sys/kern/kern_mplock.c

   1 /*
   2  * Copyright (c) 2009 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 /*
  36  * Helper functions for MP lock acquisition and release.
  37  */
  38
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/kernel.h>
  42 #include <sys/proc.h>
  43 #include <sys/rtprio.h>
  44 #include <sys/queue.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/kthread.h>
  47 #include <machine/cpu.h>
  48 #include <sys/lock.h>
  49 #include <sys/caps.h>
  50 #include <sys/spinlock.h>
  51 #include <sys/ktr.h>
  52
  53 #include <sys/thread2.h>
  54 #include <sys/mplock2.h>
  55 #include <sys/spinlock2.h>
  56
  57 #ifdef SMP
  58 static int chain_mplock = 0;
  59 static int bgl_yield = 10;
  60 static __int64_t mplock_contention_count = 0;
  61
  62 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
  63 SYSCTL_INT(_lwkt, OID_AUTO, bgl_yield_delay, CTLFLAG_RW, &bgl_yield, 0, "");
  64 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
  65         &mplock_contention_count, 0, "spinning due to MPLOCK contention");
  66
  67 /*
  68  * Kernel Trace
  69  */
  70 #if !defined(KTR_GIANT_CONTENTION)
  71 #define KTR_GIANT_CONTENTION    KTR_ALL
  72 #endif
  73
  74 KTR_INFO_MASTER(giant);
  75 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0,
  76         "thread=%p held %s:%-5d  want %s:%-5d",
  77          sizeof(void *) * 3 + sizeof(int) * 2);
  78 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1,
  79         "thread=%p held %s:%-5d  want %s:%-5d",
  80          sizeof(void *) * 3 + sizeof(int) * 2);
  81
  82 #define loggiant(name)                                          \
  83         KTR_LOG(giant_ ## name, curthread,                      \
  84                 mp_lock_holder_file, mp_lock_holder_line,       \
  85                 file, line)
  86
  87 int     mp_lock;
  88 int     cpu_contention_mask;
  89 const char *mp_lock_holder_file;        /* debugging */
  90 int     mp_lock_holder_line;            /* debugging */
  91
  92 /*
  93  * Sets up the initial MP lock state near the start of the kernel boot
  94  */
  95 void
  96 cpu_get_initial_mplock(void)
  97 {
  98         mp_lock = 0;    /* cpu 0 */
  99         curthread->td_mpcount = 1;
 100 }
 101
 102 /*
 103  * This code is called from the get_mplock() inline when the mplock
 104  * is not already held.  td_mpcount has already been predisposed
 105  * (incremented).
 106  */
 107 void
 108 _get_mplock_predisposed(const char *file, int line)
 109 {
 110         globaldata_t gd = mycpu;
 111
 112         if (gd->gd_intr_nesting_level) {
 113                 panic("Attempt to acquire mplock not already held "
 114                       "in hard section, ipi or interrupt %s:%d",
 115                       file, line);
 116         }
 117         if (atomic_cmpset_int(&mp_lock, -1, gd->gd_cpuid) == 0)
 118                 _get_mplock_contested(file, line);
 119 #ifdef INVARIANTS
 120         mp_lock_holder_file = file;
 121         mp_lock_holder_line = line;
 122 #endif
 123 }
 124
 125 /*
 126  * Called when the MP lock could not be trvially acquired.  The caller
 127  * has already bumped td_mpcount.
 128  */
 129 void
 130 _get_mplock_contested(const char *file, int line)
 131 {
 132         globaldata_t gd = mycpu;
 133         int ov;
 134         int nv;
 135         const void **stkframe = (const void **)&file;
 136
 137         ++mplock_contention_count;
 138         for (;;) {
 139                 ov = mp_lock;
 140                 nv = gd->gd_cpuid;
 141                 if (ov == gd->gd_cpuid)
 142                         break;
 143                 if (ov == -1) {
 144                         if (atomic_cmpset_int(&mp_lock, ov, gd->gd_cpuid))
 145                                 break;
 146                 } else {
 147                         gd->gd_curthread->td_mplock_stallpc = stkframe[-1];
 148                         loggiant(beg);
 149                         lwkt_switch();
 150                         loggiant(end);
 151                         KKASSERT(gd->gd_cpuid == mp_lock);
 152                         break;
 153                 }
 154         }
 155 }
 156
 157 /*
 158  * Called if td_mpcount went negative or if td_mpcount + td_xpcount is 0
 159  * and we were unable to release the MP lock.  Handles sanity checks
 160  * and conflicts.
 161  *
 162  * It is possible for the inline release to have raced an interrupt which
 163  * get/rel'd the MP lock, causing the inline's cmpset to fail.  If this
 164  * case occurs mp_lock will either already be in a released state or it
 165  * will have already been acquired by another cpu.
 166  */
 167 void
 168 _rel_mplock_contested(void)
 169 {
 170         globaldata_t gd = mycpu;
 171         thread_t td = gd->gd_curthread;
 172         int ov;
 173
 174         KKASSERT(td->td_mpcount >= 0);
 175         if (td->td_mpcount + td->td_xpcount == 0) {
 176                 for (;;) {
 177                         ov = mp_lock;
 178                         if (ov != gd->gd_cpuid)
 179                                 break;
 180                         if (atomic_cmpset_int(&mp_lock, ov, -1))
 181                                 break;
 182                 }
 183         }
 184 }
 185
 186 /*
 187  * Called when try_mplock() fails.
 188  *
 189  * The inline bumped td_mpcount so we have to undo it.
 190  *
 191  * It is possible to race an interrupt which acquired and released the
 192  * MP lock.  When combined with the td_mpcount decrement we do the MP lock
 193  * can wind up in any state and possibly not even owned by us.
 194  *
 195  * It is also possible for this function to be called even if td_mpcount > 1
 196  * if someone bumped it and raced an interrupt which then called try_mpock().
 197  */
 198 void
 199 _try_mplock_contested(const char *file, int line)
 200 {
 201         globaldata_t gd = mycpu;
 202         thread_t td = gd->gd_curthread;
 203         int ov;
 204
 205         --td->td_mpcount;
 206         KKASSERT(td->td_mpcount >= 0);
 207         ++mplock_contention_count;
 208
 209         if (td->td_mpcount + td->td_xpcount == 0) {
 210                 for (;;) {
 211                         ov = mp_lock;
 212                         if (ov != gd->gd_cpuid)
 213                                 break;
 214                         if (atomic_cmpset_int(&mp_lock, ov, -1))
 215                                 break;
 216                 }
 217         }
 218 }
 219
 220 /*
 221  * Called when cpu_try_mplock() fails.
 222  *
 223  * The inline did not touch td_mpcount so we do not either.
 224  */
 225 void
 226 _cpu_try_mplock_contested(const char *file, int line)
 227 {
 228         ++mplock_contention_count;
 229 }
 230
 231 /*
 232  * Temporarily yield the MP lock.  This is part of lwkt_user_yield()
 233  * which is kinda hackish.  The MP lock cannot be yielded if inherited
 234  * due to a preemption.
 235  */
 236 void
 237 yield_mplock(thread_t td)
 238 {
 239         int savecnt;
 240
 241         if (td->td_xpcount == 0) {
 242                 savecnt = td->td_mpcount;
 243                 td->td_mpcount = 1;
 244                 rel_mplock();
 245                 DELAY(bgl_yield);
 246                 get_mplock();
 247                 td->td_mpcount = savecnt;
 248         }
 249 }
 250
 251 #if 0
 252
 253 /*
 254  * The rel_mplock() code will call this function after releasing the
 255  * last reference on the MP lock if cpu_contention_mask is non-zero.
 256  *
 257  * We then chain an IPI to a single other cpu potentially needing the
 258  * lock.  This is a bit heuristical and we can wind up with IPIs flying
 259  * all over the place.
 260  */
 261 static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
 262
 263 void
 264 lwkt_mp_lock_uncontested(void)
 265 {
 266     globaldata_t gd;
 267     globaldata_t dgd;
 268     cpumask_t mask;
 269     cpumask_t tmpmask;
 270     int cpuid;
 271
 272     if (chain_mplock) {
 273         gd = mycpu;
 274         clr_mplock_contention_mask(gd);
 275         mask = cpu_contention_mask;
 276         tmpmask = ~((1 << gd->gd_cpuid) - 1);
 277
 278         if (mask) {
 279             if (mask & tmpmask)
 280                     cpuid = bsfl(mask & tmpmask);
 281             else
 282                     cpuid = bsfl(mask);
 283             atomic_clear_int(&cpu_contention_mask, 1 << cpuid);
 284             dgd = globaldata_find(cpuid);
 285             lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
 286         }
 287     }
 288 }
 289
 290 /*
 291  * The idea is for this IPI to interrupt a potentially lower priority
 292  * thread, such as a user thread, to allow the scheduler to reschedule
 293  * a higher priority kernel thread that needs the MP lock.
 294  *
 295  * For now we set the LWKT reschedule flag which generates an AST in
 296  * doreti, though theoretically it is also possible to possibly preempt
 297  * here if the underlying thread was operating in user mode.  Nah.
 298  */
 299 static void
 300 lwkt_mp_lock_uncontested_remote(void *arg __unused)
 301 {
 302         need_lwkt_resched();
 303 }
 304
 305 #endif
 306
 307 #endif  /* SMP */