1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/kernel/profile.c
4 * Simple profiling. Manages a direct-mapped profile hit count buffer,
5 * with configurable resolution, support for restricting the cpus on
6 * which profiling is done, and switching between cpu time and
7 * schedule() calls via kernel command line parameters passed at boot.
9 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
11 * Consolidation of architecture support code for profiling,
12 * Nadia Yvette Chambers, Oracle, July 2004
13 * Amortized hit count accounting via per-cpu open-addressed hashtables
14 * to resolve timer interrupt livelocks, Nadia Yvette Chambers,
18 #include <linux/export.h>
19 #include <linux/profile.h>
20 #include <linux/memblock.h>
21 #include <linux/notifier.h>
23 #include <linux/cpumask.h>
24 #include <linux/cpu.h>
25 #include <linux/highmem.h>
26 #include <linux/mutex.h>
27 #include <linux/slab.h>
28 #include <linux/vmalloc.h>
29 #include <linux/sched/stat.h>
31 #include <asm/sections.h>
32 #include <asm/irq_regs.h>
33 #include <asm/ptrace.h>
38 #define PROFILE_GRPSHIFT 3
39 #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
40 #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
41 #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
43 static atomic_t *prof_buffer;
44 static unsigned long prof_len;
45 static unsigned short int prof_shift;
47 int prof_on __read_mostly;
48 EXPORT_SYMBOL_GPL(prof_on);
50 int profile_setup(char *str)
52 static const char schedstr[] = "schedule";
53 static const char kvmstr[] = "kvm";
54 const char *select = NULL;
57 if (!strncmp(str, schedstr, strlen(schedstr))) {
58 prof_on = SCHED_PROFILING;
60 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
61 prof_on = KVM_PROFILING;
63 } else if (get_option(&str, &par)) {
64 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
65 prof_on = CPU_PROFILING;
66 pr_info("kernel profiling enabled (shift: %u)\n",
71 if (str[strlen(select)] == ',')
72 str += strlen(select) + 1;
73 if (get_option(&str, &par))
74 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
75 pr_info("kernel %s profiling enabled (shift: %u)\n",
81 __setup("profile=", profile_setup);
84 int __ref profile_init(void)
90 /* only text is profiled */
91 prof_len = (_etext - _stext) >> prof_shift;
94 pr_warn("profiling shift: %u too large\n", prof_shift);
99 buffer_bytes = prof_len*sizeof(atomic_t);
101 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
105 prof_buffer = alloc_pages_exact(buffer_bytes,
106 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
110 prof_buffer = vzalloc(buffer_bytes);
117 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
120 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
122 atomic_add(nr_hits, &prof_buffer[pc]);
125 void profile_hits(int type, void *__pc, unsigned int nr_hits)
127 if (prof_on != type || !prof_buffer)
129 do_profile_hits(type, __pc, nr_hits);
131 EXPORT_SYMBOL_GPL(profile_hits);
133 void profile_tick(int type)
135 struct pt_regs *regs = get_irq_regs();
137 /* This is the old kernel-only legacy profiling */
138 if (!user_mode(regs))
139 profile_hit(type, (void *)profile_pc(regs));
142 #ifdef CONFIG_PROC_FS
143 #include <linux/proc_fs.h>
144 #include <linux/seq_file.h>
145 #include <linux/uaccess.h>
148 * This function accesses profiling information. The returned data is
149 * binary: the sampling step and the actual contents of the profile
150 * buffer. Use of the program readprofile is recommended in order to
151 * get meaningful info out of these data.
154 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
156 unsigned long p = *ppos;
159 unsigned long sample_step = 1UL << prof_shift;
161 if (p >= (prof_len+1)*sizeof(unsigned int))
163 if (count > (prof_len+1)*sizeof(unsigned int) - p)
164 count = (prof_len+1)*sizeof(unsigned int) - p;
167 while (p < sizeof(unsigned int) && count > 0) {
168 if (put_user(*((char *)(&sample_step)+p), buf))
170 buf++; p++; count--; read++;
172 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
173 if (copy_to_user(buf, (void *)pnt, count))
180 /* default is to not implement this call */
181 int __weak setup_profiling_timer(unsigned mult)
187 * Writing to /proc/profile resets the counters
189 * Writing a 'profiling multiplier' value into it also re-sets the profiling
190 * interrupt frequency, on architectures that support this.
192 static ssize_t write_profile(struct file *file, const char __user *buf,
193 size_t count, loff_t *ppos)
196 if (count == sizeof(int)) {
197 unsigned int multiplier;
199 if (copy_from_user(&multiplier, buf, sizeof(int)))
202 if (setup_profiling_timer(multiplier))
206 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
210 static const struct proc_ops profile_proc_ops = {
211 .proc_read = read_profile,
212 .proc_write = write_profile,
213 .proc_lseek = default_llseek,
216 int __ref create_proc_profile(void)
218 struct proc_dir_entry *entry;
223 entry = proc_create("profile", S_IWUSR | S_IRUGO,
224 NULL, &profile_proc_ops);
226 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
229 subsys_initcall(create_proc_profile);
230 #endif /* CONFIG_PROC_FS */