Re: [RFC PATCH 2/3] LoongArch: Add larch_insn_gen_break() to generate break insn

From: WANG Xuerui
Date: Mon Apr 10 2023 - 10:16:41 EST


On 2023/4/7 20:02, Tiezhu Yang wrote:


On 04/07/2023 05:51 PM, WANG Xuerui wrote:
On 2023/4/7 10:30, Youling Tang wrote:
/* snip */

diff --git a/arch/loongarch/kernel/kprobes.c
b/arch/loongarch/kernel/kprobes.c
index 08c78d2..a5c3712 100644
--- a/arch/loongarch/kernel/kprobes.c
+++ b/arch/loongarch/kernel/kprobes.c
@@ -4,19 +4,8 @@
 #include <linux/preempt.h>
 #include <asm/break.h>

-static const union loongarch_instruction breakpoint_insn = {
-    .reg0i15_format = {
-        .opcode = break_op,
-        .immediate = BRK_KPROBE_BP,
-    }
-};
-
-static const union loongarch_instruction singlestep_insn = {
-    .reg0i15_format = {
-        .opcode = break_op,
-        .immediate = BRK_KPROBE_SSTEPBP,
-    }
-};
+#define breakpoint_insn larch_insn_gen_break(BRK_KPROBE_BP)
+#define singlestep_insn larch_insn_gen_break(BRK_KPROBE_SSTEPBP)

IMO, Defined as KPROBE_BP_INSN, KPROBE_SSTEPBP_INSN may be better.

Are you suggesting to hardcode the instruction words for those two BREAK
flavors?

I think what Youling said is:

#define KPROBE_BP_INSN         larch_insn_gen_break(BRK_KPROBE_BP)
#define KPROBE_SSTEPBP_INSN    larch_insn_gen_break(BRK_KPROBE_SSTEPBP)

I don't think it's better because even more structured info is
lost, and the compiler would generate the same code (if not, it's the
compiler that's to be fixed).

Actually, I don't know why this commit was necessary in the first place.
For the very least, it consisted of two logical changes (pass around
instruction words instead of unions; and change the BREAK insns to make
them words) that should get split;

Yes, thanks for your suggestion, I will split it into two patches
in the next version.

but again, the generated code should
be identical anyway, so it seems a lot of churn for no benefit and
reduced readability.


Define and use larch_insn_gen_break() is to avoid hardcoding the
uprobe break instruction in patch #3.

We do not like the following definitions:

#define UPROBE_SWBP_INSN    0x002a000c
#define UPROBE_XOLBP_INSN    0x002a000d

Using larch_insn_gen_break() seems better:

#define UPROBE_SWBP_INSN    larch_insn_gen_break(BRK_UPROBE_BP)
#define UPROBE_XOLBP_INSN    larch_insn_gen_break(BRK_UPROBE_XOLBP)

Sorry, I meant *not* ditching the union-typed parameters. IMO they should behave the same codegen-wise (i.e. unchanged performance), and have the benefit of being clearly typed unlike plain u32's.

--
WANG "xen0n" Xuerui

Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/