arch/sparc/kernel/perf_event.c

   1 /* Performance event support for sparc64.
   2  *
   3  * Copyright (C) 2009, 2010 David S. Miller <davem@davemloft.net>
   4  *
   5  * This code is based almost entirely upon the x86 perf event
   6  * code, which is:
   7  *
   8  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   9  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
  10  *  Copyright (C) 2009 Jaswinder Singh Rajput
  11  *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
  12  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  13  */
  14
  15 #include <linux/perf_event.h>
  16 #include <linux/kprobes.h>
  17 #include <linux/ftrace.h>
  18 #include <linux/kernel.h>
  19 #include <linux/kdebug.h>
  20 #include <linux/mutex.h>
  21
  22 #include <asm/stacktrace.h>
  23 #include <asm/cpudata.h>
  24 #include <asm/uaccess.h>
  25 #include <linux/atomic.h>
  26 #include <asm/nmi.h>
  27 #include <asm/pcr.h>
  28 #include <asm/cacheflush.h>
  29
  30 #include "kernel.h"
  31 #include "kstack.h"
  32
  33 /* Two classes of sparc64 chips currently exist.  All of which have
  34  * 32-bit counters which can generate overflow interrupts on the
  35  * transition from 0xffffffff to 0.
  36  *
  37  * All chips upto and including SPARC-T3 have two performance
  38  * counters.  The two 32-bit counters are accessed in one go using a
  39  * single 64-bit register.
  40  *
  41  * On these older chips both counters are controlled using a single
  42  * control register.  The only way to stop all sampling is to clear
  43  * all of the context (user, supervisor, hypervisor) sampling enable
  44  * bits.  But these bits apply to both counters, thus the two counters
  45  * can't be enabled/disabled individually.
  46  *
  47  * Furthermore, the control register on these older chips have two
  48  * event fields, one for each of the two counters.  It's thus nearly
  49  * impossible to have one counter going while keeping the other one
  50  * stopped.  Therefore it is possible to get overflow interrupts for
  51  * counters not currently "in use" and that condition must be checked
  52  * in the overflow interrupt handler.
  53  *
  54  * So we use a hack, in that we program inactive counters with the
  55  * "sw_count0" and "sw_count1" events.  These count how many times
  56  * the instruction "sethi %hi(0xfc000), %g0" is executed.  It's an
  57  * unusual way to encode a NOP and therefore will not trigger in
  58  * normal code.
  59  *
  60  * Starting with SPARC-T4 we have one control register per counter.
  61  * And the counters are stored in individual registers.  The registers
  62  * for the counters are 64-bit but only a 32-bit counter is
  63  * implemented.  The event selections on SPARC-T4 lack any
  64  * restrictions, therefore we can elide all of the complicated
  65  * conflict resolution code we have for SPARC-T3 and earlier chips.
  66  */
  67
  68 #define MAX_HWEVENTS                    4
  69 #define MAX_PCRS                        4
  70 #define MAX_PERIOD                      ((1UL << 32) - 1)
  71
  72 #define PIC_UPPER_INDEX                 0
  73 #define PIC_LOWER_INDEX                 1
  74 #define PIC_NO_INDEX                    -1
  75
  76 struct cpu_hw_events {
  77         /* Number of events currently scheduled onto this cpu.
  78          * This tells how many entries in the arrays below
  79          * are valid.
  80          */
  81         int                     n_events;
  82
  83         /* Number of new events added since the last hw_perf_disable().
  84          * This works because the perf event layer always adds new
  85          * events inside of a perf_{disable,enable}() sequence.
  86          */
  87         int                     n_added;
  88
  89         /* Array of events current scheduled on this cpu.  */
  90         struct perf_event       *event[MAX_HWEVENTS];
  91
  92         /* Array of encoded longs, specifying the %pcr register
  93          * encoding and the mask of PIC counters this even can
  94          * be scheduled on.  See perf_event_encode() et al.
  95          */
  96         unsigned long           events[MAX_HWEVENTS];
  97
  98         /* The current counter index assigned to an event.  When the
  99          * event hasn't been programmed into the cpu yet, this will
 100          * hold PIC_NO_INDEX.  The event->hw.idx value tells us where
 101          * we ought to schedule the event.
 102          */
 103         int                     current_idx[MAX_HWEVENTS];
 104
 105         /* Software copy of %pcr register(s) on this cpu.  */
 106         u64                     pcr[MAX_HWEVENTS];
 107
 108         /* Enabled/disable state.  */
 109         int                     enabled;
 110
 111         unsigned int            group_flag;
 112 };
 113 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
 114
 115 /* An event map describes the characteristics of a performance
 116  * counter event.  In particular it gives the encoding as well as
 117  * a mask telling which counters the event can be measured on.
 118  *
 119  * The mask is unused on SPARC-T4 and later.
 120  */
 121 struct perf_event_map {
 122         u16     encoding;
 123         u8      pic_mask;
 124 #define PIC_NONE        0x00
 125 #define PIC_UPPER       0x01
 126 #define PIC_LOWER       0x02
 127 };
 128
 129 /* Encode a perf_event_map entry into a long.  */
 130 static unsigned long perf_event_encode(const struct perf_event_map *pmap)
 131 {
 132         return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
 133 }
 134
 135 static u8 perf_event_get_msk(unsigned long val)
 136 {
 137         return val & 0xff;
 138 }
 139
 140 static u64 perf_event_get_enc(unsigned long val)
 141 {
 142         return val >> 16;
 143 }
 144
 145 #define C(x) PERF_COUNT_HW_CACHE_##x
 146
 147 #define CACHE_OP_UNSUPPORTED    0xfffe
 148 #define CACHE_OP_NONSENSE       0xffff
 149
 150 typedef struct perf_event_map cache_map_t
 151                                 [PERF_COUNT_HW_CACHE_MAX]
 152                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 153                                 [PERF_COUNT_HW_CACHE_RESULT_MAX];
 154
 155 struct sparc_pmu {
 156         const struct perf_event_map     *(*event_map)(int);
 157         const cache_map_t               *cache_map;
 158         int                             max_events;
 159         u32                             (*read_pmc)(int);
 160         void                            (*write_pmc)(int, u64);
 161         int                             upper_shift;
 162         int                             lower_shift;
 163         int                             event_mask;
 164         int                             user_bit;
 165         int                             priv_bit;
 166         int                             hv_bit;
 167         int                             irq_bit;
 168         int                             upper_nop;
 169         int                             lower_nop;
 170         unsigned int                    flags;
 171 #define SPARC_PMU_ALL_EXCLUDES_SAME     0x00000001
 172 #define SPARC_PMU_HAS_CONFLICTS         0x00000002
 173         int                             max_hw_events;
 174         int                             num_pcrs;
 175         int                             num_pic_regs;
 176 };
 177
 178 static u32 sparc_default_read_pmc(int idx)
 179 {
 180         u64 val;
 181
 182         val = pcr_ops->read_pic(0);
 183         if (idx == PIC_UPPER_INDEX)
 184                 val >>= 32;
 185
 186         return val & 0xffffffff;
 187 }
 188
 189 static void sparc_default_write_pmc(int idx, u64 val)
 190 {
 191         u64 shift, mask, pic;
 192
 193         shift = 0;
 194         if (idx == PIC_UPPER_INDEX)
 195                 shift = 32;
 196
 197         mask = ((u64) 0xffffffff) << shift;
 198         val <<= shift;
 199
 200         pic = pcr_ops->read_pic(0);
 201         pic &= ~mask;
 202         pic |= val;
 203         pcr_ops->write_pic(0, pic);
 204 }
 205
 206 static const struct perf_event_map ultra3_perfmon_event_map[] = {
 207         [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
 208         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
 209         [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
 210         [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
 211 };
 212
 213 static const struct perf_event_map *ultra3_event_map(int event_id)
 214 {
 215         return &ultra3_perfmon_event_map[event_id];
 216 }
 217
 218 static const cache_map_t ultra3_cache_map = {
 219 [C(L1D)] = {
 220         [C(OP_READ)] = {
 221                 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
 222                 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
 223         },
 224         [C(OP_WRITE)] = {
 225                 [C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
 226                 [C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
 227         },
 228         [C(OP_PREFETCH)] = {
 229                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 230                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 231         },
 232 },
 233 [C(L1I)] = {
 234         [C(OP_READ)] = {
 235                 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
 236                 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
 237         },
 238         [ C(OP_WRITE) ] = {
 239                 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
 240                 [ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
 241         },
 242         [ C(OP_PREFETCH) ] = {
 243                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 244                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 245         },
 246 },
 247 [C(LL)] = {
 248         [C(OP_READ)] = {
 249                 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
 250                 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
 251         },
 252         [C(OP_WRITE)] = {
 253                 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
 254                 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
 255         },
 256         [C(OP_PREFETCH)] = {
 257                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 258                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 259         },
 260 },
 261 [C(DTLB)] = {
 262         [C(OP_READ)] = {
 263                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 264                 [C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
 265         },
 266         [ C(OP_WRITE) ] = {
 267                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 268                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 269         },
 270         [ C(OP_PREFETCH) ] = {
 271                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 272                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 273         },
 274 },
 275 [C(ITLB)] = {
 276         [C(OP_READ)] = {
 277                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 278                 [C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
 279         },
 280         [ C(OP_WRITE) ] = {
 281                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 282                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 283         },
 284         [ C(OP_PREFETCH) ] = {
 285                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 286                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 287         },
 288 },
 289 [C(BPU)] = {
 290         [C(OP_READ)] = {
 291                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 292                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 293         },
 294         [ C(OP_WRITE) ] = {
 295                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 296                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 297         },
 298         [ C(OP_PREFETCH) ] = {
 299                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 300                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 301         },
 302 },
 303 [C(NODE)] = {
 304         [C(OP_READ)] = {
 305                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 306                 [C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
 307         },
 308         [ C(OP_WRITE) ] = {
 309                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 310                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 311         },
 312         [ C(OP_PREFETCH) ] = {
 313                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 314                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 315         },
 316 },
 317 };
 318
 319 static const struct sparc_pmu ultra3_pmu = {
 320         .event_map      = ultra3_event_map,
 321         .cache_map      = &ultra3_cache_map,
 322         .max_events     = ARRAY_SIZE(ultra3_perfmon_event_map),
 323         .read_pmc       = sparc_default_read_pmc,
 324         .write_pmc      = sparc_default_write_pmc,
 325         .upper_shift    = 11,
 326         .lower_shift    = 4,
 327         .event_mask     = 0x3f,
 328         .user_bit       = PCR_UTRACE,
 329         .priv_bit       = PCR_STRACE,
 330         .upper_nop      = 0x1c,
 331         .lower_nop      = 0x14,
 332         .flags          = (SPARC_PMU_ALL_EXCLUDES_SAME |
 333                            SPARC_PMU_HAS_CONFLICTS),
 334         .max_hw_events  = 2,
 335         .num_pcrs       = 1,
 336         .num_pic_regs   = 1,
 337 };
 338
 339 /* Niagara1 is very limited.  The upper PIC is hard-locked to count
 340  * only instructions, so it is free running which creates all kinds of
 341  * problems.  Some hardware designs make one wonder if the creator
 342  * even looked at how this stuff gets used by software.
 343  */
 344 static const struct perf_event_map niagara1_perfmon_event_map[] = {
 345         [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
 346         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
 347         [PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
 348         [PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
 349 };
 350
 351 static const struct perf_event_map *niagara1_event_map(int event_id)
 352 {
 353         return &niagara1_perfmon_event_map[event_id];
 354 }
 355
 356 static const cache_map_t niagara1_cache_map = {
 357 [C(L1D)] = {
 358         [C(OP_READ)] = {
 359                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 360                 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
 361         },
 362         [C(OP_WRITE)] = {
 363                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 364                 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
 365         },
 366         [C(OP_PREFETCH)] = {
 367                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 368                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 369         },
 370 },
 371 [C(L1I)] = {
 372         [C(OP_READ)] = {
 373                 [C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
 374                 [C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
 375         },
 376         [ C(OP_WRITE) ] = {
 377                 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
 378                 [ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
 379         },
 380         [ C(OP_PREFETCH) ] = {
 381                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 382                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 383         },
 384 },
 385 [C(LL)] = {
 386         [C(OP_READ)] = {
 387                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 388                 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
 389         },
 390         [C(OP_WRITE)] = {
 391                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 392                 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
 393         },
 394         [C(OP_PREFETCH)] = {
 395                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 396                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 397         },
 398 },
 399 [C(DTLB)] = {
 400         [C(OP_READ)] = {
 401                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 402                 [C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
 403         },
 404         [ C(OP_WRITE) ] = {
 405                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 406                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 407         },
 408         [ C(OP_PREFETCH) ] = {
 409                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 410                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 411         },
 412 },
 413 [C(ITLB)] = {
 414         [C(OP_READ)] = {
 415                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 416                 [C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
 417         },
 418         [ C(OP_WRITE) ] = {
 419                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 420                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 421         },
 422         [ C(OP_PREFETCH) ] = {
 423                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 424                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 425         },
 426 },
 427 [C(BPU)] = {
 428         [C(OP_READ)] = {
 429                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 430                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 431         },
 432         [ C(OP_WRITE) ] = {
 433                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 434                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 435         },
 436         [ C(OP_PREFETCH) ] = {
 437                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 438                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 439         },
 440 },
 441 [C(NODE)] = {
 442         [C(OP_READ)] = {
 443                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 444                 [C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
 445         },
 446         [ C(OP_WRITE) ] = {
 447                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 448                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 449         },
 450         [ C(OP_PREFETCH) ] = {
 451                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 452                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 453         },
 454 },
 455 };
 456
 457 static const struct sparc_pmu niagara1_pmu = {
 458         .event_map      = niagara1_event_map,
 459         .cache_map      = &niagara1_cache_map,
 460         .max_events     = ARRAY_SIZE(niagara1_perfmon_event_map),
 461         .read_pmc       = sparc_default_read_pmc,
 462         .write_pmc      = sparc_default_write_pmc,
 463         .upper_shift    = 0,
 464         .lower_shift    = 4,
 465         .event_mask     = 0x7,
 466         .user_bit       = PCR_UTRACE,
 467         .priv_bit       = PCR_STRACE,
 468         .upper_nop      = 0x0,
 469         .lower_nop      = 0x0,
 470         .flags          = (SPARC_PMU_ALL_EXCLUDES_SAME |
 471                            SPARC_PMU_HAS_CONFLICTS),
 472         .max_hw_events  = 2,
 473         .num_pcrs       = 1,
 474         .num_pic_regs   = 1,
 475 };
 476
 477 static const struct perf_event_map niagara2_perfmon_event_map[] = {
 478         [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
 479         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
 480         [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
 481         [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
 482         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
 483         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
 484 };
 485
 486 static const struct perf_event_map *niagara2_event_map(int event_id)
 487 {
 488         return &niagara2_perfmon_event_map[event_id];
 489 }
 490
 491 static const cache_map_t niagara2_cache_map = {
 492 [C(L1D)] = {
 493         [C(OP_READ)] = {
 494                 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
 495                 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
 496         },
 497         [C(OP_WRITE)] = {
 498                 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
 499                 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
 500         },
 501         [C(OP_PREFETCH)] = {
 502                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 503                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 504         },
 505 },
 506 [C(L1I)] = {
 507         [C(OP_READ)] = {
 508                 [C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
 509                 [C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
 510         },
 511         [ C(OP_WRITE) ] = {
 512                 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
 513                 [ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
 514         },
 515         [ C(OP_PREFETCH) ] = {
 516                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 517                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 518         },
 519 },
 520 [C(LL)] = {
 521         [C(OP_READ)] = {
 522                 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
 523                 [C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
 524         },
 525         [C(OP_WRITE)] = {
 526                 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
 527                 [C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
 528         },
 529         [C(OP_PREFETCH)] = {
 530                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 531                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 532         },
 533 },
 534 [C(DTLB)] = {
 535         [C(OP_READ)] = {
 536                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 537                 [C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
 538         },
 539         [ C(OP_WRITE) ] = {
 540                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 541                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 542         },
 543         [ C(OP_PREFETCH) ] = {
 544                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 545                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 546         },
 547 },
 548 [C(ITLB)] = {
 549         [C(OP_READ)] = {
 550                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 551                 [C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
 552         },
 553         [ C(OP_WRITE) ] = {
 554                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 555                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 556         },
 557         [ C(OP_PREFETCH) ] = {
 558                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 559                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 560         },
 561 },
 562 [C(BPU)] = {
 563         [C(OP_READ)] = {
 564                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 565                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 566         },
 567         [ C(OP_WRITE) ] = {
 568                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 569                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 570         },
 571         [ C(OP_PREFETCH) ] = {
 572                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 573                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 574         },
 575 },
 576 [C(NODE)] = {
 577         [C(OP_READ)] = {
 578                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 579                 [C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
 580         },
 581         [ C(OP_WRITE) ] = {
 582                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 583                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 584         },
 585         [ C(OP_PREFETCH) ] = {
 586                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 587                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 588         },
 589 },
 590 };
 591
 592 static const struct sparc_pmu niagara2_pmu = {
 593         .event_map      = niagara2_event_map,
 594         .cache_map      = &niagara2_cache_map,
 595         .max_events     = ARRAY_SIZE(niagara2_perfmon_event_map),
 596         .read_pmc       = sparc_default_read_pmc,
 597         .write_pmc      = sparc_default_write_pmc,
 598         .upper_shift    = 19,
 599         .lower_shift    = 6,
 600         .event_mask     = 0xfff,
 601         .user_bit       = PCR_UTRACE,
 602         .priv_bit       = PCR_STRACE,
 603         .hv_bit         = PCR_N2_HTRACE,
 604         .irq_bit        = 0x30,
 605         .upper_nop      = 0x220,
 606         .lower_nop      = 0x220,
 607         .flags          = (SPARC_PMU_ALL_EXCLUDES_SAME |
 608                            SPARC_PMU_HAS_CONFLICTS),
 609         .max_hw_events  = 2,
 610         .num_pcrs       = 1,
 611         .num_pic_regs   = 1,
 612 };
 613
 614 static const struct perf_event_map niagara4_perfmon_event_map[] = {
 615         [PERF_COUNT_HW_CPU_CYCLES] = { (26 << 6) },
 616         [PERF_COUNT_HW_INSTRUCTIONS] = { (3 << 6) | 0x3f },
 617         [PERF_COUNT_HW_CACHE_REFERENCES] = { (3 << 6) | 0x04 },
 618         [PERF_COUNT_HW_CACHE_MISSES] = { (16 << 6) | 0x07 },
 619         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { (4 << 6) | 0x01 },
 620         [PERF_COUNT_HW_BRANCH_MISSES] = { (25 << 6) | 0x0f },
 621 };
 622
 623 static const struct perf_event_map *niagara4_event_map(int event_id)
 624 {
 625         return &niagara4_perfmon_event_map[event_id];
 626 }
 627
 628 static const cache_map_t niagara4_cache_map = {
 629 [C(L1D)] = {
 630         [C(OP_READ)] = {
 631                 [C(RESULT_ACCESS)] = { (3 << 6) | 0x04 },
 632                 [C(RESULT_MISS)] = { (16 << 6) | 0x07 },
 633         },
 634         [C(OP_WRITE)] = {
 635                 [C(RESULT_ACCESS)] = { (3 << 6) | 0x08 },
 636                 [C(RESULT_MISS)] = { (16 << 6) | 0x07 },
 637         },
 638         [C(OP_PREFETCH)] = {
 639                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 640                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 641         },
 642 },
 643 [C(L1I)] = {
 644         [C(OP_READ)] = {
 645                 [C(RESULT_ACCESS)] = { (3 << 6) | 0x3f },
 646                 [C(RESULT_MISS)] = { (11 << 6) | 0x03 },
 647         },
 648         [ C(OP_WRITE) ] = {
 649                 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
 650                 [ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
 651         },
 652         [ C(OP_PREFETCH) ] = {
 653                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 654                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 655         },
 656 },
 657 [C(LL)] = {
 658         [C(OP_READ)] = {
 659                 [C(RESULT_ACCESS)] = { (3 << 6) | 0x04 },
 660                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 661         },
 662         [C(OP_WRITE)] = {
 663                 [C(RESULT_ACCESS)] = { (3 << 6) | 0x08 },
 664                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 665         },
 666         [C(OP_PREFETCH)] = {
 667                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 668                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 669         },
 670 },
 671 [C(DTLB)] = {
 672         [C(OP_READ)] = {
 673                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 674                 [C(RESULT_MISS)] = { (17 << 6) | 0x3f },
 675         },
 676         [ C(OP_WRITE) ] = {
 677                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 678                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 679         },
 680         [ C(OP_PREFETCH) ] = {
 681                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 682                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 683         },
 684 },
 685 [C(ITLB)] = {
 686         [C(OP_READ)] = {
 687                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 688                 [C(RESULT_MISS)] = { (6 << 6) | 0x3f },
 689         },
 690         [ C(OP_WRITE) ] = {
 691                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 692                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 693         },
 694         [ C(OP_PREFETCH) ] = {
 695                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 696                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 697         },
 698 },
 699 [C(BPU)] = {
 700         [C(OP_READ)] = {
 701                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 702                 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
 703         },
 704         [ C(OP_WRITE) ] = {
 705                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 706                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 707         },
 708         [ C(OP_PREFETCH) ] = {
 709                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 710                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 711         },
 712 },
 713 [C(NODE)] = {
 714         [C(OP_READ)] = {
 715                 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
 716                 [C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
 717         },
 718         [ C(OP_WRITE) ] = {
 719                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 720                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 721         },
 722         [ C(OP_PREFETCH) ] = {
 723                 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
 724                 [ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
 725         },
 726 },
 727 };
 728
 729 static u32 sparc_vt_read_pmc(int idx)
 730 {
 731         u64 val = pcr_ops->read_pic(idx);
 732
 733         return val & 0xffffffff;
 734 }
 735
 736 static void sparc_vt_write_pmc(int idx, u64 val)
 737 {
 738         u64 pcr;
 739
 740         /* There seems to be an internal latch on the overflow event
 741          * on SPARC-T4 that prevents it from triggering unless you
 742          * update the PIC exactly as we do here.  The requirement
 743          * seems to be that you have to turn off event counting in the
 744          * PCR around the PIC update.
 745          *
 746          * For example, after the following sequence:
 747          *
 748          * 1) set PIC to -1
 749          * 2) enable event counting and overflow reporting in PCR
 750          * 3) overflow triggers, softint 15 handler invoked
 751          * 4) clear OV bit in PCR
 752          * 5) write PIC to -1
 753          *
 754          * a subsequent overflow event will not trigger.  This
 755          * sequence works on SPARC-T3 and previous chips.
 756          */
 757         pcr = pcr_ops->read_pcr(idx);
 758         pcr_ops->write_pcr(idx, PCR_N4_PICNPT);
 759
 760         pcr_ops->write_pic(idx, val & 0xffffffff);
 761
 762         pcr_ops->write_pcr(idx, pcr);
 763 }
 764
 765 static const struct sparc_pmu niagara4_pmu = {
 766         .event_map      = niagara4_event_map,
 767         .cache_map      = &niagara4_cache_map,
 768         .max_events     = ARRAY_SIZE(niagara4_perfmon_event_map),
 769         .read_pmc       = sparc_vt_read_pmc,
 770         .write_pmc      = sparc_vt_write_pmc,
 771         .upper_shift    = 5,
 772         .lower_shift    = 5,
 773         .event_mask     = 0x7ff,
 774         .user_bit       = PCR_N4_UTRACE,
 775         .priv_bit       = PCR_N4_STRACE,
 776
 777         /* We explicitly don't support hypervisor tracing.  The T4
 778          * generates the overflow event for precise events via a trap
 779          * which will not be generated (ie. it's completely lost) if
 780          * we happen to be in the hypervisor when the event triggers.
 781          * Essentially, the overflow event reporting is completely
 782          * unusable when you have hypervisor mode tracing enabled.
 783          */
 784         .hv_bit         = 0,
 785
 786         .irq_bit        = PCR_N4_TOE,
 787         .upper_nop      = 0,
 788         .lower_nop      = 0,
 789         .flags          = 0,
 790         .max_hw_events  = 4,
 791         .num_pcrs       = 4,
 792         .num_pic_regs   = 4,
 793 };
 794
 795 static const struct sparc_pmu *sparc_pmu __read_mostly;
 796
 797 static u64 event_encoding(u64 event_id, int idx)
 798 {
 799         if (idx == PIC_UPPER_INDEX)
 800                 event_id <<= sparc_pmu->upper_shift;
 801         else
 802                 event_id <<= sparc_pmu->lower_shift;
 803         return event_id;
 804 }
 805
 806 static u64 mask_for_index(int idx)
 807 {
 808         return event_encoding(sparc_pmu->event_mask, idx);
 809 }
 810
 811 static u64 nop_for_index(int idx)
 812 {
 813         return event_encoding(idx == PIC_UPPER_INDEX ?
 814                               sparc_pmu->upper_nop :
 815                               sparc_pmu->lower_nop, idx);
 816 }
 817
 818 static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
 819 {
 820         u64 enc, val, mask = mask_for_index(idx);
 821         int pcr_index = 0;
 822
 823         if (sparc_pmu->num_pcrs > 1)
 824                 pcr_index = idx;
 825
 826         enc = perf_event_get_enc(cpuc->events[idx]);
 827
 828         val = cpuc->pcr[pcr_index];
 829         val &= ~mask;
 830         val |= event_encoding(enc, idx);
 831         cpuc->pcr[pcr_index] = val;
 832
 833         pcr_ops->write_pcr(pcr_index, cpuc->pcr[pcr_index]);
 834 }
 835
 836 static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
 837 {
 838         u64 mask = mask_for_index(idx);
 839         u64 nop = nop_for_index(idx);
 840         int pcr_index = 0;
 841         u64 val;
 842
 843         if (sparc_pmu->num_pcrs > 1)
 844                 pcr_index = idx;
 845
 846         val = cpuc->pcr[pcr_index];
 847         val &= ~mask;
 848         val |= nop;
 849         cpuc->pcr[pcr_index] = val;
 850
 851         pcr_ops->write_pcr(pcr_index, cpuc->pcr[pcr_index]);
 852 }
 853
 854 static u64 sparc_perf_event_update(struct perf_event *event,
 855                                    struct hw_perf_event *hwc, int idx)
 856 {
 857         int shift = 64 - 32;
 858         u64 prev_raw_count, new_raw_count;
 859         s64 delta;
 860
 861 again:
 862         prev_raw_count = local64_read(&hwc->prev_count);
 863         new_raw_count = sparc_pmu->read_pmc(idx);
 864
 865         if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 866                              new_raw_count) != prev_raw_count)
 867                 goto again;
 868
 869         delta = (new_raw_count << shift) - (prev_raw_count << shift);
 870         delta >>= shift;
 871
 872         local64_add(delta, &event->count);
 873         local64_sub(delta, &hwc->period_left);
 874
 875         return new_raw_count;
 876 }
 877
 878 static int sparc_perf_event_set_period(struct perf_event *event,
 879                                        struct hw_perf_event *hwc, int idx)
 880 {
 881         s64 left = local64_read(&hwc->period_left);
 882         s64 period = hwc->sample_period;
 883         int ret = 0;
 884
 885         if (unlikely(left <= -period)) {
 886                 left = period;
 887                 local64_set(&hwc->period_left, left);
 888                 hwc->last_period = period;
 889                 ret = 1;
 890         }
 891
 892         if (unlikely(left <= 0)) {
 893                 left += period;
 894                 local64_set(&hwc->period_left, left);
 895                 hwc->last_period = period;
 896                 ret = 1;
 897         }
 898         if (left > MAX_PERIOD)
 899                 left = MAX_PERIOD;
 900
 901         local64_set(&hwc->prev_count, (u64)-left);
 902
 903         sparc_pmu->write_pmc(idx, (u64)(-left) & 0xffffffff);
 904
 905         perf_event_update_userpage(event);
 906
 907         return ret;
 908 }
 909
 910 static void read_in_all_counters(struct cpu_hw_events *cpuc)
 911 {
 912         int i;
 913
 914         for (i = 0; i < cpuc->n_events; i++) {
 915                 struct perf_event *cp = cpuc->event[i];
 916
 917                 if (cpuc->current_idx[i] != PIC_NO_INDEX &&
 918                     cpuc->current_idx[i] != cp->hw.idx) {
 919                         sparc_perf_event_update(cp, &cp->hw,
 920                                                 cpuc->current_idx[i]);
 921                         cpuc->current_idx[i] = PIC_NO_INDEX;
 922                 }
 923         }
 924 }
 925
 926 /* On this PMU all PICs are programmed using a single PCR.  Calculate
 927  * the combined control register value.
 928  *
 929  * For such chips we require that all of the events have the same
 930  * configuration, so just fetch the settings from the first entry.
 931  */
 932 static void calculate_single_pcr(struct cpu_hw_events *cpuc)
 933 {
 934         int i;
 935
 936         if (!cpuc->n_added)
 937                 goto out;
 938
 939         /* Assign to counters all unassigned events.  */
 940         for (i = 0; i < cpuc->n_events; i++) {
 941                 struct perf_event *cp = cpuc->event[i];
 942                 struct hw_perf_event *hwc = &cp->hw;
 943                 int idx = hwc->idx;
 944                 u64 enc;
 945
 946                 if (cpuc->current_idx[i] != PIC_NO_INDEX)
 947                         continue;
 948
 949                 sparc_perf_event_set_period(cp, hwc, idx);
 950                 cpuc->current_idx[i] = idx;
 951
 952                 enc = perf_event_get_enc(cpuc->events[i]);
 953                 cpuc->pcr[0] &= ~mask_for_index(idx);
 954                 if (hwc->state & PERF_HES_STOPPED)
 955                         cpuc->pcr[0] |= nop_for_index(idx);
 956                 else
 957                         cpuc->pcr[0] |= event_encoding(enc, idx);
 958         }
 959 out:
 960         cpuc->pcr[0] |= cpuc->event[0]->hw.config_base;
 961 }
 962
 963 static void sparc_pmu_start(struct perf_event *event, int flags);
 964
 965 /* On this PMU each PIC has it's own PCR control register.  */
 966 static void calculate_multiple_pcrs(struct cpu_hw_events *cpuc)
 967 {
 968         int i;
 969
 970         if (!cpuc->n_added)
 971                 goto out;
 972
 973         for (i = 0; i < cpuc->n_events; i++) {
 974                 struct perf_event *cp = cpuc->event[i];
 975                 struct hw_perf_event *hwc = &cp->hw;
 976                 int idx = hwc->idx;
 977
 978                 if (cpuc->current_idx[i] != PIC_NO_INDEX)
 979                         continue;
 980
 981                 cpuc->current_idx[i] = idx;
 982
 983                 sparc_pmu_start(cp, PERF_EF_RELOAD);
 984         }
 985 out:
 986         for (i = 0; i < cpuc->n_events; i++) {
 987                 struct perf_event *cp = cpuc->event[i];
 988                 int idx = cp->hw.idx;
 989
 990                 cpuc->pcr[idx] |= cp->hw.config_base;
 991         }
 992 }
 993
 994 /* If performance event entries have been added, move existing events
 995  * around (if necessary) and then assign new entries to counters.
 996  */
 997 static void update_pcrs_for_enable(struct cpu_hw_events *cpuc)
 998 {
 999         if (cpuc->n_added)
1000                 read_in_all_counters(cpuc);
1001
1002         if (sparc_pmu->num_pcrs == 1) {
1003                 calculate_single_pcr(cpuc);
1004         } else {
1005                 calculate_multiple_pcrs(cpuc);
1006         }
1007 }
1008
1009 static void sparc_pmu_enable(struct pmu *pmu)
1010 {
1011         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1012         int i;
1013
1014         if (cpuc->enabled)
1015                 return;
1016
1017         cpuc->enabled = 1;
1018         barrier();
1019
1020         if (cpuc->n_events)
1021                 update_pcrs_for_enable(cpuc);
1022
1023         for (i = 0; i < sparc_pmu->num_pcrs; i++)
1024                 pcr_ops->write_pcr(i, cpuc->pcr[i]);
1025 }
1026
1027 static void sparc_pmu_disable(struct pmu *pmu)
1028 {
1029         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1030         int i;
1031
1032         if (!cpuc->enabled)
1033                 return;
1034
1035         cpuc->enabled = 0;
1036         cpuc->n_added = 0;
1037
1038         for (i = 0; i < sparc_pmu->num_pcrs; i++) {
1039                 u64 val = cpuc->pcr[i];
1040
1041                 val &= ~(sparc_pmu->user_bit | sparc_pmu->priv_bit |
1042                          sparc_pmu->hv_bit | sparc_pmu->irq_bit);
1043                 cpuc->pcr[i] = val;
1044                 pcr_ops->write_pcr(i, cpuc->pcr[i]);
1045         }
1046 }
1047
1048 static int active_event_index(struct cpu_hw_events *cpuc,
1049                               struct perf_event *event)
1050 {
1051         int i;
1052
1053         for (i = 0; i < cpuc->n_events; i++) {
1054                 if (cpuc->event[i] == event)
1055                         break;
1056         }
1057         BUG_ON(i == cpuc->n_events);
1058         return cpuc->current_idx[i];
1059 }
1060
1061 static void sparc_pmu_start(struct perf_event *event, int flags)
1062 {
1063         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1064         int idx = active_event_index(cpuc, event);
1065
1066         if (flags & PERF_EF_RELOAD) {
1067                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1068                 sparc_perf_event_set_period(event, &event->hw, idx);
1069         }
1070
1071         event->hw.state = 0;
1072
1073         sparc_pmu_enable_event(cpuc, &event->hw, idx);
1074 }
1075
1076 static void sparc_pmu_stop(struct perf_event *event, int flags)
1077 {
1078         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1079         int idx = active_event_index(cpuc, event);
1080
1081         if (!(event->hw.state & PERF_HES_STOPPED)) {
1082                 sparc_pmu_disable_event(cpuc, &event->hw, idx);
1083                 event->hw.state |= PERF_HES_STOPPED;
1084         }
1085
1086         if (!(event->hw.state & PERF_HES_UPTODATE) && (flags & PERF_EF_UPDATE)) {
1087                 sparc_perf_event_update(event, &event->hw, idx);
1088                 event->hw.state |= PERF_HES_UPTODATE;
1089         }
1090 }
1091
1092 static void sparc_pmu_del(struct perf_event *event, int _flags)
1093 {
1094         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1095         unsigned long flags;
1096         int i;
1097
1098         local_irq_save(flags);
1099
1100         for (i = 0; i < cpuc->n_events; i++) {
1101                 if (event == cpuc->event[i]) {
1102                         /* Absorb the final count and turn off the
1103                          * event.
1104                          */
1105                         sparc_pmu_stop(event, PERF_EF_UPDATE);
1106
1107                         /* Shift remaining entries down into
1108                          * the existing slot.
1109                          */
1110                         while (++i < cpuc->n_events) {
1111                                 cpuc->event[i - 1] = cpuc->event[i];
1112                                 cpuc->events[i - 1] = cpuc->events[i];
1113                                 cpuc->current_idx[i - 1] =
1114                                         cpuc->current_idx[i];
1115                         }
1116
1117                         perf_event_update_userpage(event);
1118
1119                         cpuc->n_events--;
1120                         break;
1121                 }
1122         }
1123
1124         local_irq_restore(flags);
1125 }
1126
1127 static void sparc_pmu_read(struct perf_event *event)
1128 {
1129         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1130         int idx = active_event_index(cpuc, event);
1131         struct hw_perf_event *hwc = &event->hw;
1132
1133         sparc_perf_event_update(event, hwc, idx);
1134 }
1135
1136 static atomic_t active_events = ATOMIC_INIT(0);
1137 static DEFINE_MUTEX(pmc_grab_mutex);
1138
1139 static void perf_stop_nmi_watchdog(void *unused)
1140 {
1141         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1142         int i;
1143
1144         stop_nmi_watchdog(NULL);
1145         for (i = 0; i < sparc_pmu->num_pcrs; i++)
1146                 cpuc->pcr[i] = pcr_ops->read_pcr(i);
1147 }
1148
1149 static void perf_event_grab_pmc(void)
1150 {
1151         if (atomic_inc_not_zero(&active_events))
1152                 return;
1153
1154         mutex_lock(&pmc_grab_mutex);
1155         if (atomic_read(&active_events) == 0) {
1156                 if (atomic_read(&nmi_active) > 0) {
1157                         on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
1158                         BUG_ON(atomic_read(&nmi_active) != 0);
1159                 }
1160                 atomic_inc(&active_events);
1161         }
1162         mutex_unlock(&pmc_grab_mutex);
1163 }
1164
1165 static void perf_event_release_pmc(void)
1166 {
1167         if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
1168                 if (atomic_read(&nmi_active) == 0)
1169                         on_each_cpu(start_nmi_watchdog, NULL, 1);
1170                 mutex_unlock(&pmc_grab_mutex);
1171         }
1172 }
1173
1174 static const struct perf_event_map *sparc_map_cache_event(u64 config)
1175 {
1176         unsigned int cache_type, cache_op, cache_result;
1177         const struct perf_event_map *pmap;
1178
1179         if (!sparc_pmu->cache_map)
1180                 return ERR_PTR(-ENOENT);
1181
1182         cache_type = (config >>  0) & 0xff;
1183         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
1184                 return ERR_PTR(-EINVAL);
1185
1186         cache_op = (config >>  8) & 0xff;
1187         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
1188                 return ERR_PTR(-EINVAL);
1189
1190         cache_result = (config >> 16) & 0xff;
1191         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1192                 return ERR_PTR(-EINVAL);
1193
1194         pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
1195
1196         if (pmap->encoding == CACHE_OP_UNSUPPORTED)
1197                 return ERR_PTR(-ENOENT);
1198
1199         if (pmap->encoding == CACHE_OP_NONSENSE)
1200                 return ERR_PTR(-EINVAL);
1201
1202         return pmap;
1203 }
1204
1205 static void hw_perf_event_destroy(struct perf_event *event)
1206 {
1207         perf_event_release_pmc();
1208 }
1209
1210 /* Make sure all events can be scheduled into the hardware at
1211  * the same time.  This is simplified by the fact that we only
1212  * need to support 2 simultaneous HW events.
1213  *
1214  * As a side effect, the evts[]->hw.idx values will be assigned
1215  * on success.  These are pending indexes.  When the events are
1216  * actually programmed into the chip, these values will propagate
1217  * to the per-cpu cpuc->current_idx[] slots, see the code in
1218  * maybe_change_configuration() for details.
1219  */
1220 static int sparc_check_constraints(struct perf_event **evts,
1221                                    unsigned long *events, int n_ev)
1222 {
1223         u8 msk0 = 0, msk1 = 0;
1224         int idx0 = 0;
1225
1226         /* This case is possible when we are invoked from
1227          * hw_perf_group_sched_in().
1228          */
1229         if (!n_ev)
1230                 return 0;
1231
1232         if (n_ev > sparc_pmu->max_hw_events)
1233                 return -1;
1234
1235         if (!(sparc_pmu->flags & SPARC_PMU_HAS_CONFLICTS)) {
1236                 int i;
1237
1238                 for (i = 0; i < n_ev; i++)
1239                         evts[i]->hw.idx = i;
1240                 return 0;
1241         }
1242
1243         msk0 = perf_event_get_msk(events[0]);
1244         if (n_ev == 1) {
1245                 if (msk0 & PIC_LOWER)
1246                         idx0 = 1;
1247                 goto success;
1248         }
1249         BUG_ON(n_ev != 2);
1250         msk1 = perf_event_get_msk(events[1]);
1251
1252         /* If both events can go on any counter, OK.  */
1253         if (msk0 == (PIC_UPPER | PIC_LOWER) &&
1254             msk1 == (PIC_UPPER | PIC_LOWER))
1255                 goto success;
1256
1257         /* If one event is limited to a specific counter,
1258          * and the other can go on both, OK.
1259          */
1260         if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
1261             msk1 == (PIC_UPPER | PIC_LOWER)) {
1262                 if (msk0 & PIC_LOWER)
1263                         idx0 = 1;
1264                 goto success;
1265         }
1266
1267         if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
1268             msk0 == (PIC_UPPER | PIC_LOWER)) {
1269                 if (msk1 & PIC_UPPER)
1270                         idx0 = 1;
1271                 goto success;
1272         }
1273
1274         /* If the events are fixed to different counters, OK.  */
1275         if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
1276             (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
1277                 if (msk0 & PIC_LOWER)
1278                         idx0 = 1;
1279                 goto success;
1280         }
1281
1282         /* Otherwise, there is a conflict.  */
1283         return -1;
1284
1285 success:
1286         evts[0]->hw.idx = idx0;
1287         if (n_ev == 2)
1288                 evts[1]->hw.idx = idx0 ^ 1;
1289         return 0;
1290 }
1291
1292 static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
1293 {
1294         int eu = 0, ek = 0, eh = 0;
1295         struct perf_event *event;
1296         int i, n, first;
1297
1298         if (!(sparc_pmu->flags & SPARC_PMU_ALL_EXCLUDES_SAME))
1299                 return 0;
1300
1301         n = n_prev + n_new;
1302         if (n <= 1)
1303                 return 0;
1304
1305         first = 1;
1306         for (i = 0; i < n; i++) {
1307                 event = evts[i];
1308                 if (first) {
1309                         eu = event->attr.exclude_user;
1310                         ek = event->attr.exclude_kernel;
1311                         eh = event->attr.exclude_hv;
1312                         first = 0;
1313                 } else if (event->attr.exclude_user != eu ||
1314                            event->attr.exclude_kernel != ek ||
1315                            event->attr.exclude_hv != eh) {
1316                         return -EAGAIN;
1317                 }
1318         }
1319
1320         return 0;
1321 }
1322
1323 static int collect_events(struct perf_event *group, int max_count,
1324                           struct perf_event *evts[], unsigned long *events,
1325                           int *current_idx)
1326 {
1327         struct perf_event *event;
1328         int n = 0;
1329
1330         if (!is_software_event(group)) {
1331                 if (n >= max_count)
1332                         return -1;
1333                 evts[n] = group;
1334                 events[n] = group->hw.event_base;
1335                 current_idx[n++] = PIC_NO_INDEX;
1336         }
1337         list_for_each_entry(event, &group->sibling_list, group_entry) {
1338                 if (!is_software_event(event) &&
1339                     event->state != PERF_EVENT_STATE_OFF) {
1340                         if (n >= max_count)
1341                                 return -1;
1342                         evts[n] = event;
1343                         events[n] = event->hw.event_base;
1344                         current_idx[n++] = PIC_NO_INDEX;
1345                 }
1346         }
1347         return n;
1348 }
1349
1350 static int sparc_pmu_add(struct perf_event *event, int ef_flags)
1351 {
1352         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1353         int n0, ret = -EAGAIN;
1354         unsigned long flags;
1355
1356         local_irq_save(flags);
1357
1358         n0 = cpuc->n_events;
1359         if (n0 >= sparc_pmu->max_hw_events)
1360                 goto out;
1361
1362         cpuc->event[n0] = event;
1363         cpuc->events[n0] = event->hw.event_base;
1364         cpuc->current_idx[n0] = PIC_NO_INDEX;
1365
1366         event->hw.state = PERF_HES_UPTODATE;
1367         if (!(ef_flags & PERF_EF_START))
1368                 event->hw.state |= PERF_HES_STOPPED;
1369
1370         /*
1371          * If group events scheduling transaction was started,
1372          * skip the schedulability test here, it will be performed
1373          * at commit time(->commit_txn) as a whole
1374          */
1375         if (cpuc->group_flag & PERF_EVENT_TXN)
1376                 goto nocheck;
1377
1378         if (check_excludes(cpuc->event, n0, 1))
1379                 goto out;
1380         if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1381                 goto out;
1382
1383 nocheck:
1384         cpuc->n_events++;
1385         cpuc->n_added++;
1386
1387         ret = 0;
1388 out:
1389         local_irq_restore(flags);
1390         return ret;
1391 }
1392
1393 static int sparc_pmu_event_init(struct perf_event *event)
1394 {
1395         struct perf_event_attr *attr = &event->attr;
1396         struct perf_event *evts[MAX_HWEVENTS];
1397         struct hw_perf_event *hwc = &event->hw;
1398         unsigned long events[MAX_HWEVENTS];
1399         int current_idx_dmy[MAX_HWEVENTS];
1400         const struct perf_event_map *pmap;
1401         int n;
1402
1403         if (atomic_read(&nmi_active) < 0)
1404                 return -ENODEV;
1405
1406         /* does not support taken branch sampling */
1407         if (has_branch_stack(event))
1408                 return -EOPNOTSUPP;
1409
1410         switch (attr->type) {
1411         case PERF_TYPE_HARDWARE:
1412                 if (attr->config >= sparc_pmu->max_events)
1413                         return -EINVAL;
1414                 pmap = sparc_pmu->event_map(attr->config);
1415                 break;
1416
1417         case PERF_TYPE_HW_CACHE:
1418                 pmap = sparc_map_cache_event(attr->config);
1419                 if (IS_ERR(pmap))
1420                         return PTR_ERR(pmap);
1421                 break;
1422
1423         case PERF_TYPE_RAW:
1424                 pmap = NULL;
1425                 break;
1426
1427         default:
1428                 return -ENOENT;
1429
1430         }
1431
1432         if (pmap) {
1433                 hwc->event_base = perf_event_encode(pmap);
1434         } else {
1435                 /*
1436                  * User gives us "(encoding << 16) | pic_mask" for
1437                  * PERF_TYPE_RAW events.
1438                  */
1439                 hwc->event_base = attr->config;
1440         }
1441
1442         /* We save the enable bits in the config_base.  */
1443         hwc->config_base = sparc_pmu->irq_bit;
1444         if (!attr->exclude_user)
1445                 hwc->config_base |= sparc_pmu->user_bit;
1446         if (!attr->exclude_kernel)
1447                 hwc->config_base |= sparc_pmu->priv_bit;
1448         if (!attr->exclude_hv)
1449                 hwc->config_base |= sparc_pmu->hv_bit;
1450
1451         n = 0;
1452         if (event->group_leader != event) {
1453                 n = collect_events(event->group_leader,
1454                                    sparc_pmu->max_hw_events - 1,
1455                                    evts, events, current_idx_dmy);
1456                 if (n < 0)
1457                         return -EINVAL;
1458         }
1459         events[n] = hwc->event_base;
1460         evts[n] = event;
1461
1462         if (check_excludes(evts, n, 1))
1463                 return -EINVAL;
1464
1465         if (sparc_check_constraints(evts, events, n + 1))
1466                 return -EINVAL;
1467
1468         hwc->idx = PIC_NO_INDEX;
1469
1470         /* Try to do all error checking before this point, as unwinding
1471          * state after grabbing the PMC is difficult.
1472          */
1473         perf_event_grab_pmc();
1474         event->destroy = hw_perf_event_destroy;
1475
1476         if (!hwc->sample_period) {
1477                 hwc->sample_period = MAX_PERIOD;
1478                 hwc->last_period = hwc->sample_period;
1479                 local64_set(&hwc->period_left, hwc->sample_period);
1480         }
1481
1482         return 0;
1483 }
1484
1485 /*
1486  * Start group events scheduling transaction
1487  * Set the flag to make pmu::enable() not perform the
1488  * schedulability test, it will be performed at commit time
1489  */
1490 static void sparc_pmu_start_txn(struct pmu *pmu)
1491 {
1492         struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1493
1494         perf_pmu_disable(pmu);
1495         cpuhw->group_flag |= PERF_EVENT_TXN;
1496 }
1497
1498 /*
1499  * Stop group events scheduling transaction
1500  * Clear the flag and pmu::enable() will perform the
1501  * schedulability test.
1502  */
1503 static void sparc_pmu_cancel_txn(struct pmu *pmu)
1504 {
1505         struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1506
1507         cpuhw->group_flag &= ~PERF_EVENT_TXN;
1508         perf_pmu_enable(pmu);
1509 }
1510
1511 /*
1512  * Commit group events scheduling transaction
1513  * Perform the group schedulability test as a whole
1514  * Return 0 if success
1515  */
1516 static int sparc_pmu_commit_txn(struct pmu *pmu)
1517 {
1518         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1519         int n;
1520
1521         if (!sparc_pmu)
1522                 return -EINVAL;
1523
1524         cpuc = this_cpu_ptr(&cpu_hw_events);
1525         n = cpuc->n_events;
1526         if (check_excludes(cpuc->event, 0, n))
1527                 return -EINVAL;
1528         if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1529                 return -EAGAIN;
1530
1531         cpuc->group_flag &= ~PERF_EVENT_TXN;
1532         perf_pmu_enable(pmu);
1533         return 0;
1534 }
1535
1536 static struct pmu pmu = {
1537         .pmu_enable     = sparc_pmu_enable,
1538         .pmu_disable    = sparc_pmu_disable,
1539         .event_init     = sparc_pmu_event_init,
1540         .add            = sparc_pmu_add,
1541         .del            = sparc_pmu_del,
1542         .start          = sparc_pmu_start,
1543         .stop           = sparc_pmu_stop,
1544         .read           = sparc_pmu_read,
1545         .start_txn      = sparc_pmu_start_txn,
1546         .cancel_txn     = sparc_pmu_cancel_txn,
1547         .commit_txn     = sparc_pmu_commit_txn,
1548 };
1549
1550 void perf_event_print_debug(void)
1551 {
1552         unsigned long flags;
1553         int cpu, i;
1554
1555         if (!sparc_pmu)
1556                 return;
1557
1558         local_irq_save(flags);
1559
1560         cpu = smp_processor_id();
1561
1562         pr_info("\n");
1563         for (i = 0; i < sparc_pmu->num_pcrs; i++)
1564                 pr_info("CPU#%d: PCR%d[%016llx]\n",
1565                         cpu, i, pcr_ops->read_pcr(i));
1566         for (i = 0; i < sparc_pmu->num_pic_regs; i++)
1567                 pr_info("CPU#%d: PIC%d[%016llx]\n",
1568                         cpu, i, pcr_ops->read_pic(i));
1569
1570         local_irq_restore(flags);
1571 }
1572
1573 static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1574                                             unsigned long cmd, void *__args)
1575 {
1576         struct die_args *args = __args;
1577         struct perf_sample_data data;
1578         struct cpu_hw_events *cpuc;
1579         struct pt_regs *regs;
1580         int i;
1581
1582         if (!atomic_read(&active_events))
1583                 return NOTIFY_DONE;
1584
1585         switch (cmd) {
1586         case DIE_NMI:
1587                 break;
1588
1589         default:
1590                 return NOTIFY_DONE;
1591         }
1592
1593         regs = args->regs;
1594
1595         cpuc = this_cpu_ptr(&cpu_hw_events);
1596
1597         /* If the PMU has the TOE IRQ enable bits, we need to do a
1598          * dummy write to the %pcr to clear the overflow bits and thus
1599          * the interrupt.
1600          *
1601          * Do this before we peek at the counters to determine
1602          * overflow so we don't lose any events.
1603          */
1604         if (sparc_pmu->irq_bit &&
1605             sparc_pmu->num_pcrs == 1)
1606                 pcr_ops->write_pcr(0, cpuc->pcr[0]);
1607
1608         for (i = 0; i < cpuc->n_events; i++) {
1609                 struct perf_event *event = cpuc->event[i];
1610                 int idx = cpuc->current_idx[i];
1611                 struct hw_perf_event *hwc;
1612                 u64 val;
1613
1614                 if (sparc_pmu->irq_bit &&
1615                     sparc_pmu->num_pcrs > 1)
1616                         pcr_ops->write_pcr(idx, cpuc->pcr[idx]);
1617
1618                 hwc = &event->hw;
1619                 val = sparc_perf_event_update(event, hwc, idx);
1620                 if (val & (1ULL << 31))
1621                         continue;
1622
1623                 perf_sample_data_init(&data, 0, hwc->last_period);
1624                 if (!sparc_perf_event_set_period(event, hwc, idx))
1625                         continue;
1626
1627                 if (perf_event_overflow(event, &data, regs))
1628                         sparc_pmu_stop(event, 0);
1629         }
1630
1631         return NOTIFY_STOP;
1632 }
1633
1634 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1635         .notifier_call          = perf_event_nmi_handler,
1636 };
1637
1638 static bool __init supported_pmu(void)
1639 {
1640         if (!strcmp(sparc_pmu_type, "ultra3") ||
1641             !strcmp(sparc_pmu_type, "ultra3+") ||
1642             !strcmp(sparc_pmu_type, "ultra3i") ||
1643             !strcmp(sparc_pmu_type, "ultra4+")) {
1644                 sparc_pmu = &ultra3_pmu;
1645                 return true;
1646         }
1647         if (!strcmp(sparc_pmu_type, "niagara")) {
1648                 sparc_pmu = &niagara1_pmu;
1649                 return true;
1650         }
1651         if (!strcmp(sparc_pmu_type, "niagara2") ||
1652             !strcmp(sparc_pmu_type, "niagara3")) {
1653                 sparc_pmu = &niagara2_pmu;
1654                 return true;
1655         }
1656         if (!strcmp(sparc_pmu_type, "niagara4") ||
1657             !strcmp(sparc_pmu_type, "niagara5")) {
1658                 sparc_pmu = &niagara4_pmu;
1659                 return true;
1660         }
1661         return false;
1662 }
1663
1664 static int __init init_hw_perf_events(void)
1665 {
1666         int err;
1667
1668         pr_info("Performance events: ");
1669
1670         err = pcr_arch_init();
1671         if (err || !supported_pmu()) {
1672                 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1673                 return 0;
1674         }
1675
1676         pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1677
1678         perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1679         register_die_notifier(&perf_event_nmi_notifier);
1680
1681         return 0;
1682 }
1683 pure_initcall(init_hw_perf_events);
1684
1685 void perf_callchain_kernel(struct perf_callchain_entry *entry,
1686                            struct pt_regs *regs)
1687 {
1688         unsigned long ksp, fp;
1689 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1690         int graph = 0;
1691 #endif
1692
1693         stack_trace_flush();
1694
1695         perf_callchain_store(entry, regs->tpc);
1696
1697         ksp = regs->u_regs[UREG_I6];
1698         fp = ksp + STACK_BIAS;
1699         do {
1700                 struct sparc_stackf *sf;
1701                 struct pt_regs *regs;
1702                 unsigned long pc;
1703
1704                 if (!kstack_valid(current_thread_info(), fp))
1705                         break;
1706
1707                 sf = (struct sparc_stackf *) fp;
1708                 regs = (struct pt_regs *) (sf + 1);
1709
1710                 if (kstack_is_trap_frame(current_thread_info(), regs)) {
1711                         if (user_mode(regs))
1712                                 break;
1713                         pc = regs->tpc;
1714                         fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1715                 } else {
1716                         pc = sf->callers_pc;
1717                         fp = (unsigned long)sf->fp + STACK_BIAS;
1718                 }
1719                 perf_callchain_store(entry, pc);
1720 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1721                 if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1722                         int index = current->curr_ret_stack;
1723                         if (current->ret_stack && index >= graph) {
1724                                 pc = current->ret_stack[index - graph].ret;
1725                                 perf_callchain_store(entry, pc);
1726                                 graph++;
1727                         }
1728                 }
1729 #endif
1730         } while (entry->nr < PERF_MAX_STACK_DEPTH);
1731 }
1732
1733 static void perf_callchain_user_64(struct perf_callchain_entry *entry,
1734                                    struct pt_regs *regs)
1735 {
1736         unsigned long ufp;
1737
1738         ufp = regs->u_regs[UREG_I6] + STACK_BIAS;
1739         do {
1740                 struct sparc_stackf __user *usf;
1741                 struct sparc_stackf sf;
1742                 unsigned long pc;
1743
1744                 usf = (struct sparc_stackf __user *)ufp;
1745                 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1746                         break;
1747
1748                 pc = sf.callers_pc;
1749                 ufp = (unsigned long)sf.fp + STACK_BIAS;
1750                 perf_callchain_store(entry, pc);
1751         } while (entry->nr < PERF_MAX_STACK_DEPTH);
1752 }
1753
1754 static void perf_callchain_user_32(struct perf_callchain_entry *entry,
1755                                    struct pt_regs *regs)
1756 {
1757         unsigned long ufp;
1758
1759         ufp = regs->u_regs[UREG_I6] & 0xffffffffUL;
1760         do {
1761                 unsigned long pc;
1762
1763                 if (thread32_stack_is_64bit(ufp)) {
1764                         struct sparc_stackf __user *usf;
1765                         struct sparc_stackf sf;
1766
1767                         ufp += STACK_BIAS;
1768                         usf = (struct sparc_stackf __user *)ufp;
1769                         if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1770                                 break;
1771                         pc = sf.callers_pc & 0xffffffff;
1772                         ufp = ((unsigned long) sf.fp) & 0xffffffff;
1773                 } else {
1774                         struct sparc_stackf32 __user *usf;
1775                         struct sparc_stackf32 sf;
1776                         usf = (struct sparc_stackf32 __user *)ufp;
1777                         if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1778                                 break;
1779                         pc = sf.callers_pc;
1780                         ufp = (unsigned long)sf.fp;
1781                 }
1782                 perf_callchain_store(entry, pc);
1783         } while (entry->nr < PERF_MAX_STACK_DEPTH);
1784 }
1785
1786 void
1787 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1788 {
1789         perf_callchain_store(entry, regs->tpc);
1790
1791         if (!current->mm)
1792                 return;
1793
1794         flushw_user();
1795         if (test_thread_flag(TIF_32BIT))
1796                 perf_callchain_user_32(entry, regs);
1797         else
1798                 perf_callchain_user_64(entry, regs);
1799 }