| 1 |
// SPDX-License-Identifier: GPL-2.0-only |
| 2 |
#include <linux/atomic.h> |
| 3 |
#include <linux/percpu.h> |
| 4 |
#include <linux/wait.h> |
| 5 |
#include <linux/lockdep.h> |
| 6 |
#include <linux/percpu-rwsem.h> |
| 7 |
#include <linux/rcupdate.h> |
| 8 |
#include <linux/sched.h> |
| 9 |
#include <linux/sched/task.h> |
| 10 |
#include <linux/sched/debug.h> |
| 11 |
#include <linux/errno.h> |
| 12 |
#include <trace/events/lock.h> |
| 13 |
|
| 14 |
int __percpu_init_rwsem(struct percpu_rw_semaphore *sem, |
| 15 |
const char *name, struct lock_class_key *key) |
| 16 |
{ |
| 17 |
sem->read_count = alloc_percpu(int); |
| 18 |
if (unlikely(!sem->read_count)) |
| 19 |
return -ENOMEM; |
| 20 |
|
| 21 |
rcu_sync_init(&sem->rss); |
| 22 |
rcuwait_init(&sem->writer); |
| 23 |
init_waitqueue_head(&sem->waiters); |
| 24 |
atomic_set(&sem->block, 0); |
| 25 |
#ifdef CONFIG_DEBUG_LOCK_ALLOC |
| 26 |
debug_check_no_locks_freed((void *)sem, sizeof(*sem)); |
| 27 |
lockdep_init_map(&sem->dep_map, name, key, 0); |
| 28 |
#endif |
| 29 |
return 0; |
| 30 |
} |
| 31 |
EXPORT_SYMBOL_GPL(__percpu_init_rwsem); |
| 32 |
|
| 33 |
void percpu_free_rwsem(struct percpu_rw_semaphore *sem) |
| 34 |
{ |
| 35 |
/* |
| 36 |
* XXX: temporary kludge. The error path in alloc_super() |
| 37 |
* assumes that percpu_free_rwsem() is safe after kzalloc(). |
| 38 |
*/ |
| 39 |
if (!sem->read_count) |
| 40 |
return; |
| 41 |
|
| 42 |
rcu_sync_dtor(&sem->rss); |
| 43 |
free_percpu(sem->read_count); |
| 44 |
sem->read_count = NULL; /* catch use after free bugs */ |
| 45 |
} |
| 46 |
EXPORT_SYMBOL_GPL(percpu_free_rwsem); |
| 47 |
|
| 48 |
static bool __percpu_down_read_trylock(struct percpu_rw_semaphore *sem) |
| 49 |
{ |
| 50 |
this_cpu_inc(*sem->read_count); |
| 51 |
|
| 52 |
/* |
| 53 |
* Due to having preemption disabled the decrement happens on |
| 54 |
* the same CPU as the increment, avoiding the |
| 55 |
* increment-on-one-CPU-and-decrement-on-another problem. |
| 56 |
* |
| 57 |
* If the reader misses the writer's assignment of sem->block, then the |
| 58 |
* writer is guaranteed to see the reader's increment. |
| 59 |
* |
| 60 |
* Conversely, any readers that increment their sem->read_count after |
| 61 |
* the writer looks are guaranteed to see the sem->block value, which |
| 62 |
* in turn means that they are guaranteed to immediately decrement |
| 63 |
* their sem->read_count, so that it doesn't matter that the writer |
| 64 |
* missed them. |
| 65 |
*/ |
| 66 |
|
| 67 |
smp_mb(); /* A matches D */ |
| 68 |
|
| 69 |
/* |
| 70 |
* If !sem->block the critical section starts here, matched by the |
| 71 |
* release in percpu_up_write(). |
| 72 |
*/ |
| 73 |
if (likely(!atomic_read_acquire(&sem->block))) |
| 74 |
return true; |
| 75 |
|
| 76 |
this_cpu_dec(*sem->read_count); |
| 77 |
|
| 78 |
/* Prod writer to re-evaluate readers_active_check() */ |
| 79 |
rcuwait_wake_up(&sem->writer); |
| 80 |
|
| 81 |
return false; |
| 82 |
} |
| 83 |
|
| 84 |
static inline bool __percpu_down_write_trylock(struct percpu_rw_semaphore *sem) |
| 85 |
{ |
| 86 |
if (atomic_read(&sem->block)) |
| 87 |
return false; |
| 88 |
|
| 89 |
return atomic_xchg(&sem->block, 1) == 0; |
| 90 |
} |
| 91 |
|
| 92 |
static bool __percpu_rwsem_trylock(struct percpu_rw_semaphore *sem, bool reader) |
| 93 |
{ |
| 94 |
if (reader) { |
| 95 |
bool ret; |
| 96 |
|
| 97 |
preempt_disable(); |
| 98 |
ret = __percpu_down_read_trylock(sem); |
| 99 |
preempt_enable(); |
| 100 |
|
| 101 |
return ret; |
| 102 |
} |
| 103 |
return __percpu_down_write_trylock(sem); |
| 104 |
} |
| 105 |
|
| 106 |
/* |
| 107 |
* The return value of wait_queue_entry::func means: |
| 108 |
* |
| 109 |
* <0 - error, wakeup is terminated and the error is returned |
| 110 |
* 0 - no wakeup, a next waiter is tried |
| 111 |
* >0 - woken, if EXCLUSIVE, counted towards @nr_exclusive. |
| 112 |
* |
| 113 |
* We use EXCLUSIVE for both readers and writers to preserve FIFO order, |
| 114 |
* and play games with the return value to allow waking multiple readers. |
| 115 |
* |
| 116 |
* Specifically, we wake readers until we've woken a single writer, or until a |
| 117 |
* trylock fails. |
| 118 |
*/ |
| 119 |
static int percpu_rwsem_wake_function(struct wait_queue_entry *wq_entry, |
| 120 |
unsigned int mode, int wake_flags, |
| 121 |
void *key) |
| 122 |
{ |
| 123 |
bool reader = wq_entry->flags & WQ_FLAG_CUSTOM; |
| 124 |
struct percpu_rw_semaphore *sem = key; |
| 125 |
struct task_struct *p; |
| 126 |
|
| 127 |
/* concurrent against percpu_down_write(), can get stolen */ |
| 128 |
if (!__percpu_rwsem_trylock(sem, reader)) |
| 129 |
return 1; |
| 130 |
|
| 131 |
p = get_task_struct(wq_entry->private); |
| 132 |
list_del_init(&wq_entry->entry); |
| 133 |
smp_store_release(&wq_entry->private, NULL); |
| 134 |
|
| 135 |
wake_up_process(p); |
| 136 |
put_task_struct(p); |
| 137 |
|
| 138 |
return !reader; /* wake (readers until) 1 writer */ |
| 139 |
} |
| 140 |
|
| 141 |
static void percpu_rwsem_wait(struct percpu_rw_semaphore *sem, bool reader, |
| 142 |
bool freeze) |
| 143 |
{ |
| 144 |
DEFINE_WAIT_FUNC(wq_entry, percpu_rwsem_wake_function); |
| 145 |
bool wait; |
| 146 |
|
| 147 |
spin_lock_irq(&sem->waiters.lock); |
| 148 |
/* |
| 149 |
* Serialize against the wakeup in percpu_up_write(), if we fail |
| 150 |
* the trylock, the wakeup must see us on the list. |
| 151 |
*/ |
| 152 |
wait = !__percpu_rwsem_trylock(sem, reader); |
| 153 |
if (wait) { |
| 154 |
wq_entry.flags |= WQ_FLAG_EXCLUSIVE | reader * WQ_FLAG_CUSTOM; |
| 155 |
__add_wait_queue_entry_tail(&sem->waiters, &wq_entry); |
| 156 |
} |
| 157 |
spin_unlock_irq(&sem->waiters.lock); |
| 158 |
|
| 159 |
while (wait) { |
| 160 |
set_current_state(TASK_UNINTERRUPTIBLE | |
| 161 |
(freeze ? TASK_FREEZABLE : 0)); |
| 162 |
if (!smp_load_acquire(&wq_entry.private)) |
| 163 |
break; |
| 164 |
schedule(); |
| 165 |
} |
| 166 |
__set_current_state(TASK_RUNNING); |
| 167 |
} |
| 168 |
|
| 169 |
bool __sched __percpu_down_read(struct percpu_rw_semaphore *sem, bool try, |
| 170 |
bool freeze) |
| 171 |
{ |
| 172 |
if (__percpu_down_read_trylock(sem)) |
| 173 |
return true; |
| 174 |
|
| 175 |
if (try) |
| 176 |
return false; |
| 177 |
|
| 178 |
trace_contention_begin(sem, LCB_F_PERCPU | LCB_F_READ); |
| 179 |
preempt_enable(); |
| 180 |
percpu_rwsem_wait(sem, /* .reader = */ true, freeze); |
| 181 |
preempt_disable(); |
| 182 |
trace_contention_end(sem, 0); |
| 183 |
|
| 184 |
return true; |
| 185 |
} |
| 186 |
EXPORT_SYMBOL_GPL(__percpu_down_read); |
| 187 |
|
| 188 |
#define per_cpu_sum(var) \ |
| 189 |
({ \ |
| 190 |
TYPEOF_UNQUAL(var) __sum = 0; \ |
| 191 |
int cpu; \ |
| 192 |
compiletime_assert_atomic_type(__sum); \ |
| 193 |
for_each_possible_cpu(cpu) \ |
| 194 |
__sum += per_cpu(var, cpu); \ |
| 195 |
__sum; \ |
| 196 |
}) |
| 197 |
|
| 198 |
bool percpu_is_read_locked(struct percpu_rw_semaphore *sem) |
| 199 |
{ |
| 200 |
return per_cpu_sum(*sem->read_count) != 0 && !atomic_read(&sem->block); |
| 201 |
} |
| 202 |
EXPORT_SYMBOL_GPL(percpu_is_read_locked); |
| 203 |
|
| 204 |
/* |
| 205 |
* Return true if the modular sum of the sem->read_count per-CPU variable is |
| 206 |
* zero. If this sum is zero, then it is stable due to the fact that if any |
| 207 |
* newly arriving readers increment a given counter, they will immediately |
| 208 |
* decrement that same counter. |
| 209 |
* |
| 210 |
* Assumes sem->block is set. |
| 211 |
*/ |
| 212 |
static bool readers_active_check(struct percpu_rw_semaphore *sem) |
| 213 |
{ |
| 214 |
if (per_cpu_sum(*sem->read_count) != 0) |
| 215 |
return false; |
| 216 |
|
| 217 |
/* |
| 218 |
* If we observed the decrement; ensure we see the entire critical |
| 219 |
* section. |
| 220 |
*/ |
| 221 |
|
| 222 |
smp_mb(); /* C matches B */ |
| 223 |
|
| 224 |
return true; |
| 225 |
} |
| 226 |
|
| 227 |
void __sched percpu_down_write(struct percpu_rw_semaphore *sem) |
| 228 |
{ |
| 229 |
bool contended = false; |
| 230 |
|
| 231 |
might_sleep(); |
| 232 |
rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_); |
| 233 |
|
| 234 |
/* Notify readers to take the slow path. */ |
| 235 |
rcu_sync_enter(&sem->rss); |
| 236 |
|
| 237 |
/* |
| 238 |
* Try set sem->block; this provides writer-writer exclusion. |
| 239 |
* Having sem->block set makes new readers block. |
| 240 |
*/ |
| 241 |
if (!__percpu_down_write_trylock(sem)) { |
| 242 |
trace_contention_begin(sem, LCB_F_PERCPU | LCB_F_WRITE); |
| 243 |
percpu_rwsem_wait(sem, /* .reader = */ false, false); |
| 244 |
contended = true; |
| 245 |
} |
| 246 |
|
| 247 |
/* smp_mb() implied by __percpu_down_write_trylock() on success -- D matches A */ |
| 248 |
|
| 249 |
/* |
| 250 |
* If they don't see our store of sem->block, then we are guaranteed to |
| 251 |
* see their sem->read_count increment, and therefore will wait for |
| 252 |
* them. |
| 253 |
*/ |
| 254 |
|
| 255 |
/* Wait for all active readers to complete. */ |
| 256 |
rcuwait_wait_event(&sem->writer, readers_active_check(sem), TASK_UNINTERRUPTIBLE); |
| 257 |
if (contended) |
| 258 |
trace_contention_end(sem, 0); |
| 259 |
} |
| 260 |
EXPORT_SYMBOL_GPL(percpu_down_write); |
| 261 |
|
| 262 |
void percpu_up_write(struct percpu_rw_semaphore *sem) |
| 263 |
{ |
| 264 |
rwsem_release(&sem->dep_map, _RET_IP_); |
| 265 |
|
| 266 |
/* |
| 267 |
* Signal the writer is done, no fast path yet. |
| 268 |
* |
| 269 |
* One reason that we cannot just immediately flip to readers_fast is |
| 270 |
* that new readers might fail to see the results of this writer's |
| 271 |
* critical section. |
| 272 |
* |
| 273 |
* Therefore we force it through the slow path which guarantees an |
| 274 |
* acquire and thereby guarantees the critical section's consistency. |
| 275 |
*/ |
| 276 |
atomic_set_release(&sem->block, 0); |
| 277 |
|
| 278 |
/* |
| 279 |
* Prod any pending reader/writer to make progress. |
| 280 |
*/ |
| 281 |
__wake_up(&sem->waiters, TASK_NORMAL, 1, sem); |
| 282 |
|
| 283 |
/* |
| 284 |
* Once this completes (at least one RCU-sched grace period hence) the |
| 285 |
* reader fast path will be available again. Safe to use outside the |
| 286 |
* exclusive write lock because its counting. |
| 287 |
*/ |
| 288 |
rcu_sync_exit(&sem->rss); |
| 289 |
} |
| 290 |
EXPORT_SYMBOL_GPL(percpu_up_write); |