1/*
2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
3 *
4 *   This program is free software; you can redistribute it and/or
5 *   modify it under the terms of the GNU General Public License
6 *   as published by the Free Software Foundation, version 2.
7 *
8 *   This program is distributed in the hope that it will be useful, but
9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 *   NON INFRINGEMENT.  See the GNU General Public License for
12 *   more details.
13 */
14
15#include <linux/spinlock.h>
16#include <linux/module.h>
17#include <asm/processor.h>
18#include <arch/spr_def.h>
19
20#include "spinlock_common.h"
21
22void arch_spin_lock(arch_spinlock_t *lock)
23{
24	int my_ticket;
25	int iterations = 0;
26	int delta;
27
28	while ((my_ticket = __insn_tns((void *)&lock->next_ticket)) & 1)
29		delay_backoff(iterations++);
30
31	/* Increment the next ticket number, implicitly releasing tns lock. */
32	lock->next_ticket = my_ticket + TICKET_QUANTUM;
33
34	/* Wait until it's our turn. */
35	while ((delta = my_ticket - lock->current_ticket) != 0)
36		relax((128 / CYCLES_PER_RELAX_LOOP) * delta);
37}
38EXPORT_SYMBOL(arch_spin_lock);
39
40int arch_spin_trylock(arch_spinlock_t *lock)
41{
42	/*
43	 * Grab a ticket; no need to retry if it's busy, we'll just
44	 * treat that the same as "locked", since someone else
45	 * will lock it momentarily anyway.
46	 */
47	int my_ticket = __insn_tns((void *)&lock->next_ticket);
48
49	if (my_ticket == lock->current_ticket) {
50		/* Not currently locked, so lock it by keeping this ticket. */
51		lock->next_ticket = my_ticket + TICKET_QUANTUM;
52		/* Success! */
53		return 1;
54	}
55
56	if (!(my_ticket & 1)) {
57		/* Release next_ticket. */
58		lock->next_ticket = my_ticket;
59	}
60
61	return 0;
62}
63EXPORT_SYMBOL(arch_spin_trylock);
64
65void arch_spin_unlock_wait(arch_spinlock_t *lock)
66{
67	u32 iterations = 0;
68	int curr = READ_ONCE(lock->current_ticket);
69	int next = READ_ONCE(lock->next_ticket);
70
71	/* Return immediately if unlocked. */
72	if (next == curr)
73		return;
74
75	/* Wait until the current locker has released the lock. */
76	do {
77		delay_backoff(iterations++);
78	} while (READ_ONCE(lock->current_ticket) == curr);
79}
80EXPORT_SYMBOL(arch_spin_unlock_wait);
81
82/*
83 * The low byte is always reserved to be the marker for a "tns" operation
84 * since the low bit is set to "1" by a tns.  The next seven bits are
85 * zeroes.  The next byte holds the "next" writer value, i.e. the ticket
86 * available for the next task that wants to write.  The third byte holds
87 * the current writer value, i.e. the writer who holds the current ticket.
88 * If current == next == 0, there are no interested writers.
89 */
90#define WR_NEXT_SHIFT   _WR_NEXT_SHIFT
91#define WR_CURR_SHIFT   _WR_CURR_SHIFT
92#define WR_WIDTH        _WR_WIDTH
93#define WR_MASK         ((1 << WR_WIDTH) - 1)
94
95/*
96 * The last eight bits hold the active reader count.  This has to be
97 * zero before a writer can start to write.
98 */
99#define RD_COUNT_SHIFT  _RD_COUNT_SHIFT
100#define RD_COUNT_WIDTH  _RD_COUNT_WIDTH
101#define RD_COUNT_MASK   ((1 << RD_COUNT_WIDTH) - 1)
102
103
104/*
105 * We can get the read lock if everything but the reader bits (which
106 * are in the high part of the word) is zero, i.e. no active or
107 * waiting writers, no tns.
108 *
109 * We guard the tns/store-back with an interrupt critical section to
110 * preserve the semantic that the same read lock can be acquired in an
111 * interrupt context.
112 */
113int arch_read_trylock(arch_rwlock_t *rwlock)
114{
115	u32 val;
116	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
117	val = __insn_tns((int *)&rwlock->lock);
118	if (likely((val << _RD_COUNT_WIDTH) == 0)) {
119		val += 1 << RD_COUNT_SHIFT;
120		rwlock->lock = val;
121		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
122		BUG_ON(val == 0);  /* we don't expect wraparound */
123		return 1;
124	}
125	if ((val & 1) == 0)
126		rwlock->lock = val;
127	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
128	return 0;
129}
130EXPORT_SYMBOL(arch_read_trylock);
131
132/*
133 * Spin doing arch_read_trylock() until we acquire the lock.
134 * ISSUE: This approach can permanently starve readers.  A reader who sees
135 * a writer could instead take a ticket lock (just like a writer would),
136 * and atomically enter read mode (with 1 reader) when it gets the ticket.
137 * This way both readers and writers would always make forward progress
138 * in a finite time.
139 */
140void arch_read_lock(arch_rwlock_t *rwlock)
141{
142	u32 iterations = 0;
143	while (unlikely(!arch_read_trylock(rwlock)))
144		delay_backoff(iterations++);
145}
146EXPORT_SYMBOL(arch_read_lock);
147
148void arch_read_unlock(arch_rwlock_t *rwlock)
149{
150	u32 val, iterations = 0;
151
152	mb();  /* guarantee anything modified under the lock is visible */
153	for (;;) {
154		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
155		val = __insn_tns((int *)&rwlock->lock);
156		if (likely((val & 1) == 0)) {
157			rwlock->lock = val - (1 << _RD_COUNT_SHIFT);
158			__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
159			break;
160		}
161		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
162		delay_backoff(iterations++);
163	}
164}
165EXPORT_SYMBOL(arch_read_unlock);
166
167/*
168 * We don't need an interrupt critical section here (unlike for
169 * arch_read_lock) since we should never use a bare write lock where
170 * it could be interrupted by code that could try to re-acquire it.
171 */
172void arch_write_lock(arch_rwlock_t *rwlock)
173{
174	/*
175	 * The trailing underscore on this variable (and curr_ below)
176	 * reminds us that the high bits are garbage; we mask them out
177	 * when we compare them.
178	 */
179	u32 my_ticket_;
180	u32 iterations = 0;
181	u32 val = __insn_tns((int *)&rwlock->lock);
182
183	if (likely(val == 0)) {
184		rwlock->lock = 1 << _WR_NEXT_SHIFT;
185		return;
186	}
187
188	/*
189	 * Wait until there are no readers, then bump up the next
190	 * field and capture the ticket value.
191	 */
192	for (;;) {
193		if (!(val & 1)) {
194			if ((val >> RD_COUNT_SHIFT) == 0)
195				break;
196			rwlock->lock = val;
197		}
198		delay_backoff(iterations++);
199		val = __insn_tns((int *)&rwlock->lock);
200	}
201
202	/* Take out the next ticket and extract my ticket value. */
203	rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT);
204	my_ticket_ = val >> WR_NEXT_SHIFT;
205
206	/* Wait until the "current" field matches our ticket. */
207	for (;;) {
208		u32 curr_ = val >> WR_CURR_SHIFT;
209		u32 delta = ((my_ticket_ - curr_) & WR_MASK);
210		if (likely(delta == 0))
211			break;
212
213		/* Delay based on how many lock-holders are still out there. */
214		relax((256 / CYCLES_PER_RELAX_LOOP) * delta);
215
216		/*
217		 * Get a non-tns value to check; we don't need to tns
218		 * it ourselves.  Since we're not tns'ing, we retry
219		 * more rapidly to get a valid value.
220		 */
221		while ((val = rwlock->lock) & 1)
222			relax(4);
223	}
224}
225EXPORT_SYMBOL(arch_write_lock);
226
227int arch_write_trylock(arch_rwlock_t *rwlock)
228{
229	u32 val = __insn_tns((int *)&rwlock->lock);
230
231	/*
232	 * If a tns is in progress, or there's a waiting or active locker,
233	 * or active readers, we can't take the lock, so give up.
234	 */
235	if (unlikely(val != 0)) {
236		if (!(val & 1))
237			rwlock->lock = val;
238		return 0;
239	}
240
241	/* Set the "next" field to mark it locked. */
242	rwlock->lock = 1 << _WR_NEXT_SHIFT;
243	return 1;
244}
245EXPORT_SYMBOL(arch_write_trylock);
246
247void arch_write_unlock(arch_rwlock_t *rwlock)
248{
249	u32 val, eq, mask;
250
251	mb();  /* guarantee anything modified under the lock is visible */
252	val = __insn_tns((int *)&rwlock->lock);
253	if (likely(val == (1 << _WR_NEXT_SHIFT))) {
254		rwlock->lock = 0;
255		return;
256	}
257	while (unlikely(val & 1)) {
258		/* Limited backoff since we are the highest-priority task. */
259		relax(4);
260		val = __insn_tns((int *)&rwlock->lock);
261	}
262	mask = 1 << WR_CURR_SHIFT;
263	val = __insn_addb(val, mask);
264	eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT));
265	val = __insn_mz(eq & mask, val);
266	rwlock->lock = val;
267}
268EXPORT_SYMBOL(arch_write_unlock);
269