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 #ifndef __ASSEMBLER__
  23 //l4/# include <asm/prctl.h>    /* For ARCH_SET_FS.  */
  24 # include <stdbool.h>
  25 # include <stddef.h>
  26 # include <stdint.h>
  27 # include <stdlib.h>
  28 //l4/# include <sysdep.h>
  29 # include <bits/kernel-features.h>
  30 # include <bits/wordsize.h>
  31 # include <xmmintrin.h>
  32
  33
  34 /* Type for the dtv.  */
  35 typedef union dtv
  36 {
  37   size_t counter;
  38   struct
  39   {
  40     void *val;
  41     bool is_static;
  42   } pointer;
  43 } dtv_t;
  44
  45
  46 typedef struct
  47 {
  48   void *tcb;            /* Pointer to the TCB.  Not necessarily the
  49                            thread descriptor used by libpthread.  */
  50   dtv_t *dtv;
  51   void *self;           /* Pointer to the thread descriptor.  */
  52   int multiple_threads;
  53   int gscope_flag;
  54   uintptr_t sysinfo;
  55   uintptr_t stack_guard;
  56   uintptr_t pointer_guard;
  57   unsigned long int vgetcpu_cache[2];
  58 # ifndef __ASSUME_PRIVATE_FUTEX
  59   int private_futex;
  60 # else
  61   int __unused1;
  62 # endif
  63 # if __WORDSIZE == 64
  64   int rtld_must_xmm_save;
  65 # endif
  66   /* Reservation of some values for the TM ABI.  */
  67   void *__private_tm[5];
  68 # if __WORDSIZE == 64
  69   long int __unused2;
  70   /* Have space for the post-AVX register size.  */
  71   __m128 rtld_savespace_sse[8][4];
  72
  73   void *__padding[8];
  74 # endif
  75 } tcbhead_t;
  76
  77 #else /* __ASSEMBLER__ */
  78 # include <tcb-offsets.h>
  79 #endif
  80
  81
  82 /* We require TLS support in the tools.  */
  83 #define HAVE_TLS_SUPPORT                1
  84 #define HAVE___THREAD 1
  85 #define HAVE_TLS_MODEL_ATTRIBUTE 1
  86
  87 /* Signal that TLS support is available.  */
  88 #define USE_TLS        1
  89
  90 /* Alignment requirement for the stack.  */
  91 #define STACK_ALIGN     16
  92
  93
  94 #ifndef __ASSEMBLER__
  95 /* Get system call information.  */
  96 //l4/# include <sysdep.h>
  97
  98
  99 /* Get the thread descriptor definition.  */
 100 # include <descr.h>
 101 # include <l4/sys/utcb.h>
 102 # include <l4/sys/segment.h>
 103
 104 #ifndef LOCK_PREFIX
 105 # ifdef UP
 106 #  define LOCK_PREFIX   /* nothing */
 107 # else
 108 #  define LOCK_PREFIX   "lock;"
 109 # endif
 110 #endif
 111
 112 /* This is the size of the initial TCB.  Can't be just sizeof (tcbhead_t),
 113    because NPTL getpid, __libc_alloca_cutoff etc. need (almost) the whole
 114    struct pthread even when not linked with -lpthread.  */
 115 # define TLS_INIT_TCB_SIZE sizeof (struct _pthread_descr_struct)
 116
 117 /* Alignment requirements for the initial TCB.  */
 118 # define TLS_INIT_TCB_ALIGN __alignof__ (struct _pthread_descr_struct)
 119
 120 /* This is the size of the TCB.  */
 121 # define TLS_TCB_SIZE sizeof (struct _pthread_descr_struct)
 122
 123 /* Alignment requirements for the TCB.  */
 124 //# define TLS_TCB_ALIGN __alignof__ (struct pthread)
 125 // Normally the above would be correct  But we have to store post-AVX
 126 // vector registers in the TCB and we want the storage to be aligned.
 127 // unfortunately there isn't yet a type for these values and hence no
 128 // 32-byte alignment requirement.  Make this explicit, for now.
 129 # define TLS_TCB_ALIGN 32
 130
 131 /* The TCB can have any size and the memory following the address the
 132    thread pointer points to is unspecified.  Allocate the TCB there.  */
 133 # define TLS_TCB_AT_TP  1
 134
 135
 136 /* Install the dtv pointer.  The pointer passed is to the element with
 137    index -1 which contain the length.  */
 138 # define INSTALL_DTV(descr, dtvp) \
 139   ((tcbhead_t *) (descr))->dtv = (dtvp) + 1
 140
 141 /* Install new dtv for current thread.  */
 142 # define INSTALL_NEW_DTV_ORIG(dtvp) \
 143   ({ struct pthread *__pd;                                                    \
 144      THREAD_SETMEM (__pd, header.dtv, (dtvp)); })
 145 # define INSTALL_NEW_DTV(dtvpx) \
 146   ({ struct _pthread_descr_struct *__pd;                                                      \
 147      THREAD_SETMEM (__pd, p_header.data.dtvp, (dtvpx)); })
 148
 149 /* Return dtv of given thread descriptor.  */
 150 # define GET_DTV(descr) \
 151   (((tcbhead_t *) (descr))->dtv)
 152
 153
 154 /* Macros to load from and store into segment registers.  */
 155 # define TLS_GET_FS() \
 156   ({ int __seg; __asm__ ("movl %%fs, %0" : "=q" (__seg)); __seg; })
 157 # define TLS_SET_FS(val) \
 158   __asm__ ("movl %0, %%fs" :: "q" (val))
 159
 160
 161 static inline char const *TLS_INIT_TP(void *thrdescr, int secondcall)
 162 {
 163   (void)secondcall;
 164   tcbhead_t *_head = (tcbhead_t *)thrdescr;
 165   _head->tcb = thrdescr;
 166   _head->self = thrdescr;
 167   if (fiasco_amd64_set_fs(L4_INVALID_CAP, (l4_umword_t)thrdescr, l4_utcb()) < 0)
 168     return "ERROR";
 169
 170   return NULL;
 171 }
 172 #if 0
 173 /* Code to initially initialize the thread pointer.  This might need
 174    special attention since 'errno' is not yet available and if the
 175    operation can cause a failure 'errno' must not be touched.
 176
 177    We have to make the syscall for both uses of the macro since the
 178    address might be (and probably is) different.  */
 179 # define TLS_INIT_TP(thrdescr, secondcall) \
 180   ({ void *_thrdescr = (thrdescr);                                            \
 181      tcbhead_t *_head = _thrdescr;                                            \
 182      int _result;                                                             \
 183                                                                               \
 184      _head->tcb = _thrdescr;                                                  \
 185      /* For now the thread descriptor is at the same address.  */             \
 186      _head->self = _thrdescr;                                                 \
 187                                                                               \
 188      /* It is a simple syscall to set the %fs value for the thread.  */       \
 189      __asm__ __volatile__ ("syscall"                                                  \
 190                    : "=a" (_result)                                           \
 191                    : "0" ((unsigned long int) __NR_arch_prctl),               \
 192                      "D" ((unsigned long int) ARCH_SET_FS),                   \
 193                      "S" (_thrdescr)                                          \
 194                    : "memory", "cc", "r11", "cx");                            \
 195                                                                               \
 196     _result ? "cannot set %fs base address for thread-local storage" : 0;     \
 197   })
 198 #endif
 199
 200
 201 /* Return the address of the dtv for the current thread.  */
 202 # define THREAD_DTV_ORIG() \
 203   ({ struct pthread *__pd;                                                    \
 204      THREAD_GETMEM (__pd, header.dtv); })
 205
 206 # define THREAD_DTV() \
 207   ({ struct _pthread_descr_struct *__pd;                                                      \
 208      THREAD_GETMEM (__pd, p_header.data.dtvp); })
 209
 210
 211 /* Return the thread descriptor for the current thread.
 212
 213    The contained asm must *not* be marked __volatile__ since otherwise
 214    assignments like
 215         pthread_descr self = thread_self();
 216    do not get optimized away.  */
 217 # define THREAD_SELF_ORIG \
 218   ({ struct pthread *__self;                                                  \
 219      __asm__ ("movq %%fs:%c1,%q0" : "=r" (__self)                                     \
 220           : "i" (offsetof (struct pthread, header.self)));                    \
 221      __self;})
 222 # define THREAD_SELF \
 223   ({ struct _pthread_descr_struct *__self;                                                    \
 224      __asm__ ("movq %%fs:%c1,%q0" : "=r" (__self)                                     \
 225           : "i" (offsetof (struct _pthread_descr_struct, p_header.data.self)));               \
 226      __self;})
 227
 228 /* Magic for libthread_db to know how to do THREAD_SELF.  */
 229 # define DB_THREAD_SELF_INCLUDE  <sys/reg.h> /* For the FS constant.  */
 230 # define DB_THREAD_SELF CONST_THREAD_AREA (64, FS)
 231
 232 /* Read member of the thread descriptor directly.  */
 233 # define THREAD_GETMEM(descr, member) \
 234   ({ __typeof (descr->member) __value;                                        \
 235      if (sizeof (__value) == 1)                                               \
 236        __asm__ __volatile__ ("movb %%fs:%P2,%b0"                                      \
 237                      : "=q" (__value)                                         \
 238                      : "0" (0), "i" (offsetof (struct _pthread_descr_struct, member)));     \
 239      else if (sizeof (__value) == 4)                                          \
 240        __asm__ __volatile__ ("movl %%fs:%P1,%0"                                       \
 241                      : "=r" (__value)                                         \
 242                      : "i" (offsetof (struct _pthread_descr_struct, member)));        \
 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,%q0"                                    \
 251                        : "=r" (__value)                                       \
 252                        : "i" (offsetof (struct _pthread_descr_struct, member)));              \
 253        }                                                                      \
 254      __value; })
 255
 256
 257 /* Same as THREAD_GETMEM, but the member offset can be non-constant.  */
 258 # define THREAD_GETMEM_NC(descr, member, idx) \
 259   ({ __typeof (descr->member[0]) __value;                                     \
 260      if (sizeof (__value) == 1)                                               \
 261        __asm__ __volatile__ ("movb %%fs:%P2(%q3),%b0"                                 \
 262                      : "=q" (__value)                                         \
 263                      : "0" (0), "i" (offsetof (struct _pthread_descr_struct, member[0])),   \
 264                        "r" (idx));                                            \
 265      else if (sizeof (__value) == 4)                                          \
 266        __asm__ __volatile__ ("movl %%fs:%P1(,%q2,4),%0"                               \
 267                      : "=r" (__value)                                         \
 268                      : "i" (offsetof (struct _pthread_descr_struct, member[0])), "r" (idx));\
 269      else                                                                     \
 270        {                                                                      \
 271          if (sizeof (__value) != 8)                                           \
 272            /* There should not be any value with a size other than 1,         \
 273               4 or 8.  */                                                     \
 274            abort ();                                                          \
 275                                                                               \
 276          __asm__ __volatile__ ("movq %%fs:%P1(,%q2,8),%q0"                            \
 277                        : "=r" (__value)                                       \
 278                        : "i" (offsetof (struct _pthread_descr_struct, member[0])),            \
 279                          "r" (idx));                                          \
 280        }                                                                      \
 281      __value; })
 282
 283
 284 /* Loading addresses of objects on x86-64 needs to be treated special
 285    when generating PIC code.  */
 286 #ifdef __pic__
 287 # define IMM_MODE "nr"
 288 #else
 289 # define IMM_MODE "ir"
 290 #endif
 291
 292
 293 /* Same as THREAD_SETMEM, but the member offset can be non-constant.  */
 294 # define THREAD_SETMEM(descr, member, value) \
 295   ({ if (sizeof (descr->member) == 1)                                         \
 296        __asm__ __volatile__ ("movb %b0,%%fs:%P1" :                                    \
 297                      : "iq" (value),                                          \
 298                        "i" (offsetof (struct _pthread_descr_struct, member)));        \
 299      else if (sizeof (descr->member) == 4)                                    \
 300        __asm__ __volatile__ ("movl %0,%%fs:%P1" :                                     \
 301                      : IMM_MODE (value),                                      \
 302                        "i" (offsetof (struct _pthread_descr_struct, member)));        \
 303      else                                                                     \
 304        {                                                                      \
 305          if (sizeof (descr->member) != 8)                                     \
 306            /* There should not be any value with a size other than 1,         \
 307               4 or 8.  */                                                     \
 308            abort ();                                                          \
 309                                                                               \
 310          __asm__ __volatile__ ("movq %q0,%%fs:%P1" :                                  \
 311                        : IMM_MODE ((unsigned long int) value),                \
 312                          "i" (offsetof (struct _pthread_descr_struct, member)));              \
 313        }})
 314
 315
 316 /* Set member of the thread descriptor directly.  */
 317 # define THREAD_SETMEM_NC(descr, member, idx, value) \
 318   ({ if (sizeof (descr->member[0]) == 1)                                      \
 319        __asm__ __volatile__ ("movb %b0,%%fs:%P1(%q2)" :                               \
 320                      : "iq" (value),                                          \
 321                        "i" (offsetof (struct _pthread_descr_struct, member[0])),              \
 322                        "r" (idx));                                            \
 323      else if (sizeof (descr->member[0]) == 4)                                 \
 324        __asm__ __volatile__ ("movl %0,%%fs:%P1(,%q2,4)" :                             \
 325                      : IMM_MODE (value),                                      \
 326                        "i" (offsetof (struct _pthread_descr_struct, member[0])),              \
 327                        "r" (idx));                                            \
 328      else                                                                     \
 329        {                                                                      \
 330          if (sizeof (descr->member[0]) != 8)                                  \
 331            /* There should not be any value with a size other than 1,         \
 332               4 or 8.  */                                                     \
 333            abort ();                                                          \
 334                                                                               \
 335          __asm__ __volatile__ ("movq %q0,%%fs:%P1(,%q2,8)" :                          \
 336                        : IMM_MODE ((unsigned long int) value),                \
 337                          "i" (offsetof (struct _pthread_descr_struct, member[0])),            \
 338                          "r" (idx));                                          \
 339        }})
 340
 341
 342 /* Atomic compare and exchange on TLS, returning old value.  */
 343 # define THREAD_ATOMIC_CMPXCHG_VAL(descr, member, newval, oldval) \
 344   ({ __typeof (descr->member) __ret;                                          \
 345      __typeof (oldval) __old = (oldval);                                      \
 346      if (sizeof (descr->member) == 4)                                         \
 347        __asm__ __volatile__ (LOCK_PREFIX "cmpxchgl %2, %%fs:%P3"                      \
 348                      : "=a" (__ret)                                           \
 349                      : "0" (__old), "r" (newval),                             \
 350                        "i" (offsetof (struct pthread, member)));              \
 351      else                                                                     \
 352        /* Not necessary for other sizes in the moment.  */                    \
 353        abort ();                                                              \
 354      __ret; })
 355
 356
 357 /* Atomic logical and.  */
 358 # define THREAD_ATOMIC_AND(descr, member, val) \
 359   (void) ({ if (sizeof ((descr)->member) == 4)                                \
 360               __asm__ __volatile__ (LOCK_PREFIX "andl %1, %%fs:%P0"                   \
 361                             :: "i" (offsetof (struct pthread, member)),       \
 362                                "ir" (val));                                   \
 363             else                                                              \
 364               /* Not necessary for other sizes in the moment.  */             \
 365               abort (); })
 366
 367
 368 /* Atomic set bit.  */
 369 # define THREAD_ATOMIC_BIT_SET(descr, member, bit) \
 370   (void) ({ if (sizeof ((descr)->member) == 4)                                \
 371               __asm__ __volatile__ (LOCK_PREFIX "orl %1, %%fs:%P0"                    \
 372                             :: "i" (offsetof (struct pthread, member)),       \
 373                                "ir" (1 << (bit)));                            \
 374             else                                                              \
 375               /* Not necessary for other sizes in the moment.  */             \
 376               abort (); })
 377
 378
 379 # define CALL_THREAD_FCT(descr) \
 380   ({ void *__res;                                                             \
 381      __asm__ __volatile__ ("movq %%fs:%P2, %%rdi\n\t"                                 \
 382                    "callq *%%fs:%P1"                                          \
 383                    : "=a" (__res)                                             \
 384                    : "i" (offsetof (struct pthread, start_routine)),          \
 385                      "i" (offsetof (struct pthread, arg))                     \
 386                    : "di", "si", "cx", "dx", "r8", "r9", "r10", "r11",        \
 387                      "memory", "cc");                                         \
 388      __res; })
 389
 390
 391 /* Set the stack guard field in TCB head.  */
 392 # define THREAD_SET_STACK_GUARD(value) \
 393     THREAD_SETMEM (THREAD_SELF, header.stack_guard, value)
 394 # define THREAD_COPY_STACK_GUARD(descr) \
 395     ((descr)->header.stack_guard                                              \
 396      = THREAD_GETMEM (THREAD_SELF, header.stack_guard))
 397
 398
 399 /* Set the pointer guard field in the TCB head.  */
 400 # define THREAD_SET_POINTER_GUARD(value) \
 401   THREAD_SETMEM (THREAD_SELF, header.pointer_guard, value)
 402 # define THREAD_COPY_POINTER_GUARD(descr) \
 403   ((descr)->header.pointer_guard                                              \
 404    = THREAD_GETMEM (THREAD_SELF, header.pointer_guard))
 405
 406
 407 /* Get and set the global scope generation counter in the TCB head.  */
 408 # define THREAD_GSCOPE_FLAG_UNUSED 0
 409 # define THREAD_GSCOPE_FLAG_USED   1
 410 # define THREAD_GSCOPE_FLAG_WAIT   2
 411 # define THREAD_GSCOPE_RESET_FLAG() \
 412   do                                                                          \
 413     { int __res;                                                              \
 414       __asm__ __volatile__ ("xchgl %0, %%fs:%P1"                                      \
 415                     : "=r" (__res)                                            \
 416                     : "i" (offsetof (struct pthread, header.gscope_flag)),    \
 417                       "0" (THREAD_GSCOPE_FLAG_UNUSED));                       \
 418       if (__res == THREAD_GSCOPE_FLAG_WAIT)                                   \
 419         lll_futex_wake (&THREAD_SELF->header.gscope_flag, 1, LLL_PRIVATE);    \
 420     }                                                                         \
 421   while (0)
 422 # define THREAD_GSCOPE_SET_FLAG() \
 423   THREAD_SETMEM (THREAD_SELF, header.gscope_flag, THREAD_GSCOPE_FLAG_USED)
 424 # define THREAD_GSCOPE_WAIT() \
 425   GL(dl_wait_lookup_done) ()
 426
 427
 428 # ifdef SHARED
 429 /* Defined in dl-trampoline.S.  */
 430 extern void _dl_x86_64_save_sse (void);
 431 extern void _dl_x86_64_restore_sse (void);
 432
 433 # define RTLD_CHECK_FOREIGN_CALL \
 434   (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) != 0)
 435
 436 /* NB: Don't use the xchg operation because that would imply a lock
 437    prefix which is expensive and unnecessary.  The cache line is also
 438    not contested at all.  */
 439 #  define RTLD_ENABLE_FOREIGN_CALL \
 440   int old_rtld_must_xmm_save = THREAD_GETMEM (THREAD_SELF,                    \
 441                                               header.rtld_must_xmm_save);     \
 442   THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 1)
 443
 444 #  define RTLD_PREPARE_FOREIGN_CALL \
 445   do if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save))              \
 446     {                                                                         \
 447       _dl_x86_64_save_sse ();                                                 \
 448       THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 0);              \
 449     }                                                                         \
 450   while (0)
 451
 452 #  define RTLD_FINALIZE_FOREIGN_CALL \
 453   do {                                                                        \
 454     if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) == 0)          \
 455       _dl_x86_64_restore_sse ();                                              \
 456     THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save,                    \
 457                    old_rtld_must_xmm_save);                                   \
 458   } while (0)
 459 # endif
 460
 461
 462 #endif /* __ASSEMBLER__ */
 463 #endif  /* tls.h */