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 rwbase_pre_schedule(); 75 raw_spin_lock_irq(&rtm->wait_lock); 76 77 /* 78 * Call into the slow lock path with the rtmutex->wait_lock 79 * held, so this can't result in the following race: 80 * 81 * Reader1 Reader2 Writer 82 * down_read() 83 * down_write() 84 * rtmutex_lock(m) 85 * wait() 86 * down_read() 87 * unlock(m->wait_lock) 88 * up_read() 89 * wake(Writer) 90 * lock(m->wait_lock) 91 * sem->writelocked=true 92 * unlock(m->wait_lock) 93 * 94 * up_write() 95 * sem->writelocked=false 96 * rtmutex_unlock(m) 97 * down_read() 98 * down_write() 99 * rtmutex_lock(m) 100 * wait() 101 * rtmutex_lock(m) 102 * 103 * That would put Reader1 behind the writer waiting on 104 * Reader2 to call up_read(), which might be unbound. 105 */ 106 107 trace_contention_begin(rwb, LCB_F_RT | LCB_F_READ); 108 109 /* 110 * For rwlocks this returns 0 unconditionally, so the below 111 * !ret conditionals are optimized out. 112 */ 113 ret = rwbase_rtmutex_slowlock_locked(rtm, state); 114 115 /* 116 * On success the rtmutex is held, so there can't be a writer 117 * active. Increment the reader count and immediately drop the 118 * rtmutex again. 119 * 120 * rtmutex->wait_lock has to be unlocked in any case of course. 121 */ 122 if (!ret) 123 atomic_inc(&rwb->readers); 124 raw_spin_unlock_irq(&rtm->wait_lock); 125 if (!ret) 126 rwbase_rtmutex_unlock(rtm); 127 128 trace_contention_end(rwb, ret); 129 rwbase_post_schedule(); 130 return ret; 131 } 132 133 static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb, 134 unsigned int state) 135 { 136 if (rwbase_read_trylock(rwb)) 137 return 0; 138 139 return __rwbase_read_lock(rwb, state); 140 } 141 142 static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb, 143 unsigned int state) 144 { 145 struct rt_mutex_base *rtm = &rwb->rtmutex; 146 struct task_struct *owner; 147 DEFINE_RT_WAKE_Q(wqh); 148 149 raw_spin_lock_irq(&rtm->wait_lock); 150 /* 151 * Wake the writer, i.e. the rtmutex owner. It might release the 152 * rtmutex concurrently in the fast path (due to a signal), but to 153 * clean up rwb->readers it needs to acquire rtm->wait_lock. The 154 * worst case which can happen is a spurious wakeup. 155 */ 156 owner = rt_mutex_owner(rtm); 157 if (owner) 158 rt_mutex_wake_q_add_task(&wqh, owner, state); 159 160 /* Pairs with the preempt_enable in rt_mutex_wake_up_q() */ 161 preempt_disable(); 162 raw_spin_unlock_irq(&rtm->wait_lock); 163 rt_mutex_wake_up_q(&wqh); 164 } 165 166 static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb, 167 unsigned int state) 168 { 169 /* 170 * rwb->readers can only hit 0 when a writer is waiting for the 171 * active readers to leave the critical section. 172 * 173 * dec_and_test() is fully ordered, provides RELEASE. 174 */ 175 if (unlikely(atomic_dec_and_test(&rwb->readers))) 176 __rwbase_read_unlock(rwb, state); 177 } 178 179 static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias, 180 unsigned long flags) 181 { 182 struct rt_mutex_base *rtm = &rwb->rtmutex; 183 184 /* 185 * _release() is needed in case that reader is in fast path, pairing 186 * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock(). 187 */ 188 (void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers); 189 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags); 190 rwbase_rtmutex_unlock(rtm); 191 } 192 193 static inline void rwbase_write_unlock(struct rwbase_rt *rwb) 194 { 195 struct rt_mutex_base *rtm = &rwb->rtmutex; 196 unsigned long flags; 197 198 raw_spin_lock_irqsave(&rtm->wait_lock, flags); 199 __rwbase_write_unlock(rwb, WRITER_BIAS, flags); 200 } 201 202 static inline void rwbase_write_downgrade(struct rwbase_rt *rwb) 203 { 204 struct rt_mutex_base *rtm = &rwb->rtmutex; 205 unsigned long flags; 206 207 raw_spin_lock_irqsave(&rtm->wait_lock, flags); 208 /* Release it and account current as reader */ 209 __rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags); 210 } 211 212 static inline bool __rwbase_write_trylock(struct rwbase_rt *rwb) 213 { 214 /* Can do without CAS because we're serialized by wait_lock. */ 215 lockdep_assert_held(&rwb->rtmutex.wait_lock); 216 217 /* 218 * _acquire is needed in case the reader is in the fast path, pairing 219 * with rwbase_read_unlock(), provides ACQUIRE. 220 */ 221 if (!atomic_read_acquire(&rwb->readers)) { 222 atomic_set(&rwb->readers, WRITER_BIAS); 223 return 1; 224 } 225 226 return 0; 227 } 228 229 static int __sched rwbase_write_lock(struct rwbase_rt *rwb, 230 unsigned int state) 231 { 232 struct rt_mutex_base *rtm = &rwb->rtmutex; 233 unsigned long flags; 234 235 /* Take the rtmutex as a first step */ 236 if (rwbase_rtmutex_lock_state(rtm, state)) 237 return -EINTR; 238 239 /* Force readers into slow path */ 240 atomic_sub(READER_BIAS, &rwb->readers); 241 242 rwbase_pre_schedule(); 243 244 raw_spin_lock_irqsave(&rtm->wait_lock, flags); 245 if (__rwbase_write_trylock(rwb)) 246 goto out_unlock; 247 248 rwbase_set_and_save_current_state(state); 249 trace_contention_begin(rwb, LCB_F_RT | LCB_F_WRITE); 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 rwbase_post_schedule(); 256 trace_contention_end(rwb, -EINTR); 257 return -EINTR; 258 } 259 260 if (__rwbase_write_trylock(rwb)) 261 break; 262 263 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags); 264 rwbase_schedule(); 265 raw_spin_lock_irqsave(&rtm->wait_lock, flags); 266 267 set_current_state(state); 268 } 269 rwbase_restore_current_state(); 270 trace_contention_end(rwb, 0); 271 272 out_unlock: 273 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags); 274 rwbase_post_schedule(); 275 return 0; 276 } 277 278 static inline int rwbase_write_trylock(struct rwbase_rt *rwb) 279 { 280 struct rt_mutex_base *rtm = &rwb->rtmutex; 281 unsigned long flags; 282 283 if (!rwbase_rtmutex_trylock(rtm)) 284 return 0; 285 286 atomic_sub(READER_BIAS, &rwb->readers); 287 288 raw_spin_lock_irqsave(&rtm->wait_lock, flags); 289 if (__rwbase_write_trylock(rwb)) { 290 raw_spin_unlock_irqrestore(&rtm->wait_lock, flags); 291 return 1; 292 } 293 __rwbase_write_unlock(rwb, 0, flags); 294 return 0; 295 } 296