2 * (C) Copyright 2008 Intel Corporation
4 * Andy Henroid <andrew.d.henroid@intel.com>
5 * Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
9 * Save DIMM power on Intel 7300-based platforms when all CPUs/cores
10 * are idle, using the DIMM thermal throttling capability.
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.
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>
32 #include "../dma/ioatdma_hw.h"
33 #include "../dma/ioatdma_registers.h"
35 #define I7300_IDLE_DRIVER_VERSION "1.55"
36 #define I7300_PRINT "i7300_idle:"
38 #define MAX_STOP_RETRIES 10
41 module_param_named(debug, debug, uint, 0644);
42 MODULE_PARM_DESC(debug, "Enable debug printks in this driver");
44 #define dprintk(fmt, arg...) \
45 do { if (debug) printk(KERN_INFO I7300_PRINT fmt, ##arg); } while (0)
48 * Value to set THRTLOW to when initiating 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)
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)");
62 * simple invocation and duration statistics
64 static unsigned long total_starts;
65 static unsigned long total_us;
68 static unsigned long past_skip;
71 static struct pci_dev *fbd_dev;
73 static spinlock_t i7300_idle_lock;
74 static int i7300_idle_active;
76 static u8 i7300_idle_thrtctl_saved;
77 static u8 i7300_idle_thrtlow_saved;
78 static u32 i7300_idle_mc_saved;
80 static cpumask_t idle_cpumask;
81 static ktime_t start_ktime;
82 static unsigned long avg_idle_us;
84 static struct dentry *debugfs_dir;
86 /* Begin: I/O AT Helper routines */
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)
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;
99 /* Start I/O AT memory copy */
100 static int i7300_idle_ioat_start(void)
103 /* Clear error (due to circular descriptor pointer) */
104 err = readl(ioat_chanbase + IOAT_CHANERR_OFFSET);
106 writel(err, ioat_chanbase + IOAT_CHANERR_OFFSET);
108 writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
112 /* Stop I/O AT memory copy */
113 static void i7300_idle_ioat_stop(void)
118 for (i = 0; i < MAX_STOP_RETRIES; i++) {
119 writeb(IOAT_CHANCMD_RESET,
120 ioat_chanbase + IOAT1_CHANCMD_OFFSET);
124 sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
125 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
127 if (sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE)
132 if (i == MAX_STOP_RETRIES) {
133 dprintk("failed to stop I/O AT after %d retries\n",
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)
144 memset(desc, 0, 2048);
145 memset((u8 *) desc + 2048, 0xab, 1024);
149 desc[0].src_addr = desc_phys + 2048;
150 desc[0].dst_addr = desc_phys + 1024;
153 writeb(IOAT_CHANCMD_RESET, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
154 writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
158 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
159 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
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);
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));
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,
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.
190 static int __init i7300_idle_ioat_init(void)
192 u8 ver, chan_count, ioat_chan;
195 ioat_iomap = (u8 *) ioremap_nocache(pci_resource_start(ioat_dev, 0),
196 pci_resource_len(ioat_dev, 0));
199 printk(KERN_ERR I7300_PRINT "failed to map I/O AT registers\n");
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);
210 chan_count = readb(ioat_iomap + IOAT_CHANCNT_OFFSET);
212 printk(KERN_ERR I7300_PRINT "unexpected # of I/O AT channels "
218 ioat_chan = chan_count - 1;
219 ioat_chanbase = IOAT_CHANBASE(ioat_iomap, ioat_chan);
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);
227 writew(IOAT_CHANCTRL_CHANNEL_IN_USE,
228 ioat_chanbase + IOAT_CHANCTRL_OFFSET);
230 ioat_desc = (struct ioat_dma_descriptor *)dma_alloc_coherent(
231 &dummy_dma_dev, 4096,
232 (dma_addr_t *)&ioat_desc_phys, GFP_KERNEL);
234 printk(KERN_ERR I7300_PRINT "failed to allocate I/O AT desc\n");
235 goto err_mark_unused;
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);
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");
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);
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;
264 dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
266 writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
274 static void __exit i7300_idle_ioat_exit(void)
279 i7300_idle_ioat_stop();
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);
286 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
287 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
289 if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
290 writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
296 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
297 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
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.
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");
310 dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
314 /* End: I/O AT Helper routines */
316 #define DIMM_THRTLOW 0x64
317 #define DIMM_THRTCTL 0x67
318 #define DIMM_THRTCTL_THRMHUNT (1UL << 0)
320 #define DIMM_GTW_MODE (1UL << 17)
321 #define DIMM_GBLACT 0x60
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.
328 * High weights emphasize recent history, low weights include long history.
330 #define DURATION_WEIGHT_PCT 55
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.
337 #define DURATION_THRESHOLD_US 100
340 /* Store DIMM thermal throttle configuration */
341 static int i7300_idle_thrt_save(void)
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);
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.
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).
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",
370 new_mc_val = i7300_idle_mc_saved | DIMM_GTW_MODE;
371 pci_write_config_dword(fbd_dev, DIMM_MC, new_mc_val);
374 dprintk("could not set GTW_MODE = 1 (OLTT enabled)\n");
379 /* Restore DIMM thermal throttle configuration */
380 static void i7300_idle_thrt_restore(void)
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);
387 /* Enable DIMM thermal throttling */
388 static void i7300_idle_start(void)
393 new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
394 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
396 limit = throttle_low_limit;
397 if (unlikely(limit > MAX_THROTTLE_LOW_LIMIT))
398 limit = MAX_THROTTLE_LOW_LIMIT;
400 pci_write_config_byte(fbd_dev, DIMM_THRTLOW, limit);
402 new_ctl = i7300_idle_thrtctl_saved | DIMM_THRTCTL_THRMHUNT;
403 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
406 /* Disable DIMM thermal throttling */
407 static void i7300_idle_stop(void)
412 new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
413 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
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);
423 * i7300_avg_duration_check()
424 * return 0 if the decaying average of recent idle durations is
425 * more than DURATION_THRESHOLD_US
427 static int i7300_avg_duration_check(void)
429 if (avg_idle_us >= DURATION_THRESHOLD_US)
438 /* Idle notifier to look at idle CPUs */
439 static int i7300_idle_notifier(struct notifier_block *nb, unsigned long val,
444 static ktime_t idle_begin_time;
445 static int time_init = 1;
447 if (!throttle_low_limit)
450 if (unlikely(time_init)) {
452 idle_begin_time = ktime_get();
455 spin_lock_irqsave(&i7300_idle_lock, flags);
456 if (val == IDLE_START) {
458 cpu_set(smp_processor_id(), idle_cpumask);
460 if (cpus_weight(idle_cpumask) != num_online_cpus())
463 now_ktime = ktime_get();
464 idle_begin_time = now_ktime;
466 if (i7300_avg_duration_check())
469 i7300_idle_active = 1;
471 start_ktime = now_ktime;
474 i7300_idle_ioat_start();
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;
482 now_ktime = ktime_get();
484 idle_duration_us = ktime_to_us(ktime_sub
485 (now_ktime, idle_begin_time));
488 ((100 - DURATION_WEIGHT_PCT) * avg_idle_us +
489 DURATION_WEIGHT_PCT * idle_duration_us) / 100;
491 if (i7300_idle_active) {
494 idle_ktime = ktime_sub(now_ktime, start_ktime);
495 total_us += ktime_to_us(idle_ktime);
497 i7300_idle_ioat_stop();
499 i7300_idle_active = 0;
504 spin_unlock_irqrestore(&i7300_idle_lock, flags);
508 static struct notifier_block i7300_idle_nb = {
509 .notifier_call = i7300_idle_notifier,
512 MODULE_DEVICE_TABLE(pci, pci_tbl);
514 int stats_open_generic(struct inode *inode, struct file *fp)
516 fp->private_data = inode->i_private;
520 static ssize_t stats_read_ul(struct file *fp, char __user *ubuf, size_t count,
523 unsigned long *p = fp->private_data;
527 len = snprintf(buf, 32, "%lu\n", *p);
528 return simple_read_from_buffer(ubuf, count, off, buf, len);
531 static const struct file_operations idle_fops = {
532 .open = stats_open_generic,
533 .read = stats_read_ul,
536 struct debugfs_file_info {
540 } debugfs_file_list[] = {
541 {&total_starts, "total_starts", NULL},
542 {&total_us, "total_us", NULL},
544 {&past_skip, "past_skip", NULL},
549 static int __init i7300_idle_init(void)
551 spin_lock_init(&i7300_idle_lock);
552 cpus_clear(idle_cpumask);
555 if (i7300_idle_platform_probe(&fbd_dev, &ioat_dev))
558 if (i7300_idle_thrt_save())
561 if (i7300_idle_ioat_init())
564 debugfs_dir = debugfs_create_dir("i7300_idle", NULL);
568 while (debugfs_file_list[i].ptr != NULL) {
569 debugfs_file_list[i].file = debugfs_create_file(
570 debugfs_file_list[i].name,
573 debugfs_file_list[i].ptr,
579 idle_notifier_register(&i7300_idle_nb);
581 printk(KERN_INFO "i7300_idle: loaded v%s\n", I7300_IDLE_DRIVER_VERSION);
585 static void __exit i7300_idle_exit(void)
587 idle_notifier_unregister(&i7300_idle_nb);
592 while (debugfs_file_list[i].file != NULL) {
593 debugfs_remove(debugfs_file_list[i].file);
597 debugfs_remove(debugfs_dir);
599 i7300_idle_thrt_restore();
600 i7300_idle_ioat_exit();
603 module_init(i7300_idle_init);
604 module_exit(i7300_idle_exit);
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");