Skip to contentSkip to conversation

Possible memory leak in request_firmware()

PATCH Catalin Marinas 6 participants 11 messages Jul 7, 2009 – Jul 10, 2009 Driver core

lore
Share & tools
lore
#0 lore raw

Hi,

I get a couple kmemleak reports like below which I think happen on the failure path (-ENOENT) of a request_firmware() call:

unreferenced object 0xc355fdb0 (size 256):
comm "NetworkManager", pid 2606, jiffies 4294902882 backtrace:

    [<c01e0c3a>] create_object+0xfa/0x250
    [<c01e1e7d>] kmemleak_alloc+0x5d/0x70
    [<c01dac1b>] kmem_cache_alloc+0x14b/0x190
    [<c03a0c4c>] _request_firmware+0x11c/0x530
    [<c03a1102>] request_firmware+0x12/0x20
    [<f95f6591>] iwl_mac_start+0xa1/0x850 [iwlagn]
    [<f8fb08c1>] ieee80211_open+0x2e1/0x860 [mac80211]
    [<c048459a>] dev_open+0xba/0x100
    [<c0483ab9>] dev_change_flags+0x139/0x1d0
    [<c048d392>] do_setlink+0x282/0x410
    [<c048ea81>] rtnl_setlink+0xf1/0x130
    [<c048e285>] rtnetlink_rcv_msg+0x165/0x200
    [<c049fac6>] netlink_rcv_skb+0x76/0xa0
    [<c048e10e>] rtnetlink_rcv+0x1e/0x30
    [<c049f7fb>] netlink_unicast+0x23b/0x250
    [<c04a02db>] netlink_sendmsg+0x1db/0x2d0

The f_dev in _request_firmware() is allocated via the fw_setup_device() and fw_register_device() calls and its class set to firmware_class (the class release function is fw_dev_release).

Commit 6acf70f078ca replaced the kfree(dev) in fw_dev_release() with a put_device() call but my understanding is that the release function is called via put_device -> kobject_put -> kref_put -> koject_release etc. and it should call kfree since it's the last to see this device structure alive.

Because of that, the _request_firmware() function on its -ENOENT error path only calls device_unregister(f_dev) which would eventually call fw_dev_release() but there is no kfree (the subsequent put_device call would just make the kref negative).

The patch below may fix the problem but it's only later tonight that I can test it and confirm:

drivers/base/firmware_class.c +3 -4
index ddeb819..12e6e64 100644
@@ -357,7 +357,7 @@ static void fw_dev_release(struct device *dev)
 	kfree(fw_priv->pages);
 	kfree(fw_priv->fw_id);
 	kfree(fw_priv);
-	put_device(dev);
+	kfree(dev);
 
 	module_put(THIS_MODULE);
 }
@@ -407,14 +407,13 @@ static int fw_register_device(struct device **dev_p, const char *fw_name,
 	retval = device_register(f_dev);
 	if (retval) {
 		dev_err(device, "%s: device_register failed\n", __func__);
+		kfree(fw_priv->fw_id);
 		put_device(f_dev);
-		goto error_kfree_fw_id;
+		return retval;
 	}
 	*dev_p = f_dev;
 	return 0;
 
-error_kfree_fw_id:
-	kfree(fw_priv->fw_id);
 error_kfree:
 	kfree(f_dev);
 	kfree(fw_priv);
Signature · 1 line
Catalin
#1 lore raw
Quoted 2 lines · Catalin Marinas

The patch below may fix the problem but it's only later tonight that I can test it and confirm:

Your patch looks fine to me (didn't test it either), just one minor nit:

Quoted 16 lines
drivers/base/firmware_class.c +1 -1
index ddeb819..12e6e64 100644
@@ -357,7 +357,7 @@ static void fw_dev_release(struct device *dev)
 	kfree(fw_priv->pages);
 	kfree(fw_priv->fw_id);
 	kfree(fw_priv);
-	put_device(dev);
+	kfree(dev);
 
 	module_put(THIS_MODULE);
 }

drivers/base/firmware_class.c static int fw_register_device(struct device **dev_p, const char *fw_name, · line 410

Quoted 4 lines
drivers/base/firmware_class.c +1 -0
@@ -407,14 +407,13 @@ static int fw_register_device(struct device **dev_p, const char *fw_name,
 	retval = device_register(f_dev);
 	if (retval) {
 		dev_err(device, "%s: device_register failed\n", __func__);
+		kfree(fw_priv->fw_id);

fw_priv->fw_id will be freed in the release function, so you don't need to free it here.

Quoted 14 lines
drivers/base/firmware_class.c +1 -3
 		put_device(f_dev);
-		goto error_kfree_fw_id;
+		return retval;
 	}
 	*dev_p = f_dev;
 	return 0;
 
-error_kfree_fw_id:
-	kfree(fw_priv->fw_id);
 error_kfree:
 	kfree(f_dev);
 	kfree(fw_priv);
#2 lore raw
Quoted 7 lines · Cornelia Huck
Quoted 2 lines · Catalin Marinas

The patch below may fix the problem but it's only later tonight that I can test it and confirm:

Your patch looks fine to me (didn't test it either), just one minor nit:

I tested it and it solves this leak.

Quoted 8 lines

drivers/base/firmware_class.c static int fw_register_device(struct device **dev_p, const char *fw_name, · line 410

Quoted 5 lines
drivers/base/firmware_class.c +1 -0
@@ -407,14 +407,13 @@ static int fw_register_device(struct device **dev_p, const char *fw_name,
 	retval = device_register(f_dev);
 	if (retval) {
 		dev_err(device, "%s: device_register failed\n", __func__);
+		kfree(fw_priv->fw_id);

fw_priv->fw_id will be freed in the release function, so you don't need to free it here.

OK, thanks.

There is one more leak in this area which I couldn't figure out where it should be freed:

unreferenced object 0xc353e530 (size 512):
comm "cat", pid 3130, jiffies 4294903232
backtrace:

    [<c01e6f6a>] create_object+0xfa/0x250
    [<c01e753d>] kmemleak_alloc+0x5d/0x70
    [<c01e223d>] __kmalloc+0x10d/0x210
    [<c03b2d2f>] firmware_data_write+0x1df/0x270
    [<c024163a>] write+0x13a/0x1b0
    [<c01eae1c>] vfs_write+0x9c/0x190
    [<c01eafcd>] sys_write+0x3d/0x70
    [<c010319c>] sysenter_do_call+0x12/0x38
    [<ffffffff>] 0xffffffff

Any idea? It looks like this is the kmalloc() in fw_realloc_buffer() (inlined in firmware_data_write).

Signature · 1 line
Catalin
#3 lore raw

2009/7/8 Catalin Marinas <[email protected]>:

Quoted 16 lines · Cornelia Huck

There is one more leak in this area which I couldn't figure out where it should be freed:

unreferenced object 0xc353e530 (size 512):
 comm "cat", pid 3130, jiffies 4294903232
 backtrace:

   [<c01e6f6a>] create_object+0xfa/0x250
   [<c01e753d>] kmemleak_alloc+0x5d/0x70
   [<c01e223d>] __kmalloc+0x10d/0x210
   [<c03b2d2f>] firmware_data_write+0x1df/0x270
   [<c024163a>] write+0x13a/0x1b0
   [<c01eae1c>] vfs_write+0x9c/0x190
   [<c01eafcd>] sys_write+0x3d/0x70
   [<c010319c>] sysenter_do_call+0x12/0x38
Quoted 4 lines

   [<ffffffff>] 0xffffffff

Any idea? It looks like this is the kmalloc() in fw_realloc_buffer() (inlined in firmware_data_write).

I guess the leak is introduced in commit :
commit 6e03a201bbe8137487f340d26aa662110e324b20
firmware: speed up request_firmware(), v3

The attachment patch may fix the leak, please test and verify it. Thanks.

Signature · 1 line
Lei Ming
drivers/base/firmware_class.c +0 -1
index 2da4803..271dc6b 100644
@@ -180,7 +180,6 @@ static ssize_t firmware_loading_store(struct device *dev,
 				goto err;
 			}
 			/* Pages will be freed by vfree() */
-			fw_priv->pages = NULL;
 			fw_priv->page_array_size = 0;
 			fw_priv->nr_pages = 0;
 			complete(&fw_priv->completion);
#4 lore raw
Quoted 22 lines · Ming Lei

2009/7/8 Catalin Marinas <[email protected]>:

Quoted 18 lines · Cornelia Huck

There is one more leak in this area which I couldn't figure out where it should be freed:

unreferenced object 0xc353e530 (size 512):
comm "cat", pid 3130, jiffies 4294903232
backtrace:

   [<c01e6f6a>] create_object+0xfa/0x250
   [<c01e753d>] kmemleak_alloc+0x5d/0x70
   [<c01e223d>] __kmalloc+0x10d/0x210
   [<c03b2d2f>] firmware_data_write+0x1df/0x270
   [<c024163a>] write+0x13a/0x1b0
   [<c01eae1c>] vfs_write+0x9c/0x190
   [<c01eafcd>] sys_write+0x3d/0x70
   [<c010319c>] sysenter_do_call+0x12/0x38
   [<ffffffff>] 0xffffffff
Quoted 2 lines

Any idea? It looks like this is the kmalloc() in fw_realloc_buffer() (inlined in firmware_data_write).

Quoted 6 lines

I guess the leak is introduced in commit :
commit 6e03a201bbe8137487f340d26aa662110e324b20
firmware: speed up request_firmware(), v3

The attachment patch may fix the leak, please test and verify it. Thanks.

I think you need to stop it from clearing fw_priv->nr_pages too. With that change, it looks correct. Thanks.

Signature · 2 lines
David Woodhouse                            Open Source Technology Centre
[email protected]                              Intel Corporation
#5 lore raw

2009/7/8 David Woodhouse <[email protected]>:

Quoted 27 lines · Ming Lei
Quoted 26 lines

2009/7/8 Catalin Marinas <[email protected]>:

Quoted 16 lines · Cornelia Huck

There is one more leak in this area which I couldn't figure out where it should be freed:

unreferenced object 0xc353e530 (size 512):
 comm "cat", pid 3130, jiffies 4294903232
 backtrace:

   [<c01e6f6a>] create_object+0xfa/0x250
   [<c01e753d>] kmemleak_alloc+0x5d/0x70
   [<c01e223d>] __kmalloc+0x10d/0x210
   [<c03b2d2f>] firmware_data_write+0x1df/0x270
   [<c024163a>] write+0x13a/0x1b0
   [<c01eae1c>] vfs_write+0x9c/0x190
   [<c01eafcd>] sys_write+0x3d/0x70
   [<c010319c>] sysenter_do_call+0x12/0x38
Quoted 4 lines

   [<ffffffff>] 0xffffffff

Any idea? It looks like this is the kmalloc() in fw_realloc_buffer() (inlined in firmware_data_write).

I guess the leak is introduced in commit :
             commit   6e03a201bbe8137487f340d26aa662110e324b20
             firmware: speed up request_firmware(), v3

Quoted 5 lines
Quoted 2 lines

The attachment patch may fix the leak, please test and verify it. Thanks.

I think you need to stop it from clearing fw_priv->nr_pages too. With that change, it looks correct. Thanks.

IMHO, No.
If nr_pages is not cleaned, pages pointed by pages[] will be freed by fw_dev_release, but they should be freed by vfree() in release_firmware. Right?

Quoted 2 lines
Signature · 1 line
Lei Ming
#6 lore raw
Quoted 5 lines · Ming Lei

IMHO, No.
If nr_pages is not cleaned, pages pointed by pages[] will be freed by fw_dev_release, but they should be freed by vfree() in release_firmware. Right?

Yes, sorry -- you're right, and the original patch is fine.

Signature · 2 lines
David Woodhouse                            Open Source Technology Centre
[email protected]                              Intel Corporation
#7 lore raw
In reply to Catalin Marinas

2009/7/7 Catalin Marinas <[email protected]>:

Quoted 49 lines

Hi,

I get a couple kmemleak reports like below which I think happen on the failure path (-ENOENT) of a request_firmware() call:

unreferenced object 0xc355fdb0 (size 256):
 comm "NetworkManager", pid 2606, jiffies 4294902882 backtrace:

   [<c01e0c3a>] create_object+0xfa/0x250
   [<c01e1e7d>] kmemleak_alloc+0x5d/0x70
   [<c01dac1b>] kmem_cache_alloc+0x14b/0x190
   [<c03a0c4c>] _request_firmware+0x11c/0x530
   [<c03a1102>] request_firmware+0x12/0x20
   [<f95f6591>] iwl_mac_start+0xa1/0x850 [iwlagn]
   [<f8fb08c1>] ieee80211_open+0x2e1/0x860 [mac80211]
   [<c048459a>] dev_open+0xba/0x100
   [<c0483ab9>] dev_change_flags+0x139/0x1d0
   [<c048d392>] do_setlink+0x282/0x410
   [<c048ea81>] rtnl_setlink+0xf1/0x130
   [<c048e285>] rtnetlink_rcv_msg+0x165/0x200
   [<c049fac6>] netlink_rcv_skb+0x76/0xa0
   [<c048e10e>] rtnetlink_rcv+0x1e/0x30
   [<c049f7fb>] netlink_unicast+0x23b/0x250
   [<c04a02db>] netlink_sendmsg+0x1db/0x2d0

The f_dev in _request_firmware() is allocated via the fw_setup_device() and fw_register_device() calls and its class set to firmware_class (the class release function is fw_dev_release).

Commit 6acf70f078ca replaced the kfree(dev) in fw_dev_release() with a put_device() call but my understanding is that the release function is called via put_device -> kobject_put -> kref_put -> koject_release etc. and it should call kfree since it's the last to see this device structure alive.

Because of that, the _request_firmware() function on its -ENOENT error path only calls device_unregister(f_dev) which would eventually call fw_dev_release() but there is no kfree (the subsequent put_device call would just make the kref negative).

The patch below may fix the problem but it's only later tonight that I can test it and confirm:

drivers/base/firmware_class.c +0 -0
index ddeb819..12e6e64 100644

drivers/base/firmware_class.c static void fw_dev_release(struct device *dev) · line 360

Quoted 5 lines
drivers/base/firmware_class.c +1 -1
@@ -357,7 +357,7 @@ static void fw_dev_release(struct device *dev)
        kfree(fw_priv->pages);
        kfree(fw_priv->fw_id);
        kfree(fw_priv);
-       put_device(dev);
+       kfree(dev);

Thanks.

Quoted 26 lines
drivers/base/firmware_class.c +2 -3
        module_put(THIS_MODULE);
  }
@@ -407,14 +407,13 @@ static int fw_register_device(struct device **dev_p, const char *fw_name,
        retval = device_register(f_dev);
        if (retval) {
                dev_err(device, "%s: device_register failed\n", __func__);
+               kfree(fw_priv->fw_id);
                put_device(f_dev);
-               goto error_kfree_fw_id;
+               return retval;
        }
        *dev_p = f_dev;
        return 0;
 
-error_kfree_fw_id:
-       kfree(fw_priv->fw_id);

 error_kfree:
       kfree(f_dev);
       kfree(fw_priv);

--
Catalin

Signature · 1 line
Lei Ming
#8 lore raw
In reply to Catalin Marinas
Quoted 4 lines · Catalin Marinas

Hi,

I get a couple kmemleak reports like below which I think happen on the failure path (-ENOENT) of a request_firmware() call:

<snip>

Hm, can you send me the final version of this patch, it seemed to go through a few different versions, with a signed-off-by line?

thanks,

greg k-h

#9 lore raw
Quoted 6 lines · Greg KH
Quoted 2 lines · Catalin Marinas

I get a couple kmemleak reports like below which I think happen on the failure path (-ENOENT) of a request_firmware() call:

Hm, can you send me the final version of this patch, it seemed to go through a few different versions, with a signed-off-by line?

It was merged as commit 0f2f2221b4ad8.

Signature · 1 line
Catalin
#10 lore raw
Quoted 9 lines · Catalin Marinas
Quoted 6 lines · Greg KH
Quoted 2 lines · Catalin Marinas

I get a couple kmemleak reports like below which I think happen on the failure path (-ENOENT) of a request_firmware() call:

Hm, can you send me the final version of this patch, it seemed to go through a few different versions, with a signed-off-by line?

It was merged as commit 0f2f2221b4ad8.

Ah, nice, I don't have to worry about it then, thanks.

greg k-h