mirrors/Zag
1
0
Fork 0
mirror of https://github.com/rdmsr/zag.git synced 2026-10-27 13:58:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

mm/zone: switch to linear-log bucketing

Browse Source 
While we're here, also shrink some types.
This commit is contained in:
Loïc Grégoire
2026-10-25 00:08:29 -04:00
parent 8958854a28
commit 6020484453
1 changed files with 144 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/sys/mm/zone.zig
+144 -87
View File
@@ -172,26 +172,17 @@ const excess_memory = ke.TunableType(
"mm.zone.excess_mem",
);
/// Every power-of-two-size from 8 to 2048 (inclusively).
const generic_zones_num = 9;
const generic_zone_names = [_][]const u8{
"zone-8",
"zone-16",
"zone-32",
"zone-64",
"zone-128",
"zone-256",
"zone-512",
"zone-1024",
"zone-2048",
};
/// The number of generic zones to generate.
const generic_zones_num = 33; // Up to 8192 bytes
/// 1/8th of a page.
const small_slab_size = 512;
/// Generic zones for power-of-two sizes from 8 to 2048.
/// Generic zones for general-purpose allocations.
var generic_zones: [generic_zones_num]Zone = undefined;
const generic_sizes = generate_generic_sizes();
const generic_names = generate_generic_names(generic_sizes);
const generic_max_size = bucket_to_size(generic_zones_num - 1);
var magazines_initialized = false;
var all_zones: ?*Zone = null;
@@ -361,11 +352,9 @@ const MagazineList = struct {
head: ?*Magazine,
tail: ?*Magazine,
/// Number of magazines.
num: usize,
num: u32,
/// Minimum since last update
min: usize,
/// Weighted moving average of `min`, scaled by `wma_unit`.
wma: usize,
min: u32,
};
const ReusePolicy = enum {
@@ -430,9 +419,9 @@ const Cpu = struct {
/// Magazine used for frees.
free: ?*Magazine,
/// Number of rounds in alloc magazine.
alloc_rounds: usize,
alloc_rounds: u8,
/// Number of rounds in free magazine.
free_rounds: usize,
free_rounds: u8,
/// Local per-CPU depot.
depot: Depot,
};
@@ -491,11 +480,10 @@ inline fn wma_mix(old: usize, new: usize) usize {
return (3 * old + new * wma_unit) / 4;
}
fn cpus_alloc() []rtl.CachePaddedType(Cpu) {
fn cpus_alloc() [*]rtl.CachePaddedType(Cpu) {
const ptr = cpu_zone.alloc(.{}) catch
@panic("failed to allocate zone CPU state");
const raw: [*]rtl.CachePaddedType(Cpu) = @ptrCast(@alignCast(ptr));
return raw[0..ke.ncpus];
return @ptrCast(@alignCast(ptr));
}
pub const Zone = struct {
@@ -511,8 +499,6 @@ pub const Zone = struct {
chunk_size: usize,
/// Zone lock.
lock: ke.Mutex,
/// List of full slabs (fully allocated) in this zone.
full_slabs: rtl.List,
/// List of partial slabs (partially allocated) in this zone.
partial_slabs: rtl.List,
/// List of free slabs (fully unallocated) in this zone.
@@ -525,21 +511,22 @@ pub const Zone = struct {
depot_contention_wma: usize,
depot_contention_cur: usize,
/// Size of the per-CPU depot.
cpu_depot_size: usize,
cpu_depot_size: u32,
/// Maximum size of the per-CPU depot.
cpu_depot_limit: usize,
trim_depot: bool,
/// Don't assume that the memory we're managing is actually memory.
dont_touch: bool,
cpu_depot_limit: u32,
/// Per-CPU state.
cpus: []rtl.CachePaddedType(Cpu),
cpus: [*]rtl.CachePaddedType(Cpu),
/// Various boolean flags (packed)
flags: Flags,
/// Weighted moving average of `depot.empty_mags.min`, scaled by `wma_unit`.
empty_mags_wma: usize,
/// Weighted moving average of `depot.full_mags.min`, scaled by `wma_unit`.
full_mags_wma: usize,
ctor: ?*const fn (obj: *anyopaque) void,
dtor: ?*const fn (obj: *anyopaque) void,
use_magazines: bool,
reuse_policy: ReusePolicy,
smr: ?*ke.smr.Domain,
@@ -584,6 +571,15 @@ pub const Zone = struct {
arg: ?*anyopaque = null,
};
const Flags = packed struct {
/// Don't assume that the memory we're managing is actually memory.
dont_touch: bool,
/// Whether magazines are enabled for this zone.
use_magazines: bool,
/// Internal flag set when the depot needs to be trimmed.
trim_depot: bool,
};
/// Initialize a zone.
pub fn init(
self: *Zone,
@@ -599,7 +595,6 @@ pub const Zone = struct {
else
Slab.calc_slab_size(chunk_size);
self.full_slabs.init();
self.partial_slabs.init();
self.free_slabs.init();
@@ -612,16 +607,18 @@ pub const Zone = struct {
self.dtor = options.dtor;
self.lock = ke.Mutex.init("zone");
self.depot_lock = ke.Mutex.init("zone_depot");
self.use_magazines = options.magazines;
self.flags.use_magazines = options.magazines;
self.depot = std.mem.zeroes(Depot);
self.depot_contention_cur = 0;
self.depot_contention_wma = 0;
self.cpu_depot_size = 0;
self.cpu_depot_limit = max_local_memory.load() /
(self.chunk_size * magazine_size.load());
self.trim_depot = false;
self.dont_touch = options.import != null;
self.cpu_depot_limit = @intCast(max_local_memory.load() /
(self.chunk_size * magazine_size.load()));
self.flags.trim_depot = false;
self.flags.dont_touch = options.import != null;
self.empty_mags_wma = 0;
self.full_mags_wma = 0;
// SMR zones have to be FIFO so that the minimum sequence can be
// obtained by checking the list head.
@@ -671,7 +668,7 @@ pub const Zone = struct {
/// Allocate an object from the zone.
pub fn alloc(self: *Zone, opts: AllocOpts) mm.Error!*anyopaque {
// Try grabbing an object from the magazine layer.
if (magazines_initialized and self.use_magazines) {
if (magazines_initialized and self.flags.use_magazines) {
@branchHint(.likely);
var buf: *anyopaque = undefined;
@@ -685,7 +682,7 @@ pub const Zone = struct {
cpu.alloc_rounds -= 1;
buf = alloc_mag.rounds_ptr()[cpu.alloc_rounds];
if (is_poison_enabled and !self.dont_touch) {
if (is_poison_enabled and !self.flags.dont_touch) {
if (!check_poison(buf, self.obj_size, free_poison)) {
@panic("poison mismatch: possible use-after-free");
}
@@ -708,7 +705,7 @@ pub const Zone = struct {
// swap them. The free magazine now becomes our alloc
// magazine.
std.mem.swap(?*Magazine, &cpu.alloc, &cpu.free);
std.mem.swap(usize, &cpu.alloc_rounds, &cpu.free_rounds);
std.mem.swap(u8, &cpu.alloc_rounds, &cpu.free_rounds);
continue;
}
@@ -736,7 +733,7 @@ pub const Zone = struct {
// Local depot is empty, rebalance with the zone depot.
// Purposefully race against cpu_depot_size because locking is
// tricky here, we may get a stale value but it is harmless.
const n = rtl.mem.racy_read(usize, &self.cpu_depot_size);
const n = rtl.mem.racy_read(u32, &self.cpu_depot_size);
self.alloc_depot_rebalance(n, cpu);
@@ -793,7 +790,7 @@ pub const Zone = struct {
if (cpu.alloc == null) {
cpu.alloc = nm;
cpu.alloc_rounds = cnt;
cpu.alloc_rounds = @intCast(cnt);
continue;
} else {
// We lost the race. If the magazine we got was fully
@@ -852,7 +849,7 @@ pub const Zone = struct {
/// Free an object back to the zone.
pub fn free(self: *Zone, obj: *anyopaque) void {
if (magazines_initialized and self.use_magazines) {
if (magazines_initialized and self.flags.use_magazines) {
@branchHint(.likely);
var cpu = &self.cpus[ke.cpu.current()].value;
@@ -869,7 +866,7 @@ pub const Zone = struct {
// the object.
const free_mag = cpu.free orelse unreachable;
if (is_poison_enabled and self.smr == null and
!self.dont_touch)
!self.flags.dont_touch)
{
if (self.dtor) |dtor| {
dtor(obj);
@@ -901,7 +898,7 @@ pub const Zone = struct {
// If the alloc magazine is completely empty, swap them.
// The empty alloc magazine becomes our new free magazine.
std.mem.swap(?*Magazine, &cpu.alloc, &cpu.free);
std.mem.swap(usize, &cpu.alloc_rounds, &cpu.free_rounds);
std.mem.swap(u8, &cpu.alloc_rounds, &cpu.free_rounds);
continue;
}
@@ -924,7 +921,7 @@ pub const Zone = struct {
// Local depot is empty, rebalance with the zone depot.
// Race explained in alloc.
const n = rtl.mem.racy_read(usize, &self.cpu_depot_size);
const n = rtl.mem.racy_read(u32, &self.cpu_depot_size);
self.free_depot_rebalance(n, cpu);
@@ -986,13 +983,13 @@ pub const Zone = struct {
}
fn slab_free(self: *Zone, obj: *anyopaque, poison: bool) void {
if (!is_poison_enabled or poison and !self.dont_touch) {
if (!is_poison_enabled or poison and !self.flags.dont_touch) {
if (self.dtor) |dtor| {
dtor(obj);
}
}
if (is_poison_enabled and poison and !self.dont_touch) {
if (is_poison_enabled and poison and !self.flags.dont_touch) {
fill_with_poison(obj, self.obj_size, free_poison);
}
@@ -1021,7 +1018,6 @@ pub const Zone = struct {
if (was_full) {
// There were no buffers in the slab, so it wasn't in the freelist.
// Now that there is a buffer, add it back to the freelist.
slab.link.remove();
self.partial_slabs.insert_tail(&slab.link);
}
@@ -1082,7 +1078,6 @@ pub const Zone = struct {
if (slab.refcount == slab.capacity) {
slab.link.remove();
self.full_slabs.insert_tail(&slab.link);
}
if (is_poison_enabled) {
@@ -1146,14 +1141,14 @@ pub const Zone = struct {
// interval, then the working set size is 10 and we can
// reclaim 37 magazines. We keep track of this in a weighted moving
// average (WMA) to bias recent utilization when trimming.
self.depot.empty_mags.wma = wma_mix(
self.depot.empty_mags.wma,
self.empty_mags_wma = wma_mix(
self.empty_mags_wma,
self.depot.empty_mags.min,
);
self.depot.empty_mags.min = self.depot.empty_mags.num;
self.depot.full_mags.wma = wma_mix(
self.depot.full_mags.wma,
self.full_mags_wma = wma_mix(
self.full_mags_wma,
self.depot.full_mags.min,
);
self.depot.full_mags.min = self.depot.full_mags.num;
@@ -1163,10 +1158,10 @@ pub const Zone = struct {
var cur = self.depot_contention_cur * wma_unit /
(ke.ncpus * update_interval_s);
// WMA formula.
// EWMA formula.
cur = (3 * old + cur) / 4;
if (self.use_magazines) {
if (self.flags.use_magazines) {
if (self.cpu_depot_size < self.cpu_depot_limit and
cur > depot_grow_level.load())
{
@@ -1176,31 +1171,34 @@ pub const Zone = struct {
// good or not.
cur = (depot_grow_level.load() + depot_shrink_level.load()) / 2;
const size = if (self.cpu_depot_size == 0)
const size: usize = if (self.cpu_depot_size == 0)
2
else
// Grow by 1.5x.
(3 * self.cpu_depot_size) / 2;
(3 * @as(usize, self.cpu_depot_size)) / 2;
// Clamp it.
self.cpu_depot_size = @min(size, self.cpu_depot_limit);
self.cpu_depot_size = @intCast(@min(
size,
@as(usize, self.cpu_depot_limit),
));
} else if (self.cpu_depot_size > 0 and
cur <= depot_shrink_level.load())
{
// We should shrink the depot.
cur = (depot_grow_level.load() + depot_shrink_level.load()) / 2;
self.cpu_depot_size -= 1;
self.trim_depot = true;
self.flags.trim_depot = true;
}
}
self.depot_contention_cur = 0;
self.depot_contention_wma = cur;
const trim_depot = self.trim_depot;
const trim_depot = self.flags.trim_depot;
const should_trim = self.trim_needed();
self.trim_depot = false;
self.flags.trim_depot = false;
self.depot_lock.release();
@@ -1211,12 +1209,12 @@ pub const Zone = struct {
/// Returns whether or not we should trim excess magazines.
fn trim_needed(self: *Zone) bool {
if (!self.use_magazines) return false;
if (self.trim_depot) return true;
if (!self.flags.use_magazines) return false;
if (self.flags.trim_depot) return true;
const empty = @min(
self.depot.empty_mags.wma,
self.depot.empty_mags.min * wma_unit,
self.empty_mags_wma,
@as(usize, self.depot.empty_mags.min) * wma_unit,
);
if (empty > excess_magazines.load() * wma_unit) {
@@ -1225,8 +1223,8 @@ pub const Zone = struct {
}
const full = @min(
self.depot.full_mags.wma,
self.depot.full_mags.min * wma_unit,
self.full_mags_wma,
@as(usize, self.depot.full_mags.min) * wma_unit,
);
const full_bytes = full * magazine_size.load() * self.chunk_size;
@@ -1249,8 +1247,14 @@ pub const Zone = struct {
// unlocked.
self.lock.acquire();
const empty_mags = self.trim_maglist(&self.depot.empty_mags);
const full_mags = self.trim_maglist(&self.depot.full_mags);
const empty_mags = self.trim_maglist(
&self.depot.empty_mags,
&self.empty_mags_wma,
);
const full_mags = self.trim_maglist(
&self.depot.full_mags,
&self.full_mags_wma,
);
var depot = std.mem.zeroes(Depot);
if (trim_depot) {
@@ -1287,7 +1291,7 @@ pub const Zone = struct {
/// Called when memory is low and we need to make memory ASAP.
/// Drains all magazines from the depot.
fn drain(self: *Zone) void {
if (!self.use_magazines) return;
if (!self.flags.use_magazines) return;
// Detach magazines before invoking release callbacks.
self.lock.acquire();
@@ -1297,6 +1301,8 @@ pub const Zone = struct {
// Take all elements from the zone depot.
var depot = self.depot;
self.depot = std.mem.zeroes(Depot);
self.empty_mags_wma = 0;
self.full_mags_wma = 0;
self.depot_lock.release();
@@ -1328,18 +1334,22 @@ pub const Zone = struct {
}
/// Detach excess magazines for destruction.
fn trim_maglist(self: *Zone, maglist: *MagazineList) ?*Magazine {
fn trim_maglist(
self: *Zone,
maglist: *MagazineList,
wma: *usize,
) ?*Magazine {
self.depot_lock.acquire();
// Grab the magazines and WSS under the depot lock.
const target = @min(
maglist.min * wma_unit,
maglist.wma,
@as(usize, maglist.min) * wma_unit,
wma.*,
) / wma_unit;
var removed_head: ?*Magazine = maglist.head;
var last_removed: ?*Magazine = null;
var removed: usize = 0;
var removed: u32 = 0;
for (0..target) |_| {
if (maglist.head) |head| {
@@ -1361,7 +1371,7 @@ pub const Zone = struct {
maglist.min -= removed;
maglist.num -= removed;
maglist.wma -= removed * wma_unit;
wma.* -= @as(usize, removed) * wma_unit;
self.depot_lock.release();
@@ -1790,7 +1800,7 @@ pub fn early_init() linksection(r.init) void {
zone_list_lock = ke.Mutex.init("zone_list");
for (0..generic_zones.len, &generic_zones) |i, *zone| {
zone.init(generic_zone_names[i], @as(usize, 1) << @intCast(i + 3), .{});
zone.init(generic_names[i], generic_sizes[i], .{});
}
magazine_zone.init(
@@ -1815,7 +1825,7 @@ pub fn late_init() linksection(r.init) void {
var zone = all_zones;
while (zone) |z| {
if (z.use_magazines) {
if (z.flags.use_magazines) {
z.cpus = cpus_alloc();
for (0..ke.ncpus) |i| {
@@ -1949,11 +1959,58 @@ fn gpa_free(
}
fn zone_for(size: usize) ?*Zone {
for (&generic_zones) |*zone| {
if (zone.obj_size >= size) {
return zone;
}
if (size > generic_max_size) return null;
return &generic_zones[size_to_bucket(size)];
}
/// Linear-log bucketing is used for sizing the generic buckets, see
/// https://pvk.ca/Blog/2015/06/27/linear-log-bucketing-fast-versatile-simple.
/// These formulas come from jemalloc.
const N = generic_zones_num;
fn bucket_to_size(index: usize) usize {
if (index == 0) return 8;
const i = index - 1;
const group = i / 4;
const offset = i % 4;
if (group == 0) {
return (offset + 1) * 16;
}
return null;
return (offset + 5) << (group + 3);
}
fn size_to_bucket(size: usize) usize {
if (size <= 8) return 0;
if (size <= 64) return ((size - 1) / 16) + 1;
const lg: usize = std.math.log2_int(usize, size - 1);
const group = lg - 5;
const shift: std.math.Log2Int(usize) = @intCast(lg - 2);
const offset = ((size - 1) >> shift) - 4;
return 1 + (group * 4) + offset;
}
fn generate_generic_sizes() [N]usize {
var arr: [N]usize = undefined;
for (0..N) |i| {
arr[i] = comptime bucket_to_size(i);
}
return arr;
}
fn generate_generic_names(
comptime array: [N]usize,
) [N][]const u8 {
var ret: [N][]const u8 = undefined;
for (0..N) |i| {
ret[i] = std.fmt.comptimePrint("zone-{}", .{array[i]});
}
return ret;
}