linux-2.6-590-chopstix-intern.patch

   1 diff -Nurb linux-2.6.22-580/arch/i386/Kconfig linux-2.6.22-590/arch/i386/Kconfig
   2 --- linux-2.6.22-580/arch/i386/Kconfig  2008-02-27 14:59:40.000000000 -0500
   3 +++ linux-2.6.22-590/arch/i386/Kconfig  2008-02-28 07:55:57.000000000 -0500
   4 @@ -1217,6 +1217,14 @@
   5
   6  source "arch/i386/oprofile/Kconfig"
   7
   8 +config CHOPSTIX
   9 +       bool "Chopstix (PlanetLab)"
  10 +       depends on MODULES && OPROFILE
  11 +       help
  12 +         Chopstix allows you to monitor various events by summarizing them
  13 +         in lossy data structures and transferring these data structures
  14 +         into user space. If in doubt, say "N".
  15 +
  16  config KPROBES
  17         bool "Kprobes (EXPERIMENTAL)"
  18         depends on KALLSYMS && EXPERIMENTAL && MODULES
  19 diff -Nurb linux-2.6.22-580/drivers/oprofile/cpu_buffer.c linux-2.6.22-590/drivers/oprofile/cpu_buffer.c
  20 --- linux-2.6.22-580/drivers/oprofile/cpu_buffer.c      2007-07-08 19:32:17.000000000 -0400
  21 +++ linux-2.6.22-590/drivers/oprofile/cpu_buffer.c      2008-02-28 07:40:16.000000000 -0500
  22 @@ -21,6 +21,7 @@
  23  #include <linux/oprofile.h>
  24  #include <linux/vmalloc.h>
  25  #include <linux/errno.h>
  26 +#include <linux/arrays.h>
  27
  28  #include "event_buffer.h"
  29  #include "cpu_buffer.h"
  30 @@ -143,6 +144,17 @@
  31                 b->head_pos = 0;
  32  }
  33
  34 +#ifdef CONFIG_CHOPSTIX
  35 +
  36 +struct event_spec {
  37 +       unsigned int pc;
  38 +       unsigned long dcookie;
  39 +       unsigned count;
  40 +};
  41 +
  42 +extern void (*rec_event)(void *,unsigned int);
  43 +#endif
  44 +
  45  static inline void
  46  add_sample(struct oprofile_cpu_buffer * cpu_buf,
  47             unsigned long pc, unsigned long event)
  48 @@ -151,6 +163,7 @@
  49         entry->eip = pc;
  50         entry->event = event;
  51         increment_head(cpu_buf);
  52 +
  53  }
  54
  55  static inline void
  56 @@ -237,12 +250,75 @@
  57         oprofile_end_trace(cpu_buf);
  58  }
  59
  60 +#ifdef CONFIG_CHOPSTIX
  61 +
  62 +static int proc_pid_cmdline(struct task_struct *task, char * buffer)
  63 +{
  64 +       int res = 0;
  65 +       unsigned int len;
  66 +       struct mm_struct *mm = get_task_mm(task);
  67 +       if (!mm)
  68 +               goto out;
  69 +       if (!mm->arg_end)
  70 +               goto out_mm;    /* Shh! No looking before we're done */
  71 +
  72 +       len = mm->arg_end - mm->arg_start;
  73 +
  74 +       if (len > PAGE_SIZE)
  75 +               len = PAGE_SIZE;
  76 +
  77 +       res = access_process_vm(task, mm->arg_start, buffer, len, 0);
  78 +
  79 +       // If the nul at the end of args has been overwritten, then
  80 +       // assume application is using setproctitle(3).
  81 +       if (res > 0 && buffer[res-1] != '\0' && len < PAGE_SIZE) {
  82 +               len = strnlen(buffer, res);
  83 +               if (len < res) {
  84 +                   res = len;
  85 +               } else {
  86 +                       len = mm->env_end - mm->env_start;
  87 +                       if (len > PAGE_SIZE - res)
  88 +                               len = PAGE_SIZE - res;
  89 +                       res += access_process_vm(task, mm->env_start, buffer+res, len, 0);
  90 +                       res = strnlen(buffer, res);
  91 +               }
  92 +       }
  93 +out_mm:
  94 +       mmput(mm);
  95 +out:
  96 +       return res;
  97 +}
  98 +#endif
  99 +
 100 +
 101 +
 102  void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 103  {
 104         int is_kernel = !user_mode(regs);
 105         unsigned long pc = profile_pc(regs);
 106 +       int res=0;
 107
 108 +#ifdef CONFIG_CHOPSTIX
 109 +       if (rec_event) {
 110 +               struct event esig;
 111 +               struct event_spec espec;
 112 +               /*res = proc_pid_cmdline(current, espec->appname);*/
 113 +               esig.task = current;
 114 +               espec.pc=pc;
 115 +               espec.count=1;
 116 +               esig.event_data=&espec;
 117 +               esig.event_type=event; /* index in the event array currently set up */
 118 +                                       /* make sure the counters are loaded in the order we want them to show up*/
 119 +               (*rec_event)(&esig, 1);
 120 +       }
 121 +       else {
 122         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 123 +       }
 124 +#else
 125 +       oprofile_add_ext_sample(pc, regs, event, is_kernel);
 126 +#endif
 127 +
 128 +
 129  }
 130
 131  void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 132 diff -Nurb linux-2.6.22-580/fs/exec.c linux-2.6.22-590/fs/exec.c
 133 --- linux-2.6.22-580/fs/exec.c  2008-02-27 14:59:40.000000000 -0500
 134 +++ linux-2.6.22-590/fs/exec.c  2008-02-28 07:40:48.000000000 -0500
 135 @@ -52,6 +52,7 @@
 136  #include <linux/audit.h>
 137  #include <linux/signalfd.h>
 138  #include <linux/vs_memory.h>
 139 +#include <linux/dcookies.h>
 140
 141  #include <asm/uaccess.h>
 142  #include <asm/mmu_context.h>
 143 @@ -488,6 +489,12 @@
 144
 145         if (!err) {
 146                 struct inode *inode = nd.dentry->d_inode;
 147 +#ifdef CONFIG_CHOPSTIX
 148 +               unsigned long cookie;
 149 +               if (!nd.dentry->d_cookie)
 150 +                       get_dcookie(nd.dentry, nd.mnt, &cookie);
 151 +#endif
 152 +
 153                 file = ERR_PTR(-EACCES);
 154                 if (!(nd.mnt->mnt_flags & MNT_NOEXEC) &&
 155                     S_ISREG(inode->i_mode)) {
 156 diff -Nurb linux-2.6.22-580/include/linux/arrays.h linux-2.6.22-590/include/linux/arrays.h
 157 --- linux-2.6.22-580/include/linux/arrays.h     1969-12-31 19:00:00.000000000 -0500
 158 +++ linux-2.6.22-590/include/linux/arrays.h     2008-02-27 14:59:52.000000000 -0500
 159 @@ -0,0 +1,35 @@
 160 +#ifndef __ARRAYS_H__
 161 +#define __ARRAYS_H__
 162 +#include <linux/list.h>
 163 +
 164 +#define SAMPLING_METHOD_DEFAULT 0
 165 +#define SAMPLING_METHOD_LOG 1
 166 +
 167 +/* Every probe has an array handler */
 168 +
 169 +/* XXX - Optimize this structure */
 170 +
 171 +struct array_handler {
 172 +       struct list_head link;
 173 +       unsigned int (*hash_func)(void *);
 174 +       unsigned int (*sampling_func)(void *,int,void *);
 175 +       unsigned short size;
 176 +       unsigned int threshold;
 177 +       unsigned char **expcount;
 178 +       unsigned int sampling_method;
 179 +       unsigned int **arrays;
 180 +       unsigned int arraysize;
 181 +       unsigned int num_samples[2];
 182 +       void **epoch_samples; /* size-sized lists of samples */
 183 +       unsigned int (*serialize)(void *, void *);
 184 +       unsigned char code[5];
 185 +};
 186 +
 187 +struct event {
 188 +       struct list_head link;
 189 +       void *event_data;
 190 +       unsigned int count;
 191 +       unsigned int event_type;
 192 +       struct task_struct *task;
 193 +};
 194 +#endif
 195 diff -Nurb linux-2.6.22-580/include/linux/sched.h linux-2.6.22-590/include/linux/sched.h
 196 --- linux-2.6.22-580/include/linux/sched.h      2008-02-27 14:59:41.000000000 -0500
 197 +++ linux-2.6.22-590/include/linux/sched.h      2008-02-28 07:41:28.000000000 -0500
 198 @@ -849,7 +849,11 @@
 199         unsigned int btrace_seq;
 200  #endif
 201         unsigned long sleep_avg;
 202 +#ifdef CONFIG_CHOPSTIX
 203 +       unsigned long long timestamp, last_ran, last_interrupted, last_ran_j;
 204 +#else
 205         unsigned long long timestamp, last_ran;
 206 +#endif
 207         unsigned long long sched_time; /* sched_clock time spent running */
 208         enum sleep_type sleep_type;
 209
 210 diff -Nurb linux-2.6.22-580/kernel/fork.c linux-2.6.22-590/kernel/fork.c
 211 --- linux-2.6.22-580/kernel/fork.c      2008-02-27 14:59:41.000000000 -0500
 212 +++ linux-2.6.22-590/kernel/fork.c      2008-02-28 07:42:09.000000000 -0500
 213 @@ -197,6 +197,11 @@
 214         tsk->btrace_seq = 0;
 215  #endif
 216         tsk->splice_pipe = NULL;
 217 +       //tsk->cmdline[0]='\0';
 218 +#ifdef CONFIG_CHOPSTIX
 219 +       tsk->last_interrupted = 0;
 220 +       tsk->last_ran_j = 0;
 221 +#endif
 222         return tsk;
 223  }
 224
 225 diff -Nurb linux-2.6.22-580/kernel/sched.c linux-2.6.22-590/kernel/sched.c
 226 --- linux-2.6.22-580/kernel/sched.c     2008-02-27 14:59:41.000000000 -0500
 227 +++ linux-2.6.22-590/kernel/sched.c     2008-02-28 07:45:45.000000000 -0500
 228 @@ -10,7 +10,7 @@
 229   *  1998-11-19 Implemented schedule_timeout() and related stuff
 230   *             by Andrea Arcangeli
 231   *  2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
 232 - *             hybrid priority-list and round-robin design with
 233 + *             hybrid priority-list and round-robin deventn with
 234   *             an array-switch method of distributing timeslices
 235   *             and per-CPU runqueues.  Cleanups and useful suggestions
 236   *             by Davide Libenzi, preemptible kernel bits by Robert Love.
 237 @@ -56,6 +56,7 @@
 238
 239  #include <asm/tlb.h>
 240  #include <asm/unistd.h>
 241 +#include <linux/arrays.h>
 242  #include <linux/vs_sched.h>
 243  #include <linux/vs_cvirt.h>
 244
 245 @@ -3608,6 +3609,7 @@
 246
 247  #endif
 248
 249 +
 250  static inline int interactive_sleep(enum sleep_type sleep_type)
 251  {
 252         return (sleep_type == SLEEP_INTERACTIVE ||
 253 @@ -3617,16 +3619,54 @@
 254  /*
 255   * schedule() is the main scheduler function.
 256   */
 257 +
 258 +#ifdef CONFIG_CHOPSTIX
 259 +extern void (*rec_event)(void *,unsigned int);
 260 +struct event_spec {
 261 +       unsigned long pc;
 262 +       unsigned long dcookie;
 263 +       unsigned count;
 264 +       unsigned char reason;
 265 +};
 266 +
 267 +#define top_esp                (THREAD_SIZE - sizeof(unsigned long))
 268 +#define top_ebp                (THREAD_SIZE - 2*sizeof(unsigned long))
 269 +
 270 +static inline unsigned long my_get_wchan(struct task_struct *p)
 271 +{
 272 +        unsigned long ebp, esp, eip;
 273 +        unsigned long stack_page;
 274 +        int count = 0;
 275 +        stack_page = (unsigned long)task_stack_page(p);
 276 +        esp = p->thread.esp;
 277 +        if (!stack_page || esp < stack_page || esp > top_esp+stack_page)
 278 +                return 0;
 279 +        /* include/asm-i386/system.h:switch_to() pushes ebp last. */
 280 +        ebp = *(unsigned long *) esp;
 281 +        do {
 282 +                if (ebp < stack_page || ebp > top_ebp+stack_page)
 283 +                        return 0;
 284 +                eip = *(unsigned long *) (ebp+4);
 285 +                if (!in_sched_functions(eip))
 286 +                        return eip;
 287 +                ebp = *(unsigned long *) ebp;
 288 +        } while (count++ < 16);
 289 +        return 0;
 290 +}
 291 +/* CHOPSTIX */
 292 +#endif
 293 +
 294  asmlinkage void __sched schedule(void)
 295  {
 296         struct task_struct *prev, *next;
 297         struct prio_array *array;
 298         struct list_head *queue;
 299         unsigned long long now;
 300 -       unsigned long run_time;
 301 +       unsigned long run_time, diff;
 302         int cpu, idx, new_prio;
 303         long *switch_count;
 304         struct rq *rq;
 305 +       int sampling_reason;
 306
 307         /*
 308          * Test if we are atomic.  Since do_exit() needs to call into
 309 @@ -3680,6 +3720,7 @@
 310         switch_count = &prev->nivcsw;
 311         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
 312                 switch_count = &prev->nvcsw;
 313 +
 314                 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
 315                                 unlikely(signal_pending(prev))))
 316                         prev->state = TASK_RUNNING;
 317 @@ -3689,6 +3730,11 @@
 318                                 vx_uninterruptible_inc(prev);
 319                         }
 320                         deactivate_task(prev, rq);
 321 +#ifdef CONFIG_CHOPSTIX
 322 +                       if (prev->state & TASK_INTERRUPTIBLE) {
 323 +                               prev->last_interrupted=jiffies;
 324 +                       }
 325 +#endif
 326                 }
 327         }
 328
 329 @@ -3763,8 +3809,45 @@
 330         prev->sleep_avg -= run_time;
 331         if ((long)prev->sleep_avg <= 0)
 332                 prev->sleep_avg = 0;
 333 +
 334         prev->timestamp = prev->last_ran = now;
 335 +#ifdef CONFIG_CHOPSTIX
 336
 337 +       /* CHOPSTIX */
 338 +
 339 +       prev->last_ran_j = jiffies;
 340 +       if (next->last_interrupted) {
 341 +               diff = (jiffies-next->last_interrupted);
 342 +               next->last_interrupted = 0;
 343 +               sampling_reason = 0;
 344 +       }
 345 +       else {
 346 +               diff = jiffies-next->last_ran_j;
 347 +               sampling_reason = 1;
 348 +       }
 349 +
 350 +       if (rec_event && (diff>HZ/5)) {
 351 +               struct event event;
 352 +               struct event_spec espec;
 353 +               unsigned long eip;
 354 +               unsigned int state = next->state;
 355 +
 356 +               espec.reason = sampling_reason;
 357 +
 358 +               next->state = 0;
 359 +               eip = next->thread.esp;
 360 +               next->state = state;
 361 +
 362 +               next->last_interrupted = 0;
 363 +               event.event_data=&espec;
 364 +               event.task=next;
 365 +               espec.pc=eip;
 366 +               event.event_type=2;
 367 +               /* index in the event array currently set up */
 368 +               /* make sure the counters are loaded in the order we want them to show up*/
 369 +               (*rec_event)(&event, diff);
 370 +       }
 371 +#endif
 372         sched_info_switch(prev, next);
 373         if (likely(prev != next)) {
 374                 next->timestamp = next->last_ran = now;
 375 @@ -7275,3 +7358,9 @@
 376  }
 377
 378  #endif
 379 +
 380 +#ifdef CONFIG_CHOPSTIX
 381 +void (*rec_event)(void *,unsigned int);
 382 +EXPORT_SYMBOL(rec_event);
 383 +EXPORT_SYMBOL(in_sched_functions);
 384 +#endif