Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
v5.4
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Dynamic function tracer architecture backend.
  4 *
  5 * Copyright IBM Corp. 2009,2014
  6 *
  7 *   Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>,
  8 *		Martin Schwidefsky <schwidefsky@de.ibm.com>
  9 */
 10
 11#include <linux/moduleloader.h>
 12#include <linux/hardirq.h>
 13#include <linux/uaccess.h>
 14#include <linux/ftrace.h>
 15#include <linux/kernel.h>
 16#include <linux/types.h>
 17#include <linux/kprobes.h>
 18#include <trace/syscall.h>
 19#include <asm/asm-offsets.h>
 20#include <asm/cacheflush.h>
 21#include <asm/set_memory.h>
 22#include "entry.h"
 23
 24/*
 25 * The mcount code looks like this:
 26 *	stg	%r14,8(%r15)		# offset 0
 27 *	larl	%r1,<&counter>		# offset 6
 28 *	brasl	%r14,_mcount		# offset 12
 29 *	lg	%r14,8(%r15)		# offset 18
 30 * Total length is 24 bytes. Only the first instruction will be patched
 31 * by ftrace_make_call / ftrace_make_nop.
 32 * The enabled ftrace code block looks like this:
 33 * >	brasl	%r0,ftrace_caller	# offset 0
 34 *	larl	%r1,<&counter>		# offset 6
 35 *	brasl	%r14,_mcount		# offset 12
 36 *	lg	%r14,8(%r15)		# offset 18
 37 * The ftrace function gets called with a non-standard C function call ABI
 38 * where r0 contains the return address. It is also expected that the called
 39 * function only clobbers r0 and r1, but restores r2-r15.
 40 * For module code we can't directly jump to ftrace caller, but need a
 41 * trampoline (ftrace_plt), which clobbers also r1.
 42 * The return point of the ftrace function has offset 24, so execution
 43 * continues behind the mcount block.
 44 * The disabled ftrace code block looks like this:
 45 * >	jg	.+24			# offset 0
 46 *	larl	%r1,<&counter>		# offset 6
 47 *	brasl	%r14,_mcount		# offset 12
 48 *	lg	%r14,8(%r15)		# offset 18
 49 * The jg instruction branches to offset 24 to skip as many instructions
 50 * as possible.
 51 * In case we use gcc's hotpatch feature the original and also the disabled
 52 * function prologue contains only a single six byte instruction and looks
 53 * like this:
 54 * >	brcl	0,0			# offset 0
 55 * To enable ftrace the code gets patched like above and afterwards looks
 56 * like this:
 57 * >	brasl	%r0,ftrace_caller	# offset 0
 58 */
 59
 60unsigned long ftrace_plt;
 61
 62static inline void ftrace_generate_orig_insn(struct ftrace_insn *insn)
 63{
 64#if defined(CC_USING_HOTPATCH) || defined(CC_USING_NOP_MCOUNT)
 65	/* brcl 0,0 */
 66	insn->opc = 0xc004;
 67	insn->disp = 0;
 68#else
 69	/* stg r14,8(r15) */
 70	insn->opc = 0xe3e0;
 71	insn->disp = 0xf0080024;
 72#endif
 73}
 74
 75static inline int is_kprobe_on_ftrace(struct ftrace_insn *insn)
 76{
 77#ifdef CONFIG_KPROBES
 78	if (insn->opc == BREAKPOINT_INSTRUCTION)
 79		return 1;
 80#endif
 81	return 0;
 82}
 83
 84static inline void ftrace_generate_kprobe_nop_insn(struct ftrace_insn *insn)
 85{
 86#ifdef CONFIG_KPROBES
 87	insn->opc = BREAKPOINT_INSTRUCTION;
 88	insn->disp = KPROBE_ON_FTRACE_NOP;
 89#endif
 90}
 91
 92static inline void ftrace_generate_kprobe_call_insn(struct ftrace_insn *insn)
 93{
 94#ifdef CONFIG_KPROBES
 95	insn->opc = BREAKPOINT_INSTRUCTION;
 96	insn->disp = KPROBE_ON_FTRACE_CALL;
 97#endif
 98}
 99
100int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
101		       unsigned long addr)
102{
103	return 0;
104}
105
106int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
107		    unsigned long addr)
108{
109	struct ftrace_insn orig, new, old;
110
111	if (probe_kernel_read(&old, (void *) rec->ip, sizeof(old)))
112		return -EFAULT;
113	if (addr == MCOUNT_ADDR) {
114		/* Initial code replacement */
115		ftrace_generate_orig_insn(&orig);
116		ftrace_generate_nop_insn(&new);
117	} else if (is_kprobe_on_ftrace(&old)) {
118		/*
119		 * If we find a breakpoint instruction, a kprobe has been
120		 * placed at the beginning of the function. We write the
121		 * constant KPROBE_ON_FTRACE_NOP into the remaining four
122		 * bytes of the original instruction so that the kprobes
123		 * handler can execute a nop, if it reaches this breakpoint.
124		 */
125		ftrace_generate_kprobe_call_insn(&orig);
126		ftrace_generate_kprobe_nop_insn(&new);
127	} else {
128		/* Replace ftrace call with a nop. */
129		ftrace_generate_call_insn(&orig, rec->ip);
130		ftrace_generate_nop_insn(&new);
131	}
132	/* Verify that the to be replaced code matches what we expect. */
133	if (memcmp(&orig, &old, sizeof(old)))
134		return -EINVAL;
135	s390_kernel_write((void *) rec->ip, &new, sizeof(new));
136	return 0;
137}
138
139int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
140{
141	struct ftrace_insn orig, new, old;
142
143	if (probe_kernel_read(&old, (void *) rec->ip, sizeof(old)))
144		return -EFAULT;
145	if (is_kprobe_on_ftrace(&old)) {
146		/*
147		 * If we find a breakpoint instruction, a kprobe has been
148		 * placed at the beginning of the function. We write the
149		 * constant KPROBE_ON_FTRACE_CALL into the remaining four
150		 * bytes of the original instruction so that the kprobes
151		 * handler can execute a brasl if it reaches this breakpoint.
152		 */
153		ftrace_generate_kprobe_nop_insn(&orig);
154		ftrace_generate_kprobe_call_insn(&new);
155	} else {
156		/* Replace nop with an ftrace call. */
157		ftrace_generate_nop_insn(&orig);
158		ftrace_generate_call_insn(&new, rec->ip);
159	}
160	/* Verify that the to be replaced code matches what we expect. */
161	if (memcmp(&orig, &old, sizeof(old)))
162		return -EINVAL;
163	s390_kernel_write((void *) rec->ip, &new, sizeof(new));
164	return 0;
165}
166
167int ftrace_update_ftrace_func(ftrace_func_t func)
168{
169	return 0;
170}
171
172int __init ftrace_dyn_arch_init(void)
173{
174	return 0;
175}
176
177#ifdef CONFIG_MODULES
178
179static int __init ftrace_plt_init(void)
180{
181	unsigned int *ip;
182
183	ftrace_plt = (unsigned long) module_alloc(PAGE_SIZE);
184	if (!ftrace_plt)
185		panic("cannot allocate ftrace plt\n");
186	ip = (unsigned int *) ftrace_plt;
187	ip[0] = 0x0d10e310; /* basr 1,0; lg 1,10(1); br 1 */
188	ip[1] = 0x100a0004;
189	ip[2] = 0x07f10000;
190	ip[3] = FTRACE_ADDR >> 32;
191	ip[4] = FTRACE_ADDR & 0xffffffff;
192	set_memory_ro(ftrace_plt, 1);
193	return 0;
194}
195device_initcall(ftrace_plt_init);
196
197#endif /* CONFIG_MODULES */
198
199#ifdef CONFIG_FUNCTION_GRAPH_TRACER
200/*
201 * Hook the return address and push it in the stack of return addresses
202 * in current thread info.
203 */
204unsigned long prepare_ftrace_return(unsigned long ra, unsigned long sp,
205				    unsigned long ip)
206{
 
 
207	if (unlikely(ftrace_graph_is_dead()))
208		goto out;
209	if (unlikely(atomic_read(&current->tracing_graph_pause)))
210		goto out;
211	ip -= MCOUNT_INSN_SIZE;
212	if (!function_graph_enter(ra, ip, 0, (void *) sp))
213		ra = (unsigned long) return_to_handler;
 
 
 
 
 
 
 
214out:
215	return ra;
216}
217NOKPROBE_SYMBOL(prepare_ftrace_return);
218
219/*
220 * Patch the kernel code at ftrace_graph_caller location. The instruction
221 * there is branch relative on condition. To enable the ftrace graph code
222 * block, we simply patch the mask field of the instruction to zero and
223 * turn the instruction into a nop.
224 * To disable the ftrace graph code the mask field will be patched to
225 * all ones, which turns the instruction into an unconditional branch.
226 */
227int ftrace_enable_ftrace_graph_caller(void)
228{
229	u8 op = 0x04; /* set mask field to zero */
230
231	s390_kernel_write(__va(ftrace_graph_caller)+1, &op, sizeof(op));
232	return 0;
233}
234
235int ftrace_disable_ftrace_graph_caller(void)
236{
237	u8 op = 0xf4; /* set mask field to all ones */
238
239	s390_kernel_write(__va(ftrace_graph_caller)+1, &op, sizeof(op));
240	return 0;
241}
242
243#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
v4.10.11
 
  1/*
  2 * Dynamic function tracer architecture backend.
  3 *
  4 * Copyright IBM Corp. 2009,2014
  5 *
  6 *   Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>,
  7 *		Martin Schwidefsky <schwidefsky@de.ibm.com>
  8 */
  9
 10#include <linux/moduleloader.h>
 11#include <linux/hardirq.h>
 12#include <linux/uaccess.h>
 13#include <linux/ftrace.h>
 14#include <linux/kernel.h>
 15#include <linux/types.h>
 16#include <linux/kprobes.h>
 17#include <trace/syscall.h>
 18#include <asm/asm-offsets.h>
 19#include <asm/cacheflush.h>
 
 20#include "entry.h"
 21
 22/*
 23 * The mcount code looks like this:
 24 *	stg	%r14,8(%r15)		# offset 0
 25 *	larl	%r1,<&counter>		# offset 6
 26 *	brasl	%r14,_mcount		# offset 12
 27 *	lg	%r14,8(%r15)		# offset 18
 28 * Total length is 24 bytes. Only the first instruction will be patched
 29 * by ftrace_make_call / ftrace_make_nop.
 30 * The enabled ftrace code block looks like this:
 31 * >	brasl	%r0,ftrace_caller	# offset 0
 32 *	larl	%r1,<&counter>		# offset 6
 33 *	brasl	%r14,_mcount		# offset 12
 34 *	lg	%r14,8(%r15)		# offset 18
 35 * The ftrace function gets called with a non-standard C function call ABI
 36 * where r0 contains the return address. It is also expected that the called
 37 * function only clobbers r0 and r1, but restores r2-r15.
 38 * For module code we can't directly jump to ftrace caller, but need a
 39 * trampoline (ftrace_plt), which clobbers also r1.
 40 * The return point of the ftrace function has offset 24, so execution
 41 * continues behind the mcount block.
 42 * The disabled ftrace code block looks like this:
 43 * >	jg	.+24			# offset 0
 44 *	larl	%r1,<&counter>		# offset 6
 45 *	brasl	%r14,_mcount		# offset 12
 46 *	lg	%r14,8(%r15)		# offset 18
 47 * The jg instruction branches to offset 24 to skip as many instructions
 48 * as possible.
 49 * In case we use gcc's hotpatch feature the original and also the disabled
 50 * function prologue contains only a single six byte instruction and looks
 51 * like this:
 52 * >	brcl	0,0			# offset 0
 53 * To enable ftrace the code gets patched like above and afterwards looks
 54 * like this:
 55 * >	brasl	%r0,ftrace_caller	# offset 0
 56 */
 57
 58unsigned long ftrace_plt;
 59
 60static inline void ftrace_generate_orig_insn(struct ftrace_insn *insn)
 61{
 62#ifdef CC_USING_HOTPATCH
 63	/* brcl 0,0 */
 64	insn->opc = 0xc004;
 65	insn->disp = 0;
 66#else
 67	/* stg r14,8(r15) */
 68	insn->opc = 0xe3e0;
 69	insn->disp = 0xf0080024;
 70#endif
 71}
 72
 73static inline int is_kprobe_on_ftrace(struct ftrace_insn *insn)
 74{
 75#ifdef CONFIG_KPROBES
 76	if (insn->opc == BREAKPOINT_INSTRUCTION)
 77		return 1;
 78#endif
 79	return 0;
 80}
 81
 82static inline void ftrace_generate_kprobe_nop_insn(struct ftrace_insn *insn)
 83{
 84#ifdef CONFIG_KPROBES
 85	insn->opc = BREAKPOINT_INSTRUCTION;
 86	insn->disp = KPROBE_ON_FTRACE_NOP;
 87#endif
 88}
 89
 90static inline void ftrace_generate_kprobe_call_insn(struct ftrace_insn *insn)
 91{
 92#ifdef CONFIG_KPROBES
 93	insn->opc = BREAKPOINT_INSTRUCTION;
 94	insn->disp = KPROBE_ON_FTRACE_CALL;
 95#endif
 96}
 97
 98int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
 99		       unsigned long addr)
100{
101	return 0;
102}
103
104int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
105		    unsigned long addr)
106{
107	struct ftrace_insn orig, new, old;
108
109	if (probe_kernel_read(&old, (void *) rec->ip, sizeof(old)))
110		return -EFAULT;
111	if (addr == MCOUNT_ADDR) {
112		/* Initial code replacement */
113		ftrace_generate_orig_insn(&orig);
114		ftrace_generate_nop_insn(&new);
115	} else if (is_kprobe_on_ftrace(&old)) {
116		/*
117		 * If we find a breakpoint instruction, a kprobe has been
118		 * placed at the beginning of the function. We write the
119		 * constant KPROBE_ON_FTRACE_NOP into the remaining four
120		 * bytes of the original instruction so that the kprobes
121		 * handler can execute a nop, if it reaches this breakpoint.
122		 */
123		ftrace_generate_kprobe_call_insn(&orig);
124		ftrace_generate_kprobe_nop_insn(&new);
125	} else {
126		/* Replace ftrace call with a nop. */
127		ftrace_generate_call_insn(&orig, rec->ip);
128		ftrace_generate_nop_insn(&new);
129	}
130	/* Verify that the to be replaced code matches what we expect. */
131	if (memcmp(&orig, &old, sizeof(old)))
132		return -EINVAL;
133	s390_kernel_write((void *) rec->ip, &new, sizeof(new));
134	return 0;
135}
136
137int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
138{
139	struct ftrace_insn orig, new, old;
140
141	if (probe_kernel_read(&old, (void *) rec->ip, sizeof(old)))
142		return -EFAULT;
143	if (is_kprobe_on_ftrace(&old)) {
144		/*
145		 * If we find a breakpoint instruction, a kprobe has been
146		 * placed at the beginning of the function. We write the
147		 * constant KPROBE_ON_FTRACE_CALL into the remaining four
148		 * bytes of the original instruction so that the kprobes
149		 * handler can execute a brasl if it reaches this breakpoint.
150		 */
151		ftrace_generate_kprobe_nop_insn(&orig);
152		ftrace_generate_kprobe_call_insn(&new);
153	} else {
154		/* Replace nop with an ftrace call. */
155		ftrace_generate_nop_insn(&orig);
156		ftrace_generate_call_insn(&new, rec->ip);
157	}
158	/* Verify that the to be replaced code matches what we expect. */
159	if (memcmp(&orig, &old, sizeof(old)))
160		return -EINVAL;
161	s390_kernel_write((void *) rec->ip, &new, sizeof(new));
162	return 0;
163}
164
165int ftrace_update_ftrace_func(ftrace_func_t func)
166{
167	return 0;
168}
169
170int __init ftrace_dyn_arch_init(void)
171{
172	return 0;
173}
174
 
 
175static int __init ftrace_plt_init(void)
176{
177	unsigned int *ip;
178
179	ftrace_plt = (unsigned long) module_alloc(PAGE_SIZE);
180	if (!ftrace_plt)
181		panic("cannot allocate ftrace plt\n");
182	ip = (unsigned int *) ftrace_plt;
183	ip[0] = 0x0d10e310; /* basr 1,0; lg 1,10(1); br 1 */
184	ip[1] = 0x100a0004;
185	ip[2] = 0x07f10000;
186	ip[3] = FTRACE_ADDR >> 32;
187	ip[4] = FTRACE_ADDR & 0xffffffff;
188	set_memory_ro(ftrace_plt, 1);
189	return 0;
190}
191device_initcall(ftrace_plt_init);
192
 
 
193#ifdef CONFIG_FUNCTION_GRAPH_TRACER
194/*
195 * Hook the return address and push it in the stack of return addresses
196 * in current thread info.
197 */
198unsigned long prepare_ftrace_return(unsigned long parent, unsigned long ip)
 
199{
200	struct ftrace_graph_ent trace;
201
202	if (unlikely(ftrace_graph_is_dead()))
203		goto out;
204	if (unlikely(atomic_read(&current->tracing_graph_pause)))
205		goto out;
206	ip -= MCOUNT_INSN_SIZE;
207	trace.func = ip;
208	trace.depth = current->curr_ret_stack + 1;
209	/* Only trace if the calling function expects to. */
210	if (!ftrace_graph_entry(&trace))
211		goto out;
212	if (ftrace_push_return_trace(parent, ip, &trace.depth, 0,
213				     NULL) == -EBUSY)
214		goto out;
215	parent = (unsigned long) return_to_handler;
216out:
217	return parent;
218}
219NOKPROBE_SYMBOL(prepare_ftrace_return);
220
221/*
222 * Patch the kernel code at ftrace_graph_caller location. The instruction
223 * there is branch relative on condition. To enable the ftrace graph code
224 * block, we simply patch the mask field of the instruction to zero and
225 * turn the instruction into a nop.
226 * To disable the ftrace graph code the mask field will be patched to
227 * all ones, which turns the instruction into an unconditional branch.
228 */
229int ftrace_enable_ftrace_graph_caller(void)
230{
231	u8 op = 0x04; /* set mask field to zero */
232
233	s390_kernel_write(__va(ftrace_graph_caller)+1, &op, sizeof(op));
234	return 0;
235}
236
237int ftrace_disable_ftrace_graph_caller(void)
238{
239	u8 op = 0xf4; /* set mask field to all ones */
240
241	s390_kernel_write(__va(ftrace_graph_caller)+1, &op, sizeof(op));
242	return 0;
243}
244
245#endif /* CONFIG_FUNCTION_GRAPH_TRACER */