From: Linus Torvalds on


On Mon, 21 Sep 2009, Frederik Deweerdt wrote:
<
> I'd rather wait someone picks it up for mainline inclusion. I've added
> your {Reported,Tested}-by tags.
>
> The bug its vanilla 2.6.31, and should be considered for -stable inclusion.
>
> Regards,
> Frederik
>
> ----
>
> The code introduced by commit 6e03a201bbe8137487f340d26aa662110e324b20 leads
> to a potential null deref. The following patch adds the proper locking
> around the accesses to fw_priv->fw.
> See http://bugzilla.kernel.org/show_bug.cgi?id=14185 for a full bug report.

I don't think this is correct.

I think you should protect the FW_STATUS_LOADING bit too, shouldn't you?

As it is, it does this:

if (test_bit(FW_STATUS_LOADING, &fw_priv->status)) {
mutex_lock(&fw_lock);
...
clear_bit(FW_STATUS_LOADING, &fw_priv->status);
mutex_unlock(&fw_lock);
break;
}

and if this code can race (which it obviously can, since your addition of
fw_lock mutex matters), then I think it can race on that FW_STATUS_LOADING
bit too. No?

So my gut feel is that the whole damn function should be protected by the
mutex_lock thing. IOW, the patch would be something like the appended.

UNTESTED. Somebody needs to test this, verify, and send it back to me.

Am I missing something?

Linus

---
drivers/base/firmware_class.c | 9 ++++-----
1 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/drivers/base/firmware_class.c b/drivers/base/firmware_class.c
index 7376367..1b803df 100644
--- a/drivers/base/firmware_class.c
+++ b/drivers/base/firmware_class.c
@@ -150,13 +150,12 @@ static ssize_t firmware_loading_store(struct device *dev,
int loading = simple_strtol(buf, NULL, 10);
int i;

+ mutex_lock(&fw_lock);
switch (loading) {
case 1:
- mutex_lock(&fw_lock);
- if (!fw_priv->fw) {
- mutex_unlock(&fw_lock);
+ if (!fw_priv->fw)
break;
- }
+
vfree(fw_priv->fw->data);
fw_priv->fw->data = NULL;
for (i = 0; i < fw_priv->nr_pages; i++)
@@ -167,7 +166,6 @@ static ssize_t firmware_loading_store(struct device *dev,
fw_priv->nr_pages = 0;
fw_priv->fw->size = 0;
set_bit(FW_STATUS_LOADING, &fw_priv->status);
- mutex_unlock(&fw_lock);
break;
case 0:
if (test_bit(FW_STATUS_LOADING, &fw_priv->status)) {
@@ -195,6 +193,7 @@ static ssize_t firmware_loading_store(struct device *dev,
fw_load_abort(fw_priv);
break;
}
+ mutex_unlock(&fw_lock);

return count;
}
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Sachin Sant on
Linus Torvalds wrote:
> I don't think this is correct.
>
> I think you should protect the FW_STATUS_LOADING bit too, shouldn't you?
>
> As it is, it does this:
>
> if (test_bit(FW_STATUS_LOADING, &fw_priv->status)) {
> mutex_lock(&fw_lock);
> ...
> clear_bit(FW_STATUS_LOADING, &fw_priv->status);
> mutex_unlock(&fw_lock);
> break;
> }
>
> and if this code can race (which it obviously can, since your addition of
> fw_lock mutex matters), then I think it can race on that FW_STATUS_LOADING
> bit too. No?
>
> So my gut feel is that the whole damn function should be protected by the
> mutex_lock thing. IOW, the patch would be something like the appended.
>
> UNTESTED. Somebody needs to test this, verify, and send it back to me.
>
I did a quick boot test with this patch and didn't find any issues.

But that said i haven't been able to recreate the problem reported by Lars,
so not sure how relevant would be the test results from me.

Thanks
-Sachin


--

---------------------------------
Sachin Sant
IBM Linux Technology Center
India Systems and Technology Labs
Bangalore, India
---------------------------------

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/