Skip to contentSkip to conversation

ocxl: Fix concurrent AFU open and device removal

PATCH Applied Frederic Barrat 3 participants 6 messages Jun 24, 2019 – Dec 13, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ocxl: Fix concurrent AFU open and device removal

If an ocxl device is unbound through sysfs at the same time its AFU is being opened by a user process, the open code may dereference freed stuctures, which can lead to kernel oops messages. You'd have to hit a tiny time window, but it's possible. It's fairly easy to test by making the time window bigger artificially.

Fix it with a combination of 2 changes:
- when an AFU device is found in the IDR by looking for the device minor number, we should hold a reference on the device until after the context is allocated. A reference on the AFU structure is kept when the context is allocated, so we can release the reference on the device after the context allocation.
- with the fix above, there's still another even tinier window, between the time the AFU device is found in the IDR and the reference on the device is taken. We can fix this one by removing the IDR entry earlier, when the device setup is removed, instead of waiting for the 'release' device callback. With proper locking around the IDR.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---
mpe: this fixes a commit merged in v5.2-rc1. It's late, and I don't think it's that important. If it's for the next merge window, I would add:
Cc: [email protected] # v5.2

drivers/misc/ocxl/file.c | 23 +++++++++++------------

 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/misc/ocxl/file.c +11 -12
index 2870c25da166..4d1b44de1492 100644
@@ -18,18 +18,15 @@ static struct class *ocxl_class;
 static struct mutex minors_idr_lock;
 static struct idr minors_idr;
 
-static struct ocxl_file_info *find_file_info(dev_t devno)
+static struct ocxl_file_info *find_and_get_file_info(dev_t devno)
 {
 	struct ocxl_file_info *info;
 
-	/*
-	 * We don't declare an RCU critical section here, as our AFU
-	 * is protected by a reference counter on the device. By the time the
-	 * info reference is removed from the idr, the ref count of
-	 * the device is already at 0, so no user API will access that AFU and
-	 * this function can't return it.
-	 */
+	mutex_lock(&minors_idr_lock);
 	info = idr_find(&minors_idr, MINOR(devno));
+	if (info)
+		get_device(&info->dev);
+	mutex_unlock(&minors_idr_lock);
 	return info;
 }
 
@@ -58,14 +55,16 @@ static int afu_open(struct inode *inode, struct file *file)
 
 	pr_debug("%s for device %x\n", __func__, inode->i_rdev);
 
-	info = find_file_info(inode->i_rdev);
+	info = find_and_get_file_info(inode->i_rdev);
 	if (!info)
 		return -ENODEV;
 
 	rc = ocxl_context_alloc(&ctx, info->afu, inode->i_mapping);
-	if (rc)
+	if (rc) {
+		put_device(&info->dev);
 		return rc;
-
+	}
+	put_device(&info->dev);
 	file->private_data = ctx;
 	return 0;
 }
@@ -487,7 +486,6 @@ static void info_release(struct device *dev)
 {
 	struct ocxl_file_info *info = container_of(dev, struct ocxl_file_info, dev);
 
-	free_minor(info);
 	ocxl_afu_put(info->afu);
 	kfree(info);
 }
@@ -577,6 +575,7 @@ void ocxl_file_unregister_afu(struct ocxl_afu *afu)
 
 	ocxl_file_make_invisible(info);
 	ocxl_sysfs_unregister_afu(info);
+	free_minor(info);
 	device_unregister(&info->dev);
 }
 
Signature · 1 line
2.21.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 59 lines · On Mon, 24 Jun 2019 16:41:48 +0200 Frederic Barrat <[email protected]>…

If an ocxl device is unbound through sysfs at the same time its AFU is being opened by a user process, the open code may dereference freed stuctures, which can lead to kernel oops messages. You'd have to hit a tiny time window, but it's possible. It's fairly easy to test by making the time window bigger artificially.

Fix it with a combination of 2 changes:
- when an AFU device is found in the IDR by looking for the device minor number, we should hold a reference on the device until after the context is allocated. A reference on the AFU structure is kept when the context is allocated, so we can release the reference on the device after the context allocation.
- with the fix above, there's still another even tinier window, between the time the AFU device is found in the IDR and the reference on the device is taken. We can fix this one by removing the IDR entry earlier, when the device setup is removed, instead of waiting for the 'release' device callback. With proper locking around the IDR.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---
mpe: this fixes a commit merged in v5.2-rc1. It's late, and I don't think it's that important. If it's for the next merge window, I would add:
Cc: [email protected] # v5.2

drivers/misc/ocxl/file.c | 23 +++++++++++------------

 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/misc/ocxl/file.c +5 -8
index 2870c25da166..4d1b44de1492 100644
@@ -18,18 +18,15 @@ static struct class *ocxl_class;
 static struct mutex minors_idr_lock;
 static struct idr minors_idr;
 
-static struct ocxl_file_info *find_file_info(dev_t devno)
+static struct ocxl_file_info *find_and_get_file_info(dev_t devno)
 {
 	struct ocxl_file_info *info;
 
-	/*
-	 * We don't declare an RCU critical section here, as our AFU
-	 * is protected by a reference counter on the device. By the time the
-	 * info reference is removed from the idr, the ref count of
-	 * the device is already at 0, so no user API will access that AFU and
-	 * this function can't return it.
-	 */
+	mutex_lock(&minors_idr_lock);
 	info = idr_find(&minors_idr, MINOR(devno));
+	if (info)
+		get_device(&info->dev);
+	mutex_unlock(&minors_idr_lock);
 	return info;
 }
 
@@ -58,14 +55,16 @@ static int afu_open(struct inode *inode, struct file *file)
 
 	pr_debug("%s for device %x\n", __func__, inode->i_rdev);

drivers/misc/ocxl/file.c static int afu_open(struct inode *inode, struct file *file) · line 64

Quoted 10 lines
drivers/misc/ocxl/file.c +3 -2
 
-	info = find_file_info(inode->i_rdev);
+	info = find_and_get_file_info(inode->i_rdev);
 	if (!info)
 		return -ENODEV;
 
 	rc = ocxl_context_alloc(&ctx, info->afu, inode->i_mapping);
-	if (rc)
+	if (rc) {
+		put_device(&info->dev);

You could have a single call site for put_device() since it's needed for both branches. No big deal.

Quoted 16 lines
drivers/misc/ocxl/file.c +2 -2
 		return rc;
-
+	}
+	put_device(&info->dev);
 	file->private_data = ctx;
 	return 0;
 }
@@ -487,7 +486,6 @@ static void info_release(struct device *dev)
 {
 	struct ocxl_file_info *info = container_of(dev, struct ocxl_file_info, dev);
 
-	free_minor(info);
 	ocxl_afu_put(info->afu);
 	kfree(info);
 }

drivers/misc/ocxl/file.c void ocxl_file_unregister_afu(struct ocxl_afu *afu) · line 578

Quoted 4 lines
drivers/misc/ocxl/file.c +1 -0
@@ -577,6 +575,7 @@ void ocxl_file_unregister_afu(struct ocxl_afu *afu)
 
 	ocxl_file_make_invisible(info);
 	ocxl_sysfs_unregister_afu(info);
+	free_minor(info);

Since the IDR entry is added by ocxl_file_register_afu(), it seems to make sense to undo that in ocxl_file_unregister_afu(). Out of curiosity, was there any historical reason to do this in info_release() in the first place ?

Quoted 3 lines
drivers/misc/ocxl/file.c +0 -0
 	device_unregister(&info->dev);
 }
 
#2 lore raw

Le 24/06/2019 à 17:24, Greg Kurz a écrit :

Quoted 68 lines
Quoted 65 lines · On Mon, 24 Jun 2019 16:41:48 +0200 Frederic Barrat <[email protected]>…

If an ocxl device is unbound through sysfs at the same time its AFU is being opened by a user process, the open code may dereference freed stuctures, which can lead to kernel oops messages. You'd have to hit a tiny time window, but it's possible. It's fairly easy to test by making the time window bigger artificially.

Fix it with a combination of 2 changes:
- when an AFU device is found in the IDR by looking for the device minor number, we should hold a reference on the device until after the context is allocated. A reference on the AFU structure is kept when the context is allocated, so we can release the reference on the device after the context allocation.
- with the fix above, there's still another even tinier window, between the time the AFU device is found in the IDR and the reference on the device is taken. We can fix this one by removing the IDR entry earlier, when the device setup is removed, instead of waiting for the 'release' device callback. With proper locking around the IDR.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---
mpe: this fixes a commit merged in v5.2-rc1. It's late, and I don't think it's that important. If it's for the next merge window, I would add:
Cc: [email protected] # v5.2

drivers/misc/ocxl/file.c | 23 +++++++++++------------

  1 file changed, 11 insertions(+), 12 deletions(-)
drivers/misc/ocxl/file.c +6 -9
index 2870c25da166..4d1b44de1492 100644
@@ -18,18 +18,15 @@ static struct class *ocxl_class;
  static struct mutex minors_idr_lock;
  static struct idr minors_idr;
 
-static struct ocxl_file_info *find_file_info(dev_t devno)
+static struct ocxl_file_info *find_and_get_file_info(dev_t devno)
  {
  	struct ocxl_file_info *info;
 
-	/*
-	 * We don't declare an RCU critical section here, as our AFU
-	 * is protected by a reference counter on the device. By the time the
-	 * info reference is removed from the idr, the ref count of
-	 * the device is already at 0, so no user API will access that AFU and
-	 * this function can't return it.
-	 */
+	mutex_lock(&minors_idr_lock);
  	info = idr_find(&minors_idr, MINOR(devno));
+	if (info)
+		get_device(&info->dev);
+	mutex_unlock(&minors_idr_lock);
  	return info;
  }
 
@@ -58,14 +55,16 @@ static int afu_open(struct inode *inode, struct file *file)
 
  	pr_debug("%s for device %x\n", __func__, inode->i_rdev);
 
-	info = find_file_info(inode->i_rdev);
+	info = find_and_get_file_info(inode->i_rdev);
  	if (!info)
  		return -ENODEV;
 
Quoted 7 lines

drivers/misc/ocxl/file.c static int afu_open(struct inode *inode, struct file *file) · line 64

Quoted 4 lines
drivers/misc/ocxl/file.c +2 -1
  	rc = ocxl_context_alloc(&ctx, info->afu, inode->i_mapping);
-	if (rc)
+	if (rc) {
+		put_device(&info->dev);

You could have a single call site for put_device() since it's needed for both branches. No big deal.

Agreed. Will fix if I end up respinning, but won't if it's the only complaint :-)

Quoted 17 lines
Quoted 17 lines
drivers/misc/ocxl/file.c +2 -2
  		return rc;
-
+	}
+	put_device(&info->dev);
  	file->private_data = ctx;
  	return 0;
  }
@@ -487,7 +486,6 @@ static void info_release(struct device *dev)
  {
  	struct ocxl_file_info *info = container_of(dev, struct ocxl_file_info, dev);
 
-	free_minor(info);
  	ocxl_afu_put(info->afu);
  	kfree(info);
  }
@@ -577,6 +575,7 @@ void ocxl_file_unregister_afu(struct ocxl_afu *afu)
 
Quoted 7 lines

drivers/misc/ocxl/file.c void ocxl_file_unregister_afu(struct ocxl_afu *afu) · line 578

Quoted 3 lines
drivers/misc/ocxl/file.c +1 -0
  	ocxl_file_make_invisible(info);
  	ocxl_sysfs_unregister_afu(info);
+	free_minor(info);

Since the IDR entry is added by ocxl_file_register_afu(), it seems to make sense to undo that in ocxl_file_unregister_afu(). Out of curiosity, was there any historical reason to do this in info_release() in the first place ?

Yeah, it makes a lot of sense to remove the IDR entry in ocxl_file_unregister_afu(), that's where we undo the device. I wish I had noticed during the code reviews.
I don't think there was any good reason to have it in info_release() in the first place. I remember the code went through many iterations to get the reference counting on the AFU structure and device done correctly, but we let that one slip.

I now think the pre-5.2 ocxl code was also exposed to the 2nd window mentioned in the commit log (but the first window is new with the refactoring introduced in 5.2-rc1).

Fred

Quoted 7 lines
Quoted 3 lines
drivers/misc/ocxl/file.c +0 -0
  	device_unregister(&info->dev);
  }
 
#3 lore raw
Quoted 117 lines · On Mon, 24 Jun 2019 17:39:26 +0200 Frederic Barrat <[email protected]>…

Le 24/06/2019 à 17:24, Greg Kurz a écrit :

Quoted 68 lines
Quoted 65 lines · On Mon, 24 Jun 2019 16:41:48 +0200 Frederic Barrat <[email protected]>…

If an ocxl device is unbound through sysfs at the same time its AFU is being opened by a user process, the open code may dereference freed stuctures, which can lead to kernel oops messages. You'd have to hit a tiny time window, but it's possible. It's fairly easy to test by making the time window bigger artificially.

Fix it with a combination of 2 changes:
- when an AFU device is found in the IDR by looking for the device minor number, we should hold a reference on the device until after the context is allocated. A reference on the AFU structure is kept when the context is allocated, so we can release the reference on the device after the context allocation.
- with the fix above, there's still another even tinier window, between the time the AFU device is found in the IDR and the reference on the device is taken. We can fix this one by removing the IDR entry earlier, when the device setup is removed, instead of waiting for the 'release' device callback. With proper locking around the IDR.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---
mpe: this fixes a commit merged in v5.2-rc1. It's late, and I don't think it's that important. If it's for the next merge window, I would add:
Cc: [email protected] # v5.2

drivers/misc/ocxl/file.c | 23 +++++++++++------------

  1 file changed, 11 insertions(+), 12 deletions(-)
drivers/misc/ocxl/file.c +6 -9
index 2870c25da166..4d1b44de1492 100644
@@ -18,18 +18,15 @@ static struct class *ocxl_class;
  static struct mutex minors_idr_lock;
  static struct idr minors_idr;
 
-static struct ocxl_file_info *find_file_info(dev_t devno)
+static struct ocxl_file_info *find_and_get_file_info(dev_t devno)
  {
  	struct ocxl_file_info *info;
 
-	/*
-	 * We don't declare an RCU critical section here, as our AFU
-	 * is protected by a reference counter on the device. By the time the
-	 * info reference is removed from the idr, the ref count of
-	 * the device is already at 0, so no user API will access that AFU and
-	 * this function can't return it.
-	 */
+	mutex_lock(&minors_idr_lock);
  	info = idr_find(&minors_idr, MINOR(devno));
+	if (info)
+		get_device(&info->dev);
+	mutex_unlock(&minors_idr_lock);
  	return info;
  }
 
@@ -58,14 +55,16 @@ static int afu_open(struct inode *inode, struct file *file)
 
  	pr_debug("%s for device %x\n", __func__, inode->i_rdev);
 
-	info = find_file_info(inode->i_rdev);
+	info = find_and_get_file_info(inode->i_rdev);
  	if (!info)
  		return -ENODEV;
 
Quoted 7 lines

drivers/misc/ocxl/file.c static int afu_open(struct inode *inode, struct file *file) · line 64

Quoted 4 lines
drivers/misc/ocxl/file.c +2 -1
  	rc = ocxl_context_alloc(&ctx, info->afu, inode->i_mapping);
-	if (rc)
+	if (rc) {
+		put_device(&info->dev);  

You could have a single call site for put_device() since it's needed for both branches. No big deal.

Agreed. Will fix if I end up respinning, but won't if it's the only complaint :-)

Quoted 17 lines
Quoted 17 lines
drivers/misc/ocxl/file.c +2 -2
  		return rc;
-
+	}
+	put_device(&info->dev);
  	file->private_data = ctx;
  	return 0;
  }
@@ -487,7 +486,6 @@ static void info_release(struct device *dev)
  {
  	struct ocxl_file_info *info = container_of(dev, struct ocxl_file_info, dev);
 
-	free_minor(info);
  	ocxl_afu_put(info->afu);
  	kfree(info);
  }
@@ -577,6 +575,7 @@ void ocxl_file_unregister_afu(struct ocxl_afu *afu)
 
Quoted 7 lines

drivers/misc/ocxl/file.c void ocxl_file_unregister_afu(struct ocxl_afu *afu) · line 578

Quoted 3 lines
drivers/misc/ocxl/file.c +1 -0
  	ocxl_file_make_invisible(info);
  	ocxl_sysfs_unregister_afu(info);
+	free_minor(info);  

Since the IDR entry is added by ocxl_file_register_afu(), it seems to make sense to undo that in ocxl_file_unregister_afu(). Out of curiosity, was there any historical reason to do this in info_release() in the first place ?

Yeah, it makes a lot of sense to remove the IDR entry in ocxl_file_unregister_afu(), that's where we undo the device. I wish I had noticed during the code reviews.
I don't think there was any good reason to have it in info_release() in the first place. I remember the code went through many iterations to get the reference counting on the AFU structure and device done correctly, but we let that one slip.

Quoted 4 lines

I now think the pre-5.2 ocxl code was also exposed to the 2nd window mentioned in the commit log (but the first window is new with the refactoring introduced in 5.2-rc1).

This calls for two separate patches then IMHO.

Quoted 12 lines

Fred

Quoted 7 lines
Quoted 3 lines
drivers/misc/ocxl/file.c +0 -0
  	device_unregister(&info->dev);
  }
 
#4 lore raw

Le 24/06/2019 à 17:50, Greg Kurz a écrit :

Quoted 120 lines
Quoted 117 lines · On Mon, 24 Jun 2019 17:39:26 +0200 Frederic Barrat <[email protected]>…

Le 24/06/2019 à 17:24, Greg Kurz a écrit :

Quoted 68 lines
Quoted 65 lines · On Mon, 24 Jun 2019 16:41:48 +0200 Frederic Barrat <[email protected]>…

If an ocxl device is unbound through sysfs at the same time its AFU is being opened by a user process, the open code may dereference freed stuctures, which can lead to kernel oops messages. You'd have to hit a tiny time window, but it's possible. It's fairly easy to test by making the time window bigger artificially.

Fix it with a combination of 2 changes:
- when an AFU device is found in the IDR by looking for the device minor number, we should hold a reference on the device until after the context is allocated. A reference on the AFU structure is kept when the context is allocated, so we can release the reference on the device after the context allocation.
- with the fix above, there's still another even tinier window, between the time the AFU device is found in the IDR and the reference on the device is taken. We can fix this one by removing the IDR entry earlier, when the device setup is removed, instead of waiting for the 'release' device callback. With proper locking around the IDR.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---
mpe: this fixes a commit merged in v5.2-rc1. It's late, and I don't think it's that important. If it's for the next merge window, I would add:
Cc: [email protected] # v5.2

drivers/misc/ocxl/file.c | 23 +++++++++++------------

   1 file changed, 11 insertions(+), 12 deletions(-)
drivers/misc/ocxl/file.c +6 -9
index 2870c25da166..4d1b44de1492 100644
@@ -18,18 +18,15 @@ static struct class *ocxl_class;
   static struct mutex minors_idr_lock;
   static struct idr minors_idr;
 
-static struct ocxl_file_info *find_file_info(dev_t devno)
+static struct ocxl_file_info *find_and_get_file_info(dev_t devno)
   {
   	struct ocxl_file_info *info;
 
-	/*
-	 * We don't declare an RCU critical section here, as our AFU
-	 * is protected by a reference counter on the device. By the time the
-	 * info reference is removed from the idr, the ref count of
-	 * the device is already at 0, so no user API will access that AFU and
-	 * this function can't return it.
-	 */
+	mutex_lock(&minors_idr_lock);
   	info = idr_find(&minors_idr, MINOR(devno));
+	if (info)
+		get_device(&info->dev);
+	mutex_unlock(&minors_idr_lock);
   	return info;
   }
 
@@ -58,14 +55,16 @@ static int afu_open(struct inode *inode, struct file *file)
 
   	pr_debug("%s for device %x\n", __func__, inode->i_rdev);
 
-	info = find_file_info(inode->i_rdev);
+	info = find_and_get_file_info(inode->i_rdev);
   	if (!info)
   		return -ENODEV;
 
Quoted 7 lines

drivers/misc/ocxl/file.c static int afu_open(struct inode *inode, struct file *file) · line 64

Quoted 4 lines
drivers/misc/ocxl/file.c +2 -1
   	rc = ocxl_context_alloc(&ctx, info->afu, inode->i_mapping);
-	if (rc)
+	if (rc) {
+		put_device(&info->dev);

You could have a single call site for put_device() since it's needed for both branches. No big deal.

Agreed. Will fix if I end up respinning, but won't if it's the only complaint :-)

Quoted 17 lines
Quoted 17 lines
drivers/misc/ocxl/file.c +2 -2
   		return rc;
-
+	}
+	put_device(&info->dev);
   	file->private_data = ctx;
   	return 0;
   }
@@ -487,7 +486,6 @@ static void info_release(struct device *dev)
   {
   	struct ocxl_file_info *info = container_of(dev, struct ocxl_file_info, dev);
 
-	free_minor(info);
   	ocxl_afu_put(info->afu);
   	kfree(info);
   }
@@ -577,6 +575,7 @@ void ocxl_file_unregister_afu(struct ocxl_afu *afu)
 
Quoted 7 lines

drivers/misc/ocxl/file.c void ocxl_file_unregister_afu(struct ocxl_afu *afu) · line 578

Quoted 3 lines
drivers/misc/ocxl/file.c +1 -0
   	ocxl_file_make_invisible(info);
   	ocxl_sysfs_unregister_afu(info);
+	free_minor(info);

Since the IDR entry is added by ocxl_file_register_afu(), it seems to make sense to undo that in ocxl_file_unregister_afu(). Out of curiosity, was there any historical reason to do this in info_release() in the first place ?

Yeah, it makes a lot of sense to remove the IDR entry in ocxl_file_unregister_afu(), that's where we undo the device. I wish I had noticed during the code reviews.
I don't think there was any good reason to have it in info_release() in the first place. I remember the code went through many iterations to get the reference counting on the AFU structure and device done correctly, but we let that one slip.

Quoted 6 lines
Quoted 4 lines

I now think the pre-5.2 ocxl code was also exposed to the 2nd window mentioned in the commit log (but the first window is new with the refactoring introduced in 5.2-rc1).

This calls for two separate patches then IMHO.

Well, splitting this patch in 2 wouldn't help, as the pre-5.2 code was different enough that it wouldn't apply.
I could send a different patch covering just the 2nd window to stable and backport to distros. But considering the likelyhood of hitting the problem, it's going to be low on my list.

Fred

Quoted 14 lines
Quoted 13 lines

Fred

Quoted 7 lines
Quoted 3 lines
drivers/misc/ocxl/file.c +0 -0
   	device_unregister(&info->dev);
   }
 
#5 lore raw
In reply to Frederic Barrat
Quoted 18 lines · Frederic Barrat

If an ocxl device is unbound through sysfs at the same time its AFU is being opened by a user process, the open code may dereference freed stuctures, which can lead to kernel oops messages. You'd have to hit a tiny time window, but it's possible. It's fairly easy to test by making the time window bigger artificially.

Fix it with a combination of 2 changes:
- when an AFU device is found in the IDR by looking for the device minor number, we should hold a reference on the device until after the context is allocated. A reference on the AFU structure is kept when the context is allocated, so we can release the reference on the device after the context allocation.
- with the fix above, there's still another even tinier window, between the time the AFU device is found in the IDR and the reference on the device is taken. We can fix this one by removing the IDR entry earlier, when the device setup is removed, instead of waiting for the 'release' device callback. With proper locking around the IDR.

Quoted 2 lines
Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

Applied to powerpc fixes, thanks.

https://git.kernel.org/powerpc/c/a58d37bce0d21cf7fbd589384c619e465ef2f927

cheers