When unwind instruction is 0xb2,the subsequent instructions
are uleb128 bytes.
For now,it uses only the first uleb128 byte in code.

For vsp increments of 0x204~0x400,use one uleb128 byte like below:
0xc06a00e4 <unwind_test_work>: 0x80b27fac
  Compact model index: 0
  0xb2 0x7f vsp = vsp + 1024
  0xac      pop {r4, r5, r6, r7, r8, r14}

For vsp increments larger than 0x400,use two uleb128 bytes like below:
0xc06a00e4 <unwind_test_work>: @0xc0cc9e0c
  Compact model index: 1
  0xb2 0x81 0x01 vsp = vsp + 1032
  0xac      pop {r4, r5, r6, r7, r8, r14}
The unwind works well since the decoded uleb128 byte is also 0x81.

For vsp increments larger than 0x600,use two uleb128 bytes like below:
0xc06a00e4 <unwind_test_work>: @0xc0cc9e0c
  Compact model index: 1
  0xb2 0x81 0x02 vsp = vsp + 1544
  0xac      pop {r4, r5, r6, r7, r8, r14}
In this case,the decoded uleb128 result is 0x101(vsp=0x204+(0x101<<2)).
While the uleb128 used in code is 0x81(vsp=0x204+(0x81<<2)).
The unwind aborts at this frame since it gets incorrect vsp.

To fix this,add uleb128 decode to cover all the above case.

Signed-off-by: Haibo Li <haibo.li@mediatek.com>
---
v3:
-As AngeloGioacchino Del Regno suggested,improve the comment and compress the code

v2:
- As Linus Walleij and Alexandre Mergnat suggested,add comments for unwind_decode_uleb128
- As Alexandre Mergnat suggested,change variables declaration in Alphabetical order
---
 arch/arm/kernel/unwind.c | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/arch/arm/kernel/unwind.c b/arch/arm/kernel/unwind.c
index 53be7ea6181b..9d2192156087 100644
--- a/arch/arm/kernel/unwind.c
+++ b/arch/arm/kernel/unwind.c
@@ -308,6 +308,29 @@ static int unwind_exec_pop_subset_r0_to_r3(struct unwind_ctrl_block *ctrl,
 	return URC_OK;
 }
 
+static unsigned long unwind_decode_uleb128(struct unwind_ctrl_block *ctrl)
+{
+	unsigned long bytes = 0;
+	unsigned long insn;
+	unsigned long result = 0;
+
+	/*
+	 * unwind_get_byte() will advance `ctrl` one instruction at a time, so
+	 * loop until we get an instruction byte where bit 7 is not set.
+	 *
+	 * Note: This decodes a maximum of 4 bytes to output 28 bits data where
+	 * max is 0xfffffff: that will cover a vsp increment of 1073742336, hence
+	 * it is sufficient for unwinding the stack.
+	 */
+	do {
+		insn = unwind_get_byte(ctrl);
+		result |= (insn & 0x7f) << (bytes * 7);
+		bytes++;
+	} while (!!(insn & 0x80) && (bytes != sizeof(result)));
+
+	return result;
+}
+
 /*
  * Execute the current unwind instruction.
  */
@@ -361,7 +384,7 @@ static int unwind_exec_insn(struct unwind_ctrl_block *ctrl)
 		if (ret)
 			goto error;
 	} else if (insn == 0xb2) {
-		unsigned long uleb128 = unwind_get_byte(ctrl);
+		unsigned long uleb128 = unwind_decode_uleb128(ctrl);
 
 		ctrl->vrs[SP] += 0x204 + (uleb128 << 2);
 	} else {
-- 
2.25.1

On Fri, Apr 14, 2023 at 8:10 AM Haibo Li <haibo.li@mediatek.com> wrote:

> When unwind instruction is 0xb2,the subsequent instructions
> are uleb128 bytes.
> For now,it uses only the first uleb128 byte in code.
>
> For vsp increments of 0x204~0x400,use one uleb128 byte like below:
> 0xc06a00e4 <unwind_test_work>: 0x80b27fac
>   Compact model index: 0
>   0xb2 0x7f vsp = vsp + 1024
>   0xac      pop {r4, r5, r6, r7, r8, r14}
>
> For vsp increments larger than 0x400,use two uleb128 bytes like below:
> 0xc06a00e4 <unwind_test_work>: @0xc0cc9e0c
>   Compact model index: 1
>   0xb2 0x81 0x01 vsp = vsp + 1032
>   0xac      pop {r4, r5, r6, r7, r8, r14}
> The unwind works well since the decoded uleb128 byte is also 0x81.
>
> For vsp increments larger than 0x600,use two uleb128 bytes like below:
> 0xc06a00e4 <unwind_test_work>: @0xc0cc9e0c
>   Compact model index: 1
>   0xb2 0x81 0x02 vsp = vsp + 1544
>   0xac      pop {r4, r5, r6, r7, r8, r14}
> In this case,the decoded uleb128 result is 0x101(vsp=0x204+(0x101<<2)).
> While the uleb128 used in code is 0x81(vsp=0x204+(0x81<<2)).
> The unwind aborts at this frame since it gets incorrect vsp.
>
> To fix this,add uleb128 decode to cover all the above case.
>
> Signed-off-by: Haibo Li <haibo.li@mediatek.com>

Way to go, remember to collect all the Reviewed-by tags before
you put the patch into Russell's patch tracker.

Yours,
Linus Walleij

> On Fri, Apr 14, 2023 at 8:10 AM Haibo Li <haibo.li@mediatek.com> wrote:
> 
> > When unwind instruction is 0xb2,the subsequent instructions are
> > uleb128 bytes.
> > For now,it uses only the first uleb128 byte in code.
> >
> > For vsp increments of 0x204~0x400,use one uleb128 byte like below:
> > 0xc06a00e4 <unwind_test_work>: 0x80b27fac
> >   Compact model index: 0
> >   0xb2 0x7f vsp = vsp + 1024
> >   0xac      pop {r4, r5, r6, r7, r8, r14}
> >
> > For vsp increments larger than 0x400,use two uleb128 bytes like below:
> > 0xc06a00e4 <unwind_test_work>: @0xc0cc9e0c
> >   Compact model index: 1
> >   0xb2 0x81 0x01 vsp = vsp + 1032
> >   0xac      pop {r4, r5, r6, r7, r8, r14}
> > The unwind works well since the decoded uleb128 byte is also 0x81.
> >
> > For vsp increments larger than 0x600,use two uleb128 bytes like below:
> > 0xc06a00e4 <unwind_test_work>: @0xc0cc9e0c
> >   Compact model index: 1
> >   0xb2 0x81 0x02 vsp = vsp + 1544
> >   0xac      pop {r4, r5, r6, r7, r8, r14}
> > In this case,the decoded uleb128 result is 0x101(vsp=0x204+(0x101<<2)).
> > While the uleb128 used in code is 0x81(vsp=0x204+(0x81<<2)).
> > The unwind aborts at this frame since it gets incorrect vsp.
> >
> > To fix this,add uleb128 decode to cover all the above case.
> >
> > Signed-off-by: Haibo Li <haibo.li@mediatek.com>
> 
> Way to go, remember to collect all the Reviewed-by tags before you put the
> patch into Russell's patch tracker.
> 
> Yours,
> Linus Walleij
Thanks for reminding.I put the patch into Russell's patch tracker just now,
together with the three Reviewed-by tags from you,
Alexandre Mergnat and AngeloGioacchino Del Regno.
Thank you all.