Cleanup and friction reduction
For #65355.
Change-Id: Ia14c9dc584a529a35b97801dd3e95b9acc99a511
Reviewed-on: https://go-review.googlesource.com/c/go/+/600436
Reviewed-by: Keith Randall <khr@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@golang.org>
The page tracer's functionality is now captured by the regular execution
tracer as an experimental GODEBUG variable. This is a lot more usable
and maintainable than the page tracer, which is likely to have bitrotted
by this point. There's also no tooling available for the page tracer.
Change-Id: I2408394555e01dde75a522e9a489b7e55cf12c8e
Reviewed-on: https://go-review.googlesource.com/c/go/+/583379
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Carlos Amedee <carlos@golang.org>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
During my research of pahole with Go structs, I've found couple of
structs in runtime/ pkg where we can reduce several structs' sizes
highligted by pahole tool which detect byte holes and paddings.
Overall, there are 80 bytes reduced.
Change-Id: I398e5ed6f5b199394307741981cb5ad5b875e98f
Reviewed-on: https://go-review.googlesource.com/c/go/+/578795
Auto-Submit: Keith Randall <khr@golang.org>
Reviewed-by: Joedian Reid <joedian@google.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
From the beginning of Go, the time package has had a gotcha:
if you use a select on <-time.After(1*time.Minute), even if the select
finishes immediately because some other case is ready, the underlying
timer from time.After keeps running until the minute is over. This
pins the timer in the timer heap, which keeps it from being garbage
collected and in extreme cases also slows down timer operations.
The lack of garbage collection is the more important problem.
The docs for After warn against this scenario and suggest using
NewTimer with a call to Stop after the select instead, purely to work
around this garbage collection problem.
Oddly, the docs for NewTimer and NewTicker do not mention this
problem, but they have the same issue: they cannot be collected until
either they are Stopped or, in the case of Timer, the timer expires.
(Tickers repeat, so they never expire.) People have built up a shared
knowledge that timers and tickers need to defer t.Stop even though the
docs do not mention this (it is somewhat implied by the After docs).
This CL fixes the garbage collection problem, so that a timer that is
unreferenced can be GC'ed immediately, even if it is still running.
The approach is to only insert the timer into the heap when some
channel operation is blocked on it; the last channel operation to stop
using the timer takes it back out of the heap. When a timer's channel
is no longer referenced, there are no channel operations blocked on
it, so it's not in the heap, so it can be GC'ed immediately.
This CL adds an undocumented GODEBUG asynctimerchan=1
that will disable the change. The documentation happens in
the CL 568341.
Fixes#8898.
Fixes#61542.
Change-Id: Ieb303b6de1fb3527d3256135151a9e983f3c27e6
Reviewed-on: https://go-review.googlesource.com/c/go/+/512355
Reviewed-by: Austin Clements <austin@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Auto-Submit: Russ Cox <rsc@golang.org>
The timers had evolved to the point where the state was stored as follows:
if timer in heap:
state has timerHeaped set
if heap timer is stale:
heap deadline in t.when
real deadline in t.nextWhen
state has timerNextWhen set
else:
real deadline in t.when
t.nextWhen unset
else:
real deadline in t.when
t.nextWhen unset
That made it hard to find the real deadline and just hard to think about everything.
The new state is:
real deadline in t.when (always)
if timer in heap:
state has timerHeaped set
heap deadline in t.whenHeap
if heap timer is stale:
state has timerModified set
Separately, the 'state' word itself was being used as a lock
and state bits because the code started with CAS loops,
which we abstracted into the lock/unlock methods step by step.
At this point, we can switch to a real lock, making sure to
publish the one boolean needed by timers fast paths
at each unlock.
All this simplifies various logic considerably.
Change-Id: I35766204f7a26d999206bd56cc0db60ad1b17cbe
Reviewed-on: https://go-review.googlesource.com/c/go/+/570335
Auto-Submit: Russ Cox <rsc@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
If user code has two timers t1 and t2 and does *t1 = *t2
(or *t1 = Timer{}), it creeps me out that we would be
corrupting the runtime data structures inlined in the
Timer struct. Replace that field with a pointer to the
runtime data structure instead, so that the corruption
cannot happen, even in a badly behaved program.
In fact, remove the struct definition entirely and linkname
a constructor instead. Now the runtime can evolve the struct
however it likes without needing to keep package time in sync.
Also move the workaround logic for #21874 out of
runtime and into package time.
Change-Id: Ia30f7802ee7b3a11f5d8a78dd30fd9c8633dc787
Reviewed-on: https://go-review.googlesource.com/c/go/+/568339
Reviewed-by: Ian Lance Taylor <iant@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Change-Id: I3f0b7209621b39cee69566a5cc95e4343b4f1f20
GitHub-Last-Rev: af9dbbe69ad74e8c210254dafa260a886b690853
GitHub-Pull-Request: golang/go#63321
Reviewed-on: https://go-review.googlesource.com/c/go/+/531916
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Mauri de Souza Meneguzzo <mauri870@gmail.com>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
After the previous CL, this is now all dead code. This change is
separated out to make the previous one easy to backport.
For #63334.
Related to #61718 and #59960.
Change-Id: I109673ed97c62c472bbe2717dfeeb5aa4fc883ea
Reviewed-on: https://go-review.googlesource.com/c/go/+/532117
Reviewed-by: Michael Pratt <mpratt@google.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
This has caused performance issues in production environments.
Disable it until further notice.
Fixes#63334.
Related to #61718 and #59960.
Change-Id: If84c5a8685825d43c912a71418f2597e44e867e5
Reviewed-on: https://go-review.googlesource.com/c/go/+/531816
Reviewed-by: Michael Pratt <mpratt@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Run 'unconvert -safe -apply' (https://github.com/mdempsky/unconvert)
Change-Id: I24b7cd7d286cddce86431d8470d15c5f3f0d1106
GitHub-Last-Rev: 022e75384c08bb899a8951ba0daffa0f2e14d5a7
GitHub-Pull-Request: golang/go#62662
Reviewed-on: https://go-review.googlesource.com/c/go/+/528696
Auto-Submit: Ian Lance Taylor <iant@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
The goal is supposed to be (100-reduceExtraPercent) / 100 * memoryLimit,
as stated in the original design.
Fixes#62449
Change-Id: Ia33acadc3320aa3625814595a24b9631ae8896d9
Reviewed-on: https://go-review.googlesource.com/c/go/+/525555
Reviewed-by: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Run-TryBot: Andy Pan <panjf2000@gmail.com>
Currently the runtime marks all new memory as MADV_HUGEPAGE on Linux and
manages its hugepage eligibility status. Unfortunately, the default
THP behavior on most Linux distros is that MADV_HUGEPAGE blocks while
the kernel eagerly reclaims and compacts memory to allocate a hugepage.
This direct reclaim and compaction is unbounded, and may result in
significant application thread stalls. In really bad cases, this can
exceed 100s of ms or even seconds.
Really all we want is to undo MADV_NOHUGEPAGE marks and let the default
Linux paging behavior take over, but the only way to unmark a region as
MADV_NOHUGEPAGE is to also mark it MADV_HUGEPAGE.
The overall strategy of trying to keep hugepages for the heap unbroken
however is sound. So instead let's use the new shiny MADV_COLLAPSE if it
exists.
MADV_COLLAPSE makes a best-effort synchronous attempt at collapsing the
physical memory backing a memory region into a hugepage. We'll use
MADV_COLLAPSE where we would've used MADV_HUGEPAGE, and stop using
MADV_NOHUGEPAGE altogether.
Because MADV_COLLAPSE is synchronous, it's also important to not
re-collapse huge pages if the huge pages are likely part of some large
allocation. Although in many cases it's advantageous to back these
allocations with hugepages because they're contiguous, eagerly
collapsing every hugepage means having to page in at least part of the
large allocation.
However, because we won't use MADV_NOHUGEPAGE anymore, we'll no longer
handle the fact that khugepaged might come in and back some memory we
returned to the OS with a hugepage. I've come to the conclusion that
this is basically unavoidable without a new madvise flag and that it's
just not a good default. If this change lands, advice about Linux huge
page settings will be added to the GC guide.
Verified that this change doesn't regress Sweet, at least not on my
machine with:
/sys/kernel/mm/transparent_hugepage/enabled [always or madvise]
/sys/kernel/mm/transparent_hugepage/defrag [madvise]
/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none [0 or 511]
Unfortunately, this workaround means that we only get forced hugepages
on Linux 6.1+.
Fixes#61718.
Change-Id: I7f4a7ba397847de29f800a99f9cb66cb2720a533
Reviewed-on: https://go-review.googlesource.com/c/go/+/516795
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Change-Id: I5f06a4ef1d827eb0fe32a8d98444142108b0d573
Reviewed-on: https://go-review.googlesource.com/c/go/+/508996
Run-TryBot: Ian Lance Taylor <iant@google.com>
Auto-Submit: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Auto-Submit: Ian Lance Taylor <iant@google.com>
Reviewed-by: Keith Randall <khr@google.com>
This change adds traceBlockReason which leaks fewer implementation
details of the tracer to the runtime. Currently, gopark is called with
an explicit trace event, but this leaks details about trace internals
throughout the runtime.
This change will make it easier to change out the trace implementation.
Change-Id: Id633e1704d2c8838c6abd1214d9695537c4ac7db
Reviewed-on: https://go-review.googlesource.com/c/go/+/494185
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Also, clean up atomics on released-per-cycle while we're here.
For #57069.
Change-Id: I14026e8281f01dea1e8c8de6aa8944712b7b24d9
Reviewed-on: https://go-review.googlesource.com/c/go/+/495916
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
The scavenge index currently doesn't guard against overflow, and CL
436395 removed the minHeapIdx optimization that allows the chunk scan to
skip scanning chunks that haven't been mapped for the heap, and are only
available as a consequence of chunks' mapped region being rounded out to
a page on both ends.
Because the 0'th chunk is never mapped, minHeapIdx effectively prevents
overflow, fixing the iOS breakage.
This change also refactors growth and initialization a little bit to
decouple it from pageAlloc a bit and share code across platforms.
Change-Id: If7fc3245aa81cf99451bf8468458da31986a9b0a
Reviewed-on: https://go-review.googlesource.com/c/go/+/486695
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Currently these fields are uninitialized causing failures on aix-ppc64,
which has a slightly oddly-defined address space compared to the rest.
Change-Id: I2aa46731174154dce86c2074bd0b00eef955d86d
Reviewed-on: https://go-review.googlesource.com/c/go/+/486655
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Bypass: Michael Knyszek <mknyszek@google.com>
This change makes it so that on Linux the Go runtime explicitly marks
page heap memory as either available to be backed by hugepages or not
using heuristics based on density.
The motivation behind this change is twofold:
1. In default Linux configurations, khugepaged can recoalesce hugepages
even after the scavenger breaks them up, resulting in significant
overheads for small heaps when their heaps shrink.
2. The Go runtime already has some heuristics about this, but those
heuristics appear to have bit-rotted and result in haphazard
hugepage management. Unlucky (but otherwise fairly dense) regions of
memory end up not backed by huge pages while sparse regions end up
accidentally marked MADV_HUGEPAGE and are not later broken up by the
scavenger, because it already got the memory it needed from more
dense sections (this is more likely to happen with small heaps that
go idle).
In this change, the runtime uses a new policy:
1. Mark all new memory MADV_HUGEPAGE.
2. Track whether each page chunk (4 MiB) became dense during the GC
cycle. Mark those MADV_HUGEPAGE, and hide them from the scavenger.
3. If a chunk is not dense for 1 full GC cycle, make it visible to the
scavenger.
4. The scavenger marks a chunk MADV_NOHUGEPAGE before it scavenges it.
This policy is intended to try and back memory that is a good candidate
for huge pages (high occupancy) with huge pages, and give memory that is
not (low occupancy) to the scavenger. Occupancy is defined not just by
occupancy at any instant of time, but also occupancy in the near future.
It's generally true that by the end of a GC cycle the heap gets quite
dense (from the perspective of the page allocator).
Because we want scavenging and huge page management to happen together
(the right time to MADV_NOHUGEPAGE is just before scavenging in order to
break up huge pages and keep them that way) and the cost of applying
MADV_HUGEPAGE and MADV_NOHUGEPAGE is somewhat high, the scavenger avoids
releasing memory in dense page chunks. All this together means the
scavenger will now more generally release memory on a ~1 GC cycle delay.
Notably this has implications for scavenging to maintain the memory
limit and the runtime/debug.FreeOSMemory API. This change makes it so
that in these cases all memory is visible to the scavenger regardless of
sparseness and delays the page allocator in re-marking this memory with
MADV_NOHUGEPAGE for around 1 GC cycle to mitigate churn.
The end result of this change should be little-to-no performance
difference for dense heaps (MADV_HUGEPAGE works a lot like the default
unmarked state) but should allow the scavenger to more effectively take
back fragments of huge pages. The main risk here is churn, because
MADV_HUGEPAGE usually forces the kernel to immediately back memory with
a huge page. That's the reason for the large amount of hysteresis (1
full GC cycle) and why the definition of high density is 96% occupancy.
Fixes#55328.
Change-Id: I8da7998f1a31b498a9cc9bc662c1ae1a6bf64630
Reviewed-on: https://go-review.googlesource.com/c/go/+/436395
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This change adds a new GODEBUG flag called pagetrace that writes a
low-overhead trace of how pages of memory are managed by the Go runtime.
The page tracer is kept behind a GOEXPERIMENT flag due to a potential
security risk for setuid binaries.
Change-Id: I6f4a2447d02693c25214400846a5d2832ad6e5c0
Reviewed-on: https://go-review.googlesource.com/c/go/+/444157
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This change modifies the pacer in two ways:
* It replaces the PI controller used as a smoothing function with a
simple two-cycle moving average.
* It makes the pacer use the actual GC trigger point for cons/mark (and
other) calculations instead of the precomputed one.
The second part of this change was attempted in the Go 1.19 release
cycle, but could not be done because although it resulted in a
better-behaved pacer, it exploited the PI controller's sensitivity to
history in a way that was ultimately unfavorable for most applications.
This sensitivity is complex to reason about, and forces us into choices
that don't really make sense (like using the precomputed trigger over
the actual one -- that's really a bug fix).
The net effect of this change is intended to be:
* The pacer computes a more stable estimate of the actual cons/mark
ratio, making it easier to understand.
* The pacer is much more accurate at hitting the heap goal, so the GC
respects GOGC much more often and reliably.
* The pacer forces a more stable rate of GC assists in the
steady-state overall.
See https://perf.golang.org/search?q=upload:20221106.10 for benchmark
results and #53892 for complete context. The benchmarks that regress
in memory use appear to be not worth worrying about. In all cases, it appears that the GC was triggering super early, resulting in a lack
of adherence to rule of GOGC.
The fogleman benchmarks just have a single final GC that triggers
early and also happens to be the peak heap size. The tile38
WithinCircle benchmark only has 4 GC cycles in the benchmarked region,
so it's very sensitive to pacing. In this case, the old smoothing
function is getting lucky by starting way too early, avoiding assists. Meanwhile the 2-cycle moving average is more accurate on the heap
goal, but the 1st and 2nd cycle after the initialization phase are operating on a cons/mark from the initialization phase which is much
less active, resulting in a cons/mark that's too low, causing the GC
to start too late, increasing assists, and therefore latency (but
only transiently, at this phase change). I really do think the PI
controller is just getting lucky here with a particular history,
because I've definitely observed it oscillating wildly in response to
a phase change.
This change also moves the PI controller out of mgcpacer.go, because
it's no longer used there. It now lives in mgcscavenge.go, where it's
actually used.
Fixes#53892.
Change-Id: I3f875a2e40f31f381920f91d8b090556b17a2b16
Reviewed-on: https://go-review.googlesource.com/c/go/+/417558
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This changes adds a breakdown for estimated CPU usage by time. These
estimates are not based on real on-CPU counters, so each metric has a
disclaimer explaining so. They can, however, be more reasonably
compared to a total CPU time metric that this change also adds.
Fixes#47216.
Change-Id: I125006526be9f8e0d609200e193da5a78d9935be
Reviewed-on: https://go-review.googlesource.com/c/go/+/404307
Reviewed-by: Michael Pratt <mpratt@google.com>
Reviewed-by: Josh MacDonald <jmacd@lightstep.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This is an optimization that prevents N^2 behavior within a chunk, but
was accidentally skipped. There should be no functional change as a
result of this CL.
Fixes#54892.
Change-Id: I861967a2268699fdc3464bd41bc56618b5628e6b
Reviewed-on: https://go-review.googlesource.com/c/go/+/429255
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
This change forces mark and scavenge assists to be cancelled early if
the limiter is enabled. This avoids goroutines getting stuck in really
long assists if the limiter happens to be disabled when they first come
into the assist. This can get especially bad for mark assists, which, in
dire situations, can end up "owing" the GC a really significant debt.
For #52890.
Change-Id: I4bfaa76b8de3e167d49d2ffd8bc2127b87ea566a
Reviewed-on: https://go-review.googlesource.com/c/go/+/408816
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Currently the runtime's scavenging algorithm involves running from the
top of the heap address space to the bottom (or as far as it gets) once
per GC cycle. Once it treads some ground, it doesn't tread it again
until the next GC cycle.
This works just fine for the background scavenger, for heap-growth
scavenging, and for debug.FreeOSMemory. However, it breaks down in the
face of a memory limit for small heaps in the tens of MiB. Basically,
because the scavenger never retreads old ground, it's completely
oblivious to new memory it could scavenge, and that it really *should*
in the face of a memory limit.
Also, every time some thread goes to scavenge in the runtime, it
reserves what could be a considerable amount of address space, hiding it
from other scavengers.
This change modifies and simplifies the implementation overall. It's
less code with complexities that are much better encapsulated. The
current implementation iterates optimistically over the address space
looking for memory to scavenge, keeping track of what it last saw. The
new implementation does the same, but instead of directly iterating over
pages, it iterates over chunks. It maintains an index of chunks (as a
bitmap over the address space) that indicate which chunks may contain
scavenge work. The page allocator populates this index, while scavengers
consume it and iterate over it optimistically.
This has a two key benefits:
1. Scavenging is much simpler: find a candidate chunk, and check it,
essentially just using the scavengeOne fast path. There's no need for
the complexity of iterating beyond one chunk, because the index is
lock-free and already maintains that information.
2. If pages are freed to the page allocator (always guaranteed to be
unscavenged), the page allocator immediately notifies all scavengers
of the new source of work, avoiding the hiding issues of the old
implementation.
One downside of the new implementation, however, is that it's
potentially more expensive to find pages to scavenge. In the past, if
a single page would become free high up in the address space, the
runtime's scavengers would ignore it. Now that scavengers won't, one or
more scavengers may need to iterate potentially across the whole heap to
find the next source of work. For the background scavenger, this just
means a potentially less reactive scavenger -- overall it should still
use the same amount of CPU. It means worse overheads for memory limit
scavenging, but that's not exactly something with a baseline yet.
In practice, this shouldn't be too bad, hopefully since the chunk index
is extremely compact. For a 48-bit address space, the index is only 8
MiB in size at worst, but even just one physical page in the index is
able to support up to 128 GiB heaps, provided they aren't terribly
sparse. On 32-bit platforms, the index is only 128 bytes in size.
For #48409.
Change-Id: I72b7e74365046b18c64a6417224c5d85511194fb
Reviewed-on: https://go-review.googlesource.com/c/go/+/399474
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This change does everything necessary to make the memory allocator and
the scavenger respect the memory limit. In particular, it:
- Adds a second goal for the background scavenge that's based on the
memory limit, setting a target 5% below the limit to make sure it's
working hard when the application is close to it.
- Makes span allocation assist the scavenger if the next allocation is
about to put total memory use above the memory limit.
- Measures any scavenge assist time and adds it to GC assist time for
the sake of GC CPU limiting, to avoid a death spiral as a result of
scavenging too much.
All of these changes have a relatively small impact, but each is
intimately related and thus benefit from being done together.
For #48409.
Change-Id: I35517a752f74dd12a151dd620f102c77e095d3e8
Reviewed-on: https://go-review.googlesource.com/c/go/+/397017
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Fundamentally, all of these memstats exist to serve the runtime in
managing memory. For the sake of simpler testing, couple these stats
more tightly with the GC.
This CL was mostly done automatically. The fields had to be moved
manually, but the references to the fields were updated via
gofmt -w -r 'memstats.<field> -> gcController.<field>' *.go
For #48409.
Change-Id: Ic036e875c98138d9a11e1c35f8c61b784c376134
Reviewed-on: https://go-review.googlesource.com/c/go/+/397678
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
The inconsistent heaps stats in memstats are a bit messy. Primarily,
heap_sys is non-orthogonal with heap_released and heap_inuse. In later
CLs, we're going to want heap_sys-heap_released-heap_inuse, so clean
this up by replacing heap_sys with an orthogonal metric: heapFree.
heapFree represents page heap memory that is free but not released.
I think this change also simplifies a lot of reasoning about these
stats; it's much clearer what they mean, and to obtain HeapSys for
memstats, we no longer need to do the strange subtraction from heap_sys
when allocating specifically non-heap memory from the page heap.
Because we're removing heap_sys, we need to replace it with a sysMemStat
for mem.go functions. In this case, heap_released is the most
appropriate because we increase it anyway (again, non-orthogonality). In
which case, it makes sense for heap_inuse, heap_released, and heapFree
to become more uniform, and to just represent them all as sysMemStats.
While we're here and messing with the types of heap_inuse and
heap_released, let's also fix their names (and last_heap_inuse's name)
up to the more modern Go convention of camelCase.
For #48409.
Change-Id: I87fcbf143b3e36b065c7faf9aa888d86bd11710b
Reviewed-on: https://go-review.googlesource.com/c/go/+/397677
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This change refactors the scavenger into a type whose methods represent
the actual function and scheduling of the scavenger. It also stubs out
access to global state in order to make it testable.
This change thus also adds a test for the scavenger. In writing this
test, I discovered the lack of a behavior I expected: if the
pageAlloc.scavenge returns < the bytes requested scavenged, that means
the heap is exhausted. This has been true this whole time, but was not
documented or explicitly relied upon. This change rectifies that. In
theory this means the scavenger could spin in run() indefinitely (as
happened in the test) if shouldStop never told it to stop. In practice,
shouldStop fires long before the heap is exhausted, but for future
changes it may be important. At the very least it's good to be
intentional about these things.
While we're here, I also moved the call to stopTimer out of wake and
into sleep. There's no reason to add more operations to a context that's
already precarious (running without a P on sysmon).
Change-Id: Ib31b86379fd9df84f25ae282734437afc540da5c
Reviewed-on: https://go-review.googlesource.com/c/go/+/384734
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
If something goes horribly wrong with the assumptions surrounding a
piController, its internal error state might accumulate in an unbounded
manner. In practice this means unexpected Inf and NaN values.
Avoid this by identifying cases where the error overflows and resetting
controller state.
In the scavenger, this case is much more likely. All that has to happen
is the proportional relationship between sleep time and estimated CPU
usage has to break down. Unfortunately because we're just measuring
monotonic time for all this, there are lots of ways it could happen,
especially in an oversubscribed system. In these cases, just fall back
on a conservative pace for scavenging and try to wait out the issue.
In the pacer I'm pretty sure this is impossible. Because we wire the
output of the controller to the input, the response is very directly
correlated, so it's impossible for the controller's core assumption to
break down.
While we're in the pacer, add more detail about why that controller is
even there, as well as its purpose.
Finally, let's be proactive about other sources of overflow, namely
overflow from a very large input value. This change adds a check after
the first few operations to detect overflow issues from the input,
specifically the multiplication.
No tests for the pacer because I was unable to actually break the
pacer's controller under a fuzzer, and no tests for the scavenger because
it is not really in a testable state.
However:
* This change includes a fuzz test for the piController.
* I broke out the scavenger code locally and fuzz tested it, confirming
that the patch eliminates the original failure mode.
* I tested that on a local heap-spike test, the scavenger continues
operating as expected under normal conditions.
Fixes#51061.
Change-Id: I02a01d2dbf0eb9d2a8a8e7274d4165c2b6a3415a
Reviewed-on: https://go-review.googlesource.com/c/go/+/383954
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
And then revert the bootstrap cmd directories and certain testdata.
And adjust tests as needed.
Not reverting the changes in std that are bootstrapped,
because some of those changes would appear in API docs,
and we want to use any consistently.
Instead, rewrite 'any' to 'interface{}' in cmd/dist for those directories
when preparing the bootstrap copy.
A few files changed as a result of running gofmt -w
not because of interface{} -> any but because they
hadn't been updated for the new //go:build lines.
Fixes#49884.
Change-Id: Ie8045cba995f65bd79c694ec77a1b3d1fe01bb09
Reviewed-on: https://go-review.googlesource.com/c/go/+/368254
Trust: Russ Cox <rsc@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
When using faketime, only run the scavenger for one loop.
It tries to run for 1 ms, but with faketime that calculation fails.
Prohibit write barriers in the faketime write function, in case
the GC wants to print something (e.g., with GODEBUG=gctrace=1).
Fixes#49614
Change-Id: Iab5097fe78b6e0032ea8b493088264dfb25013c6
Reviewed-on: https://go-review.googlesource.com/c/go/+/364757
Trust: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Currently "released" is not accumulated bytes released. If the last
attempt to scavenge ends up as 0, then the scavenger will go to sleep
too soon. This is an artifact from the old code where scavenge would
only be called into once.
Change-Id: I85aa2261f1504a6fb5bf086daa029eecb0e09cf4
Reviewed-on: https://go-review.googlesource.com/c/go/+/363416
Trust: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Currently the scavenger is paced to 1% of 1 CPU because it had
scalability problems. As of the last few CLs, that should be largely
resolved. This change resolves the TODO and paces the scavenger
according to 1% of overall CPU time.
This change is made separately to allow it to be more easily rolled
back.
Change-Id: I1ab4de24ba41c564960701634a128a813c55ece9
Reviewed-on: https://go-review.googlesource.com/c/go/+/358675
Trust: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Currently the scavenge rate is determined by a bunch of ad-hoc
mechanisms. Just use a controller instead, now that we have one.
To facilitate this, the scavenger now attempts to scavenge for at least
1 ms at a time, because any less and the timer system is too imprecise to
give useful feedback to the controller. Also increase the amount that we
scavenge at once, to try to reduce the overheads involved (at the
expense of a little bit of latency).
This change also modifies the controller to accept an update period,
because it's useful to allow that to be variable.
Change-Id: I8a15b2355d0a7c6cbac68c957082d5819618f7d7
Reviewed-on: https://go-review.googlesource.com/c/go/+/353975
Trust: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
This change modifies the scavenger to no longer hold the heap lock while
actively scavenging pages. To achieve this, the change also:
* Reverses the locking behavior of the (*pageAlloc).scavenge API, to
only acquire the heap lock when necessary.
* Introduces a new lock on the scavenger-related fields in a pageAlloc
so that access to those fields doesn't require the heap lock. There
are a few places in the scavenge path, notably reservation, that
requires synchronization. The heap lock is far too heavy handed for
this case.
* Changes the scavenger to marks pages that are actively being scavenged
as allocated, and "frees" them back to the page allocator the usual
way.
* Lifts the heap-growth scavenging code out of mheap.grow, where the
heap lock is held, and into allocSpan, just after the lock is
released. Releasing the lock during mheap.grow is not feasible if we
want to ensure that allocation always makes progress (post-growth,
another allocator could come in and take all that space, forcing the
goroutine that just grew the heap to do so again).
This change means that the scavenger now must do more work for each
scavenge, but it is also now much more scalable. Although in theory it's
not great by always taking the locked paths in the page allocator, it
takes advantage of some properties of the allocator:
* Most of the time, the scavenger will be working with one page at a
time. The page allocator's locked path is optimized for this case.
* On the allocation path, it doesn't need to do the find operation at
all; it can go straight to setting bits for the range and updating the
summary structure.
Change-Id: Ie941d5e7c05dcc96476795c63fef74bcafc2a0f1
Reviewed-on: https://go-review.googlesource.com/c/go/+/353974
Trust: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
The first step toward acquiring the heap lock less frequently in the
scavenger.
Change-Id: Idc69fd8602be2c83268c155951230d60e20b42fe
Reviewed-on: https://go-review.googlesource.com/c/go/+/353973
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Currently gcPaceScavenger is called by gcControllerState.commit, but it
manipulates global state which precludes testing. This change detangles
the two.
Change-Id: I10d8ebdf426d99ba49d2f2cb4fb64891e9fd6091
Reviewed-on: https://go-review.googlesource.com/c/go/+/309272
Reviewed-by: Michael Pratt <mpratt@google.com>
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
In CL 298669 we added defer/go wrapping, and, as it is not
allowed for closures to escape when compiling runtime, we worked
around it by rewriting go'd closures to argumentless
non-capturing closures, so it is not a real closure and so not
needed to escape.
Previous CL removes the restriction. Now we can undo the
workaround.
Updates #40724.
Change-Id: Ic7bf129da4aee7b7fdb7157414eca943a6a27264
Reviewed-on: https://go-review.googlesource.com/c/go/+/325110
Trust: Cherry Mui <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Currently there's a minor bug where the constant for the min fraction of
time spent scavenging is rounded down to zero. I don't think this
affects anything in practice because this case is exceedingly rare and
extreme, but currently it doesn't properly prevent the pacing parameters
from getting out of hand in these extreme cases.
Fixes#44036.
Change-Id: I7de644ab0ecac33765c337a736482a0966882780
Reviewed-on: https://go-review.googlesource.com/c/go/+/313249
Reviewed-by: Michael Pratt <mpratt@google.com>
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
This change moves next_gc and last_next_gc into gcControllerState under
the names heapGoal and lastHeapGoal respectively. These are
fundamentally GC pacer related values, and so it makes sense for them to
live here.
Partially generated by
rf '
ex . {
memstats.next_gc -> gcController.heapGoal
memstats.last_next_gc -> gcController.lastHeapGoal
}
'
except for updates to comments and gcControllerState methods, where
they're accessed through the receiver, and trace-related renames of
NextGC -> HeapGoal, while we're here.
For #44167.
Change-Id: I1e871ad78a57b01be8d9f71bd662530c84853bed
Reviewed-on: https://go-review.googlesource.com/c/go/+/306603
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Adds code to the compiler's "order" phase to rewrite go and defer
statements to always be argument-less. E.g.
defer f(x,y) => x1, y1 := x, y
defer func() { f(x1, y1) }
This transformation is not beneficial on its own, but it helps
simplify runtime defer handling for the new register ABI (when
invoking deferred functions on the panic path, the runtime doesn't
need to manage the complexity of determining which args to pass in
register vs memory).
This feature is currently enabled by default if GOEXPERIMENT=regabi or
GOEXPERIMENT=regabidefer is in effect.
Included in this CL are some workarounds in the runtime to insure that
"go" statement targets in the runtime are argument-less already (since
wrapping them can potentially introduce heap-allocated closures, which
are currently not allowed). The expectation is that these workarounds
will be temporary, and can go away once we either A) change the rules
about heap-allocated closures, or B) implement some other scheme for
handling go statements.
Change-Id: I01060d79a6b140c6f0838d6e6813f807ccdca319
Reviewed-on: https://go-review.googlesource.com/c/go/+/298669
Trust: Than McIntosh <thanm@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
This change modifies the consistent stats implementation to keep the
per-P sequence counter on each P instead of each mcache. A valid mcache
is not available everywhere that we want to call e.g. allocSpan, as per
issue #42339. By decoupling these two, we can add a mechanism to allow
contexts without a P to update stats consistently.
In this CL, we achieve that with a mutex. In practice, it will be very
rare for an M to update these stats without a P. Furthermore, the stats
reader also only needs to hold the mutex across the update to "gen"
since once that changes, writers are free to continue updating the new
stats generation. Contention could thus only arise between writers
without a P, and as mentioned earlier, those should be rare.
A nice side-effect of this change is that the consistent stats acquire
and release API becomes simpler.
Fixes#42339.
Change-Id: Ied74ab256f69abd54b550394c8ad7c4c40a5fe34
Reviewed-on: https://go-review.googlesource.com/c/go/+/267158
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Trust: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
This change moves the responsibility of throwing if an mcache is not
available to the caller, because the inlining cost of throw is set very
high in the compiler. Even if it was reduced down to the cost of a usual
function call, it would still be too expensive, so just move it out.
This choice also makes sense in the context of #42339 since we're going
to have to handle the case where we don't have an mcache to update stats
in a few contexts anyhow.
Also, add getMCache to the list of functions that should be inlined to
prevent future regressions.
getMCache is called on the allocation fast path and because its not
inlined actually causes a significant regression (~10%) in some
microbenchmarks.
Fixes#42305.
Change-Id: I64ac5e4f26b730bd4435ea1069a4a50f55411ced
Reviewed-on: https://go-review.googlesource.com/c/go/+/267157
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Some functions that required holding the heap lock _or_ world stop have
been simplified to simply requiring the heap lock. This is conceptually
simpler and taking the heap lock during world stop is guaranteed to not
contend. This was only done on functions already called on the
systemstack to avoid too many extra systemstack calls in GC.
Updates #40677
Change-Id: I15aa1dadcdd1a81aac3d2a9ecad6e7d0377befdc
Reviewed-on: https://go-review.googlesource.com/c/go/+/250262
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Trust: Michael Pratt <mpratt@google.com>
