Commit | Line | Data |
---|---|---|
8531a35e KK |
1 | /* |
2 | * This file is subject to the terms and conditions of the GNU General Public | |
3 | * License. See the file "COPYING" in the main directory of this archive | |
4 | * for more details. | |
5 | * | |
6 | * Copyright (C) 2007 MIPS Technologies, Inc. | |
7 | * Copyright (C) 2007 Ralf Baechle <ralf@linux-mips.org> | |
8 | * Copyright (C) 2008 Kevin D. Kissell, Paralogos sarl | |
9 | */ | |
10 | #include <linux/clockchips.h> | |
11 | #include <linux/interrupt.h> | |
12 | #include <linux/percpu.h> | |
631330f5 | 13 | #include <linux/smp.h> |
ca4d3e67 | 14 | #include <linux/irq.h> |
8531a35e KK |
15 | |
16 | #include <asm/smtc_ipi.h> | |
17 | #include <asm/time.h> | |
18 | #include <asm/cevt-r4k.h> | |
19 | ||
20 | /* | |
21 | * Variant clock event timer support for SMTC on MIPS 34K, 1004K | |
22 | * or other MIPS MT cores. | |
23 | * | |
24 | * Notes on SMTC Support: | |
25 | * | |
26 | * SMTC has multiple microthread TCs pretending to be Linux CPUs. | |
27 | * But there's only one Count/Compare pair per VPE, and Compare | |
28 | * interrupts are taken opportunisitically by available TCs | |
29 | * bound to the VPE with the Count register. The new timer | |
30 | * framework provides for global broadcasts, but we really | |
31 | * want VPE-level multicasts for best behavior. So instead | |
32 | * of invoking the high-level clock-event broadcast code, | |
33 | * this version of SMTC support uses the historical SMTC | |
34 | * multicast mechanisms "under the hood", appearing to the | |
35 | * generic clock layer as if the interrupts are per-CPU. | |
36 | * | |
37 | * The approach taken here is to maintain a set of NR_CPUS | |
38 | * virtual timers, and track which "CPU" needs to be alerted | |
39 | * at each event. | |
40 | * | |
41 | * It's unlikely that we'll see a MIPS MT core with more than | |
42 | * 2 VPEs, but we *know* that we won't need to handle more | |
43 | * VPEs than we have "CPUs". So NCPUs arrays of NCPUs elements | |
44 | * is always going to be overkill, but always going to be enough. | |
45 | */ | |
46 | ||
47 | unsigned long smtc_nexttime[NR_CPUS][NR_CPUS]; | |
48 | static int smtc_nextinvpe[NR_CPUS]; | |
49 | ||
50 | /* | |
51 | * Timestamps stored are absolute values to be programmed | |
52 | * into Count register. Valid timestamps will never be zero. | |
53 | * If a Zero Count value is actually calculated, it is converted | |
54 | * to be a 1, which will introduce 1 or two CPU cycles of error | |
55 | * roughly once every four billion events, which at 1000 HZ means | |
56 | * about once every 50 days. If that's actually a problem, one | |
57 | * could alternate squashing 0 to 1 and to -1. | |
58 | */ | |
59 | ||
60 | #define MAKEVALID(x) (((x) == 0L) ? 1L : (x)) | |
61 | #define ISVALID(x) ((x) != 0L) | |
62 | ||
63 | /* | |
64 | * Time comparison is subtle, as it's really truncated | |
65 | * modular arithmetic. | |
66 | */ | |
67 | ||
68 | #define IS_SOONER(a, b, reference) \ | |
69 | (((a) - (unsigned long)(reference)) < ((b) - (unsigned long)(reference))) | |
70 | ||
71 | /* | |
72 | * CATCHUP_INCREMENT, used when the function falls behind the counter. | |
73 | * Could be an increasing function instead of a constant; | |
74 | */ | |
75 | ||
76 | #define CATCHUP_INCREMENT 64 | |
77 | ||
78 | static int mips_next_event(unsigned long delta, | |
79 | struct clock_event_device *evt) | |
80 | { | |
81 | unsigned long flags; | |
82 | unsigned int mtflags; | |
83 | unsigned long timestamp, reference, previous; | |
84 | unsigned long nextcomp = 0L; | |
85 | int vpe = current_cpu_data.vpe_id; | |
86 | int cpu = smp_processor_id(); | |
87 | local_irq_save(flags); | |
88 | mtflags = dmt(); | |
89 | ||
90 | /* | |
91 | * Maintain the per-TC virtual timer | |
92 | * and program the per-VPE shared Count register | |
93 | * as appropriate here... | |
94 | */ | |
95 | reference = (unsigned long)read_c0_count(); | |
96 | timestamp = MAKEVALID(reference + delta); | |
97 | /* | |
98 | * To really model the clock, we have to catch the case | |
99 | * where the current next-in-VPE timestamp is the old | |
100 | * timestamp for the calling CPE, but the new value is | |
101 | * in fact later. In that case, we have to do a full | |
102 | * scan and discover the new next-in-VPE CPU id and | |
103 | * timestamp. | |
104 | */ | |
105 | previous = smtc_nexttime[vpe][cpu]; | |
106 | if (cpu == smtc_nextinvpe[vpe] && ISVALID(previous) | |
107 | && IS_SOONER(previous, timestamp, reference)) { | |
108 | int i; | |
109 | int soonest = cpu; | |
110 | ||
111 | /* | |
112 | * Update timestamp array here, so that new | |
113 | * value gets considered along with those of | |
114 | * other virtual CPUs on the VPE. | |
115 | */ | |
116 | smtc_nexttime[vpe][cpu] = timestamp; | |
117 | for_each_online_cpu(i) { | |
118 | if (ISVALID(smtc_nexttime[vpe][i]) | |
119 | && IS_SOONER(smtc_nexttime[vpe][i], | |
120 | smtc_nexttime[vpe][soonest], reference)) { | |
121 | soonest = i; | |
122 | } | |
123 | } | |
124 | smtc_nextinvpe[vpe] = soonest; | |
125 | nextcomp = smtc_nexttime[vpe][soonest]; | |
126 | /* | |
127 | * Otherwise, we don't have to process the whole array rank, | |
128 | * we just have to see if the event horizon has gotten closer. | |
129 | */ | |
130 | } else { | |
131 | if (!ISVALID(smtc_nexttime[vpe][smtc_nextinvpe[vpe]]) || | |
132 | IS_SOONER(timestamp, | |
133 | smtc_nexttime[vpe][smtc_nextinvpe[vpe]], reference)) { | |
134 | smtc_nextinvpe[vpe] = cpu; | |
135 | nextcomp = timestamp; | |
136 | } | |
137 | /* | |
138 | * Since next-in-VPE may me the same as the executing | |
139 | * virtual CPU, we update the array *after* checking | |
140 | * its value. | |
141 | */ | |
142 | smtc_nexttime[vpe][cpu] = timestamp; | |
143 | } | |
144 | ||
145 | /* | |
146 | * It may be that, in fact, we don't need to update Compare, | |
147 | * but if we do, we want to make sure we didn't fall into | |
148 | * a crack just behind Count. | |
149 | */ | |
150 | if (ISVALID(nextcomp)) { | |
151 | write_c0_compare(nextcomp); | |
152 | ehb(); | |
153 | /* | |
154 | * We never return an error, we just make sure | |
155 | * that we trigger the handlers as quickly as | |
156 | * we can if we fell behind. | |
157 | */ | |
158 | while ((nextcomp - (unsigned long)read_c0_count()) | |
159 | > (unsigned long)LONG_MAX) { | |
160 | nextcomp += CATCHUP_INCREMENT; | |
161 | write_c0_compare(nextcomp); | |
162 | ehb(); | |
163 | } | |
164 | } | |
165 | emt(mtflags); | |
166 | local_irq_restore(flags); | |
167 | return 0; | |
168 | } | |
169 | ||
170 | ||
171 | void smtc_distribute_timer(int vpe) | |
172 | { | |
173 | unsigned long flags; | |
174 | unsigned int mtflags; | |
175 | int cpu; | |
176 | struct clock_event_device *cd; | |
5df9d11b | 177 | unsigned long nextstamp; |
8531a35e KK |
178 | unsigned long reference; |
179 | ||
180 | ||
181 | repeat: | |
5df9d11b | 182 | nextstamp = 0L; |
8531a35e KK |
183 | for_each_online_cpu(cpu) { |
184 | /* | |
185 | * Find virtual CPUs within the current VPE who have | |
186 | * unserviced timer requests whose time is now past. | |
187 | */ | |
188 | local_irq_save(flags); | |
189 | mtflags = dmt(); | |
190 | if (cpu_data[cpu].vpe_id == vpe && | |
191 | ISVALID(smtc_nexttime[vpe][cpu])) { | |
192 | reference = (unsigned long)read_c0_count(); | |
193 | if ((smtc_nexttime[vpe][cpu] - reference) | |
194 | > (unsigned long)LONG_MAX) { | |
195 | smtc_nexttime[vpe][cpu] = 0L; | |
196 | emt(mtflags); | |
197 | local_irq_restore(flags); | |
198 | /* | |
199 | * We don't send IPIs to ourself. | |
200 | */ | |
201 | if (cpu != smp_processor_id()) { | |
202 | smtc_send_ipi(cpu, SMTC_CLOCK_TICK, 0); | |
203 | } else { | |
204 | cd = &per_cpu(mips_clockevent_device, cpu); | |
205 | cd->event_handler(cd); | |
206 | } | |
207 | } else { | |
208 | /* Local to VPE but Valid Time not yet reached. */ | |
209 | if (!ISVALID(nextstamp) || | |
210 | IS_SOONER(smtc_nexttime[vpe][cpu], nextstamp, | |
211 | reference)) { | |
212 | smtc_nextinvpe[vpe] = cpu; | |
213 | nextstamp = smtc_nexttime[vpe][cpu]; | |
214 | } | |
215 | emt(mtflags); | |
216 | local_irq_restore(flags); | |
217 | } | |
218 | } else { | |
219 | emt(mtflags); | |
220 | local_irq_restore(flags); | |
221 | ||
222 | } | |
223 | } | |
224 | /* Reprogram for interrupt at next soonest timestamp for VPE */ | |
225 | if (ISVALID(nextstamp)) { | |
226 | write_c0_compare(nextstamp); | |
227 | ehb(); | |
228 | if ((nextstamp - (unsigned long)read_c0_count()) | |
229 | > (unsigned long)LONG_MAX) | |
230 | goto repeat; | |
231 | } | |
232 | } | |
233 | ||
234 | ||
235 | irqreturn_t c0_compare_interrupt(int irq, void *dev_id) | |
236 | { | |
237 | int cpu = smp_processor_id(); | |
238 | ||
239 | /* If we're running SMTC, we've got MIPS MT and therefore MIPS32R2 */ | |
240 | handle_perf_irq(1); | |
241 | ||
242 | if (read_c0_cause() & (1 << 30)) { | |
243 | /* Clear Count/Compare Interrupt */ | |
244 | write_c0_compare(read_c0_compare()); | |
245 | smtc_distribute_timer(cpu_data[cpu].vpe_id); | |
246 | } | |
247 | return IRQ_HANDLED; | |
248 | } | |
249 | ||
250 | ||
bcf11801 | 251 | int __cpuinit smtc_clockevent_init(void) |
8531a35e KK |
252 | { |
253 | uint64_t mips_freq = mips_hpt_frequency; | |
254 | unsigned int cpu = smp_processor_id(); | |
255 | struct clock_event_device *cd; | |
256 | unsigned int irq; | |
257 | int i; | |
258 | int j; | |
259 | ||
260 | if (!cpu_has_counter || !mips_hpt_frequency) | |
261 | return -ENXIO; | |
262 | if (cpu == 0) { | |
263 | for (i = 0; i < num_possible_cpus(); i++) { | |
264 | smtc_nextinvpe[i] = 0; | |
265 | for (j = 0; j < num_possible_cpus(); j++) | |
266 | smtc_nexttime[i][j] = 0L; | |
267 | } | |
268 | /* | |
269 | * SMTC also can't have the usablility test | |
270 | * run by secondary TCs once Compare is in use. | |
271 | */ | |
272 | if (!c0_compare_int_usable()) | |
273 | return -ENXIO; | |
274 | } | |
275 | ||
276 | /* | |
277 | * With vectored interrupts things are getting platform specific. | |
278 | * get_c0_compare_int is a hook to allow a platform to return the | |
279 | * interrupt number of it's liking. | |
280 | */ | |
281 | irq = MIPS_CPU_IRQ_BASE + cp0_compare_irq; | |
282 | if (get_c0_compare_int) | |
283 | irq = get_c0_compare_int(); | |
284 | ||
285 | cd = &per_cpu(mips_clockevent_device, cpu); | |
286 | ||
287 | cd->name = "MIPS"; | |
288 | cd->features = CLOCK_EVT_FEAT_ONESHOT; | |
289 | ||
290 | /* Calculate the min / max delta */ | |
291 | cd->mult = div_sc((unsigned long) mips_freq, NSEC_PER_SEC, 32); | |
292 | cd->shift = 32; | |
293 | cd->max_delta_ns = clockevent_delta2ns(0x7fffffff, cd); | |
294 | cd->min_delta_ns = clockevent_delta2ns(0x300, cd); | |
295 | ||
296 | cd->rating = 300; | |
297 | cd->irq = irq; | |
320ab2b0 | 298 | cd->cpumask = cpumask_of(cpu); |
8531a35e KK |
299 | cd->set_next_event = mips_next_event; |
300 | cd->set_mode = mips_set_clock_mode; | |
301 | cd->event_handler = mips_event_handler; | |
302 | ||
303 | clockevents_register_device(cd); | |
304 | ||
305 | /* | |
306 | * On SMTC we only want to do the data structure | |
307 | * initialization and IRQ setup once. | |
308 | */ | |
309 | if (cpu) | |
310 | return 0; | |
311 | /* | |
312 | * And we need the hwmask associated with the c0_compare | |
313 | * vector to be initialized. | |
314 | */ | |
315 | irq_hwmask[irq] = (0x100 << cp0_compare_irq); | |
316 | if (cp0_timer_irq_installed) | |
317 | return 0; | |
318 | ||
319 | cp0_timer_irq_installed = 1; | |
320 | ||
321 | setup_irq(irq, &c0_compare_irqaction); | |
322 | ||
323 | return 0; | |
324 | } |