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 #if defined USE_TLS && USE_TLS // l4
  23 #ifndef __ASSEMBLER__
  24 # 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 # 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 # include <sysdep.h>
  98
  99
 100 /* Get the thread descriptor definition.  */
 101 # include <descr.h>
 102
 103 #ifndef LOCK_PREFIX
 104 # ifdef UP
 105 #  define LOCK_PREFIX   /* nothing */
 106 # else
 107 #  define LOCK_PREFIX   "lock;"
 108 # endif
 109 #endif
 110
 111 /* This is the size of the initial TCB.  Can't be just sizeof (tcbhead_t),
 112    because NPTL getpid, __libc_alloca_cutoff etc. need (almost) the whole
 113    struct pthread even when not linked with -lpthread.  */
 114 # define TLS_INIT_TCB_SIZE sizeof (struct pthread)
 115
 116 /* Alignment requirements for the initial TCB.  */
 117 # define TLS_INIT_TCB_ALIGN __alignof__ (struct pthread)
 118
 119 /* This is the size of the TCB.  */
 120 # define TLS_TCB_SIZE sizeof (struct pthread)
 121
 122 /* Alignment requirements for the TCB.  */
 123 //# define TLS_TCB_ALIGN __alignof__ (struct pthread)
 124 // Normally the above would be correct  But we have to store post-AVX
 125 // vector registers in the TCB and we want the storage to be aligned.
 126 // unfortunately there isn't yet a type for these values and hence no
 127 // 32-byte alignment requirement.  Make this explicit, for now.
 128 # define TLS_TCB_ALIGN 32
 129
 130 /* The TCB can have any size and the memory following the address the
 131    thread pointer points to is unspecified.  Allocate the TCB there.  */
 132 # define TLS_TCB_AT_TP  1
 133
 134
 135 /* Install the dtv pointer.  The pointer passed is to the element with
 136    index -1 which contain the length.  */
 137 # define INSTALL_DTV(descr, dtvp) \
 138   ((tcbhead_t *) (descr))->dtv = (dtvp) + 1
 139
 140 /* Install new dtv for current thread.  */
 141 # define INSTALL_NEW_DTV(dtvp) \
 142   ({ struct pthread *__pd;                                                    \
 143      THREAD_SETMEM (__pd, header.dtv, (dtvp)); })
 144
 145 /* Return dtv of given thread descriptor.  */
 146 # define GET_DTV(descr) \
 147   (((tcbhead_t *) (descr))->dtv)
 148
 149
 150 /* Macros to load from and store into segment registers.  */
 151 # define TLS_GET_FS() \
 152   ({ int __seg; __asm__ ("movl %%fs, %0" : "=q" (__seg)); __seg; })
 153 # define TLS_SET_FS(val) \
 154   __asm__ ("movl %0, %%fs" :: "q" (val))
 155
 156
 157 /* Code to initially initialize the thread pointer.  This might need
 158    special attention since 'errno' is not yet available and if the
 159    operation can cause a failure 'errno' must not be touched.
 160
 161    We have to make the syscall for both uses of the macro since the
 162    address might be (and probably is) different.  */
 163 # define TLS_INIT_TP(thrdescr, secondcall) \
 164   ({ void *_thrdescr = (thrdescr);                                            \
 165      tcbhead_t *_head = _thrdescr;                                            \
 166      int _result;                                                             \
 167                                                                               \
 168      _head->tcb = _thrdescr;                                                  \
 169      /* For now the thread descriptor is at the same address.  */             \
 170      _head->self = _thrdescr;                                                 \
 171                                                                               \
 172      /* It is a simple syscall to set the %fs value for the thread.  */       \
 173      __asm__ __volatile__ ("syscall"                                                  \
 174                    : "=a" (_result)                                           \
 175                    : "0" ((unsigned long int) __NR_arch_prctl),               \
 176                      "D" ((unsigned long int) ARCH_SET_FS),                   \
 177                      "S" (_thrdescr)                                          \
 178                    : "memory", "cc", "r11", "cx");                            \
 179                                                                               \
 180     _result ? "cannot set %fs base address for thread-local storage" : 0;     \
 181   })
 182
 183
 184 /* Return the address of the dtv for the current thread.  */
 185 # define THREAD_DTV() \
 186   ({ struct pthread *__pd;                                                    \
 187      THREAD_GETMEM (__pd, header.dtv); })
 188
 189
 190 /* Return the thread descriptor for the current thread.
 191
 192    The contained asm must *not* be marked __volatile__ since otherwise
 193    assignments like
 194         pthread_descr self = thread_self();
 195    do not get optimized away.  */
 196 # define THREAD_SELF \
 197   ({ struct pthread *__self;                                                  \
 198      __asm__ ("movq %%fs:%c1,%q0" : "=r" (__self)                                     \
 199           : "i" (offsetof (struct pthread, header.self)));                    \
 200      __self;})
 201
 202 /* Magic for libthread_db to know how to do THREAD_SELF.  */
 203 # define DB_THREAD_SELF_INCLUDE  <sys/reg.h> /* For the FS constant.  */
 204 # define DB_THREAD_SELF CONST_THREAD_AREA (64, FS)
 205
 206 /* Read member of the thread descriptor directly.  */
 207 # define THREAD_GETMEM(descr, member) \
 208   ({ __typeof (descr->member) __value;                                        \
 209      if (sizeof (__value) == 1)                                               \
 210        __asm__ __volatile__ ("movb %%fs:%P2,%b0"                                      \
 211                      : "=q" (__value)                                         \
 212                      : "0" (0), "i" (offsetof (struct pthread, member)));     \
 213      else if (sizeof (__value) == 4)                                          \
 214        __asm__ __volatile__ ("movl %%fs:%P1,%0"                                       \
 215                      : "=r" (__value)                                         \
 216                      : "i" (offsetof (struct pthread, member)));              \
 217      else                                                                     \
 218        {                                                                      \
 219          if (sizeof (__value) != 8)                                           \
 220            /* There should not be any value with a size other than 1,         \
 221               4 or 8.  */                                                     \
 222            abort ();                                                          \
 223                                                                               \
 224          __asm__ __volatile__ ("movq %%fs:%P1,%q0"                                    \
 225                        : "=r" (__value)                                       \
 226                        : "i" (offsetof (struct pthread, member)));            \
 227        }                                                                      \
 228      __value; })
 229
 230
 231 /* Same as THREAD_GETMEM, but the member offset can be non-constant.  */
 232 # define THREAD_GETMEM_NC(descr, member, idx) \
 233   ({ __typeof (descr->member[0]) __value;                                     \
 234      if (sizeof (__value) == 1)                                               \
 235        __asm__ __volatile__ ("movb %%fs:%P2(%q3),%b0"                                 \
 236                      : "=q" (__value)                                         \
 237                      : "0" (0), "i" (offsetof (struct pthread, member[0])),   \
 238                        "r" (idx));                                            \
 239      else if (sizeof (__value) == 4)                                          \
 240        __asm__ __volatile__ ("movl %%fs:%P1(,%q2,4),%0"                               \
 241                      : "=r" (__value)                                         \
 242                      : "i" (offsetof (struct pthread, member[0])), "r" (idx));\
 243      else                                                                     \
 244        {                                                                      \
 245          if (sizeof (__value) != 8)                                           \
 246            /* There should not be any value with a size other than 1,         \
 247               4 or 8.  */                                                     \
 248            abort ();                                                          \
 249                                                                               \
 250          __asm__ __volatile__ ("movq %%fs:%P1(,%q2,8),%q0"                            \
 251                        : "=r" (__value)                                       \
 252                        : "i" (offsetof (struct pthread, member[0])),          \
 253                          "r" (idx));                                          \
 254        }                                                                      \
 255      __value; })
 256
 257
 258 /* Loading addresses of objects on x86-64 needs to be treated special
 259    when generating PIC code.  */
 260 #ifdef __pic__
 261 # define IMM_MODE "nr"
 262 #else
 263 # define IMM_MODE "ir"
 264 #endif
 265
 266
 267 /* Same as THREAD_SETMEM, but the member offset can be non-constant.  */
 268 # define THREAD_SETMEM(descr, member, value) \
 269   ({ if (sizeof (descr->member) == 1)                                         \
 270        __asm__ __volatile__ ("movb %b0,%%fs:%P1" :                                    \
 271                      : "iq" (value),                                          \
 272                        "i" (offsetof (struct pthread, member)));              \
 273      else if (sizeof (descr->member) == 4)                                    \
 274        __asm__ __volatile__ ("movl %0,%%fs:%P1" :                                     \
 275                      : IMM_MODE (value),                                      \
 276                        "i" (offsetof (struct pthread, member)));              \
 277      else                                                                     \
 278        {                                                                      \
 279          if (sizeof (descr->member) != 8)                                     \
 280            /* There should not be any value with a size other than 1,         \
 281               4 or 8.  */                                                     \
 282            abort ();                                                          \
 283                                                                               \
 284          __asm__ __volatile__ ("movq %q0,%%fs:%P1" :                                  \
 285                        : IMM_MODE ((unsigned long int) value),                \
 286                          "i" (offsetof (struct pthread, member)));            \
 287        }})
 288
 289
 290 /* Set member of the thread descriptor directly.  */
 291 # define THREAD_SETMEM_NC(descr, member, idx, value) \
 292   ({ if (sizeof (descr->member[0]) == 1)                                      \
 293        __asm__ __volatile__ ("movb %b0,%%fs:%P1(%q2)" :                               \
 294                      : "iq" (value),                                          \
 295                        "i" (offsetof (struct pthread, member[0])),            \
 296                        "r" (idx));                                            \
 297      else if (sizeof (descr->member[0]) == 4)                                 \
 298        __asm__ __volatile__ ("movl %0,%%fs:%P1(,%q2,4)" :                             \
 299                      : IMM_MODE (value),                                      \
 300                        "i" (offsetof (struct pthread, member[0])),            \
 301                        "r" (idx));                                            \
 302      else                                                                     \
 303        {                                                                      \
 304          if (sizeof (descr->member[0]) != 8)                                  \
 305            /* There should not be any value with a size other than 1,         \
 306               4 or 8.  */                                                     \
 307            abort ();                                                          \
 308                                                                               \
 309          __asm__ __volatile__ ("movq %q0,%%fs:%P1(,%q2,8)" :                          \
 310                        : IMM_MODE ((unsigned long int) value),                \
 311                          "i" (offsetof (struct pthread, member[0])),          \
 312                          "r" (idx));                                          \
 313        }})
 314
 315
 316 /* Atomic compare and exchange on TLS, returning old value.  */
 317 # define THREAD_ATOMIC_CMPXCHG_VAL(descr, member, newval, oldval) \
 318   ({ __typeof (descr->member) __ret;                                          \
 319      __typeof (oldval) __old = (oldval);                                      \
 320      if (sizeof (descr->member) == 4)                                         \
 321        __asm__ __volatile__ (LOCK_PREFIX "cmpxchgl %2, %%fs:%P3"                      \
 322                      : "=a" (__ret)                                           \
 323                      : "0" (__old), "r" (newval),                             \
 324                        "i" (offsetof (struct pthread, member)));              \
 325      else                                                                     \
 326        /* Not necessary for other sizes in the moment.  */                    \
 327        abort ();                                                              \
 328      __ret; })
 329
 330
 331 /* Atomic logical and.  */
 332 # define THREAD_ATOMIC_AND(descr, member, val) \
 333   (void) ({ if (sizeof ((descr)->member) == 4)                                \
 334               __asm__ __volatile__ (LOCK_PREFIX "andl %1, %%fs:%P0"                   \
 335                             :: "i" (offsetof (struct pthread, member)),       \
 336                                "ir" (val));                                   \
 337             else                                                              \
 338               /* Not necessary for other sizes in the moment.  */             \
 339               abort (); })
 340
 341
 342 /* Atomic set bit.  */
 343 # define THREAD_ATOMIC_BIT_SET(descr, member, bit) \
 344   (void) ({ if (sizeof ((descr)->member) == 4)                                \
 345               __asm__ __volatile__ (LOCK_PREFIX "orl %1, %%fs:%P0"                    \
 346                             :: "i" (offsetof (struct pthread, member)),       \
 347                                "ir" (1 << (bit)));                            \
 348             else                                                              \
 349               /* Not necessary for other sizes in the moment.  */             \
 350               abort (); })
 351
 352
 353 # define CALL_THREAD_FCT(descr) \
 354   ({ void *__res;                                                             \
 355      __asm__ __volatile__ ("movq %%fs:%P2, %%rdi\n\t"                                 \
 356                    "callq *%%fs:%P1"                                          \
 357                    : "=a" (__res)                                             \
 358                    : "i" (offsetof (struct pthread, start_routine)),          \
 359                      "i" (offsetof (struct pthread, arg))                     \
 360                    : "di", "si", "cx", "dx", "r8", "r9", "r10", "r11",        \
 361                      "memory", "cc");                                         \
 362      __res; })
 363
 364
 365 /* Set the stack guard field in TCB head.  */
 366 # define THREAD_SET_STACK_GUARD(value) \
 367     THREAD_SETMEM (THREAD_SELF, header.stack_guard, value)
 368 # define THREAD_COPY_STACK_GUARD(descr) \
 369     ((descr)->header.stack_guard                                              \
 370      = THREAD_GETMEM (THREAD_SELF, header.stack_guard))
 371
 372
 373 /* Set the pointer guard field in the TCB head.  */
 374 # define THREAD_SET_POINTER_GUARD(value) \
 375   THREAD_SETMEM (THREAD_SELF, header.pointer_guard, value)
 376 # define THREAD_COPY_POINTER_GUARD(descr) \
 377   ((descr)->header.pointer_guard                                              \
 378    = THREAD_GETMEM (THREAD_SELF, header.pointer_guard))
 379
 380
 381 /* Get and set the global scope generation counter in the TCB head.  */
 382 # define THREAD_GSCOPE_FLAG_UNUSED 0
 383 # define THREAD_GSCOPE_FLAG_USED   1
 384 # define THREAD_GSCOPE_FLAG_WAIT   2
 385 # define THREAD_GSCOPE_RESET_FLAG() \
 386   do                                                                          \
 387     { int __res;                                                              \
 388       __asm__ __volatile__ ("xchgl %0, %%fs:%P1"                                      \
 389                     : "=r" (__res)                                            \
 390                     : "i" (offsetof (struct pthread, header.gscope_flag)),    \
 391                       "0" (THREAD_GSCOPE_FLAG_UNUSED));                       \
 392       if (__res == THREAD_GSCOPE_FLAG_WAIT)                                   \
 393         lll_futex_wake (&THREAD_SELF->header.gscope_flag, 1, LLL_PRIVATE);    \
 394     }                                                                         \
 395   while (0)
 396 # define THREAD_GSCOPE_SET_FLAG() \
 397   THREAD_SETMEM (THREAD_SELF, header.gscope_flag, THREAD_GSCOPE_FLAG_USED)
 398 # define THREAD_GSCOPE_WAIT() \
 399   GL(dl_wait_lookup_done) ()
 400
 401
 402 # ifdef SHARED
 403 /* Defined in dl-trampoline.S.  */
 404 extern void _dl_x86_64_save_sse (void);
 405 extern void _dl_x86_64_restore_sse (void);
 406
 407 # define RTLD_CHECK_FOREIGN_CALL \
 408   (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) != 0)
 409
 410 /* NB: Don't use the xchg operation because that would imply a lock
 411    prefix which is expensive and unnecessary.  The cache line is also
 412    not contested at all.  */
 413 #  define RTLD_ENABLE_FOREIGN_CALL \
 414   int old_rtld_must_xmm_save = THREAD_GETMEM (THREAD_SELF,                    \
 415                                               header.rtld_must_xmm_save);     \
 416   THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 1)
 417
 418 #  define RTLD_PREPARE_FOREIGN_CALL \
 419   do if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save))              \
 420     {                                                                         \
 421       _dl_x86_64_save_sse ();                                                 \
 422       THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 0);              \
 423     }                                                                         \
 424   while (0)
 425
 426 #  define RTLD_FINALIZE_FOREIGN_CALL \
 427   do {                                                                        \
 428     if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) == 0)          \
 429       _dl_x86_64_restore_sse ();                                              \
 430     THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save,                    \
 431                    old_rtld_must_xmm_save);                                   \
 432   } while (0)
 433 # endif
 434
 435
 436 #endif /* __ASSEMBLER__ */
 437 #endif /* L4 tls disable */
 438 #endif  /* tls.h */