Skip to content

Make HPACK safe for generational collection - #22

Merged
plajjan merged 3 commits into
mainfrom
generational-gc
Nov 9, 2026
Merged

plajjan merged 3 commits into
mainfrom
generational-gc

Conversation

@plajjan

@plajjan plajjan commented Nov 9, 2026

Copy link
Copy Markdown
Contributor

In Acton's generational GC mode a minor collection rescans only the heap blocks that the program recorded as written, so C code must put a write barrier on its pointer stores into GC memory. nghttp2 links the allocations of the HPACK deflater and inflater together by itself: it stores new entries into the dynamic table's ring buffer and hash map, and new header buffers into the inflater. That memory came from the ordinary GC heap through the allocator callbacks in src/hpack.ext.c, so a minor collection could free a table entry that the table still held.

The callbacks now allocate from acton_malloc_rescanned, which every collection rescans. The calloc callback checks for overflow, and the realloc callback allocates from the same kind when it gets a NULL pointer. The stores of the deflater and inflater pointers into their u64 attributes get ACTON_WBAR. Those two stores run while the constructor still holds the object, so the tests also pass without them, but they follow the rule for GC pointers kept in integer fields. In ordinary mode acton_malloc_rescanned returns ordinary GC memory, so nothing changes there.

The new test tables_churn_across_collections sends 20,000 header blocks through one Deflater and Inflater pair, so the dynamic tables fill, evict and refill across about 20 collections in generational mode and about 270 with the write-barrier checker. It asks for a 30 second timeout, since the checker makes it take close to two seconds. Without the fix the test crashes in every run with ACTON_GC_MODE=generational, and with acton test --gc-check it exits with status 70. With the fix all five tests pass in ordinary mode, in 10 generational runs and in 20 runs with acton test --gc-check, in both debug and --fast builds.

The repository had no CI, so this adds a workflow. On Ubuntu with Acton from the tip channel it runs acton test, and then acton test --gc-check --min-iter 0 --max-time 10000, since the checker makes some tests take longer than the default limit of 3 seconds allows. Run locally against Acton main, both steps pass with the fix, and without it the checker step fails because the new test exits with status 70.

This needs an Acton with generational collection, which provides ACTON_WBAR, acton_malloc_rescanned and acton test --gc-check. Acton main has it since actonlang/acton#3390, and CI needs a tip build from after that merge.

One Deflater and Inflater pair encodes and decodes 20,000 header
blocks. Each block repeats four headers of the previous block, which
the encoder sends as indexes into the dynamic table, and adds four new
ones. So the tables fill, evict their oldest entries and refill many
times while the collector runs. The test asks for a 30 second timeout,
since with the write-barrier checker it takes close to two seconds.

The test fails on this commit with an Acton that has generational
collection. With ACTON_GC_MODE=generational nghttp2 dereferences a NULL
pointer in a dynamic table entry that a minor collection freed. With
acton test --gc-check the checker finds new table entries and header
buffers linked into the ring buffer and the inflater without a write
barrier, and the test exits with status 70.
In generational mode a minor collection rescans only the heap blocks
that the program recorded as written. nghttp2 links the allocations of
the HPACK deflater and inflater together without a write barrier. It
stores new entries into the dynamic table's ring buffer and hash map,
and new header buffers into the inflater. That memory came from the
ordinary GC heap, so a minor collection could free a table entry that
the table still held.

The allocator callbacks that Deflater and Inflater give nghttp2 now
allocate from acton_malloc_rescanned, which every collection rescans.
The calloc callback checks the product of its sizes for overflow and
relies on acton_malloc_rescanned returning cleared memory. The realloc
callback allocates from the same kind for a NULL pointer, because
acton_realloc keeps the kind of an existing object but allocates
ordinary memory for NULL. In ordinary mode acton_malloc_rescanned
returns ordinary GC memory as before.

The deflater and inflater are kept as GC pointers in u64 attributes,
so the stores of those attributes now have the write barrier.

This needs an Acton with the write barrier of generational collection,
which provides ACTON_WBAR and acton_malloc_rescanned.
The repository had no CI. The workflow installs Acton from the tip
channel on Ubuntu and runs acton test, then acton test --gc-check. The
second run puts every test in generational GC mode under the
write-barrier checker, which exits with status 70 at a pointer store
into GC memory without its barrier. So a change that puts nghttp2's
memory back on the ordinary GC heap fails CI.

The checker makes the tests slower. --max-time 10000 gives each test
11 seconds instead of 3, and --min-iter 0 keeps a slow test at one
run.
@plajjan
plajjan merged commit 011fcc7 into main Nov 9, 2026
1 of 3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant