| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 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. |
| 8 | * |
| 9 | * Scheduler profiling support, Arjan van de Ven and Ingo Molnar, |
| 10 | * Red Hat, July 2004 |
| 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, |
| 15 | * Oracle, 2004 |
| 16 | */ |
| 17 | |
| 18 | #include <linux/export.h> |
| 19 | #include <linux/profile.h> |
| 20 | #include <linux/memblock.h> |
| 21 | #include <linux/notifier.h> |
| 22 | #include <linux/mm.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> |
| 30 | |
| 31 | #include <asm/sections.h> |
| 32 | #include <asm/irq_regs.h> |
| 33 | #include <asm/ptrace.h> |
| 34 | |
| 35 | struct profile_hit { |
| 36 | u32 pc, hits; |
| 37 | }; |
| 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) |
| 42 | |
| 43 | static atomic_t *prof_buffer; |
| 44 | static unsigned long prof_len; |
| 45 | static unsigned short int prof_shift; |
| 46 | |
| 47 | int prof_on __read_mostly; |
| 48 | EXPORT_SYMBOL_GPL(prof_on); |
| 49 | |
| 50 | int profile_setup(char *str) |
| 51 | { |
| 52 | static const char schedstr[] = "schedule" ; |
| 53 | static const char kvmstr[] = "kvm" ; |
| 54 | const char *select = NULL; |
| 55 | int par; |
| 56 | |
| 57 | if (!strncmp(str, schedstr, strlen(schedstr))) { |
| 58 | prof_on = SCHED_PROFILING; |
| 59 | select = schedstr; |
| 60 | } else if (!strncmp(str, kvmstr, strlen(kvmstr))) { |
| 61 | prof_on = KVM_PROFILING; |
| 62 | select = kvmstr; |
| 63 | } else if (get_option(str: &str, pint: &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" , |
| 67 | prof_shift); |
| 68 | } |
| 69 | |
| 70 | if (select) { |
| 71 | if (str[strlen(select)] == ',') |
| 72 | str += strlen(select) + 1; |
| 73 | if (get_option(str: &str, pint: &par)) |
| 74 | prof_shift = clamp(par, 0, BITS_PER_LONG - 1); |
| 75 | pr_info("kernel %s profiling enabled (shift: %u)\n" , |
| 76 | select, prof_shift); |
| 77 | } |
| 78 | |
| 79 | return 1; |
| 80 | } |
| 81 | __setup("profile=" , profile_setup); |
| 82 | |
| 83 | |
| 84 | int __ref profile_init(void) |
| 85 | { |
| 86 | int buffer_bytes; |
| 87 | if (!prof_on) |
| 88 | return 0; |
| 89 | |
| 90 | /* only text is profiled */ |
| 91 | prof_len = (_etext - _stext) >> prof_shift; |
| 92 | |
| 93 | if (!prof_len) { |
| 94 | pr_warn("profiling shift: %u too large\n" , prof_shift); |
| 95 | prof_on = 0; |
| 96 | return -EINVAL; |
| 97 | } |
| 98 | |
| 99 | buffer_bytes = prof_len*sizeof(atomic_t); |
| 100 | |
| 101 | prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN); |
| 102 | if (prof_buffer) |
| 103 | return 0; |
| 104 | |
| 105 | prof_buffer = alloc_pages_exact(buffer_bytes, |
| 106 | GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN); |
| 107 | if (prof_buffer) |
| 108 | return 0; |
| 109 | |
| 110 | prof_buffer = vzalloc(buffer_bytes); |
| 111 | if (prof_buffer) |
| 112 | return 0; |
| 113 | |
| 114 | return -ENOMEM; |
| 115 | } |
| 116 | |
| 117 | static void do_profile_hits(int type, void *__pc, unsigned int nr_hits) |
| 118 | { |
| 119 | unsigned long pc; |
| 120 | pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift; |
| 121 | if (pc < prof_len) |
| 122 | atomic_add(i: nr_hits, v: &prof_buffer[pc]); |
| 123 | } |
| 124 | |
| 125 | void profile_hits(int type, void *__pc, unsigned int nr_hits) |
| 126 | { |
| 127 | if (prof_on != type || !prof_buffer) |
| 128 | return; |
| 129 | do_profile_hits(type, __pc, nr_hits); |
| 130 | } |
| 131 | EXPORT_SYMBOL_GPL(profile_hits); |
| 132 | |
| 133 | void profile_tick(int type) |
| 134 | { |
| 135 | struct pt_regs *regs = get_irq_regs(); |
| 136 | |
| 137 | /* This is the old kernel-only legacy profiling */ |
| 138 | if (!user_mode(regs)) |
| 139 | profile_hit(type, ip: (void *)profile_pc(regs)); |
| 140 | } |
| 141 | |
| 142 | #ifdef CONFIG_PROC_FS |
| 143 | #include <linux/proc_fs.h> |
| 144 | #include <linux/seq_file.h> |
| 145 | #include <linux/uaccess.h> |
| 146 | |
| 147 | /* |
| 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. |
| 152 | */ |
| 153 | static ssize_t |
| 154 | read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos) |
| 155 | { |
| 156 | unsigned long p = *ppos; |
| 157 | ssize_t read; |
| 158 | char *pnt; |
| 159 | unsigned long sample_step = 1UL << prof_shift; |
| 160 | |
| 161 | if (p >= (prof_len+1)*sizeof(unsigned int)) |
| 162 | return 0; |
| 163 | if (count > (prof_len+1)*sizeof(unsigned int) - p) |
| 164 | count = (prof_len+1)*sizeof(unsigned int) - p; |
| 165 | read = 0; |
| 166 | |
| 167 | while (p < sizeof(unsigned int) && count > 0) { |
| 168 | if (put_user(*((char *)(&sample_step)+p), buf)) |
| 169 | return -EFAULT; |
| 170 | buf++; p++; count--; read++; |
| 171 | } |
| 172 | pnt = (char *)prof_buffer + p - sizeof(atomic_t); |
| 173 | if (copy_to_user(to: buf, from: (void *)pnt, n: count)) |
| 174 | return -EFAULT; |
| 175 | read += count; |
| 176 | *ppos += read; |
| 177 | return read; |
| 178 | } |
| 179 | |
| 180 | /* default is to not implement this call */ |
| 181 | int __weak setup_profiling_timer(unsigned mult) |
| 182 | { |
| 183 | return -EINVAL; |
| 184 | } |
| 185 | |
| 186 | /* |
| 187 | * Writing to /proc/profile resets the counters |
| 188 | * |
| 189 | * Writing a 'profiling multiplier' value into it also re-sets the profiling |
| 190 | * interrupt frequency, on architectures that support this. |
| 191 | */ |
| 192 | static ssize_t write_profile(struct file *file, const char __user *buf, |
| 193 | size_t count, loff_t *ppos) |
| 194 | { |
| 195 | #ifdef CONFIG_SMP |
| 196 | if (count == sizeof(int)) { |
| 197 | unsigned int multiplier; |
| 198 | |
| 199 | if (copy_from_user(to: &multiplier, from: buf, n: sizeof(int))) |
| 200 | return -EFAULT; |
| 201 | |
| 202 | if (setup_profiling_timer(multiplier)) |
| 203 | return -EINVAL; |
| 204 | } |
| 205 | #endif |
| 206 | memset(s: prof_buffer, c: 0, n: prof_len * sizeof(atomic_t)); |
| 207 | return count; |
| 208 | } |
| 209 | |
| 210 | static const struct proc_ops profile_proc_ops = { |
| 211 | .proc_read = read_profile, |
| 212 | .proc_write = write_profile, |
| 213 | .proc_lseek = default_llseek, |
| 214 | }; |
| 215 | |
| 216 | int __ref create_proc_profile(void) |
| 217 | { |
| 218 | struct proc_dir_entry *entry; |
| 219 | int err = 0; |
| 220 | |
| 221 | if (!prof_on) |
| 222 | return 0; |
| 223 | entry = proc_create(name: "profile" , S_IWUSR | S_IRUGO, |
| 224 | NULL, proc_ops: &profile_proc_ops); |
| 225 | if (entry) |
| 226 | proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t)); |
| 227 | return err; |
| 228 | } |
| 229 | subsys_initcall(create_proc_profile); |
| 230 | #endif /* CONFIG_PROC_FS */ |
| 231 | |