Skip to contentSkip to conversation

Fix MemoryRegionSection alignment and comparison

PATCH Superseded Dr. David Alan Gilbert (git) 3 participants 8 messages Aug 13, 2019 – Aug 14, 2019

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover Fix MemoryRegionSection alignment and comparison

From: "Dr. David Alan Gilbert" <[email protected]>

This fixes a symptom I've seen on vhost-user on aarch64 where the daemon would be falsely notified of memory region changes that didn't exist.
The underlying problem was me memcmp'ing MemoryRegionSections even though they had padding in.

(Discovered while getting virtiofs working on aarch)

Dave

Dr. David Alan Gilbert (2):
memory: Align and add helper for comparing MemoryRegionSections
vhost: Fix memory region section comparison

 hw/virtio/vhost.c     |  9 +++++++--
 include/exec/memory.h | 14 +++++++++++++-
 2 files changed, 20 insertions(+), 3 deletions(-)
Signature · 1 line
2.21.0
#1 lore raw

PATCH 1/2 memory: Align and add helper for comparing MemoryRegionSections

From: "Dr. David Alan Gilbert" <[email protected]>

MemoryRegionSection includes an Int128 'size' field; on some platforms the compiler causes an alignment of this to a 128bit boundary, leaving 8 bytes of dead space. This deadspace can be filled with junk.

Move the size field to the top avoiding unnecsssary alignment and provide an 'eq' routine to safely compare MRS's.

---
 include/exec/memory.h | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
include/exec/memory.h +13 -1
index 606250172a..ce62e847bd 100644
@@ -487,15 +487,27 @@ static inline FlatView *address_space_to_flatview(AddressSpace *as)
  * @nonvolatile: this section is non-volatile
  */
 struct MemoryRegionSection {
+    Int128 size;
     MemoryRegion *mr;
     FlatView *fv;
     hwaddr offset_within_region;
-    Int128 size;
     hwaddr offset_within_address_space;
     bool readonly;
     bool nonvolatile;
 };
 
+static inline bool MemoryRegionSection_eq(MemoryRegionSection *a,
+                                          MemoryRegionSection *b)
+{
+    return a->mr == b->mr &&
+           a->fv == b->fv &&
+           a->offset_within_region == b->offset_within_region &&
+           a->offset_within_address_space == b->offset_within_address_space &&
+           int128_eq(a->size, b->size) &&
+           a->readonly == b->readonly &&
+           a->nonvolatile == b->nonvolatile;
+}
+
 /**
  * memory_region_init: Initialize a memory region
  *
Signature · 1 line
2.21.0
#2 lore raw

Re: [Qemu-devel] [PATCH 1/2] memory: Align and add helper for comparing MemoryRegionSections

Hi David,

Quoted 8 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

MemoryRegionSection includes an Int128 'size' field; on some platforms the compiler causes an alignment of this to a 128bit boundary, leaving 8 bytes of dead space. This deadspace can be filled with junk.

Move the size field to the top avoiding unnecsssary alignment

"unnecessary"

This is enough change to be in its own commit.

---

Quoted 1 line

and provide an 'eq' routine to safely compare MRS's.

This is another change, and should be squashed in the next patch IMO. Doesn't Clang warn about unused 'static inline' btw?

Quoted 40 lines
---
 include/exec/memory.h | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
include/exec/memory.h +13 -1
index 606250172a..ce62e847bd 100644
@@ -487,15 +487,27 @@ static inline FlatView *address_space_to_flatview(AddressSpace *as)
  * @nonvolatile: this section is non-volatile
  */
 struct MemoryRegionSection {
+    Int128 size;
     MemoryRegion *mr;
     FlatView *fv;
     hwaddr offset_within_region;
-    Int128 size;
     hwaddr offset_within_address_space;
     bool readonly;
     bool nonvolatile;
 };
 
+static inline bool MemoryRegionSection_eq(MemoryRegionSection *a,
+                                          MemoryRegionSection *b)
+{
+    return a->mr == b->mr &&
+           a->fv == b->fv &&
+           a->offset_within_region == b->offset_within_region &&
+           a->offset_within_address_space == b->offset_within_address_space &&
+           int128_eq(a->size, b->size) &&
+           a->readonly == b->readonly &&
+           a->nonvolatile == b->nonvolatile;
+}
+
 /**
  * memory_region_init: Initialize a memory region
  *
#3 lore raw

Re: [Qemu-devel] [PATCH 1/2] memory: Align and add helper for comparing MemoryRegionSections

Quoted 5 lines · * Philippe Mathieu-Daudé ([email protected]) wrote:

Hi David,

Quoted 2 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Quoted 8 lines
Quoted 6 lines

MemoryRegionSection includes an Int128 'size' field; on some platforms the compiler causes an alignment of this to a 128bit boundary, leaving 8 bytes of dead space. This deadspace can be filled with junk.

Move the size field to the top avoiding unnecsssary alignment

"unnecessary"

Oops thanks.

Quoted 7 lines

This is enough change to be in its own commit.

---

Quoted 1 line

and provide an 'eq' routine to safely compare MRS's.

This is another change, and should be squashed in the next patch IMO.

OK, what I'll do is I'll split this one into two ; I feel better having the extra function here separate from the next commit.

Quoted 1 line

Doesn't Clang warn about unused 'static inline' btw?

I was using gcc; but we seem to have loads of static inline's - what would make this one different?

Dave

Quoted 24 lines
Quoted 24 lines
---
 include/exec/memory.h | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
include/exec/memory.h +1 -1
index 606250172a..ce62e847bd 100644
@@ -487,15 +487,27 @@ static inline FlatView *address_space_to_flatview(AddressSpace *as)
  * @nonvolatile: this section is non-volatile
  */
 struct MemoryRegionSection {
+    Int128 size;
     MemoryRegion *mr;
     FlatView *fv;
     hwaddr offset_within_region;
-    Int128 size;
     hwaddr offset_within_address_space;
     bool readonly;
     bool nonvolatile;
 };
 
Quoted 16 lines
Quoted 6 lines
include/exec/memory.h +5 -0
+static inline bool MemoryRegionSection_eq(MemoryRegionSection *a,
+                                          MemoryRegionSection *b)
+{
+    return a->mr == b->mr &&
+           a->fv == b->fv &&

include/exec/memory.h static inline FlatView *address_space_to_flatview(AddressSpace *as) · line 513

Quoted 10 lines
include/exec/memory.h +7 -0
+           a->offset_within_region == b->offset_within_region &&
+           a->offset_within_address_space == b->offset_within_address_space &&
+           int128_eq(a->size, b->size) &&
+           a->readonly == b->readonly &&
+           a->nonvolatile == b->nonvolatile;
+}
+
 /**
  * memory_region_init: Initialize a memory region
  *

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#4 lore raw

Re: [Qemu-devel] [PATCH 1/2] memory: Align and add helper for comparing MemoryRegionSections

Quoted 19 lines · Dr. David Alan Gilbert
Quoted 5 lines · * Philippe Mathieu-Daudé ([email protected]) wrote:

Hi David,

Quoted 2 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Quoted 8 lines
Quoted 6 lines

MemoryRegionSection includes an Int128 'size' field; on some platforms the compiler causes an alignment of this to a 128bit boundary, leaving 8 bytes of dead space. This deadspace can be filled with junk.

Move the size field to the top avoiding unnecsssary alignment

"unnecessary"

Oops thanks.

Quoted 2 lines

This is enough change to be in its own commit.

Quoted 8 lines
Quoted 5 lines

---

Quoted 1 line

and provide an 'eq' routine to safely compare MRS's.

This is another change, and should be squashed in the next patch IMO.

OK, what I'll do is I'll split this one into two ; I feel better having the extra function here separate from the next commit.

Thanks, feel free to add to both:

Quoted 4 lines
Quoted 1 line

Doesn't Clang warn about unused 'static inline' btw?

I was using gcc; but we seem to have loads of static inline's - what would make this one different?

I guess I was confused with 'static (no-inline)' :)

Quoted 43 lines
Quoted 24 lines
Quoted 24 lines
---
 include/exec/memory.h | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
include/exec/memory.h +1 -1
index 606250172a..ce62e847bd 100644
@@ -487,15 +487,27 @@ static inline FlatView *address_space_to_flatview(AddressSpace *as)
  * @nonvolatile: this section is non-volatile
  */
 struct MemoryRegionSection {
+    Int128 size;
     MemoryRegion *mr;
     FlatView *fv;
     hwaddr offset_within_region;
-    Int128 size;
     hwaddr offset_within_address_space;
     bool readonly;
     bool nonvolatile;
 };
 
Quoted 16 lines
Quoted 6 lines
include/exec/memory.h +5 -0
+static inline bool MemoryRegionSection_eq(MemoryRegionSection *a,
+                                          MemoryRegionSection *b)
+{
+    return a->mr == b->mr &&
+           a->fv == b->fv &&

include/exec/memory.h static inline FlatView *address_space_to_flatview(AddressSpace *as) · line 513

Quoted 10 lines
include/exec/memory.h +7 -0
+           a->offset_within_region == b->offset_within_region &&
+           a->offset_within_address_space == b->offset_within_address_space &&
+           int128_eq(a->size, b->size) &&
+           a->readonly == b->readonly &&
+           a->nonvolatile == b->nonvolatile;
+}
+
 /**
  * memory_region_init: Initialize a memory region
  *

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#5 lore raw

Re: [Qemu-devel] [PATCH 1/2] memory: Align and add helper for comparing MemoryRegionSections

Quoted 24 lines · * Philippe Mathieu-Daudé ([email protected]) wrote:
Quoted 23 lines · Dr. David Alan Gilbert
Quoted 5 lines · * Philippe Mathieu-Daudé ([email protected]) wrote:

Hi David,

Quoted 2 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Quoted 8 lines
Quoted 6 lines

MemoryRegionSection includes an Int128 'size' field; on some platforms the compiler causes an alignment of this to a 128bit boundary, leaving 8 bytes of dead space. This deadspace can be filled with junk.

Move the size field to the top avoiding unnecsssary alignment

"unnecessary"

Oops thanks.

Quoted 6 lines

This is enough change to be in its own commit.

---

Quoted 1 line

and provide an 'eq' routine to safely compare MRS's.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

This is another change, and should be squashed in the next patch IMO.

OK, what I'll do is I'll split this one into two ; I feel better having the extra function here separate from the next commit.

Thanks, feel free to add to both:

Thanks!

Quoted 50 lines
Quoted 4 lines
Quoted 1 line

Doesn't Clang warn about unused 'static inline' btw?

I was using gcc; but we seem to have loads of static inline's - what would make this one different?

I guess I was confused with 'static (no-inline)' :)

Quoted 43 lines
Quoted 24 lines
Quoted 24 lines
---
 include/exec/memory.h | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
include/exec/memory.h +1 -1
index 606250172a..ce62e847bd 100644
@@ -487,15 +487,27 @@ static inline FlatView *address_space_to_flatview(AddressSpace *as)
  * @nonvolatile: this section is non-volatile
  */
 struct MemoryRegionSection {
+    Int128 size;
     MemoryRegion *mr;
     FlatView *fv;
     hwaddr offset_within_region;
-    Int128 size;
     hwaddr offset_within_address_space;
     bool readonly;
     bool nonvolatile;
 };
 
Quoted 16 lines
Quoted 6 lines
include/exec/memory.h +5 -0
+static inline bool MemoryRegionSection_eq(MemoryRegionSection *a,
+                                          MemoryRegionSection *b)
+{
+    return a->mr == b->mr &&
+           a->fv == b->fv &&

include/exec/memory.h static inline FlatView *address_space_to_flatview(AddressSpace *as) · line 513

Quoted 10 lines
include/exec/memory.h +7 -0
+           a->offset_within_region == b->offset_within_region &&
+           a->offset_within_address_space == b->offset_within_address_space &&
+           int128_eq(a->size, b->size) &&
+           a->readonly == b->readonly &&
+           a->nonvolatile == b->nonvolatile;
+}
+
 /**
  * memory_region_init: Initialize a memory region
  *

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#6 lore raw

PATCH 2/2 vhost: Fix memory region section comparison

In reply to Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Using memcmp to compare structures wasn't safe,
as I found out on ARM when I was getting false miscompares.

Use the helper function for comparing the MRSs.

Fixes ade6d081fc33948e56e6
---
 hw/virtio/vhost.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)
hw/virtio/vhost.c +7 -2
index bc899fc60e..2ef4bc720f 100644
@@ -451,8 +451,13 @@ static void vhost_commit(MemoryListener *listener)
         changed = true;
     } else {
         /* Same size, lets check the contents */
-        changed = n_old_sections && memcmp(dev->mem_sections, old_sections,
-                         n_old_sections * sizeof(old_sections[0])) != 0;
+        for (int i = 0; i < n_old_sections; i++) {
+            if (!MemoryRegionSection_eq(&old_sections[i],
+                                        &dev->mem_sections[i])) {
+                changed = true;
+                break;
+            }
+        }
     }
 
     trace_vhost_commit(dev->started, changed);
Signature · 1 line
2.21.0
Acked-by on the v1 series
#7 lore raw
In reply to Dr. David Alan Gilbert (git)
Quoted 13 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

This fixes a symptom I've seen on vhost-user on aarch64 where the daemon would be falsely notified of memory region changes that didn't exist.
The underlying problem was me memcmp'ing MemoryRegionSections even though they had padding in.

(Discovered while getting virtiofs working on aarch)

Dave

Quoted 8 lines

Dr. David Alan Gilbert (2):
memory: Align and add helper for comparing MemoryRegionSections
vhost: Fix memory region section comparison

 hw/virtio/vhost.c     |  9 +++++++--
 include/exec/memory.h | 14 +++++++++++++-
 2 files changed, 20 insertions(+), 3 deletions(-)

I think it can be merged through the vhost tree.

Paolo