Skip to contentSkip to conversation

net: filter: export pkt_type_offset() helper

PATCH In review Denis Kirjanov 6 participants 21 messages Sep 3, 2014 – Sep 12, 2014 2 messages not in the archive

Subsystems (9)
Networking Network drivers MIPS BPF +5 more subsystems
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 net: filter: export pkt_type_offset() helper

Currently we have 2 pkt_type_offset functions doing the same thing and spread across the architecture files. Let's use the generic helper routine.

Cc Markos Chandras <[email protected]>
Cc Martin Schwidefsky <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc Alexei Starovoitov <[email protected]>

V2: Initialize the pkt_type_offset through the pure_initcall()

---
 arch/mips/net/bpf_jit.c      | 27 ++-------------------------
 arch/s390/net/bpf_jit_comp.c | 31 -------------------------------
 include/linux/filter.h       |  2 ++
 net/core/filter.c            | 16 +++++++++++++---
 4 files changed, 17 insertions(+), 59 deletions(-)
arch/mips/net/bpf_jit.c +2 -25
index 05a5661..bb9b53f 100644
@@ -765,27 +765,6 @@ static u64 jit_get_skb_w(struct sk_buff *skb, unsigned offset)
 	return (u64)err << 32 | ntohl(ret);
 }
 
-#ifdef __BIG_ENDIAN_BITFIELD
-#define PKT_TYPE_MAX	(7 << 5)
-#else
-#define PKT_TYPE_MAX	7
-#endif
-static int pkt_type_offset(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	u8 *ct = (u8 *)&skb_probe;
-	unsigned int off;
-
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
-			return off;
-	}
-	pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be found\n");
-	return -1;
-}
-
 static int build_body(struct jit_ctx *ctx)
 {
 	void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w};
@@ -1332,11 +1311,9 @@ jmp_cmp:
 		case BPF_ANC | SKF_AD_PKTTYPE:
 			ctx->flags |= SEEN_SKB;
 
-			off = pkt_type_offset();
-
-			if (off < 0)
+			if (pkt_type_offset < 0)
 				return -1;
-			emit_load_byte(r_tmp, r_skb, off, ctx);
+			emit_load_byte(r_tmp, r_skb, pkt_type_offset, ctx);
 			/* Keep only the last 3 bits */
 			emit_andi(r_A, r_tmp, PKT_TYPE_MAX, ctx);
 #ifdef __BIG_ENDIAN_BITFIELD
arch/s390/net/bpf_jit_comp.c +0 -31
index 61e45b7..caa0cab 100644
@@ -223,37 +223,6 @@ static void bpf_jit_epilogue(struct bpf_jit *jit)
 	EMIT2(0x07fe);
 }
 
-/* Helper to find the offset of pkt_type in sk_buff
- * Make sure its still a 3bit field starting at the MSBs within a byte.
- */
-#define PKT_TYPE_MAX 0xe0
-static int pkt_type_offset;
-
-static int __init bpf_pkt_type_offset_init(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	char *ct = (char *)&skb_probe;
-	int off;
-
-	pkt_type_offset = -1;
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (!ct[off])
-			continue;
-		if (ct[off] == PKT_TYPE_MAX)
-			pkt_type_offset = off;
-		else {
-			/* Found non matching bit pattern, fix needed. */
-			WARN_ON_ONCE(1);
-			pkt_type_offset = -1;
-			return -1;
-		}
-	}
-	return 0;
-}
-device_initcall(bpf_pkt_type_offset_init);
-
 /*
  * make sure we dont leak kernel information to user
  */
include/linux/filter.h +2 -0
index a5227ab..6814b54 100644
@@ -424,6 +424,8 @@ static inline void *bpf_load_pointer(const struct sk_buff *skb, int k,
 	return bpf_internal_load_pointer_neg_helper(skb, k, size);
 }
 
+extern __read_mostly int pkt_type_offset;
+
 #ifdef CONFIG_BPF_JIT
 #include <stdarg.h>
 #include <linux/linkage.h>
net/core/filter.c +13 -3
index d814b8a..edd17f1 100644
@@ -96,21 +96,29 @@ EXPORT_SYMBOL(sk_filter);
 #else
 #define PKT_TYPE_MAX	7
 #endif
-static unsigned int pkt_type_offset(void)
+
+__read_mostly int pkt_type_offset;
+
+static int __init init_pkt_type_offset(void)
 {
 	struct sk_buff skb_probe = { .pkt_type = ~0, };
 	u8 *ct = (u8 *) &skb_probe;
 	unsigned int off;
 
+	pkt_type_offset = -1;
 	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
+		if (ct[off] == PKT_TYPE_MAX) {
+			pkt_type_offset = off;
 			return off;
+		}
 	}
 
 	pr_err_once("Please fix %s, as pkt_type couldn't be found!\n", __func__);
 	return -1;
 }
 
+pure_initcall(init_pkt_type_offset);
+
 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
 {
 	return __skb_get_poff((struct sk_buff *)(unsigned long) ctx);
@@ -190,8 +198,10 @@ static bool convert_bpf_extensions(struct sock_filter *fp,
 		break;
 
 	case SKF_AD_OFF + SKF_AD_PKTTYPE:
+		if (pkt_type_offset < 0)
+			return false;
 		*insn = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX,
-				    pkt_type_offset());
+				    pkt_type_offset);
 		if (insn->off < 0)
 			return false;
 		insn++;
Signature · 1 line
2.0.0
#1 lore raw
Quoted 111 lines · Denis Kirjanov

Currently we have 2 pkt_type_offset functions doing the same thing and spread across the architecture files. Let's use the generic helper routine.

Cc Markos Chandras <[email protected]>
Cc Martin Schwidefsky <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc Alexei Starovoitov <[email protected]>

V2: Initialize the pkt_type_offset through the pure_initcall()

---
  arch/mips/net/bpf_jit.c      | 27 ++-------------------------
  arch/s390/net/bpf_jit_comp.c | 31 -------------------------------
  include/linux/filter.h       |  2 ++
  net/core/filter.c            | 16 +++++++++++++---
  4 files changed, 17 insertions(+), 59 deletions(-)
arch/mips/net/bpf_jit.c +2 -25
index 05a5661..bb9b53f 100644
@@ -765,27 +765,6 @@ static u64 jit_get_skb_w(struct sk_buff *skb, unsigned offset)
  	return (u64)err << 32 | ntohl(ret);
  }
 
-#ifdef __BIG_ENDIAN_BITFIELD
-#define PKT_TYPE_MAX	(7 << 5)
-#else
-#define PKT_TYPE_MAX	7
-#endif
-static int pkt_type_offset(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	u8 *ct = (u8 *)&skb_probe;
-	unsigned int off;
-
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
-			return off;
-	}
-	pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be found\n");
-	return -1;
-}
-
  static int build_body(struct jit_ctx *ctx)
  {
  	void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w};
@@ -1332,11 +1311,9 @@ jmp_cmp:
  		case BPF_ANC | SKF_AD_PKTTYPE:
  			ctx->flags |= SEEN_SKB;
 
-			off = pkt_type_offset();
-
-			if (off < 0)
+			if (pkt_type_offset < 0)
  				return -1;
-			emit_load_byte(r_tmp, r_skb, off, ctx);
+			emit_load_byte(r_tmp, r_skb, pkt_type_offset, ctx);
  			/* Keep only the last 3 bits */
  			emit_andi(r_A, r_tmp, PKT_TYPE_MAX, ctx);
  #ifdef __BIG_ENDIAN_BITFIELD
arch/s390/net/bpf_jit_comp.c +0 -31
index 61e45b7..caa0cab 100644
@@ -223,37 +223,6 @@ static void bpf_jit_epilogue(struct bpf_jit *jit)
  	EMIT2(0x07fe);
  }
 
-/* Helper to find the offset of pkt_type in sk_buff
- * Make sure its still a 3bit field starting at the MSBs within a byte.
- */
-#define PKT_TYPE_MAX 0xe0
-static int pkt_type_offset;
-
-static int __init bpf_pkt_type_offset_init(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	char *ct = (char *)&skb_probe;
-	int off;
-
-	pkt_type_offset = -1;
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (!ct[off])
-			continue;
-		if (ct[off] == PKT_TYPE_MAX)
-			pkt_type_offset = off;
-		else {
-			/* Found non matching bit pattern, fix needed. */
-			WARN_ON_ONCE(1);
-			pkt_type_offset = -1;
-			return -1;
-		}
-	}
-	return 0;
-}
-device_initcall(bpf_pkt_type_offset_init);
-
  /*
   * make sure we dont leak kernel information to user
   */
include/linux/filter.h +0 -0
index a5227ab..6814b54 100644

include/linux/filter.h static inline void *bpf_load_pointer(const struct sk_buff *skb, int k, · line 428

Quoted 5 lines
include/linux/filter.h +2 -0
@@ -424,6 +424,8 @@ static inline void *bpf_load_pointer(const struct sk_buff *skb, int k,
  	return bpf_internal_load_pointer_neg_helper(skb, k, size);
  }
 
+extern __read_mostly int pkt_type_offset;
+

Nit:

extern unsigned int pkt_type_offset;

(Perhaps this should also rather be named bpf_pkt_type_offset to not
cause any current/future name collisions when it's in the header?)

Quoted 8 lines
include/linux/filter.h +0 -0
  #ifdef CONFIG_BPF_JIT
  #include <stdarg.h>
  #include <linux/linkage.h>
net/core/filter.c +0 -0
index d814b8a..edd17f1 100644

net/core/filter.c EXPORT_SYMBOL(sk_filter); · line 100

Quoted 6 lines
net/core/filter.c +2 -1
@@ -96,21 +96,29 @@ EXPORT_SYMBOL(sk_filter);
  #else
  #define PKT_TYPE_MAX	7
  #endif
-static unsigned int pkt_type_offset(void)
+
+__read_mostly int pkt_type_offset;

Nit:

unsigned int pkt_type_offset __read_mostly;

Quoted 4 lines
net/core/filter.c +1 -0
+static int __init init_pkt_type_offset(void)
  {
  	struct sk_buff skb_probe = { .pkt_type = ~0, };
  	u8 *ct = (u8 *) &skb_probe;

net/core/filter.c EXPORT_SYMBOL(sk_filter); · line 114

Quoted 10 lines
net/core/filter.c +4 -1
  	unsigned int off;
 
+	pkt_type_offset = -1;
  	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
+		if (ct[off] == PKT_TYPE_MAX) {
+			pkt_type_offset = off;
  			return off;
+		}
  	}

Why not BUG_ON() when pkt_type_offset could not be found?

That way we would know that something is broken and you can spare the checks for negative offsets everywhere ...

Quoted 22 lines
net/core/filter.c +5 -1
  	pr_err_once("Please fix %s, as pkt_type couldn't be found!\n", __func__);
  	return -1;
  }
 
+pure_initcall(init_pkt_type_offset);
+
  static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
  {
  	return __skb_get_poff((struct sk_buff *)(unsigned long) ctx);
@@ -190,8 +198,10 @@ static bool convert_bpf_extensions(struct sock_filter *fp,
  		break;
 
  	case SKF_AD_OFF + SKF_AD_PKTTYPE:
+		if (pkt_type_offset < 0)
+			return false;
  		*insn = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX,
-				    pkt_type_offset());
+				    pkt_type_offset);
  		if (insn->off < 0)
  			return false;
  		insn++;
#2 lore raw
Quoted 19 lines · Daniel Borkmann
Quoted 6 lines · Denis Kirjanov
net/core/filter.c +1 -0
+static int __init init_pkt_type_offset(void)
  {
        struct sk_buff skb_probe = { .pkt_type = ~0, };
        u8 *ct = (u8 *) &skb_probe;

net/core/filter.c EXPORT_SYMBOL(sk_filter); · line 114

Quoted 10 lines
net/core/filter.c +4 -1
        unsigned int off;
 
+       pkt_type_offset = -1;
        for (off = 0; off < sizeof(struct sk_buff); off++) {
-               if (ct[off] == PKT_TYPE_MAX)
+               if (ct[off] == PKT_TYPE_MAX) {
+                       pkt_type_offset = off;
                        return off;
+               }
        }
Quoted 4 lines

Why not BUG_ON() when pkt_type_offset could not be found?

That way we would know that something is broken and you can spare the checks for negative offsets everywhere ...

also such BUG_ON will be right in the face, since the kernel won't even boot :)
I guess we can only hit it if compiler changes and starts doing crazing things with bitfields.
Eric, you're the original author of this function, thoughts?

Quoted 3 lines

net/core/filter.c static bool convert_bpf_extensions(struct sock_filter *fp, · line 202

Quoted 3 lines

case SKF_AD_OFF + SKF_AD_PKTTYPE:

net/core/filter.c +2 -0
+               if (pkt_type_offset < 0)
+                       return false;

with BUG_ON we wouldn't need this check.

Quoted 5 lines

net/core/filter.c static bool convert_bpf_extensions(struct sock_filter *fp, · line 206

Quoted 5 lines
                *insn = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX,
net/core/filter.c +1 -1
-                                   pkt_type_offset());
+                                   pkt_type_offset);
                if (insn->off < 0)
                        return false;

and this one too.
and other checks in arch/*/jit

#3 lore raw
Quoted 22 lines · Alexei Starovoitov
Quoted 21 lines · Daniel Borkmann
Quoted 6 lines · Denis Kirjanov
net/core/filter.c +1 -0
+static int __init init_pkt_type_offset(void)
  {
        struct sk_buff skb_probe = { .pkt_type = ~0, };
        u8 *ct = (u8 *) &skb_probe;

net/core/filter.c EXPORT_SYMBOL(sk_filter); · line 114

Quoted 10 lines
net/core/filter.c +4 -1
        unsigned int off;
 
+       pkt_type_offset = -1;
        for (off = 0; off < sizeof(struct sk_buff); off++) {
-               if (ct[off] == PKT_TYPE_MAX)
+               if (ct[off] == PKT_TYPE_MAX) {
+                       pkt_type_offset = off;
                        return off;
+               }
        }

Why not BUG_ON() when pkt_type_offset could not be found?

Quoted 8 lines
Quoted 2 lines

That way we would know that something is broken and you can spare the checks for negative offsets everywhere ...

also such BUG_ON will be right in the face, since the kernel won't even boot :)
I guess we can only hit it if compiler changes and starts doing crazing things with bitfields.
Eric, you're the original author of this function, thoughts?

At the time it was written, BPF JIT would have failed and we would revert to interpreter, thus BUG() was not needed : I used one pr_err_once()

Note that pkt_type could be moved easily in sk_buff definition, so a BUG() would be OK after this patch, since the check is done as a pure_initcall() at boot time.

#4 lore raw
Quoted 32 lines · Eric Dumazet
Quoted 22 lines · Alexei Starovoitov
Quoted 21 lines · Daniel Borkmann
Quoted 6 lines · Denis Kirjanov
net/core/filter.c +1 -0
+static int __init init_pkt_type_offset(void)
  {
        struct sk_buff skb_probe = { .pkt_type = ~0, };
        u8 *ct = (u8 *) &skb_probe;

net/core/filter.c EXPORT_SYMBOL(sk_filter); · line 114

Quoted 10 lines
net/core/filter.c +4 -1
        unsigned int off;
 
+       pkt_type_offset = -1;
        for (off = 0; off < sizeof(struct sk_buff); off++) {
-               if (ct[off] == PKT_TYPE_MAX)
+               if (ct[off] == PKT_TYPE_MAX) {
+                       pkt_type_offset = off;
                        return off;
+               }
        }

Why not BUG_ON() when pkt_type_offset could not be found?

Quoted 8 lines
Quoted 2 lines

That way we would know that something is broken and you can spare the checks for negative offsets everywhere ...

also such BUG_ON will be right in the face, since the kernel won't even boot :)
I guess we can only hit it if compiler changes and starts doing crazing things with bitfields.
Eric, you're the original author of this function, thoughts?

Quoted 7 lines

At the time it was written, BPF JIT would have failed and we would revert to interpreter, thus BUG() was not needed : I used one pr_err_once()

Note that pkt_type could be moved easily in sk_buff definition, so a BUG() would be OK after this patch, since the check is done as a pure_initcall() at boot time.

Can't we add an address marker to struct sk_buff?

Several possibilities are available:

ptrdiff_t pkt_type_offset[0] before the pkt_type flags field

If one wants to make it more expressive:
typedef struct {} mark_struct_offset;
and add
mark_struct_offset pkt_type_offset;
at appropriate places

Or maybe an anonymous union?

pkt_type_offset would become a simple offsetof(struct sk_buff, pkt_type_offset) then and there is no need for BUG_ON then.

Bye,
Hannes

#5 lore raw
Quoted 12 lines · Hannes Frederic Sowa

Can't we add an address marker to struct sk_buff?

Several possibilities are available:

ptrdiff_t pkt_type_offset[0] before the pkt_type flags field

If one wants to make it more expressive:
typedef struct {} mark_struct_offset;
and add
mark_struct_offset pkt_type_offset;
at appropriate places

Quoted 4 lines

Or maybe an anonymous union?

pkt_type_offset would become a simple offsetof(struct sk_buff, pkt_type_offset) then and there is no need for BUG_ON then.

You can try, and make sure this works on all gcc compilers, even the one Andrew Morton uses.

And of course, you need to not introduce holes doing so.

#6 lore raw
Quoted 16 lines · Eric Dumazet
Quoted 14 lines · Hannes Frederic Sowa

Can't we add an address marker to struct sk_buff?

Several possibilities are available:

ptrdiff_t pkt_type_offset[0] before the pkt_type flags field

If one wants to make it more expressive:
typedef struct {} mark_struct_offset;
and add
mark_struct_offset pkt_type_offset;
at appropriate places

Or maybe an anonymous union?

Quoted 8 lines
Quoted 2 lines

pkt_type_offset would become a simple offsetof(struct sk_buff, pkt_type_offset) then and there is no need for BUG_ON then.

You can try, and make sure this works on all gcc compilers, even the one Andrew Morton uses.

And of course, you need to not introduce holes doing so.

good points.
I think Hannes's idea is the best. It will help to get rid of helper altogether. For my bpf syscall proposal I've used anonymous unions and tested them with gcc 4.2 - 4.9 on x64/i386/arm32/sparc64 and clang.
All compilers were doing just fine. So it should work.

A message that is not in the archive (sent off-list or not yet received): CAHj3AVn8q-JS0Eq0QrfpBo-WfJ+-wXumhNoeDAdkPZC7LSeB5g@mail.gmail.com

#8 lore raw
Quoted 21 lines · Denis Kirjanov
Quoted 18 lines · Alexei Starovoitov
Quoted 16 lines · On Wed, Sep 3, 2014 at 7:05 PM, Eric Dumazet <[email protected]…
Quoted 14 lines · Hannes Frederic Sowa

Can't we add an address marker to struct sk_buff?

Several possibilities are available:

ptrdiff_t pkt_type_offset[0] before the pkt_type flags field

If one wants to make it more expressive:
typedef struct {} mark_struct_offset;
and add
mark_struct_offset pkt_type_offset;
at appropriate places

Or maybe an anonymous union?

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

pkt_type_offset would become a simple offsetof(struct sk_buff, pkt_type_offset) then and there is no need for BUG_ON then.

You can try, and make sure this works on all gcc compilers, even the one Andrew Morton uses.

And of course, you need to not introduce holes doing so.

Sure, it will work as expected. kmemleak uses the field[0] technique and is already used in struct sk_buff.

I checked below patch with pahole and it doesn't change anything in size or fieldoffsets.

Quoted 4 lines
Quoted 3 lines

good points.
I think Hannes's idea is the best.

Quoted 8 lines

Agreed.

Quoted 4 lines

It will help to get rid of helper altogether.
For my bpf syscall proposal I've used anonymous unions and tested them with gcc 4.2 - 4.9 on x64/i386/arm32/sparc64 and clang. All compilers were doing just fine. So it should work.

Nice. Alexei, care to send a patch then?

Denis, I thought about something like this, you can incorperate this in your patch if you can give it a test and check for other architectures, thanks!

arch/mips/net/bpf_jit.c +0 -0
index 05a5661..0fd3f95 100644
@@ -765,27 +765,6 @@ static u64 jit_get_skb_w(struct sk_buff *skb,

unsigned offset)
return (u64)err << 32 | ntohl(ret);
}

-#ifdef __BIG_ENDIAN_BITFIELD
-#define PKT_TYPE_MAX (7 << 5)
-#else
-#define PKT_TYPE_MAX 7
-#endif
-static int pkt_type_offset(void)
-{
- struct sk_buff skb_probe = {
- .pkt_type = ~0,
- };
- u8 *ct = (u8 *)&skb_probe;
- unsigned int off;
-
- for (off = 0; off < sizeof(struct sk_buff); off++) {
- if (ct[off] == PKT_TYPE_MAX)
- return off;
- }
- pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be
found\n");
- return -1;
-}
-
static int build_body(struct jit_ctx *ctx)
{
void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w};
@@ -1332,11 +1311,7 @@ jmp_cmp:
case BPF_ANC | SKF_AD_PKTTYPE:
ctx->flags |= SEEN_SKB;

- off = pkt_type_offset();
-
- if (off < 0)
- return -1;
- emit_load_byte(r_tmp, r_skb, off, ctx);
+ emit_load_byte(r_tmp, r_skb, skb_pkt_type_offset(), ctx);
/* Keep only the last 3 bits */ emit_andi(r_A, r_tmp, PKT_TYPE_MAX, ctx);
#ifdef __BIG_ENDIAN_BITFIELD

arch/s390/net/bpf_jit_comp.c +1 -34
index 61e45b7..8039bb8 100644
@@ -223,37 +223,6 @@ static void bpf_jit_epilogue(struct bpf_jit *jit)
 	EMIT2(0x07fe);
 }
 
-/* Helper to find the offset of pkt_type in sk_buff
- * Make sure its still a 3bit field starting at the MSBs within a byte.
- */
-#define PKT_TYPE_MAX 0xe0
-static int pkt_type_offset;
-
-static int __init bpf_pkt_type_offset_init(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	char *ct = (char *)&skb_probe;
-	int off;
-
-	pkt_type_offset = -1;
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (!ct[off])
-			continue;
-		if (ct[off] == PKT_TYPE_MAX)
-			pkt_type_offset = off;
-		else {
-			/* Found non matching bit pattern, fix needed. */
-			WARN_ON_ONCE(1);
-			pkt_type_offset = -1;
-			return -1;
-		}
-	}
-	return 0;
-}
-device_initcall(bpf_pkt_type_offset_init);
-
 /*
  * make sure we dont leak kernel information to user
  */
@@ -753,12 +722,10 @@ call_fn:	/* lg %r1,<d(function)>(%r13) */
 		}
 		break;
 	case BPF_ANC | SKF_AD_PKTTYPE:
-		if (pkt_type_offset < 0)
-			goto out;
 		/* lhi %r5,0 */
 		EMIT4(0xa7580000);
 		/* ic %r5,<d(pkt_type_offset)>(%r2) */
-		EMIT4_DISP(0x43502000, pkt_type_offset);
+		EMIT4_DISP(0x43502000, skb_pkt_type_offset());
 		/* srl %r5,5 */
 		EMIT4_DISP(0x88500000, 5);
 		break;
include/linux/skbuff.h +15 -0
index 02529fc..87b86aa 100644
@@ -548,6 +548,10 @@ struct sk_buff {
 				ip_summed:2,
 				nohdr:1,
 				nfctinfo:3;
+	/* __pkt_type_offset marks the offset of the bitfield pkt_type
+	 * contains, so bpf can relatively address it
+	 */
+	int			__pkt_type_offset[0];
 	__u8			pkt_type:3,
 				fclone:2,
 				ipvs_property:1,
@@ -647,6 +651,17 @@ struct sk_buff {
 #define SKB_ALLOC_FCLONE	0x01
 #define SKB_ALLOC_RX		0x02
 
+#ifdef __BIG_ENDIAN_BITFIELD
+#define PKT_TYPE_MAX   (7 << 5)
+#else
+#define PKT_TYPE_MAX   7
+#endif
+
+static inline size_t skb_pkt_type_offset(void)
+{
+	return offsetof(struct sk_buff, __pkt_type_offset);
+}
+
 /* Returns true if the skb was allocated from PFMEMALLOC reserves */
 static inline bool skb_pfmemalloc(const struct sk_buff *skb)
 {
net/core/filter.c +0 -21
index d814b8a..ca64d7a 100644
@@ -87,30 +87,6 @@ int sk_filter(struct sock *sk, struct sk_buff *skb)
 }
 EXPORT_SYMBOL(sk_filter);
 
-/* Helper to find the offset of pkt_type in sk_buff structure. We want
- * to make sure its still a 3bit field starting at a byte boundary;
- * taken from arch/x86/net/bpf_jit_comp.c.
- */
-#ifdef __BIG_ENDIAN_BITFIELD
-#define PKT_TYPE_MAX	(7 << 5)
-#else
-#define PKT_TYPE_MAX	7
-#endif
-static unsigned int pkt_type_offset(void)
-{
-	struct sk_buff skb_probe = { .pkt_type = ~0, };
-	u8 *ct = (u8 *) &skb_probe;
-	unsigned int off;
-
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
-			return off;
-	}
-
-	pr_err_once("Please fix %s, as pkt_type couldn't be found!\n",

__func__);
- return -1;
-}
-
static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5) {
return __skb_get_poff((struct sk_buff *)(unsigned long) ctx);
@@ -191,7 +167,7 @@ static bool convert_bpf_extensions(struct sock_filter *fp,

case SKF_AD_OFF + SKF_AD_PKTTYPE:
*insn = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX,
- pkt_type_offset());
+ skb_pkt_type_offset());
if (insn->off < 0)
return false;
insn++;

#9 lore raw
Quoted 12 lines · Hannes Frederic Sowa

Denis, I thought about something like this, you can incorperate this in your patch if you can give it a test and check for other architectures, thanks!

include/linux/skbuff.h +0 -0
index 02529fc..87b86aa 100644

include/linux/skbuff.h struct sk_buff { · line 554

Quoted 7 lines
include/linux/skbuff.h +4 -0
@@ -548,6 +548,10 @@ struct sk_buff {
 				ip_summed:2,
 				nohdr:1,
 				nfctinfo:3;
+	/* __pkt_type_offset marks the offset of the bitfield pkt_type
+	 * contains, so bpf can relatively address it
+	 */
+	int			__pkt_type_offset[0];

Why is it 'int', while the following uses __u8 ?

This means pkt_type has to stay at this bit offset, and nothing will detect at compile time or execution time if someone did a reorg or added a new field (it seems to be the trend lately)

__u8 bar:4,

                              pkt_type:3,

...

include/linux/skbuff.h struct sk_buff { · line 659

Quoted 13 lines
include/linux/skbuff.h +5 -0
 	__u8			pkt_type:3,
 				fclone:2,
 				ipvs_property:1,
@@ -647,6 +651,17 @@ struct sk_buff {
 #define SKB_ALLOC_FCLONE	0x01
 #define SKB_ALLOC_RX		0x02
 
 
+#ifdef __BIG_ENDIAN_BITFIELD
+#define PKT_TYPE_MAX   (7 << 5)
+#else
+#define PKT_TYPE_MAX   7
+#endif

It had a sense right before static int pkt_type_offset(void), but here, a casual reader might be lost.

I am saying this because a reorder of the bit fields is probably needed to speedup __copy_skb_header() : Grouping together bit fields could allow optimized word copies instead of bit field manipulations.

#10 lore raw
Quoted 9 lines · Eric Dumazet
Quoted 5 lines · Hannes Frederic Sowa

Denis, I thought about something like this, you can incorperate this in your patch if you can give it a test and check for other architectures, thanks!

Quoted 14 lines

include/linux/skbuff.h struct sk_buff { · line 554

Quoted 12 lines
include/linux/skbuff.h +4 -0
index 02529fc..87b86aa 100644
@@ -548,6 +548,10 @@ struct sk_buff {
 				ip_summed:2,
 				nohdr:1,
 				nfctinfo:3;
+	/* __pkt_type_offset marks the offset of the bitfield pkt_type
+	 * contains, so bpf can relatively address it
+	 */
+	int			__pkt_type_offset[0];

Why is it 'int', while the following uses __u8 ?

The type does not matter at all. Actually I wanted to use an empty struct but was afraid it might not work on older compilers and didn't want to check that with each version.

Quoted 7 lines

This means pkt_type has to stay at this bit offset, and nothing will detect at compile time or execution time if someone did a reorg or added a new field (it seems to be the trend lately)

__u8 bar:4,

                              pkt_type:3,

...

That would be fine, one just has to adapt the PKT_TYPE_MAX macros in case on reorders, __pkt_type_offset only needs to stay just in front of the bitfield pkt_type is located in. I think we should move them up into struct sk_buff, so they belong together and people see that they need to change those in case they reorder fields.

I still think about using anonymous unions, they might let us compute the pkt_type mask at compiletime, I think.

Also, maybe we can compute offset at compiletime if we switch to anonymous union.

static inline unsigned int skb_pkt_type_offset(void) {
unsigned int mask;
struct sk_buff skb = {.pkt_type = ~0U}; mask = skb.__flags2; BUILD_BUG_ON(!__builtin_constant_p(mask)); return mask;
}

something like that, haven't checked if that works.

Quoted 11 lines
Quoted 10 lines
include/linux/skbuff.h +0 -0
 	__u8			pkt_type:3,
 				fclone:2,
 				ipvs_property:1,
@@ -647,6 +651,17 @@ struct sk_buff {
 #define SKB_ALLOC_FCLONE	0x01
 #define SKB_ALLOC_RX		0x02
 
Quoted 8 lines

include/linux/skbuff.h struct sk_buff { · line 658

Quoted 5 lines
include/linux/skbuff.h +5 -0
+#ifdef __BIG_ENDIAN_BITFIELD
+#define PKT_TYPE_MAX   (7 << 5)
+#else
+#define PKT_TYPE_MAX   7
+#endif

It had a sense right before static int pkt_type_offset(void), but here, a casual reader might be lost.

Do you think we should just move PKT_TYPE_MAX macros right to the definition of __pkt_type_offset?

Quoted 3 lines

I am saying this because a reorder of the bit fields is probably needed to speedup __copy_skb_header() : Grouping together bit fields could allow optimized word copies instead of bit field manipulations.

Yes, I agree with you. It should be obvious to adapt the macros in case someones reorders sk_buff. Either BUILD_BUG_ON or very obvious comments
+ declarations.

Thanks,
Hannes

#11 lore raw
Quoted 3 lines · Hannes Frederic Sowa

The type does not matter at all. Actually I wanted to use an empty struct but was afraid it might not work on older compilers and didn't want to check that with each version.

It matters. Really.

$ cat try.c
#include <stdio.h>
struct S_int {
	char a;
	int offset[0];
	char b:1;
};
struct S_char {
	char a;
	char offset[0];
	char b:1;
};

int main()

{
	printf("sizeof(S_int) %zu\n", sizeof(struct S_int));
	printf("sizeof(S_char) %zu\n", sizeof(struct S_char));
	return 0;
}
$ gcc -o try try.c && ./try
sizeof(S_int) 8
sizeof(S_char) 2
#12 lore raw
Quoted 25 lines · Eric Dumazet
Quoted 3 lines · Hannes Frederic Sowa

The type does not matter at all. Actually I wanted to use an empty struct but was afraid it might not work on older compilers and didn't want to check that with each version.

It matters. Really.

$ cat try.c
#include <stdio.h>
struct S_int {
	char a;
	int offset[0];
	char b:1;
};
struct S_char {
	char a;
	char offset[0];
	char b:1;
};

int main()

Quoted 8 lines
{
	printf("sizeof(S_int) %zu\n", sizeof(struct S_int));
	printf("sizeof(S_char) %zu\n", sizeof(struct S_char));
	return 0;
}
$ gcc -o try try.c && ./try
sizeof(S_int) 8
sizeof(S_char) 2

Sure, compiler must make sure to correctly align structure for all members (so it must round up). But this is not an issue with struct sk_buff. I agree that switching to __u8 for offset marker seems to reduce confusion.

Statically compiling skb_pkt_type_mask as I proposed above didn't work out.

New patch, what do you think?

arch/mips/net/bpf_jit.c +1 -26
index 05a5661..0fd3f95 100644
@@ -765,27 +765,6 @@ static u64 jit_get_skb_w(struct sk_buff *skb, unsigned offset)
 	return (u64)err << 32 | ntohl(ret);
 }
 
-#ifdef __BIG_ENDIAN_BITFIELD
-#define PKT_TYPE_MAX	(7 << 5)
-#else
-#define PKT_TYPE_MAX	7
-#endif
-static int pkt_type_offset(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	u8 *ct = (u8 *)&skb_probe;
-	unsigned int off;
-
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
-			return off;
-	}
-	pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be found\n");
-	return -1;
-}
-
 static int build_body(struct jit_ctx *ctx)
 {
 	void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w};
@@ -1332,11 +1311,7 @@ jmp_cmp:
 		case BPF_ANC | SKF_AD_PKTTYPE:
 			ctx->flags |= SEEN_SKB;
 
-			off = pkt_type_offset();
-
-			if (off < 0)
-				return -1;
-			emit_load_byte(r_tmp, r_skb, off, ctx);
+			emit_load_byte(r_tmp, r_skb, skb_pkt_type_offset(), ctx);
 			/* Keep only the last 3 bits */
 			emit_andi(r_A, r_tmp, PKT_TYPE_MAX, ctx);
 #ifdef __BIG_ENDIAN_BITFIELD
arch/s390/net/bpf_jit_comp.c +1 -34
index 61e45b7..8039bb8 100644
@@ -223,37 +223,6 @@ static void bpf_jit_epilogue(struct bpf_jit *jit)
 	EMIT2(0x07fe);
 }
 
-/* Helper to find the offset of pkt_type in sk_buff
- * Make sure its still a 3bit field starting at the MSBs within a byte.
- */
-#define PKT_TYPE_MAX 0xe0
-static int pkt_type_offset;
-
-static int __init bpf_pkt_type_offset_init(void)
-{
-	struct sk_buff skb_probe = {
-		.pkt_type = ~0,
-	};
-	char *ct = (char *)&skb_probe;
-	int off;
-
-	pkt_type_offset = -1;
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (!ct[off])
-			continue;
-		if (ct[off] == PKT_TYPE_MAX)
-			pkt_type_offset = off;
-		else {
-			/* Found non matching bit pattern, fix needed. */
-			WARN_ON_ONCE(1);
-			pkt_type_offset = -1;
-			return -1;
-		}
-	}
-	return 0;
-}
-device_initcall(bpf_pkt_type_offset_init);
-
 /*
  * make sure we dont leak kernel information to user
  */
@@ -753,12 +722,10 @@ call_fn:	/* lg %r1,<d(function)>(%r13) */
 		}
 		break;
 	case BPF_ANC | SKF_AD_PKTTYPE:
-		if (pkt_type_offset < 0)
-			goto out;
 		/* lhi %r5,0 */
 		EMIT4(0xa7580000);
 		/* ic %r5,<d(pkt_type_offset)>(%r2) */
-		EMIT4_DISP(0x43502000, pkt_type_offset);
+		EMIT4_DISP(0x43502000, skb_pkt_type_offset());
 		/* srl %r5,5 */
 		EMIT4_DISP(0x88500000, 5);
 		break;
include/linux/skbuff.h +19 -0
index 02529fc..944e931 100644
@@ -548,11 +548,25 @@ struct sk_buff {
 				ip_summed:2,
 				nohdr:1,
 				nfctinfo:3;
+
+	/* pkt_type bitfield is relatively addressed from bpf engines:
+	 * if you reorder skb contents, ensure to update PKT_TYPE_MAX
+	 * and keep __pkt_type_offset marker right in front of bitfield
+	 * which does contain pkt_type member.
+	 */
+
+#ifdef __BIG_ENDIAN_BITFIELD
+#define PKT_TYPE_MAX   (7 << 5)
+#else
+#define PKT_TYPE_MAX   7
+#endif
+	__u8			__pkt_type_offset[0];
 	__u8			pkt_type:3,
 				fclone:2,
 				ipvs_property:1,
 				peeked:1,
 				nf_trace:1;
+
 	kmemcheck_bitfield_end(flags1);
 	__be16			protocol;
 
@@ -647,6 +661,11 @@ struct sk_buff {
 #define SKB_ALLOC_FCLONE	0x01
 #define SKB_ALLOC_RX		0x02
 
+static inline size_t skb_pkt_type_offset(void)
+{
+	return offsetof(struct sk_buff, __pkt_type_offset);
+}
+
 /* Returns true if the skb was allocated from PFMEMALLOC reserves */
 static inline bool skb_pfmemalloc(const struct sk_buff *skb)
 {
net/core/filter.c +1 -25
index d814b8a..ca64d7a 100644
@@ -87,30 +87,6 @@ int sk_filter(struct sock *sk, struct sk_buff *skb)
 }
 EXPORT_SYMBOL(sk_filter);
 
-/* Helper to find the offset of pkt_type in sk_buff structure. We want
- * to make sure its still a 3bit field starting at a byte boundary;
- * taken from arch/x86/net/bpf_jit_comp.c.
- */
-#ifdef __BIG_ENDIAN_BITFIELD
-#define PKT_TYPE_MAX	(7 << 5)
-#else
-#define PKT_TYPE_MAX	7
-#endif
-static unsigned int pkt_type_offset(void)
-{
-	struct sk_buff skb_probe = { .pkt_type = ~0, };
-	u8 *ct = (u8 *) &skb_probe;
-	unsigned int off;
-
-	for (off = 0; off < sizeof(struct sk_buff); off++) {
-		if (ct[off] == PKT_TYPE_MAX)
-			return off;
-	}
-
-	pr_err_once("Please fix %s, as pkt_type couldn't be found!\n", __func__);
-	return -1;
-}
-
 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
 {
 	return __skb_get_poff((struct sk_buff *)(unsigned long) ctx);
@@ -191,7 +167,7 @@ static bool convert_bpf_extensions(struct sock_filter *fp,
 
 	case SKF_AD_OFF + SKF_AD_PKTTYPE:
 		*insn = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX,
-				    pkt_type_offset());
+				    skb_pkt_type_offset());
 		if (insn->off < 0)
 			return false;
 		insn++;
#13 lore raw
In reply to Eric Dumazet

From: Eric Dumazet

Quoted 18 lines · Hannes Frederic Sowa
Quoted 4 lines

The type does not matter at all. Actually I wanted to use an empty struct but was afraid it might not work on older compilers and didn't want to check that with each version.

It matters. Really.

$ cat try.c
#include <stdio.h>
struct S_int {
	char a;
	int offset[0];
	char b:1;
};
Quoted 5 lines
struct S_char {
	char a;
	char offset[0];
	char b:1;
};

Those are also both illegal C, zero sized arrays aren't allowed.

David

#14 lore raw
Quoted 25 lines · David Laight

From: Eric Dumazet

Quoted 18 lines · Hannes Frederic Sowa
Quoted 4 lines

The type does not matter at all. Actually I wanted to use an empty struct but was afraid it might not work on older compilers and didn't want to check that with each version.

It matters. Really.

$ cat try.c
#include <stdio.h>
struct S_int {
	char a;
	int offset[0];
	char b:1;
};
Quoted 5 lines
struct S_char {
	char a;
	char offset[0];
	char b:1;
};
Quoted 3 lines

Those are also both illegal C, zero sized arrays aren't allowed.

David

This does not matter, we already use such constructs in the kernel.

Take a look at kmemcheck_bitfield_begin() for example.

#15 lore raw
Quoted 26 lines · Eric Dumazet
Quoted 25 lines · David Laight

From: Eric Dumazet

Quoted 23 lines · Hannes Frederic Sowa
Quoted 4 lines

The type does not matter at all. Actually I wanted to use an empty struct but was afraid it might not work on older compilers and didn't want to check that with each version.

It matters. Really.

$ cat try.c
#include <stdio.h>
struct S_int {
	char a;
	int offset[0];
	char b:1;
};
struct S_char {
	char a;
	char offset[0];
	char b:1;
};
Quoted 8 lines
Quoted 3 lines

Those are also both illegal C, zero sized arrays aren't allowed.

David

This does not matter, we already use such constructs in the kernel.

Take a look at kmemcheck_bitfield_begin() for example.

Which btw. also uses int, which might change alignment of structures.

#16 lore raw
Quoted 1 line · Hannes Frederic Sowa

Which btw. also uses int, which might change alignment of structures.

You missed the point .

kmemcheck wants to make sure the whole word is set, or else you could get false positives.

kmemcheck needs are quite different.

#17 lore raw
Quoted 10 lines · Eric Dumazet
Quoted 1 line · Hannes Frederic Sowa

Which btw. also uses int, which might change alignment of structures.

You missed the point .

kmemcheck wants to make sure the whole word is set, or else you could get false positives.

kmemcheck needs are quite different.

Now that you said it, I understand. :)

You were right with the int vs. u8 thing all along. gcc aligns the datatype on the next struct field and not on the whole field, as I expected. So excuse my error above.

I think the latest proposals looks good?

#18 lore raw
Quoted 12 lines · Hannes Frederic Sowa
Quoted 10 lines · Eric Dumazet
Quoted 1 line · Hannes Frederic Sowa

Which btw. also uses int, which might change alignment of structures.

You missed the point .

kmemcheck wants to make sure the whole word is set, or else you could get false positives.

kmemcheck needs are quite different.

Quoted 7 lines

Now that you said it, I understand. :)

You were right with the int vs. u8 thing all along. gcc aligns the datatype on the next struct field and not on the whole field, as I expected. So excuse my error above.

I think the latest proposals looks good?

to me: yes
and I think your latest half-patch with:
+ __u8 __pkt_type_offset[0];
also looks good.

Are you going to take it over from Denis here?

while at it would you fix sparc jit as well that has comment:

#if 0
                                /* GCC won't let us take the address of
                                 * a bit field even though we very much
                                 * know what we are doing here.
                                 */
                        case BPF_ANC | SKF_AD_PKTTYPE:
                                __emit_skb_load8(pkt_type, r_A);
                                emit_alu_K(SRL, 5);
                                break;

#endif
should be able to replace 'pkt_type' above with __pkt_type_offset...

#19 lore raw

Hi,

Quoted 20 lines · Alexei Starovoitov
Quoted 18 lines · Hannes Frederic Sowa
Quoted 10 lines · Eric Dumazet
Quoted 1 line · Hannes Frederic Sowa

Which btw. also uses int, which might change alignment of structures.

You missed the point .

kmemcheck wants to make sure the whole word is set, or else you could get false positives.

kmemcheck needs are quite different.

Now that you said it, I understand. :)

You were right with the int vs. u8 thing all along. gcc aligns the datatype on the next struct field and not on the whole field, as I expected. So excuse my error above.

Quoted 8 lines
Quoted 1 line

I think the latest proposals looks good?

to me: yes
and I think your latest half-patch with:
+ __u8 __pkt_type_offset[0];
also looks good.

Are you going to take it over from Denis here?

I don't know. Denis, do you want to incorperate my changes or should I take over here?

Quoted 10 lines

while at it would you fix sparc jit as well that has comment:

#if 0
                                /* GCC won't let us take the address of
                                 * a bit field even though we very much
                                 * know what we are doing here.
                                 */
                        case BPF_ANC | SKF_AD_PKTTYPE:
                                __emit_skb_load8(pkt_type, r_A);
                                emit_alu_K(SRL, 5);
                                break;
Quoted 2 lines

#endif
should be able to replace 'pkt_type' above with __pkt_type_offset...

I'll have a look then but cannot test sparc easily.

Bye,
Hannes

#20 lore raw
Quoted 8 lines · Hannes Frederic Sowa
Quoted 8 lines
Quoted 2 lines

I think the latest proposals looks good?

to me: yes
and I think your latest half-patch with:
+ __u8 __pkt_type_offset[0];
also looks good.

Quoted 4 lines
Quoted 1 line

Are you going to take it over from Denis here?

I don't know. Denis, do you want to incorperate my changes or should I take over here?

Hannes, Denis,
anyone of you going to continue on that?
imo it's a good cleanup.

A message that is not in the archive (sent off-list or not yet received): CAOJe8K2eQ9ejhBTk7MzJJK5EB4D62U4uKDiFFQFh+T16E7=S3A@mail.gmail.com

#22 lore raw

Hi,

Quoted 2 lines · Denis Kirjanov

Hi guys, sorry, I was on vacation and didn't notice you. I'll continue to work on that.

I have the patch still in my pipeline but just wanted to hear from you again (or you time out), if you would still like to work on it. Otherwise will send it out later today.

Bye,
Hannes