futex: Fix uninterruptible loop due to gate_area
[linux-2.6.git] / kernel / workqueue.c
index f57855f..1783aab 100644 (file)
@@ -1,19 +1,26 @@
 /*
- * linux/kernel/workqueue.c
+ * kernel/workqueue.c - generic async execution with shared worker pool
  *
- * Generic mechanism for defining kernel helper threads for running
- * arbitrary tasks in process context.
+ * Copyright (C) 2002          Ingo Molnar
  *
- * Started by Ingo Molnar, Copyright (C) 2002
+ *   Derived from the taskqueue/keventd code by:
+ *     David Woodhouse <dwmw2@infradead.org>
+ *     Andrew Morton
+ *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
+ *     Theodore Ts'o <tytso@mit.edu>
  *
- * Derived from the taskqueue/keventd code by:
+ * Made to use alloc_percpu by Christoph Lameter.
  *
- *   David Woodhouse <dwmw2@infradead.org>
- *   Andrew Morton
- *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
- *   Theodore Ts'o <tytso@mit.edu>
+ * Copyright (C) 2010          SUSE Linux Products GmbH
+ * Copyright (C) 2010          Tejun Heo <tj@kernel.org>
  *
- * Made to use alloc_percpu by Christoph Lameter.
+ * This is the generic async execution mechanism.  Work items as are
+ * executed in process context.  The worker pool is shared and
+ * automatically managed.  There is one worker pool for each CPU and
+ * one extra for works which are better served by workers which are
+ * not bound to any specific CPU.
+ *
+ * Please read Documentation/workqueue.txt for details.
  */
 
 #include <linux/module.h>
 #include <linux/lockdep.h>
 #include <linux/idr.h>
 
+#include "workqueue_sched.h"
+
 enum {
        /* global_cwq flags */
+       GCWQ_MANAGE_WORKERS     = 1 << 0,       /* need to manage workers */
+       GCWQ_MANAGING_WORKERS   = 1 << 1,       /* managing workers */
+       GCWQ_DISASSOCIATED      = 1 << 2,       /* cpu can't serve workers */
        GCWQ_FREEZING           = 1 << 3,       /* freeze in progress */
+       GCWQ_HIGHPRI_PENDING    = 1 << 4,       /* highpri works on queue */
 
        /* worker flags */
        WORKER_STARTED          = 1 << 0,       /* started */
        WORKER_DIE              = 1 << 1,       /* die die die */
        WORKER_IDLE             = 1 << 2,       /* is idle */
+       WORKER_PREP             = 1 << 3,       /* preparing to run works */
        WORKER_ROGUE            = 1 << 4,       /* not bound to any cpu */
+       WORKER_REBIND           = 1 << 5,       /* mom is home, come back */
+       WORKER_CPU_INTENSIVE    = 1 << 6,       /* cpu intensive */
+       WORKER_UNBOUND          = 1 << 7,       /* worker is unbound */
+
+       WORKER_NOT_RUNNING      = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
+                                 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
 
        /* gcwq->trustee_state */
        TRUSTEE_START           = 0,            /* start */
@@ -56,24 +76,50 @@ enum {
        BUSY_WORKER_HASH_SIZE   = 1 << BUSY_WORKER_HASH_ORDER,
        BUSY_WORKER_HASH_MASK   = BUSY_WORKER_HASH_SIZE - 1,
 
+       MAX_IDLE_WORKERS_RATIO  = 4,            /* 1/4 of busy can be idle */
+       IDLE_WORKER_TIMEOUT     = 300 * HZ,     /* keep idle ones for 5 mins */
+
+       MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
+                                               /* call for help after 10ms
+                                                  (min two ticks) */
+       MAYDAY_INTERVAL         = HZ / 10,      /* and then every 100ms */
+       CREATE_COOLDOWN         = HZ,           /* time to breath after fail */
        TRUSTEE_COOLDOWN        = HZ / 10,      /* for trustee draining */
+
+       /*
+        * Rescue workers are used only on emergencies and shared by
+        * all cpus.  Give -20.
+        */
+       RESCUER_NICE_LEVEL      = -20,
 };
 
 /*
  * Structure fields follow one of the following exclusion rules.
  *
- * I: Set during initialization and read-only afterwards.
+ * I: Modifiable by initialization/destruction paths and read-only for
+ *    everyone else.
+ *
+ * P: Preemption protected.  Disabling preemption is enough and should
+ *    only be modified and accessed from the local cpu.
  *
  * L: gcwq->lock protected.  Access with gcwq->lock held.
  *
+ * X: During normal operation, modification requires gcwq->lock and
+ *    should be done only from local cpu.  Either disabling preemption
+ *    on local cpu or grabbing gcwq->lock is enough for read access.
+ *    If GCWQ_DISASSOCIATED is set, it's identical to L.
+ *
  * F: wq->flush_mutex protected.
  *
  * W: workqueue_lock protected.
  */
 
 struct global_cwq;
-struct cpu_workqueue_struct;
 
+/*
+ * The poor guys doing the actual heavy lifting.  All on-duty workers
+ * are either serving the manager role, on idle list or on busy hash.
+ */
 struct worker {
        /* on idle list while idle, on busy hash table while busy */
        union {
@@ -82,19 +128,25 @@ struct worker {
        };
 
        struct work_struct      *current_work;  /* L: work being processed */
+       struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
        struct list_head        scheduled;      /* L: scheduled works */
        struct task_struct      *task;          /* I: worker task */
        struct global_cwq       *gcwq;          /* I: the associated gcwq */
-       struct cpu_workqueue_struct *cwq;       /* I: the associated cwq */
-       unsigned int            flags;          /* L: flags */
+       /* 64 bytes boundary on 64bit, 32 on 32bit */
+       unsigned long           last_active;    /* L: last active timestamp */
+       unsigned int            flags;          /* X: flags */
        int                     id;             /* I: worker id */
+       struct work_struct      rebind_work;    /* L: rebind worker to cpu */
 };
 
 /*
- * Global per-cpu workqueue.
+ * Global per-cpu workqueue.  There's one and only one for each cpu
+ * and all works are queued and processed here regardless of their
+ * target workqueues.
  */
 struct global_cwq {
        spinlock_t              lock;           /* the gcwq lock */
+       struct list_head        worklist;       /* L: list of pending works */
        unsigned int            cpu;            /* I: the associated cpu */
        unsigned int            flags;          /* L: GCWQ_* flags */
 
@@ -102,27 +154,28 @@ struct global_cwq {
        int                     nr_idle;        /* L: currently idle ones */
 
        /* workers are chained either in the idle_list or busy_hash */
-       struct list_head        idle_list;      /* L: list of idle workers */
+       struct list_head        idle_list;      /* X: list of idle workers */
        struct hlist_head       busy_hash[BUSY_WORKER_HASH_SIZE];
                                                /* L: hash of busy workers */
 
+       struct timer_list       idle_timer;     /* L: worker idle timeout */
+       struct timer_list       mayday_timer;   /* L: SOS timer for dworkers */
+
        struct ida              worker_ida;     /* L: for worker IDs */
 
        struct task_struct      *trustee;       /* L: for gcwq shutdown */
        unsigned int            trustee_state;  /* L: trustee state */
        wait_queue_head_t       trustee_wait;   /* trustee wait */
+       struct worker           *first_idle;    /* L: first idle worker */
 } ____cacheline_aligned_in_smp;
 
 /*
- * The per-CPU workqueue (if single thread, we always use the first
- * possible cpu).  The lower WORK_STRUCT_FLAG_BITS of
+ * The per-CPU workqueue.  The lower WORK_STRUCT_FLAG_BITS of
  * work_struct->data are used for flags and thus cwqs need to be
  * aligned at two's power of the number of flag bits.
  */
 struct cpu_workqueue_struct {
        struct global_cwq       *gcwq;          /* I: the associated gcwq */
-       struct list_head worklist;
-       struct worker           *worker;
        struct workqueue_struct *wq;            /* I: the owning workqueue */
        int                     work_color;     /* L: current color */
        int                     flush_color;    /* L: flushing color */
@@ -143,12 +196,37 @@ struct wq_flusher {
 };
 
 /*
+ * All cpumasks are assumed to be always set on UP and thus can't be
+ * used to determine whether there's something to be done.
+ */
+#ifdef CONFIG_SMP
+typedef cpumask_var_t mayday_mask_t;
+#define mayday_test_and_set_cpu(cpu, mask)     \
+       cpumask_test_and_set_cpu((cpu), (mask))
+#define mayday_clear_cpu(cpu, mask)            cpumask_clear_cpu((cpu), (mask))
+#define for_each_mayday_cpu(cpu, mask)         for_each_cpu((cpu), (mask))
+#define alloc_mayday_mask(maskp, gfp)          zalloc_cpumask_var((maskp), (gfp))
+#define free_mayday_mask(mask)                 free_cpumask_var((mask))
+#else
+typedef unsigned long mayday_mask_t;
+#define mayday_test_and_set_cpu(cpu, mask)     test_and_set_bit(0, &(mask))
+#define mayday_clear_cpu(cpu, mask)            clear_bit(0, &(mask))
+#define for_each_mayday_cpu(cpu, mask)         if ((cpu) = 0, (mask))
+#define alloc_mayday_mask(maskp, gfp)          true
+#define free_mayday_mask(mask)                 do { } while (0)
+#endif
+
+/*
  * The externally visible workqueue abstraction is an array of
  * per-CPU workqueues:
  */
 struct workqueue_struct {
-       unsigned int            flags;          /* I: WQ_* flags */
-       struct cpu_workqueue_struct *cpu_wq;    /* I: cwq's */
+       unsigned int            flags;          /* W: WQ_* flags */
+       union {
+               struct cpu_workqueue_struct __percpu    *pcpu;
+               struct cpu_workqueue_struct             *single;
+               unsigned long                           v;
+       } cpu_wq;                               /* I: cwq's */
        struct list_head        list;           /* W: list of all workqueues */
 
        struct mutex            flush_mutex;    /* protects wq flushing */
@@ -159,21 +237,93 @@ struct workqueue_struct {
        struct list_head        flusher_queue;  /* F: flush waiters */
        struct list_head        flusher_overflow; /* F: flush overflow list */
 
-       int                     saved_max_active; /* I: saved cwq max_active */
+       mayday_mask_t           mayday_mask;    /* cpus requesting rescue */
+       struct worker           *rescuer;       /* I: rescue worker */
+
+       int                     nr_drainers;    /* W: drain in progress */
+       int                     saved_max_active; /* W: saved cwq max_active */
        const char              *name;          /* I: workqueue name */
 #ifdef CONFIG_LOCKDEP
        struct lockdep_map      lockdep_map;
 #endif
 };
 
+struct workqueue_struct *system_wq __read_mostly;
+struct workqueue_struct *system_long_wq __read_mostly;
+struct workqueue_struct *system_nrt_wq __read_mostly;
+struct workqueue_struct *system_unbound_wq __read_mostly;
+struct workqueue_struct *system_freezable_wq __read_mostly;
+EXPORT_SYMBOL_GPL(system_wq);
+EXPORT_SYMBOL_GPL(system_long_wq);
+EXPORT_SYMBOL_GPL(system_nrt_wq);
+EXPORT_SYMBOL_GPL(system_unbound_wq);
+EXPORT_SYMBOL_GPL(system_freezable_wq);
+
+#define CREATE_TRACE_POINTS
+#include <trace/events/workqueue.h>
+
 #define for_each_busy_worker(worker, i, pos, gcwq)                     \
        for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)                     \
                hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
 
+static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
+                                 unsigned int sw)
+{
+       if (cpu < nr_cpu_ids) {
+               if (sw & 1) {
+                       cpu = cpumask_next(cpu, mask);
+                       if (cpu < nr_cpu_ids)
+                               return cpu;
+               }
+               if (sw & 2)
+                       return WORK_CPU_UNBOUND;
+       }
+       return WORK_CPU_NONE;
+}
+
+static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
+                               struct workqueue_struct *wq)
+{
+       return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
+}
+
+/*
+ * CPU iterators
+ *
+ * An extra gcwq is defined for an invalid cpu number
+ * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
+ * specific CPU.  The following iterators are similar to
+ * for_each_*_cpu() iterators but also considers the unbound gcwq.
+ *
+ * for_each_gcwq_cpu()         : possible CPUs + WORK_CPU_UNBOUND
+ * for_each_online_gcwq_cpu()  : online CPUs + WORK_CPU_UNBOUND
+ * for_each_cwq_cpu()          : possible CPUs for bound workqueues,
+ *                               WORK_CPU_UNBOUND for unbound workqueues
+ */
+#define for_each_gcwq_cpu(cpu)                                         \
+       for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3);         \
+            (cpu) < WORK_CPU_NONE;                                     \
+            (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
+
+#define for_each_online_gcwq_cpu(cpu)                                  \
+       for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3);           \
+            (cpu) < WORK_CPU_NONE;                                     \
+            (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
+
+#define for_each_cwq_cpu(cpu, wq)                                      \
+       for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq));        \
+            (cpu) < WORK_CPU_NONE;                                     \
+            (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
+
 #ifdef CONFIG_DEBUG_OBJECTS_WORK
 
 static struct debug_obj_descr work_debug_descr;
 
+static void *work_debug_hint(void *addr)
+{
+       return ((struct work_struct *) addr)->func;
+}
+
 /*
  * fixup_init is called when:
  * - an active object is initialized
@@ -245,6 +395,7 @@ static int work_fixup_free(void *addr, enum debug_obj_state state)
 
 static struct debug_obj_descr work_debug_descr = {
        .name           = "work_struct",
+       .debug_hint     = work_debug_hint,
        .fixup_init     = work_fixup_init,
        .fixup_activate = work_fixup_activate,
        .fixup_free     = work_fixup_free,
@@ -285,29 +436,54 @@ static DEFINE_SPINLOCK(workqueue_lock);
 static LIST_HEAD(workqueues);
 static bool workqueue_freezing;                /* W: have wqs started freezing? */
 
+/*
+ * The almighty global cpu workqueues.  nr_running is the only field
+ * which is expected to be used frequently by other cpus via
+ * try_to_wake_up().  Put it in a separate cacheline.
+ */
 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
+static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
 
-static int worker_thread(void *__worker);
+/*
+ * Global cpu workqueue and nr_running counter for unbound gcwq.  The
+ * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
+ * workers have WORKER_UNBOUND set.
+ */
+static struct global_cwq unbound_global_cwq;
+static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0);      /* always 0 */
 
-static int singlethread_cpu __read_mostly;
+static int worker_thread(void *__worker);
 
 static struct global_cwq *get_gcwq(unsigned int cpu)
 {
-       return &per_cpu(global_cwq, cpu);
+       if (cpu != WORK_CPU_UNBOUND)
+               return &per_cpu(global_cwq, cpu);
+       else
+               return &unbound_global_cwq;
 }
 
-static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
-                                           struct workqueue_struct *wq)
+static atomic_t *get_gcwq_nr_running(unsigned int cpu)
 {
-       return per_cpu_ptr(wq->cpu_wq, cpu);
+       if (cpu != WORK_CPU_UNBOUND)
+               return &per_cpu(gcwq_nr_running, cpu);
+       else
+               return &unbound_gcwq_nr_running;
 }
 
-static struct cpu_workqueue_struct *target_cwq(unsigned int cpu,
-                                              struct workqueue_struct *wq)
+static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
+                                           struct workqueue_struct *wq)
 {
-       if (unlikely(wq->flags & WQ_SINGLE_THREAD))
-               cpu = singlethread_cpu;
-       return get_cwq(cpu, wq);
+       if (!(wq->flags & WQ_UNBOUND)) {
+               if (likely(cpu < nr_cpu_ids)) {
+#ifdef CONFIG_SMP
+                       return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
+#else
+                       return wq->cpu_wq.single;
+#endif
+               }
+       } else if (likely(cpu == WORK_CPU_UNBOUND))
+               return wq->cpu_wq.single;
+       return NULL;
 }
 
 static unsigned int work_color_to_flags(int color)
@@ -327,31 +503,290 @@ static int work_next_color(int color)
 }
 
 /*
- * Set the workqueue on which a work item is to be run
- * - Must *only* be called if the pending flag is set
+ * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
+ * work is on queue.  Once execution starts, WORK_STRUCT_CWQ is
+ * cleared and the work data contains the cpu number it was last on.
+ *
+ * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
+ * cwq, cpu or clear work->data.  These functions should only be
+ * called while the work is owned - ie. while the PENDING bit is set.
+ *
+ * get_work_[g]cwq() can be used to obtain the gcwq or cwq
+ * corresponding to a work.  gcwq is available once the work has been
+ * queued anywhere after initialization.  cwq is available only from
+ * queueing until execution starts.
  */
-static inline void set_wq_data(struct work_struct *work,
-                              struct cpu_workqueue_struct *cwq,
-                              unsigned long extra_flags)
+static inline void set_work_data(struct work_struct *work, unsigned long data,
+                                unsigned long flags)
 {
        BUG_ON(!work_pending(work));
+       atomic_long_set(&work->data, data | flags | work_static(work));
+}
 
-       atomic_long_set(&work->data, (unsigned long)cwq | work_static(work) |
-                       WORK_STRUCT_PENDING | extra_flags);
+static void set_work_cwq(struct work_struct *work,
+                        struct cpu_workqueue_struct *cwq,
+                        unsigned long extra_flags)
+{
+       set_work_data(work, (unsigned long)cwq,
+                     WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
+}
+
+static void set_work_cpu(struct work_struct *work, unsigned int cpu)
+{
+       set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
+}
+
+static void clear_work_data(struct work_struct *work)
+{
+       set_work_data(work, WORK_STRUCT_NO_CPU, 0);
+}
+
+static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
+{
+       unsigned long data = atomic_long_read(&work->data);
+
+       if (data & WORK_STRUCT_CWQ)
+               return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
+       else
+               return NULL;
+}
+
+static struct global_cwq *get_work_gcwq(struct work_struct *work)
+{
+       unsigned long data = atomic_long_read(&work->data);
+       unsigned int cpu;
+
+       if (data & WORK_STRUCT_CWQ)
+               return ((struct cpu_workqueue_struct *)
+                       (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
+
+       cpu = data >> WORK_STRUCT_FLAG_BITS;
+       if (cpu == WORK_CPU_NONE)
+               return NULL;
+
+       BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
+       return get_gcwq(cpu);
 }
 
 /*
- * Clear WORK_STRUCT_PENDING and the workqueue on which it was queued.
+ * Policy functions.  These define the policies on how the global
+ * worker pool is managed.  Unless noted otherwise, these functions
+ * assume that they're being called with gcwq->lock held.
  */
-static inline void clear_wq_data(struct work_struct *work)
+
+static bool __need_more_worker(struct global_cwq *gcwq)
 {
-       atomic_long_set(&work->data, work_static(work));
+       return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
+               gcwq->flags & GCWQ_HIGHPRI_PENDING;
 }
 
-static inline struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
+/*
+ * Need to wake up a worker?  Called from anything but currently
+ * running workers.
+ */
+static bool need_more_worker(struct global_cwq *gcwq)
+{
+       return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
+}
+
+/* Can I start working?  Called from busy but !running workers. */
+static bool may_start_working(struct global_cwq *gcwq)
+{
+       return gcwq->nr_idle;
+}
+
+/* Do I need to keep working?  Called from currently running workers. */
+static bool keep_working(struct global_cwq *gcwq)
+{
+       atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
+
+       return !list_empty(&gcwq->worklist) &&
+               (atomic_read(nr_running) <= 1 ||
+                gcwq->flags & GCWQ_HIGHPRI_PENDING);
+}
+
+/* Do we need a new worker?  Called from manager. */
+static bool need_to_create_worker(struct global_cwq *gcwq)
+{
+       return need_more_worker(gcwq) && !may_start_working(gcwq);
+}
+
+/* Do I need to be the manager? */
+static bool need_to_manage_workers(struct global_cwq *gcwq)
 {
-       return (void *)(atomic_long_read(&work->data) &
-                       WORK_STRUCT_WQ_DATA_MASK);
+       return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
+}
+
+/* Do we have too many workers and should some go away? */
+static bool too_many_workers(struct global_cwq *gcwq)
+{
+       bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
+       int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
+       int nr_busy = gcwq->nr_workers - nr_idle;
+
+       return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
+}
+
+/*
+ * Wake up functions.
+ */
+
+/* Return the first worker.  Safe with preemption disabled */
+static struct worker *first_worker(struct global_cwq *gcwq)
+{
+       if (unlikely(list_empty(&gcwq->idle_list)))
+               return NULL;
+
+       return list_first_entry(&gcwq->idle_list, struct worker, entry);
+}
+
+/**
+ * wake_up_worker - wake up an idle worker
+ * @gcwq: gcwq to wake worker for
+ *
+ * Wake up the first idle worker of @gcwq.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock).
+ */
+static void wake_up_worker(struct global_cwq *gcwq)
+{
+       struct worker *worker = first_worker(gcwq);
+
+       if (likely(worker))
+               wake_up_process(worker->task);
+}
+
+/**
+ * wq_worker_waking_up - a worker is waking up
+ * @task: task waking up
+ * @cpu: CPU @task is waking up to
+ *
+ * This function is called during try_to_wake_up() when a worker is
+ * being awoken.
+ *
+ * CONTEXT:
+ * spin_lock_irq(rq->lock)
+ */
+void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
+{
+       struct worker *worker = kthread_data(task);
+
+       if (!(worker->flags & WORKER_NOT_RUNNING))
+               atomic_inc(get_gcwq_nr_running(cpu));
+}
+
+/**
+ * wq_worker_sleeping - a worker is going to sleep
+ * @task: task going to sleep
+ * @cpu: CPU in question, must be the current CPU number
+ *
+ * This function is called during schedule() when a busy worker is
+ * going to sleep.  Worker on the same cpu can be woken up by
+ * returning pointer to its task.
+ *
+ * CONTEXT:
+ * spin_lock_irq(rq->lock)
+ *
+ * RETURNS:
+ * Worker task on @cpu to wake up, %NULL if none.
+ */
+struct task_struct *wq_worker_sleeping(struct task_struct *task,
+                                      unsigned int cpu)
+{
+       struct worker *worker = kthread_data(task), *to_wakeup = NULL;
+       struct global_cwq *gcwq = get_gcwq(cpu);
+       atomic_t *nr_running = get_gcwq_nr_running(cpu);
+
+       if (worker->flags & WORKER_NOT_RUNNING)
+               return NULL;
+
+       /* this can only happen on the local cpu */
+       BUG_ON(cpu != raw_smp_processor_id());
+
+       /*
+        * The counterpart of the following dec_and_test, implied mb,
+        * worklist not empty test sequence is in insert_work().
+        * Please read comment there.
+        *
+        * NOT_RUNNING is clear.  This means that trustee is not in
+        * charge and we're running on the local cpu w/ rq lock held
+        * and preemption disabled, which in turn means that none else
+        * could be manipulating idle_list, so dereferencing idle_list
+        * without gcwq lock is safe.
+        */
+       if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
+               to_wakeup = first_worker(gcwq);
+       return to_wakeup ? to_wakeup->task : NULL;
+}
+
+/**
+ * worker_set_flags - set worker flags and adjust nr_running accordingly
+ * @worker: self
+ * @flags: flags to set
+ * @wakeup: wakeup an idle worker if necessary
+ *
+ * Set @flags in @worker->flags and adjust nr_running accordingly.  If
+ * nr_running becomes zero and @wakeup is %true, an idle worker is
+ * woken up.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock)
+ */
+static inline void worker_set_flags(struct worker *worker, unsigned int flags,
+                                   bool wakeup)
+{
+       struct global_cwq *gcwq = worker->gcwq;
+
+       WARN_ON_ONCE(worker->task != current);
+
+       /*
+        * If transitioning into NOT_RUNNING, adjust nr_running and
+        * wake up an idle worker as necessary if requested by
+        * @wakeup.
+        */
+       if ((flags & WORKER_NOT_RUNNING) &&
+           !(worker->flags & WORKER_NOT_RUNNING)) {
+               atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
+
+               if (wakeup) {
+                       if (atomic_dec_and_test(nr_running) &&
+                           !list_empty(&gcwq->worklist))
+                               wake_up_worker(gcwq);
+               } else
+                       atomic_dec(nr_running);
+       }
+
+       worker->flags |= flags;
+}
+
+/**
+ * worker_clr_flags - clear worker flags and adjust nr_running accordingly
+ * @worker: self
+ * @flags: flags to clear
+ *
+ * Clear @flags in @worker->flags and adjust nr_running accordingly.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock)
+ */
+static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
+{
+       struct global_cwq *gcwq = worker->gcwq;
+       unsigned int oflags = worker->flags;
+
+       WARN_ON_ONCE(worker->task != current);
+
+       worker->flags &= ~flags;
+
+       /*
+        * If transitioning out of NOT_RUNNING, increment nr_running.  Note
+        * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
+        * of multiple flags, not a single flag.
+        */
+       if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
+               if (!(worker->flags & WORKER_NOT_RUNNING))
+                       atomic_inc(get_gcwq_nr_running(gcwq->cpu));
 }
 
 /**
@@ -382,13 +817,104 @@ static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
 }
 
 /**
- * insert_work - insert a work into cwq
+ * __find_worker_executing_work - find worker which is executing a work
+ * @gcwq: gcwq of interest
+ * @bwh: hash head as returned by busy_worker_head()
+ * @work: work to find worker for
+ *
+ * Find a worker which is executing @work on @gcwq.  @bwh should be
+ * the hash head obtained by calling busy_worker_head() with the same
+ * work.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock).
+ *
+ * RETURNS:
+ * Pointer to worker which is executing @work if found, NULL
+ * otherwise.
+ */
+static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
+                                                  struct hlist_head *bwh,
+                                                  struct work_struct *work)
+{
+       struct worker *worker;
+       struct hlist_node *tmp;
+
+       hlist_for_each_entry(worker, tmp, bwh, hentry)
+               if (worker->current_work == work)
+                       return worker;
+       return NULL;
+}
+
+/**
+ * find_worker_executing_work - find worker which is executing a work
+ * @gcwq: gcwq of interest
+ * @work: work to find worker for
+ *
+ * Find a worker which is executing @work on @gcwq.  This function is
+ * identical to __find_worker_executing_work() except that this
+ * function calculates @bwh itself.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock).
+ *
+ * RETURNS:
+ * Pointer to worker which is executing @work if found, NULL
+ * otherwise.
+ */
+static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
+                                                struct work_struct *work)
+{
+       return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
+                                           work);
+}
+
+/**
+ * gcwq_determine_ins_pos - find insertion position
+ * @gcwq: gcwq of interest
+ * @cwq: cwq a work is being queued for
+ *
+ * A work for @cwq is about to be queued on @gcwq, determine insertion
+ * position for the work.  If @cwq is for HIGHPRI wq, the work is
+ * queued at the head of the queue but in FIFO order with respect to
+ * other HIGHPRI works; otherwise, at the end of the queue.  This
+ * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
+ * there are HIGHPRI works pending.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock).
+ *
+ * RETURNS:
+ * Pointer to inserstion position.
+ */
+static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
+                                              struct cpu_workqueue_struct *cwq)
+{
+       struct work_struct *twork;
+
+       if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
+               return &gcwq->worklist;
+
+       list_for_each_entry(twork, &gcwq->worklist, entry) {
+               struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
+
+               if (!(tcwq->wq->flags & WQ_HIGHPRI))
+                       break;
+       }
+
+       gcwq->flags |= GCWQ_HIGHPRI_PENDING;
+       return &twork->entry;
+}
+
+/**
+ * insert_work - insert a work into gcwq
  * @cwq: cwq @work belongs to
  * @work: work to insert
  * @head: insertion point
  * @extra_flags: extra WORK_STRUCT_* flags to set
  *
- * Insert @work into @cwq after @head.
+ * Insert @work which belongs to @cwq into @gcwq after @head.
+ * @extra_flags is or'd to work_struct flags.
  *
  * CONTEXT:
  * spin_lock_irq(gcwq->lock).
@@ -397,8 +923,10 @@ static void insert_work(struct cpu_workqueue_struct *cwq,
                        struct work_struct *work, struct list_head *head,
                        unsigned int extra_flags)
 {
+       struct global_cwq *gcwq = cwq->gcwq;
+
        /* we own @work, set data and link */
-       set_wq_data(work, cwq, extra_flags);
+       set_work_cwq(work, cwq, extra_flags);
 
        /*
         * Ensure that we get the right work->data if we see the
@@ -407,31 +935,121 @@ static void insert_work(struct cpu_workqueue_struct *cwq,
        smp_wmb();
 
        list_add_tail(&work->entry, head);
-       wake_up_process(cwq->worker->task);
+
+       /*
+        * Ensure either worker_sched_deactivated() sees the above
+        * list_add_tail() or we see zero nr_running to avoid workers
+        * lying around lazily while there are works to be processed.
+        */
+       smp_mb();
+
+       if (__need_more_worker(gcwq))
+               wake_up_worker(gcwq);
+}
+
+/*
+ * Test whether @work is being queued from another work executing on the
+ * same workqueue.  This is rather expensive and should only be used from
+ * cold paths.
+ */
+static bool is_chained_work(struct workqueue_struct *wq)
+{
+       unsigned long flags;
+       unsigned int cpu;
+
+       for_each_gcwq_cpu(cpu) {
+               struct global_cwq *gcwq = get_gcwq(cpu);
+               struct worker *worker;
+               struct hlist_node *pos;
+               int i;
+
+               spin_lock_irqsave(&gcwq->lock, flags);
+               for_each_busy_worker(worker, i, pos, gcwq) {
+                       if (worker->task != current)
+                               continue;
+                       spin_unlock_irqrestore(&gcwq->lock, flags);
+                       /*
+                        * I'm @worker, no locking necessary.  See if @work
+                        * is headed to the same workqueue.
+                        */
+                       return worker->current_cwq->wq == wq;
+               }
+               spin_unlock_irqrestore(&gcwq->lock, flags);
+       }
+       return false;
 }
 
 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
                         struct work_struct *work)
 {
-       struct cpu_workqueue_struct *cwq = target_cwq(cpu, wq);
-       struct global_cwq *gcwq = cwq->gcwq;
+       struct global_cwq *gcwq;
+       struct cpu_workqueue_struct *cwq;
        struct list_head *worklist;
+       unsigned int work_flags;
        unsigned long flags;
 
        debug_work_activate(work);
 
-       spin_lock_irqsave(&gcwq->lock, flags);
+       /* if dying, only works from the same workqueue are allowed */
+       if (unlikely(wq->flags & WQ_DRAINING) &&
+           WARN_ON_ONCE(!is_chained_work(wq)))
+               return;
+
+       /* determine gcwq to use */
+       if (!(wq->flags & WQ_UNBOUND)) {
+               struct global_cwq *last_gcwq;
+
+               if (unlikely(cpu == WORK_CPU_UNBOUND))
+                       cpu = raw_smp_processor_id();
+
+               /*
+                * It's multi cpu.  If @wq is non-reentrant and @work
+                * was previously on a different cpu, it might still
+                * be running there, in which case the work needs to
+                * be queued on that cpu to guarantee non-reentrance.
+                */
+               gcwq = get_gcwq(cpu);
+               if (wq->flags & WQ_NON_REENTRANT &&
+                   (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
+                       struct worker *worker;
+
+                       spin_lock_irqsave(&last_gcwq->lock, flags);
+
+                       worker = find_worker_executing_work(last_gcwq, work);
+
+                       if (worker && worker->current_cwq->wq == wq)
+                               gcwq = last_gcwq;
+                       else {
+                               /* meh... not running there, queue here */
+                               spin_unlock_irqrestore(&last_gcwq->lock, flags);
+                               spin_lock_irqsave(&gcwq->lock, flags);
+                       }
+               } else
+                       spin_lock_irqsave(&gcwq->lock, flags);
+       } else {
+               gcwq = get_gcwq(WORK_CPU_UNBOUND);
+               spin_lock_irqsave(&gcwq->lock, flags);
+       }
+
+       /* gcwq determined, get cwq and queue */
+       cwq = get_cwq(gcwq->cpu, wq);
+       trace_workqueue_queue_work(cpu, cwq, work);
+
        BUG_ON(!list_empty(&work->entry));
 
        cwq->nr_in_flight[cwq->work_color]++;
+       work_flags = work_color_to_flags(cwq->work_color);
 
        if (likely(cwq->nr_active < cwq->max_active)) {
+               trace_workqueue_activate_work(work);
                cwq->nr_active++;
-               worklist = &cwq->worklist;
-       } else
+               worklist = gcwq_determine_ins_pos(gcwq, cwq);
+       } else {
+               work_flags |= WORK_STRUCT_DELAYED;
                worklist = &cwq->delayed_works;
+       }
 
-       insert_work(cwq, work, worklist, work_color_to_flags(cwq->work_color));
+       insert_work(cwq, work, worklist, work_flags);
 
        spin_unlock_irqrestore(&gcwq->lock, flags);
 }
@@ -484,7 +1102,7 @@ EXPORT_SYMBOL_GPL(queue_work_on);
 static void delayed_work_timer_fn(unsigned long __data)
 {
        struct delayed_work *dwork = (struct delayed_work *)__data;
-       struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
+       struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
 
        __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
 }
@@ -524,13 +1142,30 @@ int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
        struct work_struct *work = &dwork->work;
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
+               unsigned int lcpu;
+
                BUG_ON(timer_pending(timer));
                BUG_ON(!list_empty(&work->entry));
 
                timer_stats_timer_set_start_info(&dwork->timer);
 
-               /* This stores cwq for the moment, for the timer_fn */
-               set_wq_data(work, target_cwq(raw_smp_processor_id(), wq), 0);
+               /*
+                * This stores cwq for the moment, for the timer_fn.
+                * Note that the work's gcwq is preserved to allow
+                * reentrance detection for delayed works.
+                */
+               if (!(wq->flags & WQ_UNBOUND)) {
+                       struct global_cwq *gcwq = get_work_gcwq(work);
+
+                       if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
+                               lcpu = gcwq->cpu;
+                       else
+                               lcpu = raw_smp_processor_id();
+               } else
+                       lcpu = WORK_CPU_UNBOUND;
+
+               set_work_cwq(work, get_cwq(lcpu, wq), 0);
+
                timer->expires = jiffies + delay;
                timer->data = (unsigned long)dwork;
                timer->function = delayed_work_timer_fn;
@@ -563,14 +1198,24 @@ static void worker_enter_idle(struct worker *worker)
        BUG_ON(!list_empty(&worker->entry) &&
               (worker->hentry.next || worker->hentry.pprev));
 
+       /* can't use worker_set_flags(), also called from start_worker() */
        worker->flags |= WORKER_IDLE;
        gcwq->nr_idle++;
+       worker->last_active = jiffies;
 
        /* idle_list is LIFO */
        list_add(&worker->entry, &gcwq->idle_list);
 
-       if (unlikely(worker->flags & WORKER_ROGUE))
+       if (likely(!(worker->flags & WORKER_ROGUE))) {
+               if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
+                       mod_timer(&gcwq->idle_timer,
+                                 jiffies + IDLE_WORKER_TIMEOUT);
+       } else
                wake_up_all(&gcwq->trustee_wait);
+
+       /* sanity check nr_running */
+       WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
+                    atomic_read(get_gcwq_nr_running(gcwq->cpu)));
 }
 
 /**
@@ -587,11 +1232,94 @@ static void worker_leave_idle(struct worker *worker)
        struct global_cwq *gcwq = worker->gcwq;
 
        BUG_ON(!(worker->flags & WORKER_IDLE));
-       worker->flags &= ~WORKER_IDLE;
+       worker_clr_flags(worker, WORKER_IDLE);
        gcwq->nr_idle--;
        list_del_init(&worker->entry);
 }
 
+/**
+ * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
+ * @worker: self
+ *
+ * Works which are scheduled while the cpu is online must at least be
+ * scheduled to a worker which is bound to the cpu so that if they are
+ * flushed from cpu callbacks while cpu is going down, they are
+ * guaranteed to execute on the cpu.
+ *
+ * This function is to be used by rogue workers and rescuers to bind
+ * themselves to the target cpu and may race with cpu going down or
+ * coming online.  kthread_bind() can't be used because it may put the
+ * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
+ * verbatim as it's best effort and blocking and gcwq may be
+ * [dis]associated in the meantime.
+ *
+ * This function tries set_cpus_allowed() and locks gcwq and verifies
+ * the binding against GCWQ_DISASSOCIATED which is set during
+ * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
+ * idle state or fetches works without dropping lock, it can guarantee
+ * the scheduling requirement described in the first paragraph.
+ *
+ * CONTEXT:
+ * Might sleep.  Called without any lock but returns with gcwq->lock
+ * held.
+ *
+ * RETURNS:
+ * %true if the associated gcwq is online (@worker is successfully
+ * bound), %false if offline.
+ */
+static bool worker_maybe_bind_and_lock(struct worker *worker)
+__acquires(&gcwq->lock)
+{
+       struct global_cwq *gcwq = worker->gcwq;
+       struct task_struct *task = worker->task;
+
+       while (true) {
+               /*
+                * The following call may fail, succeed or succeed
+                * without actually migrating the task to the cpu if
+                * it races with cpu hotunplug operation.  Verify
+                * against GCWQ_DISASSOCIATED.
+                */
+               if (!(gcwq->flags & GCWQ_DISASSOCIATED))
+                       set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
+
+               spin_lock_irq(&gcwq->lock);
+               if (gcwq->flags & GCWQ_DISASSOCIATED)
+                       return false;
+               if (task_cpu(task) == gcwq->cpu &&
+                   cpumask_equal(&current->cpus_allowed,
+                                 get_cpu_mask(gcwq->cpu)))
+                       return true;
+               spin_unlock_irq(&gcwq->lock);
+
+               /*
+                * We've raced with CPU hot[un]plug.  Give it a breather
+                * and retry migration.  cond_resched() is required here;
+                * otherwise, we might deadlock against cpu_stop trying to
+                * bring down the CPU on non-preemptive kernel.
+                */
+               cpu_relax();
+               cond_resched();
+       }
+}
+
+/*
+ * Function for worker->rebind_work used to rebind rogue busy workers
+ * to the associated cpu which is coming back online.  This is
+ * scheduled by cpu up but can race with other cpu hotplug operations
+ * and may be executed twice without intervening cpu down.
+ */
+static void worker_rebind_fn(struct work_struct *work)
+{
+       struct worker *worker = container_of(work, struct worker, rebind_work);
+       struct global_cwq *gcwq = worker->gcwq;
+
+       if (worker_maybe_bind_and_lock(worker))
+               worker_clr_flags(worker, WORKER_REBIND);
+
+       spin_unlock_irq(&gcwq->lock);
+}
+
 static struct worker *alloc_worker(void)
 {
        struct worker *worker;
@@ -600,16 +1328,19 @@ static struct worker *alloc_worker(void)
        if (worker) {
                INIT_LIST_HEAD(&worker->entry);
                INIT_LIST_HEAD(&worker->scheduled);
+               INIT_WORK(&worker->rebind_work, worker_rebind_fn);
+               /* on creation a worker is in !idle && prep state */
+               worker->flags = WORKER_PREP;
        }
        return worker;
 }
 
 /**
  * create_worker - create a new workqueue worker
- * @cwq: cwq the new worker will belong to
+ * @gcwq: gcwq the new worker will belong to
  * @bind: whether to set affinity to @cpu or not
  *
- * Create a new worker which is bound to @cwq.  The returned worker
+ * Create a new worker which is bound to @gcwq.  The returned worker
  * can be started by calling start_worker() or destroyed using
  * destroy_worker().
  *
@@ -619,11 +1350,11 @@ static struct worker *alloc_worker(void)
  * RETURNS:
  * Pointer to the newly created worker.
  */
-static struct worker *create_worker(struct cpu_workqueue_struct *cwq, bool bind)
+static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
 {
-       struct global_cwq *gcwq = cwq->gcwq;
-       int id = -1;
+       bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
        struct worker *worker = NULL;
+       int id = -1;
 
        spin_lock_irq(&gcwq->lock);
        while (ida_get_new(&gcwq->worker_ida, &id)) {
@@ -639,11 +1370,16 @@ static struct worker *create_worker(struct cpu_workqueue_struct *cwq, bool bind)
                goto fail;
 
        worker->gcwq = gcwq;
-       worker->cwq = cwq;
        worker->id = id;
 
-       worker->task = kthread_create(worker_thread, worker, "kworker/%u:%d",
-                                     gcwq->cpu, id);
+       if (!on_unbound_cpu)
+               worker->task = kthread_create_on_node(worker_thread,
+                                                     worker,
+                                                     cpu_to_node(gcwq->cpu),
+                                                     "kworker/%u:%d", gcwq->cpu, id);
+       else
+               worker->task = kthread_create(worker_thread, worker,
+                                             "kworker/u:%d", id);
        if (IS_ERR(worker->task))
                goto fail;
 
@@ -652,10 +1388,13 @@ static struct worker *create_worker(struct cpu_workqueue_struct *cwq, bool bind)
         * online later on.  Make sure every worker has
         * PF_THREAD_BOUND set.
         */
-       if (bind)
+       if (bind && !on_unbound_cpu)
                kthread_bind(worker->task, gcwq->cpu);
-       else
+       else {
                worker->task->flags |= PF_THREAD_BOUND;
+               if (on_unbound_cpu)
+                       worker->flags |= WORKER_UNBOUND;
+       }
 
        return worker;
 fail:
@@ -720,6 +1459,225 @@ static void destroy_worker(struct worker *worker)
        ida_remove(&gcwq->worker_ida, id);
 }
 
+static void idle_worker_timeout(unsigned long __gcwq)
+{
+       struct global_cwq *gcwq = (void *)__gcwq;
+
+       spin_lock_irq(&gcwq->lock);
+
+       if (too_many_workers(gcwq)) {
+               struct worker *worker;
+               unsigned long expires;
+
+               /* idle_list is kept in LIFO order, check the last one */
+               worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
+               expires = worker->last_active + IDLE_WORKER_TIMEOUT;
+
+               if (time_before(jiffies, expires))
+                       mod_timer(&gcwq->idle_timer, expires);
+               else {
+                       /* it's been idle for too long, wake up manager */
+                       gcwq->flags |= GCWQ_MANAGE_WORKERS;
+                       wake_up_worker(gcwq);
+               }
+       }
+
+       spin_unlock_irq(&gcwq->lock);
+}
+
+static bool send_mayday(struct work_struct *work)
+{
+       struct cpu_workqueue_struct *cwq = get_work_cwq(work);
+       struct workqueue_struct *wq = cwq->wq;
+       unsigned int cpu;
+
+       if (!(wq->flags & WQ_RESCUER))
+               return false;
+
+       /* mayday mayday mayday */
+       cpu = cwq->gcwq->cpu;
+       /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
+       if (cpu == WORK_CPU_UNBOUND)
+               cpu = 0;
+       if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
+               wake_up_process(wq->rescuer->task);
+       return true;
+}
+
+static void gcwq_mayday_timeout(unsigned long __gcwq)
+{
+       struct global_cwq *gcwq = (void *)__gcwq;
+       struct work_struct *work;
+
+       spin_lock_irq(&gcwq->lock);
+
+       if (need_to_create_worker(gcwq)) {
+               /*
+                * We've been trying to create a new worker but
+                * haven't been successful.  We might be hitting an
+                * allocation deadlock.  Send distress signals to
+                * rescuers.
+                */
+               list_for_each_entry(work, &gcwq->worklist, entry)
+                       send_mayday(work);
+       }
+
+       spin_unlock_irq(&gcwq->lock);
+
+       mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
+}
+
+/**
+ * maybe_create_worker - create a new worker if necessary
+ * @gcwq: gcwq to create a new worker for
+ *
+ * Create a new worker for @gcwq if necessary.  @gcwq is guaranteed to
+ * have at least one idle worker on return from this function.  If
+ * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
+ * sent to all rescuers with works scheduled on @gcwq to resolve
+ * possible allocation deadlock.
+ *
+ * On return, need_to_create_worker() is guaranteed to be false and
+ * may_start_working() true.
+ *
+ * LOCKING:
+ * spin_lock_irq(gcwq->lock) which may be released and regrabbed
+ * multiple times.  Does GFP_KERNEL allocations.  Called only from
+ * manager.
+ *
+ * RETURNS:
+ * false if no action was taken and gcwq->lock stayed locked, true
+ * otherwise.
+ */
+static bool maybe_create_worker(struct global_cwq *gcwq)
+__releases(&gcwq->lock)
+__acquires(&gcwq->lock)
+{
+       if (!need_to_create_worker(gcwq))
+               return false;
+restart:
+       spin_unlock_irq(&gcwq->lock);
+
+       /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
+       mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
+
+       while (true) {
+               struct worker *worker;
+
+               worker = create_worker(gcwq, true);
+               if (worker) {
+                       del_timer_sync(&gcwq->mayday_timer);
+                       spin_lock_irq(&gcwq->lock);
+                       start_worker(worker);
+                       BUG_ON(need_to_create_worker(gcwq));
+                       return true;
+               }
+
+               if (!need_to_create_worker(gcwq))
+                       break;
+
+               __set_current_state(TASK_INTERRUPTIBLE);
+               schedule_timeout(CREATE_COOLDOWN);
+
+               if (!need_to_create_worker(gcwq))
+                       break;
+       }
+
+       del_timer_sync(&gcwq->mayday_timer);
+       spin_lock_irq(&gcwq->lock);
+       if (need_to_create_worker(gcwq))
+               goto restart;
+       return true;
+}
+
+/**
+ * maybe_destroy_worker - destroy workers which have been idle for a while
+ * @gcwq: gcwq to destroy workers for
+ *
+ * Destroy @gcwq workers which have been idle for longer than
+ * IDLE_WORKER_TIMEOUT.
+ *
+ * LOCKING:
+ * spin_lock_irq(gcwq->lock) which may be released and regrabbed
+ * multiple times.  Called only from manager.
+ *
+ * RETURNS:
+ * false if no action was taken and gcwq->lock stayed locked, true
+ * otherwise.
+ */
+static bool maybe_destroy_workers(struct global_cwq *gcwq)
+{
+       bool ret = false;
+
+       while (too_many_workers(gcwq)) {
+               struct worker *worker;
+               unsigned long expires;
+
+               worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
+               expires = worker->last_active + IDLE_WORKER_TIMEOUT;
+
+               if (time_before(jiffies, expires)) {
+                       mod_timer(&gcwq->idle_timer, expires);
+                       break;
+               }
+
+               destroy_worker(worker);
+               ret = true;
+       }
+
+       return ret;
+}
+
+/**
+ * manage_workers - manage worker pool
+ * @worker: self
+ *
+ * Assume the manager role and manage gcwq worker pool @worker belongs
+ * to.  At any given time, there can be only zero or one manager per
+ * gcwq.  The exclusion is handled automatically by this function.
+ *
+ * The caller can safely start processing works on false return.  On
+ * true return, it's guaranteed that need_to_create_worker() is false
+ * and may_start_working() is true.
+ *
+ * CONTEXT:
+ * spin_lock_irq(gcwq->lock) which may be released and regrabbed
+ * multiple times.  Does GFP_KERNEL allocations.
+ *
+ * RETURNS:
+ * false if no action was taken and gcwq->lock stayed locked, true if
+ * some action was taken.
+ */
+static bool manage_workers(struct worker *worker)
+{
+       struct global_cwq *gcwq = worker->gcwq;
+       bool ret = false;
+
+       if (gcwq->flags & GCWQ_MANAGING_WORKERS)
+               return ret;
+
+       gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
+       gcwq->flags |= GCWQ_MANAGING_WORKERS;
+
+       /*
+        * Destroy and then create so that may_start_working() is true
+        * on return.
+        */
+       ret |= maybe_destroy_workers(gcwq);
+       ret |= maybe_create_worker(gcwq);
+
+       gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
+
+       /*
+        * The trustee might be waiting to take over the manager
+        * position, tell it we're done.
+        */
+       if (unlikely(gcwq->trustee))
+               wake_up_all(&gcwq->trustee_wait);
+
+       return ret;
+}
+
 /**
  * move_linked_works - move linked works to a list
  * @work: start of series of works to be scheduled
@@ -765,8 +1723,11 @@ static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
 {
        struct work_struct *work = list_first_entry(&cwq->delayed_works,
                                                    struct work_struct, entry);
+       struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
 
-       move_linked_works(work, &cwq->worklist, NULL);
+       trace_workqueue_activate_work(work);
+       move_linked_works(work, pos, NULL);
+       __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
        cwq->nr_active++;
 }
 
@@ -774,6 +1735,7 @@ static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
  * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
  * @cwq: cwq of interest
  * @color: color of work which left the queue
+ * @delayed: for a delayed work
  *
  * A work either has completed or is removed from pending queue,
  * decrement nr_in_flight of its cwq and handle workqueue flushing.
@@ -781,19 +1743,23 @@ static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
  * CONTEXT:
  * spin_lock_irq(gcwq->lock).
  */
-static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
+static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
+                                bool delayed)
 {
        /* ignore uncolored works */
        if (color == WORK_NO_COLOR)
                return;
 
        cwq->nr_in_flight[color]--;
-       cwq->nr_active--;
 
-       /* one down, submit a delayed one */
-       if (!list_empty(&cwq->delayed_works) &&
-           cwq->nr_active < cwq->max_active)
-               cwq_activate_first_delayed(cwq);
+       if (!delayed) {
+               cwq->nr_active--;
+               if (!list_empty(&cwq->delayed_works)) {
+                       /* one down, submit a delayed one */
+                       if (cwq->nr_active < cwq->max_active)
+                               cwq_activate_first_delayed(cwq);
+               }
+       }
 
        /* is flush in progress and are we at the flushing tip? */
        if (likely(cwq->flush_color != color))
@@ -829,12 +1795,16 @@ static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
  * spin_lock_irq(gcwq->lock) which is released and regrabbed.
  */
 static void process_one_work(struct worker *worker, struct work_struct *work)
+__releases(&gcwq->lock)
+__acquires(&gcwq->lock)
 {
-       struct cpu_workqueue_struct *cwq = worker->cwq;
+       struct cpu_workqueue_struct *cwq = get_work_cwq(work);
        struct global_cwq *gcwq = cwq->gcwq;
        struct hlist_head *bwh = busy_worker_head(gcwq, work);
+       bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
        work_func_t f = work->func;
        int work_color;
+       struct worker *collision;
 #ifdef CONFIG_LOCKDEP
        /*
         * It is permissible to free the struct work_struct from
@@ -845,20 +1815,63 @@ static void process_one_work(struct worker *worker, struct work_struct *work)
         */
        struct lockdep_map lockdep_map = work->lockdep_map;
 #endif
+       /*
+        * A single work shouldn't be executed concurrently by
+        * multiple workers on a single cpu.  Check whether anyone is
+        * already processing the work.  If so, defer the work to the
+        * currently executing one.
+        */
+       collision = __find_worker_executing_work(gcwq, bwh, work);
+       if (unlikely(collision)) {
+               move_linked_works(work, &collision->scheduled, NULL);
+               return;
+       }
+
        /* claim and process */
        debug_work_deactivate(work);
        hlist_add_head(&worker->hentry, bwh);
        worker->current_work = work;
+       worker->current_cwq = cwq;
        work_color = get_work_color(work);
+
+       /* record the current cpu number in the work data and dequeue */
+       set_work_cpu(work, gcwq->cpu);
        list_del_init(&work->entry);
 
+       /*
+        * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
+        * wake up another worker; otherwise, clear HIGHPRI_PENDING.
+        */
+       if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
+               struct work_struct *nwork = list_first_entry(&gcwq->worklist,
+                                               struct work_struct, entry);
+
+               if (!list_empty(&gcwq->worklist) &&
+                   get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
+                       wake_up_worker(gcwq);
+               else
+                       gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
+       }
+
+       /*
+        * CPU intensive works don't participate in concurrency
+        * management.  They're the scheduler's responsibility.
+        */
+       if (unlikely(cpu_intensive))
+               worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
+
        spin_unlock_irq(&gcwq->lock);
 
-       BUG_ON(get_wq_data(work) != cwq);
        work_clear_pending(work);
-       lock_map_acquire(&cwq->wq->lockdep_map);
+       lock_map_acquire_read(&cwq->wq->lockdep_map);
        lock_map_acquire(&lockdep_map);
+       trace_workqueue_execute_start(work);
        f(work);
+       /*
+        * While we must be careful to not use "work" after this, the trace
+        * point will only record its address.
+        */
+       trace_workqueue_execute_end(work);
        lock_map_release(&lockdep_map);
        lock_map_release(&cwq->wq->lockdep_map);
 
@@ -874,10 +1887,15 @@ static void process_one_work(struct worker *worker, struct work_struct *work)
 
        spin_lock_irq(&gcwq->lock);
 
+       /* clear cpu intensive status */
+       if (unlikely(cpu_intensive))
+               worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
+
        /* we're done with it, release */
        hlist_del_init(&worker->hentry);
        worker->current_work = NULL;
-       cwq_dec_nr_in_flight(cwq, work_color);
+       worker->current_cwq = NULL;
+       cwq_dec_nr_in_flight(cwq, work_color, false);
 }
 
 /**
@@ -905,25 +1923,39 @@ static void process_scheduled_works(struct worker *worker)
  * worker_thread - the worker thread function
  * @__worker: self
  *
- * The cwq worker thread function.
+ * The gcwq worker thread function.  There's a single dynamic pool of
+ * these per each cpu.  These workers process all works regardless of
+ * their specific target workqueue.  The only exception is works which
+ * belong to workqueues with a rescuer which will be explained in
+ * rescuer_thread().
  */
 static int worker_thread(void *__worker)
 {
        struct worker *worker = __worker;
        struct global_cwq *gcwq = worker->gcwq;
-       struct cpu_workqueue_struct *cwq = worker->cwq;
 
+       /* tell the scheduler that this is a workqueue worker */
+       worker->task->flags |= PF_WQ_WORKER;
 woke_up:
        spin_lock_irq(&gcwq->lock);
 
        /* DIE can be set only while we're idle, checking here is enough */
        if (worker->flags & WORKER_DIE) {
                spin_unlock_irq(&gcwq->lock);
+               worker->task->flags &= ~PF_WQ_WORKER;
                return 0;
        }
 
        worker_leave_idle(worker);
 recheck:
+       /* no more worker necessary? */
+       if (!need_more_worker(gcwq))
+               goto sleep;
+
+       /* do we need to manage? */
+       if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
+               goto recheck;
+
        /*
         * ->scheduled list can only be filled while a worker is
         * preparing to process a work or actually processing it.
@@ -931,27 +1963,18 @@ recheck:
         */
        BUG_ON(!list_empty(&worker->scheduled));
 
-       while (!list_empty(&cwq->worklist)) {
+       /*
+        * When control reaches this point, we're guaranteed to have
+        * at least one idle worker or that someone else has already
+        * assumed the manager role.
+        */
+       worker_clr_flags(worker, WORKER_PREP);
+
+       do {
                struct work_struct *work =
-                       list_first_entry(&cwq->worklist,
+                       list_first_entry(&gcwq->worklist,
                                         struct work_struct, entry);
 
-               /*
-                * The following is a rather inefficient way to close
-                * race window against cpu hotplug operations.  Will
-                * be replaced soon.
-                */
-               if (unlikely(!(worker->flags & WORKER_ROGUE) &&
-                            !cpumask_equal(&worker->task->cpus_allowed,
-                                           get_cpu_mask(gcwq->cpu)))) {
-                       spin_unlock_irq(&gcwq->lock);
-                       set_cpus_allowed_ptr(worker->task,
-                                            get_cpu_mask(gcwq->cpu));
-                       cpu_relax();
-                       spin_lock_irq(&gcwq->lock);
-                       goto recheck;
-               }
-
                if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
                        /* optimization path, not strictly necessary */
                        process_one_work(worker, work);
@@ -961,13 +1984,19 @@ recheck:
                        move_linked_works(work, &worker->scheduled, NULL);
                        process_scheduled_works(worker);
                }
-       }
+       } while (keep_working(gcwq));
+
+       worker_set_flags(worker, WORKER_PREP, false);
+sleep:
+       if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
+               goto recheck;
 
        /*
-        * gcwq->lock is held and there's no work to process, sleep.
-        * Workers are woken up only while holding gcwq->lock, so
-        * setting the current state before releasing gcwq->lock is
-        * enough to prevent losing any event.
+        * gcwq->lock is held and there's no work to process and no
+        * need to manage, sleep.  Workers are woken up only while
+        * holding gcwq->lock or from local cpu, so setting the
+        * current state before releasing gcwq->lock is enough to
+        * prevent losing any event.
         */
        worker_enter_idle(worker);
        __set_current_state(TASK_INTERRUPTIBLE);
@@ -976,6 +2005,83 @@ recheck:
        goto woke_up;
 }
 
+/**
+ * rescuer_thread - the rescuer thread function
+ * @__wq: the associated workqueue
+ *
+ * Workqueue rescuer thread function.  There's one rescuer for each
+ * workqueue which has WQ_RESCUER set.
+ *
+ * Regular work processing on a gcwq may block trying to create a new
+ * worker which uses GFP_KERNEL allocation which has slight chance of
+ * developing into deadlock if some works currently on the same queue
+ * need to be processed to satisfy the GFP_KERNEL allocation.  This is
+ * the problem rescuer solves.
+ *
+ * When such condition is possible, the gcwq summons rescuers of all
+ * workqueues which have works queued on the gcwq and let them process
+ * those works so that forward progress can be guaranteed.
+ *
+ * This should happen rarely.
+ */
+static int rescuer_thread(void *__wq)
+{
+       struct workqueue_struct *wq = __wq;
+       struct worker *rescuer = wq->rescuer;
+       struct list_head *scheduled = &rescuer->scheduled;
+       bool is_unbound = wq->flags & WQ_UNBOUND;
+       unsigned int cpu;
+
+       set_user_nice(current, RESCUER_NICE_LEVEL);
+repeat:
+       set_current_state(TASK_INTERRUPTIBLE);
+
+       if (kthread_should_stop())
+               return 0;
+
+       /*
+        * See whether any cpu is asking for help.  Unbounded
+        * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
+        */
+       for_each_mayday_cpu(cpu, wq->mayday_mask) {
+               unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
+               struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
+               struct global_cwq *gcwq = cwq->gcwq;
+               struct work_struct *work, *n;
+
+               __set_current_state(TASK_RUNNING);
+               mayday_clear_cpu(cpu, wq->mayday_mask);
+
+               /* migrate to the target cpu if possible */
+               rescuer->gcwq = gcwq;
+               worker_maybe_bind_and_lock(rescuer);
+
+               /*
+                * Slurp in all works issued via this workqueue and
+                * process'em.
+                */
+               BUG_ON(!list_empty(&rescuer->scheduled));
+               list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
+                       if (get_work_cwq(work) == cwq)
+                               move_linked_works(work, scheduled, &n);
+
+               process_scheduled_works(rescuer);
+
+               /*
+                * Leave this gcwq.  If keep_working() is %true, notify a
+                * regular worker; otherwise, we end up with 0 concurrency
+                * and stalling the execution.
+                */
+               if (keep_working(gcwq))
+                       wake_up_worker(gcwq);
+
+               spin_unlock_irq(&gcwq->lock);
+       }
+
+       schedule();
+       goto repeat;
+}
+
 struct wq_barrier {
        struct work_struct      work;
        struct completion       done;
@@ -1024,7 +2130,7 @@ static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
         * checks and call back into the fixup functions where we
         * might deadlock.
         */
-       INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
+       INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
        __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
        init_completion(&barr->done);
 
@@ -1090,7 +2196,7 @@ static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
                atomic_set(&wq->nr_cwqs_to_flush, 1);
        }
 
-       for_each_possible_cpu(cpu) {
+       for_each_cwq_cpu(cpu, wq) {
                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
                struct global_cwq *gcwq = cwq->gcwq;
 
@@ -1202,6 +2308,10 @@ void flush_workqueue(struct workqueue_struct *wq)
 
        mutex_lock(&wq->flush_mutex);
 
+       /* we might have raced, check again with mutex held */
+       if (wq->first_flusher != &this_flusher)
+               goto out_unlock;
+
        wq->first_flusher = NULL;
 
        BUG_ON(!list_empty(&this_flusher.list));
@@ -1273,58 +2383,212 @@ out_unlock:
 EXPORT_SYMBOL_GPL(flush_workqueue);
 
 /**
- * flush_work - block until a work_struct's callback has terminated
- * @work: the work which is to be flushed
- *
- * Returns false if @work has already terminated.
- *
- * It is expected that, prior to calling flush_work(), the caller has
- * arranged for the work to not be requeued, otherwise it doesn't make
- * sense to use this function.
+ * drain_workqueue - drain a workqueue
+ * @wq: workqueue to drain
+ *
+ * Wait until the workqueue becomes empty.  While draining is in progress,
+ * only chain queueing is allowed.  IOW, only currently pending or running
+ * work items on @wq can queue further work items on it.  @wq is flushed
+ * repeatedly until it becomes empty.  The number of flushing is detemined
+ * by the depth of chaining and should be relatively short.  Whine if it
+ * takes too long.
  */
-int flush_work(struct work_struct *work)
+void drain_workqueue(struct workqueue_struct *wq)
+{
+       unsigned int flush_cnt = 0;
+       unsigned int cpu;
+
+       /*
+        * __queue_work() needs to test whether there are drainers, is much
+        * hotter than drain_workqueue() and already looks at @wq->flags.
+        * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
+        */
+       spin_lock(&workqueue_lock);
+       if (!wq->nr_drainers++)
+               wq->flags |= WQ_DRAINING;
+       spin_unlock(&workqueue_lock);
+reflush:
+       flush_workqueue(wq);
+
+       for_each_cwq_cpu(cpu, wq) {
+               struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
+               bool drained;
+
+               spin_lock_irq(&cwq->gcwq->lock);
+               drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
+               spin_unlock_irq(&cwq->gcwq->lock);
+
+               if (drained)
+                       continue;
+
+               if (++flush_cnt == 10 ||
+                   (flush_cnt % 100 == 0 && flush_cnt <= 1000))
+                       pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
+                                  wq->name, flush_cnt);
+               goto reflush;
+       }
+
+       spin_lock(&workqueue_lock);
+       if (!--wq->nr_drainers)
+               wq->flags &= ~WQ_DRAINING;
+       spin_unlock(&workqueue_lock);
+}
+EXPORT_SYMBOL_GPL(drain_workqueue);
+
+static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
+                            bool wait_executing)
 {
        struct worker *worker = NULL;
-       struct cpu_workqueue_struct *cwq;
        struct global_cwq *gcwq;
-       struct wq_barrier barr;
+       struct cpu_workqueue_struct *cwq;
 
        might_sleep();
-       cwq = get_wq_data(work);
-       if (!cwq)
-               return 0;
-       gcwq = cwq->gcwq;
-
-       lock_map_acquire(&cwq->wq->lockdep_map);
-       lock_map_release(&cwq->wq->lockdep_map);
+       gcwq = get_work_gcwq(work);
+       if (!gcwq)
+               return false;
 
        spin_lock_irq(&gcwq->lock);
        if (!list_empty(&work->entry)) {
                /*
                 * See the comment near try_to_grab_pending()->smp_rmb().
-                * If it was re-queued under us we are not going to wait.
+                * If it was re-queued to a different gcwq under us, we
+                * are not going to wait.
                 */
                smp_rmb();
-               if (unlikely(cwq != get_wq_data(work)))
+               cwq = get_work_cwq(work);
+               if (unlikely(!cwq || gcwq != cwq->gcwq))
                        goto already_gone;
-       } else {
-               if (cwq->worker && cwq->worker->current_work == work)
-                       worker = cwq->worker;
+       } else if (wait_executing) {
+               worker = find_worker_executing_work(gcwq, work);
                if (!worker)
                        goto already_gone;
-       }
+               cwq = worker->current_cwq;
+       } else
+               goto already_gone;
 
-       insert_wq_barrier(cwq, &barr, work, worker);
+       insert_wq_barrier(cwq, barr, work, worker);
        spin_unlock_irq(&gcwq->lock);
-       wait_for_completion(&barr.done);
-       destroy_work_on_stack(&barr.work);
-       return 1;
+
+       /*
+        * If @max_active is 1 or rescuer is in use, flushing another work
+        * item on the same workqueue may lead to deadlock.  Make sure the
+        * flusher is not running on the same workqueue by verifying write
+        * access.
+        */
+       if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
+               lock_map_acquire(&cwq->wq->lockdep_map);
+       else
+               lock_map_acquire_read(&cwq->wq->lockdep_map);
+       lock_map_release(&cwq->wq->lockdep_map);
+
+       return true;
 already_gone:
        spin_unlock_irq(&gcwq->lock);
-       return 0;
+       return false;
+}
+
+/**
+ * flush_work - wait for a work to finish executing the last queueing instance
+ * @work: the work to flush
+ *
+ * Wait until @work has finished execution.  This function considers
+ * only the last queueing instance of @work.  If @work has been
+ * enqueued across different CPUs on a non-reentrant workqueue or on
+ * multiple workqueues, @work might still be executing on return on
+ * some of the CPUs from earlier queueing.
+ *
+ * If @work was queued only on a non-reentrant, ordered or unbound
+ * workqueue, @work is guaranteed to be idle on return if it hasn't
+ * been requeued since flush started.
+ *
+ * RETURNS:
+ * %true if flush_work() waited for the work to finish execution,
+ * %false if it was already idle.
+ */
+bool flush_work(struct work_struct *work)
+{
+       struct wq_barrier barr;
+
+       if (start_flush_work(work, &barr, true)) {
+               wait_for_completion(&barr.done);
+               destroy_work_on_stack(&barr.work);
+               return true;
+       } else
+               return false;
 }
 EXPORT_SYMBOL_GPL(flush_work);
 
+static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
+{
+       struct wq_barrier barr;
+       struct worker *worker;
+
+       spin_lock_irq(&gcwq->lock);
+
+       worker = find_worker_executing_work(gcwq, work);
+       if (unlikely(worker))
+               insert_wq_barrier(worker->current_cwq, &barr, work, worker);
+
+       spin_unlock_irq(&gcwq->lock);
+
+       if (unlikely(worker)) {
+               wait_for_completion(&barr.done);
+               destroy_work_on_stack(&barr.work);
+               return true;
+       } else
+               return false;
+}
+
+static bool wait_on_work(struct work_struct *work)
+{
+       bool ret = false;
+       int cpu;
+
+       might_sleep();
+
+       lock_map_acquire(&work->lockdep_map);
+       lock_map_release(&work->lockdep_map);
+
+       for_each_gcwq_cpu(cpu)
+               ret |= wait_on_cpu_work(get_gcwq(cpu), work);
+       return ret;
+}
+
+/**
+ * flush_work_sync - wait until a work has finished execution
+ * @work: the work to flush
+ *
+ * Wait until @work has finished execution.  On return, it's
+ * guaranteed that all queueing instances of @work which happened
+ * before this function is called are finished.  In other words, if
+ * @work hasn't been requeued since this function was called, @work is
+ * guaranteed to be idle on return.
+ *
+ * RETURNS:
+ * %true if flush_work_sync() waited for the work to finish execution,
+ * %false if it was already idle.
+ */
+bool flush_work_sync(struct work_struct *work)
+{
+       struct wq_barrier barr;
+       bool pending, waited;
+
+       /* we'll wait for executions separately, queue barr only if pending */
+       pending = start_flush_work(work, &barr, false);
+
+       /* wait for executions to finish */
+       waited = wait_on_work(work);
+
+       /* wait for the pending one */
+       if (pending) {
+               wait_for_completion(&barr.done);
+               destroy_work_on_stack(&barr.work);
+       }
+
+       return pending || waited;
+}
+EXPORT_SYMBOL_GPL(flush_work_sync);
+
 /*
  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
  * so this work can't be re-armed in any way.
@@ -1332,7 +2596,6 @@ EXPORT_SYMBOL_GPL(flush_work);
 static int try_to_grab_pending(struct work_struct *work)
 {
        struct global_cwq *gcwq;
-       struct cpu_workqueue_struct *cwq;
        int ret = -1;
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
@@ -1342,24 +2605,24 @@ static int try_to_grab_pending(struct work_struct *work)
         * The queueing is in progress, or it is already queued. Try to
         * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
         */
-
-       cwq = get_wq_data(work);
-       if (!cwq)
+       gcwq = get_work_gcwq(work);
+       if (!gcwq)
                return ret;
-       gcwq = cwq->gcwq;
 
        spin_lock_irq(&gcwq->lock);
        if (!list_empty(&work->entry)) {
                /*
-                * This work is queued, but perhaps we locked the wrong cwq.
+                * This work is queued, but perhaps we locked the wrong gcwq.
                 * In that case we must see the new value after rmb(), see
                 * insert_work()->wmb().
                 */
                smp_rmb();
-               if (cwq == get_wq_data(work)) {
+               if (gcwq == get_work_gcwq(work)) {
                        debug_work_deactivate(work);
                        list_del_init(&work->entry);
-                       cwq_dec_nr_in_flight(cwq, get_work_color(work));
+                       cwq_dec_nr_in_flight(get_work_cwq(work),
+                               get_work_color(work),
+                               *work_data_bits(work) & WORK_STRUCT_DELAYED);
                        ret = 1;
                }
        }
@@ -1368,51 +2631,7 @@ static int try_to_grab_pending(struct work_struct *work)
        return ret;
 }
 
-static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
-                               struct work_struct *work)
-{
-       struct global_cwq *gcwq = cwq->gcwq;
-       struct wq_barrier barr;
-       struct worker *worker;
-
-       spin_lock_irq(&gcwq->lock);
-
-       worker = NULL;
-       if (unlikely(cwq->worker && cwq->worker->current_work == work)) {
-               worker = cwq->worker;
-               insert_wq_barrier(cwq, &barr, work, worker);
-       }
-
-       spin_unlock_irq(&gcwq->lock);
-
-       if (unlikely(worker)) {
-               wait_for_completion(&barr.done);
-               destroy_work_on_stack(&barr.work);
-       }
-}
-
-static void wait_on_work(struct work_struct *work)
-{
-       struct cpu_workqueue_struct *cwq;
-       struct workqueue_struct *wq;
-       int cpu;
-
-       might_sleep();
-
-       lock_map_acquire(&work->lockdep_map);
-       lock_map_release(&work->lockdep_map);
-
-       cwq = get_wq_data(work);
-       if (!cwq)
-               return;
-
-       wq = cwq->wq;
-
-       for_each_possible_cpu(cpu)
-               wait_on_cpu_work(get_cwq(cpu, wq), work);
-}
-
-static int __cancel_work_timer(struct work_struct *work,
+static bool __cancel_work_timer(struct work_struct *work,
                                struct timer_list* timer)
 {
        int ret;
@@ -1424,54 +2643,91 @@ static int __cancel_work_timer(struct work_struct *work,
                wait_on_work(work);
        } while (unlikely(ret < 0));
 
-       clear_wq_data(work);
+       clear_work_data(work);
        return ret;
 }
 
 /**
- * cancel_work_sync - block until a work_struct's callback has terminated
- * @work: the work which is to be flushed
- *
- * Returns true if @work was pending.
- *
- * cancel_work_sync() will cancel the work if it is queued. If the work's
- * callback appears to be running, cancel_work_sync() will block until it
- * has completed.
+ * cancel_work_sync - cancel a work and wait for it to finish
+ * @work: the work to cancel
  *
- * It is possible to use this function if the work re-queues itself. It can
- * cancel the work even if it migrates to another workqueue, however in that
- * case it only guarantees that work->func() has completed on the last queued
- * workqueue.
+ * Cancel @work and wait for its execution to finish.  This function
+ * can be used even if the work re-queues itself or migrates to
+ * another workqueue.  On return from this function, @work is
+ * guaranteed to be not pending or executing on any CPU.
  *
- * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
- * pending, otherwise it goes into a busy-wait loop until the timer expires.
+ * cancel_work_sync(&delayed_work->work) must not be used for
+ * delayed_work's.  Use cancel_delayed_work_sync() instead.
  *
- * The caller must ensure that workqueue_struct on which this work was last
+ * The caller must ensure that the workqueue on which @work was last
  * queued can't be destroyed before this function returns.
+ *
+ * RETURNS:
+ * %true if @work was pending, %false otherwise.
  */
-int cancel_work_sync(struct work_struct *work)
+bool cancel_work_sync(struct work_struct *work)
 {
        return __cancel_work_timer(work, NULL);
 }
 EXPORT_SYMBOL_GPL(cancel_work_sync);
 
 /**
- * cancel_delayed_work_sync - reliably kill off a delayed work.
- * @dwork: the delayed work struct
+ * flush_delayed_work - wait for a dwork to finish executing the last queueing
+ * @dwork: the delayed work to flush
  *
- * Returns true if @dwork was pending.
+ * Delayed timer is cancelled and the pending work is queued for
+ * immediate execution.  Like flush_work(), this function only
+ * considers the last queueing instance of @dwork.
  *
- * It is possible to use this function if @dwork rearms itself via queue_work()
- * or queue_delayed_work(). See also the comment for cancel_work_sync().
+ * RETURNS:
+ * %true if flush_work() waited for the work to finish execution,
+ * %false if it was already idle.
+ */
+bool flush_delayed_work(struct delayed_work *dwork)
+{
+       if (del_timer_sync(&dwork->timer))
+               __queue_work(raw_smp_processor_id(),
+                            get_work_cwq(&dwork->work)->wq, &dwork->work);
+       return flush_work(&dwork->work);
+}
+EXPORT_SYMBOL(flush_delayed_work);
+
+/**
+ * flush_delayed_work_sync - wait for a dwork to finish
+ * @dwork: the delayed work to flush
+ *
+ * Delayed timer is cancelled and the pending work is queued for
+ * execution immediately.  Other than timer handling, its behavior
+ * is identical to flush_work_sync().
+ *
+ * RETURNS:
+ * %true if flush_work_sync() waited for the work to finish execution,
+ * %false if it was already idle.
  */
-int cancel_delayed_work_sync(struct delayed_work *dwork)
+bool flush_delayed_work_sync(struct delayed_work *dwork)
+{
+       if (del_timer_sync(&dwork->timer))
+               __queue_work(raw_smp_processor_id(),
+                            get_work_cwq(&dwork->work)->wq, &dwork->work);
+       return flush_work_sync(&dwork->work);
+}
+EXPORT_SYMBOL(flush_delayed_work_sync);
+
+/**
+ * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
+ * @dwork: the delayed work cancel
+ *
+ * This is cancel_work_sync() for delayed works.
+ *
+ * RETURNS:
+ * %true if @dwork was pending, %false otherwise.
+ */
+bool cancel_delayed_work_sync(struct delayed_work *dwork)
 {
        return __cancel_work_timer(&dwork->work, &dwork->timer);
 }
 EXPORT_SYMBOL(cancel_delayed_work_sync);
 
-static struct workqueue_struct *keventd_wq __read_mostly;
-
 /**
  * schedule_work - put work task in global workqueue
  * @work: job to be done
@@ -1485,7 +2741,7 @@ static struct workqueue_struct *keventd_wq __read_mostly;
  */
 int schedule_work(struct work_struct *work)
 {
-       return queue_work(keventd_wq, work);
+       return queue_work(system_wq, work);
 }
 EXPORT_SYMBOL(schedule_work);
 
@@ -1498,7 +2754,7 @@ EXPORT_SYMBOL(schedule_work);
  */
 int schedule_work_on(int cpu, struct work_struct *work)
 {
-       return queue_work_on(cpu, keventd_wq, work);
+       return queue_work_on(cpu, system_wq, work);
 }
 EXPORT_SYMBOL(schedule_work_on);
 
@@ -1513,28 +2769,11 @@ EXPORT_SYMBOL(schedule_work_on);
 int schedule_delayed_work(struct delayed_work *dwork,
                                        unsigned long delay)
 {
-       return queue_delayed_work(keventd_wq, dwork, delay);
+       return queue_delayed_work(system_wq, dwork, delay);
 }
 EXPORT_SYMBOL(schedule_delayed_work);
 
 /**
- * flush_delayed_work - block until a dwork_struct's callback has terminated
- * @dwork: the delayed work which is to be flushed
- *
- * Any timeout is cancelled, and any pending work is run immediately.
- */
-void flush_delayed_work(struct delayed_work *dwork)
-{
-       if (del_timer_sync(&dwork->timer)) {
-               __queue_work(get_cpu(), get_wq_data(&dwork->work)->wq,
-                            &dwork->work);
-               put_cpu();
-       }
-       flush_work(&dwork->work);
-}
-EXPORT_SYMBOL(flush_delayed_work);
-
-/**
  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
  * @cpu: cpu to use
  * @dwork: job to be done
@@ -1546,24 +2785,25 @@ EXPORT_SYMBOL(flush_delayed_work);
 int schedule_delayed_work_on(int cpu,
                        struct delayed_work *dwork, unsigned long delay)
 {
-       return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
+       return queue_delayed_work_on(cpu, system_wq, dwork, delay);
 }
 EXPORT_SYMBOL(schedule_delayed_work_on);
 
 /**
- * schedule_on_each_cpu - call a function on each online CPU from keventd
+ * schedule_on_each_cpu - execute a function synchronously on each online CPU
  * @func: the function to call
  *
- * Returns zero on success.
- * Returns -ve errno on failure.
- *
+ * schedule_on_each_cpu() executes @func on each online CPU using the
+ * system workqueue and blocks until all CPUs have completed.
  * schedule_on_each_cpu() is very slow.
+ *
+ * RETURNS:
+ * 0 on success, -errno on failure.
  */
 int schedule_on_each_cpu(work_func_t func)
 {
        int cpu;
-       int orig = -1;
-       struct work_struct *works;
+       struct work_struct __percpu *works;
 
        works = alloc_percpu(struct work_struct);
        if (!works)
@@ -1571,23 +2811,12 @@ int schedule_on_each_cpu(work_func_t func)
 
        get_online_cpus();
 
-       /*
-        * When running in keventd don't schedule a work item on
-        * itself.  Can just call directly because the work queue is
-        * already bound.  This also is faster.
-        */
-       if (current_is_keventd())
-               orig = raw_smp_processor_id();
-
        for_each_online_cpu(cpu) {
                struct work_struct *work = per_cpu_ptr(works, cpu);
 
                INIT_WORK(work, func);
-               if (cpu != orig)
-                       schedule_work_on(cpu, work);
+               schedule_work_on(cpu, work);
        }
-       if (orig >= 0)
-               func(per_cpu_ptr(works, orig));
 
        for_each_online_cpu(cpu)
                flush_work(per_cpu_ptr(works, cpu));
@@ -1623,7 +2852,7 @@ int schedule_on_each_cpu(work_func_t func)
  */
 void flush_scheduled_work(void)
 {
-       flush_workqueue(keventd_wq);
+       flush_workqueue(system_wq);
 }
 EXPORT_SYMBOL(flush_scheduled_work);
 
@@ -1655,26 +2884,10 @@ EXPORT_SYMBOL_GPL(execute_in_process_context);
 
 int keventd_up(void)
 {
-       return keventd_wq != NULL;
+       return system_wq != NULL;
 }
 
-int current_is_keventd(void)
-{
-       struct cpu_workqueue_struct *cwq;
-       int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
-       int ret = 0;
-
-       BUG_ON(!keventd_wq);
-
-       cwq = get_cwq(cpu, keventd_wq);
-       if (current == cwq->worker->task)
-               ret = 1;
-
-       return ret;
-
-}
-
-static struct cpu_workqueue_struct *alloc_cwqs(void)
+static int alloc_cwqs(struct workqueue_struct *wq)
 {
        /*
         * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
@@ -1684,82 +2897,108 @@ static struct cpu_workqueue_struct *alloc_cwqs(void)
        const size_t size = sizeof(struct cpu_workqueue_struct);
        const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
                                   __alignof__(unsigned long long));
-       struct cpu_workqueue_struct *cwqs;
-#ifndef CONFIG_SMP
-       void *ptr;
-
-       /*
-        * On UP, percpu allocator doesn't honor alignment parameter
-        * and simply uses arch-dependent default.  Allocate enough
-        * room to align cwq and put an extra pointer at the end
-        * pointing back to the originally allocated pointer which
-        * will be used for free.
-        *
-        * FIXME: This really belongs to UP percpu code.  Update UP
-        * percpu code to honor alignment and remove this ugliness.
-        */
-       ptr = __alloc_percpu(size + align + sizeof(void *), 1);
-       cwqs = PTR_ALIGN(ptr, align);
-       *(void **)per_cpu_ptr(cwqs + 1, 0) = ptr;
+#ifdef CONFIG_SMP
+       bool percpu = !(wq->flags & WQ_UNBOUND);
 #else
-       /* On SMP, percpu allocator can do it itself */
-       cwqs = __alloc_percpu(size, align);
+       bool percpu = false;
 #endif
+
+       if (percpu)
+               wq->cpu_wq.pcpu = __alloc_percpu(size, align);
+       else {
+               void *ptr;
+
+               /*
+                * Allocate enough room to align cwq and put an extra
+                * pointer at the end pointing back to the originally
+                * allocated pointer which will be used for free.
+                */
+               ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
+               if (ptr) {
+                       wq->cpu_wq.single = PTR_ALIGN(ptr, align);
+                       *(void **)(wq->cpu_wq.single + 1) = ptr;
+               }
+       }
+
        /* just in case, make sure it's actually aligned */
-       BUG_ON(!IS_ALIGNED((unsigned long)cwqs, align));
-       return cwqs;
+       BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
+       return wq->cpu_wq.v ? 0 : -ENOMEM;
 }
 
-static void free_cwqs(struct cpu_workqueue_struct *cwqs)
+static void free_cwqs(struct workqueue_struct *wq)
 {
-#ifndef CONFIG_SMP
-       /* on UP, the pointer to free is stored right after the cwq */
-       if (cwqs)
-               free_percpu(*(void **)per_cpu_ptr(cwqs + 1, 0));
+#ifdef CONFIG_SMP
+       bool percpu = !(wq->flags & WQ_UNBOUND);
 #else
-       free_percpu(cwqs);
+       bool percpu = false;
 #endif
+
+       if (percpu)
+               free_percpu(wq->cpu_wq.pcpu);
+       else if (wq->cpu_wq.single) {
+               /* the pointer to free is stored right after the cwq */
+               kfree(*(void **)(wq->cpu_wq.single + 1));
+       }
 }
 
-struct workqueue_struct *__create_workqueue_key(const char *name,
-                                               unsigned int flags,
-                                               int max_active,
-                                               struct lock_class_key *key,
-                                               const char *lock_name)
+static int wq_clamp_max_active(int max_active, unsigned int flags,
+                              const char *name)
+{
+       int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
+
+       if (max_active < 1 || max_active > lim)
+               printk(KERN_WARNING "workqueue: max_active %d requested for %s "
+                      "is out of range, clamping between %d and %d\n",
+                      max_active, name, 1, lim);
+
+       return clamp_val(max_active, 1, lim);
+}
+
+struct workqueue_struct *__alloc_workqueue_key(const char *name,
+                                              unsigned int flags,
+                                              int max_active,
+                                              struct lock_class_key *key,
+                                              const char *lock_name)
 {
-       bool singlethread = flags & WQ_SINGLE_THREAD;
        struct workqueue_struct *wq;
-       bool failed = false;
        unsigned int cpu;
 
-       max_active = clamp_val(max_active, 1, INT_MAX);
+       /*
+        * Workqueues which may be used during memory reclaim should
+        * have a rescuer to guarantee forward progress.
+        */
+       if (flags & WQ_MEM_RECLAIM)
+               flags |= WQ_RESCUER;
+
+       /*
+        * Unbound workqueues aren't concurrency managed and should be
+        * dispatched to workers immediately.
+        */
+       if (flags & WQ_UNBOUND)
+               flags |= WQ_HIGHPRI;
+
+       max_active = max_active ?: WQ_DFL_ACTIVE;
+       max_active = wq_clamp_max_active(max_active, flags, name);
 
        wq = kzalloc(sizeof(*wq), GFP_KERNEL);
        if (!wq)
                goto err;
 
-       wq->cpu_wq = alloc_cwqs();
-       if (!wq->cpu_wq)
-               goto err;
-
        wq->flags = flags;
        wq->saved_max_active = max_active;
        mutex_init(&wq->flush_mutex);
        atomic_set(&wq->nr_cwqs_to_flush, 0);
        INIT_LIST_HEAD(&wq->flusher_queue);
        INIT_LIST_HEAD(&wq->flusher_overflow);
+
        wq->name = name;
        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
        INIT_LIST_HEAD(&wq->list);
 
-       cpu_maps_update_begin();
-       /*
-        * We must initialize cwqs for each possible cpu even if we
-        * are going to call destroy_workqueue() finally. Otherwise
-        * cpu_up() can hit the uninitialized cwq once we drop the
-        * lock.
-        */
-       for_each_possible_cpu(cpu) {
+       if (alloc_cwqs(wq) < 0)
+               goto err;
+
+       for_each_cwq_cpu(cpu, wq) {
                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
                struct global_cwq *gcwq = get_gcwq(cpu);
 
@@ -1768,17 +3007,25 @@ struct workqueue_struct *__create_workqueue_key(const char *name,
                cwq->wq = wq;
                cwq->flush_color = -1;
                cwq->max_active = max_active;
-               INIT_LIST_HEAD(&cwq->worklist);
                INIT_LIST_HEAD(&cwq->delayed_works);
+       }
 
-               if (failed)
-                       continue;
-               cwq->worker = create_worker(cwq,
-                                           cpu_online(cpu) && !singlethread);
-               if (cwq->worker)
-                       start_worker(cwq->worker);
-               else
-                       failed = true;
+       if (flags & WQ_RESCUER) {
+               struct worker *rescuer;
+
+               if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
+                       goto err;
+
+               wq->rescuer = rescuer = alloc_worker();
+               if (!rescuer)
+                       goto err;
+
+               rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
+               if (IS_ERR(rescuer->task))
+                       goto err;
+
+               rescuer->task->flags |= PF_THREAD_BOUND;
+               wake_up_process(rescuer->task);
        }
 
        /*
@@ -1788,29 +3035,25 @@ struct workqueue_struct *__create_workqueue_key(const char *name,
         */
        spin_lock(&workqueue_lock);
 
-       if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
-               for_each_possible_cpu(cpu)
+       if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
+               for_each_cwq_cpu(cpu, wq)
                        get_cwq(cpu, wq)->max_active = 0;
 
        list_add(&wq->list, &workqueues);
 
        spin_unlock(&workqueue_lock);
 
-       cpu_maps_update_done();
-
-       if (failed) {
-               destroy_workqueue(wq);
-               wq = NULL;
-       }
        return wq;
 err:
        if (wq) {
-               free_cwqs(wq->cpu_wq);
+               free_cwqs(wq);
+               free_mayday_mask(wq->mayday_mask);
+               kfree(wq->rescuer);
                kfree(wq);
        }
        return NULL;
 }
-EXPORT_SYMBOL_GPL(__create_workqueue_key);
+EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
 
 /**
  * destroy_workqueue - safely terminate a workqueue
@@ -1822,47 +3065,160 @@ void destroy_workqueue(struct workqueue_struct *wq)
 {
        unsigned int cpu;
 
-       flush_workqueue(wq);
+       /* drain it before proceeding with destruction */
+       drain_workqueue(wq);
 
        /*
         * wq list is used to freeze wq, remove from list after
         * flushing is complete in case freeze races us.
         */
-       cpu_maps_update_begin();
        spin_lock(&workqueue_lock);
        list_del(&wq->list);
        spin_unlock(&workqueue_lock);
-       cpu_maps_update_done();
 
-       for_each_possible_cpu(cpu) {
+       /* sanity check */
+       for_each_cwq_cpu(cpu, wq) {
                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
                int i;
 
-               if (cwq->worker) {
-                       spin_lock_irq(&cwq->gcwq->lock);
-                       destroy_worker(cwq->worker);
-                       cwq->worker = NULL;
-                       spin_unlock_irq(&cwq->gcwq->lock);
-               }
-
                for (i = 0; i < WORK_NR_COLORS; i++)
                        BUG_ON(cwq->nr_in_flight[i]);
                BUG_ON(cwq->nr_active);
                BUG_ON(!list_empty(&cwq->delayed_works));
        }
 
-       free_cwqs(wq->cpu_wq);
+       if (wq->flags & WQ_RESCUER) {
+               kthread_stop(wq->rescuer->task);
+               free_mayday_mask(wq->mayday_mask);
+               kfree(wq->rescuer);
+       }
+
+       free_cwqs(wq);
        kfree(wq);
 }
 EXPORT_SYMBOL_GPL(destroy_workqueue);
 
+/**
+ * workqueue_set_max_active - adjust max_active of a workqueue
+ * @wq: target workqueue
+ * @max_active: new max_active value.
+ *
+ * Set max_active of @wq to @max_active.
+ *
+ * CONTEXT:
+ * Don't call from IRQ context.
+ */
+void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
+{
+       unsigned int cpu;
+
+       max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
+
+       spin_lock(&workqueue_lock);
+
+       wq->saved_max_active = max_active;
+
+       for_each_cwq_cpu(cpu, wq) {
+               struct global_cwq *gcwq = get_gcwq(cpu);
+
+               spin_lock_irq(&gcwq->lock);
+
+               if (!(wq->flags & WQ_FREEZABLE) ||
+                   !(gcwq->flags & GCWQ_FREEZING))
+                       get_cwq(gcwq->cpu, wq)->max_active = max_active;
+
+               spin_unlock_irq(&gcwq->lock);
+       }
+
+       spin_unlock(&workqueue_lock);
+}
+EXPORT_SYMBOL_GPL(workqueue_set_max_active);
+
+/**
+ * workqueue_congested - test whether a workqueue is congested
+ * @cpu: CPU in question
+ * @wq: target workqueue
+ *
+ * Test whether @wq's cpu workqueue for @cpu is congested.  There is
+ * no synchronization around this function and the test result is
+ * unreliable and only useful as advisory hints or for debugging.
+ *
+ * RETURNS:
+ * %true if congested, %false otherwise.
+ */
+bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
+{
+       struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
+
+       return !list_empty(&cwq->delayed_works);
+}
+EXPORT_SYMBOL_GPL(workqueue_congested);
+
+/**
+ * work_cpu - return the last known associated cpu for @work
+ * @work: the work of interest
+ *
+ * RETURNS:
+ * CPU number if @work was ever queued.  WORK_CPU_NONE otherwise.
+ */
+unsigned int work_cpu(struct work_struct *work)
+{
+       struct global_cwq *gcwq = get_work_gcwq(work);
+
+       return gcwq ? gcwq->cpu : WORK_CPU_NONE;
+}
+EXPORT_SYMBOL_GPL(work_cpu);
+
+/**
+ * work_busy - test whether a work is currently pending or running
+ * @work: the work to be tested
+ *
+ * Test whether @work is currently pending or running.  There is no
+ * synchronization around this function and the test result is
+ * unreliable and only useful as advisory hints or for debugging.
+ * Especially for reentrant wqs, the pending state might hide the
+ * running state.
+ *
+ * RETURNS:
+ * OR'd bitmask of WORK_BUSY_* bits.
+ */
+unsigned int work_busy(struct work_struct *work)
+{
+       struct global_cwq *gcwq = get_work_gcwq(work);
+       unsigned long flags;
+       unsigned int ret = 0;
+
+       if (!gcwq)
+               return false;
+
+       spin_lock_irqsave(&gcwq->lock, flags);
+
+       if (work_pending(work))
+               ret |= WORK_BUSY_PENDING;
+       if (find_worker_executing_work(gcwq, work))
+               ret |= WORK_BUSY_RUNNING;
+
+       spin_unlock_irqrestore(&gcwq->lock, flags);
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(work_busy);
+
 /*
  * CPU hotplug.
  *
- * CPU hotplug is implemented by allowing cwqs to be detached from
- * CPU, running with unbound workers and allowing them to be
- * reattached later if the cpu comes back online.  A separate thread
- * is created to govern cwqs in such state and is called the trustee.
+ * There are two challenges in supporting CPU hotplug.  Firstly, there
+ * are a lot of assumptions on strong associations among work, cwq and
+ * gcwq which make migrating pending and scheduled works very
+ * difficult to implement without impacting hot paths.  Secondly,
+ * gcwqs serve mix of short, long and very long running works making
+ * blocked draining impractical.
+ *
+ * This is solved by allowing a gcwq to be detached from CPU, running
+ * it with unbound (rogue) workers and allowing it to be reattached
+ * later if the cpu comes back online.  A separate thread is created
+ * to govern a gcwq in such state and is called the trustee of the
+ * gcwq.
  *
  * Trustee states and their descriptions.
  *
@@ -1870,11 +3226,12 @@ EXPORT_SYMBOL_GPL(destroy_workqueue);
  *             new trustee is started with this state.
  *
  * IN_CHARGE   Once started, trustee will enter this state after
- *             making all existing workers rogue.  DOWN_PREPARE waits
- *             for trustee to enter this state.  After reaching
- *             IN_CHARGE, trustee tries to execute the pending
- *             worklist until it's empty and the state is set to
- *             BUTCHER, or the state is set to RELEASE.
+ *             assuming the manager role and making all existing
+ *             workers rogue.  DOWN_PREPARE waits for trustee to
+ *             enter this state.  After reaching IN_CHARGE, trustee
+ *             tries to execute the pending worklist until it's empty
+ *             and the state is set to BUTCHER, or the state is set
+ *             to RELEASE.
  *
  * BUTCHER     Command state which is set by the cpu callback after
  *             the cpu has went down.  Once this state is set trustee
@@ -1885,7 +3242,9 @@ EXPORT_SYMBOL_GPL(destroy_workqueue);
  * RELEASE     Command state which is set by the cpu callback if the
  *             cpu down has been canceled or it has come online
  *             again.  After recognizing this state, trustee stops
- *             trying to drain or butcher and transits to DONE.
+ *             trying to drain or butcher and clears ROGUE, rebinds
+ *             all remaining workers back to the cpu and releases
+ *             manager role.
  *
  * DONE                Trustee will enter this state after BUTCHER or RELEASE
  *             is complete.
@@ -1951,25 +3310,52 @@ static int __cpuinit trustee_thread(void *__gcwq)
 {
        struct global_cwq *gcwq = __gcwq;
        struct worker *worker;
+       struct work_struct *work;
        struct hlist_node *pos;
+       long rc;
        int i;
 
        BUG_ON(gcwq->cpu != smp_processor_id());
 
        spin_lock_irq(&gcwq->lock);
        /*
-        * Make all multithread workers rogue.  Trustee must be bound
-        * to the target cpu and can't be cancelled.
+        * Claim the manager position and make all workers rogue.
+        * Trustee must be bound to the target cpu and can't be
+        * cancelled.
         */
        BUG_ON(gcwq->cpu != smp_processor_id());
+       rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
+       BUG_ON(rc < 0);
+
+       gcwq->flags |= GCWQ_MANAGING_WORKERS;
 
        list_for_each_entry(worker, &gcwq->idle_list, entry)
-               if (!(worker->cwq->wq->flags & WQ_SINGLE_THREAD))
-                       worker->flags |= WORKER_ROGUE;
+               worker->flags |= WORKER_ROGUE;
 
        for_each_busy_worker(worker, i, pos, gcwq)
-               if (!(worker->cwq->wq->flags & WQ_SINGLE_THREAD))
-                       worker->flags |= WORKER_ROGUE;
+               worker->flags |= WORKER_ROGUE;
+
+       /*
+        * Call schedule() so that we cross rq->lock and thus can
+        * guarantee sched callbacks see the rogue flag.  This is
+        * necessary as scheduler callbacks may be invoked from other
+        * cpus.
+        */
+       spin_unlock_irq(&gcwq->lock);
+       schedule();
+       spin_lock_irq(&gcwq->lock);
+
+       /*
+        * Sched callbacks are disabled now.  Zap nr_running.  After
+        * this, nr_running stays zero and need_more_worker() and
+        * keep_working() are always true as long as the worklist is
+        * not empty.
+        */
+       atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
+
+       spin_unlock_irq(&gcwq->lock);
+       del_timer_sync(&gcwq->idle_timer);
+       spin_lock_irq(&gcwq->lock);
 
        /*
         * We're now in charge.  Notify and proceed to drain.  We need
@@ -1983,18 +3369,90 @@ static int __cpuinit trustee_thread(void *__gcwq)
        /*
         * The original cpu is in the process of dying and may go away
         * anytime now.  When that happens, we and all workers would
-        * be migrated to other cpus.  Try draining any left work.
-        * Note that if the gcwq is frozen, there may be frozen works
-        * in freezeable cwqs.  Don't declare completion while frozen.
+        * be migrated to other cpus.  Try draining any left work.  We
+        * want to get it over with ASAP - spam rescuers, wake up as
+        * many idlers as necessary and create new ones till the
+        * worklist is empty.  Note that if the gcwq is frozen, there
+        * may be frozen works in freezable cwqs.  Don't declare
+        * completion while frozen.
         */
        while (gcwq->nr_workers != gcwq->nr_idle ||
               gcwq->flags & GCWQ_FREEZING ||
               gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
+               int nr_works = 0;
+
+               list_for_each_entry(work, &gcwq->worklist, entry) {
+                       send_mayday(work);
+                       nr_works++;
+               }
+
+               list_for_each_entry(worker, &gcwq->idle_list, entry) {
+                       if (!nr_works--)
+                               break;
+                       wake_up_process(worker->task);
+               }
+
+               if (need_to_create_worker(gcwq)) {
+                       spin_unlock_irq(&gcwq->lock);
+                       worker = create_worker(gcwq, false);
+                       spin_lock_irq(&gcwq->lock);
+                       if (worker) {
+                               worker->flags |= WORKER_ROGUE;
+                               start_worker(worker);
+                       }
+               }
+
                /* give a breather */
                if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
                        break;
        }
 
+       /*
+        * Either all works have been scheduled and cpu is down, or
+        * cpu down has already been canceled.  Wait for and butcher
+        * all workers till we're canceled.
+        */
+       do {
+               rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
+               while (!list_empty(&gcwq->idle_list))
+                       destroy_worker(list_first_entry(&gcwq->idle_list,
+                                                       struct worker, entry));
+       } while (gcwq->nr_workers && rc >= 0);
+
+       /*
+        * At this point, either draining has completed and no worker
+        * is left, or cpu down has been canceled or the cpu is being
+        * brought back up.  There shouldn't be any idle one left.
+        * Tell the remaining busy ones to rebind once it finishes the
+        * currently scheduled works by scheduling the rebind_work.
+        */
+       WARN_ON(!list_empty(&gcwq->idle_list));
+
+       for_each_busy_worker(worker, i, pos, gcwq) {
+               struct work_struct *rebind_work = &worker->rebind_work;
+
+               /*
+                * Rebind_work may race with future cpu hotplug
+                * operations.  Use a separate flag to mark that
+                * rebinding is scheduled.
+                */
+               worker->flags |= WORKER_REBIND;
+               worker->flags &= ~WORKER_ROGUE;
+
+               /* queue rebind_work, wq doesn't matter, use the default one */
+               if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
+                                    work_data_bits(rebind_work)))
+                       continue;
+
+               debug_work_activate(rebind_work);
+               insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
+                           worker->scheduled.next,
+                           work_color_to_flags(WORK_NO_COLOR));
+       }
+
+       /* relinquish manager role */
+       gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
+
        /* notify completion */
        gcwq->trustee = NULL;
        gcwq->trustee_state = TRUSTEE_DONE;
@@ -2015,6 +3473,8 @@ static int __cpuinit trustee_thread(void *__gcwq)
  * multiple times.  To be used by cpu_callback.
  */
 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
+__releases(&gcwq->lock)
+__acquires(&gcwq->lock)
 {
        if (!(gcwq->trustee_state == state ||
              gcwq->trustee_state == TRUSTEE_DONE)) {
@@ -2033,10 +3493,8 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
        unsigned int cpu = (unsigned long)hcpu;
        struct global_cwq *gcwq = get_gcwq(cpu);
        struct task_struct *new_trustee = NULL;
-       struct worker *worker;
-       struct hlist_node *pos;
+       struct worker *uninitialized_var(new_worker);
        unsigned long flags;
-       int i;
 
        action &= ~CPU_TASKS_FROZEN;
 
@@ -2047,6 +3505,15 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
                if (IS_ERR(new_trustee))
                        return notifier_from_errno(PTR_ERR(new_trustee));
                kthread_bind(new_trustee, cpu);
+               /* fall through */
+       case CPU_UP_PREPARE:
+               BUG_ON(gcwq->first_idle);
+               new_worker = create_worker(gcwq, false);
+               if (!new_worker) {
+                       if (new_trustee)
+                               kthread_stop(new_trustee);
+                       return NOTIFY_BAD;
+               }
        }
 
        /* some are called w/ irq disabled, don't disturb irq status */
@@ -2060,28 +3527,50 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
                gcwq->trustee_state = TRUSTEE_START;
                wake_up_process(gcwq->trustee);
                wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
+               /* fall through */
+       case CPU_UP_PREPARE:
+               BUG_ON(gcwq->first_idle);
+               gcwq->first_idle = new_worker;
+               break;
+
+       case CPU_DYING:
+               /*
+                * Before this, the trustee and all workers except for
+                * the ones which are still executing works from
+                * before the last CPU down must be on the cpu.  After
+                * this, they'll all be diasporas.
+                */
+               gcwq->flags |= GCWQ_DISASSOCIATED;
                break;
 
        case CPU_POST_DEAD:
                gcwq->trustee_state = TRUSTEE_BUTCHER;
+               /* fall through */
+       case CPU_UP_CANCELED:
+               destroy_worker(gcwq->first_idle);
+               gcwq->first_idle = NULL;
                break;
 
        case CPU_DOWN_FAILED:
        case CPU_ONLINE:
+               gcwq->flags &= ~GCWQ_DISASSOCIATED;
                if (gcwq->trustee_state != TRUSTEE_DONE) {
                        gcwq->trustee_state = TRUSTEE_RELEASE;
                        wake_up_process(gcwq->trustee);
                        wait_trustee_state(gcwq, TRUSTEE_DONE);
                }
 
-               /* clear ROGUE from all multithread workers */
-               list_for_each_entry(worker, &gcwq->idle_list, entry)
-                       if (!(worker->cwq->wq->flags & WQ_SINGLE_THREAD))
-                               worker->flags &= ~WORKER_ROGUE;
-
-               for_each_busy_worker(worker, i, pos, gcwq)
-                       if (!(worker->cwq->wq->flags & WQ_SINGLE_THREAD))
-                               worker->flags &= ~WORKER_ROGUE;
+               /*
+                * Trustee is done and there might be no worker left.
+                * Put the first_idle in and request a real manager to
+                * take a look.
+                */
+               spin_unlock_irq(&gcwq->lock);
+               kthread_bind(gcwq->first_idle->task, cpu);
+               spin_lock_irq(&gcwq->lock);
+               gcwq->flags |= GCWQ_MANAGE_WORKERS;
+               start_worker(gcwq->first_idle);
+               gcwq->first_idle = NULL;
                break;
        }
 
@@ -2142,16 +3631,15 @@ EXPORT_SYMBOL_GPL(work_on_cpu);
 /**
  * freeze_workqueues_begin - begin freezing workqueues
  *
- * Start freezing workqueues.  After this function returns, all
- * freezeable workqueues will queue new works to their frozen_works
- * list instead of the cwq ones.
+ * Start freezing workqueues.  After this function returns, all freezable
+ * workqueues will queue new works to their frozen_works list instead of
+ * gcwq->worklist.
  *
  * CONTEXT:
  * Grabs and releases workqueue_lock and gcwq->lock's.
  */
 void freeze_workqueues_begin(void)
 {
-       struct workqueue_struct *wq;
        unsigned int cpu;
 
        spin_lock(&workqueue_lock);
@@ -2159,8 +3647,9 @@ void freeze_workqueues_begin(void)
        BUG_ON(workqueue_freezing);
        workqueue_freezing = true;
 
-       for_each_possible_cpu(cpu) {
+       for_each_gcwq_cpu(cpu) {
                struct global_cwq *gcwq = get_gcwq(cpu);
+               struct workqueue_struct *wq;
 
                spin_lock_irq(&gcwq->lock);
 
@@ -2170,7 +3659,7 @@ void freeze_workqueues_begin(void)
                list_for_each_entry(wq, &workqueues, list) {
                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
 
-                       if (wq->flags & WQ_FREEZEABLE)
+                       if (cwq && wq->flags & WQ_FREEZABLE)
                                cwq->max_active = 0;
                }
 
@@ -2181,7 +3670,7 @@ void freeze_workqueues_begin(void)
 }
 
 /**
- * freeze_workqueues_busy - are freezeable workqueues still busy?
+ * freeze_workqueues_busy - are freezable workqueues still busy?
  *
  * Check whether freezing is complete.  This function must be called
  * between freeze_workqueues_begin() and thaw_workqueues().
@@ -2190,12 +3679,11 @@ void freeze_workqueues_begin(void)
  * Grabs and releases workqueue_lock.
  *
  * RETURNS:
- * %true if some freezeable workqueues are still busy.  %false if
- * freezing is complete.
+ * %true if some freezable workqueues are still busy.  %false if freezing
+ * is complete.
  */
 bool freeze_workqueues_busy(void)
 {
-       struct workqueue_struct *wq;
        unsigned int cpu;
        bool busy = false;
 
@@ -2203,7 +3691,8 @@ bool freeze_workqueues_busy(void)
 
        BUG_ON(!workqueue_freezing);
 
-       for_each_possible_cpu(cpu) {
+       for_each_gcwq_cpu(cpu) {
+               struct workqueue_struct *wq;
                /*
                 * nr_active is monotonically decreasing.  It's safe
                 * to peek without lock.
@@ -2211,7 +3700,7 @@ bool freeze_workqueues_busy(void)
                list_for_each_entry(wq, &workqueues, list) {
                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
 
-                       if (!(wq->flags & WQ_FREEZEABLE))
+                       if (!cwq || !(wq->flags & WQ_FREEZABLE))
                                continue;
 
                        BUG_ON(cwq->nr_active < 0);
@@ -2230,14 +3719,13 @@ out_unlock:
  * thaw_workqueues - thaw workqueues
  *
  * Thaw workqueues.  Normal queueing is restored and all collected
- * frozen works are transferred to their respective cwq worklists.
+ * frozen works are transferred to their respective gcwq worklists.
  *
  * CONTEXT:
  * Grabs and releases workqueue_lock and gcwq->lock's.
  */
 void thaw_workqueues(void)
 {
-       struct workqueue_struct *wq;
        unsigned int cpu;
 
        spin_lock(&workqueue_lock);
@@ -2245,8 +3733,9 @@ void thaw_workqueues(void)
        if (!workqueue_freezing)
                goto out_unlock;
 
-       for_each_possible_cpu(cpu) {
+       for_each_gcwq_cpu(cpu) {
                struct global_cwq *gcwq = get_gcwq(cpu);
+               struct workqueue_struct *wq;
 
                spin_lock_irq(&gcwq->lock);
 
@@ -2256,7 +3745,7 @@ void thaw_workqueues(void)
                list_for_each_entry(wq, &workqueues, list) {
                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
 
-                       if (!(wq->flags & WQ_FREEZEABLE))
+                       if (!cwq || !(wq->flags & WQ_FREEZABLE))
                                continue;
 
                        /* restore max_active and repopulate worklist */
@@ -2265,10 +3754,10 @@ void thaw_workqueues(void)
                        while (!list_empty(&cwq->delayed_works) &&
                               cwq->nr_active < cwq->max_active)
                                cwq_activate_first_delayed(cwq);
-
-                       wake_up_process(cwq->worker->task);
                }
 
+               wake_up_worker(gcwq);
+
                spin_unlock_irq(&gcwq->lock);
        }
 
@@ -2278,31 +3767,62 @@ out_unlock:
 }
 #endif /* CONFIG_FREEZER */
 
-void __init init_workqueues(void)
+static int __init init_workqueues(void)
 {
        unsigned int cpu;
        int i;
 
-       singlethread_cpu = cpumask_first(cpu_possible_mask);
-       hotcpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
+       cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
 
        /* initialize gcwqs */
-       for_each_possible_cpu(cpu) {
+       for_each_gcwq_cpu(cpu) {
                struct global_cwq *gcwq = get_gcwq(cpu);
 
                spin_lock_init(&gcwq->lock);
+               INIT_LIST_HEAD(&gcwq->worklist);
                gcwq->cpu = cpu;
+               gcwq->flags |= GCWQ_DISASSOCIATED;
 
                INIT_LIST_HEAD(&gcwq->idle_list);
                for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
                        INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
 
+               init_timer_deferrable(&gcwq->idle_timer);
+               gcwq->idle_timer.function = idle_worker_timeout;
+               gcwq->idle_timer.data = (unsigned long)gcwq;
+
+               setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
+                           (unsigned long)gcwq);
+
                ida_init(&gcwq->worker_ida);
 
                gcwq->trustee_state = TRUSTEE_DONE;
                init_waitqueue_head(&gcwq->trustee_wait);
        }
 
-       keventd_wq = create_workqueue("events");
-       BUG_ON(!keventd_wq);
+       /* create the initial worker */
+       for_each_online_gcwq_cpu(cpu) {
+               struct global_cwq *gcwq = get_gcwq(cpu);
+               struct worker *worker;
+
+               if (cpu != WORK_CPU_UNBOUND)
+                       gcwq->flags &= ~GCWQ_DISASSOCIATED;
+               worker = create_worker(gcwq, true);
+               BUG_ON(!worker);
+               spin_lock_irq(&gcwq->lock);
+               start_worker(worker);
+               spin_unlock_irq(&gcwq->lock);
+       }
+
+       system_wq = alloc_workqueue("events", 0, 0);
+       system_long_wq = alloc_workqueue("events_long", 0, 0);
+       system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
+       system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
+                                           WQ_UNBOUND_MAX_ACTIVE);
+       system_freezable_wq = alloc_workqueue("events_freezable",
+                                             WQ_FREEZABLE, 0);
+       BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
+              !system_unbound_wq || !system_freezable_wq);
+       return 0;
 }
+early_initcall(init_workqueues);