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 <asm/atomic.h>
  60
  61 static async_cookie_t next_cookie = 1;
  62
  63 #define MAX_THREADS     256
  64 #define MAX_WORK        32768
  65
  66 static LIST_HEAD(async_pending);
  67 static LIST_HEAD(async_running);
  68 static DEFINE_SPINLOCK(async_lock);
  69
  70 static int async_enabled = 0;
  71
  72 struct async_entry {
  73         struct list_head list;
  74         async_cookie_t   cookie;
  75         async_func_ptr   *func;
  76         void             *data;
  77         struct list_head *running;
  78 };
  79
  80 static DECLARE_WAIT_QUEUE_HEAD(async_done);
  81 static DECLARE_WAIT_QUEUE_HEAD(async_new);
  82
  83 static atomic_t entry_count;
  84 static atomic_t thread_count;
  85
  86 extern int initcall_debug;
  87
  88
  89 /*
  90  * MUST be called with the lock held!
  91  */
  92 static async_cookie_t  __lowest_in_progress(struct list_head *running)
  93 {
  94         struct async_entry *entry;
  95         async_cookie_t ret = next_cookie; /* begin with "infinity" value */
  96
  97         if (!list_empty(running)) {
  98                 entry = list_first_entry(running,
  99                         struct async_entry, list);
 100                 ret = entry->cookie;
 101         }
 102
 103         if (!list_empty(&async_pending)) {
 104                 list_for_each_entry(entry, &async_pending, list)
 105                         if (entry->running == running) {
 106                                 ret = entry->cookie;
 107                                 break;
 108                         }
 109         }
 110
 111         return ret;
 112 }
 113
 114 static async_cookie_t  lowest_in_progress(struct list_head *running)
 115 {
 116         unsigned long flags;
 117         async_cookie_t ret;
 118
 119         spin_lock_irqsave(&async_lock, flags);
 120         ret = __lowest_in_progress(running);
 121         spin_unlock_irqrestore(&async_lock, flags);
 122         return ret;
 123 }
 124 /*
 125  * pick the first pending entry and run it
 126  */
 127 static void run_one_entry(void)
 128 {
 129         unsigned long flags;
 130         struct async_entry *entry;
 131         ktime_t calltime, delta, rettime;
 132
 133         /* 1) pick one task from the pending queue */
 134
 135         spin_lock_irqsave(&async_lock, flags);
 136         if (list_empty(&async_pending))
 137                 goto out;
 138         entry = list_first_entry(&async_pending, struct async_entry, list);
 139
 140         /* 2) move it to the running queue */
 141         list_move_tail(&entry->list, entry->running);
 142         spin_unlock_irqrestore(&async_lock, flags);
 143
 144         /* 3) run it (and print duration)*/
 145         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 146                 printk("calling  %lli_%pF @ %i\n", (long long)entry->cookie,
 147                         entry->func, task_pid_nr(current));
 148                 calltime = ktime_get();
 149         }
 150         entry->func(entry->data, entry->cookie);
 151         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 152                 rettime = ktime_get();
 153                 delta = ktime_sub(rettime, calltime);
 154                 printk("initcall %lli_%pF returned 0 after %lld usecs\n",
 155                         (long long)entry->cookie,
 156                         entry->func,
 157                         (long long)ktime_to_ns(delta) >> 10);
 158         }
 159
 160         /* 4) remove it from the running queue */
 161         spin_lock_irqsave(&async_lock, flags);
 162         list_del(&entry->list);
 163
 164         /* 5) free the entry  */
 165         kfree(entry);
 166         atomic_dec(&entry_count);
 167
 168         spin_unlock_irqrestore(&async_lock, flags);
 169
 170         /* 6) wake up any waiters. */
 171         wake_up(&async_done);
 172         return;
 173
 174 out:
 175         spin_unlock_irqrestore(&async_lock, flags);
 176 }
 177
 178
 179 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
 180 {
 181         struct async_entry *entry;
 182         unsigned long flags;
 183         async_cookie_t newcookie;
 184
 185
 186         /* allow irq-off callers */
 187         entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
 188
 189         /*
 190          * If we're out of memory or if there's too much work
 191          * pending already, we execute synchronously.
 192          */
 193         if (!async_enabled || !entry || atomic_read(&entry_count) > MAX_WORK) {
 194                 kfree(entry);
 195                 spin_lock_irqsave(&async_lock, flags);
 196                 newcookie = next_cookie++;
 197                 spin_unlock_irqrestore(&async_lock, flags);
 198
 199                 /* low on memory.. run synchronously */
 200                 ptr(data, newcookie);
 201                 return newcookie;
 202         }
 203         entry->func = ptr;
 204         entry->data = data;
 205         entry->running = running;
 206
 207         spin_lock_irqsave(&async_lock, flags);
 208         newcookie = entry->cookie = next_cookie++;
 209         list_add_tail(&entry->list, &async_pending);
 210         atomic_inc(&entry_count);
 211         spin_unlock_irqrestore(&async_lock, flags);
 212         wake_up(&async_new);
 213         return newcookie;
 214 }
 215
 216 /**
 217  * async_schedule - schedule a function for asynchronous execution
 218  * @ptr: function to execute asynchronously
 219  * @data: data pointer to pass to the function
 220  *
 221  * Returns an async_cookie_t that may be used for checkpointing later.
 222  * Note: This function may be called from atomic or non-atomic contexts.
 223  */
 224 async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
 225 {
 226         return __async_schedule(ptr, data, &async_running);
 227 }
 228 EXPORT_SYMBOL_GPL(async_schedule);
 229
 230 /**
 231  * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
 232  * @ptr: function to execute asynchronously
 233  * @data: data pointer to pass to the function
 234  * @running: running list for the domain
 235  *
 236  * Returns an async_cookie_t that may be used for checkpointing later.
 237  * @running may be used in the async_synchronize_*_domain() functions
 238  * to wait within a certain synchronization domain rather than globally.
 239  * A synchronization domain is specified via the running queue @running to use.
 240  * Note: This function may be called from atomic or non-atomic contexts.
 241  */
 242 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
 243                                      struct list_head *running)
 244 {
 245         return __async_schedule(ptr, data, running);
 246 }
 247 EXPORT_SYMBOL_GPL(async_schedule_domain);
 248
 249 /**
 250  * async_synchronize_full - synchronize all asynchronous function calls
 251  *
 252  * This function waits until all asynchronous function calls have been done.
 253  */
 254 void async_synchronize_full(void)
 255 {
 256         do {
 257                 async_synchronize_cookie(next_cookie);
 258         } while (!list_empty(&async_running) || !list_empty(&async_pending));
 259 }
 260 EXPORT_SYMBOL_GPL(async_synchronize_full);
 261
 262 /**
 263  * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
 264  * @list: running list to synchronize on
 265  *
 266  * This function waits until all asynchronous function calls for the
 267  * synchronization domain specified by the running list @list have been done.
 268  */
 269 void async_synchronize_full_domain(struct list_head *list)
 270 {
 271         async_synchronize_cookie_domain(next_cookie, list);
 272 }
 273 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
 274
 275 /**
 276  * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
 277  * @cookie: async_cookie_t to use as checkpoint
 278  * @running: running list to synchronize on
 279  *
 280  * This function waits until all asynchronous function calls for the
 281  * synchronization domain specified by the running list @list submitted
 282  * prior to @cookie have been done.
 283  */
 284 void async_synchronize_cookie_domain(async_cookie_t cookie,
 285                                      struct list_head *running)
 286 {
 287         ktime_t starttime, delta, endtime;
 288
 289         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 290                 printk("async_waiting @ %i\n", task_pid_nr(current));
 291                 starttime = ktime_get();
 292         }
 293
 294         wait_event(async_done, lowest_in_progress(running) >= cookie);
 295
 296         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 297                 endtime = ktime_get();
 298                 delta = ktime_sub(endtime, starttime);
 299
 300                 printk("async_continuing @ %i after %lli usec\n",
 301                         task_pid_nr(current),
 302                         (long long)ktime_to_ns(delta) >> 10);
 303         }
 304 }
 305 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
 306
 307 /**
 308  * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
 309  * @cookie: async_cookie_t to use as checkpoint
 310  *
 311  * This function waits until all asynchronous function calls prior to @cookie
 312  * have been done.
 313  */
 314 void async_synchronize_cookie(async_cookie_t cookie)
 315 {
 316         async_synchronize_cookie_domain(cookie, &async_running);
 317 }
 318 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
 319
 320
 321 static int async_thread(void *unused)
 322 {
 323         DECLARE_WAITQUEUE(wq, current);
 324         add_wait_queue(&async_new, &wq);
 325
 326         while (!kthread_should_stop()) {
 327                 int ret = HZ;
 328                 set_current_state(TASK_INTERRUPTIBLE);
 329                 /*
 330                  * check the list head without lock.. false positives
 331                  * are dealt with inside run_one_entry() while holding
 332                  * the lock.
 333                  */
 334                 rmb();
 335                 if (!list_empty(&async_pending))
 336                         run_one_entry();
 337                 else
 338                         ret = schedule_timeout(HZ);
 339
 340                 if (ret == 0) {
 341                         /*
 342                          * we timed out, this means we as thread are redundant.
 343                          * we sign off and die, but we to avoid any races there
 344                          * is a last-straw check to see if work snuck in.
 345                          */
 346                         atomic_dec(&thread_count);
 347                         wmb(); /* manager must see our departure first */
 348                         if (list_empty(&async_pending))
 349                                 break;
 350                         /*
 351                          * woops work came in between us timing out and us
 352                          * signing off; we need to stay alive and keep working.
 353                          */
 354                         atomic_inc(&thread_count);
 355                 }
 356         }
 357         remove_wait_queue(&async_new, &wq);
 358
 359         return 0;
 360 }
 361
 362 static int async_manager_thread(void *unused)
 363 {
 364         DECLARE_WAITQUEUE(wq, current);
 365         add_wait_queue(&async_new, &wq);
 366
 367         while (!kthread_should_stop()) {
 368                 int tc, ec;
 369
 370                 set_current_state(TASK_INTERRUPTIBLE);
 371
 372                 tc = atomic_read(&thread_count);
 373                 rmb();
 374                 ec = atomic_read(&entry_count);
 375
 376                 while (tc < ec && tc < MAX_THREADS) {
 377                         if (IS_ERR(kthread_run(async_thread, NULL, "async/%i",
 378                                                tc))) {
 379                                 msleep(100);
 380                                 continue;
 381                         }
 382                         atomic_inc(&thread_count);
 383                         tc++;
 384                 }
 385
 386                 schedule();
 387         }
 388         remove_wait_queue(&async_new, &wq);
 389
 390         return 0;
 391 }
 392
 393 static int __init async_init(void)
 394 {
 395         async_enabled =
 396                 !IS_ERR(kthread_run(async_manager_thread, NULL, "async/mgr"));
 397
 398         WARN_ON(!async_enabled);
 399         return 0;
 400 }
 401
 402 core_initcall(async_init);