From 775f1686a3df68bd20370f1fabc6273883e2c5d2 Mon Sep 17 00:00:00 2001 From: Thomas Schwinge Date: Thu, 26 Mar 2020 17:34:01 +0100 Subject: [PATCH] Really fix og9 "Fix hang when running oacc exec with CUDA 9.0 nvprof" In my yesterday's commit 9ae129017c7fc1fa638d6beedd3802b515ca692b 'Fix og9 "Fix hang when running oacc exec with CUDA 9.0 nvprof"', I wrongly blamed the og9 "Fix hang when running oacc exec with CUDA 9.0 nvprof" to be responsible for the hang being back, when in fact it's only the later og9 "OpenACC Profiling Interface (incomplete)" backport from trunk that introduced the problem. The confusing thing was that the og9 "Fix hang when running oacc exec with CUDA 9.0 nvprof" commit appears *before* the og9 "OpenACC Profiling Interface (incomplete)" backport that it relates to. And, in addition to that, I pushed the wrong (incomplete) version of my fix. libgomp/ * oacc-init.c (acc_init_1): Move other 'acc_init_state' logic to where it belongs. --- libgomp/ChangeLog.omp | 5 +++++ libgomp/oacc-init.c | 12 ++++++++---- 2 files changed, 13 insertions(+), 4 deletions(-) diff --git a/libgomp/ChangeLog.omp b/libgomp/ChangeLog.omp index 75c459179981..922e00fbff5b 100644 --- a/libgomp/ChangeLog.omp +++ b/libgomp/ChangeLog.omp @@ -1,3 +1,8 @@ +2020-03-26 Thomas Schwinge + + * oacc-init.c (acc_init_1): Move other 'acc_init_state' logic to + where it belongs. + 2020-03-25 Thomas Schwinge * oacc-init.c (acc_init_1): Move 'acc_init_state' logic to where diff --git a/libgomp/oacc-init.c b/libgomp/oacc-init.c index 765fa2f3b952..40c14fa9bf2b 100644 --- a/libgomp/oacc-init.c +++ b/libgomp/oacc-init.c @@ -317,10 +317,6 @@ acc_init_1 (acc_device_t d, acc_construct_t parent_construct, int implicit) gomp_init_device (acc_dev); gomp_mutex_unlock (&acc_dev->lock); - gomp_mutex_lock (&acc_init_state_lock); - acc_init_state = initialized; - gomp_mutex_unlock (&acc_init_state_lock); - if (profiling_p) { prof_info.event_type = acc_ev_device_init_end; @@ -329,6 +325,14 @@ acc_init_1 (acc_device_t d, acc_construct_t parent_construct, int implicit) &api_info); } + /* We're setting 'initialized' *after* 'goacc_profiling_dispatch', so that a + nested 'acc_get_device_type' called from a profiling callback still sees + 'initializing', so that we don't deadlock when it then again tries to lock + 'goacc_prof_lock'. See also the discussion in 'acc_get_device_type'. */ + gomp_mutex_lock (&acc_init_state_lock); + acc_init_state = initialized; + gomp_mutex_unlock (&acc_init_state_lock); + return base_dev; } -- 2.47.2