Merge branch 'rickard/temp_alloc_check/OTP-9028' into dev

* rickard/temp_alloc_check/OTP-9028:
  Verify that temp allocated memory is released

7 files changed, 139 insertions, 0 deletions

diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 12a8022861..af7ed064d3 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -321,6 +321,7 @@ extern int count_instructions;
 #  define POST_BIF_GC_SWAPIN_0(_p, _res)				\
      ERTS_SMP_REQ_PROC_MAIN_LOCK((_p));					\
      PROCESS_MAIN_CHK_LOCKS((_p));					\
+     ERTS_VERIFY_UNUSED_TEMP_ALLOC((_p));				\
      if (((_p)->mbuf) || (MSO(_p).overhead >= BIN_VHEAP_SZ(_p)) ) {	\
        _res = erts_gc_after_bif_call((_p), (_res), NULL, 0);		\
        E = (_p)->stop;							\
@@ -328,6 +329,7 @@ extern int count_instructions;
      HTOP = HEAP_TOP((_p))
 
 #  define POST_BIF_GC_SWAPIN(_p, _res, _regs, _arity)			\
+     ERTS_VERIFY_UNUSED_TEMP_ALLOC((_p));				\
      ERTS_SMP_REQ_PROC_MAIN_LOCK((_p));					\
      PROCESS_MAIN_CHK_LOCKS((_p));					\
      if (((_p)->mbuf) || (MSO(_p).overhead >= BIN_VHEAP_SZ(_p)) ) {	\
@@ -367,6 +369,7 @@ extern int count_instructions;
            reg[0] = r(0); \
            PROCESS_MAIN_CHK_LOCKS(c_p); \
            FCALLS -= erts_garbage_collect(c_p, needed + (HeapNeed), reg, (M)); \
+           ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p); \
            PROCESS_MAIN_CHK_LOCKS(c_p); \
            r(0) = reg[0]; \
            SWAPIN; \
@@ -420,6 +423,7 @@ extern int count_instructions;
        reg[0] = r(0);                                           		\
        PROCESS_MAIN_CHK_LOCKS(c_p);                             		\
        FCALLS -= erts_garbage_collect(c_p, need, reg, (Live));  		\
+       ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);					\
        PROCESS_MAIN_CHK_LOCKS(c_p);                             		\
        r(0) = reg[0];                                           		\
        SWAPIN;                                                  		\
@@ -442,6 +446,7 @@ extern int count_instructions;
        reg[0] = r(0);                                           \
        PROCESS_MAIN_CHK_LOCKS(c_p);                             \
        FCALLS -= erts_garbage_collect(c_p, need, reg, (Live));  \
+       ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);			\
        PROCESS_MAIN_CHK_LOCKS(c_p);                             \
        r(0) = reg[0];                                           \
        SWAPIN;                                                  \
@@ -464,6 +469,7 @@ extern int count_instructions;
        reg[Live] = Extra;						\
        PROCESS_MAIN_CHK_LOCKS(c_p);					\
        FCALLS -= erts_garbage_collect(c_p, need, reg, (Live)+1);	\
+       ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);				\
        PROCESS_MAIN_CHK_LOCKS(c_p);					\
        if (Live > 0) {							\
 	   r(0) = reg[0];						\
@@ -1214,7 +1220,9 @@ void process_main(void)
  do_schedule1:
     PROCESS_MAIN_CHK_LOCKS(c_p);
     ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
+    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
     c_p = schedule(c_p, reds_used);
+    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 #ifdef DEBUG
     pid = c_p->id;
 #endif
@@ -1664,6 +1672,7 @@ void process_main(void)
 		 SWAPOUT;
 		 PROCESS_MAIN_CHK_LOCKS(c_p);
 		 FCALLS -= erts_garbage_collect(c_p, 3, reg+2, 1);
+		 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		 PROCESS_MAIN_CHK_LOCKS(c_p);
 		 SWAPIN;
 	     }
@@ -1782,6 +1791,7 @@ void process_main(void)
 				       PROCESS_MAIN_CHK_LOCKS(c_p);
 				   },
 				   {
+				       ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 				       PROCESS_MAIN_CHK_LOCKS(c_p);
 				       r(0) = reg[0];
 				       SWAPIN;
@@ -1840,6 +1850,7 @@ void process_main(void)
      CANCEL_TIMER(c_p);
      free_message(msgp);
 
+     ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
      PROCESS_MAIN_CHK_LOCKS(c_p);
 
      NextPF(0, next);
@@ -2230,6 +2241,7 @@ void process_main(void)
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 	result = (*bf)(c_p, arg);
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(result));
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_HOLE_CHECK(c_p);
 	FCALLS = c_p->fcalls;
@@ -2258,6 +2270,7 @@ void process_main(void)
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 	result = (*bf)(c_p, arg);
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(result));
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_HOLE_CHECK(c_p);
 	FCALLS = c_p->fcalls;
@@ -2287,6 +2300,7 @@ void process_main(void)
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
 	result = (*bf)(c_p, reg, live);
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	SWAPIN;
@@ -2322,6 +2336,7 @@ void process_main(void)
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
 	result = (*bf)(c_p, reg, live);
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	SWAPIN;
@@ -2360,6 +2375,7 @@ void process_main(void)
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
 	result = (*bf)(c_p, reg, live);
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	SWAPIN;
@@ -2394,6 +2410,7 @@ void process_main(void)
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 	result = (*bf)(c_p, tmp_arg1, tmp_arg2);
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(result));
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_HOLE_CHECK(c_p);
 	FCALLS = c_p->fcalls;
@@ -2417,6 +2434,7 @@ void process_main(void)
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 	result = (*bf)(c_p, tmp_arg1, tmp_arg2);
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(result));
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	ERTS_HOLE_CHECK(c_p);
 	if (is_value(result)) {
@@ -3287,6 +3305,7 @@ void process_main(void)
 	    PROCESS_MAIN_CHK_LOCKS(c_p);
 	    bif_nif_arity = I[-1];
 	    ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
+	    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 
 	    ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 	    {
@@ -3300,6 +3319,7 @@ void process_main(void)
 	    }
 	    ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(nif_bif_result));
 	    PROCESS_MAIN_CHK_LOCKS(c_p);
+	    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	    goto apply_bif_or_nif_epilogue;
 	 
 	OpCase(apply_bif):
@@ -3326,6 +3346,7 @@ void process_main(void)
 	    bif_nif_arity = I[-1];
 	    ASSERT(bif_nif_arity <= 3);
 	    ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
+	    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	    switch (bif_nif_arity) {
 	    case 3:
 		{
@@ -3334,6 +3355,7 @@ void process_main(void)
 		    nif_bif_result = (*bf)(c_p, r(0), x(1), x(2), I);
 		    ASSERT(!ERTS_PROC_IS_EXITING(c_p) ||
 			   is_non_value(nif_bif_result));
+		    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		    PROCESS_MAIN_CHK_LOCKS(c_p);
 		}
 		break;
@@ -3344,6 +3366,7 @@ void process_main(void)
 		    nif_bif_result = (*bf)(c_p, r(0), x(1), I);
 		    ASSERT(!ERTS_PROC_IS_EXITING(c_p) ||
 			   is_non_value(nif_bif_result));
+		    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		    PROCESS_MAIN_CHK_LOCKS(c_p);
 		}
 		break;
@@ -3354,6 +3377,7 @@ void process_main(void)
 		    nif_bif_result = (*bf)(c_p, r(0), I);
 		    ASSERT(!ERTS_PROC_IS_EXITING(c_p) ||
 			   is_non_value(nif_bif_result));
+		    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		    PROCESS_MAIN_CHK_LOCKS(c_p);
 		}
 		break;
@@ -3364,6 +3388,7 @@ void process_main(void)
 		    nif_bif_result = (*bf)(c_p, I);
 		    ASSERT(!ERTS_PROC_IS_EXITING(c_p) ||
 			   is_non_value(nif_bif_result));
+		    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		    PROCESS_MAIN_CHK_LOCKS(c_p);
 		    break;
 		}
@@ -4590,6 +4615,7 @@ void process_main(void)
 	 ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
 	 flags = erts_call_trace(c_p, ep->code, ep->match_prog_set, reg,
 				 0, &c_p->tracer_proc);
+	 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	 ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
 	 PROCESS_MAIN_CHK_LOCKS(c_p);
 	 ASSERT(!ERTS_PROC_IS_EXITING(c_p));
@@ -4601,6 +4627,7 @@ void process_main(void)
 		 /* SWAPOUT, SWAPIN was done and r(0) was saved above */
 		 PROCESS_MAIN_CHK_LOCKS(c_p);
 		 FCALLS -= erts_garbage_collect(c_p, 3, reg, ep->code[2]);
+		 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		 PROCESS_MAIN_CHK_LOCKS(c_p);
 		 r(0) = reg[0];
 		 SWAPIN;
@@ -4690,6 +4717,7 @@ void process_main(void)
 		 reg[0] = r(0);
 		 PROCESS_MAIN_CHK_LOCKS(c_p);
 		 FCALLS -= erts_garbage_collect(c_p, 2, reg, I[-1]);
+		 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 		 PROCESS_MAIN_CHK_LOCKS(c_p);
 		 r(0) = reg[0];
 	     }
@@ -4793,6 +4821,7 @@ void process_main(void)
 	     /* SWAPOUT was done and r(0) was saved above */
 	     PROCESS_MAIN_CHK_LOCKS(c_p);
 	     FCALLS -= erts_garbage_collect(c_p, need, reg, I[-1]);
+	     ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	     PROCESS_MAIN_CHK_LOCKS(c_p);
 	     r(0) = reg[0];
 	     SWAPIN;
@@ -6239,6 +6268,7 @@ hibernate(Process* c_p, Eterm module, Eterm function, Eterm args, Eterm* reg)
 	c_p->fvalue = NIL;
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	erts_garbage_collect_hibernate(c_p);
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
 	PROCESS_MAIN_CHK_LOCKS(c_p);
 	erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
 	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
@@ -6484,6 +6514,7 @@ new_fun(Process* p, Eterm* reg, ErlFunEntry* fe, int num_free)
     if (HEAP_LIMIT(p) - HEAP_TOP(p) <= needed) {
 	PROCESS_MAIN_CHK_LOCKS(p);
 	erts_garbage_collect(p, needed, reg, num_free);
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(p);
 	PROCESS_MAIN_CHK_LOCKS(p);
     }
     hp = p->htop;
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index e85e2d7e3f..323c422c6d 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -1411,6 +1411,33 @@ void erts_alloc_reg_scheduler_id(Uint id)
     erts_tsd_set(thr_ix_key, (void *)(long) ix);
 }
 
+static void
+no_verify(Allctr_t *allctr)
+{
+
+}
+
+erts_alloc_verify_func_t
+erts_alloc_get_verify_unused_temp_alloc(Allctr_t **allctr)
+{
+    if (erts_allctrs_info[ERTS_ALC_A_TEMPORARY].alloc_util
+	&& erts_allctrs_info[ERTS_ALC_A_TEMPORARY].thr_spec) {
+	ErtsAllocatorThrSpec_t *tspec;
+	tspec = &erts_allctr_thr_spec[ERTS_ALC_A_TEMPORARY];
+	if (!tspec->all_thr_safe) {
+	    int ix = erts_alc_get_thr_ix();
+
+	    if (ix < tspec->size) {
+		*allctr = tspec->allctr[ix];
+		return erts_alcu_verify_unused;
+	    }
+	}
+    }
+
+    *allctr = NULL;
+    return no_verify;
+}
+
 __decl_noreturn void
 erts_alc_fatal_error(int error, int func, ErtsAlcType_t n, ...)
 {
diff --git a/erts/emulator/beam/erl_alloc.h b/erts/emulator/beam/erl_alloc.h
index 3e96c76dbf..dd4cc22171 100644
--- a/erts/emulator/beam/erl_alloc.h
+++ b/erts/emulator/beam/erl_alloc.h
@@ -236,6 +236,11 @@ void *erts_realloc_fnf(ErtsAlcType_t type, void *ptr, Uint size)
 
 #endif /* #if ERTS_ALC_DO_INLINE || defined(ERTS_ALC_INTERNAL__) */
 
+typedef void (*erts_alloc_verify_func_t)(Allctr_t *);
+
+erts_alloc_verify_func_t
+erts_alloc_get_verify_unused_temp_alloc(Allctr_t **allctr);
+
 #ifndef ERTS_CACHE_LINE_SIZE
 /* Assume a cache line size of 64 bytes */
 #  define ERTS_CACHE_LINE_SIZE ((UWord) 64)
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 8b184899c9..c09f0bbd77 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -3368,6 +3368,38 @@ erts_alcu_test(unsigned long op, unsigned long a1, unsigned long a2)
  * Debug functions                                                           *
 \*                                                                           */
 
+void
+erts_alcu_verify_unused(Allctr_t *allctr)
+{
+    UWord no;
+
+    no = allctr->sbcs.curr_mseg.no;
+    no += allctr->sbcs.curr_sys_alloc.no;
+    no += allctr->mbcs.blocks.curr.no;
+
+    if (no) {
+	UWord sz = allctr->sbcs.blocks.curr.size;
+	sz += allctr->mbcs.blocks.curr.size;
+	erl_exit(ERTS_ABORT_EXIT,
+		 "%salloc() used when expected to be unused!\n"
+		 "Total amount of blocks allocated: %bpu\n"
+		 "Total amount of bytes allocated: %bpu\n",
+		 allctr->name_prefix, no, sz);
+    }
+}
+
+void
+erts_alcu_verify_unused_ts(Allctr_t *allctr)
+{
+#ifdef USE_THREADS
+    erts_mtx_lock(&allctr->mutex);
+#endif
+    erts_alcu_verify_unused(allctr);
+#ifdef USE_THREADS
+    erts_mtx_unlock(&allctr->mutex);
+#endif
+}
+
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
 
 static void
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index f2b951bca6..5a8db5e29e 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -333,6 +333,9 @@ struct Allctr_t_ {
 int	erts_alcu_start(Allctr_t *, AllctrInit_t *);
 void	erts_alcu_stop(Allctr_t *);
 
+void	erts_alcu_verify_unused(Allctr_t *);
+void	erts_alcu_verify_unused_ts(Allctr_t *allctr);
+
 unsigned long	erts_alcu_test(unsigned long, unsigned long, unsigned long);
 
 
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index ddfc27a93f..e4a871fd7e 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -2752,6 +2752,15 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online)
 
     /* init port tasks */
     erts_port_task_init();
+
+#ifndef ERTS_SMP
+#ifdef ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC
+    erts_scheduler_data->verify_unused_temp_alloc
+	= erts_alloc_get_verify_unused_temp_alloc(
+	    &erts_scheduler_data->verify_unused_temp_alloc_data);
+    ERTS_VERIFY_UNUSED_TEMP_ALLOC(NULL);
+#endif
+#endif
 }
 
 ErtsRunQueue *
@@ -3711,6 +3720,13 @@ sched_thread_func(void *vesdp)
     }
     erts_smp_mtx_unlock(&schdlr_sspnd.mtx);
 
+#ifdef ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC
+    ((ErtsSchedulerData *) vesdp)->verify_unused_temp_alloc
+	= erts_alloc_get_verify_unused_temp_alloc(
+	    &((ErtsSchedulerData *) vesdp)->verify_unused_temp_alloc_data);
+    ERTS_VERIFY_UNUSED_TEMP_ALLOC(NULL);
+#endif
+
     process_main();
     /* No schedulers should *ever* terminate */
     erl_exit(ERTS_ABORT_EXIT, "Scheduler thread number %bpu terminated\n",
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 8479e56710..e871a9834a 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -28,6 +28,12 @@
 #define ERTS_INCLUDE_SCHEDULER_INTERNALS
 #endif
 
+/* #define ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC */
+
+#if !defined(ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC) && defined(DEBUG)
+#  define ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC
+#endif
+
 typedef struct process Process;
 
 #include "sys.h"
@@ -421,6 +427,11 @@ struct ErtsSchedulerData_ {
     /* NOTE: These fields are modified under held mutexes by other threads */
     erts_smp_atomic32_t chk_cpu_bind; /* Only used when common run queue */
 #endif
+
+#ifdef ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC
+    erts_alloc_verify_func_t verify_unused_temp_alloc;
+    Allctr_t *verify_unused_temp_alloc_data;
+#endif
 };
 
 typedef union {
@@ -1145,6 +1156,20 @@ Uint erts_debug_nbalance(void);
 #  define ERTS_PROC_GET_SCHDATA(PROC) (erts_scheduler_data)
 #endif
 
+#ifdef ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC
+#  define ERTS_VERIFY_UNUSED_TEMP_ALLOC(P)					\
+do {										\
+    ErtsSchedulerData *esdp__ = ((P)						\
+				 ? ERTS_PROC_GET_SCHDATA((Process *) (P))	\
+				 : erts_get_scheduler_data());			\
+    if (esdp__)									\
+	esdp__->verify_unused_temp_alloc(					\
+	    esdp__->verify_unused_temp_alloc_data);				\
+} while (0)
+#else
+#  define ERTS_VERIFY_UNUSED_TEMP_ALLOC(ESDP)
+#endif
+
 #if defined(ERTS_SMP) || defined(USE_THREADS)
 ErtsSchedulerData *erts_get_scheduler_data(void);
 #else