projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
parisc: Enhance CPU detection code on PAT machines
[linux-2.6-block.git] / arch / parisc / kernel / perf.c
CommitLineData
1da177e4
LT		 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
06fe9fb4 27 * internally to the kernel to hopefully eliminate the possibility
1da177e4
LT		 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
a9415644 45#include <linux/capability.h>
1da177e4
LT		 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
8039de10 70static int perf_processor_interface __read_mostly = UNKNOWN_INTF;
cb6fc18e 71static int perf_enabled __read_mostly;
1da177e4 72static spinlock_t perf_lock;
cb6fc18e 73struct parisc_device *cpu_device __read_mostly;
1da177e4
LT		 74
75/* RDRs to write for PCX-W */
cb6fc18e 76static const int perf_rdrs_W[] =
1da177e4
LT		 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 */
cb6fc18e 80static const int perf_rdrs_U[] =
1da177e4
LT		 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 */
cb6fc18e 84static const struct rdr_tbl_ent perf_rdr_tbl_W[] = {
1da177e4
LT		 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 */
cb6fc18e 120static const struct rdr_tbl_ent perf_rdr_tbl_U[] = {
1da177e4
LT		 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 */
cb6fc18e 159static const uint64_t perf_bitmasks[] = {
1da177e4
LT		 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
7022672e 174 * some things have changed slightly.
1da177e4 175 */
cb6fc18e 176static const uint64_t perf_bitmasks_piranha[] = {
1da177e4
LT		 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
cb6fc18e 189static const uint64_t *bitmask_array; /* array of bitmasks to use */
1da177e4
LT		 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);
ad7dd338 200static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
1da177e4
LT		 201static void perf_start_counters(void);
202static int perf_stop_counters(uint32_t *raddr);
cb6fc18e 203static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num);
1da177e4
LT		 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
ad7dd338 441static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1da177e4
LT		 442{
443 long error_start;
ad7dd338
CH		 444 uint32_t raddr[4];
445 int error = 0;
1da177e4
LT		 446
447 switch (cmd) {
448
449 case PA_PERF_ON:
450 /* Start the counters */
451 perf_start_counters();
ad7dd338 452 break;
1da177e4
LT		 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);
ad7dd338
CH		 458 error = -EFAULT;
459 break;
1da177e4
LT		 460 }
461
462 /* copy out the Counters */
463 if (copy_to_user((void __user *)arg, raddr,
464 sizeof (raddr)) != 0) {
ad7dd338
CH		 465 error = -EFAULT;
466 break;
1da177e4 467 }
ad7dd338 468 break;
1da177e4
LT		 469
470 case PA_PERF_VERSION:
471 /* Return the version # */
ad7dd338
CH		 472 error = put_user(PERF_VERSION, (int *)arg);
473 break;
1da177e4
LT		 474
475 default:
ad7dd338 476 error = -ENOTTY;
1da177e4 477 }
ad7dd338 478
ad7dd338 479 return error;
1da177e4
LT		 480}
481
5dfe4c96 482static const struct file_operations perf_fops = {
1da177e4
LT		 483 .llseek = no_llseek,
484 .read = perf_read,
485 .write = perf_write,
ad7dd338
CH		 486 .unlocked_ioctl = perf_ioctl,
487 .compat_ioctl = perf_ioctl,
1da177e4
LT		 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 ||
2cbd42db
KM		 514 boot_cpu_data.cpu_type == mako ||
515 boot_cpu_data.cpu_type == mako2) {
1da177e4
LT		 516 perf_processor_interface = CUDA_INTF;
517 if (boot_cpu_data.cpu_type == pcxw2 ||
2cbd42db
KM		 518 boot_cpu_data.cpu_type == mako ||
519 boot_cpu_data.cpu_type == mako2)
1da177e4
LT		 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? */
ef017beb 540 cpu_device = per_cpu(cpu_data, 0).dev;
1da177e4 541 printk("Performance monitoring counters enabled for %s\n",
ef017beb 542 per_cpu(cpu_data, 0).dev->name);
1da177e4
LT		 543
544 return 0;
545}
15becabd 546device_initcall(perf_init);
1da177e4
LT		 547
548/*
549 * perf_start_counters(void)
550 *
551 * Start the counters.
552 */
553static void perf_start_counters(void)
554{
555 /* Enable performance monitor counters */
556 perf_intrigue_enable_perf_counters();
557}
558
559/*
560 * perf_stop_counters
561 *
562 * Stop the performance counters and save counts
563 * in a per_processor array.
564 */
565static int perf_stop_counters(uint32_t *raddr)
566{
567 uint64_t userbuf[MAX_RDR_WORDS];
568
569 /* Disable performance counters */
570 perf_intrigue_disable_perf_counters();
571
572 if (perf_processor_interface == ONYX_INTF) {
573 uint64_t tmp64;
574 /*
575 * Read the counters
576 */
577 if (!perf_rdr_read_ubuf(16, userbuf))
578 return -13;
579
7022672e 580 /* Counter0 is bits 1398 to 1429 */
1da177e4
LT		 581 tmp64 = (userbuf[21] << 22) & 0x00000000ffc00000;
582 tmp64 |= (userbuf[22] >> 42) & 0x00000000003fffff;
583 /* OR sticky0 (bit 1430) to counter0 bit 32 */
584 tmp64 |= (userbuf[22] >> 10) & 0x0000000080000000;
585 raddr[0] = (uint32_t)tmp64;
586
7022672e 587 /* Counter1 is bits 1431 to 1462 */
1da177e4
LT		 588 tmp64 = (userbuf[22] >> 9) & 0x00000000ffffffff;
589 /* OR sticky1 (bit 1463) to counter1 bit 32 */
590 tmp64 |= (userbuf[22] << 23) & 0x0000000080000000;
591 raddr[1] = (uint32_t)tmp64;
592
7022672e 593 /* Counter2 is bits 1464 to 1495 */
1da177e4
LT		 594 tmp64 = (userbuf[22] << 24) & 0x00000000ff000000;
595 tmp64 |= (userbuf[23] >> 40) & 0x0000000000ffffff;
596 /* OR sticky2 (bit 1496) to counter2 bit 32 */
597 tmp64 |= (userbuf[23] >> 8) & 0x0000000080000000;
598 raddr[2] = (uint32_t)tmp64;
599
7022672e 600 /* Counter3 is bits 1497 to 1528 */
1da177e4
LT		 601 tmp64 = (userbuf[23] >> 7) & 0x00000000ffffffff;
602 /* OR sticky3 (bit 1529) to counter3 bit 32 */
603 tmp64 |= (userbuf[23] << 25) & 0x0000000080000000;
604 raddr[3] = (uint32_t)tmp64;
605
606 /*
607 * Zero out the counters
608 */
609
610 /*
611 * The counters and sticky-bits comprise the last 132 bits
612 * (1398 - 1529) of RDR16 on a U chip. We'll zero these
613 * out the easy way: zero out last 10 bits of dword 21,
614 * all of dword 22 and 58 bits (plus 6 don't care bits) of
615 * dword 23.
616 */
617 userbuf[21] &= 0xfffffffffffffc00ul; /* 0 to last 10 bits */
618 userbuf[22] = 0;
619 userbuf[23] = 0;
620
621 /*
7022672e 622 * Write back the zeroed bytes + the image given
1da177e4
LT		 623 * the read was destructive.
624 */
625 perf_rdr_write(16, userbuf);
626 } else {
627
628 /*
629 * Read RDR-15 which contains the counters and sticky bits
630 */
631 if (!perf_rdr_read_ubuf(15, userbuf)) {
632 return -13;
633 }
634
635 /*
636 * Clear out the counters
637 */
638 perf_rdr_clear(15);
639
640 /*
641 * Copy the counters
642 */
643 raddr[0] = (uint32_t)((userbuf[0] >> 32) & 0x00000000ffffffffUL);
644 raddr[1] = (uint32_t)(userbuf[0] & 0x00000000ffffffffUL);
645 raddr[2] = (uint32_t)((userbuf[1] >> 32) & 0x00000000ffffffffUL);
646 raddr[3] = (uint32_t)(userbuf[1] & 0x00000000ffffffffUL);
647 }
648
649 return 0;
650}
651
652/*
653 * perf_rdr_get_entry
654 *
655 * Retrieve a pointer to the description of what this
656 * RDR contains.
657 */
cb6fc18e 658static const struct rdr_tbl_ent * perf_rdr_get_entry(uint32_t rdr_num)
1da177e4
LT		 659{
660 if (perf_processor_interface == ONYX_INTF) {
661 return &perf_rdr_tbl_U[rdr_num];
662 } else {
663 return &perf_rdr_tbl_W[rdr_num];
664 }
665}
666
667/*
668 * perf_rdr_read_ubuf
669 *
670 * Read the RDR value into the buffer specified.
671 */
672static int perf_rdr_read_ubuf(uint32_t rdr_num, uint64_t *buffer)
673{
674 uint64_t data, data_mask = 0;
675 uint32_t width, xbits, i;
cb6fc18e 676 const struct rdr_tbl_ent *tentry;
1da177e4
LT		 677
678 tentry = perf_rdr_get_entry(rdr_num);
679 if ((width = tentry->width) == 0)
680 return 0;
681
682 /* Clear out buffer */
683 i = tentry->num_words;
684 while (i--) {
685 buffer[i] = 0;
686 }
687
688 /* Check for bits an even number of 64 */
689 if ((xbits = width & 0x03f) != 0) {
690 data_mask = 1;
691 data_mask <<= (64 - xbits);
692 data_mask--;
693 }
694
695 /* Grab all of the data */
696 i = tentry->num_words;
697 while (i--) {
698
699 if (perf_processor_interface == ONYX_INTF) {
700 data = perf_rdr_shift_in_U(rdr_num, width);
701 } else {
702 data = perf_rdr_shift_in_W(rdr_num, width);
703 }
704 if (xbits) {
705 buffer[i] |= (data << (64 - xbits));
706 if (i) {
707 buffer[i-1] |= ((data >> xbits) & data_mask);
708 }
709 } else {
710 buffer[i] = data;
711 }
712 }
713
714 return 1;
715}
716
717/*
718 * perf_rdr_clear
719 *
720 * Zero out the given RDR register
721 */
722static int perf_rdr_clear(uint32_t rdr_num)
723{
cb6fc18e 724 const struct rdr_tbl_ent *tentry;
1da177e4
LT		 725 int32_t i;
726
727 tentry = perf_rdr_get_entry(rdr_num);
728
729 if (tentry->width == 0) {
730 return -1;
731 }
732
733 i = tentry->num_words;
734 while (i--) {
735 if (perf_processor_interface == ONYX_INTF) {
736 perf_rdr_shift_out_U(rdr_num, 0UL);
737 } else {
738 perf_rdr_shift_out_W(rdr_num, 0UL);
739 }
740 }
741
742 return 0;
743}
744
745
746/*
747 * perf_write_image
748 *
749 * Write the given image out to the processor
750 */
751static int perf_write_image(uint64_t *memaddr)
752{
753 uint64_t buffer[MAX_RDR_WORDS];
754 uint64_t *bptr;
755 uint32_t dwords;
cb6fc18e
HD		 756 const uint32_t *intrigue_rdr;
757 const uint64_t *intrigue_bitmask;
758 uint64_t tmp64;
53f01bba 759 void __iomem *runway;
cb6fc18e 760 const struct rdr_tbl_ent *tentry;
1da177e4
LT		 761 int i;
762
763 /* Clear out counters */
764 if (perf_processor_interface == ONYX_INTF) {
765
766 perf_rdr_clear(16);
767
768 /* Toggle performance monitor */
769 perf_intrigue_enable_perf_counters();
770 perf_intrigue_disable_perf_counters();
771
772 intrigue_rdr = perf_rdrs_U;
773 } else {
774 perf_rdr_clear(15);
775 intrigue_rdr = perf_rdrs_W;
776 }
777
778 /* Write all RDRs */
779 while (*intrigue_rdr != -1) {
780 tentry = perf_rdr_get_entry(*intrigue_rdr);
781 perf_rdr_read_ubuf(*intrigue_rdr, buffer);
782 bptr = &buffer[0];
783 dwords = tentry->num_words;
784 if (tentry->write_control) {
785 intrigue_bitmask = &bitmask_array[tentry->write_control >> 3];
786 while (dwords--) {
787 tmp64 = *intrigue_bitmask & *memaddr++;
788 tmp64 |= (~(*intrigue_bitmask++)) & *bptr;
789 *bptr++ = tmp64;
790 }
791 } else {
792 while (dwords--) {
793 *bptr++ = *memaddr++;
794 }
795 }
796
797 perf_rdr_write(*intrigue_rdr, buffer);
798 intrigue_rdr++;
799 }
800
801 /*
802 * Now copy out the Runway stuff which is not in RDRs
803 */
804
805 if (cpu_device == NULL)
806 {
807 printk(KERN_ERR "write_image: cpu_device not yet initialized!\n");
808 return -1;
809 }
810
5076c158 811 runway = ioremap_nocache(cpu_device->hpa.start, 4096);
1da177e4
LT		 812
813 /* Merge intrigue bits into Runway STATUS 0 */
53f01bba
MW		 814 tmp64 = __raw_readq(runway + RUNWAY_STATUS) & 0xffecfffffffffffful;
815 __raw_writeq(tmp64 | (*memaddr++ & 0x0013000000000000ul),
816 runway + RUNWAY_STATUS);
1da177e4
LT		 817
818 /* Write RUNWAY DEBUG registers */
819 for (i = 0; i < 8; i++) {
53f01bba 820 __raw_writeq(*memaddr++, runway + RUNWAY_DEBUG);
1da177e4
LT		 821 }
822
823 return 0;
824}
825
826/*
827 * perf_rdr_write
828 *
829 * Write the given RDR register with the contents
830 * of the given buffer.
831 */
832static void perf_rdr_write(uint32_t rdr_num, uint64_t *buffer)
833{
cb6fc18e 834 const struct rdr_tbl_ent *tentry;
1da177e4
LT		 835 int32_t i;
836
837printk("perf_rdr_write\n");
838 tentry = perf_rdr_get_entry(rdr_num);
839 if (tentry->width == 0) { return; }
840
841 i = tentry->num_words;
842 while (i--) {
843 if (perf_processor_interface == ONYX_INTF) {
844 perf_rdr_shift_out_U(rdr_num, buffer[i]);
845 } else {
846 perf_rdr_shift_out_W(rdr_num, buffer[i]);
847 }
848 }
849printk("perf_rdr_write done\n");
850}
Linux 6.x block layer and io_uring tree(s)