Idle GC workers trigger whenever there's a GC running and the
scheduler doesn't find any other work. However, they currently run for
a full scheduler quantum (~10ms) once started.
This is really bad for event-driven applications, where work may come
in on the network hundreds of times during that window. In the
go-gcbench rpc benchmark, this is bad enough to often cause effective
STWs where all Ps are in the idle worker. When this happens, we don't
even poll the network any more (except for the background 10ms poll in
sysmon), so we don't even know there's more work to do.
Fix this by making idle workers check with the scheduler roughly every
100 µs to see if there's any higher-priority work the P should be
doing. This check includes polling the network for incoming work.
Fixes#16528.
Change-Id: I6f62ebf6d36a92368da9891bafbbfd609b9bd003
Reviewed-on: https://go-review.googlesource.com/32433
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
I don't have any way to test or reproduce this problem,
but the current code is clearly wrong for Windows.
Make it better.
As I said on #17165:
But the borrowing of M's and the profiling of M's by the CPU profiler
seem not synchronized enough. This code implements the CPU profiler
on Windows:
func profileloop1(param uintptr) uint32 {
stdcall2(_SetThreadPriority, currentThread, _THREAD_PRIORITY_HIGHEST)
for {
stdcall2(_WaitForSingleObject, profiletimer, _INFINITE)
first := (*m)(atomic.Loadp(unsafe.Pointer(&allm)))
for mp := first; mp != nil; mp = mp.alllink {
thread := atomic.Loaduintptr(&mp.thread)
// Do not profile threads blocked on Notes,
// this includes idle worker threads,
// idle timer thread, idle heap scavenger, etc.
if thread == 0 || mp.profilehz == 0 || mp.blocked {
continue
}
stdcall1(_SuspendThread, thread)
if mp.profilehz != 0 && !mp.blocked {
profilem(mp)
}
stdcall1(_ResumeThread, thread)
}
}
}
func profilem(mp *m) {
var r *context
rbuf := make([]byte, unsafe.Sizeof(*r)+15)
tls := &mp.tls[0]
gp := *((**g)(unsafe.Pointer(tls)))
// align Context to 16 bytes
r = (*context)(unsafe.Pointer((uintptr(unsafe.Pointer(&rbuf[15]))) &^ 15))
r.contextflags = _CONTEXT_CONTROL
stdcall2(_GetThreadContext, mp.thread, uintptr(unsafe.Pointer(r)))
sigprof(r.ip(), r.sp(), 0, gp, mp)
}
func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
if prof.hz == 0 {
return
}
// Profiling runs concurrently with GC, so it must not allocate.
mp.mallocing++
... lots of code ...
mp.mallocing--
}
A borrowed M may migrate between threads. Between the
atomic.Loaduintptr(&mp.thread) and the SuspendThread, mp may have
moved to a new thread, so that it's in active use. In particular
it might be calling malloc, as in the crash stack trace. If so, the
mp.mallocing++ in sigprof would provoke the crash.
Those lines are trying to guard against allocation during sigprof.
But on Windows, mp is the thread being traced, not the current
thread. Those lines should really be using getg().m.mallocing, which
is the same on Unix but not on Windows. With that change, it's
possible the race on the actual thread is not a problem: the traceback
would get confused and eventually return an error, but that's fine.
The code expects that possibility.
Fixes#17165.
Change-Id: If6619731910d65ca4b1a6e7de761fa2518ef339e
Reviewed-on: https://go-review.googlesource.com/33132
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
The introduction of -buildmode=plugin means modules can be added to a
Go program while it is running. This means there exists some time
while the program is running with the module is on the moduledata
linked list, but it has not been initialized to the satisfaction of
other parts of the runtime. Notably, the GC.
This CL adds a new way of access modules, an activeModules function.
It returns a slice of modules that is built in the background and
atomically swapped in. The parts of the runtime that need to wait on
module initialization can use this slice instead of the linked list.
Fixes#17455
Change-Id: I04790fd07e40c7295beb47cea202eb439206d33d
Reviewed-on: https://go-review.googlesource.com/32357
Reviewed-by: Ian Lance Taylor <iant@golang.org>
- Adds overflow checks
- Adds parsing of negative integers
- Adds boolean return value to signal parsing errors
- Adds atoi32 for parsing of integers that fit in an int32
- Adds tests
Handling of errors to provide error messages
at the call sites is left to future CLs.
Updates #17718
Change-Id: I3cacd0ab1230b9efc5404c68edae7304d39bcbc0
Reviewed-on: https://go-review.googlesource.com/32390
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Currently, we perform write barriers after performing pointer writes.
At the moment, it simply doesn't matter what order this happens in, as
long as they appear atomic to GC. But both the hybrid barrier and ROC
are going to require a pre-write write barrier.
For the hybrid barrier, this is important because the barrier needs to
observe both the current value of the slot and the value that will be
written to it. (Alternatively, the caller could do the write and pass
in the old value, but it seems easier and more useful to just swap the
order of the barrier and the write.)
For ROC, this is necessary because, if the pointer write is going to
make the pointer reachable to some goroutine that it currently is not
visible to, the garbage collector must take some special action before
that pointer becomes more broadly visible.
This commits swaps pointer writes around so the write barrier occurs
before the pointer write.
The main subtlety here is bulk memory writes. Currently, these copy to
the destination first and then use the pointer bitmap of the
destination to find the copied pointers and invoke the write barrier.
This is necessary because the source may not have a pointer bitmap. To
handle these, we pass both the source and the destination to the bulk
memory barrier, which uses the pointer bitmap of the destination, but
reads the pointer values from the source.
Updates #17503.
Change-Id: I78ecc0c5c94ee81c29019c305b3d232069294a55
Reviewed-on: https://go-review.googlesource.com/31763
Reviewed-by: Rick Hudson <rlh@golang.org>
As for dropg, save is writing a nil pointer that will generate a write
barrier with the hybrid barrier. However, in this case, ctxt always
should already be nil, so replace the write with an assertion that
this is the case.
At this point, we're ready to disable the write barrier elision
optimizations that interfere with the hybrid barrier.
Updates #17503.
Change-Id: I83208e65aa33403d442401f355b2e013ab9a50e9
Reviewed-on: https://go-review.googlesource.com/31571
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently this contains no write barriers because it's writing nil
pointers, but with the hybrid barrier, even these will produce write
barriers. However, since these are *gs and *ms, they don't need write
barriers, so we can simply eliminate them.
Updates #17503.
Change-Id: Ib188a60492c5cfb352814bf9b2bcb2941fb7d6c0
Reviewed-on: https://go-review.googlesource.com/31570
Reviewed-by: Rick Hudson <rlh@golang.org>
The hybrid barrier requires barriers on stack-to-stack copies if
either stack is grey. There are only two instances of this in the
runtime: channel sends and starting a goroutine. Channel sends already
use typedmemmove and hence have the necessary barriers. This commits
adds barriers for the stack-to-stack copy when starting a goroutine.
Updates #17503.
Change-Id: Ibb55e08127ca4d021ac54be61cb96732efa5df5b
Reviewed-on: https://go-review.googlesource.com/31455
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently we initialize LR on a new stack by writing nil to it. But
this is an initializing write since the newly allocated stack is not
zeroed, so this is unsafe with the hybrid barrier. Change this is a
uintptr write to avoid a bad write barrier.
Updates #17503.
Change-Id: I062ac352e35df7da4644c1f2a5aaab87049d1f60
Reviewed-on: https://go-review.googlesource.com/32093
Reviewed-by: Rick Hudson <rlh@golang.org>
Since barrier-less memclr is only safe in very narrow circumstances,
this commit renames memclr to avoid accidentally calling memclr on
typed memory. This can cause subtle, non-deterministic bugs, so it's
worth some effort to prevent. In the near term, this will also prevent
bugs creeping in from any concurrent CLs that add calls to memclr; if
this happens, whichever patch hits master second will fail to compile.
This also adds the other new memclr variants to the compiler's
builtin.go to minimize the churn on that binary blob. We'll use these
in future commits.
Updates #17503.
Change-Id: I00eead049f5bd35ca107ea525966831f3d1ed9ca
Reviewed-on: https://go-review.googlesource.com/31369
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
runtime.SetMutexProfileFraction(n int) will capture 1/n-th of stack
traces of goroutines holding contended mutexes if n > 0. From runtime/pprof,
pprot.Lookup("mutex").WriteTo writes the accumulated
stack traces to w (in essentially the same format that blocking
profiling uses).
Change-Id: Ie0b54fa4226853d99aa42c14cb529ae586a8335a
Reviewed-on: https://go-review.googlesource.com/29650
Reviewed-by: Austin Clements <austin@google.com>
Fixes#16076
Change-Id: I91fa87b642592ee4604537dd8c3197cd61ec8b31
Reviewed-on: https://go-review.googlesource.com/31516
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
oneNewExtraM creates a spare M and G for use with cgo callbacks. The G
doesn't run right away, but goes directly into syscall status. For the
garbage collector, it's marked as "scan valid" and not on the rescan
list, but I forgot to also mark it as "scan done". As a result,
gcMarkRootCheck thinks that the goroutine hasn't been scanned and
panics.
This only affects GODEBUG=gccheckmark=1 mode, since we otherwise skip
the gcMarkRootCheck.
Fixes#17473.
Change-Id: I94f5671c42eb44bd5ea7dc68fbf85f0c19e2e52c
Reviewed-on: https://go-review.googlesource.com/31139
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
If morestack runs on the g0 or gsignal stack, it currently performs
some abort operation that typically produces a signal (e.g., it does
an INT $3 on x86). This is useful if you're running in a debugger, but
if you're not, the runtime tries to trap this signal, which is likely
to send the program into a deeper spiral of collapse and lead to very
confusing diagnostic output.
Help out people trying to debug without a debugger by making morestack
print an informative message before blowing up.
Change-Id: I2814c64509b137bfe20a00091d8551d18c2c4749
Reviewed-on: https://go-review.googlesource.com/31133
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Currently we use go:nowritebarrier in many places in proc.go.
go:notinheap and go:yeswritebarrierrec now let us use
go:nowritebarrierrec (the recursive form of the go:nowritebarrier
pragma) more liberally. Do so in proc.go
Change-Id: Ia7fcbc12ce6c51cb24730bf835fb7634ad53462f
Reviewed-on: https://go-review.googlesource.com/30942
Reviewed-by: Rick Hudson <rlh@golang.org>
This pragma cancels the effect of go:nowritebarrierrec. This is useful
in the scheduler because there are places where we enter a function
without a valid P (and hence cannot have write barriers), but then
obtain a P. This allows us to annotate the function with
go:nowritebarrierrec and split out the part after we've obtained a P
into a go:yeswritebarrierrec function.
Change-Id: Ic8ce4b6d3c074a1ecd8280ad90eaf39f0ffbcc2a
Reviewed-on: https://go-review.googlesource.com/30938
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
ARM direct CALL/JMP instruction has 24 bit offset, which can only
encodes jumps within +/-32M. When the target is too far, the top
bits get truncated and the program jumps wild.
This CL detects too-far jumps and automatically insert trampolines,
currently only internal linking on ARM.
It is necessary to make the following changes to the linker:
- Resolve direct jump relocs when assigning addresses to functions.
this allows trampoline insertion without moving all code that
already laid down.
- Lay down packages in dependency order, so that when resolving a
inter-package direct jump reloc, the target address is already
known. Intra-package jumps are assumed never too far.
- a linker flag -debugtramp is added for debugging trampolines:
"-debugtramp=1 -v" prints trampoline debug message
"-debugtramp=2" forces all inter-package jump to use
trampolines (currently ARM only)
"-debugtramp=2 -v" does both
- Some data structures are changed for bookkeeping.
On ARM, pseudo DIV/DIVU/MOD/MODU instructions now clobber R8
(unfortunate). In the standard library there is no ARM assembly
code that uses these instructions, and the compiler no longer emits
them (CL 29390).
all.bash passes with -debugtramp=2, except a disassembly test (this
is unavoidable as we changed the instruction).
TBD: debug info of trampolines?
Fixes#17028.
Change-Id: Idcce347ea7e0af77c4079041a160b2f6e114b474
Reviewed-on: https://go-review.googlesource.com/29397
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
If we get a SIGPROF on a non-Go thread, and the program has not called
runtime.SetCgoTraceback so we have no way to collect a stack trace, then
record a profile that is just the PC where the signal occurred. That
will at least point the user to the right area.
Retrieving the PC from the sigctxt in a signal handler on a non-G thread
required marking a number of trivial sigctxt methods as nosplit, and,
for extra safety, nowritebarrierrec.
The test shows that the existing test CgoPprofThread test does not test
the stack trace, just the profile signal. Leaving that for later.
Change-Id: I8f8f3ff09ac099fc9d9df94b5a9d210ffc20c4ab
Reviewed-on: https://go-review.googlesource.com/30252
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
PC passed to racegostart is expected to be a return PC
of the go statement. Race runtime will subtract 1 from the PC
before symbolization. Passing start PC of a function is wrong.
Add sys.PCQuantum to the function start PC.
Update #17190
Change-Id: Ia504c49e79af84ed4ea360c2aea472b370ea8bf5
Reviewed-on: https://go-review.googlesource.com/29712
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
For -buildmode=plugin, this lets the linker drop the main.main symbol
out of the binary while including most of the runtime.
(In the future it should be possible to drop the entire runtime
package from plugins.)
Change-Id: I3e7a024ddf5cc945e3d8b84bf37a0b7cb2a00eb6
Reviewed-on: https://go-review.googlesource.com/27821
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
noescape is now 0 instructions with the SSA backend.
fast atomics are no longer a TODO (at least for amd64).
Change-Id: Ib6e06f7471bef282a47ba236d8ce95404bb60a42
Reviewed-on: https://go-review.googlesource.com/28087
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
When compiling with -buildmode=shared, a map[int32]*_type is created for
each extra module mapping duplicate types back to a canonical object.
This is done in the function typelinksinit, which is called before the
init function that sets up the hash functions for the map
implementation. The result is typemap becomes unusable after
runtime initialization.
The fix in this CL is to move algorithm init before typelinksinit in
the runtime setup process. (For 1.8, we may want to turn typemap into
a sorted slice of types and use binary search.)
Manually tested on GOOS=linux with:
GOHOSTARCH=386 GOARCH=386 ./make.bash && \
go install -buildmode=shared std && \
cd ../test && \
go run run.go -linkshared
Fixes#16590
Change-Id: Idc08c50cc70d20028276fbf564509d2cd5405210
Reviewed-on: https://go-review.googlesource.com/25469
Run-TryBot: David Crawshaw <crawshaw@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
When a non-Go thread calls into Go, the runtime needs an M to run the Go
code. The runtime keeps a list of extra M's available. When the last
extra M is allocated, the needextram field is set to tell it to allocate
a new extra M as soon as it is running in Go. This ensures that an extra
M will always be available for the next thread.
However, if many threads need an extra M at the same time, this
serializes them all. One thread will get an extra M with the needextram
field set. All the other threads will see that there is no M available
and will go to sleep. The one thread that succeeded will create a new
extra M. One lucky thread will get it. All the other threads will see
that there is no M available and will go to sleep. The effect is
thundering herd, as all the threads looking for an extra M go through
the process one by one. This seems to have a particularly bad effect on
the FreeBSD scheduler for some reason.
With this change, we track the number of threads waiting for an M, and
create all of them as soon as one thread gets through. This still means
that all the threads will fight for the lock to pick up the next M. But
at least each thread that gets the lock will succeed, instead of going
to sleep only to fight again.
This smooths out the performance greatly on FreeBSD, reducing the
average wall time of `testprogcgo CgoCallbackGC` by 74%. On GNU/Linux
the average wall time goes down by 9%.
Fixes#13926Fixes#16396
Change-Id: I6dc42a4156085a7ed4e5334c60b39db8f8ef8fea
Reviewed-on: https://go-review.googlesource.com/25047
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
When the blocked field was first introduced back in
https://golang.org/cl/61250043 the scheduler trace code incorrectly used
m->blocked instead of mp->blocked. That has carried through the
conversion to Go. This CL fixes it.
Change-Id: Id81907b625221895aa5c85b9853f7c185efd8f4b
Reviewed-on: https://go-review.googlesource.com/24571
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
This has a minor performance cost, but far less than is being gained by SSA.
As an experiment, enable it during the Go 1.7 beta.
Having frame pointers on by default makes Linux's perf, Intel VTune,
and other profilers much more useful, because it lets them gather a
stack trace efficiently on profiling events.
(It doesn't help us that much, since when we walk the stack we usually
need to look up PC-specific information as well.)
Fixes#15840.
Change-Id: I4efd38412a0de4a9c87b1b6e5d11c301e63f1a2a
Reviewed-on: https://go-review.googlesource.com/23451
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Acquire and release the TSAN synchronization point when calling malloc,
just as we do when calling any other C function. If we don't do this,
TSAN will report false positive errors about races calling malloc and
free.
We used to have a special code path for malloc and free, going through
the runtime functions cmalloc and cfree. The special code path for cfree
was no longer used even before this CL. This CL stops using the special
code path for malloc, because there is no place along that path where we
could conditionally insert the TSAN synchronization. This CL removes
the support for the special code path for both functions.
Instead, cgo now automatically generates the malloc function as though
it were referenced as C.malloc. We need to automatically generate it
even if C.malloc is not called, even if malloc and size_t are not
declared, to support cgo-provided functions like C.CString.
Change-Id: I829854ec0787a80f33fa0a8a0dc2ee1d617830e2
Reviewed-on: https://go-review.googlesource.com/23260
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Currently scanstack obtains its own gcWork from the P for the duration
of the stack scan and then, if called during mark termination,
disposes the gcWork.
However, this means that the number of workbufs allocated will be at
least the number of stacks scanned during mark termination, which may
be very high (especially during a STW GC). This happens because, in
steady state, each scanstack will obtain a fresh workbuf (either from
the empty list or by allocating it), fill it with the scan results,
and then dispose it to the full list. Nothing is consuming from the
full list during this (and hence nothing is recycling them to the
empty list), so the length of the full list by the time mark
termination starts draining it is at least the number of stacks
scanned.
Fix this by pushing the gcWork acquisition up the stack to either the
gcDrain that calls markroot that calls scanstack (which batches across
many stack scans and is the path taken during STW GC) or to newstack
(which is still a single scanstack call, but this is roughly bounded
by the number of Ps).
This fix reduces the workbuf allocation for the test program from
issue #15319 from 213 MB (roughly 2KB * 1e5 goroutines) to 10 MB.
Fixes#15319.
Note that there's potentially a similar issue in write barriers during
mark 2. Fixing that will be more difficult since there's no broader
non-preemptible context, but it should also be less of a problem since
the full list is being drained during mark 2.
Some overall improvements in the go1 benchmarks, plus the usual noise.
No significant change in the garbage benchmark (time/op or GC memory).
name old time/op new time/op delta
BinaryTree17-12 2.54s ± 1% 2.51s ± 1% -1.09% (p=0.000 n=20+19)
Fannkuch11-12 2.12s ± 0% 2.17s ± 0% +2.18% (p=0.000 n=19+18)
FmtFprintfEmpty-12 45.1ns ± 1% 45.2ns ± 0% ~ (p=0.078 n=19+18)
FmtFprintfString-12 127ns ± 0% 128ns ± 0% +1.08% (p=0.000 n=19+16)
FmtFprintfInt-12 125ns ± 0% 122ns ± 1% -2.71% (p=0.000 n=14+18)
FmtFprintfIntInt-12 196ns ± 0% 190ns ± 1% -2.91% (p=0.000 n=12+20)
FmtFprintfPrefixedInt-12 196ns ± 0% 194ns ± 1% -0.94% (p=0.000 n=13+18)
FmtFprintfFloat-12 253ns ± 1% 251ns ± 1% -0.86% (p=0.000 n=19+20)
FmtManyArgs-12 807ns ± 1% 784ns ± 1% -2.85% (p=0.000 n=20+20)
GobDecode-12 7.13ms ± 1% 7.12ms ± 1% ~ (p=0.351 n=19+20)
GobEncode-12 5.89ms ± 0% 5.95ms ± 0% +0.94% (p=0.000 n=19+19)
Gzip-12 219ms ± 1% 221ms ± 1% +1.35% (p=0.000 n=18+20)
Gunzip-12 37.5ms ± 1% 37.4ms ± 0% ~ (p=0.057 n=20+19)
HTTPClientServer-12 81.4µs ± 4% 81.9µs ± 3% ~ (p=0.118 n=17+18)
JSONEncode-12 15.7ms ± 1% 15.8ms ± 1% +0.73% (p=0.000 n=17+18)
JSONDecode-12 57.9ms ± 1% 57.2ms ± 1% -1.34% (p=0.000 n=19+19)
Mandelbrot200-12 4.12ms ± 1% 4.10ms ± 0% -0.33% (p=0.000 n=19+17)
GoParse-12 3.22ms ± 2% 3.25ms ± 1% +0.72% (p=0.000 n=18+20)
RegexpMatchEasy0_32-12 70.6ns ± 1% 71.1ns ± 2% +0.63% (p=0.005 n=19+20)
RegexpMatchEasy0_1K-12 240ns ± 0% 239ns ± 1% -0.59% (p=0.000 n=19+20)
RegexpMatchEasy1_32-12 71.3ns ± 1% 71.3ns ± 1% ~ (p=0.844 n=17+17)
RegexpMatchEasy1_1K-12 384ns ± 2% 371ns ± 1% -3.45% (p=0.000 n=19+20)
RegexpMatchMedium_32-12 109ns ± 1% 108ns ± 2% -0.48% (p=0.029 n=19+19)
RegexpMatchMedium_1K-12 34.3µs ± 1% 34.5µs ± 2% ~ (p=0.160 n=18+20)
RegexpMatchHard_32-12 1.79µs ± 9% 1.72µs ± 2% -3.83% (p=0.000 n=19+19)
RegexpMatchHard_1K-12 53.3µs ± 4% 51.8µs ± 1% -2.82% (p=0.000 n=19+20)
Revcomp-12 386ms ± 0% 388ms ± 0% +0.72% (p=0.000 n=17+20)
Template-12 62.9ms ± 1% 62.5ms ± 1% -0.57% (p=0.010 n=18+19)
TimeParse-12 325ns ± 0% 331ns ± 0% +1.84% (p=0.000 n=18+19)
TimeFormat-12 338ns ± 0% 343ns ± 0% +1.34% (p=0.000 n=18+20)
[Geo mean] 52.7µs 52.5µs -0.42%
Change-Id: Ib2d34736c4ae2ec329605b0fbc44636038d8d018
Reviewed-on: https://go-review.googlesource.com/23391
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently ready always puts the readied goroutine in runnext. We're
going to have to change this for some uses, so add a flag for whether
or not to use runnext.
For now we always pass true so this is a no-op change.
For #15706.
Change-Id: Iaa66d8355ccfe4bbe347570cc1b1878c70fa25df
Reviewed-on: https://go-review.googlesource.com/23171
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
The test case in #15639 somehow causes an invalid syscall frame. The
failure is obscured because the throw occurs when throwsplit == true,
which causes a "stack split at bad time" error when trying to print the
throw message.
This CL fixes the "stack split at bad time" by using systemstack. No
test because there shouldn't be any way to trigger this error anyhow.
Update #15639.
Change-Id: I4240f3fd01bdc3c112f3ffd1316b68504222d9e1
Reviewed-on: https://go-review.googlesource.com/23153
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
If we collected a cgo traceback when entering the SIGPROF signal
handler, record it as part of the profiling stack trace.
This serves as the promised test for https://golang.org/cl/21055 .
Change-Id: I5f60cd6cea1d9b7c3932211483a6bfab60ed21d2
Reviewed-on: https://go-review.googlesource.com/22650
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Race runtime also needs local malloc caches and currently uses
a mix of per-OS-thread and per-goroutine caches. This leads to
increased memory consumption. But more importantly cache of
synchronization objects is per-goroutine and we don't always
have goroutine context when feeing memory in GC. As the result
synchronization object descriptors leak (more precisely, they
can be reused if another synchronization object is recreated
at the same address, but it does not always help). For example,
the added BenchmarkSyncLeak has effectively runaway memory
consumption (based on a real long running server).
This change updates race runtime with support for per-P contexts.
BenchmarkSyncLeak now stabilizes at ~1GB memory consumption.
Long term, this will allow us to remove race runtime dependency
on glibc (as malloc is the main cornerstone).
I've also implemented a different scheme to pass P context to
race runtime: scheduler notified race runtime about association
between G and P by calling procwire(g, p)/procunwire(g, p).
But it turned out to be very messy as we have lots of places
where the association changes (e.g. syscalls). So I dropped it
in favor of the current scheme: race runtime asks scheduler
about the current P.
Fixes#14533
Change-Id: Iad10d2f816a44affae1b9fed446b3580eafd8c69
Reviewed-on: https://go-review.googlesource.com/19970
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Runqempty is a critical predicate for scheduler. If runqempty spuriously
returns true, then scheduler can fail to schedule arbitrary number of
runnable goroutines on idle Ps for arbitrary long time. With the addition
of runnext runqempty predicate become broken (can spuriously return true).
Consider that runnext is not nil and the main array is empty. Runqempty
observes that the array is empty, then it is descheduled for some time.
Then queue owner pushes another element to the queue evicting runnext
into the array. Then queue owner pops runnext. Then runqempty resumes
and observes runnext is nil and returns true. But there were no point
in time when the queue was empty.
Fix runqempty predicate to not return true spuriously.
Change-Id: Ifb7d75a699101f3ff753c4ce7c983cf08befd31e
Reviewed-on: https://go-review.googlesource.com/20858
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently we clear gcscanvalid in both casgstatus and
casfrom_Gscanstatus if the new status is _Grunning. This is very
important to do in casgstatus. However, this is potentially wrong in
casfrom_Gscanstatus because in this case the caller doesn't own gp and
hence the write is racy. Unlike the other _Gscan statuses, during
_Gscanrunning, the G is still running. This does not indicate that
it's transitioning into a running state. The scan simply hasn't
happened yet, so it's neither valid nor invalid.
Conveniently, this also means clearing gcscanvalid is unnecessary in
this case because the G was already in _Grunning, so we can simply
remove this code. What will happen instead is that the G will be
preempted to scan itself, that scan will set gcscanvalid to true, and
then the G will return to _Grunning via casgstatus, clearing
gcscanvalid.
This fix will become necessary shortly when we start keeping track of
the set of G's with dirty stacks, since it will no longer be
idempotent to simply set gcscanvalid to false.
Change-Id: I688c82e6fbf00d5dbbbff49efa66acb99ee86785
Reviewed-on: https://go-review.googlesource.com/20669
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently all free Gs are added to one list. Split this into two
lists: one for free Gs with cached stacks and one for Gs without
cached stacks.
This lets us preferentially allocate Gs that already have a stack, but
more importantly, it sets us up to free cached G stacks concurrently.
Change-Id: Idbe486f708997e1c9d166662995283f02d1eeb3c
Reviewed-on: https://go-review.googlesource.com/20664
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently tracer uses global sequencer and it introduces
significant slowdown on parallel machines (up to 10x).
Replace the global sequencer with per-goroutine sequencer.
If we assign per-goroutine sequence numbers to only 3 types
of events (start, unblock and syscall exit), it is enough to
restore consistent partial ordering of all events. Even these
events don't need sequence numbers all the time (if goroutine
starts on the same P where it was unblocked, then start does
not need sequence number).
The burden of restoring the order is put on trace parser.
Details of the algorithm are described in the comments.
On http benchmark with GOMAXPROCS=48:
no tracing: 5026 ns/op
tracing: 27803 ns/op (+453%)
with this change: 6369 ns/op (+26%, mostly for traceback)
Also trace size is reduced by ~22%. Average event size before: 4.63
bytes/event, after: 3.62 bytes/event.
Besides running trace tests, I've also tested with manually broken
cputicks (random skew for each event, per-P skew and episodic random skew).
In all cases broken timestamps were detected and no test failures.
Change-Id: I078bde421ccc386a66f6c2051ab207bcd5613efa
Reviewed-on: https://go-review.googlesource.com/21512
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This CL introduces the typeOff type and a lookup method of the same
name that can turn a typeOff offset into an *rtype.
In a typical Go binary (built with buildmode=exe, pie, c-archive, or
c-shared), there is one moduledata and all typeOff values are offsets
relative to firstmoduledata.types. This makes computing the pointer
cheap in typical programs.
With buildmode=shared (and one day, buildmode=plugin) there are
multiple modules whose relative offset is determined at runtime.
We identify a type in the general case by the pair of the original
*rtype that references it and its typeOff value. We determine
the module from the original pointer, and then use the typeOff from
there to compute the final *rtype.
To ensure there is only one *rtype representing each type, the
runtime initializes a typemap for each module, using any identical
type from an earlier module when resolving that offset. This means
that types computed from an offset match the type mapped by the
pointer dynamic relocations.
A series of followup CLs will replace other *rtype values with typeOff
(and name/*string with nameOff).
For types created at runtime by reflect, type offsets are treated as
global IDs and reference into a reflect offset map kept by the runtime.
darwin/amd64:
cmd/go: -57KB (0.6%)
jujud: -557KB (0.8%)
linux/amd64 PIE:
cmd/go: -361KB (3.0%)
jujud: -3.5MB (4.2%)
For #6853.
Change-Id: Icf096fd884a0a0cb9f280f46f7a26c70a9006c96
Reviewed-on: https://go-review.googlesource.com/21285
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
So that all Go processes do not die on startup on a system with >256 CPUs.
I tested this by hacking osinit to set ncpu to 1000.
Updates #15131
Change-Id: I52e061a0de97be41d684dd8b748fa9087d6f1aef
Reviewed-on: https://go-review.googlesource.com/21599
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Usleep(100) in runqgrab negatively affects latency and throughput
of parallel application. We are sleeping instead of doing useful work.
This is effect is particularly visible on windows where minimal
sleep duration is 1-15ms.
Reduce sleep from 100us to 3us and use osyield on windows.
Sync chan send/recv takes ~50ns, so 3us gives us ~50x overshoot.
benchmark old ns/op new ns/op delta
BenchmarkChanSync-12 216 217 +0.46%
BenchmarkChanSyncWork-12 27213 25816 -5.13%
CPU consumption goes up from 106% to 108% in the first case,
and from 107% to 125% in the second case.
Test case from #14790 on windows:
BenchmarkDefaultResolution-8 4583372 29720 -99.35%
Benchmark1ms-8 992056 30701 -96.91%
99-th latency percentile for HTTP request serving is improved by up to 15%
(see http://golang.org/cl/20835 for details).
The following benchmarks are from the change that originally added this sleep
(see https://golang.org/s/go15gomaxprocs):
name old time/op new time/op delta
Chain 22.6µs ± 2% 22.7µs ± 6% ~ (p=0.905 n=9+10)
ChainBuf 22.4µs ± 3% 22.5µs ± 4% ~ (p=0.780 n=9+10)
Chain-2 23.5µs ± 4% 24.9µs ± 1% +5.66% (p=0.000 n=10+9)
ChainBuf-2 23.7µs ± 1% 24.4µs ± 1% +3.31% (p=0.000 n=9+10)
Chain-4 24.2µs ± 2% 25.1µs ± 3% +3.70% (p=0.000 n=9+10)
ChainBuf-4 24.4µs ± 5% 25.0µs ± 2% +2.37% (p=0.023 n=10+10)
Powser 2.37s ± 1% 2.37s ± 1% ~ (p=0.423 n=8+9)
Powser-2 2.48s ± 2% 2.57s ± 2% +3.74% (p=0.000 n=10+9)
Powser-4 2.66s ± 1% 2.75s ± 1% +3.40% (p=0.000 n=10+10)
Sieve 13.3s ± 2% 13.3s ± 2% ~ (p=1.000 n=10+9)
Sieve-2 7.00s ± 2% 7.44s ±16% ~ (p=0.408 n=8+10)
Sieve-4 4.13s ±21% 3.85s ±22% ~ (p=0.113 n=9+9)
Fixes#14790
Change-Id: Ie7c6a1c4f9c8eb2f5d65ab127a3845386d6f8b5d
Reviewed-on: https://go-review.googlesource.com/20835
Reviewed-by: Austin Clements <austin@google.com>
See #14874
This change makes the runtime register all compiler generated itabs
(as obtained from the moduledata) during init.
Change-Id: I9969a0985b99b8bda820a631f7fe4c78f1174cdf
Reviewed-on: https://go-review.googlesource.com/20900
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Michel Lespinasse <walken@google.com>
During random stealing we steal 4*GOMAXPROCS times from random procs.
One would expect that most of the time we check all procs this way,
but due to low quality PRNG we actually miss procs with frightening
probability. Below are modelling experiment results for 1e6 tries:
GOMAXPROCS = 2 : missed 1 procs 7944 times
GOMAXPROCS = 3 : missed 1 procs 101620 times
GOMAXPROCS = 3 : missed 2 procs 3571 times
GOMAXPROCS = 4 : missed 1 procs 63916 times
GOMAXPROCS = 4 : missed 2 procs 61 times
GOMAXPROCS = 4 : missed 3 procs 16 times
GOMAXPROCS = 5 : missed 1 procs 133136 times
GOMAXPROCS = 5 : missed 2 procs 1025 times
GOMAXPROCS = 5 : missed 3 procs 101 times
GOMAXPROCS = 5 : missed 4 procs 15 times
GOMAXPROCS = 8 : missed 1 procs 151765 times
GOMAXPROCS = 8 : missed 2 procs 5057 times
GOMAXPROCS = 8 : missed 3 procs 1726 times
GOMAXPROCS = 8 : missed 4 procs 68 times
GOMAXPROCS = 12 : missed 1 procs 199081 times
GOMAXPROCS = 12 : missed 2 procs 27489 times
GOMAXPROCS = 12 : missed 3 procs 3113 times
GOMAXPROCS = 12 : missed 4 procs 233 times
GOMAXPROCS = 12 : missed 5 procs 9 times
GOMAXPROCS = 16 : missed 1 procs 237477 times
GOMAXPROCS = 16 : missed 2 procs 30037 times
GOMAXPROCS = 16 : missed 3 procs 9466 times
GOMAXPROCS = 16 : missed 4 procs 1334 times
GOMAXPROCS = 16 : missed 5 procs 192 times
GOMAXPROCS = 16 : missed 6 procs 5 times
GOMAXPROCS = 16 : missed 7 procs 1 times
GOMAXPROCS = 16 : missed 8 procs 1 times
A missed proc won't lead to underutilization because we check all procs
again after dropping P. But it can lead to an unpleasant situation
when we miss a proc, drop P, check all procs, discover work, acquire P,
miss the proc again, repeat.
Improve stealing logic to cover all procs.
Also don't enter spinning mode and try to steal when there is nobody around.
Change-Id: Ibb6b122cc7fb836991bad7d0639b77c807aab4c2
Reviewed-on: https://go-review.googlesource.com/20836
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Marvin Stenger <marvin.stenger94@gmail.com>
gopark calls the unlock function after setting the G to _Gwaiting.
This means it's generally unsafe to access the G's stack from the
unlock function because the G may start running on another P. Once we
start shrinking stacks concurrently, a stack shrink could also move
the stack the moment after it enters _Gwaiting and before the unlock
function is called.
Document this restriction and fix the two places where we currently
violate it.
This is unlikely to be a problem in practice for these two places
right now, but they're already skating on thin ice. For example, the
following sequence could in principle cause corruption, deadlock, or a
panic in the select code:
On M1/P1:
1. G1 selects on channels A and B.
2. selectgoImpl calls gopark.
3. gopark puts G1 in _Gwaiting.
4. gopark calls selparkcommit.
5. selparkcommit releases the lock on channel A.
On M2/P2:
6. G2 sends to channel A.
7. The send puts G1 in _Grunnable and puts it on P2's run queue.
8. The scheduler runs, selects G1, puts it in _Grunning, and resumes G1.
9. On G1, the sellock immediately following the gopark gets called.
10. sellock grows and moves the stack.
On M1/P1:
11. selparkcommit continues to scan the lock order for the next
channel to unlock, but it's now reading from a freed (and possibly
reused) stack.
This shouldn't happen in practice because step 10 isn't the first call
to sellock, so the stack should already be big enough. However, once
we start shrinking stacks concurrently, this reasoning won't work any
more.
For #12967.
Change-Id: I3660c5be37e5be9f87433cb8141bdfdf37fadc4c
Reviewed-on: https://go-review.googlesource.com/20038
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Given a G, there's currently no way to find the channel it's blocking
on. We'll need this information to fix a (probably theoretical) bug in
select and to implement concurrent stack shrinking, so record the
channel in the sudog.
For #12967.
Change-Id: If8fb63a140f1d07175818824d08c0ebeec2bdf66
Reviewed-on: https://go-review.googlesource.com/20035
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
