1 /* $NetBSD: uvm_device.c,v 1.80 2022/07/07 13:27:02 riastradh Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * All rights reserved. 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 ``AS IS'' AND ANY EXPRESS OR 17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 * 27 * from: Id: uvm_device.c,v 1.1.2.9 1998/02/06 05:11:47 chs Exp 28 */ 29 30 /* 31 * uvm_device.c: the device pager. 32 */ 33 34 #include <sys/cdefs.h> 35 __KERNEL_RCSID(0, "$NetBSD: uvm_device.c,v 1.80 2022/07/07 13:27:02 riastradh Exp $"); 36 37 #include "opt_uvmhist.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/conf.h> 42 #include <sys/proc.h> 43 #include <sys/kmem.h> 44 45 #include <uvm/uvm.h> 46 #include <uvm/uvm_device.h> 47 #include <uvm/uvm_pmap.h> 48 49 /* 50 * private global data structure 51 * 52 * we keep a list of active device objects in the system. 53 */ 54 55 LIST_HEAD(udv_list_struct, uvm_device); 56 static struct udv_list_struct udv_list; 57 static kmutex_t udv_lock __cacheline_aligned; 58 59 /* 60 * functions 61 */ 62 63 static void udv_init(void); 64 static void udv_reference(struct uvm_object *); 65 static void udv_detach(struct uvm_object *); 66 static int udv_fault(struct uvm_faultinfo *, vaddr_t, 67 struct vm_page **, int, int, vm_prot_t, 68 int); 69 70 /* 71 * master pager structure 72 */ 73 74 const struct uvm_pagerops uvm_deviceops = { 75 .pgo_init = udv_init, 76 .pgo_reference = udv_reference, 77 .pgo_detach = udv_detach, 78 .pgo_fault = udv_fault, 79 }; 80 81 /* 82 * the ops! 83 */ 84 85 /* 86 * udv_init 87 * 88 * init pager private data structures. 89 */ 90 91 static void 92 udv_init(void) 93 { 94 LIST_INIT(&udv_list); 95 mutex_init(&udv_lock, MUTEX_DEFAULT, IPL_NONE); 96 } 97 98 /* 99 * udv_attach 100 * 101 * get a VM object that is associated with a device. allocate a new 102 * one if needed. 103 * 104 * => caller must _not_ already be holding the lock on the uvm_object. 105 * => in fact, nothing should be locked so that we can sleep here. 106 */ 107 108 struct uvm_object * 109 udv_attach(dev_t device, vm_prot_t accessprot, 110 voff_t off, /* used only for access check */ 111 vsize_t size /* used only for access check */) 112 { 113 struct uvm_device *udv, *lcv; 114 const struct cdevsw *cdev; 115 dev_mmap_t *mapfn; 116 117 UVMHIST_FUNC(__func__); 118 UVMHIST_CALLARGS(maphist, "(device=%#jx)", device,0,0,0); 119 120 KASSERT(size > 0); 121 122 /* 123 * before we do anything, ensure this device supports mmap 124 */ 125 126 cdev = cdevsw_lookup(device); 127 if (cdev == NULL) { 128 return NULL; 129 } 130 mapfn = cdev->d_mmap; 131 if (mapfn == NULL || mapfn == nommap) { 132 return NULL; 133 } 134 135 /* 136 * Negative offsets on the object are not allowed, unless the 137 * device has affirmatively set D_NEGOFFSAFE. 138 */ 139 if ((cdev->d_flag & D_NEGOFFSAFE) == 0 && off != UVM_UNKNOWN_OFFSET) { 140 if (off < 0) 141 return NULL; 142 #if SIZE_MAX > UINT32_MAX /* XXX -Wtype-limits */ 143 if (size > __type_max(voff_t)) 144 return NULL; 145 #endif 146 if (off > __type_max(voff_t) - size) 147 return NULL; 148 } 149 150 /* 151 * Check that the specified range of the device allows the 152 * desired protection. 153 * 154 * XXX assumes VM_PROT_* == PROT_* 155 * XXX clobbers off and size, but nothing else here needs them. 156 */ 157 do { 158 KASSERTMSG((off % PAGE_SIZE) == 0, "off=%jd", (intmax_t)off); 159 KASSERTMSG(size >= PAGE_SIZE, "size=%"PRIuVSIZE, size); 160 if (cdev_mmap(device, off, accessprot) == -1) 161 return NULL; 162 KASSERT(off <= __type_max(voff_t) - PAGE_SIZE || 163 (cdev->d_flag & D_NEGOFFSAFE) != 0); 164 if (__predict_false(off > __type_max(voff_t) - PAGE_SIZE)) { 165 /* 166 * off += PAGE_SIZE, with two's-complement 167 * wraparound, or 168 * 169 * off += PAGE_SIZE - 2*(VOFF_MAX + 1). 170 */ 171 CTASSERT(MIN_PAGE_SIZE >= 2); 172 off -= __type_max(voff_t); 173 off += PAGE_SIZE - 2; 174 off -= __type_max(voff_t); 175 } else { 176 off += PAGE_SIZE; 177 } 178 size -= PAGE_SIZE; 179 } while (size != 0); 180 181 /* 182 * keep looping until we get it 183 */ 184 185 for (;;) { 186 187 /* 188 * first, attempt to find it on the main list 189 */ 190 191 mutex_enter(&udv_lock); 192 LIST_FOREACH(lcv, &udv_list, u_list) { 193 if (device == lcv->u_device) 194 break; 195 } 196 197 /* 198 * got it on main list. put a hold on it and unlock udv_lock. 199 */ 200 201 if (lcv) { 202 203 /* 204 * if someone else has a hold on it, sleep and start 205 * over again. 206 */ 207 208 if (lcv->u_flags & UVM_DEVICE_HOLD) { 209 lcv->u_flags |= UVM_DEVICE_WANTED; 210 UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, false, 211 "udv_attach",0); 212 continue; 213 } 214 215 /* we are now holding it */ 216 lcv->u_flags |= UVM_DEVICE_HOLD; 217 mutex_exit(&udv_lock); 218 219 /* 220 * bump reference count, unhold, return. 221 */ 222 223 rw_enter(lcv->u_obj.vmobjlock, RW_WRITER); 224 lcv->u_obj.uo_refs++; 225 rw_exit(lcv->u_obj.vmobjlock); 226 227 mutex_enter(&udv_lock); 228 if (lcv->u_flags & UVM_DEVICE_WANTED) 229 wakeup(lcv); 230 lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD); 231 mutex_exit(&udv_lock); 232 return &lcv->u_obj; 233 } 234 235 /* 236 * Did not find it on main list. Need to allocate a new one. 237 */ 238 239 mutex_exit(&udv_lock); 240 241 /* Note: both calls may allocate memory and sleep. */ 242 udv = kmem_alloc(sizeof(*udv), KM_SLEEP); 243 uvm_obj_init(&udv->u_obj, &uvm_deviceops, true, 1); 244 245 mutex_enter(&udv_lock); 246 247 /* 248 * now we have to double check to make sure no one added it 249 * to the list while we were sleeping... 250 */ 251 252 LIST_FOREACH(lcv, &udv_list, u_list) { 253 if (device == lcv->u_device) 254 break; 255 } 256 257 /* 258 * did we lose a race to someone else? 259 * free our memory and retry. 260 */ 261 262 if (lcv) { 263 mutex_exit(&udv_lock); 264 uvm_obj_destroy(&udv->u_obj, true); 265 kmem_free(udv, sizeof(*udv)); 266 continue; 267 } 268 269 /* 270 * we have it! init the data structures, add to list 271 * and return. 272 */ 273 274 udv->u_flags = 0; 275 udv->u_device = device; 276 LIST_INSERT_HEAD(&udv_list, udv, u_list); 277 mutex_exit(&udv_lock); 278 return &udv->u_obj; 279 } 280 /*NOTREACHED*/ 281 } 282 283 /* 284 * udv_reference 285 * 286 * add a reference to a VM object. Note that the reference count must 287 * already be one (the passed in reference) so there is no chance of the 288 * udv being released or locked out here. 289 * 290 * => caller must call with object unlocked. 291 */ 292 293 static void 294 udv_reference(struct uvm_object *uobj) 295 { 296 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 297 298 rw_enter(uobj->vmobjlock, RW_WRITER); 299 uobj->uo_refs++; 300 UVMHIST_LOG(maphist, "<- done (uobj=%#jx, ref = %jd)", 301 (uintptr_t)uobj, uobj->uo_refs,0,0); 302 rw_exit(uobj->vmobjlock); 303 } 304 305 /* 306 * udv_detach 307 * 308 * remove a reference to a VM object. 309 * 310 * => caller must call with object unlocked and map locked. 311 */ 312 313 static void 314 udv_detach(struct uvm_object *uobj) 315 { 316 struct uvm_device *udv = (struct uvm_device *)uobj; 317 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 318 319 /* 320 * loop until done 321 */ 322 again: 323 rw_enter(uobj->vmobjlock, RW_WRITER); 324 if (uobj->uo_refs > 1) { 325 uobj->uo_refs--; 326 rw_exit(uobj->vmobjlock); 327 UVMHIST_LOG(maphist," <- done, uobj=%#jx, ref=%jd", 328 (uintptr_t)uobj,uobj->uo_refs,0,0); 329 return; 330 } 331 332 /* 333 * is it being held? if so, wait until others are done. 334 */ 335 336 mutex_enter(&udv_lock); 337 if (udv->u_flags & UVM_DEVICE_HOLD) { 338 udv->u_flags |= UVM_DEVICE_WANTED; 339 rw_exit(uobj->vmobjlock); 340 UVM_UNLOCK_AND_WAIT(udv, &udv_lock, false, "udv_detach",0); 341 goto again; 342 } 343 344 /* 345 * got it! nuke it now. 346 */ 347 348 LIST_REMOVE(udv, u_list); 349 if (udv->u_flags & UVM_DEVICE_WANTED) 350 wakeup(udv); 351 mutex_exit(&udv_lock); 352 rw_exit(uobj->vmobjlock); 353 354 uvm_obj_destroy(uobj, true); 355 kmem_free(udv, sizeof(*udv)); 356 UVMHIST_LOG(maphist," <- done, freed uobj=%#jx", (uintptr_t)uobj, 357 0, 0, 0); 358 } 359 360 /* 361 * udv_fault: non-standard fault routine for device "pages" 362 * 363 * => rather than having a "get" function, we have a fault routine 364 * since we don't return vm_pages we need full control over the 365 * pmap_enter map in 366 * => all the usual fault data structured are locked by the caller 367 * (i.e. maps(read), amap (if any), uobj) 368 * => on return, we unlock all fault data structures 369 * => flags: PGO_ALLPAGES: get all of the pages 370 * PGO_LOCKED: fault data structures are locked 371 * XXX: currently PGO_LOCKED is always required ... consider removing 372 * it as a flag 373 * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx] 374 */ 375 376 static int 377 udv_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr, struct vm_page **pps, 378 int npages, int centeridx, vm_prot_t access_type, 379 int flags) 380 { 381 struct vm_map_entry *entry = ufi->entry; 382 struct uvm_object *uobj = entry->object.uvm_obj; 383 struct uvm_device *udv = (struct uvm_device *)uobj; 384 vaddr_t curr_va; 385 off_t curr_offset; 386 paddr_t paddr, mdpgno; 387 u_int mmapflags; 388 int lcv, retval; 389 dev_t device; 390 vm_prot_t mapprot; 391 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 392 UVMHIST_LOG(maphist," flags=%#jx", flags,0,0,0); 393 394 /* 395 * we do not allow device mappings to be mapped copy-on-write 396 * so we kill any attempt to do so here. 397 */ 398 399 if (UVM_ET_ISCOPYONWRITE(entry)) { 400 UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=%#jx)", 401 entry->etype, 0,0,0); 402 uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj); 403 return EIO; 404 } 405 406 /* 407 * get device map function. 408 */ 409 410 device = udv->u_device; 411 if (cdevsw_lookup(device) == NULL) { 412 /* XXX This should not happen */ 413 uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj); 414 return EIO; 415 } 416 417 /* 418 * now we must determine the offset in udv to use and the VA to 419 * use for pmap_enter. note that we always use orig_map's pmap 420 * for pmap_enter (even if we have a submap). since virtual 421 * addresses in a submap must match the main map, this is ok. 422 */ 423 424 /* udv offset = (offset from start of entry) + entry's offset */ 425 curr_offset = entry->offset + (vaddr - entry->start); 426 /* pmap va = vaddr (virtual address of pps[0]) */ 427 curr_va = vaddr; 428 429 /* 430 * loop over the page range entering in as needed 431 */ 432 433 retval = 0; 434 for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE, 435 curr_va += PAGE_SIZE) { 436 if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx) 437 continue; 438 439 if (pps[lcv] == PGO_DONTCARE) 440 continue; 441 442 mdpgno = cdev_mmap(device, curr_offset, access_type); 443 if (mdpgno == -1) { 444 retval = EIO; 445 break; 446 } 447 paddr = pmap_phys_address(mdpgno); 448 mmapflags = pmap_mmap_flags(mdpgno); 449 mapprot = ufi->entry->protection; 450 UVMHIST_LOG(maphist, 451 " MAPPING: device: pm=%#jx, va=%#jx, pa=%#jx, at=%jd", 452 (uintptr_t)ufi->orig_map->pmap, curr_va, paddr, mapprot); 453 if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr, mapprot, 454 PMAP_CANFAIL | mapprot | mmapflags) != 0) { 455 /* 456 * pmap_enter() didn't have the resource to 457 * enter this mapping. Unlock everything, 458 * wait for the pagedaemon to free up some 459 * pages, and then tell uvm_fault() to start 460 * the fault again. 461 * 462 * XXX Needs some rethinking for the PGO_ALLPAGES 463 * XXX case. 464 */ 465 pmap_update(ufi->orig_map->pmap); /* sync what we have so far */ 466 uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, 467 uobj); 468 return ENOMEM; 469 } 470 } 471 472 pmap_update(ufi->orig_map->pmap); 473 uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj); 474 return retval; 475 } 476