mcache.go

     1  // Copyright 2009 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  package runtime
     6  
     7  import (
     8  	"internal/runtime/atomic"
     9  	"runtime/internal/sys"
    10  	"unsafe"
    11  )
    12  
    13  // Per-thread (in Go, per-P) cache for small objects.
    14  // This includes a small object cache and local allocation stats.
    15  // No locking needed because it is per-thread (per-P).
    16  //
    17  // mcaches are allocated from non-GC'd memory, so any heap pointers
    18  // must be specially handled.
    19  type mcache struct {
    20  	_ sys.NotInHeap
    21  
    22  	// The following members are accessed on every malloc,
    23  	// so they are grouped here for better caching.
    24  	nextSample uintptr // trigger heap sample after allocating this many bytes
    25  	scanAlloc  uintptr // bytes of scannable heap allocated
    26  
    27  	// Allocator cache for tiny objects w/o pointers.
    28  	// See "Tiny allocator" comment in malloc.go.
    29  
    30  	// tiny points to the beginning of the current tiny block, or
    31  	// nil if there is no current tiny block.
    32  	//
    33  	// tiny is a heap pointer. Since mcache is in non-GC'd memory,
    34  	// we handle it by clearing it in releaseAll during mark
    35  	// termination.
    36  	//
    37  	// tinyAllocs is the number of tiny allocations performed
    38  	// by the P that owns this mcache.
    39  	tiny       uintptr
    40  	tinyoffset uintptr
    41  	tinyAllocs uintptr
    42  
    43  	// The rest is not accessed on every malloc.
    44  
    45  	alloc [numSpanClasses]*mspan // spans to allocate from, indexed by spanClass
    46  
    47  	stackcache [_NumStackOrders]stackfreelist
    48  
    49  	// flushGen indicates the sweepgen during which this mcache
    50  	// was last flushed. If flushGen != mheap_.sweepgen, the spans
    51  	// in this mcache are stale and need to the flushed so they
    52  	// can be swept. This is done in acquirep.
    53  	flushGen atomic.Uint32
    54  }
    55  
    56  // A gclink is a node in a linked list of blocks, like mlink,
    57  // but it is opaque to the garbage collector.
    58  // The GC does not trace the pointers during collection,
    59  // and the compiler does not emit write barriers for assignments
    60  // of gclinkptr values. Code should store references to gclinks
    61  // as gclinkptr, not as *gclink.
    62  type gclink struct {
    63  	next gclinkptr
    64  }
    65  
    66  // A gclinkptr is a pointer to a gclink, but it is opaque
    67  // to the garbage collector.
    68  type gclinkptr uintptr
    69  
    70  // ptr returns the *gclink form of p.
    71  // The result should be used for accessing fields, not stored
    72  // in other data structures.
    73  func (p gclinkptr) ptr() *gclink {
    74  	return (*gclink)(unsafe.Pointer(p))
    75  }
    76  
    77  type stackfreelist struct {
    78  	list gclinkptr // linked list of free stacks
    79  	size uintptr   // total size of stacks in list
    80  }
    81  
    82  // dummy mspan that contains no free objects.
    83  var emptymspan mspan
    84  
    85  func allocmcache() *mcache {
    86  	var c *mcache
    87  	systemstack(func() {
    88  		lock(&mheap_.lock)
    89  		c = (*mcache)(mheap_.cachealloc.alloc())
    90  		c.flushGen.Store(mheap_.sweepgen)
    91  		unlock(&mheap_.lock)
    92  	})
    93  	for i := range c.alloc {
    94  		c.alloc[i] = &emptymspan
    95  	}
    96  	c.nextSample = nextSample()
    97  	return c
    98  }
    99  
   100  // freemcache releases resources associated with this
   101  // mcache and puts the object onto a free list.
   102  //
   103  // In some cases there is no way to simply release
   104  // resources, such as statistics, so donate them to
   105  // a different mcache (the recipient).
   106  func freemcache(c *mcache) {
   107  	systemstack(func() {
   108  		c.releaseAll()
   109  		stackcache_clear(c)
   110  
   111  		// NOTE(rsc,rlh): If gcworkbuffree comes back, we need to coordinate
   112  		// with the stealing of gcworkbufs during garbage collection to avoid
   113  		// a race where the workbuf is double-freed.
   114  		// gcworkbuffree(c.gcworkbuf)
   115  
   116  		lock(&mheap_.lock)
   117  		mheap_.cachealloc.free(unsafe.Pointer(c))
   118  		unlock(&mheap_.lock)
   119  	})
   120  }
   121  
   122  // getMCache is a convenience function which tries to obtain an mcache.
   123  //
   124  // Returns nil if we're not bootstrapping or we don't have a P. The caller's
   125  // P must not change, so we must be in a non-preemptible state.
   126  func getMCache(mp *m) *mcache {
   127  	// Grab the mcache, since that's where stats live.
   128  	pp := mp.p.ptr()
   129  	var c *mcache
   130  	if pp == nil {
   131  		// We will be called without a P while bootstrapping,
   132  		// in which case we use mcache0, which is set in mallocinit.
   133  		// mcache0 is cleared when bootstrapping is complete,
   134  		// by procresize.
   135  		c = mcache0
   136  	} else {
   137  		c = pp.mcache
   138  	}
   139  	return c
   140  }
   141  
   142  // refill acquires a new span of span class spc for c. This span will
   143  // have at least one free object. The current span in c must be full.
   144  //
   145  // Must run in a non-preemptible context since otherwise the owner of
   146  // c could change.
   147  func (c *mcache) refill(spc spanClass) {
   148  	// Return the current cached span to the central lists.
   149  	s := c.alloc[spc]
   150  
   151  	if s.allocCount != s.nelems {
   152  		throw("refill of span with free space remaining")
   153  	}
   154  	if s != &emptymspan {
   155  		// Mark this span as no longer cached.
   156  		if s.sweepgen != mheap_.sweepgen+3 {
   157  			throw("bad sweepgen in refill")
   158  		}
   159  		mheap_.central[spc].mcentral.uncacheSpan(s)
   160  
   161  		// Count up how many slots were used and record it.
   162  		stats := memstats.heapStats.acquire()
   163  		slotsUsed := int64(s.allocCount) - int64(s.allocCountBeforeCache)
   164  		atomic.Xadd64(&stats.smallAllocCount[spc.sizeclass()], slotsUsed)
   165  
   166  		// Flush tinyAllocs.
   167  		if spc == tinySpanClass {
   168  			atomic.Xadd64(&stats.tinyAllocCount, int64(c.tinyAllocs))
   169  			c.tinyAllocs = 0
   170  		}
   171  		memstats.heapStats.release()
   172  
   173  		// Count the allocs in inconsistent, internal stats.
   174  		bytesAllocated := slotsUsed * int64(s.elemsize)
   175  		gcController.totalAlloc.Add(bytesAllocated)
   176  
   177  		// Clear the second allocCount just to be safe.
   178  		s.allocCountBeforeCache = 0
   179  	}
   180  
   181  	// Get a new cached span from the central lists.
   182  	s = mheap_.central[spc].mcentral.cacheSpan()
   183  	if s == nil {
   184  		throw("out of memory")
   185  	}
   186  
   187  	if s.allocCount == s.nelems {
   188  		throw("span has no free space")
   189  	}
   190  
   191  	// Indicate that this span is cached and prevent asynchronous
   192  	// sweeping in the next sweep phase.
   193  	s.sweepgen = mheap_.sweepgen + 3
   194  
   195  	// Store the current alloc count for accounting later.
   196  	s.allocCountBeforeCache = s.allocCount
   197  
   198  	// Update heapLive and flush scanAlloc.
   199  	//
   200  	// We have not yet allocated anything new into the span, but we
   201  	// assume that all of its slots will get used, so this makes
   202  	// heapLive an overestimate.
   203  	//
   204  	// When the span gets uncached, we'll fix up this overestimate
   205  	// if necessary (see releaseAll).
   206  	//
   207  	// We pick an overestimate here because an underestimate leads
   208  	// the pacer to believe that it's in better shape than it is,
   209  	// which appears to lead to more memory used. See #53738 for
   210  	// more details.
   211  	usedBytes := uintptr(s.allocCount) * s.elemsize
   212  	gcController.update(int64(s.npages*pageSize)-int64(usedBytes), int64(c.scanAlloc))
   213  	c.scanAlloc = 0
   214  
   215  	c.alloc[spc] = s
   216  }
   217  
   218  // allocLarge allocates a span for a large object.
   219  func (c *mcache) allocLarge(size uintptr, noscan bool) *mspan {
   220  	if size+_PageSize < size {
   221  		throw("out of memory")
   222  	}
   223  	npages := size >> _PageShift
   224  	if size&_PageMask != 0 {
   225  		npages++
   226  	}
   227  
   228  	// Deduct credit for this span allocation and sweep if
   229  	// necessary. mHeap_Alloc will also sweep npages, so this only
   230  	// pays the debt down to npage pages.
   231  	deductSweepCredit(npages*_PageSize, npages)
   232  
   233  	spc := makeSpanClass(0, noscan)
   234  	s := mheap_.alloc(npages, spc)
   235  	if s == nil {
   236  		throw("out of memory")
   237  	}
   238  
   239  	// Count the alloc in consistent, external stats.
   240  	stats := memstats.heapStats.acquire()
   241  	atomic.Xadd64(&stats.largeAlloc, int64(npages*pageSize))
   242  	atomic.Xadd64(&stats.largeAllocCount, 1)
   243  	memstats.heapStats.release()
   244  
   245  	// Count the alloc in inconsistent, internal stats.
   246  	gcController.totalAlloc.Add(int64(npages * pageSize))
   247  
   248  	// Update heapLive.
   249  	gcController.update(int64(s.npages*pageSize), 0)
   250  
   251  	// Put the large span in the mcentral swept list so that it's
   252  	// visible to the background sweeper.
   253  	mheap_.central[spc].mcentral.fullSwept(mheap_.sweepgen).push(s)
   254  	s.limit = s.base() + size
   255  	s.initHeapBits(false)
   256  	return s
   257  }
   258  
   259  func (c *mcache) releaseAll() {
   260  	// Take this opportunity to flush scanAlloc.
   261  	scanAlloc := int64(c.scanAlloc)
   262  	c.scanAlloc = 0
   263  
   264  	sg := mheap_.sweepgen
   265  	dHeapLive := int64(0)
   266  	for i := range c.alloc {
   267  		s := c.alloc[i]
   268  		if s != &emptymspan {
   269  			slotsUsed := int64(s.allocCount) - int64(s.allocCountBeforeCache)
   270  			s.allocCountBeforeCache = 0
   271  
   272  			// Adjust smallAllocCount for whatever was allocated.
   273  			stats := memstats.heapStats.acquire()
   274  			atomic.Xadd64(&stats.smallAllocCount[spanClass(i).sizeclass()], slotsUsed)
   275  			memstats.heapStats.release()
   276  
   277  			// Adjust the actual allocs in inconsistent, internal stats.
   278  			// We assumed earlier that the full span gets allocated.
   279  			gcController.totalAlloc.Add(slotsUsed * int64(s.elemsize))
   280  
   281  			if s.sweepgen != sg+1 {
   282  				// refill conservatively counted unallocated slots in gcController.heapLive.
   283  				// Undo this.
   284  				//
   285  				// If this span was cached before sweep, then gcController.heapLive was totally
   286  				// recomputed since caching this span, so we don't do this for stale spans.
   287  				dHeapLive -= int64(s.nelems-s.allocCount) * int64(s.elemsize)
   288  			}
   289  
   290  			// Release the span to the mcentral.
   291  			mheap_.central[i].mcentral.uncacheSpan(s)
   292  			c.alloc[i] = &emptymspan
   293  		}
   294  	}
   295  	// Clear tinyalloc pool.
   296  	c.tiny = 0
   297  	c.tinyoffset = 0
   298  
   299  	// Flush tinyAllocs.
   300  	stats := memstats.heapStats.acquire()
   301  	atomic.Xadd64(&stats.tinyAllocCount, int64(c.tinyAllocs))
   302  	c.tinyAllocs = 0
   303  	memstats.heapStats.release()
   304  
   305  	// Update heapLive and heapScan.
   306  	gcController.update(dHeapLive, scanAlloc)
   307  }
   308  
   309  // prepareForSweep flushes c if the system has entered a new sweep phase
   310  // since c was populated. This must happen between the sweep phase
   311  // starting and the first allocation from c.
   312  func (c *mcache) prepareForSweep() {
   313  	// Alternatively, instead of making sure we do this on every P
   314  	// between starting the world and allocating on that P, we
   315  	// could leave allocate-black on, allow allocation to continue
   316  	// as usual, use a ragged barrier at the beginning of sweep to
   317  	// ensure all cached spans are swept, and then disable
   318  	// allocate-black. However, with this approach it's difficult
   319  	// to avoid spilling mark bits into the *next* GC cycle.
   320  	sg := mheap_.sweepgen
   321  	flushGen := c.flushGen.Load()
   322  	if flushGen == sg {
   323  		return
   324  	} else if flushGen != sg-2 {
   325  		println("bad flushGen", flushGen, "in prepareForSweep; sweepgen", sg)
   326  		throw("bad flushGen")
   327  	}
   328  	c.releaseAll()
   329  	stackcache_clear(c)
   330  	c.flushGen.Store(mheap_.sweepgen) // Synchronizes with gcStart
   331  }
   332
View as plain text