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// Scavenging free pages. 6// 7// This file implements scavenging (the release of physical pages backing mapped 8// memory) of free and unused pages in the heap as a way to deal with page-level 9// fragmentation and reduce the RSS of Go applications. 10// 11// Scavenging in Go happens on two fronts: there's the background 12// (asynchronous) scavenger and the heap-growth (synchronous) scavenger. 13// 14// The former happens on a goroutine much like the background sweeper which is 15// soft-capped at using scavengePercent of the mutator's time, based on 16// order-of-magnitude estimates of the costs of scavenging. The background 17// scavenger's primary goal is to bring the estimated heap RSS of the 18// application down to a goal. 19// 20// That goal is defined as: 21// (retainExtraPercent+100) / 100 * (next_gc / last_next_gc) * last_heap_inuse 22// 23// Essentially, we wish to have the application's RSS track the heap goal, but 24// the heap goal is defined in terms of bytes of objects, rather than pages like 25// RSS. As a result, we need to take into account for fragmentation internal to 26// spans. next_gc / last_next_gc defines the ratio between the current heap goal 27// and the last heap goal, which tells us by how much the heap is growing and 28// shrinking. We estimate what the heap will grow to in terms of pages by taking 29// this ratio and multiplying it by heap_inuse at the end of the last GC, which 30// allows us to account for this additional fragmentation. Note that this 31// procedure makes the assumption that the degree of fragmentation won't change 32// dramatically over the next GC cycle. Overestimating the amount of 33// fragmentation simply results in higher memory use, which will be accounted 34// for by the next pacing up date. Underestimating the fragmentation however 35// could lead to performance degradation. Handling this case is not within the 36// scope of the scavenger. Situations where the amount of fragmentation balloons 37// over the course of a single GC cycle should be considered pathologies, 38// flagged as bugs, and fixed appropriately. 39// 40// An additional factor of retainExtraPercent is added as a buffer to help ensure 41// that there's more unscavenged memory to allocate out of, since each allocation 42// out of scavenged memory incurs a potentially expensive page fault. 43// 44// The goal is updated after each GC and the scavenger's pacing parameters 45// (which live in mheap_) are updated to match. The pacing parameters work much 46// like the background sweeping parameters. The parameters define a line whose 47// horizontal axis is time and vertical axis is estimated heap RSS, and the 48// scavenger attempts to stay below that line at all times. 49// 50// The synchronous heap-growth scavenging happens whenever the heap grows in 51// size, for some definition of heap-growth. The intuition behind this is that 52// the application had to grow the heap because existing fragments were 53// not sufficiently large to satisfy a page-level memory allocation, so we 54// scavenge those fragments eagerly to offset the growth in RSS that results. 55 56package runtime 57 58import ( 59 "runtime/internal/atomic" 60 "runtime/internal/sys" 61 "unsafe" 62) 63 64const ( 65 // The background scavenger is paced according to these parameters. 66 // 67 // scavengePercent represents the portion of mutator time we're willing 68 // to spend on scavenging in percent. 69 scavengePercent = 1 // 1% 70 71 // retainExtraPercent represents the amount of memory over the heap goal 72 // that the scavenger should keep as a buffer space for the allocator. 73 // 74 // The purpose of maintaining this overhead is to have a greater pool of 75 // unscavenged memory available for allocation (since using scavenged memory 76 // incurs an additional cost), to account for heap fragmentation and 77 // the ever-changing layout of the heap. 78 retainExtraPercent = 10 79 80 // maxPagesPerPhysPage is the maximum number of supported runtime pages per 81 // physical page, based on maxPhysPageSize. 82 maxPagesPerPhysPage = maxPhysPageSize / pageSize 83 84 // scavengeCostRatio is the approximate ratio between the costs of using previously 85 // scavenged memory and scavenging memory. 86 // 87 // For most systems the cost of scavenging greatly outweighs the costs 88 // associated with using scavenged memory, making this constant 0. On other systems 89 // (especially ones where "sysUsed" is not just a no-op) this cost is non-trivial. 90 // 91 // This ratio is used as part of multiplicative factor to help the scavenger account 92 // for the additional costs of using scavenged memory in its pacing. 93 scavengeCostRatio = 0.7 * sys.GoosDarwin 94) 95 96// heapRetained returns an estimate of the current heap RSS. 97func heapRetained() uint64 { 98 return atomic.Load64(&memstats.heap_sys) - atomic.Load64(&memstats.heap_released) 99} 100 101// gcPaceScavenger updates the scavenger's pacing, particularly 102// its rate and RSS goal. 103// 104// The RSS goal is based on the current heap goal with a small overhead 105// to accommodate non-determinism in the allocator. 106// 107// The pacing is based on scavengePageRate, which applies to both regular and 108// huge pages. See that constant for more information. 109// 110// mheap_.lock must be held or the world must be stopped. 111func gcPaceScavenger() { 112 // If we're called before the first GC completed, disable scavenging. 113 // We never scavenge before the 2nd GC cycle anyway (we don't have enough 114 // information about the heap yet) so this is fine, and avoids a fault 115 // or garbage data later. 116 if memstats.last_next_gc == 0 { 117 mheap_.scavengeGoal = ^uint64(0) 118 return 119 } 120 // Compute our scavenging goal. 121 goalRatio := float64(memstats.next_gc) / float64(memstats.last_next_gc) 122 retainedGoal := uint64(float64(memstats.last_heap_inuse) * goalRatio) 123 // Add retainExtraPercent overhead to retainedGoal. This calculation 124 // looks strange but the purpose is to arrive at an integer division 125 // (e.g. if retainExtraPercent = 12.5, then we get a divisor of 8) 126 // that also avoids the overflow from a multiplication. 127 retainedGoal += retainedGoal / (1.0 / (retainExtraPercent / 100.0)) 128 // Align it to a physical page boundary to make the following calculations 129 // a bit more exact. 130 retainedGoal = (retainedGoal + uint64(physPageSize) - 1) &^ (uint64(physPageSize) - 1) 131 132 // Represents where we are now in the heap's contribution to RSS in bytes. 133 // 134 // Guaranteed to always be a multiple of physPageSize on systems where 135 // physPageSize <= pageSize since we map heap_sys at a rate larger than 136 // any physPageSize and released memory in multiples of the physPageSize. 137 // 138 // However, certain functions recategorize heap_sys as other stats (e.g. 139 // stack_sys) and this happens in multiples of pageSize, so on systems 140 // where physPageSize > pageSize the calculations below will not be exact. 141 // Generally this is OK since we'll be off by at most one regular 142 // physical page. 143 retainedNow := heapRetained() 144 145 // If we're already below our goal, or within one page of our goal, then disable 146 // the background scavenger. We disable the background scavenger if there's 147 // less than one physical page of work to do because it's not worth it. 148 if retainedNow <= retainedGoal || retainedNow-retainedGoal < uint64(physPageSize) { 149 mheap_.scavengeGoal = ^uint64(0) 150 return 151 } 152 mheap_.scavengeGoal = retainedGoal 153 mheap_.pages.resetScavengeAddr() 154} 155 156// Sleep/wait state of the background scavenger. 157var scavenge struct { 158 lock mutex 159 g *g 160 parked bool 161 timer *timer 162} 163 164// wakeScavenger unparks the scavenger if necessary. It must be called 165// after any pacing update. 166// 167// mheap_.lock and scavenge.lock must not be held. 168func wakeScavenger() { 169 lock(&scavenge.lock) 170 if scavenge.parked { 171 // Try to stop the timer but we don't really care if we succeed. 172 // It's possible that either a timer was never started, or that 173 // we're racing with it. 174 // In the case that we're racing with there's the low chance that 175 // we experience a spurious wake-up of the scavenger, but that's 176 // totally safe. 177 stopTimer(scavenge.timer) 178 179 // Unpark the goroutine and tell it that there may have been a pacing 180 // change. Note that we skip the scheduler's runnext slot because we 181 // want to avoid having the scavenger interfere with the fair 182 // scheduling of user goroutines. In effect, this schedules the 183 // scavenger at a "lower priority" but that's OK because it'll 184 // catch up on the work it missed when it does get scheduled. 185 scavenge.parked = false 186 systemstack(func() { 187 ready(scavenge.g, 0, false) 188 }) 189 } 190 unlock(&scavenge.lock) 191} 192 193// scavengeSleep attempts to put the scavenger to sleep for ns. 194// 195// Note that this function should only be called by the scavenger. 196// 197// The scavenger may be woken up earlier by a pacing change, and it may not go 198// to sleep at all if there's a pending pacing change. 199// 200// Returns the amount of time actually slept. 201func scavengeSleep(ns int64) int64 { 202 lock(&scavenge.lock) 203 204 // Set the timer. 205 // 206 // This must happen here instead of inside gopark 207 // because we can't close over any variables without 208 // failing escape analysis. 209 start := nanotime() 210 resetTimer(scavenge.timer, start+ns) 211 212 // Mark ourself as asleep and go to sleep. 213 scavenge.parked = true 214 goparkunlock(&scavenge.lock, waitReasonSleep, traceEvGoSleep, 2) 215 216 // Return how long we actually slept for. 217 return nanotime() - start 218} 219 220// Background scavenger. 221// 222// The background scavenger maintains the RSS of the application below 223// the line described by the proportional scavenging statistics in 224// the mheap struct. 225func bgscavenge(c chan int) { 226 scavenge.g = getg() 227 228 lock(&scavenge.lock) 229 scavenge.parked = true 230 231 scavenge.timer = new(timer) 232 scavenge.timer.f = func(_ interface{}, _ uintptr) { 233 wakeScavenger() 234 } 235 236 c <- 1 237 goparkunlock(&scavenge.lock, waitReasonGCScavengeWait, traceEvGoBlock, 1) 238 239 // Exponentially-weighted moving average of the fraction of time this 240 // goroutine spends scavenging (that is, percent of a single CPU). 241 // It represents a measure of scheduling overheads which might extend 242 // the sleep or the critical time beyond what's expected. Assume no 243 // overhead to begin with. 244 // 245 // TODO(mknyszek): Consider making this based on total CPU time of the 246 // application (i.e. scavengePercent * GOMAXPROCS). This isn't really 247 // feasible now because the scavenger acquires the heap lock over the 248 // scavenging operation, which means scavenging effectively blocks 249 // allocators and isn't scalable. However, given a scalable allocator, 250 // it makes sense to also make the scavenger scale with it; if you're 251 // allocating more frequently, then presumably you're also generating 252 // more work for the scavenger. 253 const idealFraction = scavengePercent / 100.0 254 scavengeEWMA := float64(idealFraction) 255 256 for { 257 released := uintptr(0) 258 259 // Time in scavenging critical section. 260 crit := float64(0) 261 262 // Run on the system stack since we grab the heap lock, 263 // and a stack growth with the heap lock means a deadlock. 264 systemstack(func() { 265 lock(&mheap_.lock) 266 267 // If background scavenging is disabled or if there's no work to do just park. 268 retained, goal := heapRetained(), mheap_.scavengeGoal 269 if retained <= goal { 270 unlock(&mheap_.lock) 271 return 272 } 273 unlock(&mheap_.lock) 274 275 // Scavenge one page, and measure the amount of time spent scavenging. 276 start := nanotime() 277 released = mheap_.pages.scavengeOne(physPageSize, false) 278 atomic.Xadduintptr(&mheap_.pages.scavReleased, released) 279 crit = float64(nanotime() - start) 280 }) 281 282 if released == 0 { 283 lock(&scavenge.lock) 284 scavenge.parked = true 285 goparkunlock(&scavenge.lock, waitReasonGCScavengeWait, traceEvGoBlock, 1) 286 continue 287 } 288 289 // Multiply the critical time by 1 + the ratio of the costs of using 290 // scavenged memory vs. scavenging memory. This forces us to pay down 291 // the cost of reusing this memory eagerly by sleeping for a longer period 292 // of time and scavenging less frequently. More concretely, we avoid situations 293 // where we end up scavenging so often that we hurt allocation performance 294 // because of the additional overheads of using scavenged memory. 295 crit *= 1 + scavengeCostRatio 296 297 // If we spent more than 10 ms (for example, if the OS scheduled us away, or someone 298 // put their machine to sleep) in the critical section, bound the time we use to 299 // calculate at 10 ms to avoid letting the sleep time get arbitrarily high. 300 const maxCrit = 10e6 301 if crit > maxCrit { 302 crit = maxCrit 303 } 304 305 // Compute the amount of time to sleep, assuming we want to use at most 306 // scavengePercent of CPU time. Take into account scheduling overheads 307 // that may extend the length of our sleep by multiplying by how far 308 // off we are from the ideal ratio. For example, if we're sleeping too 309 // much, then scavengeEMWA < idealFraction, so we'll adjust the sleep time 310 // down. 311 adjust := scavengeEWMA / idealFraction 312 sleepTime := int64(adjust * crit / (scavengePercent / 100.0)) 313 314 // Go to sleep. 315 slept := scavengeSleep(sleepTime) 316 317 // Compute the new ratio. 318 fraction := crit / (crit + float64(slept)) 319 320 // Set a lower bound on the fraction. 321 // Due to OS-related anomalies we may "sleep" for an inordinate amount 322 // of time. Let's avoid letting the ratio get out of hand by bounding 323 // the sleep time we use in our EWMA. 324 const minFraction = 1 / 1000 325 if fraction < minFraction { 326 fraction = minFraction 327 } 328 329 // Update scavengeEWMA by merging in the new crit/slept ratio. 330 const alpha = 0.5 331 scavengeEWMA = alpha*fraction + (1-alpha)*scavengeEWMA 332 } 333} 334 335// scavenge scavenges nbytes worth of free pages, starting with the 336// highest address first. Successive calls continue from where it left 337// off until the heap is exhausted. Call resetScavengeAddr to bring it 338// back to the top of the heap. 339// 340// Returns the amount of memory scavenged in bytes. 341// 342// If locked == false, s.mheapLock must not be locked. If locked == true, 343// s.mheapLock must be locked. 344// 345// Must run on the system stack because scavengeOne must run on the 346// system stack. 347// 348//go:systemstack 349func (s *pageAlloc) scavenge(nbytes uintptr, locked bool) uintptr { 350 released := uintptr(0) 351 for released < nbytes { 352 r := s.scavengeOne(nbytes-released, locked) 353 if r == 0 { 354 // Nothing left to scavenge! Give up. 355 break 356 } 357 released += r 358 } 359 return released 360} 361 362// printScavTrace prints a scavenge trace line to standard error. 363// 364// released should be the amount of memory released since the last time this 365// was called, and forced indicates whether the scavenge was forced by the 366// application. 367func printScavTrace(released uintptr, forced bool) { 368 printlock() 369 print("scav ", 370 released>>10, " KiB work, ", 371 atomic.Load64(&memstats.heap_released)>>10, " KiB total, ", 372 (atomic.Load64(&memstats.heap_inuse)*100)/heapRetained(), "% util", 373 ) 374 if forced { 375 print(" (forced)") 376 } 377 println() 378 printunlock() 379} 380 381// resetScavengeAddr sets the scavenge start address to the top of the heap's 382// address space. This should be called each time the scavenger's pacing 383// changes. 384// 385// s.mheapLock must be held. 386func (s *pageAlloc) resetScavengeAddr() { 387 released := atomic.Loaduintptr(&s.scavReleased) 388 if debug.scavtrace > 0 { 389 printScavTrace(released, false) 390 } 391 // Subtract from scavReleased instead of just setting it to zero because 392 // the scavenger could have increased scavReleased concurrently with the 393 // load above, and we may miss an update by just blindly zeroing the field. 394 atomic.Xadduintptr(&s.scavReleased, -released) 395 s.scavAddr = chunkBase(s.end) - 1 396} 397 398// scavengeOne starts from s.scavAddr and walks down the heap until it finds 399// a contiguous run of pages to scavenge. It will try to scavenge at most 400// max bytes at once, but may scavenge more to avoid breaking huge pages. Once 401// it scavenges some memory it returns how much it scavenged and updates s.scavAddr 402// appropriately. s.scavAddr must be reset manually and externally. 403// 404// Should it exhaust the heap, it will return 0 and set s.scavAddr to minScavAddr. 405// 406// If locked == false, s.mheapLock must not be locked. 407// If locked == true, s.mheapLock must be locked. 408// 409// Must be run on the system stack because it either acquires the heap lock 410// or executes with the heap lock acquired. 411// 412//go:systemstack 413func (s *pageAlloc) scavengeOne(max uintptr, locked bool) uintptr { 414 // Calculate the maximum number of pages to scavenge. 415 // 416 // This should be alignUp(max, pageSize) / pageSize but max can and will 417 // be ^uintptr(0), so we need to be very careful not to overflow here. 418 // Rather than use alignUp, calculate the number of pages rounded down 419 // first, then add back one if necessary. 420 maxPages := max / pageSize 421 if max%pageSize != 0 { 422 maxPages++ 423 } 424 425 // Calculate the minimum number of pages we can scavenge. 426 // 427 // Because we can only scavenge whole physical pages, we must 428 // ensure that we scavenge at least minPages each time, aligned 429 // to minPages*pageSize. 430 minPages := physPageSize / pageSize 431 if minPages < 1 { 432 minPages = 1 433 } 434 435 // Helpers for locking and unlocking only if locked == false. 436 lockHeap := func() { 437 if !locked { 438 lock(s.mheapLock) 439 } 440 } 441 unlockHeap := func() { 442 if !locked { 443 unlock(s.mheapLock) 444 } 445 } 446 447 lockHeap() 448 ci := chunkIndex(s.scavAddr) 449 if ci < s.start { 450 unlockHeap() 451 return 0 452 } 453 454 // Check the chunk containing the scav addr, starting at the addr 455 // and see if there are any free and unscavenged pages. 456 // 457 // Only check this if s.scavAddr is covered by any address range 458 // in s.inUse, so that we know our check of the summary is safe. 459 if s.inUse.contains(s.scavAddr) && s.summary[len(s.summary)-1][ci].max() >= uint(minPages) { 460 // We only bother looking for a candidate if there at least 461 // minPages free pages at all. It's important that we only 462 // continue if the summary says we can because that's how 463 // we can tell if parts of the address space are unused. 464 // See the comment on s.chunks in mpagealloc.go. 465 base, npages := s.chunkOf(ci).findScavengeCandidate(chunkPageIndex(s.scavAddr), minPages, maxPages) 466 467 // If we found something, scavenge it and return! 468 if npages != 0 { 469 s.scavengeRangeLocked(ci, base, npages) 470 unlockHeap() 471 return uintptr(npages) * pageSize 472 } 473 } 474 475 // getInUseRange returns the highest range in the 476 // intersection of [0, addr] and s.inUse. 477 // 478 // s.mheapLock must be held. 479 getInUseRange := func(addr uintptr) addrRange { 480 top := s.inUse.findSucc(addr) 481 if top == 0 { 482 return addrRange{} 483 } 484 r := s.inUse.ranges[top-1] 485 // addr is inclusive, so treat it as such when 486 // updating the limit, which is exclusive. 487 if r.limit > addr+1 { 488 r.limit = addr + 1 489 } 490 return r 491 } 492 493 // Slow path: iterate optimistically over the in-use address space 494 // looking for any free and unscavenged page. If we think we see something, 495 // lock and verify it! 496 // 497 // We iterate over the address space by taking ranges from inUse. 498newRange: 499 for { 500 r := getInUseRange(s.scavAddr) 501 if r.size() == 0 { 502 break 503 } 504 unlockHeap() 505 506 // Iterate over all of the chunks described by r. 507 // Note that r.limit is the exclusive upper bound, but what 508 // we want is the top chunk instead, inclusive, so subtract 1. 509 bot, top := chunkIndex(r.base), chunkIndex(r.limit-1) 510 for i := top; i >= bot; i-- { 511 // If this chunk is totally in-use or has no unscavenged pages, don't bother 512 // doing a more sophisticated check. 513 // 514 // Note we're accessing the summary and the chunks without a lock, but 515 // that's fine. We're being optimistic anyway. 516 517 // Check quickly if there are enough free pages at all. 518 if s.summary[len(s.summary)-1][i].max() < uint(minPages) { 519 continue 520 } 521 522 // Run over the chunk looking harder for a candidate. Again, we could 523 // race with a lot of different pieces of code, but we're just being 524 // optimistic. Make sure we load the l2 pointer atomically though, to 525 // avoid races with heap growth. It may or may not be possible to also 526 // see a nil pointer in this case if we do race with heap growth, but 527 // just defensively ignore the nils. This operation is optimistic anyway. 528 l2 := (*[1 << pallocChunksL2Bits]pallocData)(atomic.Loadp(unsafe.Pointer(&s.chunks[i.l1()]))) 529 if l2 == nil || !l2[i.l2()].hasScavengeCandidate(minPages) { 530 continue 531 } 532 533 // We found a candidate, so let's lock and verify it. 534 lockHeap() 535 536 // Find, verify, and scavenge if we can. 537 chunk := s.chunkOf(i) 538 base, npages := chunk.findScavengeCandidate(pallocChunkPages-1, minPages, maxPages) 539 if npages > 0 { 540 // We found memory to scavenge! Mark the bits and report that up. 541 // scavengeRangeLocked will update scavAddr for us, also. 542 s.scavengeRangeLocked(i, base, npages) 543 unlockHeap() 544 return uintptr(npages) * pageSize 545 } 546 547 // We were fooled, let's take this opportunity to move the scavAddr 548 // all the way down to where we searched as scavenged for future calls 549 // and keep iterating. Then, go get a new range. 550 s.scavAddr = chunkBase(i-1) + pallocChunkPages*pageSize - 1 551 continue newRange 552 } 553 lockHeap() 554 555 // Move the scavenger down the heap, past everything we just searched. 556 // Since we don't check if scavAddr moved while twe let go of the heap lock, 557 // it's possible that it moved down and we're moving it up here. This 558 // raciness could result in us searching parts of the heap unnecessarily. 559 // TODO(mknyszek): Remove this racy behavior through explicit address 560 // space reservations, which are difficult to do with just scavAddr. 561 s.scavAddr = r.base - 1 562 } 563 // We reached the end of the in-use address space and couldn't find anything, 564 // so signal that there's nothing left to scavenge. 565 s.scavAddr = minScavAddr 566 unlockHeap() 567 568 return 0 569} 570 571// scavengeRangeLocked scavenges the given region of memory. 572// 573// s.mheapLock must be held. 574func (s *pageAlloc) scavengeRangeLocked(ci chunkIdx, base, npages uint) { 575 s.chunkOf(ci).scavenged.setRange(base, npages) 576 577 // Compute the full address for the start of the range. 578 addr := chunkBase(ci) + uintptr(base)*pageSize 579 580 // Update the scav pointer. 581 s.scavAddr = addr - 1 582 583 // Only perform the actual scavenging if we're not in a test. 584 // It's dangerous to do so otherwise. 585 if s.test { 586 return 587 } 588 sysUnused(unsafe.Pointer(addr), uintptr(npages)*pageSize) 589 590 // Update global accounting only when not in test, otherwise 591 // the runtime's accounting will be wrong. 592 mSysStatInc(&memstats.heap_released, uintptr(npages)*pageSize) 593} 594 595// fillAligned returns x but with all zeroes in m-aligned 596// groups of m bits set to 1 if any bit in the group is non-zero. 597// 598// For example, fillAligned(0x0100a3, 8) == 0xff00ff. 599// 600// Note that if m == 1, this is a no-op. 601// 602// m must be a power of 2 <= maxPagesPerPhysPage. 603func fillAligned(x uint64, m uint) uint64 { 604 apply := func(x uint64, c uint64) uint64 { 605 // The technique used it here is derived from 606 // https://graphics.stanford.edu/~seander/bithacks.html#ZeroInWord 607 // and extended for more than just bytes (like nibbles 608 // and uint16s) by using an appropriate constant. 609 // 610 // To summarize the technique, quoting from that page: 611 // "[It] works by first zeroing the high bits of the [8] 612 // bytes in the word. Subsequently, it adds a number that 613 // will result in an overflow to the high bit of a byte if 614 // any of the low bits were initially set. Next the high 615 // bits of the original word are ORed with these values; 616 // thus, the high bit of a byte is set iff any bit in the 617 // byte was set. Finally, we determine if any of these high 618 // bits are zero by ORing with ones everywhere except the 619 // high bits and inverting the result." 620 return ^((((x & c) + c) | x) | c) 621 } 622 // Transform x to contain a 1 bit at the top of each m-aligned 623 // group of m zero bits. 624 switch m { 625 case 1: 626 return x 627 case 2: 628 x = apply(x, 0x5555555555555555) 629 case 4: 630 x = apply(x, 0x7777777777777777) 631 case 8: 632 x = apply(x, 0x7f7f7f7f7f7f7f7f) 633 case 16: 634 x = apply(x, 0x7fff7fff7fff7fff) 635 case 32: 636 x = apply(x, 0x7fffffff7fffffff) 637 case 64: // == maxPagesPerPhysPage 638 x = apply(x, 0x7fffffffffffffff) 639 default: 640 throw("bad m value") 641 } 642 // Now, the top bit of each m-aligned group in x is set 643 // that group was all zero in the original x. 644 645 // From each group of m bits subtract 1. 646 // Because we know only the top bits of each 647 // m-aligned group are set, we know this will 648 // set each group to have all the bits set except 649 // the top bit, so just OR with the original 650 // result to set all the bits. 651 return ^((x - (x >> (m - 1))) | x) 652} 653 654// hasScavengeCandidate returns true if there's any min-page-aligned groups of 655// min pages of free-and-unscavenged memory in the region represented by this 656// pallocData. 657// 658// min must be a non-zero power of 2 <= maxPagesPerPhysPage. 659func (m *pallocData) hasScavengeCandidate(min uintptr) bool { 660 if min&(min-1) != 0 || min == 0 { 661 print("runtime: min = ", min, "\n") 662 throw("min must be a non-zero power of 2") 663 } else if min > maxPagesPerPhysPage { 664 print("runtime: min = ", min, "\n") 665 throw("min too large") 666 } 667 668 // The goal of this search is to see if the chunk contains any free and unscavenged memory. 669 for i := len(m.scavenged) - 1; i >= 0; i-- { 670 // 1s are scavenged OR non-free => 0s are unscavenged AND free 671 // 672 // TODO(mknyszek): Consider splitting up fillAligned into two 673 // functions, since here we technically could get by with just 674 // the first half of its computation. It'll save a few instructions 675 // but adds some additional code complexity. 676 x := fillAligned(m.scavenged[i]|m.pallocBits[i], uint(min)) 677 678 // Quickly skip over chunks of non-free or scavenged pages. 679 if x != ^uint64(0) { 680 return true 681 } 682 } 683 return false 684} 685 686// findScavengeCandidate returns a start index and a size for this pallocData 687// segment which represents a contiguous region of free and unscavenged memory. 688// 689// searchIdx indicates the page index within this chunk to start the search, but 690// note that findScavengeCandidate searches backwards through the pallocData. As a 691// a result, it will return the highest scavenge candidate in address order. 692// 693// min indicates a hard minimum size and alignment for runs of pages. That is, 694// findScavengeCandidate will not return a region smaller than min pages in size, 695// or that is min pages or greater in size but not aligned to min. min must be 696// a non-zero power of 2 <= maxPagesPerPhysPage. 697// 698// max is a hint for how big of a region is desired. If max >= pallocChunkPages, then 699// findScavengeCandidate effectively returns entire free and unscavenged regions. 700// If max < pallocChunkPages, it may truncate the returned region such that size is 701// max. However, findScavengeCandidate may still return a larger region if, for 702// example, it chooses to preserve huge pages, or if max is not aligned to min (it 703// will round up). That is, even if max is small, the returned size is not guaranteed 704// to be equal to max. max is allowed to be less than min, in which case it is as if 705// max == min. 706func (m *pallocData) findScavengeCandidate(searchIdx uint, min, max uintptr) (uint, uint) { 707 if min&(min-1) != 0 || min == 0 { 708 print("runtime: min = ", min, "\n") 709 throw("min must be a non-zero power of 2") 710 } else if min > maxPagesPerPhysPage { 711 print("runtime: min = ", min, "\n") 712 throw("min too large") 713 } 714 // max may not be min-aligned, so we might accidentally truncate to 715 // a max value which causes us to return a non-min-aligned value. 716 // To prevent this, align max up to a multiple of min (which is always 717 // a power of 2). This also prevents max from ever being less than 718 // min, unless it's zero, so handle that explicitly. 719 if max == 0 { 720 max = min 721 } else { 722 max = alignUp(max, min) 723 } 724 725 i := int(searchIdx / 64) 726 // Start by quickly skipping over blocks of non-free or scavenged pages. 727 for ; i >= 0; i-- { 728 // 1s are scavenged OR non-free => 0s are unscavenged AND free 729 x := fillAligned(m.scavenged[i]|m.pallocBits[i], uint(min)) 730 if x != ^uint64(0) { 731 break 732 } 733 } 734 if i < 0 { 735 // Failed to find any free/unscavenged pages. 736 return 0, 0 737 } 738 // We have something in the 64-bit chunk at i, but it could 739 // extend further. Loop until we find the extent of it. 740 741 // 1s are scavenged OR non-free => 0s are unscavenged AND free 742 x := fillAligned(m.scavenged[i]|m.pallocBits[i], uint(min)) 743 z1 := uint(sys.LeadingZeros64(^x)) 744 run, end := uint(0), uint(i)*64+(64-z1) 745 if x<<z1 != 0 { 746 // After shifting out z1 bits, we still have 1s, 747 // so the run ends inside this word. 748 run = uint(sys.LeadingZeros64(x << z1)) 749 } else { 750 // After shifting out z1 bits, we have no more 1s. 751 // This means the run extends to the bottom of the 752 // word so it may extend into further words. 753 run = 64 - z1 754 for j := i - 1; j >= 0; j-- { 755 x := fillAligned(m.scavenged[j]|m.pallocBits[j], uint(min)) 756 run += uint(sys.LeadingZeros64(x)) 757 if x != 0 { 758 // The run stopped in this word. 759 break 760 } 761 } 762 } 763 764 // Split the run we found if it's larger than max but hold on to 765 // our original length, since we may need it later. 766 size := run 767 if size > uint(max) { 768 size = uint(max) 769 } 770 start := end - size 771 772 // Each huge page is guaranteed to fit in a single palloc chunk. 773 // 774 // TODO(mknyszek): Support larger huge page sizes. 775 // TODO(mknyszek): Consider taking pages-per-huge-page as a parameter 776 // so we can write tests for this. 777 if physHugePageSize > pageSize && physHugePageSize > physPageSize { 778 // We have huge pages, so let's ensure we don't break one by scavenging 779 // over a huge page boundary. If the range [start, start+size) overlaps with 780 // a free-and-unscavenged huge page, we want to grow the region we scavenge 781 // to include that huge page. 782 783 // Compute the huge page boundary above our candidate. 784 pagesPerHugePage := uintptr(physHugePageSize / pageSize) 785 hugePageAbove := uint(alignUp(uintptr(start), pagesPerHugePage)) 786 787 // If that boundary is within our current candidate, then we may be breaking 788 // a huge page. 789 if hugePageAbove <= end { 790 // Compute the huge page boundary below our candidate. 791 hugePageBelow := uint(alignDown(uintptr(start), pagesPerHugePage)) 792 793 if hugePageBelow >= end-run { 794 // We're in danger of breaking apart a huge page since start+size crosses 795 // a huge page boundary and rounding down start to the nearest huge 796 // page boundary is included in the full run we found. Include the entire 797 // huge page in the bound by rounding down to the huge page size. 798 size = size + (start - hugePageBelow) 799 start = hugePageBelow 800 } 801 } 802 } 803 return start, size 804} 805