xref: /linux/kernel/bpf/rqspinlock.c (revision b676ac484f847bbe5c7d29603f41475b64fefe55)
1a8fcf2a3SKumar Kartikeya Dwivedi // SPDX-License-Identifier: GPL-2.0-or-later
2a8fcf2a3SKumar Kartikeya Dwivedi /*
3a8fcf2a3SKumar Kartikeya Dwivedi  * Resilient Queued Spin Lock
4a8fcf2a3SKumar Kartikeya Dwivedi  *
5a8fcf2a3SKumar Kartikeya Dwivedi  * (C) Copyright 2013-2015 Hewlett-Packard Development Company, L.P.
6a8fcf2a3SKumar Kartikeya Dwivedi  * (C) Copyright 2013-2014,2018 Red Hat, Inc.
7a8fcf2a3SKumar Kartikeya Dwivedi  * (C) Copyright 2015 Intel Corp.
8a8fcf2a3SKumar Kartikeya Dwivedi  * (C) Copyright 2015 Hewlett-Packard Enterprise Development LP
914c48ee8SKumar Kartikeya Dwivedi  * (C) Copyright 2024-2025 Meta Platforms, Inc. and affiliates.
10a8fcf2a3SKumar Kartikeya Dwivedi  *
11a8fcf2a3SKumar Kartikeya Dwivedi  * Authors: Waiman Long <longman@redhat.com>
12a8fcf2a3SKumar Kartikeya Dwivedi  *          Peter Zijlstra <peterz@infradead.org>
1314c48ee8SKumar Kartikeya Dwivedi  *          Kumar Kartikeya Dwivedi <memxor@gmail.com>
14a8fcf2a3SKumar Kartikeya Dwivedi  */
15a8fcf2a3SKumar Kartikeya Dwivedi 
16a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/smp.h>
17a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/bug.h>
1897eb35f3SKumar Kartikeya Dwivedi #include <linux/bpf.h>
1997eb35f3SKumar Kartikeya Dwivedi #include <linux/err.h>
20a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/cpumask.h>
21a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/percpu.h>
22a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/hardirq.h>
23a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/mutex.h>
24a8fcf2a3SKumar Kartikeya Dwivedi #include <linux/prefetch.h>
25a8fcf2a3SKumar Kartikeya Dwivedi #include <asm/byteorder.h>
26c9102a68SKumar Kartikeya Dwivedi #ifdef CONFIG_QUEUED_SPINLOCKS
27a8fcf2a3SKumar Kartikeya Dwivedi #include <asm/qspinlock.h>
28c9102a68SKumar Kartikeya Dwivedi #endif
29a8fcf2a3SKumar Kartikeya Dwivedi #include <trace/events/lock.h>
3030ff1332SKumar Kartikeya Dwivedi #include <asm/rqspinlock.h>
3114c48ee8SKumar Kartikeya Dwivedi #include <linux/timekeeping.h>
32a8fcf2a3SKumar Kartikeya Dwivedi 
33a8fcf2a3SKumar Kartikeya Dwivedi /*
34a8fcf2a3SKumar Kartikeya Dwivedi  * Include queued spinlock definitions and statistics code
35a8fcf2a3SKumar Kartikeya Dwivedi  */
36c9102a68SKumar Kartikeya Dwivedi #ifdef CONFIG_QUEUED_SPINLOCKS
37a8fcf2a3SKumar Kartikeya Dwivedi #include "../locking/qspinlock.h"
38a926d099SKumar Kartikeya Dwivedi #include "../locking/lock_events.h"
3931158ad0SKumar Kartikeya Dwivedi #include "rqspinlock.h"
40c9102a68SKumar Kartikeya Dwivedi #include "../locking/mcs_spinlock.h"
41c9102a68SKumar Kartikeya Dwivedi #endif
42a8fcf2a3SKumar Kartikeya Dwivedi 
43a8fcf2a3SKumar Kartikeya Dwivedi /*
44a8fcf2a3SKumar Kartikeya Dwivedi  * The basic principle of a queue-based spinlock can best be understood
45a8fcf2a3SKumar Kartikeya Dwivedi  * by studying a classic queue-based spinlock implementation called the
46a8fcf2a3SKumar Kartikeya Dwivedi  * MCS lock. A copy of the original MCS lock paper ("Algorithms for Scalable
47a8fcf2a3SKumar Kartikeya Dwivedi  * Synchronization on Shared-Memory Multiprocessors by Mellor-Crummey and
48a8fcf2a3SKumar Kartikeya Dwivedi  * Scott") is available at
49a8fcf2a3SKumar Kartikeya Dwivedi  *
50a8fcf2a3SKumar Kartikeya Dwivedi  * https://bugzilla.kernel.org/show_bug.cgi?id=206115
51a8fcf2a3SKumar Kartikeya Dwivedi  *
52a8fcf2a3SKumar Kartikeya Dwivedi  * This queued spinlock implementation is based on the MCS lock, however to
53a8fcf2a3SKumar Kartikeya Dwivedi  * make it fit the 4 bytes we assume spinlock_t to be, and preserve its
54a8fcf2a3SKumar Kartikeya Dwivedi  * existing API, we must modify it somehow.
55a8fcf2a3SKumar Kartikeya Dwivedi  *
56a8fcf2a3SKumar Kartikeya Dwivedi  * In particular; where the traditional MCS lock consists of a tail pointer
57a8fcf2a3SKumar Kartikeya Dwivedi  * (8 bytes) and needs the next pointer (another 8 bytes) of its own node to
58a8fcf2a3SKumar Kartikeya Dwivedi  * unlock the next pending (next->locked), we compress both these: {tail,
59a8fcf2a3SKumar Kartikeya Dwivedi  * next->locked} into a single u32 value.
60a8fcf2a3SKumar Kartikeya Dwivedi  *
61a8fcf2a3SKumar Kartikeya Dwivedi  * Since a spinlock disables recursion of its own context and there is a limit
62a8fcf2a3SKumar Kartikeya Dwivedi  * to the contexts that can nest; namely: task, softirq, hardirq, nmi. As there
63a8fcf2a3SKumar Kartikeya Dwivedi  * are at most 4 nesting levels, it can be encoded by a 2-bit number. Now
64a8fcf2a3SKumar Kartikeya Dwivedi  * we can encode the tail by combining the 2-bit nesting level with the cpu
65a8fcf2a3SKumar Kartikeya Dwivedi  * number. With one byte for the lock value and 3 bytes for the tail, only a
66a8fcf2a3SKumar Kartikeya Dwivedi  * 32-bit word is now needed. Even though we only need 1 bit for the lock,
67a8fcf2a3SKumar Kartikeya Dwivedi  * we extend it to a full byte to achieve better performance for architectures
68a8fcf2a3SKumar Kartikeya Dwivedi  * that support atomic byte write.
69a8fcf2a3SKumar Kartikeya Dwivedi  *
70a8fcf2a3SKumar Kartikeya Dwivedi  * We also change the first spinner to spin on the lock bit instead of its
71a8fcf2a3SKumar Kartikeya Dwivedi  * node; whereby avoiding the need to carry a node from lock to unlock, and
72a8fcf2a3SKumar Kartikeya Dwivedi  * preserving existing lock API. This also makes the unlock code simpler and
73a8fcf2a3SKumar Kartikeya Dwivedi  * faster.
74a8fcf2a3SKumar Kartikeya Dwivedi  *
75a8fcf2a3SKumar Kartikeya Dwivedi  * N.B. The current implementation only supports architectures that allow
76a8fcf2a3SKumar Kartikeya Dwivedi  *      atomic operations on smaller 8-bit and 16-bit data types.
77a8fcf2a3SKumar Kartikeya Dwivedi  *
78a8fcf2a3SKumar Kartikeya Dwivedi  */
79a8fcf2a3SKumar Kartikeya Dwivedi 
8014c48ee8SKumar Kartikeya Dwivedi struct rqspinlock_timeout {
8114c48ee8SKumar Kartikeya Dwivedi 	u64 timeout_end;
8214c48ee8SKumar Kartikeya Dwivedi 	u64 duration;
8331158ad0SKumar Kartikeya Dwivedi 	u64 cur;
8414c48ee8SKumar Kartikeya Dwivedi 	u16 spin;
8514c48ee8SKumar Kartikeya Dwivedi };
8614c48ee8SKumar Kartikeya Dwivedi 
87164c2465SKumar Kartikeya Dwivedi #define RES_TIMEOUT_VAL	2
88164c2465SKumar Kartikeya Dwivedi 
8931158ad0SKumar Kartikeya Dwivedi DEFINE_PER_CPU_ALIGNED(struct rqspinlock_held, rqspinlock_held_locks);
9031158ad0SKumar Kartikeya Dwivedi EXPORT_SYMBOL_GPL(rqspinlock_held_locks);
9131158ad0SKumar Kartikeya Dwivedi 
is_lock_released(rqspinlock_t * lock,u32 mask,struct rqspinlock_timeout * ts)9231158ad0SKumar Kartikeya Dwivedi static bool is_lock_released(rqspinlock_t *lock, u32 mask, struct rqspinlock_timeout *ts)
9331158ad0SKumar Kartikeya Dwivedi {
9431158ad0SKumar Kartikeya Dwivedi 	if (!(atomic_read_acquire(&lock->val) & (mask)))
9531158ad0SKumar Kartikeya Dwivedi 		return true;
9631158ad0SKumar Kartikeya Dwivedi 	return false;
9731158ad0SKumar Kartikeya Dwivedi }
9831158ad0SKumar Kartikeya Dwivedi 
check_deadlock_AA(rqspinlock_t * lock,u32 mask,struct rqspinlock_timeout * ts)9931158ad0SKumar Kartikeya Dwivedi static noinline int check_deadlock_AA(rqspinlock_t *lock, u32 mask,
10031158ad0SKumar Kartikeya Dwivedi 				      struct rqspinlock_timeout *ts)
10131158ad0SKumar Kartikeya Dwivedi {
10231158ad0SKumar Kartikeya Dwivedi 	struct rqspinlock_held *rqh = this_cpu_ptr(&rqspinlock_held_locks);
10331158ad0SKumar Kartikeya Dwivedi 	int cnt = min(RES_NR_HELD, rqh->cnt);
10431158ad0SKumar Kartikeya Dwivedi 
10531158ad0SKumar Kartikeya Dwivedi 	/*
10631158ad0SKumar Kartikeya Dwivedi 	 * Return an error if we hold the lock we are attempting to acquire.
10731158ad0SKumar Kartikeya Dwivedi 	 * We'll iterate over max 32 locks; no need to do is_lock_released.
10831158ad0SKumar Kartikeya Dwivedi 	 */
10931158ad0SKumar Kartikeya Dwivedi 	for (int i = 0; i < cnt - 1; i++) {
11031158ad0SKumar Kartikeya Dwivedi 		if (rqh->locks[i] == lock)
11131158ad0SKumar Kartikeya Dwivedi 			return -EDEADLK;
11231158ad0SKumar Kartikeya Dwivedi 	}
11331158ad0SKumar Kartikeya Dwivedi 	return 0;
11431158ad0SKumar Kartikeya Dwivedi }
11531158ad0SKumar Kartikeya Dwivedi 
11631158ad0SKumar Kartikeya Dwivedi /*
11731158ad0SKumar Kartikeya Dwivedi  * This focuses on the most common case of ABBA deadlocks (or ABBA involving
11831158ad0SKumar Kartikeya Dwivedi  * more locks, which reduce to ABBA). This is not exhaustive, and we rely on
11931158ad0SKumar Kartikeya Dwivedi  * timeouts as the final line of defense.
12031158ad0SKumar Kartikeya Dwivedi  */
check_deadlock_ABBA(rqspinlock_t * lock,u32 mask,struct rqspinlock_timeout * ts)12131158ad0SKumar Kartikeya Dwivedi static noinline int check_deadlock_ABBA(rqspinlock_t *lock, u32 mask,
12231158ad0SKumar Kartikeya Dwivedi 					struct rqspinlock_timeout *ts)
12331158ad0SKumar Kartikeya Dwivedi {
12431158ad0SKumar Kartikeya Dwivedi 	struct rqspinlock_held *rqh = this_cpu_ptr(&rqspinlock_held_locks);
12531158ad0SKumar Kartikeya Dwivedi 	int rqh_cnt = min(RES_NR_HELD, rqh->cnt);
12631158ad0SKumar Kartikeya Dwivedi 	void *remote_lock;
12731158ad0SKumar Kartikeya Dwivedi 	int cpu;
12831158ad0SKumar Kartikeya Dwivedi 
12931158ad0SKumar Kartikeya Dwivedi 	/*
13031158ad0SKumar Kartikeya Dwivedi 	 * Find the CPU holding the lock that we want to acquire. If there is a
13131158ad0SKumar Kartikeya Dwivedi 	 * deadlock scenario, we will read a stable set on the remote CPU and
13231158ad0SKumar Kartikeya Dwivedi 	 * find the target. This would be a constant time operation instead of
13331158ad0SKumar Kartikeya Dwivedi 	 * O(NR_CPUS) if we could determine the owning CPU from a lock value, but
13431158ad0SKumar Kartikeya Dwivedi 	 * that requires increasing the size of the lock word.
13531158ad0SKumar Kartikeya Dwivedi 	 */
13631158ad0SKumar Kartikeya Dwivedi 	for_each_possible_cpu(cpu) {
13731158ad0SKumar Kartikeya Dwivedi 		struct rqspinlock_held *rqh_cpu = per_cpu_ptr(&rqspinlock_held_locks, cpu);
13831158ad0SKumar Kartikeya Dwivedi 		int real_cnt = READ_ONCE(rqh_cpu->cnt);
13931158ad0SKumar Kartikeya Dwivedi 		int cnt = min(RES_NR_HELD, real_cnt);
14031158ad0SKumar Kartikeya Dwivedi 
14131158ad0SKumar Kartikeya Dwivedi 		/*
14231158ad0SKumar Kartikeya Dwivedi 		 * Let's ensure to break out of this loop if the lock is available for
14331158ad0SKumar Kartikeya Dwivedi 		 * us to potentially acquire.
14431158ad0SKumar Kartikeya Dwivedi 		 */
14531158ad0SKumar Kartikeya Dwivedi 		if (is_lock_released(lock, mask, ts))
14631158ad0SKumar Kartikeya Dwivedi 			return 0;
14731158ad0SKumar Kartikeya Dwivedi 
14831158ad0SKumar Kartikeya Dwivedi 		/*
14931158ad0SKumar Kartikeya Dwivedi 		 * Skip ourselves, and CPUs whose count is less than 2, as they need at
15031158ad0SKumar Kartikeya Dwivedi 		 * least one held lock and one acquisition attempt (reflected as top
15131158ad0SKumar Kartikeya Dwivedi 		 * most entry) to participate in an ABBA deadlock.
15231158ad0SKumar Kartikeya Dwivedi 		 *
15331158ad0SKumar Kartikeya Dwivedi 		 * If cnt is more than RES_NR_HELD, it means the current lock being
15431158ad0SKumar Kartikeya Dwivedi 		 * acquired won't appear in the table, and other locks in the table are
15531158ad0SKumar Kartikeya Dwivedi 		 * already held, so we can't determine ABBA.
15631158ad0SKumar Kartikeya Dwivedi 		 */
15731158ad0SKumar Kartikeya Dwivedi 		if (cpu == smp_processor_id() || real_cnt < 2 || real_cnt > RES_NR_HELD)
15831158ad0SKumar Kartikeya Dwivedi 			continue;
15931158ad0SKumar Kartikeya Dwivedi 
16031158ad0SKumar Kartikeya Dwivedi 		/*
16131158ad0SKumar Kartikeya Dwivedi 		 * Obtain the entry at the top, this corresponds to the lock the
16231158ad0SKumar Kartikeya Dwivedi 		 * remote CPU is attempting to acquire in a deadlock situation,
16331158ad0SKumar Kartikeya Dwivedi 		 * and would be one of the locks we hold on the current CPU.
16431158ad0SKumar Kartikeya Dwivedi 		 */
16531158ad0SKumar Kartikeya Dwivedi 		remote_lock = READ_ONCE(rqh_cpu->locks[cnt - 1]);
16631158ad0SKumar Kartikeya Dwivedi 		/*
16731158ad0SKumar Kartikeya Dwivedi 		 * If it is NULL, we've raced and cannot determine a deadlock
16831158ad0SKumar Kartikeya Dwivedi 		 * conclusively, skip this CPU.
16931158ad0SKumar Kartikeya Dwivedi 		 */
17031158ad0SKumar Kartikeya Dwivedi 		if (!remote_lock)
17131158ad0SKumar Kartikeya Dwivedi 			continue;
17231158ad0SKumar Kartikeya Dwivedi 		/*
17331158ad0SKumar Kartikeya Dwivedi 		 * Find if the lock we're attempting to acquire is held by this CPU.
17431158ad0SKumar Kartikeya Dwivedi 		 * Don't consider the topmost entry, as that must be the latest lock
17531158ad0SKumar Kartikeya Dwivedi 		 * being held or acquired.  For a deadlock, the target CPU must also
17631158ad0SKumar Kartikeya Dwivedi 		 * attempt to acquire a lock we hold, so for this search only 'cnt - 1'
17731158ad0SKumar Kartikeya Dwivedi 		 * entries are important.
17831158ad0SKumar Kartikeya Dwivedi 		 */
17931158ad0SKumar Kartikeya Dwivedi 		for (int i = 0; i < cnt - 1; i++) {
18031158ad0SKumar Kartikeya Dwivedi 			if (READ_ONCE(rqh_cpu->locks[i]) != lock)
18131158ad0SKumar Kartikeya Dwivedi 				continue;
18231158ad0SKumar Kartikeya Dwivedi 			/*
18331158ad0SKumar Kartikeya Dwivedi 			 * We found our lock as held on the remote CPU.  Is the
18431158ad0SKumar Kartikeya Dwivedi 			 * acquisition attempt on the remote CPU for a lock held
18531158ad0SKumar Kartikeya Dwivedi 			 * by us?  If so, we have a deadlock situation, and need
18631158ad0SKumar Kartikeya Dwivedi 			 * to recover.
18731158ad0SKumar Kartikeya Dwivedi 			 */
18831158ad0SKumar Kartikeya Dwivedi 			for (int i = 0; i < rqh_cnt - 1; i++) {
18931158ad0SKumar Kartikeya Dwivedi 				if (rqh->locks[i] == remote_lock)
19031158ad0SKumar Kartikeya Dwivedi 					return -EDEADLK;
19131158ad0SKumar Kartikeya Dwivedi 			}
19231158ad0SKumar Kartikeya Dwivedi 			/*
19331158ad0SKumar Kartikeya Dwivedi 			 * Inconclusive; retry again later.
19431158ad0SKumar Kartikeya Dwivedi 			 */
19531158ad0SKumar Kartikeya Dwivedi 			return 0;
19631158ad0SKumar Kartikeya Dwivedi 		}
19731158ad0SKumar Kartikeya Dwivedi 	}
19831158ad0SKumar Kartikeya Dwivedi 	return 0;
19931158ad0SKumar Kartikeya Dwivedi }
20031158ad0SKumar Kartikeya Dwivedi 
check_deadlock(rqspinlock_t * lock,u32 mask,struct rqspinlock_timeout * ts)20131158ad0SKumar Kartikeya Dwivedi static noinline int check_deadlock(rqspinlock_t *lock, u32 mask,
20231158ad0SKumar Kartikeya Dwivedi 				   struct rqspinlock_timeout *ts)
20331158ad0SKumar Kartikeya Dwivedi {
20431158ad0SKumar Kartikeya Dwivedi 	int ret;
20531158ad0SKumar Kartikeya Dwivedi 
20631158ad0SKumar Kartikeya Dwivedi 	ret = check_deadlock_AA(lock, mask, ts);
20731158ad0SKumar Kartikeya Dwivedi 	if (ret)
20831158ad0SKumar Kartikeya Dwivedi 		return ret;
20931158ad0SKumar Kartikeya Dwivedi 	ret = check_deadlock_ABBA(lock, mask, ts);
21031158ad0SKumar Kartikeya Dwivedi 	if (ret)
21131158ad0SKumar Kartikeya Dwivedi 		return ret;
21231158ad0SKumar Kartikeya Dwivedi 
21331158ad0SKumar Kartikeya Dwivedi 	return 0;
21431158ad0SKumar Kartikeya Dwivedi }
21531158ad0SKumar Kartikeya Dwivedi 
check_timeout(rqspinlock_t * lock,u32 mask,struct rqspinlock_timeout * ts)21631158ad0SKumar Kartikeya Dwivedi static noinline int check_timeout(rqspinlock_t *lock, u32 mask,
21731158ad0SKumar Kartikeya Dwivedi 				  struct rqspinlock_timeout *ts)
21814c48ee8SKumar Kartikeya Dwivedi {
21914c48ee8SKumar Kartikeya Dwivedi 	u64 time = ktime_get_mono_fast_ns();
22031158ad0SKumar Kartikeya Dwivedi 	u64 prev = ts->cur;
22114c48ee8SKumar Kartikeya Dwivedi 
22214c48ee8SKumar Kartikeya Dwivedi 	if (!ts->timeout_end) {
22331158ad0SKumar Kartikeya Dwivedi 		ts->cur = time;
22414c48ee8SKumar Kartikeya Dwivedi 		ts->timeout_end = time + ts->duration;
22514c48ee8SKumar Kartikeya Dwivedi 		return 0;
22614c48ee8SKumar Kartikeya Dwivedi 	}
22714c48ee8SKumar Kartikeya Dwivedi 
22814c48ee8SKumar Kartikeya Dwivedi 	if (time > ts->timeout_end)
22914c48ee8SKumar Kartikeya Dwivedi 		return -ETIMEDOUT;
23014c48ee8SKumar Kartikeya Dwivedi 
23131158ad0SKumar Kartikeya Dwivedi 	/*
23231158ad0SKumar Kartikeya Dwivedi 	 * A millisecond interval passed from last time? Trigger deadlock
23331158ad0SKumar Kartikeya Dwivedi 	 * checks.
23431158ad0SKumar Kartikeya Dwivedi 	 */
23531158ad0SKumar Kartikeya Dwivedi 	if (prev + NSEC_PER_MSEC < time) {
23631158ad0SKumar Kartikeya Dwivedi 		ts->cur = time;
23731158ad0SKumar Kartikeya Dwivedi 		return check_deadlock(lock, mask, ts);
23831158ad0SKumar Kartikeya Dwivedi 	}
23931158ad0SKumar Kartikeya Dwivedi 
24014c48ee8SKumar Kartikeya Dwivedi 	return 0;
24114c48ee8SKumar Kartikeya Dwivedi }
24214c48ee8SKumar Kartikeya Dwivedi 
243ebababcdSKumar Kartikeya Dwivedi /*
244ebababcdSKumar Kartikeya Dwivedi  * Do not amortize with spins when res_smp_cond_load_acquire is defined,
245ebababcdSKumar Kartikeya Dwivedi  * as the macro does internal amortization for us.
246ebababcdSKumar Kartikeya Dwivedi  */
247ebababcdSKumar Kartikeya Dwivedi #ifndef res_smp_cond_load_acquire
24831158ad0SKumar Kartikeya Dwivedi #define RES_CHECK_TIMEOUT(ts, ret, mask)                              \
24914c48ee8SKumar Kartikeya Dwivedi 	({                                                            \
25014c48ee8SKumar Kartikeya Dwivedi 		if (!(ts).spin++)                                     \
25131158ad0SKumar Kartikeya Dwivedi 			(ret) = check_timeout((lock), (mask), &(ts)); \
25214c48ee8SKumar Kartikeya Dwivedi 		(ret);                                                \
25314c48ee8SKumar Kartikeya Dwivedi 	})
254ebababcdSKumar Kartikeya Dwivedi #else
255ebababcdSKumar Kartikeya Dwivedi #define RES_CHECK_TIMEOUT(ts, ret, mask)			      \
256*92b90f78SKumar Kartikeya Dwivedi 	({ (ret) = check_timeout((lock), (mask), &(ts)); })
257ebababcdSKumar Kartikeya Dwivedi #endif
25814c48ee8SKumar Kartikeya Dwivedi 
25914c48ee8SKumar Kartikeya Dwivedi /*
26014c48ee8SKumar Kartikeya Dwivedi  * Initialize the 'spin' member.
26131158ad0SKumar Kartikeya Dwivedi  * Set spin member to 0 to trigger AA/ABBA checks immediately.
26214c48ee8SKumar Kartikeya Dwivedi  */
26331158ad0SKumar Kartikeya Dwivedi #define RES_INIT_TIMEOUT(ts) ({ (ts).spin = 0; })
26414c48ee8SKumar Kartikeya Dwivedi 
26514c48ee8SKumar Kartikeya Dwivedi /*
26614c48ee8SKumar Kartikeya Dwivedi  * We only need to reset 'timeout_end', 'spin' will just wrap around as necessary.
26714c48ee8SKumar Kartikeya Dwivedi  * Duration is defined for each spin attempt, so set it here.
26814c48ee8SKumar Kartikeya Dwivedi  */
26914c48ee8SKumar Kartikeya Dwivedi #define RES_RESET_TIMEOUT(ts, _duration) ({ (ts).timeout_end = 0; (ts).duration = _duration; })
27014c48ee8SKumar Kartikeya Dwivedi 
271a8fcf2a3SKumar Kartikeya Dwivedi /*
272c9102a68SKumar Kartikeya Dwivedi  * Provide a test-and-set fallback for cases when queued spin lock support is
273c9102a68SKumar Kartikeya Dwivedi  * absent from the architecture.
274c9102a68SKumar Kartikeya Dwivedi  */
resilient_tas_spin_lock(rqspinlock_t * lock)275c9102a68SKumar Kartikeya Dwivedi int __lockfunc resilient_tas_spin_lock(rqspinlock_t *lock)
276c9102a68SKumar Kartikeya Dwivedi {
277c9102a68SKumar Kartikeya Dwivedi 	struct rqspinlock_timeout ts;
278c9102a68SKumar Kartikeya Dwivedi 	int val, ret = 0;
279c9102a68SKumar Kartikeya Dwivedi 
280c9102a68SKumar Kartikeya Dwivedi 	RES_INIT_TIMEOUT(ts);
281c9102a68SKumar Kartikeya Dwivedi 	grab_held_lock_entry(lock);
282c9102a68SKumar Kartikeya Dwivedi 
283c9102a68SKumar Kartikeya Dwivedi 	/*
284c9102a68SKumar Kartikeya Dwivedi 	 * Since the waiting loop's time is dependent on the amount of
285c9102a68SKumar Kartikeya Dwivedi 	 * contention, a short timeout unlike rqspinlock waiting loops
286c9102a68SKumar Kartikeya Dwivedi 	 * isn't enough. Choose a second as the timeout value.
287c9102a68SKumar Kartikeya Dwivedi 	 */
288c9102a68SKumar Kartikeya Dwivedi 	RES_RESET_TIMEOUT(ts, NSEC_PER_SEC);
289c9102a68SKumar Kartikeya Dwivedi retry:
290c9102a68SKumar Kartikeya Dwivedi 	val = atomic_read(&lock->val);
291c9102a68SKumar Kartikeya Dwivedi 
292c9102a68SKumar Kartikeya Dwivedi 	if (val || !atomic_try_cmpxchg(&lock->val, &val, 1)) {
293c9102a68SKumar Kartikeya Dwivedi 		if (RES_CHECK_TIMEOUT(ts, ret, ~0u))
294c9102a68SKumar Kartikeya Dwivedi 			goto out;
295c9102a68SKumar Kartikeya Dwivedi 		cpu_relax();
296c9102a68SKumar Kartikeya Dwivedi 		goto retry;
297c9102a68SKumar Kartikeya Dwivedi 	}
298c9102a68SKumar Kartikeya Dwivedi 
299c9102a68SKumar Kartikeya Dwivedi 	return 0;
300c9102a68SKumar Kartikeya Dwivedi out:
301c9102a68SKumar Kartikeya Dwivedi 	release_held_lock_entry();
302c9102a68SKumar Kartikeya Dwivedi 	return ret;
303c9102a68SKumar Kartikeya Dwivedi }
304c9102a68SKumar Kartikeya Dwivedi EXPORT_SYMBOL_GPL(resilient_tas_spin_lock);
305c9102a68SKumar Kartikeya Dwivedi 
306c9102a68SKumar Kartikeya Dwivedi #ifdef CONFIG_QUEUED_SPINLOCKS
307c9102a68SKumar Kartikeya Dwivedi 
308c9102a68SKumar Kartikeya Dwivedi /*
309a8fcf2a3SKumar Kartikeya Dwivedi  * Per-CPU queue node structures; we can never have more than 4 nested
310a8fcf2a3SKumar Kartikeya Dwivedi  * contexts: task, softirq, hardirq, nmi.
311a8fcf2a3SKumar Kartikeya Dwivedi  *
312a8fcf2a3SKumar Kartikeya Dwivedi  * Exactly fits one 64-byte cacheline on a 64-bit architecture.
313a8fcf2a3SKumar Kartikeya Dwivedi  */
314a8fcf2a3SKumar Kartikeya Dwivedi static DEFINE_PER_CPU_ALIGNED(struct qnode, rqnodes[_Q_MAX_NODES]);
315a8fcf2a3SKumar Kartikeya Dwivedi 
316ebababcdSKumar Kartikeya Dwivedi #ifndef res_smp_cond_load_acquire
317ebababcdSKumar Kartikeya Dwivedi #define res_smp_cond_load_acquire(v, c) smp_cond_load_acquire(v, c)
318ebababcdSKumar Kartikeya Dwivedi #endif
319ebababcdSKumar Kartikeya Dwivedi 
320ebababcdSKumar Kartikeya Dwivedi #define res_atomic_cond_read_acquire(v, c) res_smp_cond_load_acquire(&(v)->counter, (c))
321ebababcdSKumar Kartikeya Dwivedi 
322a8fcf2a3SKumar Kartikeya Dwivedi /**
323a8fcf2a3SKumar Kartikeya Dwivedi  * resilient_queued_spin_lock_slowpath - acquire the queued spinlock
324a8fcf2a3SKumar Kartikeya Dwivedi  * @lock: Pointer to queued spinlock structure
325a8fcf2a3SKumar Kartikeya Dwivedi  * @val: Current value of the queued spinlock 32-bit word
326a8fcf2a3SKumar Kartikeya Dwivedi  *
327337ffea5SKumar Kartikeya Dwivedi  * Return:
328337ffea5SKumar Kartikeya Dwivedi  * * 0		- Lock was acquired successfully.
32931158ad0SKumar Kartikeya Dwivedi  * * -EDEADLK	- Lock acquisition failed because of AA/ABBA deadlock.
330337ffea5SKumar Kartikeya Dwivedi  * * -ETIMEDOUT - Lock acquisition failed because of timeout.
331337ffea5SKumar Kartikeya Dwivedi  *
332a8fcf2a3SKumar Kartikeya Dwivedi  * (queue tail, pending bit, lock value)
333a8fcf2a3SKumar Kartikeya Dwivedi  *
334a8fcf2a3SKumar Kartikeya Dwivedi  *              fast     :    slow                                  :    unlock
335a8fcf2a3SKumar Kartikeya Dwivedi  *                       :                                          :
336a8fcf2a3SKumar Kartikeya Dwivedi  * uncontended  (0,0,0) -:--> (0,0,1) ------------------------------:--> (*,*,0)
337a8fcf2a3SKumar Kartikeya Dwivedi  *                       :       | ^--------.------.             /  :
338a8fcf2a3SKumar Kartikeya Dwivedi  *                       :       v           \      \            |  :
339a8fcf2a3SKumar Kartikeya Dwivedi  * pending               :    (0,1,1) +--> (0,1,0)   \           |  :
340a8fcf2a3SKumar Kartikeya Dwivedi  *                       :       | ^--'              |           |  :
341a8fcf2a3SKumar Kartikeya Dwivedi  *                       :       v                   |           |  :
342a8fcf2a3SKumar Kartikeya Dwivedi  * uncontended           :    (n,x,y) +--> (n,0,0) --'           |  :
343a8fcf2a3SKumar Kartikeya Dwivedi  *   queue               :       | ^--'                          |  :
344a8fcf2a3SKumar Kartikeya Dwivedi  *                       :       v                               |  :
345a8fcf2a3SKumar Kartikeya Dwivedi  * contended             :    (*,x,y) +--> (*,0,0) ---> (*,0,1) -'  :
346a8fcf2a3SKumar Kartikeya Dwivedi  *   queue               :         ^--'                             :
347a8fcf2a3SKumar Kartikeya Dwivedi  */
resilient_queued_spin_lock_slowpath(rqspinlock_t * lock,u32 val)348337ffea5SKumar Kartikeya Dwivedi int __lockfunc resilient_queued_spin_lock_slowpath(rqspinlock_t *lock, u32 val)
349a8fcf2a3SKumar Kartikeya Dwivedi {
350a8fcf2a3SKumar Kartikeya Dwivedi 	struct mcs_spinlock *prev, *next, *node;
35114c48ee8SKumar Kartikeya Dwivedi 	struct rqspinlock_timeout ts;
352337ffea5SKumar Kartikeya Dwivedi 	int idx, ret = 0;
353a8fcf2a3SKumar Kartikeya Dwivedi 	u32 old, tail;
354a8fcf2a3SKumar Kartikeya Dwivedi 
355a8fcf2a3SKumar Kartikeya Dwivedi 	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
356a8fcf2a3SKumar Kartikeya Dwivedi 
357ecbd8047SKumar Kartikeya Dwivedi 	if (resilient_virt_spin_lock_enabled())
358ecbd8047SKumar Kartikeya Dwivedi 		return resilient_virt_spin_lock(lock);
359ecbd8047SKumar Kartikeya Dwivedi 
36014c48ee8SKumar Kartikeya Dwivedi 	RES_INIT_TIMEOUT(ts);
36114c48ee8SKumar Kartikeya Dwivedi 
362a8fcf2a3SKumar Kartikeya Dwivedi 	/*
363a8fcf2a3SKumar Kartikeya Dwivedi 	 * Wait for in-progress pending->locked hand-overs with a bounded
364a8fcf2a3SKumar Kartikeya Dwivedi 	 * number of spins so that we guarantee forward progress.
365a8fcf2a3SKumar Kartikeya Dwivedi 	 *
366a8fcf2a3SKumar Kartikeya Dwivedi 	 * 0,1,0 -> 0,0,1
367a8fcf2a3SKumar Kartikeya Dwivedi 	 */
368a8fcf2a3SKumar Kartikeya Dwivedi 	if (val == _Q_PENDING_VAL) {
369a8fcf2a3SKumar Kartikeya Dwivedi 		int cnt = _Q_PENDING_LOOPS;
370a8fcf2a3SKumar Kartikeya Dwivedi 		val = atomic_cond_read_relaxed(&lock->val,
371a8fcf2a3SKumar Kartikeya Dwivedi 					       (VAL != _Q_PENDING_VAL) || !cnt--);
372a8fcf2a3SKumar Kartikeya Dwivedi 	}
373a8fcf2a3SKumar Kartikeya Dwivedi 
374a8fcf2a3SKumar Kartikeya Dwivedi 	/*
375a8fcf2a3SKumar Kartikeya Dwivedi 	 * If we observe any contention; queue.
376a8fcf2a3SKumar Kartikeya Dwivedi 	 */
377a8fcf2a3SKumar Kartikeya Dwivedi 	if (val & ~_Q_LOCKED_MASK)
378a8fcf2a3SKumar Kartikeya Dwivedi 		goto queue;
379a8fcf2a3SKumar Kartikeya Dwivedi 
380a8fcf2a3SKumar Kartikeya Dwivedi 	/*
381a8fcf2a3SKumar Kartikeya Dwivedi 	 * trylock || pending
382a8fcf2a3SKumar Kartikeya Dwivedi 	 *
383a8fcf2a3SKumar Kartikeya Dwivedi 	 * 0,0,* -> 0,1,* -> 0,0,1 pending, trylock
384a8fcf2a3SKumar Kartikeya Dwivedi 	 */
385a8fcf2a3SKumar Kartikeya Dwivedi 	val = queued_fetch_set_pending_acquire(lock);
386a8fcf2a3SKumar Kartikeya Dwivedi 
387a8fcf2a3SKumar Kartikeya Dwivedi 	/*
388a8fcf2a3SKumar Kartikeya Dwivedi 	 * If we observe contention, there is a concurrent locker.
389a8fcf2a3SKumar Kartikeya Dwivedi 	 *
390a8fcf2a3SKumar Kartikeya Dwivedi 	 * Undo and queue; our setting of PENDING might have made the
391a8fcf2a3SKumar Kartikeya Dwivedi 	 * n,0,0 -> 0,0,0 transition fail and it will now be waiting
392a8fcf2a3SKumar Kartikeya Dwivedi 	 * on @next to become !NULL.
393a8fcf2a3SKumar Kartikeya Dwivedi 	 */
394a8fcf2a3SKumar Kartikeya Dwivedi 	if (unlikely(val & ~_Q_LOCKED_MASK)) {
395a8fcf2a3SKumar Kartikeya Dwivedi 
396a8fcf2a3SKumar Kartikeya Dwivedi 		/* Undo PENDING if we set it. */
397a8fcf2a3SKumar Kartikeya Dwivedi 		if (!(val & _Q_PENDING_MASK))
398a8fcf2a3SKumar Kartikeya Dwivedi 			clear_pending(lock);
399a8fcf2a3SKumar Kartikeya Dwivedi 
400a8fcf2a3SKumar Kartikeya Dwivedi 		goto queue;
401a8fcf2a3SKumar Kartikeya Dwivedi 	}
402a8fcf2a3SKumar Kartikeya Dwivedi 
403a8fcf2a3SKumar Kartikeya Dwivedi 	/*
40431158ad0SKumar Kartikeya Dwivedi 	 * Grab an entry in the held locks array, to enable deadlock detection.
40531158ad0SKumar Kartikeya Dwivedi 	 */
40631158ad0SKumar Kartikeya Dwivedi 	grab_held_lock_entry(lock);
40731158ad0SKumar Kartikeya Dwivedi 
40831158ad0SKumar Kartikeya Dwivedi 	/*
409a8fcf2a3SKumar Kartikeya Dwivedi 	 * We're pending, wait for the owner to go away.
410a8fcf2a3SKumar Kartikeya Dwivedi 	 *
411a8fcf2a3SKumar Kartikeya Dwivedi 	 * 0,1,1 -> *,1,0
412a8fcf2a3SKumar Kartikeya Dwivedi 	 *
413a8fcf2a3SKumar Kartikeya Dwivedi 	 * this wait loop must be a load-acquire such that we match the
414a8fcf2a3SKumar Kartikeya Dwivedi 	 * store-release that clears the locked bit and create lock
415a8fcf2a3SKumar Kartikeya Dwivedi 	 * sequentiality; this is because not all
416a8fcf2a3SKumar Kartikeya Dwivedi 	 * clear_pending_set_locked() implementations imply full
417a8fcf2a3SKumar Kartikeya Dwivedi 	 * barriers.
418a8fcf2a3SKumar Kartikeya Dwivedi 	 */
419337ffea5SKumar Kartikeya Dwivedi 	if (val & _Q_LOCKED_MASK) {
420337ffea5SKumar Kartikeya Dwivedi 		RES_RESET_TIMEOUT(ts, RES_DEF_TIMEOUT);
42131158ad0SKumar Kartikeya Dwivedi 		res_smp_cond_load_acquire(&lock->locked, !VAL || RES_CHECK_TIMEOUT(ts, ret, _Q_LOCKED_MASK));
422337ffea5SKumar Kartikeya Dwivedi 	}
423337ffea5SKumar Kartikeya Dwivedi 
424337ffea5SKumar Kartikeya Dwivedi 	if (ret) {
425337ffea5SKumar Kartikeya Dwivedi 		/*
426337ffea5SKumar Kartikeya Dwivedi 		 * We waited for the locked bit to go back to 0, as the pending
427337ffea5SKumar Kartikeya Dwivedi 		 * waiter, but timed out. We need to clear the pending bit since
428337ffea5SKumar Kartikeya Dwivedi 		 * we own it. Once a stuck owner has been recovered, the lock
429337ffea5SKumar Kartikeya Dwivedi 		 * must be restored to a valid state, hence removing the pending
430337ffea5SKumar Kartikeya Dwivedi 		 * bit is necessary.
431337ffea5SKumar Kartikeya Dwivedi 		 *
432337ffea5SKumar Kartikeya Dwivedi 		 * *,1,* -> *,0,*
433337ffea5SKumar Kartikeya Dwivedi 		 */
434337ffea5SKumar Kartikeya Dwivedi 		clear_pending(lock);
435337ffea5SKumar Kartikeya Dwivedi 		lockevent_inc(rqspinlock_lock_timeout);
43631158ad0SKumar Kartikeya Dwivedi 		goto err_release_entry;
437337ffea5SKumar Kartikeya Dwivedi 	}
438a8fcf2a3SKumar Kartikeya Dwivedi 
439a8fcf2a3SKumar Kartikeya Dwivedi 	/*
440a8fcf2a3SKumar Kartikeya Dwivedi 	 * take ownership and clear the pending bit.
441a8fcf2a3SKumar Kartikeya Dwivedi 	 *
442a8fcf2a3SKumar Kartikeya Dwivedi 	 * 0,1,0 -> 0,0,1
443a8fcf2a3SKumar Kartikeya Dwivedi 	 */
444a8fcf2a3SKumar Kartikeya Dwivedi 	clear_pending_set_locked(lock);
445a8fcf2a3SKumar Kartikeya Dwivedi 	lockevent_inc(lock_pending);
446337ffea5SKumar Kartikeya Dwivedi 	return 0;
447a8fcf2a3SKumar Kartikeya Dwivedi 
448a8fcf2a3SKumar Kartikeya Dwivedi 	/*
449a8fcf2a3SKumar Kartikeya Dwivedi 	 * End of pending bit optimistic spinning and beginning of MCS
450a8fcf2a3SKumar Kartikeya Dwivedi 	 * queuing.
451a8fcf2a3SKumar Kartikeya Dwivedi 	 */
452a8fcf2a3SKumar Kartikeya Dwivedi queue:
453a8fcf2a3SKumar Kartikeya Dwivedi 	lockevent_inc(lock_slowpath);
45431158ad0SKumar Kartikeya Dwivedi 	/*
45531158ad0SKumar Kartikeya Dwivedi 	 * Grab deadlock detection entry for the queue path.
45631158ad0SKumar Kartikeya Dwivedi 	 */
45731158ad0SKumar Kartikeya Dwivedi 	grab_held_lock_entry(lock);
45831158ad0SKumar Kartikeya Dwivedi 
459a8fcf2a3SKumar Kartikeya Dwivedi 	node = this_cpu_ptr(&rqnodes[0].mcs);
460a8fcf2a3SKumar Kartikeya Dwivedi 	idx = node->count++;
461a8fcf2a3SKumar Kartikeya Dwivedi 	tail = encode_tail(smp_processor_id(), idx);
462a8fcf2a3SKumar Kartikeya Dwivedi 
463a8fcf2a3SKumar Kartikeya Dwivedi 	trace_contention_begin(lock, LCB_F_SPIN);
464a8fcf2a3SKumar Kartikeya Dwivedi 
465a8fcf2a3SKumar Kartikeya Dwivedi 	/*
466a8fcf2a3SKumar Kartikeya Dwivedi 	 * 4 nodes are allocated based on the assumption that there will
467a8fcf2a3SKumar Kartikeya Dwivedi 	 * not be nested NMIs taking spinlocks. That may not be true in
468a8fcf2a3SKumar Kartikeya Dwivedi 	 * some architectures even though the chance of needing more than
469a8fcf2a3SKumar Kartikeya Dwivedi 	 * 4 nodes will still be extremely unlikely. When that happens,
470a8fcf2a3SKumar Kartikeya Dwivedi 	 * we fall back to spinning on the lock directly without using
471a8fcf2a3SKumar Kartikeya Dwivedi 	 * any MCS node. This is not the most elegant solution, but is
472a8fcf2a3SKumar Kartikeya Dwivedi 	 * simple enough.
473a8fcf2a3SKumar Kartikeya Dwivedi 	 */
474a8fcf2a3SKumar Kartikeya Dwivedi 	if (unlikely(idx >= _Q_MAX_NODES)) {
475a8fcf2a3SKumar Kartikeya Dwivedi 		lockevent_inc(lock_no_node);
4763bb15936SKumar Kartikeya Dwivedi 		RES_RESET_TIMEOUT(ts, RES_DEF_TIMEOUT);
4773bb15936SKumar Kartikeya Dwivedi 		while (!queued_spin_trylock(lock)) {
47831158ad0SKumar Kartikeya Dwivedi 			if (RES_CHECK_TIMEOUT(ts, ret, ~0u)) {
4793bb15936SKumar Kartikeya Dwivedi 				lockevent_inc(rqspinlock_lock_timeout);
48031158ad0SKumar Kartikeya Dwivedi 				goto err_release_node;
4813bb15936SKumar Kartikeya Dwivedi 			}
482a8fcf2a3SKumar Kartikeya Dwivedi 			cpu_relax();
4833bb15936SKumar Kartikeya Dwivedi 		}
484a8fcf2a3SKumar Kartikeya Dwivedi 		goto release;
485a8fcf2a3SKumar Kartikeya Dwivedi 	}
486a8fcf2a3SKumar Kartikeya Dwivedi 
487a8fcf2a3SKumar Kartikeya Dwivedi 	node = grab_mcs_node(node, idx);
488a8fcf2a3SKumar Kartikeya Dwivedi 
489a8fcf2a3SKumar Kartikeya Dwivedi 	/*
490a8fcf2a3SKumar Kartikeya Dwivedi 	 * Keep counts of non-zero index values:
491a8fcf2a3SKumar Kartikeya Dwivedi 	 */
492a8fcf2a3SKumar Kartikeya Dwivedi 	lockevent_cond_inc(lock_use_node2 + idx - 1, idx);
493a8fcf2a3SKumar Kartikeya Dwivedi 
494a8fcf2a3SKumar Kartikeya Dwivedi 	/*
495a8fcf2a3SKumar Kartikeya Dwivedi 	 * Ensure that we increment the head node->count before initialising
496a8fcf2a3SKumar Kartikeya Dwivedi 	 * the actual node. If the compiler is kind enough to reorder these
497a8fcf2a3SKumar Kartikeya Dwivedi 	 * stores, then an IRQ could overwrite our assignments.
498a8fcf2a3SKumar Kartikeya Dwivedi 	 */
499a8fcf2a3SKumar Kartikeya Dwivedi 	barrier();
500a8fcf2a3SKumar Kartikeya Dwivedi 
501a8fcf2a3SKumar Kartikeya Dwivedi 	node->locked = 0;
502a8fcf2a3SKumar Kartikeya Dwivedi 	node->next = NULL;
503a8fcf2a3SKumar Kartikeya Dwivedi 
504a8fcf2a3SKumar Kartikeya Dwivedi 	/*
505a8fcf2a3SKumar Kartikeya Dwivedi 	 * We touched a (possibly) cold cacheline in the per-cpu queue node;
506a8fcf2a3SKumar Kartikeya Dwivedi 	 * attempt the trylock once more in the hope someone let go while we
507a8fcf2a3SKumar Kartikeya Dwivedi 	 * weren't watching.
508a8fcf2a3SKumar Kartikeya Dwivedi 	 */
509a8fcf2a3SKumar Kartikeya Dwivedi 	if (queued_spin_trylock(lock))
510a8fcf2a3SKumar Kartikeya Dwivedi 		goto release;
511a8fcf2a3SKumar Kartikeya Dwivedi 
512a8fcf2a3SKumar Kartikeya Dwivedi 	/*
513a8fcf2a3SKumar Kartikeya Dwivedi 	 * Ensure that the initialisation of @node is complete before we
514a8fcf2a3SKumar Kartikeya Dwivedi 	 * publish the updated tail via xchg_tail() and potentially link
515a8fcf2a3SKumar Kartikeya Dwivedi 	 * @node into the waitqueue via WRITE_ONCE(prev->next, node) below.
516a8fcf2a3SKumar Kartikeya Dwivedi 	 */
517a8fcf2a3SKumar Kartikeya Dwivedi 	smp_wmb();
518a8fcf2a3SKumar Kartikeya Dwivedi 
519a8fcf2a3SKumar Kartikeya Dwivedi 	/*
520a8fcf2a3SKumar Kartikeya Dwivedi 	 * Publish the updated tail.
521a8fcf2a3SKumar Kartikeya Dwivedi 	 * We have already touched the queueing cacheline; don't bother with
522a8fcf2a3SKumar Kartikeya Dwivedi 	 * pending stuff.
523a8fcf2a3SKumar Kartikeya Dwivedi 	 *
524a8fcf2a3SKumar Kartikeya Dwivedi 	 * p,*,* -> n,*,*
525a8fcf2a3SKumar Kartikeya Dwivedi 	 */
526a8fcf2a3SKumar Kartikeya Dwivedi 	old = xchg_tail(lock, tail);
527a8fcf2a3SKumar Kartikeya Dwivedi 	next = NULL;
528a8fcf2a3SKumar Kartikeya Dwivedi 
529a8fcf2a3SKumar Kartikeya Dwivedi 	/*
530a8fcf2a3SKumar Kartikeya Dwivedi 	 * if there was a previous node; link it and wait until reaching the
531a8fcf2a3SKumar Kartikeya Dwivedi 	 * head of the waitqueue.
532a8fcf2a3SKumar Kartikeya Dwivedi 	 */
533a8fcf2a3SKumar Kartikeya Dwivedi 	if (old & _Q_TAIL_MASK) {
534164c2465SKumar Kartikeya Dwivedi 		int val;
535164c2465SKumar Kartikeya Dwivedi 
536a8fcf2a3SKumar Kartikeya Dwivedi 		prev = decode_tail(old, rqnodes);
537a8fcf2a3SKumar Kartikeya Dwivedi 
538a8fcf2a3SKumar Kartikeya Dwivedi 		/* Link @node into the waitqueue. */
539a8fcf2a3SKumar Kartikeya Dwivedi 		WRITE_ONCE(prev->next, node);
540a8fcf2a3SKumar Kartikeya Dwivedi 
541164c2465SKumar Kartikeya Dwivedi 		val = arch_mcs_spin_lock_contended(&node->locked);
542164c2465SKumar Kartikeya Dwivedi 		if (val == RES_TIMEOUT_VAL) {
543164c2465SKumar Kartikeya Dwivedi 			ret = -EDEADLK;
544164c2465SKumar Kartikeya Dwivedi 			goto waitq_timeout;
545164c2465SKumar Kartikeya Dwivedi 		}
546a8fcf2a3SKumar Kartikeya Dwivedi 
547a8fcf2a3SKumar Kartikeya Dwivedi 		/*
548a8fcf2a3SKumar Kartikeya Dwivedi 		 * While waiting for the MCS lock, the next pointer may have
549a8fcf2a3SKumar Kartikeya Dwivedi 		 * been set by another lock waiter. We optimistically load
550a8fcf2a3SKumar Kartikeya Dwivedi 		 * the next pointer & prefetch the cacheline for writing
551a8fcf2a3SKumar Kartikeya Dwivedi 		 * to reduce latency in the upcoming MCS unlock operation.
552a8fcf2a3SKumar Kartikeya Dwivedi 		 */
553a8fcf2a3SKumar Kartikeya Dwivedi 		next = READ_ONCE(node->next);
554a8fcf2a3SKumar Kartikeya Dwivedi 		if (next)
555a8fcf2a3SKumar Kartikeya Dwivedi 			prefetchw(next);
556a8fcf2a3SKumar Kartikeya Dwivedi 	}
557a8fcf2a3SKumar Kartikeya Dwivedi 
558a8fcf2a3SKumar Kartikeya Dwivedi 	/*
559a8fcf2a3SKumar Kartikeya Dwivedi 	 * we're at the head of the waitqueue, wait for the owner & pending to
560a8fcf2a3SKumar Kartikeya Dwivedi 	 * go away.
561a8fcf2a3SKumar Kartikeya Dwivedi 	 *
562a8fcf2a3SKumar Kartikeya Dwivedi 	 * *,x,y -> *,0,0
563a8fcf2a3SKumar Kartikeya Dwivedi 	 *
564a8fcf2a3SKumar Kartikeya Dwivedi 	 * this wait loop must use a load-acquire such that we match the
565a8fcf2a3SKumar Kartikeya Dwivedi 	 * store-release that clears the locked bit and create lock
566a8fcf2a3SKumar Kartikeya Dwivedi 	 * sequentiality; this is because the set_locked() function below
567a8fcf2a3SKumar Kartikeya Dwivedi 	 * does not imply a full barrier.
568164c2465SKumar Kartikeya Dwivedi 	 *
569164c2465SKumar Kartikeya Dwivedi 	 * We use RES_DEF_TIMEOUT * 2 as the duration, as RES_DEF_TIMEOUT is
570164c2465SKumar Kartikeya Dwivedi 	 * meant to span maximum allowed time per critical section, and we may
571164c2465SKumar Kartikeya Dwivedi 	 * have both the owner of the lock and the pending bit waiter ahead of
572164c2465SKumar Kartikeya Dwivedi 	 * us.
573a8fcf2a3SKumar Kartikeya Dwivedi 	 */
574164c2465SKumar Kartikeya Dwivedi 	RES_RESET_TIMEOUT(ts, RES_DEF_TIMEOUT * 2);
575164c2465SKumar Kartikeya Dwivedi 	val = res_atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK) ||
57631158ad0SKumar Kartikeya Dwivedi 					   RES_CHECK_TIMEOUT(ts, ret, _Q_LOCKED_PENDING_MASK));
577164c2465SKumar Kartikeya Dwivedi 
578164c2465SKumar Kartikeya Dwivedi waitq_timeout:
579164c2465SKumar Kartikeya Dwivedi 	if (ret) {
580164c2465SKumar Kartikeya Dwivedi 		/*
581164c2465SKumar Kartikeya Dwivedi 		 * If the tail is still pointing to us, then we are the final waiter,
582164c2465SKumar Kartikeya Dwivedi 		 * and are responsible for resetting the tail back to 0. Otherwise, if
583164c2465SKumar Kartikeya Dwivedi 		 * the cmpxchg operation fails, we signal the next waiter to take exit
584164c2465SKumar Kartikeya Dwivedi 		 * and try the same. For a waiter with tail node 'n':
585164c2465SKumar Kartikeya Dwivedi 		 *
586164c2465SKumar Kartikeya Dwivedi 		 * n,*,* -> 0,*,*
587164c2465SKumar Kartikeya Dwivedi 		 *
588164c2465SKumar Kartikeya Dwivedi 		 * When performing cmpxchg for the whole word (NR_CPUS > 16k), it is
589164c2465SKumar Kartikeya Dwivedi 		 * possible locked/pending bits keep changing and we see failures even
590164c2465SKumar Kartikeya Dwivedi 		 * when we remain the head of wait queue. However, eventually,
591164c2465SKumar Kartikeya Dwivedi 		 * pending bit owner will unset the pending bit, and new waiters
592164c2465SKumar Kartikeya Dwivedi 		 * will queue behind us. This will leave the lock owner in
593164c2465SKumar Kartikeya Dwivedi 		 * charge, and it will eventually either set locked bit to 0, or
594164c2465SKumar Kartikeya Dwivedi 		 * leave it as 1, allowing us to make progress.
595164c2465SKumar Kartikeya Dwivedi 		 *
596164c2465SKumar Kartikeya Dwivedi 		 * We terminate the whole wait queue for two reasons. Firstly,
597164c2465SKumar Kartikeya Dwivedi 		 * we eschew per-waiter timeouts with one applied at the head of
598164c2465SKumar Kartikeya Dwivedi 		 * the wait queue.  This allows everyone to break out faster
599164c2465SKumar Kartikeya Dwivedi 		 * once we've seen the owner / pending waiter not responding for
600164c2465SKumar Kartikeya Dwivedi 		 * the timeout duration from the head.  Secondly, it avoids
601164c2465SKumar Kartikeya Dwivedi 		 * complicated synchronization, because when not leaving in FIFO
602164c2465SKumar Kartikeya Dwivedi 		 * order, prev's next pointer needs to be fixed up etc.
603164c2465SKumar Kartikeya Dwivedi 		 */
604164c2465SKumar Kartikeya Dwivedi 		if (!try_cmpxchg_tail(lock, tail, 0)) {
605164c2465SKumar Kartikeya Dwivedi 			next = smp_cond_load_relaxed(&node->next, VAL);
606164c2465SKumar Kartikeya Dwivedi 			WRITE_ONCE(next->locked, RES_TIMEOUT_VAL);
607164c2465SKumar Kartikeya Dwivedi 		}
608164c2465SKumar Kartikeya Dwivedi 		lockevent_inc(rqspinlock_lock_timeout);
60931158ad0SKumar Kartikeya Dwivedi 		goto err_release_node;
610164c2465SKumar Kartikeya Dwivedi 	}
611a8fcf2a3SKumar Kartikeya Dwivedi 
612a8fcf2a3SKumar Kartikeya Dwivedi 	/*
613a8fcf2a3SKumar Kartikeya Dwivedi 	 * claim the lock:
614a8fcf2a3SKumar Kartikeya Dwivedi 	 *
615a8fcf2a3SKumar Kartikeya Dwivedi 	 * n,0,0 -> 0,0,1 : lock, uncontended
616a8fcf2a3SKumar Kartikeya Dwivedi 	 * *,*,0 -> *,*,1 : lock, contended
617a8fcf2a3SKumar Kartikeya Dwivedi 	 *
618a8fcf2a3SKumar Kartikeya Dwivedi 	 * If the queue head is the only one in the queue (lock value == tail)
619a8fcf2a3SKumar Kartikeya Dwivedi 	 * and nobody is pending, clear the tail code and grab the lock.
620a8fcf2a3SKumar Kartikeya Dwivedi 	 * Otherwise, we only need to grab the lock.
621a8fcf2a3SKumar Kartikeya Dwivedi 	 */
622a8fcf2a3SKumar Kartikeya Dwivedi 
623a8fcf2a3SKumar Kartikeya Dwivedi 	/*
624a8fcf2a3SKumar Kartikeya Dwivedi 	 * Note: at this point: (val & _Q_PENDING_MASK) == 0, because of the
625a8fcf2a3SKumar Kartikeya Dwivedi 	 *       above wait condition, therefore any concurrent setting of
626a8fcf2a3SKumar Kartikeya Dwivedi 	 *       PENDING will make the uncontended transition fail.
627a8fcf2a3SKumar Kartikeya Dwivedi 	 */
628a8fcf2a3SKumar Kartikeya Dwivedi 	if ((val & _Q_TAIL_MASK) == tail) {
629a8fcf2a3SKumar Kartikeya Dwivedi 		if (atomic_try_cmpxchg_relaxed(&lock->val, &val, _Q_LOCKED_VAL))
630a8fcf2a3SKumar Kartikeya Dwivedi 			goto release; /* No contention */
631a8fcf2a3SKumar Kartikeya Dwivedi 	}
632a8fcf2a3SKumar Kartikeya Dwivedi 
633a8fcf2a3SKumar Kartikeya Dwivedi 	/*
634a8fcf2a3SKumar Kartikeya Dwivedi 	 * Either somebody is queued behind us or _Q_PENDING_VAL got set
635a8fcf2a3SKumar Kartikeya Dwivedi 	 * which will then detect the remaining tail and queue behind us
636a8fcf2a3SKumar Kartikeya Dwivedi 	 * ensuring we'll see a @next.
637a8fcf2a3SKumar Kartikeya Dwivedi 	 */
638a8fcf2a3SKumar Kartikeya Dwivedi 	set_locked(lock);
639a8fcf2a3SKumar Kartikeya Dwivedi 
640a8fcf2a3SKumar Kartikeya Dwivedi 	/*
641a8fcf2a3SKumar Kartikeya Dwivedi 	 * contended path; wait for next if not observed yet, release.
642a8fcf2a3SKumar Kartikeya Dwivedi 	 */
643a8fcf2a3SKumar Kartikeya Dwivedi 	if (!next)
644a8fcf2a3SKumar Kartikeya Dwivedi 		next = smp_cond_load_relaxed(&node->next, (VAL));
645a8fcf2a3SKumar Kartikeya Dwivedi 
646a8fcf2a3SKumar Kartikeya Dwivedi 	arch_mcs_spin_unlock_contended(&next->locked);
647a8fcf2a3SKumar Kartikeya Dwivedi 
648a8fcf2a3SKumar Kartikeya Dwivedi release:
649a8fcf2a3SKumar Kartikeya Dwivedi 	trace_contention_end(lock, 0);
650a8fcf2a3SKumar Kartikeya Dwivedi 
651a8fcf2a3SKumar Kartikeya Dwivedi 	/*
652a8fcf2a3SKumar Kartikeya Dwivedi 	 * release the node
653a8fcf2a3SKumar Kartikeya Dwivedi 	 */
654a8fcf2a3SKumar Kartikeya Dwivedi 	__this_cpu_dec(rqnodes[0].mcs.count);
655164c2465SKumar Kartikeya Dwivedi 	return ret;
65631158ad0SKumar Kartikeya Dwivedi err_release_node:
65731158ad0SKumar Kartikeya Dwivedi 	trace_contention_end(lock, ret);
65831158ad0SKumar Kartikeya Dwivedi 	__this_cpu_dec(rqnodes[0].mcs.count);
65931158ad0SKumar Kartikeya Dwivedi err_release_entry:
66031158ad0SKumar Kartikeya Dwivedi 	release_held_lock_entry();
66131158ad0SKumar Kartikeya Dwivedi 	return ret;
662a8fcf2a3SKumar Kartikeya Dwivedi }
663a8fcf2a3SKumar Kartikeya Dwivedi EXPORT_SYMBOL_GPL(resilient_queued_spin_lock_slowpath);
664c9102a68SKumar Kartikeya Dwivedi 
665c9102a68SKumar Kartikeya Dwivedi #endif /* CONFIG_QUEUED_SPINLOCKS */
66697eb35f3SKumar Kartikeya Dwivedi 
66797eb35f3SKumar Kartikeya Dwivedi __bpf_kfunc_start_defs();
66897eb35f3SKumar Kartikeya Dwivedi 
bpf_res_spin_lock(struct bpf_res_spin_lock * lock)66997eb35f3SKumar Kartikeya Dwivedi __bpf_kfunc int bpf_res_spin_lock(struct bpf_res_spin_lock *lock)
67097eb35f3SKumar Kartikeya Dwivedi {
67197eb35f3SKumar Kartikeya Dwivedi 	int ret;
67297eb35f3SKumar Kartikeya Dwivedi 
67397eb35f3SKumar Kartikeya Dwivedi 	BUILD_BUG_ON(sizeof(rqspinlock_t) != sizeof(struct bpf_res_spin_lock));
67497eb35f3SKumar Kartikeya Dwivedi 	BUILD_BUG_ON(__alignof__(rqspinlock_t) != __alignof__(struct bpf_res_spin_lock));
67597eb35f3SKumar Kartikeya Dwivedi 
67697eb35f3SKumar Kartikeya Dwivedi 	preempt_disable();
67797eb35f3SKumar Kartikeya Dwivedi 	ret = res_spin_lock((rqspinlock_t *)lock);
67897eb35f3SKumar Kartikeya Dwivedi 	if (unlikely(ret)) {
67997eb35f3SKumar Kartikeya Dwivedi 		preempt_enable();
68097eb35f3SKumar Kartikeya Dwivedi 		return ret;
68197eb35f3SKumar Kartikeya Dwivedi 	}
68297eb35f3SKumar Kartikeya Dwivedi 	return 0;
68397eb35f3SKumar Kartikeya Dwivedi }
68497eb35f3SKumar Kartikeya Dwivedi 
bpf_res_spin_unlock(struct bpf_res_spin_lock * lock)68597eb35f3SKumar Kartikeya Dwivedi __bpf_kfunc void bpf_res_spin_unlock(struct bpf_res_spin_lock *lock)
68697eb35f3SKumar Kartikeya Dwivedi {
68797eb35f3SKumar Kartikeya Dwivedi 	res_spin_unlock((rqspinlock_t *)lock);
68897eb35f3SKumar Kartikeya Dwivedi 	preempt_enable();
68997eb35f3SKumar Kartikeya Dwivedi }
69097eb35f3SKumar Kartikeya Dwivedi 
bpf_res_spin_lock_irqsave(struct bpf_res_spin_lock * lock,unsigned long * flags__irq_flag)69197eb35f3SKumar Kartikeya Dwivedi __bpf_kfunc int bpf_res_spin_lock_irqsave(struct bpf_res_spin_lock *lock, unsigned long *flags__irq_flag)
69297eb35f3SKumar Kartikeya Dwivedi {
69397eb35f3SKumar Kartikeya Dwivedi 	u64 *ptr = (u64 *)flags__irq_flag;
69497eb35f3SKumar Kartikeya Dwivedi 	unsigned long flags;
69597eb35f3SKumar Kartikeya Dwivedi 	int ret;
69697eb35f3SKumar Kartikeya Dwivedi 
69797eb35f3SKumar Kartikeya Dwivedi 	preempt_disable();
69897eb35f3SKumar Kartikeya Dwivedi 	local_irq_save(flags);
69997eb35f3SKumar Kartikeya Dwivedi 	ret = res_spin_lock((rqspinlock_t *)lock);
70097eb35f3SKumar Kartikeya Dwivedi 	if (unlikely(ret)) {
70197eb35f3SKumar Kartikeya Dwivedi 		local_irq_restore(flags);
70297eb35f3SKumar Kartikeya Dwivedi 		preempt_enable();
70397eb35f3SKumar Kartikeya Dwivedi 		return ret;
70497eb35f3SKumar Kartikeya Dwivedi 	}
70597eb35f3SKumar Kartikeya Dwivedi 	*ptr = flags;
70697eb35f3SKumar Kartikeya Dwivedi 	return 0;
70797eb35f3SKumar Kartikeya Dwivedi }
70897eb35f3SKumar Kartikeya Dwivedi 
bpf_res_spin_unlock_irqrestore(struct bpf_res_spin_lock * lock,unsigned long * flags__irq_flag)70997eb35f3SKumar Kartikeya Dwivedi __bpf_kfunc void bpf_res_spin_unlock_irqrestore(struct bpf_res_spin_lock *lock, unsigned long *flags__irq_flag)
71097eb35f3SKumar Kartikeya Dwivedi {
71197eb35f3SKumar Kartikeya Dwivedi 	u64 *ptr = (u64 *)flags__irq_flag;
71297eb35f3SKumar Kartikeya Dwivedi 	unsigned long flags = *ptr;
71397eb35f3SKumar Kartikeya Dwivedi 
71497eb35f3SKumar Kartikeya Dwivedi 	res_spin_unlock((rqspinlock_t *)lock);
71597eb35f3SKumar Kartikeya Dwivedi 	local_irq_restore(flags);
71697eb35f3SKumar Kartikeya Dwivedi 	preempt_enable();
71797eb35f3SKumar Kartikeya Dwivedi }
71897eb35f3SKumar Kartikeya Dwivedi 
71997eb35f3SKumar Kartikeya Dwivedi __bpf_kfunc_end_defs();
72097eb35f3SKumar Kartikeya Dwivedi 
72197eb35f3SKumar Kartikeya Dwivedi BTF_KFUNCS_START(rqspinlock_kfunc_ids)
72297eb35f3SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bpf_res_spin_lock, KF_RET_NULL)
72397eb35f3SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bpf_res_spin_unlock)
72497eb35f3SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bpf_res_spin_lock_irqsave, KF_RET_NULL)
72597eb35f3SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bpf_res_spin_unlock_irqrestore)
72697eb35f3SKumar Kartikeya Dwivedi BTF_KFUNCS_END(rqspinlock_kfunc_ids)
72797eb35f3SKumar Kartikeya Dwivedi 
72897eb35f3SKumar Kartikeya Dwivedi static const struct btf_kfunc_id_set rqspinlock_kfunc_set = {
72997eb35f3SKumar Kartikeya Dwivedi 	.owner = THIS_MODULE,
73097eb35f3SKumar Kartikeya Dwivedi 	.set = &rqspinlock_kfunc_ids,
73197eb35f3SKumar Kartikeya Dwivedi };
73297eb35f3SKumar Kartikeya Dwivedi 
rqspinlock_register_kfuncs(void)73397eb35f3SKumar Kartikeya Dwivedi static __init int rqspinlock_register_kfuncs(void)
73497eb35f3SKumar Kartikeya Dwivedi {
73597eb35f3SKumar Kartikeya Dwivedi 	return register_btf_kfunc_id_set(BPF_PROG_TYPE_UNSPEC, &rqspinlock_kfunc_set);
73697eb35f3SKumar Kartikeya Dwivedi }
73797eb35f3SKumar Kartikeya Dwivedi late_initcall(rqspinlock_register_kfuncs);
738