Changing from 4 kB to 8 kB brings significant improvement
on a variety of the Go 1 benchmarks, on both amd64
and 386 systems.
Significant runtime reductions:
amd64 386
GoParse -14% -1%
GobDecode -12% -20%
GobEncode -64% -1%
JSONDecode -9% -4%
JSONEncode -15% -5%
Template -17% -14%
In the longer term, khr's new stacks will avoid needing to
make this decision at all, but for Go 1.2 this is a reasonable
stopgap that makes performance significantly better.
Demand paging should mean that if the second 4 kB is not
used, it will not be brought into memory, so the change
should not adversely affect resident set size.
The same argument could justify bumping as high as 64 kB
on 64-bit machines, but there are diminishing returns
after 8 kB, and using 8 kB limits the possible unintended
memory overheads we are not aware of.
Benchmark graphs at
http://swtch.com/~rsc/gostackamd64.htmlhttp://swtch.com/~rsc/gostack386.html
Full data at
http://swtch.com/~rsc/gostack.zip
R=golang-dev, khr, dave, bradfitz, dvyukov
CC=golang-dev
https://golang.org/cl/14317043
It is not possible to use (there is no declaration in package syscall),
and no one seems to care.
Alex Brainman may bring this back properly for Go 1.3.
Fixes#6338.
R=golang-dev, r, alex.brainman
CC=golang-dev
https://golang.org/cl/14287043
Not scanning the stack by frames means we are reintroducing
a few false positives into the collection. Run the finalizer registration
in its own goroutine so that stack is guaranteed to be out of
consideration in a later collection.
This is working around a regression from yesterday's tip, but
it's not a regression from Go 1.1.
R=golang-dev
TBR=golang-dev
CC=golang-dev
https://golang.org/cl/14290043
Makes build unnecessarily slower. Will fix the parser instead.
««« original CL description
runtime/pprof: run TestGoroutineSwitch for longer
Short test now takes about 0.5 second here.
Fixes#6417.
The failure was also seen on our builders.
R=golang-dev, minux.ma, r
CC=golang-dev
https://golang.org/cl/13321048
»»»
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/13720048
Short test now takes about 0.5 second here.
Fixes#6417.
The failure was also seen on our builders.
R=golang-dev, minux.ma, r
CC=golang-dev
https://golang.org/cl/13321048
If a fault happens in malloc, inevitably the next thing that happens
is a deadlock trying to allocate the panic value that says the fault
happened. Stop doing that, two ways.
First, reject panic in malloc just as we reject panic in garbage collection.
Second, runtime.panicstring was using an error implementation
backed by a Go string, so the interface held an allocated *string.
Since the actual errors are C strings, define a new error
implementation backed by a C char*, which needs no indirection
and therefore no allocation.
This second fix will avoid allocation for errors like nil panic derefs
or division by zero, so it is worth doing even though the first fix
should take care of faults during malloc.
Update #6419
R=golang-dev, dvyukov, dave
CC=golang-dev
https://golang.org/cl/13774043
Keeping pointers from the pre-walk phase confuses
the race detection instrumentation.
Fixes#6418.
R=golang-dev, dvyukov, r
CC=golang-dev
https://golang.org/cl/13368057
This interface is required to use the PCDATA interface
implemented in Go 1.2. While initially entirely private, the
FUNCDATA side of the interface has been made public. This
change completes the FUNCDATA/PCDATA interface.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/13735043
The code for call site-specific pointer bitmaps was not ready in time,
but the zeroing required without it is too expensive to use by default.
We will have to wait for precise collection of stack frames until Go 1.3.
The precise collection can be re-enabled by
GOEXPERIMENT=precisestack ./all.bash
but that will not be the default for a Go 1.2 build.
Fixes#6087.
R=golang-dev, jeremyjackins, dan.kortschak, r
CC=golang-dev
https://golang.org/cl/13677045
The uint64 divide function calls _mul64x32 to do a 64x32-bit multiply
and then compares the result against the 64-bit numerator.
If the result is bigger than the numerator, must use the slow path.
Unfortunately, the 64x32 produces a 96-bit product, and only the
low 64 bits were being used in the comparison. Return all 96 bits,
the bottom 64 via the original uint64* pointer, and the top 32
as the function's return value.
Fixes 386 build (broken by ARM division tests).
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/13722044
Because we can, and because it otherwise might crash
the program if we think we're out of memory.
Fixes#6390.
R=golang-dev, iant, minux.ma
CC=golang-dev
https://golang.org/cl/13345048
The implementation of division in the 5 toolchain is a bit too magical.
Hide the magic from the traceback routines.
Also add a test for the results of the software divide routine.
Fixes#5805.
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/13239052
The kernel implementation of the fast system call path,
the one invoked by the SYSCALL instruction, is broken for
restarting system calls. A C program demonstrating this is below.
Change the system calls to use INT $0x80 instead, because
that (perhaps slightly slower) system call path actually works.
I filed http://www.freebsd.org/cgi/query-pr.cgi?pr=182161.
The C program demonstrating that it is FreeBSD's fault is below.
It reports the same "Bad address" failures from wait.
#include <sys/time.h>
#include <sys/signal.h>
#include <pthread.h>
#include <unistd.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
static void handler(int);
static void* looper(void*);
int
main(void)
{
int i;
struct sigaction sa;
pthread_cond_t cond;
pthread_mutex_t mu;
memset(&sa, 0, sizeof sa);
sa.sa_handler = handler;
sa.sa_flags = SA_RESTART;
memset(&sa.sa_mask, 0xff, sizeof sa.sa_mask);
sigaction(SIGCHLD, &sa, 0);
for(i=0; i<2; i++)
pthread_create(0, 0, looper, 0);
pthread_mutex_init(&mu, 0);
pthread_mutex_lock(&mu);
pthread_cond_init(&cond, 0);
for(;;)
pthread_cond_wait(&cond, &mu);
return 0;
}
static void
handler(int sig)
{
}
int
mywait4(int pid, int *stat, int options, struct rusage *rusage)
{
int result;
asm("movq %%rcx, %%r10; syscall"
: "=a" (result)
: "a" (7),
"D" (pid),
"S" (stat),
"d" (options),
"c" (rusage));
}
static void*
looper(void *v)
{
int pid, stat, out;
struct rusage rusage;
for(;;) {
if((pid = fork()) == 0)
_exit(0);
out = mywait4(pid, &stat, 0, &rusage);
if(out != pid) {
printf("wait4 returned %d\n", out);
}
}
}
Fixes#6372.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/13582047
The test 'gp == m->curg' is not valid on Windows,
because the goroutine being profiled is not from the
current m.
TBR=golang-dev
CC=golang-dev
https://golang.org/cl/13718043
Because profiling signals can arrive at any time, we must
handle the case where a profiling signal arrives halfway
through a goroutine switch. Luckily, although there is much
to think through, very little needs to change.
Fixes#6000.
Fixes#6015.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/13421048
Bug #1:
Issue 5406 identified an interesting case:
defer iface.M()
may end up calling a wrapper that copies an indirect receiver
from the iface value and then calls the real M method. That's
two calls down, not just one, and so recover() == nil always
in the real M method, even during a panic.
[For the purposes of this entire discussion, a wrapper's
implementation is a function containing an ordinary call, not
the optimized tail call form that is somtimes possible. The
tail call does not create a second frame, so it is already
handled correctly.]
Fix this bug by introducing g->panicwrap, which counts the
number of bytes on current stack segment that are due to
wrapper calls that should not count against the recover
check. All wrapper functions must now adjust g->panicwrap up
on entry and back down on exit. This adds slightly to their
expense; on the x86 it is a single instruction at entry and
exit; on the ARM it is three. However, the alternative is to
make a call to recover depend on being able to walk the stack,
which I very much want to avoid. We have enough problems
walking the stack for garbage collection and profiling.
Also, if performance is critical in a specific case, it is already
faster to use a pointer receiver and avoid this kind of wrapper
entirely.
Bug #2:
The old code, which did not consider the possibility of two
calls, already contained a check to see if the call had split
its stack and so the panic-created segment was one behind the
current segment. In the wrapper case, both of the two calls
might split their stacks, so the panic-created segment can be
two behind the current segment.
Fix this by propagating the Stktop.panic flag forward during
stack splits instead of looking backward during recover.
Fixes#5406.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/13367052
args is useful for printing tracebacks.
frame is not necessary anymore, but we might some day
get back to functions where the frame size does not vary
by program counter, and if so we'll need it. Avoid needing
to introduce a new struct format later by keeping it now.
Fixes#5907.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/13632051
The various throwing > 0 finish a change started
in a previous CL, which sets throwing = -1 to mean
"don't show the internals". That gets set during the
"all goroutines are asleep - deadlock!" crash, and it
should also be set during any other expected crash
that does not indicate a problem within the runtime.
Most runtime.throw do indicate a problem within the
runtime, however, so we should be able to enumerate
the ones that should be silent. The goroutine sleeping
deadlock is the only one I can think of.
Update #5139
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/13662043
Otherwise, if panic starts running deferred functions,
the code that panicked appears to be calling those
functions directly, which is not the case and can be
confusing.
For example:
main.Two()
/Users/rsc/x.go:12 +0x2a
runtime.panic(0x20dc0, 0x2100cc010)
/Users/rsc/g/go/src/pkg/runtime/panic.c:248 +0x106
main.One()
/Users/rsc/x.go:8 +0x55
This makes clear(er) that main.Two is being called during
a panic, not as a direct call from main.One.
Fixes#5832.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/13302051
This allows us to make two changes:
1. Force the argument type to be size_t, even on broken
systems that declare malloc to take a ulong.
2. Call runtime.throw if malloc fails.
(That is, the program crashes; it does not panic.)
Fixes#3403.
Fixes#5926.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/13413047
If using other gdb python scripts loaded before Go's gdb-runtime.py
and that have a different init prototype:
Traceback (most recent call last):
File "/usr/lib/go/src/pkg/runtime/runtime-gdb.py", line 446, in <module>
k()
TypeError: __init__() takes exactly 3 arguments (1 given)
The problem is that gdb keeps all python scripts in the same namespace,
so vars() contains them. To avoid that, load helpers one by one.
R=iant, rsc
CC=gobot, golang-dev
https://golang.org/cl/9752044
The code in question is trying to print a nice error message
when a Go EABI binary runs on an OABI machine.
Unfortunately, the only way to do that is to use
ARM Thumb instructions, which we otherwise don't use.
There exist ARM EABI machines that do not support Thumb.
We could run on them if not for this OABI check, so disable it.
Fixes#5685.
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/13234050
Currently lots of sys allocations are not accounted in any of XxxSys,
including GC bitmap, spans table, GC roots blocks, GC finalizer blocks,
iface table, netpoll descriptors and more. Up to ~20% can unaccounted.
This change introduces 2 new stats: GCSys and OtherSys for GC metadata
and all other misc allocations, respectively.
Also ensures that all XxxSys indeed sum up to Sys. All sys memory allocation
functions require the stat for accounting, so that it's impossible to miss something.
Also fix updating of mcache_sys/inuse, they were not updated after deallocation.
test/bench/garbage/parser before:
Sys 670064344
HeapSys 610271232
StackSys 65536
MSpanSys 14204928
MCacheSys 16384
BuckHashSys 1439992
after:
Sys 670064344
HeapSys 610271232
StackSys 65536
MSpanSys 14188544
MCacheSys 16384
BuckHashSys 3194304
GCSys 39198688
OtherSys 3129656
Fixes#5799.
R=rsc, dave, alex.brainman
CC=golang-dev
https://golang.org/cl/12946043
Make sure we never pass a timer into timerproc with
a negative duration since it will cause other timers
to never expire.
Fixes#5321.
R=golang-dev, minux.ma, remyoudompheng, mikioh.mikioh, r, bradfitz, rsc, dvyukov
CC=golang-dev
https://golang.org/cl/9035047
This replaces the mcall frame with the badmcall frame instead of
leaving the mcall frame on the stack and adding the badmcall frame.
Because mcall is no longer on the stack, traceback will now report what
called mcall, which is what we would like to see in this situation.
R=golang-dev, cshapiro
CC=golang-dev
https://golang.org/cl/13012044
When searching for an allocated bit, flushptrbuf would search
backward in the bitmap word containing the bit of pointer
being looked-up before searching the span. This extra check
was not replicated in markonly which, instead, after not
finding an allocated bit for a pointer would directly look in
the span.
Using statistics generated from godoc, before this change span
lookups were, on average, more common than word lookups. It
was common for markonly to consult spans for one third of its
pointer lookups. With this change in place, what were
previously span lookups are overwhelmingly become by the word
lookups making the total number of span lookups a relatively
small fraction of the whole.
This change also introduces some statistics gathering about
lookups guarded by the CollectStats enum.
R=golang-dev, khr
CC=golang-dev
https://golang.org/cl/13311043
#pragma textflag and #pragma dataflag directives.
Update dataflag directives to use symbols instead of integer constants.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/13310043
See how it flies. We'll disable it again if the underlying issue is not resolved.
See issue 4155 for details.
Fixes#4155.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/13368045
the use of the flag, especially for objects which actually do have
pointers but we don't want the GC to scan them.
R=golang-dev, cshapiro
CC=golang-dev
https://golang.org/cl/13181045
Go runtime support for dragonfly/amd64, largely based of the existing
FreeBSD runtime (with some clues from the varialus/godfly work).
R=bradfitz
CC=golang-dev
https://golang.org/cl/13088044
GC acquires worldsema, which is a goroutine-level semaphore
which parks goroutines. g0 can not be parked.
Fixes#6193.
R=khr, khr
CC=golang-dev
https://golang.org/cl/12880045
Update the original change but do not read interface types in
the arguments area. Once the arguments area is zeroed as the
locals area is we can safely read interface type values there
too.
««« original CL description
undo CL 12785045 / 71ce80dc4195
This has broken the 32-bit builds.
««« original CL description
cmd/gc, runtime: use type information to scan interface values
R=golang-dev, rsc, dvyukov
CC=golang-dev
https://golang.org/cl/12785045
»»»
R=khr, golang-dev, khr
CC=golang-dev
https://golang.org/cl/13010045
»»»
R=khr, khr
CC=golang-dev
https://golang.org/cl/13073045
This has broken the 32-bit builds.
««« original CL description
cmd/gc, runtime: use type information to scan interface values
R=golang-dev, rsc, dvyukov
CC=golang-dev
https://golang.org/cl/12785045
»»»
R=khr, golang-dev, khr
CC=golang-dev
https://golang.org/cl/13010045
In the crash stack trace race cgocall() calls endcgo(),
this means that m->racecall is wrong.
Indeed this can happen is a goroutine is rescheduled to another M
during race call.
Disable preemption for race calls.
Fixes#6155.
R=golang-dev, rsc, cshapiro
CC=golang-dev
https://golang.org/cl/12866045
Actually working to stay within the limit could cause subtle deadlocks.
Crashing avoids the subtlety.
Fixes#4056.
R=golang-dev, r, dvyukov
CC=golang-dev
https://golang.org/cl/13037043
Fixes#6107.
race: output goroutine 1 as main goroutine
Fixes#6130.
race: option to abort program on first detected error
R=golang-dev, mikioh.mikioh
CC=golang-dev
https://golang.org/cl/12968044
The goal is to stop only those programs that would keep
going and run the machine out of memory, but before they do that.
1 GB on 64-bit, 250 MB on 32-bit.
That seems implausibly large, and it can be adjusted.
Fixes#2556.
Fixes#4494.
Fixes#5173.
R=khr, r, dvyukov
CC=golang-dev
https://golang.org/cl/12541052
The baseline architecture had been left to the GCC configured
default which can be more accomodating than the rest of the Go
toolchain. This prevented instructions used by the 5g compiler,
like BLX, from being used in GCC compiled assembler code.
R=golang-dev, dave, rsc, elias.naur, cshapiro
CC=golang-dev
https://golang.org/cl/12954043
It doughtily misses all possible corner cases.
In particular on machines with <1GHz processors,
SetBlockProfileRate(1) disables profiling.
Fixes#6114.
R=golang-dev, bradfitz, rsc
CC=golang-dev
https://golang.org/cl/12936043
Originally the requirement was f(x) where f's argument is
exactly x's type.
CL 11858043 relaxed the requirement in a non-standard
way: f's argument must be exactly x's type or interface{}.
If we're going to relax the requirement, it should be done
in a way consistent with the rest of Go. This CL allows f's
argument to have any type for which x is assignable;
that's the same requirement the compiler would impose
if compiling f(x) directly.
Fixes#5368.
R=dvyukov, bradfitz, pieter
CC=golang-dev
https://golang.org/cl/12895043
The ARM external linking CL used BLX instructions in gcc assembler. Replace with BL to retain support on older ARM processors.
R=rsc
CC=golang-dev
https://golang.org/cl/12938043
The ARM external linking CL left missed changes to sys_freebsd_arm.s and sys_netbsd_arm.s already done to sys_linux_arm.s.
R=rsc
CC=golang-dev
https://golang.org/cl/12842044
This CL is an aggregate of 10271047, 10499043, 9733044. Descriptions of each follow:
10499043
runtime,cmd/ld: Merge TLS symbols and teach 5l about ARM TLS
This CL prepares for external linking support to ARM.
The pseudo-symbols runtime.g and runtime.m are merged into a single
runtime.tlsgm symbol. When external linking, the offset of a thread local
variable is stored at a memory location instead of being embedded into a offset
of a ldr instruction. With a single runtime.tlsgm symbol for both g and m, only
one such offset is needed.
The larger part of this CL moves TLS code from gcc compiled to internally
compiled. The TLS code now uses the modern MRC instruction, and 5l is taught
about TLS fallbacks in case the instruction is not available or appropriate.
10271047
This CL adds support for -linkmode external to 5l.
For 5l itself, use addrel to allow for D_CALL relocations to be handled by the
host linker. Of the cases listed in rsc's comment in issue 4069, only case 5 and
63 needed an update. One of the TODO: addrel cases was since replaced, and the
rest of the cases are either covered by indirection through addpool (cases with
LTO or LFROM flags) or stubs (case 74). The addpool cases are covered because
addpool emits AWORD instructions, which in turn are handled by case 11.
In the runtime, change the argv argument in the rt0* functions slightly to be a
pointer to the argv list, instead of relying on a particular location of argv.
9733044
The -shared flag to 6l outputs a shared library, implemented in Go
and callable from non-Go programs such as C.
The main part of this CL change the thread local storage model.
Go uses the fastest and least general mode, local exec. TLS data in shared
libraries normally requires at least the local dynamic mode, however, this CL
instead opts for using the initial exec mode. Initial exec mode is faster than
local dynamic mode and can be used in linux since the linker has reserved a
limited amount of TLS space for performance sensitive TLS code.
Initial exec mode requires an extra load from the GOT table to determine the
TLS offset. This penalty will not be paid if ld is not in -shared mode, since
TLS accesses will be reduced to local exec.
The elf sections .init_array and .rela.init_array are added to register the Go
runtime entry with cgo at library load time.
The "hidden" attribute is added to Cgo functions called from Go, since Go
does not generate call through the GOT table, and adding non-GOT relocations for
a global function is not supported by gcc. Cgo symbols don't need to be global
and avoiding the GOT table is also faster.
The changes to 8l are only removes code relevant to the old -shared mode where
internal linking was used.
This CL only address the low level linker work. It can be submitted by itself,
but to be useful, the runtime changes in CL 9738047 is also needed.
Design discussion at
https://groups.google.com/forum/?fromgroups#!topic/golang-nuts/zmjXkGrEx6QFixes#5590.
R=rsc
CC=golang-dev
https://golang.org/cl/12871044
Breaks the build. Old bucket arrays kept by iterators
still need to be scanned.
««« original CL description
runtime: tell GC not to scan internal hashmap structures.
We'll do it ourselves via hash_gciter, thanks.
Fixes bug 6119.
R=golang-dev, dvyukov, cookieo9, rsc
CC=golang-dev
https://golang.org/cl/12840043
»»»
R=golang-dev
CC=golang-dev
https://golang.org/cl/12884043
The NetBSD and OpenBSD failures are apparently real,
not due to the test bug fixed in 100b9fc0c46f.
««« original CL description
runtime/pprof: test netbsd and openbsd again
Maybe these will work now.
R=golang-dev, dvyukov, bradfitz
CC=golang-dev
https://golang.org/cl/12787044
»»»
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/12873043
Currently it's possible that a goroutine
that periodically executes non-blocking
cgo/syscalls is never preempted.
This change splits scheduler and syscall
ticks to prevent such situation.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12658045
Currently we lose lots of profiling signals.
Most notably, GC is not accounted at all.
But stack splits, scheduler, syscalls, etc are lost as well.
This creates seriously misleading profile.
With this change all profiling signals are accounted.
Now I see these additional entries that were previously absent:
161 29.7% 29.7% 164 30.3% syscall.Syscall
12 2.2% 50.9% 12 2.2% scanblock
11 2.0% 55.0% 11 2.0% markonly
10 1.8% 58.9% 10 1.8% sweepspan
2 0.4% 85.8% 2 0.4% runtime.newstack
It is still impossible to understand what causes stack splits,
but at least it's clear how many time is spent on them.
Update #2197.
Update #5659.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12179043
If the timer goroutine is wakeup by timeout,
other goroutines will still notewakeup because sleeping is still set.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/12763043
The original plan was to collect allocation stacks
for all memory blocks. But it was never implemented
and it's not in near plans and it's unclear how to do it at all.
R=golang-dev, dave, bradfitz
CC=golang-dev
https://golang.org/cl/12724044
Prior to this change, pointer maps encoded the disposition of
a word using a single bit. A zero signaled a non-pointer
value and a one signaled a pointer value. Interface values,
which are a effectively a union type, were conservatively
labeled as a pointer.
This change widens the logical element size of the pointer map
to two bits per word. As before, zero signals a non-pointer
value and one signals a pointer value. Additionally, a two
signals an iface pointer and a three signals an eface pointer.
Following other changes to the runtime, values two and three
will allow a type information to drive interpretation of the
subsequent word so only those interface values containing a
pointer value will be scanned.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12689046
I've placed net.runtime_Semacquire into netpoll.goc,
but netbsd does not yet use netpoll.goc.
R=golang-dev, bradfitz, iant
CC=golang-dev
https://golang.org/cl/12699045
The mutex, fdMutex, handles locking and lifetime of sysfd,
and serializes Read and Write methods.
This allows to strip 2 sync.Mutex.Lock calls,
2 sync.Mutex.Unlock calls, 1 defer and some amount
of misc overhead from every network operation.
On linux/amd64, Intel E5-2690:
benchmark old ns/op new ns/op delta
BenchmarkTCP4Persistent 9595 9454 -1.47%
BenchmarkTCP4Persistent-2 8978 8772 -2.29%
BenchmarkTCP4ConcurrentReadWrite 4900 4625 -5.61%
BenchmarkTCP4ConcurrentReadWrite-2 2603 2500 -3.96%
In general it strips 70-500 ns from every network operation depending
on processor model. On my relatively new E5-2690 it accounts to ~5%
of network op cost.
Fixes#6074.
R=golang-dev, bradfitz, alex.brainman, iant, mikioh.mikioh
CC=golang-dev
https://golang.org/cl/12418043
Introduce freezetheworld function that is a best-effort attempt to stop any concurrently running goroutines. Call it during crash.
Fixes#5873.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12054044
GetQueuedCompletionStatusEx allows to dequeue a batch of completion
notifications, which is more efficient than dequeueing one by one.
benchmark old ns/op new ns/op delta
BenchmarkClientServerParallel4 100605 90945 -9.60%
BenchmarkClientServerParallel4-2 90225 74504 -17.42%
R=golang-dev, alex.brainman
CC=golang-dev
https://golang.org/cl/12436044
The test takes up to 64 seconds on windows builders.
I've tried to reduce number of iterations in the test,
but it does not affect run time.
Fixes#6054.
R=golang-dev, alex.brainman
CC=golang-dev
https://golang.org/cl/12531043
Previously, all word aligned locations in the local variables
area were scanned as conservative roots. With this change, a
bitmap is generated describing the locations of pointer values
in local variables.
With this change the argument bitmap information has been
changed to only store information about arguments. The locals
member, has been removed. In its place, the bitmap data for
local variables is now used to store the size of locals. If
the size is negative, the magnitude indicates the size of the
local variables area.
R=rsc
CC=golang-dev
https://golang.org/cl/12328044
Remove NOPROF/DUPOK from everything.
Edits done with a script, except pclinetest.asm which depended
on the DUPOK flag on main().
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/12613044
We can then include this file in assembly to replace
cryptic constants like "7" with meaningful constants
like "(NOPROF|DUPOK|NOSPLIT)".
Converting just pkg/runtime/asm*.s for now. Dropping NOPROF
and DUPOK from lots of places where they aren't needed.
More .s files to come in a subsequent changelist.
A nonzero number in the textflag field now means
"has not been converted yet".
R=golang-dev, daniel.morsing, rsc, khr
CC=golang-dev
https://golang.org/cl/12568043
Updates #6046.
This CL just does maxstring and concatstring. There are other functions
to fix but doing them a few at a time will help isolate any (unlikely)
breakages these changes bring up in architectures I can't test
myself.
R=golang-dev, dave, iant
CC=golang-dev
https://golang.org/cl/12519044
NetBSD and OpenBSD are broken like OS X is. Good to know.
Drop required count from avg/2 to avg/3, because the
Plan 9 builder just barely missed avg/2 in one of its runs.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/12548043
Update #6046.
This CL just does findnull and findnullw. There are other functions
to fix but doing them a few at a time will help isolate any (unlikely)
breakages these changes bring up in architectures I can't test
myself.
R=golang-dev, dsymonds
CC=golang-dev
https://golang.org/cl/12520043
Embed all data necessary for read/write operations directly into netFD.
benchmark old ns/op new ns/op delta
BenchmarkTCP4Persistent 27669 23341 -15.64%
BenchmarkTCP4Persistent-2 18173 12558 -30.90%
BenchmarkTCP4Persistent-4 10390 7319 -29.56%
This change will intentionally break all builders to see
how many allocations they do per read/write.
This will be fixed soon afterwards.
R=golang-dev, alex.brainman
CC=golang-dev
https://golang.org/cl/12413043
gcpc/gcsp are used by GC in similar situation.
gcpc/gcsp are also more stable than gp->sched,
because gp->sched is mutated by entersyscall/exitsyscall
in morestack and mcall. So it has higher chances of being inconsistent.
Also, rename gcpc/gcsp to syscallpc/syscallsp.
This is the same as reverted change 12250043
with save marked as textflag 7.
The problem was that if save calls morestack,
then subsequent lessstack spoils g->sched.pc/sp.
And that bad values were remembered in g->syscallpc/sp.
Entersyscallblock had the same problem,
but it was never triggered to date.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12478043
Basically a partial rollback of 12053043 until I can
figure out what is really going on.
Fixes bug 6051.
R=golang-dev
CC=golang-dev
https://golang.org/cl/12496043
This means that pprof will no longer report profiles on OS X.
That's unfortunate, but the profiles were often wrong and, worse,
it was difficult to tell whether the profile was wrong or not.
The workarounds were making the scheduler more complex,
possibly caused a deadlock (see issue 5519), and did not actually
deliver reliable results.
It may be possible for adventurous users to apply a patch to
their kernels to get working results, or perhaps having no results
will encourage someone to do the work of creating a profiling
thread like on Windows. Issue 6047 has details.
Fixes#5519.
Fixes#6047.
R=golang-dev, bradfitz, r
CC=golang-dev
https://golang.org/cl/12429045
Break all 386 builders.
««« original CL description
runtime: use gcpc/gcsp during traceback of goroutines in syscalls
gcpc/gcsp are used by GC in similar situation.
gcpc/gcsp are also more stable than gp->sched,
because gp->sched is mutated by entersyscall/exitsyscall
in morestack and mcall. So it has higher chances of being inconsistent.
Also, rename gcpc/gcsp to syscallpc/syscallsp.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12250043
»»»
R=rsc
CC=golang-dev
https://golang.org/cl/12424045
It was needed for the old scheduler,
because there temporary could be more threads than gomaxprocs.
In the new scheduler gomaxprocs is always respected.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12438043
gcpc/gcsp are used by GC in similar situation.
gcpc/gcsp are also more stable than gp->sched,
because gp->sched is mutated by entersyscall/exitsyscall
in morestack and mcall. So it has higher chances of being inconsistent.
Also, rename gcpc/gcsp to syscallpc/syscallsp.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12250043
Windows dynamic priority boosting assumes that a process has different types
of dedicated threads -- GUI, IO, computational, etc. Go processes use
equivalent threads that all do a mix of GUI, IO, computations, etc.
In such context dynamic priority boosting does nothing but harm, so turn it off.
In particular, if 2 goroutines do heavy IO on a server uniprocessor machine,
windows rejects to schedule timer thread for 2+ seconds when priority boosting is enabled.
Fixes#5971.
R=alex.brainman
CC=golang-dev
https://golang.org/cl/12406043
It's okay to preempt at ordinary function calls because
compilers arrange that there are no live registers to save
on entry to the function call.
The software floating point routines are function calls
masquerading as individual machine instructions. They are
expected to keep all the registers intact. In particular,
they are expected not to clobber all the floating point
registers.
The floating point registers are kept per-M, because they
are not live at non-preemptive goroutine scheduling events,
and so keeping them per-M reduces the number of 132-byte
register blocks we are keeping in memory.
Because they are per-M, allowing the goroutine to be
rescheduled during software floating point simulation
would mean some other goroutine could overwrite the registers
or perhaps the goroutine would continue running on a different
M entirely.
Disallow preemption during the software floating point
routines to make sure that a function full of floating point
instructions has the same floating point registers throughout
its execution.
R=golang-dev, dave
CC=golang-dev
https://golang.org/cl/12298043
Per suggestion from Russ in February. Then strings.IndexByte
can be implemented in terms of the shared code in pkg runtime.
Update #3751
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/12289043
This allows to at least determine goroutine "identity".
Now it looks like:
goroutine 12 [running]:
goroutine running on other thread; stack unavailable
created by testing.RunTests
src/pkg/testing/testing.go:440 +0x88e
R=golang-dev, r, rsc
CC=golang-dev
https://golang.org/cl/12248043
We see timeouts in these tests on some platforms,
but not on the others. The hypothesis is that
the problematic platforms are slow uniprocessors.
Stack traces do not suggest that the process
is completely hang, and it is able to schedule
the alarm goroutine. And if it actually hangs,
we still will be able to detect that.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/12253043
Preemption during the software floating point code
could cause m (R9) to change, so that when the
original registers were restored at the end of the
floating point handler, the changed and correct m
would be replaced by the old and incorrect m.
TBR=dvyukov
CC=golang-dev
https://golang.org/cl/11883045
Operations on int64 are very stack consuming with 5c.
Fixes netbsd/arm build.
Before: TEXT runtime.timediv+0(SB),7,$52-16
After: TEXT runtime.timediv+0(SB),7,$44-16
The stack usage is unchanged on 386:
TEXT runtime.timediv+0(SB),7,$8-16
R=golang-dev, dvyukov, bradfitz
CC=golang-dev
https://golang.org/cl/12182044
Update #5139.
Double wakeup on Note was reported several times,
but no reliable reproducer.
There also was a strange report about weird value of epoll fd.
Maybe it's corruption of global data...
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/12182043
Sysmon thread parks if no goroutines are running (runtime.sched.npidle ==
runtime.gomaxprocs).
Currently it's unparked when a goroutine enters syscall, it was enough
to retake P's from blocking syscalls.
But it's not enough for reliable goroutine preemption. We need to ensure that
sysmon runs if any goroutines are running.
R=rsc
CC=golang-dev
https://golang.org/cl/12176043
Submitted with some unrelated changes that were not intended to go in.
««« original CL description
runtime: do not park sysmon thread if any goroutines are running
Sysmon thread parks if no goroutines are running (runtime.sched.npidle == runtime.gomaxprocs).
Currently it's unparked when a goroutine enters syscall, it was enough
to retake P's from blocking syscalls.
But it's not enough for reliable goroutine preemption. We need to ensure that
sysmon runs if any goroutines are running.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12167043
»»»
R=rsc
CC=golang-dev
https://golang.org/cl/12171044
This is required to properly unwind reflect.methodValueCall/makeFuncStub.
Fixes#5954.
Stats for 'go install std':
61849 total INSTCALL
24655 currently have ArgSize metadata
27278 have ArgSize metadata with this change
godoc size before: 11351888, after: 11364288
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12163043
Sysmon thread parks if no goroutines are running (runtime.sched.npidle == runtime.gomaxprocs).
Currently it's unparked when a goroutine enters syscall, it was enough
to retake P's from blocking syscalls.
But it's not enough for reliable goroutine preemption. We need to ensure that
sysmon runs if any goroutines are running.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12167043
When comparing strings, check these (in order):
- length mismatch => not equal
- string pointer equal => equal
- if length is short:
- memeq on body
- if length is long:
- compare first&last few bytes, if different => not equal
- save entry as a possible match
- after checking every entry, if there is only one possible
match, use memeq on that entry. Otherwise, fallback to hash.
benchmark old ns/op new ns/op delta
BenchmarkSameLengthMap 43 4 -89.77%
Fixes#5194.
Update #3885.
R=golang-dev, bradfitz, khr, rsc
CC=golang-dev
https://golang.org/cl/12128044
struct Hmap is the header for a map value.
CL 8377046 made flags a uint32 so that it could be updated atomically,
but that bumped the struct to 56 bytes, which allocates as 64 bytes (on amd64).
hash0 is initialized from runtime.fastrand1, which returns a uint32,
so the top 32 bits were always zero anyway. Declare it as a uint32
to reclaim 4 bytes and bring the Hmap size back down to a 48-byte allocation.
Fixes#5237.
R=golang-dev, khr, khr
CC=bradfitz, dvyukov, golang-dev
https://golang.org/cl/12034047
notetsleep: nosplit stack overflow
120 assumed on entry to notetsleep
96 after notetsleep uses 24
88 on entry to runtime.semasleep
32 after runtime.semasleep uses 56
24 on entry to runtime.nanotime
-8 after runtime.nanotime uses 32
Nanotime seems to be using only 24 bytes of stack space.
Unless I am missing something.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12041044
notetsleep: nosplit stack overflow
120 assumed on entry to notetsleep
80 after notetsleep uses 40
72 on entry to runtime.futexsleep
16 after runtime.futexsleep uses 56
8 on entry to runtime.printf
-16 after runtime.printf uses 24
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/12047043
Split stack checks (morestack) corrupt g->sched,
but g->sched must be preserved consistent for GC/traceback.
The change implements runtime.notetsleepg function,
which does entersyscall/exitsyscall and is carefully arranged
to not call any split functions in between.
R=rsc
CC=golang-dev
https://golang.org/cl/11575044
If netpoll has been told to block, it must not return with nil,
otherwise scheduler assumes that netpoll is disabled.
R=golang-dev, alex.brainman
CC=golang-dev
https://golang.org/cl/11920044
Make it accept type, combine flags.
Several reasons for the change:
1. mallocgc and settype must be atomic wrt GC
2. settype is called from only one place now
3. it will help performance (eventually settype
functionality must be combined with markallocated)
4. flags are easier to read now (no mallocgc(sz, 0, 1, 0) anymore)
R=golang-dev, iant, nightlyone, rsc, dave, khr, bradfitz, r
CC=golang-dev
https://golang.org/cl/10136043
Currently Darwin and FreeBSD support and NetBSD and OpenBSD do not
support EV_RECEIPT flag. We will drop use of EV_RECEIPT for now.
Also enables to build runtime-integrated network pollster on
freebsd/amd64,386 and openbsd/amd64,386. It just does build but never
runs pollster stuff.
This is in preparation for runtime-integrated network pollster for BSD
variants.
Update #5199
R=dvyukov, minux.ma
CC=golang-dev
https://golang.org/cl/11759044
Debugging the Windows breakage I noticed that SEH
only exists on 386, so we can balance the two stacks
a little more on amd64 and reclaim another word.
Now we're down to just one word consumed by
cgocallback_gofunc, having reclaimed 25% of the
overall budget (4 words out of 16).
Separately, fix windows/386 - the SEH must be on the
m0 stack, as must the saved SP, so we are forced to have
a three-word frame for 386. It matters much less for
386, because there 128 bytes gives 32 words to use.
R=dvyukov, alex.brainman
CC=golang-dev
https://golang.org/cl/11551044
Tying preemption to stack splits means that we have to able to
complete the call to exitsyscall (inside cgocallbackg at least for now)
without any stack split checks, meaning that the whole sequence
has to work within 128 bytes of stack, unless we increase the size
of the red zone. This CL frees up 24 bytes along that critical path
on amd64. (The 32-bit systems have plenty of space because all
their words are smaller.)
R=dvyukov
CC=golang-dev
https://golang.org/cl/11676043
Change use of x+(SP) to access the stack frame into x-(SP)
Fixes#5925.
R=golang-dev, bradfitz, dave, remyoudompheng, nick, rsc
CC=dave cheney <dave, golang-dev
https://golang.org/cl/11647043
notetsleepg is the same as notetsleep, but is called on user g.
It includes entersyscall/exitsyscall and will help to avoid
split stack functions in syscall status.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/11681043
This CL introduces a FUNCDATA number for runtime-specific
garbage collection metadata, changes the C and Go compilers
to emit that metadata, and changes the runtime to expect it.
The old pseudo-instructions that carried this information
are gone, as is the linker code to process them.
R=golang-dev, dvyukov, cshapiro
CC=golang-dev
https://golang.org/cl/11406044
whose argument size is unknown (C vararg functions, and
assembly code without an explicit specification).
We used to use 0 to mean "unknown" and 1 to mean "zero".
Now we use ArgsSizeUnknown (0x80000000) to mean "unknown".
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/11590043
If the network is not polled for 10ms, sysmon starts polling network
on every iteration (every 20us) until another thread blocks in netpoll.
Fixes#5922.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/11569043
It assumes that the m will not change, and the m may
change if the goroutine is preempted.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/11560043
If we start a garbage collection on g0 during a
stack split or unsplit, we'll see morestack or lessstack
at the top of the stack. Record an argument frame size
for those, and record that they terminate the stack.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/11533043
Deferreturn is synthesizing a new call frame.
It must not be interrupted between copying the args there
and fixing up the program counter, or else the stack will
be in an inconsistent state, one that will confuse the
garbage collector.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/11522043
With preemption, _sfloat2 can show up in stack traces.
Write the function prototype in a way that accurately
shows the frame size and the fact that it might contain
pointers.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/11523043
Windows was the only one seeing this bug reliably in the builder,
but it was easy to reproduce using 'GOGC=1 go test strconv'.
concatstring looked like it took only one string, but in fact it
takes a long list of strings. Add an explicit ... so that the traceback
will not use the "fixed" frame size and instead look at the
frame size metadata recorded by the caller.
R=golang-dev
TBR=golang-dev
CC=golang-dev
https://golang.org/cl/11531043
Otherwise the tests in pkg/runtime fail:
runtime: unknown argument frame size for runtime.deferreturn called from 0x48657b [runtime_test.func·022]
fatal error: invalid stack
...
R=golang-dev, dave
CC=golang-dev
https://golang.org/cl/11483043
Update #543
I believe the runtime is strong enough now to reenable
preemption during the function prologue.
Assuming this is or can be made stable, it will be in Go 1.2.
More aggressive preemption is not planned for Go 1.2.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/11433045
Currently preemption signal g->stackguard0==StackPreempt
can be lost if it is received when preemption is disabled
(e.g. m->lock!=0). This change duplicates the preemption
signal in g->preempt and restores g->stackguard0
when preemption is enabled.
Update #543.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/10792043
With this CL, I believe the runtime always knows
the frame size during the gc walk. There is no fallback
to "assume entire stack frame of caller" anymore.
R=golang-dev, khr, cshapiro, dvyukov
CC=golang-dev
https://golang.org/cl/11374044
I have not done the system call stubs in sys_*.s.
I hope to avoid that, because those do not block, so those
frames will not appear in stack traces during garbage
collection.
R=golang-dev, dvyukov, khr
CC=golang-dev
https://golang.org/cl/11360043
Design at http://golang.org/s/go12symtab.
This enables some cleanup of the garbage collector metadata
that will be done in future CLs.
This CL does not move the old symtab and pclntab back into
an unmapped section of the file. That's a bit tricky and will be
done separately.
Fixes#4020.
R=golang-dev, dave, cshapiro, iant, r
CC=golang-dev, nigeltao
https://golang.org/cl/11085043
A type switch on a value with map index expressions,
could get a spurious instrumentation from a OTYPESW node.
These nodes do not need instrumentation because after
walk the type switch has been turned into a sequence
of ifs.
Fixes#5890.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/11308043