Skip to contentSkip to conversation

bpftool: Enable aarch64 ISA extensions for JIT disassembly

PATCH Superseded Puranjay Mohan 2 participants 2 messages Mar 11

Subsystems (2)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 bpftool: Enable aarch64 ISA extensions for JIT disassembly

The LLVM disassembler needs ISA extension features enabled to correctly decode instructions from those extensions. On aarch64, without these features, instructions like LSE atomics (e.g. ldaddal) are silently decoded as incorrect instructions and disassembly is truncated.

Use LLVMCreateDisasmCPUFeatures() with "+all" features for aarch64 targets so that the disassembler can handle any instruction the kernel JIT might emit.

Before:

int bench_trigger_uprobe(void * ctx):
bpf_prog_538c6a43d1c6b84c_bench_trigger_uprobe:
; int cpu = bpf_get_smp_processor_id();

   0:   mov     x9, x30
   4:   nop
   8:   stp     x29, x30, [sp, #-16]!
   c:   mov     x29, sp
  10:   stp     xzr, x26, [sp, #-16]!
  14:   mov     x26, sp
  18:   mrs     x10, SP_EL0
  1c:   ldr     w7, [x10, #16]
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  20:   and     w7, w7, #0xff
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  24:   lsl     x7, x7, #7
  28:   mov     x0, #-281474976710656
  2c:   movk    x0, #32768, lsl #32
  30:   movk    x0, #35407, lsl #16
  34:   add     x0, x0, x7
  38:   mov     x1, #1

; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
3c: mov x1, #1

After:

int bench_trigger_uprobe(void * ctx):
bpf_prog_538c6a43d1c6b84c_bench_trigger_uprobe:
; int cpu = bpf_get_smp_processor_id();

   0:   mov     x9, x30
   4:   nop
   8:   stp     x29, x30, [sp, #-16]!
   c:   mov     x29, sp
  10:   stp     xzr, x26, [sp, #-16]!
  14:   mov     x26, sp
  18:   mrs     x10, SP_EL0
  1c:   ldr     w7, [x10, #16]
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  20:   and     w7, w7, #0xff
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  24:   lsl     x7, x7, #7
  28:   mov     x0, #-281474976710656
  2c:   movk    x0, #32768, lsl #32
  30:   movk    x0, #35407, lsl #16
  34:   add     x0, x0, x7
  38:   mov     x1, #1
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  3c:   ldaddal x1, x1, [x0]
; return 0;
  40:   mov     w7, #0
  44:   ldp     xzr, x26, [sp], #16
  48:   ldp     x29, x30, [sp], #16
  4c:   mov     x0, x7
  50:   ret
  54:   nop
  58:   ldr     x10, #8
  5c:   br      x10

---
Changelog:
v2: https://lore.kernel.org/all/[email protected]/
Chnages in v3:
- Fix bug in usage of strncmp() (AI)
v1: https://lore.kernel.org/all/[email protected]/
Changes in v2:
- Fix coding style issue (Quentin)
- Use strncmp() in place of strstr() for detecting aarch64 in triple. (Quentin)

---
 tools/bpf/bpftool/jit_disasm.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/bpf/bpftool/jit_disasm.c +10 -1
index 8895b4e1f690..04541155e9cc 100644
@@ -93,7 +93,16 @@ init_context(disasm_ctx_t *ctx, const char *arch,
 		p_err("Failed to retrieve triple");
 		return -1;
 	}
-	*ctx = LLVMCreateDisasm(triple, NULL, 0, NULL, symbol_lookup_callback);
+
+	/*
+	 * Enable all aarch64 ISA extensions so the disassembler can handle any
+	 * instruction the kernel JIT might emit (e.g. ARM64 LSE atomics).
+	 */
+	if (!strncmp(triple, "aarch64", 7))
+		*ctx = LLVMCreateDisasmCPUFeatures(triple, "", "+all", NULL, 0, NULL,
+						   symbol_lookup_callback);
+	else
+		*ctx = LLVMCreateDisasm(triple, NULL, 0, NULL, symbol_lookup_callback);
 	LLVMDisposeMessage(triple);
 
 	if (!*ctx) {

base-commit: ca0f39a369c5f927c3d004e63a5a778b08a9df94

Signature · 1 line
2.52.0
maintainer Acked-by on v3 1/1
#1 lore raw

2026-03-11 15:26 UTC-0700 ~ Puranjay Mohan <[email protected]>

Quoted 68 lines

The LLVM disassembler needs ISA extension features enabled to correctly decode instructions from those extensions. On aarch64, without these features, instructions like LSE atomics (e.g. ldaddal) are silently decoded as incorrect instructions and disassembly is truncated.

Use LLVMCreateDisasmCPUFeatures() with "+all" features for aarch64 targets so that the disassembler can handle any instruction the kernel JIT might emit.

Before:

int bench_trigger_uprobe(void * ctx):
bpf_prog_538c6a43d1c6b84c_bench_trigger_uprobe:

; int cpu = bpf_get_smp_processor_id();
   0:   mov     x9, x30
   4:   nop
   8:   stp     x29, x30, [sp, #-16]!
   c:   mov     x29, sp
  10:   stp     xzr, x26, [sp, #-16]!
  14:   mov     x26, sp
  18:   mrs     x10, SP_EL0
  1c:   ldr     w7, [x10, #16]
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  20:   and     w7, w7, #0xff
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  24:   lsl     x7, x7, #7
  28:   mov     x0, #-281474976710656
  2c:   movk    x0, #32768, lsl #32
  30:   movk    x0, #35407, lsl #16
  34:   add     x0, x0, x7
  38:   mov     x1, #1

; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
3c: mov x1, #1

After:

int bench_trigger_uprobe(void * ctx):
bpf_prog_538c6a43d1c6b84c_bench_trigger_uprobe:

; int cpu = bpf_get_smp_processor_id();
   0:   mov     x9, x30
   4:   nop
   8:   stp     x29, x30, [sp, #-16]!
   c:   mov     x29, sp
  10:   stp     xzr, x26, [sp, #-16]!
  14:   mov     x26, sp
  18:   mrs     x10, SP_EL0
  1c:   ldr     w7, [x10, #16]
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  20:   and     w7, w7, #0xff
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  24:   lsl     x7, x7, #7
  28:   mov     x0, #-281474976710656
  2c:   movk    x0, #32768, lsl #32
  30:   movk    x0, #35407, lsl #16
  34:   add     x0, x0, x7
  38:   mov     x1, #1
; __sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
  3c:   ldaddal x1, x1, [x0]
; return 0;
  40:   mov     w7, #0
  44:   ldp     xzr, x26, [sp], #16
  48:   ldp     x29, x30, [sp], #16
  4c:   mov     x0, x7
  50:   ret
  54:   nop
  58:   ldr     x10, #8
  5c:   br      x10
Quoted 3 lines

Thank you!