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  	"runtime/internal/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  // chunkIndex 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  	// This field is currently unused on 32-bit architectures but
   262  	// is harmless to track. We care much more about having a
   263  	// contiguous heap in these cases and take additional measures
   264  	// to ensure that, so in nearly all cases this should have just
   265  	// 1 element.
   266  	//
   267  	// All access is protected by the mheapLock.
   268  	inUse addrRanges
   269  
   270  	_ uint32 // Align scav so it's easier to reason about alignment within scav.
   271  
   272  	// scav stores the scavenger state.
   273  	scav struct {
   274  		// index is an efficient index of chunks that have pages available to
   275  		// scavenge.
   276  		index scavengeIndex
   277  
   278  		// released is the amount of memory released this generation.
   279  		//
   280  		// Updated atomically.
   281  		released uintptr
   282  
   283  		_ uint32 // Align assistTime for atomics on 32-bit platforms.
   284  
   285  		// scavengeAssistTime is the time spent scavenging in the last GC cycle.
   286  		//
   287  		// This is reset once a GC cycle ends.
   288  		assistTime atomic.Int64
   289  	}
   290  
   291  	// mheap_.lock. This level of indirection makes it possible
   292  	// to test pageAlloc indepedently of the runtime allocator.
   293  	mheapLock *mutex
   294  
   295  	// sysStat is the runtime memstat to update when new system
   296  	// memory is committed by the pageAlloc for allocation metadata.
   297  	sysStat *sysMemStat
   298  
   299  	// summaryMappedReady is the number of bytes mapped in the Ready state
   300  	// in the summary structure. Used only for testing currently.
   301  	//
   302  	// Protected by mheapLock.
   303  	summaryMappedReady uintptr
   304  
   305  	// Whether or not this struct is being used in tests.
   306  	test bool
   307  }
   308  
   309  func (p *pageAlloc) init(mheapLock *mutex, sysStat *sysMemStat) {
   310  	if levelLogPages[0] > logMaxPackedValue {
   311  		// We can't represent 1<<levelLogPages[0] pages, the maximum number
   312  		// of pages we need to represent at the root level, in a summary, which
   313  		// is a big problem. Throw.
   314  		print("runtime: root level max pages = ", 1<<levelLogPages[0], "\n")
   315  		print("runtime: summary max pages = ", maxPackedValue, "\n")
   316  		throw("root level max pages doesn't fit in summary")
   317  	}
   318  	p.sysStat = sysStat
   319  
   320  	// Initialize p.inUse.
   321  	p.inUse.init(sysStat)
   322  
   323  	// System-dependent initialization.
   324  	p.sysInit()
   325  
   326  	// Start with the searchAddr in a state indicating there's no free memory.
   327  	p.searchAddr = maxSearchAddr()
   328  
   329  	// Set the mheapLock.
   330  	p.mheapLock = mheapLock
   331  }
   332  
   333  // tryChunkOf returns the bitmap data for the given chunk.
   334  //
   335  // Returns nil if the chunk data has not been mapped.
   336  func (p *pageAlloc) tryChunkOf(ci chunkIdx) *pallocData {
   337  	l2 := p.chunks[ci.l1()]
   338  	if l2 == nil {
   339  		return nil
   340  	}
   341  	return &l2[ci.l2()]
   342  }
   343  
   344  // chunkOf returns the chunk at the given chunk index.
   345  //
   346  // The chunk index must be valid or this method may throw.
   347  func (p *pageAlloc) chunkOf(ci chunkIdx) *pallocData {
   348  	return &p.chunks[ci.l1()][ci.l2()]
   349  }
   350  
   351  // grow sets up the metadata for the address range [base, base+size).
   352  // It may allocate metadata, in which case *p.sysStat will be updated.
   353  //
   354  // p.mheapLock must be held.
   355  func (p *pageAlloc) grow(base, size uintptr) {
   356  	assertLockHeld(p.mheapLock)
   357  
   358  	// Round up to chunks, since we can't deal with increments smaller
   359  	// than chunks. Also, sysGrow expects aligned values.
   360  	limit := alignUp(base+size, pallocChunkBytes)
   361  	base = alignDown(base, pallocChunkBytes)
   362  
   363  	// Grow the summary levels in a system-dependent manner.
   364  	// We just update a bunch of additional metadata here.
   365  	p.sysGrow(base, limit)
   366  
   367  	// Update p.start and p.end.
   368  	// If no growth happened yet, start == 0. This is generally
   369  	// safe since the zero page is unmapped.
   370  	firstGrowth := p.start == 0
   371  	start, end := chunkIndex(base), chunkIndex(limit)
   372  	if firstGrowth || start < p.start {
   373  		p.start = start
   374  	}
   375  	if end > p.end {
   376  		p.end = end
   377  	}
   378  	// Note that [base, limit) will never overlap with any existing
   379  	// range inUse because grow only ever adds never-used memory
   380  	// regions to the page allocator.
   381  	p.inUse.add(makeAddrRange(base, limit))
   382  
   383  	// A grow operation is a lot like a free operation, so if our
   384  	// chunk ends up below p.searchAddr, update p.searchAddr to the
   385  	// new address, just like in free.
   386  	if b := (offAddr{base}); b.lessThan(p.searchAddr) {
   387  		p.searchAddr = b
   388  	}
   389  
   390  	// Add entries into chunks, which is sparse, if needed. Then,
   391  	// initialize the bitmap.
   392  	//
   393  	// Newly-grown memory is always considered scavenged.
   394  	// Set all the bits in the scavenged bitmaps high.
   395  	for c := chunkIndex(base); c < chunkIndex(limit); c++ {
   396  		if p.chunks[c.l1()] == nil {
   397  			// Create the necessary l2 entry.
   398  			//
   399  			// Store it atomically to avoid races with readers which
   400  			// don't acquire the heap lock.
   401  			r := sysAlloc(unsafe.Sizeof(*p.chunks[0]), p.sysStat)
   402  			if r == nil {
   403  				throw("pageAlloc: out of memory")
   404  			}
   405  			atomic.StorepNoWB(unsafe.Pointer(&p.chunks[c.l1()]), r)
   406  		}
   407  		p.chunkOf(c).scavenged.setRange(0, pallocChunkPages)
   408  	}
   409  
   410  	// Update summaries accordingly. The grow acts like a free, so
   411  	// we need to ensure this newly-free memory is visible in the
   412  	// summaries.
   413  	p.update(base, size/pageSize, true, false)
   414  }
   415  
   416  // update updates heap metadata. It must be called each time the bitmap
   417  // is updated.
   418  //
   419  // If contig is true, update does some optimizations assuming that there was
   420  // a contiguous allocation or free between addr and addr+npages. alloc indicates
   421  // whether the operation performed was an allocation or a free.
   422  //
   423  // p.mheapLock must be held.
   424  func (p *pageAlloc) update(base, npages uintptr, contig, alloc bool) {
   425  	assertLockHeld(p.mheapLock)
   426  
   427  	// base, limit, start, and end are inclusive.
   428  	limit := base + npages*pageSize - 1
   429  	sc, ec := chunkIndex(base), chunkIndex(limit)
   430  
   431  	// Handle updating the lowest level first.
   432  	if sc == ec {
   433  		// Fast path: the allocation doesn't span more than one chunk,
   434  		// so update this one and if the summary didn't change, return.
   435  		x := p.summary[len(p.summary)-1][sc]
   436  		y := p.chunkOf(sc).summarize()
   437  		if x == y {
   438  			return
   439  		}
   440  		p.summary[len(p.summary)-1][sc] = y
   441  	} else if contig {
   442  		// Slow contiguous path: the allocation spans more than one chunk
   443  		// and at least one summary is guaranteed to change.
   444  		summary := p.summary[len(p.summary)-1]
   445  
   446  		// Update the summary for chunk sc.
   447  		summary[sc] = p.chunkOf(sc).summarize()
   448  
   449  		// Update the summaries for chunks in between, which are
   450  		// either totally allocated or freed.
   451  		whole := p.summary[len(p.summary)-1][sc+1 : ec]
   452  		if alloc {
   453  			// Should optimize into a memclr.
   454  			for i := range whole {
   455  				whole[i] = 0
   456  			}
   457  		} else {
   458  			for i := range whole {
   459  				whole[i] = freeChunkSum
   460  			}
   461  		}
   462  
   463  		// Update the summary for chunk ec.
   464  		summary[ec] = p.chunkOf(ec).summarize()
   465  	} else {
   466  		// Slow general path: the allocation spans more than one chunk
   467  		// and at least one summary is guaranteed to change.
   468  		//
   469  		// We can't assume a contiguous allocation happened, so walk over
   470  		// every chunk in the range and manually recompute the summary.
   471  		summary := p.summary[len(p.summary)-1]
   472  		for c := sc; c <= ec; c++ {
   473  			summary[c] = p.chunkOf(c).summarize()
   474  		}
   475  	}
   476  
   477  	// Walk up the radix tree and update the summaries appropriately.
   478  	changed := true
   479  	for l := len(p.summary) - 2; l >= 0 && changed; l-- {
   480  		// Update summaries at level l from summaries at level l+1.
   481  		changed = false
   482  
   483  		// "Constants" for the previous level which we
   484  		// need to compute the summary from that level.
   485  		logEntriesPerBlock := levelBits[l+1]
   486  		logMaxPages := levelLogPages[l+1]
   487  
   488  		// lo and hi describe all the parts of the level we need to look at.
   489  		lo, hi := addrsToSummaryRange(l, base, limit+1)
   490  
   491  		// Iterate over each block, updating the corresponding summary in the less-granular level.
   492  		for i := lo; i < hi; i++ {
   493  			children := p.summary[l+1][i<<logEntriesPerBlock : (i+1)<<logEntriesPerBlock]
   494  			sum := mergeSummaries(children, logMaxPages)
   495  			old := p.summary[l][i]
   496  			if old != sum {
   497  				changed = true
   498  				p.summary[l][i] = sum
   499  			}
   500  		}
   501  	}
   502  }
   503  
   504  // allocRange marks the range of memory [base, base+npages*pageSize) as
   505  // allocated. It also updates the summaries to reflect the newly-updated
   506  // bitmap.
   507  //
   508  // Returns the amount of scavenged memory in bytes present in the
   509  // allocated range.
   510  //
   511  // p.mheapLock must be held.
   512  func (p *pageAlloc) allocRange(base, npages uintptr) uintptr {
   513  	assertLockHeld(p.mheapLock)
   514  
   515  	limit := base + npages*pageSize - 1
   516  	sc, ec := chunkIndex(base), chunkIndex(limit)
   517  	si, ei := chunkPageIndex(base), chunkPageIndex(limit)
   518  
   519  	scav := uint(0)
   520  	if sc == ec {
   521  		// The range doesn't cross any chunk boundaries.
   522  		chunk := p.chunkOf(sc)
   523  		scav += chunk.scavenged.popcntRange(si, ei+1-si)
   524  		chunk.allocRange(si, ei+1-si)
   525  	} else {
   526  		// The range crosses at least one chunk boundary.
   527  		chunk := p.chunkOf(sc)
   528  		scav += chunk.scavenged.popcntRange(si, pallocChunkPages-si)
   529  		chunk.allocRange(si, pallocChunkPages-si)
   530  		for c := sc + 1; c < ec; c++ {
   531  			chunk := p.chunkOf(c)
   532  			scav += chunk.scavenged.popcntRange(0, pallocChunkPages)
   533  			chunk.allocAll()
   534  		}
   535  		chunk = p.chunkOf(ec)
   536  		scav += chunk.scavenged.popcntRange(0, ei+1)
   537  		chunk.allocRange(0, ei+1)
   538  	}
   539  	p.update(base, npages, true, true)
   540  	return uintptr(scav) * pageSize
   541  }
   542  
   543  // findMappedAddr returns the smallest mapped offAddr that is
   544  // >= addr. That is, if addr refers to mapped memory, then it is
   545  // returned. If addr is higher than any mapped region, then
   546  // it returns maxOffAddr.
   547  //
   548  // p.mheapLock must be held.
   549  func (p *pageAlloc) findMappedAddr(addr offAddr) offAddr {
   550  	assertLockHeld(p.mheapLock)
   551  
   552  	// If we're not in a test, validate first by checking mheap_.arenas.
   553  	// This is a fast path which is only safe to use outside of testing.
   554  	ai := arenaIndex(addr.addr())
   555  	if p.test || mheap_.arenas[ai.l1()] == nil || mheap_.arenas[ai.l1()][ai.l2()] == nil {
   556  		vAddr, ok := p.inUse.findAddrGreaterEqual(addr.addr())
   557  		if ok {
   558  			return offAddr{vAddr}
   559  		} else {
   560  			// The candidate search address is greater than any
   561  			// known address, which means we definitely have no
   562  			// free memory left.
   563  			return maxOffAddr
   564  		}
   565  	}
   566  	return addr
   567  }
   568  
   569  // find searches for the first (address-ordered) contiguous free region of
   570  // npages in size and returns a base address for that region.
   571  //
   572  // It uses p.searchAddr to prune its search and assumes that no palloc chunks
   573  // below chunkIndex(p.searchAddr) contain any free memory at all.
   574  //
   575  // find also computes and returns a candidate p.searchAddr, which may or
   576  // may not prune more of the address space than p.searchAddr already does.
   577  // This candidate is always a valid p.searchAddr.
   578  //
   579  // find represents the slow path and the full radix tree search.
   580  //
   581  // Returns a base address of 0 on failure, in which case the candidate
   582  // searchAddr returned is invalid and must be ignored.
   583  //
   584  // p.mheapLock must be held.
   585  func (p *pageAlloc) find(npages uintptr) (uintptr, offAddr) {
   586  	assertLockHeld(p.mheapLock)
   587  
   588  	// Search algorithm.
   589  	//
   590  	// This algorithm walks each level l of the radix tree from the root level
   591  	// to the leaf level. It iterates over at most 1 << levelBits[l] of entries
   592  	// in a given level in the radix tree, and uses the summary information to
   593  	// find either:
   594  	//  1) That a given subtree contains a large enough contiguous region, at
   595  	//     which point it continues iterating on the next level, or
   596  	//  2) That there are enough contiguous boundary-crossing bits to satisfy
   597  	//     the allocation, at which point it knows exactly where to start
   598  	//     allocating from.
   599  	//
   600  	// i tracks the index into the current level l's structure for the
   601  	// contiguous 1 << levelBits[l] entries we're actually interested in.
   602  	//
   603  	// NOTE: Technically this search could allocate a region which crosses
   604  	// the arenaBaseOffset boundary, which when arenaBaseOffset != 0, is
   605  	// a discontinuity. However, the only way this could happen is if the
   606  	// page at the zero address is mapped, and this is impossible on
   607  	// every system we support where arenaBaseOffset != 0. So, the
   608  	// discontinuity is already encoded in the fact that the OS will never
   609  	// map the zero page for us, and this function doesn't try to handle
   610  	// this case in any way.
   611  
   612  	// i is the beginning of the block of entries we're searching at the
   613  	// current level.
   614  	i := 0
   615  
   616  	// firstFree is the region of address space that we are certain to
   617  	// find the first free page in the heap. base and bound are the inclusive
   618  	// bounds of this window, and both are addresses in the linearized, contiguous
   619  	// view of the address space (with arenaBaseOffset pre-added). At each level,
   620  	// this window is narrowed as we find the memory region containing the
   621  	// first free page of memory. To begin with, the range reflects the
   622  	// full process address space.
   623  	//
   624  	// firstFree is updated by calling foundFree each time free space in the
   625  	// heap is discovered.
   626  	//
   627  	// At the end of the search, base.addr() is the best new
   628  	// searchAddr we could deduce in this search.
   629  	firstFree := struct {
   630  		base, bound offAddr
   631  	}{
   632  		base:  minOffAddr,
   633  		bound: maxOffAddr,
   634  	}
   635  	// foundFree takes the given address range [addr, addr+size) and
   636  	// updates firstFree if it is a narrower range. The input range must
   637  	// either be fully contained within firstFree or not overlap with it
   638  	// at all.
   639  	//
   640  	// This way, we'll record the first summary we find with any free
   641  	// pages on the root level and narrow that down if we descend into
   642  	// that summary. But as soon as we need to iterate beyond that summary
   643  	// in a level to find a large enough range, we'll stop narrowing.
   644  	foundFree := func(addr offAddr, size uintptr) {
   645  		if firstFree.base.lessEqual(addr) && addr.add(size-1).lessEqual(firstFree.bound) {
   646  			// This range fits within the current firstFree window, so narrow
   647  			// down the firstFree window to the base and bound of this range.
   648  			firstFree.base = addr
   649  			firstFree.bound = addr.add(size - 1)
   650  		} else if !(addr.add(size-1).lessThan(firstFree.base) || firstFree.bound.lessThan(addr)) {
   651  			// This range only partially overlaps with the firstFree range,
   652  			// so throw.
   653  			print("runtime: addr = ", hex(addr.addr()), ", size = ", size, "\n")
   654  			print("runtime: base = ", hex(firstFree.base.addr()), ", bound = ", hex(firstFree.bound.addr()), "\n")
   655  			throw("range partially overlaps")
   656  		}
   657  	}
   658  
   659  	// lastSum is the summary which we saw on the previous level that made us
   660  	// move on to the next level. Used to print additional information in the
   661  	// case of a catastrophic failure.
   662  	// lastSumIdx is that summary's index in the previous level.
   663  	lastSum := packPallocSum(0, 0, 0)
   664  	lastSumIdx := -1
   665  
   666  nextLevel:
   667  	for l := 0; l < len(p.summary); l++ {
   668  		// For the root level, entriesPerBlock is the whole level.
   669  		entriesPerBlock := 1 << levelBits[l]
   670  		logMaxPages := levelLogPages[l]
   671  
   672  		// We've moved into a new level, so let's update i to our new
   673  		// starting index. This is a no-op for level 0.
   674  		i <<= levelBits[l]
   675  
   676  		// Slice out the block of entries we care about.
   677  		entries := p.summary[l][i : i+entriesPerBlock]
   678  
   679  		// Determine j0, the first index we should start iterating from.
   680  		// The searchAddr may help us eliminate iterations if we followed the
   681  		// searchAddr on the previous level or we're on the root leve, in which
   682  		// case the searchAddr should be the same as i after levelShift.
   683  		j0 := 0
   684  		if searchIdx := offAddrToLevelIndex(l, p.searchAddr); searchIdx&^(entriesPerBlock-1) == i {
   685  			j0 = searchIdx & (entriesPerBlock - 1)
   686  		}
   687  
   688  		// Run over the level entries looking for
   689  		// a contiguous run of at least npages either
   690  		// within an entry or across entries.
   691  		//
   692  		// base contains the page index (relative to
   693  		// the first entry's first page) of the currently
   694  		// considered run of consecutive pages.
   695  		//
   696  		// size contains the size of the currently considered
   697  		// run of consecutive pages.
   698  		var base, size uint
   699  		for j := j0; j < len(entries); j++ {
   700  			sum := entries[j]
   701  			if sum == 0 {
   702  				// A full entry means we broke any streak and
   703  				// that we should skip it altogether.
   704  				size = 0
   705  				continue
   706  			}
   707  
   708  			// We've encountered a non-zero summary which means
   709  			// free memory, so update firstFree.
   710  			foundFree(levelIndexToOffAddr(l, i+j), (uintptr(1)<<logMaxPages)*pageSize)
   711  
   712  			s := sum.start()
   713  			if size+s >= uint(npages) {
   714  				// If size == 0 we don't have a run yet,
   715  				// which means base isn't valid. So, set
   716  				// base to the first page in this block.
   717  				if size == 0 {
   718  					base = uint(j) << logMaxPages
   719  				}
   720  				// We hit npages; we're done!
   721  				size += s
   722  				break
   723  			}
   724  			if sum.max() >= uint(npages) {
   725  				// The entry itself contains npages contiguous
   726  				// free pages, so continue on the next level
   727  				// to find that run.
   728  				i += j
   729  				lastSumIdx = i
   730  				lastSum = sum
   731  				continue nextLevel
   732  			}
   733  			if size == 0 || s < 1<<logMaxPages {
   734  				// We either don't have a current run started, or this entry
   735  				// isn't totally free (meaning we can't continue the current
   736  				// one), so try to begin a new run by setting size and base
   737  				// based on sum.end.
   738  				size = sum.end()
   739  				base = uint(j+1)<<logMaxPages - size
   740  				continue
   741  			}
   742  			// The entry is completely free, so continue the run.
   743  			size += 1 << logMaxPages
   744  		}
   745  		if size >= uint(npages) {
   746  			// We found a sufficiently large run of free pages straddling
   747  			// some boundary, so compute the address and return it.
   748  			addr := levelIndexToOffAddr(l, i).add(uintptr(base) * pageSize).addr()
   749  			return addr, p.findMappedAddr(firstFree.base)
   750  		}
   751  		if l == 0 {
   752  			// We're at level zero, so that means we've exhausted our search.
   753  			return 0, maxSearchAddr()
   754  		}
   755  
   756  		// We're not at level zero, and we exhausted the level we were looking in.
   757  		// This means that either our calculations were wrong or the level above
   758  		// lied to us. In either case, dump some useful state and throw.
   759  		print("runtime: summary[", l-1, "][", lastSumIdx, "] = ", lastSum.start(), ", ", lastSum.max(), ", ", lastSum.end(), "\n")
   760  		print("runtime: level = ", l, ", npages = ", npages, ", j0 = ", j0, "\n")
   761  		print("runtime: p.searchAddr = ", hex(p.searchAddr.addr()), ", i = ", i, "\n")
   762  		print("runtime: levelShift[level] = ", levelShift[l], ", levelBits[level] = ", levelBits[l], "\n")
   763  		for j := 0; j < len(entries); j++ {
   764  			sum := entries[j]
   765  			print("runtime: summary[", l, "][", i+j, "] = (", sum.start(), ", ", sum.max(), ", ", sum.end(), ")\n")
   766  		}
   767  		throw("bad summary data")
   768  	}
   769  
   770  	// Since we've gotten to this point, that means we haven't found a
   771  	// sufficiently-sized free region straddling some boundary (chunk or larger).
   772  	// This means the last summary we inspected must have had a large enough "max"
   773  	// value, so look inside the chunk to find a suitable run.
   774  	//
   775  	// After iterating over all levels, i must contain a chunk index which
   776  	// is what the final level represents.
   777  	ci := chunkIdx(i)
   778  	j, searchIdx := p.chunkOf(ci).find(npages, 0)
   779  	if j == ^uint(0) {
   780  		// We couldn't find any space in this chunk despite the summaries telling
   781  		// us it should be there. There's likely a bug, so dump some state and throw.
   782  		sum := p.summary[len(p.summary)-1][i]
   783  		print("runtime: summary[", len(p.summary)-1, "][", i, "] = (", sum.start(), ", ", sum.max(), ", ", sum.end(), ")\n")
   784  		print("runtime: npages = ", npages, "\n")
   785  		throw("bad summary data")
   786  	}
   787  
   788  	// Compute the address at which the free space starts.
   789  	addr := chunkBase(ci) + uintptr(j)*pageSize
   790  
   791  	// Since we actually searched the chunk, we may have
   792  	// found an even narrower free window.
   793  	searchAddr := chunkBase(ci) + uintptr(searchIdx)*pageSize
   794  	foundFree(offAddr{searchAddr}, chunkBase(ci+1)-searchAddr)
   795  	return addr, p.findMappedAddr(firstFree.base)
   796  }
   797  
   798  // alloc allocates npages worth of memory from the page heap, returning the base
   799  // address for the allocation and the amount of scavenged memory in bytes
   800  // contained in the region [base address, base address + npages*pageSize).
   801  //
   802  // Returns a 0 base address on failure, in which case other returned values
   803  // should be ignored.
   804  //
   805  // p.mheapLock must be held.
   806  //
   807  // Must run on the system stack because p.mheapLock must be held.
   808  //
   809  //go:systemstack
   810  func (p *pageAlloc) alloc(npages uintptr) (addr uintptr, scav uintptr) {
   811  	assertLockHeld(p.mheapLock)
   812  
   813  	// If the searchAddr refers to a region which has a higher address than
   814  	// any known chunk, then we know we're out of memory.
   815  	if chunkIndex(p.searchAddr.addr()) >= p.end {
   816  		return 0, 0
   817  	}
   818  
   819  	// If npages has a chance of fitting in the chunk where the searchAddr is,
   820  	// search it directly.
   821  	searchAddr := minOffAddr
   822  	if pallocChunkPages-chunkPageIndex(p.searchAddr.addr()) >= uint(npages) {
   823  		// npages is guaranteed to be no greater than pallocChunkPages here.
   824  		i := chunkIndex(p.searchAddr.addr())
   825  		if max := p.summary[len(p.summary)-1][i].max(); max >= uint(npages) {
   826  			j, searchIdx := p.chunkOf(i).find(npages, chunkPageIndex(p.searchAddr.addr()))
   827  			if j == ^uint(0) {
   828  				print("runtime: max = ", max, ", npages = ", npages, "\n")
   829  				print("runtime: searchIdx = ", chunkPageIndex(p.searchAddr.addr()), ", p.searchAddr = ", hex(p.searchAddr.addr()), "\n")
   830  				throw("bad summary data")
   831  			}
   832  			addr = chunkBase(i) + uintptr(j)*pageSize
   833  			searchAddr = offAddr{chunkBase(i) + uintptr(searchIdx)*pageSize}
   834  			goto Found
   835  		}
   836  	}
   837  	// We failed to use a searchAddr for one reason or another, so try
   838  	// the slow path.
   839  	addr, searchAddr = p.find(npages)
   840  	if addr == 0 {
   841  		if npages == 1 {
   842  			// We failed to find a single free page, the smallest unit
   843  			// of allocation. This means we know the heap is completely
   844  			// exhausted. Otherwise, the heap still might have free
   845  			// space in it, just not enough contiguous space to
   846  			// accommodate npages.
   847  			p.searchAddr = maxSearchAddr()
   848  		}
   849  		return 0, 0
   850  	}
   851  Found:
   852  	// Go ahead and actually mark the bits now that we have an address.
   853  	scav = p.allocRange(addr, npages)
   854  
   855  	// If we found a higher searchAddr, we know that all the
   856  	// heap memory before that searchAddr in an offset address space is
   857  	// allocated, so bump p.searchAddr up to the new one.
   858  	if p.searchAddr.lessThan(searchAddr) {
   859  		p.searchAddr = searchAddr
   860  	}
   861  	return addr, scav
   862  }
   863  
   864  // free returns npages worth of memory starting at base back to the page heap.
   865  //
   866  // p.mheapLock must be held.
   867  //
   868  // Must run on the system stack because p.mheapLock must be held.
   869  //
   870  //go:systemstack
   871  func (p *pageAlloc) free(base, npages uintptr, scavenged bool) {
   872  	assertLockHeld(p.mheapLock)
   873  
   874  	// If we're freeing pages below the p.searchAddr, update searchAddr.
   875  	if b := (offAddr{base}); b.lessThan(p.searchAddr) {
   876  		p.searchAddr = b
   877  	}
   878  	limit := base + npages*pageSize - 1
   879  	if !scavenged {
   880  		p.scav.index.mark(base, limit+1)
   881  	}
   882  	if npages == 1 {
   883  		// Fast path: we're clearing a single bit, and we know exactly
   884  		// where it is, so mark it directly.
   885  		i := chunkIndex(base)
   886  		p.chunkOf(i).free1(chunkPageIndex(base))
   887  	} else {
   888  		// Slow path: we're clearing more bits so we may need to iterate.
   889  		sc, ec := chunkIndex(base), chunkIndex(limit)
   890  		si, ei := chunkPageIndex(base), chunkPageIndex(limit)
   891  
   892  		if sc == ec {
   893  			// The range doesn't cross any chunk boundaries.
   894  			p.chunkOf(sc).free(si, ei+1-si)
   895  		} else {
   896  			// The range crosses at least one chunk boundary.
   897  			p.chunkOf(sc).free(si, pallocChunkPages-si)
   898  			for c := sc + 1; c < ec; c++ {
   899  				p.chunkOf(c).freeAll()
   900  			}
   901  			p.chunkOf(ec).free(0, ei+1)
   902  		}
   903  	}
   904  	p.update(base, npages, true, false)
   905  }
   906  
   907  const (
   908  	pallocSumBytes = unsafe.Sizeof(pallocSum(0))
   909  
   910  	// maxPackedValue is the maximum value that any of the three fields in
   911  	// the pallocSum may take on.
   912  	maxPackedValue    = 1 << logMaxPackedValue
   913  	logMaxPackedValue = logPallocChunkPages + (summaryLevels-1)*summaryLevelBits
   914  
   915  	freeChunkSum = pallocSum(uint64(pallocChunkPages) |
   916  		uint64(pallocChunkPages<<logMaxPackedValue) |
   917  		uint64(pallocChunkPages<<(2*logMaxPackedValue)))
   918  )
   919  
   920  // pallocSum is a packed summary type which packs three numbers: start, max,
   921  // and end into a single 8-byte value. Each of these values are a summary of
   922  // a bitmap and are thus counts, each of which may have a maximum value of
   923  // 2^21 - 1, or all three may be equal to 2^21. The latter case is represented
   924  // by just setting the 64th bit.
   925  type pallocSum uint64
   926  
   927  // packPallocSum takes a start, max, and end value and produces a pallocSum.
   928  func packPallocSum(start, max, end uint) pallocSum {
   929  	if max == maxPackedValue {
   930  		return pallocSum(uint64(1 << 63))
   931  	}
   932  	return pallocSum((uint64(start) & (maxPackedValue - 1)) |
   933  		((uint64(max) & (maxPackedValue - 1)) << logMaxPackedValue) |
   934  		((uint64(end) & (maxPackedValue - 1)) << (2 * logMaxPackedValue)))
   935  }
   936  
   937  // start extracts the start value from a packed sum.
   938  func (p pallocSum) start() uint {
   939  	if uint64(p)&uint64(1<<63) != 0 {
   940  		return maxPackedValue
   941  	}
   942  	return uint(uint64(p) & (maxPackedValue - 1))
   943  }
   944  
   945  // max extracts the max value from a packed sum.
   946  func (p pallocSum) max() uint {
   947  	if uint64(p)&uint64(1<<63) != 0 {
   948  		return maxPackedValue
   949  	}
   950  	return uint((uint64(p) >> logMaxPackedValue) & (maxPackedValue - 1))
   951  }
   952  
   953  // end extracts the end value from a packed sum.
   954  func (p pallocSum) end() uint {
   955  	if uint64(p)&uint64(1<<63) != 0 {
   956  		return maxPackedValue
   957  	}
   958  	return uint((uint64(p) >> (2 * logMaxPackedValue)) & (maxPackedValue - 1))
   959  }
   960  
   961  // unpack unpacks all three values from the summary.
   962  func (p pallocSum) unpack() (uint, uint, uint) {
   963  	if uint64(p)&uint64(1<<63) != 0 {
   964  		return maxPackedValue, maxPackedValue, maxPackedValue
   965  	}
   966  	return uint(uint64(p) & (maxPackedValue - 1)),
   967  		uint((uint64(p) >> logMaxPackedValue) & (maxPackedValue - 1)),
   968  		uint((uint64(p) >> (2 * logMaxPackedValue)) & (maxPackedValue - 1))
   969  }
   970  
   971  // mergeSummaries merges consecutive summaries which may each represent at
   972  // most 1 << logMaxPagesPerSum pages each together into one.
   973  func mergeSummaries(sums []pallocSum, logMaxPagesPerSum uint) pallocSum {
   974  	// Merge the summaries in sums into one.
   975  	//
   976  	// We do this by keeping a running summary representing the merged
   977  	// summaries of sums[:i] in start, max, and end.
   978  	start, max, end := sums[0].unpack()
   979  	for i := 1; i < len(sums); i++ {
   980  		// Merge in sums[i].
   981  		si, mi, ei := sums[i].unpack()
   982  
   983  		// Merge in sums[i].start only if the running summary is
   984  		// completely free, otherwise this summary's start
   985  		// plays no role in the combined sum.
   986  		if start == uint(i)<<logMaxPagesPerSum {
   987  			start += si
   988  		}
   989  
   990  		// Recompute the max value of the running sum by looking
   991  		// across the boundary between the running sum and sums[i]
   992  		// and at the max sums[i], taking the greatest of those two
   993  		// and the max of the running sum.
   994  		if end+si > max {
   995  			max = end + si
   996  		}
   997  		if mi > max {
   998  			max = mi
   999  		}
  1000  
  1001  		// Merge in end by checking if this new summary is totally
  1002  		// free. If it is, then we want to extend the running sum's
  1003  		// end by the new summary. If not, then we have some alloc'd
  1004  		// pages in there and we just want to take the end value in
  1005  		// sums[i].
  1006  		if ei == 1<<logMaxPagesPerSum {
  1007  			end += 1 << logMaxPagesPerSum
  1008  		} else {
  1009  			end = ei
  1010  		}
  1011  	}
  1012  	return packPallocSum(start, max, end)
  1013  }
  1014  

View as plain text