Skip to contentSkip to conversation

Fix a deadlock with clk_pm_runtime_get()

PATCH Superseded Stephen Boyd 3 participants 15 messages Mar 25, 2024 – Sep 13, 2024

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Stephen Boyd

author maintainer
#0 lore raw

cover Fix a deadlock with clk_pm_runtime_get()

This patch series fixes a deadlock reported[1] on ChromeOS devices (Qualcomm sc7180 Trogdor). To get there, we allow __clk_release() to run without the prepare_lock held. Then we add runtime PM enabled clk_core structs to a list that we iterate and enable runtime PM for each entry before grabbing the prepare_lock to walk the clk tree. The details are in patch #4.

The patch after that is based on the analysis in the disable unused patch. We similarly resume devices from runtime suspend when walking the clk tree for the debugfs clk_summary.

Unfortunately this doesn't fix all problems with the usage of runtime PM in the clk framework. We still have a problem if preparing a clk happens in parallel to the device providing that clk runtime resuming or suspending. In that case, the task will go to sleep waiting for the runtime PM state to change, and we'll deadlock. This is primarily a problem with the global prepare_lock. I suspect we'll be able to fix this by implementing per-clk locking, because then we will be able to split up the big prepare_lock into smaller locks that don't deadlock on some device runtime PM transitions.

I'll start working on that problem in earnest now because I'm worried we're going to run into that problem very soon.

Stephen Boyd (5):
clk: Remove prepare_lock hold assertion in __clk_release()
clk: Don't hold prepare_lock when calling kref_put()
clk: Initialize struct clk_core kref earlier
clk: Get runtime PM before walking tree during disable_unused
clk: Get runtime PM before walking tree for clk_summary

 drivers/clk/clk.c | 142 +++++++++++++++++++++++++++++++++++++---------
 1 file changed, 115 insertions(+), 27 deletions(-)
Cc Douglas Anderson <[email protected]>
Cc Krzysztof Kozlowski <[email protected]>
Cc Marek Szyprowski <[email protected]>
Cc Taniya Das <[email protected]>
Cc Ulf Hansson <[email protected]>

[1] https://lore.kernel.org/all/20220922084322.RFC.2.I375b6b9e0a0a5348962f004beb3dafee6a12dfbb@changeid/

base-commit: e8f897f4afef0031fe618a8e94127a0934896aba

Signature · 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/clk/linux.git/
https://git.kernel.org/pub/scm/linux/kernel/git/sboyd/spmi.git
author maintainer
#1 lore raw

PATCH 1/5 clk: Remove prepare_lock hold assertion in __clk_release()

Removing this assertion lets us move the kref_put() call outside the prepare_lock section. We don't need to hold the prepare_lock here to free memory and destroy the clk_core structure. We've already unlinked the clk from the clk tree and by the time the release function runs nothing holds a reference to the clk_core anymore so anything with the pointer can't access the memory that's being freed anyway. Way back in commit 496eadf821c2 ("clk: Use lockdep asserts to find missing hold of prepare_lock") we didn't need to have this assertion either.

Fixes 496eadf821c2 ("clk: Use lockdep asserts to find missing hold of prepare_lock")
Cc Krzysztof Kozlowski <[email protected]>
Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 2 --
 1 file changed, 2 deletions(-)
drivers/clk/clk.c +0 -2
index 2253c154a824..44e71736477d 100644
@@ -4353,8 +4353,6 @@ static void __clk_release(struct kref *ref)
 {
 	struct clk_core *core = container_of(ref, struct clk_core, ref);
 
-	lockdep_assert_held(&prepare_lock);
-
 	clk_core_free_parent_map(core);
 	kfree_const(core->name);
 	kfree(core);
Signature · 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/clk/linux.git/
https://git.kernel.org/pub/scm/linux/kernel/git/sboyd/spmi.git
Reviewed-by on v1 1/5
#2 lore raw

Re: [PATCH 1/5] clk: Remove prepare_lock hold assertion in __clk_release()

Hi,

Quoted 10 lines · Stephen Boyd

Removing this assertion lets us move the kref_put() call outside the prepare_lock section. We don't need to hold the prepare_lock here to free memory and destroy the clk_core structure. We've already unlinked the clk from the clk tree and by the time the release function runs nothing holds a reference to the clk_core anymore so anything with the pointer can't access the memory that's being freed anyway. Way back in commit 496eadf821c2 ("clk: Use lockdep asserts to find missing hold of prepare_lock") we didn't need to have this assertion either.

Quoted 7 lines
Fixes 496eadf821c2 ("clk: Use lockdep asserts to find missing hold of prepare_lock")
Cc Krzysztof Kozlowski <[email protected]>
Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 2 --
 1 file changed, 2 deletions(-)
author maintainer
#3 lore raw

PATCH 2/5 clk: Don't hold prepare_lock when calling kref_put()

In reply to Stephen Boyd

We don't need to hold the prepare_lock when dropping a ref on a struct clk_core. The release function is only freeing memory and any code with a pointer reference has already unlinked anything pointing to the clk_core. This reduces the holding area of the prepare_lock a bit.

Note that we also don't call free_clk() with the prepare_lock held. There isn't any reason to do that.

Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)
drivers/clk/clk.c +5 -7
index 44e71736477d..9fc522c26de8 100644
@@ -4448,7 +4448,8 @@ void clk_unregister(struct clk *clk)
 	if (ops == &clk_nodrv_ops) {
 		pr_err("%s: unregistered clock: %s\n", __func__,
 		       clk->core->name);
-		goto unlock;
+		clk_prepare_unlock();
+		return;
 	}
 	/*
 	 * Assign empty clock ops for consumers that might still hold
@@ -4482,11 +4483,10 @@ void clk_unregister(struct clk *clk)
 	if (clk->core->protect_count)
 		pr_warn("%s: unregistering protected clock: %s\n",
 					__func__, clk->core->name);
+	clk_prepare_unlock();
 
 	kref_put(&clk->core->ref, __clk_release);
 	free_clk(clk);
-unlock:
-	clk_prepare_unlock();
 }
 EXPORT_SYMBOL_GPL(clk_unregister);
 
@@ -4645,13 +4645,11 @@ void __clk_put(struct clk *clk)
 	if (clk->min_rate > 0 || clk->max_rate < ULONG_MAX)
 		clk_set_rate_range_nolock(clk, 0, ULONG_MAX);
 
-	owner = clk->core->owner;
-	kref_put(&clk->core->ref, __clk_release);
-
 	clk_prepare_unlock();
 
+	owner = clk->core->owner;
+	kref_put(&clk->core->ref, __clk_release);
 	module_put(owner);
-
 	free_clk(clk);
 }
 
Signature · 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/clk/linux.git/
https://git.kernel.org/pub/scm/linux/kernel/git/sboyd/spmi.git
Reviewed-by on v1 2/5
#4 lore raw

Re: [PATCH 2/5] clk: Don't hold prepare_lock when calling kref_put()

Hi,

Quoted 6 lines · Stephen Boyd

We don't need to hold the prepare_lock when dropping a ref on a struct clk_core. The release function is only freeing memory and any code with a pointer reference has already unlinked anything pointing to the clk_core. This reduces the holding area of the prepare_lock a bit.

Quoted 8 lines

Note that we also don't call free_clk() with the prepare_lock held. There isn't any reason to do that.

Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)
author maintainer
#5 lore raw

PATCH 3/5 clk: Initialize struct clk_core kref earlier

In reply to Stephen Boyd

Initialize this kref once we allocate memory for the struct clk_core so that we can reuse the release function to free any memory associated with the structure. This mostly consolidates code, but also clarifies that the kref lifetime exists once the container structure (struct clk_core) is allocated instead of leaving it in a half-baked state for most of __clk_core_init().

Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 28 +++++++++++++---------------
 1 file changed, 13 insertions(+), 15 deletions(-)
drivers/clk/clk.c +13 -15
index 9fc522c26de8..ee80b21f2824 100644
@@ -3959,8 +3959,6 @@ static int __clk_core_init(struct clk_core *core)
 	}
 
 	clk_core_reparent_orphans_nolock();
-
-	kref_init(&core->ref);
 out:
 	clk_pm_runtime_put(core);
 unlock:
@@ -4189,6 +4187,16 @@ static void clk_core_free_parent_map(struct clk_core *core)
 	kfree(core->parents);
 }
 
+/* Free memory allocated for a struct clk_core */
+static void __clk_release(struct kref *ref)
+{
+	struct clk_core *core = container_of(ref, struct clk_core, ref);
+
+	clk_core_free_parent_map(core);
+	kfree_const(core->name);
+	kfree(core);
+}
+
 static struct clk *
 __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
 {
@@ -4209,6 +4217,8 @@ __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
 		goto fail_out;
 	}
 
+	kref_init(&core->ref);
+
 	core->name = kstrdup_const(init->name, GFP_KERNEL);
 	if (!core->name) {
 		ret = -ENOMEM;
@@ -4263,12 +4273,10 @@ __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
 	hw->clk = NULL;
 
 fail_create_clk:
-	clk_core_free_parent_map(core);
 fail_parents:
 fail_ops:
-	kfree_const(core->name);
 fail_name:
-	kfree(core);
+	kref_put(&core->ref, __clk_release);
 fail_out:
 	return ERR_PTR(ret);
 }
@@ -4348,16 +4356,6 @@ int of_clk_hw_register(struct device_node *node, struct clk_hw *hw)
 }
 EXPORT_SYMBOL_GPL(of_clk_hw_register);
 
-/* Free memory allocated for a clock. */
-static void __clk_release(struct kref *ref)
-{
-	struct clk_core *core = container_of(ref, struct clk_core, ref);
-
-	clk_core_free_parent_map(core);
-	kfree_const(core->name);
-	kfree(core);
-}
-
 /*
  * Empty clk_ops for unregistered clocks. These are used temporarily
  * after clk_unregister() was called on a clock and until last clock
Signature · 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/clk/linux.git/
https://git.kernel.org/pub/scm/linux/kernel/git/sboyd/spmi.git
Reviewed-by on v1 3/5
#6 lore raw

Re: [PATCH 3/5] clk: Initialize struct clk_core kref earlier

Hik

Quoted 56 lines · Stephen Boyd

Initialize this kref once we allocate memory for the struct clk_core so that we can reuse the release function to free any memory associated with the structure. This mostly consolidates code, but also clarifies that the kref lifetime exists once the container structure (struct clk_core) is allocated instead of leaving it in a half-baked state for most of __clk_core_init().

Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 28 +++++++++++++---------------
 1 file changed, 13 insertions(+), 15 deletions(-)
drivers/clk/clk.c +12 -2
index 9fc522c26de8..ee80b21f2824 100644
@@ -3959,8 +3959,6 @@ static int __clk_core_init(struct clk_core *core)
        }
 
        clk_core_reparent_orphans_nolock();
-
-       kref_init(&core->ref);
 out:
        clk_pm_runtime_put(core);
 unlock:
@@ -4189,6 +4187,16 @@ static void clk_core_free_parent_map(struct clk_core *core)
        kfree(core->parents);
 }
 
+/* Free memory allocated for a struct clk_core */
+static void __clk_release(struct kref *ref)
+{
+       struct clk_core *core = container_of(ref, struct clk_core, ref);
+
+       clk_core_free_parent_map(core);
+       kfree_const(core->name);
+       kfree(core);
+}
+
 static struct clk *
 __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
 {
@@ -4209,6 +4217,8 @@ __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
                goto fail_out;
        }
 
+       kref_init(&core->ref);
+
        core->name = kstrdup_const(init->name, GFP_KERNEL);
        if (!core->name) {
                ret = -ENOMEM;
@@ -4263,12 +4273,10 @@ __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
        hw->clk = NULL;
 

drivers/clk/clk.c __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw) · line 4281

Quoted 10 lines
drivers/clk/clk.c +1 -3
 fail_create_clk:
-       clk_core_free_parent_map(core);
 fail_parents:
 fail_ops:
-       kfree_const(core->name);
 fail_name:
-       kfree(core);
+       kref_put(&core->ref, __clk_release);
 fail_out:
        return ERR_PTR(ret);

If it were me, I probably would have:

* Removed "fail_out" and turned the one "goto fail_out" to just return the error.

* Consolidated the rest of the labels into a single "fail" label.

That's definitely just a style opinion though, and IMO the patch is fine as-is and overall cleans up the code.

author maintainer
#7 lore raw

PATCH 4/5 clk: Get runtime PM before walking tree during disable_unused

In reply to Stephen Boyd

Doug reported [1] the following hung task:

 INFO: task swapper/0:1 blocked for more than 122 seconds.
       Not tainted 5.15.149-21875-gf795ebc40eb8 #1
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
 task:swapper/0       state:D stack:    0 pid:    1 ppid:     0 flags:0x00000008
 Call trace:
  __switch_to+0xf4/0x1f4
  __schedule+0x418/0xb80
  schedule+0x5c/0x10c
  rpm_resume+0xe0/0x52c
  rpm_resume+0x178/0x52c
  __pm_runtime_resume+0x58/0x98
  clk_pm_runtime_get+0x30/0xb0
  clk_disable_unused_subtree+0x58/0x208
  clk_disable_unused_subtree+0x38/0x208
  clk_disable_unused_subtree+0x38/0x208
  clk_disable_unused_subtree+0x38/0x208
  clk_disable_unused_subtree+0x38/0x208
  clk_disable_unused+0x4c/0xe4
  do_one_initcall+0xcc/0x2d8
  do_initcall_level+0xa4/0x148
  do_initcalls+0x5c/0x9c
  do_basic_setup+0x24/0x30
  kernel_init_freeable+0xec/0x164
  kernel_init+0x28/0x120
  ret_from_fork+0x10/0x20
 INFO: task kworker/u16:0:9 blocked for more than 122 seconds.
       Not tainted 5.15.149-21875-gf795ebc40eb8 #1
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
 task:kworker/u16:0   state:D stack:    0 pid:    9 ppid:     2 flags:0x00000008
 Workqueue: events_unbound deferred_probe_work_func
 Call trace:
  __switch_to+0xf4/0x1f4
  __schedule+0x418/0xb80
  schedule+0x5c/0x10c
  schedule_preempt_disabled+0x2c/0x48
  __mutex_lock+0x238/0x488
  __mutex_lock_slowpath+0x1c/0x28
  mutex_lock+0x50/0x74
  clk_prepare_lock+0x7c/0x9c
  clk_core_prepare_lock+0x20/0x44
  clk_prepare+0x24/0x30
  clk_bulk_prepare+0x40/0xb0
  mdss_runtime_resume+0x54/0x1c8
  pm_generic_runtime_resume+0x30/0x44
  __genpd_runtime_resume+0x68/0x7c
  genpd_runtime_resume+0x108/0x1f4
  __rpm_callback+0x84/0x144
  rpm_callback+0x30/0x88
  rpm_resume+0x1f4/0x52c
  rpm_resume+0x178/0x52c
  __pm_runtime_resume+0x58/0x98
  __device_attach+0xe0/0x170
  device_initial_probe+0x1c/0x28
  bus_probe_device+0x3c/0x9c
  device_add+0x644/0x814
  mipi_dsi_device_register_full+0xe4/0x170
  devm_mipi_dsi_device_register_full+0x28/0x70
  ti_sn_bridge_probe+0x1dc/0x2c0
  auxiliary_bus_probe+0x4c/0x94
  really_probe+0xcc/0x2c8
  __driver_probe_device+0xa8/0x130
  driver_probe_device+0x48/0x110
  __device_attach_driver+0xa4/0xcc
  bus_for_each_drv+0x8c/0xd8
  __device_attach+0xf8/0x170
  device_initial_probe+0x1c/0x28
  bus_probe_device+0x3c/0x9c
  deferred_probe_work_func+0x9c/0xd8
  process_one_work+0x148/0x518
  worker_thread+0x138/0x350
  kthread+0x138/0x1e0
  ret_from_fork+0x10/0x20

The first thread is walking the clk tree and calling clk_pm_runtime_get() to power on devices required to read the clk hardware via struct clk_ops::is_enabled(). This thread holds the clk prepare_lock, and is trying to runtime PM resume a device, when it finds that the device is in the process of resuming so the thread schedule()s away waiting for the device to finish resuming before continuing. The second thread is runtime PM resuming the same device, but the runtime resume callback is calling clk_prepare(), trying to grab the prepare_lock waiting on the first thread.

This is a classic ABBA deadlock. To properly fix the deadlock, we must never runtime PM resume or suspend a device with the clk prepare_lock held. Actually doing that is near impossible today because the global prepare_lock would have to be dropped in the middle of the tree, the device runtime PM resumed/suspended, and then the prepare_lock grabbed again to ensure consistency of the clk tree topology. If anything changes with the clk tree in the meantime, we've lost and will need to start the operation all over again.

Luckily, most of the time we're simply incrementing or decrementing the runtime PM count on an active device, so we don't have the chance to schedule away with the prepare_lock held. Let's fix this immediate problem that can be triggered more easily by simply booting on Qualcomm sc7180.

Introduce a list of clk_core structures that have been registered, or are in the process of being registered, that require runtime PM to operate. Iterate this list and call clk_pm_runtime_get() on each of them without holding the prepare_lock during clk_disable_unused(). This way we can be certain that the runtime PM state of the devices will be active and resumed so we can't schedule away while walking the clk tree with the prepare_lock held. Similarly, call clk_pm_runtime_put() without the prepare_lock held to properly drop the runtime PM reference.

---
 drivers/clk/clk.c | 94 ++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 92 insertions(+), 2 deletions(-)
drivers/clk/clk.c +92 -2
index ee80b21f2824..31998ca67b1e 100644
@@ -37,6 +37,10 @@ static HLIST_HEAD(clk_root_list);
 static HLIST_HEAD(clk_orphan_list);
 static LIST_HEAD(clk_notifier_list);
 
+/* List of registered clks that use runtime PM */
+static HLIST_HEAD(clk_rpm_list);
+static DEFINE_MUTEX(clk_rpm_list_lock);
+
 static const struct hlist_head *all_lists[] = {
 	&clk_root_list,
 	&clk_orphan_list,
@@ -59,6 +63,7 @@ struct clk_core {
 	struct clk_hw		*hw;
 	struct module		*owner;
 	struct device		*dev;
+	struct hlist_node	rpm_node;
 	struct device_node	*of_node;
 	struct clk_core		*parent;
 	struct clk_parent_map	*parents;
@@ -122,6 +127,19 @@ static void clk_pm_runtime_put(struct clk_core *core)
 	pm_runtime_put_sync(core->dev);
 }
 
+static void clk_core_rpm_init(struct clk_core *core)
+{
+	struct device *dev = core->dev;
+
+	if (dev && pm_runtime_enabled(dev)) {
+		core->rpm_enabled = true;
+
+		mutex_lock(&clk_rpm_list_lock);
+		hlist_add_head(&core->rpm_node, &clk_rpm_list);
+		mutex_unlock(&clk_rpm_list_lock);
+	}
+}
+
 /***           locking             ***/
 static void clk_prepare_lock(void)
 {
@@ -191,6 +209,63 @@ static void clk_enable_unlock(unsigned long flags)
 	spin_unlock_irqrestore(&enable_lock, flags);
 }
 
+/*
+ * Call clk_pm_runtime_get() on all runtime PM enabled clks in the clk tree so
+ * that disabling unused clks avoids a deadlock where a device is runtime PM
+ * resuming/suspending and the runtime PM callback is trying to grab the
+ * prepare_lock for something like clk_prepare_enable() while
+ * clk_disable_unused_subtree() holds the prepare_lock and is trying to runtime
+ * PM resume/suspend the device as well.
+ */
+static int clk_pm_runtime_get_all(void)
+{
+	int ret;
+	struct clk_core *core, *failed;
+
+	/*
+	 * Grab the list lock to avoid any new clks from being registered
+	 * or unregistered.
+	 */
+	mutex_lock(&clk_rpm_list_lock);
+
+	/*
+	 * Runtime PM "get" all the devices that are needed for the clks
+	 * currently registered. Do this without holding the prepare_lock, to
+	 * avoid the deadlock.
+	 */
+	hlist_for_each_entry(core, &clk_rpm_list, rpm_node) {
+		ret = clk_pm_runtime_get(core);
+		if (ret) {
+			failed = core;
+			pr_err("clk: Failed to runtime PM get '%s' for clk '%s'\n",
+			       failed->name, dev_name(failed->dev));
+			goto out;
+		}
+	}
+
+	return 0;
+
+out:
+	hlist_for_each_entry(core, &clk_rpm_list, rpm_node) {
+		if (core == failed)
+			break;
+
+		clk_pm_runtime_put(core);
+	}
+	mutex_unlock(&clk_rpm_list_lock);
+
+	return ret;
+}
+
+static void clk_pm_runtime_put_all(void)
+{
+	struct clk_core *core;
+
+	hlist_for_each_entry(core, &clk_rpm_list, rpm_node)
+		clk_pm_runtime_put(core);
+	mutex_unlock(&clk_rpm_list_lock);
+}
+
 static bool clk_core_rate_is_protected(struct clk_core *core)
 {
 	return core->protect_count;
@@ -1431,6 +1506,7 @@ __setup("clk_ignore_unused", clk_ignore_unused_setup);
 static int __init clk_disable_unused(void)
 {
 	struct clk_core *core;
+	int ret;
 
 	if (clk_ignore_unused) {
 		pr_warn("clk: Not disabling unused clocks\n");
@@ -1439,6 +1515,13 @@ static int __init clk_disable_unused(void)
 
 	pr_info("clk: Disabling unused clocks\n");
 
+	ret = clk_pm_runtime_get_all();
+	if (ret)
+		return ret;
+	/*
+	 * Grab the prepare lock to keep the clk topology stable while iterating
+	 * over clks.
+	 */
 	clk_prepare_lock();
 
 	hlist_for_each_entry(core, &clk_root_list, child_node)
@@ -1455,6 +1538,8 @@ static int __init clk_disable_unused(void)
 
 	clk_prepare_unlock();
 
+	clk_pm_runtime_put_all();
+
 	return 0;
 }
 late_initcall_sync(clk_disable_unused);
@@ -4192,6 +4277,12 @@ static void __clk_release(struct kref *ref)
 {
 	struct clk_core *core = container_of(ref, struct clk_core, ref);
 
+	if (core->rpm_enabled) {
+		mutex_lock(&clk_rpm_list_lock);
+		hlist_del(&core->rpm_node);
+		mutex_unlock(&clk_rpm_list_lock);
+	}
+
 	clk_core_free_parent_map(core);
 	kfree_const(core->name);
 	kfree(core);
@@ -4231,9 +4322,8 @@ __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
 	}
 	core->ops = init->ops;
 
-	if (dev && pm_runtime_enabled(dev))
-		core->rpm_enabled = true;
 	core->dev = dev;
+	clk_core_rpm_init(core);
 	core->of_node = np;
 	if (dev && dev->driver)
 		core->owner = dev->driver->owner;
Signature · 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/clk/linux.git/
https://git.kernel.org/pub/scm/linux/kernel/git/sboyd/spmi.git
Reviewed-by on v1 4/5
#8 lore raw

Re: [PATCH 4/5] clk: Get runtime PM before walking tree during disable_unused

Hi,

Quoted 9 lines · Stephen Boyd

Introduce a list of clk_core structures that have been registered, or are in the process of being registered, that require runtime PM to operate. Iterate this list and call clk_pm_runtime_get() on each of them without holding the prepare_lock during clk_disable_unused(). This way we can be certain that the runtime PM state of the devices will be active and resumed so we can't schedule away while walking the clk tree with the prepare_lock held. Similarly, call clk_pm_runtime_put() without the prepare_lock held to properly drop the runtime PM reference.

There's a part of me that worries about the fact that we'll now be doing a pm_runtime get() on _all clocks_ (even those that are used) at bootup now. I worry that some device out there will be unhappy about it. ...but I guess the device passed in here is already documented to be one that the clock framework can get/put whenever it needs to prepare the clock, so that makes me feel like it should be fine.

Anyway, no action item, just documenting my thoughts...

Oh, funny. After reading the next patch, I guess I'm even less concerned. I guess we were already grabbing the pm_runtime state for all clocks while printing the clock summary. While that's a debugfs function, it's still something that many people have likely exercised and it's likely not going to introduce random/long tail problems.

drivers/clk/clk.c static void clk_enable_unlock(unsigned long flags) · line 220

Quoted 9 lines
drivers/clk/clk.c +9 -0
+/*
+ * Call clk_pm_runtime_get() on all runtime PM enabled clks in the clk tree so
+ * that disabling unused clks avoids a deadlock where a device is runtime PM
+ * resuming/suspending and the runtime PM callback is trying to grab the
+ * prepare_lock for something like clk_prepare_enable() while
+ * clk_disable_unused_subtree() holds the prepare_lock and is trying to runtime
+ * PM resume/suspend the device as well.
+ */
+static int clk_pm_runtime_get_all(void)

nit: It'd be nice if this documented that it acquired / held the lock. Could be in comments, or, might as well use the syntax like this (I think):

__acquires(&clk_rpm_list_lock);

...similar with the put function.

drivers/clk/clk.c static void clk_enable_unlock(unsigned long flags) · line 241

Quoted 11 lines
drivers/clk/clk.c +11 -0
+       /*
+        * Runtime PM "get" all the devices that are needed for the clks
+        * currently registered. Do this without holding the prepare_lock, to
+        * avoid the deadlock.
+        */
+       hlist_for_each_entry(core, &clk_rpm_list, rpm_node) {
+               ret = clk_pm_runtime_get(core);
+               if (ret) {
+                       failed = core;
+                       pr_err("clk: Failed to runtime PM get '%s' for clk '%s'\n",
+                              failed->name, dev_name(failed->dev));

If I'm reading this correctly, the strings are backward in your error print. Right now you're printing:

clk: Failed to runtime PM get '<clk_name>' for clk '<dev_name>'

With the printout fixed and some type of documentation that clk_pm_runtime_get_all() and clk_pm_runtime_put_all() grab/release the mutex:

author maintainer
#9 lore raw

Re: [PATCH 4/5] clk: Get runtime PM before walking tree during disable_unused

Quoting Doug Anderson (2024-03-25 09:19:37)

Quoted 43 lines

Hi,

Quoted 9 lines · Stephen Boyd

Introduce a list of clk_core structures that have been registered, or are in the process of being registered, that require runtime PM to operate. Iterate this list and call clk_pm_runtime_get() on each of them without holding the prepare_lock during clk_disable_unused(). This way we can be certain that the runtime PM state of the devices will be active and resumed so we can't schedule away while walking the clk tree with the prepare_lock held. Similarly, call clk_pm_runtime_put() without the prepare_lock held to properly drop the runtime PM reference.

There's a part of me that worries about the fact that we'll now be doing a pm_runtime get() on _all clocks_ (even those that are used) at bootup now. I worry that some device out there will be unhappy about it. ...but I guess the device passed in here is already documented to be one that the clock framework can get/put whenever it needs to prepare the clock, so that makes me feel like it should be fine.

Anyway, no action item, just documenting my thoughts...

Oh, funny. After reading the next patch, I guess I'm even less concerned. I guess we were already grabbing the pm_runtime state for all clocks while printing the clock summary. While that's a debugfs function, it's still something that many people have likely exercised and it's likely not going to introduce random/long tail problems.

drivers/clk/clk.c static void clk_enable_unlock(unsigned long flags) · line 220

Quoted 9 lines
drivers/clk/clk.c +9 -0
+/*
+ * Call clk_pm_runtime_get() on all runtime PM enabled clks in the clk tree so
+ * that disabling unused clks avoids a deadlock where a device is runtime PM
+ * resuming/suspending and the runtime PM callback is trying to grab the
+ * prepare_lock for something like clk_prepare_enable() while
+ * clk_disable_unused_subtree() holds the prepare_lock and is trying to runtime
+ * PM resume/suspend the device as well.
+ */
+static int clk_pm_runtime_get_all(void)

nit: It'd be nice if this documented that it acquired / held the lock. Could be in comments, or, might as well use the syntax like this (I think):

Quoted 3 lines

__acquires(&clk_rpm_list_lock);

...similar with the put function.

I had that but removed it because on the error path we drop the lock and sparse complains. I don't know how to signal that the lock is held unless an error happens, but I'm a little out of date on sparse now.

Quoted 14 lines

drivers/clk/clk.c static void clk_enable_unlock(unsigned long flags) · line 241

Quoted 13 lines
drivers/clk/clk.c +11 -0
+       /*
+        * Runtime PM "get" all the devices that are needed for the clks
+        * currently registered. Do this without holding the prepare_lock, to
+        * avoid the deadlock.
+        */
+       hlist_for_each_entry(core, &clk_rpm_list, rpm_node) {
+               ret = clk_pm_runtime_get(core);
+               if (ret) {
+                       failed = core;
+                       pr_err("clk: Failed to runtime PM get '%s' for clk '%s'\n",
+                              failed->name, dev_name(failed->dev));
Quoted 4 lines

If I'm reading this correctly, the strings are backward in your error print. Right now you're printing:

clk: Failed to runtime PM get '<clk_name>' for clk '<dev_name>'

Good catch. Thanks!

Quoted 6 lines

With the printout fixed and some type of documentation that clk_pm_runtime_get_all() and clk_pm_runtime_put_all() grab/release the mutex:

#10 lore raw

Re: [PATCH 4/5] clk: Get runtime PM before walking tree during disable_unused

Hi,

Quoted 15 lines · Stephen Boyd
Quoted 15 lines

drivers/clk/clk.c static void clk_enable_unlock(unsigned long flags) · line 220

Quoted 10 lines
drivers/clk/clk.c +9 -0
+/*
+ * Call clk_pm_runtime_get() on all runtime PM enabled clks in the clk tree so
+ * that disabling unused clks avoids a deadlock where a device is runtime PM
+ * resuming/suspending and the runtime PM callback is trying to grab the
+ * prepare_lock for something like clk_prepare_enable() while
+ * clk_disable_unused_subtree() holds the prepare_lock and is trying to runtime
+ * PM resume/suspend the device as well.
+ */
+static int clk_pm_runtime_get_all(void)

nit: It'd be nice if this documented that it acquired / held the lock. Could be in comments, or, might as well use the syntax like this (I think):

Quoted 7 lines
Quoted 3 lines

__acquires(&clk_rpm_list_lock);

...similar with the put function.

I had that but removed it because on the error path we drop the lock and sparse complains. I don't know how to signal that the lock is held unless an error happens, but I'm a little out of date on sparse now.

I'd settle for something in the comments then? Maybe tagged with "Context:" ?

Thanks!

-Doug

author maintainer
#11 lore raw

PATCH 5/5 clk: Get runtime PM before walking tree for clk_summary

In reply to Stephen Boyd

Similar to the previous commit, we should make sure that all devices are runtime resumed before printing the clk_summary through debugfs. Failure to do so would result in a deadlock if the thread is resuming a device to print clk state and that device is also runtime resuming in another thread, e.g the screen is turning on and the display driver is starting up.

Fixes 1bb294a7981c ("clk: Enable/Disable runtime PM for clk_summary")
Cc Taniya Das <[email protected]>
Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/clk/clk.c +5 -1
index 31998ca67b1e..10792599bec1 100644
@@ -3332,7 +3332,7 @@ static int clk_summary_show(struct seq_file *s, void *data)
 	seq_puts(s, "   clock                          count    count    count        rate   accuracy phase  cycle    enable   consumer                         id\n");
 	seq_puts(s, "---------------------------------------------------------------------------------------------------------------------------------------------\n");
 
-
+	clk_pm_runtime_get_all();
 	clk_prepare_lock();
 
 	for (; *lists; lists++)
@@ -3340,6 +3340,7 @@ static int clk_summary_show(struct seq_file *s, void *data)
 			clk_summary_show_subtree(s, c, 0);
 
 	clk_prepare_unlock();
+	clk_pm_runtime_put_all();
 
 	return 0;
 }
@@ -3389,6 +3390,8 @@ static int clk_dump_show(struct seq_file *s, void *data)
 	struct hlist_head **lists = s->private;
 
 	seq_putc(s, '{');
+
+	clk_pm_runtime_get_all();
 	clk_prepare_lock();
 
 	for (; *lists; lists++) {
@@ -3401,6 +3404,7 @@ static int clk_dump_show(struct seq_file *s, void *data)
 	}
 
 	clk_prepare_unlock();
+	clk_pm_runtime_put_all();
 
 	seq_puts(s, "}\n");
 	return 0;
Signature · 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/clk/linux.git/
https://git.kernel.org/pub/scm/linux/kernel/git/sboyd/spmi.git
Reviewed-by on v1 5/5
#12 lore raw

Re: [PATCH 5/5] clk: Get runtime PM before walking tree for clk_summary

Hi,

Quoted 8 lines · Stephen Boyd

Similar to the previous commit, we should make sure that all devices are runtime resumed before printing the clk_summary through debugfs. Failure to do so would result in a deadlock if the thread is resuming a device to print clk state and that device is also runtime resuming in another thread, e.g the screen is turning on and the display driver is starting up.

Quoted 7 lines
Fixes 1bb294a7981c ("clk: Enable/Disable runtime PM for clk_summary")
Cc Taniya Das <[email protected]>
Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

Shouldn't this also squash in a revert of commit 1bb294a7981c ("clk:
Enable/Disable runtime PM for clk_summary")? As it is, clk_summary_show_subtree() is left with an extra/unnecessary clk_pm_runtime_get() / clk_pm_runtime_put(), right?

Other than that, this looks good to me:

author maintainer
#13 lore raw

Re: [PATCH 5/5] clk: Get runtime PM before walking tree for clk_summary

Quoting Doug Anderson (2024-03-25 09:19:51)

Quoted 19 lines

Hi,

Quoted 8 lines · Stephen Boyd

Similar to the previous commit, we should make sure that all devices are runtime resumed before printing the clk_summary through debugfs. Failure to do so would result in a deadlock if the thread is resuming a device to print clk state and that device is also runtime resuming in another thread, e.g the screen is turning on and the display driver is starting up.

Quoted 7 lines
Fixes 1bb294a7981c ("clk: Enable/Disable runtime PM for clk_summary")
Cc Taniya Das <[email protected]>
Cc Douglas Anderson <[email protected]>
---
 drivers/clk/clk.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
Quoted 4 lines

Shouldn't this also squash in a revert of commit 1bb294a7981c ("clk:
Enable/Disable runtime PM for clk_summary")? As it is, clk_summary_show_subtree() is left with an extra/unnecessary clk_pm_runtime_get() / clk_pm_runtime_put(), right?

Sure, it is superfluous now. I suppose it means we can remove clk_pm_runtime_get()/put() calls in clk_{disable,unprepare}_unused_subtree() as well.

Quoted 4 lines

Other than that, this looks good to me:

#14 lore raw
In reply to Stephen Boyd

Hi Stephen,

I only discover this thread today, interesting read!

[email protected] wrote on Sun, 24 Mar 2024 22:43:57 -0700:

Quoted 2 lines

This patch series fixes a deadlock reported[1] on ChromeOS devices (Qualcomm sc7180 Trogdor). To get there, we allow __clk_release() to run

Not only ChromeOS devices are affected, there have been several reports with similar issues on the mailing list, especially on i.MX8MP, where the clock and power management domains are tightly connected.

Quoted 4 lines

without the prepare_lock held. Then we add runtime PM enabled clk_core structs to a list that we iterate and enable runtime PM for each entry before grabbing the prepare_lock to walk the clk tree. The details are in patch #4.

I am happy we ended-up leaning to the same solution: runtime PM calls should no longer happen after acquiring the prepare lock.

Quoted 4 lines

The patch after that is based on the analysis in the disable unused patch. We similarly resume devices from runtime suspend when walking the clk tree for the debugfs clk_summary.

Quoted 9 lines

Unfortunately this doesn't fix all problems with the usage of runtime PM in the clk framework. We still have a problem if preparing a clk happens in parallel to the device providing that clk runtime resuming or suspending. In that case, the task will go to sleep waiting for the runtime PM state to change, and we'll deadlock. This is primarily a problem with the global prepare_lock. I suspect we'll be able to fix this by implementing per-clk locking, because then we will be able to split up the big prepare_lock into smaller locks that don't deadlock on some device runtime PM transitions.

I fear splitting the locks will actually not solve the problems we encounter on i.MX8.

Let me quote some parts of your commit log in patch 4/5:

---8<---

Quoted 9 lines

This is a classic ABBA deadlock. To properly fix the deadlock, we must never runtime PM resume or suspend a device with the clk prepare_lock held. Actually doing that is near impossible today because the global prepare_lock would have to be dropped in the middle of the tree, the device runtime PM resumed/suspended, and then the prepare_lock grabbed again to ensure consistency of the clk tree topology. If anything changes with the clk tree in the meantime, we've lost and will need to start the operation all over again.

Quoted 5 lines

Luckily, most of the time we're simply incrementing or decrementing the runtime PM count on an active device, so we don't have the chance to schedule away with the prepare_lock held. Let's fix this immediate problem that can be triggered more easily by simply booting on Qualcomm sc7180.

--->8---

Regarding your former statement, I don't think it is impossible, this is what I've been trying to do recently. It is really impacting, and must be handled specifically for each situation: I am counting three of them depending on the action, where either the parents, or the children or both sides of the tree should be resumed before continuing (with optionally the new parent, when reparenting explicitly or doing some rate changes which also involve reparenting). I don't yet have a working proof-of-concept -I would have loved to before LPC- but this is promising and I believe doable.

About your second paragraph however, I am asking whether being "most of the time" incrementing or decrementing the runtime PM count is acceptable, because if we ever perform a real state change within the right conditions, we will just deadlock the platform. This is theoretical and is unlikely to happen in general, I agree. But shall we consider this situation too unlikely from happening for just ignoring it? Shall we instead fix it properly and prepare ourselves for future power-optimized architecture with a lot of dependencies between clocks and other power-related subsystems? This is an open question.

Quoted 2 lines

I'll start working on that problem in earnest now because I'm worried we're going to run into that problem very soon.

Is there any public branch I could look into?

On my side I tried to warn about this but got no feedback. I'd have loved to be pointed towards this patchset at that time :) Here is the report if you want to check it out. FYI, I was asking for feedback on very specific questions, which I consider now solved:
https://lore.kernel.org/all/20240527181928.4fc6b5f0@xps-13/

Quoted 16 lines

Stephen Boyd (5):
clk: Remove prepare_lock hold assertion in __clk_release()
clk: Don't hold prepare_lock when calling kref_put()
clk: Initialize struct clk_core kref earlier
clk: Get runtime PM before walking tree during disable_unused
clk: Get runtime PM before walking tree for clk_summary

 drivers/clk/clk.c | 142 +++++++++++++++++++++++++++++++++++++---------
 1 file changed, 115 insertions(+), 27 deletions(-)
Cc Douglas Anderson <[email protected]>
Cc Krzysztof Kozlowski <[email protected]>
Cc Marek Szyprowski <[email protected]>
Cc Taniya Das <[email protected]>
Cc Ulf Hansson <[email protected]>
Quoted 3 lines

[1] https://lore.kernel.org/all/20220922084322.RFC.2.I375b6b9e0a0a5348962f004beb3dafee6a12dfbb@changeid/

base-commit: e8f897f4afef0031fe618a8e94127a0934896aba

Thanks,
Miquèl