xref: /freebsd/sys/x86/iommu/intel_utils.c (revision 6bfca4dcab07dad45a805879d954876b353c0810)
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/bus.h>
33 #include <sys/kernel.h>
34 #include <sys/lock.h>
35 #include <sys/malloc.h>
36 #include <sys/memdesc.h>
37 #include <sys/mutex.h>
38 #include <sys/proc.h>
39 #include <sys/queue.h>
40 #include <sys/rman.h>
41 #include <sys/rwlock.h>
42 #include <sys/sched.h>
43 #include <sys/sf_buf.h>
44 #include <sys/sysctl.h>
45 #include <sys/systm.h>
46 #include <sys/taskqueue.h>
47 #include <sys/time.h>
48 #include <sys/tree.h>
49 #include <sys/vmem.h>
50 #include <vm/vm.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_kern.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_map.h>
56 #include <vm/vm_pageout.h>
57 #include <dev/pci/pcireg.h>
58 #include <dev/pci/pcivar.h>
59 #include <machine/bus.h>
60 #include <machine/cpu.h>
61 #include <machine/intr_machdep.h>
62 #include <x86/include/apicvar.h>
63 #include <x86/include/busdma_impl.h>
64 #include <dev/iommu/busdma_iommu.h>
65 #include <x86/iommu/intel_reg.h>
66 #include <x86/iommu/intel_dmar.h>
67 
68 u_int
69 dmar_nd2mask(u_int nd)
70 {
71 	static const u_int masks[] = {
72 		0x000f,	/* nd == 0 */
73 		0x002f,	/* nd == 1 */
74 		0x00ff,	/* nd == 2 */
75 		0x02ff,	/* nd == 3 */
76 		0x0fff,	/* nd == 4 */
77 		0x2fff,	/* nd == 5 */
78 		0xffff,	/* nd == 6 */
79 		0x0000,	/* nd == 7 reserved */
80 	};
81 
82 	KASSERT(nd <= 6, ("number of domains %d", nd));
83 	return (masks[nd]);
84 }
85 
86 static const struct sagaw_bits_tag {
87 	int agaw;
88 	int cap;
89 	int awlvl;
90 	int pglvl;
91 } sagaw_bits[] = {
92 	{.agaw = 30, .cap = DMAR_CAP_SAGAW_2LVL, .awlvl = DMAR_CTX2_AW_2LVL,
93 	    .pglvl = 2},
94 	{.agaw = 39, .cap = DMAR_CAP_SAGAW_3LVL, .awlvl = DMAR_CTX2_AW_3LVL,
95 	    .pglvl = 3},
96 	{.agaw = 48, .cap = DMAR_CAP_SAGAW_4LVL, .awlvl = DMAR_CTX2_AW_4LVL,
97 	    .pglvl = 4},
98 	{.agaw = 57, .cap = DMAR_CAP_SAGAW_5LVL, .awlvl = DMAR_CTX2_AW_5LVL,
99 	    .pglvl = 5}
100 	/*
101 	 * 6-level paging (DMAR_CAP_SAGAW_6LVL) is not supported on any
102 	 * current VT-d hardware and its SAGAW field value is listed as
103 	 * reserved in the VT-d spec.  If support is added in the future,
104 	 * this structure and the logic in dmar_maxaddr2mgaw() will need
105 	 * to change to avoid attempted comparison against 1ULL << 64.
106 	 */
107 };
108 
109 bool
110 dmar_pglvl_supported(struct dmar_unit *unit, int pglvl)
111 {
112 	int i;
113 
114 	for (i = 0; i < nitems(sagaw_bits); i++) {
115 		if (sagaw_bits[i].pglvl != pglvl)
116 			continue;
117 		if ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
118 			return (true);
119 	}
120 	return (false);
121 }
122 
123 int
124 domain_set_agaw(struct dmar_domain *domain, int mgaw)
125 {
126 	int sagaw, i;
127 
128 	domain->mgaw = mgaw;
129 	sagaw = DMAR_CAP_SAGAW(domain->dmar->hw_cap);
130 	for (i = 0; i < nitems(sagaw_bits); i++) {
131 		if (sagaw_bits[i].agaw >= mgaw) {
132 			domain->agaw = sagaw_bits[i].agaw;
133 			domain->pglvl = sagaw_bits[i].pglvl;
134 			domain->awlvl = sagaw_bits[i].awlvl;
135 			return (0);
136 		}
137 	}
138 	device_printf(domain->dmar->dev,
139 	    "context request mgaw %d: no agaw found, sagaw %x\n",
140 	    mgaw, sagaw);
141 	return (EINVAL);
142 }
143 
144 /*
145  * Find a best fit mgaw for the given maxaddr:
146  *   - if allow_less is false, must find sagaw which maps all requested
147  *     addresses (used by identity mappings);
148  *   - if allow_less is true, and no supported sagaw can map all requested
149  *     address space, accept the biggest sagaw, whatever is it.
150  */
151 int
152 dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr, bool allow_less)
153 {
154 	int i;
155 
156 	for (i = 0; i < nitems(sagaw_bits); i++) {
157 		if ((1ULL << sagaw_bits[i].agaw) >= maxaddr &&
158 		    (DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
159 			break;
160 	}
161 	if (allow_less && i == nitems(sagaw_bits)) {
162 		do {
163 			i--;
164 		} while ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap)
165 		    == 0);
166 	}
167 	if (i < nitems(sagaw_bits))
168 		return (sagaw_bits[i].agaw);
169 	KASSERT(0, ("no mgaw for maxaddr %jx allow_less %d",
170 	    (uintmax_t) maxaddr, allow_less));
171 	return (-1);
172 }
173 
174 /*
175  * Calculate the total amount of page table pages needed to map the
176  * whole bus address space on the context with the selected agaw.
177  */
178 vm_pindex_t
179 pglvl_max_pages(int pglvl)
180 {
181 	vm_pindex_t res;
182 	int i;
183 
184 	for (res = 0, i = pglvl; i > 0; i--) {
185 		res *= DMAR_NPTEPG;
186 		res++;
187 	}
188 	return (res);
189 }
190 
191 /*
192  * Return true if the page table level lvl supports the superpage for
193  * the context ctx.
194  */
195 int
196 domain_is_sp_lvl(struct dmar_domain *domain, int lvl)
197 {
198 	int alvl, cap_sps;
199 	static const int sagaw_sp[] = {
200 		DMAR_CAP_SPS_2M,
201 		DMAR_CAP_SPS_1G,
202 		DMAR_CAP_SPS_512G,
203 		DMAR_CAP_SPS_1T
204 	};
205 
206 	alvl = domain->pglvl - lvl - 1;
207 	cap_sps = DMAR_CAP_SPS(domain->dmar->hw_cap);
208 	return (alvl < nitems(sagaw_sp) && (sagaw_sp[alvl] & cap_sps) != 0);
209 }
210 
211 iommu_gaddr_t
212 pglvl_page_size(int total_pglvl, int lvl)
213 {
214 	int rlvl;
215 	static const iommu_gaddr_t pg_sz[] = {
216 		(iommu_gaddr_t)DMAR_PAGE_SIZE,
217 		(iommu_gaddr_t)DMAR_PAGE_SIZE << DMAR_NPTEPGSHIFT,
218 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (2 * DMAR_NPTEPGSHIFT),
219 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (3 * DMAR_NPTEPGSHIFT),
220 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (4 * DMAR_NPTEPGSHIFT),
221 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (5 * DMAR_NPTEPGSHIFT)
222 	};
223 
224 	KASSERT(lvl >= 0 && lvl < total_pglvl,
225 	    ("total %d lvl %d", total_pglvl, lvl));
226 	rlvl = total_pglvl - lvl - 1;
227 	KASSERT(rlvl < nitems(pg_sz), ("sizeof pg_sz lvl %d", lvl));
228 	return (pg_sz[rlvl]);
229 }
230 
231 iommu_gaddr_t
232 domain_page_size(struct dmar_domain *domain, int lvl)
233 {
234 
235 	return (pglvl_page_size(domain->pglvl, lvl));
236 }
237 
238 int
239 calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size,
240     iommu_gaddr_t *isizep)
241 {
242 	iommu_gaddr_t isize;
243 	int am;
244 
245 	for (am = DMAR_CAP_MAMV(unit->hw_cap);; am--) {
246 		isize = 1ULL << (am + DMAR_PAGE_SHIFT);
247 		if ((base & (isize - 1)) == 0 && size >= isize)
248 			break;
249 		if (am == 0)
250 			break;
251 	}
252 	*isizep = isize;
253 	return (am);
254 }
255 
256 iommu_haddr_t dmar_high;
257 int haw;
258 int dmar_tbl_pagecnt;
259 
260 vm_page_t
261 dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags)
262 {
263 	vm_page_t m;
264 	int zeroed, aflags;
265 
266 	zeroed = (flags & IOMMU_PGF_ZERO) != 0 ? VM_ALLOC_ZERO : 0;
267 	aflags = zeroed | VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_NODUMP |
268 	    ((flags & IOMMU_PGF_WAITOK) != 0 ? VM_ALLOC_WAITFAIL :
269 	    VM_ALLOC_NOWAIT);
270 	for (;;) {
271 		if ((flags & IOMMU_PGF_OBJL) == 0)
272 			VM_OBJECT_WLOCK(obj);
273 		m = vm_page_lookup(obj, idx);
274 		if ((flags & IOMMU_PGF_NOALLOC) != 0 || m != NULL) {
275 			if ((flags & IOMMU_PGF_OBJL) == 0)
276 				VM_OBJECT_WUNLOCK(obj);
277 			break;
278 		}
279 		m = vm_page_alloc_contig(obj, idx, aflags, 1, 0,
280 		    dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
281 		if ((flags & IOMMU_PGF_OBJL) == 0)
282 			VM_OBJECT_WUNLOCK(obj);
283 		if (m != NULL) {
284 			if (zeroed && (m->flags & PG_ZERO) == 0)
285 				pmap_zero_page(m);
286 			atomic_add_int(&dmar_tbl_pagecnt, 1);
287 			break;
288 		}
289 		if ((flags & IOMMU_PGF_WAITOK) == 0)
290 			break;
291 	}
292 	return (m);
293 }
294 
295 void
296 dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags)
297 {
298 	vm_page_t m;
299 
300 	if ((flags & IOMMU_PGF_OBJL) == 0)
301 		VM_OBJECT_WLOCK(obj);
302 	m = vm_page_grab(obj, idx, VM_ALLOC_NOCREAT);
303 	if (m != NULL) {
304 		vm_page_free(m);
305 		atomic_subtract_int(&dmar_tbl_pagecnt, 1);
306 	}
307 	if ((flags & IOMMU_PGF_OBJL) == 0)
308 		VM_OBJECT_WUNLOCK(obj);
309 }
310 
311 void *
312 dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags,
313     struct sf_buf **sf)
314 {
315 	vm_page_t m;
316 	bool allocated;
317 
318 	if ((flags & IOMMU_PGF_OBJL) == 0)
319 		VM_OBJECT_WLOCK(obj);
320 	m = vm_page_lookup(obj, idx);
321 	if (m == NULL && (flags & IOMMU_PGF_ALLOC) != 0) {
322 		m = dmar_pgalloc(obj, idx, flags | IOMMU_PGF_OBJL);
323 		allocated = true;
324 	} else
325 		allocated = false;
326 	if (m == NULL) {
327 		if ((flags & IOMMU_PGF_OBJL) == 0)
328 			VM_OBJECT_WUNLOCK(obj);
329 		return (NULL);
330 	}
331 	/* Sleepable allocations cannot fail. */
332 	if ((flags & IOMMU_PGF_WAITOK) != 0)
333 		VM_OBJECT_WUNLOCK(obj);
334 	sched_pin();
335 	*sf = sf_buf_alloc(m, SFB_CPUPRIVATE | ((flags & IOMMU_PGF_WAITOK)
336 	    == 0 ? SFB_NOWAIT : 0));
337 	if (*sf == NULL) {
338 		sched_unpin();
339 		if (allocated) {
340 			VM_OBJECT_ASSERT_WLOCKED(obj);
341 			dmar_pgfree(obj, m->pindex, flags | IOMMU_PGF_OBJL);
342 		}
343 		if ((flags & IOMMU_PGF_OBJL) == 0)
344 			VM_OBJECT_WUNLOCK(obj);
345 		return (NULL);
346 	}
347 	if ((flags & (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL)) ==
348 	    (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL))
349 		VM_OBJECT_WLOCK(obj);
350 	else if ((flags & (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL)) == 0)
351 		VM_OBJECT_WUNLOCK(obj);
352 	return ((void *)sf_buf_kva(*sf));
353 }
354 
355 void
356 dmar_unmap_pgtbl(struct sf_buf *sf)
357 {
358 
359 	sf_buf_free(sf);
360 	sched_unpin();
361 }
362 
363 static void
364 dmar_flush_transl_to_ram(struct dmar_unit *unit, void *dst, size_t sz)
365 {
366 
367 	if (DMAR_IS_COHERENT(unit))
368 		return;
369 	/*
370 	 * If DMAR does not snoop paging structures accesses, flush
371 	 * CPU cache to memory.
372 	 */
373 	pmap_force_invalidate_cache_range((uintptr_t)dst, (uintptr_t)dst + sz);
374 }
375 
376 void
377 dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst)
378 {
379 
380 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
381 }
382 
383 void
384 dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst)
385 {
386 
387 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
388 }
389 
390 void
391 dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst)
392 {
393 
394 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
395 }
396 
397 /*
398  * Load the root entry pointer into the hardware, busily waiting for
399  * the completion.
400  */
401 int
402 dmar_load_root_entry_ptr(struct dmar_unit *unit)
403 {
404 	vm_page_t root_entry;
405 	int error;
406 
407 	/*
408 	 * Access to the GCMD register must be serialized while the
409 	 * command is submitted.
410 	 */
411 	DMAR_ASSERT_LOCKED(unit);
412 
413 	VM_OBJECT_RLOCK(unit->ctx_obj);
414 	root_entry = vm_page_lookup(unit->ctx_obj, 0);
415 	VM_OBJECT_RUNLOCK(unit->ctx_obj);
416 	dmar_write8(unit, DMAR_RTADDR_REG, VM_PAGE_TO_PHYS(root_entry));
417 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SRTP);
418 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_RTPS)
419 	    != 0));
420 	return (error);
421 }
422 
423 /*
424  * Globally invalidate the context entries cache, busily waiting for
425  * the completion.
426  */
427 int
428 dmar_inv_ctx_glob(struct dmar_unit *unit)
429 {
430 	int error;
431 
432 	/*
433 	 * Access to the CCMD register must be serialized while the
434 	 * command is submitted.
435 	 */
436 	DMAR_ASSERT_LOCKED(unit);
437 	KASSERT(!unit->qi_enabled, ("QI enabled"));
438 
439 	/*
440 	 * The DMAR_CCMD_ICC bit in the upper dword should be written
441 	 * after the low dword write is completed.  Amd64
442 	 * dmar_write8() does not have this issue, i386 dmar_write8()
443 	 * writes the upper dword last.
444 	 */
445 	dmar_write8(unit, DMAR_CCMD_REG, DMAR_CCMD_ICC | DMAR_CCMD_CIRG_GLOB);
446 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_CCMD_REG + 4) & DMAR_CCMD_ICC32)
447 	    == 0));
448 	return (error);
449 }
450 
451 /*
452  * Globally invalidate the IOTLB, busily waiting for the completion.
453  */
454 int
455 dmar_inv_iotlb_glob(struct dmar_unit *unit)
456 {
457 	int error, reg;
458 
459 	DMAR_ASSERT_LOCKED(unit);
460 	KASSERT(!unit->qi_enabled, ("QI enabled"));
461 
462 	reg = 16 * DMAR_ECAP_IRO(unit->hw_ecap);
463 	/* See a comment about DMAR_CCMD_ICC in dmar_inv_ctx_glob. */
464 	dmar_write8(unit, reg + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
465 	    DMAR_IOTLB_IIRG_GLB | DMAR_IOTLB_DR | DMAR_IOTLB_DW);
466 	DMAR_WAIT_UNTIL(((dmar_read4(unit, reg + DMAR_IOTLB_REG_OFF + 4) &
467 	    DMAR_IOTLB_IVT32) == 0));
468 	return (error);
469 }
470 
471 /*
472  * Flush the chipset write buffers.  See 11.1 "Write Buffer Flushing"
473  * in the architecture specification.
474  */
475 int
476 dmar_flush_write_bufs(struct dmar_unit *unit)
477 {
478 	int error;
479 
480 	DMAR_ASSERT_LOCKED(unit);
481 
482 	/*
483 	 * DMAR_GCMD_WBF is only valid when CAP_RWBF is reported.
484 	 */
485 	KASSERT((unit->hw_cap & DMAR_CAP_RWBF) != 0,
486 	    ("dmar%d: no RWBF", unit->iommu.unit));
487 
488 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_WBF);
489 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_WBFS)
490 	    != 0));
491 	return (error);
492 }
493 
494 /*
495  * Some BIOSes protect memory region they reside in by using DMAR to
496  * prevent devices from doing any DMA transactions to that part of RAM.
497  * AMI refers to this as "DMA Control Guarantee".
498  * We need to disable this when address translation is enabled.
499  */
500 int
501 dmar_disable_protected_regions(struct dmar_unit *unit)
502 {
503 	uint32_t reg;
504 	int error;
505 
506 	DMAR_ASSERT_LOCKED(unit);
507 
508 	/* Check if we support the feature. */
509 	if ((unit->hw_cap & (DMAR_CAP_PLMR | DMAR_CAP_PHMR)) == 0)
510 		return (0);
511 
512 	reg = dmar_read4(unit, DMAR_PMEN_REG);
513 	if ((reg & DMAR_PMEN_EPM) == 0)
514 		return (0);
515 
516 	reg &= ~DMAR_PMEN_EPM;
517 	dmar_write4(unit, DMAR_PMEN_REG, reg);
518 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_PMEN_REG) & DMAR_PMEN_PRS)
519 	    != 0));
520 
521 	return (error);
522 }
523 
524 int
525 dmar_enable_translation(struct dmar_unit *unit)
526 {
527 	int error;
528 
529 	DMAR_ASSERT_LOCKED(unit);
530 	unit->hw_gcmd |= DMAR_GCMD_TE;
531 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
532 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
533 	    != 0));
534 	return (error);
535 }
536 
537 int
538 dmar_disable_translation(struct dmar_unit *unit)
539 {
540 	int error;
541 
542 	DMAR_ASSERT_LOCKED(unit);
543 	unit->hw_gcmd &= ~DMAR_GCMD_TE;
544 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
545 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
546 	    == 0));
547 	return (error);
548 }
549 
550 int
551 dmar_load_irt_ptr(struct dmar_unit *unit)
552 {
553 	uint64_t irta, s;
554 	int error;
555 
556 	DMAR_ASSERT_LOCKED(unit);
557 	irta = unit->irt_phys;
558 	if (DMAR_X2APIC(unit))
559 		irta |= DMAR_IRTA_EIME;
560 	s = fls(unit->irte_cnt) - 2;
561 	KASSERT(unit->irte_cnt >= 2 && s <= DMAR_IRTA_S_MASK &&
562 	    powerof2(unit->irte_cnt),
563 	    ("IRTA_REG_S overflow %x", unit->irte_cnt));
564 	irta |= s;
565 	dmar_write8(unit, DMAR_IRTA_REG, irta);
566 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SIRTP);
567 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRTPS)
568 	    != 0));
569 	return (error);
570 }
571 
572 int
573 dmar_enable_ir(struct dmar_unit *unit)
574 {
575 	int error;
576 
577 	DMAR_ASSERT_LOCKED(unit);
578 	unit->hw_gcmd |= DMAR_GCMD_IRE;
579 	unit->hw_gcmd &= ~DMAR_GCMD_CFI;
580 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
581 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
582 	    != 0));
583 	return (error);
584 }
585 
586 int
587 dmar_disable_ir(struct dmar_unit *unit)
588 {
589 	int error;
590 
591 	DMAR_ASSERT_LOCKED(unit);
592 	unit->hw_gcmd &= ~DMAR_GCMD_IRE;
593 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
594 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
595 	    == 0));
596 	return (error);
597 }
598 
599 #define BARRIER_F				\
600 	u_int f_done, f_inproc, f_wakeup;	\
601 						\
602 	f_done = 1 << (barrier_id * 3);		\
603 	f_inproc = 1 << (barrier_id * 3 + 1);	\
604 	f_wakeup = 1 << (barrier_id * 3 + 2)
605 
606 bool
607 dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id)
608 {
609 	BARRIER_F;
610 
611 	DMAR_LOCK(dmar);
612 	if ((dmar->barrier_flags & f_done) != 0) {
613 		DMAR_UNLOCK(dmar);
614 		return (false);
615 	}
616 
617 	if ((dmar->barrier_flags & f_inproc) != 0) {
618 		while ((dmar->barrier_flags & f_inproc) != 0) {
619 			dmar->barrier_flags |= f_wakeup;
620 			msleep(&dmar->barrier_flags, &dmar->iommu.lock, 0,
621 			    "dmarb", 0);
622 		}
623 		KASSERT((dmar->barrier_flags & f_done) != 0,
624 		    ("dmar%d barrier %d missing done", dmar->iommu.unit,
625 		    barrier_id));
626 		DMAR_UNLOCK(dmar);
627 		return (false);
628 	}
629 
630 	dmar->barrier_flags |= f_inproc;
631 	DMAR_UNLOCK(dmar);
632 	return (true);
633 }
634 
635 void
636 dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id)
637 {
638 	BARRIER_F;
639 
640 	DMAR_ASSERT_LOCKED(dmar);
641 	KASSERT((dmar->barrier_flags & (f_done | f_inproc)) == f_inproc,
642 	    ("dmar%d barrier %d missed entry", dmar->iommu.unit, barrier_id));
643 	dmar->barrier_flags |= f_done;
644 	if ((dmar->barrier_flags & f_wakeup) != 0)
645 		wakeup(&dmar->barrier_flags);
646 	dmar->barrier_flags &= ~(f_inproc | f_wakeup);
647 	DMAR_UNLOCK(dmar);
648 }
649 
650 int dmar_batch_coalesce = 100;
651 struct timespec dmar_hw_timeout = {
652 	.tv_sec = 0,
653 	.tv_nsec = 1000000
654 };
655 
656 static const uint64_t d = 1000000000;
657 
658 void
659 dmar_update_timeout(uint64_t newval)
660 {
661 
662 	/* XXXKIB not atomic */
663 	dmar_hw_timeout.tv_sec = newval / d;
664 	dmar_hw_timeout.tv_nsec = newval % d;
665 }
666 
667 uint64_t
668 dmar_get_timeout(void)
669 {
670 
671 	return ((uint64_t)dmar_hw_timeout.tv_sec * d +
672 	    dmar_hw_timeout.tv_nsec);
673 }
674 
675 static int
676 dmar_timeout_sysctl(SYSCTL_HANDLER_ARGS)
677 {
678 	uint64_t val;
679 	int error;
680 
681 	val = dmar_get_timeout();
682 	error = sysctl_handle_long(oidp, &val, 0, req);
683 	if (error != 0 || req->newptr == NULL)
684 		return (error);
685 	dmar_update_timeout(val);
686 	return (error);
687 }
688 
689 static SYSCTL_NODE(_hw_iommu, OID_AUTO, dmar, CTLFLAG_RD | CTLFLAG_MPSAFE,
690     NULL, "");
691 SYSCTL_INT(_hw_iommu_dmar, OID_AUTO, tbl_pagecnt, CTLFLAG_RD,
692     &dmar_tbl_pagecnt, 0,
693     "Count of pages used for DMAR pagetables");
694 SYSCTL_INT(_hw_iommu_dmar, OID_AUTO, batch_coalesce, CTLFLAG_RWTUN,
695     &dmar_batch_coalesce, 0,
696     "Number of qi batches between interrupt");
697 SYSCTL_PROC(_hw_iommu_dmar, OID_AUTO, timeout,
698     CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
699     dmar_timeout_sysctl, "QU",
700     "Timeout for command wait, in nanoseconds");
701