xref: /freebsd-13-stable/sys/x86/iommu/intel_utils.c (revision 3bc80996974a61a4223eae4c1ccd47b6ee32a48a)
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/cdefs.h>
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/kernel.h>
35 #include <sys/lock.h>
36 #include <sys/malloc.h>
37 #include <sys/memdesc.h>
38 #include <sys/mutex.h>
39 #include <sys/proc.h>
40 #include <sys/queue.h>
41 #include <sys/rman.h>
42 #include <sys/rwlock.h>
43 #include <sys/sched.h>
44 #include <sys/sf_buf.h>
45 #include <sys/sysctl.h>
46 #include <sys/systm.h>
47 #include <sys/taskqueue.h>
48 #include <sys/time.h>
49 #include <sys/tree.h>
50 #include <sys/vmem.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_map.h>
57 #include <vm/vm_pageout.h>
58 #include <dev/pci/pcireg.h>
59 #include <dev/pci/pcivar.h>
60 #include <machine/bus.h>
61 #include <machine/cpu.h>
62 #include <machine/intr_machdep.h>
63 #include <x86/include/apicvar.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/intel_dmar.h>
68 
69 u_int
dmar_nd2mask(u_int nd)70 dmar_nd2mask(u_int nd)
71 {
72 	static const u_int masks[] = {
73 		0x000f,	/* nd == 0 */
74 		0x002f,	/* nd == 1 */
75 		0x00ff,	/* nd == 2 */
76 		0x02ff,	/* nd == 3 */
77 		0x0fff,	/* nd == 4 */
78 		0x2fff,	/* nd == 5 */
79 		0xffff,	/* nd == 6 */
80 		0x0000,	/* nd == 7 reserved */
81 	};
82 
83 	KASSERT(nd <= 6, ("number of domains %d", nd));
84 	return (masks[nd]);
85 }
86 
87 static const struct sagaw_bits_tag {
88 	int agaw;
89 	int cap;
90 	int awlvl;
91 	int pglvl;
92 } sagaw_bits[] = {
93 	{.agaw = 30, .cap = DMAR_CAP_SAGAW_2LVL, .awlvl = DMAR_CTX2_AW_2LVL,
94 	    .pglvl = 2},
95 	{.agaw = 39, .cap = DMAR_CAP_SAGAW_3LVL, .awlvl = DMAR_CTX2_AW_3LVL,
96 	    .pglvl = 3},
97 	{.agaw = 48, .cap = DMAR_CAP_SAGAW_4LVL, .awlvl = DMAR_CTX2_AW_4LVL,
98 	    .pglvl = 4},
99 	{.agaw = 57, .cap = DMAR_CAP_SAGAW_5LVL, .awlvl = DMAR_CTX2_AW_5LVL,
100 	    .pglvl = 5}
101 	/*
102 	 * 6-level paging (DMAR_CAP_SAGAW_6LVL) is not supported on any
103 	 * current VT-d hardware and its SAGAW field value is listed as
104 	 * reserved in the VT-d spec.  If support is added in the future,
105 	 * this structure and the logic in dmar_maxaddr2mgaw() will need
106 	 * to change to avoid attempted comparison against 1ULL << 64.
107 	 */
108 };
109 
110 bool
dmar_pglvl_supported(struct dmar_unit * unit,int pglvl)111 dmar_pglvl_supported(struct dmar_unit *unit, int pglvl)
112 {
113 	int i;
114 
115 	for (i = 0; i < nitems(sagaw_bits); i++) {
116 		if (sagaw_bits[i].pglvl != pglvl)
117 			continue;
118 		if ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
119 			return (true);
120 	}
121 	return (false);
122 }
123 
124 int
domain_set_agaw(struct dmar_domain * domain,int mgaw)125 domain_set_agaw(struct dmar_domain *domain, int mgaw)
126 {
127 	int sagaw, i;
128 
129 	domain->mgaw = mgaw;
130 	sagaw = DMAR_CAP_SAGAW(domain->dmar->hw_cap);
131 	for (i = 0; i < nitems(sagaw_bits); i++) {
132 		if (sagaw_bits[i].agaw >= mgaw) {
133 			domain->agaw = sagaw_bits[i].agaw;
134 			domain->pglvl = sagaw_bits[i].pglvl;
135 			domain->awlvl = sagaw_bits[i].awlvl;
136 			return (0);
137 		}
138 	}
139 	device_printf(domain->dmar->dev,
140 	    "context request mgaw %d: no agaw found, sagaw %x\n",
141 	    mgaw, sagaw);
142 	return (EINVAL);
143 }
144 
145 /*
146  * Find a best fit mgaw for the given maxaddr:
147  *   - if allow_less is false, must find sagaw which maps all requested
148  *     addresses (used by identity mappings);
149  *   - if allow_less is true, and no supported sagaw can map all requested
150  *     address space, accept the biggest sagaw, whatever is it.
151  */
152 int
dmar_maxaddr2mgaw(struct dmar_unit * unit,iommu_gaddr_t maxaddr,bool allow_less)153 dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr, bool allow_less)
154 {
155 	int i;
156 
157 	for (i = 0; i < nitems(sagaw_bits); i++) {
158 		if ((1ULL << sagaw_bits[i].agaw) >= maxaddr &&
159 		    (DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap) != 0)
160 			break;
161 	}
162 	if (allow_less && i == nitems(sagaw_bits)) {
163 		do {
164 			i--;
165 		} while ((DMAR_CAP_SAGAW(unit->hw_cap) & sagaw_bits[i].cap)
166 		    == 0);
167 	}
168 	if (i < nitems(sagaw_bits))
169 		return (sagaw_bits[i].agaw);
170 	KASSERT(0, ("no mgaw for maxaddr %jx allow_less %d",
171 	    (uintmax_t) maxaddr, allow_less));
172 	return (-1);
173 }
174 
175 /*
176  * Calculate the total amount of page table pages needed to map the
177  * whole bus address space on the context with the selected agaw.
178  */
179 vm_pindex_t
pglvl_max_pages(int pglvl)180 pglvl_max_pages(int pglvl)
181 {
182 	vm_pindex_t res;
183 	int i;
184 
185 	for (res = 0, i = pglvl; i > 0; i--) {
186 		res *= DMAR_NPTEPG;
187 		res++;
188 	}
189 	return (res);
190 }
191 
192 /*
193  * Return true if the page table level lvl supports the superpage for
194  * the context ctx.
195  */
196 int
domain_is_sp_lvl(struct dmar_domain * domain,int lvl)197 domain_is_sp_lvl(struct dmar_domain *domain, int lvl)
198 {
199 	int alvl, cap_sps;
200 	static const int sagaw_sp[] = {
201 		DMAR_CAP_SPS_2M,
202 		DMAR_CAP_SPS_1G,
203 		DMAR_CAP_SPS_512G,
204 		DMAR_CAP_SPS_1T
205 	};
206 
207 	alvl = domain->pglvl - lvl - 1;
208 	cap_sps = DMAR_CAP_SPS(domain->dmar->hw_cap);
209 	return (alvl < nitems(sagaw_sp) && (sagaw_sp[alvl] & cap_sps) != 0);
210 }
211 
212 iommu_gaddr_t
pglvl_page_size(int total_pglvl,int lvl)213 pglvl_page_size(int total_pglvl, int lvl)
214 {
215 	int rlvl;
216 	static const iommu_gaddr_t pg_sz[] = {
217 		(iommu_gaddr_t)DMAR_PAGE_SIZE,
218 		(iommu_gaddr_t)DMAR_PAGE_SIZE << DMAR_NPTEPGSHIFT,
219 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (2 * DMAR_NPTEPGSHIFT),
220 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (3 * DMAR_NPTEPGSHIFT),
221 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (4 * DMAR_NPTEPGSHIFT),
222 		(iommu_gaddr_t)DMAR_PAGE_SIZE << (5 * DMAR_NPTEPGSHIFT)
223 	};
224 
225 	KASSERT(lvl >= 0 && lvl < total_pglvl,
226 	    ("total %d lvl %d", total_pglvl, lvl));
227 	rlvl = total_pglvl - lvl - 1;
228 	KASSERT(rlvl < nitems(pg_sz), ("sizeof pg_sz lvl %d", lvl));
229 	return (pg_sz[rlvl]);
230 }
231 
232 iommu_gaddr_t
domain_page_size(struct dmar_domain * domain,int lvl)233 domain_page_size(struct dmar_domain *domain, int lvl)
234 {
235 
236 	return (pglvl_page_size(domain->pglvl, lvl));
237 }
238 
239 int
calc_am(struct dmar_unit * unit,iommu_gaddr_t base,iommu_gaddr_t size,iommu_gaddr_t * isizep)240 calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size,
241     iommu_gaddr_t *isizep)
242 {
243 	iommu_gaddr_t isize;
244 	int am;
245 
246 	for (am = DMAR_CAP_MAMV(unit->hw_cap);; am--) {
247 		isize = 1ULL << (am + DMAR_PAGE_SHIFT);
248 		if ((base & (isize - 1)) == 0 && size >= isize)
249 			break;
250 		if (am == 0)
251 			break;
252 	}
253 	*isizep = isize;
254 	return (am);
255 }
256 
257 iommu_haddr_t dmar_high;
258 int haw;
259 int dmar_tbl_pagecnt;
260 
261 vm_page_t
dmar_pgalloc(vm_object_t obj,vm_pindex_t idx,int flags)262 dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags)
263 {
264 	vm_page_t m;
265 	int zeroed, aflags;
266 
267 	zeroed = (flags & IOMMU_PGF_ZERO) != 0 ? VM_ALLOC_ZERO : 0;
268 	aflags = zeroed | VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_NODUMP |
269 	    ((flags & IOMMU_PGF_WAITOK) != 0 ? VM_ALLOC_WAITFAIL :
270 	    VM_ALLOC_NOWAIT);
271 	for (;;) {
272 		if ((flags & IOMMU_PGF_OBJL) == 0)
273 			VM_OBJECT_WLOCK(obj);
274 		m = vm_page_lookup(obj, idx);
275 		if ((flags & IOMMU_PGF_NOALLOC) != 0 || m != NULL) {
276 			if ((flags & IOMMU_PGF_OBJL) == 0)
277 				VM_OBJECT_WUNLOCK(obj);
278 			break;
279 		}
280 		m = vm_page_alloc_contig(obj, idx, aflags, 1, 0,
281 		    dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
282 		if ((flags & IOMMU_PGF_OBJL) == 0)
283 			VM_OBJECT_WUNLOCK(obj);
284 		if (m != NULL) {
285 			if (zeroed && (m->flags & PG_ZERO) == 0)
286 				pmap_zero_page(m);
287 			atomic_add_int(&dmar_tbl_pagecnt, 1);
288 			break;
289 		}
290 		if ((flags & IOMMU_PGF_WAITOK) == 0)
291 			break;
292 	}
293 	return (m);
294 }
295 
296 void
dmar_pgfree(vm_object_t obj,vm_pindex_t idx,int flags)297 dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags)
298 {
299 	vm_page_t m;
300 
301 	if ((flags & IOMMU_PGF_OBJL) == 0)
302 		VM_OBJECT_WLOCK(obj);
303 	m = vm_page_grab(obj, idx, VM_ALLOC_NOCREAT);
304 	if (m != NULL) {
305 		vm_page_free(m);
306 		atomic_subtract_int(&dmar_tbl_pagecnt, 1);
307 	}
308 	if ((flags & IOMMU_PGF_OBJL) == 0)
309 		VM_OBJECT_WUNLOCK(obj);
310 }
311 
312 void *
dmar_map_pgtbl(vm_object_t obj,vm_pindex_t idx,int flags,struct sf_buf ** sf)313 dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags,
314     struct sf_buf **sf)
315 {
316 	vm_page_t m;
317 	bool allocated;
318 
319 	if ((flags & IOMMU_PGF_OBJL) == 0)
320 		VM_OBJECT_WLOCK(obj);
321 	m = vm_page_lookup(obj, idx);
322 	if (m == NULL && (flags & IOMMU_PGF_ALLOC) != 0) {
323 		m = dmar_pgalloc(obj, idx, flags | IOMMU_PGF_OBJL);
324 		allocated = true;
325 	} else
326 		allocated = false;
327 	if (m == NULL) {
328 		if ((flags & IOMMU_PGF_OBJL) == 0)
329 			VM_OBJECT_WUNLOCK(obj);
330 		return (NULL);
331 	}
332 	/* Sleepable allocations cannot fail. */
333 	if ((flags & IOMMU_PGF_WAITOK) != 0)
334 		VM_OBJECT_WUNLOCK(obj);
335 	sched_pin();
336 	*sf = sf_buf_alloc(m, SFB_CPUPRIVATE | ((flags & IOMMU_PGF_WAITOK)
337 	    == 0 ? SFB_NOWAIT : 0));
338 	if (*sf == NULL) {
339 		sched_unpin();
340 		if (allocated) {
341 			VM_OBJECT_ASSERT_WLOCKED(obj);
342 			dmar_pgfree(obj, m->pindex, flags | IOMMU_PGF_OBJL);
343 		}
344 		if ((flags & IOMMU_PGF_OBJL) == 0)
345 			VM_OBJECT_WUNLOCK(obj);
346 		return (NULL);
347 	}
348 	if ((flags & (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL)) ==
349 	    (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL))
350 		VM_OBJECT_WLOCK(obj);
351 	else if ((flags & (IOMMU_PGF_WAITOK | IOMMU_PGF_OBJL)) == 0)
352 		VM_OBJECT_WUNLOCK(obj);
353 	return ((void *)sf_buf_kva(*sf));
354 }
355 
356 void
dmar_unmap_pgtbl(struct sf_buf * sf)357 dmar_unmap_pgtbl(struct sf_buf *sf)
358 {
359 
360 	sf_buf_free(sf);
361 	sched_unpin();
362 }
363 
364 static void
dmar_flush_transl_to_ram(struct dmar_unit * unit,void * dst,size_t sz)365 dmar_flush_transl_to_ram(struct dmar_unit *unit, void *dst, size_t sz)
366 {
367 
368 	if (DMAR_IS_COHERENT(unit))
369 		return;
370 	/*
371 	 * If DMAR does not snoop paging structures accesses, flush
372 	 * CPU cache to memory.
373 	 */
374 	pmap_force_invalidate_cache_range((uintptr_t)dst, (uintptr_t)dst + sz);
375 }
376 
377 void
dmar_flush_pte_to_ram(struct dmar_unit * unit,dmar_pte_t * dst)378 dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst)
379 {
380 
381 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
382 }
383 
384 void
dmar_flush_ctx_to_ram(struct dmar_unit * unit,dmar_ctx_entry_t * dst)385 dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst)
386 {
387 
388 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
389 }
390 
391 void
dmar_flush_root_to_ram(struct dmar_unit * unit,dmar_root_entry_t * dst)392 dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst)
393 {
394 
395 	dmar_flush_transl_to_ram(unit, dst, sizeof(*dst));
396 }
397 
398 /*
399  * Load the root entry pointer into the hardware, busily waiting for
400  * the completion.
401  */
402 int
dmar_load_root_entry_ptr(struct dmar_unit * unit)403 dmar_load_root_entry_ptr(struct dmar_unit *unit)
404 {
405 	vm_page_t root_entry;
406 	int error;
407 
408 	/*
409 	 * Access to the GCMD register must be serialized while the
410 	 * command is submitted.
411 	 */
412 	DMAR_ASSERT_LOCKED(unit);
413 
414 	VM_OBJECT_RLOCK(unit->ctx_obj);
415 	root_entry = vm_page_lookup(unit->ctx_obj, 0);
416 	VM_OBJECT_RUNLOCK(unit->ctx_obj);
417 	dmar_write8(unit, DMAR_RTADDR_REG, VM_PAGE_TO_PHYS(root_entry));
418 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SRTP);
419 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_RTPS)
420 	    != 0));
421 	return (error);
422 }
423 
424 /*
425  * Globally invalidate the context entries cache, busily waiting for
426  * the completion.
427  */
428 int
dmar_inv_ctx_glob(struct dmar_unit * unit)429 dmar_inv_ctx_glob(struct dmar_unit *unit)
430 {
431 	int error;
432 
433 	/*
434 	 * Access to the CCMD register must be serialized while the
435 	 * command is submitted.
436 	 */
437 	DMAR_ASSERT_LOCKED(unit);
438 	KASSERT(!unit->qi_enabled, ("QI enabled"));
439 
440 	/*
441 	 * The DMAR_CCMD_ICC bit in the upper dword should be written
442 	 * after the low dword write is completed.  Amd64
443 	 * dmar_write8() does not have this issue, i386 dmar_write8()
444 	 * writes the upper dword last.
445 	 */
446 	dmar_write8(unit, DMAR_CCMD_REG, DMAR_CCMD_ICC | DMAR_CCMD_CIRG_GLOB);
447 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_CCMD_REG + 4) & DMAR_CCMD_ICC32)
448 	    == 0));
449 	return (error);
450 }
451 
452 /*
453  * Globally invalidate the IOTLB, busily waiting for the completion.
454  */
455 int
dmar_inv_iotlb_glob(struct dmar_unit * unit)456 dmar_inv_iotlb_glob(struct dmar_unit *unit)
457 {
458 	int error, reg;
459 
460 	DMAR_ASSERT_LOCKED(unit);
461 	KASSERT(!unit->qi_enabled, ("QI enabled"));
462 
463 	reg = 16 * DMAR_ECAP_IRO(unit->hw_ecap);
464 	/* See a comment about DMAR_CCMD_ICC in dmar_inv_ctx_glob. */
465 	dmar_write8(unit, reg + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
466 	    DMAR_IOTLB_IIRG_GLB | DMAR_IOTLB_DR | DMAR_IOTLB_DW);
467 	DMAR_WAIT_UNTIL(((dmar_read4(unit, reg + DMAR_IOTLB_REG_OFF + 4) &
468 	    DMAR_IOTLB_IVT32) == 0));
469 	return (error);
470 }
471 
472 /*
473  * Flush the chipset write buffers.  See 11.1 "Write Buffer Flushing"
474  * in the architecture specification.
475  */
476 int
dmar_flush_write_bufs(struct dmar_unit * unit)477 dmar_flush_write_bufs(struct dmar_unit *unit)
478 {
479 	int error;
480 
481 	DMAR_ASSERT_LOCKED(unit);
482 
483 	/*
484 	 * DMAR_GCMD_WBF is only valid when CAP_RWBF is reported.
485 	 */
486 	KASSERT((unit->hw_cap & DMAR_CAP_RWBF) != 0,
487 	    ("dmar%d: no RWBF", unit->iommu.unit));
488 
489 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_WBF);
490 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_WBFS)
491 	    != 0));
492 	return (error);
493 }
494 
495 int
dmar_enable_translation(struct dmar_unit * unit)496 dmar_enable_translation(struct dmar_unit *unit)
497 {
498 	int error;
499 
500 	DMAR_ASSERT_LOCKED(unit);
501 	unit->hw_gcmd |= DMAR_GCMD_TE;
502 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
503 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
504 	    != 0));
505 	return (error);
506 }
507 
508 int
dmar_disable_translation(struct dmar_unit * unit)509 dmar_disable_translation(struct dmar_unit *unit)
510 {
511 	int error;
512 
513 	DMAR_ASSERT_LOCKED(unit);
514 	unit->hw_gcmd &= ~DMAR_GCMD_TE;
515 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
516 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_TES)
517 	    == 0));
518 	return (error);
519 }
520 
521 int
dmar_load_irt_ptr(struct dmar_unit * unit)522 dmar_load_irt_ptr(struct dmar_unit *unit)
523 {
524 	uint64_t irta, s;
525 	int error;
526 
527 	DMAR_ASSERT_LOCKED(unit);
528 	irta = unit->irt_phys;
529 	if (DMAR_X2APIC(unit))
530 		irta |= DMAR_IRTA_EIME;
531 	s = fls(unit->irte_cnt) - 2;
532 	KASSERT(unit->irte_cnt >= 2 && s <= DMAR_IRTA_S_MASK &&
533 	    powerof2(unit->irte_cnt),
534 	    ("IRTA_REG_S overflow %x", unit->irte_cnt));
535 	irta |= s;
536 	dmar_write8(unit, DMAR_IRTA_REG, irta);
537 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd | DMAR_GCMD_SIRTP);
538 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRTPS)
539 	    != 0));
540 	return (error);
541 }
542 
543 int
dmar_enable_ir(struct dmar_unit * unit)544 dmar_enable_ir(struct dmar_unit *unit)
545 {
546 	int error;
547 
548 	DMAR_ASSERT_LOCKED(unit);
549 	unit->hw_gcmd |= DMAR_GCMD_IRE;
550 	unit->hw_gcmd &= ~DMAR_GCMD_CFI;
551 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
552 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
553 	    != 0));
554 	return (error);
555 }
556 
557 int
dmar_disable_ir(struct dmar_unit * unit)558 dmar_disable_ir(struct dmar_unit *unit)
559 {
560 	int error;
561 
562 	DMAR_ASSERT_LOCKED(unit);
563 	unit->hw_gcmd &= ~DMAR_GCMD_IRE;
564 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
565 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_IRES)
566 	    == 0));
567 	return (error);
568 }
569 
570 #define BARRIER_F				\
571 	u_int f_done, f_inproc, f_wakeup;	\
572 						\
573 	f_done = 1 << (barrier_id * 3);		\
574 	f_inproc = 1 << (barrier_id * 3 + 1);	\
575 	f_wakeup = 1 << (barrier_id * 3 + 2)
576 
577 bool
dmar_barrier_enter(struct dmar_unit * dmar,u_int barrier_id)578 dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id)
579 {
580 	BARRIER_F;
581 
582 	DMAR_LOCK(dmar);
583 	if ((dmar->barrier_flags & f_done) != 0) {
584 		DMAR_UNLOCK(dmar);
585 		return (false);
586 	}
587 
588 	if ((dmar->barrier_flags & f_inproc) != 0) {
589 		while ((dmar->barrier_flags & f_inproc) != 0) {
590 			dmar->barrier_flags |= f_wakeup;
591 			msleep(&dmar->barrier_flags, &dmar->iommu.lock, 0,
592 			    "dmarb", 0);
593 		}
594 		KASSERT((dmar->barrier_flags & f_done) != 0,
595 		    ("dmar%d barrier %d missing done", dmar->iommu.unit,
596 		    barrier_id));
597 		DMAR_UNLOCK(dmar);
598 		return (false);
599 	}
600 
601 	dmar->barrier_flags |= f_inproc;
602 	DMAR_UNLOCK(dmar);
603 	return (true);
604 }
605 
606 void
dmar_barrier_exit(struct dmar_unit * dmar,u_int barrier_id)607 dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id)
608 {
609 	BARRIER_F;
610 
611 	DMAR_ASSERT_LOCKED(dmar);
612 	KASSERT((dmar->barrier_flags & (f_done | f_inproc)) == f_inproc,
613 	    ("dmar%d barrier %d missed entry", dmar->iommu.unit, barrier_id));
614 	dmar->barrier_flags |= f_done;
615 	if ((dmar->barrier_flags & f_wakeup) != 0)
616 		wakeup(&dmar->barrier_flags);
617 	dmar->barrier_flags &= ~(f_inproc | f_wakeup);
618 	DMAR_UNLOCK(dmar);
619 }
620 
621 int dmar_batch_coalesce = 100;
622 struct timespec dmar_hw_timeout = {
623 	.tv_sec = 0,
624 	.tv_nsec = 1000000
625 };
626 
627 static const uint64_t d = 1000000000;
628 
629 void
dmar_update_timeout(uint64_t newval)630 dmar_update_timeout(uint64_t newval)
631 {
632 
633 	/* XXXKIB not atomic */
634 	dmar_hw_timeout.tv_sec = newval / d;
635 	dmar_hw_timeout.tv_nsec = newval % d;
636 }
637 
638 uint64_t
dmar_get_timeout(void)639 dmar_get_timeout(void)
640 {
641 
642 	return ((uint64_t)dmar_hw_timeout.tv_sec * d +
643 	    dmar_hw_timeout.tv_nsec);
644 }
645 
646 static int
dmar_timeout_sysctl(SYSCTL_HANDLER_ARGS)647 dmar_timeout_sysctl(SYSCTL_HANDLER_ARGS)
648 {
649 	uint64_t val;
650 	int error;
651 
652 	val = dmar_get_timeout();
653 	error = sysctl_handle_long(oidp, &val, 0, req);
654 	if (error != 0 || req->newptr == NULL)
655 		return (error);
656 	dmar_update_timeout(val);
657 	return (error);
658 }
659 
660 static SYSCTL_NODE(_hw_iommu, OID_AUTO, dmar, CTLFLAG_RD | CTLFLAG_MPSAFE,
661     NULL, "");
662 SYSCTL_INT(_hw_iommu_dmar, OID_AUTO, tbl_pagecnt, CTLFLAG_RD,
663     &dmar_tbl_pagecnt, 0,
664     "Count of pages used for DMAR pagetables");
665 SYSCTL_INT(_hw_iommu_dmar, OID_AUTO, batch_coalesce, CTLFLAG_RWTUN,
666     &dmar_batch_coalesce, 0,
667     "Number of qi batches between interrupt");
668 SYSCTL_PROC(_hw_iommu_dmar, OID_AUTO, timeout,
669     CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
670     dmar_timeout_sysctl, "QU",
671     "Timeout for command wait, in nanoseconds");
672