENQ_EXPIRE: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_EXPIRE as usize;
52const LSTAT_ENQ_REENQ: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_REENQ as usize;
53const LSTAT_ENQ_DSQ: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_DSQ as usize;
54const LSTAT_MIN_EXEC: usize = bpf_intf::layer_stat_id_LSTAT_MIN_EXEC as usize;
55const LSTAT_MIN_EXEC_NS: usize = bpf_intf::layer_stat_id_LSTAT_MIN_EXEC_NS as usize;
56const LSTAT_OPEN_IDLE: usize = bpf_intf::layer_stat_id_LSTAT_OPEN_IDLE as usize;
57const LSTAT_AFFN_VIOL: usize = bpf_intf::layer_stat_id_LSTAT_AFFN_VIOL as usize;
58const LSTAT_KEEP: usize = bpf_intf::layer_stat_id_LSTAT_KEEP as usize;
59const LSTAT_KEEP_FAIL_MAX_EXEC: usize = bpf_intf::layer_stat_id_LSTAT_KEEP_FAIL_MAX_EXEC as usize;
60const LSTAT_KEEP_FAIL_BUSY: usize = bpf_intf::layer_stat_id_LSTAT_KEEP_FAIL_BUSY as usize;
61const LSTAT_PREEMPT: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT as usize;
62const LSTAT_PREEMPT_FIRST: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_FIRST as usize;
63const LSTAT_PREEMPT_XLLC: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_XLLC as usize;
64const LSTAT_PREEMPT_XNUMA: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_XNUMA as usize;
65const LSTAT_PREEMPT_IDLE: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_IDLE as usize;
66const LSTAT_PREEMPT_FAIL: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_FAIL as usize;
67const LSTAT_EXCL_COLLISION: usize = bpf_intf::layer_stat_id_LSTAT_EXCL_COLLISION as usize;
68const LSTAT_EXCL_PREEMPT: usize = bpf_intf::layer_stat_id_LSTAT_EXCL_PREEMPT as usize;
69const LSTAT_YIELD: usize = bpf_intf::layer_stat_id_LSTAT_YIELD as usize;
70const LSTAT_YIELD_IGNORE: usize = bpf_intf::layer_stat_id_LSTAT_YIELD_IGNORE as usize;
71const LSTAT_MIGRATION: usize = bpf_intf::layer_stat_id_LSTAT_MIGRATION as usize;
72const LSTAT_XNUMA_MIGRATION: usize = bpf_intf::layer_stat_id_LSTAT_XNUMA_MIGRATION as usize;
73const LSTAT_XLLC_MIGRATION: usize = bpf_intf::layer_stat_id_LSTAT_XLLC_MIGRATION as usize;
74const LSTAT_XLLC_MIGRATION_SKIP: usize = bpf_intf::layer_stat_id_LSTAT_XLLC_MIGRATION_SKIP as usize;
75const LSTAT_XLAYER_WAKE: usize = bpf_intf::layer_stat_id_LSTAT_XLAYER_WAKE as usize;
76const LSTAT_XLAYER_REWAKE: usize = bpf_intf::layer_stat_id_LSTAT_XLAYER_REWAKE as usize;
77const LSTAT_LLC_DRAIN_TRY: usize = bpf_intf::layer_stat_id_LSTAT_LLC_DRAIN_TRY as usize;
78const LSTAT_LLC_DRAIN: usize = bpf_intf::layer_stat_id_LSTAT_LLC_DRAIN as usize;
79const LSTAT_SKIP_REMOTE_NODE: usize = bpf_intf::layer_stat_id_LSTAT_SKIP_REMOTE_NODE as usize;
80
81const LSTAT_RUNQ_LAT_BASE: usize = bpf_intf::layer_stat_id_LSTAT_RUNQ_LAT_BASE as usize;
82const NR_RUNQ_LAT_BUCKETS: usize = bpf_intf::consts_NR_RUNQ_LAT_BUCKETS as usize;
83
84const LLC_LSTAT_LAT: usize = bpf_intf::llc_layer_stat_id_LLC_LSTAT_LAT as usize;
85const LLC_LSTAT_CNT: usize = bpf_intf::llc_layer_stat_id_LLC_LSTAT_CNT as usize;
86
87fn calc_frac(a: f64, b: f64) -> f64 {
88    if b != 0.0 {
89        a / b * 100.0
90    } else {
91        0.0
92    }
93}
94
95fn fmt_pct(v: f64) -> String {
96    if v >= 99.95 {
97        format!("{:4.0}", v)
98    } else if v >= 10.0 {
99        format!("{:4.1}", v)
100    } else if v > 0.0 && v < 0.01 {
101        format!("{:4.2}", 0.01)
102    } else {
103        format!("{:4.2}", v)
104    }
105}
106
107fn fmt_duration_ms(ms: f64) -> String {
108    if ms >= 60_000.0 {
109        let min = ms / 60_000.0;
110        if min >= 100.0 {
111            format!("{:.0}min", min)
112        } else {
113            format!("{:.1}min", min)
114        }
115    } else if ms >= 1_000.0 {
116        let s = ms / 1_000.0;
117        if s >= 100.0 {
118            format!("{:.0}s", s)
119        } else {
120            format!("{:.1}s", s)
121        }
122    } else if ms >= 10.0 {
123        format!("{:.0}ms", ms)
124    } else {
125        format!("{:.1}ms", ms)
126    }
127}
128
129fn fmt_num(v: u64) -> String {
130    if v > 1_000_000 {
131        format!("{:5.1}m", v as f64 / 1_000_000.0)
132    } else if v > 1_000 {
133        format!("{:5.1}k", v as f64 / 1_000.0)
134    } else {
135        format!("{:5.0} ", v)
136    }
137}
138
139#[stat_doc]
140#[derive(Clone, Debug, Default, Serialize, Deserialize, Stats)]
141#[stat(_om_prefix = "l_", _om_label = "layer_name")]
142pub struct LayerStats {
143    #[stat(desc = "index", _om_skip)]
144    pub index: usize,
145    #[stat(desc = "Total CPU utilization (100% means one full CPU)")]
146    pub util: f64,
147    #[stat(desc = "Compensated CPU utilization (adjusted for irq/softirq/stolen)")]
148    pub util_compensated: f64,
149    #[stat(desc = "Protected CPU utilization %")]
150    pub util_protected_frac: f64,
151    #[stat(desc = "Preempt-protected CPU utilization %")]
152    pub util_protected_preempt_frac: f64,
153    #[stat(desc = "Open CPU utilization %")]
154    pub util_open_frac: f64,
155    #[stat(desc = "fraction of total CPU utilization")]
156    pub util_frac: f64,
157    #[stat(desc = "number of tasks")]
158    pub tasks: u32,
159    #[stat(desc = "count of sched events during the period")]
160    pub total: u64,
161    #[stat(desc = "% dispatched into idle CPU from select_cpu")]
162    pub sel_local: f64,
163    #[stat(desc = "% dispatched into idle CPU from enqueue")]
164    pub enq_local: f64,
165    #[stat(desc = "% enqueued after wakeup")]
166    pub enq_wakeup: f64,
167    #[stat(desc = "% enqueued after slice expiration")]
168    pub enq_expire: f64,
169    #[stat(desc = "% re-enqueued due to RT preemption")]
170    pub enq_reenq: f64,
171    #[stat(desc = "% enqueued into the layer's LLC DSQs")]
172    pub enq_dsq: f64,
173    #[stat(desc = "count of times exec duration < min_exec_us")]
174    pub min_exec: f64,
175    #[stat(desc = "total exec durations extended due to min_exec_us")]
176    pub min_exec_us: u64,
177    #[stat(desc = "% dispatched into idle CPUs occupied by other layers")]
178    pub open_idle: f64,
179    #[stat(desc = "% preempted other tasks")]
180    pub preempt: f64,
181    #[stat(desc = "% preempted XLLC tasks")]
182    pub preempt_xllc: f64,
183    #[stat(desc = "% preempted across NUMA nodes")]
184    pub preempt_xnuma: f64,
185    #[stat(desc = "% first-preempted other tasks")]
186    pub preempt_first: f64,
187    #[stat(desc = "% idle-preempted other tasks")]
188    pub preempt_idle: f64,
189    #[stat(desc = "% attempted to preempt other tasks but failed")]
190    pub preempt_fail: f64,
191    #[stat(desc = "% violated config due to CPU affinity")]
192    pub affn_viol: f64,
193    #[stat(desc = "% continued executing after slice expiration")]
194    pub keep: f64,
195    #[stat(desc = "% disallowed to continue executing due to max_exec")]
196    pub keep_fail_max_exec: f64,
197    #[stat(desc = "% disallowed to continue executing due to other tasks")]
198    pub keep_fail_busy: f64,
199    #[stat(desc = "whether is exclusive", _om_skip)]
200    pub is_excl: u32,
201    #[stat(desc = "count of times an excl task skipped a CPU as the sibling was also excl")]
202    pub excl_collision: f64,
203    #[stat(desc = "% a sibling CPU was preempted for an exclusive task")]
204    pub excl_preempt: f64,
205    #[stat(desc = "% yielded")]
206    pub yielded: f64,
207    #[stat(desc = "count of times yield was ignored")]
208    pub yield_ignore: u64,
209    #[stat(desc = "% migrated across CPUs")]
210    pub migration: f64,
211    #[stat(desc = "% migrated across NUMA nodes")]
212    pub xnuma_migration: f64,
213    #[stat(desc = "% migrated across LLCs")]
214    pub xllc_migration: f64,
215    #[stat(desc = "% migration skipped across LLCs due to xllc_mig_min_us")]
216    pub xllc_migration_skip: f64,
217    #[stat(desc = "% wakers across layers")]
218    pub xlayer_wake: f64,
219    #[stat(desc = "% rewakers across layers where waker has waken the task previously")]
220    pub xlayer_rewake: f64,
221    #[stat(desc = "% LLC draining tried")]
222    pub llc_drain_try: f64,
223    #[stat(desc = "% LLC draining succeeded")]
224    pub llc_drain: f64,
225    #[stat(desc = "% skip LLC dispatch on remote node")]
226    pub skip_remote_node: f64,
227    #[stat(desc = "mask of allocated CPUs", _om_skip)]
228    pub cpus: Vec<u64>,
229    #[stat(desc = "count of CPUs assigned")]
230    pub cur_nr_cpus: u32,
231    #[stat(desc = "minimum # of CPUs assigned")]
232    pub min_nr_cpus: u32,
233    #[stat(desc = "maximum # of CPUs assigned")]
234    pub max_nr_cpus: u32,
235    #[stat(desc = "count of CPUs assigned per LLC")]
236    pub nr_llc_cpus: Vec<u32>,
237    #[stat(desc = "slice duration config")]
238    pub slice_us: u64,
239    #[stat(desc = "Per-LLC scheduling event fractions")]
240    pub llc_fracs: Vec<f64>,
241    #[stat(desc = "Per-LLC average latency")]
242    pub llc_lats: Vec<f64>,
243    #[stat(desc = "Layer memory bandwidth as a % of total allowed (0 for \"no limit\"")]
244    pub membw_pct: f64,
245    #[stat(desc = "DSQ insertion ratio EWMA (10s window)")]
246    pub dsq_insert_ewma: f64,
247    #[stat(desc = "Per-node layer utilization (100% = one full CPU)")]
248    pub node_utils: Vec<f64>,
249    #[stat(desc = "Per-node pinned task utilization (100% = one full CPU)")]
250    pub node_pinned_utils: Vec<f64>,
251    #[stat(desc = "Per-node pinned task counts")]
252    pub node_pinned_tasks: Vec<u64>,
253    #[stat(desc = "Per-node load (100% = one full CPU, from duty cycle sum)")]
254    pub node_loads: Vec<f64>,
255    #[stat(desc = "Whether xnuma gating is active for this layer (0/1)")]
256    pub xnuma_active: u32,
257    #[stat(desc = "runqueue latency histogram, log2 us buckets 1us..32s, per stats interval")]
258    pub l_runq_lat_hist: Vec<u64>,
259}
260
261impl LayerStats {
262    pub fn new(
263        lidx: usize,
264        layer: &Layer,
265        stats: &Stats,
266        bstats: &BpfStats,
267        nr_cpus_range: (usize, usize),
268        xnuma_active: bool,
269    ) -> Self {
270        let lstat = |sidx| bstats.lstats[lidx][sidx];
271        let ltotal = lstat(LSTAT_SEL_LOCAL)
272            + lstat(LSTAT_ENQ_LOCAL)
273            + lstat(LSTAT_ENQ_WAKEUP)
274            + lstat(LSTAT_ENQ_EXPIRE)
275            + lstat(LSTAT_ENQ_REENQ)
276            + lstat(LSTAT_KEEP);
277        let lstat_pct = |sidx| {
278            if ltotal != 0 {
279                lstat(sidx) as f64 / ltotal as f64 * 100.0
280            } else {
281                0.0
282            }
283        };
284
285        let util_sum = stats.layer_utils[lidx]
286            .iter()
287            .take(LAYER_USAGE_SUM_UPTO + 1)
288            .sum::<f64>();
289
290        let util_comp_sum = stats.layer_utils_compensated[lidx]
291            .iter()
292            .take(LAYER_USAGE_SUM_UPTO + 1)
293            .sum::<f64>();
294
295        let membw_frac = match &layer.kind {
296            // Open layer's can't have a memory BW limit.
297            LayerKind::Open { .. } => 0.0,
298            LayerKind::Confined { membw_gb, .. } | LayerKind::Grouped { membw_gb, .. } => {
299                // Check if we have set a memory BW limit.
300                if let Some(membw_limit_gb) = membw_gb {
301                    stats.layer_membws[lidx]
302                        .iter()
303                        .take(LAYER_USAGE_SUM_UPTO + 1)
304                        .sum::<f64>()
305                        / (*membw_limit_gb * (1024_u64.pow(3) as f64))
306                } else {
307                    0.0
308                }
309            }
310        };
311
312        Self {
313            index: lidx,
314            util: util_sum * 100.0,
315            util_compensated: util_comp_sum * 100.0,
316            util_open_frac: calc_frac(stats.layer_utils[lidx][LAYER_USAGE_OPEN], util_sum),
317            util_protected_frac: calc_frac(
318                stats.layer_utils[lidx][LAYER_USAGE_PROTECTED],
319                util_sum,
320            ),
321            util_protected_preempt_frac: calc_frac(
322                stats.layer_utils[lidx][LAYER_USAGE_PROTECTED_PREEMPT],
323                util_sum,
324            ),
325            util_frac: calc_frac(util_sum, stats.total_util),
326            tasks: stats.nr_layer_tasks[lidx] as u32,
327            total: ltotal,
328            sel_local: lstat_pct(LSTAT_SEL_LOCAL),
329            enq_local: lstat_pct(LSTAT_ENQ_LOCAL),
330            enq_wakeup: lstat_pct(LSTAT_ENQ_WAKEUP),
331            enq_expire: lstat_pct(LSTAT_ENQ_EXPIRE),
332            enq_reenq: lstat_pct(LSTAT_ENQ_REENQ),
333            enq_dsq: lstat_pct(LSTAT_ENQ_DSQ),
334            min_exec: lstat_pct(LSTAT_MIN_EXEC),
335            min_exec_us: lstat(LSTAT_MIN_EXEC_NS) / 1000,
336            open_idle: lstat_pct(LSTAT_OPEN_IDLE),
337            preempt: lstat_pct(LSTAT_PREEMPT),
338            preempt_xllc: lstat_pct(LSTAT_PREEMPT_XLLC),
339            preempt_xnuma: lstat_pct(LSTAT_PREEMPT_XNUMA),
340            preempt_first: lstat_pct(LSTAT_PREEMPT_FIRST),
341            preempt_idle: lstat_pct(LSTAT_PREEMPT_IDLE),
342            preempt_fail: lstat_pct(LSTAT_PREEMPT_FAIL),
343            affn_viol: lstat_pct(LSTAT_AFFN_VIOL),
344            keep: lstat_pct(LSTAT_KEEP),
345            keep_fail_max_exec: lstat_pct(LSTAT_KEEP_FAIL_MAX_EXEC),
346            keep_fail_busy: lstat_pct(LSTAT_KEEP_FAIL_BUSY),
347            is_excl: layer.kind.common().exclusive as u32,
348            excl_collision: lstat_pct(LSTAT_EXCL_COLLISION),
349            excl_preempt: lstat_pct(LSTAT_EXCL_PREEMPT),
350            yielded: lstat_pct(LSTAT_YIELD),
351            yield_ignore: lstat(LSTAT_YIELD_IGNORE),
352            migration: lstat_pct(LSTAT_MIGRATION),
353            xnuma_migration: lstat_pct(LSTAT_XNUMA_MIGRATION),
354            xlayer_wake: lstat_pct(LSTAT_XLAYER_WAKE),
355            xlayer_rewake: lstat_pct(LSTAT_XLAYER_REWAKE),
356            xllc_migration: lstat_pct(LSTAT_XLLC_MIGRATION),
357            xllc_migration_skip: lstat_pct(LSTAT_XLLC_MIGRATION_SKIP),
358            llc_drain_try: lstat_pct(LSTAT_LLC_DRAIN_TRY),
359            llc_drain: lstat_pct(LSTAT_LLC_DRAIN),
360            skip_remote_node: lstat_pct(LSTAT_SKIP_REMOTE_NODE),
361            cpus: layer.cpus.as_raw_slice().to_vec(),
362            cur_nr_cpus: layer.cpus.weight() as u32,
363            min_nr_cpus: nr_cpus_range.0 as u32,
364            max_nr_cpus: nr_cpus_range.1 as u32,
365            nr_llc_cpus: layer.nr_llc_cpus.iter().map(|&v| v as u32).collect(),
366            slice_us: stats.layer_slice_us[lidx],
367            llc_fracs: {
368                let sid = LLC_LSTAT_CNT;
369                let sum = bstats.llc_lstats[lidx]
370                    .iter()
371                    .map(|lstats| lstats[sid])
372                    .sum::<u64>() as f64;
373                bstats.llc_lstats[lidx]
374                    .iter()
375                    .map(|lstats| calc_frac(lstats[sid] as f64, sum))
376                    .collect()
377            },
378            llc_lats: bstats.llc_lstats[lidx]
379                .iter()
380                .map(|lstats| lstats[LLC_LSTAT_LAT] as f64 / 1_000_000_000.0)
381                .collect(),
382            membw_pct: membw_frac * 100.0,
383            dsq_insert_ewma: stats.layer_dsq_insert_ewma[lidx] * 100.0,
384            node_utils: stats.layer_node_utils[lidx]
385                .iter()
386                .map(|u| u * 100.0)
387                .collect(),
388            node_pinned_utils: stats.layer_node_pinned_utils[lidx]
389                .iter()
390                .map(|u| u * 100.0)
391                .collect(),
392            node_pinned_tasks: stats.layer_nr_node_pinned_tasks[lidx].clone(),
393            node_loads: stats.layer_node_duty_sums[lidx]
394                .iter()
395                .map(|l| l * 100.0)
396                .collect(),
397            xnuma_active: if xnuma_active { 1 } else { 0 },
398            l_runq_lat_hist: (0..NR_RUNQ_LAT_BUCKETS)
399                .map(|b| lstat(LSTAT_RUNQ_LAT_BASE + b))
400                .collect(),
401        }
402    }
403
404    pub fn format<W: Write>(
405        &self,
406        w: &mut W,
407        name: &str,
408        topo: Option<&Topology>,
409        max_width: usize,
410        no_llc: bool,
411    ) -> Result<()> {
412        // Line 1: layer summary
413        let comp_str = if self.util > 0.1 && (self.util_compensated - self.util).abs() > 0.1 {
414            let overhead_pct = (1.0 - self.util / self.util_compensated) * 100.0;
415            format!(" comp_overhead={:.1}%", overhead_pct)
416        } else {
417            String::new()
418        };
419        writeln!(
420            w,
421            "\n\u{25B6} {} \u{2500} util/open/frac={:6.1}/{}/{:7.1}{} prot/prot_preempt={}/{} tasks={:6}",
422            name,
423            self.util,
424            fmt_pct(self.util_open_frac),
425            self.util_frac,
426            comp_str,
427            fmt_pct(self.util_protected_frac),
428            fmt_pct(self.util_protected_preempt_frac),
429            self.tasks,
430        )?;
431
432        // sched: scheduling event flow
433        writeln!(
434            w,
435            "  {:<7} tot={} dd_sel/enq={}/{} dsq/10s={}/{} wake/exp/re={}/{}/{}",
436            "sched",
437            fmt_num(self.total),
438            fmt_pct(self.sel_local),
439            fmt_pct(self.enq_local),
440            fmt_pct(self.enq_dsq),
441            fmt_pct(self.dsq_insert_ewma),
442            fmt_pct(self.enq_wakeup),
443            fmt_pct(self.enq_expire),
444            fmt_pct(self.enq_reenq),
445        )?;
446
447        // exec: execution behavior (merged keep/yield + slice/min_exec)
448        writeln!(
449            w,
450            "  {:<7} keep/max/busy={}/{}/{} yield/ign={}/{} slc={} min_ex={}/{}",
451            "exec",
452            fmt_pct(self.keep),
453            fmt_pct(self.keep_fail_max_exec),
454            fmt_pct(self.keep_fail_busy),
455            fmt_pct(self.yielded),
456            fmt_num(self.yield_ignore),
457            fmt_duration_ms(self.slice_us as f64 / 1000.0),
458            fmt_pct(self.min_exec),
459            fmt_duration_ms(self.min_exec_us as f64 / 1000.0),
460        )?;
461
462        // mig: CPU placement and movement
463        writeln!(
464            w,
465            "  {:<7} mig={} xnuma={} xllc/skip={}/{} open_idle={} affn_viol={}",
466            "mig",
467            fmt_pct(self.migration),
468            fmt_pct(self.xnuma_migration),
469            fmt_pct(self.xllc_migration),
470            fmt_pct(self.xllc_migration_skip),
471            fmt_pct(self.open_idle),
472            fmt_pct(self.affn_viol),
473        )?;
474
475        // preempt: preemption
476        writeln!(
477            w,
478            "  {:<7} preempt/first/xllc/xnuma/idle/fail={}/{}/{}/{}/{}/{}",
479            "preempt",
480            fmt_pct(self.preempt),
481            fmt_pct(self.preempt_first),
482            fmt_pct(self.preempt_xllc),
483            fmt_pct(self.preempt_xnuma),
484            fmt_pct(self.preempt_idle),
485            fmt_pct(self.preempt_fail),
486        )?;
487
488        // xlayer: cross-layer and LLC dispatch
489        writeln!(
490            w,
491            "  {:<7} wake/re={}/{} llc_drain/try={}/{} skip_rnode={}",
492            "xlayer",
493            fmt_pct(self.xlayer_wake),
494            fmt_pct(self.xlayer_rewake),
495            fmt_pct(self.llc_drain),
496            fmt_pct(self.llc_drain_try),
497            fmt_pct(self.skip_remote_node),
498        )?;
499
500        // per-node utilization, load, and pinned utilization (multi-node only)
501        if self.node_utils.len() > 1 {
502            let prefix = "  node    pin/ut/ld ";
503            // N99=999.9/999.9/99999.9 = 24 chars + 1 space = 25
504            let cell_width = 25;
505            let usable = if max_width > prefix.len() {
506                max_width - prefix.len()
507            } else {
508                60
509            };
510            let cells_per_row = (usable / cell_width).max(1);
511
512            for nid in 0..self.node_utils.len() {
513                let util = self.node_utils[nid];
514                let load = self.node_loads.get(nid).copied().unwrap_or(0.0);
515                let pin = self.node_pinned_utils.get(nid).copied().unwrap_or(0.0);
516                if nid % cells_per_row == 0 {
517                    if nid > 0 {
518                        writeln!(w)?;
519                    }
520                    write!(w, "{prefix}")?;
521                } else {
522                    write!(w, " ")?;
523                }
524                write!(w, "N{}={:5.1}/{:5.1}/{:7.1}", nid, pin, util, load)?;
525            }
526            writeln!(w)?;
527        }
528
529        // cpumask
530        let cpumask = Cpumask::from_vec(self.cpus.clone());
531
532        if let Some(topo) = topo {
533            let header = topo.format_cpumask_header(&cpumask, self.min_nr_cpus, self.max_nr_cpus);
534            writeln!(w, "  {}", header)?;
535            if cpumask.weight() > 0 {
536                topo.format_cpumask_grid(w, &cpumask, "  ", max_width)?;
537            }
538        } else {
539            writeln!(
540                w,
541                "  cpus={:3} [{:3},{:3}] {}",
542                self.cur_nr_cpus, self.min_nr_cpus, self.max_nr_cpus, cpumask,
543            )?;
544        }
545
546        // excl stats
547        if self.is_excl != 0 {
548            writeln!(
549                w,
550                "  excl_coll={} excl_preempt={}",
551                fmt_pct(self.excl_collision),
552                fmt_pct(self.excl_preempt),
553            )?;
554        } else if self.excl_collision != 0.0 || self.excl_preempt != 0.0 {
555            warn!(
556                "{}: exclusive is off but excl_coll={} excl_preempt={}",
557                name,
558                fmt_pct(self.excl_collision),
559                fmt_pct(self.excl_preempt),
560            );
561        }
562
563        // LLC stats (compact grid, skip inactive)
564        if !no_llc {
565            // Collect active LLCs (nr_cpus > 0 or frac > 0)
566            let active_llcs: Vec<(usize, f64, f64)> = self
567                .llc_fracs
568                .iter()
569                .zip(self.llc_lats.iter())
570                .enumerate()
571                .filter(|(i, (&frac, _))| {
572                    let nr_cpus = self.nr_llc_cpus.get(*i).copied().unwrap_or(0);
573                    nr_cpus > 0 || frac > 0.0
574                })
575                .map(|(i, (&frac, &lat))| (i, frac, lat))
576                .collect();
577
578            if !active_llcs.is_empty() {
579                let indent = "  ";
580                writeln!(w, "{indent}LLC sched%/lat_ms")?;
581                // Cell format: [XX]99.9/99.9 = 13 chars, + 1 space separator = 14
582                let cell_width = 14;
583                let usable = if max_width > indent.len() {
584                    max_width - indent.len()
585                } else {
586                    60
587                };
588                let cells_per_row = (usable / cell_width).max(1);
589
590                for (col, &(llc_id, frac, lat)) in active_llcs.iter().enumerate() {
591                    if col % cells_per_row == 0 {
592                        if col > 0 {
593                            writeln!(w)?;
594                        }
595                        write!(w, "{indent}")?;
596                    } else {
597                        write!(w, " ")?;
598                    }
599                    write!(w, "[{:02}]{}/{:4.1}", llc_id, fmt_pct(frac), lat * 1_000.0)?;
600                }
601                writeln!(w)?;
602            }
603        }
604
605        Ok(())
606    }
607}
608
609#[stat_doc]
610#[derive(Clone, Debug, Default, Serialize, Deserialize, Stats)]
611#[stat(top)]
612pub struct SysStats {
613    #[stat(desc = "timestamp", _om_skip)]
614    pub at: f64,
615    #[stat(desc = "# of NUMA nodes")]
616    pub nr_nodes: usize,
617    #[stat(desc = "# sched events during the period")]
618    pub total: u64,
619    #[stat(desc = "% dispatched directly into an idle CPU from select_cpu")]
620    pub local_sel: f64,
621    #[stat(desc = "% dispatched directly into an idle CPU from enqueue")]
622    pub local_enq: f64,
623    #[stat(desc = "% open layer tasks scheduled into allocated but idle CPUs")]
624    pub open_idle: f64,
625    #[stat(desc = "% violated config due to CPU affinity")]
626    pub affn_viol: f64,
627    #[stat(desc = "% sent to hi fallback DSQs")]
628    pub hi_fb: f64,
629    #[stat(desc = "% sent to lo fallback DSQs")]
630    pub lo_fb: f64,
631    #[stat(desc = "count of times an excl task skipped a CPU as the sibling was also excl")]
632    pub excl_collision: f64,
633    #[stat(desc = "count of times a sibling CPU was preempted for an excl task")]
634    pub excl_preempt: f64,
635    #[stat(desc = "count of times a CPU skipped dispatching due to an excl task on the sibling")]
636    pub excl_idle: f64,
637    #[stat(
638        desc = "count of times an idle sibling CPU was woken up after an excl task is finished"
639    )]
640    pub excl_wakeup: f64,
641    #[stat(desc = "CPU time this binary consumed during the period")]
642    pub proc_ms: u64,
643    #[stat(desc = "CPU busy % (100% means all CPU)")]
644    pub busy: f64,
645    #[stat(desc = "CPU util % (100% means one CPU)")]
646    pub util: f64,
647    #[stat(desc = "CPU util % used by hi fallback DSQs")]
648    pub hi_fb_util: f64,
649    #[stat(desc = "CPU util % used by lo fallback DSQs")]
650    pub lo_fb_util: f64,
651    #[stat(desc = "Number of tasks dispatched via antistall")]
652    pub antistall: u64,
653    #[stat(desc = "Number of times preemptions of non-scx tasks were avoided")]
654    pub skip_preempt: u64,
655    #[stat(desc = "Number of times vtime was out of range and fixed up")]
656    pub fixup_vtime: u64,
657    #[stat(desc = "Number of times cpuc->preempting_task didn't come on the CPU")]
658    pub preempting_mismatch: u64,
659    #[stat(desc = "per-node fallback CPUs")]
660    pub fallback_cpus: BTreeMap<u32, u32>,
661    #[stat(desc = "per-layer statistics")]
662    pub fallback_cpu_util: f64,
663    #[stat(desc = "fallback CPU util %")]
664    pub layers: BTreeMap<String, LayerStats>,
665    #[stat(desc = "Number of gpu tasks affinitized since scheduler start")]
666    pub gpu_tasks_affinitized: u64,
667    #[stat(desc = "Time (in ms) of last affinitization run.")]
668    pub gpu_task_affinitization_ms: u64,
669    #[stat(desc = "System CPU utilization EWMA (10s window)")]
670    pub system_cpu_util_ewma: f64,
671}
672
673impl SysStats {
674    pub fn new(
675        stats: &Stats,
676        bstats: &BpfStats,
677        fallback_cpus: &BTreeMap<usize, usize>,
678    ) -> Result<Self> {
679        let lsum = |idx| stats.bpf_stats.lstats_sums[idx];
680        let total = lsum(LSTAT_SEL_LOCAL)
681            + lsum(LSTAT_ENQ_LOCAL)
682            + lsum(LSTAT_ENQ_WAKEUP)
683            + lsum(LSTAT_ENQ_EXPIRE)
684            + lsum(LSTAT_ENQ_REENQ)
685            + lsum(LSTAT_KEEP);
686        let lsum_pct = |idx| {
687            if total != 0 {
688                lsum(idx) as f64 / total as f64 * 100.0
689            } else {
690                0.0
691            }
692        };
693
694        let elapsed_ns = stats.elapsed.as_nanos();
695
696        Ok(Self {
697            at: SystemTime::now().duration_since(UNIX_EPOCH)?.as_secs_f64(),
698            nr_nodes: stats.topo.nodes.len(),
699            total,
700            local_sel: lsum_pct(LSTAT_SEL_LOCAL),
701            local_enq: lsum_pct(LSTAT_ENQ_LOCAL),
702            open_idle: lsum_pct(LSTAT_OPEN_IDLE),
703            affn_viol: lsum_pct(LSTAT_AFFN_VIOL),
704            hi_fb: calc_frac(
705                stats.bpf_stats.gstats[GSTAT_HI_FB_EVENTS] as f64,
706                total as f64,
707            ),
708            lo_fb: calc_frac(
709                stats.bpf_stats.gstats[GSTAT_LO_FB_EVENTS] as f64,
710                total as f64,
711            ),
712            excl_collision: lsum_pct(LSTAT_EXCL_COLLISION),
713            excl_preempt: lsum_pct(LSTAT_EXCL_PREEMPT),
714            excl_idle: bstats.gstats[GSTAT_EXCL_IDLE] as f64 / total as f64,
715            excl_wakeup: bstats.gstats[GSTAT_EXCL_WAKEUP] as f64 / total as f64,
716            proc_ms: stats.processing_dur.as_millis() as u64,
717            busy: stats.cpu_busy * 100.0,
718            util: stats.total_util * 100.0,
719            hi_fb_util: stats.bpf_stats.gstats[GSTAT_HI_FB_USAGE] as f64 / elapsed_ns as f64
720                * 100.0,
721            lo_fb_util: stats.bpf_stats.gstats[GSTAT_LO_FB_USAGE] as f64 / elapsed_ns as f64
722                * 100.0,
723            antistall: stats.bpf_stats.gstats[GSTAT_ANTISTALL],
724            skip_preempt: stats.bpf_stats.gstats[GSTAT_SKIP_PREEMPT],
725            fixup_vtime: stats.bpf_stats.gstats[GSTAT_FIXUP_VTIME],
726            preempting_mismatch: stats.bpf_stats.gstats[GSTAT_PREEMPTING_MISMATCH],
727            fallback_cpus: fallback_cpus
728                .iter()
729                .map(|(&k, &v)| (k as u32, v as u32))
730                .collect(),
731            fallback_cpu_util: stats.bpf_stats.gstats[GSTAT_FB_CPU_USAGE] as f64
732                / elapsed_ns as f64
733                * 100.0,
734            layers: BTreeMap::new(),
735            gpu_tasks_affinitized: stats.gpu_tasks_affinitized,
736            gpu_task_affinitization_ms: stats.gpu_task_affinitization_ms,
737            system_cpu_util_ewma: stats.system_cpu_util_ewma * 100.0,
738        })
739    }
740
741    pub fn format<W: Write>(&self, w: &mut W) -> Result<()> {
742        writeln!(
743            w,
744            "tot={:7} local_sel/enq={}/{} open_idle={} affn_viol={} hi/lo={}/{}",
745            self.total,
746            fmt_pct(self.local_sel),
747            fmt_pct(self.local_enq),
748            fmt_pct(self.open_idle),
749            fmt_pct(self.affn_viol),
750            fmt_pct(self.hi_fb),
751            fmt_pct(self.lo_fb),
752        )?;
753
754        let single_node = self.fallback_cpus.len() == 1;
755        let fb_cpus_str: Vec<String> = self
756            .fallback_cpus
757            .iter()
758            .map(|(n, c)| {
759                if single_node {
760                    format!("{}", c)
761                } else {
762                    format!("N{}:{}", n, c)
763                }
764            })
765            .collect();
766        writeln!(
767            w,
768            "busy={:5.1} util/hi/lo={:7.1}/{}/{} fb_cpus=[{}]/util={:4.1} proc={}ms sys_util_10s={:5.1}",
769            self.busy,
770            self.util,
771            fmt_pct(self.hi_fb_util),
772            fmt_pct(self.lo_fb_util),
773            fb_cpus_str.join(","),
774            self.fallback_cpu_util,
775            self.proc_ms,
776            self.system_cpu_util_ewma,
777        )?;
778
779        writeln!(
780            w,
781            "excl_coll={:.2} excl_preempt={:.2} excl_idle={:.2} excl_wakeup={:.2}",
782            self.excl_collision, self.excl_preempt, self.excl_idle, self.excl_wakeup
783        )?;
784
785        writeln!(
786            w,
787            "skip_preempt={} antistall={} fixup_vtime={} preempting_mismatch={}",
788            self.skip_preempt, self.antistall, self.fixup_vtime, self.preempting_mismatch
789        )?;
790
791        writeln!(
792            w,
793            "gpu_tasks_affinitized={} gpu_task_affinitization_time={}",
794            self.gpu_tasks_affinitized, self.gpu_task_affinitization_ms
795        )?;
796
797        Ok(())
798    }
799
800    pub fn format_all<W: Write>(
801        &self,
802        w: &mut W,
803        topo: Option<&Topology>,
804        max_width: usize,
805        no_llc: bool,
806    ) -> Result<()> {
807        self.format(w)?;
808
809        let mut idx_to_name: Vec<(usize, &String)> =
810            self.layers.iter().map(|(k, v)| (v.index, k)).collect();
811
812        idx_to_name.sort();
813
814        for (_idx, name) in &idx_to_name {
815            self.layers[*name].format(w, name, topo, max_width, no_llc)?;
816        }
817
818        Ok(())
819    }
820}
821
822#[derive(Debug)]
823pub enum StatsReq {
824    Hello(ThreadId),
825    Refresh(ThreadId, Box<Stats>),
826    Bye(ThreadId),
827}
828
829#[derive(Debug)]
830pub enum StatsRes {
831    Hello(Box<Stats>),
832    Refreshed(Box<(Stats, SysStats)>),
833    Bye,
834}
835
836pub fn server_data() -> StatsServerData<StatsReq, StatsRes> {
837    let open: Box<dyn StatsOpener<StatsReq, StatsRes>> = Box::new(move |(req_ch, res_ch)| {
838        let tid = current().id();
839        req_ch.send(StatsReq::Hello(tid))?;
840        let mut stats = Some(match res_ch.recv()? {
841            StatsRes::Hello(v) => *v,
842            res => bail!("invalid response to Hello: {:?}", res),
843        });
844
845        let read: Box<dyn StatsReader<StatsReq, StatsRes>> =
846            Box::new(move |_args, (req_ch, res_ch)| {
847                req_ch.send(StatsReq::Refresh(tid, Box::new(stats.take().unwrap())))?;
848                let (new_stats, sys_stats) = match res_ch.recv()? {
849                    StatsRes::Refreshed(v) => *v,
850                    res => bail!("invalid response to Refresh: {:?}", res),
851                };
852                stats = Some(new_stats);
853                sys_stats.to_json()
854            });
855
856        Ok(read)
857    });
858
859    let close: Box<dyn StatsCloser<StatsReq, StatsRes>> = Box::new(move |(req_ch, res_ch)| {
860        req_ch.send(StatsReq::Bye(current().id())).unwrap();
861        match res_ch.recv().unwrap() {
862            StatsRes::Bye => {}
863            res => panic!("invalid response to Bye: {:?}", res),
864        }
865    });
866
867    StatsServerData::new()
868        .add_meta(LayerStats::meta())
869        .add_meta(SysStats::meta())
870        .add_ops(
871            "top",
872            StatsOps {
873                open,
874                close: Some(close),
875            },
876        )
877}
878
879pub fn monitor(
880    intv: Duration,
881    shutdown: Arc<AtomicBool>,
882    max_width: usize,
883    no_llc: bool,
884) -> Result<()> {
885    let topo = Topology::new().ok();
886    scx_utils::monitor_stats::<SysStats>(
887        &[],
888        intv,
889        || shutdown.load(Ordering::Relaxed),
890        |sst| {
891            let dt = DateTime::<Local>::from(UNIX_EPOCH + Duration::from_secs_f64(sst.at));
892            let header = format!("\u{2501}\u{2501} {} ", dt.to_rfc2822());
893            let pad = max_width.saturating_sub(header.chars().count());
894            println!("{}{}", header, "\u{2501}".repeat(pad));
895            sst.format_all(&mut std::io::stdout(), topo.as_ref(), max_width, no_llc)
896        },
897    )
898}