2008-02-20 02:37:27

by Yinghai Lu

Subject: [PATCH] mm: fix dma_poor_create

dev_to_node could return node that without RAM. So check it before use it in kmalloc_node

Signed-off-by: Yinghai Lu <[email protected]>

diff --git a/mm/dmapool.c b/mm/dmapool.c
index 34aaac4..6be5b5a 100644
--- a/mm/dmapool.c
+++ b/mm/dmapool.c
@@ -128,6 +128,7 @@ struct dma_pool *dma_pool_create(const char *name, struct device *dev,
{
struct dma_pool *retval;
size_t allocation;
+ int node;

if (align == 0) {
align = 1;
@@ -152,7 +153,11 @@ struct dma_pool *dma_pool_create(const char *name, struct device *dev,
return NULL;
}

- retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
+ node = dev_to_node(dev);
+ if (node == -1 || !node_online(node))
+ node = numa_node_id();
+
+ retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, node);
if (!retval)
return retval;

2008-02-20 06:52:54

by Ingo Molnar

Subject: Re: [PATCH] mm: fix dma_poor_create

* Yinghai Lu <[email protected]> wrote:

> dev_to_node could return node that without RAM. So check it before use
> it in kmalloc_node

> - retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
> + node = dev_to_node(dev);
> + if (node == -1 || !node_online(node))
> + node = numa_node_id();
> +
> + retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, node);

so this is about not crashing during bootup on nodes that have CPUs but
which have no node-specific memory attached, right?

Shouldnt kmalloc_node() be made more robust instead? I.e. push the same
code into kmalloc_node() - and make sure it will allocate _something_?
That would probably also fix a similar bug in net/core/skbuff.c's
__netdev_alloc_skb(), which too passes a dev_to_node() result to an
allocator.

Ingo

2008-02-20 07:52:25

by Yinghai Lu

Subject: Re: [PATCH] mm: fix dma_poor_create

On Tuesday 19 February 2008 10:52:30 pm Ingo Molnar wrote:
>
> * Yinghai Lu <[email protected]> wrote:
>
> > dev_to_node could return node that without RAM. So check it before use
> > it in kmalloc_node
>
> > - retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
> > + node = dev_to_node(dev);
> > + if (node == -1 || !node_online(node))
> > + node = numa_node_id();
> > +
> > + retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, node);
>
> so this is about not crashing during bootup on nodes that have CPUs but
> which have no node-specific memory attached, right?
>
> Shouldnt kmalloc_node() be made more robust instead? I.e. push the same
> code into kmalloc_node() - and make sure it will allocate _something_?
> That would probably also fix a similar bug in net/core/skbuff.c's
> __netdev_alloc_skb(), which too passes a dev_to_node() result to an
> allocator.

sound good idea to update the dev_to_node to make sure it will return -1 or the one is online.

Will send updated one.

YH