blob: cc9b1e182fcfad86bc67b56a8e172fb73d9e9ecf [file] [log] [blame]
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -07001/*
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>
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -07007#include <linux/debugfs.h>
8#include <linux/log2.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09009#include <linux/gfp.h>
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -070010
11#include <asm/paravirt.h>
12
13#include <xen/interface/xen.h>
14#include <xen/events.h>
15
16#include "xen-ops.h"
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070017#include "debugfs.h"
18
19#ifdef CONFIG_XEN_DEBUG_FS
20static 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;
Jeremy Fitzhardinge1e696f62008-08-20 17:02:20 -070027 u32 taken_slow_irqenable;
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070028
29 u64 released;
30 u32 released_slow;
31 u32 released_slow_kicked;
32
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -070033#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];
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070037
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -070038 u64 time_total;
39 u64 time_spinning;
40 u64 time_blocked;
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070041} spinlock_stats;
42
43static u8 zero_stats;
44
45static unsigned lock_timeout = 1 << 10;
46#define TIMEOUT lock_timeout
47
48static 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
59static inline u64 spin_time_start(void)
60{
61 return xen_clocksource_read();
62}
63
64static 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
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -070076static inline void spin_time_accum_spinning(u64 start)
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070077{
78 u32 delta = xen_clocksource_read() - start;
79
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -070080 __spin_time_accum(delta, spinlock_stats.histo_spin_spinning);
81 spinlock_stats.time_spinning += delta;
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070082}
83
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -070084static inline void spin_time_accum_total(u64 start)
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070085{
86 u32 delta = xen_clocksource_read() - start;
87
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -070088 __spin_time_accum(delta, spinlock_stats.histo_spin_total);
89 spinlock_stats.time_total += delta;
90}
91
92static 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;
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -070098}
99#else /* !CONFIG_XEN_DEBUG_FS */
100#define TIMEOUT (1 << 10)
101#define ADD_STATS(elem, val) do { (void)(val); } while(0)
102
103static inline u64 spin_time_start(void)
104{
105 return 0;
106}
107
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700108static inline void spin_time_accum_total(u64 start)
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700109{
110}
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700111static inline void spin_time_accum_spinning(u64 start)
112{
113}
114static inline void spin_time_accum_blocked(u64 start)
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700115{
116}
117#endif /* CONFIG_XEN_DEBUG_FS */
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700118
119struct xen_spinlock {
120 unsigned char lock; /* 0 -> free; 1 -> locked */
121 unsigned short spinners; /* count of waiting cpus */
122};
123
Thomas Gleixner445c8952009-12-02 19:49:50 +0100124static int xen_spin_is_locked(struct arch_spinlock *lock)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700125{
126 struct xen_spinlock *xl = (struct xen_spinlock *)lock;
127
128 return xl->lock != 0;
129}
130
Thomas Gleixner445c8952009-12-02 19:49:50 +0100131static int xen_spin_is_contended(struct arch_spinlock *lock)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700132{
133 struct xen_spinlock *xl = (struct xen_spinlock *)lock;
134
135 /* Not strictly true; this is only the count of contended
136 lock-takers entering the slow path. */
137 return xl->spinners != 0;
138}
139
Thomas Gleixner445c8952009-12-02 19:49:50 +0100140static int xen_spin_trylock(struct arch_spinlock *lock)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700141{
142 struct xen_spinlock *xl = (struct xen_spinlock *)lock;
143 u8 old = 1;
144
145 asm("xchgb %b0,%1"
146 : "+q" (old), "+m" (xl->lock) : : "memory");
147
148 return old == 0;
149}
150
151static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
152static DEFINE_PER_CPU(struct xen_spinlock *, lock_spinners);
153
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700154/*
155 * Mark a cpu as interested in a lock. Returns the CPU's previous
156 * lock of interest, in case we got preempted by an interrupt.
157 */
158static inline struct xen_spinlock *spinning_lock(struct xen_spinlock *xl)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700159{
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700160 struct xen_spinlock *prev;
161
Christoph Lameter780f36d2010-12-06 11:16:29 -0600162 prev = __this_cpu_read(lock_spinners);
163 __this_cpu_write(lock_spinners, xl);
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700164
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700165 wmb(); /* set lock of interest before count */
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700166
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700167 asm(LOCK_PREFIX " incw %0"
168 : "+m" (xl->spinners) : : "memory");
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700169
170 return prev;
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700171}
172
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700173/*
174 * Mark a cpu as no longer interested in a lock. Restores previous
175 * lock of interest (NULL for none).
176 */
177static inline void unspinning_lock(struct xen_spinlock *xl, struct xen_spinlock *prev)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700178{
179 asm(LOCK_PREFIX " decw %0"
180 : "+m" (xl->spinners) : : "memory");
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700181 wmb(); /* decrement count before restoring lock */
Christoph Lameter780f36d2010-12-06 11:16:29 -0600182 __this_cpu_write(lock_spinners, prev);
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700183}
184
Thomas Gleixner445c8952009-12-02 19:49:50 +0100185static noinline int xen_spin_lock_slow(struct arch_spinlock *lock, bool irq_enable)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700186{
187 struct xen_spinlock *xl = (struct xen_spinlock *)lock;
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700188 struct xen_spinlock *prev;
Christoph Lameter780f36d2010-12-06 11:16:29 -0600189 int irq = __this_cpu_read(lock_kicker_irq);
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700190 int ret;
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700191 u64 start;
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700192
193 /* If kicker interrupts not initialized yet, just spin */
194 if (irq == -1)
195 return 0;
196
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700197 start = spin_time_start();
198
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700199 /* announce we're spinning */
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700200 prev = spinning_lock(xl);
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700201
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700202 ADD_STATS(taken_slow, 1);
203 ADD_STATS(taken_slow_nested, prev != NULL);
204
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700205 do {
Jeremy Fitzhardinge4d576b52009-09-09 12:33:51 -0700206 unsigned long flags;
207
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700208 /* clear pending */
209 xen_clear_irq_pending(irq);
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700210
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700211 /* check again make sure it didn't become free while
212 we weren't looking */
213 ret = xen_spin_trylock(lock);
214 if (ret) {
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700215 ADD_STATS(taken_slow_pickup, 1);
216
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700217 /*
218 * If we interrupted another spinlock while it
219 * was blocking, make sure it doesn't block
220 * without rechecking the lock.
221 */
222 if (prev != NULL)
223 xen_set_irq_pending(irq);
224 goto out;
225 }
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700226
David Howellsdf9ee292010-10-07 14:08:55 +0100227 flags = arch_local_save_flags();
Jeremy Fitzhardinge4d576b52009-09-09 12:33:51 -0700228 if (irq_enable) {
229 ADD_STATS(taken_slow_irqenable, 1);
230 raw_local_irq_enable();
231 }
232
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700233 /*
234 * Block until irq becomes pending. If we're
235 * interrupted at this point (after the trylock but
236 * before entering the block), then the nested lock
237 * handler guarantees that the irq will be left
238 * pending if there's any chance the lock became free;
239 * xen_poll_irq() returns immediately if the irq is
240 * pending.
241 */
242 xen_poll_irq(irq);
Jeremy Fitzhardinge4d576b52009-09-09 12:33:51 -0700243
244 raw_local_irq_restore(flags);
245
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700246 ADD_STATS(taken_slow_spurious, !xen_test_irq_pending(irq));
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700247 } while (!xen_test_irq_pending(irq)); /* check for spurious wakeups */
248
Thomas Gleixnerd6c88a52008-10-15 15:27:23 +0200249 kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700250
251out:
Jeremy Fitzhardinge168d2f42008-08-20 17:02:18 -0700252 unspinning_lock(xl, prev);
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700253 spin_time_accum_blocked(start);
254
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700255 return ret;
256}
257
Thomas Gleixner445c8952009-12-02 19:49:50 +0100258static inline void __xen_spin_lock(struct arch_spinlock *lock, bool irq_enable)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700259{
260 struct xen_spinlock *xl = (struct xen_spinlock *)lock;
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700261 unsigned timeout;
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700262 u8 oldval;
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700263 u64 start_spin;
264
265 ADD_STATS(taken, 1);
266
267 start_spin = spin_time_start();
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700268
269 do {
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700270 u64 start_spin_fast = spin_time_start();
271
272 timeout = TIMEOUT;
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700273
274 asm("1: xchgb %1,%0\n"
275 " testb %1,%1\n"
276 " jz 3f\n"
277 "2: rep;nop\n"
278 " cmpb $0,%0\n"
279 " je 1b\n"
280 " dec %2\n"
281 " jnz 2b\n"
282 "3:\n"
283 : "+m" (xl->lock), "=q" (oldval), "+r" (timeout)
284 : "1" (1)
285 : "memory");
286
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700287 spin_time_accum_spinning(start_spin_fast);
Jeremy Fitzhardinge1e696f62008-08-20 17:02:20 -0700288
289 } while (unlikely(oldval != 0 &&
290 (TIMEOUT == ~0 || !xen_spin_lock_slow(lock, irq_enable))));
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700291
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700292 spin_time_accum_total(start_spin);
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700293}
294
Thomas Gleixner445c8952009-12-02 19:49:50 +0100295static void xen_spin_lock(struct arch_spinlock *lock)
Jeremy Fitzhardinge1e696f62008-08-20 17:02:20 -0700296{
297 __xen_spin_lock(lock, false);
298}
299
Thomas Gleixner445c8952009-12-02 19:49:50 +0100300static void xen_spin_lock_flags(struct arch_spinlock *lock, unsigned long flags)
Jeremy Fitzhardinge1e696f62008-08-20 17:02:20 -0700301{
302 __xen_spin_lock(lock, !raw_irqs_disabled_flags(flags));
303}
304
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700305static noinline void xen_spin_unlock_slow(struct xen_spinlock *xl)
306{
307 int cpu;
308
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700309 ADD_STATS(released_slow, 1);
310
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700311 for_each_online_cpu(cpu) {
312 /* XXX should mix up next cpu selection */
313 if (per_cpu(lock_spinners, cpu) == xl) {
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700314 ADD_STATS(released_slow_kicked, 1);
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700315 xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
316 break;
317 }
318 }
319}
320
Thomas Gleixner445c8952009-12-02 19:49:50 +0100321static void xen_spin_unlock(struct arch_spinlock *lock)
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700322{
323 struct xen_spinlock *xl = (struct xen_spinlock *)lock;
324
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700325 ADD_STATS(released, 1);
326
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700327 smp_wmb(); /* make sure no writes get moved after unlock */
328 xl->lock = 0; /* release lock */
329
Yang Xiaowei2496afb2009-09-09 12:44:52 -0700330 /*
331 * Make sure unlock happens before checking for waiting
332 * spinners. We need a strong barrier to enforce the
333 * write-read ordering to different memory locations, as the
334 * CPU makes no implied guarantees about their ordering.
335 */
336 mb();
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700337
338 if (unlikely(xl->spinners))
339 xen_spin_unlock_slow(xl);
340}
341
342static irqreturn_t dummy_handler(int irq, void *dev_id)
343{
344 BUG();
345 return IRQ_HANDLED;
346}
347
348void __cpuinit xen_init_lock_cpu(int cpu)
349{
350 int irq;
351 const char *name;
352
353 name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
354 irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
355 cpu,
356 dummy_handler,
357 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
358 name,
359 NULL);
360
361 if (irq >= 0) {
362 disable_irq(irq); /* make sure it's never delivered */
363 per_cpu(lock_kicker_irq, cpu) = irq;
364 }
365
366 printk("cpu %d spinlock event irq %d\n", cpu, irq);
367}
368
Alex Nixond68d82a2008-08-22 11:52:15 +0100369void xen_uninit_lock_cpu(int cpu)
370{
371 unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
372}
373
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700374void __init xen_init_spinlocks(void)
375{
376 pv_lock_ops.spin_is_locked = xen_spin_is_locked;
377 pv_lock_ops.spin_is_contended = xen_spin_is_contended;
378 pv_lock_ops.spin_lock = xen_spin_lock;
Jeremy Fitzhardinge1e696f62008-08-20 17:02:20 -0700379 pv_lock_ops.spin_lock_flags = xen_spin_lock_flags;
Jeremy Fitzhardinged5de8842008-07-23 13:28:58 -0700380 pv_lock_ops.spin_trylock = xen_spin_trylock;
381 pv_lock_ops.spin_unlock = xen_spin_unlock;
382}
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700383
384#ifdef CONFIG_XEN_DEBUG_FS
385
386static struct dentry *d_spin_debug;
387
388static int __init xen_spinlock_debugfs(void)
389{
390 struct dentry *d_xen = xen_init_debugfs();
391
392 if (d_xen == NULL)
393 return -ENOMEM;
394
395 d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
396
397 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
398
399 debugfs_create_u32("timeout", 0644, d_spin_debug, &lock_timeout);
400
401 debugfs_create_u64("taken", 0444, d_spin_debug, &spinlock_stats.taken);
402 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
403 &spinlock_stats.taken_slow);
404 debugfs_create_u32("taken_slow_nested", 0444, d_spin_debug,
405 &spinlock_stats.taken_slow_nested);
406 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
407 &spinlock_stats.taken_slow_pickup);
408 debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
409 &spinlock_stats.taken_slow_spurious);
Jeremy Fitzhardinge1e696f62008-08-20 17:02:20 -0700410 debugfs_create_u32("taken_slow_irqenable", 0444, d_spin_debug,
411 &spinlock_stats.taken_slow_irqenable);
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700412
413 debugfs_create_u64("released", 0444, d_spin_debug, &spinlock_stats.released);
414 debugfs_create_u32("released_slow", 0444, d_spin_debug,
415 &spinlock_stats.released_slow);
416 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
417 &spinlock_stats.released_slow_kicked);
418
419 debugfs_create_u64("time_spinning", 0444, d_spin_debug,
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700420 &spinlock_stats.time_spinning);
421 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
422 &spinlock_stats.time_blocked);
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700423 debugfs_create_u64("time_total", 0444, d_spin_debug,
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700424 &spinlock_stats.time_total);
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700425
426 xen_debugfs_create_u32_array("histo_total", 0444, d_spin_debug,
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700427 spinlock_stats.histo_spin_total, HISTO_BUCKETS + 1);
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700428 xen_debugfs_create_u32_array("histo_spinning", 0444, d_spin_debug,
Jeremy Fitzhardingef8eca412008-08-20 17:02:21 -0700429 spinlock_stats.histo_spin_spinning, HISTO_BUCKETS + 1);
430 xen_debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
431 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
Jeremy Fitzhardinge994025c2008-08-20 17:02:19 -0700432
433 return 0;
434}
435fs_initcall(xen_spinlock_debugfs);
436
437#endif /* CONFIG_XEN_DEBUG_FS */