bug#26441: Gnulib’s ‘test-lock’ fails to complete on machines with >= 32 cores

[Mathieu Othacehe]

Hi,

Here's a patch for gettext, to solve test-lock performance issues.

While rebuilding all gettext dependencies, I noticed that there are
other programs impacted by this bug :

* libunistring
* coreutils
* findutils

But the list may be bigger :

http://git.gnu.org.ua/cgit/gnulib.git/tree/users.txt

Those users are embedding gnulib source code (via git submodules for
example), so I'm not sure what to do here ...

We can consider that it's not such a big deal and ignore this problem
waiting for programs to integrate the fix, or we can patch them on
core-updates.

WDYT ?

Thank you,

Mathieu

>From e4ad9aa61fa75afa4417616de36cd25e9631fd67 Mon Sep 17 00:00:00 2001
From: Mathieu Othacehe <address@hidden>
Date: Wed, 12 Apr 2017 21:17:24 +0200
Subject: [PATCH] gnu: gettext: Fix make check issues on multi-core machines.

* gnu/packages/patches/gettext-gnulib-multi-core.patch: New file.
* gnu/packages/patches/gettext-multi-core.patch: New file.
* gnu/packages/gettext.scm (gettext-minimal)[patches]: Add a reference
to the two previous patches.
---
 gnu/packages/gettext.scm                           |   9 +-
 .../patches/gettext-gnulib-multi-core.patch        | 176 ++++++++++++++++++++
 gnu/packages/patches/gettext-multi-core.patch      | 183 +++++++++++++++++++++
 3 files changed, 367 insertions(+), 1 deletion(-)
 create mode 100644 gnu/packages/patches/gettext-gnulib-multi-core.patch
 create mode 100644 gnu/packages/patches/gettext-multi-core.patch

diff --git a/gnu/packages/gettext.scm b/gnu/packages/gettext.scm
index f583d1c2c..2a749e3a6 100644
--- a/gnu/packages/gettext.scm
+++ b/gnu/packages/gettext.scm
@@ -5,6 +5,7 @@
 ;;; Copyright © 2016 Efraim Flashner <address@hidden>
 ;;; Copyright © 2016 Alex Kost <address@hidden>
 ;;; Copyright © 2017 Marius Bakke <address@hidden>
+;;; Copyright © 2017 Mathieu Othacehe <address@hidden>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -45,7 +46,13 @@
                                  version ".tar.gz"))
              (sha256
               (base32
-               "0hsw28f9q9xaggjlsdp2qmbp2rbd1mp0njzan2ld9kiqwkq2m57z"))))
+               "0hsw28f9q9xaggjlsdp2qmbp2rbd1mp0njzan2ld9kiqwkq2m57z"))
+             ;; test-lock has performance issues on multi-core machines,
+             ;; it hangs or takes a long time to complete.
+             ;; There is one commit in gettext and one commit
+             ;; in gettext's embedded gnulib to fix this issue.
+             (patches (search-patches "gettext-multi-core.patch"
+                                      "gettext-gnulib-multi-core.patch"))))
     (build-system gnu-build-system)
     (outputs '("out"
                "doc"))                            ;8 MiB of HTML
diff --git a/gnu/packages/patches/gettext-gnulib-multi-core.patch 
b/gnu/packages/patches/gettext-gnulib-multi-core.patch
new file mode 100644
index 000000000..4393f3a1b
--- /dev/null
+++ b/gnu/packages/patches/gettext-gnulib-multi-core.patch
@@ -0,0 +1,176 @@
+This patch fixes performance problems on multi-core machines.
+See commit 480d374e596a0ee3fed168ab42cd84c313ad3c89 in gnulib
+from Bruno Haible <address@hidden>.
+
+diff --git a/gettext-tools/gnulib-tests/test-lock.c 
b/gettext-tools/gnulib-tests/test-lock.c
+index cb734b4e6..aa6de2739 100644
+--- a/gettext-tools/gnulib-tests/test-lock.c
++++ b/gettext-tools/gnulib-tests/test-lock.c
+@@ -50,6 +50,13 @@
+    Uncomment this to see if the operating system has a fair scheduler.  */
+ #define EXPLICIT_YIELD 1
+ 
++/* Whether to use 'volatile' on some variables that communicate information
++   between threads.  If set to 0, a lock is used to protect these variables.
++   If set to 1, 'volatile' is used; this is theoretically equivalent but can
++   lead to much slower execution (e.g. 30x slower total run time on a 40-core
++   machine.  */
++#define USE_VOLATILE 0
++
+ /* Whether to print debugging messages.  */
+ #define ENABLE_DEBUGGING 0
+ 
+@@ -103,6 +110,51 @@
+ # define yield()
+ #endif
+ 
++#if USE_VOLATILE
++struct atomic_int {
++  volatile int value;
++};
++static void
++init_atomic_int (struct atomic_int *ai)
++{
++}
++static int
++get_atomic_int_value (struct atomic_int *ai)
++{
++  return ai->value;
++}
++static void
++set_atomic_int_value (struct atomic_int *ai, int new_value)
++{
++  ai->value = new_value;
++}
++#else
++struct atomic_int {
++  gl_lock_define (, lock)
++  int value;
++};
++static void
++init_atomic_int (struct atomic_int *ai)
++{
++  gl_lock_init (ai->lock);
++}
++static int
++get_atomic_int_value (struct atomic_int *ai)
++{
++  gl_lock_lock (ai->lock);
++  int ret = ai->value;
++  gl_lock_unlock (ai->lock);
++  return ret;
++}
++static void
++set_atomic_int_value (struct atomic_int *ai, int new_value)
++{
++  gl_lock_lock (ai->lock);
++  ai->value = new_value;
++  gl_lock_unlock (ai->lock);
++}
++#endif
++
+ #define ACCOUNT_COUNT 4
+ 
+ static int account[ACCOUNT_COUNT];
+@@ -170,12 +222,12 @@ lock_mutator_thread (void *arg)
+   return NULL;
+ }
+ 
+-static volatile int lock_checker_done;
++static struct atomic_int lock_checker_done;
+ 
+ static void *
+ lock_checker_thread (void *arg)
+ {
+-  while (!lock_checker_done)
++  while (get_atomic_int_value (&lock_checker_done) == 0)
+     {
+       dbgprintf ("Checker %p before check lock\n", gl_thread_self_pointer ());
+       gl_lock_lock (my_lock);
+@@ -200,7 +252,8 @@ test_lock (void)
+   /* Initialization.  */
+   for (i = 0; i < ACCOUNT_COUNT; i++)
+     account[i] = 1000;
+-  lock_checker_done = 0;
++  init_atomic_int (&lock_checker_done);
++  set_atomic_int_value (&lock_checker_done, 0);
+ 
+   /* Spawn the threads.  */
+   checkerthread = gl_thread_create (lock_checker_thread, NULL);
+@@ -210,7 +263,7 @@ test_lock (void)
+   /* Wait for the threads to terminate.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (threads[i], NULL);
+-  lock_checker_done = 1;
++  set_atomic_int_value (&lock_checker_done, 1);
+   gl_thread_join (checkerthread, NULL);
+   check_accounts ();
+ }
+@@ -254,12 +307,12 @@ rwlock_mutator_thread (void *arg)
+   return NULL;
+ }
+ 
+-static volatile int rwlock_checker_done;
++static struct atomic_int rwlock_checker_done;
+ 
+ static void *
+ rwlock_checker_thread (void *arg)
+ {
+-  while (!rwlock_checker_done)
++  while (get_atomic_int_value (&rwlock_checker_done) == 0)
+     {
+       dbgprintf ("Checker %p before check rdlock\n", gl_thread_self_pointer 
());
+       gl_rwlock_rdlock (my_rwlock);
+@@ -284,7 +337,8 @@ test_rwlock (void)
+   /* Initialization.  */
+   for (i = 0; i < ACCOUNT_COUNT; i++)
+     account[i] = 1000;
+-  rwlock_checker_done = 0;
++  init_atomic_int (&rwlock_checker_done);
++  set_atomic_int_value (&rwlock_checker_done, 0);
+ 
+   /* Spawn the threads.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+@@ -295,7 +349,7 @@ test_rwlock (void)
+   /* Wait for the threads to terminate.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (threads[i], NULL);
+-  rwlock_checker_done = 1;
++  set_atomic_int_value (&rwlock_checker_done, 1);
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (checkerthreads[i], NULL);
+   check_accounts ();
+@@ -356,12 +410,12 @@ reclock_mutator_thread (void *arg)
+   return NULL;
+ }
+ 
+-static volatile int reclock_checker_done;
++static struct atomic_int reclock_checker_done;
+ 
+ static void *
+ reclock_checker_thread (void *arg)
+ {
+-  while (!reclock_checker_done)
++  while (get_atomic_int_value (&reclock_checker_done) == 0)
+     {
+       dbgprintf ("Checker %p before check lock\n", gl_thread_self_pointer ());
+       gl_recursive_lock_lock (my_reclock);
+@@ -386,7 +440,8 @@ test_recursive_lock (void)
+   /* Initialization.  */
+   for (i = 0; i < ACCOUNT_COUNT; i++)
+     account[i] = 1000;
+-  reclock_checker_done = 0;
++  init_atomic_int (&reclock_checker_done);
++  set_atomic_int_value (&reclock_checker_done, 0);
+ 
+   /* Spawn the threads.  */
+   checkerthread = gl_thread_create (reclock_checker_thread, NULL);
+@@ -396,7 +451,7 @@ test_recursive_lock (void)
+   /* Wait for the threads to terminate.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (threads[i], NULL);
+-  reclock_checker_done = 1;
++  set_atomic_int_value (&reclock_checker_done, 1);
+   gl_thread_join (checkerthread, NULL);
+   check_accounts ();
+ }
diff --git a/gnu/packages/patches/gettext-multi-core.patch 
b/gnu/packages/patches/gettext-multi-core.patch
new file mode 100644
index 000000000..3c7f0c759
--- /dev/null
+++ b/gnu/packages/patches/gettext-multi-core.patch
@@ -0,0 +1,183 @@
+This patch fixes performance problems on multi-core machines.
+See commit 1afbcb06fded2a427b761dd1615b1e48e1e853cc in gettext
+from Bruno Haible <address@hidden>.
+
+diff --git a/gettext-runtime/tests/test-lock.c 
b/gettext-runtime/tests/test-lock.c
+index d279d1d60..51cec3d6b 100644
+--- a/gettext-runtime/tests/test-lock.c
++++ b/gettext-runtime/tests/test-lock.c
+@@ -1,5 +1,5 @@
+ /* Test of locking in multithreaded situations.
+-   Copyright (C) 2005, 2008-2016 Free Software Foundation, Inc.
++   Copyright (C) 2005, 2008-2017 Free Software Foundation, Inc.
+ 
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Lesser General Public License as published by
+@@ -50,6 +50,13 @@
+    Uncomment this to see if the operating system has a fair scheduler.  */
+ #define EXPLICIT_YIELD 1
+ 
++/* Whether to use 'volatile' on some variables that communicate information
++   between threads.  If set to 0, a lock is used to protect these variables.
++   If set to 1, 'volatile' is used; this is theoretically equivalent but can
++   lead to much slower execution (e.g. 30x slower total run time on a 40-core
++   machine.  */
++#define USE_VOLATILE 0
++
+ /* Whether to print debugging messages.  */
+ #define ENABLE_DEBUGGING 0
+ 
+@@ -214,6 +221,51 @@ static inline void * gl_thread_self_pointer (void)
+ # define yield()
+ #endif
+ 
++#if USE_VOLATILE
++struct atomic_int {
++  volatile int value;
++};
++static void
++init_atomic_int (struct atomic_int *ai)
++{
++}
++static int
++get_atomic_int_value (struct atomic_int *ai)
++{
++  return ai->value;
++}
++static void
++set_atomic_int_value (struct atomic_int *ai, int new_value)
++{
++  ai->value = new_value;
++}
++#else
++struct atomic_int {
++  gl_lock_define (, lock)
++  int value;
++};
++static void
++init_atomic_int (struct atomic_int *ai)
++{
++  gl_lock_init (ai->lock);
++}
++static int
++get_atomic_int_value (struct atomic_int *ai)
++{
++  gl_lock_lock (ai->lock);
++  int ret = ai->value;
++  gl_lock_unlock (ai->lock);
++  return ret;
++}
++static void
++set_atomic_int_value (struct atomic_int *ai, int new_value)
++{
++  gl_lock_lock (ai->lock);
++  ai->value = new_value;
++  gl_lock_unlock (ai->lock);
++}
++#endif
++
+ #define ACCOUNT_COUNT 4
+ 
+ static int account[ACCOUNT_COUNT];
+@@ -281,12 +333,12 @@ lock_mutator_thread (void *arg)
+   return NULL;
+ }
+ 
+-static volatile int lock_checker_done;
++static struct atomic_int lock_checker_done;
+ 
+ static void *
+ lock_checker_thread (void *arg)
+ {
+-  while (!lock_checker_done)
++  while (get_atomic_int_value (&lock_checker_done) == 0)
+     {
+       dbgprintf ("Checker %p before check lock\n", gl_thread_self_pointer ());
+       gl_lock_lock (my_lock);
+@@ -311,7 +363,8 @@ test_lock (void)
+   /* Initialization.  */
+   for (i = 0; i < ACCOUNT_COUNT; i++)
+     account[i] = 1000;
+-  lock_checker_done = 0;
++  init_atomic_int (&lock_checker_done);
++  set_atomic_int_value (&lock_checker_done, 0);
+ 
+   /* Spawn the threads.  */
+   checkerthread = gl_thread_create (lock_checker_thread, NULL);
+@@ -321,7 +374,7 @@ test_lock (void)
+   /* Wait for the threads to terminate.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (threads[i], NULL);
+-  lock_checker_done = 1;
++  set_atomic_int_value (&lock_checker_done, 1);
+   gl_thread_join (checkerthread, NULL);
+   check_accounts ();
+ }
+@@ -365,12 +418,12 @@ rwlock_mutator_thread (void *arg)
+   return NULL;
+ }
+ 
+-static volatile int rwlock_checker_done;
++static struct atomic_int rwlock_checker_done;
+ 
+ static void *
+ rwlock_checker_thread (void *arg)
+ {
+-  while (!rwlock_checker_done)
++  while (get_atomic_int_value (&rwlock_checker_done) == 0)
+     {
+       dbgprintf ("Checker %p before check rdlock\n", gl_thread_self_pointer 
());
+       gl_rwlock_rdlock (my_rwlock);
+@@ -395,7 +448,8 @@ test_rwlock (void)
+   /* Initialization.  */
+   for (i = 0; i < ACCOUNT_COUNT; i++)
+     account[i] = 1000;
+-  rwlock_checker_done = 0;
++  init_atomic_int (&rwlock_checker_done);
++  set_atomic_int_value (&rwlock_checker_done, 0);
+ 
+   /* Spawn the threads.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+@@ -406,7 +460,7 @@ test_rwlock (void)
+   /* Wait for the threads to terminate.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (threads[i], NULL);
+-  rwlock_checker_done = 1;
++  set_atomic_int_value (&rwlock_checker_done, 1);
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (checkerthreads[i], NULL);
+   check_accounts ();
+@@ -467,12 +521,12 @@ reclock_mutator_thread (void *arg)
+   return NULL;
+ }
+ 
+-static volatile int reclock_checker_done;
++static struct atomic_int reclock_checker_done;
+ 
+ static void *
+ reclock_checker_thread (void *arg)
+ {
+-  while (!reclock_checker_done)
++  while (get_atomic_int_value (&reclock_checker_done) == 0)
+     {
+       dbgprintf ("Checker %p before check lock\n", gl_thread_self_pointer ());
+       gl_recursive_lock_lock (my_reclock);
+@@ -497,7 +551,8 @@ test_recursive_lock (void)
+   /* Initialization.  */
+   for (i = 0; i < ACCOUNT_COUNT; i++)
+     account[i] = 1000;
+-  reclock_checker_done = 0;
++  init_atomic_int (&reclock_checker_done);
++  set_atomic_int_value (&reclock_checker_done, 0);
+ 
+   /* Spawn the threads.  */
+   checkerthread = gl_thread_create (reclock_checker_thread, NULL);
+@@ -507,7 +562,7 @@ test_recursive_lock (void)
+   /* Wait for the threads to terminate.  */
+   for (i = 0; i < THREAD_COUNT; i++)
+     gl_thread_join (threads[i], NULL);
+-  reclock_checker_done = 1;
++  set_atomic_int_value (&reclock_checker_done, 1);
+   gl_thread_join (checkerthread, NULL);
+   check_accounts ();
+ }
-- 
2.11.1