ioatdma: fix race between updating ioat->head and IOAT_COMPLETION_PENDING

There is a race that can hit during __cleanup() when the ioat->head pointer is
incremented during descriptor submission. The __cleanup() can clear the
PENDING flag when it does not see any active descriptors. This causes new
submitted descriptors to be ignored because the COMPLETION_PENDING flag is
cleared. This was introduced when code was adapted from ioatdma v1 to ioatdma
v2. For v2 and v3, IOAT_COMPLETION_PENDING flag will be abandoned and a new
flag IOAT_CHAN_ACTIVE will be utilized. This flag will also be protected under
the prep_lock when being modified in order to avoid the race.

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Dan Williams <djbw@fb.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

1 files changed, 64 insertions, 49 deletions

diff --git a/drivers/dma/ioat/dma_v2.c b/drivers/dma/ioat/dma_v2.c
index b9d667851445..e2b2c70f03db 100644
--- a/drivers/dma/ioat/dma_v2.c
+++ b/drivers/dma/ioat/dma_v2.c
@@ -269,61 +269,22 @@ static void ioat2_restart_channel(struct ioat2_dma_chan *ioat)
 	__ioat2_restart_chan(ioat);
 }
 
-void ioat2_timer_event(unsigned long data)
+static void check_active(struct ioat2_dma_chan *ioat)
 {
-	struct ioat2_dma_chan *ioat = to_ioat2_chan((void *) data);
 	struct ioat_chan_common *chan = &ioat->base;
 
-	if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) {
-		dma_addr_t phys_complete;
-		u64 status;
-
-		status = ioat_chansts(chan);
-
-		/* when halted due to errors check for channel
-		 * programming errors before advancing the completion state
-		 */
-		if (is_ioat_halted(status)) {
-			u32 chanerr;
-
-			chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
-			dev_err(to_dev(chan), "%s: Channel halted (%x)\n",
-				__func__, chanerr);
-			if (test_bit(IOAT_RUN, &chan->state))
-				BUG_ON(is_ioat_bug(chanerr));
-			else /* we never got off the ground */
-				return;
-		}
-
-		/* if we haven't made progress and we have already
-		 * acknowledged a pending completion once, then be more
-		 * forceful with a restart
-		 */
-		spin_lock_bh(&chan->cleanup_lock);
-		if (ioat_cleanup_preamble(chan, &phys_complete)) {
-			__cleanup(ioat, phys_complete);
-		} else if (test_bit(IOAT_COMPLETION_ACK, &chan->state)) {
-			spin_lock_bh(&ioat->prep_lock);
-			ioat2_restart_channel(ioat);
-			spin_unlock_bh(&ioat->prep_lock);
-		} else {
-			set_bit(IOAT_COMPLETION_ACK, &chan->state);
-			mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
-		}
-		spin_unlock_bh(&chan->cleanup_lock);
-	} else {
-		u16 active;
+	if (ioat2_ring_active(ioat)) {
+		mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
+		return;
+	}
 
+	if (test_and_clear_bit(IOAT_CHAN_ACTIVE, &chan->state))
+		mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
+	else if (ioat->alloc_order > ioat_get_alloc_order()) {
 		/* if the ring is idle, empty, and oversized try to step
 		 * down the size
 		 */
-		spin_lock_bh(&chan->cleanup_lock);
-		spin_lock_bh(&ioat->prep_lock);
-		active = ioat2_ring_active(ioat);
-		if (active == 0 && ioat->alloc_order > ioat_get_alloc_order())
-			reshape_ring(ioat, ioat->alloc_order-1);
-		spin_unlock_bh(&ioat->prep_lock);
-		spin_unlock_bh(&chan->cleanup_lock);
+		reshape_ring(ioat, ioat->alloc_order - 1);
 
 		/* keep shrinking until we get back to our minimum
 		 * default size
@@ -331,6 +292,60 @@ void ioat2_timer_event(unsigned long data)
 		if (ioat->alloc_order > ioat_get_alloc_order())
 			mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
 	}
+
+}
+
+void ioat2_timer_event(unsigned long data)
+{
+	struct ioat2_dma_chan *ioat = to_ioat2_chan((void *) data);
+	struct ioat_chan_common *chan = &ioat->base;
+	dma_addr_t phys_complete;
+	u64 status;
+
+	status = ioat_chansts(chan);
+
+	/* when halted due to errors check for channel
+	 * programming errors before advancing the completion state
+	 */
+	if (is_ioat_halted(status)) {
+		u32 chanerr;
+
+		chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
+		dev_err(to_dev(chan), "%s: Channel halted (%x)\n",
+			__func__, chanerr);
+		if (test_bit(IOAT_RUN, &chan->state))
+			BUG_ON(is_ioat_bug(chanerr));
+		else /* we never got off the ground */
+			return;
+	}
+
+	/* if we haven't made progress and we have already
+	 * acknowledged a pending completion once, then be more
+	 * forceful with a restart
+	 */
+	spin_lock_bh(&chan->cleanup_lock);
+	if (ioat_cleanup_preamble(chan, &phys_complete))
+		__cleanup(ioat, phys_complete);
+	else if (test_bit(IOAT_COMPLETION_ACK, &chan->state)) {
+		spin_lock_bh(&ioat->prep_lock);
+		ioat2_restart_channel(ioat);
+		spin_unlock_bh(&ioat->prep_lock);
+		spin_unlock_bh(&chan->cleanup_lock);
+		return;
+	} else {
+		set_bit(IOAT_COMPLETION_ACK, &chan->state);
+		mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
+	}
+
+
+	if (ioat2_ring_active(ioat))
+		mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
+	else {
+		spin_lock_bh(&ioat->prep_lock);
+		check_active(ioat);
+		spin_unlock_bh(&ioat->prep_lock);
+	}
+	spin_unlock_bh(&chan->cleanup_lock);
 }
 
 static int ioat2_reset_hw(struct ioat_chan_common *chan)
@@ -404,7 +419,7 @@ static dma_cookie_t ioat2_tx_submit_unlock(struct dma_async_tx_descriptor *tx)
 	cookie = dma_cookie_assign(tx);
 	dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
 
-	if (!test_and_set_bit(IOAT_COMPLETION_PENDING, &chan->state))
+	if (!test_and_set_bit(IOAT_CHAN_ACTIVE, &chan->state))
 		mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 
 	/* make descriptor updates visible before advancing ioat->head,