diff mbox series

[RFCv2,10/15] xen/arm: mm: Allocate xen page tables in domheap rather than xenheap

Message ID 20210425201318.15447-11-julien@xen.org (mailing list archive)
State New, archived
Headers show
Series xen/arm: mm: Remove open-coding mappings | expand

Commit Message

Julien Grall April 25, 2021, 8:13 p.m. UTC
From: Julien Grall <jgrall@amazon.com>

xen_{un,}map_table() already uses the helper to map/unmap pages
on-demand (note this is currently a NOP on arm64). So switching to
domheap don't have any disavantage.

But this as the benefit:
    - to keep the page tables unmapped if an arch decided to do so
    - reduce xenheap use on arm32 which can be pretty small

Signed-off-by: Julien Grall <jgrall@amazon.com>

---
    Changes in v2:
        - New patch
---
 xen/arch/arm/mm.c | 36 +++++++++++++++++++++---------------
 1 file changed, 21 insertions(+), 15 deletions(-)

Comments

Stefano Stabellini May 12, 2021, 10:44 p.m. UTC | #1
On Sun, 25 Apr 2021, Julien Grall wrote:
> From: Julien Grall <jgrall@amazon.com>
> 
> xen_{un,}map_table() already uses the helper to map/unmap pages
> on-demand (note this is currently a NOP on arm64). So switching to
> domheap don't have any disavantage.
> 
> But this as the benefit:
>     - to keep the page tables unmapped if an arch decided to do so
>     - reduce xenheap use on arm32 which can be pretty small
> 
> Signed-off-by: Julien Grall <jgrall@amazon.com>

Thanks for the patch. It looks OK but let me ask a couple of questions
to clarify my doubts.

This change should have no impact to arm64, right?

For arm32, I wonder why we were using map_domain_page before in
xen_map_table: it wasn't necessary, was it? In fact, one could even say
that it was wrong?


> ---
>     Changes in v2:
>         - New patch
> ---
>  xen/arch/arm/mm.c | 36 +++++++++++++++++++++---------------
>  1 file changed, 21 insertions(+), 15 deletions(-)
> 
> diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c
> index 19ecf73542ce..ae5a07ea956b 100644
> --- a/xen/arch/arm/mm.c
> +++ b/xen/arch/arm/mm.c
> @@ -969,21 +969,6 @@ void *ioremap(paddr_t pa, size_t len)
>      return ioremap_attr(pa, len, PAGE_HYPERVISOR_NOCACHE);
>  }
>  
> -static int create_xen_table(lpae_t *entry)
> -{
> -    void *p;
> -    lpae_t pte;
> -
> -    p = alloc_xenheap_page();
> -    if ( p == NULL )
> -        return -ENOMEM;
> -    clear_page(p);
> -    pte = mfn_to_xen_entry(virt_to_mfn(p), MT_NORMAL);
> -    pte.pt.table = 1;
> -    write_pte(entry, pte);
> -    return 0;
> -}
> -
>  static lpae_t *xen_map_table(mfn_t mfn)
>  {
>      /*
> @@ -1024,6 +1009,27 @@ static void xen_unmap_table(const lpae_t *table)
>      unmap_domain_page(table);
>  }
>  
> +static int create_xen_table(lpae_t *entry)
> +{
> +    struct page_info *pg;
> +    void *p;
> +    lpae_t pte;
> +
> +    pg = alloc_domheap_page(NULL, 0);
> +    if ( pg == NULL )
> +        return -ENOMEM;
> +
> +    p = xen_map_table(page_to_mfn(pg));
> +    clear_page(p);
> +    xen_unmap_table(p);
> +
> +    pte = mfn_to_xen_entry(page_to_mfn(pg), MT_NORMAL);
> +    pte.pt.table = 1;
> +    write_pte(entry, pte);
> +
> +    return 0;
> +}
> +
>  #define XEN_TABLE_MAP_FAILED 0
>  #define XEN_TABLE_SUPER_PAGE 1
>  #define XEN_TABLE_NORMAL_PAGE 2
> -- 
> 2.17.1
>
Julien Grall May 13, 2021, 6:09 p.m. UTC | #2
Hi Stefano,

On 12/05/2021 23:44, Stefano Stabellini wrote:
> On Sun, 25 Apr 2021, Julien Grall wrote:
>> From: Julien Grall <jgrall@amazon.com>
>>
>> xen_{un,}map_table() already uses the helper to map/unmap pages
>> on-demand (note this is currently a NOP on arm64). So switching to
>> domheap don't have any disavantage.
>>
>> But this as the benefit:
>>      - to keep the page tables unmapped if an arch decided to do so
>>      - reduce xenheap use on arm32 which can be pretty small
>>
>> Signed-off-by: Julien Grall <jgrall@amazon.com>
> 
> Thanks for the patch. It looks OK but let me ask a couple of questions
> to clarify my doubts.
> 
> This change should have no impact to arm64, right?
> 
> For arm32, I wonder why we were using map_domain_page before in
> xen_map_table: it wasn't necessary, was it? In fact, one could even say
> that it was wrong?
In xen_map_table() we need to be able to map pages from Xen binary, 
xenheap... On arm64, we would be able to use mfn_to_virt() because 
everything is mapped in Xen. But that's not the case on arm32. So we 
need a way to map anything easily.

The only difference between xenheap and domheap are the former is always 
mapped while the latter may not be. So one can also view a xenheap page 
as a glorified domheap.

I also don't really want to create yet another interface to map pages 
(we have vmap(), map_domain_page(), map_domain_global_page()...). So, 
when I implemented xen_map_table() a couple of years ago, I came to the 
conclusion that this is a convenient (ab)use of the interface.

Cheers,
Stefano Stabellini May 13, 2021, 10:27 p.m. UTC | #3
On Thu, 13 May 2021, Julien Grall wrote:
> Hi Stefano,
> 
> On 12/05/2021 23:44, Stefano Stabellini wrote:
> > On Sun, 25 Apr 2021, Julien Grall wrote:
> > > From: Julien Grall <jgrall@amazon.com>
> > > 
> > > xen_{un,}map_table() already uses the helper to map/unmap pages
> > > on-demand (note this is currently a NOP on arm64). So switching to
> > > domheap don't have any disavantage.
> > > 
> > > But this as the benefit:
> > >      - to keep the page tables unmapped if an arch decided to do so
> > >      - reduce xenheap use on arm32 which can be pretty small
> > > 
> > > Signed-off-by: Julien Grall <jgrall@amazon.com>
> > 
> > Thanks for the patch. It looks OK but let me ask a couple of questions
> > to clarify my doubts.
> > 
> > This change should have no impact to arm64, right?
> > 
> > For arm32, I wonder why we were using map_domain_page before in
> > xen_map_table: it wasn't necessary, was it? In fact, one could even say
> > that it was wrong?
> In xen_map_table() we need to be able to map pages from Xen binary, xenheap...
> On arm64, we would be able to use mfn_to_virt() because everything is mapped
> in Xen. But that's not the case on arm32. So we need a way to map anything
> easily.
> 
> The only difference between xenheap and domheap are the former is always
> mapped while the latter may not be. So one can also view a xenheap page as a
> glorified domheap.
> 
> I also don't really want to create yet another interface to map pages (we have
> vmap(), map_domain_page(), map_domain_global_page()...). So, when I
> implemented xen_map_table() a couple of years ago, I came to the conclusion
> that this is a convenient (ab)use of the interface.

Got it. Repeating to check if I see the full picture. On ARM64 there are
no changes. On ARM32, at runtime there are no changes mapping/unmapping
pages because xen_map_table is already mapping all pages using domheap,
even xenheap pages are mapped as domheap; so this patch makes no
difference in mapping/unmapping, correct?

The only difference is that on arm32 we are using domheap to allocate
the pages, which is a different (larger) pool.
Julien Grall May 15, 2021, 8:48 a.m. UTC | #4
Hi Stefano,

On 13/05/2021 23:27, Stefano Stabellini wrote:
> On Thu, 13 May 2021, Julien Grall wrote:
>> Hi Stefano,
>>
>> On 12/05/2021 23:44, Stefano Stabellini wrote:
>>> On Sun, 25 Apr 2021, Julien Grall wrote:
>>>> From: Julien Grall <jgrall@amazon.com>
>>>>
>>>> xen_{un,}map_table() already uses the helper to map/unmap pages
>>>> on-demand (note this is currently a NOP on arm64). So switching to
>>>> domheap don't have any disavantage.
>>>>
>>>> But this as the benefit:
>>>>       - to keep the page tables unmapped if an arch decided to do so
>>>>       - reduce xenheap use on arm32 which can be pretty small
>>>>
>>>> Signed-off-by: Julien Grall <jgrall@amazon.com>
>>>
>>> Thanks for the patch. It looks OK but let me ask a couple of questions
>>> to clarify my doubts.
>>>
>>> This change should have no impact to arm64, right?
>>>
>>> For arm32, I wonder why we were using map_domain_page before in
>>> xen_map_table: it wasn't necessary, was it? In fact, one could even say
>>> that it was wrong?
>> In xen_map_table() we need to be able to map pages from Xen binary, xenheap...
>> On arm64, we would be able to use mfn_to_virt() because everything is mapped
>> in Xen. But that's not the case on arm32. So we need a way to map anything
>> easily.
>>
>> The only difference between xenheap and domheap are the former is always
>> mapped while the latter may not be. So one can also view a xenheap page as a
>> glorified domheap.
>>
>> I also don't really want to create yet another interface to map pages (we have
>> vmap(), map_domain_page(), map_domain_global_page()...). So, when I
>> implemented xen_map_table() a couple of years ago, I came to the conclusion
>> that this is a convenient (ab)use of the interface.
> 
> Got it. Repeating to check if I see the full picture. On ARM64 there are
> no changes. On ARM32, at runtime there are no changes mapping/unmapping
> pages because xen_map_table is already mapping all pages using domheap,
> even xenheap pages are mapped as domheap; so this patch makes no
> difference in mapping/unmapping, correct?

For arm32, it makes a slight difference when allocating a new page table 
(we didn't call map/unmap before) but this is not called often.

The main "drop" in performance happened when xen_{,map}_table() was 
introduced.

> 
> The only difference is that on arm32 we are using domheap to allocate
> the pages, which is a different (larger) pool.

Yes.

Would you be happy to give you acked-by/reviewed-by on this basis?

Cheers,
Stefano Stabellini May 18, 2021, 12:37 a.m. UTC | #5
On Sat, 15 May 2021, Julien Grall wrote:
> Hi Stefano,
> 
> On 13/05/2021 23:27, Stefano Stabellini wrote:
> > On Thu, 13 May 2021, Julien Grall wrote:
> > > Hi Stefano,
> > > 
> > > On 12/05/2021 23:44, Stefano Stabellini wrote:
> > > > On Sun, 25 Apr 2021, Julien Grall wrote:
> > > > > From: Julien Grall <jgrall@amazon.com>
> > > > > 
> > > > > xen_{un,}map_table() already uses the helper to map/unmap pages
> > > > > on-demand (note this is currently a NOP on arm64). So switching to
> > > > > domheap don't have any disavantage.
> > > > > 
> > > > > But this as the benefit:
> > > > >       - to keep the page tables unmapped if an arch decided to do so
> > > > >       - reduce xenheap use on arm32 which can be pretty small
> > > > > 
> > > > > Signed-off-by: Julien Grall <jgrall@amazon.com>
> > > > 
> > > > Thanks for the patch. It looks OK but let me ask a couple of questions
> > > > to clarify my doubts.
> > > > 
> > > > This change should have no impact to arm64, right?
> > > > 
> > > > For arm32, I wonder why we were using map_domain_page before in
> > > > xen_map_table: it wasn't necessary, was it? In fact, one could even say
> > > > that it was wrong?
> > > In xen_map_table() we need to be able to map pages from Xen binary,
> > > xenheap...
> > > On arm64, we would be able to use mfn_to_virt() because everything is
> > > mapped
> > > in Xen. But that's not the case on arm32. So we need a way to map anything
> > > easily.
> > > 
> > > The only difference between xenheap and domheap are the former is always
> > > mapped while the latter may not be. So one can also view a xenheap page as
> > > a
> > > glorified domheap.
> > > 
> > > I also don't really want to create yet another interface to map pages (we
> > > have
> > > vmap(), map_domain_page(), map_domain_global_page()...). So, when I
> > > implemented xen_map_table() a couple of years ago, I came to the
> > > conclusion
> > > that this is a convenient (ab)use of the interface.
> > 
> > Got it. Repeating to check if I see the full picture. On ARM64 there are
> > no changes. On ARM32, at runtime there are no changes mapping/unmapping
> > pages because xen_map_table is already mapping all pages using domheap,
> > even xenheap pages are mapped as domheap; so this patch makes no
> > difference in mapping/unmapping, correct?
> 
> For arm32, it makes a slight difference when allocating a new page table (we
> didn't call map/unmap before) but this is not called often.
> 
> The main "drop" in performance happened when xen_{,map}_table() was
> introduced.
> 
> > 
> > The only difference is that on arm32 we are using domheap to allocate
> > the pages, which is a different (larger) pool.
> 
> Yes.
> 
> Would you be happy to give you acked-by/reviewed-by on this basis?

Yes

Acked-by: Stefano Stabellini <sstabellini@kernel.org>
diff mbox series

Patch

diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c
index 19ecf73542ce..ae5a07ea956b 100644
--- a/xen/arch/arm/mm.c
+++ b/xen/arch/arm/mm.c
@@ -969,21 +969,6 @@  void *ioremap(paddr_t pa, size_t len)
     return ioremap_attr(pa, len, PAGE_HYPERVISOR_NOCACHE);
 }
 
-static int create_xen_table(lpae_t *entry)
-{
-    void *p;
-    lpae_t pte;
-
-    p = alloc_xenheap_page();
-    if ( p == NULL )
-        return -ENOMEM;
-    clear_page(p);
-    pte = mfn_to_xen_entry(virt_to_mfn(p), MT_NORMAL);
-    pte.pt.table = 1;
-    write_pte(entry, pte);
-    return 0;
-}
-
 static lpae_t *xen_map_table(mfn_t mfn)
 {
     /*
@@ -1024,6 +1009,27 @@  static void xen_unmap_table(const lpae_t *table)
     unmap_domain_page(table);
 }
 
+static int create_xen_table(lpae_t *entry)
+{
+    struct page_info *pg;
+    void *p;
+    lpae_t pte;
+
+    pg = alloc_domheap_page(NULL, 0);
+    if ( pg == NULL )
+        return -ENOMEM;
+
+    p = xen_map_table(page_to_mfn(pg));
+    clear_page(p);
+    xen_unmap_table(p);
+
+    pte = mfn_to_xen_entry(page_to_mfn(pg), MT_NORMAL);
+    pte.pt.table = 1;
+    write_pte(entry, pte);
+
+    return 0;
+}
+
 #define XEN_TABLE_MAP_FAILED 0
 #define XEN_TABLE_SUPER_PAGE 1
 #define XEN_TABLE_NORMAL_PAGE 2