rjw | 1f88458 | 2022-01-06 17:20:42 +0800 | [diff] [blame^] | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
| 2 | #ifndef _LINUX_CLOSURE_H |
| 3 | #define _LINUX_CLOSURE_H |
| 4 | |
| 5 | #include <linux/llist.h> |
| 6 | #include <linux/sched.h> |
| 7 | #include <linux/sched/task_stack.h> |
| 8 | #include <linux/workqueue.h> |
| 9 | |
| 10 | /* |
| 11 | * Closure is perhaps the most overused and abused term in computer science, but |
| 12 | * since I've been unable to come up with anything better you're stuck with it |
| 13 | * again. |
| 14 | * |
| 15 | * What are closures? |
| 16 | * |
| 17 | * They embed a refcount. The basic idea is they count "things that are in |
| 18 | * progress" - in flight bios, some other thread that's doing something else - |
| 19 | * anything you might want to wait on. |
| 20 | * |
| 21 | * The refcount may be manipulated with closure_get() and closure_put(). |
| 22 | * closure_put() is where many of the interesting things happen, when it causes |
| 23 | * the refcount to go to 0. |
| 24 | * |
| 25 | * Closures can be used to wait on things both synchronously and asynchronously, |
| 26 | * and synchronous and asynchronous use can be mixed without restriction. To |
| 27 | * wait synchronously, use closure_sync() - you will sleep until your closure's |
| 28 | * refcount hits 1. |
| 29 | * |
| 30 | * To wait asynchronously, use |
| 31 | * continue_at(cl, next_function, workqueue); |
| 32 | * |
| 33 | * passing it, as you might expect, the function to run when nothing is pending |
| 34 | * and the workqueue to run that function out of. |
| 35 | * |
| 36 | * continue_at() also, critically, requires a 'return' immediately following the |
| 37 | * location where this macro is referenced, to return to the calling function. |
| 38 | * There's good reason for this. |
| 39 | * |
| 40 | * To use safely closures asynchronously, they must always have a refcount while |
| 41 | * they are running owned by the thread that is running them. Otherwise, suppose |
| 42 | * you submit some bios and wish to have a function run when they all complete: |
| 43 | * |
| 44 | * foo_endio(struct bio *bio) |
| 45 | * { |
| 46 | * closure_put(cl); |
| 47 | * } |
| 48 | * |
| 49 | * closure_init(cl); |
| 50 | * |
| 51 | * do_stuff(); |
| 52 | * closure_get(cl); |
| 53 | * bio1->bi_endio = foo_endio; |
| 54 | * bio_submit(bio1); |
| 55 | * |
| 56 | * do_more_stuff(); |
| 57 | * closure_get(cl); |
| 58 | * bio2->bi_endio = foo_endio; |
| 59 | * bio_submit(bio2); |
| 60 | * |
| 61 | * continue_at(cl, complete_some_read, system_wq); |
| 62 | * |
| 63 | * If closure's refcount started at 0, complete_some_read() could run before the |
| 64 | * second bio was submitted - which is almost always not what you want! More |
| 65 | * importantly, it wouldn't be possible to say whether the original thread or |
| 66 | * complete_some_read()'s thread owned the closure - and whatever state it was |
| 67 | * associated with! |
| 68 | * |
| 69 | * So, closure_init() initializes a closure's refcount to 1 - and when a |
| 70 | * closure_fn is run, the refcount will be reset to 1 first. |
| 71 | * |
| 72 | * Then, the rule is - if you got the refcount with closure_get(), release it |
| 73 | * with closure_put() (i.e, in a bio->bi_endio function). If you have a refcount |
| 74 | * on a closure because you called closure_init() or you were run out of a |
| 75 | * closure - _always_ use continue_at(). Doing so consistently will help |
| 76 | * eliminate an entire class of particularly pernicious races. |
| 77 | * |
| 78 | * Lastly, you might have a wait list dedicated to a specific event, and have no |
| 79 | * need for specifying the condition - you just want to wait until someone runs |
| 80 | * closure_wake_up() on the appropriate wait list. In that case, just use |
| 81 | * closure_wait(). It will return either true or false, depending on whether the |
| 82 | * closure was already on a wait list or not - a closure can only be on one wait |
| 83 | * list at a time. |
| 84 | * |
| 85 | * Parents: |
| 86 | * |
| 87 | * closure_init() takes two arguments - it takes the closure to initialize, and |
| 88 | * a (possibly null) parent. |
| 89 | * |
| 90 | * If parent is non null, the new closure will have a refcount for its lifetime; |
| 91 | * a closure is considered to be "finished" when its refcount hits 0 and the |
| 92 | * function to run is null. Hence |
| 93 | * |
| 94 | * continue_at(cl, NULL, NULL); |
| 95 | * |
| 96 | * returns up the (spaghetti) stack of closures, precisely like normal return |
| 97 | * returns up the C stack. continue_at() with non null fn is better thought of |
| 98 | * as doing a tail call. |
| 99 | * |
| 100 | * All this implies that a closure should typically be embedded in a particular |
| 101 | * struct (which its refcount will normally control the lifetime of), and that |
| 102 | * struct can very much be thought of as a stack frame. |
| 103 | */ |
| 104 | |
| 105 | struct closure; |
| 106 | typedef void (closure_fn) (struct closure *); |
| 107 | |
| 108 | struct closure_waitlist { |
| 109 | struct llist_head list; |
| 110 | }; |
| 111 | |
| 112 | enum closure_state { |
| 113 | /* |
| 114 | * CLOSURE_WAITING: Set iff the closure is on a waitlist. Must be set by |
| 115 | * the thread that owns the closure, and cleared by the thread that's |
| 116 | * waking up the closure. |
| 117 | * |
| 118 | * CLOSURE_SLEEPING: Must be set before a thread uses a closure to sleep |
| 119 | * - indicates that cl->task is valid and closure_put() may wake it up. |
| 120 | * Only set or cleared by the thread that owns the closure. |
| 121 | * |
| 122 | * The rest are for debugging and don't affect behaviour: |
| 123 | * |
| 124 | * CLOSURE_RUNNING: Set when a closure is running (i.e. by |
| 125 | * closure_init() and when closure_put() runs then next function), and |
| 126 | * must be cleared before remaining hits 0. Primarily to help guard |
| 127 | * against incorrect usage and accidentally transferring references. |
| 128 | * continue_at() and closure_return() clear it for you, if you're doing |
| 129 | * something unusual you can use closure_set_dead() which also helps |
| 130 | * annotate where references are being transferred. |
| 131 | * |
| 132 | * CLOSURE_STACK: Sanity check - remaining should never hit 0 on a |
| 133 | * closure with this flag set |
| 134 | */ |
| 135 | |
| 136 | CLOSURE_BITS_START = (1 << 23), |
| 137 | CLOSURE_DESTRUCTOR = (1 << 23), |
| 138 | CLOSURE_WAITING = (1 << 25), |
| 139 | CLOSURE_SLEEPING = (1 << 27), |
| 140 | CLOSURE_RUNNING = (1 << 29), |
| 141 | CLOSURE_STACK = (1 << 31), |
| 142 | }; |
| 143 | |
| 144 | #define CLOSURE_GUARD_MASK \ |
| 145 | ((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_SLEEPING| \ |
| 146 | CLOSURE_RUNNING|CLOSURE_STACK) << 1) |
| 147 | |
| 148 | #define CLOSURE_REMAINING_MASK (CLOSURE_BITS_START - 1) |
| 149 | #define CLOSURE_REMAINING_INITIALIZER (1|CLOSURE_RUNNING) |
| 150 | |
| 151 | struct closure { |
| 152 | union { |
| 153 | struct { |
| 154 | struct workqueue_struct *wq; |
| 155 | struct task_struct *task; |
| 156 | struct llist_node list; |
| 157 | closure_fn *fn; |
| 158 | }; |
| 159 | struct work_struct work; |
| 160 | }; |
| 161 | |
| 162 | struct closure *parent; |
| 163 | |
| 164 | atomic_t remaining; |
| 165 | |
| 166 | #ifdef CONFIG_BCACHE_CLOSURES_DEBUG |
| 167 | #define CLOSURE_MAGIC_DEAD 0xc054dead |
| 168 | #define CLOSURE_MAGIC_ALIVE 0xc054a11e |
| 169 | |
| 170 | unsigned magic; |
| 171 | struct list_head all; |
| 172 | unsigned long ip; |
| 173 | unsigned long waiting_on; |
| 174 | #endif |
| 175 | }; |
| 176 | |
| 177 | void closure_sub(struct closure *cl, int v); |
| 178 | void closure_put(struct closure *cl); |
| 179 | void __closure_wake_up(struct closure_waitlist *list); |
| 180 | bool closure_wait(struct closure_waitlist *list, struct closure *cl); |
| 181 | void closure_sync(struct closure *cl); |
| 182 | |
| 183 | #ifdef CONFIG_BCACHE_CLOSURES_DEBUG |
| 184 | |
| 185 | void closure_debug_init(void); |
| 186 | void closure_debug_create(struct closure *cl); |
| 187 | void closure_debug_destroy(struct closure *cl); |
| 188 | |
| 189 | #else |
| 190 | |
| 191 | static inline void closure_debug_init(void) {} |
| 192 | static inline void closure_debug_create(struct closure *cl) {} |
| 193 | static inline void closure_debug_destroy(struct closure *cl) {} |
| 194 | |
| 195 | #endif |
| 196 | |
| 197 | static inline void closure_set_ip(struct closure *cl) |
| 198 | { |
| 199 | #ifdef CONFIG_BCACHE_CLOSURES_DEBUG |
| 200 | cl->ip = _THIS_IP_; |
| 201 | #endif |
| 202 | } |
| 203 | |
| 204 | static inline void closure_set_ret_ip(struct closure *cl) |
| 205 | { |
| 206 | #ifdef CONFIG_BCACHE_CLOSURES_DEBUG |
| 207 | cl->ip = _RET_IP_; |
| 208 | #endif |
| 209 | } |
| 210 | |
| 211 | static inline void closure_set_waiting(struct closure *cl, unsigned long f) |
| 212 | { |
| 213 | #ifdef CONFIG_BCACHE_CLOSURES_DEBUG |
| 214 | cl->waiting_on = f; |
| 215 | #endif |
| 216 | } |
| 217 | |
| 218 | static inline void __closure_end_sleep(struct closure *cl) |
| 219 | { |
| 220 | __set_current_state(TASK_RUNNING); |
| 221 | |
| 222 | if (atomic_read(&cl->remaining) & CLOSURE_SLEEPING) |
| 223 | atomic_sub(CLOSURE_SLEEPING, &cl->remaining); |
| 224 | } |
| 225 | |
| 226 | static inline void __closure_start_sleep(struct closure *cl) |
| 227 | { |
| 228 | closure_set_ip(cl); |
| 229 | cl->task = current; |
| 230 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 231 | |
| 232 | if (!(atomic_read(&cl->remaining) & CLOSURE_SLEEPING)) |
| 233 | atomic_add(CLOSURE_SLEEPING, &cl->remaining); |
| 234 | } |
| 235 | |
| 236 | static inline void closure_set_stopped(struct closure *cl) |
| 237 | { |
| 238 | atomic_sub(CLOSURE_RUNNING, &cl->remaining); |
| 239 | } |
| 240 | |
| 241 | static inline void set_closure_fn(struct closure *cl, closure_fn *fn, |
| 242 | struct workqueue_struct *wq) |
| 243 | { |
| 244 | BUG_ON(object_is_on_stack(cl)); |
| 245 | closure_set_ip(cl); |
| 246 | cl->fn = fn; |
| 247 | cl->wq = wq; |
| 248 | /* between atomic_dec() in closure_put() */ |
| 249 | smp_mb__before_atomic(); |
| 250 | } |
| 251 | |
| 252 | static inline void closure_queue(struct closure *cl) |
| 253 | { |
| 254 | struct workqueue_struct *wq = cl->wq; |
| 255 | if (wq) { |
| 256 | INIT_WORK(&cl->work, cl->work.func); |
| 257 | BUG_ON(!queue_work(wq, &cl->work)); |
| 258 | } else |
| 259 | cl->fn(cl); |
| 260 | } |
| 261 | |
| 262 | /** |
| 263 | * closure_get - increment a closure's refcount |
| 264 | */ |
| 265 | static inline void closure_get(struct closure *cl) |
| 266 | { |
| 267 | #ifdef CONFIG_BCACHE_CLOSURES_DEBUG |
| 268 | BUG_ON((atomic_inc_return(&cl->remaining) & |
| 269 | CLOSURE_REMAINING_MASK) <= 1); |
| 270 | #else |
| 271 | atomic_inc(&cl->remaining); |
| 272 | #endif |
| 273 | } |
| 274 | |
| 275 | /** |
| 276 | * closure_init - Initialize a closure, setting the refcount to 1 |
| 277 | * @cl: closure to initialize |
| 278 | * @parent: parent of the new closure. cl will take a refcount on it for its |
| 279 | * lifetime; may be NULL. |
| 280 | */ |
| 281 | static inline void closure_init(struct closure *cl, struct closure *parent) |
| 282 | { |
| 283 | memset(cl, 0, sizeof(struct closure)); |
| 284 | cl->parent = parent; |
| 285 | if (parent) |
| 286 | closure_get(parent); |
| 287 | |
| 288 | atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER); |
| 289 | |
| 290 | closure_debug_create(cl); |
| 291 | closure_set_ip(cl); |
| 292 | } |
| 293 | |
| 294 | static inline void closure_init_stack(struct closure *cl) |
| 295 | { |
| 296 | memset(cl, 0, sizeof(struct closure)); |
| 297 | atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER|CLOSURE_STACK); |
| 298 | } |
| 299 | |
| 300 | /** |
| 301 | * closure_wake_up - wake up all closures on a wait list. |
| 302 | */ |
| 303 | static inline void closure_wake_up(struct closure_waitlist *list) |
| 304 | { |
| 305 | smp_mb(); |
| 306 | __closure_wake_up(list); |
| 307 | } |
| 308 | |
| 309 | /** |
| 310 | * continue_at - jump to another function with barrier |
| 311 | * |
| 312 | * After @cl is no longer waiting on anything (i.e. all outstanding refs have |
| 313 | * been dropped with closure_put()), it will resume execution at @fn running out |
| 314 | * of @wq (or, if @wq is NULL, @fn will be called by closure_put() directly). |
| 315 | * |
| 316 | * This is because after calling continue_at() you no longer have a ref on @cl, |
| 317 | * and whatever @cl owns may be freed out from under you - a running closure fn |
| 318 | * has a ref on its own closure which continue_at() drops. |
| 319 | */ |
| 320 | #define continue_at(_cl, _fn, _wq) \ |
| 321 | do { \ |
| 322 | set_closure_fn(_cl, _fn, _wq); \ |
| 323 | closure_sub(_cl, CLOSURE_RUNNING + 1); \ |
| 324 | } while (0) |
| 325 | |
| 326 | /** |
| 327 | * closure_return - finish execution of a closure |
| 328 | * |
| 329 | * This is used to indicate that @cl is finished: when all outstanding refs on |
| 330 | * @cl have been dropped @cl's ref on its parent closure (as passed to |
| 331 | * closure_init()) will be dropped, if one was specified - thus this can be |
| 332 | * thought of as returning to the parent closure. |
| 333 | */ |
| 334 | #define closure_return(_cl) continue_at((_cl), NULL, NULL) |
| 335 | |
| 336 | /** |
| 337 | * continue_at_nobarrier - jump to another function without barrier |
| 338 | * |
| 339 | * Causes @fn to be executed out of @cl, in @wq context (or called directly if |
| 340 | * @wq is NULL). |
| 341 | * |
| 342 | * The ref the caller of continue_at_nobarrier() had on @cl is now owned by @fn, |
| 343 | * thus it's not safe to touch anything protected by @cl after a |
| 344 | * continue_at_nobarrier(). |
| 345 | */ |
| 346 | #define continue_at_nobarrier(_cl, _fn, _wq) \ |
| 347 | do { \ |
| 348 | set_closure_fn(_cl, _fn, _wq); \ |
| 349 | closure_queue(_cl); \ |
| 350 | } while (0) |
| 351 | |
| 352 | /** |
| 353 | * closure_return - finish execution of a closure, with destructor |
| 354 | * |
| 355 | * Works like closure_return(), except @destructor will be called when all |
| 356 | * outstanding refs on @cl have been dropped; @destructor may be used to safely |
| 357 | * free the memory occupied by @cl, and it is called with the ref on the parent |
| 358 | * closure still held - so @destructor could safely return an item to a |
| 359 | * freelist protected by @cl's parent. |
| 360 | */ |
| 361 | #define closure_return_with_destructor(_cl, _destructor) \ |
| 362 | do { \ |
| 363 | set_closure_fn(_cl, _destructor, NULL); \ |
| 364 | closure_sub(_cl, CLOSURE_RUNNING - CLOSURE_DESTRUCTOR + 1); \ |
| 365 | } while (0) |
| 366 | |
| 367 | /** |
| 368 | * closure_call - execute @fn out of a new, uninitialized closure |
| 369 | * |
| 370 | * Typically used when running out of one closure, and we want to run @fn |
| 371 | * asynchronously out of a new closure - @parent will then wait for @cl to |
| 372 | * finish. |
| 373 | */ |
| 374 | static inline void closure_call(struct closure *cl, closure_fn fn, |
| 375 | struct workqueue_struct *wq, |
| 376 | struct closure *parent) |
| 377 | { |
| 378 | closure_init(cl, parent); |
| 379 | continue_at_nobarrier(cl, fn, wq); |
| 380 | } |
| 381 | |
| 382 | #endif /* _LINUX_CLOSURE_H */ |