1 // Copyright 2015 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 // Garbage collector: write barriers. 6 // 7 // For the concurrent garbage collector, the Go compiler implements 8 // updates to pointer-valued fields that may be in heap objects by 9 // emitting calls to write barriers. The main write barrier for 10 // individual pointer writes is gcWriteBarrier and is implemented in 11 // assembly. This file contains write barrier entry points for bulk 12 // operations. See also mwbbuf.go. 13 14 package runtime 15 16 import ( 17 "internal/abi" 18 "internal/goarch" 19 "unsafe" 20 ) 21 22 // Go uses a hybrid barrier that combines a Yuasa-style deletion 23 // barrier—which shades the object whose reference is being 24 // overwritten—with Dijkstra insertion barrier—which shades the object 25 // whose reference is being written. The insertion part of the barrier 26 // is necessary while the calling goroutine's stack is grey. In 27 // pseudocode, the barrier is: 28 // 29 // writePointer(slot, ptr): 30 // shade(*slot) 31 // if current stack is grey: 32 // shade(ptr) 33 // *slot = ptr 34 // 35 // slot is the destination in Go code. 36 // ptr is the value that goes into the slot in Go code. 37 // 38 // Shade indicates that it has seen a white pointer by adding the referent 39 // to wbuf as well as marking it. 40 // 41 // The two shades and the condition work together to prevent a mutator 42 // from hiding an object from the garbage collector: 43 // 44 // 1. shade(*slot) prevents a mutator from hiding an object by moving 45 // the sole pointer to it from the heap to its stack. If it attempts 46 // to unlink an object from the heap, this will shade it. 47 // 48 // 2. shade(ptr) prevents a mutator from hiding an object by moving 49 // the sole pointer to it from its stack into a black object in the 50 // heap. If it attempts to install the pointer into a black object, 51 // this will shade it. 52 // 53 // 3. Once a goroutine's stack is black, the shade(ptr) becomes 54 // unnecessary. shade(ptr) prevents hiding an object by moving it from 55 // the stack to the heap, but this requires first having a pointer 56 // hidden on the stack. Immediately after a stack is scanned, it only 57 // points to shaded objects, so it's not hiding anything, and the 58 // shade(*slot) prevents it from hiding any other pointers on its 59 // stack. 60 // 61 // For a detailed description of this barrier and proof of 62 // correctness, see https://github.com/golang/proposal/blob/master/design/17503-eliminate-rescan.md 63 // 64 // 65 // 66 // Dealing with memory ordering: 67 // 68 // Both the Yuasa and Dijkstra barriers can be made conditional on the 69 // color of the object containing the slot. We chose not to make these 70 // conditional because the cost of ensuring that the object holding 71 // the slot doesn't concurrently change color without the mutator 72 // noticing seems prohibitive. 73 // 74 // Consider the following example where the mutator writes into 75 // a slot and then loads the slot's mark bit while the GC thread 76 // writes to the slot's mark bit and then as part of scanning reads 77 // the slot. 78 // 79 // Initially both [slot] and [slotmark] are 0 (nil) 80 // Mutator thread GC thread 81 // st [slot], ptr st [slotmark], 1 82 // 83 // ld r1, [slotmark] ld r2, [slot] 84 // 85 // Without an expensive memory barrier between the st and the ld, the final 86 // result on most HW (including 386/amd64) can be r1==r2==0. This is a classic 87 // example of what can happen when loads are allowed to be reordered with older 88 // stores (avoiding such reorderings lies at the heart of the classic 89 // Peterson/Dekker algorithms for mutual exclusion). Rather than require memory 90 // barriers, which will slow down both the mutator and the GC, we always grey 91 // the ptr object regardless of the slot's color. 92 // 93 // Another place where we intentionally omit memory barriers is when 94 // accessing mheap_.arena_used to check if a pointer points into the 95 // heap. On relaxed memory machines, it's possible for a mutator to 96 // extend the size of the heap by updating arena_used, allocate an 97 // object from this new region, and publish a pointer to that object, 98 // but for tracing running on another processor to observe the pointer 99 // but use the old value of arena_used. In this case, tracing will not 100 // mark the object, even though it's reachable. However, the mutator 101 // is guaranteed to execute a write barrier when it publishes the 102 // pointer, so it will take care of marking the object. A general 103 // consequence of this is that the garbage collector may cache the 104 // value of mheap_.arena_used. (See issue #9984.) 105 // 106 // 107 // Stack writes: 108 // 109 // The compiler omits write barriers for writes to the current frame, 110 // but if a stack pointer has been passed down the call stack, the 111 // compiler will generate a write barrier for writes through that 112 // pointer (because it doesn't know it's not a heap pointer). 113 // 114 // One might be tempted to ignore the write barrier if slot points 115 // into to the stack. Don't do it! Mark termination only re-scans 116 // frames that have potentially been active since the concurrent scan, 117 // so it depends on write barriers to track changes to pointers in 118 // stack frames that have not been active. 119 // 120 // 121 // Global writes: 122 // 123 // The Go garbage collector requires write barriers when heap pointers 124 // are stored in globals. Many garbage collectors ignore writes to 125 // globals and instead pick up global -> heap pointers during 126 // termination. This increases pause time, so we instead rely on write 127 // barriers for writes to globals so that we don't have to rescan 128 // global during mark termination. 129 // 130 // 131 // Publication ordering: 132 // 133 // The write barrier is *pre-publication*, meaning that the write 134 // barrier happens prior to the *slot = ptr write that may make ptr 135 // reachable by some goroutine that currently cannot reach it. 136 // 137 // 138 // Signal handler pointer writes: 139 // 140 // In general, the signal handler cannot safely invoke the write 141 // barrier because it may run without a P or even during the write 142 // barrier. 143 // 144 // There is exactly one exception: profbuf.go omits a barrier during 145 // signal handler profile logging. That's safe only because of the 146 // deletion barrier. See profbuf.go for a detailed argument. If we 147 // remove the deletion barrier, we'll have to work out a new way to 148 // handle the profile logging. 149 150 // typedmemmove copies a value of type typ to dst from src. 151 // Must be nosplit, see #16026. 152 // 153 // TODO: Perfect for go:nosplitrec since we can't have a safe point 154 // anywhere in the bulk barrier or memmove. 155 // 156 //go:nosplit 157 func typedmemmove(typ *_type, dst, src unsafe.Pointer) { 158 if dst == src { 159 return 160 } 161 if writeBarrier.needed && typ.ptrdata != 0 { 162 bulkBarrierPreWrite(uintptr(dst), uintptr(src), typ.ptrdata) 163 } 164 // There's a race here: if some other goroutine can write to 165 // src, it may change some pointer in src after we've 166 // performed the write barrier but before we perform the 167 // memory copy. This safe because the write performed by that 168 // other goroutine must also be accompanied by a write 169 // barrier, so at worst we've unnecessarily greyed the old 170 // pointer that was in src. 171 memmove(dst, src, typ.size) 172 if writeBarrier.cgo { 173 cgoCheckMemmove(typ, dst, src, 0, typ.size) 174 } 175 } 176 177 //go:linkname reflect_typedmemmove reflect.typedmemmove 178 func reflect_typedmemmove(typ *_type, dst, src unsafe.Pointer) { 179 if raceenabled { 180 raceWriteObjectPC(typ, dst, getcallerpc(), abi.FuncPCABIInternal(reflect_typedmemmove)) 181 raceReadObjectPC(typ, src, getcallerpc(), abi.FuncPCABIInternal(reflect_typedmemmove)) 182 } 183 if msanenabled { 184 msanwrite(dst, typ.size) 185 msanread(src, typ.size) 186 } 187 if asanenabled { 188 asanwrite(dst, typ.size) 189 asanread(src, typ.size) 190 } 191 typedmemmove(typ, dst, src) 192 } 193 194 //go:linkname reflectlite_typedmemmove internal/reflectlite.typedmemmove 195 func reflectlite_typedmemmove(typ *_type, dst, src unsafe.Pointer) { 196 reflect_typedmemmove(typ, dst, src) 197 } 198 199 // typedmemmovepartial is like typedmemmove but assumes that 200 // dst and src point off bytes into the value and only copies size bytes. 201 // off must be a multiple of goarch.PtrSize. 202 // 203 //go:linkname reflect_typedmemmovepartial reflect.typedmemmovepartial 204 func reflect_typedmemmovepartial(typ *_type, dst, src unsafe.Pointer, off, size uintptr) { 205 if writeBarrier.needed && typ.ptrdata > off && size >= goarch.PtrSize { 206 if off&(goarch.PtrSize-1) != 0 { 207 panic("reflect: internal error: misaligned offset") 208 } 209 pwsize := alignDown(size, goarch.PtrSize) 210 if poff := typ.ptrdata - off; pwsize > poff { 211 pwsize = poff 212 } 213 bulkBarrierPreWrite(uintptr(dst), uintptr(src), pwsize) 214 } 215 216 memmove(dst, src, size) 217 if writeBarrier.cgo { 218 cgoCheckMemmove(typ, dst, src, off, size) 219 } 220 } 221 222 // reflectcallmove is invoked by reflectcall to copy the return values 223 // out of the stack and into the heap, invoking the necessary write 224 // barriers. dst, src, and size describe the return value area to 225 // copy. typ describes the entire frame (not just the return values). 226 // typ may be nil, which indicates write barriers are not needed. 227 // 228 // It must be nosplit and must only call nosplit functions because the 229 // stack map of reflectcall is wrong. 230 // 231 //go:nosplit 232 func reflectcallmove(typ *_type, dst, src unsafe.Pointer, size uintptr, regs *abi.RegArgs) { 233 if writeBarrier.needed && typ != nil && typ.ptrdata != 0 && size >= goarch.PtrSize { 234 bulkBarrierPreWrite(uintptr(dst), uintptr(src), size) 235 } 236 memmove(dst, src, size) 237 238 // Move pointers returned in registers to a place where the GC can see them. 239 for i := range regs.Ints { 240 if regs.ReturnIsPtr.Get(i) { 241 regs.Ptrs[i] = unsafe.Pointer(regs.Ints[i]) 242 } 243 } 244 } 245 246 //go:nosplit 247 func typedslicecopy(typ *_type, dstPtr unsafe.Pointer, dstLen int, srcPtr unsafe.Pointer, srcLen int) int { 248 n := dstLen 249 if n > srcLen { 250 n = srcLen 251 } 252 if n == 0 { 253 return 0 254 } 255 256 // The compiler emits calls to typedslicecopy before 257 // instrumentation runs, so unlike the other copying and 258 // assignment operations, it's not instrumented in the calling 259 // code and needs its own instrumentation. 260 if raceenabled { 261 callerpc := getcallerpc() 262 pc := abi.FuncPCABIInternal(slicecopy) 263 racewriterangepc(dstPtr, uintptr(n)*typ.size, callerpc, pc) 264 racereadrangepc(srcPtr, uintptr(n)*typ.size, callerpc, pc) 265 } 266 if msanenabled { 267 msanwrite(dstPtr, uintptr(n)*typ.size) 268 msanread(srcPtr, uintptr(n)*typ.size) 269 } 270 if asanenabled { 271 asanwrite(dstPtr, uintptr(n)*typ.size) 272 asanread(srcPtr, uintptr(n)*typ.size) 273 } 274 275 if writeBarrier.cgo { 276 cgoCheckSliceCopy(typ, dstPtr, srcPtr, n) 277 } 278 279 if dstPtr == srcPtr { 280 return n 281 } 282 283 // Note: No point in checking typ.ptrdata here: 284 // compiler only emits calls to typedslicecopy for types with pointers, 285 // and growslice and reflect_typedslicecopy check for pointers 286 // before calling typedslicecopy. 287 size := uintptr(n) * typ.size 288 if writeBarrier.needed { 289 pwsize := size - typ.size + typ.ptrdata 290 bulkBarrierPreWrite(uintptr(dstPtr), uintptr(srcPtr), pwsize) 291 } 292 // See typedmemmove for a discussion of the race between the 293 // barrier and memmove. 294 memmove(dstPtr, srcPtr, size) 295 return n 296 } 297 298 //go:linkname reflect_typedslicecopy reflect.typedslicecopy 299 func reflect_typedslicecopy(elemType *_type, dst, src slice) int { 300 if elemType.ptrdata == 0 { 301 return slicecopy(dst.array, dst.len, src.array, src.len, elemType.size) 302 } 303 return typedslicecopy(elemType, dst.array, dst.len, src.array, src.len) 304 } 305 306 // typedmemclr clears the typed memory at ptr with type typ. The 307 // memory at ptr must already be initialized (and hence in type-safe 308 // state). If the memory is being initialized for the first time, see 309 // memclrNoHeapPointers. 310 // 311 // If the caller knows that typ has pointers, it can alternatively 312 // call memclrHasPointers. 313 // 314 //go:nosplit 315 func typedmemclr(typ *_type, ptr unsafe.Pointer) { 316 if writeBarrier.needed && typ.ptrdata != 0 { 317 bulkBarrierPreWrite(uintptr(ptr), 0, typ.ptrdata) 318 } 319 memclrNoHeapPointers(ptr, typ.size) 320 } 321 322 //go:linkname reflect_typedmemclr reflect.typedmemclr 323 func reflect_typedmemclr(typ *_type, ptr unsafe.Pointer) { 324 typedmemclr(typ, ptr) 325 } 326 327 //go:linkname reflect_typedmemclrpartial reflect.typedmemclrpartial 328 func reflect_typedmemclrpartial(typ *_type, ptr unsafe.Pointer, off, size uintptr) { 329 if writeBarrier.needed && typ.ptrdata != 0 { 330 bulkBarrierPreWrite(uintptr(ptr), 0, size) 331 } 332 memclrNoHeapPointers(ptr, size) 333 } 334 335 // memclrHasPointers clears n bytes of typed memory starting at ptr. 336 // The caller must ensure that the type of the object at ptr has 337 // pointers, usually by checking typ.ptrdata. However, ptr 338 // does not have to point to the start of the allocation. 339 // 340 //go:nosplit 341 func memclrHasPointers(ptr unsafe.Pointer, n uintptr) { 342 bulkBarrierPreWrite(uintptr(ptr), 0, n) 343 memclrNoHeapPointers(ptr, n) 344 } 345