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 | ||
80bd58fe JF |
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 | }; | |
994025ca | 28 | |
994025ca | 29 | |
80bd58fe | 30 | #ifdef CONFIG_XEN_DEBUG_FS |
f8eca41f | 31 | #define HISTO_BUCKETS 30 |
80bd58fe JF |
32 | static struct xen_spinlock_stats |
33 | { | |
34 | u32 contention_stats[NR_CONTENTION_STATS]; | |
f8eca41f | 35 | u32 histo_spin_blocked[HISTO_BUCKETS+1]; |
f8eca41f | 36 | u64 time_blocked; |
994025ca JF |
37 | } spinlock_stats; |
38 | ||
39 | static u8 zero_stats; | |
40 | ||
994025ca JF |
41 | static inline void check_zero(void) |
42 | { | |
80bd58fe | 43 | u8 ret; |
d6abfdb2 | 44 | u8 old = READ_ONCE(zero_stats); |
80bd58fe JF |
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)); | |
994025ca JF |
50 | } |
51 | } | |
52 | ||
80bd58fe JF |
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 | } | |
994025ca JF |
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 | ||
f8eca41f | 76 | static inline void spin_time_accum_blocked(u64 start) |
994025ca JF |
77 | { |
78 | u32 delta = xen_clocksource_read() - start; | |
79 | ||
f8eca41f JF |
80 | __spin_time_accum(delta, spinlock_stats.histo_spin_blocked); |
81 | spinlock_stats.time_blocked += delta; | |
994025ca JF |
82 | } |
83 | #else /* !CONFIG_XEN_DEBUG_FS */ | |
80bd58fe JF |
84 | static inline void add_stats(enum xen_contention_stat var, u32 val) |
85 | { | |
86 | } | |
994025ca JF |
87 | |
88 | static inline u64 spin_time_start(void) | |
89 | { | |
90 | return 0; | |
91 | } | |
92 | ||
f8eca41f | 93 | static inline void spin_time_accum_blocked(u64 start) |
994025ca JF |
94 | { |
95 | } | |
96 | #endif /* CONFIG_XEN_DEBUG_FS */ | |
d5de8841 | 97 | |
80bd58fe JF |
98 | struct xen_lock_waiting { |
99 | struct arch_spinlock *lock; | |
100 | __ticket_t want; | |
d5de8841 JF |
101 | }; |
102 | ||
545ac138 | 103 | static DEFINE_PER_CPU(int, lock_kicker_irq) = -1; |
354e7b76 | 104 | static DEFINE_PER_CPU(char *, irq_name); |
80bd58fe JF |
105 | static DEFINE_PER_CPU(struct xen_lock_waiting, lock_waiting); |
106 | static cpumask_t waiting_cpus; | |
d5de8841 | 107 | |
c3b7cb1f | 108 | static bool xen_pvspin = true; |
dd41f818 | 109 | __visible void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want) |
d5de8841 | 110 | { |
780f36d8 | 111 | int irq = __this_cpu_read(lock_kicker_irq); |
89cbc767 | 112 | struct xen_lock_waiting *w = this_cpu_ptr(&lock_waiting); |
80bd58fe | 113 | int cpu = smp_processor_id(); |
f8eca41f | 114 | u64 start; |
d6abfdb2 | 115 | __ticket_t head; |
80bd58fe | 116 | unsigned long flags; |
d5de8841 JF |
117 | |
118 | /* If kicker interrupts not initialized yet, just spin */ | |
119 | if (irq == -1) | |
80bd58fe | 120 | return; |
d5de8841 | 121 | |
f8eca41f JF |
122 | start = spin_time_start(); |
123 | ||
80bd58fe JF |
124 | /* |
125 | * Make sure an interrupt handler can't upset things in a | |
126 | * partially setup state. | |
127 | */ | |
128 | local_irq_save(flags); | |
1ed7bf5f JF |
129 | /* |
130 | * We don't really care if we're overwriting some other | |
131 | * (lock,want) pair, as that would mean that we're currently | |
132 | * in an interrupt context, and the outer context had | |
133 | * interrupts enabled. That has already kicked the VCPU out | |
134 | * of xen_poll_irq(), so it will just return spuriously and | |
135 | * retry with newly setup (lock,want). | |
136 | * | |
137 | * The ordering protocol on this is that the "lock" pointer | |
138 | * may only be set non-NULL if the "want" ticket is correct. | |
139 | * If we're updating "want", we must first clear "lock". | |
140 | */ | |
141 | w->lock = NULL; | |
142 | smp_wmb(); | |
80bd58fe JF |
143 | w->want = want; |
144 | smp_wmb(); | |
145 | w->lock = lock; | |
4d576b57 | 146 | |
80bd58fe JF |
147 | /* This uses set_bit, which atomic and therefore a barrier */ |
148 | cpumask_set_cpu(cpu, &waiting_cpus); | |
149 | add_stats(TAKEN_SLOW, 1); | |
4d576b57 | 150 | |
80bd58fe JF |
151 | /* clear pending */ |
152 | xen_clear_irq_pending(irq); | |
4d576b57 | 153 | |
80bd58fe JF |
154 | /* Only check lock once pending cleared */ |
155 | barrier(); | |
d5de8841 | 156 | |
1ed7bf5f JF |
157 | /* |
158 | * Mark entry to slowpath before doing the pickup test to make | |
159 | * sure we don't deadlock with an unlocker. | |
160 | */ | |
96f853ea JF |
161 | __ticket_enter_slowpath(lock); |
162 | ||
d6abfdb2 R |
163 | /* make sure enter_slowpath, which is atomic does not cross the read */ |
164 | smp_mb__after_atomic(); | |
165 | ||
1ed7bf5f JF |
166 | /* |
167 | * check again make sure it didn't become free while | |
168 | * we weren't looking | |
169 | */ | |
d6abfdb2 R |
170 | head = READ_ONCE(lock->tickets.head); |
171 | if (__tickets_equal(head, want)) { | |
80bd58fe JF |
172 | add_stats(TAKEN_SLOW_PICKUP, 1); |
173 | goto out; | |
174 | } | |
1ed7bf5f JF |
175 | |
176 | /* Allow interrupts while blocked */ | |
177 | local_irq_restore(flags); | |
178 | ||
179 | /* | |
180 | * If an interrupt happens here, it will leave the wakeup irq | |
181 | * pending, which will cause xen_poll_irq() to return | |
182 | * immediately. | |
183 | */ | |
184 | ||
80bd58fe JF |
185 | /* Block until irq becomes pending (or perhaps a spurious wakeup) */ |
186 | xen_poll_irq(irq); | |
187 | add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq)); | |
1ed7bf5f JF |
188 | |
189 | local_irq_save(flags); | |
190 | ||
770144ea | 191 | kstat_incr_irq_this_cpu(irq); |
d5de8841 | 192 | out: |
80bd58fe JF |
193 | cpumask_clear_cpu(cpu, &waiting_cpus); |
194 | w->lock = NULL; | |
1ed7bf5f | 195 | |
80bd58fe | 196 | local_irq_restore(flags); |
1ed7bf5f | 197 | |
f8eca41f | 198 | spin_time_accum_blocked(start); |
d5de8841 | 199 | } |
354714dd | 200 | PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning); |
d5de8841 | 201 | |
80bd58fe | 202 | static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next) |
d5de8841 JF |
203 | { |
204 | int cpu; | |
205 | ||
80bd58fe JF |
206 | add_stats(RELEASED_SLOW, 1); |
207 | ||
208 | for_each_cpu(cpu, &waiting_cpus) { | |
209 | const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu); | |
994025ca | 210 | |
1ed7bf5f | 211 | /* Make sure we read lock before want */ |
d6abfdb2 R |
212 | if (READ_ONCE(w->lock) == lock && |
213 | READ_ONCE(w->want) == next) { | |
80bd58fe | 214 | add_stats(RELEASED_SLOW_KICKED, 1); |
d5de8841 | 215 | xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR); |
80bd58fe | 216 | break; |
d5de8841 JF |
217 | } |
218 | } | |
219 | } | |
220 | ||
d5de8841 JF |
221 | static irqreturn_t dummy_handler(int irq, void *dev_id) |
222 | { | |
223 | BUG(); | |
224 | return IRQ_HANDLED; | |
225 | } | |
226 | ||
148f9bb8 | 227 | void xen_init_lock_cpu(int cpu) |
d5de8841 JF |
228 | { |
229 | int irq; | |
354e7b76 | 230 | char *name; |
d5de8841 | 231 | |
3310bbed KRW |
232 | if (!xen_pvspin) |
233 | return; | |
234 | ||
cb91f8f4 | 235 | WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n", |
cb9c6f15 KRW |
236 | cpu, per_cpu(lock_kicker_irq, cpu)); |
237 | ||
d5de8841 JF |
238 | name = kasprintf(GFP_KERNEL, "spinlock%d", cpu); |
239 | irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR, | |
240 | cpu, | |
241 | dummy_handler, | |
9d71cee6 | 242 | IRQF_PERCPU|IRQF_NOBALANCING, |
d5de8841 JF |
243 | name, |
244 | NULL); | |
245 | ||
246 | if (irq >= 0) { | |
247 | disable_irq(irq); /* make sure it's never delivered */ | |
248 | per_cpu(lock_kicker_irq, cpu) = irq; | |
354e7b76 | 249 | per_cpu(irq_name, cpu) = name; |
d5de8841 JF |
250 | } |
251 | ||
252 | printk("cpu %d spinlock event irq %d\n", cpu, irq); | |
253 | } | |
254 | ||
d68d82af AN |
255 | void xen_uninit_lock_cpu(int cpu) |
256 | { | |
3310bbed KRW |
257 | if (!xen_pvspin) |
258 | return; | |
259 | ||
d68d82af | 260 | unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL); |
cb9c6f15 | 261 | per_cpu(lock_kicker_irq, cpu) = -1; |
354e7b76 KRW |
262 | kfree(per_cpu(irq_name, cpu)); |
263 | per_cpu(irq_name, cpu) = NULL; | |
d68d82af AN |
264 | } |
265 | ||
b8fa70b5 | 266 | |
a945928e KRW |
267 | /* |
268 | * Our init of PV spinlocks is split in two init functions due to us | |
269 | * using paravirt patching and jump labels patching and having to do | |
270 | * all of this before SMP code is invoked. | |
271 | * | |
272 | * The paravirt patching needs to be done _before_ the alternative asm code | |
273 | * is started, otherwise we would not patch the core kernel code. | |
274 | */ | |
d5de8841 JF |
275 | void __init xen_init_spinlocks(void) |
276 | { | |
70dd4998 | 277 | |
b8fa70b5 JF |
278 | if (!xen_pvspin) { |
279 | printk(KERN_DEBUG "xen: PV spinlocks disabled\n"); | |
280 | return; | |
281 | } | |
e0fc17a9 | 282 | printk(KERN_DEBUG "xen: PV spinlocks enabled\n"); |
354714dd | 283 | pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning); |
80bd58fe | 284 | pv_lock_ops.unlock_kick = xen_unlock_kick; |
d5de8841 | 285 | } |
994025ca | 286 | |
a945928e KRW |
287 | /* |
288 | * While the jump_label init code needs to happend _after_ the jump labels are | |
289 | * enabled and before SMP is started. Hence we use pre-SMP initcall level | |
290 | * init. We cannot do it in xen_init_spinlocks as that is done before | |
291 | * jump labels are activated. | |
292 | */ | |
293 | static __init int xen_init_spinlocks_jump(void) | |
294 | { | |
295 | if (!xen_pvspin) | |
296 | return 0; | |
297 | ||
e0fc17a9 KRW |
298 | if (!xen_domain()) |
299 | return 0; | |
300 | ||
a945928e KRW |
301 | static_key_slow_inc(¶virt_ticketlocks_enabled); |
302 | return 0; | |
303 | } | |
304 | early_initcall(xen_init_spinlocks_jump); | |
305 | ||
b8fa70b5 JF |
306 | static __init int xen_parse_nopvspin(char *arg) |
307 | { | |
308 | xen_pvspin = false; | |
309 | return 0; | |
310 | } | |
311 | early_param("xen_nopvspin", xen_parse_nopvspin); | |
312 | ||
994025ca JF |
313 | #ifdef CONFIG_XEN_DEBUG_FS |
314 | ||
315 | static struct dentry *d_spin_debug; | |
316 | ||
317 | static int __init xen_spinlock_debugfs(void) | |
318 | { | |
319 | struct dentry *d_xen = xen_init_debugfs(); | |
320 | ||
321 | if (d_xen == NULL) | |
322 | return -ENOMEM; | |
323 | ||
3310bbed KRW |
324 | if (!xen_pvspin) |
325 | return 0; | |
326 | ||
994025ca JF |
327 | d_spin_debug = debugfs_create_dir("spinlocks", d_xen); |
328 | ||
329 | debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats); | |
330 | ||
994025ca | 331 | debugfs_create_u32("taken_slow", 0444, d_spin_debug, |
80bd58fe | 332 | &spinlock_stats.contention_stats[TAKEN_SLOW]); |
994025ca | 333 | debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug, |
80bd58fe | 334 | &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]); |
994025ca | 335 | debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug, |
80bd58fe | 336 | &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]); |
994025ca | 337 | |
994025ca | 338 | debugfs_create_u32("released_slow", 0444, d_spin_debug, |
80bd58fe | 339 | &spinlock_stats.contention_stats[RELEASED_SLOW]); |
994025ca | 340 | debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug, |
80bd58fe | 341 | &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]); |
994025ca | 342 | |
f8eca41f JF |
343 | debugfs_create_u64("time_blocked", 0444, d_spin_debug, |
344 | &spinlock_stats.time_blocked); | |
994025ca | 345 | |
9fe2a701 SV |
346 | debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug, |
347 | spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1); | |
994025ca JF |
348 | |
349 | return 0; | |
350 | } | |
351 | fs_initcall(xen_spinlock_debugfs); | |
352 | ||
353 | #endif /* CONFIG_XEN_DEBUG_FS */ |