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/atomic.h>
  53 #include <linux/ktime.h>
  54 #include <linux/export.h>
  55 #include <linux/wait.h>
  56 #include <linux/sched.h>
  57 #include <linux/slab.h>
  58 #include <linux/workqueue.h>
  59
  60 static async_cookie_t next_cookie = 1;
  61
  62 #define MAX_WORK        32768
  63
  64 static LIST_HEAD(async_pending);
  65 static ASYNC_DOMAIN(async_running);
  66 static LIST_HEAD(async_domains);
  67 static DEFINE_SPINLOCK(async_lock);
  68 static DEFINE_MUTEX(async_register_mutex);
  69
  70 struct async_entry {
  71         struct list_head        list;
  72         struct work_struct      work;
  73         async_cookie_t          cookie;
  74         async_func_ptr          *func;
  75         void                    *data;
  76         struct async_domain     *running;
  77 };
  78
  79 static DECLARE_WAIT_QUEUE_HEAD(async_done);
  80
  81 static atomic_t entry_count;
  82
  83
  84 /*
  85  * MUST be called with the lock held!
  86  */
  87 static async_cookie_t  __lowest_in_progress(struct async_domain *running)
  88 {
  89         struct async_entry *entry;
  90
  91         if (!list_empty(&running->domain)) {
  92                 entry = list_first_entry(&running->domain, typeof(*entry), list);
  93                 return entry->cookie;
  94         }
  95
  96         list_for_each_entry(entry, &async_pending, list)
  97                 if (entry->running == running)
  98                         return entry->cookie;
  99
 100         return next_cookie;     /* "infinity" value */
 101 }
 102
 103 static async_cookie_t  lowest_in_progress(struct async_domain *running)
 104 {
 105         unsigned long flags;
 106         async_cookie_t ret;
 107
 108         spin_lock_irqsave(&async_lock, flags);
 109         ret = __lowest_in_progress(running);
 110         spin_unlock_irqrestore(&async_lock, flags);
 111         return ret;
 112 }
 113
 114 /*
 115  * pick the first pending entry and run it
 116  */
 117 static void async_run_entry_fn(struct work_struct *work)
 118 {
 119         struct async_entry *entry =
 120                 container_of(work, struct async_entry, work);
 121         unsigned long flags;
 122         ktime_t uninitialized_var(calltime), delta, rettime;
 123         struct async_domain *running = entry->running;
 124
 125         /* 1) move self to the running queue */
 126         spin_lock_irqsave(&async_lock, flags);
 127         list_move_tail(&entry->list, &running->domain);
 128         spin_unlock_irqrestore(&async_lock, flags);
 129
 130         /* 2) run (and print duration) */
 131         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 132                 printk(KERN_DEBUG "calling  %lli_%pF @ %i\n",
 133                         (long long)entry->cookie,
 134                         entry->func, task_pid_nr(current));
 135                 calltime = ktime_get();
 136         }
 137         entry->func(entry->data, entry->cookie);
 138         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 139                 rettime = ktime_get();
 140                 delta = ktime_sub(rettime, calltime);
 141                 printk(KERN_DEBUG "initcall %lli_%pF returned 0 after %lld usecs\n",
 142                         (long long)entry->cookie,
 143                         entry->func,
 144                         (long long)ktime_to_ns(delta) >> 10);
 145         }
 146
 147         /* 3) remove self from the running queue */
 148         spin_lock_irqsave(&async_lock, flags);
 149         list_del(&entry->list);
 150         if (running->registered && --running->count == 0)
 151                 list_del_init(&running->node);
 152
 153         /* 4) free the entry */
 154         kfree(entry);
 155         atomic_dec(&entry_count);
 156
 157         spin_unlock_irqrestore(&async_lock, flags);
 158
 159         /* 5) wake up any waiters */
 160         wake_up(&async_done);
 161 }
 162
 163 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct async_domain *running)
 164 {
 165         struct async_entry *entry;
 166         unsigned long flags;
 167         async_cookie_t newcookie;
 168
 169         /* allow irq-off callers */
 170         entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
 171
 172         /*
 173          * If we're out of memory or if there's too much work
 174          * pending already, we execute synchronously.
 175          */
 176         if (!entry || atomic_read(&entry_count) > MAX_WORK) {
 177                 kfree(entry);
 178                 spin_lock_irqsave(&async_lock, flags);
 179                 newcookie = next_cookie++;
 180                 spin_unlock_irqrestore(&async_lock, flags);
 181
 182                 /* low on memory.. run synchronously */
 183                 ptr(data, newcookie);
 184                 return newcookie;
 185         }
 186         INIT_WORK(&entry->work, async_run_entry_fn);
 187         entry->func = ptr;
 188         entry->data = data;
 189         entry->running = running;
 190
 191         spin_lock_irqsave(&async_lock, flags);
 192         newcookie = entry->cookie = next_cookie++;
 193         list_add_tail(&entry->list, &async_pending);
 194         if (running->registered && running->count++ == 0)
 195                 list_add_tail(&running->node, &async_domains);
 196         atomic_inc(&entry_count);
 197         spin_unlock_irqrestore(&async_lock, flags);
 198
 199         /* mark that this task has queued an async job, used by module init */
 200         current->flags |= PF_USED_ASYNC;
 201
 202         /* schedule for execution */
 203         queue_work(system_unbound_wq, &entry->work);
 204
 205         return newcookie;
 206 }
 207
 208 /**
 209  * async_schedule - schedule a function for asynchronous execution
 210  * @ptr: function to execute asynchronously
 211  * @data: data pointer to pass to the function
 212  *
 213  * Returns an async_cookie_t that may be used for checkpointing later.
 214  * Note: This function may be called from atomic or non-atomic contexts.
 215  */
 216 async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
 217 {
 218         return __async_schedule(ptr, data, &async_running);
 219 }
 220 EXPORT_SYMBOL_GPL(async_schedule);
 221
 222 /**
 223  * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
 224  * @ptr: function to execute asynchronously
 225  * @data: data pointer to pass to the function
 226  * @running: running list for the domain
 227  *
 228  * Returns an async_cookie_t that may be used for checkpointing later.
 229  * @running may be used in the async_synchronize_*_domain() functions
 230  * to wait within a certain synchronization domain rather than globally.
 231  * A synchronization domain is specified via the running queue @running to use.
 232  * Note: This function may be called from atomic or non-atomic contexts.
 233  */
 234 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
 235                                      struct async_domain *running)
 236 {
 237         return __async_schedule(ptr, data, running);
 238 }
 239 EXPORT_SYMBOL_GPL(async_schedule_domain);
 240
 241 /**
 242  * async_synchronize_full - synchronize all asynchronous function calls
 243  *
 244  * This function waits until all asynchronous function calls have been done.
 245  */
 246 void async_synchronize_full(void)
 247 {
 248         mutex_lock(&async_register_mutex);
 249         do {
 250                 struct async_domain *domain = NULL;
 251
 252                 spin_lock_irq(&async_lock);
 253                 if (!list_empty(&async_domains))
 254                         domain = list_first_entry(&async_domains, typeof(*domain), node);
 255                 spin_unlock_irq(&async_lock);
 256
 257                 async_synchronize_cookie_domain(next_cookie, domain);
 258         } while (!list_empty(&async_domains));
 259         mutex_unlock(&async_register_mutex);
 260 }
 261 EXPORT_SYMBOL_GPL(async_synchronize_full);
 262
 263 /**
 264  * async_unregister_domain - ensure no more anonymous waiters on this domain
 265  * @domain: idle domain to flush out of any async_synchronize_full instances
 266  *
 267  * async_synchronize_{cookie|full}_domain() are not flushed since callers
 268  * of these routines should know the lifetime of @domain
 269  *
 270  * Prefer ASYNC_DOMAIN_EXCLUSIVE() declarations over flushing
 271  */
 272 void async_unregister_domain(struct async_domain *domain)
 273 {
 274         mutex_lock(&async_register_mutex);
 275         spin_lock_irq(&async_lock);
 276         WARN_ON(!domain->registered || !list_empty(&domain->node) ||
 277                 !list_empty(&domain->domain));
 278         domain->registered = 0;
 279         spin_unlock_irq(&async_lock);
 280         mutex_unlock(&async_register_mutex);
 281 }
 282 EXPORT_SYMBOL_GPL(async_unregister_domain);
 283
 284 /**
 285  * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
 286  * @domain: running list to synchronize on
 287  *
 288  * This function waits until all asynchronous function calls for the
 289  * synchronization domain specified by the running list @domain have been done.
 290  */
 291 void async_synchronize_full_domain(struct async_domain *domain)
 292 {
 293         async_synchronize_cookie_domain(next_cookie, domain);
 294 }
 295 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
 296
 297 /**
 298  * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
 299  * @cookie: async_cookie_t to use as checkpoint
 300  * @running: running list to synchronize on
 301  *
 302  * This function waits until all asynchronous function calls for the
 303  * synchronization domain specified by running list @running submitted
 304  * prior to @cookie have been done.
 305  */
 306 void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *running)
 307 {
 308         ktime_t uninitialized_var(starttime), delta, endtime;
 309
 310         if (!running)
 311                 return;
 312
 313         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 314                 printk(KERN_DEBUG "async_waiting @ %i\n", task_pid_nr(current));
 315                 starttime = ktime_get();
 316         }
 317
 318         wait_event(async_done, lowest_in_progress(running) >= cookie);
 319
 320         if (initcall_debug && system_state == SYSTEM_BOOTING) {
 321                 endtime = ktime_get();
 322                 delta = ktime_sub(endtime, starttime);
 323
 324                 printk(KERN_DEBUG "async_continuing @ %i after %lli usec\n",
 325                         task_pid_nr(current),
 326                         (long long)ktime_to_ns(delta) >> 10);
 327         }
 328 }
 329 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
 330
 331 /**
 332  * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
 333  * @cookie: async_cookie_t to use as checkpoint
 334  *
 335  * This function waits until all asynchronous function calls prior to @cookie
 336  * have been done.
 337  */
 338 void async_synchronize_cookie(async_cookie_t cookie)
 339 {
 340         async_synchronize_cookie_domain(cookie, &async_running);
 341 }
 342 EXPORT_SYMBOL_GPL(async_synchronize_cookie);