Skip to contentSkip to conversation

kmemleak reports firmware loader funnies in iwlwifi

DISCUSSION Dave Jones 3 participants 6 messages Jun 26, 2009 – Jun 29, 2009

Subsystems (3)
lore
Share & tools
lore

Dave Jones

author
#0 lore raw

After enabling kmemleak in the Fedora rawhide kernels, we've mostly seen a flood of what appear to be false positives, but the below looks really suspect..

https://bugzilla.redhat.com/show_bug.cgi?id=507971

Here's the summary...

iwlagn 0000:03:00.0: loaded firmware version 8.24.2.12
kmemleak: Freeing unknown object at 0xffffc90018070000
Pid: 1034, comm: NetworkManager Not tainted 2.6.31-0.25.rc0.git22.fc12.x86_64 #1

Call Trace:
 [<ffffffff81139f74>] delete_object+0x5b/0x13b
 [<ffffffff8113b012>] kmemleak_free+0x5b/0xb5
 [<ffffffff8111dc51>] vfree+0x40/0x68
 [<ffffffff813485e6>] release_firmware+0x49/0x6c
 [<ffffffffa021997c>] ? iwl_mac_start+0xc5c/0x106b [iwlagn]
 [<ffffffffa0219adc>] iwl_mac_start+0xdbc/0x106b [iwlagn]
 [<ffffffff8109df9b>] ? __module_text_address+0x25/0x85

So it appears to be vfree'ing something that it had no knowledge of ever allocating. afaict _request_firmware only vmallocs when it's using a firmware image built into the driver, which isn't the case here, so I'm not sure why we end up trying to vfree instead of kfree when we call release_firmware

anyone know what's going on here?

Dave

#1 lore raw
Quoted 14 lines · Dave Jones

iwlagn 0000:03:00.0: loaded firmware version 8.24.2.12
kmemleak: Freeing unknown object at 0xffffc90018070000
Pid: 1034, comm: NetworkManager Not tainted 2.6.31-0.25.rc0.git22.fc12.x86_64 #1

Call Trace:
 [<ffffffff81139f74>] delete_object+0x5b/0x13b
 [<ffffffff8113b012>] kmemleak_free+0x5b/0xb5
 [<ffffffff8111dc51>] vfree+0x40/0x68
 [<ffffffff813485e6>] release_firmware+0x49/0x6c
 [<ffffffffa021997c>] ? iwl_mac_start+0xc5c/0x106b [iwlagn]
 [<ffffffffa0219adc>] iwl_mac_start+0xdbc/0x106b [iwlagn]
 [<ffffffff8109df9b>] ? __module_text_address+0x25/0x85
Quoted 5 lines

So it appears to be vfree'ing something that it had no knowledge of ever allocating. afaict _request_firmware only vmallocs when it's using a firmware image built into the driver, which isn't the case here, so I'm not sure why we end up trying to vfree instead of kfree when we call release_firmware

It seems that firmware_loading_store uses vmap() to set fw->data and it later calls vfree() on this (maybe as a short-cut to vunmap + __free_page).

Kmemleak doesn't track vmap allocations but tracks vmalloc/vfree calls. A solution here is to track vmap/vunmap calls (setting the block size to 0 or adding extra checks to make sure it doesn't scan I/O or user pages).

Another solution (my preferred one) is to annotate some of the few places where vmap may be used with vfree. In this particular case:

drivers/base/firmware_class.c +7 -0
index ddeb819..26fb808 100644
@@ -179,6 +179,13 @@ static ssize_t firmware_loading_store(struct device *dev,
 				dev_err(dev, "%s: vmap() failed\n", __func__);
 				goto err;
 			}
+			/*
+			 * This block of memory is later freed using vfree.
+			 * Since kmemleak does not track vmap calls, just
+			 * inform it about this block but ignore it during
+			 * scanning.
+			 */
+			kmemleak_alloc(fw_priv->fw->data, 0, -1, GFP_KERNEL);
 			/* Pages will be freed by vfree() */
 			fw_priv->pages = NULL;
 			fw_priv->page_array_size = 0;
Signature · 1 line
Catalin
#2 lore raw

Hi Catalin,

Quoted 8 lines · Catalin Marinas
drivers/base/firmware_class.c +0 -0
index ddeb819..26fb808 100644
@@ -179,6 +179,13 @@ static ssize_t firmware_loading_store(struct device *dev,
                                dev_err(dev, "%s: vmap() failed\n", __func__);
                                goto err;
                        }

drivers/base/firmware_class.c static ssize_t firmware_loading_store(struct device *dev, · line 191

Quoted 10 lines
drivers/base/firmware_class.c +7 -0
+                       /*
+                        * This block of memory is later freed using vfree.
+                        * Since kmemleak does not track vmap calls, just
+                        * inform it about this block but ignore it during
+                        * scanning.
+                        */
+                       kmemleak_alloc(fw_priv->fw->data, 0, -1, GFP_KERNEL);
                        /* Pages will be freed by vfree() */
                        fw_priv->pages = NULL;
                        fw_priv->page_array_size = 0;

Would it be possible to put this hook in vmap() somehow?

#3 lore raw
Quoted 16 lines · Pekka Enberg

Hi Catalin,

Quoted 12 lines · Catalin Marinas
drivers/base/firmware_class.c +4 -0
index ddeb819..26fb808 100644
@@ -179,6 +179,13 @@ static ssize_t firmware_loading_store(struct device *dev,
                               dev_err(dev, "%s: vmap() failed\n", __func__);
                               goto err;
                       }
+                       /*
+                        * This block of memory is later freed using vfree.
+                        * Since kmemleak does not track vmap calls, just
+                        * inform it about this block but ignore it during
Quoted 8 lines

drivers/base/firmware_class.c static ssize_t firmware_loading_store(struct device *dev, · line 191

Quoted 6 lines
drivers/base/firmware_class.c +3 -0
+                        * scanning.
+                        */
+                       kmemleak_alloc(fw_priv->fw->data, 0, -1, GFP_KERNEL);
                       /* Pages will be freed by vfree() */
                       fw_priv->pages = NULL;
                       fw_priv->page_array_size = 0;

Would it be possible to put this hook in vmap() somehow?

It can be (and it could track vmap leaks as well). BTW, is there any use-case where vmap'ed memory may contain pointers? I did a grep but none of the vmap'ed blocks seem to have pointers.

Signature · 1 line
Catalin
#4 lore raw
In reply to Pekka Enberg
Quoted 16 lines · Pekka Enberg

Hi Catalin,

Quoted 12 lines · Catalin Marinas
drivers/base/firmware_class.c +4 -0
index ddeb819..26fb808 100644
@@ -179,6 +179,13 @@ static ssize_t firmware_loading_store(struct device *dev,
                               dev_err(dev, "%s: vmap() failed\n", __func__);
                               goto err;
                       }
+                       /*
+                        * This block of memory is later freed using vfree.
+                        * Since kmemleak does not track vmap calls, just
+                        * inform it about this block but ignore it during
Quoted 8 lines

drivers/base/firmware_class.c static ssize_t firmware_loading_store(struct device *dev, · line 191

Quoted 6 lines
drivers/base/firmware_class.c +3 -0
+                        * scanning.
+                        */
+                       kmemleak_alloc(fw_priv->fw->data, 0, -1, GFP_KERNEL);
                       /* Pages will be freed by vfree() */
                       fw_priv->pages = NULL;
                       fw_priv->page_array_size = 0;

Would it be possible to put this hook in vmap() somehow?

I tried to do this but it has some other implications. If I add the vmap hook, I would need to add vunmap as well. On ARM, at least, iounmap calls vunmap (but not vmap) which means that I would need to add ioremap support as well. That's not a bug issue but this is more like a new feature than a bug fix.

I propose that for now I disable the kmemleak warning for unknown pointers and add a patch to linux-next which tracks ioremap and vmap mappings. Does this sound fine?

Thanks.

Signature · 1 line
Catalin
#5 lore raw

Hi Catalin,

Quoted 9 lines · Catalin Marinas

I tried to do this but it has some other implications. If I add the vmap hook, I would need to add vunmap as well. On ARM, at least, iounmap calls vunmap (but not vmap) which means that I would need to add ioremap support as well. That's not a bug issue but this is more like a new feature than a bug fix.

I propose that for now I disable the kmemleak warning for unknown pointers and add a patch to linux-next which tracks ioremap and vmap mappings. Does this sound fine?

Yup, makes sense.