Merge commit 'refs/top-bases/linux-configure-branch' into linux-configure-branch

6 files changed, 136 insertions, 135 deletions

diff --git a/ChangeLog b/ChangeLog
index 2e9d58288..73d92055d 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -54,9 +54,6 @@
 	* module/zcommon/zfs_prop.c: Export new quota related symbols.
 	* lib/libzfs/libzfs_util.c: Increase buffer size for nvlist which
 	is needed for large configurations.
-	* lib/libzpool/kernel.c: Reimplement user kthreads based on pthread
-	thread specific data.  This resolves previous scalabily concerns about
-	the cost of calling curthread which previously required a list walk.
 
 	* Bug fixes:
 	* lib/libspl/include/sys/zfs_debug.h: Removed duplicate file.
diff --git a/cmd/ztest/ztest.c b/cmd/ztest/ztest.c
index 64e6eb690..7da8f1f4d 100644
--- a/cmd/ztest/ztest.c
+++ b/cmd/ztest/ztest.c
@@ -141,6 +141,7 @@ typedef struct ztest_args {
 	objset_t	*za_os;
 	zilog_t		*za_zilog;
 	kthread_t	*za_thread;
+	kt_did_t	za_threadid;
 	uint64_t	za_instance;
 	uint64_t	za_random;
 	uint64_t	za_diroff;
@@ -156,7 +157,6 @@ typedef struct ztest_args {
 	ztest_block_tag_t za_wbt;
 	dmu_object_info_t za_doi;
 	dmu_buf_t	*za_dbuf;
-	boolean_t	za_exited;
 } ztest_args_t;
 
 typedef void ztest_func_t(ztest_args_t *);
@@ -253,8 +253,6 @@ typedef struct ztest_shared {
 	kmutex_t		zs_sync_lock[ZTEST_SYNC_LOCKS];
 	uint64_t		zs_seq[ZTEST_SYNC_LOCKS];
 	ztest_cb_list_t		zs_cb_list;
-	kmutex_t		zs_thr_lock;
-	kcondvar_t		zs_thr_cv;
 } ztest_shared_t;
 
 static char ztest_dev_template[] = "%s/%s.%llua";
@@ -266,7 +264,6 @@ static int ztest_dump_core = 1;
 
 static uint64_t metaslab_sz;
 static boolean_t ztest_exiting;
-static boolean_t resume_thr_exited;
 
 extern uint64_t metaslab_gang_bang;
 extern uint64_t metaslab_df_alloc_threshold;
@@ -3812,8 +3809,6 @@ ztest_resume_thread(void *arg)
 		ztest_resume(spa);
 	}
 
-	resume_thr_exited = B_TRUE;
-
 	thread_exit();
 	return (NULL);
 }
@@ -3879,13 +3874,6 @@ ztest_thread(void *arg)
 			break;
 	}
 
-	mutex_enter(&zs->zs_thr_lock);
-	za->za_exited = B_TRUE;
-	mutex_exit(&zs->zs_thr_lock);
-
-	/* Announce that the thread has finished */
-	cv_broadcast(&zs->zs_thr_cv);
-
 	thread_exit();
 	return (NULL);
 }
@@ -3902,14 +3890,13 @@ ztest_run(char *pool)
 	spa_t *spa;
 	char name[100];
 	kthread_t *resume_thread;
+	kt_did_t resume_id;
 
 	ztest_exiting = B_FALSE;
 
 	mutex_init(&zs->zs_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
 	rw_init(&zs->zs_name_lock, NULL, RW_DEFAULT, NULL);
 	mutex_init(&zs->zs_cb_list.zcl_callbacks_lock,NULL,MUTEX_DEFAULT,NULL);
-	mutex_init(&zs->zs_thr_lock, NULL, MUTEX_DEFAULT, NULL);
-	cv_init(&zs->zs_thr_cv, NULL, CV_DEFAULT, NULL);
 
 	list_create(&zs->zs_cb_list.zcl_callbacks, sizeof (ztest_cb_data_t),
 	    offsetof(ztest_cb_data_t, zcd_node));
@@ -3982,9 +3969,9 @@ ztest_run(char *pool)
 	/*
 	 * Create a thread to periodically resume suspended I/O.
 	 */
-	resume_thr_exited = B_FALSE;
 	VERIFY3P((resume_thread = thread_create(NULL, 0, ztest_resume_thread,
-	    spa, TS_RUN, NULL, 0, 0)), !=, NULL);
+	         spa, THR_BOUND, NULL, 0, 0)), !=, NULL);
+	resume_id = resume_thread->t_tid;
 
 	/*
 	 * Verify that we can safely inquire about about any object,
@@ -4060,18 +4047,13 @@ ztest_run(char *pool)
 			za[d].za_zilog = zil_open(za[d].za_os, NULL);
 		}
 
-		za[t].za_exited = B_FALSE;
-
 		VERIFY3P((za[t].za_thread = thread_create(NULL, 0, ztest_thread,
-		    &za[t], TS_RUN, NULL, 0, 0)), !=, NULL);
+		         &za[t], THR_BOUND, NULL, 0, 0)), !=, NULL);
+		za[t].za_threadid = za[t].za_thread->t_tid;
 	}
 
 	while (--t >= 0) {
-		mutex_enter(&zs->zs_thr_lock);
-		while (!za[t].za_exited)
-			cv_wait(&zs->zs_thr_cv, &zs->zs_thr_lock);
-		mutex_exit(&zs->zs_thr_lock);
-
+		VERIFY(thread_join(za[t].za_threadid, NULL, NULL) == 0);
 		if (t < zopt_datasets) {
 			zil_close(za[t].za_zilog);
 			dmu_objset_close(za[t].za_os);
@@ -4110,11 +4092,7 @@ ztest_run(char *pool)
 
 	/* Kill the resume thread */
 	ztest_exiting = B_TRUE;
-
-	/* Wait for the resume thread to exit */
-	while (!resume_thr_exited)
-		(void) poll(NULL, 0, 200);
-
+	VERIFY(thread_join(resume_id, NULL, NULL) == 0);
 	ztest_resume(spa);
 
 	/*
@@ -4130,8 +4108,6 @@ ztest_run(char *pool)
 
 	list_destroy(&zs->zs_cb_list.zcl_callbacks);
 
-	cv_destroy(&zs->zs_thr_cv);
-	mutex_destroy(&zs->zs_thr_lock);
 	mutex_destroy(&zs->zs_cb_list.zcl_callbacks_lock);
 	rw_destroy(&zs->zs_name_lock);
 	mutex_destroy(&zs->zs_vdev_lock);
diff --git a/lib/libzpool/include/sys/zfs_context.h b/lib/libzpool/include/sys/zfs_context.h
index 0475ce093..9377dab2f 100644
--- a/lib/libzpool/include/sys/zfs_context.h
+++ b/lib/libzpool/include/sys/zfs_context.h
@@ -153,34 +153,34 @@ extern void vpanic(const char *, __va_list);
 /*
  * Threads
  */
+#define THR_BOUND		0x00000001
 #define TS_RUN			0x00000002
 
-#define STACK_SIZE 8192 /* x86/x64 */
-
-typedef void (*thread_func_t)(void);
-typedef void (*thread_func_arg_t)(void *);
+typedef void (*thread_func_t)(void *);
 typedef pthread_t kt_did_t;
 
 typedef struct kthread {
+        list_node_t	t_node;
 	kt_did_t	t_tid;
-	thread_func_t	t_func;
-	void *		t_arg;
+	pthread_attr_t	t_attr;
 } kthread_t;
 
-/* XXX tsd_create()/tsd_destroy() missing */
 #define	tsd_get(key)		pthread_getspecific(key)
 #define	tsd_set(key, val)	pthread_setspecific(key, val)
 #define	curthread		zk_thread_current()
 #define thread_exit		zk_thread_exit
 #define thread_create(stk, stksize, func, arg, len, pp, state, pri)	\
-	zk_thread_create(stk, stksize, (thread_func_t) func, arg, len,  \
-	    NULL, state, pri)
+	zk_thread_create(stk, stksize, (thread_func_t)func, arg,	\
+			 len, NULL, state, pri)
+#define thread_join(tid, dtid, status)					\
+	zk_thread_join(tid, dtid, status)
 
 extern kthread_t *zk_thread_current(void);
 extern void zk_thread_exit(void);
 extern kthread_t *zk_thread_create(caddr_t stk, size_t  stksize,
 	thread_func_t func, void *arg, size_t len,
 	void *pp, int state, pri_t pri);
+extern int zk_thread_join(kt_did_t tid, kthread_t *dtid, void **status);
 
 #define	issig(why)	(FALSE)
 #define	ISSIG(thr, why)	(FALSE)
@@ -315,7 +315,7 @@ extern taskq_t	*taskq_create(const char *, int, pri_t, int, int, uint_t);
 extern taskqid_t taskq_dispatch(taskq_t *, task_func_t, void *, uint_t);
 extern void	taskq_destroy(taskq_t *);
 extern void	taskq_wait(taskq_t *);
-extern int	taskq_member(taskq_t *, kthread_t *);
+extern int	taskq_member(taskq_t *, void *);
 extern void	system_taskq_init(void);
 extern void	system_taskq_fini(void);
 
diff --git a/lib/libzpool/kernel.c b/lib/libzpool/kernel.c
index fcd0f2871..ab97636ba 100644
--- a/lib/libzpool/kernel.c
+++ b/lib/libzpool/kernel.c
@@ -57,141 +57,155 @@ struct utsname utsname = {
  * =========================================================================
  */
 
+/* NOTE: Tracking each tid on a list and using it for curthread lookups
+ *       is slow at best but it provides an easy way to provide a kthread
+ *       style API on top of pthreads.  For now we just want ztest to work
+ *       to validate correctness.  Performance is not much of an issue
+ *       since that is what the in-kernel version is for.  That said
+ *       reworking this to track the kthread_t structure as thread
+ *       specific data would be probably the best way to speed this up.
+ */
+
 pthread_cond_t kthread_cond = PTHREAD_COND_INITIALIZER;
 pthread_mutex_t kthread_lock = PTHREAD_MUTEX_INITIALIZER;
-pthread_key_t kthread_key;
-int kthread_nr = 0;
+list_t kthread_list;
+
+static int
+thread_count(void)
+{
+	kthread_t *kt;
+	int count = 0;
+
+	for (kt = list_head(&kthread_list); kt != NULL;
+	     kt = list_next(&kthread_list, kt))
+		count++;
+
+	return count;
+}
 
 static void
 thread_init(void)
 {
 	kthread_t *kt;
 
-	VERIFY3S(pthread_key_create(&kthread_key, NULL), ==, 0);
+	/* Initialize list for tracking kthreads */
+	list_create(&kthread_list, sizeof (kthread_t),
+		    offsetof(kthread_t, t_node));
 
 	/* Create entry for primary kthread */
 	kt = umem_zalloc(sizeof(kthread_t), UMEM_NOFAIL);
-	kt->t_tid = pthread_self();
-	kt->t_func = NULL;
-
-	VERIFY3S(pthread_setspecific(kthread_key, kt), ==, 0);
-
-	/* Only the main thread should be running at the moment */
-	ASSERT3S(kthread_nr, ==, 0);
-	kthread_nr = 1;
+	list_link_init(&kt->t_node);
+	VERIFY3U(kt->t_tid = pthread_self(), !=, 0);
+        VERIFY3S(pthread_attr_init(&kt->t_attr), ==, 0);
+	VERIFY3S(pthread_mutex_lock(&kthread_lock), ==, 0);
+	list_insert_head(&kthread_list, kt);
+	VERIFY3S(pthread_mutex_unlock(&kthread_lock), ==, 0);
 }
 
 static void
 thread_fini(void)
 {
-	kthread_t *kt = curthread;
-
-	ASSERT(pthread_equal(kt->t_tid, pthread_self()));
-	ASSERT3P(kt->t_func, ==, NULL);
-
-	umem_free(kt, sizeof(kthread_t));
+	kthread_t *kt;
+	struct timespec ts = { 0 };
+	int count;
 
 	/* Wait for all threads to exit via thread_exit() */
 	VERIFY3S(pthread_mutex_lock(&kthread_lock), ==, 0);
+	while ((count = thread_count()) > 1) {
+		clock_gettime(CLOCK_REALTIME, &ts);
+		ts.tv_sec += 1;
+		pthread_cond_timedwait(&kthread_cond, &kthread_lock, &ts);
+	}
 
-	kthread_nr--; /* Main thread is exiting */
-
-	while (kthread_nr > 0)
-		VERIFY3S(pthread_cond_wait(&kthread_cond, &kthread_lock), ==,
-		    0);
-
-	ASSERT3S(kthread_nr, ==, 0);
+	ASSERT3S(thread_count(), ==, 1);
+	kt = list_head(&kthread_list);
+	list_remove(&kthread_list, kt);
 	VERIFY3S(pthread_mutex_unlock(&kthread_lock), ==, 0);
 
-	VERIFY3S(pthread_key_delete(kthread_key), ==, 0);
+	VERIFY(pthread_attr_destroy(&kt->t_attr) == 0);
+	umem_free(kt, sizeof(kthread_t));
+
+	/* Cleanup list for tracking kthreads */
+	list_destroy(&kthread_list);
 }
 
 kthread_t *
 zk_thread_current(void)
 {
-	kthread_t *kt = pthread_getspecific(kthread_key);
-
-	ASSERT3P(kt, !=, NULL);
-
-	return kt;
-}
-
-void *
-zk_thread_helper(void *arg)
-{
-	kthread_t *kt = (kthread_t *) arg;
-
-	VERIFY3S(pthread_setspecific(kthread_key, kt), ==, 0);
-
-	VERIFY3S(pthread_mutex_lock(&kthread_lock), ==, 0);
-	kthread_nr++;
-	VERIFY3S(pthread_mutex_unlock(&kthread_lock), ==, 0);
-
-	kt->t_tid = pthread_self();
-	((thread_func_arg_t) kt->t_func)(kt->t_arg);
+	kt_did_t tid = pthread_self();
+	kthread_t *kt;
+	int count = 1;
 
-	/* Unreachable, thread must exit with thread_exit() */
-	abort();
+	/*
+	 * Because a newly created thread may call zk_thread_current()
+	 * before the thread parent has had time to add the thread's tid
+	 * to our lookup list.  We will loop as long as there are tid
+	 * which have not yet been set which must be one of ours.
+	 * Yes it's a hack, at some point we can just use native pthreads.
+	 */
+	while (count > 0) {
+		count = 0;
+		VERIFY3S(pthread_mutex_lock(&kthread_lock), ==, 0);
+		for (kt = list_head(&kthread_list); kt != NULL;
+		     kt = list_next(&kthread_list, kt)) {
+
+			if (kt->t_tid == tid) {
+				VERIFY3S(pthread_mutex_unlock(
+				         &kthread_lock), ==, 0);
+				return kt;
+			}
+
+			if (kt->t_tid == (kt_did_t)-1)
+				count++;
+		}
+		VERIFY3S(pthread_mutex_unlock(&kthread_lock), ==, 0);
+	}
 
+	/* Unreachable */
+	ASSERT(0);
 	return NULL;
 }
 
 kthread_t *
-zk_thread_create(caddr_t stk, size_t stksize, thread_func_t func, void *arg,
+zk_thread_create(caddr_t stk, size_t  stksize, thread_func_t func, void *arg,
 	      size_t len, void *pp, int state, pri_t pri)
 {
 	kthread_t *kt;
-	pthread_t tid;
-	pthread_attr_t attr;
-	size_t stack;
-
-	/*
-	 * Due to a race when getting/setting the thread ID, currently only
-	 * detached threads are supported.
-	 */
-	ASSERT3S(state & ~TS_RUN, ==, 0);
 
 	kt = umem_zalloc(sizeof(kthread_t), UMEM_NOFAIL);
-	kt->t_func = func;
-	kt->t_arg = arg;
-
-	/*
-	 * The Solaris kernel stack size in x86/x64 is 8K, so we reduce the
-	 * default stack size in userspace, for sanity checking.
-	 *
-	 * PTHREAD_STACK_MIN is the stack required for a NULL procedure in
-	 * userspace.
-	 *
-	 * XXX: Stack size for other architectures is not being taken into
-	 * account.
-	 */
-	stack = PTHREAD_STACK_MIN + MAX(stksize, STACK_SIZE);
-	
-	VERIFY3S(pthread_attr_init(&attr), ==, 0);
-	VERIFY3S(pthread_attr_setstacksize(&attr, stack), ==, 0);
-	VERIFY3S(pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED),
-	    ==, 0);
+	kt->t_tid = (kt_did_t)-1;
+	list_link_init(&kt->t_node);
+	VERIFY(pthread_attr_init(&kt->t_attr) == 0);
 
-	VERIFY3S(pthread_create(&tid, &attr, &zk_thread_helper, kt), ==, 0);
+	VERIFY3S(pthread_mutex_lock(&kthread_lock), ==, 0);
+	list_insert_head(&kthread_list, kt);
+	VERIFY3S(pthread_mutex_unlock(&kthread_lock), ==, 0);
 
-	VERIFY3S(pthread_attr_destroy(&attr), ==, 0);
+	VERIFY3U(pthread_create(&kt->t_tid, &kt->t_attr,
+			      (void *(*)(void *))func, arg), ==, 0);
 
 	return kt;
 }
 
+int
+zk_thread_join(kt_did_t tid, kthread_t *dtid, void **status)
+{
+	return pthread_join(tid, status);
+}
+
 void
 zk_thread_exit(void)
 {
-	kthread_t *kt = curthread;
+	kthread_t *kt;
 
-	ASSERT(pthread_equal(kt->t_tid, pthread_self()));
+	VERIFY3P(kt = curthread, !=, NULL);
+	VERIFY3S(pthread_mutex_lock(&kthread_lock), ==, 0);
+	list_remove(&kthread_list, kt);
+	VERIFY3S(pthread_mutex_unlock(&kthread_lock), ==, 0);
 
+	VERIFY(pthread_attr_destroy(&kt->t_attr) == 0);
 	umem_free(kt, sizeof(kthread_t));
 
-	pthread_mutex_lock(&kthread_lock);
-	kthread_nr--;
-	pthread_mutex_unlock(&kthread_lock);
-
 	pthread_cond_broadcast(&kthread_cond);
 	pthread_exit(NULL);
 }
diff --git a/lib/libzpool/taskq.c b/lib/libzpool/taskq.c
index 1efdf1d6f..42e2dd3f4 100644
--- a/lib/libzpool/taskq.c
+++ b/lib/libzpool/taskq.c
@@ -43,6 +43,7 @@ struct taskq {
 	kcondvar_t	tq_dispatch_cv;
 	kcondvar_t	tq_wait_cv;
 	kthread_t	**tq_threadlist;
+	kt_did_t	*tq_idlist;
 	int		tq_flags;
 	int		tq_active;
 	int		tq_nthreads;
@@ -134,7 +135,7 @@ taskq_wait(taskq_t *tq)
 	mutex_exit(&tq->tq_lock);
 }
 
-static void
+static void *
 taskq_thread(void *arg)
 {
 	taskq_t *tq = arg;
@@ -164,6 +165,7 @@ taskq_thread(void *arg)
 	cv_broadcast(&tq->tq_wait_cv);
 	mutex_exit(&tq->tq_lock);
 	thread_exit();
+	return (NULL);
 }
 
 /*ARGSUSED*/
@@ -198,8 +200,10 @@ taskq_create(const char *name, int nthreads, pri_t pri,
 	tq->tq_maxalloc = maxalloc;
 	tq->tq_task.task_next = &tq->tq_task;
 	tq->tq_task.task_prev = &tq->tq_task;
-	tq->tq_threadlist = kmem_alloc(tq->tq_nthreads * sizeof(kthread_t *),
-	    KM_SLEEP);
+	VERIFY3P((tq->tq_threadlist = kmem_alloc(tq->tq_nthreads *
+	         sizeof(kthread_t *), KM_SLEEP)), !=, NULL);
+	VERIFY3P((tq->tq_idlist = kmem_alloc(tq->tq_nthreads *
+	         sizeof(kt_did_t), KM_SLEEP)), !=, NULL);
 
 	if (flags & TASKQ_PREPOPULATE) {
 		mutex_enter(&tq->tq_lock);
@@ -210,7 +214,8 @@ taskq_create(const char *name, int nthreads, pri_t pri,
 
 	for (t = 0; t < tq->tq_nthreads; t++) {
 		VERIFY((tq->tq_threadlist[t] = thread_create(NULL, 0,
-		    taskq_thread, tq, TS_RUN, NULL, 0, 0)) != NULL);
+		       taskq_thread, tq, THR_BOUND, NULL, 0, 0)) != NULL);
+		tq->tq_idlist[t] = tq->tq_threadlist[t]->t_tid;
 	}
 
 	return (tq);
@@ -219,6 +224,7 @@ taskq_create(const char *name, int nthreads, pri_t pri,
 void
 taskq_destroy(taskq_t *tq)
 {
+	int t;
 	int nthreads = tq->tq_nthreads;
 
 	taskq_wait(tq);
@@ -239,7 +245,11 @@ taskq_destroy(taskq_t *tq)
 
 	mutex_exit(&tq->tq_lock);
 
+	for (t = 0; t < nthreads; t++)
+		VERIFY3S(thread_join(tq->tq_idlist[t], NULL, NULL), ==, 0);
+
 	kmem_free(tq->tq_threadlist, nthreads * sizeof(kthread_t *));
+	kmem_free(tq->tq_idlist, nthreads * sizeof(kt_did_t));
 
 	rw_destroy(&tq->tq_threadlock);
 	mutex_destroy(&tq->tq_lock);
@@ -250,7 +260,7 @@ taskq_destroy(taskq_t *tq)
 }
 
 int
-taskq_member(taskq_t *tq, kthread_t *t)
+taskq_member(taskq_t *tq, void *t)
 {
 	int i;
 
@@ -258,7 +268,7 @@ taskq_member(taskq_t *tq, kthread_t *t)
 		return (1);
 
 	for (i = 0; i < tq->tq_nthreads; i++)
-		if (tq->tq_threadlist[i] == t)
+		if (tq->tq_threadlist[i] == (kthread_t *)t)
 			return (1);
 
 	return (0);
diff --git a/module/zfs/txg.c b/module/zfs/txg.c
index afc4c8332..3d82990f5 100644
--- a/module/zfs/txg.c
+++ b/module/zfs/txg.c
@@ -446,6 +446,8 @@ txg_sync_thread(dsl_pool_t *dp)
 		rw_exit(&tx->tx_suspend);
 		cv_broadcast(&tx->tx_sync_done_cv);
 	}
+
+	thread_exit();
 }
 
 static void
@@ -490,6 +492,8 @@ txg_quiesce_thread(dsl_pool_t *dp)
 		cv_broadcast(&tx->tx_sync_more_cv);
 		cv_broadcast(&tx->tx_quiesce_done_cv);
 	}
+
+	thread_exit();
 }
 
 /*