Skip to contentSkip to conversation

x86/EPT: avoid undue forcing of MMIO accesses to UC

PATCH In review Jan Beulich 3 participants 17 messages Jun 12, 2024 – Jun 13, 2024

lore
Share & tools
lore
#0 lore raw

v2 cover x86/EPT: avoid undue forcing of MMIO accesses to UC

..., getting in the way of, in particular, PVH Dom0 accessing its video frame buffer (if it has a console).

While especially the 1st one may not appear to be so, both of the earlier patches are strictly prereqs to the last one.

1: correct special page checking in epte_get_entry_emt() 2: avoid marking non-present entries for re-configuring 3: drop questionable mfn_valid() from epte_get_entry_emt()

Jan

#1 lore raw

PATCH v2 1/3 x86/EPT: correct special page checking in epte_get_entry_emt()

mfn_valid() granularity is (currently) 256Mb. Therefore the start of a 1Gb page passing the test doesn't necessarily mean all parts of such a range would also pass. Yet using the result of mfn_to_page() on an MFN which doesn't pass mfn_valid() checking is liable to result in a crash (the invocation of mfn_to_page() alone is presumably "just" UB in such a case).

Fixes ca24b2ffdbd9 ("x86/hvm: set 'ipat' in EPT for special pages")

---
Of course we could leverage mfn_valid() granularity here to do an increment by more than 1 if mfn_valid() returned false. Yet doing so likely would want a suitable helper to be introduced first, rather than open-coding such logic here.
---
v2: New.

xen/arch/x86/mm/p2m-ept.c +5 -1
@@ -519,8 +519,12 @@ int epte_get_entry_emt(struct domain *d,
     }
 
     for ( special_pgs = i = 0; i < (1ul << order); i++ )
-        if ( is_special_page(mfn_to_page(mfn_add(mfn, i))) )
+    {
+        mfn_t cur = mfn_add(mfn, i);
+
+        if ( mfn_valid(cur) && is_special_page(mfn_to_page(cur)) )
             special_pgs++;
+    }
 
     if ( special_pgs )
     {
Reviewed-by on v2 1/3
#2 lore raw

Re: [PATCH v2 for-4.19 1/3] x86/EPT: correct special page checking in epte_get_entry_emt()

Quoted 3 lines · Jan Beulich

mfn_valid() granularity is (currently) 256Mb. Therefore the start of a 1Gb page passing the test doesn't necessarily mean all parts of such a range would also pass.

How would such a superpage end up in the EPT?

I would assume this can only happen when adding a superpage MMIO that has part of it return success from mfn_valid()?

Quoted 7 lines

Yet using the result of mfn_to_page() on an MFN
which doesn't pass mfn_valid() checking is liable to result in a crash (the invocation of mfn_to_page() alone is presumably "just" UB in such a case).

Fixes ca24b2ffdbd9 ("x86/hvm: set 'ipat' in EPT for special pages")
Quoted 5 lines

---
Of course we could leverage mfn_valid() granularity here to do an increment by more than 1 if mfn_valid() returned false. Yet doing so likely would want a suitable helper to be introduced first, rather than open-coding such logic here.

We would still need to call is_special_page() on each 4K chunk, at which point taking advantage of the mfn_valid() granularity is likely to make the code more complicated to follow IMO.

Thanks, Roger.

#3 lore raw

Re: [PATCH v2 for-4.19 1/3] x86/EPT: correct special page checking in epte_get_entry_emt()

Quoted 9 lines · Roger Pau Monné
Quoted 3 lines · Jan Beulich

mfn_valid() granularity is (currently) 256Mb. Therefore the start of a 1Gb page passing the test doesn't necessarily mean all parts of such a range would also pass.

How would such a superpage end up in the EPT?

I would assume this can only happen when adding a superpage MMIO that has part of it return success from mfn_valid()?

Yes, that's the only way I can think of.

Quoted 9 lines
Quoted 7 lines

Yet using the result of mfn_to_page() on an MFN
which doesn't pass mfn_valid() checking is liable to result in a crash (the invocation of mfn_to_page() alone is presumably "just" UB in such a case).

Fixes ca24b2ffdbd9 ("x86/hvm: set 'ipat' in EPT for special pages")

Thanks.

Quoted 7 lines
Quoted 5 lines

---
Of course we could leverage mfn_valid() granularity here to do an increment by more than 1 if mfn_valid() returned false. Yet doing so likely would want a suitable helper to be introduced first, rather than open-coding such logic here.

We would still need to call is_special_page() on each 4K chunk,

Why? Within any block for which mfn_valid() returns false, there can be no RAM pages and hence also no special ones. It's only blocks where mfn_valid() returns true that we'd need to iterate through page-by-page.

Quoted 3 lines

at
which point taking advantage of the mfn_valid() granularity is likely to make the code more complicated to follow IMO.

Right, this making it more complicated is the main counter argument. Hence why I think that if to go such a route at all, it would need some new helper(s) such that at the use sites things still would remain reasonably clear.

Jan

#4 lore raw

Re: [PATCH v2 for-4.19 1/3] x86/EPT: correct special page checking in epte_get_entry_emt()

Quoted 31 lines · Jan Beulich
Quoted 9 lines · Roger Pau Monné
Quoted 3 lines · Jan Beulich

mfn_valid() granularity is (currently) 256Mb. Therefore the start of a 1Gb page passing the test doesn't necessarily mean all parts of such a range would also pass.

How would such a superpage end up in the EPT?

I would assume this can only happen when adding a superpage MMIO that has part of it return success from mfn_valid()?

Yes, that's the only way I can think of.

Quoted 9 lines
Quoted 7 lines

Yet using the result of mfn_to_page() on an MFN
which doesn't pass mfn_valid() checking is liable to result in a crash (the invocation of mfn_to_page() alone is presumably "just" UB in such a case).

Fixes ca24b2ffdbd9 ("x86/hvm: set 'ipat' in EPT for special pages")

Thanks.

Quoted 6 lines
Quoted 5 lines

---
Of course we could leverage mfn_valid() granularity here to do an increment by more than 1 if mfn_valid() returned false. Yet doing so likely would want a suitable helper to be introduced first, rather than open-coding such logic here.

Quoted 5 lines
Quoted 1 line

We would still need to call is_special_page() on each 4K chunk,

Why? Within any block for which mfn_valid() returns false, there can be no RAM pages and hence also no special ones. It's only blocks where mfn_valid() returns true that we'd need to iterate through page-by-page.

Oh right, I was thinking the other way around (mfn_valid() returning true), never mind.

Quoted 8 lines
Quoted 3 lines

at
which point taking advantage of the mfn_valid() granularity is likely to make the code more complicated to follow IMO.

Right, this making it more complicated is the main counter argument. Hence why I think that if to go such a route at all, it would need some new helper(s) such that at the use sites things still would remain reasonably clear.

We could also add an extra check to exit the loop early if special pages have been found but don't match the current loop index, as it's all special pages or none.

Thanks, Roger.

#5 lore raw

Re: [PATCH v2 for-4.19 1/3] x86/EPT: correct special page checking in epte_get_entry_emt()

Quoted 3 lines · Roger Pau Monné

We could also add an extra check to exit the loop early if special pages have been found but don't match the current loop index, as it's all special pages or none.

I was actually considering to make such a change, but then concluded that in the common case there'll be no special pages anyway, and hence we need to run the loop to completion anyway.

Jan

#6 lore raw

Re: [PATCH v2 for-4.19 1/3] x86/EPT: correct special page checking in epte_get_entry_emt()

In reply to Jan Beulich
Quoted 12 lines · Jan Beulich

mfn_valid() granularity is (currently) 256Mb. Therefore the start of a
1Gb page passing the test doesn't necessarily mean all parts of such a
range would also pass. Yet using the result of mfn_to_page() on an MFN
which doesn't pass mfn_valid() checking is liable to result in a crash
(the invocation of mfn_to_page() alone is presumably "just" UB in such a
case).

Quoted 2 lines
Fixes ca24b2ffdbd9 ("x86/hvm: set 'ipat' in EPT for special pages")

~ Oleksii

Quoted 26 lines

---
Of course we could leverage mfn_valid() granularity here to do an increment by more than 1 if mfn_valid() returned false. Yet doing so likely would want a suitable helper to be introduced first, rather than
open-coding such logic here.
---
v2: New.

xen/arch/x86/mm/p2m-ept.c +0 -0
@@ -519,8 +519,12 @@ int epte_get_entry_emt(struct domain *d,
      }


     for ( special_pgs = i = 0; i < (1ul << order); i++ ) -        if ( is_special_page(mfn_to_page(mfn_add(mfn, i))) ) +    {
+        mfn_t cur = mfn_add(mfn, i);
+
+        if ( mfn_valid(cur) && is_special_page(mfn_to_page(cur)) ) special_pgs++;
+    }

     if ( special_pgs )
     {

#7 lore raw

PATCH v2 2/3 x86/EPT: avoid marking non-present entries for re-configuring

In reply to Jan Beulich

For non-present entries EMT, like most other fields, is meaningless to hardware. Make the logic in ept_set_entry() setting the field (and iPAT) conditional upon dealing with a present entry, leaving the value at 0 otherwise. This has two effects for epte_get_entry_emt() which we'll want to leverage subsequently:
1) The call moved here now won't be issued with INVALID_MFN anymore (a respective BUG_ON() is being added).
2) Neither of the other two calls could now be issued with a truncated form of INVALID_MFN anymore (as long as there's no bug anywhere marking an entry present when that was populated using INVALID_MFN).

---
v2: New.

xen/arch/x86/mm/p2m-ept.c +18 -11
@@ -650,6 +650,8 @@ static int cf_check resolve_misconfig(st
             if ( e.emt != MTRR_NUM_TYPES )
                 break;
 
+            ASSERT(is_epte_present(&e));
+
             if ( level == 0 )
             {
                 for ( gfn -= i, i = 0; i < EPT_PAGETABLE_ENTRIES; ++i )
@@ -915,17 +917,6 @@ ept_set_entry(struct p2m_domain *p2m, gf
 
     if ( mfn_valid(mfn) || p2m_allows_invalid_mfn(p2mt) )
     {
-        bool ipat;
-        int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
-                                     i * EPT_TABLE_ORDER, &ipat,
-                                     p2mt);
-
-        if ( emt >= 0 )
-            new_entry.emt = emt;
-        else /* ept_handle_misconfig() will need to take care of this. */
-            new_entry.emt = MTRR_NUM_TYPES;
-
-        new_entry.ipat = ipat;
         new_entry.sp = !!i;
         new_entry.sa_p2mt = p2mt;
         new_entry.access = p2ma;
@@ -941,6 +932,22 @@ ept_set_entry(struct p2m_domain *p2m, gf
             need_modify_vtd_table = 0;
 
         ept_p2m_type_to_flags(p2m, &new_entry);
+
+        if ( is_epte_present(&new_entry) )
+        {
+            bool ipat;
+            int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
+                                         i * EPT_TABLE_ORDER, &ipat,
+                                         p2mt);
+
+            BUG_ON(mfn_eq(mfn, INVALID_MFN));
+
+            if ( emt >= 0 )
+                new_entry.emt = emt;
+            else /* ept_handle_misconfig() will need to take care of this. */
+                new_entry.emt = MTRR_NUM_TYPES;
+            new_entry.ipat = ipat;
+        }
     }
 
     if ( sve != -1 )
#8 lore raw

Re: [PATCH v2 for-4.19 2/3] x86/EPT: avoid marking non-present entries for re-configuring

Quoted 18 lines · Jan Beulich

For non-present entries EMT, like most other fields, is meaningless to hardware. Make the logic in ept_set_entry() setting the field (and iPAT) conditional upon dealing with a present entry, leaving the value at 0 otherwise. This has two effects for epte_get_entry_emt() which we'll want to leverage subsequently:
1) The call moved here now won't be issued with INVALID_MFN anymore (a respective BUG_ON() is being added).
2) Neither of the other two calls could now be issued with a truncated form of INVALID_MFN anymore (as long as there's no bug anywhere marking an entry present when that was populated using INVALID_MFN).

---
v2: New.

xen/arch/x86/mm/p2m-ept.c static int cf_check resolve_misconfig(st · line 653

Quoted 4 lines
xen/arch/x86/mm/p2m-ept.c +1 -0
@@ -650,6 +650,8 @@ static int cf_check resolve_misconfig(st
             if ( e.emt != MTRR_NUM_TYPES )
                 break;
 
+            ASSERT(is_epte_present(&e));

If this is added here, then there's a condition further below:

if ( !is_epte_valid(&e) || !is_epte_present(&e) )

That needs adjusting AFAICT.

However, in ept_set_entry() we seem to unconditionally call resolve_misconfig() against the new entry to be populated, won't this possibly cause resolve_misconfig() to be called against non-present EPT entries? I think this is fine because such non-present entries will have emt == 0, and hence will take the break just ahead of the added ASSERT().

Quoted 32 lines
xen/arch/x86/mm/p2m-ept.c +7 -11
+
             if ( level == 0 )
             {
                 for ( gfn -= i, i = 0; i < EPT_PAGETABLE_ENTRIES; ++i )
@@ -915,17 +917,6 @@ ept_set_entry(struct p2m_domain *p2m, gf
 
     if ( mfn_valid(mfn) || p2m_allows_invalid_mfn(p2mt) )
     {
-        bool ipat;
-        int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
-                                     i * EPT_TABLE_ORDER, &ipat,
-                                     p2mt);
-
-        if ( emt >= 0 )
-            new_entry.emt = emt;
-        else /* ept_handle_misconfig() will need to take care of this. */
-            new_entry.emt = MTRR_NUM_TYPES;
-
-        new_entry.ipat = ipat;
         new_entry.sp = !!i;
         new_entry.sa_p2mt = p2mt;
         new_entry.access = p2ma;
@@ -941,6 +932,22 @@ ept_set_entry(struct p2m_domain *p2m, gf
             need_modify_vtd_table = 0;
 
         ept_p2m_type_to_flags(p2m, &new_entry);
+
+        if ( is_epte_present(&new_entry) )
+        {
+            bool ipat;
+            int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
+                                         i * EPT_TABLE_ORDER, &ipat,

xen/arch/x86/mm/p2m-ept.c ept_set_entry(struct p2m_domain *p2m, gf · line 950

Quoted 10 lines
xen/arch/x86/mm/p2m-ept.c +10 -0
+                                         p2mt);
+
+            BUG_ON(mfn_eq(mfn, INVALID_MFN));
+
+            if ( emt >= 0 )
+                new_entry.emt = emt;
+            else /* ept_handle_misconfig() will need to take care of this. */
+                new_entry.emt = MTRR_NUM_TYPES;
+            new_entry.ipat = ipat;
+        }

Should we assert that if new_entry.emt == MTRR_NUM_TYPES the entry must have the present bit set before the atomic_write_ept_entry() call?

Thanks, Roger.

#9 lore raw

Re: [PATCH v2 for-4.19 2/3] x86/EPT: avoid marking non-present entries for re-configuring

Quoted 22 lines · Roger Pau Monné
Quoted 21 lines · Jan Beulich

For non-present entries EMT, like most other fields, is meaningless to hardware. Make the logic in ept_set_entry() setting the field (and iPAT) conditional upon dealing with a present entry, leaving the value at 0 otherwise. This has two effects for epte_get_entry_emt() which we'll want to leverage subsequently:
1) The call moved here now won't be issued with INVALID_MFN anymore (a respective BUG_ON() is being added).
2) Neither of the other two calls could now be issued with a truncated form of INVALID_MFN anymore (as long as there's no bug anywhere marking an entry present when that was populated using INVALID_MFN).

---
v2: New.

xen/arch/x86/mm/p2m-ept.c +0 -0
@@ -650,6 +650,8 @@ static int cf_check resolve_misconfig(st
             if ( e.emt != MTRR_NUM_TYPES )
                 break;
 
Quoted 7 lines

xen/arch/x86/mm/p2m-ept.c static int cf_check resolve_misconfig(st · line 653

Quoted 1 line
xen/arch/x86/mm/p2m-ept.c +1 -0
+            ASSERT(is_epte_present(&e));

If this is added here, then there's a condition further below:

if ( !is_epte_valid(&e) || !is_epte_present(&e) )

That needs adjusting AFAICT.

I don't think so, because e was re-fetched in between.

Quoted 6 lines

However, in ept_set_entry() we seem to unconditionally call resolve_misconfig() against the new entry to be populated, won't this possibly cause resolve_misconfig() to be called against non-present EPT entries? I think this is fine because such non-present entries will have emt == 0, and hence will take the break just ahead of the added ASSERT().

Right, hence how I placed this assertion.

Quoted 16 lines
Quoted 16 lines
xen/arch/x86/mm/p2m-ept.c +12 -0
@@ -941,6 +932,22 @@ ept_set_entry(struct p2m_domain *p2m, gf
             need_modify_vtd_table = 0;
 
         ept_p2m_type_to_flags(p2m, &new_entry);
+
+        if ( is_epte_present(&new_entry) )
+        {
+            bool ipat;
+            int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
+                                         i * EPT_TABLE_ORDER, &ipat,
+                                         p2mt);
+
+            BUG_ON(mfn_eq(mfn, INVALID_MFN));
+
+            if ( emt >= 0 )
+                new_entry.emt = emt;
Quoted 8 lines

xen/arch/x86/mm/p2m-ept.c ept_set_entry(struct p2m_domain *p2m, gf · line 950

Quoted 4 lines
xen/arch/x86/mm/p2m-ept.c +4 -0
+            else /* ept_handle_misconfig() will need to take care of this. */
+                new_entry.emt = MTRR_NUM_TYPES;
+            new_entry.ipat = ipat;
+        }

Should we assert that if new_entry.emt == MTRR_NUM_TYPES the entry must have the present bit set before the atomic_write_ept_entry() call?

This would feel excessive to me. All writing to new_entry is close together, immediately ahead of that atomic_write_ept_entry(). And we're (now) writing MTRR_NUM_TYPES only when is_epte_present() is true (note that it's not "the present bit").

Jan

Reviewed-by
#10 lore raw

Re: [PATCH v2 for-4.19 2/3] x86/EPT: avoid marking non-present entries for re-configuring

Quoted 13 lines · Jan Beulich
Quoted 12 lines · Roger Pau Monné
Quoted 11 lines · Jan Beulich

For non-present entries EMT, like most other fields, is meaningless to hardware. Make the logic in ept_set_entry() setting the field (and iPAT) conditional upon dealing with a present entry, leaving the value at 0 otherwise. This has two effects for epte_get_entry_emt() which we'll want to leverage subsequently:
1) The call moved here now won't be issued with INVALID_MFN anymore (a respective BUG_ON() is being added).
2) Neither of the other two calls could now be issued with a truncated form of INVALID_MFN anymore (as long as there's no bug anywhere marking an entry present when that was populated using INVALID_MFN).

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 13 lines
Quoted 13 lines

xen/arch/x86/mm/p2m-ept.c static int cf_check resolve_misconfig(st · line 653

Quoted 10 lines

---
v2: New.

xen/arch/x86/mm/p2m-ept.c +1 -0
@@ -650,6 +650,8 @@ static int cf_check resolve_misconfig(st
             if ( e.emt != MTRR_NUM_TYPES )
                 break;
 
+            ASSERT(is_epte_present(&e));

If this is added here, then there's a condition further below:

Quoted 5 lines
Quoted 3 lines

if ( !is_epte_valid(&e) || !is_epte_present(&e) )

That needs adjusting AFAICT.

I don't think so, because e was re-fetched in between.

Oh, I see, we take the opportunity to do the recalculation for all the EPT entries that share the same page table.

Quoted 8 lines
Quoted 6 lines

However, in ept_set_entry() we seem to unconditionally call resolve_misconfig() against the new entry to be populated, won't this possibly cause resolve_misconfig() to be called against non-present EPT entries? I think this is fine because such non-present entries will have emt == 0, and hence will take the break just ahead of the added ASSERT().

Right, hence how I placed this assertion.

OK, just wanted to double check.

Quoted 25 lines
Quoted 16 lines
Quoted 16 lines
xen/arch/x86/mm/p2m-ept.c +12 -0
@@ -941,6 +932,22 @@ ept_set_entry(struct p2m_domain *p2m, gf
             need_modify_vtd_table = 0;
 
         ept_p2m_type_to_flags(p2m, &new_entry);
+
+        if ( is_epte_present(&new_entry) )
+        {
+            bool ipat;
+            int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
+                                         i * EPT_TABLE_ORDER, &ipat,
+                                         p2mt);
+
+            BUG_ON(mfn_eq(mfn, INVALID_MFN));
+
+            if ( emt >= 0 )
+                new_entry.emt = emt;
Quoted 8 lines

xen/arch/x86/mm/p2m-ept.c ept_set_entry(struct p2m_domain *p2m, gf · line 950

Quoted 4 lines
xen/arch/x86/mm/p2m-ept.c +4 -0
+            else /* ept_handle_misconfig() will need to take care of this. */
+                new_entry.emt = MTRR_NUM_TYPES;
+            new_entry.ipat = ipat;
+        }

Should we assert that if new_entry.emt == MTRR_NUM_TYPES the entry must have the present bit set before the atomic_write_ept_entry() call?

Quoted 4 lines

This would feel excessive to me. All writing to new_entry is close together, immediately ahead of that atomic_write_ept_entry(). And we're (now) writing MTRR_NUM_TYPES only when is_epte_present() is true (note that it's not "the present bit").

Fair enough.

Thanks, Roger.

#11 lore raw

Re: [PATCH v2 for-4.19 2/3] x86/EPT: avoid marking non-present entries for re-configuring

In reply to Jan Beulich
Quoted 16 lines · Jan Beulich

For non-present entries EMT, like most other fields, is meaningless to
hardware. Make the logic in ept_set_entry() setting the field (and iPAT)
conditional upon dealing with a present entry, leaving the value at 0 otherwise. This has two effects for epte_get_entry_emt() which we'll want to leverage subsequently:
1) The call moved here now won't be issued with INVALID_MFN anymore (a
   respective BUG_ON() is being added).
2) Neither of the other two calls could now be issued with a truncated
   form of INVALID_MFN anymore (as long as there's no bug anywhere marking an entry present when that was populated using INVALID_MFN).

Quoted 1 line

~ Oleksii

Quoted 60 lines

---
v2: New.

xen/arch/x86/mm/p2m-ept.c +0 -0
@@ -650,6 +650,8 @@ static int cf_check resolve_misconfig(st
              if ( e.emt != MTRR_NUM_TYPES )
                  break;


+            ASSERT(is_epte_present(&e));
+
             if ( level == 0 )
             {
                 for ( gfn -= i, i = 0; i < EPT_PAGETABLE_ENTRIES; ++i )

xen/arch/x86/mm/p2m-ept.c +0 -0
@@ -915,17 +917,6 @@ ept_set_entry(struct p2m_domain *p2m, gf


     if ( mfn_valid(mfn) || p2m_allows_invalid_mfn(p2mt) ) {
-        bool ipat;
-        int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn, -                                     i * EPT_TABLE_ORDER, &ipat, -                                     p2mt);
-
-        if ( emt >= 0 )
-            new_entry.emt = emt;
-        else /* ept_handle_misconfig() will need to take care of this. */
-            new_entry.emt = MTRR_NUM_TYPES;
-
-        new_entry.ipat = ipat;
         new_entry.sp = !!i;
         new_entry.sa_p2mt = p2mt;
         new_entry.access = p2ma;

xen/arch/x86/mm/p2m-ept.c +0 -0
@@ -941,6 +932,22 @@ ept_set_entry(struct p2m_domain *p2m, gf
              need_modify_vtd_table = 0;


         ept_p2m_type_to_flags(p2m, &new_entry);
+
+        if ( is_epte_present(&new_entry) )
+        {
+            bool ipat;
+            int emt = epte_get_entry_emt(p2m->domain, _gfn(gfn), mfn,
+                                         i * EPT_TABLE_ORDER, &ipat, +                                         p2mt);

xen/arch/x86/mm/p2m-ept.c +6 -0
+
+            BUG_ON(mfn_eq(mfn, INVALID_MFN));
+
+            if ( emt >= 0 )
+                new_entry.emt = emt;
+            else /* ept_handle_misconfig() will need to take care of

this. */
+                new_entry.emt = MTRR_NUM_TYPES;
+            new_entry.ipat = ipat;
+        }
     }

     if ( sve != -1 )

#12 lore raw

PATCH v2 3/3 x86/EPT: drop questionable mfn_valid() from epte_get_entry_emt()

In reply to Jan Beulich

mfn_valid() is RAM-focused; it will often return false for MMIO. Yet access to actual MMIO space should not generally be restricted to UC only; especially video frame buffer accesses are unduly affected by such a restriction.

Since, as of ???????????? ("x86/EPT: avoid marking non-present entries for re-configuring"), the function won't be called with INVALID_MFN or, worse, truncated forms thereof anymore, we call fully drop that check.

Fixes 81fd0d3ca4b2 ("x86/hvm: simplify 'mmio_direct' check in epte_get_entry_emt()")

---
Considering that we've just declared PVH Dom0 "supported", this may well qualify for 4.19. The issue was specifically very noticeable there.
---
v2: Different approach (and hence different title and description).

xen/arch/x86/mm/p2m-ept.c +0 -6
@@ -501,12 +501,6 @@ int epte_get_entry_emt(struct domain *d,
         return -1;
     }
 
-    if ( !mfn_valid(mfn) )
-    {
-        *ipat = true;
-        return X86_MT_UC;
-    }
-
     /*
      * Conditional must be kept in sync with the code in
      * {iomem,ioports}_{permit,deny}_access().
Reviewed-by on v2 3/3
#13 lore raw

Re: [PATCH v2 for-4.19 3/3] x86/EPT: drop questionable mfn_valid() from epte_get_entry_emt()

Quoted 5 lines · Jan Beulich

mfn_valid() is RAM-focused; it will often return false for MMIO. Yet access to actual MMIO space should not generally be restricted to UC only; especially video frame buffer accesses are unduly affected by such a restriction.

Quoted 7 lines

Since, as of ???????????? ("x86/EPT: avoid marking non-present entries for re-configuring"), the function won't be called with INVALID_MFN or, worse, truncated forms thereof anymore, we call fully drop that check.

Fixes 81fd0d3ca4b2 ("x86/hvm: simplify 'mmio_direct' check in epte_get_entry_emt()")

I do think this is the way to go (removing quirks from epte_get_entry_emt()), however it's a risky change to make at this point in the release.

If this turns out to cause some unexpected damage, it would only affect HVM guests with PCI passthrough and PVH dom0, which I consider not great, but tolerable.

I would be more comfortable with making the change just not so close to the release, but that's where we are.

I wonder if you should explicitly mention that if adding the mfn_valid() check was done to ensure all mappings to MMIO are created with effective UC caching attribute it won't be fully correct either. Xen could map those using a different effective caching attribute by virtue of host MTRRs being in effect plus Xen chosen PAT attributes.

Thanks, Roger.

#14 lore raw

Re: [PATCH v2 for-4.19 3/3] x86/EPT: drop questionable mfn_valid() from epte_get_entry_emt()

Quoted 18 lines · Roger Pau Monné
Quoted 5 lines · Jan Beulich

mfn_valid() is RAM-focused; it will often return false for MMIO. Yet access to actual MMIO space should not generally be restricted to UC only; especially video frame buffer accesses are unduly affected by such a restriction.

Quoted 7 lines

Since, as of ???????????? ("x86/EPT: avoid marking non-present entries for re-configuring"), the function won't be called with INVALID_MFN or, worse, truncated forms thereof anymore, we call fully drop that check.

Fixes 81fd0d3ca4b2 ("x86/hvm: simplify 'mmio_direct' check in epte_get_entry_emt()")

I do think this is the way to go (removing quirks from epte_get_entry_emt()), however it's a risky change to make at this point in the release.

Quoted 6 lines

If this turns out to cause some unexpected damage, it would only affect HVM guests with PCI passthrough and PVH dom0, which I consider not great, but tolerable.

I would be more comfortable with making the change just not so close to the release, but that's where we are.

Certainly, and I could live with Oleksii revoking his R-a-b (or simply not offering it for either of the two prereq changes). Main thing for me is - PVH Dom0 finally isn't so horribly slow anymore. However, if it doesn't go into the release, then I'd also be unsure about eventual backporting.

Quoted 1 line

Thanks.

Quoted 5 lines

I wonder if you should explicitly mention that if adding the mfn_valid() check was done to ensure all mappings to MMIO are created with effective UC caching attribute it won't be fully correct either. Xen could map those using a different effective caching attribute by virtue of host MTRRs being in effect plus Xen chosen PAT attributes.

Well, the mfn_valid() can't have been there to cover _all_ MMIO. It was maybe a flawed initial attempt at doing so, and then wasn't properly adjusted / dropped. So overall - no, I don't think extending the description with anything along the lines of the above would make a lot of sense.

Jan

#15 lore raw

Re: [PATCH v2 for-4.19 3/3] x86/EPT: drop questionable mfn_valid() from epte_get_entry_emt()

Quoted 7 lines · Jan Beulich
Quoted 5 lines · Roger Pau Monné

I wonder if you should explicitly mention that if adding the mfn_valid() check was done to ensure all mappings to MMIO are created with effective UC caching attribute it won't be fully correct either. Xen could map those using a different effective caching attribute by virtue of host MTRRs being in effect plus Xen chosen PAT attributes.

Quoted 5 lines

Well, the mfn_valid() can't have been there to cover _all_ MMIO. It was maybe a flawed initial attempt at doing so, and then wasn't properly adjusted / dropped. So overall - no, I don't think extending the description with anything along the lines of the above would make a lot of sense.

I realized myself when writing the paragraph that I wouldn't even know how to word it properly, neither it would be much helpful without knowing the exact intention the mfn_valid() check was added for.

Thanks, Roger.

#16 lore raw

Re: [PATCH v2 for-4.19 3/3] x86/EPT: drop questionable mfn_valid() from epte_get_entry_emt()

In reply to Jan Beulich
Quoted 23 lines · Jan Beulich
Quoted 22 lines · Roger Pau Monné
Quoted 5 lines · Jan Beulich

mfn_valid() is RAM-focused; it will often return false for MMIO. Yet access to actual MMIO space should not generally be restricted to UC only; especially video frame buffer accesses are unduly affected by such a restriction.

Quoted 7 lines

Since, as of ???????????? ("x86/EPT: avoid marking non-present entries for re-configuring"), the function won't be called with INVALID_MFN or, worse, truncated forms thereof anymore, we call fully drop that check.

Fixes 81fd0d3ca4b2 ("x86/hvm: simplify 'mmio_direct' check in epte_get_entry_emt()")

I do think this is the way to go (removing quirks from epte_get_entry_emt()), however it's a risky change to make at this point in the release.

If this turns out to cause some unexpected damage, it would only affect HVM guests with PCI passthrough and PVH dom0, which I consider not great, but tolerable.

Quoted 8 lines
Quoted 2 lines

I would be more comfortable with making the change just not so close to the release, but that's where we are.

Certainly, and I could live with Oleksii revoking his R-a-b (or simply not offering it for either of the two prereq changes). Main thing for me is - PVH Dom0 finally isn't so horribly slow anymore. However, if it doesn't go into the release, then I'd also be unsure about eventual backporting.

Thinking about this, it's also likely to fix issues with PCI passthrough to HVM guests, so I'm quite sure we would need to backport it.

David Woodhouse already had to fix it once:

https://xenbits.xen.org/gitweb/?p=xen.git;a=commit;h=30921dc2df3665ca1b2593595aa6725ff013d386

And I'm quite sure this fix was not related to PVH dom0.

Thanks, Roger.