Skip to contentSkip to conversation

nvme: fix memory leak for power latency tolerance req

PATCH New Yufen Yu 3 participants 4 messages May 16, 2019 – May 17, 2019 NVMe

lore
Share & tools
lore

Yufen Yu

author
#0 lore raw

PATCH 1/1 nvme: fix memory leak for power latency tolerance req

nvme_init_ctrl() initialize latency tolerance controls and kzalloc structdev_pm_qos_request 'req' in nvme_init_ctrl(). The only way to free the 'req' is nvme_init_identify(), when apst is disable and prev_apst is enable.

But, prev_apst is disable when device probe, and we will miss to free it when the device does't support APST.

To fix the problem, we call dev_pm_qos_update_user_latency_tolerance() to nvme_init_identify() only when the device actually supports APST. And it needs to be hidden when remove nvme ctrl.

Fixes c5552fde102fc("nvme: Enable autonomous power state transitions")
Signed-off-by Yufen Yu <yuyufen at huawei.com>
---
 drivers/nvme/host/core.c | 34 ++++++++++++++++++++++++----------
 1 file changed, 24 insertions(+), 10 deletions(-)
drivers/nvme/host/core.c +24 -10
index 15d114787d6a..e10756b7c13f 100644
@@ -2485,6 +2485,21 @@ static int nvme_get_effects_log(struct nvme_ctrl *ctrl)
 	return ret;
 }
 
+static void nvme_expose_latency_tolerance(struct nvme_ctrl *ctrl)
+{
+	ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
+	dev_pm_qos_update_user_latency_tolerance(ctrl->device,
+		min(default_ps_max_latency_us, (unsigned long)S32_MAX));
+
+	dev_pm_qos_expose_latency_tolerance(ctrl->device);
+}
+
+static void nvme_hide_latency_tolerance(struct nvme_ctrl *ctrl)
+{
+	ctrl->device->power.set_latency_tolerance = NULL;
+	dev_pm_qos_hide_latency_tolerance(ctrl->device);
+}
+
 /*
  * Initialize the cached copies of the Identify data and various controller
  * register in our nvme_ctrl structure.  This should be called as soon as
@@ -2643,9 +2658,9 @@ int nvme_init_identify(struct nvme_ctrl *ctrl)
 		return ret;
 
 	if (ctrl->apst_enabled && !prev_apst_enabled)
-		dev_pm_qos_expose_latency_tolerance(ctrl->device);
+		nvme_expose_latency_tolerance(ctrl);
 	else if (!ctrl->apst_enabled && prev_apst_enabled)
-		dev_pm_qos_hide_latency_tolerance(ctrl->device);
+		nvme_hide_latency_tolerance(ctrl);
 
 	ret = nvme_configure_apst(ctrl);
 	if (ret < 0)
@@ -3687,6 +3702,13 @@ EXPORT_SYMBOL_GPL(nvme_start_ctrl);
 
 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
 {
+	/*
+	 * If userspace latency tolerance have been exposed,
+	 * we need to hide it.
+	 */
+	if (ctrl->device->power.set_latency_tolerance)
+		nvme_hide_latency_tolerance(ctrl);
+
 	cdev_device_del(&ctrl->cdev, ctrl->device);
 }
 EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
@@ -3773,14 +3795,6 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
 	if (ret)
 		goto out_free_name;
 
-	/*
-	 * Initialize latency tolerance controls.  The sysfs files won't
-	 * be visible to userspace unless the device actually supports APST.
-	 */
-	ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
-	dev_pm_qos_update_user_latency_tolerance(ctrl->device,
-		min(default_ps_max_latency_us, (unsigned long)S32_MAX));
-
 	return 0;
 out_free_name:
 	kfree_const(ctrl->device->kobj.name);
Signature · 1 line
2.16.2.dirty
#1 lore raw

PATCH 1/1 nvme: fix memory leak for power latency tolerance req

Hi,

Quoted 77 lines · Yufen Yu

nvme_init_ctrl() initialize latency tolerance controls and kzalloc structdev_pm_qos_request 'req' in nvme_init_ctrl(). The only way to free the 'req' is nvme_init_identify(), when apst is disable and prev_apst is enable.

But, prev_apst is disable when device probe, and we will miss to free it when the device does't support APST.

To fix the problem, we call dev_pm_qos_update_user_latency_tolerance() to nvme_init_identify() only when the device actually supports APST. And it needs to be hidden when remove nvme ctrl.

Fixes c5552fde102fc("nvme: Enable autonomous power state transitions")
Signed-off-by Yufen Yu <yuyufen at huawei.com>
---
  drivers/nvme/host/core.c | 34 ++++++++++++++++++++++++----------
  1 file changed, 24 insertions(+), 10 deletions(-)
drivers/nvme/host/core.c +24 -3
index 15d114787d6a..e10756b7c13f 100644
@@ -2485,6 +2485,21 @@ static int nvme_get_effects_log(struct nvme_ctrl *ctrl)
  	return ret;
  }
 
+static void nvme_expose_latency_tolerance(struct nvme_ctrl *ctrl)
+{
+	ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
+	dev_pm_qos_update_user_latency_tolerance(ctrl->device,
+		min(default_ps_max_latency_us, (unsigned long)S32_MAX));
+
+	dev_pm_qos_expose_latency_tolerance(ctrl->device);
+}
+
+static void nvme_hide_latency_tolerance(struct nvme_ctrl *ctrl)
+{
+	ctrl->device->power.set_latency_tolerance = NULL;
+	dev_pm_qos_hide_latency_tolerance(ctrl->device);
+}
+
  /*
   * Initialize the cached copies of the Identify data and various controller
   * register in our nvme_ctrl structure.  This should be called as soon as
@@ -2643,9 +2658,9 @@ int nvme_init_identify(struct nvme_ctrl *ctrl)
  		return ret;
 
  	if (ctrl->apst_enabled && !prev_apst_enabled)
-		dev_pm_qos_expose_latency_tolerance(ctrl->device);
+		nvme_expose_latency_tolerance(ctrl);
  	else if (!ctrl->apst_enabled && prev_apst_enabled)
-		dev_pm_qos_hide_latency_tolerance(ctrl->device);
+		nvme_hide_latency_tolerance(ctrl);
 
  	ret = nvme_configure_apst(ctrl);
  	if (ret < 0)
@@ -3687,6 +3702,13 @@ EXPORT_SYMBOL_GPL(nvme_start_ctrl);
 
  void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
  {
+	/*
+	 * If userspace latency tolerance have been exposed,
+	 * we need to hide it.
+	 */
+	if (ctrl->device->power.set_latency_tolerance)
+		nvme_hide_latency_tolerance(ctrl);
+
  	cdev_device_del(&ctrl->cdev, ctrl->device);
  }
  EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
@@ -3773,14 +3795,6 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
  	if (ret)
  		goto out_free_name;
 
-	/*

drivers/nvme/host/core.c int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, · line 3800

Quoted 10 lines
drivers/nvme/host/core.c +0 -7
-	 * Initialize latency tolerance controls.  The sysfs files won't
-	 * be visible to userspace unless the device actually supports APST.
-	 */
-	ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
-	dev_pm_qos_update_user_latency_tolerance(ctrl->device,
-		min(default_ps_max_latency_us, (unsigned long)S32_MAX));
-
  	return 0;
  out_free_name:
  	kfree_const(ctrl->device->kobj.name);
Tested-by David Milburn <dmilburn at redhat.com>

Before patch running v5.1

$ cat /sys/kernel/debug/kmemleak
unreferenced object 0xffff924143eebfc0 (size 64):
   comm "kworker/1:2", pid 418, jiffies 4294679082 (age 275.968s)
   hex dump (first 32 bytes):
     00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
     00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
   backtrace:
     [<000000001db7845a>] dev_pm_qos_update_user_latency_tolerance+0x63/0xe0
     [<000000008e36e7c0>] nvme_init_ctrl+0x2ec/0x310 [nvme_core]
     [<000000000177edcb>] nvme_probe+0x224/0x3e0 [nvme]
     [<00000000d2397746>] local_pci_probe+0x42/0x80
     [<0000000068201d8c>] work_for_cpu_fn+0x16/0x20
     [<00000000ccf116ec>] process_one_work+0x1a1/0x3a0
     [<000000007b61fbe2>] worker_thread+0x1c9/0x380
     [<0000000073ebfed5>] kthread+0x10c/0x130
     [<000000007f82e90d>] ret_from_fork+0x35/0x40

After patch no more memory leak as indicated above.

Thanks,
David

#2 lore raw

PATCH 1/1 nvme: fix memory leak for power latency tolerance req

In reply to Yufen Yu

It's perfectly safe to call dev_pm_qos_hide_latency_tolerance even if we never exposed it, so I think you're making this more complicated than necessary. Would the following one-liner work for you instead?

---

drivers/nvme/host/core.c +1 -0
index f2411d50e764..0899dc5645a9 100644
@@ -3697,6 +3697,7 @@ EXPORT_SYMBOL_GPL(nvme_start_ctrl);
 
 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
 {
+	dev_pm_qos_hide_latency_tolerance(ctrl->device);
 	cdev_device_del(&ctrl->cdev, ctrl->device);
 }
 EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);

--

author
#3 lore raw

PATCH 1/1 nvme: fix memory leak for power latency tolerance req

Quoted 10 lines · Keith Busch

It's perfectly safe to call dev_pm_qos_hide_latency_tolerance even if we never exposed it, so I think you're making this more complicated than necessary. Would the following one-liner work for you instead?

---

drivers/nvme/host/core.c +0 -0
index f2411d50e764..0899dc5645a9 100644

drivers/nvme/host/core.c EXPORT_SYMBOL_GPL(nvme_start_ctrl); · line 3702

Quoted 6 lines
drivers/nvme/host/core.c +1 -0
@@ -3697,6 +3697,7 @@ EXPORT_SYMBOL_GPL(nvme_start_ctrl);
 
  void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
  {
+	dev_pm_qos_hide_latency_tolerance(ctrl->device);
  	cdev_device_del(&ctrl->cdev, ctrl->device);
  }

Yes, you are right.
I have worried that it will create a new 'req' if it has been free and set 'NULL'.
Thus, I try to hide it based on device->power.set_latency_tolerance.

Actually, I ignore the fact that dev_pm_qos_hide_latency_tolerance() will pass '-1' to
dev_pm_qos_update_user_latency_tolerance(). It will not create a new 'req'.

I will send patch v2 with the suggestion.

Yufen
Thanks a lot

Quoted 5 lines
drivers/nvme/host/core.c +0 -1
  EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
--

.