1911537 - Write a memory allocator for JSObject slots/elements buffers

Status: RESOLVED FIXED

(Core :: JavaScript: GC, enhancement, P3)

Description

[Jon Coppeard (:jonco)]

Currently these are allocated with jemalloc. Using our own allocator would have a number of advantages. It would:

• eliminate lock contention between foreground allocation and background sweeping
• reduce the work necessary in the JSObject finalizer
• optimise the nursery promotion check for whether something is in the tenured heap

This is also a step towards the future goal of allocating more internal data in the GC heap.

Comment 1

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 1: Tidy some inline nursery methods into Nursery-inl.h r?sfink

Comment 2

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 2: Make mark bitmap methods take void* so they don't have to be used with Cells r?sfink

Comment 3

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 3: Add a check for data protected by a specific mutex r?sfink

Comment 4

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 4: Add a dobuly linked list implementation that only uses a single word for the list r?sfink

The allocator uses a lot of lists because it has arrays of free lists per
supported allocation size. mozilla::LinkedList uses three words for the list
itself which would make this an unnecessarily large memory overhead. This
implementation only uses a single word for the list by storing a pointer to the
first list element rather than having the list be a special node in the list
itself.

Like mozilla::LinkedList it is a circular doubly linked list. Another
difference however is that it uses pointer tagging to remove the need for a
sentinel flag in each node, which ends up taking up a whole word due to
alignement constraints.

Comment 5

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 5: Rename nursery methods that can allocate malloc buffers to make that clear r?sfink

The eventual plan is to replace these with equivalent calls that use the new allcoator.

Comment 6

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 6: Add buffer allocator class and support for small allocations r?sfink

The allocator supports three size classes: small, medium and large. This adds
support for the small class for allocations of up to 128 bytes. This is backed
by our existing cell allocator.

This adds a new trace kind 'SmallBuffer' and new alloc kinds in powers of two.

This memory is accounted to the malloc heap size so as to disturb GC scheduling
as little as possible. Eventually the distiction between GC heap and malloc
heap will be removed.

Comment 7

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 7: Support allocating medium size allocations r?sfink

Comment 8

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 8: Support freeing medium size allocations r?sfink

Comment 9

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 9: Support resizing medium allocations in place where possible r?sfink

Comment 10

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 10: Add a way to allocate buffers of 1MB or greater r?sfink

This adds large buffers using the OS allocator. Buffers of this size are
relatively rare in the JS engine and so are not particularly optimised.

Comment 11

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 11: Add memory reporting for the buffer allocator and a way to dump statistics r?sfink

Comment 12

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 12: Add jsapi tests for the buffer allocator r?sfink

Comment 13

[Jeff Muizelaar [:jrmuizel]]

(In reply to Jon Coppeard (:jonco) from comment #0)

• reduce the work necessary in the JSObject finalizer
• optimise the nursery promotion check for whether something is in the tenured heap

How does this help with those things? Does this come from knowing something about the addresses?

Comment 14

[Jon Coppeard (:jonco)]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #13)
With the new allocator we can sweep to discard unmarked allocations in 1MB chunks, rather than having to run a finalizer to call free for each one individually. This is more efficient because each free operations updates allocator state and with the new allocator we can do these updates in bulk instead.

Managing the memory ourselves means we can use tricks to work out what kind of memory it is (e.g. for the nursery vs tenured check) such looking at the chunk kind stored at the start of each 1MB chunk. We can't do this for malloced memory because we don't want to make assumptions about the layout of their data structures.

Comment 15

[Jon Coppeard (:jonco)]

Attachment: Bug 1911537 - Part 13: Add magic check values for various data structures r?sfink

Comment 16

[Phabricator Automation]

Comment on attachment 9445569 [details]
Bug 1911537 - Part 13: Add magic check values for various data structures r?sfink

Revision D233106 was moved to bug 1940065. Setting attachment 9445569 [details] to obsolete.

Comment 17

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/20791092320d
Part 1: Tidy some inline nursery methods into Nursery-inl.h r=sfink
https://hg.mozilla.org/integration/autoland/rev/ba658f0d1824
Part 2: Make mark bitmap methods take void* so they don't have to be used with Cells r=sfink
https://hg.mozilla.org/integration/autoland/rev/a15f54d57882
Part 3: Add a check for data protected by a specific mutex r=sfink
https://hg.mozilla.org/integration/autoland/rev/6c3ba1086317
Part 4: Add a dobuly linked list implementation that only uses a single word for the list r=sfink
https://hg.mozilla.org/integration/autoland/rev/543c9847cc28
Part 5: Rename nursery methods that can allocate malloc buffers to make that clear r=sfink
https://hg.mozilla.org/integration/autoland/rev/ae0acd4fa39c
Part 6: Add buffer allocator class and support for small allocations r=sfink
https://hg.mozilla.org/integration/autoland/rev/af074fe74bd9
Part 7: Support allocating medium size allocations r=sfink
https://hg.mozilla.org/integration/autoland/rev/a1bd5de92497
Part 8: Support freeing medium size allocations r=sfink
https://hg.mozilla.org/integration/autoland/rev/88009dbfaf57
Part 9: Support resizing medium allocations in place where possible r=sfink
https://hg.mozilla.org/integration/autoland/rev/f63df5aa379b
Part 10: Add a way to allocate buffers of 1MB or greater r=sfink
https://hg.mozilla.org/integration/autoland/rev/45ae405157e2
Part 11: Add memory reporting for the buffer allocator and a way to dump statistics r=sfink
https://hg.mozilla.org/integration/autoland/rev/ab05c4f378df
Part 12: Add jsapi tests for the buffer allocator r=sfink

Comment 18

[Cristina Horotan [:chorotan]]

https://hg.mozilla.org/mozilla-central/rev/20791092320d
https://hg.mozilla.org/mozilla-central/rev/ba658f0d1824
https://hg.mozilla.org/mozilla-central/rev/a15f54d57882
https://hg.mozilla.org/mozilla-central/rev/6c3ba1086317
https://hg.mozilla.org/mozilla-central/rev/543c9847cc28
https://hg.mozilla.org/mozilla-central/rev/ae0acd4fa39c
https://hg.mozilla.org/mozilla-central/rev/af074fe74bd9
https://hg.mozilla.org/mozilla-central/rev/a1bd5de92497
https://hg.mozilla.org/mozilla-central/rev/88009dbfaf57
https://hg.mozilla.org/mozilla-central/rev/f63df5aa379b
https://hg.mozilla.org/mozilla-central/rev/45ae405157e2
https://hg.mozilla.org/mozilla-central/rev/ab05c4f378df

Comment 19

[Florin Bilt]

(In reply to Cristina Horotan [:chorotan] from comment #18)

https://hg.mozilla.org/mozilla-central/rev/20791092320d
https://hg.mozilla.org/mozilla-central/rev/ba658f0d1824
https://hg.mozilla.org/mozilla-central/rev/a15f54d57882
https://hg.mozilla.org/mozilla-central/rev/6c3ba1086317
https://hg.mozilla.org/mozilla-central/rev/543c9847cc28
https://hg.mozilla.org/mozilla-central/rev/ae0acd4fa39c
https://hg.mozilla.org/mozilla-central/rev/af074fe74bd9
https://hg.mozilla.org/mozilla-central/rev/a1bd5de92497
https://hg.mozilla.org/mozilla-central/rev/88009dbfaf57
https://hg.mozilla.org/mozilla-central/rev/f63df5aa379b
https://hg.mozilla.org/mozilla-central/rev/45ae405157e2
https://hg.mozilla.org/mozilla-central/rev/ab05c4f378df

Perfherder has detected a awsy performance change from push ab05c4f378df2225e26149a3fd8fe1e08886482e.

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)
4%	JS	windows11-64-2009-shippable-qr	fission tp6	170,187,023.94 -> 162,785,355.84
2%	JS	windows11-64-2009-shippable-qr	fission tp6	166,389,041.85 -> 162,804,663.42

Details of the alert can be found in the alert summary, including links to graphs and comparisons for each of the affected tests.

If you need the profiling jobs you can trigger them yourself from treeherder job view or ask a sheriff to do that for you.

You can run these tests on try with ./mach try perf --alert 43303

For more information on performance sheriffing please see our FAQ.