Skip to contentSkip to conversation

xenbus: avoid stack overflow warning

PATCH In review Arnd Bergmann 3 participants 9 messages May 5, 2020 – May 6, 2020 Xen hypervisor interface

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xenbus: avoid stack overflow warning

The __xenbus_map_ring() function has two large arrays, 'map' and 'unmap' on its stack. When clang decides to inline it into its caller, xenbus_map_ring_valloc_hvm(), the total stack usage exceeds the warning limit for stack size on 32-bit architectures.

drivers/xen/xenbus/xenbus_client.c:592:12: error: stack frame size of 1104 bytes in function 'xenbus_map_ring_valloc_hvm' [-Werror,-Wframe-larger-than=]

As far as I can tell, other compilers don't inline it here, so we get no warning, but the stack usage is actually the same. It is possible for both arrays to use the same location on the stack, but the compiler cannot prove that this is safe because they get passed to external functions that may end up using them until they go out of scope.

Move the two arrays into separate basic blocks to limit the scope and force them to occupy less stack in total, regardless of the inlining decision.

---
 drivers/xen/xenbus/xenbus_client.c | 74 +++++++++++++++++-------------
 1 file changed, 41 insertions(+), 33 deletions(-)
drivers/xen/xenbus/xenbus_client.c +41 -33
index 040d2a43e8e3..23ca70378e36 100644
@@ -470,54 +470,62 @@ static int __xenbus_map_ring(struct xenbus_device *dev,
 			     unsigned int flags,
 			     bool *leaked)
 {
-	struct gnttab_map_grant_ref map[XENBUS_MAX_RING_GRANTS];
-	struct gnttab_unmap_grant_ref unmap[XENBUS_MAX_RING_GRANTS];
 	int i, j;
 	int err = GNTST_okay;
 
-	if (nr_grefs > XENBUS_MAX_RING_GRANTS)
-		return -EINVAL;
+	{
+		struct gnttab_map_grant_ref map[XENBUS_MAX_RING_GRANTS];
 
-	for (i = 0; i < nr_grefs; i++) {
-		memset(&map[i], 0, sizeof(map[i]));
-		gnttab_set_map_op(&map[i], addrs[i], flags, gnt_refs[i],
-				  dev->otherend_id);
-		handles[i] = INVALID_GRANT_HANDLE;
-	}
+		if (nr_grefs > XENBUS_MAX_RING_GRANTS)
+			return -EINVAL;
 
-	gnttab_batch_map(map, i);
+		for (i = 0; i < nr_grefs; i++) {
+			memset(&map[i], 0, sizeof(map[i]));
+			gnttab_set_map_op(&map[i], addrs[i], flags,
+					  gnt_refs[i], dev->otherend_id);
+			handles[i] = INVALID_GRANT_HANDLE;
+		}
+
+		gnttab_batch_map(map, i);
 
-	for (i = 0; i < nr_grefs; i++) {
-		if (map[i].status != GNTST_okay) {
-			err = map[i].status;
-			xenbus_dev_fatal(dev, map[i].status,
+		for (i = 0; i < nr_grefs; i++) {
+			if (map[i].status != GNTST_okay) {
+				err = map[i].status;
+				xenbus_dev_fatal(dev, map[i].status,
 					 "mapping in shared page %d from domain %d",
 					 gnt_refs[i], dev->otherend_id);
-			goto fail;
-		} else
-			handles[i] = map[i].handle;
+				goto fail;
+			} else
+				handles[i] = map[i].handle;
+		}
 	}
-
 	return GNTST_okay;
 
  fail:
-	for (i = j = 0; i < nr_grefs; i++) {
-		if (handles[i] != INVALID_GRANT_HANDLE) {
-			memset(&unmap[j], 0, sizeof(unmap[j]));
-			gnttab_set_unmap_op(&unmap[j], (phys_addr_t)addrs[i],
-					    GNTMAP_host_map, handles[i]);
-			j++;
+	{
+		struct gnttab_unmap_grant_ref unmap[XENBUS_MAX_RING_GRANTS];
+
+		for (i = j = 0; i < nr_grefs; i++) {
+			if (handles[i] != INVALID_GRANT_HANDLE) {
+				memset(&unmap[j], 0, sizeof(unmap[j]));
+				gnttab_set_unmap_op(&unmap[j],
+						    (phys_addr_t)addrs[i],
+						    GNTMAP_host_map,
+						    handles[i]);
+				j++;
+			}
 		}
-	}
 
-	if (HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref, unmap, j))
-		BUG();
+		if (HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
+					      unmap, j))
+			BUG();
 
-	*leaked = false;
-	for (i = 0; i < j; i++) {
-		if (unmap[i].status != GNTST_okay) {
-			*leaked = true;
-			break;
+		*leaked = false;
+		for (i = 0; i < j; i++) {
+			if (unmap[i].status != GNTST_okay) {
+				*leaked = true;
+				break;
+			}
 		}
 	}
 
Signature · 1 line
2.26.0
maintainer
#1 lore raw
Quoted 13 lines · Arnd Bergmann

The __xenbus_map_ring() function has two large arrays, 'map' and 'unmap' on its stack. When clang decides to inline it into its caller, xenbus_map_ring_valloc_hvm(), the total stack usage exceeds the warning limit for stack size on 32-bit architectures.

drivers/xen/xenbus/xenbus_client.c:592:12: error: stack frame size of 1104 bytes in function 'xenbus_map_ring_valloc_hvm' [-Werror,-Wframe-larger-than=]

As far as I can tell, other compilers don't inline it here, so we get no warning, but the stack usage is actually the same. It is possible for both arrays to use the same location on the stack, but the compiler cannot prove that this is safe because they get passed to external functions that may end up using them until they go out of scope.

Quoted 3 lines

Move the two arrays into separate basic blocks to limit the scope and force them to occupy less stack in total, regardless of the inlining decision.

Why don't you put both arrays into a union?

Juergen

Quoted 107 lines
---
  drivers/xen/xenbus/xenbus_client.c | 74 +++++++++++++++++-------------
  1 file changed, 41 insertions(+), 33 deletions(-)
drivers/xen/xenbus/xenbus_client.c +41 -33
index 040d2a43e8e3..23ca70378e36 100644
@@ -470,54 +470,62 @@ static int __xenbus_map_ring(struct xenbus_device *dev,
  			     unsigned int flags,
  			     bool *leaked)
  {
-	struct gnttab_map_grant_ref map[XENBUS_MAX_RING_GRANTS];
-	struct gnttab_unmap_grant_ref unmap[XENBUS_MAX_RING_GRANTS];
  	int i, j;
  	int err = GNTST_okay;
 
-	if (nr_grefs > XENBUS_MAX_RING_GRANTS)
-		return -EINVAL;
+	{
+		struct gnttab_map_grant_ref map[XENBUS_MAX_RING_GRANTS];
 
-	for (i = 0; i < nr_grefs; i++) {
-		memset(&map[i], 0, sizeof(map[i]));
-		gnttab_set_map_op(&map[i], addrs[i], flags, gnt_refs[i],
-				  dev->otherend_id);
-		handles[i] = INVALID_GRANT_HANDLE;
-	}
+		if (nr_grefs > XENBUS_MAX_RING_GRANTS)
+			return -EINVAL;
 
-	gnttab_batch_map(map, i);
+		for (i = 0; i < nr_grefs; i++) {
+			memset(&map[i], 0, sizeof(map[i]));
+			gnttab_set_map_op(&map[i], addrs[i], flags,
+					  gnt_refs[i], dev->otherend_id);
+			handles[i] = INVALID_GRANT_HANDLE;
+		}
+
+		gnttab_batch_map(map, i);
 
-	for (i = 0; i < nr_grefs; i++) {
-		if (map[i].status != GNTST_okay) {
-			err = map[i].status;
-			xenbus_dev_fatal(dev, map[i].status,
+		for (i = 0; i < nr_grefs; i++) {
+			if (map[i].status != GNTST_okay) {
+				err = map[i].status;
+				xenbus_dev_fatal(dev, map[i].status,
  					 "mapping in shared page %d from domain %d",
  					 gnt_refs[i], dev->otherend_id);
-			goto fail;
-		} else
-			handles[i] = map[i].handle;
+				goto fail;
+			} else
+				handles[i] = map[i].handle;
+		}
  	}
-
  	return GNTST_okay;
 
   fail:
-	for (i = j = 0; i < nr_grefs; i++) {
-		if (handles[i] != INVALID_GRANT_HANDLE) {
-			memset(&unmap[j], 0, sizeof(unmap[j]));
-			gnttab_set_unmap_op(&unmap[j], (phys_addr_t)addrs[i],
-					    GNTMAP_host_map, handles[i]);
-			j++;
+	{
+		struct gnttab_unmap_grant_ref unmap[XENBUS_MAX_RING_GRANTS];
+
+		for (i = j = 0; i < nr_grefs; i++) {
+			if (handles[i] != INVALID_GRANT_HANDLE) {
+				memset(&unmap[j], 0, sizeof(unmap[j]));
+				gnttab_set_unmap_op(&unmap[j],
+						    (phys_addr_t)addrs[i],
+						    GNTMAP_host_map,
+						    handles[i]);
+				j++;
+			}
  		}
-	}
 
-	if (HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref, unmap, j))
-		BUG();
+		if (HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref,
+					      unmap, j))
+			BUG();
 
-	*leaked = false;
-	for (i = 0; i < j; i++) {
-		if (unmap[i].status != GNTST_okay) {
-			*leaked = true;
-			break;
+		*leaked = false;
+		for (i = 0; i < j; i++) {
+			if (unmap[i].status != GNTST_okay) {
+				*leaked = true;
+				break;
+			}
  		}
  	}
 
#2 lore raw
Quoted 14 lines · Jürgen Groß
Quoted 13 lines · Arnd Bergmann

The __xenbus_map_ring() function has two large arrays, 'map' and 'unmap' on its stack. When clang decides to inline it into its caller, xenbus_map_ring_valloc_hvm(), the total stack usage exceeds the warning limit for stack size on 32-bit architectures.

drivers/xen/xenbus/xenbus_client.c:592:12: error: stack frame size of 1104 bytes in function 'xenbus_map_ring_valloc_hvm' [-Werror,-Wframe-larger-than=]

As far as I can tell, other compilers don't inline it here, so we get no warning, but the stack usage is actually the same. It is possible for both arrays to use the same location on the stack, but the compiler cannot prove that this is safe because they get passed to external functions that may end up using them until they go out of scope.

Quoted 5 lines
Quoted 3 lines

Move the two arrays into separate basic blocks to limit the scope and force them to occupy less stack in total, regardless of the inlining decision.

Why don't you put both arrays into a union?

I considered that as well, and don't really mind either way. I think it does get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Arnd

maintainer
#3 lore raw
Quoted 19 lines · Arnd Bergmann
Quoted 18 lines · Jürgen Groß
Quoted 16 lines · Arnd Bergmann

The __xenbus_map_ring() function has two large arrays, 'map' and 'unmap' on its stack. When clang decides to inline it into its caller, xenbus_map_ring_valloc_hvm(), the total stack usage exceeds the warning limit for stack size on 32-bit architectures.

drivers/xen/xenbus/xenbus_client.c:592:12: error: stack frame size of 1104 bytes in function 'xenbus_map_ring_valloc_hvm' [-Werror,-Wframe-larger-than=]

As far as I can tell, other compilers don't inline it here, so we get no warning, but the stack usage is actually the same. It is possible for both arrays to use the same location on the stack, but the compiler cannot prove that this is safe because they get passed to external functions that may end up using them until they go out of scope.

Move the two arrays into separate basic blocks to limit the scope and force them to occupy less stack in total, regardless of the inlining decision.

Quoted 5 lines
Quoted 1 line

Why don't you put both arrays into a union?

I considered that as well, and don't really mind either way. I think it does get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Hmm, thinking more about it I think the real clean solution would be to extend struct map_ring_valloc_hvm to cover the pv case, too, to add the map and unmap arrays (possibly as a union) to it and to allocate it dynamically instead of having it on the stack.

Would you be fine doing this?

Juergen

#4 lore raw
Quoted 31 lines · Jürgen Groß
Quoted 23 lines · Arnd Bergmann
Quoted 22 lines · Jürgen Groß
Quoted 20 lines · Arnd Bergmann

The __xenbus_map_ring() function has two large arrays, 'map' and 'unmap' on its stack. When clang decides to inline it into its caller, xenbus_map_ring_valloc_hvm(), the total stack usage exceeds the warning
limit for stack size on 32-bit architectures.

drivers/xen/xenbus/xenbus_client.c:592:12: error: stack frame size of 1104 bytes in function 'xenbus_map_ring_valloc_hvm' [-Werror,-Wframe-larger-than=]

As far as I can tell, other compilers don't inline it here, so we get no warning, but the stack usage is actually the same. It is possible for both arrays to use the same location on the stack, but the compiler
cannot prove that this is safe because they get passed to external functions that may end up using them until they go out of scope.

Move the two arrays into separate basic blocks to limit the scope and force them to occupy less stack in total, regardless of the inlining decision.

Quoted 6 lines
Quoted 1 line

Why don't you put both arrays into a union?

I considered that as well, and don't really mind either way. I think it does
get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Quoted 6 lines

Hmm, thinking more about it I think the real clean solution would be to extend struct map_ring_valloc_hvm to cover the pv case, too, to add the map and unmap arrays (possibly as a union) to it and to allocate it dynamically instead of having it on the stack.

Would you be fine doing this?

Another option might be to factor out/modify code from xenbus_unmap_ring() and call the resulting code from __xenbus_map_ring()'s fail path.

-boris

maintainer
#5 lore raw
Quoted 38 lines · Boris Ostrovsky
Quoted 36 lines · Jürgen Groß
Quoted 23 lines · Arnd Bergmann
Quoted 22 lines · Jürgen Groß
Quoted 20 lines · Arnd Bergmann

The __xenbus_map_ring() function has two large arrays, 'map' and 'unmap' on its stack. When clang decides to inline it into its caller, xenbus_map_ring_valloc_hvm(), the total stack usage exceeds the warning
limit for stack size on 32-bit architectures.

drivers/xen/xenbus/xenbus_client.c:592:12: error: stack frame size of 1104 bytes in function 'xenbus_map_ring_valloc_hvm' [-Werror,-Wframe-larger-than=]

As far as I can tell, other compilers don't inline it here, so we get no warning, but the stack usage is actually the same. It is possible for both arrays to use the same location on the stack, but the compiler
cannot prove that this is safe because they get passed to external functions that may end up using them until they go out of scope.

Move the two arrays into separate basic blocks to limit the scope and force them to occupy less stack in total, regardless of the inlining decision.

Quoted 6 lines
Quoted 1 line

Why don't you put both arrays into a union?

I considered that as well, and don't really mind either way. I think it does
get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Hmm, thinking more about it I think the real clean solution would be to extend struct map_ring_valloc_hvm to cover the pv case, too, to add the map and unmap arrays (possibly as a union) to it and to allocate it dynamically instead of having it on the stack.

Quoted 7 lines
Quoted 1 line

Would you be fine doing this?

Another option might be to factor out/modify code from xenbus_unmap_ring() and call the resulting code from __xenbus_map_ring()'s fail path.

This will still allocate large arrays on the stack. If we ever increase the max ring page order it will explode.

Juergen

#6 lore raw
In reply to Jürgen Groß
Quoted 8 lines · Jürgen Groß
Quoted 6 lines · Arnd Bergmann
Quoted 1 line · Jürgen Groß

On 05.05.20 16:15, Arnd Bergmann wrote:

I considered that as well, and don't really mind either way. I think it does get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Quoted 6 lines

Hmm, thinking more about it I think the real clean solution would be to extend struct map_ring_valloc_hvm to cover the pv case, too, to add the map and unmap arrays (possibly as a union) to it and to allocate it dynamically instead of having it on the stack.

Would you be fine doing this?

This is a little more complex than I'd want to do without doing any testing (and no, I don't want to do the testing either) ;-)

It does sound like a better approach though.

Arnd

maintainer
#7 lore raw
Quoted 14 lines · Arnd Bergmann
Quoted 13 lines · Jürgen Groß
Quoted 6 lines · Arnd Bergmann
Quoted 1 line · Jürgen Groß

On 05.05.20 16:15, Arnd Bergmann wrote:

I considered that as well, and don't really mind either way. I think it does get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Hmm, thinking more about it I think the real clean solution would be to extend struct map_ring_valloc_hvm to cover the pv case, too, to add the map and unmap arrays (possibly as a union) to it and to allocate it dynamically instead of having it on the stack.

Quoted 6 lines
Quoted 1 line

Would you be fine doing this?

This is a little more complex than I'd want to do without doing any testing (and no, I don't want to do the testing either) ;-)

It does sound like a better approach though.

I take this as you are fine with me writing the patch and adding you as "Reported-by:"?

Juergen

#8 lore raw
Quoted 18 lines · Jürgen Groß
Quoted 16 lines · Arnd Bergmann
Quoted 14 lines · Jürgen Groß
Quoted 6 lines · Arnd Bergmann
Quoted 1 line · Jürgen Groß

On 05.05.20 16:15, Arnd Bergmann wrote:

I considered that as well, and don't really mind either way. I think it does get a bit ugly whatever we do. If you prefer the union, I can respin the patch that way.

Hmm, thinking more about it I think the real clean solution would be to extend struct map_ring_valloc_hvm to cover the pv case, too, to add the map and unmap arrays (possibly as a union) to it and to allocate it dynamically instead of having it on the stack.

Would you be fine doing this?

Quoted 7 lines
Quoted 4 lines

This is a little more complex than I'd want to do without doing any testing (and no, I don't want to do the testing either) ;-)

It does sound like a better approach though.

I take this as you are fine with me writing the patch and adding you as "Reported-by:"?

Yes, definitely. Thanks!

Arnd