xref: /linux/arch/x86/xen/spinlock.c (revision ff5599816711d2e67da2d7561fd36ac48debd433)
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 #include <linux/gfp.h>
10 #include <linux/slab.h>
11 
12 #include <asm/paravirt.h>
13 
14 #include <xen/interface/xen.h>
15 #include <xen/events.h>
16 
17 #include "xen-ops.h"
18 #include "debugfs.h"
19 
20 #ifdef CONFIG_XEN_DEBUG_FS
21 static struct xen_spinlock_stats
22 {
23 	u64 taken;
24 	u32 taken_slow;
25 	u32 taken_slow_nested;
26 	u32 taken_slow_pickup;
27 	u32 taken_slow_spurious;
28 	u32 taken_slow_irqenable;
29 
30 	u64 released;
31 	u32 released_slow;
32 	u32 released_slow_kicked;
33 
34 #define HISTO_BUCKETS	30
35 	u32 histo_spin_total[HISTO_BUCKETS+1];
36 	u32 histo_spin_spinning[HISTO_BUCKETS+1];
37 	u32 histo_spin_blocked[HISTO_BUCKETS+1];
38 
39 	u64 time_total;
40 	u64 time_spinning;
41 	u64 time_blocked;
42 } spinlock_stats;
43 
44 static u8 zero_stats;
45 
46 static unsigned lock_timeout = 1 << 10;
47 #define TIMEOUT lock_timeout
48 
49 static inline void check_zero(void)
50 {
51 	if (unlikely(zero_stats)) {
52 		memset(&spinlock_stats, 0, sizeof(spinlock_stats));
53 		zero_stats = 0;
54 	}
55 }
56 
57 #define ADD_STATS(elem, val)			\
58 	do { check_zero(); spinlock_stats.elem += (val); } while(0)
59 
60 static inline u64 spin_time_start(void)
61 {
62 	return xen_clocksource_read();
63 }
64 
65 static void __spin_time_accum(u64 delta, u32 *array)
66 {
67 	unsigned index = ilog2(delta);
68 
69 	check_zero();
70 
71 	if (index < HISTO_BUCKETS)
72 		array[index]++;
73 	else
74 		array[HISTO_BUCKETS]++;
75 }
76 
77 static inline void spin_time_accum_spinning(u64 start)
78 {
79 	u32 delta = xen_clocksource_read() - start;
80 
81 	__spin_time_accum(delta, spinlock_stats.histo_spin_spinning);
82 	spinlock_stats.time_spinning += delta;
83 }
84 
85 static inline void spin_time_accum_total(u64 start)
86 {
87 	u32 delta = xen_clocksource_read() - start;
88 
89 	__spin_time_accum(delta, spinlock_stats.histo_spin_total);
90 	spinlock_stats.time_total += delta;
91 }
92 
93 static inline void spin_time_accum_blocked(u64 start)
94 {
95 	u32 delta = xen_clocksource_read() - start;
96 
97 	__spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
98 	spinlock_stats.time_blocked += delta;
99 }
100 #else  /* !CONFIG_XEN_DEBUG_FS */
101 #define TIMEOUT			(1 << 10)
102 #define ADD_STATS(elem, val)	do { (void)(val); } while(0)
103 
104 static inline u64 spin_time_start(void)
105 {
106 	return 0;
107 }
108 
109 static inline void spin_time_accum_total(u64 start)
110 {
111 }
112 static inline void spin_time_accum_spinning(u64 start)
113 {
114 }
115 static inline void spin_time_accum_blocked(u64 start)
116 {
117 }
118 #endif  /* CONFIG_XEN_DEBUG_FS */
119 
120 /*
121  * Size struct xen_spinlock so it's the same as arch_spinlock_t.
122  */
123 #if NR_CPUS < 256
124 typedef u8 xen_spinners_t;
125 # define inc_spinners(xl) \
126 	asm(LOCK_PREFIX " incb %0" : "+m" ((xl)->spinners) : : "memory");
127 # define dec_spinners(xl) \
128 	asm(LOCK_PREFIX " decb %0" : "+m" ((xl)->spinners) : : "memory");
129 #else
130 typedef u16 xen_spinners_t;
131 # define inc_spinners(xl) \
132 	asm(LOCK_PREFIX " incw %0" : "+m" ((xl)->spinners) : : "memory");
133 # define dec_spinners(xl) \
134 	asm(LOCK_PREFIX " decw %0" : "+m" ((xl)->spinners) : : "memory");
135 #endif
136 
137 struct xen_spinlock {
138 	unsigned char lock;		/* 0 -> free; 1 -> locked */
139 	xen_spinners_t spinners;	/* count of waiting cpus */
140 };
141 
142 static int xen_spin_is_locked(struct arch_spinlock *lock)
143 {
144 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
145 
146 	return xl->lock != 0;
147 }
148 
149 static int xen_spin_is_contended(struct arch_spinlock *lock)
150 {
151 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
152 
153 	/* Not strictly true; this is only the count of contended
154 	   lock-takers entering the slow path. */
155 	return xl->spinners != 0;
156 }
157 
158 static int xen_spin_trylock(struct arch_spinlock *lock)
159 {
160 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
161 	u8 old = 1;
162 
163 	asm("xchgb %b0,%1"
164 	    : "+q" (old), "+m" (xl->lock) : : "memory");
165 
166 	return old == 0;
167 }
168 
169 static DEFINE_PER_CPU(char *, irq_name);
170 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
171 static DEFINE_PER_CPU(struct xen_spinlock *, lock_spinners);
172 
173 /*
174  * Mark a cpu as interested in a lock.  Returns the CPU's previous
175  * lock of interest, in case we got preempted by an interrupt.
176  */
177 static inline struct xen_spinlock *spinning_lock(struct xen_spinlock *xl)
178 {
179 	struct xen_spinlock *prev;
180 
181 	prev = __this_cpu_read(lock_spinners);
182 	__this_cpu_write(lock_spinners, xl);
183 
184 	wmb();			/* set lock of interest before count */
185 
186 	inc_spinners(xl);
187 
188 	return prev;
189 }
190 
191 /*
192  * Mark a cpu as no longer interested in a lock.  Restores previous
193  * lock of interest (NULL for none).
194  */
195 static inline void unspinning_lock(struct xen_spinlock *xl, struct xen_spinlock *prev)
196 {
197 	dec_spinners(xl);
198 	wmb();			/* decrement count before restoring lock */
199 	__this_cpu_write(lock_spinners, prev);
200 }
201 
202 static noinline int xen_spin_lock_slow(struct arch_spinlock *lock, bool irq_enable)
203 {
204 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
205 	struct xen_spinlock *prev;
206 	int irq = __this_cpu_read(lock_kicker_irq);
207 	int ret;
208 	u64 start;
209 
210 	/* If kicker interrupts not initialized yet, just spin */
211 	if (irq == -1)
212 		return 0;
213 
214 	start = spin_time_start();
215 
216 	/* announce we're spinning */
217 	prev = spinning_lock(xl);
218 
219 	ADD_STATS(taken_slow, 1);
220 	ADD_STATS(taken_slow_nested, prev != NULL);
221 
222 	do {
223 		unsigned long flags;
224 
225 		/* clear pending */
226 		xen_clear_irq_pending(irq);
227 
228 		/* check again make sure it didn't become free while
229 		   we weren't looking  */
230 		ret = xen_spin_trylock(lock);
231 		if (ret) {
232 			ADD_STATS(taken_slow_pickup, 1);
233 
234 			/*
235 			 * If we interrupted another spinlock while it
236 			 * was blocking, make sure it doesn't block
237 			 * without rechecking the lock.
238 			 */
239 			if (prev != NULL)
240 				xen_set_irq_pending(irq);
241 			goto out;
242 		}
243 
244 		flags = arch_local_save_flags();
245 		if (irq_enable) {
246 			ADD_STATS(taken_slow_irqenable, 1);
247 			raw_local_irq_enable();
248 		}
249 
250 		/*
251 		 * Block until irq becomes pending.  If we're
252 		 * interrupted at this point (after the trylock but
253 		 * before entering the block), then the nested lock
254 		 * handler guarantees that the irq will be left
255 		 * pending if there's any chance the lock became free;
256 		 * xen_poll_irq() returns immediately if the irq is
257 		 * pending.
258 		 */
259 		xen_poll_irq(irq);
260 
261 		raw_local_irq_restore(flags);
262 
263 		ADD_STATS(taken_slow_spurious, !xen_test_irq_pending(irq));
264 	} while (!xen_test_irq_pending(irq)); /* check for spurious wakeups */
265 
266 	kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
267 
268 out:
269 	unspinning_lock(xl, prev);
270 	spin_time_accum_blocked(start);
271 
272 	return ret;
273 }
274 
275 static inline void __xen_spin_lock(struct arch_spinlock *lock, bool irq_enable)
276 {
277 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
278 	unsigned timeout;
279 	u8 oldval;
280 	u64 start_spin;
281 
282 	ADD_STATS(taken, 1);
283 
284 	start_spin = spin_time_start();
285 
286 	do {
287 		u64 start_spin_fast = spin_time_start();
288 
289 		timeout = TIMEOUT;
290 
291 		asm("1: xchgb %1,%0\n"
292 		    "   testb %1,%1\n"
293 		    "   jz 3f\n"
294 		    "2: rep;nop\n"
295 		    "   cmpb $0,%0\n"
296 		    "   je 1b\n"
297 		    "   dec %2\n"
298 		    "   jnz 2b\n"
299 		    "3:\n"
300 		    : "+m" (xl->lock), "=q" (oldval), "+r" (timeout)
301 		    : "1" (1)
302 		    : "memory");
303 
304 		spin_time_accum_spinning(start_spin_fast);
305 
306 	} while (unlikely(oldval != 0 &&
307 			  (TIMEOUT == ~0 || !xen_spin_lock_slow(lock, irq_enable))));
308 
309 	spin_time_accum_total(start_spin);
310 }
311 
312 static void xen_spin_lock(struct arch_spinlock *lock)
313 {
314 	__xen_spin_lock(lock, false);
315 }
316 
317 static void xen_spin_lock_flags(struct arch_spinlock *lock, unsigned long flags)
318 {
319 	__xen_spin_lock(lock, !raw_irqs_disabled_flags(flags));
320 }
321 
322 static noinline void xen_spin_unlock_slow(struct xen_spinlock *xl)
323 {
324 	int cpu;
325 
326 	ADD_STATS(released_slow, 1);
327 
328 	for_each_online_cpu(cpu) {
329 		/* XXX should mix up next cpu selection */
330 		if (per_cpu(lock_spinners, cpu) == xl) {
331 			ADD_STATS(released_slow_kicked, 1);
332 			xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
333 		}
334 	}
335 }
336 
337 static void xen_spin_unlock(struct arch_spinlock *lock)
338 {
339 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
340 
341 	ADD_STATS(released, 1);
342 
343 	smp_wmb();		/* make sure no writes get moved after unlock */
344 	xl->lock = 0;		/* release lock */
345 
346 	/*
347 	 * Make sure unlock happens before checking for waiting
348 	 * spinners.  We need a strong barrier to enforce the
349 	 * write-read ordering to different memory locations, as the
350 	 * CPU makes no implied guarantees about their ordering.
351 	 */
352 	mb();
353 
354 	if (unlikely(xl->spinners))
355 		xen_spin_unlock_slow(xl);
356 }
357 
358 static irqreturn_t dummy_handler(int irq, void *dev_id)
359 {
360 	BUG();
361 	return IRQ_HANDLED;
362 }
363 
364 void __cpuinit xen_init_lock_cpu(int cpu)
365 {
366 	int irq;
367 	char *name;
368 
369 	WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
370 	     cpu, per_cpu(lock_kicker_irq, cpu));
371 
372 	/*
373 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
374 	 * (xen: disable PV spinlocks on HVM)
375 	 */
376 	if (xen_hvm_domain())
377 		return;
378 
379 	name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
380 	irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
381 				     cpu,
382 				     dummy_handler,
383 				     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
384 				     name,
385 				     NULL);
386 
387 	if (irq >= 0) {
388 		disable_irq(irq); /* make sure it's never delivered */
389 		per_cpu(lock_kicker_irq, cpu) = irq;
390 		per_cpu(irq_name, cpu) = name;
391 	}
392 
393 	printk("cpu %d spinlock event irq %d\n", cpu, irq);
394 }
395 
396 void xen_uninit_lock_cpu(int cpu)
397 {
398 	/*
399 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
400 	 * (xen: disable PV spinlocks on HVM)
401 	 */
402 	if (xen_hvm_domain())
403 		return;
404 
405 	unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
406 	per_cpu(lock_kicker_irq, cpu) = -1;
407 	kfree(per_cpu(irq_name, cpu));
408 	per_cpu(irq_name, cpu) = NULL;
409 }
410 
411 void __init xen_init_spinlocks(void)
412 {
413 	/*
414 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
415 	 * (xen: disable PV spinlocks on HVM)
416 	 */
417 	if (xen_hvm_domain())
418 		return;
419 
420 	BUILD_BUG_ON(sizeof(struct xen_spinlock) > sizeof(arch_spinlock_t));
421 
422 	pv_lock_ops.spin_is_locked = xen_spin_is_locked;
423 	pv_lock_ops.spin_is_contended = xen_spin_is_contended;
424 	pv_lock_ops.spin_lock = xen_spin_lock;
425 	pv_lock_ops.spin_lock_flags = xen_spin_lock_flags;
426 	pv_lock_ops.spin_trylock = xen_spin_trylock;
427 	pv_lock_ops.spin_unlock = xen_spin_unlock;
428 }
429 
430 #ifdef CONFIG_XEN_DEBUG_FS
431 
432 static struct dentry *d_spin_debug;
433 
434 static int __init xen_spinlock_debugfs(void)
435 {
436 	struct dentry *d_xen = xen_init_debugfs();
437 
438 	if (d_xen == NULL)
439 		return -ENOMEM;
440 
441 	d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
442 
443 	debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
444 
445 	debugfs_create_u32("timeout", 0644, d_spin_debug, &lock_timeout);
446 
447 	debugfs_create_u64("taken", 0444, d_spin_debug, &spinlock_stats.taken);
448 	debugfs_create_u32("taken_slow", 0444, d_spin_debug,
449 			   &spinlock_stats.taken_slow);
450 	debugfs_create_u32("taken_slow_nested", 0444, d_spin_debug,
451 			   &spinlock_stats.taken_slow_nested);
452 	debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
453 			   &spinlock_stats.taken_slow_pickup);
454 	debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
455 			   &spinlock_stats.taken_slow_spurious);
456 	debugfs_create_u32("taken_slow_irqenable", 0444, d_spin_debug,
457 			   &spinlock_stats.taken_slow_irqenable);
458 
459 	debugfs_create_u64("released", 0444, d_spin_debug, &spinlock_stats.released);
460 	debugfs_create_u32("released_slow", 0444, d_spin_debug,
461 			   &spinlock_stats.released_slow);
462 	debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
463 			   &spinlock_stats.released_slow_kicked);
464 
465 	debugfs_create_u64("time_spinning", 0444, d_spin_debug,
466 			   &spinlock_stats.time_spinning);
467 	debugfs_create_u64("time_blocked", 0444, d_spin_debug,
468 			   &spinlock_stats.time_blocked);
469 	debugfs_create_u64("time_total", 0444, d_spin_debug,
470 			   &spinlock_stats.time_total);
471 
472 	debugfs_create_u32_array("histo_total", 0444, d_spin_debug,
473 				spinlock_stats.histo_spin_total, HISTO_BUCKETS + 1);
474 	debugfs_create_u32_array("histo_spinning", 0444, d_spin_debug,
475 				spinlock_stats.histo_spin_spinning, HISTO_BUCKETS + 1);
476 	debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
477 				spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
478 
479 	return 0;
480 }
481 fs_initcall(xen_spinlock_debugfs);
482 
483 #endif	/* CONFIG_XEN_DEBUG_FS */
484