Re: [RFC PATCH 21/25] kvx: Add support for ftrace

[Mark Rutland <mark.rutland@arm.com>]

Hi,

As others have said, I think it'd be worth splitting this series up, with this
being part of a follow-up, since it's not vital to boot and use the kernel.
That'll aid review of the smaller initial series, and later we can focus on the
bits and pieces here.

That said, I have a few comments below.

On Tue, Jan 03, 2023 at 05:43:55PM +0100, Yann Sionneau wrote:
> Add support for ftrace to kvx arch.
> 
> CC: Steven Rostedt <rostedt@goodmis.org>
> CC: Masami Hiramatsu <mhiramat@kernel.org>
> CC: Mark Rutland <mark.rutland@arm.com>
> CC: linux-kernel@vger.kernel.org
> Co-developed-by: Clement Leger <clement.leger@bootlin.com>
> Signed-off-by: Clement Leger <clement.leger@bootlin.com>
> Co-developed-by: Guillaume Thouvenin <gthouvenin@kalray.eu>
> Signed-off-by: Guillaume Thouvenin <gthouvenin@kalray.eu>
> Co-developed-by: Julian Vetter <jvetter@kalray.eu>
> Signed-off-by: Julian Vetter <jvetter@kalray.eu>
> Co-developed-by: Marius Gligor <mgligor@kalray.eu>
> Signed-off-by: Marius Gligor <mgligor@kalray.eu>
> Co-developed-by: Yann Sionneau <ysionneau@kalray.eu>
> Signed-off-by: Yann Sionneau <ysionneau@kalray.eu>
> ---
>  arch/kvx/include/asm/ftrace.h    |  41 ++++
>  arch/kvx/kernel/ftrace.c         | 339 ++++++++++++++++++++++++++++++
>  arch/kvx/kernel/mcount.S         | 340 +++++++++++++++++++++++++++++++
>  arch/kvx/kernel/return_address.c |  55 +++++
>  4 files changed, 775 insertions(+)
>  create mode 100644 arch/kvx/include/asm/ftrace.h
>  create mode 100644 arch/kvx/kernel/ftrace.c
>  create mode 100644 arch/kvx/kernel/mcount.S
>  create mode 100644 arch/kvx/kernel/return_address.c
> 
> diff --git a/arch/kvx/include/asm/ftrace.h b/arch/kvx/include/asm/ftrace.h
> new file mode 100644
> index 000000000000..f7bcb325d255
> --- /dev/null
> +++ b/arch/kvx/include/asm/ftrace.h
> @@ -0,0 +1,41 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Guillaume Thouvenin
> + *            Clement Leger
> + */
> +
> +#ifndef _ASM_KVX_FTRACE_H
> +#define _ASM_KVX_FTRACE_H
> +
> +#include <asm/insns_defs.h>

The <asm/insns_defs.h> header file gets added by a subsequent patch, in:

  "[RFC PATCH 23/25] kvx: Add debugging related support"

Obviously that means it won't build as-is, but it's also confusing for
reviewers.

> +
> +#define INSN_MAKE_IMM64_SYLLABLE_SIZE  INSN_SIZE(MAKE_IMM64)

IIUC this is 12 bytes.

> +#define INSN_ICALL_SYLLABLE_SIZE       INSN_SIZE(ICALL)
> +#define INSN_IGOTO_SYLLABLE_SIZE       INSN_SIZE(IGOTO)
> +#define INSN_CALL_SYLLABLE_SIZE        INSN_SIZE(CALL)
> +#define INSN_NOP_SYLLABLE_SIZE         INSN_SIZE(NOP)

IIUC all these are all 4 bytes.

> +
> +#define INSN_ICALL_REG_MASK  0x3f
> +
> +#define MCOUNT_ADDR	((unsigned long)(__mcount))
> +#define MCOUNT_INSN_SIZE   INSN_CALL_SYLLABLE_SIZE /* sizeof mcount call */

Does your GCC port support -fpatchable-function-entry?

If it does, I would strongly recommend moving over to using that, as it'll
allow you to implement FTRACE_WITH_ARGS and all the fancy features dependent
upon that.

If it doesn't, I'd strongly recommend implementing that on the GCC side now so
that you can make use of it in future without having to retain support for an
mcount based ftrace. That'll make things easier to maintain going forwards, and
it'll give you a lot of flexibility w.r.t. your ftrace implementation.

> +extern void ftrace_graph_call(void);
> +
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +extern unsigned long ftrace_call_adjust(unsigned long addr);
> +struct dyn_arch_ftrace {
> +	unsigned int insn;
> +};
> +#endif
> +
> +extern void *return_address(unsigned int level);
> +#define ftrace_return_address(n) return_address(n)
> +
> +#ifdef CONFIG_FUNCTION_TRACER
> +extern void __mcount(void);
> +#define mcount __mcount
> +#endif /* CONFIG_FUNCTION_TRACER */
> +
> +#endif /* _ASM_KVX_FTRACE_H */
> diff --git a/arch/kvx/kernel/ftrace.c b/arch/kvx/kernel/ftrace.c
> new file mode 100644
> index 000000000000..4c9e3ef62714
> --- /dev/null
> +++ b/arch/kvx/kernel/ftrace.c
> @@ -0,0 +1,339 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Guillaume Thouvenin
> + *            Marius Gligor
> + *            Clement Leger
> + */
> +
> +#include <linux/ftrace.h>
> +#include <linux/atomic.h>
> +#include <linux/stop_machine.h>
> +#include <asm/insns.h>
> +#include <asm/insns_defs.h>
> +#include <asm/cacheflush.h>
> +
> +/* The longest insns we check are for the far call: make + icall */
> +#define MAX_SYLLABLES_TO_CHECK (KVX_INSN_MAKE_IMM64_SIZE + INSN_ICALL_SYLLABLE_SIZE)

What's the range on your CALL instruction? Is that smaller than your kernel
image size, or is this just for modules?

> +
> +static int read_insns_and_check(u32 *insns, u8 insns_len, u32 *addr)
> +{
> +	u32 insns_read[MAX_SYLLABLES_TO_CHECK];
> +	int syllables = insns_len / KVX_INSN_SYLLABLE_WIDTH;
> +	int i;
> +
> +	if (syllables > MAX_SYLLABLES_TO_CHECK) {
> +		pr_err("%s: shouldn't have more than %d syllables to check\n",
> +		       __func__, MAX_SYLLABLES_TO_CHECK);
> +		return -EINVAL;
> +	}
> +
> +	if (kvx_insns_read(insns_read, insns_len, addr)) {
> +		pr_err("%s: error when trying to read syllable\n", __func__);
> +		return -EFAULT;
> +	}
> +
> +	for (i = 0; i < syllables; i++) {
> +		if (insns[i] != insns_read[i]) {
> +			pr_err("%s: Instruction verification failed at PC 0x%lx\n",
> +			       __func__,
> +			       (unsigned long)addr + i * KVX_INSN_SYLLABLE_WIDTH);
> +			pr_err("%s: \tExpect  0x%x\n", __func__, insns[i]);
> +			pr_err("%s: \tRead    0x%x\n", __func__, insns_read[i]);
> +			return -EFAULT;
> +		}
> +	}
> +
> +	return 0;
> +}
> +
> +static int write_insns_and_check(u32 *insns, u8 insns_len, u32 *insn_addr)
> +{
> +	int ret;
> +
> +	ret = kvx_insns_write_nostop(insns, insns_len, insn_addr);
> +	if (ret)
> +		return ret;

Again, kvx_insns_write_nostop() is added by a subsequenbt patch, which makes
this painful to review.

AFAICT this assumees that you can concurrently modify instructions while those
are being executed. Does your architecture provide a strong guanratee there?
For example, on arm64 we can only do this safely for a defined small set of
instructions (including B/BL/BLR/NOP), and even then it doesn't gurantee that
updates to multiple instructions will be seen in order.

I see that ftrace_enable_ftrace_graph_caller() writes multiple instructions,
which strikes me as likely to be unsafe depending on the ordering and atomicity
of both those writes and the instruction fetches which will consume them.

I see that for arch_ftrace_update_code() you use stop_machine(), and I'd expect
similar here, but maybe I've missed something that inhibits concurrent
exxecution?

> +
> +	/* Check that what have been written is correct. */
> +	return read_insns_and_check(insns, insns_len, insn_addr);
> +}
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
> +			   unsigned long frame_pointer)
> +{
> +	unsigned long return_hooker = (unsigned long)&return_to_handler;
> +	unsigned long old;
> +
> +	if (unlikely(atomic_read(&current->tracing_graph_pause)))
> +		return;
> +
> +	old = *parent;
> +	*parent = return_hooker;
> +
> +	if (function_graph_enter(old, self_addr, frame_pointer, NULL))
> +		*parent = old;
> +}
> +
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +int ftrace_enable_ftrace_graph_caller(void)
> +{
> +	unsigned int insn[KVX_INSN_MAKE_IMM64_SIZE + KVX_INSN_IGOTO_SIZE];
> +	void *ip = (void *)ftrace_call;
> +
> +	KVX_INSN_MAKE_IMM64(insn, KVX_INSN_PARALLEL_EOB, KVX_REG_R32,
> +			    (unsigned long)&ftrace_graph_call);
> +	KVX_INSN_IGOTO(&insn[KVX_INSN_MAKE_IMM64_SIZE],
> +		       KVX_INSN_PARALLEL_EOB,
> +		       KVX_REG_R32);
> +
> +	return write_insns_and_check(insn,
> +				     INSN_MAKE_IMM64_SYLLABLE_SIZE
> +				     + INSN_IGOTO_SYLLABLE_SIZE,
> +				     ip);
> +}
> +
> +int ftrace_disable_ftrace_graph_caller(void)
> +{
> +	unsigned int nop;
> +	void *ip = (void *)ftrace_call;
> +
> +	KVX_INSN_NOP(&nop, KVX_INSN_PARALLEL_EOB);
> +	return write_insns_and_check(&nop,
> +				     INSN_NOP_SYLLABLE_SIZE,
> +				     ip + INSN_MAKE_IMM64_SYLLABLE_SIZE);
> +}
> +#endif /* CONFIG_DYNAMIC_FTRACE */
> +
> +#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
> +
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +struct kvx_ftrace_modify_param {
> +	atomic_t	cpu_ack;
> +	int		cpu_master;
> +	int		cmd;
> +};
> +
> +static int __ftrace_modify_code_kvx(void *data)
> +{
> +	struct kvx_ftrace_modify_param *mp = data;
> +	int no_cpus = num_online_cpus();
> +	int cpu = smp_processor_id();
> +
> +	if (cpu == mp->cpu_master) {
> +		ftrace_modify_all_code(mp->cmd);
> +
> +		/* Inform the other cpus that they can invalidate I-cache */
> +		atomic_inc(&mp->cpu_ack);
> +
> +		/* param (which contains the cpu_ack counter) is allocated on the
> +		 * master stack: the master must wait for all other CPUs to leave
> +		 * this function before returning and releasing the stack
> +		 * allocation.
> +		 */
> +		while (atomic_read(&mp->cpu_ack) < no_cpus)
> +			cpu_relax();
> +	} else {
> +		/* Wait for the master cpu to finish the code modification */
> +		while (atomic_read(&mp->cpu_ack) == 0)
> +			cpu_relax();
> +		atomic_inc(&mp->cpu_ack);
> +
> +		l1_inval_icache_all();

Wouldn't it make more sense to do the icache invalidate *before* signalling the
master?

This function (and its callees) will need to be noinstr to be safe. There's
some latent issues there on arm64 which I'm slowly working through at the
moment, but there's no need for you to start with the same problems.

> +	}
> +
> +	return 0;
> +}
> +
> +void arch_ftrace_update_code(int command)
> +{
> +	const struct cpumask *cpumask = cpu_online_mask;

Below you use cpu_online_mask directly, so this isn't necessary.

> +	struct kvx_ftrace_modify_param mp = {
> +		.cpu_ack = ATOMIC_INIT(0),
> +		.cpu_master = smp_processor_id(),
> +		.cmd = command,
> +	};
> +
> +	stop_machine(__ftrace_modify_code_kvx, &mp, cpu_online_mask);
> +}
> +
> +unsigned long ftrace_call_adjust(unsigned long addr)
> +{
> +	/*
> +	 * Module are using far call and kernel functions are using
> +	 * pcrel. If it is a call we don't need to adjust the address but
> +	 * if it is an icall the address is on the make. The generated code
> +	 * looks like:
> +	 *
> +	 * 1c:   e0 00 c4 8f                             get $r32 = $ra
> +	 * 20:   00 00 84 e0 00 00 00 80 00 00 00 00     make $r33 = 0 (0x0);;
> +	 *
> +	 *            20: R_KVX_S64_LO10      __mcount
> +	 *            24: R_KVX_S64_UP27      __mcount
> +	 *            28: R_KVX_S64_EX27      __mcount
> +	 * 2c:   21 00 dc 0f                             icall $r33;;
> +	 *
> +	 * So we just need to add INSN_MAKE_IMM64_SYLLABLE_SIZE (0xc) to the
> +	 * address.
> +	 */

You could cheat here, and place a small veneer/PLT next to each module to do
the indirect branch to ftrace. That way, each call site would only need a pcrel
rather than a far call, which might simplfiy your runtime patching logic.

On arm64 we a special ftrace PLT for this.

> +	unsigned int insn;
> +
> +	/*
> +	 * The CALL is 1 syllable while the MAKE IMM64 is 3. But we just
> +	 * need to check that the first syllable of the MAKE IMM64 is the
> +	 * LO10. So finally we just need to read one syllable to adjust the
> +	 * call.
> +	 */
> +	if (kvx_insns_read(&insn, KVX_INSN_SYLLABLE_WIDTH, (void *)addr)) {
> +		pr_err("%s: error when trying to read syllable\n", __func__);
> +		return 0;
> +	}
> +
> +	if (IS_INSN(insn, CALL))
> +		return addr;
> +
> +	if (IS_INSN(insn, MAKE_IMM64))
> +		return addr + INSN_MAKE_IMM64_SYLLABLE_SIZE;
> +
> +	pr_err("%s: syllable is neither a CALL nor a MAKE\n", __func__);
> +	return 0;
> +}
> +
> +/*
> + * Do runtime patching of the active tracer.
> + * This will be modifying the assembly code at the location of the
> + * ftrace_call symbol inside of the ftrace_caller() function.
> + */
> +int ftrace_update_ftrace_func(ftrace_func_t func)
> +{
> +	void *ip;
> +	unsigned int insn[KVX_INSN_MAKE_IMM64_SIZE + KVX_INSN_ICALL_SIZE];
> +
> +	ip = (void *)ftrace_call;
> +	KVX_INSN_MAKE_IMM64(insn, KVX_INSN_PARALLEL_EOB, KVX_REG_R32,
> +			    (unsigned long)func);
> +	KVX_INSN_ICALL(&insn[KVX_INSN_MAKE_IMM64_SIZE],
> +		       KVX_INSN_PARALLEL_EOB,
> +		       KVX_REG_R32);
> +	return write_insns_and_check(insn,
> +				     INSN_MAKE_IMM64_SYLLABLE_SIZE
> +				     + INSN_ICALL_SYLLABLE_SIZE,
> +				     ip);
> +}
> +
> +/*
> + * Turn the mcount call site into a call to an arbitrary location (but
> + * typically that is ftrace_caller()) at runtime.
> + */
> +int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
> +{
> +	void *ip = (void *)rec->ip;
> +	unsigned int insn;
> +	int ret;
> +
> +	/* Ensure that a NOP will be replaced */
> +	if (kvx_insns_read(&insn, KVX_INSN_SYLLABLE_WIDTH, ip)) {
> +		pr_err("%s: failed to read insn\n", __func__);
> +		return -EFAULT;
> +	}
> +
> +	if (!IS_INSN(insn, NOP)) {
> +		pr_err("%s: insn 0x%x is not a NOP\n", __func__, insn);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * Now we can replace the instruction depending of what has been
> +	 * nopified (call or icall)
> +	 */
> +	insn = rec->arch.insn;
> +
> +	if (IS_INSN(insn, CALL)) {
> +		s32 pcrel = addr - (unsigned long) ip;
> +		u32 insn_call;
> +
> +		BUG_ON(KVX_INSN_GOTO_PCREL27_CHECK(pcrel));
> +		KVX_INSN_CALL(&insn_call, KVX_INSN_PARALLEL_EOB, pcrel);
> +
> +		return write_insns_and_check(&insn_call,
> +					     INSN_CALL_SYLLABLE_SIZE, ip);
> +	}
> +
> +	if (IS_INSN(insn, ICALL)) {
> +		u32 insn_make[KVX_INSN_MAKE_IMM64_SIZE];
> +		u32 r = insn & INSN_ICALL_REG_MASK;
> +
> +		KVX_INSN_MAKE_IMM64(insn_make, KVX_INSN_PARALLEL_EOB, r, addr);
> +		ret = write_insns_and_check(insn_make,
> +					    INSN_MAKE_IMM64_SYLLABLE_SIZE,
> +					    ip - INSN_MAKE_IMM64_SYLLABLE_SIZE);
> +		if (ret)
> +			return ret;
> +
> +		return write_insns_and_check(&insn,
> +					      INSN_ICALL_SYLLABLE_SIZE, ip);
> +	}
> +
> +	/* It is neither a call nor an icall */
> +	pr_err("%s: insn 0x%x is neither a CALL nor ICALL\n", __func__, insn);
> +	return -EINVAL;
> +}
> +
> +/*
> + * Turn the mcount call site into a nop at runtime
> + */
> +int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
> +		    unsigned long addr)
> +{
> +	unsigned long ip = rec->ip;
> +	unsigned int insn;
> +	unsigned int nop;
> +
> +	/*
> +	 * Ensure that the instruction that will be replaced is a call or an
> +	 * icall to addr.
> +	 */
> +	if (kvx_insns_read(&insn, KVX_INSN_SYLLABLE_WIDTH, (void *)ip)) {
> +		pr_err("%s: error when trying to read syllable\n", __func__);
> +		return -EFAULT;
> +	}
> +
> +	if (IS_INSN(insn, CALL)) {
> +		int pcrel = ((int)(insn & 0x7ffffff) << 5) >> 3;
> +
> +		if ((ip + pcrel != addr)) {
> +			pr_err("%s: failed to check call addr 0x%lx != 0x%lx\n",
> +			       __func__, ip + pcrel, addr);
> +			return -EINVAL;
> +		}
> +	} else if (IS_INSN(insn, ICALL)) {
> +		unsigned int insn_make[KVX_INSN_MAKE_IMM64_SIZE];
> +		unsigned int reg = insn & INSN_ICALL_REG_MASK;
> +		int ret;
> +
> +		KVX_INSN_MAKE_IMM64(insn_make,
> +				    KVX_INSN_PARALLEL_EOB, reg,
> +				    addr);
> +
> +		ret = read_insns_and_check(insn_make,
> +					   INSN_MAKE_IMM64_SYLLABLE_SIZE,
> +					   (void *)ip -
> +					   INSN_MAKE_IMM64_SYLLABLE_SIZE);
> +		if (ret)
> +			return ret;
> +	} else {
> +		pr_err("%s: insn 0x%x is neither a CALL nor an ICALL\n",
> +		       __func__, insn);
> +		return -EINVAL;
> +	}
> +
> +	rec->arch.insn = insn;
> +	KVX_INSN_NOP(&nop, KVX_INSN_PARALLEL_EOB);
> +	return write_insns_and_check(&nop, INSN_NOP_SYLLABLE_SIZE, (void *)ip);
> +}
> +
> +#endif /* CONFIG_DYNAMIC_FTRACE */
> +
> +/* __mcount is defined in mcount.S */
> +EXPORT_SYMBOL(__mcount);
> diff --git a/arch/kvx/kernel/mcount.S b/arch/kvx/kernel/mcount.S
> new file mode 100644
> index 000000000000..317797aeed42
> --- /dev/null
> +++ b/arch/kvx/kernel/mcount.S
> @@ -0,0 +1,340 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Guillaume Thouvenin
> + *            Marius Gligor
> + *            Clement Leger
> + */
> +
> +#include <linux/linkage.h>
> +#include <asm/insns_defs.h>
> +#include <asm/sfr_defs.h>
> +
> +.altmacro
> +
> +/*
> + * !!! The call to __mcount is special and breaks the ABI !!!
> + * It is because with dynamic ftrace the call can be replaced by any
> + * tracer and GCC cannot assume which caller registers must be saved.
> + * It's because the call is being inserted in the prologue of functions
> + * that is out of GCC knowledge.
> + * We could make this known to GCC, but then we would hit the problem of not
> + * being able to easily decide which insns are emitted and how they are
> + * bundled. So we choose to save nothing when calling __mcount and we will
> + * save all caller registers once __mcount called.

What about the return address? I assume that this is a branch-and-link
architecture, and CALL/ICALL clobber that?

Thanks,
Mark.

> + *
> + * Another specificity is that parameter is passed in $r32 instead of $r0.
> + *
> + * Our stack will be:
> + *	0  | $FP $RA $r15 -
> + *	32 | $r0 $r1 $r2 $r3
> + *	64 | $r4 $r5 $r6 $r7
> + *	96 | $r8 $r9 $r10 $r11
> + *	128| $r32 $r33 $r34 $r35
> + *	160| $r36 $r37 $r38 $r39
> + *	192| $r40 $r41 $r42 $r43
> + *	224| $r44 $r45 $r46 $r47
> + *	256| $r48 $r49 $r50 $r51
> + *	288| $r52 $r53 $r54 $r55
> + *	320| $r56 $r57 $r58 $r59
> + *	352| $r60 $r61 $r62 $r63
> + *	384| $r16 $r17
> + */
> +
> +#define STACK_OFFSET_FP   0
> +#define STACK_OFFSET_RA   (STACK_OFFSET_FP + 8)
> +#define STACK_OFFSET_R15  (STACK_OFFSET_RA + 8)
> +#define STACK_OFFSET_Q0   (STACK_OFFSET_FP + 32)
> +#define STACK_OFFSET_Q4   (STACK_OFFSET_Q0 + 32)
> +#define STACK_OFFSET_Q8   (STACK_OFFSET_Q4 + 32)
> +#define STACK_OFFSET_Q32  (STACK_OFFSET_Q8 + 32)
> +#define STACK_OFFSET_Q36  (STACK_OFFSET_Q32 + 32)
> +#define STACK_OFFSET_Q40  (STACK_OFFSET_Q36 + 32)
> +#define STACK_OFFSET_Q44  (STACK_OFFSET_Q40 + 32)
> +#define STACK_OFFSET_Q48  (STACK_OFFSET_Q44 + 32)
> +#define STACK_OFFSET_Q52  (STACK_OFFSET_Q48 + 32)
> +#define STACK_OFFSET_Q56  (STACK_OFFSET_Q52 + 32)
> +#define STACK_OFFSET_Q60  (STACK_OFFSET_Q56 + 32)
> +#define STACK_OFFSET_P16  (STACK_OFFSET_Q60 + 32)
> +#define STACK_SIZE        (STACK_OFFSET_P16 + 16)
> +
> +#define RA_DWARF_REGNO    (64 + KVX_SFR_RA)
> +#define R14_DWARF_REGNO   14
> +
> +.macro save_all_context_but_P16
> +	/*
> +	 * r16 and r17 are already saved. r16 is already used to store the
> +	 * address of the function tracer if any.
> +	 * so !!! USE r17 as scratch register !!!
> +	 */
> +	get $r17 = $ra
> +	sd STACK_OFFSET_FP[$r12] = $r14
> +	;;
> +	.cfi_offset R14_DWARF_REGNO, -STACK_SIZE + STACK_OFFSET_FP
> +	sd STACK_OFFSET_RA[$r12] = $r17
> +	;;
> +	.cfi_offset RA_DWARF_REGNO, -STACK_SIZE + STACK_OFFSET_RA
> +	sd STACK_OFFSET_R15[$r12] = $r15
> +	;;
> +	so STACK_OFFSET_Q0[$r12] = $r0r1r2r3
> +	;;
> +	so STACK_OFFSET_Q4[$r12] = $r4r5r6r7
> +	;;
> +	so STACK_OFFSET_Q8[$r12] = $r8r9r10r11
> +	;;
> +	so STACK_OFFSET_Q32[$r12] = $r32r33r34r35
> +	;;
> +	so STACK_OFFSET_Q36[$r12] = $r36r37r38r39
> +	;;
> +	so STACK_OFFSET_Q40[$r12] = $r40r41r42r43
> +	;;
> +	so STACK_OFFSET_Q44[$r12] = $r44r45r46r47
> +	;;
> +	so STACK_OFFSET_Q48[$r12] = $r48r49r50r51
> +	;;
> +	so STACK_OFFSET_Q52[$r12] = $r52r53r54r55
> +	;;
> +	so STACK_OFFSET_Q56[$r12] = $r56r57r58r59
> +	;;
> +	so STACK_OFFSET_Q60[$r12] = $r60r61r62r63
> +	;;
> +.endm
> +
> +.macro restore_all_context
> +	ld $r15 = STACK_OFFSET_RA[$r12]
> +	;;
> +	ld $r14 = STACK_OFFSET_FP[$r12]
> +	set $ra = $r15
> +	;;
> +	.cfi_restore RA_DWARF_REGNO
> +	.cfi_restore R14_DWARF_REGNO
> +	ld $r15 = STACK_OFFSET_R15[$r12]
> +	;;
> +	lq $r16r17 = STACK_OFFSET_P16[$r12]
> +	;;
> +	lo $r0r1r2r3 = STACK_OFFSET_Q0[$r12]
> +	;;
> +	lo $r4r5r6r7 = STACK_OFFSET_Q4[$r12]
> +	;;
> +	lo $r8r9r10r11 = STACK_OFFSET_Q8[$r12]
> +	;;
> +	lo $r32r33r34r35 = STACK_OFFSET_Q32[$r12]
> +	;;
> +	lo $r36r37r38r39 = STACK_OFFSET_Q36[$r12]
> +	;;
> +	lo $r40r41r42r43 = STACK_OFFSET_Q40[$r12]
> +	;;
> +	lo $r44r45r46r47 = STACK_OFFSET_Q44[$r12]
> +	;;
> +	lo $r48r49r50r51 = STACK_OFFSET_Q48[$r12]
> +	;;
> +	lo $r52r53r54r55 = STACK_OFFSET_Q52[$r12]
> +	;;
> +	lo $r56r57r58r59 = STACK_OFFSET_Q56[$r12]
> +	;;
> +	lo $r60r61r62r63 = STACK_OFFSET_Q60[$r12]
> +	;;
> +	addd $r12 = $r12, STACK_SIZE
> +	;;
> +	.cfi_def_cfa_offset 0
> +.endm
> +
> +ENTRY(__mcount)
> +.cfi_startproc
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +	/*
> +	 * With dynamic ftrace mcount() is only used during boot and then all
> +	 * references to it will be patched out never to return. ftrace_caller
> +	 * is used.
> +	 */
> +	ret
> +	;;
> +#endif
> +	addd $r12 = $r12, -STACK_SIZE
> +	;;
> +	.cfi_def_cfa_offset STACK_SIZE
> +	sq STACK_OFFSET_P16[$r12] = $r16r17
> +	;;
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +	/*
> +	 * Check the function pointers ftrace_graph_return
> +	 * (compare to ftrace_stub) and ftrace_graph_entry (compare to
> +	 * ftrace_graph_entry_stub).  If either of those is not set to the
> +	 * relevant stub function, call the arch-specific function
> +	 * ftrace_graph_caller.
> +	 */
> +	make $r16 = ftrace_graph_return
> +	make $r17 = ftrace_graph_entry
> +	;;
> +	ld $r16 = 0[$r16]
> +	;;
> +	compd.eq $r16 = $r16, ftrace_stub
> +	ld $r17 = 0[$r17]
> +	;;
> +	/*
> +	 * If ftrace_graph_return is not equal to ftrace_stub ($r16 == 0) then
> +	 * jump to ftrace_graph_caller. Otherwise compare ftrace_graph_entry.
> +	 */
> +	cb.deqz $r16? ftrace_graph_caller
> +	compd.eq $r17 = $r17, ftrace_graph_entry_stub
> +	;;
> +	/*
> +	 * If ftrace_graph_entry is equal to ftrace_graph_entry_stub ($r17 == 1)
> +	 * then jump to ftrace_graph_end. If not we continue.
> +	 */
> +	cb.dnez $r17? ftrace_graph_end
> +	;;
> +ftrace_graph_caller:
> +	save_all_context_but_P16
> +	;;
> +GLOBAL(ftrace_graph_call)
> +	ld $r16 = 0[$r12]
> +	;;
> +	addd $r0 = $r16, 0x8 /* a pointer to the frompc */
> +	get $r1 = $ra /* selfpc */
> +	copyd $r2 = $r16 /* FP of the function that called __mcount */
> +	;;
> +	/* get the address of the bundle */
> +	addd $r1 = $r1, -KVX_INSN_SYLLABLE_WIDTH
> +	;;
> +	call prepare_ftrace_return
> +	;;
> +	restore_all_context
> +	;;
> +	ret
> +	;;
> +ftrace_graph_end:
> +	.cfi_def_cfa_offset STACK_SIZE
> +#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
> +	/*
> +	 * Check if function pointer ftrace_trace_function is set to
> +	 * ftrace_stub.
> +	 * If it isn't, then call that function in the same way the mcount
> +	 * function normally calls __mcount_internal
> +	 * - the first argument is the "frompc" ($r0)
> +	 * - the second argument is the "selfpc" ($ra - mcount size adjustement)
> +	 * As we don't use any callee saved registers we don't need to save
> +	 * them.
> +	 */
> +	make $r16 = ftrace_trace_function
> +	;;
> +	ld $r16 = 0[$r16]
> +	;;
> +	/*
> +	 * NOTE: keep value in $r16 (the function we may call if set). See
> +	 *       comment in save_all_context_but_P16.
> +	 */
> +	compd.eq $r17 = $r16, ftrace_stub
> +	;;
> +	/*
> +	 * If ftrace_trace_function is equal to ftrace_stub ($r17 == 1), there
> +	 * is nothing to do, so return immediately otherwise call the function
> +	 * in $r16 with correct parameters.
> +	 */
> +	cb.deqz $r17? 1f
> +	;;
> +	/* Only need to restore P16 */
> +	lq $r16r17 = STACK_OFFSET_P16[$r12]
> +	;;
> +	addd $r12 = $r12, STACK_SIZE
> +	;;
> +	.cfi_def_cfa_offset 0
> +	ret
> +	;;
> +
> +1:
> +	save_all_context_but_P16
> +	;;
> +	copyd $r0 = $r32
> +	get $r1 = $ra
> +	;;
> +	addd $r1 = $r1, -KVX_INSN_SYLLABLE_WIDTH
> +	;;
> +	icall $r16
> +	;;
> +	restore_all_context
> +	;;
> +	ret
> +	;;
> +.cfi_endproc
> +ENDPROC(__mcount)
> +
> +/* Required by ftrace_graph_ret_addr */
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +ENTRY(return_to_handler)
> +	addd $r12 = $r12, -128
> +	;;
> +	so 0[$r12] = $r0r1r2r3
> +	;;
> +	so 32[$r12] = $r4r5r6r7
> +	;;
> +	so 64[$r12] = $r8r9r10r11
> +	;;
> +	call ftrace_return_to_handler
> +	;;
> +	set $ra = $r0
> +	lo $r0r1r2r3 = 0[$r12]
> +	;;
> +	lo $r4r5r6r7 = 32[$r12]
> +	;;
> +	lo $r8r9r10r11 = 64[$r12]
> +	addd $r12 = $r12, 128
> +	;;
> +	ret
> +	;;
> +ENDPROC(return_to_handler)
> +#endif
> +
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +ENTRY(ftrace_caller)
> +.cfi_startproc
> +
> +	addd $r12 = $r12, -STACK_SIZE
> +	;;
> +	.cfi_def_cfa_offset STACK_SIZE
> +	sq STACK_OFFSET_P16[$r12] = $r16r17
> +	;;
> +	save_all_context_but_P16
> +	;;
> +	copyd $r0 = $r32
> +	get $r1 = $ra
> +	;;
> +	addd $r1 = $r1, -KVX_INSN_SYLLABLE_WIDTH
> +	;;
> +	goto ftrace_call
> +	;;
> +.align 64
> +GLOBAL(ftrace_call)
> +	/*
> +	 * To work with module we need to use far call. So prepare the
> +	 * space for them:
> +	 *  - make: 3 syllables
> +	 *  - icall/igoto: 1 syllable
> +	 *
> +	 * We are using igoto in ftrace_enable_ftrace_graph_caller(). This is
> +	 * because this path is a tweak to enable function graph tracer with
> +	 * dynamic ftrace. We need to jump to ftrace_graph_call without updating
> +	 * RA and FP. Once in ftrace_graph_call we will return to the correct
> +	 * address from the classical path of the function graph tracer.
> +	 */
> +	nop
> +	;;
> +	nop
> +	;;
> +	nop
> +	;;
> +	nop
> +	;;
> +	restore_all_context
> +	;;
> +	ret
> +	;;
> +.cfi_endproc
> +ENDPROC(ftrace_caller)
> +#endif // CONFIG_DYNAMIC_FTRACE
> +
> +ENTRY(ftrace_stub)
> +	ret
> +	;;
> +ENDPROC(ftrace_stub)
> +