Scott Baker | 1fe7273 | 2019-10-21 10:58:51 -0700 | [diff] [blame] | 1 | // Copyright 2016 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 http2 |
| 6 | |
| 7 | import ( |
| 8 | "fmt" |
| 9 | "math" |
| 10 | "sort" |
| 11 | ) |
| 12 | |
| 13 | // RFC 7540, Section 5.3.5: the default weight is 16. |
| 14 | const priorityDefaultWeight = 15 // 16 = 15 + 1 |
| 15 | |
| 16 | // PriorityWriteSchedulerConfig configures a priorityWriteScheduler. |
| 17 | type PriorityWriteSchedulerConfig struct { |
| 18 | // MaxClosedNodesInTree controls the maximum number of closed streams to |
| 19 | // retain in the priority tree. Setting this to zero saves a small amount |
| 20 | // of memory at the cost of performance. |
| 21 | // |
| 22 | // See RFC 7540, Section 5.3.4: |
| 23 | // "It is possible for a stream to become closed while prioritization |
| 24 | // information ... is in transit. ... This potentially creates suboptimal |
| 25 | // prioritization, since the stream could be given a priority that is |
| 26 | // different from what is intended. To avoid these problems, an endpoint |
| 27 | // SHOULD retain stream prioritization state for a period after streams |
| 28 | // become closed. The longer state is retained, the lower the chance that |
| 29 | // streams are assigned incorrect or default priority values." |
| 30 | MaxClosedNodesInTree int |
| 31 | |
| 32 | // MaxIdleNodesInTree controls the maximum number of idle streams to |
| 33 | // retain in the priority tree. Setting this to zero saves a small amount |
| 34 | // of memory at the cost of performance. |
| 35 | // |
| 36 | // See RFC 7540, Section 5.3.4: |
| 37 | // Similarly, streams that are in the "idle" state can be assigned |
| 38 | // priority or become a parent of other streams. This allows for the |
| 39 | // creation of a grouping node in the dependency tree, which enables |
| 40 | // more flexible expressions of priority. Idle streams begin with a |
| 41 | // default priority (Section 5.3.5). |
| 42 | MaxIdleNodesInTree int |
| 43 | |
| 44 | // ThrottleOutOfOrderWrites enables write throttling to help ensure that |
| 45 | // data is delivered in priority order. This works around a race where |
| 46 | // stream B depends on stream A and both streams are about to call Write |
| 47 | // to queue DATA frames. If B wins the race, a naive scheduler would eagerly |
| 48 | // write as much data from B as possible, but this is suboptimal because A |
| 49 | // is a higher-priority stream. With throttling enabled, we write a small |
| 50 | // amount of data from B to minimize the amount of bandwidth that B can |
| 51 | // steal from A. |
| 52 | ThrottleOutOfOrderWrites bool |
| 53 | } |
| 54 | |
| 55 | // NewPriorityWriteScheduler constructs a WriteScheduler that schedules |
| 56 | // frames by following HTTP/2 priorities as described in RFC 7540 Section 5.3. |
| 57 | // If cfg is nil, default options are used. |
| 58 | func NewPriorityWriteScheduler(cfg *PriorityWriteSchedulerConfig) WriteScheduler { |
| 59 | if cfg == nil { |
| 60 | // For justification of these defaults, see: |
| 61 | // https://docs.google.com/document/d/1oLhNg1skaWD4_DtaoCxdSRN5erEXrH-KnLrMwEpOtFY |
| 62 | cfg = &PriorityWriteSchedulerConfig{ |
| 63 | MaxClosedNodesInTree: 10, |
| 64 | MaxIdleNodesInTree: 10, |
| 65 | ThrottleOutOfOrderWrites: false, |
| 66 | } |
| 67 | } |
| 68 | |
| 69 | ws := &priorityWriteScheduler{ |
| 70 | nodes: make(map[uint32]*priorityNode), |
| 71 | maxClosedNodesInTree: cfg.MaxClosedNodesInTree, |
| 72 | maxIdleNodesInTree: cfg.MaxIdleNodesInTree, |
| 73 | enableWriteThrottle: cfg.ThrottleOutOfOrderWrites, |
| 74 | } |
| 75 | ws.nodes[0] = &ws.root |
| 76 | if cfg.ThrottleOutOfOrderWrites { |
| 77 | ws.writeThrottleLimit = 1024 |
| 78 | } else { |
| 79 | ws.writeThrottleLimit = math.MaxInt32 |
| 80 | } |
| 81 | return ws |
| 82 | } |
| 83 | |
| 84 | type priorityNodeState int |
| 85 | |
| 86 | const ( |
| 87 | priorityNodeOpen priorityNodeState = iota |
| 88 | priorityNodeClosed |
| 89 | priorityNodeIdle |
| 90 | ) |
| 91 | |
| 92 | // priorityNode is a node in an HTTP/2 priority tree. |
| 93 | // Each node is associated with a single stream ID. |
| 94 | // See RFC 7540, Section 5.3. |
| 95 | type priorityNode struct { |
| 96 | q writeQueue // queue of pending frames to write |
| 97 | id uint32 // id of the stream, or 0 for the root of the tree |
| 98 | weight uint8 // the actual weight is weight+1, so the value is in [1,256] |
| 99 | state priorityNodeState // open | closed | idle |
| 100 | bytes int64 // number of bytes written by this node, or 0 if closed |
| 101 | subtreeBytes int64 // sum(node.bytes) of all nodes in this subtree |
| 102 | |
| 103 | // These links form the priority tree. |
| 104 | parent *priorityNode |
| 105 | kids *priorityNode // start of the kids list |
| 106 | prev, next *priorityNode // doubly-linked list of siblings |
| 107 | } |
| 108 | |
| 109 | func (n *priorityNode) setParent(parent *priorityNode) { |
| 110 | if n == parent { |
| 111 | panic("setParent to self") |
| 112 | } |
| 113 | if n.parent == parent { |
| 114 | return |
| 115 | } |
| 116 | // Unlink from current parent. |
| 117 | if parent := n.parent; parent != nil { |
| 118 | if n.prev == nil { |
| 119 | parent.kids = n.next |
| 120 | } else { |
| 121 | n.prev.next = n.next |
| 122 | } |
| 123 | if n.next != nil { |
| 124 | n.next.prev = n.prev |
| 125 | } |
| 126 | } |
| 127 | // Link to new parent. |
| 128 | // If parent=nil, remove n from the tree. |
| 129 | // Always insert at the head of parent.kids (this is assumed by walkReadyInOrder). |
| 130 | n.parent = parent |
| 131 | if parent == nil { |
| 132 | n.next = nil |
| 133 | n.prev = nil |
| 134 | } else { |
| 135 | n.next = parent.kids |
| 136 | n.prev = nil |
| 137 | if n.next != nil { |
| 138 | n.next.prev = n |
| 139 | } |
| 140 | parent.kids = n |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | func (n *priorityNode) addBytes(b int64) { |
| 145 | n.bytes += b |
| 146 | for ; n != nil; n = n.parent { |
| 147 | n.subtreeBytes += b |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | // walkReadyInOrder iterates over the tree in priority order, calling f for each node |
| 152 | // with a non-empty write queue. When f returns true, this function returns true and the |
| 153 | // walk halts. tmp is used as scratch space for sorting. |
| 154 | // |
| 155 | // f(n, openParent) takes two arguments: the node to visit, n, and a bool that is true |
| 156 | // if any ancestor p of n is still open (ignoring the root node). |
| 157 | func (n *priorityNode) walkReadyInOrder(openParent bool, tmp *[]*priorityNode, f func(*priorityNode, bool) bool) bool { |
| 158 | if !n.q.empty() && f(n, openParent) { |
| 159 | return true |
| 160 | } |
| 161 | if n.kids == nil { |
| 162 | return false |
| 163 | } |
| 164 | |
| 165 | // Don't consider the root "open" when updating openParent since |
| 166 | // we can't send data frames on the root stream (only control frames). |
| 167 | if n.id != 0 { |
| 168 | openParent = openParent || (n.state == priorityNodeOpen) |
| 169 | } |
| 170 | |
| 171 | // Common case: only one kid or all kids have the same weight. |
| 172 | // Some clients don't use weights; other clients (like web browsers) |
| 173 | // use mostly-linear priority trees. |
| 174 | w := n.kids.weight |
| 175 | needSort := false |
| 176 | for k := n.kids.next; k != nil; k = k.next { |
| 177 | if k.weight != w { |
| 178 | needSort = true |
| 179 | break |
| 180 | } |
| 181 | } |
| 182 | if !needSort { |
| 183 | for k := n.kids; k != nil; k = k.next { |
| 184 | if k.walkReadyInOrder(openParent, tmp, f) { |
| 185 | return true |
| 186 | } |
| 187 | } |
| 188 | return false |
| 189 | } |
| 190 | |
| 191 | // Uncommon case: sort the child nodes. We remove the kids from the parent, |
| 192 | // then re-insert after sorting so we can reuse tmp for future sort calls. |
| 193 | *tmp = (*tmp)[:0] |
| 194 | for n.kids != nil { |
| 195 | *tmp = append(*tmp, n.kids) |
| 196 | n.kids.setParent(nil) |
| 197 | } |
| 198 | sort.Sort(sortPriorityNodeSiblings(*tmp)) |
| 199 | for i := len(*tmp) - 1; i >= 0; i-- { |
| 200 | (*tmp)[i].setParent(n) // setParent inserts at the head of n.kids |
| 201 | } |
| 202 | for k := n.kids; k != nil; k = k.next { |
| 203 | if k.walkReadyInOrder(openParent, tmp, f) { |
| 204 | return true |
| 205 | } |
| 206 | } |
| 207 | return false |
| 208 | } |
| 209 | |
| 210 | type sortPriorityNodeSiblings []*priorityNode |
| 211 | |
| 212 | func (z sortPriorityNodeSiblings) Len() int { return len(z) } |
| 213 | func (z sortPriorityNodeSiblings) Swap(i, k int) { z[i], z[k] = z[k], z[i] } |
| 214 | func (z sortPriorityNodeSiblings) Less(i, k int) bool { |
| 215 | // Prefer the subtree that has sent fewer bytes relative to its weight. |
| 216 | // See sections 5.3.2 and 5.3.4. |
| 217 | wi, bi := float64(z[i].weight+1), float64(z[i].subtreeBytes) |
| 218 | wk, bk := float64(z[k].weight+1), float64(z[k].subtreeBytes) |
| 219 | if bi == 0 && bk == 0 { |
| 220 | return wi >= wk |
| 221 | } |
| 222 | if bk == 0 { |
| 223 | return false |
| 224 | } |
| 225 | return bi/bk <= wi/wk |
| 226 | } |
| 227 | |
| 228 | type priorityWriteScheduler struct { |
| 229 | // root is the root of the priority tree, where root.id = 0. |
| 230 | // The root queues control frames that are not associated with any stream. |
| 231 | root priorityNode |
| 232 | |
| 233 | // nodes maps stream ids to priority tree nodes. |
| 234 | nodes map[uint32]*priorityNode |
| 235 | |
| 236 | // maxID is the maximum stream id in nodes. |
| 237 | maxID uint32 |
| 238 | |
| 239 | // lists of nodes that have been closed or are idle, but are kept in |
| 240 | // the tree for improved prioritization. When the lengths exceed either |
| 241 | // maxClosedNodesInTree or maxIdleNodesInTree, old nodes are discarded. |
| 242 | closedNodes, idleNodes []*priorityNode |
| 243 | |
| 244 | // From the config. |
| 245 | maxClosedNodesInTree int |
| 246 | maxIdleNodesInTree int |
| 247 | writeThrottleLimit int32 |
| 248 | enableWriteThrottle bool |
| 249 | |
| 250 | // tmp is scratch space for priorityNode.walkReadyInOrder to reduce allocations. |
| 251 | tmp []*priorityNode |
| 252 | |
| 253 | // pool of empty queues for reuse. |
| 254 | queuePool writeQueuePool |
| 255 | } |
| 256 | |
| 257 | func (ws *priorityWriteScheduler) OpenStream(streamID uint32, options OpenStreamOptions) { |
| 258 | // The stream may be currently idle but cannot be opened or closed. |
| 259 | if curr := ws.nodes[streamID]; curr != nil { |
| 260 | if curr.state != priorityNodeIdle { |
| 261 | panic(fmt.Sprintf("stream %d already opened", streamID)) |
| 262 | } |
| 263 | curr.state = priorityNodeOpen |
| 264 | return |
| 265 | } |
| 266 | |
| 267 | // RFC 7540, Section 5.3.5: |
| 268 | // "All streams are initially assigned a non-exclusive dependency on stream 0x0. |
| 269 | // Pushed streams initially depend on their associated stream. In both cases, |
| 270 | // streams are assigned a default weight of 16." |
| 271 | parent := ws.nodes[options.PusherID] |
| 272 | if parent == nil { |
| 273 | parent = &ws.root |
| 274 | } |
| 275 | n := &priorityNode{ |
| 276 | q: *ws.queuePool.get(), |
| 277 | id: streamID, |
| 278 | weight: priorityDefaultWeight, |
| 279 | state: priorityNodeOpen, |
| 280 | } |
| 281 | n.setParent(parent) |
| 282 | ws.nodes[streamID] = n |
| 283 | if streamID > ws.maxID { |
| 284 | ws.maxID = streamID |
| 285 | } |
| 286 | } |
| 287 | |
| 288 | func (ws *priorityWriteScheduler) CloseStream(streamID uint32) { |
| 289 | if streamID == 0 { |
| 290 | panic("violation of WriteScheduler interface: cannot close stream 0") |
| 291 | } |
| 292 | if ws.nodes[streamID] == nil { |
| 293 | panic(fmt.Sprintf("violation of WriteScheduler interface: unknown stream %d", streamID)) |
| 294 | } |
| 295 | if ws.nodes[streamID].state != priorityNodeOpen { |
| 296 | panic(fmt.Sprintf("violation of WriteScheduler interface: stream %d already closed", streamID)) |
| 297 | } |
| 298 | |
| 299 | n := ws.nodes[streamID] |
| 300 | n.state = priorityNodeClosed |
| 301 | n.addBytes(-n.bytes) |
| 302 | |
| 303 | q := n.q |
| 304 | ws.queuePool.put(&q) |
| 305 | n.q.s = nil |
| 306 | if ws.maxClosedNodesInTree > 0 { |
| 307 | ws.addClosedOrIdleNode(&ws.closedNodes, ws.maxClosedNodesInTree, n) |
| 308 | } else { |
| 309 | ws.removeNode(n) |
| 310 | } |
| 311 | } |
| 312 | |
| 313 | func (ws *priorityWriteScheduler) AdjustStream(streamID uint32, priority PriorityParam) { |
| 314 | if streamID == 0 { |
| 315 | panic("adjustPriority on root") |
| 316 | } |
| 317 | |
| 318 | // If streamID does not exist, there are two cases: |
| 319 | // - A closed stream that has been removed (this will have ID <= maxID) |
| 320 | // - An idle stream that is being used for "grouping" (this will have ID > maxID) |
| 321 | n := ws.nodes[streamID] |
| 322 | if n == nil { |
| 323 | if streamID <= ws.maxID || ws.maxIdleNodesInTree == 0 { |
| 324 | return |
| 325 | } |
| 326 | ws.maxID = streamID |
| 327 | n = &priorityNode{ |
| 328 | q: *ws.queuePool.get(), |
| 329 | id: streamID, |
| 330 | weight: priorityDefaultWeight, |
| 331 | state: priorityNodeIdle, |
| 332 | } |
| 333 | n.setParent(&ws.root) |
| 334 | ws.nodes[streamID] = n |
| 335 | ws.addClosedOrIdleNode(&ws.idleNodes, ws.maxIdleNodesInTree, n) |
| 336 | } |
| 337 | |
| 338 | // Section 5.3.1: A dependency on a stream that is not currently in the tree |
| 339 | // results in that stream being given a default priority (Section 5.3.5). |
| 340 | parent := ws.nodes[priority.StreamDep] |
| 341 | if parent == nil { |
| 342 | n.setParent(&ws.root) |
| 343 | n.weight = priorityDefaultWeight |
| 344 | return |
| 345 | } |
| 346 | |
| 347 | // Ignore if the client tries to make a node its own parent. |
| 348 | if n == parent { |
| 349 | return |
| 350 | } |
| 351 | |
| 352 | // Section 5.3.3: |
| 353 | // "If a stream is made dependent on one of its own dependencies, the |
| 354 | // formerly dependent stream is first moved to be dependent on the |
| 355 | // reprioritized stream's previous parent. The moved dependency retains |
| 356 | // its weight." |
| 357 | // |
| 358 | // That is: if parent depends on n, move parent to depend on n.parent. |
| 359 | for x := parent.parent; x != nil; x = x.parent { |
| 360 | if x == n { |
| 361 | parent.setParent(n.parent) |
| 362 | break |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | // Section 5.3.3: The exclusive flag causes the stream to become the sole |
| 367 | // dependency of its parent stream, causing other dependencies to become |
| 368 | // dependent on the exclusive stream. |
| 369 | if priority.Exclusive { |
| 370 | k := parent.kids |
| 371 | for k != nil { |
| 372 | next := k.next |
| 373 | if k != n { |
| 374 | k.setParent(n) |
| 375 | } |
| 376 | k = next |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | n.setParent(parent) |
| 381 | n.weight = priority.Weight |
| 382 | } |
| 383 | |
| 384 | func (ws *priorityWriteScheduler) Push(wr FrameWriteRequest) { |
| 385 | var n *priorityNode |
| 386 | if id := wr.StreamID(); id == 0 { |
| 387 | n = &ws.root |
| 388 | } else { |
| 389 | n = ws.nodes[id] |
| 390 | if n == nil { |
| 391 | // id is an idle or closed stream. wr should not be a HEADERS or |
| 392 | // DATA frame. However, wr can be a RST_STREAM. In this case, we |
| 393 | // push wr onto the root, rather than creating a new priorityNode, |
| 394 | // since RST_STREAM is tiny and the stream's priority is unknown |
| 395 | // anyway. See issue #17919. |
| 396 | if wr.DataSize() > 0 { |
| 397 | panic("add DATA on non-open stream") |
| 398 | } |
| 399 | n = &ws.root |
| 400 | } |
| 401 | } |
| 402 | n.q.push(wr) |
| 403 | } |
| 404 | |
| 405 | func (ws *priorityWriteScheduler) Pop() (wr FrameWriteRequest, ok bool) { |
| 406 | ws.root.walkReadyInOrder(false, &ws.tmp, func(n *priorityNode, openParent bool) bool { |
| 407 | limit := int32(math.MaxInt32) |
| 408 | if openParent { |
| 409 | limit = ws.writeThrottleLimit |
| 410 | } |
| 411 | wr, ok = n.q.consume(limit) |
| 412 | if !ok { |
| 413 | return false |
| 414 | } |
| 415 | n.addBytes(int64(wr.DataSize())) |
| 416 | // If B depends on A and B continuously has data available but A |
| 417 | // does not, gradually increase the throttling limit to allow B to |
| 418 | // steal more and more bandwidth from A. |
| 419 | if openParent { |
| 420 | ws.writeThrottleLimit += 1024 |
| 421 | if ws.writeThrottleLimit < 0 { |
| 422 | ws.writeThrottleLimit = math.MaxInt32 |
| 423 | } |
| 424 | } else if ws.enableWriteThrottle { |
| 425 | ws.writeThrottleLimit = 1024 |
| 426 | } |
| 427 | return true |
| 428 | }) |
| 429 | return wr, ok |
| 430 | } |
| 431 | |
| 432 | func (ws *priorityWriteScheduler) addClosedOrIdleNode(list *[]*priorityNode, maxSize int, n *priorityNode) { |
| 433 | if maxSize == 0 { |
| 434 | return |
| 435 | } |
| 436 | if len(*list) == maxSize { |
| 437 | // Remove the oldest node, then shift left. |
| 438 | ws.removeNode((*list)[0]) |
| 439 | x := (*list)[1:] |
| 440 | copy(*list, x) |
| 441 | *list = (*list)[:len(x)] |
| 442 | } |
| 443 | *list = append(*list, n) |
| 444 | } |
| 445 | |
| 446 | func (ws *priorityWriteScheduler) removeNode(n *priorityNode) { |
| 447 | for k := n.kids; k != nil; k = k.next { |
| 448 | k.setParent(n.parent) |
| 449 | } |
| 450 | n.setParent(nil) |
| 451 | delete(ws.nodes, n.id) |
| 452 | } |