[PATCH 1/3] net: bpf jit: ppc: optimize choose_load_func error path

[PATCH 1/3] net: bpf jit: ppc: optimize choose_load_func error path

[Vladimir Murzin]

Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
were not passed. At the same time handlers for "any offset" cases make
the same checks against r_addr at run-time, that will always lead to
bpf_error.

Run-time checks are still necessary for indirect load operations, but
error path for absolute and mesh loads are worth to optimize during bpf
compile time.

Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>

Cc: Jan Seiffert <kaffeemonster@googlemail.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Daniel Borkmann <dborkman@redhat.com>
Cc: Matt Evans <matt@ozlabs.org>

---
 arch/powerpc/net/bpf_jit_comp.c |    7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
index bf56e33..754320a 100644
--- a/arch/powerpc/net/bpf_jit_comp.c
+++ b/arch/powerpc/net/bpf_jit_comp.c
@@ -132,7 +132,7 @@ static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
 }
 
 #define CHOOSE_LOAD_FUNC(K, func) \
-	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
+	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
 
 /* Assemble the body code between the prologue & epilogue. */
 static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
@@ -427,6 +427,11 @@ static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
 		case BPF_S_LD_B_ABS:
 			func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
 		common_load:
+			if (!func) {
+				PPC_LI(r_ret, 0);
+				PPC_JMP(exit_addr);
+				break;
+			}
 			/* Load from [K]. */
 			ctx->seen |= SEEN_DATAREF;
 			PPC_LI64(r_scratch1, func);
-- 
1.7.10.4

Re: [PATCH 1/3] net: bpf jit: ppc: optimize choose_load_func error path

[Jan Seiffert]

Vladimir Murzin schrieb:
> Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
> were not passed. At the same time handlers for "any offset" cases make
> the same checks against r_addr at run-time, that will always lead to
> bpf_error.
> 

Hmmm, if i only would remember why i wrote it that way....
I memory serves me right the idea was to always have a solid fall back, no
matter what, to the generic load function which works more like the load_pointer
from filter.c.
This way the COOSE-macro may could have been used at more places, but that
never played out.

And since all i wanted was to get the negative indirect load fixed,
optimizing the constant error case was not on my plate.
That you can get your negative K filter JITed in the first place, even
if the constant error case was slower than necessary, was good enough ;)

The ARM JIT is broken till this date...

You can have my
I'm-OK-with-this: Jan Seiffert <kaffeemonster@googlemail.com>

for all three patches, -ENOTIME for a full review ATM.

> Run-time checks are still necessary for indirect load operations, but
> error path for absolute and mesh loads are worth to optimize during bpf
> compile time.
> 
> Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>
> 
> Cc: Jan Seiffert <kaffeemonster@googlemail.com>
> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Cc: Paul Mackerras <paulus@samba.org>
> Cc: Daniel Borkmann <dborkman@redhat.com>
> Cc: Matt Evans <matt@ozlabs.org>
> 
> ---
>  arch/powerpc/net/bpf_jit_comp.c |    7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
> index bf56e33..754320a 100644
> --- a/arch/powerpc/net/bpf_jit_comp.c
> +++ b/arch/powerpc/net/bpf_jit_comp.c
> @@ -132,7 +132,7 @@ static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
>  }
>  
>  #define CHOOSE_LOAD_FUNC(K, func) \
> -	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
> +	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
>  
>  /* Assemble the body code between the prologue & epilogue. */
>  static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
> @@ -427,6 +427,11 @@ static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
>  		case BPF_S_LD_B_ABS:
>  			func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
>  		common_load:
> +			if (!func) {
> +				PPC_LI(r_ret, 0);
> +				PPC_JMP(exit_addr);
> +				break;
> +			}
>  			/* Load from [K]. */
>  			ctx->seen |= SEEN_DATAREF;
>  			PPC_LI64(r_scratch1, func);
> 


-- 
An UDP packet walks into a

Re: [PATCH 1/3] net: bpf jit: ppc: optimize choose_load_func error path

[Vladimir Murzin]

On Wed, Oct 09, 2013 at 12:50:32AM +0200, Jan Seiffert wrote:
> Vladimir Murzin schrieb:
> > Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
> > were not passed. At the same time handlers for "any offset" cases make
> > the same checks against r_addr at run-time, that will always lead to
> > bpf_error.
> > 
> 
> Hmmm, if i only would remember why i wrote it that way....
> I memory serves me right the idea was to always have a solid fall back, no
> matter what, to the generic load function which works more like the load_pointer
> from filter.c.
> This way the COOSE-macro may could have been used at more places, but that
> never played out.
> 
> And since all i wanted was to get the negative indirect load fixed,
> optimizing the constant error case was not on my plate.
> That you can get your negative K filter JITed in the first place, even
> if the constant error case was slower than necessary, was good enough ;)
> 
> The ARM JIT is broken till this date...

... and s390 too.

> 
> You can have my
> I'm-OK-with-this: Jan Seiffert <kaffeemonster@googlemail.com>
> 
> for all three patches, -ENOTIME for a full review ATM.
> 

Thanks for feedback, Jan!

> > Run-time checks are still necessary for indirect load operations, but
> > error path for absolute and mesh loads are worth to optimize during bpf
> > compile time.
> > 
> > Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>
> > 
> > Cc: Jan Seiffert <kaffeemonster@googlemail.com>
> > Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> > Cc: Paul Mackerras <paulus@samba.org>
> > Cc: Daniel Borkmann <dborkman@redhat.com>
> > Cc: Matt Evans <matt@ozlabs.org>
> > 
> > ---
> >  arch/powerpc/net/bpf_jit_comp.c |    7 ++++++-
> >  1 file changed, 6 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
> > index bf56e33..754320a 100644
> > --- a/arch/powerpc/net/bpf_jit_comp.c
> > +++ b/arch/powerpc/net/bpf_jit_comp.c
> > @@ -132,7 +132,7 @@ static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
> >  }
> >  
> >  #define CHOOSE_LOAD_FUNC(K, func) \
> > -	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
> > +	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
> >  
> >  /* Assemble the body code between the prologue & epilogue. */
> >  static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
> > @@ -427,6 +427,11 @@ static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
> >  		case BPF_S_LD_B_ABS:
> >  			func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
> >  		common_load:
> > +			if (!func) {
> > +				PPC_LI(r_ret, 0);
> > +				PPC_JMP(exit_addr);
> > +				break;
> > +			}
> >  			/* Load from [K]. */
> >  			ctx->seen |= SEEN_DATAREF;
> >  			PPC_LI64(r_scratch1, func);
> > 
> 
> 
> -- 
> An UDP packet walks into a

Re: [PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[David Miller]

From: Vladimir Murzin <murzin.v@gmail.com>
Date: Tue,  8 Oct 2013 20:31:49 +0400

> Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
> were not passed. At the same time handlers for "any offset" cases make
> the same checks against r_addr at run-time, that will always lead to
> bpf_error.
> 
> Run-time checks are still necessary for indirect load operations, but
> error path for absolute and mesh loads are worth to optimize during bpf
> compile time.
> 
> Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>
> 
> Cc: Jan Seiffert <kaffeemonster@googlemail.com>
> Cc: Eric Dumazet <edumazet@google.com>
> Cc: "David S. Miller" <davem@davemloft.net
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> 
> ---
>  arch/x86/net/bpf_jit_comp.c |   15 +++++++++------
>  1 file changed, 9 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index 79c216a..28ac17f 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -123,7 +123,7 @@ static inline void bpf_flush_icache(void *start, void *end)
>  }
>  
>  #define CHOOSE_LOAD_FUNC(K, func) \
> -	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
> +	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
>  
>  /* Helper to find the offset of pkt_type in sk_buff
>   * We want to make sure its still a 3bit field starting at a byte boundary.
> @@ -611,7 +611,13 @@ void bpf_jit_compile(struct sk_filter *fp)
>  			}
>  			case BPF_S_LD_W_ABS:
>  				func = CHOOSE_LOAD_FUNC(K, sk_load_word);
> -common_load:			seen |= SEEN_DATAREF;
> +common_load:
> +				if (!func) {
> +					CLEAR_A();
> +					EMIT_JMP(cleanup_addr - addrs[i]);
> +					break;
> +				}
> +				seen |= SEEN_DATAREF;
>  				t_offset = func - (image + addrs[i]);
>  				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
>  				EMIT1_off32(0xe8, t_offset); /* call */
> @@ -625,10 +631,7 @@ common_load:			seen |= SEEN_DATAREF;
>  			case BPF_S_LDX_B_MSH:
>  				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
>  				seen |= SEEN_DATAREF | SEEN_XREG;
> -				t_offset = func - (image + addrs[i]);
> -				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
> -				EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */
> -				break;
> +				goto common_load;

This second hunk will set SEEN_DATAREF even if common_load takes the
!func path, that is not the intention at all here.

There's a reason why these two code blocks aren't shared.

Re: [PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[Vladimir Murzin]

On Fri, Oct 11, 2013 at 02:56:13PM -0400, David Miller wrote:
> From: Vladimir Murzin <murzin.v@gmail.com>
> Date: Tue,  8 Oct 2013 20:31:49 +0400
> 
> > Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
> > were not passed. At the same time handlers for "any offset" cases make
> > the same checks against r_addr at run-time, that will always lead to
> > bpf_error.
> > 
> > Run-time checks are still necessary for indirect load operations, but
> > error path for absolute and mesh loads are worth to optimize during bpf
> > compile time.
> > 
> > Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>
> > 
> > Cc: Jan Seiffert <kaffeemonster@googlemail.com>
> > Cc: Eric Dumazet <edumazet@google.com>
> > Cc: "David S. Miller" <davem@davemloft.net
> > Cc: "H. Peter Anvin" <hpa@zytor.com>
> > Cc: Ingo Molnar <mingo@redhat.com>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > 
> > ---
> >  arch/x86/net/bpf_jit_comp.c |   15 +++++++++------
> >  1 file changed, 9 insertions(+), 6 deletions(-)
> > 
> > diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> > index 79c216a..28ac17f 100644
> > --- a/arch/x86/net/bpf_jit_comp.c
> > +++ b/arch/x86/net/bpf_jit_comp.c
> > @@ -123,7 +123,7 @@ static inline void bpf_flush_icache(void *start, void *end)
> >  }
> >  
> >  #define CHOOSE_LOAD_FUNC(K, func) \
> > -	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
> > +	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
> >  
> >  /* Helper to find the offset of pkt_type in sk_buff
> >   * We want to make sure its still a 3bit field starting at a byte boundary.
> > @@ -611,7 +611,13 @@ void bpf_jit_compile(struct sk_filter *fp)
> >  			}
> >  			case BPF_S_LD_W_ABS:
> >  				func = CHOOSE_LOAD_FUNC(K, sk_load_word);
> > -common_load:			seen |= SEEN_DATAREF;
> > +common_load:
> > +				if (!func) {
> > +					CLEAR_A();
> > +					EMIT_JMP(cleanup_addr - addrs[i]);
> > +					break;
> > +				}
> > +				seen |= SEEN_DATAREF;
> >  				t_offset = func - (image + addrs[i]);
> >  				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
> >  				EMIT1_off32(0xe8, t_offset); /* call */
> > @@ -625,10 +631,7 @@ common_load:			seen |= SEEN_DATAREF;
> >  			case BPF_S_LDX_B_MSH:
> >  				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
> >  				seen |= SEEN_DATAREF | SEEN_XREG;
> > -				t_offset = func - (image + addrs[i]);
> > -				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
> > -				EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */
> > -				break;
> > +				goto common_load;
> 
> This second hunk will set SEEN_DATAREF even if common_load takes the
> !func path, that is not the intention at all here.
> 
> There's a reason why these two code blocks aren't shared.

Thanks for review, David!

What about patch bellow?

---
 arch/x86/net/bpf_jit_comp.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index 79c216a..92128fe 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -123,7 +123,7 @@ static inline void bpf_flush_icache(void *start, void *end)
 }
 
 #define CHOOSE_LOAD_FUNC(K, func) \
-	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
+	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
 
 /* Helper to find the offset of pkt_type in sk_buff
  * We want to make sure its still a 3bit field starting at a byte boundary.
@@ -611,7 +611,14 @@ void bpf_jit_compile(struct sk_filter *fp)
 			}
 			case BPF_S_LD_W_ABS:
 				func = CHOOSE_LOAD_FUNC(K, sk_load_word);
-common_load:			seen |= SEEN_DATAREF;
+common_load:
+				if (!func) {
+					CLEAR_A();
+					EIT_JMP(cleanup_addr - addrs[i]);
+					break;
+				}
+
+				seen |= SEEN_DATAREF;
 				t_offset = func - (image + addrs[i]);
 				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
 				EMIT1_off32(0xe8, t_offset); /* call */
@@ -624,6 +631,13 @@ common_load:			seen |= SEEN_DATAREF;
 				goto common_load;
 			case BPF_S_LDX_B_MSH:
 				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
+
+				if (!func) {
+					CLEAR_A();
+					EIT_JMP(cleanup_addr - addrs[i]);
+					break;
+				}
+
 				seen |= SEEN_DATAREF | SEEN_XREG;
 				t_offset = func - (image + addrs[i]);
 				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
-- 
1.8.1.5

Vladimir

Re: [PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[malc]

On Sun, 13 Oct 2013, Vladimir Murzin wrote:

> On Fri, Oct 11, 2013 at 02:56:13PM -0400, David Miller wrote:
> > From: Vladimir Murzin <murzin.v@gmail.com>
> > Date: Tue,  8 Oct 2013 20:31:49 +0400
> > 

[..snip..]

> -common_load:			seen |= SEEN_DATAREF;
> +common_load:
> +				if (!func) {
> +					CLEAR_A();
> +					EIT_JMP(cleanup_addr - addrs[i]);

                                        EMIT? (likewise elsewhere)

> +					break;
> +				}
> +
> +				seen |= SEEN_DATAREF;
>  				t_offset = func - (image + addrs[i]);
>  				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
>  				EMIT1_off32(0xe8, t_offset); /* call */
> @@ -624,6 +631,13 @@ common_load:			seen |= SEEN_DATAREF;
>  				goto common_load;
>  			case BPF_S_LDX_B_MSH:
>  				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
> +
> +				if (!func) {
> +					CLEAR_A();
> +					EIT_JMP(cleanup_addr - addrs[i]);
> +					break;
> +				}
> +
>  				seen |= SEEN_DATAREF | SEEN_XREG;
>  				t_offset = func - (image + addrs[i]);
>  				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
> 

-- 
mailto:av1474@comtv.ru

Re: [PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[Vladimir Murzin]

On Sun, Oct 13, 2013 at 06:25:34PM +0400, malc wrote:
> On Sun, 13 Oct 2013, Vladimir Murzin wrote:
> 
> > On Fri, Oct 11, 2013 at 02:56:13PM -0400, David Miller wrote:
> > > From: Vladimir Murzin <murzin.v@gmail.com>
> > > Date: Tue,  8 Oct 2013 20:31:49 +0400
> > > 
> 
> [..snip..]
> 
> > -common_load:			seen |= SEEN_DATAREF;
> > +common_load:
> > +				if (!func) {
> > +					CLEAR_A();
> > +					EIT_JMP(cleanup_addr - addrs[i]);
> 
>                                         EMIT? (likewise elsewhere)
Oops... Thanks for quick response!

I'd better send the patch as a separate message.

> 
> > +					break;
> > +				}
> > +
> > +				seen |= SEEN_DATAREF;
> >  				t_offset = func - (image + addrs[i]);
> >  				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
> >  				EMIT1_off32(0xe8, t_offset); /* call */
> > @@ -624,6 +631,13 @@ common_load:			seen |= SEEN_DATAREF;
> >  				goto common_load;
> >  			case BPF_S_LDX_B_MSH:
> >  				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
> > +
> > +				if (!func) {
> > +					CLEAR_A();
> > +					EIT_JMP(cleanup_addr - addrs[i]);
> > +					break;
> > +				}
> > +
> >  				seen |= SEEN_DATAREF | SEEN_XREG;
> >  				t_offset = func - (image + addrs[i]);
> >  				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
> > 
> 
> -- 
> mailto:av1474@comtv.ru

[PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[Vladimir Murzin]

Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
were not passed. At the same time handlers for "any offset" cases make
the same checks against r_addr at run-time, that will always lead to
bpf_error.

Run-time checks are still necessary for indirect load operations, but
error path for absolute and mesh loads are worth to optimize during bpf
compile time.

Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>
---

David pointed at inability to merge mesh load with common load code. This
patch is updated according to this note.

 arch/x86/net/bpf_jit_comp.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index 79c216a..92128fe 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -123,7 +123,7 @@ static inline void bpf_flush_icache(void *start, void *end)
 }
 
 #define CHOOSE_LOAD_FUNC(K, func) \
-	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
+	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
 
 /* Helper to find the offset of pkt_type in sk_buff
  * We want to make sure its still a 3bit field starting at a byte boundary.
@@ -611,7 +611,14 @@ void bpf_jit_compile(struct sk_filter *fp)
 			}
 			case BPF_S_LD_W_ABS:
 				func = CHOOSE_LOAD_FUNC(K, sk_load_word);
-common_load:			seen |= SEEN_DATAREF;
+common_load:
+				if (!func) {
+					CLEAR_A();
+					EMIT_JMP(cleanup_addr - addrs[i]);
+					break;
+				}
+
+				seen |= SEEN_DATAREF;
 				t_offset = func - (image + addrs[i]);
 				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
 				EMIT1_off32(0xe8, t_offset); /* call */
@@ -624,6 +631,13 @@ common_load:			seen |= SEEN_DATAREF;
 				goto common_load;
 			case BPF_S_LDX_B_MSH:
 				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
+
+				if (!func) {
+					CLEAR_A();
+					EMIT_JMP(cleanup_addr - addrs[i]);
+					break;
+				}
+
 				seen |= SEEN_DATAREF | SEEN_XREG;
 				t_offset = func - (image + addrs[i]);
 				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
-- 
1.8.1.5

[PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[Vladimir Murzin]

Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
were not passed. At the same time handlers for "any offset" cases make
the same checks against r_addr at run-time, that will always lead to
bpf_error.

Run-time checks are still necessary for indirect load operations, but
error path for absolute and mesh loads are worth to optimize during bpf
compile time.

Signed-off-by: Vladimir Murzin <murzin.v@gmail.com>
---

David pointed at inability to merge mesh load with common load code. This
patch is updated according to this note.

 arch/x86/net/bpf_jit_comp.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index 79c216a..92128fe 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -123,7 +123,7 @@ static inline void bpf_flush_icache(void *start, void *end)
 }
 
 #define CHOOSE_LOAD_FUNC(K, func) \
-	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
+	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : NULL) : func##_positive_offset)
 
 /* Helper to find the offset of pkt_type in sk_buff
  * We want to make sure its still a 3bit field starting at a byte boundary.
@@ -611,7 +611,14 @@ void bpf_jit_compile(struct sk_filter *fp)
 			}
 			case BPF_S_LD_W_ABS:
 				func = CHOOSE_LOAD_FUNC(K, sk_load_word);
-common_load:			seen |= SEEN_DATAREF;
+common_load:
+				if (!func) {
+					CLEAR_A();
+					EMIT_JMP(cleanup_addr - addrs[i]);
+					break;
+				}
+
+				seen |= SEEN_DATAREF;
 				t_offset = func - (image + addrs[i]);
 				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
 				EMIT1_off32(0xe8, t_offset); /* call */
@@ -624,6 +631,13 @@ common_load:			seen |= SEEN_DATAREF;
 				goto common_load;
 			case BPF_S_LDX_B_MSH:
 				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
+
+				if (!func) {
+					CLEAR_A();
+					EMIT_JMP(cleanup_addr - addrs[i]);
+					break;
+				}
+
 				seen |= SEEN_DATAREF | SEEN_XREG;
 				t_offset = func - (image + addrs[i]);
 				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
-- 
1.8.1.5

Re: [PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[Eric Dumazet]

On Sun, 2013-10-13 at 16:54 +0200, Vladimir Murzin wrote:
> Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
> were not passed. At the same time handlers for "any offset" cases make
> the same checks against r_addr at run-time, that will always lead to
> bpf_error.
> 
> Run-time checks are still necessary for indirect load operations, but
> error path for absolute and mesh loads are worth to optimize during bpf
> compile time.

I don't get the point.

What real world use case or problem are you trying to handle ?

bpf_error returns 0, so it seems your patch does the same.

A buggy BPF program should not expect us to 'save' a few cycles.

Re: [PATCH 2/3] net: bpf jit: x86: optimize choose_load_func error path

[Vladimir Murzin]

On Sun, Oct 13, 2013 at 09:36:34AM -0700, Eric Dumazet wrote:
> On Sun, 2013-10-13 at 16:54 +0200, Vladimir Murzin wrote:
> > Macro CHOOSE_LOAD_FUNC returns handler for "any offset" if checks for K
> > were not passed. At the same time handlers for "any offset" cases make
> > the same checks against r_addr at run-time, that will always lead to
> > bpf_error.
> > 
> > Run-time checks are still necessary for indirect load operations, but
> > error path for absolute and mesh loads are worth to optimize during bpf
> > compile time.
> 
> I don't get the point.
> 
> What real world use case or problem are you trying to handle ?
> 
> bpf_error returns 0, so it seems your patch does the same.
> 
> A buggy BPF program should not expect us to 'save' a few cycles.
> 
> 
> 

Hi Eric!

There is no real world use case for me - it was eliminated by plain code
reading. The patch is not supposed to change behavior of BPF program - only
optimization of the error path. 
I agree with, you there is no significant reason for optimizations of rarely
used pice of code. However, it is not only saving pipeline cycles and I-cache
lines for usually "never taken" branch. In case this "never taken" branch is a
buggy part of BPF program, we can avoid extra instructions and save space for
the rest of BPF program - there is no need to care about seen flags of buggy
part anymore.

Anyway, if you still think it's not good enough - just throw it away ;)

Thanks
Vladimir