xref: /openbmc/linux/arch/x86/xen/spinlock.c (revision 39b6f3aa)
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 
11 #include <asm/paravirt.h>
12 
13 #include <xen/interface/xen.h>
14 #include <xen/events.h>
15 
16 #include "xen-ops.h"
17 #include "debugfs.h"
18 
19 #ifdef CONFIG_XEN_DEBUG_FS
20 static struct xen_spinlock_stats
21 {
22 	u64 taken;
23 	u32 taken_slow;
24 	u32 taken_slow_nested;
25 	u32 taken_slow_pickup;
26 	u32 taken_slow_spurious;
27 	u32 taken_slow_irqenable;
28 
29 	u64 released;
30 	u32 released_slow;
31 	u32 released_slow_kicked;
32 
33 #define HISTO_BUCKETS	30
34 	u32 histo_spin_total[HISTO_BUCKETS+1];
35 	u32 histo_spin_spinning[HISTO_BUCKETS+1];
36 	u32 histo_spin_blocked[HISTO_BUCKETS+1];
37 
38 	u64 time_total;
39 	u64 time_spinning;
40 	u64 time_blocked;
41 } spinlock_stats;
42 
43 static u8 zero_stats;
44 
45 static unsigned lock_timeout = 1 << 10;
46 #define TIMEOUT lock_timeout
47 
48 static inline void check_zero(void)
49 {
50 	if (unlikely(zero_stats)) {
51 		memset(&spinlock_stats, 0, sizeof(spinlock_stats));
52 		zero_stats = 0;
53 	}
54 }
55 
56 #define ADD_STATS(elem, val)			\
57 	do { check_zero(); spinlock_stats.elem += (val); } while(0)
58 
59 static inline u64 spin_time_start(void)
60 {
61 	return xen_clocksource_read();
62 }
63 
64 static void __spin_time_accum(u64 delta, u32 *array)
65 {
66 	unsigned index = ilog2(delta);
67 
68 	check_zero();
69 
70 	if (index < HISTO_BUCKETS)
71 		array[index]++;
72 	else
73 		array[HISTO_BUCKETS]++;
74 }
75 
76 static inline void spin_time_accum_spinning(u64 start)
77 {
78 	u32 delta = xen_clocksource_read() - start;
79 
80 	__spin_time_accum(delta, spinlock_stats.histo_spin_spinning);
81 	spinlock_stats.time_spinning += delta;
82 }
83 
84 static inline void spin_time_accum_total(u64 start)
85 {
86 	u32 delta = xen_clocksource_read() - start;
87 
88 	__spin_time_accum(delta, spinlock_stats.histo_spin_total);
89 	spinlock_stats.time_total += delta;
90 }
91 
92 static inline void spin_time_accum_blocked(u64 start)
93 {
94 	u32 delta = xen_clocksource_read() - start;
95 
96 	__spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
97 	spinlock_stats.time_blocked += delta;
98 }
99 #else  /* !CONFIG_XEN_DEBUG_FS */
100 #define TIMEOUT			(1 << 10)
101 #define ADD_STATS(elem, val)	do { (void)(val); } while(0)
102 
103 static inline u64 spin_time_start(void)
104 {
105 	return 0;
106 }
107 
108 static inline void spin_time_accum_total(u64 start)
109 {
110 }
111 static inline void spin_time_accum_spinning(u64 start)
112 {
113 }
114 static inline void spin_time_accum_blocked(u64 start)
115 {
116 }
117 #endif  /* CONFIG_XEN_DEBUG_FS */
118 
119 /*
120  * Size struct xen_spinlock so it's the same as arch_spinlock_t.
121  */
122 #if NR_CPUS < 256
123 typedef u8 xen_spinners_t;
124 # define inc_spinners(xl) \
125 	asm(LOCK_PREFIX " incb %0" : "+m" ((xl)->spinners) : : "memory");
126 # define dec_spinners(xl) \
127 	asm(LOCK_PREFIX " decb %0" : "+m" ((xl)->spinners) : : "memory");
128 #else
129 typedef u16 xen_spinners_t;
130 # define inc_spinners(xl) \
131 	asm(LOCK_PREFIX " incw %0" : "+m" ((xl)->spinners) : : "memory");
132 # define dec_spinners(xl) \
133 	asm(LOCK_PREFIX " decw %0" : "+m" ((xl)->spinners) : : "memory");
134 #endif
135 
136 struct xen_spinlock {
137 	unsigned char lock;		/* 0 -> free; 1 -> locked */
138 	xen_spinners_t spinners;	/* count of waiting cpus */
139 };
140 
141 static int xen_spin_is_locked(struct arch_spinlock *lock)
142 {
143 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
144 
145 	return xl->lock != 0;
146 }
147 
148 static int xen_spin_is_contended(struct arch_spinlock *lock)
149 {
150 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
151 
152 	/* Not strictly true; this is only the count of contended
153 	   lock-takers entering the slow path. */
154 	return xl->spinners != 0;
155 }
156 
157 static int xen_spin_trylock(struct arch_spinlock *lock)
158 {
159 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
160 	u8 old = 1;
161 
162 	asm("xchgb %b0,%1"
163 	    : "+q" (old), "+m" (xl->lock) : : "memory");
164 
165 	return old == 0;
166 }
167 
168 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
169 static DEFINE_PER_CPU(struct xen_spinlock *, lock_spinners);
170 
171 /*
172  * Mark a cpu as interested in a lock.  Returns the CPU's previous
173  * lock of interest, in case we got preempted by an interrupt.
174  */
175 static inline struct xen_spinlock *spinning_lock(struct xen_spinlock *xl)
176 {
177 	struct xen_spinlock *prev;
178 
179 	prev = __this_cpu_read(lock_spinners);
180 	__this_cpu_write(lock_spinners, xl);
181 
182 	wmb();			/* set lock of interest before count */
183 
184 	inc_spinners(xl);
185 
186 	return prev;
187 }
188 
189 /*
190  * Mark a cpu as no longer interested in a lock.  Restores previous
191  * lock of interest (NULL for none).
192  */
193 static inline void unspinning_lock(struct xen_spinlock *xl, struct xen_spinlock *prev)
194 {
195 	dec_spinners(xl);
196 	wmb();			/* decrement count before restoring lock */
197 	__this_cpu_write(lock_spinners, prev);
198 }
199 
200 static noinline int xen_spin_lock_slow(struct arch_spinlock *lock, bool irq_enable)
201 {
202 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
203 	struct xen_spinlock *prev;
204 	int irq = __this_cpu_read(lock_kicker_irq);
205 	int ret;
206 	u64 start;
207 
208 	/* If kicker interrupts not initialized yet, just spin */
209 	if (irq == -1)
210 		return 0;
211 
212 	start = spin_time_start();
213 
214 	/* announce we're spinning */
215 	prev = spinning_lock(xl);
216 
217 	ADD_STATS(taken_slow, 1);
218 	ADD_STATS(taken_slow_nested, prev != NULL);
219 
220 	do {
221 		unsigned long flags;
222 
223 		/* clear pending */
224 		xen_clear_irq_pending(irq);
225 
226 		/* check again make sure it didn't become free while
227 		   we weren't looking  */
228 		ret = xen_spin_trylock(lock);
229 		if (ret) {
230 			ADD_STATS(taken_slow_pickup, 1);
231 
232 			/*
233 			 * If we interrupted another spinlock while it
234 			 * was blocking, make sure it doesn't block
235 			 * without rechecking the lock.
236 			 */
237 			if (prev != NULL)
238 				xen_set_irq_pending(irq);
239 			goto out;
240 		}
241 
242 		flags = arch_local_save_flags();
243 		if (irq_enable) {
244 			ADD_STATS(taken_slow_irqenable, 1);
245 			raw_local_irq_enable();
246 		}
247 
248 		/*
249 		 * Block until irq becomes pending.  If we're
250 		 * interrupted at this point (after the trylock but
251 		 * before entering the block), then the nested lock
252 		 * handler guarantees that the irq will be left
253 		 * pending if there's any chance the lock became free;
254 		 * xen_poll_irq() returns immediately if the irq is
255 		 * pending.
256 		 */
257 		xen_poll_irq(irq);
258 
259 		raw_local_irq_restore(flags);
260 
261 		ADD_STATS(taken_slow_spurious, !xen_test_irq_pending(irq));
262 	} while (!xen_test_irq_pending(irq)); /* check for spurious wakeups */
263 
264 	kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
265 
266 out:
267 	unspinning_lock(xl, prev);
268 	spin_time_accum_blocked(start);
269 
270 	return ret;
271 }
272 
273 static inline void __xen_spin_lock(struct arch_spinlock *lock, bool irq_enable)
274 {
275 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
276 	unsigned timeout;
277 	u8 oldval;
278 	u64 start_spin;
279 
280 	ADD_STATS(taken, 1);
281 
282 	start_spin = spin_time_start();
283 
284 	do {
285 		u64 start_spin_fast = spin_time_start();
286 
287 		timeout = TIMEOUT;
288 
289 		asm("1: xchgb %1,%0\n"
290 		    "   testb %1,%1\n"
291 		    "   jz 3f\n"
292 		    "2: rep;nop\n"
293 		    "   cmpb $0,%0\n"
294 		    "   je 1b\n"
295 		    "   dec %2\n"
296 		    "   jnz 2b\n"
297 		    "3:\n"
298 		    : "+m" (xl->lock), "=q" (oldval), "+r" (timeout)
299 		    : "1" (1)
300 		    : "memory");
301 
302 		spin_time_accum_spinning(start_spin_fast);
303 
304 	} while (unlikely(oldval != 0 &&
305 			  (TIMEOUT == ~0 || !xen_spin_lock_slow(lock, irq_enable))));
306 
307 	spin_time_accum_total(start_spin);
308 }
309 
310 static void xen_spin_lock(struct arch_spinlock *lock)
311 {
312 	__xen_spin_lock(lock, false);
313 }
314 
315 static void xen_spin_lock_flags(struct arch_spinlock *lock, unsigned long flags)
316 {
317 	__xen_spin_lock(lock, !raw_irqs_disabled_flags(flags));
318 }
319 
320 static noinline void xen_spin_unlock_slow(struct xen_spinlock *xl)
321 {
322 	int cpu;
323 
324 	ADD_STATS(released_slow, 1);
325 
326 	for_each_online_cpu(cpu) {
327 		/* XXX should mix up next cpu selection */
328 		if (per_cpu(lock_spinners, cpu) == xl) {
329 			ADD_STATS(released_slow_kicked, 1);
330 			xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
331 		}
332 	}
333 }
334 
335 static void xen_spin_unlock(struct arch_spinlock *lock)
336 {
337 	struct xen_spinlock *xl = (struct xen_spinlock *)lock;
338 
339 	ADD_STATS(released, 1);
340 
341 	smp_wmb();		/* make sure no writes get moved after unlock */
342 	xl->lock = 0;		/* release lock */
343 
344 	/*
345 	 * Make sure unlock happens before checking for waiting
346 	 * spinners.  We need a strong barrier to enforce the
347 	 * write-read ordering to different memory locations, as the
348 	 * CPU makes no implied guarantees about their ordering.
349 	 */
350 	mb();
351 
352 	if (unlikely(xl->spinners))
353 		xen_spin_unlock_slow(xl);
354 }
355 
356 static irqreturn_t dummy_handler(int irq, void *dev_id)
357 {
358 	BUG();
359 	return IRQ_HANDLED;
360 }
361 
362 void __cpuinit xen_init_lock_cpu(int cpu)
363 {
364 	int irq;
365 	const char *name;
366 
367 	WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
368 	     cpu, per_cpu(lock_kicker_irq, cpu));
369 
370 	/*
371 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
372 	 * (xen: disable PV spinlocks on HVM)
373 	 */
374 	if (xen_hvm_domain())
375 		return;
376 
377 	name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
378 	irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
379 				     cpu,
380 				     dummy_handler,
381 				     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
382 				     name,
383 				     NULL);
384 
385 	if (irq >= 0) {
386 		disable_irq(irq); /* make sure it's never delivered */
387 		per_cpu(lock_kicker_irq, cpu) = irq;
388 	}
389 
390 	printk("cpu %d spinlock event irq %d\n", cpu, irq);
391 }
392 
393 void xen_uninit_lock_cpu(int cpu)
394 {
395 	/*
396 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
397 	 * (xen: disable PV spinlocks on HVM)
398 	 */
399 	if (xen_hvm_domain())
400 		return;
401 
402 	unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
403 	per_cpu(lock_kicker_irq, cpu) = -1;
404 }
405 
406 void __init xen_init_spinlocks(void)
407 {
408 	/*
409 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
410 	 * (xen: disable PV spinlocks on HVM)
411 	 */
412 	if (xen_hvm_domain())
413 		return;
414 
415 	BUILD_BUG_ON(sizeof(struct xen_spinlock) > sizeof(arch_spinlock_t));
416 
417 	pv_lock_ops.spin_is_locked = xen_spin_is_locked;
418 	pv_lock_ops.spin_is_contended = xen_spin_is_contended;
419 	pv_lock_ops.spin_lock = xen_spin_lock;
420 	pv_lock_ops.spin_lock_flags = xen_spin_lock_flags;
421 	pv_lock_ops.spin_trylock = xen_spin_trylock;
422 	pv_lock_ops.spin_unlock = xen_spin_unlock;
423 }
424 
425 #ifdef CONFIG_XEN_DEBUG_FS
426 
427 static struct dentry *d_spin_debug;
428 
429 static int __init xen_spinlock_debugfs(void)
430 {
431 	struct dentry *d_xen = xen_init_debugfs();
432 
433 	if (d_xen == NULL)
434 		return -ENOMEM;
435 
436 	d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
437 
438 	debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
439 
440 	debugfs_create_u32("timeout", 0644, d_spin_debug, &lock_timeout);
441 
442 	debugfs_create_u64("taken", 0444, d_spin_debug, &spinlock_stats.taken);
443 	debugfs_create_u32("taken_slow", 0444, d_spin_debug,
444 			   &spinlock_stats.taken_slow);
445 	debugfs_create_u32("taken_slow_nested", 0444, d_spin_debug,
446 			   &spinlock_stats.taken_slow_nested);
447 	debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
448 			   &spinlock_stats.taken_slow_pickup);
449 	debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
450 			   &spinlock_stats.taken_slow_spurious);
451 	debugfs_create_u32("taken_slow_irqenable", 0444, d_spin_debug,
452 			   &spinlock_stats.taken_slow_irqenable);
453 
454 	debugfs_create_u64("released", 0444, d_spin_debug, &spinlock_stats.released);
455 	debugfs_create_u32("released_slow", 0444, d_spin_debug,
456 			   &spinlock_stats.released_slow);
457 	debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
458 			   &spinlock_stats.released_slow_kicked);
459 
460 	debugfs_create_u64("time_spinning", 0444, d_spin_debug,
461 			   &spinlock_stats.time_spinning);
462 	debugfs_create_u64("time_blocked", 0444, d_spin_debug,
463 			   &spinlock_stats.time_blocked);
464 	debugfs_create_u64("time_total", 0444, d_spin_debug,
465 			   &spinlock_stats.time_total);
466 
467 	debugfs_create_u32_array("histo_total", 0444, d_spin_debug,
468 				spinlock_stats.histo_spin_total, HISTO_BUCKETS + 1);
469 	debugfs_create_u32_array("histo_spinning", 0444, d_spin_debug,
470 				spinlock_stats.histo_spin_spinning, HISTO_BUCKETS + 1);
471 	debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
472 				spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
473 
474 	return 0;
475 }
476 fs_initcall(xen_spinlock_debugfs);
477 
478 #endif	/* CONFIG_XEN_DEBUG_FS */
479