| 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 | #include "workqueue_internal.h" |
| 61 | |
| 62 | static async_cookie_t next_cookie = 1; |
| 63 | |
| 64 | #define MAX_WORK 32768 |
| 65 | |
| 66 | static LIST_HEAD(async_pending); |
| 67 | static ASYNC_DOMAIN(async_dfl_domain); |
| 68 | static LIST_HEAD(async_domains); |
| 69 | static DEFINE_SPINLOCK(async_lock); |
| 70 | static DEFINE_MUTEX(async_register_mutex); |
| 71 | |
| 72 | struct async_entry { |
| 73 | struct list_head list; |
| 74 | struct work_struct work; |
| 75 | async_cookie_t cookie; |
| 76 | async_func_ptr *func; |
| 77 | void *data; |
| 78 | struct async_domain *domain; |
| 79 | }; |
| 80 | |
| 81 | static DECLARE_WAIT_QUEUE_HEAD(async_done); |
| 82 | |
| 83 | static atomic_t entry_count; |
| 84 | |
| 85 | |
| 86 | /* |
| 87 | * MUST be called with the lock held! |
| 88 | */ |
| 89 | static async_cookie_t __lowest_in_progress(struct async_domain *domain) |
| 90 | { |
| 91 | async_cookie_t first_running = next_cookie; /* infinity value */ |
| 92 | async_cookie_t first_pending = next_cookie; /* ditto */ |
| 93 | struct async_entry *entry; |
| 94 | |
| 95 | /* |
| 96 | * Both running and pending lists are sorted but not disjoint. |
| 97 | * Take the first cookies from both and return the min. |
| 98 | */ |
| 99 | if (!list_empty(&domain->running)) { |
| 100 | entry = list_first_entry(&domain->running, typeof(*entry), list); |
| 101 | first_running = entry->cookie; |
| 102 | } |
| 103 | |
| 104 | list_for_each_entry(entry, &async_pending, list) { |
| 105 | if (entry->domain == domain) { |
| 106 | first_pending = entry->cookie; |
| 107 | break; |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | return min(first_running, first_pending); |
| 112 | } |
| 113 | |
| 114 | static async_cookie_t lowest_in_progress(struct async_domain *domain) |
| 115 | { |
| 116 | unsigned long flags; |
| 117 | async_cookie_t ret; |
| 118 | |
| 119 | spin_lock_irqsave(&async_lock, flags); |
| 120 | ret = __lowest_in_progress(domain); |
| 121 | spin_unlock_irqrestore(&async_lock, flags); |
| 122 | return ret; |
| 123 | } |
| 124 | |
| 125 | /* |
| 126 | * pick the first pending entry and run it |
| 127 | */ |
| 128 | static void async_run_entry_fn(struct work_struct *work) |
| 129 | { |
| 130 | struct async_entry *entry = |
| 131 | container_of(work, struct async_entry, work); |
| 132 | struct async_entry *pos; |
| 133 | unsigned long flags; |
| 134 | ktime_t uninitialized_var(calltime), delta, rettime; |
| 135 | struct async_domain *domain = entry->domain; |
| 136 | |
| 137 | /* 1) move self to the running queue, make sure it stays sorted */ |
| 138 | spin_lock_irqsave(&async_lock, flags); |
| 139 | list_for_each_entry_reverse(pos, &domain->running, list) |
| 140 | if (entry->cookie < pos->cookie) |
| 141 | break; |
| 142 | list_move_tail(&entry->list, &pos->list); |
| 143 | spin_unlock_irqrestore(&async_lock, flags); |
| 144 | |
| 145 | /* 2) run (and print duration) */ |
| 146 | if (initcall_debug && system_state == SYSTEM_BOOTING) { |
| 147 | printk(KERN_DEBUG "calling %lli_%pF @ %i\n", |
| 148 | (long long)entry->cookie, |
| 149 | entry->func, task_pid_nr(current)); |
| 150 | calltime = ktime_get(); |
| 151 | } |
| 152 | entry->func(entry->data, entry->cookie); |
| 153 | if (initcall_debug && system_state == SYSTEM_BOOTING) { |
| 154 | rettime = ktime_get(); |
| 155 | delta = ktime_sub(rettime, calltime); |
| 156 | printk(KERN_DEBUG "initcall %lli_%pF returned 0 after %lld usecs\n", |
| 157 | (long long)entry->cookie, |
| 158 | entry->func, |
| 159 | (long long)ktime_to_ns(delta) >> 10); |
| 160 | } |
| 161 | |
| 162 | /* 3) remove self from the running queue */ |
| 163 | spin_lock_irqsave(&async_lock, flags); |
| 164 | list_del(&entry->list); |
| 165 | if (domain->registered && --domain->count == 0) |
| 166 | list_del_init(&domain->node); |
| 167 | |
| 168 | /* 4) free the entry */ |
| 169 | kfree(entry); |
| 170 | atomic_dec(&entry_count); |
| 171 | |
| 172 | spin_unlock_irqrestore(&async_lock, flags); |
| 173 | |
| 174 | /* 5) wake up any waiters */ |
| 175 | wake_up(&async_done); |
| 176 | } |
| 177 | |
| 178 | static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct async_domain *domain) |
| 179 | { |
| 180 | struct async_entry *entry; |
| 181 | unsigned long flags; |
| 182 | async_cookie_t newcookie; |
| 183 | |
| 184 | /* allow irq-off callers */ |
| 185 | entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC); |
| 186 | |
| 187 | /* |
| 188 | * If we're out of memory or if there's too much work |
| 189 | * pending already, we execute synchronously. |
| 190 | */ |
| 191 | if (!entry || atomic_read(&entry_count) > MAX_WORK) { |
| 192 | kfree(entry); |
| 193 | spin_lock_irqsave(&async_lock, flags); |
| 194 | newcookie = next_cookie++; |
| 195 | spin_unlock_irqrestore(&async_lock, flags); |
| 196 | |
| 197 | /* low on memory.. run synchronously */ |
| 198 | ptr(data, newcookie); |
| 199 | return newcookie; |
| 200 | } |
| 201 | INIT_WORK(&entry->work, async_run_entry_fn); |
| 202 | entry->func = ptr; |
| 203 | entry->data = data; |
| 204 | entry->domain = domain; |
| 205 | |
| 206 | spin_lock_irqsave(&async_lock, flags); |
| 207 | newcookie = entry->cookie = next_cookie++; |
| 208 | list_add_tail(&entry->list, &async_pending); |
| 209 | if (domain->registered && domain->count++ == 0) |
| 210 | list_add_tail(&domain->node, &async_domains); |
| 211 | atomic_inc(&entry_count); |
| 212 | spin_unlock_irqrestore(&async_lock, flags); |
| 213 | |
| 214 | /* mark that this task has queued an async job, used by module init */ |
| 215 | current->flags |= PF_USED_ASYNC; |
| 216 | |
| 217 | /* schedule for execution */ |
| 218 | queue_work(system_unbound_wq, &entry->work); |
| 219 | |
| 220 | return newcookie; |
| 221 | } |
| 222 | |
| 223 | /** |
| 224 | * async_schedule - schedule a function for asynchronous execution |
| 225 | * @ptr: function to execute asynchronously |
| 226 | * @data: data pointer to pass to the function |
| 227 | * |
| 228 | * Returns an async_cookie_t that may be used for checkpointing later. |
| 229 | * Note: This function may be called from atomic or non-atomic contexts. |
| 230 | */ |
| 231 | async_cookie_t async_schedule(async_func_ptr *ptr, void *data) |
| 232 | { |
| 233 | return __async_schedule(ptr, data, &async_dfl_domain); |
| 234 | } |
| 235 | EXPORT_SYMBOL_GPL(async_schedule); |
| 236 | |
| 237 | /** |
| 238 | * async_schedule_domain - schedule a function for asynchronous execution within a certain domain |
| 239 | * @ptr: function to execute asynchronously |
| 240 | * @data: data pointer to pass to the function |
| 241 | * @domain: the domain |
| 242 | * |
| 243 | * Returns an async_cookie_t that may be used for checkpointing later. |
| 244 | * @domain may be used in the async_synchronize_*_domain() functions to |
| 245 | * wait within a certain synchronization domain rather than globally. A |
| 246 | * synchronization domain is specified via @domain. Note: This function |
| 247 | * may be called from atomic or non-atomic contexts. |
| 248 | */ |
| 249 | async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data, |
| 250 | struct async_domain *domain) |
| 251 | { |
| 252 | return __async_schedule(ptr, data, domain); |
| 253 | } |
| 254 | EXPORT_SYMBOL_GPL(async_schedule_domain); |
| 255 | |
| 256 | /** |
| 257 | * async_synchronize_full - synchronize all asynchronous function calls |
| 258 | * |
| 259 | * This function waits until all asynchronous function calls have been done. |
| 260 | */ |
| 261 | void async_synchronize_full(void) |
| 262 | { |
| 263 | mutex_lock(&async_register_mutex); |
| 264 | do { |
| 265 | struct async_domain *domain = NULL; |
| 266 | |
| 267 | spin_lock_irq(&async_lock); |
| 268 | if (!list_empty(&async_domains)) |
| 269 | domain = list_first_entry(&async_domains, typeof(*domain), node); |
| 270 | spin_unlock_irq(&async_lock); |
| 271 | |
| 272 | async_synchronize_cookie_domain(next_cookie, domain); |
| 273 | } while (!list_empty(&async_domains)); |
| 274 | mutex_unlock(&async_register_mutex); |
| 275 | } |
| 276 | EXPORT_SYMBOL_GPL(async_synchronize_full); |
| 277 | |
| 278 | /** |
| 279 | * async_unregister_domain - ensure no more anonymous waiters on this domain |
| 280 | * @domain: idle domain to flush out of any async_synchronize_full instances |
| 281 | * |
| 282 | * async_synchronize_{cookie|full}_domain() are not flushed since callers |
| 283 | * of these routines should know the lifetime of @domain |
| 284 | * |
| 285 | * Prefer ASYNC_DOMAIN_EXCLUSIVE() declarations over flushing |
| 286 | */ |
| 287 | void async_unregister_domain(struct async_domain *domain) |
| 288 | { |
| 289 | mutex_lock(&async_register_mutex); |
| 290 | spin_lock_irq(&async_lock); |
| 291 | WARN_ON(!domain->registered || !list_empty(&domain->node) || |
| 292 | !list_empty(&domain->running)); |
| 293 | domain->registered = 0; |
| 294 | spin_unlock_irq(&async_lock); |
| 295 | mutex_unlock(&async_register_mutex); |
| 296 | } |
| 297 | EXPORT_SYMBOL_GPL(async_unregister_domain); |
| 298 | |
| 299 | /** |
| 300 | * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain |
| 301 | * @domain: the domain to synchronize |
| 302 | * |
| 303 | * This function waits until all asynchronous function calls for the |
| 304 | * synchronization domain specified by @domain have been done. |
| 305 | */ |
| 306 | void async_synchronize_full_domain(struct async_domain *domain) |
| 307 | { |
| 308 | async_synchronize_cookie_domain(next_cookie, domain); |
| 309 | } |
| 310 | EXPORT_SYMBOL_GPL(async_synchronize_full_domain); |
| 311 | |
| 312 | /** |
| 313 | * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing |
| 314 | * @cookie: async_cookie_t to use as checkpoint |
| 315 | * @domain: the domain to synchronize |
| 316 | * |
| 317 | * This function waits until all asynchronous function calls for the |
| 318 | * synchronization domain specified by @domain submitted prior to @cookie |
| 319 | * have been done. |
| 320 | */ |
| 321 | void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain) |
| 322 | { |
| 323 | ktime_t uninitialized_var(starttime), delta, endtime; |
| 324 | |
| 325 | if (!domain) |
| 326 | return; |
| 327 | |
| 328 | if (initcall_debug && system_state == SYSTEM_BOOTING) { |
| 329 | printk(KERN_DEBUG "async_waiting @ %i\n", task_pid_nr(current)); |
| 330 | starttime = ktime_get(); |
| 331 | } |
| 332 | |
| 333 | wait_event(async_done, lowest_in_progress(domain) >= cookie); |
| 334 | |
| 335 | if (initcall_debug && system_state == SYSTEM_BOOTING) { |
| 336 | endtime = ktime_get(); |
| 337 | delta = ktime_sub(endtime, starttime); |
| 338 | |
| 339 | printk(KERN_DEBUG "async_continuing @ %i after %lli usec\n", |
| 340 | task_pid_nr(current), |
| 341 | (long long)ktime_to_ns(delta) >> 10); |
| 342 | } |
| 343 | } |
| 344 | EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain); |
| 345 | |
| 346 | /** |
| 347 | * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing |
| 348 | * @cookie: async_cookie_t to use as checkpoint |
| 349 | * |
| 350 | * This function waits until all asynchronous function calls prior to @cookie |
| 351 | * have been done. |
| 352 | */ |
| 353 | void async_synchronize_cookie(async_cookie_t cookie) |
| 354 | { |
| 355 | async_synchronize_cookie_domain(cookie, &async_dfl_domain); |
| 356 | } |
| 357 | EXPORT_SYMBOL_GPL(async_synchronize_cookie); |
| 358 | |
| 359 | /** |
| 360 | * current_is_async - is %current an async worker task? |
| 361 | * |
| 362 | * Returns %true if %current is an async worker task. |
| 363 | */ |
| 364 | bool current_is_async(void) |
| 365 | { |
| 366 | struct worker *worker = current_wq_worker(); |
| 367 | |
| 368 | return worker && worker->current_func == async_run_entry_fn; |
| 369 | } |