Skip to contentSkip to conversation

use new API for Xen page tables

PATCH Superseded Hongyan Xia 3 participants 22 messages Apr 8, 2020 – Apr 28, 2020

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v2 cover use new API for Xen page tables

From: Hongyan Xia <[email protected]>

This small series is basically just rewriting functions using the new API to map and unmap PTEs. Each patch is independent.

Apart from mapping and unmapping page tables, no other functional change intended.

---
Changed in v2:
- I kept UNMAP_DOMAIN_PAGE() for now in v2, but I if people say in some cases it is an overkill and unmap_domain_page() should be used, I am okay with that and can make the change.
- code cleanup and style fixes.
- unmap as early as possible.

Wei Liu (5):
x86/shim: map and unmap page tables in replace_va_mapping x86_64/mm: map and unmap page tables in m2p_mapped x86_64/mm: map and unmap page tables in share_hotadd_m2p_table x86_64/mm: map and unmap page tables in destroy_compat_m2p_mapping x86_64/mm: map and unmap page tables in destroy_m2p_mapping

 xen/arch/x86/pv/shim.c     |  9 ++++---
 xen/arch/x86/x86_64/mm.c   | 55 ++++++++++++++++++++++----------------
 xen/include/asm-x86/page.h | 13 +++++++++
 3 files changed, 50 insertions(+), 27 deletions(-)
Signature · 1 line
2.24.1.AMZN
#1 lore raw

PATCH v2 1/5 x86/shim: map and unmap page tables in replace_va_mapping

From: Wei Liu <[email protected]>

Also, introduce lYe_from_lXe() macros which do not rely on the direct map when walking page tables. Unfortunately, they cannot be inline functions due to the header dependency on domain_page.h, so keep them as macros just like map_lYt_from_lXe().

---
Changed in v2:
- unmap as early as possible instead of unmapping all at the end.
- use lYe_from_lXe() macros and lift them from a later patch to this patch.
- const qualify pointers in new macros.

---
 xen/arch/x86/pv/shim.c     |  9 +++++----
 xen/include/asm-x86/page.h | 13 +++++++++++++
 2 files changed, 18 insertions(+), 4 deletions(-)
xen/arch/x86/pv/shim.c +5 -4
index ed2ece8a8a..28d2076065 100644
@@ -168,16 +168,17 @@ const struct platform_bad_page *__init pv_shim_reserved_pages(unsigned int *size
 static void __init replace_va_mapping(struct domain *d, l4_pgentry_t *l4start,
                                       unsigned long va, mfn_t mfn)
 {
-    l4_pgentry_t *pl4e = l4start + l4_table_offset(va);
-    l3_pgentry_t *pl3e = l4e_to_l3e(*pl4e) + l3_table_offset(va);
-    l2_pgentry_t *pl2e = l3e_to_l2e(*pl3e) + l2_table_offset(va);
-    l1_pgentry_t *pl1e = l2e_to_l1e(*pl2e) + l1_table_offset(va);
+    l4_pgentry_t l4e = l4start[l4_table_offset(va)];
+    l3_pgentry_t l3e = l3e_from_l4e(l4e, l3_table_offset(va));
+    l2_pgentry_t l2e = l2e_from_l3e(l3e, l2_table_offset(va));
+    l1_pgentry_t *pl1e = map_l1t_from_l2e(l2e) + l1_table_offset(va);
     struct page_info *page = mfn_to_page(l1e_get_mfn(*pl1e));
 
     put_page_and_type(page);
 
     *pl1e = l1e_from_mfn(mfn, (!is_pv_32bit_domain(d) ? L1_PROT
                                                       : COMPAT_L1_PROT));
+    UNMAP_DOMAIN_PAGE(pl1e);
 }
 
 static void evtchn_reserve(struct domain *d, unsigned int port)
xen/include/asm-x86/page.h +13 -0
index c98d8f5ede..1e16f3980d 100644
@@ -196,6 +196,19 @@ static inline l4_pgentry_t l4e_from_paddr(paddr_t pa, unsigned int flags)
 #define map_l2t_from_l3e(x)        (l2_pgentry_t *)map_domain_page(l3e_get_mfn(x))
 #define map_l3t_from_l4e(x)        (l3_pgentry_t *)map_domain_page(l4e_get_mfn(x))
 
+/* Unlike lYe_to_lXe(), lXe_from_lYe() do not rely on the direct map. */
+#define l2e_from_l3e(l3e, offset) ({                        \
+        const l2_pgentry_t *l2t = map_l2t_from_l3e(l3e);    \
+        l2_pgentry_t l2e = l2t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l2t);                             \
+        l2e; })
+
+#define l3e_from_l4e(l4e, offset) ({                        \
+        const l3_pgentry_t *l3t = map_l3t_from_l4e(l4e);    \
+        l3_pgentry_t l3e = l3t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l3t);                             \
+        l3e; })
+
 /* Given a virtual address, get an entry offset into a page table. */
 #define l1_table_offset(a)         \
     (((a) >> L1_PAGETABLE_SHIFT) & (L1_PAGETABLE_ENTRIES - 1))
Signature · 1 line
2.24.1.AMZN
#2 lore raw

Re: [PATCH v2 1/5] x86/shim: map and unmap page tables in replace_va_mapping

Quoted 12 lines · Hongyan Xia
xen/arch/x86/pv/shim.c +2 -4
@@ -168,16 +168,17 @@ const struct platform_bad_page *__init pv_shim_reserved_pages(unsigned int *size
 static void __init replace_va_mapping(struct domain *d, l4_pgentry_t *l4start,
                                       unsigned long va, mfn_t mfn)
 {
-    l4_pgentry_t *pl4e = l4start + l4_table_offset(va);
-    l3_pgentry_t *pl3e = l4e_to_l3e(*pl4e) + l3_table_offset(va);
-    l2_pgentry_t *pl2e = l3e_to_l2e(*pl3e) + l2_table_offset(va);
-    l1_pgentry_t *pl1e = l2e_to_l1e(*pl2e) + l1_table_offset(va);
+    l4_pgentry_t l4e = l4start[l4_table_offset(va)];
+    l3_pgentry_t l3e = l3e_from_l4e(l4e, l3_table_offset(va));

xen/arch/x86/pv/shim.c const struct platform_bad_page *__init pv_shim_reserved_pages(unsigned int *size · line 182

Quoted 10 lines
xen/arch/x86/pv/shim.c +3 -0
+    l2_pgentry_t l2e = l2e_from_l3e(l3e, l2_table_offset(va));
+    l1_pgentry_t *pl1e = map_l1t_from_l2e(l2e) + l1_table_offset(va);
     struct page_info *page = mfn_to_page(l1e_get_mfn(*pl1e));
 
     put_page_and_type(page);
 
     *pl1e = l1e_from_mfn(mfn, (!is_pv_32bit_domain(d) ? L1_PROT
                                                       : COMPAT_L1_PROT));
+    UNMAP_DOMAIN_PAGE(pl1e);
 }

As said before, here and below I think it should be unmap_domain_page().

Quoted 8 lines
xen/include/asm-x86/page.h +2 -0
@@ -196,6 +196,19 @@ static inline l4_pgentry_t l4e_from_paddr(paddr_t pa, unsigned int flags)
 #define map_l2t_from_l3e(x)        (l2_pgentry_t *)map_domain_page(l3e_get_mfn(x))
 #define map_l3t_from_l4e(x)        (l3_pgentry_t *)map_domain_page(l4e_get_mfn(x))
 
+/* Unlike lYe_to_lXe(), lXe_from_lYe() do not rely on the direct map. */
+#define l2e_from_l3e(l3e, offset) ({                        \

xen/include/asm-x86/page.h static inline l4_pgentry_t l4e_from_paddr(paddr_t pa, unsigned int flags) · line 210

Quoted 10 lines
xen/include/asm-x86/page.h +10 -0
+        const l2_pgentry_t *l2t = map_l2t_from_l3e(l3e);    \
+        l2_pgentry_t l2e = l2t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l2t);                             \
+        l2e; })
+
+#define l3e_from_l4e(l4e, offset) ({                        \
+        const l3_pgentry_t *l3t = map_l3t_from_l4e(l4e);    \
+        l3_pgentry_t l3e = l3t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l3t);                             \
+        l3e; })

I think l1e_from_l2e() should be introduced at the same time, even if for now it's unused. I also think, like we do elsewhere, that macro-local variables would better have _ suffixes, to avoid possible variable aliasing issues.

Jan

#3 lore raw

Re: [PATCH v2 1/5] x86/shim: map and unmap page tables in replace_va_mapping

Quoted 50 lines · Jan Beulich
Quoted 23 lines · Hongyan Xia
xen/arch/x86/pv/shim.c +0 -0
@@ -168,16 +168,17 @@ const struct platform_bad_page *__init

pv_shim_reserved_pages(unsigned int *size
static void __init replace_va_mapping(struct domain *d,
l4_pgentry_t *l4start,
unsigned long va, mfn_t mfn)
{

xen/arch/x86/pv/shim.c +4 -4
-    l4_pgentry_t *pl4e = l4start + l4_table_offset(va);
-    l3_pgentry_t *pl3e = l4e_to_l3e(*pl4e) + l3_table_offset(va);
-    l2_pgentry_t *pl2e = l3e_to_l2e(*pl3e) + l2_table_offset(va);
-    l1_pgentry_t *pl1e = l2e_to_l1e(*pl2e) + l1_table_offset(va);
+    l4_pgentry_t l4e = l4start[l4_table_offset(va)];
+    l3_pgentry_t l3e = l3e_from_l4e(l4e, l3_table_offset(va));
+    l2_pgentry_t l2e = l2e_from_l3e(l3e, l2_table_offset(va));
+    l1_pgentry_t *pl1e = map_l1t_from_l2e(l2e) +
l1_table_offset(va);
     struct page_info *page = mfn_to_page(l1e_get_mfn(*pl1e));
 
     put_page_and_type(page);
 
     *pl1e = l1e_from_mfn(mfn, (!is_pv_32bit_domain(d) ? L1_PROT
                                                       :
Quoted 3 lines

COMPAT_L1_PROT));
+ UNMAP_DOMAIN_PAGE(pl1e);
}

As said before, here and below I think it should be unmap_domain_page().

Quoted 19 lines
xen/include/asm-x86/page.h +1 -0
@@ -196,6 +196,19 @@ static inline l4_pgentry_t
 l4e_from_paddr(paddr_t pa, unsigned int flags)
 #define map_l2t_from_l3e(x)        (l2_pgentry_t
 *)map_domain_page(l3e_get_mfn(x))
 #define map_l3t_from_l4e(x)        (l3_pgentry_t
 *)map_domain_page(l4e_get_mfn(x))
 
+/* Unlike lYe_to_lXe(), lXe_from_lYe() do not rely on the direct

map. */

xen/include/asm-x86/page.h +8 -0
+#define l2e_from_l3e(l3e, offset) ({                        \
+        const l2_pgentry_t *l2t = map_l2t_from_l3e(l3e);    \
+        l2_pgentry_t l2e = l2t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l2t);                             \
+        l2e; })
+
+#define l3e_from_l4e(l4e, offset) ({                        \
+        const l3_pgentry_t *l3t = map_l3t_from_l4e(l4e);    \
Quoted 8 lines

xen/include/asm-x86/page.h static inline l4_pgentry_t l4e_from_paddr(paddr_t pa, unsigned int flags) · line 210

Quoted 3 lines
xen/include/asm-x86/page.h +3 -0
+        l3_pgentry_t l3e = l3t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l3t);                             \
+        l3e; })

I think l1e_from_l2e() should be introduced at the same time, even if for now it's unused. I also think, like we do elsewhere, that macro-local variables would better have _ suffixes, to avoid possible variable aliasing issues.

Shall I address the comments and send a new rev now, or is this small series still being reviewed?

Hongyan

#4 lore raw

Re: [PATCH v2 1/5] x86/shim: map and unmap page tables in replace_va_mapping

Quoted 60 lines · Hongyan Xia
Quoted 50 lines · Jan Beulich
Quoted 23 lines · Hongyan Xia
xen/arch/x86/pv/shim.c +0 -0
@@ -168,16 +168,17 @@ const struct platform_bad_page *__init

pv_shim_reserved_pages(unsigned int *size
static void __init replace_va_mapping(struct domain *d,
l4_pgentry_t *l4start,
unsigned long va, mfn_t mfn)
{

xen/arch/x86/pv/shim.c +4 -4
-    l4_pgentry_t *pl4e = l4start + l4_table_offset(va);
-    l3_pgentry_t *pl3e = l4e_to_l3e(*pl4e) + l3_table_offset(va);
-    l2_pgentry_t *pl2e = l3e_to_l2e(*pl3e) + l2_table_offset(va);
-    l1_pgentry_t *pl1e = l2e_to_l1e(*pl2e) + l1_table_offset(va);
+    l4_pgentry_t l4e = l4start[l4_table_offset(va)];
+    l3_pgentry_t l3e = l3e_from_l4e(l4e, l3_table_offset(va));
+    l2_pgentry_t l2e = l2e_from_l3e(l3e, l2_table_offset(va));
+    l1_pgentry_t *pl1e = map_l1t_from_l2e(l2e) +
l1_table_offset(va);
     struct page_info *page = mfn_to_page(l1e_get_mfn(*pl1e));
 
     put_page_and_type(page);
 
     *pl1e = l1e_from_mfn(mfn, (!is_pv_32bit_domain(d) ? L1_PROT
                                                       :
Quoted 3 lines

COMPAT_L1_PROT));
+ UNMAP_DOMAIN_PAGE(pl1e);
}

As said before, here and below I think it should be unmap_domain_page().

Quoted 19 lines
xen/include/asm-x86/page.h +1 -0
@@ -196,6 +196,19 @@ static inline l4_pgentry_t
 l4e_from_paddr(paddr_t pa, unsigned int flags)
 #define map_l2t_from_l3e(x)        (l2_pgentry_t
 *)map_domain_page(l3e_get_mfn(x))
 #define map_l3t_from_l4e(x)        (l3_pgentry_t
 *)map_domain_page(l4e_get_mfn(x))
 
+/* Unlike lYe_to_lXe(), lXe_from_lYe() do not rely on the direct

map. */

xen/include/asm-x86/page.h +8 -0
+#define l2e_from_l3e(l3e, offset) ({                        \
+        const l2_pgentry_t *l2t = map_l2t_from_l3e(l3e);    \
+        l2_pgentry_t l2e = l2t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l2t);                             \
+        l2e; })
+
+#define l3e_from_l4e(l4e, offset) ({                        \
+        const l3_pgentry_t *l3t = map_l3t_from_l4e(l4e);    \
Quoted 8 lines

xen/include/asm-x86/page.h static inline l4_pgentry_t l4e_from_paddr(paddr_t pa, unsigned int flags) · line 210

Quoted 3 lines
xen/include/asm-x86/page.h +3 -0
+        l3_pgentry_t l3e = l3t[offset];                     \
+        UNMAP_DOMAIN_PAGE(l3t);                             \
+        l3e; })

I think l1e_from_l2e() should be introduced at the same time, even if for now it's unused. I also think, like we do elsewhere, that macro-local variables would better have _ suffixes, to avoid possible variable aliasing issues.

Quoted 2 lines

Shall I address the comments and send a new rev now, or is this small series still being reviewed?

I didn't get to look at patches 2 thru 5 yet, if this (partly) answers the question.

Jan

#5 lore raw

PATCH v2 2/5 x86_64/mm: map and unmap page tables in m2p_mapped

In reply to Hongyan Xia

From: Wei Liu <[email protected]>

---
Changed in v2:
- avoid adding goto labels, simply get the PTE and unmap quickly.
- code style fixes.

---
 xen/arch/x86/x86_64/mm.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
xen/arch/x86/x86_64/mm.c +7 -7
index cee836ec37..8e0caa7327 100644
@@ -129,14 +129,14 @@ static mfn_t alloc_hotadd_mfn(struct mem_hotadd_info *info)
 static int m2p_mapped(unsigned long spfn)
 {
     unsigned long va;
-    l3_pgentry_t *l3_ro_mpt;
-    l2_pgentry_t *l2_ro_mpt;
+    l3_pgentry_t l3e_ro_mpt;
+    l2_pgentry_t l2e_ro_mpt;
 
     va = RO_MPT_VIRT_START + spfn * sizeof(*machine_to_phys_mapping);
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(va)]);
+    l3e_ro_mpt = l3e_from_l4e(idle_pg_table[l4_table_offset(va)],
+                              l3_table_offset(va));
 
-    switch ( l3e_get_flags(l3_ro_mpt[l3_table_offset(va)]) &
-             (_PAGE_PRESENT |_PAGE_PSE))
+    switch ( l3e_get_flags(l3e_ro_mpt) & (_PAGE_PRESENT | _PAGE_PSE) )
     {
         case _PAGE_PSE|_PAGE_PRESENT:
             return M2P_1G_MAPPED;
@@ -146,9 +146,9 @@ static int m2p_mapped(unsigned long spfn)
         default:
             return M2P_NO_MAPPED;
     }
-    l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
+    l2e_ro_mpt = l2e_from_l3e(l3e_ro_mpt, l2_table_offset(va));
 
-    if (l2e_get_flags(l2_ro_mpt[l2_table_offset(va)]) & _PAGE_PRESENT)
+    if ( l2e_get_flags(l2e_ro_mpt) & _PAGE_PRESENT )
         return M2P_2M_MAPPED;
 
     return M2P_NO_MAPPED;
Signature · 1 line
2.24.1.AMZN
Reviewed-by on v2 2/5
#6 lore raw

Re: [PATCH v2 2/5] x86_64/mm: map and unmap page tables in m2p_mapped

xen/arch/x86/x86_64/mm.c static mfn_t alloc_hotadd_mfn(struct mem_hotadd_info *info) · line 133

Quoted 10 lines · Hongyan Xia
xen/arch/x86/x86_64/mm.c +2 -2
@@ -129,14 +129,14 @@ static mfn_t alloc_hotadd_mfn(struct mem_hotadd_info *info)
 static int m2p_mapped(unsigned long spfn)
 {
     unsigned long va;
-    l3_pgentry_t *l3_ro_mpt;
-    l2_pgentry_t *l2_ro_mpt;
+    l3_pgentry_t l3e_ro_mpt;
+    l2_pgentry_t l2e_ro_mpt;

Preferably with the _ro_mpt tags here dropped

Jan

#7 lore raw

PATCH v2 3/5 x86_64/mm: map and unmap page tables in share_hotadd_m2p_table

In reply to Hongyan Xia

From: Wei Liu <[email protected]>

Fetch lYe by mapping and unmapping lXe instead of using the direct map, which is now done via the lYe_from_lXe() helpers.

---
Changed in v2:
- the introduction of the macros is now lifted to a previous patch.

---
 xen/arch/x86/x86_64/mm.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
xen/arch/x86/x86_64/mm.c +6 -6
index 8e0caa7327..d23c7e4f5b 100644
@@ -167,14 +167,14 @@ static int share_hotadd_m2p_table(struct mem_hotadd_info *info)
           v += n << PAGE_SHIFT )
     {
         n = L2_PAGETABLE_ENTRIES * L1_PAGETABLE_ENTRIES;
-        l3e = l4e_to_l3e(idle_pg_table[l4_table_offset(v)])[
-            l3_table_offset(v)];
+        l3e = l3e_from_l4e(idle_pg_table[l4_table_offset(v)],
+                           l3_table_offset(v));
         if ( !(l3e_get_flags(l3e) & _PAGE_PRESENT) )
             continue;
         if ( !(l3e_get_flags(l3e) & _PAGE_PSE) )
         {
             n = L1_PAGETABLE_ENTRIES;
-            l2e = l3e_to_l2e(l3e)[l2_table_offset(v)];
+            l2e = l2e_from_l3e(l3e, l2_table_offset(v));
             if ( !(l2e_get_flags(l2e) & _PAGE_PRESENT) )
                 continue;
             m2p_start_mfn = l2e_get_mfn(l2e);
@@ -195,11 +195,11 @@ static int share_hotadd_m2p_table(struct mem_hotadd_info *info)
           v != RDWR_COMPAT_MPT_VIRT_END;
           v += 1 << L2_PAGETABLE_SHIFT )
     {
-        l3e = l4e_to_l3e(idle_pg_table[l4_table_offset(v)])[
-            l3_table_offset(v)];
+        l3e = l3e_from_l4e(idle_pg_table[l4_table_offset(v)],
+                           l3_table_offset(v));
         if ( !(l3e_get_flags(l3e) & _PAGE_PRESENT) )
             continue;
-        l2e = l3e_to_l2e(l3e)[l2_table_offset(v)];
+        l2e = l2e_from_l3e(l3e, l2_table_offset(v));
         if ( !(l2e_get_flags(l2e) & _PAGE_PRESENT) )
             continue;
         m2p_start_mfn = l2e_get_mfn(l2e);
Signature · 1 line
2.24.1.AMZN
#9 lore raw

PATCH v2 4/5 x86_64/mm: map and unmap page tables in destroy_compat_m2p_mapping

In reply to Hongyan Xia

From: Wei Liu <[email protected]>

---
Changed in v2:
- there is pretty much no point in keeping l3_ro_mpt mapped, just fetch the l3e instead, which also cleans up the code.

---
 xen/arch/x86/x86_64/mm.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
xen/arch/x86/x86_64/mm.c +8 -4
index d23c7e4f5b..ae8f4dd0b9 100644
@@ -220,7 +220,7 @@ static void destroy_compat_m2p_mapping(struct mem_hotadd_info *info)
     unsigned long i, va, rwva, pt_pfn;
     unsigned long smap = info->spfn, emap = info->spfn;
 
-    l3_pgentry_t *l3_ro_mpt;
+    l3_pgentry_t l3e_ro_mpt;
     l2_pgentry_t *l2_ro_mpt;
 
     if ( smap > ((RDWR_COMPAT_MPT_VIRT_END - RDWR_COMPAT_MPT_VIRT_START) >> 2) )
@@ -229,11 +229,13 @@ static void destroy_compat_m2p_mapping(struct mem_hotadd_info *info)
     if ( emap > ((RDWR_COMPAT_MPT_VIRT_END - RDWR_COMPAT_MPT_VIRT_START) >> 2) )
         emap = (RDWR_COMPAT_MPT_VIRT_END - RDWR_COMPAT_MPT_VIRT_START) >> 2;
 
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(HIRO_COMPAT_MPT_VIRT_START)]);
+    l3e_ro_mpt = l3e_from_l4e(idle_pg_table[
+                                  l4_table_offset(HIRO_COMPAT_MPT_VIRT_START)],
+                              l3_table_offset(HIRO_COMPAT_MPT_VIRT_START));
 
-    ASSERT(l3e_get_flags(l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)]) & _PAGE_PRESENT);
+    ASSERT(l3e_get_flags(l3e_ro_mpt) & _PAGE_PRESENT);
 
-    l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)]);
+    l2_ro_mpt = map_l2t_from_l3e(l3e_ro_mpt);
 
     for ( i = smap; i < emap; )
     {
@@ -255,6 +257,8 @@ static void destroy_compat_m2p_mapping(struct mem_hotadd_info *info)
         i += 1UL << (L2_PAGETABLE_SHIFT - 2);
     }
 
+    UNMAP_DOMAIN_PAGE(l2_ro_mpt);
+
     return;
 }
 
Signature · 1 line
2.24.1.AMZN
#10 lore raw

cover x86: high compat r/o M2P table handling adjustments

While looking at "x86_64/mm: map and unmap page tables in destroy_compat_m2p_mapping" it occurred to me that the mappings changed there can be dropped altogether, as can the linear address range used for this. Note that both patches have only been lightly tested so far, I guess in particular the 2nd one may still have issues.

1: x86: drop unnecessary page table walking in compat r/o M2P handling 2: x86: drop high compat r/o M2P table address range

Jan

#11 lore raw

PATCH 1/2 x86: drop unnecessary page table walking in compat r/o M2P handling

We have a global variable where the necessary L2 table is recorded; no need to inspect L4 and L3 tables (and this way a few less places will eventually need adjustment when we want to support 5-level page tables). Also avoid setting up the L3 entry, as the address range never gets used anyway (it'll be dropped altogether in a subsequent patch).

xen/arch/x86/x86_64/mm.c +2 -23
@@ -219,9 +219,7 @@ static void destroy_compat_m2p_mapping(s
 {
     unsigned long i, va, rwva, pt_pfn;
     unsigned long smap = info->spfn, emap = info->spfn;
-
-    l3_pgentry_t *l3_ro_mpt;
-    l2_pgentry_t *l2_ro_mpt;
+    l2_pgentry_t *l2_ro_mpt = compat_idle_pg_table_l2;
 
     if ( smap > ((RDWR_COMPAT_MPT_VIRT_END - RDWR_COMPAT_MPT_VIRT_START) >> 2) )
         return;
@@ -229,12 +227,6 @@ static void destroy_compat_m2p_mapping(s
     if ( emap > ((RDWR_COMPAT_MPT_VIRT_END - RDWR_COMPAT_MPT_VIRT_START) >> 2) )
         emap = (RDWR_COMPAT_MPT_VIRT_END - RDWR_COMPAT_MPT_VIRT_START) >> 2;
 
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(HIRO_COMPAT_MPT_VIRT_START)]);
-
-    ASSERT(l3e_get_flags(l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)]) & _PAGE_PRESENT);
-
-    l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)]);
-
     for ( i = smap; i < emap; )
     {
         va = HIRO_COMPAT_MPT_VIRT_START +
@@ -323,7 +315,6 @@ static int setup_compat_m2p_table(struct
     unsigned long i, va, smap, emap, rwva, epfn = info->epfn;
     mfn_t mfn;
     unsigned int n;
-    l3_pgentry_t *l3_ro_mpt = NULL;
     l2_pgentry_t *l2_ro_mpt = NULL;
     int err = 0;
 
@@ -342,13 +333,7 @@ static int setup_compat_m2p_table(struct
     emap = ( (epfn + ((1UL << (L2_PAGETABLE_SHIFT - 2)) - 1 )) &
                 ~((1UL << (L2_PAGETABLE_SHIFT - 2)) - 1) );
 
-    va = HIRO_COMPAT_MPT_VIRT_START +
-         smap * sizeof(*compat_machine_to_phys_mapping);
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(va)]);
-
-    ASSERT(l3e_get_flags(l3_ro_mpt[l3_table_offset(va)]) & _PAGE_PRESENT);
-
-    l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
+    l2_ro_mpt = compat_idle_pg_table_l2;
 
 #define MFN(x) (((x) << L2_PAGETABLE_SHIFT) / sizeof(unsigned int))
 #define CNT ((sizeof(*frame_table) & -sizeof(*frame_table)) / \
@@ -627,16 +612,10 @@ void __init paging_init(void)
 #undef MFN
 
     /* Create user-accessible L2 directory to map the MPT for compat guests. */
-    BUILD_BUG_ON(l4_table_offset(RDWR_MPT_VIRT_START) !=
-                 l4_table_offset(HIRO_COMPAT_MPT_VIRT_START));
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(
-        HIRO_COMPAT_MPT_VIRT_START)]);
     if ( (l2_ro_mpt = alloc_xen_pagetable()) == NULL )
         goto nomem;
     compat_idle_pg_table_l2 = l2_ro_mpt;
     clear_page(l2_ro_mpt);
-    l3e_write(&l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)],
-              l3e_from_paddr(__pa(l2_ro_mpt), __PAGE_HYPERVISOR_RO));
     l2_ro_mpt += l2_table_offset(HIRO_COMPAT_MPT_VIRT_START);
     /* Allocate and map the compatibility mode machine-to-phys table. */
     mpt_size = (mpt_size >> 1) + (1UL << (L2_PAGETABLE_SHIFT - 1));
author Reviewed-by on v1 1/2
#12 lore raw

Re: [PATCH 1/2] x86: drop unnecessary page table walking in compat r/o M2P handling

Quoted 84 lines · Jan Beulich

We have a global variable where the necessary L2 table is recorded; no
need to inspect L4 and L3 tables (and this way a few less places will eventually need adjustment when we want to support 5-level page tables).
Also avoid setting up the L3 entry, as the address range never gets used
anyway (it'll be dropped altogether in a subsequent patch).

xen/arch/x86/x86_64/mm.c +1 -3
@@ -219,9 +219,7 @@ static void destroy_compat_m2p_mapping(s
 {
     unsigned long i, va, rwva, pt_pfn;
     unsigned long smap = info->spfn, emap = info->spfn;
-
-    l3_pgentry_t *l3_ro_mpt;
-    l2_pgentry_t *l2_ro_mpt;
+    l2_pgentry_t *l2_ro_mpt = compat_idle_pg_table_l2;
 
     if ( smap > ((RDWR_COMPAT_MPT_VIRT_END -
 RDWR_COMPAT_MPT_VIRT_START) >> 2) )
         return;
xen/arch/x86/x86_64/mm.c +0 -1
@@ -229,12 +227,6 @@ static void destroy_compat_m2p_mapping(s
     if ( emap > ((RDWR_COMPAT_MPT_VIRT_END -
 RDWR_COMPAT_MPT_VIRT_START) >> 2) )
         emap = (RDWR_COMPAT_MPT_VIRT_END -
 RDWR_COMPAT_MPT_VIRT_START) >> 2;
 
-    l3_ro_mpt =

l4e_to_l3e(idle_pg_table[l4_table_offset(HIRO_COMPAT_MPT_VIRT_START)] );

xen/arch/x86/x86_64/mm.c +0 -2
-
-    ASSERT(l3e_get_flags(l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_V

IRT_START)]) & _PAGE_PRESENT);

xen/arch/x86/x86_64/mm.c +0 -2
-
-    l2_ro_mpt =

l3e_to_l2e(l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)]);

xen/arch/x86/x86_64/mm.c +1 -9
-
     for ( i = smap; i < emap; )
     {
         va = HIRO_COMPAT_MPT_VIRT_START +
@@ -323,7 +315,6 @@ static int setup_compat_m2p_table(struct
     unsigned long i, va, smap, emap, rwva, epfn = info->epfn;
     mfn_t mfn;
     unsigned int n;
-    l3_pgentry_t *l3_ro_mpt = NULL;
     l2_pgentry_t *l2_ro_mpt = NULL;
     int err = 0;
 
@@ -342,13 +333,7 @@ static int setup_compat_m2p_table(struct
     emap = ( (epfn + ((1UL << (L2_PAGETABLE_SHIFT - 2)) - 1 )) &
                 ~((1UL << (L2_PAGETABLE_SHIFT - 2)) - 1) );
 
-    va = HIRO_COMPAT_MPT_VIRT_START +
-         smap * sizeof(*compat_machine_to_phys_mapping);
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(va)]);
-
-    ASSERT(l3e_get_flags(l3_ro_mpt[l3_table_offset(va)]) &
 _PAGE_PRESENT);
-
-    l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
+    l2_ro_mpt = compat_idle_pg_table_l2;
 
 #define MFN(x) (((x) << L2_PAGETABLE_SHIFT) / sizeof(unsigned int))

#define CNT ((sizeof(*frame_table) & -sizeof(*frame_table)) / \

xen/arch/x86/x86_64/mm.c +0 -0
@@ -627,16 +612,10 @@ void __init paging_init(void)
 #undef MFN
 
     /* Create user-accessible L2 directory to map the MPT for compat

guests. */

xen/arch/x86/x86_64/mm.c +0 -5
-    BUILD_BUG_ON(l4_table_offset(RDWR_MPT_VIRT_START) !=
-                 l4_table_offset(HIRO_COMPAT_MPT_VIRT_START));
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(
-        HIRO_COMPAT_MPT_VIRT_START)]);
     if ( (l2_ro_mpt = alloc_xen_pagetable()) == NULL )
         goto nomem;
     compat_idle_pg_table_l2 = l2_ro_mpt;
     clear_page(l2_ro_mpt);
-    l3e_write(&l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)

],
- l3e_from_paddr(__pa(l2_ro_mpt),

Quoted 5 lines
__PAGE_HYPERVISOR_RO));
     l2_ro_mpt += l2_table_offset(HIRO_COMPAT_MPT_VIRT_START);
     /* Allocate and map the compatibility mode machine-to-phys
table. */
     mpt_size = (mpt_size >> 1) + (1UL << (L2_PAGETABLE_SHIFT - 1));

The code around here, I am wondering if there is a reason to put it in this patch. If we bisect, we can end up in a commit which says the address range of compat is still there in config.h, but actually it's gone, so this probably belongs to the 2nd patch.

Apart from that,

I would like to drop relevant map/unmap patches and replace them with the new clean-up ones (and place them at the beginning of the series), if there is no objection with that.

Hongyan

#13 lore raw

Re: [PATCH 1/2] x86: drop unnecessary page table walking in compat r/o M2P handling

Quoted 19 lines · Hongyan Xia
Quoted 18 lines · Jan Beulich
xen/arch/x86/x86_64/mm.c +0 -0
@@ -627,16 +612,10 @@ void __init paging_init(void)
 #undef MFN
 
     /* Create user-accessible L2 directory to map the MPT for compat

guests. */

xen/arch/x86/x86_64/mm.c +0 -5
-    BUILD_BUG_ON(l4_table_offset(RDWR_MPT_VIRT_START) !=
-                 l4_table_offset(HIRO_COMPAT_MPT_VIRT_START));
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(
-        HIRO_COMPAT_MPT_VIRT_START)]);
     if ( (l2_ro_mpt = alloc_xen_pagetable()) == NULL )
         goto nomem;
     compat_idle_pg_table_l2 = l2_ro_mpt;
     clear_page(l2_ro_mpt);
-    l3e_write(&l3_ro_mpt[l3_table_offset(HIRO_COMPAT_MPT_VIRT_START)

],
- l3e_from_paddr(__pa(l2_ro_mpt),

__PAGE_HYPERVISOR_RO));
     l2_ro_mpt += l2_table_offset(HIRO_COMPAT_MPT_VIRT_START);
Quoted 8 lines
Quoted 3 lines
     /* Allocate and map the compatibility mode machine-to-phys
table. */
     mpt_size = (mpt_size >> 1) + (1UL << (L2_PAGETABLE_SHIFT - 1));

The code around here, I am wondering if there is a reason to put it in this patch. If we bisect, we can end up in a commit which says the address range of compat is still there in config.h, but actually it's gone, so this probably belongs to the 2nd patch.

I could be done either way, I guess. Note though how patch 2's description starts with "Now that we don't properly hook things up into the page tables anymore". I.e. after this patch the address range continues to exist solely for calculation purposes.

Quoted 2 lines

Apart from that,

Thanks.

Quoted 3 lines

I would like to drop relevant map/unmap patches and replace them with the new clean-up ones (and place them at the beginning of the series), if there is no objection with that.

Depending on turnaround, I'd much rather see this go in before you re-post. But if you feel like making it part of your series, go ahead.

Jan

#14 lore raw

Re: [PATCH 1/2] x86: drop unnecessary page table walking in compat r/o M2P handling

Quoted 11 lines · Jan Beulich
Quoted 6 lines · Hongyan Xia

...
I would like to drop relevant map/unmap patches and replace them with
the new clean-up ones (and place them at the beginning of the series),
if there is no objection with that.

Depending on turnaround, I'd much rather see this go in before you re-post. But if you feel like making it part of your series, go ahead.

I actually also very much prefer to see those clean-up patches go in before I post the next revision, so please go ahead.

I will post the next revision minus patch 4 then to avoid conflict.

Hongyan

Reviewed-by on v1 1/2
#15 lore raw

Re: [PATCH 1/2] x86: drop unnecessary page table walking in compat r/o M2P handling

In reply to Jan Beulich
Quoted 7 lines · Jan Beulich

We have a global variable where the necessary L2 table is recorded; no need to inspect L4 and L3 tables (and this way a few less places will eventually need adjustment when we want to support 5-level page tables). Also avoid setting up the L3 entry, as the address range never gets used anyway (it'll be dropped altogether in a subsequent patch).

#16 lore raw

PATCH 2/2 x86: drop high compat r/o M2P table address range

In reply to Jan Beulich

Now that we don't properly hook things up into the page tables anymore we also don't need to set aside an address range. Drop it, using compat_idle_pg_table_l2[] simply (explicitly) from slot 0.

While doing the re-arrangement, which is accompanied by the dropping or replacing of some local variables, restrict the scopes of some further ones at the same time.

---
TBD: With the changed usage perhaps we want to also rename

     compat_idle_pg_table_l2[] (to e.g. compat_idle_l2_entries[])?
xen/arch/x86/mm.c +1 -2
@@ -1454,8 +1454,7 @@ static bool pae_xen_mappings_check(const
 void init_xen_pae_l2_slots(l2_pgentry_t *l2t, const struct domain *d)
 {
     memcpy(&l2t[COMPAT_L2_PAGETABLE_FIRST_XEN_SLOT(d)],
-           &compat_idle_pg_table_l2[
-               l2_table_offset(HIRO_COMPAT_MPT_VIRT_START)],
+           compat_idle_pg_table_l2,
            COMPAT_L2_PAGETABLE_XEN_SLOTS(d) * sizeof(*l2t));
 }
 
xen/arch/x86/x86_64/mm.c +17 -25

90 lines not on this page · raw message

xen/include/asm-x86/config.h +1 -4
@@ -215,11 +215,8 @@ extern unsigned char boot_edid_info[128]
 /* Slot 261: compatibility machine-to-phys conversion table (1GB). */
 #define RDWR_COMPAT_MPT_VIRT_START VMAP_VIRT_END
 #define RDWR_COMPAT_MPT_VIRT_END (RDWR_COMPAT_MPT_VIRT_START + GB(1))
-/* Slot 261: high read-only compat machine-to-phys conversion table (1GB). */
-#define HIRO_COMPAT_MPT_VIRT_START RDWR_COMPAT_MPT_VIRT_END
-#define HIRO_COMPAT_MPT_VIRT_END (HIRO_COMPAT_MPT_VIRT_START + GB(1))
 /* Slot 261: xen text, static data, bss, per-cpu stubs and executable fixmap (1GB). */
-#define XEN_VIRT_START          (HIRO_COMPAT_MPT_VIRT_END)
+#define XEN_VIRT_START          RDWR_COMPAT_MPT_VIRT_END
 #define XEN_VIRT_END            (XEN_VIRT_START + GB(1))
 
 #ifndef CONFIG_BIGMEM
Reviewed-by on v1 2/2
#17 lore raw

Re: [PATCH 2/2] x86: drop high compat r/o M2P table address range

Quoted 9 lines · Jan Beulich

Now that we don't properly hook things up into the page tables anymore we also don't need to set aside an address range. Drop it, using compat_idle_pg_table_l2[] simply (explicitly) from slot 0.

While doing the re-arrangement, which is accompanied by the dropping or replacing of some local variables, restrict the scopes of some further ones at the same time.

Quoted 3 lines

---
TBD: With the changed usage perhaps we want to also rename

     compat_idle_pg_table_l2[] (to e.g. compat_idle_l2_entries[])?

No opinion on this.

#18 lore raw

Re: [PATCH 0/2] x86: high compat r/o M2P table handling adjustments

In reply to Jan Beulich
Quoted 9 lines · Jan Beulich

While looking at "x86_64/mm: map and unmap page tables in destroy_compat_m2p_mapping" it occurred to me that the mappings changed there can be dropped altogether, as can the linear address range used for this. Note that both patches have only been lightly tested so far, I guess in particular the 2nd one may still have issues.

1: x86: drop unnecessary page table walking in compat r/o M2P handling 2: x86: drop high compat r/o M2P table address range

Just fyi - with the reviews I've got I'm intending to commit this later this week, unless I hear otherwise soon.

Jan

#19 lore raw

Re: [PATCH v2 4/5] x86_64/mm: map and unmap page tables in destroy_compat_m2p_mapping

In reply to Hongyan Xia
Quoted 9 lines · Hongyan Xia

From: Wei Liu <[email protected]>

---
Changed in v2:
- there is pretty much no point in keeping l3_ro_mpt mapped, just fetch the l3e instead, which also cleans up the code.

Even better, there's no need to do any mapping her at all. I've posted a pair of patches, the first of which can be seen as a suggested replacement for the one here.

Jan

#20 lore raw

PATCH v2 5/5 x86_64/mm: map and unmap page tables in destroy_m2p_mapping

In reply to Hongyan Xia

From: Wei Liu <[email protected]>

---
Changed in v2:
- no point in re-mapping l2t because it is exactly the same as l2_ro_mpt.
- point l2_ro_mpt to the entry instead of doing l2_table_offset() all the time.

---
 xen/arch/x86/x86_64/mm.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
xen/arch/x86/x86_64/mm.c +11 -6
index ae8f4dd0b9..ec5269e7fc 100644
@@ -268,7 +268,8 @@ static void destroy_m2p_mapping(struct mem_hotadd_info *info)
     unsigned long i, va, rwva;
     unsigned long smap = info->spfn, emap = info->epfn;
 
-    l3_ro_mpt = l4e_to_l3e(idle_pg_table[l4_table_offset(RO_MPT_VIRT_START)]);
+    l3_ro_mpt = map_l3t_from_l4e(
+                    idle_pg_table[l4_table_offset(RO_MPT_VIRT_START)]);
 
     /*
      * No need to clean m2p structure existing before the hotplug
@@ -290,26 +291,30 @@ static void destroy_m2p_mapping(struct mem_hotadd_info *info)
             continue;
         }
 
-        l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
-        if (!(l2e_get_flags(l2_ro_mpt[l2_table_offset(va)]) & _PAGE_PRESENT))
+        l2_ro_mpt = map_l2t_from_l3e(l3_ro_mpt[l3_table_offset(va)]) +
+                    l2_table_offset(va);
+        if ( !(l2e_get_flags(*l2_ro_mpt) & _PAGE_PRESENT) )
         {
             i = ( i & ~((1UL << (L2_PAGETABLE_SHIFT - 3)) - 1)) +
                     (1UL << (L2_PAGETABLE_SHIFT - 3)) ;
+            UNMAP_DOMAIN_PAGE(l2_ro_mpt);
             continue;
         }
 
-        pt_pfn = l2e_get_pfn(l2_ro_mpt[l2_table_offset(va)]);
+        pt_pfn = l2e_get_pfn(*l2_ro_mpt);
         if ( hotadd_mem_valid(pt_pfn, info) )
         {
             destroy_xen_mappings(rwva, rwva + (1UL << L2_PAGETABLE_SHIFT));
 
-            l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
-            l2e_write(&l2_ro_mpt[l2_table_offset(va)], l2e_empty());
+            l2e_write(l2_ro_mpt, l2e_empty());
         }
         i = ( i & ~((1UL << (L2_PAGETABLE_SHIFT - 3)) - 1)) +
               (1UL << (L2_PAGETABLE_SHIFT - 3));
+        UNMAP_DOMAIN_PAGE(l2_ro_mpt);
     }
 
+    UNMAP_DOMAIN_PAGE(l3_ro_mpt);
+
     destroy_compat_m2p_mapping(info);
 
     /* Brute-Force flush all TLB */
Signature · 1 line
2.24.1.AMZN
#21 lore raw

Re: [PATCH v2 5/5] x86_64/mm: map and unmap page tables in destroy_m2p_mapping

xen/arch/x86/x86_64/mm.c static void destroy_m2p_mapping(struct mem_hotadd_info *info) · line 295

Quoted 8 lines · Hongyan Xia
xen/arch/x86/x86_64/mm.c +2 -2
@@ -290,26 +291,30 @@ static void destroy_m2p_mapping(struct mem_hotadd_info *info)
             continue;
         }
 
-        l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
-        if (!(l2e_get_flags(l2_ro_mpt[l2_table_offset(va)]) & _PAGE_PRESENT))
+        l2_ro_mpt = map_l2t_from_l3e(l3_ro_mpt[l3_table_offset(va)]) +
+                    l2_table_offset(va);

Either the name of the variable should be changed or you shouldn't add in l2_table_offset(va) here. Personally I'd go with renaming to just pl2e.

Jan