Thread (18 messages) flat view 18 messages, 4 authors, 2008-06-23

Re: [PATCH] net: via-velocity.c fix sleep-with-spinlock bug during MTU change

From: Andrew Morton <akpm@linux-foundation.org>
Date: 2008-06-03 20:48:05
Also in: lkml

On Sat, 31 May 2008 18:46:15 -0700
Arjan van de Ven [off-list ref] wrote:
quoted hunk ↗ jump to hunk
From: Arjan van de Ven <redacted>
Subject: [PATCH] net: via-velocity.c fix sleep-with-spinlock bug during MTU change

The via-velocity.c driver reinitializes (frees/allocates) several
metadata structures during an MTU change. Unfortunately the allocations
of the new versions of the metadata is done with GFP_KERNEL, even
though this change of datastructures is (and needs to be) done while
holding a spinlock (with irqs off).

Clearly that isn't a good thing, and kerneloops.org has trapped a large
deal of the resulting warnings. The fix is to use GFP_ATOMIC.
While not elegant, avoiding the lock is going to be extremely complex.
In addition, this is a "static", long lived allocation (after all, how
often do you actually change your mtu) and not something that happens
on an ongoing basis.

...
diff --git a/drivers/net/via-velocity.c b/drivers/net/via-velocity.c
index 6b8d882..4bf08fd 100644
--- a/drivers/net/via-velocity.c
+++ b/drivers/net/via-velocity.c
@@ -1,4 +1,4 @@
-/*
+;/*
Cat sat on your keyboard?
quoted hunk ↗ jump to hunk
  * This code is derived from the VIA reference driver (copyright message
  * below) provided to Red Hat by VIA Networking Technologies, Inc. for
  * addition to the Linux kernel.
@@ -1241,6 +1241,9 @@ static int velocity_rx_refill(struct velocity_info *vptr)
  *
  *	Allocate and set up the receive buffers for each ring slot and
  *	assign them to the network adapter.
+ *
+ *	Note: This function gets called with irqs off/lock held
+ *	from velocity_change_mtu()
  */
 
 static int velocity_init_rd_ring(struct velocity_info *vptr)
@@ -1251,7 +1254,7 @@ static int velocity_init_rd_ring(struct velocity_info *vptr)
 	vptr->rx_buf_sz = (mtu <= ETH_DATA_LEN) ? PKT_BUF_SZ : mtu + 32;
 
 	vptr->rd_info = kcalloc(vptr->options.numrx,
-				sizeof(struct velocity_rd_info), GFP_KERNEL);
+				sizeof(struct velocity_rd_info), GFP_ATOMIC);
What happens if this allocation fails?  I think the driver is dead?

We've gone and freed the rd_ring and the td_ring and we _might_ have
allocated a new rd_ring and not a new td_ring.

And we've set vptr->rx_buf_sz, which may or may not be a problem.

And we've gone and left the interface in a downed state.

So hrm.  It could all be a lot better.  Just looking quickly at the
code I _think_ we might be able to do all the needed allocations
outside the lock and then swizzle them into place after taking the
lock. ie, something as simple as:

	struct velocity_info *temp_vptr;

	...

	velocity_init_rd_ring(temp_vptr);	/* Can use GFP_KERNEL! */

	spin_lock_irqsave(&vptr->lock, flags);

	velocity_free_td_ring(vptr);
	velocity_free_rd_ring(vptr);

	vptr->foo = temp_vptr->foo;
	vptr->bar = temp_vptr->bar;
	...

	spin_unlock_irqrestore(&vptr->lock, flags);

?
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help