[PATCH bpf-next v3 0/2] bpf, arm64: fix bpf line info

[PATCH bpf-next v3 0/2] bpf, arm64: fix bpf line info

[Hou Tao]

Hi,

The patchset addresses two issues in bpf line info for arm64:

(1) insn_to_jit_off only considers the body itself and ignores
    prologue before the body. Fixed in patch #1.

(2) insn_to_jit_off passed to bpf_prog_fill_jited_linfo() is
    calculated in instruction granularity instead of bytes
    granularity. Fixed in patch #2.

Comments are always welcome.

Regards,
Tao

Change Log:
v3:
 * patch #2: explain why bpf2a64_offset() needs update
 * add Fixes tags in both patches

v2: https://lore.kernel.org/bpf/20220125105707.292449-1-houtao1@huawei.com
 * split into two independent patches (from Daniel)
 * use AARCH64_INSN_SIZE instead of defining INSN_SIZE

v1: https://lore.kernel.org/bpf/20220104014236.1512639-1-houtao1@huawei.com

Hou Tao (2):
  bpf, arm64: call build_prologue() first in first JIT pass
  bpf, arm64: calculate offset as byte-offset for bpf line info

 arch/arm64/net/bpf_jit_comp.c | 29 +++++++++++++++++++----------
 1 file changed, 19 insertions(+), 10 deletions(-)

-- 
2.27.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

[PATCH bpf-next v3 1/2] bpf, arm64: call build_prologue() first in first JIT pass

[Hou Tao]

BPF line info needs ctx->offset to be the instruction offset
in the whole jited image instead of the body itself, so also
call build_prologue() first in first JIT pass.

Fixes: 37ab566c178d ("bpf: arm64: Enable arm64 jit to provide bpf_line_info")
Signed-off-by: Hou Tao <houtao1@huawei.com>
---
 arch/arm64/net/bpf_jit_comp.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index 2375ed3e4c8a..68b35c83e637 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -1190,15 +1190,18 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 		goto out_off;
 	}
 
-	/* 1. Initial fake pass to compute ctx->idx. */
-
-	/* Fake pass to fill in ctx->offset. */
-	if (build_body(&ctx, extra_pass)) {
+	/*
+	 * 1. Initial fake pass to compute ctx->idx and ctx->offset.
+	 *
+	 * BPF line info needs ctx->offset[i] to be the offset of
+	 * instruction[i] in jited image, so build prologue first.
+	 */
+	if (build_prologue(&ctx, was_classic)) {
 		prog = orig_prog;
 		goto out_off;
 	}
 
-	if (build_prologue(&ctx, was_classic)) {
+	if (build_body(&ctx, extra_pass)) {
 		prog = orig_prog;
 		goto out_off;
 	}
-- 
2.27.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

[PATCH bpf-next v3 2/2] bpf, arm64: calculate offset as byte-offset for bpf line info

[Hou Tao]

insn_to_jit_off passed to bpf_prog_fill_jited_linfo() is calculated
in instruction granularity instead of bytes granularity, but bpf
line info requires byte offset, so fixing it by calculating ctx->offset
as byte-offset. bpf2a64_offset() needs to return relative instruction
offset by using ctx->offfset, so update it accordingly.

Fixes: 37ab566c178d ("bpf: arm64: Enable arm64 jit to provide bpf_line_info")
Signed-off-by: Hou Tao <houtao1@huawei.com>
---
 arch/arm64/net/bpf_jit_comp.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index 68b35c83e637..aed07cba78ec 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -164,9 +164,14 @@ static inline int bpf2a64_offset(int bpf_insn, int off,
 	/*
 	 * Whereas arm64 branch instructions encode the offset
 	 * from the branch itself, so we must subtract 1 from the
-	 * instruction offset.
+	 * instruction offset. The unit of ctx->offset is byte, so
+	 * subtract AARCH64_INSN_SIZE from it. bpf2a64_offset()
+	 * returns instruction offset, so divide by AARCH64_INSN_SIZE
+	 * at the end.
 	 */
-	return ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1);
+	return (ctx->offset[bpf_insn + off] -
+		(ctx->offset[bpf_insn] - AARCH64_INSN_SIZE)) /
+		AARCH64_INSN_SIZE;
 }
 
 static void jit_fill_hole(void *area, unsigned int size)
@@ -1087,13 +1092,14 @@ static int build_body(struct jit_ctx *ctx, bool extra_pass)
 		const struct bpf_insn *insn = &prog->insnsi[i];
 		int ret;
 
+		/* BPF line info needs byte-offset instead of insn-offset */
 		if (ctx->image == NULL)
-			ctx->offset[i] = ctx->idx;
+			ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;
 		ret = build_insn(insn, ctx, extra_pass);
 		if (ret > 0) {
 			i++;
 			if (ctx->image == NULL)
-				ctx->offset[i] = ctx->idx;
+				ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;
 			continue;
 		}
 		if (ret)
@@ -1105,7 +1111,7 @@ static int build_body(struct jit_ctx *ctx, bool extra_pass)
 	 * instruction (end of program)
 	 */
 	if (ctx->image == NULL)
-		ctx->offset[i] = ctx->idx;
+		ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;
 
 	return 0;
 }
-- 
2.27.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH bpf-next v3 0/2] bpf, arm64: fix bpf line info

[Hou Tao]

ping ?

On 2/8/2022 9:25 AM, Hou Tao wrote:
> Hi,
>
> The patchset addresses two issues in bpf line info for arm64:
>
> (1) insn_to_jit_off only considers the body itself and ignores
>     prologue before the body. Fixed in patch #1.
>
> (2) insn_to_jit_off passed to bpf_prog_fill_jited_linfo() is
>     calculated in instruction granularity instead of bytes
>     granularity. Fixed in patch #2.
>
> Comments are always welcome.
>
> Regards,
> Tao
>
> Change Log:
> v3:
>  * patch #2: explain why bpf2a64_offset() needs update
>  * add Fixes tags in both patches
>
> v2: https://lore.kernel.org/bpf/20220125105707.292449-1-houtao1@huawei.com
>  * split into two independent patches (from Daniel)
>  * use AARCH64_INSN_SIZE instead of defining INSN_SIZE
>
> v1: https://lore.kernel.org/bpf/20220104014236.1512639-1-houtao1@huawei.com
>
> Hou Tao (2):
>   bpf, arm64: call build_prologue() first in first JIT pass
>   bpf, arm64: calculate offset as byte-offset for bpf line info
>
>  arch/arm64/net/bpf_jit_comp.c | 29 +++++++++++++++++++----------
>  1 file changed, 19 insertions(+), 10 deletions(-)
>


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH bpf-next v3 0/2] bpf, arm64: fix bpf line info

[Martin KaFai Lau]

On Tue, Feb 08, 2022 at 09:25:37AM +0800, Hou Tao wrote:
> Hi,
> 
> The patchset addresses two issues in bpf line info for arm64:
> 
> (1) insn_to_jit_off only considers the body itself and ignores
>     prologue before the body. Fixed in patch #1.
> 
> (2) insn_to_jit_off passed to bpf_prog_fill_jited_linfo() is
>     calculated in instruction granularity instead of bytes
>     granularity. Fixed in patch #2.
> 
> Comments are always welcome.
> 
> Regards,
> Tao
> 
> Change Log:
> v3:
>  * patch #2: explain why bpf2a64_offset() needs update
>  * add Fixes tags in both patches
It makes sense to me on the bpf linfo expectation.  It will be
useful to also get some eyes from the arm experts.

Acked-by: Martin KaFai Lau <kafai@fb.com>

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH bpf-next v3 2/2] bpf, arm64: calculate offset as byte-offset for bpf line info

[Daniel Borkmann]

On 2/8/22 2:25 AM, Hou Tao wrote:
> insn_to_jit_off passed to bpf_prog_fill_jited_linfo() is calculated
> in instruction granularity instead of bytes granularity, but bpf
> line info requires byte offset, so fixing it by calculating ctx->offset
> as byte-offset. bpf2a64_offset() needs to return relative instruction
> offset by using ctx->offfset, so update it accordingly.
> 
> Fixes: 37ab566c178d ("bpf: arm64: Enable arm64 jit to provide bpf_line_info")
> Signed-off-by: Hou Tao <houtao1@huawei.com>
> ---
>   arch/arm64/net/bpf_jit_comp.c | 16 +++++++++++-----
>   1 file changed, 11 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
> index 68b35c83e637..aed07cba78ec 100644
> --- a/arch/arm64/net/bpf_jit_comp.c
> +++ b/arch/arm64/net/bpf_jit_comp.c
> @@ -164,9 +164,14 @@ static inline int bpf2a64_offset(int bpf_insn, int off,
>   	/*
>   	 * Whereas arm64 branch instructions encode the offset
>   	 * from the branch itself, so we must subtract 1 from the
> -	 * instruction offset.
> +	 * instruction offset. The unit of ctx->offset is byte, so
> +	 * subtract AARCH64_INSN_SIZE from it. bpf2a64_offset()
> +	 * returns instruction offset, so divide by AARCH64_INSN_SIZE
> +	 * at the end.
>   	 */
> -	return ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1);
> +	return (ctx->offset[bpf_insn + off] -
> +		(ctx->offset[bpf_insn] - AARCH64_INSN_SIZE)) /
> +		AARCH64_INSN_SIZE;
>   }
>   
>   static void jit_fill_hole(void *area, unsigned int size)
> @@ -1087,13 +1092,14 @@ static int build_body(struct jit_ctx *ctx, bool extra_pass)
>   		const struct bpf_insn *insn = &prog->insnsi[i];
>   		int ret;
>   
> +		/* BPF line info needs byte-offset instead of insn-offset */
>   		if (ctx->image == NULL)
> -			ctx->offset[i] = ctx->idx;
> +			ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;
>   		ret = build_insn(insn, ctx, extra_pass);
>   		if (ret > 0) {
>   			i++;
>   			if (ctx->image == NULL)
> -				ctx->offset[i] = ctx->idx;
> +				ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;
>   			continue;
>   		}
>   		if (ret)
> @@ -1105,7 +1111,7 @@ static int build_body(struct jit_ctx *ctx, bool extra_pass)
>   	 * instruction (end of program)
>   	 */
>   	if (ctx->image == NULL)
> -		ctx->offset[i] = ctx->idx;
> +		ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;

Both patches look good to me. For this one specifically, given bpf2a64_offset()
needs to return relative instruction offset via ctx->offfset, can't we just
simplify it like this w/o the AARCH64_INSN_SIZE back/forth dance?

diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index 74f9a9b6a053..72f4702a9d01 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -999,7 +999,7 @@ struct arm64_jit_data {

  struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
  {
-	int image_size, prog_size, extable_size;
+	int image_size, prog_size, extable_size, i;
  	struct bpf_prog *tmp, *orig_prog = prog;
  	struct bpf_binary_header *header;
  	struct arm64_jit_data *jit_data;
@@ -1130,6 +1130,9 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
  	prog->jited_len = prog_size;

  	if (!prog->is_func || extra_pass) {
+		/* BPF line info needs byte-offset instead of insn-offset. */
+		for (i = 0; i < prog->len + 1; i++)
+			ctx.offset[i] *= AARCH64_INSN_SIZE;
  		bpf_prog_fill_jited_linfo(prog, ctx.offset + 1);
  out_off:
  		kfree(ctx.offset);
-- 
2.21.0

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH bpf-next v3 2/2] bpf, arm64: calculate offset as byte-offset for bpf line info

[Hou Tao]

Hi,

On 2/19/2022 7:20 AM, Daniel Borkmann wrote:
> On 2/8/22 2:25 AM, Hou Tao wrote:
>> insn_to_jit_off passed to bpf_prog_fill_jited_linfo() is calculated
>> in instruction granularity instead of bytes granularity, but bpf
>> line info requires byte offset, so fixing it by calculating ctx->offset
>> as byte-offset. bpf2a64_offset() needs to return relative instruction
>> offset by using ctx->offfset, so update it accordingly.
>>
>> Fixes: 37ab566c178d ("bpf: arm64: Enable arm64 jit to provide bpf_line_info")
>> Signed-off-by: Hou Tao <houtao1@huawei.com>
>> ---
>>   arch/arm64/net/bpf_jit_comp.c | 16 +++++++++++-----
>>   1 file changed, 11 insertions(+), 5 deletions(-)
>>
[snip]
>>           if (ret)
>> @@ -1105,7 +1111,7 @@ static int build_body(struct jit_ctx *ctx, bool
>> extra_pass)
>>        * instruction (end of program)
>>        */
>>       if (ctx->image == NULL)
>> -        ctx->offset[i] = ctx->idx;
>> +        ctx->offset[i] = ctx->idx * AARCH64_INSN_SIZE;
>
> Both patches look good to me. For this one specifically, given bpf2a64_offset()
> needs to return relative instruction offset via ctx->offfset, can't we just
> simplify it like this w/o the AARCH64_INSN_SIZE back/forth dance?
>
> diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
> index 74f9a9b6a053..72f4702a9d01 100644
> --- a/arch/arm64/net/bpf_jit_comp.c
> +++ b/arch/arm64/net/bpf_jit_comp.c
> @@ -999,7 +999,7 @@ struct arm64_jit_data {
>
>  struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
>  {
> -    int image_size, prog_size, extable_size;
> +    int image_size, prog_size, extable_size, i;
>      struct bpf_prog *tmp, *orig_prog = prog;
>      struct bpf_binary_header *header;
>      struct arm64_jit_data *jit_data;
> @@ -1130,6 +1130,9 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
>      prog->jited_len = prog_size;
>
>      if (!prog->is_func || extra_pass) {
> +        /* BPF line info needs byte-offset instead of insn-offset. */
> +        for (i = 0; i < prog->len + 1; i++)
> +            ctx.offset[i] *= AARCH64_INSN_SIZE;
>          bpf_prog_fill_jited_linfo(prog, ctx.offset + 1);
>  out_off:
>          kfree(ctx.offset);
The fix is much simpler. I will check whether or not it works.

Thanks,
Tao


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel