Skip to contentSkip to conversation

lightnvm: missing nvm_lock acquire

PATCH In review Wenwei Tao 3 participants 9 messages Nov 24, 2015 – Nov 25, 2015

lore
Share & tools
lore

Wenwei Tao

author
#0 lore raw

PATCH 1/1 lightnvm: missing nvm_lock acquire

To avoid race conditions, traverse dev, media manager, and targeet lists and also register, unregister entries to/from them, should be always under the nvm_lock control.

---
 drivers/lightnvm/core.c | 19 +++++++++++++------
 1 file changed, 13 insertions(+), 6 deletions(-)
drivers/lightnvm/core.c +13 -6
index f659e60..39aec3a 100644
@@ -333,19 +333,19 @@ EXPORT_SYMBOL(nvm_register);
 
 void nvm_unregister(char *disk_name)
 {
+	down_write(&nvm_lock);
 	struct nvm_dev *dev = nvm_find_nvm_dev(disk_name);
 
 	if (!dev) {
 		pr_err("nvm: could not find device %s to unregister\n",
 								disk_name);
+		up_write(&nvm_lock);
 		return;
 	}
 
-	nvm_exit(dev);
-
-	down_write(&nvm_lock);
 	list_del(&dev->devices);
 	up_write(&nvm_lock);
+	nvm_exit(dev);
 }
 EXPORT_SYMBOL(nvm_unregister);
 
@@ -365,12 +365,15 @@ static int nvm_create_target(struct nvm_dev *dev,
 	void *targetdata;
 	int ret = 0;
 
+	down_write(&nvm_lock);
 	if (!dev->mt) {
 		/* register with device with a supported NVM manager */
 		list_for_each_entry(mt, &nvm_mgrs, list) {
 			ret = mt->register_mgr(dev);
-			if (ret < 0)
+			if (ret < 0) {
+				up_write(&nvm_lock);
 				return ret; /* initialization failed */
+			}
 			if (ret > 0) {
 				dev->mt = mt;
 				break; /* successfully initialized */
@@ -379,6 +382,7 @@ static int nvm_create_target(struct nvm_dev *dev,
 
 		if (!ret) {
 			pr_info("nvm: no compatible nvm manager found.\n");
+			up_write(&nvm_lock);
 			return -ENODEV;
 		}
 	}
@@ -386,10 +390,10 @@ static int nvm_create_target(struct nvm_dev *dev,
 	tt = nvm_find_target_type(create->tgttype);
 	if (!tt) {
 		pr_err("nvm: target type %s not found\n", create->tgttype);
+		up_write(&nvm_lock);
 		return -EINVAL;
 	}
 
-	down_write(&nvm_lock);
 	list_for_each_entry(t, &dev->online_targets, list) {
 		if (!strcmp(create->tgtname, t->disk->disk_name)) {
 			pr_err("nvm: target name already exists.\n");
@@ -472,8 +476,9 @@ static int __nvm_configure_create(struct nvm_ioctl_create *create)
 {
 	struct nvm_dev *dev;
 	struct nvm_ioctl_create_simple *s;
-
+	down_write(&nvm_lock);
 	dev = nvm_find_nvm_dev(create->dev);
+	up_write(&nvm_lock);
 	if (!dev) {
 		pr_err("nvm: device not found\n");
 		return -EINVAL;
@@ -532,7 +537,9 @@ static int nvm_configure_show(const char *val)
 		return -EINVAL;
 	}
 
+	down_write(&nvm_lock);
 	dev = nvm_find_nvm_dev(devname);
+	up_write(&nvm_lock);
 	if (!dev) {
 		pr_err("nvm: device not found\n");
 		return -EINVAL;
Signature · 1 line
1.9.1
#1 lore raw
reproduce:

Hi Wenwei,

[auto build test WARNING on v4.4-rc2]
[also build test WARNING on next-20151124]

url:    https://github.com/0day-ci/linux/commits/Wenwei-Tao/lightnvm-missing-nvm_lock-acquire/20151124-232858
config: x86_64-randconfig-x016-11241713 (attached as .config)
reproduce:
        # save the attached .config to linux build tree
        make ARCH=x86_64 

All warnings (new ones prefixed by >>):

drivers/lightnvm/core.c: In function 'nvm_unregister':

>> drivers/lightnvm/core.c:337:2: warning: ISO C90 forbids mixed declarations and code [-Wdeclaration-after-statement]
     struct nvm_dev *dev = nvm_find_nvm_dev(disk_name);
     ^

vim +337 drivers/lightnvm/core.c

cd9e9808 Matias Bjørling 2015-10-28  321  		}
cd9e9808 Matias Bjørling 2015-10-28  322  	} else if (dev->ops->max_phys_sect > 256) {
cd9e9808 Matias Bjørling 2015-10-28  323  		pr_info("nvm: max sectors supported is 256.\n");
cd9e9808 Matias Bjørling 2015-10-28  324  		return -EINVAL;
cd9e9808 Matias Bjørling 2015-10-28  325  	}
cd9e9808 Matias Bjørling 2015-10-28  326  
cd9e9808 Matias Bjørling 2015-10-28  327  	return 0;
cd9e9808 Matias Bjørling 2015-10-28  328  err_init:
cd9e9808 Matias Bjørling 2015-10-28  329  	kfree(dev);
cd9e9808 Matias Bjørling 2015-10-28  330  	return ret;
cd9e9808 Matias Bjørling 2015-10-28  331  }
cd9e9808 Matias Bjørling 2015-10-28  332  EXPORT_SYMBOL(nvm_register);
cd9e9808 Matias Bjørling 2015-10-28  333  
cd9e9808 Matias Bjørling 2015-10-28  334  void nvm_unregister(char *disk_name)
cd9e9808 Matias Bjørling 2015-10-28  335  {
0d5267d8 Wenwei Tao      2015-11-24  336  	down_write(&nvm_lock);
cd9e9808 Matias Bjørling 2015-10-28 @337  	struct nvm_dev *dev = nvm_find_nvm_dev(disk_name);
cd9e9808 Matias Bjørling 2015-10-28  338  
cd9e9808 Matias Bjørling 2015-10-28  339  	if (!dev) {
cd9e9808 Matias Bjørling 2015-10-28  340  		pr_err("nvm: could not find device %s to unregister\n",
cd9e9808 Matias Bjørling 2015-10-28  341  								disk_name);
0d5267d8 Wenwei Tao      2015-11-24  342  		up_write(&nvm_lock);
cd9e9808 Matias Bjørling 2015-10-28  343  		return;
cd9e9808 Matias Bjørling 2015-10-28  344  	}
cd9e9808 Matias Bjørling 2015-10-28  345  

:::::: The code at line 337 was first introduced by commit
:::::: cd9e9808d18fe7107c306f6e71c8be7230ee42b4 lightnvm: Support for Open-Channel SSDs

:::::: TO: Matias Bjørling <[email protected]>
:::::: CC: Jens Axboe <[email protected]>

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation

#2 lore raw
In reply to Wenwei Tao
Quoted 14 lines · Wenwei Tao

To avoid race conditions, traverse dev, media manager, and targeet lists and also register, unregister entries to/from them, should be always under the nvm_lock control.

---
  drivers/lightnvm/core.c | 19 +++++++++++++------
  1 file changed, 13 insertions(+), 6 deletions(-)
drivers/lightnvm/core.c +0 -0
index f659e60..39aec3a 100644

drivers/lightnvm/core.c EXPORT_SYMBOL(nvm_register); · line 337

Quoted 5 lines
drivers/lightnvm/core.c +1 -0
@@ -333,19 +333,19 @@ EXPORT_SYMBOL(nvm_register);
 
  void nvm_unregister(char *disk_name)
  {
+	down_write(&nvm_lock);
  	struct nvm_dev *dev = nvm_find_nvm_dev(disk_name);

In nvm_find_nvm_dev, can we instead replace it with list_for_each_entry_safe. Then we are sure that we won't hit a missing element, and will make the other cases significantly cleaner.

Quoted 18 lines
drivers/lightnvm/core.c +2 -3
  	if (!dev) {
  		pr_err("nvm: could not find device %s to unregister\n",
  								disk_name);
+		up_write(&nvm_lock);
  		return;
  	}
 
-	nvm_exit(dev);
-
-	down_write(&nvm_lock);
  	list_del(&dev->devices);
  	up_write(&nvm_lock);
+	nvm_exit(dev);
  }
  EXPORT_SYMBOL(nvm_unregister);
 

drivers/lightnvm/core.c static int nvm_create_target(struct nvm_dev *dev, · line 371

Quoted 7 lines
drivers/lightnvm/core.c +1 -0
@@ -365,12 +365,15 @@ static int nvm_create_target(struct nvm_dev *dev,
  	void *targetdata;
  	int ret = 0;
 
+	down_write(&nvm_lock);
  	if (!dev->mt) {
  		/* register with device with a supported NVM manager */
  		list_for_each_entry(mt, &nvm_mgrs, list) {

Same with this one?

author
#3 lore raw

Hi Matias
I think list_for_each_entry_safe(pos, n, head, member) cannot avoid race condition
the item point by ‘n’ can be deleted and freed in the same time we operate on 'pos'
so lock is still necessary.

2015-11-25 2:36 GMT+08:00 Matias <[email protected]>:

Quoted 56 lines · Wenwei Tao
Quoted 15 lines

To avoid race conditions, traverse dev, media manager, and targeet lists and also register, unregister entries to/from them, should be always under the nvm_lock control.

---
  drivers/lightnvm/core.c | 19 +++++++++++++------
  1 file changed, 13 insertions(+), 6 deletions(-)
drivers/lightnvm/core.c +0 -0
index f659e60..39aec3a 100644

drivers/lightnvm/core.c EXPORT_SYMBOL(nvm_register); · line 337

Quoted 5 lines
drivers/lightnvm/core.c +1 -0
@@ -333,19 +333,19 @@ EXPORT_SYMBOL(nvm_register);
 
  void nvm_unregister(char *disk_name)
  {
+       down_write(&nvm_lock);
        struct nvm_dev *dev = nvm_find_nvm_dev(disk_name);

In nvm_find_nvm_dev, can we instead replace it with list_for_each_entry_safe. Then we are sure that we won't hit a missing element, and will make the other cases significantly cleaner.

Quoted 19 lines
drivers/lightnvm/core.c +2 -3
        if (!dev) {
                pr_err("nvm: could not find device %s to unregister\n",
 
 disk_name);
+               up_write(&nvm_lock);
                return;
        }
 
-       nvm_exit(dev);
-
-       down_write(&nvm_lock);
        list_del(&dev->devices);
        up_write(&nvm_lock);
+       nvm_exit(dev);
  }
  EXPORT_SYMBOL(nvm_unregister);
 

drivers/lightnvm/core.c static int nvm_create_target(struct nvm_dev *dev, · line 371

Quoted 7 lines
drivers/lightnvm/core.c +1 -0
@@ -365,12 +365,15 @@ static int nvm_create_target(struct nvm_dev *dev,
        void *targetdata;
        int ret = 0;
 
+       down_write(&nvm_lock);
        if (!dev->mt) {
                /* register with device with a supported NVM manager */
                list_for_each_entry(mt, &nvm_mgrs, list) {

Same with this one?

#4 lore raw
Quoted 6 lines · Wenwei Tao

Hi Matias
I think list_for_each_entry_safe(pos, n, head, member) cannot avoid race condition
the item point by ‘n’ can be deleted and freed in the same time we operate on 'pos'
so lock is still necessary.

You're right. Brainfart...

I'll apply it with the fixes that kbuild proposed.

#5 lore raw
In reply to Wenwei Tao
Quoted 7 lines · Wenwei Tao

Hi Matias
I think list_for_each_entry_safe(pos, n, head, member) cannot avoid race condition
the item point by ‘n’ can be deleted and freed in the same time we operate on 'pos'
so lock is still necessary.

I've made the patch like this, ok?

To avoid race conditions, traverse dev, media manager, and target lists and also register, unregister entries to/from them, should be always under the nvm_lock control.

---
 drivers/lightnvm/core.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/lightnvm/core.c +4 -1
index ea6dba5..9f901f6 100644
@@ -127,9 +127,11 @@ static struct nvm_dev *nvm_find_nvm_dev(const char *name)
 {
 	struct nvm_dev *dev;
 
+	down_write(&nvm_lock);
 	list_for_each_entry(dev, &nvm_devices, devices)
 		if (!strcmp(name, dev->name))
 			return dev;
+	up_write(&nvm_lock);
 
 	return NULL;
 }
@@ -388,13 +390,14 @@ static int nvm_create_target(struct nvm_dev *dev,
 		}
 	}
 
+	down_write(&nvm_lock);
 	tt = nvm_find_target_type(create->tgttype);
 	if (!tt) {
 		pr_err("nvm: target type %s not found\n", create->tgttype);
+		up_write(&nvm_lock);
 		return -EINVAL;
 	}
 
-	down_write(&nvm_lock);
 	list_for_each_entry(t, &dev->online_targets, list) {
 		if (!strcmp(create->tgtname, t->disk->disk_name)) {
 			pr_err("nvm: target name already exists.\n");
Signature · 1 line
2.1.4
author
#6 lore raw

Hi Matias

Place the lock inside nvm_find_nvm_dev can make code cleaner, but sometime may lack of flexibility, cause redundant lock and unlock, just my opinion. Place the lock inside the function is okay with me, just make sure do the proper up_write(&nvm_lock) in nvm_create_target(), it has a lot 'return'.

I just submit a patch set covering the lock issue and the target issue mentioned before, you can have a look before you apply this patch. Thanks.

2015-11-25 19:37 GMT+08:00 Matias Bjørling <[email protected]>:

Quoted 59 lines · Wenwei Tao
Quoted 8 lines

Hi Matias
I think list_for_each_entry_safe(pos, n, head, member) cannot avoid race condition
the item point by ‘n’ can be deleted and freed in the same time we operate on 'pos'
so lock is still necessary.

I've made the patch like this, ok?

To avoid race conditions, traverse dev, media manager, and target lists and also register, unregister entries to/from them, should be always under the nvm_lock control.

---
 drivers/lightnvm/core.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/lightnvm/core.c +2 -0
index ea6dba5..9f901f6 100644
@@ -127,9 +127,11 @@ static struct nvm_dev *nvm_find_nvm_dev(const char
 *name)
 {
        struct nvm_dev *dev;
 
+       down_write(&nvm_lock);
        list_for_each_entry(dev, &nvm_devices, devices)
                if (!strcmp(name, dev->name))
                        return dev;
+       up_write(&nvm_lock);
 
        return NULL;

}

drivers/lightnvm/core.c +2 -1
@@ -388,13 +390,14 @@ static int nvm_create_target(struct nvm_dev *dev,
                }
        }
 
+       down_write(&nvm_lock);
 
        tt = nvm_find_target_type(create->tgttype);
        if (!tt) {
                pr_err("nvm: target type %s not found\n", create->tgttype);
+               up_write(&nvm_lock);
                return -EINVAL;
        }
 
-       down_write(&nvm_lock);
        list_for_each_entry(t, &dev->online_targets, list) {
                if (!strcmp(create->tgtname, t->disk->disk_name)) {
                        pr_err("nvm: target name already exists.\n");

--
2.1.4

author
#7 lore raw

Or we can add a new function

static struct nvm_dev *nvm_find_nvm_dev_locked(const char *name)

 {
        struct nvm_dev *dev;
+       down_write(&nvm_lock);
        list_for_each_entry(dev, &nvm_devices, devices)
                if (!strcmp(name, dev->name))
                        return dev;

+ up_write(&nvm_lock);

        return NULL;

}

use nvm_find_nvm_dev_locked() and nvm_find_nvm_dev() properly.

2015-11-25 21:11 GMT+08:00 Wenwei Tao <[email protected]>:

Quoted 72 lines

Hi Matias

Place the lock inside nvm_find_nvm_dev can make code cleaner, but sometime may lack of flexibility, cause redundant lock and unlock, just my opinion. Place the lock inside the function is okay with me, just make sure do the proper up_write(&nvm_lock) in nvm_create_target(), it has a lot 'return'.

I just submit a patch set covering the lock issue and the target issue mentioned before, you can have a look before you apply this patch. Thanks.

2015-11-25 19:37 GMT+08:00 Matias Bjørling <[email protected]>:

Quoted 59 lines · Wenwei Tao
Quoted 8 lines

Hi Matias
I think list_for_each_entry_safe(pos, n, head, member) cannot avoid race condition
the item point by ‘n’ can be deleted and freed in the same time we operate on 'pos'
so lock is still necessary.

I've made the patch like this, ok?

To avoid race conditions, traverse dev, media manager, and target lists and also register, unregister entries to/from them, should be always under the nvm_lock control.

---
 drivers/lightnvm/core.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/lightnvm/core.c +2 -0
index ea6dba5..9f901f6 100644
@@ -127,9 +127,11 @@ static struct nvm_dev *nvm_find_nvm_dev(const char
 *name)
 {
        struct nvm_dev *dev;
 
+       down_write(&nvm_lock);
        list_for_each_entry(dev, &nvm_devices, devices)
                if (!strcmp(name, dev->name))
                        return dev;
+       up_write(&nvm_lock);
 
        return NULL;

}

drivers/lightnvm/core.c +2 -1
@@ -388,13 +390,14 @@ static int nvm_create_target(struct nvm_dev *dev,
                }
        }
 
+       down_write(&nvm_lock);
 
        tt = nvm_find_target_type(create->tgttype);
        if (!tt) {
                pr_err("nvm: target type %s not found\n", create->tgttype);
+               up_write(&nvm_lock);
                return -EINVAL;
        }
 
-       down_write(&nvm_lock);
        list_for_each_entry(t, &dev->online_targets, list) {
                if (!strcmp(create->tgtname, t->disk->disk_name)) {
                        pr_err("nvm: target name already exists.\n");

--
2.1.4

#8 lore raw
Quoted 15 lines · Wenwei Tao

Or we can add a new function

static struct nvm_dev *nvm_find_nvm_dev_locked(const char *name)

  {
         struct nvm_dev *dev;
drivers/lightnvm/core.c +2 -0
+       down_write(&nvm_lock);
         list_for_each_entry(dev, &nvm_devices, devices)
                 if (!strcmp(name, dev->name))
                         return dev;
+       up_write(&nvm_lock);
 
         return NULL;
  }
Quoted 2 lines

use nvm_find_nvm_dev_locked() and nvm_find_nvm_dev() properly.

That could also work. Let's go with your original solution. I don't like it completely, but on the other hand. The other iterators build up to the locks in that way. So making a special case for the nvm_find_nvm_dev properly makes the code harder to understand.