aboutsummaryrefslogtreecommitdiffstats
path: root/erts
diff options
context:
space:
mode:
authorJohn Högberg <[email protected]>2018-06-25 14:38:31 +0200
committerJohn Högberg <[email protected]>2018-06-25 14:38:31 +0200
commit38f3ea86a09252213f06ff427cceafba1f33caaa (patch)
tree87427521ce70c56131018e91efd2f45f7129ccfb /erts
parentebb767a2656a4d604e6f860cc03cbb4c0b0aecb8 (diff)
parent56a94efd43c1e5784ee97aac80b903e62155af4d (diff)
downloadotp-38f3ea86a09252213f06ff427cceafba1f33caaa.tar.gz
otp-38f3ea86a09252213f06ff427cceafba1f33caaa.tar.bz2
otp-38f3ea86a09252213f06ff427cceafba1f33caaa.zip
Merge branch 'john/erts/merge-OTP-15067' into maint
* john/erts/merge-OTP-15067: Don't enqueue system tasks if target process is in fail_state Fix erroneous schedule of freed/exiting processes Fix deadlock in run queue evacuation Fix memory leak of processes that died in the run queue
Diffstat (limited to 'erts')
-rw-r--r--erts/emulator/beam/erl_process.c284
-rw-r--r--erts/emulator/test/process_SUITE.erl54
2 files changed, 187 insertions, 151 deletions
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 8253ec4f40..11b52526d5 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -3761,6 +3761,8 @@ dequeue_process(ErtsRunQueue *runq, int prio_q, erts_aint32_t *statep)
ERTS_THR_DATA_DEPENDENCY_READ_MEMORY_BARRIER;
state = erts_atomic32_read_nob(&p->state);
+ ASSERT(state & ERTS_PSFLG_IN_RUNQ);
+
if (statep)
*statep = state;
@@ -3768,8 +3770,7 @@ dequeue_process(ErtsRunQueue *runq, int prio_q, erts_aint32_t *statep)
rqi = &runq->procs.prio_info[prio];
- if (p)
- unqueue_process(runq, rpq, rqi, prio, NULL, p);
+ unqueue_process(runq, rpq, rqi, prio, NULL, p);
return p;
}
@@ -4088,7 +4089,7 @@ evacuate_run_queue(ErtsRunQueue *rq,
erts_runq_unlock(to_rq);
smp_notify_inc_runq(to_rq);
- erts_runq_lock(to_rq);
+ erts_runq_lock(rq);
}
if (rq->ports.start) {
@@ -4157,22 +4158,17 @@ evacuate_run_queue(ErtsRunQueue *rq,
free_proxy_proc(proc);
else {
erts_aint32_t clr_bits;
-#ifdef DEBUG
- erts_aint32_t old;
-#endif
clr_bits = ERTS_PSFLG_IN_RUNQ;
clr_bits |= qbit << ERTS_PSFLGS_IN_PRQ_MASK_OFFSET;
-#ifdef DEBUG
- old =
-#else
- (void)
-#endif
- erts_atomic32_read_band_mb(&proc->state,
- ~clr_bits);
- ASSERT((old & clr_bits) == clr_bits);
+ state = erts_atomic32_read_band_mb(&proc->state, ~clr_bits);
+ ASSERT((state & clr_bits) == clr_bits);
+ if (state & ERTS_PSFLG_FREE) {
+ /* free and not queued by proxy */
+ erts_proc_dec_refc(proc);
+ }
}
goto handle_next_proc;
@@ -6208,13 +6204,14 @@ fin_dirty_enq_s_change(Process *p,
/* Already enqueue by someone else... */
if (pstruct_reserved) {
/* We reserved process struct for enqueue; clear it... */
-#ifdef DEBUG
- erts_aint32_t old =
-#else
- (void)
-#endif
- erts_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_IN_RUNQ);
- ASSERT(old & ERTS_PSFLG_IN_RUNQ);
+ erts_aint32_t state;
+
+ state = erts_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_IN_RUNQ);
+ ASSERT(state & ERTS_PSFLG_IN_RUNQ);
+
+ if (state & ERTS_PSFLG_FREE) {
+ erts_proc_dec_refc(p);
+ }
}
return 0;
}
@@ -6407,8 +6404,9 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p,
== ERTS_PSFLG_ACTIVE));
n &= ~running_flgs;
- if ((a & (ERTS_PSFLG_ACTIVE_SYS|ERTS_PSFLG_DIRTY_ACTIVE_SYS))
- || (a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE) {
+ if ((!!(a & (ERTS_PSFLG_ACTIVE_SYS|ERTS_PSFLG_DIRTY_ACTIVE_SYS))
+ | ((a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE))
+ & !(a & ERTS_PSFLG_FREE)) {
enqueue = check_enqueue_in_prio_queue(p, &enq_prio, &n, a);
}
a = erts_atomic32_cmpxchg_mb(&p->state, n, e);
@@ -6655,62 +6653,72 @@ erts_schedule_process(Process *p, erts_aint32_t state, ErtsProcLocks locks)
schedule_process(p, state, locks);
}
+/* Enqueues the given sys task on the process and schedules it. The task may be
+ * NULL if only scheduling is desired. */
static ERTS_INLINE erts_aint32_t
-active_sys_enqueue(Process *p, erts_aint32_t state,
- erts_aint32_t enable_flags, int status_locked)
-{
- /*
- * This function may or may not be called with status locke held.
- * It always returns without the status lock held!
- */
- unsigned int prof_runnable_procs = erts_system_profile_flags.runnable_procs;
- erts_aint32_t n, a = state, enq_prio = -1;
- int slocked = status_locked;
+active_sys_enqueue(Process *p, ErtsProcSysTask *sys_task,
+ erts_aint32_t task_prio, erts_aint32_t enable_flags,
+ erts_aint32_t state, erts_aint32_t *fail_state_p)
+{
+ int runnable_procs = erts_system_profile_flags.runnable_procs;
+ erts_aint32_t n, a, enq_prio, fail_state;
+ int already_scheduled;
+ int status_locked;
int enqueue; /* < 0 -> use proxy */
- /* Status lock prevents out of order "runnable proc" trace msgs */
- ERTS_LC_ASSERT(slocked || !(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)));
- ERTS_LC_ASSERT(!slocked || (ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)));
-
- if (!prof_runnable_procs) {
- if (slocked) {
- erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
- slocked = 0;
- }
- }
- else {
- if (!slocked) {
- erts_proc_lock(p, ERTS_PROC_LOCK_STATUS);
- slocked = !0;
- }
- }
+ enable_flags |= ERTS_PSFLG_ACTIVE_SYS;
+ fail_state = *fail_state_p;
+ already_scheduled = 0;
+ status_locked = 0;
+ enq_prio = -1;
+ a = state;
+ ERTS_LC_ASSERT(!(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)));
+ ASSERT(fail_state & (ERTS_PSFLG_EXITING | ERTS_PSFLG_FREE));
+ ASSERT(!(fail_state & enable_flags));
ASSERT(!(state & ERTS_PSFLG_PROXY));
+ /* When runnable_procs is enabled, we need to take the status lock to
+ * prevent trace messages from being sent in the wrong order. The lock must
+ * be held over the call to add2runq.
+ *
+ * Otherwise, we only need to take it when we're enqueuing a task and can
+ * safely release it before add2runq. */
+ if (sys_task || runnable_procs) {
+ erts_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+ status_locked = 1;
+ }
+
while (1) {
erts_aint32_t e;
n = e = a;
- if (a & ERTS_PSFLG_FREE)
- goto cleanup; /* We don't want to schedule free processes... */
+ if (a & fail_state) {
+ *fail_state_p = a & fail_state;
+ goto cleanup;
+ }
enqueue = ERTS_ENQUEUE_NOT;
- n |= enable_flags;
- n |= ERTS_PSFLG_ACTIVE_SYS;
+ n |= enable_flags;
+
if (!(a & (ERTS_PSFLG_RUNNING
| ERTS_PSFLG_RUNNING_SYS
| ERTS_PSFLG_DIRTY_RUNNING
- | ERTS_PSFLG_DIRTY_RUNNING_SYS)))
+ | ERTS_PSFLG_DIRTY_RUNNING_SYS))) {
enqueue = check_enqueue_in_prio_queue(p, &enq_prio, &n, a);
+ }
+
a = erts_atomic32_cmpxchg_mb(&p->state, n, e);
- if (a == e)
+ if (a == e) {
break;
- if (a == n && enqueue == ERTS_ENQUEUE_NOT)
- goto cleanup;
+ }
+ else if (a == n && enqueue == ERTS_ENQUEUE_NOT) {
+ already_scheduled = 1;
+ break;
+ }
}
- if (prof_runnable_procs) {
-
+ if (!already_scheduled && runnable_procs) {
if (!(a & (ERTS_PSFLG_ACTIVE_SYS
| ERTS_PSFLG_RUNNING
| ERTS_PSFLG_RUNNING_SYS
@@ -6720,19 +6728,56 @@ active_sys_enqueue(Process *p, erts_aint32_t state,
/* We activated a prevously inactive process */
profile_runnable_proc(p, am_active);
}
-
- erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
- slocked = 0;
}
- add2runq(enqueue, enq_prio, p, n, NULL);
+ if (sys_task) {
+ ErtsProcSysTaskQs *stqs = p->sys_task_qs;
-cleanup:
+ if (!stqs) {
+ sys_task->next = sys_task->prev = sys_task;
- if (slocked)
- erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+ stqs = proc_sys_task_queues_alloc();
- ERTS_LC_ASSERT(!(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)));
+ stqs->qmask = 1 << task_prio;
+ stqs->ncount = 0;
+ stqs->q[PRIORITY_MAX] = NULL;
+ stqs->q[PRIORITY_HIGH] = NULL;
+ stqs->q[PRIORITY_NORMAL] = NULL;
+ stqs->q[PRIORITY_LOW] = NULL;
+ stqs->q[task_prio] = sys_task;
+
+ p->sys_task_qs = stqs;
+ }
+ else {
+ if (!stqs->q[task_prio]) {
+ sys_task->next = sys_task->prev = sys_task;
+
+ stqs->q[task_prio] = sys_task;
+ stqs->qmask |= 1 << task_prio;
+ }
+ else {
+ sys_task->next = stqs->q[task_prio];
+ sys_task->prev = stqs->q[task_prio]->prev;
+ sys_task->next->prev = sys_task;
+ sys_task->prev->next = sys_task;
+ ASSERT(stqs->qmask & (1 << task_prio));
+ }
+ }
+ }
+
+ if (status_locked && !runnable_procs) {
+ erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+ status_locked = 0;
+ }
+
+ if (!already_scheduled) {
+ add2runq(enqueue, enq_prio, p, n, NULL);
+ }
+
+cleanup:
+ if (status_locked) {
+ erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+ }
return n;
}
@@ -6740,103 +6785,41 @@ cleanup:
erts_aint32_t
erts_proc_sys_schedule(Process *p, erts_aint32_t state, erts_aint32_t enable_flag)
{
- /* We are not allowed to call this function with status lock held... */
- return active_sys_enqueue(p, state, enable_flag, 0);
+ erts_aint32_t fail_state = ERTS_PSFLG_FREE;
+
+ return active_sys_enqueue(p, NULL, 0, enable_flag, state, &fail_state);
}
static int
schedule_process_sys_task(Process *p, erts_aint32_t prio, ErtsProcSysTask *st,
erts_aint32_t *fail_state_p)
{
- int res;
- int locked;
- ErtsProcSysTaskQs *stqs, *free_stqs;
erts_aint32_t fail_state, state;
- fail_state = *fail_state_p;
-
- res = 1; /* prepare for success */
- st->next = st->prev = st; /* Prep for empty prio queue */
+ /* Elevate priority if needed. */
state = erts_atomic32_read_nob(&p->state);
- locked = 0;
- free_stqs = NULL;
- if (state & ERTS_PSFLG_SYS_TASKS)
- stqs = NULL;
- else {
- alloc_qs:
- stqs = proc_sys_task_queues_alloc();
- stqs->qmask = 1 << prio;
- stqs->ncount = 0;
- stqs->q[PRIORITY_MAX] = NULL;
- stqs->q[PRIORITY_HIGH] = NULL;
- stqs->q[PRIORITY_NORMAL] = NULL;
- stqs->q[PRIORITY_LOW] = NULL;
- stqs->q[prio] = st;
- }
-
- if (!locked) {
- locked = 1;
- erts_proc_lock(p, ERTS_PROC_LOCK_STATUS);
-
- state = erts_atomic32_read_nob(&p->state);
- if (state & fail_state) {
- erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
- *fail_state_p = (state & fail_state);
- free_stqs = stqs;
- res = 0;
- goto cleanup;
- }
- }
-
- if (!p->sys_task_qs) {
- if (stqs)
- p->sys_task_qs = stqs;
- else
- goto alloc_qs;
- }
- else {
- free_stqs = stqs;
- stqs = p->sys_task_qs;
- if (!stqs->q[prio]) {
- stqs->q[prio] = st;
- stqs->qmask |= 1 << prio;
- }
- else {
- st->next = stqs->q[prio];
- st->prev = stqs->q[prio]->prev;
- st->next->prev = st;
- st->prev->next = st;
- ASSERT(stqs->qmask & (1 << prio));
- }
- }
-
if (ERTS_PSFLGS_GET_ACT_PRIO(state) > prio) {
- erts_aint32_t n, a, e;
- /* Need to elevate actual prio */
+ erts_aint32_t n, a, e;
- a = state;
- do {
- if (ERTS_PSFLGS_GET_ACT_PRIO(a) <= prio) {
- n = a;
- break;
- }
- n = e = a;
- n &= ~ERTS_PSFLGS_ACT_PRIO_MASK;
- n |= (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET);
- a = erts_atomic32_cmpxchg_nob(&p->state, n, e);
- } while (a != e);
- state = n;
- }
-
- /* active_sys_enqueue() always return with status lock unlocked */
- (void) active_sys_enqueue(p, state, ERTS_PSFLG_SYS_TASKS, locked);
+ a = state;
+ do {
+ if (ERTS_PSFLGS_GET_ACT_PRIO(a) <= prio) {
+ n = a;
+ break;
+ }
+ n = e = a;
+ n &= ~ERTS_PSFLGS_ACT_PRIO_MASK;
+ n |= (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET);
+ a = erts_atomic32_cmpxchg_nob(&p->state, n, e);
+ } while (a != e);
-cleanup:
+ state = n;
+ }
- if (free_stqs)
- proc_sys_task_queues_free(free_stqs);
+ fail_state = *fail_state_p;
- return res;
+ return !(active_sys_enqueue(p, st, prio, ERTS_PSFLG_SYS_TASKS,
+ state, fail_state_p) & fail_state);
}
static ERTS_INLINE int
@@ -9563,6 +9546,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
}
else if (state & ERTS_PSFLG_FREE) {
/* free and not queued by proxy */
+ ASSERT(state & ERTS_PSFLG_IN_RUNQ);
erts_proc_dec_refc(p);
}
if (!is_normal_sched)
diff --git a/erts/emulator/test/process_SUITE.erl b/erts/emulator/test/process_SUITE.erl
index 585c5a1871..57eb082d64 100644
--- a/erts/emulator/test/process_SUITE.erl
+++ b/erts/emulator/test/process_SUITE.erl
@@ -59,6 +59,7 @@
no_priority_inversion2/1,
system_task_blast/1,
system_task_on_suspended/1,
+ system_task_failed_enqueue/1,
gc_request_when_gc_disabled/1,
gc_request_blast_when_gc_disabled/1]).
-export([prio_server/2, prio_client/2, init/1, handle_event/2]).
@@ -106,7 +107,7 @@ groups() ->
otp_7738_resume]},
{system_task, [],
[no_priority_inversion, no_priority_inversion2,
- system_task_blast, system_task_on_suspended,
+ system_task_blast, system_task_on_suspended, system_task_failed_enqueue,
gc_request_when_gc_disabled, gc_request_blast_when_gc_disabled]}].
init_per_suite(Config) ->
@@ -2625,6 +2626,57 @@ system_task_on_suspended(Config) when is_list(Config) ->
ok
end.
+%% When a system task couldn't be enqueued due to the process being in an
+%% incompatible state, it would linger in the system task list and get executed
+%% anyway the next time the process was scheduled. This would result in a
+%% double-free at best.
+%%
+%% This test continuously purges modules while other processes run dirty code,
+%% which will provoke this error as ERTS_PSTT_CPC can't be enqueued while a
+%% process is running dirty code.
+system_task_failed_enqueue(Config) when is_list(Config) ->
+ case erlang:system_info(dirty_cpu_schedulers) of
+ N when N > 0 ->
+ system_task_failed_enqueue_1(Config);
+ _ ->
+ {skipped, "No dirty scheduler support"}
+ end.
+
+system_task_failed_enqueue_1(Config) ->
+ Priv = proplists:get_value(priv_dir, Config),
+
+ Purgers = [spawn_link(fun() -> purge_loop(Priv, Id) end)
+ || Id <- lists:seq(1, erlang:system_info(schedulers))],
+ Hogs = [spawn_link(fun() -> dirty_loop() end)
+ || _ <- lists:seq(1, erlang:system_info(dirty_cpu_schedulers))],
+
+ ct:sleep(5000),
+
+ [begin
+ unlink(Pid),
+ exit(Pid, kill)
+ end || Pid <- (Purgers ++ Hogs)],
+
+ ok.
+
+purge_loop(PrivDir, Id) ->
+ Mod = "failed_enq_" ++ integer_to_list(Id),
+ Path = PrivDir ++ "/" ++ Mod,
+ file:write_file(Path ++ ".erl",
+ "-module('" ++ Mod ++ "').\n" ++
+ "-export([t/0]).\n" ++
+ "t() -> ok."),
+ purge_loop_1(Path).
+purge_loop_1(Path) ->
+ {ok, Mod} = compile:file(Path, []),
+ erlang:delete_module(Mod),
+ erts_code_purger:purge(Mod),
+ purge_loop_1(Path).
+
+dirty_loop() ->
+ ok = erts_debug:dirty_cpu(reschedule, 10000),
+ dirty_loop().
+
gc_request_when_gc_disabled(Config) when is_list(Config) ->
AIS = erts_debug:set_internal_state(available_internal_state, true),
gc_request_when_gc_disabled_do(ref),