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>
31 #include "../dma/ioatdma_hw.h"
32 #include "../dma/ioatdma_registers.h"
34 #define I7300_IDLE_DRIVER_VERSION "1.55"
35 #define I7300_PRINT "i7300_idle:"
38 module_param_named(debug, debug, uint, 0644);
39 MODULE_PARM_DESC(debug, "Enable debug printks in this driver");
41 #define dprintk(fmt, arg...) \
42 do { if (debug) printk(KERN_INFO I7300_PRINT fmt, ##arg); } while (0)
45 * Value to set THRTLOW to when initiating throttling
47 * 1 = Throttle when > 4 activations per eval window (Maximum throttling)
48 * 2 = Throttle when > 8 activations
49 * 168 = Throttle when > 168 activations (Minimum throttling)
51 #define MAX_THRTLWLIMIT 168
52 static uint i7300_idle_thrtlowlm = 1;
53 module_param_named(thrtlwlimit, i7300_idle_thrtlowlm, uint, 0644);
54 MODULE_PARM_DESC(thrtlwlimit,
55 "Value for THRTLOWLM activation field "
56 "(0 = disable throttle, 1 = Max throttle, 168 = Min throttle)");
59 * simple invocation and duration statistics
61 static unsigned long total_starts;
62 static unsigned long total_us;
65 static unsigned long past_skip;
68 static struct pci_dev *fbd_dev;
70 static spinlock_t i7300_idle_lock;
71 static int i7300_idle_active;
73 static u8 i7300_idle_thrtctl_saved;
74 static u8 i7300_idle_thrtlow_saved;
75 static u32 i7300_idle_mc_saved;
77 static cpumask_t idle_cpumask;
78 static ktime_t start_ktime;
79 static unsigned long avg_idle_us;
81 static struct dentry *debugfs_dir;
83 /* Begin: I/O AT Helper routines */
85 #define IOAT_CHANBASE(ioat_ctl, chan) (ioat_ctl + 0x80 + 0x80 * chan)
86 /* Snoop control (disable snoops when coherency is not important) */
87 #define IOAT_DESC_SADDR_SNP_CTL (1UL << 1)
88 #define IOAT_DESC_DADDR_SNP_CTL (1UL << 2)
90 static struct pci_dev *ioat_dev;
91 static struct ioat_dma_descriptor *ioat_desc; /* I/O AT desc & data (1 page) */
92 static unsigned long ioat_desc_phys;
93 static u8 *ioat_iomap; /* I/O AT memory-mapped control regs (aka CB_BAR) */
94 static u8 *ioat_chanbase;
96 /* Start I/O AT memory copy */
97 static int i7300_idle_ioat_start(void)
100 /* Clear error (due to circular descriptor pointer) */
101 err = readl(ioat_chanbase + IOAT_CHANERR_OFFSET);
103 writel(err, ioat_chanbase + IOAT_CHANERR_OFFSET);
105 writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
109 /* Stop I/O AT memory copy */
110 static void i7300_idle_ioat_stop(void)
115 for (i = 0; i < 5; i++) {
116 writeb(IOAT_CHANCMD_RESET,
117 ioat_chanbase + IOAT1_CHANCMD_OFFSET);
121 sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
122 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
124 if (sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE)
130 dprintk("failed to suspend+reset I/O AT after 5 retries\n");
134 /* Test I/O AT by copying 1024 byte from 2k to 1k */
135 static int __init i7300_idle_ioat_selftest(u8 *ctl,
136 struct ioat_dma_descriptor *desc, unsigned long desc_phys)
140 memset(desc, 0, 2048);
141 memset((u8 *) desc + 2048, 0xab, 1024);
145 desc[0].src_addr = desc_phys + 2048;
146 desc[0].dst_addr = desc_phys + 1024;
149 writeb(IOAT_CHANCMD_RESET, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
150 writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
154 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
155 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
157 if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_DONE) {
158 /* Not complete, reset the channel */
159 writeb(IOAT_CHANCMD_RESET,
160 ioat_chanbase + IOAT1_CHANCMD_OFFSET);
164 if (*(u32 *) ((u8 *) desc + 3068) != 0xabababab ||
165 *(u32 *) ((u8 *) desc + 2044) != 0xabababab) {
166 dprintk("Data values src 0x%x, dest 0x%x, memset 0x%x\n",
167 *(u32 *) ((u8 *) desc + 2048),
168 *(u32 *) ((u8 *) desc + 1024),
169 *(u32 *) ((u8 *) desc + 3072));
175 static struct device dummy_dma_dev = {
176 .bus_id = "fallback device",
177 .coherent_dma_mask = DMA_64BIT_MASK,
178 .dma_mask = &dummy_dma_dev.coherent_dma_mask,
181 /* Setup and initialize I/O AT */
182 /* This driver needs I/O AT as the throttling takes effect only when there is
183 * some memory activity. We use I/O AT to set up a dummy copy, while all CPUs
184 * go idle and memory is throttled.
186 static int __init i7300_idle_ioat_init(void)
188 u8 ver, chan_count, ioat_chan;
191 ioat_iomap = (u8 *) ioremap_nocache(pci_resource_start(ioat_dev, 0),
192 pci_resource_len(ioat_dev, 0));
195 printk(KERN_ERR I7300_PRINT "failed to map I/O AT registers\n");
199 ver = readb(ioat_iomap + IOAT_VER_OFFSET);
200 if (ver != IOAT_VER_1_2) {
201 printk(KERN_ERR I7300_PRINT "unknown I/O AT version (%u.%u)\n",
202 ver >> 4, ver & 0xf);
206 chan_count = readb(ioat_iomap + IOAT_CHANCNT_OFFSET);
208 printk(KERN_ERR I7300_PRINT "unexpected # of I/O AT channels "
214 ioat_chan = chan_count - 1;
215 ioat_chanbase = IOAT_CHANBASE(ioat_iomap, ioat_chan);
217 chan_ctl = readw(ioat_chanbase + IOAT_CHANCTRL_OFFSET);
218 if (chan_ctl & IOAT_CHANCTRL_CHANNEL_IN_USE) {
219 printk(KERN_ERR I7300_PRINT "channel %d in use\n", ioat_chan);
223 writew(IOAT_CHANCTRL_CHANNEL_IN_USE,
224 ioat_chanbase + IOAT_CHANCTRL_OFFSET);
226 ioat_desc = (struct ioat_dma_descriptor *)dma_alloc_coherent(
227 &dummy_dma_dev, 4096,
228 (dma_addr_t *)&ioat_desc_phys, GFP_KERNEL);
230 printk(KERN_ERR I7300_PRINT "failed to allocate I/O AT desc\n");
231 goto err_mark_unused;
234 writel(ioat_desc_phys & 0xffffffffUL,
235 ioat_chanbase + IOAT1_CHAINADDR_OFFSET_LOW);
236 writel(ioat_desc_phys >> 32,
237 ioat_chanbase + IOAT1_CHAINADDR_OFFSET_HIGH);
239 if (i7300_idle_ioat_selftest(ioat_iomap, ioat_desc, ioat_desc_phys)) {
240 printk(KERN_ERR I7300_PRINT "I/O AT self-test failed\n");
244 /* Setup circular I/O AT descriptor chain */
245 ioat_desc[0].ctl = IOAT_DESC_SADDR_SNP_CTL | IOAT_DESC_DADDR_SNP_CTL;
246 ioat_desc[0].src_addr = ioat_desc_phys + 2048;
247 ioat_desc[0].dst_addr = ioat_desc_phys + 3072;
248 ioat_desc[0].size = 128;
249 ioat_desc[0].next = ioat_desc_phys + sizeof(struct ioat_dma_descriptor);
251 ioat_desc[1].ctl = ioat_desc[0].ctl;
252 ioat_desc[1].src_addr = ioat_desc[0].src_addr;
253 ioat_desc[1].dst_addr = ioat_desc[0].dst_addr;
254 ioat_desc[1].size = ioat_desc[0].size;
255 ioat_desc[1].next = ioat_desc_phys;
260 dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
262 writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
270 static void __exit i7300_idle_ioat_exit(void)
275 i7300_idle_ioat_stop();
277 /* Wait for a while for the channel to halt before releasing */
278 for (i = 0; i < 10; i++) {
279 writeb(IOAT_CHANCMD_RESET,
280 ioat_chanbase + IOAT1_CHANCMD_OFFSET);
282 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
283 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
285 if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
286 writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
292 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
293 IOAT_CHANSTS_DMA_TRANSFER_STATUS;
296 * We tried to reset multiple times. If IO A/T channel is still active
297 * flag an error and return without cleanup. Memory leak is better
298 * than random corruption in that extreme error situation.
300 if (chan_sts == IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
301 printk(KERN_ERR I7300_PRINT "Unable to stop IO A/T channels."
302 " Not freeing resources\n");
306 dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
310 /* End: I/O AT Helper routines */
312 #define DIMM_THRTLOW 0x64
313 #define DIMM_THRTCTL 0x67
314 #define DIMM_THRTCTL_THRMHUNT (1UL << 0)
316 #define DIMM_GTW_MODE (1UL << 17)
317 #define DIMM_GBLACT 0x60
320 * Keep track of an exponential-decaying average of recent idle durations.
321 * The latest duration gets DURATION_WEIGHT_PCT percentage weight
322 * in this average, with the old average getting the remaining weight.
324 * High weights emphasize recent history, low weights include long history.
326 #define DURATION_WEIGHT_PCT 55
329 * When the decaying average of recent durations or the predicted duration
330 * of the next timer interrupt is shorter than duration_threshold, the
331 * driver will decline to throttle.
333 #define DURATION_THRESHOLD_US 100
336 /* Store DIMM thermal throttle configuration */
337 static int i7300_idle_thrt_save(void)
342 pci_read_config_byte(fbd_dev, DIMM_THRTCTL, &i7300_idle_thrtctl_saved);
343 pci_read_config_byte(fbd_dev, DIMM_THRTLOW, &i7300_idle_thrtlow_saved);
344 pci_read_config_dword(fbd_dev, DIMM_MC, &i7300_idle_mc_saved);
346 * Make sure we have Global Throttling Window Mode set to have a
347 * "short" window. This (mostly) works around an issue where
348 * throttling persists until the end of the global throttling window
349 * size. On the tested system, this was resulting in a maximum of
350 * 64 ms to exit throttling (average 32 ms). The actual numbers
351 * depends on system frequencies. Setting the short window reduces
352 * this by a factor of 4096.
354 * We will only do this only if the system is set for
355 * unlimited-activations while in open-loop throttling (i.e., when
356 * Global Activation Throttle Limit is zero).
358 pci_read_config_byte(fbd_dev, DIMM_GBLACT, &gblactlm);
359 dprintk("thrtctl_saved = 0x%02x, thrtlow_saved = 0x%02x\n",
360 i7300_idle_thrtctl_saved,
361 i7300_idle_thrtlow_saved);
362 dprintk("mc_saved = 0x%08x, gblactlm = 0x%02x\n",
366 new_mc_val = i7300_idle_mc_saved | DIMM_GTW_MODE;
367 pci_write_config_dword(fbd_dev, DIMM_MC, new_mc_val);
370 dprintk("could not set GTW_MODE = 1 (OLTT enabled)\n");
375 /* Restore DIMM thermal throttle configuration */
376 static void i7300_idle_thrt_restore(void)
378 pci_write_config_dword(fbd_dev, DIMM_MC, i7300_idle_mc_saved);
379 pci_write_config_byte(fbd_dev, DIMM_THRTLOW, i7300_idle_thrtlow_saved);
380 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, i7300_idle_thrtctl_saved);
383 /* Enable DIMM thermal throttling */
384 static void i7300_idle_start(void)
389 new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
390 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
392 limit = i7300_idle_thrtlowlm;
393 if (unlikely(limit > MAX_THRTLWLIMIT))
394 limit = MAX_THRTLWLIMIT;
396 pci_write_config_byte(fbd_dev, DIMM_THRTLOW, limit);
398 new_ctl = i7300_idle_thrtctl_saved | DIMM_THRTCTL_THRMHUNT;
399 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
402 /* Disable DIMM thermal throttling */
403 static void i7300_idle_stop(void)
408 new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
409 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
411 pci_write_config_byte(fbd_dev, DIMM_THRTLOW, i7300_idle_thrtlow_saved);
412 pci_write_config_byte(fbd_dev, DIMM_THRTCTL, i7300_idle_thrtctl_saved);
413 pci_read_config_byte(fbd_dev, DIMM_THRTCTL, &got_ctl);
414 WARN_ON_ONCE(got_ctl != i7300_idle_thrtctl_saved);
419 * i7300_avg_duration_check()
420 * return 0 if the decaying average of recent idle durations is
421 * more than DURATION_THRESHOLD_US
423 static int i7300_avg_duration_check(void)
425 if (avg_idle_us >= DURATION_THRESHOLD_US)
434 /* Idle notifier to look at idle CPUs */
435 static int i7300_idle_notifier(struct notifier_block *nb, unsigned long val,
440 static ktime_t idle_begin_time;
441 static int time_init = 1;
443 if (!i7300_idle_thrtlowlm)
446 if (unlikely(time_init)) {
448 idle_begin_time = ktime_get();
451 spin_lock_irqsave(&i7300_idle_lock, flags);
452 if (val == IDLE_START) {
454 cpu_set(smp_processor_id(), idle_cpumask);
456 if (cpus_weight(idle_cpumask) != num_online_cpus())
459 now_ktime = ktime_get();
460 idle_begin_time = now_ktime;
462 if (i7300_avg_duration_check())
465 i7300_idle_active = 1;
467 start_ktime = now_ktime;
470 i7300_idle_ioat_start();
472 } else if (val == IDLE_END) {
473 cpu_clear(smp_processor_id(), idle_cpumask);
474 if (cpus_weight(idle_cpumask) == (num_online_cpus() - 1)) {
475 /* First CPU coming out of idle */
476 u64 idle_duration_us;
478 now_ktime = ktime_get();
480 idle_duration_us = ktime_to_us(ktime_sub
481 (now_ktime, idle_begin_time));
484 ((100 - DURATION_WEIGHT_PCT) * avg_idle_us +
485 DURATION_WEIGHT_PCT * idle_duration_us) / 100;
487 if (i7300_idle_active) {
490 idle_ktime = ktime_sub(now_ktime, start_ktime);
491 total_us += ktime_to_us(idle_ktime);
493 i7300_idle_ioat_stop();
495 i7300_idle_active = 0;
500 spin_unlock_irqrestore(&i7300_idle_lock, flags);
504 static struct notifier_block i7300_idle_nb = {
505 .notifier_call = i7300_idle_notifier,
509 * I/O AT controls (PCI bus 0 device 8 function 0)
510 * DIMM controls (PCI bus 0 device 16 function 1)
513 #define IOAT_DEVFN PCI_DEVFN(8, 0)
515 #define MEMCTL_DEVFN PCI_DEVFN(16, 1)
519 unsigned int ioat_dev;
523 * The i5000 chip-set has the same hooks as the i7300
524 * but support is disabled by default because this driver
525 * has not been validated on that platform.
527 #define SUPPORT_I5000 0
529 static const struct fbd_ioat fbd_ioat_list[] = {
530 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IOAT_CNB},
532 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IOAT},
537 /* table of devices that work with this driver */
538 static const struct pci_device_id pci_tbl[] = {
539 { PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_FBD_CNB) },
541 { PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_5000_ERR) },
543 { } /* Terminating entry */
546 MODULE_DEVICE_TABLE(pci, pci_tbl);
548 /* Check for known platforms with I/O-AT */
549 static int __init i7300_idle_platform_probe(void)
553 fbd_dev = pci_get_bus_and_slot(MEMCTL_BUS, MEMCTL_DEVFN);
557 for (i = 0; pci_tbl[i].vendor != 0; i++) {
558 if (fbd_dev->vendor == pci_tbl[i].vendor &&
559 fbd_dev->device == pci_tbl[i].device) {
563 if (pci_tbl[i].vendor == 0)
566 ioat_dev = pci_get_bus_and_slot(IOAT_BUS, IOAT_DEVFN);
570 for (i = 0; fbd_ioat_list[i].vendor != 0; i++) {
571 if (ioat_dev->vendor == fbd_ioat_list[i].vendor &&
572 ioat_dev->device == fbd_ioat_list[i].ioat_dev) {
579 int stats_open_generic(struct inode *inode, struct file *fp)
581 fp->private_data = inode->i_private;
585 static ssize_t stats_read_ul(struct file *fp, char __user *ubuf, size_t count,
588 unsigned long *p = fp->private_data;
592 len = snprintf(buf, 32, "%lu\n", *p);
593 return simple_read_from_buffer(ubuf, count, off, buf, len);
596 static const struct file_operations idle_fops = {
597 .open = stats_open_generic,
598 .read = stats_read_ul,
601 struct debugfs_file_info {
605 } debugfs_file_list[] = {
606 {&total_starts, "total_starts", NULL},
607 {&total_us, "total_us", NULL},
609 {&past_skip, "past_skip", NULL},
614 static int __init i7300_idle_init(void)
616 spin_lock_init(&i7300_idle_lock);
617 cpus_clear(idle_cpumask);
620 if (i7300_idle_platform_probe())
623 if (i7300_idle_thrt_save())
626 if (i7300_idle_ioat_init())
629 debugfs_dir = debugfs_create_dir("i7300_idle", NULL);
633 while (debugfs_file_list[i].ptr != NULL) {
634 debugfs_file_list[i].file = debugfs_create_file(
635 debugfs_file_list[i].name,
638 debugfs_file_list[i].ptr,
644 idle_notifier_register(&i7300_idle_nb);
646 printk(KERN_INFO "i7300_idle: loaded v%s\n", I7300_IDLE_DRIVER_VERSION);
650 static void __exit i7300_idle_exit(void)
652 idle_notifier_unregister(&i7300_idle_nb);
657 while (debugfs_file_list[i].file != NULL) {
658 debugfs_remove(debugfs_file_list[i].file);
662 debugfs_remove(debugfs_dir);
664 i7300_idle_thrt_restore();
665 i7300_idle_ioat_exit();
668 module_init(i7300_idle_init);
669 module_exit(i7300_idle_exit);
671 MODULE_AUTHOR("Andy Henroid <andrew.d.henroid@intel.com>");
672 MODULE_DESCRIPTION("Intel Chipset DIMM Idle Power Saving Driver v"
673 I7300_IDLE_DRIVER_VERSION);
674 MODULE_LICENSE("GPL");