Cuts the number of calls from 6 to 2 in the non-debug case.
LGTM=iant
R=golang-codereviews, iant
CC=0intro, aram, golang-codereviews, khr
https://golang.org/cl/86040043
Given
type Outer struct {
*Inner
...
}
the compiler generates the implementation of (*Outer).M dispatching to
the embedded Inner. The implementation is logically:
func (p *Outer) M() {
(p.Inner).M()
}
but since the only change here is the replacement of one pointer
receiver with another, the actual generated code overwrites the
original receiver with the p.Inner pointer and then jumps to the M
method expecting the *Inner receiver.
During reflect.Value.Call, we create an argument frame and the
associated data structures to describe it to the garbage collector,
populate the frame, call reflect.call to run a function call using
that frame, and then copy the results back out of the frame. The
reflect.call function does a memmove of the frame structure onto the
stack (to set up the inputs), runs the call, and the memmoves the
stack back to the frame structure (to preserve the outputs).
Originally reflect.call did not distinguish inputs from outputs: both
memmoves were for the full stack frame. However, in the case where the
called function was one of these wrappers, the rewritten receiver is
almost certainly a different type than the original receiver. This is
not a problem on the stack, where we use the program counter to
determine the type information and understand that during (*Outer).M
the receiver is an *Outer while during (*Inner).M the receiver in the
same memory word is now an *Inner. But in the statically typed
argument frame created by reflect, the receiver is always an *Outer.
Copying the modified receiver pointer off the stack into the frame
will store an *Inner there, and then if a garbage collection happens
to scan that argument frame before it is discarded, it will scan the
*Inner memory as if it were an *Outer. If the two have different
memory layouts, the collection will intepret the memory incorrectly.
Fix by only copying back the results.
Fixes#7725.
LGTM=khr
R=khr
CC=dave, golang-codereviews
https://golang.org/cl/85180043
Trying to make GODEBUG=gcdead=1 work with liveness
and in particular ambiguously live variables.
1. In the liveness computation, mark all ambiguously live
variables as live for the entire function, except the entry.
They are zeroed directly after entry, and we need them not
to be poisoned thereafter.
2. In the liveness computation, compute liveness (and deadness)
for all parameters, not just pointer-containing parameters.
Otherwise gcdead poisons untracked scalar parameters and results.
3. Fix liveness debugging print for -live=2 to use correct bitmaps.
(Was not updated for compaction during compaction CL.)
4. Correct varkill during map literal initialization.
Was killing the map itself instead of the inserted value temp.
5. Disable aggressive varkill cleanup for call arguments if
the call appears in a defer or go statement.
6. In the garbage collector, avoid bug scanning empty
strings. An empty string is two zeros. The multiword
code only looked at the first zero and then interpreted
the next two bits in the bitmap as an ordinary word bitmap.
For a string the bits are 11 00, so if a live string was zero
length with a 0 base pointer, the poisoning code treated
the length as an ordinary word with code 00, meaning it
needed poisoning, turning the string into a poison-length
string with base pointer 0. By the same logic I believe that
a live nil slice (bits 11 01 00) will have its cap poisoned.
Always scan full multiword struct.
7. In the runtime, treat both poison words (PoisonGC and
PoisonStack) as invalid pointers that warrant crashes.
Manual testing as follows:
- Create a script called gcdead on your PATH containing:
#!/bin/bash
GODEBUG=gcdead=1 GOGC=10 GOTRACEBACK=2 exec "$@"
- Now you can build a test and then run 'gcdead ./foo.test'.
- More importantly, you can run 'go test -short -exec gcdead std'
to run all the tests.
Fixes#7676.
While here, enable the precise scanning of slices, since that was
disabled due to bugs like these. That now works, both with and
without gcdead.
Fixes#7549.
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/83410044
Reduce footprint of liveness bitmaps by about 5x.
1. Mark all liveness bitmap symbols as 4-byte aligned
(they were aligned to a larger size by default).
2. The bitmap data is a bitmap count n followed by n bitmaps.
Each bitmap begins with its own count m giving the number
of bits. All the m's are the same for the n bitmaps.
Emit this bitmap length once instead of n times.
3. Many bitmaps within a function have the same bit values,
but each call site was given a distinct bitmap. Merge duplicate
bitmaps so that no bitmap is written more than once.
4. Many functions end up with the same aggregate bitmap data.
We used to name the bitmap data funcname.gcargs and funcname.gclocals.
Instead, name it gclocals.<md5 of data> and mark it dupok so
that the linker coalesces duplicate sets. This cut the bitmap
data remaining after step 3 by 40%; I was not expecting it to
be quite so dramatic.
Applied to "go build -ldflags -w code.google.com/p/go.tools/cmd/godoc":
bitmaps pclntab binary on disk
before this CL 1326600 1985854 12738268
4-byte align 1154288 (0.87x) 1985854 (1.00x) 12566236 (0.99x)
one bitmap len 782528 (0.54x) 1985854 (1.00x) 12193500 (0.96x)
dedup bitmap 414748 (0.31x) 1948478 (0.98x) 11787996 (0.93x)
dedup bitmap set 245580 (0.19x) 1948478 (0.98x) 11620060 (0.91x)
While here, remove various dead blocks of code from plive.c.
Fixes#6929.
Fixes#7568.
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/83630044
GODEBUG=allocfreetrace=1:
The allocfreetrace=1 mode prints a stack trace for each block
allocated and freed, and also a stack trace for each garbage collection.
It was implemented by reusing the heap profiling support: if allocfreetrace=1
then the heap profile was effectively running at 1 sample per 1 byte allocated
(always sample). The stack being shown at allocation was the stack gathered
for profiling, meaning it was derived only from the program counters and
did not include information about function arguments or frame pointers.
The stack being shown at free was the allocation stack, not the free stack.
If you are generating this log, you can find the allocation stack yourself, but
it can be useful to see exactly the sequence that led to freeing the block:
was it the garbage collector or an explicit free? Now that the garbage collector
runs on an m0 stack, the stack trace for the garbage collector was never interesting.
Fix all these problems:
1. Decouple allocfreetrace=1 from heap profiling.
2. Print the standard goroutine stack traces instead of a custom format.
3. Print the stack trace at time of allocation for an allocation,
and print the stack trace at time of free (not the allocation trace again)
for a free.
4. Print all goroutine stacks at garbage collection. Having all the stacks
means that you can see the exact point at which each goroutine was
preempted, which is often useful for identifying liveness-related errors.
GODEBUG=gcdead=1:
This mode overwrites dead pointers with a poison value.
Detect the poison value as an invalid pointer during collection,
the same way that small integers are invalid pointers.
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/81670043
This is the same check we use during stack copying.
The check cannot be applied to C stack frames, even
though we do emit pointer bitmaps for the arguments,
because (1) the pointer bitmaps assume all arguments
are always live, not true of outputs during the prologue,
and (2) the pointer bitmaps encode interface values as
pointer pairs, not true of interfaces holding integers.
For the rest of the frames, however, we should hold ourselves
to the rule that a pointer marked live really is initialized.
The interface scanning already implicitly checks this
because it interprets the type word as a valid type pointer.
This may slow things down a little because of the extra loads.
Or it may speed things up because we don't bother enqueuing
nil pointers anymore. Enough of the rest of the system is slow
right now that we can't measure it meaningfully.
Enable for now, even if it is slow, to shake out bugs in the
liveness bitmaps, and then decide whether to turn it off
for the Go 1.3 release (issue 7650 reminds us to do this).
The new m->traceback field lets us force printing of fp=
values on all goroutine stack traces when we detect a
bad pointer. This makes it easier to understand exactly
where in the frame the bad pointer is, so that we can trace
it back to a specific variable and determine what is wrong.
Update #7650
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/80860044
Currently it's possible that bgsweep finishes before all spans
have been swept (we only know that sweeping of all spans has *started*).
In such case bgsweep may fail wake up runfinq goroutine when it needs to.
finq may still be nil at this point, but some finalizers may be queued later.
Make bgsweep to wait for sweeping to *complete*, then it can decide
whether it needs to wake up runfinq for sure.
Update #7533
LGTM=rsc
R=rsc
CC=golang-codereviews
https://golang.org/cl/75960043
If we set obj, then it will be enqueued for marking at the end of the scanning loop.
This is not necessary, since we've already marked it.
This can wait for 1.4 if you wish.
LGTM=rsc
R=rsc
CC=golang-codereviews
https://golang.org/cl/80030043
Change two-bit stack map entries to encode:
0 = dead
1 = scalar
2 = pointer
3 = multiword
If multiword, the two-bit entry for the following word encodes:
0 = string
1 = slice
2 = iface
3 = eface
That way, during stack scanning we can check if a string
is zero length or a slice has zero capacity. We can avoid
following the contained pointer in those cases. It is safe
to do so because it can never be dereferenced, and it is
desirable to do so because it may cause false retention
of the following block in memory.
Slice feature turned off until issue 7564 is fixed.
Update #7549
LGTM=rsc
R=golang-codereviews, bradfitz, rsc
CC=golang-codereviews
https://golang.org/cl/76380043
The existing code did not have a clear notion of whether
memory has been actually reserved. It checked based on
whether in 32-bit mode or 64-bit mode and (on GNU/Linux) the
requested address, but it confused the requested address and
the returned address.
LGTM=rsc
R=rsc, dvyukov
CC=golang-codereviews, michael.hudson
https://golang.org/cl/79610043
The nproc and ndone fields are uint32. This makes the type
consistent.
LGTM=minux.ma
R=golang-codereviews, minux.ma
CC=golang-codereviews
https://golang.org/cl/79340044
It's possible that bgsweep constantly does not catch up for some reason,
in this case runfinq was not woken at all.
R=rsc
CC=golang-codereviews
https://golang.org/cl/75940043
The problem was that spans end up in wrong lists after split
(e.g. in h->busy instead of h->central->empty).
Also the span can be non-swept before split,
I don't know what it can cause, but it's safer to operate on swept spans.
Fixes#7544.
R=golang-codereviews, rsc
CC=golang-codereviews, khr
https://golang.org/cl/76160043
The garbage collector uses type information to guide the
traversal of the heap. If it sees a field that should be a string,
it marks the object pointed at by the string data pointer as
visited but does not bother to look at the data, because
strings contain bytes, not pointers.
If you save s[len(s):] somewhere, though, the string data pointer
actually points just beyond the string data; if the string data
were exactly the size of an allocated block, the string data
pointer would actually point at the next block. It is incorrect
to mark that next block as visited and not bother to look at
the data, because the next block may be some other type
entirely.
The fix is to ignore strings with zero length during collection:
they are empty and can never become non-empty: the base
pointer will never be used again. The handling of slices already
does this (but using cap instead of len).
This was not a bug in Go 1.2, because until January all string
allocations included a trailing NUL byte not included in the
length, so s[len(s):] still pointed inside the string allocation
(at the NUL).
This bug was causing the crashes in test/run.go. Specifically,
the parsing of a regexp in package regexp/syntax allocated a
[]syntax.Inst with rounded size 1152 bytes. In fact it
allocated many such slices, because during the processing of
test/index2.go it creates thousands of regexps that are all
approximately the same complexity. That takes a long time, and
test/run works on other tests in other goroutines. One such
other test is chan/perm.go, which uses an 1152-byte source
file. test/run reads that file into a []byte and then calls
strings.Split(string(src), "\n"). The string(src) creates an
1152-byte string - and there's a very good chance of it
landing next to one of the many many regexp slices already
allocated - and then because the file ends in a \n,
strings.Split records the tail empty string as the final
element in the slice. A garbage collection happens at this
point, the collection finds that string before encountering
the []syntax.Inst data it now inadvertently points to, and the
[]syntax.Inst data is not scanned for the pointers that it
contains. Each syntax.Inst contains a []rune, those are
missed, and the backing rune arrays are freed for reuse. When
the regexp is later executed, the runes being searched for are
no longer runes at all, and there is no match, even on text
that should match.
On 64-bit machines the pointer in the []rune inside the
syntax.Inst is larger (along with a few other pointers),
pushing the []syntax.Inst backing array into a larger size
class, avoiding the collision with chan/perm.go's
inadvertently sized file.
I expect this was more prevalent on OS X than on Linux or
Windows because those managed to run faster or slower and
didn't overlap index2.go with chan/perm.go as often. On the
ARM systems, we only run one errorcheck test at a time, so
index2 and chan/perm would never overlap.
It is possible that this bug is the root cause of other crashes
as well. For now we only know it is the cause of the test/run crash.
Many thanks to Dmitriy for help debugging.
Fixes#7344.
Fixes#7455.
LGTM=r, dvyukov, dave, iant
R=golang-codereviews, dave, r, dvyukov, delpontej, iant
CC=golang-codereviews, khr
https://golang.org/cl/74250043
Spans are now private to threads, and the loop
is removed from all other functions.
Remove it from marknogc for consistency.
LGTM=khr, rsc
R=golang-codereviews, bradfitz, khr
CC=golang-codereviews, khr, rsc
https://golang.org/cl/72520043
One reason the sync.Pool finalizer test can fail is that
this function's ef1 contains uninitialized data that just
happens to point at some of the old pool. I've seen this cause
retention of a single pool cache line (32 elements) on arm.
Really we need liveness information for C functions, but
for now we can be more careful about data in long-lived
C functions that block.
LGTM=bradfitz, dvyukov
R=golang-codereviews, bradfitz, dvyukov
CC=golang-codereviews, iant, khr
https://golang.org/cl/72490043
Two memory allocator bug fixes.
- efence is not maintaining the proper heap metadata
to make eventual memory reuse safe, so use SysFault.
- now that our heap PageSize is 8k but most hardware
uses 4k pages, SysAlloc and SysReserve results must be
explicitly aligned. Do that in a few more call sites and
document this fact in malloc.h.
Fixes#7448.
LGTM=iant
R=golang-codereviews, josharian, iant
CC=dvyukov, golang-codereviews
https://golang.org/cl/71750048
Before GC, we flush all the per-P allocation caches. Doing
stack shrinking mid-GC causes these caches to fill up. At the
end of gc, the sweepgen is incremented which causes all of the
data in these caches to be in a bad state (cached but not yet
swept).
Move the stack shrinking until after sweepgen is incremented,
so any caching that happens as part of shrinking is done with
already-swept data.
Reenable stack copying.
LGTM=bradfitz
R=golang-codereviews, bradfitz
CC=golang-codereviews
https://golang.org/cl/69620043
On stack overflow, if all frames on the stack are
copyable, we copy the frames to a new stack twice
as large as the old one. During GC, if a G is using
less than 1/4 of its stack, copy the stack to a stack
half its size.
TODO
- Do something about C frames. When a C frame is in the
stack segment, it isn't copyable. We allocate a new segment
in this case.
- For idempotent C code, we can abort it, copy the stack,
then retry. I'm working on a separate CL for this.
- For other C code, we can raise the stackguard
to the lowest Go frame so the next call that Go frame
makes triggers a copy, which will then succeed.
- Pick a starting stack size?
The plan is that eventually we reach a point where the
stack contains only copyable frames.
LGTM=rsc
R=dvyukov, rsc
CC=golang-codereviews
https://golang.org/cl/54650044
MCaches now hold a MSpan for each sizeclass which they have
exclusive access to allocate from, so no lock is needed.
Modifying the heap bitmaps also no longer requires a cas.
runtime.free gets more expensive. But we don't use it
much any more.
It's not much faster on 1 processor, but it's a lot
faster on multiple processors.
benchmark old ns/op new ns/op delta
BenchmarkSetTypeNoPtr1 24 23 -0.42%
BenchmarkSetTypeNoPtr2 33 34 +0.89%
BenchmarkSetTypePtr1 51 49 -3.72%
BenchmarkSetTypePtr2 55 54 -1.98%
benchmark old ns/op new ns/op delta
BenchmarkAllocation 52739 50770 -3.73%
BenchmarkAllocation-2 33957 34141 +0.54%
BenchmarkAllocation-3 33326 29015 -12.94%
BenchmarkAllocation-4 38105 25795 -32.31%
BenchmarkAllocation-5 68055 24409 -64.13%
BenchmarkAllocation-6 71544 23488 -67.17%
BenchmarkAllocation-7 68374 23041 -66.30%
BenchmarkAllocation-8 70117 20758 -70.40%
LGTM=rsc, dvyukov
R=dvyukov, bradfitz, khr, rsc
CC=golang-codereviews
https://golang.org/cl/46810043
See golang.org/s/go13nacl for design overview.
This CL is the mostly mechanical changes from rsc's Go 1.2 based NaCl branch, specifically 39cb35750369 to 500771b477cf from https://code.google.com/r/rsc-go13nacl. This CL does not include working NaCl support, there are probably two or three more large merges to come.
CL 15750044 is not included as it involves more invasive changes to the linker which will need to be merged separately.
The exact change lists included are
15050047: syscall: support for Native Client
15360044: syscall: unzip implementation for Native Client
15370044: syscall: Native Client SRPC implementation
15400047: cmd/dist, cmd/go, go/build, test: support for Native Client
15410048: runtime: support for Native Client
15410049: syscall: file descriptor table for Native Client
15410050: syscall: in-memory file system for Native Client
15440048: all: update +build lines for Native Client port
15540045: cmd/6g, cmd/8g, cmd/gc: support for Native Client
15570045: os: support for Native Client
15680044: crypto/..., hash/crc32, reflect, sync/atomic: support for amd64p32
15690044: net: support for Native Client
15690048: runtime: support for fake time like on Go Playground
15690051: build: disable various tests on Native Client
LGTM=rsc
R=rsc
CC=golang-codereviews
https://golang.org/cl/68150047
Reinforce the guarantee that MSpan_EnsureSwept actually ensures that the span is swept.
I have not observed crashes related to this, but I do not see why it can't crash as well.
LGTM=rsc
R=golang-codereviews
CC=golang-codereviews, khr, rsc
https://golang.org/cl/67990043
runfinqv is already defined the same way on line 271.
There may also be something to fix in compiler/linker wrt diagnostics.
Fixes#7375.
LGTM=bradfitz
R=golang-codereviews, dave, bradfitz
CC=golang-codereviews
https://golang.org/cl/67850044
Package runtime's C functions written to be called from Go
started out written in C using carefully constructed argument
lists and the FLUSH macro to write a result back to memory.
For some functions, the appropriate parameter list ended up
being architecture-dependent due to differences in alignment,
so we added 'goc2c', which takes a .goc file containing Go func
declarations but C bodies, rewrites the Go func declaration to
equivalent C declarations for the target architecture, adds the
needed FLUSH statements, and writes out an equivalent C file.
That C file is compiled as part of package runtime.
Native Client's x86-64 support introduces the most complex
alignment rules yet, breaking many functions that could until
now be portably written in C. Using goc2c for those avoids the
breakage.
Separately, Keith's work on emitting stack information from
the C compiler would require the hand-written functions
to add #pragmas specifying how many arguments are result
parameters. Using goc2c for those avoids maintaining #pragmas.
For both reasons, use goc2c for as many Go-called C functions
as possible.
This CL is a replay of the bulk of CL 15400047 and CL 15790043,
both of which were reviewed as part of the NaCl port and are
checked in to the NaCl branch. This CL is part of bringing the
NaCl code into the main tree.
No new code here, just reformatting and occasional movement
into .h files.
LGTM=r
R=dave, alex.brainman, r
CC=golang-codereviews
https://golang.org/cl/65220044
This cleans up the code significantly, and it avoids any
possible problems with madvise zeroing out some but
not all of the data.
Fixes#6400.
LGTM=dave
R=dvyukov, dave
CC=golang-codereviews
https://golang.org/cl/57680046
The issue was that one of the MSpan_Sweep callers
was doing sweep with preemption enabled.
Additional checks are added.
LGTM=rsc
R=rsc, dave
CC=golang-codereviews
https://golang.org/cl/62990043
State of the world:
CL 46430043 introduced a new concurrent sweep but is broken.
CL 62360043 made the new sweep non-concurrent
to try to fix the world while we understand what's wrong with
the concurrent version.
This CL fixes the non-concurrent form to run finalizers.
This CL is just a band-aid to get the build green again.
Dmitriy is working on understanding and then fixing what's
wrong with the concurrent sweep.
TBR=dvyukov
CC=golang-codereviews
https://golang.org/cl/62370043
Moves sweep phase out of stoptheworld by adding
background sweeper goroutine and lazy on-demand sweeping.
It turned out to be somewhat trickier than I expected,
because there is no point in time when we know size of live heap
nor consistent number of mallocs and frees.
So everything related to next_gc, mprof, memstats, etc becomes trickier.
At the end of GC next_gc is conservatively set to heap_alloc*GOGC,
which is much larger than real value. But after every sweep
next_gc is decremented by freed*GOGC. So when everything is swept
next_gc becomes what it should be.
For mprof I had to introduce 3-generation scheme (allocs, revent_allocs, prev_allocs),
because by the end of GC we know number of frees for the *previous* GC.
Significant caution is required to not cross yet-unknown real value of next_gc.
This is achieved by 2 means:
1. Whenever I allocate a span from MCentral, I sweep a span in that MCentral.
2. Whenever I allocate N pages from MHeap, I sweep until at least N pages are
returned to heap.
This provides quite strong guarantees that heap does not grow when it should now.
http-1
allocated 7036 7033 -0.04%
allocs 60 60 +0.00%
cputime 51050 46700 -8.52%
gc-pause-one 34060569 1777993 -94.78%
gc-pause-total 2554 133 -94.79%
latency-50 178448 170926 -4.22%
latency-95 284350 198294 -30.26%
latency-99 345191 220652 -36.08%
rss 101564416 101007360 -0.55%
sys-gc 6606832 6541296 -0.99%
sys-heap 88801280 87752704 -1.18%
sys-other 7334208 7405928 +0.98%
sys-stack 524288 524288 +0.00%
sys-total 103266608 102224216 -1.01%
time 50339 46533 -7.56%
virtual-mem 292990976 293728256 +0.25%
garbage-1
allocated 2983818 2990889 +0.24%
allocs 62880 62902 +0.03%
cputime 16480000 16190000 -1.76%
gc-pause-one 828462467 487875135 -41.11%
gc-pause-total 4142312 2439375 -41.11%
rss 1151709184 1153712128 +0.17%
sys-gc 66068352 66068352 +0.00%
sys-heap 1039728640 1039728640 +0.00%
sys-other 37776064 40770176 +7.93%
sys-stack 8781824 8781824 +0.00%
sys-total 1152354880 1155348992 +0.26%
time 16496998 16199876 -1.80%
virtual-mem 1409564672 1402281984 -0.52%
LGTM=rsc
R=golang-codereviews, sameer, rsc, iant, jeremyjackins, gobot
CC=golang-codereviews, khr
https://golang.org/cl/46430043
Currently windows crashes because early allocs in schedinit
try to allocate tiny memory blocks, but m->p is not yet setup.
I've considered calling procresize(1) earlier in schedinit,
but this refactoring is better and must fix the issue as well.
Fixes#7218.
R=golang-codereviews, r
CC=golang-codereviews
https://golang.org/cl/54570045
There is more zeroing than I would like right now -
temporaries used for the new map and channel runtime
calls need to be eliminated - but it will do for now.
This CL only has an effect if you are building with
GOEXPERIMENT=precisestack ./all.bash
(or make.bash). It costs about 5% in the overall time
spent in all.bash. That number will come down before
we make it on by default, but this should be enough for
Keith to try using the precise maps for copying stacks.
amd64 only (and it's not really great generated code).
TBR=khr, iant
CC=golang-codereviews
https://golang.org/cl/56430043
Introduces two-phase goroutine parking mechanism -- prepare to park, commit park.
This mechanism does not require backing mutex to protect wait predicate.
Use it in netpoll. See comment in netpoll.goc for details.
This slightly reduces contention between reader, writer and read/write io notifications;
and just eliminates a bunch of mutex operations from hotpaths, thus making then faster.
benchmark old ns/op new ns/op delta
BenchmarkTCP4ConcurrentReadWrite 2109 1945 -7.78%
BenchmarkTCP4ConcurrentReadWrite-2 1162 1113 -4.22%
BenchmarkTCP4ConcurrentReadWrite-4 798 755 -5.39%
BenchmarkTCP4ConcurrentReadWrite-8 803 748 -6.85%
BenchmarkTCP4Persistent 9411 9240 -1.82%
BenchmarkTCP4Persistent-2 5888 5813 -1.27%
BenchmarkTCP4Persistent-4 4016 3968 -1.20%
BenchmarkTCP4Persistent-8 3943 3857 -2.18%
R=golang-codereviews, mikioh.mikioh, gobot, iant, rsc
CC=golang-codereviews, khr
https://golang.org/cl/45700043
Currently we collect (add) all roots into a global array in a single-threaded GC phase.
This hinders parallelism.
With this change we just kick off parallel for for number_of_goroutines+5 iterations.
Then parallel for callback decides whether it needs to scan stack of a goroutine
scan data segment, scan finalizers, etc. This eliminates the single-threaded phase entirely.
This requires to store all goroutines in an array instead of a linked list
(to allow direct indexing).
This CL also removes DebugScan functionality. It is broken because it uses
unbounded stack, so it can not run on g0. When it was working, I've found
it helpless for debugging issues because the two algorithms are too different now.
This change would require updating the DebugScan, so it's simpler to just delete it.
With 8 threads this change reduces GC pause by ~6%, while keeping cputime roughly the same.
garbage-8
allocated 2987886 2989221 +0.04%
allocs 62885 62887 +0.00%
cputime 21286000 21272000 -0.07%
gc-pause-one 26633247 24885421 -6.56%
gc-pause-total 873570 811264 -7.13%
rss 242089984 242515968 +0.18%
sys-gc 13934336 13869056 -0.47%
sys-heap 205062144 205062144 +0.00%
sys-other 12628288 12628288 +0.00%
sys-stack 11534336 11927552 +3.41%
sys-total 243159104 243487040 +0.13%
time 2809477 2740795 -2.44%
R=golang-codereviews, rsc
CC=cshapiro, golang-codereviews, khr
https://golang.org/cl/46860043
Instead of a per-goroutine stack of defers for all sizes,
introduce per-P defer pool for argument sizes 8, 24, 40, 56, 72 bytes.
For a program that starts 1e6 goroutines and then joins then:
old: rss=6.6g virtmem=10.2g time=4.85s
new: rss=4.5g virtmem= 8.2g time=3.48s
R=golang-codereviews, rsc
CC=golang-codereviews
https://golang.org/cl/42750044
Currently for 2-word blocks we set the flag to clear the flag. Makes no sense.
In particular on 32-bits we call memclr always.
R=golang-codereviews, dave, iant
CC=golang-codereviews, khr, rsc
https://golang.org/cl/41170044
Example of output:
goroutine 4 [sleep for 3 min]:
time.Sleep(0x34630b8a000)
src/pkg/runtime/time.goc:31 +0x31
main.func·002()
block.go:16 +0x2c
created by main.main
block.go:17 +0x33
Full program and output are here:
http://play.golang.org/p/NEZdADI3TdFixes#6809.
R=golang-codereviews, khr, kamil.kisiel, bradfitz, rsc
CC=golang-codereviews
https://golang.org/cl/50420043