patches/pthreads.patch

   1 Use POSIX threads in userspace.
   2
   3 Index: zfs+chaos4/cmd/lztest/ztest.c
   4 ===================================================================
   5 --- zfs+chaos4.orig/cmd/lztest/ztest.c
   6 +++ zfs+chaos4/cmd/lztest/ztest.c
   7 @@ -141,7 +141,7 @@ typedef struct ztest_args {
   8         spa_t           *za_spa;
   9         objset_t        *za_os;
  10         zilog_t         *za_zilog;
  11 -       thread_t        za_thread;
  12 +       pthread_t       za_thread;
  13         uint64_t        za_instance;
  14         uint64_t        za_random;
  15         uint64_t        za_diroff;
  16 @@ -224,17 +224,17 @@ ztest_info_t ztest_info[] = {
  17   * Stuff we need to share writably between parent and child.
  18   */
  19  typedef struct ztest_shared {
  20 -       mutex_t         zs_vdev_lock;
  21 -       rwlock_t        zs_name_lock;
  22 -       uint64_t        zs_vdev_primaries;
  23 -       uint64_t        zs_enospc_count;
  24 -       hrtime_t        zs_start_time;
  25 -       hrtime_t        zs_stop_time;
  26 -       uint64_t        zs_alloc;
  27 -       uint64_t        zs_space;
  28 -       ztest_info_t    zs_info[ZTEST_FUNCS];
  29 -       mutex_t         zs_sync_lock[ZTEST_SYNC_LOCKS];
  30 -       uint64_t        zs_seq[ZTEST_SYNC_LOCKS];
  31 +       pthread_mutex_t         zs_vdev_lock;
  32 +       pthread_rwlock_t        zs_name_lock;
  33 +       uint64_t                zs_vdev_primaries;
  34 +       uint64_t                zs_enospc_count;
  35 +       hrtime_t                zs_start_time;
  36 +       hrtime_t                zs_stop_time;
  37 +       uint64_t                zs_alloc;
  38 +       uint64_t                zs_space;
  39 +       ztest_info_t            zs_info[ZTEST_FUNCS];
  40 +       pthread_mutex_t         zs_sync_lock[ZTEST_SYNC_LOCKS];
  41 +       uint64_t                zs_seq[ZTEST_SYNC_LOCKS];
  42  } ztest_shared_t;
  43
  44  static char ztest_dev_template[] = "%s/%s.%llua";
  45 @@ -818,7 +818,7 @@ ztest_spa_create_destroy(ztest_args_t *z
  46          * Attempt to create an existing pool.  It shouldn't matter
  47          * what's in the nvroot; we should fail with EEXIST.
  48          */
  49 -       (void) rw_rdlock(&ztest_shared->zs_name_lock);
  50 +       (void) pthread_rwlock_rdlock(&ztest_shared->zs_name_lock);
  51         nvroot = make_vdev_root(0, 0, 0, 0, 1);
  52         error = spa_create(za->za_pool, nvroot, NULL, NULL);
  53         nvlist_free(nvroot);
  54 @@ -834,7 +834,7 @@ ztest_spa_create_destroy(ztest_args_t *z
  55                 fatal(0, "spa_destroy() = %d", error);
  56
  57         spa_close(spa, FTAG);
  58 -       (void) rw_unlock(&ztest_shared->zs_name_lock);
  59 +       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
  60  }
  61
  62  /*
  63 @@ -851,7 +851,7 @@ ztest_vdev_add_remove(ztest_args_t *za)
  64         if (zopt_verbose >= 6)
  65                 (void) printf("adding vdev\n");
  66
  67 -       (void) mutex_lock(&ztest_shared->zs_vdev_lock);
  68 +       (void) pthread_mutex_lock(&ztest_shared->zs_vdev_lock);
  69
  70         spa_config_enter(spa, RW_READER, FTAG);
  71
  72 @@ -869,7 +869,7 @@ ztest_vdev_add_remove(ztest_args_t *za)
  73         error = spa_vdev_add(spa, nvroot);
  74         nvlist_free(nvroot);
  75
  76 -       (void) mutex_unlock(&ztest_shared->zs_vdev_lock);
  77 +       (void) pthread_mutex_unlock(&ztest_shared->zs_vdev_lock);
  78
  79         if (error == ENOSPC)
  80                 ztest_record_enospc("spa_vdev_add");
  81 @@ -927,7 +927,7 @@ ztest_vdev_attach_detach(ztest_args_t *z
  82         int error, expected_error;
  83         int fd;
  84
  85 -       (void) mutex_lock(&ztest_shared->zs_vdev_lock);
  86 +       (void) pthread_mutex_lock(&ztest_shared->zs_vdev_lock);
  87
  88         spa_config_enter(spa, RW_READER, FTAG);
  89
  90 @@ -1054,7 +1054,7 @@ ztest_vdev_attach_detach(ztest_args_t *z
  91                     oldpath, newpath, replacing, error, expected_error);
  92         }
  93
  94 -       (void) mutex_unlock(&ztest_shared->zs_vdev_lock);
  95 +       (void) pthread_mutex_unlock(&ztest_shared->zs_vdev_lock);
  96  }
  97
  98  /*
  99 @@ -1071,7 +1071,7 @@ ztest_vdev_LUN_growth(ztest_args_t *za)
 100         size_t fsize;
 101         int fd;
 102
 103 -       (void) mutex_lock(&ztest_shared->zs_vdev_lock);
 104 +       (void) pthread_mutex_lock(&ztest_shared->zs_vdev_lock);
 105
 106         /*
 107          * Pick a random leaf vdev.
 108 @@ -1102,7 +1102,7 @@ ztest_vdev_LUN_growth(ztest_args_t *za)
 109                 (void) close(fd);
 110         }
 111
 112 -       (void) mutex_unlock(&ztest_shared->zs_vdev_lock);
 113 +       (void) pthread_mutex_unlock(&ztest_shared->zs_vdev_lock);
 114  }
 115
 116  /* ARGSUSED */
 117 @@ -1198,7 +1198,7 @@ ztest_dmu_objset_create_destroy(ztest_ar
 118         uint64_t objects;
 119         ztest_replay_t zr;
 120
 121 -       (void) rw_rdlock(&ztest_shared->zs_name_lock);
 122 +       (void) pthread_rwlock_rdlock(&ztest_shared->zs_name_lock);
 123         (void) snprintf(name, 100, "%s/%s_temp_%llu", za->za_pool, za->za_pool,
 124             (u_longlong_t)za->za_instance);
 125
 126 @@ -1242,7 +1242,7 @@ ztest_dmu_objset_create_destroy(ztest_ar
 127         if (error) {
 128                 if (error == ENOSPC) {
 129                         ztest_record_enospc("dmu_objset_create");
 130 -                       (void) rw_unlock(&ztest_shared->zs_name_lock);
 131 +                       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 132                         return;
 133                 }
 134                 fatal(0, "dmu_objset_create(%s) = %d", name, error);
 135 @@ -1321,7 +1321,7 @@ ztest_dmu_objset_create_destroy(ztest_ar
 136         if (error)
 137                 fatal(0, "dmu_objset_destroy(%s) = %d", name, error);
 138
 139 -       (void) rw_unlock(&ztest_shared->zs_name_lock);
 140 +       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 141  }
 142
 143  /*
 144 @@ -1335,7 +1335,7 @@ ztest_dmu_snapshot_create_destroy(ztest_
 145         char snapname[100];
 146         char osname[MAXNAMELEN];
 147
 148 -       (void) rw_rdlock(&ztest_shared->zs_name_lock);
 149 +       (void) pthread_rwlock_rdlock(&ztest_shared->zs_name_lock);
 150         dmu_objset_name(os, osname);
 151         (void) snprintf(snapname, 100, "%s@%llu", osname,
 152             (u_longlong_t)za->za_instance);
 153 @@ -1348,7 +1348,7 @@ ztest_dmu_snapshot_create_destroy(ztest_
 154                 ztest_record_enospc("dmu_take_snapshot");
 155         else if (error != 0 && error != EEXIST)
 156                 fatal(0, "dmu_take_snapshot() = %d", error);
 157 -       (void) rw_unlock(&ztest_shared->zs_name_lock);
 158 +       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 159  }
 160
 161  #define        ZTEST_TRAVERSE_BLOCKS   1000
 162 @@ -1992,7 +1992,7 @@ ztest_dmu_write_parallel(ztest_args_t *z
 163         int bs = ZTEST_DIROBJ_BLOCKSIZE;
 164         int do_free = 0;
 165         uint64_t off, txg_how;
 166 -       mutex_t *lp;
 167 +       pthread_mutex_t *lp;
 168         char osname[MAXNAMELEN];
 169         char iobuf[SPA_MAXBLOCKSIZE];
 170         blkptr_t blk = { 0 };
 171 @@ -2041,7 +2041,7 @@ ztest_dmu_write_parallel(ztest_args_t *z
 172         }
 173
 174         lp = &ztest_shared->zs_sync_lock[b];
 175 -       (void) mutex_lock(lp);
 176 +       (void) pthread_mutex_lock(lp);
 177
 178         wbt->bt_objset = dmu_objset_id(os);
 179         wbt->bt_object = ZTEST_DIROBJ;
 180 @@ -2087,7 +2087,7 @@ ztest_dmu_write_parallel(ztest_args_t *z
 181                 dmu_write(os, ZTEST_DIROBJ, off, btsize, wbt, tx);
 182         }
 183
 184 -       (void) mutex_unlock(lp);
 185 +       (void) pthread_mutex_unlock(lp);
 186
 187         if (ztest_random(1000) == 0)
 188                 (void) poll(NULL, 0, 1); /* open dn_notxholds window */
 189 @@ -2106,7 +2106,7 @@ ztest_dmu_write_parallel(ztest_args_t *z
 190         /*
 191          * dmu_sync() the block we just wrote.
 192          */
 193 -       (void) mutex_lock(lp);
 194 +       (void) pthread_mutex_lock(lp);
 195
 196         blkoff = P2ALIGN_TYPED(off, bs, uint64_t);
 197         error = dmu_buf_hold(os, ZTEST_DIROBJ, blkoff, FTAG, &db);
 198 @@ -2114,7 +2114,7 @@ ztest_dmu_write_parallel(ztest_args_t *z
 199         if (error) {
 200                 dprintf("dmu_buf_hold(%s, %d, %llx) = %d\n",
 201                     osname, ZTEST_DIROBJ, blkoff, error);
 202 -               (void) mutex_unlock(lp);
 203 +               (void) pthread_mutex_unlock(lp);
 204                 return;
 205         }
 206         blkoff = off - blkoff;
 207 @@ -2122,7 +2122,7 @@ ztest_dmu_write_parallel(ztest_args_t *z
 208         dmu_buf_rele(db, FTAG);
 209         za->za_dbuf = NULL;
 210
 211 -       (void) mutex_unlock(lp);
 212 +       (void) pthread_mutex_unlock(lp);
 213
 214         if (error) {
 215                 dprintf("dmu_sync(%s, %d, %llx) = %d\n",
 216 @@ -2502,7 +2502,7 @@ ztest_dsl_prop_get_set(ztest_args_t *za)
 217         char osname[MAXNAMELEN];
 218         int error;
 219
 220 -       (void) rw_rdlock(&ztest_shared->zs_name_lock);
 221 +       (void) pthread_rwlock_rdlock(&ztest_shared->zs_name_lock);
 222
 223         dmu_objset_name(os, osname);
 224
 225 @@ -2541,7 +2541,7 @@ ztest_dsl_prop_get_set(ztest_args_t *za)
 226                 }
 227         }
 228
 229 -       (void) rw_unlock(&ztest_shared->zs_name_lock);
 230 +       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 231  }
 232
 233  static void
 234 @@ -2693,7 +2693,7 @@ ztest_spa_rename(ztest_args_t *za)
 235         int error;
 236         spa_t *spa;
 237
 238 -       (void) rw_wrlock(&ztest_shared->zs_name_lock);
 239 +       (void) pthread_rwlock_wrlock(&ztest_shared->zs_name_lock);
 240
 241         oldname = za->za_pool;
 242         newname = umem_alloc(strlen(oldname) + 5, UMEM_NOFAIL);
 243 @@ -2745,7 +2745,7 @@ ztest_spa_rename(ztest_args_t *za)
 244
 245         umem_free(newname, strlen(newname) + 1);
 246
 247 -       (void) rw_unlock(&ztest_shared->zs_name_lock);
 248 +       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 249  }
 250
 251
 252 @@ -3090,13 +3090,13 @@ ztest_run(char *pool)
 253         ztest_args_t *za;
 254         spa_t *spa;
 255         char name[100];
 256 -       thread_t tid;
 257 +       pthread_t tid;
 258
 259 -       (void) _mutex_init(&zs->zs_vdev_lock, USYNC_THREAD, NULL);
 260 -       (void) rwlock_init(&zs->zs_name_lock, USYNC_THREAD, NULL);
 261 +       (void) pthread_mutex_init(&zs->zs_vdev_lock, NULL);
 262 +       (void) pthread_rwlock_init(&zs->zs_name_lock, NULL);
 263
 264         for (t = 0; t < ZTEST_SYNC_LOCKS; t++)
 265 -               (void) _mutex_init(&zs->zs_sync_lock[t], USYNC_THREAD, NULL);
 266 +               (void) pthread_mutex_init(&zs->zs_sync_lock[t], NULL);
 267
 268         /*
 269          * Destroy one disk before we even start.
 270 @@ -3153,7 +3153,7 @@ ztest_run(char *pool)
 271          * start the thread before setting the zio_io_fail_shift, which
 272          * will indicate our failure rate.
 273          */
 274 -       error = thr_create(0, 0, ztest_suspend_monitor, NULL, THR_BOUND, &tid);
 275 +       error = pthread_create(&tid, NULL, ztest_suspend_monitor, NULL);
 276         if (error) {
 277                 fatal(0, "can't create suspend monitor thread: error %d",
 278                     t, error);
 279 @@ -3217,7 +3217,7 @@ ztest_run(char *pool)
 280                 if (t < zopt_datasets) {
 281                         ztest_replay_t zr;
 282                         int test_future = FALSE;
 283 -                       (void) rw_rdlock(&ztest_shared->zs_name_lock);
 284 +                       (void) pthread_rwlock_rdlock(&ztest_shared->zs_name_lock);
 285                         (void) snprintf(name, 100, "%s/%s_%d", pool, pool, d);
 286                         error = dmu_objset_create(name, DMU_OST_OTHER, NULL, 0,
 287                             ztest_create_cb, NULL);
 288 @@ -3225,7 +3225,7 @@ ztest_run(char *pool)
 289                                 test_future = TRUE;
 290                         } else if (error == ENOSPC) {
 291                                 zs->zs_enospc_count++;
 292 -                               (void) rw_unlock(&ztest_shared->zs_name_lock);
 293 +                               (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 294                                 break;
 295                         } else if (error != 0) {
 296                                 fatal(0, "dmu_objset_create(%s) = %d",
 297 @@ -3236,7 +3236,7 @@ ztest_run(char *pool)
 298                         if (error)
 299                                 fatal(0, "dmu_objset_open('%s') = %d",
 300                                     name, error);
 301 -                       (void) rw_unlock(&ztest_shared->zs_name_lock);
 302 +                       (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 303                         if (test_future)
 304                                 ztest_dmu_check_future_leak(&za[t]);
 305                         zr.zr_os = za[d].za_os;
 306 @@ -3245,15 +3245,15 @@ ztest_run(char *pool)
 307                         za[d].za_zilog = zil_open(za[d].za_os, NULL);
 308                 }
 309
 310 -               error = thr_create(0, 0, ztest_thread, &za[t], THR_BOUND,
 311 -                   &za[t].za_thread);
 312 +               error = pthread_create(&za[t].za_thread, NULL, ztest_thread,
 313 +                   &za[t]);
 314                 if (error)
 315                         fatal(0, "can't create thread %d: error %d",
 316                             t, error);
 317         }
 318
 319         while (--t >= 0) {
 320 -               error = thr_join(za[t].za_thread, NULL, NULL);
 321 +               error = pthread_join(za[t].za_thread, NULL);
 322                 if (error)
 323                         fatal(0, "thr_join(%d) = %d", t, error);
 324                 if (za[t].za_th)
 325 @@ -3276,14 +3276,14 @@ ztest_run(char *pool)
 326          * If we had out-of-space errors, destroy a random objset.
 327          */
 328         if (zs->zs_enospc_count != 0) {
 329 -               (void) rw_rdlock(&ztest_shared->zs_name_lock);
 330 +               (void) pthread_rwlock_rdlock(&ztest_shared->zs_name_lock);
 331                 d = (int)ztest_random(zopt_datasets);
 332                 (void) snprintf(name, 100, "%s/%s_%d", pool, pool, d);
 333                 if (zopt_verbose >= 3)
 334                         (void) printf("Destroying %s to free up space\n", name);
 335                 (void) dmu_objset_find(name, ztest_destroy_cb, &za[d],
 336                     DS_FIND_SNAPSHOTS | DS_FIND_CHILDREN);
 337 -               (void) rw_unlock(&ztest_shared->zs_name_lock);
 338 +               (void) pthread_rwlock_unlock(&ztest_shared->zs_name_lock);
 339         }
 340
 341         txg_wait_synced(spa_get_dsl(spa), 0);
 342 @@ -3301,7 +3301,7 @@ ztest_run(char *pool)
 343         mutex_enter(&spa->spa_zio_lock);
 344         cv_broadcast(&spa->spa_zio_cv);
 345         mutex_exit(&spa->spa_zio_lock);
 346 -       error = thr_join(tid, NULL, NULL);
 347 +       error = pthread_join(tid, NULL);
 348         if (error)
 349                 fatal(0, "thr_join(%d) = %d", tid, error);
 350
 351 Index: zfs+chaos4/lib/libuutil/uu_misc.c
 352 ===================================================================
 353 --- zfs+chaos4.orig/lib/libuutil/uu_misc.c
 354 +++ zfs+chaos4/lib/libuutil/uu_misc.c
 355 @@ -37,7 +37,6 @@
 356  #include <stdlib.h>
 357  #include <string.h>
 358  #include <sys/debug.h>
 359 -#include <thread.h>
 360  #include <unistd.h>
 361
 362  #if !defined(TEXT_DOMAIN)
 363 @@ -70,11 +69,12 @@ static va_list              uu_panic_args;
 364  static pthread_t       uu_panic_thread;
 365
 366  static uint32_t                _uu_main_error;
 367 +static __thread int    _uu_main_thread = 0;
 368
 369  void
 370  uu_set_error(uint_t code)
 371  {
 372 -       if (thr_main() != 0) {
 373 +       if (_uu_main_thread) {
 374                 _uu_main_error = code;
 375                 return;
 376         }
 377 @@ -103,7 +103,7 @@ uu_set_error(uint_t code)
 378  uint32_t
 379  uu_error(void)
 380  {
 381 -       if (thr_main() != 0)
 382 +       if (_uu_main_thread)
 383                 return (_uu_main_error);
 384
 385         if (uu_error_key_setup < 0)     /* can't happen? */
 386 @@ -255,5 +255,6 @@ uu_release_child(void)
 387  static void
 388  uu_init(void)
 389  {
 390 +       _uu_main_thread = 1;
 391         (void) pthread_atfork(uu_lockup, uu_release, uu_release_child);
 392  }
 393 Index: zfs+chaos4/lib/libzfscommon/include/sys/zfs_context_user.h
 394 ===================================================================
 395 --- zfs+chaos4.orig/lib/libzfscommon/include/sys/zfs_context_user.h
 396 +++ zfs+chaos4/lib/libzfscommon/include/sys/zfs_context_user.h
 397 @@ -52,8 +52,7 @@ extern "C" {
 398  #include <errno.h>
 399  #include <string.h>
 400  #include <strings.h>
 401 -#include <synch.h>
 402 -#include <thread.h>
 403 +#include <pthread.h>
 404  #include <assert.h>
 405  #include <alloca.h>
 406  #include <umem.h>
 407 @@ -191,13 +190,15 @@ _NOTE(CONSTCOND) } while (0)
 408  /*
 409   * Threads
 410   */
 411 -#define        curthread       ((void *)(uintptr_t)thr_self())
 412 +
 413 +/* XXX: not portable */
 414 +#define        curthread       ((void *)(uintptr_t)pthread_self())
 415
 416  typedef struct kthread kthread_t;
 417
 418  #define        thread_create(stk, stksize, func, arg, len, pp, state, pri)     \
 419         zk_thread_create(func, arg)
 420 -#define        thread_exit() thr_exit(NULL)
 421 +#define        thread_exit() pthread_exit(NULL)
 422
 423  extern kthread_t *zk_thread_create(void (*func)(), void *arg);
 424
 425 @@ -207,28 +208,18 @@ extern kthread_t *zk_thread_create(void
 426  /*
 427   * Mutexes
 428   */
 429 +#define MTX_MAGIC 0x9522f51362a6e326ull
 430  typedef struct kmutex {
 431         void            *m_owner;
 432 -       boolean_t       initialized;
 433 -       mutex_t         m_lock;
 434 +       uint64_t        m_magic;
 435 +       pthread_mutex_t m_lock;
 436  } kmutex_t;
 437
 438 -#define        MUTEX_DEFAULT   USYNC_THREAD
 439 -#undef MUTEX_HELD
 440 -#define        MUTEX_HELD(m) _mutex_held(&(m)->m_lock)
 441 -
 442 -/*
 443 - * Argh -- we have to get cheesy here because the kernel and userland
 444 - * have different signatures for the same routine.
 445 - */
 446 -extern int _mutex_init(mutex_t *mp, int type, void *arg);
 447 -extern int _mutex_destroy(mutex_t *mp);
 448 -
 449 -#define        mutex_init(mp, b, c, d)         zmutex_init((kmutex_t *)(mp))
 450 -#define        mutex_destroy(mp)               zmutex_destroy((kmutex_t *)(mp))
 451 +#define        MUTEX_DEFAULT   0
 452 +#define        MUTEX_HELD(m)   ((m)->m_owner == curthread)
 453
 454 -extern void zmutex_init(kmutex_t *mp);
 455 -extern void zmutex_destroy(kmutex_t *mp);
 456 +extern void mutex_init(kmutex_t *mp, char *name, int type, void *cookie);
 457 +extern void mutex_destroy(kmutex_t *mp);
 458  extern void mutex_enter(kmutex_t *mp);
 459  extern void mutex_exit(kmutex_t *mp);
 460  extern int mutex_tryenter(kmutex_t *mp);
 461 @@ -237,23 +228,24 @@ extern void *mutex_owner(kmutex_t *mp);
 462  /*
 463   * RW locks
 464   */
 465 +#define RW_MAGIC 0x4d31fb123648e78aull
 466  typedef struct krwlock {
 467 -       void            *rw_owner;
 468 -       boolean_t       initialized;
 469 -       rwlock_t        rw_lock;
 470 +       void                    *rw_owner;
 471 +       void                    *rw_wr_owner;
 472 +       uint64_t                rw_magic;
 473 +       pthread_rwlock_t        rw_lock;
 474 +       uint_t                  rw_readers;
 475  } krwlock_t;
 476
 477  typedef int krw_t;
 478
 479  #define        RW_READER       0
 480  #define        RW_WRITER       1
 481 -#define        RW_DEFAULT      USYNC_THREAD
 482 -
 483 -#undef RW_READ_HELD
 484 -#define        RW_READ_HELD(x)         _rw_read_held(&(x)->rw_lock)
 485 +#define        RW_DEFAULT      0
 486
 487 -#undef RW_WRITE_HELD
 488 -#define        RW_WRITE_HELD(x)        _rw_write_held(&(x)->rw_lock)
 489 +#define        RW_READ_HELD(x)         ((x)->rw_readers > 0)
 490 +#define        RW_WRITE_HELD(x)        ((x)->rw_wr_owner == curthread)
 491 +#define        RW_LOCK_HELD(x)         (RW_READ_HELD(x) || RW_WRITE_HELD(x))
 492
 493  extern void rw_init(krwlock_t *rwlp, char *name, int type, void *arg);
 494  extern void rw_destroy(krwlock_t *rwlp);
 495 @@ -271,9 +263,13 @@ extern gid_t *crgetgroups(cred_t *cr);
 496  /*
 497   * Condition variables
 498   */
 499 -typedef cond_t kcondvar_t;
 500 +#define CV_MAGIC 0xd31ea9a83b1b30c4ull
 501 +typedef struct kcondvar {
 502 +       uint64_t cv_magic;
 503 +       pthread_cond_t cv;
 504 +} kcondvar_t;
 505
 506 -#define        CV_DEFAULT      USYNC_THREAD
 507 +#define        CV_DEFAULT      0
 508
 509  extern void cv_init(kcondvar_t *cv, char *name, int type, void *arg);
 510  extern void cv_destroy(kcondvar_t *cv);
 511 @@ -444,7 +440,8 @@ extern void delay(clock_t ticks);
 512  #define        minclsyspri     60
 513  #define        maxclsyspri     99
 514
 515 -#define        CPU_SEQID       (thr_self() & (max_ncpus - 1))
 516 +/* XXX: not portable */
 517 +#define        CPU_SEQID       (pthread_self() & (max_ncpus - 1))
 518
 519  #define        kcred           NULL
 520  #define        CRED()          NULL
 521 Index: zfs+chaos4/lib/libzpool/kernel.c
 522 ===================================================================
 523 --- zfs+chaos4.orig/lib/libzpool/kernel.c
 524 +++ zfs+chaos4/lib/libzpool/kernel.c
 525 @@ -38,6 +38,7 @@
 526  #include <sys/zfs_context.h>
 527  #include <sys/zmod.h>
 528  #include <sys/utsname.h>
 529 +#include <sys/time.h>
 530
 531  /*
 532   * Emulation of kernel services in userland.
 533 @@ -60,11 +61,15 @@ struct utsname utsname = {
 534  kthread_t *
 535  zk_thread_create(void (*func)(), void *arg)
 536  {
 537 -       thread_t tid;
 538 +       pthread_t tid;
 539
 540 -       VERIFY(thr_create(0, 0, (void *(*)(void *))func, arg, THR_DETACHED,
 541 -           &tid) == 0);
 542 +       pthread_attr_t attr;
 543 +       VERIFY(pthread_attr_init(&attr) == 0);
 544 +       VERIFY(pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED) == 0);
 545
 546 +       VERIFY(pthread_create(&tid, &attr, (void *(*)(void *))func, arg) == 0);
 547 +
 548 +       /* XXX: not portable */
 549         return ((void *)(uintptr_t)tid);
 550  }
 551
 552 @@ -97,30 +102,37 @@ kstat_delete(kstat_t *ksp)
 553   * =========================================================================
 554   */
 555  void
 556 -zmutex_init(kmutex_t *mp)
 557 +mutex_init(kmutex_t *mp, char *name, int type, void *cookie)
 558  {
 559 +       ASSERT(type == MUTEX_DEFAULT);
 560 +       ASSERT(cookie == NULL);
 561 +
 562 +#ifdef IM_FEELING_LUCKY
 563 +       ASSERT(mp->m_magic != MTX_MAGIC);
 564 +#endif
 565 +
 566         mp->m_owner = NULL;
 567 -       mp->initialized = B_TRUE;
 568 -       (void) _mutex_init(&mp->m_lock, USYNC_THREAD, NULL);
 569 +       mp->m_magic = MTX_MAGIC;
 570 +       VERIFY3S(pthread_mutex_init(&mp->m_lock, NULL), ==, 0);
 571  }
 572
 573  void
 574 -zmutex_destroy(kmutex_t *mp)
 575 +mutex_destroy(kmutex_t *mp)
 576  {
 577 -       ASSERT(mp->initialized == B_TRUE);
 578 +       ASSERT(mp->m_magic == MTX_MAGIC);
 579         ASSERT(mp->m_owner == NULL);
 580 -       (void) _mutex_destroy(&(mp)->m_lock);
 581 +       VERIFY3S(pthread_mutex_destroy(&(mp)->m_lock), ==, 0);
 582         mp->m_owner = (void *)-1UL;
 583 -       mp->initialized = B_FALSE;
 584 +       mp->m_magic = 0;
 585  }
 586
 587  void
 588  mutex_enter(kmutex_t *mp)
 589  {
 590 -       ASSERT(mp->initialized == B_TRUE);
 591 +       ASSERT(mp->m_magic == MTX_MAGIC);
 592         ASSERT(mp->m_owner != (void *)-1UL);
 593         ASSERT(mp->m_owner != curthread);
 594 -       VERIFY(mutex_lock(&mp->m_lock) == 0);
 595 +       VERIFY3S(pthread_mutex_lock(&mp->m_lock), ==, 0);
 596         ASSERT(mp->m_owner == NULL);
 597         mp->m_owner = curthread;
 598  }
 599 @@ -128,9 +140,9 @@ mutex_enter(kmutex_t *mp)
 600  int
 601  mutex_tryenter(kmutex_t *mp)
 602  {
 603 -       ASSERT(mp->initialized == B_TRUE);
 604 +       ASSERT(mp->m_magic == MTX_MAGIC);
 605         ASSERT(mp->m_owner != (void *)-1UL);
 606 -       if (0 == mutex_trylock(&mp->m_lock)) {
 607 +       if (0 == pthread_mutex_trylock(&mp->m_lock)) {
 608                 ASSERT(mp->m_owner == NULL);
 609                 mp->m_owner = curthread;
 610                 return (1);
 611 @@ -142,16 +154,16 @@ mutex_tryenter(kmutex_t *mp)
 612  void
 613  mutex_exit(kmutex_t *mp)
 614  {
 615 -       ASSERT(mp->initialized == B_TRUE);
 616 +       ASSERT(mp->m_magic == MTX_MAGIC);
 617         ASSERT(mutex_owner(mp) == curthread);
 618         mp->m_owner = NULL;
 619 -       VERIFY(mutex_unlock(&mp->m_lock) == 0);
 620 +       VERIFY3S(pthread_mutex_unlock(&mp->m_lock), ==, 0);
 621  }
 622
 623  void *
 624  mutex_owner(kmutex_t *mp)
 625  {
 626 -       ASSERT(mp->initialized == B_TRUE);
 627 +       ASSERT(mp->m_magic == MTX_MAGIC);
 628         return (mp->m_owner);
 629  }
 630
 631 @@ -164,31 +176,48 @@ mutex_owner(kmutex_t *mp)
 632  void
 633  rw_init(krwlock_t *rwlp, char *name, int type, void *arg)
 634  {
 635 -       rwlock_init(&rwlp->rw_lock, USYNC_THREAD, NULL);
 636 +       ASSERT(type == RW_DEFAULT);
 637 +       ASSERT(arg == NULL);
 638 +
 639 +#ifdef IM_FEELING_LUCKY
 640 +       ASSERT(rwlp->rw_magic != RW_MAGIC);
 641 +#endif
 642 +
 643 +       VERIFY3S(pthread_rwlock_init(&rwlp->rw_lock, NULL), ==, 0);
 644         rwlp->rw_owner = NULL;
 645 -       rwlp->initialized = B_TRUE;
 646 +       rwlp->rw_wr_owner = NULL;
 647 +       rwlp->rw_readers = 0;
 648 +       rwlp->rw_magic = RW_MAGIC;
 649  }
 650
 651  void
 652  rw_destroy(krwlock_t *rwlp)
 653  {
 654 -       rwlock_destroy(&rwlp->rw_lock);
 655 -       rwlp->rw_owner = (void *)-1UL;
 656 -       rwlp->initialized = B_FALSE;
 657 +       ASSERT(rwlp->rw_magic == RW_MAGIC);
 658 +
 659 +       VERIFY3S(pthread_rwlock_destroy(&rwlp->rw_lock), ==, 0);
 660 +       rwlp->rw_magic = 0;
 661  }
 662
 663  void
 664  rw_enter(krwlock_t *rwlp, krw_t rw)
 665  {
 666 -       ASSERT(!RW_LOCK_HELD(rwlp));
 667 -       ASSERT(rwlp->initialized == B_TRUE);
 668 -       ASSERT(rwlp->rw_owner != (void *)-1UL);
 669 +       ASSERT(rwlp->rw_magic == RW_MAGIC);
 670         ASSERT(rwlp->rw_owner != curthread);
 671 +       ASSERT(rwlp->rw_wr_owner != curthread);
 672
 673 -       if (rw == RW_READER)
 674 -               (void) rw_rdlock(&rwlp->rw_lock);
 675 -       else
 676 -               (void) rw_wrlock(&rwlp->rw_lock);
 677 +       if (rw == RW_READER) {
 678 +               VERIFY3S(pthread_rwlock_rdlock(&rwlp->rw_lock), ==, 0);
 679 +               ASSERT(rwlp->rw_wr_owner == NULL);
 680 +
 681 +               atomic_inc_uint(&rwlp->rw_readers);
 682 +       } else {
 683 +               VERIFY3S(pthread_rwlock_wrlock(&rwlp->rw_lock), ==, 0);
 684 +               ASSERT(rwlp->rw_wr_owner == NULL);
 685 +               ASSERT3U(rwlp->rw_readers, ==, 0);
 686 +
 687 +               rwlp->rw_wr_owner = curthread;
 688 +       }
 689
 690         rwlp->rw_owner = curthread;
 691  }
 692 @@ -196,11 +225,16 @@ rw_enter(krwlock_t *rwlp, krw_t rw)
 693  void
 694  rw_exit(krwlock_t *rwlp)
 695  {
 696 -       ASSERT(rwlp->initialized == B_TRUE);
 697 -       ASSERT(rwlp->rw_owner != (void *)-1UL);
 698 +       ASSERT(rwlp->rw_magic == RW_MAGIC);
 699 +       ASSERT(RW_LOCK_HELD(rwlp));
 700 +
 701 +       if (RW_READ_HELD(rwlp))
 702 +               atomic_dec_uint(&rwlp->rw_readers);
 703 +       else
 704 +               rwlp->rw_wr_owner = NULL;
 705
 706         rwlp->rw_owner = NULL;
 707 -       (void) rw_unlock(&rwlp->rw_lock);
 708 +       VERIFY3S(pthread_rwlock_unlock(&rwlp->rw_lock), ==, 0);
 709  }
 710
 711  int
 712 @@ -208,19 +242,29 @@ rw_tryenter(krwlock_t *rwlp, krw_t rw)
 713  {
 714         int rv;
 715
 716 -       ASSERT(rwlp->initialized == B_TRUE);
 717 -       ASSERT(rwlp->rw_owner != (void *)-1UL);
 718 +       ASSERT(rwlp->rw_magic == RW_MAGIC);
 719
 720         if (rw == RW_READER)
 721 -               rv = rw_tryrdlock(&rwlp->rw_lock);
 722 +               rv = pthread_rwlock_tryrdlock(&rwlp->rw_lock);
 723         else
 724 -               rv = rw_trywrlock(&rwlp->rw_lock);
 725 +               rv = pthread_rwlock_trywrlock(&rwlp->rw_lock);
 726
 727         if (rv == 0) {
 728 +               ASSERT(rwlp->rw_wr_owner == NULL);
 729 +
 730 +               if (rw == RW_READER)
 731 +                       atomic_inc_uint(&rwlp->rw_readers);
 732 +               else {
 733 +                       ASSERT3U(rwlp->rw_readers, ==, 0);
 734 +                       rwlp->rw_wr_owner = curthread;
 735 +               }
 736 +
 737                 rwlp->rw_owner = curthread;
 738                 return (1);
 739         }
 740
 741 +       VERIFY3S(rv, ==, EBUSY);
 742 +
 743         return (0);
 744  }
 745
 746 @@ -228,8 +272,7 @@ rw_tryenter(krwlock_t *rwlp, krw_t rw)
 747  int
 748  rw_tryupgrade(krwlock_t *rwlp)
 749  {
 750 -       ASSERT(rwlp->initialized == B_TRUE);
 751 -       ASSERT(rwlp->rw_owner != (void *)-1UL);
 752 +       ASSERT(rwlp->rw_magic == RW_MAGIC);
 753
 754         return (0);
 755  }
 756 @@ -243,22 +286,34 @@ rw_tryupgrade(krwlock_t *rwlp)
 757  void
 758  cv_init(kcondvar_t *cv, char *name, int type, void *arg)
 759  {
 760 -       VERIFY(cond_init(cv, type, NULL) == 0);
 761 +       ASSERT(type == CV_DEFAULT);
 762 +
 763 +#ifdef IM_FEELING_LUCKY
 764 +       ASSERT(cv->cv_magic != CV_MAGIC);
 765 +#endif
 766 +
 767 +       cv->cv_magic = CV_MAGIC;
 768 +
 769 +       VERIFY3S(pthread_cond_init(&cv->cv, NULL), ==, 0);
 770  }
 771
 772  void
 773  cv_destroy(kcondvar_t *cv)
 774  {
 775 -       VERIFY(cond_destroy(cv) == 0);
 776 +       ASSERT(cv->cv_magic == CV_MAGIC);
 777 +       VERIFY3S(pthread_cond_destroy(&cv->cv), ==, 0);
 778 +       cv->cv_magic = 0;
 779  }
 780
 781  void
 782  cv_wait(kcondvar_t *cv, kmutex_t *mp)
 783  {
 784 +       ASSERT(cv->cv_magic == CV_MAGIC);
 785         ASSERT(mutex_owner(mp) == curthread);
 786         mp->m_owner = NULL;
 787 -       int ret = cond_wait(cv, &mp->m_lock);
 788 -       VERIFY(ret == 0 || ret == EINTR);
 789 +       int ret = pthread_cond_wait(&cv->cv, &mp->m_lock);
 790 +       if (ret != 0)
 791 +               VERIFY3S(ret, ==, EINTR);
 792         mp->m_owner = curthread;
 793  }
 794
 795 @@ -266,29 +321,38 @@ clock_t
 796  cv_timedwait(kcondvar_t *cv, kmutex_t *mp, clock_t abstime)
 797  {
 798         int error;
 799 +       struct timeval tv;
 800         timestruc_t ts;
 801         clock_t delta;
 802
 803 +       ASSERT(cv->cv_magic == CV_MAGIC);
 804 +
 805  top:
 806         delta = abstime - lbolt;
 807         if (delta <= 0)
 808                 return (-1);
 809
 810 -       ts.tv_sec = delta / hz;
 811 -       ts.tv_nsec = (delta % hz) * (NANOSEC / hz);
 812 +       VERIFY(gettimeofday(&tv, NULL) == 0);
 813 +
 814 +       ts.tv_sec = tv.tv_sec + delta / hz;
 815 +       ts.tv_nsec = tv.tv_usec * 1000 + (delta % hz) * (NANOSEC / hz);
 816 +       if (ts.tv_nsec >= NANOSEC) {
 817 +               ts.tv_sec++;
 818 +               ts.tv_nsec -= NANOSEC;
 819 +       }
 820
 821         ASSERT(mutex_owner(mp) == curthread);
 822         mp->m_owner = NULL;
 823 -       error = cond_reltimedwait(cv, &mp->m_lock, &ts);
 824 +       error = pthread_cond_timedwait(&cv->cv, &mp->m_lock, &ts);
 825         mp->m_owner = curthread;
 826
 827 -       if (error == ETIME)
 828 +       if (error == ETIMEDOUT)
 829                 return (-1);
 830
 831         if (error == EINTR)
 832                 goto top;
 833
 834 -       ASSERT(error == 0);
 835 +       VERIFY3S(error, ==, 0);
 836
 837         return (1);
 838  }
 839 @@ -296,13 +360,15 @@ top:
 840  void
 841  cv_signal(kcondvar_t *cv)
 842  {
 843 -       VERIFY(cond_signal(cv) == 0);
 844 +       ASSERT(cv->cv_magic == CV_MAGIC);
 845 +       VERIFY3S(pthread_cond_signal(&cv->cv), ==, 0);
 846  }
 847
 848  void
 849  cv_broadcast(kcondvar_t *cv)
 850  {
 851 -       VERIFY(cond_broadcast(cv) == 0);
 852 +       ASSERT(cv->cv_magic == CV_MAGIC);
 853 +       VERIFY3S(pthread_cond_broadcast(&cv->cv), ==, 0);
 854  }
 855
 856  /*
 857 @@ -549,11 +615,11 @@ __dprintf(const char *file, const char *
 858             dprintf_find_string(func)) {
 859                 /* Print out just the function name if requested */
 860                 flockfile(stdout);
 861 -               /* XXX: the following printf may not be portable */
 862 +               /* XXX: the following 2 printfs may not be portable */
 863                 if (dprintf_find_string("pid"))
 864                         (void) printf("%llu ", (u_longlong_t) getpid());
 865                 if (dprintf_find_string("tid"))
 866 -                       (void) printf("%u ", (uint_t) thr_self());
 867 +                       (void) printf("%u ", (uint_t) pthread_self());
 868                 if (dprintf_find_string("cpu"))
 869                         (void) printf("%u ", getcpuid());
 870                 if (dprintf_find_string("time"))
 871 Index: zfs+chaos4/lib/libzpool/taskq.c
 872 ===================================================================
 873 --- zfs+chaos4.orig/lib/libzpool/taskq.c
 874 +++ zfs+chaos4/lib/libzpool/taskq.c
 875 @@ -43,7 +43,7 @@ struct taskq {
 876         krwlock_t       tq_threadlock;
 877         kcondvar_t      tq_dispatch_cv;
 878         kcondvar_t      tq_wait_cv;
 879 -       thread_t        *tq_threadlist;
 880 +       pthread_t       *tq_threadlist;
 881         int             tq_flags;
 882         int             tq_active;
 883         int             tq_nthreads;
 884 @@ -186,7 +186,7 @@ taskq_create(const char *name, int nthre
 885         tq->tq_maxalloc = maxalloc;
 886         tq->tq_task.task_next = &tq->tq_task;
 887         tq->tq_task.task_prev = &tq->tq_task;
 888 -       tq->tq_threadlist = kmem_alloc(nthreads * sizeof (thread_t), KM_SLEEP);
 889 +       tq->tq_threadlist = kmem_alloc(nthreads * sizeof (pthread_t), KM_SLEEP);
 890
 891         if (flags & TASKQ_PREPOPULATE) {
 892                 mutex_enter(&tq->tq_lock);
 893 @@ -196,8 +196,8 @@ taskq_create(const char *name, int nthre
 894         }
 895
 896         for (t = 0; t < nthreads; t++)
 897 -               VERIFY(thr_create(0, 0, taskq_thread,
 898 -                   tq, THR_BOUND, &tq->tq_threadlist[t]) == 0);
 899 +               VERIFY(pthread_create(&tq->tq_threadlist[t],
 900 +                   NULL, taskq_thread, tq) == 0);
 901
 902         return (tq);
 903  }
 904 @@ -227,9 +227,9 @@ taskq_destroy(taskq_t *tq)
 905         mutex_exit(&tq->tq_lock);
 906
 907         for (t = 0; t < nthreads; t++)
 908 -               VERIFY(thr_join(tq->tq_threadlist[t], NULL, NULL) == 0);
 909 +               VERIFY(pthread_join(tq->tq_threadlist[t], NULL) == 0);
 910
 911 -       kmem_free(tq->tq_threadlist, nthreads * sizeof (thread_t));
 912 +       kmem_free(tq->tq_threadlist, nthreads * sizeof (pthread_t));
 913
 914         rw_destroy(&tq->tq_threadlock);
 915         mutex_destroy(&tq->tq_lock);
 916 @@ -248,7 +248,7 @@ taskq_member(taskq_t *tq, void *t)
 917                 return (1);
 918
 919         for (i = 0; i < tq->tq_nthreads; i++)
 920 -               if (tq->tq_threadlist[i] == (thread_t)(uintptr_t)t)
 921 +               if (tq->tq_threadlist[i] == (pthread_t)(uintptr_t)t)
 922                         return (1);
 923
 924         return (0);