~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/workqueue.c

Version: ~ [ linux-5.16 ] ~ [ linux-5.15.13 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.90 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.170 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.224 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.261 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.296 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.298 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * kernel/workqueue.c - generic async execution with shared worker pool
  3  *
  4  * Copyright (C) 2002           Ingo Molnar
  5  *
  6  *   Derived from the taskqueue/keventd code by:
  7  *     David Woodhouse <dwmw2@infradead.org>
  8  *     Andrew Morton
  9  *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
 10  *     Theodore Ts'o <tytso@mit.edu>
 11  *
 12  * Made to use alloc_percpu by Christoph Lameter.
 13  *
 14  * Copyright (C) 2010           SUSE Linux Products GmbH
 15  * Copyright (C) 2010           Tejun Heo <tj@kernel.org>
 16  *
 17  * This is the generic async execution mechanism.  Work items as are
 18  * executed in process context.  The worker pool is shared and
 19  * automatically managed.  There are two worker pools for each CPU (one for
 20  * normal work items and the other for high priority ones) and some extra
 21  * pools for workqueues which are not bound to any specific CPU - the
 22  * number of these backing pools is dynamic.
 23  *
 24  * Please read Documentation/workqueue.txt for details.
 25  */
 26 
 27 #include <linux/export.h>
 28 #include <linux/kernel.h>
 29 #include <linux/sched.h>
 30 #include <linux/init.h>
 31 #include <linux/signal.h>
 32 #include <linux/completion.h>
 33 #include <linux/workqueue.h>
 34 #include <linux/slab.h>
 35 #include <linux/cpu.h>
 36 #include <linux/notifier.h>
 37 #include <linux/kthread.h>
 38 #include <linux/hardirq.h>
 39 #include <linux/mempolicy.h>
 40 #include <linux/freezer.h>
 41 #include <linux/kallsyms.h>
 42 #include <linux/debug_locks.h>
 43 #include <linux/lockdep.h>
 44 #include <linux/idr.h>
 45 #include <linux/jhash.h>
 46 #include <linux/hashtable.h>
 47 #include <linux/rculist.h>
 48 #include <linux/nodemask.h>
 49 #include <linux/moduleparam.h>
 50 #include <linux/uaccess.h>
 51 
 52 #include "workqueue_internal.h"
 53 
 54 enum {
 55         /*
 56          * worker_pool flags
 57          *
 58          * A bound pool is either associated or disassociated with its CPU.
 59          * While associated (!DISASSOCIATED), all workers are bound to the
 60          * CPU and none has %WORKER_UNBOUND set and concurrency management
 61          * is in effect.
 62          *
 63          * While DISASSOCIATED, the cpu may be offline and all workers have
 64          * %WORKER_UNBOUND set and concurrency management disabled, and may
 65          * be executing on any CPU.  The pool behaves as an unbound one.
 66          *
 67          * Note that DISASSOCIATED should be flipped only while holding
 68          * attach_mutex to avoid changing binding state while
 69          * worker_attach_to_pool() is in progress.
 70          */
 71         POOL_MANAGER_ACTIVE     = 1 << 0,       /* being managed */
 72         POOL_DISASSOCIATED      = 1 << 2,       /* cpu can't serve workers */
 73 
 74         /* worker flags */
 75         WORKER_DIE              = 1 << 1,       /* die die die */
 76         WORKER_IDLE             = 1 << 2,       /* is idle */
 77         WORKER_PREP             = 1 << 3,       /* preparing to run works */
 78         WORKER_CPU_INTENSIVE    = 1 << 6,       /* cpu intensive */
 79         WORKER_UNBOUND          = 1 << 7,       /* worker is unbound */
 80         WORKER_REBOUND          = 1 << 8,       /* worker was rebound */
 81 
 82         WORKER_NOT_RUNNING      = WORKER_PREP | WORKER_CPU_INTENSIVE |
 83                                   WORKER_UNBOUND | WORKER_REBOUND,
 84 
 85         NR_STD_WORKER_POOLS     = 2,            /* # standard pools per cpu */
 86 
 87         UNBOUND_POOL_HASH_ORDER = 6,            /* hashed by pool->attrs */
 88         BUSY_WORKER_HASH_ORDER  = 6,            /* 64 pointers */
 89 
 90         MAX_IDLE_WORKERS_RATIO  = 4,            /* 1/4 of busy can be idle */
 91         IDLE_WORKER_TIMEOUT     = 300 * HZ,     /* keep idle ones for 5 mins */
 92 
 93         MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
 94                                                 /* call for help after 10ms
 95                                                    (min two ticks) */
 96         MAYDAY_INTERVAL         = HZ / 10,      /* and then every 100ms */
 97         CREATE_COOLDOWN         = HZ,           /* time to breath after fail */
 98 
 99         /*
100          * Rescue workers are used only on emergencies and shared by
101          * all cpus.  Give MIN_NICE.
102          */
103         RESCUER_NICE_LEVEL      = MIN_NICE,
104         HIGHPRI_NICE_LEVEL      = MIN_NICE,
105 
106         WQ_NAME_LEN             = 24,
107 };
108 
109 /*
110  * Structure fields follow one of the following exclusion rules.
111  *
112  * I: Modifiable by initialization/destruction paths and read-only for
113  *    everyone else.
114  *
115  * P: Preemption protected.  Disabling preemption is enough and should
116  *    only be modified and accessed from the local cpu.
117  *
118  * L: pool->lock protected.  Access with pool->lock held.
119  *
120  * X: During normal operation, modification requires pool->lock and should
121  *    be done only from local cpu.  Either disabling preemption on local
122  *    cpu or grabbing pool->lock is enough for read access.  If
123  *    POOL_DISASSOCIATED is set, it's identical to L.
124  *
125  * A: pool->attach_mutex protected.
126  *
127  * PL: wq_pool_mutex protected.
128  *
129  * PR: wq_pool_mutex protected for writes.  Sched-RCU protected for reads.
130  *
131  * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
132  *
133  * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
134  *      sched-RCU for reads.
135  *
136  * WQ: wq->mutex protected.
137  *
138  * WR: wq->mutex protected for writes.  Sched-RCU protected for reads.
139  *
140  * MD: wq_mayday_lock protected.
141  */
142 
143 /* struct worker is defined in workqueue_internal.h */
144 
145 struct worker_pool {
146         spinlock_t              lock;           /* the pool lock */
147         int                     cpu;            /* I: the associated cpu */
148         int                     node;           /* I: the associated node ID */
149         int                     id;             /* I: pool ID */
150         unsigned int            flags;          /* X: flags */
151 
152         unsigned long           watchdog_ts;    /* L: watchdog timestamp */
153 
154         struct list_head        worklist;       /* L: list of pending works */
155         int                     nr_workers;     /* L: total number of workers */
156 
157         /* nr_idle includes the ones off idle_list for rebinding */
158         int                     nr_idle;        /* L: currently idle ones */
159 
160         struct list_head        idle_list;      /* X: list of idle workers */
161         struct timer_list       idle_timer;     /* L: worker idle timeout */
162         struct timer_list       mayday_timer;   /* L: SOS timer for workers */
163 
164         /* a workers is either on busy_hash or idle_list, or the manager */
165         DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
166                                                 /* L: hash of busy workers */
167 
168         /* see manage_workers() for details on the two manager mutexes */
169         struct worker           *manager;       /* L: purely informational */
170         struct mutex            attach_mutex;   /* attach/detach exclusion */
171         struct list_head        workers;        /* A: attached workers */
172         struct completion       *detach_completion; /* all workers detached */
173 
174         struct ida              worker_ida;     /* worker IDs for task name */
175 
176         struct workqueue_attrs  *attrs;         /* I: worker attributes */
177         struct hlist_node       hash_node;      /* PL: unbound_pool_hash node */
178         int                     refcnt;         /* PL: refcnt for unbound pools */
179 
180         /*
181          * The current concurrency level.  As it's likely to be accessed
182          * from other CPUs during try_to_wake_up(), put it in a separate
183          * cacheline.
184          */
185         atomic_t                nr_running ____cacheline_aligned_in_smp;
186 
187         /*
188          * Destruction of pool is sched-RCU protected to allow dereferences
189          * from get_work_pool().
190          */
191         struct rcu_head         rcu;
192 } ____cacheline_aligned_in_smp;
193 
194 /*
195  * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
196  * of work_struct->data are used for flags and the remaining high bits
197  * point to the pwq; thus, pwqs need to be aligned at two's power of the
198  * number of flag bits.
199  */
200 struct pool_workqueue {
201         struct worker_pool      *pool;          /* I: the associated pool */
202         struct workqueue_struct *wq;            /* I: the owning workqueue */
203         int                     work_color;     /* L: current color */
204         int                     flush_color;    /* L: flushing color */
205         int                     refcnt;         /* L: reference count */
206         int                     nr_in_flight[WORK_NR_COLORS];
207                                                 /* L: nr of in_flight works */
208         int                     nr_active;      /* L: nr of active works */
209         int                     max_active;     /* L: max active works */
210         struct list_head        delayed_works;  /* L: delayed works */
211         struct list_head        pwqs_node;      /* WR: node on wq->pwqs */
212         struct list_head        mayday_node;    /* MD: node on wq->maydays */
213 
214         /*
215          * Release of unbound pwq is punted to system_wq.  See put_pwq()
216          * and pwq_unbound_release_workfn() for details.  pool_workqueue
217          * itself is also sched-RCU protected so that the first pwq can be
218          * determined without grabbing wq->mutex.
219          */
220         struct work_struct      unbound_release_work;
221         struct rcu_head         rcu;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
223 
224 /*
225  * Structure used to wait for workqueue flush.
226  */
227 struct wq_flusher {
228         struct list_head        list;           /* WQ: list of flushers */
229         int                     flush_color;    /* WQ: flush color waiting for */
230         struct completion       done;           /* flush completion */
231 };
232 
233 struct wq_device;
234 
235 /*
236  * The externally visible workqueue.  It relays the issued work items to
237  * the appropriate worker_pool through its pool_workqueues.
238  */
239 struct workqueue_struct {
240         struct list_head        pwqs;           /* WR: all pwqs of this wq */
241         struct list_head        list;           /* PR: list of all workqueues */
242 
243         struct mutex            mutex;          /* protects this wq */
244         int                     work_color;     /* WQ: current work color */
245         int                     flush_color;    /* WQ: current flush color */
246         atomic_t                nr_pwqs_to_flush; /* flush in progress */
247         struct wq_flusher       *first_flusher; /* WQ: first flusher */
248         struct list_head        flusher_queue;  /* WQ: flush waiters */
249         struct list_head        flusher_overflow; /* WQ: flush overflow list */
250 
251         struct list_head        maydays;        /* MD: pwqs requesting rescue */
252         struct worker           *rescuer;       /* I: rescue worker */
253 
254         int                     nr_drainers;    /* WQ: drain in progress */
255         int                     saved_max_active; /* WQ: saved pwq max_active */
256 
257         struct workqueue_attrs  *unbound_attrs; /* PW: only for unbound wqs */
258         struct pool_workqueue   *dfl_pwq;       /* PW: only for unbound wqs */
259 
260 #ifdef CONFIG_SYSFS
261         struct wq_device        *wq_dev;        /* I: for sysfs interface */
262 #endif
263 #ifdef CONFIG_LOCKDEP
264         struct lockdep_map      lockdep_map;
265 #endif
266         char                    name[WQ_NAME_LEN]; /* I: workqueue name */
267 
268         /*
269          * Destruction of workqueue_struct is sched-RCU protected to allow
270          * walking the workqueues list without grabbing wq_pool_mutex.
271          * This is used to dump all workqueues from sysrq.
272          */
273         struct rcu_head         rcu;
274 
275         /* hot fields used during command issue, aligned to cacheline */
276         unsigned int            flags ____cacheline_aligned; /* WQ: WQ_* flags */
277         struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278         struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279 };
280 
281 static struct kmem_cache *pwq_cache;
282 
283 static cpumask_var_t *wq_numa_possible_cpumask;
284                                         /* possible CPUs of each node */
285 
286 static bool wq_disable_numa;
287 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288 
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292 
293 static bool wq_online;                  /* can kworkers be created yet? */
294 
295 static bool wq_numa_enabled;            /* unbound NUMA affinity enabled */
296 
297 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
299 
300 static DEFINE_MUTEX(wq_pool_mutex);     /* protects pools and workqueues list */
301 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
302 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
303 
304 static LIST_HEAD(workqueues);           /* PR: list of all workqueues */
305 static bool workqueue_freezing;         /* PL: have wqs started freezing? */
306 
307 /* PL: allowable cpus for unbound wqs and work items */
308 static cpumask_var_t wq_unbound_cpumask;
309 
310 /* CPU where unbound work was last round robin scheduled from this CPU */
311 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
312 
313 /*
314  * Local execution of unbound work items is no longer guaranteed.  The
315  * following always forces round-robin CPU selection on unbound work items
316  * to uncover usages which depend on it.
317  */
318 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
319 static bool wq_debug_force_rr_cpu = true;
320 #else
321 static bool wq_debug_force_rr_cpu = false;
322 #endif
323 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
324 
325 /* the per-cpu worker pools */
326 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
327 
328 static DEFINE_IDR(worker_pool_idr);     /* PR: idr of all pools */
329 
330 /* PL: hash of all unbound pools keyed by pool->attrs */
331 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
332 
333 /* I: attributes used when instantiating standard unbound pools on demand */
334 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
335 
336 /* I: attributes used when instantiating ordered pools on demand */
337 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
338 
339 struct workqueue_struct *system_wq __read_mostly;
340 EXPORT_SYMBOL(system_wq);
341 struct workqueue_struct *system_highpri_wq __read_mostly;
342 EXPORT_SYMBOL_GPL(system_highpri_wq);
343 struct workqueue_struct *system_long_wq __read_mostly;
344 EXPORT_SYMBOL_GPL(system_long_wq);
345 struct workqueue_struct *system_unbound_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_unbound_wq);
347 struct workqueue_struct *system_freezable_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_freezable_wq);
349 struct workqueue_struct *system_power_efficient_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
351 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
352 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
353 
354 static int worker_thread(void *__worker);
355 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
356 
357 #define CREATE_TRACE_POINTS
358 #include <trace/events/workqueue.h>
359 
360 #define assert_rcu_or_pool_mutex()                                      \
361         RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&                 \
362                          !lockdep_is_held(&wq_pool_mutex),              \
363                          "sched RCU or wq_pool_mutex should be held")
364 
365 #define assert_rcu_or_wq_mutex(wq)                                      \
366         RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&                 \
367                          !lockdep_is_held(&wq->mutex),                  \
368                          "sched RCU or wq->mutex should be held")
369 
370 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq)                        \
371         RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&                 \
372                          !lockdep_is_held(&wq->mutex) &&                \
373                          !lockdep_is_held(&wq_pool_mutex),              \
374                          "sched RCU, wq->mutex or wq_pool_mutex should be held")
375 
376 #define for_each_cpu_worker_pool(pool, cpu)                             \
377         for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];               \
378              (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
379              (pool)++)
380 
381 /**
382  * for_each_pool - iterate through all worker_pools in the system
383  * @pool: iteration cursor
384  * @pi: integer used for iteration
385  *
386  * This must be called either with wq_pool_mutex held or sched RCU read
387  * locked.  If the pool needs to be used beyond the locking in effect, the
388  * caller is responsible for guaranteeing that the pool stays online.
389  *
390  * The if/else clause exists only for the lockdep assertion and can be
391  * ignored.
392  */
393 #define for_each_pool(pool, pi)                                         \
394         idr_for_each_entry(&worker_pool_idr, pool, pi)                  \
395                 if (({ assert_rcu_or_pool_mutex(); false; })) { }       \
396                 else
397 
398 /**
399  * for_each_pool_worker - iterate through all workers of a worker_pool
400  * @worker: iteration cursor
401  * @pool: worker_pool to iterate workers of
402  *
403  * This must be called with @pool->attach_mutex.
404  *
405  * The if/else clause exists only for the lockdep assertion and can be
406  * ignored.
407  */
408 #define for_each_pool_worker(worker, pool)                              \
409         list_for_each_entry((worker), &(pool)->workers, node)           \
410                 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
411                 else
412 
413 /**
414  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
415  * @pwq: iteration cursor
416  * @wq: the target workqueue
417  *
418  * This must be called either with wq->mutex held or sched RCU read locked.
419  * If the pwq needs to be used beyond the locking in effect, the caller is
420  * responsible for guaranteeing that the pwq stays online.
421  *
422  * The if/else clause exists only for the lockdep assertion and can be
423  * ignored.
424  */
425 #define for_each_pwq(pwq, wq)                                           \
426         list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)          \
427                 if (({ assert_rcu_or_wq_mutex(wq); false; })) { }       \
428                 else
429 
430 #ifdef CONFIG_DEBUG_OBJECTS_WORK
431 
432 static struct debug_obj_descr work_debug_descr;
433 
434 static void *work_debug_hint(void *addr)
435 {
436         return ((struct work_struct *) addr)->func;
437 }
438 
439 static bool work_is_static_object(void *addr)
440 {
441         struct work_struct *work = addr;
442 
443         return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
444 }
445 
446 /*
447  * fixup_init is called when:
448  * - an active object is initialized
449  */
450 static bool work_fixup_init(void *addr, enum debug_obj_state state)
451 {
452         struct work_struct *work = addr;
453 
454         switch (state) {
455         case ODEBUG_STATE_ACTIVE:
456                 cancel_work_sync(work);
457                 debug_object_init(work, &work_debug_descr);
458                 return true;
459         default:
460                 return false;
461         }
462 }
463 
464 /*
465  * fixup_free is called when:
466  * - an active object is freed
467  */
468 static bool work_fixup_free(void *addr, enum debug_obj_state state)
469 {
470         struct work_struct *work = addr;
471 
472         switch (state) {
473         case ODEBUG_STATE_ACTIVE:
474                 cancel_work_sync(work);
475                 debug_object_free(work, &work_debug_descr);
476                 return true;
477         default:
478                 return false;
479         }
480 }
481 
482 static struct debug_obj_descr work_debug_descr = {
483         .name           = "work_struct",
484         .debug_hint     = work_debug_hint,
485         .is_static_object = work_is_static_object,
486         .fixup_init     = work_fixup_init,
487         .fixup_free     = work_fixup_free,
488 };
489 
490 static inline void debug_work_activate(struct work_struct *work)
491 {
492         debug_object_activate(work, &work_debug_descr);
493 }
494 
495 static inline void debug_work_deactivate(struct work_struct *work)
496 {
497         debug_object_deactivate(work, &work_debug_descr);
498 }
499 
500 void __init_work(struct work_struct *work, int onstack)
501 {
502         if (onstack)
503                 debug_object_init_on_stack(work, &work_debug_descr);
504         else
505                 debug_object_init(work, &work_debug_descr);
506 }
507 EXPORT_SYMBOL_GPL(__init_work);
508 
509 void destroy_work_on_stack(struct work_struct *work)
510 {
511         debug_object_free(work, &work_debug_descr);
512 }
513 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
514 
515 void destroy_delayed_work_on_stack(struct delayed_work *work)
516 {
517         destroy_timer_on_stack(&work->timer);
518         debug_object_free(&work->work, &work_debug_descr);
519 }
520 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
521 
522 #else
523 static inline void debug_work_activate(struct work_struct *work) { }
524 static inline void debug_work_deactivate(struct work_struct *work) { }
525 #endif
526 
527 /**
528  * worker_pool_assign_id - allocate ID and assing it to @pool
529  * @pool: the pool pointer of interest
530  *
531  * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
532  * successfully, -errno on failure.
533  */
534 static int worker_pool_assign_id(struct worker_pool *pool)
535 {
536         int ret;
537 
538         lockdep_assert_held(&wq_pool_mutex);
539 
540         ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
541                         GFP_KERNEL);
542         if (ret >= 0) {
543                 pool->id = ret;
544                 return 0;
545         }
546         return ret;
547 }
548 
549 /**
550  * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
551  * @wq: the target workqueue
552  * @node: the node ID
553  *
554  * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
555  * read locked.
556  * If the pwq needs to be used beyond the locking in effect, the caller is
557  * responsible for guaranteeing that the pwq stays online.
558  *
559  * Return: The unbound pool_workqueue for @node.
560  */
561 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
562                                                   int node)
563 {
564         assert_rcu_or_wq_mutex_or_pool_mutex(wq);
565 
566         /*
567          * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
568          * delayed item is pending.  The plan is to keep CPU -> NODE
569          * mapping valid and stable across CPU on/offlines.  Once that
570          * happens, this workaround can be removed.
571          */
572         if (unlikely(node == NUMA_NO_NODE))
573                 return wq->dfl_pwq;
574 
575         return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
576 }
577 
578 static unsigned int work_color_to_flags(int color)
579 {
580         return color << WORK_STRUCT_COLOR_SHIFT;
581 }
582 
583 static int get_work_color(struct work_struct *work)
584 {
585         return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
586                 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
587 }
588 
589 static int work_next_color(int color)
590 {
591         return (color + 1) % WORK_NR_COLORS;
592 }
593 
594 /*
595  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
596  * contain the pointer to the queued pwq.  Once execution starts, the flag
597  * is cleared and the high bits contain OFFQ flags and pool ID.
598  *
599  * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
600  * and clear_work_data() can be used to set the pwq, pool or clear
601  * work->data.  These functions should only be called while the work is
602  * owned - ie. while the PENDING bit is set.
603  *
604  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
605  * corresponding to a work.  Pool is available once the work has been
606  * queued anywhere after initialization until it is sync canceled.  pwq is
607  * available only while the work item is queued.
608  *
609  * %WORK_OFFQ_CANCELING is used to mark a work item which is being
610  * canceled.  While being canceled, a work item may have its PENDING set
611  * but stay off timer and worklist for arbitrarily long and nobody should
612  * try to steal the PENDING bit.
613  */
614 static inline void set_work_data(struct work_struct *work, unsigned long data,
615                                  unsigned long flags)
616 {
617         WARN_ON_ONCE(!work_pending(work));
618         atomic_long_set(&work->data, data | flags | work_static(work));
619 }
620 
621 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
622                          unsigned long extra_flags)
623 {
624         set_work_data(work, (unsigned long)pwq,
625                       WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
626 }
627 
628 static void set_work_pool_and_keep_pending(struct work_struct *work,
629                                            int pool_id)
630 {
631         set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
632                       WORK_STRUCT_PENDING);
633 }
634 
635 static void set_work_pool_and_clear_pending(struct work_struct *work,
636                                             int pool_id)
637 {
638         /*
639          * The following wmb is paired with the implied mb in
640          * test_and_set_bit(PENDING) and ensures all updates to @work made
641          * here are visible to and precede any updates by the next PENDING
642          * owner.
643          */
644         smp_wmb();
645         set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
646         /*
647          * The following mb guarantees that previous clear of a PENDING bit
648          * will not be reordered with any speculative LOADS or STORES from
649          * work->current_func, which is executed afterwards.  This possible
650          * reordering can lead to a missed execution on attempt to qeueue
651          * the same @work.  E.g. consider this case:
652          *
653          *   CPU#0                         CPU#1
654          *   ----------------------------  --------------------------------
655          *
656          * 1  STORE event_indicated
657          * 2  queue_work_on() {
658          * 3    test_and_set_bit(PENDING)
659          * 4 }                             set_..._and_clear_pending() {
660          * 5                                 set_work_data() # clear bit
661          * 6                                 smp_mb()
662          * 7                               work->current_func() {
663          * 8                                  LOAD event_indicated
664          *                                 }
665          *
666          * Without an explicit full barrier speculative LOAD on line 8 can
667          * be executed before CPU#0 does STORE on line 1.  If that happens,
668          * CPU#0 observes the PENDING bit is still set and new execution of
669          * a @work is not queued in a hope, that CPU#1 will eventually
670          * finish the queued @work.  Meanwhile CPU#1 does not see
671          * event_indicated is set, because speculative LOAD was executed
672          * before actual STORE.
673          */
674         smp_mb();
675 }
676 
677 static void clear_work_data(struct work_struct *work)
678 {
679         smp_wmb();      /* see set_work_pool_and_clear_pending() */
680         set_work_data(work, WORK_STRUCT_NO_POOL, 0);
681 }
682 
683 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
684 {
685         unsigned long data = atomic_long_read(&work->data);
686 
687         if (data & WORK_STRUCT_PWQ)
688                 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
689         else
690                 return NULL;
691 }
692 
693 /**
694  * get_work_pool - return the worker_pool a given work was associated with
695  * @work: the work item of interest
696  *
697  * Pools are created and destroyed under wq_pool_mutex, and allows read
698  * access under sched-RCU read lock.  As such, this function should be
699  * called under wq_pool_mutex or with preemption disabled.
700  *
701  * All fields of the returned pool are accessible as long as the above
702  * mentioned locking is in effect.  If the returned pool needs to be used
703  * beyond the critical section, the caller is responsible for ensuring the
704  * returned pool is and stays online.
705  *
706  * Return: The worker_pool @work was last associated with.  %NULL if none.
707  */
708 static struct worker_pool *get_work_pool(struct work_struct *work)
709 {
710         unsigned long data = atomic_long_read(&work->data);
711         int pool_id;
712 
713         assert_rcu_or_pool_mutex();
714 
715         if (data & WORK_STRUCT_PWQ)
716                 return ((struct pool_workqueue *)
717                         (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
718 
719         pool_id = data >> WORK_OFFQ_POOL_SHIFT;
720         if (pool_id == WORK_OFFQ_POOL_NONE)
721                 return NULL;
722 
723         return idr_find(&worker_pool_idr, pool_id);
724 }
725 
726 /**
727  * get_work_pool_id - return the worker pool ID a given work is associated with
728  * @work: the work item of interest
729  *
730  * Return: The worker_pool ID @work was last associated with.
731  * %WORK_OFFQ_POOL_NONE if none.
732  */
733 static int get_work_pool_id(struct work_struct *work)
734 {
735         unsigned long data = atomic_long_read(&work->data);
736 
737         if (data & WORK_STRUCT_PWQ)
738                 return ((struct pool_workqueue *)
739                         (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
740 
741         return data >> WORK_OFFQ_POOL_SHIFT;
742 }
743 
744 static void mark_work_canceling(struct work_struct *work)
745 {
746         unsigned long pool_id = get_work_pool_id(work);
747 
748         pool_id <<= WORK_OFFQ_POOL_SHIFT;
749         set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
750 }
751 
752 static bool work_is_canceling(struct work_struct *work)
753 {
754         unsigned long data = atomic_long_read(&work->data);
755 
756         return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
757 }
758 
759 /*
760  * Policy functions.  These define the policies on how the global worker
761  * pools are managed.  Unless noted otherwise, these functions assume that
762  * they're being called with pool->lock held.
763  */
764 
765 static bool __need_more_worker(struct worker_pool *pool)
766 {
767         return !atomic_read(&pool->nr_running);
768 }
769 
770 /*
771  * Need to wake up a worker?  Called from anything but currently
772  * running workers.
773  *
774  * Note that, because unbound workers never contribute to nr_running, this
775  * function will always return %true for unbound pools as long as the
776  * worklist isn't empty.
777  */
778 static bool need_more_worker(struct worker_pool *pool)
779 {
780         return !list_empty(&pool->worklist) && __need_more_worker(pool);
781 }
782 
783 /* Can I start working?  Called from busy but !running workers. */
784 static bool may_start_working(struct worker_pool *pool)
785 {
786         return pool->nr_idle;
787 }
788 
789 /* Do I need to keep working?  Called from currently running workers. */
790 static bool keep_working(struct worker_pool *pool)
791 {
792         return !list_empty(&pool->worklist) &&
793                 atomic_read(&pool->nr_running) <= 1;
794 }
795 
796 /* Do we need a new worker?  Called from manager. */
797 static bool need_to_create_worker(struct worker_pool *pool)
798 {
799         return need_more_worker(pool) && !may_start_working(pool);
800 }
801 
802 /* Do we have too many workers and should some go away? */
803 static bool too_many_workers(struct worker_pool *pool)
804 {
805         bool managing = pool->flags & POOL_MANAGER_ACTIVE;
806         int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
807         int nr_busy = pool->nr_workers - nr_idle;
808 
809         return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
810 }
811 
812 /*
813  * Wake up functions.
814  */
815 
816 /* Return the first idle worker.  Safe with preemption disabled */
817 static struct worker *first_idle_worker(struct worker_pool *pool)
818 {
819         if (unlikely(list_empty(&pool->idle_list)))
820                 return NULL;
821 
822         return list_first_entry(&pool->idle_list, struct worker, entry);
823 }
824 
825 /**
826  * wake_up_worker - wake up an idle worker
827  * @pool: worker pool to wake worker from
828  *
829  * Wake up the first idle worker of @pool.
830  *
831  * CONTEXT:
832  * spin_lock_irq(pool->lock).
833  */
834 static void wake_up_worker(struct worker_pool *pool)
835 {
836         struct worker *worker = first_idle_worker(pool);
837 
838         if (likely(worker))
839                 wake_up_process(worker->task);
840 }
841 
842 /**
843  * wq_worker_waking_up - a worker is waking up
844  * @task: task waking up
845  * @cpu: CPU @task is waking up to
846  *
847  * This function is called during try_to_wake_up() when a worker is
848  * being awoken.
849  *
850  * CONTEXT:
851  * spin_lock_irq(rq->lock)
852  */
853 void wq_worker_waking_up(struct task_struct *task, int cpu)
854 {
855         struct worker *worker = kthread_data(task);
856 
857         if (!(worker->flags & WORKER_NOT_RUNNING)) {
858                 WARN_ON_ONCE(worker->pool->cpu != cpu);
859                 atomic_inc(&worker->pool->nr_running);
860         }
861 }
862 
863 /**
864  * wq_worker_sleeping - a worker is going to sleep
865  * @task: task going to sleep
866  *
867  * This function is called during schedule() when a busy worker is
868  * going to sleep.  Worker on the same cpu can be woken up by
869  * returning pointer to its task.
870  *
871  * CONTEXT:
872  * spin_lock_irq(rq->lock)
873  *
874  * Return:
875  * Worker task on @cpu to wake up, %NULL if none.
876  */
877 struct task_struct *wq_worker_sleeping(struct task_struct *task)
878 {
879         struct worker *worker = kthread_data(task), *to_wakeup = NULL;
880         struct worker_pool *pool;
881 
882         /*
883          * Rescuers, which may not have all the fields set up like normal
884          * workers, also reach here, let's not access anything before
885          * checking NOT_RUNNING.
886          */
887         if (worker->flags & WORKER_NOT_RUNNING)
888                 return NULL;
889 
890         pool = worker->pool;
891 
892         /* this can only happen on the local cpu */
893         if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
894                 return NULL;
895 
896         /*
897          * The counterpart of the following dec_and_test, implied mb,
898          * worklist not empty test sequence is in insert_work().
899          * Please read comment there.
900          *
901          * NOT_RUNNING is clear.  This means that we're bound to and
902          * running on the local cpu w/ rq lock held and preemption
903          * disabled, which in turn means that none else could be
904          * manipulating idle_list, so dereferencing idle_list without pool
905          * lock is safe.
906          */
907         if (atomic_dec_and_test(&pool->nr_running) &&
908             !list_empty(&pool->worklist))
909                 to_wakeup = first_idle_worker(pool);
910         return to_wakeup ? to_wakeup->task : NULL;
911 }
912 
913 /**
914  * worker_set_flags - set worker flags and adjust nr_running accordingly
915  * @worker: self
916  * @flags: flags to set
917  *
918  * Set @flags in @worker->flags and adjust nr_running accordingly.
919  *
920  * CONTEXT:
921  * spin_lock_irq(pool->lock)
922  */
923 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
924 {
925         struct worker_pool *pool = worker->pool;
926 
927         WARN_ON_ONCE(worker->task != current);
928 
929         /* If transitioning into NOT_RUNNING, adjust nr_running. */
930         if ((flags & WORKER_NOT_RUNNING) &&
931             !(worker->flags & WORKER_NOT_RUNNING)) {
932                 atomic_dec(&pool->nr_running);
933         }
934 
935         worker->flags |= flags;
936 }
937 
938 /**
939  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
940  * @worker: self
941  * @flags: flags to clear
942  *
943  * Clear @flags in @worker->flags and adjust nr_running accordingly.
944  *
945  * CONTEXT:
946  * spin_lock_irq(pool->lock)
947  */
948 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
949 {
950         struct worker_pool *pool = worker->pool;
951         unsigned int oflags = worker->flags;
952 
953         WARN_ON_ONCE(worker->task != current);
954 
955         worker->flags &= ~flags;
956 
957         /*
958          * If transitioning out of NOT_RUNNING, increment nr_running.  Note
959          * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
960          * of multiple flags, not a single flag.
961          */
962         if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
963                 if (!(worker->flags & WORKER_NOT_RUNNING))
964                         atomic_inc(&pool->nr_running);
965 }
966 
967 /**
968  * find_worker_executing_work - find worker which is executing a work
969  * @pool: pool of interest
970  * @work: work to find worker for
971  *
972  * Find a worker which is executing @work on @pool by searching
973  * @pool->busy_hash which is keyed by the address of @work.  For a worker
974  * to match, its current execution should match the address of @work and
975  * its work function.  This is to avoid unwanted dependency between
976  * unrelated work executions through a work item being recycled while still
977  * being executed.
978  *
979  * This is a bit tricky.  A work item may be freed once its execution
980  * starts and nothing prevents the freed area from being recycled for
981  * another work item.  If the same work item address ends up being reused
982  * before the original execution finishes, workqueue will identify the
983  * recycled work item as currently executing and make it wait until the
984  * current execution finishes, introducing an unwanted dependency.
985  *
986  * This function checks the work item address and work function to avoid
987  * false positives.  Note that this isn't complete as one may construct a
988  * work function which can introduce dependency onto itself through a
989  * recycled work item.  Well, if somebody wants to shoot oneself in the
990  * foot that badly, there's only so much we can do, and if such deadlock
991  * actually occurs, it should be easy to locate the culprit work function.
992  *
993  * CONTEXT:
994  * spin_lock_irq(pool->lock).
995  *
996  * Return:
997  * Pointer to worker which is executing @work if found, %NULL
998  * otherwise.
999  */
1000 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1001                                                  struct work_struct *work)
1002 {
1003         struct worker *worker;
1004 
1005         hash_for_each_possible(pool->busy_hash, worker, hentry,
1006                                (unsigned long)work)
1007                 if (worker->current_work == work &&
1008                     worker->current_func == work->func)
1009                         return worker;
1010 
1011         return NULL;
1012 }
1013 
1014 /**
1015  * move_linked_works - move linked works to a list
1016  * @work: start of series of works to be scheduled
1017  * @head: target list to append @work to
1018  * @nextp: out parameter for nested worklist walking
1019  *
1020  * Schedule linked works starting from @work to @head.  Work series to
1021  * be scheduled starts at @work and includes any consecutive work with
1022  * WORK_STRUCT_LINKED set in its predecessor.
1023  *
1024  * If @nextp is not NULL, it's updated to point to the next work of
1025  * the last scheduled work.  This allows move_linked_works() to be
1026  * nested inside outer list_for_each_entry_safe().
1027  *
1028  * CONTEXT:
1029  * spin_lock_irq(pool->lock).
1030  */
1031 static void move_linked_works(struct work_struct *work, struct list_head *head,
1032                               struct work_struct **nextp)
1033 {
1034         struct work_struct *n;
1035 
1036         /*
1037          * Linked worklist will always end before the end of the list,
1038          * use NULL for list head.
1039          */
1040         list_for_each_entry_safe_from(work, n, NULL, entry) {
1041                 list_move_tail(&work->entry, head);
1042                 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1043                         break;
1044         }
1045 
1046         /*
1047          * If we're already inside safe list traversal and have moved
1048          * multiple works to the scheduled queue, the next position
1049          * needs to be updated.
1050          */
1051         if (nextp)
1052                 *nextp = n;
1053 }
1054 
1055 /**
1056  * get_pwq - get an extra reference on the specified pool_workqueue
1057  * @pwq: pool_workqueue to get
1058  *
1059  * Obtain an extra reference on @pwq.  The caller should guarantee that
1060  * @pwq has positive refcnt and be holding the matching pool->lock.
1061  */
1062 static void get_pwq(struct pool_workqueue *pwq)
1063 {
1064         lockdep_assert_held(&pwq->pool->lock);
1065         WARN_ON_ONCE(pwq->refcnt <= 0);
1066         pwq->refcnt++;
1067 }
1068 
1069 /**
1070  * put_pwq - put a pool_workqueue reference
1071  * @pwq: pool_workqueue to put
1072  *
1073  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1074  * destruction.  The caller should be holding the matching pool->lock.
1075  */
1076 static void put_pwq(struct pool_workqueue *pwq)
1077 {
1078         lockdep_assert_held(&pwq->pool->lock);
1079         if (likely(--pwq->refcnt))
1080                 return;
1081         if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1082                 return;
1083         /*
1084          * @pwq can't be released under pool->lock, bounce to
1085          * pwq_unbound_release_workfn().  This never recurses on the same
1086          * pool->lock as this path is taken only for unbound workqueues and
1087          * the release work item is scheduled on a per-cpu workqueue.  To
1088          * avoid lockdep warning, unbound pool->locks are given lockdep
1089          * subclass of 1 in get_unbound_pool().
1090          */
1091         schedule_work(&pwq->unbound_release_work);
1092 }
1093 
1094 /**
1095  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1096  * @pwq: pool_workqueue to put (can be %NULL)
1097  *
1098  * put_pwq() with locking.  This function also allows %NULL @pwq.
1099  */
1100 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1101 {
1102         if (pwq) {
1103                 /*
1104                  * As both pwqs and pools are sched-RCU protected, the
1105                  * following lock operations are safe.
1106                  */
1107                 spin_lock_irq(&pwq->pool->lock);
1108                 put_pwq(pwq);
1109                 spin_unlock_irq(&pwq->pool->lock);
1110         }
1111 }
1112 
1113 static void pwq_activate_delayed_work(struct work_struct *work)
1114 {
1115         struct pool_workqueue *pwq = get_work_pwq(work);
1116 
1117         trace_workqueue_activate_work(work);
1118         if (list_empty(&pwq->pool->worklist))
1119                 pwq->pool->watchdog_ts = jiffies;
1120         move_linked_works(work, &pwq->pool->worklist, NULL);
1121         __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1122         pwq->nr_active++;
1123 }
1124 
1125 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1126 {
1127         struct work_struct *work = list_first_entry(&pwq->delayed_works,
1128                                                     struct work_struct, entry);
1129 
1130         pwq_activate_delayed_work(work);
1131 }
1132 
1133 /**
1134  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1135  * @pwq: pwq of interest
1136  * @color: color of work which left the queue
1137  *
1138  * A work either has completed or is removed from pending queue,
1139  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1140  *
1141  * CONTEXT:
1142  * spin_lock_irq(pool->lock).
1143  */
1144 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1145 {
1146         /* uncolored work items don't participate in flushing or nr_active */
1147         if (color == WORK_NO_COLOR)
1148                 goto out_put;
1149 
1150         pwq->nr_in_flight[color]--;
1151 
1152         pwq->nr_active--;
1153         if (!list_empty(&pwq->delayed_works)) {
1154                 /* one down, submit a delayed one */
1155                 if (pwq->nr_active < pwq->max_active)
1156                         pwq_activate_first_delayed(pwq);
1157         }
1158 
1159         /* is flush in progress and are we at the flushing tip? */
1160         if (likely(pwq->flush_color != color))
1161                 goto out_put;
1162 
1163         /* are there still in-flight works? */
1164         if (pwq->nr_in_flight[color])
1165                 goto out_put;
1166 
1167         /* this pwq is done, clear flush_color */
1168         pwq->flush_color = -1;
1169 
1170         /*
1171          * If this was the last pwq, wake up the first flusher.  It
1172          * will handle the rest.
1173          */
1174         if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1175                 complete(&pwq->wq->first_flusher->done);
1176 out_put:
1177         put_pwq(pwq);
1178 }
1179 
1180 /**
1181  * try_to_grab_pending - steal work item from worklist and disable irq
1182  * @work: work item to steal
1183  * @is_dwork: @work is a delayed_work
1184  * @flags: place to store irq state
1185  *
1186  * Try to grab PENDING bit of @work.  This function can handle @work in any
1187  * stable state - idle, on timer or on worklist.
1188  *
1189  * Return:
1190  *  1           if @work was pending and we successfully stole PENDING
1191  *  0           if @work was idle and we claimed PENDING
1192  *  -EAGAIN     if PENDING couldn't be grabbed at the moment, safe to busy-retry
1193  *  -ENOENT     if someone else is canceling @work, this state may persist
1194  *              for arbitrarily long
1195  *
1196  * Note:
1197  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1198  * interrupted while holding PENDING and @work off queue, irq must be
1199  * disabled on entry.  This, combined with delayed_work->timer being
1200  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1201  *
1202  * On successful return, >= 0, irq is disabled and the caller is
1203  * responsible for releasing it using local_irq_restore(*@flags).
1204  *
1205  * This function is safe to call from any context including IRQ handler.
1206  */
1207 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1208                                unsigned long *flags)
1209 {
1210         struct worker_pool *pool;
1211         struct pool_workqueue *pwq;
1212 
1213         local_irq_save(*flags);
1214 
1215         /* try to steal the timer if it exists */
1216         if (is_dwork) {
1217                 struct delayed_work *dwork = to_delayed_work(work);
1218 
1219                 /*
1220                  * dwork->timer is irqsafe.  If del_timer() fails, it's
1221                  * guaranteed that the timer is not queued anywhere and not
1222                  * running on the local CPU.
1223                  */
1224                 if (likely(del_timer(&dwork->timer)))
1225                         return 1;
1226         }
1227 
1228         /* try to claim PENDING the normal way */
1229         if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1230                 return 0;
1231 
1232         /*
1233          * The queueing is in progress, or it is already queued. Try to
1234          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1235          */
1236         pool = get_work_pool(work);
1237         if (!pool)
1238                 goto fail;
1239 
1240         spin_lock(&pool->lock);
1241         /*
1242          * work->data is guaranteed to point to pwq only while the work
1243          * item is queued on pwq->wq, and both updating work->data to point
1244          * to pwq on queueing and to pool on dequeueing are done under
1245          * pwq->pool->lock.  This in turn guarantees that, if work->data
1246          * points to pwq which is associated with a locked pool, the work
1247          * item is currently queued on that pool.
1248          */
1249         pwq = get_work_pwq(work);
1250         if (pwq && pwq->pool == pool) {
1251                 debug_work_deactivate(work);
1252 
1253                 /*
1254                  * A delayed work item cannot be grabbed directly because
1255                  * it might have linked NO_COLOR work items which, if left
1256                  * on the delayed_list, will confuse pwq->nr_active
1257                  * management later on and cause stall.  Make sure the work
1258                  * item is activated before grabbing.
1259                  */
1260                 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1261                         pwq_activate_delayed_work(work);
1262 
1263                 list_del_init(&work->entry);
1264                 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1265 
1266                 /* work->data points to pwq iff queued, point to pool */
1267                 set_work_pool_and_keep_pending(work, pool->id);
1268 
1269                 spin_unlock(&pool->lock);
1270                 return 1;
1271         }
1272         spin_unlock(&pool->lock);
1273 fail:
1274         local_irq_restore(*flags);
1275         if (work_is_canceling(work))
1276                 return -ENOENT;
1277         cpu_relax();
1278         return -EAGAIN;
1279 }
1280 
1281 /**
1282  * insert_work - insert a work into a pool
1283  * @pwq: pwq @work belongs to
1284  * @work: work to insert
1285  * @head: insertion point
1286  * @extra_flags: extra WORK_STRUCT_* flags to set
1287  *
1288  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1289  * work_struct flags.
1290  *
1291  * CONTEXT:
1292  * spin_lock_irq(pool->lock).
1293  */
1294 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1295                         struct list_head *head, unsigned int extra_flags)
1296 {
1297         struct worker_pool *pool = pwq->pool;
1298 
1299         /* we own @work, set data and link */
1300         set_work_pwq(work, pwq, extra_flags);
1301         list_add_tail(&work->entry, head);
1302         get_pwq(pwq);
1303 
1304         /*
1305          * Ensure either wq_worker_sleeping() sees the above
1306          * list_add_tail() or we see zero nr_running to avoid workers lying
1307          * around lazily while there are works to be processed.
1308          */
1309         smp_mb();
1310 
1311         if (__need_more_worker(pool))
1312                 wake_up_worker(pool);
1313 }
1314 
1315 /*
1316  * Test whether @work is being queued from another work executing on the
1317  * same workqueue.
1318  */
1319 static bool is_chained_work(struct workqueue_struct *wq)
1320 {
1321         struct worker *worker;
1322 
1323         worker = current_wq_worker();
1324         /*
1325          * Return %true iff I'm a worker execuing a work item on @wq.  If
1326          * I'm @worker, it's safe to dereference it without locking.
1327          */
1328         return worker && worker->current_pwq->wq == wq;
1329 }
1330 
1331 /*
1332  * When queueing an unbound work item to a wq, prefer local CPU if allowed
1333  * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
1334  * avoid perturbing sensitive tasks.
1335  */
1336 static int wq_select_unbound_cpu(int cpu)
1337 {
1338         static bool printed_dbg_warning;
1339         int new_cpu;
1340 
1341         if (likely(!wq_debug_force_rr_cpu)) {
1342                 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1343                         return cpu;
1344         } else if (!printed_dbg_warning) {
1345                 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1346                 printed_dbg_warning = true;
1347         }
1348 
1349         if (cpumask_empty(wq_unbound_cpumask))
1350                 return cpu;
1351 
1352         new_cpu = __this_cpu_read(wq_rr_cpu_last);
1353         new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1354         if (unlikely(new_cpu >= nr_cpu_ids)) {
1355                 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1356                 if (unlikely(new_cpu >= nr_cpu_ids))
1357                         return cpu;
1358         }
1359         __this_cpu_write(wq_rr_cpu_last, new_cpu);
1360 
1361         return new_cpu;
1362 }
1363 
1364 static void __queue_work(int cpu, struct workqueue_struct *wq,
1365                          struct work_struct *work)
1366 {
1367         struct pool_workqueue *pwq;
1368         struct worker_pool *last_pool;
1369         struct list_head *worklist;
1370         unsigned int work_flags;
1371         unsigned int req_cpu = cpu;
1372 
1373         /*
1374          * While a work item is PENDING && off queue, a task trying to
1375          * steal the PENDING will busy-loop waiting for it to either get
1376          * queued or lose PENDING.  Grabbing PENDING and queueing should
1377          * happen with IRQ disabled.
1378          */
1379         WARN_ON_ONCE(!irqs_disabled());
1380 
1381         debug_work_activate(work);
1382 
1383         /* if draining, only works from the same workqueue are allowed */
1384         if (unlikely(wq->flags & __WQ_DRAINING) &&
1385             WARN_ON_ONCE(!is_chained_work(wq)))
1386                 return;
1387 retry:
1388         if (req_cpu == WORK_CPU_UNBOUND)
1389                 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1390 
1391         /* pwq which will be used unless @work is executing elsewhere */
1392         if (!(wq->flags & WQ_UNBOUND))
1393                 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1394         else
1395                 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1396 
1397         /*
1398          * If @work was previously on a different pool, it might still be
1399          * running there, in which case the work needs to be queued on that
1400          * pool to guarantee non-reentrancy.
1401          */
1402         last_pool = get_work_pool(work);
1403         if (last_pool && last_pool != pwq->pool) {
1404                 struct worker *worker;
1405 
1406                 spin_lock(&last_pool->lock);
1407 
1408                 worker = find_worker_executing_work(last_pool, work);
1409 
1410                 if (worker && worker->current_pwq->wq == wq) {
1411                         pwq = worker->current_pwq;
1412                 } else {
1413                         /* meh... not running there, queue here */
1414                         spin_unlock(&last_pool->lock);
1415                         spin_lock(&pwq->pool->lock);
1416                 }
1417         } else {
1418                 spin_lock(&pwq->pool->lock);
1419         }
1420 
1421         /*
1422          * pwq is determined and locked.  For unbound pools, we could have
1423          * raced with pwq release and it could already be dead.  If its
1424          * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1425          * without another pwq replacing it in the numa_pwq_tbl or while
1426          * work items are executing on it, so the retrying is guaranteed to
1427          * make forward-progress.
1428          */
1429         if (unlikely(!pwq->refcnt)) {
1430                 if (wq->flags & WQ_UNBOUND) {
1431                         spin_unlock(&pwq->pool->lock);
1432                         cpu_relax();
1433                         goto retry;
1434                 }
1435                 /* oops */
1436                 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1437                           wq->name, cpu);
1438         }
1439 
1440         /* pwq determined, queue */
1441         trace_workqueue_queue_work(req_cpu, pwq, work);
1442 
1443         if (WARN_ON(!list_empty(&work->entry))) {
1444                 spin_unlock(&pwq->pool->lock);
1445                 return;
1446         }
1447 
1448         pwq->nr_in_flight[pwq->work_color]++;
1449         work_flags = work_color_to_flags(pwq->work_color);
1450 
1451         if (likely(pwq->nr_active < pwq->max_active)) {
1452                 trace_workqueue_activate_work(work);
1453                 pwq->nr_active++;
1454                 worklist = &pwq->pool->worklist;
1455                 if (list_empty(worklist))
1456                         pwq->pool->watchdog_ts = jiffies;
1457         } else {
1458                 work_flags |= WORK_STRUCT_DELAYED;
1459                 worklist = &pwq->delayed_works;
1460         }
1461 
1462         insert_work(pwq, work, worklist, work_flags);
1463 
1464         spin_unlock(&pwq->pool->lock);
1465 }
1466 
1467 /**
1468  * queue_work_on - queue work on specific cpu
1469  * @cpu: CPU number to execute work on
1470  * @wq: workqueue to use
1471  * @work: work to queue
1472  *
1473  * We queue the work to a specific CPU, the caller must ensure it
1474  * can't go away.
1475  *
1476  * Return: %false if @work was already on a queue, %true otherwise.
1477  */
1478 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1479                    struct work_struct *work)
1480 {
1481         bool ret = false;
1482         unsigned long flags;
1483 
1484         local_irq_save(flags);
1485 
1486         if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1487                 __queue_work(cpu, wq, work);
1488                 ret = true;
1489         }
1490 
1491         local_irq_restore(flags);
1492         return ret;
1493 }
1494 EXPORT_SYMBOL(queue_work_on);
1495 
1496 void delayed_work_timer_fn(unsigned long __data)
1497 {
1498         struct delayed_work *dwork = (struct delayed_work *)__data;
1499 
1500         /* should have been called from irqsafe timer with irq already off */
1501         __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1502 }
1503 EXPORT_SYMBOL(delayed_work_timer_fn);
1504 
1505 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1506                                 struct delayed_work *dwork, unsigned long delay)
1507 {
1508         struct timer_list *timer = &dwork->timer;
1509         struct work_struct *work = &dwork->work;
1510 
1511         WARN_ON_ONCE(!wq);
1512         WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1513                      timer->data != (unsigned long)dwork);
1514         WARN_ON_ONCE(timer_pending(timer));
1515         WARN_ON_ONCE(!list_empty(&work->entry));
1516 
1517         /*
1518          * If @delay is 0, queue @dwork->work immediately.  This is for
1519          * both optimization and correctness.  The earliest @timer can
1520          * expire is on the closest next tick and delayed_work users depend
1521          * on that there's no such delay when @delay is 0.
1522          */
1523         if (!delay) {
1524                 __queue_work(cpu, wq, &dwork->work);
1525                 return;
1526         }
1527 
1528         dwork->wq = wq;
1529         dwork->cpu = cpu;
1530         timer->expires = jiffies + delay;
1531 
1532         if (unlikely(cpu != WORK_CPU_UNBOUND))
1533                 add_timer_on(timer, cpu);
1534         else
1535                 add_timer(timer);
1536 }
1537 
1538 /**
1539  * queue_delayed_work_on - queue work on specific CPU after delay
1540  * @cpu: CPU number to execute work on
1541  * @wq: workqueue to use
1542  * @dwork: work to queue
1543  * @delay: number of jiffies to wait before queueing
1544  *
1545  * Return: %false if @work was already on a queue, %true otherwise.  If
1546  * @delay is zero and @dwork is idle, it will be scheduled for immediate
1547  * execution.
1548  */
1549 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1550                            struct delayed_work *dwork, unsigned long delay)
1551 {
1552         struct work_struct *work = &dwork->work;
1553         bool ret = false;
1554         unsigned long flags;
1555 
1556         /* read the comment in __queue_work() */
1557         local_irq_save(flags);
1558 
1559         if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1560                 __queue_delayed_work(cpu, wq, dwork, delay);
1561                 ret = true;
1562         }
1563 
1564         local_irq_restore(flags);
1565         return ret;
1566 }
1567 EXPORT_SYMBOL(queue_delayed_work_on);
1568 
1569 /**
1570  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1571  * @cpu: CPU number to execute work on
1572  * @wq: workqueue to use
1573  * @dwork: work to queue
1574  * @delay: number of jiffies to wait before queueing
1575  *
1576  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1577  * modify @dwork's timer so that it expires after @delay.  If @delay is
1578  * zero, @work is guaranteed to be scheduled immediately regardless of its
1579  * current state.
1580  *
1581  * Return: %false if @dwork was idle and queued, %true if @dwork was
1582  * pending and its timer was modified.
1583  *
1584  * This function is safe to call from any context including IRQ handler.
1585  * See try_to_grab_pending() for details.
1586  */
1587 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1588                          struct delayed_work *dwork, unsigned long delay)
1589 {
1590         unsigned long flags;
1591         int ret;
1592 
1593         do {
1594                 ret = try_to_grab_pending(&dwork->work, true, &flags);
1595         } while (unlikely(ret == -EAGAIN));
1596 
1597         if (likely(ret >= 0)) {
1598                 __queue_delayed_work(cpu, wq, dwork, delay);
1599                 local_irq_restore(flags);
1600         }
1601 
1602         /* -ENOENT from try_to_grab_pending() becomes %true */
1603         return ret;
1604 }
1605 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1606 
1607 /**
1608  * worker_enter_idle - enter idle state
1609  * @worker: worker which is entering idle state
1610  *
1611  * @worker is entering idle state.  Update stats and idle timer if
1612  * necessary.
1613  *
1614  * LOCKING:
1615  * spin_lock_irq(pool->lock).
1616  */
1617 static void worker_enter_idle(struct worker *worker)
1618 {
1619         struct worker_pool *pool = worker->pool;
1620 
1621         if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1622             WARN_ON_ONCE(!list_empty(&worker->entry) &&
1623                          (worker->hentry.next || worker->hentry.pprev)))
1624                 return;
1625 
1626         /* can't use worker_set_flags(), also called from create_worker() */
1627         worker->flags |= WORKER_IDLE;
1628         pool->nr_idle++;
1629         worker->last_active = jiffies;
1630 
1631         /* idle_list is LIFO */
1632         list_add(&worker->entry, &pool->idle_list);
1633 
1634         if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1635                 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1636 
1637         /*
1638          * Sanity check nr_running.  Because wq_unbind_fn() releases
1639          * pool->lock between setting %WORKER_UNBOUND and zapping
1640          * nr_running, the warning may trigger spuriously.  Check iff
1641          * unbind is not in progress.
1642          */
1643         WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1644                      pool->nr_workers == pool->nr_idle &&
1645                      atomic_read(&pool->nr_running));
1646 }
1647 
1648 /**
1649  * worker_leave_idle - leave idle state
1650  * @worker: worker which is leaving idle state
1651  *
1652  * @worker is leaving idle state.  Update stats.
1653  *
1654  * LOCKING:
1655  * spin_lock_irq(pool->lock).
1656  */
1657 static void worker_leave_idle(struct worker *worker)
1658 {
1659         struct worker_pool *pool = worker->pool;
1660 
1661         if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1662                 return;
1663         worker_clr_flags(worker, WORKER_IDLE);
1664         pool->nr_idle--;
1665         list_del_init(&worker->entry);
1666 }
1667 
1668 static struct worker *alloc_worker(int node)
1669 {
1670         struct worker *worker;
1671 
1672         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1673         if (worker) {
1674                 INIT_LIST_HEAD(&worker->entry);
1675                 INIT_LIST_HEAD(&worker->scheduled);
1676                 INIT_LIST_HEAD(&worker->node);
1677                 /* on creation a worker is in !idle && prep state */
1678                 worker->flags = WORKER_PREP;
1679         }
1680         return worker;
1681 }
1682 
1683 /**
1684  * worker_attach_to_pool() - attach a worker to a pool
1685  * @worker: worker to be attached
1686  * @pool: the target pool
1687  *
1688  * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
1689  * cpu-binding of @worker are kept coordinated with the pool across
1690  * cpu-[un]hotplugs.
1691  */
1692 static void worker_attach_to_pool(struct worker *worker,
1693                                    struct worker_pool *pool)
1694 {
1695         mutex_lock(&pool->attach_mutex);
1696 
1697         /*
1698          * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1699          * online CPUs.  It'll be re-applied when any of the CPUs come up.
1700          */
1701         set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1702 
1703         /*
1704          * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1705          * stable across this function.  See the comments above the
1706          * flag definition for details.
1707          */
1708         if (pool->flags & POOL_DISASSOCIATED)
1709                 worker->flags |= WORKER_UNBOUND;
1710 
1711         list_add_tail(&worker->node, &pool->workers);
1712 
1713         mutex_unlock(&pool->attach_mutex);
1714 }
1715 
1716 /**
1717  * worker_detach_from_pool() - detach a worker from its pool
1718  * @worker: worker which is attached to its pool
1719  * @pool: the pool @worker is attached to
1720  *
1721  * Undo the attaching which had been done in worker_attach_to_pool().  The
1722  * caller worker shouldn't access to the pool after detached except it has
1723  * other reference to the pool.
1724  */
1725 static void worker_detach_from_pool(struct worker *worker,
1726                                     struct worker_pool *pool)
1727 {
1728         struct completion *detach_completion = NULL;
1729 
1730         mutex_lock(&pool->attach_mutex);
1731         list_del(&worker->node);
1732         if (list_empty(&pool->workers))
1733                 detach_completion = pool->detach_completion;
1734         mutex_unlock(&pool->attach_mutex);
1735 
1736         /* clear leftover flags without pool->lock after it is detached */
1737         worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1738 
1739         if (detach_completion)
1740                 complete(detach_completion);
1741 }
1742 
1743 /**
1744  * create_worker - create a new workqueue worker
1745  * @pool: pool the new worker will belong to
1746  *
1747  * Create and start a new worker which is attached to @pool.
1748  *
1749  * CONTEXT:
1750  * Might sleep.  Does GFP_KERNEL allocations.
1751  *
1752  * Return:
1753  * Pointer to the newly created worker.
1754  */
1755 static struct worker *create_worker(struct worker_pool *pool)
1756 {
1757         struct worker *worker = NULL;
1758         int id = -1;
1759         char id_buf[16];
1760 
1761         /* ID is needed to determine kthread name */
1762         id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1763         if (id < 0)
1764                 goto fail;
1765 
1766         worker = alloc_worker(pool->node);
1767         if (!worker)
1768                 goto fail;
1769 
1770         worker->pool = pool;
1771         worker->id = id;
1772 
1773         if (pool->cpu >= 0)
1774                 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1775                          pool->attrs->nice < 0  ? "H" : "");
1776         else
1777                 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1778 
1779         worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1780                                               "kworker/%s", id_buf);
1781         if (IS_ERR(worker->task))
1782                 goto fail;
1783 
1784         set_user_nice(worker->task, pool->attrs->nice);
1785         kthread_bind_mask(worker->task, pool->attrs->cpumask);
1786 
1787         /* successful, attach the worker to the pool */
1788         worker_attach_to_pool(worker, pool);
1789 
1790         /* start the newly created worker */
1791         spin_lock_irq(&pool->lock);
1792         worker->pool->nr_workers++;
1793         worker_enter_idle(worker);
1794         wake_up_process(worker->task);
1795         spin_unlock_irq(&pool->lock);
1796 
1797         return worker;
1798 
1799 fail:
1800         if (id >= 0)
1801                 ida_simple_remove(&pool->worker_ida, id);
1802         kfree(worker);
1803         return NULL;
1804 }
1805 
1806 /**
1807  * destroy_worker - destroy a workqueue worker
1808  * @worker: worker to be destroyed
1809  *
1810  * Destroy @worker and adjust @pool stats accordingly.  The worker should
1811  * be idle.
1812  *
1813  * CONTEXT:
1814  * spin_lock_irq(pool->lock).
1815  */
1816 static void destroy_worker(struct worker *worker)
1817 {
1818         struct worker_pool *pool = worker->pool;
1819 
1820         lockdep_assert_held(&pool->lock);
1821 
1822         /* sanity check frenzy */
1823         if (WARN_ON(worker->current_work) ||
1824             WARN_ON(!list_empty(&worker->scheduled)) ||
1825             WARN_ON(!(worker->flags & WORKER_IDLE)))
1826                 return;
1827 
1828         pool->nr_workers--;
1829         pool->nr_idle--;
1830 
1831         list_del_init(&worker->entry);
1832         worker->flags |= WORKER_DIE;
1833         wake_up_process(worker->task);
1834 }
1835 
1836 static void idle_worker_timeout(unsigned long __pool)
1837 {
1838         struct worker_pool *pool = (void *)__pool;
1839 
1840         spin_lock_irq(&pool->lock);
1841 
1842         while (too_many_workers(pool)) {
1843                 struct worker *worker;
1844                 unsigned long expires;
1845 
1846                 /* idle_list is kept in LIFO order, check the last one */
1847                 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1848                 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1849 
1850                 if (time_before(jiffies, expires)) {
1851                         mod_timer(&pool->idle_timer, expires);
1852                         break;
1853                 }
1854 
1855                 destroy_worker(worker);
1856         }
1857 
1858         spin_unlock_irq(&pool->lock);
1859 }
1860 
1861 static void send_mayday(struct work_struct *work)
1862 {
1863         struct pool_workqueue *pwq = get_work_pwq(work);
1864         struct workqueue_struct *wq = pwq->wq;
1865 
1866         lockdep_assert_held(&wq_mayday_lock);
1867 
1868         if (!wq->rescuer)
1869                 return;
1870 
1871         /* mayday mayday mayday */
1872         if (list_empty(&pwq->mayday_node)) {
1873                 /*
1874                  * If @pwq is for an unbound wq, its base ref may be put at
1875                  * any time due to an attribute change.  Pin @pwq until the
1876                  * rescuer is done with it.
1877                  */
1878                 get_pwq(pwq);
1879                 list_add_tail(&pwq->mayday_node, &wq->maydays);
1880                 wake_up_process(wq->rescuer->task);
1881         }
1882 }
1883 
1884 static void pool_mayday_timeout(unsigned long __pool)
1885 {
1886         struct worker_pool *pool = (void *)__pool;
1887         struct work_struct *work;
1888 
1889         spin_lock_irq(&pool->lock);
1890         spin_lock(&wq_mayday_lock);             /* for wq->maydays */
1891 
1892         if (need_to_create_worker(pool)) {
1893                 /*
1894                  * We've been trying to create a new worker but
1895                  * haven't been successful.  We might be hitting an
1896                  * allocation deadlock.  Send distress signals to
1897                  * rescuers.
1898                  */
1899                 list_for_each_entry(work, &pool->worklist, entry)
1900                         send_mayday(work);
1901         }
1902 
1903         spin_unlock(&wq_mayday_lock);
1904         spin_unlock_irq(&pool->lock);
1905 
1906         mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1907 }
1908 
1909 /**
1910  * maybe_create_worker - create a new worker if necessary
1911  * @pool: pool to create a new worker for
1912  *
1913  * Create a new worker for @pool if necessary.  @pool is guaranteed to
1914  * have at least one idle worker on return from this function.  If
1915  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1916  * sent to all rescuers with works scheduled on @pool to resolve
1917  * possible allocation deadlock.
1918  *
1919  * On return, need_to_create_worker() is guaranteed to be %false and
1920  * may_start_working() %true.
1921  *
1922  * LOCKING:
1923  * spin_lock_irq(pool->lock) which may be released and regrabbed
1924  * multiple times.  Does GFP_KERNEL allocations.  Called only from
1925  * manager.
1926  */
1927 static void maybe_create_worker(struct worker_pool *pool)
1928 __releases(&pool->lock)
1929 __acquires(&pool->lock)
1930 {
1931 restart:
1932         spin_unlock_irq(&pool->lock);
1933 
1934         /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1935         mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1936 
1937         while (true) {
1938                 if (create_worker(pool) || !need_to_create_worker(pool))
1939                         break;
1940 
1941                 schedule_timeout_interruptible(CREATE_COOLDOWN);
1942 
1943                 if (!need_to_create_worker(pool))
1944                         break;
1945         }
1946 
1947         del_timer_sync(&pool->mayday_timer);
1948         spin_lock_irq(&pool->lock);
1949         /*
1950          * This is necessary even after a new worker was just successfully
1951          * created as @pool->lock was dropped and the new worker might have
1952          * already become busy.
1953          */
1954         if (need_to_create_worker(pool))
1955                 goto restart;
1956 }
1957 
1958 /**
1959  * manage_workers - manage worker pool
1960  * @worker: self
1961  *
1962  * Assume the manager role and manage the worker pool @worker belongs
1963  * to.  At any given time, there can be only zero or one manager per
1964  * pool.  The exclusion is handled automatically by this function.
1965  *
1966  * The caller can safely start processing works on false return.  On
1967  * true return, it's guaranteed that need_to_create_worker() is false
1968  * and may_start_working() is true.
1969  *
1970  * CONTEXT:
1971  * spin_lock_irq(pool->lock) which may be released and regrabbed
1972  * multiple times.  Does GFP_KERNEL allocations.
1973  *
1974  * Return:
1975  * %false if the pool doesn't need management and the caller can safely
1976  * start processing works, %true if management function was performed and
1977  * the conditions that the caller verified before calling the function may
1978  * no longer be true.
1979  */
1980 static bool manage_workers(struct worker *worker)
1981 {
1982         struct worker_pool *pool = worker->pool;
1983 
1984         if (pool->flags & POOL_MANAGER_ACTIVE)
1985                 return false;
1986 
1987         pool->flags |= POOL_MANAGER_ACTIVE;
1988         pool->manager = worker;
1989 
1990         maybe_create_worker(pool);
1991 
1992         pool->manager = NULL;
1993         pool->flags &= ~POOL_MANAGER_ACTIVE;
1994         wake_up(&wq_manager_wait);
1995         return true;
1996 }
1997 
1998 /**
1999  * process_one_work - process single work
2000  * @worker: self
2001  * @work: work to process
2002  *
2003  * Process @work.  This function contains all the logics necessary to
2004  * process a single work including synchronization against and
2005  * interaction with other workers on the same cpu, queueing and
2006  * flushing.  As long as context requirement is met, any worker can
2007  * call this function to process a work.
2008  *
2009  * CONTEXT:
2010  * spin_lock_irq(pool->lock) which is released and regrabbed.
2011  */
2012 static void process_one_work(struct worker *worker, struct work_struct *work)
2013 __releases(&pool->lock)
2014 __acquires(&pool->lock)
2015 {
2016         struct pool_workqueue *pwq = get_work_pwq(work);
2017         struct worker_pool *pool = worker->pool;
2018         bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2019         int work_color;
2020         struct worker *collision;
2021 #ifdef CONFIG_LOCKDEP
2022         /*
2023          * It is permissible to free the struct work_struct from
2024          * inside the function that is called from it, this we need to
2025          * take into account for lockdep too.  To avoid bogus "held
2026          * lock freed" warnings as well as problems when looking into
2027          * work->lockdep_map, make a copy and use that here.
2028          */
2029         struct lockdep_map lockdep_map;
2030 
2031         lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2032 #endif
2033         /* ensure we're on the correct CPU */
2034         WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2035                      raw_smp_processor_id() != pool->cpu);
2036 
2037         /*
2038          * A single work shouldn't be executed concurrently by
2039          * multiple workers on a single cpu.  Check whether anyone is
2040          * already processing the work.  If so, defer the work to the
2041          * currently executing one.
2042          */
2043         collision = find_worker_executing_work(pool, work);
2044         if (unlikely(collision)) {
2045                 move_linked_works(work, &collision->scheduled, NULL);
2046                 return;
2047         }
2048 
2049         /* claim and dequeue */
2050         debug_work_deactivate(work);
2051         hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2052         worker->current_work = work;
2053         worker->current_func = work->func;
2054         worker->current_pwq = pwq;
2055         work_color = get_work_color(work);
2056 
2057         list_del_init(&work->entry);
2058 
2059         /*
2060          * CPU intensive works don't participate in concurrency management.
2061          * They're the scheduler's responsibility.  This takes @worker out
2062          * of concurrency management and the next code block will chain
2063          * execution of the pending work items.
2064          */
2065         if (unlikely(cpu_intensive))
2066                 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2067 
2068         /*
2069          * Wake up another worker if necessary.  The condition is always
2070          * false for normal per-cpu workers since nr_running would always
2071          * be >= 1 at this point.  This is used to chain execution of the
2072          * pending work items for WORKER_NOT_RUNNING workers such as the
2073          * UNBOUND and CPU_INTENSIVE ones.
2074          */
2075         if (need_more_worker(pool))
2076                 wake_up_worker(pool);
2077 
2078         /*
2079          * Record the last pool and clear PENDING which should be the last
2080          * update to @work.  Also, do this inside @pool->lock so that
2081          * PENDING and queued state changes happen together while IRQ is
2082          * disabled.
2083          */
2084         set_work_pool_and_clear_pending(work, pool->id);
2085 
2086         spin_unlock_irq(&pool->lock);
2087 
2088         lock_map_acquire_read(&pwq->wq->lockdep_map);
2089         lock_map_acquire(&lockdep_map);
2090         trace_workqueue_execute_start(work);
2091         worker->current_func(work);
2092         /*
2093          * While we must be careful to not use "work" after this, the trace
2094          * point will only record its address.
2095          */
2096         trace_workqueue_execute_end(work);
2097         lock_map_release(&lockdep_map);
2098         lock_map_release(&pwq->wq->lockdep_map);
2099 
2100         if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2101                 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2102                        "     last function: %pf\n",
2103                        current->comm, preempt_count(), task_pid_nr(current),
2104                        worker->current_func);
2105                 debug_show_held_locks(current);
2106                 dump_stack();
2107         }
2108 
2109         /*
2110          * The following prevents a kworker from hogging CPU on !PREEMPT
2111          * kernels, where a requeueing work item waiting for something to
2112          * happen could deadlock with stop_machine as such work item could
2113          * indefinitely requeue itself while all other CPUs are trapped in
2114          * stop_machine. At the same time, report a quiescent RCU state so
2115          * the same condition doesn't freeze RCU.
2116          */
2117         cond_resched_rcu_qs();
2118 
2119         spin_lock_irq(&pool->lock);
2120 
2121         /* clear cpu intensive status */
2122         if (unlikely(cpu_intensive))
2123                 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2124 
2125         /* we're done with it, release */
2126         hash_del(&worker->hentry);
2127         worker->current_work = NULL;
2128         worker->current_func = NULL;
2129         worker->current_pwq = NULL;
2130         worker->desc_valid = false;
2131         pwq_dec_nr_in_flight(pwq, work_color);
2132 }
2133 
2134 /**
2135  * process_scheduled_works - process scheduled works
2136  * @worker: self
2137  *
2138  * Process all scheduled works.  Please note that the scheduled list
2139  * may change while processing a work, so this function repeatedly
2140  * fetches a work from the top and executes it.
2141  *
2142  * CONTEXT:
2143  * spin_lock_irq(pool->lock) which may be released and regrabbed
2144  * multiple times.
2145  */
2146 static void process_scheduled_works(struct worker *worker)
2147 {
2148         while (!list_empty(&worker->scheduled)) {
2149                 struct work_struct *work = list_first_entry(&worker->scheduled,
2150                                                 struct work_struct, entry);
2151                 process_one_work(worker, work);
2152         }
2153 }
2154 
2155 /**
2156  * worker_thread - the worker thread function
2157  * @__worker: self
2158  *
2159  * The worker thread function.  All workers belong to a worker_pool -
2160  * either a per-cpu one or dynamic unbound one.  These workers process all
2161  * work items regardless of their specific target workqueue.  The only
2162  * exception is work items which belong to workqueues with a rescuer which
2163  * will be explained in rescuer_thread().
2164  *
2165  * Return: 0
2166  */
2167 static int worker_thread(void *__worker)
2168 {
2169         struct worker *worker = __worker;
2170         struct worker_pool *pool = worker->pool;
2171 
2172         /* tell the scheduler that this is a workqueue worker */
2173         worker->task->flags |= PF_WQ_WORKER;
2174 woke_up:
2175         spin_lock_irq(&pool->lock);
2176 
2177         /* am I supposed to die? */
2178         if (unlikely(worker->flags & WORKER_DIE)) {
2179                 spin_unlock_irq(&pool->lock);
2180                 WARN_ON_ONCE(!list_empty(&worker->entry));
2181                 worker->task->flags &= ~PF_WQ_WORKER;
2182 
2183                 set_task_comm(worker->task, "kworker/dying");
2184                 ida_simple_remove(&pool->worker_ida, worker->id);
2185                 worker_detach_from_pool(worker, pool);
2186                 kfree(worker);
2187                 return 0;
2188         }
2189 
2190         worker_leave_idle(worker);
2191 recheck:
2192         /* no more worker necessary? */
2193         if (!need_more_worker(pool))
2194                 goto sleep;
2195 
2196         /* do we need to manage? */
2197         if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2198                 goto recheck;
2199 
2200         /*
2201          * ->scheduled list can only be filled while a worker is
2202          * preparing to process a work or actually processing it.
2203          * Make sure nobody diddled with it while I was sleeping.
2204          */
2205         WARN_ON_ONCE(!list_empty(&worker->scheduled));
2206 
2207         /*
2208          * Finish PREP stage.  We're guaranteed to have at least one idle
2209          * worker or that someone else has already assumed the manager
2210          * role.  This is where @worker starts participating in concurrency
2211          * management if applicable and concurrency management is restored
2212          * after being rebound.  See rebind_workers() for details.
2213          */
2214         worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2215 
2216         do {
2217                 struct work_struct *work =
2218                         list_first_entry(&pool->worklist,
2219                                          struct work_struct, entry);
2220 
2221                 pool->watchdog_ts = jiffies;
2222 
2223                 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2224                         /* optimization path, not strictly necessary */
2225                         process_one_work(worker, work);
2226                         if (unlikely(!list_empty(&worker->scheduled)))
2227                                 process_scheduled_works(worker);
2228                 } else {
2229                         move_linked_works(work, &worker->scheduled, NULL);
2230                         process_scheduled_works(worker);
2231                 }
2232         } while (keep_working(pool));
2233 
2234         worker_set_flags(worker, WORKER_PREP);
2235 sleep:
2236         /*
2237          * pool->lock is held and there's no work to process and no need to
2238          * manage, sleep.  Workers are woken up only while holding
2239          * pool->lock or from local cpu, so setting the current state
2240          * before releasing pool->lock is enough to prevent losing any
2241          * event.
2242          */
2243         worker_enter_idle(worker);
2244         __set_current_state(TASK_INTERRUPTIBLE);
2245         spin_unlock_irq(&pool->lock);
2246         schedule();
2247         goto woke_up;
2248 }
2249 
2250 /**
2251  * rescuer_thread - the rescuer thread function
2252  * @__rescuer: self
2253  *
2254  * Workqueue rescuer thread function.  There's one rescuer for each
2255  * workqueue which has WQ_MEM_RECLAIM set.
2256  *
2257  * Regular work processing on a pool may block trying to create a new
2258  * worker which uses GFP_KERNEL allocation which has slight chance of
2259  * developing into deadlock if some works currently on the same queue
2260  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2261  * the problem rescuer solves.
2262  *
2263  * When such condition is possible, the pool summons rescuers of all
2264  * workqueues which have works queued on the pool and let them process
2265  * those works so that forward progress can be guaranteed.
2266  *
2267  * This should happen rarely.
2268  *
2269  * Return: 0
2270  */
2271 static int rescuer_thread(void *__rescuer)
2272 {
2273         struct worker *rescuer = __rescuer;
2274         struct workqueue_struct *wq = rescuer->rescue_wq;
2275         struct list_head *scheduled = &rescuer->scheduled;
2276         bool should_stop;
2277 
2278         set_user_nice(current, RESCUER_NICE_LEVEL);
2279 
2280         /*
2281          * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2282          * doesn't participate in concurrency management.
2283          */
2284         rescuer->task->flags |= PF_WQ_WORKER;
2285 repeat:
2286         set_current_state(TASK_INTERRUPTIBLE);
2287 
2288         /*
2289          * By the time the rescuer is requested to stop, the workqueue
2290          * shouldn't have any work pending, but @wq->maydays may still have
2291          * pwq(s) queued.  This can happen by non-rescuer workers consuming
2292          * all the work items before the rescuer got to them.  Go through
2293          * @wq->maydays processing before acting on should_stop so that the
2294          * list is always empty on exit.
2295          */
2296         should_stop = kthread_should_stop();
2297 
2298         /* see whether any pwq is asking for help */
2299         spin_lock_irq(&wq_mayday_lock);
2300 
2301         while (!list_empty(&wq->maydays)) {
2302                 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2303                                         struct pool_workqueue, mayday_node);
2304                 struct worker_pool *pool = pwq->pool;
2305                 struct work_struct *work, *n;
2306                 bool first = true;
2307 
2308                 __set_current_state(TASK_RUNNING);
2309                 list_del_init(&pwq->mayday_node);
2310 
2311                 spin_unlock_irq(&wq_mayday_lock);
2312 
2313                 worker_attach_to_pool(rescuer, pool);
2314 
2315                 spin_lock_irq(&pool->lock);
2316                 rescuer->pool = pool;
2317 
2318                 /*
2319                  * Slurp in all works issued via this workqueue and
2320                  * process'em.
2321                  */
2322                 WARN_ON_ONCE(!list_empty(scheduled));
2323                 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2324                         if (get_work_pwq(work) == pwq) {
2325                                 if (first)
2326                                         pool->watchdog_ts = jiffies;
2327                                 move_linked_works(work, scheduled, &n);
2328                         }
2329                         first = false;
2330                 }
2331 
2332                 if (!list_empty(scheduled)) {
2333                         process_scheduled_works(rescuer);
2334 
2335                         /*
2336                          * The above execution of rescued work items could
2337                          * have created more to rescue through
2338                          * pwq_activate_first_delayed() or chained
2339                          * queueing.  Let's put @pwq back on mayday list so
2340                          * that such back-to-back work items, which may be
2341                          * being used to relieve memory pressure, don't
2342                          * incur MAYDAY_INTERVAL delay inbetween.
2343                          */
2344                         if (need_to_create_worker(pool)) {
2345                                 spin_lock(&wq_mayday_lock);
2346                                 get_pwq(pwq);
2347                                 list_move_tail(&pwq->mayday_node, &wq->maydays);
2348                                 spin_unlock(&wq_mayday_lock);
2349                         }
2350                 }
2351 
2352                 /*
2353                  * Put the reference grabbed by send_mayday().  @pool won't
2354                  * go away while we're still attached to it.
2355                  */
2356                 put_pwq(pwq);
2357 
2358                 /*
2359                  * Leave this pool.  If need_more_worker() is %true, notify a
2360                  * regular worker; otherwise, we end up with 0 concurrency
2361                  * and stalling the execution.
2362                  */
2363                 if (need_more_worker(pool))
2364                         wake_up_worker(pool);
2365 
2366                 rescuer->pool = NULL;
2367                 spin_unlock_irq(&pool->lock);
2368 
2369                 worker_detach_from_pool(rescuer, pool);
2370 
2371                 spin_lock_irq(&wq_mayday_lock);
2372         }
2373 
2374         spin_unlock_irq(&wq_mayday_lock);
2375 
2376         if (should_stop) {
2377                 __set_current_state(TASK_RUNNING);
2378                 rescuer->task->flags &= ~PF_WQ_WORKER;
2379                 return 0;
2380         }
2381 
2382         /* rescuers should never participate in concurrency management */
2383         WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2384         schedule();
2385         goto repeat;
2386 }
2387 
2388 /**
2389  * check_flush_dependency - check for flush dependency sanity
2390  * @target_wq: workqueue being flushed
2391  * @target_work: work item being flushed (NULL for workqueue flushes)
2392  *
2393  * %current is trying to flush the whole @target_wq or @target_work on it.
2394  * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2395  * reclaiming memory or running on a workqueue which doesn't have
2396  * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2397  * a deadlock.
2398  */
2399 static void check_flush_dependency(struct workqueue_struct *target_wq,
2400                                    struct work_struct *target_work)
2401 {
2402         work_func_t target_func = target_work ? target_work->func : NULL;
2403         struct worker *worker;
2404 
2405         if (target_wq->flags & WQ_MEM_RECLAIM)
2406                 return;
2407 
2408         worker = current_wq_worker();
2409 
2410         WARN_ONCE(current->flags & PF_MEMALLOC,
2411                   "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2412                   current->pid, current->comm, target_wq->name, target_func);
2413         WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2414                               (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2415                   "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2416                   worker->current_pwq->wq->name, worker->current_func,
2417                   target_wq->name, target_func);
2418 }
2419 
2420 struct wq_barrier {
2421         struct work_struct      work;
2422         struct completion       done;
2423         struct task_struct      *task;  /* purely informational */
2424 };
2425 
2426 static void wq_barrier_func(struct work_struct *work)
2427 {
2428         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2429         complete(&barr->done);
2430 }
2431 
2432 /**
2433  * insert_wq_barrier - insert a barrier work
2434  * @pwq: pwq to insert barrier into
2435  * @barr: wq_barrier to insert
2436  * @target: target work to attach @barr to
2437  * @worker: worker currently executing @target, NULL if @target is not executing
2438  *
2439  * @barr is linked to @target such that @barr is completed only after
2440  * @target finishes execution.  Please note that the ordering
2441  * guarantee is observed only with respect to @target and on the local
2442  * cpu.
2443  *
2444  * Currently, a queued barrier can't be canceled.  This is because
2445  * try_to_grab_pending() can't determine whether the work to be
2446  * grabbed is at the head of the queue and thus can't clear LINKED
2447  * flag of the previous work while there must be a valid next work
2448  * after a work with LINKED flag set.
2449  *
2450  * Note that when @worker is non-NULL, @target may be modified
2451  * underneath us, so we can't reliably determine pwq from @target.
2452  *
2453  * CONTEXT:
2454  * spin_lock_irq(pool->lock).
2455  */
2456 static void insert_wq_barrier(struct pool_workqueue *pwq,
2457                               struct wq_barrier *barr,
2458                               struct work_struct *target, struct worker *worker)
2459 {
2460         struct list_head *head;
2461         unsigned int linked = 0;
2462 
2463         /*
2464          * debugobject calls are safe here even with pool->lock locked
2465          * as we know for sure that this will not trigger any of the
2466          * checks and call back into the fixup functions where we
2467          * might deadlock.
2468          */
2469         INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2470         __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2471         init_completion(&barr->done);
2472         barr->task = current;
2473 
2474         /*
2475          * If @target is currently being executed, schedule the
2476          * barrier to the worker; otherwise, put it after @target.
2477          */
2478         if (worker)
2479                 head = worker->scheduled.next;
2480         else {
2481                 unsigned long *bits = work_data_bits(target);
2482 
2483                 head = target->entry.next;
2484                 /* there can already be other linked works, inherit and set */
2485                 linked = *bits & WORK_STRUCT_LINKED;
2486                 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2487         }
2488 
2489         debug_work_activate(&barr->work);
2490         insert_work(pwq, &barr->work, head,
2491                     work_color_to_flags(WORK_NO_COLOR) | linked);
2492 }
2493 
2494 /**
2495  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2496  * @wq: workqueue being flushed
2497  * @flush_color: new flush color, < 0 for no-op
2498  * @work_color: new work color, < 0 for no-op
2499  *
2500  * Prepare pwqs for workqueue flushing.
2501  *
2502  * If @flush_color is non-negative, flush_color on all pwqs should be
2503  * -1.  If no pwq has in-flight commands at the specified color, all
2504  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2505  * has in flight commands, its pwq->flush_color is set to
2506  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2507  * wakeup logic is armed and %true is returned.
2508  *
2509  * The caller should have initialized @wq->first_flusher prior to
2510  * calling this function with non-negative @flush_color.  If
2511  * @flush_color is negative, no flush color update is done and %false
2512  * is returned.
2513  *
2514  * If @work_color is non-negative, all pwqs should have the same
2515  * work_color which is previous to @work_color and all will be
2516  * advanced to @work_color.
2517  *
2518  * CONTEXT:
2519  * mutex_lock(wq->mutex).
2520  *
2521  * Return:
2522  * %true if @flush_color >= 0 and there's something to flush.  %false
2523  * otherwise.
2524  */
2525 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2526                                       int flush_color, int work_color)
2527 {
2528         bool wait = false;
2529         struct pool_workqueue *pwq;
2530 
2531         if (flush_color >= 0) {
2532                 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2533                 atomic_set(&wq->nr_pwqs_to_flush, 1);
2534         }
2535 
2536         for_each_pwq(pwq, wq) {
2537                 struct worker_pool *pool = pwq->pool;
2538 
2539                 spin_lock_irq(&pool->lock);
2540 
2541                 if (flush_color >= 0) {
2542                         WARN_ON_ONCE(pwq->flush_color != -1);
2543 
2544                         if (pwq->nr_in_flight[flush_color]) {
2545                                 pwq->flush_color = flush_color;
2546                                 atomic_inc(&wq->nr_pwqs_to_flush);
2547                                 wait = true;
2548                         }
2549                 }
2550 
2551                 if (work_color >= 0) {
2552                         WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2553                         pwq->work_color = work_color;
2554                 }
2555 
2556                 spin_unlock_irq(&pool->lock);
2557         }
2558 
2559         if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2560                 complete(&wq->first_flusher->done);
2561 
2562         return wait;
2563 }
2564 
2565 /**
2566  * flush_workqueue - ensure that any scheduled work has run to completion.
2567  * @wq: workqueue to flush
2568  *
2569  * This function sleeps until all work items which were queued on entry
2570  * have finished execution, but it is not livelocked by new incoming ones.
2571  */
2572 void flush_workqueue(struct workqueue_struct *wq)
2573 {
2574         struct wq_flusher this_flusher = {
2575                 .list = LIST_HEAD_INIT(this_flusher.list),
2576                 .flush_color = -1,
2577                 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2578         };
2579         int next_color;
2580 
2581         if (WARN_ON(!wq_online))
2582                 return;
2583 
2584         lock_map_acquire(&wq->lockdep_map);
2585         lock_map_release(&wq->lockdep_map);
2586 
2587         mutex_lock(&wq->mutex);
2588 
2589         /*
2590          * Start-to-wait phase
2591          */
2592         next_color = work_next_color(wq->work_color);
2593 
2594         if (next_color != wq->flush_color) {
2595                 /*
2596                  * Color space is not full.  The current work_color
2597                  * becomes our flush_color and work_color is advanced
2598                  * by one.
2599                  */
2600                 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2601                 this_flusher.flush_color = wq->work_color;
2602                 wq->work_color = next_color;
2603 
2604                 if (!wq->first_flusher) {
2605                         /* no flush in progress, become the first flusher */
2606                         WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2607 
2608                         wq->first_flusher = &this_flusher;
2609 
2610                         if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2611                                                        wq->work_color)) {
2612                                 /* nothing to flush, done */
2613                                 wq->flush_color = next_color;
2614                                 wq->first_flusher = NULL;
2615                                 goto out_unlock;
2616                         }
2617                 } else {
2618                         /* wait in queue */
2619                         WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2620                         list_add_tail(&this_flusher.list, &wq->flusher_queue);
2621                         flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2622                 }
2623         } else {
2624                 /*
2625                  * Oops, color space is full, wait on overflow queue.
2626                  * The next flush completion will assign us
2627                  * flush_color and transfer to flusher_queue.
2628                  */
2629                 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2630         }
2631 
2632         check_flush_dependency(wq, NULL);
2633 
2634         mutex_unlock(&wq->mutex);
2635 
2636         wait_for_completion(&this_flusher.done);
2637 
2638         /*
2639          * Wake-up-and-cascade phase
2640          *
2641          * First flushers are responsible for cascading flushes and
2642          * handling overflow.  Non-first flushers can simply return.
2643          */
2644         if (wq->first_flusher != &this_flusher)
2645                 return;
2646 
2647         mutex_lock(&wq->mutex);
2648 
2649         /* we might have raced, check again with mutex held */
2650         if (wq->first_flusher != &this_flusher)
2651                 goto out_unlock;
2652 
2653         wq->first_flusher = NULL;
2654 
2655         WARN_ON_ONCE(!list_empty(&this_flusher.list));
2656         WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2657 
2658         while (true) {
2659                 struct wq_flusher *next, *tmp;
2660 
2661                 /* complete all the flushers sharing the current flush color */
2662                 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2663                         if (next->flush_color != wq->flush_color)
2664                                 break;
2665                         list_del_init(&next->list);
2666                         complete(&next->done);
2667                 }
2668 
2669                 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2670                              wq->flush_color != work_next_color(wq->work_color));
2671 
2672                 /* this flush_color is finished, advance by one */
2673                 wq->flush_color = work_next_color(wq->flush_color);
2674 
2675                 /* one color has been freed, handle overflow queue */
2676                 if (!list_empty(&wq->flusher_overflow)) {
2677                         /*
2678                          * Assign the same color to all overflowed
2679                          * flushers, advance work_color and append to
2680                          * flusher_queue.  This is the start-to-wait
2681                          * phase for these overflowed flushers.
2682                          */
2683                         list_for_each_entry(tmp, &wq->flusher_overflow, list)
2684                                 tmp->flush_color = wq->work_color;
2685 
2686                         wq->work_color = work_next_color(wq->work_color);
2687 
2688                         list_splice_tail_init(&wq->flusher_overflow,
2689                                               &wq->flusher_queue);
2690                         flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2691                 }
2692 
2693                 if (list_empty(&wq->flusher_queue)) {
2694                         WARN_ON_ONCE(wq->flush_color != wq->work_color);
2695                         break;
2696                 }
2697 
2698                 /*
2699                  * Need to flush more colors.  Make the next flusher
2700                  * the new first flusher and arm pwqs.
2701                  */
2702                 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2703                 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2704 
2705                 list_del_init(&next->list);
2706                 wq->first_flusher = next;
2707 
2708                 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2709                         break;
2710 
2711                 /*
2712                  * Meh... this color is already done, clear first
2713                  * flusher and repeat cascading.
2714                  */
2715                 wq->first_flusher = NULL;
2716         }
2717 
2718 out_unlock:
2719         mutex_unlock(&wq->mutex);
2720 }
2721 EXPORT_SYMBOL(flush_workqueue);
2722 
2723 /**
2724  * drain_workqueue - drain a workqueue
2725  * @wq: workqueue to drain
2726  *
2727  * Wait until the workqueue becomes empty.  While draining is in progress,
2728  * only chain queueing is allowed.  IOW, only currently pending or running
2729  * work items on @wq can queue further work items on it.  @wq is flushed
2730  * repeatedly until it becomes empty.  The number of flushing is determined
2731  * by the depth of chaining and should be relatively short.  Whine if it
2732  * takes too long.
2733  */
2734 void drain_workqueue(struct workqueue_struct *wq)
2735 {
2736         unsigned int flush_cnt = 0;
2737         struct pool_workqueue *pwq;
2738 
2739         /*
2740          * __queue_work() needs to test whether there are drainers, is much
2741          * hotter than drain_workqueue() and already looks at @wq->flags.
2742          * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2743          */
2744         mutex_lock(&wq->mutex);
2745         if (!wq->nr_drainers++)
2746                 wq->flags |= __WQ_DRAINING;
2747         mutex_unlock(&wq->mutex);
2748 reflush:
2749         flush_workqueue(wq);
2750 
2751         mutex_lock(&wq->mutex);
2752 
2753         for_each_pwq(pwq, wq) {
2754                 bool drained;
2755 
2756                 spin_lock_irq(&pwq->pool->lock);
2757                 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2758                 spin_unlock_irq(&pwq->pool->lock);
2759 
2760                 if (drained)
2761                         continue;
2762 
2763                 if (++flush_cnt == 10 ||
2764                     (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2765                         pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2766                                 wq->name, flush_cnt);
2767 
2768                 mutex_unlock(&wq->mutex);
2769                 goto reflush;
2770         }
2771 
2772         if (!--wq->nr_drainers)
2773                 wq->flags &= ~__WQ_DRAINING;
2774         mutex_unlock(&wq->mutex);
2775 }
2776 EXPORT_SYMBOL_GPL(drain_workqueue);
2777 
2778 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2779 {
2780         struct worker *worker = NULL;
2781         struct worker_pool *pool;
2782         struct pool_workqueue *pwq;
2783 
2784         might_sleep();
2785 
2786         local_irq_disable();
2787         pool = get_work_pool(work);
2788         if (!pool) {
2789                 local_irq_enable();
2790                 return false;
2791         }
2792 
2793         spin_lock(&pool->lock);
2794         /* see the comment in try_to_grab_pending() with the same code */
2795         pwq = get_work_pwq(work);
2796         if (pwq) {
2797                 if (unlikely(pwq->pool != pool))
2798                         goto already_gone;
2799         } else {
2800                 worker = find_worker_executing_work(pool, work);
2801                 if (!worker)
2802                         goto already_gone;
2803                 pwq = worker->current_pwq;
2804         }
2805 
2806         check_flush_dependency(pwq->wq, work);
2807 
2808         insert_wq_barrier(pwq, barr, work, worker);
2809         spin_unlock_irq(&pool->lock);
2810 
2811         /*
2812          * If @max_active is 1 or rescuer is in use, flushing another work
2813          * item on the same workqueue may lead to deadlock.  Make sure the
2814          * flusher is not running on the same workqueue by verifying write
2815          * access.
2816          */
2817         if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2818                 lock_map_acquire(&pwq->wq->lockdep_map);
2819         else
2820                 lock_map_acquire_read(&pwq->wq->lockdep_map);
2821         lock_map_release(&pwq->wq->lockdep_map);
2822 
2823         return true;
2824 already_gone:
2825         spin_unlock_irq(&pool->lock);
2826         return false;
2827 }
2828 
2829 /**
2830  * flush_work - wait for a work to finish executing the last queueing instance
2831  * @work: the work to flush
2832  *
2833  * Wait until @work has finished execution.  @work is guaranteed to be idle
2834  * on return if it hasn't been requeued since flush started.
2835  *
2836  * Return:
2837  * %true if flush_work() waited for the work to finish execution,
2838  * %false if it was already idle.
2839  */
2840 bool flush_work(struct work_struct *work)
2841 {
2842         struct wq_barrier barr;
2843 
2844         if (WARN_ON(!wq_online))
2845                 return false;
2846 
2847         lock_map_acquire(&work->lockdep_map);
2848         lock_map_release(&work->lockdep_map);
2849 
2850         if (start_flush_work(work, &barr)) {
2851                 wait_for_completion(&barr.done);
2852                 destroy_work_on_stack(&barr.work);
2853                 return true;
2854         } else {
2855                 return false;
2856         }
2857 }
2858 EXPORT_SYMBOL_GPL(flush_work);
2859 
2860 struct cwt_wait {
2861         wait_queue_entry_t              wait;
2862         struct work_struct      *work;
2863 };
2864 
2865 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2866 {
2867         struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2868 
2869         if (cwait->work != key)
2870                 return 0;
2871         return autoremove_wake_function(wait, mode, sync, key);
2872 }
2873 
2874 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2875 {
2876         static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2877         unsigned long flags;
2878         int ret;
2879 
2880         do {
2881                 ret = try_to_grab_pending(work, is_dwork, &flags);
2882                 /*
2883                  * If someone else is already canceling, wait for it to
2884                  * finish.  flush_work() doesn't work for PREEMPT_NONE
2885                  * because we may get scheduled between @work's completion
2886                  * and the other canceling task resuming and clearing
2887                  * CANCELING - flush_work() will return false immediately
2888                  * as @work is no longer busy, try_to_grab_pending() will
2889                  * return -ENOENT as @work is still being canceled and the
2890                  * other canceling task won't be able to clear CANCELING as
2891                  * we're hogging the CPU.
2892                  *
2893                  * Let's wait for completion using a waitqueue.  As this
2894                  * may lead to the thundering herd problem, use a custom
2895                  * wake function which matches @work along with exclusive
2896                  * wait and wakeup.
2897                  */
2898                 if (unlikely(ret == -ENOENT)) {
2899                         struct cwt_wait cwait;
2900 
2901                         init_wait(&cwait.wait);
2902                         cwait.wait.func = cwt_wakefn;
2903                         cwait.work = work;
2904 
2905                         prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2906                                                   TASK_UNINTERRUPTIBLE);
2907                         if (work_is_canceling(work))
2908                                 schedule();
2909                         finish_wait(&cancel_waitq, &cwait.wait);
2910                 }
2911         } while (unlikely(ret < 0));
2912 
2913         /* tell other tasks trying to grab @work to back off */
2914         mark_work_canceling(work);
2915         local_irq_restore(flags);
2916 
2917         /*
2918          * This allows canceling during early boot.  We know that @work
2919          * isn't executing.
2920          */
2921         if (wq_online)
2922                 flush_work(work);
2923 
2924         clear_work_data(work);
2925 
2926         /*
2927          * Paired with prepare_to_wait() above so that either
2928          * waitqueue_active() is visible here or !work_is_canceling() is
2929          * visible there.
2930          */
2931         smp_mb();
2932         if (waitqueue_active(&cancel_waitq))
2933                 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2934 
2935         return ret;
2936 }
2937 
2938 /**
2939  * cancel_work_sync - cancel a work and wait for it to finish
2940  * @work: the work to cancel
2941  *
2942  * Cancel @work and wait for its execution to finish.  This function
2943  * can be used even if the work re-queues itself or migrates to
2944  * another workqueue.  On return from this function, @work is
2945  * guaranteed to be not pending or executing on any CPU.
2946  *
2947  * cancel_work_sync(&delayed_work->work) must not be used for
2948  * delayed_work's.  Use cancel_delayed_work_sync() instead.
2949  *
2950  * The caller must ensure that the workqueue on which @work was last
2951  * queued can't be destroyed before this function returns.
2952  *
2953  * Return:
2954  * %true if @work was pending, %false otherwise.
2955  */
2956 bool cancel_work_sync(struct work_struct *work)
2957 {
2958         return __cancel_work_timer(work, false);
2959 }
2960 EXPORT_SYMBOL_GPL(cancel_work_sync);
2961 
2962 /**
2963  * flush_delayed_work - wait for a dwork to finish executing the last queueing
2964  * @dwork: the delayed work to flush
2965  *
2966  * Delayed timer is cancelled and the pending work is queued for
2967  * immediate execution.  Like flush_work(), this function only
2968  * considers the last queueing instance of @dwork.
2969  *
2970  * Return:
2971  * %true if flush_work() waited for the work to finish execution,
2972  * %false if it was already idle.
2973  */
2974 bool flush_delayed_work(struct delayed_work *dwork)
2975 {
2976         local_irq_disable();
2977         if (del_timer_sync(&dwork->timer))
2978                 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2979         local_irq_enable();
2980         return flush_work(&dwork->work);
2981 }
2982 EXPORT_SYMBOL(flush_delayed_work);
2983 
2984 static bool __cancel_work(struct work_struct *work, bool is_dwork)
2985 {
2986         unsigned long flags;
2987         int ret;
2988 
2989         do {
2990                 ret = try_to_grab_pending(work, is_dwork, &flags);
2991         } while (unlikely(ret == -EAGAIN));
2992 
2993         if (unlikely(ret < 0))
2994                 return false;
2995 
2996         set_work_pool_and_clear_pending(work, get_work_pool_id(work));
2997         local_irq_restore(flags);
2998         return ret;
2999 }
3000 
3001 /*
3002  * See cancel_delayed_work()
3003  */
3004 bool cancel_work(struct work_struct *work)
3005 {
3006         return __cancel_work(work, false);
3007 }
3008 
3009 /**
3010  * cancel_delayed_work - cancel a delayed work
3011  * @dwork: delayed_work to cancel
3012  *
3013  * Kill off a pending delayed_work.
3014  *
3015  * Return: %true if @dwork was pending and canceled; %false if it wasn't
3016  * pending.
3017  *
3018  * Note:
3019  * The work callback function may still be running on return, unless
3020  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
3021  * use cancel_delayed_work_sync() to wait on it.
3022  *
3023  * This function is safe to call from any context including IRQ handler.
3024  */
3025 bool cancel_delayed_work(struct delayed_work *dwork)
3026 {
3027         return __cancel_work(&dwork->work, true);
3028 }
3029 EXPORT_SYMBOL(cancel_delayed_work);
3030 
3031 /**
3032  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3033  * @dwork: the delayed work cancel
3034  *
3035  * This is cancel_work_sync() for delayed works.
3036  *
3037  * Return:
3038  * %true if @dwork was pending, %false otherwise.
3039  */
3040 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3041 {
3042         return __cancel_work_timer(&dwork->work, true);
3043 }
3044 EXPORT_SYMBOL(cancel_delayed_work_sync);
3045 
3046 /**
3047  * schedule_on_each_cpu - execute a function synchronously on each online CPU
3048  * @func: the function to call
3049  *
3050  * schedule_on_each_cpu() executes @func on each online CPU using the
3051  * system workqueue and blocks until all CPUs have completed.
3052  * schedule_on_each_cpu() is very slow.
3053  *
3054  * Return:
3055  * 0 on success, -errno on failure.
3056  */
3057 int schedule_on_each_cpu(work_func_t func)
3058 {
3059         int cpu;
3060         struct work_struct __percpu *works;
3061 
3062         works = alloc_percpu(struct work_struct);
3063         if (!works)
3064                 return -ENOMEM;
3065 
3066         get_online_cpus();
3067 
3068         for_each_online_cpu(cpu) {
3069                 struct work_struct *work = per_cpu_ptr(works, cpu);
3070 
3071                 INIT_WORK(work, func);
3072                 schedule_work_on(cpu, work);
3073         }
3074 
3075         for_each_online_cpu(cpu)
3076                 flush_work(per_cpu_ptr(works, cpu));
3077 
3078         put_online_cpus();
3079         free_percpu(works);
3080         return 0;
3081 }
3082 
3083 /**
3084  * execute_in_process_context - reliably execute the routine with user context
3085  * @fn:         the function to execute
3086  * @ew:         guaranteed storage for the execute work structure (must
3087  *              be available when the work executes)
3088  *
3089  * Executes the function immediately if process context is available,
3090  * otherwise schedules the function for delayed execution.
3091  *
3092  * Return:      0 - function was executed
3093  *              1 - function was scheduled for execution
3094  */
3095 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3096 {
3097         if (!in_interrupt()) {
3098                 fn(&ew->work);
3099                 return 0;
3100         }
3101 
3102         INIT_WORK(&ew->work, fn);
3103         schedule_work(&ew->work);
3104 
3105         return 1;
3106 }
3107 EXPORT_SYMBOL_GPL(execute_in_process_context);
3108 
3109 /**
3110  * free_workqueue_attrs - free a workqueue_attrs
3111  * @attrs: workqueue_attrs to free
3112  *
3113  * Undo alloc_workqueue_attrs().
3114  */
3115 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3116 {
3117         if (attrs) {
3118                 free_cpumask_var(attrs->cpumask);
3119                 kfree(attrs);
3120         }
3121 }
3122 
3123 /**
3124  * alloc_workqueue_attrs - allocate a workqueue_attrs
3125  * @gfp_mask: allocation mask to use
3126  *
3127  * Allocate a new workqueue_attrs, initialize with default settings and
3128  * return it.
3129  *
3130  * Return: The allocated new workqueue_attr on success. %NULL on failure.
3131  */
3132 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3133 {
3134         struct workqueue_attrs *attrs;
3135 
3136         attrs = kzalloc(sizeof(*attrs), gfp_mask);
3137         if (!attrs)
3138                 goto fail;
3139         if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3140                 goto fail;
3141 
3142         cpumask_copy(attrs->cpumask, cpu_possible_mask);
3143         return attrs;
3144 fail:
3145         free_workqueue_attrs(attrs);
3146         return NULL;
3147 }
3148 
3149 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3150                                  const struct workqueue_attrs *from)
3151 {
3152         to->nice = from->nice;
3153         cpumask_copy(to->cpumask, from->cpumask);
3154         /*
3155          * Unlike hash and equality test, this function doesn't ignore
3156          * ->no_numa as it is used for both pool and wq attrs.  Instead,
3157          * get_unbound_pool() explicitly clears ->no_numa after copying.
3158          */
3159         to->no_numa = from->no_numa;
3160 }
3161 
3162 /* hash value of the content of @attr */
3163 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3164 {
3165         u32 hash = 0;
3166 
3167         hash = jhash_1word(attrs->nice, hash);
3168         hash = jhash(cpumask_bits(attrs->cpumask),
3169                      BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3170         return hash;
3171 }
3172 
3173 /* content equality test */
3174 static bool wqattrs_equal(const struct workqueue_attrs *a,
3175                           const struct workqueue_attrs *b)
3176 {
3177         if (a->nice != b->nice)
3178                 return false;
3179         if (!cpumask_equal(a->cpumask, b->cpumask))
3180                 return false;
3181         return true;
3182 }
3183 
3184 /**
3185  * init_worker_pool - initialize a newly zalloc'd worker_pool
3186  * @pool: worker_pool to initialize
3187  *
3188  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3189  *
3190  * Return: 0 on success, -errno on failure.  Even on failure, all fields
3191  * inside @pool proper are initialized and put_unbound_pool() can be called
3192  * on @pool safely to release it.
3193  */
3194 static int init_worker_pool(struct worker_pool *pool)
3195 {
3196         spin_lock_init(&pool->lock);
3197         pool->id = -1;
3198         pool->cpu = -1;
3199         pool->node = NUMA_NO_NODE;
3200         pool->flags |= POOL_DISASSOCIATED;
3201         pool->watchdog_ts = jiffies;
3202         INIT_LIST_HEAD(&pool->worklist);
3203         INIT_LIST_HEAD(&pool->idle_list);
3204         hash_init(pool->busy_hash);
3205 
3206         setup_deferrable_timer(&pool->idle_timer, idle_worker_timeout,
3207                                (unsigned long)pool);
3208 
3209         setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3210                     (unsigned long)pool);
3211 
3212         mutex_init(&pool->attach_mutex);
3213         INIT_LIST_HEAD(&pool->workers);
3214 
3215         ida_init(&pool->worker_ida);
3216         INIT_HLIST_NODE(&pool->hash_node);
3217         pool->refcnt = 1;
3218 
3219         /* shouldn't fail above this point */
3220         pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3221         if (!pool->attrs)
3222                 return -ENOMEM;
3223         return 0;
3224 }
3225 
3226 static void rcu_free_wq(struct rcu_head *rcu)
3227 {
3228         struct workqueue_struct *wq =
3229                 container_of(rcu, struct workqueue_struct, rcu);
3230 
3231         if (!(wq->flags & WQ_UNBOUND))
3232                 free_percpu(wq->cpu_pwqs);
3233         else
3234                 free_workqueue_attrs(wq->unbound_attrs);
3235 
3236         kfree(wq->rescuer);
3237         kfree(wq);
3238 }
3239 
3240 static void rcu_free_pool(struct rcu_head *rcu)
3241 {
3242         struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3243 
3244         ida_destroy(&pool->worker_ida);
3245         free_workqueue_attrs(pool->attrs);
3246         kfree(pool);
3247 }
3248 
3249 /**
3250  * put_unbound_pool - put a worker_pool
3251  * @pool: worker_pool to put
3252  *
3253  * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3254  * safe manner.  get_unbound_pool() calls this function on its failure path
3255  * and this function should be able to release pools which went through,
3256  * successfully or not, init_worker_pool().
3257  *
3258  * Should be called with wq_pool_mutex held.
3259  */
3260 static void put_unbound_pool(struct worker_pool *pool)
3261 {
3262         DECLARE_COMPLETION_ONSTACK(detach_completion);
3263         struct worker *worker;
3264 
3265         lockdep_assert_held(&wq_pool_mutex);
3266 
3267         if (--pool->refcnt)
3268                 return;
3269 
3270         /* sanity checks */
3271         if (WARN_ON(!(pool->cpu < 0)) ||
3272             WARN_ON(!list_empty(&pool->worklist)))
3273                 return;
3274 
3275         /* release id and unhash */
3276         if (pool->id >= 0)
3277                 idr_remove(&worker_pool_idr, pool->id);
3278         hash_del(&pool->hash_node);
3279 
3280         /*
3281          * Become the manager and destroy all workers.  This prevents
3282          * @pool's workers from blocking on attach_mutex.  We're the last
3283          * manager and @pool gets freed with the flag set.
3284          */
3285         spin_lock_irq(&pool->lock);
3286         wait_event_lock_irq(wq_manager_wait,
3287                             !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3288         pool->flags |= POOL_MANAGER_ACTIVE;
3289 
3290         while ((worker = first_idle_worker(pool)))
3291                 destroy_worker(worker);
3292         WARN_ON(pool->nr_workers || pool->nr_idle);
3293         spin_unlock_irq(&pool->lock);
3294 
3295         mutex_lock(&pool->attach_mutex);
3296         if (!list_empty(&pool->workers))
3297                 pool->detach_completion = &detach_completion;
3298         mutex_unlock(&pool->attach_mutex);
3299 
3300         if (pool->detach_completion)
3301                 wait_for_completion(pool->detach_completion);
3302 
3303         /* shut down the timers */
3304         del_timer_sync(&pool->idle_timer);
3305         del_timer_sync(&pool->mayday_timer);
3306 
3307         /* sched-RCU protected to allow dereferences from get_work_pool() */
3308         call_rcu_sched(&pool->rcu, rcu_free_pool);
3309 }
3310 
3311 /**
3312  * get_unbound_pool - get a worker_pool with the specified attributes
3313  * @attrs: the attributes of the worker_pool to get
3314  *
3315  * Obtain a worker_pool which has the same attributes as @attrs, bump the
3316  * reference count and return it.  If there already is a matching
3317  * worker_pool, it will be used; otherwise, this function attempts to
3318  * create a new one.
3319  *
3320  * Should be called with wq_pool_mutex held.
3321  *
3322  * Return: On success, a worker_pool with the same attributes as @attrs.
3323  * On failure, %NULL.
3324  */
3325 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3326 {
3327         u32 hash = wqattrs_hash(attrs);
3328         struct worker_pool *pool;
3329         int node;
3330         int target_node = NUMA_NO_NODE;
3331 
3332         lockdep_assert_held(&wq_pool_mutex);
3333 
3334         /* do we already have a matching pool? */
3335         hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3336                 if (wqattrs_equal(pool->attrs, attrs)) {
3337                         pool->refcnt++;
3338                         return pool;
3339                 }
3340         }
3341 
3342         /* if cpumask is contained inside a NUMA node, we belong to that node */
3343         if (wq_numa_enabled) {
3344                 for_each_node(node) {
3345                         if (cpumask_subset(attrs->cpumask,
3346                                            wq_numa_possible_cpumask[node])) {
3347                                 target_node = node;
3348                                 break;
3349                         }
3350                 }
3351         }
3352 
3353         /* nope, create a new one */
3354         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3355         if (!pool || init_worker_pool(pool) < 0)
3356                 goto fail;
3357 
3358         lockdep_set_subclass(&pool->lock, 1);   /* see put_pwq() */
3359         copy_workqueue_attrs(pool->attrs, attrs);
3360         pool->node = target_node;
3361 
3362         /*
3363          * no_numa isn't a worker_pool attribute, always clear it.  See
3364          * 'struct workqueue_attrs' comments for detail.
3365          */
3366         pool->attrs->no_numa = false;
3367 
3368         if (worker_pool_assign_id(pool) < 0)
3369                 goto fail;
3370 
3371         /* create and start the initial worker */
3372         if (wq_online && !create_worker(pool))
3373                 goto fail;
3374 
3375         /* install */
3376         hash_add(unbound_pool_hash, &pool->hash_node, hash);
3377 
3378         return pool;
3379 fail:
3380         if (pool)
3381                 put_unbound_pool(pool);
3382         return NULL;
3383 }
3384 
3385 static void rcu_free_pwq(struct rcu_head *rcu)
3386 {
3387         kmem_cache_free(pwq_cache,
3388                         container_of(rcu, struct pool_workqueue, rcu));
3389 }
3390 
3391 /*
3392  * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3393  * and needs to be destroyed.
3394  */
3395 static void pwq_unbound_release_workfn(struct work_struct *work)
3396 {
3397         struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3398                                                   unbound_release_work);
3399         struct workqueue_struct *wq = pwq->wq;
3400         struct worker_pool *pool = pwq->pool;
3401         bool is_last;
3402 
3403         if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3404                 return;
3405 
3406         mutex_lock(&wq->mutex);
3407         list_del_rcu(&pwq->pwqs_node);
3408         is_last = list_empty(&wq->pwqs);
3409         mutex_unlock(&wq->mutex);
3410 
3411         mutex_lock(&wq_pool_mutex);
3412         put_unbound_pool(pool);
3413         mutex_unlock(&wq_pool_mutex);
3414 
3415         call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3416 
3417         /*
3418          * If we're the last pwq going away, @wq is already dead and no one
3419          * is gonna access it anymore.  Schedule RCU free.
3420          */
3421         if (is_last)
3422                 call_rcu_sched(&wq->rcu, rcu_free_wq);
3423 }
3424 
3425 /**
3426  * pwq_adjust_max_active - update a pwq's max_active to the current setting
3427  * @pwq: target pool_workqueue
3428  *
3429  * If @pwq isn't freezing, set @pwq->max_active to the associated
3430  * workqueue's saved_max_active and activate delayed work items
3431  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3432  */
3433 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3434 {
3435         struct workqueue_struct *wq = pwq->wq;
3436         bool freezable = wq->flags & WQ_FREEZABLE;
3437         unsigned long flags;
3438 
3439         /* for @wq->saved_max_active */
3440         lockdep_assert_held(&wq->mutex);
3441 
3442         /* fast exit for non-freezable wqs */
3443         if (!freezable && pwq->max_active == wq->saved_max_active)
3444                 return;
3445 
3446         /* this function can be called during early boot w/ irq disabled */
3447         spin_lock_irqsave(&pwq->pool->lock, flags);
3448 
3449         /*
3450          * During [un]freezing, the caller is responsible for ensuring that
3451          * this function is called at least once after @workqueue_freezing
3452          * is updated and visible.
3453          */
3454         if (!freezable || !workqueue_freezing) {
3455                 pwq->max_active = wq->saved_max_active;
3456 
3457                 while (!list_empty(&pwq->delayed_works) &&
3458                        pwq->nr_active < pwq->max_active)
3459                         pwq_activate_first_delayed(pwq);
3460 
3461                 /*
3462                  * Need to kick a worker after thawed or an unbound wq's
3463                  * max_active is bumped.  It's a slow path.  Do it always.
3464                  */
3465                 wake_up_worker(pwq->pool);
3466         } else {
3467                 pwq->max_active = 0;
3468         }
3469 
3470         spin_unlock_irqrestore(&pwq->pool->lock, flags);
3471 }
3472 
3473 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3474 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3475                      struct worker_pool *pool)
3476 {
3477         BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3478 
3479         memset(pwq, 0, sizeof(*pwq));
3480 
3481         pwq->pool = pool;
3482         pwq->wq = wq;
3483         pwq->flush_color = -1;
3484         pwq->refcnt = 1;
3485         INIT_LIST_HEAD(&pwq->delayed_works);
3486         INIT_LIST_HEAD(&pwq->pwqs_node);
3487         INIT_LIST_HEAD(&pwq->mayday_node);
3488         INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3489 }
3490 
3491 /* sync @pwq with the current state of its associated wq and link it */
3492 static void link_pwq(struct pool_workqueue *pwq)
3493 {
3494         struct workqueue_struct *wq = pwq->wq;
3495 
3496         lockdep_assert_held(&wq->mutex);
3497 
3498         /* may be called multiple times, ignore if already linked */
3499         if (!list_empty(&pwq->pwqs_node))
3500                 return;
3501 
3502         /* set the matching work_color */
3503         pwq->work_color = wq->work_color;
3504 
3505         /* sync max_active to the current setting */
3506         pwq_adjust_max_active(pwq);
3507 
3508         /* link in @pwq */
3509         list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3510 }
3511 
3512 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3513 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3514                                         const struct workqueue_attrs *attrs)
3515 {
3516         struct worker_pool *pool;
3517         struct pool_workqueue *pwq;
3518 
3519         lockdep_assert_held(&wq_pool_mutex);
3520 
3521         pool = get_unbound_pool(attrs);
3522         if (!pool)
3523                 return NULL;
3524 
3525         pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3526         if (!pwq) {
3527                 put_unbound_pool(pool);
3528                 return NULL;
3529         }
3530 
3531         init_pwq(pwq, wq, pool);
3532         return pwq;
3533 }
3534 
3535 /**
3536  * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3537  * @attrs: the wq_attrs of the default pwq of the target workqueue
3538  * @node: the target NUMA node
3539  * @cpu_going_down: if >= 0, the CPU to consider as offline
3540  * @cpumask: outarg, the resulting cpumask
3541  *
3542  * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3543  * @cpu_going_down is >= 0, that cpu is considered offline during
3544  * calculation.  The result is stored in @cpumask.
3545  *
3546  * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3547  * enabled and @node has online CPUs requested by @attrs, the returned
3548  * cpumask is the intersection of the possible CPUs of @node and
3549  * @attrs->cpumask.
3550  *
3551  * The caller is responsible for ensuring that the cpumask of @node stays
3552  * stable.
3553  *
3554  * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3555  * %false if equal.
3556  */
3557 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3558                                  int cpu_going_down, cpumask_t *cpumask)
3559 {
3560         if (!wq_numa_enabled || attrs->no_numa)
3561                 goto use_dfl;
3562 
3563         /* does @node have any online CPUs @attrs wants? */
3564         cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3565         if (cpu_going_down >= 0)
3566                 cpumask_clear_cpu(cpu_going_down, cpumask);
3567 
3568         if (cpumask_empty(cpumask))
3569                 goto use_dfl;
3570 
3571         /* yeap, return possible CPUs in @node that @attrs wants */
3572         cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3573 
3574         if (cpumask_empty(cpumask)) {
3575                 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3576                                 "possible intersect\n");
3577                 return false;
3578         }
3579 
3580         return !cpumask_equal(cpumask, attrs->cpumask);
3581 
3582 use_dfl:
3583         cpumask_copy(cpumask, attrs->cpumask);
3584         return false;
3585 }
3586 
3587 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3589                                                    int node,
3590                                                    struct pool_workqueue *pwq)
3591 {
3592         struct pool_workqueue *old_pwq;
3593 
3594         lockdep_assert_held(&wq_pool_mutex);
3595         lockdep_assert_held(&wq->mutex);
3596 
3597         /* link_pwq() can handle duplicate calls */
3598         link_pwq(pwq);
3599 
3600         old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601         rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602         return old_pwq;
3603 }
3604 
3605 /* context to store the prepared attrs & pwqs before applying */
3606 struct apply_wqattrs_ctx {
3607         struct workqueue_struct *wq;            /* target workqueue */
3608         struct workqueue_attrs  *attrs;         /* attrs to apply */
3609         struct list_head        list;           /* queued for batching commit */
3610         struct pool_workqueue   *dfl_pwq;
3611         struct pool_workqueue   *pwq_tbl[];
3612 };
3613 
3614 /* free the resources after success or abort */
3615 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616 {
3617         if (ctx) {
3618                 int node;
3619 
3620                 for_each_node(node)
3621                         put_pwq_unlocked(ctx->pwq_tbl[node]);
3622                 put_pwq_unlocked(ctx->dfl_pwq);
3623 
3624                 free_workqueue_attrs(ctx->attrs);
3625 
3626                 kfree(ctx);
3627         }
3628 }
3629 
3630 /* allocate the attrs and pwqs for later installation */
3631 static struct apply_wqattrs_ctx *
3632 apply_wqattrs_prepare(struct workqueue_struct *wq,
3633                       const struct workqueue_attrs *attrs)
3634 {
3635         struct apply_wqattrs_ctx *ctx;
3636         struct workqueue_attrs *new_attrs, *tmp_attrs;
3637         int node;
3638 
3639         lockdep_assert_held(&wq_pool_mutex);
3640 
3641         ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3642                       GFP_KERNEL);
3643 
3644         new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645         tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646         if (!ctx || !new_attrs || !tmp_attrs)
3647                 goto out_free;
3648 
3649         /*
3650          * Calculate the attrs of the default pwq.
3651          * If the user configured cpumask doesn't overlap with the
3652          * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653          */
3654         copy_workqueue_attrs(new_attrs, attrs);
3655         cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3656         if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657                 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3658 
3659         /*
3660          * We may create multiple pwqs with differing cpumasks.  Make a
3661          * copy of @new_attrs which will be modified and used to obtain
3662          * pools.
3663          */
3664         copy_workqueue_attrs(tmp_attrs, new_attrs);
3665 
3666         /*
3667          * If something goes wrong during CPU up/down, we'll fall back to
3668          * the default pwq covering whole @attrs->cpumask.  Always create
3669          * it even if we don't use it immediately.
3670          */
3671         ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672         if (!ctx->dfl_pwq)
3673                 goto out_free;
3674 
3675         for_each_node(node) {
3676                 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3677                         ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678                         if (!ctx->pwq_tbl[node])
3679                                 goto out_free;
3680                 } else {
3681                         ctx->dfl_pwq->refcnt++;
3682                         ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683                 }
3684         }
3685 
3686         /* save the user configured attrs and sanitize it. */
3687         copy_workqueue_attrs(new_attrs, attrs);
3688         cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3689         ctx->attrs = new_attrs;
3690 
3691         ctx->wq = wq;
3692         free_workqueue_attrs(tmp_attrs);
3693         return ctx;
3694 
3695 out_free:
3696         free_workqueue_attrs(tmp_attrs);
3697         free_workqueue_attrs(new_attrs);
3698         apply_wqattrs_cleanup(ctx);
3699         return NULL;
3700 }
3701 
3702 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704 {
3705         int node;
3706 
3707         /* all pwqs have been created successfully, let's install'em */
3708         mutex_lock(&ctx->wq->mutex);
3709 
3710         copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711 
3712         /* save the previous pwq and install the new one */
3713         for_each_node(node)
3714                 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715                                                           ctx->pwq_tbl[node]);
3716 
3717         /* @dfl_pwq might not have been used, ensure it's linked */
3718         link_pwq(ctx->dfl_pwq);
3719         swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720 
3721         mutex_unlock(&ctx->wq->mutex);
3722 }
3723 
3724 static void apply_wqattrs_lock(void)
3725 {
3726         /* CPUs should stay stable across pwq creations and installations */
3727         get_online_cpus();
3728         mutex_lock(&wq_pool_mutex);
3729 }
3730 
3731 static void apply_wqattrs_unlock(void)
3732 {
3733         mutex_unlock(&wq_pool_mutex);
3734         put_online_cpus();
3735 }
3736 
3737 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738                                         const struct workqueue_attrs *attrs)
3739 {
3740         struct apply_wqattrs_ctx *ctx;
3741 
3742         /* only unbound workqueues can change attributes */
3743         if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744                 return -EINVAL;
3745 
3746         /* creating multiple pwqs breaks ordering guarantee */
3747         if (!list_empty(&wq->pwqs)) {
3748                 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3749                         return -EINVAL;
3750 
3751                 wq->flags &= ~__WQ_ORDERED;
3752         }
3753 
3754         ctx = apply_wqattrs_prepare(wq, attrs);
3755         if (!ctx)
3756                 return -ENOMEM;
3757 
3758         /* the ctx has been prepared successfully, let's commit it */
3759         apply_wqattrs_commit(ctx);
3760         apply_wqattrs_cleanup(ctx);
3761 
3762         return 0;
3763 }
3764 
3765 /**
3766  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3767  * @wq: the target workqueue
3768  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3769  *
3770  * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3771  * machines, this function maps a separate pwq to each NUMA node with
3772  * possibles CPUs in @attrs->cpumask so that work items are affine to the
3773  * NUMA node it was issued on.  Older pwqs are released as in-flight work
3774  * items finish.  Note that a work item which repeatedly requeues itself
3775  * back-to-back will stay on its current pwq.
3776  *
3777  * Performs GFP_KERNEL allocations.
3778  *
3779  * Return: 0 on success and -errno on failure.
3780  */
3781 int apply_workqueue_attrs(struct workqueue_struct *wq,
3782                           const struct workqueue_attrs *attrs)
3783 {
3784         int ret;
3785 
3786         apply_wqattrs_lock();
3787         ret = apply_workqueue_attrs_locked(wq, attrs);
3788         apply_wqattrs_unlock();
3789 
3790         return ret;
3791 }
3792 
3793 /**
3794  * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3795  * @wq: the target workqueue
3796  * @cpu: the CPU coming up or going down
3797  * @online: whether @cpu is coming up or going down
3798  *
3799  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3800  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3801  * @wq accordingly.
3802  *
3803  * If NUMA affinity can't be adjusted due to memory allocation failure, it
3804  * falls back to @wq->dfl_pwq which may not be optimal but is always
3805  * correct.
3806  *
3807  * Note that when the last allowed CPU of a NUMA node goes offline for a
3808  * workqueue with a cpumask spanning multiple nodes, the workers which were
3809  * already executing the work items for the workqueue will lose their CPU
3810  * affinity and may execute on any CPU.  This is similar to how per-cpu
3811  * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3812  * affinity, it's the user's responsibility to flush the work item from
3813  * CPU_DOWN_PREPARE.
3814  */
3815 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3816                                    bool online)
3817 {
3818         int node = cpu_to_node(cpu);
3819         int cpu_off = online ? -1 : cpu;
3820         struct pool_workqueue *old_pwq = NULL, *pwq;
3821         struct workqueue_attrs *target_attrs;
3822         cpumask_t *cpumask;
3823 
3824         lockdep_assert_held(&wq_pool_mutex);
3825 
3826         if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3827             wq->unbound_attrs->no_numa)
3828                 return;
3829 
3830         /*
3831          * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3832          * Let's use a preallocated one.  The following buf is protected by
3833          * CPU hotplug exclusion.
3834          */
3835         target_attrs = wq_update_unbound_numa_attrs_buf;
3836         cpumask = target_attrs->cpumask;
3837 
3838         copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3839         pwq = unbound_pwq_by_node(wq, node);
3840 
3841         /*
3842          * Let's determine what needs to be done.  If the target cpumask is
3843          * different from the default pwq's, we need to compare it to @pwq's
3844          * and create a new one if they don't match.  If the target cpumask
3845          * equals the default pwq's, the default pwq should be used.
3846          */
3847         if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3848                 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3849                         return;
3850         } else {
3851                 goto use_dfl_pwq;
3852         }
3853 
3854         /* create a new pwq */
3855         pwq = alloc_unbound_pwq(wq, target_attrs);
3856         if (!pwq) {
3857                 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3858                         wq->name);
3859                 goto use_dfl_pwq;
3860         }
3861 
3862         /* Install the new pwq. */
3863         mutex_lock(&wq->mutex);
3864         old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3865         goto out_unlock;
3866 
3867 use_dfl_pwq:
3868         mutex_lock(&wq->mutex);
3869         spin_lock_irq(&wq->dfl_pwq->pool->lock);
3870         get_pwq(wq->dfl_pwq);
3871         spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3872         old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3873 out_unlock:
3874         mutex_unlock(&wq->mutex);
3875         put_pwq_unlocked(old_pwq);
3876 }
3877 
3878 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3879 {
3880         bool highpri = wq->flags & WQ_HIGHPRI;
3881         int cpu, ret;
3882 
3883         if (!(wq->flags & WQ_UNBOUND)) {
3884                 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3885                 if (!wq->cpu_pwqs)
3886                         return -ENOMEM;
3887 
3888                 for_each_possible_cpu(cpu) {
3889                         struct pool_workqueue *pwq =
3890                                 per_cpu_ptr(wq->cpu_pwqs, cpu);
3891                         struct worker_pool *cpu_pools =
3892                                 per_cpu(cpu_worker_pools, cpu);
3893 
3894                         init_pwq(pwq, wq, &cpu_pools[highpri]);
3895 
3896                         mutex_lock(&wq->mutex);
3897                         link_pwq(pwq);
3898                         mutex_unlock(&wq->mutex);
3899                 }
3900                 return 0;
3901         } else if (wq->flags & __WQ_ORDERED) {
3902                 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3903                 /* there should only be single pwq for ordering guarantee */
3904                 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3905                               wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3906                      "ordering guarantee broken for workqueue %s\n", wq->name);
3907                 return ret;
3908         } else {
3909                 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3910         }
3911 }
3912 
3913 static int wq_clamp_max_active(int max_active, unsigned int flags,
3914                                const char *name)
3915 {
3916         int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3917 
3918         if (max_active < 1 || max_active > lim)
3919                 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3920                         max_active, name, 1, lim);
3921 
3922         return clamp_val(max_active, 1, lim);
3923 }
3924 
3925 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3926                                                unsigned int flags,
3927                                                int max_active,
3928                                                struct lock_class_key *key,
3929                                                const char *lock_name, ...)
3930 {
3931         size_t tbl_size = 0;
3932         va_list args;
3933         struct workqueue_struct *wq;
3934         struct pool_workqueue *pwq;
3935 
3936         /*
3937          * Unbound && max_active == 1 used to imply ordered, which is no
3938          * longer the case on NUMA machines due to per-node pools.  While
3939          * alloc_ordered_workqueue() is the right way to create an ordered
3940          * workqueue, keep the previous behavior to avoid subtle breakages
3941          * on NUMA.
3942          */
3943         if ((flags & WQ_UNBOUND) && max_active == 1)
3944                 flags |= __WQ_ORDERED;
3945 
3946         /* see the comment above the definition of WQ_POWER_EFFICIENT */
3947         if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3948                 flags |= WQ_UNBOUND;
3949 
3950         /* allocate wq and format name */
3951         if (flags & WQ_UNBOUND)
3952                 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3953 
3954         wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3955         if (!wq)
3956                 return NULL;
3957 
3958         if (flags & WQ_UNBOUND) {
3959                 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3960                 if (!wq->unbound_attrs)
3961                         goto err_free_wq;
3962         }
3963 
3964         va_start(args, lock_name);
3965         vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3966         va_end(args);
3967 
3968         max_active = max_active ?: WQ_DFL_ACTIVE;
3969         max_active = wq_clamp_max_active(max_active, flags, wq->name);
3970 
3971         /* init wq */
3972         wq->flags = flags;
3973         wq->saved_max_active = max_active;
3974         mutex_init(&wq->mutex);
3975         atomic_set(&wq->nr_pwqs_to_flush, 0);
3976         INIT_LIST_HEAD(&wq->pwqs);
3977         INIT_LIST_HEAD(&wq->flusher_queue);
3978         INIT_LIST_HEAD(&wq->flusher_overflow);
3979         INIT_LIST_HEAD(&wq->maydays);
3980 
3981         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3982         INIT_LIST_HEAD(&wq->list);
3983 
3984         if (alloc_and_link_pwqs(wq) < 0)
3985                 goto err_free_wq;
3986 
3987         /*
3988          * Workqueues which may be used during memory reclaim should
3989          * have a rescuer to guarantee forward progress.
3990          */
3991         if (flags & WQ_MEM_RECLAIM) {
3992                 struct worker *rescuer;
3993 
3994                 rescuer = alloc_worker(NUMA_NO_NODE);
3995                 if (!rescuer)
3996                         goto err_destroy;
3997 
3998                 rescuer->rescue_wq = wq;
3999                 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4000                                                wq->name);
4001                 if (IS_ERR(rescuer->task)) {
4002                         kfree(rescuer);
4003                         goto err_destroy;
4004                 }
4005 
4006                 wq->rescuer = rescuer;
4007                 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4008                 wake_up_process(rescuer->task);
4009         }
4010 
4011         if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4012                 goto err_destroy;
4013 
4014         /*
4015          * wq_pool_mutex protects global freeze state and workqueues list.
4016          * Grab it, adjust max_active and add the new @wq to workqueues
4017          * list.
4018          */
4019         mutex_lock(&wq_pool_mutex);
4020 
4021         mutex_lock(&wq->mutex);
4022         for_each_pwq(pwq, wq)
4023                 pwq_adjust_max_active(pwq);
4024         mutex_unlock(&wq->mutex);
4025 
4026         list_add_tail_rcu(&wq->list, &workqueues);
4027 
4028         mutex_unlock(&wq_pool_mutex);
4029 
4030         return wq;
4031 
4032 err_free_wq:
4033         free_workqueue_attrs(wq->unbound_attrs);
4034         kfree(wq);
4035         return NULL;
4036 err_destroy:
4037         destroy_workqueue(wq);
4038         return NULL;
4039 }
4040 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4041 
4042 /**
4043  * destroy_workqueue - safely terminate a workqueue
4044  * @wq: target workqueue
4045  *
4046  * Safely destroy a workqueue. All work currently pending will be done first.
4047  */
4048 void destroy_workqueue(struct workqueue_struct *wq)
4049 {
4050         struct pool_workqueue *pwq;
4051         int node;
4052 
4053         /* drain it before proceeding with destruction */
4054         drain_workqueue(wq);
4055 
4056         /* sanity checks */
4057         mutex_lock(&wq->mutex);
4058         for_each_pwq(pwq, wq) {
4059                 int i;
4060 
4061                 for (i = 0; i < WORK_NR_COLORS; i++) {
4062                         if (WARN_ON(pwq->nr_in_flight[i])) {
4063                                 mutex_unlock(&wq->mutex);
4064                                 show_workqueue_state();
4065                                 return;
4066                         }
4067                 }
4068 
4069                 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4070                     WARN_ON(pwq->nr_active) ||
4071                     WARN_ON(!list_empty(&pwq->delayed_works))) {
4072                         mutex_unlock(&wq->mutex);
4073                         show_workqueue_state();
4074                         return;
4075                 }
4076         }
4077         mutex_unlock(&wq->mutex);
4078 
4079         /*
4080          * wq list is used to freeze wq, remove from list after
4081          * flushing is complete in case freeze races us.
4082          */
4083         mutex_lock(&wq_pool_mutex);
4084         list_del_rcu(&wq->list);
4085         mutex_unlock(&wq_pool_mutex);
4086 
4087         workqueue_sysfs_unregister(wq);
4088 
4089         if (wq->rescuer)
4090                 kthread_stop(wq->rescuer->task);
4091 
4092         if (!(wq->flags & WQ_UNBOUND)) {
4093                 /*
4094                  * The base ref is never dropped on per-cpu pwqs.  Directly
4095                  * schedule RCU free.
4096                  */
4097                 call_rcu_sched(&wq->rcu, rcu_free_wq);
4098         } else {
4099                 /*
4100                  * We're the sole accessor of @wq at this point.  Directly
4101                  * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4102                  * @wq will be freed when the last pwq is released.
4103                  */
4104                 for_each_node(node) {
4105                         pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4106                         RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4107                         put_pwq_unlocked(pwq);
4108                 }
4109 
4110                 /*
4111                  * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4112                  * put.  Don't access it afterwards.
4113                  */
4114                 pwq = wq->dfl_pwq;
4115                 wq->dfl_pwq = NULL;
4116                 put_pwq_unlocked(pwq);
4117         }
4118 }
4119 EXPORT_SYMBOL_GPL(destroy_workqueue);
4120 
4121 /**
4122  * workqueue_set_max_active - adjust max_active of a workqueue
4123  * @wq: target workqueue
4124  * @max_active: new max_active value.
4125  *
4126  * Set max_active of @wq to @max_active.
4127  *
4128  * CONTEXT:
4129  * Don't call from IRQ context.
4130  */
4131 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4132 {
4133         struct pool_workqueue *pwq;
4134 
4135         /* disallow meddling with max_active for ordered workqueues */
4136         if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4137                 return;
4138 
4139         max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4140 
4141         mutex_lock(&wq->mutex);
4142 
4143         wq->flags &= ~__WQ_ORDERED;
4144         wq->saved_max_active = max_active;
4145 
4146         for_each_pwq(pwq, wq)
4147                 pwq_adjust_max_active(pwq);
4148 
4149         mutex_unlock(&wq->mutex);
4150 }
4151 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4152 
4153 /**
4154  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4155  *
4156  * Determine whether %current is a workqueue rescuer.  Can be used from
4157  * work functions to determine whether it's being run off the rescuer task.
4158  *
4159  * Return: %true if %current is a workqueue rescuer. %false otherwise.
4160  */
4161 bool current_is_workqueue_rescuer(void)
4162 {
4163         struct worker *worker = current_wq_worker();
4164 
4165         return worker && worker->rescue_wq;
4166 }
4167 
4168 /**
4169  * workqueue_congested - test whether a workqueue is congested
4170  * @cpu: CPU in question
4171  * @wq: target workqueue
4172  *
4173  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4174  * no synchronization around this function and the test result is
4175  * unreliable and only useful as advisory hints or for debugging.
4176  *
4177  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4178  * Note that both per-cpu and unbound workqueues may be associated with
4179  * multiple pool_workqueues which have separate congested states.  A
4180  * workqueue being congested on one CPU doesn't mean the workqueue is also
4181  * contested on other CPUs / NUMA nodes.
4182  *
4183  * Return:
4184  * %true if congested, %false otherwise.
4185  */
4186 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4187 {
4188         struct pool_workqueue *pwq;
4189         bool ret;
4190 
4191         rcu_read_lock_sched();
4192 
4193         if (cpu == WORK_CPU_UNBOUND)
4194                 cpu = smp_processor_id();
4195 
4196         if (!(wq->flags & WQ_UNBOUND))
4197                 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4198         else
4199                 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4200 
4201         ret = !list_empty(&pwq->delayed_works);
4202         rcu_read_unlock_sched();
4203 
4204         return ret;
4205 }
4206 EXPORT_SYMBOL_GPL(workqueue_congested);
4207 
4208 /**
4209  * work_busy - test whether a work is currently pending or running
4210  * @work: the work to be tested
4211  *
4212  * Test whether @work is currently pending or running.  There is no
4213  * synchronization around this function and the test result is
4214  * unreliable and only useful as advisory hints or for debugging.
4215  *
4216  * Return:
4217  * OR'd bitmask of WORK_BUSY_* bits.
4218  */
4219 unsigned int work_busy(struct work_struct *work)
4220 {
4221         struct worker_pool *pool;
4222         unsigned long flags;
4223         unsigned int ret = 0;
4224 
4225         if (work_pending(work))
4226                 ret |= WORK_BUSY_PENDING;
4227 
4228         local_irq_save(flags);
4229         pool = get_work_pool(work);
4230         if (pool) {
4231                 spin_lock(&pool->lock);
4232                 if (find_worker_executing_work(pool, work))
4233                         ret |= WORK_BUSY_RUNNING;
4234                 spin_unlock(&pool->lock);
4235         }
4236         local_irq_restore(flags);
4237 
4238         return ret;
4239 }
4240 EXPORT_SYMBOL_GPL(work_busy);
4241 
4242 /**
4243  * set_worker_desc - set description for the current work item
4244  * @fmt: printf-style format string
4245  * @...: arguments for the format string
4246  *
4247  * This function can be called by a running work function to describe what
4248  * the work item is about.  If the worker task gets dumped, this
4249  * information will be printed out together to help debugging.  The
4250  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4251  */
4252 void set_worker_desc(const char *fmt, ...)
4253 {
4254         struct worker *worker = current_wq_worker();
4255         va_list args;
4256 
4257         if (worker) {
4258                 va_start(args, fmt);
4259                 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4260                 va_end(args);
4261                 worker->desc_valid = true;
4262         }
4263 }
4264 
4265 /**
4266  * print_worker_info - print out worker information and description
4267  * @log_lvl: the log level to use when printing
4268  * @task: target task
4269  *
4270  * If @task is a worker and currently executing a work item, print out the
4271  * name of the workqueue being serviced and worker description set with
4272  * set_worker_desc() by the currently executing work item.
4273  *
4274  * This function can be safely called on any task as long as the
4275  * task_struct itself is accessible.  While safe, this function isn't
4276  * synchronized and may print out mixups or garbages of limited length.
4277  */
4278 void print_worker_info(const char *log_lvl, struct task_struct *task)
4279 {
4280         work_func_t *fn = NULL;
4281         char name[WQ_NAME_LEN] = { };
4282         char desc[WORKER_DESC_LEN] = { };
4283         struct pool_workqueue *pwq = NULL;
4284         struct workqueue_struct *wq = NULL;
4285         bool desc_valid = false;
4286         struct worker *worker;
4287 
4288         if (!(task->flags & PF_WQ_WORKER))
4289                 return;
4290 
4291         /*
4292          * This function is called without any synchronization and @task
4293          * could be in any state.  Be careful with dereferences.
4294          */
4295         worker = kthread_probe_data(task);
4296 
4297         /*
4298          * Carefully copy the associated workqueue's workfn and name.  Keep
4299          * the original last '\0' in case the original contains garbage.
4300          */
4301         probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4302         probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4303         probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4304         probe_kernel_read(name, wq->name, sizeof(name) - 1);
4305 
4306         /* copy worker description */
4307         probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4308         if (desc_valid)
4309                 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4310 
4311         if (fn || name[0] || desc[0]) {
4312                 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4313                 if (desc[0])
4314                         pr_cont(" (%s)", desc);
4315                 pr_cont("\n");
4316         }
4317 }
4318 
4319 static void pr_cont_pool_info(struct worker_pool *pool)
4320 {
4321         pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4322         if (pool->node != NUMA_NO_NODE)
4323                 pr_cont(" node=%d", pool->node);
4324         pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4325 }
4326 
4327 static void pr_cont_work(bool comma, struct work_struct *work)
4328 {
4329         if (work->func == wq_barrier_func) {
4330                 struct wq_barrier *barr;
4331 
4332                 barr = container_of(work, struct wq_barrier, work);
4333 
4334                 pr_cont("%s BAR(%d)", comma ? "," : "",
4335                         task_pid_nr(barr->task));
4336         } else {
4337                 pr_cont("%s %pf", comma ? "," : "", work->func);
4338         }
4339 }
4340 
4341 static void show_pwq(struct pool_workqueue *pwq)
4342 {
4343         struct worker_pool *pool = pwq->pool;
4344         struct work_struct *work;
4345         struct worker *worker;
4346         bool has_in_flight = false, has_pending = false;
4347         int bkt;
4348 
4349         pr_info("  pwq %d:", pool->id);
4350         pr_cont_pool_info(pool);
4351 
4352         pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4353                 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4354 
4355         hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4356                 if (worker->current_pwq == pwq) {
4357                         has_in_flight = true;
4358                         break;
4359                 }
4360         }
4361         if (has_in_flight) {
4362                 bool comma = false;
4363 
4364                 pr_info("    in-flight:");
4365                 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4366                         if (worker->current_pwq != pwq)
4367                                 continue;
4368 
4369                         pr_cont("%s %d%s:%pf", comma ? "," : "",
4370                                 task_pid_nr(worker->task),
4371                                 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4372                                 worker->current_func);
4373                         list_for_each_entry(work, &worker->scheduled, entry)
4374                                 pr_cont_work(false, work);
4375                         comma = true;
4376                 }
4377                 pr_cont("\n");
4378         }
4379 
4380         list_for_each_entry(work, &pool->worklist, entry) {
4381                 if (get_work_pwq(work) == pwq) {
4382                         has_pending = true;
4383                         break;
4384                 }
4385         }
4386         if (has_pending) {
4387                 bool comma = false;
4388 
4389                 pr_info("    pending:");
4390                 list_for_each_entry(work, &pool->worklist, entry) {
4391                         if (get_work_pwq(work) != pwq)
4392                                 continue;
4393 
4394                         pr_cont_work(comma, work);
4395                         comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4396                 }
4397                 pr_cont("\n");
4398         }
4399 
4400         if (!list_empty(&pwq->delayed_works)) {
4401                 bool comma = false;
4402 
4403                 pr_info("    delayed:");
4404                 list_for_each_entry(work, &pwq->delayed_works, entry) {
4405                         pr_cont_work(comma, work);
4406                         comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4407                 }
4408                 pr_cont("\n");
4409         }
4410 }
4411 
4412 /**
4413  * show_workqueue_state - dump workqueue state
4414  *
4415  * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4416  * all busy workqueues and pools.
4417  */
4418 void show_workqueue_state(void)
4419 {
4420         struct workqueue_struct *wq;
4421         struct worker_pool *pool;
4422         unsigned long flags;
4423         int pi;
4424 
4425         rcu_read_lock_sched();
4426 
4427         pr_info("Showing busy workqueues and worker pools:\n");
4428 
4429         list_for_each_entry_rcu(wq, &workqueues, list) {
4430                 struct pool_workqueue *pwq;
4431                 bool idle = true;
4432 
4433                 for_each_pwq(pwq, wq) {
4434                         if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4435                                 idle = false;
4436                                 break;
4437                         }
4438                 }
4439                 if (idle)
4440                         continue;
4441 
4442                 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4443 
4444                 for_each_pwq(pwq, wq) {
4445                         spin_lock_irqsave(&pwq->pool->lock, flags);
4446                         if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4447                                 show_pwq(pwq);
4448                         spin_unlock_irqrestore(&pwq->pool->lock, flags);
4449                 }
4450         }
4451 
4452         for_each_pool(pool, pi) {
4453                 struct worker *worker;
4454                 bool first = true;
4455 
4456                 spin_lock_irqsave(&pool->lock, flags);
4457                 if (pool->nr_workers == pool->nr_idle)
4458                         goto next_pool;
4459 
4460                 pr_info("pool %d:", pool->id);
4461                 pr_cont_pool_info(pool);
4462                 pr_cont(" hung=%us workers=%d",
4463                         jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4464                         pool->nr_workers);
4465                 if (pool->manager)
4466                         pr_cont(" manager: %d",
4467                                 task_pid_nr(pool->manager->task));
4468                 list_for_each_entry(worker, &pool->idle_list, entry) {
4469                         pr_cont(" %s%d", first ? "idle: " : "",
4470                                 task_pid_nr(worker->task));
4471                         first = false;
4472                 }
4473                 pr_cont("\n");
4474         next_pool:
4475                 spin_unlock_irqrestore(&pool->lock, flags);
4476         }
4477 
4478         rcu_read_unlock_sched();
4479 }
4480 
4481 /*
4482  * CPU hotplug.
4483  *
4484  * There are two challenges in supporting CPU hotplug.  Firstly, there
4485  * are a lot of assumptions on strong associations among work, pwq and
4486  * pool which make migrating pending and scheduled works very
4487  * difficult to implement without impacting hot paths.  Secondly,
4488  * worker pools serve mix of short, long and very long running works making
4489  * blocked draining impractical.
4490  *
4491  * This is solved by allowing the pools to be disassociated from the CPU
4492  * running as an unbound one and allowing it to be reattached later if the
4493  * cpu comes back online.
4494  */
4495 
4496 static void wq_unbind_fn(struct work_struct *work)
4497 {
4498         int cpu = smp_processor_id();
4499         struct worker_pool *pool;
4500         struct worker *worker;
4501 
4502         for_each_cpu_worker_pool(pool, cpu) {
4503                 mutex_lock(&pool->attach_mutex);
4504                 spin_lock_irq(&pool->lock);
4505 
4506                 /*
4507                  * We've blocked all attach/detach operations. Make all workers
4508                  * unbound and set DISASSOCIATED.  Before this, all workers
4509                  * except for the ones which are still executing works from
4510                  * before the last CPU down must be on the cpu.  After
4511                  * this, they may become diasporas.
4512                  */
4513                 for_each_pool_worker(worker, pool)
4514                         worker->flags |= WORKER_UNBOUND;
4515 
4516                 pool->flags |= POOL_DISASSOCIATED;
4517 
4518                 spin_unlock_irq(&pool->lock);
4519                 mutex_unlock(&pool->attach_mutex);
4520 
4521                 /*
4522                  * Call schedule() so that we cross rq->lock and thus can
4523                  * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4524                  * This is necessary as scheduler callbacks may be invoked
4525                  * from other cpus.
4526                  */
4527                 schedule();
4528 
4529                 /*
4530                  * Sched callbacks are disabled now.  Zap nr_running.
4531                  * After this, nr_running stays zero and need_more_worker()
4532                  * and keep_working() are always true as long as the
4533                  * worklist is not empty.  This pool now behaves as an
4534                  * unbound (in terms of concurrency management) pool which
4535                  * are served by workers tied to the pool.
4536                  */
4537                 atomic_set(&pool->nr_running, 0);
4538 
4539                 /*
4540                  * With concurrency management just turned off, a busy
4541                  * worker blocking could lead to lengthy stalls.  Kick off
4542                  * unbound chain execution of currently pending work items.
4543                  */
4544                 spin_lock_irq(&pool->lock);
4545                 wake_up_worker(pool);
4546                 spin_unlock_irq(&pool->lock);
4547         }
4548 }
4549 
4550 /**
4551  * rebind_workers - rebind all workers of a pool to the associated CPU
4552  * @pool: pool of interest
4553  *
4554  * @pool->cpu is coming online.  Rebind all workers to the CPU.
4555  */
4556 static void rebind_workers(struct worker_pool *pool)
4557 {
4558         struct worker *worker;
4559 
4560         lockdep_assert_held(&pool->attach_mutex);
4561 
4562         /*
4563          * Restore CPU affinity of all workers.  As all idle workers should
4564          * be on the run-queue of the associated CPU before any local
4565          * wake-ups for concurrency management happen, restore CPU affinity
4566          * of all workers first and then clear UNBOUND.  As we're called
4567          * from CPU_ONLINE, the following shouldn't fail.
4568          */
4569         for_each_pool_worker(worker, pool)
4570                 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4571                                                   pool->attrs->cpumask) < 0);
4572 
4573         spin_lock_irq(&pool->lock);
4574 
4575         /*
4576          * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4577          * w/o preceding DOWN_PREPARE.  Work around it.  CPU hotplug is
4578          * being reworked and this can go away in time.
4579          */
4580         if (!(pool->flags & POOL_DISASSOCIATED)) {
4581                 spin_unlock_irq(&pool->lock);
4582                 return;
4583         }
4584 
4585         pool->flags &= ~POOL_DISASSOCIATED;
4586 
4587         for_each_pool_worker(worker, pool) {
4588                 unsigned int worker_flags = worker->flags;
4589 
4590                 /*
4591                  * A bound idle worker should actually be on the runqueue
4592                  * of the associated CPU for local wake-ups targeting it to
4593                  * work.  Kick all idle workers so that they migrate to the
4594                  * associated CPU.  Doing this in the same loop as
4595                  * replacing UNBOUND with REBOUND is safe as no worker will
4596                  * be bound before @pool->lock is released.
4597                  */
4598                 if (worker_flags & WORKER_IDLE)
4599                         wake_up_process(worker->task);
4600 
4601                 /*
4602                  * We want to clear UNBOUND but can't directly call
4603                  * worker_clr_flags() or adjust nr_running.  Atomically
4604                  * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4605                  * @worker will clear REBOUND using worker_clr_flags() when
4606                  * it initiates the next execution cycle thus restoring
4607                  * concurrency management.  Note that when or whether
4608                  * @worker clears REBOUND doesn't affect correctness.
4609                  *
4610                  * ACCESS_ONCE() is necessary because @worker->flags may be
4611                  * tested without holding any lock in
4612                  * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4613                  * fail incorrectly leading to premature concurrency
4614                  * management operations.
4615                  */
4616                 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4617                 worker_flags |= WORKER_REBOUND;
4618                 worker_flags &= ~WORKER_UNBOUND;
4619                 ACCESS_ONCE(worker->flags) = worker_flags;
4620         }
4621 
4622         spin_unlock_irq(&pool->lock);
4623 }
4624 
4625 /**
4626  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4627  * @pool: unbound pool of interest
4628  * @cpu: the CPU which is coming up
4629  *
4630  * An unbound pool may end up with a cpumask which doesn't have any online
4631  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4632  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4633  * online CPU before, cpus_allowed of all its workers should be restored.
4634  */
4635 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4636 {
4637         static cpumask_t cpumask;
4638         struct worker *worker;
4639 
4640         lockdep_assert_held(&pool->attach_mutex);
4641 
4642         /* is @cpu allowed for @pool? */
4643         if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4644                 return;
4645 
4646         cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4647 
4648         /* as we're called from CPU_ONLINE, the following shouldn't fail */
4649         for_each_pool_worker(worker, pool)
4650                 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4651 }
4652 
4653 int workqueue_prepare_cpu(unsigned int cpu)
4654 {
4655         struct worker_pool *pool;
4656 
4657         for_each_cpu_worker_pool(pool, cpu) {
4658                 if (pool->nr_workers)
4659                         continue;
4660                 if (!create_worker(pool))
4661                         return -ENOMEM;
4662         }
4663         return 0;
4664 }
4665 
4666 int workqueue_online_cpu(unsigned int cpu)
4667 {
4668         struct worker_pool *pool;
4669         struct workqueue_struct *wq;
4670         int pi;
4671 
4672         mutex_lock(&wq_pool_mutex);
4673 
4674         for_each_pool(pool, pi) {
4675                 mutex_lock(&pool->attach_mutex);
4676 
4677                 if (pool->cpu == cpu)
4678                         rebind_workers(pool);
4679                 else if (pool->cpu < 0)
4680                         restore_unbound_workers_cpumask(pool, cpu);
4681 
4682                 mutex_unlock(&pool->attach_mutex);
4683         }
4684 
4685         /* update NUMA affinity of unbound workqueues */
4686         list_for_each_entry(wq, &workqueues, list)
4687                 wq_update_unbound_numa(wq, cpu, true);
4688 
4689         mutex_unlock(&wq_pool_mutex);
4690         return 0;
4691 }
4692 
4693 int workqueue_offline_cpu(unsigned int cpu)
4694 {
4695         struct work_struct unbind_work;
4696         struct workqueue_struct *wq;
4697 
4698         /* unbinding per-cpu workers should happen on the local CPU */
4699         INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4700         queue_work_on(cpu, system_highpri_wq, &unbind_work);
4701 
4702         /* update NUMA affinity of unbound workqueues */
4703         mutex_lock(&wq_pool_mutex);
4704         list_for_each_entry(wq, &workqueues, list)
4705                 wq_update_unbound_numa(wq, cpu, false);
4706         mutex_unlock(&wq_pool_mutex);
4707 
4708         /* wait for per-cpu unbinding to finish */
4709         flush_work(&unbind_work);
4710         destroy_work_on_stack(&unbind_work);
4711         return 0;
4712 }
4713 
4714 #ifdef CONFIG_SMP
4715 
4716 struct work_for_cpu {
4717         struct work_struct work;
4718         long (*fn)(void *);
4719         void *arg;
4720         long ret;
4721 };
4722 
4723 static void work_for_cpu_fn(struct work_struct *work)
4724 {
4725         struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4726 
4727         wfc->ret = wfc->fn(wfc->arg);
4728 }
4729 
4730 /**
4731  * work_on_cpu - run a function in thread context on a particular cpu
4732  * @cpu: the cpu to run on
4733  * @fn: the function to run
4734  * @arg: the function arg
4735  *
4736  * It is up to the caller to ensure that the cpu doesn't go offline.
4737  * The caller must not hold any locks which would prevent @fn from completing.
4738  *
4739  * Return: The value @fn returns.
4740  */
4741 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4742 {
4743         struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4744 
4745         INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4746         schedule_work_on(cpu, &wfc.work);
4747         flush_work(&wfc.work);
4748         destroy_work_on_stack(&wfc.work);
4749         return wfc.ret;
4750 }
4751 EXPORT_SYMBOL_GPL(work_on_cpu);
4752 
4753 /**
4754  * work_on_cpu_safe - run a function in thread context on a particular cpu
4755  * @cpu: the cpu to run on
4756  * @fn:  the function to run
4757  * @arg: the function argument
4758  *
4759  * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4760  * any locks which would prevent @fn from completing.
4761  *
4762  * Return: The value @fn returns.
4763  */
4764 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4765 {
4766         long ret = -ENODEV;
4767 
4768         get_online_cpus();
4769         if (cpu_online(cpu))
4770                 ret = work_on_cpu(cpu, fn, arg);
4771         put_online_cpus();
4772         return ret;
4773 }
4774 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
4775 #endif /* CONFIG_SMP */
4776 
4777 #ifdef CONFIG_FREEZER
4778 
4779 /**
4780  * freeze_workqueues_begin - begin freezing workqueues
4781  *
4782  * Start freezing workqueues.  After this function returns, all freezable
4783  * workqueues will queue new works to their delayed_works list instead of
4784  * pool->worklist.
4785  *
4786  * CONTEXT:
4787  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4788  */
4789 void freeze_workqueues_begin(void)
4790 {
4791         struct workqueue_struct *wq;
4792         struct pool_workqueue *pwq;
4793 
4794         mutex_lock(&wq_pool_mutex);
4795 
4796         WARN_ON_ONCE(workqueue_freezing);
4797         workqueue_freezing = true;
4798 
4799         list_for_each_entry(wq, &workqueues, list) {
4800                 mutex_lock(&wq->mutex);
4801                 for_each_pwq(pwq, wq)
4802                         pwq_adjust_max_active(pwq);
4803                 mutex_unlock(&wq->mutex);
4804         }
4805 
4806         mutex_unlock(&wq_pool_mutex);
4807 }
4808 
4809 /**
4810  * freeze_workqueues_busy - are freezable workqueues still busy?
4811  *
4812  * Check whether freezing is complete.  This function must be called
4813  * between freeze_workqueues_begin() and thaw_workqueues().
4814  *
4815  * CONTEXT:
4816  * Grabs and releases wq_pool_mutex.
4817  *
4818  * Return:
4819  * %true if some freezable workqueues are still busy.  %false if freezing
4820  * is complete.
4821  */
4822 bool freeze_workqueues_busy(void)
4823 {
4824         bool busy = false;
4825         struct workqueue_struct *wq;
4826         struct pool_workqueue *pwq;
4827 
4828         mutex_lock(&wq_pool_mutex);
4829 
4830         WARN_ON_ONCE(!workqueue_freezing);
4831 
4832         list_for_each_entry(wq, &workqueues, list) {
4833                 if (!(wq->flags & WQ_FREEZABLE))
4834                         continue;
4835                 /*
4836                  * nr_active is monotonically decreasing.  It's safe
4837                  * to peek without lock.
4838                  */
4839                 rcu_read_lock_sched();
4840                 for_each_pwq(pwq, wq) {
4841                         WARN_ON_ONCE(pwq->nr_active < 0);
4842                         if (pwq->nr_active) {
4843                                 busy = true;
4844                                 rcu_read_unlock_sched();
4845                                 goto out_unlock;
4846                         }
4847                 }
4848                 rcu_read_unlock_sched();
4849         }
4850 out_unlock:
4851         mutex_unlock(&wq_pool_mutex);
4852         return busy;
4853 }
4854 
4855 /**
4856  * thaw_workqueues - thaw workqueues
4857  *
4858  * Thaw workqueues.  Normal queueing is restored and all collected
4859  * frozen works are transferred to their respective pool worklists.
4860  *
4861  * CONTEXT:
4862  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4863  */
4864 void thaw_workqueues(void)
4865 {
4866         struct workqueue_struct *wq;
4867         struct pool_workqueue *pwq;
4868 
4869         mutex_lock(&wq_pool_mutex);
4870 
4871         if (!workqueue_freezing)
4872                 goto out_unlock;
4873 
4874         workqueue_freezing = false;
4875 
4876         /* restore max_active and repopulate worklist */
4877         list_for_each_entry(wq, &workqueues, list) {
4878                 mutex_lock(&wq->mutex);
4879                 for_each_pwq(pwq, wq)
4880                         pwq_adjust_max_active(pwq);
4881                 mutex_unlock(&wq->mutex);
4882         }
4883 
4884 out_unlock:
4885         mutex_unlock(&wq_pool_mutex);
4886 }
4887 #endif /* CONFIG_FREEZER */
4888 
4889 static int workqueue_apply_unbound_cpumask(void)
4890 {
4891         LIST_HEAD(ctxs);
4892         int ret = 0;
4893         struct workqueue_struct *wq;
4894         struct apply_wqattrs_ctx *ctx, *n;
4895 
4896         lockdep_assert_held(&wq_pool_mutex);
4897 
4898         list_for_each_entry(wq, &workqueues, list) {
4899                 if (!(wq->flags & WQ_UNBOUND))
4900                         continue;
4901                 /* creating multiple pwqs breaks ordering guarantee */
4902                 if (wq->flags & __WQ_ORDERED)
4903                         continue;
4904 
4905                 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4906                 if (!ctx) {
4907                         ret = -ENOMEM;
4908                         break;
4909                 }
4910 
4911                 list_add_tail(&ctx->list, &ctxs);
4912         }
4913 
4914         list_for_each_entry_safe(ctx, n, &ctxs, list) {
4915                 if (!ret)
4916                         apply_wqattrs_commit(ctx);
4917                 apply_wqattrs_cleanup(ctx);
4918         }
4919 
4920         return ret;
4921 }
4922 
4923 /**
4924  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4925  *  @cpumask: the cpumask to set
4926  *
4927  *  The low-level workqueues cpumask is a global cpumask that limits
4928  *  the affinity of all unbound workqueues.  This function check the @cpumask
4929  *  and apply it to all unbound workqueues and updates all pwqs of them.
4930  *
4931  *  Retun:      0       - Success
4932  *              -EINVAL - Invalid @cpumask
4933  *              -ENOMEM - Failed to allocate memory for attrs or pwqs.
4934  */
4935 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4936 {
4937         int ret = -EINVAL;
4938         cpumask_var_t saved_cpumask;
4939 
4940         if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4941                 return -ENOMEM;
4942 
4943         cpumask_and(cpumask, cpumask, cpu_possible_mask);
4944         if (!cpumask_empty(cpumask)) {
4945                 apply_wqattrs_lock();
4946 
4947                 /* save the old wq_unbound_cpumask. */
4948                 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4949 
4950                 /* update wq_unbound_cpumask at first and apply it to wqs. */
4951                 cpumask_copy(wq_unbound_cpumask, cpumask);
4952                 ret = workqueue_apply_unbound_cpumask();
4953 
4954                 /* restore the wq_unbound_cpumask when failed. */
4955                 if (ret < 0)
4956                         cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4957 
4958                 apply_wqattrs_unlock();
4959         }
4960 
4961         free_cpumask_var(saved_cpumask);
4962         return ret;
4963 }
4964 
4965 #ifdef CONFIG_SYSFS
4966 /*
4967  * Workqueues with WQ_SYSFS flag set is visible to userland via
4968  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
4969  * following attributes.
4970  *
4971  *  per_cpu     RO bool : whether the workqueue is per-cpu or unbound
4972  *  max_active  RW int  : maximum number of in-flight work items
4973  *
4974  * Unbound workqueues have the following extra attributes.
4975  *
4976  *  id          RO int  : the associated pool ID
4977  *  nice        RW int  : nice value of the workers
4978  *  cpumask     RW mask : bitmask of allowed CPUs for the workers
4979  */
4980 struct wq_device {
4981         struct workqueue_struct         *wq;
4982         struct device                   dev;
4983 };
4984 
4985 static struct workqueue_struct *dev_to_wq(struct device *dev)
4986 {
4987         struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4988 
4989         return wq_dev->wq;
4990 }
4991 
4992 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4993                             char *buf)
4994 {
4995         struct workqueue_struct *wq = dev_to_wq(dev);
4996 
4997         return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4998 }
4999 static DEVICE_ATTR_RO(per_cpu);
5000 
5001 static ssize_t max_active_show(struct device *dev,
5002                                struct device_attribute *attr, char *buf)
5003 {
5004         struct workqueue_struct *wq = dev_to_wq(dev);
5005 
5006         return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5007 }
5008 
5009 static ssize_t max_active_store(struct device *dev,
5010                                 struct device_attribute *attr, const char *buf,
5011                                 size_t count)
5012 {
5013         struct workqueue_struct *wq = dev_to_wq(dev);
5014         int val;
5015 
5016         if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5017                 return -EINVAL;
5018 
5019         workqueue_set_max_active(wq, val);
5020         return count;
5021 }
5022 static DEVICE_ATTR_RW(max_active);
5023 
5024 static struct attribute *wq_sysfs_attrs[] = {
5025         &dev_attr_per_cpu.attr,
5026         &dev_attr_max_active.attr,
5027         NULL,
5028 };
5029 ATTRIBUTE_GROUPS(wq_sysfs);
5030 
5031 static ssize_t wq_pool_ids_show(struct device *dev,
5032                                 struct device_attribute *attr, char *buf)
5033 {
5034         struct workqueue_struct *wq = dev_to_wq(dev);
5035         const char *delim = "";
5036         int node, written = 0;
5037 
5038         rcu_read_lock_sched();
5039         for_each_node(node) {
5040                 written += scnprintf(buf + written, PAGE_SIZE - written,
5041                                      "%s%d:%d", delim, node,
5042                                      unbound_pwq_by_node(wq, node)->pool->id);
5043                 delim = " ";
5044         }
5045         written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5046         rcu_read_unlock_sched();
5047 
5048         return written;
5049 }
5050 
5051 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5052                             char *buf)
5053 {
5054         struct workqueue_struct *wq = dev_to_wq(dev);
5055         int written;
5056 
5057         mutex_lock(&wq->mutex);
5058         written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5059         mutex_unlock(&wq->mutex);
5060 
5061         return written;
5062 }
5063 
5064 /* prepare workqueue_attrs for sysfs store operations */
5065 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5066 {
5067         struct workqueue_attrs *attrs;
5068 
5069         lockdep_assert_held(&wq_pool_mutex);
5070 
5071         attrs = alloc_workqueue_attrs(GFP_KERNEL);
5072         if (!attrs)
5073                 return NULL;
5074 
5075         copy_workqueue_attrs(attrs, wq->unbound_attrs);
5076         return attrs;
5077 }
5078 
5079 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5080                              const char *buf, size_t count)
5081 {
5082         struct workqueue_struct *wq = dev_to_wq(dev);
5083         struct workqueue_attrs *attrs;
5084         int ret = -ENOMEM;
5085 
5086         apply_wqattrs_lock();
5087 
5088         attrs = wq_sysfs_prep_attrs(wq);
5089         if (!attrs)
5090                 goto out_unlock;
5091 
5092         if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5093             attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5094                 ret = apply_workqueue_attrs_locked(wq, attrs);
5095         else
5096                 ret = -EINVAL;
5097 
5098 out_unlock:
5099         apply_wqattrs_unlock();
5100         free_workqueue_attrs(attrs);
5101         return ret ?: count;
5102 }
5103 
5104 static ssize_t wq_cpumask_show(struct device *dev,
5105                                struct device_attribute *attr, char *buf)
5106 {
5107         struct workqueue_struct *wq = dev_to_wq(dev);
5108         int written;
5109 
5110         mutex_lock(&wq->mutex);
5111         written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5112                             cpumask_pr_args(wq->unbound_attrs->cpumask));
5113         mutex_unlock(&wq->mutex);
5114         return written;
5115 }
5116 
5117 static ssize_t wq_cpumask_store(struct device *dev,
5118                                 struct device_attribute *attr,
5119                                 const char *buf, size_t count)
5120 {
5121         struct workqueue_struct *wq = dev_to_wq(dev);
5122         struct workqueue_attrs *attrs;
5123         int ret = -ENOMEM;
5124 
5125         apply_wqattrs_lock();
5126 
5127         attrs = wq_sysfs_prep_attrs(wq);
5128         if (!attrs)
5129                 goto out_unlock;
5130 
5131         ret = cpumask_parse(buf, attrs->cpumask);
5132         if (!ret)
5133                 ret = apply_workqueue_attrs_locked(wq, attrs);
5134 
5135 out_unlock:
5136         apply_wqattrs_unlock();
5137         free_workqueue_attrs(attrs);
5138         return ret ?: count;
5139 }
5140 
5141 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5142                             char *buf)
5143 {
5144         struct workqueue_struct *wq = dev_to_wq(dev);
5145         int written;
5146 
5147         mutex_lock(&wq->mutex);
5148         written = scnprintf(buf, PAGE_SIZE, "%d\n",
5149                             !wq->unbound_attrs->no_numa);
5150         mutex_unlock(&wq->mutex);
5151 
5152         return written;
5153 }
5154 
5155 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5156                              const char *buf, size_t count)
5157 {
5158         struct workqueue_struct *wq = dev_to_wq(dev);
5159         struct workqueue_attrs *attrs;
5160         int v, ret = -ENOMEM;
5161 
5162         apply_wqattrs_lock();
5163 
5164         attrs = wq_sysfs_prep_attrs(wq);
5165         if (!attrs)
5166                 goto out_unlock;
5167 
5168         ret = -EINVAL;
5169         if (sscanf(buf, "%d", &v) == 1) {
5170                 attrs->no_numa = !v;
5171                 ret = apply_workqueue_attrs_locked(wq, attrs);
5172         }
5173 
5174 out_unlock:
5175         apply_wqattrs_unlock();
5176         free_workqueue_attrs(attrs);
5177         return ret ?: count;
5178 }
5179 
5180 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5181         __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5182         __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5183         __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5184         __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5185         __ATTR_NULL,
5186 };
5187 
5188 static struct bus_type wq_subsys = {
5189         .name                           = "workqueue",
5190         .dev_groups                     = wq_sysfs_groups,
5191 };
5192 
5193 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5194                 struct device_attribute *attr, char *buf)
5195 {
5196         int written;
5197 
5198         mutex_lock(&wq_pool_mutex);
5199         written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5200                             cpumask_pr_args(wq_unbound_cpumask));
5201         mutex_unlock(&wq_pool_mutex);
5202 
5203         return written;
5204 }
5205 
5206 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5207                 struct device_attribute *attr, const char *buf, size_t count)
5208 {
5209         cpumask_var_t cpumask;
5210         int ret;
5211 
5212         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5213                 return -ENOMEM;
5214 
5215         ret = cpumask_parse(buf, cpumask);
5216         if (!ret)
5217                 ret = workqueue_set_unbound_cpumask(cpumask);
5218 
5219         free_cpumask_var(cpumask);
5220         return ret ? ret : count;
5221 }
5222 
5223 static struct device_attribute wq_sysfs_cpumask_attr =
5224         __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5225                wq_unbound_cpumask_store);
5226 
5227 static int __init wq_sysfs_init(void)
5228 {
5229         int err;
5230 
5231         err = subsys_virtual_register(&wq_subsys, NULL);
5232         if (err)
5233                 return err;
5234 
5235         return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5236 }
5237 core_initcall(wq_sysfs_init);
5238 
5239 static void wq_device_release(struct device *dev)
5240 {
5241         struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5242 
5243         kfree(wq_dev);
5244 }
5245 
5246 /**
5247  * workqueue_sysfs_register - make a workqueue visible in sysfs
5248  * @wq: the workqueue to register
5249  *
5250  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5251  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5252  * which is the preferred method.
5253  *
5254  * Workqueue user should use this function directly iff it wants to apply
5255  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5256  * apply_workqueue_attrs() may race against userland updating the
5257  * attributes.
5258  *
5259  * Return: 0 on success, -errno on failure.
5260  */
5261 int workqueue_sysfs_register(struct workqueue_struct *wq)
5262 {
5263         struct wq_device *wq_dev;
5264         int ret;
5265 
5266         /*
5267          * Adjusting max_active or creating new pwqs by applying
5268          * attributes breaks ordering guarantee.  Disallow exposing ordered
5269          * workqueues.
5270          */
5271         if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5272                 return -EINVAL;
5273 
5274         wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5275         if (!wq_dev)
5276                 return -ENOMEM;
5277 
5278         wq_dev->wq = wq;
5279         wq_dev->dev.bus = &wq_subsys;
5280         wq_dev->dev.release = wq_device_release;
5281         dev_set_name(&wq_dev->dev, "%s", wq->name);
5282 
5283         /*
5284          * unbound_attrs are created separately.  Suppress uevent until
5285          * everything is ready.
5286          */
5287         dev_set_uevent_suppress(&wq_dev->dev, true);
5288 
5289         ret = device_register(&wq_dev->dev);
5290         if (ret) {
5291                 kfree(wq_dev);
5292                 wq->wq_dev = NULL;
5293                 return ret;
5294         }
5295 
5296         if (wq->flags & WQ_UNBOUND) {
5297                 struct device_attribute *attr;
5298 
5299                 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5300                         ret = device_create_file(&wq_dev->dev, attr);
5301                         if (ret) {
5302                                 device_unregister(&wq_dev->dev);
5303                                 wq->wq_dev = NULL;
5304                                 return ret;
5305                         }
5306                 }
5307         }
5308 
5309         dev_set_uevent_suppress(&wq_dev->dev, false);
5310         kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5311         return 0;
5312 }
5313 
5314 /**
5315  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5316  * @wq: the workqueue to unregister
5317  *
5318  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5319  */
5320 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5321 {
5322         struct wq_device *wq_dev = wq->wq_dev;
5323 
5324         if (!wq->wq_dev)
5325                 return;
5326 
5327         wq->wq_dev = NULL;
5328         device_unregister(&wq_dev->dev);
5329 }
5330 #else   /* CONFIG_SYSFS */
5331 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)     { }
5332 #endif  /* CONFIG_SYSFS */
5333 
5334 /*
5335  * Workqueue watchdog.
5336  *
5337  * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5338  * flush dependency, a concurrency managed work item which stays RUNNING
5339  * indefinitely.  Workqueue stalls can be very difficult to debug as the
5340  * usual warning mechanisms don't trigger and internal workqueue state is
5341  * largely opaque.
5342  *
5343  * Workqueue watchdog monitors all worker pools periodically and dumps
5344  * state if some pools failed to make forward progress for a while where
5345  * forward progress is defined as the first item on ->worklist changing.
5346  *
5347  * This mechanism is controlled through the kernel parameter
5348  * "workqueue.watchdog_thresh" which can be updated at runtime through the
5349  * corresponding sysfs parameter file.
5350  */
5351 #ifdef CONFIG_WQ_WATCHDOG
5352 
5353 static void wq_watchdog_timer_fn(unsigned long data);
5354 
5355 static unsigned long wq_watchdog_thresh = 30;
5356 static struct timer_list wq_watchdog_timer =
5357         TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5358 
5359 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5360 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5361 
5362 static void wq_watchdog_reset_touched(void)
5363 {
5364         int cpu;
5365 
5366         wq_watchdog_touched = jiffies;
5367         for_each_possible_cpu(cpu)
5368                 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5369 }
5370 
5371 static void wq_watchdog_timer_fn(unsigned long data)
5372 {
5373         unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5374         bool lockup_detected = false;
5375         struct worker_pool *pool;
5376         int pi;
5377 
5378         if (!thresh)
5379                 return;
5380 
5381         rcu_read_lock();
5382 
5383         for_each_pool(pool, pi) {
5384                 unsigned long pool_ts, touched, ts;
5385 
5386                 if (list_empty(&pool->worklist))
5387                         continue;
5388 
5389                 /* get the latest of pool and touched timestamps */
5390                 pool_ts = READ_ONCE(pool->watchdog_ts);
5391                 touched = READ_ONCE(wq_watchdog_touched);
5392 
5393                 if (time_after(pool_ts, touched))
5394                         ts = pool_ts;
5395                 else
5396                         ts = touched;
5397 
5398                 if (pool->cpu >= 0) {
5399                         unsigned long cpu_touched =
5400                                 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5401                                                   pool->cpu));
5402                         if (time_after(cpu_touched, ts))
5403                                 ts = cpu_touched;
5404                 }
5405 
5406                 /* did we stall? */
5407                 if (time_after(jiffies, ts + thresh)) {
5408                         lockup_detected = true;
5409                         pr_emerg("BUG: workqueue lockup - pool");
5410                         pr_cont_pool_info(pool);
5411                         pr_cont(" stuck for %us!\n",
5412                                 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5413                 }
5414         }
5415 
5416         rcu_read_unlock();
5417 
5418         if (lockup_detected)
5419                 show_workqueue_state();
5420 
5421         wq_watchdog_reset_touched();
5422         mod_timer(&wq_watchdog_timer, jiffies + thresh);
5423 }
5424 
5425 void wq_watchdog_touch(int cpu)
5426 {
5427         if (cpu >= 0)
5428                 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5429         else
5430                 wq_watchdog_touched = jiffies;
5431 }
5432 
5433 static void wq_watchdog_set_thresh(unsigned long thresh)
5434 {
5435         wq_watchdog_thresh = 0;
5436         del_timer_sync(&wq_watchdog_timer);
5437 
5438         if (thresh) {
5439                 wq_watchdog_thresh = thresh;
5440                 wq_watchdog_reset_touched();
5441                 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5442         }
5443 }
5444 
5445 static int wq_watchdog_param_set_thresh(const char *val,
5446                                         const struct kernel_param *kp)
5447 {
5448         unsigned long thresh;
5449         int ret;
5450 
5451         ret = kstrtoul(val, 0, &thresh);
5452         if (ret)
5453                 return ret;
5454 
5455         if (system_wq)
5456                 wq_watchdog_set_thresh(thresh);
5457         else
5458                 wq_watchdog_thresh = thresh;
5459 
5460         return 0;
5461 }
5462 
5463 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5464         .set    = wq_watchdog_param_set_thresh,
5465         .get    = param_get_ulong,
5466 };
5467 
5468 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5469                 0644);
5470 
5471 static void wq_watchdog_init(void)
5472 {
5473         wq_watchdog_set_thresh(wq_watchdog_thresh);
5474 }
5475 
5476 #else   /* CONFIG_WQ_WATCHDOG */
5477 
5478 static inline void wq_watchdog_init(void) { }
5479 
5480 #endif  /* CONFIG_WQ_WATCHDOG */
5481 
5482 static void __init wq_numa_init(void)
5483 {
5484         cpumask_var_t *tbl;
5485         int node, cpu;
5486 
5487         if (num_possible_nodes() <= 1)
5488                 return;
5489 
5490         if (wq_disable_numa) {
5491                 pr_info("workqueue: NUMA affinity support disabled\n");
5492                 return;
5493         }
5494 
5495         wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5496         BUG_ON(!wq_update_unbound_numa_attrs_buf);
5497 
5498         /*
5499          * We want masks of possible CPUs of each node which isn't readily
5500          * available.  Build one from cpu_to_node() which should have been
5501          * fully initialized by now.
5502          */
5503         tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5504         BUG_ON(!tbl);
5505 
5506         for_each_node(node)
5507                 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5508                                 node_online(node) ? node : NUMA_NO_NODE));
5509 
5510         for_each_possible_cpu(cpu) {
5511                 node = cpu_to_node(cpu);
5512                 if (WARN_ON(node == NUMA_NO_NODE)) {
5513                         pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5514                         /* happens iff arch is bonkers, let's just proceed */
5515                         return;
5516                 }
5517                 cpumask_set_cpu(cpu, tbl[node]);
5518         }
5519 
5520         wq_numa_possible_cpumask = tbl;
5521         wq_numa_enabled = true;
5522 }
5523 
5524 /**
5525  * workqueue_init_early - early init for workqueue subsystem
5526  *
5527  * This is the first half of two-staged workqueue subsystem initialization
5528  * and invoked as soon as the bare basics - memory allocation, cpumasks and
5529  * idr are up.  It sets up all the data structures and system workqueues
5530  * and allows early boot code to create workqueues and queue/cancel work
5531  * items.  Actual work item execution starts only after kthreads can be
5532  * created and scheduled right before early initcalls.
5533  */
5534 int __init workqueue_init_early(void)
5535 {
5536         int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5537         int i, cpu;
5538 
5539         WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5540 
5541         BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5542         cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5543 
5544         pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5545 
5546         /* initialize CPU pools */
5547         for_each_possible_cpu(cpu) {
5548                 struct worker_pool *pool;
5549 
5550                 i = 0;
5551                 for_each_cpu_worker_pool(pool, cpu) {
5552                         BUG_ON(init_worker_pool(pool));
5553                         pool->cpu = cpu;
5554                         cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5555                         pool->attrs->nice = std_nice[i++];
5556                         pool->node = cpu_to_node(cpu);
5557 
5558                         /* alloc pool ID */
5559                         mutex_lock(&wq_pool_mutex);
5560                         BUG_ON(worker_pool_assign_id(pool));
5561                         mutex_unlock(&wq_pool_mutex);
5562                 }
5563         }
5564 
5565         /* create default unbound and ordered wq attrs */
5566         for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5567                 struct workqueue_attrs *attrs;
5568 
5569                 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5570                 attrs->nice = std_nice[i];
5571                 unbound_std_wq_attrs[i] = attrs;
5572 
5573                 /*
5574                  * An ordered wq should have only one pwq as ordering is
5575                  * guaranteed by max_active which is enforced by pwqs.
5576                  * Turn off NUMA so that dfl_pwq is used for all nodes.
5577                  */
5578                 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5579                 attrs->nice = std_nice[i];
5580                 attrs->no_numa = true;
5581                 ordered_wq_attrs[i] = attrs;
5582         }
5583 
5584         system_wq = alloc_workqueue("events", 0, 0);
5585         system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5586         system_long_wq = alloc_workqueue("events_long", 0, 0);
5587         system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5588                                             WQ_UNBOUND_MAX_ACTIVE);
5589         system_freezable_wq = alloc_workqueue("events_freezable",
5590                                               WQ_FREEZABLE, 0);
5591         system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5592                                               WQ_POWER_EFFICIENT, 0);
5593         system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5594                                               WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5595                                               0);
5596         BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5597                !system_unbound_wq || !system_freezable_wq ||
5598                !system_power_efficient_wq ||
5599                !system_freezable_power_efficient_wq);
5600 
5601         return 0;
5602 }
5603 
5604 /**
5605  * workqueue_init - bring workqueue subsystem fully online
5606  *
5607  * This is the latter half of two-staged workqueue subsystem initialization
5608  * and invoked as soon as kthreads can be created and scheduled.
5609  * Workqueues have been created and work items queued on them, but there
5610  * are no kworkers executing the work items yet.  Populate the worker pools
5611  * with the initial workers and enable future kworker creations.
5612  */
5613 int __init workqueue_init(void)
5614 {
5615         struct workqueue_struct *wq;
5616         struct worker_pool *pool;
5617         int cpu, bkt;
5618 
5619         /*
5620          * It'd be simpler to initialize NUMA in workqueue_init_early() but
5621          * CPU to node mapping may not be available that early on some
5622          * archs such as power and arm64.  As per-cpu pools created
5623          * previously could be missing node hint and unbound pools NUMA
5624          * affinity, fix them up.
5625          */
5626         wq_numa_init();
5627 
5628         mutex_lock(&wq_pool_mutex);
5629 
5630         for_each_possible_cpu(cpu) {
5631                 for_each_cpu_worker_pool(pool, cpu) {
5632                         pool->node = cpu_to_node(cpu);
5633                 }
5634         }
5635 
5636         list_for_each_entry(wq, &workqueues, list)
5637                 wq_update_unbound_numa(wq, smp_processor_id(), true);
5638 
5639         mutex_unlock(&wq_pool_mutex);
5640 
5641         /* create the initial workers */
5642         for_each_online_cpu(cpu) {
5643                 for_each_cpu_worker_pool(pool, cpu) {
5644                         pool->flags &= ~POOL_DISASSOCIATED;
5645                         BUG_ON(!create_worker(pool));
5646                 }
5647         }
5648 
5649         hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5650                 BUG_ON(!create_worker(pool));
5651 
5652         wq_online = true;
5653         wq_watchdog_init();
5654 
5655         return 0;
5656 }
5657 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp