diff --git a/src/fibers/fiber.cpp b/src/fibers/fiber.cpp
index 8863872..aa7dab9 100644
--- a/src/fibers/fiber.cpp
+++ b/src/fibers/fiber.cpp
@@ -586,6 +586,8 @@ struct FiberScheduler::ProcessorState
     void wakeThread() noexcept;
     void parkThread(uint64_t waitNs, CpuTimer * timer) noexcept;
     bool hasWork() const noexcept;
+    uint32_t sqReady() const noexcept;
+    uint32_t cqReady() const noexcept;
 
     void enqueueDoorbell() noexcept;
     void postWakeup(ProcessorState * target) noexcept;
@@ -770,6 +772,13 @@ void FiberScheduler::ProcessorState::wakeThread() noexcept
 
 void FiberScheduler::ProcessorState::parkThread(uint64_t waitNs, CpuTimer * timer) noexcept
 {
+    // Flush deferred SQEs: the idle path has no other submit, and parking
+    // passes to_submit=0, so a deferred doorbell rearm, MSG_RING wakeup, or
+    // remote cancel would sit queued until unrelated activity lands on this
+    // ring. If the submit defers again: on EBUSY hasWork() sees the full CQ
+    // and skips the park; on EAGAIN the timed park is the retry backoff.
+    submitIo(true);
+
     // Announce that we are about to park, then a seq_cst fence pairing with the one in wakeThread:
     // release alone is not a StoreLoad barrier, so without this the store could reorder past the
     // hasWork() re-check below while a concurrent wakeThread reads sleeping=false - both miss, and
@@ -822,13 +831,30 @@ bool FiberScheduler::ProcessorState::hasWork() const noexcept
     {
         return true;
     }
+    if (cqReady())
+    {
+        return true;
+    }
+    return false;
+}
 
-    // Ignore ring->cq.khead atomic/non-atomic mismatch.
+// The kernel and other submitters update the SQ/CQ ring counters with plain
+// stores; a racy read is benign - a stale count only defers or repeats work -
+// so hide it from TSan here rather than annotating every call site.
+uint32_t FiberScheduler::ProcessorState::sqReady() const noexcept
+{
     TSAN_IGNORE_BEGIN();
-    uint32_t count = ::io_uring_cq_ready(&ring);
+    uint32_t count = ::io_uring_sq_ready(&ring);
     TSAN_IGNORE_END();
+    return count;
+}
 
-    return count > 0;
+uint32_t FiberScheduler::ProcessorState::cqReady() const noexcept
+{
+    TSAN_IGNORE_BEGIN();
+    uint32_t count = ::io_uring_cq_ready(&ring);
+    TSAN_IGNORE_END();
+    return count;
 }
 
 void FiberScheduler::ProcessorState::enqueueDoorbell() noexcept
@@ -943,9 +969,7 @@ bool FiberScheduler::ProcessorState::submitIo(bool flush) noexcept
     // the submission lock when there's nothing to submit or the count/staleness
     // thresholds haven't been met. Kept small so it inlines into runFiber;
     // the rest lives in submitIoSlow.
-    TSAN_IGNORE_BEGIN();
-    uint32_t count = ::io_uring_sq_ready(&ring);
-    TSAN_IGNORE_END();
+    uint32_t count = sqReady();
     if (count == 0)
     {
         return false;
@@ -970,7 +994,7 @@ __attribute__((noinline)) bool FiberScheduler::ProcessorState::submitIoSlow(uint
 {
     std::lock_guard lock(submissionLock);
 
-    uint32_t count = ::io_uring_sq_ready(&ring);
+    uint32_t count = sqReady();
     if (count == 0)
     {
         return false;
@@ -1970,7 +1994,8 @@ bool FiberScheduler::handleReadyQueue(ProcessorState * processor, CpuTimer * tim
 bool FiberScheduler::handleCompletionQueue(ProcessorState * processor) noexcept
 {
     // Fast path: CQ ring is empty, nothing to drain.
-    if (::io_uring_cq_ready(&processor->ring) == 0)
+    uint32_t count = processor->cqReady();
+    if (count == 0)
     {
         return false;
     }