arch/arm/oprofile/op_model_xscale.c

   1 /**
   2  * @file op_model_xscale.c
   3  * XScale Performance Monitor Driver
   4  *
   5  * @remark Copyright 2000-2004 Deepak Saxena <dsaxena@mvista.com>
   6  * @remark Copyright 2000-2004 MontaVista Software Inc
   7  * @remark Copyright 2004 Dave Jiang <dave.jiang@intel.com>
   8  * @remark Copyright 2004 Intel Corporation
   9  * @remark Copyright 2004 Zwane Mwaikambo <zwane@arm.linux.org.uk>
  10  * @remark Copyright 2004 OProfile Authors
  11  *
  12  * @remark Read the file COPYING
  13  *
  14  * @author Zwane Mwaikambo
  15  */
  16
  17 /* #define DEBUG */
  18 #include <linux/types.h>
  19 #include <linux/errno.h>
  20 #include <linux/sched.h>
  21 #include <linux/oprofile.h>
  22 #include <linux/interrupt.h>
  23 #include <asm/irq.h>
  24 #include <asm/system.h>
  25
  26 #include "op_counter.h"
  27 #include "op_arm_model.h"
  28
  29 #define PMU_ENABLE      0x001   /* Enable counters */
  30 #define PMN_RESET       0x002   /* Reset event counters */
  31 #define CCNT_RESET      0x004   /* Reset clock counter */
  32 #define PMU_RESET       (CCNT_RESET | PMN_RESET)
  33 #define PMU_CNT64       0x008   /* Make CCNT count every 64th cycle */
  34
  35 /* TODO do runtime detection */
  36 #ifdef CONFIG_ARCH_IOP310
  37 #define XSCALE_PMU_IRQ  IRQ_XS80200_PMU
  38 #endif
  39 #ifdef CONFIG_ARCH_IOP321
  40 #define XSCALE_PMU_IRQ  IRQ_IOP321_CORE_PMU
  41 #endif
  42 #ifdef CONFIG_ARCH_IOP331
  43 #define XSCALE_PMU_IRQ  IRQ_IOP331_CORE_PMU
  44 #endif
  45 #ifdef CONFIG_ARCH_PXA
  46 #define XSCALE_PMU_IRQ  IRQ_PMU
  47 #endif
  48
  49 /*
  50  * Different types of events that can be counted by the XScale PMU
  51  * as used by Oprofile userspace. Here primarily for documentation
  52  * purposes.
  53  */
  54
  55 #define EVT_ICACHE_MISS                 0x00
  56 #define EVT_ICACHE_NO_DELIVER           0x01
  57 #define EVT_DATA_STALL                  0x02
  58 #define EVT_ITLB_MISS                   0x03
  59 #define EVT_DTLB_MISS                   0x04
  60 #define EVT_BRANCH                      0x05
  61 #define EVT_BRANCH_MISS                 0x06
  62 #define EVT_INSTRUCTION                 0x07
  63 #define EVT_DCACHE_FULL_STALL           0x08
  64 #define EVT_DCACHE_FULL_STALL_CONTIG    0x09
  65 #define EVT_DCACHE_ACCESS               0x0A
  66 #define EVT_DCACHE_MISS                 0x0B
  67 #define EVT_DCACE_WRITE_BACK            0x0C
  68 #define EVT_PC_CHANGED                  0x0D
  69 #define EVT_BCU_REQUEST                 0x10
  70 #define EVT_BCU_FULL                    0x11
  71 #define EVT_BCU_DRAIN                   0x12
  72 #define EVT_BCU_ECC_NO_ELOG             0x14
  73 #define EVT_BCU_1_BIT_ERR               0x15
  74 #define EVT_RMW                         0x16
  75 /* EVT_CCNT is not hardware defined */
  76 #define EVT_CCNT                        0xFE
  77 #define EVT_UNUSED                      0xFF
  78
  79 struct pmu_counter {
  80         volatile unsigned long ovf;
  81         unsigned long reset_counter;
  82 };
  83
  84 enum { CCNT, PMN0, PMN1, PMN2, PMN3, MAX_COUNTERS };
  85
  86 static struct pmu_counter results[MAX_COUNTERS];
  87
  88 /*
  89  * There are two versions of the PMU in current XScale processors
  90  * with differing register layouts and number of performance counters.
  91  * e.g. IOP321 is xsc1 whilst IOP331 is xsc2.
  92  * We detect which register layout to use in xscale_detect_pmu()
  93  */
  94 enum { PMU_XSC1, PMU_XSC2 };
  95
  96 struct pmu_type {
  97         int id;
  98         char *name;
  99         int num_counters;
 100         unsigned int int_enable;
 101         unsigned int cnt_ovf[MAX_COUNTERS];
 102         unsigned int int_mask[MAX_COUNTERS];
 103 };
 104
 105 static struct pmu_type pmu_parms[] = {
 106         {
 107                 .id             = PMU_XSC1,
 108                 .name           = "arm/xscale1",
 109                 .num_counters   = 3,
 110                 .int_mask       = { [PMN0] = 0x10, [PMN1] = 0x20,
 111                                     [CCNT] = 0x40 },
 112                 .cnt_ovf        = { [CCNT] = 0x400, [PMN0] = 0x100,
 113                                     [PMN1] = 0x200},
 114         },
 115         {
 116                 .id             = PMU_XSC2,
 117                 .name           = "arm/xscale2",
 118                 .num_counters   = 5,
 119                 .int_mask       = { [CCNT] = 0x01, [PMN0] = 0x02,
 120                                     [PMN1] = 0x04, [PMN2] = 0x08,
 121                                     [PMN3] = 0x10 },
 122                 .cnt_ovf        = { [CCNT] = 0x01, [PMN0] = 0x02,
 123                                     [PMN1] = 0x04, [PMN2] = 0x08,
 124                                     [PMN3] = 0x10 },
 125         },
 126 };
 127
 128 static struct pmu_type *pmu;
 129
 130 static void write_pmnc(u32 val)
 131 {
 132         if (pmu->id == PMU_XSC1) {
 133                 /* upper 4bits and 7, 11 are write-as-0 */
 134                 val &= 0xffff77f;
 135                 __asm__ __volatile__ ("mcr p14, 0, %0, c0, c0, 0" : : "r" (val));
 136         } else {
 137                 /* bits 4-23 are write-as-0, 24-31 are write ignored */
 138                 val &= 0xf;
 139                 __asm__ __volatile__ ("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
 140         }
 141 }
 142
 143 static u32 read_pmnc(void)
 144 {
 145         u32 val;
 146
 147         if (pmu->id == PMU_XSC1)
 148                 __asm__ __volatile__ ("mrc p14, 0, %0, c0, c0, 0" : "=r" (val));
 149         else {
 150                 __asm__ __volatile__ ("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
 151                 /* bits 1-2 and 4-23 are read-unpredictable */
 152                 val &= 0xff000009;
 153         }
 154
 155         return val;
 156 }
 157
 158 static u32 __xsc1_read_counter(int counter)
 159 {
 160         u32 val = 0;
 161
 162         switch (counter) {
 163         case CCNT:
 164                 __asm__ __volatile__ ("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
 165                 break;
 166         case PMN0:
 167                 __asm__ __volatile__ ("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
 168                 break;
 169         case PMN1:
 170                 __asm__ __volatile__ ("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
 171                 break;
 172         }
 173         return val;
 174 }
 175
 176 static u32 __xsc2_read_counter(int counter)
 177 {
 178         u32 val = 0;
 179
 180         switch (counter) {
 181         case CCNT:
 182                 __asm__ __volatile__ ("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
 183                 break;
 184         case PMN0:
 185                 __asm__ __volatile__ ("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
 186                 break;
 187         case PMN1:
 188                 __asm__ __volatile__ ("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
 189                 break;
 190         case PMN2:
 191                 __asm__ __volatile__ ("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
 192                 break;
 193         case PMN3:
 194                 __asm__ __volatile__ ("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
 195                 break;
 196         }
 197         return val;
 198 }
 199
 200 static u32 read_counter(int counter)
 201 {
 202         u32 val;
 203
 204         if (pmu->id == PMU_XSC1)
 205                 val = __xsc1_read_counter(counter);
 206         else
 207                 val = __xsc2_read_counter(counter);
 208
 209         return val;
 210 }
 211
 212 static void __xsc1_write_counter(int counter, u32 val)
 213 {
 214         switch (counter) {
 215         case CCNT:
 216                 __asm__ __volatile__ ("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
 217                 break;
 218         case PMN0:
 219                 __asm__ __volatile__ ("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
 220                 break;
 221         case PMN1:
 222                 __asm__ __volatile__ ("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
 223                 break;
 224         }
 225 }
 226
 227 static void __xsc2_write_counter(int counter, u32 val)
 228 {
 229         switch (counter) {
 230         case CCNT:
 231                 __asm__ __volatile__ ("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
 232                 break;
 233         case PMN0:
 234                 __asm__ __volatile__ ("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
 235                 break;
 236         case PMN1:
 237                 __asm__ __volatile__ ("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
 238                 break;
 239         case PMN2:
 240                 __asm__ __volatile__ ("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
 241                 break;
 242         case PMN3:
 243                 __asm__ __volatile__ ("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
 244                 break;
 245         }
 246 }
 247
 248 static void write_counter(int counter, u32 val)
 249 {
 250         if (pmu->id == PMU_XSC1)
 251                 __xsc1_write_counter(counter, val);
 252         else
 253                 __xsc2_write_counter(counter, val);
 254 }
 255
 256 static int xscale_setup_ctrs(void)
 257 {
 258         u32 evtsel, pmnc;
 259         int i;
 260
 261         for (i = CCNT; i < MAX_COUNTERS; i++) {
 262                 if (counter_config[i].enabled)
 263                         continue;
 264
 265                 counter_config[i].event = EVT_UNUSED;
 266         }
 267
 268         switch (pmu->id) {
 269         case PMU_XSC1:
 270                 pmnc = (counter_config[PMN1].event << 20) | (counter_config[PMN0].event << 12);
 271                 pr_debug("xscale_setup_ctrs: pmnc: %#08x\n", pmnc);
 272                 write_pmnc(pmnc);
 273                 break;
 274
 275         case PMU_XSC2:
 276                 evtsel = counter_config[PMN0].event | (counter_config[PMN1].event << 8) |
 277                         (counter_config[PMN2].event << 16) | (counter_config[PMN3].event << 24);
 278
 279                 pr_debug("xscale_setup_ctrs: evtsel %#08x\n", evtsel);
 280                 __asm__ __volatile__ ("mcr p14, 0, %0, c8, c1, 0" : : "r" (evtsel));
 281                 break;
 282         }
 283
 284         for (i = CCNT; i < MAX_COUNTERS; i++) {
 285                 if (counter_config[i].event == EVT_UNUSED) {
 286                         counter_config[i].event = 0;
 287                         pmu->int_enable &= ~pmu->int_mask[i];
 288                         continue;
 289                 }
 290
 291                 results[i].reset_counter = counter_config[i].count;
 292                 write_counter(i, -(u32)counter_config[i].count);
 293                 pmu->int_enable |= pmu->int_mask[i];
 294                 pr_debug("xscale_setup_ctrs: counter%d %#08x from %#08lx\n", i,
 295                         read_counter(i), counter_config[i].count);
 296         }
 297
 298         return 0;
 299 }
 300
 301 static void inline __xsc1_check_ctrs(void)
 302 {
 303         int i;
 304         u32 pmnc = read_pmnc();
 305
 306         /* NOTE: there's an A stepping errata that states if an overflow */
 307         /*       bit already exists and another occurs, the previous     */
 308         /*       Overflow bit gets cleared. There's no workaround.       */
 309         /*       Fixed in B stepping or later                            */
 310
 311         /* Write the value back to clear the overflow flags. Overflow */
 312         /* flags remain in pmnc for use below */
 313         write_pmnc(pmnc & ~PMU_ENABLE);
 314
 315         for (i = CCNT; i <= PMN1; i++) {
 316                 if (!(pmu->int_mask[i] & pmu->int_enable))
 317                         continue;
 318
 319                 if (pmnc & pmu->cnt_ovf[i])
 320                         results[i].ovf++;
 321         }
 322 }
 323
 324 static void inline __xsc2_check_ctrs(void)
 325 {
 326         int i;
 327         u32 flag = 0, pmnc = read_pmnc();
 328
 329         pmnc &= ~PMU_ENABLE;
 330         write_pmnc(pmnc);
 331
 332         /* read overflow flag register */
 333         __asm__ __volatile__ ("mrc p14, 0, %0, c5, c1, 0" : "=r" (flag));
 334
 335         for (i = CCNT; i <= PMN3; i++) {
 336                 if (!(pmu->int_mask[i] & pmu->int_enable))
 337                         continue;
 338
 339                 if (flag & pmu->cnt_ovf[i])
 340                         results[i].ovf++;
 341         }
 342
 343         /* writeback clears overflow bits */
 344         __asm__ __volatile__ ("mcr p14, 0, %0, c5, c1, 0" : : "r" (flag));
 345 }
 346
 347 static irqreturn_t xscale_pmu_interrupt(int irq, void *arg, struct pt_regs *regs)
 348 {
 349         int i;
 350         u32 pmnc;
 351
 352         if (pmu->id == PMU_XSC1)
 353                 __xsc1_check_ctrs();
 354         else
 355                 __xsc2_check_ctrs();
 356
 357         for (i = CCNT; i < MAX_COUNTERS; i++) {
 358                 if (!results[i].ovf)
 359                         continue;
 360
 361                 write_counter(i, -(u32)results[i].reset_counter);
 362                 oprofile_add_sample(regs, i);
 363                 results[i].ovf--;
 364         }
 365
 366         pmnc = read_pmnc() | PMU_ENABLE;
 367         write_pmnc(pmnc);
 368
 369         return IRQ_HANDLED;
 370 }
 371
 372 static void xscale_pmu_stop(void)
 373 {
 374         u32 pmnc = read_pmnc();
 375
 376         pmnc &= ~PMU_ENABLE;
 377         write_pmnc(pmnc);
 378
 379         free_irq(XSCALE_PMU_IRQ, results);
 380 }
 381
 382 static int xscale_pmu_start(void)
 383 {
 384         int ret;
 385         u32 pmnc = read_pmnc();
 386
 387         ret = request_irq(XSCALE_PMU_IRQ, xscale_pmu_interrupt, SA_INTERRUPT,
 388                         "XScale PMU", (void *)results);
 389
 390         if (ret < 0) {
 391                 printk(KERN_ERR "oprofile: unable to request IRQ%d for XScale PMU\n",
 392                         XSCALE_PMU_IRQ);
 393                 return ret;
 394         }
 395
 396         if (pmu->id == PMU_XSC1)
 397                 pmnc |= pmu->int_enable;
 398         else {
 399                 __asm__ __volatile__ ("mcr p14, 0, %0, c4, c1, 0" : : "r" (pmu->int_enable));
 400                 pmnc &= ~PMU_CNT64;
 401         }
 402
 403         pmnc |= PMU_ENABLE;
 404         write_pmnc(pmnc);
 405         pr_debug("xscale_pmu_start: pmnc: %#08x mask: %08x\n", pmnc, pmu->int_enable);
 406         return 0;
 407 }
 408
 409 static int xscale_detect_pmu(void)
 410 {
 411         int ret = 0;
 412         u32 id;
 413
 414         id = (read_cpuid(CPUID_ID) >> 13) & 0x7;
 415
 416         switch (id) {
 417         case 1:
 418                 pmu = &pmu_parms[PMU_XSC1];
 419                 break;
 420         case 2:
 421                 pmu = &pmu_parms[PMU_XSC2];
 422                 break;
 423         default:
 424                 ret = -ENODEV;
 425                 break;
 426         }
 427
 428         if (!ret) {
 429                 op_xscale_spec.name = pmu->name;
 430                 op_xscale_spec.num_counters = pmu->num_counters;
 431                 pr_debug("xscale_detect_pmu: detected %s PMU\n", pmu->name);
 432         }
 433
 434         return ret;
 435 }
 436
 437 struct op_arm_model_spec op_xscale_spec = {
 438         .init           = xscale_detect_pmu,
 439         .setup_ctrs     = xscale_setup_ctrs,
 440         .start          = xscale_pmu_start,
 441         .stop           = xscale_pmu_stop,
 442 };
 443