While we're here, update the documentation and delete variables with no effect.
Change-Id: I4df0d266dff880df61b488ed547c2870205862f0
Reviewed-on: https://go-review.googlesource.com/10790
Reviewed-by: Austin Clements <austin@google.com>
These were found by grepping the comments from the go code and feeding
the output to aspell.
Change-Id: Id734d6c8d1938ec3c36bd94a4dbbad577e3ad395
Reviewed-on: https://go-review.googlesource.com/10941
Reviewed-by: Aamir Khan <syst3m.w0rm@gmail.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Stack barriers assume that writes through pointers to frames above the
current frame will get write barriers, and hence these frames do not
need to be re-scanned to pick up these changes. For normal writes,
this is true. However, there are places in the runtime that use
typedmemmove to potentially write through pointers to higher frames
(such as mapassign1). Currently, typedmemmove does not execute write
barriers if the destination is on the stack. If there's a stack
barrier between the current frame and the frame being modified with
typedmemmove, and the stack barrier is not otherwise hit, it's
possible that the garbage collector will never see the updated pointer
and incorrectly reclaim the object.
Fix this by making heapBitsBulkBarrier (which lies behind typedmemmove
and its variants) detect when the destination is in the stack and
unwind stack barriers up to the point, forcing mark termination to
later rescan the effected frame and collect these pointers.
Fixes#11084. Might be related to #10240, #10541, #10941, #11023,
#11027 and possibly others.
Change-Id: I323d6cd0f1d29fa01f8fc946f4b90e04ef210efd
Reviewed-on: https://go-review.googlesource.com/10791
Reviewed-by: Russ Cox <rsc@golang.org>
Currently the stack barriers are installed at the next frame boundary
after gp.sched.sp + 1024*2^n for n=0,1,2,... However, when a G is in a
system call, we set gp.sched.sp to 0, which causes stack barriers to
be installed at *every* frame. This easily overflows the slice we've
reserved for storing the stack barrier information, and causes a
"slice bounds out of range" panic in gcInstallStackBarrier.
Fix this by using gp.syscallsp instead of gp.sched.sp if it's
non-zero. This is the same logic that gentraceback uses to determine
the current SP.
Fixes#11049.
Change-Id: Ie40eeee5bec59b7c1aa715a7c17aa63b1f1cf4e8
Reviewed-on: https://go-review.googlesource.com/10755
Reviewed-by: Russ Cox <rsc@golang.org>
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes#10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently there's a race between stopg scanning another G's stack and
the G reaching a preemption point and scanning its own stack. When
this race occurs, the G's stack is scanned twice. Currently this is
okay, so this race is benign.
However, we will shortly be adding stack barriers during the first
stack scan, so scanning will no longer be idempotent. To prepare for
this, this change ensures that each stack is scanned only once during
each GC phase by checking the flag that indicates that the stack has
been scanned in this phase before scanning the stack.
Change-Id: Id9f4d5e2e5b839bc3f200ec1723a4a12dd677ab4
Reviewed-on: https://go-review.googlesource.com/10458
Reviewed-by: Rick Hudson <rlh@golang.org>
The stack barrier code will need a bookkeeping structure to keep track
of the overwritten return PCs. This commit introduces and allocates
this structure, but does not yet use the structure.
We don't want to allocate space for this structure during garbage
collection, so this commit allocates it along with the allocation of
the corresponding stack. However, we can't do a regular allocation in
newstack because mallocgc may itself grow the stack (which would lead
to a recursive allocation). Hence, this commit makes the bookkeeping
structure part of the stack allocation itself by stealing the
necessary space from the top of the stack allocation. Since the size
of this bookkeeping structure is logarithmic in the size of the stack,
this has minimal impact on stack behavior.
Change-Id: Ia14408be06aafa9ca4867f4e70bddb3fe0e96665
Reviewed-on: https://go-review.googlesource.com/10313
Reviewed-by: Russ Cox <rsc@golang.org>
This is dead code. If you want to quiesce the system the
preferred way is to use forEachP(func(*p){}).
Change-Id: Ic7677a5dd55e3639b99e78ddeb2c71dd1dd091fa
Reviewed-on: https://go-review.googlesource.com/10267
Reviewed-by: Austin Clements <austin@google.com>
For the conversion of the heap bitmap from 4-bit to 2-bit fields,
I replaced heapBitsSetType with the dumbest thing that could possibly work:
two atomic operations (atomicand8+atomicor8) per 2-bit field.
This CL replaces that code with a proper implementation that
avoids the atomics whenever possible. Benchmarks vs base CL
(before the conversion to 2-bit heap bitmap) and vs Go 1.4 below.
Compared to Go 1.4, SetTypePtr (a 1-pointer allocation)
is 10ns slower because a race against the concurrent GC requires the
use of an atomicor8 that used to be an ordinary write. This slowdown
was present even in the base CL.
Compared to both Go 1.4 and base, SetTypeNode8 (a 10-word allocation)
is 10ns slower because it too needs a new atomic, because with the
denser representation, the byte on the end of the allocation is now shared
with the object next to it; this was not true with the 4-bit representation.
Excluding these two (fundamental) slowdowns due to the use of atomics,
the new code is noticeably faster than both Go 1.4 and the base CL.
The next CL will reintroduce the ``typeDead'' optimization.
Stats are from 5 runs on a MacBookPro10,2 (late 2012 Core i5).
Compared to base CL (** = new atomic)
name old mean new mean delta
SetTypePtr 14.1ns × (0.99,1.02) 14.7ns × (0.93,1.10) ~ (p=0.175)
SetTypePtr8 18.4ns × (1.00,1.01) 18.6ns × (0.81,1.21) ~ (p=0.866)
SetTypePtr16 28.7ns × (1.00,1.00) 22.4ns × (0.90,1.27) -21.88% (p=0.015)
SetTypePtr32 52.3ns × (1.00,1.00) 33.8ns × (0.93,1.24) -35.37% (p=0.001)
SetTypePtr64 79.2ns × (1.00,1.00) 55.1ns × (1.00,1.01) -30.43% (p=0.000)
SetTypePtr126 118ns × (1.00,1.00) 100ns × (1.00,1.00) -15.97% (p=0.000)
SetTypePtr128 130ns × (0.92,1.19) 98ns × (1.00,1.00) -24.36% (p=0.008)
SetTypePtrSlice 726ns × (0.96,1.08) 760ns × (1.00,1.00) ~ (p=0.152)
SetTypeNode1 14.1ns × (0.94,1.15) 12.0ns × (1.00,1.01) -14.60% (p=0.020)
SetTypeNode1Slice 135ns × (0.96,1.07) 88ns × (1.00,1.00) -34.53% (p=0.000)
SetTypeNode8 20.9ns × (1.00,1.01) 32.6ns × (1.00,1.00) +55.37% (p=0.000) **
SetTypeNode8Slice 414ns × (0.99,1.02) 244ns × (1.00,1.00) -41.09% (p=0.000)
SetTypeNode64 80.0ns × (1.00,1.00) 57.4ns × (1.00,1.00) -28.23% (p=0.000)
SetTypeNode64Slice 2.15µs × (1.00,1.01) 1.56µs × (1.00,1.00) -27.43% (p=0.000)
SetTypeNode124 119ns × (0.99,1.00) 100ns × (1.00,1.00) -16.11% (p=0.000)
SetTypeNode124Slice 3.40µs × (1.00,1.00) 2.93µs × (1.00,1.00) -13.80% (p=0.000)
SetTypeNode126 120ns × (1.00,1.01) 98ns × (1.00,1.00) -18.19% (p=0.000)
SetTypeNode126Slice 3.53µs × (0.98,1.08) 3.02µs × (1.00,1.00) -14.49% (p=0.002)
SetTypeNode1024 726ns × (0.97,1.09) 740ns × (1.00,1.00) ~ (p=0.451)
SetTypeNode1024Slice 24.9µs × (0.89,1.37) 23.1µs × (1.00,1.00) ~ (p=0.476)
Compared to Go 1.4 (** = new atomic)
name old mean new mean delta
SetTypePtr 5.71ns × (0.89,1.19) 14.68ns × (0.93,1.10) +157.24% (p=0.000) **
SetTypePtr8 19.3ns × (0.96,1.10) 18.6ns × (0.81,1.21) ~ (p=0.638)
SetTypePtr16 30.7ns × (0.99,1.03) 22.4ns × (0.90,1.27) -26.88% (p=0.005)
SetTypePtr32 51.5ns × (1.00,1.00) 33.8ns × (0.93,1.24) -34.40% (p=0.001)
SetTypePtr64 83.6ns × (0.94,1.12) 55.1ns × (1.00,1.01) -34.12% (p=0.001)
SetTypePtr126 137ns × (0.87,1.26) 100ns × (1.00,1.00) -27.10% (p=0.028)
SetTypePtrSlice 865ns × (0.80,1.23) 760ns × (1.00,1.00) ~ (p=0.243)
SetTypeNode1 15.2ns × (0.88,1.12) 12.0ns × (1.00,1.01) -20.89% (p=0.014)
SetTypeNode1Slice 156ns × (0.93,1.16) 88ns × (1.00,1.00) -43.57% (p=0.001)
SetTypeNode8 23.8ns × (0.90,1.18) 32.6ns × (1.00,1.00) +36.76% (p=0.003) **
SetTypeNode8Slice 502ns × (0.92,1.10) 244ns × (1.00,1.00) -51.46% (p=0.000)
SetTypeNode64 85.6ns × (0.94,1.11) 57.4ns × (1.00,1.00) -32.89% (p=0.001)
SetTypeNode64Slice 2.36µs × (0.91,1.14) 1.56µs × (1.00,1.00) -33.96% (p=0.002)
SetTypeNode124 130ns × (0.91,1.12) 100ns × (1.00,1.00) -23.49% (p=0.004)
SetTypeNode124Slice 3.81µs × (0.90,1.22) 2.93µs × (1.00,1.00) -23.09% (p=0.025)
There are fewer benchmarks vs Go 1.4 because unrolling directly
into the heap bitmap is not yet implemented, so those would not
be meaningful comparisons.
These benchmarks were not present in Go 1.4 as distributed.
The backport to Go 1.4 is in github.com/rsc/go's go14bench branch,
commit 71d5ee5.
Change-Id: I95ed05a22bf484b0fc9efad549279e766c98d2b6
Reviewed-on: https://go-review.googlesource.com/9704
Reviewed-by: Rick Hudson <rlh@golang.org>
Previous CLs changed the representation of the non-heap type bitmaps
to be 1-bit bitmaps (pointer or not). Before this CL, the heap bitmap
stored a 2-bit type for each word and a mark bit and checkmark bit
for the first word of the object. (There used to be additional per-word bits.)
Reduce heap bitmap to 2-bit, with 1 dedicated to pointer or not,
and the other used for mark, checkmark, and "keep scanning forward
to find pointers in this object." See comments for details.
This CL replaces heapBitsSetType with very slow but obviously correct code.
A followup CL will optimize it. (Spoiler: the new code is faster than Go 1.4 was.)
Change-Id: I999577a133f3cfecacebdec9cdc3573c235c7fb9
Reviewed-on: https://go-review.googlesource.com/9703
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
It was testing the mark bits on what roots pointed at,
but not the remainder of the live heap, because in
CL 2991 I accidentally inverted this check during
refactoring.
The next CL will turn it back off by default again,
but I want one run on the builders with the full
checkmark checks.
Change-Id: Ic166458cea25c0a56e5387fc527cb166ff2e5ada
Reviewed-on: https://go-review.googlesource.com/9824
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
This tracks the number of scannable bytes in the allocated heap. That
is, bytes that the garbage collector must scan before reaching the
last pointer field in each object.
This will be used to compute a more robust estimate of the GC scan
work.
Change-Id: I1eecd45ef9cdd65b69d2afb5db5da885c80086bb
Reviewed-on: https://go-review.googlesource.com/9695
Reviewed-by: Russ Cox <rsc@golang.org>
The garbage collector predicts how much "scan work" must be done in a
cycle to determine how much work should be done by mutators when they
allocate. Most code doesn't care what units the scan work is in: it
simply knows that a certain amount of scan work has to be done in the
cycle. Currently, the GC uses the number of pointer slots scanned as
the scan work on the theory that this is the bulk of the time spent in
the garbage collector and hence reflects real CPU resource usage.
However, this metric is difficult to estimate at the beginning of a
cycle.
Switch to counting the total number of bytes scanned, including both
pointer and scalar slots. This is still less than the total marked
heap since it omits no-scan objects and no-scan tails of objects. This
metric may not reflect absolute performance as well as the count of
scanned pointer slots (though it still takes time to scan scalar
fields), but it will be much easier to estimate robustly, which is
more important.
Change-Id: Ie3a5eeeb0384a1ca566f61b2f11e9ff3a75ca121
Reviewed-on: https://go-review.googlesource.com/9694
Reviewed-by: Russ Cox <rsc@golang.org>
(1) Count pointer-free objects found during scanning roots
as marked bytes, by not zeroing the mark total after scanning roots.
(2) Don't count the bytes for the roots themselves, by not adding
them to the mark total in scanblock (the zeroing removed by (1)
was aimed at that add but hitting more).
Combined, (1) and (2) fix the calculation of the marked heap size.
This makes the GC trigger much less often in the Go 1 benchmarks,
which have a global []byte pointing at 256 MB of data.
That 256 MB allocation was not being included in the heap size
in the current code, but was included in Go 1.4.
This is the source of much of the relative slowdown in that directory.
(3) Count the bytes for the roots as scanned work, by not zeroing
the scan total after scanning roots. There is no strict justification
for this, and it probably doesn't matter much either way,
but it was always combined with another buggy zeroing
(removed in (1)), so guilty by association.
Austin noticed this.
name old mean new mean delta
BenchmarkBinaryTree17 13.1s × (0.97,1.03) 5.9s × (0.97,1.05) -55.19% (p=0.000)
BenchmarkFannkuch11 4.35s × (0.99,1.01) 4.37s × (1.00,1.01) +0.47% (p=0.032)
BenchmarkFmtFprintfEmpty 84.6ns × (0.95,1.14) 85.7ns × (0.94,1.05) ~ (p=0.521)
BenchmarkFmtFprintfString 320ns × (0.95,1.06) 283ns × (0.99,1.02) -11.48% (p=0.000)
BenchmarkFmtFprintfInt 311ns × (0.98,1.03) 288ns × (0.99,1.02) -7.26% (p=0.000)
BenchmarkFmtFprintfIntInt 554ns × (0.96,1.05) 478ns × (0.99,1.02) -13.70% (p=0.000)
BenchmarkFmtFprintfPrefixedInt 434ns × (0.96,1.06) 393ns × (0.98,1.04) -9.60% (p=0.000)
BenchmarkFmtFprintfFloat 620ns × (0.99,1.03) 584ns × (0.99,1.01) -5.73% (p=0.000)
BenchmarkFmtManyArgs 2.19µs × (0.98,1.03) 1.94µs × (0.99,1.01) -11.62% (p=0.000)
BenchmarkGobDecode 21.2ms × (0.97,1.06) 15.2ms × (0.99,1.01) -28.17% (p=0.000)
BenchmarkGobEncode 18.1ms × (0.94,1.06) 11.8ms × (0.99,1.01) -35.00% (p=0.000)
BenchmarkGzip 650ms × (0.98,1.01) 649ms × (0.99,1.02) ~ (p=0.802)
BenchmarkGunzip 143ms × (1.00,1.01) 143ms × (1.00,1.01) ~ (p=0.438)
BenchmarkHTTPClientServer 110µs × (0.98,1.04) 101µs × (0.98,1.02) -8.79% (p=0.000)
BenchmarkJSONEncode 40.3ms × (0.97,1.03) 31.8ms × (0.98,1.03) -20.92% (p=0.000)
BenchmarkJSONDecode 119ms × (0.97,1.02) 108ms × (0.99,1.02) -9.15% (p=0.000)
BenchmarkMandelbrot200 6.03ms × (1.00,1.01) 6.03ms × (0.99,1.01) ~ (p=0.750)
BenchmarkGoParse 8.58ms × (0.89,1.10) 6.80ms × (1.00,1.00) -20.71% (p=0.000)
BenchmarkRegexpMatchEasy0_32 162ns × (1.00,1.01) 162ns × (0.99,1.02) ~ (p=0.131)
BenchmarkRegexpMatchEasy0_1K 540ns × (0.99,1.02) 559ns × (0.99,1.02) +3.58% (p=0.000)
BenchmarkRegexpMatchEasy1_32 139ns × (0.98,1.04) 139ns × (1.00,1.00) ~ (p=0.466)
BenchmarkRegexpMatchEasy1_1K 889ns × (0.99,1.01) 885ns × (0.99,1.01) -0.50% (p=0.022)
BenchmarkRegexpMatchMedium_32 252ns × (0.99,1.02) 252ns × (0.99,1.01) ~ (p=0.469)
BenchmarkRegexpMatchMedium_1K 72.9µs × (0.99,1.01) 73.6µs × (0.99,1.03) ~ (p=0.168)
BenchmarkRegexpMatchHard_32 3.87µs × (1.00,1.01) 3.86µs × (1.00,1.00) ~ (p=0.055)
BenchmarkRegexpMatchHard_1K 118µs × (0.99,1.01) 117µs × (0.99,1.00) ~ (p=0.133)
BenchmarkRevcomp 995ms × (0.94,1.10) 949ms × (0.99,1.01) -4.64% (p=0.000)
BenchmarkTemplate 141ms × (0.97,1.02) 127ms × (0.99,1.01) -10.00% (p=0.000)
BenchmarkTimeParse 641ns × (0.99,1.01) 623ns × (0.99,1.01) -2.79% (p=0.000)
BenchmarkTimeFormat 729ns × (0.98,1.03) 679ns × (0.99,1.00) -6.93% (p=0.000)
Change-Id: I839bd7356630d18377989a0748763414e15ed057
Reviewed-on: https://go-review.googlesource.com/9602
Reviewed-by: Austin Clements <austin@google.com>
gcDumpObject is used to print the source and destination objects when
checkmark find a missing mark. However, gcDumpObject currently assumes
the given pointer will point to a heap object. This is not true of the
source object during root marking and may not even be true of the
destination object in the limited situations where the heap points
back in to the stack.
If the pointer isn't a heap object, gcDumpObject will attempt an
out-of-bounds access to h_spans. This will cause a panicslice, which
will attempt to construct a useful panic message. This will cause a
string allocation, which will lead mallocgc to panic because the GC is
in mark termination (checkmark only happens during mark termination).
Fix this by checking that the pointer points into the heap arena
before attempting to use it as an arena pointer.
Change-Id: I09da600c380d4773f1f8f38e45b82cb229ea6382
Reviewed-on: https://go-review.googlesource.com/9498
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, we use a full stop-the-world around enabling write
barriers. This is to ensure that all Gs have enabled write barriers
before any blackening occurs (either in gcBgMarkWorker() or in
gcAssistAlloc()).
However, there's no need to bring the whole world to a synchronous
stop to ensure this. This change replaces the STW with a ragged
barrier that ensures each P has individually observed that write
barriers should be enabled before GC performs any blackening.
Change-Id: If2f129a6a55bd8bdd4308067af2b739f3fb41955
Reviewed-on: https://go-review.googlesource.com/8207
Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, each M has a cache of the most recently used *workbuf. This
is used primarily by the write barrier so it doesn't have to access
the global workbuf lists on every write barrier. It's also used by
stack scanning because it's convenient.
This cache is important for write barrier performance, but this
particular approach has several downsides. It's faster than no cache,
but far from optimal (as the benchmarks below show). It's complex:
access to the cache is sprinkled through most of the workbuf list
operations and it requires special care to transform into and back out
of the gcWork cache that's actually used for scanning and marking. It
requires atomic exchanges to take ownership of the cached workbuf and
to return it to the M's cache even though it's almost always used by
only the current M. Since it's per-M, flushing these caches is O(# of
Ms), which may be high. And it has some significant subtleties: for
example, in general the cache shouldn't be used after the
harvestwbufs() in mark termination because it could hide work from
mark termination, but stack scanning can happen after this and *will*
use the cache (but it turns out this is okay because it will always be
followed by a getfull(), which drains the cache).
This change replaces this cache with a per-P gcWork object. This
gcWork cache can be used directly by scanning and marking (as long as
preemption is disabled, which is a general requirement of gcWork).
Since it's per-P, it doesn't require synchronization, which simplifies
things and means the only atomic operations in the write barrier are
occasionally fetching new work buffers and setting a mark bit if the
object isn't already marked. This cache can be flushed in O(# of Ps),
which is generally small. It follows a simple flushing rule: the cache
can be used during any phase, but during mark termination it must be
flushed before allowing preemption. This also makes the dispose during
mutator assist no longer necessary, which eliminates the vast majority
of gcWork dispose calls and reduces contention on the global workbuf
lists. And it's a lot faster on some benchmarks:
benchmark old ns/op new ns/op delta
BenchmarkBinaryTree17 11963668673 11206112763 -6.33%
BenchmarkFannkuch11 2643217136 2649182499 +0.23%
BenchmarkFmtFprintfEmpty 70.4 70.2 -0.28%
BenchmarkFmtFprintfString 364 307 -15.66%
BenchmarkFmtFprintfInt 317 282 -11.04%
BenchmarkFmtFprintfIntInt 512 483 -5.66%
BenchmarkFmtFprintfPrefixedInt 404 380 -5.94%
BenchmarkFmtFprintfFloat 521 479 -8.06%
BenchmarkFmtManyArgs 2164 1894 -12.48%
BenchmarkGobDecode 30366146 22429593 -26.14%
BenchmarkGobEncode 29867472 26663152 -10.73%
BenchmarkGzip 391236616 396779490 +1.42%
BenchmarkGunzip 96639491 96297024 -0.35%
BenchmarkHTTPClientServer 100110 70763 -29.31%
BenchmarkJSONEncode 51866051 52511382 +1.24%
BenchmarkJSONDecode 103813138 86094963 -17.07%
BenchmarkMandelbrot200 4121834 4120886 -0.02%
BenchmarkGoParse 16472789 5879949 -64.31%
BenchmarkRegexpMatchEasy0_32 140 140 +0.00%
BenchmarkRegexpMatchEasy0_1K 394 394 +0.00%
BenchmarkRegexpMatchEasy1_32 120 120 +0.00%
BenchmarkRegexpMatchEasy1_1K 621 614 -1.13%
BenchmarkRegexpMatchMedium_32 209 202 -3.35%
BenchmarkRegexpMatchMedium_1K 54889 55175 +0.52%
BenchmarkRegexpMatchHard_32 2682 2675 -0.26%
BenchmarkRegexpMatchHard_1K 79383 79524 +0.18%
BenchmarkRevcomp 584116718 584595320 +0.08%
BenchmarkTemplate 125400565 109620196 -12.58%
BenchmarkTimeParse 386 387 +0.26%
BenchmarkTimeFormat 580 447 -22.93%
(Best out of 10 runs. The delta of averages is similar.)
This also puts us in a good position to flush these caches when
nearing the end of concurrent marking, which will let us increase the
size of the work buffers while still controlling mark termination
pause time.
Change-Id: I2dd94c8517a19297a98ec280203cccaa58792522
Reviewed-on: https://go-review.googlesource.com/9178
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Currently, the main GC goroutine sleeps on a note during concurrent
mark and the first background mark worker or assist to finish marking
use wakes up that note to let the main goroutine proceed into mark
termination. Unfortunately, the latency of this wakeup can be quite
high, since the GC goroutine will typically have lost its P while in
the futex sleep, meaning it will be placed on the global run queue and
will wait there until some P is kind enough to pick it up. This delay
gives the mutator more time to allocate and create floating garbage,
growing the heap unnecessarily. Worse, it's likely that background
marking has stopped at this point (unless GOMAXPROCS>4), so anything
that's allocated and published to the heap during this window will
have to be scanned during mark termination while the world is stopped.
This change replaces the note sleep/wakeup with a gopark/ready
scheme. This keeps the wakeup inside the Go scheduler and lets the
garbage collector take advantage of the new scheduler semantics that
run the ready()d goroutine immediately when the ready()ing goroutine
sleeps.
For the json benchmark from x/benchmarks with GOMAXPROCS=4, this
reduces the delay in waking up the GC goroutine and entering mark
termination once concurrent marking is done from ~100ms to typically
<100µs.
Change-Id: Ib11f8b581b8914f2d68e0094f121e49bac3bb384
Reviewed-on: https://go-review.googlesource.com/9291
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
To achieve a 2% improvement in the garbage benchmark this CL removes
an unneeded assert and avoids one hbits.next() call per object
being scanned.
Change-Id: Ibd542d01e9c23eace42228886f9edc488354df0d
Reviewed-on: https://go-review.googlesource.com/9244
Reviewed-by: Austin Clements <austin@google.com>
Currently, the concurrent mark phase is performed by the main GC
goroutine. Prior to the previous commit enabling preemption, this
caused marking to always consume 1/GOMAXPROCS of the available CPU
time. If GOMAXPROCS=1, this meant background GC would consume 100% of
the CPU (effectively a STW). If GOMAXPROCS>4, background GC would use
less than the goal of 25%. If GOMAXPROCS=4, background GC would use
the goal 25%, but if the mutator wasn't using the remaining 75%,
background marking wouldn't take advantage of the idle time. Enabling
preemption in the previous commit made GC miss CPU targets in
completely different ways, but set us up to bring everything back in
line.
This change replaces the fixed GC goroutine with per-P background mark
goroutines. Once started, these goroutines don't go in the standard
run queues; instead, they are scheduled specially such that the time
spent in mutator assists and the background mark goroutines totals 25%
of the CPU time available to the program. Furthermore, this lets
background marking take advantage of idle Ps, which significantly
boosts GC performance for applications that under-utilize the CPU.
This requires also changing how time is reported for gctrace, so this
change splits the concurrent mark CPU time into assist/background/idle
scanning.
This also requires increasing the size of the StackRecord slice used
in a GoroutineProfile test.
Change-Id: I0936ff907d2cee6cb687a208f2df47e8988e3157
Reviewed-on: https://go-review.googlesource.com/8850
Reviewed-by: Rick Hudson <rlh@golang.org>
This time is tracked per P and periodically flushed to the global
controller state. This will be used to compute mutator assist
utilization in order to schedule background GC work.
Change-Id: Ib94f90903d426a02cf488bf0e2ef67a068eb3eec
Reviewed-on: https://go-review.googlesource.com/8837
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, mutator allocation periodically assists the garbage
collector by performing a small, fixed amount of scanning work.
However, to control heap growth, mutators need to perform scanning
work *proportional* to their allocation rate.
This change implements proportional mutator assists. This uses the
scan work estimate computed by the garbage collector at the beginning
of each cycle to compute how much scan work must be performed per
allocation byte to complete the estimated scan work by the time the
heap reaches the goal size. When allocation triggers an assist, it
uses this ratio and the amount allocated since the last assist to
compute the assist work, then attempts to steal as much of this work
as possible from the background collector's credit, and then performs
any remaining scan work itself.
Change-Id: I98b2078147a60d01d6228b99afd414ef857e4fba
Reviewed-on: https://go-review.googlesource.com/8836
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, the "n" in gcDrainN is in terms of objects to scan. This is
used by gchelpwork to perform a limited amount of work on allocation,
but is a pretty arbitrary way to bound this amount of work since the
number of objects has little relation to how long they take to scan.
Modify gcDrainN to perform a fixed amount of scan work instead. For
now, gchelpwork still performs a fairly arbitrary amount of scan work,
but at least this is much more closely related to how long the work
will take. Shortly, we'll use this to precisely control the scan work
performed by mutator assists during allocation to achieve the heap
size goal.
Change-Id: I3cd07fe0516304298a0af188d0ccdf621d4651cc
Reviewed-on: https://go-review.googlesource.com/8835
Reviewed-by: Rick Hudson <rlh@golang.org>
This tracks scan work done by background GC in a global pool. Mutator
assists will draw on this credit to avoid doing work when background
GC is staying ahead.
Unlike the other GC controller tracking variables, this will be both
written and read throughout the cycle. Hence, we can't arbitrarily
delay updates like we can for scan work and bytes marked. However, we
still want to minimize contention, so this global credit pool is
allowed some error from the "true" amount of credit. Background GC
accumulates credit locally up to a limit and only then flushes to the
global pool. Similarly, mutator assists will draw from the credit pool
in batches.
Change-Id: I1aa4fc604b63bf53d1ee2a967694dffdfc3e255e
Reviewed-on: https://go-review.googlesource.com/8834
Reviewed-by: Rick Hudson <rlh@golang.org>
This tracks the amount of scan work in terms of scanned pointers
during the concurrent mark phase. We'll use this information to
estimate scan work for the next cycle.
Currently this aggregates the work counter in gcWork and dispose
atomically aggregates this into a global work counter. dispose happens
relatively infrequently, so the contention on the global counter
should be low. If this turns out to be an issue, we can reduce the
number of disposes, and if it's still a problem, we can switch to
per-P counters.
Change-Id: Iac0364c466ee35fab781dbbbe7970a5f3c4e1fc1
Reviewed-on: https://go-review.googlesource.com/8832
Reviewed-by: Rick Hudson <rlh@golang.org>
'themoduledata' doesn't really make sense now we support multiple moduledata
objects.
Change-Id: I8263045d8f62a42cb523502b37289b0fba054f62
Reviewed-on: https://go-review.googlesource.com/8521
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This changes all the places that consult themoduledata to consult a
linked list of moduledata objects, as will be necessary for
-linkshared to work.
Obviously, as there is as yet no way of adding moduledata objects to
this list, all this change achieves right now is wasting a few
instructions here and there.
Change-Id: I397af7f60d0849b76aaccedf72238fe664867051
Reviewed-on: https://go-review.googlesource.com/8231
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This tracks the number of heap bytes marked by a GC cycle. We'll use
this information to precisely trigger the next GC cycle.
Currently this aggregates the work counter in gcWork and dispose
atomically aggregates this into a global work counter. dispose happens
relatively infrequently, so the contention on the global counter
should be low. If this turns out to be an issue, we can reduce the
number of disposes, and if it's still a problem, we can switch to
per-P counters.
Change-Id: I1bc377cb2e802ef61c2968602b63146d52e7f5db
Reviewed-on: https://go-review.googlesource.com/8388
Reviewed-by: Russ Cox <rsc@golang.org>
In preparation for being able to run a go program that has code
in several objects, this changes from having several linker
symbols used by the runtime into having one linker symbol that
points at a structure containing the needed data. Multiple
object support will construct a linked list of such structures.
A follow up will initialize the slices in the themoduledata
structure directly from the linker but I was aiming for a minimal
diff for now.
Change-Id: I613cce35309801cf265a1d5ae5aaca8d689c5cbf
Reviewed-on: https://go-review.googlesource.com/7441
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Currently, gcDrainN is documented saying that it must be run on the
system stack. In fact, the problem and solution here are somewhat
subtler. First, it doesn't have to happen on the system stack, it just
has to be non-stoppable (that is, non-preemptible). Second, this isn't
specific to gcDrainN (though gcDrainN is perhaps the most surprising
instance); it's general to anything that uses the gcWork structure.
Move the comment to gcWork and generalize it.
Change-Id: I5277b5abb070e47f8d783bc15a310b379c6adc22
Reviewed-on: https://go-review.googlesource.com/8247
Reviewed-by: Rick Hudson <rlh@golang.org>
gcDrain used to be passed a *workbuf to start draining from, but now
it takes a gcWork, which hides whether or not there's an initial
workbuf. Update the comment to match this.
Change-Id: I976b58e5bfebc451cfd4fa75e770113067b5cc07
Reviewed-on: https://go-review.googlesource.com/8246
Reviewed-by: Rick Hudson <rlh@golang.org>
The barrier in gcDrain does not account for concurrent gcDrainNs
happening in gchelpwork, so it can actually return while there is
still work being done. It turns out this is okay, but for subtle
reasons involving gcDrainN always being run on the system
stack. Document these reasons.
Change-Id: Ib07b3753cc4e2b54533ab3081a359cbd1c3c08fb
Reviewed-on: https://go-review.googlesource.com/7736
Reviewed-by: Rick Hudson <rlh@golang.org>
The distinction between gcWorkProducer and gcWork (producer and
consumer) is not serving us as originally intended, so merge these
into just gcWork.
The original intent was to replace the currentwbuf cache with a
gcWorkProducer. However, with gchelpwork (aka mutator assists),
mutators can both produce and consume work, so it will make more sense
to cache a whole gcWork.
Change-Id: I6e633e96db7cb23a64fbadbfc4607e3ad32bcfb3
Reviewed-on: https://go-review.googlesource.com/7733
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently markroot fetches the wbuf to fill from the per-M wbuf
cache. The wbuf cache is primarily meant for the write barrier because
it produces very little work on each call. There's little point to
using the cache in mark root, since each call to markroot is likely to
produce a large amount of work (so the slight win on getting it from
the cache instead of from the central wbuf lists doesn't matter), and
markroot does not dispose the wbuf back to the cache (so most markroot
calls won't get anything from the wbuf cache anyway).
Instead, just get the wbuf from the central wbuf lists like other work
producers. This will simplify later changes.
Change-Id: I07a18a4335a41e266a6d70aa3a0911a40babce23
Reviewed-on: https://go-review.googlesource.com/7732
Reviewed-by: Rick Hudson <rlh@golang.org>
When checkmark fails, greyobject dumps both the object that pointed to
the unmarked object and the unmarked object. This code cluttered up
greyobject, was copy-pasted for the two objects, and the copy for
dumping the unmarked object was not entirely correct.
Extract object dumping out to a new function. This declutters
greyobject and fixes the bugs in dumping the unmarked object. The new
function is slightly cleaned up from the original code to have more
natural control flow and shows a marker on the field in the base
object that points to the unmarked object to make it easy to find.
Change-Id: Ib51318a943f50b0b99995f0941d03ee8876b9fcf
Reviewed-on: https://go-review.googlesource.com/7506
Reviewed-by: Rick Hudson <rlh@golang.org>
scanobject no longer returns the new wbuf.
Change-Id: I0da335ae5cd7ef7ea0e0fa965cf0e9f3a650d0e6
Reviewed-on: https://go-review.googlesource.com/7505
Reviewed-by: Rick Hudson <rlh@golang.org>
Everything has moved to Go, but comments still refer to .c/.h files.
Fix all of those up, at least for these three directories.
Fixes#10138
Change-Id: Ie5efe89b247841e0b3f82aac5256b2c606ef67dc
Reviewed-on: https://go-review.googlesource.com/7431
Reviewed-by: Russ Cox <rsc@golang.org>
This is an experiment to see if removing the boundary bit logic will
lead to fewer cache misses and improved performance. Instead of using
boundary bits we use the span information to get element size and use
some bit whacking to get the boundary without having to touch the
random heap bits which cause cache misses.
Furthermore once the boundary bit is removed we can either use that
bit for a simpler checkmark routine or we can reduce the number of
bits in the GC bitmap to 2 bits per pointer sized work. For example
the 2 bits at the boundary can be used for marking and pointer/scalar
differentiation. Since we don't need the mark bit except at the
boundary nibble of the object other nibbles can use this bit
as a noscan bit to indicate that there are no more pointers in
the object.
Currently the changed included in this CL slows down the garbage
benchmark. With the boundary bits garbage gives 5.78 and without
(this CL) it gives 5.88 which is a 2% slowdown.
Change-Id: Id68f831ad668176f7dc9f7b57b339e4ebb6dc4c2
Reviewed-on: https://go-review.googlesource.com/6665
Reviewed-by: Austin Clements <austin@google.com>
Previously, the typeDead check in greyobject was under a separate
!useCheckmark conditional. Put it with the rest of the !useCheckmark
code. Also move a comment about atomic update of the marked bit to
where we actually do that update now.
Change-Id: Ief5f16401a25739ad57d959607b8d81ffe0bc211
Reviewed-on: https://go-review.googlesource.com/6271
Reviewed-by: Rick Hudson <rlh@golang.org>
Previously, we had three loops in the garbage collector that all
cleared the per-G GC flags. Consolidate these into one function.
This one function is designed to work in a concurrent setting. As a
result, it's slightly more expensive than the loops it replaces during
STW phases, but these happen at most twice per GC.
Change-Id: Id1ec0074fd58865eb0112b8a0547b267802d0df1
Reviewed-on: https://go-review.googlesource.com/5881
Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
This is a nice split but more importantly it provides a better
way to fit the checkmark phase into the sequencing.
Also factor out common span copying into gcSpanCopy.
Change-Id: Ia058644974e4ed4ac3cf4b017a3446eb2284d053
Reviewed-on: https://go-review.googlesource.com/5333
Reviewed-by: Austin Clements <austin@google.com>
Move code from malloc1.go, malloc2.go, mem.go, mgc0.go into
appropriate locations.
Factor mgc.go into mgc.go, mgcmark.go, mgcsweep.go, mstats.go.
A lot of this code was in certain files because the right place was in
a C file but it was written in Go, or vice versa. This is one step toward
making things actually well-organized again.
Change-Id: I6741deb88a7cfb1c17ffe0bcca3989e10207968f
Reviewed-on: https://go-review.googlesource.com/5300
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Rick Hudson <rlh@golang.org>
