Re: [PATCH v5 3/5] LoongArch: BPF: Implement dynamic code modification support

[Chenghao Duan <duanchenghao@kylinos.cn>]

On Tue, Aug 05, 2025 at 12:10:05PM +0800, Huacai Chen wrote:
> On Mon, Aug 4, 2025 at 10:02 AM Hengqi Chen <hengqi.chen@gmail.com> wrote:
> >
> > On Wed, Jul 30, 2025 at 9:13 PM Chenghao Duan <duanchenghao@kylinos.cn> wrote:
> > >
> > > This commit adds support for BPF dynamic code modification on the
> > > LoongArch architecture.:
> > > 1. Implement bpf_arch_text_poke() for runtime instruction patching.
> > > 2. Add bpf_arch_text_copy() for instruction block copying.
> > > 3. Create bpf_arch_text_invalidate() for code invalidation.
> > >
> > > On LoongArch, since symbol addresses in the direct mapping
> > > region cannot be reached via relative jump instructions from the paged
> > > mapping region, we use the move_imm+jirl instruction pair as absolute
> > > jump instructions. These require 2-5 instructions, so we reserve 5 NOP
> > > instructions in the program as placeholders for function jumps.
> > >
> > > larch_insn_text_copy is solely used for BPF. The use of
> > > larch_insn_text_copy() requires page_size alignment. Currently, only
> > > the size of the trampoline is page-aligned.
> > >
> > > Co-developed-by: George Guo <guodongtai@kylinos.cn>
> > > Signed-off-by: George Guo <guodongtai@kylinos.cn>
> > > Signed-off-by: Chenghao Duan <duanchenghao@kylinos.cn>
> > > ---
> > >  arch/loongarch/include/asm/inst.h |   1 +
> > >  arch/loongarch/kernel/inst.c      |  27 ++++++++
> > >  arch/loongarch/net/bpf_jit.c      | 104 ++++++++++++++++++++++++++++++
> > >  3 files changed, 132 insertions(+)
> > >
> > > diff --git a/arch/loongarch/include/asm/inst.h b/arch/loongarch/include/asm/inst.h
> > > index 2ae96a35d..88bb73e46 100644
> > > --- a/arch/loongarch/include/asm/inst.h
> > > +++ b/arch/loongarch/include/asm/inst.h
> > > @@ -497,6 +497,7 @@ void arch_simulate_insn(union loongarch_instruction insn, struct pt_regs *regs);
> > >  int larch_insn_read(void *addr, u32 *insnp);
> > >  int larch_insn_write(void *addr, u32 insn);
> > >  int larch_insn_patch_text(void *addr, u32 insn);
> > > +int larch_insn_text_copy(void *dst, void *src, size_t len);
> > >
> > >  u32 larch_insn_gen_nop(void);
> > >  u32 larch_insn_gen_b(unsigned long pc, unsigned long dest);
> > > diff --git a/arch/loongarch/kernel/inst.c b/arch/loongarch/kernel/inst.c
> > > index 674e3b322..7df63a950 100644
> > > --- a/arch/loongarch/kernel/inst.c
> > > +++ b/arch/loongarch/kernel/inst.c
> > > @@ -4,6 +4,7 @@
> > >   */
> > >  #include <linux/sizes.h>
> > >  #include <linux/uaccess.h>
> > > +#include <linux/set_memory.h>
> > >
> > >  #include <asm/cacheflush.h>
> > >  #include <asm/inst.h>
> > > @@ -218,6 +219,32 @@ int larch_insn_patch_text(void *addr, u32 insn)
> > >         return ret;
> > >  }
> > >
> > > +int larch_insn_text_copy(void *dst, void *src, size_t len)
> > > +{
> > > +       int ret;
> > > +       unsigned long flags;
> > > +       unsigned long dst_start, dst_end, dst_len;
> > > +
> > > +       dst_start = round_down((unsigned long)dst, PAGE_SIZE);
> > > +       dst_end = round_up((unsigned long)dst + len, PAGE_SIZE);
> > > +       dst_len = dst_end - dst_start;
> > > +
> > > +       set_memory_rw(dst_start, dst_len / PAGE_SIZE);
> > > +       raw_spin_lock_irqsave(&patch_lock, flags);
> > > +
> > > +       ret = copy_to_kernel_nofault(dst, src, len);
> > > +       if (ret)
> > > +               pr_err("%s: operation failed\n", __func__);
> > > +
> > > +       raw_spin_unlock_irqrestore(&patch_lock, flags);
> > > +       set_memory_rox(dst_start, dst_len / PAGE_SIZE);
> > > +
> > > +       if (!ret)
> > > +               flush_icache_range((unsigned long)dst, (unsigned long)dst + len);
> > > +
> > > +       return ret;
> > > +}
> > > +
> > >  u32 larch_insn_gen_nop(void)
> > >  {
> > >         return INSN_NOP;
> > > diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
> > > index 7032f11d3..5e6ae7e0e 100644
> > > --- a/arch/loongarch/net/bpf_jit.c
> > > +++ b/arch/loongarch/net/bpf_jit.c
> > > @@ -4,8 +4,12 @@
> > >   *
> > >   * Copyright (C) 2022 Loongson Technology Corporation Limited
> > >   */
> > > +#include <linux/memory.h>
> > >  #include "bpf_jit.h"
> > >
> > > +#define LOONGARCH_LONG_JUMP_NINSNS 5
> > > +#define LOONGARCH_LONG_JUMP_NBYTES (LOONGARCH_LONG_JUMP_NINSNS * 4)
> > > +
> > >  #define REG_TCC                LOONGARCH_GPR_A6
> > >  #define TCC_SAVED      LOONGARCH_GPR_S5
> > >
> > > @@ -88,6 +92,7 @@ static u8 tail_call_reg(struct jit_ctx *ctx)
> > >   */
> > >  static void build_prologue(struct jit_ctx *ctx)
> > >  {
> > > +       int i;
> > >         int stack_adjust = 0, store_offset, bpf_stack_adjust;
> > >
> > >         bpf_stack_adjust = round_up(ctx->prog->aux->stack_depth, 16);
> > > @@ -98,6 +103,10 @@ static void build_prologue(struct jit_ctx *ctx)
> > >         stack_adjust = round_up(stack_adjust, 16);
> > >         stack_adjust += bpf_stack_adjust;
> > >
> > > +       /* Reserve space for the move_imm + jirl instruction */
> > > +       for (i = 0; i < LOONGARCH_LONG_JUMP_NINSNS; i++)
> > > +               emit_insn(ctx, nop);
> > > +
> > >         /*
> > >          * First instruction initializes the tail call count (TCC).
> > >          * On tail call we skip this instruction, and the TCC is
> > > @@ -1367,3 +1376,98 @@ bool bpf_jit_supports_subprog_tailcalls(void)
> > >  {
> > >         return true;
> > >  }
> > > +
> > > +static int emit_jump_and_link(struct jit_ctx *ctx, u8 rd, u64 target)
> > > +{
> > > +       if (!target) {
> > > +               pr_err("bpf_jit: jump target address is error\n");
> > > +               return -EFAULT;
> > > +       }
> > > +
> > > +       move_imm(ctx, LOONGARCH_GPR_T1, target, false);
> > > +       emit_insn(ctx, jirl, rd, LOONGARCH_GPR_T1, 0);
> > > +
> > > +       return 0;
> > > +}
> > > +
> > > +static int gen_jump_or_nops(void *target, void *ip, u32 *insns, bool is_call)
> > > +{
> > > +       struct jit_ctx ctx;
> > > +
> > > +       ctx.idx = 0;
> > > +       ctx.image = (union loongarch_instruction *)insns;
> > > +
> > > +       if (!target) {
> > > +               emit_insn((&ctx), nop);
> > > +               emit_insn((&ctx), nop);
> >
> > There should be 5 nops, no ?
> Chenghao,
> 
> We have already fixed the concurrent problem, now this is the only
> issue, please reply tas soon as possible.
> 
> Huacai

Hi Hengqi & Huacai,

I'm sorry I just saw the email.
This position can be configured with 5 NOP instructions, and I have
tested it successfully.

sudo ./test_progs -a fentry_test/fentry
sudo ./test_progs -a fexit_test/fexit
sudo ./test_progs -a fentry_fexit
sudo ./test_progs -a modify_return
sudo ./test_progs -a fexit_sleep
sudo ./test_progs -a test_overhead
sudo ./test_progs -a trampoline_count
sudo ./test_progs -a fexit_bpf2bpf

if (!target) {
	int i;
	for (i = 0; i < LOONGARCH_LONG_JUMP_NINSNS; i++)
		emit_insn((&ctx), nop);
	return 0;
}


Chenghao

> 
> >
> > > +               return 0;
> > > +       }
> > > +
> > > +       return emit_jump_and_link(&ctx, is_call ? LOONGARCH_GPR_T0 : LOONGARCH_GPR_ZERO,
> > > +                                 (unsigned long)target);
> > > +}
> > > +
> > > +int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
> > > +                      void *old_addr, void *new_addr)
> > > +{
> > > +       u32 old_insns[LOONGARCH_LONG_JUMP_NINSNS] = {[0 ... 4] = INSN_NOP};
> > > +       u32 new_insns[LOONGARCH_LONG_JUMP_NINSNS] = {[0 ... 4] = INSN_NOP};
> > > +       bool is_call = poke_type == BPF_MOD_CALL;
> > > +       int ret;
> > > +
> > > +       if (!is_kernel_text((unsigned long)ip) &&
> > > +               !is_bpf_text_address((unsigned long)ip))
> > > +               return -ENOTSUPP;
> > > +
> > > +       ret = gen_jump_or_nops(old_addr, ip, old_insns, is_call);
> > > +       if (ret)
> > > +               return ret;
> > > +
> > > +       if (memcmp(ip, old_insns, LOONGARCH_LONG_JUMP_NBYTES))
> > > +               return -EFAULT;
> > > +
> > > +       ret = gen_jump_or_nops(new_addr, ip, new_insns, is_call);
> > > +       if (ret)
> > > +               return ret;
> > > +
> > > +       mutex_lock(&text_mutex);
> > > +       if (memcmp(ip, new_insns, LOONGARCH_LONG_JUMP_NBYTES))
> > > +               ret = larch_insn_text_copy(ip, new_insns, LOONGARCH_LONG_JUMP_NBYTES);
> > > +       mutex_unlock(&text_mutex);
> > > +       return ret;
> > > +}
> > > +
> > > +int bpf_arch_text_invalidate(void *dst, size_t len)
> > > +{
> > > +       int i;
> > > +       int ret = 0;
> > > +       u32 *inst;
> > > +
> > > +       inst = kvmalloc(len, GFP_KERNEL);
> > > +       if (!inst)
> > > +               return -ENOMEM;
> > > +
> > > +       for (i = 0; i < (len/sizeof(u32)); i++)
> > > +               inst[i] = INSN_BREAK;
> > > +
> > > +       mutex_lock(&text_mutex);
> > > +       if (larch_insn_text_copy(dst, inst, len))
> > > +               ret = -EINVAL;
> > > +       mutex_unlock(&text_mutex);
> > > +
> > > +       kvfree(inst);
> > > +       return ret;
> > > +}
> > > +
> > > +void *bpf_arch_text_copy(void *dst, void *src, size_t len)
> > > +{
> > > +       int ret;
> > > +
> > > +       mutex_lock(&text_mutex);
> > > +       ret = larch_insn_text_copy(dst, src, len);
> > > +       mutex_unlock(&text_mutex);
> > > +       if (ret)
> > > +               return ERR_PTR(-EINVAL);
> > > +
> > > +       return dst;
> > > +}
> > > --
> >
> > bpf_arch_text_invalidate() and bpf_arch_text_copy() is not related to
> > BPF trampoline, right ?

From the perspective of BPF core source code calls, the two functions
bpf_arch_text_invalidate() and bpf_arch_text_copy() are not only used for
trampolines.

> >
> > > 2.25.1
> > >