While it is okay for IOMMU page tables to be set up for guests starting
in PoD mode, actual device assignment may only occur once all PoD
entries have been removed from the P2M. So far this was enforced only
for boot-time assignment, and only in the tool stack.

Also use the new function to replace p2m_pod_entry_count(): Its unlocked
access to p2m->pod.entry_count wasn't really okay (irrespective of the
result being stale by the time the caller gets to see it). Nor was the
use of that function in line with the immediately preceding comment: A
PoD guest isn't just one with a non-zero entry count, but also one with
a non-empty cache (e.g. prior to actually launching the guest).

To allow the tool stack to see a consistent snapshot of PoD state, move
the tail of XENMEM_{get,set}_pod_target handling into a function, adding
proper locking there.

In libxl take the liberty to use the new local variable r also for a
pre-existing call into libxc.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
---
If p2m->pod.entry_count == p2m->pod.count it is in principle possible to
permit device assignment by actively resolving all remaining PoD entries.

Initially I thought this was introduced by f89f555827a6 ("remove late
(on-demand) construction of IOMMU page tables"), but without
arch_iommu_use_permitted() checking for PoD I think the issue has been
there before that.
---
v4: Drop tool stack side change (superseded by 07449ecfa425). Extend VM
    event related paragraph of description.
v3: In p2m_pod_set_mem_target() move check down.
v2: New.

--- a/xen/arch/x86/mm/p2m-pod.c
+++ b/xen/arch/x86/mm/p2m-pod.c
@@ -20,6 +20,7 @@
  */
 
 #include <xen/event.h>
+#include <xen/iocap.h>
 #include <xen/ioreq.h>
 #include <xen/mm.h>
 #include <xen/sched.h>
@@ -360,7 +361,10 @@ p2m_pod_set_mem_target(struct domain *d,
 
     ASSERT( pod_target >= p2m->pod.count );
 
-    ret = p2m_pod_set_cache_target(p2m, pod_target, 1/*preemptible*/);
+    if ( has_arch_pdevs(d) || cache_flush_permitted(d) )
+        ret = -ENOTEMPTY;
+    else
+        ret = p2m_pod_set_cache_target(p2m, pod_target, 1/*preemptible*/);
 
 out:
     pod_unlock(p2m);
@@ -368,6 +372,23 @@ out:
     return ret;
 }
 
+void p2m_pod_get_mem_target(const struct domain *d, xen_pod_target_t *target)
+{
+    struct p2m_domain *p2m = p2m_get_hostp2m(d);
+
+    ASSERT(is_hvm_domain(d));
+
+    pod_lock(p2m);
+    lock_page_alloc(p2m);
+
+    target->tot_pages       = domain_tot_pages(d);
+    target->pod_cache_pages = p2m->pod.count;
+    target->pod_entries     = p2m->pod.entry_count;
+
+    unlock_page_alloc(p2m);
+    pod_unlock(p2m);
+}
+
 int p2m_pod_empty_cache(struct domain *d)
 {
     struct p2m_domain *p2m = p2m_get_hostp2m(d);
@@ -1391,6 +1412,9 @@ guest_physmap_mark_populate_on_demand(st
     if ( !paging_mode_translate(d) )
         return -EINVAL;
 
+    if ( has_arch_pdevs(d) || cache_flush_permitted(d) )
+        return -ENOTEMPTY;
+
     do {
         rc = mark_populate_on_demand(d, gfn, chunk_order);
 
@@ -1412,3 +1436,20 @@ void p2m_pod_init(struct p2m_domain *p2m
     for ( i = 0; i < ARRAY_SIZE(p2m->pod.mrp.list); ++i )
         p2m->pod.mrp.list[i] = gfn_x(INVALID_GFN);
 }
+
+bool p2m_pod_active(const struct domain *d)
+{
+    struct p2m_domain *p2m;
+    bool res;
+
+    if ( !is_hvm_domain(d) )
+        return false;
+
+    p2m = p2m_get_hostp2m(d);
+
+    pod_lock(p2m);
+    res = p2m->pod.entry_count | p2m->pod.count;
+    pod_unlock(p2m);
+
+    return res;
+}
--- a/xen/arch/x86/mm.c
+++ b/xen/arch/x86/mm.c
@@ -4804,7 +4804,6 @@ long arch_memory_op(unsigned long cmd, X
     {
         xen_pod_target_t target;
         struct domain *d;
-        struct p2m_domain *p2m;
 
         if ( copy_from_guest(&target, arg, 1) )
             return -EFAULT;
@@ -4835,10 +4834,7 @@ long arch_memory_op(unsigned long cmd, X
         }
         else if ( rc >= 0 )
         {
-            p2m = p2m_get_hostp2m(d);
-            target.tot_pages       = domain_tot_pages(d);
-            target.pod_cache_pages = p2m->pod.count;
-            target.pod_entries     = p2m->pod.entry_count;
+            p2m_pod_get_mem_target(d, &target);
 
             if ( __copy_to_guest(arg, &target, 1) )
                 rc = -EFAULT;
--- a/xen/common/vm_event.c
+++ b/xen/common/vm_event.c
@@ -639,7 +639,7 @@ int vm_event_domctl(struct domain *d, st
 
             rc = -EXDEV;
             /* Disallow paging in a PoD guest */
-            if ( p2m_pod_entry_count(p2m_get_hostp2m(d)) )
+            if ( p2m_pod_active(d) )
                 break;
 
             /* domain_pause() not required here, see XSA-99 */
--- a/xen/drivers/passthrough/x86/iommu.c
+++ b/xen/drivers/passthrough/x86/iommu.c
@@ -513,11 +513,12 @@ bool arch_iommu_use_permitted(const stru
 {
     /*
      * Prevent device assign if mem paging, mem sharing or log-dirty
-     * have been enabled for this domain.
+     * have been enabled for this domain, or if PoD is still in active use.
      */
     return d == dom_io ||
            (likely(!mem_sharing_enabled(d)) &&
             likely(!mem_paging_enabled(d)) &&
+            likely(!p2m_pod_active(d)) &&
             likely(!p2m_is_global_logdirty(d)));
 }
 
--- a/xen/arch/x86/include/asm/p2m.h
+++ b/xen/arch/x86/include/asm/p2m.h
@@ -701,6 +701,12 @@ int p2m_pod_empty_cache(struct domain *d
  * domain matches target */
 int p2m_pod_set_mem_target(struct domain *d, unsigned long target);
 
+/* Obtain a consistent snapshot of PoD related domain state. */
+void p2m_pod_get_mem_target(const struct domain *d, xen_pod_target_t *target);
+
+/* Check whether PoD is (still) active in a domain. */
+bool p2m_pod_active(const struct domain *d);
+
 /* Scan pod cache when offline/broken page triggered */
 int
 p2m_pod_offline_or_broken_hit(struct page_info *p);
@@ -709,11 +715,6 @@ p2m_pod_offline_or_broken_hit(struct pag
 void
 p2m_pod_offline_or_broken_replace(struct page_info *p);
 
-static inline long p2m_pod_entry_count(const struct p2m_domain *p2m)
-{
-    return p2m->pod.entry_count;
-}
-
 #else
 
 static inline bool
@@ -727,6 +728,11 @@ static inline int p2m_pod_empty_cache(st
     return 0;
 }
 
+static inline bool p2m_pod_active(const struct domain *d)
+{
+    return false;
+}
+
 static inline int p2m_pod_offline_or_broken_hit(struct page_info *p)
 {
     return 0;
@@ -737,11 +743,6 @@ static inline void p2m_pod_offline_or_br
     ASSERT_UNREACHABLE();
 }
 
-static inline long p2m_pod_entry_count(const struct p2m_domain *p2m)
-{
-    return 0;
-}
-
 #endif

On Mon, Apr 11, 2022 at 11:47:46AM +0200, Jan Beulich wrote:
> While it is okay for IOMMU page tables to be set up for guests starting
> in PoD mode, actual device assignment may only occur once all PoD
> entries have been removed from the P2M. So far this was enforced only
> for boot-time assignment, and only in the tool stack.
> 
> Also use the new function to replace p2m_pod_entry_count(): Its unlocked
> access to p2m->pod.entry_count wasn't really okay (irrespective of the
> result being stale by the time the caller gets to see it). Nor was the
> use of that function in line with the immediately preceding comment: A
> PoD guest isn't just one with a non-zero entry count, but also one with
> a non-empty cache (e.g. prior to actually launching the guest).
> 
> To allow the tool stack to see a consistent snapshot of PoD state, move
> the tail of XENMEM_{get,set}_pod_target handling into a function, adding
> proper locking there.
> 
> In libxl take the liberty to use the new local variable r also for a
> pre-existing call into libxc.
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>

Just one comment below.

> ---
> If p2m->pod.entry_count == p2m->pod.count it is in principle possible to
> permit device assignment by actively resolving all remaining PoD entries.
> 
> Initially I thought this was introduced by f89f555827a6 ("remove late
> (on-demand) construction of IOMMU page tables"), but without
> arch_iommu_use_permitted() checking for PoD I think the issue has been
> there before that.
> ---
> v4: Drop tool stack side change (superseded by 07449ecfa425). Extend VM
>     event related paragraph of description.
> v3: In p2m_pod_set_mem_target() move check down.
> v2: New.
> 
> --- a/xen/arch/x86/mm/p2m-pod.c
> +++ b/xen/arch/x86/mm/p2m-pod.c
> @@ -20,6 +20,7 @@
>   */
>  
>  #include <xen/event.h>
> +#include <xen/iocap.h>
>  #include <xen/ioreq.h>
>  #include <xen/mm.h>
>  #include <xen/sched.h>
> @@ -360,7 +361,10 @@ p2m_pod_set_mem_target(struct domain *d,
>  
>      ASSERT( pod_target >= p2m->pod.count );
>  
> -    ret = p2m_pod_set_cache_target(p2m, pod_target, 1/*preemptible*/);
> +    if ( has_arch_pdevs(d) || cache_flush_permitted(d) )
> +        ret = -ENOTEMPTY;

ENOTEMPTY seems weird here.  I think the reasoning is that the set of
passthrough devices is not empty? IMO it's confusing as the function
itself is related to buffer management, so returning ENOTEMPTY could
be confused with some other condition.

Might be less ambiguous to use EXDEV.

Thanks, Roger.

On 11.04.2022 12:29, Roger Pau Monné wrote:
> On Mon, Apr 11, 2022 at 11:47:46AM +0200, Jan Beulich wrote:
>> While it is okay for IOMMU page tables to be set up for guests starting
>> in PoD mode, actual device assignment may only occur once all PoD
>> entries have been removed from the P2M. So far this was enforced only
>> for boot-time assignment, and only in the tool stack.
>>
>> Also use the new function to replace p2m_pod_entry_count(): Its unlocked
>> access to p2m->pod.entry_count wasn't really okay (irrespective of the
>> result being stale by the time the caller gets to see it). Nor was the
>> use of that function in line with the immediately preceding comment: A
>> PoD guest isn't just one with a non-zero entry count, but also one with
>> a non-empty cache (e.g. prior to actually launching the guest).
>>
>> To allow the tool stack to see a consistent snapshot of PoD state, move
>> the tail of XENMEM_{get,set}_pod_target handling into a function, adding
>> proper locking there.
>>
>> In libxl take the liberty to use the new local variable r also for a
>> pre-existing call into libxc.
>>
>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>

Thanks.

>> --- a/xen/arch/x86/mm/p2m-pod.c
>> +++ b/xen/arch/x86/mm/p2m-pod.c
>> @@ -20,6 +20,7 @@
>>   */
>>  
>>  #include <xen/event.h>
>> +#include <xen/iocap.h>
>>  #include <xen/ioreq.h>
>>  #include <xen/mm.h>
>>  #include <xen/sched.h>
>> @@ -360,7 +361,10 @@ p2m_pod_set_mem_target(struct domain *d,
>>  
>>      ASSERT( pod_target >= p2m->pod.count );
>>  
>> -    ret = p2m_pod_set_cache_target(p2m, pod_target, 1/*preemptible*/);
>> +    if ( has_arch_pdevs(d) || cache_flush_permitted(d) )
>> +        ret = -ENOTEMPTY;
> 
> ENOTEMPTY seems weird here.  I think the reasoning is that the set of
> passthrough devices is not empty?

Yes.

> IMO it's confusing as the function
> itself is related to buffer management, so returning ENOTEMPTY could
> be confused with some other condition.
> 
> Might be less ambiguous to use EXDEV.

I don't think there's any particularly good error code to use here.
Hence I've tried to pick one that makes some sense, and which isn't
widely used (this latter aspect would be EXDEV slightly less desirable).

Jan