sys/kern/kern_umtx.c

   1 /*
   2  * (MPSAFE)
   3  *
   4  * Copyright (c) 2003,2004,2010 The DragonFly Project.  All rights reserved.
   5  *
   6  * This code is derived from software contributed to The DragonFly Project
   7  * by Matthew Dillon <dillon@backplane.com> and David Xu <davidxu@freebsd.org>
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  *
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice, this list of conditions and the following disclaimer in
  17  *    the documentation and/or other materials provided with the
  18  *    distribution.
  19  * 3. Neither the name of The DragonFly Project nor the names of its
  20  *    contributors may be used to endorse or promote products derived
  21  *    from this software without specific, prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  26  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  27  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  28  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  29  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  30  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  31  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  32  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  33  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  */
  36
  37 /*
  38  * This module implements userland mutex helper functions.  umtx_sleep()
  39  * handling blocking and umtx_wakeup() handles wakeups.  The sleep/wakeup
  40  * functions operate on user addresses.
  41  */
  42
  43 #include <sys/param.h>
  44 #include <sys/systm.h>
  45 #include <sys/kernel.h>
  46 #include <sys/sysproto.h>
  47 #include <sys/sysunion.h>
  48 #include <sys/sysent.h>
  49 #include <sys/syscall.h>
  50 #include <sys/module.h>
  51
  52 #include <cpu/lwbuf.h>
  53
  54 #include <vm/vm.h>
  55 #include <vm/vm_param.h>
  56 #include <sys/lock.h>
  57 #include <vm/pmap.h>
  58 #include <vm/vm_map.h>
  59 #include <vm/vm_object.h>
  60 #include <vm/vm_page.h>
  61 #include <vm/vm_pager.h>
  62 #include <vm/vm_pageout.h>
  63 #include <vm/vm_extern.h>
  64 #include <vm/vm_kern.h>
  65
  66 #include <vm/vm_page2.h>
  67
  68 #include <machine/vmm.h>
  69
  70 static void umtx_sleep_page_action_cow(vm_page_t m, vm_page_action_t action);
  71
  72 /*
  73  * If the contents of the userland-supplied pointer matches the specified
  74  * value enter an interruptable sleep for up to <timeout> microseconds.
  75  * If the contents does not match then return immediately.
  76  *
  77  * Returns 0 if we slept and were woken up, -1 and EWOULDBLOCK if we slept
  78  * and timed out, and EBUSY if the contents of the pointer already does
  79  * not match the specified value.  A timeout of 0 indicates an unlimited sleep.
  80  * EINTR is returned if the call was interrupted by a signal (even if
  81  * the signal specifies that the system call should restart).
  82  *
  83  * This function interlocks against call to umtx_wakeup.  It does NOT interlock
  84  * against changes in *ptr.  However, it does not have to.  The standard use
  85  * of *ptr is to differentiate between an uncontested and a contested mutex
  86  * and call umtx_wakeup when releasing a contested mutex.  Therefore we can
  87  * safely race against changes in *ptr as long as we are properly interlocked
  88  * against the umtx_wakeup() call.
  89  *
  90  * The VM page associated with the mutex is held in an attempt to keep
  91  * the mutex's physical address consistent, allowing umtx_sleep() and
  92  * umtx_wakeup() to use the physical address as their rendezvous.  BUT
  93  * situations can arise where the physical address may change, particularly
  94  * if a threaded program fork()'s and the mutex's memory becomes
  95  * copy-on-write.  We register an event on the VM page to catch COWs.
  96  *
  97  * umtx_sleep { const int *ptr, int value, int timeout }
  98  */
  99 int
 100 sys_umtx_sleep(struct umtx_sleep_args *uap)
 101 {
 102     struct lwbuf lwb_cache;
 103     struct lwbuf *lwb;
 104     struct vm_page_action action;
 105     vm_page_t m;
 106     void *waddr;
 107     int offset;
 108     int timeout;
 109     int error = EBUSY;
 110
 111     if (uap->timeout < 0)
 112         return (EINVAL);
 113
 114     if (curthread->td_vmm) {
 115         register_t gpa;
 116         vmm_vm_get_gpa(curproc, &gpa, (register_t) uap->ptr);
 117         uap->ptr = (const int *)gpa;
 118     }
 119
 120     if ((vm_offset_t)uap->ptr & (sizeof(int) - 1))
 121         return (EFAULT);
 122
 123     /*
 124      * When faulting in the page, force any COW pages to be resolved.
 125      * Otherwise the physical page we sleep on my not match the page
 126      * being woken up.
 127      *
 128      * WARNING! We can only use vm_fault_page*() for reading data.  We
 129      *          cannot use it for writing data because there is no pmap
 130      *          interlock to protect against flushes/pageouts.
 131      */
 132     m = vm_fault_page_quick((vm_offset_t)uap->ptr,
 133                             VM_PROT_READ|VM_PROT_WRITE, &error, NULL);
 134     if (m == NULL) {
 135         error = EFAULT;
 136         goto done;
 137     }
 138     lwb = lwbuf_alloc(m, &lwb_cache);
 139     offset = (vm_offset_t)uap->ptr & PAGE_MASK;
 140
 141     /*
 142      * The critical section is required to interlock the tsleep against
 143      * a wakeup from another cpu.  The lfence forces synchronization.
 144      */
 145     if (*(int *)(lwbuf_kva(lwb) + offset) == uap->value) {
 146         if ((timeout = uap->timeout) != 0) {
 147             timeout = (timeout / 1000000) * hz +
 148                       ((timeout % 1000000) * hz + 999999) / 1000000;
 149         }
 150         waddr = (void *)((intptr_t)VM_PAGE_TO_PHYS(m) + offset);
 151
 152         /*
 153          * Wake us up if the memory location COWs while we are sleeping.
 154          */
 155         crit_enter();
 156         vm_page_init_action(m, &action, umtx_sleep_page_action_cow, waddr);
 157         vm_page_register_action(&action, VMEVENT_COW);
 158
 159         /*
 160          * We must interlock just before sleeping.  If we interlock before
 161          * registration the lock operations done by the registration can
 162          * interfere with it.
 163          */
 164         tsleep_interlock(waddr, PCATCH | PDOMAIN_UMTX);
 165         if (*(int *)(lwbuf_kva(lwb) + offset) == uap->value &&
 166             action.event == VMEVENT_COW) {
 167                 error = tsleep(waddr, PCATCH | PINTERLOCKED | PDOMAIN_UMTX,
 168                                "umtxsl", timeout);
 169         } else {
 170                 error = EBUSY;
 171         }
 172         vm_page_unregister_action(&action);
 173         crit_exit();
 174         /* Always break out in case of signal, even if restartable */
 175         if (error == ERESTART)
 176                 error = EINTR;
 177     } else {
 178         error = EBUSY;
 179     }
 180
 181     lwbuf_free(lwb);
 182     vm_page_unhold(m);
 183 done:
 184     return(error);
 185 }
 186
 187 /*
 188  * If this page is being copied it may no longer represent the page
 189  * underlying our virtual address.  Wake up any umtx_sleep()'s
 190  * that were waiting on its physical address to force them to retry.
 191  */
 192 static void
 193 umtx_sleep_page_action_cow(vm_page_t m, vm_page_action_t action)
 194 {
 195     wakeup_domain(action->data, PDOMAIN_UMTX);
 196 }
 197
 198 /*
 199  * umtx_wakeup { const int *ptr, int count }
 200  *
 201  * Wakeup the specified number of processes held in umtx_sleep() on the
 202  * specified user address.  A count of 0 wakes up all waiting processes.
 203  *
 204  * XXX assumes that the physical address space does not exceed the virtual
 205  * address space.
 206  */
 207 int
 208 sys_umtx_wakeup(struct umtx_wakeup_args *uap)
 209 {
 210     vm_page_t m;
 211     int offset;
 212     int error;
 213     void *waddr;
 214
 215     if (curthread->td_vmm) {
 216         register_t gpa;
 217         vmm_vm_get_gpa(curproc, &gpa, (register_t) uap->ptr);
 218         uap->ptr = (const int *)gpa;
 219     }
 220
 221     /*
 222      * WARNING! We can only use vm_fault_page*() for reading data.  We
 223      *          cannot use it for writing data because there is no pmap
 224      *          interlock to protect against flushes/pageouts.
 225      */
 226     cpu_mfence();
 227     if ((vm_offset_t)uap->ptr & (sizeof(int) - 1))
 228         return (EFAULT);
 229     m = vm_fault_page_quick((vm_offset_t)uap->ptr,
 230                             VM_PROT_READ, &error, NULL);
 231     if (m == NULL) {
 232         error = EFAULT;
 233         goto done;
 234     }
 235     offset = (vm_offset_t)uap->ptr & PAGE_MASK;
 236     waddr = (void *)((intptr_t)VM_PAGE_TO_PHYS(m) + offset);
 237
 238     if (uap->count == 1) {
 239         wakeup_domain_one(waddr, PDOMAIN_UMTX);
 240     } else {
 241         /* XXX wakes them all up for now */
 242         wakeup_domain(waddr, PDOMAIN_UMTX);
 243     }
 244     vm_page_unhold(m);
 245     error = 0;
 246 done:
 247     return(error);
 248 }