mirror of
https://gitlab.winehq.org/wine/wine-gecko.git
synced 2024-09-13 09:24:08 -07:00
Bug 936514 - Improve GC documentation comments r=billm DONTBUILD
This commit is contained in:
Jon Coppeard
parent
511751e611
commit
d1410db013
180
js/src/jsgc.cpp
180
js/src/jsgc.cpp
@ -5,32 +5,170 @@
|
||||
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
|
||||
|
||||
/*
|
||||
* This code implements a mark-and-sweep garbage collector. The mark phase is
|
||||
* incremental. Most sweeping is done on a background thread. A GC is divided
|
||||
* into slices as follows:
|
||||
* This code implements an incremental mark-and-sweep garbage collector, with
|
||||
* most sweeping carried out in the background on a parallel thread.
|
||||
*
|
||||
* Slice 1: Roots pushed onto the mark stack. The mark stack is processed by
|
||||
* popping an element, marking it, and pushing its children.
|
||||
* ... JS code runs ...
|
||||
* Slice 2: More mark stack processing.
|
||||
* ... JS code runs ...
|
||||
* Slice n-1: More mark stack processing.
|
||||
* ... JS code runs ...
|
||||
* Slice n: Mark stack is completely drained. Some sweeping is done.
|
||||
* ... JS code runs, remaining sweeping done on background thread ...
|
||||
* Full vs. zone GC
|
||||
* ----------------
|
||||
*
|
||||
* The collector can collect all zones at once, or a subset. These types of
|
||||
* collection are referred to as a full GC and a zone GC respectively.
|
||||
*
|
||||
* The atoms zone is only collected in a full GC since objects in any zone may
|
||||
* have pointers to atoms, and these are not recorded in the cross compartment
|
||||
* pointer map. Also, the atoms zone is not collected if any thread has an
|
||||
* AutoKeepAtoms instance on the stack, or there are any exclusive threads using
|
||||
* the runtime.
|
||||
*
|
||||
* It is possible for an incremental collection that started out as a full GC to
|
||||
* become a zone GC if new zones are created during the course of the
|
||||
* collection.
|
||||
*
|
||||
* Incremental collection
|
||||
* ----------------------
|
||||
*
|
||||
* For a collection to be carried out incrementally the following conditions
|
||||
* must be met:
|
||||
* - the collection must be run by calling js::GCSlice() rather than js::GC()
|
||||
* - the GC mode must have been set to JSGC_MODE_INCREMENTAL with
|
||||
* JS_SetGCParameter()
|
||||
* - no thread may have an AutoKeepAtoms instance on the stack
|
||||
* - all native objects that have their own trace hook must indicate that they
|
||||
* implement read and write barriers with the JSCLASS_IMPLEMENTS_BARRIERS
|
||||
* flag
|
||||
*
|
||||
* The last condition is an engine-internal mechanism to ensure that incremental
|
||||
* collection is not carried out without the correct barriers being implemented.
|
||||
* For more information see 'Incremental marking' below.
|
||||
*
|
||||
* If the collection is not incremental, all foreground activity happens inside
|
||||
* a single call to GC() or GCSlice(). However the collection is not complete
|
||||
* until the background sweeping activity has finished.
|
||||
*
|
||||
* An incremental collection proceeds as a series of slices, interleaved with
|
||||
* mutator activity, i.e. running JavaScript code. Slices are limited by a time
|
||||
* budget. The slice finishes as soon as possible after the requested time has
|
||||
* passed.
|
||||
*
|
||||
* Collector states
|
||||
* ----------------
|
||||
*
|
||||
* The collector proceeds through the following states, the current state being
|
||||
* held in JSRuntime::gcIncrementalState:
|
||||
*
|
||||
* - MARK_ROOTS - marks the stack and other roots
|
||||
* - MARK - incrementally marks reachable things
|
||||
* - SWEEP - sweeps zones in groups and continues marking unswept zones
|
||||
*
|
||||
* The MARK_ROOTS activity always takes place in the first slice. The next two
|
||||
* states can take place over one or more slices.
|
||||
*
|
||||
* In other words an incremental collection proceeds like this:
|
||||
*
|
||||
* Slice 1: MARK_ROOTS: Roots pushed onto the mark stack.
|
||||
* MARK: The mark stack is processed by popping an element,
|
||||
* marking it, and pushing its children.
|
||||
*
|
||||
* ... JS code runs ...
|
||||
*
|
||||
* Slice 2: MARK: More mark stack processing.
|
||||
*
|
||||
* ... JS code runs ...
|
||||
*
|
||||
* Slice n-1: MARK: More mark stack processing.
|
||||
*
|
||||
* ... JS code runs ...
|
||||
*
|
||||
* Slice n: MARK: Mark stack is completely drained.
|
||||
* SWEEP: Select first group of zones to sweep and sweep them.
|
||||
*
|
||||
* ... JS code runs ...
|
||||
*
|
||||
* Slice n+1: SWEEP: Mark objects in unswept zones that were newly
|
||||
* identified as alive (see below). Then sweep more zone
|
||||
* groups.
|
||||
*
|
||||
* ... JS code runs ...
|
||||
*
|
||||
* Slice n+2: SWEEP: Mark objects in unswept zones that were newly
|
||||
* identified as alive. Then sweep more zone groups.
|
||||
*
|
||||
* ... JS code runs ...
|
||||
*
|
||||
* Slice m: SWEEP: Sweeping is finished, and background sweeping
|
||||
* started on the helper thread.
|
||||
*
|
||||
* ... JS code runs, remaining sweeping done on background thread ...
|
||||
*
|
||||
* When background sweeping finishes the GC is complete.
|
||||
*
|
||||
* Incremental GC requires close collaboration with the mutator (i.e., JS code):
|
||||
* Incremental marking
|
||||
* -------------------
|
||||
*
|
||||
* 1. During an incremental GC, if a memory location (except a root) is written
|
||||
* to, then the value it previously held must be marked. Write barriers ensure
|
||||
* this.
|
||||
* 2. Any object that is allocated during incremental GC must start out marked.
|
||||
* 3. Roots are special memory locations that don't need write
|
||||
* barriers. However, they must be marked in the first slice. Roots are things
|
||||
* like the C stack and the VM stack, since it would be too expensive to put
|
||||
* barriers on them.
|
||||
* Incremental collection requires close collaboration with the mutator (i.e.,
|
||||
* JS code) to guarantee correctness.
|
||||
*
|
||||
* - During an incremental GC, if a memory location (except a root) is written
|
||||
* to, then the value it previously held must be marked. Write barriers
|
||||
* ensure this.
|
||||
*
|
||||
* - Any object that is allocated during incremental GC must start out marked.
|
||||
*
|
||||
* - Roots are marked in the first slice and hence don't need write barriers.
|
||||
* Roots are things like the C stack and the VM stack.
|
||||
*
|
||||
* The problem that write barriers solve is that between slices the mutator can
|
||||
* change the object graph. We must ensure that it cannot do this in such a way
|
||||
* that makes us fail to mark a reachable object (marking an unreachable object
|
||||
* is tolerable).
|
||||
*
|
||||
* We use a snapshot-at-the-beginning algorithm to do this. This means that we
|
||||
* promise to mark at least everything that is reachable at the beginning of
|
||||
* collection. To implement it we mark the old contents of every non-root memory
|
||||
* location written to by the mutator while the collection is in progress, using
|
||||
* write barriers. This is described in gc/Barrier.h.
|
||||
*
|
||||
* Incremental sweeping
|
||||
* --------------------
|
||||
*
|
||||
* Sweeping is difficult to do incrementally because object finalizers must be
|
||||
* run at the start of sweeping, before any mutator code runs. The reason is
|
||||
* that some objects use their finalizers to remove themselves from caches. If
|
||||
* mutator code was allowed to run after the start of sweeping, it could observe
|
||||
* the state of the cache and create a new reference to an object that was just
|
||||
* about to be destroyed.
|
||||
*
|
||||
* Sweeping all finalizable objects in one go would introduce long pauses, so
|
||||
* instead sweeping broken up into groups of zones. Zones which are not yet
|
||||
* being swept are still marked, so the issue above does not apply.
|
||||
*
|
||||
* The order of sweeping is restricted by cross compartment pointers - for
|
||||
* example say that object |a| from zone A points to object |b| in zone B and
|
||||
* neither object was marked when we transitioned to the SWEEP phase. Imagine we
|
||||
* sweep B first and then return to the mutator. It's possible that the mutator
|
||||
* could cause |a| to become alive through a read barrier (perhaps it was a
|
||||
* shape that was accessed via a shape table). Then we would need to mark |b|,
|
||||
* which |a| points to, but |b| has already been swept.
|
||||
*
|
||||
* So if there is such a pointer then marking of zone B must not finish before
|
||||
* marking of zone A. Pointers which form a cycle between zones therefore
|
||||
* restrict those zones to being swept at the same time, and these are found
|
||||
* using Tarjan's algorithm for finding the strongly connected components of a
|
||||
* graph.
|
||||
*
|
||||
* GC things without finalizers, and things with finalizers that are able to run
|
||||
* in the background, are swept on the background thread. This accounts for most
|
||||
* of the sweeping work.
|
||||
*
|
||||
* Reset
|
||||
* -----
|
||||
*
|
||||
* During incremental collection it is possible, although unlikely, for
|
||||
* conditions to change such that incremental collection is no longer safe. In
|
||||
* this case, the collection is 'reset' by ResetIncrementalGC(). If we are in
|
||||
* the mark state, this just stops marking, but if we have started sweeping
|
||||
* already, we continue until we have swept the current zone group. Following a
|
||||
* reset, a new non-incremental collection is started.
|
||||
*/
|
||||
|
||||
#include "jsgcinlines.h"
|
||||
|
||||
Loading…
Reference in New Issue
Block a user