Source file src/runtime/mpagealloc.go
1 // Copyright 2019 The Go Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style 3 // license that can be found in the LICENSE file. 4 5 // Page allocator. 6 // 7 // The page allocator manages mapped pages (defined by pageSize, NOT 8 // physPageSize) for allocation and re-use. It is embedded into mheap. 9 // 10 // Pages are managed using a bitmap that is sharded into chunks. 11 // In the bitmap, 1 means in-use, and 0 means free. The bitmap spans the 12 // process's address space. Chunks are managed in a sparse-array-style structure 13 // similar to mheap.arenas, since the bitmap may be large on some systems. 14 // 15 // The bitmap is efficiently searched by using a radix tree in combination 16 // with fast bit-wise intrinsics. Allocation is performed using an address-ordered 17 // first-fit approach. 18 // 19 // Each entry in the radix tree is a summary that describes three properties of 20 // a particular region of the address space: the number of contiguous free pages 21 // at the start and end of the region it represents, and the maximum number of 22 // contiguous free pages found anywhere in that region. 23 // 24 // Each level of the radix tree is stored as one contiguous array, which represents 25 // a different granularity of subdivision of the processes' address space. Thus, this 26 // radix tree is actually implicit in these large arrays, as opposed to having explicit 27 // dynamically-allocated pointer-based node structures. Naturally, these arrays may be 28 // quite large for system with large address spaces, so in these cases they are mapped 29 // into memory as needed. The leaf summaries of the tree correspond to a bitmap chunk. 30 // 31 // The root level (referred to as L0 and index 0 in pageAlloc.summary) has each 32 // summary represent the largest section of address space (16 GiB on 64-bit systems), 33 // with each subsequent level representing successively smaller subsections until we 34 // reach the finest granularity at the leaves, a chunk. 35 // 36 // More specifically, each summary in each level (except for leaf summaries) 37 // represents some number of entries in the following level. For example, each 38 // summary in the root level may represent a 16 GiB region of address space, 39 // and in the next level there could be 8 corresponding entries which represent 2 40 // GiB subsections of that 16 GiB region, each of which could correspond to 8 41 // entries in the next level which each represent 256 MiB regions, and so on. 42 // 43 // Thus, this design only scales to heaps so large, but can always be extended to 44 // larger heaps by simply adding levels to the radix tree, which mostly costs 45 // additional virtual address space. The choice of managing large arrays also means 46 // that a large amount of virtual address space may be reserved by the runtime. 47 48 package runtime 49 50 import ( 51 "internal/runtime/atomic" 52 "unsafe" 53 ) 54 55 const ( 56 // The size of a bitmap chunk, i.e. the amount of bits (that is, pages) to consider 57 // in the bitmap at once. 58 pallocChunkPages = 1 << logPallocChunkPages 59 pallocChunkBytes = pallocChunkPages * pageSize 60 logPallocChunkPages = 9 61 logPallocChunkBytes = logPallocChunkPages + pageShift 62 63 // The number of radix bits for each level. 64 // 65 // The value of 3 is chosen such that the block of summaries we need to scan at 66 // each level fits in 64 bytes (2^3 summaries * 8 bytes per summary), which is 67 // close to the L1 cache line width on many systems. Also, a value of 3 fits 4 tree 68 // levels perfectly into the 21-bit pallocBits summary field at the root level. 69 // 70 // The following equation explains how each of the constants relate: 71 // summaryL0Bits + (summaryLevels-1)*summaryLevelBits + logPallocChunkBytes = heapAddrBits 72 // 73 // summaryLevels is an architecture-dependent value defined in mpagealloc_*.go. 74 summaryLevelBits = 3 75 summaryL0Bits = heapAddrBits - logPallocChunkBytes - (summaryLevels-1)*summaryLevelBits 76 77 // pallocChunksL2Bits is the number of bits of the chunk index number 78 // covered by the second level of the chunks map. 79 // 80 // See (*pageAlloc).chunks for more details. Update the documentation 81 // there should this change. 82 pallocChunksL2Bits = heapAddrBits - logPallocChunkBytes - pallocChunksL1Bits 83 pallocChunksL1Shift = pallocChunksL2Bits 84 ) 85 86 // maxSearchAddr returns the maximum searchAddr value, which indicates 87 // that the heap has no free space. 88 // 89 // This function exists just to make it clear that this is the maximum address 90 // for the page allocator's search space. See maxOffAddr for details. 91 // 92 // It's a function (rather than a variable) because it needs to be 93 // usable before package runtime's dynamic initialization is complete. 94 // See #51913 for details. 95 func maxSearchAddr() offAddr { return maxOffAddr } 96 97 // Global chunk index. 98 // 99 // Represents an index into the leaf level of the radix tree. 100 // Similar to arenaIndex, except instead of arenas, it divides the address 101 // space into chunks. 102 type chunkIdx uint 103 104 // chunkIndex returns the global index of the palloc chunk containing the 105 // pointer p. 106 func chunkIndex(p uintptr) chunkIdx { 107 return chunkIdx((p - arenaBaseOffset) / pallocChunkBytes) 108 } 109 110 // chunkBase returns the base address of the palloc chunk at index ci. 111 func chunkBase(ci chunkIdx) uintptr { 112 return uintptr(ci)*pallocChunkBytes + arenaBaseOffset 113 } 114 115 // chunkPageIndex computes the index of the page that contains p, 116 // relative to the chunk which contains p. 117 func chunkPageIndex(p uintptr) uint { 118 return uint(p % pallocChunkBytes / pageSize) 119 } 120 121 // l1 returns the index into the first level of (*pageAlloc).chunks. 122 func (i chunkIdx) l1() uint { 123 if pallocChunksL1Bits == 0 { 124 // Let the compiler optimize this away if there's no 125 // L1 map. 126 return 0 127 } else { 128 return uint(i) >> pallocChunksL1Shift 129 } 130 } 131 132 // l2 returns the index into the second level of (*pageAlloc).chunks. 133 func (i chunkIdx) l2() uint { 134 if pallocChunksL1Bits == 0 { 135 return uint(i) 136 } else { 137 return uint(i) & (1<<pallocChunksL2Bits - 1) 138 } 139 } 140 141 // offAddrToLevelIndex converts an address in the offset address space 142 // to the index into summary[level] containing addr. 143 func offAddrToLevelIndex(level int, addr offAddr) int { 144 return int((addr.a - arenaBaseOffset) >> levelShift[level]) 145 } 146 147 // levelIndexToOffAddr converts an index into summary[level] into 148 // the corresponding address in the offset address space. 149 func levelIndexToOffAddr(level, idx int) offAddr { 150 return offAddr{(uintptr(idx) << levelShift[level]) + arenaBaseOffset} 151 } 152 153 // addrsToSummaryRange converts base and limit pointers into a range 154 // of entries for the given summary level. 155 // 156 // The returned range is inclusive on the lower bound and exclusive on 157 // the upper bound. 158 func addrsToSummaryRange(level int, base, limit uintptr) (lo int, hi int) { 159 // This is slightly more nuanced than just a shift for the exclusive 160 // upper-bound. Note that the exclusive upper bound may be within a 161 // summary at this level, meaning if we just do the obvious computation 162 // hi will end up being an inclusive upper bound. Unfortunately, just 163 // adding 1 to that is too broad since we might be on the very edge 164 // of a summary's max page count boundary for this level 165 // (1 << levelLogPages[level]). So, make limit an inclusive upper bound 166 // then shift, then add 1, so we get an exclusive upper bound at the end. 167 lo = int((base - arenaBaseOffset) >> levelShift[level]) 168 hi = int(((limit-1)-arenaBaseOffset)>>levelShift[level]) + 1 169 return 170 } 171 172 // blockAlignSummaryRange aligns indices into the given level to that 173 // level's block width (1 << levelBits[level]). It assumes lo is inclusive 174 // and hi is exclusive, and so aligns them down and up respectively. 175 func blockAlignSummaryRange(level int, lo, hi int) (int, int) { 176 e := uintptr(1) << levelBits[level] 177 return int(alignDown(uintptr(lo), e)), int(alignUp(uintptr(hi), e)) 178 } 179 180 type pageAlloc struct { 181 // Radix tree of summaries. 182 // 183 // Each slice's cap represents the whole memory reservation. 184 // Each slice's len reflects the allocator's maximum known 185 // mapped heap address for that level. 186 // 187 // The backing store of each summary level is reserved in init 188 // and may or may not be committed in grow (small address spaces 189 // may commit all the memory in init). 190 // 191 // The purpose of keeping len <= cap is to enforce bounds checks 192 // on the top end of the slice so that instead of an unknown 193 // runtime segmentation fault, we get a much friendlier out-of-bounds 194 // error. 195 // 196 // To iterate over a summary level, use inUse to determine which ranges 197 // are currently available. Otherwise one might try to access 198 // memory which is only Reserved which may result in a hard fault. 199 // 200 // We may still get segmentation faults < len since some of that 201 // memory may not be committed yet. 202 summary [summaryLevels][]pallocSum 203 204 // chunks is a slice of bitmap chunks. 205 // 206 // The total size of chunks is quite large on most 64-bit platforms 207 // (O(GiB) or more) if flattened, so rather than making one large mapping 208 // (which has problems on some platforms, even when PROT_NONE) we use a 209 // two-level sparse array approach similar to the arena index in mheap. 210 // 211 // To find the chunk containing a memory address `a`, do: 212 // chunkOf(chunkIndex(a)) 213 // 214 // Below is a table describing the configuration for chunks for various 215 // heapAddrBits supported by the runtime. 216 // 217 // heapAddrBits | L1 Bits | L2 Bits | L2 Entry Size 218 // ------------------------------------------------ 219 // 32 | 0 | 10 | 128 KiB 220 // 33 (iOS) | 0 | 11 | 256 KiB 221 // 48 | 13 | 13 | 1 MiB 222 // 223 // There's no reason to use the L1 part of chunks on 32-bit, the 224 // address space is small so the L2 is small. For platforms with a 225 // 48-bit address space, we pick the L1 such that the L2 is 1 MiB 226 // in size, which is a good balance between low granularity without 227 // making the impact on BSS too high (note the L1 is stored directly 228 // in pageAlloc). 229 // 230 // To iterate over the bitmap, use inUse to determine which ranges 231 // are currently available. Otherwise one might iterate over unused 232 // ranges. 233 // 234 // Protected by mheapLock. 235 // 236 // TODO(mknyszek): Consider changing the definition of the bitmap 237 // such that 1 means free and 0 means in-use so that summaries and 238 // the bitmaps align better on zero-values. 239 chunks [1 << pallocChunksL1Bits]*[1 << pallocChunksL2Bits]pallocData 240 241 // The address to start an allocation search with. It must never 242 // point to any memory that is not contained in inUse, i.e. 243 // inUse.contains(searchAddr.addr()) must always be true. The one 244 // exception to this rule is that it may take on the value of 245 // maxOffAddr to indicate that the heap is exhausted. 246 // 247 // We guarantee that all valid heap addresses below this value 248 // are allocated and not worth searching. 249 searchAddr offAddr 250 251 // start and end represent the chunk indices 252 // which pageAlloc knows about. It assumes 253 // chunks in the range [start, end) are 254 // currently ready to use. 255 start, end chunkIdx 256 257 // inUse is a slice of ranges of address space which are 258 // known by the page allocator to be currently in-use (passed 259 // to grow). 260 // 261 // We care much more about having a contiguous heap in these cases 262 // and take additional measures to ensure that, so in nearly all 263 // cases this should have just 1 element. 264 // 265 // All access is protected by the mheapLock. 266 inUse addrRanges 267 268 // scav stores the scavenger state. 269 scav struct { 270 // index is an efficient index of chunks that have pages available to 271 // scavenge. 272 index scavengeIndex 273 274 // releasedBg is the amount of memory released in the background this 275 // scavenge cycle. 276 releasedBg atomic.Uintptr 277 278 // releasedEager is the amount of memory released eagerly this scavenge 279 // cycle. 280 releasedEager atomic.Uintptr 281 } 282 283 // mheap_.lock. This level of indirection makes it possible 284 // to test pageAlloc independently of the runtime allocator. 285 mheapLock *mutex 286 287 // sysStat is the runtime memstat to update when new system 288 // memory is committed by the pageAlloc for allocation metadata. 289 sysStat *sysMemStat 290 291 // summaryMappedReady is the number of bytes mapped in the Ready state 292 // in the summary structure. Used only for testing currently. 293 // 294 // Protected by mheapLock. 295 summaryMappedReady uintptr 296 297 // chunkHugePages indicates whether page bitmap chunks should be backed 298 // by huge pages. 299 chunkHugePages bool 300 301 // Whether or not this struct is being used in tests. 302 test bool 303 } 304 305 func (p *pageAlloc) init(mheapLock *mutex, sysStat *sysMemStat, test bool) { 306 if levelLogPages[0] > logMaxPackedValue { 307 // We can't represent 1<<levelLogPages[0] pages, the maximum number 308 // of pages we need to represent at the root level, in a summary, which 309 // is a big problem. Throw. 310 print("runtime: root level max pages = ", 1<<levelLogPages[0], "\n") 311 print("runtime: summary max pages = ", maxPackedValue, "\n") 312 throw("root level max pages doesn't fit in summary") 313 } 314 p.sysStat = sysStat 315 316 // Initialize p.inUse. 317 p.inUse.init(sysStat) 318 319 // System-dependent initialization. 320 p.sysInit(test) 321 322 // Start with the searchAddr in a state indicating there's no free memory. 323 p.searchAddr = maxSearchAddr() 324 325 // Set the mheapLock. 326 p.mheapLock = mheapLock 327 328 // Initialize the scavenge index. 329 p.summaryMappedReady += p.scav.index.init(test, sysStat) 330 331 // Set if we're in a test. 332 p.test = test 333 } 334 335 // tryChunkOf returns the bitmap data for the given chunk. 336 // 337 // Returns nil if the chunk data has not been mapped. 338 func (p *pageAlloc) tryChunkOf(ci chunkIdx) *pallocData { 339 l2 := p.chunks[ci.l1()] 340 if l2 == nil { 341 return nil 342 } 343 return &l2[ci.l2()] 344 } 345 346 // chunkOf returns the chunk at the given chunk index. 347 // 348 // The chunk index must be valid or this method may throw. 349 func (p *pageAlloc) chunkOf(ci chunkIdx) *pallocData { 350 return &p.chunks[ci.l1()][ci.l2()] 351 } 352 353 // grow sets up the metadata for the address range [base, base+size). 354 // It may allocate metadata, in which case *p.sysStat will be updated. 355 // 356 // p.mheapLock must be held. 357 func (p *pageAlloc) grow(base, size uintptr) { 358 assertLockHeld(p.mheapLock) 359 360 // Round up to chunks, since we can't deal with increments smaller 361 // than chunks. Also, sysGrow expects aligned values. 362 limit := alignUp(base+size, pallocChunkBytes) 363 base = alignDown(base, pallocChunkBytes) 364 365 // Grow the summary levels in a system-dependent manner. 366 // We just update a bunch of additional metadata here. 367 p.sysGrow(base, limit) 368 369 // Grow the scavenge index. 370 p.summaryMappedReady += p.scav.index.grow(base, limit, p.sysStat) 371 372 // Update p.start and p.end. 373 // If no growth happened yet, start == 0. This is generally 374 // safe since the zero page is unmapped. 375 firstGrowth := p.start == 0 376 start, end := chunkIndex(base), chunkIndex(limit) 377 if firstGrowth || start < p.start { 378 p.start = start 379 } 380 if end > p.end { 381 p.end = end 382 } 383 // Note that [base, limit) will never overlap with any existing 384 // range inUse because grow only ever adds never-used memory 385 // regions to the page allocator. 386 p.inUse.add(makeAddrRange(base, limit)) 387 388 // A grow operation is a lot like a free operation, so if our 389 // chunk ends up below p.searchAddr, update p.searchAddr to the 390 // new address, just like in free. 391 if b := (offAddr{base}); b.lessThan(p.searchAddr) { 392 p.searchAddr = b 393 } 394 395 // Add entries into chunks, which is sparse, if needed. Then, 396 // initialize the bitmap. 397 // 398 // Newly-grown memory is always considered scavenged. 399 // Set all the bits in the scavenged bitmaps high. 400 for c := chunkIndex(base); c < chunkIndex(limit); c++ { 401 if p.chunks[c.l1()] == nil { 402 // Create the necessary l2 entry. 403 const l2Size = unsafe.Sizeof(*p.chunks[0]) 404 r := sysAlloc(l2Size, p.sysStat) 405 if r == nil { 406 throw("pageAlloc: out of memory") 407 } 408 if !p.test { 409 // Make the chunk mapping eligible or ineligible 410 // for huge pages, depending on what our current 411 // state is. 412 if p.chunkHugePages { 413 sysHugePage(r, l2Size) 414 } else { 415 sysNoHugePage(r, l2Size) 416 } 417 } 418 // Store the new chunk block but avoid a write barrier. 419 // grow is used in call chains that disallow write barriers. 420 *(*uintptr)(unsafe.Pointer(&p.chunks[c.l1()])) = uintptr(r) 421 } 422 p.chunkOf(c).scavenged.setRange(0, pallocChunkPages) 423 } 424 425 // Update summaries accordingly. The grow acts like a free, so 426 // we need to ensure this newly-free memory is visible in the 427 // summaries. 428 p.update(base, size/pageSize, true, false) 429 } 430 431 // enableChunkHugePages enables huge pages for the chunk bitmap mappings (disabled by default). 432 // 433 // This function is idempotent. 434 // 435 // A note on latency: for sufficiently small heaps (<10s of GiB) this function will take constant 436 // time, but may take time proportional to the size of the mapped heap beyond that. 437 // 438 // The heap lock must not be held over this operation, since it will briefly acquire 439 // the heap lock. 440 // 441 // Must be called on the system stack because it acquires the heap lock. 442 // 443 //go:systemstack 444 func (p *pageAlloc) enableChunkHugePages() { 445 // Grab the heap lock to turn on huge pages for new chunks and clone the current 446 // heap address space ranges. 447 // 448 // After the lock is released, we can be sure that bitmaps for any new chunks may 449 // be backed with huge pages, and we have the address space for the rest of the 450 // chunks. At the end of this function, all chunk metadata should be backed by huge 451 // pages. 452 lock(&mheap_.lock) 453 if p.chunkHugePages { 454 unlock(&mheap_.lock) 455 return 456 } 457 p.chunkHugePages = true 458 var inUse addrRanges 459 inUse.sysStat = p.sysStat 460 p.inUse.cloneInto(&inUse) 461 unlock(&mheap_.lock) 462 463 // This might seem like a lot of work, but all these loops are for generality. 464 // 465 // For a 1 GiB contiguous heap, a 48-bit address space, 13 L1 bits, a palloc chunk size 466 // of 4 MiB, and adherence to the default set of heap address hints, this will result in 467 // exactly 1 call to sysHugePage. 468 for _, r := range p.inUse.ranges { 469 for i := chunkIndex(r.base.addr()).l1(); i < chunkIndex(r.limit.addr()-1).l1(); i++ { 470 // N.B. We can assume that p.chunks[i] is non-nil and in a mapped part of p.chunks 471 // because it's derived from inUse, which never shrinks. 472 sysHugePage(unsafe.Pointer(p.chunks[i]), unsafe.Sizeof(*p.chunks[0])) 473 } 474 } 475 } 476 477 // update updates heap metadata. It must be called each time the bitmap 478 // is updated. 479 // 480 // If contig is true, update does some optimizations assuming that there was 481 // a contiguous allocation or free between addr and addr+npages. alloc indicates 482 // whether the operation performed was an allocation or a free. 483 // 484 // p.mheapLock must be held. 485 func (p *pageAlloc) update(base, npages uintptr, contig, alloc bool) { 486 assertLockHeld(p.mheapLock) 487 488 // base, limit, start, and end are inclusive. 489 limit := base + npages*pageSize - 1 490 sc, ec := chunkIndex(base), chunkIndex(limit) 491 492 // Handle updating the lowest level first. 493 if sc == ec { 494 // Fast path: the allocation doesn't span more than one chunk, 495 // so update this one and if the summary didn't change, return. 496 x := p.summary[len(p.summary)-1][sc] 497 y := p.chunkOf(sc).summarize() 498 if x == y { 499 return 500 } 501 p.summary[len(p.summary)-1][sc] = y 502 } else if contig { 503 // Slow contiguous path: the allocation spans more than one chunk 504 // and at least one summary is guaranteed to change. 505 summary := p.summary[len(p.summary)-1] 506 507 // Update the summary for chunk sc. 508 summary[sc] = p.chunkOf(sc).summarize() 509 510 // Update the summaries for chunks in between, which are 511 // either totally allocated or freed. 512 whole := p.summary[len(p.summary)-1][sc+1 : ec] 513 if alloc { 514 clear(whole) 515 } else { 516 for i := range whole { 517 whole[i] = freeChunkSum 518 } 519 } 520 521 // Update the summary for chunk ec. 522 summary[ec] = p.chunkOf(ec).summarize() 523 } else { 524 // Slow general path: the allocation spans more than one chunk 525 // and at least one summary is guaranteed to change. 526 // 527 // We can't assume a contiguous allocation happened, so walk over 528 // every chunk in the range and manually recompute the summary. 529 summary := p.summary[len(p.summary)-1] 530 for c := sc; c <= ec; c++ { 531 summary[c] = p.chunkOf(c).summarize() 532 } 533 } 534 535 // Walk up the radix tree and update the summaries appropriately. 536 changed := true 537 for l := len(p.summary) - 2; l >= 0 && changed; l-- { 538 // Update summaries at level l from summaries at level l+1. 539 changed = false 540 541 // "Constants" for the previous level which we 542 // need to compute the summary from that level. 543 logEntriesPerBlock := levelBits[l+1] 544 logMaxPages := levelLogPages[l+1] 545 546 // lo and hi describe all the parts of the level we need to look at. 547 lo, hi := addrsToSummaryRange(l, base, limit+1) 548 549 // Iterate over each block, updating the corresponding summary in the less-granular level. 550 for i := lo; i < hi; i++ { 551 children := p.summary[l+1][i<<logEntriesPerBlock : (i+1)<<logEntriesPerBlock] 552 sum := mergeSummaries(children, logMaxPages) 553 old := p.summary[l][i] 554 if old != sum { 555 changed = true 556 p.summary[l][i] = sum 557 } 558 } 559 } 560 } 561 562 // allocRange marks the range of memory [base, base+npages*pageSize) as 563 // allocated. It also updates the summaries to reflect the newly-updated 564 // bitmap. 565 // 566 // Returns the amount of scavenged memory in bytes present in the 567 // allocated range. 568 // 569 // p.mheapLock must be held. 570 func (p *pageAlloc) allocRange(base, npages uintptr) uintptr { 571 assertLockHeld(p.mheapLock) 572 573 limit := base + npages*pageSize - 1 574 sc, ec := chunkIndex(base), chunkIndex(limit) 575 si, ei := chunkPageIndex(base), chunkPageIndex(limit) 576 577 scav := uint(0) 578 if sc == ec { 579 // The range doesn't cross any chunk boundaries. 580 chunk := p.chunkOf(sc) 581 scav += chunk.scavenged.popcntRange(si, ei+1-si) 582 chunk.allocRange(si, ei+1-si) 583 p.scav.index.alloc(sc, ei+1-si) 584 } else { 585 // The range crosses at least one chunk boundary. 586 chunk := p.chunkOf(sc) 587 scav += chunk.scavenged.popcntRange(si, pallocChunkPages-si) 588 chunk.allocRange(si, pallocChunkPages-si) 589 p.scav.index.alloc(sc, pallocChunkPages-si) 590 for c := sc + 1; c < ec; c++ { 591 chunk := p.chunkOf(c) 592 scav += chunk.scavenged.popcntRange(0, pallocChunkPages) 593 chunk.allocAll() 594 p.scav.index.alloc(c, pallocChunkPages) 595 } 596 chunk = p.chunkOf(ec) 597 scav += chunk.scavenged.popcntRange(0, ei+1) 598 chunk.allocRange(0, ei+1) 599 p.scav.index.alloc(ec, ei+1) 600 } 601 p.update(base, npages, true, true) 602 return uintptr(scav) * pageSize 603 } 604 605 // findMappedAddr returns the smallest mapped offAddr that is 606 // >= addr. That is, if addr refers to mapped memory, then it is 607 // returned. If addr is higher than any mapped region, then 608 // it returns maxOffAddr. 609 // 610 // p.mheapLock must be held. 611 func (p *pageAlloc) findMappedAddr(addr offAddr) offAddr { 612 assertLockHeld(p.mheapLock) 613 614 // If we're not in a test, validate first by checking mheap_.arenas. 615 // This is a fast path which is only safe to use outside of testing. 616 ai := arenaIndex(addr.addr()) 617 if p.test || mheap_.arenas[ai.l1()] == nil || mheap_.arenas[ai.l1()][ai.l2()] == nil { 618 vAddr, ok := p.inUse.findAddrGreaterEqual(addr.addr()) 619 if ok { 620 return offAddr{vAddr} 621 } else { 622 // The candidate search address is greater than any 623 // known address, which means we definitely have no 624 // free memory left. 625 return maxOffAddr 626 } 627 } 628 return addr 629 } 630 631 // find searches for the first (address-ordered) contiguous free region of 632 // npages in size and returns a base address for that region. 633 // 634 // It uses p.searchAddr to prune its search and assumes that no palloc chunks 635 // below chunkIndex(p.searchAddr) contain any free memory at all. 636 // 637 // find also computes and returns a candidate p.searchAddr, which may or 638 // may not prune more of the address space than p.searchAddr already does. 639 // This candidate is always a valid p.searchAddr. 640 // 641 // find represents the slow path and the full radix tree search. 642 // 643 // Returns a base address of 0 on failure, in which case the candidate 644 // searchAddr returned is invalid and must be ignored. 645 // 646 // p.mheapLock must be held. 647 func (p *pageAlloc) find(npages uintptr) (uintptr, offAddr) { 648 assertLockHeld(p.mheapLock) 649 650 // Search algorithm. 651 // 652 // This algorithm walks each level l of the radix tree from the root level 653 // to the leaf level. It iterates over at most 1 << levelBits[l] of entries 654 // in a given level in the radix tree, and uses the summary information to 655 // find either: 656 // 1) That a given subtree contains a large enough contiguous region, at 657 // which point it continues iterating on the next level, or 658 // 2) That there are enough contiguous boundary-crossing bits to satisfy 659 // the allocation, at which point it knows exactly where to start 660 // allocating from. 661 // 662 // i tracks the index into the current level l's structure for the 663 // contiguous 1 << levelBits[l] entries we're actually interested in. 664 // 665 // NOTE: Technically this search could allocate a region which crosses 666 // the arenaBaseOffset boundary, which when arenaBaseOffset != 0, is 667 // a discontinuity. However, the only way this could happen is if the 668 // page at the zero address is mapped, and this is impossible on 669 // every system we support where arenaBaseOffset != 0. So, the 670 // discontinuity is already encoded in the fact that the OS will never 671 // map the zero page for us, and this function doesn't try to handle 672 // this case in any way. 673 674 // i is the beginning of the block of entries we're searching at the 675 // current level. 676 i := 0 677 678 // firstFree is the region of address space that we are certain to 679 // find the first free page in the heap. base and bound are the inclusive 680 // bounds of this window, and both are addresses in the linearized, contiguous 681 // view of the address space (with arenaBaseOffset pre-added). At each level, 682 // this window is narrowed as we find the memory region containing the 683 // first free page of memory. To begin with, the range reflects the 684 // full process address space. 685 // 686 // firstFree is updated by calling foundFree each time free space in the 687 // heap is discovered. 688 // 689 // At the end of the search, base.addr() is the best new 690 // searchAddr we could deduce in this search. 691 firstFree := struct { 692 base, bound offAddr 693 }{ 694 base: minOffAddr, 695 bound: maxOffAddr, 696 } 697 // foundFree takes the given address range [addr, addr+size) and 698 // updates firstFree if it is a narrower range. The input range must 699 // either be fully contained within firstFree or not overlap with it 700 // at all. 701 // 702 // This way, we'll record the first summary we find with any free 703 // pages on the root level and narrow that down if we descend into 704 // that summary. But as soon as we need to iterate beyond that summary 705 // in a level to find a large enough range, we'll stop narrowing. 706 foundFree := func(addr offAddr, size uintptr) { 707 if firstFree.base.lessEqual(addr) && addr.add(size-1).lessEqual(firstFree.bound) { 708 // This range fits within the current firstFree window, so narrow 709 // down the firstFree window to the base and bound of this range. 710 firstFree.base = addr 711 firstFree.bound = addr.add(size - 1) 712 } else if !(addr.add(size-1).lessThan(firstFree.base) || firstFree.bound.lessThan(addr)) { 713 // This range only partially overlaps with the firstFree range, 714 // so throw. 715 print("runtime: addr = ", hex(addr.addr()), ", size = ", size, "\n") 716 print("runtime: base = ", hex(firstFree.base.addr()), ", bound = ", hex(firstFree.bound.addr()), "\n") 717 throw("range partially overlaps") 718 } 719 } 720 721 // lastSum is the summary which we saw on the previous level that made us 722 // move on to the next level. Used to print additional information in the 723 // case of a catastrophic failure. 724 // lastSumIdx is that summary's index in the previous level. 725 lastSum := packPallocSum(0, 0, 0) 726 lastSumIdx := -1 727 728 nextLevel: 729 for l := 0; l < len(p.summary); l++ { 730 // For the root level, entriesPerBlock is the whole level. 731 entriesPerBlock := 1 << levelBits[l] 732 logMaxPages := levelLogPages[l] 733 734 // We've moved into a new level, so let's update i to our new 735 // starting index. This is a no-op for level 0. 736 i <<= levelBits[l] 737 738 // Slice out the block of entries we care about. 739 entries := p.summary[l][i : i+entriesPerBlock] 740 741 // Determine j0, the first index we should start iterating from. 742 // The searchAddr may help us eliminate iterations if we followed the 743 // searchAddr on the previous level or we're on the root level, in which 744 // case the searchAddr should be the same as i after levelShift. 745 j0 := 0 746 if searchIdx := offAddrToLevelIndex(l, p.searchAddr); searchIdx&^(entriesPerBlock-1) == i { 747 j0 = searchIdx & (entriesPerBlock - 1) 748 } 749 750 // Run over the level entries looking for 751 // a contiguous run of at least npages either 752 // within an entry or across entries. 753 // 754 // base contains the page index (relative to 755 // the first entry's first page) of the currently 756 // considered run of consecutive pages. 757 // 758 // size contains the size of the currently considered 759 // run of consecutive pages. 760 var base, size uint 761 for j := j0; j < len(entries); j++ { 762 sum := entries[j] 763 if sum == 0 { 764 // A full entry means we broke any streak and 765 // that we should skip it altogether. 766 size = 0 767 continue 768 } 769 770 // We've encountered a non-zero summary which means 771 // free memory, so update firstFree. 772 foundFree(levelIndexToOffAddr(l, i+j), (uintptr(1)<<logMaxPages)*pageSize) 773 774 s := sum.start() 775 if size+s >= uint(npages) { 776 // If size == 0 we don't have a run yet, 777 // which means base isn't valid. So, set 778 // base to the first page in this block. 779 if size == 0 { 780 base = uint(j) << logMaxPages 781 } 782 // We hit npages; we're done! 783 size += s 784 break 785 } 786 if sum.max() >= uint(npages) { 787 // The entry itself contains npages contiguous 788 // free pages, so continue on the next level 789 // to find that run. 790 i += j 791 lastSumIdx = i 792 lastSum = sum 793 continue nextLevel 794 } 795 if size == 0 || s < 1<<logMaxPages { 796 // We either don't have a current run started, or this entry 797 // isn't totally free (meaning we can't continue the current 798 // one), so try to begin a new run by setting size and base 799 // based on sum.end. 800 size = sum.end() 801 base = uint(j+1)<<logMaxPages - size 802 continue 803 } 804 // The entry is completely free, so continue the run. 805 size += 1 << logMaxPages 806 } 807 if size >= uint(npages) { 808 // We found a sufficiently large run of free pages straddling 809 // some boundary, so compute the address and return it. 810 addr := levelIndexToOffAddr(l, i).add(uintptr(base) * pageSize).addr() 811 return addr, p.findMappedAddr(firstFree.base) 812 } 813 if l == 0 { 814 // We're at level zero, so that means we've exhausted our search. 815 return 0, maxSearchAddr() 816 } 817 818 // We're not at level zero, and we exhausted the level we were looking in. 819 // This means that either our calculations were wrong or the level above 820 // lied to us. In either case, dump some useful state and throw. 821 print("runtime: summary[", l-1, "][", lastSumIdx, "] = ", lastSum.start(), ", ", lastSum.max(), ", ", lastSum.end(), "\n") 822 print("runtime: level = ", l, ", npages = ", npages, ", j0 = ", j0, "\n") 823 print("runtime: p.searchAddr = ", hex(p.searchAddr.addr()), ", i = ", i, "\n") 824 print("runtime: levelShift[level] = ", levelShift[l], ", levelBits[level] = ", levelBits[l], "\n") 825 for j := 0; j < len(entries); j++ { 826 sum := entries[j] 827 print("runtime: summary[", l, "][", i+j, "] = (", sum.start(), ", ", sum.max(), ", ", sum.end(), ")\n") 828 } 829 throw("bad summary data") 830 } 831 832 // Since we've gotten to this point, that means we haven't found a 833 // sufficiently-sized free region straddling some boundary (chunk or larger). 834 // This means the last summary we inspected must have had a large enough "max" 835 // value, so look inside the chunk to find a suitable run. 836 // 837 // After iterating over all levels, i must contain a chunk index which 838 // is what the final level represents. 839 ci := chunkIdx(i) 840 j, searchIdx := p.chunkOf(ci).find(npages, 0) 841 if j == ^uint(0) { 842 // We couldn't find any space in this chunk despite the summaries telling 843 // us it should be there. There's likely a bug, so dump some state and throw. 844 sum := p.summary[len(p.summary)-1][i] 845 print("runtime: summary[", len(p.summary)-1, "][", i, "] = (", sum.start(), ", ", sum.max(), ", ", sum.end(), ")\n") 846 print("runtime: npages = ", npages, "\n") 847 throw("bad summary data") 848 } 849 850 // Compute the address at which the free space starts. 851 addr := chunkBase(ci) + uintptr(j)*pageSize 852 853 // Since we actually searched the chunk, we may have 854 // found an even narrower free window. 855 searchAddr := chunkBase(ci) + uintptr(searchIdx)*pageSize 856 foundFree(offAddr{searchAddr}, chunkBase(ci+1)-searchAddr) 857 return addr, p.findMappedAddr(firstFree.base) 858 } 859 860 // alloc allocates npages worth of memory from the page heap, returning the base 861 // address for the allocation and the amount of scavenged memory in bytes 862 // contained in the region [base address, base address + npages*pageSize). 863 // 864 // Returns a 0 base address on failure, in which case other returned values 865 // should be ignored. 866 // 867 // p.mheapLock must be held. 868 // 869 // Must run on the system stack because p.mheapLock must be held. 870 // 871 //go:systemstack 872 func (p *pageAlloc) alloc(npages uintptr) (addr uintptr, scav uintptr) { 873 assertLockHeld(p.mheapLock) 874 875 // If the searchAddr refers to a region which has a higher address than 876 // any known chunk, then we know we're out of memory. 877 if chunkIndex(p.searchAddr.addr()) >= p.end { 878 return 0, 0 879 } 880 881 // If npages has a chance of fitting in the chunk where the searchAddr is, 882 // search it directly. 883 searchAddr := minOffAddr 884 if pallocChunkPages-chunkPageIndex(p.searchAddr.addr()) >= uint(npages) { 885 // npages is guaranteed to be no greater than pallocChunkPages here. 886 i := chunkIndex(p.searchAddr.addr()) 887 if max := p.summary[len(p.summary)-1][i].max(); max >= uint(npages) { 888 j, searchIdx := p.chunkOf(i).find(npages, chunkPageIndex(p.searchAddr.addr())) 889 if j == ^uint(0) { 890 print("runtime: max = ", max, ", npages = ", npages, "\n") 891 print("runtime: searchIdx = ", chunkPageIndex(p.searchAddr.addr()), ", p.searchAddr = ", hex(p.searchAddr.addr()), "\n") 892 throw("bad summary data") 893 } 894 addr = chunkBase(i) + uintptr(j)*pageSize 895 searchAddr = offAddr{chunkBase(i) + uintptr(searchIdx)*pageSize} 896 goto Found 897 } 898 } 899 // We failed to use a searchAddr for one reason or another, so try 900 // the slow path. 901 addr, searchAddr = p.find(npages) 902 if addr == 0 { 903 if npages == 1 { 904 // We failed to find a single free page, the smallest unit 905 // of allocation. This means we know the heap is completely 906 // exhausted. Otherwise, the heap still might have free 907 // space in it, just not enough contiguous space to 908 // accommodate npages. 909 p.searchAddr = maxSearchAddr() 910 } 911 return 0, 0 912 } 913 Found: 914 // Go ahead and actually mark the bits now that we have an address. 915 scav = p.allocRange(addr, npages) 916 917 // If we found a higher searchAddr, we know that all the 918 // heap memory before that searchAddr in an offset address space is 919 // allocated, so bump p.searchAddr up to the new one. 920 if p.searchAddr.lessThan(searchAddr) { 921 p.searchAddr = searchAddr 922 } 923 return addr, scav 924 } 925 926 // free returns npages worth of memory starting at base back to the page heap. 927 // 928 // p.mheapLock must be held. 929 // 930 // Must run on the system stack because p.mheapLock must be held. 931 // 932 //go:systemstack 933 func (p *pageAlloc) free(base, npages uintptr) { 934 assertLockHeld(p.mheapLock) 935 936 // If we're freeing pages below the p.searchAddr, update searchAddr. 937 if b := (offAddr{base}); b.lessThan(p.searchAddr) { 938 p.searchAddr = b 939 } 940 limit := base + npages*pageSize - 1 941 if npages == 1 { 942 // Fast path: we're clearing a single bit, and we know exactly 943 // where it is, so mark it directly. 944 i := chunkIndex(base) 945 pi := chunkPageIndex(base) 946 p.chunkOf(i).free1(pi) 947 p.scav.index.free(i, pi, 1) 948 } else { 949 // Slow path: we're clearing more bits so we may need to iterate. 950 sc, ec := chunkIndex(base), chunkIndex(limit) 951 si, ei := chunkPageIndex(base), chunkPageIndex(limit) 952 953 if sc == ec { 954 // The range doesn't cross any chunk boundaries. 955 p.chunkOf(sc).free(si, ei+1-si) 956 p.scav.index.free(sc, si, ei+1-si) 957 } else { 958 // The range crosses at least one chunk boundary. 959 p.chunkOf(sc).free(si, pallocChunkPages-si) 960 p.scav.index.free(sc, si, pallocChunkPages-si) 961 for c := sc + 1; c < ec; c++ { 962 p.chunkOf(c).freeAll() 963 p.scav.index.free(c, 0, pallocChunkPages) 964 } 965 p.chunkOf(ec).free(0, ei+1) 966 p.scav.index.free(ec, 0, ei+1) 967 } 968 } 969 p.update(base, npages, true, false) 970 } 971 972 const ( 973 pallocSumBytes = unsafe.Sizeof(pallocSum(0)) 974 975 // maxPackedValue is the maximum value that any of the three fields in 976 // the pallocSum may take on. 977 maxPackedValue = 1 << logMaxPackedValue 978 logMaxPackedValue = logPallocChunkPages + (summaryLevels-1)*summaryLevelBits 979 980 freeChunkSum = pallocSum(uint64(pallocChunkPages) | 981 uint64(pallocChunkPages<<logMaxPackedValue) | 982 uint64(pallocChunkPages<<(2*logMaxPackedValue))) 983 ) 984 985 // pallocSum is a packed summary type which packs three numbers: start, max, 986 // and end into a single 8-byte value. Each of these values are a summary of 987 // a bitmap and are thus counts, each of which may have a maximum value of 988 // 2^21 - 1, or all three may be equal to 2^21. The latter case is represented 989 // by just setting the 64th bit. 990 type pallocSum uint64 991 992 // packPallocSum takes a start, max, and end value and produces a pallocSum. 993 func packPallocSum(start, max, end uint) pallocSum { 994 if max == maxPackedValue { 995 return pallocSum(uint64(1 << 63)) 996 } 997 return pallocSum((uint64(start) & (maxPackedValue - 1)) | 998 ((uint64(max) & (maxPackedValue - 1)) << logMaxPackedValue) | 999 ((uint64(end) & (maxPackedValue - 1)) << (2 * logMaxPackedValue))) 1000 } 1001 1002 // start extracts the start value from a packed sum. 1003 func (p pallocSum) start() uint { 1004 if uint64(p)&uint64(1<<63) != 0 { 1005 return maxPackedValue 1006 } 1007 return uint(uint64(p) & (maxPackedValue - 1)) 1008 } 1009 1010 // max extracts the max value from a packed sum. 1011 func (p pallocSum) max() uint { 1012 if uint64(p)&uint64(1<<63) != 0 { 1013 return maxPackedValue 1014 } 1015 return uint((uint64(p) >> logMaxPackedValue) & (maxPackedValue - 1)) 1016 } 1017 1018 // end extracts the end value from a packed sum. 1019 func (p pallocSum) end() uint { 1020 if uint64(p)&uint64(1<<63) != 0 { 1021 return maxPackedValue 1022 } 1023 return uint((uint64(p) >> (2 * logMaxPackedValue)) & (maxPackedValue - 1)) 1024 } 1025 1026 // unpack unpacks all three values from the summary. 1027 func (p pallocSum) unpack() (uint, uint, uint) { 1028 if uint64(p)&uint64(1<<63) != 0 { 1029 return maxPackedValue, maxPackedValue, maxPackedValue 1030 } 1031 return uint(uint64(p) & (maxPackedValue - 1)), 1032 uint((uint64(p) >> logMaxPackedValue) & (maxPackedValue - 1)), 1033 uint((uint64(p) >> (2 * logMaxPackedValue)) & (maxPackedValue - 1)) 1034 } 1035 1036 // mergeSummaries merges consecutive summaries which may each represent at 1037 // most 1 << logMaxPagesPerSum pages each together into one. 1038 func mergeSummaries(sums []pallocSum, logMaxPagesPerSum uint) pallocSum { 1039 // Merge the summaries in sums into one. 1040 // 1041 // We do this by keeping a running summary representing the merged 1042 // summaries of sums[:i] in start, most, and end. 1043 start, most, end := sums[0].unpack() 1044 for i := 1; i < len(sums); i++ { 1045 // Merge in sums[i]. 1046 si, mi, ei := sums[i].unpack() 1047 1048 // Merge in sums[i].start only if the running summary is 1049 // completely free, otherwise this summary's start 1050 // plays no role in the combined sum. 1051 if start == uint(i)<<logMaxPagesPerSum { 1052 start += si 1053 } 1054 1055 // Recompute the max value of the running sum by looking 1056 // across the boundary between the running sum and sums[i] 1057 // and at the max sums[i], taking the greatest of those two 1058 // and the max of the running sum. 1059 most = max(most, end+si, mi) 1060 1061 // Merge in end by checking if this new summary is totally 1062 // free. If it is, then we want to extend the running sum's 1063 // end by the new summary. If not, then we have some alloc'd 1064 // pages in there and we just want to take the end value in 1065 // sums[i]. 1066 if ei == 1<<logMaxPagesPerSum { 1067 end += 1 << logMaxPagesPerSum 1068 } else { 1069 end = ei 1070 } 1071 } 1072 return packPallocSum(start, most, end) 1073 } 1074