memory.qdoc

Go to the documentation of this file.

// Copyright (C) 2023 The Qt Company Ltd.
// SPDX-License-Identifier: LicenseRef-Qt-Commercial OR GFDL-1.3-no-invariants-only
/*!
 
\page qtqml-javascript-memory.html
\title Memory Management in the JavaScript Engine
\brief Describes how the JavaScript Engine manages memory.
 
\section1 Introduction
 
This document describes the \e dynamic memory management of the JavaScript
Engine in QML. It is a rather technical, in depth description. You only need
to read this if you care about the exact characteristics of JavaScript memory
management in QML. In particular, it can be helpful if you're trying to
optimize your application for maximum performance.
 
\note By compiling your QML code to C++ using the \l{Qt Quick Compiler}
you can avoid much of the JavaScript heap usage. The generated C++ code uses the
familiar C++ stack and heap for storing objects and values. The
\l{JavaScript Host Environment}, however, always uses some JavaScript-managed
memory, no matter if you use it or not. If you use features that cannot be
compiled to C++, the engine will fall back to interpretation or JIT compilation
and use JavaScript objects stored on the JavaScript heap, though.
 
\section1 Basic Principles
 
The JavaScript engine in QML has a dedicated memory manager that requests
address space in units of multiple pages from the operating system. Objects,
strings, and other managed values created in JavaScript are then placed in this
address space, using the JavaScript engine's own allocation scheme. The
JavaScript engine does not use the C library's malloc() and free(), nor the
default implementations of C++'s new and delete to allocate memory for
JavaScript objects.
 
Requests for address space are generally done with mmap() on Unix-like systems
and with VirtualAlloc() on windows. There are several platform-specific
implementations of those primitives. Address space reserved this way is not
immediately committed to physical memory. Rather, the operating system notices
when a page of memory is actually accessed and only then commits it. Therefore,
the address space is practically free and having a lot of it gives the
JavaScript memory manager the leverage it needs to place objects in an efficient
way on the JavaScript heap. Furthermore, there are platform-specific techniques
to tell the operating system that a chunk of address space, though still
reserved, does not have to be mapped into physical memory for the time being.
The operating system can then decommit the memory as needed and use it for other
tasks. Crucially, most operating systems do not guarantee immediate action on
such a decommit request. They will only decommit the memory when it is actually
needed for something else. On Unix-like systems we generally use madvise() for
this. Windows has specific flags to VirtualFree() to do the equivalent.
 
\note There are memory profiling tools that do not understand this mechanism and
over-report JavaScript memory usage.
 
All values stored on the JavaScript heap are subject to garbage collection.
None of the values are immediately "deleted" when they go out of scope or are
otherwise "dropped". Only the garbage collector may remove values from the
JavaScript heap and return memory (see \l{Garbage Collection} below for how
this works).
 
\section1 QObject-based Types
 
QObject-based types, and in particular everything you can phrase as a QML
element, are allocated on the C++ heap. Only a small wrapper around the pointer
is placed on the JavaScript heap when a QObject is accessed from JavaScript.
Such a wrapper, however, can own the QObject it points to. See
\l{QJSEngine::ObjectOwnership}. If the wrapper owns the object, it will be
deleted when the wrapper is garbage-collected. You can then also manually
trigger the deletion by calling the destroy() method on it. destroy() internally
calls \l{QObject::deleteLater()}. It will therefore not immediately delete the
object, but wait for the next event loop iteration.
 
QML-declared \e properties of objects are stored on the JavaScript heap. They
live as long as the object they belong to lives. Afterwards they are removed the
next time the garbage collector runs.
 
\section1 Object Allocation
 
In JavaScript, any structured type is an object. This includes function objects,
arrays, regular expressions, date objects and much more. QML has a number of
internal object types, such as the above mentioned QObject wrapper. Whenever
an object is created, the memory manager locates some storage for it on the
JavaScript heap.
 
JavaScript strings are also managed values, but their string data is not
allocated on the JavaScript heap. Similar to QObject wrappers, the heap objects
for strings are just thin wrappers around a pointer to string data.
 
When allocating memory for an object, the size of the object is first rounded up
to 32 byte alignment. Each 32 byte piece of address space is called a "slot".
For objects smaller than a "huge size" threshold, the memory manager performs
a series of attempts to place the object in memory:
\list
\li The memory manager keeps linked lists of previously freed pieces of heap,
    called "bins". Each bin holds pieces of heap with a fixed per-bin size in
    slots. If the bin for the right size is not empty, it picks the first entry
    and places the object there.
\li The memory that hasn't been used yet is managed via a bumper allocator. A
    bumper pointer points to the byte beyond the occupied address space. If
    there is still enough unused address space, the bumper is increased
    accordingly, and the object is placed in unused space.
\li A separate bin is kept for previously freed pieces of heap of varying sizes
    larger than the specific sizes mentioned above. The memory manager
    traverses this list and tries to find a piece it can split to accommodate
    the new object.
\li The memory manager searches the lists of specifically sized bins
    larger than the object to be allocated and tries to split one of those.
\li Finally, if none of the above works, the memory manager reserves more
    address space and allocates the object using the bumper allocator.
\endlist
 
Huge objects are handled by their own allocator. For each of those one or more
separate memory pages are obtained from the OS and managed separately.
 
Additionally, each new chunk of address space the memory manager obtains from
the OS gets a header that holds a number of flags for each slot:
\list
\li \e{object}: The first slot occupied by an object is flagged with this bit.
\li \e{extends}: Any further slots occupied by an object are flagged with this
    bit.
\li \e{mark}: When the garbage collector runs, it sets this bit if the object is
    still in use.
\endlist
 
\section1 Internal Classes
 
In order to minimize the required storage for metadata on what members
an object holds, the JavaScript engine assigns an "internal class" to each
object. Other JavaScript engines call this "hidden class" or "shape".
Internal classes are deduplicated and kept in a tree. If a property is
added to an object, the children of the current internal class are checked to
see if the same object layout has occurred before. If so, we can use the
resulting internal class right away. Otherwise we have to create a new one.
 
Internal classes are stored in their own section of the JavaScript heap that
otherwise works the same way as the general object allocation described above.
This is because internal classes have to be kept alive while the objects using
them are collected. Internal classes are then collected in a separate pass.
 
The actual property attributes stored in internal classes are \e not kept on
the JavaScript heap, though, but rather managed using new and delete.
 
\section1 Garbage Collection
 
The garbage collector used in the JavaScript engine is a non-moving,
stop-the-world Mark and Sweep design. In the \e mark phase we traverse all the
known places where live references to objects can be found. In particular:
 
\list
\li JavaScript globals
\li Undeletable parts of QML and JavaScript compilation units
\li The JavaScript stack
\li The persistent value storage. This is where QJSValue and similar classes
    keep references to JavaScript objects.
\endlist
 
For any object found in those places the mark bits are set recursively for
anything it references.
 
In the \e sweep phase the garbage collector then traverses the whole heap and
frees any objects not marked before. The resulting released memory is sorted
into the bins to be used for further allocations. If a chunk of address space
is completely empty, it is decommitted, but the address space is retained
(see \l{Basic Principles} above). If the memory usage grows again, the same
address space is re-used.
 
The garbage collector is triggered either manually by calling the \l [QML] {Qt::}{gc()} function
or by a heuristic that takes the following aspects into account:
 
\list
\li The amount of memory managed by object on the JavaScript heap, but not
    directly allocated on the JavaScript heap, such as strings and internal
    class member data. A dynamic threshold is maintained for those. If it is
    surpassed, the garbage collector runs and the threshold is increased. If
    the amount of managed external memory falls far below the threshold, the
    threshold is decreased.
\li The total address space reserved. The internal memory allocation on the
    JavaScript heap is only considered after at least some address space has
    been reserved.
\li The additional address space reservation since the last garbage collector
    run. If the amount of address space is more than double the amount of
    used memory after the last garbage collector run, we run the garbage
    collector again.
\endlist
 
\section1 Analyzing Memory Usage
 
In order to observe the development of both the address space and the number
of objects allocated in it, it is best to use a specialized tool. The
\l{Qt Creator: QML Profiler}{QML Profiler} provides a visualization that
helps here. More generic tools cannot see what the JavaScript memory manager
does within the address space it reserves and may not even notice that part
of the address space is not committed to physical memory.
 
Another way to debug memory usage are the
\l{QLoggingCategory}{logging categories} \e{qt.qml.gc.statistics} and
\e{qt.qml.gc.allocatorStats}. If you enable the \e{Debug} level for
qt.qml.gc.statistics, the garbage collector will print some information every
time it runs:
 
\list
\li How much total address space is reserved
\li How much memory was in use before and after the garbage collection
\li How many objects of various sizes were allocated so far
\endlist
 
The \e{Debug} level for qt.qml.gc.allocatorStats prints more detailed
statistics that also include how the garbage collector was triggered, timings
for the mark and sweep phases and a detailed breakdown of memory usage by bytes
and chunks of address space.
 
*/