kernel/locking/rwbase_rt.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 /*
   4  * RT-specific reader/writer semaphores and reader/writer locks
   5  *
   6  * down_write/write_lock()
   7  *  1) Lock rtmutex
   8  *  2) Remove the reader BIAS to force readers into the slow path
   9  *  3) Wait until all readers have left the critical section
  10  *  4) Mark it write locked
  11  *
  12  * up_write/write_unlock()
  13  *  1) Remove the write locked marker
  14  *  2) Set the reader BIAS, so readers can use the fast path again
  15  *  3) Unlock rtmutex, to release blocked readers
  16  *
  17  * down_read/read_lock()
  18  *  1) Try fast path acquisition (reader BIAS is set)
  19  *  2) Take tmutex::wait_lock, which protects the writelocked flag
  20  *  3) If !writelocked, acquire it for read
  21  *  4) If writelocked, block on tmutex
  22  *  5) unlock rtmutex, goto 1)
  23  *
  24  * up_read/read_unlock()
  25  *  1) Try fast path release (reader count != 1)
  26  *  2) Wake the writer waiting in down_write()/write_lock() #3
  27  *
  28  * down_read/read_lock()#3 has the consequence, that rw semaphores and rw
  29  * locks on RT are not writer fair, but writers, which should be avoided in
  30  * RT tasks (think mmap_sem), are subject to the rtmutex priority/DL
  31  * inheritance mechanism.
  32  *
  33  * It's possible to make the rw primitives writer fair by keeping a list of
  34  * active readers. A blocked writer would force all newly incoming readers
  35  * to block on the rtmutex, but the rtmutex would have to be proxy locked
  36  * for one reader after the other. We can't use multi-reader inheritance
  37  * because there is no way to support that with SCHED_DEADLINE.
  38  * Implementing the one by one reader boosting/handover mechanism is a
  39  * major surgery for a very dubious value.
  40  *
  41  * The risk of writer starvation is there, but the pathological use cases
  42  * which trigger it are not necessarily the typical RT workloads.
  43  *
  44  * Fast-path orderings:
  45  * The lock/unlock of readers can run in fast paths: lock and unlock are only
  46  * atomic ops, and there is no inner lock to provide ACQUIRE and RELEASE
  47  * semantics of rwbase_rt. Atomic ops should thus provide _acquire()
  48  * and _release() (or stronger).
  49  *
  50  * Common code shared between RT rw_semaphore and rwlock
  51  */
  52
  53 static __always_inline int rwbase_read_trylock(struct rwbase_rt *rwb)
  54 {
  55         int r;
  56
  57         /*
  58          * Increment reader count, if sem->readers < 0, i.e. READER_BIAS is
  59          * set.
  60          */
  61         for (r = atomic_read(&rwb->readers); r < 0;) {
  62                 if (likely(atomic_try_cmpxchg_acquire(&rwb->readers, &r, r + 1)))
  63                         return 1;
  64         }
  65         return 0;
  66 }
  67
  68 static int __sched __rwbase_read_lock(struct rwbase_rt *rwb,
  69                                       unsigned int state)
  70 {
  71         struct rt_mutex_base *rtm = &rwb->rtmutex;
  72         int ret;
  73
  74         raw_spin_lock_irq(&rtm->wait_lock);
  75
  76         /*
  77          * Call into the slow lock path with the rtmutex->wait_lock
  78          * held, so this can't result in the following race:
  79          *
  80          * Reader1              Reader2         Writer
  81          *                      down_read()
  82          *                                      down_write()
  83          *                                      rtmutex_lock(m)
  84          *                                      wait()
  85          * down_read()
  86          * unlock(m->wait_lock)
  87          *                      up_read()
  88          *                      wake(Writer)
  89          *                                      lock(m->wait_lock)
  90          *                                      sem->writelocked=true
  91          *                                      unlock(m->wait_lock)
  92          *
  93          *                                      up_write()
  94          *                                      sem->writelocked=false
  95          *                                      rtmutex_unlock(m)
  96          *                      down_read()
  97          *                                      down_write()
  98          *                                      rtmutex_lock(m)
  99          *                                      wait()
 100          * rtmutex_lock(m)
 101          *
 102          * That would put Reader1 behind the writer waiting on
 103          * Reader2 to call up_read(), which might be unbound.
 104          */
 105
 106         trace_contention_begin(rwb, LCB_F_RT | LCB_F_READ);
 107
 108         /*
 109          * For rwlocks this returns 0 unconditionally, so the below
 110          * !ret conditionals are optimized out.
 111          */
 112         ret = rwbase_rtmutex_slowlock_locked(rtm, state);
 113
 114         /*
 115          * On success the rtmutex is held, so there can't be a writer
 116          * active. Increment the reader count and immediately drop the
 117          * rtmutex again.
 118          *
 119          * rtmutex->wait_lock has to be unlocked in any case of course.
 120          */
 121         if (!ret)
 122                 atomic_inc(&rwb->readers);
 123         raw_spin_unlock_irq(&rtm->wait_lock);
 124         if (!ret)
 125                 rwbase_rtmutex_unlock(rtm);
 126
 127         trace_contention_end(rwb, ret);
 128         return ret;
 129 }
 130
 131 static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb,
 132                                             unsigned int state)
 133 {
 134         if (rwbase_read_trylock(rwb))
 135                 return 0;
 136
 137         return __rwbase_read_lock(rwb, state);
 138 }
 139
 140 static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
 141                                          unsigned int state)
 142 {
 143         struct rt_mutex_base *rtm = &rwb->rtmutex;
 144         struct task_struct *owner;
 145         DEFINE_RT_WAKE_Q(wqh);
 146
 147         raw_spin_lock_irq(&rtm->wait_lock);
 148         /*
 149          * Wake the writer, i.e. the rtmutex owner. It might release the
 150          * rtmutex concurrently in the fast path (due to a signal), but to
 151          * clean up rwb->readers it needs to acquire rtm->wait_lock. The
 152          * worst case which can happen is a spurious wakeup.
 153          */
 154         owner = rt_mutex_owner(rtm);
 155         if (owner)
 156                 rt_mutex_wake_q_add_task(&wqh, owner, state);
 157
 158         /* Pairs with the preempt_enable in rt_mutex_wake_up_q() */
 159         preempt_disable();
 160         raw_spin_unlock_irq(&rtm->wait_lock);
 161         rt_mutex_wake_up_q(&wqh);
 162 }
 163
 164 static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
 165                                                unsigned int state)
 166 {
 167         /*
 168          * rwb->readers can only hit 0 when a writer is waiting for the
 169          * active readers to leave the critical section.
 170          *
 171          * dec_and_test() is fully ordered, provides RELEASE.
 172          */
 173         if (unlikely(atomic_dec_and_test(&rwb->readers)))
 174                 __rwbase_read_unlock(rwb, state);
 175 }
 176
 177 static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
 178                                          unsigned long flags)
 179 {
 180         struct rt_mutex_base *rtm = &rwb->rtmutex;
 181
 182         /*
 183          * _release() is needed in case that reader is in fast path, pairing
 184          * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
 185          */
 186         (void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
 187         raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 188         rwbase_rtmutex_unlock(rtm);
 189 }
 190
 191 static inline void rwbase_write_unlock(struct rwbase_rt *rwb)
 192 {
 193         struct rt_mutex_base *rtm = &rwb->rtmutex;
 194         unsigned long flags;
 195
 196         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 197         __rwbase_write_unlock(rwb, WRITER_BIAS, flags);
 198 }
 199
 200 static inline void rwbase_write_downgrade(struct rwbase_rt *rwb)
 201 {
 202         struct rt_mutex_base *rtm = &rwb->rtmutex;
 203         unsigned long flags;
 204
 205         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 206         /* Release it and account current as reader */
 207         __rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags);
 208 }
 209
 210 static inline bool __rwbase_write_trylock(struct rwbase_rt *rwb)
 211 {
 212         /* Can do without CAS because we're serialized by wait_lock. */
 213         lockdep_assert_held(&rwb->rtmutex.wait_lock);
 214
 215         /*
 216          * _acquire is needed in case the reader is in the fast path, pairing
 217          * with rwbase_read_unlock(), provides ACQUIRE.
 218          */
 219         if (!atomic_read_acquire(&rwb->readers)) {
 220                 atomic_set(&rwb->readers, WRITER_BIAS);
 221                 return 1;
 222         }
 223
 224         return 0;
 225 }
 226
 227 static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
 228                                      unsigned int state)
 229 {
 230         struct rt_mutex_base *rtm = &rwb->rtmutex;
 231         unsigned long flags;
 232
 233         /* Take the rtmutex as a first step */
 234         if (rwbase_rtmutex_lock_state(rtm, state))
 235                 return -EINTR;
 236
 237         /* Force readers into slow path */
 238         atomic_sub(READER_BIAS, &rwb->readers);
 239
 240         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 241         if (__rwbase_write_trylock(rwb))
 242                 goto out_unlock;
 243
 244         rwbase_set_and_save_current_state(state);
 245         trace_contention_begin(rwb, LCB_F_RT | LCB_F_WRITE);
 246         for (;;) {
 247                 /* Optimized out for rwlocks */
 248                 if (rwbase_signal_pending_state(state, current)) {
 249                         rwbase_restore_current_state();
 250                         __rwbase_write_unlock(rwb, 0, flags);
 251                         trace_contention_end(rwb, -EINTR);
 252                         return -EINTR;
 253                 }
 254
 255                 if (__rwbase_write_trylock(rwb))
 256                         break;
 257
 258                 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 259                 rwbase_schedule();
 260                 raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 261
 262                 set_current_state(state);
 263         }
 264         rwbase_restore_current_state();
 265         trace_contention_end(rwb, 0);
 266
 267 out_unlock:
 268         raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 269         return 0;
 270 }
 271
 272 static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
 273 {
 274         struct rt_mutex_base *rtm = &rwb->rtmutex;
 275         unsigned long flags;
 276
 277         if (!rwbase_rtmutex_trylock(rtm))
 278                 return 0;
 279
 280         atomic_sub(READER_BIAS, &rwb->readers);
 281
 282         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 283         if (__rwbase_write_trylock(rwb)) {
 284                 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 285                 return 1;
 286         }
 287         __rwbase_write_unlock(rwb, 0, flags);
 288         return 0;
 289 }