github.com/mtsmfm/go/src@v0.0.0-20221020090648-44bdcb9f8fde/runtime/mpagealloc.go (about)

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