[Feature] (P14): key the backend twin tables by session and share group

The last piece of state ownership this phase owed: the eight DirectGLES twin
tables resolved a handle process-wide, so two contexts of one session that do
NOT share a share group - which have independent name spaces, so the same
{slot, gen} in both is legal - were handed the SAME twin, driver ids,
persistent map and pooled store included. Two contexts that DO share a group
were served by whichever of them twinned first. Design and per-slice shape:
docs/Disaggregated/design/11-state-ownership.md (S6).

Key: DirectGLES.h's TwinKey{SessionKey, ShareGroupKey} - the same two words
the applier's object records are filed under, with the group derived by
MGPipeApplierShareGroupKeyFor so a record and the twin describing it can never
land in different buckets. Session is in the key because a twin names a driver
id and no driver id crosses a connection.

Resolution: SessionRuntime.cpp's ResolveThreadTwinKey, installed by
InstallSessionRuntimeHooks beside the S2 PipeInputs probe, the S4 native
tuple key and the S5 applier key. It reads the ThreadSessionScope's session
and CurrentContextToken() - apply-thread state only, never MG_State, so the
frontend-keyed-lookup prohibition on the apply thread still holds. A thread
that belongs to no session answers {0, 0}, which is the process singleton this
replaces; the monolith and every unit case are unchanged by construction.

Shape: SlotTables.h's SlotTableRegistry holds one SharedPtr bucket per key,
with a thread-local last-key memo. It forwards every handle-keyed entry -
GetOrCreate, FindByHandle, LiveGenAt, HandleOf, StateForHandle,
NoteStateForHandle, ReleaseByHandle, ForEachLive, the counters - to the
CALLING THREAD's bucket, so the ~150 call sites in Managers.cpp and
DirectGLES.cpp are literally unchanged. The holder list stays process-wide: a
death notice is about an object, not a bucket, and each shared holder is one
bucket. ForEachLive now walks the calling thread's bucket (what a context can
see is its own group); the death and diagnostic walks take the new
ForEachBucket.

DropEveryTwinForEndedServerSession is session-scoped: it runs on the apply
thread of the session that is ending, while S2 lets neighbour sessions keep
rendering in the same process. No session on the thread keeps the old
process-wide drop verbatim. Managers.h's "a texture is owned by exactly one
context (share groups are not implemented)" is corrected: a texture belongs to
a share group and its sibling contexts legitimately mutate it.

Tests: MG_Test/Wire/TwinOwnershipTest.cpp, five cases driving the real twin
tables with no EGL and no wire - the same handle in two nonshared groups is
two twins and neither overwrites the other; two contexts of one group resolve
ONE twin; a per-session drop takes only that session's; a thread with no
session still gets the process-wide table; and the six-registry template
repeats the first three. Negative control run and reverted: making the
registry answer a single {0, 0} bucket turns four of the five red.

Full host ctest against a stashed baseline, compared by test name: 0 new
failures. The one name in the difference set
(LogForwardChannel.APeerThatStopsReadingCostsWarnLinesNotTheCallersTime) is a
pre-existing flake, failing on the baseline too with the identical assertion.
The 103-case acceptance set plus the five new cases is 108/108 green.

Remaining debts (documented): the unit shadows (g_activeTextureUnit,
g_boundTexturesCache, g_boundSamplersCache) are still process-global; the
driver-level glIsBuffer probe is still not usable on this host, so sharing is
proven at the twin layer (the twin IS the object holding the driver id); the
backend context/sync generations are still process-wide; the per-record
ContextSlot is still S1's edge declaration.
This commit is contained in:
2026-10-01 20:11:53 -04:00
parent 5c6399e240
commit ff2120616d
11 changed files with 831 additions and 30 deletions
+267
View File
@@ -11,6 +11,12 @@
#include <MG_Pipe/MGPipeHandles.h>
// P14 S6: the key this table is filed under includes the calling thread's {session, share
// group}, and the probe that answers it lives on the backend's public surface beside S4's
// native-tuple key (DirectGLES.h). Included here so this header is self-contained for the two
// translation units that include it without copying Managers.h's include order.
#include "DirectGLES.h"
#if MOBILEGL_PIPE_PUSH
#include <MG_Impl/Pipe/SlotAllocator.h>
#endif
@@ -155,6 +161,249 @@ namespace MobileGL::MG_Backend::DirectGLES {
return enabled;
}
// =====================================================================================
// P14 S6 (docs/Disaggregated/design/11-state-ownership.md): THE TWIN TABLE IS PER
// {session, share group}, NOT PER PROCESS.
// =====================================================================================
//
// WHAT WAS WRONG, in one sentence: a twin is a DRIVER object - a GL buffer id, a program, a
// texture, the storage behind them - so it belongs to one native context's object namespace,
// and a table keyed by {slot, gen} ALONE resolves the same handle value to the same twin
// whichever context asked. The frontend gives two contexts that do NOT share a group two
// independent name spaces, so the same {slot, gen} in both is legal and means two different
// objects; one process-wide table made the second context adopt the first's twin (including
// its driver ids, its persistent map and its pooled store). At the same time two contexts of
// ONE group - which must see one set of objects - were served by whichever of them twinned
// first, i.e. by luck rather than by the group the two agree on.
//
// THE KEY IS the same pair the applier files its OBJECT RECORDS under: {session, share
// group}. Same probe, same `0 means this context started its own group` rule (see
// DirectGLES.h's TwinKey and SessionRuntime.cpp's ResolveThreadTwinKey), so a record and the
// twin it describes are always in the same bucket. Session is in the key because a twin
// names a driver id and no driver id crosses a connection - two connections in one process
// are two clients, and nothing may be shared with either from the other.
//
// THE HOLDER LIST STAYS PROCESS-WIDE AND THAT IS NOT A COMPROMISE. A notice is about an
// OBJECT, not about a bucket: OnFrontendObjectDestroyed and ReleaseTwinByHandle must reach
// the by-value copies a fixture or a context reset is holding, whichever bucket they sit in,
// and "which bucket holds it" is exactly the question that produced the two-holder leak the
// list exists to fix. Each SHARED holder of the list is ONE bucket; the walk therefore covers
// every bucket of every table.
//
// THE SAME TABLE OBJECT IS REUSED, NOT REPLACED, and that is what keeps ~200 call sites
// alive: `g_backendTextureObjects.Find(...)` still answers from the bucket the calling thread
// is in, because the object the global names multiplexes on the key at each lookup. A caller
// that must address a bucket OTHER than the calling thread's - the death walks, the
// session-teardown drop - takes Shared() or calls the static entries, which walk buckets;
// and a caller that wants the whole process's view (the teardown, the tests) takes Snapshot().
// FORWARD-DECLARED, because the registry below is DEFINED before the table it holds; the
// members that name the table are defined out of line at the bottom of this header.
template <typename StateObject, typename BackendObject, MG_Pipe::MGPipeKind kKind>
class BackendSlotTable;
template <typename StateObject, typename BackendObject, MG_Pipe::MGPipeKind kKind>
class SlotTableRegistry {
public:
using Table = BackendSlotTable<StateObject, BackendObject, kKind>;
using Key = TwinKey;
using StatePtr = SharedPtr<StateObject>;
using BackendPtr = SharedPtr<BackendObject>;
// The bound the two tables that ARE a bucket registry have always spelled by name; the
// six registries reach it through their own `SlotTable` alias.
static constexpr Uint32 kMaxHandleSlot = 1u << 20;
// One bucket. SHARED, so a caller may keep it across other lookups: the map owns the
// pointee and never moves it.
SharedPtr<Table> Shared(const Key& key) const {
// The common case on the hot path: the same key as last time. One compare, no probe.
if (m_memoTable != nullptr && m_memoKey == key) return m_memoTable;
const auto found = m_buckets.find(key);
if (found != m_buckets.end()) {
m_memoKey = key;
m_memoTable = found->second;
return found->second;
}
SharedPtr<Table> fresh = MakeShared<Table>();
m_buckets.emplace(key, fresh);
m_memoKey = key;
m_memoTable = fresh;
return fresh;
}
// The calling thread's bucket - the ONE entry every in-place call site below resolves
// through, and the process-wide answer on a thread that belongs to no session.
//
// CONST, AND IT ANSWERS A MUTABLE BUCKET: the index is `mutable` because a LOOKUP may
// have to create the bucket it is about to miss in (a handle with no twin yet is the
// ordinary case), and the bucket's own entries are the table's business rather than the
// registry's. The alternative - a const overload returning "no table" - would make every
// read-only caller handle a case that cannot arise.
Table& ForCallingThread() const {
Key key;
CurrentTwinKey(&key);
return *Shared(key);
}
// Every bucket, in no particular order.
Vector<SharedPtr<Table>> Snapshot() const {
Vector<SharedPtr<Table>> all;
all.reserve(m_buckets.size());
for (const auto& bucket : m_buckets) all.push_back(bucket.second);
return all;
}
SizeT BucketCount() const { return m_buckets.size(); }
// THE PER-BUCKET ONE-SHOT, because a bucket held by SharedPtr is never dropped by an
// assignment: a caller that means "this context/group is gone" has to be able to say so,
// and the key is how it says which one. The bucket LEAVES the index before its twins are
// released, so a twin destructor that re-entered Shared() cannot find a half-erased one.
Bool Drop(const Key& key) {
const auto found = m_buckets.find(key);
if (found == m_buckets.end()) return false;
SharedPtr<Table> dropped = found->second;
m_buckets.erase(found);
if (m_memoTable == dropped) {
m_memoKey = Key{};
m_memoTable = nullptr;
}
dropped->Reset();
return true;
}
// EVERY bucket of one session, for the teardown of a session that owned several groups.
// The session is the key's high half, so this is one pass and no per-group bookkeeping -
// and it is the difference between "this session's twins go" and "the process's do".
Uint32 DropSession(Uint64 sessionKey) {
Uint32 dropped = 0;
for (auto bucket = m_buckets.begin(); bucket != m_buckets.end();) {
if (bucket->first.SessionKey != sessionKey) {
++bucket;
continue;
}
SharedPtr<Table> table = bucket->second;
if (m_memoTable == table) {
m_memoKey = Key{};
m_memoTable = nullptr;
}
bucket = m_buckets.erase(bucket);
table->Reset();
++dropped;
}
return dropped;
}
// Replace this registry with an empty one, releasing every bucket's twins. This is what
// `registry = {}` meant when the registry WAS one table; a caller that must keep the
// tables alive past the call takes a Snapshot() first.
void Reset() {
const Vector<SharedPtr<Table>> all = Snapshot();
m_buckets.clear();
m_memoKey = Key{};
m_memoTable = nullptr;
for (const SharedPtr<Table>& table : all) table->Reset();
}
// ---- the calling thread's bucket, member for member ---------------------------------
//
// THIS IS WHY THE SHAPE COST THE CALL SITES NOTHING. Every handle-keyed entry the table
// has is repeated here and forwarded to the bucket the calling thread resolves to, so
// `g_backendTextureObjects.FindByHandle(h)` and `g_backendBufferResources.GetOrCreate(h)`
// - unchanged source - answer from the right context's object namespace. A caller that
// must name a bucket OTHER than its own (a death walk, a case) takes Shared()/Snapshot()
// and calls the table directly.
BackendPtr& GetOrCreate(MG_Pipe::MGPipeHandle handle) {
return ForCallingThread().GetOrCreate(handle);
}
BackendPtr* FindByHandle(MG_Pipe::MGPipeHandle handle) {
return ForCallingThread().FindByHandle(handle);
}
const BackendPtr* FindByHandle(MG_Pipe::MGPipeHandle handle) const {
return ForCallingThread().FindByHandle(handle);
}
// Resolve-or-create in a bucket the caller NAMES rather than the one it is in. The
// monolith-glue frontend-keyed entry (StateBackendObjectRegistry::GetOrCreate below)
// needs it: a state object handed over on the client side of a sync is described by a
// record in the CALLING thread's group, and a twin built for it must land in the same
// bucket as that record.
BackendPtr& GetOrCreateIn(const Key& key, MG_Pipe::MGPipeHandle handle) {
return Shared(key)->GetOrCreate(handle);
}
Uint32 LiveGenAt(Uint32 slot) const { return ForCallingThread().LiveGenAt(slot); }
MG_Pipe::MGPipeHandle HandleOf(const StateObject* stateObj) const {
return ForCallingThread().HandleOf(stateObj);
}
void NoteStateForHandle(MG_Pipe::MGPipeHandle handle, const StatePtr& stateObj) {
ForCallingThread().NoteStateForHandle(handle, stateObj);
}
StatePtr StateForHandle(MG_Pipe::MGPipeHandle handle) const {
return ForCallingThread().StateForHandle(handle);
}
template <typename Fn>
void ForEachLive(Fn&& fn) const {
// THE CALLING THREAD'S BUCKET, not every bucket, and that is the semantic answer
// rather than a shortcut: a caller walking twins is asking "which of the objects I
// can see need this", and what it can see is its own share group. Every bucket in
// one process would walk a neighbour session's objects.
ForCallingThread().ForEachLive(fn);
}
// Every bucket, in one pass. The caller receives the key with the handle so it can tell
// which context a twin belongs to - the only walks that need the whole process's view
// are a diagnostic and a teardown.
template <typename Fn>
void ForEachBucket(Fn&& fn) const {
for (const auto& bucket : m_buckets) {
bucket.second->ForEachLive(
[&](MG_Pipe::MGPipeHandle handle, const BackendPtr& twin) { fn(bucket.first, handle, twin); });
}
}
// The process-wide answer for the two statics: a notice is about an OBJECT, so it must
// reach the twin in EVERY bucket of EVERY holder. They forward to the table's own
// statics, which walk the holder list - each SHARED holder of it is one bucket.
static Bool ReleaseTwinByHandle(MG_Pipe::MGPipeHandle handle) {
return Table::ReleaseTwinByHandle(handle);
}
static Bool OnFrontendObjectDestroyed(Uint64 lifetimeId) {
return Table::OnFrontendObjectDestroyed(lifetimeId);
}
// The instance spelling, and it HANDS THE TWIN OUT rather than answering a Bool: the two
// tables that ARE a bucket registry (buffers, sampler views) have always been called as
// `g_backendBufferResources.ReleaseByHandle(res)` and their callers still have to decide
// what happens to the driver id the twin owns. The bucket is the CALLING THREAD'S, which
// is the bucket that twin was created in.
BackendPtr ReleaseByHandle(MG_Pipe::MGPipeHandle handle) {
return ForCallingThread().ReleaseByHandle(handle);
}
static Uint32 HolderCount() { return Table::HolderCount(); }
Uint32 LiveCount() const { return ForCallingThread().LiveCount(); }
Uint32 CompositeLiveCount() const { return ForCallingThread().CompositeLiveCount(); }
SizeT OrdinaryCapacityForTest() const { return ForCallingThread().OrdinaryCapacityForTest(); }
SizeT CompositeCapacityForTest() const { return ForCallingThread().CompositeCapacityForTest(); }
// ---- the bucket index's own controls, for the two tables that ARE a registry --------
//
// The six TwinRegistry registries carry the same three under the names in Managers.h
// (DropBucket / DropBucketsOfSession / ResetBuckets); they are repeated here because
// SlotTableRegistry IS the public type of the buffer and sampler-view tables, so a caller
// of those has no registry object in between.
Bool DropBucket(const Key& key) { return Drop(key); }
// Drop every bucket of one session; answers how many groups that took with it.
Uint32 DropBucketsOfSession(Uint64 sessionKey) { return DropSession(sessionKey); }
// Empty every bucket (the process-shaped world's teardown, key {0, 0}).
void ResetBuckets() { Reset(); }
private:
// ONE BUCKET PER LIVE {session, share group}, held by SHARED pointer so a caller may
// snapshot one across other lookups and so `Drop` can take a bucket's twins away
// explicitly - an embedded table could do neither. `std::unordered_map` rather than the
// tree's own flat Vector: this map is probed once per twin lookup on the draw path.
mutable std::unordered_map<Key, SharedPtr<Table>, TwinKeyHash> m_buckets;
// Hot-path memo. A registry is per KIND and each kind is touched several times per draw;
// a thread only ever moves between buckets when its session binds another context.
mutable SharedPtr<Table> m_memoTable;
mutable Key m_memoKey;
};
template <typename StateObject, typename BackendObject, MG_Pipe::MGPipeKind kKind>
class BackendSlotTable {
public:
@@ -616,6 +865,24 @@ namespace MobileGL::MG_Backend::DirectGLES {
}
}
// Empty this table, releasing every twin it holds. THE SPACES ARE MOVED OUT FIRST and
// that is the whole of the care taken here: a twin's destructor is a driver call that
// could re-enter GetOrCreate on this table (SlotTables.h's ReleaseTwinAt says why), and
// destroying the entries in place would leave that re-entry resizing the vector the
// destructors are being walked out of. Here the table is already empty when the entries
// die, so a re-entrant insert grows a fresh vector.
//
// This is what `table = {}` used to mean, spelled so a caller cannot miss the ordering.
void Reset() {
Vector<Entry> slots = std::move(m_slots);
Vector<Entry> band = std::move(m_band);
BackendPtr nullTwin = std::move(m_nullTwin);
ForgetHandle();
slots.clear();
band.clear();
nullTwin.reset();
}
Uint32 LiveCount() const {
Uint32 count = 0;
for (const Entry& entry : m_slots) {