sys/kern/kern_sharedpage.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2010, 2012 Konstantin Belousov <kib@FreeBSD.org>
   5  * Copyright (c) 2015 The FreeBSD Foundation
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by Konstantin Belousov
   9  * under sponsorship from the FreeBSD Foundation.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  */
  32
  33 #include <sys/cdefs.h>
  34 __FBSDID("$FreeBSD$");
  35
  36 #include "opt_vm.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/kernel.h>
  41 #include <sys/lock.h>
  42 #include <sys/malloc.h>
  43 #include <sys/rwlock.h>
  44 #include <sys/sysent.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/vdso.h>
  47
  48 #include <vm/vm.h>
  49 #include <vm/vm_param.h>
  50 #include <vm/pmap.h>
  51 #include <vm/vm_extern.h>
  52 #include <vm/vm_kern.h>
  53 #include <vm/vm_map.h>
  54 #include <vm/vm_object.h>
  55 #include <vm/vm_page.h>
  56 #include <vm/vm_pager.h>
  57
  58 static struct sx shared_page_alloc_sx;
  59 static vm_object_t shared_page_obj;
  60 static int shared_page_free;
  61 char *shared_page_mapping;
  62
  63 void
  64 shared_page_write(int base, int size, const void *data)
  65 {
  66
  67         bcopy(data, shared_page_mapping + base, size);
  68 }
  69
  70 static int
  71 shared_page_alloc_locked(int size, int align)
  72 {
  73         int res;
  74
  75         res = roundup(shared_page_free, align);
  76         if (res + size >= IDX_TO_OFF(shared_page_obj->size))
  77                 res = -1;
  78         else
  79                 shared_page_free = res + size;
  80         return (res);
  81 }
  82
  83 int
  84 shared_page_alloc(int size, int align)
  85 {
  86         int res;
  87
  88         sx_xlock(&shared_page_alloc_sx);
  89         res = shared_page_alloc_locked(size, align);
  90         sx_xunlock(&shared_page_alloc_sx);
  91         return (res);
  92 }
  93
  94 int
  95 shared_page_fill(int size, int align, const void *data)
  96 {
  97         int res;
  98
  99         sx_xlock(&shared_page_alloc_sx);
 100         res = shared_page_alloc_locked(size, align);
 101         if (res != -1)
 102                 shared_page_write(res, size, data);
 103         sx_xunlock(&shared_page_alloc_sx);
 104         return (res);
 105 }
 106
 107 static void
 108 shared_page_init(void *dummy __unused)
 109 {
 110         vm_page_t m;
 111         vm_offset_t addr;
 112
 113         sx_init(&shared_page_alloc_sx, "shpsx");
 114         shared_page_obj = vm_pager_allocate(OBJT_PHYS, 0, PAGE_SIZE,
 115             VM_PROT_DEFAULT, 0, NULL);
 116         VM_OBJECT_WLOCK(shared_page_obj);
 117         m = vm_page_grab(shared_page_obj, 0, VM_ALLOC_NOBUSY | VM_ALLOC_ZERO);
 118         m->valid = VM_PAGE_BITS_ALL;
 119         VM_OBJECT_WUNLOCK(shared_page_obj);
 120         addr = kva_alloc(PAGE_SIZE);
 121         pmap_qenter(addr, &m, 1);
 122         shared_page_mapping = (char *)addr;
 123 }
 124
 125 SYSINIT(shp, SI_SUB_EXEC, SI_ORDER_FIRST, (sysinit_cfunc_t)shared_page_init,
 126     NULL);
 127
 128 /*
 129  * Push the timehands update to the shared page.
 130  *
 131  * The lockless update scheme is similar to the one used to update the
 132  * in-kernel timehands, see sys/kern/kern_tc.c:tc_windup() (which
 133  * calls us after the timehands are updated).
 134  */
 135 static void
 136 timehands_update(struct vdso_sv_tk *svtk)
 137 {
 138         struct vdso_timehands th;
 139         struct vdso_timekeep *tk;
 140         uint32_t enabled, idx;
 141
 142         enabled = tc_fill_vdso_timehands(&th);
 143         th.th_gen = 0;
 144         idx = svtk->sv_timekeep_curr;
 145         if (++idx >= VDSO_TH_NUM)
 146                 idx = 0;
 147         svtk->sv_timekeep_curr = idx;
 148         if (++svtk->sv_timekeep_gen == 0)
 149                 svtk->sv_timekeep_gen = 1;
 150
 151         tk = (struct vdso_timekeep *)(shared_page_mapping +
 152             svtk->sv_timekeep_off);
 153         tk->tk_th[idx].th_gen = 0;
 154         atomic_thread_fence_rel();
 155         if (enabled)
 156                 tk->tk_th[idx] = th;
 157         atomic_store_rel_32(&tk->tk_th[idx].th_gen, svtk->sv_timekeep_gen);
 158         atomic_store_rel_32(&tk->tk_current, idx);
 159
 160         /*
 161          * The ordering of the assignment to tk_enabled relative to
 162          * the update of the vdso_timehands is not important.
 163          */
 164         tk->tk_enabled = enabled;
 165 }
 166
 167 #ifdef COMPAT_FREEBSD32
 168 static void
 169 timehands_update32(struct vdso_sv_tk *svtk)
 170 {
 171         struct vdso_timehands32 th;
 172         struct vdso_timekeep32 *tk;
 173         uint32_t enabled, idx;
 174
 175         enabled = tc_fill_vdso_timehands32(&th);
 176         th.th_gen = 0;
 177         idx = svtk->sv_timekeep_curr;
 178         if (++idx >= VDSO_TH_NUM)
 179                 idx = 0;
 180         svtk->sv_timekeep_curr = idx;
 181         if (++svtk->sv_timekeep_gen == 0)
 182                 svtk->sv_timekeep_gen = 1;
 183
 184         tk = (struct vdso_timekeep32 *)(shared_page_mapping +
 185             svtk->sv_timekeep_off);
 186         tk->tk_th[idx].th_gen = 0;
 187         atomic_thread_fence_rel();
 188         if (enabled)
 189                 tk->tk_th[idx] = th;
 190         atomic_store_rel_32(&tk->tk_th[idx].th_gen, svtk->sv_timekeep_gen);
 191         atomic_store_rel_32(&tk->tk_current, idx);
 192         tk->tk_enabled = enabled;
 193 }
 194 #endif
 195
 196 /*
 197  * This is hackish, but easiest way to avoid creating list structures
 198  * that needs to be iterated over from the hardclock interrupt
 199  * context.
 200  */
 201 static struct vdso_sv_tk *host_svtk;
 202 #ifdef COMPAT_FREEBSD32
 203 static struct vdso_sv_tk *compat32_svtk;
 204 #endif
 205
 206 void
 207 timekeep_push_vdso(void)
 208 {
 209
 210         if (host_svtk != NULL)
 211                 timehands_update(host_svtk);
 212 #ifdef COMPAT_FREEBSD32
 213         if (compat32_svtk != NULL)
 214                 timehands_update32(compat32_svtk);
 215 #endif
 216 }
 217
 218 struct vdso_sv_tk *
 219 alloc_sv_tk(void)
 220 {
 221         struct vdso_sv_tk *svtk;
 222         int tk_base;
 223         uint32_t tk_ver;
 224
 225         tk_ver = VDSO_TK_VER_CURR;
 226         svtk = malloc(sizeof(struct vdso_sv_tk), M_TEMP, M_WAITOK | M_ZERO);
 227         tk_base = shared_page_alloc(sizeof(struct vdso_timekeep) +
 228             sizeof(struct vdso_timehands) * VDSO_TH_NUM, 16);
 229         KASSERT(tk_base != -1, ("tk_base -1 for native"));
 230         shared_page_write(tk_base + offsetof(struct vdso_timekeep, tk_ver),
 231             sizeof(uint32_t), &tk_ver);
 232         svtk->sv_timekeep_off = tk_base;
 233         timekeep_push_vdso();
 234         return (svtk);
 235 }
 236
 237 #ifdef COMPAT_FREEBSD32
 238 struct vdso_sv_tk *
 239 alloc_sv_tk_compat32(void)
 240 {
 241         struct vdso_sv_tk *svtk;
 242         int tk_base;
 243         uint32_t tk_ver;
 244
 245         svtk = malloc(sizeof(struct vdso_sv_tk), M_TEMP, M_WAITOK | M_ZERO);
 246         tk_ver = VDSO_TK_VER_CURR;
 247         tk_base = shared_page_alloc(sizeof(struct vdso_timekeep32) +
 248             sizeof(struct vdso_timehands32) * VDSO_TH_NUM, 16);
 249         KASSERT(tk_base != -1, ("tk_base -1 for 32bit"));
 250         shared_page_write(tk_base + offsetof(struct vdso_timekeep32,
 251             tk_ver), sizeof(uint32_t), &tk_ver);
 252         svtk->sv_timekeep_off = tk_base;
 253         timekeep_push_vdso();
 254         return (svtk);
 255 }
 256 #endif
 257
 258 void
 259 exec_sysvec_init(void *param)
 260 {
 261         struct sysentvec *sv;
 262
 263         sv = (struct sysentvec *)param;
 264         if ((sv->sv_flags & SV_SHP) == 0)
 265                 return;
 266         sv->sv_shared_page_obj = shared_page_obj;
 267         sv->sv_sigcode_base = sv->sv_shared_page_base +
 268             shared_page_fill(*(sv->sv_szsigcode), 16, sv->sv_sigcode);
 269         if ((sv->sv_flags & SV_ABI_MASK) != SV_ABI_FREEBSD)
 270                 return;
 271         if ((sv->sv_flags & SV_TIMEKEEP) != 0) {
 272 #ifdef COMPAT_FREEBSD32
 273                 if ((sv->sv_flags & SV_ILP32) != 0) {
 274                         KASSERT(compat32_svtk == NULL,
 275                             ("Compat32 already registered"));
 276                         compat32_svtk = alloc_sv_tk_compat32();
 277                         sv->sv_timekeep_base = sv->sv_shared_page_base +
 278                             compat32_svtk->sv_timekeep_off;
 279                 } else {
 280 #endif
 281                         KASSERT(host_svtk == NULL, ("Host already registered"));
 282                         host_svtk = alloc_sv_tk();
 283                         sv->sv_timekeep_base = sv->sv_shared_page_base +
 284                             host_svtk->sv_timekeep_off;
 285 #ifdef COMPAT_FREEBSD32
 286                 }
 287 #endif
 288         }
 289 }