1 /*- 2 * Copyright (c) 2010 Isilon Systems, Inc. 3 * Copyright (c) 2016 Matthew Macy (mmacy@mattmacy.io) 4 * Copyright (c) 2017 Mellanox Technologies, Ltd. 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 unmodified, this list of conditions, and the following 12 * disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29 #include <sys/param.h> 30 #include <sys/systm.h> 31 #include <sys/malloc.h> 32 #include <sys/kernel.h> 33 #include <sys/sysctl.h> 34 #include <sys/lock.h> 35 #include <sys/mutex.h> 36 #include <sys/rwlock.h> 37 #include <sys/proc.h> 38 #include <sys/sched.h> 39 #include <sys/memrange.h> 40 41 #include <machine/bus.h> 42 43 #include <vm/vm.h> 44 #include <vm/pmap.h> 45 #include <vm/vm_param.h> 46 #include <vm/vm_kern.h> 47 #include <vm/vm_object.h> 48 #include <vm/vm_map.h> 49 #include <vm/vm_page.h> 50 #include <vm/vm_pageout.h> 51 #include <vm/vm_pager.h> 52 #include <vm/vm_radix.h> 53 #include <vm/vm_reserv.h> 54 #include <vm/vm_extern.h> 55 56 #include <vm/uma.h> 57 #include <vm/uma_int.h> 58 59 #include <linux/gfp.h> 60 #include <linux/mm.h> 61 #include <linux/preempt.h> 62 #include <linux/fs.h> 63 #include <linux/shmem_fs.h> 64 #include <linux/kernel.h> 65 #include <linux/idr.h> 66 #include <linux/io.h> 67 #include <linux/io-mapping.h> 68 69 #ifdef __i386__ 70 DEFINE_IDR(mtrr_idr); 71 static MALLOC_DEFINE(M_LKMTRR, "idr", "Linux MTRR compat"); 72 extern int pat_works; 73 #endif 74 75 void 76 si_meminfo(struct sysinfo *si) 77 { 78 si->totalram = physmem; 79 si->freeram = vm_free_count(); 80 si->totalhigh = 0; 81 si->freehigh = 0; 82 si->mem_unit = PAGE_SIZE; 83 } 84 85 void * 86 linux_page_address(const struct page *page) 87 { 88 89 if (page->object != kernel_object) { 90 return (PMAP_HAS_DMAP ? 91 ((void *)(uintptr_t)PHYS_TO_DMAP(page_to_phys(page))) : 92 NULL); 93 } 94 return ((void *)(uintptr_t)(VM_MIN_KERNEL_ADDRESS + 95 IDX_TO_OFF(page->pindex))); 96 } 97 98 struct page * 99 linux_alloc_pages(gfp_t flags, unsigned int order) 100 { 101 struct page *page; 102 103 if (PMAP_HAS_DMAP) { 104 unsigned long npages = 1UL << order; 105 int req = VM_ALLOC_WIRED; 106 107 if ((flags & M_ZERO) != 0) 108 req |= VM_ALLOC_ZERO; 109 if (order == 0 && (flags & GFP_DMA32) == 0) { 110 page = vm_page_alloc_noobj(req); 111 if (page == NULL) 112 return (NULL); 113 } else { 114 vm_paddr_t pmax = (flags & GFP_DMA32) ? 115 BUS_SPACE_MAXADDR_32BIT : BUS_SPACE_MAXADDR; 116 retry: 117 page = vm_page_alloc_noobj_contig(req, npages, 0, pmax, 118 PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); 119 if (page == NULL) { 120 if ((flags & (M_WAITOK | __GFP_NORETRY)) == 121 M_WAITOK) { 122 int err = vm_page_reclaim_contig(req, 123 npages, 0, pmax, PAGE_SIZE, 0); 124 if (err == ENOMEM) 125 vm_wait(NULL); 126 else if (err != 0) 127 return (NULL); 128 flags &= ~M_WAITOK; 129 goto retry; 130 } 131 return (NULL); 132 } 133 } 134 } else { 135 vm_offset_t vaddr; 136 137 vaddr = linux_alloc_kmem(flags, order); 138 if (vaddr == 0) 139 return (NULL); 140 141 page = virt_to_page((void *)vaddr); 142 143 KASSERT(vaddr == (vm_offset_t)page_address(page), 144 ("Page address mismatch")); 145 } 146 147 return (page); 148 } 149 150 static void 151 _linux_free_kmem(vm_offset_t addr, unsigned int order) 152 { 153 size_t size = ((size_t)PAGE_SIZE) << order; 154 155 kmem_free((void *)addr, size); 156 } 157 158 void 159 linux_free_pages(struct page *page, unsigned int order) 160 { 161 if (PMAP_HAS_DMAP) { 162 unsigned long npages = 1UL << order; 163 unsigned long x; 164 165 for (x = 0; x != npages; x++) { 166 vm_page_t pgo = page + x; 167 168 if (vm_page_unwire_noq(pgo)) 169 vm_page_free(pgo); 170 } 171 } else { 172 vm_offset_t vaddr; 173 174 vaddr = (vm_offset_t)page_address(page); 175 176 _linux_free_kmem(vaddr, order); 177 } 178 } 179 180 void 181 linux_release_pages(struct page **pages, int nr) 182 { 183 int i; 184 185 for (i = 0; i < nr; i++) 186 put_page(pages[i]); 187 } 188 189 vm_offset_t 190 linux_alloc_kmem(gfp_t flags, unsigned int order) 191 { 192 size_t size = ((size_t)PAGE_SIZE) << order; 193 void *addr; 194 195 addr = kmem_alloc_contig(size, flags & GFP_NATIVE_MASK, 0, 196 ((flags & GFP_DMA32) == 0) ? -1UL : BUS_SPACE_MAXADDR_32BIT, 197 PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); 198 199 return ((vm_offset_t)addr); 200 } 201 202 void 203 linux_free_kmem(vm_offset_t addr, unsigned int order) 204 { 205 KASSERT((addr & ~PAGE_MASK) == 0, 206 ("%s: addr %p is not page aligned", __func__, (void *)addr)); 207 208 if (addr >= VM_MIN_KERNEL_ADDRESS && addr < VM_MAX_KERNEL_ADDRESS) { 209 _linux_free_kmem(addr, order); 210 } else { 211 vm_page_t page; 212 213 page = PHYS_TO_VM_PAGE(DMAP_TO_PHYS(addr)); 214 linux_free_pages(page, order); 215 } 216 } 217 218 static int 219 linux_get_user_pages_internal(vm_map_t map, unsigned long start, int nr_pages, 220 int write, struct page **pages) 221 { 222 vm_prot_t prot; 223 size_t len; 224 int count; 225 226 prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ; 227 len = ptoa((vm_offset_t)nr_pages); 228 count = vm_fault_quick_hold_pages(map, start, len, prot, pages, nr_pages); 229 return (count == -1 ? -EFAULT : nr_pages); 230 } 231 232 int 233 __get_user_pages_fast(unsigned long start, int nr_pages, int write, 234 struct page **pages) 235 { 236 vm_map_t map; 237 vm_page_t *mp; 238 vm_offset_t va; 239 vm_offset_t end; 240 vm_prot_t prot; 241 int count; 242 243 if (nr_pages == 0 || in_interrupt()) 244 return (0); 245 246 MPASS(pages != NULL); 247 map = &curthread->td_proc->p_vmspace->vm_map; 248 end = start + ptoa((vm_offset_t)nr_pages); 249 if (!vm_map_range_valid(map, start, end)) 250 return (-EINVAL); 251 prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ; 252 for (count = 0, mp = pages, va = start; va < end; 253 mp++, va += PAGE_SIZE, count++) { 254 *mp = pmap_extract_and_hold(map->pmap, va, prot); 255 if (*mp == NULL) 256 break; 257 258 if ((prot & VM_PROT_WRITE) != 0 && 259 (*mp)->dirty != VM_PAGE_BITS_ALL) { 260 /* 261 * Explicitly dirty the physical page. Otherwise, the 262 * caller's changes may go unnoticed because they are 263 * performed through an unmanaged mapping or by a DMA 264 * operation. 265 * 266 * The object lock is not held here. 267 * See vm_page_clear_dirty_mask(). 268 */ 269 vm_page_dirty(*mp); 270 } 271 } 272 return (count); 273 } 274 275 long 276 get_user_pages_remote(struct task_struct *task, struct mm_struct *mm, 277 unsigned long start, unsigned long nr_pages, unsigned int gup_flags, 278 struct page **pages, struct vm_area_struct **vmas) 279 { 280 vm_map_t map; 281 282 map = &task->task_thread->td_proc->p_vmspace->vm_map; 283 return (linux_get_user_pages_internal(map, start, nr_pages, 284 !!(gup_flags & FOLL_WRITE), pages)); 285 } 286 287 long 288 lkpi_get_user_pages(unsigned long start, unsigned long nr_pages, 289 unsigned int gup_flags, struct page **pages) 290 { 291 vm_map_t map; 292 293 map = &curthread->td_proc->p_vmspace->vm_map; 294 return (linux_get_user_pages_internal(map, start, nr_pages, 295 !!(gup_flags & FOLL_WRITE), pages)); 296 } 297 298 int 299 is_vmalloc_addr(const void *addr) 300 { 301 return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL); 302 } 303 304 vm_fault_t 305 lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr, 306 unsigned long pfn, pgprot_t prot) 307 { 308 struct pctrie_iter pages; 309 vm_object_t vm_obj = vma->vm_obj; 310 vm_object_t tmp_obj; 311 vm_page_t page; 312 vm_pindex_t pindex; 313 314 VM_OBJECT_ASSERT_WLOCKED(vm_obj); 315 vm_page_iter_init(&pages, vm_obj); 316 pindex = OFF_TO_IDX(addr - vma->vm_start); 317 if (vma->vm_pfn_count == 0) 318 vma->vm_pfn_first = pindex; 319 MPASS(pindex <= OFF_TO_IDX(vma->vm_end)); 320 321 retry: 322 page = vm_page_grab_iter(vm_obj, pindex, VM_ALLOC_NOCREAT, &pages); 323 if (page == NULL) { 324 page = PHYS_TO_VM_PAGE(IDX_TO_OFF(pfn)); 325 if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) { 326 pctrie_iter_reset(&pages); 327 goto retry; 328 } 329 if (page->object != NULL) { 330 tmp_obj = page->object; 331 vm_page_xunbusy(page); 332 VM_OBJECT_WUNLOCK(vm_obj); 333 VM_OBJECT_WLOCK(tmp_obj); 334 if (page->object == tmp_obj && 335 vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) { 336 KASSERT(page->object == tmp_obj, 337 ("page has changed identity")); 338 KASSERT((page->oflags & VPO_UNMANAGED) == 0, 339 ("page does not belong to shmem")); 340 vm_pager_page_unswapped(page); 341 if (pmap_page_is_mapped(page)) { 342 vm_page_xunbusy(page); 343 VM_OBJECT_WUNLOCK(tmp_obj); 344 printf("%s: page rename failed: page " 345 "is mapped\n", __func__); 346 VM_OBJECT_WLOCK(vm_obj); 347 return (VM_FAULT_NOPAGE); 348 } 349 vm_page_remove(page); 350 } 351 VM_OBJECT_WUNLOCK(tmp_obj); 352 pctrie_iter_reset(&pages); 353 VM_OBJECT_WLOCK(vm_obj); 354 goto retry; 355 } 356 if (vm_page_iter_insert(page, vm_obj, pindex, &pages) != 0) { 357 vm_page_xunbusy(page); 358 return (VM_FAULT_OOM); 359 } 360 vm_page_valid(page); 361 } 362 pmap_page_set_memattr(page, pgprot2cachemode(prot)); 363 vma->vm_pfn_count++; 364 365 return (VM_FAULT_NOPAGE); 366 } 367 368 int 369 lkpi_remap_pfn_range(struct vm_area_struct *vma, unsigned long start_addr, 370 unsigned long start_pfn, unsigned long size, pgprot_t prot) 371 { 372 vm_object_t vm_obj; 373 unsigned long addr, pfn; 374 int err = 0; 375 376 vm_obj = vma->vm_obj; 377 378 VM_OBJECT_WLOCK(vm_obj); 379 for (addr = start_addr, pfn = start_pfn; 380 addr < start_addr + size; 381 addr += PAGE_SIZE) { 382 vm_fault_t ret; 383 retry: 384 ret = lkpi_vmf_insert_pfn_prot_locked(vma, addr, pfn, prot); 385 386 if ((ret & VM_FAULT_OOM) != 0) { 387 VM_OBJECT_WUNLOCK(vm_obj); 388 vm_wait(NULL); 389 VM_OBJECT_WLOCK(vm_obj); 390 goto retry; 391 } 392 393 if ((ret & VM_FAULT_ERROR) != 0) { 394 err = -EFAULT; 395 break; 396 } 397 398 pfn++; 399 } 400 VM_OBJECT_WUNLOCK(vm_obj); 401 402 if (unlikely(err)) { 403 zap_vma_ptes(vma, start_addr, 404 (pfn - start_pfn) << PAGE_SHIFT); 405 return (err); 406 } 407 408 return (0); 409 } 410 411 int 412 lkpi_io_mapping_map_user(struct io_mapping *iomap, 413 struct vm_area_struct *vma, unsigned long addr, 414 unsigned long pfn, unsigned long size) 415 { 416 pgprot_t prot; 417 int ret; 418 419 prot = cachemode2protval(iomap->attr); 420 ret = lkpi_remap_pfn_range(vma, addr, pfn, size, prot); 421 422 return (ret); 423 } 424 425 /* 426 * Although FreeBSD version of unmap_mapping_range has semantics and types of 427 * parameters compatible with Linux version, the values passed in are different 428 * @obj should match to vm_private_data field of vm_area_struct returned by 429 * mmap file operation handler, see linux_file_mmap_single() sources 430 * @holelen should match to size of area to be munmapped. 431 */ 432 void 433 lkpi_unmap_mapping_range(void *obj, loff_t const holebegin __unused, 434 loff_t const holelen __unused, int even_cows __unused) 435 { 436 vm_object_t devobj; 437 438 devobj = cdev_pager_lookup(obj); 439 if (devobj != NULL) { 440 cdev_mgtdev_pager_free_pages(devobj); 441 vm_object_deallocate(devobj); 442 } 443 } 444 445 int 446 lkpi_arch_phys_wc_add(unsigned long base, unsigned long size) 447 { 448 #ifdef __i386__ 449 struct mem_range_desc *mrdesc; 450 int error, id, act; 451 452 /* If PAT is available, do nothing */ 453 if (pat_works) 454 return (0); 455 456 mrdesc = malloc(sizeof(*mrdesc), M_LKMTRR, M_WAITOK); 457 mrdesc->mr_base = base; 458 mrdesc->mr_len = size; 459 mrdesc->mr_flags = MDF_WRITECOMBINE; 460 strlcpy(mrdesc->mr_owner, "drm", sizeof(mrdesc->mr_owner)); 461 act = MEMRANGE_SET_UPDATE; 462 error = mem_range_attr_set(mrdesc, &act); 463 if (error == 0) { 464 error = idr_get_new(&mtrr_idr, mrdesc, &id); 465 MPASS(idr_find(&mtrr_idr, id) == mrdesc); 466 if (error != 0) { 467 act = MEMRANGE_SET_REMOVE; 468 mem_range_attr_set(mrdesc, &act); 469 } 470 } 471 if (error != 0) { 472 free(mrdesc, M_LKMTRR); 473 pr_warn( 474 "Failed to add WC MTRR for [%p-%p]: %d; " 475 "performance may suffer\n", 476 (void *)base, (void *)(base + size - 1), error); 477 } else 478 pr_warn("Successfully added WC MTRR for [%p-%p]\n", 479 (void *)base, (void *)(base + size - 1)); 480 481 return (error != 0 ? -error : id + __MTRR_ID_BASE); 482 #else 483 return (0); 484 #endif 485 } 486 487 void 488 lkpi_arch_phys_wc_del(int reg) 489 { 490 #ifdef __i386__ 491 struct mem_range_desc *mrdesc; 492 int act; 493 494 /* Check if arch_phys_wc_add() failed. */ 495 if (reg < __MTRR_ID_BASE) 496 return; 497 498 mrdesc = idr_find(&mtrr_idr, reg - __MTRR_ID_BASE); 499 MPASS(mrdesc != NULL); 500 idr_remove(&mtrr_idr, reg - __MTRR_ID_BASE); 501 act = MEMRANGE_SET_REMOVE; 502 mem_range_attr_set(mrdesc, &act); 503 free(mrdesc, M_LKMTRR); 504 #endif 505 } 506 507 /* 508 * This is a highly simplified version of the Linux page_frag_cache. 509 * We only support up-to 1 single page as fragment size and we will 510 * always return a full page. This may be wasteful on small objects 511 * but the only known consumer (mt76) is either asking for a half-page 512 * or a full page. If this was to become a problem we can implement 513 * a more elaborate version. 514 */ 515 void * 516 linuxkpi_page_frag_alloc(struct page_frag_cache *pfc, 517 size_t fragsz, gfp_t gfp) 518 { 519 vm_page_t pages; 520 521 if (fragsz == 0) 522 return (NULL); 523 524 KASSERT(fragsz <= PAGE_SIZE, ("%s: fragsz %zu > PAGE_SIZE not yet " 525 "supported", __func__, fragsz)); 526 527 pages = alloc_pages(gfp, flsl(howmany(fragsz, PAGE_SIZE) - 1)); 528 if (pages == NULL) 529 return (NULL); 530 pfc->va = linux_page_address(pages); 531 532 /* Passed in as "count" to __page_frag_cache_drain(). Unused by us. */ 533 pfc->pagecnt_bias = 0; 534 535 return (pfc->va); 536 } 537 538 void 539 linuxkpi_page_frag_free(void *addr) 540 { 541 vm_page_t page; 542 543 page = virt_to_page(addr); 544 linux_free_pages(page, 0); 545 } 546 547 void 548 linuxkpi__page_frag_cache_drain(struct page *page, size_t count __unused) 549 { 550 551 linux_free_pages(page, 0); 552 } 553