1 /*- 2 * Copyright (c) 2013 The FreeBSD Foundation 3 * All rights reserved. 4 * 5 * This software was developed by Konstantin Belousov <kib@FreeBSD.org> 6 * under sponsorship from the FreeBSD Foundation. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following 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 AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include <sys/param.h> 34 #include <sys/systm.h> 35 #include <sys/malloc.h> 36 #include <sys/bus.h> 37 #include <sys/interrupt.h> 38 #include <sys/kernel.h> 39 #include <sys/ktr.h> 40 #include <sys/lock.h> 41 #include <sys/memdesc.h> 42 #include <sys/mutex.h> 43 #include <sys/proc.h> 44 #include <sys/rwlock.h> 45 #include <sys/rman.h> 46 #include <sys/sf_buf.h> 47 #include <sys/sysctl.h> 48 #include <sys/taskqueue.h> 49 #include <sys/tree.h> 50 #include <sys/uio.h> 51 #include <sys/vmem.h> 52 #include <vm/vm.h> 53 #include <vm/vm_extern.h> 54 #include <vm/vm_kern.h> 55 #include <vm/vm_object.h> 56 #include <vm/vm_page.h> 57 #include <vm/vm_pager.h> 58 #include <vm/vm_map.h> 59 #include <machine/atomic.h> 60 #include <machine/bus.h> 61 #include <machine/cpu.h> 62 #include <machine/md_var.h> 63 #include <machine/specialreg.h> 64 #include <x86/include/busdma_impl.h> 65 #include <x86/iommu/intel_reg.h> 66 #include <x86/iommu/busdma_dmar.h> 67 #include <x86/iommu/intel_dmar.h> 68 69 static int ctx_unmap_buf_locked(struct dmar_ctx *ctx, dmar_gaddr_t base, 70 dmar_gaddr_t size, int flags); 71 72 /* 73 * The cache of the identity mapping page tables for the DMARs. Using 74 * the cache saves significant amount of memory for page tables by 75 * reusing the page tables, since usually DMARs are identical and have 76 * the same capabilities. Still, cache records the information needed 77 * to match DMAR capabilities and page table format, to correctly 78 * handle different DMARs. 79 */ 80 81 struct idpgtbl { 82 dmar_gaddr_t maxaddr; /* Page table covers the guest address 83 range [0..maxaddr) */ 84 int pglvl; /* Total page table levels ignoring 85 superpages */ 86 int leaf; /* The last materialized page table 87 level, it is non-zero if superpages 88 are supported */ 89 vm_object_t pgtbl_obj; /* The page table pages */ 90 LIST_ENTRY(idpgtbl) link; 91 }; 92 93 static struct sx idpgtbl_lock; 94 SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl"); 95 static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls); 96 static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl", 97 "Intel DMAR Identity mappings cache elements"); 98 99 /* 100 * Build the next level of the page tables for the identity mapping. 101 * - lvl is the level to build; 102 * - idx is the index of the page table page in the pgtbl_obj, which is 103 * being allocated filled now; 104 * - addr is the starting address in the bus address space which is 105 * mapped by the page table page. 106 */ 107 static void 108 ctx_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx, 109 dmar_gaddr_t addr) 110 { 111 vm_page_t m, m1; 112 dmar_pte_t *pte; 113 struct sf_buf *sf; 114 dmar_gaddr_t f, pg_sz; 115 vm_pindex_t base; 116 int i; 117 118 VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj); 119 if (addr >= tbl->maxaddr) 120 return; 121 m = dmar_pgalloc(tbl->pgtbl_obj, idx, DMAR_PGF_OBJL | DMAR_PGF_WAITOK | 122 DMAR_PGF_ZERO); 123 base = idx * DMAR_NPTEPG + 1; /* Index of the first child page of idx */ 124 pg_sz = pglvl_page_size(tbl->pglvl, lvl); 125 if (lvl != tbl->leaf) { 126 for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) 127 ctx_idmap_nextlvl(tbl, lvl + 1, base + i, f); 128 } 129 VM_OBJECT_WUNLOCK(tbl->pgtbl_obj); 130 pte = dmar_map_pgtbl(tbl->pgtbl_obj, idx, DMAR_PGF_WAITOK, &sf); 131 if (lvl == tbl->leaf) { 132 for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) { 133 if (f >= tbl->maxaddr) 134 break; 135 pte[i].pte = (DMAR_PTE_ADDR_MASK & f) | 136 DMAR_PTE_R | DMAR_PTE_W; 137 } 138 } else { 139 for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) { 140 if (f >= tbl->maxaddr) 141 break; 142 m1 = dmar_pgalloc(tbl->pgtbl_obj, base + i, 143 DMAR_PGF_NOALLOC); 144 KASSERT(m1 != NULL, ("lost page table page")); 145 pte[i].pte = (DMAR_PTE_ADDR_MASK & 146 VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W; 147 } 148 } 149 /* ctx_get_idmap_pgtbl flushes CPU cache if needed. */ 150 dmar_unmap_pgtbl(sf); 151 VM_OBJECT_WLOCK(tbl->pgtbl_obj); 152 } 153 154 /* 155 * Find a ready and compatible identity-mapping page table in the 156 * cache. If not found, populate the identity-mapping page table for 157 * the context, up to the maxaddr. The maxaddr byte is allowed to be 158 * not mapped, which is aligned with the definition of Maxmem as the 159 * highest usable physical address + 1. If superpages are used, the 160 * maxaddr is typically mapped. 161 */ 162 vm_object_t 163 ctx_get_idmap_pgtbl(struct dmar_ctx *ctx, dmar_gaddr_t maxaddr) 164 { 165 struct dmar_unit *unit; 166 struct idpgtbl *tbl; 167 vm_object_t res; 168 vm_page_t m; 169 int leaf, i; 170 171 leaf = 0; /* silence gcc */ 172 173 /* 174 * First, determine where to stop the paging structures. 175 */ 176 for (i = 0; i < ctx->pglvl; i++) { 177 if (i == ctx->pglvl - 1 || ctx_is_sp_lvl(ctx, i)) { 178 leaf = i; 179 break; 180 } 181 } 182 183 /* 184 * Search the cache for a compatible page table. Qualified 185 * page table must map up to maxaddr, its level must be 186 * supported by the DMAR and leaf should be equal to the 187 * calculated value. The later restriction could be lifted 188 * but I believe it is currently impossible to have any 189 * deviations for existing hardware. 190 */ 191 sx_slock(&idpgtbl_lock); 192 LIST_FOREACH(tbl, &idpgtbls, link) { 193 if (tbl->maxaddr >= maxaddr && 194 dmar_pglvl_supported(ctx->dmar, tbl->pglvl) && 195 tbl->leaf == leaf) { 196 res = tbl->pgtbl_obj; 197 vm_object_reference(res); 198 sx_sunlock(&idpgtbl_lock); 199 ctx->pglvl = tbl->pglvl; /* XXXKIB ? */ 200 goto end; 201 } 202 } 203 204 /* 205 * Not found in cache, relock the cache into exclusive mode to 206 * be able to add element, and recheck cache again after the 207 * relock. 208 */ 209 sx_sunlock(&idpgtbl_lock); 210 sx_xlock(&idpgtbl_lock); 211 LIST_FOREACH(tbl, &idpgtbls, link) { 212 if (tbl->maxaddr >= maxaddr && 213 dmar_pglvl_supported(ctx->dmar, tbl->pglvl) && 214 tbl->leaf == leaf) { 215 res = tbl->pgtbl_obj; 216 vm_object_reference(res); 217 sx_xunlock(&idpgtbl_lock); 218 ctx->pglvl = tbl->pglvl; /* XXXKIB ? */ 219 return (res); 220 } 221 } 222 223 /* 224 * Still not found, create new page table. 225 */ 226 tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK); 227 tbl->pglvl = ctx->pglvl; 228 tbl->leaf = leaf; 229 tbl->maxaddr = maxaddr; 230 tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL, 231 IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL); 232 VM_OBJECT_WLOCK(tbl->pgtbl_obj); 233 ctx_idmap_nextlvl(tbl, 0, 0, 0); 234 VM_OBJECT_WUNLOCK(tbl->pgtbl_obj); 235 LIST_INSERT_HEAD(&idpgtbls, tbl, link); 236 res = tbl->pgtbl_obj; 237 vm_object_reference(res); 238 sx_xunlock(&idpgtbl_lock); 239 240 end: 241 /* 242 * Table was found or created. 243 * 244 * If DMAR does not snoop paging structures accesses, flush 245 * CPU cache to memory. Note that dmar_unmap_pgtbl() coherent 246 * argument was possibly invalid at the time of the identity 247 * page table creation, since DMAR which was passed at the 248 * time of creation could be coherent, while current DMAR is 249 * not. 250 * 251 * If DMAR cannot look into the chipset write buffer, flush it 252 * as well. 253 */ 254 unit = ctx->dmar; 255 if (!DMAR_IS_COHERENT(unit)) { 256 VM_OBJECT_WLOCK(res); 257 for (m = vm_page_lookup(res, 0); m != NULL; 258 m = vm_page_next(m)) 259 pmap_invalidate_cache_pages(&m, 1); 260 VM_OBJECT_WUNLOCK(res); 261 } 262 if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) { 263 DMAR_LOCK(unit); 264 dmar_flush_write_bufs(unit); 265 DMAR_UNLOCK(unit); 266 } 267 268 return (res); 269 } 270 271 /* 272 * Return a reference to the identity mapping page table to the cache. 273 */ 274 void 275 put_idmap_pgtbl(vm_object_t obj) 276 { 277 struct idpgtbl *tbl, *tbl1; 278 vm_object_t rmobj; 279 280 sx_slock(&idpgtbl_lock); 281 KASSERT(obj->ref_count >= 2, ("lost cache reference")); 282 vm_object_deallocate(obj); 283 284 /* 285 * Cache always owns one last reference on the page table object. 286 * If there is an additional reference, object must stay. 287 */ 288 if (obj->ref_count > 1) { 289 sx_sunlock(&idpgtbl_lock); 290 return; 291 } 292 293 /* 294 * Cache reference is the last, remove cache element and free 295 * page table object, returning the page table pages to the 296 * system. 297 */ 298 sx_sunlock(&idpgtbl_lock); 299 sx_xlock(&idpgtbl_lock); 300 LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) { 301 rmobj = tbl->pgtbl_obj; 302 if (rmobj->ref_count == 1) { 303 LIST_REMOVE(tbl, link); 304 atomic_subtract_int(&dmar_tbl_pagecnt, 305 rmobj->resident_page_count); 306 vm_object_deallocate(rmobj); 307 free(tbl, M_DMAR_IDPGTBL); 308 } 309 } 310 sx_xunlock(&idpgtbl_lock); 311 } 312 313 /* 314 * The core routines to map and unmap host pages at the given guest 315 * address. Support superpages. 316 */ 317 318 /* 319 * Index of the pte for the guest address base in the page table at 320 * the level lvl. 321 */ 322 static int 323 ctx_pgtbl_pte_off(struct dmar_ctx *ctx, dmar_gaddr_t base, int lvl) 324 { 325 326 base >>= DMAR_PAGE_SHIFT + (ctx->pglvl - lvl - 1) * DMAR_NPTEPGSHIFT; 327 return (base & DMAR_PTEMASK); 328 } 329 330 /* 331 * Returns the page index of the page table page in the page table 332 * object, which maps the given address base at the page table level 333 * lvl. 334 */ 335 static vm_pindex_t 336 ctx_pgtbl_get_pindex(struct dmar_ctx *ctx, dmar_gaddr_t base, int lvl) 337 { 338 vm_pindex_t idx, pidx; 339 int i; 340 341 KASSERT(lvl >= 0 && lvl < ctx->pglvl, ("wrong lvl %p %d", ctx, lvl)); 342 343 for (pidx = idx = 0, i = 0; i < lvl; i++, pidx = idx) 344 idx = ctx_pgtbl_pte_off(ctx, base, i) + pidx * DMAR_NPTEPG + 1; 345 return (idx); 346 } 347 348 static dmar_pte_t * 349 ctx_pgtbl_map_pte(struct dmar_ctx *ctx, dmar_gaddr_t base, int lvl, int flags, 350 vm_pindex_t *idxp, struct sf_buf **sf) 351 { 352 vm_page_t m; 353 struct sf_buf *sfp; 354 dmar_pte_t *pte, *ptep; 355 vm_pindex_t idx, idx1; 356 357 DMAR_CTX_ASSERT_PGLOCKED(ctx); 358 KASSERT((flags & DMAR_PGF_OBJL) != 0, ("lost PGF_OBJL")); 359 360 idx = ctx_pgtbl_get_pindex(ctx, base, lvl); 361 if (*sf != NULL && idx == *idxp) { 362 pte = (dmar_pte_t *)sf_buf_kva(*sf); 363 } else { 364 if (*sf != NULL) 365 dmar_unmap_pgtbl(*sf); 366 *idxp = idx; 367 retry: 368 pte = dmar_map_pgtbl(ctx->pgtbl_obj, idx, flags, sf); 369 if (pte == NULL) { 370 KASSERT(lvl > 0, ("lost root page table page %p", ctx)); 371 /* 372 * Page table page does not exists, allocate 373 * it and create pte in the up level. 374 */ 375 m = dmar_pgalloc(ctx->pgtbl_obj, idx, flags | 376 DMAR_PGF_ZERO); 377 if (m == NULL) 378 return (NULL); 379 380 /* 381 * Prevent potential free while pgtbl_obj is 382 * unlocked in the recursive call to 383 * ctx_pgtbl_map_pte(), if other thread did 384 * pte write and clean while the lock if 385 * dropped. 386 */ 387 m->wire_count++; 388 389 sfp = NULL; 390 ptep = ctx_pgtbl_map_pte(ctx, base, lvl - 1, flags, 391 &idx1, &sfp); 392 if (ptep == NULL) { 393 KASSERT(m->pindex != 0, 394 ("loosing root page %p", ctx)); 395 m->wire_count--; 396 dmar_pgfree(ctx->pgtbl_obj, m->pindex, flags); 397 return (NULL); 398 } 399 dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W | 400 VM_PAGE_TO_PHYS(m)); 401 dmar_flush_pte_to_ram(ctx->dmar, ptep); 402 sf_buf_page(sfp)->wire_count += 1; 403 m->wire_count--; 404 dmar_unmap_pgtbl(sfp); 405 /* Only executed once. */ 406 goto retry; 407 } 408 } 409 pte += ctx_pgtbl_pte_off(ctx, base, lvl); 410 return (pte); 411 } 412 413 static int 414 ctx_map_buf_locked(struct dmar_ctx *ctx, dmar_gaddr_t base, dmar_gaddr_t size, 415 vm_page_t *ma, uint64_t pflags, int flags) 416 { 417 dmar_pte_t *pte; 418 struct sf_buf *sf; 419 dmar_gaddr_t pg_sz, base1, size1; 420 vm_pindex_t pi, c, idx, run_sz; 421 int lvl; 422 bool superpage; 423 424 DMAR_CTX_ASSERT_PGLOCKED(ctx); 425 426 base1 = base; 427 size1 = size; 428 flags |= DMAR_PGF_OBJL; 429 TD_PREP_PINNED_ASSERT; 430 431 for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz, 432 pi += run_sz) { 433 for (lvl = 0, c = 0, superpage = false;; lvl++) { 434 pg_sz = ctx_page_size(ctx, lvl); 435 run_sz = pg_sz >> DMAR_PAGE_SHIFT; 436 if (lvl == ctx->pglvl - 1) 437 break; 438 /* 439 * Check if the current base suitable for the 440 * superpage mapping. First, verify the level. 441 */ 442 if (!ctx_is_sp_lvl(ctx, lvl)) 443 continue; 444 /* 445 * Next, look at the size of the mapping and 446 * alignment of both guest and host addresses. 447 */ 448 if (size < pg_sz || (base & (pg_sz - 1)) != 0 || 449 (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0) 450 continue; 451 /* All passed, check host pages contiguouty. */ 452 if (c == 0) { 453 for (c = 1; c < run_sz; c++) { 454 if (VM_PAGE_TO_PHYS(ma[pi + c]) != 455 VM_PAGE_TO_PHYS(ma[pi + c - 1]) + 456 PAGE_SIZE) 457 break; 458 } 459 } 460 if (c >= run_sz) { 461 superpage = true; 462 break; 463 } 464 } 465 KASSERT(size >= pg_sz, 466 ("mapping loop overflow %p %jx %jx %jx", ctx, 467 (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz)); 468 KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl)); 469 pte = ctx_pgtbl_map_pte(ctx, base, lvl, flags, &idx, &sf); 470 if (pte == NULL) { 471 KASSERT((flags & DMAR_PGF_WAITOK) == 0, 472 ("failed waitable pte alloc %p", ctx)); 473 if (sf != NULL) 474 dmar_unmap_pgtbl(sf); 475 ctx_unmap_buf_locked(ctx, base1, base - base1, flags); 476 TD_PINNED_ASSERT; 477 return (ENOMEM); 478 } 479 dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags | 480 (superpage ? DMAR_PTE_SP : 0)); 481 dmar_flush_pte_to_ram(ctx->dmar, pte); 482 sf_buf_page(sf)->wire_count += 1; 483 } 484 if (sf != NULL) 485 dmar_unmap_pgtbl(sf); 486 TD_PINNED_ASSERT; 487 return (0); 488 } 489 490 int 491 ctx_map_buf(struct dmar_ctx *ctx, dmar_gaddr_t base, dmar_gaddr_t size, 492 vm_page_t *ma, uint64_t pflags, int flags) 493 { 494 struct dmar_unit *unit; 495 int error; 496 497 unit = ctx->dmar; 498 499 KASSERT((ctx->flags & DMAR_CTX_IDMAP) == 0, 500 ("modifying idmap pagetable ctx %p", ctx)); 501 KASSERT((base & DMAR_PAGE_MASK) == 0, 502 ("non-aligned base %p %jx %jx", ctx, (uintmax_t)base, 503 (uintmax_t)size)); 504 KASSERT((size & DMAR_PAGE_MASK) == 0, 505 ("non-aligned size %p %jx %jx", ctx, (uintmax_t)base, 506 (uintmax_t)size)); 507 KASSERT(size > 0, ("zero size %p %jx %jx", ctx, (uintmax_t)base, 508 (uintmax_t)size)); 509 KASSERT(base < (1ULL << ctx->agaw), 510 ("base too high %p %jx %jx agaw %d", ctx, (uintmax_t)base, 511 (uintmax_t)size, ctx->agaw)); 512 KASSERT(base + size < (1ULL << ctx->agaw), 513 ("end too high %p %jx %jx agaw %d", ctx, (uintmax_t)base, 514 (uintmax_t)size, ctx->agaw)); 515 KASSERT(base + size > base, 516 ("size overflow %p %jx %jx", ctx, (uintmax_t)base, 517 (uintmax_t)size)); 518 KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0, 519 ("neither read nor write %jx", (uintmax_t)pflags)); 520 KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP | 521 DMAR_PTE_TM)) == 0, 522 ("invalid pte flags %jx", (uintmax_t)pflags)); 523 KASSERT((pflags & DMAR_PTE_SNP) == 0 || 524 (unit->hw_ecap & DMAR_ECAP_SC) != 0, 525 ("PTE_SNP for dmar without snoop control %p %jx", 526 ctx, (uintmax_t)pflags)); 527 KASSERT((pflags & DMAR_PTE_TM) == 0 || 528 (unit->hw_ecap & DMAR_ECAP_DI) != 0, 529 ("PTE_TM for dmar without DIOTLB %p %jx", 530 ctx, (uintmax_t)pflags)); 531 KASSERT((flags & ~DMAR_PGF_WAITOK) == 0, ("invalid flags %x", flags)); 532 533 DMAR_CTX_PGLOCK(ctx); 534 error = ctx_map_buf_locked(ctx, base, size, ma, pflags, flags); 535 DMAR_CTX_PGUNLOCK(ctx); 536 if (error != 0) 537 return (error); 538 539 if ((unit->hw_cap & DMAR_CAP_CM) != 0) 540 ctx_flush_iotlb_sync(ctx, base, size); 541 else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) { 542 /* See 11.1 Write Buffer Flushing. */ 543 DMAR_LOCK(unit); 544 dmar_flush_write_bufs(unit); 545 DMAR_UNLOCK(unit); 546 } 547 return (0); 548 } 549 550 static void ctx_unmap_clear_pte(struct dmar_ctx *ctx, dmar_gaddr_t base, 551 int lvl, int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_fs); 552 553 static void 554 ctx_free_pgtbl_pde(struct dmar_ctx *ctx, dmar_gaddr_t base, int lvl, int flags) 555 { 556 struct sf_buf *sf; 557 dmar_pte_t *pde; 558 vm_pindex_t idx; 559 560 sf = NULL; 561 pde = ctx_pgtbl_map_pte(ctx, base, lvl, flags, &idx, &sf); 562 ctx_unmap_clear_pte(ctx, base, lvl, flags, pde, &sf, true); 563 } 564 565 static void 566 ctx_unmap_clear_pte(struct dmar_ctx *ctx, dmar_gaddr_t base, int lvl, 567 int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_sf) 568 { 569 vm_page_t m; 570 571 dmar_pte_clear(&pte->pte); 572 dmar_flush_pte_to_ram(ctx->dmar, pte); 573 m = sf_buf_page(*sf); 574 if (free_sf) { 575 dmar_unmap_pgtbl(*sf); 576 *sf = NULL; 577 } 578 m->wire_count--; 579 if (m->wire_count != 0) 580 return; 581 KASSERT(lvl != 0, 582 ("lost reference (lvl) on root pg ctx %p base %jx lvl %d", 583 ctx, (uintmax_t)base, lvl)); 584 KASSERT(m->pindex != 0, 585 ("lost reference (idx) on root pg ctx %p base %jx lvl %d", 586 ctx, (uintmax_t)base, lvl)); 587 dmar_pgfree(ctx->pgtbl_obj, m->pindex, flags); 588 ctx_free_pgtbl_pde(ctx, base, lvl - 1, flags); 589 } 590 591 /* 592 * Assumes that the unmap is never partial. 593 */ 594 static int 595 ctx_unmap_buf_locked(struct dmar_ctx *ctx, dmar_gaddr_t base, 596 dmar_gaddr_t size, int flags) 597 { 598 dmar_pte_t *pte; 599 struct sf_buf *sf; 600 vm_pindex_t idx; 601 dmar_gaddr_t pg_sz, base1, size1; 602 int lvl; 603 604 DMAR_CTX_ASSERT_PGLOCKED(ctx); 605 if (size == 0) 606 return (0); 607 608 KASSERT((ctx->flags & DMAR_CTX_IDMAP) == 0, 609 ("modifying idmap pagetable ctx %p", ctx)); 610 KASSERT((base & DMAR_PAGE_MASK) == 0, 611 ("non-aligned base %p %jx %jx", ctx, (uintmax_t)base, 612 (uintmax_t)size)); 613 KASSERT((size & DMAR_PAGE_MASK) == 0, 614 ("non-aligned size %p %jx %jx", ctx, (uintmax_t)base, 615 (uintmax_t)size)); 616 KASSERT(base < (1ULL << ctx->agaw), 617 ("base too high %p %jx %jx agaw %d", ctx, (uintmax_t)base, 618 (uintmax_t)size, ctx->agaw)); 619 KASSERT(base + size < (1ULL << ctx->agaw), 620 ("end too high %p %jx %jx agaw %d", ctx, (uintmax_t)base, 621 (uintmax_t)size, ctx->agaw)); 622 KASSERT(base + size > base, 623 ("size overflow %p %jx %jx", ctx, (uintmax_t)base, 624 (uintmax_t)size)); 625 KASSERT((flags & ~DMAR_PGF_WAITOK) == 0, ("invalid flags %x", flags)); 626 627 pg_sz = 0; /* silence gcc */ 628 base1 = base; 629 size1 = size; 630 flags |= DMAR_PGF_OBJL; 631 TD_PREP_PINNED_ASSERT; 632 633 for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) { 634 for (lvl = 0; lvl < ctx->pglvl; lvl++) { 635 if (lvl != ctx->pglvl - 1 && !ctx_is_sp_lvl(ctx, lvl)) 636 continue; 637 pg_sz = ctx_page_size(ctx, lvl); 638 if (pg_sz > size) 639 continue; 640 pte = ctx_pgtbl_map_pte(ctx, base, lvl, flags, 641 &idx, &sf); 642 KASSERT(pte != NULL, 643 ("sleeping or page missed %p %jx %d 0x%x", 644 ctx, (uintmax_t)base, lvl, flags)); 645 if ((pte->pte & DMAR_PTE_SP) != 0 || 646 lvl == ctx->pglvl - 1) { 647 ctx_unmap_clear_pte(ctx, base, lvl, flags, 648 pte, &sf, false); 649 break; 650 } 651 } 652 KASSERT(size >= pg_sz, 653 ("unmapping loop overflow %p %jx %jx %jx", ctx, 654 (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz)); 655 } 656 if (sf != NULL) 657 dmar_unmap_pgtbl(sf); 658 /* 659 * See 11.1 Write Buffer Flushing for an explanation why RWBF 660 * can be ignored there. 661 */ 662 663 TD_PINNED_ASSERT; 664 return (0); 665 } 666 667 int 668 ctx_unmap_buf(struct dmar_ctx *ctx, dmar_gaddr_t base, dmar_gaddr_t size, 669 int flags) 670 { 671 int error; 672 673 DMAR_CTX_PGLOCK(ctx); 674 error = ctx_unmap_buf_locked(ctx, base, size, flags); 675 DMAR_CTX_PGUNLOCK(ctx); 676 return (error); 677 } 678 679 int 680 ctx_alloc_pgtbl(struct dmar_ctx *ctx) 681 { 682 vm_page_t m; 683 684 KASSERT(ctx->pgtbl_obj == NULL, ("already initialized %p", ctx)); 685 686 ctx->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL, 687 IDX_TO_OFF(pglvl_max_pages(ctx->pglvl)), 0, 0, NULL); 688 DMAR_CTX_PGLOCK(ctx); 689 m = dmar_pgalloc(ctx->pgtbl_obj, 0, DMAR_PGF_WAITOK | 690 DMAR_PGF_ZERO | DMAR_PGF_OBJL); 691 /* No implicit free of the top level page table page. */ 692 m->wire_count = 1; 693 DMAR_CTX_PGUNLOCK(ctx); 694 return (0); 695 } 696 697 void 698 ctx_free_pgtbl(struct dmar_ctx *ctx) 699 { 700 vm_object_t obj; 701 vm_page_t m; 702 703 obj = ctx->pgtbl_obj; 704 if (obj == NULL) { 705 KASSERT((ctx->dmar->hw_ecap & DMAR_ECAP_PT) != 0 && 706 (ctx->flags & DMAR_CTX_IDMAP) != 0, 707 ("lost pagetable object ctx %p", ctx)); 708 return; 709 } 710 DMAR_CTX_ASSERT_PGLOCKED(ctx); 711 ctx->pgtbl_obj = NULL; 712 713 if ((ctx->flags & DMAR_CTX_IDMAP) != 0) { 714 put_idmap_pgtbl(obj); 715 ctx->flags &= ~DMAR_CTX_IDMAP; 716 return; 717 } 718 719 /* Obliterate wire_counts */ 720 VM_OBJECT_ASSERT_WLOCKED(obj); 721 for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m)) 722 m->wire_count = 0; 723 VM_OBJECT_WUNLOCK(obj); 724 vm_object_deallocate(obj); 725 } 726 727 static inline uint64_t 728 ctx_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro) 729 { 730 uint64_t iotlbr; 731 732 dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT | 733 DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt); 734 for (;;) { 735 iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF); 736 if ((iotlbr & DMAR_IOTLB_IVT) == 0) 737 break; 738 cpu_spinwait(); 739 } 740 return (iotlbr); 741 } 742 743 void 744 ctx_flush_iotlb_sync(struct dmar_ctx *ctx, dmar_gaddr_t base, dmar_gaddr_t size) 745 { 746 struct dmar_unit *unit; 747 dmar_gaddr_t isize; 748 uint64_t iotlbr; 749 int am, iro; 750 751 unit = ctx->dmar; 752 KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call", 753 unit->unit)); 754 iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16; 755 DMAR_LOCK(unit); 756 if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) { 757 iotlbr = ctx_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM | 758 DMAR_IOTLB_DID(ctx->domain), iro); 759 KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) != 760 DMAR_IOTLB_IAIG_INVLD, 761 ("dmar%d: invalidation failed %jx", unit->unit, 762 (uintmax_t)iotlbr)); 763 } else { 764 for (; size > 0; base += isize, size -= isize) { 765 am = calc_am(unit, base, size, &isize); 766 dmar_write8(unit, iro, base | am); 767 iotlbr = ctx_wait_iotlb_flush(unit, 768 DMAR_IOTLB_IIRG_PAGE | DMAR_IOTLB_DID(ctx->domain), 769 iro); 770 KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) != 771 DMAR_IOTLB_IAIG_INVLD, 772 ("dmar%d: PSI invalidation failed " 773 "iotlbr 0x%jx base 0x%jx size 0x%jx am %d", 774 unit->unit, (uintmax_t)iotlbr, 775 (uintmax_t)base, (uintmax_t)size, am)); 776 /* 777 * Any non-page granularity covers whole guest 778 * address space for the domain. 779 */ 780 if ((iotlbr & DMAR_IOTLB_IAIG_MASK) != 781 DMAR_IOTLB_IAIG_PAGE) 782 break; 783 } 784 } 785 DMAR_UNLOCK(unit); 786 } 787