xref: /linux/arch/parisc/kernel/perf.c (revision 80d443e8876602be2c130f79c4de81e12e2a700d)
1 /*
2  *  Parisc performance counters
3  *  Copyright (C) 2001 Randolph Chung <tausq@debian.org>
4  *
5  *  This code is derived, with permission, from HP/UX sources.
6  *
7  *    This program is free software; you can redistribute it and/or modify
8  *    it under the terms of the GNU General Public License as published by
9  *    the Free Software Foundation; either version 2, or (at your option)
10  *    any later version.
11  *
12  *    This program is distributed in the hope that it will be useful,
13  *    but WITHOUT ANY WARRANTY; without even the implied warranty of
14  *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  *    GNU General Public License for more details.
16  *
17  *    You should have received a copy of the GNU General Public License
18  *    along with this program; if not, write to the Free Software
19  *    Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  */
21 
22 /*
23  *  Edited comment from original sources:
24  *
25  *  This driver programs the PCX-U/PCX-W performance counters
26  *  on the PA-RISC 2.0 chips.  The driver keeps all images now
27  *  internally to the kernel to hopefully eliminate the possibility
28  *  of a bad image halting the CPU.  Also, there are different
29  *  images for the PCX-W and later chips vs the PCX-U chips.
30  *
31  *  Only 1 process is allowed to access the driver at any time,
32  *  so the only protection that is needed is at open and close.
33  *  A variable "perf_enabled" is used to hold the state of the
34  *  driver.  The spinlock "perf_lock" is used to protect the
35  *  modification of the state during open/close operations so
36  *  multiple processes don't get into the driver simultaneously.
37  *
38  *  This driver accesses the processor directly vs going through
39  *  the PDC INTRIGUE calls.  This is done to eliminate bugs introduced
40  *  in various PDC revisions.  The code is much more maintainable
41  *  and reliable this way vs having to debug on every version of PDC
42  *  on every box.
43  */
44 
45 #include <linux/capability.h>
46 #include <linux/init.h>
47 #include <linux/proc_fs.h>
48 #include <linux/miscdevice.h>
49 #include <linux/spinlock.h>
50 
51 #include <linux/uaccess.h>
52 #include <asm/perf.h>
53 #include <asm/parisc-device.h>
54 #include <asm/processor.h>
55 #include <asm/runway.h>
56 #include <asm/io.h>		/* for __raw_read() */
57 
58 #include "perf_images.h"
59 
60 #define MAX_RDR_WORDS	24
61 #define PERF_VERSION	2	/* derived from hpux's PI v2 interface */
62 
63 /* definition of RDR regs */
64 struct rdr_tbl_ent {
65 	uint16_t	width;
66 	uint8_t		num_words;
67 	uint8_t		write_control;
68 };
69 
70 static int perf_processor_interface __read_mostly = UNKNOWN_INTF;
71 static int perf_enabled __read_mostly;
72 static spinlock_t perf_lock;
73 struct parisc_device *cpu_device __read_mostly;
74 
75 /* RDRs to write for PCX-W */
76 static const int perf_rdrs_W[] =
77 	{ 0, 1, 4, 5, 6, 15, 16, 17, 18, 20, 21, 22, 23, 24, 25, -1 };
78 
79 /* RDRs to write for PCX-U */
80 static const int perf_rdrs_U[] =
81 	{ 0, 1, 4, 5, 6, 7, 16, 17, 18, 20, 21, 22, 23, 24, 25, -1 };
82 
83 /* RDR register descriptions for PCX-W */
84 static const struct rdr_tbl_ent perf_rdr_tbl_W[] = {
85 	{ 19,	1,	8 },   /* RDR 0 */
86 	{ 16,	1,	16 },  /* RDR 1 */
87 	{ 72,	2,	0 },   /* RDR 2 */
88 	{ 81,	2,	0 },   /* RDR 3 */
89 	{ 328,	6,	0 },   /* RDR 4 */
90 	{ 160,	3,	0 },   /* RDR 5 */
91 	{ 336,	6,	0 },   /* RDR 6 */
92 	{ 164,	3,	0 },   /* RDR 7 */
93 	{ 0,	0,	0 },   /* RDR 8 */
94 	{ 35,	1,	0 },   /* RDR 9 */
95 	{ 6,	1,	0 },   /* RDR 10 */
96 	{ 18,	1,	0 },   /* RDR 11 */
97 	{ 13,	1,	0 },   /* RDR 12 */
98 	{ 8,	1,	0 },   /* RDR 13 */
99 	{ 8,	1,	0 },   /* RDR 14 */
100 	{ 8,	1,	0 },   /* RDR 15 */
101 	{ 1530,	24,	0 },   /* RDR 16 */
102 	{ 16,	1,	0 },   /* RDR 17 */
103 	{ 4,	1,	0 },   /* RDR 18 */
104 	{ 0,	0,	0 },   /* RDR 19 */
105 	{ 152,	3,	24 },  /* RDR 20 */
106 	{ 152,	3,	24 },  /* RDR 21 */
107 	{ 233,	4,	48 },  /* RDR 22 */
108 	{ 233,	4,	48 },  /* RDR 23 */
109 	{ 71,	2,	0 },   /* RDR 24 */
110 	{ 71,	2,	0 },   /* RDR 25 */
111 	{ 11,	1,	0 },   /* RDR 26 */
112 	{ 18,	1,	0 },   /* RDR 27 */
113 	{ 128,	2,	0 },   /* RDR 28 */
114 	{ 0,	0,	0 },   /* RDR 29 */
115 	{ 16,	1,	0 },   /* RDR 30 */
116 	{ 16,	1,	0 },   /* RDR 31 */
117 };
118 
119 /* RDR register descriptions for PCX-U */
120 static const struct rdr_tbl_ent perf_rdr_tbl_U[] = {
121 	{ 19,	1,	8 },              /* RDR 0 */
122 	{ 32,	1,	16 },             /* RDR 1 */
123 	{ 20,	1,	0 },              /* RDR 2 */
124 	{ 0,	0,	0 },              /* RDR 3 */
125 	{ 344,	6,	0 },              /* RDR 4 */
126 	{ 176,	3,	0 },              /* RDR 5 */
127 	{ 336,	6,	0 },              /* RDR 6 */
128 	{ 0,	0,	0 },              /* RDR 7 */
129 	{ 0,	0,	0 },              /* RDR 8 */
130 	{ 0,	0,	0 },              /* RDR 9 */
131 	{ 28,	1,	0 },              /* RDR 10 */
132 	{ 33,	1,	0 },              /* RDR 11 */
133 	{ 0,	0,	0 },              /* RDR 12 */
134 	{ 230,	4,	0 },              /* RDR 13 */
135 	{ 32,	1,	0 },              /* RDR 14 */
136 	{ 128,	2,	0 },              /* RDR 15 */
137 	{ 1494,	24,	0 },              /* RDR 16 */
138 	{ 18,	1,	0 },              /* RDR 17 */
139 	{ 4,	1,	0 },              /* RDR 18 */
140 	{ 0,	0,	0 },              /* RDR 19 */
141 	{ 158,	3,	24 },             /* RDR 20 */
142 	{ 158,	3,	24 },             /* RDR 21 */
143 	{ 194,	4,	48 },             /* RDR 22 */
144 	{ 194,	4,	48 },             /* RDR 23 */
145 	{ 71,	2,	0 },              /* RDR 24 */
146 	{ 71,	2,	0 },              /* RDR 25 */
147 	{ 28,	1,	0 },              /* RDR 26 */
148 	{ 33,	1,	0 },              /* RDR 27 */
149 	{ 88,	2,	0 },              /* RDR 28 */
150 	{ 32,	1,	0 },              /* RDR 29 */
151 	{ 24,	1,	0 },              /* RDR 30 */
152 	{ 16,	1,	0 },              /* RDR 31 */
153 };
154 
155 /*
156  * A non-zero write_control in the above tables is a byte offset into
157  * this array.
158  */
159 static const uint64_t perf_bitmasks[] = {
160 	0x0000000000000000ul,     /* first dbl word must be zero */
161 	0xfdffe00000000000ul,     /* RDR0 bitmask */
162 	0x003f000000000000ul,     /* RDR1 bitmask */
163 	0x00fffffffffffffful,     /* RDR20-RDR21 bitmask (152 bits) */
164 	0xfffffffffffffffful,
165 	0xfffffffc00000000ul,
166 	0xfffffffffffffffful,     /* RDR22-RDR23 bitmask (233 bits) */
167 	0xfffffffffffffffful,
168 	0xfffffffffffffffcul,
169 	0xff00000000000000ul
170 };
171 
172 /*
173  * Write control bitmasks for Pa-8700 processor given
174  * some things have changed slightly.
175  */
176 static const uint64_t perf_bitmasks_piranha[] = {
177 	0x0000000000000000ul,     /* first dbl word must be zero */
178 	0xfdffe00000000000ul,     /* RDR0 bitmask */
179 	0x003f000000000000ul,     /* RDR1 bitmask */
180 	0x00fffffffffffffful,     /* RDR20-RDR21 bitmask (158 bits) */
181 	0xfffffffffffffffful,
182 	0xfffffffc00000000ul,
183 	0xfffffffffffffffful,     /* RDR22-RDR23 bitmask (210 bits) */
184 	0xfffffffffffffffful,
185 	0xfffffffffffffffful,
186 	0xfffc000000000000ul
187 };
188 
189 static const uint64_t *bitmask_array;   /* array of bitmasks to use */
190 
191 /******************************************************************************
192  * Function Prototypes
193  *****************************************************************************/
194 static int perf_config(uint32_t *image_ptr);
195 static int perf_release(struct inode *inode, struct file *file);
196 static int perf_open(struct inode *inode, struct file *file);
197 static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t *ppos);
198 static ssize_t perf_write(struct file *file, const char __user *buf, size_t count,
199 	loff_t *ppos);
200 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
201 static void perf_start_counters(void);
202 static int perf_stop_counters(uint32_t *raddr);
203 static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num);
204 static int perf_rdr_read_ubuf(uint32_t	rdr_num, uint64_t *buffer);
205 static int perf_rdr_clear(uint32_t rdr_num);
206 static int perf_write_image(uint64_t *memaddr);
207 static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer);
208 
209 /* External Assembly Routines */
210 extern uint64_t perf_rdr_shift_in_W (uint32_t rdr_num, uint16_t width);
211 extern uint64_t perf_rdr_shift_in_U (uint32_t rdr_num, uint16_t width);
212 extern void perf_rdr_shift_out_W (uint32_t rdr_num, uint64_t buffer);
213 extern void perf_rdr_shift_out_U (uint32_t rdr_num, uint64_t buffer);
214 extern void perf_intrigue_enable_perf_counters (void);
215 extern void perf_intrigue_disable_perf_counters (void);
216 
217 /******************************************************************************
218  * Function Definitions
219  *****************************************************************************/
220 
221 
222 /*
223  * configure:
224  *
225  * Configure the cpu with a given data image.  First turn off the counters,
226  * then download the image, then turn the counters back on.
227  */
228 static int perf_config(uint32_t *image_ptr)
229 {
230 	long error;
231 	uint32_t raddr[4];
232 
233 	/* Stop the counters*/
234 	error = perf_stop_counters(raddr);
235 	if (error != 0) {
236 		printk("perf_config: perf_stop_counters = %ld\n", error);
237 		return -EINVAL;
238 	}
239 
240 printk("Preparing to write image\n");
241 	/* Write the image to the chip */
242 	error = perf_write_image((uint64_t *)image_ptr);
243 	if (error != 0) {
244 		printk("perf_config: DOWNLOAD = %ld\n", error);
245 		return -EINVAL;
246 	}
247 
248 printk("Preparing to start counters\n");
249 
250 	/* Start the counters */
251 	perf_start_counters();
252 
253 	return sizeof(uint32_t);
254 }
255 
256 /*
257  * Open the device and initialize all of its memory.  The device is only
258  * opened once, but can be "queried" by multiple processes that know its
259  * file descriptor.
260  */
261 static int perf_open(struct inode *inode, struct file *file)
262 {
263 	spin_lock(&perf_lock);
264 	if (perf_enabled) {
265 		spin_unlock(&perf_lock);
266 		return -EBUSY;
267 	}
268 	perf_enabled = 1;
269  	spin_unlock(&perf_lock);
270 
271 	return 0;
272 }
273 
274 /*
275  * Close the device.
276  */
277 static int perf_release(struct inode *inode, struct file *file)
278 {
279 	spin_lock(&perf_lock);
280 	perf_enabled = 0;
281 	spin_unlock(&perf_lock);
282 
283 	return 0;
284 }
285 
286 /*
287  * Read does nothing for this driver
288  */
289 static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t *ppos)
290 {
291 	return 0;
292 }
293 
294 /*
295  * write:
296  *
297  * This routine downloads the image to the chip.  It must be
298  * called on the processor that the download should happen
299  * on.
300  */
301 static ssize_t perf_write(struct file *file, const char __user *buf, size_t count,
302 	loff_t *ppos)
303 {
304 	size_t image_size;
305 	uint32_t image_type;
306 	uint32_t interface_type;
307 	uint32_t test;
308 
309 	if (perf_processor_interface == ONYX_INTF)
310 		image_size = PCXU_IMAGE_SIZE;
311 	else if (perf_processor_interface == CUDA_INTF)
312 		image_size = PCXW_IMAGE_SIZE;
313 	else
314 		return -EFAULT;
315 
316 	if (!capable(CAP_SYS_ADMIN))
317 		return -EACCES;
318 
319 	if (count != sizeof(uint32_t))
320 		return -EIO;
321 
322 	if (copy_from_user(&image_type, buf, sizeof(uint32_t)))
323 		return -EFAULT;
324 
325 	/* Get the interface type and test type */
326    	interface_type = (image_type >> 16) & 0xffff;
327 	test           = (image_type & 0xffff);
328 
329 	/* Make sure everything makes sense */
330 
331 	/* First check the machine type is correct for
332 	   the requested image */
333         if (((perf_processor_interface == CUDA_INTF) &&
334 		       (interface_type != CUDA_INTF)) ||
335 	    ((perf_processor_interface == ONYX_INTF) &&
336 	               (interface_type != ONYX_INTF)))
337 		return -EINVAL;
338 
339 	/* Next check to make sure the requested image
340 	   is valid */
341 	if (((interface_type == CUDA_INTF) &&
342 		       (test >= MAX_CUDA_IMAGES)) ||
343 	    ((interface_type == ONYX_INTF) &&
344 		       (test >= MAX_ONYX_IMAGES)))
345 		return -EINVAL;
346 
347 	/* Copy the image into the processor */
348 	if (interface_type == CUDA_INTF)
349 		return perf_config(cuda_images[test]);
350 	else
351 		return perf_config(onyx_images[test]);
352 
353 	return count;
354 }
355 
356 /*
357  * Patch the images that need to know the IVA addresses.
358  */
359 static void perf_patch_images(void)
360 {
361 #if 0 /* FIXME!! */
362 /*
363  * NOTE:  this routine is VERY specific to the current TLB image.
364  * If the image is changed, this routine might also need to be changed.
365  */
366 	extern void $i_itlb_miss_2_0();
367 	extern void $i_dtlb_miss_2_0();
368 	extern void PA2_0_iva();
369 
370 	/*
371 	 * We can only use the lower 32-bits, the upper 32-bits should be 0
372 	 * anyway given this is in the kernel
373 	 */
374 	uint32_t itlb_addr  = (uint32_t)&($i_itlb_miss_2_0);
375 	uint32_t dtlb_addr  = (uint32_t)&($i_dtlb_miss_2_0);
376 	uint32_t IVAaddress = (uint32_t)&PA2_0_iva;
377 
378 	if (perf_processor_interface == ONYX_INTF) {
379 		/* clear last 2 bytes */
380 		onyx_images[TLBMISS][15] &= 0xffffff00;
381 		/* set 2 bytes */
382 		onyx_images[TLBMISS][15] |= (0x000000ff&((dtlb_addr) >> 24));
383 		onyx_images[TLBMISS][16] = (dtlb_addr << 8)&0xffffff00;
384 		onyx_images[TLBMISS][17] = itlb_addr;
385 
386 		/* clear last 2 bytes */
387 		onyx_images[TLBHANDMISS][15] &= 0xffffff00;
388 		/* set 2 bytes */
389 		onyx_images[TLBHANDMISS][15] |= (0x000000ff&((dtlb_addr) >> 24));
390 		onyx_images[TLBHANDMISS][16] = (dtlb_addr << 8)&0xffffff00;
391 		onyx_images[TLBHANDMISS][17] = itlb_addr;
392 
393 		/* clear last 2 bytes */
394 		onyx_images[BIG_CPI][15] &= 0xffffff00;
395 		/* set 2 bytes */
396 		onyx_images[BIG_CPI][15] |= (0x000000ff&((dtlb_addr) >> 24));
397 		onyx_images[BIG_CPI][16] = (dtlb_addr << 8)&0xffffff00;
398 		onyx_images[BIG_CPI][17] = itlb_addr;
399 
400 	    onyx_images[PANIC][15] &= 0xffffff00;  /* clear last 2 bytes */
401 	 	onyx_images[PANIC][15] |= (0x000000ff&((IVAaddress) >> 24)); /* set 2 bytes */
402 		onyx_images[PANIC][16] = (IVAaddress << 8)&0xffffff00;
403 
404 
405 	} else if (perf_processor_interface == CUDA_INTF) {
406 		/* Cuda interface */
407 		cuda_images[TLBMISS][16] =
408 			(cuda_images[TLBMISS][16]&0xffff0000) |
409 			((dtlb_addr >> 8)&0x0000ffff);
410 		cuda_images[TLBMISS][17] =
411 			((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
412 		cuda_images[TLBMISS][18] = (itlb_addr << 16)&0xffff0000;
413 
414 		cuda_images[TLBHANDMISS][16] =
415 			(cuda_images[TLBHANDMISS][16]&0xffff0000) |
416 			((dtlb_addr >> 8)&0x0000ffff);
417 		cuda_images[TLBHANDMISS][17] =
418 			((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
419 		cuda_images[TLBHANDMISS][18] = (itlb_addr << 16)&0xffff0000;
420 
421 		cuda_images[BIG_CPI][16] =
422 			(cuda_images[BIG_CPI][16]&0xffff0000) |
423 			((dtlb_addr >> 8)&0x0000ffff);
424 		cuda_images[BIG_CPI][17] =
425 			((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
426 		cuda_images[BIG_CPI][18] = (itlb_addr << 16)&0xffff0000;
427 	} else {
428 		/* Unknown type */
429 	}
430 #endif
431 }
432 
433 
434 /*
435  * ioctl routine
436  * All routines effect the processor that they are executed on.  Thus you
437  * must be running on the processor that you wish to change.
438  */
439 
440 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
441 {
442 	long error_start;
443 	uint32_t raddr[4];
444 	int error = 0;
445 
446 	switch (cmd) {
447 
448 	    case PA_PERF_ON:
449 			/* Start the counters */
450 			perf_start_counters();
451 			break;
452 
453 	    case PA_PERF_OFF:
454 			error_start = perf_stop_counters(raddr);
455 			if (error_start != 0) {
456 				printk(KERN_ERR "perf_off: perf_stop_counters = %ld\n", error_start);
457 				error = -EFAULT;
458 				break;
459 			}
460 
461 			/* copy out the Counters */
462 			if (copy_to_user((void __user *)arg, raddr,
463 					sizeof (raddr)) != 0) {
464 				error =  -EFAULT;
465 				break;
466 			}
467 			break;
468 
469 	    case PA_PERF_VERSION:
470   	  		/* Return the version # */
471 			error = put_user(PERF_VERSION, (int *)arg);
472 			break;
473 
474 	    default:
475   	 		error = -ENOTTY;
476 	}
477 
478 	return error;
479 }
480 
481 static const struct file_operations perf_fops = {
482 	.llseek = no_llseek,
483 	.read = perf_read,
484 	.write = perf_write,
485 	.unlocked_ioctl = perf_ioctl,
486 	.compat_ioctl = perf_ioctl,
487 	.open = perf_open,
488 	.release = perf_release
489 };
490 
491 static struct miscdevice perf_dev = {
492 	MISC_DYNAMIC_MINOR,
493 	PA_PERF_DEV,
494 	&perf_fops
495 };
496 
497 /*
498  * Initialize the module
499  */
500 static int __init perf_init(void)
501 {
502 	int ret;
503 
504 	/* Determine correct processor interface to use */
505 	bitmask_array = perf_bitmasks;
506 
507 	if (boot_cpu_data.cpu_type == pcxu ||
508 	    boot_cpu_data.cpu_type == pcxu_) {
509 		perf_processor_interface = ONYX_INTF;
510 	} else if (boot_cpu_data.cpu_type == pcxw ||
511 		 boot_cpu_data.cpu_type == pcxw_ ||
512 		 boot_cpu_data.cpu_type == pcxw2 ||
513 		 boot_cpu_data.cpu_type == mako ||
514 		 boot_cpu_data.cpu_type == mako2) {
515 		perf_processor_interface = CUDA_INTF;
516 		if (boot_cpu_data.cpu_type == pcxw2 ||
517 		    boot_cpu_data.cpu_type == mako ||
518 		    boot_cpu_data.cpu_type == mako2)
519 			bitmask_array = perf_bitmasks_piranha;
520 	} else {
521 		perf_processor_interface = UNKNOWN_INTF;
522 		printk("Performance monitoring counters not supported on this processor\n");
523 		return -ENODEV;
524 	}
525 
526 	ret = misc_register(&perf_dev);
527 	if (ret) {
528 		printk(KERN_ERR "Performance monitoring counters: "
529 			"cannot register misc device.\n");
530 		return ret;
531 	}
532 
533 	/* Patch the images to match the system */
534     	perf_patch_images();
535 
536 	spin_lock_init(&perf_lock);
537 
538 	/* TODO: this only lets us access the first cpu.. what to do for SMP? */
539 	cpu_device = per_cpu(cpu_data, 0).dev;
540 	printk("Performance monitoring counters enabled for %s\n",
541 		per_cpu(cpu_data, 0).dev->name);
542 
543 	return 0;
544 }
545 device_initcall(perf_init);
546 
547 /*
548  * perf_start_counters(void)
549  *
550  * Start the counters.
551  */
552 static void perf_start_counters(void)
553 {
554 	/* Enable performance monitor counters */
555 	perf_intrigue_enable_perf_counters();
556 }
557 
558 /*
559  * perf_stop_counters
560  *
561  * Stop the performance counters and save counts
562  * in a per_processor array.
563  */
564 static int perf_stop_counters(uint32_t *raddr)
565 {
566 	uint64_t userbuf[MAX_RDR_WORDS];
567 
568 	/* Disable performance counters */
569 	perf_intrigue_disable_perf_counters();
570 
571 	if (perf_processor_interface == ONYX_INTF) {
572 		uint64_t tmp64;
573 		/*
574 		 * Read the counters
575 		 */
576 		if (!perf_rdr_read_ubuf(16, userbuf))
577 			return -13;
578 
579 		/* Counter0 is bits 1398 to 1429 */
580 		tmp64 =  (userbuf[21] << 22) & 0x00000000ffc00000;
581 		tmp64 |= (userbuf[22] >> 42) & 0x00000000003fffff;
582 		/* OR sticky0 (bit 1430) to counter0 bit 32 */
583 		tmp64 |= (userbuf[22] >> 10) & 0x0000000080000000;
584 		raddr[0] = (uint32_t)tmp64;
585 
586 		/* Counter1 is bits 1431 to 1462 */
587 		tmp64 =  (userbuf[22] >> 9) & 0x00000000ffffffff;
588 		/* OR sticky1 (bit 1463) to counter1 bit 32 */
589 		tmp64 |= (userbuf[22] << 23) & 0x0000000080000000;
590 		raddr[1] = (uint32_t)tmp64;
591 
592 		/* Counter2 is bits 1464 to 1495 */
593 		tmp64 =  (userbuf[22] << 24) & 0x00000000ff000000;
594 		tmp64 |= (userbuf[23] >> 40) & 0x0000000000ffffff;
595 		/* OR sticky2 (bit 1496) to counter2 bit 32 */
596 		tmp64 |= (userbuf[23] >> 8) & 0x0000000080000000;
597 		raddr[2] = (uint32_t)tmp64;
598 
599 		/* Counter3 is bits 1497 to 1528 */
600 		tmp64 =  (userbuf[23] >> 7) & 0x00000000ffffffff;
601 		/* OR sticky3 (bit 1529) to counter3 bit 32 */
602 		tmp64 |= (userbuf[23] << 25) & 0x0000000080000000;
603 		raddr[3] = (uint32_t)tmp64;
604 
605 		/*
606 		 * Zero out the counters
607 		 */
608 
609 		/*
610 		 * The counters and sticky-bits comprise the last 132 bits
611 		 * (1398 - 1529) of RDR16 on a U chip.  We'll zero these
612 		 * out the easy way: zero out last 10 bits of dword 21,
613 		 * all of dword 22 and 58 bits (plus 6 don't care bits) of
614 		 * dword 23.
615 		 */
616 		userbuf[21] &= 0xfffffffffffffc00ul;	/* 0 to last 10 bits */
617 		userbuf[22] = 0;
618 		userbuf[23] = 0;
619 
620 		/*
621 		 * Write back the zeroed bytes + the image given
622 		 * the read was destructive.
623 		 */
624 		perf_rdr_write(16, userbuf);
625 	} else {
626 
627 		/*
628 		 * Read RDR-15 which contains the counters and sticky bits
629 		 */
630 		if (!perf_rdr_read_ubuf(15, userbuf)) {
631 			return -13;
632 		}
633 
634 		/*
635 		 * Clear out the counters
636 		 */
637 		perf_rdr_clear(15);
638 
639 		/*
640 		 * Copy the counters
641 		 */
642 		raddr[0] = (uint32_t)((userbuf[0] >> 32) & 0x00000000ffffffffUL);
643 		raddr[1] = (uint32_t)(userbuf[0] & 0x00000000ffffffffUL);
644 		raddr[2] = (uint32_t)((userbuf[1] >> 32) & 0x00000000ffffffffUL);
645 		raddr[3] = (uint32_t)(userbuf[1] & 0x00000000ffffffffUL);
646 	}
647 
648 	return 0;
649 }
650 
651 /*
652  * perf_rdr_get_entry
653  *
654  * Retrieve a pointer to the description of what this
655  * RDR contains.
656  */
657 static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num)
658 {
659 	if (perf_processor_interface == ONYX_INTF) {
660 		return &perf_rdr_tbl_U[rdr_num];
661 	} else {
662 		return &perf_rdr_tbl_W[rdr_num];
663 	}
664 }
665 
666 /*
667  * perf_rdr_read_ubuf
668  *
669  * Read the RDR value into the buffer specified.
670  */
671 static int perf_rdr_read_ubuf(uint32_t	rdr_num, uint64_t *buffer)
672 {
673 	uint64_t	data, data_mask = 0;
674 	uint32_t	width, xbits, i;
675 	const struct rdr_tbl_ent *tentry;
676 
677 	tentry = perf_rdr_get_entry(rdr_num);
678 	if ((width = tentry->width) == 0)
679 		return 0;
680 
681 	/* Clear out buffer */
682 	i = tentry->num_words;
683 	while (i--) {
684 		buffer[i] = 0;
685 	}
686 
687 	/* Check for bits an even number of 64 */
688 	if ((xbits = width & 0x03f) != 0) {
689 		data_mask = 1;
690 		data_mask <<= (64 - xbits);
691 		data_mask--;
692 	}
693 
694 	/* Grab all of the data */
695 	i = tentry->num_words;
696 	while (i--) {
697 
698 		if (perf_processor_interface == ONYX_INTF) {
699 			data = perf_rdr_shift_in_U(rdr_num, width);
700 		} else {
701 			data = perf_rdr_shift_in_W(rdr_num, width);
702 		}
703 		if (xbits) {
704 			buffer[i] |= (data << (64 - xbits));
705 			if (i) {
706 				buffer[i-1] |= ((data >> xbits) & data_mask);
707 			}
708 		} else {
709 			buffer[i] = data;
710 		}
711 	}
712 
713 	return 1;
714 }
715 
716 /*
717  * perf_rdr_clear
718  *
719  * Zero out the given RDR register
720  */
721 static int perf_rdr_clear(uint32_t	rdr_num)
722 {
723 	const struct rdr_tbl_ent *tentry;
724 	int32_t		i;
725 
726 	tentry = perf_rdr_get_entry(rdr_num);
727 
728 	if (tentry->width == 0) {
729 		return -1;
730 	}
731 
732 	i = tentry->num_words;
733 	while (i--) {
734 		if (perf_processor_interface == ONYX_INTF) {
735 			perf_rdr_shift_out_U(rdr_num, 0UL);
736 		} else {
737 			perf_rdr_shift_out_W(rdr_num, 0UL);
738 		}
739 	}
740 
741 	return 0;
742 }
743 
744 
745 /*
746  * perf_write_image
747  *
748  * Write the given image out to the processor
749  */
750 static int perf_write_image(uint64_t *memaddr)
751 {
752 	uint64_t buffer[MAX_RDR_WORDS];
753 	uint64_t *bptr;
754 	uint32_t dwords;
755 	const uint32_t *intrigue_rdr;
756 	const uint64_t *intrigue_bitmask;
757 	uint64_t tmp64;
758 	void __iomem *runway;
759 	const struct rdr_tbl_ent *tentry;
760 	int i;
761 
762 	/* Clear out counters */
763 	if (perf_processor_interface == ONYX_INTF) {
764 
765 		perf_rdr_clear(16);
766 
767 		/* Toggle performance monitor */
768 		perf_intrigue_enable_perf_counters();
769 		perf_intrigue_disable_perf_counters();
770 
771 		intrigue_rdr = perf_rdrs_U;
772 	} else {
773 		perf_rdr_clear(15);
774 		intrigue_rdr = perf_rdrs_W;
775 	}
776 
777 	/* Write all RDRs */
778 	while (*intrigue_rdr != -1) {
779 		tentry = perf_rdr_get_entry(*intrigue_rdr);
780 		perf_rdr_read_ubuf(*intrigue_rdr, buffer);
781 		bptr   = &buffer[0];
782 		dwords = tentry->num_words;
783 		if (tentry->write_control) {
784 			intrigue_bitmask = &bitmask_array[tentry->write_control >> 3];
785 			while (dwords--) {
786 				tmp64 = *intrigue_bitmask & *memaddr++;
787 				tmp64 |= (~(*intrigue_bitmask++)) & *bptr;
788 				*bptr++ = tmp64;
789 			}
790 		} else {
791 			while (dwords--) {
792 				*bptr++ = *memaddr++;
793 			}
794 		}
795 
796 		perf_rdr_write(*intrigue_rdr, buffer);
797 		intrigue_rdr++;
798 	}
799 
800 	/*
801 	 * Now copy out the Runway stuff which is not in RDRs
802 	 */
803 
804 	if (cpu_device == NULL)
805 	{
806 		printk(KERN_ERR "write_image: cpu_device not yet initialized!\n");
807 		return -1;
808 	}
809 
810 	runway = ioremap_nocache(cpu_device->hpa.start, 4096);
811 
812 	/* Merge intrigue bits into Runway STATUS 0 */
813 	tmp64 = __raw_readq(runway + RUNWAY_STATUS) & 0xffecfffffffffffful;
814 	__raw_writeq(tmp64 | (*memaddr++ & 0x0013000000000000ul),
815 		     runway + RUNWAY_STATUS);
816 
817 	/* Write RUNWAY DEBUG registers */
818 	for (i = 0; i < 8; i++) {
819 		__raw_writeq(*memaddr++, runway + RUNWAY_DEBUG);
820 	}
821 
822 	return 0;
823 }
824 
825 /*
826  * perf_rdr_write
827  *
828  * Write the given RDR register with the contents
829  * of the given buffer.
830  */
831 static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer)
832 {
833 	const struct rdr_tbl_ent *tentry;
834 	int32_t		i;
835 
836 printk("perf_rdr_write\n");
837 	tentry = perf_rdr_get_entry(rdr_num);
838 	if (tentry->width == 0) { return; }
839 
840 	i = tentry->num_words;
841 	while (i--) {
842 		if (perf_processor_interface == ONYX_INTF) {
843 			perf_rdr_shift_out_U(rdr_num, buffer[i]);
844 		} else {
845 			perf_rdr_shift_out_W(rdr_num, buffer[i]);
846 		}
847 	}
848 printk("perf_rdr_write done\n");
849 }
850