Skip to main content

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    /// Additional heap allocated by `PinMode::Copy`. Already-heap ANN routing
108    /// structures are resident but do not contribute here.
109    pub heap_copy_bytes: u64,
110}
111
112/// RAII owner for heap pages locked on behalf of one immutable ANN artifact
113/// generation. The referenced allocations are owned by the same
114/// `TrainedVectorStructures`; its field order drops this set before the
115/// artifact `Arc`s, so every address remains valid through `munlock`.
116struct HeapPinGuard {
117    page_start: *mut libc::c_void,
118    page_len: usize,
119}
120
121// The guard never dereferences its pointer. The immutable artifact allocations
122// it describes are safe to share, and mlock/munlock operate on process mappings.
123unsafe impl Send for HeapPinGuard {}
124unsafe impl Sync for HeapPinGuard {}
125
126impl Drop for HeapPinGuard {
127    fn drop(&mut self) {
128        if unsafe { libc::munlock(self.page_start, self.page_len) } != 0 {
129            log::warn!(
130                "[pin] munlock failed for {} of ANN heap: {}",
131                crate::format_bytes(self.page_len as u64),
132                std::io::Error::last_os_error()
133            );
134        }
135    }
136}
137
138/// Locked heap allocations associated with one index-global ANN generation.
139/// Segment-local vector/code payloads are intentionally excluded.
140#[derive(Default)]
141pub(crate) struct HeapPinSet {
142    guards: Vec<HeapPinGuard>,
143    /// Keep every allocation owner alive until after its guards are dropped,
144    /// even if a cloned `TrainedVectorStructures` has its public maps mutated.
145    owners: Vec<Arc<dyn std::any::Any + Send + Sync>>,
146    report: PinReport,
147}
148
149impl HeapPinSet {
150    pub(crate) fn report(&self) -> PinReport {
151        self.report
152    }
153
154    pub(crate) fn retain_owner<T: std::any::Any + Send + Sync>(&mut self, owner: Arc<T>) {
155        self.owners.push(owner);
156    }
157
158    /// Keep one immutable heap slice resident, subject to the generation
159    /// budget. `Copy` mode needs no allocation: trained artifacts are already
160    /// heap-owned, which is exactly the residency guarantee that mode provides
161    /// on the supported swapless deployment.
162    pub(crate) fn pin_slice<T>(
163        &mut self,
164        slice: &[T],
165        label: &str,
166        mode: PinMode,
167        remaining: &mut u64,
168    ) {
169        let len = std::mem::size_of_val(slice);
170        if len == 0 {
171            return;
172        }
173        let Ok(len_u64) = u64::try_from(len) else {
174            self.report.failed_bytes = u64::MAX;
175            log::warn!("[pin] ANN region {label} is too large to account");
176            return;
177        };
178        self.report.intended_bytes = self.report.intended_bytes.saturating_add(len_u64);
179        if len_u64 > *remaining {
180            self.report.skipped_budget_bytes =
181                self.report.skipped_budget_bytes.saturating_add(len_u64);
182            log::debug!(
183                "[pin] ANN budget exhausted: skipping {} ({}, {} remaining)",
184                label,
185                crate::format_bytes(len_u64),
186                crate::format_bytes(*remaining)
187            );
188            return;
189        }
190
191        if mode == PinMode::Copy {
192            *remaining -= len_u64;
193            self.report.pinned_bytes = self.report.pinned_bytes.saturating_add(len_u64);
194            return;
195        }
196
197        let page_size = unsafe { libc::sysconf(libc::_SC_PAGESIZE) };
198        let page_size = usize::try_from(page_size).ok().filter(|&size| size > 0);
199        let Some(page_size) = page_size else {
200            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
201            log::warn!("[pin] cannot determine page size while locking {label}");
202            return;
203        };
204        let address = slice.as_ptr() as usize;
205        let page_start = address / page_size * page_size;
206        let Some(end) = address.checked_add(len) else {
207            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
208            log::warn!("[pin] ANN region address overflow while locking {label}");
209            return;
210        };
211        let Some(rounded_end) = end
212            .checked_add(page_size - 1)
213            .map(|value| value / page_size * page_size)
214        else {
215            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
216            log::warn!("[pin] ANN region page range overflow while locking {label}");
217            return;
218        };
219        let page_len = rounded_end - page_start;
220        let page_start = page_start as *mut libc::c_void;
221        if unsafe { libc::mlock(page_start.cast_const(), page_len) } == 0 {
222            self.guards.push(HeapPinGuard {
223                page_start,
224                page_len,
225            });
226            *remaining -= len_u64;
227            self.report.pinned_bytes = self.report.pinned_bytes.saturating_add(len_u64);
228        } else {
229            self.report.failed_bytes = self.report.failed_bytes.saturating_add(len_u64);
230            log::warn!(
231                "[pin] mlock failed for ANN {} ({}): {} — check RLIMIT_MEMLOCK/CAP_IPC_LOCK; continuing unpinned",
232                label,
233                crate::format_bytes(len_u64),
234                std::io::Error::last_os_error()
235            );
236        }
237    }
238}
239
240/// Pin one metadata section, updating `remaining` budget and the report.
241///
242/// In `Copy` mode the section is replaced with a heap copy (heap memory is
243/// not page-cache-evictable). In `Mlock` mode the mmap pages are locked in
244/// place. Non-mmap-backed sections (RAM directories) are already resident
245/// and are skipped silently.
246pub(crate) fn pin_section(
247    bytes: &mut OwnedBytes,
248    label: &str,
249    mode: PinMode,
250    remaining: &mut u64,
251    report: &mut PinReport,
252) {
253    if !bytes.is_mmap() || bytes.is_empty() {
254        return;
255    }
256    let len = bytes.len() as u64;
257    report.intended_bytes += len;
258
259    if len > *remaining {
260        report.skipped_budget_bytes += len;
261        log::debug!(
262            "[pin] budget exhausted: skipping {} ({}, {} remaining)",
263            label,
264            crate::format_bytes(len),
265            crate::format_bytes(*remaining)
266        );
267        return;
268    }
269
270    match mode {
271        PinMode::Mlock => {
272            if bytes.mlock() {
273                *remaining -= len;
274                report.pinned_bytes += len;
275            } else {
276                report.failed_bytes += len;
277                log::warn!(
278                    "[pin] mlock failed for {} ({}) — check RLIMIT_MEMLOCK; \
279                     continuing unpinned",
280                    label,
281                    crate::format_bytes(len)
282                );
283            }
284        }
285        PinMode::Copy => {
286            *bytes = OwnedBytes::new(bytes.to_vec());
287            *remaining -= len;
288            report.pinned_bytes += len;
289            report.heap_copy_bytes += len;
290        }
291    }
292}