Skip to contentSkip to conversation

Fix domheap structure allocation when NUMA=on

PATCH In review Yang, Xiaowei 3 participants 9 messages Mar 20, 2009 – Mar 24, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix domheap structure allocation when NUMA=on

DIRECTMAP_VIRT_END can't be passed to virt_to_mfn(), as it's just beyond direct map boundary and triggers ASSERT very early at boot time.

xen/common/page_alloc.c +1 -1
diff -r 78ed606cb1ab xen/common/page_alloc.c
@@ -293,7 +297,7 @@ static unsigned long init_node_heap(int 
     }
 #ifdef DIRECTMAP_VIRT_END
     else if ( nr >= needed &&
-              mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END) )
+              mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE) )
     {
         _heap[node] = mfn_to_virt(mfn);
         avail[node] = mfn_to_virt(mfn + needed) - sizeof(**avail) * NR_ZONES;
xen/include/asm-x86/x86_64/page.h +1 -1
diff -r 78ed606cb1ab xen/include/asm-x86/x86_64/page.h
@@ -40,7 +40,7 @@ static inline unsigned long __virt_to_ma
     ASSERT(va >= XEN_VIRT_START);
     ASSERT(va < DIRECTMAP_VIRT_END);
     ASSERT((va < XEN_VIRT_END) || (va >= DIRECTMAP_VIRT_START));
-    if ( va > DIRECTMAP_VIRT_START )
+    if ( va >= DIRECTMAP_VIRT_START )
         return va - DIRECTMAP_VIRT_START;
     return va - XEN_VIRT_START + xen_phys_start;
 }

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#1 lore raw
Quoted 3 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 20.03.09 06:05 >>>

DIRECTMAP_VIRT_END can't be passed to virt_to_mfn(), as it's just beyond direct map boundary and triggers ASSERT very early at boot time.

While I agree to the analysis, I would think that this

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE) )

should rather be

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - 1) + 1 )

Jan

#2 lore raw
Quoted 11 lines · Jan Beulich
Quoted 2 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 20.03.09 06:05 >>>

DIRECTMAP_VIRT_END can't be passed to virt_to_mfn(), as it's just beyond direct map boundary and triggers ASSERT very early at boot time.

While I agree to the analysis, I would think that this

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE) )

should rather be

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - 1) + 1 )

virt_to_mfn(DIRECTMAP_VIRT_END - 1) is equal to virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE). Why +1? We use '<=' here.

Thanks,
xiaowei

#3 lore raw
Quoted 8 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 20.03.09 09:34 >>>

Quoted 6 lines · Jan Beulich
Quoted 3 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 20.03.09 06:05 >>>

DIRECTMAP_VIRT_END can't be passed to virt_to_mfn(), as it's just beyond direct map boundary and triggers ASSERT very early at boot time.

While I agree to the analysis, I would think that this

Quoted 7 lines
Quoted 6 lines

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE) )

should rather be

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - 1) + 1 )

virt_to_mfn(DIRECTMAP_VIRT_END - 1) is equal to

Depending on whether DIRECTMAP_VIRT_END is the last byte or the first following byte. Using "- 1" avoids such a dependency.

Quoted 1 line

virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE). Why +1? We use '<=' here.

Because on the left side of the comparison we also calculate the first following mfn, not the last included one.

Jan

#4 lore raw
Quoted 9 lines · Jan Beulich

virt_to_mfn(DIRECTMAP_VIRT_END - 1) is equal to

Depending on whether DIRECTMAP_VIRT_END is the last byte or the first following byte. Using "- 1" avoids such a dependency.

Quoted 1 line

virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE). Why +1? We use '<=' here.

Because on the left side of the comparison we also calculate the first following mfn, not the last included one.

Since DIRECTMAP_VIRT_END is always the following byte not the last byte, it sounds like you think the page_alloc.c chunk is not needed at all. I'll just check in the header fix for now.

-- Keir

#5 lore raw
Quoted 8 lines
Quoted 1 line
Quoted 1 line

Keir Fraser <[email protected]> 20.03.09 09:56 >>>

Quoted 5 lines · Jan Beulich

virt_to_mfn(DIRECTMAP_VIRT_END - 1) is equal to

Depending on whether DIRECTMAP_VIRT_END is the last byte or the first following byte. Using "- 1" avoids such a dependency.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE). Why +1? We use '<=' here.

Because on the left side of the comparison we also calculate the first following mfn, not the last included one.

Since DIRECTMAP_VIRT_END is always the following byte not the last byte, it sounds like you think the page_alloc.c chunk is not needed at all. I'll just check in the header fix for now.

No, that chunk is necessary afaics - as I said I agree to Xiaowei's diagnosis that virt_to_mfn() should not be used on DIRECTMAP_VIRT_END alone. Subtracting 1 from it, and then adding back 1 to the result is indeed what I think is the right thing to do here.

Jan

#6 lore raw
Quoted 8 lines · Jan Beulich
Quoted 3 lines

Since DIRECTMAP_VIRT_END is always the following byte not the last byte, it sounds like you think the page_alloc.c chunk is not needed at all. I'll just check in the header fix for now.

No, that chunk is necessary afaics - as I said I agree to Xiaowei's diagnosis that virt_to_mfn() should not be used on DIRECTMAP_VIRT_END alone. Subtracting 1 from it, and then adding back 1 to the result is indeed what I think is the right thing to do here.

Ah... virt_to_mfn() has a discontinuity at DIRECTMAP_VIRT_END. I get you. I'll apply that chunk as a separate changeset then.

-- Keir

#7 lore raw
In reply to Jan Beulich
Quoted 14 lines · Jan Beulich
Quoted 6 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 20.03.09 09:34 >>>

Quoted 4 lines · Jan Beulich
Quoted 2 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 20.03.09 06:05 >>>

DIRECTMAP_VIRT_END can't be passed to virt_to_mfn(), as it's just beyond direct map boundary and triggers ASSERT very early at boot time.

While I agree to the analysis, I would think that this

Quoted 7 lines
Quoted 6 lines

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE) )

should rather be

+ mfn + needed <= virt_to_mfn(DIRECTMAP_VIRT_END - 1) + 1 )

virt_to_mfn(DIRECTMAP_VIRT_END - 1) is equal to

Quoted 8 lines

Depending on whether DIRECTMAP_VIRT_END is the last byte or the first following byte. Using "- 1" avoids such a dependency.

Quoted 1 line

virt_to_mfn(DIRECTMAP_VIRT_END - PAGE_SIZE). Why +1? We use '<=' here.

Because on the left side of the comparison we also calculate the first following mfn, not the last included one.

Jan,
Thanks for the clarification!

Now I found another potential issue. Since mfn+needed could be equal to virt_to_mfn(DIRECTMAP_VIRT_END-1)+1, it can't be passed to mfn_to_virt() directly - the valid range is [0,DIRECT_MAP_BYTES>>PAGE_SHIFT).

Here is the patch to fix it. More assert and 32 bit counterpart are added.

Thanks,
xiaowei

xen/common/page_alloc.c +2 -1
diff -r 0477f9061c8a xen/common/page_alloc.c
@@ -302,7 +302,8 @@ static unsigned long init_node_heap(int 
               (mfn + needed) <= (virt_to_mfn(DIRECTMAP_VIRT_END - 1) + 1) )
     {
         _heap[node] = mfn_to_virt(mfn);
-        avail[node] = mfn_to_virt(mfn + needed) - sizeof(**avail) * NR_ZONES;
+        avail[node] = mfn_to_virt(mfn + needed - 1) +
+                      PAGE_SIZE - sizeof(**avail) * NR_ZONES;
     }
 #endif
     else if ( get_order_from_bytes(sizeof(**_heap)) ==
xen/include/asm-x86/x86_32/page.h +16 -3
diff -r 0477f9061c8a xen/include/asm-x86/x86_32/page.h
@@ -27,9 +27,6 @@
 #define __PAGE_OFFSET           (0xFF000000)
 #define __XEN_VIRT_START        __PAGE_OFFSET
 
-#define virt_to_maddr(va) ((unsigned long)(va)-DIRECTMAP_VIRT_START)
-#define maddr_to_virt(ma) ((void *)((unsigned long)(ma)+DIRECTMAP_VIRT_START))
-
 #define VADDR_BITS              32
 #define VADDR_MASK              (~0UL)
 
@@ -43,6 +40,22 @@
 
 #include <xen/config.h>
 #include <asm/types.h>
+
+static inline unsigned long __virt_to_maddr(unsigned long va)
+{
+    ASSERT(va >= DIRECTMAP_VIRT_START && va < DIRECTMAP_VIRT_END);
+    return va - DIRECTMAP_VIRT_START;
+}
+#define virt_to_maddr(va)       \
+    (__virt_to_maddr((unsigned long)(va)))
+
+static inline void *__maddr_to_virt(unsigned long ma)
+{
+    ASSERT(ma < DIRECTMAP_VIRT_END - DIRECTMAP_VIRT_START);
+    return (void *)(ma + DIRECTMAP_VIRT_START);
+}
+#define maddr_to_virt(ma)       \
+    (__maddr_to_virt((unsigned long)(ma)))
 
 /* read access (should only be used for debug printk's) */
 typedef u64 intpte_t;
xen/include/asm-x86/x86_64/page.h +7 -1
diff -r 0477f9061c8a xen/include/asm-x86/x86_64/page.h
@@ -46,8 +46,14 @@ static inline unsigned long __virt_to_ma
 }
 #define virt_to_maddr(va)       \
     (__virt_to_maddr((unsigned long)(va)))
+
+static inline void *__maddr_to_virt(unsigned long ma)
+{
+    ASSERT(ma < DIRECTMAP_VIRT_END - DIRECTMAP_VIRT_START);
+    return (void *)(ma + DIRECTMAP_VIRT_START);
+}
 #define maddr_to_virt(ma)       \
-    ((void *)((unsigned long)(ma)+DIRECTMAP_VIRT_START))
+    (__maddr_to_virt((unsigned long)(ma)))
 
 /* read access (should only be used for debug printk's) */
 typedef u64 intpte_t;

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#8 lore raw
Quoted 4 lines
Quoted 1 line
Quoted 1 line

"Yang, Xiaowei" <[email protected]> 24.03.09 11:48 >>>

Now I found another potential issue. Since mfn+needed could be equal to virt_to_mfn(DIRECTMAP_VIRT_END-1)+1, it can't be passed to mfn_to_virt() directly - the valid range is [0,DIRECT_MAP_BYTES>>PAGE_SHIFT).

Oh, indeed. Thanks for noticing and fixing.

Jan