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::{Arc, 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/// RAII owner for heap pages locked on behalf of one immutable ANN artifact
110/// generation. The referenced allocations are owned by the same
111/// `TrainedVectorStructures`; its field order drops this set before the
112/// artifact `Arc`s, so every address remains valid through `munlock`.
113struct HeapPinGuard {
114    page_start: *mut libc::c_void,
115    page_len: usize,
116}
117
118// The guard never dereferences its pointer. The immutable artifact allocations
119// it describes are safe to share, and mlock/munlock operate on process mappings.
120unsafe impl Send for HeapPinGuard {}
121unsafe impl Sync for HeapPinGuard {}
122
123impl Drop for HeapPinGuard {
124    fn drop(&mut self) {
125        if unsafe { libc::munlock(self.page_start, self.page_len) } != 0 {
126            log::warn!(
127                "[pin] munlock failed for {} ANN heap bytes: {}",
128                self.page_len,
129                std::io::Error::last_os_error()
130            );
131        }
132    }
133}
134
135/// Locked heap allocations associated with one index-global ANN generation.
136/// Segment-local vector/code payloads are intentionally excluded.
137#[derive(Default)]
138pub(crate) struct HeapPinSet {
139    guards: Vec<HeapPinGuard>,
140    /// Keep every allocation owner alive until after its guards are dropped,
141    /// even if a cloned `TrainedVectorStructures` has its public maps mutated.
142    owners: Vec<Arc<dyn std::any::Any + Send + Sync>>,
143    report: PinReport,
144}
145
146impl HeapPinSet {
147    pub(crate) fn report(&self) -> PinReport {
148        self.report
149    }
150
151    pub(crate) fn retain_owner<T: std::any::Any + Send + Sync>(&mut self, owner: Arc<T>) {
152        self.owners.push(owner);
153    }
154
155    /// Keep one immutable heap slice resident, subject to the generation
156    /// budget. `Copy` mode needs no allocation: trained artifacts are already
157    /// heap-owned, which is exactly the residency guarantee that mode provides
158    /// on the supported swapless deployment.
159    pub(crate) fn pin_slice<T>(
160        &mut self,
161        slice: &[T],
162        label: &str,
163        mode: PinMode,
164        remaining: &mut u64,
165    ) {
166        let len = std::mem::size_of_val(slice);
167        if len == 0 {
168            return;
169        }
170        let Ok(len_u64) = u64::try_from(len) else {
171            self.report.failed_bytes = u64::MAX;
172            log::warn!("[pin] ANN region {label} is too large to account");
173            return;
174        };
175        self.report.intended_bytes = self.report.intended_bytes.saturating_add(len_u64);
176        if len_u64 > *remaining {
177            self.report.skipped_budget_bytes =
178                self.report.skipped_budget_bytes.saturating_add(len_u64);
179            log::debug!(
180                "[pin] ANN budget exhausted: skipping {} ({} bytes, {} remaining)",
181                label,
182                len_u64,
183                *remaining
184            );
185            return;
186        }
187
188        if mode == PinMode::Copy {
189            *remaining -= len_u64;
190            self.report.pinned_bytes = self.report.pinned_bytes.saturating_add(len_u64);
191            return;
192        }
193
194        let page_size = unsafe { libc::sysconf(libc::_SC_PAGESIZE) };
195        let page_size = usize::try_from(page_size).ok().filter(|&size| size > 0);
196        let Some(page_size) = page_size else {
197            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
198            log::warn!("[pin] cannot determine page size while locking {label}");
199            return;
200        };
201        let address = slice.as_ptr() as usize;
202        let page_start = address / page_size * page_size;
203        let Some(end) = address.checked_add(len) else {
204            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
205            log::warn!("[pin] ANN region address overflow while locking {label}");
206            return;
207        };
208        let Some(rounded_end) = end
209            .checked_add(page_size - 1)
210            .map(|value| value / page_size * page_size)
211        else {
212            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
213            log::warn!("[pin] ANN region page range overflow while locking {label}");
214            return;
215        };
216        let page_len = rounded_end - page_start;
217        let page_start = page_start as *mut libc::c_void;
218        if unsafe { libc::mlock(page_start.cast_const(), page_len) } == 0 {
219            self.guards.push(HeapPinGuard {
220                page_start,
221                page_len,
222            });
223            *remaining -= len_u64;
224            self.report.pinned_bytes = self.report.pinned_bytes.saturating_add(len_u64);
225        } else {
226            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
227            log::warn!(
228                "[pin] mlock failed for ANN {} ({} bytes): {} — check RLIMIT_MEMLOCK/CAP_IPC_LOCK; continuing unpinned",
229                label,
230                len_u64,
231                std::io::Error::last_os_error()
232            );
233        }
234    }
235}
236
237/// Pin one metadata section, updating `remaining` budget and the report.
238///
239/// In `Copy` mode the section is replaced with a heap copy (heap memory is
240/// not page-cache-evictable). In `Mlock` mode the mmap pages are locked in
241/// place. Non-mmap-backed sections (RAM directories) are already resident
242/// and are skipped silently.
243pub(crate) fn pin_section(
244    bytes: &mut OwnedBytes,
245    label: &str,
246    mode: PinMode,
247    remaining: &mut u64,
248    report: &mut PinReport,
249) {
250    if !bytes.is_mmap() || bytes.is_empty() {
251        return;
252    }
253    let len = bytes.len() as u64;
254    report.intended_bytes += len;
255
256    if len > *remaining {
257        report.skipped_budget_bytes += len;
258        log::debug!(
259            "[pin] budget exhausted: skipping {} ({} bytes, {} remaining)",
260            label,
261            len,
262            *remaining
263        );
264        return;
265    }
266
267    match mode {
268        PinMode::Mlock => {
269            if bytes.mlock() {
270                *remaining -= len;
271                report.pinned_bytes += len;
272            } else {
273                report.failed_bytes += len;
274                log::warn!(
275                    "[pin] mlock failed for {} ({} bytes) — check RLIMIT_MEMLOCK; \
276                     continuing unpinned",
277                    label,
278                    len
279                );
280            }
281        }
282        PinMode::Copy => {
283            *bytes = OwnedBytes::new(bytes.to_vec());
284            *remaining -= len;
285            report.pinned_bytes += len;
286        }
287    }
288}
hermes_core/segment/pin.rs

hermes_core/segment/
pin.rs
