[PATCH 0/2] xen/balloon: fix balloon driver initial target

[PATCH 0/2] xen/balloon: fix balloon driver initial target

[Roger Pau Monne]

Hello,

The fist patch is a bugfix to revert to the previous way of setting the
balloon memory target for PV domains.  The new way is off by ~97 pages
on domUs, because page at PFN 0 and the ISA range are ignored by Linux,
but populated from Xen's perspective.

Second patch aims to improve the initial memory target used by dom0.
With this new approach the target set by the balloon driver matches
exactly the target Xen by the toolstack when late-initializing it.

Thanks, Roger.

Roger Pau Monne (2):
  Partial revert "x86/xen: fix balloon target initialization for PVH
    dom0"
  xen/balloon: improve accuracy of initial balloon target for dom0

 arch/x86/xen/enlighten.c        |  2 +-
 drivers/xen/balloon.c           | 26 ++++++++++++++++++++++----
 drivers/xen/unpopulated-alloc.c |  3 +++
 include/xen/xen.h               |  2 ++
 4 files changed, 28 insertions(+), 5 deletions(-)

-- 
2.51.0

[PATCH 1/2] Partial revert "x86/xen: fix balloon target initialization for PVH dom0"

[Roger Pau Monne]

This partially reverts commit 87af633689ce16ddb166c80f32b120e50b1295de so
the current memory target for PV guests is still fetched from
start_info->nr_pages, which matches exactly what the toolstack sets the
initial memory target to.

Using get_num_physpages() is possible on PV also, but needs adjusting to
take into account the ISA hole and the PFN at 0 not considered usable
memory despite being populated, and hence would need extra adjustments.
Instead of carrying those extra adjustments switch back to the previous
code.  That leaves Linux with a difference in how current memory target is
obtained for HVM vs PV, but that's better than adding extra logic just for
PV.

However if switching to start_info->nr_pages for PV domains we need to
differentiate between released pages (freed back to the hypervisor) as
opposed to pages in the physmap which are not populated to start with.
Introduce a new xen_unpopulated_pages to account for papges that have
never been populated, and hence in the PV case don't need subtracting.

Fixes: 87af633689ce ("x86/xen: fix balloon target initialization for PVH dom0")
Reported-by: James Dingwall <james@dingwall.me.uk>
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
---
 arch/x86/xen/enlighten.c        |  2 +-
 drivers/xen/balloon.c           | 19 +++++++++++++++----
 drivers/xen/unpopulated-alloc.c |  3 +++
 include/xen/xen.h               |  2 ++
 4 files changed, 21 insertions(+), 5 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 53282dc7d5ac..23b91bf9b663 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -470,7 +470,7 @@ int __init arch_xen_unpopulated_init(struct resource **res)
 		 * driver to know how much of the physmap is unpopulated and
 		 * set an accurate initial memory target.
 		 */
-		xen_released_pages += xen_extra_mem[i].n_pfns;
+		xen_unpopulated_pages += xen_extra_mem[i].n_pfns;
 		/* Zero so region is not also added to the balloon driver. */
 		xen_extra_mem[i].n_pfns = 0;
 	}
diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
index 49c3f9926394..8c44a25a7d2b 100644
--- a/drivers/xen/balloon.c
+++ b/drivers/xen/balloon.c
@@ -724,6 +724,7 @@ static int __init balloon_add_regions(void)
 static int __init balloon_init(void)
 {
 	struct task_struct *task;
+	unsigned long current_pages;
 	int rc;
 
 	if (!xen_domain())
@@ -731,12 +732,18 @@ static int __init balloon_init(void)
 
 	pr_info("Initialising balloon driver\n");
 
-	if (xen_released_pages >= get_num_physpages()) {
-		WARN(1, "Released pages underflow current target");
-		return -ERANGE;
+	if (xen_pv_domain()) {
+		if (xen_released_pages >= xen_start_info->nr_pages)
+			goto underflow;
+		current_pages = min(xen_start_info->nr_pages -
+		                    xen_released_pages, max_pfn);
+	} else {
+		if (xen_unpopulated_pages >= get_num_physpages())
+			goto underflow;
+		current_pages = get_num_physpages() - xen_unpopulated_pages;
 	}
 
-	balloon_stats.current_pages = get_num_physpages() - xen_released_pages;
+	balloon_stats.current_pages = current_pages;
 	balloon_stats.target_pages  = balloon_stats.current_pages;
 	balloon_stats.balloon_low   = 0;
 	balloon_stats.balloon_high  = 0;
@@ -767,6 +774,10 @@ static int __init balloon_init(void)
 	xen_balloon_init();
 
 	return 0;
+
+ underflow:
+	WARN(1, "Released pages underflow current target");
+	return -ERANGE;
 }
 subsys_initcall(balloon_init);
 
diff --git a/drivers/xen/unpopulated-alloc.c b/drivers/xen/unpopulated-alloc.c
index d6fc2aefe264..1dc0b495c8e5 100644
--- a/drivers/xen/unpopulated-alloc.c
+++ b/drivers/xen/unpopulated-alloc.c
@@ -18,6 +18,9 @@ static unsigned int list_count;
 
 static struct resource *target_resource;
 
+/* Pages to subtract from the memory count when setting balloon target. */
+unsigned long xen_unpopulated_pages __initdata;
+
 /*
  * If arch is not happy with system "iomem_resource" being used for
  * the region allocation it can provide it's own view by creating specific
diff --git a/include/xen/xen.h b/include/xen/xen.h
index 61854e3f2837..f280c5dcf923 100644
--- a/include/xen/xen.h
+++ b/include/xen/xen.h
@@ -69,11 +69,13 @@ extern u64 xen_saved_max_mem_size;
 #endif
 
 #ifdef CONFIG_XEN_UNPOPULATED_ALLOC
+extern unsigned long xen_unpopulated_pages;
 int xen_alloc_unpopulated_pages(unsigned int nr_pages, struct page **pages);
 void xen_free_unpopulated_pages(unsigned int nr_pages, struct page **pages);
 #include <linux/ioport.h>
 int arch_xen_unpopulated_init(struct resource **res);
 #else
+#define xen_unpopulated_pages 0UL
 #include <xen/balloon.h>
 static inline int xen_alloc_unpopulated_pages(unsigned int nr_pages,
 		struct page **pages)
-- 
2.51.0

Re: [PATCH 1/2] Partial revert "x86/xen: fix balloon target initialization for PVH dom0"

[Jürgen Groß]

[-- Attachment #1.1.1: Type: text/plain, Size: 1388 bytes --]

On 28.01.26 12:05, Roger Pau Monne wrote:
> This partially reverts commit 87af633689ce16ddb166c80f32b120e50b1295de so
> the current memory target for PV guests is still fetched from
> start_info->nr_pages, which matches exactly what the toolstack sets the
> initial memory target to.
> 
> Using get_num_physpages() is possible on PV also, but needs adjusting to
> take into account the ISA hole and the PFN at 0 not considered usable
> memory despite being populated, and hence would need extra adjustments.
> Instead of carrying those extra adjustments switch back to the previous
> code.  That leaves Linux with a difference in how current memory target is
> obtained for HVM vs PV, but that's better than adding extra logic just for
> PV.
> 
> However if switching to start_info->nr_pages for PV domains we need to
> differentiate between released pages (freed back to the hypervisor) as
> opposed to pages in the physmap which are not populated to start with.
> Introduce a new xen_unpopulated_pages to account for papges that have
> never been populated, and hence in the PV case don't need subtracting.
> 
> Fixes: 87af633689ce ("x86/xen: fix balloon target initialization for PVH dom0")
> Reported-by: James Dingwall <james@dingwall.me.uk>
> Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>

Reviewed-by: Juergen Gross <jgross@suse.com>


Juergen

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3743 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]

[PATCH 2/2] xen/balloon: improve accuracy of initial balloon target for dom0

[Roger Pau Monne]

The dom0 balloon target set by the toolstack is the value returned by
XENMEM_current_reservation.  Do the same in the kernel balloon driver and
set the current allocation to the value returned by
XENMEM_current_reservation.  On my test system this causes the kernel
balloon driver target to exactly match the value set by the toolstack in
xenstore.

Note this approach can be used by both PV and PVH dom0s, as the toolstack
always uses XENMEM_current_reservation to set the initial target regardless
of the dom0 type.

Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
---
 drivers/xen/balloon.c | 27 +++++++++++++++++----------
 1 file changed, 17 insertions(+), 10 deletions(-)

diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
index 8c44a25a7d2b..9b6531eb28b6 100644
--- a/drivers/xen/balloon.c
+++ b/drivers/xen/balloon.c
@@ -724,7 +724,8 @@ static int __init balloon_add_regions(void)
 static int __init balloon_init(void)
 {
 	struct task_struct *task;
-	unsigned long current_pages;
+	long current_pages = 0;
+	domid_t domid = DOMID_SELF;
 	int rc;
 
 	if (!xen_domain())
@@ -732,15 +733,21 @@ static int __init balloon_init(void)
 
 	pr_info("Initialising balloon driver\n");
 
-	if (xen_pv_domain()) {
-		if (xen_released_pages >= xen_start_info->nr_pages)
-			goto underflow;
-		current_pages = min(xen_start_info->nr_pages -
-		                    xen_released_pages, max_pfn);
-	} else {
-		if (xen_unpopulated_pages >= get_num_physpages())
-			goto underflow;
-		current_pages = get_num_physpages() - xen_unpopulated_pages;
+	if (xen_initial_domain())
+		current_pages = HYPERVISOR_memory_op(XENMEM_current_reservation,
+		                                     &domid);
+	if (current_pages <= 0) {
+		if (xen_pv_domain()) {
+			if (xen_released_pages >= xen_start_info->nr_pages)
+				goto underflow;
+			current_pages = min(xen_start_info->nr_pages -
+			                    xen_released_pages, max_pfn);
+		} else {
+			if (xen_unpopulated_pages >= get_num_physpages())
+				goto underflow;
+			current_pages = get_num_physpages() -
+			                xen_unpopulated_pages;
+		}
 	}
 
 	balloon_stats.current_pages = current_pages;
-- 
2.51.0

Re: [PATCH 2/2] xen/balloon: improve accuracy of initial balloon target for dom0

[Jürgen Groß]

[-- Attachment #1.1.1: Type: text/plain, Size: 2542 bytes --]

On 28.01.26 12:05, Roger Pau Monne wrote:
> The dom0 balloon target set by the toolstack is the value returned by
> XENMEM_current_reservation.  Do the same in the kernel balloon driver and
> set the current allocation to the value returned by
> XENMEM_current_reservation.  On my test system this causes the kernel
> balloon driver target to exactly match the value set by the toolstack in
> xenstore.
> 
> Note this approach can be used by both PV and PVH dom0s, as the toolstack
> always uses XENMEM_current_reservation to set the initial target regardless
> of the dom0 type.
> 
> Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
> ---
>   drivers/xen/balloon.c | 27 +++++++++++++++++----------
>   1 file changed, 17 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
> index 8c44a25a7d2b..9b6531eb28b6 100644
> --- a/drivers/xen/balloon.c
> +++ b/drivers/xen/balloon.c
> @@ -724,7 +724,8 @@ static int __init balloon_add_regions(void)
>   static int __init balloon_init(void)
>   {
>   	struct task_struct *task;
> -	unsigned long current_pages;
> +	long current_pages = 0;
> +	domid_t domid = DOMID_SELF;
>   	int rc;
>   
>   	if (!xen_domain())
> @@ -732,15 +733,21 @@ static int __init balloon_init(void)
>   
>   	pr_info("Initialising balloon driver\n");
>   
> -	if (xen_pv_domain()) {
> -		if (xen_released_pages >= xen_start_info->nr_pages)
> -			goto underflow;
> -		current_pages = min(xen_start_info->nr_pages -
> -		                    xen_released_pages, max_pfn);
> -	} else {
> -		if (xen_unpopulated_pages >= get_num_physpages())
> -			goto underflow;
> -		current_pages = get_num_physpages() - xen_unpopulated_pages;
> +	if (xen_initial_domain())
> +		current_pages = HYPERVISOR_memory_op(XENMEM_current_reservation,
> +		                                     &domid);

Is there any specific reason why this should be limited to dom0?

I _think_ this should work for other domains, too.


Juergen

> +	if (current_pages <= 0) {
> +		if (xen_pv_domain()) {
> +			if (xen_released_pages >= xen_start_info->nr_pages)
> +				goto underflow;
> +			current_pages = min(xen_start_info->nr_pages -
> +			                    xen_released_pages, max_pfn);
> +		} else {
> +			if (xen_unpopulated_pages >= get_num_physpages())
> +				goto underflow;
> +			current_pages = get_num_physpages() -
> +			                xen_unpopulated_pages;
> +		}
>   	}
>   
>   	balloon_stats.current_pages = current_pages;


[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3743 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]

Re: [PATCH 2/2] xen/balloon: improve accuracy of initial balloon target for dom0

[Roger Pau Monné]

On Wed, Jan 28, 2026 at 12:31:13PM +0100, Jürgen Groß wrote:
> On 28.01.26 12:05, Roger Pau Monne wrote:
> > The dom0 balloon target set by the toolstack is the value returned by
> > XENMEM_current_reservation.  Do the same in the kernel balloon driver and
> > set the current allocation to the value returned by
> > XENMEM_current_reservation.  On my test system this causes the kernel
> > balloon driver target to exactly match the value set by the toolstack in
> > xenstore.
> > 
> > Note this approach can be used by both PV and PVH dom0s, as the toolstack
> > always uses XENMEM_current_reservation to set the initial target regardless
> > of the dom0 type.
> > 
> > Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
> > ---
> >   drivers/xen/balloon.c | 27 +++++++++++++++++----------
> >   1 file changed, 17 insertions(+), 10 deletions(-)
> > 
> > diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
> > index 8c44a25a7d2b..9b6531eb28b6 100644
> > --- a/drivers/xen/balloon.c
> > +++ b/drivers/xen/balloon.c
> > @@ -724,7 +724,8 @@ static int __init balloon_add_regions(void)
> >   static int __init balloon_init(void)
> >   {
> >   	struct task_struct *task;
> > -	unsigned long current_pages;
> > +	long current_pages = 0;
> > +	domid_t domid = DOMID_SELF;
> >   	int rc;
> >   	if (!xen_domain())
> > @@ -732,15 +733,21 @@ static int __init balloon_init(void)
> >   	pr_info("Initialising balloon driver\n");
> > -	if (xen_pv_domain()) {
> > -		if (xen_released_pages >= xen_start_info->nr_pages)
> > -			goto underflow;
> > -		current_pages = min(xen_start_info->nr_pages -
> > -		                    xen_released_pages, max_pfn);
> > -	} else {
> > -		if (xen_unpopulated_pages >= get_num_physpages())
> > -			goto underflow;
> > -		current_pages = get_num_physpages() - xen_unpopulated_pages;
> > +	if (xen_initial_domain())
> > +		current_pages = HYPERVISOR_memory_op(XENMEM_current_reservation,
> > +		                                     &domid);
> 
> Is there any specific reason why this should be limited to dom0?
> 
> I _think_ this should work for other domains, too.

Sadly it doesn't, I've already tested.  The value returned by
XENMEM_current_reservation on PV guests is slightly different than
what's in xen_start_info->nr_pages, which exactly what the toolstack
writes in xenstore.  I assume there's some other stuff that's
accounted for in d->tot_pages, but don't really know what I'm afraid.

And in the HVM/PVH case using XENMEM_current_reservation for domUs
would also take into account the Video memory, which skews the target.

This is the best I could do I'm afraid, at the expense of having so
many different way to fetch the information.

Thanks, Roger.

Re: [PATCH 2/2] xen/balloon: improve accuracy of initial balloon target for dom0

[Jürgen Groß]

[-- Attachment #1.1.1: Type: text/plain, Size: 3237 bytes --]

On 28.01.26 12:54, Roger Pau Monné wrote:
> On Wed, Jan 28, 2026 at 12:31:13PM +0100, Jürgen Groß wrote:
>> On 28.01.26 12:05, Roger Pau Monne wrote:
>>> The dom0 balloon target set by the toolstack is the value returned by
>>> XENMEM_current_reservation.  Do the same in the kernel balloon driver and
>>> set the current allocation to the value returned by
>>> XENMEM_current_reservation.  On my test system this causes the kernel
>>> balloon driver target to exactly match the value set by the toolstack in
>>> xenstore.
>>>
>>> Note this approach can be used by both PV and PVH dom0s, as the toolstack
>>> always uses XENMEM_current_reservation to set the initial target regardless
>>> of the dom0 type.
>>>
>>> Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
>>> ---
>>>    drivers/xen/balloon.c | 27 +++++++++++++++++----------
>>>    1 file changed, 17 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
>>> index 8c44a25a7d2b..9b6531eb28b6 100644
>>> --- a/drivers/xen/balloon.c
>>> +++ b/drivers/xen/balloon.c
>>> @@ -724,7 +724,8 @@ static int __init balloon_add_regions(void)
>>>    static int __init balloon_init(void)
>>>    {
>>>    	struct task_struct *task;
>>> -	unsigned long current_pages;
>>> +	long current_pages = 0;
>>> +	domid_t domid = DOMID_SELF;
>>>    	int rc;
>>>    	if (!xen_domain())
>>> @@ -732,15 +733,21 @@ static int __init balloon_init(void)
>>>    	pr_info("Initialising balloon driver\n");
>>> -	if (xen_pv_domain()) {
>>> -		if (xen_released_pages >= xen_start_info->nr_pages)
>>> -			goto underflow;
>>> -		current_pages = min(xen_start_info->nr_pages -
>>> -		                    xen_released_pages, max_pfn);
>>> -	} else {
>>> -		if (xen_unpopulated_pages >= get_num_physpages())
>>> -			goto underflow;
>>> -		current_pages = get_num_physpages() - xen_unpopulated_pages;
>>> +	if (xen_initial_domain())
>>> +		current_pages = HYPERVISOR_memory_op(XENMEM_current_reservation,
>>> +		                                     &domid);
>>
>> Is there any specific reason why this should be limited to dom0?
>>
>> I _think_ this should work for other domains, too.
> 
> Sadly it doesn't, I've already tested.  The value returned by
> XENMEM_current_reservation on PV guests is slightly different than
> what's in xen_start_info->nr_pages, which exactly what the toolstack
> writes in xenstore.  I assume there's some other stuff that's
> accounted for in d->tot_pages, but don't really know what I'm afraid.
> 
> And in the HVM/PVH case using XENMEM_current_reservation for domUs
> would also take into account the Video memory, which skews the target.
> 
> This is the best I could do I'm afraid, at the expense of having so
> many different way to fetch the information.

Meh, too bad.

For now I think we need to live with that. But for the future I'd like to
suggest to add a new Xenstore node memory/reservation-diff containing the
difference between XENMEM_current_reservation output and the tools' view
of the domain's memory size.

The balloon driver could read that node in balloon_init() and use
XENMEM_current_reservation in case the node is found.


Juergen

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3743 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]

Re: [PATCH 2/2] xen/balloon: improve accuracy of initial balloon target for dom0

[Jürgen Groß]

[-- Attachment #1.1.1: Type: text/plain, Size: 714 bytes --]

On 28.01.26 12:05, Roger Pau Monne wrote:
> The dom0 balloon target set by the toolstack is the value returned by
> XENMEM_current_reservation.  Do the same in the kernel balloon driver and
> set the current allocation to the value returned by
> XENMEM_current_reservation.  On my test system this causes the kernel
> balloon driver target to exactly match the value set by the toolstack in
> xenstore.
> 
> Note this approach can be used by both PV and PVH dom0s, as the toolstack
> always uses XENMEM_current_reservation to set the initial target regardless
> of the dom0 type.
> 
> Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>

Reviewed-by: Juergen Gross <jgross@suse.com>


Juergen

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3743 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]