RE: [PATCH v3 3/4] fsl-dma: change release process of dma descriptor for supporting async_tx

-----Original Message-----
From: Ira W. Snyder [mailto:iws@ovro.caltech.edu]
Sent: Tuesday, July 17, 2012 4:01 AM
To: Liu Qiang-B32616
Cc: linux-crypto@vger.kernel.org; linuxppc-dev@lists.ozlabs.org; Phillips
Kim-R1AAHA; herbert@gondor.hengli.com.au; davem@davemloft.net; Dan
Williams; Vinod Koul; Li Yang-R58472
Subject: Re: [PATCH v3 3/4] fsl-dma: change release process of dma
descriptor for supporting async_tx

On Mon, Jul 16, 2012 at 12:08:29PM +0800, Qiang Liu wrote:

quoted

Fix the potential risk when enable config NET_DMA and ASYNC_TX.
Async_tx is lack of support in current release process of dma
descriptor, all descriptors will be released whatever is acked or
no-acked by async_tx, so there is a potential race condition when dma
engine is uesd by others clients (e.g. when enable NET_DMA to offload

TCP).

quoted

In our case, a race condition which is raised when use both of talitos
and dmaengine to offload xor is because napi scheduler will sync all
pending requests in dma channels, it affects the process of raid
operations due to ack_tx is not checked in fsl dma. The no-acked
descriptor is freed which is submitted just now, as a dependent tx,
this freed descriptor trigger
BUG_ON(async_tx_test_ack(depend_tx)) in async_tx_submit().

Cc: Dan Williams <redacted>
Cc: Vinod Koul <redacted>
Cc: Li Yang <redacted>
Cc: Ira W. Snyder <redacted>
Signed-off-by: Qiang Liu <redacted>
---
 drivers/dma/fsldma.c |  378 +++++++++++++++++++++++++++++-------------

-------

quoted

 drivers/dma/fsldma.h |    1 +
 2 files changed, 225 insertions(+), 154 deletions(-)

diff --git a/drivers/dma/fsldma.c b/drivers/dma/fsldma.c index

4f2f212..4ee1b8f 100644

--- a/drivers/dma/fsldma.c
+++ b/drivers/dma/fsldma.c

@@ -400,6 +400,217 @@ out_splice:
 	list_splice_tail_init(&desc->tx_list, &chan->ld_pending);  }

+/**
+ * fsl_chan_xfer_ld_queue - transfer any pending transactions
+ * @chan : Freescale DMA channel
+ *
+ * HARDWARE STATE: idle
+ * LOCKING: must hold chan->desc_lock  */ static void
+fsl_chan_xfer_ld_queue(struct fsldma_chan *chan) {
+	struct fsl_desc_sw *desc;
+
+	/*
+	 * If the list of pending descriptors is empty, then we
+	 * don't need to do any work at all
+	 */
+	if (list_empty(&chan->ld_pending)) {
+		chan_dbg(chan, "no pending LDs\n");
+		return;
+	}
+
+	/*
+	 * The DMA controller is not idle, which means that the interrupt
+	 * handler will start any queued transactions when it runs after
+	 * this transaction finishes
+	 */
+	if (!chan->idle) {
+		chan_dbg(chan, "DMA controller still busy\n");
+		return;
+	}
+
+	/*
+	 * If there are some link descriptors which have not been
+	 * transferred, we need to start the controller
+	 */
+
+	/*
+	 * Move all elements from the queue of pending transactions
+	 * onto the list of running transactions
+	 */
+	chan_dbg(chan, "idle, starting controller\n");
+	desc = list_first_entry(&chan->ld_pending, struct fsl_desc_sw,

node);

quoted

+	list_splice_tail_init(&chan->ld_pending, &chan->ld_running);
+
+	/*
+	 * The 85xx DMA controller doesn't clear the channel start bit
+	 * automatically at the end of a transfer. Therefore we must clear
+	 * it in software before starting the transfer.
+	 */
+	if ((chan->feature & FSL_DMA_IP_MASK) == FSL_DMA_IP_85XX) {
+		u32 mode;
+
+		mode = DMA_IN(chan, &chan->regs->mr, 32);
+		mode &= ~FSL_DMA_MR_CS;
+		DMA_OUT(chan, &chan->regs->mr, mode, 32);
+	}
+
+	/*
+	 * Program the descriptor's address into the DMA controller,
+	 * then start the DMA transaction
+	 */
+	set_cdar(chan, desc->async_tx.phys);
+	get_cdar(chan);
+
+	dma_start(chan);
+	chan->idle = false;
+}
+

Please add a note about the locking requirements here, and for the other
new functions you've added.

You call this function from two places:

1) fsldma_cleanup_descriptor() - called with mod->desc_lock held
2) fsl_tx_status() - WITHOUT mod->desc_lock held

One of them is definitely wrong, and I'd bet that it is #2. You're
modifying ld_completed without a lock.

Yes, My bad, I will correct it.

quoted

+static int
+fsldma_clean_completed_descriptor(struct fsldma_chan *chan) {
+	struct fsl_desc_sw *desc, *_desc;
+
+	/* Run the callback for each descriptor, in order */
+	list_for_each_entry_safe(desc, _desc, &chan->ld_completed, node) {
+
+		if (async_tx_test_ack(&desc->async_tx)) {
+			/* Remove from the list of transactions */
+			list_del(&desc->node);
+#ifdef FSL_DMA_LD_DEBUG
+			chan_dbg(chan, "LD %p free\n", desc); #endif
+			dma_pool_free(chan->desc_pool, desc,
+					desc->async_tx.phys);
+		}
+	}
+
+	return 0;
+}
+
+/**
+ * fsldma_run_tx_complete_actions - cleanup and free a single link
+descriptor
+ * @chan: Freescale DMA channel
+ * @desc: descriptor to cleanup and free
+ * @cookie: Freescale DMA transaction identifier
+ *
+ * This function is used on a descriptor which has been executed by
+the DMA
+ * controller. It will run any callbacks, submit any dependencies,
+and then
+ * free the descriptor.
+ */
+static dma_cookie_t fsldma_run_tx_complete_actions(struct fsl_desc_sw

*desc,

quoted

+		struct fsldma_chan *chan, dma_cookie_t cookie) {
+	struct dma_async_tx_descriptor *txd = &desc->async_tx;
+	struct device *dev = chan->common.device->dev;
+	dma_addr_t src = get_desc_src(chan, desc);
+	dma_addr_t dst = get_desc_dst(chan, desc);
+	u32 len = get_desc_cnt(chan, desc);
+
+	BUG_ON(txd->cookie < 0);
+
+	if (txd->cookie > 0) {
+		cookie = txd->cookie;
+
+		/* Run the link descriptor callback function */
+		if (txd->callback) {
+#ifdef FSL_DMA_LD_DEBUG
+			chan_dbg(chan, "LD %p callback\n", desc); #endif
+			txd->callback(txd->callback_param);
+		}
+
+		/* Unmap the dst buffer, if requested */
+		if (!(txd->flags & DMA_COMPL_SKIP_DEST_UNMAP)) {
+			if (txd->flags & DMA_COMPL_DEST_UNMAP_SINGLE)
+				dma_unmap_single(dev, dst, len, DMA_FROM_DEVICE);
+			else
+				dma_unmap_page(dev, dst, len, DMA_FROM_DEVICE);
+		}
+
+		/* Unmap the src buffer, if requested */
+		if (!(txd->flags & DMA_COMPL_SKIP_SRC_UNMAP)) {
+			if (txd->flags & DMA_COMPL_SRC_UNMAP_SINGLE)
+				dma_unmap_single(dev, src, len, DMA_TO_DEVICE);
+			else
+				dma_unmap_page(dev, src, len, DMA_TO_DEVICE);
+		}
+	}
+
+	/* Run any dependencies */
+	dma_run_dependencies(txd);
+
+	return cookie;
+}
+
+static int
+fsldma_clean_running_descriptor(struct fsl_desc_sw *desc,
+			struct fsldma_chan *chan)
+{
+	/* Remove from the list of transactions */
+	list_del(&desc->node);
+	/* the client is allowed to attach dependent operations
+	 * until 'ack' is set
+	 */

This comment is does not follow the coding style. It should be:

/*
 * the client is allowed to attech dependent operations
 * until 'ack' is set
 */

Fine, I will correct it in v4. Include another 2 places related to coding style.
Thanks.

quoted

+	if (!async_tx_test_ack(&desc->async_tx)) {
+		/* move this slot to the completed */

Perhaps a better comment would be:

Move this descriptor to the list of descriptors which is complete, but
still awaiting the 'ack' bit to be set.

Accept.

quoted

+		list_add_tail(&desc->node, &chan->ld_completed);
+		return 0;
+	}
+
+	dma_pool_free(chan->desc_pool, desc,
+			desc->async_tx.phys);

This should all be on one line. It is less than 80 characters wide.

Accept.

quoted

+	return 0;
+}
+

Locking notes please.

I will add it in v4, please check. Thanks.

quoted

+static void fsldma_cleanup_descriptor(struct fsldma_chan *chan) {
+	struct fsl_desc_sw *desc, *_desc;
+	dma_cookie_t cookie = 0;
+	dma_addr_t curr_phys = get_cdar(chan);
+	int idle = dma_is_idle(chan);
+	int seen_current = 0;
+
+	fsldma_clean_completed_descriptor(chan);
+
+	/* Run the callback for each descriptor, in order */
+	list_for_each_entry_safe(desc, _desc, &chan->ld_running, node) {
+		/* do not advance past the current descriptor loaded into the
+		 * hardware channel, subsequent descriptors are either in
+		 * process or have not been submitted
+		 */

Coding style.

quoted

+		if (seen_current)
+			break;
+
+		/* stop the search if we reach the current descriptor and the
+		 * channel is busy
+		 */

Coding style.

quoted

+		if (desc->async_tx.phys == curr_phys) {
+			seen_current = 1;
+			if (!idle)
+				break;
+		}
+

This trick where you try to look at the hardware state to determine how
much to clean up has been a source of headaches in the past versions of
this driver.

I know a little about the history of this driver. Could you tell me what kind
headaches in the past versions, I can have a test and fix it in my patch. And
I also think use current phys addr should be more explicit.
Thanks.

It is much easier to reason about the state of the hardware and avoid
race conditions if you complete entire blocks of descriptors after the
hardware interrupts you to tell you it is finished.

In current driver, it's ok, but there is a potential issue when enable
async_tx api. How can you determine these descriptors in ld_running whether
have been completed in fsl_tx_status()? (I mean in my patch.)

This is the philosophy employed by the driver before your modifications:
ld_pending: a block of descriptors which is ready to run as soon as the
hardware becomes idle.
ld_running: a block of descriptors which is being executed by the
hardware.

These 2 lists are not enough, async_tx descriptors must be kept order as
expected of its submitter, we only can process the descriptor which has been
acked by async_tx api, but not free all descriptors in ld_running.
For example,
Step 1, in async_tx memcpy api,
tx2 = device->device_prep_dma_memcpy(...),
tx2->depend_tx = tx1;
step 2, in async_tx submit api,
async_tx_submit() will check tx2->depend_tx whether has been acked,
unfortunately, tx2->depend_tx is freed before step 2 (dma channels is flushed
by other drivers), the contents of tx2->depend_tx is set to POOL_POISON_FREED
if DMAPOOL_DEBUG is enabled (illegal pointer will be used if disable this config
option);
step 3, BUG_ON(async_tx_test_ack(depend_tx)) is triggered.

For resolve this potential risk, first, ack flag must be checked in dma driver,
second, ld_completed is added to restore all descriptors which have been acked
and completed.
In my following answer, most of all are around this idea.

quoted

+		cookie = fsldma_run_tx_complete_actions(desc, chan, cookie);
+
+		if (fsldma_clean_running_descriptor(desc, chan))
+			break;
+
+	}
+
+	/*
+	 * Start any pending transactions automatically
+	 *
+	 * In the ideal case, we keep the DMA controller busy while we go
+	 * ahead and free the descriptors below.
+	 */
+	fsl_chan_xfer_ld_queue(chan);
+
+	if (cookie > 0)
+		chan->common.completed_cookie = cookie; }
+
 static dma_cookie_t fsl_dma_tx_submit(struct dma_async_tx_descriptor
*tx)  {
 	struct fsldma_chan *chan = to_fsl_chan(tx->chan); @@ -534,8 +745,10

@@ static void fsl_dma_free_chan_resources(struct dma_chan *dchan)

 	chan_dbg(chan, "free all channel resources\n");
 	spin_lock_irqsave(&chan->desc_lock, flags);
+	fsldma_cleanup_descriptor(chan);
 	fsldma_free_desc_list(chan, &chan->ld_pending);
 	fsldma_free_desc_list(chan, &chan->ld_running);
+	fsldma_free_desc_list(chan, &chan->ld_completed);
 	spin_unlock_irqrestore(&chan->desc_lock, flags);

 	dma_pool_destroy(chan->desc_pool);

@@ -811,124 +1024,6 @@ static int fsl_dma_device_control(struct

dma_chan *dchan,  }

 /**
- * fsldma_cleanup_descriptor - cleanup and free a single link
descriptor
- * @chan: Freescale DMA channel
- * @desc: descriptor to cleanup and free
- *
- * This function is used on a descriptor which has been executed by
the DMA
- * controller. It will run any callbacks, submit any dependencies,
and then
- * free the descriptor.
- */
-static void fsldma_cleanup_descriptor(struct fsldma_chan *chan,
-				      struct fsl_desc_sw *desc)
-{
-	struct dma_async_tx_descriptor *txd = &desc->async_tx;
-	struct device *dev = chan->common.device->dev;
-	dma_addr_t src = get_desc_src(chan, desc);
-	dma_addr_t dst = get_desc_dst(chan, desc);
-	u32 len = get_desc_cnt(chan, desc);
-
-	/* Run the link descriptor callback function */
-	if (txd->callback) {
-#ifdef FSL_DMA_LD_DEBUG
-		chan_dbg(chan, "LD %p callback\n", desc);
-#endif
-		txd->callback(txd->callback_param);
-	}
-
-	/* Run any dependencies */
-	dma_run_dependencies(txd);
-
-	/* Unmap the dst buffer, if requested */
-	if (!(txd->flags & DMA_COMPL_SKIP_DEST_UNMAP)) {
-		if (txd->flags & DMA_COMPL_DEST_UNMAP_SINGLE)
-			dma_unmap_single(dev, dst, len, DMA_FROM_DEVICE);
-		else
-			dma_unmap_page(dev, dst, len, DMA_FROM_DEVICE);
-	}
-
-	/* Unmap the src buffer, if requested */
-	if (!(txd->flags & DMA_COMPL_SKIP_SRC_UNMAP)) {
-		if (txd->flags & DMA_COMPL_SRC_UNMAP_SINGLE)
-			dma_unmap_single(dev, src, len, DMA_TO_DEVICE);
-		else
-			dma_unmap_page(dev, src, len, DMA_TO_DEVICE);
-	}
-
-#ifdef FSL_DMA_LD_DEBUG
-	chan_dbg(chan, "LD %p free\n", desc);
-#endif
-	dma_pool_free(chan->desc_pool, desc, txd->phys);
-}
-
-/**
- * fsl_chan_xfer_ld_queue - transfer any pending transactions
- * @chan : Freescale DMA channel
- *
- * HARDWARE STATE: idle
- * LOCKING: must hold chan->desc_lock
- */
-static void fsl_chan_xfer_ld_queue(struct fsldma_chan *chan) -{
-	struct fsl_desc_sw *desc;
-
-	/*
-	 * If the list of pending descriptors is empty, then we
-	 * don't need to do any work at all
-	 */
-	if (list_empty(&chan->ld_pending)) {
-		chan_dbg(chan, "no pending LDs\n");
-		return;
-	}
-
-	/*
-	 * The DMA controller is not idle, which means that the interrupt
-	 * handler will start any queued transactions when it runs after
-	 * this transaction finishes
-	 */
-	if (!chan->idle) {
-		chan_dbg(chan, "DMA controller still busy\n");
-		return;
-	}
-
-	/*
-	 * If there are some link descriptors which have not been
-	 * transferred, we need to start the controller
-	 */
-
-	/*
-	 * Move all elements from the queue of pending transactions
-	 * onto the list of running transactions
-	 */
-	chan_dbg(chan, "idle, starting controller\n");
-	desc = list_first_entry(&chan->ld_pending, struct fsl_desc_sw,

node);

quoted

-	list_splice_tail_init(&chan->ld_pending, &chan->ld_running);
-
-	/*
-	 * The 85xx DMA controller doesn't clear the channel start bit
-	 * automatically at the end of a transfer. Therefore we must clear
-	 * it in software before starting the transfer.
-	 */
-	if ((chan->feature & FSL_DMA_IP_MASK) == FSL_DMA_IP_85XX) {
-		u32 mode;
-
-		mode = DMA_IN(chan, &chan->regs->mr, 32);
-		mode &= ~FSL_DMA_MR_CS;
-		DMA_OUT(chan, &chan->regs->mr, mode, 32);
-	}
-
-	/*
-	 * Program the descriptor's address into the DMA controller,
-	 * then start the DMA transaction
-	 */
-	set_cdar(chan, desc->async_tx.phys);
-	get_cdar(chan);
-
-	dma_start(chan);
-	chan->idle = false;
-}
-
-/**
  * fsl_dma_memcpy_issue_pending - Issue the DMA start command
  * @chan : Freescale DMA channel
  */

@@ -954,11 +1049,17 @@ static enum dma_status fsl_tx_status(struct

dma_chan *dchan,

quoted

 	enum dma_status ret;
 	unsigned long flags;

-	spin_lock_irqsave(&chan->desc_lock, flags);
 	ret = dma_cookie_status(dchan, cookie, txstate);
+	if (ret == DMA_SUCCESS) {
+		fsldma_clean_completed_descriptor(chan);
+		return ret;
+	}
+
+	spin_lock_irqsave(&chan->desc_lock, flags);
+	fsldma_cleanup_descriptor(chan);

You've made status from a very quick operation (compare two cookies) into
a potentially long running operation. It may now run callbacks, unmap
pages, etc.

Yes, that's right, callbacks and unmap pages will be invoked if DMA status
is not DMA_SUCCESS.

I note that Documentation/crypto/async-tx-api.txt section 3.6
"Constraints" does say that async_*() functions cannot be called from IRQ
context. However, the DMAEngine API itself does not have anything to say
about IRQ context.

I expect fsl_tx_status() to be quick, and not potentially call other
arbitrary code.

fsl_tx_status() is not in IRQ context. It's reasonable to unmap pages and
run dependency of descriptor as fsldma_run_tx_complete_actions() does.

Is it possible to leave this function unchanged?

According to my knowledge, it is unlikely to be left unchanged, or it will violate
the design of this interface. If DMA status is not DMA_SUCCESS, that means there
are new descriptors completed, we should move them to ld_completed, and do actions
to its async_tx descriptors, then dma descriptor will be freed at another time.
Most of my idea is from iop-adma.c, the original interface is not align with it.
Async_tx flags (expecially ack_test) is not considered when dma descriptor is freed,
so some random errors happened, I think you must familiar with this history.
As an important "synchronization flag", async_tx api must control the order of tx
descriptor. Dma driver also should make sure keeping order when free the descriptor.
That's the reason why I add ld_completed list.

Also note that if you leave this function unchanged, the locking error
pointed out above (fsldma_clean_completed_descriptor() is not called with
the lock held) goes away.

I know. Thanks.

quoted

 	spin_unlock_irqrestore(&chan->desc_lock, flags);

-	return ret;
+	return dma_cookie_status(dchan, cookie, txstate);
 }


/*--------------------------------------------------------------------
--------*/ @@ -1035,53 +1136,21 @@ static irqreturn_t
fsldma_chan_irq(int irq, void *data)  static void
dma_do_tasklet(unsigned long data)  {
 	struct fsldma_chan *chan = (struct fsldma_chan *)data;
-	struct fsl_desc_sw *desc, *_desc;
-	LIST_HEAD(ld_cleanup);
 	unsigned long flags;

 	chan_dbg(chan, "tasklet entry\n");

 	spin_lock_irqsave(&chan->desc_lock, flags);

-	/* update the cookie if we have some descriptors to cleanup */
-	if (!list_empty(&chan->ld_running)) {
-		dma_cookie_t cookie;
-
-		desc = to_fsl_desc(chan->ld_running.prev);
-		cookie = desc->async_tx.cookie;
-		dma_cookie_complete(&desc->async_tx);
-
-		chan_dbg(chan, "completed_cookie=%d\n", cookie);
-	}
-
-	/*
-	 * move the descriptors to a temporary list so we can drop the lock
-	 * during the entire cleanup operation
-	 */
-	list_splice_tail_init(&chan->ld_running, &ld_cleanup);
+	/* Run all cleanup for this descriptor */
+	fsldma_cleanup_descriptor(chan);

 	/* the hardware is now idle and ready for more */
 	chan->idle = true;

-	/*
-	 * Start any pending transactions automatically
-	 *
-	 * In the ideal case, we keep the DMA controller busy while we go
-	 * ahead and free the descriptors below.
-	 */
 	fsl_chan_xfer_ld_queue(chan);
 	spin_unlock_irqrestore(&chan->desc_lock, flags);

-	/* Run the callback for each descriptor, in order */
-	list_for_each_entry_safe(desc, _desc, &ld_cleanup, node) {
-
-		/* Remove from the list of transactions */
-		list_del(&desc->node);
-
-		/* Run all cleanup for this descriptor */
-		fsldma_cleanup_descriptor(chan, desc);
-	}
-
 	chan_dbg(chan, "tasklet exit\n");
 }

@@ -1262,6 +1331,7 @@ static int __devinit fsl_dma_chan_probe(struct

fsldma_device *fdev,

quoted

 	spin_lock_init(&chan->desc_lock);
 	INIT_LIST_HEAD(&chan->ld_pending);
 	INIT_LIST_HEAD(&chan->ld_running);
+	INIT_LIST_HEAD(&chan->ld_completed);
 	chan->idle = true;

 	chan->common.device = &fdev->common; diff --git
a/drivers/dma/fsldma.h b/drivers/dma/fsldma.h index f5c3879..7ede908
100644

--- a/drivers/dma/fsldma.h
+++ b/drivers/dma/fsldma.h

@@ -140,6 +140,7 @@ struct fsldma_chan {
 	spinlock_t desc_lock;		/* Descriptor operation lock */
 	struct list_head ld_pending;	/* Link descriptors queue */
 	struct list_head ld_running;	/* Link descriptors queue */
+	struct list_head ld_completed;	/* Link descriptors queue */
 	struct dma_chan common;		/* DMA common channel */
 	struct dma_pool *desc_pool;	/* Descriptors pool */
 	struct device *dev;		/* Channel device */
--

1.7.5.1