On 5 Aug 2020, at 20:43, John Paul Adrian Glaubitz <glaubitz@xxxxxxxxxxxxxxxxxxx> wrote:
>
> Hi Christoph!
>
> On 8/5/19 9:10 AM, Christoph Hellwig wrote:
>> Seems like we dropped the ball on this..
>>
>> Did I give you a patch like this (for 5.2 and probably earlier, won't
>> apply to 5.3-rc) to test before as that is anther idea?
>>
>> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
>> index 2c2772e9702a..e471158c7c6e 100644
>> --- a/kernel/dma/direct.c
>> +++ b/kernel/dma/direct.c
>> @@ -119,7 +119,8 @@ struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
>> }
>> }
>> if (!page)
>> - page = alloc_pages_node(dev_to_node(dev), gfp, page_order);
>> + page = alloc_pages_node(local_memory_node(dev_to_node(dev)),
>> + gfp, page_order);
>>
>> if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
>> __free_pages(page, page_order);
>
> I just applied this patch on top of 4.19.137 and it's crashing with this trace
> when trying to load the "hpsa" module. It definitely looks like an issue with
> dma_direct_alloc():
My guess is dev_to_node gave NUMA_NO_NODE for a random PCI device. Try:
int nid = dev_to_node(dev);
if (nid >= 0)
nid = local_memory_node(nid);
and then pass nid to alloc_pages_node instead.
Jess
|