kernel/async.c

   1 /*
   2  * async.c: Asynchronous function calls for boot performance
   3  *
   4  * (C) Copyright 2009 Intel Corporation
   5  * Author: Arjan van de Ven <arjan@linux.intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; version 2
  10  * of the License.
  11  */
  12
  13
  14 /*
  15
  16 Goals and Theory of Operation
  17
  18 The primary goal of this feature is to reduce the kernel boot time,
  19 by doing various independent hardware delays and discovery operations
  20 decoupled and not strictly serialized.
  21
  22 More specifically, the asynchronous function call concept allows
  23 certain operations (primarily during system boot) to happen
  24 asynchronously, out of order, while these operations still
  25 have their externally visible parts happen sequentially and in-order.
  26 (not unlike how out-of-order CPUs retire their instructions in order)
  27
  28 Key to the asynchronous function call implementation is the concept of
  29 a "sequence cookie" (which, although it has an abstracted type, can be
  30 thought of as a monotonically incrementing number).
  31
  32 The async core will assign each scheduled event such a sequence cookie and
  33 pass this to the called functions.
  34
  35 The asynchronously called function should before doing a globally visible
  36 operation, such as registering device numbers, call the
  37 async_synchronize_cookie() function and pass in its own cookie. The
  38 async_synchronize_cookie() function will make sure that all asynchronous
  39 operations that were scheduled prior to the operation corresponding with the
  40 cookie have completed.
  41
  42 Subsystem/driver initialization code that scheduled asynchronous probe
  43 functions, but which shares global resources with other drivers/subsystems
  44 that do not use the asynchronous call feature, need to do a full
  45 synchronization with the async_synchronize_full() function, before returning
  46 from their init function. This is to maintain strict ordering between the
  47 asynchronous and synchronous parts of the kernel.
  48
  49 */
  50
  51 #include <linux/async.h>
  52 #include <linux/bug.h>
  53 #include <linux/module.h>
  54 #include <linux/wait.h>
  55 #include <linux/sched.h>
  56 #include <linux/init.h>
  57 #include <linux/kthread.h>
  58 #include <linux/delay.h>
  59 #include <linux/slab.h>
  60 #include <asm/atomic.h>
  61
  62 static async_cookie_t next_cookie = 1;
  63
  64 #define MAX_THREADS     256
  65 #define MAX_WORK        32768
  66
  67 static LIST_HEAD(async_pending);
  68 static LIST_HEAD(async_running);
  69 static DEFINE_SPINLOCK(async_lock);
  70
  71 static int async_enabled = 0;
  72
  73 struct async_entry {
  74         struct list_head list;
  75         async_cookie_t   cookie;
  76         async_func_ptr   *func;
  77         void             *data;
  78         struct list_head *running;
  79 };
  80
  81 static DECLARE_WAIT_QUEUE_HEAD(async_done);
  82 static DECLARE_WAIT_QUEUE_HEAD(async_new);
  83
  84 static atomic_t entry_count;
  85 static atomic_t thread_count;
  86
  87 extern int initcall_debug;
  88
  89
  90 /*
  91  * MUST be called with the lock held!
  92  */
  93 static async_cookie_t  __lowest_in_progress(struct list_head *running)
  94 {
  95         struct async_entry *entry;
  96
  97         if (!list_empty(running)) {
  98                 entry = list_first_entry(running,
  99                         struct async_entry, list);
 100                 return entry->cookie;
 101         }
 102
 103         list_for_each_entry(entry, &async_pending, list)
 104                 if (entry->running == running)
 105                         return entry->cookie;
 106
 107         return next_cookie;     /* "infinity" value */
 108 }
 109
 110 static async_cookie_t  lowest_in_progress(struct list_head *running)
 111 {
 112         unsigned long flags;
 113         async_cookie_t ret;
 114
 115         spin_lock_irqsave(&async_lock, flags);
 116         ret = __lowest_in_progress(running);
 117         spin_unlock_irqrestore(&async_lock, flags);
 118         return ret;
 119 }
 120 /*
 121  * pick the first pending entry and run it
 122  */
 123 static void run_one_entry(void)
 124 {
 125         unsigned long flags;
 126         struct async_entry *entry;
 127         ktime_t calltime, delta, rettime;
 128
 129         /* 1) pick one task from the pending queue */
 130
 131         spin_lock_irqsave(&async_lock, flags);
 132         if (list_empty(&async_pending))
 133                 goto out;
 134         entry = list_first_entry(&async_pending, struct async_entry, list);
 135
 136         /* 2) move it to the running queue */
 137         list_move_tail(&entry->list, entry->running);
 138         spin_unlock_irqrestore(&async_lock, flags);
 139
 140         /* 3) run it (and print duration)*/
 141         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 142                 printk("calling  %lli_%pF @ %i\n", (long long)entry->cookie,
 143                         entry->func, task_pid_nr(current));
 144                 calltime = ktime_get();
 145         }
 146         entry->func(entry->data, entry->cookie);
 147         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 148                 rettime = ktime_get();
 149                 delta = ktime_sub(rettime, calltime);
 150                 printk("initcall %lli_%pF returned 0 after %lld usecs\n",
 151                         (long long)entry->cookie,
 152                         entry->func,
 153                         (long long)ktime_to_ns(delta) >> 10);
 154         }
 155
 156         /* 4) remove it from the running queue */
 157         spin_lock_irqsave(&async_lock, flags);
 158         list_del(&entry->list);
 159
 160         /* 5) free the entry  */
 161         kfree(entry);
 162         atomic_dec(&entry_count);
 163
 164         spin_unlock_irqrestore(&async_lock, flags);
 165
 166         /* 6) wake up any waiters. */
 167         wake_up(&async_done);
 168         return;
 169
 170 out:
 171         spin_unlock_irqrestore(&async_lock, flags);
 172 }
 173
 174
 175 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
 176 {
 177         struct async_entry *entry;
 178         unsigned long flags;
 179         async_cookie_t newcookie;
 180
 181
 182         /* allow irq-off callers */
 183         entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
 184
 185         /*
 186          * If we're out of memory or if there's too much work
 187          * pending already, we execute synchronously.
 188          */
 189         if (!async_enabled || !entry || atomic_read(&entry_count) > MAX_WORK) {
 190                 kfree(entry);
 191                 spin_lock_irqsave(&async_lock, flags);
 192                 newcookie = next_cookie++;
 193                 spin_unlock_irqrestore(&async_lock, flags);
 194
 195                 /* low on memory.. run synchronously */
 196                 ptr(data, newcookie);
 197                 return newcookie;
 198         }
 199         entry->func = ptr;
 200         entry->data = data;
 201         entry->running = running;
 202
 203         spin_lock_irqsave(&async_lock, flags);
 204         newcookie = entry->cookie = next_cookie++;
 205         list_add_tail(&entry->list, &async_pending);
 206         atomic_inc(&entry_count);
 207         spin_unlock_irqrestore(&async_lock, flags);
 208         wake_up(&async_new);
 209         return newcookie;
 210 }
 211
 212 /**
 213  * async_schedule - schedule a function for asynchronous execution
 214  * @ptr: function to execute asynchronously
 215  * @data: data pointer to pass to the function
 216  *
 217  * Returns an async_cookie_t that may be used for checkpointing later.
 218  * Note: This function may be called from atomic or non-atomic contexts.
 219  */
 220 async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
 221 {
 222         return __async_schedule(ptr, data, &async_running);
 223 }
 224 EXPORT_SYMBOL_GPL(async_schedule);
 225
 226 /**
 227  * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
 228  * @ptr: function to execute asynchronously
 229  * @data: data pointer to pass to the function
 230  * @running: running list for the domain
 231  *
 232  * Returns an async_cookie_t that may be used for checkpointing later.
 233  * @running may be used in the async_synchronize_*_domain() functions
 234  * to wait within a certain synchronization domain rather than globally.
 235  * A synchronization domain is specified via the running queue @running to use.
 236  * Note: This function may be called from atomic or non-atomic contexts.
 237  */
 238 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
 239                                      struct list_head *running)
 240 {
 241         return __async_schedule(ptr, data, running);
 242 }
 243 EXPORT_SYMBOL_GPL(async_schedule_domain);
 244
 245 /**
 246  * async_synchronize_full - synchronize all asynchronous function calls
 247  *
 248  * This function waits until all asynchronous function calls have been done.
 249  */
 250 void async_synchronize_full(void)
 251 {
 252         do {
 253                 async_synchronize_cookie(next_cookie);
 254         } while (!list_empty(&async_running) || !list_empty(&async_pending));
 255 }
 256 EXPORT_SYMBOL_GPL(async_synchronize_full);
 257
 258 /**
 259  * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
 260  * @list: running list to synchronize on
 261  *
 262  * This function waits until all asynchronous function calls for the
 263  * synchronization domain specified by the running list @list have been done.
 264  */
 265 void async_synchronize_full_domain(struct list_head *list)
 266 {
 267         async_synchronize_cookie_domain(next_cookie, list);
 268 }
 269 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
 270
 271 /**
 272  * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
 273  * @cookie: async_cookie_t to use as checkpoint
 274  * @running: running list to synchronize on
 275  *
 276  * This function waits until all asynchronous function calls for the
 277  * synchronization domain specified by the running list @list submitted
 278  * prior to @cookie have been done.
 279  */
 280 void async_synchronize_cookie_domain(async_cookie_t cookie,
 281                                      struct list_head *running)
 282 {
 283         ktime_t starttime, delta, endtime;
 284
 285         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 286                 printk("async_waiting @ %i\n", task_pid_nr(current));
 287                 starttime = ktime_get();
 288         }
 289
 290         wait_event(async_done, lowest_in_progress(running) >= cookie);
 291
 292         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 293                 endtime = ktime_get();
 294                 delta = ktime_sub(endtime, starttime);
 295
 296                 printk("async_continuing @ %i after %lli usec\n",
 297                         task_pid_nr(current),
 298                         (long long)ktime_to_ns(delta) >> 10);
 299         }
 300 }
 301 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
 302
 303 /**
 304  * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
 305  * @cookie: async_cookie_t to use as checkpoint
 306  *
 307  * This function waits until all asynchronous function calls prior to @cookie
 308  * have been done.
 309  */
 310 void async_synchronize_cookie(async_cookie_t cookie)
 311 {
 312         async_synchronize_cookie_domain(cookie, &async_running);
 313 }
 314 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
 315
 316
 317 static int async_thread(void *unused)
 318 {
 319         DECLARE_WAITQUEUE(wq, current);
 320         add_wait_queue(&async_new, &wq);
 321
 322         while (!kthread_should_stop()) {
 323                 int ret = HZ;
 324                 set_current_state(TASK_INTERRUPTIBLE);
 325                 /*
 326                  * check the list head without lock.. false positives
 327                  * are dealt with inside run_one_entry() while holding
 328                  * the lock.
 329                  */
 330                 rmb();
 331                 if (!list_empty(&async_pending))
 332                         run_one_entry();
 333                 else
 334                         ret = schedule_timeout(HZ);
 335
 336                 if (ret == 0) {
 337                         /*
 338                          * we timed out, this means we as thread are redundant.
 339                          * we sign off and die, but we to avoid any races there
 340                          * is a last-straw check to see if work snuck in.
 341                          */
 342                         atomic_dec(&thread_count);
 343                         wmb(); /* manager must see our departure first */
 344                         if (list_empty(&async_pending))
 345                                 break;
 346                         /*
 347                          * woops work came in between us timing out and us
 348                          * signing off; we need to stay alive and keep working.
 349                          */
 350                         atomic_inc(&thread_count);
 351                 }
 352         }
 353         remove_wait_queue(&async_new, &wq);
 354
 355         return 0;
 356 }
 357
 358 static int async_manager_thread(void *unused)
 359 {
 360         DECLARE_WAITQUEUE(wq, current);
 361         add_wait_queue(&async_new, &wq);
 362
 363         while (!kthread_should_stop()) {
 364                 int tc, ec;
 365
 366                 set_current_state(TASK_INTERRUPTIBLE);
 367
 368                 tc = atomic_read(&thread_count);
 369                 rmb();
 370                 ec = atomic_read(&entry_count);
 371
 372                 while (tc < ec && tc < MAX_THREADS) {
 373                         if (IS_ERR(kthread_run(async_thread, NULL, "async/%i",
 374                                                tc))) {
 375                                 msleep(100);
 376                                 continue;
 377                         }
 378                         atomic_inc(&thread_count);
 379                         tc++;
 380                 }
 381
 382                 schedule();
 383         }
 384         remove_wait_queue(&async_new, &wq);
 385
 386         return 0;
 387 }
 388
 389 static int __init async_init(void)
 390 {
 391         async_enabled =
 392                 !IS_ERR(kthread_run(async_manager_thread, NULL, "async/mgr"));
 393
 394         WARN_ON(!async_enabled);
 395         return 0;
 396 }
 397
 398 core_initcall(async_init);