Add AlignedBlockList #6410

jercaianu · 2018-04-02T20:27:25Z

Implemented AlignedBlockList, an allocator which holds a list of other allocators and allows for constant time deallocations, using aligned allocations.

Internally, this holds a doubly linked list of other allocators. Each node contains an allocator, followed by its payload. Each node will be allocated by the parent allocator at addresses multiple of 'theAlignment'.
Whenever a deallocation happens, the buffer candidate for deallocation will be rounded to the nearest multiple of 'theAlignment', in order to find in constant time the allocator from which the buffer originated.

This will play a very important part in creating a fast and safe allocator.

Thanks,
Alex

dlang-bot · 2018-04-02T20:27:26Z

Thanks for your pull request and interest in making D better, @jercaianu! We are looking forward to reviewing it, and you should be hearing from a maintainer soon.
Please verify that your PR follows this checklist:

My PR is fully covered with tests (you can see the annotated coverage diff directly on GitHub with CodeCov's browser extension
My PR is as minimal as possible (smaller, focused PRs are easier to review than big ones)
I have provided a detailed rationale explaining my changes
New or modified functions have Ddoc comments (with Params: and Returns:)

Please see CONTRIBUTING.md for more information.

If you have addressed all reviews or aren't sure how to proceed, don't hesitate to ping us with a simple comment.

Bugzilla references

Your PR doesn't reference any Bugzilla issue.

If your PR contains non-trivial changes, please reference a Bugzilla issue or create a manual changelog.

Testing this PR locally

If you don't have a local development environment setup, you can use Digger to test this PR:

dub fetch digger
dub run digger -- build "master + phobos#6410"

jercaianu · 2018-04-02T20:27:52Z

ping @andralex @edi33416

wilzbach

A quick initial feedback - mostly on the style. This is not a review.
I think you could save a lot of code if you would use a custom, local atomicOp - maybe this can even be done with alias this'ing the respective integral type?

wilzbach · 2018-04-02T21:42:26Z

std/experimental/allocator/building_blocks/aligned_block_list.d

@@ -0,0 +1,629 @@
+// Written in the D programming language.
+/**


Please provide a description here.

wilzbach · 2018-04-02T21:43:43Z

std/experimental/allocator/building_blocks/aligned_block_list.d

@@ -0,0 +1,629 @@
+// Written in the D programming language.
+/**
+Source: $(PHOBOSSRC std/experimental/allocator/building_blocks/_aligned_block_list.d)


Preventing Ddoc's auto-hightlighting is no longer required - dlang/dlang.org#2307

wilzbach · 2018-04-02T21:43:56Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    static if (isShared)
+    SpinLock lock = SpinLock(SpinLock.Contention.brief);
+
+    private void moveToFront(AlignedBlockNode *tmp)


wilzbach · 2018-04-02T21:44:03Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+        }
+    }
+
+    AlignedBlockNode *root;


wilzbach · 2018-04-02T21:44:16Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    enum maxNodes = 50;
+
+    static if (isShared)
+    SpinLock lock = SpinLock(SpinLock.Contention.brief);


Indent (DStyle)

wilzbach · 2018-04-02T21:51:54Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+
+    void fun()
+    {
+        auto rnd = Random();


No. Use a fixed seed and a fixed engine, everything else might lead to random coverage changes.

wilzbach · 2018-04-02T21:52:54Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    a.allocatorForSize!512.parent = &pageAlloc;
+    a.allocatorForSize!256.parent = &pageAlloc;
+
+    auto rnd = Random();


Seed + fix the generator

wilzbach · 2018-04-02T21:53:20Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+Although allocations are in theory served in linear searching time, `deallocate` calls take
+$(BIGOH 1) time, by using aligned allocations. `ParentAllocator` must implement `alignedAllocate`
+and it must be able to allocate `theAlignment` bytes at the same alignment. Each aligned allocation
+done by `ParentAllocator` will contain metadata for an `Allocator`, followed by its payload.


Needs a ddoc Params section.

wilzbach · 2018-04-02T21:54:15Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    size_t goodAllocSize(const size_t n)
+    {
+        Allocator a = null;
+        return a.goodAllocSize(n);


Wouldn't this segfault?

If goodAllocSize does not use internal allocator state (and it shouldn't) it won't segfault.
Currently, this would work with all allocators in experimental

wilzbach · 2018-04-02T21:54:47Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+                static if (isShared)
+                {
+                    atomicOp!"+="(tmp.bytesUsed, n);
+                    lock.lock();


You never unlock this

A bit above I have a scope(exit) lock.unlock()

n8sh · 2018-04-03T07:02:18Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    import core.internal.spinlock : SpinLock;
+
+private:
+    struct AlignedBlockNode


static struct?

jercaianu · 2018-04-03T13:16:51Z

@wilzbach Thanks for comments

wilzbach

import std.std doesn't work.

wilzbach · 2018-04-04T04:06:06Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+{
+    import std.experimental.allocator.building_blocks.region : SharedRegion;
+    import std.experimental.allocator.building_blocks.ascending_page_allocator : SharedAscendingPageAllocator;
+    import std.std.experimental.allocator.building_blocks.null_allocator : NullAllocator;


std/experimental/allocator/building_blocks/aligned_block_list.d(512): Error: module null_allocator is in file 'std/std/experimental/allocator/building_blocks/null_allocator.d' which cannot be read
import path[0] = ../druntime/import
posix.mak:350: recipe for target 'generated/linux/debug/32/unittest/std/experimental/allocator/building_blocks/aligned_block_list.o' failed
make[1]: *** [generated/linux/debug/32/unittest/std/experimental/allocator/building_blocks/aligned_block_list.o] Error 1

-> You have a std.std

n8sh · 2018-04-04T17:35:54Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    int numNodes;
+
+    // If the numNodes exceeds this limit, we will start deallocating nodes
+    enum maxNodes = 50;


I'm curious as to where this number came from. Was it just a gut feeling? Did you make a decision to bias towards "probably not too high" or "probably not too low"?

Depending on what the caller passes, maxNodes should be different.
The deallocation of the blocks provided by ParentAllocator happen lazily inside allocate.
I didn't want to delete nodes too often, but also I didn't want to have high memory usage so I hardcoded something for my tests.
I'll investigate here a bit more to see if I can get rid of this, or maybe pass it as a parameter somehow.
Thanks!

n8sh · 2018-04-04T17:45:44Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+        auto next = tmp.next;
+        if (tmp.prev) tmp.prev.next = tmp.next;
+        if (tmp.next) tmp.next.prev = tmp.prev;
+        assert(parent.deallocate((cast(void*) tmp)[0 .. alignment]));


In release mode assertions are omitted and this code won't be run.

Example: https://run.dlang.io/is/m8gzG3

import std.stdio; auto foo(ref int x) { return x &= 1; } void main(string[] args) { int x = 255; assert(foo(x)); writeln(x); }

This prints 1 when run with no compiler arguments but prints 255 with -release.

Thanks, good catch!

n8sh · 2018-04-04T19:04:40Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    static if (stateSize!ParentAllocator) ParentAllocator parent;
+    else alias parent = ParentAllocator.instance;
+
+    enum ulong alignment = theAlignment;


I don't see how this could be correct if theAlignment is the size of each node and memory is allocated at an offset from the start of the node.

Since the metadata for each node and the payload are adjacent in memory, I considered the alignment to be for the whole block (metadata + payload).
Inside the payload, an Allocator can manage the allocations with its own alignment.

Edit: nevermind I see the problem

At the same time, I think it's not great for the caller to expect each allocation to have an alignment, but we silently add some padding before.
I think it's better for the alignment to be (metadata + payload), but clearly specify that and expose the length of the metadata.

Because other allocators in std.experiment.allocator use alignment to specifically mean the alignment of memory returned by allocate, use a different name here and everything will be fine. Otherwise, templated code that uses design-by-introspection can leap to a wrong conclusion -- the compiler doesn't read the DDOC, so a comment will not prevent that!

n8sh · 2018-04-04T19:08:19Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+done by `ParentAllocator` will contain metadata for an `Allocator`, followed by its payload.
+
+Params:
+    Allocator = the allocator which is used to manage each node


The documentation does an insufficient job of explaining what needs to be true of the Allocator!

Yes, you're right, I haven't explained here too much about the Allocator.
Allocator can be anything, as long as it has a constructor which receives ubyte[] and its ParentAllocator is the NullAllocator.

For the moment, these are the only constraints which come to mind.
I'll update the docs
Thanks!

n8sh · 2018-04-04T22:36:27Z

It isn't clear to me that the shared version is correct. Does using a spinlock prevent stale reads of the root pointer and prevent stale reads of AlignedBlockNode's next/prev pointers?

jercaianu · 2018-04-05T14:03:54Z

@n8sh
Thanks for the review!
The list only is changed when the lock is taken. Only under the lock new nodes are added/removed and root updated.
To increase performance, allocate happens when the lock is not taken. Each node has keepAlive, which basically acts as a refcount to make sure no node is removed while it is used for allocations, outside the lock.

Besides the unittests, I stress tested this and I haven't found any races.

n8sh · 2018-04-08T20:30:53Z

std/experimental/allocator/building_blocks/aligned_block_list.d

    else alias parent = ParentAllocator.instance;

-    enum ulong alignment = theAlignment;
+    enum ulong alignment = ParentAllocator.alignment;


If alignment means the alignment of the memory returned by allocate, it would appear this should be Allocator.alignment instead of ParentAllocator.alignment.

Also maybe it should have a guard like:

static if (hasStaticallyKnownAlignment!Allocator) enum alignment = Allocator.alignment;

The existence of std.experimental.allocator.common : hasStaticallyKnownAlignment implies that not all allocators have statically known alignments, but this may be obsolete or not applicable to the specific allocators required here.

andralex

please mind @wilzbach 's and my nits

andralex · 2018-04-16T15:16:57Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+import std.experimental.allocator.building_blocks.null_allocator;
+
+// Common function implementation for thread local and shared AlignedBlockList
+private mixin template AlignedBlockListImpl(bool isShared)


bool is OK for a private artifact

andralex · 2018-04-16T15:19:34Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    AlignedBlockNode* root;
+
+    // Number of active nodes
+    int numNodes;


Wouldn't size_t be better?

size_t is size of objects in memory. For counts we can use that of course, but just incidentally.

andralex · 2018-04-16T15:19:43Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+    int numNodes;
+
+    // If the numNodes exceeds this limit, we will start deallocating nodes
+    enum maxNodes = 64;


enum uint

andralex · 2018-04-16T15:21:00Z

std/experimental/allocator/building_blocks/aligned_block_list.d

+
+    enum ulong alignment = Allocator.alignment;
+
+    static if (hasMember!(ParentAllocator, "owns"))


jercaianu requested review from CyberShadow, MartinNowak and wilzbach as code owners April 2, 2018 20:27

wilzbach reviewed Apr 2, 2018

View reviewed changes

n8sh reviewed Apr 3, 2018

View reviewed changes

n8sh mentioned this pull request Apr 3, 2018

Issue 18715 - Non-documented unittests should not use unpredictableSeed or default Random alias #6414

Merged

jercaianu force-pushed the alignedBlockList branch 4 times, most recently from 090e429 to 2f9396d Compare April 3, 2018 22:02

wilzbach reviewed Apr 4, 2018

View reviewed changes

jercaianu force-pushed the alignedBlockList branch 2 times, most recently from 5fa6a6b to 478ecd5 Compare April 4, 2018 11:31

n8sh reviewed Apr 4, 2018

View reviewed changes

jercaianu force-pushed the alignedBlockList branch from 478ecd5 to 1704139 Compare April 8, 2018 16:40

n8sh reviewed Apr 8, 2018

View reviewed changes

jercaianu force-pushed the alignedBlockList branch from 1704139 to a16b23b Compare April 11, 2018 09:54

JackStouffer added Severity:Enhancement Allocators @andralex labels Apr 13, 2018

n8sh approved these changes Apr 13, 2018

View reviewed changes

andralex approved these changes Apr 16, 2018

View reviewed changes

andralex removed the @andralex label Apr 16, 2018

Add AlignedBlockList

0fb44b0

jercaianu force-pushed the alignedBlockList branch from a16b23b to 0fb44b0 Compare April 16, 2018 16:37

n8sh added the Merge:auto-merge label Apr 17, 2018

dlang-bot merged commit c5129be into dlang:master Apr 17, 2018

		@@ -0,0 +1,629 @@
		// Written in the D programming language.
		/**


		enum ulong alignment = Allocator.alignment;

		static if (hasMember!(ParentAllocator, "owns"))

Uh oh!

Add AlignedBlockList #6410

Add AlignedBlockList #6410

Uh oh!

Conversation

jercaianu commented Apr 2, 2018

Uh oh!

dlang-bot commented Apr 2, 2018

Bugzilla references

Testing this PR locally

Uh oh!

jercaianu commented Apr 2, 2018

Uh oh!

wilzbach left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jercaianu commented Apr 3, 2018

Uh oh!

wilzbach left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jercaianu Apr 5, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jercaianu Apr 5, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

n8sh commented Apr 4, 2018

Uh oh!

jercaianu commented Apr 5, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

andralex left a comment

jercaianu Apr 5, 2018 •

edited

Loading

jercaianu Apr 5, 2018 •

edited

Loading