xref: /illumos-gate/usr/src/uts/i86pc/io/rootnex.c (revision f9c7ecc9013918bf789e154e2ed5f2e8fb6ce24c)
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  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
26  * Copyright (c) 2011 Bayard G. Bell.  All rights reserved.
27  * Copyright 2012 Garrett D'Amore <garrett@damore.org>.  All rights reserved.
28  * Copyright 2017 Joyent, Inc.
29  */
30 
31 /*
32  * x86 root nexus driver
33  */
34 
35 #include <sys/sysmacros.h>
36 #include <sys/conf.h>
37 #include <sys/autoconf.h>
38 #include <sys/sysmacros.h>
39 #include <sys/debug.h>
40 #include <sys/psw.h>
41 #include <sys/ddidmareq.h>
42 #include <sys/promif.h>
43 #include <sys/devops.h>
44 #include <sys/kmem.h>
45 #include <sys/cmn_err.h>
46 #include <vm/seg.h>
47 #include <vm/seg_kmem.h>
48 #include <vm/seg_dev.h>
49 #include <sys/vmem.h>
50 #include <sys/mman.h>
51 #include <vm/hat.h>
52 #include <vm/as.h>
53 #include <vm/page.h>
54 #include <sys/avintr.h>
55 #include <sys/errno.h>
56 #include <sys/modctl.h>
57 #include <sys/ddi_impldefs.h>
58 #include <sys/sunddi.h>
59 #include <sys/sunndi.h>
60 #include <sys/mach_intr.h>
61 #include <sys/psm.h>
62 #include <sys/ontrap.h>
63 #include <sys/atomic.h>
64 #include <sys/sdt.h>
65 #include <sys/rootnex.h>
66 #include <vm/hat_i86.h>
67 #include <sys/ddifm.h>
68 #include <sys/ddi_isa.h>
69 #include <sys/apic.h>
70 
71 #ifdef __xpv
72 #include <sys/bootinfo.h>
73 #include <sys/hypervisor.h>
74 #include <sys/bootconf.h>
75 #include <vm/kboot_mmu.h>
76 #endif
77 
78 #if defined(__amd64) && !defined(__xpv)
79 #include <sys/immu.h>
80 #endif
81 
82 
83 /*
84  * enable/disable extra checking of function parameters. Useful for debugging
85  * drivers.
86  */
87 #ifdef	DEBUG
88 int rootnex_alloc_check_parms = 1;
89 int rootnex_bind_check_parms = 1;
90 int rootnex_bind_check_inuse = 1;
91 int rootnex_unbind_verify_buffer = 0;
92 int rootnex_sync_check_parms = 1;
93 #else
94 int rootnex_alloc_check_parms = 0;
95 int rootnex_bind_check_parms = 0;
96 int rootnex_bind_check_inuse = 0;
97 int rootnex_unbind_verify_buffer = 0;
98 int rootnex_sync_check_parms = 0;
99 #endif
100 
101 boolean_t rootnex_dmar_not_setup;
102 
103 /* Master Abort and Target Abort panic flag */
104 int rootnex_fm_ma_ta_panic_flag = 0;
105 
106 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
107 int rootnex_bind_fail = 1;
108 int rootnex_bind_warn = 1;
109 uint8_t *rootnex_warn_list;
110 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
111 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
112 
113 /*
114  * revert back to old broken behavior of always sync'ing entire copy buffer.
115  * This is useful if be have a buggy driver which doesn't correctly pass in
116  * the offset and size into ddi_dma_sync().
117  */
118 int rootnex_sync_ignore_params = 0;
119 
120 /*
121  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
122  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
123  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
124  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
125  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
126  * (< 8K). We will still need to allocate the copy buffer during bind though
127  * (if we need one). These can only be modified in /etc/system before rootnex
128  * attach.
129  */
130 #if defined(__amd64)
131 int rootnex_prealloc_cookies = 65;
132 int rootnex_prealloc_windows = 4;
133 int rootnex_prealloc_copybuf = 2;
134 #else
135 int rootnex_prealloc_cookies = 33;
136 int rootnex_prealloc_windows = 4;
137 int rootnex_prealloc_copybuf = 2;
138 #endif
139 
140 /* driver global state */
141 static rootnex_state_t *rootnex_state;
142 
143 #ifdef DEBUG
144 /* shortcut to rootnex counters */
145 static uint64_t *rootnex_cnt;
146 #endif
147 
148 /*
149  * XXX - does x86 even need these or are they left over from the SPARC days?
150  */
151 /* statically defined integer/boolean properties for the root node */
152 static rootnex_intprop_t rootnex_intprp[] = {
153 	{ "PAGESIZE",			PAGESIZE },
154 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
155 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
156 	{ DDI_RELATIVE_ADDRESSING,	1 },
157 };
158 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
159 
160 /*
161  * If we're dom0, we're using a real device so we need to load
162  * the cookies with MFNs instead of PFNs.
163  */
164 #ifdef __xpv
165 typedef maddr_t rootnex_addr_t;
166 #define	ROOTNEX_PADDR_TO_RBASE(pa)	\
167 	(DOMAIN_IS_INITDOMAIN(xen_info) ? pa_to_ma(pa) : (pa))
168 #else
169 typedef paddr_t rootnex_addr_t;
170 #define	ROOTNEX_PADDR_TO_RBASE(pa)	(pa)
171 #endif
172 
173 static struct cb_ops rootnex_cb_ops = {
174 	nodev,		/* open */
175 	nodev,		/* close */
176 	nodev,		/* strategy */
177 	nodev,		/* print */
178 	nodev,		/* dump */
179 	nodev,		/* read */
180 	nodev,		/* write */
181 	nodev,		/* ioctl */
182 	nodev,		/* devmap */
183 	nodev,		/* mmap */
184 	nodev,		/* segmap */
185 	nochpoll,	/* chpoll */
186 	ddi_prop_op,	/* cb_prop_op */
187 	NULL,		/* struct streamtab */
188 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
189 	CB_REV,		/* Rev */
190 	nodev,		/* cb_aread */
191 	nodev		/* cb_awrite */
192 };
193 
194 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
195     off_t offset, off_t len, caddr_t *vaddrp);
196 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
197     struct hat *hat, struct seg *seg, caddr_t addr,
198     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
199 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
200     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
201     ddi_dma_handle_t *handlep);
202 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
203     ddi_dma_handle_t handle);
204 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
205     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
206     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
207 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
208     ddi_dma_handle_t handle);
209 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
210     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
211 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
212     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
213     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
214 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
215     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
216     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
217 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
218     ddi_ctl_enum_t ctlop, void *arg, void *result);
219 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
220     ddi_iblock_cookie_t *ibc);
221 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
222     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
223 static int rootnex_alloc_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *,
224     void *);
225 static int rootnex_free_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *);
226 
227 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
228     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
229     ddi_dma_handle_t *handlep);
230 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
231     ddi_dma_handle_t handle);
232 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
233     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
234     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
235 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
236     ddi_dma_handle_t handle);
237 #if defined(__amd64) && !defined(__xpv)
238 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
239     ddi_dma_handle_t handle);
240 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
241     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
242 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
243     ddi_dma_cookie_t *cookiep, uint_t ccount);
244 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
245     ddi_dma_handle_t handle);
246 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
247 #endif
248 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
249     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
250 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
251     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
252     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
253 
254 #if defined(__amd64) && !defined(__xpv)
255 static int rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
256     ddi_dma_handle_t handle, void *v);
257 static void *rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
258     ddi_dma_handle_t handle);
259 #endif
260 
261 
262 static struct bus_ops rootnex_bus_ops = {
263 	BUSO_REV,
264 	rootnex_map,
265 	NULL,
266 	NULL,
267 	NULL,
268 	rootnex_map_fault,
269 	0,
270 	rootnex_dma_allochdl,
271 	rootnex_dma_freehdl,
272 	rootnex_dma_bindhdl,
273 	rootnex_dma_unbindhdl,
274 	rootnex_dma_sync,
275 	rootnex_dma_win,
276 	rootnex_dma_mctl,
277 	rootnex_ctlops,
278 	ddi_bus_prop_op,
279 	i_ddi_rootnex_get_eventcookie,
280 	i_ddi_rootnex_add_eventcall,
281 	i_ddi_rootnex_remove_eventcall,
282 	i_ddi_rootnex_post_event,
283 	0,			/* bus_intr_ctl */
284 	0,			/* bus_config */
285 	0,			/* bus_unconfig */
286 	rootnex_fm_init,	/* bus_fm_init */
287 	NULL,			/* bus_fm_fini */
288 	NULL,			/* bus_fm_access_enter */
289 	NULL,			/* bus_fm_access_exit */
290 	NULL,			/* bus_powr */
291 	rootnex_intr_ops	/* bus_intr_op */
292 };
293 
294 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
295 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
296 static int rootnex_quiesce(dev_info_t *dip);
297 
298 static struct dev_ops rootnex_ops = {
299 	DEVO_REV,
300 	0,
301 	ddi_no_info,
302 	nulldev,
303 	nulldev,
304 	rootnex_attach,
305 	rootnex_detach,
306 	nulldev,
307 	&rootnex_cb_ops,
308 	&rootnex_bus_ops,
309 	NULL,
310 	rootnex_quiesce,		/* quiesce */
311 };
312 
313 static struct modldrv rootnex_modldrv = {
314 	&mod_driverops,
315 	"i86pc root nexus",
316 	&rootnex_ops
317 };
318 
319 static struct modlinkage rootnex_modlinkage = {
320 	MODREV_1,
321 	(void *)&rootnex_modldrv,
322 	NULL
323 };
324 
325 #if defined(__amd64) && !defined(__xpv)
326 static iommulib_nexops_t iommulib_nexops = {
327 	IOMMU_NEXOPS_VERSION,
328 	"Rootnex IOMMU ops Vers 1.1",
329 	NULL,
330 	rootnex_coredma_allochdl,
331 	rootnex_coredma_freehdl,
332 	rootnex_coredma_bindhdl,
333 	rootnex_coredma_unbindhdl,
334 	rootnex_coredma_reset_cookies,
335 	rootnex_coredma_get_cookies,
336 	rootnex_coredma_set_cookies,
337 	rootnex_coredma_clear_cookies,
338 	rootnex_coredma_get_sleep_flags,
339 	rootnex_coredma_sync,
340 	rootnex_coredma_win,
341 	rootnex_coredma_hdl_setprivate,
342 	rootnex_coredma_hdl_getprivate
343 };
344 #endif
345 
346 /*
347  *  extern hacks
348  */
349 extern struct seg_ops segdev_ops;
350 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
351 #ifdef	DDI_MAP_DEBUG
352 extern int ddi_map_debug_flag;
353 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
354 #endif
355 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
356 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
357 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
358     psm_intr_op_t, int *);
359 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
360 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
361 
362 /*
363  * Use device arena to use for device control register mappings.
364  * Various kernel memory walkers (debugger, dtrace) need to know
365  * to avoid this address range to prevent undesired device activity.
366  */
367 extern void *device_arena_alloc(size_t size, int vm_flag);
368 extern void device_arena_free(void * vaddr, size_t size);
369 
370 
371 /*
372  *  Internal functions
373  */
374 static int rootnex_dma_init();
375 static void rootnex_add_props(dev_info_t *);
376 static int rootnex_ctl_reportdev(dev_info_t *dip);
377 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
378 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
379 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
380 static int rootnex_map_handle(ddi_map_req_t *mp);
381 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
382 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
383 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
384     ddi_dma_attr_t *attr);
385 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
386     rootnex_sglinfo_t *sglinfo);
387 static void rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object,
388     ddi_dma_cookie_t *sgl, rootnex_sglinfo_t *sglinfo);
389 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
390     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
391 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
392     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
393 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
394 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
395     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
396 static void rootnex_teardown_windows(rootnex_dma_t *dma);
397 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
398     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
399 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
400     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
401     size_t *copybuf_used, page_t **cur_pp);
402 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
403     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
404     ddi_dma_attr_t *attr, off_t cur_offset);
405 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
406     rootnex_dma_t *dma, rootnex_window_t **windowp,
407     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
408 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
409     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
410 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
411     off_t offset, size_t size, uint_t cache_flags);
412 static int rootnex_verify_buffer(rootnex_dma_t *dma);
413 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
414     const void *comp_addr, const void *not_used);
415 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
416     rootnex_sglinfo_t *sglinfo);
417 static struct as *rootnex_get_as(ddi_dma_obj_t *dmar_object);
418 
419 /*
420  * _init()
421  *
422  */
423 int
424 _init(void)
425 {
426 
427 	rootnex_state = NULL;
428 	return (mod_install(&rootnex_modlinkage));
429 }
430 
431 
432 /*
433  * _info()
434  *
435  */
436 int
437 _info(struct modinfo *modinfop)
438 {
439 	return (mod_info(&rootnex_modlinkage, modinfop));
440 }
441 
442 
443 /*
444  * _fini()
445  *
446  */
447 int
448 _fini(void)
449 {
450 	return (EBUSY);
451 }
452 
453 
454 /*
455  * rootnex_attach()
456  *
457  */
458 static int
459 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
460 {
461 	int fmcap;
462 	int e;
463 
464 	switch (cmd) {
465 	case DDI_ATTACH:
466 		break;
467 	case DDI_RESUME:
468 #if defined(__amd64) && !defined(__xpv)
469 		return (immu_unquiesce());
470 #else
471 		return (DDI_SUCCESS);
472 #endif
473 	default:
474 		return (DDI_FAILURE);
475 	}
476 
477 	/*
478 	 * We should only have one instance of rootnex. Save it away since we
479 	 * don't have an easy way to get it back later.
480 	 */
481 	ASSERT(rootnex_state == NULL);
482 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
483 
484 	rootnex_state->r_dip = dip;
485 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
486 	rootnex_state->r_reserved_msg_printed = B_FALSE;
487 #ifdef DEBUG
488 	rootnex_cnt = &rootnex_state->r_counters[0];
489 #endif
490 
491 	/*
492 	 * Set minimum fm capability level for i86pc platforms and then
493 	 * initialize error handling. Since we're the rootnex, we don't
494 	 * care what's returned in the fmcap field.
495 	 */
496 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
497 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
498 	fmcap = ddi_system_fmcap;
499 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
500 
501 	/* initialize DMA related state */
502 	e = rootnex_dma_init();
503 	if (e != DDI_SUCCESS) {
504 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
505 		return (DDI_FAILURE);
506 	}
507 
508 	/* Add static root node properties */
509 	rootnex_add_props(dip);
510 
511 	/* since we can't call ddi_report_dev() */
512 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
513 
514 	/* Initialize rootnex event handle */
515 	i_ddi_rootnex_init_events(dip);
516 
517 #if defined(__amd64) && !defined(__xpv)
518 	e = iommulib_nexus_register(dip, &iommulib_nexops,
519 	    &rootnex_state->r_iommulib_handle);
520 
521 	ASSERT(e == DDI_SUCCESS);
522 #endif
523 
524 	return (DDI_SUCCESS);
525 }
526 
527 
528 /*
529  * rootnex_detach()
530  *
531  */
532 /*ARGSUSED*/
533 static int
534 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
535 {
536 	switch (cmd) {
537 	case DDI_SUSPEND:
538 #if defined(__amd64) && !defined(__xpv)
539 		return (immu_quiesce());
540 #else
541 		return (DDI_SUCCESS);
542 #endif
543 	default:
544 		return (DDI_FAILURE);
545 	}
546 	/*NOTREACHED*/
547 
548 }
549 
550 
551 /*
552  * rootnex_dma_init()
553  *
554  */
555 /*ARGSUSED*/
556 static int
557 rootnex_dma_init()
558 {
559 	size_t bufsize;
560 
561 
562 	/*
563 	 * size of our cookie/window/copybuf state needed in dma bind that we
564 	 * pre-alloc in dma_alloc_handle
565 	 */
566 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
567 	rootnex_state->r_prealloc_size =
568 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
569 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
570 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
571 
572 	/*
573 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
574 	 * allocate 16 extra bytes for struct pointer alignment
575 	 * (p->dmai_private & dma->dp_prealloc_buffer)
576 	 */
577 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
578 	    rootnex_state->r_prealloc_size + 0x10;
579 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
580 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
581 	if (rootnex_state->r_dmahdl_cache == NULL) {
582 		return (DDI_FAILURE);
583 	}
584 
585 	/*
586 	 * allocate array to track which major numbers we have printed warnings
587 	 * for.
588 	 */
589 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
590 	    KM_SLEEP);
591 
592 	return (DDI_SUCCESS);
593 }
594 
595 
596 /*
597  * rootnex_add_props()
598  *
599  */
600 static void
601 rootnex_add_props(dev_info_t *dip)
602 {
603 	rootnex_intprop_t *rpp;
604 	int i;
605 
606 	/* Add static integer/boolean properties to the root node */
607 	rpp = rootnex_intprp;
608 	for (i = 0; i < NROOT_INTPROPS; i++) {
609 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
610 		    rpp[i].prop_name, rpp[i].prop_value);
611 	}
612 }
613 
614 
615 
616 /*
617  * *************************
618  *  ctlops related routines
619  * *************************
620  */
621 
622 /*
623  * rootnex_ctlops()
624  *
625  */
626 /*ARGSUSED*/
627 static int
628 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
629     void *arg, void *result)
630 {
631 	int n, *ptr;
632 	struct ddi_parent_private_data *pdp;
633 
634 	switch (ctlop) {
635 	case DDI_CTLOPS_DMAPMAPC:
636 		/*
637 		 * Return 'partial' to indicate that dma mapping
638 		 * has to be done in the main MMU.
639 		 */
640 		return (DDI_DMA_PARTIAL);
641 
642 	case DDI_CTLOPS_BTOP:
643 		/*
644 		 * Convert byte count input to physical page units.
645 		 * (byte counts that are not a page-size multiple
646 		 * are rounded down)
647 		 */
648 		*(ulong_t *)result = btop(*(ulong_t *)arg);
649 		return (DDI_SUCCESS);
650 
651 	case DDI_CTLOPS_PTOB:
652 		/*
653 		 * Convert size in physical pages to bytes
654 		 */
655 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
656 		return (DDI_SUCCESS);
657 
658 	case DDI_CTLOPS_BTOPR:
659 		/*
660 		 * Convert byte count input to physical page units
661 		 * (byte counts that are not a page-size multiple
662 		 * are rounded up)
663 		 */
664 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
665 		return (DDI_SUCCESS);
666 
667 	case DDI_CTLOPS_INITCHILD:
668 		return (impl_ddi_sunbus_initchild(arg));
669 
670 	case DDI_CTLOPS_UNINITCHILD:
671 		impl_ddi_sunbus_removechild(arg);
672 		return (DDI_SUCCESS);
673 
674 	case DDI_CTLOPS_REPORTDEV:
675 		return (rootnex_ctl_reportdev(rdip));
676 
677 	case DDI_CTLOPS_IOMIN:
678 		/*
679 		 * Nothing to do here but reflect back..
680 		 */
681 		return (DDI_SUCCESS);
682 
683 	case DDI_CTLOPS_REGSIZE:
684 	case DDI_CTLOPS_NREGS:
685 		break;
686 
687 	case DDI_CTLOPS_SIDDEV:
688 		if (ndi_dev_is_prom_node(rdip))
689 			return (DDI_SUCCESS);
690 		if (ndi_dev_is_persistent_node(rdip))
691 			return (DDI_SUCCESS);
692 		return (DDI_FAILURE);
693 
694 	case DDI_CTLOPS_POWER:
695 		return ((*pm_platform_power)((power_req_t *)arg));
696 
697 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
698 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
699 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
700 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
701 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
702 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
703 		if (!rootnex_state->r_reserved_msg_printed) {
704 			rootnex_state->r_reserved_msg_printed = B_TRUE;
705 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
706 			    "1 or more reserved/obsolete operations.");
707 		}
708 		return (DDI_FAILURE);
709 
710 	default:
711 		return (DDI_FAILURE);
712 	}
713 	/*
714 	 * The rest are for "hardware" properties
715 	 */
716 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
717 		return (DDI_FAILURE);
718 
719 	if (ctlop == DDI_CTLOPS_NREGS) {
720 		ptr = (int *)result;
721 		*ptr = pdp->par_nreg;
722 	} else {
723 		off_t *size = (off_t *)result;
724 
725 		ptr = (int *)arg;
726 		n = *ptr;
727 		if (n >= pdp->par_nreg) {
728 			return (DDI_FAILURE);
729 		}
730 		*size = (off_t)pdp->par_reg[n].regspec_size;
731 	}
732 	return (DDI_SUCCESS);
733 }
734 
735 
736 /*
737  * rootnex_ctl_reportdev()
738  *
739  */
740 static int
741 rootnex_ctl_reportdev(dev_info_t *dev)
742 {
743 	int i, n, len, f_len = 0;
744 	char *buf;
745 
746 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
747 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
748 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
749 	len = strlen(buf);
750 
751 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
752 
753 		struct regspec *rp = sparc_pd_getreg(dev, i);
754 
755 		if (i == 0)
756 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
757 			    ": ");
758 		else
759 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
760 			    " and ");
761 		len = strlen(buf);
762 
763 		switch (rp->regspec_bustype) {
764 
765 		case BTEISA:
766 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
767 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
768 			break;
769 
770 		case BTISA:
771 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
772 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
773 			break;
774 
775 		default:
776 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
777 			    "space %x offset %x",
778 			    rp->regspec_bustype, rp->regspec_addr);
779 			break;
780 		}
781 		len = strlen(buf);
782 	}
783 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
784 		int pri;
785 
786 		if (i != 0) {
787 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
788 			    ",");
789 			len = strlen(buf);
790 		}
791 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
792 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
793 		    " sparc ipl %d", pri);
794 		len = strlen(buf);
795 	}
796 #ifdef DEBUG
797 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
798 		cmn_err(CE_NOTE, "next message is truncated: "
799 		    "printed length 1024, real length %d", f_len);
800 	}
801 #endif /* DEBUG */
802 	cmn_err(CE_CONT, "?%s\n", buf);
803 	kmem_free(buf, REPORTDEV_BUFSIZE);
804 	return (DDI_SUCCESS);
805 }
806 
807 
808 /*
809  * ******************
810  *  map related code
811  * ******************
812  */
813 
814 /*
815  * rootnex_map()
816  *
817  */
818 static int
819 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
820     off_t len, caddr_t *vaddrp)
821 {
822 	struct regspec *orp = NULL;
823 	struct regspec64 rp = { 0 };
824 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
825 
826 	mp = &mr;
827 
828 	switch (mp->map_op)  {
829 	case DDI_MO_MAP_LOCKED:
830 	case DDI_MO_UNMAP:
831 	case DDI_MO_MAP_HANDLE:
832 		break;
833 	default:
834 #ifdef	DDI_MAP_DEBUG
835 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
836 		    mp->map_op);
837 #endif	/* DDI_MAP_DEBUG */
838 		return (DDI_ME_UNIMPLEMENTED);
839 	}
840 
841 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
842 #ifdef	DDI_MAP_DEBUG
843 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
844 #endif	/* DDI_MAP_DEBUG */
845 		return (DDI_ME_UNIMPLEMENTED);
846 	}
847 
848 	/*
849 	 * First, we need to get the original regspec out before we convert it
850 	 * to the extended format. If we have a register number, then we need to
851 	 * convert that to a regspec.
852 	 */
853 	if (mp->map_type == DDI_MT_RNUMBER)  {
854 
855 		int rnumber = mp->map_obj.rnumber;
856 #ifdef	DDI_MAP_DEBUG
857 		static char *out_of_range =
858 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
859 #endif	/* DDI_MAP_DEBUG */
860 
861 		orp = i_ddi_rnumber_to_regspec(rdip, rnumber);
862 		if (orp == NULL) {
863 #ifdef	DDI_MAP_DEBUG
864 			cmn_err(CE_WARN, out_of_range, rnumber,
865 			    ddi_get_name(rdip));
866 #endif	/* DDI_MAP_DEBUG */
867 			return (DDI_ME_RNUMBER_RANGE);
868 		}
869 	} else if (!(mp->map_flags & DDI_MF_EXT_REGSPEC)) {
870 		orp = mp->map_obj.rp;
871 	}
872 
873 	/*
874 	 * Ensure that we are always using a 64-bit extended regspec regardless
875 	 * of what was passed into us. If the child driver is using a 64-bit
876 	 * regspec, then we need to make sure that we copy this to the local
877 	 * regspec64, rp.
878 	 */
879 	if (orp != NULL) {
880 		rp.regspec_bustype = orp->regspec_bustype;
881 		rp.regspec_addr = orp->regspec_addr;
882 		rp.regspec_size = orp->regspec_size;
883 	} else {
884 		struct regspec64 *rp64;
885 		rp64 = (struct regspec64 *)mp->map_obj.rp;
886 		rp = *rp64;
887 	}
888 
889 	mp->map_type = DDI_MT_REGSPEC;
890 	mp->map_flags |= DDI_MF_EXT_REGSPEC;
891 	mp->map_obj.rp = (struct regspec *)&rp;
892 
893 	/*
894 	 * Adjust offset and length correspnding to called values...
895 	 * XXX: A non-zero length means override the one in the regspec
896 	 * XXX: (regardless of what's in the parent's range?)
897 	 */
898 
899 #ifdef	DDI_MAP_DEBUG
900 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
901 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
902 	    rp.regspec_bustype, rp.regspec_addr, rp.regspec_size, offset,
903 	    len, mp->map_handlep);
904 #endif	/* DDI_MAP_DEBUG */
905 
906 	/*
907 	 * I/O or memory mapping:
908 	 *
909 	 *	<bustype=0, addr=x, len=x>: memory
910 	 *	<bustype=1, addr=x, len=x>: i/o
911 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
912 	 */
913 
914 	if (rp.regspec_bustype > 1 && rp.regspec_addr != 0) {
915 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
916 		    " <0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 ">",
917 		    ddi_get_name(dip), ddi_get_name(rdip), rp.regspec_bustype,
918 		    rp.regspec_addr, rp.regspec_size);
919 		return (DDI_ME_INVAL);
920 	}
921 
922 	if (rp.regspec_bustype > 1 && rp.regspec_addr == 0) {
923 		/*
924 		 * compatibility i/o mapping
925 		 */
926 		rp.regspec_bustype += offset;
927 	} else {
928 		/*
929 		 * Normal memory or i/o mapping
930 		 */
931 		rp.regspec_addr += offset;
932 	}
933 
934 	if (len != 0)
935 		rp.regspec_size = len;
936 
937 #ifdef	DDI_MAP_DEBUG
938 	cmn_err(CE_CONT, "             <%s,%s> <0x%" PRIx64 ", 0x%" PRIx64
939 	    ", 0x%" PRId64 "> offset %d len %d handle 0x%x\n",
940 	    ddi_get_name(dip), ddi_get_name(rdip), rp.regspec_bustype,
941 	    rp.regspec_addr, rp.regspec_size, offset, len, mp->map_handlep);
942 #endif	/* DDI_MAP_DEBUG */
943 
944 
945 	/*
946 	 * The x86 root nexus does not have any notion of valid ranges of
947 	 * addresses. Its children have valid ranges, but because there are none
948 	 * for the nexus, we don't need to call i_ddi_apply_range().  Verify
949 	 * that is the case.
950 	 */
951 	ASSERT0(sparc_pd_getnrng(dip));
952 
953 	switch (mp->map_op)  {
954 	case DDI_MO_MAP_LOCKED:
955 
956 		/*
957 		 * Set up the locked down kernel mapping to the regspec...
958 		 */
959 
960 		return (rootnex_map_regspec(mp, vaddrp));
961 
962 	case DDI_MO_UNMAP:
963 
964 		/*
965 		 * Release mapping...
966 		 */
967 
968 		return (rootnex_unmap_regspec(mp, vaddrp));
969 
970 	case DDI_MO_MAP_HANDLE:
971 
972 		return (rootnex_map_handle(mp));
973 
974 	default:
975 		return (DDI_ME_UNIMPLEMENTED);
976 	}
977 }
978 
979 
980 /*
981  * rootnex_map_fault()
982  *
983  *	fault in mappings for requestors
984  */
985 /*ARGSUSED*/
986 static int
987 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
988     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
989     uint_t lock)
990 {
991 
992 #ifdef	DDI_MAP_DEBUG
993 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
994 	ddi_map_debug(" Seg <%s>\n",
995 	    seg->s_ops == &segdev_ops ? "segdev" :
996 	    seg == &kvseg ? "segkmem" : "NONE!");
997 #endif	/* DDI_MAP_DEBUG */
998 
999 	/*
1000 	 * This is all terribly broken, but it is a start
1001 	 *
1002 	 * XXX	Note that this test means that segdev_ops
1003 	 *	must be exported from seg_dev.c.
1004 	 * XXX	What about devices with their own segment drivers?
1005 	 */
1006 	if (seg->s_ops == &segdev_ops) {
1007 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
1008 
1009 		if (hat == NULL) {
1010 			/*
1011 			 * This is one plausible interpretation of
1012 			 * a null hat i.e. use the first hat on the
1013 			 * address space hat list which by convention is
1014 			 * the hat of the system MMU.  At alternative
1015 			 * would be to panic .. this might well be better ..
1016 			 */
1017 			ASSERT(AS_READ_HELD(seg->s_as));
1018 			hat = seg->s_as->a_hat;
1019 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
1020 		}
1021 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
1022 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
1023 	} else if (seg == &kvseg && dp == NULL) {
1024 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
1025 		    HAT_LOAD_LOCK);
1026 	} else
1027 		return (DDI_FAILURE);
1028 	return (DDI_SUCCESS);
1029 }
1030 
1031 
1032 static int
1033 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1034 {
1035 	rootnex_addr_t rbase;
1036 	void *cvaddr;
1037 	uint64_t npages, pgoffset;
1038 	struct regspec64 *rp;
1039 	ddi_acc_hdl_t *hp;
1040 	ddi_acc_impl_t *ap;
1041 	uint_t	hat_acc_flags;
1042 	paddr_t pbase;
1043 
1044 	ASSERT(mp->map_flags & DDI_MF_EXT_REGSPEC);
1045 	rp = (struct regspec64 *)mp->map_obj.rp;
1046 	hp = mp->map_handlep;
1047 
1048 #ifdef	DDI_MAP_DEBUG
1049 	ddi_map_debug(
1050 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1051 	    rp->regspec_bustype, rp->regspec_addr,
1052 	    rp->regspec_size, mp->map_handlep);
1053 #endif	/* DDI_MAP_DEBUG */
1054 
1055 	/*
1056 	 * I/O or memory mapping
1057 	 *
1058 	 *	<bustype=0, addr=x, len=x>: memory
1059 	 *	<bustype=1, addr=x, len=x>: i/o
1060 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1061 	 */
1062 
1063 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1064 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1065 		    " <0x%" PRIx64 ", 0x%" PRIx64", 0x%" PRIx64">",
1066 		    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size);
1067 		return (DDI_FAILURE);
1068 	}
1069 
1070 	if (rp->regspec_bustype != 0) {
1071 		/*
1072 		 * I/O space - needs a handle.
1073 		 */
1074 		if (hp == NULL) {
1075 			return (DDI_FAILURE);
1076 		}
1077 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1078 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1079 		impl_acc_hdl_init(hp);
1080 
1081 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1082 #ifdef  DDI_MAP_DEBUG
1083 			ddi_map_debug("rootnex_map_regspec: mmap() "
1084 			    "to I/O space is not supported.\n");
1085 #endif  /* DDI_MAP_DEBUG */
1086 			return (DDI_ME_INVAL);
1087 		} else {
1088 			/*
1089 			 * 1275-compliant vs. compatibility i/o mapping
1090 			 */
1091 			*vaddrp =
1092 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1093 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1094 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1095 #ifdef __xpv
1096 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1097 				hp->ah_pfn = xen_assign_pfn(
1098 				    mmu_btop((ulong_t)rp->regspec_addr &
1099 				    MMU_PAGEMASK));
1100 			} else {
1101 				hp->ah_pfn = mmu_btop(
1102 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1103 			}
1104 #else
1105 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1106 			    MMU_PAGEMASK);
1107 #endif
1108 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1109 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1110 		}
1111 
1112 #ifdef	DDI_MAP_DEBUG
1113 		ddi_map_debug(
1114 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1115 		    rp->regspec_size, *vaddrp);
1116 #endif	/* DDI_MAP_DEBUG */
1117 		return (DDI_SUCCESS);
1118 	}
1119 
1120 	/*
1121 	 * Memory space
1122 	 */
1123 
1124 	if (hp != NULL) {
1125 		/*
1126 		 * hat layer ignores
1127 		 * hp->ah_acc.devacc_attr_endian_flags.
1128 		 */
1129 		switch (hp->ah_acc.devacc_attr_dataorder) {
1130 		case DDI_STRICTORDER_ACC:
1131 			hat_acc_flags = HAT_STRICTORDER;
1132 			break;
1133 		case DDI_UNORDERED_OK_ACC:
1134 			hat_acc_flags = HAT_UNORDERED_OK;
1135 			break;
1136 		case DDI_MERGING_OK_ACC:
1137 			hat_acc_flags = HAT_MERGING_OK;
1138 			break;
1139 		case DDI_LOADCACHING_OK_ACC:
1140 			hat_acc_flags = HAT_LOADCACHING_OK;
1141 			break;
1142 		case DDI_STORECACHING_OK_ACC:
1143 			hat_acc_flags = HAT_STORECACHING_OK;
1144 			break;
1145 		default:
1146 			return (DDI_ME_INVAL);
1147 		}
1148 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1149 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1150 		impl_acc_hdl_init(hp);
1151 		hp->ah_hat_flags = hat_acc_flags;
1152 	} else {
1153 		hat_acc_flags = HAT_STRICTORDER;
1154 	}
1155 
1156 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1157 #ifdef __xpv
1158 	/*
1159 	 * If we're dom0, we're using a real device so we need to translate
1160 	 * the MA to a PA.
1161 	 */
1162 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1163 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1164 	} else {
1165 		pbase = rbase;
1166 	}
1167 #else
1168 	pbase = rbase;
1169 #endif
1170 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1171 
1172 	if (rp->regspec_size == 0) {
1173 #ifdef  DDI_MAP_DEBUG
1174 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1175 #endif  /* DDI_MAP_DEBUG */
1176 		return (DDI_ME_INVAL);
1177 	}
1178 
1179 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1180 		/* extra cast to make gcc happy */
1181 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1182 	} else {
1183 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1184 
1185 #ifdef	DDI_MAP_DEBUG
1186 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1187 		    "physical %llx", npages, pbase);
1188 #endif	/* DDI_MAP_DEBUG */
1189 
1190 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1191 		if (cvaddr == NULL)
1192 			return (DDI_ME_NORESOURCES);
1193 
1194 		/*
1195 		 * Now map in the pages we've allocated...
1196 		 */
1197 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1198 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1199 		    HAT_LOAD_LOCK);
1200 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1201 
1202 		/* save away pfn and npages for FMA */
1203 		hp = mp->map_handlep;
1204 		if (hp) {
1205 			hp->ah_pfn = mmu_btop(pbase);
1206 			hp->ah_pnum = npages;
1207 		}
1208 	}
1209 
1210 #ifdef	DDI_MAP_DEBUG
1211 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1212 #endif	/* DDI_MAP_DEBUG */
1213 	return (DDI_SUCCESS);
1214 }
1215 
1216 
1217 static int
1218 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1219 {
1220 	caddr_t addr = (caddr_t)*vaddrp;
1221 	uint64_t npages, pgoffset;
1222 	struct regspec64 *rp;
1223 
1224 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1225 		return (0);
1226 
1227 	ASSERT(mp->map_flags & DDI_MF_EXT_REGSPEC);
1228 	rp = (struct regspec64 *)mp->map_obj.rp;
1229 
1230 	if (rp->regspec_size == 0) {
1231 #ifdef  DDI_MAP_DEBUG
1232 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1233 #endif  /* DDI_MAP_DEBUG */
1234 		return (DDI_ME_INVAL);
1235 	}
1236 
1237 	/*
1238 	 * I/O or memory mapping:
1239 	 *
1240 	 *	<bustype=0, addr=x, len=x>: memory
1241 	 *	<bustype=1, addr=x, len=x>: i/o
1242 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1243 	 */
1244 	if (rp->regspec_bustype != 0) {
1245 		/*
1246 		 * This is I/O space, which requires no particular
1247 		 * processing on unmap since it isn't mapped in the
1248 		 * first place.
1249 		 */
1250 		return (DDI_SUCCESS);
1251 	}
1252 
1253 	/*
1254 	 * Memory space
1255 	 */
1256 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1257 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1258 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1259 	device_arena_free(addr - pgoffset, ptob(npages));
1260 
1261 	/*
1262 	 * Destroy the pointer - the mapping has logically gone
1263 	 */
1264 	*vaddrp = NULL;
1265 
1266 	return (DDI_SUCCESS);
1267 }
1268 
1269 static int
1270 rootnex_map_handle(ddi_map_req_t *mp)
1271 {
1272 	rootnex_addr_t rbase;
1273 	ddi_acc_hdl_t *hp;
1274 	uint64_t pgoffset;
1275 	struct regspec64 *rp;
1276 	paddr_t pbase;
1277 
1278 	rp = (struct regspec64 *)mp->map_obj.rp;
1279 
1280 #ifdef	DDI_MAP_DEBUG
1281 	ddi_map_debug(
1282 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1283 	    rp->regspec_bustype, rp->regspec_addr,
1284 	    rp->regspec_size, mp->map_handlep);
1285 #endif	/* DDI_MAP_DEBUG */
1286 
1287 	/*
1288 	 * I/O or memory mapping:
1289 	 *
1290 	 *	<bustype=0, addr=x, len=x>: memory
1291 	 *	<bustype=1, addr=x, len=x>: i/o
1292 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1293 	 */
1294 	if (rp->regspec_bustype != 0) {
1295 		/*
1296 		 * This refers to I/O space, and we don't support "mapping"
1297 		 * I/O space to a user.
1298 		 */
1299 		return (DDI_FAILURE);
1300 	}
1301 
1302 	/*
1303 	 * Set up the hat_flags for the mapping.
1304 	 */
1305 	hp = mp->map_handlep;
1306 
1307 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1308 	case DDI_NEVERSWAP_ACC:
1309 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1310 		break;
1311 	case DDI_STRUCTURE_LE_ACC:
1312 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1313 		break;
1314 	case DDI_STRUCTURE_BE_ACC:
1315 		return (DDI_FAILURE);
1316 	default:
1317 		return (DDI_REGS_ACC_CONFLICT);
1318 	}
1319 
1320 	switch (hp->ah_acc.devacc_attr_dataorder) {
1321 	case DDI_STRICTORDER_ACC:
1322 		break;
1323 	case DDI_UNORDERED_OK_ACC:
1324 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1325 		break;
1326 	case DDI_MERGING_OK_ACC:
1327 		hp->ah_hat_flags |= HAT_MERGING_OK;
1328 		break;
1329 	case DDI_LOADCACHING_OK_ACC:
1330 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1331 		break;
1332 	case DDI_STORECACHING_OK_ACC:
1333 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1334 		break;
1335 	default:
1336 		return (DDI_FAILURE);
1337 	}
1338 
1339 	rbase = (rootnex_addr_t)rp->regspec_addr &
1340 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1341 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1342 
1343 	if (rp->regspec_size == 0)
1344 		return (DDI_ME_INVAL);
1345 
1346 #ifdef __xpv
1347 	/*
1348 	 * If we're dom0, we're using a real device so we need to translate
1349 	 * the MA to a PA.
1350 	 */
1351 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1352 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1353 		    (rbase & MMU_PAGEOFFSET);
1354 	} else {
1355 		pbase = rbase;
1356 	}
1357 #else
1358 	pbase = rbase;
1359 #endif
1360 
1361 	hp->ah_pfn = mmu_btop(pbase);
1362 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1363 
1364 	return (DDI_SUCCESS);
1365 }
1366 
1367 
1368 
1369 /*
1370  * ************************
1371  *  interrupt related code
1372  * ************************
1373  */
1374 
1375 /*
1376  * rootnex_intr_ops()
1377  *	bus_intr_op() function for interrupt support
1378  */
1379 /* ARGSUSED */
1380 static int
1381 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1382     ddi_intr_handle_impl_t *hdlp, void *result)
1383 {
1384 	struct intrspec			*ispec;
1385 
1386 	DDI_INTR_NEXDBG((CE_CONT,
1387 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1388 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1389 
1390 	/* Process the interrupt operation */
1391 	switch (intr_op) {
1392 	case DDI_INTROP_GETCAP:
1393 		/* First check with pcplusmp */
1394 		if (psm_intr_ops == NULL)
1395 			return (DDI_FAILURE);
1396 
1397 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1398 			*(int *)result = 0;
1399 			return (DDI_FAILURE);
1400 		}
1401 		break;
1402 	case DDI_INTROP_SETCAP:
1403 		if (psm_intr_ops == NULL)
1404 			return (DDI_FAILURE);
1405 
1406 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1407 			return (DDI_FAILURE);
1408 		break;
1409 	case DDI_INTROP_ALLOC:
1410 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1411 		return (rootnex_alloc_intr_fixed(rdip, hdlp, result));
1412 	case DDI_INTROP_FREE:
1413 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1414 		return (rootnex_free_intr_fixed(rdip, hdlp));
1415 	case DDI_INTROP_GETPRI:
1416 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1417 			return (DDI_FAILURE);
1418 		*(int *)result = ispec->intrspec_pri;
1419 		break;
1420 	case DDI_INTROP_SETPRI:
1421 		/* Validate the interrupt priority passed to us */
1422 		if (*(int *)result > LOCK_LEVEL)
1423 			return (DDI_FAILURE);
1424 
1425 		/* Ensure that PSM is all initialized and ispec is ok */
1426 		if ((psm_intr_ops == NULL) ||
1427 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1428 			return (DDI_FAILURE);
1429 
1430 		/* Change the priority */
1431 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1432 		    PSM_FAILURE)
1433 			return (DDI_FAILURE);
1434 
1435 		/* update the ispec with the new priority */
1436 		ispec->intrspec_pri =  *(int *)result;
1437 		break;
1438 	case DDI_INTROP_ADDISR:
1439 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1440 			return (DDI_FAILURE);
1441 		ispec->intrspec_func = hdlp->ih_cb_func;
1442 		break;
1443 	case DDI_INTROP_REMISR:
1444 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1445 			return (DDI_FAILURE);
1446 		ispec->intrspec_func = (uint_t (*)()) 0;
1447 		break;
1448 	case DDI_INTROP_ENABLE:
1449 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1450 			return (DDI_FAILURE);
1451 
1452 		/* Call psmi to translate irq with the dip */
1453 		if (psm_intr_ops == NULL)
1454 			return (DDI_FAILURE);
1455 
1456 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1457 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1458 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1459 			return (DDI_FAILURE);
1460 
1461 		/* Add the interrupt handler */
1462 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1463 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1464 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1465 			return (DDI_FAILURE);
1466 		break;
1467 	case DDI_INTROP_DISABLE:
1468 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1469 			return (DDI_FAILURE);
1470 
1471 		/* Call psm_ops() to translate irq with the dip */
1472 		if (psm_intr_ops == NULL)
1473 			return (DDI_FAILURE);
1474 
1475 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1476 		(void) (*psm_intr_ops)(rdip, hdlp,
1477 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1478 
1479 		/* Remove the interrupt handler */
1480 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1481 		    hdlp->ih_cb_func, hdlp->ih_vector);
1482 		break;
1483 	case DDI_INTROP_SETMASK:
1484 		if (psm_intr_ops == NULL)
1485 			return (DDI_FAILURE);
1486 
1487 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1488 			return (DDI_FAILURE);
1489 		break;
1490 	case DDI_INTROP_CLRMASK:
1491 		if (psm_intr_ops == NULL)
1492 			return (DDI_FAILURE);
1493 
1494 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1495 			return (DDI_FAILURE);
1496 		break;
1497 	case DDI_INTROP_GETPENDING:
1498 		if (psm_intr_ops == NULL)
1499 			return (DDI_FAILURE);
1500 
1501 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1502 		    result)) {
1503 			*(int *)result = 0;
1504 			return (DDI_FAILURE);
1505 		}
1506 		break;
1507 	case DDI_INTROP_NAVAIL:
1508 	case DDI_INTROP_NINTRS:
1509 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1510 		if (*(int *)result == 0) {
1511 			/*
1512 			 * Special case for 'pcic' driver' only. This driver
1513 			 * driver is a child of 'isa' and 'rootnex' drivers.
1514 			 *
1515 			 * See detailed comments on this in the function
1516 			 * rootnex_get_ispec().
1517 			 *
1518 			 * Children of 'pcic' send 'NINITR' request all the
1519 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1520 			 * field may not initialized. So, we fake it here
1521 			 * to return 1 (a la what PCMCIA nexus does).
1522 			 */
1523 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1524 				*(int *)result = 1;
1525 			else
1526 				return (DDI_FAILURE);
1527 		}
1528 		break;
1529 	case DDI_INTROP_SUPPORTED_TYPES:
1530 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1531 		break;
1532 	default:
1533 		return (DDI_FAILURE);
1534 	}
1535 
1536 	return (DDI_SUCCESS);
1537 }
1538 
1539 
1540 /*
1541  * rootnex_get_ispec()
1542  *	convert an interrupt number to an interrupt specification.
1543  *	The interrupt number determines which interrupt spec will be
1544  *	returned if more than one exists.
1545  *
1546  *	Look into the parent private data area of the 'rdip' to find out
1547  *	the interrupt specification.  First check to make sure there is
1548  *	one that matchs "inumber" and then return a pointer to it.
1549  *
1550  *	Return NULL if one could not be found.
1551  *
1552  *	NOTE: This is needed for rootnex_intr_ops()
1553  */
1554 static struct intrspec *
1555 rootnex_get_ispec(dev_info_t *rdip, int inum)
1556 {
1557 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1558 
1559 	/*
1560 	 * Special case handling for drivers that provide their own
1561 	 * intrspec structures instead of relying on the DDI framework.
1562 	 *
1563 	 * A broken hardware driver in ON could potentially provide its
1564 	 * own intrspec structure, instead of relying on the hardware.
1565 	 * If these drivers are children of 'rootnex' then we need to
1566 	 * continue to provide backward compatibility to them here.
1567 	 *
1568 	 * Following check is a special case for 'pcic' driver which
1569 	 * was found to have broken hardwre andby provides its own intrspec.
1570 	 *
1571 	 * Verbatim comments from this driver are shown here:
1572 	 * "Don't use the ddi_add_intr since we don't have a
1573 	 * default intrspec in all cases."
1574 	 *
1575 	 * Since an 'ispec' may not be always created for it,
1576 	 * check for that and create one if so.
1577 	 *
1578 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1579 	 */
1580 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1581 		pdp->par_nintr = 1;
1582 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1583 		    pdp->par_nintr, KM_SLEEP);
1584 	}
1585 
1586 	/* Validate the interrupt number */
1587 	if (inum >= pdp->par_nintr)
1588 		return (NULL);
1589 
1590 	/* Get the interrupt structure pointer and return that */
1591 	return ((struct intrspec *)&pdp->par_intr[inum]);
1592 }
1593 
1594 /*
1595  * Allocate interrupt vector for FIXED (legacy) type.
1596  */
1597 static int
1598 rootnex_alloc_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp,
1599     void *result)
1600 {
1601 	struct intrspec		*ispec;
1602 	ddi_intr_handle_impl_t	info_hdl;
1603 	int			ret;
1604 	int			free_phdl = 0;
1605 	apic_get_type_t		type_info;
1606 
1607 	if (psm_intr_ops == NULL)
1608 		return (DDI_FAILURE);
1609 
1610 	if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1611 		return (DDI_FAILURE);
1612 
1613 	/*
1614 	 * If the PSM module is "APIX" then pass the request for it
1615 	 * to allocate the vector now.
1616 	 */
1617 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1618 	info_hdl.ih_private = &type_info;
1619 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1620 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1621 		if (hdlp->ih_private == NULL) { /* allocate phdl structure */
1622 			free_phdl = 1;
1623 			i_ddi_alloc_intr_phdl(hdlp);
1624 		}
1625 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1626 		ret = (*psm_intr_ops)(rdip, hdlp,
1627 		    PSM_INTR_OP_ALLOC_VECTORS, result);
1628 		if (free_phdl) { /* free up the phdl structure */
1629 			free_phdl = 0;
1630 			i_ddi_free_intr_phdl(hdlp);
1631 			hdlp->ih_private = NULL;
1632 		}
1633 	} else {
1634 		/*
1635 		 * No APIX module; fall back to the old scheme where the
1636 		 * interrupt vector is allocated during ddi_enable_intr() call.
1637 		 */
1638 		hdlp->ih_pri = ispec->intrspec_pri;
1639 		*(int *)result = hdlp->ih_scratch1;
1640 		ret = DDI_SUCCESS;
1641 	}
1642 
1643 	return (ret);
1644 }
1645 
1646 /*
1647  * Free up interrupt vector for FIXED (legacy) type.
1648  */
1649 static int
1650 rootnex_free_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp)
1651 {
1652 	struct intrspec			*ispec;
1653 	struct ddi_parent_private_data	*pdp;
1654 	ddi_intr_handle_impl_t		info_hdl;
1655 	int				ret;
1656 	apic_get_type_t			type_info;
1657 
1658 	if (psm_intr_ops == NULL)
1659 		return (DDI_FAILURE);
1660 
1661 	/*
1662 	 * If the PSM module is "APIX" then pass the request for it
1663 	 * to free up the vector now.
1664 	 */
1665 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1666 	info_hdl.ih_private = &type_info;
1667 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1668 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1669 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1670 			return (DDI_FAILURE);
1671 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1672 		ret = (*psm_intr_ops)(rdip, hdlp,
1673 		    PSM_INTR_OP_FREE_VECTORS, NULL);
1674 	} else {
1675 		/*
1676 		 * No APIX module; fall back to the old scheme where
1677 		 * the interrupt vector was already freed during
1678 		 * ddi_disable_intr() call.
1679 		 */
1680 		ret = DDI_SUCCESS;
1681 	}
1682 
1683 	pdp = ddi_get_parent_data(rdip);
1684 
1685 	/*
1686 	 * Special case for 'pcic' driver' only.
1687 	 * If an intrspec was created for it, clean it up here
1688 	 * See detailed comments on this in the function
1689 	 * rootnex_get_ispec().
1690 	 */
1691 	if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1692 		kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1693 		    pdp->par_nintr);
1694 		/*
1695 		 * Set it to zero; so that
1696 		 * DDI framework doesn't free it again
1697 		 */
1698 		pdp->par_intr = NULL;
1699 		pdp->par_nintr = 0;
1700 	}
1701 
1702 	return (ret);
1703 }
1704 
1705 
1706 /*
1707  * ******************
1708  *  dma related code
1709  * ******************
1710  */
1711 
1712 /*ARGSUSED*/
1713 static int
1714 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1715     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1716     ddi_dma_handle_t *handlep)
1717 {
1718 	uint64_t maxsegmentsize_ll;
1719 	uint_t maxsegmentsize;
1720 	ddi_dma_impl_t *hp;
1721 	rootnex_dma_t *dma;
1722 	uint64_t count_max;
1723 	uint64_t seg;
1724 	int kmflag;
1725 	int e;
1726 
1727 
1728 	/* convert our sleep flags */
1729 	if (waitfp == DDI_DMA_SLEEP) {
1730 		kmflag = KM_SLEEP;
1731 	} else {
1732 		kmflag = KM_NOSLEEP;
1733 	}
1734 
1735 	/*
1736 	 * We try to do only one memory allocation here. We'll do a little
1737 	 * pointer manipulation later. If the bind ends up taking more than
1738 	 * our prealloc's space, we'll have to allocate more memory in the
1739 	 * bind operation. Not great, but much better than before and the
1740 	 * best we can do with the current bind interfaces.
1741 	 */
1742 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1743 	if (hp == NULL)
1744 		return (DDI_DMA_NORESOURCES);
1745 
1746 	/* Do our pointer manipulation now, align the structures */
1747 	hp->dmai_private = (void *)(((uintptr_t)hp +
1748 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1749 	dma = (rootnex_dma_t *)hp->dmai_private;
1750 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1751 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1752 
1753 	/* setup the handle */
1754 	rootnex_clean_dmahdl(hp);
1755 	hp->dmai_error.err_fep = NULL;
1756 	hp->dmai_error.err_cf = NULL;
1757 	dma->dp_dip = rdip;
1758 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1759 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1760 
1761 	/*
1762 	 * The BOUNCE_ON_SEG workaround is not needed when an IOMMU
1763 	 * is being used. Set the upper limit to the seg value.
1764 	 * There will be enough DVMA space to always get addresses
1765 	 * that will match the constraints.
1766 	 */
1767 	if (IOMMU_USED(rdip) &&
1768 	    (attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG)) {
1769 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_seg;
1770 		dma->dp_sglinfo.si_flags &= ~_DDI_DMA_BOUNCE_ON_SEG;
1771 	} else
1772 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1773 
1774 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1775 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1776 	hp->dmai_rdip = rdip;
1777 	hp->dmai_attr = *attr;
1778 
1779 	if (attr->dma_attr_seg >= dma->dp_sglinfo.si_max_addr)
1780 		dma->dp_sglinfo.si_cancross = B_FALSE;
1781 	else
1782 		dma->dp_sglinfo.si_cancross = B_TRUE;
1783 
1784 	/* we don't need to worry about the SPL since we do a tryenter */
1785 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1786 
1787 	/*
1788 	 * Figure out our maximum segment size. If the segment size is greater
1789 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1790 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1791 	 * dma_attr_count_max are size-1 type values.
1792 	 *
1793 	 * Maximum segment size is the largest physically contiguous chunk of
1794 	 * memory that we can return from a bind (i.e. the maximum size of a
1795 	 * single cookie).
1796 	 */
1797 
1798 	/* handle the rollover cases */
1799 	seg = attr->dma_attr_seg + 1;
1800 	if (seg < attr->dma_attr_seg) {
1801 		seg = attr->dma_attr_seg;
1802 	}
1803 	count_max = attr->dma_attr_count_max + 1;
1804 	if (count_max < attr->dma_attr_count_max) {
1805 		count_max = attr->dma_attr_count_max;
1806 	}
1807 
1808 	/*
1809 	 * granularity may or may not be a power of two. If it isn't, we can't
1810 	 * use a simple mask.
1811 	 */
1812 	if (!ISP2(attr->dma_attr_granular)) {
1813 		dma->dp_granularity_power_2 = B_FALSE;
1814 	} else {
1815 		dma->dp_granularity_power_2 = B_TRUE;
1816 	}
1817 
1818 	/*
1819 	 * maxxfer should be a whole multiple of granularity. If we're going to
1820 	 * break up a window because we're greater than maxxfer, we might as
1821 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1822 	 * worry about triming the window later on for this case.
1823 	 */
1824 	if (attr->dma_attr_granular > 1) {
1825 		if (dma->dp_granularity_power_2) {
1826 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1827 			    (attr->dma_attr_maxxfer &
1828 			    (attr->dma_attr_granular - 1));
1829 		} else {
1830 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1831 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1832 		}
1833 	} else {
1834 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1835 	}
1836 
1837 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1838 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1839 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1840 		maxsegmentsize = 0xFFFFFFFF;
1841 	} else {
1842 		maxsegmentsize = maxsegmentsize_ll;
1843 	}
1844 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1845 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1846 
1847 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1848 	if (rootnex_alloc_check_parms) {
1849 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1850 		if (e != DDI_SUCCESS) {
1851 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1852 			(void) rootnex_dma_freehdl(dip, rdip,
1853 			    (ddi_dma_handle_t)hp);
1854 			return (e);
1855 		}
1856 	}
1857 
1858 	*handlep = (ddi_dma_handle_t)hp;
1859 
1860 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1861 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1862 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1863 
1864 	return (DDI_SUCCESS);
1865 }
1866 
1867 
1868 /*
1869  * rootnex_dma_allochdl()
1870  *    called from ddi_dma_alloc_handle().
1871  */
1872 static int
1873 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1874     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1875 {
1876 	int retval = DDI_SUCCESS;
1877 #if defined(__amd64) && !defined(__xpv)
1878 
1879 	if (IOMMU_UNITIALIZED(rdip)) {
1880 		retval = iommulib_nex_open(dip, rdip);
1881 
1882 		if (retval != DDI_SUCCESS && retval != DDI_ENOTSUP)
1883 			return (retval);
1884 	}
1885 
1886 	if (IOMMU_UNUSED(rdip)) {
1887 		retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1888 		    handlep);
1889 	} else {
1890 		retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1891 		    waitfp, arg, handlep);
1892 	}
1893 #else
1894 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1895 	    handlep);
1896 #endif
1897 	switch (retval) {
1898 	case DDI_DMA_NORESOURCES:
1899 		if (waitfp != DDI_DMA_DONTWAIT) {
1900 			ddi_set_callback(waitfp, arg,
1901 			    &rootnex_state->r_dvma_call_list_id);
1902 		}
1903 		break;
1904 	case DDI_SUCCESS:
1905 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1906 		break;
1907 	default:
1908 		break;
1909 	}
1910 	return (retval);
1911 }
1912 
1913 /*ARGSUSED*/
1914 static int
1915 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1916     ddi_dma_handle_t handle)
1917 {
1918 	ddi_dma_impl_t *hp;
1919 	rootnex_dma_t *dma;
1920 
1921 
1922 	hp = (ddi_dma_impl_t *)handle;
1923 	dma = (rootnex_dma_t *)hp->dmai_private;
1924 
1925 	/* unbind should have been called first */
1926 	ASSERT(!dma->dp_inuse);
1927 
1928 	mutex_destroy(&dma->dp_mutex);
1929 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1930 
1931 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1932 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1933 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1934 
1935 	return (DDI_SUCCESS);
1936 }
1937 
1938 /*
1939  * rootnex_dma_freehdl()
1940  *    called from ddi_dma_free_handle().
1941  */
1942 static int
1943 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1944 {
1945 	int ret;
1946 
1947 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1948 #if defined(__amd64) && !defined(__xpv)
1949 	if (IOMMU_USED(rdip))
1950 		ret = iommulib_nexdma_freehdl(dip, rdip, handle);
1951 	else
1952 #endif
1953 	ret = rootnex_coredma_freehdl(dip, rdip, handle);
1954 
1955 	if (rootnex_state->r_dvma_call_list_id)
1956 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1957 
1958 	return (ret);
1959 }
1960 
1961 /*ARGSUSED*/
1962 static int
1963 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1964     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1965     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1966 {
1967 	rootnex_sglinfo_t *sinfo;
1968 	ddi_dma_obj_t *dmao;
1969 #if defined(__amd64) && !defined(__xpv)
1970 	struct dvmaseg *dvs;
1971 	ddi_dma_cookie_t *cookie;
1972 #endif
1973 	ddi_dma_attr_t *attr;
1974 	ddi_dma_impl_t *hp;
1975 	rootnex_dma_t *dma;
1976 	int kmflag;
1977 	int e;
1978 	uint_t ncookies;
1979 
1980 	hp = (ddi_dma_impl_t *)handle;
1981 	dma = (rootnex_dma_t *)hp->dmai_private;
1982 	dmao = &dma->dp_dma;
1983 	sinfo = &dma->dp_sglinfo;
1984 	attr = &hp->dmai_attr;
1985 
1986 	/* convert the sleep flags */
1987 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1988 		dma->dp_sleep_flags = kmflag = KM_SLEEP;
1989 	} else {
1990 		dma->dp_sleep_flags = kmflag = KM_NOSLEEP;
1991 	}
1992 
1993 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
1994 
1995 	/*
1996 	 * This is useful for debugging a driver. Not as useful in a production
1997 	 * system. The only time this will fail is if you have a driver bug.
1998 	 */
1999 	if (rootnex_bind_check_inuse) {
2000 		/*
2001 		 * No one else should ever have this lock unless someone else
2002 		 * is trying to use this handle. So contention on the lock
2003 		 * is the same as inuse being set.
2004 		 */
2005 		e = mutex_tryenter(&dma->dp_mutex);
2006 		if (e == 0) {
2007 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2008 			return (DDI_DMA_INUSE);
2009 		}
2010 		if (dma->dp_inuse) {
2011 			mutex_exit(&dma->dp_mutex);
2012 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2013 			return (DDI_DMA_INUSE);
2014 		}
2015 		dma->dp_inuse = B_TRUE;
2016 		mutex_exit(&dma->dp_mutex);
2017 	}
2018 
2019 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
2020 	if (rootnex_bind_check_parms) {
2021 		e = rootnex_valid_bind_parms(dmareq, attr);
2022 		if (e != DDI_SUCCESS) {
2023 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2024 			rootnex_clean_dmahdl(hp);
2025 			return (e);
2026 		}
2027 	}
2028 
2029 	/* save away the original bind info */
2030 	dma->dp_dma = dmareq->dmar_object;
2031 
2032 #if defined(__amd64) && !defined(__xpv)
2033 	if (IOMMU_USED(rdip)) {
2034 		dmao = &dma->dp_dvma;
2035 		e = iommulib_nexdma_mapobject(dip, rdip, handle, dmareq, dmao);
2036 		switch (e) {
2037 		case DDI_SUCCESS:
2038 			if (sinfo->si_cancross ||
2039 			    dmao->dmao_obj.dvma_obj.dv_nseg != 1 ||
2040 			    dmao->dmao_size > sinfo->si_max_cookie_size) {
2041 				dma->dp_dvma_used = B_TRUE;
2042 				break;
2043 			}
2044 			sinfo->si_sgl_size = 1;
2045 			hp->dmai_rflags |= DMP_NOSYNC;
2046 
2047 			dma->dp_dvma_used = B_TRUE;
2048 			dma->dp_need_to_free_cookie = B_FALSE;
2049 
2050 			dvs = &dmao->dmao_obj.dvma_obj.dv_seg[0];
2051 			cookie = hp->dmai_cookie = dma->dp_cookies =
2052 			    (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2053 			cookie->dmac_laddress = dvs->dvs_start +
2054 			    dmao->dmao_obj.dvma_obj.dv_off;
2055 			cookie->dmac_size = dvs->dvs_len;
2056 			cookie->dmac_type = 0;
2057 
2058 			ROOTNEX_DPROBE1(rootnex__bind__dvmafast, dev_info_t *,
2059 			    rdip);
2060 			goto fast;
2061 		case DDI_ENOTSUP:
2062 			break;
2063 		default:
2064 			rootnex_clean_dmahdl(hp);
2065 			return (e);
2066 		}
2067 	}
2068 #endif
2069 
2070 	/*
2071 	 * Figure out a rough estimate of what maximum number of pages
2072 	 * this buffer could use (a high estimate of course).
2073 	 */
2074 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
2075 
2076 	if (dma->dp_dvma_used) {
2077 		/*
2078 		 * The number of physical pages is the worst case.
2079 		 *
2080 		 * For DVMA, the worst case is the length divided
2081 		 * by the maximum cookie length, plus 1. Add to that
2082 		 * the number of segment boundaries potentially crossed, and
2083 		 * the additional number of DVMA segments that was returned.
2084 		 *
2085 		 * In the normal case, for modern devices, si_cancross will
2086 		 * be false, and dv_nseg will be 1, and the fast path will
2087 		 * have been taken above.
2088 		 */
2089 		ncookies = (dma->dp_dma.dmao_size / sinfo->si_max_cookie_size)
2090 		    + 1;
2091 		if (sinfo->si_cancross)
2092 			ncookies +=
2093 			    (dma->dp_dma.dmao_size / attr->dma_attr_seg) + 1;
2094 		ncookies += (dmao->dmao_obj.dvma_obj.dv_nseg - 1);
2095 
2096 		sinfo->si_max_pages = MIN(sinfo->si_max_pages, ncookies);
2097 	}
2098 
2099 	/*
2100 	 * We'll use the pre-allocated cookies for any bind that will *always*
2101 	 * fit (more important to be consistent, we don't want to create
2102 	 * additional degenerate cases).
2103 	 */
2104 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
2105 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2106 		dma->dp_need_to_free_cookie = B_FALSE;
2107 		ROOTNEX_DPROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
2108 		    uint_t, sinfo->si_max_pages);
2109 
2110 	/*
2111 	 * For anything larger than that, we'll go ahead and allocate the
2112 	 * maximum number of pages we expect to see. Hopefuly, we won't be
2113 	 * seeing this path in the fast path for high performance devices very
2114 	 * frequently.
2115 	 *
2116 	 * a ddi bind interface that allowed the driver to provide storage to
2117 	 * the bind interface would speed this case up.
2118 	 */
2119 	} else {
2120 		/*
2121 		 * Save away how much memory we allocated. If we're doing a
2122 		 * nosleep, the alloc could fail...
2123 		 */
2124 		dma->dp_cookie_size = sinfo->si_max_pages *
2125 		    sizeof (ddi_dma_cookie_t);
2126 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
2127 		if (dma->dp_cookies == NULL) {
2128 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2129 			rootnex_clean_dmahdl(hp);
2130 			return (DDI_DMA_NORESOURCES);
2131 		}
2132 		dma->dp_need_to_free_cookie = B_TRUE;
2133 		ROOTNEX_DPROBE2(rootnex__bind__alloc, dev_info_t *, rdip,
2134 		    uint_t, sinfo->si_max_pages);
2135 	}
2136 	hp->dmai_cookie = dma->dp_cookies;
2137 
2138 	/*
2139 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
2140 	 * looking at the constraints in the dma structure. It will then put
2141 	 * some additional state about the sgl in the dma struct (i.e. is
2142 	 * the sgl clean, or do we need to do some munging; how many pages
2143 	 * need to be copied, etc.)
2144 	 */
2145 	if (dma->dp_dvma_used)
2146 		rootnex_dvma_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2147 	else
2148 		rootnex_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2149 
2150 out:
2151 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
2152 	/* if we don't need a copy buffer, we don't need to sync */
2153 	if (sinfo->si_copybuf_req == 0) {
2154 		hp->dmai_rflags |= DMP_NOSYNC;
2155 	}
2156 
2157 	/*
2158 	 * if we don't need the copybuf and we don't need to do a partial,  we
2159 	 * hit the fast path. All the high performance devices should be trying
2160 	 * to hit this path. To hit this path, a device should be able to reach
2161 	 * all of memory, shouldn't try to bind more than it can transfer, and
2162 	 * the buffer shouldn't require more cookies than the driver/device can
2163 	 * handle [sgllen]).
2164 	 *
2165 	 * Note that negative values of dma_attr_sgllen are supposed
2166 	 * to mean unlimited, but we just cast them to mean a
2167 	 * "ridiculous large limit".  This saves some extra checks on
2168 	 * hot paths.
2169 	 */
2170 	if ((sinfo->si_copybuf_req == 0) &&
2171 	    (sinfo->si_sgl_size <= (unsigned)attr->dma_attr_sgllen) &&
2172 	    (dmao->dmao_size <= dma->dp_maxxfer)) {
2173 fast:
2174 		/*
2175 		 * If the driver supports FMA, insert the handle in the FMA DMA
2176 		 * handle cache.
2177 		 */
2178 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2179 			hp->dmai_error.err_cf = rootnex_dma_check;
2180 
2181 		/*
2182 		 * copy out the first cookie and ccountp, set the cookie
2183 		 * pointer to the second cookie. The first cookie is passed
2184 		 * back on the stack. Additional cookies are accessed via
2185 		 * ddi_dma_nextcookie()
2186 		 */
2187 		*cookiep = dma->dp_cookies[0];
2188 		*ccountp = sinfo->si_sgl_size;
2189 		hp->dmai_cookie++;
2190 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2191 		hp->dmai_ncookies = *ccountp;
2192 		hp->dmai_curcookie = 1;
2193 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2194 		ROOTNEX_DPROBE4(rootnex__bind__fast, dev_info_t *, rdip,
2195 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2196 		    uint_t, dmao->dmao_size, uint_t, *ccountp);
2197 
2198 
2199 		return (DDI_DMA_MAPPED);
2200 	}
2201 
2202 	/*
2203 	 * go to the slow path, we may need to alloc more memory, create
2204 	 * multiple windows, and munge up a sgl to make the device happy.
2205 	 */
2206 
2207 	/*
2208 	 * With the IOMMU mapobject method used, we should never hit
2209 	 * the slow path. If we do, something is seriously wrong.
2210 	 * Clean up and return an error.
2211 	 */
2212 
2213 #if defined(__amd64) && !defined(__xpv)
2214 
2215 	if (dma->dp_dvma_used) {
2216 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2217 		    &dma->dp_dvma);
2218 		e = DDI_DMA_NOMAPPING;
2219 	} else {
2220 #endif
2221 		e = rootnex_bind_slowpath(hp, dmareq, dma, attr, &dma->dp_dma,
2222 		    kmflag);
2223 #if defined(__amd64) && !defined(__xpv)
2224 	}
2225 #endif
2226 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2227 		if (dma->dp_need_to_free_cookie) {
2228 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2229 		}
2230 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2231 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2232 		return (e);
2233 	}
2234 
2235 	/*
2236 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2237 	 * cache.
2238 	 */
2239 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2240 		hp->dmai_error.err_cf = rootnex_dma_check;
2241 
2242 	/* if the first window uses the copy buffer, sync it for the device */
2243 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2244 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2245 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2246 		    DDI_DMA_SYNC_FORDEV);
2247 	}
2248 
2249 	/*
2250 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2251 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2252 	 * If we have a partial map (i.e. multiple windows), the number of
2253 	 * cookies we return is the number of cookies in the first window.
2254 	 */
2255 	if (e == DDI_DMA_MAPPED) {
2256 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2257 		*ccountp = sinfo->si_sgl_size;
2258 		hp->dmai_nwin = 1;
2259 	} else {
2260 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2261 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2262 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2263 	}
2264 	*cookiep = dma->dp_cookies[0];
2265 	hp->dmai_cookie++;
2266 	hp->dmai_ncookies = *ccountp;
2267 	hp->dmai_curcookie = 1;
2268 
2269 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2270 	ROOTNEX_DPROBE4(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2271 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2272 	    dmao->dmao_size, uint_t, *ccountp);
2273 	return (e);
2274 }
2275 
2276 /*
2277  * rootnex_dma_bindhdl()
2278  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2279  */
2280 static int
2281 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2282     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2283     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2284 {
2285 	int ret;
2286 #if defined(__amd64) && !defined(__xpv)
2287 	if (IOMMU_USED(rdip))
2288 		ret = iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2289 		    cookiep, ccountp);
2290 	else
2291 #endif
2292 	ret = rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2293 	    cookiep, ccountp);
2294 
2295 	if (ret == DDI_DMA_NORESOURCES && dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
2296 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
2297 		    &rootnex_state->r_dvma_call_list_id);
2298 	}
2299 
2300 	return (ret);
2301 }
2302 
2303 
2304 
2305 /*ARGSUSED*/
2306 static int
2307 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2308     ddi_dma_handle_t handle)
2309 {
2310 	ddi_dma_impl_t *hp;
2311 	rootnex_dma_t *dma;
2312 	int e;
2313 
2314 	hp = (ddi_dma_impl_t *)handle;
2315 	dma = (rootnex_dma_t *)hp->dmai_private;
2316 
2317 	/* make sure the buffer wasn't free'd before calling unbind */
2318 	if (rootnex_unbind_verify_buffer) {
2319 		e = rootnex_verify_buffer(dma);
2320 		if (e != DDI_SUCCESS) {
2321 			ASSERT(0);
2322 			return (DDI_FAILURE);
2323 		}
2324 	}
2325 
2326 	/* sync the current window before unbinding the buffer */
2327 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2328 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2329 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2330 		    DDI_DMA_SYNC_FORCPU);
2331 	}
2332 
2333 	/*
2334 	 * cleanup and copy buffer or window state. if we didn't use the copy
2335 	 * buffer or windows, there won't be much to do :-)
2336 	 */
2337 	rootnex_teardown_copybuf(dma);
2338 	rootnex_teardown_windows(dma);
2339 
2340 #if defined(__amd64) && !defined(__xpv)
2341 	if (IOMMU_USED(rdip) && dma->dp_dvma_used)
2342 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2343 		    &dma->dp_dvma);
2344 #endif
2345 
2346 	/*
2347 	 * If we had to allocate space to for the worse case sgl (it didn't
2348 	 * fit into our pre-allocate buffer), free that up now
2349 	 */
2350 	if (dma->dp_need_to_free_cookie) {
2351 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2352 	}
2353 
2354 	/*
2355 	 * clean up the handle so it's ready for the next bind (i.e. if the
2356 	 * handle is reused).
2357 	 */
2358 	rootnex_clean_dmahdl(hp);
2359 	hp->dmai_error.err_cf = NULL;
2360 
2361 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2362 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2363 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2364 
2365 	return (DDI_SUCCESS);
2366 }
2367 
2368 /*
2369  * rootnex_dma_unbindhdl()
2370  *    called from ddi_dma_unbind_handle()
2371  */
2372 /*ARGSUSED*/
2373 static int
2374 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2375     ddi_dma_handle_t handle)
2376 {
2377 	int ret;
2378 
2379 #if defined(__amd64) && !defined(__xpv)
2380 	if (IOMMU_USED(rdip))
2381 		ret = iommulib_nexdma_unbindhdl(dip, rdip, handle);
2382 	else
2383 #endif
2384 	ret = rootnex_coredma_unbindhdl(dip, rdip, handle);
2385 
2386 	if (rootnex_state->r_dvma_call_list_id)
2387 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2388 
2389 	return (ret);
2390 }
2391 
2392 #if defined(__amd64) && !defined(__xpv)
2393 
2394 static int
2395 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2396 {
2397 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2398 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2399 
2400 	if (dma->dp_sleep_flags != KM_SLEEP &&
2401 	    dma->dp_sleep_flags != KM_NOSLEEP)
2402 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2403 	return (dma->dp_sleep_flags);
2404 }
2405 /*ARGSUSED*/
2406 static void
2407 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2408 {
2409 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2410 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2411 	rootnex_window_t *window;
2412 
2413 	if (dma->dp_window) {
2414 		window = &dma->dp_window[dma->dp_current_win];
2415 		hp->dmai_cookie = window->wd_first_cookie;
2416 	} else {
2417 		hp->dmai_cookie = dma->dp_cookies;
2418 	}
2419 	hp->dmai_cookie++;
2420 	hp->dmai_curcookie = 1;
2421 }
2422 
2423 /*ARGSUSED*/
2424 static int
2425 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2426     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2427 {
2428 	int i;
2429 	int km_flags;
2430 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2431 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2432 	rootnex_window_t *window;
2433 	ddi_dma_cookie_t *cp;
2434 	ddi_dma_cookie_t *cookie;
2435 
2436 	ASSERT(*cookiepp == NULL);
2437 	ASSERT(*ccountp == 0);
2438 
2439 	if (dma->dp_window) {
2440 		window = &dma->dp_window[dma->dp_current_win];
2441 		cp = window->wd_first_cookie;
2442 		*ccountp = window->wd_cookie_cnt;
2443 	} else {
2444 		cp = dma->dp_cookies;
2445 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2446 	}
2447 
2448 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2449 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2450 	if (cookie == NULL) {
2451 		return (DDI_DMA_NORESOURCES);
2452 	}
2453 
2454 	for (i = 0; i < *ccountp; i++) {
2455 		cookie[i].dmac_notused = cp[i].dmac_notused;
2456 		cookie[i].dmac_type = cp[i].dmac_type;
2457 		cookie[i].dmac_address = cp[i].dmac_address;
2458 		cookie[i].dmac_size = cp[i].dmac_size;
2459 	}
2460 
2461 	*cookiepp = cookie;
2462 
2463 	return (DDI_SUCCESS);
2464 }
2465 
2466 /*ARGSUSED*/
2467 static int
2468 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2469     ddi_dma_cookie_t *cookiep, uint_t ccount)
2470 {
2471 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2472 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2473 	rootnex_window_t *window;
2474 	ddi_dma_cookie_t *cur_cookiep;
2475 
2476 	ASSERT(cookiep);
2477 	ASSERT(ccount != 0);
2478 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2479 
2480 	if (dma->dp_window) {
2481 		window = &dma->dp_window[dma->dp_current_win];
2482 		dma->dp_saved_cookies = window->wd_first_cookie;
2483 		window->wd_first_cookie = cookiep;
2484 		ASSERT(ccount == window->wd_cookie_cnt);
2485 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2486 		    + window->wd_first_cookie;
2487 	} else {
2488 		dma->dp_saved_cookies = dma->dp_cookies;
2489 		dma->dp_cookies = cookiep;
2490 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2491 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2492 		    + dma->dp_cookies;
2493 	}
2494 
2495 	dma->dp_need_to_switch_cookies = B_TRUE;
2496 	hp->dmai_cookie = cur_cookiep;
2497 
2498 	return (DDI_SUCCESS);
2499 }
2500 
2501 /*ARGSUSED*/
2502 static int
2503 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2504 {
2505 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2506 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2507 	rootnex_window_t *window;
2508 	ddi_dma_cookie_t *cur_cookiep;
2509 	ddi_dma_cookie_t *cookie_array;
2510 	uint_t ccount;
2511 
2512 	/* check if cookies have not been switched */
2513 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2514 		return (DDI_SUCCESS);
2515 
2516 	ASSERT(dma->dp_saved_cookies);
2517 
2518 	if (dma->dp_window) {
2519 		window = &dma->dp_window[dma->dp_current_win];
2520 		cookie_array = window->wd_first_cookie;
2521 		window->wd_first_cookie = dma->dp_saved_cookies;
2522 		dma->dp_saved_cookies = NULL;
2523 		ccount = window->wd_cookie_cnt;
2524 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2525 		    + window->wd_first_cookie;
2526 	} else {
2527 		cookie_array = dma->dp_cookies;
2528 		dma->dp_cookies = dma->dp_saved_cookies;
2529 		dma->dp_saved_cookies = NULL;
2530 		ccount = dma->dp_sglinfo.si_sgl_size;
2531 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2532 		    + dma->dp_cookies;
2533 	}
2534 
2535 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2536 
2537 	hp->dmai_cookie = cur_cookiep;
2538 
2539 	dma->dp_need_to_switch_cookies = B_FALSE;
2540 
2541 	return (DDI_SUCCESS);
2542 }
2543 
2544 #endif
2545 
2546 static struct as *
2547 rootnex_get_as(ddi_dma_obj_t *dmao)
2548 {
2549 	struct as *asp;
2550 
2551 	switch (dmao->dmao_type) {
2552 	case DMA_OTYP_VADDR:
2553 	case DMA_OTYP_BUFVADDR:
2554 		asp = dmao->dmao_obj.virt_obj.v_as;
2555 		if (asp == NULL)
2556 			asp = &kas;
2557 		break;
2558 	default:
2559 		asp = NULL;
2560 		break;
2561 	}
2562 	return (asp);
2563 }
2564 
2565 /*
2566  * rootnex_verify_buffer()
2567  *   verify buffer wasn't free'd
2568  */
2569 static int
2570 rootnex_verify_buffer(rootnex_dma_t *dma)
2571 {
2572 	page_t **pplist;
2573 	caddr_t vaddr;
2574 	uint_t pcnt;
2575 	uint_t poff;
2576 	page_t *pp;
2577 	char b;
2578 	int i;
2579 
2580 	/* Figure out how many pages this buffer occupies */
2581 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2582 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2583 	} else {
2584 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2585 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2586 	}
2587 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2588 
2589 	switch (dma->dp_dma.dmao_type) {
2590 	case DMA_OTYP_PAGES:
2591 		/*
2592 		 * for a linked list of pp's walk through them to make sure
2593 		 * they're locked and not free.
2594 		 */
2595 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2596 		for (i = 0; i < pcnt; i++) {
2597 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2598 				return (DDI_FAILURE);
2599 			}
2600 			pp = pp->p_next;
2601 		}
2602 		break;
2603 
2604 	case DMA_OTYP_VADDR:
2605 	case DMA_OTYP_BUFVADDR:
2606 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2607 		/*
2608 		 * for an array of pp's walk through them to make sure they're
2609 		 * not free. It's possible that they may not be locked.
2610 		 */
2611 		if (pplist) {
2612 			for (i = 0; i < pcnt; i++) {
2613 				if (PP_ISFREE(pplist[i])) {
2614 					return (DDI_FAILURE);
2615 				}
2616 			}
2617 
2618 		/* For a virtual address, try to peek at each page */
2619 		} else {
2620 			if (rootnex_get_as(&dma->dp_dma) == &kas) {
2621 				for (i = 0; i < pcnt; i++) {
2622 					if (ddi_peek8(NULL, vaddr, &b) ==
2623 					    DDI_FAILURE)
2624 						return (DDI_FAILURE);
2625 					vaddr += MMU_PAGESIZE;
2626 				}
2627 			}
2628 		}
2629 		break;
2630 
2631 	default:
2632 		cmn_err(CE_PANIC, "rootnex_verify_buffer: bad DMA object");
2633 		break;
2634 	}
2635 
2636 	return (DDI_SUCCESS);
2637 }
2638 
2639 
2640 /*
2641  * rootnex_clean_dmahdl()
2642  *    Clean the dma handle. This should be called on a handle alloc and an
2643  *    unbind handle. Set the handle state to the default settings.
2644  */
2645 static void
2646 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2647 {
2648 	rootnex_dma_t *dma;
2649 
2650 
2651 	dma = (rootnex_dma_t *)hp->dmai_private;
2652 
2653 	hp->dmai_nwin = 0;
2654 	dma->dp_current_cookie = 0;
2655 	dma->dp_copybuf_size = 0;
2656 	dma->dp_window = NULL;
2657 	dma->dp_cbaddr = NULL;
2658 	dma->dp_inuse = B_FALSE;
2659 	dma->dp_dvma_used = B_FALSE;
2660 	dma->dp_need_to_free_cookie = B_FALSE;
2661 	dma->dp_need_to_switch_cookies = B_FALSE;
2662 	dma->dp_saved_cookies = NULL;
2663 	dma->dp_sleep_flags = KM_PANIC;
2664 	dma->dp_need_to_free_window = B_FALSE;
2665 	dma->dp_partial_required = B_FALSE;
2666 	dma->dp_trim_required = B_FALSE;
2667 	dma->dp_sglinfo.si_copybuf_req = 0;
2668 #if !defined(__amd64)
2669 	dma->dp_cb_remaping = B_FALSE;
2670 	dma->dp_kva = NULL;
2671 #endif
2672 
2673 	/* FMA related initialization */
2674 	hp->dmai_fault = 0;
2675 	hp->dmai_fault_check = NULL;
2676 	hp->dmai_fault_notify = NULL;
2677 	hp->dmai_error.err_ena = 0;
2678 	hp->dmai_error.err_status = DDI_FM_OK;
2679 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2680 	hp->dmai_error.err_ontrap = NULL;
2681 
2682 	/* Cookie tracking */
2683 	hp->dmai_ncookies = 0;
2684 	hp->dmai_curcookie = 0;
2685 }
2686 
2687 
2688 /*
2689  * rootnex_valid_alloc_parms()
2690  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2691  */
2692 static int
2693 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2694 {
2695 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2696 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2697 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2698 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2699 		return (DDI_DMA_BADATTR);
2700 	}
2701 
2702 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2703 		return (DDI_DMA_BADATTR);
2704 	}
2705 
2706 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2707 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2708 	    attr->dma_attr_sgllen == 0) {
2709 		return (DDI_DMA_BADATTR);
2710 	}
2711 
2712 	/* We should be able to DMA into every byte offset in a page */
2713 	if (maxsegmentsize < MMU_PAGESIZE) {
2714 		return (DDI_DMA_BADATTR);
2715 	}
2716 
2717 	/* if we're bouncing on seg, seg must be <= addr_hi */
2718 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2719 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2720 		return (DDI_DMA_BADATTR);
2721 	}
2722 	return (DDI_SUCCESS);
2723 }
2724 
2725 /*
2726  * rootnex_valid_bind_parms()
2727  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2728  */
2729 /* ARGSUSED */
2730 static int
2731 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2732 {
2733 #if !defined(__amd64)
2734 	/*
2735 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2736 	 * we can track the offset for the obsoleted interfaces.
2737 	 */
2738 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2739 		return (DDI_DMA_TOOBIG);
2740 	}
2741 #endif
2742 
2743 	return (DDI_SUCCESS);
2744 }
2745 
2746 
2747 /*
2748  * rootnex_need_bounce_seg()
2749  *    check to see if the buffer lives on both side of the seg.
2750  */
2751 static boolean_t
2752 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2753 {
2754 	ddi_dma_atyp_t buftype;
2755 	rootnex_addr_t raddr;
2756 	boolean_t lower_addr;
2757 	boolean_t upper_addr;
2758 	uint64_t offset;
2759 	page_t **pplist;
2760 	uint64_t paddr;
2761 	uint32_t psize;
2762 	uint32_t size;
2763 	caddr_t vaddr;
2764 	uint_t pcnt;
2765 	page_t *pp;
2766 
2767 	pp = NULL;
2768 	/* shortcuts */
2769 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2770 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2771 	buftype = dmar_object->dmao_type;
2772 	size = dmar_object->dmao_size;
2773 
2774 	lower_addr = B_FALSE;
2775 	upper_addr = B_FALSE;
2776 	pcnt = 0;
2777 
2778 	/*
2779 	 * Process the first page to handle the initial offset of the buffer.
2780 	 * We'll use the base address we get later when we loop through all
2781 	 * the pages.
2782 	 */
2783 	if (buftype == DMA_OTYP_PAGES) {
2784 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2785 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2786 		    MMU_PAGEOFFSET;
2787 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2788 		psize = MIN(size, (MMU_PAGESIZE - offset));
2789 		pp = pp->p_next;
2790 		sglinfo->si_asp = NULL;
2791 	} else if (pplist != NULL) {
2792 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2793 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2794 		if (sglinfo->si_asp == NULL) {
2795 			sglinfo->si_asp = &kas;
2796 		}
2797 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2798 		paddr += offset;
2799 		psize = MIN(size, (MMU_PAGESIZE - offset));
2800 		pcnt++;
2801 	} else {
2802 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2803 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2804 		if (sglinfo->si_asp == NULL) {
2805 			sglinfo->si_asp = &kas;
2806 		}
2807 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2808 		paddr += offset;
2809 		psize = MIN(size, (MMU_PAGESIZE - offset));
2810 		vaddr += psize;
2811 	}
2812 
2813 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2814 
2815 	if ((raddr + psize) > sglinfo->si_segmask) {
2816 		upper_addr = B_TRUE;
2817 	} else {
2818 		lower_addr = B_TRUE;
2819 	}
2820 	size -= psize;
2821 
2822 	/*
2823 	 * Walk through the rest of the pages in the buffer. Track to see
2824 	 * if we have pages on both sides of the segment boundary.
2825 	 */
2826 	while (size > 0) {
2827 		/* partial or full page */
2828 		psize = MIN(size, MMU_PAGESIZE);
2829 
2830 		if (buftype == DMA_OTYP_PAGES) {
2831 			/* get the paddr from the page_t */
2832 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2833 			paddr = pfn_to_pa(pp->p_pagenum);
2834 			pp = pp->p_next;
2835 		} else if (pplist != NULL) {
2836 			/* index into the array of page_t's to get the paddr */
2837 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2838 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2839 			pcnt++;
2840 		} else {
2841 			/* call into the VM to get the paddr */
2842 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2843 			    vaddr));
2844 			vaddr += psize;
2845 		}
2846 
2847 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2848 
2849 		if ((raddr + psize) > sglinfo->si_segmask) {
2850 			upper_addr = B_TRUE;
2851 		} else {
2852 			lower_addr = B_TRUE;
2853 		}
2854 		/*
2855 		 * if the buffer lives both above and below the segment
2856 		 * boundary, or the current page is the page immediately
2857 		 * after the segment, we will use a copy/bounce buffer for
2858 		 * all pages > seg.
2859 		 */
2860 		if ((lower_addr && upper_addr) ||
2861 		    (raddr == (sglinfo->si_segmask + 1))) {
2862 			return (B_TRUE);
2863 		}
2864 
2865 		size -= psize;
2866 	}
2867 
2868 	return (B_FALSE);
2869 }
2870 
2871 /*
2872  * rootnex_get_sgl()
2873  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2874  *    with a call to the vm layer when vm2.0 comes around...
2875  */
2876 static void
2877 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2878     rootnex_sglinfo_t *sglinfo)
2879 {
2880 	ddi_dma_atyp_t buftype;
2881 	rootnex_addr_t raddr;
2882 	uint64_t last_page;
2883 	uint64_t offset;
2884 	uint64_t addrhi;
2885 	uint64_t addrlo;
2886 	uint64_t maxseg;
2887 	page_t **pplist;
2888 	uint64_t paddr;
2889 	uint32_t psize;
2890 	uint32_t size;
2891 	caddr_t vaddr;
2892 	uint_t pcnt;
2893 	page_t *pp;
2894 	uint_t cnt;
2895 
2896 	pp = NULL;
2897 	/* shortcuts */
2898 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2899 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2900 	maxseg = sglinfo->si_max_cookie_size;
2901 	buftype = dmar_object->dmao_type;
2902 	addrhi = sglinfo->si_max_addr;
2903 	addrlo = sglinfo->si_min_addr;
2904 	size = dmar_object->dmao_size;
2905 
2906 	pcnt = 0;
2907 	cnt = 0;
2908 
2909 
2910 	/*
2911 	 * check to see if we need to use the copy buffer for pages over
2912 	 * the segment attr.
2913 	 */
2914 	sglinfo->si_bounce_on_seg = B_FALSE;
2915 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2916 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2917 		    dmar_object, sglinfo);
2918 	}
2919 
2920 	/*
2921 	 * if we were passed down a linked list of pages, i.e. pointer to
2922 	 * page_t, use this to get our physical address and buf offset.
2923 	 */
2924 	if (buftype == DMA_OTYP_PAGES) {
2925 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2926 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2927 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2928 		    MMU_PAGEOFFSET;
2929 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2930 		psize = MIN(size, (MMU_PAGESIZE - offset));
2931 		pp = pp->p_next;
2932 		sglinfo->si_asp = NULL;
2933 
2934 	/*
2935 	 * We weren't passed down a linked list of pages, but if we were passed
2936 	 * down an array of pages, use this to get our physical address and buf
2937 	 * offset.
2938 	 */
2939 	} else if (pplist != NULL) {
2940 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2941 		    (buftype == DMA_OTYP_BUFVADDR));
2942 
2943 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2944 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2945 		if (sglinfo->si_asp == NULL) {
2946 			sglinfo->si_asp = &kas;
2947 		}
2948 
2949 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2950 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2951 		paddr += offset;
2952 		psize = MIN(size, (MMU_PAGESIZE - offset));
2953 		pcnt++;
2954 
2955 	/*
2956 	 * All we have is a virtual address, we'll need to call into the VM
2957 	 * to get the physical address.
2958 	 */
2959 	} else {
2960 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2961 		    (buftype == DMA_OTYP_BUFVADDR));
2962 
2963 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2964 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2965 		if (sglinfo->si_asp == NULL) {
2966 			sglinfo->si_asp = &kas;
2967 		}
2968 
2969 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2970 		paddr += offset;
2971 		psize = MIN(size, (MMU_PAGESIZE - offset));
2972 		vaddr += psize;
2973 	}
2974 
2975 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2976 
2977 	/*
2978 	 * Setup the first cookie with the physical address of the page and the
2979 	 * size of the page (which takes into account the initial offset into
2980 	 * the page.
2981 	 */
2982 	sgl[cnt].dmac_laddress = raddr;
2983 	sgl[cnt].dmac_size = psize;
2984 	sgl[cnt].dmac_type = 0;
2985 
2986 	/*
2987 	 * Save away the buffer offset into the page. We'll need this later in
2988 	 * the copy buffer code to help figure out the page index within the
2989 	 * buffer and the offset into the current page.
2990 	 */
2991 	sglinfo->si_buf_offset = offset;
2992 
2993 	/*
2994 	 * If we are using the copy buffer for anything over the segment
2995 	 * boundary, and this page is over the segment boundary.
2996 	 *   OR
2997 	 * if the DMA engine can't reach the physical address.
2998 	 */
2999 	if (((sglinfo->si_bounce_on_seg) &&
3000 	    ((raddr + psize) > sglinfo->si_segmask)) ||
3001 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
3002 		/*
3003 		 * Increase how much copy buffer we use. We always increase by
3004 		 * pagesize so we don't have to worry about converting offsets.
3005 		 * Set a flag in the cookies dmac_type to indicate that it uses
3006 		 * the copy buffer. If this isn't the last cookie, go to the
3007 		 * next cookie (since we separate each page which uses the copy
3008 		 * buffer in case the copy buffer is not physically contiguous.
3009 		 */
3010 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
3011 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3012 		if ((cnt + 1) < sglinfo->si_max_pages) {
3013 			cnt++;
3014 			sgl[cnt].dmac_laddress = 0;
3015 			sgl[cnt].dmac_size = 0;
3016 			sgl[cnt].dmac_type = 0;
3017 		}
3018 	}
3019 
3020 	/*
3021 	 * save this page's physical address so we can figure out if the next
3022 	 * page is physically contiguous. Keep decrementing size until we are
3023 	 * done with the buffer.
3024 	 */
3025 	last_page = raddr & MMU_PAGEMASK;
3026 	size -= psize;
3027 
3028 	while (size > 0) {
3029 		/* Get the size for this page (i.e. partial or full page) */
3030 		psize = MIN(size, MMU_PAGESIZE);
3031 
3032 		if (buftype == DMA_OTYP_PAGES) {
3033 			/* get the paddr from the page_t */
3034 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
3035 			paddr = pfn_to_pa(pp->p_pagenum);
3036 			pp = pp->p_next;
3037 		} else if (pplist != NULL) {
3038 			/* index into the array of page_t's to get the paddr */
3039 			ASSERT(!PP_ISFREE(pplist[pcnt]));
3040 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
3041 			pcnt++;
3042 		} else {
3043 			/* call into the VM to get the paddr */
3044 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
3045 			    vaddr));
3046 			vaddr += psize;
3047 		}
3048 
3049 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
3050 
3051 		/*
3052 		 * If we are using the copy buffer for anything over the
3053 		 * segment boundary, and this page is over the segment
3054 		 * boundary.
3055 		 *   OR
3056 		 * if the DMA engine can't reach the physical address.
3057 		 */
3058 		if (((sglinfo->si_bounce_on_seg) &&
3059 		    ((raddr + psize) > sglinfo->si_segmask)) ||
3060 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
3061 
3062 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
3063 
3064 			/*
3065 			 * if there is something in the current cookie, go to
3066 			 * the next one. We only want one page in a cookie which
3067 			 * uses the copybuf since the copybuf doesn't have to
3068 			 * be physically contiguous.
3069 			 */
3070 			if (sgl[cnt].dmac_size != 0) {
3071 				cnt++;
3072 			}
3073 			sgl[cnt].dmac_laddress = raddr;
3074 			sgl[cnt].dmac_size = psize;
3075 #if defined(__amd64)
3076 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3077 #else
3078 			/*
3079 			 * save the buf offset for 32-bit kernel. used in the
3080 			 * obsoleted interfaces.
3081 			 */
3082 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
3083 			    (dmar_object->dmao_size - size);
3084 #endif
3085 			/* if this isn't the last cookie, go to the next one */
3086 			if ((cnt + 1) < sglinfo->si_max_pages) {
3087 				cnt++;
3088 				sgl[cnt].dmac_laddress = 0;
3089 				sgl[cnt].dmac_size = 0;
3090 				sgl[cnt].dmac_type = 0;
3091 			}
3092 
3093 		/*
3094 		 * this page didn't need the copy buffer, if it's not physically
3095 		 * contiguous, or it would put us over a segment boundary, or it
3096 		 * puts us over the max cookie size, or the current sgl doesn't
3097 		 * have anything in it.
3098 		 */
3099 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
3100 		    !(raddr & sglinfo->si_segmask) ||
3101 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3102 		    (sgl[cnt].dmac_size == 0)) {
3103 			/*
3104 			 * if we're not already in a new cookie, go to the next
3105 			 * cookie.
3106 			 */
3107 			if (sgl[cnt].dmac_size != 0) {
3108 				cnt++;
3109 			}
3110 
3111 			/* save the cookie information */
3112 			sgl[cnt].dmac_laddress = raddr;
3113 			sgl[cnt].dmac_size = psize;
3114 #if defined(__amd64)
3115 			sgl[cnt].dmac_type = 0;
3116 #else
3117 			/*
3118 			 * save the buf offset for 32-bit kernel. used in the
3119 			 * obsoleted interfaces.
3120 			 */
3121 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
3122 #endif
3123 
3124 		/*
3125 		 * this page didn't need the copy buffer, it is physically
3126 		 * contiguous with the last page, and it's <= the max cookie
3127 		 * size.
3128 		 */
3129 		} else {
3130 			sgl[cnt].dmac_size += psize;
3131 
3132 			/*
3133 			 * if this exactly ==  the maximum cookie size, and
3134 			 * it isn't the last cookie, go to the next cookie.
3135 			 */
3136 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3137 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3138 				cnt++;
3139 				sgl[cnt].dmac_laddress = 0;
3140 				sgl[cnt].dmac_size = 0;
3141 				sgl[cnt].dmac_type = 0;
3142 			}
3143 		}
3144 
3145 		/*
3146 		 * save this page's physical address so we can figure out if the
3147 		 * next page is physically contiguous. Keep decrementing size
3148 		 * until we are done with the buffer.
3149 		 */
3150 		last_page = raddr;
3151 		size -= psize;
3152 	}
3153 
3154 	/* we're done, save away how many cookies the sgl has */
3155 	if (sgl[cnt].dmac_size == 0) {
3156 		ASSERT(cnt < sglinfo->si_max_pages);
3157 		sglinfo->si_sgl_size = cnt;
3158 	} else {
3159 		sglinfo->si_sgl_size = cnt + 1;
3160 	}
3161 }
3162 
3163 static void
3164 rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
3165     rootnex_sglinfo_t *sglinfo)
3166 {
3167 	uint64_t offset;
3168 	uint64_t maxseg;
3169 	uint64_t dvaddr;
3170 	struct dvmaseg *dvs;
3171 	uint64_t paddr;
3172 	uint32_t psize, ssize;
3173 	uint32_t size;
3174 	uint_t cnt;
3175 	int physcontig;
3176 
3177 	ASSERT(dmar_object->dmao_type == DMA_OTYP_DVADDR);
3178 
3179 	/* shortcuts */
3180 	maxseg = sglinfo->si_max_cookie_size;
3181 	size = dmar_object->dmao_size;
3182 
3183 	cnt = 0;
3184 	sglinfo->si_bounce_on_seg = B_FALSE;
3185 
3186 	dvs = dmar_object->dmao_obj.dvma_obj.dv_seg;
3187 	offset = dmar_object->dmao_obj.dvma_obj.dv_off;
3188 	ssize = dvs->dvs_len;
3189 	paddr = dvs->dvs_start;
3190 	paddr += offset;
3191 	psize = MIN(ssize, (maxseg - offset));
3192 	dvaddr = paddr + psize;
3193 	ssize -= psize;
3194 
3195 	sgl[cnt].dmac_laddress = paddr;
3196 	sgl[cnt].dmac_size = psize;
3197 	sgl[cnt].dmac_type = 0;
3198 
3199 	size -= psize;
3200 	while (size > 0) {
3201 		if (ssize == 0) {
3202 			dvs++;
3203 			ssize = dvs->dvs_len;
3204 			dvaddr = dvs->dvs_start;
3205 			physcontig = 0;
3206 		} else
3207 			physcontig = 1;
3208 
3209 		paddr = dvaddr;
3210 		psize = MIN(ssize, maxseg);
3211 		dvaddr += psize;
3212 		ssize -= psize;
3213 
3214 		if (!physcontig || !(paddr & sglinfo->si_segmask) ||
3215 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3216 		    (sgl[cnt].dmac_size == 0)) {
3217 			/*
3218 			 * if we're not already in a new cookie, go to the next
3219 			 * cookie.
3220 			 */
3221 			if (sgl[cnt].dmac_size != 0) {
3222 				cnt++;
3223 			}
3224 
3225 			/* save the cookie information */
3226 			sgl[cnt].dmac_laddress = paddr;
3227 			sgl[cnt].dmac_size = psize;
3228 			sgl[cnt].dmac_type = 0;
3229 		} else {
3230 			sgl[cnt].dmac_size += psize;
3231 
3232 			/*
3233 			 * if this exactly ==  the maximum cookie size, and
3234 			 * it isn't the last cookie, go to the next cookie.
3235 			 */
3236 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3237 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3238 				cnt++;
3239 				sgl[cnt].dmac_laddress = 0;
3240 				sgl[cnt].dmac_size = 0;
3241 				sgl[cnt].dmac_type = 0;
3242 			}
3243 		}
3244 		size -= psize;
3245 	}
3246 
3247 	/* we're done, save away how many cookies the sgl has */
3248 	if (sgl[cnt].dmac_size == 0) {
3249 		sglinfo->si_sgl_size = cnt;
3250 	} else {
3251 		sglinfo->si_sgl_size = cnt + 1;
3252 	}
3253 }
3254 
3255 /*
3256  * rootnex_bind_slowpath()
3257  *    Call in the bind path if the calling driver can't use the sgl without
3258  *    modifying it. We either need to use the copy buffer and/or we will end up
3259  *    with a partial bind.
3260  */
3261 static int
3262 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3263     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3264 {
3265 	rootnex_sglinfo_t *sinfo;
3266 	rootnex_window_t *window;
3267 	ddi_dma_cookie_t *cookie;
3268 	size_t copybuf_used;
3269 	size_t dmac_size;
3270 	boolean_t partial;
3271 	off_t cur_offset;
3272 	page_t *cur_pp;
3273 	major_t mnum;
3274 	int e;
3275 	int i;
3276 
3277 
3278 	sinfo = &dma->dp_sglinfo;
3279 	copybuf_used = 0;
3280 	partial = B_FALSE;
3281 
3282 	/*
3283 	 * If we're using the copybuf, set the copybuf state in dma struct.
3284 	 * Needs to be first since it sets the copy buffer size.
3285 	 */
3286 	if (sinfo->si_copybuf_req != 0) {
3287 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
3288 		if (e != DDI_SUCCESS) {
3289 			return (e);
3290 		}
3291 	} else {
3292 		dma->dp_copybuf_size = 0;
3293 	}
3294 
3295 	/*
3296 	 * Figure out if we need to do a partial mapping. If so, figure out
3297 	 * if we need to trim the buffers when we munge the sgl.
3298 	 */
3299 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
3300 	    (dmao->dmao_size > dma->dp_maxxfer) ||
3301 	    ((unsigned)attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
3302 		dma->dp_partial_required = B_TRUE;
3303 		if (attr->dma_attr_granular != 1) {
3304 			dma->dp_trim_required = B_TRUE;
3305 		}
3306 	} else {
3307 		dma->dp_partial_required = B_FALSE;
3308 		dma->dp_trim_required = B_FALSE;
3309 	}
3310 
3311 	/* If we need to do a partial bind, make sure the driver supports it */
3312 	if (dma->dp_partial_required &&
3313 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3314 
3315 		mnum = ddi_driver_major(dma->dp_dip);
3316 		/*
3317 		 * patchable which allows us to print one warning per major
3318 		 * number.
3319 		 */
3320 		if ((rootnex_bind_warn) &&
3321 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3322 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3323 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3324 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3325 			    "There is a small risk of data corruption in "
3326 			    "particular with large I/Os. The driver should be "
3327 			    "replaced with a corrected version for proper "
3328 			    "system operation. To disable this warning, add "
3329 			    "'set rootnex:rootnex_bind_warn=0' to "
3330 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3331 		}
3332 		return (DDI_DMA_TOOBIG);
3333 	}
3334 
3335 	/*
3336 	 * we might need multiple windows, setup state to handle them. In this
3337 	 * code path, we will have at least one window.
3338 	 */
3339 	e = rootnex_setup_windows(hp, dma, attr, dmao, kmflag);
3340 	if (e != DDI_SUCCESS) {
3341 		rootnex_teardown_copybuf(dma);
3342 		return (e);
3343 	}
3344 
3345 	window = &dma->dp_window[0];
3346 	cookie = &dma->dp_cookies[0];
3347 	cur_offset = 0;
3348 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3349 	if (dmao->dmao_type == DMA_OTYP_PAGES) {
3350 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3351 	}
3352 
3353 	/* loop though all the cookies we got back from get_sgl() */
3354 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3355 		/*
3356 		 * If we're using the copy buffer, check this cookie and setup
3357 		 * its associated copy buffer state. If this cookie uses the
3358 		 * copy buffer, make sure we sync this window during dma_sync.
3359 		 */
3360 		if (dma->dp_copybuf_size > 0) {
3361 			rootnex_setup_cookie(dmao, dma, cookie,
3362 			    cur_offset, &copybuf_used, &cur_pp);
3363 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3364 				window->wd_dosync = B_TRUE;
3365 			}
3366 		}
3367 
3368 		/*
3369 		 * save away the cookie size, since it could be modified in
3370 		 * the windowing code.
3371 		 */
3372 		dmac_size = cookie->dmac_size;
3373 
3374 		/* if we went over max copybuf size */
3375 		if (dma->dp_copybuf_size &&
3376 		    (copybuf_used > dma->dp_copybuf_size)) {
3377 			partial = B_TRUE;
3378 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3379 			    cookie, cur_offset, &copybuf_used);
3380 			if (e != DDI_SUCCESS) {
3381 				rootnex_teardown_copybuf(dma);
3382 				rootnex_teardown_windows(dma);
3383 				return (e);
3384 			}
3385 
3386 			/*
3387 			 * if the coookie uses the copy buffer, make sure the
3388 			 * new window we just moved to is set to sync.
3389 			 */
3390 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3391 				window->wd_dosync = B_TRUE;
3392 			}
3393 			ROOTNEX_DPROBE1(rootnex__copybuf__window, dev_info_t *,
3394 			    dma->dp_dip);
3395 
3396 		/* if the cookie cnt == max sgllen, move to the next window */
3397 		} else if (window->wd_cookie_cnt >=
3398 		    (unsigned)attr->dma_attr_sgllen) {
3399 			partial = B_TRUE;
3400 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3401 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3402 			    cookie, attr, cur_offset);
3403 			if (e != DDI_SUCCESS) {
3404 				rootnex_teardown_copybuf(dma);
3405 				rootnex_teardown_windows(dma);
3406 				return (e);
3407 			}
3408 
3409 			/*
3410 			 * if the coookie uses the copy buffer, make sure the
3411 			 * new window we just moved to is set to sync.
3412 			 */
3413 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3414 				window->wd_dosync = B_TRUE;
3415 			}
3416 			ROOTNEX_DPROBE1(rootnex__sgllen__window, dev_info_t *,
3417 			    dma->dp_dip);
3418 
3419 		/* else if we will be over maxxfer */
3420 		} else if ((window->wd_size + dmac_size) >
3421 		    dma->dp_maxxfer) {
3422 			partial = B_TRUE;
3423 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3424 			    cookie);
3425 			if (e != DDI_SUCCESS) {
3426 				rootnex_teardown_copybuf(dma);
3427 				rootnex_teardown_windows(dma);
3428 				return (e);
3429 			}
3430 
3431 			/*
3432 			 * if the coookie uses the copy buffer, make sure the
3433 			 * new window we just moved to is set to sync.
3434 			 */
3435 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3436 				window->wd_dosync = B_TRUE;
3437 			}
3438 			ROOTNEX_DPROBE1(rootnex__maxxfer__window, dev_info_t *,
3439 			    dma->dp_dip);
3440 
3441 		/* else this cookie fits in the current window */
3442 		} else {
3443 			window->wd_cookie_cnt++;
3444 			window->wd_size += dmac_size;
3445 		}
3446 
3447 		/* track our offset into the buffer, go to the next cookie */
3448 		ASSERT(dmac_size <= dmao->dmao_size);
3449 		ASSERT(cookie->dmac_size <= dmac_size);
3450 		cur_offset += dmac_size;
3451 		cookie++;
3452 	}
3453 
3454 	/* if we ended up with a zero sized window in the end, clean it up */
3455 	if (window->wd_size == 0) {
3456 		hp->dmai_nwin--;
3457 		window--;
3458 	}
3459 
3460 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3461 
3462 	if (!partial) {
3463 		return (DDI_DMA_MAPPED);
3464 	}
3465 
3466 	ASSERT(dma->dp_partial_required);
3467 	return (DDI_DMA_PARTIAL_MAP);
3468 }
3469 
3470 /*
3471  * rootnex_setup_copybuf()
3472  *    Called in bind slowpath. Figures out if we're going to use the copy
3473  *    buffer, and if we do, sets up the basic state to handle it.
3474  */
3475 static int
3476 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3477     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3478 {
3479 	rootnex_sglinfo_t *sinfo;
3480 	ddi_dma_attr_t lattr;
3481 	size_t max_copybuf;
3482 	int cansleep;
3483 	int e;
3484 #if !defined(__amd64)
3485 	int vmflag;
3486 #endif
3487 
3488 	ASSERT(!dma->dp_dvma_used);
3489 
3490 	sinfo = &dma->dp_sglinfo;
3491 
3492 	/* read this first so it's consistent through the routine  */
3493 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3494 
3495 	/* We need to call into the rootnex on ddi_dma_sync() */
3496 	hp->dmai_rflags &= ~DMP_NOSYNC;
3497 
3498 	/* make sure the copybuf size <= the max size */
3499 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3500 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3501 
3502 #if !defined(__amd64)
3503 	/*
3504 	 * if we don't have kva space to copy to/from, allocate the KVA space
3505 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3506 	 * the 64-bit kernel.
3507 	 */
3508 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3509 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3510 
3511 		/* convert the sleep flags */
3512 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3513 			vmflag = VM_SLEEP;
3514 		} else {
3515 			vmflag = VM_NOSLEEP;
3516 		}
3517 
3518 		/* allocate Kernel VA space that we can bcopy to/from */
3519 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3520 		    vmflag);
3521 		if (dma->dp_kva == NULL) {
3522 			return (DDI_DMA_NORESOURCES);
3523 		}
3524 	}
3525 #endif
3526 
3527 	/* convert the sleep flags */
3528 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3529 		cansleep = 1;
3530 	} else {
3531 		cansleep = 0;
3532 	}
3533 
3534 	/*
3535 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3536 	 * engine limits, so we can't use kmem_alloc... We don't need
3537 	 * contiguous memory (sgllen) since we will be forcing windows on
3538 	 * sgllen anyway.
3539 	 */
3540 	lattr = *attr;
3541 	lattr.dma_attr_align = MMU_PAGESIZE;
3542 	lattr.dma_attr_sgllen = -1;	/* no limit */
3543 	/*
3544 	 * if we're using the copy buffer because of seg, use that for our
3545 	 * upper address limit.
3546 	 */
3547 	if (sinfo->si_bounce_on_seg) {
3548 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3549 	}
3550 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3551 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3552 	if (e != DDI_SUCCESS) {
3553 #if !defined(__amd64)
3554 		if (dma->dp_kva != NULL) {
3555 			vmem_free(heap_arena, dma->dp_kva,
3556 			    dma->dp_copybuf_size);
3557 		}
3558 #endif
3559 		return (DDI_DMA_NORESOURCES);
3560 	}
3561 
3562 	ROOTNEX_DPROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3563 	    size_t, dma->dp_copybuf_size);
3564 
3565 	return (DDI_SUCCESS);
3566 }
3567 
3568 
3569 /*
3570  * rootnex_setup_windows()
3571  *    Called in bind slowpath to setup the window state. We always have windows
3572  *    in the slowpath. Even if the window count = 1.
3573  */
3574 static int
3575 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3576     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3577 {
3578 	rootnex_window_t *windowp;
3579 	rootnex_sglinfo_t *sinfo;
3580 	size_t copy_state_size;
3581 	size_t win_state_size;
3582 	size_t state_available;
3583 	size_t space_needed;
3584 	uint_t copybuf_win;
3585 	uint_t maxxfer_win;
3586 	size_t space_used;
3587 	uint_t sglwin;
3588 
3589 
3590 	sinfo = &dma->dp_sglinfo;
3591 
3592 	dma->dp_current_win = 0;
3593 	hp->dmai_nwin = 0;
3594 
3595 	/* If we don't need to do a partial, we only have one window */
3596 	if (!dma->dp_partial_required) {
3597 		dma->dp_max_win = 1;
3598 
3599 	/*
3600 	 * we need multiple windows, need to figure out the worse case number
3601 	 * of windows.
3602 	 */
3603 	} else {
3604 		/*
3605 		 * if we need windows because we need more copy buffer that
3606 		 * we allow, the worse case number of windows we could need
3607 		 * here would be (copybuf space required / copybuf space that
3608 		 * we have) plus one for remainder, and plus 2 to handle the
3609 		 * extra pages on the trim for the first and last pages of the
3610 		 * buffer (a page is the minimum window size so under the right
3611 		 * attr settings, you could have a window for each page).
3612 		 * The last page will only be hit here if the size is not a
3613 		 * multiple of the granularity (which theoretically shouldn't
3614 		 * be the case but never has been enforced, so we could have
3615 		 * broken things without it).
3616 		 */
3617 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3618 			ASSERT(dma->dp_copybuf_size > 0);
3619 			copybuf_win = (sinfo->si_copybuf_req /
3620 			    dma->dp_copybuf_size) + 1 + 2;
3621 		} else {
3622 			copybuf_win = 0;
3623 		}
3624 
3625 		/*
3626 		 * if we need windows because we have more cookies than the H/W
3627 		 * can handle, the number of windows we would need here would
3628 		 * be (cookie count / cookies count H/W supports minus 1[for
3629 		 * trim]) plus one for remainder.
3630 		 */
3631 		if ((unsigned)attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3632 			sglwin = (sinfo->si_sgl_size /
3633 			    (attr->dma_attr_sgllen - 1)) + 1;
3634 		} else {
3635 			sglwin = 0;
3636 		}
3637 
3638 		/*
3639 		 * if we need windows because we're binding more memory than the
3640 		 * H/W can transfer at once, the number of windows we would need
3641 		 * here would be (xfer count / max xfer H/W supports) plus one
3642 		 * for remainder, and plus 2 to handle the extra pages on the
3643 		 * trim (see above comment about trim)
3644 		 */
3645 		if (dmao->dmao_size > dma->dp_maxxfer) {
3646 			maxxfer_win = (dmao->dmao_size /
3647 			    dma->dp_maxxfer) + 1 + 2;
3648 		} else {
3649 			maxxfer_win = 0;
3650 		}
3651 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3652 		ASSERT(dma->dp_max_win > 0);
3653 	}
3654 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3655 
3656 	/*
3657 	 * Get space for window and potential copy buffer state. Before we
3658 	 * go and allocate memory, see if we can get away with using what's
3659 	 * left in the pre-allocted state or the dynamically allocated sgl.
3660 	 */
3661 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3662 	    sizeof (ddi_dma_cookie_t));
3663 
3664 	/* if we dynamically allocated space for the cookies */
3665 	if (dma->dp_need_to_free_cookie) {
3666 		/* if we have more space in the pre-allocted buffer, use it */
3667 		ASSERT(space_used <= dma->dp_cookie_size);
3668 		if ((dma->dp_cookie_size - space_used) <=
3669 		    rootnex_state->r_prealloc_size) {
3670 			state_available = rootnex_state->r_prealloc_size;
3671 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3672 
3673 		/*
3674 		 * else, we have more free space in the dynamically allocated
3675 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3676 		 * didn't need a lot of cookies.
3677 		 */
3678 		} else {
3679 			state_available = dma->dp_cookie_size - space_used;
3680 			windowp = (rootnex_window_t *)
3681 			    &dma->dp_cookies[sinfo->si_sgl_size];
3682 		}
3683 
3684 	/* we used the pre-alloced buffer */
3685 	} else {
3686 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3687 		state_available = rootnex_state->r_prealloc_size - space_used;
3688 		windowp = (rootnex_window_t *)
3689 		    &dma->dp_cookies[sinfo->si_sgl_size];
3690 	}
3691 
3692 	/*
3693 	 * figure out how much state we need to track the copy buffer. Add an
3694 	 * addition 8 bytes for pointer alignemnt later.
3695 	 */
3696 	if (dma->dp_copybuf_size > 0) {
3697 		copy_state_size = sinfo->si_max_pages *
3698 		    sizeof (rootnex_pgmap_t);
3699 	} else {
3700 		copy_state_size = 0;
3701 	}
3702 	/* add an additional 8 bytes for pointer alignment */
3703 	space_needed = win_state_size + copy_state_size + 0x8;
3704 
3705 	/* if we have enough space already, use it */
3706 	if (state_available >= space_needed) {
3707 		dma->dp_window = windowp;
3708 		dma->dp_need_to_free_window = B_FALSE;
3709 
3710 	/* not enough space, need to allocate more. */
3711 	} else {
3712 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3713 		if (dma->dp_window == NULL) {
3714 			return (DDI_DMA_NORESOURCES);
3715 		}
3716 		dma->dp_need_to_free_window = B_TRUE;
3717 		dma->dp_window_size = space_needed;
3718 		ROOTNEX_DPROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3719 		    dma->dp_dip, size_t, space_needed);
3720 	}
3721 
3722 	/*
3723 	 * we allocate copy buffer state and window state at the same time.
3724 	 * setup our copy buffer state pointers. Make sure it's aligned.
3725 	 */
3726 	if (dma->dp_copybuf_size > 0) {
3727 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3728 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3729 
3730 #if !defined(__amd64)
3731 		/*
3732 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3733 		 * false/NULL. Should be quicker to bzero vs loop and set.
3734 		 */
3735 		bzero(dma->dp_pgmap, copy_state_size);
3736 #endif
3737 	} else {
3738 		dma->dp_pgmap = NULL;
3739 	}
3740 
3741 	return (DDI_SUCCESS);
3742 }
3743 
3744 
3745 /*
3746  * rootnex_teardown_copybuf()
3747  *    cleans up after rootnex_setup_copybuf()
3748  */
3749 static void
3750 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3751 {
3752 #if !defined(__amd64)
3753 	int i;
3754 
3755 	/*
3756 	 * if we allocated kernel heap VMEM space, go through all the pages and
3757 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3758 	 * arena. Then free the VMEM space.
3759 	 */
3760 	if (dma->dp_kva != NULL) {
3761 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3762 			if (dma->dp_pgmap[i].pm_mapped) {
3763 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3764 				    MMU_PAGESIZE, HAT_UNLOAD);
3765 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3766 			}
3767 		}
3768 
3769 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3770 	}
3771 
3772 #endif
3773 
3774 	/* if we allocated a copy buffer, free it */
3775 	if (dma->dp_cbaddr != NULL) {
3776 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3777 	}
3778 }
3779 
3780 
3781 /*
3782  * rootnex_teardown_windows()
3783  *    cleans up after rootnex_setup_windows()
3784  */
3785 static void
3786 rootnex_teardown_windows(rootnex_dma_t *dma)
3787 {
3788 	/*
3789 	 * if we had to allocate window state on the last bind (because we
3790 	 * didn't have enough pre-allocated space in the handle), free it.
3791 	 */
3792 	if (dma->dp_need_to_free_window) {
3793 		kmem_free(dma->dp_window, dma->dp_window_size);
3794 	}
3795 }
3796 
3797 
3798 /*
3799  * rootnex_init_win()
3800  *    Called in bind slow path during creation of a new window. Initializes
3801  *    window state to default values.
3802  */
3803 /*ARGSUSED*/
3804 static void
3805 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3806     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3807 {
3808 	hp->dmai_nwin++;
3809 	window->wd_dosync = B_FALSE;
3810 	window->wd_offset = cur_offset;
3811 	window->wd_size = 0;
3812 	window->wd_first_cookie = cookie;
3813 	window->wd_cookie_cnt = 0;
3814 	window->wd_trim.tr_trim_first = B_FALSE;
3815 	window->wd_trim.tr_trim_last = B_FALSE;
3816 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3817 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3818 #if !defined(__amd64)
3819 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3820 #endif
3821 }
3822 
3823 
3824 /*
3825  * rootnex_setup_cookie()
3826  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3827  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3828  *    out if it uses the copy buffer, and if it does, save away everything we'll
3829  *    need during sync.
3830  */
3831 static void
3832 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3833     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3834     page_t **cur_pp)
3835 {
3836 	boolean_t copybuf_sz_power_2;
3837 	rootnex_sglinfo_t *sinfo;
3838 	paddr_t paddr;
3839 	uint_t pidx;
3840 	uint_t pcnt;
3841 	off_t poff;
3842 #if defined(__amd64)
3843 	pfn_t pfn;
3844 #else
3845 	page_t **pplist;
3846 #endif
3847 
3848 	ASSERT(dmar_object->dmao_type != DMA_OTYP_DVADDR);
3849 
3850 	sinfo = &dma->dp_sglinfo;
3851 
3852 	/*
3853 	 * Calculate the page index relative to the start of the buffer. The
3854 	 * index to the current page for our buffer is the offset into the
3855 	 * first page of the buffer plus our current offset into the buffer
3856 	 * itself, shifted of course...
3857 	 */
3858 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3859 	ASSERT(pidx < sinfo->si_max_pages);
3860 
3861 	/* if this cookie uses the copy buffer */
3862 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3863 		/*
3864 		 * NOTE: we know that since this cookie uses the copy buffer, it
3865 		 * is <= MMU_PAGESIZE.
3866 		 */
3867 
3868 		/*
3869 		 * get the offset into the page. For the 64-bit kernel, get the
3870 		 * pfn which we'll use with seg kpm.
3871 		 */
3872 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3873 #if defined(__amd64)
3874 		/* mfn_to_pfn() is a NOP on i86pc */
3875 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3876 #endif /* __amd64 */
3877 
3878 		/* figure out if the copybuf size is a power of 2 */
3879 		if (!ISP2(dma->dp_copybuf_size)) {
3880 			copybuf_sz_power_2 = B_FALSE;
3881 		} else {
3882 			copybuf_sz_power_2 = B_TRUE;
3883 		}
3884 
3885 		/* This page uses the copy buffer */
3886 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3887 
3888 		/*
3889 		 * save the copy buffer KVA that we'll use with this page.
3890 		 * if we still fit within the copybuf, it's a simple add.
3891 		 * otherwise, we need to wrap over using & or % accordingly.
3892 		 */
3893 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3894 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3895 			    *copybuf_used;
3896 		} else {
3897 			if (copybuf_sz_power_2) {
3898 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3899 				    (uintptr_t)dma->dp_cbaddr +
3900 				    (*copybuf_used &
3901 				    (dma->dp_copybuf_size - 1)));
3902 			} else {
3903 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3904 				    (uintptr_t)dma->dp_cbaddr +
3905 				    (*copybuf_used % dma->dp_copybuf_size));
3906 			}
3907 		}
3908 
3909 		/*
3910 		 * over write the cookie physical address with the address of
3911 		 * the physical address of the copy buffer page that we will
3912 		 * use.
3913 		 */
3914 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3915 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3916 
3917 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
3918 
3919 		/* if we have a kernel VA, it's easy, just save that address */
3920 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3921 		    (sinfo->si_asp == &kas)) {
3922 			/*
3923 			 * save away the page aligned virtual address of the
3924 			 * driver buffer. Offsets are handled in the sync code.
3925 			 */
3926 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3927 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3928 			    & MMU_PAGEMASK);
3929 #if !defined(__amd64)
3930 			/*
3931 			 * we didn't need to, and will never need to map this
3932 			 * page.
3933 			 */
3934 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3935 #endif
3936 
3937 		/* we don't have a kernel VA. We need one for the bcopy. */
3938 		} else {
3939 #if defined(__amd64)
3940 			/*
3941 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3942 			 * get a Kernel VA for the corresponding pfn.
3943 			 */
3944 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3945 #else
3946 			/*
3947 			 * for the 32-bit kernel, this is a pain. First we'll
3948 			 * save away the page_t or user VA for this page. This
3949 			 * is needed in rootnex_dma_win() when we switch to a
3950 			 * new window which requires us to re-map the copy
3951 			 * buffer.
3952 			 */
3953 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3954 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3955 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3956 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3957 			} else if (pplist != NULL) {
3958 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3959 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3960 			} else {
3961 				dma->dp_pgmap[pidx].pm_pp = NULL;
3962 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3963 				    (((uintptr_t)
3964 				    dmar_object->dmao_obj.virt_obj.v_addr +
3965 				    cur_offset) & MMU_PAGEMASK);
3966 			}
3967 
3968 			/*
3969 			 * save away the page aligned virtual address which was
3970 			 * allocated from the kernel heap arena (taking into
3971 			 * account if we need more copy buffer than we alloced
3972 			 * and use multiple windows to handle this, i.e. &,%).
3973 			 * NOTE: there isn't and physical memory backing up this
3974 			 * virtual address space currently.
3975 			 */
3976 			if ((*copybuf_used + MMU_PAGESIZE) <=
3977 			    dma->dp_copybuf_size) {
3978 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3979 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3980 				    MMU_PAGEMASK);
3981 			} else {
3982 				if (copybuf_sz_power_2) {
3983 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3984 					    (((uintptr_t)dma->dp_kva +
3985 					    (*copybuf_used &
3986 					    (dma->dp_copybuf_size - 1))) &
3987 					    MMU_PAGEMASK);
3988 				} else {
3989 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3990 					    (((uintptr_t)dma->dp_kva +
3991 					    (*copybuf_used %
3992 					    dma->dp_copybuf_size)) &
3993 					    MMU_PAGEMASK);
3994 				}
3995 			}
3996 
3997 			/*
3998 			 * if we haven't used up the available copy buffer yet,
3999 			 * map the kva to the physical page.
4000 			 */
4001 			if (!dma->dp_cb_remaping && ((*copybuf_used +
4002 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
4003 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
4004 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
4005 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
4006 					    dma->dp_pgmap[pidx].pm_kaddr);
4007 				} else {
4008 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
4009 					    sinfo->si_asp,
4010 					    dma->dp_pgmap[pidx].pm_kaddr);
4011 				}
4012 
4013 			/*
4014 			 * we've used up the available copy buffer, this page
4015 			 * will have to be mapped during rootnex_dma_win() when
4016 			 * we switch to a new window which requires a re-map
4017 			 * the copy buffer. (32-bit kernel only)
4018 			 */
4019 			} else {
4020 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4021 			}
4022 #endif
4023 			/* go to the next page_t */
4024 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4025 				*cur_pp = (*cur_pp)->p_next;
4026 			}
4027 		}
4028 
4029 		/* add to the copy buffer count */
4030 		*copybuf_used += MMU_PAGESIZE;
4031 
4032 	/*
4033 	 * This cookie doesn't use the copy buffer. Walk through the pages this
4034 	 * cookie occupies to reflect this.
4035 	 */
4036 	} else {
4037 		/*
4038 		 * figure out how many pages the cookie occupies. We need to
4039 		 * use the original page offset of the buffer and the cookies
4040 		 * offset in the buffer to do this.
4041 		 */
4042 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
4043 		pcnt = mmu_btopr(cookie->dmac_size + poff);
4044 
4045 		while (pcnt > 0) {
4046 #if !defined(__amd64)
4047 			/*
4048 			 * the 32-bit kernel doesn't have seg kpm, so we need
4049 			 * to map in the driver buffer (if it didn't come down
4050 			 * with a kernel VA) on the fly. Since this page doesn't
4051 			 * use the copy buffer, it's not, or will it ever, have
4052 			 * to be mapped in.
4053 			 */
4054 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4055 #endif
4056 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
4057 
4058 			/*
4059 			 * we need to update pidx and cur_pp or we'll loose
4060 			 * track of where we are.
4061 			 */
4062 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4063 				*cur_pp = (*cur_pp)->p_next;
4064 			}
4065 			pidx++;
4066 			pcnt--;
4067 		}
4068 	}
4069 }
4070 
4071 
4072 /*
4073  * rootnex_sgllen_window_boundary()
4074  *    Called in the bind slow path when the next cookie causes us to exceed (in
4075  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
4076  *    length supported by the DMA H/W.
4077  */
4078 static int
4079 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4080     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
4081     off_t cur_offset)
4082 {
4083 	off_t new_offset;
4084 	size_t trim_sz;
4085 	off_t coffset;
4086 
4087 
4088 	/*
4089 	 * if we know we'll never have to trim, it's pretty easy. Just move to
4090 	 * the next window and init it. We're done.
4091 	 */
4092 	if (!dma->dp_trim_required) {
4093 		(*windowp)++;
4094 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4095 		(*windowp)->wd_cookie_cnt++;
4096 		(*windowp)->wd_size = cookie->dmac_size;
4097 		return (DDI_SUCCESS);
4098 	}
4099 
4100 	/* figure out how much we need to trim from the window */
4101 	ASSERT(attr->dma_attr_granular != 0);
4102 	if (dma->dp_granularity_power_2) {
4103 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
4104 	} else {
4105 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
4106 	}
4107 
4108 	/* The window's a whole multiple of granularity. We're done */
4109 	if (trim_sz == 0) {
4110 		(*windowp)++;
4111 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4112 		(*windowp)->wd_cookie_cnt++;
4113 		(*windowp)->wd_size = cookie->dmac_size;
4114 		return (DDI_SUCCESS);
4115 	}
4116 
4117 	/*
4118 	 * The window's not a whole multiple of granularity, since we know this
4119 	 * is due to the sgllen, we need to go back to the last cookie and trim
4120 	 * that one, add the left over part of the old cookie into the new
4121 	 * window, and then add in the new cookie into the new window.
4122 	 */
4123 
4124 	/*
4125 	 * make sure the driver isn't making us do something bad... Trimming and
4126 	 * sgllen == 1 don't go together.
4127 	 */
4128 	if (attr->dma_attr_sgllen == 1) {
4129 		return (DDI_DMA_NOMAPPING);
4130 	}
4131 
4132 	/*
4133 	 * first, setup the current window to account for the trim. Need to go
4134 	 * back to the last cookie for this.
4135 	 */
4136 	cookie--;
4137 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4138 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4139 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4140 	ASSERT(cookie->dmac_size > trim_sz);
4141 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4142 	(*windowp)->wd_size -= trim_sz;
4143 
4144 	/* save the buffer offsets for the next window */
4145 	coffset = cookie->dmac_size - trim_sz;
4146 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4147 
4148 	/*
4149 	 * set this now in case this is the first window. all other cases are
4150 	 * set in dma_win()
4151 	 */
4152 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4153 
4154 	/*
4155 	 * initialize the next window using what's left over in the previous
4156 	 * cookie.
4157 	 */
4158 	(*windowp)++;
4159 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4160 	(*windowp)->wd_cookie_cnt++;
4161 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4162 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4163 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4164 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4165 		(*windowp)->wd_dosync = B_TRUE;
4166 	}
4167 
4168 	/*
4169 	 * now go back to the current cookie and add it to the new window. set
4170 	 * the new window size to the what was left over from the previous
4171 	 * cookie and what's in the current cookie.
4172 	 */
4173 	cookie++;
4174 	(*windowp)->wd_cookie_cnt++;
4175 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4176 
4177 	/*
4178 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
4179 	 * a max size of maxxfer). Handle that case.
4180 	 */
4181 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
4182 		/*
4183 		 * maxxfer is already a whole multiple of granularity, and this
4184 		 * trim will be <= the previous trim (since a cookie can't be
4185 		 * larger than maxxfer). Make things simple here.
4186 		 */
4187 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
4188 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4189 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4190 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4191 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4192 		(*windowp)->wd_size -= trim_sz;
4193 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
4194 
4195 		/* save the buffer offsets for the next window */
4196 		coffset = cookie->dmac_size - trim_sz;
4197 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4198 
4199 		/* setup the next window */
4200 		(*windowp)++;
4201 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4202 		(*windowp)->wd_cookie_cnt++;
4203 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4204 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4205 		    coffset;
4206 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4207 	}
4208 
4209 	return (DDI_SUCCESS);
4210 }
4211 
4212 
4213 /*
4214  * rootnex_copybuf_window_boundary()
4215  *    Called in bind slowpath when we get to a window boundary because we used
4216  *    up all the copy buffer that we have.
4217  */
4218 static int
4219 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4220     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
4221     size_t *copybuf_used)
4222 {
4223 	rootnex_sglinfo_t *sinfo;
4224 	off_t new_offset;
4225 	size_t trim_sz;
4226 	paddr_t paddr;
4227 	off_t coffset;
4228 	uint_t pidx;
4229 	off_t poff;
4230 
4231 	pidx = 0;
4232 	sinfo = &dma->dp_sglinfo;
4233 
4234 	/*
4235 	 * the copy buffer should be a whole multiple of page size. We know that
4236 	 * this cookie is <= MMU_PAGESIZE.
4237 	 */
4238 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
4239 
4240 	/*
4241 	 * from now on, all new windows in this bind need to be re-mapped during
4242 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
4243 	 * space...
4244 	 */
4245 #if !defined(__amd64)
4246 	dma->dp_cb_remaping = B_TRUE;
4247 #endif
4248 
4249 	/* reset copybuf used */
4250 	*copybuf_used = 0;
4251 
4252 	/*
4253 	 * if we don't have to trim (since granularity is set to 1), go to the
4254 	 * next window and add the current cookie to it. We know the current
4255 	 * cookie uses the copy buffer since we're in this code path.
4256 	 */
4257 	if (!dma->dp_trim_required) {
4258 		(*windowp)++;
4259 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4260 
4261 		/* Add this cookie to the new window */
4262 		(*windowp)->wd_cookie_cnt++;
4263 		(*windowp)->wd_size += cookie->dmac_size;
4264 		*copybuf_used += MMU_PAGESIZE;
4265 		return (DDI_SUCCESS);
4266 	}
4267 
4268 	/*
4269 	 * *** may need to trim, figure it out.
4270 	 */
4271 
4272 	/* figure out how much we need to trim from the window */
4273 	if (dma->dp_granularity_power_2) {
4274 		trim_sz = (*windowp)->wd_size &
4275 		    (hp->dmai_attr.dma_attr_granular - 1);
4276 	} else {
4277 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
4278 	}
4279 
4280 	/*
4281 	 * if the window's a whole multiple of granularity, go to the next
4282 	 * window, init it, then add in the current cookie. We know the current
4283 	 * cookie uses the copy buffer since we're in this code path.
4284 	 */
4285 	if (trim_sz == 0) {
4286 		(*windowp)++;
4287 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4288 
4289 		/* Add this cookie to the new window */
4290 		(*windowp)->wd_cookie_cnt++;
4291 		(*windowp)->wd_size += cookie->dmac_size;
4292 		*copybuf_used += MMU_PAGESIZE;
4293 		return (DDI_SUCCESS);
4294 	}
4295 
4296 	/*
4297 	 * *** We figured it out, we definitly need to trim
4298 	 */
4299 
4300 	/*
4301 	 * make sure the driver isn't making us do something bad...
4302 	 * Trimming and sgllen == 1 don't go together.
4303 	 */
4304 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4305 		return (DDI_DMA_NOMAPPING);
4306 	}
4307 
4308 	/*
4309 	 * first, setup the current window to account for the trim. Need to go
4310 	 * back to the last cookie for this. Some of the last cookie will be in
4311 	 * the current window, and some of the last cookie will be in the new
4312 	 * window. All of the current cookie will be in the new window.
4313 	 */
4314 	cookie--;
4315 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4316 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4317 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4318 	ASSERT(cookie->dmac_size > trim_sz);
4319 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4320 	(*windowp)->wd_size -= trim_sz;
4321 
4322 	/*
4323 	 * we're trimming the last cookie (not the current cookie). So that
4324 	 * last cookie may have or may not have been using the copy buffer (
4325 	 * we know the cookie passed in uses the copy buffer since we're in
4326 	 * this code path).
4327 	 *
4328 	 * If the last cookie doesn't use the copy buffer, nothing special to
4329 	 * do. However, if it does uses the copy buffer, it will be both the
4330 	 * last page in the current window and the first page in the next
4331 	 * window. Since we are reusing the copy buffer (and KVA space on the
4332 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4333 	 * current window, and the start of the copy buffer in the next window.
4334 	 * Track that info... The cookie physical address was already set to
4335 	 * the copy buffer physical address in setup_cookie..
4336 	 */
4337 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4338 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4339 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4340 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4341 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4342 		(*windowp)->wd_trim.tr_last_cbaddr =
4343 		    dma->dp_pgmap[pidx].pm_cbaddr;
4344 #if !defined(__amd64)
4345 		(*windowp)->wd_trim.tr_last_kaddr =
4346 		    dma->dp_pgmap[pidx].pm_kaddr;
4347 #endif
4348 	}
4349 
4350 	/* save the buffer offsets for the next window */
4351 	coffset = cookie->dmac_size - trim_sz;
4352 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4353 
4354 	/*
4355 	 * set this now in case this is the first window. all other cases are
4356 	 * set in dma_win()
4357 	 */
4358 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4359 
4360 	/*
4361 	 * initialize the next window using what's left over in the previous
4362 	 * cookie.
4363 	 */
4364 	(*windowp)++;
4365 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4366 	(*windowp)->wd_cookie_cnt++;
4367 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4368 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4369 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4370 
4371 	/*
4372 	 * again, we're tracking if the last cookie uses the copy buffer.
4373 	 * read the comment above for more info on why we need to track
4374 	 * additional state.
4375 	 *
4376 	 * For the first cookie in the new window, we need reset the physical
4377 	 * address to DMA into to the start of the copy buffer plus any
4378 	 * initial page offset which may be present.
4379 	 */
4380 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4381 		(*windowp)->wd_dosync = B_TRUE;
4382 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4383 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4384 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4385 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4386 
4387 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4388 		    poff;
4389 		(*windowp)->wd_trim.tr_first_paddr =
4390 		    ROOTNEX_PADDR_TO_RBASE(paddr);
4391 
4392 #if !defined(__amd64)
4393 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4394 #endif
4395 		/* account for the cookie copybuf usage in the new window */
4396 		*copybuf_used += MMU_PAGESIZE;
4397 
4398 		/*
4399 		 * every piece of code has to have a hack, and here is this
4400 		 * ones :-)
4401 		 *
4402 		 * There is a complex interaction between setup_cookie and the
4403 		 * copybuf window boundary. The complexity had to be in either
4404 		 * the maxxfer window, or the copybuf window, and I chose the
4405 		 * copybuf code.
4406 		 *
4407 		 * So in this code path, we have taken the last cookie,
4408 		 * virtually broken it in half due to the trim, and it happens
4409 		 * to use the copybuf which further complicates life. At the
4410 		 * same time, we have already setup the current cookie, which
4411 		 * is now wrong. More background info: the current cookie uses
4412 		 * the copybuf, so it is only a page long max. So we need to
4413 		 * fix the current cookies copy buffer address, physical
4414 		 * address, and kva for the 32-bit kernel. We due this by
4415 		 * bumping them by page size (of course, we can't due this on
4416 		 * the physical address since the copy buffer may not be
4417 		 * physically contiguous).
4418 		 */
4419 		cookie++;
4420 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4421 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4422 
4423 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4424 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4425 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
4426 
4427 #if !defined(__amd64)
4428 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4429 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4430 #endif
4431 	} else {
4432 		/* go back to the current cookie */
4433 		cookie++;
4434 	}
4435 
4436 	/*
4437 	 * add the current cookie to the new window. set the new window size to
4438 	 * the what was left over from the previous cookie and what's in the
4439 	 * current cookie.
4440 	 */
4441 	(*windowp)->wd_cookie_cnt++;
4442 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4443 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4444 
4445 	/*
4446 	 * we know that the cookie passed in always uses the copy buffer. We
4447 	 * wouldn't be here if it didn't.
4448 	 */
4449 	*copybuf_used += MMU_PAGESIZE;
4450 
4451 	return (DDI_SUCCESS);
4452 }
4453 
4454 
4455 /*
4456  * rootnex_maxxfer_window_boundary()
4457  *    Called in bind slowpath when we get to a window boundary because we will
4458  *    go over maxxfer.
4459  */
4460 static int
4461 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4462     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4463 {
4464 	size_t dmac_size;
4465 	off_t new_offset;
4466 	size_t trim_sz;
4467 	off_t coffset;
4468 
4469 
4470 	/*
4471 	 * calculate how much we have to trim off of the current cookie to equal
4472 	 * maxxfer. We don't have to account for granularity here since our
4473 	 * maxxfer already takes that into account.
4474 	 */
4475 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4476 	ASSERT(trim_sz <= cookie->dmac_size);
4477 	ASSERT(trim_sz <= dma->dp_maxxfer);
4478 
4479 	/* save cookie size since we need it later and we might change it */
4480 	dmac_size = cookie->dmac_size;
4481 
4482 	/*
4483 	 * if we're not trimming the entire cookie, setup the current window to
4484 	 * account for the trim.
4485 	 */
4486 	if (trim_sz < cookie->dmac_size) {
4487 		(*windowp)->wd_cookie_cnt++;
4488 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4489 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4490 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4491 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4492 		(*windowp)->wd_size = dma->dp_maxxfer;
4493 
4494 		/*
4495 		 * set the adjusted cookie size now in case this is the first
4496 		 * window. All other windows are taken care of in get win
4497 		 */
4498 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4499 	}
4500 
4501 	/*
4502 	 * coffset is the current offset within the cookie, new_offset is the
4503 	 * current offset with the entire buffer.
4504 	 */
4505 	coffset = dmac_size - trim_sz;
4506 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4507 
4508 	/* initialize the next window */
4509 	(*windowp)++;
4510 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4511 	(*windowp)->wd_cookie_cnt++;
4512 	(*windowp)->wd_size = trim_sz;
4513 	if (trim_sz < dmac_size) {
4514 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4515 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4516 		    coffset;
4517 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4518 	}
4519 
4520 	return (DDI_SUCCESS);
4521 }
4522 
4523 
4524 /*ARGSUSED*/
4525 static int
4526 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4527     off_t off, size_t len, uint_t cache_flags)
4528 {
4529 	rootnex_sglinfo_t *sinfo;
4530 	rootnex_pgmap_t *cbpage;
4531 	rootnex_window_t *win;
4532 	ddi_dma_impl_t *hp;
4533 	rootnex_dma_t *dma;
4534 	caddr_t fromaddr;
4535 	caddr_t toaddr;
4536 	uint_t psize;
4537 	off_t offset;
4538 	uint_t pidx;
4539 	size_t size;
4540 	off_t poff;
4541 	int e;
4542 
4543 
4544 	hp = (ddi_dma_impl_t *)handle;
4545 	dma = (rootnex_dma_t *)hp->dmai_private;
4546 	sinfo = &dma->dp_sglinfo;
4547 
4548 	/*
4549 	 * if we don't have any windows, we don't need to sync. A copybuf
4550 	 * will cause us to have at least one window.
4551 	 */
4552 	if (dma->dp_window == NULL) {
4553 		return (DDI_SUCCESS);
4554 	}
4555 
4556 	/* This window may not need to be sync'd */
4557 	win = &dma->dp_window[dma->dp_current_win];
4558 	if (!win->wd_dosync) {
4559 		return (DDI_SUCCESS);
4560 	}
4561 
4562 	/* handle off and len special cases */
4563 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4564 		offset = win->wd_offset;
4565 	} else {
4566 		offset = off;
4567 	}
4568 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4569 		size = win->wd_size;
4570 	} else {
4571 		size = len;
4572 	}
4573 
4574 	/* check the sync args to make sure they make a little sense */
4575 	if (rootnex_sync_check_parms) {
4576 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4577 		    cache_flags);
4578 		if (e != DDI_SUCCESS) {
4579 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4580 			return (DDI_FAILURE);
4581 		}
4582 	}
4583 
4584 	/*
4585 	 * special case the first page to handle the offset into the page. The
4586 	 * offset to the current page for our buffer is the offset into the
4587 	 * first page of the buffer plus our current offset into the buffer
4588 	 * itself, masked of course.
4589 	 */
4590 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4591 	psize = MIN((MMU_PAGESIZE - poff), size);
4592 
4593 	/* go through all the pages that we want to sync */
4594 	while (size > 0) {
4595 		/*
4596 		 * Calculate the page index relative to the start of the buffer.
4597 		 * The index to the current page for our buffer is the offset
4598 		 * into the first page of the buffer plus our current offset
4599 		 * into the buffer itself, shifted of course...
4600 		 */
4601 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4602 		ASSERT(pidx < sinfo->si_max_pages);
4603 
4604 		/*
4605 		 * if this page uses the copy buffer, we need to sync it,
4606 		 * otherwise, go on to the next page.
4607 		 */
4608 		cbpage = &dma->dp_pgmap[pidx];
4609 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4610 		    (cbpage->pm_uses_copybuf == B_FALSE));
4611 		if (cbpage->pm_uses_copybuf) {
4612 			/* cbaddr and kaddr should be page aligned */
4613 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4614 			    MMU_PAGEOFFSET) == 0);
4615 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4616 			    MMU_PAGEOFFSET) == 0);
4617 
4618 			/*
4619 			 * if we're copying for the device, we are going to
4620 			 * copy from the drivers buffer and to the rootnex
4621 			 * allocated copy buffer.
4622 			 */
4623 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4624 				fromaddr = cbpage->pm_kaddr + poff;
4625 				toaddr = cbpage->pm_cbaddr + poff;
4626 				ROOTNEX_DPROBE2(rootnex__sync__dev,
4627 				    dev_info_t *, dma->dp_dip, size_t, psize);
4628 
4629 			/*
4630 			 * if we're copying for the cpu/kernel, we are going to
4631 			 * copy from the rootnex allocated copy buffer to the
4632 			 * drivers buffer.
4633 			 */
4634 			} else {
4635 				fromaddr = cbpage->pm_cbaddr + poff;
4636 				toaddr = cbpage->pm_kaddr + poff;
4637 				ROOTNEX_DPROBE2(rootnex__sync__cpu,
4638 				    dev_info_t *, dma->dp_dip, size_t, psize);
4639 			}
4640 
4641 			bcopy(fromaddr, toaddr, psize);
4642 		}
4643 
4644 		/*
4645 		 * decrement size until we're done, update our offset into the
4646 		 * buffer, and get the next page size.
4647 		 */
4648 		size -= psize;
4649 		offset += psize;
4650 		psize = MIN(MMU_PAGESIZE, size);
4651 
4652 		/* page offset is zero for the rest of this loop */
4653 		poff = 0;
4654 	}
4655 
4656 	return (DDI_SUCCESS);
4657 }
4658 
4659 /*
4660  * rootnex_dma_sync()
4661  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4662  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4663  *    is set, ddi_dma_sync() returns immediately passing back success.
4664  */
4665 /*ARGSUSED*/
4666 static int
4667 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4668     off_t off, size_t len, uint_t cache_flags)
4669 {
4670 #if defined(__amd64) && !defined(__xpv)
4671 	if (IOMMU_USED(rdip)) {
4672 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4673 		    cache_flags));
4674 	}
4675 #endif
4676 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4677 	    cache_flags));
4678 }
4679 
4680 /*
4681  * rootnex_valid_sync_parms()
4682  *    checks the parameters passed to sync to verify they are correct.
4683  */
4684 static int
4685 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4686     off_t offset, size_t size, uint_t cache_flags)
4687 {
4688 	off_t woffset;
4689 
4690 
4691 	/*
4692 	 * the first part of the test to make sure the offset passed in is
4693 	 * within the window.
4694 	 */
4695 	if (offset < win->wd_offset) {
4696 		return (DDI_FAILURE);
4697 	}
4698 
4699 	/*
4700 	 * second and last part of the test to make sure the offset and length
4701 	 * passed in is within the window.
4702 	 */
4703 	woffset = offset - win->wd_offset;
4704 	if ((woffset + size) > win->wd_size) {
4705 		return (DDI_FAILURE);
4706 	}
4707 
4708 	/*
4709 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4710 	 * be set too.
4711 	 */
4712 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4713 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4714 		return (DDI_SUCCESS);
4715 	}
4716 
4717 	/*
4718 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4719 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4720 	 */
4721 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4722 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4723 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4724 		return (DDI_SUCCESS);
4725 	}
4726 
4727 	return (DDI_FAILURE);
4728 }
4729 
4730 
4731 /*ARGSUSED*/
4732 static int
4733 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4734     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4735     uint_t *ccountp)
4736 {
4737 	rootnex_window_t *window;
4738 	rootnex_trim_t *trim;
4739 	ddi_dma_impl_t *hp;
4740 	rootnex_dma_t *dma;
4741 	ddi_dma_obj_t *dmao;
4742 #if !defined(__amd64)
4743 	rootnex_sglinfo_t *sinfo;
4744 	rootnex_pgmap_t *pmap;
4745 	uint_t pidx;
4746 	uint_t pcnt;
4747 	off_t poff;
4748 	int i;
4749 #endif
4750 
4751 
4752 	hp = (ddi_dma_impl_t *)handle;
4753 	dma = (rootnex_dma_t *)hp->dmai_private;
4754 #if !defined(__amd64)
4755 	sinfo = &dma->dp_sglinfo;
4756 #endif
4757 
4758 	/* If we try and get a window which doesn't exist, return failure */
4759 	if (win >= hp->dmai_nwin) {
4760 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4761 		return (DDI_FAILURE);
4762 	}
4763 
4764 	dmao = dma->dp_dvma_used ? &dma->dp_dvma : &dma->dp_dma;
4765 
4766 	/*
4767 	 * if we don't have any windows, and they're asking for the first
4768 	 * window, setup the cookie pointer to the first cookie in the bind.
4769 	 * setup our return values, then increment the cookie since we return
4770 	 * the first cookie on the stack.
4771 	 */
4772 	if (dma->dp_window == NULL) {
4773 		if (win != 0) {
4774 			ROOTNEX_DPROF_INC(
4775 			    &rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4776 			return (DDI_FAILURE);
4777 		}
4778 		hp->dmai_cookie = dma->dp_cookies;
4779 		*offp = 0;
4780 		*lenp = dmao->dmao_size;
4781 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4782 		*cookiep = hp->dmai_cookie[0];
4783 		hp->dmai_cookie++;
4784 		hp->dmai_ncookies = *ccountp;
4785 		hp->dmai_curcookie = 1;
4786 		return (DDI_SUCCESS);
4787 	}
4788 
4789 	/* sync the old window before moving on to the new one */
4790 	window = &dma->dp_window[dma->dp_current_win];
4791 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4792 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4793 		    DDI_DMA_SYNC_FORCPU);
4794 	}
4795 
4796 #if !defined(__amd64)
4797 	/*
4798 	 * before we move to the next window, if we need to re-map, unmap all
4799 	 * the pages in this window.
4800 	 */
4801 	if (dma->dp_cb_remaping) {
4802 		/*
4803 		 * If we switch to this window again, we'll need to map in
4804 		 * on the fly next time.
4805 		 */
4806 		window->wd_remap_copybuf = B_TRUE;
4807 
4808 		/*
4809 		 * calculate the page index into the buffer where this window
4810 		 * starts, and the number of pages this window takes up.
4811 		 */
4812 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4813 		    MMU_PAGESHIFT;
4814 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4815 		    MMU_PAGEOFFSET;
4816 		pcnt = mmu_btopr(window->wd_size + poff);
4817 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4818 
4819 		/* unmap pages which are currently mapped in this window */
4820 		for (i = 0; i < pcnt; i++) {
4821 			if (dma->dp_pgmap[pidx].pm_mapped) {
4822 				hat_unload(kas.a_hat,
4823 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4824 				    HAT_UNLOAD);
4825 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4826 			}
4827 			pidx++;
4828 		}
4829 	}
4830 #endif
4831 
4832 	/*
4833 	 * Move to the new window.
4834 	 * NOTE: current_win must be set for sync to work right
4835 	 */
4836 	dma->dp_current_win = win;
4837 	window = &dma->dp_window[win];
4838 
4839 	/* if needed, adjust the first and/or last cookies for trim */
4840 	trim = &window->wd_trim;
4841 	if (trim->tr_trim_first) {
4842 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4843 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4844 #if !defined(__amd64)
4845 		window->wd_first_cookie->dmac_type =
4846 		    (window->wd_first_cookie->dmac_type &
4847 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4848 #endif
4849 		if (trim->tr_first_copybuf_win) {
4850 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4851 			    trim->tr_first_cbaddr;
4852 #if !defined(__amd64)
4853 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4854 			    trim->tr_first_kaddr;
4855 #endif
4856 		}
4857 	}
4858 	if (trim->tr_trim_last) {
4859 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4860 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4861 		if (trim->tr_last_copybuf_win) {
4862 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4863 			    trim->tr_last_cbaddr;
4864 #if !defined(__amd64)
4865 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4866 			    trim->tr_last_kaddr;
4867 #endif
4868 		}
4869 	}
4870 
4871 	/*
4872 	 * setup the cookie pointer to the first cookie in the window. setup
4873 	 * our return values, then increment the cookie since we return the
4874 	 * first cookie on the stack.
4875 	 */
4876 	hp->dmai_cookie = window->wd_first_cookie;
4877 	*offp = window->wd_offset;
4878 	*lenp = window->wd_size;
4879 	*ccountp = window->wd_cookie_cnt;
4880 	*cookiep = hp->dmai_cookie[0];
4881 	hp->dmai_ncookies = *ccountp;
4882 	hp->dmai_curcookie = 1;
4883 	hp->dmai_cookie++;
4884 
4885 #if !defined(__amd64)
4886 	/* re-map copybuf if required for this window */
4887 	if (dma->dp_cb_remaping) {
4888 		/*
4889 		 * calculate the page index into the buffer where this
4890 		 * window starts.
4891 		 */
4892 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4893 		    MMU_PAGESHIFT;
4894 		ASSERT(pidx < sinfo->si_max_pages);
4895 
4896 		/*
4897 		 * the first page can get unmapped if it's shared with the
4898 		 * previous window. Even if the rest of this window is already
4899 		 * mapped in, we need to still check this one.
4900 		 */
4901 		pmap = &dma->dp_pgmap[pidx];
4902 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4903 			if (pmap->pm_pp != NULL) {
4904 				pmap->pm_mapped = B_TRUE;
4905 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4906 			} else if (pmap->pm_vaddr != NULL) {
4907 				pmap->pm_mapped = B_TRUE;
4908 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4909 				    pmap->pm_kaddr);
4910 			}
4911 		}
4912 		pidx++;
4913 
4914 		/* map in the rest of the pages if required */
4915 		if (window->wd_remap_copybuf) {
4916 			window->wd_remap_copybuf = B_FALSE;
4917 
4918 			/* figure out many pages this window takes up */
4919 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4920 			    MMU_PAGEOFFSET;
4921 			pcnt = mmu_btopr(window->wd_size + poff);
4922 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4923 
4924 			/* map pages which require it */
4925 			for (i = 1; i < pcnt; i++) {
4926 				pmap = &dma->dp_pgmap[pidx];
4927 				if (pmap->pm_uses_copybuf) {
4928 					ASSERT(pmap->pm_mapped == B_FALSE);
4929 					if (pmap->pm_pp != NULL) {
4930 						pmap->pm_mapped = B_TRUE;
4931 						i86_pp_map(pmap->pm_pp,
4932 						    pmap->pm_kaddr);
4933 					} else if (pmap->pm_vaddr != NULL) {
4934 						pmap->pm_mapped = B_TRUE;
4935 						i86_va_map(pmap->pm_vaddr,
4936 						    sinfo->si_asp,
4937 						    pmap->pm_kaddr);
4938 					}
4939 				}
4940 				pidx++;
4941 			}
4942 		}
4943 	}
4944 #endif
4945 
4946 	/* if the new window uses the copy buffer, sync it for the device */
4947 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4948 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4949 		    DDI_DMA_SYNC_FORDEV);
4950 	}
4951 
4952 	return (DDI_SUCCESS);
4953 }
4954 
4955 /*
4956  * rootnex_dma_win()
4957  *    called from ddi_dma_getwin()
4958  */
4959 /*ARGSUSED*/
4960 static int
4961 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4962     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4963     uint_t *ccountp)
4964 {
4965 #if defined(__amd64) && !defined(__xpv)
4966 	if (IOMMU_USED(rdip)) {
4967 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4968 		    cookiep, ccountp));
4969 	}
4970 #endif
4971 
4972 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4973 	    cookiep, ccountp));
4974 }
4975 
4976 #if defined(__amd64) && !defined(__xpv)
4977 /*ARGSUSED*/
4978 static int
4979 rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
4980     ddi_dma_handle_t handle, void *v)
4981 {
4982 	ddi_dma_impl_t *hp;
4983 	rootnex_dma_t *dma;
4984 
4985 	hp = (ddi_dma_impl_t *)handle;
4986 	dma = (rootnex_dma_t *)hp->dmai_private;
4987 	dma->dp_iommu_private = v;
4988 
4989 	return (DDI_SUCCESS);
4990 }
4991 
4992 /*ARGSUSED*/
4993 static void *
4994 rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
4995     ddi_dma_handle_t handle)
4996 {
4997 	ddi_dma_impl_t *hp;
4998 	rootnex_dma_t *dma;
4999 
5000 	hp = (ddi_dma_impl_t *)handle;
5001 	dma = (rootnex_dma_t *)hp->dmai_private;
5002 
5003 	return (dma->dp_iommu_private);
5004 }
5005 #endif
5006 
5007 /*
5008  * ************************
5009  *  obsoleted dma routines
5010  * ************************
5011  */
5012 
5013 /*
5014  * rootnex_dma_mctl()
5015  *
5016  * We don't support this legacy interface any more on x86.
5017  */
5018 /* ARGSUSED */
5019 static int
5020 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
5021     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
5022     uint_t cache_flags)
5023 {
5024 	/*
5025 	 * The only thing dma_mctl is usef for anymore is legacy SPARC
5026 	 * dvma and sbus-specific routines.
5027 	 */
5028 	return (DDI_FAILURE);
5029 }
5030 
5031 /*
5032  * *********
5033  *  FMA Code
5034  * *********
5035  */
5036 
5037 /*
5038  * rootnex_fm_init()
5039  *    FMA init busop
5040  */
5041 /* ARGSUSED */
5042 static int
5043 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
5044     ddi_iblock_cookie_t *ibc)
5045 {
5046 	*ibc = rootnex_state->r_err_ibc;
5047 
5048 	return (ddi_system_fmcap);
5049 }
5050 
5051 /*
5052  * rootnex_dma_check()
5053  *    Function called after a dma fault occurred to find out whether the
5054  *    fault address is associated with a driver that is able to handle faults
5055  *    and recover from faults.
5056  */
5057 /* ARGSUSED */
5058 static int
5059 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
5060     const void *not_used)
5061 {
5062 	rootnex_window_t *window;
5063 	uint64_t start_addr;
5064 	uint64_t fault_addr;
5065 	ddi_dma_impl_t *hp;
5066 	rootnex_dma_t *dma;
5067 	uint64_t end_addr;
5068 	size_t csize;
5069 	int i;
5070 	int j;
5071 
5072 
5073 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5074 	hp = (ddi_dma_impl_t *)handle;
5075 	ASSERT(hp);
5076 
5077 	dma = (rootnex_dma_t *)hp->dmai_private;
5078 
5079 	/* Get the address that we need to search for */
5080 	fault_addr = *(uint64_t *)addr;
5081 
5082 	/*
5083 	 * if we don't have any windows, we can just walk through all the
5084 	 * cookies.
5085 	 */
5086 	if (dma->dp_window == NULL) {
5087 		/* for each cookie */
5088 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5089 			/*
5090 			 * if the faulted address is within the physical address
5091 			 * range of the cookie, return DDI_FM_NONFATAL.
5092 			 */
5093 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5094 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5095 			    dma->dp_cookies[i].dmac_size))) {
5096 				return (DDI_FM_NONFATAL);
5097 			}
5098 		}
5099 
5100 		/* fault_addr not within this DMA handle */
5101 		return (DDI_FM_UNKNOWN);
5102 	}
5103 
5104 	/* we have mutiple windows, walk through each window */
5105 	for (i = 0; i < hp->dmai_nwin; i++) {
5106 		window = &dma->dp_window[i];
5107 
5108 		/* Go through all the cookies in the window */
5109 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5110 
5111 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5112 			csize = window->wd_first_cookie[j].dmac_size;
5113 
5114 			/*
5115 			 * if we are trimming the first cookie in the window,
5116 			 * and this is the first cookie, adjust the start
5117 			 * address and size of the cookie to account for the
5118 			 * trim.
5119 			 */
5120 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5121 				start_addr = window->wd_trim.tr_first_paddr;
5122 				csize = window->wd_trim.tr_first_size;
5123 			}
5124 
5125 			/*
5126 			 * if we are trimming the last cookie in the window,
5127 			 * and this is the last cookie, adjust the start
5128 			 * address and size of the cookie to account for the
5129 			 * trim.
5130 			 */
5131 			if (window->wd_trim.tr_trim_last &&
5132 			    (j == (window->wd_cookie_cnt - 1))) {
5133 				start_addr = window->wd_trim.tr_last_paddr;
5134 				csize = window->wd_trim.tr_last_size;
5135 			}
5136 
5137 			end_addr = start_addr + csize;
5138 
5139 			/*
5140 			 * if the faulted address is within the physical
5141 			 * address of the cookie, return DDI_FM_NONFATAL.
5142 			 */
5143 			if ((fault_addr >= start_addr) &&
5144 			    (fault_addr <= end_addr)) {
5145 				return (DDI_FM_NONFATAL);
5146 			}
5147 		}
5148 	}
5149 
5150 	/* fault_addr not within this DMA handle */
5151 	return (DDI_FM_UNKNOWN);
5152 }
5153 
5154 /*ARGSUSED*/
5155 static int
5156 rootnex_quiesce(dev_info_t *dip)
5157 {
5158 #if defined(__amd64) && !defined(__xpv)
5159 	return (immu_quiesce());
5160 #else
5161 	return (DDI_SUCCESS);
5162 #endif
5163 }
5164 
5165 #if defined(__xpv)
5166 void
5167 immu_init(void)
5168 {
5169 	;
5170 }
5171 
5172 void
5173 immu_startup(void)
5174 {
5175 	;
5176 }
5177 /*ARGSUSED*/
5178 void
5179 immu_physmem_update(uint64_t addr, uint64_t size)
5180 {
5181 	;
5182 }
5183 #endif
5184