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          * Allow readers, as long as the writer has not completely
  77          * acquired the semaphore for write.
  78          */
  79         if (atomic_read(&rwb->readers) != WRITER_BIAS) {
  80                 atomic_inc(&rwb->readers);
  81                 raw_spin_unlock_irq(&rtm->wait_lock);
  82                 return 0;
  83         }
  84
  85         /*
  86          * Call into the slow lock path with the rtmutex->wait_lock
  87          * held, so this can't result in the following race:
  88          *
  89          * Reader1              Reader2         Writer
  90          *                      down_read()
  91          *                                      down_write()
  92          *                                      rtmutex_lock(m)
  93          *                                      wait()
  94          * down_read()
  95          * unlock(m->wait_lock)
  96          *                      up_read()
  97          *                      wake(Writer)
  98          *                                      lock(m->wait_lock)
  99          *                                      sem->writelocked=true
 100          *                                      unlock(m->wait_lock)
 101          *
 102          *                                      up_write()
 103          *                                      sem->writelocked=false
 104          *                                      rtmutex_unlock(m)
 105          *                      down_read()
 106          *                                      down_write()
 107          *                                      rtmutex_lock(m)
 108          *                                      wait()
 109          * rtmutex_lock(m)
 110          *
 111          * That would put Reader1 behind the writer waiting on
 112          * Reader2 to call up_read(), which might be unbound.
 113          */
 114
 115         /*
 116          * For rwlocks this returns 0 unconditionally, so the below
 117          * !ret conditionals are optimized out.
 118          */
 119         ret = rwbase_rtmutex_slowlock_locked(rtm, state);
 120
 121         /*
 122          * On success the rtmutex is held, so there can't be a writer
 123          * active. Increment the reader count and immediately drop the
 124          * rtmutex again.
 125          *
 126          * rtmutex->wait_lock has to be unlocked in any case of course.
 127          */
 128         if (!ret)
 129                 atomic_inc(&rwb->readers);
 130         raw_spin_unlock_irq(&rtm->wait_lock);
 131         if (!ret)
 132                 rwbase_rtmutex_unlock(rtm);
 133         return ret;
 134 }
 135
 136 static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb,
 137                                             unsigned int state)
 138 {
 139         if (rwbase_read_trylock(rwb))
 140                 return 0;
 141
 142         return __rwbase_read_lock(rwb, state);
 143 }
 144
 145 static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
 146                                          unsigned int state)
 147 {
 148         struct rt_mutex_base *rtm = &rwb->rtmutex;
 149         struct task_struct *owner;
 150         DEFINE_RT_WAKE_Q(wqh);
 151
 152         raw_spin_lock_irq(&rtm->wait_lock);
 153         /*
 154          * Wake the writer, i.e. the rtmutex owner. It might release the
 155          * rtmutex concurrently in the fast path (due to a signal), but to
 156          * clean up rwb->readers it needs to acquire rtm->wait_lock. The
 157          * worst case which can happen is a spurious wakeup.
 158          */
 159         owner = rt_mutex_owner(rtm);
 160         if (owner)
 161                 rt_mutex_wake_q_add_task(&wqh, owner, state);
 162
 163         /* Pairs with the preempt_enable in rt_mutex_wake_up_q() */
 164         preempt_disable();
 165         raw_spin_unlock_irq(&rtm->wait_lock);
 166         rt_mutex_wake_up_q(&wqh);
 167 }
 168
 169 static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
 170                                                unsigned int state)
 171 {
 172         /*
 173          * rwb->readers can only hit 0 when a writer is waiting for the
 174          * active readers to leave the critical section.
 175          *
 176          * dec_and_test() is fully ordered, provides RELEASE.
 177          */
 178         if (unlikely(atomic_dec_and_test(&rwb->readers)))
 179                 __rwbase_read_unlock(rwb, state);
 180 }
 181
 182 static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
 183                                          unsigned long flags)
 184 {
 185         struct rt_mutex_base *rtm = &rwb->rtmutex;
 186
 187         /*
 188          * _release() is needed in case that reader is in fast path, pairing
 189          * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
 190          */
 191         (void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
 192         raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 193         rwbase_rtmutex_unlock(rtm);
 194 }
 195
 196 static inline void rwbase_write_unlock(struct rwbase_rt *rwb)
 197 {
 198         struct rt_mutex_base *rtm = &rwb->rtmutex;
 199         unsigned long flags;
 200
 201         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 202         __rwbase_write_unlock(rwb, WRITER_BIAS, flags);
 203 }
 204
 205 static inline void rwbase_write_downgrade(struct rwbase_rt *rwb)
 206 {
 207         struct rt_mutex_base *rtm = &rwb->rtmutex;
 208         unsigned long flags;
 209
 210         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 211         /* Release it and account current as reader */
 212         __rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags);
 213 }
 214
 215 static inline bool __rwbase_write_trylock(struct rwbase_rt *rwb)
 216 {
 217         /* Can do without CAS because we're serialized by wait_lock. */
 218         lockdep_assert_held(&rwb->rtmutex.wait_lock);
 219
 220         /*
 221          * _acquire is needed in case the reader is in the fast path, pairing
 222          * with rwbase_read_unlock(), provides ACQUIRE.
 223          */
 224         if (!atomic_read_acquire(&rwb->readers)) {
 225                 atomic_set(&rwb->readers, WRITER_BIAS);
 226                 return 1;
 227         }
 228
 229         return 0;
 230 }
 231
 232 static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
 233                                      unsigned int state)
 234 {
 235         struct rt_mutex_base *rtm = &rwb->rtmutex;
 236         unsigned long flags;
 237
 238         /* Take the rtmutex as a first step */
 239         if (rwbase_rtmutex_lock_state(rtm, state))
 240                 return -EINTR;
 241
 242         /* Force readers into slow path */
 243         atomic_sub(READER_BIAS, &rwb->readers);
 244
 245         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 246         if (__rwbase_write_trylock(rwb))
 247                 goto out_unlock;
 248
 249         rwbase_set_and_save_current_state(state);
 250         for (;;) {
 251                 /* Optimized out for rwlocks */
 252                 if (rwbase_signal_pending_state(state, current)) {
 253                         rwbase_restore_current_state();
 254                         __rwbase_write_unlock(rwb, 0, flags);
 255                         return -EINTR;
 256                 }
 257
 258                 if (__rwbase_write_trylock(rwb))
 259                         break;
 260
 261                 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 262                 rwbase_schedule();
 263                 raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 264
 265                 set_current_state(state);
 266         }
 267         rwbase_restore_current_state();
 268
 269 out_unlock:
 270         raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 271         return 0;
 272 }
 273
 274 static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
 275 {
 276         struct rt_mutex_base *rtm = &rwb->rtmutex;
 277         unsigned long flags;
 278
 279         if (!rwbase_rtmutex_trylock(rtm))
 280                 return 0;
 281
 282         atomic_sub(READER_BIAS, &rwb->readers);
 283
 284         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
 285         if (__rwbase_write_trylock(rwb)) {
 286                 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
 287                 return 1;
 288         }
 289         __rwbase_write_unlock(rwb, 0, flags);
 290         return 0;
 291 }