xref: /freebsd/sys/x86/iommu/intel_utils.c (revision 123af6ec70016f5556da5972d4d63c7d175c06d3)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/bus.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/malloc.h>
40 #include <sys/memdesc.h>
41 #include <sys/mutex.h>
42 #include <sys/proc.h>
43 #include <sys/queue.h>
44 #include <sys/rman.h>
45 #include <sys/rwlock.h>
46 #include <sys/sched.h>
47 #include <sys/sf_buf.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 #include <sys/taskqueue.h>
51 #include <sys/time.h>
52 #include <sys/tree.h>
53 #include <sys/vmem.h>
54 #include <dev/pci/pcivar.h>
55 #include <vm/vm.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_object.h>
59 #include <vm/vm_page.h>
60 #include <vm/vm_map.h>
61 #include <vm/vm_pageout.h>
62 #include <machine/bus.h>
63 #include <machine/cpu.h>
64 #include <machine/intr_machdep.h>
65 #include <x86/include/apicvar.h>
66 #include <x86/include/busdma_impl.h>
67 #include <x86/iommu/intel_reg.h>
68 #include <x86/iommu/busdma_dmar.h>
69 #include <x86/iommu/intel_dmar.h>
70 
71 u_int
72 dmar_nd2mask(u_int nd)
73 {
74 	static const u_int masks[] = {
75 		0x000f,	/* nd == 0 */
76 		0x002f,	/* nd == 1 */
77 		0x00ff,	/* nd == 2 */
78 		0x02ff,	/* nd == 3 */
79 		0x0fff,	/* nd == 4 */
80 		0x2fff,	/* nd == 5 */
81 		0xffff,	/* nd == 6 */
82 		0x0000,	/* nd == 7 reserved */
83 	};
84 
85 	KASSERT(nd <= 6, ("number of domains %d", nd));
86 	return (masks[nd]);
87 }
88 
89 static const struct sagaw_bits_tag {
90 	int agaw;
91 	int cap;
92 	int awlvl;
93 	int pglvl;
94 } sagaw_bits[] = {
95 	{.agaw = 30, .cap = DMAR_CAP_SAGAW_2LVL, .awlvl = DMAR_CTX2_AW_2LVL,
96 	    .pglvl = 2},
97 	{.agaw = 39, .cap = DMAR_CAP_SAGAW_3LVL, .awlvl = DMAR_CTX2_AW_3LVL,
98 	    .pglvl = 3},
99 	{.agaw = 48, .cap = DMAR_CAP_SAGAW_4LVL, .awlvl = DMAR_CTX2_AW_4LVL,
100 	    .pglvl = 4},
101 	{.agaw = 57, .cap = DMAR_CAP_SAGAW_5LVL, .awlvl = DMAR_CTX2_AW_5LVL,
102 	    .pglvl = 5},
103 	{.agaw = 64, .cap = DMAR_CAP_SAGAW_6LVL, .awlvl = DMAR_CTX2_AW_6LVL,
104 	    .pglvl = 6}
105 };
106 
107 bool
108 dmar_pglvl_supported(struct dmar_unit *unit, int pglvl)
109 {
110 	int i;
111 
112 	for (i = 0; i < nitems(sagaw_bits); i++) {
113 		if (sagaw_bits[i].pglvl != pglvl)
114 			continue;
115 		if ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
116 			return (true);
117 	}
118 	return (false);
119 }
120 
121 int
122 domain_set_agaw(struct dmar_domain *domain, int mgaw)
123 {
124 	int sagaw, i;
125 
126 	domain->mgaw = mgaw;
127 	sagaw = DMAR_CAP_SAGAW(domain->dmar->hw_cap);
128 	for (i = 0; i < nitems(sagaw_bits); i++) {
129 		if (sagaw_bits[i].agaw >= mgaw) {
130 			domain->agaw = sagaw_bits[i].agaw;
131 			domain->pglvl = sagaw_bits[i].pglvl;
132 			domain->awlvl = sagaw_bits[i].awlvl;
133 			return (0);
134 		}
135 	}
136 	device_printf(domain->dmar->dev,
137 	    "context request mgaw %d: no agaw found, sagaw %x\n",
138 	    mgaw, sagaw);
139 	return (EINVAL);
140 }
141 
142 /*
143  * Find a best fit mgaw for the given maxaddr:
144  *   - if allow_less is false, must find sagaw which maps all requested
145  *     addresses (used by identity mappings);
146  *   - if allow_less is true, and no supported sagaw can map all requested
147  *     address space, accept the biggest sagaw, whatever is it.
148  */
149 int
150 dmar_maxaddr2mgaw(struct dmar_unit *unit, dmar_gaddr_t maxaddr, bool allow_less)
151 {
152 	int i;
153 
154 	for (i = 0; i < nitems(sagaw_bits); i++) {
155 		if ((1ULL << sagaw_bits[i].agaw) >= maxaddr &&
156 		    (DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
157 			break;
158 	}
159 	if (allow_less && i == nitems(sagaw_bits)) {
160 		do {
161 			i--;
162 		} while ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap)
163 		    == 0);
164 	}
165 	if (i < nitems(sagaw_bits))
166 		return (sagaw_bits[i].agaw);
167 	KASSERT(0, ("no mgaw for maxaddr %jx allow_less %d",
168 	    (uintmax_t) maxaddr, allow_less));
169 	return (-1);
170 }
171 
172 /*
173  * Calculate the total amount of page table pages needed to map the
174  * whole bus address space on the context with the selected agaw.
175  */
176 vm_pindex_t
177 pglvl_max_pages(int pglvl)
178 {
179 	vm_pindex_t res;
180 	int i;
181 
182 	for (res = 0, i = pglvl; i > 0; i--) {
183 		res *= DMAR_NPTEPG;
184 		res++;
185 	}
186 	return (res);
187 }
188 
189 /*
190  * Return true if the page table level lvl supports the superpage for
191  * the context ctx.
192  */
193 int
194 domain_is_sp_lvl(struct dmar_domain *domain, int lvl)
195 {
196 	int alvl, cap_sps;
197 	static const int sagaw_sp[] = {
198 		DMAR_CAP_SPS_2M,
199 		DMAR_CAP_SPS_1G,
200 		DMAR_CAP_SPS_512G,
201 		DMAR_CAP_SPS_1T
202 	};
203 
204 	alvl = domain->pglvl - lvl - 1;
205 	cap_sps = DMAR_CAP_SPS(domain->dmar->hw_cap);
206 	return (alvl < nitems(sagaw_sp) && (sagaw_sp[alvl] & cap_sps) != 0);
207 }
208 
209 dmar_gaddr_t
210 pglvl_page_size(int total_pglvl, int lvl)
211 {
212 	int rlvl;
213 	static const dmar_gaddr_t pg_sz[] = {
214 		(dmar_gaddr_t)DMAR_PAGE_SIZE,
215 		(dmar_gaddr_t)DMAR_PAGE_SIZE << DMAR_NPTEPGSHIFT,
216 		(dmar_gaddr_t)DMAR_PAGE_SIZE << (2 * DMAR_NPTEPGSHIFT),
217 		(dmar_gaddr_t)DMAR_PAGE_SIZE << (3 * DMAR_NPTEPGSHIFT),
218 		(dmar_gaddr_t)DMAR_PAGE_SIZE << (4 * DMAR_NPTEPGSHIFT),
219 		(dmar_gaddr_t)DMAR_PAGE_SIZE << (5 * DMAR_NPTEPGSHIFT)
220 	};
221 
222 	KASSERT(lvl >= 0 && lvl < total_pglvl,
223 	    ("total %d lvl %d", total_pglvl, lvl));
224 	rlvl = total_pglvl - lvl - 1;
225 	KASSERT(rlvl < nitems(pg_sz), ("sizeof pg_sz lvl %d", lvl));
226 	return (pg_sz[rlvl]);
227 }
228 
229 dmar_gaddr_t
230 domain_page_size(struct dmar_domain *domain, int lvl)
231 {
232 
233 	return (pglvl_page_size(domain->pglvl, lvl));
234 }
235 
236 int
237 calc_am(struct dmar_unit *unit, dmar_gaddr_t base, dmar_gaddr_t size,
238     dmar_gaddr_t *isizep)
239 {
240 	dmar_gaddr_t isize;
241 	int am;
242 
243 	for (am = DMAR_CAP_MAMV(unit->hw_cap);; am--) {
244 		isize = 1ULL << (am + DMAR_PAGE_SHIFT);
245 		if ((base & (isize - 1)) == 0 && size >= isize)
246 			break;
247 		if (am == 0)
248 			break;
249 	}
250 	*isizep = isize;
251 	return (am);
252 }
253 
254 dmar_haddr_t dmar_high;
255 int haw;
256 int dmar_tbl_pagecnt;
257 
258 vm_page_t
259 dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags)
260 {
261 	vm_page_t m;
262 	int zeroed, aflags;
263 
264 	zeroed = (flags & DMAR_PGF_ZERO) != 0 ? VM_ALLOC_ZERO : 0;
265 	aflags = zeroed | VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_NODUMP |
266 	    ((flags & DMAR_PGF_WAITOK) != 0 ? VM_ALLOC_WAITFAIL :
267 	    VM_ALLOC_NOWAIT);
268 	for (;;) {
269 		if ((flags & DMAR_PGF_OBJL) == 0)
270 			VM_OBJECT_WLOCK(obj);
271 		m = vm_page_lookup(obj, idx);
272 		if ((flags & DMAR_PGF_NOALLOC) != 0 || m != NULL) {
273 			if ((flags & DMAR_PGF_OBJL) == 0)
274 				VM_OBJECT_WUNLOCK(obj);
275 			break;
276 		}
277 		m = vm_page_alloc_contig(obj, idx, aflags, 1, 0,
278 		    dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
279 		if ((flags & DMAR_PGF_OBJL) == 0)
280 			VM_OBJECT_WUNLOCK(obj);
281 		if (m != NULL) {
282 			if (zeroed && (m->flags & PG_ZERO) == 0)
283 				pmap_zero_page(m);
284 			atomic_add_int(&dmar_tbl_pagecnt, 1);
285 			break;
286 		}
287 		if ((flags & DMAR_PGF_WAITOK) == 0)
288 			break;
289 	}
290 	return (m);
291 }
292 
293 void
294 dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags)
295 {
296 	vm_page_t m;
297 
298 	if ((flags & DMAR_PGF_OBJL) == 0)
299 		VM_OBJECT_WLOCK(obj);
300 	m = vm_page_lookup(obj, idx);
301 	if (m != NULL) {
302 		vm_page_free(m);
303 		atomic_subtract_int(&dmar_tbl_pagecnt, 1);
304 	}
305 	if ((flags & DMAR_PGF_OBJL) == 0)
306 		VM_OBJECT_WUNLOCK(obj);
307 }
308 
309 void *
310 dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags,
311     struct sf_buf **sf)
312 {
313 	vm_page_t m;
314 	bool allocated;
315 
316 	if ((flags & DMAR_PGF_OBJL) == 0)
317 		VM_OBJECT_WLOCK(obj);
318 	m = vm_page_lookup(obj, idx);
319 	if (m == NULL && (flags & DMAR_PGF_ALLOC) != 0) {
320 		m = dmar_pgalloc(obj, idx, flags | DMAR_PGF_OBJL);
321 		allocated = true;
322 	} else
323 		allocated = false;
324 	if (m == NULL) {
325 		if ((flags & DMAR_PGF_OBJL) == 0)
326 			VM_OBJECT_WUNLOCK(obj);
327 		return (NULL);
328 	}
329 	/* Sleepable allocations cannot fail. */
330 	if ((flags & DMAR_PGF_WAITOK) != 0)
331 		VM_OBJECT_WUNLOCK(obj);
332 	sched_pin();
333 	*sf = sf_buf_alloc(m, SFB_CPUPRIVATE | ((flags & DMAR_PGF_WAITOK)
334 	    == 0 ? SFB_NOWAIT : 0));
335 	if (*sf == NULL) {
336 		sched_unpin();
337 		if (allocated) {
338 			VM_OBJECT_ASSERT_WLOCKED(obj);
339 			dmar_pgfree(obj, m->pindex, flags | DMAR_PGF_OBJL);
340 		}
341 		if ((flags & DMAR_PGF_OBJL) == 0)
342 			VM_OBJECT_WUNLOCK(obj);
343 		return (NULL);
344 	}
345 	if ((flags & (DMAR_PGF_WAITOK | DMAR_PGF_OBJL)) ==
346 	    (DMAR_PGF_WAITOK | DMAR_PGF_OBJL))
347 		VM_OBJECT_WLOCK(obj);
348 	else if ((flags & (DMAR_PGF_WAITOK | DMAR_PGF_OBJL)) == 0)
349 		VM_OBJECT_WUNLOCK(obj);
350 	return ((void *)sf_buf_kva(*sf));
351 }
352 
353 void
354 dmar_unmap_pgtbl(struct sf_buf *sf)
355 {
356 
357 	sf_buf_free(sf);
358 	sched_unpin();
359 }
360 
361 static void
362 dmar_flush_transl_to_ram(struct dmar_unit *unit, void *dst, size_t sz)
363 {
364 
365 	if (DMAR_IS_COHERENT(unit))
366 		return;
367 	/*
368 	 * If DMAR does not snoop paging structures accesses, flush
369 	 * CPU cache to memory.
370 	 */
371 	pmap_force_invalidate_cache_range((uintptr_t)dst, (uintptr_t)dst + sz);
372 }
373 
374 void
375 dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst)
376 {
377 
378 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
379 }
380 
381 void
382 dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst)
383 {
384 
385 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
386 }
387 
388 void
389 dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst)
390 {
391 
392 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
393 }
394 
395 /*
396  * Load the root entry pointer into the hardware, busily waiting for
397  * the completion.
398  */
399 int
400 dmar_load_root_entry_ptr(struct dmar_unit *unit)
401 {
402 	vm_page_t root_entry;
403 	int error;
404 
405 	/*
406 	 * Access to the GCMD register must be serialized while the
407 	 * command is submitted.
408 	 */
409 	DMAR_ASSERT_LOCKED(unit);
410 
411 	VM_OBJECT_RLOCK(unit->ctx_obj);
412 	root_entry = vm_page_lookup(unit->ctx_obj, 0);
413 	VM_OBJECT_RUNLOCK(unit->ctx_obj);
414 	dmar_write8(unit, DMAR_RTADDR_REG, VM_PAGE_TO_PHYS(root_entry));
415 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SRTP);
416 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_RTPS)
417 	    != 0));
418 	return (error);
419 }
420 
421 /*
422  * Globally invalidate the context entries cache, busily waiting for
423  * the completion.
424  */
425 int
426 dmar_inv_ctx_glob(struct dmar_unit *unit)
427 {
428 	int error;
429 
430 	/*
431 	 * Access to the CCMD register must be serialized while the
432 	 * command is submitted.
433 	 */
434 	DMAR_ASSERT_LOCKED(unit);
435 	KASSERT(!unit->qi_enabled, ("QI enabled"));
436 
437 	/*
438 	 * The DMAR_CCMD_ICC bit in the upper dword should be written
439 	 * after the low dword write is completed.  Amd64
440 	 * dmar_write8() does not have this issue, i386 dmar_write8()
441 	 * writes the upper dword last.
442 	 */
443 	dmar_write8(unit, DMAR_CCMD_REG, DMAR_CCMD_ICC | DMAR_CCMD_CIRG_GLOB);
444 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_CCMD_REG + 4) & DMAR_CCMD_ICC32)
445 	    == 0));
446 	return (error);
447 }
448 
449 /*
450  * Globally invalidate the IOTLB, busily waiting for the completion.
451  */
452 int
453 dmar_inv_iotlb_glob(struct dmar_unit *unit)
454 {
455 	int error, reg;
456 
457 	DMAR_ASSERT_LOCKED(unit);
458 	KASSERT(!unit->qi_enabled, ("QI enabled"));
459 
460 	reg = 16 * DMAR_ECAP_IRO(unit->hw_ecap);
461 	/* See a comment about DMAR_CCMD_ICC in dmar_inv_ctx_glob. */
462 	dmar_write8(unit, reg + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
463 	    DMAR_IOTLB_IIRG_GLB | DMAR_IOTLB_DR | DMAR_IOTLB_DW);
464 	DMAR_WAIT_UNTIL(((dmar_read4(unit, reg + DMAR_IOTLB_REG_OFF + 4) &
465 	    DMAR_IOTLB_IVT32) == 0));
466 	return (error);
467 }
468 
469 /*
470  * Flush the chipset write buffers.  See 11.1 "Write Buffer Flushing"
471  * in the architecture specification.
472  */
473 int
474 dmar_flush_write_bufs(struct dmar_unit *unit)
475 {
476 	int error;
477 
478 	DMAR_ASSERT_LOCKED(unit);
479 
480 	/*
481 	 * DMAR_GCMD_WBF is only valid when CAP_RWBF is reported.
482 	 */
483 	KASSERT((unit->hw_cap & DMAR_CAP_RWBF) != 0,
484 	    ("dmar%d: no RWBF", unit->unit));
485 
486 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_WBF);
487 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_WBFS)
488 	    != 0));
489 	return (error);
490 }
491 
492 int
493 dmar_enable_translation(struct dmar_unit *unit)
494 {
495 	int error;
496 
497 	DMAR_ASSERT_LOCKED(unit);
498 	unit->hw_gcmd |= DMAR_GCMD_TE;
499 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
500 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
501 	    != 0));
502 	return (error);
503 }
504 
505 int
506 dmar_disable_translation(struct dmar_unit *unit)
507 {
508 	int error;
509 
510 	DMAR_ASSERT_LOCKED(unit);
511 	unit->hw_gcmd &= ~DMAR_GCMD_TE;
512 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
513 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
514 	    == 0));
515 	return (error);
516 }
517 
518 int
519 dmar_load_irt_ptr(struct dmar_unit *unit)
520 {
521 	uint64_t irta, s;
522 	int error;
523 
524 	DMAR_ASSERT_LOCKED(unit);
525 	irta = unit->irt_phys;
526 	if (DMAR_X2APIC(unit))
527 		irta |= DMAR_IRTA_EIME;
528 	s = fls(unit->irte_cnt) - 2;
529 	KASSERT(unit->irte_cnt >= 2 && s <= DMAR_IRTA_S_MASK &&
530 	    powerof2(unit->irte_cnt),
531 	    ("IRTA_REG_S overflow %x", unit->irte_cnt));
532 	irta |= s;
533 	dmar_write8(unit, DMAR_IRTA_REG, irta);
534 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SIRTP);
535 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRTPS)
536 	    != 0));
537 	return (error);
538 }
539 
540 int
541 dmar_enable_ir(struct dmar_unit *unit)
542 {
543 	int error;
544 
545 	DMAR_ASSERT_LOCKED(unit);
546 	unit->hw_gcmd |= DMAR_GCMD_IRE;
547 	unit->hw_gcmd &= ~DMAR_GCMD_CFI;
548 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
549 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
550 	    != 0));
551 	return (error);
552 }
553 
554 int
555 dmar_disable_ir(struct dmar_unit *unit)
556 {
557 	int error;
558 
559 	DMAR_ASSERT_LOCKED(unit);
560 	unit->hw_gcmd &= ~DMAR_GCMD_IRE;
561 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
562 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
563 	    == 0));
564 	return (error);
565 }
566 
567 #define BARRIER_F				\
568 	u_int f_done, f_inproc, f_wakeup;	\
569 						\
570 	f_done = 1 << (barrier_id * 3);		\
571 	f_inproc = 1 << (barrier_id * 3 + 1);	\
572 	f_wakeup = 1 << (barrier_id * 3 + 2)
573 
574 bool
575 dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id)
576 {
577 	BARRIER_F;
578 
579 	DMAR_LOCK(dmar);
580 	if ((dmar->barrier_flags & f_done) != 0) {
581 		DMAR_UNLOCK(dmar);
582 		return (false);
583 	}
584 
585 	if ((dmar->barrier_flags & f_inproc) != 0) {
586 		while ((dmar->barrier_flags & f_inproc) != 0) {
587 			dmar->barrier_flags |= f_wakeup;
588 			msleep(&dmar->barrier_flags, &dmar->lock, 0,
589 			    "dmarb", 0);
590 		}
591 		KASSERT((dmar->barrier_flags & f_done) != 0,
592 		    ("dmar%d barrier %d missing done", dmar->unit, barrier_id));
593 		DMAR_UNLOCK(dmar);
594 		return (false);
595 	}
596 
597 	dmar->barrier_flags |= f_inproc;
598 	DMAR_UNLOCK(dmar);
599 	return (true);
600 }
601 
602 void
603 dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id)
604 {
605 	BARRIER_F;
606 
607 	DMAR_ASSERT_LOCKED(dmar);
608 	KASSERT((dmar->barrier_flags & (f_done | f_inproc)) == f_inproc,
609 	    ("dmar%d barrier %d missed entry", dmar->unit, barrier_id));
610 	dmar->barrier_flags |= f_done;
611 	if ((dmar->barrier_flags & f_wakeup) != 0)
612 		wakeup(&dmar->barrier_flags);
613 	dmar->barrier_flags &= ~(f_inproc | f_wakeup);
614 	DMAR_UNLOCK(dmar);
615 }
616 
617 int dmar_match_verbose;
618 int dmar_batch_coalesce = 100;
619 struct timespec dmar_hw_timeout = {
620 	.tv_sec = 0,
621 	.tv_nsec = 1000000
622 };
623 
624 static const uint64_t d = 1000000000;
625 
626 void
627 dmar_update_timeout(uint64_t newval)
628 {
629 
630 	/* XXXKIB not atomic */
631 	dmar_hw_timeout.tv_sec = newval / d;
632 	dmar_hw_timeout.tv_nsec = newval % d;
633 }
634 
635 uint64_t
636 dmar_get_timeout(void)
637 {
638 
639 	return ((uint64_t)dmar_hw_timeout.tv_sec * d +
640 	    dmar_hw_timeout.tv_nsec);
641 }
642 
643 static int
644 dmar_timeout_sysctl(SYSCTL_HANDLER_ARGS)
645 {
646 	uint64_t val;
647 	int error;
648 
649 	val = dmar_get_timeout();
650 	error = sysctl_handle_long(oidp, &val, 0, req);
651 	if (error != 0 || req->newptr == NULL)
652 		return (error);
653 	dmar_update_timeout(val);
654 	return (error);
655 }
656 
657 static SYSCTL_NODE(_hw, OID_AUTO, dmar, CTLFLAG_RD, NULL, "");
658 SYSCTL_INT(_hw_dmar, OID_AUTO, tbl_pagecnt, CTLFLAG_RD,
659     &dmar_tbl_pagecnt, 0,
660     "Count of pages used for DMAR pagetables");
661 SYSCTL_INT(_hw_dmar, OID_AUTO, match_verbose, CTLFLAG_RWTUN,
662     &dmar_match_verbose, 0,
663     "Verbose matching of the PCI devices to DMAR paths");
664 SYSCTL_INT(_hw_dmar, OID_AUTO, batch_coalesce, CTLFLAG_RWTUN,
665     &dmar_batch_coalesce, 0,
666     "Number of qi batches between interrupt");
667 SYSCTL_PROC(_hw_dmar, OID_AUTO, timeout,
668     CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
669     dmar_timeout_sysctl, "QU",
670     "Timeout for command wait, in nanoseconds");
671 #ifdef INVARIANTS
672 int dmar_check_free;
673 SYSCTL_INT(_hw_dmar, OID_AUTO, check_free, CTLFLAG_RWTUN,
674     &dmar_check_free, 0,
675     "Check the GPA RBtree for free_down and free_after validity");
676 #endif
677 
678