Skip to contentSkip to conversation

avoid unaligned access when accessing poll stack

PATCH In review Jes Sorensen 5 participants 14 messages Mar 31, 2006 – Apr 2, 2006 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 avoid unaligned access when accessing poll stack

Hi,

Patch 70674f95c0a2ea694d5c39f4e514f538a09be36f
[PATCH] Optimize select/poll by putting small data sets on the stack
resulted in the poll stack being 4-byte aligned on 64-bit architectures, causing misaligned accesses to elements in the array.

This patch fixes it by declaring the stack in terms of 'long' instead of 'char'.

Cheers,
Jes

Force alignment of poll stack to long to avoid unaligned access on 64 bit architectures.

---
 fs/select.c |    6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
fs/select.c +4 -2
Index: linux-2.6/fs/select.c
===================================================================
@@ -639,8 +639,10 @@
  	struct poll_list *walk;
 	struct fdtable *fdt;
 	int max_fdset;
-	/* Allocate small arguments on the stack to save memory and be faster */
-	char stack_pps[POLL_STACK_ALLOC];
+	/* Allocate small arguments on the stack to save memory and be 
+	   faster - use long to make sure the buffer is aligned properly
+	   on 64 bit archs to avoid unaligned access */
+	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 	struct poll_list *stack_pp = NULL;
 
 	/* Do a sanity check on nfds ... */
#1 lore raw
Quoted 9 lines · Jes Sorensen

Hi,

Patch 70674f95c0a2ea694d5c39f4e514f538a09be36f
[PATCH] Optimize select/poll by putting small data sets on the stack
resulted in the poll stack being 4-byte aligned on 64-bit architectures, causing misaligned accesses to elements in the array.

This patch fixes it by declaring the stack in terms of 'long' instead of 'char'.

You should do that for poll too then.

-Andi

#2 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Andi" == Andi Kleen <[email protected]> writes:

Quoted 9 lines · Jes Sorensen
Quoted 9 lines

Hi,

Patch 70674f95c0a2ea694d5c39f4e514f538a09be36f [PATCH] Optimize select/poll by putting small data sets on the stack resulted in the poll stack being 4-byte aligned on 64-bit architectures, causing misaligned accesses to elements in the array.

This patch fixes it by declaring the stack in terms of 'long' instead of 'char'.

Andi> You should do that for poll too then.

I assume you mean select().

Updated patch attached.

Jes

Force alignment of poll and select stacks to long to avoid unaligned access on 64 bit architectures.

---
 fs/select.c |    8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
fs/select.c +5 -3
Index: linux-2.6/fs/select.c
===================================================================
@@ -314,7 +314,7 @@
 	int ret, size, max_fdset;
 	struct fdtable *fdt;
 	/* Allocate small arguments on the stack to save memory and be faster */
-	char stack_fds[SELECT_STACK_ALLOC];
+	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 
 	ret = -EINVAL;
 	if (n < 0)
@@ -639,8 +639,10 @@
  	struct poll_list *walk;
 	struct fdtable *fdt;
 	int max_fdset;
-	/* Allocate small arguments on the stack to save memory and be faster */
-	char stack_pps[POLL_STACK_ALLOC];
+	/* Allocate small arguments on the stack to save memory and be
+	   faster - use long to make sure the buffer is aligned properly
+	   on 64 bit archs to avoid unaligned access */
+	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 	struct poll_list *stack_pp = NULL;
 
 	/* Do a sanity check on nfds ... */
#3 lore raw
Quoted 3 lines · Jes Sorensen

I assume you mean select().

Updated patch attached.

This fixes a few problems introduced by this patch.

* Fixes two warnings caused by mixing "char *" and "long *" pointers

* If SELECT_STACK_ALLOC is not a multiple of sizeof(long) then stack_fds[] would be less than SELECT_STACK_ALLOC bytes and could overflow later in the function. Fixed by simply rearranging the test later to work on sizeof(stack_fds)

Currently SELECT_STACK_ALLOC is 256 so this doesn't happen, but it's nasty to have things like this hidden in the code. What if later someone decides to change SELECT_STACK_ALLOC to 300?

* I also changed "size" to be unsigned since that makes more sense and is less prone to overflow bugs. I'm also a little scared by the "kmalloc(6 * size)" since that type of call is a classic multiply-overflow security hole (hence libc's calloc() API). However "size" is constrained by fdt->max_fdset so I think it isn't exploitable. The kernel doesn't have an overflow-safe API for kmalloc'ing arrays, does it?

Patch is vs current git HEAD. Only compile/boot tested.

fs/select.c +5 -4
index 071660f..bd9c7db 100644
@@ -311,7 +311,8 @@ static int core_sys_select(int n, fd_set
 {
 	fd_set_bits fds;
 	char *bits;
-	int ret, size, max_fdset;
+	int ret, max_fdset;
+	unsigned int size;
 	struct fdtable *fdt;
 	/* Allocate small arguments on the stack to save memory and be faster */
 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
@@ -335,8 +336,8 @@ static int core_sys_select(int n, fd_set
 	 */
 	ret = -ENOMEM;
 	size = FDS_BYTES(n);
-	if (6*size < SELECT_STACK_ALLOC)
-		bits = stack_fds;
+	if (size < sizeof(stack_fds) / 6)
+		bits = (char *) stack_fds;
 	else
 		bits = kmalloc(6 * size, GFP_KERNEL);
 	if (!bits)
@@ -373,7 +374,7 @@ static int core_sys_select(int n, fd_set
 		ret = -EFAULT;
 
 out:
-	if (bits != stack_fds)
+	if (bits != (char *) stack_fds)
 		kfree(bits);
 out_nofds:
 	return ret;
#4 lore raw
Quoted 6 lines · Mitchell Blank Jr

* I also changed "size" to be unsigned since that makes more sense and is less prone to overflow bugs. I'm also a little scared by the "kmalloc(6 * size)" since that type of call is a classic multiply-overflow security hole (hence libc's calloc() API). However "size" is constrained by fdt->max_fdset so I think it isn't exploitable. The kernel doesn't have an overflow-safe API for kmalloc'ing arrays, does it?

kcalloc. But it's slow since it memsets.

-Andi

#5 lore raw
In reply to Mitchell Blank Jr

Here's a slightly updated version of my patch: it changes the

if (size < sizeof(stack_fds) / 6)
to:

		if (size <= sizeof(stack_fds) / 6)

Otherwise this is exactly the same as the version I just posted. The old code had this problem too but before it only mattered if SELECT_STACK_ALLOC was a multiple of six.

fs/select.c +5 -4
index 071660f..c46d40c 100644
@@ -311,7 +311,8 @@ static int core_sys_select(int n, fd_set
 {
 	fd_set_bits fds;
 	char *bits;
-	int ret, size, max_fdset;
+	int ret, max_fdset;
+	unsigned int size;
 	struct fdtable *fdt;
 	/* Allocate small arguments on the stack to save memory and be faster */
 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
@@ -335,8 +336,8 @@ static int core_sys_select(int n, fd_set
 	 */
 	ret = -ENOMEM;
 	size = FDS_BYTES(n);
-	if (6*size < SELECT_STACK_ALLOC)
-		bits = stack_fds;
+	if (size <= sizeof(stack_fds) / 6)
+		bits = (char *) stack_fds;
 	else
 		bits = kmalloc(6 * size, GFP_KERNEL);
 	if (!bits)
@@ -373,7 +374,7 @@ static int core_sys_select(int n, fd_set
 		ret = -EFAULT;
 
 out:
-	if (bits != stack_fds)
+	if (bits != (char *) stack_fds)
 		kfree(bits);
 out_nofds:
 	return ret;
#6 lore raw
In reply to Jes Sorensen

fs/select.c line 646

Quoted 10 lines · Jes Sorensen

struct poll_list *walk;

fs/select.c +4 -2
 	struct fdtable *fdt;
 	int max_fdset;
-	/* Allocate small arguments on the stack to save memory and be faster */
-	char stack_pps[POLL_STACK_ALLOC];
+	/* Allocate small arguments on the stack to save memory and be 
+	   faster - use long to make sure the buffer is aligned properly
+	   on 64 bit archs to avoid unaligned access */
+	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 	struct poll_list *stack_pp = NULL;

struct poll_list stack_pps[POLL_STACK_ALLOC / sizeof(struct poll_list)];

is more readable, and probably gcc align it rightly?

Signature · 1 line
OGAWA Hirofumi <[email protected]>
#7 lore raw
Quoted 15 lines · OGAWA Hirofumi

fs/select.c line 646

Quoted 10 lines · Jes Sorensen

struct poll_list *walk;

fs/select.c +4 -2
 	struct fdtable *fdt;
 	int max_fdset;
-	/* Allocate small arguments on the stack to save memory and be faster */
-	char stack_pps[POLL_STACK_ALLOC];
+	/* Allocate small arguments on the stack to save memory and be 
+	   faster - use long to make sure the buffer is aligned properly
+	   on 64 bit archs to avoid unaligned access */
+	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 	struct poll_list *stack_pp = NULL;

struct poll_list stack_pps[POLL_STACK_ALLOC / sizeof(struct poll_list)];

Quoted 1 line

is more readable, and probably gcc align it rightly?

Yes, but it would be wrong

-Andi

#8 lore raw
Quoted 16 lines · Andi Kleen
Quoted 15 lines · OGAWA Hirofumi

fs/select.c line 646

Quoted 10 lines · Jes Sorensen

struct poll_list *walk;

fs/select.c +4 -2
 	struct fdtable *fdt;
 	int max_fdset;
-	/* Allocate small arguments on the stack to save memory and be faster */
-	char stack_pps[POLL_STACK_ALLOC];
+	/* Allocate small arguments on the stack to save memory and be 
+	   faster - use long to make sure the buffer is aligned properly
+	   on 64 bit archs to avoid unaligned access */
+	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 	struct poll_list *stack_pp = NULL;

struct poll_list stack_pps[POLL_STACK_ALLOC / sizeof(struct poll_list)];

Quoted 3 lines
Quoted 1 line

is more readable, and probably gcc align it rightly?

Yes, but it would be wrong

OK. So how about this?

	char stack_pps[POLL_STACK_ALLOC]
        	__attribute__((aligned (sizeof(struct poll_list))));
Signature · 1 line
OGAWA Hirofumi <[email protected]>
#9 lore raw
Quoted 5 lines · OGAWA Hirofumi

OK. So how about this?

	char stack_pps[POLL_STACK_ALLOC]
        	__attribute__((aligned (sizeof(struct poll_list))));

This would require much more alignment than really necessary. Probably you meant s/sizeof/alignof/. But Jes' patch is fine I think.

-Andi

#10 lore raw
Quoted 5 lines · Andi Kleen
Quoted 2 lines
	char stack_pps[POLL_STACK_ALLOC]
        	__attribute__((aligned (sizeof(struct poll_list))));

This would require much more alignment than really necessary. Probably you meant s/sizeof/alignof/. But Jes' patch is fine I think.

I read your this patch now, so, I agree. Sorry for noise.

Signature · 1 line
OGAWA Hirofumi <[email protected]>
#12 lore raw
In reply to Jes Sorensen
Quoted 4 lines · Jes Sorensen

[PATCH] Optimize select/poll by putting small data sets on the stack
resulted in the poll stack being 4-byte aligned on 64-bit architectures, causing misaligned accesses to elements in the array.

How come you noticed this but I did not?

#13 lore raw
Quoted 6 lines · Andrew Morton
Quoted 3 lines · Jes Sorensen

[PATCH] Optimize select/poll by putting small data sets on the stack
resulted in the poll stack being 4-byte aligned on 64-bit architectures, causing misaligned accesses to elements in the array.

How come you noticed this but I did not?

Heh,

The ia64 kernel spits out nifty little messages moaning about this since it goes to firmware emulation. They were all in the middle of the bootup messages making them a bit harder to notice as well.

Cheers,
Jes