Re: [ath9k-devel] [3.3-rc2+] Thousands of ath9k warnings on dmesg before laptop froze

[Pavel Roskin <proski@gnu.org>]

[-- Attachment #1: Type: text/plain, Size: 2004 bytes --]

On Mon, 6 Feb 2012 00:29:07 +0000
"Carlos R. Mafra" <crmafra@gmail.com> wrote:

> 
> I'm testing the latest kernel 3.3.0-rc2+ I pulled from git 
> this morning.
> 
> My laptop just froze, and when I rebooted I noticed
> that /var/log/messages contained 48 thousand (!) warnings coming from
> ath9k since a few hours ago. I'm pasting the first one:

> 
>  ------------[ cut here ]------------
>  WARNING:
> at /home/mafra/linux-2.6/drivers/net/wireless/ath/ath9k/rc.c:697
> ath_rc_get_highest_rix+0x156/0x210 [ath9k]() Hardware name: VPCEB4X1E

I believe I found a solution for this today.  Please see this bug
tracker: https://bugzilla.redhat.com/show_bug.cgi?id=768639

While Fedora users report a warning, I've seen panic reports in the
list.  It's a memory corruption bug, so it can manifest in different
ways.  Please test the latest patch (attached).

Here's my comment to the patch:

This patch is based on my analysis of printk() output I added to the
ath9k driver.  I didn't have a chance to test the patch, so testing
would be greatly appreciated.

The corruption must be happening in ath_debug_stat_rc(), which is given
the result of ath_rc_get_rateindex().  ath_rc_get_rateindex() can
return -1, which causes ath_debug_stat_rc() to increment the value that
lies 16 bytes before rcstats in struct ath_rate_priv.  On 64-bit
systems, that happens to be rate_table.  Once the rate_table pointer is
incremented, all data there becomes invalid, which leads to the
warning.  On 32-bit systems, the corruption should happen in
neg_ht_rates.

The -1 value of idx in struct ieee80211_tx_rate is described in
net/mac80211.h. I don't know why we have -1 there and how to reproduce
the problem reliably. But -1 can be there and ath9k has no checks for
it.

The patch introduces two protections: ath_rc_get_rateindex() never
returns a negative value and ath_debug_stat_rc() checks the array
bounds.

It may not be good enough for the kernel, but it may be good enough for
Fedora.

-- 
Regards,
Pavel Roskin

[-- Attachment #2: 01-rix-check.patch --]
[-- Type: text/x-patch, Size: 1280 bytes --]

Prevent memory corruption in ath9k rate control algorithm

From: Pavel Roskin <proski@gnu.org>

Check final_rate in ath_debug_stat_rc().  Don't return negative values
from ath_rc_get_rateindex(), callers don't expect it.

Signed-off-by: Pavel Roskin <proski@gnu.org>
---

 drivers/net/wireless/ath/ath9k/rc.c |   10 ++++++++++
 1 files changed, 10 insertions(+), 0 deletions(-)


diff --git a/drivers/net/wireless/ath/ath9k/rc.c b/drivers/net/wireless/ath/ath9k/rc.c
index 635b592..afe22f4 100644
--- a/drivers/net/wireless/ath/ath9k/rc.c
+++ b/drivers/net/wireless/ath/ath9k/rc.c
@@ -385,6 +385,11 @@ static int ath_rc_get_rateindex(const struct ath_rate_table *rate_table,
 	int rix = 0, i = 0;
 	static const int mcs_rix_off[] = { 7, 15, 20, 21, 22, 23 };
 
+	if (rate->idx < 0) {
+		printk(KERN_ERR "%s: rate->idx = %d\n", __func__, rate->idx);
+		return 0;
+	}
+
 	if (!(rate->flags & IEEE80211_TX_RC_MCS))
 		return rate->idx;
 
@@ -1324,6 +1329,11 @@ static void ath_debug_stat_rc(struct ath_rate_priv *rc, int final_rate)
 {
 	struct ath_rc_stats *stats;
 
+	if (final_rate < 0 || final_rate >= RATE_TABLE_SIZE) {
+		printk(KERN_ERR "%s: invalid final_rate: %d\n", __func__,
+		       final_rate);
+		return;
+	}
 	stats = &rc->rcstats[final_rate];
 	stats->success++;
 }