Repository navigation
fix: free pooled secure allocations with their allocated size - #8115
Conversation
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 50 minutes. View limit details
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. |
|
✅ Final review complete — no blockers (commit fe26212) · triage: low · Phase 2 only (queue backlog) |
thepastaclaw
left a comment
There was a problem hiding this comment.
Final validation — Phase 2 only (queue backlog)
The PR correctly restores allocation/deallocation size symmetry in both secure allocator layers and explicitly cleanses PBKDF2 and serialized ECDH key material. The added regression coverage exercises allocator reuse, both BLS serialization schemes, both DKG decryption APIs, and truncated ciphertext handling. No in-scope correctness, security, compatibility, or test issues were found; the remaining CI build/test jobs were still pending, while completed code checks passed.
Review provenance
Source: reviewer 1: gpt-6.1-sol (agent: phase2-reviewer, role: general); reviewer 2: gpt-6.1-sol (agent: phase2-reviewer, role: dash-core-commit-history); final verifier: gpt-6.1-sol (agent: sol-verifier, role: final-verifier)
- Triage:
lowbygpt-6.1-sol(effort low) — The diff makes small, well-contained allocation-size corrections and explicit buffer-cleansing additions with regression tests; despite touching key handling, it is not large or intricate enough to qualify as critical. - Phase 1 reviewers: not run (skipped for throughput: 18 PRs queued, above the 10 limit)
- Fresh verifier:
gpt-6.1-sol— final-verifier; agentsol-verifier - Phase 2 reviewers:
gpt-6.1-sol— general (completed, effort medium); agentphase2-reviewer,gpt-6.1-sol— dash-core-commit-history (completed, effort medium); agentphase2-reviewer
mt_pooled_secure_allocator and the BLS secure_allocate wrapper each prepend a size_t header to the allocation, but their deallocation paths passed the size without it. The pool rounds sizes up to whole chunks, so when the header pushes an allocation across a chunk boundary, fewer chunks were freed and cleansed than were allocated. Current BLS callers request 32, 144 or 288 bytes, which round to the same single 304-byte chunk either way, so this is a latent asymmetry rather than an observed leak. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
e37d0ae to
fe26212
Compare
thepastaclaw
left a comment
There was a problem hiding this comment.
Re-review — Final validation — Phase 2 only (queue backlog)
The changes correctly restore allocation/deallocation size symmetry in both the pooled secure allocator and the BLS secure allocation wrapper. The boundary-crossing regression test exercises the generic allocator, and the available CI snapshot shows successful builds and linting with no failed code checks; remaining source-test jobs were still pending.
🔴 0 blocking | 🟡 0 suggestion(s) | 💬 0 nitpick(s)
Review provenance
Source: reviewer 1: gpt-6.1-sol (agent: phase2-reviewer, role: general); reviewer 2: gpt-6.1-sol (agent: phase2-reviewer, role: dash-core-commit-history); final verifier: gpt-6.1-sol (agent: sol-verifier, role: final-verifier)
- Triage:
lowbygpt-6.1-sol(effort low) — The diff makes two small, symmetric deallocation-size fixes in contained allocator and BLS wrapper code, with a focused regression test and no changes to cryptographic behavior or key handling. - Phase 1 reviewers: not run (skipped for throughput: 20 PRs queued, above the 10 limit)
- Fresh verifier:
gpt-6.1-sol— final-verifier; agentsol-verifier - Phase 2 reviewers:
gpt-6.1-sol— general (completed, effort medium); agentphase2-reviewer,gpt-6.1-sol— dash-core-commit-history (completed, effort medium); agentphase2-reviewer
🤖 Prompt for all review comments with AI agents
These findings are from an automated code review. Verify the current code and confirm that no unresolved issues remain.
No unresolved findings remain from the prior review on this head.
… size fe26212 fix: free pooled secure allocations with their allocated size (pasta) Pull request description: ## Issue being fixed or feature implemented `mt_pooled_secure_allocator` and the BLS `secure_allocate` wrapper each prepend a `size_t` header to the allocation, but their deallocation paths pass the size without that header. The underlying pool rounds sizes up to whole chunks, so when the header pushes an allocation across a chunk boundary, fewer chunks are freed and cleansed than were allocated, and each allocate/free cycle strands a chunk of locked memory. This is a latent asymmetry, not an observed leak: the pool chunk is 304 bytes, and current BLS callers request 32, 144 or 288 bytes, which round to one chunk on both the allocation and the free side. Only a request of 289..296 bytes (mod 304) through the BLS wrapper would hit it. The generic allocator is exposed for any size that crosses a chunk boundary once its own header is added. ## What was done? - Include each layer's header in its deallocation size, so it matches what the same layer requested on allocation. - Add an allocator unit test that repeats a boundary-crossing allocate/free cycle and checks that locked-pool usage stays flat. An earlier revision of this PR also cleansed the serialized ECDH key in `bls_ies.cpp` and the PBKDF2 temporaries. That was dropped: it wiped only some copies of the secrets involved (the DH point and the plaintext share buffers stayed uncleansed) and no test can observe it, so it does not belong with this fix. ## How Has This Been Tested? On macOS arm64 with depends, `--enable-debug --enable-werror`: - `./src/test/test_dash --run_test=allocator_tests,bls_tests` passes. - With the `mt_pooled_secure.h` change reverted, `allocator_tests/mt_pooled_secure_allocator_reuses_all_chunks` fails: locked-pool usage grows from 384 to 82,288 bytes over 1,000 cycles of a 33-byte allocation (two 40-byte chunks allocated, one freed). The `secure_free` change in `bls.cpp` has no test of its own; it is the same one-line symmetry fix one layer up, and no current caller reaches the affected sizes. ## Breaking Changes None. ## Checklist: - [x] I have performed a self-review of my own code - [ ] I have made corresponding changes to the documentation (not needed) - [x] I have assigned this pull request to a milestone 🤖 Generated with [Claude Code](https://claude.com/claude-code) Top commit has no ACKs. Tree-SHA512: 0539c70a4d421207b0aded68f6a1b0a37a3a7df78c74327effa78270ea0981bf5162ffd7c591780b148f48636c81f03d7e00dccb7c34defac95e27e38554e314 (cherry picked from commit ec0c1cd)
b534385 chore: prepare v24.0.2 release (pasta) 70e06ae Merge #8117: fix: races in watch-connection seed, partial sig-share batches, and ChainLock shutdown (merge-script) d50ea34 Merge #8115: fix: free pooled secure allocations with their allocated size (merge-script) 824e58a Merge #8113: fix(wallet): schedule CoinJoin auto-denomination per wallet and guard an empty round list (merge-script) 068529f Merge #8112: fix: answer govsync only to the requesting peer and index replayed orphan votes (merge-script) 484ec63 Merge #8119: fix(stats): wrong inputValue gauge, counter data race, and null-tip crash (merge-script) 28d5711 Merge #8120: fix(utils): accept -printcrashinfo in dash-cli and dash-tx (merge-script) 578b914 Merge bitcoin-core/gui#770: Revert "gui: provide wallet controller context to wallet actions" (Konstantin Akimov) d6902fa Merge bitcoin-core/gui#765: Fix wallet list hover crash on shutdown (Konstantin Akimov) Pull request description: ## Issue being fixed or feature implemented Backports and release prep for v24.0.2, on top of `v24.0.x` at 137d999 (v24.0.1). One PR carries both, so the release branch gets a single merge. Two fixes are the reason for the release: - **Memory leak (#8115).** Since v24.0.0 a node's locked-memory use grows for as long as it runs. On `v24.0.x` BLS still uses relic, where a G2 element is 296 bytes; the pooled secure allocator frees such an allocation with one chunk fewer than it allocated. The batch-verification weighting added in v24.0.0 makes one such allocation per verified message, so the pool keeps growing until restart. On develop (blst, 288-byte elements) the same code path does not leak, which is why #8115 describes the bug as latent. - **macOS crash on close (#8073's bitcoin-core/gui#765 and gui#770).** On macOS the menu bar stays usable after the main window is hidden for shutdown. `File > Open Wallet`, `Close Wallet` and the other wallet items then call the wallet controller after it has been deleted. The rest are small fixes that were already merged on develop and are low risk to carry. ## What was done? ### Picked In develop merge order. The Dash PRs are picked from their develop merge commits with `git cherry-pick -m1 -x`; the two GUI backports are picked with `-x` from the commits #8073 merged, like earlier partial picks on this branch. - bitcoin-core/gui#765 Fix wallet list hover crash on shutdown, and bitcoin-core/gui#770, which reverts its first commit (from #8073). Together they leave one change: `appMenuBar->clear()` when shutdown starts. - #8120 fix(utils): accept -printcrashinfo in dash-cli and dash-tx - #8119 fix(stats): wrong inputValue gauge, counter data race, and null-tip crash - #8112 fix: answer govsync only to the requesting peer and index replayed orphan votes - #8113 fix(wallet): schedule CoinJoin auto-denomination per wallet and guard an empty round list - #8115 fix: free pooled secure allocations with their allocated size - #8117 fix: races in watch-connection seed, partial sig-share batches, and ChainLock shutdown Then one commit, `chore: prepare v24.0.2 release`: - `configure.ac`: `_CLIENT_VERSION_BUILD` 1 → 2. - `doc/release-notes.md`: v24.0.2 notes. The v24.0.1 notes are archived verbatim to `doc/release-notes/dash/release-notes-24.0.1.md` and added to "Older releases". - `contrib/flatpak/org.dash.dash-core.metainfo.xml`: `24.0.2` entry dated 2026-11-13. The date needs changing if the tag is cut on another day. - `doc/man/*.1`: version string only. `-printcrashinfo` (#8120) is a hidden option, so no help text changed. ### Adaptations Every pick changes the same lines as its develop counterpart except: - **#8073:** only the two GUI commits are taken. The other two backports in that PR (bitcoin#28629, partial bitcoin#28639) are test and refactor changes for code `v24.0.x` doesn't need. - **#8112:** `test/functional/feature_governance.py` differs on `v24.0.x` (no pruned node from #7860). The `v24.0.x` file is kept and only #8112's additions are applied: the imports it needs, `GovernanceSyncPeer`, `check_govsync_is_peer_local()` and its call. The source change is identical. - **#8113:** one line of the new unit test. `v24.0.x` has the five-argument `CWallet` constructor and `CreateMockWalletDatabase()`, so the second wallet is built the way the fixture in the same file builds the first. The source change is identical. - **#8117:** in `src/llmq/utils.cpp` the neighbouring includes differ; the pick removes `<atomic>` and adds nothing. The source change is otherwise identical. ### Not picked - #8111 (check every BIP39 checksum bit): it can refuse a mnemonic that v24.0.1 accepts, so it should come with a major or minor release and a release note, not a patch release. - #8118 (RPC argument and locking fixes): changes RPC error behaviour and needs a hand-port to `rpc/masternode.cpp`. - #8122 (CacheMultiMap): the use-after-free can't be reached on `v24.0.x`, and the capacity part is hardening with no behaviour change. - #8047 (index startup after an unclean stop): on `v24.0.x` a synced node never records the uncommitted tip it handles, and the governance index it mainly covers doesn't exist here. - #8092 (wording): changes two translatable strings. - #8101, #8102, #8098, #8095 with #8121, #8116 with #7815, #8085: unchanged from the v24.0.1 decision in #8128. - Features, refactors, Bitcoin Core backports, CI and test-only changes merged since the branch point. #8103 (EvoDB left ahead of the coins DB after a reindex) is a candidate that is still open on develop. If it merges before this PR does, its pick will be added here. ## How Has This Been Tested? Linux x86_64, depends build with `--enable-debug --enable-werror`, GUI enabled. Head is b534385. - `dashd`, `dash-qt`, `test_dash` and `test_dash-qt` build cleanly; `dashd -version` reports `v24.0.2`. - `test_dash` (1106 cases), `test_dash-qt` and `test/util/test_runner.py` pass. That includes the picked unit tests: `allocator_tests/mt_pooled_secure_allocator_reuses_all_chunks` (#8115), `coinjoin_tests/coinjoin_maintenance_runs_for_every_wallet` (#8113) and the orphan-vote case in `governance_vote_processing_tests` (#8112). - The full functional suite (`test_runner.py -j14`): 312 passed, 12 skipped, 1 failed. The skipped ones are the usual environment-dependent tests (`feature_bind_port_*`, `rpc_bind.py --ipv6`, `interface_usdt_*`, previous-release compatibility and `feature_unsupported_utxo_db.py`). The run includes the scenarios the picks add or extend: `feature_governance.py` (#8112), `feature_stats.py` (#8119) and `interface_bitcoin_cli.py` (#8120). - The failure was `feature_llmq_is_cl_conflicts.py`, in `mine_cycle_quorum`: the node logs show the block for the final commitments being built about 40 ms before the second quorum's final commitment reached the mining node, so that quorum was not mined. It is a timing race in the test helper under load, in code none of the picks touch (develop reworked those waits in #7892). Run again on its own four times in parallel, it passed all four. - **Leak, end to end.** A regtest script (not part of the tree) mines a single-member InstantSend quorum, sends InstantSend-locked transactions and reads `getmemoryinfo` `locked.used` on each node every 50 locks. - The published v24.0.1 Linux binary: the four non-masternode nodes go from 39,568 bytes to 467,664 bytes over 600 locks, and are still growing at the end. - This branch: the same nodes stay at 20,096 bytes for all 600 locks. - The man pages were regenerated with `contrib/devtools/gen-manpages.py` from this build. They are byte-identical to the committed ones except for the `lock` debug category in `dashd.1` and `dash-qt.1`, which only an `--enable-debug` build lists. - The archived v24.0.1 notes are byte-identical to the previous `doc/release-notes.md`. Not tested: the macOS shutdown crash itself. The GUI change is the upstream one-liner and was built and run through `test_dash-qt` on Linux only. ## Breaking Changes None. No consensus change, and no change to which P2P messages a node accepts or how it scores peers. A node that syncs governance from a v24.0.2 node receives the same inventories as before; other peers of that node no longer receive them. ## Checklist: - [x] I have performed a self-review of my own code - [x] I have made corresponding changes to the documentation - [x] I have assigned this pull request to a milestone _(if you have permission to set one)_ 🤖 Generated with [Claude Code](https://claude.com/claude-code) Top commit has no ACKs. Tree-SHA512: 6d9861f4a8c2341889bb0ce5dd500a7083ea59395777a21c5e5d01e70ec8e8b489cf6425bc26ca79f15b1919610ceab10baada361043322dc689b7dfcae7cb5d
Issue being fixed or feature implemented
mt_pooled_secure_allocatorand the BLSsecure_allocatewrapper each prepend asize_theader to the allocation, but their deallocation paths pass the size without that header. The underlying pool rounds sizes up to whole chunks, so when the header pushes an allocation across a chunk boundary, fewer chunks are freed and cleansed than were allocated, and each allocate/free cycle strands a chunk of locked memory.This is a latent asymmetry, not an observed leak: the pool chunk is 304 bytes, and current BLS callers request 32, 144 or 288 bytes, which round to one chunk on both the allocation and the free side. Only a request of 289..296 bytes (mod 304) through the BLS wrapper would hit it. The generic allocator is exposed for any size that crosses a chunk boundary once its own header is added.
What was done?
An earlier revision of this PR also cleansed the serialized ECDH key in
bls_ies.cppand the PBKDF2 temporaries. That was dropped: it wiped only some copies of the secrets involved (the DH point and the plaintext share buffers stayed uncleansed) and no test can observe it, so it does not belong with this fix.How Has This Been Tested?
On macOS arm64 with depends,
--enable-debug --enable-werror:./src/test/test_dash --run_test=allocator_tests,bls_testspasses.mt_pooled_secure.hchange reverted,allocator_tests/mt_pooled_secure_allocator_reuses_all_chunksfails: locked-pool usage grows from 384 to 82,288 bytes over 1,000 cycles of a 33-byte allocation (two 40-byte chunks allocated, one freed).The
secure_freechange inbls.cpphas no test of its own; it is the same one-line symmetry fix one layer up, and no current caller reaches the affected sizes.Breaking Changes
None.
Checklist:
🤖 Generated with Claude Code