Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Parisc performance counters
4 * Copyright (C) 2001 Randolph Chung <tausq@debian.org>
5 *
6 * This code is derived, with permission, from HP/UX sources.
7 */
8
9/*
10 * Edited comment from original sources:
11 *
12 * This driver programs the PCX-U/PCX-W performance counters
13 * on the PA-RISC 2.0 chips. The driver keeps all images now
14 * internally to the kernel to hopefully eliminate the possibility
15 * of a bad image halting the CPU. Also, there are different
16 * images for the PCX-W and later chips vs the PCX-U chips.
17 *
18 * Only 1 process is allowed to access the driver at any time,
19 * so the only protection that is needed is at open and close.
20 * A variable "perf_enabled" is used to hold the state of the
21 * driver. The spinlock "perf_lock" is used to protect the
22 * modification of the state during open/close operations so
23 * multiple processes don't get into the driver simultaneously.
24 *
25 * This driver accesses the processor directly vs going through
26 * the PDC INTRIGUE calls. This is done to eliminate bugs introduced
27 * in various PDC revisions. The code is much more maintainable
28 * and reliable this way vs having to debug on every version of PDC
29 * on every box.
30 */
31
32#include <linux/capability.h>
33#include <linux/init.h>
34#include <linux/proc_fs.h>
35#include <linux/miscdevice.h>
36#include <linux/spinlock.h>
37
38#include <linux/uaccess.h>
39#include <asm/perf.h>
40#include <asm/parisc-device.h>
41#include <asm/processor.h>
42#include <asm/runway.h>
43#include <asm/io.h> /* for __raw_read() */
44
45#include "perf_images.h"
46
47#define MAX_RDR_WORDS 24
48#define PERF_VERSION 2 /* derived from hpux's PI v2 interface */
49
50/* definition of RDR regs */
51struct rdr_tbl_ent {
52 uint16_t width;
53 uint8_t num_words;
54 uint8_t write_control;
55};
56
57static int perf_processor_interface __read_mostly = UNKNOWN_INTF;
58static int perf_enabled __read_mostly;
59static DEFINE_SPINLOCK(perf_lock);
60struct parisc_device *cpu_device __read_mostly;
61
62/* RDRs to write for PCX-W */
63static const int perf_rdrs_W[] =
64 { 0, 1, 4, 5, 6, 15, 16, 17, 18, 20, 21, 22, 23, 24, 25, -1 };
65
66/* RDRs to write for PCX-U */
67static const int perf_rdrs_U[] =
68 { 0, 1, 4, 5, 6, 7, 16, 17, 18, 20, 21, 22, 23, 24, 25, -1 };
69
70/* RDR register descriptions for PCX-W */
71static const struct rdr_tbl_ent perf_rdr_tbl_W[] = {
72 { 19, 1, 8 }, /* RDR 0 */
73 { 16, 1, 16 }, /* RDR 1 */
74 { 72, 2, 0 }, /* RDR 2 */
75 { 81, 2, 0 }, /* RDR 3 */
76 { 328, 6, 0 }, /* RDR 4 */
77 { 160, 3, 0 }, /* RDR 5 */
78 { 336, 6, 0 }, /* RDR 6 */
79 { 164, 3, 0 }, /* RDR 7 */
80 { 0, 0, 0 }, /* RDR 8 */
81 { 35, 1, 0 }, /* RDR 9 */
82 { 6, 1, 0 }, /* RDR 10 */
83 { 18, 1, 0 }, /* RDR 11 */
84 { 13, 1, 0 }, /* RDR 12 */
85 { 8, 1, 0 }, /* RDR 13 */
86 { 8, 1, 0 }, /* RDR 14 */
87 { 8, 1, 0 }, /* RDR 15 */
88 { 1530, 24, 0 }, /* RDR 16 */
89 { 16, 1, 0 }, /* RDR 17 */
90 { 4, 1, 0 }, /* RDR 18 */
91 { 0, 0, 0 }, /* RDR 19 */
92 { 152, 3, 24 }, /* RDR 20 */
93 { 152, 3, 24 }, /* RDR 21 */
94 { 233, 4, 48 }, /* RDR 22 */
95 { 233, 4, 48 }, /* RDR 23 */
96 { 71, 2, 0 }, /* RDR 24 */
97 { 71, 2, 0 }, /* RDR 25 */
98 { 11, 1, 0 }, /* RDR 26 */
99 { 18, 1, 0 }, /* RDR 27 */
100 { 128, 2, 0 }, /* RDR 28 */
101 { 0, 0, 0 }, /* RDR 29 */
102 { 16, 1, 0 }, /* RDR 30 */
103 { 16, 1, 0 }, /* RDR 31 */
104};
105
106/* RDR register descriptions for PCX-U */
107static const struct rdr_tbl_ent perf_rdr_tbl_U[] = {
108 { 19, 1, 8 }, /* RDR 0 */
109 { 32, 1, 16 }, /* RDR 1 */
110 { 20, 1, 0 }, /* RDR 2 */
111 { 0, 0, 0 }, /* RDR 3 */
112 { 344, 6, 0 }, /* RDR 4 */
113 { 176, 3, 0 }, /* RDR 5 */
114 { 336, 6, 0 }, /* RDR 6 */
115 { 0, 0, 0 }, /* RDR 7 */
116 { 0, 0, 0 }, /* RDR 8 */
117 { 0, 0, 0 }, /* RDR 9 */
118 { 28, 1, 0 }, /* RDR 10 */
119 { 33, 1, 0 }, /* RDR 11 */
120 { 0, 0, 0 }, /* RDR 12 */
121 { 230, 4, 0 }, /* RDR 13 */
122 { 32, 1, 0 }, /* RDR 14 */
123 { 128, 2, 0 }, /* RDR 15 */
124 { 1494, 24, 0 }, /* RDR 16 */
125 { 18, 1, 0 }, /* RDR 17 */
126 { 4, 1, 0 }, /* RDR 18 */
127 { 0, 0, 0 }, /* RDR 19 */
128 { 158, 3, 24 }, /* RDR 20 */
129 { 158, 3, 24 }, /* RDR 21 */
130 { 194, 4, 48 }, /* RDR 22 */
131 { 194, 4, 48 }, /* RDR 23 */
132 { 71, 2, 0 }, /* RDR 24 */
133 { 71, 2, 0 }, /* RDR 25 */
134 { 28, 1, 0 }, /* RDR 26 */
135 { 33, 1, 0 }, /* RDR 27 */
136 { 88, 2, 0 }, /* RDR 28 */
137 { 32, 1, 0 }, /* RDR 29 */
138 { 24, 1, 0 }, /* RDR 30 */
139 { 16, 1, 0 }, /* RDR 31 */
140};
141
142/*
143 * A non-zero write_control in the above tables is a byte offset into
144 * this array.
145 */
146static const uint64_t perf_bitmasks[] = {
147 0x0000000000000000ul, /* first dbl word must be zero */
148 0xfdffe00000000000ul, /* RDR0 bitmask */
149 0x003f000000000000ul, /* RDR1 bitmask */
150 0x00fffffffffffffful, /* RDR20-RDR21 bitmask (152 bits) */
151 0xfffffffffffffffful,
152 0xfffffffc00000000ul,
153 0xfffffffffffffffful, /* RDR22-RDR23 bitmask (233 bits) */
154 0xfffffffffffffffful,
155 0xfffffffffffffffcul,
156 0xff00000000000000ul
157};
158
159/*
160 * Write control bitmasks for Pa-8700 processor given
161 * some things have changed slightly.
162 */
163static const uint64_t perf_bitmasks_piranha[] = {
164 0x0000000000000000ul, /* first dbl word must be zero */
165 0xfdffe00000000000ul, /* RDR0 bitmask */
166 0x003f000000000000ul, /* RDR1 bitmask */
167 0x00fffffffffffffful, /* RDR20-RDR21 bitmask (158 bits) */
168 0xfffffffffffffffful,
169 0xfffffffc00000000ul,
170 0xfffffffffffffffful, /* RDR22-RDR23 bitmask (210 bits) */
171 0xfffffffffffffffful,
172 0xfffffffffffffffful,
173 0xfffc000000000000ul
174};
175
176static const uint64_t *bitmask_array; /* array of bitmasks to use */
177
178/******************************************************************************
179 * Function Prototypes
180 *****************************************************************************/
181static int perf_config(uint32_t *image_ptr);
182static int perf_release(struct inode *inode, struct file *file);
183static int perf_open(struct inode *inode, struct file *file);
184static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t *ppos);
185static ssize_t perf_write(struct file *file, const char __user *buf,
186 size_t count, loff_t *ppos);
187static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
188static void perf_start_counters(void);
189static int perf_stop_counters(uint32_t *raddr);
190static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num);
191static int perf_rdr_read_ubuf(uint32_t rdr_num, uint64_t *buffer);
192static int perf_rdr_clear(uint32_t rdr_num);
193static int perf_write_image(uint64_t *memaddr);
194static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer);
195
196/* External Assembly Routines */
197extern uint64_t perf_rdr_shift_in_W (uint32_t rdr_num, uint16_t width);
198extern uint64_t perf_rdr_shift_in_U (uint32_t rdr_num, uint16_t width);
199extern void perf_rdr_shift_out_W (uint32_t rdr_num, uint64_t buffer);
200extern void perf_rdr_shift_out_U (uint32_t rdr_num, uint64_t buffer);
201extern void perf_intrigue_enable_perf_counters (void);
202extern void perf_intrigue_disable_perf_counters (void);
203
204/******************************************************************************
205 * Function Definitions
206 *****************************************************************************/
207
208
209/*
210 * configure:
211 *
212 * Configure the cpu with a given data image. First turn off the counters,
213 * then download the image, then turn the counters back on.
214 */
215static int perf_config(uint32_t *image_ptr)
216{
217 long error;
218 uint32_t raddr[4];
219
220 /* Stop the counters*/
221 error = perf_stop_counters(raddr);
222 if (error != 0) {
223 printk("perf_config: perf_stop_counters = %ld\n", error);
224 return -EINVAL;
225 }
226
227printk("Preparing to write image\n");
228 /* Write the image to the chip */
229 error = perf_write_image((uint64_t *)image_ptr);
230 if (error != 0) {
231 printk("perf_config: DOWNLOAD = %ld\n", error);
232 return -EINVAL;
233 }
234
235printk("Preparing to start counters\n");
236
237 /* Start the counters */
238 perf_start_counters();
239
240 return sizeof(uint32_t);
241}
242
243/*
244 * Open the device and initialize all of its memory. The device is only
245 * opened once, but can be "queried" by multiple processes that know its
246 * file descriptor.
247 */
248static int perf_open(struct inode *inode, struct file *file)
249{
250 spin_lock(&perf_lock);
251 if (perf_enabled) {
252 spin_unlock(&perf_lock);
253 return -EBUSY;
254 }
255 perf_enabled = 1;
256 spin_unlock(&perf_lock);
257
258 return 0;
259}
260
261/*
262 * Close the device.
263 */
264static int perf_release(struct inode *inode, struct file *file)
265{
266 spin_lock(&perf_lock);
267 perf_enabled = 0;
268 spin_unlock(&perf_lock);
269
270 return 0;
271}
272
273/*
274 * Read does nothing for this driver
275 */
276static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t *ppos)
277{
278 return 0;
279}
280
281/*
282 * write:
283 *
284 * This routine downloads the image to the chip. It must be
285 * called on the processor that the download should happen
286 * on.
287 */
288static ssize_t perf_write(struct file *file, const char __user *buf,
289 size_t count, loff_t *ppos)
290{
291 size_t image_size;
292 uint32_t image_type;
293 uint32_t interface_type;
294 uint32_t test;
295
296 if (perf_processor_interface == ONYX_INTF)
297 image_size = PCXU_IMAGE_SIZE;
298 else if (perf_processor_interface == CUDA_INTF)
299 image_size = PCXW_IMAGE_SIZE;
300 else
301 return -EFAULT;
302
303 if (!capable(CAP_SYS_ADMIN))
304 return -EACCES;
305
306 if (count != sizeof(uint32_t))
307 return -EIO;
308
309 if (copy_from_user(&image_type, buf, sizeof(uint32_t)))
310 return -EFAULT;
311
312 /* Get the interface type and test type */
313 interface_type = (image_type >> 16) & 0xffff;
314 test = (image_type & 0xffff);
315
316 /* Make sure everything makes sense */
317
318 /* First check the machine type is correct for
319 the requested image */
320 if (((perf_processor_interface == CUDA_INTF) &&
321 (interface_type != CUDA_INTF)) ||
322 ((perf_processor_interface == ONYX_INTF) &&
323 (interface_type != ONYX_INTF)))
324 return -EINVAL;
325
326 /* Next check to make sure the requested image
327 is valid */
328 if (((interface_type == CUDA_INTF) &&
329 (test >= MAX_CUDA_IMAGES)) ||
330 ((interface_type == ONYX_INTF) &&
331 (test >= MAX_ONYX_IMAGES)))
332 return -EINVAL;
333
334 /* Copy the image into the processor */
335 if (interface_type == CUDA_INTF)
336 return perf_config(cuda_images[test]);
337 else
338 return perf_config(onyx_images[test]);
339
340 return count;
341}
342
343/*
344 * Patch the images that need to know the IVA addresses.
345 */
346static void perf_patch_images(void)
347{
348#if 0 /* FIXME!! */
349/*
350 * NOTE: this routine is VERY specific to the current TLB image.
351 * If the image is changed, this routine might also need to be changed.
352 */
353 extern void $i_itlb_miss_2_0();
354 extern void $i_dtlb_miss_2_0();
355 extern void PA2_0_iva();
356
357 /*
358 * We can only use the lower 32-bits, the upper 32-bits should be 0
359 * anyway given this is in the kernel
360 */
361 uint32_t itlb_addr = (uint32_t)&($i_itlb_miss_2_0);
362 uint32_t dtlb_addr = (uint32_t)&($i_dtlb_miss_2_0);
363 uint32_t IVAaddress = (uint32_t)&PA2_0_iva;
364
365 if (perf_processor_interface == ONYX_INTF) {
366 /* clear last 2 bytes */
367 onyx_images[TLBMISS][15] &= 0xffffff00;
368 /* set 2 bytes */
369 onyx_images[TLBMISS][15] |= (0x000000ff&((dtlb_addr) >> 24));
370 onyx_images[TLBMISS][16] = (dtlb_addr << 8)&0xffffff00;
371 onyx_images[TLBMISS][17] = itlb_addr;
372
373 /* clear last 2 bytes */
374 onyx_images[TLBHANDMISS][15] &= 0xffffff00;
375 /* set 2 bytes */
376 onyx_images[TLBHANDMISS][15] |= (0x000000ff&((dtlb_addr) >> 24));
377 onyx_images[TLBHANDMISS][16] = (dtlb_addr << 8)&0xffffff00;
378 onyx_images[TLBHANDMISS][17] = itlb_addr;
379
380 /* clear last 2 bytes */
381 onyx_images[BIG_CPI][15] &= 0xffffff00;
382 /* set 2 bytes */
383 onyx_images[BIG_CPI][15] |= (0x000000ff&((dtlb_addr) >> 24));
384 onyx_images[BIG_CPI][16] = (dtlb_addr << 8)&0xffffff00;
385 onyx_images[BIG_CPI][17] = itlb_addr;
386
387 onyx_images[PANIC][15] &= 0xffffff00; /* clear last 2 bytes */
388 onyx_images[PANIC][15] |= (0x000000ff&((IVAaddress) >> 24)); /* set 2 bytes */
389 onyx_images[PANIC][16] = (IVAaddress << 8)&0xffffff00;
390
391
392 } else if (perf_processor_interface == CUDA_INTF) {
393 /* Cuda interface */
394 cuda_images[TLBMISS][16] =
395 (cuda_images[TLBMISS][16]&0xffff0000) |
396 ((dtlb_addr >> 8)&0x0000ffff);
397 cuda_images[TLBMISS][17] =
398 ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
399 cuda_images[TLBMISS][18] = (itlb_addr << 16)&0xffff0000;
400
401 cuda_images[TLBHANDMISS][16] =
402 (cuda_images[TLBHANDMISS][16]&0xffff0000) |
403 ((dtlb_addr >> 8)&0x0000ffff);
404 cuda_images[TLBHANDMISS][17] =
405 ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
406 cuda_images[TLBHANDMISS][18] = (itlb_addr << 16)&0xffff0000;
407
408 cuda_images[BIG_CPI][16] =
409 (cuda_images[BIG_CPI][16]&0xffff0000) |
410 ((dtlb_addr >> 8)&0x0000ffff);
411 cuda_images[BIG_CPI][17] =
412 ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
413 cuda_images[BIG_CPI][18] = (itlb_addr << 16)&0xffff0000;
414 } else {
415 /* Unknown type */
416 }
417#endif
418}
419
420
421/*
422 * ioctl routine
423 * All routines effect the processor that they are executed on. Thus you
424 * must be running on the processor that you wish to change.
425 */
426
427static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
428{
429 long error_start;
430 uint32_t raddr[4];
431 int error = 0;
432
433 switch (cmd) {
434
435 case PA_PERF_ON:
436 /* Start the counters */
437 perf_start_counters();
438 break;
439
440 case PA_PERF_OFF:
441 error_start = perf_stop_counters(raddr);
442 if (error_start != 0) {
443 printk(KERN_ERR "perf_off: perf_stop_counters = %ld\n", error_start);
444 error = -EFAULT;
445 break;
446 }
447
448 /* copy out the Counters */
449 if (copy_to_user((void __user *)arg, raddr,
450 sizeof (raddr)) != 0) {
451 error = -EFAULT;
452 break;
453 }
454 break;
455
456 case PA_PERF_VERSION:
457 /* Return the version # */
458 error = put_user(PERF_VERSION, (int *)arg);
459 break;
460
461 default:
462 error = -ENOTTY;
463 }
464
465 return error;
466}
467
468static const struct file_operations perf_fops = {
469 .llseek = no_llseek,
470 .read = perf_read,
471 .write = perf_write,
472 .unlocked_ioctl = perf_ioctl,
473 .compat_ioctl = perf_ioctl,
474 .open = perf_open,
475 .release = perf_release
476};
477
478static struct miscdevice perf_dev = {
479 MISC_DYNAMIC_MINOR,
480 PA_PERF_DEV,
481 &perf_fops
482};
483
484/*
485 * Initialize the module
486 */
487static int __init perf_init(void)
488{
489 int ret;
490
491 /* Determine correct processor interface to use */
492 bitmask_array = perf_bitmasks;
493
494 if (boot_cpu_data.cpu_type == pcxu ||
495 boot_cpu_data.cpu_type == pcxu_) {
496 perf_processor_interface = ONYX_INTF;
497 } else if (boot_cpu_data.cpu_type == pcxw ||
498 boot_cpu_data.cpu_type == pcxw_ ||
499 boot_cpu_data.cpu_type == pcxw2 ||
500 boot_cpu_data.cpu_type == mako ||
501 boot_cpu_data.cpu_type == mako2) {
502 perf_processor_interface = CUDA_INTF;
503 if (boot_cpu_data.cpu_type == pcxw2 ||
504 boot_cpu_data.cpu_type == mako ||
505 boot_cpu_data.cpu_type == mako2)
506 bitmask_array = perf_bitmasks_piranha;
507 } else {
508 perf_processor_interface = UNKNOWN_INTF;
509 printk("Performance monitoring counters not supported on this processor\n");
510 return -ENODEV;
511 }
512
513 ret = misc_register(&perf_dev);
514 if (ret) {
515 printk(KERN_ERR "Performance monitoring counters: "
516 "cannot register misc device.\n");
517 return ret;
518 }
519
520 /* Patch the images to match the system */
521 perf_patch_images();
522
523 /* TODO: this only lets us access the first cpu.. what to do for SMP? */
524 cpu_device = per_cpu(cpu_data, 0).dev;
525 printk("Performance monitoring counters enabled for %s\n",
526 per_cpu(cpu_data, 0).dev->name);
527
528 return 0;
529}
530device_initcall(perf_init);
531
532/*
533 * perf_start_counters(void)
534 *
535 * Start the counters.
536 */
537static void perf_start_counters(void)
538{
539 /* Enable performance monitor counters */
540 perf_intrigue_enable_perf_counters();
541}
542
543/*
544 * perf_stop_counters
545 *
546 * Stop the performance counters and save counts
547 * in a per_processor array.
548 */
549static int perf_stop_counters(uint32_t *raddr)
550{
551 uint64_t userbuf[MAX_RDR_WORDS];
552
553 /* Disable performance counters */
554 perf_intrigue_disable_perf_counters();
555
556 if (perf_processor_interface == ONYX_INTF) {
557 uint64_t tmp64;
558 /*
559 * Read the counters
560 */
561 if (!perf_rdr_read_ubuf(16, userbuf))
562 return -13;
563
564 /* Counter0 is bits 1398 to 1429 */
565 tmp64 = (userbuf[21] << 22) & 0x00000000ffc00000;
566 tmp64 |= (userbuf[22] >> 42) & 0x00000000003fffff;
567 /* OR sticky0 (bit 1430) to counter0 bit 32 */
568 tmp64 |= (userbuf[22] >> 10) & 0x0000000080000000;
569 raddr[0] = (uint32_t)tmp64;
570
571 /* Counter1 is bits 1431 to 1462 */
572 tmp64 = (userbuf[22] >> 9) & 0x00000000ffffffff;
573 /* OR sticky1 (bit 1463) to counter1 bit 32 */
574 tmp64 |= (userbuf[22] << 23) & 0x0000000080000000;
575 raddr[1] = (uint32_t)tmp64;
576
577 /* Counter2 is bits 1464 to 1495 */
578 tmp64 = (userbuf[22] << 24) & 0x00000000ff000000;
579 tmp64 |= (userbuf[23] >> 40) & 0x0000000000ffffff;
580 /* OR sticky2 (bit 1496) to counter2 bit 32 */
581 tmp64 |= (userbuf[23] >> 8) & 0x0000000080000000;
582 raddr[2] = (uint32_t)tmp64;
583
584 /* Counter3 is bits 1497 to 1528 */
585 tmp64 = (userbuf[23] >> 7) & 0x00000000ffffffff;
586 /* OR sticky3 (bit 1529) to counter3 bit 32 */
587 tmp64 |= (userbuf[23] << 25) & 0x0000000080000000;
588 raddr[3] = (uint32_t)tmp64;
589
590 /*
591 * Zero out the counters
592 */
593
594 /*
595 * The counters and sticky-bits comprise the last 132 bits
596 * (1398 - 1529) of RDR16 on a U chip. We'll zero these
597 * out the easy way: zero out last 10 bits of dword 21,
598 * all of dword 22 and 58 bits (plus 6 don't care bits) of
599 * dword 23.
600 */
601 userbuf[21] &= 0xfffffffffffffc00ul; /* 0 to last 10 bits */
602 userbuf[22] = 0;
603 userbuf[23] = 0;
604
605 /*
606 * Write back the zeroed bytes + the image given
607 * the read was destructive.
608 */
609 perf_rdr_write(16, userbuf);
610 } else {
611
612 /*
613 * Read RDR-15 which contains the counters and sticky bits
614 */
615 if (!perf_rdr_read_ubuf(15, userbuf)) {
616 return -13;
617 }
618
619 /*
620 * Clear out the counters
621 */
622 perf_rdr_clear(15);
623
624 /*
625 * Copy the counters
626 */
627 raddr[0] = (uint32_t)((userbuf[0] >> 32) & 0x00000000ffffffffUL);
628 raddr[1] = (uint32_t)(userbuf[0] & 0x00000000ffffffffUL);
629 raddr[2] = (uint32_t)((userbuf[1] >> 32) & 0x00000000ffffffffUL);
630 raddr[3] = (uint32_t)(userbuf[1] & 0x00000000ffffffffUL);
631 }
632
633 return 0;
634}
635
636/*
637 * perf_rdr_get_entry
638 *
639 * Retrieve a pointer to the description of what this
640 * RDR contains.
641 */
642static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num)
643{
644 if (perf_processor_interface == ONYX_INTF) {
645 return &perf_rdr_tbl_U[rdr_num];
646 } else {
647 return &perf_rdr_tbl_W[rdr_num];
648 }
649}
650
651/*
652 * perf_rdr_read_ubuf
653 *
654 * Read the RDR value into the buffer specified.
655 */
656static int perf_rdr_read_ubuf(uint32_t rdr_num, uint64_t *buffer)
657{
658 uint64_t data, data_mask = 0;
659 uint32_t width, xbits, i;
660 const struct rdr_tbl_ent *tentry;
661
662 tentry = perf_rdr_get_entry(rdr_num);
663 if ((width = tentry->width) == 0)
664 return 0;
665
666 /* Clear out buffer */
667 i = tentry->num_words;
668 while (i--) {
669 buffer[i] = 0;
670 }
671
672 /* Check for bits an even number of 64 */
673 if ((xbits = width & 0x03f) != 0) {
674 data_mask = 1;
675 data_mask <<= (64 - xbits);
676 data_mask--;
677 }
678
679 /* Grab all of the data */
680 i = tentry->num_words;
681 while (i--) {
682
683 if (perf_processor_interface == ONYX_INTF) {
684 data = perf_rdr_shift_in_U(rdr_num, width);
685 } else {
686 data = perf_rdr_shift_in_W(rdr_num, width);
687 }
688 if (xbits) {
689 buffer[i] |= (data << (64 - xbits));
690 if (i) {
691 buffer[i-1] |= ((data >> xbits) & data_mask);
692 }
693 } else {
694 buffer[i] = data;
695 }
696 }
697
698 return 1;
699}
700
701/*
702 * perf_rdr_clear
703 *
704 * Zero out the given RDR register
705 */
706static int perf_rdr_clear(uint32_t rdr_num)
707{
708 const struct rdr_tbl_ent *tentry;
709 int32_t i;
710
711 tentry = perf_rdr_get_entry(rdr_num);
712
713 if (tentry->width == 0) {
714 return -1;
715 }
716
717 i = tentry->num_words;
718 while (i--) {
719 if (perf_processor_interface == ONYX_INTF) {
720 perf_rdr_shift_out_U(rdr_num, 0UL);
721 } else {
722 perf_rdr_shift_out_W(rdr_num, 0UL);
723 }
724 }
725
726 return 0;
727}
728
729
730/*
731 * perf_write_image
732 *
733 * Write the given image out to the processor
734 */
735static int perf_write_image(uint64_t *memaddr)
736{
737 uint64_t buffer[MAX_RDR_WORDS];
738 uint64_t *bptr;
739 uint32_t dwords;
740 const uint32_t *intrigue_rdr;
741 const uint64_t *intrigue_bitmask;
742 uint64_t tmp64;
743 void __iomem *runway;
744 const struct rdr_tbl_ent *tentry;
745 int i;
746
747 /* Clear out counters */
748 if (perf_processor_interface == ONYX_INTF) {
749
750 perf_rdr_clear(16);
751
752 /* Toggle performance monitor */
753 perf_intrigue_enable_perf_counters();
754 perf_intrigue_disable_perf_counters();
755
756 intrigue_rdr = perf_rdrs_U;
757 } else {
758 perf_rdr_clear(15);
759 intrigue_rdr = perf_rdrs_W;
760 }
761
762 /* Write all RDRs */
763 while (*intrigue_rdr != -1) {
764 tentry = perf_rdr_get_entry(*intrigue_rdr);
765 perf_rdr_read_ubuf(*intrigue_rdr, buffer);
766 bptr = &buffer[0];
767 dwords = tentry->num_words;
768 if (tentry->write_control) {
769 intrigue_bitmask = &bitmask_array[tentry->write_control >> 3];
770 while (dwords--) {
771 tmp64 = *intrigue_bitmask & *memaddr++;
772 tmp64 |= (~(*intrigue_bitmask++)) & *bptr;
773 *bptr++ = tmp64;
774 }
775 } else {
776 while (dwords--) {
777 *bptr++ = *memaddr++;
778 }
779 }
780
781 perf_rdr_write(*intrigue_rdr, buffer);
782 intrigue_rdr++;
783 }
784
785 /*
786 * Now copy out the Runway stuff which is not in RDRs
787 */
788
789 if (cpu_device == NULL)
790 {
791 printk(KERN_ERR "write_image: cpu_device not yet initialized!\n");
792 return -1;
793 }
794
795 runway = ioremap_nocache(cpu_device->hpa.start, 4096);
796 if (!runway) {
797 pr_err("perf_write_image: ioremap failed!\n");
798 return -ENOMEM;
799 }
800
801 /* Merge intrigue bits into Runway STATUS 0 */
802 tmp64 = __raw_readq(runway + RUNWAY_STATUS) & 0xffecfffffffffffful;
803 __raw_writeq(tmp64 | (*memaddr++ & 0x0013000000000000ul),
804 runway + RUNWAY_STATUS);
805
806 /* Write RUNWAY DEBUG registers */
807 for (i = 0; i < 8; i++) {
808 __raw_writeq(*memaddr++, runway + RUNWAY_DEBUG);
809 }
810
811 return 0;
812}
813
814/*
815 * perf_rdr_write
816 *
817 * Write the given RDR register with the contents
818 * of the given buffer.
819 */
820static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer)
821{
822 const struct rdr_tbl_ent *tentry;
823 int32_t i;
824
825printk("perf_rdr_write\n");
826 tentry = perf_rdr_get_entry(rdr_num);
827 if (tentry->width == 0) { return; }
828
829 i = tentry->num_words;
830 while (i--) {
831 if (perf_processor_interface == ONYX_INTF) {
832 perf_rdr_shift_out_U(rdr_num, buffer[i]);
833 } else {
834 perf_rdr_shift_out_W(rdr_num, buffer[i]);
835 }
836 }
837printk("perf_rdr_write done\n");
838}
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 <asm/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 */
64struct rdr_tbl_ent {
65 uint16_t width;
66 uint8_t num_words;
67 uint8_t write_control;
68};
69
70static int perf_processor_interface __read_mostly = UNKNOWN_INTF;
71static int perf_enabled __read_mostly;
72static spinlock_t perf_lock;
73struct parisc_device *cpu_device __read_mostly;
74
75/* RDRs to write for PCX-W */
76static 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 */
80static 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 */
84static 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 */
120static 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 */
159static 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 */
176static 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
189static const uint64_t *bitmask_array; /* array of bitmasks to use */
190
191/******************************************************************************
192 * Function Prototypes
193 *****************************************************************************/
194static int perf_config(uint32_t *image_ptr);
195static int perf_release(struct inode *inode, struct file *file);
196static int perf_open(struct inode *inode, struct file *file);
197static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t *ppos);
198static ssize_t perf_write(struct file *file, const char __user *buf, size_t count,
199 loff_t *ppos);
200static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
201static void perf_start_counters(void);
202static int perf_stop_counters(uint32_t *raddr);
203static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num);
204static int perf_rdr_read_ubuf(uint32_t rdr_num, uint64_t *buffer);
205static int perf_rdr_clear(uint32_t rdr_num);
206static int perf_write_image(uint64_t *memaddr);
207static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer);
208
209/* External Assembly Routines */
210extern uint64_t perf_rdr_shift_in_W (uint32_t rdr_num, uint16_t width);
211extern uint64_t perf_rdr_shift_in_U (uint32_t rdr_num, uint16_t width);
212extern void perf_rdr_shift_out_W (uint32_t rdr_num, uint64_t buffer);
213extern void perf_rdr_shift_out_U (uint32_t rdr_num, uint64_t buffer);
214extern void perf_intrigue_enable_perf_counters (void);
215extern 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 */
228static 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
240printk("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
248printk("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 */
261static 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 */
277static 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 */
289static 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 */
301static ssize_t perf_write(struct file *file, const char __user *buf, size_t count,
302 loff_t *ppos)
303{
304 int err;
305 size_t image_size;
306 uint32_t image_type;
307 uint32_t interface_type;
308 uint32_t test;
309
310 if (perf_processor_interface == ONYX_INTF)
311 image_size = PCXU_IMAGE_SIZE;
312 else if (perf_processor_interface == CUDA_INTF)
313 image_size = PCXW_IMAGE_SIZE;
314 else
315 return -EFAULT;
316
317 if (!capable(CAP_SYS_ADMIN))
318 return -EACCES;
319
320 if (count != sizeof(uint32_t))
321 return -EIO;
322
323 if ((err = copy_from_user(&image_type, buf, sizeof(uint32_t))) != 0)
324 return err;
325
326 /* Get the interface type and test type */
327 interface_type = (image_type >> 16) & 0xffff;
328 test = (image_type & 0xffff);
329
330 /* Make sure everything makes sense */
331
332 /* First check the machine type is correct for
333 the requested image */
334 if (((perf_processor_interface == CUDA_INTF) &&
335 (interface_type != CUDA_INTF)) ||
336 ((perf_processor_interface == ONYX_INTF) &&
337 (interface_type != ONYX_INTF)))
338 return -EINVAL;
339
340 /* Next check to make sure the requested image
341 is valid */
342 if (((interface_type == CUDA_INTF) &&
343 (test >= MAX_CUDA_IMAGES)) ||
344 ((interface_type == ONYX_INTF) &&
345 (test >= MAX_ONYX_IMAGES)))
346 return -EINVAL;
347
348 /* Copy the image into the processor */
349 if (interface_type == CUDA_INTF)
350 return perf_config(cuda_images[test]);
351 else
352 return perf_config(onyx_images[test]);
353
354 return count;
355}
356
357/*
358 * Patch the images that need to know the IVA addresses.
359 */
360static void perf_patch_images(void)
361{
362#if 0 /* FIXME!! */
363/*
364 * NOTE: this routine is VERY specific to the current TLB image.
365 * If the image is changed, this routine might also need to be changed.
366 */
367 extern void $i_itlb_miss_2_0();
368 extern void $i_dtlb_miss_2_0();
369 extern void PA2_0_iva();
370
371 /*
372 * We can only use the lower 32-bits, the upper 32-bits should be 0
373 * anyway given this is in the kernel
374 */
375 uint32_t itlb_addr = (uint32_t)&($i_itlb_miss_2_0);
376 uint32_t dtlb_addr = (uint32_t)&($i_dtlb_miss_2_0);
377 uint32_t IVAaddress = (uint32_t)&PA2_0_iva;
378
379 if (perf_processor_interface == ONYX_INTF) {
380 /* clear last 2 bytes */
381 onyx_images[TLBMISS][15] &= 0xffffff00;
382 /* set 2 bytes */
383 onyx_images[TLBMISS][15] |= (0x000000ff&((dtlb_addr) >> 24));
384 onyx_images[TLBMISS][16] = (dtlb_addr << 8)&0xffffff00;
385 onyx_images[TLBMISS][17] = itlb_addr;
386
387 /* clear last 2 bytes */
388 onyx_images[TLBHANDMISS][15] &= 0xffffff00;
389 /* set 2 bytes */
390 onyx_images[TLBHANDMISS][15] |= (0x000000ff&((dtlb_addr) >> 24));
391 onyx_images[TLBHANDMISS][16] = (dtlb_addr << 8)&0xffffff00;
392 onyx_images[TLBHANDMISS][17] = itlb_addr;
393
394 /* clear last 2 bytes */
395 onyx_images[BIG_CPI][15] &= 0xffffff00;
396 /* set 2 bytes */
397 onyx_images[BIG_CPI][15] |= (0x000000ff&((dtlb_addr) >> 24));
398 onyx_images[BIG_CPI][16] = (dtlb_addr << 8)&0xffffff00;
399 onyx_images[BIG_CPI][17] = itlb_addr;
400
401 onyx_images[PANIC][15] &= 0xffffff00; /* clear last 2 bytes */
402 onyx_images[PANIC][15] |= (0x000000ff&((IVAaddress) >> 24)); /* set 2 bytes */
403 onyx_images[PANIC][16] = (IVAaddress << 8)&0xffffff00;
404
405
406 } else if (perf_processor_interface == CUDA_INTF) {
407 /* Cuda interface */
408 cuda_images[TLBMISS][16] =
409 (cuda_images[TLBMISS][16]&0xffff0000) |
410 ((dtlb_addr >> 8)&0x0000ffff);
411 cuda_images[TLBMISS][17] =
412 ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
413 cuda_images[TLBMISS][18] = (itlb_addr << 16)&0xffff0000;
414
415 cuda_images[TLBHANDMISS][16] =
416 (cuda_images[TLBHANDMISS][16]&0xffff0000) |
417 ((dtlb_addr >> 8)&0x0000ffff);
418 cuda_images[TLBHANDMISS][17] =
419 ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
420 cuda_images[TLBHANDMISS][18] = (itlb_addr << 16)&0xffff0000;
421
422 cuda_images[BIG_CPI][16] =
423 (cuda_images[BIG_CPI][16]&0xffff0000) |
424 ((dtlb_addr >> 8)&0x0000ffff);
425 cuda_images[BIG_CPI][17] =
426 ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff);
427 cuda_images[BIG_CPI][18] = (itlb_addr << 16)&0xffff0000;
428 } else {
429 /* Unknown type */
430 }
431#endif
432}
433
434
435/*
436 * ioctl routine
437 * All routines effect the processor that they are executed on. Thus you
438 * must be running on the processor that you wish to change.
439 */
440
441static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
442{
443 long error_start;
444 uint32_t raddr[4];
445 int error = 0;
446
447 switch (cmd) {
448
449 case PA_PERF_ON:
450 /* Start the counters */
451 perf_start_counters();
452 break;
453
454 case PA_PERF_OFF:
455 error_start = perf_stop_counters(raddr);
456 if (error_start != 0) {
457 printk(KERN_ERR "perf_off: perf_stop_counters = %ld\n", error_start);
458 error = -EFAULT;
459 break;
460 }
461
462 /* copy out the Counters */
463 if (copy_to_user((void __user *)arg, raddr,
464 sizeof (raddr)) != 0) {
465 error = -EFAULT;
466 break;
467 }
468 break;
469
470 case PA_PERF_VERSION:
471 /* Return the version # */
472 error = put_user(PERF_VERSION, (int *)arg);
473 break;
474
475 default:
476 error = -ENOTTY;
477 }
478
479 return error;
480}
481
482static const struct file_operations perf_fops = {
483 .llseek = no_llseek,
484 .read = perf_read,
485 .write = perf_write,
486 .unlocked_ioctl = perf_ioctl,
487 .compat_ioctl = perf_ioctl,
488 .open = perf_open,
489 .release = perf_release
490};
491
492static struct miscdevice perf_dev = {
493 MISC_DYNAMIC_MINOR,
494 PA_PERF_DEV,
495 &perf_fops
496};
497
498/*
499 * Initialize the module
500 */
501static int __init perf_init(void)
502{
503 int ret;
504
505 /* Determine correct processor interface to use */
506 bitmask_array = perf_bitmasks;
507
508 if (boot_cpu_data.cpu_type == pcxu ||
509 boot_cpu_data.cpu_type == pcxu_) {
510 perf_processor_interface = ONYX_INTF;
511 } else if (boot_cpu_data.cpu_type == pcxw ||
512 boot_cpu_data.cpu_type == pcxw_ ||
513 boot_cpu_data.cpu_type == pcxw2 ||
514 boot_cpu_data.cpu_type == mako ||
515 boot_cpu_data.cpu_type == mako2) {
516 perf_processor_interface = CUDA_INTF;
517 if (boot_cpu_data.cpu_type == pcxw2 ||
518 boot_cpu_data.cpu_type == mako ||
519 boot_cpu_data.cpu_type == mako2)
520 bitmask_array = perf_bitmasks_piranha;
521 } else {
522 perf_processor_interface = UNKNOWN_INTF;
523 printk("Performance monitoring counters not supported on this processor\n");
524 return -ENODEV;
525 }
526
527 ret = misc_register(&perf_dev);
528 if (ret) {
529 printk(KERN_ERR "Performance monitoring counters: "
530 "cannot register misc device.\n");
531 return ret;
532 }
533
534 /* Patch the images to match the system */
535 perf_patch_images();
536
537 spin_lock_init(&perf_lock);
538
539 /* TODO: this only lets us access the first cpu.. what to do for SMP? */
540 cpu_device = per_cpu(cpu_data, 0).dev;
541 printk("Performance monitoring counters enabled for %s\n",
542 per_cpu(cpu_data, 0).dev->name);
543
544 return 0;
545}
546
547/*
548 * perf_start_counters(void)
549 *
550 * Start the counters.
551 */
552static 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 */
564static 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 */
657static 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 */
671static 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 */
721static 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 */
750static 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 */
831static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer)
832{
833 const struct rdr_tbl_ent *tentry;
834 int32_t i;
835
836printk("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 }
848printk("perf_rdr_write done\n");
849}
850
851module_init(perf_init);