Source file src/runtime/mpagealloc_64bit.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  //go:build amd64 || arm64 || loong64 || mips64 || mips64le || ppc64 || ppc64le || riscv64 || s390x
     6  
     7  package runtime
     8  
     9  import (
    10  	"unsafe"
    11  )
    12  
    13  const (
    14  	// The number of levels in the radix tree.
    15  	summaryLevels = 5
    16  
    17  	// Constants for testing.
    18  	pageAlloc32Bit = 0
    19  	pageAlloc64Bit = 1
    20  
    21  	// Number of bits needed to represent all indices into the L1 of the
    22  	// chunks map.
    23  	//
    24  	// See (*pageAlloc).chunks for more details. Update the documentation
    25  	// there should this number change.
    26  	pallocChunksL1Bits = 13
    27  )
    28  
    29  // levelBits is the number of bits in the radix for a given level in the super summary
    30  // structure.
    31  //
    32  // The sum of all the entries of levelBits should equal heapAddrBits.
    33  var levelBits = [summaryLevels]uint{
    34  	summaryL0Bits,
    35  	summaryLevelBits,
    36  	summaryLevelBits,
    37  	summaryLevelBits,
    38  	summaryLevelBits,
    39  }
    40  
    41  // levelShift is the number of bits to shift to acquire the radix for a given level
    42  // in the super summary structure.
    43  //
    44  // With levelShift, one can compute the index of the summary at level l related to a
    45  // pointer p by doing:
    46  //
    47  //	p >> levelShift[l]
    48  var levelShift = [summaryLevels]uint{
    49  	heapAddrBits - summaryL0Bits,
    50  	heapAddrBits - summaryL0Bits - 1*summaryLevelBits,
    51  	heapAddrBits - summaryL0Bits - 2*summaryLevelBits,
    52  	heapAddrBits - summaryL0Bits - 3*summaryLevelBits,
    53  	heapAddrBits - summaryL0Bits - 4*summaryLevelBits,
    54  }
    55  
    56  // levelLogPages is log2 the maximum number of runtime pages in the address space
    57  // a summary in the given level represents.
    58  //
    59  // The leaf level always represents exactly log2 of 1 chunk's worth of pages.
    60  var levelLogPages = [summaryLevels]uint{
    61  	logPallocChunkPages + 4*summaryLevelBits,
    62  	logPallocChunkPages + 3*summaryLevelBits,
    63  	logPallocChunkPages + 2*summaryLevelBits,
    64  	logPallocChunkPages + 1*summaryLevelBits,
    65  	logPallocChunkPages,
    66  }
    67  
    68  // sysInit performs architecture-dependent initialization of fields
    69  // in pageAlloc. pageAlloc should be uninitialized except for sysStat
    70  // if any runtime statistic should be updated.
    71  func (p *pageAlloc) sysInit(test bool) {
    72  	// Reserve memory for each level. This will get mapped in
    73  	// as R/W by setArenas.
    74  	for l, shift := range levelShift {
    75  		entries := 1 << (heapAddrBits - shift)
    76  
    77  		// Reserve b bytes of memory anywhere in the address space.
    78  		b := alignUp(uintptr(entries)*pallocSumBytes, physPageSize)
    79  		r := sysReserve(nil, b)
    80  		if r == nil {
    81  			throw("failed to reserve page summary memory")
    82  		}
    83  
    84  		// Put this reservation into a slice.
    85  		sl := notInHeapSlice{(*notInHeap)(r), 0, entries}
    86  		p.summary[l] = *(*[]pallocSum)(unsafe.Pointer(&sl))
    87  	}
    88  }
    89  
    90  // sysGrow performs architecture-dependent operations on heap
    91  // growth for the page allocator, such as mapping in new memory
    92  // for summaries. It also updates the length of the slices in
    93  // p.summary.
    94  //
    95  // base is the base of the newly-added heap memory and limit is
    96  // the first address past the end of the newly-added heap memory.
    97  // Both must be aligned to pallocChunkBytes.
    98  //
    99  // The caller must update p.start and p.end after calling sysGrow.
   100  func (p *pageAlloc) sysGrow(base, limit uintptr) {
   101  	if base%pallocChunkBytes != 0 || limit%pallocChunkBytes != 0 {
   102  		print("runtime: base = ", hex(base), ", limit = ", hex(limit), "\n")
   103  		throw("sysGrow bounds not aligned to pallocChunkBytes")
   104  	}
   105  
   106  	// addrRangeToSummaryRange converts a range of addresses into a range
   107  	// of summary indices which must be mapped to support those addresses
   108  	// in the summary range.
   109  	addrRangeToSummaryRange := func(level int, r addrRange) (int, int) {
   110  		sumIdxBase, sumIdxLimit := addrsToSummaryRange(level, r.base.addr(), r.limit.addr())
   111  		return blockAlignSummaryRange(level, sumIdxBase, sumIdxLimit)
   112  	}
   113  
   114  	// summaryRangeToSumAddrRange converts a range of indices in any
   115  	// level of p.summary into page-aligned addresses which cover that
   116  	// range of indices.
   117  	summaryRangeToSumAddrRange := func(level, sumIdxBase, sumIdxLimit int) addrRange {
   118  		baseOffset := alignDown(uintptr(sumIdxBase)*pallocSumBytes, physPageSize)
   119  		limitOffset := alignUp(uintptr(sumIdxLimit)*pallocSumBytes, physPageSize)
   120  		base := unsafe.Pointer(&p.summary[level][0])
   121  		return addrRange{
   122  			offAddr{uintptr(add(base, baseOffset))},
   123  			offAddr{uintptr(add(base, limitOffset))},
   124  		}
   125  	}
   126  
   127  	// addrRangeToSumAddrRange is a convenience function that converts
   128  	// an address range r to the address range of the given summary level
   129  	// that stores the summaries for r.
   130  	addrRangeToSumAddrRange := func(level int, r addrRange) addrRange {
   131  		sumIdxBase, sumIdxLimit := addrRangeToSummaryRange(level, r)
   132  		return summaryRangeToSumAddrRange(level, sumIdxBase, sumIdxLimit)
   133  	}
   134  
   135  	// Find the first inUse index which is strictly greater than base.
   136  	//
   137  	// Because this function will never be asked remap the same memory
   138  	// twice, this index is effectively the index at which we would insert
   139  	// this new growth, and base will never overlap/be contained within
   140  	// any existing range.
   141  	//
   142  	// This will be used to look at what memory in the summary array is already
   143  	// mapped before and after this new range.
   144  	inUseIndex := p.inUse.findSucc(base)
   145  
   146  	// Walk up the radix tree and map summaries in as needed.
   147  	for l := range p.summary {
   148  		// Figure out what part of the summary array this new address space needs.
   149  		needIdxBase, needIdxLimit := addrRangeToSummaryRange(l, makeAddrRange(base, limit))
   150  
   151  		// Update the summary slices with a new upper-bound. This ensures
   152  		// we get tight bounds checks on at least the top bound.
   153  		//
   154  		// We must do this regardless of whether we map new memory.
   155  		if needIdxLimit > len(p.summary[l]) {
   156  			p.summary[l] = p.summary[l][:needIdxLimit]
   157  		}
   158  
   159  		// Compute the needed address range in the summary array for level l.
   160  		need := summaryRangeToSumAddrRange(l, needIdxBase, needIdxLimit)
   161  
   162  		// Prune need down to what needs to be newly mapped. Some parts of it may
   163  		// already be mapped by what inUse describes due to page alignment requirements
   164  		// for mapping. Because this function will never be asked to remap the same
   165  		// memory twice, it should never be possible to prune in such a way that causes
   166  		// need to be split.
   167  		if inUseIndex > 0 {
   168  			need = need.subtract(addrRangeToSumAddrRange(l, p.inUse.ranges[inUseIndex-1]))
   169  		}
   170  		if inUseIndex < len(p.inUse.ranges) {
   171  			need = need.subtract(addrRangeToSumAddrRange(l, p.inUse.ranges[inUseIndex]))
   172  		}
   173  		// It's possible that after our pruning above, there's nothing new to map.
   174  		if need.size() == 0 {
   175  			continue
   176  		}
   177  
   178  		// Map and commit need.
   179  		sysMap(unsafe.Pointer(need.base.addr()), need.size(), p.sysStat)
   180  		sysUsed(unsafe.Pointer(need.base.addr()), need.size(), need.size())
   181  		p.summaryMappedReady += need.size()
   182  	}
   183  
   184  	// Update the scavenge index.
   185  	p.summaryMappedReady += p.scav.index.sysGrow(base, limit, p.sysStat)
   186  }
   187  
   188  // sysGrow increases the index's backing store in response to a heap growth.
   189  //
   190  // Returns the amount of memory added to sysStat.
   191  func (s *scavengeIndex) sysGrow(base, limit uintptr, sysStat *sysMemStat) uintptr {
   192  	if base%pallocChunkBytes != 0 || limit%pallocChunkBytes != 0 {
   193  		print("runtime: base = ", hex(base), ", limit = ", hex(limit), "\n")
   194  		throw("sysGrow bounds not aligned to pallocChunkBytes")
   195  	}
   196  	scSize := unsafe.Sizeof(atomicScavChunkData{})
   197  	// Map and commit the pieces of chunks that we need.
   198  	//
   199  	// We always map the full range of the minimum heap address to the
   200  	// maximum heap address. We don't do this for the summary structure
   201  	// because it's quite large and a discontiguous heap could cause a
   202  	// lot of memory to be used. In this situation, the worst case overhead
   203  	// is in the single-digit MiB if we map the whole thing.
   204  	//
   205  	// The base address of the backing store is always page-aligned,
   206  	// because it comes from the OS, so it's sufficient to align the
   207  	// index.
   208  	haveMin := s.min.Load()
   209  	haveMax := s.max.Load()
   210  	needMin := alignDown(uintptr(chunkIndex(base)), physPageSize/scSize)
   211  	needMax := alignUp(uintptr(chunkIndex(limit)), physPageSize/scSize)
   212  
   213  	// We need a contiguous range, so extend the range if there's no overlap.
   214  	if needMax < haveMin {
   215  		needMax = haveMin
   216  	}
   217  	if haveMax != 0 && needMin > haveMax {
   218  		needMin = haveMax
   219  	}
   220  
   221  	// Avoid a panic from indexing one past the last element.
   222  	chunksBase := uintptr(unsafe.Pointer(&s.chunks[0]))
   223  	have := makeAddrRange(chunksBase+haveMin*scSize, chunksBase+haveMax*scSize)
   224  	need := makeAddrRange(chunksBase+needMin*scSize, chunksBase+needMax*scSize)
   225  
   226  	// Subtract any overlap from rounding. We can't re-map memory because
   227  	// it'll be zeroed.
   228  	need = need.subtract(have)
   229  
   230  	// If we've got something to map, map it, and update the slice bounds.
   231  	if need.size() != 0 {
   232  		sysMap(unsafe.Pointer(need.base.addr()), need.size(), sysStat)
   233  		sysUsed(unsafe.Pointer(need.base.addr()), need.size(), need.size())
   234  		// Update the indices only after the new memory is valid.
   235  		if haveMax == 0 || needMin < haveMin {
   236  			s.min.Store(needMin)
   237  		}
   238  		if needMax > haveMax {
   239  			s.max.Store(needMax)
   240  		}
   241  	}
   242  	return need.size()
   243  }
   244  
   245  // sysInit initializes the scavengeIndex' chunks array.
   246  //
   247  // Returns the amount of memory added to sysStat.
   248  func (s *scavengeIndex) sysInit(test bool, sysStat *sysMemStat) uintptr {
   249  	n := uintptr(1<<heapAddrBits) / pallocChunkBytes
   250  	nbytes := n * unsafe.Sizeof(atomicScavChunkData{})
   251  	r := sysReserve(nil, nbytes)
   252  	sl := notInHeapSlice{(*notInHeap)(r), int(n), int(n)}
   253  	s.chunks = *(*[]atomicScavChunkData)(unsafe.Pointer(&sl))
   254  	return 0 // All memory above is mapped Reserved.
   255  }
   256  

View as plain text