This converts several unsynchronized reads (reads without holding
prof.signalLock) into atomic reads.
For #53821.
For #52912.
Change-Id: I421b96a22fbe26d699bcc21010c8a9e0f4efc276
Reviewed-on: https://go-review.googlesource.com/c/go/+/420196
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
I've dropped the note that sched.timeToRun is protected by sched.lock,
as it does not seem to be true.
For #53821.
Change-Id: I03f8dc6ca0bcd4ccf3ec113010a0aa39c6f7d6ef
Reviewed-on: https://go-review.googlesource.com/c/go/+/419449
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Pratt <mpratt@google.com>
This converts a few unsynchronized accesses.
For #53821.
Change-Id: Ie2728779111e3e042696f15648981c5d5a86ca6d
Reviewed-on: https://go-review.googlesource.com/c/go/+/419448
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Note that this replaces numerous unsynchronized loads throughout the
scheduler.
For #53821.
Change-Id: Ica80b04c9e8c184bfef186e549526fc3f117c387
Reviewed-on: https://go-review.googlesource.com/c/go/+/419447
Run-TryBot: Michael Pratt <mpratt@google.com>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Note that this converts nmspinning from uint32 to int32 for consistency
with the other count fields in schedt.
For #53821.
Change-Id: Ia6ca7a2b476128eda3b68e9f0c7775ae66c0c744
Reviewed-on: https://go-review.googlesource.com/c/go/+/419446
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Pratt <mpratt@google.com>
Reviewed-by: Austin Clements <austin@google.com>
Note that this converts npidle from uint32 to int32 for consistency with
the other count fields in schedt and the type of p.id.
Note that this changes previously unsynchronized operations to
synchronized operations in:
* handoffp
* injectglist
* schedtrace
* schedEnableUser
* sync_runtime_canSpin
For #53821.
Change-Id: I36d1b3b4a28131c9d47884fade6bc44439dd6937
Reviewed-on: https://go-review.googlesource.com/c/go/+/419445
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
Note that this converts ngsys from uint32 to int32 to match the other
(non-atomic) counters.
For #53821.
Change-Id: I3acbfbbd1dabc59b0ea5ddc86a97e0d0afa9f80c
Reviewed-on: https://go-review.googlesource.com/c/go/+/419444
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Note that this converts pollUntil from uint64 to int64, the type used by
nanotime().
For #53821.
Change-Id: Iec9ec7e09d3350552561d0708ba6ea9e8a8ae7ab
Reviewed-on: https://go-review.googlesource.com/c/go/+/419443
Run-TryBot: Michael Pratt <mpratt@google.com>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Note that this changes the type from uint64 to int64, the type used by
nanotime(). It also adds an atomic load in pollWork(), which used to use
a non-atomic load.
For #53821.
Change-Id: I6173c90f20bfdc0e0a4bc3a7b1c798d1c429fff5
Reviewed-on: https://go-review.googlesource.com/c/go/+/419442
Run-TryBot: Michael Pratt <mpratt@google.com>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
We currently print these as -1, but some are technically uint64. We can
be more explicit about their irrelevance by printing 'nil' rather than
-1.
Change-Id: I267fd8830564c75032bfe9176af59047f5a90202
Reviewed-on: https://go-review.googlesource.com/c/go/+/419441
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Pratt <mpratt@google.com>
schedt.goidgen and p.goidcache are already uint64, this makes all cases
consistent.
The only oddball here is schedtrace which prints -1 as an equivalent for
N/A or nil. A future CL will make this more explicit.
Change-Id: I489626f3232799f6ca333d0d103b71d9d3aa7494
Reviewed-on: https://go-review.googlesource.com/c/go/+/419440
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
We're about to require that all uses of trace.lock be on the system
stack. That's mostly easy, except that it's involving parking the
trace reader. Fix this by changing that parking protocol so it instead
synchronizes through an atomic.
For #53979.
Change-Id: Icd6db8678dd01094029d7ad1c612029f571b4cbb
Reviewed-on: https://go-review.googlesource.com/c/go/+/418955
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Assumingly this appears to have been a bug in c2go, as this appeared in
https://go.dev/cl/172260043.
Change-Id: I8477ec226ffb31268b9c479af13b0fcf401dbeec
Reviewed-on: https://go-review.googlesource.com/c/go/+/418276
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
This manually replaces uses of _g_ that cannot be trivially switched to
gp since there is another gp variable in scope.
Most of these functions only use the current g to reach the m, so this
helps with clarity by switching all accesses directly to an mp variable.
Change-Id: I96a4fc1c32470a7f3d12ddec9f147c2743210e71
Reviewed-on: https://go-review.googlesource.com/c/go/+/418577
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
_g_, _p_, and _m_ are primarily vestiges of the C version of the
runtime, while today we prefer Go-style variable names (generally gp,
pp, and mp).
This change replaces all remaining uses of _p_ with pp. These are all
trivial replacements (i.e., no conflicts). That said, there are several
functions that refer to two different Ps at once. There the naming
convention is generally that pp refers to the local P, and p2 refers to
the other P we are accessing.
Change-Id: I205b801be839216972e7644b1fbeacdbf2612859
Reviewed-on: https://go-review.googlesource.com/c/go/+/306674
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
In faketime mode, checkdead is responsible for jumping time forward to
the next timer expiration, and waking an M to handle the newly ready
timer.
Currently it pulls the exact P that owns the next timer off of the pidle
list. In theory this is efficient because that P is immediately eligible
to run the timer without stealing. Unfortunately it is also fraught with
peril because we are skipping all of the bookkeeping in pidleget:
* Skipped updates to timerpMask mean that our timers may not be eligible
for stealing, as they should be.
* Skipped updates to idlepMask mean that our runq may not be eligible
for stealing, as they should be.
* Skipped updates to sched.npidle may break tracking of spinning Ms,
potentially resulting in lost work.
* Finally, as of CL 410122, skipped updates to p.limiterEvent may affect
the GC limiter, or cause a fatal throw when another event occurs.
The last case has finally undercovered this issue since it quickly
results in a hard crash.
We could add all of these updates into checkdead, but it is much more
maintainable to keep this logic in one place and use pidleget here like
everywhere else in the runtime. This means we probably won't wake the
P owning the timer, meaning that the P will need to steal the timer,
which is less efficient, but faketime is not a performance-sensitive
build mode. Note that the M will automatically make itself a spinning M
to make it eligible to steal since it is the only one running.
Fixes#53294
For #52890
Change-Id: I4acc3d259b9b4d7dc02608581c8b4fd259f272e9
Reviewed-on: https://go-review.googlesource.com/c/go/+/411119
Run-TryBot: Michael Pratt <mpratt@google.com>
Auto-Submit: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Currently the GC CPU limiter only tracks idle GC work time. However, in
very undersubscribed situations, it's possible that all this extra idle
time prevents the enabling of the limiter, since it all gets account for
as mutator time. Fix this by tracking all idle time via pidleget and
pidleput. To support this, pidleget and pidleput also accept and return
"now" parameters like the timer code.
While we're here, let's clean up some incorrect assumptions that some of
the scheduling code makes about "now."
Fixes#52890.
Change-Id: I4a97893d2e5ad1e8c821f8773c2a1d449267c951
Reviewed-on: https://go-review.googlesource.com/c/go/+/410122
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
NetBSD appears to have the same issue OpenBSD had in runqgrab. See
issue #52475 for more details.
For #35166.
Change-Id: Ie53192d26919b4717bc0d61cadd88d688ff38bb4
Reviewed-on: https://go-review.googlesource.com/c/go/+/407139
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Measure the average stack size used by goroutines at every GC. When
starting a new goroutine, allocate an initial goroutine stack of that
average size. Intuition is that we'll waste at most 2x in stack space
because only half the goroutines can be below average. In turn, we
avoid some of the early stack growth / copying needed in the average
case.
More details in the design doc at: https://docs.google.com/document/d/1YDlGIdVTPnmUiTAavlZxBI1d9pwGQgZT7IKFKlIXohQ/edit?usp=sharing
name old time/op new time/op delta
Issue18138 95.3µs ± 0% 67.3µs ±13% -29.35% (p=0.000 n=9+10)
Fixes#18138
Change-Id: Iba34d22ed04279da7e718bbd569bbf2734922eaa
Reviewed-on: https://go-review.googlesource.com/c/go/+/345889
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
The CPU profiler adds goroutine labels to its samples based on
getg().m.curg. That allows the profile to correctly attribute work that
the runtime does on behalf of that goroutine on the M's g0 stack via
systemstack calls, such as using runtime.Callers to record the call
stack.
Those labels also cover work on the g0 stack via mcall. When the active
goroutine calls runtime.Gosched, it will receive attribution of its
share of the scheduler work necessary to find the next runnable
goroutine.
The execution tracer's attribution of CPU samples to specific goroutines
should match. When curg is set, attribute the CPU samples to that
goroutine's ID.
Fixes#52693
Change-Id: Ic9af92e153abd8477559e48bc8ebaf3739527b94
Reviewed-on: https://go-review.googlesource.com/c/go/+/404055
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Bryan Mills <bcmills@google.com>
Run-TryBot: Rhys Hiltner <rhys@justin.tv>
TryBot-Result: Gopher Robot <gobot@golang.org>
When the CPU profiler and execution tracer are both active, report the
CPU profile samples in the execution trace data stream.
Include only samples that arrive on the threads known to the runtime,
but include them even when running g0 (such as near the scheduler) or if
there's no P (such as near syscalls).
Render them in "go tool trace" as instantaneous events.
For #16895
Change-Id: I0aa501a7b450c971e510961c0290838729033f7f
Reviewed-on: https://go-review.googlesource.com/c/go/+/400795
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Rhys Hiltner <rhys@justin.tv>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
The race annotations for goroutine label maps covered the special type
of read necessary to create CPU profiles. Extend that to include
goroutine profiles. Annotate the copy involved in creating new
goroutines.
Fixes#50292
Change-Id: I10f69314e4f4eba85c506590fe4781f4d6b8ec2d
Reviewed-on: https://go-review.googlesource.com/c/go/+/385660
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Austin Clements <austin@google.com>
This change adds a GC CPU utilization limiter to the GC. It disables
assists to ensure GC CPU utilization remains under 50%. It uses a leaky
bucket mechanism that will only fill if GC CPU utilization exceeds 50%.
Once the bucket begins to overflow, GC assists are limited until the
bucket empties, at the risk of GC overshoot. The limiter is primarily
updated by assists. The scheduler may also update it, but only if the
GC is on and a few milliseconds have passed since the last update. This
second case exists to ensure that if the limiter is on, and no assists
are happening, we're still updating the limiter regularly.
The purpose of this limiter is to mitigate GC death spirals, opting to
use more memory instead.
This change turns the limiter on always. In practice, 50% overall GC CPU
utilization is very difficult to hit unless you're trying; even the most
allocation-heavy applications with complex heaps still need to do
something with that memory. Note that small GOGC values (i.e.
single-digit, or low teens) are more likely to trigger the limiter,
which means the GOGC tradeoff may no longer be respected. Even so, it
should still be relatively rare.
This change also introduces the feature flag for code to support the
memory limit feature.
For #48409.
Change-Id: Ia30f914e683e491a00900fd27868446c65e5d3c2
Reviewed-on: https://go-review.googlesource.com/c/go/+/353989
Reviewed-by: Michael Pratt <mpratt@google.com>
This gives explicit names to the possible states of throwing (-1, 0, 1).
m.throwing is now one of:
throwTypeOff: not throwing, previously == 0
throwTypeUser: user throw, previously == -1
throwTypeRuntime: runtime throw, previously == 1
For runtime throws, we now always include frame metadata and system
goroutines regardless of GOTRACEBACK to aid in debugging the runtime.
For user throws, we no longer include frame metadata or runtime frames,
unless GOTRACEBACK=system or higher.
For #51485.
Change-Id: If252e2377a0b6385ce7756b937929be4273a56c0
Reviewed-on: https://go-review.googlesource.com/c/go/+/390421
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Austin Clements <austin@google.com>
"User" throws are throws due to some invariant broken by the application.
"System" throws are due to some invariant broken by the runtime,
environment, etc (i.e., not the fault of the application).
This CL sends "user" throws through the new fatal. Currently this
function is identical to throw, but with a different name to clearly
differentiate the throw type in the stack trace, and hopefully be a bit
more clear to users what it means.
This CL changes a few categories of throw to fatal:
1. Concurrent map read/write.
2. Deadlock detection.
3. Unlock of unlocked sync.Mutex.
4. Inconsistent results from syscall.AllThreadsSyscall.
"Thread exhaustion" and "out of memory" (usually address space full)
throws are additional throws that are arguably the fault of user code,
but I've left off for now because there is no specific invariant that
they have broken to get into these states.
For #51485
Change-Id: I713276a6c290fd34a6563e6e9ef378669d74ae32
Reviewed-on: https://go-review.googlesource.com/c/go/+/390420
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
This change refactors the scavenger into a type whose methods represent
the actual function and scheduling of the scavenger. It also stubs out
access to global state in order to make it testable.
This change thus also adds a test for the scavenger. In writing this
test, I discovered the lack of a behavior I expected: if the
pageAlloc.scavenge returns < the bytes requested scavenged, that means
the heap is exhausted. This has been true this whole time, but was not
documented or explicitly relied upon. This change rectifies that. In
theory this means the scavenger could spin in run() indefinitely (as
happened in the test) if shouldStop never told it to stop. In practice,
shouldStop fires long before the heap is exhausted, but for future
changes it may be important. At the very least it's good to be
intentional about these things.
While we're here, I also moved the call to stopTimer out of wake and
into sleep. There's no reason to add more operations to a context that's
already precarious (running without a P on sysmon).
Change-Id: Ib31b86379fd9df84f25ae282734437afc540da5c
Reviewed-on: https://go-review.googlesource.com/c/go/+/384734
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This change moves several scheduling decisions made by schedule into
findrunnable. The main motivation behind this change is the fact that
stopped Ms can't become dedicated or fractional GC workers. The main
reason for this is that when a stopped M wakes up, it stays in
findrunnable until it finds work, which means it will never consider GC
work. On that note, it'll also never consider becoming the trace reader,
either.
Another way of looking at it is that this change tries to make
findrunnable aware of more sources of work than it was before. With this
change, any M in findrunnable should be capable of becoming a GC worker,
resolving #44313. While we're here, let's also make more sources of
work, such as the trace reader, visible to handoffp, which should really
be checking all sources of work. With that, we also now correctly handle
the case where StopTrace is called from the last live M that is also
locked (#39004). stoplockedm calls handoffp to start a new M and handle
the work it cannot, and once we include the trace reader in that, we
ensure that the trace reader gets scheduled.
This change attempts to preserve the exact same ordering of work
checking to reduce its impact.
One consequence of this change is that upon entering schedule, some
sources of work won't be checked twice (i.e. the local and global
runqs, and timers) as they do now, which in some sense gives them a
lower priority than they had before.
Fixes#39004.
Fixes#44313.
Change-Id: I5d8b7f63839db8d9a3e47cdda604baac1fe615ce
Reviewed-on: https://go-review.googlesource.com/c/go/+/393880
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This change reduces the maximum number of idle mark workers during
periodic (currently every 2 minutes) GC cycles to 1.
Idle mark workers soak up all available and unused Ps, up to GOMAXPROCS.
While this provides some throughput and latency benefit in general, it
can cause what appear to be massive CPU utilization spikes in otherwise
idle applications. This is mostly an issue for *very* idle applications,
ones idle enough to trigger periodic GC cycles. This spike also tends to
interact poorly with auto-scaling systems, as the system might assume
the load average is very low and suddenly see a massive burst in
activity.
The result of this change is not to bring down this 100% (of GOMAXPROCS)
CPU utilization spike to 0%, but rather
min(25% + 1/GOMAXPROCS*100%, 100%)
Idle mark workers also do incur a small latency penalty as they must be
descheduled for other work that might pop up. Luckily the runtime is
pretty good about getting idle mark workers off of Ps, so in general
the latency benefit from shorter GC cycles outweighs this cost. But, the
cost is still non-zero and may be more significant in idle applications
that aren't invoking assists and write barriers quite as often.
We can't completely eliminate idle mark workers because they're
currently necessary for GC progress in some circumstances. Namely,
they're critical for progress when all we have is fractional workers. If
a fractional worker meets its quota, and all user goroutines are blocked
directly or indirectly on a GC cycle (via runtime.GOMAXPROCS, or
runtime.GC), the program may deadlock without GC workers, since the
fractional worker will go to sleep with nothing to wake it.
Fixes#37116.
For #44163.
Change-Id: Ib74793bb6b88d1765c52d445831310b0d11ef423
Reviewed-on: https://go-review.googlesource.com/c/go/+/393394
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
OpenBSD has a coarse sleep granularity that rounds up to 10 ms
increments. This can cause significant STW delays, among other issues.
As far as I can tell, there's only 1 tightly timed sleep without an
explicit wakeup for which this actually matters.
Fixes#52475.
Change-Id: Ic69fc11096ddbbafd79b2dcdf3f912fde242db24
Reviewed-on: https://go-review.googlesource.com/c/go/+/401638
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This is a re-do of CL 388477, fixing #52472.
It is unsafe to call syscall.RawSyscall from syscall.Syscall with
-coverpkg=all and -race. This is because:
1. Coverage adds a sync/atomic call in RawSyscall to increment the
coverage counter.
2. Race mode instruments sync/atomic calls with TSAN runtime calls. TSAN
eventually calls runtime.racecallbackfunc, which expects
getg().m.p != 0, which is no longer true after entersyscall().
cmd/go actually avoids adding coverage instrumention to package runtime
in race mode entirely to avoid these kinds of problems. Rather than also
excluding all of syscall for this one function, work around by calling
RawSyscall6 instead, which avoids coverage instrumention both by being
written in assembly and in package runtime/*.
For #51087Fixes#52472
Change-Id: Iaffd27df03753020c4716059a455d6ca7b62f347
Reviewed-on: https://go-review.googlesource.com/c/go/+/401654
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Michael Pratt <mpratt@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
[This CL is part of a sequence implementing the proposal #51082.
The design doc is at https://go.dev/s/godocfmt-design.]
Run the updated gofmt, which reformats doc comments,
on the main repository. Vendored files are excluded.
For #51082.
Change-Id: I7332f099b60f716295fb34719c98c04eb1a85407
Reviewed-on: https://go-review.googlesource.com/c/go/+/384268
Reviewed-by: Jonathan Amsterdam <jba@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
A future change to gofmt will rewrite
// Doc comment.
//go:foo
to
// Doc comment.
//
//go:foo
Apply that change preemptively to all comments (not necessarily just doc comments).
For #51082.
Change-Id: Iffe0285418d1e79d34526af3520b415a12203ca9
Reviewed-on: https://go-review.googlesource.com/c/go/+/384260
Trust: Russ Cox <rsc@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
A run of lines that are indented with any number of spaces or tabs
format as a <pre> block. This commit fixes various doc comments
that format badly according to that (standard) rule.
For example, consider:
// - List item.
// Second line.
// - Another item.
Because the - lines are unindented, this is actually two paragraphs
separated by a one-line <pre> block. This CL rewrites it to:
// - List item.
// Second line.
// - Another item.
Today, that will format as a single <pre> block.
In a future release, we hope to format it as a bulleted list.
Various other minor fixes as well, all in preparation for reformatting.
For #51082.
Change-Id: I95cf06040d4186830e571cd50148be3bf8daf189
Reviewed-on: https://go-review.googlesource.com/c/go/+/384257
Trust: Russ Cox <rsc@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
For certain values of GOMAXPROCS, the current code is less random than
it looks. For example with GOMAXPROCS=12, there are 4 coprimes: 1 5 7 11.
That's bad, as 12 and 4 are not relatively prime. So if pos == 2, then we
always pick 7 as the inc. We want to pick pos and inc independently
at random.
Change-Id: I5c7e4f01f9223cbc2db12a685dc0bced2cf39abf
Reviewed-on: https://go-review.googlesource.com/c/go/+/369976
Run-TryBot: Keith Randall <khr@golang.org>
Trust: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
In issue 50113, we see that a thread blocked in a system call can result
in a hang of AllThreadsSyscall. To resolve this, we must send a signal
to these threads to knock them out of the system call long enough to run
the per-thread syscall.
Stepping back, if we need to send signals anyway, it should be possible
to implement this entire mechanism on top of signals. This CL does so,
vastly simplifying the mechanism, both as a direct result of
newly-unnecessary code as well as some ancillary simplifications to make
things simpler to follow.
Major changes:
* The rest of the mechanism is moved to os_linux.go, with fields in mOS
instead of m itself.
* 'Fixup' fields and functions are renamed to 'perThreadSyscall' so they
are more precise about their purpose.
* Rather than getting passed a closure, doAllThreadsSyscall takes the
syscall number and arguments. This avoids a lot of hairy behavior:
* The closure may potentially only be live in fields in the M,
hidden from the GC. Not necessary with no closure.
* The need to loan out the race context. A direct RawSyscall6 call
does not require any race context.
* The closure previously conditionally panicked in strange
locations, like a signal handler. Now we simply throw.
* All manual fixup synchronization with mPark, sysmon, templateThread,
sigqueue, etc is gone. The core approach is much simpler:
doAllThreadsSyscall sends a signal to every thread in allm, which
executes the system call from the signal handler. We use (SIGRTMIN +
1), aka SIGSETXID, the same signal used by glibc for this purpose. As
such, we are careful to only handle this signal on non-cgo binaries.
Synchronization with thread creation is a key part of this CL. The
comment near the top of doAllThreadsSyscall describes the required
synchronization semantics and how they are achieved.
Note that current use of allocmLock protects the state mutations of allm
that are also protected by sched.lock. allocmLock is used instead of
sched.lock simply to avoid holding sched.lock for so long.
Fixes#50113
Change-Id: Ic7ea856dc66cf711731540a54996e08fc986ce84
Reviewed-on: https://go-review.googlesource.com/c/go/+/383434
Reviewed-by: Austin Clements <austin@google.com>
Trust: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
syscall_runtime_doAllThreadsSyscall is only used on Linux. In
preparation of a follow-up CL that will modify the function to use other
Linux-only functions, move it to os_linux.go with no changes.
For #50113.
Change-Id: I348b6130038603aa0a917be1f1debbca5a5a073f
Reviewed-on: https://go-review.googlesource.com/c/go/+/383996
Trust: Michael Pratt <mpratt@google.com>
Reviewed-by: Andrew G. Morgan <agm@google.com>
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
CL 337350 changed mp.fastrand from a [2]uint32 to a uint64 and changed
the initialization to a single call of int64Hash. However, int64Hash
returns uintptr, so 32-bit systems this always left the most
significant 32 bits of mp.fastrand initialized to 0. The new code also
did not protect against initializing mp.fastrand to 0, which on a
system that does not implement math.Mul64 (most 32-bit systems) would
lead fastrand to always return 0.
This CL restores the mp.fastrand initialization to what it was before
CL 337350, adjusted for the change from [2]uint32 to uint64.
Change-Id: I663b415d9424d967e8e665ce2d017604dcd5b204
Reviewed-on: https://go-review.googlesource.com/c/go/+/383916
Trust: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
