sys/kern/kern_membarrier.c

   1 /*-
   2  * Copyright (c) 2021 The FreeBSD Foundation
   3  *
   4  * This software were developed by Konstantin Belousov <kib@FreeBSD.org>
   5  * under sponsorship from the FreeBSD Foundation.
   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  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/param.h>
  30 #include <sys/systm.h>
  31 #include <sys/cpuset.h>
  32 #include <sys/lock.h>
  33 #include <sys/membarrier.h>
  34 #include <sys/mutex.h>
  35 #include <sys/proc.h>
  36 #include <sys/sched.h>
  37 #include <sys/smp.h>
  38 #include <sys/syscallsubr.h>
  39 #include <sys/sysproto.h>
  40
  41 #include <vm/vm_param.h>
  42 #include <vm/vm.h>
  43 #include <vm/pmap.h>
  44 #include <vm/vm_map.h>
  45
  46 #define MEMBARRIER_SUPPORTED_CMDS       (                       \
  47     MEMBARRIER_CMD_GLOBAL |                                     \
  48     MEMBARRIER_CMD_GLOBAL_EXPEDITED |                           \
  49     MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED |                  \
  50     MEMBARRIER_CMD_PRIVATE_EXPEDITED |                          \
  51     MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED |                 \
  52     MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE |                \
  53     MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE)
  54
  55 static void
  56 membarrier_action_seqcst(void *arg __unused)
  57 {
  58         atomic_thread_fence_seq_cst();
  59 }
  60
  61 static void
  62 membarrier_action_seqcst_sync_core(void *arg __unused)
  63 {
  64         atomic_thread_fence_seq_cst();
  65         cpu_sync_core();
  66 }
  67
  68 static void
  69 do_membarrier_ipi(cpuset_t *csp, void (*func)(void *))
  70 {
  71         atomic_thread_fence_seq_cst();
  72         smp_rendezvous_cpus(*csp, smp_no_rendezvous_barrier, func,
  73             smp_no_rendezvous_barrier, NULL);
  74         atomic_thread_fence_seq_cst();
  75 }
  76
  77 static void
  78 check_cpu_switched(int c, cpuset_t *csp, uint64_t *swt, bool init)
  79 {
  80         struct pcpu *pc;
  81         uint64_t sw;
  82
  83         if (CPU_ISSET(c, csp))
  84                 return;
  85
  86         pc = cpuid_to_pcpu[c];
  87         if (pc->pc_curthread == pc->pc_idlethread) {
  88                 CPU_SET(c, csp);
  89                 return;
  90         }
  91
  92         /*
  93          * Sync with context switch to ensure that override of
  94          * pc_curthread with non-idle thread pointer is visible before
  95          * reading of pc_switchtime.
  96          */
  97         atomic_thread_fence_acq();
  98
  99         sw = pc->pc_switchtime;
 100         if (init)
 101                 swt[c] = sw;
 102         else if (sw != swt[c])
 103                 CPU_SET(c, csp);
 104 }
 105
 106 /*
 107  *
 108  * XXXKIB: We execute the requested action (seq_cst and possibly
 109  * sync_core) on current CPU as well.  There is no guarantee that
 110  * current thread executes anything with the full fence semantics
 111  * during syscall execution.  Similarly, cpu_core_sync() semantics
 112  * might be not provided by the syscall return.  E.g. on amd64 we
 113  * typically return without IRET.
 114  */
 115 int
 116 kern_membarrier(struct thread *td, int cmd, unsigned flags, int cpu_id)
 117 {
 118         struct proc *p, *p1;
 119         struct thread *td1;
 120         cpuset_t cs;
 121         uint64_t *swt;
 122         int c, error;
 123         bool first;
 124
 125         if (flags != 0 || (cmd & ~MEMBARRIER_SUPPORTED_CMDS) != 0)
 126                 return (EINVAL);
 127
 128         if (cmd == MEMBARRIER_CMD_QUERY) {
 129                 td->td_retval[0] = MEMBARRIER_SUPPORTED_CMDS;
 130                 return (0);
 131         }
 132
 133         p = td->td_proc;
 134         error = 0;
 135
 136         switch (cmd) {
 137         case MEMBARRIER_CMD_GLOBAL:
 138                 swt = malloc((mp_maxid + 1) * sizeof(*swt), M_TEMP, M_WAITOK);
 139                 CPU_ZERO(&cs);
 140                 sched_pin();
 141                 CPU_SET(PCPU_GET(cpuid), &cs);
 142                 for (first = true; error == 0; first = false) {
 143                         CPU_FOREACH(c)
 144                                 check_cpu_switched(c, &cs, swt, first);
 145                         if (CPU_CMP(&cs, &all_cpus) == 0)
 146                                 break;
 147                         error = pause_sig("mmbr", 1);
 148                         if (error == EWOULDBLOCK)
 149                                 error = 0;
 150                 }
 151                 sched_unpin();
 152                 free(swt, M_TEMP);
 153                 atomic_thread_fence_seq_cst();
 154                 break;
 155
 156         case MEMBARRIER_CMD_GLOBAL_EXPEDITED:
 157                 if ((td->td_proc->p_flag2 & P2_MEMBAR_GLOBE) == 0) {
 158                         error = EPERM;
 159                 } else {
 160                         CPU_ZERO(&cs);
 161                         CPU_FOREACH(c) {
 162                                 td1 = cpuid_to_pcpu[c]->pc_curthread;
 163                                 p1 = td1->td_proc;
 164                                 if (p1 != NULL &&
 165                                     (p1->p_flag2 & P2_MEMBAR_GLOBE) != 0)
 166                                         CPU_SET(c, &cs);
 167                         }
 168                         do_membarrier_ipi(&cs, membarrier_action_seqcst);
 169                 }
 170                 break;
 171
 172         case MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED:
 173                 if ((p->p_flag2 & P2_MEMBAR_GLOBE) == 0) {
 174                         PROC_LOCK(p);
 175                         p->p_flag2 |= P2_MEMBAR_GLOBE;
 176                         PROC_UNLOCK(p);
 177                 }
 178                 break;
 179
 180         case MEMBARRIER_CMD_PRIVATE_EXPEDITED:
 181                 if ((td->td_proc->p_flag2 & P2_MEMBAR_PRIVE) == 0) {
 182                         error = EPERM;
 183                 } else {
 184                         pmap_active_cpus(vmspace_pmap(p->p_vmspace), &cs);
 185                         do_membarrier_ipi(&cs, membarrier_action_seqcst);
 186                 }
 187                 break;
 188
 189         case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED:
 190                 if ((p->p_flag2 & P2_MEMBAR_PRIVE) == 0) {
 191                         PROC_LOCK(p);
 192                         p->p_flag2 |= P2_MEMBAR_PRIVE;
 193                         PROC_UNLOCK(p);
 194                 }
 195                 break;
 196
 197         case MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE:
 198                 if ((td->td_proc->p_flag2 & P2_MEMBAR_PRIVE_SYNCORE) == 0) {
 199                         error = EPERM;
 200                 } else {
 201                         /*
 202                          * Calculating the IPI multicast mask from
 203                          * pmap active mask means that we do not call
 204                          * cpu_sync_core() on CPUs that were missed
 205                          * from pmap active mask but could be switched
 206                          * from or to meantime.  This is fine at least
 207                          * on amd64 because threads always use slow
 208                          * (IRETQ) path to return from syscall after
 209                          * context switch.
 210                          */
 211                         pmap_active_cpus(vmspace_pmap(p->p_vmspace), &cs);
 212
 213                         do_membarrier_ipi(&cs,
 214                             membarrier_action_seqcst_sync_core);
 215                 }
 216                 break;
 217
 218         case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE:
 219                 if ((p->p_flag2 & P2_MEMBAR_PRIVE_SYNCORE) == 0) {
 220                         PROC_LOCK(p);
 221                         p->p_flag2 |= P2_MEMBAR_PRIVE_SYNCORE;
 222                         PROC_UNLOCK(p);
 223                 }
 224                 break;
 225
 226         default:
 227                 error = EINVAL;
 228                 break;
 229         }
 230
 231         return (error);
 232 }
 233
 234 int
 235 sys_membarrier(struct thread *td, struct membarrier_args *uap)
 236 {
 237         return (kern_membarrier(td, uap->cmd, uap->flags, uap->cpu_id));
 238 }