Implement performance monitor

Cargo.lock

@@ -3197,7 +3197,7 @@ checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
 
 [[package]]
 name = "hello-agent"
-version = "0.1.6"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "env_logger 0.10.2",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "hello-agent"
-version = "0.1.6"
+version = "0.1.7"
 edition = "2021"
 rust-version = "1.75"
 description = "Headless RustDesk-protocol-compatible support agent for Windows"

src/main.rs

@@ -30,6 +30,8 @@ mod cm_popup;
 #[cfg(target_os = "windows")]
 mod exec;
 mod login_events;
+mod perf;
+mod perf_events;
 #[cfg(target_os = "windows")]
 mod service;
 #[cfg(target_os = "windows")]

src/perf.rs

@@ -0,0 +1,311 @@
+// Continuous performance sampling.
+//
+// One sample per minute: overall CPU%, memory used / total, top
+// process by CPU, top process by memory, uptime, process count.
+// Posted in batches to /api/agent/metrics; surfaced on the admin
+// device detail page as a 24 h sparkline plus a "live" snapshot card.
+//
+// Architecture mirrors `login_events`:
+//   * One background thread, its own current-thread Tokio runtime.
+//   * Sysinfo crate (vendored under hbb_common) does the cross-cutting
+//     work — we keep one `System` instance alive across iterations so
+//     its per-process CPU% accounting (which is differential against
+//     the previous refresh) stays accurate.
+//   * In-memory queue with retry-with-backoff and a hard drop cap so
+//     a permanently-misconfigured agent can't balloon memory.
+//   * Server-side `INSERT OR IGNORE` keyed on (peer_id, at) dedups any
+//     retries that get there twice.
+//
+// Process CPU% on Windows is reported by sysinfo as "% of one core",
+// so a busy multi-threaded process can read >100%. We normalise by
+// `cpus().len()` so the snapshot card's "top CPU: chrome.exe 18%"
+// is comparable to the overall CPU%. Memory is in bytes from sysinfo;
+// we convert to MB on the wire.
+
+use anyhow::{anyhow, Result};
+use std::sync::Mutex;
+use std::time::Duration;
+
+/// Sampling cadence. 60 s strikes a balance between resolution (enough
+/// granularity to spot a 2-minute CPU spike) and storage (~1440 rows /
+/// device / day on the server side).
+const SAMPLE_INTERVAL: Duration = Duration::from_secs(60);
+
+/// Flush cadence on the happy path. The reporter tries to flush after
+/// every sample anyway; this is the floor used by the network-error
+/// backoff before it doubles.
+const FLUSH_INTERVAL_BASE: Duration = Duration::from_secs(60);
+const FLUSH_INTERVAL_MAX: Duration = Duration::from_secs(15 * 60);
+
+/// At backoff cap = 15 min, this is ~3 days of trying — enough to ride
+/// out a long server-side outage. Beyond that we drop, on the same
+/// reasoning as `login_events::DROP_AFTER`.
+const DROP_AFTER: u32 = 300;
+
+/// Must match the server's MAX_SAMPLES_PER_POST.
+const MAX_SAMPLES_PER_POST: usize = 512;
+
+#[derive(Clone, Debug, Default)]
+struct PendingSample {
+    at: i64,
+    cpu_pct: f64,
+    mem_used_mb: i64,
+    mem_total_mb: i64,
+    proc_count: i64,
+    uptime_secs: i64,
+    top_cpu_name: String,
+    top_cpu_pct: f64,
+    top_mem_name: String,
+    top_mem_mb: i64,
+    attempts: u32,
+}
+
+#[cfg(target_os = "windows")]
+static QUEUE: Mutex<Vec<PendingSample>> = Mutex::new(Vec::new());
+
+/// Kick off the metrics sampler. Safe to call multiple times — guarded
+/// by an `AtomicBool` so a stray second call is a no-op.
+pub fn start() {
+    #[cfg(not(target_os = "windows"))]
+    {
+        // Cross-platform stub; the implementation only runs on Windows
+        // because that's the only OS hello-agent ships on. Sysinfo is
+        // cross-platform, so a future Linux build can drop the cfg-gate
+        // and reuse this module unchanged.
+    }
+    #[cfg(target_os = "windows")]
+    {
+        use std::sync::atomic::{AtomicBool, Ordering};
+        static STARTED: AtomicBool = AtomicBool::new(false);
+        if STARTED.swap(true, Ordering::SeqCst) {
+            return;
+        }
+        std::thread::spawn(move || {
+            let rt = match tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+            {
+                Ok(rt) => rt,
+                Err(e) => {
+                    log::warn!("perf: build runtime: {e}");
+                    return;
+                }
+            };
+            rt.block_on(run_loop());
+        });
+    }
+}
+
+#[cfg(target_os = "windows")]
+async fn run_loop() {
+    use hbb_common::sysinfo::System;
+
+    // One System instance, refreshed across iterations. Sysinfo's
+    // per-process CPU% is computed against the previous refresh, so a
+    // throwaway `System::new()` each tick would always read 0%.
+    let mut sys = System::new();
+    // Prime CPU + processes once so the first real sample isn't NaN.
+    // The MINIMUM_CPU_UPDATE_INTERVAL sleep is what sysinfo's docs
+    // recommend between back-to-back CPU refreshes; here we have the
+    // full SAMPLE_INTERVAL coming up so we just refresh-and-wait.
+    sys.refresh_cpu();
+    sys.refresh_processes();
+    tokio::time::sleep(SAMPLE_INTERVAL).await;
+
+    let mut flush_backoff = FLUSH_INTERVAL_BASE;
+    loop {
+        if let Some(sample) = collect_sample(&mut sys) {
+            QUEUE.lock().unwrap().push(sample);
+        }
+
+        match flush_once().await {
+            FlushOutcome::Idle | FlushOutcome::AllSent => {
+                flush_backoff = FLUSH_INTERVAL_BASE;
+            }
+            FlushOutcome::Failed => {
+                flush_backoff = (flush_backoff * 2).min(FLUSH_INTERVAL_MAX);
+            }
+        }
+
+        // We always sleep SAMPLE_INTERVAL — the network backoff slows
+        // the *retry* of pending events, not observation. Sampling
+        // continues at full cadence so a transient outage doesn't punch
+        // a hole in the chart on either side of the dropped window.
+        tokio::time::sleep(SAMPLE_INTERVAL.min(flush_backoff)).await;
+    }
+}
+
+#[cfg(target_os = "windows")]
+fn collect_sample(sys: &mut hbb_common::sysinfo::System) -> Option<PendingSample> {
+    sys.refresh_cpu();
+    sys.refresh_memory();
+    sys.refresh_processes();
+
+    let at = hbb_common::chrono::Utc::now().timestamp();
+    let cpu_pct = sys.global_cpu_info().cpu_usage() as f64;
+    // Sysinfo returns bytes on every platform; the server schema stores
+    // MB to keep row sizes small at scale. Divide-then-cast bounds
+    // arithmetic to u64 territory.
+    let mem_total_mb = (sys.total_memory() / 1024 / 1024) as i64;
+    let mem_used_mb = (sys.used_memory() / 1024 / 1024) as i64;
+    let uptime_secs = sys.uptime() as i64;
+
+    // Per-process CPU% from sysinfo is normalised per core: a 4-thread
+    // process pinning 4 cores on a 4-core machine reads 400%. Divide
+    // by core count so the snapshot card's number matches the overall
+    // CPU% reading (both 0-100 of the whole machine).
+    let cpu_count = sys.cpus().len().max(1) as f64;
+
+    let mut top_cpu: Option<(&str, f32)> = None;
+    let mut top_mem: Option<(&str, u64)> = None;
+    let mut proc_count = 0i64;
+    for proc in sys.processes().values() {
+        proc_count += 1;
+        let name = proc.name();
+        // Some kernel-side rows show up with empty names on Windows;
+        // skip them so we don't ever render a top-CPU row with no
+        // label.
+        if name.is_empty() {
+            continue;
+        }
+        let cu = proc.cpu_usage();
+        if cu.is_finite() && cu > top_cpu.map(|(_, v)| v).unwrap_or(0.0) {
+            top_cpu = Some((name, cu));
+        }
+        let mu = proc.memory();
+        if mu > top_mem.map(|(_, v)| v).unwrap_or(0) {
+            top_mem = Some((name, mu));
+        }
+    }
+
+    let (top_cpu_name, top_cpu_pct) = top_cpu
+        .map(|(n, v)| (n.to_string(), (v as f64 / cpu_count).min(100.0)))
+        .unwrap_or_default();
+    let (top_mem_name, top_mem_mb) = top_mem
+        .map(|(n, v)| (n.to_string(), (v / 1024 / 1024) as i64))
+        .unwrap_or_default();
+
+    Some(PendingSample {
+        at,
+        cpu_pct,
+        mem_used_mb,
+        mem_total_mb,
+        proc_count,
+        uptime_secs,
+        top_cpu_name,
+        top_cpu_pct,
+        top_mem_name,
+        top_mem_mb,
+        attempts: 0,
+    })
+}
+
+#[cfg(target_os = "windows")]
+enum FlushOutcome {
+    Idle,
+    AllSent,
+    Failed,
+}
+
+#[cfg(target_os = "windows")]
+async fn flush_once() -> FlushOutcome {
+    let batch: Vec<PendingSample> = {
+        let mut q = QUEUE.lock().unwrap();
+        if q.is_empty() {
+            return FlushOutcome::Idle;
+        }
+        let take = q.len().min(MAX_SAMPLES_PER_POST);
+        q.drain(..take).collect()
+    };
+
+    match post_batch(&batch).await {
+        Ok(()) => FlushOutcome::AllSent,
+        Err(e) => {
+            log::warn!(
+                "perf: flush of {} sample(s) failed: {e:#}",
+                batch.len(),
+            );
+            let mut requeued: Vec<PendingSample> = batch
+                .into_iter()
+                .filter_map(|mut s| {
+                    s.attempts = s.attempts.saturating_add(1);
+                    if s.attempts >= DROP_AFTER {
+                        log::warn!(
+                            "perf: dropping sample after {} attempts: at={}",
+                            s.attempts, s.at,
+                        );
+                        None
+                    } else {
+                        Some(s)
+                    }
+                })
+                .collect();
+            let mut q = QUEUE.lock().unwrap();
+            let tail: Vec<PendingSample> = q.drain(..).collect();
+            requeued.extend(tail);
+            *q = requeued;
+            FlushOutcome::Failed
+        }
+    }
+}
+
+#[cfg(target_os = "windows")]
+async fn post_batch(batch: &[PendingSample]) -> Result<()> {
+    let api = librustdesk::common::get_api_server(
+        hbb_common::config::Config::get_option("api-server"),
+        hbb_common::config::Config::get_option("custom-rendezvous-server"),
+    );
+    if api.is_empty() {
+        return Err(anyhow!("no api-server configured yet"));
+    }
+
+    let url = format!("{api}/api/agent/metrics");
+    let id = hbb_common::config::Config::get_id();
+    let uuid = librustdesk::common::encode64(hbb_common::get_uuid());
+
+    let samples: Vec<hbb_common::serde_json::Value> = batch
+        .iter()
+        .map(|s| {
+            hbb_common::serde_json::json!({
+                "at":           s.at,
+                "cpu_pct":      s.cpu_pct,
+                "mem_used_mb":  s.mem_used_mb,
+                "mem_total_mb": s.mem_total_mb,
+                "proc_count":   s.proc_count,
+                "uptime_secs":  s.uptime_secs,
+                "top_cpu_name": s.top_cpu_name,
+                "top_cpu_pct":  s.top_cpu_pct,
+                "top_mem_name": s.top_mem_name,
+                "top_mem_mb":   s.top_mem_mb,
+            })
+        })
+        .collect();
+    let body = hbb_common::serde_json::json!({
+        "id":      id,
+        "uuid":    uuid,
+        "samples": samples,
+    })
+    .to_string();
+
+    let headers = librustdesk::hbbs_http::sign::build_signed_headers(
+        "POST",
+        "/api/agent/metrics",
+        body.as_bytes(),
+    )
+    .unwrap_or_default();
+
+    let resp = librustdesk::common::post_request(url, body, &headers)
+        .await
+        .map_err(|e| anyhow!("post: {e}"))?;
+    let trimmed = resp.trim();
+    if trimmed == "OK" || trimmed == "ID_NOT_FOUND" {
+        // ID_NOT_FOUND mirrors the unattended_password / login_events
+        // contract: server doesn't know the peer yet (rendezvous race
+        // on first boot). We drop the batch rather than retry forever;
+        // the next sample lands once /api/heartbeat has created the
+        // peer row.
+        Ok(())
+    } else {
+        Err(anyhow!("unexpected response: {trimmed}"))
+    }
+}

src/perf_events.rs

@@ -0,0 +1,507 @@
+// Windows event-log scraper for performance-related events.
+//
+// Pulls fresh entries from three channels Microsoft itself uses to
+// flag "the OS noticed this machine was slow":
+//
+//   * `Microsoft-Windows-Diagnostics-Performance/Operational` — boot,
+//     shutdown, standby, resume degradation (Event IDs 100/200/300/400
+//     family), each carrying total time + the component that caused it.
+//   * `Microsoft-Windows-Resource-Exhaustion-Detector/Operational` —
+//     Event ID 2004 fires when virtual memory hits the wall, with the
+//     top processes by working set.
+//   * `System` — IDs 41 (Kernel-Power unexpected reboot), 6008 (dirty
+//     shutdown), 1001 (BugCheck / BSOD).
+//
+// Cadence is intentionally low (5 min): these events are sparse — a
+// healthy machine may produce zero per day, a sick one a handful.
+//
+// State machine per channel: a numeric RecordId cursor persisted in
+// the agent config (`Config::set_option`). First run with no cursor
+// pulls the trailing 7 days bounded by `MAX_FIRST_RUN_EVENTS`; every
+// subsequent run pulls everything with `EventRecordID > cursor`. The
+// cursor advances on observation, not on successful POST — see the
+// `flush_once` comment for the tradeoff.
+
+#![cfg_attr(not(target_os = "windows"), allow(dead_code))]
+
+use anyhow::{anyhow, Result};
+use serde::Deserialize;
+use std::sync::Mutex;
+use std::time::Duration;
+
+/// How often to scrape. 5 min keeps the UI freshness reasonable
+/// (operators looking at a complaint won't usually see a stale view)
+/// without burning host CPU on PowerShell startup every minute.
+const SCRAPE_INTERVAL: Duration = Duration::from_secs(5 * 60);
+
+const FLUSH_INTERVAL_BASE: Duration = Duration::from_secs(60);
+const FLUSH_INTERVAL_MAX: Duration = Duration::from_secs(15 * 60);
+
+/// First-run lookback. A box that's been alive for months would
+/// otherwise dump thousands of `Diagnostics-Performance` events on a
+/// fresh install — useful in theory, but the UI shows the most recent
+/// 20 and the older entries are mostly noise.
+const MAX_FIRST_RUN_EVENTS: u32 = 100;
+
+/// Per-scrape cap so a misbehaving event log can't blow up an
+/// individual run. Anything beyond this on a single pass is dropped on
+/// the floor and picked up on the next scrape (cursor advances to the
+/// last seen, so we don't oscillate).
+const MAX_PER_SCRAPE: u32 = 200;
+
+/// Must match the server's MAX_EVENTS_PER_POST.
+const MAX_EVENTS_PER_POST: usize = 128;
+
+/// Drop pending events after this many retries — at 15 min cap this
+/// is ~5 days, plenty for any realistic outage window.
+const DROP_AFTER: u32 = 480;
+
+/// One channel config: the WEL log name, the short provider tag stored
+/// server-side (matches the `devices.perf_src_*` i18n keys), and an
+/// optional ID allow-list. `None` means "everything in this channel";
+/// `Some(&[…])` restricts to those event IDs (used for `System` to
+/// avoid pulling boot / service-start chatter).
+struct ChannelCfg {
+    provider: &'static str,
+    log_name: &'static str,
+    event_ids: Option<&'static [u32]>,
+}
+
+const CHANNELS: &[ChannelCfg] = &[
+    ChannelCfg {
+        provider: "diag-perf",
+        log_name: "Microsoft-Windows-Diagnostics-Performance/Operational",
+        event_ids: None,
+    },
+    ChannelCfg {
+        provider: "res-exh",
+        log_name: "Microsoft-Windows-Resource-Exhaustion-Detector/Operational",
+        event_ids: None,
+    },
+    ChannelCfg {
+        // Subset of `System` that's actually performance-relevant —
+        // 41 = Kernel-Power unexpected reboot, 6008 = dirty shutdown,
+        // 1001 = BugCheck (BSOD report). Adding more IDs is just a
+        // matter of extending this slice.
+        provider: "system",
+        log_name: "System",
+        event_ids: Some(&[41, 6008, 1001]),
+    },
+];
+
+/// JSON shape we ask PowerShell to emit. Mirrors `PerfEventIn` on the
+/// server.
+#[derive(Debug, Clone, Deserialize)]
+struct RawEvent {
+    at: i64,
+    event_id: i64,
+    level: i64,
+    record_id: i64,
+    #[serde(default)]
+    summary: String,
+    #[serde(default)]
+    detail: String,
+}
+
+#[derive(Clone, Debug)]
+struct PendingPerfEvent {
+    provider: &'static str,
+    at: i64,
+    event_id: i64,
+    level: i64,
+    record_id: i64,
+    summary: String,
+    detail: String,
+    attempts: u32,
+}
+
+#[cfg(target_os = "windows")]
+static QUEUE: Mutex<Vec<PendingPerfEvent>> = Mutex::new(Vec::new());
+
+pub fn start() {
+    #[cfg(not(target_os = "windows"))]
+    {}
+    #[cfg(target_os = "windows")]
+    {
+        use std::sync::atomic::{AtomicBool, Ordering};
+        static STARTED: AtomicBool = AtomicBool::new(false);
+        if STARTED.swap(true, Ordering::SeqCst) {
+            return;
+        }
+        std::thread::spawn(move || {
+            let rt = match tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+            {
+                Ok(rt) => rt,
+                Err(e) => {
+                    log::warn!("perf-events: build runtime: {e}");
+                    return;
+                }
+            };
+            rt.block_on(run_loop());
+        });
+    }
+}
+
+#[cfg(target_os = "windows")]
+async fn run_loop() {
+    // Stagger the first scrape so we don't pile on the
+    // unattended_password POST + login_events first poll + sysinfo
+    // upload all at the same boot moment. PowerShell first invocation
+    // is ~1 s of CPU; doing it 30 s in instead of immediately is
+    // friendlier on cold-boot CPU load.
+    tokio::time::sleep(Duration::from_secs(30)).await;
+
+    let mut flush_backoff = FLUSH_INTERVAL_BASE;
+    loop {
+        for ch in CHANNELS {
+            match scrape_channel(ch) {
+                Ok(events) if !events.is_empty() => {
+                    log::info!(
+                        "perf-events: {} new from {}",
+                        events.len(),
+                        ch.provider,
+                    );
+                    QUEUE.lock().unwrap().extend(events);
+                }
+                Ok(_) => {}
+                Err(e) => {
+                    log::warn!(
+                        "perf-events: scrape {} failed: {e:#}",
+                        ch.provider,
+                    );
+                }
+            }
+        }
+
+        match flush_once().await {
+            FlushOutcome::Idle | FlushOutcome::AllSent => {
+                flush_backoff = FLUSH_INTERVAL_BASE;
+            }
+            FlushOutcome::Failed => {
+                flush_backoff = (flush_backoff * 2).min(FLUSH_INTERVAL_MAX);
+            }
+        }
+
+        tokio::time::sleep(SCRAPE_INTERVAL.min(flush_backoff)).await;
+    }
+}
+
+// ─────────────────────────── per-channel scrape ───────────────────────────
+
+#[cfg(target_os = "windows")]
+fn scrape_channel(ch: &ChannelCfg) -> Result<Vec<PendingPerfEvent>> {
+    let cursor = read_cursor(ch.provider);
+    let script = build_script(ch, cursor);
+    let stdout = run_powershell(&script)?;
+    if stdout.is_empty() {
+        return Ok(Vec::new());
+    }
+    // ConvertTo-Json with a single-element array still emits a JSON
+    // object (PowerShell's "unrolling" quirk); coerce both shapes.
+    let trimmed = stdout.trim();
+    if trimmed == "[]" || trimmed.is_empty() {
+        return Ok(Vec::new());
+    }
+    let raw: Vec<RawEvent> = match hbb_common::serde_json::from_str(trimmed) {
+        Ok(v) => v,
+        Err(_) => {
+            // Single-object case
+            match hbb_common::serde_json::from_str::<RawEvent>(trimmed) {
+                Ok(single) => vec![single],
+                Err(e) => {
+                    return Err(anyhow!(
+                        "PowerShell output is not valid JSON: {e}; first 200 chars: {:.200}",
+                        trimmed
+                    ));
+                }
+            }
+        }
+    };
+
+    let mut max_record_id = cursor;
+    let mut out = Vec::with_capacity(raw.len());
+    for ev in raw {
+        if (ev.record_id as u64) > max_record_id {
+            max_record_id = ev.record_id as u64;
+        }
+        out.push(PendingPerfEvent {
+            provider: ch.provider,
+            at: ev.at,
+            event_id: ev.event_id,
+            level: ev.level,
+            record_id: ev.record_id,
+            summary: ev.summary,
+            detail: ev.detail,
+            attempts: 0,
+        });
+    }
+
+    // Advance the cursor on observation, not on successful POST. The
+    // tradeoff: an agent that crashes between observing and POSTing
+    // loses those rows from the UI. Windows still has them in the
+    // event log, so the operator can fall back to Event Viewer if
+    // they really need them; we prefer that to repeatedly re-observing
+    // a backlog the server has already taken (unique-index dedup
+    // would absorb it, but the agent's queue would grow unbounded
+    // every scrape).
+    if max_record_id > cursor {
+        write_cursor(ch.provider, max_record_id);
+    }
+
+    Ok(out)
+}
+
+/// Build the PowerShell script for one channel. Two shapes:
+///
+///   * `cursor == 0` (first run): `FilterHashtable` with a 7-day
+///     `StartTime` and the optional ID allow-list. The hashtable form
+///     is the only one that accepts both a time bound and an `Id`
+///     array in one call.
+///   * `cursor > 0`: `FilterXPath` with `EventRecordID > $cursor` and
+///     the optional ID-list expanded to `(EventID=A or EventID=B …)`.
+#[cfg(target_os = "windows")]
+fn build_script(ch: &ChannelCfg, cursor: u64) -> String {
+    // PowerShell-quote the log name (single-quote escape = doubled
+    // single quote).
+    let q = |s: &str| s.replace('\'', "''");
+    let log_name_quoted = q(ch.log_name);
+
+    let filter_clause = if cursor == 0 {
+        // FilterHashtable + StartTime + optional Id array.
+        match ch.event_ids {
+            Some(ids) => {
+                let id_list = ids
+                    .iter()
+                    .map(|i| i.to_string())
+                    .collect::<Vec<_>>()
+                    .join(",");
+                format!(
+                    "-FilterHashtable @{{LogName='{ln}'; StartTime=(Get-Date).AddDays(-7); Id=@({ids})}} -MaxEvents {max}",
+                    ln = log_name_quoted,
+                    ids = id_list,
+                    max = MAX_FIRST_RUN_EVENTS,
+                )
+            }
+            None => format!(
+                "-FilterHashtable @{{LogName='{ln}'; StartTime=(Get-Date).AddDays(-7)}} -MaxEvents {max}",
+                ln = log_name_quoted,
+                max = MAX_FIRST_RUN_EVENTS,
+            ),
+        }
+    } else {
+        // FilterXPath. Note PowerShell expands `$cursor` from the
+        // outer script, so we splat the literal value into the XPath.
+        let id_clause = match ch.event_ids {
+            Some(ids) => {
+                let or = ids
+                    .iter()
+                    .map(|i| format!("System/EventID={i}"))
+                    .collect::<Vec<_>>()
+                    .join(" or ");
+                format!(" and ({or})")
+            }
+            None => String::new(),
+        };
+        // Double-curly to escape format!'s own `{}` interpolation.
+        format!(
+            "-LogName '{ln}' -FilterXPath \"*[System/EventRecordID>{cur}{id}]\" -MaxEvents {max}",
+            ln = log_name_quoted,
+            cur = cursor,
+            id = id_clause,
+            max = MAX_PER_SCRAPE,
+        )
+    };
+
+    // Single-quoted PowerShell here-doc would escape too aggressively;
+    // we stick to plain string concatenation. The `try / catch` block
+    // returns '[]' on any failure so the Rust side gets a parseable
+    // empty array rather than a stderr blob.
+    format!(
+        r#"$ErrorActionPreference = 'SilentlyContinue'
+try {{
+  $events = Get-WinEvent {filter} 2>$null
+  if ($null -eq $events) {{ '[]' ; exit 0 }}
+  $arr = @($events | Sort-Object RecordId | ForEach-Object {{
+    $msg = $_.Message
+    if ($null -eq $msg) {{ $msg = '' }}
+    $oneline = ($msg -replace "(`r?`n)+", " ").Trim()
+    if ($oneline.Length -gt 300) {{ $oneline = $oneline.Substring(0,300) }}
+    $detail = $msg
+    if ($detail.Length -gt 4000) {{ $detail = $detail.Substring(0,4000) }}
+    [PSCustomObject]@{{
+      at        = [int64](([System.DateTimeOffset]$_.TimeCreated.ToUniversalTime()).ToUnixTimeSeconds())
+      event_id  = $_.Id
+      level     = $_.Level
+      record_id = $_.RecordId
+      summary   = $oneline
+      detail    = $detail
+    }}
+  }})
+  $arr | ConvertTo-Json -Depth 3 -Compress
+}} catch {{
+  '[]'
+}}
+"#,
+        filter = filter_clause,
+    )
+}
+
+#[cfg(target_os = "windows")]
+fn run_powershell(script: &str) -> Result<String> {
+    use std::os::windows::process::CommandExt;
+    use std::process::Command;
+    // Matches inventory.rs — hides the brief console flash when the
+    // agent runs interactively (dev mode); no effect in service mode.
+    const CREATE_NO_WINDOW: u32 = 0x08000000;
+
+    let output = Command::new("powershell.exe")
+        .args([
+            "-NoProfile",
+            "-NonInteractive",
+            "-ExecutionPolicy",
+            "Bypass",
+            "-Command",
+            script,
+        ])
+        .creation_flags(CREATE_NO_WINDOW)
+        .output()
+        .map_err(|e| anyhow!("spawn powershell: {e}"))?;
+
+    if !output.status.success() {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        return Err(anyhow!(
+            "powershell exited {:?}: {}",
+            output.status.code(),
+            stderr.trim()
+        ));
+    }
+    Ok(String::from_utf8_lossy(&output.stdout).trim().to_string())
+}
+
+// ─────────────────────────────── cursor I/O ───────────────────────────────
+
+#[cfg(target_os = "windows")]
+fn cursor_key(provider: &str) -> String {
+    format!("perf-event-cursor-{provider}")
+}
+
+#[cfg(target_os = "windows")]
+fn read_cursor(provider: &str) -> u64 {
+    let raw = hbb_common::config::Config::get_option(&cursor_key(provider));
+    raw.parse::<u64>().unwrap_or(0)
+}
+
+#[cfg(target_os = "windows")]
+fn write_cursor(provider: &str, value: u64) {
+    hbb_common::config::Config::set_option(
+        cursor_key(provider),
+        value.to_string(),
+    );
+}
+
+// ───────────────────────────────── flush ──────────────────────────────────
+
+#[cfg(target_os = "windows")]
+enum FlushOutcome {
+    Idle,
+    AllSent,
+    Failed,
+}
+
+#[cfg(target_os = "windows")]
+async fn flush_once() -> FlushOutcome {
+    let batch: Vec<PendingPerfEvent> = {
+        let mut q = QUEUE.lock().unwrap();
+        if q.is_empty() {
+            return FlushOutcome::Idle;
+        }
+        let take = q.len().min(MAX_EVENTS_PER_POST);
+        q.drain(..take).collect()
+    };
+
+    match post_batch(&batch).await {
+        Ok(()) => FlushOutcome::AllSent,
+        Err(e) => {
+            log::warn!(
+                "perf-events: flush of {} event(s) failed: {e:#}",
+                batch.len(),
+            );
+            let mut requeued: Vec<PendingPerfEvent> = batch
+                .into_iter()
+                .filter_map(|mut ev| {
+                    ev.attempts = ev.attempts.saturating_add(1);
+                    if ev.attempts >= DROP_AFTER {
+                        log::warn!(
+                            "perf-events: dropping event after {} attempts: \
+                             provider={} record_id={}",
+                            ev.attempts, ev.provider, ev.record_id,
+                        );
+                        None
+                    } else {
+                        Some(ev)
+                    }
+                })
+                .collect();
+            let mut q = QUEUE.lock().unwrap();
+            let tail: Vec<PendingPerfEvent> = q.drain(..).collect();
+            requeued.extend(tail);
+            *q = requeued;
+            FlushOutcome::Failed
+        }
+    }
+}
+
+#[cfg(target_os = "windows")]
+async fn post_batch(batch: &[PendingPerfEvent]) -> Result<()> {
+    let api = librustdesk::common::get_api_server(
+        hbb_common::config::Config::get_option("api-server"),
+        hbb_common::config::Config::get_option("custom-rendezvous-server"),
+    );
+    if api.is_empty() {
+        return Err(anyhow!("no api-server configured yet"));
+    }
+
+    let url = format!("{api}/api/agent/perf-events");
+    let id = hbb_common::config::Config::get_id();
+    let uuid = librustdesk::common::encode64(hbb_common::get_uuid());
+
+    let events: Vec<hbb_common::serde_json::Value> = batch
+        .iter()
+        .map(|ev| {
+            hbb_common::serde_json::json!({
+                "at":          ev.at,
+                "provider":    ev.provider,
+                "event_id":    ev.event_id,
+                "level":       ev.level,
+                "record_id":   ev.record_id,
+                "summary":     ev.summary,
+                "detail_json": ev.detail,
+            })
+        })
+        .collect();
+    let body = hbb_common::serde_json::json!({
+        "id":     id,
+        "uuid":   uuid,
+        "events": events,
+    })
+    .to_string();
+
+    let headers = librustdesk::hbbs_http::sign::build_signed_headers(
+        "POST",
+        "/api/agent/perf-events",
+        body.as_bytes(),
+    )
+    .unwrap_or_default();
+
+    let resp = librustdesk::common::post_request(url, body, &headers)
+        .await
+        .map_err(|e| anyhow!("post: {e}"))?;
+    let trimmed = resp.trim();
+    if trimmed == "OK" || trimmed == "ID_NOT_FOUND" {
+        Ok(())
+    } else {
+        Err(anyhow!("unexpected response: {trimmed}"))
+    }
+}

src/service.rs

@@ -711,6 +711,20 @@ fn service_main_inner() -> Result<()> {
     // shutdown hook (the SCM termination is enough).
     crate::login_events::start();
 
+    // Start the continuous performance sampler: one CPU / memory /
+    // top-process sample per minute, posted in batches to
+    // /api/agent/metrics. Powers the device-detail Performance card
+    // and 24 h sparkline.
+    crate::perf::start();
+
+    // Start the Windows-event-log perf-event scraper: pulls boot /
+    // shutdown / sleep degradation, memory exhaustion, BSOD and
+    // unexpected-reboot events from the OS-managed channels and
+    // POSTs them to /api/agent/perf-events. Persists a per-channel
+    // RecordId cursor in the agent config so a restart doesn't
+    // re-emit the whole history.
+    crate::perf_events::start();
+
     // Worker process handle. Killed on Stop, replaced on session change.
     // `last_state` carries (session_id, had_user). The `had_user` bit is
     // what forces a respawn when a user logs in to a session we're