hermes_core/segment/
pin.rs

1//! Hot-metadata pinning: budgeted residency for per-query-mandatory
2//! structures (a meta/data residency split).
3//!
4//! Every query must touch certain small metadata sections — BMP block-offset
5//! tables, sparse skip sections, doc-id maps, superblock grids. Under memory
6//! pressure the kernel evicts them like bulk data, and queries then pay major
7//! faults on structures they cannot skip. This module pins them, in priority
8//! order (smallest/hottest first), until a per-segment budget is exhausted.
9//!
10//! Design: `docs/hot-metadata-pinning.md`. Bulk data (BMP 4-bit grid, block
11//! data, raw vectors) is never pinned — it is covered by the
12//! `MADV_RANDOM`/`MADV_WILLNEED` discipline instead.
13
14use std::sync::OnceLock;
15
16use crate::directories::OwnedBytes;
17
18/// How pinned bytes are kept resident.
19#[derive(Debug, Clone, Copy, PartialEq, Eq)]
20pub enum PinMode {
21    /// `mlock` the mmap pages in place — zero-copy, but requires
22    /// RLIMIT_MEMLOCK headroom (containers often need CAP_IPC_LOCK or an
23    /// explicit ulimit). Failures are logged and counted, never fatal.
24    Mlock,
25    /// Copy the section to the heap — no permissions needed, duplicates the
26    /// bytes. Immune to page-cache eviction (production runs swapless).
27    Copy,
28}
29
30/// Per-segment metadata pinning policy.
31#[derive(Debug, Clone, Copy)]
32pub struct PinPolicy {
33    /// Metadata bytes to pin per segment. 0 = pinning disabled (default).
34    pub budget_bytes: u64,
35    pub mode: PinMode,
36}
37
38impl PinPolicy {
39    pub const fn disabled() -> Self {
40        Self {
41            budget_bytes: 0,
42            mode: PinMode::Mlock,
43        }
44    }
45
46    pub fn is_enabled(&self) -> bool {
47        self.budget_bytes > 0
48    }
49
50    /// Read policy from environment:
51    /// `HERMES_PIN_METADATA_BUDGET_MB` (default 0 = off),
52    /// `HERMES_PIN_MODE` = `mlock` (default) | `copy`.
53    ///
54    /// Used as the default initializer for [`pin_policy`], and as the
55    /// backward-compatible fallback for consumers (e.g. hermes-server) that
56    /// surface these as CLI flags but still honor the env vars when the flags
57    /// are left unset.
58    pub fn from_env() -> Self {
59        let budget_mb: u64 = std::env::var("HERMES_PIN_METADATA_BUDGET_MB")
60            .ok()
61            .and_then(|v| v.parse().ok())
62            .unwrap_or(0);
63        let mode = match std::env::var("HERMES_PIN_MODE").as_deref() {
64            Ok("copy") => PinMode::Copy,
65            Ok("mlock") | Err(_) => PinMode::Mlock,
66            Ok(other) => {
67                log::warn!("HERMES_PIN_MODE '{}' unknown; using mlock", other);
68                PinMode::Mlock
69            }
70        };
71        Self {
72            budget_bytes: budget_mb * 1024 * 1024,
73            mode,
74        }
75    }
76}
77
78static PIN_POLICY: OnceLock<PinPolicy> = OnceLock::new();
79
80/// Override the process-wide pin policy. Must be called before the first
81/// segment is opened; returns false (and warns) if the policy was already
82/// initialized.
83pub fn set_pin_policy(policy: PinPolicy) -> bool {
84    let ok = PIN_POLICY.set(policy).is_ok();
85    if !ok {
86        log::warn!("pin policy already initialized; set_pin_policy ignored");
87    }
88    ok
89}
90
91/// The process-wide pin policy (env-initialized on first use).
92pub fn pin_policy() -> &'static PinPolicy {
93    PIN_POLICY.get_or_init(PinPolicy::from_env)
94}
95
96/// Accumulates pin accounting for one segment.
97#[derive(Debug, Default, Clone, Copy)]
98pub struct PinReport {
99    /// Bytes of pinnable metadata found (regardless of budget/failures)
100    pub intended_bytes: u64,
101    /// Bytes actually pinned
102    pub pinned_bytes: u64,
103    /// Bytes skipped because the budget was exhausted
104    pub skipped_budget_bytes: u64,
105    /// Bytes where mlock failed (RLIMIT_MEMLOCK etc.)
106    pub failed_bytes: u64,
107}
108
109/// Pin one metadata section, updating `remaining` budget and the report.
110///
111/// In `Copy` mode the section is replaced with a heap copy (heap memory is
112/// not page-cache-evictable). In `Mlock` mode the mmap pages are locked in
113/// place. Non-mmap-backed sections (RAM directories) are already resident
114/// and are skipped silently.
115pub(crate) fn pin_section(
116    bytes: &mut OwnedBytes,
117    label: &str,
118    mode: PinMode,
119    remaining: &mut u64,
120    report: &mut PinReport,
121) {
122    if !bytes.is_mmap() || bytes.is_empty() {
123        return;
124    }
125    let len = bytes.len() as u64;
126    report.intended_bytes += len;
127
128    if len > *remaining {
129        report.skipped_budget_bytes += len;
130        log::debug!(
131            "[pin] budget exhausted: skipping {} ({} bytes, {} remaining)",
132            label,
133            len,
134            *remaining
135        );
136        return;
137    }
138
139    match mode {
140        PinMode::Mlock => {
141            if bytes.mlock() {
142                *remaining -= len;
143                report.pinned_bytes += len;
144            } else {
145                report.failed_bytes += len;
146                log::warn!(
147                    "[pin] mlock failed for {} ({} bytes) — check RLIMIT_MEMLOCK; \
148                     continuing unpinned",
149                    label,
150                    len
151                );
152            }
153        }
154        PinMode::Copy => {
155            *bytes = OwnedBytes::new(bytes.to_vec());
156            *remaining -= len;
157            report.pinned_bytes += len;
158        }
159    }
160}
hermes_core/segment/pin.rs

hermes_core/segment/
pin.rs
