Andy Wingo
d2e745ac23
Rework large object space
...
Store allocations in a splay tree so that we can efficiently map from an
edge originating in the lospace to its object. Defer returning memory
to the OS to a periodic background thread, using a similar strategy as
for nofl and copy-space pages. Use a size-segregated freelist instead
of requiring a full best-fit search for those pages that haven't yet
been returned to the OS.
2025-01-06 16:59:54 +01:00
Andy Wingo
8e631ca3f3
Remove embedder requirement for per-object remset bits
...
Since we now have a field-logging write barrier, we don't need
per-object log bits.
2025-01-06 16:59:54 +01:00
Andy Wingo
4be3e69ac1
Add asserts to address map / address set
2025-01-06 16:59:54 +01:00
Andy Wingo
95868c70a2
Add splay tree
2025-01-06 16:59:54 +01:00
Andy Wingo
2dcdfc24bc
Field set: per-edge clear callback is optional
2025-01-06 16:59:54 +01:00
Andy Wingo
6a6f5b09e3
Use PROT_NONE to reserve memory, then remap within that memory
...
Should play better with the kernel's overcommit heuristics.
2025-01-06 16:59:54 +01:00
Andy Wingo
922c13a183
Move mmap / munmap / madvise to gc-platform
2025-01-06 16:59:54 +01:00
Andy Wingo
ac016d5f83
nofl: Fix hole count / size computation for promoted blocks
2024-10-07 15:00:45 +02:00
Andy Wingo
cff99c75a8
Fix bug clearing field set
2024-10-04 14:00:47 +02:00
Andy Wingo
da4f1ec806
Fix bug in which head byte's logged bits were not cleared
2024-10-04 13:51:49 +02:00
Andy Wingo
b4ea55b9c4
Don't clear log bits when marking
...
This happens in a post-pass.
2024-10-04 13:51:27 +02:00
Andy Wingo
6d48e12f78
Add assertions when forgetting nofl edges
2024-10-04 13:51:17 +02:00
Andy Wingo
745a5ab558
Don't clear remembered set in non-generational mode
2024-10-04 13:50:57 +02:00
Andy Wingo
478b9de798
Add assertions when pushing edge buffers
2024-10-04 13:50:35 +02:00
Andy Wingo
3c8c956f4c
Add gc_edge_address
2024-10-04 13:49:27 +02:00
Andy Wingo
1a79c3a451
mmc: only serialize root-tracing if there are pinned roots
2024-10-04 11:41:08 +02:00
Andy Wingo
b5c36b9fd8
Explicitly support immediate values
...
Because we have to deref edges ourselves, as part of generational
marking, we need to ignore edges that don't point to heap objects.
2024-10-04 11:40:09 +02:00
Andy Wingo
1ecb45a437
Switch mmc to field-logging write barrier
...
Instead of the card table, use metadata bytes for field-logging. More
precision should lead to less work during the pause.
2024-10-02 21:25:09 +02:00
Andy Wingo
1493bf6398
Add gc_object_is_old_generation
...
Will be useful for write barriers.
2024-10-01 15:48:55 +02:00
Andy Wingo
42bf36d7cc
Add nursery for lospace
2024-10-01 14:38:08 +02:00
Andy Wingo
4aa5d04f08
Fix sense of "large_object_space_is_copied".
...
* src/large-object-space.h (large_object_space_is_copied): I don't
understand why or how it was like it was!
2024-10-01 13:18:15 +02:00
Andy Wingo
8e1574491a
Fix ephemerons test for mmc
...
Change to avoid detecting OOM based on no allocation since last GC, if
the collection was explicitly triggered by the user.
2024-10-01 13:16:43 +02:00
Andy Wingo
1f4e3bdf37
Add field-logging write barrier (fast path only)
...
Add a new kind of write barrier, one which has a bit per field; the
mutator that sets the bit will need to add the field's location (the
edge) to a remembered set. Here we just have the fast-path
implementation.
2024-10-01 10:34:27 +02:00
Andy Wingo
3955d2ad96
Factor out locking utils to separate header
2024-09-30 20:52:45 +02:00
Andy Wingo
691c777e7b
Fix ABA problem in the copy space
...
Same concerns as the previous fix to the nofl space.
2024-09-30 12:29:35 +02:00
Andy Wingo
326e925f4c
Fix an ABA problem in the nofl space
...
We use Treiber stacks to represent sets of blocks: blocks to sweep, full
blocks, and so on. This is fine as long as we are only adding to or
only removing from those sets, but as soon as we have concurrent add and
remove, we need to avoid the ABA problem.
Concurrent add and remove occurs for partly-full blocks, which are both
acquired and released by mutators; empty blocks, which can be added to
by heap growth at the same time as the mutator acquires them; and the
paged-out queue, which is also concurrent with heap growth/shrinkage.
2024-09-30 10:59:47 +02:00
Andy Wingo
8fba0e5322
Implement cooperative safepoint API
...
Fixes https://github.com/wingo/whippet/issues/9 .
2024-09-18 11:54:36 +02:00
Andy Wingo
9f26dbb1fc
Implement per-object pinning API
...
Fixes https://github.com/wingo/whippet/issues/6 .
2024-09-18 10:55:02 +02:00
Andy Wingo
a722b9c13f
Force major GC before signalling OOM
2024-09-16 15:46:49 +02:00
Andy Wingo
2b59efd9fc
Fix semi-space with fixed-sized heaps
2024-09-16 15:06:50 +02:00
Andy Wingo
b7306950bc
Implement adaptive heap sizing for semi
2024-09-16 14:19:54 +02:00
Andy Wingo
1bf250f62a
Heap growth can compete with lospace for nofl blocks
2024-09-16 13:40:09 +02:00
Andy Wingo
317039d952
Relax assertion when expanding the heap
...
It could be that newly mapped blocks were already acquired by other threads.
2024-09-16 13:07:25 +02:00
Andy Wingo
dcfdc547f6
Whoops, fix refactor-induced locking problem
2024-09-16 11:45:01 +02:00
Andy Wingo
cf570d0206
Don't release shared worklist buffers when less than 256 kB
...
Fixes https://github.com/wingo/whippet/issues/8 .
2024-09-16 10:51:07 +02:00
Andy Wingo
7984f60eae
MMC and PCC defer actual page-out operations to background thread
...
Should avoid excessive VM traffic when allocating large objects, or when
the adaptive heap sizer is on and we see frequent expansions and
resizes.
2024-09-16 10:00:01 +02:00
Andy Wingo
d785f082b1
Factor out adapative heap sizer background thread to own file
...
This will let us piggy-back on the thread to asynchronously release
memory to the OS.
2024-09-16 10:00:01 +02:00
Andy Wingo
2818958c59
First version of adaptive heap sizing for pcc and mcc
2024-09-16 09:59:55 +02:00
Andy Wingo
1ff082705e
Remove scc
...
PCC with GC_PARALLEL=0 is exactly equivalent to SCC. Also now that PCC
will dynamically avoid atomic forwarding if parallelism is disabled at
run-time, there is no need to keep SCC around.
2024-09-10 11:10:47 +02:00
Andy Wingo
6545b34073
Reorder events in event listener; refactors to mmc and pcc
...
In GC, request mutators to stop before doing anything else; changes the
order of the event listener interface. Also, refactor mmc to look more
like pcc.
2024-09-10 10:55:38 +02:00
Andy Wingo
9f437485ec
MMC marks roots in parallel during the pause, not while stopping
...
Following the analysis in
https://wingolog.org/archives/2024/09/06/on-taking-advantage-of-ragged-stops ,
we simplify MMC by traversing roots only during the pause. This lets us
use gc_tracer parallel root-tracing.
2024-09-09 15:03:49 +02:00
Andy Wingo
8604ad6beb
mmc reformatting
2024-09-08 09:54:16 +02:00
Andy Wingo
44a7240e16
Rename "whippet" collector to "mmc": mostly marking collector
2024-09-02 13:19:05 +02:00
Andy Wingo
cf129f10de
nofl: Block marks are bytes
...
There was no need to use a bitvector, and the marks were only being
partially cleared. More straightforward (and still low overhead) as
bytes.
2024-08-30 21:20:34 +02:00
Andy Wingo
8d6db735fd
Nofl space can have discontiguous slabs
2024-08-30 16:53:20 +02:00
Andy Wingo
1a7d08baac
Add separate extents header
...
Will be useful to let slab heaps grow.
2024-08-30 16:53:20 +02:00
Andy Wingo
c949e4e4a9
Refactor representation of blocks in nofl space
2024-08-30 16:53:20 +02:00
Andy Wingo
59b85abbda
Fix regarding memset of block marks
2024-08-25 20:46:05 +02:00
Andy Wingo
59c9f5dff9
Mark blocks that are targets of evacuation
2024-08-25 08:58:55 +02:00
Andy Wingo
7b4a56c51a
nofl: Fix sticky mark bit treatment for block marks
2024-08-25 08:45:17 +02:00
