Re: [TIP] BUG kmalloc-4096: Poison overwritten (ath5k_rx_skb_alloc)

[Bob Copeland <me@bobcopeland.com>]

On Thu, Feb 26, 2009 at 05:03:39PM +0000, Sitsofe Wheeler wrote:
> 
> Note that I was able to reproduce it again : )

By the way, here's the theoretical race I was alluding to.  ath5k_reset 
happens pretty frequently when scanning, and it's possible that the rx
tasklet is run on another cpu after interrupts are turned off, but
it's a small window and I couldn't trigger it with any mdelays.

[PATCH] ath5k: manipulate rxlink and descriptor address under rxbuf lock

Grabbing an ath5k_buf then dropping the lock is racy because the
referenced descriptor can be obtained in another thread and released
before the buffer is handed to the hardware.  Likewise, manipulating
sc->rxlink without the lock can lead to having multiple self-linked
hardware descriptors.

Signed-off-by: Bob Copeland <me@bobcopeland.com>
---
 drivers/net/wireless/ath5k/base.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/drivers/net/wireless/ath5k/base.c b/drivers/net/wireless/ath5k/base.c
index 8d4b11c..0d3adb5 100644
--- a/drivers/net/wireless/ath5k/base.c
+++ b/drivers/net/wireless/ath5k/base.c
@@ -1586,9 +1586,8 @@ ath5k_rx_start(struct ath5k_softc *sc)
 	ATH5K_DBG(sc, ATH5K_DEBUG_RESET, "cachelsz %u rxbufsize %u\n",
 		sc->cachelsz, sc->rxbufsize);
 
-	sc->rxlink = NULL;
-
 	spin_lock_bh(&sc->rxbuflock);
+	sc->rxlink = NULL;
 	list_for_each_entry(bf, &sc->rxbuf, list) {
 		ret = ath5k_rxbuf_setup(sc, bf);
 		if (ret != 0) {
@@ -1597,9 +1596,9 @@ ath5k_rx_start(struct ath5k_softc *sc)
 		}
 	}
 	bf = list_first_entry(&sc->rxbuf, struct ath5k_buf, list);
+	ath5k_hw_set_rxdp(ah, bf->daddr);
 	spin_unlock_bh(&sc->rxbuflock);
 
-	ath5k_hw_set_rxdp(ah, bf->daddr);
 	ath5k_hw_start_rx_dma(ah);	/* enable recv descriptors */
 	ath5k_mode_setup(sc);		/* set filters, etc. */
 	ath5k_hw_start_rx_pcu(ah);	/* re-enable PCU/DMA engine */
-- 
1.6.0.6



-- 
Bob Copeland %% www.bobcopeland.com