]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - c_src/include/linux/closure.h
move Rust sources to top level, C sources into c_src
[bcachefs-tools-debian] / c_src / include / linux / closure.h
diff --git a/c_src/include/linux/closure.h b/c_src/include/linux/closure.h
new file mode 100644 (file)
index 0000000..c554c6a
--- /dev/null
@@ -0,0 +1,415 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _LINUX_CLOSURE_H
+#define _LINUX_CLOSURE_H
+
+#include <linux/llist.h>
+#include <linux/sched.h>
+#include <linux/sched/task_stack.h>
+#include <linux/workqueue.h>
+
+/*
+ * Closure is perhaps the most overused and abused term in computer science, but
+ * since I've been unable to come up with anything better you're stuck with it
+ * again.
+ *
+ * What are closures?
+ *
+ * They embed a refcount. The basic idea is they count "things that are in
+ * progress" - in flight bios, some other thread that's doing something else -
+ * anything you might want to wait on.
+ *
+ * The refcount may be manipulated with closure_get() and closure_put().
+ * closure_put() is where many of the interesting things happen, when it causes
+ * the refcount to go to 0.
+ *
+ * Closures can be used to wait on things both synchronously and asynchronously,
+ * and synchronous and asynchronous use can be mixed without restriction. To
+ * wait synchronously, use closure_sync() - you will sleep until your closure's
+ * refcount hits 1.
+ *
+ * To wait asynchronously, use
+ *   continue_at(cl, next_function, workqueue);
+ *
+ * passing it, as you might expect, the function to run when nothing is pending
+ * and the workqueue to run that function out of.
+ *
+ * continue_at() also, critically, requires a 'return' immediately following the
+ * location where this macro is referenced, to return to the calling function.
+ * There's good reason for this.
+ *
+ * To use safely closures asynchronously, they must always have a refcount while
+ * they are running owned by the thread that is running them. Otherwise, suppose
+ * you submit some bios and wish to have a function run when they all complete:
+ *
+ * foo_endio(struct bio *bio)
+ * {
+ *     closure_put(cl);
+ * }
+ *
+ * closure_init(cl);
+ *
+ * do_stuff();
+ * closure_get(cl);
+ * bio1->bi_endio = foo_endio;
+ * bio_submit(bio1);
+ *
+ * do_more_stuff();
+ * closure_get(cl);
+ * bio2->bi_endio = foo_endio;
+ * bio_submit(bio2);
+ *
+ * continue_at(cl, complete_some_read, system_wq);
+ *
+ * If closure's refcount started at 0, complete_some_read() could run before the
+ * second bio was submitted - which is almost always not what you want! More
+ * importantly, it wouldn't be possible to say whether the original thread or
+ * complete_some_read()'s thread owned the closure - and whatever state it was
+ * associated with!
+ *
+ * So, closure_init() initializes a closure's refcount to 1 - and when a
+ * closure_fn is run, the refcount will be reset to 1 first.
+ *
+ * Then, the rule is - if you got the refcount with closure_get(), release it
+ * with closure_put() (i.e, in a bio->bi_endio function). If you have a refcount
+ * on a closure because you called closure_init() or you were run out of a
+ * closure - _always_ use continue_at(). Doing so consistently will help
+ * eliminate an entire class of particularly pernicious races.
+ *
+ * Lastly, you might have a wait list dedicated to a specific event, and have no
+ * need for specifying the condition - you just want to wait until someone runs
+ * closure_wake_up() on the appropriate wait list. In that case, just use
+ * closure_wait(). It will return either true or false, depending on whether the
+ * closure was already on a wait list or not - a closure can only be on one wait
+ * list at a time.
+ *
+ * Parents:
+ *
+ * closure_init() takes two arguments - it takes the closure to initialize, and
+ * a (possibly null) parent.
+ *
+ * If parent is non null, the new closure will have a refcount for its lifetime;
+ * a closure is considered to be "finished" when its refcount hits 0 and the
+ * function to run is null. Hence
+ *
+ * continue_at(cl, NULL, NULL);
+ *
+ * returns up the (spaghetti) stack of closures, precisely like normal return
+ * returns up the C stack. continue_at() with non null fn is better thought of
+ * as doing a tail call.
+ *
+ * All this implies that a closure should typically be embedded in a particular
+ * struct (which its refcount will normally control the lifetime of), and that
+ * struct can very much be thought of as a stack frame.
+ */
+
+struct closure;
+struct closure_syncer;
+typedef void (closure_fn) (struct work_struct *);
+extern struct dentry *bcache_debug;
+
+struct closure_waitlist {
+       struct llist_head       list;
+};
+
+enum closure_state {
+       /*
+        * CLOSURE_WAITING: Set iff the closure is on a waitlist. Must be set by
+        * the thread that owns the closure, and cleared by the thread that's
+        * waking up the closure.
+        *
+        * The rest are for debugging and don't affect behaviour:
+        *
+        * CLOSURE_RUNNING: Set when a closure is running (i.e. by
+        * closure_init() and when closure_put() runs then next function), and
+        * must be cleared before remaining hits 0. Primarily to help guard
+        * against incorrect usage and accidentally transferring references.
+        * continue_at() and closure_return() clear it for you, if you're doing
+        * something unusual you can use closure_set_dead() which also helps
+        * annotate where references are being transferred.
+        */
+
+       CLOSURE_BITS_START      = (1U << 26),
+       CLOSURE_DESTRUCTOR      = (1U << 26),
+       CLOSURE_WAITING         = (1U << 28),
+       CLOSURE_RUNNING         = (1U << 30),
+};
+
+#define CLOSURE_GUARD_MASK                                     \
+       ((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_RUNNING) << 1)
+
+#define CLOSURE_REMAINING_MASK         (CLOSURE_BITS_START - 1)
+#define CLOSURE_REMAINING_INITIALIZER  (1|CLOSURE_RUNNING)
+
+struct closure {
+       union {
+               struct {
+                       struct workqueue_struct *wq;
+                       struct closure_syncer   *s;
+                       struct llist_node       list;
+                       closure_fn              *fn;
+               };
+               struct work_struct      work;
+       };
+
+       struct closure          *parent;
+
+       atomic_t                remaining;
+       bool                    closure_get_happened;
+
+#ifdef CONFIG_DEBUG_CLOSURES
+#define CLOSURE_MAGIC_DEAD     0xc054dead
+#define CLOSURE_MAGIC_ALIVE    0xc054a11e
+
+       unsigned int            magic;
+       struct list_head        all;
+       unsigned long           ip;
+       unsigned long           waiting_on;
+#endif
+};
+
+void closure_sub(struct closure *cl, int v);
+void closure_put(struct closure *cl);
+void __closure_wake_up(struct closure_waitlist *list);
+bool closure_wait(struct closure_waitlist *list, struct closure *cl);
+void __closure_sync(struct closure *cl);
+
+static inline unsigned closure_nr_remaining(struct closure *cl)
+{
+       return atomic_read(&cl->remaining) & CLOSURE_REMAINING_MASK;
+}
+
+/**
+ * closure_sync - sleep until a closure a closure has nothing left to wait on
+ *
+ * Sleeps until the refcount hits 1 - the thread that's running the closure owns
+ * the last refcount.
+ */
+static inline void closure_sync(struct closure *cl)
+{
+#ifdef CONFIG_DEBUG_CLOSURES
+       BUG_ON(closure_nr_remaining(cl) != 1 && !cl->closure_get_happened);
+#endif
+
+       if (cl->closure_get_happened)
+               __closure_sync(cl);
+}
+
+#ifdef CONFIG_DEBUG_CLOSURES
+
+void closure_debug_create(struct closure *cl);
+void closure_debug_destroy(struct closure *cl);
+
+#else
+
+static inline void closure_debug_create(struct closure *cl) {}
+static inline void closure_debug_destroy(struct closure *cl) {}
+
+#endif
+
+static inline void closure_set_ip(struct closure *cl)
+{
+#ifdef CONFIG_DEBUG_CLOSURES
+       cl->ip = _THIS_IP_;
+#endif
+}
+
+static inline void closure_set_ret_ip(struct closure *cl)
+{
+#ifdef CONFIG_DEBUG_CLOSURES
+       cl->ip = _RET_IP_;
+#endif
+}
+
+static inline void closure_set_waiting(struct closure *cl, unsigned long f)
+{
+#ifdef CONFIG_DEBUG_CLOSURES
+       cl->waiting_on = f;
+#endif
+}
+
+static inline void closure_set_stopped(struct closure *cl)
+{
+       atomic_sub(CLOSURE_RUNNING, &cl->remaining);
+}
+
+static inline void set_closure_fn(struct closure *cl, closure_fn *fn,
+                                 struct workqueue_struct *wq)
+{
+       closure_set_ip(cl);
+       cl->fn = fn;
+       cl->wq = wq;
+}
+
+static inline void closure_queue(struct closure *cl)
+{
+       struct workqueue_struct *wq = cl->wq;
+       /**
+        * Changes made to closure, work_struct, or a couple of other structs
+        * may cause work.func not pointing to the right location.
+        */
+       BUILD_BUG_ON(offsetof(struct closure, fn)
+                    != offsetof(struct work_struct, func));
+
+       if (wq) {
+               INIT_WORK(&cl->work, cl->work.func);
+               BUG_ON(!queue_work(wq, &cl->work));
+       } else
+               cl->fn(&cl->work);
+}
+
+/**
+ * closure_get - increment a closure's refcount
+ */
+static inline void closure_get(struct closure *cl)
+{
+       cl->closure_get_happened = true;
+
+#ifdef CONFIG_DEBUG_CLOSURES
+       BUG_ON((atomic_inc_return(&cl->remaining) &
+               CLOSURE_REMAINING_MASK) <= 1);
+#else
+       atomic_inc(&cl->remaining);
+#endif
+}
+
+/**
+ * closure_init - Initialize a closure, setting the refcount to 1
+ * @cl:                closure to initialize
+ * @parent:    parent of the new closure. cl will take a refcount on it for its
+ *             lifetime; may be NULL.
+ */
+static inline void closure_init(struct closure *cl, struct closure *parent)
+{
+       cl->fn = NULL;
+       cl->parent = parent;
+       if (parent)
+               closure_get(parent);
+
+       atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
+       cl->closure_get_happened = false;
+
+       closure_debug_create(cl);
+       closure_set_ip(cl);
+}
+
+static inline void closure_init_stack(struct closure *cl)
+{
+       memset(cl, 0, sizeof(struct closure));
+       atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
+}
+
+/**
+ * closure_wake_up - wake up all closures on a wait list,
+ *                  with memory barrier
+ */
+static inline void closure_wake_up(struct closure_waitlist *list)
+{
+       /* Memory barrier for the wait list */
+       smp_mb();
+       __closure_wake_up(list);
+}
+
+#define CLOSURE_CALLBACK(name) void name(struct work_struct *ws)
+#define closure_type(name, type, member)                               \
+       struct closure *cl = container_of(ws, struct closure, work);    \
+       type *name = container_of(cl, type, member)
+
+/**
+ * continue_at - jump to another function with barrier
+ *
+ * After @cl is no longer waiting on anything (i.e. all outstanding refs have
+ * been dropped with closure_put()), it will resume execution at @fn running out
+ * of @wq (or, if @wq is NULL, @fn will be called by closure_put() directly).
+ *
+ * This is because after calling continue_at() you no longer have a ref on @cl,
+ * and whatever @cl owns may be freed out from under you - a running closure fn
+ * has a ref on its own closure which continue_at() drops.
+ *
+ * Note you are expected to immediately return after using this macro.
+ */
+#define continue_at(_cl, _fn, _wq)                                     \
+do {                                                                   \
+       set_closure_fn(_cl, _fn, _wq);                                  \
+       closure_sub(_cl, CLOSURE_RUNNING + 1);                          \
+} while (0)
+
+/**
+ * closure_return - finish execution of a closure
+ *
+ * This is used to indicate that @cl is finished: when all outstanding refs on
+ * @cl have been dropped @cl's ref on its parent closure (as passed to
+ * closure_init()) will be dropped, if one was specified - thus this can be
+ * thought of as returning to the parent closure.
+ */
+#define closure_return(_cl)    continue_at((_cl), NULL, NULL)
+
+/**
+ * continue_at_nobarrier - jump to another function without barrier
+ *
+ * Causes @fn to be executed out of @cl, in @wq context (or called directly if
+ * @wq is NULL).
+ *
+ * The ref the caller of continue_at_nobarrier() had on @cl is now owned by @fn,
+ * thus it's not safe to touch anything protected by @cl after a
+ * continue_at_nobarrier().
+ */
+#define continue_at_nobarrier(_cl, _fn, _wq)                           \
+do {                                                                   \
+       set_closure_fn(_cl, _fn, _wq);                                  \
+       closure_queue(_cl);                                             \
+} while (0)
+
+/**
+ * closure_return_with_destructor - finish execution of a closure,
+ *                                 with destructor
+ *
+ * Works like closure_return(), except @destructor will be called when all
+ * outstanding refs on @cl have been dropped; @destructor may be used to safely
+ * free the memory occupied by @cl, and it is called with the ref on the parent
+ * closure still held - so @destructor could safely return an item to a
+ * freelist protected by @cl's parent.
+ */
+#define closure_return_with_destructor(_cl, _destructor)               \
+do {                                                                   \
+       set_closure_fn(_cl, _destructor, NULL);                         \
+       closure_sub(_cl, CLOSURE_RUNNING - CLOSURE_DESTRUCTOR + 1);     \
+} while (0)
+
+/**
+ * closure_call - execute @fn out of a new, uninitialized closure
+ *
+ * Typically used when running out of one closure, and we want to run @fn
+ * asynchronously out of a new closure - @parent will then wait for @cl to
+ * finish.
+ */
+static inline void closure_call(struct closure *cl, closure_fn fn,
+                               struct workqueue_struct *wq,
+                               struct closure *parent)
+{
+       closure_init(cl, parent);
+       continue_at_nobarrier(cl, fn, wq);
+}
+
+#define __closure_wait_event(waitlist, _cond)                          \
+do {                                                                   \
+       struct closure cl;                                              \
+                                                                       \
+       closure_init_stack(&cl);                                        \
+                                                                       \
+       while (1) {                                                     \
+               closure_wait(waitlist, &cl);                            \
+               if (_cond)                                              \
+                       break;                                          \
+               closure_sync(&cl);                                      \
+       }                                                               \
+       closure_wake_up(waitlist);                                      \
+       closure_sync(&cl);                                              \
+} while (0)
+
+#define closure_wait_event(waitlist, _cond)                            \
+do {                                                                   \
+       if (!(_cond))                                                   \
+               __closure_wait_event(waitlist, _cond);                  \
+} while (0)
+
+#endif /* _LINUX_CLOSURE_H */