2 * Dynamic function tracer architecture backend.
4 * Copyright IBM Corp. 2009,2014
6 * Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>,
7 * Martin Schwidefsky <schwidefsky@de.ibm.com>
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>
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
51 unsigned long ftrace_plt;
53 int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
59 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
62 struct ftrace_insn orig, new, old;
64 if (probe_kernel_read(&old, (void *) rec->ip, sizeof(old)))
66 if (addr == MCOUNT_ADDR) {
67 /* Initial code replacement; we expect to see stg r14,8(r15) */
69 orig.disp = 0xf0080024;
70 ftrace_generate_nop_insn(&new);
71 } else if (old.opc == BREAKPOINT_INSTRUCTION) {
73 * If we find a breakpoint instruction, a kprobe has been
74 * placed at the beginning of the function. We write the
75 * constant KPROBE_ON_FTRACE_NOP into the remaining four
76 * bytes of the original instruction so that the kprobes
77 * handler can execute a nop, if it reaches this breakpoint.
79 new.opc = orig.opc = BREAKPOINT_INSTRUCTION;
80 orig.disp = KPROBE_ON_FTRACE_CALL;
81 new.disp = KPROBE_ON_FTRACE_NOP;
83 /* Replace ftrace call with a nop. */
84 ftrace_generate_call_insn(&orig, rec->ip);
85 ftrace_generate_nop_insn(&new);
87 /* Verify that the to be replaced code matches what we expect. */
88 if (memcmp(&orig, &old, sizeof(old)))
90 if (probe_kernel_write((void *) rec->ip, &new, sizeof(new)))
95 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
97 struct ftrace_insn orig, new, old;
99 if (probe_kernel_read(&old, (void *) rec->ip, sizeof(old)))
101 if (old.opc == BREAKPOINT_INSTRUCTION) {
103 * If we find a breakpoint instruction, a kprobe has been
104 * placed at the beginning of the function. We write the
105 * constant KPROBE_ON_FTRACE_CALL into the remaining four
106 * bytes of the original instruction so that the kprobes
107 * handler can execute a brasl if it reaches this breakpoint.
109 new.opc = orig.opc = BREAKPOINT_INSTRUCTION;
110 orig.disp = KPROBE_ON_FTRACE_NOP;
111 new.disp = KPROBE_ON_FTRACE_CALL;
113 /* Replace nop with an ftrace call. */
114 ftrace_generate_nop_insn(&orig);
115 ftrace_generate_call_insn(&new, rec->ip);
117 /* Verify that the to be replaced code matches what we expect. */
118 if (memcmp(&orig, &old, sizeof(old)))
120 if (probe_kernel_write((void *) rec->ip, &new, sizeof(new)))
125 int ftrace_update_ftrace_func(ftrace_func_t func)
130 int __init ftrace_dyn_arch_init(void)
135 static int __init ftrace_plt_init(void)
139 ftrace_plt = (unsigned long) module_alloc(PAGE_SIZE);
141 panic("cannot allocate ftrace plt\n");
142 ip = (unsigned int *) ftrace_plt;
143 ip[0] = 0x0d10e310; /* basr 1,0; lg 1,10(1); br 1 */
146 ip[3] = FTRACE_ADDR >> 32;
147 ip[4] = FTRACE_ADDR & 0xffffffff;
148 set_memory_ro(ftrace_plt, 1);
151 device_initcall(ftrace_plt_init);
153 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
155 * Hook the return address and push it in the stack of return addresses
156 * in current thread info.
158 unsigned long prepare_ftrace_return(unsigned long parent, unsigned long ip)
160 struct ftrace_graph_ent trace;
162 if (unlikely(ftrace_graph_is_dead()))
164 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
166 ip = (ip & PSW_ADDR_INSN) - MCOUNT_INSN_SIZE;
168 trace.depth = current->curr_ret_stack + 1;
169 /* Only trace if the calling function expects to. */
170 if (!ftrace_graph_entry(&trace))
172 if (ftrace_push_return_trace(parent, ip, &trace.depth, 0) == -EBUSY)
174 parent = (unsigned long) return_to_handler;
178 NOKPROBE_SYMBOL(prepare_ftrace_return);
181 * Patch the kernel code at ftrace_graph_caller location. The instruction
182 * there is branch relative on condition. To enable the ftrace graph code
183 * block, we simply patch the mask field of the instruction to zero and
184 * turn the instruction into a nop.
185 * To disable the ftrace graph code the mask field will be patched to
186 * all ones, which turns the instruction into an unconditional branch.
188 int ftrace_enable_ftrace_graph_caller(void)
190 u8 op = 0x04; /* set mask field to zero */
192 return probe_kernel_write(__va(ftrace_graph_caller)+1, &op, sizeof(op));
195 int ftrace_disable_ftrace_graph_caller(void)
197 u8 op = 0xf4; /* set mask field to all ones */
199 return probe_kernel_write(__va(ftrace_graph_caller)+1, &op, sizeof(op));
202 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */