[MTRDevice] Defer first Thread subscription until operational advertisement seen

On a cold start where the Home app foregrounds before the Thread interface
has finished coming up, _delegateAdded -> _ensureSubscriptionForExistingDelegates
-> _setupSubscriptionWithReason fires the very first SubscribeRequest before
threadradiod's [WED_START] completes.  The send fails immediately with
"OS Error 0x02000041: No route to host" (errno 65) from
UDPEndPointImplSockets.cpp.  We then back off and retry, wasting 3-6 seconds
of coldstart on a doomed attempt.

Gate the first subscription attempt: if _deviceUsesThread is true and we have
no cached _lastSubscriptionIPAddress, defer scheduling the pool work until
either (a) nodeMayBeAdvertisingOperational fires for this device -- the strong
signal that the Thread interface is up and the accessory is reachable -- or
(b) a 1-second watchdog elapses -- the safety net for devices that never
advertise during the deferral window.

Implementation:

  - New deferringFirstThreadSubscription / hasDeferredFirstThreadSubscription
    properties on MTRDevice_Concrete.  The first is YES only during the
    deferral window; the second is a one-shot latch so subsequent re-entries
    of _ensureSubscriptionForExistingDelegates (notably the one
    nodeMayBeAdvertisingOperational triggers after clearing the deferral)
    don't re-enter the deferral path.

  - _ensureSubscriptionForExistingDelegates factors the existing pool-work
    scheduling out into a local block scheduleSubscriptionPoolWork that the
    deferral watchdog reuses verbatim.

  - nodeMayBeAdvertisingOperational, when it observes a pending first-Thread
    deferral, clears the flag and re-enters _ensureSubscriptionForExistingDelegates
    with reason "operational advertisement seen" instead of going through the
    normal _triggerResubscribeWithReason path (which would no-op because we
    aren't yet in MTRInternalDeviceStateSubscribing).

  - 1s watchdog runs on the device queue, takes _lock, double-checks the
    deferral flag (the advertisement path may have cleared it racingly), and
    if still set proceeds with the original scheduling.

Only Thread devices are affected; Wi-Fi/Ethernet devices retain the existing
immediate behavior.  In the worst case (device that never advertises), the
first subscribe is delayed by exactly 1s vs. today's "fire immediately, fail
in 3-6s, then retry" behavior -- still a substantial coldstart win.

Adds behavior tests covering the deferral, the advertisement-clears-deferral
path, the 1s watchdog, and the invalidate-during-deferral safety path.

rdar://168903097

Author: woody-apple

src/darwin/Framework/CHIP/MTRDevice_Concrete.mm

@@ -345,6 +345,21 @@ @interface MTRDevice_Concrete ()
 @property (nonatomic) MTRInternalDeviceState internalDeviceState;
 @property (nonatomic) BOOL doingCASEAttemptForDeviceMayBeReachable;
 
+// Set to YES while we are waiting to fire our very first subscribe attempt on
+// a Thread device for which we have no cached IP address.  The actual subscribe
+// fires when either (a) an operational mDNS advertisement arrives for this
+// device, or (b) a 1-second watchdog elapses, whichever comes first.  This
+// avoids racing the very first SubscribeRequest against threadradiod
+// [WED_START], which would otherwise produce a No-route-to-host failure and
+// burn 3-6 seconds of coldstart on a retry.
+@property (nonatomic) BOOL deferringFirstThreadSubscription;
+
+// Whether we have already gone through the first-Thread-subscribe deferral
+// path for this MTRDevice instance.  Used so we only defer once: subsequent
+// re-entries (e.g. the one nodeMayBeAdvertisingOperational triggers after
+// clearing the deferral) must not defer again.
+@property (nonatomic) BOOL hasDeferredFirstThreadSubscription;
+
 #define MTRDEVICE_SUBSCRIPTION_ATTEMPT_MIN_WAIT_SECONDS (1)
 #define MTRDEVICE_SUBSCRIPTION_ATTEMPT_MAX_WAIT_SECONDS (3600)
 @property (nonatomic) uint32_t lastSubscriptionAttemptWait;
@@ -1067,27 +1082,109 @@ - (void)_ensureSubscriptionForExistingDelegates:(NSString *)reason
         mtr_weakify(self);
         if ([self _deviceUsesThread]) {
             MTR_LOG(" => %@ - device is a thread device, scheduling in pool", self);
+
+            // If a first-Thread-subscribe deferral is already in flight, do
+            // nothing on re-entry.  The watchdog (or operational advertisement)
+            // that originally armed the gate is still on the hook to schedule
+            // the subscribe; coming back through here during the 1s window
+            // would otherwise either skip the gate (because
+            // hasDeferredFirstThreadSubscription is already YES, so
+            // shouldDeferForThreadColdstart evaluates to NO) and fall through
+            // to scheduleSubscriptionPoolWork(), bypassing the deferral.
+            if (_deferringFirstThreadSubscription) {
+                MTR_LOG("%@ - first Thread subscribe deferral still in flight; ignoring re-entry (%@)", self, reason);
+                return;
+            }
+
             NSString * description = [NSString stringWithFormat:@"MTRDevice setDelegate first subscription / controller resume (%p)", self];
-            [self _scheduleSubscriptionPoolWork:^{
+            void (^scheduleSubscriptionPoolWork)(void) = ^{
                 mtr_strongify(self);
-                VerifyOrReturn(self, MTR_LOG_DEBUG("_ensureSubscriptionForExistingDelegates _scheduleSubscriptionPoolWork called back with nil MTRDevice"));
-
-                [self->_deviceController asyncDispatchToMatterQueue:^{
+                if (!self) {
+                    return;
+                }
+                // _scheduleSubscriptionPoolWork: requires _lock be held by the
+                // caller.  Most callers below already do this; the watchdog
+                // path takes the lock explicitly before invoking this block.
+                os_unfair_lock_assert_owner(&self->_lock);
+                [self _scheduleSubscriptionPoolWork:^{
                     mtr_strongify(self);
-                    VerifyOrReturn(self, MTR_LOG_DEBUG("_ensureSubscriptionForExistingDelegates asyncDispatchToMatterQueue called back with nil MTRDevice"));
+                    VerifyOrReturn(self, MTR_LOG_DEBUG("_ensureSubscriptionForExistingDelegates _scheduleSubscriptionPoolWork called back with nil MTRDevice"));
+
+                    [self->_deviceController asyncDispatchToMatterQueue:^{
+                        mtr_strongify(self);
+                        VerifyOrReturn(self, MTR_LOG_DEBUG("_ensureSubscriptionForExistingDelegates asyncDispatchToMatterQueue called back with nil MTRDevice"));
+
+                        std::lock_guard lock(self->_lock);
+                        [self _setupSubscriptionWithReason:[NSString stringWithFormat:@"%@ and scheduled subscription is happening", reason]];
+                    } errorHandler:^(NSError * _Nonnull error) {
+                        mtr_strongify(self);
+                        VerifyOrReturn(self, MTR_LOG_DEBUG("_ensureSubscriptionForExistingDelegates asyncDispatchToMatterQueue errored with nil MTRDevice"));
+
+                        // If controller is not running, clear work item from the subscription queue
+                        MTR_LOG_ERROR("%@ could not dispatch to matter queue for resubscription - error %@", self, error);
+                        std::lock_guard lock(self->_lock);
+                        [self _clearSubscriptionPoolWork];
+                    }];
+                } inNanoseconds:0 description:description];
+            };
 
-                    std::lock_guard lock(self->_lock);
-                    [self _setupSubscriptionWithReason:[NSString stringWithFormat:@"%@ and scheduled subscription is happening", reason]];
-                } errorHandler:^(NSError * _Nonnull error) {
+            // Coldstart gate: if this is our first subscription attempt for a
+            // Thread device and we have no cached IP address yet, the Thread
+            // interface is very likely still coming up.  Firing immediately
+            // would produce a No-route-to-host (errno 65) from the first
+            // SubscribeRequest before threadradiod's [WED_START] completes,
+            // and we'd burn 3-6s of coldstart on a wasted attempt.  Defer
+            // until either an operational advertisement arrives for this
+            // device (the strong signal that the interface is up) or a 1s
+            // watchdog elapses (the safety net), whichever comes first.
+            //
+            // Short-circuit the gate if we have evidence the device is
+            // already reachable: a recent report (within the last 60s) means
+            // the Thread interface is up and the device is talking to us, so
+            // there is no benefit to paying the 1s watchdog penalty (this is
+            // the warm-Thread cold-app-launch path).
+            NSDate * lastReportTime = [_mostRecentReportTimes lastObject];
+            BOOL hasRecentReport = (lastReportTime != nil
+                && -[lastReportTime timeIntervalSinceNow] < 60.0);
+            BOOL shouldDeferForThreadColdstart = (!_lastSubscriptionIPAddress.has_value()
+                && !hasRecentReport
+                && !_deferringFirstThreadSubscription
+                && !_hasDeferredFirstThreadSubscription);
+            if (shouldDeferForThreadColdstart) {
+                MTR_LOG("%@ - deferring first Thread subscribe up to 1s; waiting for operational advertisement", self);
+                _deferringFirstThreadSubscription = YES;
+                _hasDeferredFirstThreadSubscription = YES;
+
+                static constexpr int64_t kFirstThreadSubscribeWatchdogNs = static_cast<int64_t>(NSEC_PER_SEC);
+                dispatch_after(dispatch_time(DISPATCH_TIME_NOW, kFirstThreadSubscribeWatchdogNs), self.queue, ^{
                     mtr_strongify(self);
-                    VerifyOrReturn(self, MTR_LOG_DEBUG("_ensureSubscriptionForExistingDelegates asyncDispatchToMatterQueue errored with nil MTRDevice"));
-
-                    // If controller is not running, clear work item from the subscription queue
-                    MTR_LOG_ERROR("%@ could not dispatch to matter queue for resubscription - error %@", self, error);
+                    if (!self) {
+                        return;
+                    }
                     std::lock_guard lock(self->_lock);
-                    [self _clearSubscriptionPoolWork];
-                }];
-            } inNanoseconds:0 description:description];
+                    if (!self->_deferringFirstThreadSubscription) {
+                        // Operational advertisement already triggered the subscribe.
+                        return;
+                    }
+                    self->_deferringFirstThreadSubscription = NO;
+                    // Defensive: invalidate / controllerSuspended could have
+                    // landed during the 1s window and torn down the
+                    // subscription path even though they also clear
+                    // deferringFirstThreadSubscription (so the early-return
+                    // above usually catches this).  Belt-and-suspenders: re-
+                    // check that subscriptions are still allowed before
+                    // scheduling the pool work.
+                    if (![self _subscriptionsAllowed]) {
+                        MTR_LOG("%@ - first Thread subscribe watchdog: subscriptions no longer allowed; aborting", self);
+                        return;
+                    }
+                    MTR_LOG("%@ - first Thread subscribe watchdog fired; proceeding without operational advertisement", self);
+                    scheduleSubscriptionPoolWork();
+                });
+                return;
+            }
+
+            scheduleSubscriptionPoolWork();
         } else {
             [_deviceController asyncDispatchToMatterQueue:^{
                 mtr_strongify(self);
@@ -1119,6 +1216,13 @@ - (void)invalidate
     // attempt, since we now have no delegate.
     _reattemptingSubscription = NO;
 
+    // Clear the first-Thread-subscribe deferral state under _lock so that any
+    // already-armed watchdog short-circuits at its early-return when it
+    // observes deferringFirstThreadSubscription == NO, and so a future
+    // re-attach starts from a clean slate.
+    _deferringFirstThreadSubscription = NO;
+    _hasDeferredFirstThreadSubscription = NO;
+
     // Clear subscription pool work item if it's in progress, to avoid forever
     // taking up a slot in the controller's work queue.
     [self _clearSubscriptionPoolWork];
@@ -1152,6 +1256,23 @@ - (void)nodeMayBeAdvertisingOperational
 
     MTR_LOG("%@ saw new operational advertisement", self);
 
+    // If we were deferring our very first Thread subscribe waiting for exactly
+    // this signal, clear the deferral flag and trigger the subscribe via the
+    // normal path immediately rather than waiting for the 1s watchdog.
+    {
+        std::lock_guard lock(_lock);
+        if (_deferringFirstThreadSubscription) {
+            _deferringFirstThreadSubscription = NO;
+            MTR_LOG("%@ - operational advertisement cleared first-Thread-subscribe deferral", self);
+            [self _ensureSubscriptionForExistingDelegates:@"operational advertisement seen"];
+            return;
+        }
+            MTR_LOG("%@ - operational advertisement cleared first-Thread-subscribe deferral", self);
+            [self _ensureSubscriptionForExistingDelegates:@"operational advertisement seen"];
+            return;
+        }
+    }
+
     [self _triggerResubscribeWithReason:@"operational advertisement seen"
                     nodeLikelyReachable:YES];
 }
@@ -5164,6 +5285,14 @@ - (void)controllerSuspended
     self.suspended = YES;
     [self _resetSubscriptionWithReasonString:@"Controller suspended"];
 
+    // Clear any pending first-Thread-subscribe deferral so an already-armed
+    // watchdog short-circuits at its early-return rather than scheduling a
+    // subscribe that _subscriptionsAllowed will refuse anyway.  The latch is
+    // also cleared so that a controllerResumed -> _ensureSubscriptionForExistingDelegates
+    // re-entry can re-arm the deferral path against a fresh resume.
+    _deferringFirstThreadSubscription = NO;
+    _hasDeferredFirstThreadSubscription = NO;
+
     // Ensure that any pre-existing resubscribe attempts we control don't try to
     // do anything.
     _reattemptingSubscription = NO;