PM: clk: make PM clock layer compatible with clocks that must sleep

The clock API splits its interface into sleepable ant atomic contexts:

- clk_prepare/clk_unprepare for stuff that might sleep

- clk_enable_clk_disable for anything that may be done in atomic context

The code handling runtime PM for clocks only calls clk_disable() on
suspend requests, and clk_enable on resume requests. This means that
runtime PM with clock providers that only have the prepare/unprepare
methods implemented is basically useless.

Many clock implementations can't accommodate atomic contexts. This is
often the case when communication with the clock happens through another
subsystem like I2C or SCMI.

Let's make the clock PM code useful with such clocks by safely invoking
clk_prepare/clk_unprepare upon resume/suspend requests. Of course, when
such clocks are registered with the PM layer then pm_runtime_irq_safe()
can't be used, and neither pm_runtime_suspend() nor pm_runtime_resume()
may be invoked in atomic context.

For clocks that do implement the enable and disable methods then
everything just works as before.

A note on sparse:
According to https://lwn.net/Articles/109066/ there are things
that sparse can't cope with. In particular, pm_clk_op_lock() and
pm_clk_op_unlock() may or may not lock/unlock psd->lock depending on
some runtime condition. To work around that we tell it the lock is
always untaken for the purpose of static analisys.

Thanks to Naresh Kamboju for reporting issues with the initial patch.

Signed-off-by: Nicolas Pitre <[e-mail]>
Tested-by: Naresh Kamboju <[e-mail]>
Signed-off-by: Rafael J. Wysocki <[e-mail]>

drivers/base/power/clock_ops.c

···
 enum pce_status {
 	PCE_STATUS_NONE = 0,
 	PCE_STATUS_ACQUIRED,
+	PCE_STATUS_PREPARED,
 	PCE_STATUS_ENABLED,
 	PCE_STATUS_ERROR,
 };
···
 	char *con_id;
 	struct clk *clk;
 	enum pce_status status;
+	bool enabled_when_prepared;
 };
+
+/**
+ * pm_clk_list_lock - ensure exclusive access for modifying the PM clock
+ *		      entry list.
+ * @psd: pm_subsys_data instance corresponding to the PM clock entry list
+ *	 and clk_op_might_sleep count to be modified.
+ *
+ * Get exclusive access before modifying the PM clock entry list and the
+ * clock_op_might_sleep count to guard against concurrent modifications.
+ * This also protects against a concurrent clock_op_might_sleep and PM clock
+ * entry list usage in pm_clk_suspend()/pm_clk_resume() that may or may not
+ * happen in atomic context, hence both the mutex and the spinlock must be
+ * taken here.
+ */
+static void pm_clk_list_lock(struct pm_subsys_data *psd)
+	__acquires(&psd->lock)
+{
+	mutex_lock(&psd->clock_mutex);
+	spin_lock_irq(&psd->lock);
+}
+
+/**
+ * pm_clk_list_unlock - counterpart to pm_clk_list_lock().
+ * @psd: the same pm_subsys_data instance previously passed to
+ *	 pm_clk_list_lock().
+ */
+static void pm_clk_list_unlock(struct pm_subsys_data *psd)
+	__releases(&psd->lock)
+{
+	spin_unlock_irq(&psd->lock);
+	mutex_unlock(&psd->clock_mutex);
+}
+
+/**
+ * pm_clk_op_lock - ensure exclusive access for performing clock operations.
+ * @psd: pm_subsys_data instance corresponding to the PM clock entry list
+ *	 and clk_op_might_sleep count being used.
+ * @flags: stored irq flags.
+ * @fn: string for the caller function's name.
+ *
+ * This is used by pm_clk_suspend() and pm_clk_resume() to guard
+ * against concurrent modifications to the clock entry list and the
+ * clock_op_might_sleep count. If clock_op_might_sleep is != 0 then
+ * only the mutex can be locked and those functions can only be used in
+ * non atomic context. If clock_op_might_sleep == 0 then these functions
+ * may be used in any context and only the spinlock can be locked.
+ * Returns -EINVAL if called in atomic context when clock ops might sleep.
+ */
+static int pm_clk_op_lock(struct pm_subsys_data *psd, unsigned long *flags,
+			  const char *fn)
+	/* sparse annotations don't work here as exit state isn't static */
+{
+	bool atomic_context = in_atomic() || irqs_disabled();
+
+try_again:
+	spin_lock_irqsave(&psd->lock, *flags);
+	if (!psd->clock_op_might_sleep) {
+		/* the __release is there to work around sparse limitations */
+		__release(&psd->lock);
+		return 0;
+	}
+
+	/* bail out if in atomic context */
+	if (atomic_context) {
+		pr_err("%s: atomic context with clock_ops_might_sleep = %d",
+		       fn, psd->clock_op_might_sleep);
+		spin_unlock_irqrestore(&psd->lock, *flags);
+		might_sleep();
+		return -EPERM;
+	}
+
+	/* we must switch to the mutex */
+	spin_unlock_irqrestore(&psd->lock, *flags);
+	mutex_lock(&psd->clock_mutex);
+
+	/*
+	 * There was a possibility for psd->clock_op_might_sleep
+	 * to become 0 above. Keep the mutex only if not the case.
+	 */
+	if (likely(psd->clock_op_might_sleep))
+		return 0;
+
+	mutex_unlock(&psd->clock_mutex);
+	goto try_again;
+}
+
+/**
+ * pm_clk_op_unlock - counterpart to pm_clk_op_lock().
+ * @psd: the same pm_subsys_data instance previously passed to
+ *	 pm_clk_op_lock().
+ * @flags: irq flags provided by pm_clk_op_lock().
+ */
+static void pm_clk_op_unlock(struct pm_subsys_data *psd, unsigned long *flags)
+	/* sparse annotations don't work here as entry state isn't static */
+{
+	if (psd->clock_op_might_sleep) {
+		mutex_unlock(&psd->clock_mutex);
+	} else {
+		/* the __acquire is there to work around sparse limitations */
+		__acquire(&psd->lock);
+		spin_unlock_irqrestore(&psd->lock, *flags);
+	}
+}
 
 /**
  * pm_clk_enable - Enable a clock, reporting any errors
···
 {
 	int ret;
 
-	if (ce->status < PCE_STATUS_ERROR) {
+	switch (ce->status) {
+	case PCE_STATUS_ACQUIRED:
+		ret = clk_prepare_enable(ce->clk);
+		break;
+	case PCE_STATUS_PREPARED:
 		ret = clk_enable(ce->clk);
-		if (!ret)
-			ce->status = PCE_STATUS_ENABLED;
-		else
-			dev_err(dev, "%s: failed to enable clk %p, error %d\n",
-				__func__, ce->clk, ret);
+		break;
+	default:
+		return;
 	}
+	if (!ret)
+		ce->status = PCE_STATUS_ENABLED;
+	else
+		dev_err(dev, "%s: failed to enable clk %p, error %d\n",
+			__func__, ce->clk, ret);
 }
 
 /**
···
 		ce->clk = clk_get(dev, ce->con_id);
 	if (IS_ERR(ce->clk)) {
 		ce->status = PCE_STATUS_ERROR;
+		return;
+	} else if (clk_is_enabled_when_prepared(ce->clk)) {
+		/* we defer preparing the clock in that case */
+		ce->status = PCE_STATUS_ACQUIRED;
+		ce->enabled_when_prepared = true;
+	} else if (clk_prepare(ce->clk)) {
+		ce->status = PCE_STATUS_ERROR;
+		dev_err(dev, "clk_prepare() failed\n");
+		return;
 	} else {
-		if (clk_prepare(ce->clk)) {
-			ce->status = PCE_STATUS_ERROR;
-			dev_err(dev, "clk_prepare() failed\n");
-		} else {
-			ce->status = PCE_STATUS_ACQUIRED;
-			dev_dbg(dev,
-				"Clock %pC con_id %s managed by runtime PM.\n",
-				ce->clk, ce->con_id);
-		}
+		ce->status = PCE_STATUS_PREPARED;
 	}
+	dev_dbg(dev, "Clock %pC con_id %s managed by runtime PM.\n",
+		ce->clk, ce->con_id);
 }
 
 static int __pm_clk_add(struct device *dev, const char *con_id,
···
 
 	pm_clk_acquire(dev, ce);
 
-	spin_lock_irq(&psd->lock);
+	pm_clk_list_lock(psd);
 	list_add_tail(&ce->node, &psd->clock_list);
-	spin_unlock_irq(&psd->lock);
+	if (ce->enabled_when_prepared)
+		psd->clock_op_might_sleep++;
+	pm_clk_list_unlock(psd);
 	return 0;
 }
 
···
 	if (!ce)
 		return;
 
-	if (ce->status < PCE_STATUS_ERROR) {
-		if (ce->status == PCE_STATUS_ENABLED)
-			clk_disable(ce->clk);
-
-		if (ce->status >= PCE_STATUS_ACQUIRED) {
-			clk_unprepare(ce->clk);
+	switch (ce->status) {
+	case PCE_STATUS_ENABLED:
+		clk_disable(ce->clk);
+		fallthrough;
+	case PCE_STATUS_PREPARED:
+		clk_unprepare(ce->clk);
+		fallthrough;
+	case PCE_STATUS_ACQUIRED:
+	case PCE_STATUS_ERROR:
+		if (!IS_ERR(ce->clk))
 			clk_put(ce->clk);
-		}
+		break;
+	default:
+		break;
 	}
 
 	kfree(ce->con_id);
···
 	if (!psd)
 		return;
 
-	spin_lock_irq(&psd->lock);
+	pm_clk_list_lock(psd);
 
 	list_for_each_entry(ce, &psd->clock_list, node) {
 		if (!con_id && !ce->con_id)
···
 			goto remove;
 	}
 
-	spin_unlock_irq(&psd->lock);
+	pm_clk_list_unlock(psd);
 	return;
 
  remove:
 	list_del(&ce->node);
-	spin_unlock_irq(&psd->lock);
+	if (ce->enabled_when_prepared)
+		psd->clock_op_might_sleep--;
+	pm_clk_list_unlock(psd);
 
 	__pm_clk_remove(ce);
 }
···
 	if (!psd || !clk)
 		return;
 
-	spin_lock_irq(&psd->lock);
+	pm_clk_list_lock(psd);
 
 	list_for_each_entry(ce, &psd->clock_list, node) {
 		if (clk == ce->clk)
 			goto remove;
 	}
 
-	spin_unlock_irq(&psd->lock);
+	pm_clk_list_unlock(psd);
 	return;
 
  remove:
 	list_del(&ce->node);
-	spin_unlock_irq(&psd->lock);
+	if (ce->enabled_when_prepared)
+		psd->clock_op_might_sleep--;
+	pm_clk_list_unlock(psd);
 
 	__pm_clk_remove(ce);
 }
···
  * @dev: Device to initialize the list of PM clocks for.
  *
  * Initialize the lock and clock_list members of the device's pm_subsys_data
- * object.
+ * object, set the count of clocks that might sleep to 0.
  */
 void pm_clk_init(struct device *dev)
 {
 	struct pm_subsys_data *psd = dev_to_psd(dev);
-	if (psd)
+	if (psd) {
 		INIT_LIST_HEAD(&psd->clock_list);
+		mutex_init(&psd->clock_mutex);
+		psd->clock_op_might_sleep = 0;
+	}
 }
 EXPORT_SYMBOL_GPL(pm_clk_init);
 
···
 
 	INIT_LIST_HEAD(&list);
 
-	spin_lock_irq(&psd->lock);
+	pm_clk_list_lock(psd);
 
 	list_for_each_entry_safe_reverse(ce, c, &psd->clock_list, node)
 		list_move(&ce->node, &list);
+	psd->clock_op_might_sleep = 0;
 
-	spin_unlock_irq(&psd->lock);
+	pm_clk_list_unlock(psd);
 
 	dev_pm_put_subsys_data(dev);
 
···
 	struct pm_subsys_data *psd = dev_to_psd(dev);
 	struct pm_clock_entry *ce;
 	unsigned long flags;
+	int ret;
 
 	dev_dbg(dev, "%s()\n", __func__);
 
 	if (!psd)
 		return 0;
 
-	spin_lock_irqsave(&psd->lock, flags);
+	ret = pm_clk_op_lock(psd, &flags, __func__);
+	if (ret)
+		return ret;
 
 	list_for_each_entry_reverse(ce, &psd->clock_list, node) {
-		if (ce->status < PCE_STATUS_ERROR) {
-			if (ce->status == PCE_STATUS_ENABLED)
+		if (ce->status == PCE_STATUS_ENABLED) {
+			if (ce->enabled_when_prepared) {
+				clk_disable_unprepare(ce->clk);
+				ce->status = PCE_STATUS_ACQUIRED;
+			} else {
 				clk_disable(ce->clk);
-			ce->status = PCE_STATUS_ACQUIRED;
+				ce->status = PCE_STATUS_PREPARED;
+			}
 		}
 	}
 
-	spin_unlock_irqrestore(&psd->lock, flags);
+	pm_clk_op_unlock(psd, &flags);
 
 	return 0;
 }
···
 	struct pm_subsys_data *psd = dev_to_psd(dev);
 	struct pm_clock_entry *ce;
 	unsigned long flags;
+	int ret;
 
 	dev_dbg(dev, "%s()\n", __func__);
 
 	if (!psd)
 		return 0;
 
-	spin_lock_irqsave(&psd->lock, flags);
+	ret = pm_clk_op_lock(psd, &flags, __func__);
+	if (ret)
+		return ret;
 
 	list_for_each_entry(ce, &psd->clock_list, node)
 		__pm_clk_enable(dev, ce);
 
-	spin_unlock_irqrestore(&psd->lock, flags);
+	pm_clk_op_unlock(psd, &flags);
 
 	return 0;
 }

drivers/clk/clk.c

···
 }
 EXPORT_SYMBOL_GPL(clk_enable);
 
+/**
+ * clk_is_enabled_when_prepared - indicate if preparing a clock also enables it.
+ * @clk: clock source
+ *
+ * Returns true if clk_prepare() implicitly enables the clock, effectively
+ * making clk_enable()/clk_disable() no-ops, false otherwise.
+ *
+ * This is of interest mainly to power management code where actually
+ * disabling the clock also requires unpreparing it to have any material
+ * effect.
+ *
+ * Regardless of the value returned here, the caller must always invoke
+ * clk_enable() or clk_prepare_enable()  and counterparts for usage counts
+ * to be right.
+ */
+bool clk_is_enabled_when_prepared(struct clk *clk)
+{
+	return clk && !(clk->core->ops->enable && clk->core->ops->disable);
+}
+EXPORT_SYMBOL_GPL(clk_is_enabled_when_prepared);
+
 static int clk_core_prepare_enable(struct clk_core *core)
 {
 	int ret;

include/linux/clk.h

···
 
 #endif
 
+#ifdef CONFIG_HAVE_CLK_PREPARE
 /**
  * clk_prepare - prepare a clock source
  * @clk: clock source
···
  *
  * Must not be called from within atomic context.
  */
-#ifdef CONFIG_HAVE_CLK_PREPARE
 int clk_prepare(struct clk *clk);
 int __must_check clk_bulk_prepare(int num_clks,
 				  const struct clk_bulk_data *clks);
+
+/**
+ * clk_is_enabled_when_prepared - indicate if preparing a clock also enables it.
+ * @clk: clock source
+ *
+ * Returns true if clk_prepare() implicitly enables the clock, effectively
+ * making clk_enable()/clk_disable() no-ops, false otherwise.
+ *
+ * This is of interest mainly to the power management code where actually
+ * disabling the clock also requires unpreparing it to have any material
+ * effect.
+ *
+ * Regardless of the value returned here, the caller must always invoke
+ * clk_enable() or clk_prepare_enable()  and counterparts for usage counts
+ * to be right.
+ */
+bool clk_is_enabled_when_prepared(struct clk *clk);
 #else
 static inline int clk_prepare(struct clk *clk)
 {
···
 {
 	might_sleep();
 	return 0;
+}
+
+static inline bool clk_is_enabled_when_prepared(struct clk *clk)
+{
+	return false;
 }
 #endif
 

include/linux/pm.h

···
 	spinlock_t lock;
 	unsigned int refcount;
 #ifdef CONFIG_PM_CLK
+	unsigned int clock_op_might_sleep;
+	struct mutex clock_mutex;
 	struct list_head clock_list;
 #endif
 #ifdef CONFIG_PM_GENERIC_DOMAINS