ARM: PL08x: move callback outside spinlock'd region
authorRussell King - ARM Linux <linux@arm.linux.org.uk>
Mon, 3 Jan 2011 22:41:34 +0000 (22:41 +0000)
committerDan Williams <dan.j.williams@intel.com>
Wed, 5 Jan 2011 03:16:13 +0000 (19:16 -0800)
Calling the callback handler with spinlocks in the tasklet held leads
to deadlock when dmaengine functions are called:

BUG: spinlock lockup on CPU#0, sh/417, c1870a08
Backtrace:
...
[<c017b408>] (do_raw_spin_lock+0x0/0x154) from [<c02c4b98>] (_raw_spin_lock_irqsave+0x54/0x60)
[<c02c4b44>] (_raw_spin_lock_irqsave+0x0/0x60) from [<c01f5828>] (pl08x_prep_channel_resources+0x718/0x8b4)
[<c01f5110>] (pl08x_prep_channel_resources+0x0/0x8b4) from [<c01f5bb4>] (pl08x_prep_slave_sg+0x120/0x19c)
[<c01f5a94>] (pl08x_prep_slave_sg+0x0/0x19c) from [<c01be7a0>] (pl011_dma_tx_refill+0x164/0x224)
[<c01be63c>] (pl011_dma_tx_refill+0x0/0x224) from [<c01bf1c8>] (pl011_dma_tx_callback+0x7c/0xc4)
[<c01bf14c>] (pl011_dma_tx_callback+0x0/0xc4) from [<c01f4d34>] (pl08x_tasklet+0x60/0x368)
[<c01f4cd4>] (pl08x_tasklet+0x0/0x368) from [<c004d978>] (tasklet_action+0xa0/0x100)

Dan quoted the documentation:
> 2/ Completion callback routines cannot submit new operations.  This
>    results in recursion in the synchronous case and spin_locks being
>    acquired twice in the asynchronous case.

but then followed up to say:
> I should clarify, this is the async_memcpy() api requirement which is
> not used outside of md/raid5.  DMA drivers can and do allow new
> submissions from callbacks, and the ones that do so properly move the
> callback outside of the driver lock.

So let's fix it by moving the callback out of the spinlocked region.

Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Acked-by: Linus Walleij <linus.walleij@stericsson.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/dma/amba-pl08x.c

index 986c12775b0b8c8127236f198a119f895764aa50..660165dd945a592c91bae2b132641d8623593b4a 100644 (file)
@@ -1540,32 +1540,29 @@ static void pl08x_tasklet(unsigned long data)
 {
        struct pl08x_dma_chan *plchan = (struct pl08x_dma_chan *) data;
        struct pl08x_driver_data *pl08x = plchan->host;
+       struct pl08x_txd *txd;
+       dma_async_tx_callback callback = NULL;
+       void *callback_param = NULL;
        unsigned long flags;
 
        spin_lock_irqsave(&plchan->lock, flags);
 
-       if (plchan->at) {
-               dma_async_tx_callback callback =
-                       plchan->at->tx.callback;
-               void *callback_param =
-                       plchan->at->tx.callback_param;
+       txd = plchan->at;
+       plchan->at = NULL;
 
-               /*
-                * Update last completed
-                */
-               plchan->lc = plchan->at->tx.cookie;
+       if (txd) {
+               callback = txd->tx.callback;
+               callback_param = txd->tx.callback_param;
 
                /*
-                * Callback to signal completion
+                * Update last completed
                 */
-               if (callback)
-                       callback(callback_param);
+               plchan->lc = txd->tx.cookie;
 
                /*
                 * Free the descriptor
                 */
-               pl08x_free_txd(pl08x, plchan->at);
-               plchan->at = NULL;
+               pl08x_free_txd(pl08x, txd);
        }
        /*
         * If a new descriptor is queued, set it up
@@ -1616,6 +1613,10 @@ static void pl08x_tasklet(unsigned long data)
        }
 
        spin_unlock_irqrestore(&plchan->lock, flags);
+
+       /* Callback to signal completion */
+       if (callback)
+               callback(callback_param);
 }
 
 static irqreturn_t pl08x_irq(int irq, void *dev)