sysdeps/unix/sysv/linux/s390/elision-lock.c

   1 /* Elided pthread mutex lock.
   2    Copyright (C) 2014-2016 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, see
  17    <http://www.gnu.org/licenses/>.  */
  18
  19 #include <pthread.h>
  20 #include <pthreadP.h>
  21 #include <lowlevellock.h>
  22 #include <htmintrin.h>
  23 #include <elision-conf.h>
  24 #include <stdint.h>
  25
  26 #if !defined(LLL_LOCK) && !defined(EXTRAARG)
  27 /* Make sure the configuration code is always linked in for static
  28    libraries.  */
  29 #include "elision-conf.c"
  30 #endif
  31
  32 #ifndef EXTRAARG
  33 #define EXTRAARG
  34 #endif
  35 #ifndef LLL_LOCK
  36 #define LLL_LOCK(a,b) lll_lock(a,b), 0
  37 #endif
  38
  39 #define aconf __elision_aconf
  40
  41 /* Adaptive lock using transactions.
  42    By default the lock region is run as a transaction, and when it
  43    aborts or the lock is busy the lock adapts itself.  */
  44
  45 int
  46 __lll_lock_elision (int *futex, short *adapt_count, EXTRAARG int private)
  47 {
  48   /* adapt_count can be accessed concurrently; these accesses can be both
  49      inside of transactions (if critical sections are nested and the outer
  50      critical section uses lock elision) and outside of transactions.  Thus,
  51      we need to use atomic accesses to avoid data races.  However, the
  52      value of adapt_count is just a hint, so relaxed MO accesses are
  53      sufficient.  */
  54   if (atomic_load_relaxed (adapt_count) > 0)
  55     {
  56       /* Lost updates are possible, but harmless.  Due to races this might lead
  57          to *adapt_count becoming less than zero.  */
  58       atomic_store_relaxed (adapt_count,
  59                             atomic_load_relaxed (adapt_count) - 1);
  60       goto use_lock;
  61     }
  62
  63   __asm__ volatile (".machinemode \"zarch_nohighgprs\"\n\t"
  64                     ".machine \"all\""
  65                     : : : "memory");
  66
  67   int try_tbegin;
  68   for (try_tbegin = aconf.try_tbegin;
  69        try_tbegin > 0;
  70        try_tbegin--)
  71     {
  72       unsigned status;
  73       if (__builtin_expect
  74           ((status = __builtin_tbegin((void *)0)) == _HTM_TBEGIN_STARTED, 1))
  75         {
  76           if (*futex == 0)
  77             return 0;
  78           /* Lock was busy.  Fall back to normal locking.  */
  79           if (__builtin_expect (__builtin_tx_nesting_depth (), 1))
  80             {
  81               /* In a non-nested transaction there is no need to abort,
  82                  which is expensive.  */
  83               __builtin_tend ();
  84               /* Don't try to use transactions for the next couple of times.
  85                  See above for why relaxed MO is sufficient.  */
  86               if (aconf.skip_lock_busy > 0)
  87                 atomic_store_relaxed (adapt_count, aconf.skip_lock_busy);
  88               goto use_lock;
  89             }
  90           else /* nesting depth is > 1 */
  91             {
  92               /* A nested transaction will abort eventually because it
  93                  cannot make any progress before *futex changes back to 0.
  94                  So we may as well abort immediately.
  95                  This persistently aborts the outer transaction to force
  96                  the outer mutex use the default lock instead of retrying
  97                  with transactions until the try_tbegin of the outer mutex
  98                  is zero.
  99                  The adapt_count of this inner mutex is not changed,
 100                  because using the default lock with the inner mutex
 101                  would abort the outer transaction.
 102               */
 103               __builtin_tabort (_HTM_FIRST_USER_ABORT_CODE | 1);
 104             }
 105         }
 106       else
 107         {
 108           if (status != _HTM_TBEGIN_TRANSIENT)
 109             {
 110               /* A persistent abort (cc 1 or 3) indicates that a retry is
 111                  probably futile.  Use the normal locking now and for the
 112                  next couple of calls.
 113                  Be careful to avoid writing to the lock.  See above for why
 114                  relaxed MO is sufficient.  */
 115               if (aconf.skip_lock_internal_abort > 0)
 116                 atomic_store_relaxed (adapt_count,
 117                                       aconf.skip_lock_internal_abort);
 118               goto use_lock;
 119             }
 120         }
 121     }
 122
 123   /* Same logic as above, but for for a number of temporary failures in a
 124      row.  See above for why relaxed MO is sufficient.  */
 125   if (aconf.skip_lock_out_of_tbegin_retries > 0 && aconf.try_tbegin > 0)
 126     atomic_store_relaxed (adapt_count, aconf.skip_lock_out_of_tbegin_retries);
 127
 128   use_lock:
 129   return LLL_LOCK ((*futex), private);
 130 }