History

futex: Split out PI futex

Move the PI futex implementation into it's own file.

Peter Zijlstra committed 5 years ago

Commit 85dc28f1 parent 966cb754 files · +1452 −1408Browse files

Unified · Split

4 files
  1. kernel/futex/Makefile +1 −1
  2. kernel/futex/core.c +101 −1407
  3. kernel/futex/futex.h +117 −0
  4. kernel/futex/pi.c +1233 −0
kernel/futex/Makefile+1−1
@@ -1,3 +1,3 @@
11 # SPDX-License-Identifier: GPL-2.0
22
3−obj-y += core.o syscalls.o
3+obj-y += core.o syscalls.o pi.o
kernel/futex/core.c+101−1407
@@ -147,67 +147,6 @@ int __read_mostly futex_cmpxchg_enabled; · 146 unmodified lines
147147 #endif
148148
149149
150−/*
151− * Priority Inheritance state:
152− */
153−struct futex_pi_state {
154− /*
155− * list of 'owned' pi_state instances - these have to be
156− * cleaned up in do_exit() if the task exits prematurely:
157− */
158− struct list_head list;
159−
160− /*
161− * The PI object:
162− */
163− struct rt_mutex_base pi_mutex;
164−
165− struct task_struct *owner;
166− refcount_t refcount;
167−
168− union futex_key key;
169−} __randomize_layout;
170−
171−/**
172− * struct futex_q - The hashed futex queue entry, one per waiting task
173− * @list: priority-sorted list of tasks waiting on this futex
174− * @task: the task waiting on the futex
175− * @lock_ptr: the hash bucket lock
176− * @key: the key the futex is hashed on
177− * @pi_state: optional priority inheritance state
178− * @rt_waiter: rt_waiter storage for use with requeue_pi
179− * @requeue_pi_key: the requeue_pi target futex key
180− * @bitset: bitset for the optional bitmasked wakeup
181− * @requeue_state: State field for futex_requeue_pi()
182− * @requeue_wait: RCU wait for futex_requeue_pi() (RT only)
183− *
184− * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
185− * we can wake only the relevant ones (hashed queues may be shared).
186− *
187− * A futex_q has a woken state, just like tasks have TASK_RUNNING.
188− * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
189− * The order of wakeup is always to make the first condition true, then
190− * the second.
191− *
192− * PI futexes are typically woken before they are removed from the hash list via
193− * the rt_mutex code. See futex_unqueue_pi().
194− */
195−struct futex_q {
196− struct plist_node list;
197−
198− struct task_struct *task;
199− spinlock_t *lock_ptr;
200− union futex_key key;
201− struct futex_pi_state *pi_state;
202− struct rt_mutex_waiter *rt_waiter;
203− union futex_key *requeue_pi_key;
204− u32 bitset;
205− atomic_t requeue_state;
206−#ifdef CONFIG_PREEMPT_RT
207− struct rcuwait requeue_wait;
208−#endif
209−} __randomize_layout;
210−
211150 /*
212151 * On PREEMPT_RT, the hash bucket lock is a 'sleeping' spinlock with an
213152 * underlying rtmutex. The task which is about to be requeued could have
@@ -259,24 +198,13 @@ enum { · 45 unmodified lines
259198 Q_REQUEUE_PI_LOCKED,
260199 };
261200
262−static const struct futex_q futex_q_init = {
201+const struct futex_q futex_q_init = {
263202 /* list gets initialized in futex_queue()*/
264203 .key = FUTEX_KEY_INIT,
265204 .bitset = FUTEX_BITSET_MATCH_ANY,
266205 .requeue_state = ATOMIC_INIT(Q_REQUEUE_PI_NONE),
267206 };
268207
269−/*
270− * Hash buckets are shared by all the futex_keys that hash to the same
271− * location. Each key may have multiple futex_q structures, one for each task
272− * waiting on a futex.
273− */
274−struct futex_hash_bucket {
275− atomic_t waiters;
276− spinlock_t lock;
277− struct plist_head chain;
278−} ____cacheline_aligned_in_smp;
279−
280208 /*
281209 * The base of the bucket array and its size are always used together
282210 * (after initialization only in futex_hash()), so ensure that they
@@ -386,7 +314,7 @@ static inline int hb_waiters_pending(struct futex_hash_bucket *hb) · 103 unmodified lines
386314 * We hash on the keys returned from get_futex_key (see below) and return the
387315 * corresponding hash bucket in the global hash.
388316 */
389−static struct futex_hash_bucket *futex_hash(union futex_key *key)
317+struct futex_hash_bucket *futex_hash(union futex_key *key)
390318 {
391319 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
392320 key->both.offset);
@@ -410,11 +338,6 @@ static inline int match_futex(union futex_key *key1, union futex_key *key2) · 17 unmodified lines
410338 && key1->both.offset == key2->both.offset);
411339 }
412340
413−enum futex_access {
414− FUTEX_READ,
415− FUTEX_WRITE
416−};
417−
418341 /**
419342 * futex_setup_timer - set up the sleeping hrtimer.
420343 * @time: ptr to the given timeout value
@@ -425,7 +348,7 @@ enum futex_access { · 4 unmodified lines
425348 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
426349 * value given
427350 */
428−static inline struct hrtimer_sleeper *
351+struct hrtimer_sleeper *
429352 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
430353 int flags, u64 range_ns)
431354 {
@@ -511,8 +434,8 @@ static u64 get_inode_sequence_number(struct inode *inode) · 79 unmodified lines
511434 *
512435 * lock_page() might sleep, the caller should not hold a spinlock.
513436 */
514−static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
515− enum futex_access rw)
437+int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
438+ enum futex_access rw)
516439 {
517440 unsigned long address = (unsigned long)uaddr;
518441 struct mm_struct *mm = current->mm;
@@ -700,7 +623,7 @@ out: · 181 unmodified lines
700623 * disabled section so we can as well avoid the #PF overhead by
701624 * calling get_user_pages() right away.
702625 */
703−static int fault_in_user_writeable(u32 __user *uaddr)
626+int fault_in_user_writeable(u32 __user *uaddr)
704627 {
705628 struct mm_struct *mm = current->mm;
706629 int ret;
@@ -720,8 +643,7 @@ static int fault_in_user_writeable(u32 __user *uaddr) · 13 unmodified lines
720643 *
721644 * Must be called with the hb lock held.
722645 */
723−static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
724− union futex_key *key)
646+struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, union futex_key *key)
725647 {
726648 struct futex_q *this;
727649
@@ -732,8 +654,7 @@ static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, · 4 unmodified lines
732654 return NULL;
733655 }
734656
735−static int futex_cmpxchg_value_locked(u32 *curval, u32 __user *uaddr,
736− u32 uval, u32 newval)
657+int futex_cmpxchg_value_locked(u32 *curval, u32 __user *uaddr, u32 uval, u32 newval)
737658 {
738659 int ret;
739660
@@ -744,7 +665,7 @@ static int futex_cmpxchg_value_locked(u32 *curval, u32 __user *uaddr, · 4 unmodified lines
744665 return ret;
745666 }
746667
747−static int futex_get_value_locked(u32 *dest, u32 __user *from)
668+int futex_get_value_locked(u32 *dest, u32 __user *from)
748669 {
749670 int ret;
750671
@@ -755,399 +676,6 @@ static int futex_get_value_locked(u32 *dest, u32 __user *from) · 4 unmodified lines
755676 return ret ? -EFAULT : 0;
756677 }
757678
758−
759−/*
760− * PI code:
761− */
762−static int refill_pi_state_cache(void)
763−{
764− struct futex_pi_state *pi_state;
765−
766− if (likely(current->pi_state_cache))
767− return 0;
768−
769− pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
770−
771− if (!pi_state)
772− return -ENOMEM;
773−
774− INIT_LIST_HEAD(&pi_state->list);
775− /* pi_mutex gets initialized later */
776− pi_state->owner = NULL;
777− refcount_set(&pi_state->refcount, 1);
778− pi_state->key = FUTEX_KEY_INIT;
779−
780− current->pi_state_cache = pi_state;
781−
782− return 0;
783−}
784−
785−static struct futex_pi_state *alloc_pi_state(void)
786−{
787− struct futex_pi_state *pi_state = current->pi_state_cache;
788−
789− WARN_ON(!pi_state);
790− current->pi_state_cache = NULL;
791−
792− return pi_state;
793−}
794−
795−static void pi_state_update_owner(struct futex_pi_state *pi_state,
796− struct task_struct *new_owner)
797−{
798− struct task_struct *old_owner = pi_state->owner;
799−
800− lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
801−
802− if (old_owner) {
803− raw_spin_lock(&old_owner->pi_lock);
804− WARN_ON(list_empty(&pi_state->list));
805− list_del_init(&pi_state->list);
806− raw_spin_unlock(&old_owner->pi_lock);
807− }
808−
809− if (new_owner) {
810− raw_spin_lock(&new_owner->pi_lock);
811− WARN_ON(!list_empty(&pi_state->list));
812− list_add(&pi_state->list, &new_owner->pi_state_list);
813− pi_state->owner = new_owner;
814− raw_spin_unlock(&new_owner->pi_lock);
815− }
816−}
817−
818−static void get_pi_state(struct futex_pi_state *pi_state)
819−{
820− WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
821−}
822−
823−/*
824− * Drops a reference to the pi_state object and frees or caches it
825− * when the last reference is gone.
826− */
827−static void put_pi_state(struct futex_pi_state *pi_state)
828−{
829− if (!pi_state)
830− return;
831−
832− if (!refcount_dec_and_test(&pi_state->refcount))
833− return;
834−
835− /*
836− * If pi_state->owner is NULL, the owner is most probably dying
837− * and has cleaned up the pi_state already
838− */
839− if (pi_state->owner) {
840− unsigned long flags;
841−
842− raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
843− pi_state_update_owner(pi_state, NULL);
844− rt_mutex_proxy_unlock(&pi_state->pi_mutex);
845− raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
846− }
847−
848− if (current->pi_state_cache) {
849− kfree(pi_state);
850− } else {
851− /*
852− * pi_state->list is already empty.
853− * clear pi_state->owner.
854− * refcount is at 0 - put it back to 1.
855− */
856− pi_state->owner = NULL;
857− refcount_set(&pi_state->refcount, 1);
858− current->pi_state_cache = pi_state;
859− }
860−}
861−
862−#ifdef CONFIG_FUTEX_PI
863−
864−/*
865− * This task is holding PI mutexes at exit time => bad.
866− * Kernel cleans up PI-state, but userspace is likely hosed.
867− * (Robust-futex cleanup is separate and might save the day for userspace.)
868− */
869−static void exit_pi_state_list(struct task_struct *curr)
870−{
871− struct list_head *next, *head = &curr->pi_state_list;
872− struct futex_pi_state *pi_state;
873− struct futex_hash_bucket *hb;
874− union futex_key key = FUTEX_KEY_INIT;
875−
876− if (!futex_cmpxchg_enabled)
877− return;
878− /*
879− * We are a ZOMBIE and nobody can enqueue itself on
880− * pi_state_list anymore, but we have to be careful
881− * versus waiters unqueueing themselves:
882− */
883− raw_spin_lock_irq(&curr->pi_lock);
884− while (!list_empty(head)) {
885− next = head->next;
886− pi_state = list_entry(next, struct futex_pi_state, list);
887− key = pi_state->key;
888− hb = futex_hash(&key);
889−
890− /*
891− * We can race against put_pi_state() removing itself from the
892− * list (a waiter going away). put_pi_state() will first
893− * decrement the reference count and then modify the list, so
894− * its possible to see the list entry but fail this reference
895− * acquire.
896− *
897− * In that case; drop the locks to let put_pi_state() make
898− * progress and retry the loop.
899− */
900− if (!refcount_inc_not_zero(&pi_state->refcount)) {
901− raw_spin_unlock_irq(&curr->pi_lock);
902− cpu_relax();
903− raw_spin_lock_irq(&curr->pi_lock);
904− continue;
905− }
906− raw_spin_unlock_irq(&curr->pi_lock);
907−
908− spin_lock(&hb->lock);
909− raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
910− raw_spin_lock(&curr->pi_lock);
911− /*
912− * We dropped the pi-lock, so re-check whether this
913− * task still owns the PI-state:
914− */
915− if (head->next != next) {
916− /* retain curr->pi_lock for the loop invariant */
917− raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
918− spin_unlock(&hb->lock);
919− put_pi_state(pi_state);
920− continue;
921− }
922−
923− WARN_ON(pi_state->owner != curr);
924− WARN_ON(list_empty(&pi_state->list));
925− list_del_init(&pi_state->list);
926− pi_state->owner = NULL;
927−
928− raw_spin_unlock(&curr->pi_lock);
929− raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
930− spin_unlock(&hb->lock);
931−
932− rt_mutex_futex_unlock(&pi_state->pi_mutex);
933− put_pi_state(pi_state);
934−
935− raw_spin_lock_irq(&curr->pi_lock);
936− }
937− raw_spin_unlock_irq(&curr->pi_lock);
938−}
939−#else
940−static inline void exit_pi_state_list(struct task_struct *curr) { }
941−#endif
942−
943−/*
944− * We need to check the following states:
945− *
946− * Waiter | pi_state | pi->owner | uTID | uODIED | ?
947− *
948− * [1] NULL | --- | --- | 0 | 0/1 | Valid
949− * [2] NULL | --- | --- | >0 | 0/1 | Valid
950− *
951− * [3] Found | NULL | -- | Any | 0/1 | Invalid
952− *
953− * [4] Found | Found | NULL | 0 | 1 | Valid
954− * [5] Found | Found | NULL | >0 | 1 | Invalid
955− *
956− * [6] Found | Found | task | 0 | 1 | Valid
957− *
958− * [7] Found | Found | NULL | Any | 0 | Invalid
959− *
960− * [8] Found | Found | task | ==taskTID | 0/1 | Valid
961− * [9] Found | Found | task | 0 | 0 | Invalid
962− * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
963− *
964− * [1] Indicates that the kernel can acquire the futex atomically. We
965− * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
966− *
967− * [2] Valid, if TID does not belong to a kernel thread. If no matching
968− * thread is found then it indicates that the owner TID has died.
969− *
970− * [3] Invalid. The waiter is queued on a non PI futex
971− *
972− * [4] Valid state after exit_robust_list(), which sets the user space
973− * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
974− *
975− * [5] The user space value got manipulated between exit_robust_list()
976− * and exit_pi_state_list()
977− *
978− * [6] Valid state after exit_pi_state_list() which sets the new owner in
979− * the pi_state but cannot access the user space value.
980− *
981− * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
982− *
983− * [8] Owner and user space value match
984− *
985− * [9] There is no transient state which sets the user space TID to 0
986− * except exit_robust_list(), but this is indicated by the
987− * FUTEX_OWNER_DIED bit. See [4]
988− *
989− * [10] There is no transient state which leaves owner and user space
990− * TID out of sync. Except one error case where the kernel is denied
991− * write access to the user address, see fixup_pi_state_owner().
992− *
993− *
994− * Serialization and lifetime rules:
995− *
996− * hb->lock:
997− *
998− * hb -> futex_q, relation
999− * futex_q -> pi_state, relation
1000− *
1001− * (cannot be raw because hb can contain arbitrary amount
1002− * of futex_q's)
1003− *
1004− * pi_mutex->wait_lock:
1005− *
1006− * {uval, pi_state}
1007− *
1008− * (and pi_mutex 'obviously')
1009− *
1010− * p->pi_lock:
1011− *
1012− * p->pi_state_list -> pi_state->list, relation
1013− * pi_mutex->owner -> pi_state->owner, relation
1014− *
1015− * pi_state->refcount:
1016− *
1017− * pi_state lifetime
1018− *
1019− *
1020− * Lock order:
1021− *
1022− * hb->lock
1023− * pi_mutex->wait_lock
1024− * p->pi_lock
1025− *
1026− */
1027−
1028−/*
1029− * Validate that the existing waiter has a pi_state and sanity check
1030− * the pi_state against the user space value. If correct, attach to
1031− * it.
1032− */
1033−static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1034− struct futex_pi_state *pi_state,
1035− struct futex_pi_state **ps)
1036−{
1037− pid_t pid = uval & FUTEX_TID_MASK;
1038− u32 uval2;
1039− int ret;
1040−
1041− /*
1042− * Userspace might have messed up non-PI and PI futexes [3]
1043− */
1044− if (unlikely(!pi_state))
1045− return -EINVAL;
1046−
1047− /*
1048− * We get here with hb->lock held, and having found a
1049− * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1050− * has dropped the hb->lock in between futex_queue() and futex_unqueue_pi(),
1051− * which in turn means that futex_lock_pi() still has a reference on
1052− * our pi_state.
1053− *
1054− * The waiter holding a reference on @pi_state also protects against
1055− * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1056− * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1057− * free pi_state before we can take a reference ourselves.
1058− */
1059− WARN_ON(!refcount_read(&pi_state->refcount));
1060−
1061− /*
1062− * Now that we have a pi_state, we can acquire wait_lock
1063− * and do the state validation.
1064− */
1065− raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1066−
1067− /*
1068− * Since {uval, pi_state} is serialized by wait_lock, and our current
1069− * uval was read without holding it, it can have changed. Verify it
1070− * still is what we expect it to be, otherwise retry the entire
1071− * operation.
1072− */
1073− if (futex_get_value_locked(&uval2, uaddr))
1074− goto out_efault;
1075−
1076− if (uval != uval2)
1077− goto out_eagain;
1078−
1079− /*
1080− * Handle the owner died case:
1081− */
1082− if (uval & FUTEX_OWNER_DIED) {
1083− /*
1084− * exit_pi_state_list sets owner to NULL and wakes the
1085− * topmost waiter. The task which acquires the
1086− * pi_state->rt_mutex will fixup owner.
1087− */
1088− if (!pi_state->owner) {
1089− /*
1090− * No pi state owner, but the user space TID
1091− * is not 0. Inconsistent state. [5]
1092− */
1093− if (pid)
1094− goto out_einval;
1095− /*
1096− * Take a ref on the state and return success. [4]
1097− */
1098− goto out_attach;
1099− }
1100−
1101− /*
1102− * If TID is 0, then either the dying owner has not
1103− * yet executed exit_pi_state_list() or some waiter
1104− * acquired the rtmutex in the pi state, but did not
1105− * yet fixup the TID in user space.
1106− *
1107− * Take a ref on the state and return success. [6]
1108− */
1109− if (!pid)
1110− goto out_attach;
1111− } else {
1112− /*
1113− * If the owner died bit is not set, then the pi_state
1114− * must have an owner. [7]
1115− */
1116− if (!pi_state->owner)
1117− goto out_einval;
1118− }
1119−
1120− /*
1121− * Bail out if user space manipulated the futex value. If pi
1122− * state exists then the owner TID must be the same as the
1123− * user space TID. [9/10]
1124− */
1125− if (pid != task_pid_vnr(pi_state->owner))
1126− goto out_einval;
1127−
1128−out_attach:
1129− get_pi_state(pi_state);
1130− raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1131− *ps = pi_state;
1132− return 0;
1133−
1134−out_einval:
1135− ret = -EINVAL;
1136− goto out_error;
1137−
1138−out_eagain:
1139− ret = -EAGAIN;
1140− goto out_error;
1141−
1142−out_efault:
1143− ret = -EFAULT;
1144− goto out_error;
1145−
1146−out_error:
1147− raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1148− return ret;
1149−}
1150−
1151679 /**
1152680 * wait_for_owner_exiting - Block until the owner has exited
1153681 * @ret: owner's current futex lock status
@@ -1155,7 +683,7 @@ out_error: · 1 unmodified line
1155683 *
1156684 * Caller must hold a refcount on @exiting.
1157685 */
1158−static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
686+void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1159687 {
1160688 if (ret != -EBUSY) {
1161689 WARN_ON_ONCE(exiting);
@@ -1179,296 +707,6 @@ static void wait_for_owner_exiting(int ret, struct task_struct *exiting) · 17 unmodified lines
1179707 put_task_struct(exiting);
1180708 }
1181709
1182−static int handle_exit_race(u32 __user *uaddr, u32 uval,
1183− struct task_struct *tsk)
1184−{
1185− u32 uval2;
1186−
1187− /*
1188− * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1189− * caller that the alleged owner is busy.
1190− */
1191− if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1192− return -EBUSY;
1193−
1194− /*
1195− * Reread the user space value to handle the following situation:
1196− *
1197− * CPU0 CPU1
1198− *
1199− * sys_exit() sys_futex()
1200− * do_exit() futex_lock_pi()
1201− * futex_lock_pi_atomic()
1202− * exit_signals(tsk) No waiters:
1203− * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1204− * mm_release(tsk) Set waiter bit
1205− * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1206− * Set owner died attach_to_pi_owner() {
1207− * *uaddr = 0xC0000000; tsk = get_task(PID);
1208− * } if (!tsk->flags & PF_EXITING) {
1209− * ... attach();
1210− * tsk->futex_state = } else {
1211− * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1212− * FUTEX_STATE_DEAD)
1213− * return -EAGAIN;
1214− * return -ESRCH; <--- FAIL
1215− * }
1216− *
1217− * Returning ESRCH unconditionally is wrong here because the
1218− * user space value has been changed by the exiting task.
1219− *
1220− * The same logic applies to the case where the exiting task is
1221− * already gone.
1222− */
1223− if (futex_get_value_locked(&uval2, uaddr))
1224− return -EFAULT;
1225−
1226− /* If the user space value has changed, try again. */
1227− if (uval2 != uval)
1228− return -EAGAIN;
1229−
1230− /*
1231− * The exiting task did not have a robust list, the robust list was
1232− * corrupted or the user space value in *uaddr is simply bogus.
1233− * Give up and tell user space.
1234− */
1235− return -ESRCH;
1236−}
1237−
1238−static void __attach_to_pi_owner(struct task_struct *p, union futex_key *key,
1239− struct futex_pi_state **ps)
1240−{
1241− /*
1242− * No existing pi state. First waiter. [2]
1243− *
1244− * This creates pi_state, we have hb->lock held, this means nothing can
1245− * observe this state, wait_lock is irrelevant.
1246− */
1247− struct futex_pi_state *pi_state = alloc_pi_state();
1248−
1249− /*
1250− * Initialize the pi_mutex in locked state and make @p
1251− * the owner of it:
1252− */
1253− rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1254−
1255− /* Store the key for possible exit cleanups: */
1256− pi_state->key = *key;
1257−
1258− WARN_ON(!list_empty(&pi_state->list));
1259− list_add(&pi_state->list, &p->pi_state_list);
1260− /*
1261− * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1262− * because there is no concurrency as the object is not published yet.
1263− */
1264− pi_state->owner = p;
1265−
1266− *ps = pi_state;
1267−}
1268−/*
1269− * Lookup the task for the TID provided from user space and attach to
1270− * it after doing proper sanity checks.
1271− */
1272−static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1273− struct futex_pi_state **ps,
1274− struct task_struct **exiting)
1275−{
1276− pid_t pid = uval & FUTEX_TID_MASK;
1277− struct task_struct *p;
1278−
1279− /*
1280− * We are the first waiter - try to look up the real owner and attach
1281− * the new pi_state to it, but bail out when TID = 0 [1]
1282− *
1283− * The !pid check is paranoid. None of the call sites should end up
1284− * with pid == 0, but better safe than sorry. Let the caller retry
1285− */
1286− if (!pid)
1287− return -EAGAIN;
1288− p = find_get_task_by_vpid(pid);
1289− if (!p)
1290− return handle_exit_race(uaddr, uval, NULL);
1291−
1292− if (unlikely(p->flags & PF_KTHREAD)) {
1293− put_task_struct(p);
1294− return -EPERM;
1295− }
1296−
1297− /*
1298− * We need to look at the task state to figure out, whether the
1299− * task is exiting. To protect against the change of the task state
1300− * in futex_exit_release(), we do this protected by p->pi_lock:
1301− */
1302− raw_spin_lock_irq(&p->pi_lock);
1303− if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1304− /*
1305− * The task is on the way out. When the futex state is
1306− * FUTEX_STATE_DEAD, we know that the task has finished
1307− * the cleanup:
1308− */
1309− int ret = handle_exit_race(uaddr, uval, p);
1310−
1311− raw_spin_unlock_irq(&p->pi_lock);
1312− /*
1313− * If the owner task is between FUTEX_STATE_EXITING and
1314− * FUTEX_STATE_DEAD then store the task pointer and keep
1315− * the reference on the task struct. The calling code will
1316− * drop all locks, wait for the task to reach
1317− * FUTEX_STATE_DEAD and then drop the refcount. This is
1318− * required to prevent a live lock when the current task
1319− * preempted the exiting task between the two states.
1320− */
1321− if (ret == -EBUSY)
1322− *exiting = p;
1323− else
1324− put_task_struct(p);
1325− return ret;
1326− }
1327−
1328− __attach_to_pi_owner(p, key, ps);
1329− raw_spin_unlock_irq(&p->pi_lock);
1330−
1331− put_task_struct(p);
1332−
1333− return 0;
1334−}
1335−
1336−static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1337−{
1338− int err;
1339− u32 curval;
1340−
1341− if (unlikely(should_fail_futex(true)))
1342− return -EFAULT;
1343−
1344− err = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
1345− if (unlikely(err))
1346− return err;
1347−
1348− /* If user space value changed, let the caller retry */
1349− return curval != uval ? -EAGAIN : 0;
1350−}
1351−
1352−/**
1353− * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1354− * @uaddr: the pi futex user address
1355− * @hb: the pi futex hash bucket
1356− * @key: the futex key associated with uaddr and hb
1357− * @ps: the pi_state pointer where we store the result of the
1358− * lookup
1359− * @task: the task to perform the atomic lock work for. This will
1360− * be "current" except in the case of requeue pi.
1361− * @exiting: Pointer to store the task pointer of the owner task
1362− * which is in the middle of exiting
1363− * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1364− *
1365− * Return:
1366− * - 0 - ready to wait;
1367− * - 1 - acquired the lock;
1368− * - <0 - error
1369− *
1370− * The hb->lock must be held by the caller.
1371− *
1372− * @exiting is only set when the return value is -EBUSY. If so, this holds
1373− * a refcount on the exiting task on return and the caller needs to drop it
1374− * after waiting for the exit to complete.
1375− */
1376−static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1377− union futex_key *key,
1378− struct futex_pi_state **ps,
1379− struct task_struct *task,
1380− struct task_struct **exiting,
1381− int set_waiters)
1382−{
1383− u32 uval, newval, vpid = task_pid_vnr(task);
1384− struct futex_q *top_waiter;
1385− int ret;
1386−
1387− /*
1388− * Read the user space value first so we can validate a few
1389− * things before proceeding further.
1390− */
1391− if (futex_get_value_locked(&uval, uaddr))
1392− return -EFAULT;
1393−
1394− if (unlikely(should_fail_futex(true)))
1395− return -EFAULT;
1396−
1397− /*
1398− * Detect deadlocks.
1399− */
1400− if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1401− return -EDEADLK;
1402−
1403− if ((unlikely(should_fail_futex(true))))
1404− return -EDEADLK;
1405−
1406− /*
1407− * Lookup existing state first. If it exists, try to attach to
1408− * its pi_state.
1409− */
1410− top_waiter = futex_top_waiter(hb, key);
1411− if (top_waiter)
1412− return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1413−
1414− /*
1415− * No waiter and user TID is 0. We are here because the
1416− * waiters or the owner died bit is set or called from
1417− * requeue_cmp_pi or for whatever reason something took the
1418− * syscall.
1419− */
1420− if (!(uval & FUTEX_TID_MASK)) {
1421− /*
1422− * We take over the futex. No other waiters and the user space
1423− * TID is 0. We preserve the owner died bit.
1424− */
1425− newval = uval & FUTEX_OWNER_DIED;
1426− newval |= vpid;
1427−
1428− /* The futex requeue_pi code can enforce the waiters bit */
1429− if (set_waiters)
1430− newval |= FUTEX_WAITERS;
1431−
1432− ret = lock_pi_update_atomic(uaddr, uval, newval);
1433− if (ret)
1434− return ret;
1435−
1436− /*
1437− * If the waiter bit was requested the caller also needs PI
1438− * state attached to the new owner of the user space futex.
1439− *
1440− * @task is guaranteed to be alive and it cannot be exiting
1441− * because it is either sleeping or waiting in
1442− * futex_requeue_pi_wakeup_sync().
1443− *
1444− * No need to do the full attach_to_pi_owner() exercise
1445− * because @task is known and valid.
1446− */
1447− if (set_waiters) {
1448− raw_spin_lock_irq(&task->pi_lock);
1449− __attach_to_pi_owner(task, key, ps);
1450− raw_spin_unlock_irq(&task->pi_lock);
1451− }
1452− return 1;
1453− }
1454−
1455− /*
1456− * First waiter. Set the waiters bit before attaching ourself to
1457− * the owner. If owner tries to unlock, it will be forced into
1458− * the kernel and blocked on hb->lock.
1459− */
1460− newval = uval | FUTEX_WAITERS;
1461− ret = lock_pi_update_atomic(uaddr, uval, newval);
1462− if (ret)
1463− return ret;
1464− /*
1465− * If the update of the user space value succeeded, we try to
1466− * attach to the owner. If that fails, no harm done, we only
1467− * set the FUTEX_WAITERS bit in the user space variable.
1468− */
1469− return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1470−}
1471−
1472710 /**
1473711 * __futex_unqueue() - Remove the futex_q from its futex_hash_bucket
1474712 * @q: The futex_q to unqueue
@@ -1519,79 +757,6 @@ static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q) · 44 unmodified lines
1519757 wake_q_add_safe(wake_q, p);
1520758 }
1521759
1522−/*
1523− * Caller must hold a reference on @pi_state.
1524− */
1525−static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1526−{
1527− struct rt_mutex_waiter *top_waiter;
1528− struct task_struct *new_owner;
1529− bool postunlock = false;
1530− DEFINE_RT_WAKE_Q(wqh);
1531− u32 curval, newval;
1532− int ret = 0;
1533−
1534− top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1535− if (WARN_ON_ONCE(!top_waiter)) {
1536− /*
1537− * As per the comment in futex_unlock_pi() this should not happen.
1538− *
1539− * When this happens, give up our locks and try again, giving
1540− * the futex_lock_pi() instance time to complete, either by
1541− * waiting on the rtmutex or removing itself from the futex
1542− * queue.
1543− */
1544− ret = -EAGAIN;
1545− goto out_unlock;
1546− }
1547−
1548− new_owner = top_waiter->task;
1549−
1550− /*
1551− * We pass it to the next owner. The WAITERS bit is always kept
1552− * enabled while there is PI state around. We cleanup the owner
1553− * died bit, because we are the owner.
1554− */
1555− newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1556−
1557− if (unlikely(should_fail_futex(true))) {
1558− ret = -EFAULT;
1559− goto out_unlock;
1560− }
1561−
1562− ret = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
1563− if (!ret && (curval != uval)) {
1564− /*
1565− * If a unconditional UNLOCK_PI operation (user space did not
1566− * try the TID->0 transition) raced with a waiter setting the
1567− * FUTEX_WAITERS flag between get_user() and locking the hash
1568− * bucket lock, retry the operation.
1569− */
1570− if ((FUTEX_TID_MASK & curval) == uval)
1571− ret = -EAGAIN;
1572− else
1573− ret = -EINVAL;
1574− }
1575−
1576− if (!ret) {
1577− /*
1578− * This is a point of no return; once we modified the uval
1579− * there is no going back and subsequent operations must
1580− * not fail.
1581− */
1582− pi_state_update_owner(pi_state, new_owner);
1583− postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wqh);
1584− }
1585−
1586−out_unlock:
1587− raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1588−
1589− if (postunlock)
1590− rt_mutex_postunlock(&wqh);
1591−
1592− return ret;
1593−}
1594−
1595760 /*
1596761 * Express the locking dependencies for lockdep:
1597762 */
@@ -2410,7 +1575,7 @@ out_unlock: · 812 unmodified lines
24101575 }
24111576
24121577 /* The key must be already stored in q->key. */
2413−static inline struct futex_hash_bucket *futex_q_lock(struct futex_q *q)
1578+struct futex_hash_bucket *futex_q_lock(struct futex_q *q)
24141579 __acquires(&hb->lock)
24151580 {
24161581 struct futex_hash_bucket *hb;
@@ -2433,15 +1598,14 @@ static inline struct futex_hash_bucket *futex_q_lock(struct futex_q *q) · 16 unmodified lines
24331598 return hb;
24341599 }
24351600
2436−static inline void
2437−futex_q_unlock(struct futex_hash_bucket *hb)
1601+void futex_q_unlock(struct futex_hash_bucket *hb)
24381602 __releases(&hb->lock)
24391603 {
24401604 spin_unlock(&hb->lock);
24411605 hb_waiters_dec(hb);
24421606 }
24431607
2444−static inline void __futex_queue(struct futex_q *q, struct futex_hash_bucket *hb)
1608+void __futex_queue(struct futex_q *q, struct futex_hash_bucket *hb)
24451609 {
24461610 int prio;
24471611
@@ -2537,255 +1701,17 @@ retry: · 89 unmodified lines
25371701 * PI futexes can not be requeued and must remove themselves from the
25381702 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
25391703 */
2540−static void futex_unqueue_pi(struct futex_q *q)
1704+void futex_unqueue_pi(struct futex_q *q)
25411705 {
25421706 __futex_unqueue(q);
25431707
25441708 BUG_ON(!q->pi_state);
25451709 put_pi_state(q->pi_state);
2546− q->pi_state = NULL;
2547−}
2548−
2549−static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2550− struct task_struct *argowner)
2551−{
2552− struct futex_pi_state *pi_state = q->pi_state;
2553− struct task_struct *oldowner, *newowner;
2554− u32 uval, curval, newval, newtid;
2555− int err = 0;
2556−
2557− oldowner = pi_state->owner;
2558−
2559− /*
2560− * We are here because either:
2561− *
2562− * - we stole the lock and pi_state->owner needs updating to reflect
2563− * that (@argowner == current),
2564− *
2565− * or:
2566− *
2567− * - someone stole our lock and we need to fix things to point to the
2568− * new owner (@argowner == NULL).
2569− *
2570− * Either way, we have to replace the TID in the user space variable.
2571− * This must be atomic as we have to preserve the owner died bit here.
2572− *
2573− * Note: We write the user space value _before_ changing the pi_state
2574− * because we can fault here. Imagine swapped out pages or a fork
2575− * that marked all the anonymous memory readonly for cow.
2576− *
2577− * Modifying pi_state _before_ the user space value would leave the
2578− * pi_state in an inconsistent state when we fault here, because we
2579− * need to drop the locks to handle the fault. This might be observed
2580− * in the PID checks when attaching to PI state .
2581− */
2582−retry:
2583− if (!argowner) {
2584− if (oldowner != current) {
2585− /*
2586− * We raced against a concurrent self; things are
2587− * already fixed up. Nothing to do.
2588− */
2589− return 0;
2590− }
2591−
2592− if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2593− /* We got the lock. pi_state is correct. Tell caller. */
2594− return 1;
2595− }
2596−
2597− /*
2598− * The trylock just failed, so either there is an owner or
2599− * there is a higher priority waiter than this one.
2600− */
2601− newowner = rt_mutex_owner(&pi_state->pi_mutex);
2602− /*
2603− * If the higher priority waiter has not yet taken over the
2604− * rtmutex then newowner is NULL. We can't return here with
2605− * that state because it's inconsistent vs. the user space
2606− * state. So drop the locks and try again. It's a valid
2607− * situation and not any different from the other retry
2608− * conditions.
2609− */
2610− if (unlikely(!newowner)) {
2611− err = -EAGAIN;
2612− goto handle_err;
2613− }
2614− } else {
2615− WARN_ON_ONCE(argowner != current);
2616− if (oldowner == current) {
2617− /*
2618− * We raced against a concurrent self; things are
2619− * already fixed up. Nothing to do.
2620− */
2621− return 1;
2622− }
2623− newowner = argowner;
2624− }
2625−
2626− newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2627− /* Owner died? */
2628− if (!pi_state->owner)
2629− newtid |= FUTEX_OWNER_DIED;
2630−
2631− err = futex_get_value_locked(&uval, uaddr);
2632− if (err)
2633− goto handle_err;
2634−
2635− for (;;) {
2636− newval = (uval & FUTEX_OWNER_DIED) | newtid;
2637−
2638− err = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
2639− if (err)
2640− goto handle_err;
2641−
2642− if (curval == uval)
2643− break;
2644− uval = curval;
2645− }
2646−
2647− /*
2648− * We fixed up user space. Now we need to fix the pi_state
2649− * itself.
2650− */
2651− pi_state_update_owner(pi_state, newowner);
2652−
2653− return argowner == current;
2654−
2655− /*
2656− * In order to reschedule or handle a page fault, we need to drop the
2657− * locks here. In the case of a fault, this gives the other task
2658− * (either the highest priority waiter itself or the task which stole
2659− * the rtmutex) the chance to try the fixup of the pi_state. So once we
2660− * are back from handling the fault we need to check the pi_state after
2661− * reacquiring the locks and before trying to do another fixup. When
2662− * the fixup has been done already we simply return.
2663− *
2664− * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2665− * drop hb->lock since the caller owns the hb -> futex_q relation.
2666− * Dropping the pi_mutex->wait_lock requires the state revalidate.
2667− */
2668−handle_err:
2669− raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2670− spin_unlock(q->lock_ptr);
2671−
2672− switch (err) {
2673− case -EFAULT:
2674− err = fault_in_user_writeable(uaddr);
2675− break;
2676−
2677− case -EAGAIN:
2678− cond_resched();
2679− err = 0;
2680− break;
2681−
2682− default:
2683− WARN_ON_ONCE(1);
2684− break;
2685− }
2686−
2687− spin_lock(q->lock_ptr);
2688− raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2689−
2690− /*
2691− * Check if someone else fixed it for us:
2692− */
2693− if (pi_state->owner != oldowner)
2694− return argowner == current;
2695−
2696− /* Retry if err was -EAGAIN or the fault in succeeded */
2697− if (!err)
2698− goto retry;
2699−
2700− /*
2701− * fault_in_user_writeable() failed so user state is immutable. At
2702− * best we can make the kernel state consistent but user state will
2703− * be most likely hosed and any subsequent unlock operation will be
2704− * rejected due to PI futex rule [10].
2705− *
2706− * Ensure that the rtmutex owner is also the pi_state owner despite
2707− * the user space value claiming something different. There is no
2708− * point in unlocking the rtmutex if current is the owner as it
2709− * would need to wait until the next waiter has taken the rtmutex
2710− * to guarantee consistent state. Keep it simple. Userspace asked
2711− * for this wreckaged state.
2712− *
2713− * The rtmutex has an owner - either current or some other
2714− * task. See the EAGAIN loop above.
2715− */
2716− pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2717−
2718− return err;
2719−}
2720−
2721−static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2722− struct task_struct *argowner)
2723−{
2724− struct futex_pi_state *pi_state = q->pi_state;
2725− int ret;
2726−
2727− lockdep_assert_held(q->lock_ptr);
2728−
2729− raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2730− ret = __fixup_pi_state_owner(uaddr, q, argowner);
2731− raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2732− return ret;
2733−}
2734−
2735−static long futex_wait_restart(struct restart_block *restart);
2736−
2737−/**
2738− * fixup_owner() - Post lock pi_state and corner case management
2739− * @uaddr: user address of the futex
2740− * @q: futex_q (contains pi_state and access to the rt_mutex)
2741− * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2742− *
2743− * After attempting to lock an rt_mutex, this function is called to cleanup
2744− * the pi_state owner as well as handle race conditions that may allow us to
2745− * acquire the lock. Must be called with the hb lock held.
2746− *
2747− * Return:
2748− * - 1 - success, lock taken;
2749− * - 0 - success, lock not taken;
2750− * - <0 - on error (-EFAULT)
2751− */
2752−static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2753−{
2754− if (locked) {
2755− /*
2756− * Got the lock. We might not be the anticipated owner if we
2757− * did a lock-steal - fix up the PI-state in that case:
2758− *
2759− * Speculative pi_state->owner read (we don't hold wait_lock);
2760− * since we own the lock pi_state->owner == current is the
2761− * stable state, anything else needs more attention.
2762− */
2763− if (q->pi_state->owner != current)
2764− return fixup_pi_state_owner(uaddr, q, current);
2765− return 1;
2766− }
2767−
2768− /*
2769− * If we didn't get the lock; check if anybody stole it from us. In
2770− * that case, we need to fix up the uval to point to them instead of
2771− * us, otherwise bad things happen. [10]
2772− *
2773− * Another speculative read; pi_state->owner == current is unstable
2774− * but needs our attention.
2775− */
2776− if (q->pi_state->owner == current)
2777− return fixup_pi_state_owner(uaddr, q, NULL);
2778−
2779− /*
2780− * Paranoia check. If we did not take the lock, then we should not be
2781− * the owner of the rt_mutex. Warn and establish consistent state.
2782− */
2783− if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2784− return fixup_pi_state_owner(uaddr, q, current);
2785−
2786− return 0;
1710+ q->pi_state = NULL;
27871711 }
27881712
1713+static long futex_wait_restart(struct restart_block *restart);
1714+
27891715 /**
27901716 * futex_wait_queue() - futex_queue() and wait for wakeup, timeout, or signal
27911717 * @hb: the futex hash bucket, must be locked by the caller
@@ -2974,319 +1900,6 @@ static long futex_wait_restart(struct restart_block *restart) · 182 unmodified lines
29741900 }
29751901
29761902
2977−/*
2978− * Userspace tried a 0 -> TID atomic transition of the futex value
2979− * and failed. The kernel side here does the whole locking operation:
2980− * if there are waiters then it will block as a consequence of relying
2981− * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2982− * a 0 value of the futex too.).
2983− *
2984− * Also serves as futex trylock_pi()'ing, and due semantics.
2985− */
2986−int futex_lock_pi(u32 __user *uaddr, unsigned int flags, ktime_t *time, int trylock)
2987−{
2988− struct hrtimer_sleeper timeout, *to;
2989− struct task_struct *exiting = NULL;
2990− struct rt_mutex_waiter rt_waiter;
2991− struct futex_hash_bucket *hb;
2992− struct futex_q q = futex_q_init;
2993− int res, ret;
2994−
2995− if (!IS_ENABLED(CONFIG_FUTEX_PI))
2996− return -ENOSYS;
2997−
2998− if (refill_pi_state_cache())
2999− return -ENOMEM;
3000−
3001− to = futex_setup_timer(time, &timeout, flags, 0);
3002−
3003−retry:
3004− ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
3005− if (unlikely(ret != 0))
3006− goto out;
3007−
3008−retry_private:
3009− hb = futex_q_lock(&q);
3010−
3011− ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
3012− &exiting, 0);
3013− if (unlikely(ret)) {
3014− /*
3015− * Atomic work succeeded and we got the lock,
3016− * or failed. Either way, we do _not_ block.
3017− */
3018− switch (ret) {
3019− case 1:
3020− /* We got the lock. */
3021− ret = 0;
3022− goto out_unlock_put_key;
3023− case -EFAULT:
3024− goto uaddr_faulted;
3025− case -EBUSY:
3026− case -EAGAIN:
3027− /*
3028− * Two reasons for this:
3029− * - EBUSY: Task is exiting and we just wait for the
3030− * exit to complete.
3031− * - EAGAIN: The user space value changed.
3032− */
3033− futex_q_unlock(hb);
3034− /*
3035− * Handle the case where the owner is in the middle of
3036− * exiting. Wait for the exit to complete otherwise
3037− * this task might loop forever, aka. live lock.
3038− */
3039− wait_for_owner_exiting(ret, exiting);
3040− cond_resched();
3041− goto retry;
3042− default:
3043− goto out_unlock_put_key;
3044− }
3045− }
3046−
3047− WARN_ON(!q.pi_state);
3048−
3049− /*
3050− * Only actually queue now that the atomic ops are done:
3051− */
3052− __futex_queue(&q, hb);
3053−
3054− if (trylock) {
3055− ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
3056− /* Fixup the trylock return value: */
3057− ret = ret ? 0 : -EWOULDBLOCK;
3058− goto no_block;
3059− }
3060−
3061− rt_mutex_init_waiter(&rt_waiter);
3062−
3063− /*
3064− * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
3065− * hold it while doing rt_mutex_start_proxy(), because then it will
3066− * include hb->lock in the blocking chain, even through we'll not in
3067− * fact hold it while blocking. This will lead it to report -EDEADLK
3068− * and BUG when futex_unlock_pi() interleaves with this.
3069− *
3070− * Therefore acquire wait_lock while holding hb->lock, but drop the
3071− * latter before calling __rt_mutex_start_proxy_lock(). This
3072− * interleaves with futex_unlock_pi() -- which does a similar lock
3073− * handoff -- such that the latter can observe the futex_q::pi_state
3074− * before __rt_mutex_start_proxy_lock() is done.
3075− */
3076− raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
3077− spin_unlock(q.lock_ptr);
3078− /*
3079− * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
3080− * such that futex_unlock_pi() is guaranteed to observe the waiter when
3081− * it sees the futex_q::pi_state.
3082− */
3083− ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
3084− raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
3085−
3086− if (ret) {
3087− if (ret == 1)
3088− ret = 0;
3089− goto cleanup;
3090− }
3091−
3092− if (unlikely(to))
3093− hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
3094−
3095− ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
3096−
3097−cleanup:
3098− spin_lock(q.lock_ptr);
3099− /*
3100− * If we failed to acquire the lock (deadlock/signal/timeout), we must
3101− * first acquire the hb->lock before removing the lock from the
3102− * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3103− * lists consistent.
3104− *
3105− * In particular; it is important that futex_unlock_pi() can not
3106− * observe this inconsistency.
3107− */
3108− if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
3109− ret = 0;
3110−
3111−no_block:
3112− /*
3113− * Fixup the pi_state owner and possibly acquire the lock if we
3114− * haven't already.
3115− */
3116− res = fixup_owner(uaddr, &q, !ret);
3117− /*
3118− * If fixup_owner() returned an error, propagate that. If it acquired
3119− * the lock, clear our -ETIMEDOUT or -EINTR.
3120− */
3121− if (res)
3122− ret = (res < 0) ? res : 0;
3123−
3124− futex_unqueue_pi(&q);
3125− spin_unlock(q.lock_ptr);
3126− goto out;
3127−
3128−out_unlock_put_key:
3129− futex_q_unlock(hb);
3130−
3131−out:
3132− if (to) {
3133− hrtimer_cancel(&to->timer);
3134− destroy_hrtimer_on_stack(&to->timer);
3135− }
3136− return ret != -EINTR ? ret : -ERESTARTNOINTR;
3137−
3138−uaddr_faulted:
3139− futex_q_unlock(hb);
3140−
3141− ret = fault_in_user_writeable(uaddr);
3142− if (ret)
3143− goto out;
3144−
3145− if (!(flags & FLAGS_SHARED))
3146− goto retry_private;
3147−
3148− goto retry;
3149−}
3150−
3151−/*
3152− * Userspace attempted a TID -> 0 atomic transition, and failed.
3153− * This is the in-kernel slowpath: we look up the PI state (if any),
3154− * and do the rt-mutex unlock.
3155− */
3156−int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
3157−{
3158− u32 curval, uval, vpid = task_pid_vnr(current);
3159− union futex_key key = FUTEX_KEY_INIT;
3160− struct futex_hash_bucket *hb;
3161− struct futex_q *top_waiter;
3162− int ret;
3163−
3164− if (!IS_ENABLED(CONFIG_FUTEX_PI))
3165− return -ENOSYS;
3166−
3167−retry:
3168− if (get_user(uval, uaddr))
3169− return -EFAULT;
3170− /*
3171− * We release only a lock we actually own:
3172− */
3173− if ((uval & FUTEX_TID_MASK) != vpid)
3174− return -EPERM;
3175−
3176− ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
3177− if (ret)
3178− return ret;
3179−
3180− hb = futex_hash(&key);
3181− spin_lock(&hb->lock);
3182−
3183− /*
3184− * Check waiters first. We do not trust user space values at
3185− * all and we at least want to know if user space fiddled
3186− * with the futex value instead of blindly unlocking.
3187− */
3188− top_waiter = futex_top_waiter(hb, &key);
3189− if (top_waiter) {
3190− struct futex_pi_state *pi_state = top_waiter->pi_state;
3191−
3192− ret = -EINVAL;
3193− if (!pi_state)
3194− goto out_unlock;
3195−
3196− /*
3197− * If current does not own the pi_state then the futex is
3198− * inconsistent and user space fiddled with the futex value.
3199− */
3200− if (pi_state->owner != current)
3201− goto out_unlock;
3202−
3203− get_pi_state(pi_state);
3204− /*
3205− * By taking wait_lock while still holding hb->lock, we ensure
3206− * there is no point where we hold neither; and therefore
3207− * wake_futex_pi() must observe a state consistent with what we
3208− * observed.
3209− *
3210− * In particular; this forces __rt_mutex_start_proxy() to
3211− * complete such that we're guaranteed to observe the
3212− * rt_waiter. Also see the WARN in wake_futex_pi().
3213− */
3214− raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3215− spin_unlock(&hb->lock);
3216−
3217− /* drops pi_state->pi_mutex.wait_lock */
3218− ret = wake_futex_pi(uaddr, uval, pi_state);
3219−
3220− put_pi_state(pi_state);
3221−
3222− /*
3223− * Success, we're done! No tricky corner cases.
3224− */
3225− if (!ret)
3226− return ret;
3227− /*
3228− * The atomic access to the futex value generated a
3229− * pagefault, so retry the user-access and the wakeup:
3230− */
3231− if (ret == -EFAULT)
3232− goto pi_faulted;
3233− /*
3234− * A unconditional UNLOCK_PI op raced against a waiter
3235− * setting the FUTEX_WAITERS bit. Try again.
3236− */
3237− if (ret == -EAGAIN)
3238− goto pi_retry;
3239− /*
3240− * wake_futex_pi has detected invalid state. Tell user
3241− * space.
3242− */
3243− return ret;
3244− }
3245−
3246− /*
3247− * We have no kernel internal state, i.e. no waiters in the
3248− * kernel. Waiters which are about to queue themselves are stuck
3249− * on hb->lock. So we can safely ignore them. We do neither
3250− * preserve the WAITERS bit not the OWNER_DIED one. We are the
3251− * owner.
3252− */
3253− if ((ret = futex_cmpxchg_value_locked(&curval, uaddr, uval, 0))) {
3254− spin_unlock(&hb->lock);
3255− switch (ret) {
3256− case -EFAULT:
3257− goto pi_faulted;
3258−
3259− case -EAGAIN:
3260− goto pi_retry;
3261−
3262− default:
3263− WARN_ON_ONCE(1);
3264− return ret;
3265− }
3266− }
3267−
3268− /*
3269− * If uval has changed, let user space handle it.
3270− */
3271− ret = (curval == uval) ? 0 : -EAGAIN;
3272−
3273−out_unlock:
3274− spin_unlock(&hb->lock);
3275− return ret;
3276−
3277−pi_retry:
3278− cond_resched();
3279− goto retry;
3280−
3281−pi_faulted:
3282−
3283− ret = fault_in_user_writeable(uaddr);
3284− if (!ret)
3285− goto retry;
3286−
3287− return ret;
3288−}
3289−
32901903 /**
32911904 * handle_early_requeue_pi_wakeup() - Handle early wakeup on the initial futex
32921905 * @hb: the hash_bucket futex_q was original enqueued on
@@ -3441,7 +2054,7 @@ int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags, · 148 unmodified lines
34412054 /* The requeue acquired the lock */
34422055 if (q.pi_state && (q.pi_state->owner != current)) {
34432056 spin_lock(q.lock_ptr);
3444− ret = fixup_owner(uaddr2, &q, true);
2057+ ret = fixup_pi_owner(uaddr2, &q, true);
34452058 /*
34462059 * Drop the reference to the pi state which the
34472060 * requeue_pi() code acquired for us.
@@ -3471,9 +2084,9 @@ int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags, · 23 unmodified lines
34712084 * Fixup the pi_state owner and possibly acquire the lock if we
34722085 * haven't already.
34732086 */
3474− res = fixup_owner(uaddr2, &q, !ret);
2087+ res = fixup_pi_owner(uaddr2, &q, !ret);
34752088 /*
3476− * If fixup_owner() returned an error, propagate that. If it
2089+ * If fixup_pi_owner() returned an error, propagate that. If it
34772090 * acquired the lock, clear -ETIMEDOUT or -EINTR.
34782091 */
34792092 if (res)
@@ -3811,6 +2424,87 @@ static void compat_exit_robust_list(struct task_struct *curr) · 331 unmodified lines
38112424 }
38122425 #endif
38132426
2427+#ifdef CONFIG_FUTEX_PI
2428+
2429+/*
2430+ * This task is holding PI mutexes at exit time => bad.
2431+ * Kernel cleans up PI-state, but userspace is likely hosed.
2432+ * (Robust-futex cleanup is separate and might save the day for userspace.)
2433+ */
2434+static void exit_pi_state_list(struct task_struct *curr)
2435+{
2436+ struct list_head *next, *head = &curr->pi_state_list;
2437+ struct futex_pi_state *pi_state;
2438+ struct futex_hash_bucket *hb;
2439+ union futex_key key = FUTEX_KEY_INIT;
2440+
2441+ if (!futex_cmpxchg_enabled)
2442+ return;
2443+ /*
2444+ * We are a ZOMBIE and nobody can enqueue itself on
2445+ * pi_state_list anymore, but we have to be careful
2446+ * versus waiters unqueueing themselves:
2447+ */
2448+ raw_spin_lock_irq(&curr->pi_lock);
2449+ while (!list_empty(head)) {
2450+ next = head->next;
2451+ pi_state = list_entry(next, struct futex_pi_state, list);
2452+ key = pi_state->key;
2453+ hb = futex_hash(&key);
2454+
2455+ /*
2456+ * We can race against put_pi_state() removing itself from the
2457+ * list (a waiter going away). put_pi_state() will first
2458+ * decrement the reference count and then modify the list, so
2459+ * its possible to see the list entry but fail this reference
2460+ * acquire.
2461+ *
2462+ * In that case; drop the locks to let put_pi_state() make
2463+ * progress and retry the loop.
2464+ */
2465+ if (!refcount_inc_not_zero(&pi_state->refcount)) {
2466+ raw_spin_unlock_irq(&curr->pi_lock);
2467+ cpu_relax();
2468+ raw_spin_lock_irq(&curr->pi_lock);
2469+ continue;
2470+ }
2471+ raw_spin_unlock_irq(&curr->pi_lock);
2472+
2473+ spin_lock(&hb->lock);
2474+ raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2475+ raw_spin_lock(&curr->pi_lock);
2476+ /*
2477+ * We dropped the pi-lock, so re-check whether this
2478+ * task still owns the PI-state:
2479+ */
2480+ if (head->next != next) {
2481+ /* retain curr->pi_lock for the loop invariant */
2482+ raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
2483+ spin_unlock(&hb->lock);
2484+ put_pi_state(pi_state);
2485+ continue;
2486+ }
2487+
2488+ WARN_ON(pi_state->owner != curr);
2489+ WARN_ON(list_empty(&pi_state->list));
2490+ list_del_init(&pi_state->list);
2491+ pi_state->owner = NULL;
2492+
2493+ raw_spin_unlock(&curr->pi_lock);
2494+ raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2495+ spin_unlock(&hb->lock);
2496+
2497+ rt_mutex_futex_unlock(&pi_state->pi_mutex);
2498+ put_pi_state(pi_state);
2499+
2500+ raw_spin_lock_irq(&curr->pi_lock);
2501+ }
2502+ raw_spin_unlock_irq(&curr->pi_lock);
2503+}
2504+#else
2505+static inline void exit_pi_state_list(struct task_struct *curr) { }
2506+#endif
2507+
38142508 static void futex_cleanup(struct task_struct *tsk)
38152509 {
38162510 if (unlikely(tsk->robust_list)) {
kernel/futex/futex.h+117−0
@@ -2,6 +2,7 @@ · 1 unmodified line
22 #ifndef _FUTEX_H
33 #define _FUTEX_H
44
5+#include <linux/futex.h>
56 #include <asm/futex.h>
67
78 /*
@@ -35,6 +36,122 @@ static inline bool should_fail_futex(bool fshared) · 27 unmodified lines
3536 }
3637 #endif
3738
39+/*
40+ * Hash buckets are shared by all the futex_keys that hash to the same
41+ * location. Each key may have multiple futex_q structures, one for each task
42+ * waiting on a futex.
43+ */
44+struct futex_hash_bucket {
45+ atomic_t waiters;
46+ spinlock_t lock;
47+ struct plist_head chain;
48+} ____cacheline_aligned_in_smp;
49+
50+/*
51+ * Priority Inheritance state:
52+ */
53+struct futex_pi_state {
54+ /*
55+ * list of 'owned' pi_state instances - these have to be
56+ * cleaned up in do_exit() if the task exits prematurely:
57+ */
58+ struct list_head list;
59+
60+ /*
61+ * The PI object:
62+ */
63+ struct rt_mutex_base pi_mutex;
64+
65+ struct task_struct *owner;
66+ refcount_t refcount;
67+
68+ union futex_key key;
69+} __randomize_layout;
70+
71+/**
72+ * struct futex_q - The hashed futex queue entry, one per waiting task
73+ * @list: priority-sorted list of tasks waiting on this futex
74+ * @task: the task waiting on the futex
75+ * @lock_ptr: the hash bucket lock
76+ * @key: the key the futex is hashed on
77+ * @pi_state: optional priority inheritance state
78+ * @rt_waiter: rt_waiter storage for use with requeue_pi
79+ * @requeue_pi_key: the requeue_pi target futex key
80+ * @bitset: bitset for the optional bitmasked wakeup
81+ * @requeue_state: State field for futex_requeue_pi()
82+ * @requeue_wait: RCU wait for futex_requeue_pi() (RT only)
83+ *
84+ * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
85+ * we can wake only the relevant ones (hashed queues may be shared).
86+ *
87+ * A futex_q has a woken state, just like tasks have TASK_RUNNING.
88+ * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
89+ * The order of wakeup is always to make the first condition true, then
90+ * the second.
91+ *
92+ * PI futexes are typically woken before they are removed from the hash list via
93+ * the rt_mutex code. See futex_unqueue_pi().
94+ */
95+struct futex_q {
96+ struct plist_node list;
97+
98+ struct task_struct *task;
99+ spinlock_t *lock_ptr;
100+ union futex_key key;
101+ struct futex_pi_state *pi_state;
102+ struct rt_mutex_waiter *rt_waiter;
103+ union futex_key *requeue_pi_key;
104+ u32 bitset;
105+ atomic_t requeue_state;
106+#ifdef CONFIG_PREEMPT_RT
107+ struct rcuwait requeue_wait;
108+#endif
109+} __randomize_layout;
110+
111+extern const struct futex_q futex_q_init;
112+
113+enum futex_access {
114+ FUTEX_READ,
115+ FUTEX_WRITE
116+};
117+
118+extern int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
119+ enum futex_access rw);
120+
121+extern struct futex_hash_bucket *futex_hash(union futex_key *key);
122+
123+extern struct hrtimer_sleeper *
124+futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
125+ int flags, u64 range_ns);
126+
127+extern int fault_in_user_writeable(u32 __user *uaddr);
128+extern int futex_cmpxchg_value_locked(u32 *curval, u32 __user *uaddr, u32 uval, u32 newval);
129+extern int futex_get_value_locked(u32 *dest, u32 __user *from);
130+extern struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, union futex_key *key);
131+
132+extern void __futex_queue(struct futex_q *q, struct futex_hash_bucket *hb);
133+extern void futex_unqueue_pi(struct futex_q *q);
134+
135+extern void wait_for_owner_exiting(int ret, struct task_struct *exiting);
136+
137+extern struct futex_hash_bucket *futex_q_lock(struct futex_q *q);
138+extern void futex_q_unlock(struct futex_hash_bucket *hb);
139+
140+
141+extern int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
142+ union futex_key *key,
143+ struct futex_pi_state **ps,
144+ struct task_struct *task,
145+ struct task_struct **exiting,
146+ int set_waiters);
147+
148+extern int refill_pi_state_cache(void);
149+extern void get_pi_state(struct futex_pi_state *pi_state);
150+extern void put_pi_state(struct futex_pi_state *pi_state);
151+extern int fixup_pi_owner(u32 __user *uaddr, struct futex_q *q, int locked);
152+
153+/* syscalls */
154+
38155 extern int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags, u32
39156 val, ktime_t *abs_time, u32 bitset, u32 __user
40157 *uaddr2);
kernel/futex/pi.c+1233−0
@@ -0,0 +1,1233 @@
1+// SPDX-License-Identifier: GPL-2.0-or-later
2+
3+#include <linux/slab.h>
4+#include <linux/sched/task.h>
5+
6+#include "futex.h"
7+#include "../locking/rtmutex_common.h"
8+
9+/*
10+ * PI code:
11+ */
12+int refill_pi_state_cache(void)
13+{
14+ struct futex_pi_state *pi_state;
15+
16+ if (likely(current->pi_state_cache))
17+ return 0;
18+
19+ pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
20+
21+ if (!pi_state)
22+ return -ENOMEM;
23+
24+ INIT_LIST_HEAD(&pi_state->list);
25+ /* pi_mutex gets initialized later */
26+ pi_state->owner = NULL;
27+ refcount_set(&pi_state->refcount, 1);
28+ pi_state->key = FUTEX_KEY_INIT;
29+
30+ current->pi_state_cache = pi_state;
31+
32+ return 0;
33+}
34+
35+static struct futex_pi_state *alloc_pi_state(void)
36+{
37+ struct futex_pi_state *pi_state = current->pi_state_cache;
38+
39+ WARN_ON(!pi_state);
40+ current->pi_state_cache = NULL;
41+
42+ return pi_state;
43+}
44+
45+static void pi_state_update_owner(struct futex_pi_state *pi_state,
46+ struct task_struct *new_owner)
47+{
48+ struct task_struct *old_owner = pi_state->owner;
49+
50+ lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
51+
52+ if (old_owner) {
53+ raw_spin_lock(&old_owner->pi_lock);
54+ WARN_ON(list_empty(&pi_state->list));
55+ list_del_init(&pi_state->list);
56+ raw_spin_unlock(&old_owner->pi_lock);
57+ }
58+
59+ if (new_owner) {
60+ raw_spin_lock(&new_owner->pi_lock);
61+ WARN_ON(!list_empty(&pi_state->list));
62+ list_add(&pi_state->list, &new_owner->pi_state_list);
63+ pi_state->owner = new_owner;
64+ raw_spin_unlock(&new_owner->pi_lock);
65+ }
66+}
67+
68+void get_pi_state(struct futex_pi_state *pi_state)
69+{
70+ WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
71+}
72+
73+/*
74+ * Drops a reference to the pi_state object and frees or caches it
75+ * when the last reference is gone.
76+ */
77+void put_pi_state(struct futex_pi_state *pi_state)
78+{
79+ if (!pi_state)
80+ return;
81+
82+ if (!refcount_dec_and_test(&pi_state->refcount))
83+ return;
84+
85+ /*
86+ * If pi_state->owner is NULL, the owner is most probably dying
87+ * and has cleaned up the pi_state already
88+ */
89+ if (pi_state->owner) {
90+ unsigned long flags;
91+
92+ raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
93+ pi_state_update_owner(pi_state, NULL);
94+ rt_mutex_proxy_unlock(&pi_state->pi_mutex);
95+ raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
96+ }
97+
98+ if (current->pi_state_cache) {
99+ kfree(pi_state);
100+ } else {
101+ /*
102+ * pi_state->list is already empty.
103+ * clear pi_state->owner.
104+ * refcount is at 0 - put it back to 1.
105+ */
106+ pi_state->owner = NULL;
107+ refcount_set(&pi_state->refcount, 1);
108+ current->pi_state_cache = pi_state;
109+ }
110+}
111+
112+/*
113+ * We need to check the following states:
114+ *
115+ * Waiter | pi_state | pi->owner | uTID | uODIED | ?
116+ *
117+ * [1] NULL | --- | --- | 0 | 0/1 | Valid
118+ * [2] NULL | --- | --- | >0 | 0/1 | Valid
119+ *
120+ * [3] Found | NULL | -- | Any | 0/1 | Invalid
121+ *
122+ * [4] Found | Found | NULL | 0 | 1 | Valid
123+ * [5] Found | Found | NULL | >0 | 1 | Invalid
124+ *
125+ * [6] Found | Found | task | 0 | 1 | Valid
126+ *
127+ * [7] Found | Found | NULL | Any | 0 | Invalid
128+ *
129+ * [8] Found | Found | task | ==taskTID | 0/1 | Valid
130+ * [9] Found | Found | task | 0 | 0 | Invalid
131+ * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
132+ *
133+ * [1] Indicates that the kernel can acquire the futex atomically. We
134+ * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
135+ *
136+ * [2] Valid, if TID does not belong to a kernel thread. If no matching
137+ * thread is found then it indicates that the owner TID has died.
138+ *
139+ * [3] Invalid. The waiter is queued on a non PI futex
140+ *
141+ * [4] Valid state after exit_robust_list(), which sets the user space
142+ * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
143+ *
144+ * [5] The user space value got manipulated between exit_robust_list()
145+ * and exit_pi_state_list()
146+ *
147+ * [6] Valid state after exit_pi_state_list() which sets the new owner in
148+ * the pi_state but cannot access the user space value.
149+ *
150+ * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
151+ *
152+ * [8] Owner and user space value match
153+ *
154+ * [9] There is no transient state which sets the user space TID to 0
155+ * except exit_robust_list(), but this is indicated by the
156+ * FUTEX_OWNER_DIED bit. See [4]
157+ *
158+ * [10] There is no transient state which leaves owner and user space
159+ * TID out of sync. Except one error case where the kernel is denied
160+ * write access to the user address, see fixup_pi_state_owner().
161+ *
162+ *
163+ * Serialization and lifetime rules:
164+ *
165+ * hb->lock:
166+ *
167+ * hb -> futex_q, relation
168+ * futex_q -> pi_state, relation
169+ *
170+ * (cannot be raw because hb can contain arbitrary amount
171+ * of futex_q's)
172+ *
173+ * pi_mutex->wait_lock:
174+ *
175+ * {uval, pi_state}
176+ *
177+ * (and pi_mutex 'obviously')
178+ *
179+ * p->pi_lock:
180+ *
181+ * p->pi_state_list -> pi_state->list, relation
182+ * pi_mutex->owner -> pi_state->owner, relation
183+ *
184+ * pi_state->refcount:
185+ *
186+ * pi_state lifetime
187+ *
188+ *
189+ * Lock order:
190+ *
191+ * hb->lock
192+ * pi_mutex->wait_lock
193+ * p->pi_lock
194+ *
195+ */
196+
197+/*
198+ * Validate that the existing waiter has a pi_state and sanity check
199+ * the pi_state against the user space value. If correct, attach to
200+ * it.
201+ */
202+static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
203+ struct futex_pi_state *pi_state,
204+ struct futex_pi_state **ps)
205+{
206+ pid_t pid = uval & FUTEX_TID_MASK;
207+ u32 uval2;
208+ int ret;
209+
210+ /*
211+ * Userspace might have messed up non-PI and PI futexes [3]
212+ */
213+ if (unlikely(!pi_state))
214+ return -EINVAL;
215+
216+ /*
217+ * We get here with hb->lock held, and having found a
218+ * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
219+ * has dropped the hb->lock in between futex_queue() and futex_unqueue_pi(),
220+ * which in turn means that futex_lock_pi() still has a reference on
221+ * our pi_state.
222+ *
223+ * The waiter holding a reference on @pi_state also protects against
224+ * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
225+ * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
226+ * free pi_state before we can take a reference ourselves.
227+ */
228+ WARN_ON(!refcount_read(&pi_state->refcount));
229+
230+ /*
231+ * Now that we have a pi_state, we can acquire wait_lock
232+ * and do the state validation.
233+ */
234+ raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
235+
236+ /*
237+ * Since {uval, pi_state} is serialized by wait_lock, and our current
238+ * uval was read without holding it, it can have changed. Verify it
239+ * still is what we expect it to be, otherwise retry the entire
240+ * operation.
241+ */
242+ if (futex_get_value_locked(&uval2, uaddr))
243+ goto out_efault;
244+
245+ if (uval != uval2)
246+ goto out_eagain;
247+
248+ /*
249+ * Handle the owner died case:
250+ */
251+ if (uval & FUTEX_OWNER_DIED) {
252+ /*
253+ * exit_pi_state_list sets owner to NULL and wakes the
254+ * topmost waiter. The task which acquires the
255+ * pi_state->rt_mutex will fixup owner.
256+ */
257+ if (!pi_state->owner) {
258+ /*
259+ * No pi state owner, but the user space TID
260+ * is not 0. Inconsistent state. [5]
261+ */
262+ if (pid)
263+ goto out_einval;
264+ /*
265+ * Take a ref on the state and return success. [4]
266+ */
267+ goto out_attach;
268+ }
269+
270+ /*
271+ * If TID is 0, then either the dying owner has not
272+ * yet executed exit_pi_state_list() or some waiter
273+ * acquired the rtmutex in the pi state, but did not
274+ * yet fixup the TID in user space.
275+ *
276+ * Take a ref on the state and return success. [6]
277+ */
278+ if (!pid)
279+ goto out_attach;
280+ } else {
281+ /*
282+ * If the owner died bit is not set, then the pi_state
283+ * must have an owner. [7]
284+ */
285+ if (!pi_state->owner)
286+ goto out_einval;
287+ }
288+
289+ /*
290+ * Bail out if user space manipulated the futex value. If pi
291+ * state exists then the owner TID must be the same as the
292+ * user space TID. [9/10]
293+ */
294+ if (pid != task_pid_vnr(pi_state->owner))
295+ goto out_einval;
296+
297+out_attach:
298+ get_pi_state(pi_state);
299+ raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
300+ *ps = pi_state;
301+ return 0;
302+
303+out_einval:
304+ ret = -EINVAL;
305+ goto out_error;
306+
307+out_eagain:
308+ ret = -EAGAIN;
309+ goto out_error;
310+
311+out_efault:
312+ ret = -EFAULT;
313+ goto out_error;
314+
315+out_error:
316+ raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
317+ return ret;
318+}
319+
320+static int handle_exit_race(u32 __user *uaddr, u32 uval,
321+ struct task_struct *tsk)
322+{
323+ u32 uval2;
324+
325+ /*
326+ * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
327+ * caller that the alleged owner is busy.
328+ */
329+ if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
330+ return -EBUSY;
331+
332+ /*
333+ * Reread the user space value to handle the following situation:
334+ *
335+ * CPU0 CPU1
336+ *
337+ * sys_exit() sys_futex()
338+ * do_exit() futex_lock_pi()
339+ * futex_lock_pi_atomic()
340+ * exit_signals(tsk) No waiters:
341+ * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
342+ * mm_release(tsk) Set waiter bit
343+ * exit_robust_list(tsk) { *uaddr = 0x80000PID;
344+ * Set owner died attach_to_pi_owner() {
345+ * *uaddr = 0xC0000000; tsk = get_task(PID);
346+ * } if (!tsk->flags & PF_EXITING) {
347+ * ... attach();
348+ * tsk->futex_state = } else {
349+ * FUTEX_STATE_DEAD; if (tsk->futex_state !=
350+ * FUTEX_STATE_DEAD)
351+ * return -EAGAIN;
352+ * return -ESRCH; <--- FAIL
353+ * }
354+ *
355+ * Returning ESRCH unconditionally is wrong here because the
356+ * user space value has been changed by the exiting task.
357+ *
358+ * The same logic applies to the case where the exiting task is
359+ * already gone.
360+ */
361+ if (futex_get_value_locked(&uval2, uaddr))
362+ return -EFAULT;
363+
364+ /* If the user space value has changed, try again. */
365+ if (uval2 != uval)
366+ return -EAGAIN;
367+
368+ /*
369+ * The exiting task did not have a robust list, the robust list was
370+ * corrupted or the user space value in *uaddr is simply bogus.
371+ * Give up and tell user space.
372+ */
373+ return -ESRCH;
374+}
375+
376+static void __attach_to_pi_owner(struct task_struct *p, union futex_key *key,
377+ struct futex_pi_state **ps)
378+{
379+ /*
380+ * No existing pi state. First waiter. [2]
381+ *
382+ * This creates pi_state, we have hb->lock held, this means nothing can
383+ * observe this state, wait_lock is irrelevant.
384+ */
385+ struct futex_pi_state *pi_state = alloc_pi_state();
386+
387+ /*
388+ * Initialize the pi_mutex in locked state and make @p
389+ * the owner of it:
390+ */
391+ rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
392+
393+ /* Store the key for possible exit cleanups: */
394+ pi_state->key = *key;
395+
396+ WARN_ON(!list_empty(&pi_state->list));
397+ list_add(&pi_state->list, &p->pi_state_list);
398+ /*
399+ * Assignment without holding pi_state->pi_mutex.wait_lock is safe
400+ * because there is no concurrency as the object is not published yet.
401+ */
402+ pi_state->owner = p;
403+
404+ *ps = pi_state;
405+}
406+/*
407+ * Lookup the task for the TID provided from user space and attach to
408+ * it after doing proper sanity checks.
409+ */
410+static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
411+ struct futex_pi_state **ps,
412+ struct task_struct **exiting)
413+{
414+ pid_t pid = uval & FUTEX_TID_MASK;
415+ struct task_struct *p;
416+
417+ /*
418+ * We are the first waiter - try to look up the real owner and attach
419+ * the new pi_state to it, but bail out when TID = 0 [1]
420+ *
421+ * The !pid check is paranoid. None of the call sites should end up
422+ * with pid == 0, but better safe than sorry. Let the caller retry
423+ */
424+ if (!pid)
425+ return -EAGAIN;
426+ p = find_get_task_by_vpid(pid);
427+ if (!p)
428+ return handle_exit_race(uaddr, uval, NULL);
429+
430+ if (unlikely(p->flags & PF_KTHREAD)) {
431+ put_task_struct(p);
432+ return -EPERM;
433+ }
434+
435+ /*
436+ * We need to look at the task state to figure out, whether the
437+ * task is exiting. To protect against the change of the task state
438+ * in futex_exit_release(), we do this protected by p->pi_lock:
439+ */
440+ raw_spin_lock_irq(&p->pi_lock);
441+ if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
442+ /*
443+ * The task is on the way out. When the futex state is
444+ * FUTEX_STATE_DEAD, we know that the task has finished
445+ * the cleanup:
446+ */
447+ int ret = handle_exit_race(uaddr, uval, p);
448+
449+ raw_spin_unlock_irq(&p->pi_lock);
450+ /*
451+ * If the owner task is between FUTEX_STATE_EXITING and
452+ * FUTEX_STATE_DEAD then store the task pointer and keep
453+ * the reference on the task struct. The calling code will
454+ * drop all locks, wait for the task to reach
455+ * FUTEX_STATE_DEAD and then drop the refcount. This is
456+ * required to prevent a live lock when the current task
457+ * preempted the exiting task between the two states.
458+ */
459+ if (ret == -EBUSY)
460+ *exiting = p;
461+ else
462+ put_task_struct(p);
463+ return ret;
464+ }
465+
466+ __attach_to_pi_owner(p, key, ps);
467+ raw_spin_unlock_irq(&p->pi_lock);
468+
469+ put_task_struct(p);
470+
471+ return 0;
472+}
473+
474+static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
475+{
476+ int err;
477+ u32 curval;
478+
479+ if (unlikely(should_fail_futex(true)))
480+ return -EFAULT;
481+
482+ err = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
483+ if (unlikely(err))
484+ return err;
485+
486+ /* If user space value changed, let the caller retry */
487+ return curval != uval ? -EAGAIN : 0;
488+}
489+
490+/**
491+ * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
492+ * @uaddr: the pi futex user address
493+ * @hb: the pi futex hash bucket
494+ * @key: the futex key associated with uaddr and hb
495+ * @ps: the pi_state pointer where we store the result of the
496+ * lookup
497+ * @task: the task to perform the atomic lock work for. This will
498+ * be "current" except in the case of requeue pi.
499+ * @exiting: Pointer to store the task pointer of the owner task
500+ * which is in the middle of exiting
501+ * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
502+ *
503+ * Return:
504+ * - 0 - ready to wait;
505+ * - 1 - acquired the lock;
506+ * - <0 - error
507+ *
508+ * The hb->lock must be held by the caller.
509+ *
510+ * @exiting is only set when the return value is -EBUSY. If so, this holds
511+ * a refcount on the exiting task on return and the caller needs to drop it
512+ * after waiting for the exit to complete.
513+ */
514+int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
515+ union futex_key *key,
516+ struct futex_pi_state **ps,
517+ struct task_struct *task,
518+ struct task_struct **exiting,
519+ int set_waiters)
520+{
521+ u32 uval, newval, vpid = task_pid_vnr(task);
522+ struct futex_q *top_waiter;
523+ int ret;
524+
525+ /*
526+ * Read the user space value first so we can validate a few
527+ * things before proceeding further.
528+ */
529+ if (futex_get_value_locked(&uval, uaddr))
530+ return -EFAULT;
531+
532+ if (unlikely(should_fail_futex(true)))
533+ return -EFAULT;
534+
535+ /*
536+ * Detect deadlocks.
537+ */
538+ if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
539+ return -EDEADLK;
540+
541+ if ((unlikely(should_fail_futex(true))))
542+ return -EDEADLK;
543+
544+ /*
545+ * Lookup existing state first. If it exists, try to attach to
546+ * its pi_state.
547+ */
548+ top_waiter = futex_top_waiter(hb, key);
549+ if (top_waiter)
550+ return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
551+
552+ /*
553+ * No waiter and user TID is 0. We are here because the
554+ * waiters or the owner died bit is set or called from
555+ * requeue_cmp_pi or for whatever reason something took the
556+ * syscall.
557+ */
558+ if (!(uval & FUTEX_TID_MASK)) {
559+ /*
560+ * We take over the futex. No other waiters and the user space
561+ * TID is 0. We preserve the owner died bit.
562+ */
563+ newval = uval & FUTEX_OWNER_DIED;
564+ newval |= vpid;
565+
566+ /* The futex requeue_pi code can enforce the waiters bit */
567+ if (set_waiters)
568+ newval |= FUTEX_WAITERS;
569+
570+ ret = lock_pi_update_atomic(uaddr, uval, newval);
571+ if (ret)
572+ return ret;
573+
574+ /*
575+ * If the waiter bit was requested the caller also needs PI
576+ * state attached to the new owner of the user space futex.
577+ *
578+ * @task is guaranteed to be alive and it cannot be exiting
579+ * because it is either sleeping or waiting in
580+ * futex_requeue_pi_wakeup_sync().
581+ *
582+ * No need to do the full attach_to_pi_owner() exercise
583+ * because @task is known and valid.
584+ */
585+ if (set_waiters) {
586+ raw_spin_lock_irq(&task->pi_lock);
587+ __attach_to_pi_owner(task, key, ps);
588+ raw_spin_unlock_irq(&task->pi_lock);
589+ }
590+ return 1;
591+ }
592+
593+ /*
594+ * First waiter. Set the waiters bit before attaching ourself to
595+ * the owner. If owner tries to unlock, it will be forced into
596+ * the kernel and blocked on hb->lock.
597+ */
598+ newval = uval | FUTEX_WAITERS;
599+ ret = lock_pi_update_atomic(uaddr, uval, newval);
600+ if (ret)
601+ return ret;
602+ /*
603+ * If the update of the user space value succeeded, we try to
604+ * attach to the owner. If that fails, no harm done, we only
605+ * set the FUTEX_WAITERS bit in the user space variable.
606+ */
607+ return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
608+}
609+
610+/*
611+ * Caller must hold a reference on @pi_state.
612+ */
613+static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
614+{
615+ struct rt_mutex_waiter *top_waiter;
616+ struct task_struct *new_owner;
617+ bool postunlock = false;
618+ DEFINE_RT_WAKE_Q(wqh);
619+ u32 curval, newval;
620+ int ret = 0;
621+
622+ top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
623+ if (WARN_ON_ONCE(!top_waiter)) {
624+ /*
625+ * As per the comment in futex_unlock_pi() this should not happen.
626+ *
627+ * When this happens, give up our locks and try again, giving
628+ * the futex_lock_pi() instance time to complete, either by
629+ * waiting on the rtmutex or removing itself from the futex
630+ * queue.
631+ */
632+ ret = -EAGAIN;
633+ goto out_unlock;
634+ }
635+
636+ new_owner = top_waiter->task;
637+
638+ /*
639+ * We pass it to the next owner. The WAITERS bit is always kept
640+ * enabled while there is PI state around. We cleanup the owner
641+ * died bit, because we are the owner.
642+ */
643+ newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
644+
645+ if (unlikely(should_fail_futex(true))) {
646+ ret = -EFAULT;
647+ goto out_unlock;
648+ }
649+
650+ ret = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
651+ if (!ret && (curval != uval)) {
652+ /*
653+ * If a unconditional UNLOCK_PI operation (user space did not
654+ * try the TID->0 transition) raced with a waiter setting the
655+ * FUTEX_WAITERS flag between get_user() and locking the hash
656+ * bucket lock, retry the operation.
657+ */
658+ if ((FUTEX_TID_MASK & curval) == uval)
659+ ret = -EAGAIN;
660+ else
661+ ret = -EINVAL;
662+ }
663+
664+ if (!ret) {
665+ /*
666+ * This is a point of no return; once we modified the uval
667+ * there is no going back and subsequent operations must
668+ * not fail.
669+ */
670+ pi_state_update_owner(pi_state, new_owner);
671+ postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wqh);
672+ }
673+
674+out_unlock:
675+ raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
676+
677+ if (postunlock)
678+ rt_mutex_postunlock(&wqh);
679+
680+ return ret;
681+}
682+
683+static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
684+ struct task_struct *argowner)
685+{
686+ struct futex_pi_state *pi_state = q->pi_state;
687+ struct task_struct *oldowner, *newowner;
688+ u32 uval, curval, newval, newtid;
689+ int err = 0;
690+
691+ oldowner = pi_state->owner;
692+
693+ /*
694+ * We are here because either:
695+ *
696+ * - we stole the lock and pi_state->owner needs updating to reflect
697+ * that (@argowner == current),
698+ *
699+ * or:
700+ *
701+ * - someone stole our lock and we need to fix things to point to the
702+ * new owner (@argowner == NULL).
703+ *
704+ * Either way, we have to replace the TID in the user space variable.
705+ * This must be atomic as we have to preserve the owner died bit here.
706+ *
707+ * Note: We write the user space value _before_ changing the pi_state
708+ * because we can fault here. Imagine swapped out pages or a fork
709+ * that marked all the anonymous memory readonly for cow.
710+ *
711+ * Modifying pi_state _before_ the user space value would leave the
712+ * pi_state in an inconsistent state when we fault here, because we
713+ * need to drop the locks to handle the fault. This might be observed
714+ * in the PID checks when attaching to PI state .
715+ */
716+retry:
717+ if (!argowner) {
718+ if (oldowner != current) {
719+ /*
720+ * We raced against a concurrent self; things are
721+ * already fixed up. Nothing to do.
722+ */
723+ return 0;
724+ }
725+
726+ if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
727+ /* We got the lock. pi_state is correct. Tell caller. */
728+ return 1;
729+ }
730+
731+ /*
732+ * The trylock just failed, so either there is an owner or
733+ * there is a higher priority waiter than this one.
734+ */
735+ newowner = rt_mutex_owner(&pi_state->pi_mutex);
736+ /*
737+ * If the higher priority waiter has not yet taken over the
738+ * rtmutex then newowner is NULL. We can't return here with
739+ * that state because it's inconsistent vs. the user space
740+ * state. So drop the locks and try again. It's a valid
741+ * situation and not any different from the other retry
742+ * conditions.
743+ */
744+ if (unlikely(!newowner)) {
745+ err = -EAGAIN;
746+ goto handle_err;
747+ }
748+ } else {
749+ WARN_ON_ONCE(argowner != current);
750+ if (oldowner == current) {
751+ /*
752+ * We raced against a concurrent self; things are
753+ * already fixed up. Nothing to do.
754+ */
755+ return 1;
756+ }
757+ newowner = argowner;
758+ }
759+
760+ newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
761+ /* Owner died? */
762+ if (!pi_state->owner)
763+ newtid |= FUTEX_OWNER_DIED;
764+
765+ err = futex_get_value_locked(&uval, uaddr);
766+ if (err)
767+ goto handle_err;
768+
769+ for (;;) {
770+ newval = (uval & FUTEX_OWNER_DIED) | newtid;
771+
772+ err = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
773+ if (err)
774+ goto handle_err;
775+
776+ if (curval == uval)
777+ break;
778+ uval = curval;
779+ }
780+
781+ /*
782+ * We fixed up user space. Now we need to fix the pi_state
783+ * itself.
784+ */
785+ pi_state_update_owner(pi_state, newowner);
786+
787+ return argowner == current;
788+
789+ /*
790+ * In order to reschedule or handle a page fault, we need to drop the
791+ * locks here. In the case of a fault, this gives the other task
792+ * (either the highest priority waiter itself or the task which stole
793+ * the rtmutex) the chance to try the fixup of the pi_state. So once we
794+ * are back from handling the fault we need to check the pi_state after
795+ * reacquiring the locks and before trying to do another fixup. When
796+ * the fixup has been done already we simply return.
797+ *
798+ * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
799+ * drop hb->lock since the caller owns the hb -> futex_q relation.
800+ * Dropping the pi_mutex->wait_lock requires the state revalidate.
801+ */
802+handle_err:
803+ raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
804+ spin_unlock(q->lock_ptr);
805+
806+ switch (err) {
807+ case -EFAULT:
808+ err = fault_in_user_writeable(uaddr);
809+ break;
810+
811+ case -EAGAIN:
812+ cond_resched();
813+ err = 0;
814+ break;
815+
816+ default:
817+ WARN_ON_ONCE(1);
818+ break;
819+ }
820+
821+ spin_lock(q->lock_ptr);
822+ raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
823+
824+ /*
825+ * Check if someone else fixed it for us:
826+ */
827+ if (pi_state->owner != oldowner)
828+ return argowner == current;
829+
830+ /* Retry if err was -EAGAIN or the fault in succeeded */
831+ if (!err)
832+ goto retry;
833+
834+ /*
835+ * fault_in_user_writeable() failed so user state is immutable. At
836+ * best we can make the kernel state consistent but user state will
837+ * be most likely hosed and any subsequent unlock operation will be
838+ * rejected due to PI futex rule [10].
839+ *
840+ * Ensure that the rtmutex owner is also the pi_state owner despite
841+ * the user space value claiming something different. There is no
842+ * point in unlocking the rtmutex if current is the owner as it
843+ * would need to wait until the next waiter has taken the rtmutex
844+ * to guarantee consistent state. Keep it simple. Userspace asked
845+ * for this wreckaged state.
846+ *
847+ * The rtmutex has an owner - either current or some other
848+ * task. See the EAGAIN loop above.
849+ */
850+ pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
851+
852+ return err;
853+}
854+
855+static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
856+ struct task_struct *argowner)
857+{
858+ struct futex_pi_state *pi_state = q->pi_state;
859+ int ret;
860+
861+ lockdep_assert_held(q->lock_ptr);
862+
863+ raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
864+ ret = __fixup_pi_state_owner(uaddr, q, argowner);
865+ raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
866+ return ret;
867+}
868+
869+/**
870+ * fixup_pi_owner() - Post lock pi_state and corner case management
871+ * @uaddr: user address of the futex
872+ * @q: futex_q (contains pi_state and access to the rt_mutex)
873+ * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
874+ *
875+ * After attempting to lock an rt_mutex, this function is called to cleanup
876+ * the pi_state owner as well as handle race conditions that may allow us to
877+ * acquire the lock. Must be called with the hb lock held.
878+ *
879+ * Return:
880+ * - 1 - success, lock taken;
881+ * - 0 - success, lock not taken;
882+ * - <0 - on error (-EFAULT)
883+ */
884+int fixup_pi_owner(u32 __user *uaddr, struct futex_q *q, int locked)
885+{
886+ if (locked) {
887+ /*
888+ * Got the lock. We might not be the anticipated owner if we
889+ * did a lock-steal - fix up the PI-state in that case:
890+ *
891+ * Speculative pi_state->owner read (we don't hold wait_lock);
892+ * since we own the lock pi_state->owner == current is the
893+ * stable state, anything else needs more attention.
894+ */
895+ if (q->pi_state->owner != current)
896+ return fixup_pi_state_owner(uaddr, q, current);
897+ return 1;
898+ }
899+
900+ /*
901+ * If we didn't get the lock; check if anybody stole it from us. In
902+ * that case, we need to fix up the uval to point to them instead of
903+ * us, otherwise bad things happen. [10]
904+ *
905+ * Another speculative read; pi_state->owner == current is unstable
906+ * but needs our attention.
907+ */
908+ if (q->pi_state->owner == current)
909+ return fixup_pi_state_owner(uaddr, q, NULL);
910+
911+ /*
912+ * Paranoia check. If we did not take the lock, then we should not be
913+ * the owner of the rt_mutex. Warn and establish consistent state.
914+ */
915+ if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
916+ return fixup_pi_state_owner(uaddr, q, current);
917+
918+ return 0;
919+}
920+
921+/*
922+ * Userspace tried a 0 -> TID atomic transition of the futex value
923+ * and failed. The kernel side here does the whole locking operation:
924+ * if there are waiters then it will block as a consequence of relying
925+ * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
926+ * a 0 value of the futex too.).
927+ *
928+ * Also serves as futex trylock_pi()'ing, and due semantics.
929+ */
930+int futex_lock_pi(u32 __user *uaddr, unsigned int flags, ktime_t *time, int trylock)
931+{
932+ struct hrtimer_sleeper timeout, *to;
933+ struct task_struct *exiting = NULL;
934+ struct rt_mutex_waiter rt_waiter;
935+ struct futex_hash_bucket *hb;
936+ struct futex_q q = futex_q_init;
937+ int res, ret;
938+
939+ if (!IS_ENABLED(CONFIG_FUTEX_PI))
940+ return -ENOSYS;
941+
942+ if (refill_pi_state_cache())
943+ return -ENOMEM;
944+
945+ to = futex_setup_timer(time, &timeout, flags, 0);
946+
947+retry:
948+ ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
949+ if (unlikely(ret != 0))
950+ goto out;
951+
952+retry_private:
953+ hb = futex_q_lock(&q);
954+
955+ ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
956+ &exiting, 0);
957+ if (unlikely(ret)) {
958+ /*
959+ * Atomic work succeeded and we got the lock,
960+ * or failed. Either way, we do _not_ block.
961+ */
962+ switch (ret) {
963+ case 1:
964+ /* We got the lock. */
965+ ret = 0;
966+ goto out_unlock_put_key;
967+ case -EFAULT:
968+ goto uaddr_faulted;
969+ case -EBUSY:
970+ case -EAGAIN:
971+ /*
972+ * Two reasons for this:
973+ * - EBUSY: Task is exiting and we just wait for the
974+ * exit to complete.
975+ * - EAGAIN: The user space value changed.
976+ */
977+ futex_q_unlock(hb);
978+ /*
979+ * Handle the case where the owner is in the middle of
980+ * exiting. Wait for the exit to complete otherwise
981+ * this task might loop forever, aka. live lock.
982+ */
983+ wait_for_owner_exiting(ret, exiting);
984+ cond_resched();
985+ goto retry;
986+ default:
987+ goto out_unlock_put_key;
988+ }
989+ }
990+
991+ WARN_ON(!q.pi_state);
992+
993+ /*
994+ * Only actually queue now that the atomic ops are done:
995+ */
996+ __futex_queue(&q, hb);
997+
998+ if (trylock) {
999+ ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
1000+ /* Fixup the trylock return value: */
1001+ ret = ret ? 0 : -EWOULDBLOCK;
1002+ goto no_block;
1003+ }
1004+
1005+ rt_mutex_init_waiter(&rt_waiter);
1006+
1007+ /*
1008+ * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
1009+ * hold it while doing rt_mutex_start_proxy(), because then it will
1010+ * include hb->lock in the blocking chain, even through we'll not in
1011+ * fact hold it while blocking. This will lead it to report -EDEADLK
1012+ * and BUG when futex_unlock_pi() interleaves with this.
1013+ *
1014+ * Therefore acquire wait_lock while holding hb->lock, but drop the
1015+ * latter before calling __rt_mutex_start_proxy_lock(). This
1016+ * interleaves with futex_unlock_pi() -- which does a similar lock
1017+ * handoff -- such that the latter can observe the futex_q::pi_state
1018+ * before __rt_mutex_start_proxy_lock() is done.
1019+ */
1020+ raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
1021+ spin_unlock(q.lock_ptr);
1022+ /*
1023+ * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
1024+ * such that futex_unlock_pi() is guaranteed to observe the waiter when
1025+ * it sees the futex_q::pi_state.
1026+ */
1027+ ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
1028+ raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
1029+
1030+ if (ret) {
1031+ if (ret == 1)
1032+ ret = 0;
1033+ goto cleanup;
1034+ }
1035+
1036+ if (unlikely(to))
1037+ hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
1038+
1039+ ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
1040+
1041+cleanup:
1042+ spin_lock(q.lock_ptr);
1043+ /*
1044+ * If we failed to acquire the lock (deadlock/signal/timeout), we must
1045+ * first acquire the hb->lock before removing the lock from the
1046+ * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
1047+ * lists consistent.
1048+ *
1049+ * In particular; it is important that futex_unlock_pi() can not
1050+ * observe this inconsistency.
1051+ */
1052+ if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
1053+ ret = 0;
1054+
1055+no_block:
1056+ /*
1057+ * Fixup the pi_state owner and possibly acquire the lock if we
1058+ * haven't already.
1059+ */
1060+ res = fixup_pi_owner(uaddr, &q, !ret);
1061+ /*
1062+ * If fixup_pi_owner() returned an error, propagate that. If it acquired
1063+ * the lock, clear our -ETIMEDOUT or -EINTR.
1064+ */
1065+ if (res)
1066+ ret = (res < 0) ? res : 0;
1067+
1068+ futex_unqueue_pi(&q);
1069+ spin_unlock(q.lock_ptr);
1070+ goto out;
1071+
1072+out_unlock_put_key:
1073+ futex_q_unlock(hb);
1074+
1075+out:
1076+ if (to) {
1077+ hrtimer_cancel(&to->timer);
1078+ destroy_hrtimer_on_stack(&to->timer);
1079+ }
1080+ return ret != -EINTR ? ret : -ERESTARTNOINTR;
1081+
1082+uaddr_faulted:
1083+ futex_q_unlock(hb);
1084+
1085+ ret = fault_in_user_writeable(uaddr);
1086+ if (ret)
1087+ goto out;
1088+
1089+ if (!(flags & FLAGS_SHARED))
1090+ goto retry_private;
1091+
1092+ goto retry;
1093+}
1094+
1095+/*
1096+ * Userspace attempted a TID -> 0 atomic transition, and failed.
1097+ * This is the in-kernel slowpath: we look up the PI state (if any),
1098+ * and do the rt-mutex unlock.
1099+ */
1100+int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
1101+{
1102+ u32 curval, uval, vpid = task_pid_vnr(current);
1103+ union futex_key key = FUTEX_KEY_INIT;
1104+ struct futex_hash_bucket *hb;
1105+ struct futex_q *top_waiter;
1106+ int ret;
1107+
1108+ if (!IS_ENABLED(CONFIG_FUTEX_PI))
1109+ return -ENOSYS;
1110+
1111+retry:
1112+ if (get_user(uval, uaddr))
1113+ return -EFAULT;
1114+ /*
1115+ * We release only a lock we actually own:
1116+ */
1117+ if ((uval & FUTEX_TID_MASK) != vpid)
1118+ return -EPERM;
1119+
1120+ ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
1121+ if (ret)
1122+ return ret;
1123+
1124+ hb = futex_hash(&key);
1125+ spin_lock(&hb->lock);
1126+
1127+ /*
1128+ * Check waiters first. We do not trust user space values at
1129+ * all and we at least want to know if user space fiddled
1130+ * with the futex value instead of blindly unlocking.
1131+ */
1132+ top_waiter = futex_top_waiter(hb, &key);
1133+ if (top_waiter) {
1134+ struct futex_pi_state *pi_state = top_waiter->pi_state;
1135+
1136+ ret = -EINVAL;
1137+ if (!pi_state)
1138+ goto out_unlock;
1139+
1140+ /*
1141+ * If current does not own the pi_state then the futex is
1142+ * inconsistent and user space fiddled with the futex value.
1143+ */
1144+ if (pi_state->owner != current)
1145+ goto out_unlock;
1146+
1147+ get_pi_state(pi_state);
1148+ /*
1149+ * By taking wait_lock while still holding hb->lock, we ensure
1150+ * there is no point where we hold neither; and therefore
1151+ * wake_futex_p() must observe a state consistent with what we
1152+ * observed.
1153+ *
1154+ * In particular; this forces __rt_mutex_start_proxy() to
1155+ * complete such that we're guaranteed to observe the
1156+ * rt_waiter. Also see the WARN in wake_futex_pi().
1157+ */
1158+ raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1159+ spin_unlock(&hb->lock);
1160+
1161+ /* drops pi_state->pi_mutex.wait_lock */
1162+ ret = wake_futex_pi(uaddr, uval, pi_state);
1163+
1164+ put_pi_state(pi_state);
1165+
1166+ /*
1167+ * Success, we're done! No tricky corner cases.
1168+ */
1169+ if (!ret)
1170+ return ret;
1171+ /*
1172+ * The atomic access to the futex value generated a
1173+ * pagefault, so retry the user-access and the wakeup:
1174+ */
1175+ if (ret == -EFAULT)
1176+ goto pi_faulted;
1177+ /*
1178+ * A unconditional UNLOCK_PI op raced against a waiter
1179+ * setting the FUTEX_WAITERS bit. Try again.
1180+ */
1181+ if (ret == -EAGAIN)
1182+ goto pi_retry;
1183+ /*
1184+ * wake_futex_pi has detected invalid state. Tell user
1185+ * space.
1186+ */
1187+ return ret;
1188+ }
1189+
1190+ /*
1191+ * We have no kernel internal state, i.e. no waiters in the
1192+ * kernel. Waiters which are about to queue themselves are stuck
1193+ * on hb->lock. So we can safely ignore them. We do neither
1194+ * preserve the WAITERS bit not the OWNER_DIED one. We are the
1195+ * owner.
1196+ */
1197+ if ((ret = futex_cmpxchg_value_locked(&curval, uaddr, uval, 0))) {
1198+ spin_unlock(&hb->lock);
1199+ switch (ret) {
1200+ case -EFAULT:
1201+ goto pi_faulted;
1202+
1203+ case -EAGAIN:
1204+ goto pi_retry;
1205+
1206+ default:
1207+ WARN_ON_ONCE(1);
1208+ return ret;
1209+ }
1210+ }
1211+
1212+ /*
1213+ * If uval has changed, let user space handle it.
1214+ */
1215+ ret = (curval == uval) ? 0 : -EAGAIN;
1216+
1217+out_unlock:
1218+ spin_unlock(&hb->lock);
1219+ return ret;
1220+
1221+pi_retry:
1222+ cond_resched();
1223+ goto retry;
1224+
1225+pi_faulted:
1226+
1227+ ret = fault_in_user_writeable(uaddr);
1228+ if (!ret)
1229+ goto retry;
1230+
1231+ return ret;
1232+}
1233+