*
   * In part because Cells are relatively large, we avoid creating
   * them until they are needed.  When there is no contention, all
   * updates are made to the base field.  Upon first contention (a
   * failed CAS on base update), the table is initialized to size 2.
   * The table size is doubled upon further contention until
   * reaching the nearest power of two greater than or equal to the

   *
   * In part because Cells are relatively large, we avoid creating
   * them until they are needed.  When there is no contention, all
   * updates are made to the base field.  Upon first contention (a
   * failed CAS on base update), the table is initialized to size 2.
   * The table size is doubled upon further contention until
   * reaching the nearest power of two greater than or equal to the

 * sum that is used for purposes such as collecting statistics, not for fine-grained synchronization
 * control. Under low update contention, the two classes have similar characteristics. But under
 * high contention, expected throughput of this class is significantly higher, at the expense of
 * higher space consumption.
 *
 * <p>This class extends {@link Number}, but does not define methods such as {@code

 * sum that is used for purposes such as collecting statistics, not for fine-grained synchronization
 * control. Under low update contention, the two classes have similar characteristics. But under
 * high contention, expected throughput of this class is significantly higher, at the expense of
 * higher space consumption.
 *
 * <p>This class extends {@link Number}, but does not define methods such as {@code

 * sum that is used for purposes such as collecting statistics, not for fine-grained synchronization
 * control. Under low update contention, the two classes have similar characteristics. But under
 * high contention, expected throughput of this class is significantly higher, at the expense of
 * higher space consumption.
 *
 * <p>This class extends {@link Number}, but does not define methods such as {@code

 * sum that is used for purposes such as collecting statistics, not for fine-grained synchronization
 * control. Under low update contention, the two classes have similar characteristics. But under
 * high contention, expected throughput of this class is significantly higher, at the expense of
 * higher space consumption.
 *
 * <p>This class extends {@link Number}, but does not define methods such as {@code

#### Per-Connection Locks

Each connection has its own lock. The connections in the pool are all in a `ConcurrentLinkedQueue`. Due to data races, iterators of this queue may return removed connections. Callers must check the connection's `noNewExchanges` property before using connections from the pool.

The connection lock is never held while doing I/O (even closing a socket) to prevent contention.

[GC guide](/doc/gc-guide#Linux_transparent_huge_pages), or switch to a Linux
distribution that disables transparent huge pages altogether.

Go 1.22 added contention on runtime-internal locks to the [`mutex`
profile](/pkg/runtime/pprof#Profile). Contention on these locks is always
reported at `runtime._LostContendedRuntimeLock`. Complete stack traces of
runtime locks can be enabled with the [`runtimecontentionstacks`

  /**
   * Guides the allowed concurrency among update operations. Used as a hint for internal sizing. The
   * table is internally partitioned to try to permit the indicated number of concurrent updates
   * without contention. Because assignment of entries to these partitions is not necessarily
   * uniform, the actual concurrency observed may vary. Ideally, you should choose a value to

 * <i>stripes</i>, increasing the granularity of a single lock and allowing independent operations
 * to lock different stripes and proceed concurrently, instead of creating contention for a single
 * lock.
 *
 * <p>The guarantee provided by this class is that equal keys lead to the same lock (or semaphore),
 * i.e. {@code if (key1.equals(key2))} then {@code striped.get(key1) == striped.get(key2)} (assuming