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 log::warn;
17use scx_stats::prelude::*;
18use scx_stats_derive::stat_doc;
19use scx_stats_derive::Stats;
20use scx_utils::Cpumask;
21use serde::Deserialize;
22use serde::Serialize;
23
24use crate::bpf_intf;
25use crate::BpfStats;
26use crate::Layer;
27use crate::LayerKind;
28use crate::Stats;
29use crate::LAYER_USAGE_OPEN;
30use crate::LAYER_USAGE_PROTECTED;
31use crate::LAYER_USAGE_PROTECTED_PREEMPT;
32use crate::LAYER_USAGE_SUM_UPTO;
33
34const GSTAT_EXCL_IDLE: usize = bpf_intf::global_stat_id_GSTAT_EXCL_IDLE as usize;
35const GSTAT_EXCL_WAKEUP: usize = bpf_intf::global_stat_id_GSTAT_EXCL_WAKEUP as usize;
36const GSTAT_HI_FB_EVENTS: usize = bpf_intf::global_stat_id_GSTAT_HI_FB_EVENTS as usize;
37const GSTAT_HI_FB_USAGE: usize = bpf_intf::global_stat_id_GSTAT_HI_FB_USAGE as usize;
38const GSTAT_LO_FB_EVENTS: usize = bpf_intf::global_stat_id_GSTAT_LO_FB_EVENTS as usize;
39const GSTAT_LO_FB_USAGE: usize = bpf_intf::global_stat_id_GSTAT_LO_FB_USAGE as usize;
40const GSTAT_FB_CPU_USAGE: usize = bpf_intf::global_stat_id_GSTAT_FB_CPU_USAGE as usize;
41const GSTAT_ANTISTALL: usize = bpf_intf::global_stat_id_GSTAT_ANTISTALL as usize;
42const GSTAT_SKIP_PREEMPT: usize = bpf_intf::global_stat_id_GSTAT_SKIP_PREEMPT as usize;
43const GSTAT_FIXUP_VTIME: usize = bpf_intf::global_stat_id_GSTAT_FIXUP_VTIME as usize;
44const GSTAT_PREEMPTING_MISMATCH: usize =
45    bpf_intf::global_stat_id_GSTAT_PREEMPTING_MISMATCH as usize;
46
47const LSTAT_SEL_LOCAL: usize = bpf_intf::layer_stat_id_LSTAT_SEL_LOCAL as usize;
48const LSTAT_ENQ_LOCAL: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_LOCAL as usize;
49const LSTAT_ENQ_WAKEUP: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_WAKEUP as usize;
50const LSTAT_ENQ_EXPIRE: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_EXPIRE as usize;
51const LSTAT_ENQ_REENQ: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_REENQ as usize;
52const LSTAT_ENQ_DSQ: usize = bpf_intf::layer_stat_id_LSTAT_ENQ_DSQ as usize;
53const LSTAT_MIN_EXEC: usize = bpf_intf::layer_stat_id_LSTAT_MIN_EXEC as usize;
54const LSTAT_MIN_EXEC_NS: usize = bpf_intf::layer_stat_id_LSTAT_MIN_EXEC_NS as usize;
55const LSTAT_OPEN_IDLE: usize = bpf_intf::layer_stat_id_LSTAT_OPEN_IDLE as usize;
56const LSTAT_AFFN_VIOL: usize = bpf_intf::layer_stat_id_LSTAT_AFFN_VIOL as usize;
57const LSTAT_KEEP: usize = bpf_intf::layer_stat_id_LSTAT_KEEP as usize;
58const LSTAT_KEEP_FAIL_MAX_EXEC: usize = bpf_intf::layer_stat_id_LSTAT_KEEP_FAIL_MAX_EXEC as usize;
59const LSTAT_KEEP_FAIL_BUSY: usize = bpf_intf::layer_stat_id_LSTAT_KEEP_FAIL_BUSY as usize;
60const LSTAT_PREEMPT: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT as usize;
61const LSTAT_PREEMPT_FIRST: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_FIRST as usize;
62const LSTAT_PREEMPT_XLLC: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_XLLC as usize;
63const LSTAT_PREEMPT_XNUMA: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_XNUMA as usize;
64const LSTAT_PREEMPT_IDLE: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_IDLE as usize;
65const LSTAT_PREEMPT_FAIL: usize = bpf_intf::layer_stat_id_LSTAT_PREEMPT_FAIL as usize;
66const LSTAT_EXCL_COLLISION: usize = bpf_intf::layer_stat_id_LSTAT_EXCL_COLLISION as usize;
67const LSTAT_EXCL_PREEMPT: usize = bpf_intf::layer_stat_id_LSTAT_EXCL_PREEMPT as usize;
68const LSTAT_YIELD: usize = bpf_intf::layer_stat_id_LSTAT_YIELD as usize;
69const LSTAT_YIELD_IGNORE: usize = bpf_intf::layer_stat_id_LSTAT_YIELD_IGNORE as usize;
70const LSTAT_MIGRATION: usize = bpf_intf::layer_stat_id_LSTAT_MIGRATION as usize;
71const LSTAT_XNUMA_MIGRATION: usize = bpf_intf::layer_stat_id_LSTAT_XNUMA_MIGRATION as usize;
72const LSTAT_XLLC_MIGRATION: usize = bpf_intf::layer_stat_id_LSTAT_XLLC_MIGRATION as usize;
73const LSTAT_XLLC_MIGRATION_SKIP: usize = bpf_intf::layer_stat_id_LSTAT_XLLC_MIGRATION_SKIP as usize;
74const LSTAT_XLAYER_WAKE: usize = bpf_intf::layer_stat_id_LSTAT_XLAYER_WAKE as usize;
75const LSTAT_XLAYER_REWAKE: usize = bpf_intf::layer_stat_id_LSTAT_XLAYER_REWAKE as usize;
76const LSTAT_LLC_DRAIN_TRY: usize = bpf_intf::layer_stat_id_LSTAT_LLC_DRAIN_TRY as usize;
77const LSTAT_LLC_DRAIN: usize = bpf_intf::layer_stat_id_LSTAT_LLC_DRAIN as usize;
78const LSTAT_SKIP_REMOTE_NODE: usize = bpf_intf::layer_stat_id_LSTAT_SKIP_REMOTE_NODE as usize;
79
80const LLC_LSTAT_LAT: usize = bpf_intf::llc_layer_stat_id_LLC_LSTAT_LAT as usize;
81const LLC_LSTAT_CNT: usize = bpf_intf::llc_layer_stat_id_LLC_LSTAT_CNT as usize;
82
83fn calc_frac(a: f64, b: f64) -> f64 {
84    if b != 0.0 {
85        a / b * 100.0
86    } else {
87        0.0
88    }
89}
90
91fn fmt_pct(v: f64) -> String {
92    if v >= 99.995 {
93        format!("{:5.1}", v)
94    } else if v > 0.0 && v < 0.01 {
95        format!("{:5.2}", 0.01)
96    } else {
97        format!("{:5.2}", v)
98    }
99}
100
101fn fmt_num(v: u64) -> String {
102    if v > 1_000_000 {
103        format!("{:5.1}m", v as f64 / 1_000_000.0)
104    } else if v > 1_000 {
105        format!("{:5.1}k", v as f64 / 1_000.0)
106    } else {
107        format!("{:5.0} ", v)
108    }
109}
110
111#[stat_doc]
112#[derive(Clone, Debug, Default, Serialize, Deserialize, Stats)]
113#[stat(_om_prefix = "l_", _om_label = "layer_name")]
114pub struct LayerStats {
115    #[stat(desc = "index", _om_skip)]
116    pub index: usize,
117    #[stat(desc = "Total CPU utilization (100% means one full CPU)")]
118    pub util: f64,
119    #[stat(desc = "Protected CPU utilization %")]
120    pub util_protected_frac: f64,
121    #[stat(desc = "Preempt-protected CPU utilization %")]
122    pub util_protected_preempt_frac: f64,
123    #[stat(desc = "Open CPU utilization %")]
124    pub util_open_frac: f64,
125    #[stat(desc = "fraction of total CPU utilization")]
126    pub util_frac: f64,
127    #[stat(desc = "number of tasks")]
128    pub tasks: u32,
129    #[stat(desc = "count of sched events during the period")]
130    pub total: u64,
131    #[stat(desc = "% dispatched into idle CPU from select_cpu")]
132    pub sel_local: f64,
133    #[stat(desc = "% dispatched into idle CPU from enqueue")]
134    pub enq_local: f64,
135    #[stat(desc = "% enqueued after wakeup")]
136    pub enq_wakeup: f64,
137    #[stat(desc = "% enqueued after slice expiration")]
138    pub enq_expire: f64,
139    #[stat(desc = "% re-enqueued due to RT preemption")]
140    pub enq_reenq: f64,
141    #[stat(desc = "% enqueued into the layer's LLC DSQs")]
142    pub enq_dsq: f64,
143    #[stat(desc = "count of times exec duration < min_exec_us")]
144    pub min_exec: f64,
145    #[stat(desc = "total exec durations extended due to min_exec_us")]
146    pub min_exec_us: u64,
147    #[stat(desc = "% dispatched into idle CPUs occupied by other layers")]
148    pub open_idle: f64,
149    #[stat(desc = "% preempted other tasks")]
150    pub preempt: f64,
151    #[stat(desc = "% preempted XLLC tasks")]
152    pub preempt_xllc: f64,
153    #[stat(desc = "% preempted XNUMA tasks")]
154    pub preempt_xnuma: f64,
155    #[stat(desc = "% first-preempted other tasks")]
156    pub preempt_first: f64,
157    #[stat(desc = "% idle-preempted other tasks")]
158    pub preempt_idle: f64,
159    #[stat(desc = "% attempted to preempt other tasks but failed")]
160    pub preempt_fail: f64,
161    #[stat(desc = "% violated config due to CPU affinity")]
162    pub affn_viol: f64,
163    #[stat(desc = "% continued executing after slice expiration")]
164    pub keep: f64,
165    #[stat(desc = "% disallowed to continue executing due to max_exec")]
166    pub keep_fail_max_exec: f64,
167    #[stat(desc = "% disallowed to continue executing due to other tasks")]
168    pub keep_fail_busy: f64,
169    #[stat(desc = "whether is exclusive", _om_skip)]
170    pub is_excl: u32,
171    #[stat(desc = "count of times an excl task skipped a CPU as the sibling was also excl")]
172    pub excl_collision: f64,
173    #[stat(desc = "% a sibling CPU was preempted for an exclusive task")]
174    pub excl_preempt: f64,
175    #[stat(desc = "% yielded")]
176    pub yielded: f64,
177    #[stat(desc = "count of times yield was ignored")]
178    pub yield_ignore: u64,
179    #[stat(desc = "% migrated across CPUs")]
180    pub migration: f64,
181    #[stat(desc = "% migrated across NUMA nodes")]
182    pub xnuma_migration: f64,
183    #[stat(desc = "% migrated across LLCs")]
184    pub xllc_migration: f64,
185    #[stat(desc = "% migration skipped across LLCs due to xllc_mig_min_us")]
186    pub xllc_migration_skip: f64,
187    #[stat(desc = "% wakers across layers")]
188    pub xlayer_wake: f64,
189    #[stat(desc = "% rewakers across layers where waker has waken the task previously")]
190    pub xlayer_rewake: f64,
191    #[stat(desc = "% LLC draining tried")]
192    pub llc_drain_try: f64,
193    #[stat(desc = "% LLC draining succeeded")]
194    pub llc_drain: f64,
195    #[stat(desc = "% skip LLC dispatch on remote node")]
196    pub skip_remote_node: f64,
197    #[stat(desc = "mask of allocated CPUs", _om_skip)]
198    pub cpus: Vec<u64>,
199    #[stat(desc = "count of CPUs assigned")]
200    pub cur_nr_cpus: u32,
201    #[stat(desc = "minimum # of CPUs assigned")]
202    pub min_nr_cpus: u32,
203    #[stat(desc = "maximum # of CPUs assigned")]
204    pub max_nr_cpus: u32,
205    #[stat(desc = "count of CPUs assigned per LLC")]
206    pub nr_llc_cpus: Vec<u32>,
207    #[stat(desc = "slice duration config")]
208    pub slice_us: u64,
209    #[stat(desc = "Per-LLC scheduling event fractions")]
210    pub llc_fracs: Vec<f64>,
211    #[stat(desc = "Per-LLC average latency")]
212    pub llc_lats: Vec<f64>,
213    #[stat(desc = "Layer memory bandwidth as a % of total allowed (0 for \"no limit\"")]
214    pub membw_pct: f64,
215}
216
217impl LayerStats {
218    pub fn new(
219        lidx: usize,
220        layer: &Layer,
221        stats: &Stats,
222        bstats: &BpfStats,
223        nr_cpus_range: (usize, usize),
224    ) -> Self {
225        let lstat = |sidx| bstats.lstats[lidx][sidx];
226        let ltotal = lstat(LSTAT_SEL_LOCAL)
227            + lstat(LSTAT_ENQ_LOCAL)
228            + lstat(LSTAT_ENQ_WAKEUP)
229            + lstat(LSTAT_ENQ_EXPIRE)
230            + lstat(LSTAT_ENQ_REENQ)
231            + lstat(LSTAT_KEEP);
232        let lstat_pct = |sidx| {
233            if ltotal != 0 {
234                lstat(sidx) as f64 / ltotal as f64 * 100.0
235            } else {
236                0.0
237            }
238        };
239
240        let util_sum = stats.layer_utils[lidx]
241            .iter()
242            .take(LAYER_USAGE_SUM_UPTO + 1)
243            .sum::<f64>();
244
245        let membw_frac = match &layer.kind {
246            // Open layer's can't have a memory BW limit.
247            LayerKind::Open { .. } => 0.0,
248            LayerKind::Confined { membw_gb, .. } | LayerKind::Grouped { membw_gb, .. } => {
249                // Check if we have set a memory BW limit.
250                if let Some(membw_limit_gb) = membw_gb {
251                    stats.layer_membws[lidx]
252                        .iter()
253                        .take(LAYER_USAGE_SUM_UPTO + 1)
254                        .sum::<f64>()
255                        / ((*membw_limit_gb * (1024_u64.pow(3) as f64)) as f64)
256                } else {
257                    0.0
258                }
259            }
260        };
261
262        Self {
263            index: lidx,
264            util: util_sum * 100.0,
265            util_open_frac: calc_frac(stats.layer_utils[lidx][LAYER_USAGE_OPEN], util_sum),
266            util_protected_frac: calc_frac(
267                stats.layer_utils[lidx][LAYER_USAGE_PROTECTED],
268                util_sum,
269            ),
270            util_protected_preempt_frac: calc_frac(
271                stats.layer_utils[lidx][LAYER_USAGE_PROTECTED_PREEMPT],
272                util_sum,
273            ),
274            util_frac: calc_frac(util_sum, stats.total_util),
275            tasks: stats.nr_layer_tasks[lidx] as u32,
276            total: ltotal,
277            sel_local: lstat_pct(LSTAT_SEL_LOCAL),
278            enq_local: lstat_pct(LSTAT_ENQ_LOCAL),
279            enq_wakeup: lstat_pct(LSTAT_ENQ_WAKEUP),
280            enq_expire: lstat_pct(LSTAT_ENQ_EXPIRE),
281            enq_reenq: lstat_pct(LSTAT_ENQ_REENQ),
282            enq_dsq: lstat_pct(LSTAT_ENQ_DSQ),
283            min_exec: lstat_pct(LSTAT_MIN_EXEC),
284            min_exec_us: (lstat(LSTAT_MIN_EXEC_NS) / 1000) as u64,
285            open_idle: lstat_pct(LSTAT_OPEN_IDLE),
286            preempt: lstat_pct(LSTAT_PREEMPT),
287            preempt_xllc: lstat_pct(LSTAT_PREEMPT_XLLC),
288            preempt_xnuma: lstat_pct(LSTAT_PREEMPT_XNUMA),
289            preempt_first: lstat_pct(LSTAT_PREEMPT_FIRST),
290            preempt_idle: lstat_pct(LSTAT_PREEMPT_IDLE),
291            preempt_fail: lstat_pct(LSTAT_PREEMPT_FAIL),
292            affn_viol: lstat_pct(LSTAT_AFFN_VIOL),
293            keep: lstat_pct(LSTAT_KEEP),
294            keep_fail_max_exec: lstat_pct(LSTAT_KEEP_FAIL_MAX_EXEC),
295            keep_fail_busy: lstat_pct(LSTAT_KEEP_FAIL_BUSY),
296            is_excl: layer.kind.common().exclusive as u32,
297            excl_collision: lstat_pct(LSTAT_EXCL_COLLISION),
298            excl_preempt: lstat_pct(LSTAT_EXCL_PREEMPT),
299            yielded: lstat_pct(LSTAT_YIELD),
300            yield_ignore: lstat(LSTAT_YIELD_IGNORE) as u64,
301            migration: lstat_pct(LSTAT_MIGRATION),
302            xnuma_migration: lstat_pct(LSTAT_XNUMA_MIGRATION),
303            xlayer_wake: lstat_pct(LSTAT_XLAYER_WAKE),
304            xlayer_rewake: lstat_pct(LSTAT_XLAYER_REWAKE),
305            xllc_migration: lstat_pct(LSTAT_XLLC_MIGRATION),
306            xllc_migration_skip: lstat_pct(LSTAT_XLLC_MIGRATION_SKIP),
307            llc_drain_try: lstat_pct(LSTAT_LLC_DRAIN_TRY),
308            llc_drain: lstat_pct(LSTAT_LLC_DRAIN),
309            skip_remote_node: lstat_pct(LSTAT_SKIP_REMOTE_NODE),
310            cpus: layer.cpus.as_raw_slice().to_vec(),
311            cur_nr_cpus: layer.cpus.weight() as u32,
312            min_nr_cpus: nr_cpus_range.0 as u32,
313            max_nr_cpus: nr_cpus_range.1 as u32,
314            nr_llc_cpus: layer.nr_llc_cpus.iter().map(|&v| v as u32).collect(),
315            slice_us: stats.layer_slice_us[lidx],
316            llc_fracs: {
317                let sid = LLC_LSTAT_CNT;
318                let sum = bstats.llc_lstats[lidx]
319                    .iter()
320                    .map(|lstats| lstats[sid])
321                    .sum::<u64>() as f64;
322                bstats.llc_lstats[lidx]
323                    .iter()
324                    .map(|lstats| calc_frac(lstats[sid] as f64, sum))
325                    .collect()
326            },
327            llc_lats: bstats.llc_lstats[lidx]
328                .iter()
329                .map(|lstats| lstats[LLC_LSTAT_LAT] as f64 / 1_000_000_000.0)
330                .collect(),
331            membw_pct: membw_frac * 100.0,
332        }
333    }
334
335    pub fn format<W: Write>(&self, w: &mut W, name: &str, header_width: usize) -> Result<()> {
336        writeln!(
337            w,
338            "  {:<width$}: util/open/frac={:6.1}/{}/{:7.1} prot/prot_preempt={}/{} tasks={:6}",
339            name,
340            self.util,
341            fmt_pct(self.util_open_frac),
342            self.util_frac,
343            fmt_pct(self.util_protected_frac),
344            fmt_pct(self.util_protected_preempt_frac),
345            self.tasks,
346            width = header_width,
347        )?;
348
349        writeln!(
350            w,
351            "  {:<width$}  tot={:7} local_sel/enq={}/{} enq_dsq={} wake/exp/reenq={}/{}/{}",
352            "",
353            self.total,
354            fmt_pct(self.sel_local),
355            fmt_pct(self.enq_local),
356            fmt_pct(self.enq_dsq),
357            fmt_pct(self.enq_wakeup),
358            fmt_pct(self.enq_expire),
359            fmt_pct(self.enq_reenq),
360            width = header_width,
361        )?;
362
363        writeln!(
364            w,
365            "  {:<width$}  keep/max/busy={}/{}/{} yield/ign={}/{}",
366            "",
367            fmt_pct(self.keep),
368            fmt_pct(self.keep_fail_max_exec),
369            fmt_pct(self.keep_fail_busy),
370            fmt_pct(self.yielded),
371            fmt_num(self.yield_ignore),
372            width = header_width,
373        )?;
374
375        writeln!(
376            w,
377            "  {:<width$}  open_idle={} mig={} xnuma_mig={} xllc_mig/skip={}/{} affn_viol={}",
378            "",
379            fmt_pct(self.open_idle),
380            fmt_pct(self.migration),
381            fmt_pct(self.xnuma_migration),
382            fmt_pct(self.xllc_migration),
383            fmt_pct(self.xllc_migration_skip),
384            fmt_pct(self.affn_viol),
385            width = header_width,
386        )?;
387
388        writeln!(
389            w,
390            "  {:<width$}  preempt/first/xllc/xnuma/idle/fail={}/{}/{}/{}/{}/{}",
391            "",
392            fmt_pct(self.preempt),
393            fmt_pct(self.preempt_first),
394            fmt_pct(self.preempt_xllc),
395            fmt_pct(self.preempt_xnuma),
396            fmt_pct(self.preempt_idle),
397            fmt_pct(self.preempt_fail),
398            width = header_width,
399        )?;
400
401        writeln!(
402            w,
403            "  {:<width$}  xlayer_wake/re={}/{} llc_drain/try={}/{} skip_rnode={}",
404            "",
405            fmt_pct(self.xlayer_wake),
406            fmt_pct(self.xlayer_rewake),
407            fmt_pct(self.llc_drain),
408            fmt_pct(self.llc_drain_try),
409            fmt_pct(self.skip_remote_node),
410            width = header_width,
411        )?;
412
413        writeln!(
414            w,
415            "  {:<width$}  slice={}ms min_exec={}/{:7.2}ms",
416            "",
417            self.slice_us as f64 / 1000.0,
418            fmt_pct(self.min_exec),
419            self.min_exec_us as f64 / 1000.0,
420            width = header_width
421        )?;
422
423        let cpumask = Cpumask::from_vec(self.cpus.clone());
424
425        writeln!(
426            w,
427            "  {:<width$}  cpus={:3} [{:3},{:3}] {}",
428            "",
429            self.cur_nr_cpus,
430            self.min_nr_cpus,
431            self.max_nr_cpus,
432            &cpumask,
433            width = header_width
434        )?;
435
436        write!(
437            w,
438            "  {:<width$}  [LLC] nr_cpus: sched% lat_ms",
439            "",
440            width = header_width
441        )?;
442
443        for (i, (&frac, &lat)) in self.llc_fracs.iter().zip(self.llc_lats.iter()).enumerate() {
444            if (i % 4) == 0 {
445                writeln!(w, "")?;
446                write!(w, "  {:<width$}  [{:03}]", "", i, width = header_width)?;
447            } else {
448                write!(w, " |")?;
449            }
450            write!(
451                w,
452                " {:2}:{}%{:7.2}",
453                self.nr_llc_cpus[i],
454                fmt_pct(frac),
455                lat * 1_000.0
456            )?;
457        }
458        writeln!(w, "")?;
459
460        if self.is_excl != 0 {
461            writeln!(
462                w,
463                "  {:<width$}  excl_coll={} excl_preempt={}",
464                "",
465                fmt_pct(self.excl_collision),
466                fmt_pct(self.excl_preempt),
467                width = header_width,
468            )?;
469        } else if self.excl_collision != 0.0 || self.excl_preempt != 0.0 {
470            warn!(
471                "{}: exclusive is off but excl_coll={} excl_preempt={}",
472                name,
473                fmt_pct(self.excl_collision),
474                fmt_pct(self.excl_preempt),
475            );
476        }
477
478        Ok(())
479    }
480}
481
482#[stat_doc]
483#[derive(Clone, Debug, Default, Serialize, Deserialize, Stats)]
484#[stat(top)]
485pub struct SysStats {
486    #[stat(desc = "timestamp", _om_skip)]
487    pub at: f64,
488    #[stat(desc = "# of NUMA nodes")]
489    pub nr_nodes: usize,
490    #[stat(desc = "# sched events during the period")]
491    pub total: u64,
492    #[stat(desc = "% dispatched directly into an idle CPU from select_cpu")]
493    pub local_sel: f64,
494    #[stat(desc = "% dispatched directly into an idle CPU from enqueue")]
495    pub local_enq: f64,
496    #[stat(desc = "% open layer tasks scheduled into allocated but idle CPUs")]
497    pub open_idle: f64,
498    #[stat(desc = "% violated config due to CPU affinity")]
499    pub affn_viol: f64,
500    #[stat(desc = "% sent to hi fallback DSQs")]
501    pub hi_fb: f64,
502    #[stat(desc = "% sent to lo fallback DSQs")]
503    pub lo_fb: f64,
504    #[stat(desc = "count of times an excl task skipped a CPU as the sibling was also excl")]
505    pub excl_collision: f64,
506    #[stat(desc = "count of times a sibling CPU was preempted for an excl task")]
507    pub excl_preempt: f64,
508    #[stat(desc = "count of times a CPU skipped dispatching due to an excl task on the sibling")]
509    pub excl_idle: f64,
510    #[stat(
511        desc = "count of times an idle sibling CPU was woken up after an excl task is finished"
512    )]
513    pub excl_wakeup: f64,
514    #[stat(desc = "CPU time this binary consumed during the period")]
515    pub proc_ms: u64,
516    #[stat(desc = "CPU busy % (100% means all CPU)")]
517    pub busy: f64,
518    #[stat(desc = "CPU util % (100% means one CPU)")]
519    pub util: f64,
520    #[stat(desc = "CPU util % used by hi fallback DSQs")]
521    pub hi_fb_util: f64,
522    #[stat(desc = "CPU util % used by lo fallback DSQs")]
523    pub lo_fb_util: f64,
524    #[stat(desc = "Number of tasks dispatched via antistall")]
525    pub antistall: u64,
526    #[stat(desc = "Number of times preemptions of non-scx tasks were avoided")]
527    pub skip_preempt: u64,
528    #[stat(desc = "Number of times vtime was out of range and fixed up")]
529    pub fixup_vtime: u64,
530    #[stat(desc = "Number of times cpuc->preempting_task didn't come on the CPU")]
531    pub preempting_mismatch: u64,
532    #[stat(desc = "fallback CPU")]
533    pub fallback_cpu: u32,
534    #[stat(desc = "per-layer statistics")]
535    pub fallback_cpu_util: f64,
536    #[stat(desc = "fallback CPU util %")]
537    pub layers: BTreeMap<String, LayerStats>,
538    #[stat(desc = "Number of gpu tasks affinitized since scheduler start")]
539    pub gpu_tasks_affinitized: u64,
540    #[stat(desc = "Time (in ms) of last affinitization run.")]
541    pub gpu_task_affinitization_ms: u64,
542}
543
544impl SysStats {
545    pub fn new(stats: &Stats, bstats: &BpfStats, fallback_cpu: usize) -> Result<Self> {
546        let lsum = |idx| stats.bpf_stats.lstats_sums[idx];
547        let total = lsum(LSTAT_SEL_LOCAL)
548            + lsum(LSTAT_ENQ_LOCAL)
549            + lsum(LSTAT_ENQ_WAKEUP)
550            + lsum(LSTAT_ENQ_EXPIRE)
551            + lsum(LSTAT_ENQ_REENQ)
552            + lsum(LSTAT_KEEP);
553        let lsum_pct = |idx| {
554            if total != 0 {
555                lsum(idx) as f64 / total as f64 * 100.0
556            } else {
557                0.0
558            }
559        };
560
561        let elapsed_ns = stats.elapsed.as_nanos();
562
563        Ok(Self {
564            at: SystemTime::now().duration_since(UNIX_EPOCH)?.as_secs_f64(),
565            nr_nodes: stats.nr_nodes,
566            total,
567            local_sel: lsum_pct(LSTAT_SEL_LOCAL),
568            local_enq: lsum_pct(LSTAT_ENQ_LOCAL),
569            open_idle: lsum_pct(LSTAT_OPEN_IDLE),
570            affn_viol: lsum_pct(LSTAT_AFFN_VIOL),
571            hi_fb: calc_frac(
572                stats.bpf_stats.gstats[GSTAT_HI_FB_EVENTS] as f64,
573                total as f64,
574            ),
575            lo_fb: calc_frac(
576                stats.bpf_stats.gstats[GSTAT_LO_FB_EVENTS] as f64,
577                total as f64,
578            ),
579            excl_collision: lsum_pct(LSTAT_EXCL_COLLISION),
580            excl_preempt: lsum_pct(LSTAT_EXCL_PREEMPT),
581            excl_idle: bstats.gstats[GSTAT_EXCL_IDLE] as f64 / total as f64,
582            excl_wakeup: bstats.gstats[GSTAT_EXCL_WAKEUP] as f64 / total as f64,
583            proc_ms: stats.processing_dur.as_millis() as u64,
584            busy: stats.cpu_busy * 100.0,
585            util: stats.total_util * 100.0,
586            hi_fb_util: stats.bpf_stats.gstats[GSTAT_HI_FB_USAGE] as f64 / elapsed_ns as f64
587                * 100.0,
588            lo_fb_util: stats.bpf_stats.gstats[GSTAT_LO_FB_USAGE] as f64 / elapsed_ns as f64
589                * 100.0,
590            antistall: stats.bpf_stats.gstats[GSTAT_ANTISTALL],
591            skip_preempt: stats.bpf_stats.gstats[GSTAT_SKIP_PREEMPT],
592            fixup_vtime: stats.bpf_stats.gstats[GSTAT_FIXUP_VTIME],
593            preempting_mismatch: stats.bpf_stats.gstats[GSTAT_PREEMPTING_MISMATCH],
594            fallback_cpu: fallback_cpu as u32,
595            fallback_cpu_util: stats.bpf_stats.gstats[GSTAT_FB_CPU_USAGE] as f64
596                / elapsed_ns as f64
597                * 100.0,
598            layers: BTreeMap::new(),
599            gpu_tasks_affinitized: stats.gpu_tasks_affinitized,
600            gpu_task_affinitization_ms: stats.gpu_task_affinitization_ms,
601        })
602    }
603
604    pub fn format<W: Write>(&self, w: &mut W) -> Result<()> {
605        writeln!(
606            w,
607            "tot={:7} local_sel/enq={}/{} open_idle={} affn_viol={} hi/lo={}/{}",
608            self.total,
609            fmt_pct(self.local_sel),
610            fmt_pct(self.local_enq),
611            fmt_pct(self.open_idle),
612            fmt_pct(self.affn_viol),
613            fmt_pct(self.hi_fb),
614            fmt_pct(self.lo_fb),
615        )?;
616
617        writeln!(
618            w,
619            "busy={:5.1} util/hi/lo={:7.1}/{}/{} fallback_cpu/util={:3}/{:4.1} proc={:?}ms",
620            self.busy,
621            self.util,
622            fmt_pct(self.hi_fb_util),
623            fmt_pct(self.lo_fb_util),
624            self.fallback_cpu,
625            self.fallback_cpu_util,
626            self.proc_ms,
627        )?;
628
629        writeln!(
630            w,
631            "excl_coll={:.2} excl_preempt={:.2} excl_idle={:.2} excl_wakeup={:.2}",
632            self.excl_collision, self.excl_preempt, self.excl_idle, self.excl_wakeup
633        )?;
634
635        writeln!(
636            w,
637            "skip_preempt={} antistall={} fixup_vtime={} preempting_mismatch={}",
638            self.skip_preempt, self.antistall, self.fixup_vtime, self.preempting_mismatch
639        )?;
640
641        writeln!(
642            w,
643            "gpu_tasks_affinitized={} gpu_task_affinitization_time={}",
644            self.gpu_tasks_affinitized, self.gpu_task_affinitization_ms
645        )?;
646
647        Ok(())
648    }
649
650    pub fn format_all<W: Write>(&self, w: &mut W) -> Result<()> {
651        self.format(w)?;
652
653        let header_width = self
654            .layers
655            .keys()
656            .map(|name| name.len())
657            .max()
658            .unwrap_or(0)
659            .max(4);
660
661        let mut idx_to_name: Vec<(usize, &String)> =
662            self.layers.iter().map(|(k, v)| (v.index, k)).collect();
663
664        idx_to_name.sort();
665
666        for (_idx, name) in &idx_to_name {
667            self.layers[*name].format(w, name, header_width)?;
668        }
669
670        Ok(())
671    }
672}
673
674#[derive(Debug)]
675pub enum StatsReq {
676    Hello(ThreadId),
677    Refresh(ThreadId, Stats),
678    Bye(ThreadId),
679}
680
681#[derive(Debug)]
682pub enum StatsRes {
683    Hello(Stats),
684    Refreshed((Stats, SysStats)),
685    Bye,
686}
687
688pub fn server_data() -> StatsServerData<StatsReq, StatsRes> {
689    let open: Box<dyn StatsOpener<StatsReq, StatsRes>> = Box::new(move |(req_ch, res_ch)| {
690        let tid = current().id();
691        req_ch.send(StatsReq::Hello(tid))?;
692        let mut stats = Some(match res_ch.recv()? {
693            StatsRes::Hello(v) => v,
694            res => bail!("invalid response to Hello: {:?}", res),
695        });
696
697        let read: Box<dyn StatsReader<StatsReq, StatsRes>> =
698            Box::new(move |_args, (req_ch, res_ch)| {
699                req_ch.send(StatsReq::Refresh(tid, stats.take().unwrap()))?;
700                let (new_stats, sys_stats) = match res_ch.recv()? {
701                    StatsRes::Refreshed(v) => v,
702                    res => bail!("invalid response to Refresh: {:?}", res),
703                };
704                stats = Some(new_stats);
705                sys_stats.to_json()
706            });
707
708        Ok(read)
709    });
710
711    let close: Box<dyn StatsCloser<StatsReq, StatsRes>> = Box::new(move |(req_ch, res_ch)| {
712        req_ch.send(StatsReq::Bye(current().id())).unwrap();
713        match res_ch.recv().unwrap() {
714            StatsRes::Bye => {}
715            res => panic!("invalid response to Bye: {:?}", res),
716        }
717    });
718
719    StatsServerData::new()
720        .add_meta(LayerStats::meta())
721        .add_meta(SysStats::meta())
722        .add_ops(
723            "top",
724            StatsOps {
725                open,
726                close: Some(close),
727            },
728        )
729}
730
731pub fn monitor(intv: Duration, shutdown: Arc<AtomicBool>) -> Result<()> {
732    scx_utils::monitor_stats::<SysStats>(
733        &vec![],
734        intv,
735        || shutdown.load(Ordering::Relaxed),
736        |sst| {
737            let dt = DateTime::<Local>::from(UNIX_EPOCH + Duration::from_secs_f64(sst.at));
738            println!("###### {} ######", dt.to_rfc2822());
739            sst.format_all(&mut std::io::stdout())
740        },
741    )
742}