coregrind/m_scheduler/ticket-lock-linux.c

   1
   2 /*--------------------------------------------------------------------*/
   3 /*--- Linux ticket lock implementation         ticket-lock-linux.c ---*/
   4 /*---                                                              ---*/
   5 /*--- Guarantees fair scheduling even if multiple threads are      ---*/
   6 /*--- runnable at the same time on a multicore system. Has been    ---*/
   7 /*--- observed to cause a slow-down compared to the generic        ---*/
   8 /*--- scheduler lock with CPU frequency scaling enabled. Makes     ---*/
   9 /*--- Valgrind slightly faster if CPU frequency scaling has been   ---*/
  10 /*--- disabled. See also http://bugs.kde.org/show_bug.cgi?id=270006---*/
  11 /*--------------------------------------------------------------------*/
  12
  13 /*
  14    This file is part of Valgrind, a dynamic binary instrumentation
  15    framework.
  16
  17    Copyright (C) 2011-2013 Bart Van Assche <bvanassche@acm.org>.
  18
  19    This program is free software; you can redistribute it and/or
  20    modify it under the terms of the GNU General Public License as
  21    published by the Free Software Foundation; either version 2 of the
  22    License, or (at your option) any later version.
  23
  24    This program is distributed in the hope that it will be useful, but
  25    WITHOUT ANY WARRANTY; without even the implied warranty of
  26    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  27    General Public License for more details.
  28
  29    You should have received a copy of the GNU General Public License
  30    along with this program; if not, write to the Free Software
  31    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  32    02111-1307, USA.
  33
  34    The GNU General Public License is contained in the file COPYING.
  35 */
  36
  37 #include "pub_core_basics.h"
  38 #include "pub_core_libcassert.h"
  39 #include "pub_core_libcbase.h"     // VG_(memset)()
  40 #include "pub_core_libcprint.h"
  41 #include "pub_core_syscall.h"
  42 #include "pub_core_vki.h"
  43 #include "pub_core_vkiscnums.h"    // __NR_futex
  44 #include "pub_core_libcproc.h"
  45 #include "pub_core_mallocfree.h"
  46 #include "pub_core_threadstate.h"
  47 #include "pub_core_inner.h"
  48 #if defined(ENABLE_INNER_CLIENT_REQUEST)
  49 #include "helgrind/helgrind.h"
  50 #endif
  51 #include "priv_sched-lock.h"
  52 #include "priv_sched-lock-impl.h"
  53
  54 #define TL_FUTEX_COUNT_LOG2 4
  55 #define TL_FUTEX_COUNT (1U << TL_FUTEX_COUNT_LOG2)
  56 #define TL_FUTEX_MASK (TL_FUTEX_COUNT - 1)
  57
  58 struct sched_lock {
  59    volatile unsigned head;
  60    volatile unsigned tail;
  61    volatile unsigned futex[TL_FUTEX_COUNT];
  62    int owner;
  63 };
  64
  65 #if 1
  66 static Bool s_debug;
  67 #else
  68 static Bool s_debug = True;
  69 #endif
  70
  71 static const HChar *get_sched_lock_name(void)
  72 {
  73    return "ticket lock";
  74 }
  75
  76 static struct sched_lock *create_sched_lock(void)
  77 {
  78    struct sched_lock *p;
  79
  80    p = VG_(malloc)("sched_lock", sizeof(*p));
  81
  82    // The futex syscall requires that a futex takes four bytes.
  83    vg_assert(sizeof(p->futex[0]) == 4);
  84
  85    VG_(memset)(p, 0, sizeof(*p));
  86
  87    INNER_REQUEST(ANNOTATE_RWLOCK_CREATE(p));
  88    INNER_REQUEST(ANNOTATE_BENIGN_RACE_SIZED(&p->futex, sizeof(p->futex), ""));
  89    return p;
  90 }
  91
  92 static void destroy_sched_lock(struct sched_lock *p)
  93 {
  94    INNER_REQUEST(ANNOTATE_RWLOCK_DESTROY(p));
  95    VG_(free)(p);
  96 }
  97
  98 static int get_sched_lock_owner(struct sched_lock *p)
  99 {
 100    return p->owner;
 101 }
 102
 103 /*
 104  * Acquire ticket lock. Increment the tail of the queue and use the original
 105  * value as the ticket value. Wait until the head of the queue equals the
 106  * ticket value. The futex used to wait depends on the ticket value in order
 107  * to avoid that all threads get woken up every time a ticket lock is
 108  * released. That last effect is sometimes called the "thundering herd"
 109  * effect.
 110  *
 111  * See also Nick Piggin, x86: FIFO ticket spinlocks, Linux kernel mailing list
 112  * (http://lkml.org/lkml/2007/11/1/125) for more info.
 113  */
 114 static void acquire_sched_lock(struct sched_lock *p)
 115 {
 116    unsigned ticket, futex_value;
 117    volatile unsigned *futex;
 118    SysRes sres;
 119
 120    ticket = __sync_fetch_and_add(&p->tail, 1);
 121    futex = &p->futex[ticket & TL_FUTEX_MASK];
 122    if (s_debug)
 123       VG_(printf)("[%d/%d] acquire: ticket %d\n", VG_(getpid)(),
 124                   VG_(gettid)(), ticket);
 125    for (;;) {
 126       futex_value = *futex;
 127       __sync_synchronize();
 128       if (ticket == p->head)
 129          break;
 130       if (s_debug)
 131          VG_(printf)("[%d/%d] acquire: ticket %d - waiting until"
 132                      " futex[%ld] != %d\n", VG_(getpid)(),
 133                      VG_(gettid)(), ticket, (long)(futex - p->futex),
 134                      futex_value);
 135       sres = VG_(do_syscall3)(__NR_futex, (UWord)futex,
 136                               VKI_FUTEX_WAIT | VKI_FUTEX_PRIVATE_FLAG,
 137                               futex_value);
 138       if (sr_isError(sres) && sres._val != VKI_EAGAIN) {
 139          VG_(printf)("futex_wait() returned error code %ld\n", sres._val);
 140          vg_assert(False);
 141       }
 142    }
 143    __sync_synchronize();
 144    INNER_REQUEST(ANNOTATE_RWLOCK_ACQUIRED(p, /*is_w*/1));
 145    vg_assert(p->owner == 0);
 146    p->owner = VG_(gettid)();
 147 }
 148
 149 /*
 150  * Release a ticket lock by incrementing the head of the queue. Only generate
 151  * a thread wakeup signal if at least one thread is waiting. If the queue tail
 152  * matches the wakeup_ticket value, no threads have to be woken up.
 153  *
 154  * Note: tail will only be read after head has been incremented since both are
 155  * declared as volatile and since the __sync...() functions imply a memory
 156  * barrier.
 157  */
 158 static void release_sched_lock(struct sched_lock *p)
 159 {
 160    unsigned wakeup_ticket, futex_value;
 161    volatile unsigned *futex;
 162    SysRes sres;
 163
 164    vg_assert(p->owner != 0);
 165    p->owner = 0;
 166    INNER_REQUEST(ANNOTATE_RWLOCK_RELEASED(p, /*is_w*/1));
 167    wakeup_ticket = __sync_fetch_and_add(&p->head, 1) + 1;
 168    if (p->tail != wakeup_ticket) {
 169       futex = &p->futex[wakeup_ticket & TL_FUTEX_MASK];
 170       futex_value = __sync_fetch_and_add(futex, 1);
 171       if (s_debug)
 172          VG_(printf)("[%d/%d] release: waking up ticket %d (futex[%ld] = %d)"
 173                      "\n", VG_(getpid)(), VG_(gettid)(), wakeup_ticket,
 174                      (long)(futex - p->futex), futex_value);
 175       sres = VG_(do_syscall3)(__NR_futex, (UWord)futex,
 176                               VKI_FUTEX_WAKE | VKI_FUTEX_PRIVATE_FLAG,
 177                               0x7fffffff);
 178       vg_assert(!sr_isError(sres));
 179    } else {
 180       if (s_debug)
 181          VG_(printf)("[%d/%d] release: no thread is waiting for ticket %d\n",
 182                      VG_(getpid)(), VG_(gettid)(), wakeup_ticket);
 183    }
 184 }
 185
 186 const struct sched_lock_ops ML_(linux_ticket_lock_ops) = {
 187    .get_sched_lock_name  = get_sched_lock_name,
 188    .create_sched_lock    = create_sched_lock,
 189    .destroy_sched_lock   = destroy_sched_lock,
 190    .get_sched_lock_owner = get_sched_lock_owner,
 191    .acquire_sched_lock   = acquire_sched_lock,
 192    .release_sched_lock   = release_sched_lock,
 193 };