child.c

   1 /*
   2  * Each process that gets forked runs this code.
   3  */
   4
   5 #include <fcntl.h>
   6 #include <errno.h>
   7 #include <signal.h>
   8 #include <stdlib.h>
   9 #include <string.h>
  10 #include <unistd.h>
  11 #include <sched.h>
  12 #include <sys/time.h>
  13 #include <sys/resource.h>
  14 #include <sys/prctl.h>
  15
  16 #include "arch.h"
  17 #include "child.h"
  18 #include "list.h"
  19 #include "log.h"
  20 #include "maps.h"
  21 #include "params.h"     // for 'debug'
  22 #include "pids.h"
  23 #include "random.h"
  24 #include "shm.h"
  25 #include "signals.h"
  26 #include "syscall.h"
  27 #include "tables.h"
  28 #include "trinity.h"    // ARRAY_SIZE
  29 #include "utils.h"      // zmalloc
  30
  31 static void disable_coredumps(void)
  32 {
  33         struct rlimit limit = { .rlim_cur = 0, .rlim_max = 0 };
  34
  35         if (debug == TRUE) {
  36                 (void)signal(SIGABRT, SIG_DFL);
  37                 (void)signal(SIGSEGV, SIG_DFL);
  38                 return;
  39         }
  40
  41         if (setrlimit(RLIMIT_CORE, &limit) != 0)
  42                 perror( "setrlimit(RLIMIT_CORE)" );
  43
  44         prctl(PR_SET_DUMPABLE, FALSE);
  45 }
  46
  47 static void enable_coredumps(void)
  48 {
  49         struct rlimit limit = {
  50                 .rlim_cur = RLIM_INFINITY,
  51                 .rlim_max = RLIM_INFINITY
  52         };
  53
  54         if (debug == TRUE)
  55                 return;
  56
  57         prctl(PR_SET_DUMPABLE, TRUE);
  58
  59         (void) setrlimit(RLIMIT_CORE, &limit);
  60 }
  61 static void set_make_it_fail(void)
  62 {
  63         int fd;
  64         const char *buf = "1";
  65
  66         /* If we failed last time, don't bother trying in future. */
  67         if (shm->do_make_it_fail == TRUE)
  68                 return;
  69
  70         fd = open("/proc/self/make-it-fail", O_WRONLY);
  71         if (fd == -1)
  72                 return;
  73
  74         if (write(fd, buf, 1) == -1) {
  75                 if (errno != EPERM)
  76                         outputerr("writing to /proc/self/make-it-fail failed! (%s)\n", strerror(errno));
  77                 else
  78                         shm->do_make_it_fail = TRUE;
  79         }
  80         close(fd);
  81 }
  82
  83 /*
  84  * We call this occasionally to set some FPU state, in the hopes that we
  85  * might tickle some weird FPU/scheduler related bugs
  86  */
  87 static void use_fpu(void)
  88 {
  89         double x = 0;
  90         asm volatile("":"+m" (x));
  91         x += 1;
  92         asm volatile("":"+m" (x));
  93 }
  94
  95 int this_child = 0;
  96
  97 static void setup_page_maps(void)
  98 {
  99         unsigned long *page;
 100         unsigned int i;
 101
 102         page = (void *) page_maps;
 103
 104         for (i = 0; i < page_size / sizeof(unsigned long); i++) {
 105                 struct map *map;
 106
 107                 map = get_map();
 108                 page[i] = (unsigned long) map->ptr;
 109         }
 110 }
 111
 112 static void oom_score_adj(int adj)
 113 {
 114         FILE *fp;
 115
 116         fp = fopen("/proc/self/oom_score_adj", "w");
 117         if (!fp)
 118                 return;
 119
 120         fprintf(fp, "%d", adj);
 121         fclose(fp);
 122 }
 123
 124 void init_child(int childno)
 125 {
 126         cpu_set_t set;
 127         pid_t pid = getpid();
 128         char childname[17];
 129
 130         this_child = childno;
 131
 132         set_seed(childno);
 133
 134         shm->kill_count[childno] = 0;
 135
 136         shm->num_mappings[childno] = 0;
 137         shm->mappings[childno] = zmalloc(sizeof(struct map));
 138         INIT_LIST_HEAD(&shm->mappings[childno]->list);
 139
 140         setup_page_maps();
 141
 142         if (sched_getaffinity(pid, sizeof(set), &set) == 0) {
 143                 CPU_ZERO(&set);
 144                 CPU_SET(childno, &set);
 145                 sched_setaffinity(pid, sizeof(set), &set);
 146         }
 147
 148         shm->child_syscall_count[childno] = 0;
 149
 150         memset(childname, 0, sizeof(childname));
 151         sprintf(childname, "trinity-c%d", childno);
 152         prctl(PR_SET_NAME, (unsigned long) &childname);
 153
 154         oom_score_adj(500);
 155
 156         /* Wait for parent to set our pidslot */
 157         while (shm->pids[childno] != getpid()) {
 158                 int ret = 0;
 159
 160                 /* Make sure parent is actually alive to wait for us. */
 161                 ret = pid_alive(shm->mainpid);
 162                 if (ret != 0) {
 163                         shm->exit_reason = EXIT_SHM_CORRUPTION;
 164                         outputerr(BUGTXT "parent (%d) went away!\n", shm->mainpid);
 165                         sleep(20000);
 166                 }
 167         }
 168
 169         /* Wait for all the children to start up. */
 170         while (shm->ready == FALSE)
 171                 sleep(1);
 172
 173         set_make_it_fail();
 174
 175         if (rand() % 100 < 50)
 176                 use_fpu();
 177
 178         mask_signals_child();
 179
 180         disable_coredumps();
 181 }
 182
 183 void check_parent_pid(void)
 184 {
 185         pid_t pid;
 186         unsigned int i;
 187         static unsigned int parent_check_time = 10;
 188
 189         parent_check_time--;
 190         if (parent_check_time != 0)
 191                 return;
 192
 193         parent_check_time = 10;
 194
 195         if (getppid() == shm->mainpid)
 196                 return;
 197
 198         pid = getpid();
 199
 200         //FIXME: Add locking so only one child does this output.
 201         output(0, BUGTXT "CHILD (pid:%d) GOT REPARENTED! "
 202                 "parent pid:%d. Watchdog pid:%d\n",
 203                 pid, shm->mainpid, watchdog_pid);
 204         output(0, BUGTXT "Last syscalls:\n");
 205
 206         for_each_pidslot(i) {
 207                 // Skip over 'boring' entries.
 208                 if ((shm->pids[i] == EMPTY_PIDSLOT) &&
 209                     (shm->previous[i].nr == 0) &&
 210                     (shm->child_syscall_count[i] == 0))
 211                         continue;
 212
 213                 output(0, "[%d]  pid:%d call:%s callno:%d\n",
 214                         i, shm->pids[i],
 215                         print_syscall_name(shm->previous[i].nr, shm->previous[i].do32bit),
 216                         shm->child_syscall_count[i]);
 217         }
 218         shm->exit_reason = EXIT_REPARENT_PROBLEM;
 219         exit(EXIT_FAILURE);
 220         //TODO: Emergency logging.
 221 }
 222
 223 struct child_funcs {
 224         const char *name;
 225         int (*func)(int childno);
 226         unsigned char likelyhood;
 227 };
 228
 229 static const struct child_funcs child_ops[] = {
 230         { .name = "rand_syscalls", .func = child_random_syscalls, .likelyhood = 100 },
 231 };
 232
 233
 234 // FIXME: when we have different child ops, we're going to need to redo the progress detector.
 235 static unsigned int handle_sigreturn(int childno)
 236 {
 237         static unsigned int count = 0;
 238         static unsigned int last = -1;
 239
 240         output(2, "<timed out>\n");     /* Flush out the previous syscall output. */
 241
 242         /* Check if we're making any progress at all. */
 243         if (shm->child_syscall_count[childno] == last) {
 244                 count++;
 245                 //output(1, "no progress for %d tries.\n", count);
 246         } else {
 247                 count = 0;
 248                 last = shm->child_syscall_count[childno];
 249         }
 250         if (count == 3) {
 251                 output(1, "no progress for 3 tries, exiting child.\n");
 252                 return 0;
 253         }
 254
 255         if (shm->kill_count[childno] > 0) {
 256                 output(1, "[%d] Missed a kill signal, exiting\n", getpid());
 257                 return 0;
 258         }
 259
 260         if (sigwas != SIGALRM)
 261                 output(1, "[%d] Back from signal handler! (sig was %s)\n", getpid(), strsignal(sigwas));
 262
 263         return 1;
 264 }
 265
 266 void child_process(int childno)
 267 {
 268         int ret;
 269         unsigned int i;
 270         const char *lastop = NULL;
 271
 272         ret = sigsetjmp(ret_jump, 1);
 273         if (ret != 0) {
 274                 if (handle_sigreturn(childno) == 0)
 275                         return; // Exit the child, things are getting too weird.
 276         }
 277
 278         while (shm->exit_reason == STILL_RUNNING) {
 279
 280                 check_parent_pid();
 281
 282                 while (shm->regenerating == TRUE)
 283                         sleep(1);
 284
 285                 /* If the parent reseeded, we should reflect the latest seed too. */
 286                 if (shm->seed != shm->seeds[childno])
 287                         set_seed(childno);
 288
 289                 /* Choose operations for this iteration. */
 290                 i = rand() % ARRAY_SIZE(child_ops);
 291
 292                 if (rand() % 100 <= child_ops[i].likelyhood) {
 293                         if (lastop != child_ops[i].name) {
 294                                 output(0, "Chose %s.\n", child_ops[i].name);
 295                                 lastop = child_ops[i].name;
 296                         }
 297
 298                         ret = child_ops[i].func(childno);       // Do we care about the return code ? Right now, no.
 299                 }
 300         }
 301
 302         enable_coredumps();
 303 }