drivers/md/bcache/closure.h

   1 #ifndef _LINUX_CLOSURE_H
   2 #define _LINUX_CLOSURE_H
   3
   4 #include <linux/llist.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/task_stack.h>
   7 #include <linux/workqueue.h>
   8
   9 /*
  10  * Closure is perhaps the most overused and abused term in computer science, but
  11  * since I've been unable to come up with anything better you're stuck with it
  12  * again.
  13  *
  14  * What are closures?
  15  *
  16  * They embed a refcount. The basic idea is they count "things that are in
  17  * progress" - in flight bios, some other thread that's doing something else -
  18  * anything you might want to wait on.
  19  *
  20  * The refcount may be manipulated with closure_get() and closure_put().
  21  * closure_put() is where many of the interesting things happen, when it causes
  22  * the refcount to go to 0.
  23  *
  24  * Closures can be used to wait on things both synchronously and asynchronously,
  25  * and synchronous and asynchronous use can be mixed without restriction. To
  26  * wait synchronously, use closure_sync() - you will sleep until your closure's
  27  * refcount hits 1.
  28  *
  29  * To wait asynchronously, use
  30  *   continue_at(cl, next_function, workqueue);
  31  *
  32  * passing it, as you might expect, the function to run when nothing is pending
  33  * and the workqueue to run that function out of.
  34  *
  35  * continue_at() also, critically, requires a 'return' immediately following the
  36  * location where this macro is referenced, to return to the calling function.
  37  * There's good reason for this.
  38  *
  39  * To use safely closures asynchronously, they must always have a refcount while
  40  * they are running owned by the thread that is running them. Otherwise, suppose
  41  * you submit some bios and wish to have a function run when they all complete:
  42  *
  43  * foo_endio(struct bio *bio)
  44  * {
  45  *      closure_put(cl);
  46  * }
  47  *
  48  * closure_init(cl);
  49  *
  50  * do_stuff();
  51  * closure_get(cl);
  52  * bio1->bi_endio = foo_endio;
  53  * bio_submit(bio1);
  54  *
  55  * do_more_stuff();
  56  * closure_get(cl);
  57  * bio2->bi_endio = foo_endio;
  58  * bio_submit(bio2);
  59  *
  60  * continue_at(cl, complete_some_read, system_wq);
  61  *
  62  * If closure's refcount started at 0, complete_some_read() could run before the
  63  * second bio was submitted - which is almost always not what you want! More
  64  * importantly, it wouldn't be possible to say whether the original thread or
  65  * complete_some_read()'s thread owned the closure - and whatever state it was
  66  * associated with!
  67  *
  68  * So, closure_init() initializes a closure's refcount to 1 - and when a
  69  * closure_fn is run, the refcount will be reset to 1 first.
  70  *
  71  * Then, the rule is - if you got the refcount with closure_get(), release it
  72  * with closure_put() (i.e, in a bio->bi_endio function). If you have a refcount
  73  * on a closure because you called closure_init() or you were run out of a
  74  * closure - _always_ use continue_at(). Doing so consistently will help
  75  * eliminate an entire class of particularly pernicious races.
  76  *
  77  * Lastly, you might have a wait list dedicated to a specific event, and have no
  78  * need for specifying the condition - you just want to wait until someone runs
  79  * closure_wake_up() on the appropriate wait list. In that case, just use
  80  * closure_wait(). It will return either true or false, depending on whether the
  81  * closure was already on a wait list or not - a closure can only be on one wait
  82  * list at a time.
  83  *
  84  * Parents:
  85  *
  86  * closure_init() takes two arguments - it takes the closure to initialize, and
  87  * a (possibly null) parent.
  88  *
  89  * If parent is non null, the new closure will have a refcount for its lifetime;
  90  * a closure is considered to be "finished" when its refcount hits 0 and the
  91  * function to run is null. Hence
  92  *
  93  * continue_at(cl, NULL, NULL);
  94  *
  95  * returns up the (spaghetti) stack of closures, precisely like normal return
  96  * returns up the C stack. continue_at() with non null fn is better thought of
  97  * as doing a tail call.
  98  *
  99  * All this implies that a closure should typically be embedded in a particular
 100  * struct (which its refcount will normally control the lifetime of), and that
 101  * struct can very much be thought of as a stack frame.
 102  */
 103
 104 struct closure;
 105 typedef void (closure_fn) (struct closure *);
 106
 107 struct closure_waitlist {
 108         struct llist_head       list;
 109 };
 110
 111 enum closure_state {
 112         /*
 113          * CLOSURE_WAITING: Set iff the closure is on a waitlist. Must be set by
 114          * the thread that owns the closure, and cleared by the thread that's
 115          * waking up the closure.
 116          *
 117          * CLOSURE_SLEEPING: Must be set before a thread uses a closure to sleep
 118          * - indicates that cl->task is valid and closure_put() may wake it up.
 119          * Only set or cleared by the thread that owns the closure.
 120          *
 121          * The rest are for debugging and don't affect behaviour:
 122          *
 123          * CLOSURE_RUNNING: Set when a closure is running (i.e. by
 124          * closure_init() and when closure_put() runs then next function), and
 125          * must be cleared before remaining hits 0. Primarily to help guard
 126          * against incorrect usage and accidentally transferring references.
 127          * continue_at() and closure_return() clear it for you, if you're doing
 128          * something unusual you can use closure_set_dead() which also helps
 129          * annotate where references are being transferred.
 130          *
 131          * CLOSURE_STACK: Sanity check - remaining should never hit 0 on a
 132          * closure with this flag set
 133          */
 134
 135         CLOSURE_BITS_START      = (1 << 23),
 136         CLOSURE_DESTRUCTOR      = (1 << 23),
 137         CLOSURE_WAITING         = (1 << 25),
 138         CLOSURE_SLEEPING        = (1 << 27),
 139         CLOSURE_RUNNING         = (1 << 29),
 140         CLOSURE_STACK           = (1 << 31),
 141 };
 142
 143 #define CLOSURE_GUARD_MASK                                      \
 144         ((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_SLEEPING|  \
 145           CLOSURE_RUNNING|CLOSURE_STACK) << 1)
 146
 147 #define CLOSURE_REMAINING_MASK          (CLOSURE_BITS_START - 1)
 148 #define CLOSURE_REMAINING_INITIALIZER   (1|CLOSURE_RUNNING)
 149
 150 struct closure {
 151         union {
 152                 struct {
 153                         struct workqueue_struct *wq;
 154                         struct task_struct      *task;
 155                         struct llist_node       list;
 156                         closure_fn              *fn;
 157                 };
 158                 struct work_struct      work;
 159         };
 160
 161         struct closure          *parent;
 162
 163         atomic_t                remaining;
 164
 165 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 166 #define CLOSURE_MAGIC_DEAD      0xc054dead
 167 #define CLOSURE_MAGIC_ALIVE     0xc054a11e
 168
 169         unsigned                magic;
 170         struct list_head        all;
 171         unsigned long           ip;
 172         unsigned long           waiting_on;
 173 #endif
 174 };
 175
 176 void closure_sub(struct closure *cl, int v);
 177 void closure_put(struct closure *cl);
 178 void __closure_wake_up(struct closure_waitlist *list);
 179 bool closure_wait(struct closure_waitlist *list, struct closure *cl);
 180 void closure_sync(struct closure *cl);
 181
 182 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 183
 184 void closure_debug_init(void);
 185 void closure_debug_create(struct closure *cl);
 186 void closure_debug_destroy(struct closure *cl);
 187
 188 #else
 189
 190 static inline void closure_debug_init(void) {}
 191 static inline void closure_debug_create(struct closure *cl) {}
 192 static inline void closure_debug_destroy(struct closure *cl) {}
 193
 194 #endif
 195
 196 static inline void closure_set_ip(struct closure *cl)
 197 {
 198 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 199         cl->ip = _THIS_IP_;
 200 #endif
 201 }
 202
 203 static inline void closure_set_ret_ip(struct closure *cl)
 204 {
 205 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 206         cl->ip = _RET_IP_;
 207 #endif
 208 }
 209
 210 static inline void closure_set_waiting(struct closure *cl, unsigned long f)
 211 {
 212 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 213         cl->waiting_on = f;
 214 #endif
 215 }
 216
 217 static inline void __closure_end_sleep(struct closure *cl)
 218 {
 219         __set_current_state(TASK_RUNNING);
 220
 221         if (atomic_read(&cl->remaining) & CLOSURE_SLEEPING)
 222                 atomic_sub(CLOSURE_SLEEPING, &cl->remaining);
 223 }
 224
 225 static inline void __closure_start_sleep(struct closure *cl)
 226 {
 227         closure_set_ip(cl);
 228         cl->task = current;
 229         set_current_state(TASK_UNINTERRUPTIBLE);
 230
 231         if (!(atomic_read(&cl->remaining) & CLOSURE_SLEEPING))
 232                 atomic_add(CLOSURE_SLEEPING, &cl->remaining);
 233 }
 234
 235 static inline void closure_set_stopped(struct closure *cl)
 236 {
 237         atomic_sub(CLOSURE_RUNNING, &cl->remaining);
 238 }
 239
 240 static inline void set_closure_fn(struct closure *cl, closure_fn *fn,
 241                                   struct workqueue_struct *wq)
 242 {
 243         BUG_ON(object_is_on_stack(cl));
 244         closure_set_ip(cl);
 245         cl->fn = fn;
 246         cl->wq = wq;
 247         /* between atomic_dec() in closure_put() */
 248         smp_mb__before_atomic();
 249 }
 250
 251 static inline void closure_queue(struct closure *cl)
 252 {
 253         struct workqueue_struct *wq = cl->wq;
 254         if (wq) {
 255                 INIT_WORK(&cl->work, cl->work.func);
 256                 BUG_ON(!queue_work(wq, &cl->work));
 257         } else
 258                 cl->fn(cl);
 259 }
 260
 261 /**
 262  * closure_get - increment a closure's refcount
 263  */
 264 static inline void closure_get(struct closure *cl)
 265 {
 266 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 267         BUG_ON((atomic_inc_return(&cl->remaining) &
 268                 CLOSURE_REMAINING_MASK) <= 1);
 269 #else
 270         atomic_inc(&cl->remaining);
 271 #endif
 272 }
 273
 274 /**
 275  * closure_init - Initialize a closure, setting the refcount to 1
 276  * @cl:         closure to initialize
 277  * @parent:     parent of the new closure. cl will take a refcount on it for its
 278  *              lifetime; may be NULL.
 279  */
 280 static inline void closure_init(struct closure *cl, struct closure *parent)
 281 {
 282         memset(cl, 0, sizeof(struct closure));
 283         cl->parent = parent;
 284         if (parent)
 285                 closure_get(parent);
 286
 287         atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
 288
 289         closure_debug_create(cl);
 290         closure_set_ip(cl);
 291 }
 292
 293 static inline void closure_init_stack(struct closure *cl)
 294 {
 295         memset(cl, 0, sizeof(struct closure));
 296         atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER|CLOSURE_STACK);
 297 }
 298
 299 /**
 300  * closure_wake_up - wake up all closures on a wait list.
 301  */
 302 static inline void closure_wake_up(struct closure_waitlist *list)
 303 {
 304         smp_mb();
 305         __closure_wake_up(list);
 306 }
 307
 308 /**
 309  * continue_at - jump to another function with barrier
 310  *
 311  * After @cl is no longer waiting on anything (i.e. all outstanding refs have
 312  * been dropped with closure_put()), it will resume execution at @fn running out
 313  * of @wq (or, if @wq is NULL, @fn will be called by closure_put() directly).
 314  *
 315  * NOTE: This macro expands to a return in the calling function!
 316  *
 317  * This is because after calling continue_at() you no longer have a ref on @cl,
 318  * and whatever @cl owns may be freed out from under you - a running closure fn
 319  * has a ref on its own closure which continue_at() drops.
 320  */
 321 #define continue_at(_cl, _fn, _wq)                                      \
 322 do {                                                                    \
 323         set_closure_fn(_cl, _fn, _wq);                                  \
 324         closure_sub(_cl, CLOSURE_RUNNING + 1);                          \
 325 } while (0)
 326
 327 /**
 328  * closure_return - finish execution of a closure
 329  *
 330  * This is used to indicate that @cl is finished: when all outstanding refs on
 331  * @cl have been dropped @cl's ref on its parent closure (as passed to
 332  * closure_init()) will be dropped, if one was specified - thus this can be
 333  * thought of as returning to the parent closure.
 334  */
 335 #define closure_return(_cl)     continue_at((_cl), NULL, NULL)
 336
 337 /**
 338  * continue_at_nobarrier - jump to another function without barrier
 339  *
 340  * Causes @fn to be executed out of @cl, in @wq context (or called directly if
 341  * @wq is NULL).
 342  *
 343  * NOTE: like continue_at(), this macro expands to a return in the caller!
 344  *
 345  * The ref the caller of continue_at_nobarrier() had on @cl is now owned by @fn,
 346  * thus it's not safe to touch anything protected by @cl after a
 347  * continue_at_nobarrier().
 348  */
 349 #define continue_at_nobarrier(_cl, _fn, _wq)                            \
 350 do {                                                                    \
 351         set_closure_fn(_cl, _fn, _wq);                                  \
 352         closure_queue(_cl);                                             \
 353 } while (0)
 354
 355 /**
 356  * closure_return - finish execution of a closure, with destructor
 357  *
 358  * Works like closure_return(), except @destructor will be called when all
 359  * outstanding refs on @cl have been dropped; @destructor may be used to safely
 360  * free the memory occupied by @cl, and it is called with the ref on the parent
 361  * closure still held - so @destructor could safely return an item to a
 362  * freelist protected by @cl's parent.
 363  */
 364 #define closure_return_with_destructor(_cl, _destructor)                \
 365 do {                                                                    \
 366         set_closure_fn(_cl, _destructor, NULL);                         \
 367         closure_sub(_cl, CLOSURE_RUNNING - CLOSURE_DESTRUCTOR + 1);     \
 368 } while (0)
 369
 370 /**
 371  * closure_call - execute @fn out of a new, uninitialized closure
 372  *
 373  * Typically used when running out of one closure, and we want to run @fn
 374  * asynchronously out of a new closure - @parent will then wait for @cl to
 375  * finish.
 376  */
 377 static inline void closure_call(struct closure *cl, closure_fn fn,
 378                                 struct workqueue_struct *wq,
 379                                 struct closure *parent)
 380 {
 381         closure_init(cl, parent);
 382         continue_at_nobarrier(cl, fn, wq);
 383 }
 384
 385 #endif /* _LINUX_CLOSURE_H */