xref: /openbmc/linux/arch/x86/xen/spinlock.c (revision b627b4ed)
1 /*
2  * Split spinlock implementation out into its own file, so it can be
3  * compiled in a FTRACE-compatible way.
4  */
5 #include <linux/kernel_stat.h>
6 #include <linux/spinlock.h>
7 #include <linux/debugfs.h>
8 #include <linux/log2.h>
9 
10 #include <asm/paravirt.h>
11 
12 #include <xen/interface/xen.h>
13 #include <xen/events.h>
14 
15 #include "xen-ops.h"
16 #include "debugfs.h"
17 
18 #ifdef CONFIG_XEN_DEBUG_FS
19 static struct xen_spinlock_stats
20 {
21 	u64 taken;
22 	u32 taken_slow;
23 	u32 taken_slow_nested;
24 	u32 taken_slow_pickup;
25 	u32 taken_slow_spurious;
26 	u32 taken_slow_irqenable;
27 
28 	u64 released;
29 	u32 released_slow;
30 	u32 released_slow_kicked;
31 
32 #define HISTO_BUCKETS	30
33 	u32 histo_spin_total[HISTO_BUCKETS+1];
34 	u32 histo_spin_spinning[HISTO_BUCKETS+1];
35 	u32 histo_spin_blocked[HISTO_BUCKETS+1];
36 
37 	u64 time_total;
38 	u64 time_spinning;
39 	u64 time_blocked;
40 } spinlock_stats;
41 
42 static u8 zero_stats;
43 
44 static unsigned lock_timeout = 1 << 10;
45 #define TIMEOUT lock_timeout
46 
47 static inline void check_zero(void)
48 {
49 	if (unlikely(zero_stats)) {
50 		memset(&spinlock_stats, 0, sizeof(spinlock_stats));
51 		zero_stats = 0;
52 	}
53 }
54 
55 #define ADD_STATS(elem, val)			\
56 	do { check_zero(); spinlock_stats.elem += (val); } while(0)
57 
58 static inline u64 spin_time_start(void)
59 {
60 	return xen_clocksource_read();
61 }
62 
63 static void __spin_time_accum(u64 delta, u32 *array)
64 {
65 	unsigned index = ilog2(delta);
66 
67 	check_zero();
68 
69 	if (index < HISTO_BUCKETS)
70 		array[index]++;
71 	else
72 		array[HISTO_BUCKETS]++;
73 }
74 
75 static inline void spin_time_accum_spinning(u64 start)
76 {
77 	u32 delta = xen_clocksource_read() - start;
78 
79 	__spin_time_accum(delta, spinlock_stats.histo_spin_spinning);
80 	spinlock_stats.time_spinning += delta;
81 }
82 
83 static inline void spin_time_accum_total(u64 start)
84 {
85 	u32 delta = xen_clocksource_read() - start;
86 
87 	__spin_time_accum(delta, spinlock_stats.histo_spin_total);
88 	spinlock_stats.time_total += delta;
89 }
90 
91 static inline void spin_time_accum_blocked(u64 start)
92 {
93 	u32 delta = xen_clocksource_read() - start;
94 
95 	__spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
96 	spinlock_stats.time_blocked += delta;
97 }
98 #else  /* !CONFIG_XEN_DEBUG_FS */
99 #define TIMEOUT			(1 << 10)
100 #define ADD_STATS(elem, val)	do { (void)(val); } while(0)
101 
102 static inline u64 spin_time_start(void)
103 {
104 	return 0;
105 }
106 
107 static inline void spin_time_accum_total(u64 start)
108 {
109 }
110 static inline void spin_time_accum_spinning(u64 start)
111 {
112 }
113 static inline void spin_time_accum_blocked(u64 start)
114 {
115 }
116 #endif  /* CONFIG_XEN_DEBUG_FS */
117 
118 struct xen_spinlock {
119 	unsigned char lock;		/* 0 -> free; 1 -> locked */
120 	unsigned short spinners;	/* count of waiting cpus */
121 };
122 
123 static int xen_spin_is_locked(struct raw_spinlock *lock)
124 {
125 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
126 
127 	return xl->lock != 0;
128 }
129 
130 static int xen_spin_is_contended(struct raw_spinlock *lock)
131 {
132 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
133 
134 	/* Not strictly true; this is only the count of contended
135 	   lock-takers entering the slow path. */
136 	return xl->spinners != 0;
137 }
138 
139 static int xen_spin_trylock(struct raw_spinlock *lock)
140 {
141 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
142 	u8 old = 1;
143 
144 	asm("xchgb %b0,%1"
145 	    : "+q" (old), "+m" (xl->lock) : : "memory");
146 
147 	return old == 0;
148 }
149 
150 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
151 static DEFINE_PER_CPU(struct xen_spinlock *, lock_spinners);
152 
153 /*
154  * Mark a cpu as interested in a lock.  Returns the CPU's previous
155  * lock of interest, in case we got preempted by an interrupt.
156  */
157 static inline struct xen_spinlock *spinning_lock(struct xen_spinlock *xl)
158 {
159 	struct xen_spinlock *prev;
160 
161 	prev = __get_cpu_var(lock_spinners);
162 	__get_cpu_var(lock_spinners) = xl;
163 
164 	wmb();			/* set lock of interest before count */
165 
166 	asm(LOCK_PREFIX " incw %0"
167 	    : "+m" (xl->spinners) : : "memory");
168 
169 	return prev;
170 }
171 
172 /*
173  * Mark a cpu as no longer interested in a lock.  Restores previous
174  * lock of interest (NULL for none).
175  */
176 static inline void unspinning_lock(struct xen_spinlock *xl, struct xen_spinlock *prev)
177 {
178 	asm(LOCK_PREFIX " decw %0"
179 	    : "+m" (xl->spinners) : : "memory");
180 	wmb();			/* decrement count before restoring lock */
181 	__get_cpu_var(lock_spinners) = prev;
182 }
183 
184 static noinline int xen_spin_lock_slow(struct raw_spinlock *lock, bool irq_enable)
185 {
186 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
187 	struct xen_spinlock *prev;
188 	int irq = __get_cpu_var(lock_kicker_irq);
189 	int ret;
190 	unsigned long flags;
191 	u64 start;
192 
193 	/* If kicker interrupts not initialized yet, just spin */
194 	if (irq == -1)
195 		return 0;
196 
197 	start = spin_time_start();
198 
199 	/* announce we're spinning */
200 	prev = spinning_lock(xl);
201 
202 	flags = __raw_local_save_flags();
203 	if (irq_enable) {
204 		ADD_STATS(taken_slow_irqenable, 1);
205 		raw_local_irq_enable();
206 	}
207 
208 	ADD_STATS(taken_slow, 1);
209 	ADD_STATS(taken_slow_nested, prev != NULL);
210 
211 	do {
212 		/* clear pending */
213 		xen_clear_irq_pending(irq);
214 
215 		/* check again make sure it didn't become free while
216 		   we weren't looking  */
217 		ret = xen_spin_trylock(lock);
218 		if (ret) {
219 			ADD_STATS(taken_slow_pickup, 1);
220 
221 			/*
222 			 * If we interrupted another spinlock while it
223 			 * was blocking, make sure it doesn't block
224 			 * without rechecking the lock.
225 			 */
226 			if (prev != NULL)
227 				xen_set_irq_pending(irq);
228 			goto out;
229 		}
230 
231 		/*
232 		 * Block until irq becomes pending.  If we're
233 		 * interrupted at this point (after the trylock but
234 		 * before entering the block), then the nested lock
235 		 * handler guarantees that the irq will be left
236 		 * pending if there's any chance the lock became free;
237 		 * xen_poll_irq() returns immediately if the irq is
238 		 * pending.
239 		 */
240 		xen_poll_irq(irq);
241 		ADD_STATS(taken_slow_spurious, !xen_test_irq_pending(irq));
242 	} while (!xen_test_irq_pending(irq)); /* check for spurious wakeups */
243 
244 	kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
245 
246 out:
247 	raw_local_irq_restore(flags);
248 	unspinning_lock(xl, prev);
249 	spin_time_accum_blocked(start);
250 
251 	return ret;
252 }
253 
254 static inline void __xen_spin_lock(struct raw_spinlock *lock, bool irq_enable)
255 {
256 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
257 	unsigned timeout;
258 	u8 oldval;
259 	u64 start_spin;
260 
261 	ADD_STATS(taken, 1);
262 
263 	start_spin = spin_time_start();
264 
265 	do {
266 		u64 start_spin_fast = spin_time_start();
267 
268 		timeout = TIMEOUT;
269 
270 		asm("1: xchgb %1,%0\n"
271 		    "   testb %1,%1\n"
272 		    "   jz 3f\n"
273 		    "2: rep;nop\n"
274 		    "   cmpb $0,%0\n"
275 		    "   je 1b\n"
276 		    "   dec %2\n"
277 		    "   jnz 2b\n"
278 		    "3:\n"
279 		    : "+m" (xl->lock), "=q" (oldval), "+r" (timeout)
280 		    : "1" (1)
281 		    : "memory");
282 
283 		spin_time_accum_spinning(start_spin_fast);
284 
285 	} while (unlikely(oldval != 0 &&
286 			  (TIMEOUT == ~0 || !xen_spin_lock_slow(lock, irq_enable))));
287 
288 	spin_time_accum_total(start_spin);
289 }
290 
291 static void xen_spin_lock(struct raw_spinlock *lock)
292 {
293 	__xen_spin_lock(lock, false);
294 }
295 
296 static void xen_spin_lock_flags(struct raw_spinlock *lock, unsigned long flags)
297 {
298 	__xen_spin_lock(lock, !raw_irqs_disabled_flags(flags));
299 }
300 
301 static noinline void xen_spin_unlock_slow(struct xen_spinlock *xl)
302 {
303 	int cpu;
304 
305 	ADD_STATS(released_slow, 1);
306 
307 	for_each_online_cpu(cpu) {
308 		/* XXX should mix up next cpu selection */
309 		if (per_cpu(lock_spinners, cpu) == xl) {
310 			ADD_STATS(released_slow_kicked, 1);
311 			xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
312 			break;
313 		}
314 	}
315 }
316 
317 static void xen_spin_unlock(struct raw_spinlock *lock)
318 {
319 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
320 
321 	ADD_STATS(released, 1);
322 
323 	smp_wmb();		/* make sure no writes get moved after unlock */
324 	xl->lock = 0;		/* release lock */
325 
326 	/* make sure unlock happens before kick */
327 	barrier();
328 
329 	if (unlikely(xl->spinners))
330 		xen_spin_unlock_slow(xl);
331 }
332 
333 static irqreturn_t dummy_handler(int irq, void *dev_id)
334 {
335 	BUG();
336 	return IRQ_HANDLED;
337 }
338 
339 void __cpuinit xen_init_lock_cpu(int cpu)
340 {
341 	int irq;
342 	const char *name;
343 
344 	name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
345 	irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
346 				     cpu,
347 				     dummy_handler,
348 				     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
349 				     name,
350 				     NULL);
351 
352 	if (irq >= 0) {
353 		disable_irq(irq); /* make sure it's never delivered */
354 		per_cpu(lock_kicker_irq, cpu) = irq;
355 	}
356 
357 	printk("cpu %d spinlock event irq %d\n", cpu, irq);
358 }
359 
360 void xen_uninit_lock_cpu(int cpu)
361 {
362 	unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
363 }
364 
365 void __init xen_init_spinlocks(void)
366 {
367 	pv_lock_ops.spin_is_locked = xen_spin_is_locked;
368 	pv_lock_ops.spin_is_contended = xen_spin_is_contended;
369 	pv_lock_ops.spin_lock = xen_spin_lock;
370 	pv_lock_ops.spin_lock_flags = xen_spin_lock_flags;
371 	pv_lock_ops.spin_trylock = xen_spin_trylock;
372 	pv_lock_ops.spin_unlock = xen_spin_unlock;
373 }
374 
375 #ifdef CONFIG_XEN_DEBUG_FS
376 
377 static struct dentry *d_spin_debug;
378 
379 static int __init xen_spinlock_debugfs(void)
380 {
381 	struct dentry *d_xen = xen_init_debugfs();
382 
383 	if (d_xen == NULL)
384 		return -ENOMEM;
385 
386 	d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
387 
388 	debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
389 
390 	debugfs_create_u32("timeout", 0644, d_spin_debug, &lock_timeout);
391 
392 	debugfs_create_u64("taken", 0444, d_spin_debug, &spinlock_stats.taken);
393 	debugfs_create_u32("taken_slow", 0444, d_spin_debug,
394 			   &spinlock_stats.taken_slow);
395 	debugfs_create_u32("taken_slow_nested", 0444, d_spin_debug,
396 			   &spinlock_stats.taken_slow_nested);
397 	debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
398 			   &spinlock_stats.taken_slow_pickup);
399 	debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
400 			   &spinlock_stats.taken_slow_spurious);
401 	debugfs_create_u32("taken_slow_irqenable", 0444, d_spin_debug,
402 			   &spinlock_stats.taken_slow_irqenable);
403 
404 	debugfs_create_u64("released", 0444, d_spin_debug, &spinlock_stats.released);
405 	debugfs_create_u32("released_slow", 0444, d_spin_debug,
406 			   &spinlock_stats.released_slow);
407 	debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
408 			   &spinlock_stats.released_slow_kicked);
409 
410 	debugfs_create_u64("time_spinning", 0444, d_spin_debug,
411 			   &spinlock_stats.time_spinning);
412 	debugfs_create_u64("time_blocked", 0444, d_spin_debug,
413 			   &spinlock_stats.time_blocked);
414 	debugfs_create_u64("time_total", 0444, d_spin_debug,
415 			   &spinlock_stats.time_total);
416 
417 	xen_debugfs_create_u32_array("histo_total", 0444, d_spin_debug,
418 				     spinlock_stats.histo_spin_total, HISTO_BUCKETS + 1);
419 	xen_debugfs_create_u32_array("histo_spinning", 0444, d_spin_debug,
420 				     spinlock_stats.histo_spin_spinning, HISTO_BUCKETS + 1);
421 	xen_debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
422 				     spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
423 
424 	return 0;
425 }
426 fs_initcall(xen_spinlock_debugfs);
427 
428 #endif	/* CONFIG_XEN_DEBUG_FS */
429