xref: /titanic_41/usr/src/uts/i86pc/io/immu_dvma.c (revision 3fb517f786391b507780c78aabb8d98bfea9efe9)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Portions Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /*
26  * Copyright (c) 2009, Intel Corporation.
27  * All rights reserved.
28  */
29 
30 /*
31  * DVMA code
32  * This file contains Intel IOMMU code that deals with DVMA
33  * i.e. DMA remapping.
34  */
35 
36 #include <sys/sysmacros.h>
37 #include <sys/pcie.h>
38 #include <sys/pci_cfgspace.h>
39 #include <vm/hat_i86.h>
40 #include <sys/memlist.h>
41 #include <sys/acpi/acpi.h>
42 #include <sys/acpica.h>
43 #include <sys/modhash.h>
44 #include <sys/immu.h>
45 
46 #undef	TEST
47 
48 /*
49  * Macros based on PCI spec
50  */
51 #define	IMMU_PCI_REV2CLASS(r)   ((r) >> 8)  /* classcode from revid */
52 #define	IMMU_PCI_CLASS2BASE(c)  ((c) >> 16) /* baseclass from classcode */
53 #define	IMMU_PCI_CLASS2SUB(c)   (((c) >> 8) & 0xff); /* classcode */
54 
55 #define	IMMU_CONTIG_PADDR(d, p) \
56 	((d).dck_paddr && ((d).dck_paddr + IMMU_PAGESIZE) == (p))
57 
58 typedef struct dvma_arg {
59 	immu_t *dva_immu;
60 	dev_info_t *dva_rdip;
61 	dev_info_t *dva_ddip;
62 	domain_t *dva_domain;
63 	int dva_level;
64 	immu_flags_t dva_flags;
65 	list_t *dva_list;
66 	int dva_error;
67 } dvma_arg_t;
68 
69 static domain_t *domain_create(immu_t *immu, dev_info_t *ddip,
70     dev_info_t *rdip, immu_flags_t immu_flags);
71 static immu_devi_t *create_immu_devi(dev_info_t *rdip, int bus,
72     int dev, int func, immu_flags_t immu_flags);
73 static void destroy_immu_devi(immu_devi_t *immu_devi);
74 static boolean_t dvma_map(immu_t *immu, domain_t *domain, uint64_t sdvma,
75     uint64_t nvpages, dcookie_t *dcookies, int dcount, dev_info_t *rdip,
76     immu_flags_t immu_flags);
77 
78 /* Extern globals */
79 extern struct memlist  *phys_install;
80 
81 
82 /* static Globals */
83 
84 /*
85  * Used to setup DMA objects (memory regions)
86  * for DMA reads by IOMMU units
87  */
88 static ddi_dma_attr_t immu_dma_attr = {
89 	DMA_ATTR_V0,
90 	0U,
91 	0xffffffffU,
92 	0xffffffffU,
93 	MMU_PAGESIZE, /* MMU page aligned */
94 	0x1,
95 	0x1,
96 	0xffffffffU,
97 	0xffffffffU,
98 	1,
99 	4,
100 	0
101 };
102 
103 static ddi_device_acc_attr_t immu_acc_attr = {
104 	DDI_DEVICE_ATTR_V0,
105 	DDI_NEVERSWAP_ACC,
106 	DDI_STRICTORDER_ACC
107 };
108 
109 
110 /* globals private to this file */
111 static kmutex_t immu_domain_lock;
112 static list_t immu_unity_domain_list;
113 static list_t immu_xlate_domain_list;
114 
115 /* structure used to store idx into each level of the page tables */
116 typedef struct xlate {
117 	int xlt_level;
118 	uint_t xlt_idx;
119 	pgtable_t *xlt_pgtable;
120 } xlate_t;
121 
122 /* 0 is reserved by Vt-d spec. Solaris reserves 1 */
123 #define	IMMU_UNITY_DID   1
124 
125 static mod_hash_t *bdf_domain_hash;
126 
127 static domain_t *
128 bdf_domain_lookup(immu_devi_t *immu_devi)
129 {
130 	domain_t *domain;
131 	int16_t seg = immu_devi->imd_seg;
132 	int16_t bus = immu_devi->imd_bus;
133 	int16_t devfunc = immu_devi->imd_devfunc;
134 	uintptr_t bdf = (seg << 16 | bus << 8 | devfunc);
135 
136 	if (seg < 0 || bus < 0 || devfunc < 0) {
137 		return (NULL);
138 	}
139 
140 	domain = NULL;
141 	if (mod_hash_find(bdf_domain_hash,
142 	    (void *)bdf, (void *)&domain) == 0) {
143 		ASSERT(domain);
144 		ASSERT(domain->dom_did > 0);
145 		return (domain);
146 	} else {
147 		return (NULL);
148 	}
149 }
150 
151 static void
152 bdf_domain_insert(immu_devi_t *immu_devi, domain_t *domain)
153 {
154 	int16_t seg = immu_devi->imd_seg;
155 	int16_t bus = immu_devi->imd_bus;
156 	int16_t devfunc = immu_devi->imd_devfunc;
157 	uintptr_t bdf = (seg << 16 | bus << 8 | devfunc);
158 	int r;
159 
160 	if (seg < 0 || bus < 0 || devfunc < 0) {
161 		return;
162 	}
163 
164 	r = mod_hash_insert(bdf_domain_hash, (void *)bdf, (void *)domain);
165 	ASSERT(r != MH_ERR_DUPLICATE);
166 	ASSERT(r == 0);
167 }
168 
169 static int
170 match_lpc(dev_info_t *pdip, void *arg)
171 {
172 	immu_devi_t *immu_devi;
173 	dvma_arg_t *dvap = (dvma_arg_t *)arg;
174 
175 	ASSERT(dvap->dva_error == DDI_FAILURE);
176 	ASSERT(dvap->dva_ddip == NULL);
177 	ASSERT(dvap->dva_list);
178 
179 	if (list_is_empty(dvap->dva_list)) {
180 		return (DDI_WALK_TERMINATE);
181 	}
182 
183 	immu_devi = list_head(dvap->dva_list);
184 	for (; immu_devi; immu_devi = list_next(dvap->dva_list,
185 	    immu_devi)) {
186 		ASSERT(immu_devi->imd_dip);
187 		if (immu_devi->imd_dip == pdip) {
188 			dvap->dva_ddip = pdip;
189 			dvap->dva_error = DDI_SUCCESS;
190 			return (DDI_WALK_TERMINATE);
191 		}
192 	}
193 
194 	return (DDI_WALK_CONTINUE);
195 }
196 
197 static void
198 immu_devi_set_spclist(dev_info_t *dip, immu_t *immu)
199 {
200 	list_t *spclist = NULL;
201 	immu_devi_t *immu_devi;
202 
203 	ASSERT(MUTEX_HELD(&(DEVI(dip)->devi_lock)));
204 
205 	immu_devi = IMMU_DEVI(dip);
206 	if (immu_devi->imd_display == B_TRUE) {
207 		spclist = &(immu->immu_dvma_gfx_list);
208 	} else if (immu_devi->imd_lpc == B_TRUE) {
209 		spclist = &(immu->immu_dvma_lpc_list);
210 	}
211 
212 	if (spclist) {
213 		mutex_enter(&(immu->immu_lock));
214 		list_insert_head(spclist, immu_devi);
215 		mutex_exit(&(immu->immu_lock));
216 	}
217 }
218 
219 /*
220  * Set the immu_devi struct in the immu_devi field of a devinfo node
221  */
222 int
223 immu_devi_set(dev_info_t *dip, immu_flags_t immu_flags)
224 {
225 	int bus, dev, func;
226 	immu_devi_t *new_imd;
227 	immu_devi_t *immu_devi;
228 
229 	ASSERT(root_devinfo);
230 	ASSERT(dip);
231 	ASSERT(dip != root_devinfo);
232 
233 	immu_devi = immu_devi_get(dip);
234 	if (immu_devi != NULL) {
235 		return (DDI_SUCCESS);
236 	}
237 
238 	bus = dev = func = -1;
239 
240 	/*
241 	 * Assume a new immu_devi struct is needed
242 	 */
243 	if (!DEVI_IS_PCI(dip) || acpica_get_bdf(dip, &bus, &dev, &func) != 0) {
244 		/*
245 		 * No BDF. Set bus = -1 to indicate this.
246 		 * We still need to create a immu_devi struct
247 		 * though
248 		 */
249 		bus = -1;
250 		dev = 0;
251 		func = 0;
252 	}
253 
254 	new_imd = create_immu_devi(dip, bus, dev, func, immu_flags);
255 	if (new_imd  == NULL) {
256 		ddi_err(DER_WARN, dip, "Failed to create immu_devi "
257 		    "structure");
258 		return (DDI_FAILURE);
259 	}
260 
261 	/*
262 	 * Check if some other thread allocated a immu_devi while we
263 	 * didn't own the lock.
264 	 */
265 	mutex_enter(&(DEVI(dip)->devi_lock));
266 	if (IMMU_DEVI(dip) == NULL) {
267 		IMMU_DEVI_SET(dip, new_imd);
268 	} else {
269 		destroy_immu_devi(new_imd);
270 	}
271 	mutex_exit(&(DEVI(dip)->devi_lock));
272 
273 	return (DDI_SUCCESS);
274 }
275 
276 static dev_info_t *
277 get_lpc_devinfo(immu_t *immu, dev_info_t *rdip, immu_flags_t immu_flags)
278 {
279 	dvma_arg_t dvarg = {0};
280 	dvarg.dva_list = &(immu->immu_dvma_lpc_list);
281 	dvarg.dva_rdip = rdip;
282 	dvarg.dva_error = DDI_FAILURE;
283 
284 	if (immu_walk_ancestor(rdip, NULL, match_lpc,
285 	    &dvarg, NULL, immu_flags) != DDI_SUCCESS) {
286 		ddi_err(DER_MODE, rdip, "Could not walk ancestors to "
287 		    "find lpc_devinfo for ISA device");
288 		return (NULL);
289 	}
290 
291 	if (dvarg.dva_error != DDI_SUCCESS || dvarg.dva_ddip == NULL) {
292 		ddi_err(DER_MODE, rdip, "Could not find lpc_devinfo for "
293 		    "ISA device");
294 		return (NULL);
295 	}
296 
297 	return (dvarg.dva_ddip);
298 }
299 
300 static dev_info_t *
301 get_gfx_devinfo(dev_info_t *rdip)
302 {
303 	immu_t *immu;
304 	immu_devi_t *immu_devi;
305 	list_t *list_gfx;
306 
307 	/*
308 	 * The GFX device may not be on the same IMMU unit as "agpgart"
309 	 * so search globally
310 	 */
311 	immu_devi = NULL;
312 	immu = list_head(&immu_list);
313 	for (; immu; immu = list_next(&immu_list, immu)) {
314 		list_gfx = &(immu->immu_dvma_gfx_list);
315 		if (!list_is_empty(list_gfx)) {
316 			immu_devi = list_head(list_gfx);
317 			break;
318 		}
319 	}
320 
321 	if (immu_devi == NULL) {
322 		ddi_err(DER_WARN, rdip, "IMMU: No GFX device. "
323 		    "Cannot redirect agpgart");
324 		return (NULL);
325 	}
326 
327 	/* list is not empty we checked above */
328 	ASSERT(immu_devi);
329 	ASSERT(immu_devi->imd_dip);
330 
331 	ddi_err(DER_LOG, rdip, "IMMU: GFX redirect to %s",
332 	    ddi_node_name(immu_devi->imd_dip));
333 
334 	return (immu_devi->imd_dip);
335 }
336 
337 static immu_flags_t
338 dma_to_immu_flags(struct ddi_dma_req *dmareq)
339 {
340 	immu_flags_t flags = 0;
341 
342 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
343 		flags |= IMMU_FLAGS_SLEEP;
344 	} else {
345 		flags |= IMMU_FLAGS_NOSLEEP;
346 	}
347 
348 #ifdef BUGGY_DRIVERS
349 
350 	flags |= (IMMU_FLAGS_READ | IMMU_FLAGS_WRITE);
351 
352 #else
353 	/*
354 	 * Read and write flags need to be reversed.
355 	 * DMA_READ means read from device and write
356 	 * to memory. So DMA read means DVMA write.
357 	 */
358 	if (dmareq->dmar_flags & DDI_DMA_READ)
359 		flags |= IMMU_FLAGS_WRITE;
360 
361 	if (dmareq->dmar_flags & DDI_DMA_WRITE)
362 		flags |= IMMU_FLAGS_READ;
363 
364 	/*
365 	 * Some buggy drivers specify neither READ or WRITE
366 	 * For such drivers set both read and write permissions
367 	 */
368 	if ((dmareq->dmar_flags & (DDI_DMA_READ | DDI_DMA_WRITE)) == 0) {
369 		flags |= (IMMU_FLAGS_READ | IMMU_FLAGS_WRITE);
370 	}
371 #endif
372 
373 	return (flags);
374 }
375 
376 int
377 pgtable_ctor(void *buf, void *arg, int kmflag)
378 {
379 	size_t actual_size = 0;
380 	pgtable_t *pgtable;
381 	int (*dmafp)(caddr_t);
382 	caddr_t vaddr;
383 	void *next;
384 
385 	ASSERT(buf);
386 	ASSERT(arg == NULL);
387 
388 	pgtable = (pgtable_t *)buf;
389 
390 	dmafp = (kmflag & KM_NOSLEEP) ? DDI_DMA_DONTWAIT : DDI_DMA_SLEEP;
391 
392 	next = kmem_zalloc(IMMU_PAGESIZE, kmflag);
393 	if (next == NULL) {
394 		return (-1);
395 	}
396 
397 	ASSERT(root_devinfo);
398 	if (ddi_dma_alloc_handle(root_devinfo, &immu_dma_attr,
399 	    dmafp, NULL, &pgtable->hwpg_dmahdl) != DDI_SUCCESS) {
400 		kmem_free(next, IMMU_PAGESIZE);
401 		return (-1);
402 	}
403 
404 	if (ddi_dma_mem_alloc(pgtable->hwpg_dmahdl, IMMU_PAGESIZE,
405 	    &immu_acc_attr, DDI_DMA_CONSISTENT | IOMEM_DATA_UNCACHED,
406 	    dmafp, NULL, &vaddr, &actual_size,
407 	    &pgtable->hwpg_memhdl) != DDI_SUCCESS) {
408 		ddi_dma_free_handle(&pgtable->hwpg_dmahdl);
409 		kmem_free(next, IMMU_PAGESIZE);
410 		return (-1);
411 	}
412 
413 	/*
414 	 * Memory allocation failure. Maybe a temporary condition
415 	 * so return error rather than panic, so we can try again
416 	 */
417 	if (actual_size < IMMU_PAGESIZE) {
418 		ddi_dma_mem_free(&pgtable->hwpg_memhdl);
419 		ddi_dma_free_handle(&pgtable->hwpg_dmahdl);
420 		kmem_free(next, IMMU_PAGESIZE);
421 		return (-1);
422 	}
423 
424 	pgtable->hwpg_paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, vaddr));
425 	pgtable->hwpg_vaddr = vaddr;
426 	pgtable->swpg_next_array = next;
427 
428 	rw_init(&(pgtable->swpg_rwlock), NULL, RW_DEFAULT, NULL);
429 
430 	return (0);
431 }
432 
433 void
434 pgtable_dtor(void *buf, void *arg)
435 {
436 	pgtable_t *pgtable;
437 
438 	ASSERT(buf);
439 	ASSERT(arg == NULL);
440 
441 	pgtable = (pgtable_t *)buf;
442 	ASSERT(pgtable->swpg_next_array);
443 
444 	/* destroy will panic if lock is held. */
445 	rw_destroy(&(pgtable->swpg_rwlock));
446 
447 	ddi_dma_mem_free(&pgtable->hwpg_memhdl);
448 	ddi_dma_free_handle(&pgtable->hwpg_dmahdl);
449 	kmem_free(pgtable->swpg_next_array, IMMU_PAGESIZE);
450 
451 	/* don't zero out hwpg_vaddr and swpg_next_array for debugging */
452 }
453 
454 /*
455  * pgtable_alloc()
456  *	alloc a IOMMU pgtable structure.
457  *	This same struct is used for root and context tables as well.
458  *	This routine allocs the f/ollowing:
459  *	- a pgtable_t struct
460  *	- a HW page which holds PTEs/entries which is accesssed by HW
461  *        so we set up DMA for this page
462  *	- a SW page which is only for our bookeeping
463  *        (for example to  hold pointers to the next level pgtable).
464  *        So a simple kmem_alloc suffices
465  */
466 static pgtable_t *
467 pgtable_alloc(immu_t *immu, immu_flags_t immu_flags)
468 {
469 	pgtable_t *pgtable;
470 	int kmflags;
471 
472 	ASSERT(immu);
473 
474 	kmflags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
475 
476 	pgtable = kmem_cache_alloc(immu_pgtable_cache, kmflags);
477 	if (pgtable == NULL) {
478 		return (NULL);
479 	}
480 	return (pgtable);
481 }
482 
483 static void
484 pgtable_zero(immu_t *immu, pgtable_t *pgtable)
485 {
486 	bzero(pgtable->hwpg_vaddr, IMMU_PAGESIZE);
487 	bzero(pgtable->swpg_next_array, IMMU_PAGESIZE);
488 
489 	/* Dont need to flush the write we will flush when we use the entry */
490 	immu_regs_cpu_flush(immu, pgtable->hwpg_vaddr, IMMU_PAGESIZE);
491 }
492 
493 static void
494 pgtable_free(immu_t *immu, pgtable_t *pgtable)
495 {
496 	ASSERT(immu);
497 	ASSERT(pgtable);
498 
499 	kmem_cache_free(immu_pgtable_cache, pgtable);
500 }
501 
502 /*
503  * Function to identify a display device from the PCI class code
504  */
505 static boolean_t
506 device_is_display(uint_t classcode)
507 {
508 	static uint_t disp_classes[] = {
509 		0x000100,
510 		0x030000,
511 		0x030001
512 	};
513 	int i, nclasses = sizeof (disp_classes) / sizeof (uint_t);
514 
515 	for (i = 0; i < nclasses; i++) {
516 		if (classcode == disp_classes[i])
517 			return (B_TRUE);
518 	}
519 	return (B_FALSE);
520 }
521 
522 /*
523  * Function that determines if device is PCIEX and/or PCIEX bridge
524  */
525 static boolean_t
526 device_is_pciex(
527 	uchar_t bus, uchar_t dev, uchar_t func, boolean_t *is_pcib)
528 {
529 	ushort_t cap;
530 	ushort_t capsp;
531 	ushort_t cap_count = PCI_CAP_MAX_PTR;
532 	ushort_t status;
533 	boolean_t is_pciex = B_FALSE;
534 
535 	*is_pcib = B_FALSE;
536 
537 	status = pci_getw_func(bus, dev, func, PCI_CONF_STAT);
538 	if (!(status & PCI_STAT_CAP))
539 		return (B_FALSE);
540 
541 	capsp = pci_getb_func(bus, dev, func, PCI_CONF_CAP_PTR);
542 	while (cap_count-- && capsp >= PCI_CAP_PTR_OFF) {
543 		capsp &= PCI_CAP_PTR_MASK;
544 		cap = pci_getb_func(bus, dev, func, capsp);
545 
546 		if (cap == PCI_CAP_ID_PCI_E) {
547 			status = pci_getw_func(bus, dev, func, capsp + 2);
548 			/*
549 			 * See section 7.8.2 of PCI-Express Base Spec v1.0a
550 			 * for Device/Port Type.
551 			 * PCIE_PCIECAP_DEV_TYPE_PCIE2PCI implies that the
552 			 * device is a PCIE2PCI bridge
553 			 */
554 			*is_pcib =
555 			    ((status & PCIE_PCIECAP_DEV_TYPE_MASK) ==
556 			    PCIE_PCIECAP_DEV_TYPE_PCIE2PCI) ? B_TRUE : B_FALSE;
557 			is_pciex = B_TRUE;
558 		}
559 
560 		capsp = (*pci_getb_func)(bus, dev, func,
561 		    capsp + PCI_CAP_NEXT_PTR);
562 	}
563 
564 	return (is_pciex);
565 }
566 
567 
568 /*
569  * immu_dvma_get_immu()
570  *   get the immu unit structure for a dev_info node
571  */
572 immu_t *
573 immu_dvma_get_immu(dev_info_t *dip, immu_flags_t immu_flags)
574 {
575 	immu_devi_t *immu_devi;
576 	immu_t *immu;
577 
578 	/*
579 	 * check if immu unit was already found earlier.
580 	 * If yes, then it will be stashed in immu_devi struct.
581 	 */
582 	immu_devi = immu_devi_get(dip);
583 	if (immu_devi == NULL) {
584 		if (immu_devi_set(dip, immu_flags) != DDI_SUCCESS) {
585 			/*
586 			 * May fail because of low memory. Return error rather
587 			 * than panic as we want driver to rey again later
588 			 */
589 			ddi_err(DER_PANIC, dip, "immu_dvma_get_immu: "
590 			    "No immu_devi structure");
591 			/*NOTREACHED*/
592 		}
593 		immu_devi = immu_devi_get(dip);
594 		ASSERT(immu_devi);
595 	}
596 
597 	mutex_enter(&(DEVI(dip)->devi_lock));
598 	if (immu_devi->imd_immu) {
599 		immu = immu_devi->imd_immu;
600 		mutex_exit(&(DEVI(dip)->devi_lock));
601 		return (immu);
602 	}
603 	mutex_exit(&(DEVI(dip)->devi_lock));
604 
605 	immu = immu_dmar_get_immu(dip);
606 	if (immu == NULL) {
607 		ddi_err(DER_PANIC, dip, "immu_dvma_get_immu: "
608 		    "Cannot find immu_t for device");
609 		/*NOTREACHED*/
610 	}
611 
612 	/*
613 	 * Check if some other thread found immu
614 	 * while lock was not held
615 	 */
616 	immu_devi = immu_devi_get(dip);
617 	/* immu_devi should be present as we found it earlier */
618 	if (immu_devi == NULL) {
619 		ddi_err(DER_PANIC, dip,
620 		    "immu_dvma_get_immu: No immu_devi structure");
621 		/*NOTREACHED*/
622 	}
623 
624 	mutex_enter(&(DEVI(dip)->devi_lock));
625 	if (immu_devi->imd_immu == NULL) {
626 		/* nobody else set it, so we should do it */
627 		immu_devi->imd_immu = immu;
628 		immu_devi_set_spclist(dip, immu);
629 	} else {
630 		/*
631 		 * if some other thread got immu before
632 		 * us, it should get the same results
633 		 */
634 		if (immu_devi->imd_immu != immu) {
635 			ddi_err(DER_PANIC, dip, "Multiple "
636 			    "immu units found for device. Expected (%p), "
637 			    "actual (%p)", (void *)immu,
638 			    (void *)immu_devi->imd_immu);
639 			mutex_exit(&(DEVI(dip)->devi_lock));
640 			/*NOTREACHED*/
641 		}
642 	}
643 	mutex_exit(&(DEVI(dip)->devi_lock));
644 
645 	return (immu);
646 }
647 
648 
649 /* ############################# IMMU_DEVI code ############################ */
650 
651 /*
652  * Allocate a immu_devi structure and initialize it
653  */
654 static immu_devi_t *
655 create_immu_devi(dev_info_t *rdip, int bus, int dev, int func,
656     immu_flags_t immu_flags)
657 {
658 	uchar_t baseclass, subclass;
659 	uint_t classcode, revclass;
660 	immu_devi_t *immu_devi;
661 	boolean_t pciex = B_FALSE;
662 	int kmflags;
663 	boolean_t is_pcib = B_FALSE;
664 
665 	/* bus ==  -1 indicate non-PCI device (no BDF) */
666 	ASSERT(bus == -1 || bus >= 0);
667 	ASSERT(dev >= 0);
668 	ASSERT(func >= 0);
669 
670 	kmflags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
671 	immu_devi = kmem_zalloc(sizeof (immu_devi_t), kmflags);
672 	if (immu_devi == NULL) {
673 		ddi_err(DER_WARN, rdip, "Failed to allocate memory for "
674 		    "Intel IOMMU immu_devi structure");
675 		return (NULL);
676 	}
677 	immu_devi->imd_dip = rdip;
678 	immu_devi->imd_seg = 0; /* Currently seg can only be 0 */
679 	immu_devi->imd_bus = bus;
680 	immu_devi->imd_pcib_type = IMMU_PCIB_BAD;
681 
682 	if (bus == -1) {
683 		immu_devi->imd_pcib_type = IMMU_PCIB_NOBDF;
684 		return (immu_devi);
685 	}
686 
687 	immu_devi->imd_devfunc = IMMU_PCI_DEVFUNC(dev, func);
688 	immu_devi->imd_sec = 0;
689 	immu_devi->imd_sub = 0;
690 
691 	revclass = pci_getl_func(bus, dev, func, PCI_CONF_REVID);
692 
693 	classcode = IMMU_PCI_REV2CLASS(revclass);
694 	baseclass = IMMU_PCI_CLASS2BASE(classcode);
695 	subclass = IMMU_PCI_CLASS2SUB(classcode);
696 
697 	if (baseclass == PCI_CLASS_BRIDGE && subclass == PCI_BRIDGE_PCI) {
698 
699 		immu_devi->imd_sec = pci_getb_func(bus, dev, func,
700 		    PCI_BCNF_SECBUS);
701 		immu_devi->imd_sub = pci_getb_func(bus, dev, func,
702 		    PCI_BCNF_SUBBUS);
703 
704 		pciex = device_is_pciex(bus, dev, func, &is_pcib);
705 		if (pciex  == B_TRUE && is_pcib == B_TRUE) {
706 			immu_devi->imd_pcib_type = IMMU_PCIB_PCIE_PCI;
707 		} else if (pciex == B_TRUE) {
708 			immu_devi->imd_pcib_type = IMMU_PCIB_PCIE_PCIE;
709 		} else {
710 			immu_devi->imd_pcib_type = IMMU_PCIB_PCI_PCI;
711 		}
712 	} else {
713 		immu_devi->imd_pcib_type = IMMU_PCIB_ENDPOINT;
714 	}
715 
716 	/* check for certain special devices */
717 	immu_devi->imd_display = device_is_display(classcode);
718 
719 	immu_devi->imd_lpc = ((baseclass == PCI_CLASS_BRIDGE) &&
720 	    (subclass == PCI_BRIDGE_ISA)) ? B_TRUE : B_FALSE;
721 
722 	immu_devi->imd_domain = NULL;
723 
724 	return (immu_devi);
725 }
726 
727 static void
728 destroy_immu_devi(immu_devi_t *immu_devi)
729 {
730 	kmem_free(immu_devi, sizeof (immu_devi_t));
731 }
732 
733 static domain_t *
734 immu_devi_domain(dev_info_t *rdip, dev_info_t **ddipp)
735 {
736 	immu_devi_t *immu_devi;
737 	domain_t *domain;
738 	dev_info_t *ddip;
739 
740 	ASSERT(rdip);
741 	ASSERT(ddipp);
742 
743 	*ddipp = NULL;
744 
745 	immu_devi = immu_devi_get(rdip);
746 	if (immu_devi == NULL) {
747 		return (NULL);
748 	}
749 
750 	mutex_enter(&(DEVI(rdip)->devi_lock));
751 	domain = immu_devi->imd_domain;
752 	ddip = immu_devi->imd_ddip;
753 	mutex_exit(&(DEVI(rdip)->devi_lock));
754 
755 	if (domain) {
756 		ASSERT(domain->dom_did > 0);
757 		ASSERT(ddip);
758 		*ddipp = ddip;
759 	}
760 
761 	return (domain);
762 
763 }
764 
765 /* ############################# END IMMU_DEVI code ######################## */
766 /* ############################# DOMAIN code ############################### */
767 
768 /*
769  * This routine always succeeds
770  */
771 static int
772 did_alloc(immu_t *immu, dev_info_t *rdip,
773     dev_info_t *ddip, immu_flags_t immu_flags)
774 {
775 	int did;
776 
777 	ASSERT(immu);
778 	ASSERT(rdip);
779 	ASSERT(rdip != root_devinfo);
780 
781 	did = (uintptr_t)vmem_alloc(immu->immu_did_arena, 1,
782 	    (immu_flags & IMMU_FLAGS_NOSLEEP) ? VM_NOSLEEP : VM_SLEEP);
783 
784 	if (did == 0) {
785 		ASSERT(immu->immu_unity_domain);
786 		ASSERT(immu->immu_unity_domain->dom_did > 0);
787 		ddi_err(DER_WARN, rdip, "device domain-id alloc error"
788 		    " domain-device: %s%d. immu unit is %s. Using "
789 		    "unity domain with domain-id (%d)",
790 		    ddi_driver_name(ddip), ddi_get_instance(ddip),
791 		    immu->immu_name, immu->immu_unity_domain->dom_did);
792 		did = immu->immu_unity_domain->dom_did;
793 	}
794 
795 	return (did);
796 }
797 
798 static int
799 get_branch_domain(dev_info_t *pdip, void *arg)
800 {
801 	immu_devi_t *immu_devi;
802 	domain_t *domain;
803 	dev_info_t *ddip;
804 	immu_t *immu;
805 	dvma_arg_t *dvp = (dvma_arg_t *)arg;
806 
807 	ASSERT(pdip);
808 	ASSERT(dvp);
809 	ASSERT(dvp->dva_rdip);
810 
811 	/*
812 	 * The field dvp->dva_rdip is a work-in-progress
813 	 * and gets updated as we walk up the ancestor
814 	 * tree. The final ddip is set only when we reach
815 	 * the top of the tree. So the dvp->dva_ddip field cannot
816 	 * be relied on until we reach the top of the field.
817 	 */
818 
819 	/* immu_devi may not be set. */
820 	immu_devi = immu_devi_get(pdip);
821 	if (immu_devi == NULL) {
822 		if (immu_devi_set(pdip, dvp->dva_flags) != DDI_SUCCESS) {
823 			dvp->dva_error = DDI_FAILURE;
824 			return (DDI_WALK_TERMINATE);
825 		}
826 	}
827 
828 	immu_devi = immu_devi_get(pdip);
829 	ASSERT(immu_devi);
830 	immu = immu_devi->imd_immu;
831 	if (immu == NULL) {
832 		immu = immu_dvma_get_immu(pdip, dvp->dva_flags);
833 		ASSERT(immu);
834 	}
835 
836 	/*
837 	 * If we encounter a PCIE_PCIE bridge *ANCESTOR* we need to
838 	 * terminate the walk (since the device under the PCIE bridge
839 	 * is a PCIE device and has an independent entry in the
840 	 * root/context table)
841 	 */
842 	if (dvp->dva_rdip != pdip &&
843 	    immu_devi->imd_pcib_type == IMMU_PCIB_PCIE_PCIE) {
844 		return (DDI_WALK_TERMINATE);
845 	}
846 
847 	/*
848 	 * In order to be a domain-dim, it must be a PCI device i.e.
849 	 * must have valid BDF. This also eliminates the root complex.
850 	 */
851 	if (immu_devi->imd_pcib_type != IMMU_PCIB_BAD &&
852 	    immu_devi->imd_pcib_type != IMMU_PCIB_NOBDF) {
853 		ASSERT(immu_devi->imd_bus >= 0);
854 		ASSERT(immu_devi->imd_devfunc >= 0);
855 		dvp->dva_ddip = pdip;
856 	}
857 
858 	if (immu_devi->imd_display == B_TRUE ||
859 	    (dvp->dva_flags & IMMU_FLAGS_UNITY)) {
860 		dvp->dva_domain = immu->immu_unity_domain;
861 		/* continue walking to find ddip */
862 		return (DDI_WALK_CONTINUE);
863 	}
864 
865 	mutex_enter(&(DEVI(pdip)->devi_lock));
866 	domain = immu_devi->imd_domain;
867 	ddip = immu_devi->imd_ddip;
868 	mutex_exit(&(DEVI(pdip)->devi_lock));
869 
870 	if (domain && ddip) {
871 		/* if domain is set, it must be the same */
872 		if (dvp->dva_domain) {
873 			ASSERT(domain == dvp->dva_domain);
874 		}
875 		dvp->dva_domain = domain;
876 		dvp->dva_ddip = ddip;
877 		return (DDI_WALK_TERMINATE);
878 	}
879 
880 	/* immu_devi either has both set or both clear */
881 	ASSERT(domain == NULL);
882 	ASSERT(ddip == NULL);
883 
884 	/* Domain may already be set, continue walking so that ddip gets set */
885 	if (dvp->dva_domain) {
886 		return (DDI_WALK_CONTINUE);
887 	}
888 
889 	/* domain is not set in either immu_devi or dvp */
890 	domain = bdf_domain_lookup(immu_devi);
891 	if (domain == NULL) {
892 		return (DDI_WALK_CONTINUE);
893 	}
894 
895 	/* ok, the BDF hash had a domain for this BDF. */
896 
897 	/* Grab lock again to check if something else set immu_devi fields */
898 	mutex_enter(&(DEVI(pdip)->devi_lock));
899 	if (immu_devi->imd_domain != NULL) {
900 		ASSERT(immu_devi->imd_domain == domain);
901 		dvp->dva_domain = domain;
902 	} else {
903 		dvp->dva_domain = domain;
904 	}
905 	mutex_exit(&(DEVI(pdip)->devi_lock));
906 
907 	/*
908 	 * walk upwards until the topmost PCI bridge is found
909 	 */
910 	return (DDI_WALK_CONTINUE);
911 
912 }
913 
914 static void
915 map_unity_domain(domain_t *domain)
916 {
917 	struct memlist *mp;
918 	uint64_t start;
919 	uint64_t npages;
920 	dcookie_t dcookies[1] = {0};
921 	int dcount = 0;
922 
923 	ASSERT(domain);
924 	ASSERT(domain->dom_did == IMMU_UNITY_DID);
925 
926 	/*
927 	 * We call into routines that grab the lock so we should
928 	 * not be called with the lock held. This does not matter
929 	 * much since, no else has a reference to this domain
930 	 */
931 	ASSERT(!rw_lock_held(&(domain->dom_pgtable_rwlock)));
932 
933 	/*
934 	 * UNITY arenas are a mirror of the physical memory
935 	 * installed on the system.
936 	 */
937 
938 #ifdef BUGGY_DRIVERS
939 	/*
940 	 * Dont skip page0. Some broken HW/FW access it.
941 	 */
942 	dcookies[0].dck_paddr = 0;
943 	dcookies[0].dck_npages = 1;
944 	dcount = 1;
945 	(void) dvma_map(domain->dom_immu, domain, 0, 1, dcookies, dcount, NULL,
946 	    IMMU_FLAGS_READ | IMMU_FLAGS_WRITE | IMMU_FLAGS_PAGE1);
947 #endif
948 
949 	memlist_read_lock();
950 
951 	mp = phys_install;
952 
953 	if (mp->ml_address == 0) {
954 		/* since we already mapped page1 above */
955 		start = IMMU_PAGESIZE;
956 	} else {
957 		start = mp->ml_address;
958 	}
959 	npages = mp->ml_size/IMMU_PAGESIZE + 1;
960 
961 	dcookies[0].dck_paddr = start;
962 	dcookies[0].dck_npages = npages;
963 	dcount = 1;
964 	(void) dvma_map(domain->dom_immu, domain, start, npages, dcookies,
965 	    dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE);
966 
967 	ddi_err(DER_LOG, NULL, "IMMU: mapping PHYS span [0x%" PRIx64
968 	    " - 0x%" PRIx64 "]", start, start + mp->ml_size);
969 
970 	mp = mp->ml_next;
971 	while (mp) {
972 		ddi_err(DER_LOG, NULL, "IMMU: mapping PHYS span [0x%" PRIx64
973 		    " - 0x%" PRIx64 "]", mp->ml_address,
974 		    mp->ml_address + mp->ml_size);
975 
976 		start = mp->ml_address;
977 		npages = mp->ml_size/IMMU_PAGESIZE + 1;
978 
979 		dcookies[0].dck_paddr = start;
980 		dcookies[0].dck_npages = npages;
981 		dcount = 1;
982 		(void) dvma_map(domain->dom_immu, domain, start, npages,
983 		    dcookies, dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE);
984 		mp = mp->ml_next;
985 	}
986 
987 	mp = bios_rsvd;
988 	while (mp) {
989 		ddi_err(DER_LOG, NULL, "IMMU: mapping PHYS span [0x%" PRIx64
990 		    " - 0x%" PRIx64 "]", mp->ml_address,
991 		    mp->ml_address + mp->ml_size);
992 
993 		start = mp->ml_address;
994 		npages = mp->ml_size/IMMU_PAGESIZE + 1;
995 
996 		dcookies[0].dck_paddr = start;
997 		dcookies[0].dck_npages = npages;
998 		dcount = 1;
999 		(void) dvma_map(domain->dom_immu, domain, start, npages,
1000 		    dcookies, dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE);
1001 
1002 		mp = mp->ml_next;
1003 	}
1004 
1005 	memlist_read_unlock();
1006 }
1007 
1008 /*
1009  * create_xlate_arena()
1010  * 	Create the dvma arena for a domain with translation
1011  *	mapping
1012  */
1013 static void
1014 create_xlate_arena(immu_t *immu, domain_t *domain,
1015     dev_info_t *rdip, immu_flags_t immu_flags)
1016 {
1017 	char *arena_name;
1018 	struct memlist *mp;
1019 	int vmem_flags;
1020 	uint64_t start;
1021 	uint_t mgaw;
1022 	uint64_t size;
1023 	uint64_t maxaddr;
1024 	void *vmem_ret;
1025 
1026 	arena_name = domain->dom_dvma_arena_name;
1027 
1028 	/* Note, don't do sizeof (arena_name) - it is just a pointer */
1029 	(void) snprintf(arena_name,
1030 	    sizeof (domain->dom_dvma_arena_name),
1031 	    "%s-domain-%d-xlate-DVMA-arena", immu->immu_name,
1032 	    domain->dom_did);
1033 
1034 	vmem_flags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? VM_NOSLEEP : VM_SLEEP;
1035 
1036 	/*
1037 	 * No one else has access to this domain.
1038 	 * So no domain locks needed
1039 	 */
1040 	ASSERT(!rw_lock_held(&(domain->dom_pgtable_rwlock)));
1041 
1042 	/* Restrict mgaddr (max guest addr) to MGAW */
1043 	mgaw = IMMU_CAP_MGAW(immu->immu_regs_cap);
1044 
1045 	/*
1046 	 * To ensure we avoid ioapic and PCI MMIO ranges we just
1047 	 * use the physical memory address range of the system as the
1048 	 * range
1049 	 */
1050 	maxaddr = ((uint64_t)1 << mgaw);
1051 
1052 	memlist_read_lock();
1053 
1054 	mp = phys_install;
1055 
1056 	if (mp->ml_address == 0)
1057 		start = MMU_PAGESIZE;
1058 	else
1059 		start = mp->ml_address;
1060 
1061 	if (start + mp->ml_size > maxaddr)
1062 		size = maxaddr - start;
1063 	else
1064 		size = mp->ml_size;
1065 
1066 	ddi_err(DER_VERB, rdip,
1067 	    "%s: Creating dvma vmem arena [0x%" PRIx64
1068 	    " - 0x%" PRIx64 "]", arena_name, start, start + size);
1069 
1070 	ASSERT(domain->dom_dvma_arena == NULL);
1071 
1072 	/*
1073 	 * We always allocate in quanta of IMMU_PAGESIZE
1074 	 */
1075 	domain->dom_dvma_arena = vmem_create(arena_name,
1076 	    (void *)(uintptr_t)start,	/* start addr */
1077 	    size,			/* size */
1078 	    IMMU_PAGESIZE,		/* quantum */
1079 	    NULL,			/* afunc */
1080 	    NULL,			/* ffunc */
1081 	    NULL,			/* source */
1082 	    0,				/* qcache_max */
1083 	    vmem_flags);
1084 
1085 	if (domain->dom_dvma_arena == NULL) {
1086 		ddi_err(DER_PANIC, rdip,
1087 		    "Failed to allocate DVMA arena(%s) "
1088 		    "for domain ID (%d)", arena_name, domain->dom_did);
1089 		/*NOTREACHED*/
1090 	}
1091 
1092 	mp = mp->ml_next;
1093 	while (mp) {
1094 
1095 		if (mp->ml_address == 0)
1096 			start = MMU_PAGESIZE;
1097 		else
1098 			start = mp->ml_address;
1099 
1100 		if (start + mp->ml_size > maxaddr)
1101 			size = maxaddr - start;
1102 		else
1103 			size = mp->ml_size;
1104 
1105 		ddi_err(DER_VERB, rdip,
1106 		    "%s: Adding dvma vmem span [0x%" PRIx64
1107 		    " - 0x%" PRIx64 "]", arena_name, start,
1108 		    start + size);
1109 
1110 		vmem_ret = vmem_add(domain->dom_dvma_arena,
1111 		    (void *)(uintptr_t)start, size,  vmem_flags);
1112 
1113 		if (vmem_ret == NULL) {
1114 			ddi_err(DER_PANIC, rdip,
1115 			    "Failed to allocate DVMA arena(%s) "
1116 			    "for domain ID (%d)",
1117 			    arena_name, domain->dom_did);
1118 			/*NOTREACHED*/
1119 		}
1120 		mp = mp->ml_next;
1121 	}
1122 	memlist_read_unlock();
1123 }
1124 
1125 /* ################################### DOMAIN CODE ######################### */
1126 
1127 /*
1128  * Set the domain and domain-dip for a dip
1129  */
1130 static void
1131 set_domain(
1132 	dev_info_t *dip,
1133 	dev_info_t *ddip,
1134 	domain_t *domain)
1135 {
1136 	immu_devi_t *immu_devi;
1137 	domain_t *fdomain;
1138 	dev_info_t *fddip;
1139 
1140 	ASSERT(dip);
1141 	ASSERT(ddip);
1142 	ASSERT(domain);
1143 	ASSERT(domain->dom_did > 0); /* must be an initialized domain */
1144 
1145 	immu_devi = immu_devi_get(dip);
1146 	ASSERT(immu_devi);
1147 
1148 	mutex_enter(&(DEVI(dip)->devi_lock));
1149 	fddip = immu_devi->imd_ddip;
1150 	fdomain = immu_devi->imd_domain;
1151 
1152 	if (fddip) {
1153 		ASSERT(fddip == ddip);
1154 	} else {
1155 		immu_devi->imd_ddip = ddip;
1156 	}
1157 
1158 	if (fdomain) {
1159 		ASSERT(fdomain == domain);
1160 	} else {
1161 		immu_devi->imd_domain = domain;
1162 	}
1163 	mutex_exit(&(DEVI(dip)->devi_lock));
1164 }
1165 
1166 /*
1167  * device_domain()
1168  * 	Get domain for a device. The domain may be global in which case it
1169  *	is shared between all IOMMU units. Due to potential AGAW differences
1170  *      between IOMMU units, such global domains *have to be* UNITY mapping
1171  *      domains. Alternatively, the domain may be local to a IOMMU unit.
1172  *	Local domains may be shared or immu_devi, although the
1173  *      scope of sharing
1174  *	is restricted to devices controlled by the IOMMU unit to
1175  *      which the domain
1176  *	belongs. If shared, they (currently) have to be UNITY domains. If
1177  *      immu_devi a domain may be either UNITY or translation (XLATE) domain.
1178  */
1179 static domain_t *
1180 device_domain(dev_info_t *rdip, dev_info_t **ddipp, immu_flags_t immu_flags)
1181 {
1182 	dev_info_t *ddip; /* topmost dip in domain i.e. domain owner */
1183 	immu_t *immu;
1184 	domain_t *domain;
1185 	dvma_arg_t dvarg = {0};
1186 	int level;
1187 
1188 	ASSERT(rdip);
1189 
1190 	*ddipp = NULL;
1191 
1192 	/*
1193 	 * Check if the domain is already set. This is usually true
1194 	 * if this is not the first DVMA transaction.
1195 	 */
1196 	ddip = NULL;
1197 	domain = immu_devi_domain(rdip, &ddip);
1198 	if (domain) {
1199 		ASSERT(domain->dom_did > 0);
1200 		ASSERT(ddip);
1201 		*ddipp = ddip;
1202 		return (domain);
1203 	}
1204 
1205 	immu = immu_dvma_get_immu(rdip, immu_flags);
1206 	if (immu == NULL) {
1207 		/*
1208 		 * possible that there is no IOMMU unit for this device
1209 		 * - BIOS bugs are one example.
1210 		 */
1211 		ddi_err(DER_WARN, rdip, "No IMMU unit found for device");
1212 		return (NULL);
1213 	}
1214 
1215 	dvarg.dva_rdip = rdip;
1216 	dvarg.dva_ddip = NULL;
1217 	dvarg.dva_domain = NULL;
1218 	dvarg.dva_flags = immu_flags;
1219 	level = 0;
1220 	if (immu_walk_ancestor(rdip, NULL, get_branch_domain,
1221 	    &dvarg, &level, immu_flags) != DDI_SUCCESS) {
1222 		/*
1223 		 * maybe low memory. return error,
1224 		 * so driver tries again later
1225 		 */
1226 		return (NULL);
1227 	}
1228 
1229 	/* should have walked at least 1 dip (i.e. edip) */
1230 	ASSERT(level > 0);
1231 
1232 	ddip = dvarg.dva_ddip;	/* must be present */
1233 	domain = dvarg.dva_domain;	/* may be NULL */
1234 
1235 	/*
1236 	 * We may find the domain during our ancestor walk on any one of our
1237 	 * ancestor dips, If the domain is found then the domain-dip
1238 	 * (i.e. ddip) will also be found in the same immu_devi struct.
1239 	 * The domain-dip is the highest ancestor dip which shares the
1240 	 * same domain with edip.
1241 	 * The domain may or may not be found, but the domain dip must
1242 	 * be found.
1243 	 */
1244 	if (ddip == NULL) {
1245 		ddi_err(DER_MODE, rdip, "Cannot find domain dip for device.");
1246 		return (NULL);
1247 	}
1248 
1249 	/*
1250 	 * Did we find a domain ?
1251 	 */
1252 	if (domain) {
1253 		goto found;
1254 	}
1255 
1256 	/* nope, so allocate */
1257 	domain = domain_create(immu, ddip, rdip, immu_flags);
1258 	if (domain == NULL) {
1259 		return (NULL);
1260 	}
1261 	ASSERT(domain->dom_did > 0);
1262 
1263 	/*FALLTHROUGH*/
1264 found:
1265 	/*
1266 	 * We know *domain *is* the right domain, so panic if
1267 	 * another domain is set for either the request-dip or
1268 	 * effective dip.
1269 	 */
1270 	set_domain(ddip, ddip, domain);
1271 	set_domain(rdip, ddip, domain);
1272 
1273 	*ddipp = ddip;
1274 	return (domain);
1275 }
1276 
1277 static void
1278 create_unity_domain(immu_t *immu)
1279 {
1280 	domain_t *domain;
1281 
1282 	/* 0 is reserved by Vt-d */
1283 	/*LINTED*/
1284 	ASSERT(IMMU_UNITY_DID > 0);
1285 
1286 	/* domain created during boot and always use sleep flag */
1287 	domain = kmem_zalloc(sizeof (domain_t), KM_SLEEP);
1288 
1289 	rw_init(&(domain->dom_pgtable_rwlock), NULL, RW_DEFAULT, NULL);
1290 
1291 	domain->dom_did = IMMU_UNITY_DID;
1292 	domain->dom_maptype = IMMU_MAPTYPE_UNITY;
1293 
1294 	domain->dom_immu = immu;
1295 	immu->immu_unity_domain = domain;
1296 
1297 	/*
1298 	 * Setup the domain's initial page table
1299 	 * should never fail.
1300 	 */
1301 	domain->dom_pgtable_root = pgtable_alloc(immu, IMMU_FLAGS_SLEEP);
1302 	ASSERT(domain->dom_pgtable_root);
1303 	pgtable_zero(immu, domain->dom_pgtable_root);
1304 
1305 	map_unity_domain(domain);
1306 
1307 	/*
1308 	 * put it on the system-wide UNITY domain list
1309 	 */
1310 	mutex_enter(&(immu_domain_lock));
1311 	list_insert_tail(&immu_unity_domain_list, domain);
1312 	mutex_exit(&(immu_domain_lock));
1313 }
1314 
1315 /*
1316  * ddip is the domain-dip - the topmost dip in a domain
1317  * rdip is the requesting-dip - the device which is
1318  * requesting DVMA setup
1319  * if domain is a non-shared domain rdip == ddip
1320  */
1321 static domain_t *
1322 domain_create(immu_t *immu, dev_info_t *ddip, dev_info_t *rdip,
1323     immu_flags_t immu_flags)
1324 {
1325 	int kmflags;
1326 	domain_t *domain;
1327 	char mod_hash_name[128];
1328 	immu_devi_t *immu_devi;
1329 	int did;
1330 	dcookie_t dcookies[1] = {0};
1331 	int dcount = 0;
1332 
1333 	ASSERT(immu);
1334 	ASSERT(ddip);
1335 
1336 	immu_devi = immu_devi_get(rdip);
1337 
1338 	ASSERT(immu_devi);
1339 
1340 	/*
1341 	 * First allocate a domainid.
1342 	 * This routine will never fail, since if we run out
1343 	 * of domains the unity domain will be allocated.
1344 	 */
1345 	did = did_alloc(immu, rdip, ddip, immu_flags);
1346 	ASSERT(did > 0);
1347 	if (did == IMMU_UNITY_DID) {
1348 		/* domain overflow */
1349 		ASSERT(immu->immu_unity_domain);
1350 		return (immu->immu_unity_domain);
1351 	}
1352 
1353 	kmflags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
1354 	domain = kmem_zalloc(sizeof (domain_t), kmflags);
1355 	if (domain == NULL) {
1356 		ddi_err(DER_PANIC, rdip, "Failed to alloc DVMA domain "
1357 		    "structure for device. IOMMU unit: %s", immu->immu_name);
1358 		/*NOTREACHED*/
1359 	}
1360 
1361 	rw_init(&(domain->dom_pgtable_rwlock), NULL, RW_DEFAULT, NULL);
1362 
1363 	(void) snprintf(mod_hash_name, sizeof (mod_hash_name),
1364 	    "immu%s-domain%d-pava-hash", immu->immu_name, did);
1365 
1366 	domain->dom_did = did;
1367 	domain->dom_immu = immu;
1368 	domain->dom_maptype = IMMU_MAPTYPE_XLATE;
1369 
1370 	/*
1371 	 * Create xlate DVMA arena for this domain.
1372 	 */
1373 	create_xlate_arena(immu, domain, rdip, immu_flags);
1374 
1375 	/*
1376 	 * Setup the domain's initial page table
1377 	 */
1378 	domain->dom_pgtable_root = pgtable_alloc(immu, immu_flags);
1379 	if (domain->dom_pgtable_root == NULL) {
1380 		ddi_err(DER_PANIC, rdip, "Failed to alloc root "
1381 		    "pgtable for domain (%d). IOMMU unit: %s",
1382 		    domain->dom_did, immu->immu_name);
1383 		/*NOTREACHED*/
1384 	}
1385 	pgtable_zero(immu, domain->dom_pgtable_root);
1386 
1387 	/*
1388 	 * Since this is a immu unit-specific domain, put it on
1389 	 * the per-immu domain list.
1390 	 */
1391 	mutex_enter(&(immu->immu_lock));
1392 	list_insert_head(&immu->immu_domain_list, domain);
1393 	mutex_exit(&(immu->immu_lock));
1394 
1395 	/*
1396 	 * Also put it on the system-wide xlate domain list
1397 	 */
1398 	mutex_enter(&(immu_domain_lock));
1399 	list_insert_head(&immu_xlate_domain_list, domain);
1400 	mutex_exit(&(immu_domain_lock));
1401 
1402 	bdf_domain_insert(immu_devi, domain);
1403 
1404 #ifdef BUGGY_DRIVERS
1405 	/*
1406 	 * Map page0. Some broken HW/FW access it.
1407 	 */
1408 	dcookies[0].dck_paddr = 0;
1409 	dcookies[0].dck_npages = 1;
1410 	dcount = 1;
1411 	(void) dvma_map(domain->dom_immu, domain, 0, 1, dcookies, dcount, NULL,
1412 	    IMMU_FLAGS_READ | IMMU_FLAGS_WRITE | IMMU_FLAGS_PAGE1);
1413 #endif
1414 	return (domain);
1415 }
1416 
1417 /*
1418  * Create domainid arena.
1419  * Domainid 0 is reserved by Vt-d spec and cannot be used by
1420  * system software.
1421  * Domainid 1 is reserved by solaris and used for *all* of the following:
1422  *	as the "uninitialized" domain - For devices not yet controlled
1423  *	by Solaris
1424  *	as the "unity" domain - For devices that will always belong
1425  *	to the unity domain
1426  *	as the "overflow" domain - Used for any new device after we
1427  *	run out of domains
1428  * All of the above domains map into a single domain with
1429  * domainid 1 and UNITY DVMA mapping
1430  * Each IMMU unity has its own unity/uninit/overflow domain
1431  */
1432 static void
1433 did_init(immu_t *immu)
1434 {
1435 	(void) snprintf(immu->immu_did_arena_name,
1436 	    sizeof (immu->immu_did_arena_name),
1437 	    "%s_domainid_arena", immu->immu_name);
1438 
1439 	ddi_err(DER_VERB, NULL, "%s: Creating domainid arena %s",
1440 	    immu->immu_name, immu->immu_did_arena_name);
1441 
1442 	immu->immu_did_arena = vmem_create(
1443 	    immu->immu_did_arena_name,
1444 	    (void *)(uintptr_t)(IMMU_UNITY_DID + 1),   /* start addr */
1445 	    immu->immu_max_domains - IMMU_UNITY_DID,
1446 	    1,				/* quantum */
1447 	    NULL,			/* afunc */
1448 	    NULL,			/* ffunc */
1449 	    NULL,			/* source */
1450 	    0,				/* qcache_max */
1451 	    VM_SLEEP);
1452 
1453 	/* Even with SLEEP flag, vmem_create() can fail */
1454 	if (immu->immu_did_arena == NULL) {
1455 		ddi_err(DER_PANIC, NULL, "%s: Failed to create Intel "
1456 		    "IOMMU domainid allocator: %s", immu->immu_name,
1457 		    immu->immu_did_arena_name);
1458 	}
1459 }
1460 
1461 /* #########################  CONTEXT CODE ################################# */
1462 
1463 static void
1464 context_set(immu_t *immu, domain_t *domain, pgtable_t *root_table,
1465     int bus, int devfunc)
1466 {
1467 	pgtable_t *context;
1468 	pgtable_t *pgtable_root;
1469 	pgtable_t *unity_pgtable_root;
1470 	hw_rce_t *hw_rent;
1471 	hw_rce_t *hw_cent;
1472 	hw_rce_t *ctxp;
1473 	int sid;
1474 	krw_t rwtype;
1475 	boolean_t fill_root;
1476 	boolean_t fill_ctx;
1477 
1478 	ASSERT(immu);
1479 	ASSERT(domain);
1480 	ASSERT(root_table);
1481 	ASSERT(bus >= 0);
1482 	ASSERT(devfunc >= 0);
1483 	ASSERT(domain->dom_pgtable_root);
1484 
1485 	pgtable_root = domain->dom_pgtable_root;
1486 
1487 	ctxp = (hw_rce_t *)(root_table->swpg_next_array);
1488 	context = *(pgtable_t **)(ctxp + bus);
1489 	hw_rent = (hw_rce_t *)(root_table->hwpg_vaddr) + bus;
1490 
1491 	fill_root = B_FALSE;
1492 	fill_ctx = B_FALSE;
1493 
1494 	/* Check the most common case first with reader lock */
1495 	rw_enter(&(immu->immu_ctx_rwlock), RW_READER);
1496 	rwtype = RW_READER;
1497 again:
1498 	if (ROOT_GET_P(hw_rent)) {
1499 		ASSERT(ROOT_GET_CONT(hw_rent) == context->hwpg_paddr);
1500 		hw_cent = (hw_rce_t *)(context->hwpg_vaddr) + devfunc;
1501 		if (CONT_GET_AVAIL(hw_cent) == IMMU_CONT_INITED) {
1502 			ASSERT(CONT_GET_P(hw_cent));
1503 			ASSERT(CONT_GET_DID(hw_cent) == domain->dom_did);
1504 			ASSERT(CONT_GET_AW(hw_cent) == immu->immu_dvma_agaw);
1505 			ASSERT(CONT_GET_TTYPE(hw_cent) == TTYPE_XLATE_ONLY);
1506 			ASSERT(CONT_GET_ASR(hw_cent) ==
1507 			    pgtable_root->hwpg_paddr);
1508 			rw_exit(&(immu->immu_ctx_rwlock));
1509 			return;
1510 		} else {
1511 			fill_ctx = B_TRUE;
1512 		}
1513 	} else {
1514 		fill_root = B_TRUE;
1515 		fill_ctx = B_TRUE;
1516 	}
1517 
1518 	if (rwtype == RW_READER &&
1519 	    rw_tryupgrade(&(immu->immu_ctx_rwlock)) == 0) {
1520 		rw_exit(&(immu->immu_ctx_rwlock));
1521 		rw_enter(&(immu->immu_ctx_rwlock), RW_WRITER);
1522 		rwtype = RW_WRITER;
1523 		goto again;
1524 	}
1525 	rwtype = RW_WRITER;
1526 
1527 	if (fill_root == B_TRUE) {
1528 		ROOT_SET_CONT(hw_rent, context->hwpg_paddr);
1529 		ROOT_SET_P(hw_rent);
1530 		immu_regs_cpu_flush(immu, (caddr_t)hw_rent, sizeof (hw_rce_t));
1531 	}
1532 
1533 	if (fill_ctx == B_TRUE) {
1534 		hw_cent = (hw_rce_t *)(context->hwpg_vaddr) + devfunc;
1535 		unity_pgtable_root = immu->immu_unity_domain->dom_pgtable_root;
1536 		ASSERT(CONT_GET_AVAIL(hw_cent) == IMMU_CONT_UNINITED);
1537 		ASSERT(CONT_GET_P(hw_cent));
1538 		ASSERT(CONT_GET_DID(hw_cent) ==
1539 		    immu->immu_unity_domain->dom_did);
1540 		ASSERT(CONT_GET_AW(hw_cent) == immu->immu_dvma_agaw);
1541 		ASSERT(CONT_GET_TTYPE(hw_cent) == TTYPE_XLATE_ONLY);
1542 		ASSERT(CONT_GET_ASR(hw_cent) ==
1543 		    unity_pgtable_root->hwpg_paddr);
1544 
1545 		/* need to disable context entry before reprogramming it */
1546 		bzero(hw_cent, sizeof (hw_rce_t));
1547 
1548 		/* flush caches */
1549 		immu_regs_cpu_flush(immu, (caddr_t)hw_cent, sizeof (hw_rce_t));
1550 		ASSERT(rw_write_held(&(immu->immu_ctx_rwlock)));
1551 
1552 		sid = ((bus << 8) | devfunc);
1553 		immu_regs_context_flush(immu, 0, sid, domain->dom_did,
1554 		    CONTEXT_FSI);
1555 
1556 		immu_regs_wbf_flush(immu);
1557 
1558 		CONT_SET_AVAIL(hw_cent, IMMU_CONT_INITED);
1559 		CONT_SET_DID(hw_cent, domain->dom_did);
1560 		CONT_SET_AW(hw_cent, immu->immu_dvma_agaw);
1561 		CONT_SET_ASR(hw_cent, pgtable_root->hwpg_paddr);
1562 		/*LINTED*/
1563 		CONT_SET_TTYPE(hw_cent, TTYPE_XLATE_ONLY);
1564 		CONT_SET_P(hw_cent);
1565 		immu_regs_cpu_flush(immu, (caddr_t)hw_cent, sizeof (hw_rce_t));
1566 	}
1567 	rw_exit(&(immu->immu_ctx_rwlock));
1568 }
1569 
1570 static pgtable_t *
1571 context_create(immu_t *immu)
1572 {
1573 	int	bus;
1574 	int	devfunc;
1575 	pgtable_t *root_table;
1576 	pgtable_t *context;
1577 	pgtable_t *pgtable_root;
1578 	hw_rce_t *ctxp;
1579 	hw_rce_t *hw_rent;
1580 	hw_rce_t *hw_cent;
1581 
1582 	/* Allocate a zeroed root table (4K 256b entries) */
1583 	root_table = pgtable_alloc(immu, IMMU_FLAGS_SLEEP);
1584 	pgtable_zero(immu, root_table);
1585 
1586 	/*
1587 	 * Setup context tables for all possible root table entries.
1588 	 * Start out with unity domains for all entries.
1589 	 */
1590 	ctxp = (hw_rce_t *)(root_table->swpg_next_array);
1591 	hw_rent = (hw_rce_t *)(root_table->hwpg_vaddr);
1592 	for (bus = 0; bus < IMMU_ROOT_NUM; bus++, ctxp++, hw_rent++) {
1593 		context = pgtable_alloc(immu, IMMU_FLAGS_SLEEP);
1594 		pgtable_zero(immu, context);
1595 		ASSERT(ROOT_GET_P(hw_rent) == 0);
1596 		ROOT_SET_P(hw_rent);
1597 		ROOT_SET_CONT(hw_rent, context->hwpg_paddr);
1598 		hw_cent = (hw_rce_t *)(context->hwpg_vaddr);
1599 		for (devfunc = 0; devfunc < IMMU_CONT_NUM;
1600 		    devfunc++, hw_cent++) {
1601 			ASSERT(CONT_GET_P(hw_cent) == 0);
1602 			pgtable_root =
1603 			    immu->immu_unity_domain->dom_pgtable_root;
1604 			CONT_SET_DID(hw_cent,
1605 			    immu->immu_unity_domain->dom_did);
1606 			CONT_SET_AW(hw_cent, immu->immu_dvma_agaw);
1607 			CONT_SET_ASR(hw_cent, pgtable_root->hwpg_paddr);
1608 			/*LINTED*/
1609 			CONT_SET_TTYPE(hw_cent, TTYPE_XLATE_ONLY);
1610 			CONT_SET_AVAIL(hw_cent, IMMU_CONT_UNINITED);
1611 			CONT_SET_P(hw_cent);
1612 		}
1613 		immu_regs_cpu_flush(immu, context->hwpg_vaddr, IMMU_PAGESIZE);
1614 		*((pgtable_t **)ctxp) = context;
1615 	}
1616 	immu_regs_cpu_flush(immu, root_table->hwpg_vaddr, IMMU_PAGESIZE);
1617 
1618 	return (root_table);
1619 }
1620 
1621 /*
1622  * Called during rootnex attach, so no locks needed
1623  */
1624 static void
1625 context_init(immu_t *immu)
1626 {
1627 	ASSERT(immu);
1628 	ASSERT(immu->immu_ctx_root == NULL);
1629 
1630 	rw_init(&(immu->immu_ctx_rwlock), NULL, RW_DEFAULT, NULL);
1631 
1632 	immu_regs_wbf_flush(immu);
1633 
1634 	immu->immu_ctx_root = context_create(immu);
1635 
1636 	immu_regs_set_root_table(immu);
1637 
1638 	rw_enter(&(immu->immu_ctx_rwlock), RW_WRITER);
1639 	immu_regs_context_flush(immu, 0, 0, 0, CONTEXT_GLOBAL);
1640 	rw_exit(&(immu->immu_ctx_rwlock));
1641 	immu_regs_iotlb_flush(immu, 0, 0, 0, 0, IOTLB_GLOBAL);
1642 	immu_regs_wbf_flush(immu);
1643 }
1644 
1645 
1646 /*
1647  * Find top pcib
1648  */
1649 static int
1650 find_top_pcib(dev_info_t *dip, void *arg)
1651 {
1652 	immu_devi_t *immu_devi;
1653 	dev_info_t **pcibdipp = (dev_info_t **)arg;
1654 
1655 	ASSERT(dip);
1656 
1657 	immu_devi = immu_devi_get(dip);
1658 	ASSERT(immu_devi);
1659 
1660 	if (immu_devi->imd_pcib_type == IMMU_PCIB_PCI_PCI) {
1661 		*pcibdipp = dip;
1662 	}
1663 
1664 	return (DDI_WALK_CONTINUE);
1665 }
1666 
1667 static int
1668 immu_context_update(immu_t *immu, domain_t *domain, dev_info_t *ddip,
1669     dev_info_t *rdip, immu_flags_t immu_flags)
1670 {
1671 	immu_devi_t *r_immu_devi;
1672 	immu_devi_t *d_immu_devi;
1673 	int r_bus;
1674 	int d_bus;
1675 	int r_devfunc;
1676 	int d_devfunc;
1677 	immu_pcib_t d_pcib_type;
1678 	immu_pcib_t r_pcib_type;
1679 	dev_info_t *pcibdip;
1680 
1681 	if (ddip == NULL || rdip == NULL ||
1682 	    ddip == root_devinfo || rdip == root_devinfo) {
1683 		ddi_err(DER_MODE, rdip, "immu_contexts_update: domain-dip or "
1684 		    "request-dip are NULL or are root devinfo");
1685 		return (DDI_FAILURE);
1686 	}
1687 
1688 	/*
1689 	 * We need to set the context fields
1690 	 * based on what type of device rdip and ddip are.
1691 	 * To do that we need the immu_devi field.
1692 	 * Set the immu_devi field (if not already set)
1693 	 */
1694 	if (immu_devi_set(ddip, immu_flags) == DDI_FAILURE) {
1695 		ddi_err(DER_MODE, rdip,
1696 		    "immu_context_update: failed to set immu_devi for ddip");
1697 		return (DDI_FAILURE);
1698 	}
1699 
1700 	if (immu_devi_set(rdip, immu_flags) == DDI_FAILURE) {
1701 		ddi_err(DER_MODE, rdip,
1702 		    "immu_context_update: failed to set immu_devi for rdip");
1703 		return (DDI_FAILURE);
1704 	}
1705 
1706 	d_immu_devi = immu_devi_get(ddip);
1707 	r_immu_devi = immu_devi_get(rdip);
1708 	ASSERT(r_immu_devi);
1709 	ASSERT(d_immu_devi);
1710 
1711 	d_bus = d_immu_devi->imd_bus;
1712 	d_devfunc = d_immu_devi->imd_devfunc;
1713 	d_pcib_type = d_immu_devi->imd_pcib_type;
1714 	r_bus = r_immu_devi->imd_bus;
1715 	r_devfunc = r_immu_devi->imd_devfunc;
1716 	r_pcib_type = r_immu_devi->imd_pcib_type;
1717 
1718 	ASSERT(d_bus >= 0);
1719 
1720 	if (rdip == ddip) {
1721 		ASSERT(d_pcib_type == IMMU_PCIB_ENDPOINT ||
1722 		    d_pcib_type == IMMU_PCIB_PCIE_PCIE);
1723 		ASSERT(r_bus >= 0);
1724 		ASSERT(r_devfunc >= 0);
1725 		/* rdip is a PCIE device. set context for it only */
1726 		context_set(immu, domain, immu->immu_ctx_root, r_bus,
1727 		    r_devfunc);
1728 #ifdef BUGGY_DRIVERS
1729 	} else if (r_immu_devi == d_immu_devi) {
1730 #ifdef TEST
1731 		ddi_err(DER_WARN, rdip, "Driver bug: Devices 0x%lx and "
1732 		    "0x%lx are identical", rdip, ddip);
1733 #endif
1734 		ASSERT(d_pcib_type == IMMU_PCIB_ENDPOINT);
1735 		ASSERT(r_bus >= 0);
1736 		ASSERT(r_devfunc >= 0);
1737 		/* rdip is a PCIE device. set context for it only */
1738 		context_set(immu, domain, immu->immu_ctx_root, r_bus,
1739 		    r_devfunc);
1740 #endif
1741 	} else if (d_pcib_type == IMMU_PCIB_PCIE_PCI) {
1742 		/*
1743 		 * ddip is a PCIE_PCI bridge. Set context for ddip's
1744 		 * secondary bus. If rdip is on ddip's secondary
1745 		 * bus, set context for rdip. Else, set context
1746 		 * for rdip's PCI bridge on ddip's secondary bus.
1747 		 */
1748 		context_set(immu, domain, immu->immu_ctx_root,
1749 		    d_immu_devi->imd_sec, 0);
1750 		if (d_immu_devi->imd_sec == r_bus) {
1751 			context_set(immu, domain, immu->immu_ctx_root,
1752 			    r_bus, r_devfunc);
1753 		} else {
1754 			pcibdip = NULL;
1755 			if (immu_walk_ancestor(rdip, ddip, find_top_pcib,
1756 			    &pcibdip, NULL, immu_flags) == DDI_SUCCESS &&
1757 			    pcibdip != NULL) {
1758 				ASSERT(pcibdip);
1759 				r_immu_devi = immu_devi_get(pcibdip);
1760 				ASSERT(d_immu_devi);
1761 				ASSERT(d_immu_devi->imd_pcib_type ==
1762 				    IMMU_PCIB_PCI_PCI);
1763 				r_bus = r_immu_devi->imd_bus;
1764 				r_devfunc = r_immu_devi->imd_devfunc;
1765 				context_set(immu, domain, immu->immu_ctx_root,
1766 				    r_bus, r_devfunc);
1767 			} else {
1768 				ddi_err(DER_PANIC, rdip, "Failed to find PCI "
1769 				    " bridge for PCI device");
1770 				/*NOTREACHED*/
1771 			}
1772 		}
1773 	} else if (d_pcib_type == IMMU_PCIB_PCI_PCI) {
1774 		context_set(immu, domain, immu->immu_ctx_root, d_bus,
1775 		    d_devfunc);
1776 	} else if (d_pcib_type == IMMU_PCIB_ENDPOINT) {
1777 		ASSERT(r_pcib_type == IMMU_PCIB_NOBDF);
1778 		/*
1779 		 * ddip is a PCIE device which has a non-PCI device under it
1780 		 * i.e. it is a PCI-nonPCI bridge. Example: pciicde-ata
1781 		 */
1782 		context_set(immu, domain, immu->immu_ctx_root, d_bus,
1783 		    d_devfunc);
1784 	} else {
1785 		ddi_err(DER_PANIC, rdip, "unknown device type. Cannot "
1786 		    "set IMMU context.");
1787 		/*NOTREACHED*/
1788 	}
1789 
1790 	/* XXX do we need a membar_producer() here */
1791 	return (DDI_SUCCESS);
1792 }
1793 
1794 /* ##################### END CONTEXT CODE ################################## */
1795 /* ##################### MAPPING CODE ################################## */
1796 
1797 
1798 static boolean_t
1799 PDTE_check(immu_t *immu, hw_pdte_t pdte, pgtable_t *next, paddr_t paddr,
1800     dev_info_t *rdip, immu_flags_t immu_flags)
1801 {
1802 	if (immu_flags & IMMU_FLAGS_PAGE1) {
1803 		ASSERT(paddr == 0);
1804 	} else {
1805 		ASSERT((next == NULL) ^ (paddr == 0));
1806 	}
1807 
1808 	/* The PDTE must be set i.e. present bit is set */
1809 	if (!PDTE_P(pdte)) {
1810 		ddi_err(DER_MODE, rdip, "No present flag");
1811 		return (B_FALSE);
1812 	}
1813 
1814 	/*
1815 	 * Just assert to check most significant system software field
1816 	 * (PDTE_SW4) as it is same as present bit and we
1817 	 * checked that above
1818 	 */
1819 	ASSERT(PDTE_SW4(pdte));
1820 
1821 	/*
1822 	 * TM field should be clear if not reserved.
1823 	 * non-leaf is always reserved
1824 	 */
1825 	if (next == NULL && immu->immu_TM_reserved == B_FALSE) {
1826 		if (PDTE_TM(pdte)) {
1827 			ddi_err(DER_MODE, rdip, "TM flag set");
1828 			return (B_FALSE);
1829 		}
1830 	}
1831 
1832 	/*
1833 	 * The SW3 field is not used and must be clear
1834 	 */
1835 	if (PDTE_SW3(pdte)) {
1836 		ddi_err(DER_MODE, rdip, "SW3 set");
1837 		return (B_FALSE);
1838 	}
1839 
1840 	/*
1841 	 * PFN (for PTE) or next level pgtable-paddr (for PDE) must be set
1842 	 */
1843 	if (next == NULL) {
1844 		ASSERT(paddr % IMMU_PAGESIZE == 0);
1845 		if (PDTE_PADDR(pdte) != paddr) {
1846 			ddi_err(DER_MODE, rdip,
1847 			    "PTE paddr mismatch: %lx != %lx",
1848 			    PDTE_PADDR(pdte), paddr);
1849 			return (B_FALSE);
1850 		}
1851 	} else {
1852 		if (PDTE_PADDR(pdte) != next->hwpg_paddr) {
1853 			ddi_err(DER_MODE, rdip,
1854 			    "PDE paddr mismatch: %lx != %lx",
1855 			    PDTE_PADDR(pdte), next->hwpg_paddr);
1856 			return (B_FALSE);
1857 		}
1858 	}
1859 
1860 	/*
1861 	 * SNP field should be clear if not reserved.
1862 	 * non-leaf is always reserved
1863 	 */
1864 	if (next == NULL && immu->immu_SNP_reserved == B_FALSE) {
1865 		if (PDTE_SNP(pdte)) {
1866 			ddi_err(DER_MODE, rdip, "SNP set");
1867 			return (B_FALSE);
1868 		}
1869 	}
1870 
1871 	/* second field available for system software should be clear */
1872 	if (PDTE_SW2(pdte)) {
1873 		ddi_err(DER_MODE, rdip, "SW2 set");
1874 		return (B_FALSE);
1875 	}
1876 
1877 	/* Super pages field should be clear */
1878 	if (PDTE_SP(pdte)) {
1879 		ddi_err(DER_MODE, rdip, "SP set");
1880 		return (B_FALSE);
1881 	}
1882 
1883 	/*
1884 	 * least significant field available for
1885 	 * system software should be clear
1886 	 */
1887 	if (PDTE_SW1(pdte)) {
1888 		ddi_err(DER_MODE, rdip, "SW1 set");
1889 		return (B_FALSE);
1890 	}
1891 
1892 	if ((immu_flags & IMMU_FLAGS_READ) && !PDTE_READ(pdte)) {
1893 		ddi_err(DER_MODE, rdip, "READ not set");
1894 		return (B_FALSE);
1895 	}
1896 
1897 	if ((immu_flags & IMMU_FLAGS_WRITE) && !PDTE_WRITE(pdte)) {
1898 		ddi_err(DER_MODE, rdip, "WRITE not set");
1899 		return (B_FALSE);
1900 	}
1901 
1902 	return (B_TRUE);
1903 }
1904 /*ARGSUSED*/
1905 static void
1906 PTE_clear_all(immu_t *immu, domain_t *domain, xlate_t *xlate,
1907     uint64_t *dvma_ptr, uint64_t *npages_ptr, dev_info_t *rdip)
1908 {
1909 	uint64_t npages;
1910 	uint64_t dvma;
1911 	pgtable_t *pgtable;
1912 	hw_pdte_t *hwp;
1913 	hw_pdte_t *shwp;
1914 	int idx;
1915 	hw_pdte_t pte;
1916 
1917 	ASSERT(xlate->xlt_level == 1);
1918 
1919 	pgtable = xlate->xlt_pgtable;
1920 	idx = xlate->xlt_idx;
1921 
1922 	ASSERT(pgtable);
1923 	ASSERT(idx <= IMMU_PGTABLE_MAXIDX);
1924 
1925 	dvma = *dvma_ptr;
1926 	npages = *npages_ptr;
1927 
1928 	ASSERT(dvma);
1929 	ASSERT(dvma % IMMU_PAGESIZE == 0);
1930 	ASSERT(npages);
1931 
1932 	/*
1933 	 * since a caller gets a unique dvma for a physical address,
1934 	 * no other concurrent thread will be writing to the same
1935 	 * PTE even if it has the same paddr. So no locks needed.
1936 	 */
1937 	shwp = (hw_pdte_t *)(pgtable->hwpg_vaddr) + idx;
1938 
1939 	hwp = shwp;
1940 	for (; npages > 0 && idx <= IMMU_PGTABLE_MAXIDX; idx++, hwp++) {
1941 
1942 		pte = *hwp;
1943 
1944 		/* Cannot clear a HW PTE that is aleady clear */
1945 		ASSERT(PDTE_P(pte));
1946 		PDTE_CLEAR_P(pte);
1947 		*hwp = pte;
1948 
1949 		dvma += IMMU_PAGESIZE;
1950 		npages--;
1951 	}
1952 
1953 
1954 #ifdef TEST
1955 	/* dont need to flush write during unmap */
1956 	immu_regs_cpu_flush(immu, (caddr_t)shwp,
1957 	    (hwp - shwp) * sizeof (hw_pdte_t));
1958 #endif
1959 
1960 	*dvma_ptr = dvma;
1961 	*npages_ptr = npages;
1962 
1963 	xlate->xlt_idx = idx;
1964 }
1965 
1966 /*ARGSUSED*/
1967 static void
1968 xlate_setup(immu_t *immu, uint64_t dvma, xlate_t *xlate,
1969     int nlevels, dev_info_t *rdip)
1970 {
1971 	int level;
1972 	uint64_t offbits;
1973 
1974 	/* level 0 is never used. Sanity check */
1975 	ASSERT(xlate->xlt_level == 0);
1976 	ASSERT(xlate->xlt_idx == 0);
1977 	ASSERT(xlate->xlt_pgtable == NULL);
1978 	ASSERT(dvma % IMMU_PAGESIZE == 0);
1979 
1980 	/*
1981 	 * Skip the first 12 bits which is the offset into
1982 	 * 4K PFN (phys page frame based on IMMU_PAGESIZE)
1983 	 */
1984 	offbits = dvma >> IMMU_PAGESHIFT;
1985 
1986 	/* skip to level 1 i.e. leaf PTE */
1987 	for (level = 1, xlate++; level <= nlevels; level++, xlate++) {
1988 		xlate->xlt_level = level;
1989 		xlate->xlt_idx = (offbits & IMMU_PGTABLE_LEVEL_MASK);
1990 		ASSERT(xlate->xlt_idx <= IMMU_PGTABLE_MAXIDX);
1991 		xlate->xlt_pgtable = NULL;
1992 		offbits >>= IMMU_PGTABLE_LEVEL_STRIDE;
1993 	}
1994 }
1995 
1996 /*
1997  * Read the pgtables
1998  */
1999 static void
2000 PDE_lookup(immu_t *immu, domain_t *domain, xlate_t *xlate, int nlevels,
2001     dev_info_t *rdip)
2002 {
2003 	pgtable_t *pgtable;
2004 	pgtable_t *next;
2005 	hw_pdte_t pde;
2006 	uint_t idx;
2007 
2008 	/* xlate should be at level 0 */
2009 	ASSERT(xlate->xlt_level == 0);
2010 	ASSERT(xlate->xlt_idx == 0);
2011 
2012 	/* start with highest level pgtable i.e. root */
2013 	xlate += nlevels;
2014 	ASSERT(xlate->xlt_level == nlevels);
2015 
2016 	if (xlate->xlt_pgtable == NULL) {
2017 		xlate->xlt_pgtable = domain->dom_pgtable_root;
2018 	}
2019 
2020 	for (; xlate->xlt_level > 1; xlate--) {
2021 
2022 		idx = xlate->xlt_idx;
2023 		pgtable = xlate->xlt_pgtable;
2024 
2025 		ASSERT(pgtable);
2026 		ASSERT(idx <= IMMU_PGTABLE_MAXIDX);
2027 
2028 		if ((xlate - 1)->xlt_pgtable) {
2029 			continue;
2030 		}
2031 
2032 		/* xlate's leafier level is not set, set it now */
2033 
2034 		/* Lock the pgtable in read mode */
2035 		rw_enter(&(pgtable->swpg_rwlock), RW_READER);
2036 
2037 		/*
2038 		 * since we are unmapping, the pgtable should
2039 		 * already point to a leafier pgtable.
2040 		 */
2041 		next = *(pgtable->swpg_next_array + idx);
2042 		ASSERT(next);
2043 
2044 		pde = *((hw_pdte_t *)(pgtable->hwpg_vaddr) + idx);
2045 
2046 		ASSERT(PDTE_check(immu, pde, next, 0, rdip, 0) == B_TRUE);
2047 
2048 		(xlate - 1)->xlt_pgtable = next;
2049 
2050 		rw_exit(&(pgtable->swpg_rwlock));
2051 	}
2052 }
2053 
2054 /*ARGSUSED*/
2055 static void
2056 PTE_set_one(immu_t *immu, hw_pdte_t *hwp, paddr_t paddr,
2057     dev_info_t *rdip, immu_flags_t immu_flags)
2058 {
2059 	hw_pdte_t pte;
2060 
2061 	pte = *hwp;
2062 
2063 #ifndef DEBUG
2064 	/* Set paddr */
2065 	ASSERT(paddr % IMMU_PAGESIZE == 0);
2066 	pte = 0;
2067 	PDTE_SET_PADDR(pte, paddr);
2068 	PDTE_SET_READ(pte);
2069 	PDTE_SET_WRITE(pte);
2070 	*hwp = pte;
2071 #else
2072 
2073 	if (PDTE_P(pte)) {
2074 		if (PDTE_PADDR(pte) != paddr) {
2075 			ddi_err(DER_MODE, rdip, "PTE paddr %lx != paddr %lx",
2076 			    PDTE_PADDR(pte), paddr);
2077 		}
2078 #ifdef BUGGY_DRIVERS
2079 		return;
2080 #else
2081 		goto out;
2082 #endif
2083 	}
2084 
2085 	/* Don't touch SW4. It is the present field */
2086 
2087 	/* clear TM field if not reserved */
2088 	if (immu->immu_TM_reserved == B_FALSE) {
2089 		PDTE_CLEAR_TM(pte);
2090 	}
2091 
2092 #ifdef DEBUG
2093 	/* Clear 3rd field for system software  - not used */
2094 	PDTE_CLEAR_SW3(pte);
2095 #endif
2096 
2097 	/* Set paddr */
2098 	ASSERT(paddr % IMMU_PAGESIZE == 0);
2099 	PDTE_CLEAR_PADDR(pte);
2100 	PDTE_SET_PADDR(pte, paddr);
2101 
2102 	/*  clear SNP field if not reserved. */
2103 	if (immu->immu_SNP_reserved == B_FALSE) {
2104 		PDTE_CLEAR_SNP(pte);
2105 	}
2106 
2107 #ifdef DEBUG
2108 	/* Clear SW2 field available for software */
2109 	PDTE_CLEAR_SW2(pte);
2110 #endif
2111 
2112 
2113 #ifdef DEBUG
2114 	/* SP is don't care for PTEs. Clear it for cleanliness */
2115 	PDTE_CLEAR_SP(pte);
2116 #endif
2117 
2118 #ifdef DEBUG
2119 	/* Clear SW1 field available for software */
2120 	PDTE_CLEAR_SW1(pte);
2121 #endif
2122 
2123 	/*
2124 	 * Now that we are done writing the PTE
2125 	 * set the "present" flag. Note this present
2126 	 * flag is a bit in the PDE/PTE that the
2127 	 * spec says is available for system software.
2128 	 * This is an implementation detail of Solaris
2129 	 * bare-metal Intel IOMMU.
2130 	 * The present field in a PDE/PTE is not defined
2131 	 * by the Vt-d spec
2132 	 */
2133 
2134 	PDTE_SET_P(pte);
2135 
2136 out:
2137 #ifdef BUGGY_DRIVERS
2138 	PDTE_SET_READ(pte);
2139 	PDTE_SET_WRITE(pte);
2140 #else
2141 	if (immu_flags & IMMU_FLAGS_READ)
2142 		PDTE_SET_READ(pte);
2143 	if (immu_flags & IMMU_FLAGS_WRITE)
2144 		PDTE_SET_WRITE(pte);
2145 #endif
2146 
2147 	*hwp = pte;
2148 #endif
2149 }
2150 
2151 /*ARGSUSED*/
2152 static void
2153 PTE_set_all(immu_t *immu, domain_t *domain, xlate_t *xlate,
2154     uint64_t *dvma_ptr, uint64_t *nvpages_ptr, dcookie_t *dcookies,
2155     int dcount, dev_info_t *rdip, immu_flags_t immu_flags)
2156 {
2157 	paddr_t paddr;
2158 	uint64_t nvpages;
2159 	uint64_t nppages;
2160 	uint64_t dvma;
2161 	pgtable_t *pgtable;
2162 	hw_pdte_t *hwp;
2163 	hw_pdte_t *shwp;
2164 	int idx;
2165 	int j;
2166 
2167 	ASSERT(xlate->xlt_level == 1);
2168 
2169 	pgtable = xlate->xlt_pgtable;
2170 	idx = xlate->xlt_idx;
2171 
2172 	ASSERT(idx <= IMMU_PGTABLE_MAXIDX);
2173 	ASSERT(pgtable);
2174 
2175 	dvma = *dvma_ptr;
2176 	nvpages = *nvpages_ptr;
2177 
2178 	ASSERT(dvma || (immu_flags & IMMU_FLAGS_PAGE1));
2179 	ASSERT(nvpages);
2180 
2181 	/*
2182 	 * since a caller gets a unique dvma for a physical address,
2183 	 * no other concurrent thread will be writing to the same
2184 	 * PTE even if it has the same paddr. So no locks needed.
2185 	 */
2186 	shwp = (hw_pdte_t *)(pgtable->hwpg_vaddr) + idx;
2187 
2188 	hwp = shwp;
2189 	for (j = dcount - 1; j >= 0; j--) {
2190 		if (nvpages <= dcookies[j].dck_npages)
2191 			break;
2192 		nvpages -= dcookies[j].dck_npages;
2193 	}
2194 
2195 	ASSERT(j >= 0);
2196 	ASSERT(nvpages);
2197 	ASSERT(nvpages <= dcookies[j].dck_npages);
2198 	nppages = nvpages;
2199 	paddr = dcookies[j].dck_paddr +
2200 	    (dcookies[j].dck_npages - nppages) * IMMU_PAGESIZE;
2201 
2202 	nvpages = *nvpages_ptr;
2203 	for (; nvpages > 0 && idx <= IMMU_PGTABLE_MAXIDX; idx++, hwp++) {
2204 
2205 		ASSERT(paddr || (immu_flags & IMMU_FLAGS_PAGE1));
2206 
2207 		PTE_set_one(immu, hwp, paddr, rdip, immu_flags);
2208 
2209 		ASSERT(PDTE_check(immu, *hwp, NULL, paddr, rdip, immu_flags)
2210 		    == B_TRUE);
2211 		nppages--;
2212 		nvpages--;
2213 		paddr += IMMU_PAGESIZE;
2214 		dvma += IMMU_PAGESIZE;
2215 
2216 		if (nppages == 0) {
2217 			j++;
2218 		}
2219 
2220 		if (j == dcount) {
2221 			ASSERT(nvpages == 0);
2222 			break;
2223 		}
2224 
2225 		ASSERT(nvpages);
2226 		if (nppages == 0) {
2227 			nppages = dcookies[j].dck_npages;
2228 			paddr = dcookies[j].dck_paddr;
2229 		}
2230 	}
2231 
2232 	/* flush writes to HW PTE table */
2233 	immu_regs_cpu_flush(immu, (caddr_t)shwp, (hwp - shwp) *
2234 	    sizeof (hw_pdte_t));
2235 
2236 	if (nvpages) {
2237 		*dvma_ptr = dvma;
2238 		*nvpages_ptr = nvpages;
2239 	} else {
2240 		*dvma_ptr = 0;
2241 		*nvpages_ptr = 0;
2242 	}
2243 
2244 	xlate->xlt_idx = idx;
2245 }
2246 
2247 /*ARGSUSED*/
2248 static void
2249 PDE_set_one(immu_t *immu, hw_pdte_t *hwp, pgtable_t *next,
2250     dev_info_t *rdip, immu_flags_t immu_flags)
2251 {
2252 	hw_pdte_t pde;
2253 
2254 	pde = *hwp;
2255 
2256 	/* if PDE is already set, make sure it is correct */
2257 	if (PDTE_P(pde)) {
2258 		ASSERT(PDTE_PADDR(pde) == next->hwpg_paddr);
2259 #ifdef BUGGY_DRIVERS
2260 		return;
2261 #else
2262 		goto out;
2263 #endif
2264 	}
2265 
2266 	/* Dont touch SW4, it is the present bit */
2267 
2268 	/* don't touch TM field it is reserved for PDEs */
2269 
2270 	/* 3rd field available for system software is not used */
2271 	PDTE_CLEAR_SW3(pde);
2272 
2273 	/* Set next level pgtable-paddr for PDE */
2274 	ASSERT(next->hwpg_paddr % IMMU_PAGESIZE == 0);
2275 	PDTE_CLEAR_PADDR(pde);
2276 	PDTE_SET_PADDR(pde, next->hwpg_paddr);
2277 
2278 	/* don't touch SNP field it is reserved for PDEs */
2279 
2280 	/* Clear second field available for system software */
2281 	PDTE_CLEAR_SW2(pde);
2282 
2283 	/* No super pages for PDEs */
2284 	PDTE_CLEAR_SP(pde);
2285 
2286 	/* Clear SW1 for software */
2287 	PDTE_CLEAR_SW1(pde);
2288 
2289 	/*
2290 	 * Now that we are done writing the PDE
2291 	 * set the "present" flag. Note this present
2292 	 * flag is a bit in the PDE/PTE that the
2293 	 * spec says is available for system software.
2294 	 * This is an implementation detail of Solaris
2295 	 * base-metal Intel IOMMU.
2296 	 * The present field in a PDE/PTE is not defined
2297 	 * by the Vt-d spec
2298 	 */
2299 
2300 out:
2301 #ifdef  BUGGY_DRIVERS
2302 	PDTE_SET_READ(pde);
2303 	PDTE_SET_WRITE(pde);
2304 #else
2305 	if (immu_flags & IMMU_FLAGS_READ)
2306 		PDTE_SET_READ(pde);
2307 	if (immu_flags & IMMU_FLAGS_WRITE)
2308 		PDTE_SET_WRITE(pde);
2309 #endif
2310 
2311 	PDTE_SET_P(pde);
2312 
2313 	*hwp = pde;
2314 
2315 	immu_regs_cpu_flush(immu, (caddr_t)hwp, sizeof (hw_pdte_t));
2316 }
2317 
2318 /*
2319  * Used to set PDEs
2320  */
2321 static boolean_t
2322 PDE_set_all(immu_t *immu, domain_t *domain, xlate_t *xlate, int nlevels,
2323     dev_info_t *rdip, immu_flags_t immu_flags)
2324 {
2325 	pgtable_t *pgtable;
2326 	pgtable_t *new;
2327 	pgtable_t *next;
2328 	hw_pdte_t *hwp;
2329 	int level;
2330 	uint_t idx;
2331 	krw_t rwtype;
2332 	boolean_t set = B_FALSE;
2333 
2334 	/* xlate should be at level 0 */
2335 	ASSERT(xlate->xlt_level == 0);
2336 	ASSERT(xlate->xlt_idx == 0);
2337 
2338 	/* start with highest level pgtable i.e. root */
2339 	xlate += nlevels;
2340 	ASSERT(xlate->xlt_level == nlevels);
2341 
2342 	new = NULL;
2343 	xlate->xlt_pgtable = domain->dom_pgtable_root;
2344 	for (level = nlevels; level > 1; level--, xlate--) {
2345 
2346 		ASSERT(xlate->xlt_level == level);
2347 
2348 		idx = xlate->xlt_idx;
2349 		pgtable = xlate->xlt_pgtable;
2350 
2351 		ASSERT(pgtable);
2352 		ASSERT(idx <= IMMU_PGTABLE_MAXIDX);
2353 
2354 		/* speculative alloc */
2355 		if (new == NULL) {
2356 			new = pgtable_alloc(immu, immu_flags);
2357 			if (new == NULL) {
2358 				ddi_err(DER_PANIC, rdip, "pgtable alloc err");
2359 			}
2360 		}
2361 
2362 		/* Lock the pgtable in READ mode first */
2363 		rw_enter(&(pgtable->swpg_rwlock), RW_READER);
2364 		rwtype = RW_READER;
2365 again:
2366 		hwp = (hw_pdte_t *)(pgtable->hwpg_vaddr) + idx;
2367 
2368 		ASSERT(pgtable->swpg_next_array);
2369 
2370 		next = (pgtable->swpg_next_array)[idx];
2371 
2372 		/*
2373 		 * check if leafier level already has a pgtable
2374 		 * if yes, verify
2375 		 */
2376 		if (next == NULL) {
2377 			/* Change to a write lock */
2378 			if (rwtype == RW_READER &&
2379 			    rw_tryupgrade(&(pgtable->swpg_rwlock)) == 0) {
2380 				rw_exit(&(pgtable->swpg_rwlock));
2381 				rw_enter(&(pgtable->swpg_rwlock), RW_WRITER);
2382 				rwtype = RW_WRITER;
2383 				goto again;
2384 			}
2385 			rwtype = RW_WRITER;
2386 			pgtable_zero(immu, new);
2387 			next = new;
2388 			new = NULL;
2389 			(pgtable->swpg_next_array)[idx] = next;
2390 			PDE_set_one(immu, hwp, next, rdip, immu_flags);
2391 			set = B_TRUE;
2392 			rw_downgrade(&(pgtable->swpg_rwlock));
2393 			rwtype = RW_READER;
2394 		} else {
2395 			hw_pdte_t pde = *hwp;
2396 
2397 #ifndef  BUGGY_DRIVERS
2398 			/*
2399 			 * If buggy driver we already set permission
2400 			 * READ+WRITE so nothing to do for that case
2401 			 * XXX Check that read writer perms change before
2402 			 * actually setting perms. Also need to hold lock
2403 			 */
2404 			if (immu_flags & IMMU_FLAGS_READ)
2405 				PDTE_SET_READ(pde);
2406 			if (immu_flags & IMMU_FLAGS_WRITE)
2407 				PDTE_SET_WRITE(pde);
2408 
2409 #endif
2410 
2411 			*hwp = pde;
2412 		}
2413 
2414 		ASSERT(PDTE_check(immu, *hwp, next, 0, rdip, immu_flags)
2415 		    == B_TRUE);
2416 
2417 		(xlate - 1)->xlt_pgtable = next;
2418 		ASSERT(rwtype == RW_READER);
2419 		rw_exit(&(pgtable->swpg_rwlock));
2420 	}
2421 
2422 	if (new) {
2423 		pgtable_free(immu, new);
2424 	}
2425 
2426 	return (set);
2427 }
2428 
2429 /*
2430  * dvma_map()
2431  *     map a contiguous range of DVMA pages
2432  *
2433  *     immu: IOMMU unit for which we are generating DVMA cookies
2434  *   domain: domain
2435  *    sdvma: Starting dvma
2436  *   spaddr: Starting paddr
2437  *   npages: Number of pages
2438  *     rdip: requesting device
2439  *     immu_flags: flags
2440  */
2441 static boolean_t
2442 dvma_map(immu_t *immu, domain_t *domain, uint64_t sdvma, uint64_t snvpages,
2443     dcookie_t *dcookies, int dcount, dev_info_t *rdip, immu_flags_t immu_flags)
2444 {
2445 	uint64_t dvma;
2446 	uint64_t n;
2447 	int nlevels = immu->immu_dvma_nlevels;
2448 	xlate_t xlate[IMMU_PGTABLE_MAX_LEVELS + 1] = {0};
2449 	boolean_t pde_set = B_FALSE;
2450 
2451 	ASSERT(nlevels <= IMMU_PGTABLE_MAX_LEVELS);
2452 	ASSERT(sdvma % IMMU_PAGESIZE == 0);
2453 	ASSERT(snvpages);
2454 
2455 	n = snvpages;
2456 	dvma = sdvma;
2457 
2458 	while (n > 0) {
2459 		xlate_setup(immu, dvma, xlate, nlevels, rdip);
2460 
2461 		/* Lookup or allocate PGDIRs and PGTABLEs if necessary */
2462 		if (PDE_set_all(immu, domain, xlate, nlevels, rdip, immu_flags)
2463 		    == B_TRUE) {
2464 			pde_set = B_TRUE;
2465 		}
2466 
2467 		/* set all matching ptes that fit into this leaf pgtable */
2468 		PTE_set_all(immu, domain, &xlate[1], &dvma, &n, dcookies,
2469 		    dcount, rdip, immu_flags);
2470 	}
2471 
2472 	return (pde_set);
2473 }
2474 
2475 /*
2476  * dvma_unmap()
2477  *   unmap a range of DVMAs
2478  *
2479  * immu: IOMMU unit state
2480  * domain: domain for requesting device
2481  * ddip: domain-dip
2482  * dvma: starting DVMA
2483  * npages: Number of IMMU pages to be unmapped
2484  * rdip: requesting device
2485  */
2486 static void
2487 dvma_unmap(immu_t *immu, domain_t *domain, uint64_t sdvma, uint64_t snpages,
2488     dev_info_t *rdip)
2489 {
2490 	int nlevels = immu->immu_dvma_nlevels;
2491 	xlate_t xlate[IMMU_PGTABLE_MAX_LEVELS + 1] = {0};
2492 	uint64_t n;
2493 	uint64_t dvma;
2494 
2495 	ASSERT(nlevels <= IMMU_PGTABLE_MAX_LEVELS);
2496 	ASSERT(sdvma != 0);
2497 	ASSERT(sdvma % IMMU_PAGESIZE == 0);
2498 	ASSERT(snpages);
2499 
2500 	dvma = sdvma;
2501 	n = snpages;
2502 
2503 	while (n > 0) {
2504 		/* setup the xlate array */
2505 		xlate_setup(immu, dvma, xlate, nlevels, rdip);
2506 
2507 		/* just lookup existing pgtables. Should never fail */
2508 		PDE_lookup(immu, domain, xlate, nlevels, rdip);
2509 
2510 		/* clear all matching ptes that fit into this leaf pgtable */
2511 		PTE_clear_all(immu, domain, &xlate[1], &dvma, &n, rdip);
2512 	}
2513 
2514 	/* No need to flush IOTLB after unmap */
2515 }
2516 
2517 static uint64_t
2518 dvma_alloc(ddi_dma_impl_t *hp, domain_t *domain, uint_t npages)
2519 {
2520 	ddi_dma_attr_t *dma_attr;
2521 	uint64_t dvma;
2522 	size_t xsize, align;
2523 	uint64_t minaddr, maxaddr;
2524 
2525 	ASSERT(domain->dom_maptype != IMMU_MAPTYPE_UNITY);
2526 
2527 	/* shotcuts */
2528 	dma_attr = &(hp->dmai_attr);
2529 
2530 	/* parameters */
2531 	xsize = npages * IMMU_PAGESIZE;
2532 	align = MAX((size_t)(dma_attr->dma_attr_align), IMMU_PAGESIZE);
2533 	minaddr = dma_attr->dma_attr_addr_lo;
2534 	maxaddr = dma_attr->dma_attr_addr_hi + 1;
2535 	/* nocross is checked in cookie_update() */
2536 
2537 	/* handle the rollover cases */
2538 	if (maxaddr < dma_attr->dma_attr_addr_hi) {
2539 		maxaddr = dma_attr->dma_attr_addr_hi;
2540 	}
2541 
2542 	/*
2543 	 * allocate from vmem arena.
2544 	 */
2545 	dvma = (uint64_t)(uintptr_t)vmem_xalloc(domain->dom_dvma_arena,
2546 	    xsize, align, 0, 0, (void *)(uintptr_t)minaddr,
2547 	    (void *)(uintptr_t)maxaddr, VM_NOSLEEP);
2548 
2549 	ASSERT(dvma);
2550 	ASSERT(dvma >= minaddr);
2551 	ASSERT(dvma + xsize - 1 < maxaddr);
2552 
2553 	return (dvma);
2554 }
2555 
2556 static void
2557 dvma_free(domain_t *domain, uint64_t dvma, uint64_t npages)
2558 {
2559 	uint64_t size = npages * IMMU_PAGESIZE;
2560 
2561 	ASSERT(domain);
2562 	ASSERT(domain->dom_did > 0);
2563 	ASSERT(dvma);
2564 	ASSERT(npages);
2565 
2566 	if (domain->dom_maptype != IMMU_MAPTYPE_XLATE) {
2567 		ASSERT(domain->dom_maptype == IMMU_MAPTYPE_UNITY);
2568 		return;
2569 	}
2570 
2571 	vmem_free(domain->dom_dvma_arena, (void *)(uintptr_t)dvma, size);
2572 }
2573 /*ARGSUSED*/
2574 static void
2575 cookie_free(rootnex_dma_t *dma, immu_t *immu, domain_t *domain,
2576     dev_info_t *rdip)
2577 {
2578 	int i;
2579 	uint64_t dvma;
2580 	uint64_t npages;
2581 	dvcookie_t  *dvcookies = dma->dp_dvcookies;
2582 
2583 	ASSERT(dma->dp_max_cookies);
2584 	ASSERT(dma->dp_max_dcookies);
2585 	ASSERT(dma->dp_dvmax < dma->dp_max_cookies);
2586 	ASSERT(dma->dp_dmax < dma->dp_max_dcookies);
2587 
2588 	/*
2589 	 * we allocated DVMA in a single chunk. Calculate total number
2590 	 * of pages
2591 	 */
2592 	for (i = 0, npages = 0; i <= dma->dp_dvmax; i++) {
2593 		npages += dvcookies[i].dvck_npages;
2594 	}
2595 	dvma = dvcookies[0].dvck_dvma;
2596 #ifdef DEBUG
2597 	/* Unmap only in DEBUG mode */
2598 	dvma_unmap(immu, domain, dvma, npages, rdip);
2599 #endif
2600 	dvma_free(domain, dvma, npages);
2601 
2602 	kmem_free(dma->dp_dvcookies, sizeof (dvcookie_t) * dma->dp_max_cookies);
2603 	dma->dp_dvcookies = NULL;
2604 	kmem_free(dma->dp_dcookies, sizeof (dcookie_t) * dma->dp_max_dcookies);
2605 	dma->dp_dcookies = NULL;
2606 	if (dma->dp_need_to_free_cookie == B_TRUE) {
2607 		kmem_free(dma->dp_cookies, sizeof (ddi_dma_cookie_t) *
2608 		    dma->dp_max_cookies);
2609 		dma->dp_dcookies = NULL;
2610 		dma->dp_need_to_free_cookie = B_FALSE;
2611 	}
2612 
2613 	dma->dp_max_cookies = 0;
2614 	dma->dp_max_dcookies = 0;
2615 	dma->dp_cookie_size = 0;
2616 	dma->dp_dvmax = 0;
2617 	dma->dp_dmax = 0;
2618 }
2619 
2620 /*
2621  * cookie_alloc()
2622  */
2623 static int
2624 cookie_alloc(rootnex_dma_t *dma, struct ddi_dma_req *dmareq,
2625     ddi_dma_attr_t *attr, uint_t prealloc)
2626 {
2627 	int kmflag;
2628 	rootnex_sglinfo_t *sinfo = &(dma->dp_sglinfo);
2629 	dvcookie_t *dvcookies = dma->dp_dvcookies;
2630 	dcookie_t *dcookies = dma->dp_dcookies;
2631 	ddi_dma_cookie_t *cookies = dma->dp_cookies;
2632 	uint64_t max_cookies;
2633 	uint64_t max_dcookies;
2634 	uint64_t cookie_size;
2635 
2636 	/* we need to allocate new array */
2637 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
2638 		kmflag =  KM_SLEEP;
2639 	} else {
2640 		kmflag =  KM_NOSLEEP;
2641 	}
2642 
2643 	/*
2644 	 * XXX make sure cookies size doen't exceed sinfo->si_max_cookie_size;
2645 	 */
2646 
2647 	/*
2648 	 * figure out the rough estimate of array size
2649 	 * At a minimum, each cookie must hold 1 page.
2650 	 * At a maximum, it cannot exceed dma_attr_sgllen
2651 	 */
2652 	max_dcookies = dmareq->dmar_object.dmao_size + IMMU_PAGEOFFSET;
2653 	max_dcookies /= IMMU_PAGESIZE;
2654 	max_dcookies++;
2655 	max_cookies = MIN(max_dcookies, attr->dma_attr_sgllen);
2656 
2657 	/* allocate the dvma cookie array */
2658 	dvcookies = kmem_zalloc(sizeof (dvcookie_t) * max_cookies, kmflag);
2659 	if (dvcookies == NULL) {
2660 		return (DDI_FAILURE);
2661 	}
2662 
2663 	/* allocate the "phys" cookie array */
2664 	dcookies = kmem_zalloc(sizeof (dcookie_t) * max_dcookies, kmflag);
2665 	if (dcookies == NULL) {
2666 		kmem_free(dvcookies, sizeof (dvcookie_t) * max_cookies);
2667 		dvcookies = NULL;
2668 		return (DDI_FAILURE);
2669 	}
2670 
2671 	/* allocate the "real" cookie array  - the one given to users */
2672 	cookie_size = sizeof (ddi_dma_cookie_t) * max_cookies;
2673 	if (max_cookies > prealloc) {
2674 		cookies = kmem_zalloc(cookie_size, kmflag);
2675 		if (cookies == NULL) {
2676 			kmem_free(dvcookies, sizeof (dvcookie_t) * max_cookies);
2677 			kmem_free(dcookies, sizeof (dcookie_t) * max_dcookies);
2678 			goto fail;
2679 		}
2680 		dma->dp_need_to_free_cookie = B_TRUE;
2681 	} else {
2682 		/* the preallocated buffer fits this size */
2683 		cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2684 		bzero(cookies, sizeof (ddi_dma_cookie_t)* max_cookies);
2685 		dma->dp_need_to_free_cookie = B_FALSE;
2686 	}
2687 
2688 	dma->dp_dvcookies = dvcookies;
2689 	dma->dp_dcookies = dcookies;
2690 	dma->dp_cookies = cookies;
2691 	dma->dp_cookie_size = cookie_size;
2692 	dma->dp_max_cookies = max_cookies;
2693 	dma->dp_max_dcookies = max_dcookies;
2694 	dma->dp_dvmax = 0;
2695 	dma->dp_dmax = 0;
2696 	sinfo->si_max_pages = dma->dp_max_cookies;
2697 
2698 	return (DDI_SUCCESS);
2699 
2700 fail:
2701 	dma->dp_dvcookies = NULL;
2702 	dma->dp_dcookies = NULL;
2703 	dma->dp_cookies = NULL;
2704 	dma->dp_cookie_size = 0;
2705 	dma->dp_max_cookies = 0;
2706 	dma->dp_max_dcookies = 0;
2707 	dma->dp_dvmax = 0;
2708 	dma->dp_dmax = 0;
2709 	dma->dp_need_to_free_cookie = B_FALSE;
2710 	sinfo->si_max_pages = 0;
2711 
2712 	return (DDI_FAILURE);
2713 }
2714 
2715 /*ARGSUSED*/
2716 static void
2717 cookie_update(domain_t *domain, rootnex_dma_t *dma, paddr_t paddr,
2718     int64_t psize, uint64_t maxseg, size_t nocross)
2719 {
2720 	dvcookie_t *dvcookies = dma->dp_dvcookies;
2721 	dcookie_t *dcookies = dma->dp_dcookies;
2722 	ddi_dma_cookie_t *cookies = dma->dp_cookies;
2723 	uint64_t dvmax = dma->dp_dvmax;
2724 	uint64_t dmax = dma->dp_dmax;
2725 
2726 	ASSERT(dvmax < dma->dp_max_cookies);
2727 	ASSERT(dmax < dma->dp_max_dcookies);
2728 
2729 	paddr &= IMMU_PAGEMASK;
2730 
2731 	ASSERT(paddr);
2732 	ASSERT(psize);
2733 	ASSERT(maxseg);
2734 
2735 	/*
2736 	 * check to see if this page would put us
2737 	 * over the max cookie size.
2738 	 */
2739 	if (cookies[dvmax].dmac_size + psize > maxseg) {
2740 		dvmax++;    /* use the next dvcookie */
2741 		dmax++;    /* also means we use the next dcookie */
2742 		ASSERT(dvmax < dma->dp_max_cookies);
2743 		ASSERT(dmax < dma->dp_max_dcookies);
2744 	}
2745 
2746 	/*
2747 	 * check to see if this page would make us larger than
2748 	 * the nocross boundary. If yes, create a new cookie
2749 	 * otherwise we will fail later with vmem_xalloc()
2750 	 * due to overconstrained alloc requests
2751 	 * nocross == 0 implies no nocross constraint.
2752 	 */
2753 	if (nocross > 0) {
2754 		ASSERT((dvcookies[dvmax].dvck_npages) * IMMU_PAGESIZE
2755 		    <= nocross);
2756 		if ((dvcookies[dvmax].dvck_npages + 1) * IMMU_PAGESIZE
2757 		    > nocross) {
2758 			dvmax++;    /* use the next dvcookie */
2759 			dmax++;    /* also means we use the next dcookie */
2760 			ASSERT(dvmax < dma->dp_max_cookies);
2761 			ASSERT(dmax < dma->dp_max_dcookies);
2762 		}
2763 		ASSERT((dvcookies[dvmax].dvck_npages) * IMMU_PAGESIZE
2764 		    <= nocross);
2765 	}
2766 
2767 	/*
2768 	 * If the cookie is empty
2769 	 */
2770 	if (dvcookies[dvmax].dvck_npages == 0) {
2771 		ASSERT(cookies[dvmax].dmac_size == 0);
2772 		ASSERT(dvcookies[dvmax].dvck_dvma == 0);
2773 		ASSERT(dvcookies[dvmax].dvck_npages
2774 		    == 0);
2775 		ASSERT(dcookies[dmax].dck_paddr == 0);
2776 		ASSERT(dcookies[dmax].dck_npages == 0);
2777 
2778 		dvcookies[dvmax].dvck_dvma = 0;
2779 		dvcookies[dvmax].dvck_npages = 1;
2780 		dcookies[dmax].dck_paddr = paddr;
2781 		dcookies[dmax].dck_npages = 1;
2782 		cookies[dvmax].dmac_size = psize;
2783 	} else {
2784 		/* Cookie not empty. Add to it */
2785 		cookies[dma->dp_dvmax].dmac_size += psize;
2786 		ASSERT(dvcookies[dma->dp_dvmax].dvck_dvma == 0);
2787 		dvcookies[dma->dp_dvmax].dvck_npages++;
2788 		ASSERT(dcookies[dmax].dck_paddr != 0);
2789 		ASSERT(dcookies[dmax].dck_npages != 0);
2790 
2791 		/* Check if this paddr is contiguous */
2792 		if (IMMU_CONTIG_PADDR(dcookies[dmax], paddr)) {
2793 			dcookies[dmax].dck_npages++;
2794 		} else {
2795 			/* No, we need a new dcookie */
2796 			dmax++;
2797 			ASSERT(dcookies[dmax].dck_paddr == 0);
2798 			ASSERT(dcookies[dmax].dck_npages == 0);
2799 			dcookies[dmax].dck_paddr = paddr;
2800 			dcookies[dmax].dck_npages = 1;
2801 		}
2802 	}
2803 
2804 	dma->dp_dvmax = dvmax;
2805 	dma->dp_dmax = dmax;
2806 }
2807 
2808 static void
2809 cookie_finalize(ddi_dma_impl_t *hp, immu_t *immu, domain_t *domain,
2810     dev_info_t *rdip, immu_flags_t immu_flags)
2811 {
2812 	int i;
2813 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2814 	dvcookie_t *dvcookies = dma->dp_dvcookies;
2815 	dcookie_t *dcookies = dma->dp_dcookies;
2816 	ddi_dma_cookie_t *cookies = dma->dp_cookies;
2817 	uint64_t npages;
2818 	uint64_t dvma;
2819 	boolean_t pde_set;
2820 
2821 	/* First calculate the total number of pages required */
2822 	for (i = 0, npages = 0; i <= dma->dp_dvmax; i++) {
2823 		npages += dvcookies[i].dvck_npages;
2824 	}
2825 
2826 	/* Now allocate dvma */
2827 	dvma = dvma_alloc(hp, domain, npages);
2828 
2829 	/* Now map the dvma */
2830 	pde_set = dvma_map(immu, domain, dvma, npages, dcookies,
2831 	    dma->dp_dmax + 1, rdip, immu_flags);
2832 
2833 	/* Invalidate the IOTLB */
2834 	immu_regs_iotlb_flush(immu, domain->dom_did, dvma, npages,
2835 	    pde_set == B_TRUE ? TLB_IVA_WHOLE : TLB_IVA_LEAF, IOTLB_PSI);
2836 
2837 	/* Now setup dvcookies and real cookie addresses */
2838 	for (i = 0; i <= dma->dp_dvmax; i++) {
2839 		dvcookies[i].dvck_dvma = dvma;
2840 		cookies[i].dmac_laddress = dvma;
2841 		ASSERT(cookies[i].dmac_size != 0);
2842 		cookies[i].dmac_type = 0;
2843 		dvma += (dvcookies[i].dvck_npages * IMMU_PAGESIZE);
2844 	}
2845 
2846 #ifdef TEST
2847 	immu_regs_iotlb_flush(immu, domain->dom_did, 0, 0, 0, IOTLB_DSI);
2848 #endif
2849 }
2850 
2851 /*
2852  * cookie_create()
2853  */
2854 static int
2855 cookie_create(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
2856     ddi_dma_attr_t *a, immu_t *immu, domain_t *domain, dev_info_t *rdip,
2857     uint_t prealloc_count, immu_flags_t immu_flags)
2858 {
2859 	ddi_dma_atyp_t buftype;
2860 	uint64_t offset;
2861 	page_t **pparray;
2862 	uint64_t paddr;
2863 	uint_t psize;
2864 	uint_t size;
2865 	uint64_t maxseg;
2866 	caddr_t vaddr;
2867 	uint_t pcnt;
2868 	page_t *page;
2869 	rootnex_sglinfo_t *sglinfo;
2870 	ddi_dma_obj_t *dmar_object;
2871 	rootnex_dma_t *dma;
2872 	size_t nocross;
2873 
2874 	dma = (rootnex_dma_t *)hp->dmai_private;
2875 	sglinfo = &(dma->dp_sglinfo);
2876 	dmar_object = &(dmareq->dmar_object);
2877 	maxseg = sglinfo->si_max_cookie_size;
2878 	pparray = dmar_object->dmao_obj.virt_obj.v_priv;
2879 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2880 	buftype = dmar_object->dmao_type;
2881 	size = dmar_object->dmao_size;
2882 	nocross = (size_t)(a->dma_attr_seg + 1);
2883 
2884 	/*
2885 	 * Allocate cookie, dvcookie and dcookie
2886 	 */
2887 	if (cookie_alloc(dma, dmareq, a, prealloc_count) != DDI_SUCCESS) {
2888 		return (DDI_FAILURE);
2889 	}
2890 	hp->dmai_cookie = dma->dp_cookies;
2891 
2892 	pcnt = 0;
2893 
2894 	/* retrieve paddr, psize, offset from dmareq */
2895 	if (buftype == DMA_OTYP_PAGES) {
2896 		page = dmar_object->dmao_obj.pp_obj.pp_pp;
2897 		ASSERT(!PP_ISFREE(page) && PAGE_LOCKED(page));
2898 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2899 		    MMU_PAGEOFFSET;
2900 		paddr = pfn_to_pa(page->p_pagenum) + offset;
2901 		psize = MIN((MMU_PAGESIZE - offset), size);
2902 		sglinfo->si_asp = NULL;
2903 		page = page->p_next;
2904 	} else {
2905 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2906 		    (buftype == DMA_OTYP_BUFVADDR));
2907 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2908 		if (sglinfo->si_asp == NULL) {
2909 			sglinfo->si_asp = &kas;
2910 		}
2911 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2912 		if (pparray != NULL) {
2913 			ASSERT(!PP_ISFREE(pparray[pcnt]));
2914 			paddr = pfn_to_pa(pparray[pcnt]->p_pagenum) + offset;
2915 			psize = MIN((MMU_PAGESIZE - offset), size);
2916 			pcnt++;
2917 		} else {
2918 			paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2919 			    vaddr)) + offset;
2920 			psize = MIN(size, (MMU_PAGESIZE - offset));
2921 			vaddr += psize;
2922 		}
2923 	}
2924 
2925 	/* save the iommu page offset */
2926 	sglinfo->si_buf_offset = offset & IMMU_PAGEOFFSET;
2927 
2928 	/*
2929 	 * setup dvcookie and dcookie for [paddr, paddr+psize)
2930 	 */
2931 	cookie_update(domain, dma, paddr, psize, maxseg, nocross);
2932 
2933 	size -= psize;
2934 	while (size > 0) {
2935 		/* get the size for this page (i.e. partial or full page) */
2936 		psize = MIN(size, MMU_PAGESIZE);
2937 		if (buftype == DMA_OTYP_PAGES) {
2938 			/* get the paddr from the page_t */
2939 			ASSERT(!PP_ISFREE(page) && PAGE_LOCKED(page));
2940 			paddr = pfn_to_pa(page->p_pagenum);
2941 			page = page->p_next;
2942 		} else if (pparray != NULL) {
2943 			/* index into the array of page_t's to get the paddr */
2944 			ASSERT(!PP_ISFREE(pparray[pcnt]));
2945 			paddr = pfn_to_pa(pparray[pcnt]->p_pagenum);
2946 			pcnt++;
2947 		} else {
2948 			/* call into the VM to get the paddr */
2949 			paddr = pfn_to_pa(hat_getpfnum
2950 			    (sglinfo->si_asp->a_hat, vaddr));
2951 			vaddr += psize;
2952 		}
2953 		/*
2954 		 * set dvcookie and dcookie for [paddr, paddr+psize)
2955 		 */
2956 		cookie_update(domain, dma, paddr, psize, maxseg, nocross);
2957 		size -= psize;
2958 	}
2959 
2960 	cookie_finalize(hp, immu, domain, rdip, immu_flags);
2961 
2962 	/* take account in the offset into the first page */
2963 	dma->dp_cookies[0].dmac_laddress += sglinfo->si_buf_offset;
2964 
2965 	/* save away how many cookies we have */
2966 	sglinfo->si_sgl_size = dma->dp_dvmax + 1;
2967 
2968 	return (DDI_SUCCESS);
2969 }
2970 
2971 /* ############################# Functions exported ######################## */
2972 
2973 /*
2974  * setup the DVMA subsystem
2975  * this code runs only for the first IOMMU unit
2976  */
2977 void
2978 immu_dvma_setup(list_t *listp)
2979 {
2980 	immu_t *immu;
2981 	uint_t kval;
2982 	size_t nchains;
2983 
2984 	/* locks */
2985 	mutex_init(&immu_domain_lock, NULL, MUTEX_DEFAULT, NULL);
2986 
2987 	/* Create lists */
2988 	list_create(&immu_unity_domain_list, sizeof (domain_t),
2989 	    offsetof(domain_t, dom_maptype_node));
2990 	list_create(&immu_xlate_domain_list, sizeof (domain_t),
2991 	    offsetof(domain_t, dom_maptype_node));
2992 
2993 	/* Setup BDF domain hash */
2994 	nchains = 0xff;
2995 	kval = mod_hash_iddata_gen(nchains);
2996 
2997 	bdf_domain_hash = mod_hash_create_extended("BDF-DOMAIN_HASH",
2998 	    nchains, mod_hash_null_keydtor, mod_hash_null_valdtor,
2999 	    mod_hash_byid, (void *)(uintptr_t)kval, mod_hash_idkey_cmp,
3000 	    KM_NOSLEEP);
3001 	ASSERT(bdf_domain_hash);
3002 
3003 	immu = list_head(listp);
3004 	for (; immu; immu = list_next(listp, immu)) {
3005 		create_unity_domain(immu);
3006 		did_init(immu);
3007 		context_init(immu);
3008 		immu->immu_dvma_setup = B_TRUE;
3009 	}
3010 }
3011 
3012 /*
3013  * Startup up one DVMA unit
3014  */
3015 void
3016 immu_dvma_startup(immu_t *immu)
3017 {
3018 	ASSERT(immu);
3019 	ASSERT(immu->immu_dvma_running == B_FALSE);
3020 
3021 	if (immu_gfxdvma_enable == B_FALSE &&
3022 	    immu->immu_dvma_gfx_only == B_TRUE) {
3023 		return;
3024 	}
3025 
3026 	/*
3027 	 * DVMA will start once IOMMU is "running"
3028 	 */
3029 	ASSERT(immu->immu_dvma_running == B_FALSE);
3030 	immu->immu_dvma_running = B_TRUE;
3031 }
3032 
3033 /*
3034  * immu_dvma_physmem_update()
3035  *       called when the installed memory on a
3036  *       system increases, to expand domain DVMA
3037  *       for domains with UNITY mapping
3038  */
3039 void
3040 immu_dvma_physmem_update(uint64_t addr, uint64_t size)
3041 {
3042 	uint64_t start;
3043 	uint64_t npages;
3044 	int dcount;
3045 	dcookie_t dcookies[1] = {0};
3046 	domain_t *domain;
3047 
3048 	/*
3049 	 * Just walk the system-wide list of domains with
3050 	 * UNITY mapping. Both the list of *all* domains
3051 	 * and *UNITY* domains is protected by the same
3052 	 * single lock
3053 	 */
3054 	mutex_enter(&immu_domain_lock);
3055 	domain = list_head(&immu_unity_domain_list);
3056 	for (; domain; domain = list_next(&immu_unity_domain_list, domain)) {
3057 
3058 		/* There is no vmem_arena for unity domains. Just map it */
3059 		ddi_err(DER_LOG, NULL, "IMMU: unity-domain: Adding map "
3060 		    "[0x%" PRIx64 " - 0x%" PRIx64 "]", addr, addr + size);
3061 
3062 		start = IMMU_ROUNDOWN(addr);
3063 		npages = (IMMU_ROUNDUP(size) / IMMU_PAGESIZE) + 1;
3064 
3065 		dcookies[0].dck_paddr = start;
3066 		dcookies[0].dck_npages = npages;
3067 		dcount = 1;
3068 		(void) dvma_map(domain->dom_immu, domain, start, npages,
3069 		    dcookies, dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE);
3070 
3071 	}
3072 	mutex_exit(&immu_domain_lock);
3073 }
3074 
3075 
3076 int
3077 immu_dvma_map(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq, memrng_t *mrng,
3078     uint_t prealloc_count, dev_info_t *rdip, immu_flags_t immu_flags)
3079 {
3080 	ddi_dma_attr_t *attr;
3081 	dev_info_t *ddip;
3082 	domain_t *domain;
3083 	immu_t *immu;
3084 	dcookie_t dcookies[1] = {0};
3085 	int dcount = 0;
3086 	boolean_t pde_set = B_TRUE;
3087 	int r = DDI_FAILURE;
3088 
3089 	ASSERT(immu_enable == B_TRUE);
3090 	ASSERT(immu_running == B_TRUE || !(immu_flags & IMMU_FLAGS_DMAHDL));
3091 	ASSERT(hp || !(immu_flags & IMMU_FLAGS_DMAHDL));
3092 
3093 	/*
3094 	 * Intel IOMMU will only be turned on if IOMMU
3095 	 * page size is a multiple of IOMMU page size
3096 	 */
3097 
3098 	/*LINTED*/
3099 	ASSERT(MMU_PAGESIZE % IMMU_PAGESIZE == 0);
3100 
3101 	/* Can only do DVMA if dip is attached */
3102 	if (rdip == NULL) {
3103 		ddi_err(DER_PANIC, rdip, "DVMA map: No device specified");
3104 		/*NOTREACHED*/
3105 	}
3106 
3107 	immu_flags |= dma_to_immu_flags(dmareq);
3108 
3109 
3110 	immu = immu_dvma_get_immu(rdip, immu_flags);
3111 	if (immu == NULL) {
3112 		/*
3113 		 * possible that there is no IOMMU unit for this device
3114 		 * - BIOS bugs are one example.
3115 		 */
3116 		ddi_err(DER_WARN, rdip, "No IMMU unit found for device");
3117 		return (DDI_DMA_NORESOURCES);
3118 	}
3119 
3120 
3121 	/*
3122 	 * redirect isa devices attached under lpc to lpc dip
3123 	 */
3124 	if (strcmp(ddi_node_name(ddi_get_parent(rdip)), "isa") == 0) {
3125 		rdip = get_lpc_devinfo(immu, rdip, immu_flags);
3126 		if (rdip == NULL) {
3127 			ddi_err(DER_PANIC, rdip, "IMMU redirect failed");
3128 			/*NOTREACHED*/
3129 		}
3130 	}
3131 
3132 	/* Reset immu, as redirection can change IMMU */
3133 	immu = NULL;
3134 
3135 	/*
3136 	 * for gart, redirect to the real graphic devinfo
3137 	 */
3138 	if (strcmp(ddi_node_name(rdip), "agpgart") == 0) {
3139 		rdip = get_gfx_devinfo(rdip);
3140 		if (rdip == NULL) {
3141 			ddi_err(DER_PANIC, rdip, "IMMU redirect failed");
3142 			/*NOTREACHED*/
3143 		}
3144 	}
3145 
3146 	/*
3147 	 * Setup DVMA domain for the device. This does
3148 	 * work only the first time we do DVMA for a
3149 	 * device.
3150 	 */
3151 	ddip = NULL;
3152 	domain = device_domain(rdip, &ddip, immu_flags);
3153 	if (domain == NULL) {
3154 		ASSERT(ddip == NULL);
3155 		ddi_err(DER_MODE, rdip, "Intel IOMMU setup failed for device");
3156 		return (DDI_DMA_NORESOURCES);
3157 	}
3158 
3159 	/*
3160 	 * If a domain is found, we must also have a domain dip
3161 	 * which is the topmost ancestor dip of rdip that shares
3162 	 * the same domain with rdip.
3163 	 */
3164 	if (domain->dom_did == 0 || ddip == NULL) {
3165 		ddi_err(DER_MODE, rdip, "domain did 0(%d) or ddip NULL(%p)",
3166 		    domain->dom_did, ddip);
3167 		return (DDI_DMA_NORESOURCES);
3168 	}
3169 
3170 	immu = domain->dom_immu;
3171 	ASSERT(immu);
3172 	if (domain->dom_did == IMMU_UNITY_DID) {
3173 		ASSERT(domain == immu->immu_unity_domain);
3174 		/* mapping already done. Let rootnex create cookies */
3175 		r = DDI_DMA_USE_PHYSICAL;
3176 	} else  if (immu_flags & IMMU_FLAGS_DMAHDL) {
3177 
3178 		/* if we have a DMA handle, the IOMMUs must be running */
3179 		ASSERT(immu->immu_regs_running == B_TRUE);
3180 		ASSERT(immu->immu_dvma_running == B_TRUE);
3181 
3182 		attr = &hp->dmai_attr;
3183 		if (attr == NULL) {
3184 			ddi_err(DER_PANIC, rdip,
3185 			    "DMA handle (%p): NULL attr", hp);
3186 			/*NOTREACHED*/
3187 		}
3188 
3189 		if (cookie_create(hp, dmareq, attr, immu, domain, rdip,
3190 		    prealloc_count, immu_flags) != DDI_SUCCESS) {
3191 			ddi_err(DER_MODE, rdip, "dvcookie_alloc: failed");
3192 			return (DDI_DMA_NORESOURCES);
3193 		}
3194 		r = DDI_DMA_MAPPED;
3195 	} else if (immu_flags & IMMU_FLAGS_MEMRNG) {
3196 		dcookies[0].dck_paddr = mrng->mrng_start;
3197 		dcookies[0].dck_npages = mrng->mrng_npages;
3198 		dcount = 1;
3199 		pde_set = dvma_map(immu, domain, mrng->mrng_start,
3200 		    mrng->mrng_npages, dcookies, dcount, rdip, immu_flags);
3201 		immu_regs_iotlb_flush(immu, domain->dom_did, mrng->mrng_start,
3202 		    mrng->mrng_npages, pde_set == B_TRUE ?
3203 		    TLB_IVA_WHOLE : TLB_IVA_LEAF, IOTLB_PSI);
3204 		r = DDI_DMA_MAPPED;
3205 	} else {
3206 		ddi_err(DER_PANIC, rdip, "invalid flags for immu_dvma_map()");
3207 		/*NOTREACHED*/
3208 	}
3209 
3210 	/*
3211 	 * Update the root and context entries
3212 	 */
3213 	if (immu_context_update(immu, domain, ddip, rdip, immu_flags)
3214 	    != DDI_SUCCESS) {
3215 		ddi_err(DER_MODE, rdip, "DVMA map: context update failed");
3216 		return (DDI_DMA_NORESOURCES);
3217 	}
3218 
3219 	immu_regs_wbf_flush(immu);
3220 
3221 	return (r);
3222 }
3223 
3224 int
3225 immu_dvma_unmap(ddi_dma_impl_t *hp, dev_info_t *rdip)
3226 {
3227 	ddi_dma_attr_t *attr;
3228 	rootnex_dma_t *dma;
3229 	domain_t *domain;
3230 	immu_t *immu;
3231 	dev_info_t *ddip;
3232 	immu_flags_t immu_flags;
3233 
3234 	ASSERT(immu_enable == B_TRUE);
3235 	ASSERT(immu_running == B_TRUE);
3236 	ASSERT(hp);
3237 
3238 	/*
3239 	 * Intel IOMMU will only be turned on if IOMMU
3240 	 * page size is same as MMU page size
3241 	 */
3242 	/*LINTED*/
3243 	ASSERT(MMU_PAGESIZE == IMMU_PAGESIZE);
3244 
3245 	/* rdip need not be attached */
3246 	if (rdip == NULL) {
3247 		ddi_err(DER_PANIC, rdip, "DVMA unmap: No device specified");
3248 		return (DDI_DMA_NORESOURCES);
3249 	}
3250 
3251 	/*
3252 	 * Get the device domain, this should always
3253 	 * succeed since there had to be a domain to
3254 	 * setup DVMA.
3255 	 */
3256 	dma = (rootnex_dma_t *)hp->dmai_private;
3257 	attr = &hp->dmai_attr;
3258 	if (attr == NULL) {
3259 		ddi_err(DER_PANIC, rdip, "DMA handle (%p) has NULL attr", hp);
3260 		/*NOTREACHED*/
3261 	}
3262 	immu_flags = dma->dp_sleep_flags;
3263 
3264 	immu = immu_dvma_get_immu(rdip, immu_flags);
3265 	if (immu == NULL) {
3266 		/*
3267 		 * possible that there is no IOMMU unit for this device
3268 		 * - BIOS bugs are one example.
3269 		 */
3270 		ddi_err(DER_WARN, rdip, "No IMMU unit found for device");
3271 		return (DDI_DMA_NORESOURCES);
3272 	}
3273 
3274 
3275 	/*
3276 	 * redirect isa devices attached under lpc to lpc dip
3277 	 */
3278 	if (strcmp(ddi_node_name(ddi_get_parent(rdip)), "isa") == 0) {
3279 		rdip = get_lpc_devinfo(immu, rdip, immu_flags);
3280 		if (rdip == NULL) {
3281 			ddi_err(DER_PANIC, rdip, "IMMU redirect failed");
3282 			/*NOTREACHED*/
3283 		}
3284 	}
3285 
3286 	/* Reset immu, as redirection can change IMMU */
3287 	immu = NULL;
3288 
3289 	/*
3290 	 * for gart, redirect to the real graphic devinfo
3291 	 */
3292 	if (strcmp(ddi_node_name(rdip), "agpgart") == 0) {
3293 		rdip = get_gfx_devinfo(rdip);
3294 		if (rdip == NULL) {
3295 			ddi_err(DER_PANIC, rdip, "IMMU redirect failed");
3296 			/*NOTREACHED*/
3297 		}
3298 	}
3299 
3300 	ddip = NULL;
3301 	domain = device_domain(rdip, &ddip, immu_flags);
3302 	if (domain == NULL || domain->dom_did == 0 || ddip == NULL) {
3303 		ddi_err(DER_MODE, rdip, "Attempt to unmap DVMA for "
3304 		    "a device without domain or with an uninitialized "
3305 		    "domain");
3306 		return (DDI_DMA_NORESOURCES);
3307 	}
3308 
3309 	/*
3310 	 * immu must be set in the domain.
3311 	 */
3312 	immu = domain->dom_immu;
3313 	ASSERT(immu);
3314 	if (domain->dom_did == IMMU_UNITY_DID) {
3315 		ASSERT(domain == immu->immu_unity_domain);
3316 		/*
3317 		 * domain is unity, nothing to do here, let the rootnex
3318 		 * code free the cookies.
3319 		 */
3320 		return (DDI_DMA_USE_PHYSICAL);
3321 	}
3322 
3323 	dma = hp->dmai_private;
3324 	if (dma == NULL) {
3325 		ddi_err(DER_PANIC, rdip, "DVMA unmap: DMA handle (%p) has "
3326 		    "no private dma structure", hp);
3327 		/*NOTREACHED*/
3328 	}
3329 
3330 	cookie_free(dma, immu, domain, rdip);
3331 
3332 	/* No invalidation needed for unmap */
3333 	immu_regs_wbf_flush(immu);
3334 
3335 	return (DDI_SUCCESS);
3336 }
3337 
3338 immu_devi_t *
3339 immu_devi_get(dev_info_t *rdip)
3340 {
3341 	immu_devi_t *immu_devi;
3342 	volatile uintptr_t *vptr = (uintptr_t *)&(DEVI(rdip)->devi_iommu);
3343 
3344 	/* Just want atomic reads. No need for lock */
3345 	immu_devi = (immu_devi_t *)(uintptr_t)atomic_or_64_nv((uint64_t *)vptr,
3346 	    0);
3347 	return (immu_devi);
3348 }
3349