arch/mips/sibyte/sb1250/bcm1250_tbprof.c

   1 /*
   2  * Copyright (C) 2001, 2002, 2003 Broadcom Corporation
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public License
   6  * as published by the Free Software Foundation; either version 2
   7  * of the License, or (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  17  */
  18
  19 #define SBPROF_TB_DEBUG 0
  20
  21 #include <linux/module.h>
  22 #include <linux/kernel.h>
  23 #include <linux/types.h>
  24 #include <linux/init.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/slab.h>
  27 #include <linux/vmalloc.h>
  28 #include <linux/fs.h>
  29 #include <linux/errno.h>
  30 #include <linux/reboot.h>
  31 #include <linux/wait.h>
  32 #include <asm/uaccess.h>
  33 #include <asm/io.h>
  34 #include <asm/sibyte/sb1250.h>
  35 #include <asm/sibyte/sb1250_regs.h>
  36 #include <asm/sibyte/sb1250_scd.h>
  37 #include <asm/sibyte/sb1250_int.h>
  38 #include <asm/sibyte/trace_prof.h>
  39
  40 #define DEVNAME "bcm1250_tbprof"
  41
  42 static struct sbprof_tb sbp;
  43
  44 #define TB_FULL (sbp.next_tb_sample == MAX_TB_SAMPLES)
  45
  46 /************************************************************************
  47  * Support for ZBbus sampling using the trace buffer
  48  *
  49  * We use the SCD performance counter interrupt, caused by a Zclk counter
  50  * overflow, to trigger the start of tracing.
  51  *
  52  * We set the trace buffer to sample everything and freeze on
  53  * overflow.
  54  *
  55  * We map the interrupt for trace_buffer_freeze to handle it on CPU 0.
  56  *
  57  ************************************************************************/
  58
  59 static u_int64_t tb_period;
  60
  61 static void arm_tb(void)
  62 {
  63         u_int64_t scdperfcnt;
  64         u_int64_t next = (1ULL << 40) - tb_period;
  65         u_int64_t tb_options = M_SCD_TRACE_CFG_FREEZE_FULL;
  66         /* Generate an SCD_PERFCNT interrupt in TB_PERIOD Zclks to
  67            trigger start of trace.  XXX vary sampling period */
  68         bus_writeq(0, IOADDR(A_SCD_PERF_CNT_1));
  69         scdperfcnt = bus_readq(IOADDR(A_SCD_PERF_CNT_CFG));
  70         /* Unfortunately, in Pass 2 we must clear all counters to knock down
  71            a previous interrupt request.  This means that bus profiling
  72            requires ALL of the SCD perf counters. */
  73         bus_writeq((scdperfcnt & ~M_SPC_CFG_SRC1) | // keep counters 0,2,3 as is
  74                    M_SPC_CFG_ENABLE |            // enable counting
  75                    M_SPC_CFG_CLEAR |             // clear all counters
  76                    V_SPC_CFG_SRC1(1),            // counter 1 counts cycles
  77                    IOADDR(A_SCD_PERF_CNT_CFG));
  78         bus_writeq(next, IOADDR(A_SCD_PERF_CNT_1));
  79         /* Reset the trace buffer */
  80         bus_writeq(M_SCD_TRACE_CFG_RESET, IOADDR(A_SCD_TRACE_CFG));
  81 #if 0 && defined(M_SCD_TRACE_CFG_FORCECNT)
  82         /* XXXKW may want to expose control to the data-collector */
  83         tb_options |= M_SCD_TRACE_CFG_FORCECNT;
  84 #endif
  85         bus_writeq(tb_options, IOADDR(A_SCD_TRACE_CFG));
  86         sbp.tb_armed = 1;
  87 }
  88
  89 static irqreturn_t sbprof_tb_intr(int irq, void *dev_id, struct pt_regs *regs)
  90 {
  91         int i;
  92         DBG(printk(DEVNAME ": tb_intr\n"));
  93         if (sbp.next_tb_sample < MAX_TB_SAMPLES) {
  94                 /* XXX should use XKPHYS to make writes bypass L2 */
  95                 u_int64_t *p = sbp.sbprof_tbbuf[sbp.next_tb_sample++];
  96                 /* Read out trace */
  97                 bus_writeq(M_SCD_TRACE_CFG_START_READ, IOADDR(A_SCD_TRACE_CFG));
  98                 __asm__ __volatile__ ("sync" : : : "memory");
  99                 /* Loop runs backwards because bundles are read out in reverse order */
 100                 for (i = 256 * 6; i > 0; i -= 6) {
 101                         // Subscripts decrease to put bundle in the order
 102                         //   t0 lo, t0 hi, t1 lo, t1 hi, t2 lo, t2 hi
 103                         p[i-1] = bus_readq(IOADDR(A_SCD_TRACE_READ)); // read t2 hi
 104                         p[i-2] = bus_readq(IOADDR(A_SCD_TRACE_READ)); // read t2 lo
 105                         p[i-3] = bus_readq(IOADDR(A_SCD_TRACE_READ)); // read t1 hi
 106                         p[i-4] = bus_readq(IOADDR(A_SCD_TRACE_READ)); // read t1 lo
 107                         p[i-5] = bus_readq(IOADDR(A_SCD_TRACE_READ)); // read t0 hi
 108                         p[i-6] = bus_readq(IOADDR(A_SCD_TRACE_READ)); // read t0 lo
 109                 }
 110                 if (!sbp.tb_enable) {
 111                         DBG(printk(DEVNAME ": tb_intr shutdown\n"));
 112                         bus_writeq(M_SCD_TRACE_CFG_RESET,
 113                                    IOADDR(A_SCD_TRACE_CFG));
 114                         sbp.tb_armed = 0;
 115                         wake_up(&sbp.tb_sync);
 116                 } else {
 117                         arm_tb();       // knock down current interrupt and get another one later
 118                 }
 119         } else {
 120                 /* No more trace buffer samples */
 121                 DBG(printk(DEVNAME ": tb_intr full\n"));
 122                 bus_writeq(M_SCD_TRACE_CFG_RESET, IOADDR(A_SCD_TRACE_CFG));
 123                 sbp.tb_armed = 0;
 124                 if (!sbp.tb_enable) {
 125                         wake_up(&sbp.tb_sync);
 126                 }
 127                 wake_up(&sbp.tb_read);
 128         }
 129         return IRQ_HANDLED;
 130 }
 131
 132 static irqreturn_t sbprof_pc_intr(int irq, void *dev_id, struct pt_regs *regs)
 133 {
 134         printk(DEVNAME ": unexpected pc_intr");
 135         return IRQ_NONE;
 136 }
 137
 138 int sbprof_zbprof_start(struct file *filp)
 139 {
 140         u_int64_t scdperfcnt;
 141
 142         if (sbp.tb_enable)
 143                 return -EBUSY;
 144
 145         DBG(printk(DEVNAME ": starting\n"));
 146
 147         sbp.tb_enable = 1;
 148         sbp.next_tb_sample = 0;
 149         filp->f_pos = 0;
 150
 151         if (request_irq
 152             (K_INT_TRACE_FREEZE, sbprof_tb_intr, 0, DEVNAME " trace freeze", &sbp)) {
 153                 return -EBUSY;
 154         }
 155         /* Make sure there isn't a perf-cnt interrupt waiting */
 156         scdperfcnt = bus_readq(IOADDR(A_SCD_PERF_CNT_CFG));
 157         /* Disable and clear counters, override SRC_1 */
 158         bus_writeq((scdperfcnt & ~(M_SPC_CFG_SRC1 | M_SPC_CFG_ENABLE)) |
 159                    M_SPC_CFG_ENABLE |
 160                    M_SPC_CFG_CLEAR |
 161                    V_SPC_CFG_SRC1(1),
 162                    IOADDR(A_SCD_PERF_CNT_CFG));
 163
 164         /* We grab this interrupt to prevent others from trying to use
 165            it, even though we don't want to service the interrupts
 166            (they only feed into the trace-on-interrupt mechanism) */
 167         if (request_irq
 168             (K_INT_PERF_CNT, sbprof_pc_intr, 0, DEVNAME " scd perfcnt", &sbp)) {
 169                 free_irq(K_INT_TRACE_FREEZE, &sbp);
 170                 return -EBUSY;
 171         }
 172
 173         /* I need the core to mask these, but the interrupt mapper to
 174            pass them through.  I am exploiting my knowledge that
 175            cp0_status masks out IP[5]. krw */
 176         bus_writeq(K_INT_MAP_I3,
 177                    IOADDR(A_IMR_REGISTER(0, R_IMR_INTERRUPT_MAP_BASE) +
 178                           (K_INT_PERF_CNT << 3)));
 179
 180         /* Initialize address traps */
 181         bus_writeq(0, IOADDR(A_ADDR_TRAP_UP_0));
 182         bus_writeq(0, IOADDR(A_ADDR_TRAP_UP_1));
 183         bus_writeq(0, IOADDR(A_ADDR_TRAP_UP_2));
 184         bus_writeq(0, IOADDR(A_ADDR_TRAP_UP_3));
 185
 186         bus_writeq(0, IOADDR(A_ADDR_TRAP_DOWN_0));
 187         bus_writeq(0, IOADDR(A_ADDR_TRAP_DOWN_1));
 188         bus_writeq(0, IOADDR(A_ADDR_TRAP_DOWN_2));
 189         bus_writeq(0, IOADDR(A_ADDR_TRAP_DOWN_3));
 190
 191         bus_writeq(0, IOADDR(A_ADDR_TRAP_CFG_0));
 192         bus_writeq(0, IOADDR(A_ADDR_TRAP_CFG_1));
 193         bus_writeq(0, IOADDR(A_ADDR_TRAP_CFG_2));
 194         bus_writeq(0, IOADDR(A_ADDR_TRAP_CFG_3));
 195
 196         /* Initialize Trace Event 0-7 */
 197         //                              when interrupt
 198         bus_writeq(M_SCD_TREVT_INTERRUPT, IOADDR(A_SCD_TRACE_EVENT_0));
 199         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_1));
 200         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_2));
 201         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_3));
 202         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_4));
 203         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_5));
 204         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_6));
 205         bus_writeq(0, IOADDR(A_SCD_TRACE_EVENT_7));
 206
 207         /* Initialize Trace Sequence 0-7 */
 208         //                                   Start on event 0 (interrupt)
 209         bus_writeq(V_SCD_TRSEQ_FUNC_START | 0x0fff,
 210                    IOADDR(A_SCD_TRACE_SEQUENCE_0));
 211         //                        dsamp when d used | asamp when a used
 212         bus_writeq(M_SCD_TRSEQ_ASAMPLE | M_SCD_TRSEQ_DSAMPLE |
 213                    K_SCD_TRSEQ_TRIGGER_ALL,
 214                    IOADDR(A_SCD_TRACE_SEQUENCE_1));
 215         bus_writeq(0, IOADDR(A_SCD_TRACE_SEQUENCE_2));
 216         bus_writeq(0, IOADDR(A_SCD_TRACE_SEQUENCE_3));
 217         bus_writeq(0, IOADDR(A_SCD_TRACE_SEQUENCE_4));
 218         bus_writeq(0, IOADDR(A_SCD_TRACE_SEQUENCE_5));
 219         bus_writeq(0, IOADDR(A_SCD_TRACE_SEQUENCE_6));
 220         bus_writeq(0, IOADDR(A_SCD_TRACE_SEQUENCE_7));
 221
 222         /* Now indicate the PERF_CNT interrupt as a trace-relevant interrupt */
 223         bus_writeq((1ULL << K_INT_PERF_CNT),
 224                    IOADDR(A_IMR_REGISTER(0, R_IMR_INTERRUPT_TRACE)));
 225
 226         arm_tb();
 227
 228         DBG(printk(DEVNAME ": done starting\n"));
 229
 230         return 0;
 231 }
 232
 233 int sbprof_zbprof_stop(void)
 234 {
 235         DEFINE_WAIT(wait);
 236         DBG(printk(DEVNAME ": stopping\n"));
 237
 238         if (sbp.tb_enable) {
 239                 sbp.tb_enable = 0;
 240                 /* XXXKW there is a window here where the intr handler
 241                    may run, see the disable, and do the wake_up before
 242                    this sleep happens. */
 243                 if (sbp.tb_armed) {
 244                         DBG(printk(DEVNAME ": wait for disarm\n"));
 245                         prepare_to_wait(&sbp.tb_sync, &wait, TASK_INTERRUPTIBLE);
 246                         schedule();
 247                         finish_wait(&sbp.tb_sync, &wait);
 248                         DBG(printk(DEVNAME ": disarm complete\n"));
 249                 }
 250                 free_irq(K_INT_TRACE_FREEZE, &sbp);
 251                 free_irq(K_INT_PERF_CNT, &sbp);
 252         }
 253
 254         DBG(printk(DEVNAME ": done stopping\n"));
 255
 256         return 0;
 257 }
 258
 259 static int sbprof_tb_open(struct inode *inode, struct file *filp)
 260 {
 261         int minor;
 262
 263         minor = iminor(inode);
 264         if (minor != 0) {
 265                 return -ENODEV;
 266         }
 267         if (sbp.open) {
 268                 return -EBUSY;
 269         }
 270
 271         memset(&sbp, 0, sizeof(struct sbprof_tb));
 272         sbp.sbprof_tbbuf = vmalloc(MAX_TBSAMPLE_BYTES);
 273         if (!sbp.sbprof_tbbuf) {
 274                 return -ENOMEM;
 275         }
 276         memset(sbp.sbprof_tbbuf, 0, MAX_TBSAMPLE_BYTES);
 277         init_waitqueue_head(&sbp.tb_sync);
 278         init_waitqueue_head(&sbp.tb_read);
 279         sbp.open = 1;
 280
 281         return 0;
 282 }
 283
 284 static int sbprof_tb_release(struct inode *inode, struct file *filp)
 285 {
 286         int minor;
 287
 288         minor = iminor(inode);
 289         if (minor != 0 || !sbp.open) {
 290                 return -ENODEV;
 291         }
 292
 293         if (sbp.tb_armed || sbp.tb_enable) {
 294                 sbprof_zbprof_stop();
 295         }
 296
 297         vfree(sbp.sbprof_tbbuf);
 298         sbp.open = 0;
 299
 300         return 0;
 301 }
 302
 303 static ssize_t sbprof_tb_read(struct file *filp, char *buf,
 304                               size_t size, loff_t *offp)
 305 {
 306         int cur_sample, sample_off, cur_count, sample_left;
 307         char *src;
 308         int   count   =  0;
 309         char *dest    =  buf;
 310         long  cur_off = *offp;
 311
 312         count = 0;
 313         cur_sample = cur_off / TB_SAMPLE_SIZE;
 314         sample_off = cur_off % TB_SAMPLE_SIZE;
 315         sample_left = TB_SAMPLE_SIZE - sample_off;
 316         while (size && (cur_sample < sbp.next_tb_sample)) {
 317                 cur_count = size < sample_left ? size : sample_left;
 318                 src = (char *)(((long)sbp.sbprof_tbbuf[cur_sample])+sample_off);
 319                 copy_to_user(dest, src, cur_count);
 320                 DBG(printk(DEVNAME ": read from sample %d, %d bytes\n",
 321                            cur_sample, cur_count));
 322                 size -= cur_count;
 323                 sample_left -= cur_count;
 324                 if (!sample_left) {
 325                         cur_sample++;
 326                         sample_off = 0;
 327                         sample_left = TB_SAMPLE_SIZE;
 328                 } else {
 329                         sample_off += cur_count;
 330                 }
 331                 cur_off += cur_count;
 332                 dest += cur_count;
 333                 count += cur_count;
 334         }
 335         *offp = cur_off;
 336
 337         return count;
 338 }
 339
 340 static int sbprof_tb_ioctl(struct inode *inode,
 341                            struct file *filp,
 342                            unsigned int command,
 343                            unsigned long arg)
 344 {
 345         int error = 0;
 346
 347         switch (command) {
 348         case SBPROF_ZBSTART:
 349                 error = sbprof_zbprof_start(filp);
 350                 break;
 351         case SBPROF_ZBSTOP:
 352                 error = sbprof_zbprof_stop();
 353                 break;
 354         case SBPROF_ZBWAITFULL:
 355                 DEFINE_WAIT(wait);
 356                 prepare_to_wait(&sbp.tb_read, &wait, TASK_INTERRUPTIBLE);
 357                 schedule();
 358                 finish_wait(&sbp.tb_read, &wait);
 359                 /* XXXKW check if interrupted? */
 360                 return put_user(TB_FULL, (int *) arg);
 361         default:
 362                 error = -EINVAL;
 363                 break;
 364         }
 365
 366         return error;
 367 }
 368
 369 static struct file_operations sbprof_tb_fops = {
 370         .owner          = THIS_MODULE,
 371         .open           = sbprof_tb_open,
 372         .release        = sbprof_tb_release,
 373         .read           = sbprof_tb_read,
 374         .ioctl          = sbprof_tb_ioctl,
 375         .mmap           = NULL,
 376 };
 377
 378 static int __init sbprof_tb_init(void)
 379 {
 380         if (register_chrdev(SBPROF_TB_MAJOR, DEVNAME, &sbprof_tb_fops)) {
 381                 printk(KERN_WARNING DEVNAME ": initialization failed (dev %d)\n",
 382                        SBPROF_TB_MAJOR);
 383                 return -EIO;
 384         }
 385         sbp.open = 0;
 386         tb_period = zbbus_mhz * 10000LL;
 387         printk(KERN_INFO DEVNAME ": initialized - tb_period = %lld\n", tb_period);
 388         return 0;
 389 }
 390
 391 static void __exit sbprof_tb_cleanup(void)
 392 {
 393         unregister_chrdev(SBPROF_TB_MAJOR, DEVNAME);
 394 }
 395
 396 module_init(sbprof_tb_init);
 397 module_exit(sbprof_tb_cleanup);