| ... | @@ -99,12 +99,14 @@ const math = std.math; | ... | @@ -99,12 +99,14 @@ const math = std.math; |
| 99 | const assert = std.debug.assert; | 99 | const assert = std.debug.assert; |
| 100 | const mem = std.mem; | 100 | const mem = std.mem; |
| 101 | const Allocator = std.mem.Allocator; | 101 | const Allocator = std.mem.Allocator; |
| | 102 | const StackTrace = std.builtin.StackTrace; |
| | 103 | |
| 102 | const page_size: usize = @max(std.heap.page_size_max, switch (builtin.os.tag) { | 104 | const page_size: usize = @max(std.heap.page_size_max, switch (builtin.os.tag) { |
| 103 | .windows => 64 * 1024, // Makes `std.heap.PageAllocator` take the happy path. | 105 | .windows => 64 * 1024, // Makes `std.heap.PageAllocator` take the happy path. |
| 104 | .wasi => 64 * 1024, // Max alignment supported by `std.heap.WasmAllocator`. | 106 | .wasi => 64 * 1024, // Max alignment supported by `std.heap.WasmAllocator`. |
| 105 | else => 2 * 1024 * 1024, // Avoids too many active mappings when `page_size_max` is low. | 107 | else => 2 * 1024 * 1024, // Avoids too many active mappings when `page_size_max` is low. |
| 106 | }); | 108 | }); |
| 107 | const StackTrace = std.builtin.StackTrace; | 109 | const page_align: mem.Alignment = .fromByteUnits(page_size); |
| 108 | | 110 | |
| 109 | /// Integer type for pointing to slots in a small allocation | 111 | /// Integer type for pointing to slots in a small allocation |
| 110 | const SlotIndex = std.meta.Int(.unsigned, math.log2(page_size) + 1); | 112 | const SlotIndex = std.meta.Int(.unsigned, math.log2(page_size) + 1); |
| ... | @@ -168,29 +170,23 @@ pub const Check = enum { ok, leak }; | ... | @@ -168,29 +170,23 @@ pub const Check = enum { ok, leak }; |
| 168 | pub fn GeneralPurposeAllocator(comptime config: Config) type { | 170 | pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 169 | return struct { | 171 | return struct { |
| 170 | backing_allocator: Allocator = std.heap.page_allocator, | 172 | backing_allocator: Allocator = std.heap.page_allocator, |
| 171 | buckets: [small_bucket_count]Buckets = [1]Buckets{Buckets{}} ** small_bucket_count, | 173 | /// Tracks the active bucket, which is the one that has free slots in it. |
| 172 | cur_buckets: [small_bucket_count]?*BucketHeader = [1]?*BucketHeader{null} ** small_bucket_count, | 174 | buckets: [small_bucket_count]?*BucketHeader = [1]?*BucketHeader{null} ** small_bucket_count, |
| 173 | large_allocations: LargeAllocTable = .{}, | 175 | large_allocations: LargeAllocTable = .empty, |
| 174 | empty_buckets: if (config.retain_metadata) Buckets else void = | | |
| 175 | if (config.retain_metadata) Buckets{} else {}, | | |
| 176 | bucket_node_pool: std.heap.MemoryPool(Buckets.Node) = std.heap.MemoryPool(Buckets.Node).init(std.heap.page_allocator), | | |
| 177 | | | |
| 178 | total_requested_bytes: @TypeOf(total_requested_bytes_init) = total_requested_bytes_init, | 176 | total_requested_bytes: @TypeOf(total_requested_bytes_init) = total_requested_bytes_init, |
| 179 | requested_memory_limit: @TypeOf(requested_memory_limit_init) = requested_memory_limit_init, | 177 | requested_memory_limit: @TypeOf(requested_memory_limit_init) = requested_memory_limit_init, |
| 180 | | | |
| 181 | mutex: @TypeOf(mutex_init) = mutex_init, | 178 | mutex: @TypeOf(mutex_init) = mutex_init, |
| 182 | | 179 | |
| 183 | const Self = @This(); | 180 | const Self = @This(); |
| 184 | | 181 | |
| 185 | /// The initial state of a `GeneralPurposeAllocator`, containing no | 182 | pub const init: Self = .{}; |
| 186 | /// allocations and backed by the system page allocator. | 183 | |
| 187 | pub const init: Self = .{ | 184 | /// These can be derived from size_class_index but the calculation is nontrivial. |
| 188 | .backing_allocator = std.heap.page_allocator, | 185 | const slot_counts: [small_bucket_count]SlotIndex = init: { |
| 189 | .buckets = [1]Buckets{.{}} ** small_bucket_count, | 186 | @setEvalBranchQuota(10000); |
| 190 | .cur_buckets = [1]?*BucketHeader{null} ** small_bucket_count, | 187 | var result: [small_bucket_count]SlotIndex = undefined; |
| 191 | .large_allocations = .{}, | 188 | for (&result, 0..) |*elem, i| elem.* = calculateSlotCount(i); |
| 192 | .empty_buckets = if (config.retain_metadata) .{} else {}, | 189 | break :init result; |
| 193 | .bucket_node_pool = .init(std.heap.page_allocator), | | |
| 194 | }; | 190 | }; |
| 195 | | 191 | |
| 196 | const total_requested_bytes_init = if (config.enable_memory_limit) @as(usize, 0) else {}; | 192 | const total_requested_bytes_init = if (config.enable_memory_limit) @as(usize, 0) else {}; |
| ... | @@ -204,8 +200,8 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -204,8 +200,8 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 204 | DummyMutex{}; | 200 | DummyMutex{}; |
| 205 | | 201 | |
| 206 | const DummyMutex = struct { | 202 | const DummyMutex = struct { |
| 207 | fn lock(_: *DummyMutex) void {} | 203 | inline fn lock(_: *DummyMutex) void {} |
| 208 | fn unlock(_: *DummyMutex) void {} | 204 | inline fn unlock(_: *DummyMutex) void {} |
| 209 | }; | 205 | }; |
| 210 | | 206 | |
| 211 | const stack_n = config.stack_trace_frames; | 207 | const stack_n = config.stack_trace_frames; |
| ... | @@ -223,7 +219,6 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -223,7 +219,6 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 223 | return std.math.order(@intFromPtr(a.page), @intFromPtr(b.page)); | 219 | return std.math.order(@intFromPtr(a.page), @intFromPtr(b.page)); |
| 224 | } | 220 | } |
| 225 | }.compare; | 221 | }.compare; |
| 226 | const Buckets = std.Treap(*BucketHeader, bucketCompare); | | |
| 227 | | 222 | |
| 228 | const LargeAlloc = struct { | 223 | const LargeAlloc = struct { |
| 229 | bytes: []u8, | 224 | bytes: []u8, |
| ... | @@ -259,46 +254,50 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -259,46 +254,50 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 259 | }; | 254 | }; |
| 260 | const LargeAllocTable = std.AutoHashMapUnmanaged(usize, LargeAlloc); | 255 | const LargeAllocTable = std.AutoHashMapUnmanaged(usize, LargeAlloc); |
| 261 | | 256 | |
| 262 | // Bucket: In memory, in order: | 257 | /// Bucket: In memory, in order: |
| 263 | // * BucketHeader | 258 | /// * BucketHeader |
| 264 | // * bucket_used_bits: [N]u8, // 1 bit for every slot; 1 byte for every 8 slots | 259 | /// * bucket_used_bits: [N]u8, // 1 bit for every slot; 1 byte for every 8 slots |
| 265 | // -- below only exists when config.safety is true -- | 260 | /// -- below only exists when config.safety is true -- |
| 266 | // * requested_sizes: [N]LargestSizeClassInt // 1 int for every slot | 261 | /// * requested_sizes: [N]LargestSizeClassInt // 1 int for every slot |
| 267 | // * log2_ptr_aligns: [N]u8 // 1 byte for every slot | 262 | /// * log2_ptr_aligns: [N]u8 // 1 byte for every slot |
| 268 | // -- above only exists when config.safety is true -- | 263 | /// -- above only exists when config.safety is true -- |
| 269 | // * stack_trace_addresses: [N]usize, // traces_per_slot for every allocation | 264 | /// * stack_trace_addresses: [N]usize, // traces_per_slot for every allocation |
| 270 | | | |
| 271 | const BucketHeader = struct { | 265 | const BucketHeader = struct { |
| 272 | page: [*]align(page_size) u8, | 266 | allocated_count: SlotIndex, |
| 273 | alloc_cursor: SlotIndex, | 267 | freed_count: SlotIndex, |
| 274 | used_count: SlotIndex, | 268 | prev: ?*BucketHeader, |
| | 269 | |
| | 270 | fn fromPage(page_addr: usize, slot_count: usize) *BucketHeader { |
| | 271 | const unaligned = page_addr + page_size - bucketSize(slot_count); |
| | 272 | return @ptrFromInt(unaligned & ~(@as(usize, @alignOf(BucketHeader)) - 1)); |
| | 273 | } |
| 275 | | 274 | |
| | 275 | // TODO use usize instead of u8 |
| 276 | fn usedBits(bucket: *BucketHeader, index: usize) *u8 { | 276 | fn usedBits(bucket: *BucketHeader, index: usize) *u8 { |
| 277 | return @as(*u8, @ptrFromInt(@intFromPtr(bucket) + @sizeOf(BucketHeader) + index)); | 277 | // TODO avoid ptr to int |
| | 278 | return @ptrFromInt(@intFromPtr(bucket) + @sizeOf(BucketHeader) + index); |
| 278 | } | 279 | } |
| 279 | | 280 | |
| 280 | fn requestedSizes(bucket: *BucketHeader, size_class: usize) []LargestSizeClassInt { | 281 | fn requestedSizes(bucket: *BucketHeader, slot_count: usize) []LargestSizeClassInt { |
| 281 | if (!config.safety) @compileError("requested size is only stored when safety is enabled"); | 282 | if (!config.safety) @compileError("requested size is only stored when safety is enabled"); |
| 282 | const start_ptr = @as([*]u8, @ptrCast(bucket)) + bucketRequestedSizesStart(size_class); | 283 | const start_ptr = @as([*]u8, @ptrCast(bucket)) + bucketRequestedSizesStart(slot_count); |
| 283 | const sizes = @as([*]LargestSizeClassInt, @ptrCast(@alignCast(start_ptr))); | 284 | const sizes = @as([*]LargestSizeClassInt, @ptrCast(@alignCast(start_ptr))); |
| 284 | const slot_count = @divExact(page_size, size_class); | | |
| 285 | return sizes[0..slot_count]; | 285 | return sizes[0..slot_count]; |
| 286 | } | 286 | } |
| 287 | | 287 | |
| 288 | fn log2PtrAligns(bucket: *BucketHeader, size_class: usize) []mem.Alignment { | 288 | fn log2PtrAligns(bucket: *BucketHeader, slot_count: usize) []mem.Alignment { |
| 289 | if (!config.safety) @compileError("requested size is only stored when safety is enabled"); | 289 | if (!config.safety) @compileError("requested size is only stored when safety is enabled"); |
| 290 | const aligns_ptr = @as([*]u8, @ptrCast(bucket)) + bucketAlignsStart(size_class); | 290 | const aligns_ptr = @as([*]u8, @ptrCast(bucket)) + bucketAlignsStart(slot_count); |
| 291 | const slot_count = @divExact(page_size, size_class); | | |
| 292 | return @ptrCast(aligns_ptr[0..slot_count]); | 291 | return @ptrCast(aligns_ptr[0..slot_count]); |
| 293 | } | 292 | } |
| 294 | | 293 | |
| 295 | fn stackTracePtr( | 294 | fn stackTracePtr( |
| 296 | bucket: *BucketHeader, | 295 | bucket: *BucketHeader, |
| 297 | size_class: usize, | 296 | slot_count: usize, |
| 298 | slot_index: SlotIndex, | 297 | slot_index: SlotIndex, |
| 299 | trace_kind: TraceKind, | 298 | trace_kind: TraceKind, |
| 300 | ) *[stack_n]usize { | 299 | ) *[stack_n]usize { |
| 301 | const start_ptr = @as([*]u8, @ptrCast(bucket)) + bucketStackFramesStart(size_class); | 300 | const start_ptr = @as([*]u8, @ptrCast(bucket)) + bucketStackFramesStart(slot_count); |
| 302 | const addr = start_ptr + one_trace_size * traces_per_slot * slot_index + | 301 | const addr = start_ptr + one_trace_size * traces_per_slot * slot_index + |
| 303 | @intFromEnum(trace_kind) * @as(usize, one_trace_size); | 302 | @intFromEnum(trace_kind) * @as(usize, one_trace_size); |
| 304 | return @ptrCast(@alignCast(addr)); | 303 | return @ptrCast(@alignCast(addr)); |
| ... | @@ -307,21 +306,15 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -307,21 +306,15 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 307 | fn captureStackTrace( | 306 | fn captureStackTrace( |
| 308 | bucket: *BucketHeader, | 307 | bucket: *BucketHeader, |
| 309 | ret_addr: usize, | 308 | ret_addr: usize, |
| 310 | size_class: usize, | 309 | slot_count: usize, |
| 311 | slot_index: SlotIndex, | 310 | slot_index: SlotIndex, |
| 312 | trace_kind: TraceKind, | 311 | trace_kind: TraceKind, |
| 313 | ) void { | 312 | ) void { |
| 314 | // Initialize them to 0. When determining the count we must look | 313 | // Initialize them to 0. When determining the count we must look |
| 315 | // for non zero addresses. | 314 | // for non zero addresses. |
| 316 | const stack_addresses = bucket.stackTracePtr(size_class, slot_index, trace_kind); | 315 | const stack_addresses = bucket.stackTracePtr(slot_count, slot_index, trace_kind); |
| 317 | collectStackTrace(ret_addr, stack_addresses); | 316 | collectStackTrace(ret_addr, stack_addresses); |
| 318 | } | 317 | } |
| 319 | | | |
| 320 | /// Only valid for buckets within `empty_buckets`, and relies on the `alloc_cursor` | | |
| 321 | /// of empty buckets being set to `slot_count` when they are added to `empty_buckets` | | |
| 322 | fn emptyBucketSizeClass(bucket: *BucketHeader) usize { | | |
| 323 | return @divExact(page_size, bucket.alloc_cursor); | | |
| 324 | } | | |
| 325 | }; | 318 | }; |
| 326 | | 319 | |
| 327 | pub fn allocator(self: *Self) Allocator { | 320 | pub fn allocator(self: *Self) Allocator { |
| ... | @@ -338,64 +331,76 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -338,64 +331,76 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 338 | | 331 | |
| 339 | fn bucketStackTrace( | 332 | fn bucketStackTrace( |
| 340 | bucket: *BucketHeader, | 333 | bucket: *BucketHeader, |
| 341 | size_class: usize, | 334 | slot_count: usize, |
| 342 | slot_index: SlotIndex, | 335 | slot_index: SlotIndex, |
| 343 | trace_kind: TraceKind, | 336 | trace_kind: TraceKind, |
| 344 | ) StackTrace { | 337 | ) StackTrace { |
| 345 | const stack_addresses = bucket.stackTracePtr(size_class, slot_index, trace_kind); | 338 | const stack_addresses = bucket.stackTracePtr(slot_count, slot_index, trace_kind); |
| 346 | var len: usize = 0; | 339 | var len: usize = 0; |
| 347 | while (len < stack_n and stack_addresses[len] != 0) { | 340 | while (len < stack_n and stack_addresses[len] != 0) { |
| 348 | len += 1; | 341 | len += 1; |
| 349 | } | 342 | } |
| 350 | return StackTrace{ | 343 | return .{ |
| 351 | .instruction_addresses = stack_addresses, | 344 | .instruction_addresses = stack_addresses, |
| 352 | .index = len, | 345 | .index = len, |
| 353 | }; | 346 | }; |
| 354 | } | 347 | } |
| 355 | | 348 | |
| 356 | fn bucketRequestedSizesStart(size_class: usize) usize { | 349 | fn bucketRequestedSizesStart(slot_count: usize) usize { |
| 357 | if (!config.safety) @compileError("requested sizes are not stored unless safety is enabled"); | 350 | if (!config.safety) @compileError("requested sizes are not stored unless safety is enabled"); |
| 358 | return mem.alignForward( | 351 | return mem.alignForward( |
| 359 | usize, | 352 | usize, |
| 360 | @sizeOf(BucketHeader) + usedBitsCount(size_class), | 353 | @sizeOf(BucketHeader) + usedBitsCount(slot_count), |
| 361 | @alignOf(LargestSizeClassInt), | 354 | @alignOf(LargestSizeClassInt), |
| 362 | ); | 355 | ); |
| 363 | } | 356 | } |
| 364 | | 357 | |
| 365 | fn bucketAlignsStart(size_class: usize) usize { | 358 | fn bucketAlignsStart(slot_count: usize) usize { |
| 366 | if (!config.safety) @compileError("requested sizes are not stored unless safety is enabled"); | 359 | if (!config.safety) @compileError("requested sizes are not stored unless safety is enabled"); |
| 367 | const slot_count = @divExact(page_size, size_class); | 360 | return bucketRequestedSizesStart(slot_count) + (@sizeOf(LargestSizeClassInt) * slot_count); |
| 368 | return bucketRequestedSizesStart(size_class) + (@sizeOf(LargestSizeClassInt) * slot_count); | | |
| 369 | } | 361 | } |
| 370 | | 362 | |
| 371 | fn bucketStackFramesStart(size_class: usize) usize { | 363 | fn bucketStackFramesStart(slot_count: usize) usize { |
| 372 | const unaligned_start = if (config.safety) blk: { | 364 | const unaligned_start = if (config.safety) |
| 373 | const slot_count = @divExact(page_size, size_class); | 365 | bucketAlignsStart(slot_count) + slot_count |
| 374 | break :blk bucketAlignsStart(size_class) + slot_count; | 366 | else |
| 375 | } else @sizeOf(BucketHeader) + usedBitsCount(size_class); | 367 | @sizeOf(BucketHeader) + usedBitsCount(slot_count); |
| 376 | return mem.alignForward( | 368 | return mem.alignForward(usize, unaligned_start, @alignOf(usize)); |
| 377 | usize, | | |
| 378 | unaligned_start, | | |
| 379 | @alignOf(usize), | | |
| 380 | ); | | |
| 381 | } | 369 | } |
| 382 | | 370 | |
| 383 | fn bucketSize(size_class: usize) usize { | 371 | fn bucketSize(slot_count: usize) usize { |
| 384 | const slot_count = @divExact(page_size, size_class); | 372 | return bucketStackFramesStart(slot_count) + one_trace_size * traces_per_slot * slot_count; |
| 385 | return bucketStackFramesStart(size_class) + one_trace_size * traces_per_slot * slot_count; | | |
| 386 | } | 373 | } |
| 387 | | 374 | |
| 388 | fn usedBitsCount(size_class: usize) usize { | 375 | /// This is executed only at compile-time to prepopulate a lookup table. |
| 389 | const slot_count = @divExact(page_size, size_class); | 376 | fn calculateSlotCount(size_class_index: usize) SlotIndex { |
| 390 | if (slot_count < 8) return 1; | 377 | const size_class = @as(usize, 1) << @as(u6, @intCast(size_class_index)); |
| 391 | return @divExact(slot_count, 8); | 378 | var lower: usize = 8; |
| | 379 | var upper: usize = (page_size - bucketSize(lower)) / size_class; |
| | 380 | while (upper > lower) { |
| | 381 | const proposed: usize = lower + (upper - lower) / 2; |
| | 382 | if (proposed == lower) return lower; |
| | 383 | const slots_end = proposed * size_class; |
| | 384 | const header_begin = mem.alignForward(usize, slots_end, @alignOf(BucketHeader)); |
| | 385 | const bucket_size = bucketSize(proposed); |
| | 386 | const end = header_begin + bucket_size; |
| | 387 | if (end > page_size) { |
| | 388 | upper = proposed - 1; |
| | 389 | } else { |
| | 390 | lower = proposed; |
| | 391 | } |
| | 392 | } |
| | 393 | return lower; |
| 392 | } | 394 | } |
| 393 | | 395 | |
| 394 | fn detectLeaksInBucket( | 396 | fn usedBitsCount(slot_count: usize) usize { |
| 395 | bucket: *BucketHeader, | 397 | assert(slot_count >= 8); |
| 396 | size_class: usize, | 398 | return (slot_count + 7) / 8; |
| 397 | used_bits_count: usize, | 399 | } |
| 398 | ) bool { | 400 | |
| | 401 | fn detectLeaksInBucket(bucket: *BucketHeader, size_class_index: usize, used_bits_count: usize) bool { |
| | 402 | const size_class = @as(usize, 1) << @as(u6, @intCast(size_class_index)); |
| | 403 | const slot_count = slot_counts[size_class_index]; |
| 399 | var leaks = false; | 404 | var leaks = false; |
| 400 | var used_bits_byte: usize = 0; | 405 | var used_bits_byte: usize = 0; |
| 401 | while (used_bits_byte < used_bits_count) : (used_bits_byte += 1) { | 406 | while (used_bits_byte < used_bits_count) : (used_bits_byte += 1) { |
| ... | @@ -405,12 +410,11 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -405,12 +410,11 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 405 | while (true) : (bit_index += 1) { | 410 | while (true) : (bit_index += 1) { |
| 406 | const is_used = @as(u1, @truncate(used_byte >> bit_index)) != 0; | 411 | const is_used = @as(u1, @truncate(used_byte >> bit_index)) != 0; |
| 407 | if (is_used) { | 412 | if (is_used) { |
| 408 | const slot_index = @as(SlotIndex, @intCast(used_bits_byte * 8 + bit_index)); | 413 | const slot_index: SlotIndex = @intCast(used_bits_byte * 8 + bit_index); |
| 409 | const stack_trace = bucketStackTrace(bucket, size_class, slot_index, .alloc); | 414 | const stack_trace = bucketStackTrace(bucket, slot_count, slot_index, .alloc); |
| 410 | const addr = bucket.page + slot_index * size_class; | 415 | const page_addr = @intFromPtr(bucket) & ~(page_size - 1); |
| 411 | log.err("memory address 0x{x} leaked: {}", .{ | 416 | const addr = page_addr + slot_index * size_class; |
| 412 | @intFromPtr(addr), stack_trace, | 417 | log.err("memory address 0x{x} leaked: {}", .{ addr, stack_trace }); |
| 413 | }); | | |
| 414 | leaks = true; | 418 | leaks = true; |
| 415 | } | 419 | } |
| 416 | if (bit_index == math.maxInt(u3)) | 420 | if (bit_index == math.maxInt(u3)) |
| ... | @@ -425,16 +429,16 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -425,16 +429,16 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 425 | pub fn detectLeaks(self: *Self) bool { | 429 | pub fn detectLeaks(self: *Self) bool { |
| 426 | var leaks = false; | 430 | var leaks = false; |
| 427 | | 431 | |
| 428 | for (&self.buckets, 0..) |*buckets, bucket_i| { | 432 | for (self.buckets, 0..) |init_optional_bucket, size_class_index| { |
| 429 | if (buckets.root == null) continue; | 433 | var optional_bucket = init_optional_bucket; |
| 430 | const size_class = @as(usize, 1) << @as(math.Log2Int(usize), @intCast(bucket_i)); | 434 | const slot_count = slot_counts[size_class_index]; |
| 431 | const used_bits_count = usedBitsCount(size_class); | 435 | const used_bits_count = usedBitsCount(slot_count); |
| 432 | var it = buckets.inorderIterator(); | 436 | while (optional_bucket) |bucket| { |
| 433 | while (it.next()) |node| { | 437 | leaks = detectLeaksInBucket(bucket, size_class_index, used_bits_count) or leaks; |
| 434 | const bucket = node.key; | 438 | optional_bucket = bucket.prev; |
| 435 | leaks = detectLeaksInBucket(bucket, size_class, used_bits_count) or leaks; | | |
| 436 | } | 439 | } |
| 437 | } | 440 | } |
| | 441 | |
| 438 | var it = self.large_allocations.valueIterator(); | 442 | var it = self.large_allocations.valueIterator(); |
| 439 | while (it.next()) |large_alloc| { | 443 | while (it.next()) |large_alloc| { |
| 440 | if (config.retain_metadata and large_alloc.freed) continue; | 444 | if (config.retain_metadata and large_alloc.freed) continue; |
| ... | @@ -447,46 +451,21 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -447,46 +451,21 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 447 | return leaks; | 451 | return leaks; |
| 448 | } | 452 | } |
| 449 | | 453 | |
| 450 | fn freeBucket(self: *Self, bucket: *BucketHeader, size_class: usize) void { | | |
| 451 | const bucket_size = bucketSize(size_class); | | |
| 452 | const bucket_slice = @as([*]align(@alignOf(BucketHeader)) u8, @ptrCast(bucket))[0..bucket_size]; | | |
| 453 | self.backing_allocator.free(bucket_slice); | | |
| 454 | } | | |
| 455 | | | |
| 456 | fn freeRetainedMetadata(self: *Self) void { | 454 | fn freeRetainedMetadata(self: *Self) void { |
| 457 | if (config.retain_metadata) { | 455 | comptime assert(config.retain_metadata); |
| 458 | if (config.never_unmap) { | 456 | if (config.never_unmap) { |
| 459 | // free large allocations that were intentionally leaked by never_unmap | 457 | // free large allocations that were intentionally leaked by never_unmap |
| 460 | var it = self.large_allocations.iterator(); | 458 | var it = self.large_allocations.iterator(); |
| 461 | while (it.next()) |large| { | 459 | while (it.next()) |large| { |
| 462 | if (large.value_ptr.freed) { | 460 | if (large.value_ptr.freed) { |
| 463 | self.backing_allocator.rawFree(large.value_ptr.bytes, large.value_ptr.alignment, @returnAddress()); | 461 | self.backing_allocator.rawFree(large.value_ptr.bytes, large.value_ptr.alignment, @returnAddress()); |
| 464 | } | | |
| 465 | } | | |
| 466 | } | | |
| 467 | // free retained metadata for small allocations | | |
| 468 | while (self.empty_buckets.getMin()) |node| { | | |
| 469 | // remove the node from the tree before destroying it | | |
| 470 | var entry = self.empty_buckets.getEntryForExisting(node); | | |
| 471 | entry.set(null); | | |
| 472 | | | |
| 473 | var bucket = node.key; | | |
| 474 | if (config.never_unmap) { | | |
| 475 | // free page that was intentionally leaked by never_unmap | | |
| 476 | self.backing_allocator.free(bucket.page[0..page_size]); | | |
| 477 | } | 462 | } |
| 478 | // alloc_cursor was set to slot count when bucket added to empty_buckets | | |
| 479 | self.freeBucket(bucket, bucket.emptyBucketSizeClass()); | | |
| 480 | self.bucket_node_pool.destroy(node); | | |
| 481 | } | 463 | } |
| 482 | self.empty_buckets.root = null; | | |
| 483 | } | 464 | } |
| 484 | } | 465 | } |
| 485 | | 466 | |
| 486 | pub fn flushRetainedMetadata(self: *Self) void { | 467 | pub fn flushRetainedMetadata(self: *Self) void { |
| 487 | if (!config.retain_metadata) { | 468 | comptime assert(config.retain_metadata); |
| 488 | @compileError("'flushRetainedMetadata' requires 'config.retain_metadata = true'"); | | |
| 489 | } | | |
| 490 | self.freeRetainedMetadata(); | 469 | self.freeRetainedMetadata(); |
| 491 | // also remove entries from large_allocations | 470 | // also remove entries from large_allocations |
| 492 | var it = self.large_allocations.iterator(); | 471 | var it = self.large_allocations.iterator(); |
| ... | @@ -500,13 +479,10 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -500,13 +479,10 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 500 | /// Returns `Check.leak` if there were leaks; `Check.ok` otherwise. | 479 | /// Returns `Check.leak` if there were leaks; `Check.ok` otherwise. |
| 501 | pub fn deinit(self: *Self) Check { | 480 | pub fn deinit(self: *Self) Check { |
| 502 | const leaks = if (config.safety) self.detectLeaks() else false; | 481 | const leaks = if (config.safety) self.detectLeaks() else false; |
| 503 | if (config.retain_metadata) { | 482 | if (config.retain_metadata) self.freeRetainedMetadata(); |
| 504 | self.freeRetainedMetadata(); | | |
| 505 | } | | |
| 506 | self.large_allocations.deinit(self.backing_allocator); | 483 | self.large_allocations.deinit(self.backing_allocator); |
| 507 | self.bucket_node_pool.deinit(); | | |
| 508 | self.* = undefined; | 484 | self.* = undefined; |
| 509 | return @as(Check, @enumFromInt(@intFromBool(leaks))); | 485 | return if (leaks) .leak else .ok; |
| 510 | } | 486 | } |
| 511 | | 487 | |
| 512 | fn collectStackTrace(first_trace_addr: usize, addresses: *[stack_n]usize) void { | 488 | fn collectStackTrace(first_trace_addr: usize, addresses: *[stack_n]usize) void { |
| ... | @@ -520,8 +496,8 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -520,8 +496,8 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 520 | } | 496 | } |
| 521 | | 497 | |
| 522 | fn reportDoubleFree(ret_addr: usize, alloc_stack_trace: StackTrace, free_stack_trace: StackTrace) void { | 498 | fn reportDoubleFree(ret_addr: usize, alloc_stack_trace: StackTrace, free_stack_trace: StackTrace) void { |
| 523 | var addresses: [stack_n]usize = [1]usize{0} ** stack_n; | 499 | var addresses: [stack_n]usize = @splat(0); |
| 524 | var second_free_stack_trace = StackTrace{ | 500 | var second_free_stack_trace: StackTrace = .{ |
| 525 | .instruction_addresses = &addresses, | 501 | .instruction_addresses = &addresses, |
| 526 | .index = 0, | 502 | .index = 0, |
| 527 | }; | 503 | }; |
| ... | @@ -531,58 +507,6 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -531,58 +507,6 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 531 | }); | 507 | }); |
| 532 | } | 508 | } |
| 533 | | 509 | |
| 534 | const Slot = struct { | | |
| 535 | bucket: *BucketHeader, | | |
| 536 | slot_index: usize, | | |
| 537 | ptr: [*]u8, | | |
| 538 | }; | | |
| 539 | | | |
| 540 | fn allocSlot(self: *Self, size_class: usize, trace_addr: usize) Error!Slot { | | |
| 541 | const bucket_index = math.log2(size_class); | | |
| 542 | var buckets = &self.buckets[bucket_index]; | | |
| 543 | const slot_count = @divExact(page_size, size_class); | | |
| 544 | if (self.cur_buckets[bucket_index] == null or self.cur_buckets[bucket_index].?.alloc_cursor == slot_count) { | | |
| 545 | const new_bucket = try self.createBucket(size_class); | | |
| 546 | errdefer self.freeBucket(new_bucket, size_class); | | |
| 547 | const node = try self.bucket_node_pool.create(); | | |
| 548 | node.key = new_bucket; | | |
| 549 | var entry = buckets.getEntryFor(new_bucket); | | |
| 550 | std.debug.assert(entry.node == null); | | |
| 551 | entry.set(node); | | |
| 552 | self.cur_buckets[bucket_index] = node.key; | | |
| 553 | } | | |
| 554 | const bucket = self.cur_buckets[bucket_index].?; | | |
| 555 | | | |
| 556 | const slot_index = bucket.alloc_cursor; | | |
| 557 | bucket.alloc_cursor += 1; | | |
| 558 | | | |
| 559 | const used_bits_byte = bucket.usedBits(slot_index / 8); | | |
| 560 | const used_bit_index: u3 = @as(u3, @intCast(slot_index % 8)); // TODO cast should be unnecessary | | |
| 561 | used_bits_byte.* |= (@as(u8, 1) << used_bit_index); | | |
| 562 | bucket.used_count += 1; | | |
| 563 | bucket.captureStackTrace(trace_addr, size_class, slot_index, .alloc); | | |
| 564 | return .{ | | |
| 565 | .bucket = bucket, | | |
| 566 | .slot_index = slot_index, | | |
| 567 | .ptr = bucket.page + slot_index * size_class, | | |
| 568 | }; | | |
| 569 | } | | |
| 570 | | | |
| 571 | fn searchBucket( | | |
| 572 | buckets: *Buckets, | | |
| 573 | addr: usize, | | |
| 574 | current_bucket: ?*BucketHeader, | | |
| 575 | ) ?*BucketHeader { | | |
| 576 | const search_page: [*]align(page_size) u8 = @ptrFromInt(mem.alignBackward(usize, addr, page_size)); | | |
| 577 | if (current_bucket != null and current_bucket.?.page == search_page) { | | |
| 578 | return current_bucket; | | |
| 579 | } | | |
| 580 | var search_header: BucketHeader = undefined; | | |
| 581 | search_header.page = search_page; | | |
| 582 | const entry = buckets.getEntryFor(&search_header); | | |
| 583 | return if (entry.node) |node| node.key else null; | | |
| 584 | } | | |
| 585 | | | |
| 586 | /// This function assumes the object is in the large object storage regardless | 510 | /// This function assumes the object is in the large object storage regardless |
| 587 | /// of the parameters. | 511 | /// of the parameters. |
| 588 | fn resizeLarge( | 512 | fn resizeLarge( |
| ... | @@ -752,201 +676,177 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -752,201 +676,177 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 752 | } | 676 | } |
| 753 | } | 677 | } |
| 754 | | 678 | |
| 755 | pub fn setRequestedMemoryLimit(self: *Self, limit: usize) void { | 679 | fn alloc(context: *anyopaque, len: usize, alignment: mem.Alignment, ret_addr: usize) ?[*]u8 { |
| 756 | self.requested_memory_limit = limit; | | |
| 757 | } | | |
| 758 | | | |
| 759 | fn resize( | | |
| 760 | context: *anyopaque, | | |
| 761 | memory: []u8, | | |
| 762 | alignment: mem.Alignment, | | |
| 763 | new_len: usize, | | |
| 764 | return_address: usize, | | |
| 765 | ) bool { | | |
| 766 | return realloc(context, memory, alignment, new_len, return_address, false) != null; | | |
| 767 | } | | |
| 768 | | | |
| 769 | fn remap( | | |
| 770 | context: *anyopaque, | | |
| 771 | memory: []u8, | | |
| 772 | alignment: mem.Alignment, | | |
| 773 | new_len: usize, | | |
| 774 | return_address: usize, | | |
| 775 | ) ?[*]u8 { | | |
| 776 | return realloc(context, memory, alignment, new_len, return_address, true); | | |
| 777 | } | | |
| 778 | | | |
| 779 | fn realloc( | | |
| 780 | context: *anyopaque, | | |
| 781 | old_mem: []u8, | | |
| 782 | alignment: mem.Alignment, | | |
| 783 | new_len: usize, | | |
| 784 | ret_addr: usize, | | |
| 785 | may_move: bool, | | |
| 786 | ) ?[*]u8 { | | |
| 787 | const self: *Self = @ptrCast(@alignCast(context)); | 680 | const self: *Self = @ptrCast(@alignCast(context)); |
| 788 | self.mutex.lock(); | 681 | self.mutex.lock(); |
| 789 | defer self.mutex.unlock(); | 682 | defer self.mutex.unlock(); |
| 790 | | 683 | |
| 791 | assert(old_mem.len != 0); | 684 | if (config.enable_memory_limit) { |
| 792 | | 685 | const new_req_bytes = self.total_requested_bytes + len; |
| 793 | const aligned_size = @max(old_mem.len, alignment.toByteUnits()); | 686 | if (new_req_bytes > self.requested_memory_limit) return null; |
| 794 | if (aligned_size > largest_bucket_object_size) { | 687 | self.total_requested_bytes = new_req_bytes; |
| 795 | return self.resizeLarge(old_mem, alignment, new_len, ret_addr, may_move); | | |
| 796 | } | 688 | } |
| 797 | const size_class_hint = math.ceilPowerOfTwoAssert(usize, aligned_size); | | |
| 798 | | 689 | |
| 799 | var bucket_index = math.log2(size_class_hint); | 690 | const size_class_index: usize = @max(@bitSizeOf(usize) - @clz(len - 1), @intFromEnum(alignment)); |
| 800 | var size_class: usize = size_class_hint; | 691 | if (size_class_index >= self.buckets.len) { |
| 801 | const bucket = while (bucket_index < small_bucket_count) : (bucket_index += 1) { | 692 | @branchHint(.unlikely); |
| 802 | if (searchBucket(&self.buckets[bucket_index], @intFromPtr(old_mem.ptr), self.cur_buckets[bucket_index])) |bucket| { | 693 | self.large_allocations.ensureUnusedCapacity(self.backing_allocator, 1) catch return null; |
| 803 | break bucket; | 694 | const ptr = self.backing_allocator.rawAlloc(len, alignment, ret_addr) orelse return null; |
| | 695 | const slice = ptr[0..len]; |
| | 696 | |
| | 697 | const gop = self.large_allocations.getOrPutAssumeCapacity(@intFromPtr(slice.ptr)); |
| | 698 | if (config.retain_metadata and !config.never_unmap) { |
| | 699 | // Backing allocator may be reusing memory that we're retaining metadata for |
| | 700 | assert(!gop.found_existing or gop.value_ptr.freed); |
| | 701 | } else { |
| | 702 | assert(!gop.found_existing); // This would mean the kernel double-mapped pages. |
| 804 | } | 703 | } |
| 805 | size_class *= 2; | 704 | gop.value_ptr.bytes = slice; |
| 806 | } else blk: { | 705 | if (config.enable_memory_limit) |
| | 706 | gop.value_ptr.requested_size = len; |
| | 707 | gop.value_ptr.captureStackTrace(ret_addr, .alloc); |
| 807 | if (config.retain_metadata) { | 708 | if (config.retain_metadata) { |
| 808 | if (!self.large_allocations.contains(@intFromPtr(old_mem.ptr))) { | 709 | gop.value_ptr.freed = false; |
| 809 | // object not in active buckets or a large allocation, so search empty buckets | 710 | if (config.never_unmap) { |
| 810 | if (searchBucket(&self.empty_buckets, @intFromPtr(old_mem.ptr), null)) |bucket| { | 711 | gop.value_ptr.alignment = alignment; |
| 811 | size_class = bucket.emptyBucketSizeClass(); | | |
| 812 | // bucket is empty so is_used below will always be false and we exit there | | |
| 813 | break :blk bucket; | | |
| 814 | } else { | | |
| 815 | @panic("Invalid free"); | | |
| 816 | } | | |
| 817 | } | 712 | } |
| 818 | } | 713 | } |
| 819 | return self.resizeLarge(old_mem, alignment, new_len, ret_addr, may_move); | 714 | |
| 820 | }; | 715 | if (config.verbose_log) { |
| 821 | const byte_offset = @intFromPtr(old_mem.ptr) - @intFromPtr(bucket.page); | 716 | log.info("large alloc {d} bytes at {*}", .{ slice.len, slice.ptr }); |
| 822 | const slot_index = @as(SlotIndex, @intCast(byte_offset / size_class)); | | |
| 823 | const used_byte_index = slot_index / 8; | | |
| 824 | const used_bit_index = @as(u3, @intCast(slot_index % 8)); | | |
| 825 | const used_byte = bucket.usedBits(used_byte_index); | | |
| 826 | const is_used = @as(u1, @truncate(used_byte.* >> used_bit_index)) != 0; | | |
| 827 | if (!is_used) { | | |
| 828 | if (config.safety) { | | |
| 829 | reportDoubleFree(ret_addr, bucketStackTrace(bucket, size_class, slot_index, .alloc), bucketStackTrace(bucket, size_class, slot_index, .free)); | | |
| 830 | @panic("Unrecoverable double free"); | | |
| 831 | } else { | | |
| 832 | unreachable; | | |
| 833 | } | 717 | } |
| | 718 | return slice.ptr; |
| 834 | } | 719 | } |
| 835 | | 720 | |
| 836 | // Definitely an in-use small alloc now. | 721 | const slot_count = slot_counts[size_class_index]; |
| 837 | if (config.safety) { | 722 | |
| 838 | const requested_size = bucket.requestedSizes(size_class)[slot_index]; | 723 | if (self.buckets[size_class_index]) |bucket| { |
| 839 | if (requested_size == 0) @panic("Invalid free"); | 724 | @branchHint(.likely); |
| 840 | const slot_alignment = bucket.log2PtrAligns(size_class)[slot_index]; | 725 | const slot_index = bucket.allocated_count; |
| 841 | if (old_mem.len != requested_size or alignment != slot_alignment) { | 726 | if (slot_index < slot_count) { |
| 842 | var addresses: [stack_n]usize = [1]usize{0} ** stack_n; | 727 | @branchHint(.likely); |
| 843 | var free_stack_trace = StackTrace{ | 728 | bucket.allocated_count = slot_index + 1; |
| 844 | .instruction_addresses = &addresses, | 729 | const used_bits_byte = bucket.usedBits(slot_index / 8); |
| 845 | .index = 0, | 730 | const used_bit_index: u3 = @intCast(slot_index % 8); |
| 846 | }; | 731 | used_bits_byte.* |= (@as(u8, 1) << used_bit_index); |
| 847 | std.debug.captureStackTrace(ret_addr, &free_stack_trace); | 732 | const size_class = @as(usize, 1) << @as(u6, @intCast(size_class_index)); |
| 848 | if (old_mem.len != requested_size) { | 733 | if (config.stack_trace_frames > 0) { |
| 849 | log.err("Allocation size {d} bytes does not match resize size {d}. Allocation: {} Resize: {}", .{ | 734 | bucket.captureStackTrace(ret_addr, slot_count, slot_index, .alloc); |
| 850 | requested_size, | | |
| 851 | old_mem.len, | | |
| 852 | bucketStackTrace(bucket, size_class, slot_index, .alloc), | | |
| 853 | free_stack_trace, | | |
| 854 | }); | | |
| 855 | } | 735 | } |
| 856 | if (alignment != slot_alignment) { | 736 | if (config.safety) { |
| 857 | log.err("Allocation alignment {d} does not match resize alignment {d}. Allocation: {} Resize: {}", .{ | 737 | bucket.requestedSizes(slot_count)[slot_index] = @intCast(len); |
| 858 | slot_alignment.toByteUnits(), | 738 | bucket.log2PtrAligns(slot_count)[slot_index] = alignment; |
| 859 | alignment.toByteUnits(), | | |
| 860 | bucketStackTrace(bucket, size_class, slot_index, .alloc), | | |
| 861 | free_stack_trace, | | |
| 862 | }); | | |
| 863 | } | 739 | } |
| | 740 | const page_addr = @intFromPtr(bucket) & ~(page_size - 1); |
| | 741 | const addr = page_addr + slot_index * size_class; |
| | 742 | if (config.verbose_log) { |
| | 743 | log.info("small alloc {d} bytes at 0x{x}", .{ len, addr }); |
| | 744 | } |
| | 745 | return @ptrFromInt(addr); |
| 864 | } | 746 | } |
| 865 | } | 747 | } |
| 866 | const prev_req_bytes = self.total_requested_bytes; | 748 | |
| 867 | if (config.enable_memory_limit) { | 749 | const page = self.backing_allocator.rawAlloc(page_size, page_align, @returnAddress()) orelse |
| 868 | const new_req_bytes = prev_req_bytes + new_len - old_mem.len; | 750 | return null; |
| 869 | if (new_req_bytes > prev_req_bytes and new_req_bytes > self.requested_memory_limit) { | 751 | const bucket: *BucketHeader = .fromPage(@intFromPtr(page), slot_count); |
| 870 | return null; | 752 | bucket.* = .{ |
| 871 | } | 753 | .allocated_count = 1, |
| 872 | self.total_requested_bytes = new_req_bytes; | 754 | .freed_count = 0, |
| | 755 | .prev = self.buckets[size_class_index], |
| | 756 | }; |
| | 757 | self.buckets[size_class_index] = bucket; |
| | 758 | |
| | 759 | if (!config.backing_allocator_zeroes) { |
| | 760 | @memset(@as([*]u8, @as(*[1]u8, bucket.usedBits(0)))[0..usedBitsCount(slot_count)], 0); |
| | 761 | if (config.safety) @memset(bucket.requestedSizes(slot_count), 0); |
| 873 | } | 762 | } |
| 874 | | 763 | |
| 875 | const new_aligned_size = @max(new_len, alignment.toByteUnits()); | 764 | bucket.usedBits(0).* = 0b1; |
| 876 | const new_size_class = math.ceilPowerOfTwoAssert(usize, new_aligned_size); | 765 | |
| 877 | if (new_size_class <= size_class) { | 766 | if (config.stack_trace_frames > 0) { |
| 878 | if (old_mem.len > new_len) { | 767 | bucket.captureStackTrace(ret_addr, slot_count, 0, .alloc); |
| 879 | @memset(old_mem[new_len..], undefined); | | |
| 880 | } | | |
| 881 | if (config.verbose_log) { | | |
| 882 | log.info("small resize {d} bytes at {*} to {d}", .{ | | |
| 883 | old_mem.len, old_mem.ptr, new_len, | | |
| 884 | }); | | |
| 885 | } | | |
| 886 | if (config.safety) { | | |
| 887 | bucket.requestedSizes(size_class)[slot_index] = @intCast(new_len); | | |
| 888 | } | | |
| 889 | return old_mem.ptr; | | |
| 890 | } | 768 | } |
| 891 | | 769 | |
| 892 | if (config.enable_memory_limit) { | 770 | if (config.safety) { |
| 893 | self.total_requested_bytes = prev_req_bytes; | 771 | bucket.requestedSizes(slot_count)[0] = @intCast(len); |
| | 772 | bucket.log2PtrAligns(slot_count)[0] = alignment; |
| 894 | } | 773 | } |
| | 774 | |
| | 775 | if (config.verbose_log) { |
| | 776 | log.info("small alloc {d} bytes at 0x{x}", .{ len, @intFromPtr(page) }); |
| | 777 | } |
| | 778 | |
| | 779 | return page; |
| | 780 | } |
| | 781 | |
| | 782 | fn resize( |
| | 783 | context: *anyopaque, |
| | 784 | memory: []u8, |
| | 785 | alignment: mem.Alignment, |
| | 786 | new_len: usize, |
| | 787 | return_address: usize, |
| | 788 | ) bool { |
| | 789 | _ = context; |
| | 790 | _ = memory; |
| | 791 | _ = alignment; |
| | 792 | _ = new_len; |
| | 793 | _ = return_address; |
| | 794 | return false; |
| | 795 | } |
| | 796 | |
| | 797 | fn remap( |
| | 798 | context: *anyopaque, |
| | 799 | memory: []u8, |
| | 800 | alignment: mem.Alignment, |
| | 801 | new_len: usize, |
| | 802 | return_address: usize, |
| | 803 | ) ?[*]u8 { |
| | 804 | _ = context; |
| | 805 | _ = memory; |
| | 806 | _ = alignment; |
| | 807 | _ = new_len; |
| | 808 | _ = return_address; |
| 895 | return null; | 809 | return null; |
| 896 | } | 810 | } |
| 897 | | 811 | |
| 898 | fn free( | 812 | fn free( |
| 899 | ctx: *anyopaque, | 813 | context: *anyopaque, |
| 900 | old_mem: []u8, | 814 | old_memory: []u8, |
| 901 | alignment: mem.Alignment, | 815 | alignment: mem.Alignment, |
| 902 | ret_addr: usize, | 816 | return_address: usize, |
| 903 | ) void { | 817 | ) void { |
| 904 | const self: *Self = @ptrCast(@alignCast(ctx)); | 818 | const self: *Self = @ptrCast(@alignCast(context)); |
| 905 | self.mutex.lock(); | 819 | self.mutex.lock(); |
| 906 | defer self.mutex.unlock(); | 820 | defer self.mutex.unlock(); |
| 907 | | 821 | |
| 908 | assert(old_mem.len != 0); | 822 | assert(old_memory.len != 0); |
| 909 | | 823 | |
| 910 | const aligned_size = @max(old_mem.len, alignment.toByteUnits()); | 824 | const size_class_index: usize = @max(@bitSizeOf(usize) - @clz(old_memory.len - 1), @intFromEnum(alignment)); |
| 911 | if (aligned_size > largest_bucket_object_size) { | 825 | if (size_class_index >= self.buckets.len) { |
| 912 | self.freeLarge(old_mem, alignment, ret_addr); | 826 | @branchHint(.unlikely); |
| | 827 | self.freeLarge(old_memory, alignment, return_address); |
| 913 | return; | 828 | return; |
| 914 | } | 829 | } |
| 915 | const size_class_hint = math.ceilPowerOfTwoAssert(usize, aligned_size); | | |
| 916 | | 830 | |
| 917 | var bucket_index = math.log2(size_class_hint); | 831 | const slot_count = slot_counts[size_class_index]; |
| 918 | var size_class: usize = size_class_hint; | 832 | const freed_addr = @intFromPtr(old_memory.ptr); |
| 919 | const bucket = while (bucket_index < small_bucket_count) : (bucket_index += 1) { | 833 | const page_addr = freed_addr & ~(page_size - 1); |
| 920 | if (searchBucket(&self.buckets[bucket_index], @intFromPtr(old_mem.ptr), self.cur_buckets[bucket_index])) |bucket| { | 834 | const bucket: *BucketHeader = .fromPage(page_addr, slot_count); |
| 921 | break bucket; | 835 | const page_offset = freed_addr - page_addr; |
| 922 | } | 836 | const size_class = @as(usize, 1) << @as(u6, @intCast(size_class_index)); |
| 923 | size_class *= 2; | 837 | const slot_index: SlotIndex = @intCast(page_offset / size_class); |
| 924 | } else blk: { | | |
| 925 | if (config.retain_metadata) { | | |
| 926 | if (!self.large_allocations.contains(@intFromPtr(old_mem.ptr))) { | | |
| 927 | // object not in active buckets or a large allocation, so search empty buckets | | |
| 928 | if (searchBucket(&self.empty_buckets, @intFromPtr(old_mem.ptr), null)) |bucket| { | | |
| 929 | size_class = bucket.emptyBucketSizeClass(); | | |
| 930 | // bucket is empty so is_used below will always be false and we exit there | | |
| 931 | break :blk bucket; | | |
| 932 | } else { | | |
| 933 | @panic("Invalid free"); | | |
| 934 | } | | |
| 935 | } | | |
| 936 | } | | |
| 937 | self.freeLarge(old_mem, alignment, ret_addr); | | |
| 938 | return; | | |
| 939 | }; | | |
| 940 | const byte_offset = @intFromPtr(old_mem.ptr) - @intFromPtr(bucket.page); | | |
| 941 | const slot_index = @as(SlotIndex, @intCast(byte_offset / size_class)); | | |
| 942 | const used_byte_index = slot_index / 8; | 838 | const used_byte_index = slot_index / 8; |
| 943 | const used_bit_index = @as(u3, @intCast(slot_index % 8)); | 839 | const used_bit_index: u3 = @intCast(slot_index % 8); |
| 944 | const used_byte = bucket.usedBits(used_byte_index); | 840 | const used_byte = bucket.usedBits(used_byte_index); |
| 945 | const is_used = @as(u1, @truncate(used_byte.* >> used_bit_index)) != 0; | 841 | const is_used = @as(u1, @truncate(used_byte.* >> used_bit_index)) != 0; |
| 946 | if (!is_used) { | 842 | if (!is_used) { |
| 947 | if (config.safety) { | 843 | if (config.safety) { |
| 948 | reportDoubleFree(ret_addr, bucketStackTrace(bucket, size_class, slot_index, .alloc), bucketStackTrace(bucket, size_class, slot_index, .free)); | 844 | reportDoubleFree( |
| 949 | // Recoverable if this is a free. | 845 | return_address, |
| | 846 | bucketStackTrace(bucket, slot_count, slot_index, .alloc), |
| | 847 | bucketStackTrace(bucket, slot_count, slot_index, .free), |
| | 848 | ); |
| | 849 | // Recoverable since this is a free. |
| 950 | return; | 850 | return; |
| 951 | } else { | 851 | } else { |
| 952 | unreachable; | 852 | unreachable; |
| ... | @@ -955,21 +855,21 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -955,21 +855,21 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 955 | | 855 | |
| 956 | // Definitely an in-use small alloc now. | 856 | // Definitely an in-use small alloc now. |
| 957 | if (config.safety) { | 857 | if (config.safety) { |
| 958 | const requested_size = bucket.requestedSizes(size_class)[slot_index]; | 858 | const requested_size = bucket.requestedSizes(slot_count)[slot_index]; |
| 959 | if (requested_size == 0) @panic("Invalid free"); | 859 | if (requested_size == 0) @panic("Invalid free"); |
| 960 | const slot_alignment = bucket.log2PtrAligns(size_class)[slot_index]; | 860 | const slot_alignment = bucket.log2PtrAligns(slot_count)[slot_index]; |
| 961 | if (old_mem.len != requested_size or alignment != slot_alignment) { | 861 | if (old_memory.len != requested_size or alignment != slot_alignment) { |
| 962 | var addresses: [stack_n]usize = [1]usize{0} ** stack_n; | 862 | var addresses: [stack_n]usize = [1]usize{0} ** stack_n; |
| 963 | var free_stack_trace = StackTrace{ | 863 | var free_stack_trace: StackTrace = .{ |
| 964 | .instruction_addresses = &addresses, | 864 | .instruction_addresses = &addresses, |
| 965 | .index = 0, | 865 | .index = 0, |
| 966 | }; | 866 | }; |
| 967 | std.debug.captureStackTrace(ret_addr, &free_stack_trace); | 867 | std.debug.captureStackTrace(return_address, &free_stack_trace); |
| 968 | if (old_mem.len != requested_size) { | 868 | if (old_memory.len != requested_size) { |
| 969 | log.err("Allocation size {d} bytes does not match free size {d}. Allocation: {} Free: {}", .{ | 869 | log.err("Allocation size {d} bytes does not match free size {d}. Allocation: {} Free: {}", .{ |
| 970 | requested_size, | 870 | requested_size, |
| 971 | old_mem.len, | 871 | old_memory.len, |
| 972 | bucketStackTrace(bucket, size_class, slot_index, .alloc), | 872 | bucketStackTrace(bucket, slot_count, slot_index, .alloc), |
| 973 | free_stack_trace, | 873 | free_stack_trace, |
| 974 | }); | 874 | }); |
| 975 | } | 875 | } |
| ... | @@ -977,7 +877,7 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -977,7 +877,7 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 977 | log.err("Allocation alignment {d} does not match free alignment {d}. Allocation: {} Free: {}", .{ | 877 | log.err("Allocation alignment {d} does not match free alignment {d}. Allocation: {} Free: {}", .{ |
| 978 | slot_alignment.toByteUnits(), | 878 | slot_alignment.toByteUnits(), |
| 979 | alignment.toByteUnits(), | 879 | alignment.toByteUnits(), |
| 980 | bucketStackTrace(bucket, size_class, slot_index, .alloc), | 880 | bucketStackTrace(bucket, slot_count, slot_index, .alloc), |
| 981 | free_stack_trace, | 881 | free_stack_trace, |
| 982 | }); | 882 | }); |
| 983 | } | 883 | } |
| ... | @@ -985,142 +885,29 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { | ... | @@ -985,142 +885,29 @@ pub fn GeneralPurposeAllocator(comptime config: Config) type { |
| 985 | } | 885 | } |
| 986 | | 886 | |
| 987 | if (config.enable_memory_limit) { | 887 | if (config.enable_memory_limit) { |
| 988 | self.total_requested_bytes -= old_mem.len; | 888 | self.total_requested_bytes -= old_memory.len; |
| 989 | } | 889 | } |
| 990 | | 890 | |
| 991 | // Capture stack trace to be the "first free", in case a double free happens. | 891 | // Capture stack trace to be the "first free", in case a double free happens. |
| 992 | bucket.captureStackTrace(ret_addr, size_class, slot_index, .free); | 892 | bucket.captureStackTrace(return_address, slot_count, slot_index, .free); |
| 993 | | 893 | |
| 994 | used_byte.* &= ~(@as(u8, 1) << used_bit_index); | 894 | used_byte.* &= ~(@as(u8, 1) << used_bit_index); |
| 995 | bucket.used_count -= 1; | | |
| 996 | if (config.safety) { | 895 | if (config.safety) { |
| 997 | bucket.requestedSizes(size_class)[slot_index] = 0; | 896 | bucket.requestedSizes(slot_count)[slot_index] = 0; |
| 998 | } | 897 | } |
| 999 | if (bucket.used_count == 0) { | 898 | bucket.freed_count += 1; |
| 1000 | var entry = self.buckets[bucket_index].getEntryFor(bucket); | 899 | if (bucket.freed_count == bucket.allocated_count) { |
| 1001 | // save the node for destruction/insertion into in empty_buckets | 900 | if (self.buckets[size_class_index] == bucket) { |
| 1002 | const node = entry.node.?; | 901 | self.buckets[size_class_index] = null; |
| 1003 | entry.set(null); | | |
| 1004 | if (self.cur_buckets[bucket_index] == bucket) { | | |
| 1005 | self.cur_buckets[bucket_index] = null; | | |
| 1006 | } | 902 | } |
| 1007 | if (!config.never_unmap) { | 903 | if (!config.never_unmap) { |
| 1008 | self.backing_allocator.free(bucket.page[0..page_size]); | 904 | const page: [*]align(page_size) u8 = @ptrFromInt(page_addr); |
| 1009 | } | 905 | self.backing_allocator.rawFree(page[0..page_size], page_align, @returnAddress()); |
| 1010 | if (!config.retain_metadata) { | | |
| 1011 | self.freeBucket(bucket, size_class); | | |
| 1012 | self.bucket_node_pool.destroy(node); | | |
| 1013 | } else { | | |
| 1014 | // move alloc_cursor to end so we can tell size_class later | | |
| 1015 | const slot_count = @divExact(page_size, size_class); | | |
| 1016 | bucket.alloc_cursor = @as(SlotIndex, @truncate(slot_count)); | | |
| 1017 | var empty_entry = self.empty_buckets.getEntryFor(node.key); | | |
| 1018 | empty_entry.set(node); | | |
| 1019 | } | | |
| 1020 | } else { | | |
| 1021 | @memset(old_mem, undefined); | | |
| 1022 | } | | |
| 1023 | if (config.verbose_log) { | | |
| 1024 | log.info("small free {d} bytes at {*}", .{ old_mem.len, old_mem.ptr }); | | |
| 1025 | } | | |
| 1026 | } | | |
| 1027 | | | |
| 1028 | // Returns true if an allocation of `size` bytes is within the specified | | |
| 1029 | // limits if enable_memory_limit is true | | |
| 1030 | fn isAllocationAllowed(self: *Self, size: usize) bool { | | |
| 1031 | if (config.enable_memory_limit) { | | |
| 1032 | const new_req_bytes = self.total_requested_bytes + size; | | |
| 1033 | if (new_req_bytes > self.requested_memory_limit) | | |
| 1034 | return false; | | |
| 1035 | self.total_requested_bytes = new_req_bytes; | | |
| 1036 | } | | |
| 1037 | | | |
| 1038 | return true; | | |
| 1039 | } | | |
| 1040 | | | |
| 1041 | fn alloc(ctx: *anyopaque, len: usize, alignment: mem.Alignment, ret_addr: usize) ?[*]u8 { | | |
| 1042 | const self: *Self = @ptrCast(@alignCast(ctx)); | | |
| 1043 | self.mutex.lock(); | | |
| 1044 | defer self.mutex.unlock(); | | |
| 1045 | if (!self.isAllocationAllowed(len)) return null; | | |
| 1046 | return allocInner(self, len, alignment, ret_addr) catch return null; | | |
| 1047 | } | | |
| 1048 | | | |
| 1049 | fn allocInner( | | |
| 1050 | self: *Self, | | |
| 1051 | len: usize, | | |
| 1052 | alignment: mem.Alignment, | | |
| 1053 | ret_addr: usize, | | |
| 1054 | ) Allocator.Error![*]u8 { | | |
| 1055 | const new_aligned_size = @max(len, alignment.toByteUnits()); | | |
| 1056 | if (new_aligned_size > largest_bucket_object_size) { | | |
| 1057 | try self.large_allocations.ensureUnusedCapacity(self.backing_allocator, 1); | | |
| 1058 | const ptr = self.backing_allocator.rawAlloc(len, alignment, ret_addr) orelse | | |
| 1059 | return error.OutOfMemory; | | |
| 1060 | const slice = ptr[0..len]; | | |
| 1061 | | | |
| 1062 | const gop = self.large_allocations.getOrPutAssumeCapacity(@intFromPtr(slice.ptr)); | | |
| 1063 | if (config.retain_metadata and !config.never_unmap) { | | |
| 1064 | // Backing allocator may be reusing memory that we're retaining metadata for | | |
| 1065 | assert(!gop.found_existing or gop.value_ptr.freed); | | |
| 1066 | } else { | | |
| 1067 | assert(!gop.found_existing); // This would mean the kernel double-mapped pages. | | |
| 1068 | } | 906 | } |
| 1069 | gop.value_ptr.bytes = slice; | | |
| 1070 | if (config.enable_memory_limit) | | |
| 1071 | gop.value_ptr.requested_size = len; | | |
| 1072 | gop.value_ptr.captureStackTrace(ret_addr, .alloc); | | |
| 1073 | if (config.retain_metadata) { | | |
| 1074 | gop.value_ptr.freed = false; | | |
| 1075 | if (config.never_unmap) { | | |
| 1076 | gop.value_ptr.alignment = alignment; | | |
| 1077 | } | | |
| 1078 | } | | |
| 1079 | | | |
| 1080 | if (config.verbose_log) { | | |
| 1081 | log.info("large alloc {d} bytes at {*}", .{ slice.len, slice.ptr }); | | |
| 1082 | } | | |
| 1083 | return slice.ptr; | | |
| 1084 | } | | |
| 1085 | | | |
| 1086 | const new_size_class = math.ceilPowerOfTwoAssert(usize, new_aligned_size); | | |
| 1087 | const slot = try self.allocSlot(new_size_class, ret_addr); | | |
| 1088 | if (config.safety) { | | |
| 1089 | slot.bucket.requestedSizes(new_size_class)[slot.slot_index] = @intCast(len); | | |
| 1090 | slot.bucket.log2PtrAligns(new_size_class)[slot.slot_index] = alignment; | | |
| 1091 | } | 907 | } |
| 1092 | if (config.verbose_log) { | 908 | if (config.verbose_log) { |
| 1093 | log.info("small alloc {d} bytes at {*}", .{ len, slot.ptr }); | 909 | log.info("small free {d} bytes at {*}", .{ old_memory.len, old_memory.ptr }); |
| 1094 | } | | |
| 1095 | return slot.ptr; | | |
| 1096 | } | | |
| 1097 | | | |
| 1098 | fn createBucket(self: *Self, size_class: usize) Error!*BucketHeader { | | |
| 1099 | const alignment: mem.Alignment = .fromByteUnits(page_size); | | |
| 1100 | const page = self.backing_allocator.rawAlloc(page_size, alignment, @returnAddress()) orelse | | |
| 1101 | return error.OutOfMemory; | | |
| 1102 | errdefer self.backing_allocator.rawFree(page[0..page_size], alignment, @returnAddress()); | | |
| 1103 | | | |
| 1104 | const bucket_size = bucketSize(size_class); | | |
| 1105 | const header_align: mem.Alignment = .fromByteUnits(@alignOf(BucketHeader)); | | |
| 1106 | const ptr: *BucketHeader = @alignCast(@ptrCast(self.backing_allocator.rawAlloc( | | |
| 1107 | bucket_size, | | |
| 1108 | header_align, | | |
| 1109 | @returnAddress(), | | |
| 1110 | ) orelse return error.OutOfMemory)); | | |
| 1111 | ptr.* = .{ | | |
| 1112 | .page = @alignCast(page), | | |
| 1113 | .alloc_cursor = 0, | | |
| 1114 | .used_count = 0, | | |
| 1115 | }; | | |
| 1116 | if (!config.backing_allocator_zeroes) { | | |
| 1117 | @memset(@as([*]u8, @as(*[1]u8, ptr.usedBits(0)))[0..usedBitsCount(size_class)], 0); | | |
| 1118 | if (config.safety) { | | |
| 1119 | // Set the requested sizes to zeroes | | |
| 1120 | @memset(mem.sliceAsBytes(ptr.requestedSizes(size_class)), 0); | | |
| 1121 | } | | |
| 1122 | } | 910 | } |
| 1123 | return ptr; | | |
| 1124 | } | 911 | } |
| 1125 | }; | 912 | }; |
| 1126 | } | 913 | } |
| ... | @@ -1460,7 +1247,7 @@ test "setting a memory cap" { | ... | @@ -1460,7 +1247,7 @@ test "setting a memory cap" { |
| 1460 | defer std.testing.expect(gpa.deinit() == .ok) catch @panic("leak"); | 1247 | defer std.testing.expect(gpa.deinit() == .ok) catch @panic("leak"); |
| 1461 | const allocator = gpa.allocator(); | 1248 | const allocator = gpa.allocator(); |
| 1462 | | 1249 | |
| 1463 | gpa.setRequestedMemoryLimit(1010); | 1250 | gpa.requested_memory_limit = 1010; |
| 1464 | | 1251 | |
| 1465 | const small = try allocator.create(i32); | 1252 | const small = try allocator.create(i32); |
| 1466 | try std.testing.expect(gpa.total_requested_bytes == 4); | 1253 | try std.testing.expect(gpa.total_requested_bytes == 4); |
| ... | @@ -1481,63 +1268,6 @@ test "setting a memory cap" { | ... | @@ -1481,63 +1268,6 @@ test "setting a memory cap" { |
| 1481 | allocator.free(exact); | 1268 | allocator.free(exact); |
| 1482 | } | 1269 | } |
| 1483 | | 1270 | |
| 1484 | test "double frees" { | | |
| 1485 | // use a GPA to back a GPA to check for leaks of the latter's metadata | | |
| 1486 | var backing_gpa = GeneralPurposeAllocator(.{ .safety = true }){}; | | |
| 1487 | defer std.testing.expect(backing_gpa.deinit() == .ok) catch @panic("leak"); | | |
| 1488 | | | |
| 1489 | const GPA = GeneralPurposeAllocator(.{ .safety = true, .never_unmap = true, .retain_metadata = true }); | | |
| 1490 | var gpa = GPA{ .backing_allocator = backing_gpa.allocator() }; | | |
| 1491 | defer std.testing.expect(gpa.deinit() == .ok) catch @panic("leak"); | | |
| 1492 | const allocator = gpa.allocator(); | | |
| 1493 | | | |
| 1494 | // detect a small allocation double free, even though bucket is emptied | | |
| 1495 | const index: usize = 6; | | |
| 1496 | const size_class: usize = @as(usize, 1) << 6; | | |
| 1497 | const small = try allocator.alloc(u8, size_class); | | |
| 1498 | try std.testing.expect(GPA.searchBucket(&gpa.buckets[index], @intFromPtr(small.ptr), gpa.cur_buckets[index]) != null); | | |
| 1499 | allocator.free(small); | | |
| 1500 | try std.testing.expect(GPA.searchBucket(&gpa.buckets[index], @intFromPtr(small.ptr), gpa.cur_buckets[index]) == null); | | |
| 1501 | try std.testing.expect(GPA.searchBucket(&gpa.empty_buckets, @intFromPtr(small.ptr), null) != null); | | |
| 1502 | | | |
| 1503 | // detect a large allocation double free | | |
| 1504 | const large = try allocator.alloc(u8, 2 * page_size); | | |
| 1505 | try std.testing.expect(gpa.large_allocations.contains(@intFromPtr(large.ptr))); | | |
| 1506 | try std.testing.expectEqual(gpa.large_allocations.getEntry(@intFromPtr(large.ptr)).?.value_ptr.bytes, large); | | |
| 1507 | allocator.free(large); | | |
| 1508 | try std.testing.expect(gpa.large_allocations.contains(@intFromPtr(large.ptr))); | | |
| 1509 | try std.testing.expect(gpa.large_allocations.getEntry(@intFromPtr(large.ptr)).?.value_ptr.freed); | | |
| 1510 | | | |
| 1511 | const normal_small = try allocator.alloc(u8, size_class); | | |
| 1512 | defer allocator.free(normal_small); | | |
| 1513 | const normal_large = try allocator.alloc(u8, 2 * page_size); | | |
| 1514 | defer allocator.free(normal_large); | | |
| 1515 | | | |
| 1516 | // check that flushing retained metadata doesn't disturb live allocations | | |
| 1517 | gpa.flushRetainedMetadata(); | | |
| 1518 | try std.testing.expect(gpa.empty_buckets.root == null); | | |
| 1519 | try std.testing.expect(GPA.searchBucket(&gpa.buckets[index], @intFromPtr(normal_small.ptr), gpa.cur_buckets[index]) != null); | | |
| 1520 | try std.testing.expect(gpa.large_allocations.contains(@intFromPtr(normal_large.ptr))); | | |
| 1521 | try std.testing.expect(!gpa.large_allocations.contains(@intFromPtr(large.ptr))); | | |
| 1522 | } | | |
| 1523 | | | |
| 1524 | test "empty bucket size class" { | | |
| 1525 | const GPA = GeneralPurposeAllocator(.{ .safety = true, .never_unmap = true, .retain_metadata = true }); | | |
| 1526 | var gpa = GPA{}; | | |
| 1527 | defer std.testing.expect(gpa.deinit() == .ok) catch @panic("leak"); | | |
| 1528 | const allocator = gpa.allocator(); | | |
| 1529 | | | |
| 1530 | // allocate and free to create an empty bucket | | |
| 1531 | const size_class: usize = @as(usize, 1) << 6; | | |
| 1532 | const small = try allocator.alloc(u8, size_class); | | |
| 1533 | allocator.free(small); | | |
| 1534 | | | |
| 1535 | // the metadata tracking system relies on alloc_cursor of empty buckets | | |
| 1536 | // being set to the slot count so that we can get back the size class. | | |
| 1537 | const empty_bucket = GPA.searchBucket(&gpa.empty_buckets, @intFromPtr(small.ptr), null).?; | | |
| 1538 | try std.testing.expect(empty_bucket.emptyBucketSizeClass() == size_class); | | |
| 1539 | } | | |
| 1540 | | | |
| 1541 | test "bug 9995 fix, large allocs count requested size not backing size" { | 1271 | test "bug 9995 fix, large allocs count requested size not backing size" { |
| 1542 | // with AtLeast, buffer likely to be larger than requested, especially when shrinking | 1272 | // with AtLeast, buffer likely to be larger than requested, especially when shrinking |
| 1543 | var gpa = GeneralPurposeAllocator(.{ .enable_memory_limit = true }){}; | 1273 | var gpa = GeneralPurposeAllocator(.{ .enable_memory_limit = true }){}; |
Andrew Kelley <andrew@ziglang.org> 2025-02-04 23:05:06-08:00
lib