diff options
Diffstat (limited to 'arch/mips/kernel/cevt-smtc.c')
| -rw-r--r-- | arch/mips/kernel/cevt-smtc.c | 321 |
1 files changed, 321 insertions, 0 deletions
diff --git a/arch/mips/kernel/cevt-smtc.c b/arch/mips/kernel/cevt-smtc.c new file mode 100644 index 000000000000..5162fe4b5952 --- /dev/null +++ b/arch/mips/kernel/cevt-smtc.c | |||
| @@ -0,0 +1,321 @@ | |||
| 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> | ||
| 13 | |||
| 14 | #include <asm/smtc_ipi.h> | ||
| 15 | #include <asm/time.h> | ||
| 16 | #include <asm/cevt-r4k.h> | ||
| 17 | |||
| 18 | /* | ||
| 19 | * Variant clock event timer support for SMTC on MIPS 34K, 1004K | ||
| 20 | * or other MIPS MT cores. | ||
| 21 | * | ||
| 22 | * Notes on SMTC Support: | ||
| 23 | * | ||
| 24 | * SMTC has multiple microthread TCs pretending to be Linux CPUs. | ||
| 25 | * But there's only one Count/Compare pair per VPE, and Compare | ||
| 26 | * interrupts are taken opportunisitically by available TCs | ||
| 27 | * bound to the VPE with the Count register. The new timer | ||
| 28 | * framework provides for global broadcasts, but we really | ||
| 29 | * want VPE-level multicasts for best behavior. So instead | ||
| 30 | * of invoking the high-level clock-event broadcast code, | ||
| 31 | * this version of SMTC support uses the historical SMTC | ||
| 32 | * multicast mechanisms "under the hood", appearing to the | ||
| 33 | * generic clock layer as if the interrupts are per-CPU. | ||
| 34 | * | ||
| 35 | * The approach taken here is to maintain a set of NR_CPUS | ||
| 36 | * virtual timers, and track which "CPU" needs to be alerted | ||
| 37 | * at each event. | ||
| 38 | * | ||
| 39 | * It's unlikely that we'll see a MIPS MT core with more than | ||
| 40 | * 2 VPEs, but we *know* that we won't need to handle more | ||
| 41 | * VPEs than we have "CPUs". So NCPUs arrays of NCPUs elements | ||
| 42 | * is always going to be overkill, but always going to be enough. | ||
| 43 | */ | ||
| 44 | |||
| 45 | unsigned long smtc_nexttime[NR_CPUS][NR_CPUS]; | ||
| 46 | static int smtc_nextinvpe[NR_CPUS]; | ||
| 47 | |||
| 48 | /* | ||
| 49 | * Timestamps stored are absolute values to be programmed | ||
| 50 | * into Count register. Valid timestamps will never be zero. | ||
| 51 | * If a Zero Count value is actually calculated, it is converted | ||
| 52 | * to be a 1, which will introduce 1 or two CPU cycles of error | ||
| 53 | * roughly once every four billion events, which at 1000 HZ means | ||
| 54 | * about once every 50 days. If that's actually a problem, one | ||
| 55 | * could alternate squashing 0 to 1 and to -1. | ||
| 56 | */ | ||
| 57 | |||
| 58 | #define MAKEVALID(x) (((x) == 0L) ? 1L : (x)) | ||
| 59 | #define ISVALID(x) ((x) != 0L) | ||
| 60 | |||
| 61 | /* | ||
| 62 | * Time comparison is subtle, as it's really truncated | ||
| 63 | * modular arithmetic. | ||
| 64 | */ | ||
| 65 | |||
| 66 | #define IS_SOONER(a, b, reference) \ | ||
| 67 | (((a) - (unsigned long)(reference)) < ((b) - (unsigned long)(reference))) | ||
| 68 | |||
| 69 | /* | ||
| 70 | * CATCHUP_INCREMENT, used when the function falls behind the counter. | ||
| 71 | * Could be an increasing function instead of a constant; | ||
| 72 | */ | ||
| 73 | |||
| 74 | #define CATCHUP_INCREMENT 64 | ||
| 75 | |||
| 76 | static int mips_next_event(unsigned long delta, | ||
| 77 | struct clock_event_device *evt) | ||
| 78 | { | ||
| 79 | unsigned long flags; | ||
| 80 | unsigned int mtflags; | ||
| 81 | unsigned long timestamp, reference, previous; | ||
| 82 | unsigned long nextcomp = 0L; | ||
| 83 | int vpe = current_cpu_data.vpe_id; | ||
| 84 | int cpu = smp_processor_id(); | ||
| 85 | local_irq_save(flags); | ||
| 86 | mtflags = dmt(); | ||
| 87 | |||
| 88 | /* | ||
| 89 | * Maintain the per-TC virtual timer | ||
| 90 | * and program the per-VPE shared Count register | ||
| 91 | * as appropriate here... | ||
| 92 | */ | ||
| 93 | reference = (unsigned long)read_c0_count(); | ||
| 94 | timestamp = MAKEVALID(reference + delta); | ||
| 95 | /* | ||
| 96 | * To really model the clock, we have to catch the case | ||
| 97 | * where the current next-in-VPE timestamp is the old | ||
| 98 | * timestamp for the calling CPE, but the new value is | ||
| 99 | * in fact later. In that case, we have to do a full | ||
| 100 | * scan and discover the new next-in-VPE CPU id and | ||
| 101 | * timestamp. | ||
| 102 | */ | ||
| 103 | previous = smtc_nexttime[vpe][cpu]; | ||
| 104 | if (cpu == smtc_nextinvpe[vpe] && ISVALID(previous) | ||
| 105 | && IS_SOONER(previous, timestamp, reference)) { | ||
| 106 | int i; | ||
| 107 | int soonest = cpu; | ||
| 108 | |||
| 109 | /* | ||
| 110 | * Update timestamp array here, so that new | ||
| 111 | * value gets considered along with those of | ||
| 112 | * other virtual CPUs on the VPE. | ||
| 113 | */ | ||
| 114 | smtc_nexttime[vpe][cpu] = timestamp; | ||
| 115 | for_each_online_cpu(i) { | ||
| 116 | if (ISVALID(smtc_nexttime[vpe][i]) | ||
| 117 | && IS_SOONER(smtc_nexttime[vpe][i], | ||
| 118 | smtc_nexttime[vpe][soonest], reference)) { | ||
| 119 | soonest = i; | ||
| 120 | } | ||
| 121 | } | ||
| 122 | smtc_nextinvpe[vpe] = soonest; | ||
| 123 | nextcomp = smtc_nexttime[vpe][soonest]; | ||
| 124 | /* | ||
| 125 | * Otherwise, we don't have to process the whole array rank, | ||
| 126 | * we just have to see if the event horizon has gotten closer. | ||
| 127 | */ | ||
| 128 | } else { | ||
| 129 | if (!ISVALID(smtc_nexttime[vpe][smtc_nextinvpe[vpe]]) || | ||
| 130 | IS_SOONER(timestamp, | ||
| 131 | smtc_nexttime[vpe][smtc_nextinvpe[vpe]], reference)) { | ||
| 132 | smtc_nextinvpe[vpe] = cpu; | ||
| 133 | nextcomp = timestamp; | ||
| 134 | } | ||
| 135 | /* | ||
| 136 | * Since next-in-VPE may me the same as the executing | ||
| 137 | * virtual CPU, we update the array *after* checking | ||
| 138 | * its value. | ||
| 139 | */ | ||
| 140 | smtc_nexttime[vpe][cpu] = timestamp; | ||
| 141 | } | ||
| 142 | |||
| 143 | /* | ||
| 144 | * It may be that, in fact, we don't need to update Compare, | ||
| 145 | * but if we do, we want to make sure we didn't fall into | ||
| 146 | * a crack just behind Count. | ||
| 147 | */ | ||
| 148 | if (ISVALID(nextcomp)) { | ||
| 149 | write_c0_compare(nextcomp); | ||
| 150 | ehb(); | ||
| 151 | /* | ||
| 152 | * We never return an error, we just make sure | ||
| 153 | * that we trigger the handlers as quickly as | ||
| 154 | * we can if we fell behind. | ||
| 155 | */ | ||
| 156 | while ((nextcomp - (unsigned long)read_c0_count()) | ||
| 157 | > (unsigned long)LONG_MAX) { | ||
| 158 | nextcomp += CATCHUP_INCREMENT; | ||
| 159 | write_c0_compare(nextcomp); | ||
| 160 | ehb(); | ||
| 161 | } | ||
| 162 | } | ||
| 163 | emt(mtflags); | ||
| 164 | local_irq_restore(flags); | ||
| 165 | return 0; | ||
| 166 | } | ||
| 167 | |||
| 168 | |||
| 169 | void smtc_distribute_timer(int vpe) | ||
| 170 | { | ||
| 171 | unsigned long flags; | ||
| 172 | unsigned int mtflags; | ||
| 173 | int cpu; | ||
| 174 | struct clock_event_device *cd; | ||
| 175 | unsigned long nextstamp = 0L; | ||
| 176 | unsigned long reference; | ||
| 177 | |||
| 178 | |||
| 179 | repeat: | ||
| 180 | for_each_online_cpu(cpu) { | ||
| 181 | /* | ||
| 182 | * Find virtual CPUs within the current VPE who have | ||
| 183 | * unserviced timer requests whose time is now past. | ||
| 184 | */ | ||
| 185 | local_irq_save(flags); | ||
| 186 | mtflags = dmt(); | ||
| 187 | if (cpu_data[cpu].vpe_id == vpe && | ||
| 188 | ISVALID(smtc_nexttime[vpe][cpu])) { | ||
| 189 | reference = (unsigned long)read_c0_count(); | ||
| 190 | if ((smtc_nexttime[vpe][cpu] - reference) | ||
| 191 | > (unsigned long)LONG_MAX) { | ||
| 192 | smtc_nexttime[vpe][cpu] = 0L; | ||
| 193 | emt(mtflags); | ||
| 194 | local_irq_restore(flags); | ||
| 195 | /* | ||
| 196 | * We don't send IPIs to ourself. | ||
| 197 | */ | ||
| 198 | if (cpu != smp_processor_id()) { | ||
| 199 | smtc_send_ipi(cpu, SMTC_CLOCK_TICK, 0); | ||
| 200 | } else { | ||
| 201 | cd = &per_cpu(mips_clockevent_device, cpu); | ||
| 202 | cd->event_handler(cd); | ||
| 203 | } | ||
| 204 | } else { | ||
| 205 | /* Local to VPE but Valid Time not yet reached. */ | ||
| 206 | if (!ISVALID(nextstamp) || | ||
| 207 | IS_SOONER(smtc_nexttime[vpe][cpu], nextstamp, | ||
| 208 | reference)) { | ||
| 209 | smtc_nextinvpe[vpe] = cpu; | ||
| 210 | nextstamp = smtc_nexttime[vpe][cpu]; | ||
| 211 | } | ||
| 212 | emt(mtflags); | ||
| 213 | local_irq_restore(flags); | ||
| 214 | } | ||
| 215 | } else { | ||
| 216 | emt(mtflags); | ||
| 217 | local_irq_restore(flags); | ||
| 218 | |||
| 219 | } | ||
| 220 | } | ||
| 221 | /* Reprogram for interrupt at next soonest timestamp for VPE */ | ||
| 222 | if (ISVALID(nextstamp)) { | ||
| 223 | write_c0_compare(nextstamp); | ||
| 224 | ehb(); | ||
| 225 | if ((nextstamp - (unsigned long)read_c0_count()) | ||
| 226 | > (unsigned long)LONG_MAX) | ||
| 227 | goto repeat; | ||
| 228 | } | ||
| 229 | } | ||
| 230 | |||
| 231 | |||
| 232 | irqreturn_t c0_compare_interrupt(int irq, void *dev_id) | ||
| 233 | { | ||
| 234 | int cpu = smp_processor_id(); | ||
| 235 | |||
| 236 | /* If we're running SMTC, we've got MIPS MT and therefore MIPS32R2 */ | ||
| 237 | handle_perf_irq(1); | ||
| 238 | |||
| 239 | if (read_c0_cause() & (1 << 30)) { | ||
| 240 | /* Clear Count/Compare Interrupt */ | ||
| 241 | write_c0_compare(read_c0_compare()); | ||
| 242 | smtc_distribute_timer(cpu_data[cpu].vpe_id); | ||
| 243 | } | ||
| 244 | return IRQ_HANDLED; | ||
| 245 | } | ||
| 246 | |||
| 247 | |||
| 248 | int __cpuinit mips_clockevent_init(void) | ||
| 249 | { | ||
| 250 | uint64_t mips_freq = mips_hpt_frequency; | ||
| 251 | unsigned int cpu = smp_processor_id(); | ||
| 252 | struct clock_event_device *cd; | ||
| 253 | unsigned int irq; | ||
| 254 | int i; | ||
| 255 | int j; | ||
| 256 | |||
| 257 | if (!cpu_has_counter || !mips_hpt_frequency) | ||
| 258 | return -ENXIO; | ||
| 259 | if (cpu == 0) { | ||
| 260 | for (i = 0; i < num_possible_cpus(); i++) { | ||
| 261 | smtc_nextinvpe[i] = 0; | ||
| 262 | for (j = 0; j < num_possible_cpus(); j++) | ||
| 263 | smtc_nexttime[i][j] = 0L; | ||
| 264 | } | ||
| 265 | /* | ||
| 266 | * SMTC also can't have the usablility test | ||
| 267 | * run by secondary TCs once Compare is in use. | ||
| 268 | */ | ||
| 269 | if (!c0_compare_int_usable()) | ||
| 270 | return -ENXIO; | ||
| 271 | } | ||
| 272 | |||
| 273 | /* | ||
| 274 | * With vectored interrupts things are getting platform specific. | ||
| 275 | * get_c0_compare_int is a hook to allow a platform to return the | ||
| 276 | * interrupt number of it's liking. | ||
| 277 | */ | ||
| 278 | irq = MIPS_CPU_IRQ_BASE + cp0_compare_irq; | ||
| 279 | if (get_c0_compare_int) | ||
| 280 | irq = get_c0_compare_int(); | ||
| 281 | |||
| 282 | cd = &per_cpu(mips_clockevent_device, cpu); | ||
| 283 | |||
| 284 | cd->name = "MIPS"; | ||
| 285 | cd->features = CLOCK_EVT_FEAT_ONESHOT; | ||
| 286 | |||
| 287 | /* Calculate the min / max delta */ | ||
| 288 | cd->mult = div_sc((unsigned long) mips_freq, NSEC_PER_SEC, 32); | ||
| 289 | cd->shift = 32; | ||
| 290 | cd->max_delta_ns = clockevent_delta2ns(0x7fffffff, cd); | ||
| 291 | cd->min_delta_ns = clockevent_delta2ns(0x300, cd); | ||
| 292 | |||
| 293 | cd->rating = 300; | ||
| 294 | cd->irq = irq; | ||
| 295 | cd->cpumask = cpumask_of_cpu(cpu); | ||
| 296 | cd->set_next_event = mips_next_event; | ||
| 297 | cd->set_mode = mips_set_clock_mode; | ||
| 298 | cd->event_handler = mips_event_handler; | ||
| 299 | |||
| 300 | clockevents_register_device(cd); | ||
| 301 | |||
| 302 | /* | ||
| 303 | * On SMTC we only want to do the data structure | ||
| 304 | * initialization and IRQ setup once. | ||
| 305 | */ | ||
| 306 | if (cpu) | ||
| 307 | return 0; | ||
| 308 | /* | ||
| 309 | * And we need the hwmask associated with the c0_compare | ||
| 310 | * vector to be initialized. | ||
| 311 | */ | ||
| 312 | irq_hwmask[irq] = (0x100 << cp0_compare_irq); | ||
| 313 | if (cp0_timer_irq_installed) | ||
| 314 | return 0; | ||
| 315 | |||
| 316 | cp0_timer_irq_installed = 1; | ||
| 317 | |||
| 318 | setup_irq(irq, &c0_compare_irqaction); | ||
| 319 | |||
| 320 | return 0; | ||
| 321 | } | ||
