From: Eric Dumazet <[email protected]>
Some devices (like mlx4) try hard to allocate memory on selected
NUMA node, but it turns out intel_alloc_coherent() is not NUMA
aware yet.
Note that dma_generic_alloc_coherent() in arch/x86/kernel/pci-dma.c
gets this right.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Benjamin Serebrin <[email protected]>
Cc: David Woodhouse <[email protected]>
Cc: Joerg Roedel <[email protected]>
---
drivers/iommu/intel-iommu.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index a1373cf343269455808f66ad18dc0a2fb7aa73f2..0efef077abc099eb29ebc5cefdd1b996f025dffd 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -3734,8 +3734,11 @@ static void *intel_alloc_coherent(struct device *dev, size_t size,
}
}
- if (!page)
- page = alloc_pages(flags, order);
+ if (!page) {
+ page = alloc_pages_node(dev_to_node(dev), flags, order);
+ if (!page)
+ page = alloc_pages(flags, order);
+ }
if (!page)
return NULL;
memset(page_address(page), 0, size);
On Wed, 2018-01-31 at 14:45 -0800, Eric Dumazet wrote:
> From: Eric Dumazet <[email protected]>
>
> Some devices (like mlx4) try hard to allocate memory on selected
> NUMA node, but it turns out intel_alloc_coherent() is not NUMA
> aware yet.
>
> Note that dma_generic_alloc_coherent() in arch/x86/kernel/pci-dma.c
> gets this right.
>
> Signed-off-by: Eric Dumazet <[email protected]>
> Cc: Benjamin Serebrin <[email protected]>
> Cc: David Woodhouse <[email protected]>
> Cc: Joerg Roedel <[email protected]>
> ---
> drivers/iommu/intel-iommu.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
> index a1373cf343269455808f66ad18dc0a2fb7aa73f2..0efef077abc099eb29ebc5cefdd1b996f025dffd 100644
> --- a/drivers/iommu/intel-iommu.c
> +++ b/drivers/iommu/intel-iommu.c
> @@ -3734,8 +3734,11 @@ static void *intel_alloc_coherent(struct device *dev, size_t size,
> }
> }
>
> - if (!page)
> - page = alloc_pages(flags, order);
> + if (!page) {
> + page = alloc_pages_node(dev_to_node(dev), flags, order);
> + if (!page)
> + page = alloc_pages(flags, order);
I'll send a V2 without the fallback to alloc_pages()
This seems not necessary at all.
> + }
> if (!page)
> return NULL;
> memset(page_address(page), 0, size);
From: Eric Dumazet <[email protected]>
Some devices (like mlx4) try hard to allocate memory on selected
NUMA node, but it turns out intel_alloc_coherent() is not NUMA
aware yet.
Note that dma_generic_alloc_coherent() in arch/x86/kernel/pci-dma.c
gets this right.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Benjamin Serebrin <[email protected]>
Cc: David Woodhouse <[email protected]>
Cc: Joerg Roedel <[email protected]>
---
v2: no fallback to alloc_pages(), this is not needed and might even
hurt in OOM cases.
drivers/iommu/intel-iommu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index a1373cf343269455808f66ad18dc0a2fb7aa73f2..3c538466a98bdb8fffdca688462b1350d536791b 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -3735,7 +3735,7 @@ static void *intel_alloc_coherent(struct device *dev, size_t size,
}
if (!page)
- page = alloc_pages(flags, order);
+ page = alloc_pages_node(dev_to_node(dev), flags, order);
if (!page)
return NULL;
memset(page_address(page), 0, size);
I've got patches pending to replace all that code with
dma_direct_alloc, which will do the right thing. They were
submitted for 4.16, and I will resend them after -rc1.
On Fri, Feb 2, 2018 at 10:53 AM, Christoph Hellwig <[email protected]> wrote:
> I've got patches pending to replace all that code with
> dma_direct_alloc, which will do the right thing. They were
> submitted for 4.16, and I will resend them after -rc1.
I see, thanks Christoph !
On 2/2/18 10:59 AM, Eric Dumazet wrote:
> On Fri, Feb 2, 2018 at 10:53 AM, Christoph Hellwig <[email protected]> wrote:
>> I've got patches pending to replace all that code with
>> dma_direct_alloc, which will do the right thing. They were
>> submitted for 4.16, and I will resend them after -rc1.
>
> I see, thanks Christoph !
>
Hi Christoph
It seems 4.16 has shipped ( :) ) , and intel_alloc_coherent() still has no NUMA awareness.
Should I respin https://lore.kernel.org/patchwork/patch/884326/
Thanks !
On Wed, Apr 01, 2020 at 03:53:38PM -0700, Eric Dumazet wrote:
>
>
> On 2/2/18 10:59 AM, Eric Dumazet wrote:
> > On Fri, Feb 2, 2018 at 10:53 AM, Christoph Hellwig <[email protected]> wrote:
> >> I've got patches pending to replace all that code with
> >> dma_direct_alloc, which will do the right thing. They were
> >> submitted for 4.16, and I will resend them after -rc1.
> >
> > I see, thanks Christoph !
> >
>
> Hi Christoph
>
> It seems 4.16 has shipped ( :) ) , and intel_alloc_coherent() still has no NUMA awareness.
Actually, that code went in and then got reverted again..
> Should I respin https://lore.kernel.org/patchwork/patch/884326/
Maybe. We are still hoping to convert intel-iommu to the dma-iommu
framework, but I'm not sure how long that is going to take, so maybe
just respin it for now.