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
 80	/*
 81	 * The TILE architecture doesn't do read speculation; therefore
 82	 * a control dependency guarantees a LOAD->{LOAD,STORE} order.
 83	 */
 84	barrier();
 85}
 86EXPORT_SYMBOL(arch_spin_unlock_wait);
 87
 88/*
 89 * The low byte is always reserved to be the marker for a "tns" operation
 90 * since the low bit is set to "1" by a tns.  The next seven bits are
 91 * zeroes.  The next byte holds the "next" writer value, i.e. the ticket
 92 * available for the next task that wants to write.  The third byte holds
 93 * the current writer value, i.e. the writer who holds the current ticket.
 94 * If current == next == 0, there are no interested writers.
 95 */
 96#define WR_NEXT_SHIFT   _WR_NEXT_SHIFT
 97#define WR_CURR_SHIFT   _WR_CURR_SHIFT
 98#define WR_WIDTH        _WR_WIDTH
 99#define WR_MASK         ((1 << WR_WIDTH) - 1)
100
101/*
102 * The last eight bits hold the active reader count.  This has to be
103 * zero before a writer can start to write.
104 */
105#define RD_COUNT_SHIFT  _RD_COUNT_SHIFT
106#define RD_COUNT_WIDTH  _RD_COUNT_WIDTH
107#define RD_COUNT_MASK   ((1 << RD_COUNT_WIDTH) - 1)
108
109
110/*
111 * We can get the read lock if everything but the reader bits (which
112 * are in the high part of the word) is zero, i.e. no active or
113 * waiting writers, no tns.
114 *
115 * We guard the tns/store-back with an interrupt critical section to
116 * preserve the semantic that the same read lock can be acquired in an
117 * interrupt context.
118 */
119int arch_read_trylock(arch_rwlock_t *rwlock)
120{
121	u32 val;
122	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
123	val = __insn_tns((int *)&rwlock->lock);
124	if (likely((val << _RD_COUNT_WIDTH) == 0)) {
125		val += 1 << RD_COUNT_SHIFT;
126		rwlock->lock = val;
127		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
128		BUG_ON(val == 0);  /* we don't expect wraparound */
129		return 1;
130	}
131	if ((val & 1) == 0)
132		rwlock->lock = val;
133	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
134	return 0;
135}
136EXPORT_SYMBOL(arch_read_trylock);
137
138/*
139 * Spin doing arch_read_trylock() until we acquire the lock.
140 * ISSUE: This approach can permanently starve readers.  A reader who sees
141 * a writer could instead take a ticket lock (just like a writer would),
142 * and atomically enter read mode (with 1 reader) when it gets the ticket.
143 * This way both readers and writers would always make forward progress
144 * in a finite time.
145 */
146void arch_read_lock(arch_rwlock_t *rwlock)
147{
148	u32 iterations = 0;
149	while (unlikely(!arch_read_trylock(rwlock)))
150		delay_backoff(iterations++);
151}
152EXPORT_SYMBOL(arch_read_lock);
153
154void arch_read_unlock(arch_rwlock_t *rwlock)
155{
156	u32 val, iterations = 0;
157
158	mb();  /* guarantee anything modified under the lock is visible */
159	for (;;) {
160		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
161		val = __insn_tns((int *)&rwlock->lock);
162		if (likely((val & 1) == 0)) {
163			rwlock->lock = val - (1 << _RD_COUNT_SHIFT);
164			__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
165			break;
166		}
167		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
168		delay_backoff(iterations++);
169	}
170}
171EXPORT_SYMBOL(arch_read_unlock);
172
173/*
174 * We don't need an interrupt critical section here (unlike for
175 * arch_read_lock) since we should never use a bare write lock where
176 * it could be interrupted by code that could try to re-acquire it.
177 */
178void arch_write_lock(arch_rwlock_t *rwlock)
179{
180	/*
181	 * The trailing underscore on this variable (and curr_ below)
182	 * reminds us that the high bits are garbage; we mask them out
183	 * when we compare them.
184	 */
185	u32 my_ticket_;
186	u32 iterations = 0;
187	u32 val = __insn_tns((int *)&rwlock->lock);
188
189	if (likely(val == 0)) {
190		rwlock->lock = 1 << _WR_NEXT_SHIFT;
191		return;
192	}
193
194	/*
195	 * Wait until there are no readers, then bump up the next
196	 * field and capture the ticket value.
197	 */
198	for (;;) {
199		if (!(val & 1)) {
200			if ((val >> RD_COUNT_SHIFT) == 0)
201				break;
202			rwlock->lock = val;
203		}
204		delay_backoff(iterations++);
205		val = __insn_tns((int *)&rwlock->lock);
206	}
207
208	/* Take out the next ticket and extract my ticket value. */
209	rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT);
210	my_ticket_ = val >> WR_NEXT_SHIFT;
211
212	/* Wait until the "current" field matches our ticket. */
213	for (;;) {
214		u32 curr_ = val >> WR_CURR_SHIFT;
215		u32 delta = ((my_ticket_ - curr_) & WR_MASK);
216		if (likely(delta == 0))
217			break;
218
219		/* Delay based on how many lock-holders are still out there. */
220		relax((256 / CYCLES_PER_RELAX_LOOP) * delta);
221
222		/*
223		 * Get a non-tns value to check; we don't need to tns
224		 * it ourselves.  Since we're not tns'ing, we retry
225		 * more rapidly to get a valid value.
226		 */
227		while ((val = rwlock->lock) & 1)
228			relax(4);
229	}
230}
231EXPORT_SYMBOL(arch_write_lock);
232
233int arch_write_trylock(arch_rwlock_t *rwlock)
234{
235	u32 val = __insn_tns((int *)&rwlock->lock);
236
237	/*
238	 * If a tns is in progress, or there's a waiting or active locker,
239	 * or active readers, we can't take the lock, so give up.
240	 */
241	if (unlikely(val != 0)) {
242		if (!(val & 1))
243			rwlock->lock = val;
244		return 0;
245	}
246
247	/* Set the "next" field to mark it locked. */
248	rwlock->lock = 1 << _WR_NEXT_SHIFT;
249	return 1;
250}
251EXPORT_SYMBOL(arch_write_trylock);
252
253void arch_write_unlock(arch_rwlock_t *rwlock)
254{
255	u32 val, eq, mask;
256
257	mb();  /* guarantee anything modified under the lock is visible */
258	val = __insn_tns((int *)&rwlock->lock);
259	if (likely(val == (1 << _WR_NEXT_SHIFT))) {
260		rwlock->lock = 0;
261		return;
262	}
263	while (unlikely(val & 1)) {
264		/* Limited backoff since we are the highest-priority task. */
265		relax(4);
266		val = __insn_tns((int *)&rwlock->lock);
267	}
268	mask = 1 << WR_CURR_SHIFT;
269	val = __insn_addb(val, mask);
270	eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT));
271	val = __insn_mz(eq & mask, val);
272	rwlock->lock = val;
273}
274EXPORT_SYMBOL(arch_write_unlock);