HPA: Do deferred operations on background threads.

include/jemalloc/internal/background_thread_externs.h

@@ -2,6 +2,7 @@
 #define JEMALLOC_INTERNAL_BACKGROUND_THREAD_EXTERNS_H
 
 extern bool opt_background_thread;
+extern ssize_t opt_background_thread_hpa_interval_max_ms;
 extern size_t opt_max_background_threads;
 extern malloc_mutex_t background_thread_lock;
 extern atomic_b_t background_thread_enabled_state;

include/jemalloc/internal/background_thread_structs.h

@@ -11,6 +11,14 @@
 #define MAX_BACKGROUND_THREAD_LIMIT MALLOCX_ARENA_LIMIT
 #define DEFAULT_NUM_BACKGROUND_THREAD 4
 
+/*
+ * These exist only as a transitional state.  Eventually, deferral should be
+ * part of the PAI, and each implementation can indicate wait times with more
+ * specificity.
+ */
+#define BACKGROUND_THREAD_HPA_INTERVAL_MAX_UNINITIALIZED (-2)
+#define BACKGROUND_THREAD_HPA_INTERVAL_MAX_DEFAULT_WHEN_ENABLED 5000
+
 typedef enum {
 	background_thread_stopped,
 	background_thread_started,

include/jemalloc/internal/pa.h

@@ -172,11 +172,21 @@ bool pa_shrink(tsdn_t *tsdn, pa_shard_t *shard, edata_t *edata, size_t old_size,
  */
 void pa_dalloc(tsdn_t *tsdn, pa_shard_t *shard, edata_t *edata,
     bool *generated_dirty);
-
 bool pa_decay_ms_set(tsdn_t *tsdn, pa_shard_t *shard, extent_state_t state,
     ssize_t decay_ms, pac_purge_eagerness_t eagerness);
 ssize_t pa_decay_ms_get(pa_shard_t *shard, extent_state_t state);
 
+/*
+ * Do deferred work on this PA shard.
+ *
+ * Morally, this should do both PAC decay and the HPA deferred work.  For now,
+ * though, the arena, background thread, and PAC modules are tightly interwoven
+ * in a way that's tricky to extricate, so we only do the HPA-specific parts.
+ */
+void pa_shard_set_deferral_allowed(tsdn_t *tsdn, pa_shard_t *shard,
+    bool deferral_allowed);
+void pa_shard_do_deferred_work(tsdn_t *tsdn, pa_shard_t *shard);
+
 /******************************************************************************/
 /*
  * Various bits of "boring" functionality that are still part of this module,