xref: /freebsd/sys/x86/iommu/intel_idpgtbl.c (revision e47161e5f1f01ef300c6e7efdb9c92e3a6c497ff)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2013 The FreeBSD Foundation
5  *
6  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
7  * under sponsorship from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/malloc.h>
34 #include <sys/bus.h>
35 #include <sys/interrupt.h>
36 #include <sys/kernel.h>
37 #include <sys/ktr.h>
38 #include <sys/lock.h>
39 #include <sys/memdesc.h>
40 #include <sys/mutex.h>
41 #include <sys/proc.h>
42 #include <sys/rwlock.h>
43 #include <sys/rman.h>
44 #include <sys/sf_buf.h>
45 #include <sys/sysctl.h>
46 #include <sys/taskqueue.h>
47 #include <sys/tree.h>
48 #include <sys/uio.h>
49 #include <sys/vmem.h>
50 #include <sys/vmmeter.h>
51 #include <vm/vm.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_kern.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_pager.h>
57 #include <vm/vm_map.h>
58 #include <dev/pci/pcireg.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 <dev/iommu/busdma_iommu.h>
66 #include <x86/iommu/intel_reg.h>
67 #include <x86/iommu/x86_iommu.h>
68 #include <x86/iommu/intel_dmar.h>
69 
70 static int domain_unmap_buf_locked(struct dmar_domain *domain,
71     iommu_gaddr_t base, iommu_gaddr_t size, int flags);
72 
73 /*
74  * The cache of the identity mapping page tables for the DMARs.  Using
75  * the cache saves significant amount of memory for page tables by
76  * reusing the page tables, since usually DMARs are identical and have
77  * the same capabilities.  Still, cache records the information needed
78  * to match DMAR capabilities and page table format, to correctly
79  * handle different DMARs.
80  */
81 
82 struct idpgtbl {
83 	iommu_gaddr_t maxaddr;	/* Page table covers the guest address
84 				   range [0..maxaddr) */
85 	int pglvl;		/* Total page table levels ignoring
86 				   superpages */
87 	int leaf;		/* The last materialized page table
88 				   level, it is non-zero if superpages
89 				   are supported */
90 	vm_object_t pgtbl_obj;	/* The page table pages */
91 	LIST_ENTRY(idpgtbl) link;
92 };
93 
94 static struct sx idpgtbl_lock;
95 SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl");
96 static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls);
97 static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl",
98     "Intel DMAR Identity mappings cache elements");
99 
100 /*
101  * Build the next level of the page tables for the identity mapping.
102  * - lvl is the level to build;
103  * - idx is the index of the page table page in the pgtbl_obj, which is
104  *   being allocated filled now;
105  * - addr is the starting address in the bus address space which is
106  *   mapped by the page table page.
107  */
108 static void
109 domain_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx,
110     iommu_gaddr_t addr)
111 {
112 	vm_page_t m1;
113 	iommu_pte_t *pte;
114 	struct sf_buf *sf;
115 	iommu_gaddr_t f, pg_sz;
116 	vm_pindex_t base;
117 	int i;
118 
119 	VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj);
120 	if (addr >= tbl->maxaddr)
121 		return;
122 	(void)iommu_pgalloc(tbl->pgtbl_obj, idx, IOMMU_PGF_OBJL |
123 	    IOMMU_PGF_WAITOK | IOMMU_PGF_ZERO);
124 	base = idx * IOMMU_NPTEPG + 1; /* Index of the first child page of idx */
125 	pg_sz = pglvl_page_size(tbl->pglvl, lvl);
126 	if (lvl != tbl->leaf) {
127 		for (i = 0, f = addr; i < IOMMU_NPTEPG; i++, f += pg_sz)
128 			domain_idmap_nextlvl(tbl, lvl + 1, base + i, f);
129 	}
130 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
131 	pte = iommu_map_pgtbl(tbl->pgtbl_obj, idx, IOMMU_PGF_WAITOK, &sf);
132 	if (lvl == tbl->leaf) {
133 		for (i = 0, f = addr; i < IOMMU_NPTEPG; i++, f += pg_sz) {
134 			if (f >= tbl->maxaddr)
135 				break;
136 			pte[i].pte = (DMAR_PTE_ADDR_MASK & f) |
137 			    DMAR_PTE_R | DMAR_PTE_W;
138 		}
139 	} else {
140 		for (i = 0, f = addr; i < IOMMU_NPTEPG; i++, f += pg_sz) {
141 			if (f >= tbl->maxaddr)
142 				break;
143 			m1 = iommu_pgalloc(tbl->pgtbl_obj, base + i,
144 			    IOMMU_PGF_NOALLOC);
145 			KASSERT(m1 != NULL, ("lost page table page"));
146 			pte[i].pte = (DMAR_PTE_ADDR_MASK &
147 			    VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W;
148 		}
149 	}
150 	/* domain_get_idmap_pgtbl flushes CPU cache if needed. */
151 	iommu_unmap_pgtbl(sf);
152 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
153 }
154 
155 /*
156  * Find a ready and compatible identity-mapping page table in the
157  * cache. If not found, populate the identity-mapping page table for
158  * the context, up to the maxaddr. The maxaddr byte is allowed to be
159  * not mapped, which is aligned with the definition of Maxmem as the
160  * highest usable physical address + 1.  If superpages are used, the
161  * maxaddr is typically mapped.
162  */
163 vm_object_t
164 domain_get_idmap_pgtbl(struct dmar_domain *domain, iommu_gaddr_t maxaddr)
165 {
166 	struct dmar_unit *unit;
167 	struct idpgtbl *tbl;
168 	vm_object_t res;
169 	vm_page_t m;
170 	int leaf, i;
171 
172 	leaf = 0; /* silence gcc */
173 
174 	/*
175 	 * First, determine where to stop the paging structures.
176 	 */
177 	for (i = 0; i < domain->pglvl; i++) {
178 		if (i == domain->pglvl - 1 || domain_is_sp_lvl(domain, i)) {
179 			leaf = i;
180 			break;
181 		}
182 	}
183 
184 	/*
185 	 * Search the cache for a compatible page table.  Qualified
186 	 * page table must map up to maxaddr, its level must be
187 	 * supported by the DMAR and leaf should be equal to the
188 	 * calculated value.  The later restriction could be lifted
189 	 * but I believe it is currently impossible to have any
190 	 * deviations for existing hardware.
191 	 */
192 	sx_slock(&idpgtbl_lock);
193 	LIST_FOREACH(tbl, &idpgtbls, link) {
194 		if (tbl->maxaddr >= maxaddr &&
195 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
196 		    tbl->leaf == leaf) {
197 			res = tbl->pgtbl_obj;
198 			vm_object_reference(res);
199 			sx_sunlock(&idpgtbl_lock);
200 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
201 			goto end;
202 		}
203 	}
204 
205 	/*
206 	 * Not found in cache, relock the cache into exclusive mode to
207 	 * be able to add element, and recheck cache again after the
208 	 * relock.
209 	 */
210 	sx_sunlock(&idpgtbl_lock);
211 	sx_xlock(&idpgtbl_lock);
212 	LIST_FOREACH(tbl, &idpgtbls, link) {
213 		if (tbl->maxaddr >= maxaddr &&
214 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
215 		    tbl->leaf == leaf) {
216 			res = tbl->pgtbl_obj;
217 			vm_object_reference(res);
218 			sx_xunlock(&idpgtbl_lock);
219 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
220 			return (res);
221 		}
222 	}
223 
224 	/*
225 	 * Still not found, create new page table.
226 	 */
227 	tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK);
228 	tbl->pglvl = domain->pglvl;
229 	tbl->leaf = leaf;
230 	tbl->maxaddr = maxaddr;
231 	tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
232 	    IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL);
233 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
234 	domain_idmap_nextlvl(tbl, 0, 0, 0);
235 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
236 	LIST_INSERT_HEAD(&idpgtbls, tbl, link);
237 	res = tbl->pgtbl_obj;
238 	vm_object_reference(res);
239 	sx_xunlock(&idpgtbl_lock);
240 
241 end:
242 	/*
243 	 * Table was found or created.
244 	 *
245 	 * If DMAR does not snoop paging structures accesses, flush
246 	 * CPU cache to memory.  Note that dmar_unmap_pgtbl() coherent
247 	 * argument was possibly invalid at the time of the identity
248 	 * page table creation, since DMAR which was passed at the
249 	 * time of creation could be coherent, while current DMAR is
250 	 * not.
251 	 *
252 	 * If DMAR cannot look into the chipset write buffer, flush it
253 	 * as well.
254 	 */
255 	unit = domain->dmar;
256 	if (!DMAR_IS_COHERENT(unit)) {
257 		VM_OBJECT_WLOCK(res);
258 		for (m = vm_page_lookup(res, 0); m != NULL;
259 		     m = vm_page_next(m))
260 			pmap_invalidate_cache_pages(&m, 1);
261 		VM_OBJECT_WUNLOCK(res);
262 	}
263 	if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
264 		DMAR_LOCK(unit);
265 		dmar_flush_write_bufs(unit);
266 		DMAR_UNLOCK(unit);
267 	}
268 
269 	return (res);
270 }
271 
272 /*
273  * Return a reference to the identity mapping page table to the cache.
274  */
275 void
276 put_idmap_pgtbl(vm_object_t obj)
277 {
278 	struct idpgtbl *tbl, *tbl1;
279 	vm_object_t rmobj;
280 
281 	sx_slock(&idpgtbl_lock);
282 	KASSERT(obj->ref_count >= 2, ("lost cache reference"));
283 	vm_object_deallocate(obj);
284 
285 	/*
286 	 * Cache always owns one last reference on the page table object.
287 	 * If there is an additional reference, object must stay.
288 	 */
289 	if (obj->ref_count > 1) {
290 		sx_sunlock(&idpgtbl_lock);
291 		return;
292 	}
293 
294 	/*
295 	 * Cache reference is the last, remove cache element and free
296 	 * page table object, returning the page table pages to the
297 	 * system.
298 	 */
299 	sx_sunlock(&idpgtbl_lock);
300 	sx_xlock(&idpgtbl_lock);
301 	LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) {
302 		rmobj = tbl->pgtbl_obj;
303 		if (rmobj->ref_count == 1) {
304 			LIST_REMOVE(tbl, link);
305 			atomic_subtract_int(&iommu_tbl_pagecnt,
306 			    rmobj->resident_page_count);
307 			vm_object_deallocate(rmobj);
308 			free(tbl, M_DMAR_IDPGTBL);
309 		}
310 	}
311 	sx_xunlock(&idpgtbl_lock);
312 }
313 
314 /*
315  * The core routines to map and unmap host pages at the given guest
316  * address.  Support superpages.
317  */
318 
319 static iommu_pte_t *
320 domain_pgtbl_map_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
321     int flags, vm_pindex_t *idxp, struct sf_buf **sf)
322 {
323 	vm_page_t m;
324 	struct sf_buf *sfp;
325 	iommu_pte_t *pte, *ptep;
326 	vm_pindex_t idx, idx1;
327 
328 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
329 	KASSERT((flags & IOMMU_PGF_OBJL) != 0, ("lost PGF_OBJL"));
330 
331 	idx = pglvl_pgtbl_get_pindex(domain->pglvl, base, lvl);
332 	if (*sf != NULL && idx == *idxp) {
333 		pte = (iommu_pte_t *)sf_buf_kva(*sf);
334 	} else {
335 		if (*sf != NULL)
336 			iommu_unmap_pgtbl(*sf);
337 		*idxp = idx;
338 retry:
339 		pte = iommu_map_pgtbl(domain->pgtbl_obj, idx, flags, sf);
340 		if (pte == NULL) {
341 			KASSERT(lvl > 0,
342 			    ("lost root page table page %p", domain));
343 			/*
344 			 * Page table page does not exist, allocate
345 			 * it and create a pte in the preceeding page level
346 			 * to reference the allocated page table page.
347 			 */
348 			m = iommu_pgalloc(domain->pgtbl_obj, idx, flags |
349 			    IOMMU_PGF_ZERO);
350 			if (m == NULL)
351 				return (NULL);
352 
353 			/*
354 			 * Prevent potential free while pgtbl_obj is
355 			 * unlocked in the recursive call to
356 			 * domain_pgtbl_map_pte(), if other thread did
357 			 * pte write and clean while the lock is
358 			 * dropped.
359 			 */
360 			vm_page_wire(m);
361 
362 			sfp = NULL;
363 			ptep = domain_pgtbl_map_pte(domain, base, lvl - 1,
364 			    flags, &idx1, &sfp);
365 			if (ptep == NULL) {
366 				KASSERT(m->pindex != 0,
367 				    ("loosing root page %p", domain));
368 				vm_page_unwire_noq(m);
369 				iommu_pgfree(domain->pgtbl_obj, m->pindex,
370 				    flags);
371 				return (NULL);
372 			}
373 			dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W |
374 			    VM_PAGE_TO_PHYS(m));
375 			dmar_flush_pte_to_ram(domain->dmar, ptep);
376 			vm_page_wire(sf_buf_page(sfp));
377 			vm_page_unwire_noq(m);
378 			iommu_unmap_pgtbl(sfp);
379 			/* Only executed once. */
380 			goto retry;
381 		}
382 	}
383 	pte += pglvl_pgtbl_pte_off(domain->pglvl, base, lvl);
384 	return (pte);
385 }
386 
387 static int
388 domain_map_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
389     iommu_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags)
390 {
391 	iommu_pte_t *pte;
392 	struct sf_buf *sf;
393 	iommu_gaddr_t pg_sz, base1;
394 	vm_pindex_t pi, c, idx, run_sz;
395 	int lvl;
396 	bool superpage;
397 
398 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
399 
400 	base1 = base;
401 	flags |= IOMMU_PGF_OBJL;
402 	TD_PREP_PINNED_ASSERT;
403 
404 	for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz,
405 	    pi += run_sz) {
406 		for (lvl = 0, c = 0, superpage = false;; lvl++) {
407 			pg_sz = domain_page_size(domain, lvl);
408 			run_sz = pg_sz >> IOMMU_PAGE_SHIFT;
409 			if (lvl == domain->pglvl - 1)
410 				break;
411 			/*
412 			 * Check if the current base suitable for the
413 			 * superpage mapping.  First, verify the level.
414 			 */
415 			if (!domain_is_sp_lvl(domain, lvl))
416 				continue;
417 			/*
418 			 * Next, look at the size of the mapping and
419 			 * alignment of both guest and host addresses.
420 			 */
421 			if (size < pg_sz || (base & (pg_sz - 1)) != 0 ||
422 			    (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0)
423 				continue;
424 			/* All passed, check host pages contiguouty. */
425 			if (c == 0) {
426 				for (c = 1; c < run_sz; c++) {
427 					if (VM_PAGE_TO_PHYS(ma[pi + c]) !=
428 					    VM_PAGE_TO_PHYS(ma[pi + c - 1]) +
429 					    PAGE_SIZE)
430 						break;
431 				}
432 			}
433 			if (c >= run_sz) {
434 				superpage = true;
435 				break;
436 			}
437 		}
438 		KASSERT(size >= pg_sz,
439 		    ("mapping loop overflow %p %jx %jx %jx", domain,
440 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
441 		KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl));
442 		pte = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
443 		if (pte == NULL) {
444 			KASSERT((flags & IOMMU_PGF_WAITOK) == 0,
445 			    ("failed waitable pte alloc %p", domain));
446 			if (sf != NULL)
447 				iommu_unmap_pgtbl(sf);
448 			domain_unmap_buf_locked(domain, base1, base - base1,
449 			    flags);
450 			TD_PINNED_ASSERT;
451 			return (ENOMEM);
452 		}
453 		dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags |
454 		    (superpage ? DMAR_PTE_SP : 0));
455 		dmar_flush_pte_to_ram(domain->dmar, pte);
456 		vm_page_wire(sf_buf_page(sf));
457 	}
458 	if (sf != NULL)
459 		iommu_unmap_pgtbl(sf);
460 	TD_PINNED_ASSERT;
461 	return (0);
462 }
463 
464 static int
465 domain_map_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
466     iommu_gaddr_t size, vm_page_t *ma, uint64_t eflags, int flags)
467 {
468 	struct dmar_domain *domain;
469 	struct dmar_unit *unit;
470 	uint64_t pflags;
471 	int error;
472 
473 	pflags = ((eflags & IOMMU_MAP_ENTRY_READ) != 0 ? DMAR_PTE_R : 0) |
474 	    ((eflags & IOMMU_MAP_ENTRY_WRITE) != 0 ? DMAR_PTE_W : 0) |
475 	    ((eflags & IOMMU_MAP_ENTRY_SNOOP) != 0 ? DMAR_PTE_SNP : 0) |
476 	    ((eflags & IOMMU_MAP_ENTRY_TM) != 0 ? DMAR_PTE_TM : 0);
477 
478 	domain = IODOM2DOM(iodom);
479 	unit = domain->dmar;
480 
481 	KASSERT((iodom->flags & IOMMU_DOMAIN_IDMAP) == 0,
482 	    ("modifying idmap pagetable domain %p", domain));
483 	KASSERT((base & IOMMU_PAGE_MASK) == 0,
484 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
485 	    (uintmax_t)size));
486 	KASSERT((size & IOMMU_PAGE_MASK) == 0,
487 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
488 	    (uintmax_t)size));
489 	KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base,
490 	    (uintmax_t)size));
491 	KASSERT(base < (1ULL << domain->agaw),
492 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
493 	    (uintmax_t)size, domain->agaw));
494 	KASSERT(base + size < (1ULL << domain->agaw),
495 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
496 	    (uintmax_t)size, domain->agaw));
497 	KASSERT(base + size > base,
498 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
499 	    (uintmax_t)size));
500 	KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0,
501 	    ("neither read nor write %jx", (uintmax_t)pflags));
502 	KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP |
503 	    DMAR_PTE_TM)) == 0,
504 	    ("invalid pte flags %jx", (uintmax_t)pflags));
505 	KASSERT((pflags & DMAR_PTE_SNP) == 0 ||
506 	    (unit->hw_ecap & DMAR_ECAP_SC) != 0,
507 	    ("PTE_SNP for dmar without snoop control %p %jx",
508 	    domain, (uintmax_t)pflags));
509 	KASSERT((pflags & DMAR_PTE_TM) == 0 ||
510 	    (unit->hw_ecap & DMAR_ECAP_DI) != 0,
511 	    ("PTE_TM for dmar without DIOTLB %p %jx",
512 	    domain, (uintmax_t)pflags));
513 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
514 
515 	DMAR_DOMAIN_PGLOCK(domain);
516 	error = domain_map_buf_locked(domain, base, size, ma, pflags, flags);
517 	DMAR_DOMAIN_PGUNLOCK(domain);
518 	if (error != 0)
519 		return (error);
520 
521 	if ((unit->hw_cap & DMAR_CAP_CM) != 0)
522 		domain_flush_iotlb_sync(domain, base, size);
523 	else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
524 		/* See 11.1 Write Buffer Flushing. */
525 		DMAR_LOCK(unit);
526 		dmar_flush_write_bufs(unit);
527 		DMAR_UNLOCK(unit);
528 	}
529 	return (0);
530 }
531 
532 static void domain_unmap_clear_pte(struct dmar_domain *domain,
533     iommu_gaddr_t base, int lvl, int flags, iommu_pte_t *pte,
534     struct sf_buf **sf, bool free_fs);
535 
536 static void
537 domain_free_pgtbl_pde(struct dmar_domain *domain, iommu_gaddr_t base,
538     int lvl, int flags)
539 {
540 	struct sf_buf *sf;
541 	iommu_pte_t *pde;
542 	vm_pindex_t idx;
543 
544 	sf = NULL;
545 	pde = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
546 	domain_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, true);
547 }
548 
549 static void
550 domain_unmap_clear_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
551     int flags, iommu_pte_t *pte, struct sf_buf **sf, bool free_sf)
552 {
553 	vm_page_t m;
554 
555 	dmar_pte_clear(&pte->pte);
556 	dmar_flush_pte_to_ram(domain->dmar, pte);
557 	m = sf_buf_page(*sf);
558 	if (free_sf) {
559 		iommu_unmap_pgtbl(*sf);
560 		*sf = NULL;
561 	}
562 	if (!vm_page_unwire_noq(m))
563 		return;
564 	KASSERT(lvl != 0,
565 	    ("lost reference (lvl) on root pg domain %p base %jx lvl %d",
566 	    domain, (uintmax_t)base, lvl));
567 	KASSERT(m->pindex != 0,
568 	    ("lost reference (idx) on root pg domain %p base %jx lvl %d",
569 	    domain, (uintmax_t)base, lvl));
570 	iommu_pgfree(domain->pgtbl_obj, m->pindex, flags);
571 	domain_free_pgtbl_pde(domain, base, lvl - 1, flags);
572 }
573 
574 /*
575  * Assumes that the unmap is never partial.
576  */
577 static int
578 domain_unmap_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
579     iommu_gaddr_t size, int flags)
580 {
581 	iommu_pte_t *pte;
582 	struct sf_buf *sf;
583 	vm_pindex_t idx;
584 	iommu_gaddr_t pg_sz;
585 	int lvl;
586 
587 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
588 	if (size == 0)
589 		return (0);
590 
591 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
592 	    ("modifying idmap pagetable domain %p", domain));
593 	KASSERT((base & IOMMU_PAGE_MASK) == 0,
594 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
595 	    (uintmax_t)size));
596 	KASSERT((size & IOMMU_PAGE_MASK) == 0,
597 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
598 	    (uintmax_t)size));
599 	KASSERT(base < (1ULL << domain->agaw),
600 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
601 	    (uintmax_t)size, domain->agaw));
602 	KASSERT(base + size < (1ULL << domain->agaw),
603 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
604 	    (uintmax_t)size, domain->agaw));
605 	KASSERT(base + size > base,
606 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
607 	    (uintmax_t)size));
608 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
609 
610 	pg_sz = 0; /* silence gcc */
611 	flags |= IOMMU_PGF_OBJL;
612 	TD_PREP_PINNED_ASSERT;
613 
614 	for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) {
615 		for (lvl = 0; lvl < domain->pglvl; lvl++) {
616 			if (lvl != domain->pglvl - 1 &&
617 			    !domain_is_sp_lvl(domain, lvl))
618 				continue;
619 			pg_sz = domain_page_size(domain, lvl);
620 			if (pg_sz > size)
621 				continue;
622 			pte = domain_pgtbl_map_pte(domain, base, lvl, flags,
623 			    &idx, &sf);
624 			KASSERT(pte != NULL,
625 			    ("sleeping or page missed %p %jx %d 0x%x",
626 			    domain, (uintmax_t)base, lvl, flags));
627 			if ((pte->pte & DMAR_PTE_SP) != 0 ||
628 			    lvl == domain->pglvl - 1) {
629 				domain_unmap_clear_pte(domain, base, lvl,
630 				    flags, pte, &sf, false);
631 				break;
632 			}
633 		}
634 		KASSERT(size >= pg_sz,
635 		    ("unmapping loop overflow %p %jx %jx %jx", domain,
636 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
637 	}
638 	if (sf != NULL)
639 		iommu_unmap_pgtbl(sf);
640 	/*
641 	 * See 11.1 Write Buffer Flushing for an explanation why RWBF
642 	 * can be ignored there.
643 	 */
644 
645 	TD_PINNED_ASSERT;
646 	return (0);
647 }
648 
649 static int
650 domain_unmap_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
651     iommu_gaddr_t size, int flags)
652 {
653 	struct dmar_domain *domain;
654 	int error;
655 
656 	domain = IODOM2DOM(iodom);
657 
658 	DMAR_DOMAIN_PGLOCK(domain);
659 	error = domain_unmap_buf_locked(domain, base, size, flags);
660 	DMAR_DOMAIN_PGUNLOCK(domain);
661 	return (error);
662 }
663 
664 int
665 dmar_domain_alloc_pgtbl(struct dmar_domain *domain)
666 {
667 	vm_page_t m;
668 
669 	KASSERT(domain->pgtbl_obj == NULL,
670 	    ("already initialized %p", domain));
671 
672 	domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
673 	    IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL);
674 	DMAR_DOMAIN_PGLOCK(domain);
675 	m = iommu_pgalloc(domain->pgtbl_obj, 0, IOMMU_PGF_WAITOK |
676 	    IOMMU_PGF_ZERO | IOMMU_PGF_OBJL);
677 	/* No implicit free of the top level page table page. */
678 	vm_page_wire(m);
679 	DMAR_DOMAIN_PGUNLOCK(domain);
680 	DMAR_LOCK(domain->dmar);
681 	domain->iodom.flags |= IOMMU_DOMAIN_PGTBL_INITED;
682 	DMAR_UNLOCK(domain->dmar);
683 	return (0);
684 }
685 
686 void
687 dmar_domain_free_pgtbl(struct dmar_domain *domain)
688 {
689 	vm_object_t obj;
690 	vm_page_t m;
691 
692 	obj = domain->pgtbl_obj;
693 	if (obj == NULL) {
694 		KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 &&
695 		    (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0,
696 		    ("lost pagetable object domain %p", domain));
697 		return;
698 	}
699 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
700 	domain->pgtbl_obj = NULL;
701 
702 	if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0) {
703 		put_idmap_pgtbl(obj);
704 		domain->iodom.flags &= ~IOMMU_DOMAIN_IDMAP;
705 		return;
706 	}
707 
708 	/* Obliterate ref_counts */
709 	VM_OBJECT_ASSERT_WLOCKED(obj);
710 	for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m)) {
711 		vm_page_clearref(m);
712 		vm_wire_sub(1);
713 	}
714 	VM_OBJECT_WUNLOCK(obj);
715 	vm_object_deallocate(obj);
716 }
717 
718 static inline uint64_t
719 domain_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro)
720 {
721 	uint64_t iotlbr;
722 
723 	dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
724 	    DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt);
725 	for (;;) {
726 		iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF);
727 		if ((iotlbr & DMAR_IOTLB_IVT) == 0)
728 			break;
729 		cpu_spinwait();
730 	}
731 	return (iotlbr);
732 }
733 
734 void
735 domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
736     iommu_gaddr_t size)
737 {
738 	struct dmar_unit *unit;
739 	iommu_gaddr_t isize;
740 	uint64_t iotlbr;
741 	int am, iro;
742 
743 	unit = domain->dmar;
744 	KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call",
745 	    unit->iommu.unit));
746 	iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16;
747 	DMAR_LOCK(unit);
748 	if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) {
749 		iotlbr = domain_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM |
750 		    DMAR_IOTLB_DID(domain->domain), iro);
751 		KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
752 		    DMAR_IOTLB_IAIG_INVLD,
753 		    ("dmar%d: invalidation failed %jx", unit->iommu.unit,
754 		    (uintmax_t)iotlbr));
755 	} else {
756 		for (; size > 0; base += isize, size -= isize) {
757 			am = calc_am(unit, base, size, &isize);
758 			dmar_write8(unit, iro, base | am);
759 			iotlbr = domain_wait_iotlb_flush(unit,
760 			    DMAR_IOTLB_IIRG_PAGE |
761 			    DMAR_IOTLB_DID(domain->domain), iro);
762 			KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
763 			    DMAR_IOTLB_IAIG_INVLD,
764 			    ("dmar%d: PSI invalidation failed "
765 			    "iotlbr 0x%jx base 0x%jx size 0x%jx am %d",
766 			    unit->iommu.unit, (uintmax_t)iotlbr,
767 			    (uintmax_t)base, (uintmax_t)size, am));
768 			/*
769 			 * Any non-page granularity covers whole guest
770 			 * address space for the domain.
771 			 */
772 			if ((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
773 			    DMAR_IOTLB_IAIG_PAGE)
774 				break;
775 		}
776 	}
777 	DMAR_UNLOCK(unit);
778 }
779 
780 const struct iommu_domain_map_ops dmar_domain_map_ops = {
781 	.map = domain_map_buf,
782 	.unmap = domain_unmap_buf,
783 };
784