Files
MobileGL/docs/archive/notes/p2/p2-results/fix-aba-review-v1.md
T

22 KiB
Raw Blame History

fix-aba review v1 — adversarial review of c52364cc

Reviewer ran everything below itself, in ~/w7/p2-fix-aba (the tree has since been rebased onto 2d690754; HEAD is 55d2af9b and git show 55d2af9b --format="" == git show c52364cc --format="", byte-identical, so every command below is equally a command against c52364cc). Builds used: build-linux (pull), build-push (push), and build-verify, which this reviewer configured and built (-DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON, rc 0) because the implementer had not. Tree left unmodified (git status --porcelain empty; only build-verify/ was added).

Verdict: 1 major (a coverage claim that the lane does not deliver), 6 minors. No functional defect was found: the control works, it is load-bearing, it is deterministic, it touches only identity guards, the pull build is untouched, and every gate command is green.


1. Verdict summary

# claim under review outcome
1 the control is load-bearing on BOTH arms holds for the vertex-array case; the review brief's own wording about the Texture/Framebuffer cases is refuted (§4.1)
2 the knob touches nothing but identity guards; pull build untouched; push build unchanged knob-off holds (one vestigial 4th consumer, minor 1)
3 the reproducer is deterministic; the old one depended on an allocator that never obliged holds (10/10, zero flips); "depends on no allocator behaviour" needs one qualification (minor 3)
4 the second lane is registered in every build shape the others are; G8 satisfied or restated holds; G8's command is green in build-verify, and the restatement is declared — but what it restates to is overstated (MAJOR 1)
5 build-verify: ctest -R HandleRecycle, ctest -L unit holds: 40/40 and 1562/1562
— P1's ScopedPipeVerb/no-verb concern: public GL only holds

2. MAJOR 1 — the AbaControlHandles lane does not defeat the generation, because the scenario never constructs a slot reuse

Where the claim is made (five places, all as fact):

  • commit message: "That is what makes the control cover the key P2 SHIPS … the old control said nothing about the generation in {slot, gen} - the whole of what makes the re-keyed memos ABA-safe"
  • MobileGL/MG_Backend/DirectVulkan/Renderer/MagmaPipeArms.h:170-173 — "on the handle arm the constant defeats the GENERATION in {slot, gen}, which is the whole of what makes the re-keyed memos ABA-safe"
  • MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp:102-105 — "claimed without the Owner compare, which is precisely 'the slot was recycled and Gen did not move'"
  • MobileGL/MG_IntegrationTest/CMakeLists.txt:834-836 — "AbaControlHandles … defeats the {slot, gen} GENERATION, which is what makes the re-keyed memos ABA-safe"
  • ~/w7/notes/p2/p2-results/fix-aba-v1.md §2, and §4.1's declared deviation, which is what the integrator will amend BRIEF-P2.md D18/G8 to say.

Why it is not what the lane demonstrates. Magma's mint has no death notification; slots are returned to the free list only by an age sweep:

  • MagmaPipeArms.h:316-317 — kSweepInterval = 256, kRetireAgeBoundaries = 1024;
  • MagmaPipeArms.h:285-309 — OnFrameBoundary() is the only producer of m_freeSlots, and it sweeps only when (m_boundary % 256) == 0, retiring entries idle for more than 1024 boundaries;
  • MagmaPipeArms.h:327-367 — ClaimSlot() reuses a slot only from m_freeSlots; otherwise it appends a fresh entry with Gen = 1;
  • nothing in MobileGL/MG_Backend/DirectVulkan consumes NotifyStateObjectDestroyed (grep -rn NotifyStateObjectDestroyed MobileGL/MG_Backend/DirectVulkan → no hits; the only site is MG_State/GLState/VertexArrayState/VertexArrayObject.cpp:53, which feeds the client-side tracker).

AVertexArrayAtARecycledAddressDoesNotInheritItsPredecessorsVertexInput issues 5 EndFrame()s (1 primer + kWarmupFrames=3 + 1 ABA frame). So when the replacement VAO calls Acquire, the free list is empty and it gets a brand-new slot with Gen = 1; the dead VAO's slot is still occupied. The handle compare that the knob-off run passes is decided by the slot, and the generation never participates.

Reproduce (standalone, outside the tree — the file is slotprobe.cpp, reproduced in §6):

$ clang++ -std=c++23 $(<flags from build-push/compile_commands.json>) \
      -I MobileGL/MG_Backend/DirectVulkan/Renderer -o /tmp/slotprobe slotprobe.cpp && /tmp/slotprobe
A) HandleRecycleScenario's shape (5 frame boundaries, then the replacement):
  primerVao (lifetimeId 1)     slot=1 gen=1
  redVao    (lifetimeId 2)     slot=2 gen=1
  greenVao  (lifetimeId 3)     slot=3 gen=1          <-- a NEW slot, not the dead one's
B) what a slot REUSE looks like (the ABA the generation exists to stop):
  first     (lifetimeId 1)     slot=1 gen=1
  second    (lifetimeId 2)     slot=1 gen=2          <-- needs >1024 idle boundaries

Consequences.

  1. A regression that removes ++m_entries[index].Gen (MagmaPipeArms.h:343) leaves every arm of this scenario green — Handles, Legacy, and both AbaControl lanes. That is the exact result magma-v3's probe B already measured (green over 432 integration cases) and which the result file §4.4 attributes to the cross-frame gate alone; slot non-reuse is the second, independent reason, and it survives the same-frame rewrite this round performed.
  2. What the knob really defeats on the handle arm is the whole handle compare (Owner == handle / vaoHandle == handle), of which the generation is one field that this scenario can never make load-bearing. "Defeats the handle compare" is true and worth having; "defeats the generation, which is the whole of what makes the re-keyed memos ABA-safe" is not what the lane shows.
  3. The failure shape is the same one this round was convened to fix, one level up: §1.1 of the result file correctly refuses to trust an unverified assumption about the heap allocator, and then makes an unverified assumption about the slot allocator in its replacement claim.

And the two requirements are mutually exclusive as the code stands, which the fix must face: the pixel-visible memo is ResolvedVertexBindings, which declines across frames by design (VulkanRenderer.cpp:3635-3644), so the ABA window must be one frame; a genuine slot reuse needs ≥1024 frame boundaries of idleness after the dead VAO's last draw, which necessarily puts the two draws in different frames. So no same-frame pixel reproducer can ever cover the generation for this memo — the claim is not merely unproven here, it is unreachable by this shape.

Remedies (either is acceptable; the first is a paragraph).

  • (a) Restate at all five sites: the control defeats the identity compare (slot and generation together) on whichever arm is running, and record — beside result-file §4.4, in MEASUREMENTS.md — that a generation-only regression is out of this scenario's reach because Magma's mint retires slots by age (256/1024) and the pixel-visible memo is same-frame. Amend BRIEF-P2.md D18/G8 to the same words, not to the stronger ones.
  • (b) Cover the generation where it is expressible: a unit test over MagmaPipeIdentityTable + MemosFor/LookupVaoDrawMemo that drives a real retire→reuse and asserts the memo is not inherited; that costs one MG_Test file and it is the only place a ++Gen deletion can be caught today.

3. Minors

Minor 1 — a fourth consumer of the knob, now dead, that bypasses the "one question" accessor. VulkanRenderer.cpp:3737:

const Bool compareLifetimeId = !MG_Config::Features.PipeHandleAbaControl;

The control's early return at VulkanRenderer.cpp:3653-3668 fires first, so this line can never be false. It contradicts the commit message ("Three sites, all behind one question (MagmaPipeAbaControlDefeatsIdentity)"), the result file's three-row table (§2), and the discipline MagmaPipeArms.h:151-154 sells ("three sites deciding separately could disagree"). It is also a trap: narrow the early return later and this site silently returns to D18's retired semantics. Delete it (restoring constexpr Bool compareLifetimeId = true; unconditionally), or route it through the accessor and say why it is unreachable.

Minor 2 — a stale comment the file's own doctrine forbids. HandleRecycleScenario.cpp:478-482 still says "BOTH the VAO and the buffer are recycled here so that a key built out of raw addresses matches while the bytes behind it do not." After this change the buffer is not recycled: both buffers are created up front (:494-495) and neither is deleted inside the ABA window (:524-536); only the VAO is. The header 30 lines above now says the opposite.

Minor 3 — "the reproducer no longer depends on the allocator" needs one qualification. It no longer depends on the C++ heap allocator (correct, and well evidenced). It still depends on the GL name allocator: :547-550 skips when glGenVertexArrays does not hand the name back, and a skip in an AbaControl lane is a control that stopped controlling while ctest stays rc 0. That is survivable only because TheReproducerRecyclesEveryName (:440-475) reds the same lane in that case — worth one sentence at :543-546, since the name check there now advertises itself as vestigial.

Minor 4 — the knob costs the shipped push arm a per-draw branch it did not cost before. D18's consumer sat on the legacy path (:3737), after the handle arm had returned, so the arm P2 measures carried no control branch. Now MagmaPipeAbaControlDefeatsIdentity() is a global load + branch at the top of LookupVaoDrawMemo (once per draw) and at the top of MemosFor (up to 3× per draw: VertexInputStateFactory.cpp:124,143,171,182), plus one per enabled attribute in ComputeHash (:67, which the old else if skipped entirely on the handle arm). Tiny, but P2's whole subject is per-draw overhead and D.4.2 will measure it. [[unlikely]], or hoisting the question once per draw, would remove the question.

Minor 5 — the aliased entry hands out vkBuffers belonging to whatever VAO last used slot 0. With the knob on, LookupVaoDrawMemo returns entry 0 for every VAO uncleared, so TryBindResolvedVertexBindings' fast path (VulkanRenderer.cpp:3602-3605) can bind a stale entry.vkBuffers[...] — including one whose BufferObject has since been deleted, if a lane's other cases ever delete a buffer between two draws of one frame. It cannot happen in the three cases as written (buffers outlive the frame), the frame-serial gate bounds it to one frame, and lavapipe is green 10/10 — but it is a live-VkBuffer hazard confined to the two control lanes, and worth one line next to :3661-3664's "nothing else about the entry is relaxed".

Minor 6 — G8's build-verify command does not arm the comparator over these lanes. The AbaControl* lanes do not set MOBILEGL_PIPE_VERIFY=1 (only the *.Verify.* lanes do, MG_IntegrationTest/CMakeLists.txt:1012-1062), so running G8 in build-verify proves the build, not that the control survives an armed comparator. Pre-existing lane design, not this change; the result file's item 2 ("build-verify should be run before the gate is signed off") is satisfied by §5 below, but the gate's prose implies more than the lane does.


4. Refutations of the review brief's own premises

4.1 "the VertexArray/Texture/Framebuffer AbaControl cases see STALE pixels on both arms" — false, and correctly so. The knob steers DirectVulkan's vertex-input keys only. Only the vertex-array case asserts the corruption (:553, armExpectsCorruption=true); the texture case passes armExpectsCorruption=false (:604) and the framebuffer case never calls ExpectPixelsFor at all (:684-692), so both expect FRESH in every arm — stated at :570-571. Measured in build-verify:

2468: [ HandleRecycle ] arm=AbaControl expected=STALE observed=STALE  - the draw after the VAO was recycled inside one frame
2469: [ HandleRecycle ] arm=AbaControl expected=FRESH observed=FRESH  - the draw after the texture was recycled
2472: [ HandleRecycle ] arm=AbaControl expected=STALE observed=STALE  - (AbaControlHandles) the VAO case
2473: [ HandleRecycle ] arm=AbaControl expected=FRESH observed=FRESH  - (AbaControlHandles) the texture case

So "with the knob off the AbaControl lanes FAIL" means one entry of four per lane fails; the other three pass unchanged. G8's "ABA-control arm red" is carried by a single case.

4.2 "the old version depended on the allocator handing back a 3920-byte block" — confirmed from the diff. Old ComputeHash set bufferKey = (Uint64)(SizeT)attr.Buffer.get(), and old LookupVaoDrawMemo only dropped the lifetime-id half of first.vaoKey == vao && first.vaoLifetimeId == lifetimeId — so the address was still required to match on both the buffer and the VAO. The recorded red is ~/w7/logs/aba-baseline.log (96% tests passed, 1 tests failed out of 28, the vertex-array entry, "should be all red, but 11625 of 11625 pixels (100%) are not").


5. Everything this reviewer ran

All from /home/swung/w7/p2-fix-aba. Logs under ~/w7/logs/rv/ (deleted at the end of this round except where cited).

5.1 The four states (claim 1), reproduced without ctest so the knob can be removed

BIN=build-push/MobileGL/MG_IntegrationTest/MobileGLIntegrationTest
F=--gtest_filter=HandleRecycleScenario.AVertexArrayAtARecycledAddressDoesNotInheritItsPredecessorsVertexInput
COMMON="MOBILEGL_BACKEND_TYPE=DirectVulkan MGITEST_HANDLE_ARM=aba MGITEST_PIPE_PUSH_BUILD=1 \
        MGITEST_HANDLE_ABA_IMPLEMENTED=1 \
        __EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json \
        VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/lvp_icd.json \
        MOBILEGL_MAGMA_FIX_ITERATIONRP_SUBGROUP_SCRATCH=1 MOBILEGL_MAGMA_DERIVE_NUM_SUBGROUPS=1 \
        MOBILEGL_MAGMA_ITERATIONRP_FIX_BARRIER=1"

env $COMMON MOBILEGL_PIPE_PUSH=0        MOBILEGL_PIPE_HANDLE_ABA_CONTROL=1 $BIN $F   # A
env $COMMON MOBILEGL_PIPE_PUSH=0                                           $BIN $F   # B
env $COMMON MOBILEGL_PIPE_LEGACY_MEMOS=0 MOBILEGL_PIPE_HANDLE_ABA_CONTROL=1 $BIN $F  # C
env $COMMON MOBILEGL_PIPE_LEGACY_MEMOS=0                                    $BIN $F  # D
arm knob printed line result
A pre-handle ON expected=STALE observed=STALE [ PASSED ] 1 test.
B pre-handle OFF expected=STALE observed=FRESH [ FAILED ], rc 1
C handle ON expected=STALE observed=STALE [ PASSED ] 1 test.
D handle OFF expected=STALE observed=FRESH [ FAILED ], rc 1

(D also proves the handle arm is genuinely entered: with MOBILEGL_PIPE_LEGACY_MEMOS=0 a clear Track-H bit 6 would have been a startup Fatal{PipeLegacyMemosDisabled} (MagmaPipeArms.h:108-116), so the process running at all means bit 6 is set and MagmaPipeTrackHArmIsHandles is true.)

5.2 The knob's blast radius (claim 2)

grep -rn "PipeHandleAbaControl\|MagmaPipeAbaControlDefeatsIdentity\|kMagmaPipeAbaControlSlotIndex" over the tree gives exactly four executable consumers, all #if MOBILEGL_PIPE_PUSH: VertexInputStateFactory.cpp:67 (bufferKey = 0), :101 (MemosFor returns entry 0), VulkanRenderer.cpp:3653 (aliased VaoDrawMemo), and the vestigial :3737 (minor 1). Plus Config.h:371 (the field), ConfigLoader.cpp:267 (the parse), and comments.

Untouched by the diff, and verified by reading: TryBindResolvedVertexBindings' frame-serial gate (VulkanRenderer.cpp:3635-3644), its manager-wide slice-epoch compare (:3602), its per-binding live-buffer compare (:3611-3613), its per-resource slice-epoch compare (:3624-3628), its host-map check (entry.anyBufferMapped, :3602) and SyncPersistentMappedRange (:3621); SetupDrawSnapshot's own identity compare (:6360) is also untouched. m_vaoMemos[0] is safe on an empty table: MagmaPipeSlotTable::operator[] grows on demand (MagmaPipeArms.h:430-436).

Pull build (G1), --threshold 0:

$ python3 scripts/symbol_report.py --before ~/w7/p2-before-libMobileGL.so \
      --after build-linux/libMobileGL.so --threshold 0
.text 10792579 -> 10792739 (+160, +0.001%)
27799 -> 27799 defined symbols: 0 added, 0 removed, 4 resized, 0 renamed
  RenderState::RenderState()                1700 -> 1848   (+148)
  RenderState::SetCapability(...)            850 ->  927   (+77)
  RenderState::IsCapabilityEnabled(...)      239 ->  268   (+29)
  _GLOBAL__sub_I_DirectGLES.cpp             1340 -> 1331   (-9)

Exactly ID-3's four, nothing else — reproduced independently of the implementer's run.

Push build, knob off:

$ ctest --test-dir build-push -L integration-gpu -R DirectVulkan --no-tests=error -j 1
rc=0    100% tests passed, 0 tests failed out of 455

5.3 Determinism (claim 3)

$ for i in $(seq 1 10); do ctest --test-dir build-push -R 'HandleRecycle\.AbaControl' \
      --no-tests=error -j 1; done
run 1..10: rc=0 | 100% tests passed, 0 tests failed out of 8      # 2 lanes x 4 entries, zero flips

Sampled with -V: expected=STALE observed=STALE on both 2468 (pre-handle) and 2472 (AbaControlHandles) every time.

5.4 Lane registration and G2 (claim 4)

$ for d in build-linux build-push; do ctest --test-dir $d -N | grep -E '^[[:space:]]*Test[[:space:]]+#[0-9]+:' \
      | sed -E 's/^ *Test +#[0-9]+: //' | LC_ALL=C sort > names-$d.txt; done
$ diff names-build-linux.txt names-build-push.txt    # empty; 2478 entries each
$ grep -c HandleRecycle names-build-linux.txt        # 32, including all four AbaControlHandles entries
$ ctest --test-dir build-linux -R HandleRecycle --no-tests=error -j 4
rc=0    100% tests passed, 0 tests failed out of 32   (36 skip lines; the push arms skip)

The pull build's skip reason is the designed one (HandleRecycleScenario.cpp:329): "the AbaControl arm needs a library built with MOBILEGL_PIPE_PUSH, and this one was not …". CI needs no edit: the G8 step selects by regex (.github/workflows/test.yml:600, -R 'HandleRecycle|CsoContentAddressing'), so the new lane is picked up automatically.

5.5 build-verify (claim 5) — configured and built by this reviewer

$ cmake -S . -B build-verify -G Ninja ... -DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON
$ cmake --build build-verify -j 12                                   # rc 0
$ ctest --test-dir build-verify -R HandleRecycle --no-tests=error -j 4 --output-on-failure
rc=0    100% tests passed, 0 tests failed out of 40                   # G8
$ ctest --test-dir build-verify -L unit --no-tests=error -j 8
rc=0    100% tests passed, 0 tests failed out of 1562
$ for i in 1 2 3; do ctest --test-dir build-verify -L integration-gpu \
      -R 'HandleRecycle|CsoContentAddressing' --no-tests=error -j 4; done   # the CI step's shape
run 1..3: 100% tests passed, 0 tests failed out of 48

The configure log confirms both probes arm: "DirectVulkan's vertex input is keyed on {slot, gen}" and "MOBILEGL_PIPE_HANDLE_ABA_CONTROL has a consumer" (both now resolve to MagmaPipeArms.h).

5.6 P1's ScopedPipeVerb / no-verb concern

HandleRecycleScenario.cpp includes only <Harness/HeadlessGL.h>, <Harness/ScenarioFixture.h>, <GL/gl.h>, <GL/glcorearb.h> and five standard headers (:106-123); every operation in the three cases is a gl* entry point or a harness helper that is itself public GL. No backend type, no MG_Pipe symbol, no verb is named. Nothing regressed here, and the build-verify build compiles the comparator alongside without a divergence (§5.5) — see minor 6 for what that does and does not prove.

5.7 The -j 8 arming-lane race (result file V5.6) — independently corroborated as pre-existing

$ for i in 1 2 3; do ctest --test-dir build-push -R 'PrimGenReroute\.|PointSizeDemotion\.' \
      --no-tests=error -j 8; done
run 1: 89% tests passed, 2 tests failed out of 18
run 2: 89% tests passed, 2 tests failed out of 18
run 3: 94% tests passed, 1 tests failed out of 18

No HandleRecycle entry is in that selection, and the same lane is 455/455 at -j 1. The diagnosis in V5.6 (a shared fopen(path,"w") log truncated by lane-mates) stands, and the hazard is not this change's.


6. slotprobe.cpp (the MAJOR 1 probe, standalone, outside the tree)

#include <cstdio>
#include "MagmaPipeArms.h"
namespace MobileGL::MG_Util::Debug { void Log(const char*, int, const char*, ...) {} }  // link stub
using namespace MobileGL::MG_Backend::DirectVulkan;
using MobileGL::MG_Pipe::MGPipeHandle;
static void show(const char* what, MGPipeHandle h) {
    std::printf("  %-28s slot=%u gen=%u\n", what, h.Slot, h.Gen);
}
int main() {
    std::printf("A) HandleRecycleScenario's shape (5 frame boundaries, then the replacement):\n");
    MagmaPipeIdentityTable t("VertexElementsCso");
    show("primerVao (lifetimeId 1)", t.Acquire(1));
    show("redVao    (lifetimeId 2)", t.Acquire(2));
    for (int i = 0; i < 5; ++i) t.OnFrameBoundary();      // primer + 3 warm-ups + the ABA frame
    show("greenVao  (lifetimeId 3)", t.Acquire(3));       // redVao died here; the table is not told
    std::printf("B) what a slot REUSE looks like (the ABA the generation exists to stop):\n");
    MagmaPipeIdentityTable t2("VertexElementsCso");
    show("first     (lifetimeId 1)", t2.Acquire(1));
    for (int i = 0; i < 1281; ++i) t2.OnFrameBoundary();  // > kRetireAgeBoundaries, landing on a sweep
    show("second    (lifetimeId 2)", t2.Acquire(2));
    return 0;
}

Build flags are the -D/-I set of VertexInputStateFactory.cpp from build-push/compile_commands.json, plus -std=c++23 and -I MobileGL/MG_Backend/DirectVulkan/Renderer.