Print this page
XXXX cpudrv attach error handling is leaky
XXXX cpudrv attach is racy
Split |
Close |
Expand all |
Collapse all |
--- old/usr/src/uts/common/io/cpudrv.c
+++ new/usr/src/uts/common/io/cpudrv.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 25 /*
26 26 * Copyright (c) 2009, Intel Corporation.
27 27 * All Rights Reserved.
28 28 */
29 29
30 30 /*
31 31 * CPU Device driver. The driver is not DDI-compliant.
32 32 *
33 33 * The driver supports following features:
34 34 * - Power management.
35 35 */
36 36
37 37 #include <sys/types.h>
38 38 #include <sys/param.h>
39 39 #include <sys/errno.h>
40 40 #include <sys/modctl.h>
41 41 #include <sys/kmem.h>
42 42 #include <sys/conf.h>
43 43 #include <sys/cmn_err.h>
44 44 #include <sys/stat.h>
45 45 #include <sys/debug.h>
46 46 #include <sys/systm.h>
47 47 #include <sys/ddi.h>
48 48 #include <sys/sunddi.h>
49 49 #include <sys/sdt.h>
50 50 #include <sys/epm.h>
51 51 #include <sys/machsystm.h>
52 52 #include <sys/x_call.h>
53 53 #include <sys/cpudrv_mach.h>
54 54 #include <sys/msacct.h>
55 55
56 56 /*
57 57 * CPU power management
58 58 *
59 59 * The supported power saving model is to slow down the CPU (on SPARC by
60 60 * dividing the CPU clock and on x86 by dropping down a P-state).
61 61 * Periodically we determine the amount of time the CPU is running
62 62 * idle thread and threads in user mode during the last quantum. If the idle
63 63 * thread was running less than its low water mark for current speed for
64 64 * number of consecutive sampling periods, or number of running threads in
65 65 * user mode are above its high water mark, we arrange to go to the higher
66 66 * speed. If the idle thread was running more than its high water mark without
67 67 * dropping a number of consecutive times below the mark, and number of threads
68 68 * running in user mode are below its low water mark, we arrange to go to the
69 69 * next lower speed. While going down, we go through all the speeds. While
70 70 * going up we go to the maximum speed to minimize impact on the user, but have
71 71 * provisions in the driver to go to other speeds.
72 72 *
73 73 * The driver does not have knowledge of a particular implementation of this
74 74 * scheme and will work with all CPUs supporting this model. On SPARC, the
75 75 * driver determines supported speeds by looking at 'clock-divisors' property
76 76 * created by OBP. On x86, the driver retrieves the supported speeds from
77 77 * ACPI.
78 78 */
79 79
80 80 /*
81 81 * Configuration function prototypes and data structures
82 82 */
83 83 static int cpudrv_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
84 84 static int cpudrv_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
85 85 static int cpudrv_power(dev_info_t *dip, int comp, int level);
86 86
87 87 struct dev_ops cpudrv_ops = {
88 88 DEVO_REV, /* rev */
89 89 0, /* refcnt */
90 90 nodev, /* getinfo */
91 91 nulldev, /* identify */
92 92 nulldev, /* probe */
93 93 cpudrv_attach, /* attach */
94 94 cpudrv_detach, /* detach */
95 95 nodev, /* reset */
96 96 (struct cb_ops *)NULL, /* cb_ops */
97 97 (struct bus_ops *)NULL, /* bus_ops */
98 98 cpudrv_power, /* power */
99 99 ddi_quiesce_not_needed, /* quiesce */
100 100 };
101 101
102 102 static struct modldrv modldrv = {
103 103 &mod_driverops, /* modops */
104 104 "CPU Driver", /* linkinfo */
105 105 &cpudrv_ops, /* dev_ops */
106 106 };
107 107
108 108 static struct modlinkage modlinkage = {
109 109 MODREV_1, /* rev */
110 110 &modldrv, /* linkage */
111 111 NULL
112 112 };
113 113
114 114 /*
115 115 * Function prototypes
116 116 */
117 117 static int cpudrv_init(cpudrv_devstate_t *cpudsp);
118 118 static void cpudrv_free(cpudrv_devstate_t *cpudsp);
119 119 static int cpudrv_comp_create(cpudrv_devstate_t *cpudsp);
120 120 static void cpudrv_monitor_disp(void *arg);
121 121 static void cpudrv_monitor(void *arg);
122 122
123 123 /*
124 124 * Driver global variables
125 125 */
126 126 uint_t cpudrv_debug = 0;
127 127 void *cpudrv_state;
128 128 static uint_t cpudrv_idle_hwm = CPUDRV_IDLE_HWM;
129 129 static uint_t cpudrv_idle_lwm = CPUDRV_IDLE_LWM;
130 130 static uint_t cpudrv_idle_buf_zone = CPUDRV_IDLE_BUF_ZONE;
131 131 static uint_t cpudrv_idle_bhwm_cnt_max = CPUDRV_IDLE_BHWM_CNT_MAX;
132 132 static uint_t cpudrv_idle_blwm_cnt_max = CPUDRV_IDLE_BLWM_CNT_MAX;
133 133 static uint_t cpudrv_user_hwm = CPUDRV_USER_HWM;
134 134
135 135 boolean_t cpudrv_enabled = B_TRUE;
136 136
137 137 /*
138 138 * cpudrv_direct_pm allows user applications to directly control the
139 139 * power state transitions (direct pm) without following the normal
140 140 * direct pm protocol. This is needed because the normal protocol
141 141 * requires that a device only be lowered when it is idle, and be
142 142 * brought up when it request to do so by calling pm_raise_power().
143 143 * Ignoring this protocol is harmless for CPU (other than speed).
144 144 * Moreover it might be the case that CPU is never idle or wants
145 145 * to be at higher speed because of the addition CPU cycles required
146 146 * to run the user application.
147 147 *
148 148 * The driver will still report idle/busy status to the framework. Although
149 149 * framework will ignore this information for direct pm devices and not
150 150 * try to bring them down when idle, user applications can still use this
151 151 * information if they wants.
152 152 *
153 153 * In the future, provide an ioctl to control setting of this mode. In
154 154 * that case, this variable should move to the state structure and
155 155 * be protected by the lock in the state structure.
156 156 */
157 157 int cpudrv_direct_pm = 0;
158 158
159 159 /*
160 160 * Arranges for the handler function to be called at the interval suitable
161 161 * for current speed.
162 162 */
163 163 #define CPUDRV_MONITOR_INIT(cpudsp) { \
164 164 if (cpudrv_is_enabled(cpudsp)) { \
165 165 ASSERT(mutex_owned(&(cpudsp)->lock)); \
166 166 (cpudsp)->cpudrv_pm.timeout_id = \
167 167 timeout(cpudrv_monitor_disp, \
168 168 (cpudsp), (((cpudsp)->cpudrv_pm.cur_spd == NULL) ? \
169 169 CPUDRV_QUANT_CNT_OTHR : \
170 170 (cpudsp)->cpudrv_pm.cur_spd->quant_cnt)); \
171 171 } \
172 172 }
173 173
174 174 /*
175 175 * Arranges for the handler function not to be called back.
176 176 */
177 177 #define CPUDRV_MONITOR_FINI(cpudsp) { \
178 178 timeout_id_t tmp_tid; \
179 179 ASSERT(mutex_owned(&(cpudsp)->lock)); \
180 180 tmp_tid = (cpudsp)->cpudrv_pm.timeout_id; \
181 181 (cpudsp)->cpudrv_pm.timeout_id = 0; \
182 182 mutex_exit(&(cpudsp)->lock); \
183 183 if (tmp_tid != 0) { \
184 184 (void) untimeout(tmp_tid); \
185 185 mutex_enter(&(cpudsp)->cpudrv_pm.timeout_lock); \
186 186 while ((cpudsp)->cpudrv_pm.timeout_count != 0) \
187 187 cv_wait(&(cpudsp)->cpudrv_pm.timeout_cv, \
188 188 &(cpudsp)->cpudrv_pm.timeout_lock); \
189 189 mutex_exit(&(cpudsp)->cpudrv_pm.timeout_lock); \
190 190 } \
191 191 mutex_enter(&(cpudsp)->lock); \
192 192 }
193 193
194 194 int
195 195 _init(void)
196 196 {
197 197 int error;
198 198
199 199 DPRINTF(D_INIT, (" _init: function called\n"));
200 200 if ((error = ddi_soft_state_init(&cpudrv_state,
201 201 sizeof (cpudrv_devstate_t), 0)) != 0) {
202 202 return (error);
203 203 }
204 204
205 205 if ((error = mod_install(&modlinkage)) != 0) {
206 206 ddi_soft_state_fini(&cpudrv_state);
207 207 }
208 208
209 209 /*
210 210 * Callbacks used by the PPM driver.
211 211 */
212 212 CPUDRV_SET_PPM_CALLBACKS();
213 213 return (error);
214 214 }
215 215
216 216 int
217 217 _fini(void)
218 218 {
219 219 int error;
220 220
221 221 DPRINTF(D_FINI, (" _fini: function called\n"));
222 222 if ((error = mod_remove(&modlinkage)) == 0) {
223 223 ddi_soft_state_fini(&cpudrv_state);
224 224 }
225 225
226 226 return (error);
227 227 }
228 228
229 229 int
230 230 _info(struct modinfo *modinfop)
231 231 {
232 232 return (mod_info(&modlinkage, modinfop));
233 233 }
234 234
235 235 /*
236 236 * Driver attach(9e) entry point.
237 237 */
238 238 static int
239 239 cpudrv_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
240 240 {
241 241 int instance;
242 242 cpudrv_devstate_t *cpudsp;
243 243
244 244 instance = ddi_get_instance(dip);
245 245
246 246 switch (cmd) {
247 247 case DDI_ATTACH:
248 248 DPRINTF(D_ATTACH, ("cpudrv_attach: instance %d: "
249 249 "DDI_ATTACH called\n", instance));
250 250 if (!cpudrv_is_enabled(NULL))
251 251 return (DDI_FAILURE);
252 252 if (ddi_soft_state_zalloc(cpudrv_state, instance) !=
253 253 DDI_SUCCESS) {
254 254 cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
255 255 "can't allocate state", instance);
256 256 cpudrv_enabled = B_FALSE;
257 257 return (DDI_FAILURE);
258 258 }
259 259 if ((cpudsp = ddi_get_soft_state(cpudrv_state, instance)) ==
260 260 NULL) {
261 261 cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
262 262 "can't get state", instance);
263 263 ddi_soft_state_free(cpudrv_state, instance);
264 264 cpudrv_enabled = B_FALSE;
265 265 return (DDI_FAILURE);
266 266 }
267 267 cpudsp->dip = dip;
268 268
269 269 /*
↓ open down ↓ |
269 lines elided |
↑ open up ↑ |
270 270 * Find CPU number for this dev_info node.
271 271 */
272 272 if (!cpudrv_get_cpu_id(dip, &(cpudsp->cpu_id))) {
273 273 cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
274 274 "can't convert dip to cpu_id", instance);
275 275 ddi_soft_state_free(cpudrv_state, instance);
276 276 cpudrv_enabled = B_FALSE;
277 277 return (DDI_FAILURE);
278 278 }
279 279
280 - mutex_init(&cpudsp->lock, NULL, MUTEX_DRIVER, NULL);
281 - if (cpudrv_is_enabled(cpudsp)) {
282 - if (cpudrv_init(cpudsp) != DDI_SUCCESS) {
283 - cpudrv_enabled = B_FALSE;
284 - cpudrv_free(cpudsp);
285 - ddi_soft_state_free(cpudrv_state, instance);
286 - return (DDI_FAILURE);
287 - }
288 - if (cpudrv_comp_create(cpudsp) != DDI_SUCCESS) {
289 - cpudrv_enabled = B_FALSE;
290 - cpudrv_free(cpudsp);
291 - ddi_soft_state_free(cpudrv_state, instance);
292 - return (DDI_FAILURE);
293 - }
294 - if (ddi_prop_update_string(DDI_DEV_T_NONE,
295 - dip, "pm-class", "CPU") != DDI_PROP_SUCCESS) {
296 - cpudrv_enabled = B_FALSE;
297 - cpudrv_free(cpudsp);
298 - ddi_soft_state_free(cpudrv_state, instance);
299 - return (DDI_FAILURE);
300 - }
280 + if (!cpudrv_is_enabled(cpudsp)) {
281 + cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
282 + "not supported or it got disabled on us",
283 + instance);
284 + cpudrv_enabled = B_FALSE;
285 + ddi_soft_state_free(cpudrv_state, instance);
286 + return (DDI_FAILURE);
287 + }
301 288
302 - /*
303 - * Taskq is used to dispatch routine to monitor CPU
304 - * activities.
305 - */
306 - cpudsp->cpudrv_pm.tq = ddi_taskq_create(dip,
307 - "cpudrv_monitor", CPUDRV_TASKQ_THREADS,
308 - TASKQ_DEFAULTPRI, 0);
289 + mutex_init(&cpudsp->lock, NULL, MUTEX_DRIVER, NULL);
290 + if (cpudrv_init(cpudsp) != DDI_SUCCESS) {
291 + cpudrv_enabled = B_FALSE;
292 + cpudrv_free(cpudsp);
293 + ddi_soft_state_free(cpudrv_state, instance);
294 + return (DDI_FAILURE);
295 + }
296 + if (cpudrv_comp_create(cpudsp) != DDI_SUCCESS) {
297 + cpudrv_enabled = B_FALSE;
298 + cpudrv_free(cpudsp);
299 + ddi_soft_state_free(cpudrv_state, instance);
300 + return (DDI_FAILURE);
301 + }
302 + if (ddi_prop_update_string(DDI_DEV_T_NONE,
303 + dip, "pm-class", "CPU") != DDI_PROP_SUCCESS) {
304 + cpudrv_enabled = B_FALSE;
305 + cpudrv_free(cpudsp);
306 + ddi_soft_state_free(cpudrv_state, instance);
307 + return (DDI_FAILURE);
308 + }
309 309
310 - mutex_init(&cpudsp->cpudrv_pm.timeout_lock, NULL,
311 - MUTEX_DRIVER, NULL);
312 - cv_init(&cpudsp->cpudrv_pm.timeout_cv, NULL,
313 - CV_DEFAULT, NULL);
310 + /*
311 + * Taskq is used to dispatch routine to monitor CPU
312 + * activities.
313 + */
314 + cpudsp->cpudrv_pm.tq = ddi_taskq_create(dip,
315 + "cpudrv_monitor", CPUDRV_TASKQ_THREADS,
316 + TASKQ_DEFAULTPRI, 0);
317 + if (cpudsp->cpudrv_pm.tq == NULL) {
318 + cpudrv_enabled = B_FALSE;
319 + cpudrv_free(cpudsp);
320 + ddi_soft_state_free(cpudrv_state, instance);
321 + return (DDI_FAILURE);
322 + }
314 323
315 - /*
316 - * Driver needs to assume that CPU is running at
317 - * unknown speed at DDI_ATTACH and switch it to the
318 - * needed speed. We assume that initial needed speed
319 - * is full speed for us.
320 - */
321 - /*
322 - * We need to take the lock because cpudrv_monitor()
323 - * will start running in parallel with attach().
324 - */
325 - mutex_enter(&cpudsp->lock);
326 - cpudsp->cpudrv_pm.cur_spd = NULL;
327 - cpudsp->cpudrv_pm.pm_started = B_FALSE;
328 - /*
329 - * We don't call pm_raise_power() directly from attach
330 - * because driver attach for a slave CPU node can
331 - * happen before the CPU is even initialized. We just
332 - * start the monitoring system which understands
333 - * unknown speed and moves CPU to top speed when it
334 - * has been initialized.
335 - */
336 - CPUDRV_MONITOR_INIT(cpudsp);
337 - mutex_exit(&cpudsp->lock);
324 + mutex_init(&cpudsp->cpudrv_pm.timeout_lock, NULL,
325 + MUTEX_DRIVER, NULL);
326 + cv_init(&cpudsp->cpudrv_pm.timeout_cv, NULL,
327 + CV_DEFAULT, NULL);
338 328
339 - }
329 + /*
330 + * Driver needs to assume that CPU is running at
331 + * unknown speed at DDI_ATTACH and switch it to the
332 + * needed speed. We assume that initial needed speed
333 + * is full speed for us.
334 + */
335 + /*
336 + * We need to take the lock because cpudrv_monitor()
337 + * will start running in parallel with attach().
338 + */
339 + mutex_enter(&cpudsp->lock);
340 + cpudsp->cpudrv_pm.cur_spd = NULL;
341 + cpudsp->cpudrv_pm.pm_started = B_FALSE;
342 + /*
343 + * We don't call pm_raise_power() directly from attach
344 + * because driver attach for a slave CPU node can
345 + * happen before the CPU is even initialized. We just
346 + * start the monitoring system which understands
347 + * unknown speed and moves CPU to top speed when it
348 + * has been initialized.
349 + */
350 + CPUDRV_MONITOR_INIT(cpudsp);
351 + mutex_exit(&cpudsp->lock);
340 352
341 353 if (!cpudrv_mach_init(cpudsp)) {
342 354 cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
343 355 "cpudrv_mach_init failed", instance);
344 356 cpudrv_enabled = B_FALSE;
357 + ddi_taskq_destroy(cpudsp->cpudrv_pm.tq);
345 358 cpudrv_free(cpudsp);
346 359 ddi_soft_state_free(cpudrv_state, instance);
347 360 return (DDI_FAILURE);
348 361 }
349 362
350 363 CPUDRV_INSTALL_MAX_CHANGE_HANDLER(cpudsp);
351 364
352 365 (void) ddi_prop_update_int(DDI_DEV_T_NONE, dip,
353 366 DDI_NO_AUTODETACH, 1);
354 367 ddi_report_dev(dip);
355 368 return (DDI_SUCCESS);
356 369
357 370 case DDI_RESUME:
358 371 DPRINTF(D_ATTACH, ("cpudrv_attach: instance %d: "
359 372 "DDI_RESUME called\n", instance));
360 373
361 374 cpudsp = ddi_get_soft_state(cpudrv_state, instance);
362 375 ASSERT(cpudsp != NULL);
363 376
364 377 /*
365 378 * Nothing to do for resume, if not doing active PM.
366 379 */
367 380 if (!cpudrv_is_enabled(cpudsp))
368 381 return (DDI_SUCCESS);
369 382
370 383 mutex_enter(&cpudsp->lock);
371 384 /*
372 385 * Driver needs to assume that CPU is running at unknown speed
373 386 * at DDI_RESUME and switch it to the needed speed. We assume
374 387 * that the needed speed is full speed for us.
375 388 */
376 389 cpudsp->cpudrv_pm.cur_spd = NULL;
377 390 CPUDRV_MONITOR_INIT(cpudsp);
378 391 mutex_exit(&cpudsp->lock);
379 392 CPUDRV_REDEFINE_TOPSPEED(dip);
380 393 return (DDI_SUCCESS);
381 394
382 395 default:
383 396 return (DDI_FAILURE);
384 397 }
385 398 }
386 399
387 400 /*
388 401 * Driver detach(9e) entry point.
389 402 */
390 403 static int
391 404 cpudrv_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
392 405 {
393 406 int instance;
394 407 cpudrv_devstate_t *cpudsp;
395 408 cpudrv_pm_t *cpupm;
396 409
397 410 instance = ddi_get_instance(dip);
398 411
399 412 switch (cmd) {
400 413 case DDI_DETACH:
401 414 DPRINTF(D_DETACH, ("cpudrv_detach: instance %d: "
402 415 "DDI_DETACH called\n", instance));
403 416
404 417 #if defined(__x86)
405 418 cpudsp = ddi_get_soft_state(cpudrv_state, instance);
406 419 ASSERT(cpudsp != NULL);
407 420
408 421 /*
409 422 * Nothing to do for detach, if no doing active PM.
410 423 */
411 424 if (!cpudrv_is_enabled(cpudsp))
412 425 return (DDI_SUCCESS);
413 426
414 427 /*
415 428 * uninstall PPC/_TPC change notification handler
416 429 */
417 430 CPUDRV_UNINSTALL_MAX_CHANGE_HANDLER(cpudsp);
418 431
419 432 /*
420 433 * destruct platform specific resource
421 434 */
422 435 if (!cpudrv_mach_fini(cpudsp))
423 436 return (DDI_FAILURE);
424 437
425 438 mutex_enter(&cpudsp->lock);
426 439 CPUDRV_MONITOR_FINI(cpudsp);
427 440 cv_destroy(&cpudsp->cpudrv_pm.timeout_cv);
428 441 mutex_destroy(&cpudsp->cpudrv_pm.timeout_lock);
429 442 ddi_taskq_destroy(cpudsp->cpudrv_pm.tq);
430 443 cpudrv_free(cpudsp);
431 444 mutex_exit(&cpudsp->lock);
432 445 mutex_destroy(&cpudsp->lock);
433 446 ddi_soft_state_free(cpudrv_state, instance);
434 447 (void) ddi_prop_update_int(DDI_DEV_T_NONE, dip,
435 448 DDI_NO_AUTODETACH, 0);
436 449 return (DDI_SUCCESS);
437 450
438 451 #else
439 452 /*
440 453 * If the only thing supported by the driver is power
441 454 * management, we can in future enhance the driver and
442 455 * framework that loads it to unload the driver when
443 456 * user has disabled CPU power management.
444 457 */
445 458 return (DDI_FAILURE);
446 459 #endif
447 460
448 461 case DDI_SUSPEND:
449 462 DPRINTF(D_DETACH, ("cpudrv_detach: instance %d: "
450 463 "DDI_SUSPEND called\n", instance));
451 464
452 465 cpudsp = ddi_get_soft_state(cpudrv_state, instance);
453 466 ASSERT(cpudsp != NULL);
454 467
455 468 /*
456 469 * Nothing to do for suspend, if not doing active PM.
457 470 */
458 471 if (!cpudrv_is_enabled(cpudsp))
459 472 return (DDI_SUCCESS);
460 473
461 474 /*
462 475 * During a checkpoint-resume sequence, framework will
463 476 * stop interrupts to quiesce kernel activity. This will
464 477 * leave our monitoring system ineffective. Handle this
465 478 * by stopping our monitoring system and bringing CPU
466 479 * to full speed. In case we are in special direct pm
467 480 * mode, we leave the CPU at whatever speed it is. This
468 481 * is harmless other than speed.
469 482 */
470 483 mutex_enter(&cpudsp->lock);
471 484 cpupm = &(cpudsp->cpudrv_pm);
472 485
473 486 DPRINTF(D_DETACH, ("cpudrv_detach: instance %d: DDI_SUSPEND - "
474 487 "cur_spd %d, topspeed %d\n", instance,
475 488 cpupm->cur_spd->pm_level,
476 489 CPUDRV_TOPSPEED(cpupm)->pm_level));
477 490
478 491 CPUDRV_MONITOR_FINI(cpudsp);
479 492
480 493 if (!cpudrv_direct_pm && (cpupm->cur_spd !=
481 494 CPUDRV_TOPSPEED(cpupm))) {
482 495 if (cpupm->pm_busycnt < 1) {
483 496 if ((pm_busy_component(dip, CPUDRV_COMP_NUM)
484 497 == DDI_SUCCESS)) {
485 498 cpupm->pm_busycnt++;
486 499 } else {
487 500 CPUDRV_MONITOR_INIT(cpudsp);
488 501 mutex_exit(&cpudsp->lock);
489 502 cmn_err(CE_WARN, "cpudrv_detach: "
490 503 "instance %d: can't busy CPU "
491 504 "component", instance);
492 505 return (DDI_FAILURE);
493 506 }
494 507 }
495 508 mutex_exit(&cpudsp->lock);
496 509 if (pm_raise_power(dip, CPUDRV_COMP_NUM,
497 510 CPUDRV_TOPSPEED(cpupm)->pm_level) !=
498 511 DDI_SUCCESS) {
499 512 mutex_enter(&cpudsp->lock);
500 513 CPUDRV_MONITOR_INIT(cpudsp);
501 514 mutex_exit(&cpudsp->lock);
502 515 cmn_err(CE_WARN, "cpudrv_detach: instance %d: "
503 516 "can't raise CPU power level to %d",
504 517 instance,
505 518 CPUDRV_TOPSPEED(cpupm)->pm_level);
506 519 return (DDI_FAILURE);
507 520 } else {
508 521 return (DDI_SUCCESS);
509 522 }
510 523 } else {
511 524 mutex_exit(&cpudsp->lock);
512 525 return (DDI_SUCCESS);
513 526 }
514 527
515 528 default:
516 529 return (DDI_FAILURE);
517 530 }
518 531 }
519 532
520 533 /*
521 534 * Driver power(9e) entry point.
522 535 *
523 536 * Driver's notion of current power is set *only* in power(9e) entry point
524 537 * after actual power change operation has been successfully completed.
525 538 */
526 539 /* ARGSUSED */
527 540 static int
528 541 cpudrv_power(dev_info_t *dip, int comp, int level)
529 542 {
530 543 int instance;
531 544 cpudrv_devstate_t *cpudsp;
532 545 cpudrv_pm_t *cpudrvpm;
533 546 cpudrv_pm_spd_t *new_spd;
534 547 boolean_t is_ready;
535 548 int ret;
536 549
537 550 instance = ddi_get_instance(dip);
538 551
539 552 DPRINTF(D_POWER, ("cpudrv_power: instance %d: level %d\n",
540 553 instance, level));
541 554
542 555 if ((cpudsp = ddi_get_soft_state(cpudrv_state, instance)) == NULL) {
543 556 cmn_err(CE_WARN, "cpudrv_power: instance %d: can't "
544 557 "get state", instance);
545 558 return (DDI_FAILURE);
546 559 }
547 560
548 561 /*
549 562 * We're not ready until we can get a cpu_t
550 563 */
551 564 is_ready = (cpudrv_get_cpu(cpudsp) == DDI_SUCCESS);
552 565
553 566 mutex_enter(&cpudsp->lock);
554 567 cpudrvpm = &(cpudsp->cpudrv_pm);
555 568
556 569 /*
557 570 * In normal operation, we fail if we are busy and request is
558 571 * to lower the power level. We let this go through if the driver
559 572 * is in special direct pm mode. On x86, we also let this through
560 573 * if the change is due to a request to govern the max speed.
561 574 */
562 575 if (!cpudrv_direct_pm && (cpudrvpm->pm_busycnt >= 1) &&
563 576 !cpudrv_is_governor_thread(cpudrvpm)) {
564 577 if ((cpudrvpm->cur_spd != NULL) &&
565 578 (level < cpudrvpm->cur_spd->pm_level)) {
566 579 mutex_exit(&cpudsp->lock);
567 580 return (DDI_FAILURE);
568 581 }
569 582 }
570 583
571 584 for (new_spd = cpudrvpm->head_spd; new_spd; new_spd =
572 585 new_spd->down_spd) {
573 586 if (new_spd->pm_level == level)
574 587 break;
575 588 }
576 589 if (!new_spd) {
577 590 CPUDRV_RESET_GOVERNOR_THREAD(cpudrvpm);
578 591 mutex_exit(&cpudsp->lock);
579 592 cmn_err(CE_WARN, "cpudrv_power: instance %d: "
580 593 "can't locate new CPU speed", instance);
581 594 return (DDI_FAILURE);
582 595 }
583 596
584 597 /*
585 598 * We currently refuse to power manage if the CPU is not ready to
586 599 * take cross calls (cross calls fail silently if CPU is not ready
587 600 * for it).
588 601 *
589 602 * Additionally, for x86 platforms we cannot power manage an instance,
590 603 * until it has been initialized.
591 604 */
592 605 if (is_ready) {
593 606 is_ready = CPUDRV_XCALL_IS_READY(cpudsp->cpu_id);
594 607 if (!is_ready) {
595 608 DPRINTF(D_POWER, ("cpudrv_power: instance %d: "
596 609 "CPU not ready for x-calls\n", instance));
597 610 } else if (!(is_ready = cpudrv_power_ready(cpudsp->cp))) {
598 611 DPRINTF(D_POWER, ("cpudrv_power: instance %d: "
599 612 "waiting for all CPUs to be power manageable\n",
600 613 instance));
601 614 }
602 615 }
603 616 if (!is_ready) {
604 617 CPUDRV_RESET_GOVERNOR_THREAD(cpudrvpm);
605 618 mutex_exit(&cpudsp->lock);
606 619 return (DDI_FAILURE);
607 620 }
608 621
609 622 /*
610 623 * Execute CPU specific routine on the requested CPU to
611 624 * change its speed to normal-speed/divisor.
612 625 */
613 626 if ((ret = cpudrv_change_speed(cpudsp, new_spd)) != DDI_SUCCESS) {
614 627 cmn_err(CE_WARN, "cpudrv_power: "
615 628 "cpudrv_change_speed() return = %d", ret);
616 629 mutex_exit(&cpudsp->lock);
617 630 return (DDI_FAILURE);
618 631 }
619 632
620 633 /*
621 634 * Reset idle threshold time for the new power level.
622 635 */
623 636 if ((cpudrvpm->cur_spd != NULL) && (level <
624 637 cpudrvpm->cur_spd->pm_level)) {
625 638 if (pm_idle_component(dip, CPUDRV_COMP_NUM) ==
626 639 DDI_SUCCESS) {
627 640 if (cpudrvpm->pm_busycnt >= 1)
628 641 cpudrvpm->pm_busycnt--;
629 642 } else {
630 643 cmn_err(CE_WARN, "cpudrv_power: instance %d: "
631 644 "can't idle CPU component",
632 645 ddi_get_instance(dip));
633 646 }
634 647 }
635 648 /*
636 649 * Reset various parameters because we are now running at new speed.
637 650 */
638 651 cpudrvpm->lastquan_mstate[CMS_IDLE] = 0;
639 652 cpudrvpm->lastquan_mstate[CMS_SYSTEM] = 0;
640 653 cpudrvpm->lastquan_mstate[CMS_USER] = 0;
641 654 cpudrvpm->lastquan_ticks = 0;
642 655 cpudrvpm->cur_spd = new_spd;
643 656 CPUDRV_RESET_GOVERNOR_THREAD(cpudrvpm);
644 657 mutex_exit(&cpudsp->lock);
645 658
646 659 return (DDI_SUCCESS);
647 660 }
648 661
649 662 /*
650 663 * Initialize power management data.
651 664 */
652 665 static int
653 666 cpudrv_init(cpudrv_devstate_t *cpudsp)
654 667 {
655 668 cpudrv_pm_t *cpupm = &(cpudsp->cpudrv_pm);
656 669 cpudrv_pm_spd_t *cur_spd;
657 670 cpudrv_pm_spd_t *prev_spd = NULL;
658 671 int *speeds;
659 672 uint_t nspeeds;
660 673 int idle_cnt_percent;
661 674 int user_cnt_percent;
662 675 int i;
663 676
664 677 CPUDRV_GET_SPEEDS(cpudsp, speeds, nspeeds);
665 678 if (nspeeds < 2) {
666 679 /* Need at least two speeds to power manage */
667 680 CPUDRV_FREE_SPEEDS(speeds, nspeeds);
668 681 return (DDI_FAILURE);
669 682 }
670 683 cpupm->num_spd = nspeeds;
671 684
672 685 /*
673 686 * Calculate the watermarks and other parameters based on the
674 687 * supplied speeds.
675 688 *
676 689 * One of the basic assumption is that for X amount of CPU work,
677 690 * if CPU is slowed down by a factor of N, the time it takes to
678 691 * do the same work will be N * X.
679 692 *
680 693 * The driver declares that a CPU is idle and ready for slowed down,
681 694 * if amount of idle thread is more than the current speed idle_hwm
682 695 * without dropping below idle_hwm a number of consecutive sampling
683 696 * intervals and number of running threads in user mode are below
684 697 * user_lwm. We want to set the current user_lwm such that if we
685 698 * just switched to the next slower speed with no change in real work
686 699 * load, the amount of user threads at the slower speed will be such
687 700 * that it falls below the slower speed's user_hwm. If we didn't do
688 701 * that then we will just come back to the higher speed as soon as we
689 702 * go down even with no change in work load.
690 703 * The user_hwm is a fixed precentage and not calculated dynamically.
691 704 *
692 705 * We bring the CPU up if idle thread at current speed is less than
693 706 * the current speed idle_lwm for a number of consecutive sampling
694 707 * intervals or user threads are above the user_hwm for the current
695 708 * speed.
696 709 */
697 710 for (i = 0; i < nspeeds; i++) {
698 711 cur_spd = kmem_zalloc(sizeof (cpudrv_pm_spd_t), KM_SLEEP);
699 712 cur_spd->speed = speeds[i];
700 713 if (i == 0) { /* normal speed */
701 714 cpupm->head_spd = cur_spd;
702 715 CPUDRV_TOPSPEED(cpupm) = cur_spd;
703 716 cur_spd->quant_cnt = CPUDRV_QUANT_CNT_NORMAL;
704 717 cur_spd->idle_hwm =
705 718 (cpudrv_idle_hwm * cur_spd->quant_cnt) / 100;
706 719 /* can't speed anymore */
707 720 cur_spd->idle_lwm = 0;
708 721 cur_spd->user_hwm = UINT_MAX;
709 722 } else {
710 723 cur_spd->quant_cnt = CPUDRV_QUANT_CNT_OTHR;
711 724 ASSERT(prev_spd != NULL);
712 725 prev_spd->down_spd = cur_spd;
713 726 cur_spd->up_spd = cpupm->head_spd;
714 727
715 728 /*
716 729 * Let's assume CPU is considered idle at full speed
717 730 * when it is spending I% of time in running the idle
718 731 * thread. At full speed, CPU will be busy (100 - I) %
719 732 * of times. This % of busyness increases by factor of
720 733 * N as CPU slows down. CPU that is idle I% of times
721 734 * in full speed, it is idle (100 - ((100 - I) * N)) %
722 735 * of times in N speed. The idle_lwm is a fixed
723 736 * percentage. A large value of N may result in
724 737 * idle_hwm to go below idle_lwm. We need to make sure
725 738 * that there is at least a buffer zone seperation
726 739 * between the idle_lwm and idle_hwm values.
727 740 */
728 741 idle_cnt_percent = CPUDRV_IDLE_CNT_PERCENT(
729 742 cpudrv_idle_hwm, speeds, i);
730 743 idle_cnt_percent = max(idle_cnt_percent,
731 744 (cpudrv_idle_lwm + cpudrv_idle_buf_zone));
732 745 cur_spd->idle_hwm =
733 746 (idle_cnt_percent * cur_spd->quant_cnt) / 100;
734 747 cur_spd->idle_lwm =
735 748 (cpudrv_idle_lwm * cur_spd->quant_cnt) / 100;
736 749
737 750 /*
738 751 * The lwm for user threads are determined such that
739 752 * if CPU slows down, the load of work in the
740 753 * new speed would still keep the CPU at or below the
741 754 * user_hwm in the new speed. This is to prevent
742 755 * the quick jump back up to higher speed.
743 756 */
744 757 cur_spd->user_hwm = (cpudrv_user_hwm *
745 758 cur_spd->quant_cnt) / 100;
746 759 user_cnt_percent = CPUDRV_USER_CNT_PERCENT(
747 760 cpudrv_user_hwm, speeds, i);
748 761 prev_spd->user_lwm =
749 762 (user_cnt_percent * prev_spd->quant_cnt) / 100;
750 763 }
751 764 prev_spd = cur_spd;
752 765 }
753 766 /* Slowest speed. Can't slow down anymore */
754 767 cur_spd->idle_hwm = UINT_MAX;
755 768 cur_spd->user_lwm = -1;
756 769 #ifdef DEBUG
757 770 DPRINTF(D_PM_INIT, ("cpudrv_init: instance %d: head_spd spd %d, "
758 771 "num_spd %d\n", ddi_get_instance(cpudsp->dip),
759 772 cpupm->head_spd->speed, cpupm->num_spd));
760 773 for (cur_spd = cpupm->head_spd; cur_spd; cur_spd = cur_spd->down_spd) {
761 774 DPRINTF(D_PM_INIT, ("cpudrv_init: instance %d: speed %d, "
762 775 "down_spd spd %d, idle_hwm %d, user_lwm %d, "
763 776 "up_spd spd %d, idle_lwm %d, user_hwm %d, "
764 777 "quant_cnt %d\n", ddi_get_instance(cpudsp->dip),
765 778 cur_spd->speed,
766 779 (cur_spd->down_spd ? cur_spd->down_spd->speed : 0),
767 780 cur_spd->idle_hwm, cur_spd->user_lwm,
768 781 (cur_spd->up_spd ? cur_spd->up_spd->speed : 0),
769 782 cur_spd->idle_lwm, cur_spd->user_hwm,
770 783 cur_spd->quant_cnt));
771 784 }
772 785 #endif /* DEBUG */
773 786 CPUDRV_FREE_SPEEDS(speeds, nspeeds);
774 787 return (DDI_SUCCESS);
775 788 }
776 789
777 790 /*
778 791 * Free CPU power management data.
779 792 */
780 793 static void
781 794 cpudrv_free(cpudrv_devstate_t *cpudsp)
782 795 {
783 796 cpudrv_pm_t *cpupm = &(cpudsp->cpudrv_pm);
784 797 cpudrv_pm_spd_t *cur_spd, *next_spd;
785 798
786 799 cur_spd = cpupm->head_spd;
787 800 while (cur_spd) {
788 801 next_spd = cur_spd->down_spd;
789 802 kmem_free(cur_spd, sizeof (cpudrv_pm_spd_t));
790 803 cur_spd = next_spd;
791 804 }
792 805 bzero(cpupm, sizeof (cpudrv_pm_t));
793 806 }
794 807
795 808 /*
796 809 * Create pm-components property.
797 810 */
798 811 static int
799 812 cpudrv_comp_create(cpudrv_devstate_t *cpudsp)
800 813 {
801 814 cpudrv_pm_t *cpupm = &(cpudsp->cpudrv_pm);
802 815 cpudrv_pm_spd_t *cur_spd;
803 816 char **pmc;
804 817 int size;
805 818 char name[] = "NAME=CPU Speed";
806 819 int i, j;
807 820 uint_t comp_spd;
808 821 int result = DDI_FAILURE;
809 822
810 823 pmc = kmem_zalloc((cpupm->num_spd + 1) * sizeof (char *), KM_SLEEP);
811 824 size = CPUDRV_COMP_SIZE();
812 825 if (cpupm->num_spd > CPUDRV_COMP_MAX_VAL) {
813 826 cmn_err(CE_WARN, "cpudrv_comp_create: instance %d: "
814 827 "number of speeds exceeded limits",
815 828 ddi_get_instance(cpudsp->dip));
816 829 kmem_free(pmc, (cpupm->num_spd + 1) * sizeof (char *));
817 830 return (result);
818 831 }
819 832
820 833 for (i = cpupm->num_spd, cur_spd = cpupm->head_spd; i > 0;
821 834 i--, cur_spd = cur_spd->down_spd) {
822 835 cur_spd->pm_level = i;
823 836 pmc[i] = kmem_zalloc((size * sizeof (char)), KM_SLEEP);
824 837 comp_spd = CPUDRV_COMP_SPEED(cpupm, cur_spd);
825 838 if (comp_spd > CPUDRV_COMP_MAX_VAL) {
826 839 cmn_err(CE_WARN, "cpudrv_comp_create: "
827 840 "instance %d: speed exceeded limits",
828 841 ddi_get_instance(cpudsp->dip));
829 842 for (j = cpupm->num_spd; j >= i; j--) {
830 843 kmem_free(pmc[j], size * sizeof (char));
831 844 }
832 845 kmem_free(pmc, (cpupm->num_spd + 1) *
833 846 sizeof (char *));
834 847 return (result);
835 848 }
836 849 CPUDRV_COMP_SPRINT(pmc[i], cpupm, cur_spd, comp_spd)
837 850 DPRINTF(D_PM_COMP_CREATE, ("cpudrv_comp_create: "
838 851 "instance %d: pm-components power level %d string '%s'\n",
839 852 ddi_get_instance(cpudsp->dip), i, pmc[i]));
840 853 }
841 854 pmc[0] = kmem_zalloc(sizeof (name), KM_SLEEP);
842 855 (void) strcat(pmc[0], name);
843 856 DPRINTF(D_PM_COMP_CREATE, ("cpudrv_comp_create: instance %d: "
844 857 "pm-components component name '%s'\n",
845 858 ddi_get_instance(cpudsp->dip), pmc[0]));
846 859
847 860 if (ddi_prop_update_string_array(DDI_DEV_T_NONE, cpudsp->dip,
848 861 "pm-components", pmc, cpupm->num_spd + 1) == DDI_PROP_SUCCESS) {
849 862 result = DDI_SUCCESS;
850 863 } else {
851 864 cmn_err(CE_WARN, "cpudrv_comp_create: instance %d: "
852 865 "can't create pm-components property",
853 866 ddi_get_instance(cpudsp->dip));
854 867 }
855 868
856 869 for (i = cpupm->num_spd; i > 0; i--) {
857 870 kmem_free(pmc[i], size * sizeof (char));
858 871 }
859 872 kmem_free(pmc[0], sizeof (name));
860 873 kmem_free(pmc, (cpupm->num_spd + 1) * sizeof (char *));
861 874 return (result);
862 875 }
863 876
864 877 /*
865 878 * Mark a component idle.
866 879 */
867 880 #define CPUDRV_MONITOR_PM_IDLE_COMP(dip, cpupm) { \
868 881 if ((cpupm)->pm_busycnt >= 1) { \
869 882 if (pm_idle_component((dip), CPUDRV_COMP_NUM) == \
870 883 DDI_SUCCESS) { \
871 884 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: " \
872 885 "instance %d: pm_idle_component called\n", \
873 886 ddi_get_instance((dip)))); \
874 887 (cpupm)->pm_busycnt--; \
875 888 } else { \
876 889 cmn_err(CE_WARN, "cpudrv_monitor: instance %d: " \
877 890 "can't idle CPU component", \
878 891 ddi_get_instance((dip))); \
879 892 } \
880 893 } \
881 894 }
882 895
883 896 /*
884 897 * Marks a component busy in both PM framework and driver state structure.
885 898 */
886 899 #define CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm) { \
887 900 if ((cpupm)->pm_busycnt < 1) { \
888 901 if (pm_busy_component((dip), CPUDRV_COMP_NUM) == \
889 902 DDI_SUCCESS) { \
890 903 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: " \
891 904 "instance %d: pm_busy_component called\n", \
892 905 ddi_get_instance((dip)))); \
893 906 (cpupm)->pm_busycnt++; \
894 907 } else { \
895 908 cmn_err(CE_WARN, "cpudrv_monitor: instance %d: " \
896 909 "can't busy CPU component", \
897 910 ddi_get_instance((dip))); \
898 911 } \
899 912 } \
900 913 }
901 914
902 915 /*
903 916 * Marks a component busy and calls pm_raise_power().
904 917 */
905 918 #define CPUDRV_MONITOR_PM_BUSY_AND_RAISE(dip, cpudsp, cpupm, new_spd) { \
906 919 int ret; \
907 920 /* \
908 921 * Mark driver and PM framework busy first so framework doesn't try \
909 922 * to bring CPU to lower speed when we need to be at higher speed. \
910 923 */ \
911 924 CPUDRV_MONITOR_PM_BUSY_COMP((dip), (cpupm)); \
912 925 mutex_exit(&(cpudsp)->lock); \
913 926 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: " \
914 927 "pm_raise_power called to %d\n", ddi_get_instance((dip)), \
915 928 (new_spd->pm_level))); \
916 929 ret = pm_raise_power((dip), CPUDRV_COMP_NUM, (new_spd->pm_level)); \
917 930 if (ret != DDI_SUCCESS) { \
918 931 cmn_err(CE_WARN, "cpudrv_monitor: instance %d: can't " \
919 932 "raise CPU power level", ddi_get_instance((dip))); \
920 933 } \
921 934 mutex_enter(&(cpudsp)->lock); \
922 935 if (ret == DDI_SUCCESS && cpudsp->cpudrv_pm.cur_spd == NULL) { \
923 936 cpudsp->cpudrv_pm.cur_spd = new_spd; \
924 937 } \
925 938 }
926 939
927 940 /*
928 941 * In order to monitor a CPU, we need to hold cpu_lock to access CPU
929 942 * statistics. Holding cpu_lock is not allowed from a callout routine.
930 943 * We dispatch a taskq to do that job.
931 944 */
932 945 static void
933 946 cpudrv_monitor_disp(void *arg)
934 947 {
935 948 cpudrv_devstate_t *cpudsp = (cpudrv_devstate_t *)arg;
936 949
937 950 /*
938 951 * We are here because the last task has scheduled a timeout.
939 952 * The queue should be empty at this time.
940 953 */
941 954 mutex_enter(&cpudsp->cpudrv_pm.timeout_lock);
942 955 if ((ddi_taskq_dispatch(cpudsp->cpudrv_pm.tq, cpudrv_monitor, arg,
943 956 DDI_NOSLEEP)) != DDI_SUCCESS) {
944 957 mutex_exit(&cpudsp->cpudrv_pm.timeout_lock);
945 958 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor_disp: failed to "
946 959 "dispatch the cpudrv_monitor taskq\n"));
947 960 mutex_enter(&cpudsp->lock);
948 961 CPUDRV_MONITOR_INIT(cpudsp);
949 962 mutex_exit(&cpudsp->lock);
950 963 return;
951 964 }
952 965 cpudsp->cpudrv_pm.timeout_count++;
953 966 mutex_exit(&cpudsp->cpudrv_pm.timeout_lock);
954 967 }
955 968
956 969 /*
957 970 * Monitors each CPU for the amount of time idle thread was running in the
958 971 * last quantum and arranges for the CPU to go to the lower or higher speed.
959 972 * Called at the time interval appropriate for the current speed. The
960 973 * time interval for normal speed is CPUDRV_QUANT_CNT_NORMAL. The time
961 974 * interval for other speeds (including unknown speed) is
962 975 * CPUDRV_QUANT_CNT_OTHR.
963 976 */
964 977 static void
965 978 cpudrv_monitor(void *arg)
966 979 {
967 980 cpudrv_devstate_t *cpudsp = (cpudrv_devstate_t *)arg;
968 981 cpudrv_pm_t *cpupm;
969 982 cpudrv_pm_spd_t *cur_spd, *new_spd;
970 983 dev_info_t *dip;
971 984 uint_t idle_cnt, user_cnt, system_cnt;
972 985 clock_t ticks;
973 986 uint_t tick_cnt;
974 987 hrtime_t msnsecs[NCMSTATES];
975 988 boolean_t is_ready;
976 989
977 990 #define GET_CPU_MSTATE_CNT(state, cnt) \
978 991 msnsecs[state] = NSEC_TO_TICK(msnsecs[state]); \
979 992 if (cpupm->lastquan_mstate[state] > msnsecs[state]) \
980 993 msnsecs[state] = cpupm->lastquan_mstate[state]; \
981 994 cnt = msnsecs[state] - cpupm->lastquan_mstate[state]; \
982 995 cpupm->lastquan_mstate[state] = msnsecs[state]
983 996
984 997 /*
985 998 * We're not ready until we can get a cpu_t
986 999 */
987 1000 is_ready = (cpudrv_get_cpu(cpudsp) == DDI_SUCCESS);
988 1001
989 1002 mutex_enter(&cpudsp->lock);
990 1003 cpupm = &(cpudsp->cpudrv_pm);
991 1004 if (cpupm->timeout_id == 0) {
992 1005 mutex_exit(&cpudsp->lock);
993 1006 goto do_return;
994 1007 }
995 1008 cur_spd = cpupm->cur_spd;
996 1009 dip = cpudsp->dip;
997 1010
998 1011 /*
999 1012 * We assume that a CPU is initialized and has a valid cpu_t
1000 1013 * structure, if it is ready for cross calls. If this changes,
1001 1014 * additional checks might be needed.
1002 1015 *
1003 1016 * Additionally, for x86 platforms we cannot power manage an
1004 1017 * instance, until it has been initialized.
1005 1018 */
1006 1019 if (is_ready) {
1007 1020 is_ready = CPUDRV_XCALL_IS_READY(cpudsp->cpu_id);
1008 1021 if (!is_ready) {
1009 1022 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: "
1010 1023 "CPU not ready for x-calls\n",
1011 1024 ddi_get_instance(dip)));
1012 1025 } else if (!(is_ready = cpudrv_power_ready(cpudsp->cp))) {
1013 1026 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: "
1014 1027 "waiting for all CPUs to be power manageable\n",
1015 1028 ddi_get_instance(dip)));
1016 1029 }
1017 1030 }
1018 1031 if (!is_ready) {
1019 1032 /*
1020 1033 * Make sure that we are busy so that framework doesn't
1021 1034 * try to bring us down in this situation.
1022 1035 */
1023 1036 CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm);
1024 1037 CPUDRV_MONITOR_INIT(cpudsp);
1025 1038 mutex_exit(&cpudsp->lock);
1026 1039 goto do_return;
1027 1040 }
1028 1041
1029 1042 /*
1030 1043 * Make sure that we are still not at unknown power level.
1031 1044 */
1032 1045 if (cur_spd == NULL) {
1033 1046 DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: "
1034 1047 "cur_spd is unknown\n", ddi_get_instance(dip)));
1035 1048 CPUDRV_MONITOR_PM_BUSY_AND_RAISE(dip, cpudsp, cpupm,
1036 1049 CPUDRV_TOPSPEED(cpupm));
1037 1050 /*
1038 1051 * We just changed the speed. Wait till at least next
1039 1052 * call to this routine before proceeding ahead.
1040 1053 */
1041 1054 CPUDRV_MONITOR_INIT(cpudsp);
1042 1055 mutex_exit(&cpudsp->lock);
1043 1056 goto do_return;
1044 1057 }
1045 1058
1046 1059 if (!cpupm->pm_started) {
1047 1060 cpupm->pm_started = B_TRUE;
1048 1061 cpudrv_set_supp_freqs(cpudsp);
1049 1062 }
1050 1063
1051 1064 get_cpu_mstate(cpudsp->cp, msnsecs);
1052 1065 GET_CPU_MSTATE_CNT(CMS_IDLE, idle_cnt);
1053 1066 GET_CPU_MSTATE_CNT(CMS_USER, user_cnt);
1054 1067 GET_CPU_MSTATE_CNT(CMS_SYSTEM, system_cnt);
1055 1068
1056 1069 /*
1057 1070 * We can't do anything when we have just switched to a state
1058 1071 * because there is no valid timestamp.
1059 1072 */
1060 1073 if (cpupm->lastquan_ticks == 0) {
1061 1074 cpupm->lastquan_ticks = NSEC_TO_TICK(gethrtime());
1062 1075 CPUDRV_MONITOR_INIT(cpudsp);
1063 1076 mutex_exit(&cpudsp->lock);
1064 1077 goto do_return;
1065 1078 }
1066 1079
1067 1080 /*
1068 1081 * Various watermarks are based on this routine being called back
1069 1082 * exactly at the requested period. This is not guaranteed
1070 1083 * because this routine is called from a taskq that is dispatched
1071 1084 * from a timeout routine. Handle this by finding out how many
1072 1085 * ticks have elapsed since the last call and adjusting
1073 1086 * the idle_cnt based on the delay added to the requested period
1074 1087 * by timeout and taskq.
1075 1088 */
1076 1089 ticks = NSEC_TO_TICK(gethrtime());
1077 1090 tick_cnt = ticks - cpupm->lastquan_ticks;
1078 1091 ASSERT(tick_cnt != 0);
1079 1092 cpupm->lastquan_ticks = ticks;
1080 1093
1081 1094 /*
1082 1095 * Time taken between recording the current counts and
1083 1096 * arranging the next call of this routine is an error in our
1084 1097 * calculation. We minimize the error by calling
1085 1098 * CPUDRV_MONITOR_INIT() here instead of end of this routine.
1086 1099 */
1087 1100 CPUDRV_MONITOR_INIT(cpudsp);
1088 1101 DPRINTF(D_PM_MONITOR_VERBOSE, ("cpudrv_monitor: instance %d: "
1089 1102 "idle count %d, user count %d, system count %d, pm_level %d, "
1090 1103 "pm_busycnt %d\n", ddi_get_instance(dip), idle_cnt, user_cnt,
1091 1104 system_cnt, cur_spd->pm_level, cpupm->pm_busycnt));
1092 1105
1093 1106 #ifdef DEBUG
1094 1107 /*
1095 1108 * Notify that timeout and taskq has caused delays and we need to
1096 1109 * scale our parameters accordingly.
1097 1110 *
1098 1111 * To get accurate result, don't turn on other DPRINTFs with
1099 1112 * the following DPRINTF. PROM calls generated by other
1100 1113 * DPRINTFs changes the timing.
1101 1114 */
1102 1115 if (tick_cnt > cur_spd->quant_cnt) {
1103 1116 DPRINTF(D_PM_MONITOR_DELAY, ("cpudrv_monitor: instance %d: "
1104 1117 "tick count %d > quantum_count %u\n",
1105 1118 ddi_get_instance(dip), tick_cnt, cur_spd->quant_cnt));
1106 1119 }
1107 1120 #endif /* DEBUG */
1108 1121
1109 1122 /*
1110 1123 * Adjust counts based on the delay added by timeout and taskq.
1111 1124 */
1112 1125 idle_cnt = (idle_cnt * cur_spd->quant_cnt) / tick_cnt;
1113 1126 user_cnt = (user_cnt * cur_spd->quant_cnt) / tick_cnt;
1114 1127
1115 1128 if ((user_cnt > cur_spd->user_hwm) || (idle_cnt < cur_spd->idle_lwm &&
1116 1129 cur_spd->idle_blwm_cnt >= cpudrv_idle_blwm_cnt_max)) {
1117 1130 cur_spd->idle_blwm_cnt = 0;
1118 1131 cur_spd->idle_bhwm_cnt = 0;
1119 1132 /*
1120 1133 * In normal situation, arrange to go to next higher speed.
1121 1134 * If we are running in special direct pm mode, we just stay
1122 1135 * at the current speed.
1123 1136 */
1124 1137 if (cur_spd == cur_spd->up_spd || cpudrv_direct_pm) {
1125 1138 CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm);
1126 1139 } else {
1127 1140 new_spd = cur_spd->up_spd;
1128 1141 CPUDRV_MONITOR_PM_BUSY_AND_RAISE(dip, cpudsp, cpupm,
1129 1142 new_spd);
1130 1143 }
1131 1144 } else if ((user_cnt <= cur_spd->user_lwm) &&
1132 1145 (idle_cnt >= cur_spd->idle_hwm) || !CPU_ACTIVE(cpudsp->cp)) {
1133 1146 cur_spd->idle_blwm_cnt = 0;
1134 1147 cur_spd->idle_bhwm_cnt = 0;
1135 1148 /*
1136 1149 * Arrange to go to next lower speed by informing our idle
1137 1150 * status to the power management framework.
1138 1151 */
1139 1152 CPUDRV_MONITOR_PM_IDLE_COMP(dip, cpupm);
1140 1153 } else {
1141 1154 /*
1142 1155 * If we are between the idle water marks and have not
1143 1156 * been here enough consecutive times to be considered
1144 1157 * busy, just increment the count and return.
1145 1158 */
1146 1159 if ((idle_cnt < cur_spd->idle_hwm) &&
1147 1160 (idle_cnt >= cur_spd->idle_lwm) &&
1148 1161 (cur_spd->idle_bhwm_cnt < cpudrv_idle_bhwm_cnt_max)) {
1149 1162 cur_spd->idle_blwm_cnt = 0;
1150 1163 cur_spd->idle_bhwm_cnt++;
1151 1164 mutex_exit(&cpudsp->lock);
1152 1165 goto do_return;
1153 1166 }
1154 1167 if (idle_cnt < cur_spd->idle_lwm) {
1155 1168 cur_spd->idle_blwm_cnt++;
1156 1169 cur_spd->idle_bhwm_cnt = 0;
1157 1170 }
1158 1171 /*
1159 1172 * Arranges to stay at the current speed.
1160 1173 */
1161 1174 CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm);
1162 1175 }
1163 1176 mutex_exit(&cpudsp->lock);
1164 1177 do_return:
1165 1178 mutex_enter(&cpupm->timeout_lock);
1166 1179 ASSERT(cpupm->timeout_count > 0);
1167 1180 cpupm->timeout_count--;
1168 1181 cv_signal(&cpupm->timeout_cv);
1169 1182 mutex_exit(&cpupm->timeout_lock);
1170 1183 }
1171 1184
1172 1185 /*
1173 1186 * get cpu_t structure for cpudrv_devstate_t
1174 1187 */
1175 1188 int
1176 1189 cpudrv_get_cpu(cpudrv_devstate_t *cpudsp)
1177 1190 {
1178 1191 ASSERT(cpudsp != NULL);
1179 1192
1180 1193 /*
1181 1194 * return DDI_SUCCESS if cpudrv_devstate_t
1182 1195 * already contains cpu_t structure
1183 1196 */
1184 1197 if (cpudsp->cp != NULL)
1185 1198 return (DDI_SUCCESS);
1186 1199
1187 1200 if (MUTEX_HELD(&cpu_lock)) {
1188 1201 cpudsp->cp = cpu_get(cpudsp->cpu_id);
1189 1202 } else {
1190 1203 mutex_enter(&cpu_lock);
1191 1204 cpudsp->cp = cpu_get(cpudsp->cpu_id);
1192 1205 mutex_exit(&cpu_lock);
1193 1206 }
1194 1207
1195 1208 if (cpudsp->cp == NULL)
1196 1209 return (DDI_FAILURE);
1197 1210
1198 1211 return (DDI_SUCCESS);
1199 1212 }
↓ open down ↓ |
845 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX