Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux...
[linux-2.6] / drivers / idle / i7300_idle.c
1 /*
2  * (C) Copyright 2008 Intel Corporation
3  * Authors:
4  * Andy Henroid <andrew.d.henroid@intel.com>
5  * Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
6  */
7
8 /*
9  * Save DIMM power on Intel 7300-based platforms when all CPUs/cores
10  * are idle, using the DIMM thermal throttling capability.
11  *
12  * This driver depends on the Intel integrated DMA controller (I/O AT).
13  * If the driver for I/O AT (drivers/dma/ioatdma*) is also enabled,
14  * this driver should work cooperatively.
15  */
16
17 /* #define DEBUG */
18
19 #include <linux/module.h>
20 #include <linux/pci.h>
21 #include <linux/sched.h>
22 #include <linux/notifier.h>
23 #include <linux/cpumask.h>
24 #include <linux/ktime.h>
25 #include <linux/delay.h>
26 #include <linux/debugfs.h>
27 #include <linux/stop_machine.h>
28 #include <linux/i7300_idle.h>
29
30 #include <asm/idle.h>
31
32 #include "../dma/ioatdma_hw.h"
33 #include "../dma/ioatdma_registers.h"
34
35 #define I7300_IDLE_DRIVER_VERSION       "1.55"
36 #define I7300_PRINT                     "i7300_idle:"
37
38 #define MAX_STOP_RETRIES        10
39
40 static int debug;
41 module_param_named(debug, debug, uint, 0644);
42 MODULE_PARM_DESC(debug, "Enable debug printks in this driver");
43
44 #define dprintk(fmt, arg...) \
45         do { if (debug) printk(KERN_INFO I7300_PRINT fmt, ##arg); } while (0)
46
47 /*
48  * Value to set THRTLOW to when initiating throttling
49  *  0 = No throttling
50  *  1 = Throttle when > 4 activations per eval window (Maximum throttling)
51  *  2 = Throttle when > 8 activations
52  *  168 = Throttle when > 672 activations (Minimum throttling)
53  */
54 #define MAX_THROTTLE_LOW_LIMIT          168
55 static uint throttle_low_limit = 1;
56 module_param_named(throttle_low_limit, throttle_low_limit, uint, 0644);
57 MODULE_PARM_DESC(throttle_low_limit,
58                 "Value for THRTLOWLM activation field "
59                 "(0 = disable throttle, 1 = Max throttle, 168 = Min throttle)");
60
61 /*
62  * simple invocation and duration statistics
63  */
64 static unsigned long total_starts;
65 static unsigned long total_us;
66
67 #ifdef DEBUG
68 static unsigned long past_skip;
69 #endif
70
71 static struct pci_dev *fbd_dev;
72
73 static spinlock_t i7300_idle_lock;
74 static int i7300_idle_active;
75
76 static u8 i7300_idle_thrtctl_saved;
77 static u8 i7300_idle_thrtlow_saved;
78 static u32 i7300_idle_mc_saved;
79
80 static cpumask_t idle_cpumask;
81 static ktime_t start_ktime;
82 static unsigned long avg_idle_us;
83
84 static struct dentry *debugfs_dir;
85
86 /* Begin: I/O AT Helper routines */
87
88 #define IOAT_CHANBASE(ioat_ctl, chan) (ioat_ctl + 0x80 + 0x80 * chan)
89 /* Snoop control (disable snoops when coherency is not important) */
90 #define IOAT_DESC_SADDR_SNP_CTL (1UL << 1)
91 #define IOAT_DESC_DADDR_SNP_CTL (1UL << 2)
92
93 static struct pci_dev *ioat_dev;
94 static struct ioat_dma_descriptor *ioat_desc; /* I/O AT desc & data (1 page) */
95 static unsigned long ioat_desc_phys;
96 static u8 *ioat_iomap; /* I/O AT memory-mapped control regs (aka CB_BAR) */
97 static u8 *ioat_chanbase;
98
99 /* Start I/O AT memory copy */
100 static int i7300_idle_ioat_start(void)
101 {
102         u32 err;
103         /* Clear error (due to circular descriptor pointer) */
104         err = readl(ioat_chanbase + IOAT_CHANERR_OFFSET);
105         if (err)
106                 writel(err, ioat_chanbase + IOAT_CHANERR_OFFSET);
107
108         writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
109         return 0;
110 }
111
112 /* Stop I/O AT memory copy */
113 static void i7300_idle_ioat_stop(void)
114 {
115         int i;
116         u64 sts;
117
118         for (i = 0; i < MAX_STOP_RETRIES; i++) {
119                 writeb(IOAT_CHANCMD_RESET,
120                         ioat_chanbase + IOAT1_CHANCMD_OFFSET);
121
122                 udelay(10);
123
124                 sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
125                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
126
127                 if (sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE)
128                         break;
129
130         }
131
132         if (i == MAX_STOP_RETRIES) {
133                 dprintk("failed to stop I/O AT after %d retries\n",
134                         MAX_STOP_RETRIES);
135         }
136 }
137
138 /* Test I/O AT by copying 1024 byte from 2k to 1k */
139 static int __init i7300_idle_ioat_selftest(u8 *ctl,
140                 struct ioat_dma_descriptor *desc, unsigned long desc_phys)
141 {
142         u64 chan_sts;
143
144         memset(desc, 0, 2048);
145         memset((u8 *) desc + 2048, 0xab, 1024);
146
147         desc[0].size = 1024;
148         desc[0].ctl = 0;
149         desc[0].src_addr = desc_phys + 2048;
150         desc[0].dst_addr = desc_phys + 1024;
151         desc[0].next = 0;
152
153         writeb(IOAT_CHANCMD_RESET, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
154         writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
155
156         udelay(1000);
157
158         chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
159                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
160
161         if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_DONE) {
162                 /* Not complete, reset the channel */
163                 writeb(IOAT_CHANCMD_RESET,
164                        ioat_chanbase + IOAT1_CHANCMD_OFFSET);
165                 return -1;
166         }
167
168         if (*(u32 *) ((u8 *) desc + 3068) != 0xabababab ||
169             *(u32 *) ((u8 *) desc + 2044) != 0xabababab) {
170                 dprintk("Data values src 0x%x, dest 0x%x, memset 0x%x\n",
171                         *(u32 *) ((u8 *) desc + 2048),
172                         *(u32 *) ((u8 *) desc + 1024),
173                         *(u32 *) ((u8 *) desc + 3072));
174                 return -1;
175         }
176         return 0;
177 }
178
179 static struct device dummy_dma_dev = {
180         .init_name = "fallback device",
181         .coherent_dma_mask = DMA_64BIT_MASK,
182         .dma_mask = &dummy_dma_dev.coherent_dma_mask,
183 };
184
185 /* Setup and initialize I/O AT */
186 /* This driver needs I/O AT as the throttling takes effect only when there is
187  * some memory activity. We use I/O AT to set up a dummy copy, while all CPUs
188  * go idle and memory is throttled.
189  */
190 static int __init i7300_idle_ioat_init(void)
191 {
192         u8 ver, chan_count, ioat_chan;
193         u16 chan_ctl;
194
195         ioat_iomap = (u8 *) ioremap_nocache(pci_resource_start(ioat_dev, 0),
196                                             pci_resource_len(ioat_dev, 0));
197
198         if (!ioat_iomap) {
199                 printk(KERN_ERR I7300_PRINT "failed to map I/O AT registers\n");
200                 goto err_ret;
201         }
202
203         ver = readb(ioat_iomap + IOAT_VER_OFFSET);
204         if (ver != IOAT_VER_1_2) {
205                 printk(KERN_ERR I7300_PRINT "unknown I/O AT version (%u.%u)\n",
206                         ver >> 4, ver & 0xf);
207                 goto err_unmap;
208         }
209
210         chan_count = readb(ioat_iomap + IOAT_CHANCNT_OFFSET);
211         if (!chan_count) {
212                 printk(KERN_ERR I7300_PRINT "unexpected # of I/O AT channels "
213                         "(%u)\n",
214                         chan_count);
215                 goto err_unmap;
216         }
217
218         ioat_chan = chan_count - 1;
219         ioat_chanbase = IOAT_CHANBASE(ioat_iomap, ioat_chan);
220
221         chan_ctl = readw(ioat_chanbase + IOAT_CHANCTRL_OFFSET);
222         if (chan_ctl & IOAT_CHANCTRL_CHANNEL_IN_USE) {
223                 printk(KERN_ERR I7300_PRINT "channel %d in use\n", ioat_chan);
224                 goto err_unmap;
225         }
226
227         writew(IOAT_CHANCTRL_CHANNEL_IN_USE,
228                 ioat_chanbase + IOAT_CHANCTRL_OFFSET);
229
230         ioat_desc = (struct ioat_dma_descriptor *)dma_alloc_coherent(
231                         &dummy_dma_dev, 4096,
232                         (dma_addr_t *)&ioat_desc_phys, GFP_KERNEL);
233         if (!ioat_desc) {
234                 printk(KERN_ERR I7300_PRINT "failed to allocate I/O AT desc\n");
235                 goto err_mark_unused;
236         }
237
238         writel(ioat_desc_phys & 0xffffffffUL,
239                ioat_chanbase + IOAT1_CHAINADDR_OFFSET_LOW);
240         writel(ioat_desc_phys >> 32,
241                ioat_chanbase + IOAT1_CHAINADDR_OFFSET_HIGH);
242
243         if (i7300_idle_ioat_selftest(ioat_iomap, ioat_desc, ioat_desc_phys)) {
244                 printk(KERN_ERR I7300_PRINT "I/O AT self-test failed\n");
245                 goto err_free;
246         }
247
248         /* Setup circular I/O AT descriptor chain */
249         ioat_desc[0].ctl = IOAT_DESC_SADDR_SNP_CTL | IOAT_DESC_DADDR_SNP_CTL;
250         ioat_desc[0].src_addr = ioat_desc_phys + 2048;
251         ioat_desc[0].dst_addr = ioat_desc_phys + 3072;
252         ioat_desc[0].size = 128;
253         ioat_desc[0].next = ioat_desc_phys + sizeof(struct ioat_dma_descriptor);
254
255         ioat_desc[1].ctl = ioat_desc[0].ctl;
256         ioat_desc[1].src_addr = ioat_desc[0].src_addr;
257         ioat_desc[1].dst_addr = ioat_desc[0].dst_addr;
258         ioat_desc[1].size = ioat_desc[0].size;
259         ioat_desc[1].next = ioat_desc_phys;
260
261         return 0;
262
263 err_free:
264         dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
265 err_mark_unused:
266         writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
267 err_unmap:
268         iounmap(ioat_iomap);
269 err_ret:
270         return -ENODEV;
271 }
272
273 /* Cleanup I/O AT */
274 static void __exit i7300_idle_ioat_exit(void)
275 {
276         int i;
277         u64 chan_sts;
278
279         i7300_idle_ioat_stop();
280
281         /* Wait for a while for the channel to halt before releasing */
282         for (i = 0; i < MAX_STOP_RETRIES; i++) {
283                 writeb(IOAT_CHANCMD_RESET,
284                        ioat_chanbase + IOAT1_CHANCMD_OFFSET);
285
286                 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
287                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
288
289                 if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
290                         writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
291                         break;
292                 }
293                 udelay(1000);
294         }
295
296         chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
297                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
298
299         /*
300          * We tried to reset multiple times. If IO A/T channel is still active
301          * flag an error and return without cleanup. Memory leak is better
302          * than random corruption in that extreme error situation.
303          */
304         if (chan_sts == IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
305                 printk(KERN_ERR I7300_PRINT "Unable to stop IO A/T channels."
306                         " Not freeing resources\n");
307                 return;
308         }
309
310         dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
311         iounmap(ioat_iomap);
312 }
313
314 /* End: I/O AT Helper routines */
315
316 #define DIMM_THRTLOW 0x64
317 #define DIMM_THRTCTL 0x67
318 #define DIMM_THRTCTL_THRMHUNT (1UL << 0)
319 #define DIMM_MC 0x40
320 #define DIMM_GTW_MODE (1UL << 17)
321 #define DIMM_GBLACT 0x60
322
323 /*
324  * Keep track of an exponential-decaying average of recent idle durations.
325  * The latest duration gets DURATION_WEIGHT_PCT percentage weight
326  * in this average, with the old average getting the remaining weight.
327  *
328  * High weights emphasize recent history, low weights include long history.
329  */
330 #define DURATION_WEIGHT_PCT 55
331
332 /*
333  * When the decaying average of recent durations or the predicted duration
334  * of the next timer interrupt is shorter than duration_threshold, the
335  * driver will decline to throttle.
336  */
337 #define DURATION_THRESHOLD_US 100
338
339
340 /* Store DIMM thermal throttle configuration */
341 static int i7300_idle_thrt_save(void)
342 {
343         u32 new_mc_val;
344         u8 gblactlm;
345
346         pci_read_config_byte(fbd_dev, DIMM_THRTCTL, &i7300_idle_thrtctl_saved);
347         pci_read_config_byte(fbd_dev, DIMM_THRTLOW, &i7300_idle_thrtlow_saved);
348         pci_read_config_dword(fbd_dev, DIMM_MC, &i7300_idle_mc_saved);
349         /*
350          * Make sure we have Global Throttling Window Mode set to have a
351          * "short" window. This (mostly) works around an issue where
352          * throttling persists until the end of the global throttling window
353          * size. On the tested system, this was resulting in a maximum of
354          * 64 ms to exit throttling (average 32 ms). The actual numbers
355          * depends on system frequencies. Setting the short window reduces
356          * this by a factor of 4096.
357          *
358          * We will only do this only if the system is set for
359          * unlimited-activations while in open-loop throttling (i.e., when
360          * Global Activation Throttle Limit is zero).
361          */
362         pci_read_config_byte(fbd_dev, DIMM_GBLACT, &gblactlm);
363         dprintk("thrtctl_saved = 0x%02x, thrtlow_saved = 0x%02x\n",
364                 i7300_idle_thrtctl_saved,
365                 i7300_idle_thrtlow_saved);
366         dprintk("mc_saved = 0x%08x, gblactlm = 0x%02x\n",
367                 i7300_idle_mc_saved,
368                 gblactlm);
369         if (gblactlm == 0) {
370                 new_mc_val = i7300_idle_mc_saved | DIMM_GTW_MODE;
371                 pci_write_config_dword(fbd_dev, DIMM_MC, new_mc_val);
372                 return 0;
373         } else {
374                 dprintk("could not set GTW_MODE = 1 (OLTT enabled)\n");
375                 return -ENODEV;
376         }
377 }
378
379 /* Restore DIMM thermal throttle configuration */
380 static void i7300_idle_thrt_restore(void)
381 {
382         pci_write_config_dword(fbd_dev, DIMM_MC, i7300_idle_mc_saved);
383         pci_write_config_byte(fbd_dev, DIMM_THRTLOW, i7300_idle_thrtlow_saved);
384         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, i7300_idle_thrtctl_saved);
385 }
386
387 /* Enable DIMM thermal throttling */
388 static void i7300_idle_start(void)
389 {
390         u8 new_ctl;
391         u8 limit;
392
393         new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
394         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
395
396         limit = throttle_low_limit;
397         if (unlikely(limit > MAX_THROTTLE_LOW_LIMIT))
398                 limit = MAX_THROTTLE_LOW_LIMIT;
399
400         pci_write_config_byte(fbd_dev, DIMM_THRTLOW, limit);
401
402         new_ctl = i7300_idle_thrtctl_saved | DIMM_THRTCTL_THRMHUNT;
403         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
404 }
405
406 /* Disable DIMM thermal throttling */
407 static void i7300_idle_stop(void)
408 {
409         u8 new_ctl;
410         u8 got_ctl;
411
412         new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
413         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
414
415         pci_write_config_byte(fbd_dev, DIMM_THRTLOW, i7300_idle_thrtlow_saved);
416         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, i7300_idle_thrtctl_saved);
417         pci_read_config_byte(fbd_dev, DIMM_THRTCTL, &got_ctl);
418         WARN_ON_ONCE(got_ctl != i7300_idle_thrtctl_saved);
419 }
420
421
422 /*
423  * i7300_avg_duration_check()
424  * return 0 if the decaying average of recent idle durations is
425  * more than DURATION_THRESHOLD_US
426  */
427 static int i7300_avg_duration_check(void)
428 {
429         if (avg_idle_us >= DURATION_THRESHOLD_US)
430                 return 0;
431
432 #ifdef DEBUG
433         past_skip++;
434 #endif
435         return 1;
436 }
437
438 /* Idle notifier to look at idle CPUs */
439 static int i7300_idle_notifier(struct notifier_block *nb, unsigned long val,
440                                 void *data)
441 {
442         unsigned long flags;
443         ktime_t now_ktime;
444         static ktime_t idle_begin_time;
445         static int time_init = 1;
446
447         if (!throttle_low_limit)
448                 return 0;
449
450         if (unlikely(time_init)) {
451                 time_init = 0;
452                 idle_begin_time = ktime_get();
453         }
454
455         spin_lock_irqsave(&i7300_idle_lock, flags);
456         if (val == IDLE_START) {
457
458                 cpu_set(smp_processor_id(), idle_cpumask);
459
460                 if (cpus_weight(idle_cpumask) != num_online_cpus())
461                         goto end;
462
463                 now_ktime = ktime_get();
464                 idle_begin_time = now_ktime;
465
466                 if (i7300_avg_duration_check())
467                         goto end;
468
469                 i7300_idle_active = 1;
470                 total_starts++;
471                 start_ktime = now_ktime;
472
473                 i7300_idle_start();
474                 i7300_idle_ioat_start();
475
476         } else if (val == IDLE_END) {
477                 cpu_clear(smp_processor_id(), idle_cpumask);
478                 if (cpus_weight(idle_cpumask) == (num_online_cpus() - 1)) {
479                         /* First CPU coming out of idle */
480                         u64 idle_duration_us;
481
482                         now_ktime = ktime_get();
483
484                         idle_duration_us = ktime_to_us(ktime_sub
485                                                 (now_ktime, idle_begin_time));
486
487                         avg_idle_us =
488                                 ((100 - DURATION_WEIGHT_PCT) * avg_idle_us +
489                                  DURATION_WEIGHT_PCT * idle_duration_us) / 100;
490
491                         if (i7300_idle_active) {
492                                 ktime_t idle_ktime;
493
494                                 idle_ktime = ktime_sub(now_ktime, start_ktime);
495                                 total_us += ktime_to_us(idle_ktime);
496
497                                 i7300_idle_ioat_stop();
498                                 i7300_idle_stop();
499                                 i7300_idle_active = 0;
500                         }
501                 }
502         }
503 end:
504         spin_unlock_irqrestore(&i7300_idle_lock, flags);
505         return 0;
506 }
507
508 static struct notifier_block i7300_idle_nb = {
509         .notifier_call = i7300_idle_notifier,
510 };
511
512 MODULE_DEVICE_TABLE(pci, pci_tbl);
513
514 int stats_open_generic(struct inode *inode, struct file *fp)
515 {
516         fp->private_data = inode->i_private;
517         return 0;
518 }
519
520 static ssize_t stats_read_ul(struct file *fp, char __user *ubuf, size_t count,
521                                 loff_t *off)
522 {
523         unsigned long *p = fp->private_data;
524         char buf[32];
525         int len;
526
527         len = snprintf(buf, 32, "%lu\n", *p);
528         return simple_read_from_buffer(ubuf, count, off, buf, len);
529 }
530
531 static const struct file_operations idle_fops = {
532         .open   = stats_open_generic,
533         .read   = stats_read_ul,
534 };
535
536 struct debugfs_file_info {
537         void *ptr;
538         char name[32];
539         struct dentry *file;
540 } debugfs_file_list[] = {
541                                 {&total_starts, "total_starts", NULL},
542                                 {&total_us, "total_us", NULL},
543 #ifdef DEBUG
544                                 {&past_skip, "past_skip", NULL},
545 #endif
546                                 {NULL, "", NULL}
547                         };
548
549 static int __init i7300_idle_init(void)
550 {
551         spin_lock_init(&i7300_idle_lock);
552         cpus_clear(idle_cpumask);
553         total_us = 0;
554
555         if (i7300_idle_platform_probe(&fbd_dev, &ioat_dev))
556                 return -ENODEV;
557
558         if (i7300_idle_thrt_save())
559                 return -ENODEV;
560
561         if (i7300_idle_ioat_init())
562                 return -ENODEV;
563
564         debugfs_dir = debugfs_create_dir("i7300_idle", NULL);
565         if (debugfs_dir) {
566                 int i = 0;
567
568                 while (debugfs_file_list[i].ptr != NULL) {
569                         debugfs_file_list[i].file = debugfs_create_file(
570                                         debugfs_file_list[i].name,
571                                         S_IRUSR,
572                                         debugfs_dir,
573                                         debugfs_file_list[i].ptr,
574                                         &idle_fops);
575                         i++;
576                 }
577         }
578
579         idle_notifier_register(&i7300_idle_nb);
580
581         printk(KERN_INFO "i7300_idle: loaded v%s\n", I7300_IDLE_DRIVER_VERSION);
582         return 0;
583 }
584
585 static void __exit i7300_idle_exit(void)
586 {
587         idle_notifier_unregister(&i7300_idle_nb);
588
589         if (debugfs_dir) {
590                 int i = 0;
591
592                 while (debugfs_file_list[i].file != NULL) {
593                         debugfs_remove(debugfs_file_list[i].file);
594                         i++;
595                 }
596
597                 debugfs_remove(debugfs_dir);
598         }
599         i7300_idle_thrt_restore();
600         i7300_idle_ioat_exit();
601 }
602
603 module_init(i7300_idle_init);
604 module_exit(i7300_idle_exit);
605
606 MODULE_AUTHOR("Andy Henroid <andrew.d.henroid@intel.com>");
607 MODULE_DESCRIPTION("Intel Chipset DIMM Idle Power Saving Driver v"
608                         I7300_IDLE_DRIVER_VERSION);
609 MODULE_LICENSE("GPL");