| 1 |
/* -*- linux-c -*- |
| 2 |
* |
| 3 |
* /proc transport and control |
| 4 |
* Copyright (C) 2005-2018 Red Hat Inc. |
| 5 |
* |
| 6 |
* This file is part of systemtap, and is free software. You can |
| 7 |
* redistribute it and/or modify it under the terms of the GNU General |
| 8 |
* Public License (GPL); either version 2, or (at your option) any |
| 9 |
* later version. |
| 10 |
*/ |
| 11 |
|
| 12 |
#include "relay_compat.h" |
| 13 |
#include "proc_fs_compatibility.h" |
| 14 |
|
| 15 |
#if defined(STAPCONF_PATH_LOOKUP) && !defined(STAPCONF_KERN_PATH_PARENT) |
| 16 |
#define kern_path_parent(name, nameidata) \ |
| 17 |
path_lookup(name, LOOKUP_PARENT, nameidata) |
| 18 |
#endif |
| 19 |
|
| 20 |
/* _stp_procfs_module_dir is the '/proc/systemtap/{module_name}' directory. */ |
| 21 |
static struct proc_dir_entry *_stp_procfs_module_dir = NULL; |
| 22 |
static struct path _stp_procfs_module_dir_path; |
| 23 |
|
| 24 |
static bool _stp_proc_dir_exists(void) |
| 25 |
{ |
| 26 |
int found = 0; |
| 27 |
#if defined(STAPCONF_PATH_LOOKUP) || defined(STAPCONF_KERN_PATH_PARENT) |
| 28 |
struct nameidata nd; |
| 29 |
#else /* STAPCONF_VFS_PATH_LOOKUP or STAPCONF_KERN_PATH */ |
| 30 |
struct path path; |
| 31 |
#if defined(STAPCONF_VFS_PATH_LOOKUP) |
| 32 |
struct vfsmount *mnt; |
| 33 |
#endif |
| 34 |
int rc; |
| 35 |
#endif /* STAPCONF_VFS_PATH_LOOKUP or STAPCONF_KERN_PATH */ |
| 36 |
|
| 37 |
#if defined(STAPCONF_PATH_LOOKUP) || defined(STAPCONF_KERN_PATH_PARENT) |
| 38 |
/* Why "/proc/systemtap/foo"? kern_path_parent() is basically |
| 39 |
* the same thing as calling the old path_lookup() with flags |
| 40 |
* set to LOOKUP_PARENT, which means to look up the parent of |
| 41 |
* the path, which in this case is "/proc/systemtap". */ |
| 42 |
if (! kern_path_parent("/proc/systemtap/foo", &nd)) { |
| 43 |
found = 1; |
| 44 |
#ifdef STAPCONF_NAMEIDATA_CLEANUP |
| 45 |
path_put(&nd.path); |
| 46 |
#else /* !STAPCONF_NAMEIDATA_CLEANUP */ |
| 47 |
path_release(&nd); |
| 48 |
#endif /* !STAPCONF_NAMEIDATA_CLEANUP */ |
| 49 |
} |
| 50 |
|
| 51 |
#elif defined(STAPCONF_KERN_PATH) |
| 52 |
/* Prefer kern_path() over vfs_path_lookup(), since on some |
| 53 |
* kernels the declaration for vfs_path_lookup() was moved to |
| 54 |
* a private header. */ |
| 55 |
|
| 56 |
/* See if '/proc/systemtap' exists. */ |
| 57 |
rc = kern_path("/proc/systemtap", 0, &path); |
| 58 |
if (rc == 0) { |
| 59 |
found = 1; |
| 60 |
path_put (&path); |
| 61 |
} |
| 62 |
|
| 63 |
#else /* STAPCONF_VFS_PATH_LOOKUP */ |
| 64 |
/* See if '/proc/systemtap' exists. */ |
| 65 |
mnt = init_pid_ns.proc_mnt; |
| 66 |
rc = vfs_path_lookup(mnt->mnt_root, mnt, "systemtap", 0, &path); |
| 67 |
if (rc == 0) { |
| 68 |
found = 1; |
| 69 |
path_put (&path); |
| 70 |
} |
| 71 |
#endif /* STAPCONF_VFS_PATH_LOOKUP */ |
| 72 |
|
| 73 |
return found; |
| 74 |
} |
| 75 |
|
| 76 |
/* |
| 77 |
* Safely creates '/proc/systemtap' (if necessary) and |
| 78 |
* '/proc/systemtap/{module_name}'. |
| 79 |
* |
| 80 |
* NB: this function is suitable to call from early in the the |
| 81 |
* module-init function, and doesn't rely on any other facilities |
| 82 |
* in our runtime. PR19833. See also PR15408. |
| 83 |
*/ |
| 84 |
static int _stp_mkdir_proc_module(void) |
| 85 |
{ |
| 86 |
static char proc_root_name[STP_MODULE_NAME_LEN + sizeof("systemtap/")]; |
| 87 |
int rc; |
| 88 |
|
| 89 |
if (_stp_procfs_module_dir != NULL) |
| 90 |
return 0; |
| 91 |
|
| 92 |
/* If we couldn't find "/proc/systemtap", create it. */ |
| 93 |
if (!_stp_proc_dir_exists()) { |
| 94 |
/* |
| 95 |
* We need some sleepable way to synchronize with other stap |
| 96 |
* modules which are also being loaded for the first time on |
| 97 |
* this system. The `/proc/systemtap` directory is never removed |
| 98 |
* after it's made, so this race only happens briefly the first |
| 99 |
* time stap is used on the current boot. On 3.19+ kernels, the |
| 100 |
* race results in only a WARN and proc_mkdir() failing; nothing |
| 101 |
* more than that. However, on kernels <3.19, proc_mkdir() |
| 102 |
* doesn't error out when a duplicate directory is made, and |
| 103 |
* instead there are leaks in addition to the WARN (see kernel |
| 104 |
* commit b208d54b7539 for details). We'd like to fix the leak |
| 105 |
* and ideally not scare sysadmins with WARNs, so we abuse |
| 106 |
* `module_mutex` in the kernel for mutual exclusion between all |
| 107 |
* stap modules. Since `module_mutex` isn't ours to abuse |
| 108 |
* freely, we elide it by checking if `/proc/systemtap` exists |
| 109 |
* first, and if it doesn't, then we check again after taking |
| 110 |
* the lock. This means we'll only use `module_mutex` and |
| 111 |
* redundantly check for the existence of `/proc/systemtap` just |
| 112 |
* once for each of the first stap modules loaded on the system, |
| 113 |
* and only for those stap modules which encounter the race. |
| 114 |
* After the race window, we're back to just the single check |
| 115 |
* for `/proc/systemtap` and nothing more. |
| 116 |
* |
| 117 |
* This doesn't work on 5.12+ kernels though, as `module_mutex` |
| 118 |
* is no longer exported, but that isn't a big deal. Since |
| 119 |
* there's no risk of a leak on 5.12+ kernels, the worst that |
| 120 |
* can happen is a cosmetic WARN. |
| 121 |
* |
| 122 |
* We never need to check proc_mkdir() for an error because, if |
| 123 |
* it fails on 5.12+ without `module_mutex` due to the directory |
| 124 |
* already existing, then it's guaranteed that the directory |
| 125 |
* will be immediately available to use since procfs serializes |
| 126 |
* the existence check and the registration under the same hold |
| 127 |
* of a global lock. And if there's a proc_mkdir() error even |
| 128 |
* with `module_mutex`, then the other proc_mkdir() attempt |
| 129 |
* below which *is* checked for errors will fail anyway and |
| 130 |
* produce a fatal error message. |
| 131 |
*/ |
| 132 |
#ifdef STAPCONF_MODULE_MUTEX |
| 133 |
mutex_lock(&module_mutex); |
| 134 |
if (!_stp_proc_dir_exists()) |
| 135 |
proc_mkdir("systemtap", NULL); |
| 136 |
mutex_unlock(&module_mutex); |
| 137 |
#else |
| 138 |
proc_mkdir("systemtap", NULL); |
| 139 |
#endif |
| 140 |
} |
| 141 |
|
| 142 |
/* Create the "systemtap/{module_name} directory in procfs. */ |
| 143 |
strlcpy(proc_root_name, "/proc/systemtap/", sizeof(proc_root_name)); |
| 144 |
strlcat(proc_root_name, THIS_MODULE->name, sizeof(proc_root_name)); |
| 145 |
_stp_procfs_module_dir = proc_mkdir(&proc_root_name[6], NULL); // skip the /proc/ |
| 146 |
#ifdef STAPCONF_PROCFS_OWNER |
| 147 |
if (_stp_procfs_module_dir != NULL) |
| 148 |
_stp_procfs_module_dir->owner = THIS_MODULE; |
| 149 |
#endif |
| 150 |
if (_stp_procfs_module_dir == NULL) |
| 151 |
errk("Unable to create '/proc/systemap/%s':" |
| 152 |
" proc_mkdir failed.\n", THIS_MODULE->name); |
| 153 |
else { |
| 154 |
rc = kern_path(proc_root_name, 0, &_stp_procfs_module_dir_path); |
| 155 |
if (rc != 0) { |
| 156 |
errk("Unable to resolve /proc/systemap/%s':" |
| 157 |
" to path.\n", THIS_MODULE->name); |
| 158 |
proc_remove(_stp_procfs_module_dir); |
| 159 |
_stp_procfs_module_dir = NULL; |
| 160 |
return rc; |
| 161 |
} |
| 162 |
} |
| 163 |
|
| 164 |
done: |
| 165 |
return (_stp_procfs_module_dir) ? 0 : -EINVAL; |
| 166 |
} |
| 167 |
|
| 168 |
|
| 169 |
/* |
| 170 |
* Removes '/proc/systemtap/{module_name}'. Notice we're leaving |
| 171 |
* '/proc/systemtap' behind. There is no way on newer kernels to know |
| 172 |
* if a procfs directory is empty. |
| 173 |
* |
| 174 |
* NB: this is suitable to call late in the module cleanup function, |
| 175 |
* and does not rely on any other facilities in the runtime. PR19833. |
| 176 |
* See also PR15408. |
| 177 |
*/ |
| 178 |
static void _stp_rmdir_proc_module(void) |
| 179 |
{ |
| 180 |
if (_stp_procfs_module_dir) { |
| 181 |
path_put(& _stp_procfs_module_dir_path); |
| 182 |
proc_remove(_stp_procfs_module_dir); |
| 183 |
_stp_procfs_module_dir = NULL; |
| 184 |
} |
| 185 |
} |
| 186 |
|
| 187 |
|
| 188 |
inline static int _stp_procfs_ctl_write_fs(int type, void *data, unsigned len) |
| 189 |
{ |
| 190 |
struct _stp_buffer *bptr; |
| 191 |
unsigned long flags; |
| 192 |
|
| 193 |
#define WRITE_AGG |
| 194 |
#ifdef WRITE_AGG |
| 195 |
stp_spin_lock_irqsave(&_stp_ctl_ready_lock, flags); |
| 196 |
if (!list_empty(&_stp_ctl_ready_q)) { |
| 197 |
bptr = (struct _stp_buffer *)_stp_ctl_ready_q.prev; |
| 198 |
if ((bptr->len + len) <= STP_CTL_BUFFER_SIZE |
| 199 |
&& type == STP_REALTIME_DATA |
| 200 |
&& bptr->type == STP_REALTIME_DATA) { |
| 201 |
memcpy(bptr->buf + bptr->len, data, len); |
| 202 |
bptr->len += len; |
| 203 |
stp_spin_unlock_irqrestore(&_stp_ctl_ready_lock, flags); |
| 204 |
return len; |
| 205 |
} |
| 206 |
} |
| 207 |
stp_spin_unlock_irqrestore(&_stp_ctl_ready_lock, flags); |
| 208 |
#endif |
| 209 |
return 0; |
| 210 |
} |
| 211 |
|
| 212 |
static int _stp_proc_ctl_read_bufsize(char *page, char **start, off_t off, int count, int *eof, void *data) |
| 213 |
{ |
| 214 |
int len = sprintf(page, "%d,%d\n", _stp_nsubbufs, _stp_subbuf_size); |
| 215 |
if (len <= off + count) |
| 216 |
*eof = 1; |
| 217 |
*start = page + off; |
| 218 |
len -= off; |
| 219 |
if (len > count) |
| 220 |
len = count; |
| 221 |
if (len < 0) |
| 222 |
len = 0; |
| 223 |
return len; |
| 224 |
} |
| 225 |
|
| 226 |
|
| 227 |
static struct file_operations _stp_ctl_fops_cmd; |
| 228 |
#ifdef STAPCONF_PROC_OPS /* control.c */ |
| 229 |
static struct proc_ops _stp_ctl_proc_ops_cmd; |
| 230 |
#endif |
| 231 |
|
| 232 |
|
| 233 |
static int _stp_procfs_register_ctl_channel_fs(void) |
| 234 |
{ |
| 235 |
struct proc_dir_entry *bs = NULL; |
| 236 |
struct proc_dir_entry *de; |
| 237 |
|
| 238 |
if (_stp_mkdir_proc_module()) |
| 239 |
goto err0; |
| 240 |
|
| 241 |
/* create /proc/systemtap/module_name/.cmd */ |
| 242 |
#ifdef STAPCONF_PROC_OPS |
| 243 |
de = proc_create(".cmd", 0600, _stp_procfs_module_dir, &_stp_ctl_proc_ops_cmd); |
| 244 |
#else |
| 245 |
de = proc_create(".cmd", 0600, _stp_procfs_module_dir, &_stp_ctl_fops_cmd); |
| 246 |
#endif |
| 247 |
if (de == NULL) |
| 248 |
goto err1; |
| 249 |
proc_set_user(de, KUIDT_INIT(_stp_uid), KGIDT_INIT(_stp_gid)); |
| 250 |
|
| 251 |
return 0; |
| 252 |
|
| 253 |
err1: |
| 254 |
_stp_rmdir_proc_module(); |
| 255 |
err0: |
| 256 |
return -1; |
| 257 |
} |
| 258 |
|
| 259 |
static void _stp_procfs_unregister_ctl_channel_fs(void) |
| 260 |
{ |
| 261 |
remove_proc_entry(".cmd", _stp_procfs_module_dir); |
| 262 |
_stp_rmdir_proc_module(); |
| 263 |
} |
| 264 |
|
| 265 |
|
| 266 |
|
| 267 |
#ifdef STAPCONF_PROC_OPS |
| 268 |
struct proc_ops relay_procfs_operations; |
| 269 |
#else |
| 270 |
struct file_operations relay_procfs_operations; |
| 271 |
#endif |
| 272 |
|
| 273 |
|
| 274 |
// We need to map procfs concepts of proc_dir_entry* and relayfs/vfs of path/dentry*. |
| 275 |
struct procfs_relay_file |
| 276 |
{ |
| 277 |
struct path p; // contains the dentry* |
| 278 |
struct proc_dir_entry *pde; // entry valid if this pointer non-NULL |
| 279 |
}; |
| 280 |
struct procfs_relay_file *p_r_files; |
| 281 |
|
| 282 |
|
| 283 |
static int _stp_procfs_transport_fs_init(const char *module_name) |
| 284 |
{ |
| 285 |
p_r_files = _stp_vzalloc(num_possible_cpus() |
| 286 |
* sizeof(struct procfs_relay_file)); |
| 287 |
if (unlikely(p_r_files == NULL)) |
| 288 |
return -ENOMEM; |
| 289 |
|
| 290 |
#ifdef STAPCONF_PROC_OPS |
| 291 |
relay_procfs_operations.proc_open = __stp_relay_file_open; |
| 292 |
relay_procfs_operations.proc_poll = __stp_relay_file_poll; |
| 293 |
relay_procfs_operations.proc_mmap = relay_file_operations.mmap; |
| 294 |
relay_procfs_operations.proc_read = __stp_relay_file_read; |
| 295 |
relay_procfs_operations.proc_lseek = relay_file_operations.llseek; |
| 296 |
relay_procfs_operations.proc_release = relay_file_operations.release; |
| 297 |
#else |
| 298 |
relay_procfs_operations = relay_file_operations; |
| 299 |
relay_procfs_operations.open = __stp_relay_file_open; |
| 300 |
relay_procfs_operations.owner = THIS_MODULE; |
| 301 |
relay_procfs_operations.poll = __stp_relay_file_poll; |
| 302 |
relay_procfs_operations.read = __stp_relay_file_read; |
| 303 |
#endif |
| 304 |
|
| 305 |
if (_stp_mkdir_proc_module()) { // get the _stp_procfs_module_dir* created |
| 306 |
_stp_vfree (p_r_files); |
| 307 |
p_r_files = NULL; |
| 308 |
return -1; |
| 309 |
} |
| 310 |
|
| 311 |
dbug_trans(1, "transport_fs_init dentry=%08lx pde=%08lx ", |
| 312 |
(unsigned long) _stp_procfs_module_dir_path.dentry, |
| 313 |
(unsigned long) _stp_procfs_module_dir); |
| 314 |
|
| 315 |
if (_stp_transport_data_fs_init() != 0) { |
| 316 |
_stp_rmdir_proc_module(); |
| 317 |
_stp_vfree (p_r_files); |
| 318 |
p_r_files = NULL; |
| 319 |
return -1; |
| 320 |
} |
| 321 |
|
| 322 |
return 0; |
| 323 |
} |
| 324 |
|
| 325 |
|
| 326 |
static void _stp_procfs_transport_fs_close(void) |
| 327 |
{ |
| 328 |
_stp_transport_data_fs_close(); |
| 329 |
|
| 330 |
if (likely(p_r_files)) { |
| 331 |
_stp_vfree (p_r_files); |
| 332 |
p_r_files = NULL; |
| 333 |
} |
| 334 |
} |
| 335 |
|
| 336 |
|
| 337 |
static struct dentry *_stp_procfs_get_module_dir(void) |
| 338 |
{ |
| 339 |
return _stp_procfs_module_dir_path.dentry; |
| 340 |
} |
| 341 |
|
| 342 |
|
| 343 |
static int __stp_procfs_relay_remove_buf_file_callback(struct dentry *dentry) |
| 344 |
{ |
| 345 |
unsigned i; |
| 346 |
struct proc_dir_entry *pde = NULL; |
| 347 |
|
| 348 |
// find the corresponding pde* |
| 349 |
|
| 350 |
/* NB We cannot use the for_each_online_cpu() here since online |
| 351 |
* CPUs may get changed on-the-fly through the CPU hotplug feature |
| 352 |
* of the kernel. |
| 353 |
*/ |
| 354 |
for_each_possible_cpu(i) |
| 355 |
{ |
| 356 |
if (p_r_files[i].pde != NULL && |
| 357 |
p_r_files[i].p.dentry == dentry) |
| 358 |
break; |
| 359 |
} |
| 360 |
|
| 361 |
if (i != num_possible_cpus()) |
| 362 |
{ |
| 363 |
pde = p_r_files[i].pde; |
| 364 |
proc_remove (pde); |
| 365 |
path_put (& p_r_files[i].p); |
| 366 |
p_r_files[i].pde = NULL; |
| 367 |
} |
| 368 |
|
| 369 |
dbug_trans(1, "remove-buf dentry=%08lx pde=%08lx i=%u", |
| 370 |
(unsigned long) dentry, (unsigned long) pde, i); |
| 371 |
return 0; |
| 372 |
} |
| 373 |
|
| 374 |
|
| 375 |
static struct dentry * |
| 376 |
__stp_procfs_relay_create_buf_file_callback(const char *filename, |
| 377 |
struct dentry *parent, |
| 378 |
#ifdef STAPCONF_RELAY_UMODE_T |
| 379 |
umode_t mode, |
| 380 |
#else |
| 381 |
int mode, |
| 382 |
#endif |
| 383 |
struct rchan_buf *buf, |
| 384 |
int *is_global) |
| 385 |
{ |
| 386 |
int rc = 0; |
| 387 |
struct dentry* de = NULL; |
| 388 |
char fullpath[sizeof("/proc/systemtap") + STP_MODULE_NAME_LEN + sizeof("/traceNNNNN") + 42]; |
| 389 |
struct proc_dir_entry *pde; |
| 390 |
unsigned i = 0; |
| 391 |
struct inode* in; |
| 392 |
|
| 393 |
if (is_global) |
| 394 |
*is_global = 0; |
| 395 |
|
| 396 |
if (parent != _stp_procfs_module_dir_path.dentry) |
| 397 |
goto out; |
| 398 |
|
| 399 |
pde = proc_create (filename, 0400, |
| 400 |
_stp_procfs_module_dir, |
| 401 |
& relay_procfs_operations); |
| 402 |
if (pde == NULL) |
| 403 |
goto out; |
| 404 |
|
| 405 |
proc_set_user(pde, KUIDT_INIT(_stp_uid), KGIDT_INIT(_stp_gid)); |
| 406 |
|
| 407 |
rc = snprintf(fullpath, sizeof(fullpath), "/proc/systemtap/%s/%s", |
| 408 |
THIS_MODULE->name, filename); |
| 409 |
|
| 410 |
// find spot to plop this |
| 411 |
|
| 412 |
/* NB We cannot use the for_each_online_cpu() here since online |
| 413 |
* CPUs may get changed on-the-fly through the CPU hotplug feature |
| 414 |
* of the kernel. |
| 415 |
*/ |
| 416 |
for_each_possible_cpu(i) |
| 417 |
{ |
| 418 |
if (p_r_files[i].pde == NULL) |
| 419 |
break; |
| 420 |
} |
| 421 |
if (i == num_possible_cpus()) |
| 422 |
goto out1; |
| 423 |
|
| 424 |
rc = kern_path (fullpath, 0, &p_r_files[i].p); |
| 425 |
if (rc) |
| 426 |
goto out1; |
| 427 |
p_r_files[i].pde = pde; |
| 428 |
de = p_r_files[i].p.dentry; |
| 429 |
|
| 430 |
// fill in the relayfs i_private |
| 431 |
in = de->d_inode; |
| 432 |
in->i_private = buf; |
| 433 |
|
| 434 |
// success! |
| 435 |
goto out; |
| 436 |
|
| 437 |
out1: |
| 438 |
proc_remove (pde); |
| 439 |
|
| 440 |
out: |
| 441 |
dbug_trans(1, "create-buf name=%s parent=%08lx -> i=%u rc=%d de=%08lx", |
| 442 |
filename, (unsigned long) parent, |
| 443 |
i, rc, (unsigned long) de); |
| 444 |
return de; |
| 445 |
} |