Received-SPF: pass (google.com: domain of linux-kernel+bounces-66140-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
From: Elizabeth Figura <zfigura@codeweavers.com>
To: Arnd Bergmann <arnd@arndb.de>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Jonathan Corbet <corbet@lwn.net>,
	Shuah Khan <shuah@kernel.org>
Cc: linux-kernel@vger.kernel.org,
	linux-api@vger.kernel.org,
	wine-devel@winehq.org,
	=?UTF-8?q?Andr=C3=A9=20Almeida?= <andrealmeid@igalia.com>,
	Wolfram Sang <wsa@kernel.org>,
	Arkadiusz Hiler <ahiler@codeweavers.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Andy Lutomirski <luto@kernel.org>,
	linux-doc@vger.kernel.org,
	linux-kselftest@vger.kernel.org,
	Elizabeth Figura <zfigura@codeweavers.com>
Subject: [PATCH 05/31] ntsync: Introduce NTSYNC_IOC_WAIT_ALL.
Date: Wed, 14 Feb 2024 17:36:41 -0600
Message-ID: <20240214233645.9273-6-zfigura@codeweavers.com>
In-Reply-To: <20240214233645.9273-1-zfigura@codeweavers.com>
References: <20240214233645.9273-1-zfigura@codeweavers.com>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit

This is similar to NTSYNC_IOC_WAIT_ANY, but waits until all of the objects are
simultaneously signaled, and then acquires all of them as a single atomic
operation.

Signed-off-by: Elizabeth Figura <zfigura@codeweavers.com>
---
 drivers/misc/ntsync.c       | 242 ++++++++++++++++++++++++++++++++++--
 include/uapi/linux/ntsync.h |   1 +
 2 files changed, 235 insertions(+), 8 deletions(-)

diff --git a/drivers/misc/ntsync.c b/drivers/misc/ntsync.c
index 123af239f52e..e173774eea52 100644
--- a/drivers/misc/ntsync.c
+++ b/drivers/misc/ntsync.c
@@ -42,7 +42,34 @@ struct ntsync_obj {
 		} sem;
 	} u;
 
+	/*
+	 * any_waiters is protected by the object lock, but all_waiters is
+	 * protected by the device wait_all_lock.
+	 */
 	struct list_head any_waiters;
+	struct list_head all_waiters;
+
+	/*
+	 * Hint describing how many tasks are queued on this object in a
+	 * wait-all operation.
+	 *
+	 * Any time we do a wake, we may need to wake "all" waiters as well as
+	 * "any" waiters. In order to atomically wake "all" waiters, we must
+	 * lock all of the objects, and that means grabbing the wait_all_lock
+	 * below (and, due to lock ordering rules, before locking this object).
+	 * However, wait-all is a rare operation, and grabbing the wait-all
+	 * lock for every wake would create unnecessary contention.
+	 * Therefore we first check whether all_hint is zero, and, if it is,
+	 * we skip trying to wake "all" waiters.
+	 *
+	 * This hint isn't protected by any lock. It might change during the
+	 * course of a wake, but there's no meaningful race there; it's only a
+	 * hint.
+	 *
+	 * Since wait requests must originate from user-space threads, we're
+	 * limited here by PID_MAX_LIMIT, so there's no risk of overflow.
+	 */
+	atomic_t all_hint;
 };
 
 struct ntsync_q_entry {
@@ -63,14 +90,99 @@ struct ntsync_q {
 	 */
 	atomic_t signaled;
 
+	bool all;
 	__u32 count;
 	struct ntsync_q_entry entries[];
 };
 
 struct ntsync_device {
+	/*
+	 * Wait-all operations must atomically grab all objects, and be totally
+	 * ordered with respect to each other and wait-any operations.
+	 * If one thread is trying to acquire several objects, another thread
+	 * cannot touch the object at the same time.
+	 *
+	 * We achieve this by grabbing multiple object locks at the same time.
+	 * However, this creates a lock ordering problem. To solve that problem,
+	 * wait_all_lock is taken first whenever multiple objects must be locked
+	 * at the same time.
+	 */
+	spinlock_t wait_all_lock;
+
 	struct file *file;
 };
 
+static bool is_signaled(struct ntsync_obj *obj, __u32 owner)
+{
+	lockdep_assert_held(&obj->lock);
+
+	switch (obj->type) {
+	case NTSYNC_TYPE_SEM:
+		return !!obj->u.sem.count;
+	}
+
+	WARN(1, "bad object type %#x\n", obj->type);
+	return false;
+}
+
+/*
+ * "locked_obj" is an optional pointer to an object which is already locked and
+ * should not be locked again. This is necessary so that changing an object's
+ * state and waking it can be a single atomic operation.
+ */
+static void try_wake_all(struct ntsync_device *dev, struct ntsync_q *q,
+			 struct ntsync_obj *locked_obj)
+{
+	__u32 count = q->count;
+	bool can_wake = true;
+	__u32 i;
+
+	lockdep_assert_held(&dev->wait_all_lock);
+	if (locked_obj)
+		lockdep_assert_held(&locked_obj->lock);
+
+	for (i = 0; i < count; i++) {
+		if (q->entries[i].obj != locked_obj)
+			spin_lock_nest_lock(&q->entries[i].obj->lock, &dev->wait_all_lock);
+	}
+
+	for (i = 0; i < count; i++) {
+		if (!is_signaled(q->entries[i].obj, q->owner)) {
+			can_wake = false;
+			break;
+		}
+	}
+
+	if (can_wake && atomic_cmpxchg(&q->signaled, -1, 0) == -1) {
+		for (i = 0; i < count; i++) {
+			struct ntsync_obj *obj = q->entries[i].obj;
+
+			switch (obj->type) {
+			case NTSYNC_TYPE_SEM:
+				obj->u.sem.count--;
+				break;
+			}
+		}
+		wake_up_process(q->task);
+	}
+
+	for (i = 0; i < count; i++) {
+		if (q->entries[i].obj != locked_obj)
+			spin_unlock(&q->entries[i].obj->lock);
+	}
+}
+
+static void try_wake_all_obj(struct ntsync_device *dev, struct ntsync_obj *obj)
+{
+	struct ntsync_q_entry *entry;
+
+	lockdep_assert_held(&dev->wait_all_lock);
+	lockdep_assert_held(&obj->lock);
+
+	list_for_each_entry(entry, &obj->all_waiters, node)
+		try_wake_all(dev, entry->q, obj);
+}
+
 static void try_wake_any_sem(struct ntsync_obj *sem)
 {
 	struct ntsync_q_entry *entry;
@@ -110,6 +222,7 @@ static int post_sem_state(struct ntsync_obj *sem, __u32 count)
 
 static int ntsync_sem_post(struct ntsync_obj *sem, void __user *argp)
 {
+	struct ntsync_device *dev = sem->dev;
 	__u32 __user *user_args = argp;
 	__u32 prev_count;
 	__u32 args;
@@ -121,14 +234,29 @@ static int ntsync_sem_post(struct ntsync_obj *sem, void __user *argp)
 	if (sem->type != NTSYNC_TYPE_SEM)
 		return -EINVAL;
 
-	spin_lock(&sem->lock);
+	if (atomic_read(&sem->all_hint) > 0) {
+		spin_lock(&dev->wait_all_lock);
+		spin_lock_nest_lock(&sem->lock, &dev->wait_all_lock);
 
-	prev_count = sem->u.sem.count;
-	ret = post_sem_state(sem, args);
-	if (!ret)
-		try_wake_any_sem(sem);
+		prev_count = sem->u.sem.count;
+		ret = post_sem_state(sem, args);
+		if (!ret) {
+			try_wake_all_obj(dev, sem);
+			try_wake_any_sem(sem);
+		}
 
-	spin_unlock(&sem->lock);
+		spin_unlock(&sem->lock);
+		spin_unlock(&dev->wait_all_lock);
+	} else {
+		spin_lock(&sem->lock);
+
+		prev_count = sem->u.sem.count;
+		ret = post_sem_state(sem, args);
+		if (!ret)
+			try_wake_any_sem(sem);
+
+		spin_unlock(&sem->lock);
+	}
 
 	if (!ret && put_user(prev_count, user_args))
 		ret = -EFAULT;
@@ -181,6 +309,8 @@ static struct ntsync_obj *ntsync_alloc_obj(struct ntsync_device *dev,
 	get_file(dev->file);
 	spin_lock_init(&obj->lock);
 	INIT_LIST_HEAD(&obj->any_waiters);
+	INIT_LIST_HEAD(&obj->all_waiters);
+	atomic_set(&obj->all_hint, 0);
 
 	return obj;
 }
@@ -285,7 +415,7 @@ static int ntsync_schedule(const struct ntsync_q *q, const struct ntsync_wait_ar
  * Allocate and initialize the ntsync_q structure, but do not queue us yet.
  */
 static int setup_wait(struct ntsync_device *dev,
-		      const struct ntsync_wait_args *args,
+		      const struct ntsync_wait_args *args, bool all,
 		      struct ntsync_q **ret_q)
 {
 	const __u32 count = args->count;
@@ -309,6 +439,7 @@ static int setup_wait(struct ntsync_device *dev,
 	q->task = current;
 	q->owner = args->owner;
 	atomic_set(&q->signaled, -1);
+	q->all = all;
 	q->count = count;
 
 	for (i = 0; i < count; i++) {
@@ -318,6 +449,16 @@ static int setup_wait(struct ntsync_device *dev,
 		if (!obj)
 			goto err;
 
+		if (all) {
+			/* Check that the objects are all distinct. */
+			for (j = 0; j < i; j++) {
+				if (obj == q->entries[j].obj) {
+					put_obj(obj);
+					goto err;
+				}
+			}
+		}
+
 		entry->obj = obj;
 		entry->q = q;
 		entry->index = i;
@@ -353,7 +494,7 @@ static int ntsync_wait_any(struct ntsync_device *dev, void __user *argp)
 	if (copy_from_user(&args, argp, sizeof(args)))
 		return -EFAULT;
 
-	ret = setup_wait(dev, &args, &q);
+	ret = setup_wait(dev, &args, false, &q);
 	if (ret < 0)
 		return ret;
 
@@ -415,6 +556,87 @@ static int ntsync_wait_any(struct ntsync_device *dev, void __user *argp)
 	return ret;
 }
 
+static int ntsync_wait_all(struct ntsync_device *dev, void __user *argp)
+{
+	struct ntsync_wait_args args;
+	struct ntsync_q *q;
+	int signaled;
+	__u32 i;
+	int ret;
+
+	if (copy_from_user(&args, argp, sizeof(args)))
+		return -EFAULT;
+
+	ret = setup_wait(dev, &args, true, &q);
+	if (ret < 0)
+		return ret;
+
+	/* queue ourselves */
+
+	spin_lock(&dev->wait_all_lock);
+
+	for (i = 0; i < args.count; i++) {
+		struct ntsync_q_entry *entry = &q->entries[i];
+		struct ntsync_obj *obj = entry->obj;
+
+		atomic_inc(&obj->all_hint);
+
+		/*
+		 * obj->all_waiters is protected by dev->wait_all_lock rather
+		 * than obj->lock, so there is no need to acquire obj->lock
+		 * here.
+		 */
+		list_add_tail(&entry->node, &obj->all_waiters);
+	}
+
+	/* check if we are already signaled */
+
+	try_wake_all(dev, q, NULL);
+
+	spin_unlock(&dev->wait_all_lock);
+
+	/* sleep */
+
+	ret = ntsync_schedule(q, &args);
+
+	/* and finally, unqueue */
+
+	spin_lock(&dev->wait_all_lock);
+
+	for (i = 0; i < args.count; i++) {
+		struct ntsync_q_entry *entry = &q->entries[i];
+		struct ntsync_obj *obj = entry->obj;
+
+		/*
+		 * obj->all_waiters is protected by dev->wait_all_lock rather
+		 * than obj->lock, so there is no need to acquire it here.
+		 */
+		list_del(&entry->node);
+
+		atomic_dec(&obj->all_hint);
+
+		put_obj(obj);
+	}
+
+	spin_unlock(&dev->wait_all_lock);
+
+	signaled = atomic_read(&q->signaled);
+	if (signaled != -1) {
+		struct ntsync_wait_args __user *user_args = argp;
+
+		/* even if we caught a signal, we need to communicate success */
+		ret = 0;
+
+		if (put_user(signaled, &user_args->index))
+			ret = -EFAULT;
+	} else if (!ret) {
+		ret = -ETIMEDOUT;
+	}
+
+	kfree(q);
+	return ret;
+}
+
 static int ntsync_char_open(struct inode *inode, struct file *file)
 {
 	struct ntsync_device *dev;
@@ -423,6 +645,8 @@ static int ntsync_char_open(struct inode *inode, struct file *file)
 	if (!dev)
 		return -ENOMEM;
 
+	spin_lock_init(&dev->wait_all_lock);
+
 	file->private_data = dev;
 	dev->file = file;
 	return nonseekable_open(inode, file);
@@ -446,6 +670,8 @@ static long ntsync_char_ioctl(struct file *file, unsigned int cmd,
 	switch (cmd) {
 	case NTSYNC_IOC_CREATE_SEM:
 		return ntsync_create_sem(dev, argp);
+	case NTSYNC_IOC_WAIT_ALL:
+		return ntsync_wait_all(dev, argp);
 	case NTSYNC_IOC_WAIT_ANY:
 		return ntsync_wait_any(dev, argp);
 	default:
diff --git a/include/uapi/linux/ntsync.h b/include/uapi/linux/ntsync.h
index 9cd1dd05d971..524404f6aceb 100644
--- a/include/uapi/linux/ntsync.h
+++ b/include/uapi/linux/ntsync.h
@@ -29,6 +29,7 @@ struct ntsync_wait_args {
 
 #define NTSYNC_IOC_CREATE_SEM		_IOWR('N', 0x80, struct ntsync_sem_args)
 #define NTSYNC_IOC_WAIT_ANY		_IOWR('N', 0x82, struct ntsync_wait_args)
+#define NTSYNC_IOC_WAIT_ALL		_IOWR('N', 0x83, struct ntsync_wait_args)
 
 #define NTSYNC_IOC_SEM_POST		_IOWR('N', 0x81, __u32)
 
-- 
2.43.0