kmemleak reports firmware loader funnies in iwlwifi

kmemleak reports firmware loader funnies in iwlwifi

[Dave Jones]

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

Re: kmemleak reports firmware loader funnies in iwlwifi

[Catalin Marinas]

Dave Jones <davej@redhat.com> wrote:
> 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

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:


diff --git a/drivers/base/firmware_class.c b/drivers/base/firmware_class.c
index ddeb819..26fb808 100644
--- a/drivers/base/firmware_class.c
+++ b/drivers/base/firmware_class.c
@@ -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;


-- 
Catalin

Re: kmemleak reports firmware loader funnies in iwlwifi

[Pekka Enberg]

Hi Catalin,

On Mon, Jun 29, 2009 at 12:39 PM, Catalin
Marinas<catalin.marinas@arm.com> wrote:
> diff --git a/drivers/base/firmware_class.c b/drivers/base/firmware_class.c
> index ddeb819..26fb808 100644
> --- a/drivers/base/firmware_class.c
> +++ b/drivers/base/firmware_class.c
> @@ -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;

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

Re: kmemleak reports firmware loader funnies in iwlwifi

[Catalin Marinas]

On Mon, 2009-06-29 at 12:48 +0300, Pekka Enberg wrote:
> Hi Catalin,
> 
> On Mon, Jun 29, 2009 at 12:39 PM, Catalin
> Marinas<catalin.marinas@arm.com> wrote:
> > diff --git a/drivers/base/firmware_class.c b/drivers/base/firmware_class.c
> > index ddeb819..26fb808 100644
> > --- a/drivers/base/firmware_class.c
> > +++ b/drivers/base/firmware_class.c
> > @@ -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;
> 
> 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.

-- 
Catalin

Re: kmemleak reports firmware loader funnies in iwlwifi

[Catalin Marinas]

On Mon, 2009-06-29 at 12:48 +0300, Pekka Enberg wrote:
> Hi Catalin,
> 
> On Mon, Jun 29, 2009 at 12:39 PM, Catalin
> Marinas<catalin.marinas@arm.com> wrote:
> > diff --git a/drivers/base/firmware_class.c b/drivers/base/firmware_class.c
> > index ddeb819..26fb808 100644
> > --- a/drivers/base/firmware_class.c
> > +++ b/drivers/base/firmware_class.c
> > @@ -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;
> 
> 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.

-- 
Catalin

Re: kmemleak reports firmware loader funnies in iwlwifi

[Pekka Enberg]

Hi Catalin,

On Mon, Jun 29, 2009 at 5:33 PM, Catalin Marinas<catalin.marinas@arm.com> wrote:
> 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.