Commit | Line | Data |
---|---|---|
d5de8841 JF |
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> | |
994025ca JF |
7 | #include <linux/debugfs.h> |
8 | #include <linux/log2.h> | |
5a0e3ad6 | 9 | #include <linux/gfp.h> |
354e7b76 | 10 | #include <linux/slab.h> |
d5de8841 JF |
11 | |
12 | #include <asm/paravirt.h> | |
13 | ||
14 | #include <xen/interface/xen.h> | |
15 | #include <xen/events.h> | |
16 | ||
17 | #include "xen-ops.h" | |
994025ca JF |
18 | #include "debugfs.h" |
19 | ||
e95e6f17 DV |
20 | static DEFINE_PER_CPU(int, lock_kicker_irq) = -1; |
21 | static DEFINE_PER_CPU(char *, irq_name); | |
22 | static bool xen_pvspin = true; | |
23 | ||
62c7a1e9 | 24 | #ifdef CONFIG_QUEUED_SPINLOCKS |
e95e6f17 DV |
25 | |
26 | #include <asm/qspinlock.h> | |
27 | ||
28 | static void xen_qlock_kick(int cpu) | |
29 | { | |
707e59ba RL |
30 | int irq = per_cpu(lock_kicker_irq, cpu); |
31 | ||
32 | /* Don't kick if the target's kicker interrupt is not initialized. */ | |
33 | if (irq == -1) | |
34 | return; | |
35 | ||
e95e6f17 DV |
36 | xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR); |
37 | } | |
38 | ||
39 | /* | |
40 | * Halt the current CPU & release it back to the host | |
41 | */ | |
42 | static void xen_qlock_wait(u8 *byte, u8 val) | |
43 | { | |
44 | int irq = __this_cpu_read(lock_kicker_irq); | |
45 | ||
46 | /* If kicker interrupts not initialized yet, just spin */ | |
47 | if (irq == -1) | |
48 | return; | |
49 | ||
50 | /* clear pending */ | |
51 | xen_clear_irq_pending(irq); | |
52 | barrier(); | |
53 | ||
54 | /* | |
55 | * We check the byte value after clearing pending IRQ to make sure | |
56 | * that we won't miss a wakeup event because of the clearing. | |
57 | * | |
58 | * The sync_clear_bit() call in xen_clear_irq_pending() is atomic. | |
59 | * So it is effectively a memory barrier for x86. | |
60 | */ | |
61 | if (READ_ONCE(*byte) != val) | |
62 | return; | |
63 | ||
64 | /* | |
65 | * If an interrupt happens here, it will leave the wakeup irq | |
66 | * pending, which will cause xen_poll_irq() to return | |
67 | * immediately. | |
68 | */ | |
69 | ||
70 | /* Block until irq becomes pending (or perhaps a spurious wakeup) */ | |
71 | xen_poll_irq(irq); | |
72 | } | |
73 | ||
62c7a1e9 | 74 | #else /* CONFIG_QUEUED_SPINLOCKS */ |
e95e6f17 | 75 | |
80bd58fe JF |
76 | enum xen_contention_stat { |
77 | TAKEN_SLOW, | |
78 | TAKEN_SLOW_PICKUP, | |
79 | TAKEN_SLOW_SPURIOUS, | |
80 | RELEASED_SLOW, | |
81 | RELEASED_SLOW_KICKED, | |
82 | NR_CONTENTION_STATS | |
83 | }; | |
994025ca | 84 | |
994025ca | 85 | |
80bd58fe | 86 | #ifdef CONFIG_XEN_DEBUG_FS |
f8eca41f | 87 | #define HISTO_BUCKETS 30 |
80bd58fe JF |
88 | static struct xen_spinlock_stats |
89 | { | |
90 | u32 contention_stats[NR_CONTENTION_STATS]; | |
f8eca41f | 91 | u32 histo_spin_blocked[HISTO_BUCKETS+1]; |
f8eca41f | 92 | u64 time_blocked; |
994025ca JF |
93 | } spinlock_stats; |
94 | ||
95 | static u8 zero_stats; | |
96 | ||
994025ca JF |
97 | static inline void check_zero(void) |
98 | { | |
80bd58fe | 99 | u8 ret; |
d6abfdb2 | 100 | u8 old = READ_ONCE(zero_stats); |
80bd58fe JF |
101 | if (unlikely(old)) { |
102 | ret = cmpxchg(&zero_stats, old, 0); | |
103 | /* This ensures only one fellow resets the stat */ | |
104 | if (ret == old) | |
105 | memset(&spinlock_stats, 0, sizeof(spinlock_stats)); | |
994025ca JF |
106 | } |
107 | } | |
108 | ||
80bd58fe JF |
109 | static inline void add_stats(enum xen_contention_stat var, u32 val) |
110 | { | |
111 | check_zero(); | |
112 | spinlock_stats.contention_stats[var] += val; | |
113 | } | |
994025ca JF |
114 | |
115 | static inline u64 spin_time_start(void) | |
116 | { | |
117 | return xen_clocksource_read(); | |
118 | } | |
119 | ||
120 | static void __spin_time_accum(u64 delta, u32 *array) | |
121 | { | |
122 | unsigned index = ilog2(delta); | |
123 | ||
124 | check_zero(); | |
125 | ||
126 | if (index < HISTO_BUCKETS) | |
127 | array[index]++; | |
128 | else | |
129 | array[HISTO_BUCKETS]++; | |
130 | } | |
131 | ||
f8eca41f | 132 | static inline void spin_time_accum_blocked(u64 start) |
994025ca JF |
133 | { |
134 | u32 delta = xen_clocksource_read() - start; | |
135 | ||
f8eca41f JF |
136 | __spin_time_accum(delta, spinlock_stats.histo_spin_blocked); |
137 | spinlock_stats.time_blocked += delta; | |
994025ca JF |
138 | } |
139 | #else /* !CONFIG_XEN_DEBUG_FS */ | |
80bd58fe JF |
140 | static inline void add_stats(enum xen_contention_stat var, u32 val) |
141 | { | |
142 | } | |
994025ca JF |
143 | |
144 | static inline u64 spin_time_start(void) | |
145 | { | |
146 | return 0; | |
147 | } | |
148 | ||
f8eca41f | 149 | static inline void spin_time_accum_blocked(u64 start) |
994025ca JF |
150 | { |
151 | } | |
152 | #endif /* CONFIG_XEN_DEBUG_FS */ | |
d5de8841 | 153 | |
80bd58fe JF |
154 | struct xen_lock_waiting { |
155 | struct arch_spinlock *lock; | |
156 | __ticket_t want; | |
d5de8841 JF |
157 | }; |
158 | ||
80bd58fe JF |
159 | static DEFINE_PER_CPU(struct xen_lock_waiting, lock_waiting); |
160 | static cpumask_t waiting_cpus; | |
d5de8841 | 161 | |
dd41f818 | 162 | __visible void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want) |
d5de8841 | 163 | { |
780f36d8 | 164 | int irq = __this_cpu_read(lock_kicker_irq); |
89cbc767 | 165 | struct xen_lock_waiting *w = this_cpu_ptr(&lock_waiting); |
80bd58fe | 166 | int cpu = smp_processor_id(); |
f8eca41f | 167 | u64 start; |
d6abfdb2 | 168 | __ticket_t head; |
80bd58fe | 169 | unsigned long flags; |
d5de8841 JF |
170 | |
171 | /* If kicker interrupts not initialized yet, just spin */ | |
172 | if (irq == -1) | |
80bd58fe | 173 | return; |
d5de8841 | 174 | |
f8eca41f JF |
175 | start = spin_time_start(); |
176 | ||
80bd58fe JF |
177 | /* |
178 | * Make sure an interrupt handler can't upset things in a | |
179 | * partially setup state. | |
180 | */ | |
181 | local_irq_save(flags); | |
1ed7bf5f JF |
182 | /* |
183 | * We don't really care if we're overwriting some other | |
184 | * (lock,want) pair, as that would mean that we're currently | |
185 | * in an interrupt context, and the outer context had | |
186 | * interrupts enabled. That has already kicked the VCPU out | |
187 | * of xen_poll_irq(), so it will just return spuriously and | |
188 | * retry with newly setup (lock,want). | |
189 | * | |
190 | * The ordering protocol on this is that the "lock" pointer | |
191 | * may only be set non-NULL if the "want" ticket is correct. | |
192 | * If we're updating "want", we must first clear "lock". | |
193 | */ | |
194 | w->lock = NULL; | |
195 | smp_wmb(); | |
80bd58fe JF |
196 | w->want = want; |
197 | smp_wmb(); | |
198 | w->lock = lock; | |
4d576b57 | 199 | |
80bd58fe JF |
200 | /* This uses set_bit, which atomic and therefore a barrier */ |
201 | cpumask_set_cpu(cpu, &waiting_cpus); | |
202 | add_stats(TAKEN_SLOW, 1); | |
4d576b57 | 203 | |
80bd58fe JF |
204 | /* clear pending */ |
205 | xen_clear_irq_pending(irq); | |
4d576b57 | 206 | |
80bd58fe JF |
207 | /* Only check lock once pending cleared */ |
208 | barrier(); | |
d5de8841 | 209 | |
1ed7bf5f JF |
210 | /* |
211 | * Mark entry to slowpath before doing the pickup test to make | |
212 | * sure we don't deadlock with an unlocker. | |
213 | */ | |
96f853ea JF |
214 | __ticket_enter_slowpath(lock); |
215 | ||
d6abfdb2 R |
216 | /* make sure enter_slowpath, which is atomic does not cross the read */ |
217 | smp_mb__after_atomic(); | |
218 | ||
1ed7bf5f JF |
219 | /* |
220 | * check again make sure it didn't become free while | |
221 | * we weren't looking | |
222 | */ | |
d6abfdb2 R |
223 | head = READ_ONCE(lock->tickets.head); |
224 | if (__tickets_equal(head, want)) { | |
80bd58fe JF |
225 | add_stats(TAKEN_SLOW_PICKUP, 1); |
226 | goto out; | |
227 | } | |
1ed7bf5f JF |
228 | |
229 | /* Allow interrupts while blocked */ | |
230 | local_irq_restore(flags); | |
231 | ||
232 | /* | |
233 | * If an interrupt happens here, it will leave the wakeup irq | |
234 | * pending, which will cause xen_poll_irq() to return | |
235 | * immediately. | |
236 | */ | |
237 | ||
80bd58fe JF |
238 | /* Block until irq becomes pending (or perhaps a spurious wakeup) */ |
239 | xen_poll_irq(irq); | |
240 | add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq)); | |
1ed7bf5f JF |
241 | |
242 | local_irq_save(flags); | |
243 | ||
770144ea | 244 | kstat_incr_irq_this_cpu(irq); |
d5de8841 | 245 | out: |
80bd58fe JF |
246 | cpumask_clear_cpu(cpu, &waiting_cpus); |
247 | w->lock = NULL; | |
1ed7bf5f | 248 | |
80bd58fe | 249 | local_irq_restore(flags); |
1ed7bf5f | 250 | |
f8eca41f | 251 | spin_time_accum_blocked(start); |
d5de8841 | 252 | } |
354714dd | 253 | PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning); |
d5de8841 | 254 | |
80bd58fe | 255 | static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next) |
d5de8841 JF |
256 | { |
257 | int cpu; | |
258 | ||
80bd58fe JF |
259 | add_stats(RELEASED_SLOW, 1); |
260 | ||
261 | for_each_cpu(cpu, &waiting_cpus) { | |
262 | const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu); | |
994025ca | 263 | |
1ed7bf5f | 264 | /* Make sure we read lock before want */ |
d6abfdb2 R |
265 | if (READ_ONCE(w->lock) == lock && |
266 | READ_ONCE(w->want) == next) { | |
80bd58fe | 267 | add_stats(RELEASED_SLOW_KICKED, 1); |
d5de8841 | 268 | xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR); |
80bd58fe | 269 | break; |
d5de8841 JF |
270 | } |
271 | } | |
272 | } | |
62c7a1e9 | 273 | #endif /* CONFIG_QUEUED_SPINLOCKS */ |
d5de8841 | 274 | |
d5de8841 JF |
275 | static irqreturn_t dummy_handler(int irq, void *dev_id) |
276 | { | |
277 | BUG(); | |
278 | return IRQ_HANDLED; | |
279 | } | |
280 | ||
148f9bb8 | 281 | void xen_init_lock_cpu(int cpu) |
d5de8841 JF |
282 | { |
283 | int irq; | |
354e7b76 | 284 | char *name; |
d5de8841 | 285 | |
3310bbed KRW |
286 | if (!xen_pvspin) |
287 | return; | |
288 | ||
cb91f8f4 | 289 | WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n", |
cb9c6f15 KRW |
290 | cpu, per_cpu(lock_kicker_irq, cpu)); |
291 | ||
d5de8841 JF |
292 | name = kasprintf(GFP_KERNEL, "spinlock%d", cpu); |
293 | irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR, | |
294 | cpu, | |
295 | dummy_handler, | |
9d71cee6 | 296 | IRQF_PERCPU|IRQF_NOBALANCING, |
d5de8841 JF |
297 | name, |
298 | NULL); | |
299 | ||
300 | if (irq >= 0) { | |
301 | disable_irq(irq); /* make sure it's never delivered */ | |
302 | per_cpu(lock_kicker_irq, cpu) = irq; | |
354e7b76 | 303 | per_cpu(irq_name, cpu) = name; |
d5de8841 JF |
304 | } |
305 | ||
306 | printk("cpu %d spinlock event irq %d\n", cpu, irq); | |
307 | } | |
308 | ||
d68d82af AN |
309 | void xen_uninit_lock_cpu(int cpu) |
310 | { | |
3310bbed KRW |
311 | if (!xen_pvspin) |
312 | return; | |
313 | ||
d68d82af | 314 | unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL); |
cb9c6f15 | 315 | per_cpu(lock_kicker_irq, cpu) = -1; |
354e7b76 KRW |
316 | kfree(per_cpu(irq_name, cpu)); |
317 | per_cpu(irq_name, cpu) = NULL; | |
d68d82af AN |
318 | } |
319 | ||
b8fa70b5 | 320 | |
a945928e KRW |
321 | /* |
322 | * Our init of PV spinlocks is split in two init functions due to us | |
323 | * using paravirt patching and jump labels patching and having to do | |
324 | * all of this before SMP code is invoked. | |
325 | * | |
326 | * The paravirt patching needs to be done _before_ the alternative asm code | |
327 | * is started, otherwise we would not patch the core kernel code. | |
328 | */ | |
d5de8841 JF |
329 | void __init xen_init_spinlocks(void) |
330 | { | |
70dd4998 | 331 | |
b8fa70b5 JF |
332 | if (!xen_pvspin) { |
333 | printk(KERN_DEBUG "xen: PV spinlocks disabled\n"); | |
334 | return; | |
335 | } | |
e0fc17a9 | 336 | printk(KERN_DEBUG "xen: PV spinlocks enabled\n"); |
62c7a1e9 | 337 | #ifdef CONFIG_QUEUED_SPINLOCKS |
e95e6f17 DV |
338 | __pv_init_lock_hash(); |
339 | pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath; | |
340 | pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock); | |
341 | pv_lock_ops.wait = xen_qlock_wait; | |
342 | pv_lock_ops.kick = xen_qlock_kick; | |
343 | #else | |
354714dd | 344 | pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning); |
80bd58fe | 345 | pv_lock_ops.unlock_kick = xen_unlock_kick; |
e95e6f17 | 346 | #endif |
d5de8841 | 347 | } |
994025ca | 348 | |
a945928e KRW |
349 | /* |
350 | * While the jump_label init code needs to happend _after_ the jump labels are | |
351 | * enabled and before SMP is started. Hence we use pre-SMP initcall level | |
352 | * init. We cannot do it in xen_init_spinlocks as that is done before | |
353 | * jump labels are activated. | |
354 | */ | |
355 | static __init int xen_init_spinlocks_jump(void) | |
356 | { | |
357 | if (!xen_pvspin) | |
358 | return 0; | |
359 | ||
e0fc17a9 KRW |
360 | if (!xen_domain()) |
361 | return 0; | |
362 | ||
a945928e KRW |
363 | static_key_slow_inc(¶virt_ticketlocks_enabled); |
364 | return 0; | |
365 | } | |
366 | early_initcall(xen_init_spinlocks_jump); | |
367 | ||
b8fa70b5 JF |
368 | static __init int xen_parse_nopvspin(char *arg) |
369 | { | |
370 | xen_pvspin = false; | |
371 | return 0; | |
372 | } | |
373 | early_param("xen_nopvspin", xen_parse_nopvspin); | |
374 | ||
62c7a1e9 | 375 | #if defined(CONFIG_XEN_DEBUG_FS) && !defined(CONFIG_QUEUED_SPINLOCKS) |
994025ca JF |
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 | ||
3310bbed KRW |
386 | if (!xen_pvspin) |
387 | return 0; | |
388 | ||
994025ca JF |
389 | d_spin_debug = debugfs_create_dir("spinlocks", d_xen); |
390 | ||
391 | debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats); | |
392 | ||
994025ca | 393 | debugfs_create_u32("taken_slow", 0444, d_spin_debug, |
80bd58fe | 394 | &spinlock_stats.contention_stats[TAKEN_SLOW]); |
994025ca | 395 | debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug, |
80bd58fe | 396 | &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]); |
994025ca | 397 | debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug, |
80bd58fe | 398 | &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]); |
994025ca | 399 | |
994025ca | 400 | debugfs_create_u32("released_slow", 0444, d_spin_debug, |
80bd58fe | 401 | &spinlock_stats.contention_stats[RELEASED_SLOW]); |
994025ca | 402 | debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug, |
80bd58fe | 403 | &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]); |
994025ca | 404 | |
f8eca41f JF |
405 | debugfs_create_u64("time_blocked", 0444, d_spin_debug, |
406 | &spinlock_stats.time_blocked); | |
994025ca | 407 | |
9fe2a701 SV |
408 | debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug, |
409 | spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1); | |
994025ca JF |
410 | |
411 | return 0; | |
412 | } | |
413 | fs_initcall(xen_spinlock_debugfs); | |
414 | ||
415 | #endif /* CONFIG_XEN_DEBUG_FS */ |