xref: /linux/arch/x86/xen/spinlock.c (revision 4d7696f1b05f4aeb586c74868fe3da2731daca4b)
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 enum xen_contention_stat {
21 	TAKEN_SLOW,
22 	TAKEN_SLOW_PICKUP,
23 	TAKEN_SLOW_SPURIOUS,
24 	RELEASED_SLOW,
25 	RELEASED_SLOW_KICKED,
26 	NR_CONTENTION_STATS
27 };
28 
29 
30 #ifdef CONFIG_XEN_DEBUG_FS
31 #define HISTO_BUCKETS	30
32 static struct xen_spinlock_stats
33 {
34 	u32 contention_stats[NR_CONTENTION_STATS];
35 	u32 histo_spin_blocked[HISTO_BUCKETS+1];
36 	u64 time_blocked;
37 } spinlock_stats;
38 
39 static u8 zero_stats;
40 
41 static inline void check_zero(void)
42 {
43 	u8 ret;
44 	u8 old = ACCESS_ONCE(zero_stats);
45 	if (unlikely(old)) {
46 		ret = cmpxchg(&zero_stats, old, 0);
47 		/* This ensures only one fellow resets the stat */
48 		if (ret == old)
49 			memset(&spinlock_stats, 0, sizeof(spinlock_stats));
50 	}
51 }
52 
53 static inline void add_stats(enum xen_contention_stat var, u32 val)
54 {
55 	check_zero();
56 	spinlock_stats.contention_stats[var] += val;
57 }
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_blocked(u64 start)
77 {
78 	u32 delta = xen_clocksource_read() - start;
79 
80 	__spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
81 	spinlock_stats.time_blocked += delta;
82 }
83 #else  /* !CONFIG_XEN_DEBUG_FS */
84 #define TIMEOUT			(1 << 10)
85 static inline void add_stats(enum xen_contention_stat var, u32 val)
86 {
87 }
88 
89 static inline u64 spin_time_start(void)
90 {
91 	return 0;
92 }
93 
94 static inline void spin_time_accum_blocked(u64 start)
95 {
96 }
97 #endif  /* CONFIG_XEN_DEBUG_FS */
98 
99 /*
100  * Size struct xen_spinlock so it's the same as arch_spinlock_t.
101  */
102 #if NR_CPUS < 256
103 typedef u8 xen_spinners_t;
104 # define inc_spinners(xl) \
105 	asm(LOCK_PREFIX " incb %0" : "+m" ((xl)->spinners) : : "memory");
106 # define dec_spinners(xl) \
107 	asm(LOCK_PREFIX " decb %0" : "+m" ((xl)->spinners) : : "memory");
108 #else
109 typedef u16 xen_spinners_t;
110 # define inc_spinners(xl) \
111 	asm(LOCK_PREFIX " incw %0" : "+m" ((xl)->spinners) : : "memory");
112 # define dec_spinners(xl) \
113 	asm(LOCK_PREFIX " decw %0" : "+m" ((xl)->spinners) : : "memory");
114 #endif
115 
116 struct xen_lock_waiting {
117 	struct arch_spinlock *lock;
118 	__ticket_t want;
119 };
120 
121 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
122 static DEFINE_PER_CPU(char *, irq_name);
123 static DEFINE_PER_CPU(struct xen_lock_waiting, lock_waiting);
124 static cpumask_t waiting_cpus;
125 
126 static void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
127 {
128 	int irq = __this_cpu_read(lock_kicker_irq);
129 	struct xen_lock_waiting *w = &__get_cpu_var(lock_waiting);
130 	int cpu = smp_processor_id();
131 	u64 start;
132 	unsigned long flags;
133 
134 	/* If kicker interrupts not initialized yet, just spin */
135 	if (irq == -1)
136 		return;
137 
138 	start = spin_time_start();
139 
140 	/*
141 	 * Make sure an interrupt handler can't upset things in a
142 	 * partially setup state.
143 	 */
144 	local_irq_save(flags);
145 	/*
146 	 * We don't really care if we're overwriting some other
147 	 * (lock,want) pair, as that would mean that we're currently
148 	 * in an interrupt context, and the outer context had
149 	 * interrupts enabled.  That has already kicked the VCPU out
150 	 * of xen_poll_irq(), so it will just return spuriously and
151 	 * retry with newly setup (lock,want).
152 	 *
153 	 * The ordering protocol on this is that the "lock" pointer
154 	 * may only be set non-NULL if the "want" ticket is correct.
155 	 * If we're updating "want", we must first clear "lock".
156 	 */
157 	w->lock = NULL;
158 	smp_wmb();
159 	w->want = want;
160 	smp_wmb();
161 	w->lock = lock;
162 
163 	/* This uses set_bit, which atomic and therefore a barrier */
164 	cpumask_set_cpu(cpu, &waiting_cpus);
165 	add_stats(TAKEN_SLOW, 1);
166 
167 	/* clear pending */
168 	xen_clear_irq_pending(irq);
169 
170 	/* Only check lock once pending cleared */
171 	barrier();
172 
173 	/*
174 	 * Mark entry to slowpath before doing the pickup test to make
175 	 * sure we don't deadlock with an unlocker.
176 	 */
177 	__ticket_enter_slowpath(lock);
178 
179 	/*
180 	 * check again make sure it didn't become free while
181 	 * we weren't looking
182 	 */
183 	if (ACCESS_ONCE(lock->tickets.head) == want) {
184 		add_stats(TAKEN_SLOW_PICKUP, 1);
185 		goto out;
186 	}
187 
188 	/* Allow interrupts while blocked */
189 	local_irq_restore(flags);
190 
191 	/*
192 	 * If an interrupt happens here, it will leave the wakeup irq
193 	 * pending, which will cause xen_poll_irq() to return
194 	 * immediately.
195 	 */
196 
197 	/* Block until irq becomes pending (or perhaps a spurious wakeup) */
198 	xen_poll_irq(irq);
199 	add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq));
200 
201 	local_irq_save(flags);
202 
203 	kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
204 out:
205 	cpumask_clear_cpu(cpu, &waiting_cpus);
206 	w->lock = NULL;
207 
208 	local_irq_restore(flags);
209 
210 	spin_time_accum_blocked(start);
211 }
212 PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning);
213 
214 static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next)
215 {
216 	int cpu;
217 
218 	add_stats(RELEASED_SLOW, 1);
219 
220 	for_each_cpu(cpu, &waiting_cpus) {
221 		const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu);
222 
223 		/* Make sure we read lock before want */
224 		if (ACCESS_ONCE(w->lock) == lock &&
225 		    ACCESS_ONCE(w->want) == next) {
226 			add_stats(RELEASED_SLOW_KICKED, 1);
227 			xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
228 			break;
229 		}
230 	}
231 }
232 
233 static irqreturn_t dummy_handler(int irq, void *dev_id)
234 {
235 	BUG();
236 	return IRQ_HANDLED;
237 }
238 
239 void xen_init_lock_cpu(int cpu)
240 {
241 	int irq;
242 	char *name;
243 
244 	WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
245 	     cpu, per_cpu(lock_kicker_irq, cpu));
246 
247 	/*
248 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
249 	 * (xen: disable PV spinlocks on HVM)
250 	 */
251 	if (xen_hvm_domain())
252 		return;
253 
254 	name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
255 	irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
256 				     cpu,
257 				     dummy_handler,
258 				     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
259 				     name,
260 				     NULL);
261 
262 	if (irq >= 0) {
263 		disable_irq(irq); /* make sure it's never delivered */
264 		per_cpu(lock_kicker_irq, cpu) = irq;
265 		per_cpu(irq_name, cpu) = name;
266 	}
267 
268 	printk("cpu %d spinlock event irq %d\n", cpu, irq);
269 }
270 
271 void xen_uninit_lock_cpu(int cpu)
272 {
273 	/*
274 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
275 	 * (xen: disable PV spinlocks on HVM)
276 	 */
277 	if (xen_hvm_domain())
278 		return;
279 
280 	unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
281 	per_cpu(lock_kicker_irq, cpu) = -1;
282 	kfree(per_cpu(irq_name, cpu));
283 	per_cpu(irq_name, cpu) = NULL;
284 }
285 
286 static bool xen_pvspin __initdata = true;
287 
288 void __init xen_init_spinlocks(void)
289 {
290 	/*
291 	 * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
292 	 * (xen: disable PV spinlocks on HVM)
293 	 */
294 	if (xen_hvm_domain())
295 		return;
296 
297 	if (!xen_pvspin) {
298 		printk(KERN_DEBUG "xen: PV spinlocks disabled\n");
299 		return;
300 	}
301 
302 	static_key_slow_inc(&paravirt_ticketlocks_enabled);
303 
304 	pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
305 	pv_lock_ops.unlock_kick = xen_unlock_kick;
306 }
307 
308 static __init int xen_parse_nopvspin(char *arg)
309 {
310 	xen_pvspin = false;
311 	return 0;
312 }
313 early_param("xen_nopvspin", xen_parse_nopvspin);
314 
315 #ifdef CONFIG_XEN_DEBUG_FS
316 
317 static struct dentry *d_spin_debug;
318 
319 static int __init xen_spinlock_debugfs(void)
320 {
321 	struct dentry *d_xen = xen_init_debugfs();
322 
323 	if (d_xen == NULL)
324 		return -ENOMEM;
325 
326 	d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
327 
328 	debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
329 
330 	debugfs_create_u32("taken_slow", 0444, d_spin_debug,
331 			   &spinlock_stats.contention_stats[TAKEN_SLOW]);
332 	debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
333 			   &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
334 	debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
335 			   &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]);
336 
337 	debugfs_create_u32("released_slow", 0444, d_spin_debug,
338 			   &spinlock_stats.contention_stats[RELEASED_SLOW]);
339 	debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
340 			   &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
341 
342 	debugfs_create_u64("time_blocked", 0444, d_spin_debug,
343 			   &spinlock_stats.time_blocked);
344 
345 	debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
346 				spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
347 
348 	return 0;
349 }
350 fs_initcall(xen_spinlock_debugfs);
351 
352 #endif	/* CONFIG_XEN_DEBUG_FS */
353