| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
| |
After a THROWTO message has been handle the message closure is
overwritten by a NULL message. We must ensure that the original
closure's pointers continue to be visible to the nonmoving GC.
|
| |
|
|
|
| |
This is necessary since emptyInbox may read from to_cap->inbox without
taking cap->lock.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
| |
executeMessage previously had a write barrier at the beginning of its
loop apparently in an attempt to synchronize with another thread's
writes to the Message. I would guess that the author had intended to use
a load barrier here given that there are no globally-visible writes done
in executeMessage.
I've removed the redundant barrier since the necessary load barrier is
now provided by the ACQUIRE_LOAD.
|
| | |
|
| |
|
|
|
|
|
|
|
|
| |
Previously we would assert that threads which are sending a
`MSG_THROWTO` message must have their blocking status be blocked on the
message. In the usual case of a thread throwing to another thread this
is guaranteed by `stg_killThreadzh`. However, `throwToSelf`, used by
the GC to kill threads which ran out of heap, failed to guarantee this.
Noted while debugging #17785.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This extends the non-moving collector to allow concurrent collection.
The full design of the collector implemented here is described in detail
in a technical note
B. Gamari. "A Concurrent Garbage Collector For the Glasgow Haskell
Compiler" (2018)
This extension involves the introduction of a capability-local
remembered set, known as the /update remembered set/, which tracks
objects which may no longer be visible to the collector due to mutation.
To maintain this remembered set we introduce a write barrier on
mutations which is enabled while a concurrent mark is underway.
The update remembered set representation is similar to that of the
nonmoving mark queue, being a chunked array of `MarkEntry`s. Each
`Capability` maintains a single accumulator chunk, which it flushed
when it (a) is filled, or (b) when the nonmoving collector enters its
post-mark synchronization phase.
While the write barrier touches a significant amount of code it is
conceptually straightforward: the mutator must ensure that the referee
of any pointer it overwrites is added to the update remembered set.
However, there are a few details:
* In the case of objects with a dirty flag (e.g. `MVar`s) we can
exploit the fact that only the *first* mutation requires a write
barrier.
* Weak references, as usual, complicate things. In particular, we must
ensure that the referee of a weak object is marked if dereferenced by
the mutator. For this we (unfortunately) must introduce a read
barrier, as described in Note [Concurrent read barrier on deRefWeak#]
(in `NonMovingMark.c`).
* Stable names are also a bit tricky as described in Note [Sweeping
stable names in the concurrent collector] (`NonMovingSweep.c`).
We take quite some pains to ensure that the high thread count often seen
in parallel Haskell applications doesn't affect pause times. To this end
we allow thread stacks to be marked either by the thread itself (when it
is executed or stack-underflows) or the concurrent mark thread (if the
thread owning the stack is never scheduled). There is a non-trivial
handshake to ensure that this happens without racing which is described
in Note [StgStack dirtiness flags and concurrent marking].
Co-Authored-by: Ömer Sinan Ağacan <omer@well-typed.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Here the following changes are introduced:
- A read barrier machine op is added to Cmm.
- The order in which a closure's fields are read and written is changed.
- Memory barriers are added to RTS code to ensure correctness on
out-or-order machines with weak memory ordering.
Cmm has a new CallishMachOp called MO_ReadBarrier. On weak memory machines, this
is lowered to an instruction that ensures memory reads that occur after said
instruction in program order are not performed before reads coming before said
instruction in program order. On machines with strong memory ordering properties
(e.g. X86, SPARC in TSO mode) no such instruction is necessary, so
MO_ReadBarrier is simply erased. However, such an instruction is necessary on
weakly ordered machines, e.g. ARM and PowerPC.
Weam memory ordering has consequences for how closures are observed and mutated.
For example, consider a closure that needs to be updated to an indirection. In
order for the indirection to be safe for concurrent observers to enter, said
observers must read the indirection's info table before they read the
indirectee. Furthermore, the entering observer makes assumptions about the
closure based on its info table contents, e.g. an INFO_TYPE of IND imples the
closure has an indirectee pointer that is safe to follow.
When a closure is updated with an indirection, both its info table and its
indirectee must be written. With weak memory ordering, these two writes can be
arbitrarily reordered, and perhaps even interleaved with other threads' reads
and writes (in the absence of memory barrier instructions). Consider this
example of a bad reordering:
- An updater writes to a closure's info table (INFO_TYPE is now IND).
- A concurrent observer branches upon reading the closure's INFO_TYPE as IND.
- A concurrent observer reads the closure's indirectee and enters it. (!!!)
- An updater writes the closure's indirectee.
Here the update to the indirectee comes too late and the concurrent observer has
jumped off into the abyss. Speculative execution can also cause us issues,
consider:
- An observer is about to case on a value in closure's info table.
- The observer speculatively reads one or more of closure's fields.
- An updater writes to closure's info table.
- The observer takes a branch based on the new info table value, but with the
old closure fields!
- The updater writes to the closure's other fields, but its too late.
Because of these effects, reads and writes to a closure's info table must be
ordered carefully with respect to reads and writes to the closure's other
fields, and memory barriers must be placed to ensure that reads and writes occur
in program order. Specifically, updates to a closure must follow the following
pattern:
- Update the closure's (non-info table) fields.
- Write barrier.
- Update the closure's info table.
Observing a closure's fields must follow the following pattern:
- Read the closure's info pointer.
- Read barrier.
- Read the closure's (non-info table) fields.
This patch updates RTS code to obey this pattern. This should fix long-standing
SMP bugs on ARM (specifically newer aarch64 microarchitectures supporting
out-of-order execution) and PowerPC. This fixes issue #15449.
Co-Authored-By: Ben Gamari <ben@well-typed.com>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Test Plan: ci
Reviewers: osa1, bgamari, erikd
Reviewed By: osa1
Subscribers: rwbarton, thomie, carter
Differential Revision: https://phabricator.haskell.org/D4517
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The existing internal counters:
* gc_alloc_block_sync
* whitehole_spin
* gen[g].sync
* gen[1].sync
are now not shown in the -s report unless --internal-counters is also passed.
If --internal-counters is passed we now show the counters above, reformatted, as
well as several other counters. In particular, we now count the yieldThread()
calls that SpinLocks do as well as their spins.
The added counters are:
* gc_spin (spin and yield)
* mut_spin (spin and yield)
* whitehole_threadPaused (spin only)
* whitehole_executeMessage (spin only)
* whitehole_lockClosure (spin only)
* waitForGcThreadsd (spin and yield)
As well as the following, which are not SpinLock-like things:
* any_work
* do_work
* scav_find_work
See the Note for descriptions of what these counters are.
We add busy_wait_nops in these loops along with the counter increment where it
was absent.
Old internal counters output:
```
gc_alloc_block_sync: 0
whitehole_gc_spin: 0
gen[0].sync: 0
gen[1].sync: 0
```
New internal counters output:
```
Internal Counters:
Spins Yields
gc_alloc_block_sync 323 0
gc_spin 9016713 752
mut_spin 57360944 47716
whitehole_gc 0 n/a
whitehole_threadPaused 0 n/a
whitehole_executeMessage 0 n/a
whitehole_lockClosure 0 0
waitForGcThreads 2 415
gen[0].sync 6 0
gen[1].sync 1 0
any_work 2017
no_work 2014
scav_find_work 1004
```
Test Plan:
./validate
Check it builds with #define PROF_SPIN removed from includes/rts/Config.h
Reviewers: bgamari, erikd, simonmar, hvr
Reviewed By: simonmar
Subscribers: rwbarton, thomie, carter
GHC Trac Issues: #3553, #9221
Differential Revision: https://phabricator.haskell.org/D4302
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Summary:
The problem occurred when
* Threads A & B evaluate the same thunk
* Thread A context-switches, so the thunk gets blackholed
* Thread C enters the blackhole, creates a BLOCKING_QUEUE attached to
the blackhole and thread A's `tso->bq` queue
* Thread B updates the blackhole with a value, overwriting the BLOCKING_QUEUE
* We GC, replacing A's update frame with stg_enter_checkbh
* Throw an exception in A, which ignores the stg_enter_checkbh frame
Now we have C blocked on A's tso->bq queue, but we forgot to check the
queue because the stg_enter_checkbh frame has been thrown away by the
exception.
The solution and alternative designs are discussed in Note [upd-black-hole].
This also exposed a bug in the interpreter, whereby we were sometimes
context-switching without calling `threadPaused()`. I've fixed this
and added some Notes.
Test Plan:
* `cd testsuite/tests/concurrent && make slow`
* validate
Reviewers: niteria, bgamari, austin, erikd
Reviewed By: erikd
Subscribers: rwbarton, thomie
GHC Trac Issues: #13751
Differential Revision: https://phabricator.haskell.org/D3630
|
| |
|
|
| |
Our new CPP linter enforces this.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Test Plan: Validate on lots of platforms
Reviewers: erikd, simonmar, austin
Reviewed By: erikd, simonmar
Subscribers: michalt, thomie
Differential Revision: https://phabricator.haskell.org/D2699
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The `nat` type was an alias for `unsigned int` with a comment saying
it was at least 32 bits. We keep the typedef in case client code is
using it but mark it as deprecated.
Test Plan: Validated on Linux, OS X and Windows
Reviewers: simonmar, austin, thomie, hvr, bgamari, hsyl20
Differential Revision: https://phabricator.haskell.org/D2166
|
| |
|
|
| |
This reverts commit 39b5c1cbd8950755de400933cecca7b8deb4ffcd.
|
| |
|
|
|
|
|
|
| |
This will hopefully help ensure some basic consistency in the forward by
overriding buffer variables. In particular, it sets the wrap length, the
offset to 4, and turns off tabs.
Signed-off-by: Austin Seipp <austin@well-typed.com>
|
| |
|
|
| |
Signed-off-by: Austin Seipp <austin@well-typed.com>
|
| |
|
|
| |
Signed-off-by: Austin Seipp <austin@well-typed.com>
|
| |
|
|
|
|
|
|
|
| |
This adds some new functions: peekRunQueue, promoteInRunQueue,
singletonRunQueue and truncateRunQueue which help abstract away
manual linked list manipulation, making it easier to swap in
a new queue implementation.
Signed-off-by: Edward Z. Yang <ezyang@mit.edu>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
lnat was originally "long unsigned int" but we were using it when we
wanted a 64-bit type on a 64-bit machine. This broke on Windows x64,
where long == int == 32 bits. Using types of unspecified size is bad,
but what we really wanted was a type with N bits on an N-bit machine.
StgWord is exactly that.
lnat was mentioned in some APIs that clients might be using
(e.g. StackOverflowHook()), so we leave it defined but with a comment
to say that it's deprecated.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The parallel GC was using setContextSwitches() to stop all the other
threads, which sets the context_switch flag on every Capability. That
had the side effect of causing every Capability to also switch
threads, and since GCs can be much more frequent than context
switches, this increased the context switch frequency. When context
switches are expensive (because the switch is between two bound
threads or a bound and unbound thread), the difference is quite
noticeable.
The fix is to have a separate flag to indicate that a Capability
should stop and return to the scheduler, but not switch threads. I've
called this the "interrupt" flag.
|
| |
|
|
|
|
|
|
|
| |
So we can now get these in ThreadScope:
19487000: cap 1: stopping thread 6 (blocked on black hole owned by thread 4)
Note: needs an update to ghc-events. Older ThreadScopes will just
ignore the new information.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch makes two changes to the way stacks are managed:
1. The stack is now stored in a separate object from the TSO.
This means that it is easier to replace the stack object for a thread
when the stack overflows or underflows; we don't have to leave behind
the old TSO as an indirection any more. Consequently, we can remove
ThreadRelocated and deRefTSO(), which were a pain.
This is obviously the right thing, but the last time I tried to do it
it made performance worse. This time I seem to have cracked it.
2. Stacks are now represented as a chain of chunks, rather than
a single monolithic object.
The big advantage here is that individual chunks are marked clean or
dirty according to whether they contain pointers to the young
generation, and the GC can avoid traversing clean stack chunks during
a young-generation collection. This means that programs with deep
stacks will see a big saving in GC overhead when using the default GC
settings.
A secondary advantage is that there is much less copying involved as
the stack grows. Programs that quickly grow a deep stack will see big
improvements.
In some ways the implementation is simpler, as nothing special needs
to be done to reclaim stack as the stack shrinks (the GC just recovers
the dead stack chunks). On the other hand, we have to manage stack
underflow between chunks, so there's a new stack frame
(UNDERFLOW_FRAME), and we now have separate TSO and STACK objects.
The total amount of code is probably about the same as before.
There are new RTS flags:
-ki<size> Sets the initial thread stack size (default 1k) Egs: -ki4k -ki2m
-kc<size> Sets the stack chunk size (default 32k)
-kb<size> Sets the stack chunk buffer size (default 1k)
-ki was previously called just -k, and the old name is still accepted
for backwards compatibility. These new options are documented.
|
| |
|
|
|
|
|
|
| |
This was leading to looping and excessive allocation, when the
computation should have just blocked on the black hole.
Reported by Christian Höner zu Siederdissen <choener@tbi.univie.ac.at>
on glasgow-haskell-users.
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
| |
We just about got away with this on x86 which isn't
alignment-sensitive. The result of the memory load is compared
against a few different values, but there is a fallback case that
happened to be the right thing when the pointer was tagged. A good
bug to find, nonetheless.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The list of threads blocked on an MVar is now represented as a list of
separately allocated objects rather than being linked through the TSOs
themselves. This lets us remove a TSO from the list in O(1) time
rather than O(n) time, by marking the list object. Removing this
linear component fixes some pathalogical performance cases where many
threads were blocked on an MVar and became unreachable simultaneously
(nofib/smp/threads007), or when sending an asynchronous exception to a
TSO in a long list of thread blocked on an MVar.
MVar performance has actually improved by a few percent as a result of
this change, slightly to my surprise.
This is the final cleanup in the sequence, which let me remove the old
way of waking up threads (unblockOne(), MSG_WAKEUP) in favour of the
new way (tryWakeupThread and MSG_TRY_WAKEUP, which is idempotent). It
is now the case that only the Capability that owns a TSO may modify
its state (well, almost), and this simplifies various things. More of
the RTS is based on message-passing between Capabilities now.
|
| |
|
|
|
|
|
| |
This fixes #3838, and was made possible by the new BLACKHOLE
infrastructure. To allow reording of the run queue I had to make it
doubly-linked, which entails some extra trickiness with regard to
GC write barriers and suchlike.
|
| | |
|
|
|
This replaces the global blackhole_queue with a clever scheme that
enables us to queue up blocked threads on the closure that they are
blocked on, while still avoiding atomic instructions in the common
case.
Advantages:
- gets rid of a locked global data structure and some tricky GC code
(replacing it with some per-thread data structures and different
tricky GC code :)
- wakeups are more prompt: parallel/concurrent performance should
benefit. I haven't seen anything dramatic in the parallel
benchmarks so far, but a couple of threading benchmarks do improve
a bit.
- waking up a thread blocked on a blackhole is now O(1) (e.g. if
it is the target of throwTo).
- less sharing and better separation of Capabilities: communication
is done with messages, the data structures are strictly owned by a
Capability and cannot be modified except by sending messages.
- this change will utlimately enable us to do more intelligent
scheduling when threads block on each other. This is what started
off the whole thing, but it isn't done yet (#3838).
I'll be documenting all this on the wiki in due course.
|
