[RFC PATCH] ARM 32: Workaround for memory problem >1gb on last section

[RFC PATCH] ARM 32: Workaround for memory problem >1gb on last section

[Frediano Ziglio]

setup_xenheap_mappings setup head memory on Arm 32 has a limit of 1GB.
On system with large memory is possible that there are no blocks of memory
smaller than 1gb leading xenheap_pages to be more than 1gb.
This cause memory errors trying to access heap after the 1gb limit.

I actually consider this patch as RFC as the pages are allocated at the end
of the block found however I don't think is safe to assume that the end is
aligned to 32mb as required.

Signed-off-by: Frediano Ziglio <frediano.ziglio@huawei.com>
---
 xen/arch/arm/setup.c | 5 +++++
 1 file changed, 5 insertions(+)


diff --git a/xen/arch/arm/setup.c b/xen/arch/arm/setup.c
index 446de8a..34b55b4 100644
--- a/xen/arch/arm/setup.c
+++ b/xen/arch/arm/setup.c
@@ -524,6 +524,11 @@ static void __init setup_mm(unsigned long
dtb_paddr, size_t dtb_size)
     if ( ! e )
         panic("Not not enough space for xenheap");

+#ifdef CONFIG_ARM_32
+    if ( xenheap_pages > 1<<(30-PAGE_SHIFT) )
+        xenheap_pages = 1<<(30-PAGE_SHIFT);
+#endif
+
     domheap_pages = heap_pages - xenheap_pages;

     printk("Xen heap: %"PRIpaddr"-%"PRIpaddr" (%lu pages)\n",
--
1.9.1

Re: [RFC PATCH] ARM 32: Workaround for memory problem >1gb on last section

[Ian Campbell]

On Mon, 2014-09-29 at 17:02 +0100, Frediano Ziglio wrote:
> setup_xenheap_mappings setup head memory on Arm 32 has a limit of 1GB.
> On system with large memory is possible that there are no blocks of memory
> smaller than 1gb leading xenheap_pages to be more than 1gb.
> This cause memory errors trying to access heap after the 1gb limit.
> 
> I actually consider this patch as RFC as the pages are allocated at the end
> of the block found however I don't think is safe to assume that the end is
> aligned to 32mb as required.
> 
> Signed-off-by: Frediano Ziglio <frediano.ziglio@huawei.com>

Acked-by: Ian Campbell <ian.campbell@citrix.com>

Re: [RFC PATCH] ARM 32: Workaround for memory problem >1gb on last section

[Ian Campbell]

On Wed, 2014-10-01 at 11:37 +0100, Ian Campbell wrote:
> On Mon, 2014-09-29 at 17:02 +0100, Frediano Ziglio wrote:
> > setup_xenheap_mappings setup head memory on Arm 32 has a limit of 1GB.
> > On system with large memory is possible that there are no blocks of memory
> > smaller than 1gb leading xenheap_pages to be more than 1gb.
> > This cause memory errors trying to access heap after the 1gb limit.
> > 
> > I actually consider this patch as RFC as the pages are allocated at the end
> > of the block found however I don't think is safe to assume that the end is
> > aligned to 32mb as required.
> > 
> > Signed-off-by: Frediano Ziglio <frediano.ziglio@huawei.com>
> 
> Acked-by: Ian Campbell <ian.campbell@citrix.com>

WRT the RFC and comment I think this patch is in and of itself correct.
I think the comment about no blocks of memory smaller than 1gb is a bit
misleading though, we will just look for a big enough region, so it's
actually the presence of large enough blocks not the absence of small
ones which is the problem.

The issue which you raise regarding the 32MB alignment is handled by the
call to consider_modules which takes 32MB as the required alignment
parameter, AFAIK it should return something which is suitably aligned,
since it aligns the end address on entry and eventually returns it. I'm
not 100% sure that it doesn't rely on size being 32MB aligned, but I my
99% sure it doesn't ;-).

Ian

Re: [RFC PATCH] ARM 32: Workaround for memory problem >1gb on last section

[Julien Grall]

Hello Frediano,

On 29/09/2014 17:02, Frediano Ziglio wrote:
> setup_xenheap_mappings setup head memory on Arm 32 has a limit of 1GB.
> On system with large memory is possible that there are no blocks of memory

it's

> smaller than 1gb leading xenheap_pages to be more than 1gb.

Can you be consistent with the way to write 1GB?

> This cause memory errors trying to access heap after the 1gb limit.
>
> I actually consider this patch as RFC as the pages are allocated at the end
> of the block found however I don't think is safe to assume that the end is
> aligned to 32mb as required.

After Ian's comment, I guess this paragraph should be dropped.

> Signed-off-by: Frediano Ziglio <frediano.ziglio@huawei.com>
> ---
>   xen/arch/arm/setup.c | 5 +++++
>   1 file changed, 5 insertions(+)
>
>
> diff --git a/xen/arch/arm/setup.c b/xen/arch/arm/setup.c
> index 446de8a..34b55b4 100644
> --- a/xen/arch/arm/setup.c
> +++ b/xen/arch/arm/setup.c
> @@ -524,6 +524,11 @@ static void __init setup_mm(unsigned long
> dtb_paddr, size_t dtb_size)
>       if ( ! e )
>           panic("Not not enough space for xenheap");
>
> +#ifdef CONFIG_ARM_32

This version of setup_mm is only compiled for arm32, therefore the 
#ifdef is not necessary.

Regards,


-- 
Julien Grall

[RFC PATCH] ARM 32: Workaround for memory problem >1gb on last section

[Frediano Ziglio]

setup_xenheap_mappings setup head memory on Arm 32 has a limit of 1GB.
On system with large memory is possible that there are no blocks of memory
smaller than 1gb leading xenheap_pages to be more than 1gb.
This cause memory errors trying to access heap after the 1gb limit.

I actually consider this patch as RFC as the pages are allocated at the end
of the block found however I don't think is safe to assume that the end is
aligned to 32mb as required.

Signed-off-by: Frediano Ziglio <frediano.ziglio@huawei.com>
---
 xen/arch/arm/setup.c | 5 +++++
 1 file changed, 5 insertions(+)


diff --git a/xen/arch/arm/setup.c b/xen/arch/arm/setup.c
index 446de8a..34b55b4 100644
--- a/xen/arch/arm/setup.c
+++ b/xen/arch/arm/setup.c
@@ -524,6 +524,11 @@ static void __init setup_mm(unsigned long
dtb_paddr, size_t dtb_size)
     if ( ! e )
         panic("Not not enough space for xenheap");

+#ifdef CONFIG_ARM_32
+    if ( xenheap_pages > 1<<(30-PAGE_SHIFT) )
+        xenheap_pages = 1<<(30-PAGE_SHIFT);
+#endif
+
     domheap_pages = heap_pages - xenheap_pages;

     printk("Xen heap: %"PRIpaddr"-%"PRIpaddr" (%lu pages)\n",
--
1.9.1