mm: add node-preferring page allocation (alloc_page_on_node)

Makefile

@@ -1131,6 +1131,7 @@ endif
 objects += linux.o
 objects += core/commands.o
 objects += core/sched.o
+objects += core/numa.o
 objects += core/mmio.o
 objects += core/kprintf.o
 ifeq ($(conf_tracepoints),1)

core/mempool.cc

@@ -23,6 +23,7 @@
 #endif
 #include <atomic>
 #include <osv/mmu.hh>
+#include <osv/numa.hh>
 #include <osv/trace.hh>
 #include <lockfree/ring.hh>
 #include <osv/percpu-worker.hh>
@@ -597,6 +598,11 @@ class page_range_allocator {
     page_range* alloc(size_t size, bool contiguous = true);
     page_range* alloc_aligned(size_t size, size_t offset, size_t alignment,
                               bool fill = false);
+    // Carve a single page from a free range that lies on NUMA node `node`,
+    // returning it, or nullptr if no free memory on that node is available.
+    // Used by the node-preferring allocator; the caller falls back to a normal
+    // allocation when this returns nullptr.
+    page_range* alloc_page_from_node(int node);
     void free(page_range* pr);
 
     void initial_add(page_range* pr);
@@ -796,6 +802,38 @@ page_range* page_range_allocator::alloc(size_t size, bool contiguous)
     return &pr;
 }
 
+page_range* page_range_allocator::alloc_page_from_node(int node)
+{
+    // Find a free range whose memory lies on the requested NUMA node, then
+    // carve a single page from its front (mirroring alloc()).  This is a
+    // best-effort preference: a linear scan over the free lists, returning the
+    // first range on the node.  Returns nullptr if none is found, so the caller
+    // can fall back to a node-agnostic allocation.
+    page_range* found = nullptr;
+    for_each(0, [&] (page_range& pr) {
+        auto phys = mmu::virt_to_phys(static_cast<void*>(&pr));
+        if (numa::node_of_phys(phys) == node) {
+            found = &pr;
+            return false;   // stop iterating
+        }
+        return true;
+    });
+    if (!found) {
+        return nullptr;
+    }
+
+    auto& pr = *found;
+    remove(pr);
+    if (pr.size > page_size) {
+        auto& np = *new (static_cast<void*>(&pr) + page_size)
+                        page_range(pr.size - page_size);
+        insert(np);
+        pr.size = page_size;
+    }
+    set_bits(pr, false);
+    return &pr;
+}
+
 page_range* page_range_allocator::alloc_aligned(size_t size, size_t offset,
                                                 size_t alignment, bool fill)
 {
@@ -1789,6 +1827,40 @@ void* alloc_page()
     return p;
 }
 
+// Allocate one page, preferring memory local to NUMA node `node`.  Falls back
+// to a node-agnostic allocation when the node has no free memory local to it
+// in the global page-range allocator (or NUMA is not available).
+//
+// This is a best-effort node-preferring hint that the scheduler and
+// mbind/set_mempolicy can build on.  It only sees memory still held by the
+// global page-range allocator: OSv drains much of physical memory into per-CPU
+// L1/L2 pools that are not (yet) node-partitioned, so once a node's global
+// ranges are exhausted this falls back.  Full per-node pools are future work
+// (roadmap B2.2-full); this step gives correct placement while node-local
+// global memory is available and never fails or misplaces.
+void* alloc_page_on_node(int node)
+{
+    if (node < 0 || !numa::available() || (unsigned)node >= numa::nr_nodes()) {
+        return alloc_page();
+    }
+    void* ret = nullptr;
+    WITH_LOCK(free_page_ranges_lock) {
+        page_range* pr = free_page_ranges.alloc_page_from_node(node);
+        if (pr) {
+            on_alloc(page_size);
+            ret = static_cast<void*>(pr);
+        }
+    }
+    if (!ret) {
+        // No free memory on that node: fall back rather than fail.
+        return alloc_page();
+    }
+#if CONF_memory_tracker
+    tracker_remember(ret, page_size);
+#endif
+    return ret;
+}
+
 static inline void untracked_free_page(void *v)
 {
     trace_memory_page_free(v);

core/numa.cc

@@ -0,0 +1,188 @@
+/*
+ * Copyright (C) 2026 Waldemar Kozaczuk
+ *
+ * This work is open source software, licensed under the terms of the
+ * BSD license as described in the LICENSE file in the top-level directory.
+ */
+
+#include <osv/numa.hh>
+#include <osv/sched.hh>
+#include <osv/debug.hh>
+
+#include <unordered_map>
+#include <algorithm>
+
+#include <osv/drivers_config.h>
+
+#if CONF_drivers_acpi
+extern "C" {
+#include "acpi.h"
+}
+#include <boost/intrusive/parent_from_member.hpp>
+#endif
+
+namespace numa {
+
+static bool s_available = false;
+static unsigned s_nr_nodes = 1;
+// Map from a raw APIC id to the NUMA node it belongs to (from SRAT).
+static std::unordered_map<uint32_t, unsigned> s_apic_to_node;
+// Map from a sched cpu id to its NUMA node (resolved via APIC id).
+static std::unordered_map<unsigned, unsigned> s_cpu_to_node;
+// SLIT distance matrix, row-major, s_nr_nodes x s_nr_nodes; empty if no SLIT.
+static std::vector<uint8_t> s_distances;
+static std::vector<mem_range> s_mem_ranges;
+static bool s_initialized = false;
+
+unsigned nr_nodes() { return s_nr_nodes; }
+bool available() { return s_available; }
+const std::vector<mem_range>& memory_ranges() { return s_mem_ranges; }
+
+unsigned node_of_cpu(unsigned cpu_id)
+{
+    auto it = s_cpu_to_node.find(cpu_id);
+    return it == s_cpu_to_node.end() ? 0 : it->second;
+}
+
+int node_of_phys(uint64_t phys)
+{
+    for (auto& r : s_mem_ranges) {
+        if (phys >= r.base && phys < r.base + r.length) {
+            return (int)r.node;
+        }
+    }
+    return -1;
+}
+
+unsigned distance(unsigned from, unsigned to)
+{
+    if (from == to) {
+        return 10;   // ACPI convention: 10 == local.
+    }
+    if (!s_distances.empty() && from < s_nr_nodes && to < s_nr_nodes) {
+        return s_distances[from * s_nr_nodes + to];
+    }
+    return 20;       // Default remote distance when no SLIT is present.
+}
+
+#if CONF_drivers_acpi
+using boost::intrusive::get_parent_from_member;
+
+// Record the (apic id -> node) mapping and track the highest node seen.
+static void record_cpu_affinity(uint32_t apic_id, unsigned node, unsigned& max_node)
+{
+    s_apic_to_node[apic_id] = node;
+    max_node = std::max(max_node, node);
+}
+
+static void parse_srat()
+{
+    char sig[] = ACPI_SIG_SRAT;
+    ACPI_TABLE_HEADER* header;
+    if (AcpiGetTable(sig, 0, &header) != AE_OK) {
+        return;   // No SRAT: leave the single-node fallback in place.
+    }
+    auto srat = get_parent_from_member(header, &ACPI_TABLE_SRAT::Header);
+    void* sub = srat + 1;
+    void* end = static_cast<void*>(srat) + srat->Header.Length;
+    unsigned max_node = 0;
+
+    while (sub < end) {
+        auto s = static_cast<ACPI_SUBTABLE_HEADER*>(sub);
+        if (s->Length == 0) {
+            break;   // Guard against a malformed zero-length subtable.
+        }
+        switch (s->Type) {
+        case ACPI_SRAT_TYPE_CPU_AFFINITY: {
+            auto a = get_parent_from_member(s, &ACPI_SRAT_CPU_AFFINITY::Header);
+            if (a->Flags & ACPI_SRAT_CPU_ENABLED) {
+                unsigned node = a->ProximityDomainLo |
+                    (a->ProximityDomainHi[0] << 8) |
+                    (a->ProximityDomainHi[1] << 16) |
+                    (a->ProximityDomainHi[2] << 24);
+                record_cpu_affinity(a->ApicId, node, max_node);
+            }
+            break;
+        }
+        case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY: {
+            auto a = get_parent_from_member(s, &ACPI_SRAT_X2APIC_CPU_AFFINITY::Header);
+            if (a->Flags & ACPI_SRAT_CPU_ENABLED) {
+                record_cpu_affinity(a->ApicId, a->ProximityDomain, max_node);
+            }
+            break;
+        }
+        case ACPI_SRAT_TYPE_MEMORY_AFFINITY: {
+            auto m = get_parent_from_member(s, &ACPI_SRAT_MEM_AFFINITY::Header);
+            if (m->Flags & ACPI_SRAT_MEM_ENABLED) {
+                s_mem_ranges.push_back(mem_range{
+                    m->BaseAddress, m->Length, m->ProximityDomain,
+                    (m->Flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) != 0});
+                max_node = std::max(max_node, (unsigned)m->ProximityDomain);
+            }
+            break;
+        }
+        default:
+            break;
+        }
+        sub = static_cast<void*>(sub) + s->Length;
+    }
+
+    if (!s_apic_to_node.empty() || !s_mem_ranges.empty()) {
+        s_available = true;
+        s_nr_nodes = max_node + 1;
+    }
+}
+
+static void parse_slit()
+{
+    char sig[] = ACPI_SIG_SLIT;
+    ACPI_TABLE_HEADER* header;
+    if (AcpiGetTable(sig, 0, &header) != AE_OK) {
+        return;
+    }
+    auto slit = get_parent_from_member(header, &ACPI_TABLE_SLIT::Header);
+    uint64_t n = slit->LocalityCount;
+    // Only trust SLIT if it agrees with the node count we saw in SRAT.
+    if (n == 0 || n != s_nr_nodes) {
+        return;
+    }
+    s_distances.assign(slit->Entry, slit->Entry + n * n);
+}
+
+// Resolve the (apic id -> node) map into a (sched cpu id -> node) map.
+static void resolve_cpus()
+{
+    for (auto* c : sched::cpus) {
+        auto it = s_apic_to_node.find(c->arch.apic_id);
+        if (it != s_apic_to_node.end()) {
+            s_cpu_to_node[c->id] = it->second;
+        }
+    }
+}
+#endif
+
+void init()
+{
+    if (s_initialized) {
+        return;
+    }
+    s_initialized = true;
+
+#if CONF_drivers_acpi
+    parse_srat();
+    if (s_available) {
+        parse_slit();
+        resolve_cpus();
+    }
+#endif
+
+    if (s_available) {
+        debugf("NUMA: %u node(s), %zu CPU(s) mapped, %zu memory range(s)%s\n",
+               s_nr_nodes, s_cpu_to_node.size(), s_mem_ranges.size(),
+               s_distances.empty() ? ", no SLIT" : "");
+    } else {
+        debugf("NUMA: no SRAT, assuming a single flat node\n");
+    }
+}
+
+}

include/osv/numa.hh

@@ -0,0 +1,67 @@
+/*
+ * Copyright (C) 2026 Waldemar Kozaczuk
+ *
+ * This work is open source software, licensed under the terms of the
+ * BSD license as described in the LICENSE file in the top-level directory.
+ */
+
+#ifndef OSV_NUMA_HH
+#define OSV_NUMA_HH
+
+#include <cstdint>
+#include <vector>
+#include <cstddef>
+
+// NUMA topology discovery.
+//
+// This module parses the ACPI SRAT (System Resource Affinity Table) and SLIT
+// (System Locality Distance Information Table) to learn the machine's NUMA
+// layout: which node each CPU belongs to, which physical memory ranges belong
+// to each node, and the relative distances between nodes.
+//
+// This is discovery only: it does not change how memory is allocated or how
+// threads are scheduled.  It exposes the topology so later work (a node-aware
+// allocator, scheduler affinity, mbind/get_mempolicy) can use it.  On a machine
+// with no SRAT (the common single-node virtual machine), the whole system is
+// reported as one node (node 0) containing every CPU.
+
+namespace numa {
+
+// A contiguous physical memory range assigned to a NUMA node.
+struct mem_range {
+    uint64_t base;
+    uint64_t length;
+    unsigned node;
+    bool     hotpluggable;
+};
+
+// Discover the topology by parsing SRAT/SLIT.  Safe to call once, after ACPI is
+// initialized and the CPUs have been enumerated (so APIC ids are known).  If no
+// SRAT is present, initializes a single flat node.  Idempotent.
+void init();
+
+// Number of NUMA nodes (>= 1).
+unsigned nr_nodes();
+
+// True if the topology came from a real SRAT (as opposed to the synthesized
+// single-node fallback).
+bool available();
+
+// The node a CPU (by sched cpu id) belongs to, or 0 if unknown.
+unsigned node_of_cpu(unsigned cpu_id);
+
+// The node that owns a physical address, or -1 if it falls in no known range.
+// Only meaningful when available() is true.
+int node_of_phys(uint64_t phys);
+
+// The SLIT distance from node `from` to node `to`.  Linux/ACPI convention:
+// 10 == local (same node), higher == farther.  Returns 10 for the diagonal and
+// a default (10 local / 20 remote) when no SLIT is present.
+unsigned distance(unsigned from, unsigned to);
+
+// The memory ranges discovered from SRAT (empty if none).
+const std::vector<mem_range>& memory_ranges();
+
+}
+
+#endif

include/osv/pagealloc.hh

@@ -13,6 +13,10 @@
 namespace memory {
 
 void* alloc_page();
+// Allocate one page preferring memory local to NUMA node `node`, falling back
+// to a node-agnostic allocation when the node has no free memory or NUMA is
+// unavailable.
+void* alloc_page_on_node(int node);
 void free_page(void* page);
 void* alloc_huge_page(size_t bytes);
 void free_huge_page(void *page, size_t bytes);