xref: /freebsd/sys/x86/iommu/intel_utils.c (revision 73577bf01de5c4677dc54d97f93310286c254780)
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 <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_kern.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_map.h>
60 #include <vm/vm_pageout.h>
61 #include <dev/pci/pcireg.h>
62 #include <dev/pci/pcivar.h>
63 #include <machine/bus.h>
64 #include <machine/cpu.h>
65 #include <machine/intr_machdep.h>
66 #include <x86/include/apicvar.h>
67 #include <x86/include/busdma_impl.h>
68 #include <dev/iommu/busdma_iommu.h>
69 #include <x86/iommu/intel_reg.h>
70 #include <x86/iommu/intel_dmar.h>
71 
72 u_int
73 dmar_nd2mask(u_int nd)
74 {
75 	static const u_int masks[] = {
76 		0x000f,	/* nd == 0 */
77 		0x002f,	/* nd == 1 */
78 		0x00ff,	/* nd == 2 */
79 		0x02ff,	/* nd == 3 */
80 		0x0fff,	/* nd == 4 */
81 		0x2fff,	/* nd == 5 */
82 		0xffff,	/* nd == 6 */
83 		0x0000,	/* nd == 7 reserved */
84 	};
85 
86 	KASSERT(nd <= 6, ("number of domains %d", nd));
87 	return (masks[nd]);
88 }
89 
90 static const struct sagaw_bits_tag {
91 	int agaw;
92 	int cap;
93 	int awlvl;
94 	int pglvl;
95 } sagaw_bits[] = {
96 	{.agaw = 30, .cap = DMAR_CAP_SAGAW_2LVL, .awlvl = DMAR_CTX2_AW_2LVL,
97 	    .pglvl = 2},
98 	{.agaw = 39, .cap = DMAR_CAP_SAGAW_3LVL, .awlvl = DMAR_CTX2_AW_3LVL,
99 	    .pglvl = 3},
100 	{.agaw = 48, .cap = DMAR_CAP_SAGAW_4LVL, .awlvl = DMAR_CTX2_AW_4LVL,
101 	    .pglvl = 4},
102 	{.agaw = 57, .cap = DMAR_CAP_SAGAW_5LVL, .awlvl = DMAR_CTX2_AW_5LVL,
103 	    .pglvl = 5},
104 	{.agaw = 64, .cap = DMAR_CAP_SAGAW_6LVL, .awlvl = DMAR_CTX2_AW_6LVL,
105 	    .pglvl = 6}
106 };
107 
108 bool
109 dmar_pglvl_supported(struct dmar_unit *unit, int pglvl)
110 {
111 	int i;
112 
113 	for (i = 0; i < nitems(sagaw_bits); i++) {
114 		if (sagaw_bits[i].pglvl != pglvl)
115 			continue;
116 		if ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
117 			return (true);
118 	}
119 	return (false);
120 }
121 
122 int
123 domain_set_agaw(struct dmar_domain *domain, int mgaw)
124 {
125 	int sagaw, i;
126 
127 	domain->mgaw = mgaw;
128 	sagaw = DMAR_CAP_SAGAW(domain->dmar->hw_cap);
129 	for (i = 0; i < nitems(sagaw_bits); i++) {
130 		if (sagaw_bits[i].agaw >= mgaw) {
131 			domain->agaw = sagaw_bits[i].agaw;
132 			domain->pglvl = sagaw_bits[i].pglvl;
133 			domain->awlvl = sagaw_bits[i].awlvl;
134 			return (0);
135 		}
136 	}
137 	device_printf(domain->dmar->dev,
138 	    "context request mgaw %d: no agaw found, sagaw %x\n",
139 	    mgaw, sagaw);
140 	return (EINVAL);
141 }
142 
143 /*
144  * Find a best fit mgaw for the given maxaddr:
145  *   - if allow_less is false, must find sagaw which maps all requested
146  *     addresses (used by identity mappings);
147  *   - if allow_less is true, and no supported sagaw can map all requested
148  *     address space, accept the biggest sagaw, whatever is it.
149  */
150 int
151 dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr, bool allow_less)
152 {
153 	int i;
154 
155 	for (i = 0; i < nitems(sagaw_bits); i++) {
156 		if ((1ULL << sagaw_bits[i].agaw) >= maxaddr &&
157 		    (DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
158 			break;
159 	}
160 	if (allow_less && i == nitems(sagaw_bits)) {
161 		do {
162 			i--;
163 		} while ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap)
164 		    == 0);
165 	}
166 	if (i < nitems(sagaw_bits))
167 		return (sagaw_bits[i].agaw);
168 	KASSERT(0, ("no mgaw for maxaddr %jx allow_less %d",
169 	    (uintmax_t) maxaddr, allow_less));
170 	return (-1);
171 }
172 
173 /*
174  * Calculate the total amount of page table pages needed to map the
175  * whole bus address space on the context with the selected agaw.
176  */
177 vm_pindex_t
178 pglvl_max_pages(int pglvl)
179 {
180 	vm_pindex_t res;
181 	int i;
182 
183 	for (res = 0, i = pglvl; i > 0; i--) {
184 		res *= DMAR_NPTEPG;
185 		res++;
186 	}
187 	return (res);
188 }
189 
190 /*
191  * Return true if the page table level lvl supports the superpage for
192  * the context ctx.
193  */
194 int
195 domain_is_sp_lvl(struct dmar_domain *domain, int lvl)
196 {
197 	int alvl, cap_sps;
198 	static const int sagaw_sp[] = {
199 		DMAR_CAP_SPS_2M,
200 		DMAR_CAP_SPS_1G,
201 		DMAR_CAP_SPS_512G,
202 		DMAR_CAP_SPS_1T
203 	};
204 
205 	alvl = domain->pglvl - lvl - 1;
206 	cap_sps = DMAR_CAP_SPS(domain->dmar->hw_cap);
207 	return (alvl < nitems(sagaw_sp) && (sagaw_sp[alvl] & cap_sps) != 0);
208 }
209 
210 iommu_gaddr_t
211 pglvl_page_size(int total_pglvl, int lvl)
212 {
213 	int rlvl;
214 	static const iommu_gaddr_t pg_sz[] = {
215 		(iommu_gaddr_t)DMAR_PAGE_SIZE,
216 		(iommu_gaddr_t)DMAR_PAGE_SIZE << DMAR_NPTEPGSHIFT,
217 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (2 * DMAR_NPTEPGSHIFT),
218 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (3 * DMAR_NPTEPGSHIFT),
219 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (4 * DMAR_NPTEPGSHIFT),
220 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (5 * DMAR_NPTEPGSHIFT)
221 	};
222 
223 	KASSERT(lvl >= 0 && lvl < total_pglvl,
224 	    ("total %d lvl %d", total_pglvl, lvl));
225 	rlvl = total_pglvl - lvl - 1;
226 	KASSERT(rlvl < nitems(pg_sz), ("sizeof pg_sz lvl %d", lvl));
227 	return (pg_sz[rlvl]);
228 }
229 
230 iommu_gaddr_t
231 domain_page_size(struct dmar_domain *domain, int lvl)
232 {
233 
234 	return (pglvl_page_size(domain->pglvl, lvl));
235 }
236 
237 int
238 calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size,
239     iommu_gaddr_t *isizep)
240 {
241 	iommu_gaddr_t isize;
242 	int am;
243 
244 	for (am = DMAR_CAP_MAMV(unit->hw_cap);; am--) {
245 		isize = 1ULL << (am + DMAR_PAGE_SHIFT);
246 		if ((base & (isize - 1)) == 0 && size >= isize)
247 			break;
248 		if (am == 0)
249 			break;
250 	}
251 	*isizep = isize;
252 	return (am);
253 }
254 
255 iommu_haddr_t dmar_high;
256 int haw;
257 int dmar_tbl_pagecnt;
258 
259 vm_page_t
260 dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags)
261 {
262 	vm_page_t m;
263 	int zeroed, aflags;
264 
265 	zeroed = (flags & IOMMU_PGF_ZERO) != 0 ? VM_ALLOC_ZERO : 0;
266 	aflags = zeroed | VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_NODUMP |
267 	    ((flags & IOMMU_PGF_WAITOK) != 0 ? VM_ALLOC_WAITFAIL :
268 	    VM_ALLOC_NOWAIT);
269 	for (;;) {
270 		if ((flags & IOMMU_PGF_OBJL) == 0)
271 			VM_OBJECT_WLOCK(obj);
272 		m = vm_page_lookup(obj, idx);
273 		if ((flags & IOMMU_PGF_NOALLOC) != 0 || m != NULL) {
274 			if ((flags & IOMMU_PGF_OBJL) == 0)
275 				VM_OBJECT_WUNLOCK(obj);
276 			break;
277 		}
278 		m = vm_page_alloc_contig(obj, idx, aflags, 1, 0,
279 		    dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
280 		if ((flags & IOMMU_PGF_OBJL) == 0)
281 			VM_OBJECT_WUNLOCK(obj);
282 		if (m != NULL) {
283 			if (zeroed && (m->flags & PG_ZERO) == 0)
284 				pmap_zero_page(m);
285 			atomic_add_int(&dmar_tbl_pagecnt, 1);
286 			break;
287 		}
288 		if ((flags & IOMMU_PGF_WAITOK) == 0)
289 			break;
290 	}
291 	return (m);
292 }
293 
294 void
295 dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags)
296 {
297 	vm_page_t m;
298 
299 	if ((flags & IOMMU_PGF_OBJL) == 0)
300 		VM_OBJECT_WLOCK(obj);
301 	m = vm_page_grab(obj, idx, VM_ALLOC_NOCREAT);
302 	if (m != NULL) {
303 		vm_page_free(m);
304 		atomic_subtract_int(&dmar_tbl_pagecnt, 1);
305 	}
306 	if ((flags & IOMMU_PGF_OBJL) == 0)
307 		VM_OBJECT_WUNLOCK(obj);
308 }
309 
310 void *
311 dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags,
312     struct sf_buf **sf)
313 {
314 	vm_page_t m;
315 	bool allocated;
316 
317 	if ((flags & IOMMU_PGF_OBJL) == 0)
318 		VM_OBJECT_WLOCK(obj);
319 	m = vm_page_lookup(obj, idx);
320 	if (m == NULL && (flags & IOMMU_PGF_ALLOC) != 0) {
321 		m = dmar_pgalloc(obj, idx, flags | IOMMU_PGF_OBJL);
322 		allocated = true;
323 	} else
324 		allocated = false;
325 	if (m == NULL) {
326 		if ((flags & IOMMU_PGF_OBJL) == 0)
327 			VM_OBJECT_WUNLOCK(obj);
328 		return (NULL);
329 	}
330 	/* Sleepable allocations cannot fail. */
331 	if ((flags & IOMMU_PGF_WAITOK) != 0)
332 		VM_OBJECT_WUNLOCK(obj);
333 	sched_pin();
334 	*sf = sf_buf_alloc(m, SFB_CPUPRIVATE | ((flags & IOMMU_PGF_WAITOK)
335 	    == 0 ? SFB_NOWAIT : 0));
336 	if (*sf == NULL) {
337 		sched_unpin();
338 		if (allocated) {
339 			VM_OBJECT_ASSERT_WLOCKED(obj);
340 			dmar_pgfree(obj, m->pindex, flags | IOMMU_PGF_OBJL);
341 		}
342 		if ((flags & IOMMU_PGF_OBJL) == 0)
343 			VM_OBJECT_WUNLOCK(obj);
344 		return (NULL);
345 	}
346 	if ((flags & (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL)) ==
347 	    (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL))
348 		VM_OBJECT_WLOCK(obj);
349 	else if ((flags & (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL)) == 0)
350 		VM_OBJECT_WUNLOCK(obj);
351 	return ((void *)sf_buf_kva(*sf));
352 }
353 
354 void
355 dmar_unmap_pgtbl(struct sf_buf *sf)
356 {
357 
358 	sf_buf_free(sf);
359 	sched_unpin();
360 }
361 
362 static void
363 dmar_flush_transl_to_ram(struct dmar_unit *unit, void *dst, size_t sz)
364 {
365 
366 	if (DMAR_IS_COHERENT(unit))
367 		return;
368 	/*
369 	 * If DMAR does not snoop paging structures accesses, flush
370 	 * CPU cache to memory.
371 	 */
372 	pmap_force_invalidate_cache_range((uintptr_t)dst, (uintptr_t)dst + sz);
373 }
374 
375 void
376 dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst)
377 {
378 
379 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
380 }
381 
382 void
383 dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst)
384 {
385 
386 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
387 }
388 
389 void
390 dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst)
391 {
392 
393 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
394 }
395 
396 /*
397  * Load the root entry pointer into the hardware, busily waiting for
398  * the completion.
399  */
400 int
401 dmar_load_root_entry_ptr(struct dmar_unit *unit)
402 {
403 	vm_page_t root_entry;
404 	int error;
405 
406 	/*
407 	 * Access to the GCMD register must be serialized while the
408 	 * command is submitted.
409 	 */
410 	DMAR_ASSERT_LOCKED(unit);
411 
412 	VM_OBJECT_RLOCK(unit->ctx_obj);
413 	root_entry = vm_page_lookup(unit->ctx_obj, 0);
414 	VM_OBJECT_RUNLOCK(unit->ctx_obj);
415 	dmar_write8(unit, DMAR_RTADDR_REG, VM_PAGE_TO_PHYS(root_entry));
416 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SRTP);
417 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_RTPS)
418 	    != 0));
419 	return (error);
420 }
421 
422 /*
423  * Globally invalidate the context entries cache, busily waiting for
424  * the completion.
425  */
426 int
427 dmar_inv_ctx_glob(struct dmar_unit *unit)
428 {
429 	int error;
430 
431 	/*
432 	 * Access to the CCMD register must be serialized while the
433 	 * command is submitted.
434 	 */
435 	DMAR_ASSERT_LOCKED(unit);
436 	KASSERT(!unit->qi_enabled, ("QI enabled"));
437 
438 	/*
439 	 * The DMAR_CCMD_ICC bit in the upper dword should be written
440 	 * after the low dword write is completed.  Amd64
441 	 * dmar_write8() does not have this issue, i386 dmar_write8()
442 	 * writes the upper dword last.
443 	 */
444 	dmar_write8(unit, DMAR_CCMD_REG, DMAR_CCMD_ICC | DMAR_CCMD_CIRG_GLOB);
445 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_CCMD_REG + 4) & DMAR_CCMD_ICC32)
446 	    == 0));
447 	return (error);
448 }
449 
450 /*
451  * Globally invalidate the IOTLB, busily waiting for the completion.
452  */
453 int
454 dmar_inv_iotlb_glob(struct dmar_unit *unit)
455 {
456 	int error, reg;
457 
458 	DMAR_ASSERT_LOCKED(unit);
459 	KASSERT(!unit->qi_enabled, ("QI enabled"));
460 
461 	reg = 16 * DMAR_ECAP_IRO(unit->hw_ecap);
462 	/* See a comment about DMAR_CCMD_ICC in dmar_inv_ctx_glob. */
463 	dmar_write8(unit, reg + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
464 	    DMAR_IOTLB_IIRG_GLB | DMAR_IOTLB_DR | DMAR_IOTLB_DW);
465 	DMAR_WAIT_UNTIL(((dmar_read4(unit, reg + DMAR_IOTLB_REG_OFF + 4) &
466 	    DMAR_IOTLB_IVT32) == 0));
467 	return (error);
468 }
469 
470 /*
471  * Flush the chipset write buffers.  See 11.1 "Write Buffer Flushing"
472  * in the architecture specification.
473  */
474 int
475 dmar_flush_write_bufs(struct dmar_unit *unit)
476 {
477 	int error;
478 
479 	DMAR_ASSERT_LOCKED(unit);
480 
481 	/*
482 	 * DMAR_GCMD_WBF is only valid when CAP_RWBF is reported.
483 	 */
484 	KASSERT((unit->hw_cap & DMAR_CAP_RWBF) != 0,
485 	    ("dmar%d: no RWBF", unit->iommu.unit));
486 
487 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_WBF);
488 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_WBFS)
489 	    != 0));
490 	return (error);
491 }
492 
493 int
494 dmar_enable_translation(struct dmar_unit *unit)
495 {
496 	int error;
497 
498 	DMAR_ASSERT_LOCKED(unit);
499 	unit->hw_gcmd |= DMAR_GCMD_TE;
500 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
501 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
502 	    != 0));
503 	return (error);
504 }
505 
506 int
507 dmar_disable_translation(struct dmar_unit *unit)
508 {
509 	int error;
510 
511 	DMAR_ASSERT_LOCKED(unit);
512 	unit->hw_gcmd &= ~DMAR_GCMD_TE;
513 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
514 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
515 	    == 0));
516 	return (error);
517 }
518 
519 int
520 dmar_load_irt_ptr(struct dmar_unit *unit)
521 {
522 	uint64_t irta, s;
523 	int error;
524 
525 	DMAR_ASSERT_LOCKED(unit);
526 	irta = unit->irt_phys;
527 	if (DMAR_X2APIC(unit))
528 		irta |= DMAR_IRTA_EIME;
529 	s = fls(unit->irte_cnt) - 2;
530 	KASSERT(unit->irte_cnt >= 2 && s <= DMAR_IRTA_S_MASK &&
531 	    powerof2(unit->irte_cnt),
532 	    ("IRTA_REG_S overflow %x", unit->irte_cnt));
533 	irta |= s;
534 	dmar_write8(unit, DMAR_IRTA_REG, irta);
535 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SIRTP);
536 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRTPS)
537 	    != 0));
538 	return (error);
539 }
540 
541 int
542 dmar_enable_ir(struct dmar_unit *unit)
543 {
544 	int error;
545 
546 	DMAR_ASSERT_LOCKED(unit);
547 	unit->hw_gcmd |= DMAR_GCMD_IRE;
548 	unit->hw_gcmd &= ~DMAR_GCMD_CFI;
549 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
550 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
551 	    != 0));
552 	return (error);
553 }
554 
555 int
556 dmar_disable_ir(struct dmar_unit *unit)
557 {
558 	int error;
559 
560 	DMAR_ASSERT_LOCKED(unit);
561 	unit->hw_gcmd &= ~DMAR_GCMD_IRE;
562 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
563 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
564 	    == 0));
565 	return (error);
566 }
567 
568 #define BARRIER_F				\
569 	u_int f_done, f_inproc, f_wakeup;	\
570 						\
571 	f_done = 1 << (barrier_id * 3);		\
572 	f_inproc = 1 << (barrier_id * 3 + 1);	\
573 	f_wakeup = 1 << (barrier_id * 3 + 2)
574 
575 bool
576 dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id)
577 {
578 	BARRIER_F;
579 
580 	DMAR_LOCK(dmar);
581 	if ((dmar->barrier_flags & f_done) != 0) {
582 		DMAR_UNLOCK(dmar);
583 		return (false);
584 	}
585 
586 	if ((dmar->barrier_flags & f_inproc) != 0) {
587 		while ((dmar->barrier_flags & f_inproc) != 0) {
588 			dmar->barrier_flags |= f_wakeup;
589 			msleep(&dmar->barrier_flags, &dmar->iommu.lock, 0,
590 			    "dmarb", 0);
591 		}
592 		KASSERT((dmar->barrier_flags & f_done) != 0,
593 		    ("dmar%d barrier %d missing done", dmar->iommu.unit,
594 		    barrier_id));
595 		DMAR_UNLOCK(dmar);
596 		return (false);
597 	}
598 
599 	dmar->barrier_flags |= f_inproc;
600 	DMAR_UNLOCK(dmar);
601 	return (true);
602 }
603 
604 void
605 dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id)
606 {
607 	BARRIER_F;
608 
609 	DMAR_ASSERT_LOCKED(dmar);
610 	KASSERT((dmar->barrier_flags & (f_done | f_inproc)) == f_inproc,
611 	    ("dmar%d barrier %d missed entry", dmar->iommu.unit, barrier_id));
612 	dmar->barrier_flags |= f_done;
613 	if ((dmar->barrier_flags & f_wakeup) != 0)
614 		wakeup(&dmar->barrier_flags);
615 	dmar->barrier_flags &= ~(f_inproc | f_wakeup);
616 	DMAR_UNLOCK(dmar);
617 }
618 
619 int dmar_batch_coalesce = 100;
620 struct timespec dmar_hw_timeout = {
621 	.tv_sec = 0,
622 	.tv_nsec = 1000000
623 };
624 
625 static const uint64_t d = 1000000000;
626 
627 void
628 dmar_update_timeout(uint64_t newval)
629 {
630 
631 	/* XXXKIB not atomic */
632 	dmar_hw_timeout.tv_sec = newval / d;
633 	dmar_hw_timeout.tv_nsec = newval % d;
634 }
635 
636 uint64_t
637 dmar_get_timeout(void)
638 {
639 
640 	return ((uint64_t)dmar_hw_timeout.tv_sec * d +
641 	    dmar_hw_timeout.tv_nsec);
642 }
643 
644 static int
645 dmar_timeout_sysctl(SYSCTL_HANDLER_ARGS)
646 {
647 	uint64_t val;
648 	int error;
649 
650 	val = dmar_get_timeout();
651 	error = sysctl_handle_long(oidp, &val, 0, req);
652 	if (error != 0 || req->newptr == NULL)
653 		return (error);
654 	dmar_update_timeout(val);
655 	return (error);
656 }
657 
658 static SYSCTL_NODE(_hw_iommu, OID_AUTO, dmar, CTLFLAG_RD | CTLFLAG_MPSAFE,
659     NULL, "");
660 SYSCTL_INT(_hw_iommu_dmar, OID_AUTO, tbl_pagecnt, CTLFLAG_RD,
661     &dmar_tbl_pagecnt, 0,
662     "Count of pages used for DMAR pagetables");
663 SYSCTL_INT(_hw_iommu_dmar, OID_AUTO, batch_coalesce, CTLFLAG_RWTUN,
664     &dmar_batch_coalesce, 0,
665     "Number of qi batches between interrupt");
666 SYSCTL_PROC(_hw_iommu_dmar, OID_AUTO, timeout,
667     CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
668     dmar_timeout_sysctl, "QU",
669     "Timeout for command wait, in nanoseconds");
670