l4/pkg/uclibc/lib/libpthread/src/sysdeps/x86_64/tls.h

   1 /* Definition for thread-local data handling.  nptl/x86_64 version.
   2    Copyright (C) 2002-2007, 2008, 2009 Free Software Foundation, Inc.
   3    This file is part of the GNU C Library.
   4
   5    The GNU C Library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 2.1 of the License, or (at your option) any later version.
   9
  10    The GNU C Library is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with the GNU C Library; if not, write to the Free
  17    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  18    02111-1307 USA.  */
  19
  20 #ifndef _TLS_H
  21 #define _TLS_H  1
  22
  23 #ifndef __ASSEMBLER__
  24 //l4/# include <asm/prctl.h>    /* For ARCH_SET_FS.  */
  25 # include <stdbool.h>
  26 # include <stddef.h>
  27 # include <stdint.h>
  28 # include <stdlib.h>
  29 //l4/# include <sysdep.h>
  30 # include <bits/kernel-features.h>
  31 # include <bits/wordsize.h>
  32 # include <xmmintrin.h>
  33
  34
  35 /* Type for the dtv.  */
  36 typedef union dtv
  37 {
  38   size_t counter;
  39   struct
  40   {
  41     void *val;
  42     bool is_static;
  43   } pointer;
  44 } dtv_t;
  45
  46
  47 typedef struct
  48 {
  49   void *tcb;            /* Pointer to the TCB.  Not necessarily the
  50                            thread descriptor used by libpthread.  */
  51   dtv_t *dtv;
  52   void *self;           /* Pointer to the thread descriptor.  */
  53   int multiple_threads;
  54   int gscope_flag;
  55   uintptr_t sysinfo;
  56   uintptr_t stack_guard;
  57   uintptr_t pointer_guard;
  58   unsigned long int vgetcpu_cache[2];
  59 # ifndef __ASSUME_PRIVATE_FUTEX
  60   int private_futex;
  61 # else
  62   int __unused1;
  63 # endif
  64 # if __WORDSIZE == 64
  65   int rtld_must_xmm_save;
  66 # endif
  67   /* Reservation of some values for the TM ABI.  */
  68   void *__private_tm[5];
  69 # if __WORDSIZE == 64
  70   long int __unused2;
  71   /* Have space for the post-AVX register size.  */
  72   __m128 rtld_savespace_sse[8][4];
  73
  74   void *__padding[8];
  75 # endif
  76 } tcbhead_t;
  77
  78 #else /* __ASSEMBLER__ */
  79 # include <tcb-offsets.h>
  80 #endif
  81
  82
  83 /* We require TLS support in the tools.  */
  84 #define HAVE_TLS_SUPPORT                1
  85 #define HAVE___THREAD 1
  86 #define HAVE_TLS_MODEL_ATTRIBUTE 1
  87
  88 /* Signal that TLS support is available.  */
  89 #define USE_TLS        1
  90
  91 /* Alignment requirement for the stack.  */
  92 #define STACK_ALIGN     16
  93
  94
  95 #ifndef __ASSEMBLER__
  96 /* Get system call information.  */
  97 //l4/# include <sysdep.h>
  98
  99
 100 /* Get the thread descriptor definition.  */
 101 # include <descr.h>
 102 # include <l4/sys/utcb.h>
 103 # include <l4/sys/segment.h>
 104
 105 #ifndef LOCK_PREFIX
 106 # ifdef UP
 107 #  define LOCK_PREFIX   /* nothing */
 108 # else
 109 #  define LOCK_PREFIX   "lock;"
 110 # endif
 111 #endif
 112
 113 /* This is the size of the initial TCB.  Can't be just sizeof (tcbhead_t),
 114    because NPTL getpid, __libc_alloca_cutoff etc. need (almost) the whole
 115    struct pthread even when not linked with -lpthread.  */
 116 # define TLS_INIT_TCB_SIZE sizeof (struct pthread)
 117
 118 /* Alignment requirements for the initial TCB.  */
 119 # define TLS_INIT_TCB_ALIGN __alignof__ (struct pthread)
 120
 121 /* This is the size of the TCB.  */
 122 # define TLS_TCB_SIZE sizeof (struct pthread)
 123
 124 /* Alignment requirements for the TCB.  */
 125 //# define TLS_TCB_ALIGN __alignof__ (struct pthread)
 126 // Normally the above would be correct  But we have to store post-AVX
 127 // vector registers in the TCB and we want the storage to be aligned.
 128 // unfortunately there isn't yet a type for these values and hence no
 129 // 32-byte alignment requirement.  Make this explicit, for now.
 130 # define TLS_TCB_ALIGN 32
 131
 132 /* The TCB can have any size and the memory following the address the
 133    thread pointer points to is unspecified.  Allocate the TCB there.  */
 134 # define TLS_TCB_AT_TP  1
 135
 136
 137 /* Install the dtv pointer.  The pointer passed is to the element with
 138    index -1 which contain the length.  */
 139 # define INSTALL_DTV(descr, dtvp) \
 140   ((tcbhead_t *) (descr))->dtv = (dtvp) + 1
 141
 142 /* Install new dtv for current thread.  */
 143 # define INSTALL_NEW_DTV(dtvp) \
 144   ({ struct pthread *__pd;                                                    \
 145      THREAD_SETMEM (__pd, header.dtv, (dtvp)); })
 146
 147 /* Return dtv of given thread descriptor.  */
 148 # define GET_DTV(descr) \
 149   (((tcbhead_t *) (descr))->dtv)
 150
 151
 152 /* Macros to load from and store into segment registers.  */
 153 # define TLS_GET_FS() \
 154   ({ int __seg; __asm__ ("movl %%fs, %0" : "=q" (__seg)); __seg; })
 155 # define TLS_SET_FS(val) \
 156   __asm__ ("movl %0, %%fs" :: "q" (val))
 157
 158
 159 static inline char const *TLS_INIT_TP(void *thrdescr, int secondcall)
 160 {
 161   (void)secondcall;
 162   tcbhead_t *_head = (tcbhead_t *)thrdescr;
 163   _head->tcb = thrdescr;
 164   _head->self = thrdescr;
 165   if (fiasco_amd64_set_fs(L4_INVALID_CAP, (l4_umword_t)thrdescr, l4_utcb()) < 0)
 166     return "ERROR";
 167
 168   return NULL;
 169 }
 170 #if 0
 171 /* Code to initially initialize the thread pointer.  This might need
 172    special attention since 'errno' is not yet available and if the
 173    operation can cause a failure 'errno' must not be touched.
 174
 175    We have to make the syscall for both uses of the macro since the
 176    address might be (and probably is) different.  */
 177 # define TLS_INIT_TP(thrdescr, secondcall) \
 178   ({ void *_thrdescr = (thrdescr);                                            \
 179      tcbhead_t *_head = _thrdescr;                                            \
 180      int _result;                                                             \
 181                                                                               \
 182      _head->tcb = _thrdescr;                                                  \
 183      /* For now the thread descriptor is at the same address.  */             \
 184      _head->self = _thrdescr;                                                 \
 185                                                                               \
 186      /* It is a simple syscall to set the %fs value for the thread.  */       \
 187      __asm__ __volatile__ ("syscall"                                                  \
 188                    : "=a" (_result)                                           \
 189                    : "0" ((unsigned long int) __NR_arch_prctl),               \
 190                      "D" ((unsigned long int) ARCH_SET_FS),                   \
 191                      "S" (_thrdescr)                                          \
 192                    : "memory", "cc", "r11", "cx");                            \
 193                                                                               \
 194     _result ? "cannot set %fs base address for thread-local storage" : 0;     \
 195   })
 196 #endif
 197
 198
 199 /* Return the address of the dtv for the current thread.  */
 200 # define THREAD_DTV() \
 201   ({ struct pthread *__pd;                                                    \
 202      THREAD_GETMEM (__pd, header.dtv); })
 203
 204
 205 /* Return the thread descriptor for the current thread.
 206
 207    The contained asm must *not* be marked __volatile__ since otherwise
 208    assignments like
 209         pthread_descr self = thread_self();
 210    do not get optimized away.  */
 211 # define THREAD_SELF \
 212   ({ struct pthread *__self;                                                  \
 213      __asm__ ("movq %%fs:%c1,%q0" : "=r" (__self)                                     \
 214           : "i" (offsetof (struct pthread, header.self)));                    \
 215      __self;})
 216
 217 /* Magic for libthread_db to know how to do THREAD_SELF.  */
 218 # define DB_THREAD_SELF_INCLUDE  <sys/reg.h> /* For the FS constant.  */
 219 # define DB_THREAD_SELF CONST_THREAD_AREA (64, FS)
 220
 221 /* Read member of the thread descriptor directly.  */
 222 # define THREAD_GETMEM(descr, member) \
 223   ({ __typeof (descr->member) __value;                                        \
 224      if (sizeof (__value) == 1)                                               \
 225        __asm__ __volatile__ ("movb %%fs:%P2,%b0"                                      \
 226                      : "=q" (__value)                                         \
 227                      : "0" (0), "i" (offsetof (struct pthread, member)));     \
 228      else if (sizeof (__value) == 4)                                          \
 229        __asm__ __volatile__ ("movl %%fs:%P1,%0"                                       \
 230                      : "=r" (__value)                                         \
 231                      : "i" (offsetof (struct pthread, member)));              \
 232      else                                                                     \
 233        {                                                                      \
 234          if (sizeof (__value) != 8)                                           \
 235            /* There should not be any value with a size other than 1,         \
 236               4 or 8.  */                                                     \
 237            abort ();                                                          \
 238                                                                               \
 239          __asm__ __volatile__ ("movq %%fs:%P1,%q0"                                    \
 240                        : "=r" (__value)                                       \
 241                        : "i" (offsetof (struct pthread, member)));            \
 242        }                                                                      \
 243      __value; })
 244
 245
 246 /* Same as THREAD_GETMEM, but the member offset can be non-constant.  */
 247 # define THREAD_GETMEM_NC(descr, member, idx) \
 248   ({ __typeof (descr->member[0]) __value;                                     \
 249      if (sizeof (__value) == 1)                                               \
 250        __asm__ __volatile__ ("movb %%fs:%P2(%q3),%b0"                                 \
 251                      : "=q" (__value)                                         \
 252                      : "0" (0), "i" (offsetof (struct pthread, member[0])),   \
 253                        "r" (idx));                                            \
 254      else if (sizeof (__value) == 4)                                          \
 255        __asm__ __volatile__ ("movl %%fs:%P1(,%q2,4),%0"                               \
 256                      : "=r" (__value)                                         \
 257                      : "i" (offsetof (struct pthread, member[0])), "r" (idx));\
 258      else                                                                     \
 259        {                                                                      \
 260          if (sizeof (__value) != 8)                                           \
 261            /* There should not be any value with a size other than 1,         \
 262               4 or 8.  */                                                     \
 263            abort ();                                                          \
 264                                                                               \
 265          __asm__ __volatile__ ("movq %%fs:%P1(,%q2,8),%q0"                            \
 266                        : "=r" (__value)                                       \
 267                        : "i" (offsetof (struct pthread, member[0])),          \
 268                          "r" (idx));                                          \
 269        }                                                                      \
 270      __value; })
 271
 272
 273 /* Loading addresses of objects on x86-64 needs to be treated special
 274    when generating PIC code.  */
 275 #ifdef __pic__
 276 # define IMM_MODE "nr"
 277 #else
 278 # define IMM_MODE "ir"
 279 #endif
 280
 281
 282 /* Same as THREAD_SETMEM, but the member offset can be non-constant.  */
 283 # define THREAD_SETMEM(descr, member, value) \
 284   ({ if (sizeof (descr->member) == 1)                                         \
 285        __asm__ __volatile__ ("movb %b0,%%fs:%P1" :                                    \
 286                      : "iq" (value),                                          \
 287                        "i" (offsetof (struct pthread, member)));              \
 288      else if (sizeof (descr->member) == 4)                                    \
 289        __asm__ __volatile__ ("movl %0,%%fs:%P1" :                                     \
 290                      : IMM_MODE (value),                                      \
 291                        "i" (offsetof (struct pthread, member)));              \
 292      else                                                                     \
 293        {                                                                      \
 294          if (sizeof (descr->member) != 8)                                     \
 295            /* There should not be any value with a size other than 1,         \
 296               4 or 8.  */                                                     \
 297            abort ();                                                          \
 298                                                                               \
 299          __asm__ __volatile__ ("movq %q0,%%fs:%P1" :                                  \
 300                        : IMM_MODE ((unsigned long int) value),                \
 301                          "i" (offsetof (struct pthread, member)));            \
 302        }})
 303
 304
 305 /* Set member of the thread descriptor directly.  */
 306 # define THREAD_SETMEM_NC(descr, member, idx, value) \
 307   ({ if (sizeof (descr->member[0]) == 1)                                      \
 308        __asm__ __volatile__ ("movb %b0,%%fs:%P1(%q2)" :                               \
 309                      : "iq" (value),                                          \
 310                        "i" (offsetof (struct pthread, member[0])),            \
 311                        "r" (idx));                                            \
 312      else if (sizeof (descr->member[0]) == 4)                                 \
 313        __asm__ __volatile__ ("movl %0,%%fs:%P1(,%q2,4)" :                             \
 314                      : IMM_MODE (value),                                      \
 315                        "i" (offsetof (struct pthread, member[0])),            \
 316                        "r" (idx));                                            \
 317      else                                                                     \
 318        {                                                                      \
 319          if (sizeof (descr->member[0]) != 8)                                  \
 320            /* There should not be any value with a size other than 1,         \
 321               4 or 8.  */                                                     \
 322            abort ();                                                          \
 323                                                                               \
 324          __asm__ __volatile__ ("movq %q0,%%fs:%P1(,%q2,8)" :                          \
 325                        : IMM_MODE ((unsigned long int) value),                \
 326                          "i" (offsetof (struct pthread, member[0])),          \
 327                          "r" (idx));                                          \
 328        }})
 329
 330
 331 /* Atomic compare and exchange on TLS, returning old value.  */
 332 # define THREAD_ATOMIC_CMPXCHG_VAL(descr, member, newval, oldval) \
 333   ({ __typeof (descr->member) __ret;                                          \
 334      __typeof (oldval) __old = (oldval);                                      \
 335      if (sizeof (descr->member) == 4)                                         \
 336        __asm__ __volatile__ (LOCK_PREFIX "cmpxchgl %2, %%fs:%P3"                      \
 337                      : "=a" (__ret)                                           \
 338                      : "0" (__old), "r" (newval),                             \
 339                        "i" (offsetof (struct pthread, member)));              \
 340      else                                                                     \
 341        /* Not necessary for other sizes in the moment.  */                    \
 342        abort ();                                                              \
 343      __ret; })
 344
 345
 346 /* Atomic logical and.  */
 347 # define THREAD_ATOMIC_AND(descr, member, val) \
 348   (void) ({ if (sizeof ((descr)->member) == 4)                                \
 349               __asm__ __volatile__ (LOCK_PREFIX "andl %1, %%fs:%P0"                   \
 350                             :: "i" (offsetof (struct pthread, member)),       \
 351                                "ir" (val));                                   \
 352             else                                                              \
 353               /* Not necessary for other sizes in the moment.  */             \
 354               abort (); })
 355
 356
 357 /* Atomic set bit.  */
 358 # define THREAD_ATOMIC_BIT_SET(descr, member, bit) \
 359   (void) ({ if (sizeof ((descr)->member) == 4)                                \
 360               __asm__ __volatile__ (LOCK_PREFIX "orl %1, %%fs:%P0"                    \
 361                             :: "i" (offsetof (struct pthread, member)),       \
 362                                "ir" (1 << (bit)));                            \
 363             else                                                              \
 364               /* Not necessary for other sizes in the moment.  */             \
 365               abort (); })
 366
 367
 368 # define CALL_THREAD_FCT(descr) \
 369   ({ void *__res;                                                             \
 370      __asm__ __volatile__ ("movq %%fs:%P2, %%rdi\n\t"                                 \
 371                    "callq *%%fs:%P1"                                          \
 372                    : "=a" (__res)                                             \
 373                    : "i" (offsetof (struct pthread, start_routine)),          \
 374                      "i" (offsetof (struct pthread, arg))                     \
 375                    : "di", "si", "cx", "dx", "r8", "r9", "r10", "r11",        \
 376                      "memory", "cc");                                         \
 377      __res; })
 378
 379
 380 /* Set the stack guard field in TCB head.  */
 381 # define THREAD_SET_STACK_GUARD(value) \
 382     THREAD_SETMEM (THREAD_SELF, header.stack_guard, value)
 383 # define THREAD_COPY_STACK_GUARD(descr) \
 384     ((descr)->header.stack_guard                                              \
 385      = THREAD_GETMEM (THREAD_SELF, header.stack_guard))
 386
 387
 388 /* Set the pointer guard field in the TCB head.  */
 389 # define THREAD_SET_POINTER_GUARD(value) \
 390   THREAD_SETMEM (THREAD_SELF, header.pointer_guard, value)
 391 # define THREAD_COPY_POINTER_GUARD(descr) \
 392   ((descr)->header.pointer_guard                                              \
 393    = THREAD_GETMEM (THREAD_SELF, header.pointer_guard))
 394
 395
 396 /* Get and set the global scope generation counter in the TCB head.  */
 397 # define THREAD_GSCOPE_FLAG_UNUSED 0
 398 # define THREAD_GSCOPE_FLAG_USED   1
 399 # define THREAD_GSCOPE_FLAG_WAIT   2
 400 # define THREAD_GSCOPE_RESET_FLAG() \
 401   do                                                                          \
 402     { int __res;                                                              \
 403       __asm__ __volatile__ ("xchgl %0, %%fs:%P1"                                      \
 404                     : "=r" (__res)                                            \
 405                     : "i" (offsetof (struct pthread, header.gscope_flag)),    \
 406                       "0" (THREAD_GSCOPE_FLAG_UNUSED));                       \
 407       if (__res == THREAD_GSCOPE_FLAG_WAIT)                                   \
 408         lll_futex_wake (&THREAD_SELF->header.gscope_flag, 1, LLL_PRIVATE);    \
 409     }                                                                         \
 410   while (0)
 411 # define THREAD_GSCOPE_SET_FLAG() \
 412   THREAD_SETMEM (THREAD_SELF, header.gscope_flag, THREAD_GSCOPE_FLAG_USED)
 413 # define THREAD_GSCOPE_WAIT() \
 414   GL(dl_wait_lookup_done) ()
 415
 416
 417 # ifdef SHARED
 418 /* Defined in dl-trampoline.S.  */
 419 extern void _dl_x86_64_save_sse (void);
 420 extern void _dl_x86_64_restore_sse (void);
 421
 422 # define RTLD_CHECK_FOREIGN_CALL \
 423   (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) != 0)
 424
 425 /* NB: Don't use the xchg operation because that would imply a lock
 426    prefix which is expensive and unnecessary.  The cache line is also
 427    not contested at all.  */
 428 #  define RTLD_ENABLE_FOREIGN_CALL \
 429   int old_rtld_must_xmm_save = THREAD_GETMEM (THREAD_SELF,                    \
 430                                               header.rtld_must_xmm_save);     \
 431   THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 1)
 432
 433 #  define RTLD_PREPARE_FOREIGN_CALL \
 434   do if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save))              \
 435     {                                                                         \
 436       _dl_x86_64_save_sse ();                                                 \
 437       THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 0);              \
 438     }                                                                         \
 439   while (0)
 440
 441 #  define RTLD_FINALIZE_FOREIGN_CALL \
 442   do {                                                                        \
 443     if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) == 0)          \
 444       _dl_x86_64_restore_sse ();                                              \
 445     THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save,                    \
 446                    old_rtld_must_xmm_save);                                   \
 447   } while (0)
 448 # endif
 449
 450
 451 #endif /* __ASSEMBLER__ */
 452
 453 #endif  /* tls.h */