Merge branch 'bpf-trampoline-support-jmp-mode'

Menglong Dong says:

====================
bpf trampoline support "jmp" mode

For now, the bpf trampoline is called by the "call" instruction. However,
it break the RSB and introduce extra overhead in x86_64 arch.

For example, we hook the function "foo" with fexit, the call and return
logic will be like this:
call foo -> call trampoline -> call foo-body ->
return foo-body -> return foo

As we can see above, there are 3 call, but 2 return, which break the RSB
balance. We can pseudo a "return" here, but it's not the best choice,
as it will still cause once RSB miss:
call foo -> call trampoline -> call foo-body ->
return foo-body -> return dummy -> return foo

The "return dummy" doesn't pair the "call trampoline", which can also
cause the RSB miss.

Therefore, we introduce the "jmp" mode for bpf trampoline, as advised by
Alexei in [1]. And the logic will become this:
call foo -> jmp trampoline -> call foo-body ->
return foo-body -> return foo

As we can see above, the RSB is totally balanced after this series.

In this series, we introduce the FTRACE_OPS_FL_JMP for ftrace to make it
use the "jmp" instruction instead of "call".

And we also do some adjustment to bpf_arch_text_poke() to allow us specify
the old and new poke_type.

For the BPF_TRAMP_F_SHARE_IPMODIFY case, we will fallback to the "call"
mode, as it need to get the function address from the stack, which is not
supported in "jmp" mode.

Before this series, we have the following performance with the bpf
benchmark:

$ cd tools/testing/selftests/bpf
$ ./benchs/run_bench_trigger.sh
usermode-count : 890.171 ± 1.522M/s
kernel-count : 409.184 ± 0.330M/s
syscall-count : 26.792 ± 0.010M/s
fentry : 171.242 ± 0.322M/s
fexit : 80.544 ± 0.045M/s
fmodret : 78.301 ± 0.065M/s
rawtp : 192.906 ± 0.900M/s
tp : 81.883 ± 0.209M/s
kprobe : 52.029 ± 0.113M/s
kprobe-multi : 62.237 ± 0.060M/s
kprobe-multi-all: 4.761 ± 0.014M/s
kretprobe : 23.779 ± 0.046M/s
kretprobe-multi: 29.134 ± 0.012M/s
kretprobe-multi-all: 3.822 ± 0.003M/

And after this series, we have the following performance:

usermode-count : 890.443 ± 0.307M/s
kernel-count : 416.139 ± 0.055M/s
syscall-count : 31.037 ± 0.813M/s
fentry : 169.549 ± 0.519M/s
fexit : 136.540 ± 0.518M/s
fmodret : 159.248 ± 0.188M/s
rawtp : 194.475 ± 0.144M/s
tp : 84.505 ± 0.041M/s
kprobe : 59.951 ± 0.071M/s
kprobe-multi : 63.153 ± 0.177M/s
kprobe-multi-all: 4.699 ± 0.012M/s
kretprobe : 23.740 ± 0.015M/s
kretprobe-multi: 29.301 ± 0.022M/s
kretprobe-multi-all: 3.869 ± 0.005M/s

As we can see above, the performance of fexit increase from 80.544M/s to
136.540M/s, and the "fmodret" increase from 78.301M/s to 159.248M/s.

Link: https://lore.kernel.org/bpf/[e-mail]/
Changes since v2:
* reject if the addr is already "jmp" in register_ftrace_direct() and
__modify_ftrace_direct() in the 1st patch.
* fix compile error in powerpc in the 5th patch.
* changes in the 6th patch:
- fix the compile error by wrapping the write to tr->fops->flags with
CONFIG_DYNAMIC_FTRACE_WITH_JMP
- reset BPF_TRAMP_F_SKIP_FRAME when the second try of modify_fentry in
bpf_trampoline_update()

Link: https://lore.kernel.org/bpf/[e-mail]/
Changes since v1:
* change the bool parameter that we add to save_args() to "u32 flags"
* rename bpf_trampoline_need_jmp() to bpf_trampoline_use_jmp()
* add new function parameter to bpf_arch_text_poke instead of introduce
bpf_arch_text_poke_type()
* rename bpf_text_poke to bpf_trampoline_update_fentry
* remove the BPF_TRAMP_F_JMPED and check the current mode with the origin
flags instead.

Link: https://lore.kernel.org/bpf/[e-mail]/[1]
====================

Acked-by: Steven Rostedt (Google) <[e-mail]>
Link: https://patch.msgid.link/[e-mail]
Signed-off-by: Alexei Starovoitov <[e-mail]>

arch/arm64/net/bpf_jit_comp.c

···
  * The dummy_tramp is used to prevent another CPU from jumping to unknown
  * locations during the patching process, making the patching process easier.
  */
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
-		       void *old_addr, void *new_addr)
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr)
 {
 	int ret;
 	u32 old_insn;
···
 		    !poking_bpf_entry))
 		return -EINVAL;
 
-	if (poke_type == BPF_MOD_CALL)
-		branch_type = AARCH64_INSN_BRANCH_LINK;
-	else
-		branch_type = AARCH64_INSN_BRANCH_NOLINK;
-
+	branch_type = old_t == BPF_MOD_CALL ? AARCH64_INSN_BRANCH_LINK :
+					      AARCH64_INSN_BRANCH_NOLINK;
 	if (gen_branch_or_nop(branch_type, ip, old_addr, plt, &old_insn) < 0)
 		return -EFAULT;
 
+	branch_type = new_t == BPF_MOD_CALL ? AARCH64_INSN_BRANCH_LINK :
+					      AARCH64_INSN_BRANCH_NOLINK;
 	if (gen_branch_or_nop(branch_type, ip, new_addr, plt, &new_insn) < 0)
 		return -EFAULT;
 

arch/loongarch/net/bpf_jit.c

···
 	return ret ? ERR_PTR(-EINVAL) : dst;
 }
 
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
-		       void *old_addr, void *new_addr)
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr)
 {
 	int ret;
-	bool is_call = (poke_type == BPF_MOD_CALL);
+	bool is_call;
 	u32 old_insns[LOONGARCH_LONG_JUMP_NINSNS] = {[0 ... 4] = INSN_NOP};
 	u32 new_insns[LOONGARCH_LONG_JUMP_NINSNS] = {[0 ... 4] = INSN_NOP};
 
···
 	if (!is_bpf_text_address((unsigned long)ip))
 		return -ENOTSUPP;
 
+	is_call = old_t == BPF_MOD_CALL;
 	ret = emit_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;
 
+	is_call = new_t == BPF_MOD_CALL;
 	ret = emit_jump_or_nops(new_addr, ip, new_insns, is_call);
 	if (ret)
 		return ret;

arch/powerpc/net/bpf_jit_comp.c

···
  * execute isync (or some CSI) so that they don't go back into the
  * trampoline again.
  */
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
-		       void *old_addr, void *new_addr)
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr)
 {
 	unsigned long bpf_func, bpf_func_end, size, offset;
 	ppc_inst_t old_inst, new_inst;
···
 		return -EOPNOTSUPP;
 
 	bpf_func = (unsigned long)ip;
-	branch_flags = poke_type == BPF_MOD_CALL ? BRANCH_SET_LINK : 0;
 
 	/* We currently only support poking bpf programs */
 	if (!__bpf_address_lookup(bpf_func, &size, &offset, name)) {
···
 	 * an unconditional branch instruction at im->ip_after_call
 	 */
 	if (offset) {
-		if (poke_type != BPF_MOD_JUMP) {
+		if (old_t == BPF_MOD_CALL || new_t == BPF_MOD_CALL) {
 			pr_err("%s (0x%lx): calls are not supported in bpf prog body\n", __func__,
 			       bpf_func);
 			return -EOPNOTSUPP;
···
 	}
 
 	old_inst = ppc_inst(PPC_RAW_NOP());
+	branch_flags = old_t == BPF_MOD_CALL ? BRANCH_SET_LINK : 0;
 	if (old_addr) {
 		if (is_offset_in_branch_range(ip - old_addr))
 			create_branch(&old_inst, ip, (unsigned long)old_addr, branch_flags);
···
 				      branch_flags);
 	}
 	new_inst = ppc_inst(PPC_RAW_NOP());
+	branch_flags = new_t == BPF_MOD_CALL ? BRANCH_SET_LINK : 0;
 	if (new_addr) {
 		if (is_offset_in_branch_range(ip - new_addr))
 			create_branch(&new_inst, ip, (unsigned long)new_addr, branch_flags);

arch/riscv/net/bpf_jit_comp64.c

···
 	return emit_jump_and_link(is_call ? RV_REG_T0 : RV_REG_ZERO, rvoff, false, &ctx);
 }
 
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
-		       void *old_addr, void *new_addr)
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr)
 {
 	u32 old_insns[RV_FENTRY_NINSNS], new_insns[RV_FENTRY_NINSNS];
-	bool is_call = poke_type == BPF_MOD_CALL;
+	bool is_call;
 	int ret;
 
 	if (!is_kernel_text((unsigned long)ip) &&
 	    !is_bpf_text_address((unsigned long)ip))
 		return -ENOTSUPP;
 
+	is_call = old_t == BPF_MOD_CALL;
 	ret = gen_jump_or_nops(old_addr, ip, old_insns, is_call);
 	if (ret)
 		return ret;
···
 	if (memcmp(ip, old_insns, RV_FENTRY_NBYTES))
 		return -EFAULT;
 
+	is_call = new_t == BPF_MOD_CALL;
 	ret = gen_jump_or_nops(new_addr, ip, new_insns, is_call);
 	if (ret)
 		return ret;
···
 	store_args(nr_arg_slots, args_off, ctx);
 
 	/* skip to actual body of traced function */
-	if (flags & BPF_TRAMP_F_SKIP_FRAME)
+	if (flags & BPF_TRAMP_F_ORIG_STACK)
 		orig_call += RV_FENTRY_NINSNS * 4;
 
 	if (flags & BPF_TRAMP_F_CALL_ORIG) {

arch/s390/net/bpf_jit_comp.c

···
 	return true;
 }
 
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
-		       void *old_addr, void *new_addr)
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr)
 {
 	struct bpf_plt expected_plt, current_plt, new_plt, *plt;
 	struct {
···
 	if (insn.opc != (0xc004 | (old_addr ? 0xf0 : 0)))
 		return -EINVAL;
 
-	if (t == BPF_MOD_JUMP &&
+	if ((new_t == BPF_MOD_JUMP || old_t == BPF_MOD_JUMP) &&
 	    insn.disp == ((char *)new_addr - (char *)ip) >> 1) {
 		/*
 		 * The branch already points to the destination,

arch/x86/Kconfig

···
 	select HAVE_DYNAMIC_FTRACE_WITH_ARGS	if X86_64
 	select HAVE_FTRACE_REGS_HAVING_PT_REGS	if X86_64
 	select HAVE_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
+	select HAVE_DYNAMIC_FTRACE_WITH_JMP	if X86_64
 	select HAVE_SAMPLE_FTRACE_DIRECT	if X86_64
 	select HAVE_SAMPLE_FTRACE_DIRECT_MULTI	if X86_64
 	select HAVE_EBPF_JIT

arch/x86/kernel/ftrace.c

···
 	 * No need to translate into a callthunk. The trampoline does
 	 * the depth accounting itself.
 	 */
-	return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr);
+	if (ftrace_is_jmp(addr)) {
+		addr = ftrace_jmp_get(addr);
+		return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
+	} else {
+		return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr);
+	}
 }
 
 static int ftrace_verify_code(unsigned long ip, const char *old_code)

arch/x86/kernel/ftrace_64.S

···
 	ANNOTATE_NOENDBR
 	RET
 
+1:
+	testb	$1, %al
+	jz	2f
+	andq $0xfffffffffffffffe, %rax
+	movq %rax, MCOUNT_REG_SIZE+8(%rsp)
+	restore_mcount_regs
+	/* Restore flags */
+	popfq
+	RET
+
 	/* Swap the flags with orig_rax */
-1:	movq MCOUNT_REG_SIZE(%rsp), %rdi
+2:	movq MCOUNT_REG_SIZE(%rsp), %rdi
 	movq %rdi, MCOUNT_REG_SIZE-8(%rsp)
 	movq %rax, MCOUNT_REG_SIZE(%rsp)
 

arch/x86/net/bpf_jit_comp.c

···
 	return emit_patch(pprog, func, ip, 0xE9);
 }
 
-static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
+static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+				enum bpf_text_poke_type new_t,
 				void *old_addr, void *new_addr)
 {
 	const u8 *nop_insn = x86_nops[5];
···
 	int ret;
 
 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
-	if (old_addr) {
+	if (old_t != BPF_MOD_NOP && old_addr) {
 		prog = old_insn;
-		ret = t == BPF_MOD_CALL ?
+		ret = old_t == BPF_MOD_CALL ?
 		      emit_call(&prog, old_addr, ip) :
 		      emit_jump(&prog, old_addr, ip);
 		if (ret)
···
 	}
 
 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
-	if (new_addr) {
+	if (new_t != BPF_MOD_NOP && new_addr) {
 		prog = new_insn;
-		ret = t == BPF_MOD_CALL ?
+		ret = new_t == BPF_MOD_CALL ?
 		      emit_call(&prog, new_addr, ip) :
 		      emit_jump(&prog, new_addr, ip);
 		if (ret)
···
 	return ret;
 }
 
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
-		       void *old_addr, void *new_addr)
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr)
 {
 	if (!is_kernel_text((long)ip) &&
 	    !is_bpf_text_address((long)ip))
···
 	if (is_endbr(ip))
 		ip += ENDBR_INSN_SIZE;
 
-	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
+	return __bpf_arch_text_poke(ip, old_t, new_t, old_addr, new_addr);
 }
 
 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
···
 		target = array->ptrs[poke->tail_call.key];
 		if (target) {
 			ret = __bpf_arch_text_poke(poke->tailcall_target,
-						   BPF_MOD_JUMP, NULL,
+						   BPF_MOD_NOP, BPF_MOD_JUMP,
+						   NULL,
 						   (u8 *)target->bpf_func +
 						   poke->adj_off);
 			BUG_ON(ret < 0);
 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
-						   BPF_MOD_JUMP,
+						   BPF_MOD_JUMP, BPF_MOD_NOP,
 						   (u8 *)poke->tailcall_target +
 						   X86_PATCH_SIZE, NULL);
 			BUG_ON(ret < 0);
···
 }
 
 static void save_args(const struct btf_func_model *m, u8 **prog,
-		      int stack_size, bool for_call_origin)
+		      int stack_size, bool for_call_origin, u32 flags)
 {
 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
+	bool use_jmp = bpf_trampoline_use_jmp(flags);
 	int i, j;
 
 	/* Store function arguments to stack.
···
 			 */
 			for (j = 0; j < arg_regs; j++) {
 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
-					 nr_stack_slots * 8 + 0x18);
+					 nr_stack_slots * 8 + 16 + (!use_jmp) * 8);
 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
 					 -stack_size);
 
···
 		 * should be 16-byte aligned. Following code depend on
 		 * that stack_size is already 8-byte aligned.
 		 */
-		stack_size += (stack_size % 16) ? 0 : 8;
+		if (bpf_trampoline_use_jmp(flags)) {
+			/* no rip in the "jmp" case */
+			stack_size += (stack_size % 16) ? 8 : 0;
+		} else {
+			stack_size += (stack_size % 16) ? 0 : 8;
+		}
 	}
 
 	arg_stack_off = stack_size;
 
-	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
+	if (flags & BPF_TRAMP_F_CALL_ORIG) {
 		/* skip patched call instruction and point orig_call to actual
 		 * body of the kernel function.
 		 */
···
 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
 	}
 
-	save_args(m, &prog, regs_off, false);
+	save_args(m, &prog, regs_off, false, flags);
 
 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
 		/* arg1: mov rdi, im */
···
 
 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
 		restore_regs(m, &prog, regs_off);
-		save_args(m, &prog, arg_stack_off, true);
+		save_args(m, &prog, arg_stack_off, true, flags);
 
 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
 			/* Before calling the original function, load the
···
 			       struct bpf_prog *new, struct bpf_prog *old)
 {
 	u8 *old_addr, *new_addr, *old_bypass_addr;
+	enum bpf_text_poke_type t;
 	int ret;
 
 	old_bypass_addr = old ? NULL : poke->bypass_addr;
···
 	 * the kallsyms check.
 	 */
 	if (new) {
+		t = old_addr ? BPF_MOD_JUMP : BPF_MOD_NOP;
 		ret = __bpf_arch_text_poke(poke->tailcall_target,
-					   BPF_MOD_JUMP,
+					   t, BPF_MOD_JUMP,
 					   old_addr, new_addr);
 		BUG_ON(ret < 0);
 		if (!old) {
 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
-						   BPF_MOD_JUMP,
+						   BPF_MOD_JUMP, BPF_MOD_NOP,
 						   poke->bypass_addr,
 						   NULL);
 			BUG_ON(ret < 0);
 		}
 	} else {
+		t = old_bypass_addr ? BPF_MOD_JUMP : BPF_MOD_NOP;
 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
-					   BPF_MOD_JUMP,
-					   old_bypass_addr,
+					   t, BPF_MOD_JUMP, old_bypass_addr,
 					   poke->bypass_addr);
 		BUG_ON(ret < 0);
 		/* let other CPUs finish the execution of program
···
 		 */
 		if (!ret)
 			synchronize_rcu();
+		t = old_addr ? BPF_MOD_JUMP : BPF_MOD_NOP;
 		ret = __bpf_arch_text_poke(poke->tailcall_target,
-					   BPF_MOD_JUMP,
-					   old_addr, NULL);
+					   t, BPF_MOD_NOP, old_addr, NULL);
 		BUG_ON(ret < 0);
 	}
 }

include/linux/bpf.h

···
 bpf_trampoline_enter_t bpf_trampoline_enter(const struct bpf_prog *prog);
 bpf_trampoline_exit_t bpf_trampoline_exit(const struct bpf_prog *prog);
 
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_JMP
+static inline bool bpf_trampoline_use_jmp(u64 flags)
+{
+	return flags & BPF_TRAMP_F_CALL_ORIG && !(flags & BPF_TRAMP_F_SKIP_FRAME);
+}
+#else
+static inline bool bpf_trampoline_use_jmp(u64 flags)
+{
+	return false;
+}
+#endif
+
 struct bpf_ksym {
 	unsigned long		 start;
 	unsigned long		 end;
···
 #endif /* CONFIG_INET */
 
 enum bpf_text_poke_type {
+	BPF_MOD_NOP,
 	BPF_MOD_CALL,
 	BPF_MOD_JUMP,
 };
 
-int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
-		       void *addr1, void *addr2);
+int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+		       enum bpf_text_poke_type new_t, void *old_addr,
+		       void *new_addr);
 
 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
 			       struct bpf_prog *new, struct bpf_prog *old);

include/linux/ftrace.h

···
 	FTRACE_OPS_FL_DIRECT			= BIT(17),
 	FTRACE_OPS_FL_SUBOP			= BIT(18),
 	FTRACE_OPS_FL_GRAPH			= BIT(19),
+	FTRACE_OPS_FL_JMP			= BIT(20),
 };
 
 #ifndef CONFIG_DYNAMIC_FTRACE_WITH_ARGS
···
 static inline void arch_ftrace_set_direct_caller(struct ftrace_regs *fregs,
 						 unsigned long addr) { }
 #endif /* CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS */
+
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_JMP
+static inline bool ftrace_is_jmp(unsigned long addr)
+{
+	return addr & 1;
+}
+
+static inline unsigned long ftrace_jmp_set(unsigned long addr)
+{
+	return addr | 1UL;
+}
+
+static inline unsigned long ftrace_jmp_get(unsigned long addr)
+{
+	return addr & ~1UL;
+}
+#else
+static inline bool ftrace_is_jmp(unsigned long addr)
+{
+	return false;
+}
+
+static inline unsigned long ftrace_jmp_set(unsigned long addr)
+{
+	return addr;
+}
+
+static inline unsigned long ftrace_jmp_get(unsigned long addr)
+{
+	return addr;
+}
+#endif /* CONFIG_DYNAMIC_FTRACE_WITH_JMP */
 
 #ifdef CONFIG_STACK_TRACER
 

kernel/bpf/core.c

···
 	return -EFAULT;
 }
 
-int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
-			      void *addr1, void *addr2)
+int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
+			      enum bpf_text_poke_type new_t, void *old_addr,
+			      void *new_addr)
 {
 	return -ENOTSUPP;
 }

kernel/bpf/trampoline.c

···
 	return tr;
 }
 
-static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
+static int bpf_trampoline_update_fentry(struct bpf_trampoline *tr, u32 orig_flags,
+					void *old_addr, void *new_addr)
 {
+	enum bpf_text_poke_type new_t = BPF_MOD_CALL, old_t = BPF_MOD_CALL;
 	void *ip = tr->func.addr;
+
+	if (!new_addr)
+		new_t = BPF_MOD_NOP;
+	else if (bpf_trampoline_use_jmp(tr->flags))
+		new_t = BPF_MOD_JUMP;
+
+	if (!old_addr)
+		old_t = BPF_MOD_NOP;
+	else if (bpf_trampoline_use_jmp(orig_flags))
+		old_t = BPF_MOD_JUMP;
+
+	return bpf_arch_text_poke(ip, old_t, new_t, old_addr, new_addr);
+}
+
+static int unregister_fentry(struct bpf_trampoline *tr, u32 orig_flags,
+			     void *old_addr)
+{
 	int ret;
 
 	if (tr->func.ftrace_managed)
 		ret = unregister_ftrace_direct(tr->fops, (long)old_addr, false);
 	else
-		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
+		ret = bpf_trampoline_update_fentry(tr, orig_flags, old_addr, NULL);
 
 	return ret;
 }
 
-static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr,
+static int modify_fentry(struct bpf_trampoline *tr, u32 orig_flags,
+			 void *old_addr, void *new_addr,
 			 bool lock_direct_mutex)
 {
-	void *ip = tr->func.addr;
 	int ret;
 
 	if (tr->func.ftrace_managed) {
···
 		else
 			ret = modify_ftrace_direct_nolock(tr->fops, (long)new_addr);
 	} else {
-		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
+		ret = bpf_trampoline_update_fentry(tr, orig_flags, old_addr,
+						   new_addr);
 	}
 	return ret;
 }
···
 			return ret;
 		ret = register_ftrace_direct(tr->fops, (long)new_addr);
 	} else {
-		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
+		ret = bpf_trampoline_update_fentry(tr, 0, NULL, new_addr);
 	}
 
 	return ret;
···
 	 * call_rcu_tasks() is not necessary.
 	 */
 	if (im->ip_after_call) {
-		int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_JUMP,
-					     NULL, im->ip_epilogue);
+		int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_NOP,
+					     BPF_MOD_JUMP, NULL,
+					     im->ip_epilogue);
 		WARN_ON(err);
 		if (IS_ENABLED(CONFIG_TASKS_RCU))
 			call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
···
 		return PTR_ERR(tlinks);
 
 	if (total == 0) {
-		err = unregister_fentry(tr, tr->cur_image->image);
+		err = unregister_fentry(tr, orig_flags, tr->cur_image->image);
 		bpf_tramp_image_put(tr->cur_image);
 		tr->cur_image = NULL;
 		goto out;
···
 
 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 again:
-	if ((tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) &&
-	    (tr->flags & BPF_TRAMP_F_CALL_ORIG))
-		tr->flags |= BPF_TRAMP_F_ORIG_STACK;
+	if (tr->flags & BPF_TRAMP_F_CALL_ORIG) {
+		if (tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) {
+			/* The BPF_TRAMP_F_SKIP_FRAME can be cleared in the
+			 * first try, reset it in the second try.
+			 */
+			tr->flags |= BPF_TRAMP_F_ORIG_STACK | BPF_TRAMP_F_SKIP_FRAME;
+		} else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_JMP)) {
+			/* Use "jmp" instead of "call" for the trampoline
+			 * in the origin call case, and we don't need to
+			 * skip the frame.
+			 */
+			tr->flags &= ~BPF_TRAMP_F_SKIP_FRAME;
+		}
+	}
 #endif
 
 	size = arch_bpf_trampoline_size(&tr->func.model, tr->flags,
···
 	if (err)
 		goto out_free;
 
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_JMP
+	if (bpf_trampoline_use_jmp(tr->flags))
+		tr->fops->flags |= FTRACE_OPS_FL_JMP;
+	else
+		tr->fops->flags &= ~FTRACE_OPS_FL_JMP;
+#endif
+
 	WARN_ON(tr->cur_image && total == 0);
 	if (tr->cur_image)
 		/* progs already running at this address */
-		err = modify_fentry(tr, tr->cur_image->image, im->image, lock_direct_mutex);
+		err = modify_fentry(tr, orig_flags, tr->cur_image->image,
+				    im->image, lock_direct_mutex);
 	else
 		/* first time registering */
 		err = register_fentry(tr, im->image);
···
 	tr->cur_image = im;
 out:
 	/* If any error happens, restore previous flags */
-	if (err)
+	if (err) {
 		tr->flags = orig_flags;
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_JMP
+		if (bpf_trampoline_use_jmp(tr->flags))
+			tr->fops->flags |= FTRACE_OPS_FL_JMP;
+		else
+			tr->fops->flags &= ~FTRACE_OPS_FL_JMP;
+#endif
+	}
 	kfree(tlinks);
 	return err;
 
···
 		if (err)
 			return err;
 		tr->extension_prog = link->link.prog;
-		return bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
+		return bpf_arch_text_poke(tr->func.addr, BPF_MOD_NOP,
+					  BPF_MOD_JUMP, NULL,
 					  link->link.prog->bpf_func);
 	}
 	if (cnt >= BPF_MAX_TRAMP_LINKS)
···
 	if (kind == BPF_TRAMP_REPLACE) {
 		WARN_ON_ONCE(!tr->extension_prog);
 		err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
+					 BPF_MOD_NOP,
 					 tr->extension_prog->bpf_func, NULL);
 		tr->extension_prog = NULL;
 		guard(mutex)(&tgt_prog->aux->ext_mutex);

kernel/trace/Kconfig

···
 	  If the architecture generates __patchable_function_entries sections
 	  but does not want them included in the ftrace locations.
 
+config HAVE_DYNAMIC_FTRACE_WITH_JMP
+	bool
+	help
+	  If the architecture supports to replace the __fentry__ with a
+	  "jmp" instruction.
+
 config HAVE_SYSCALL_TRACEPOINTS
 	bool
 	help
···
 	def_bool y
 	depends on DYNAMIC_FTRACE
 	depends on HAVE_DYNAMIC_FTRACE_WITH_ARGS
+
+config DYNAMIC_FTRACE_WITH_JMP
+	def_bool y
+	depends on DYNAMIC_FTRACE
+	depends on DYNAMIC_FTRACE_WITH_DIRECT_CALLS
+	depends on HAVE_DYNAMIC_FTRACE_WITH_JMP
 
 config FPROBE
 	bool "Kernel Function Probe (fprobe)"

kernel/trace/ftrace.c

···
 	for (i = 0; i < size; i++) {
 		hlist_for_each_entry(entry, &hash->buckets[i], hlist) {
 			del = __ftrace_lookup_ip(direct_functions, entry->ip);
-			if (del && del->direct == addr) {
+			if (del && ftrace_jmp_get(del->direct) ==
+				   ftrace_jmp_get(addr)) {
 				remove_hash_entry(direct_functions, del);
 				kfree(del);
 			}
···
 	if (ftrace_hash_empty(hash))
 		return -EINVAL;
 
+	/* This is a "raw" address, and this should never happen. */
+	if (WARN_ON_ONCE(ftrace_is_jmp(addr)))
+		return -EINVAL;
+
 	mutex_lock(&direct_mutex);
+
+	if (ops->flags & FTRACE_OPS_FL_JMP)
+		addr = ftrace_jmp_set(addr);
 
 	/* Make sure requested entries are not already registered.. */
 	size = 1 << hash->size_bits;
···
 	int err;
 
 	lockdep_assert_held_once(&direct_mutex);
+
+	/* This is a "raw" address, and this should never happen. */
+	if (WARN_ON_ONCE(ftrace_is_jmp(addr)))
+		return -EINVAL;
+
+	if (ops->flags & FTRACE_OPS_FL_JMP)
+		addr = ftrace_jmp_set(addr);
 
 	/* Enable the tmp_ops to have the same functions as the direct ops */
 	ftrace_ops_init(&tmp_ops);