sys/vm/device_pager.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1990 University of Utah.
   5  * Copyright (c) 1991, 1993
   6  *      The Regents of the University of California.  All rights reserved.
   7  *
   8  * This code is derived from software contributed to Berkeley by
   9  * the Systems Programming Group of the University of Utah Computer
  10  * Science Department.
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  * 1. Redistributions of source code must retain the above copyright
  16  *    notice, this list of conditions and the following disclaimer.
  17  * 2. Redistributions in binary form must reproduce the above copyright
  18  *    notice, this list of conditions and the following disclaimer in the
  19  *    documentation and/or other materials provided with the distribution.
  20  * 3. Neither the name of the University nor the names of its contributors
  21  *    may be used to endorse or promote products derived from this software
  22  *    without specific prior written permission.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  *
  36  *      @(#)device_pager.c      8.1 (Berkeley) 6/11/93
  37  */
  38
  39 #include <sys/cdefs.h>
  40 #include <sys/param.h>
  41 #include <sys/systm.h>
  42 #include <sys/conf.h>
  43 #include <sys/lock.h>
  44 #include <sys/proc.h>
  45 #include <sys/mutex.h>
  46 #include <sys/mman.h>
  47 #include <sys/rwlock.h>
  48 #include <sys/sx.h>
  49 #include <sys/user.h>
  50 #include <sys/vmmeter.h>
  51
  52 #include <vm/vm.h>
  53 #include <vm/vm_param.h>
  54 #include <vm/vm_object.h>
  55 #include <vm/vm_page.h>
  56 #include <vm/vm_pager.h>
  57 #include <vm/vm_phys.h>
  58 #include <vm/uma.h>
  59
  60 static void dev_pager_init(void);
  61 static vm_object_t dev_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
  62     vm_ooffset_t, struct ucred *);
  63 static void dev_pager_dealloc(vm_object_t);
  64 static int dev_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
  65 static void dev_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
  66 static boolean_t dev_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
  67 static void dev_pager_free_page(vm_object_t object, vm_page_t m);
  68 static int dev_pager_populate(vm_object_t object, vm_pindex_t pidx,
  69     int fault_type, vm_prot_t, vm_pindex_t *first, vm_pindex_t *last);
  70
  71 /* list of device pager objects */
  72 static struct pagerlst dev_pager_object_list;
  73 /* protect list manipulation */
  74 static struct mtx dev_pager_mtx;
  75
  76 const struct pagerops devicepagerops = {
  77         .pgo_kvme_type = KVME_TYPE_DEVICE,
  78         .pgo_init =     dev_pager_init,
  79         .pgo_alloc =    dev_pager_alloc,
  80         .pgo_dealloc =  dev_pager_dealloc,
  81         .pgo_getpages = dev_pager_getpages,
  82         .pgo_putpages = dev_pager_putpages,
  83         .pgo_haspage =  dev_pager_haspage,
  84 };
  85
  86 const struct pagerops mgtdevicepagerops = {
  87         .pgo_kvme_type = KVME_TYPE_MGTDEVICE,
  88         .pgo_alloc =    dev_pager_alloc,
  89         .pgo_dealloc =  dev_pager_dealloc,
  90         .pgo_getpages = dev_pager_getpages,
  91         .pgo_putpages = dev_pager_putpages,
  92         .pgo_haspage =  dev_pager_haspage,
  93         .pgo_populate = dev_pager_populate,
  94 };
  95
  96 static int old_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
  97     vm_ooffset_t foff, struct ucred *cred, u_short *color);
  98 static void old_dev_pager_dtor(void *handle);
  99 static int old_dev_pager_fault(vm_object_t object, vm_ooffset_t offset,
 100     int prot, vm_page_t *mres);
 101
 102 static const struct cdev_pager_ops old_dev_pager_ops = {
 103         .cdev_pg_ctor = old_dev_pager_ctor,
 104         .cdev_pg_dtor = old_dev_pager_dtor,
 105         .cdev_pg_fault = old_dev_pager_fault
 106 };
 107
 108 static void
 109 dev_pager_init(void)
 110 {
 111
 112         TAILQ_INIT(&dev_pager_object_list);
 113         mtx_init(&dev_pager_mtx, "dev_pager list", NULL, MTX_DEF);
 114 }
 115
 116 vm_object_t
 117 cdev_pager_lookup(void *handle)
 118 {
 119         vm_object_t object;
 120
 121         mtx_lock(&dev_pager_mtx);
 122         object = vm_pager_object_lookup(&dev_pager_object_list, handle);
 123         mtx_unlock(&dev_pager_mtx);
 124         return (object);
 125 }
 126
 127 vm_object_t
 128 cdev_pager_allocate(void *handle, enum obj_type tp,
 129     const struct cdev_pager_ops *ops, vm_ooffset_t size, vm_prot_t prot,
 130     vm_ooffset_t foff, struct ucred *cred)
 131 {
 132         vm_object_t object, object1;
 133         vm_pindex_t pindex;
 134         u_short color;
 135
 136         if (tp != OBJT_DEVICE && tp != OBJT_MGTDEVICE)
 137                 return (NULL);
 138         KASSERT(tp == OBJT_MGTDEVICE || ops->cdev_pg_populate == NULL,
 139             ("populate on unmanaged device pager"));
 140
 141         /*
 142          * Offset should be page aligned.
 143          */
 144         if (foff & PAGE_MASK)
 145                 return (NULL);
 146
 147         /*
 148          * Treat the mmap(2) file offset as an unsigned value for a
 149          * device mapping.  This, in effect, allows a user to pass all
 150          * possible off_t values as the mapping cookie to the driver.  At
 151          * this point, we know that both foff and size are a multiple
 152          * of the page size.  Do a check to avoid wrap.
 153          */
 154         size = round_page(size);
 155         pindex = OFF_TO_IDX(foff) + OFF_TO_IDX(size);
 156         if (pindex > OBJ_MAX_SIZE || pindex < OFF_TO_IDX(foff) ||
 157             pindex < OFF_TO_IDX(size))
 158                 return (NULL);
 159
 160         if (ops->cdev_pg_ctor(handle, size, prot, foff, cred, &color) != 0)
 161                 return (NULL);
 162         mtx_lock(&dev_pager_mtx);
 163
 164         /*
 165          * Look up pager, creating as necessary.
 166          */
 167         object1 = NULL;
 168         object = vm_pager_object_lookup(&dev_pager_object_list, handle);
 169         if (object == NULL) {
 170                 /*
 171                  * Allocate object and associate it with the pager.  Initialize
 172                  * the object's pg_color based upon the physical address of the
 173                  * device's memory.
 174                  */
 175                 mtx_unlock(&dev_pager_mtx);
 176                 object1 = vm_object_allocate(tp, pindex);
 177                 object1->flags |= OBJ_COLORED;
 178                 object1->pg_color = color;
 179                 object1->handle = handle;
 180                 object1->un_pager.devp.ops = ops;
 181                 object1->un_pager.devp.dev = handle;
 182                 TAILQ_INIT(&object1->un_pager.devp.devp_pglist);
 183                 mtx_lock(&dev_pager_mtx);
 184                 object = vm_pager_object_lookup(&dev_pager_object_list, handle);
 185                 if (object != NULL) {
 186                         /*
 187                          * We raced with other thread while allocating object.
 188                          */
 189                         if (pindex > object->size)
 190                                 object->size = pindex;
 191                         KASSERT(object->type == tp,
 192                             ("Inconsistent device pager type %p %d",
 193                             object, tp));
 194                         KASSERT(object->un_pager.devp.ops == ops,
 195                             ("Inconsistent devops %p %p", object, ops));
 196                 } else {
 197                         object = object1;
 198                         object1 = NULL;
 199                         object->handle = handle;
 200                         TAILQ_INSERT_TAIL(&dev_pager_object_list, object,
 201                             pager_object_list);
 202                         if (ops->cdev_pg_populate != NULL)
 203                                 vm_object_set_flag(object, OBJ_POPULATE);
 204                 }
 205         } else {
 206                 if (pindex > object->size)
 207                         object->size = pindex;
 208                 KASSERT(object->type == tp,
 209                     ("Inconsistent device pager type %p %d", object, tp));
 210         }
 211         mtx_unlock(&dev_pager_mtx);
 212         if (object1 != NULL) {
 213                 object1->handle = object1;
 214                 mtx_lock(&dev_pager_mtx);
 215                 TAILQ_INSERT_TAIL(&dev_pager_object_list, object1,
 216                     pager_object_list);
 217                 mtx_unlock(&dev_pager_mtx);
 218                 vm_object_deallocate(object1);
 219         }
 220         return (object);
 221 }
 222
 223 static vm_object_t
 224 dev_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 225     vm_ooffset_t foff, struct ucred *cred)
 226 {
 227
 228         return (cdev_pager_allocate(handle, OBJT_DEVICE, &old_dev_pager_ops,
 229             size, prot, foff, cred));
 230 }
 231
 232 void
 233 cdev_pager_free_page(vm_object_t object, vm_page_t m)
 234 {
 235
 236         VM_OBJECT_ASSERT_WLOCKED(object);
 237         if (object->type == OBJT_MGTDEVICE) {
 238                 KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("unmanaged %p", m));
 239                 pmap_remove_all(m);
 240                 (void)vm_page_remove(m);
 241         } else if (object->type == OBJT_DEVICE)
 242                 dev_pager_free_page(object, m);
 243 }
 244
 245 static void
 246 dev_pager_free_page(vm_object_t object, vm_page_t m)
 247 {
 248
 249         VM_OBJECT_ASSERT_WLOCKED(object);
 250         KASSERT((object->type == OBJT_DEVICE &&
 251             (m->oflags & VPO_UNMANAGED) != 0),
 252             ("Managed device or page obj %p m %p", object, m));
 253         TAILQ_REMOVE(&object->un_pager.devp.devp_pglist, m, plinks.q);
 254         vm_page_putfake(m);
 255 }
 256
 257 static void
 258 dev_pager_dealloc(vm_object_t object)
 259 {
 260         vm_page_t m;
 261
 262         VM_OBJECT_WUNLOCK(object);
 263         object->un_pager.devp.ops->cdev_pg_dtor(object->un_pager.devp.dev);
 264
 265         mtx_lock(&dev_pager_mtx);
 266         TAILQ_REMOVE(&dev_pager_object_list, object, pager_object_list);
 267         mtx_unlock(&dev_pager_mtx);
 268         VM_OBJECT_WLOCK(object);
 269
 270         if (object->type == OBJT_DEVICE) {
 271                 /*
 272                  * Free up our fake pages.
 273                  */
 274                 while ((m = TAILQ_FIRST(&object->un_pager.devp.devp_pglist))
 275                     != NULL) {
 276                         if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0)
 277                                 continue;
 278
 279                         dev_pager_free_page(object, m);
 280                 }
 281         }
 282         object->handle = NULL;
 283         object->type = OBJT_DEAD;
 284 }
 285
 286 static int
 287 dev_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
 288     int *rahead)
 289 {
 290         int error;
 291
 292         /* Since our haspage reports zero after/before, the count is 1. */
 293         KASSERT(count == 1, ("%s: count %d", __func__, count));
 294         if (object->un_pager.devp.ops->cdev_pg_fault == NULL)
 295                 return (VM_PAGER_FAIL);
 296         VM_OBJECT_WLOCK(object);
 297         error = object->un_pager.devp.ops->cdev_pg_fault(object,
 298             IDX_TO_OFF(ma[0]->pindex), PROT_READ, &ma[0]);
 299
 300         VM_OBJECT_ASSERT_WLOCKED(object);
 301
 302         if (error == VM_PAGER_OK) {
 303                 KASSERT((object->type == OBJT_DEVICE &&
 304                      (ma[0]->oflags & VPO_UNMANAGED) != 0) ||
 305                     (object->type == OBJT_MGTDEVICE &&
 306                      (ma[0]->oflags & VPO_UNMANAGED) == 0),
 307                     ("Wrong page type %p %p", ma[0], object));
 308                 if (object->type == OBJT_DEVICE) {
 309                         TAILQ_INSERT_TAIL(&object->un_pager.devp.devp_pglist,
 310                             ma[0], plinks.q);
 311                 }
 312                 if (rbehind)
 313                         *rbehind = 0;
 314                 if (rahead)
 315                         *rahead = 0;
 316         }
 317         VM_OBJECT_WUNLOCK(object);
 318
 319         return (error);
 320 }
 321
 322 static int
 323 dev_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type,
 324     vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
 325 {
 326
 327         VM_OBJECT_ASSERT_WLOCKED(object);
 328         if (object->un_pager.devp.ops->cdev_pg_populate == NULL)
 329                 return (VM_PAGER_FAIL);
 330         return (object->un_pager.devp.ops->cdev_pg_populate(object, pidx,
 331             fault_type, max_prot, first, last));
 332 }
 333
 334 static int
 335 old_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, int prot,
 336     vm_page_t *mres)
 337 {
 338         vm_paddr_t paddr;
 339         vm_page_t m_paddr, page;
 340         struct cdev *dev;
 341         struct cdevsw *csw;
 342         struct file *fpop;
 343         struct thread *td;
 344         vm_memattr_t memattr, memattr1;
 345         int ref, ret;
 346
 347         memattr = object->memattr;
 348
 349         VM_OBJECT_WUNLOCK(object);
 350
 351         dev = object->handle;
 352         csw = dev_refthread(dev, &ref);
 353         if (csw == NULL) {
 354                 VM_OBJECT_WLOCK(object);
 355                 return (VM_PAGER_FAIL);
 356         }
 357         td = curthread;
 358         fpop = td->td_fpop;
 359         td->td_fpop = NULL;
 360         ret = csw->d_mmap(dev, offset, &paddr, prot, &memattr);
 361         td->td_fpop = fpop;
 362         dev_relthread(dev, ref);
 363         if (ret != 0) {
 364                 printf(
 365             "WARNING: dev_pager_getpage: map function returns error %d", ret);
 366                 VM_OBJECT_WLOCK(object);
 367                 return (VM_PAGER_FAIL);
 368         }
 369
 370         /* If "paddr" is a real page, perform a sanity check on "memattr". */
 371         if ((m_paddr = vm_phys_paddr_to_vm_page(paddr)) != NULL &&
 372             (memattr1 = pmap_page_get_memattr(m_paddr)) != memattr) {
 373                 /*
 374                  * For the /dev/mem d_mmap routine to return the
 375                  * correct memattr, pmap_page_get_memattr() needs to
 376                  * be called, which we do there.
 377                  */
 378                 if ((csw->d_flags & D_MEM) == 0) {
 379                         printf("WARNING: Device driver %s has set "
 380                             "\"memattr\" inconsistently (drv %u pmap %u).\n",
 381                             csw->d_name, memattr, memattr1);
 382                 }
 383                 memattr = memattr1;
 384         }
 385         if (((*mres)->flags & PG_FICTITIOUS) != 0) {
 386                 /*
 387                  * If the passed in result page is a fake page, update it with
 388                  * the new physical address.
 389                  */
 390                 page = *mres;
 391                 VM_OBJECT_WLOCK(object);
 392                 vm_page_updatefake(page, paddr, memattr);
 393         } else {
 394                 /*
 395                  * Replace the passed in reqpage page with our own fake page and
 396                  * free up the all of the original pages.
 397                  */
 398                 page = vm_page_getfake(paddr, memattr);
 399                 VM_OBJECT_WLOCK(object);
 400                 vm_page_replace(page, object, (*mres)->pindex, *mres);
 401                 *mres = page;
 402         }
 403         vm_page_valid(page);
 404         return (VM_PAGER_OK);
 405 }
 406
 407 static void
 408 dev_pager_putpages(vm_object_t object, vm_page_t *m, int count, int flags,
 409     int *rtvals)
 410 {
 411
 412         panic("dev_pager_putpage called");
 413 }
 414
 415 static boolean_t
 416 dev_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 417     int *after)
 418 {
 419
 420         if (before != NULL)
 421                 *before = 0;
 422         if (after != NULL)
 423                 *after = 0;
 424         return (TRUE);
 425 }
 426
 427 static int
 428 old_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
 429     vm_ooffset_t foff, struct ucred *cred, u_short *color)
 430 {
 431         struct cdev *dev;
 432         struct cdevsw *csw;
 433         vm_memattr_t dummy;
 434         vm_ooffset_t off;
 435         vm_paddr_t paddr;
 436         unsigned int npages;
 437         int ref;
 438
 439         /*
 440          * Make sure this device can be mapped.
 441          */
 442         dev = handle;
 443         csw = dev_refthread(dev, &ref);
 444         if (csw == NULL)
 445                 return (ENXIO);
 446
 447         /*
 448          * Check that the specified range of the device allows the desired
 449          * protection.
 450          *
 451          * XXX assumes VM_PROT_* == PROT_*
 452          */
 453         npages = OFF_TO_IDX(size);
 454         paddr = 0; /* Make paddr initialized for the case of size == 0. */
 455         for (off = foff; npages--; off += PAGE_SIZE) {
 456                 if (csw->d_mmap(dev, off, &paddr, (int)prot, &dummy) != 0) {
 457                         dev_relthread(dev, ref);
 458                         return (EINVAL);
 459                 }
 460         }
 461
 462         dev_ref(dev);
 463         dev_relthread(dev, ref);
 464         *color = atop(paddr) - OFF_TO_IDX(off - PAGE_SIZE);
 465         return (0);
 466 }
 467
 468 static void
 469 old_dev_pager_dtor(void *handle)
 470 {
 471
 472         dev_rel(handle);
 473 }