Merge branch 'for-linus' of master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband
[linux-2.6] / arch / i386 / kernel / vmitime.c
1 /*
2  * VMI paravirtual timer support routines.
3  *
4  * Copyright (C) 2005, VMware, Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License as published by
8  * the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful, but
12  * WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
14  * NON INFRINGEMENT.  See the GNU General Public License for more
15  * details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program; if not, write to the Free Software
19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  *
21  * Send feedback to dhecht@vmware.com
22  *
23  */
24
25 /*
26  * Portions of this code from arch/i386/kernel/timers/timer_tsc.c.
27  * Portions of the CONFIG_NO_IDLE_HZ code from arch/s390/kernel/time.c.
28  * See comments there for proper credits.
29  */
30
31 #include <linux/spinlock.h>
32 #include <linux/init.h>
33 #include <linux/errno.h>
34 #include <linux/jiffies.h>
35 #include <linux/interrupt.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/rcupdate.h>
38 #include <linux/clocksource.h>
39
40 #include <asm/timer.h>
41 #include <asm/io.h>
42 #include <asm/apic.h>
43 #include <asm/div64.h>
44 #include <asm/timer.h>
45 #include <asm/desc.h>
46
47 #include <asm/vmi.h>
48 #include <asm/vmi_time.h>
49
50 #include <mach_timer.h>
51 #include <io_ports.h>
52
53 #ifdef CONFIG_X86_LOCAL_APIC
54 #define VMI_ALARM_WIRING VMI_ALARM_WIRED_LVTT
55 #else
56 #define VMI_ALARM_WIRING VMI_ALARM_WIRED_IRQ0
57 #endif
58
59 /* Cached VMI operations */
60 struct vmi_timer_ops vmi_timer_ops;
61
62 #ifdef CONFIG_NO_IDLE_HZ
63
64 /* /proc/sys/kernel/hz_timer state. */
65 int sysctl_hz_timer;
66
67 /* Some stats */
68 static DEFINE_PER_CPU(unsigned long, vmi_idle_no_hz_irqs);
69 static DEFINE_PER_CPU(unsigned long, vmi_idle_no_hz_jiffies);
70 static DEFINE_PER_CPU(unsigned long, idle_start_jiffies);
71
72 #endif /* CONFIG_NO_IDLE_HZ */
73
74 /* Number of alarms per second. By default this is CONFIG_VMI_ALARM_HZ. */
75 static int alarm_hz = CONFIG_VMI_ALARM_HZ;
76
77 /* Cache of the value get_cycle_frequency / HZ. */
78 static signed long long cycles_per_jiffy;
79
80 /* Cache of the value get_cycle_frequency / alarm_hz. */
81 static signed long long cycles_per_alarm;
82
83 /* The number of cycles accounted for by the 'jiffies'/'xtime' count.
84  * Protected by xtime_lock. */
85 static unsigned long long real_cycles_accounted_system;
86
87 /* The number of cycles accounted for by update_process_times(), per cpu. */
88 static DEFINE_PER_CPU(unsigned long long, process_times_cycles_accounted_cpu);
89
90 /* The number of stolen cycles accounted, per cpu. */
91 static DEFINE_PER_CPU(unsigned long long, stolen_cycles_accounted_cpu);
92
93 /* Clock source. */
94 static cycle_t read_real_cycles(void)
95 {
96         return vmi_timer_ops.get_cycle_counter(VMI_CYCLES_REAL);
97 }
98
99 static cycle_t read_available_cycles(void)
100 {
101         return vmi_timer_ops.get_cycle_counter(VMI_CYCLES_AVAILABLE);
102 }
103
104 #if 0
105 static cycle_t read_stolen_cycles(void)
106 {
107         return vmi_timer_ops.get_cycle_counter(VMI_CYCLES_STOLEN);
108 }
109 #endif  /*  0  */
110
111 static struct clocksource clocksource_vmi = {
112         .name                   = "vmi-timer",
113         .rating                 = 450,
114         .read                   = read_real_cycles,
115         .mask                   = CLOCKSOURCE_MASK(64),
116         .mult                   = 0, /* to be set */
117         .shift                  = 22,
118         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS,
119 };
120
121
122 /* Timer interrupt handler. */
123 static irqreturn_t vmi_timer_interrupt(int irq, void *dev_id);
124
125 static struct irqaction vmi_timer_irq  = {
126         .handler = vmi_timer_interrupt,
127         .flags = IRQF_DISABLED,
128         .mask = CPU_MASK_NONE,
129         .name = "VMI-alarm",
130 };
131
132 /* Alarm rate */
133 static int __init vmi_timer_alarm_rate_setup(char* str)
134 {
135         int alarm_rate;
136         if (get_option(&str, &alarm_rate) == 1 && alarm_rate > 0) {
137                 alarm_hz = alarm_rate;
138                 printk(KERN_WARNING "VMI timer alarm HZ set to %d\n", alarm_hz);
139         }
140         return 1;
141 }
142 __setup("vmi_timer_alarm_hz=", vmi_timer_alarm_rate_setup);
143
144
145 /* Initialization */
146 static void vmi_get_wallclock_ts(struct timespec *ts)
147 {
148         unsigned long long wallclock;
149         wallclock = vmi_timer_ops.get_wallclock(); // nsec units
150         ts->tv_nsec = do_div(wallclock, 1000000000);
151         ts->tv_sec = wallclock;
152 }
153
154 unsigned long vmi_get_wallclock(void)
155 {
156         struct timespec ts;
157         vmi_get_wallclock_ts(&ts);
158         return ts.tv_sec;
159 }
160
161 int vmi_set_wallclock(unsigned long now)
162 {
163         return -1;
164 }
165
166 unsigned long long vmi_get_sched_cycles(void)
167 {
168         return read_available_cycles();
169 }
170
171 unsigned long vmi_cpu_khz(void)
172 {
173         unsigned long long khz;
174
175         khz = vmi_timer_ops.get_cycle_frequency();
176         (void)do_div(khz, 1000);
177         return khz;
178 }
179
180 void __init vmi_time_init(void)
181 {
182         unsigned long long cycles_per_sec, cycles_per_msec;
183         unsigned long flags;
184
185         local_irq_save(flags);
186         setup_irq(0, &vmi_timer_irq);
187 #ifdef CONFIG_X86_LOCAL_APIC
188         set_intr_gate(LOCAL_TIMER_VECTOR, apic_vmi_timer_interrupt);
189 #endif
190
191         real_cycles_accounted_system = read_real_cycles();
192         per_cpu(process_times_cycles_accounted_cpu, 0) = read_available_cycles();
193
194         cycles_per_sec = vmi_timer_ops.get_cycle_frequency();
195         cycles_per_jiffy = cycles_per_sec;
196         (void)do_div(cycles_per_jiffy, HZ);
197         cycles_per_alarm = cycles_per_sec;
198         (void)do_div(cycles_per_alarm, alarm_hz);
199         cycles_per_msec = cycles_per_sec;
200         (void)do_div(cycles_per_msec, 1000);
201
202         printk(KERN_WARNING "VMI timer cycles/sec = %llu ; cycles/jiffy = %llu ;"
203                "cycles/alarm = %llu\n", cycles_per_sec, cycles_per_jiffy,
204                cycles_per_alarm);
205
206         clocksource_vmi.mult = clocksource_khz2mult(cycles_per_msec,
207                                                     clocksource_vmi.shift);
208         if (clocksource_register(&clocksource_vmi))
209                 printk(KERN_WARNING "Error registering VMITIME clocksource.");
210
211         /* Disable PIT. */
212         outb_p(0x3a, PIT_MODE); /* binary, mode 5, LSB/MSB, ch 0 */
213
214         /* schedule the alarm. do this in phase with process_times_cycles_accounted_cpu
215          * reduce the latency calling update_process_times. */
216         vmi_timer_ops.set_alarm(
217                       VMI_ALARM_WIRED_IRQ0 | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
218                       per_cpu(process_times_cycles_accounted_cpu, 0) + cycles_per_alarm,
219                       cycles_per_alarm);
220
221         local_irq_restore(flags);
222 }
223
224 #ifdef CONFIG_X86_LOCAL_APIC
225
226 void __init vmi_timer_setup_boot_alarm(void)
227 {
228         local_irq_disable();
229
230         /* Route the interrupt to the correct vector. */
231         apic_write_around(APIC_LVTT, LOCAL_TIMER_VECTOR);
232
233         /* Cancel the IRQ0 wired alarm, and setup the LVTT alarm. */
234         vmi_timer_ops.cancel_alarm(VMI_CYCLES_AVAILABLE);
235         vmi_timer_ops.set_alarm(
236                       VMI_ALARM_WIRED_LVTT | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
237                       per_cpu(process_times_cycles_accounted_cpu, 0) + cycles_per_alarm,
238                       cycles_per_alarm);
239         local_irq_enable();
240 }
241
242 /* Initialize the time accounting variables for an AP on an SMP system.
243  * Also, set the local alarm for the AP. */
244 void __devinit vmi_timer_setup_secondary_alarm(void)
245 {
246         int cpu = smp_processor_id();
247
248         /* Route the interrupt to the correct vector. */
249         apic_write_around(APIC_LVTT, LOCAL_TIMER_VECTOR);
250
251         per_cpu(process_times_cycles_accounted_cpu, cpu) = read_available_cycles();
252
253         vmi_timer_ops.set_alarm(
254                       VMI_ALARM_WIRED_LVTT | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
255                       per_cpu(process_times_cycles_accounted_cpu, cpu) + cycles_per_alarm,
256                       cycles_per_alarm);
257 }
258
259 #endif
260
261 /* Update system wide (real) time accounting (e.g. jiffies, xtime). */
262 static void vmi_account_real_cycles(unsigned long long cur_real_cycles)
263 {
264         long long cycles_not_accounted;
265
266         write_seqlock(&xtime_lock);
267
268         cycles_not_accounted = cur_real_cycles - real_cycles_accounted_system;
269         while (cycles_not_accounted >= cycles_per_jiffy) {
270                 /* systems wide jiffies. */
271                 do_timer(1);
272
273                 cycles_not_accounted -= cycles_per_jiffy;
274                 real_cycles_accounted_system += cycles_per_jiffy;
275         }
276
277         write_sequnlock(&xtime_lock);
278 }
279
280 /* Update per-cpu process times. */
281 static void vmi_account_process_times_cycles(struct pt_regs *regs, int cpu,
282                                              unsigned long long cur_process_times_cycles)
283 {
284         long long cycles_not_accounted;
285         cycles_not_accounted = cur_process_times_cycles -
286                 per_cpu(process_times_cycles_accounted_cpu, cpu);
287
288         while (cycles_not_accounted >= cycles_per_jiffy) {
289                 /* Account time to the current process.  This includes
290                  * calling into the scheduler to decrement the timeslice
291                  * and possibly reschedule.*/
292                 update_process_times(user_mode(regs));
293                 /* XXX handle /proc/profile multiplier.  */
294                 profile_tick(CPU_PROFILING);
295
296                 cycles_not_accounted -= cycles_per_jiffy;
297                 per_cpu(process_times_cycles_accounted_cpu, cpu) += cycles_per_jiffy;
298         }
299 }
300
301 #ifdef CONFIG_NO_IDLE_HZ
302 /* Update per-cpu idle times.  Used when a no-hz halt is ended. */
303 static void vmi_account_no_hz_idle_cycles(int cpu,
304                                           unsigned long long cur_process_times_cycles)
305 {
306         long long cycles_not_accounted;
307         unsigned long no_idle_hz_jiffies = 0;
308
309         cycles_not_accounted = cur_process_times_cycles -
310                 per_cpu(process_times_cycles_accounted_cpu, cpu);
311
312         while (cycles_not_accounted >= cycles_per_jiffy) {
313                 no_idle_hz_jiffies++;
314                 cycles_not_accounted -= cycles_per_jiffy;
315                 per_cpu(process_times_cycles_accounted_cpu, cpu) += cycles_per_jiffy;
316         }
317         /* Account time to the idle process. */
318         account_steal_time(idle_task(cpu), jiffies_to_cputime(no_idle_hz_jiffies));
319 }
320 #endif
321
322 /* Update per-cpu stolen time. */
323 static void vmi_account_stolen_cycles(int cpu,
324                                       unsigned long long cur_real_cycles,
325                                       unsigned long long cur_avail_cycles)
326 {
327         long long stolen_cycles_not_accounted;
328         unsigned long stolen_jiffies = 0;
329
330         if (cur_real_cycles < cur_avail_cycles)
331                 return;
332
333         stolen_cycles_not_accounted = cur_real_cycles - cur_avail_cycles -
334                 per_cpu(stolen_cycles_accounted_cpu, cpu);
335
336         while (stolen_cycles_not_accounted >= cycles_per_jiffy) {
337                 stolen_jiffies++;
338                 stolen_cycles_not_accounted -= cycles_per_jiffy;
339                 per_cpu(stolen_cycles_accounted_cpu, cpu) += cycles_per_jiffy;
340         }
341         /* HACK: pass NULL to force time onto cpustat->steal. */
342         account_steal_time(NULL, jiffies_to_cputime(stolen_jiffies));
343 }
344
345 /* Body of either IRQ0 interrupt handler (UP no local-APIC) or
346  * local-APIC LVTT interrupt handler (UP & local-APIC or SMP). */
347 static void vmi_local_timer_interrupt(int cpu)
348 {
349         unsigned long long cur_real_cycles, cur_process_times_cycles;
350
351         cur_real_cycles = read_real_cycles();
352         cur_process_times_cycles = read_available_cycles();
353         /* Update system wide (real) time state (xtime, jiffies). */
354         vmi_account_real_cycles(cur_real_cycles);
355         /* Update per-cpu process times. */
356         vmi_account_process_times_cycles(get_irq_regs(), cpu, cur_process_times_cycles);
357         /* Update time stolen from this cpu by the hypervisor. */
358         vmi_account_stolen_cycles(cpu, cur_real_cycles, cur_process_times_cycles);
359 }
360
361 #ifdef CONFIG_NO_IDLE_HZ
362
363 /* Must be called only from idle loop, with interrupts disabled. */
364 int vmi_stop_hz_timer(void)
365 {
366         /* Note that cpu_set, cpu_clear are (SMP safe) atomic on x86. */
367
368         unsigned long seq, next;
369         unsigned long long real_cycles_expiry;
370         int cpu = smp_processor_id();
371
372         BUG_ON(!irqs_disabled());
373         if (sysctl_hz_timer != 0)
374                 return 0;
375
376         cpu_set(cpu, nohz_cpu_mask);
377         smp_mb();
378
379         if (rcu_needs_cpu(cpu) || local_softirq_pending() ||
380             (next = next_timer_interrupt(),
381              time_before_eq(next, jiffies + HZ/CONFIG_VMI_ALARM_HZ))) {
382                 cpu_clear(cpu, nohz_cpu_mask);
383                 return 0;
384         }
385
386         /* Convert jiffies to the real cycle counter. */
387         do {
388                 seq = read_seqbegin(&xtime_lock);
389                 real_cycles_expiry = real_cycles_accounted_system +
390                         (long)(next - jiffies) * cycles_per_jiffy;
391         } while (read_seqretry(&xtime_lock, seq));
392
393         /* This cpu is going idle. Disable the periodic alarm. */
394         vmi_timer_ops.cancel_alarm(VMI_CYCLES_AVAILABLE);
395         per_cpu(idle_start_jiffies, cpu) = jiffies;
396         /* Set the real time alarm to expire at the next event. */
397         vmi_timer_ops.set_alarm(
398                 VMI_ALARM_WIRING | VMI_ALARM_IS_ONESHOT | VMI_CYCLES_REAL,
399                 real_cycles_expiry, 0);
400         return 1;
401 }
402
403 static void vmi_reenable_hz_timer(int cpu)
404 {
405         /* For /proc/vmi/info idle_hz stat. */
406         per_cpu(vmi_idle_no_hz_jiffies, cpu) += jiffies - per_cpu(idle_start_jiffies, cpu);
407         per_cpu(vmi_idle_no_hz_irqs, cpu)++;
408
409         /* Don't bother explicitly cancelling the one-shot alarm -- at
410          * worse we will receive a spurious timer interrupt. */
411         vmi_timer_ops.set_alarm(
412                       VMI_ALARM_WIRING | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
413                       per_cpu(process_times_cycles_accounted_cpu, cpu) + cycles_per_alarm,
414                       cycles_per_alarm);
415         /* Indicate this cpu is no longer nohz idle. */
416         cpu_clear(cpu, nohz_cpu_mask);
417 }
418
419 /* Called from interrupt handlers when (local) HZ timer is disabled. */
420 void vmi_account_time_restart_hz_timer(void)
421 {
422         unsigned long long cur_real_cycles, cur_process_times_cycles;
423         int cpu = smp_processor_id();
424
425         BUG_ON(!irqs_disabled());
426         /* Account the time during which the HZ timer was disabled. */
427         cur_real_cycles = read_real_cycles();
428         cur_process_times_cycles = read_available_cycles();
429         /* Update system wide (real) time state (xtime, jiffies). */
430         vmi_account_real_cycles(cur_real_cycles);
431         /* Update per-cpu idle times. */
432         vmi_account_no_hz_idle_cycles(cpu, cur_process_times_cycles);
433         /* Update time stolen from this cpu by the hypervisor. */
434         vmi_account_stolen_cycles(cpu, cur_real_cycles, cur_process_times_cycles);
435         /* Reenable the hz timer. */
436         vmi_reenable_hz_timer(cpu);
437 }
438
439 #endif /* CONFIG_NO_IDLE_HZ */
440
441 /* UP (and no local-APIC) VMI-timer alarm interrupt handler.
442  * Handler for IRQ0. Not used when SMP or X86_LOCAL_APIC after
443  * APIC setup and setup_boot_vmi_alarm() is called.  */
444 static irqreturn_t vmi_timer_interrupt(int irq, void *dev_id)
445 {
446         vmi_local_timer_interrupt(smp_processor_id());
447         return IRQ_HANDLED;
448 }
449
450 #ifdef CONFIG_X86_LOCAL_APIC
451
452 /* SMP VMI-timer alarm interrupt handler. Handler for LVTT vector.
453  * Also used in UP when CONFIG_X86_LOCAL_APIC.
454  * The wrapper code is from arch/i386/kernel/apic.c#smp_apic_timer_interrupt. */
455 void smp_apic_vmi_timer_interrupt(struct pt_regs *regs)
456 {
457         struct pt_regs *old_regs = set_irq_regs(regs);
458         int cpu = smp_processor_id();
459
460         /*
461          * the NMI deadlock-detector uses this.
462          */
463         per_cpu(irq_stat,cpu).apic_timer_irqs++;
464
465         /*
466          * NOTE! We'd better ACK the irq immediately,
467          * because timer handling can be slow.
468          */
469         ack_APIC_irq();
470
471         /*
472          * update_process_times() expects us to have done irq_enter().
473          * Besides, if we don't timer interrupts ignore the global
474          * interrupt lock, which is the WrongThing (tm) to do.
475          */
476         irq_enter();
477         vmi_local_timer_interrupt(cpu);
478         irq_exit();
479         set_irq_regs(old_regs);
480 }
481
482 #endif  /* CONFIG_X86_LOCAL_APIC */