Merge pull request #986 from LibreQoE/ecore2

Ecore2

src/rust/Cargo.lock

@@ -2021,6 +2021,7 @@ dependencies = [
  "ip_network",
  "ip_network_table",
  "lqos_utils",
+ "nix",
  "once_cell",
  "serde",
  "serde_json",

src/rust/lqos_config/Cargo.toml

@@ -19,6 +19,7 @@ toml = {  workspace = true }
 lqos_utils = { path = "../lqos_utils" }
 arc-swap = { workspace = true }
 once_cell = { workspace = true }
+nix = { workspace = true, features = ["sched"] }
 
 # For memory debugging
 allocative.workspace = true

src/rust/lqos_config/src/cpu_topology.rs

@@ -2,15 +2,20 @@
 //!
 //! LibreQoS primarily runs on Linux and may be deployed on hybrid CPUs that
 //! include both performance cores ("P-cores") and efficiency cores ("E-cores").
-//! When possible, we detect E-cores via sysfs and provide a list of CPUs that
-//! should be used for shaping / CPU binning.
+//! When enabled, detection probes several Linux and CPU-specific sources in
+//! descending order of confidence, persists the resolved split to the LibreQoS
+//! runtime directory, and reuses that cached result across processes.
 
 use crate::Config;
-use std::collections::HashSet;
-use std::path::Path;
+use serde::{Deserialize, Serialize};
+use std::collections::{BTreeMap, HashSet};
+use std::path::{Path, PathBuf};
 use thiserror::Error;
+use tracing::debug;
 
 const POSSIBLE_CPUS_PATH: &str = "/sys/devices/system/cpu/possible";
+const CPU_TOPOLOGY_CACHE_FILE: &str = "cpu_topology_cache.json";
+const CPU_TOPOLOGY_CACHE_VERSION: u32 = 1;
 
 const CPU_CORE_SYSFS_PATHS: [&str; 2] = [
     "/sys/bus/event_source/devices/cpu_core/cpus",
@@ -22,16 +27,31 @@ const CPU_ATOM_SYSFS_PATHS: [&str; 2] = [
     "/sys/devices/cpu_atom/cpus",
 ];
 
+const CORE_TYPE_PERFORMANCE_VALUES: [u32; 2] = [1, 2];
+const CORE_TYPE_EFFICIENCY_VALUE: u32 = 3;
+const CPU_CAPACITY_EFFICIENCY_PERCENT: u32 = 80;
+const CPU_FREQ_EFFICIENCY_PERCENT: u32 = 90;
+const CPUID_CORE_TYPE_ATOM: u32 = 0x20;
+const CPUID_CORE_TYPE_CORE: u32 = 0x40;
+
 /// Indicates which mechanism was used to choose shaping CPUs.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 pub enum ShapingCpuSource {
     /// Exclusion is disabled; use all possible CPUs.
     ExclusionDisabled,
+    /// Used `/sys/devices/system/cpu/cpu*/topology/core_type`.
+    CoreTypeSysfs,
+    /// Used `/sys/devices/system/cpu/cpu*/cpu_capacity`.
+    CpuCapacitySysfs,
     /// Used `/sys/.../cpu_core/cpus` to identify P-cores.
     CpuCoreSysfs,
     /// Used `/sys/.../cpu_atom/cpus` to identify E-cores, then computed `possible - atom`.
     PossibleMinusAtomSysfs,
-    /// Hybrid sysfs paths were unavailable; used `/sys/devices/system/cpu/possible`.
+    /// Used Intel hybrid CPUID leaf `0x1A`.
+    CpuidHybrid,
+    /// Used `cpuinfo_max_freq` as a last-resort heuristic.
+    MaxFrequencyHeuristic,
+    /// Hybrid detection failed or was not trustworthy; used `/sys/devices/system/cpu/possible`.
     FallbackAllPossible,
     /// Could not read `/sys/devices/system/cpu/possible`; used Rust's available parallelism.
     FallbackAvailableParallelism,
@@ -44,16 +64,82 @@ pub struct ShapingCpuDetection {
     pub exclude_efficiency_cores: bool,
     /// Source used to determine the shaping CPU list.
     pub source: ShapingCpuSource,
+    /// Whether the resolved topology was loaded from the persisted cache file.
+    pub from_cache: bool,
+    /// Whether a trustworthy hybrid split was detected.
+    pub has_hybrid_split: bool,
+    /// Human-readable detail used for logs and tests.
+    pub detail: String,
     /// All possible CPU IDs (best effort).
     pub possible: Vec<u32>,
-    /// Detected performance-core CPU IDs (may be empty if unknown).
+    /// Detected performance-core CPU IDs.
     pub performance: Vec<u32>,
-    /// Detected efficiency-core CPU IDs (may be empty if unknown).
+    /// Detected efficiency-core CPU IDs.
     pub efficiency: Vec<u32>,
     /// Final CPU list that should be used for shaping/bins.
     pub shaping: Vec<u32>,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+struct ResolvedHybridCpuTopology {
+    source: ShapingCpuSource,
+    from_cache: bool,
+    has_hybrid_split: bool,
+    detail: String,
+    possible: Vec<u32>,
+    performance: Vec<u32>,
+    efficiency: Vec<u32>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+struct CpuTopologyCacheFile {
+    version: u32,
+    fingerprint: CpuTopologyFingerprint,
+    topology: ResolvedHybridCpuTopology,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+struct CpuTopologyFingerprint {
+    kernel_release: Option<String>,
+    possible: Vec<u32>,
+    vendor_id: Option<String>,
+    model_name: Option<String>,
+    cpu_family: Option<String>,
+    model: Option<String>,
+}
+
+impl ResolvedHybridCpuTopology {
+    fn hybrid(
+        source: ShapingCpuSource,
+        possible: &[u32],
+        performance: Vec<u32>,
+        efficiency: Vec<u32>,
+        detail: impl Into<String>,
+    ) -> Self {
+        Self {
+            source,
+            from_cache: false,
+            has_hybrid_split: true,
+            detail: detail.into(),
+            possible: possible.to_vec(),
+            performance,
+            efficiency,
+        }
+    }
+
+    fn fallback(source: ShapingCpuSource, possible: &[u32], detail: impl Into<String>) -> Self {
+        Self {
+            source,
+            from_cache: false,
+            has_hybrid_split: false,
+            detail: detail.into(),
+            possible: possible.to_vec(),
+            performance: possible.to_vec(),
+            efficiency: Vec::new(),
+        }
+    }
+}
+
 /// Errors that can occur when parsing a Linux cpulist string.
 #[derive(Error, Debug, Clone, PartialEq, Eq)]
 pub enum CpuListParseError {
@@ -138,20 +224,436 @@ fn possible_cpu_list() -> (Vec<u32>, ShapingCpuSource) {
     (fallback, ShapingCpuSource::FallbackAvailableParallelism)
 }
 
+fn try_read_u32(path: impl AsRef<Path>) -> Option<u32> {
+    std::fs::read_to_string(path)
+        .ok()?
+        .trim()
+        .parse::<u32>()
+        .ok()
+}
+
+fn read_per_cpu_values(possible: &[u32], suffix: &str) -> BTreeMap<u32, u32> {
+    possible
+        .iter()
+        .filter_map(|cpu| {
+            let path = format!("/sys/devices/system/cpu/cpu{cpu}/{suffix}");
+            try_read_u32(&path).map(|value| (*cpu, value))
+        })
+        .collect()
+}
+
+fn filter_subset(possible: &[u32], cpus: Vec<u32>) -> Vec<u32> {
+    let possible_set: HashSet<u32> = possible.iter().copied().collect();
+    let mut filtered: Vec<u32> = cpus
+        .into_iter()
+        .filter(|cpu| possible_set.contains(cpu))
+        .collect();
+    filtered.sort_unstable();
+    filtered.dedup();
+    filtered
+}
+
+fn validate_split(
+    possible: &[u32],
+    performance: Vec<u32>,
+    efficiency: Vec<u32>,
+) -> Option<(Vec<u32>, Vec<u32>)> {
+    let possible_set: HashSet<u32> = possible.iter().copied().collect();
+    let mut perf = filter_subset(possible, performance);
+    let eff = filter_subset(possible, efficiency);
+    if perf.is_empty() || eff.is_empty() {
+        return None;
+    }
+
+    let eff_set: HashSet<u32> = eff.iter().copied().collect();
+    perf.retain(|cpu| !eff_set.contains(cpu));
+    if perf.is_empty() {
+        return None;
+    }
+
+    let union: HashSet<u32> = perf.iter().chain(eff.iter()).copied().collect();
+    if union != possible_set {
+        return None;
+    }
+
+    Some((perf, eff))
+}
+
+fn detect_from_core_types(
+    possible: &[u32],
+    values: &BTreeMap<u32, u32>,
+) -> Option<ResolvedHybridCpuTopology> {
+    if values.is_empty() {
+        return None;
+    }
+
+    let performance: Vec<u32> = values
+        .iter()
+        .filter_map(|(cpu, value)| CORE_TYPE_PERFORMANCE_VALUES.contains(value).then_some(*cpu))
+        .collect();
+    let efficiency: Vec<u32> = values
+        .iter()
+        .filter_map(|(cpu, value)| (*value == CORE_TYPE_EFFICIENCY_VALUE).then_some(*cpu))
+        .collect();
+
+    let (performance, efficiency) = validate_split(possible, performance, efficiency)?;
+    Some(ResolvedHybridCpuTopology::hybrid(
+        ShapingCpuSource::CoreTypeSysfs,
+        possible,
+        performance,
+        efficiency,
+        "Detected hybrid CPU split via topology/core_type",
+    ))
+}
+
+fn detect_from_capacity(
+    possible: &[u32],
+    values: &BTreeMap<u32, u32>,
+) -> Option<ResolvedHybridCpuTopology> {
+    if values.len() != possible.len() || values.is_empty() {
+        return None;
+    }
+
+    let max_capacity = values.values().copied().max()?;
+    let min_capacity = values.values().copied().min()?;
+    if max_capacity == 0 || max_capacity == min_capacity {
+        return None;
+    }
+
+    let threshold = max_capacity.saturating_mul(CPU_CAPACITY_EFFICIENCY_PERCENT) / 100;
+    let mut performance = Vec::new();
+    let mut efficiency = Vec::new();
+    for (cpu, value) in values {
+        if *value < threshold {
+            efficiency.push(*cpu);
+        } else {
+            performance.push(*cpu);
+        }
+    }
+
+    let (performance, efficiency) = validate_split(possible, performance, efficiency)?;
+    Some(ResolvedHybridCpuTopology::hybrid(
+        ShapingCpuSource::CpuCapacitySysfs,
+        possible,
+        performance,
+        efficiency,
+        format!(
+            "Detected hybrid CPU split via cpu_capacity (max={max_capacity}, min={min_capacity})"
+        ),
+    ))
+}
+
+fn detect_from_cpu_core_atom_lists(
+    possible: &[u32],
+    perf_opt: Option<Vec<u32>>,
+    eff_opt: Option<Vec<u32>>,
+) -> Option<ResolvedHybridCpuTopology> {
+    let filtered_efficiency = eff_opt
+        .clone()
+        .map(|efficiency| filter_subset(possible, efficiency));
+
+    if let Some(mut performance) = perf_opt {
+        performance = filter_subset(possible, performance);
+        if let Some(efficiency) = filtered_efficiency.as_ref() {
+            let eff_set: HashSet<u32> = efficiency.iter().copied().collect();
+            performance.retain(|cpu| !eff_set.contains(cpu));
+        }
+
+        if !performance.is_empty() {
+            return Some(ResolvedHybridCpuTopology::hybrid(
+                ShapingCpuSource::CpuCoreSysfs,
+                possible,
+                performance,
+                filtered_efficiency.unwrap_or_default(),
+                "Detected performance CPUs via cpu_core cpulist",
+            ));
+        }
+    }
+
+    if let Some(efficiency) = filtered_efficiency {
+        let possible_set: HashSet<u32> = possible.iter().copied().collect();
+        let eff_set: HashSet<u32> = efficiency.iter().copied().collect();
+        let performance: Vec<u32> = possible_set.difference(&eff_set).copied().collect();
+        let (performance, efficiency) = validate_split(possible, performance, efficiency)?;
+        return Some(ResolvedHybridCpuTopology::hybrid(
+            ShapingCpuSource::PossibleMinusAtomSysfs,
+            possible,
+            performance,
+            efficiency,
+            "Detected hybrid CPU split via cpu_atom cpulist",
+        ));
+    }
+
+    None
+}
+
+fn detect_from_frequency(
+    possible: &[u32],
+    values: &BTreeMap<u32, u32>,
+) -> Option<ResolvedHybridCpuTopology> {
+    if values.len() != possible.len() || values.is_empty() {
+        return None;
+    }
+
+    let max_freq = values.values().copied().max()?;
+    let min_freq = values.values().copied().min()?;
+    if max_freq == 0 || max_freq == min_freq {
+        return None;
+    }
+
+    let threshold = max_freq.saturating_mul(CPU_FREQ_EFFICIENCY_PERCENT) / 100;
+    let mut performance = Vec::new();
+    let mut efficiency = Vec::new();
+    for (cpu, value) in values {
+        if *value < threshold {
+            efficiency.push(*cpu);
+        } else {
+            performance.push(*cpu);
+        }
+    }
+
+    let (performance, efficiency) = validate_split(possible, performance, efficiency)?;
+    Some(ResolvedHybridCpuTopology::hybrid(
+        ShapingCpuSource::MaxFrequencyHeuristic,
+        possible,
+        performance,
+        efficiency,
+        format!("Detected hybrid CPU split via cpuinfo_max_freq (max={max_freq}, min={min_freq})"),
+    ))
+}
+
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn detect_from_cpuid(possible: &[u32]) -> Option<ResolvedHybridCpuTopology> {
+    use nix::sched::{CpuSet, sched_getaffinity, sched_setaffinity};
+    use nix::unistd::Pid;
+
+    #[cfg(target_arch = "x86")]
+    use std::arch::x86::__cpuid_count;
+    #[cfg(target_arch = "x86_64")]
+    use std::arch::x86_64::__cpuid_count;
+
+    let original_affinity = sched_getaffinity(Pid::from_raw(0)).ok()?;
+    let restore_affinity = original_affinity;
+    let mut performance = Vec::new();
+    let mut efficiency = Vec::new();
+
+    for cpu in possible {
+        let mut affinity = CpuSet::new();
+        if affinity.set(*cpu as usize).is_err() {
+            let _ = sched_setaffinity(Pid::from_raw(0), &restore_affinity);
+            return None;
+        }
+        if sched_setaffinity(Pid::from_raw(0), &affinity).is_err() {
+            let _ = sched_setaffinity(Pid::from_raw(0), &restore_affinity);
+            return None;
+        }
+
+        let leaf = unsafe { __cpuid_count(0x1a, 0) };
+        let core_type = (leaf.eax >> 24) & 0xff;
+        match core_type {
+            CPUID_CORE_TYPE_CORE => performance.push(*cpu),
+            CPUID_CORE_TYPE_ATOM => efficiency.push(*cpu),
+            _ => {}
+        }
+    }
+
+    let _ = sched_setaffinity(Pid::from_raw(0), &restore_affinity);
+
+    let (performance, efficiency) = validate_split(possible, performance, efficiency)?;
+    Some(ResolvedHybridCpuTopology::hybrid(
+        ShapingCpuSource::CpuidHybrid,
+        possible,
+        performance,
+        efficiency,
+        "Detected hybrid CPU split via Intel hybrid CPUID leaf 0x1A",
+    ))
+}
+
+#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
+fn detect_from_cpuid(_possible: &[u32]) -> Option<ResolvedHybridCpuTopology> {
+    None
+}
+
+fn detect_hybrid_topology(
+    possible: &[u32],
+    possible_source: ShapingCpuSource,
+) -> ResolvedHybridCpuTopology {
+    let core_types = read_per_cpu_values(possible, "topology/core_type");
+    if let Some(resolved) = detect_from_core_types(possible, &core_types) {
+        return resolved;
+    }
+
+    let capacities = read_per_cpu_values(possible, "cpu_capacity");
+    if let Some(resolved) = detect_from_capacity(possible, &capacities) {
+        return resolved;
+    }
+
+    let perf_opt = first_cpu_list(&CPU_CORE_SYSFS_PATHS);
+    let eff_opt = first_cpu_list(&CPU_ATOM_SYSFS_PATHS);
+    if let Some(resolved) = detect_from_cpu_core_atom_lists(possible, perf_opt, eff_opt) {
+        return resolved;
+    }
+
+    if let Some(resolved) = detect_from_cpuid(possible) {
+        return resolved;
+    }
+
+    let frequencies = read_per_cpu_values(possible, "cpufreq/cpuinfo_max_freq");
+    if let Some(resolved) = detect_from_frequency(possible, &frequencies) {
+        return resolved;
+    }
+
+    ResolvedHybridCpuTopology::fallback(
+        possible_source,
+        possible,
+        "No trustworthy hybrid CPU split detected; using all possible CPUs",
+    )
+}
+
+fn cache_file_path(cfg: &Config) -> PathBuf {
+    Path::new(&cfg.lqos_directory).join(CPU_TOPOLOGY_CACHE_FILE)
+}
+
+fn read_cpuinfo_identity() -> (
+    Option<String>,
+    Option<String>,
+    Option<String>,
+    Option<String>,
+) {
+    let cpuinfo = match std::fs::read_to_string("/proc/cpuinfo") {
+        Ok(cpuinfo) => cpuinfo,
+        Err(_) => return (None, None, None, None),
+    };
+
+    let mut vendor_id = None;
+    let mut model_name = None;
+    let mut cpu_family = None;
+    let mut model = None;
+    for line in cpuinfo.lines() {
+        let Some((key, value)) = line.split_once(':') else {
+            continue;
+        };
+        let key = key.trim();
+        let value = value.trim().to_string();
+        match key {
+            "vendor_id" if vendor_id.is_none() => vendor_id = Some(value),
+            "model name" | "Processor" | "Hardware" if model_name.is_none() => {
+                model_name = Some(value);
+            }
+            "cpu family" if cpu_family.is_none() => cpu_family = Some(value),
+            "model" if model.is_none() => model = Some(value),
+            _ => {}
+        }
+    }
+
+    (vendor_id, model_name, cpu_family, model)
+}
+
+impl CpuTopologyFingerprint {
+    fn gather(possible: &[u32]) -> Self {
+        let kernel_release = std::fs::read_to_string("/proc/sys/kernel/osrelease")
+            .ok()
+            .map(|s| s.trim().to_string())
+            .filter(|s| !s.is_empty());
+        let (vendor_id, model_name, cpu_family, model) = read_cpuinfo_identity();
+        Self {
+            kernel_release,
+            possible: possible.to_vec(),
+            vendor_id,
+            model_name,
+            cpu_family,
+            model,
+        }
+    }
+}
+
+fn load_topology_cache(
+    cfg: &Config,
+    fingerprint: &CpuTopologyFingerprint,
+) -> Option<ResolvedHybridCpuTopology> {
+    let path = cache_file_path(cfg);
+    let raw = std::fs::read_to_string(&path).ok()?;
+    let cache: CpuTopologyCacheFile = serde_json::from_str(&raw).ok()?;
+    if cache.version != CPU_TOPOLOGY_CACHE_VERSION || cache.fingerprint != *fingerprint {
+        return None;
+    }
+
+    let mut topology = cache.topology;
+    topology.from_cache = true;
+    Some(topology)
+}
+
+fn store_topology_cache(
+    cfg: &Config,
+    fingerprint: &CpuTopologyFingerprint,
+    topology: &ResolvedHybridCpuTopology,
+) {
+    let path = cache_file_path(cfg);
+    if let Some(parent) = path.parent()
+        && let Err(e) = std::fs::create_dir_all(parent)
+    {
+        debug!("Unable to create CPU topology cache directory {parent:?}: {e}");
+        return;
+    }
+
+    let temp_path = path.with_extension("json.tmp");
+    let cache = CpuTopologyCacheFile {
+        version: CPU_TOPOLOGY_CACHE_VERSION,
+        fingerprint: fingerprint.clone(),
+        topology: ResolvedHybridCpuTopology {
+            from_cache: false,
+            ..topology.clone()
+        },
+    };
+
+    let serialized = match serde_json::to_string_pretty(&cache) {
+        Ok(serialized) => serialized,
+        Err(e) => {
+            debug!("Unable to serialize CPU topology cache: {e}");
+            return;
+        }
+    };
+
+    if let Err(e) = std::fs::write(&temp_path, serialized.as_bytes()) {
+        debug!("Unable to write temporary CPU topology cache {temp_path:?}: {e}");
+        return;
+    }
+
+    if let Err(e) = std::fs::rename(&temp_path, &path) {
+        debug!("Unable to atomically replace CPU topology cache {path:?}: {e}");
+        let _ = std::fs::remove_file(&temp_path);
+    }
+}
+
+fn cached_or_detected_topology(cfg: &Config) -> ResolvedHybridCpuTopology {
+    let (possible, possible_source) = possible_cpu_list();
+    let fingerprint = CpuTopologyFingerprint::gather(&possible);
+
+    load_topology_cache(cfg, &fingerprint).unwrap_or_else(|| {
+        let detected = detect_hybrid_topology(&possible, possible_source);
+        store_topology_cache(cfg, &fingerprint, &detected);
+        detected
+    })
+}
+
 /// Detects the set of CPUs that should be used for shaping/CPU binning.
 ///
-/// When `cfg.exclude_efficiency_cores` is enabled, this attempts to identify
-/// performance cores via sysfs. If detection fails or yields an empty set,
-/// this falls back to "all possible CPUs".
+/// When `cfg.exclude_efficiency_cores` is enabled, this may read Linux sysfs,
+/// CPU information, and a cached topology file under the LibreQoS runtime
+/// directory. If hybrid detection fails or yields an untrustworthy split,
+/// LibreQoS falls back to all possible CPUs so non-hybrid systems are
+/// unaffected.
 pub fn detect_shaping_cpus(cfg: &Config) -> ShapingCpuDetection {
     let exclude = cfg.exclude_efficiency_cores;
 
-    let (possible, possible_source) = possible_cpu_list();
-
     if !exclude {
+        let (possible, _) = possible_cpu_list();
         return ShapingCpuDetection {
-            exclude_efficiency_cores: exclude,
+            exclude_efficiency_cores: false,
             source: ShapingCpuSource::ExclusionDisabled,
+            from_cache: false,
+            has_hybrid_split: false,
+            detail: "Efficiency-core exclusion disabled".to_string(),
             possible: possible.clone(),
             performance: possible.clone(),
             efficiency: Vec::new(),
@@ -159,75 +661,44 @@ pub fn detect_shaping_cpus(cfg: &Config) -> ShapingCpuDetection {
         };
     }
 
-    let perf_opt = first_cpu_list(&CPU_CORE_SYSFS_PATHS);
-    let eff_opt = first_cpu_list(&CPU_ATOM_SYSFS_PATHS);
+    let topology = cached_or_detected_topology(cfg);
+    let performance = if topology.has_hybrid_split {
+        topology.performance.clone()
+    } else {
+        topology.possible.clone()
+    };
+    let efficiency = if topology.has_hybrid_split {
+        topology.efficiency.clone()
+    } else {
+        Vec::new()
+    };
 
-    // Prefer cpu_core list if present.
-    if let Some(mut perf) = perf_opt.clone() {
-        // If both are present, remove overlaps just in case.
-        if let Some(eff) = eff_opt.clone()
-            && !eff.is_empty()
-            && !perf.is_empty()
-        {
-            let eff_set: HashSet<u32> = eff.iter().copied().collect();
-            perf.retain(|c| !eff_set.contains(c));
-        }
-
-        // Ensure it is a subset of possible CPUs, if that list exists.
-        if !possible.is_empty() {
-            let poss_set: HashSet<u32> = possible.iter().copied().collect();
-            perf.retain(|c| poss_set.contains(c));
-        }
-
-        if !perf.is_empty() {
-            return ShapingCpuDetection {
-                exclude_efficiency_cores: exclude,
-                source: ShapingCpuSource::CpuCoreSysfs,
-                possible,
-                performance: perf.clone(),
-                efficiency: eff_opt.unwrap_or_default(),
-                shaping: perf,
-            };
-        }
-    }
-
-    // Next-best: if we know efficiency CPUs, compute possible - efficiency.
-    if let Some(eff) = eff_opt.clone()
-        && !eff.is_empty()
-        && !possible.is_empty()
-    {
-        let eff_set: HashSet<u32> = eff.iter().copied().collect();
-        let perf: Vec<u32> = possible
-            .iter()
-            .copied()
-            .filter(|c| !eff_set.contains(c))
-            .collect();
-        if !perf.is_empty() {
-            return ShapingCpuDetection {
-                exclude_efficiency_cores: exclude,
-                source: ShapingCpuSource::PossibleMinusAtomSysfs,
-                possible,
-                performance: perf.clone(),
-                efficiency: eff,
-                shaping: perf,
-            };
-        }
-    }
-
-    // Final fallback: use all possible CPUs.
     ShapingCpuDetection {
-        exclude_efficiency_cores: exclude,
-        source: possible_source,
-        possible: possible.clone(),
-        performance: possible.clone(),
-        efficiency: eff_opt.unwrap_or_default(),
-        shaping: possible,
+        exclude_efficiency_cores: true,
+        source: topology.source,
+        from_cache: topology.from_cache,
+        has_hybrid_split: topology.has_hybrid_split,
+        detail: topology.detail,
+        possible: topology.possible.clone(),
+        performance: performance.clone(),
+        efficiency,
+        shaping: performance,
     }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
+    use std::sync::atomic::{AtomicU64, Ordering};
+
+    static TEST_COUNTER: AtomicU64 = AtomicU64::new(0);
+
+    fn temp_dir(name: &str) -> PathBuf {
+        let unique = TEST_COUNTER.fetch_add(1, Ordering::Relaxed);
+        let dir = std::env::temp_dir().join(format!("libreqos-{name}-{unique}"));
+        std::fs::create_dir_all(&dir).expect("test temp dir should be creatable");
+        dir
+    }
 
     #[test]
     fn parse_cpu_list_single() {
@@ -276,4 +747,196 @@ mod tests {
             CpuListParseError::Empty
         );
     }
+
+    #[test]
+    fn detects_core_type_split() {
+        let possible = vec![0, 1, 2, 3];
+        let values = BTreeMap::from([(0, 1), (1, 2), (2, 3), (3, 3)]);
+        let detection =
+            detect_from_core_types(&possible, &values).expect("core_type should detect hybrid");
+        assert_eq!(detection.source, ShapingCpuSource::CoreTypeSysfs);
+        assert_eq!(detection.performance, vec![0, 1]);
+        assert_eq!(detection.efficiency, vec![2, 3]);
+    }
+
+    #[test]
+    fn detects_cpu_capacity_split() {
+        let possible = vec![0, 1, 2, 3];
+        let values = BTreeMap::from([(0, 1024), (1, 1024), (2, 512), (3, 512)]);
+        let detection =
+            detect_from_capacity(&possible, &values).expect("cpu_capacity should detect hybrid");
+        assert_eq!(detection.source, ShapingCpuSource::CpuCapacitySysfs);
+        assert_eq!(detection.performance, vec![0, 1]);
+        assert_eq!(detection.efficiency, vec![2, 3]);
+    }
+
+    #[test]
+    fn capacity_rejects_non_partitioned_split() {
+        let possible = vec![0, 1, 2, 3];
+        let values = BTreeMap::from([(0, 1024), (1, 1024), (2, 512)]);
+        assert!(
+            detect_from_capacity(&possible, &values).is_none(),
+            "partial capacity data should not be trusted"
+        );
+    }
+
+    #[test]
+    fn detects_cpu_core_atom_lists() {
+        let possible = vec![0, 1, 2, 3];
+        let detection =
+            detect_from_cpu_core_atom_lists(&possible, Some(vec![0, 1]), Some(vec![2, 3]))
+                .expect("cpu_core/cpu_atom should detect hybrid");
+        assert_eq!(detection.source, ShapingCpuSource::CpuCoreSysfs);
+        assert_eq!(detection.performance, vec![0, 1]);
+        assert_eq!(detection.efficiency, vec![2, 3]);
+    }
+
+    #[test]
+    fn detects_cpu_core_list_without_atom_list() {
+        let possible = vec![0, 1, 2, 3];
+        let detection = detect_from_cpu_core_atom_lists(&possible, Some(vec![0, 1]), None)
+            .expect("cpu_core without cpu_atom should still detect performance CPUs");
+        assert_eq!(detection.source, ShapingCpuSource::CpuCoreSysfs);
+        assert!(detection.has_hybrid_split);
+        assert_eq!(detection.performance, vec![0, 1]);
+        assert!(detection.efficiency.is_empty());
+    }
+
+    #[test]
+    fn detects_frequency_split() {
+        let possible = vec![0, 1, 2, 3];
+        let values = BTreeMap::from([(0, 5200000), (1, 5200000), (2, 3900000), (3, 3900000)]);
+        let detection = detect_from_frequency(&possible, &values)
+            .expect("frequency split should detect hybrid");
+        assert_eq!(detection.source, ShapingCpuSource::MaxFrequencyHeuristic);
+        assert_eq!(detection.performance, vec![0, 1]);
+        assert_eq!(detection.efficiency, vec![2, 3]);
+    }
+
+    #[test]
+    fn fallback_uses_all_possible_cpus() {
+        let possible = vec![9_999];
+        let detection = detect_hybrid_topology(&possible, ShapingCpuSource::FallbackAllPossible);
+        assert_eq!(detection.source, ShapingCpuSource::FallbackAllPossible);
+        assert!(!detection.has_hybrid_split);
+        assert_eq!(detection.performance, possible);
+        assert!(detection.efficiency.is_empty());
+    }
+
+    #[test]
+    fn topology_cache_round_trip() {
+        let cfg = Config {
+            lqos_directory: temp_dir("cpu-topology-cache").display().to_string(),
+            ..Config::default()
+        };
+        let path = cache_file_path(&cfg);
+        let fingerprint = CpuTopologyFingerprint {
+            kernel_release: Some("6.9.0".to_string()),
+            possible: vec![0, 1, 2, 3],
+            vendor_id: Some("GenuineIntel".to_string()),
+            model_name: Some("Test CPU".to_string()),
+            cpu_family: Some("6".to_string()),
+            model: Some("151".to_string()),
+        };
+        let topology = ResolvedHybridCpuTopology::hybrid(
+            ShapingCpuSource::CoreTypeSysfs,
+            &[0, 1, 2, 3],
+            vec![0, 1],
+            vec![2, 3],
+            "test topology",
+        );
+
+        store_topology_cache(&cfg, &fingerprint, &topology);
+        let loaded = load_topology_cache(&cfg, &fingerprint).expect("cache should round-trip");
+        assert!(loaded.from_cache);
+        assert_eq!(loaded.source, topology.source);
+        assert_eq!(loaded.performance, topology.performance);
+        assert_eq!(loaded.efficiency, topology.efficiency);
+
+        let _ = std::fs::remove_file(path);
+    }
+
+    #[test]
+    fn topology_cache_invalidates_on_fingerprint_change() {
+        let cfg = Config {
+            lqos_directory: temp_dir("cpu-topology-cache-mismatch")
+                .display()
+                .to_string(),
+            ..Config::default()
+        };
+        let path = cache_file_path(&cfg);
+        let fingerprint = CpuTopologyFingerprint {
+            kernel_release: Some("6.9.0".to_string()),
+            possible: vec![0, 1, 2, 3],
+            vendor_id: Some("GenuineIntel".to_string()),
+            model_name: Some("Test CPU".to_string()),
+            cpu_family: Some("6".to_string()),
+            model: Some("151".to_string()),
+        };
+        let topology = ResolvedHybridCpuTopology::hybrid(
+            ShapingCpuSource::CoreTypeSysfs,
+            &[0, 1, 2, 3],
+            vec![0, 1],
+            vec![2, 3],
+            "test topology",
+        );
+        store_topology_cache(&cfg, &fingerprint, &topology);
+
+        let mut mismatched = fingerprint.clone();
+        mismatched.kernel_release = Some("6.10.0".to_string());
+        assert!(
+            load_topology_cache(&cfg, &mismatched).is_none(),
+            "fingerprint mismatch should invalidate the cache"
+        );
+
+        let _ = std::fs::remove_file(path);
+    }
+
+    #[test]
+    fn cached_detection_uses_each_config_directory_independently() {
+        let (possible, possible_source) = possible_cpu_list();
+        let fingerprint = CpuTopologyFingerprint::gather(&possible);
+
+        let cfg_a = Config {
+            lqos_directory: temp_dir("cpu-topology-cache-a").display().to_string(),
+            ..Config::default()
+        };
+        let cfg_b = Config {
+            lqos_directory: temp_dir("cpu-topology-cache-b").display().to_string(),
+            ..Config::default()
+        };
+
+        let path_a = cache_file_path(&cfg_a);
+        let path_b = cache_file_path(&cfg_b);
+
+        let topology_a = ResolvedHybridCpuTopology::fallback(
+            possible_source,
+            &possible,
+            "cache entry from directory a",
+        );
+        let topology_b = ResolvedHybridCpuTopology::fallback(
+            ShapingCpuSource::FallbackAvailableParallelism,
+            &possible,
+            "cache entry from directory b",
+        );
+
+        store_topology_cache(&cfg_a, &fingerprint, &topology_a);
+        store_topology_cache(&cfg_b, &fingerprint, &topology_b);
+
+        let loaded_a = cached_or_detected_topology(&cfg_a);
+        let loaded_b = cached_or_detected_topology(&cfg_b);
+
+        assert!(loaded_a.from_cache);
+        assert!(loaded_b.from_cache);
+        assert_eq!(loaded_a.detail, "cache entry from directory a");
+        assert_eq!(loaded_b.detail, "cache entry from directory b");
+        assert_eq!(loaded_a.source, possible_source);
+        assert_eq!(
+            loaded_b.source,
+            ShapingCpuSource::FallbackAvailableParallelism
+        );
+
+        let _ = std::fs::remove_file(path_a);
+        let _ = std::fs::remove_file(path_b);
+    }
 }

src/rust/lqos_config/src/etc/v15/top_config.rs

@@ -163,8 +163,9 @@ pub struct Config {
 
     /// Exclude efficiency cores (E-cores) from CPU assignment / shaping where possible.
     ///
-    /// On hybrid CPUs, this attempts to detect performance cores (P-cores) via sysfs and
-    /// restricts LibreQoS CPU binning and XDP redirection to those CPUs.
+    /// On hybrid CPUs, this tries several detection methods, caches the resolved
+    /// P-core/E-core split under the LibreQoS runtime directory, and restricts
+    /// shaping/XDP CPU assignment to performance cores when the split is trustworthy.
     #[serde(default = "default_true")]
     pub exclude_efficiency_cores: bool,
 

src/rust/lqos_python/src/lib.rs

@@ -1427,7 +1427,8 @@ fn monitor_mode_only() -> PyResult<bool> {
 
 /// Returns the number of CPUs that should be used for shaping / binning.
 ///
-/// On hybrid CPUs, this may exclude efficiency cores (E-cores) when configured.
+/// On hybrid CPUs, this may exclude efficiency cores (E-cores) when configured
+/// using the cached multi-method topology detector.
 #[pyfunction]
 fn shaping_cpu_count() -> PyResult<u32> {
     let config = lqos_config::load_config().unwrap();
@@ -1440,12 +1441,11 @@ fn shaping_cpu_count() -> PyResult<u32> {
 fn efficiency_core_ids() -> PyResult<Vec<u32>> {
     let config = lqos_config::load_config().unwrap();
     let det = lqos_config::detect_shaping_cpus(config.as_ref());
-    let efficiency = match det.source {
-        lqos_config::ShapingCpuSource::CpuCoreSysfs
-        | lqos_config::ShapingCpuSource::PossibleMinusAtomSysfs => det.efficiency,
-        _ => Vec::new(),
-    };
-    Ok(efficiency)
+    Ok(if det.has_hybrid_split {
+        det.efficiency
+    } else {
+        Vec::new()
+    })
 }
 
 #[pyfunction]

src/rust/lqos_sys/src/lqos_kernel.rs

@@ -278,13 +278,16 @@ pub fn attach_xdp_and_tc_to_interface(
         Ok(cfg) => {
             let det = lqos_config::detect_shaping_cpus(cfg.as_ref());
             info!(
-                "CPU topology: exclude_efficiency_cores={} source={:?} possible={} performance={} efficiency={} shaping={}",
+                "CPU topology: exclude_efficiency_cores={} source={:?} from_cache={} hybrid_split={} possible={} performance={} efficiency={} shaping={} detail={}",
                 det.exclude_efficiency_cores,
                 det.source,
+                det.from_cache,
+                det.has_hybrid_split,
                 det.possible.len(),
                 det.performance.len(),
                 det.efficiency.len(),
-                det.shaping.len()
+                det.shaping.len(),
+                det.detail
             );
             det.shaping
         }

src/rust/lqosd/src/node_manager/static2/config_general.html

@@ -92,7 +92,7 @@
             <div class="mb-3 form-check">
                 <input type="checkbox" class="form-check-input" id="excludeEfficiencyCores">
                 <label class="form-check-label" for="excludeEfficiencyCores">Exclude Efficiency Cores</label>
-                <div class="form-text">Avoid assigning shaping work to E-cores on hybrid CPUs (recommended)</div>
+                <div class="form-text">Use cached hybrid-core detection to keep shaping on P-cores and leave E-cores for background work when available</div>
             </div>
 
             <button type="button" id="saveButton" class="btn btn-outline-primary">Save Changes</button>