xref: /titanic_52/usr/src/uts/i86pc/io/rootnex.c (revision 8e30ce7b864214c672f5378abebdcce0490c47ec)
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 /*
26  * x86 root nexus driver
27  */
28 
29 #include <sys/sysmacros.h>
30 #include <sys/conf.h>
31 #include <sys/autoconf.h>
32 #include <sys/sysmacros.h>
33 #include <sys/debug.h>
34 #include <sys/psw.h>
35 #include <sys/ddidmareq.h>
36 #include <sys/promif.h>
37 #include <sys/devops.h>
38 #include <sys/kmem.h>
39 #include <sys/cmn_err.h>
40 #include <vm/seg.h>
41 #include <vm/seg_kmem.h>
42 #include <vm/seg_dev.h>
43 #include <sys/vmem.h>
44 #include <sys/mman.h>
45 #include <vm/hat.h>
46 #include <vm/as.h>
47 #include <vm/page.h>
48 #include <sys/avintr.h>
49 #include <sys/errno.h>
50 #include <sys/modctl.h>
51 #include <sys/ddi_impldefs.h>
52 #include <sys/sunddi.h>
53 #include <sys/sunndi.h>
54 #include <sys/mach_intr.h>
55 #include <sys/psm.h>
56 #include <sys/ontrap.h>
57 #include <sys/atomic.h>
58 #include <sys/sdt.h>
59 #include <sys/rootnex.h>
60 #include <vm/hat_i86.h>
61 #include <sys/ddifm.h>
62 #include <sys/ddi_isa.h>
63 #include <sys/apic.h>
64 
65 #ifdef __xpv
66 #include <sys/bootinfo.h>
67 #include <sys/hypervisor.h>
68 #include <sys/bootconf.h>
69 #include <vm/kboot_mmu.h>
70 #endif
71 
72 #if defined(__amd64) && !defined(__xpv)
73 #include <sys/immu.h>
74 #endif
75 
76 
77 /*
78  * enable/disable extra checking of function parameters. Useful for debugging
79  * drivers.
80  */
81 #ifdef	DEBUG
82 int rootnex_alloc_check_parms = 1;
83 int rootnex_bind_check_parms = 1;
84 int rootnex_bind_check_inuse = 1;
85 int rootnex_unbind_verify_buffer = 0;
86 int rootnex_sync_check_parms = 1;
87 #else
88 int rootnex_alloc_check_parms = 0;
89 int rootnex_bind_check_parms = 0;
90 int rootnex_bind_check_inuse = 0;
91 int rootnex_unbind_verify_buffer = 0;
92 int rootnex_sync_check_parms = 0;
93 #endif
94 
95 boolean_t rootnex_dmar_not_setup;
96 
97 /* Master Abort and Target Abort panic flag */
98 int rootnex_fm_ma_ta_panic_flag = 0;
99 
100 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
101 int rootnex_bind_fail = 1;
102 int rootnex_bind_warn = 1;
103 uint8_t *rootnex_warn_list;
104 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
105 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
106 
107 /*
108  * revert back to old broken behavior of always sync'ing entire copy buffer.
109  * This is useful if be have a buggy driver which doesn't correctly pass in
110  * the offset and size into ddi_dma_sync().
111  */
112 int rootnex_sync_ignore_params = 0;
113 
114 /*
115  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
116  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
117  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
118  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
119  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
120  * (< 8K). We will still need to allocate the copy buffer during bind though
121  * (if we need one). These can only be modified in /etc/system before rootnex
122  * attach.
123  */
124 #if defined(__amd64)
125 int rootnex_prealloc_cookies = 65;
126 int rootnex_prealloc_windows = 4;
127 int rootnex_prealloc_copybuf = 2;
128 #else
129 int rootnex_prealloc_cookies = 33;
130 int rootnex_prealloc_windows = 4;
131 int rootnex_prealloc_copybuf = 2;
132 #endif
133 
134 /* driver global state */
135 static rootnex_state_t *rootnex_state;
136 
137 /* shortcut to rootnex counters */
138 static uint64_t *rootnex_cnt;
139 
140 /*
141  * XXX - does x86 even need these or are they left over from the SPARC days?
142  */
143 /* statically defined integer/boolean properties for the root node */
144 static rootnex_intprop_t rootnex_intprp[] = {
145 	{ "PAGESIZE",			PAGESIZE },
146 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
147 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
148 	{ DDI_RELATIVE_ADDRESSING,	1 },
149 };
150 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
151 
152 #ifdef __xpv
153 typedef maddr_t rootnex_addr_t;
154 #define	ROOTNEX_PADDR_TO_RBASE(xinfo, pa)	\
155 	(DOMAIN_IS_INITDOMAIN(xinfo) ? pa_to_ma(pa) : (pa))
156 #else
157 typedef paddr_t rootnex_addr_t;
158 #endif
159 
160 #if !defined(__xpv)
161 char _depends_on[] = "misc/iommulib misc/acpica";
162 #endif
163 
164 static struct cb_ops rootnex_cb_ops = {
165 	nodev,		/* open */
166 	nodev,		/* close */
167 	nodev,		/* strategy */
168 	nodev,		/* print */
169 	nodev,		/* dump */
170 	nodev,		/* read */
171 	nodev,		/* write */
172 	nodev,		/* ioctl */
173 	nodev,		/* devmap */
174 	nodev,		/* mmap */
175 	nodev,		/* segmap */
176 	nochpoll,	/* chpoll */
177 	ddi_prop_op,	/* cb_prop_op */
178 	NULL,		/* struct streamtab */
179 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
180 	CB_REV,		/* Rev */
181 	nodev,		/* cb_aread */
182 	nodev		/* cb_awrite */
183 };
184 
185 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
186     off_t offset, off_t len, caddr_t *vaddrp);
187 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
188     struct hat *hat, struct seg *seg, caddr_t addr,
189     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
190 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
191     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
192 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
193     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
194     ddi_dma_handle_t *handlep);
195 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
196     ddi_dma_handle_t handle);
197 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
198     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
199     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
200 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
201     ddi_dma_handle_t handle);
202 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
203     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
204 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
205     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
206     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
207 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
208     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
209     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
210 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
211     ddi_ctl_enum_t ctlop, void *arg, void *result);
212 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
213     ddi_iblock_cookie_t *ibc);
214 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
215     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
216 static int rootnex_alloc_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *,
217     void *);
218 static int rootnex_free_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *);
219 
220 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
221     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
222     ddi_dma_handle_t *handlep);
223 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
224     ddi_dma_handle_t handle);
225 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
226     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
227     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
228 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
229     ddi_dma_handle_t handle);
230 #if defined(__amd64) && !defined(__xpv)
231 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
232     ddi_dma_handle_t handle);
233 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
234     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
235 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
236     ddi_dma_cookie_t *cookiep, uint_t ccount);
237 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
238     ddi_dma_handle_t handle);
239 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
240 #endif
241 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
242     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
243 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
244     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
245     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
246 
247 static struct bus_ops rootnex_bus_ops = {
248 	BUSO_REV,
249 	rootnex_map,
250 	NULL,
251 	NULL,
252 	NULL,
253 	rootnex_map_fault,
254 	rootnex_dma_map,
255 	rootnex_dma_allochdl,
256 	rootnex_dma_freehdl,
257 	rootnex_dma_bindhdl,
258 	rootnex_dma_unbindhdl,
259 	rootnex_dma_sync,
260 	rootnex_dma_win,
261 	rootnex_dma_mctl,
262 	rootnex_ctlops,
263 	ddi_bus_prop_op,
264 	i_ddi_rootnex_get_eventcookie,
265 	i_ddi_rootnex_add_eventcall,
266 	i_ddi_rootnex_remove_eventcall,
267 	i_ddi_rootnex_post_event,
268 	0,			/* bus_intr_ctl */
269 	0,			/* bus_config */
270 	0,			/* bus_unconfig */
271 	rootnex_fm_init,	/* bus_fm_init */
272 	NULL,			/* bus_fm_fini */
273 	NULL,			/* bus_fm_access_enter */
274 	NULL,			/* bus_fm_access_exit */
275 	NULL,			/* bus_powr */
276 	rootnex_intr_ops	/* bus_intr_op */
277 };
278 
279 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
280 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
281 static int rootnex_quiesce(dev_info_t *dip);
282 
283 static struct dev_ops rootnex_ops = {
284 	DEVO_REV,
285 	0,
286 	ddi_no_info,
287 	nulldev,
288 	nulldev,
289 	rootnex_attach,
290 	rootnex_detach,
291 	nulldev,
292 	&rootnex_cb_ops,
293 	&rootnex_bus_ops,
294 	NULL,
295 	rootnex_quiesce,		/* quiesce */
296 };
297 
298 static struct modldrv rootnex_modldrv = {
299 	&mod_driverops,
300 	"i86pc root nexus",
301 	&rootnex_ops
302 };
303 
304 static struct modlinkage rootnex_modlinkage = {
305 	MODREV_1,
306 	(void *)&rootnex_modldrv,
307 	NULL
308 };
309 
310 #if defined(__amd64) && !defined(__xpv)
311 static iommulib_nexops_t iommulib_nexops = {
312 	IOMMU_NEXOPS_VERSION,
313 	"Rootnex IOMMU ops Vers 1.1",
314 	NULL,
315 	rootnex_coredma_allochdl,
316 	rootnex_coredma_freehdl,
317 	rootnex_coredma_bindhdl,
318 	rootnex_coredma_unbindhdl,
319 	rootnex_coredma_reset_cookies,
320 	rootnex_coredma_get_cookies,
321 	rootnex_coredma_set_cookies,
322 	rootnex_coredma_clear_cookies,
323 	rootnex_coredma_get_sleep_flags,
324 	rootnex_coredma_sync,
325 	rootnex_coredma_win,
326 	rootnex_dma_map,
327 	rootnex_dma_mctl
328 };
329 #endif
330 
331 /*
332  *  extern hacks
333  */
334 extern struct seg_ops segdev_ops;
335 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
336 #ifdef	DDI_MAP_DEBUG
337 extern int ddi_map_debug_flag;
338 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
339 #endif
340 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
341 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
342 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
343     psm_intr_op_t, int *);
344 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
345 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
346 
347 /*
348  * Use device arena to use for device control register mappings.
349  * Various kernel memory walkers (debugger, dtrace) need to know
350  * to avoid this address range to prevent undesired device activity.
351  */
352 extern void *device_arena_alloc(size_t size, int vm_flag);
353 extern void device_arena_free(void * vaddr, size_t size);
354 
355 
356 /*
357  *  Internal functions
358  */
359 static int rootnex_dma_init();
360 static void rootnex_add_props(dev_info_t *);
361 static int rootnex_ctl_reportdev(dev_info_t *dip);
362 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
363 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
364 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
365 static int rootnex_map_handle(ddi_map_req_t *mp);
366 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
367 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
368 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
369     ddi_dma_attr_t *attr);
370 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
371     rootnex_sglinfo_t *sglinfo);
372 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
373     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag);
374 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
375     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
376 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
377 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
378     ddi_dma_attr_t *attr, int kmflag);
379 static void rootnex_teardown_windows(rootnex_dma_t *dma);
380 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
381     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
382 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
383     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
384     size_t *copybuf_used, page_t **cur_pp);
385 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
386     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
387     ddi_dma_attr_t *attr, off_t cur_offset);
388 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
389     rootnex_dma_t *dma, rootnex_window_t **windowp,
390     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
391 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
392     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
393 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
394     off_t offset, size_t size, uint_t cache_flags);
395 static int rootnex_verify_buffer(rootnex_dma_t *dma);
396 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
397     const void *comp_addr, const void *not_used);
398 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
399     rootnex_sglinfo_t *sglinfo);
400 
401 /*
402  * _init()
403  *
404  */
405 int
406 _init(void)
407 {
408 
409 	rootnex_state = NULL;
410 	return (mod_install(&rootnex_modlinkage));
411 }
412 
413 
414 /*
415  * _info()
416  *
417  */
418 int
419 _info(struct modinfo *modinfop)
420 {
421 	return (mod_info(&rootnex_modlinkage, modinfop));
422 }
423 
424 
425 /*
426  * _fini()
427  *
428  */
429 int
430 _fini(void)
431 {
432 	return (EBUSY);
433 }
434 
435 
436 /*
437  * rootnex_attach()
438  *
439  */
440 static int
441 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
442 {
443 	int fmcap;
444 	int e;
445 
446 	switch (cmd) {
447 	case DDI_ATTACH:
448 		break;
449 	case DDI_RESUME:
450 #if defined(__amd64) && !defined(__xpv)
451 		return (immu_unquiesce());
452 #else
453 		return (DDI_SUCCESS);
454 #endif
455 	default:
456 		return (DDI_FAILURE);
457 	}
458 
459 	/*
460 	 * We should only have one instance of rootnex. Save it away since we
461 	 * don't have an easy way to get it back later.
462 	 */
463 	ASSERT(rootnex_state == NULL);
464 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
465 
466 	rootnex_state->r_dip = dip;
467 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
468 	rootnex_state->r_reserved_msg_printed = B_FALSE;
469 	rootnex_cnt = &rootnex_state->r_counters[0];
470 
471 	/*
472 	 * Set minimum fm capability level for i86pc platforms and then
473 	 * initialize error handling. Since we're the rootnex, we don't
474 	 * care what's returned in the fmcap field.
475 	 */
476 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
477 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
478 	fmcap = ddi_system_fmcap;
479 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
480 
481 	/* initialize DMA related state */
482 	e = rootnex_dma_init();
483 	if (e != DDI_SUCCESS) {
484 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
485 		return (DDI_FAILURE);
486 	}
487 
488 	/* Add static root node properties */
489 	rootnex_add_props(dip);
490 
491 	/* since we can't call ddi_report_dev() */
492 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
493 
494 	/* Initialize rootnex event handle */
495 	i_ddi_rootnex_init_events(dip);
496 
497 #if defined(__amd64) && !defined(__xpv)
498 	e = iommulib_nexus_register(dip, &iommulib_nexops,
499 	    &rootnex_state->r_iommulib_handle);
500 
501 	ASSERT(e == DDI_SUCCESS);
502 #endif
503 
504 	return (DDI_SUCCESS);
505 }
506 
507 
508 /*
509  * rootnex_detach()
510  *
511  */
512 /*ARGSUSED*/
513 static int
514 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
515 {
516 	switch (cmd) {
517 	case DDI_SUSPEND:
518 #if defined(__amd64) && !defined(__xpv)
519 		return (immu_quiesce());
520 #else
521 		return (DDI_SUCCESS);
522 #endif
523 	default:
524 		return (DDI_FAILURE);
525 	}
526 	/*NOTREACHED*/
527 
528 }
529 
530 
531 /*
532  * rootnex_dma_init()
533  *
534  */
535 /*ARGSUSED*/
536 static int
537 rootnex_dma_init()
538 {
539 	size_t bufsize;
540 
541 
542 	/*
543 	 * size of our cookie/window/copybuf state needed in dma bind that we
544 	 * pre-alloc in dma_alloc_handle
545 	 */
546 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
547 	rootnex_state->r_prealloc_size =
548 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
549 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
550 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
551 
552 	/*
553 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
554 	 * allocate 16 extra bytes for struct pointer alignment
555 	 * (p->dmai_private & dma->dp_prealloc_buffer)
556 	 */
557 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
558 	    rootnex_state->r_prealloc_size + 0x10;
559 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
560 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
561 	if (rootnex_state->r_dmahdl_cache == NULL) {
562 		return (DDI_FAILURE);
563 	}
564 
565 	/*
566 	 * allocate array to track which major numbers we have printed warnings
567 	 * for.
568 	 */
569 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
570 	    KM_SLEEP);
571 
572 	return (DDI_SUCCESS);
573 }
574 
575 
576 /*
577  * rootnex_add_props()
578  *
579  */
580 static void
581 rootnex_add_props(dev_info_t *dip)
582 {
583 	rootnex_intprop_t *rpp;
584 	int i;
585 
586 	/* Add static integer/boolean properties to the root node */
587 	rpp = rootnex_intprp;
588 	for (i = 0; i < NROOT_INTPROPS; i++) {
589 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
590 		    rpp[i].prop_name, rpp[i].prop_value);
591 	}
592 }
593 
594 
595 
596 /*
597  * *************************
598  *  ctlops related routines
599  * *************************
600  */
601 
602 /*
603  * rootnex_ctlops()
604  *
605  */
606 /*ARGSUSED*/
607 static int
608 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
609     void *arg, void *result)
610 {
611 	int n, *ptr;
612 	struct ddi_parent_private_data *pdp;
613 
614 	switch (ctlop) {
615 	case DDI_CTLOPS_DMAPMAPC:
616 		/*
617 		 * Return 'partial' to indicate that dma mapping
618 		 * has to be done in the main MMU.
619 		 */
620 		return (DDI_DMA_PARTIAL);
621 
622 	case DDI_CTLOPS_BTOP:
623 		/*
624 		 * Convert byte count input to physical page units.
625 		 * (byte counts that are not a page-size multiple
626 		 * are rounded down)
627 		 */
628 		*(ulong_t *)result = btop(*(ulong_t *)arg);
629 		return (DDI_SUCCESS);
630 
631 	case DDI_CTLOPS_PTOB:
632 		/*
633 		 * Convert size in physical pages to bytes
634 		 */
635 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
636 		return (DDI_SUCCESS);
637 
638 	case DDI_CTLOPS_BTOPR:
639 		/*
640 		 * Convert byte count input to physical page units
641 		 * (byte counts that are not a page-size multiple
642 		 * are rounded up)
643 		 */
644 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
645 		return (DDI_SUCCESS);
646 
647 	case DDI_CTLOPS_INITCHILD:
648 		return (impl_ddi_sunbus_initchild(arg));
649 
650 	case DDI_CTLOPS_UNINITCHILD:
651 		impl_ddi_sunbus_removechild(arg);
652 		return (DDI_SUCCESS);
653 
654 	case DDI_CTLOPS_REPORTDEV:
655 		return (rootnex_ctl_reportdev(rdip));
656 
657 	case DDI_CTLOPS_IOMIN:
658 		/*
659 		 * Nothing to do here but reflect back..
660 		 */
661 		return (DDI_SUCCESS);
662 
663 	case DDI_CTLOPS_REGSIZE:
664 	case DDI_CTLOPS_NREGS:
665 		break;
666 
667 	case DDI_CTLOPS_SIDDEV:
668 		if (ndi_dev_is_prom_node(rdip))
669 			return (DDI_SUCCESS);
670 		if (ndi_dev_is_persistent_node(rdip))
671 			return (DDI_SUCCESS);
672 		return (DDI_FAILURE);
673 
674 	case DDI_CTLOPS_POWER:
675 		return ((*pm_platform_power)((power_req_t *)arg));
676 
677 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
678 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
679 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
680 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
681 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
682 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
683 		if (!rootnex_state->r_reserved_msg_printed) {
684 			rootnex_state->r_reserved_msg_printed = B_TRUE;
685 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
686 			    "1 or more reserved/obsolete operations.");
687 		}
688 		return (DDI_FAILURE);
689 
690 	default:
691 		return (DDI_FAILURE);
692 	}
693 	/*
694 	 * The rest are for "hardware" properties
695 	 */
696 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
697 		return (DDI_FAILURE);
698 
699 	if (ctlop == DDI_CTLOPS_NREGS) {
700 		ptr = (int *)result;
701 		*ptr = pdp->par_nreg;
702 	} else {
703 		off_t *size = (off_t *)result;
704 
705 		ptr = (int *)arg;
706 		n = *ptr;
707 		if (n >= pdp->par_nreg) {
708 			return (DDI_FAILURE);
709 		}
710 		*size = (off_t)pdp->par_reg[n].regspec_size;
711 	}
712 	return (DDI_SUCCESS);
713 }
714 
715 
716 /*
717  * rootnex_ctl_reportdev()
718  *
719  */
720 static int
721 rootnex_ctl_reportdev(dev_info_t *dev)
722 {
723 	int i, n, len, f_len = 0;
724 	char *buf;
725 
726 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
727 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
728 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
729 	len = strlen(buf);
730 
731 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
732 
733 		struct regspec *rp = sparc_pd_getreg(dev, i);
734 
735 		if (i == 0)
736 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
737 			    ": ");
738 		else
739 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
740 			    " and ");
741 		len = strlen(buf);
742 
743 		switch (rp->regspec_bustype) {
744 
745 		case BTEISA:
746 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
747 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
748 			break;
749 
750 		case BTISA:
751 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
752 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
753 			break;
754 
755 		default:
756 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
757 			    "space %x offset %x",
758 			    rp->regspec_bustype, rp->regspec_addr);
759 			break;
760 		}
761 		len = strlen(buf);
762 	}
763 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
764 		int pri;
765 
766 		if (i != 0) {
767 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
768 			    ",");
769 			len = strlen(buf);
770 		}
771 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
772 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
773 		    " sparc ipl %d", pri);
774 		len = strlen(buf);
775 	}
776 #ifdef DEBUG
777 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
778 		cmn_err(CE_NOTE, "next message is truncated: "
779 		    "printed length 1024, real length %d", f_len);
780 	}
781 #endif /* DEBUG */
782 	cmn_err(CE_CONT, "?%s\n", buf);
783 	kmem_free(buf, REPORTDEV_BUFSIZE);
784 	return (DDI_SUCCESS);
785 }
786 
787 
788 /*
789  * ******************
790  *  map related code
791  * ******************
792  */
793 
794 /*
795  * rootnex_map()
796  *
797  */
798 static int
799 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
800     off_t len, caddr_t *vaddrp)
801 {
802 	struct regspec *rp, tmp_reg;
803 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
804 	int error;
805 
806 	mp = &mr;
807 
808 	switch (mp->map_op)  {
809 	case DDI_MO_MAP_LOCKED:
810 	case DDI_MO_UNMAP:
811 	case DDI_MO_MAP_HANDLE:
812 		break;
813 	default:
814 #ifdef	DDI_MAP_DEBUG
815 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
816 		    mp->map_op);
817 #endif	/* DDI_MAP_DEBUG */
818 		return (DDI_ME_UNIMPLEMENTED);
819 	}
820 
821 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
822 #ifdef	DDI_MAP_DEBUG
823 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
824 #endif	/* DDI_MAP_DEBUG */
825 		return (DDI_ME_UNIMPLEMENTED);
826 	}
827 
828 	/*
829 	 * First, if given an rnumber, convert it to a regspec...
830 	 * (Presumably, this is on behalf of a child of the root node?)
831 	 */
832 
833 	if (mp->map_type == DDI_MT_RNUMBER)  {
834 
835 		int rnumber = mp->map_obj.rnumber;
836 #ifdef	DDI_MAP_DEBUG
837 		static char *out_of_range =
838 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
839 #endif	/* DDI_MAP_DEBUG */
840 
841 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
842 		if (rp == NULL)  {
843 #ifdef	DDI_MAP_DEBUG
844 			cmn_err(CE_WARN, out_of_range, rnumber,
845 			    ddi_get_name(rdip));
846 #endif	/* DDI_MAP_DEBUG */
847 			return (DDI_ME_RNUMBER_RANGE);
848 		}
849 
850 		/*
851 		 * Convert the given ddi_map_req_t from rnumber to regspec...
852 		 */
853 
854 		mp->map_type = DDI_MT_REGSPEC;
855 		mp->map_obj.rp = rp;
856 	}
857 
858 	/*
859 	 * Adjust offset and length correspnding to called values...
860 	 * XXX: A non-zero length means override the one in the regspec
861 	 * XXX: (regardless of what's in the parent's range?)
862 	 */
863 
864 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
865 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
866 
867 #ifdef	DDI_MAP_DEBUG
868 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
869 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
870 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
871 	    len, mp->map_handlep);
872 #endif	/* DDI_MAP_DEBUG */
873 
874 	/*
875 	 * I/O or memory mapping:
876 	 *
877 	 *	<bustype=0, addr=x, len=x>: memory
878 	 *	<bustype=1, addr=x, len=x>: i/o
879 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
880 	 */
881 
882 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
883 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
884 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
885 		    ddi_get_name(rdip), rp->regspec_bustype,
886 		    rp->regspec_addr, rp->regspec_size);
887 		return (DDI_ME_INVAL);
888 	}
889 
890 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
891 		/*
892 		 * compatibility i/o mapping
893 		 */
894 		rp->regspec_bustype += (uint_t)offset;
895 	} else {
896 		/*
897 		 * Normal memory or i/o mapping
898 		 */
899 		rp->regspec_addr += (uint_t)offset;
900 	}
901 
902 	if (len != 0)
903 		rp->regspec_size = (uint_t)len;
904 
905 #ifdef	DDI_MAP_DEBUG
906 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
907 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
908 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
909 	    offset, len, mp->map_handlep);
910 #endif	/* DDI_MAP_DEBUG */
911 
912 	/*
913 	 * Apply any parent ranges at this level, if applicable.
914 	 * (This is where nexus specific regspec translation takes place.
915 	 * Use of this function is implicit agreement that translation is
916 	 * provided via ddi_apply_range.)
917 	 */
918 
919 #ifdef	DDI_MAP_DEBUG
920 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
921 	    ddi_get_name(dip), ddi_get_name(rdip));
922 #endif	/* DDI_MAP_DEBUG */
923 
924 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
925 		return (error);
926 
927 	switch (mp->map_op)  {
928 	case DDI_MO_MAP_LOCKED:
929 
930 		/*
931 		 * Set up the locked down kernel mapping to the regspec...
932 		 */
933 
934 		return (rootnex_map_regspec(mp, vaddrp));
935 
936 	case DDI_MO_UNMAP:
937 
938 		/*
939 		 * Release mapping...
940 		 */
941 
942 		return (rootnex_unmap_regspec(mp, vaddrp));
943 
944 	case DDI_MO_MAP_HANDLE:
945 
946 		return (rootnex_map_handle(mp));
947 
948 	default:
949 		return (DDI_ME_UNIMPLEMENTED);
950 	}
951 }
952 
953 
954 /*
955  * rootnex_map_fault()
956  *
957  *	fault in mappings for requestors
958  */
959 /*ARGSUSED*/
960 static int
961 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
962     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
963     uint_t lock)
964 {
965 
966 #ifdef	DDI_MAP_DEBUG
967 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
968 	ddi_map_debug(" Seg <%s>\n",
969 	    seg->s_ops == &segdev_ops ? "segdev" :
970 	    seg == &kvseg ? "segkmem" : "NONE!");
971 #endif	/* DDI_MAP_DEBUG */
972 
973 	/*
974 	 * This is all terribly broken, but it is a start
975 	 *
976 	 * XXX	Note that this test means that segdev_ops
977 	 *	must be exported from seg_dev.c.
978 	 * XXX	What about devices with their own segment drivers?
979 	 */
980 	if (seg->s_ops == &segdev_ops) {
981 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
982 
983 		if (hat == NULL) {
984 			/*
985 			 * This is one plausible interpretation of
986 			 * a null hat i.e. use the first hat on the
987 			 * address space hat list which by convention is
988 			 * the hat of the system MMU.  At alternative
989 			 * would be to panic .. this might well be better ..
990 			 */
991 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
992 			hat = seg->s_as->a_hat;
993 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
994 		}
995 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
996 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
997 	} else if (seg == &kvseg && dp == NULL) {
998 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
999 		    HAT_LOAD_LOCK);
1000 	} else
1001 		return (DDI_FAILURE);
1002 	return (DDI_SUCCESS);
1003 }
1004 
1005 
1006 /*
1007  * rootnex_map_regspec()
1008  *     we don't support mapping of I/O cards above 4Gb
1009  */
1010 static int
1011 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1012 {
1013 	rootnex_addr_t rbase;
1014 	void *cvaddr;
1015 	uint_t npages, pgoffset;
1016 	struct regspec *rp;
1017 	ddi_acc_hdl_t *hp;
1018 	ddi_acc_impl_t *ap;
1019 	uint_t	hat_acc_flags;
1020 	paddr_t pbase;
1021 
1022 	rp = mp->map_obj.rp;
1023 	hp = mp->map_handlep;
1024 
1025 #ifdef	DDI_MAP_DEBUG
1026 	ddi_map_debug(
1027 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1028 	    rp->regspec_bustype, rp->regspec_addr,
1029 	    rp->regspec_size, mp->map_handlep);
1030 #endif	/* DDI_MAP_DEBUG */
1031 
1032 	/*
1033 	 * I/O or memory mapping
1034 	 *
1035 	 *	<bustype=0, addr=x, len=x>: memory
1036 	 *	<bustype=1, addr=x, len=x>: i/o
1037 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1038 	 */
1039 
1040 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1041 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1042 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1043 		    rp->regspec_addr, rp->regspec_size);
1044 		return (DDI_FAILURE);
1045 	}
1046 
1047 	if (rp->regspec_bustype != 0) {
1048 		/*
1049 		 * I/O space - needs a handle.
1050 		 */
1051 		if (hp == NULL) {
1052 			return (DDI_FAILURE);
1053 		}
1054 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1055 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1056 		impl_acc_hdl_init(hp);
1057 
1058 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1059 #ifdef  DDI_MAP_DEBUG
1060 			ddi_map_debug("rootnex_map_regspec: mmap() "
1061 			    "to I/O space is not supported.\n");
1062 #endif  /* DDI_MAP_DEBUG */
1063 			return (DDI_ME_INVAL);
1064 		} else {
1065 			/*
1066 			 * 1275-compliant vs. compatibility i/o mapping
1067 			 */
1068 			*vaddrp =
1069 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1070 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1071 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1072 #ifdef __xpv
1073 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1074 				hp->ah_pfn = xen_assign_pfn(
1075 				    mmu_btop((ulong_t)rp->regspec_addr &
1076 				    MMU_PAGEMASK));
1077 			} else {
1078 				hp->ah_pfn = mmu_btop(
1079 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1080 			}
1081 #else
1082 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1083 			    MMU_PAGEMASK);
1084 #endif
1085 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1086 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1087 		}
1088 
1089 #ifdef	DDI_MAP_DEBUG
1090 		ddi_map_debug(
1091 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1092 		    rp->regspec_size, *vaddrp);
1093 #endif	/* DDI_MAP_DEBUG */
1094 		return (DDI_SUCCESS);
1095 	}
1096 
1097 	/*
1098 	 * Memory space
1099 	 */
1100 
1101 	if (hp != NULL) {
1102 		/*
1103 		 * hat layer ignores
1104 		 * hp->ah_acc.devacc_attr_endian_flags.
1105 		 */
1106 		switch (hp->ah_acc.devacc_attr_dataorder) {
1107 		case DDI_STRICTORDER_ACC:
1108 			hat_acc_flags = HAT_STRICTORDER;
1109 			break;
1110 		case DDI_UNORDERED_OK_ACC:
1111 			hat_acc_flags = HAT_UNORDERED_OK;
1112 			break;
1113 		case DDI_MERGING_OK_ACC:
1114 			hat_acc_flags = HAT_MERGING_OK;
1115 			break;
1116 		case DDI_LOADCACHING_OK_ACC:
1117 			hat_acc_flags = HAT_LOADCACHING_OK;
1118 			break;
1119 		case DDI_STORECACHING_OK_ACC:
1120 			hat_acc_flags = HAT_STORECACHING_OK;
1121 			break;
1122 		}
1123 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1124 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1125 		impl_acc_hdl_init(hp);
1126 		hp->ah_hat_flags = hat_acc_flags;
1127 	} else {
1128 		hat_acc_flags = HAT_STRICTORDER;
1129 	}
1130 
1131 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1132 #ifdef __xpv
1133 	/*
1134 	 * If we're dom0, we're using a real device so we need to translate
1135 	 * the MA to a PA.
1136 	 */
1137 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1138 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1139 	} else {
1140 		pbase = rbase;
1141 	}
1142 #else
1143 	pbase = rbase;
1144 #endif
1145 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1146 
1147 	if (rp->regspec_size == 0) {
1148 #ifdef  DDI_MAP_DEBUG
1149 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1150 #endif  /* DDI_MAP_DEBUG */
1151 		return (DDI_ME_INVAL);
1152 	}
1153 
1154 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1155 		/* extra cast to make gcc happy */
1156 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1157 	} else {
1158 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1159 
1160 #ifdef	DDI_MAP_DEBUG
1161 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1162 		    "physical %llx", npages, pbase);
1163 #endif	/* DDI_MAP_DEBUG */
1164 
1165 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1166 		if (cvaddr == NULL)
1167 			return (DDI_ME_NORESOURCES);
1168 
1169 		/*
1170 		 * Now map in the pages we've allocated...
1171 		 */
1172 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1173 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1174 		    HAT_LOAD_LOCK);
1175 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1176 
1177 		/* save away pfn and npages for FMA */
1178 		hp = mp->map_handlep;
1179 		if (hp) {
1180 			hp->ah_pfn = mmu_btop(pbase);
1181 			hp->ah_pnum = npages;
1182 		}
1183 	}
1184 
1185 #ifdef	DDI_MAP_DEBUG
1186 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1187 #endif	/* DDI_MAP_DEBUG */
1188 	return (DDI_SUCCESS);
1189 }
1190 
1191 
1192 /*
1193  * rootnex_unmap_regspec()
1194  *
1195  */
1196 static int
1197 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1198 {
1199 	caddr_t addr = (caddr_t)*vaddrp;
1200 	uint_t npages, pgoffset;
1201 	struct regspec *rp;
1202 
1203 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1204 		return (0);
1205 
1206 	rp = mp->map_obj.rp;
1207 
1208 	if (rp->regspec_size == 0) {
1209 #ifdef  DDI_MAP_DEBUG
1210 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1211 #endif  /* DDI_MAP_DEBUG */
1212 		return (DDI_ME_INVAL);
1213 	}
1214 
1215 	/*
1216 	 * I/O or memory mapping:
1217 	 *
1218 	 *	<bustype=0, addr=x, len=x>: memory
1219 	 *	<bustype=1, addr=x, len=x>: i/o
1220 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1221 	 */
1222 	if (rp->regspec_bustype != 0) {
1223 		/*
1224 		 * This is I/O space, which requires no particular
1225 		 * processing on unmap since it isn't mapped in the
1226 		 * first place.
1227 		 */
1228 		return (DDI_SUCCESS);
1229 	}
1230 
1231 	/*
1232 	 * Memory space
1233 	 */
1234 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1235 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1236 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1237 	device_arena_free(addr - pgoffset, ptob(npages));
1238 
1239 	/*
1240 	 * Destroy the pointer - the mapping has logically gone
1241 	 */
1242 	*vaddrp = NULL;
1243 
1244 	return (DDI_SUCCESS);
1245 }
1246 
1247 
1248 /*
1249  * rootnex_map_handle()
1250  *
1251  */
1252 static int
1253 rootnex_map_handle(ddi_map_req_t *mp)
1254 {
1255 	rootnex_addr_t rbase;
1256 	ddi_acc_hdl_t *hp;
1257 	uint_t pgoffset;
1258 	struct regspec *rp;
1259 	paddr_t pbase;
1260 
1261 	rp = mp->map_obj.rp;
1262 
1263 #ifdef	DDI_MAP_DEBUG
1264 	ddi_map_debug(
1265 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1266 	    rp->regspec_bustype, rp->regspec_addr,
1267 	    rp->regspec_size, mp->map_handlep);
1268 #endif	/* DDI_MAP_DEBUG */
1269 
1270 	/*
1271 	 * I/O or memory mapping:
1272 	 *
1273 	 *	<bustype=0, addr=x, len=x>: memory
1274 	 *	<bustype=1, addr=x, len=x>: i/o
1275 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1276 	 */
1277 	if (rp->regspec_bustype != 0) {
1278 		/*
1279 		 * This refers to I/O space, and we don't support "mapping"
1280 		 * I/O space to a user.
1281 		 */
1282 		return (DDI_FAILURE);
1283 	}
1284 
1285 	/*
1286 	 * Set up the hat_flags for the mapping.
1287 	 */
1288 	hp = mp->map_handlep;
1289 
1290 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1291 	case DDI_NEVERSWAP_ACC:
1292 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1293 		break;
1294 	case DDI_STRUCTURE_LE_ACC:
1295 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1296 		break;
1297 	case DDI_STRUCTURE_BE_ACC:
1298 		return (DDI_FAILURE);
1299 	default:
1300 		return (DDI_REGS_ACC_CONFLICT);
1301 	}
1302 
1303 	switch (hp->ah_acc.devacc_attr_dataorder) {
1304 	case DDI_STRICTORDER_ACC:
1305 		break;
1306 	case DDI_UNORDERED_OK_ACC:
1307 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1308 		break;
1309 	case DDI_MERGING_OK_ACC:
1310 		hp->ah_hat_flags |= HAT_MERGING_OK;
1311 		break;
1312 	case DDI_LOADCACHING_OK_ACC:
1313 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1314 		break;
1315 	case DDI_STORECACHING_OK_ACC:
1316 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1317 		break;
1318 	default:
1319 		return (DDI_FAILURE);
1320 	}
1321 
1322 	rbase = (rootnex_addr_t)rp->regspec_addr &
1323 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1324 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1325 
1326 	if (rp->regspec_size == 0)
1327 		return (DDI_ME_INVAL);
1328 
1329 #ifdef __xpv
1330 	/*
1331 	 * If we're dom0, we're using a real device so we need to translate
1332 	 * the MA to a PA.
1333 	 */
1334 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1335 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1336 		    (rbase & MMU_PAGEOFFSET);
1337 	} else {
1338 		pbase = rbase;
1339 	}
1340 #else
1341 	pbase = rbase;
1342 #endif
1343 
1344 	hp->ah_pfn = mmu_btop(pbase);
1345 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1346 
1347 	return (DDI_SUCCESS);
1348 }
1349 
1350 
1351 
1352 /*
1353  * ************************
1354  *  interrupt related code
1355  * ************************
1356  */
1357 
1358 /*
1359  * rootnex_intr_ops()
1360  *	bus_intr_op() function for interrupt support
1361  */
1362 /* ARGSUSED */
1363 static int
1364 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1365     ddi_intr_handle_impl_t *hdlp, void *result)
1366 {
1367 	struct intrspec			*ispec;
1368 
1369 	DDI_INTR_NEXDBG((CE_CONT,
1370 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1371 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1372 
1373 	/* Process the interrupt operation */
1374 	switch (intr_op) {
1375 	case DDI_INTROP_GETCAP:
1376 		/* First check with pcplusmp */
1377 		if (psm_intr_ops == NULL)
1378 			return (DDI_FAILURE);
1379 
1380 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1381 			*(int *)result = 0;
1382 			return (DDI_FAILURE);
1383 		}
1384 		break;
1385 	case DDI_INTROP_SETCAP:
1386 		if (psm_intr_ops == NULL)
1387 			return (DDI_FAILURE);
1388 
1389 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1390 			return (DDI_FAILURE);
1391 		break;
1392 	case DDI_INTROP_ALLOC:
1393 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1394 		return (rootnex_alloc_intr_fixed(rdip, hdlp, result));
1395 	case DDI_INTROP_FREE:
1396 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1397 		return (rootnex_free_intr_fixed(rdip, hdlp));
1398 	case DDI_INTROP_GETPRI:
1399 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1400 			return (DDI_FAILURE);
1401 		*(int *)result = ispec->intrspec_pri;
1402 		break;
1403 	case DDI_INTROP_SETPRI:
1404 		/* Validate the interrupt priority passed to us */
1405 		if (*(int *)result > LOCK_LEVEL)
1406 			return (DDI_FAILURE);
1407 
1408 		/* Ensure that PSM is all initialized and ispec is ok */
1409 		if ((psm_intr_ops == NULL) ||
1410 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1411 			return (DDI_FAILURE);
1412 
1413 		/* Change the priority */
1414 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1415 		    PSM_FAILURE)
1416 			return (DDI_FAILURE);
1417 
1418 		/* update the ispec with the new priority */
1419 		ispec->intrspec_pri =  *(int *)result;
1420 		break;
1421 	case DDI_INTROP_ADDISR:
1422 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1423 			return (DDI_FAILURE);
1424 		ispec->intrspec_func = hdlp->ih_cb_func;
1425 		break;
1426 	case DDI_INTROP_REMISR:
1427 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1428 			return (DDI_FAILURE);
1429 		ispec->intrspec_func = (uint_t (*)()) 0;
1430 		break;
1431 	case DDI_INTROP_ENABLE:
1432 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1433 			return (DDI_FAILURE);
1434 
1435 		/* Call psmi to translate irq with the dip */
1436 		if (psm_intr_ops == NULL)
1437 			return (DDI_FAILURE);
1438 
1439 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1440 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1441 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1442 			return (DDI_FAILURE);
1443 
1444 		/* Add the interrupt handler */
1445 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1446 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1447 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1448 			return (DDI_FAILURE);
1449 		break;
1450 	case DDI_INTROP_DISABLE:
1451 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1452 			return (DDI_FAILURE);
1453 
1454 		/* Call psm_ops() to translate irq with the dip */
1455 		if (psm_intr_ops == NULL)
1456 			return (DDI_FAILURE);
1457 
1458 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1459 		(void) (*psm_intr_ops)(rdip, hdlp,
1460 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1461 
1462 		/* Remove the interrupt handler */
1463 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1464 		    hdlp->ih_cb_func, hdlp->ih_vector);
1465 		break;
1466 	case DDI_INTROP_SETMASK:
1467 		if (psm_intr_ops == NULL)
1468 			return (DDI_FAILURE);
1469 
1470 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1471 			return (DDI_FAILURE);
1472 		break;
1473 	case DDI_INTROP_CLRMASK:
1474 		if (psm_intr_ops == NULL)
1475 			return (DDI_FAILURE);
1476 
1477 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1478 			return (DDI_FAILURE);
1479 		break;
1480 	case DDI_INTROP_GETPENDING:
1481 		if (psm_intr_ops == NULL)
1482 			return (DDI_FAILURE);
1483 
1484 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1485 		    result)) {
1486 			*(int *)result = 0;
1487 			return (DDI_FAILURE);
1488 		}
1489 		break;
1490 	case DDI_INTROP_NAVAIL:
1491 	case DDI_INTROP_NINTRS:
1492 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1493 		if (*(int *)result == 0) {
1494 			/*
1495 			 * Special case for 'pcic' driver' only. This driver
1496 			 * driver is a child of 'isa' and 'rootnex' drivers.
1497 			 *
1498 			 * See detailed comments on this in the function
1499 			 * rootnex_get_ispec().
1500 			 *
1501 			 * Children of 'pcic' send 'NINITR' request all the
1502 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1503 			 * field may not initialized. So, we fake it here
1504 			 * to return 1 (a la what PCMCIA nexus does).
1505 			 */
1506 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1507 				*(int *)result = 1;
1508 			else
1509 				return (DDI_FAILURE);
1510 		}
1511 		break;
1512 	case DDI_INTROP_SUPPORTED_TYPES:
1513 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1514 		break;
1515 	default:
1516 		return (DDI_FAILURE);
1517 	}
1518 
1519 	return (DDI_SUCCESS);
1520 }
1521 
1522 
1523 /*
1524  * rootnex_get_ispec()
1525  *	convert an interrupt number to an interrupt specification.
1526  *	The interrupt number determines which interrupt spec will be
1527  *	returned if more than one exists.
1528  *
1529  *	Look into the parent private data area of the 'rdip' to find out
1530  *	the interrupt specification.  First check to make sure there is
1531  *	one that matchs "inumber" and then return a pointer to it.
1532  *
1533  *	Return NULL if one could not be found.
1534  *
1535  *	NOTE: This is needed for rootnex_intr_ops()
1536  */
1537 static struct intrspec *
1538 rootnex_get_ispec(dev_info_t *rdip, int inum)
1539 {
1540 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1541 
1542 	/*
1543 	 * Special case handling for drivers that provide their own
1544 	 * intrspec structures instead of relying on the DDI framework.
1545 	 *
1546 	 * A broken hardware driver in ON could potentially provide its
1547 	 * own intrspec structure, instead of relying on the hardware.
1548 	 * If these drivers are children of 'rootnex' then we need to
1549 	 * continue to provide backward compatibility to them here.
1550 	 *
1551 	 * Following check is a special case for 'pcic' driver which
1552 	 * was found to have broken hardwre andby provides its own intrspec.
1553 	 *
1554 	 * Verbatim comments from this driver are shown here:
1555 	 * "Don't use the ddi_add_intr since we don't have a
1556 	 * default intrspec in all cases."
1557 	 *
1558 	 * Since an 'ispec' may not be always created for it,
1559 	 * check for that and create one if so.
1560 	 *
1561 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1562 	 */
1563 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1564 		pdp->par_nintr = 1;
1565 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1566 		    pdp->par_nintr, KM_SLEEP);
1567 	}
1568 
1569 	/* Validate the interrupt number */
1570 	if (inum >= pdp->par_nintr)
1571 		return (NULL);
1572 
1573 	/* Get the interrupt structure pointer and return that */
1574 	return ((struct intrspec *)&pdp->par_intr[inum]);
1575 }
1576 
1577 /*
1578  * Allocate interrupt vector for FIXED (legacy) type.
1579  */
1580 static int
1581 rootnex_alloc_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp,
1582     void *result)
1583 {
1584 	struct intrspec		*ispec;
1585 	ddi_intr_handle_impl_t	info_hdl;
1586 	int			ret;
1587 	int			free_phdl = 0;
1588 	apic_get_type_t		type_info;
1589 
1590 	if (psm_intr_ops == NULL)
1591 		return (DDI_FAILURE);
1592 
1593 	if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1594 		return (DDI_FAILURE);
1595 
1596 	/*
1597 	 * If the PSM module is "APIX" then pass the request for it
1598 	 * to allocate the vector now.
1599 	 */
1600 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1601 	info_hdl.ih_private = &type_info;
1602 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1603 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1604 		if (hdlp->ih_private == NULL) { /* allocate phdl structure */
1605 			free_phdl = 1;
1606 			i_ddi_alloc_intr_phdl(hdlp);
1607 		}
1608 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1609 		ret = (*psm_intr_ops)(rdip, hdlp,
1610 		    PSM_INTR_OP_ALLOC_VECTORS, result);
1611 		if (free_phdl) { /* free up the phdl structure */
1612 			free_phdl = 0;
1613 			i_ddi_free_intr_phdl(hdlp);
1614 			hdlp->ih_private = NULL;
1615 		}
1616 	} else {
1617 		/*
1618 		 * No APIX module; fall back to the old scheme where the
1619 		 * interrupt vector is allocated during ddi_enable_intr() call.
1620 		 */
1621 		hdlp->ih_pri = ispec->intrspec_pri;
1622 		*(int *)result = hdlp->ih_scratch1;
1623 		ret = DDI_SUCCESS;
1624 	}
1625 
1626 	return (ret);
1627 }
1628 
1629 /*
1630  * Free up interrupt vector for FIXED (legacy) type.
1631  */
1632 static int
1633 rootnex_free_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp)
1634 {
1635 	struct intrspec			*ispec;
1636 	struct ddi_parent_private_data	*pdp;
1637 	ddi_intr_handle_impl_t		info_hdl;
1638 	int				ret;
1639 	apic_get_type_t			type_info;
1640 
1641 	if (psm_intr_ops == NULL)
1642 		return (DDI_FAILURE);
1643 
1644 	/*
1645 	 * If the PSM module is "APIX" then pass the request for it
1646 	 * to free up the vector now.
1647 	 */
1648 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1649 	info_hdl.ih_private = &type_info;
1650 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1651 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1652 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1653 			return (DDI_FAILURE);
1654 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1655 		ret = (*psm_intr_ops)(rdip, hdlp,
1656 		    PSM_INTR_OP_FREE_VECTORS, NULL);
1657 	} else {
1658 		/*
1659 		 * No APIX module; fall back to the old scheme where
1660 		 * the interrupt vector was already freed during
1661 		 * ddi_disable_intr() call.
1662 		 */
1663 		ret = DDI_SUCCESS;
1664 	}
1665 
1666 	pdp = ddi_get_parent_data(rdip);
1667 
1668 	/*
1669 	 * Special case for 'pcic' driver' only.
1670 	 * If an intrspec was created for it, clean it up here
1671 	 * See detailed comments on this in the function
1672 	 * rootnex_get_ispec().
1673 	 */
1674 	if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1675 		kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1676 		    pdp->par_nintr);
1677 		/*
1678 		 * Set it to zero; so that
1679 		 * DDI framework doesn't free it again
1680 		 */
1681 		pdp->par_intr = NULL;
1682 		pdp->par_nintr = 0;
1683 	}
1684 
1685 	return (ret);
1686 }
1687 
1688 
1689 /*
1690  * ******************
1691  *  dma related code
1692  * ******************
1693  */
1694 
1695 /*ARGSUSED*/
1696 static int
1697 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1698     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1699     ddi_dma_handle_t *handlep)
1700 {
1701 	uint64_t maxsegmentsize_ll;
1702 	uint_t maxsegmentsize;
1703 	ddi_dma_impl_t *hp;
1704 	rootnex_dma_t *dma;
1705 	uint64_t count_max;
1706 	uint64_t seg;
1707 	int kmflag;
1708 	int e;
1709 
1710 
1711 	/* convert our sleep flags */
1712 	if (waitfp == DDI_DMA_SLEEP) {
1713 		kmflag = KM_SLEEP;
1714 	} else {
1715 		kmflag = KM_NOSLEEP;
1716 	}
1717 
1718 	/*
1719 	 * We try to do only one memory allocation here. We'll do a little
1720 	 * pointer manipulation later. If the bind ends up taking more than
1721 	 * our prealloc's space, we'll have to allocate more memory in the
1722 	 * bind operation. Not great, but much better than before and the
1723 	 * best we can do with the current bind interfaces.
1724 	 */
1725 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1726 	if (hp == NULL) {
1727 		if (waitfp != DDI_DMA_DONTWAIT) {
1728 			ddi_set_callback(waitfp, arg,
1729 			    &rootnex_state->r_dvma_call_list_id);
1730 		}
1731 		return (DDI_DMA_NORESOURCES);
1732 	}
1733 
1734 	/* Do our pointer manipulation now, align the structures */
1735 	hp->dmai_private = (void *)(((uintptr_t)hp +
1736 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1737 	dma = (rootnex_dma_t *)hp->dmai_private;
1738 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1739 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1740 
1741 	/* setup the handle */
1742 	rootnex_clean_dmahdl(hp);
1743 	hp->dmai_error.err_fep = NULL;
1744 	hp->dmai_error.err_cf = NULL;
1745 	dma->dp_dip = rdip;
1746 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1747 	dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1748 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1749 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1750 	hp->dmai_rdip = rdip;
1751 	hp->dmai_attr = *attr;
1752 
1753 	/* we don't need to worry about the SPL since we do a tryenter */
1754 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1755 
1756 	/*
1757 	 * Figure out our maximum segment size. If the segment size is greater
1758 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1759 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1760 	 * dma_attr_count_max are size-1 type values.
1761 	 *
1762 	 * Maximum segment size is the largest physically contiguous chunk of
1763 	 * memory that we can return from a bind (i.e. the maximum size of a
1764 	 * single cookie).
1765 	 */
1766 
1767 	/* handle the rollover cases */
1768 	seg = attr->dma_attr_seg + 1;
1769 	if (seg < attr->dma_attr_seg) {
1770 		seg = attr->dma_attr_seg;
1771 	}
1772 	count_max = attr->dma_attr_count_max + 1;
1773 	if (count_max < attr->dma_attr_count_max) {
1774 		count_max = attr->dma_attr_count_max;
1775 	}
1776 
1777 	/*
1778 	 * granularity may or may not be a power of two. If it isn't, we can't
1779 	 * use a simple mask.
1780 	 */
1781 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1782 		dma->dp_granularity_power_2 = B_FALSE;
1783 	} else {
1784 		dma->dp_granularity_power_2 = B_TRUE;
1785 	}
1786 
1787 	/*
1788 	 * maxxfer should be a whole multiple of granularity. If we're going to
1789 	 * break up a window because we're greater than maxxfer, we might as
1790 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1791 	 * worry about triming the window later on for this case.
1792 	 */
1793 	if (attr->dma_attr_granular > 1) {
1794 		if (dma->dp_granularity_power_2) {
1795 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1796 			    (attr->dma_attr_maxxfer &
1797 			    (attr->dma_attr_granular - 1));
1798 		} else {
1799 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1800 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1801 		}
1802 	} else {
1803 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1804 	}
1805 
1806 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1807 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1808 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1809 		maxsegmentsize = 0xFFFFFFFF;
1810 	} else {
1811 		maxsegmentsize = maxsegmentsize_ll;
1812 	}
1813 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1814 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1815 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1816 
1817 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1818 	if (rootnex_alloc_check_parms) {
1819 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1820 		if (e != DDI_SUCCESS) {
1821 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1822 			(void) rootnex_dma_freehdl(dip, rdip,
1823 			    (ddi_dma_handle_t)hp);
1824 			return (e);
1825 		}
1826 	}
1827 
1828 	*handlep = (ddi_dma_handle_t)hp;
1829 
1830 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1831 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1832 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1833 
1834 	return (DDI_SUCCESS);
1835 }
1836 
1837 
1838 /*
1839  * rootnex_dma_allochdl()
1840  *    called from ddi_dma_alloc_handle().
1841  */
1842 static int
1843 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1844     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1845 {
1846 	int retval;
1847 #if defined(__amd64) && !defined(__xpv)
1848 	uint_t error = ENOTSUP;
1849 
1850 	retval = iommulib_nex_open(rdip, &error);
1851 
1852 	if (retval != DDI_SUCCESS && error == ENOTSUP) {
1853 		/* No IOMMU */
1854 		return (rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1855 		    handlep));
1856 	} else if (retval != DDI_SUCCESS) {
1857 		return (DDI_FAILURE);
1858 	}
1859 
1860 	ASSERT(IOMMU_USED(rdip));
1861 
1862 	/* has an IOMMU */
1863 	retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1864 	    waitfp, arg, handlep);
1865 #else
1866 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1867 	    handlep);
1868 #endif
1869 	if (retval == DDI_SUCCESS)
1870 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1871 	return (retval);
1872 }
1873 
1874 /*ARGSUSED*/
1875 static int
1876 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1877     ddi_dma_handle_t handle)
1878 {
1879 	ddi_dma_impl_t *hp;
1880 	rootnex_dma_t *dma;
1881 
1882 
1883 	hp = (ddi_dma_impl_t *)handle;
1884 	dma = (rootnex_dma_t *)hp->dmai_private;
1885 
1886 	/* unbind should have been called first */
1887 	ASSERT(!dma->dp_inuse);
1888 
1889 	mutex_destroy(&dma->dp_mutex);
1890 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1891 
1892 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1893 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1894 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1895 
1896 	if (rootnex_state->r_dvma_call_list_id)
1897 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1898 
1899 	return (DDI_SUCCESS);
1900 }
1901 
1902 /*
1903  * rootnex_dma_freehdl()
1904  *    called from ddi_dma_free_handle().
1905  */
1906 static int
1907 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1908 {
1909 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1910 #if defined(__amd64) && !defined(__xpv)
1911 	if (IOMMU_USED(rdip)) {
1912 		return (iommulib_nexdma_freehdl(dip, rdip, handle));
1913 	}
1914 #endif
1915 	return (rootnex_coredma_freehdl(dip, rdip, handle));
1916 }
1917 
1918 /*ARGSUSED*/
1919 static int
1920 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1921     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1922     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1923 {
1924 	rootnex_sglinfo_t *sinfo;
1925 	ddi_dma_attr_t *attr;
1926 	ddi_dma_impl_t *hp;
1927 	rootnex_dma_t *dma;
1928 	int kmflag;
1929 	int e;
1930 
1931 	hp = (ddi_dma_impl_t *)handle;
1932 	dma = (rootnex_dma_t *)hp->dmai_private;
1933 	sinfo = &dma->dp_sglinfo;
1934 	attr = &hp->dmai_attr;
1935 
1936 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1937 		dma->dp_sleep_flags = KM_SLEEP;
1938 	} else {
1939 		dma->dp_sleep_flags = KM_NOSLEEP;
1940 	}
1941 
1942 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
1943 
1944 	/*
1945 	 * This is useful for debugging a driver. Not as useful in a production
1946 	 * system. The only time this will fail is if you have a driver bug.
1947 	 */
1948 	if (rootnex_bind_check_inuse) {
1949 		/*
1950 		 * No one else should ever have this lock unless someone else
1951 		 * is trying to use this handle. So contention on the lock
1952 		 * is the same as inuse being set.
1953 		 */
1954 		e = mutex_tryenter(&dma->dp_mutex);
1955 		if (e == 0) {
1956 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1957 			return (DDI_DMA_INUSE);
1958 		}
1959 		if (dma->dp_inuse) {
1960 			mutex_exit(&dma->dp_mutex);
1961 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1962 			return (DDI_DMA_INUSE);
1963 		}
1964 		dma->dp_inuse = B_TRUE;
1965 		mutex_exit(&dma->dp_mutex);
1966 	}
1967 
1968 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1969 	if (rootnex_bind_check_parms) {
1970 		e = rootnex_valid_bind_parms(dmareq, attr);
1971 		if (e != DDI_SUCCESS) {
1972 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1973 			rootnex_clean_dmahdl(hp);
1974 			return (e);
1975 		}
1976 	}
1977 
1978 	/* save away the original bind info */
1979 	dma->dp_dma = dmareq->dmar_object;
1980 
1981 #if defined(__amd64) && !defined(__xpv)
1982 	e = immu_map_sgl(hp, dmareq, rootnex_prealloc_cookies, rdip);
1983 	switch (e) {
1984 	case DDI_DMA_MAPPED:
1985 		goto out;
1986 	case DDI_DMA_USE_PHYSICAL:
1987 		break;
1988 	case DDI_DMA_PARTIAL:
1989 		ddi_err(DER_PANIC, rdip, "Partial DVMA map");
1990 		e = DDI_DMA_NORESOURCES;
1991 		/*FALLTHROUGH*/
1992 	default:
1993 		ddi_err(DER_MODE, rdip, "DVMA map failed");
1994 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1995 		rootnex_clean_dmahdl(hp);
1996 		return (e);
1997 	}
1998 #endif
1999 
2000 	/*
2001 	 * Figure out a rough estimate of what maximum number of pages this
2002 	 * buffer could use (a high estimate of course).
2003 	 */
2004 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
2005 
2006 	/*
2007 	 * We'll use the pre-allocated cookies for any bind that will *always*
2008 	 * fit (more important to be consistent, we don't want to create
2009 	 * additional degenerate cases).
2010 	 */
2011 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
2012 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2013 		dma->dp_need_to_free_cookie = B_FALSE;
2014 		DTRACE_PROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
2015 		    uint_t, sinfo->si_max_pages);
2016 
2017 	/*
2018 	 * For anything larger than that, we'll go ahead and allocate the
2019 	 * maximum number of pages we expect to see. Hopefuly, we won't be
2020 	 * seeing this path in the fast path for high performance devices very
2021 	 * frequently.
2022 	 *
2023 	 * a ddi bind interface that allowed the driver to provide storage to
2024 	 * the bind interface would speed this case up.
2025 	 */
2026 	} else {
2027 		/* convert the sleep flags */
2028 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
2029 			kmflag =  KM_SLEEP;
2030 		} else {
2031 			kmflag =  KM_NOSLEEP;
2032 		}
2033 
2034 		/*
2035 		 * Save away how much memory we allocated. If we're doing a
2036 		 * nosleep, the alloc could fail...
2037 		 */
2038 		dma->dp_cookie_size = sinfo->si_max_pages *
2039 		    sizeof (ddi_dma_cookie_t);
2040 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
2041 		if (dma->dp_cookies == NULL) {
2042 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2043 			rootnex_clean_dmahdl(hp);
2044 			return (DDI_DMA_NORESOURCES);
2045 		}
2046 		dma->dp_need_to_free_cookie = B_TRUE;
2047 		DTRACE_PROBE2(rootnex__bind__alloc, dev_info_t *, rdip, uint_t,
2048 		    sinfo->si_max_pages);
2049 	}
2050 	hp->dmai_cookie = dma->dp_cookies;
2051 
2052 	/*
2053 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
2054 	 * looking at the constraints in the dma structure. It will then put
2055 	 * some additional state about the sgl in the dma struct (i.e. is
2056 	 * the sgl clean, or do we need to do some munging; how many pages
2057 	 * need to be copied, etc.)
2058 	 */
2059 	rootnex_get_sgl(&dmareq->dmar_object, dma->dp_cookies,
2060 	    &dma->dp_sglinfo);
2061 
2062 out:
2063 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
2064 	/* if we don't need a copy buffer, we don't need to sync */
2065 	if (sinfo->si_copybuf_req == 0) {
2066 		hp->dmai_rflags |= DMP_NOSYNC;
2067 	}
2068 
2069 	/*
2070 	 * if we don't need the copybuf and we don't need to do a partial,  we
2071 	 * hit the fast path. All the high performance devices should be trying
2072 	 * to hit this path. To hit this path, a device should be able to reach
2073 	 * all of memory, shouldn't try to bind more than it can transfer, and
2074 	 * the buffer shouldn't require more cookies than the driver/device can
2075 	 * handle [sgllen]).
2076 	 */
2077 	if ((sinfo->si_copybuf_req == 0) &&
2078 	    (sinfo->si_sgl_size <= attr->dma_attr_sgllen) &&
2079 	    (dma->dp_dma.dmao_size < dma->dp_maxxfer)) {
2080 		/*
2081 		 * If the driver supports FMA, insert the handle in the FMA DMA
2082 		 * handle cache.
2083 		 */
2084 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2085 			hp->dmai_error.err_cf = rootnex_dma_check;
2086 
2087 		/*
2088 		 * copy out the first cookie and ccountp, set the cookie
2089 		 * pointer to the second cookie. The first cookie is passed
2090 		 * back on the stack. Additional cookies are accessed via
2091 		 * ddi_dma_nextcookie()
2092 		 */
2093 		*cookiep = dma->dp_cookies[0];
2094 		*ccountp = sinfo->si_sgl_size;
2095 		hp->dmai_cookie++;
2096 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2097 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2098 		DTRACE_PROBE3(rootnex__bind__fast, dev_info_t *, rdip,
2099 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2100 		    uint_t, dma->dp_dma.dmao_size);
2101 
2102 
2103 		return (DDI_DMA_MAPPED);
2104 	}
2105 
2106 	/*
2107 	 * go to the slow path, we may need to alloc more memory, create
2108 	 * multiple windows, and munge up a sgl to make the device happy.
2109 	 */
2110 	e = rootnex_bind_slowpath(hp, dmareq, dma, attr, kmflag);
2111 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2112 		if (dma->dp_need_to_free_cookie) {
2113 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2114 		}
2115 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2116 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2117 		return (e);
2118 	}
2119 
2120 	/*
2121 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2122 	 * cache.
2123 	 */
2124 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2125 		hp->dmai_error.err_cf = rootnex_dma_check;
2126 
2127 	/* if the first window uses the copy buffer, sync it for the device */
2128 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2129 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2130 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2131 		    DDI_DMA_SYNC_FORDEV);
2132 	}
2133 
2134 	/*
2135 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2136 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2137 	 * If we have a partial map (i.e. multiple windows), the number of
2138 	 * cookies we return is the number of cookies in the first window.
2139 	 */
2140 	if (e == DDI_DMA_MAPPED) {
2141 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2142 		*ccountp = sinfo->si_sgl_size;
2143 		hp->dmai_nwin = 1;
2144 	} else {
2145 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2146 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2147 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2148 	}
2149 	*cookiep = dma->dp_cookies[0];
2150 	hp->dmai_cookie++;
2151 
2152 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2153 	ROOTNEX_DPROBE3(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2154 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2155 	    dma->dp_dma.dmao_size);
2156 	return (e);
2157 }
2158 
2159 /*
2160  * rootnex_dma_bindhdl()
2161  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2162  */
2163 static int
2164 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2165     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2166     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2167 {
2168 #if defined(__amd64) && !defined(__xpv)
2169 	if (IOMMU_USED(rdip)) {
2170 		return (iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2171 		    cookiep, ccountp));
2172 	}
2173 #endif
2174 	return (rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2175 	    cookiep, ccountp));
2176 }
2177 
2178 
2179 
2180 /*ARGSUSED*/
2181 static int
2182 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2183     ddi_dma_handle_t handle)
2184 {
2185 	ddi_dma_impl_t *hp;
2186 	rootnex_dma_t *dma;
2187 	int e;
2188 
2189 	hp = (ddi_dma_impl_t *)handle;
2190 	dma = (rootnex_dma_t *)hp->dmai_private;
2191 
2192 	/* make sure the buffer wasn't free'd before calling unbind */
2193 	if (rootnex_unbind_verify_buffer) {
2194 		e = rootnex_verify_buffer(dma);
2195 		if (e != DDI_SUCCESS) {
2196 			ASSERT(0);
2197 			return (DDI_FAILURE);
2198 		}
2199 	}
2200 
2201 	/* sync the current window before unbinding the buffer */
2202 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2203 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2204 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2205 		    DDI_DMA_SYNC_FORCPU);
2206 	}
2207 
2208 	/*
2209 	 * cleanup and copy buffer or window state. if we didn't use the copy
2210 	 * buffer or windows, there won't be much to do :-)
2211 	 */
2212 	rootnex_teardown_copybuf(dma);
2213 	rootnex_teardown_windows(dma);
2214 
2215 #if defined(__amd64) && !defined(__xpv)
2216 	/*
2217 	 * Clean up the page tables and free the dvma
2218 	 */
2219 	e = immu_unmap_sgl(hp, rdip);
2220 	if (e != DDI_DMA_USE_PHYSICAL && e != DDI_SUCCESS) {
2221 		return (e);
2222 	}
2223 #endif
2224 
2225 	/*
2226 	 * If we had to allocate space to for the worse case sgl (it didn't
2227 	 * fit into our pre-allocate buffer), free that up now
2228 	 */
2229 	if (dma->dp_need_to_free_cookie) {
2230 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2231 	}
2232 
2233 	/*
2234 	 * clean up the handle so it's ready for the next bind (i.e. if the
2235 	 * handle is reused).
2236 	 */
2237 	rootnex_clean_dmahdl(hp);
2238 	hp->dmai_error.err_cf = NULL;
2239 
2240 	if (rootnex_state->r_dvma_call_list_id)
2241 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2242 
2243 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2244 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2245 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2246 
2247 	return (DDI_SUCCESS);
2248 }
2249 
2250 /*
2251  * rootnex_dma_unbindhdl()
2252  *    called from ddi_dma_unbind_handle()
2253  */
2254 /*ARGSUSED*/
2255 static int
2256 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2257     ddi_dma_handle_t handle)
2258 {
2259 #if defined(__amd64) && !defined(__xpv)
2260 	if (IOMMU_USED(rdip)) {
2261 		return (iommulib_nexdma_unbindhdl(dip, rdip, handle));
2262 	}
2263 #endif
2264 	return (rootnex_coredma_unbindhdl(dip, rdip, handle));
2265 }
2266 
2267 #if defined(__amd64) && !defined(__xpv)
2268 
2269 static int
2270 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2271 {
2272 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2273 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2274 
2275 	if (dma->dp_sleep_flags != KM_SLEEP &&
2276 	    dma->dp_sleep_flags != KM_NOSLEEP)
2277 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2278 	return (dma->dp_sleep_flags);
2279 }
2280 /*ARGSUSED*/
2281 static void
2282 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2283 {
2284 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2285 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2286 	rootnex_window_t *window;
2287 
2288 	if (dma->dp_window) {
2289 		window = &dma->dp_window[dma->dp_current_win];
2290 		hp->dmai_cookie = window->wd_first_cookie;
2291 	} else {
2292 		hp->dmai_cookie = dma->dp_cookies;
2293 	}
2294 	hp->dmai_cookie++;
2295 }
2296 
2297 /*ARGSUSED*/
2298 static int
2299 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2300     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2301 {
2302 	int i;
2303 	int km_flags;
2304 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2305 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2306 	rootnex_window_t *window;
2307 	ddi_dma_cookie_t *cp;
2308 	ddi_dma_cookie_t *cookie;
2309 
2310 	ASSERT(*cookiepp == NULL);
2311 	ASSERT(*ccountp == 0);
2312 
2313 	if (dma->dp_window) {
2314 		window = &dma->dp_window[dma->dp_current_win];
2315 		cp = window->wd_first_cookie;
2316 		*ccountp = window->wd_cookie_cnt;
2317 	} else {
2318 		cp = dma->dp_cookies;
2319 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2320 	}
2321 
2322 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2323 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2324 	if (cookie == NULL) {
2325 		return (DDI_DMA_NORESOURCES);
2326 	}
2327 
2328 	for (i = 0; i < *ccountp; i++) {
2329 		cookie[i].dmac_notused = cp[i].dmac_notused;
2330 		cookie[i].dmac_type = cp[i].dmac_type;
2331 		cookie[i].dmac_address = cp[i].dmac_address;
2332 		cookie[i].dmac_size = cp[i].dmac_size;
2333 	}
2334 
2335 	*cookiepp = cookie;
2336 
2337 	return (DDI_SUCCESS);
2338 }
2339 
2340 /*ARGSUSED*/
2341 static int
2342 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2343     ddi_dma_cookie_t *cookiep, uint_t ccount)
2344 {
2345 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2346 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2347 	rootnex_window_t *window;
2348 	ddi_dma_cookie_t *cur_cookiep;
2349 
2350 	ASSERT(cookiep);
2351 	ASSERT(ccount != 0);
2352 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2353 
2354 	if (dma->dp_window) {
2355 		window = &dma->dp_window[dma->dp_current_win];
2356 		dma->dp_saved_cookies = window->wd_first_cookie;
2357 		window->wd_first_cookie = cookiep;
2358 		ASSERT(ccount == window->wd_cookie_cnt);
2359 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2360 		    + window->wd_first_cookie;
2361 	} else {
2362 		dma->dp_saved_cookies = dma->dp_cookies;
2363 		dma->dp_cookies = cookiep;
2364 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2365 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2366 		    + dma->dp_cookies;
2367 	}
2368 
2369 	dma->dp_need_to_switch_cookies = B_TRUE;
2370 	hp->dmai_cookie = cur_cookiep;
2371 
2372 	return (DDI_SUCCESS);
2373 }
2374 
2375 /*ARGSUSED*/
2376 static int
2377 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2378 {
2379 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2380 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2381 	rootnex_window_t *window;
2382 	ddi_dma_cookie_t *cur_cookiep;
2383 	ddi_dma_cookie_t *cookie_array;
2384 	uint_t ccount;
2385 
2386 	/* check if cookies have not been switched */
2387 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2388 		return (DDI_SUCCESS);
2389 
2390 	ASSERT(dma->dp_saved_cookies);
2391 
2392 	if (dma->dp_window) {
2393 		window = &dma->dp_window[dma->dp_current_win];
2394 		cookie_array = window->wd_first_cookie;
2395 		window->wd_first_cookie = dma->dp_saved_cookies;
2396 		dma->dp_saved_cookies = NULL;
2397 		ccount = window->wd_cookie_cnt;
2398 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2399 		    + window->wd_first_cookie;
2400 	} else {
2401 		cookie_array = dma->dp_cookies;
2402 		dma->dp_cookies = dma->dp_saved_cookies;
2403 		dma->dp_saved_cookies = NULL;
2404 		ccount = dma->dp_sglinfo.si_sgl_size;
2405 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2406 		    + dma->dp_cookies;
2407 	}
2408 
2409 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2410 
2411 	hp->dmai_cookie = cur_cookiep;
2412 
2413 	dma->dp_need_to_switch_cookies = B_FALSE;
2414 
2415 	return (DDI_SUCCESS);
2416 }
2417 
2418 #endif
2419 
2420 /*
2421  * rootnex_verify_buffer()
2422  *   verify buffer wasn't free'd
2423  */
2424 static int
2425 rootnex_verify_buffer(rootnex_dma_t *dma)
2426 {
2427 	page_t **pplist;
2428 	caddr_t vaddr;
2429 	uint_t pcnt;
2430 	uint_t poff;
2431 	page_t *pp;
2432 	char b;
2433 	int i;
2434 
2435 	/* Figure out how many pages this buffer occupies */
2436 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2437 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2438 	} else {
2439 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2440 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2441 	}
2442 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2443 
2444 	switch (dma->dp_dma.dmao_type) {
2445 	case DMA_OTYP_PAGES:
2446 		/*
2447 		 * for a linked list of pp's walk through them to make sure
2448 		 * they're locked and not free.
2449 		 */
2450 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2451 		for (i = 0; i < pcnt; i++) {
2452 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2453 				return (DDI_FAILURE);
2454 			}
2455 			pp = pp->p_next;
2456 		}
2457 		break;
2458 
2459 	case DMA_OTYP_VADDR:
2460 	case DMA_OTYP_BUFVADDR:
2461 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2462 		/*
2463 		 * for an array of pp's walk through them to make sure they're
2464 		 * not free. It's possible that they may not be locked.
2465 		 */
2466 		if (pplist) {
2467 			for (i = 0; i < pcnt; i++) {
2468 				if (PP_ISFREE(pplist[i])) {
2469 					return (DDI_FAILURE);
2470 				}
2471 			}
2472 
2473 		/* For a virtual address, try to peek at each page */
2474 		} else {
2475 			if (dma->dp_sglinfo.si_asp == &kas) {
2476 				for (i = 0; i < pcnt; i++) {
2477 					if (ddi_peek8(NULL, vaddr, &b) ==
2478 					    DDI_FAILURE)
2479 						return (DDI_FAILURE);
2480 					vaddr += MMU_PAGESIZE;
2481 				}
2482 			}
2483 		}
2484 		break;
2485 
2486 	default:
2487 		ASSERT(0);
2488 		break;
2489 	}
2490 
2491 	return (DDI_SUCCESS);
2492 }
2493 
2494 
2495 /*
2496  * rootnex_clean_dmahdl()
2497  *    Clean the dma handle. This should be called on a handle alloc and an
2498  *    unbind handle. Set the handle state to the default settings.
2499  */
2500 static void
2501 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2502 {
2503 	rootnex_dma_t *dma;
2504 
2505 
2506 	dma = (rootnex_dma_t *)hp->dmai_private;
2507 
2508 	hp->dmai_nwin = 0;
2509 	dma->dp_current_cookie = 0;
2510 	dma->dp_copybuf_size = 0;
2511 	dma->dp_window = NULL;
2512 	dma->dp_cbaddr = NULL;
2513 	dma->dp_inuse = B_FALSE;
2514 	dma->dp_need_to_free_cookie = B_FALSE;
2515 	dma->dp_need_to_switch_cookies = B_FALSE;
2516 	dma->dp_saved_cookies = NULL;
2517 	dma->dp_sleep_flags = KM_PANIC;
2518 	dma->dp_need_to_free_window = B_FALSE;
2519 	dma->dp_partial_required = B_FALSE;
2520 	dma->dp_trim_required = B_FALSE;
2521 	dma->dp_sglinfo.si_copybuf_req = 0;
2522 #if !defined(__amd64)
2523 	dma->dp_cb_remaping = B_FALSE;
2524 	dma->dp_kva = NULL;
2525 #endif
2526 
2527 	/* FMA related initialization */
2528 	hp->dmai_fault = 0;
2529 	hp->dmai_fault_check = NULL;
2530 	hp->dmai_fault_notify = NULL;
2531 	hp->dmai_error.err_ena = 0;
2532 	hp->dmai_error.err_status = DDI_FM_OK;
2533 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2534 	hp->dmai_error.err_ontrap = NULL;
2535 }
2536 
2537 
2538 /*
2539  * rootnex_valid_alloc_parms()
2540  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2541  */
2542 static int
2543 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2544 {
2545 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2546 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2547 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2548 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2549 		return (DDI_DMA_BADATTR);
2550 	}
2551 
2552 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2553 		return (DDI_DMA_BADATTR);
2554 	}
2555 
2556 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2557 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2558 	    attr->dma_attr_sgllen <= 0) {
2559 		return (DDI_DMA_BADATTR);
2560 	}
2561 
2562 	/* We should be able to DMA into every byte offset in a page */
2563 	if (maxsegmentsize < MMU_PAGESIZE) {
2564 		return (DDI_DMA_BADATTR);
2565 	}
2566 
2567 	/* if we're bouncing on seg, seg must be <= addr_hi */
2568 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2569 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2570 		return (DDI_DMA_BADATTR);
2571 	}
2572 	return (DDI_SUCCESS);
2573 }
2574 
2575 /*
2576  * rootnex_valid_bind_parms()
2577  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2578  */
2579 /* ARGSUSED */
2580 static int
2581 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2582 {
2583 #if !defined(__amd64)
2584 	/*
2585 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2586 	 * we can track the offset for the obsoleted interfaces.
2587 	 */
2588 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2589 		return (DDI_DMA_TOOBIG);
2590 	}
2591 #endif
2592 
2593 	return (DDI_SUCCESS);
2594 }
2595 
2596 
2597 /*
2598  * rootnex_need_bounce_seg()
2599  *    check to see if the buffer lives on both side of the seg.
2600  */
2601 static boolean_t
2602 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2603 {
2604 	ddi_dma_atyp_t buftype;
2605 	rootnex_addr_t raddr;
2606 	boolean_t lower_addr;
2607 	boolean_t upper_addr;
2608 	uint64_t offset;
2609 	page_t **pplist;
2610 	uint64_t paddr;
2611 	uint32_t psize;
2612 	uint32_t size;
2613 	caddr_t vaddr;
2614 	uint_t pcnt;
2615 	page_t *pp;
2616 
2617 
2618 	/* shortcuts */
2619 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2620 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2621 	buftype = dmar_object->dmao_type;
2622 	size = dmar_object->dmao_size;
2623 
2624 	lower_addr = B_FALSE;
2625 	upper_addr = B_FALSE;
2626 	pcnt = 0;
2627 
2628 	/*
2629 	 * Process the first page to handle the initial offset of the buffer.
2630 	 * We'll use the base address we get later when we loop through all
2631 	 * the pages.
2632 	 */
2633 	if (buftype == DMA_OTYP_PAGES) {
2634 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2635 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2636 		    MMU_PAGEOFFSET;
2637 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2638 		psize = MIN(size, (MMU_PAGESIZE - offset));
2639 		pp = pp->p_next;
2640 		sglinfo->si_asp = NULL;
2641 	} else if (pplist != NULL) {
2642 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2643 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2644 		if (sglinfo->si_asp == NULL) {
2645 			sglinfo->si_asp = &kas;
2646 		}
2647 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2648 		paddr += offset;
2649 		psize = MIN(size, (MMU_PAGESIZE - offset));
2650 		pcnt++;
2651 	} else {
2652 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2653 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2654 		if (sglinfo->si_asp == NULL) {
2655 			sglinfo->si_asp = &kas;
2656 		}
2657 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2658 		paddr += offset;
2659 		psize = MIN(size, (MMU_PAGESIZE - offset));
2660 		vaddr += psize;
2661 	}
2662 
2663 #ifdef __xpv
2664 	/*
2665 	 * If we're dom0, we're using a real device so we need to load
2666 	 * the cookies with MFNs instead of PFNs.
2667 	 */
2668 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2669 #else
2670 	raddr = paddr;
2671 #endif
2672 
2673 	if ((raddr + psize) > sglinfo->si_segmask) {
2674 		upper_addr = B_TRUE;
2675 	} else {
2676 		lower_addr = B_TRUE;
2677 	}
2678 	size -= psize;
2679 
2680 	/*
2681 	 * Walk through the rest of the pages in the buffer. Track to see
2682 	 * if we have pages on both sides of the segment boundary.
2683 	 */
2684 	while (size > 0) {
2685 		/* partial or full page */
2686 		psize = MIN(size, MMU_PAGESIZE);
2687 
2688 		if (buftype == DMA_OTYP_PAGES) {
2689 			/* get the paddr from the page_t */
2690 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2691 			paddr = pfn_to_pa(pp->p_pagenum);
2692 			pp = pp->p_next;
2693 		} else if (pplist != NULL) {
2694 			/* index into the array of page_t's to get the paddr */
2695 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2696 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2697 			pcnt++;
2698 		} else {
2699 			/* call into the VM to get the paddr */
2700 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2701 			    vaddr));
2702 			vaddr += psize;
2703 		}
2704 
2705 #ifdef __xpv
2706 		/*
2707 		 * If we're dom0, we're using a real device so we need to load
2708 		 * the cookies with MFNs instead of PFNs.
2709 		 */
2710 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2711 #else
2712 		raddr = paddr;
2713 #endif
2714 
2715 		if ((raddr + psize) > sglinfo->si_segmask) {
2716 			upper_addr = B_TRUE;
2717 		} else {
2718 			lower_addr = B_TRUE;
2719 		}
2720 		/*
2721 		 * if the buffer lives both above and below the segment
2722 		 * boundary, or the current page is the page immediately
2723 		 * after the segment, we will use a copy/bounce buffer for
2724 		 * all pages > seg.
2725 		 */
2726 		if ((lower_addr && upper_addr) ||
2727 		    (raddr == (sglinfo->si_segmask + 1))) {
2728 			return (B_TRUE);
2729 		}
2730 
2731 		size -= psize;
2732 	}
2733 
2734 	return (B_FALSE);
2735 }
2736 
2737 
2738 /*
2739  * rootnex_get_sgl()
2740  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2741  *    with a call to the vm layer when vm2.0 comes around...
2742  */
2743 static void
2744 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2745     rootnex_sglinfo_t *sglinfo)
2746 {
2747 	ddi_dma_atyp_t buftype;
2748 	rootnex_addr_t raddr;
2749 	uint64_t last_page;
2750 	uint64_t offset;
2751 	uint64_t addrhi;
2752 	uint64_t addrlo;
2753 	uint64_t maxseg;
2754 	page_t **pplist;
2755 	uint64_t paddr;
2756 	uint32_t psize;
2757 	uint32_t size;
2758 	caddr_t vaddr;
2759 	uint_t pcnt;
2760 	page_t *pp;
2761 	uint_t cnt;
2762 
2763 
2764 	/* shortcuts */
2765 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2766 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2767 	maxseg = sglinfo->si_max_cookie_size;
2768 	buftype = dmar_object->dmao_type;
2769 	addrhi = sglinfo->si_max_addr;
2770 	addrlo = sglinfo->si_min_addr;
2771 	size = dmar_object->dmao_size;
2772 
2773 	pcnt = 0;
2774 	cnt = 0;
2775 
2776 
2777 	/*
2778 	 * check to see if we need to use the copy buffer for pages over
2779 	 * the segment attr.
2780 	 */
2781 	sglinfo->si_bounce_on_seg = B_FALSE;
2782 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2783 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2784 		    dmar_object, sglinfo);
2785 	}
2786 
2787 	/*
2788 	 * if we were passed down a linked list of pages, i.e. pointer to
2789 	 * page_t, use this to get our physical address and buf offset.
2790 	 */
2791 	if (buftype == DMA_OTYP_PAGES) {
2792 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2793 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2794 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2795 		    MMU_PAGEOFFSET;
2796 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2797 		psize = MIN(size, (MMU_PAGESIZE - offset));
2798 		pp = pp->p_next;
2799 		sglinfo->si_asp = NULL;
2800 
2801 	/*
2802 	 * We weren't passed down a linked list of pages, but if we were passed
2803 	 * down an array of pages, use this to get our physical address and buf
2804 	 * offset.
2805 	 */
2806 	} else if (pplist != NULL) {
2807 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2808 		    (buftype == DMA_OTYP_BUFVADDR));
2809 
2810 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2811 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2812 		if (sglinfo->si_asp == NULL) {
2813 			sglinfo->si_asp = &kas;
2814 		}
2815 
2816 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2817 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2818 		paddr += offset;
2819 		psize = MIN(size, (MMU_PAGESIZE - offset));
2820 		pcnt++;
2821 
2822 	/*
2823 	 * All we have is a virtual address, we'll need to call into the VM
2824 	 * to get the physical address.
2825 	 */
2826 	} else {
2827 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2828 		    (buftype == DMA_OTYP_BUFVADDR));
2829 
2830 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2831 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2832 		if (sglinfo->si_asp == NULL) {
2833 			sglinfo->si_asp = &kas;
2834 		}
2835 
2836 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2837 		paddr += offset;
2838 		psize = MIN(size, (MMU_PAGESIZE - offset));
2839 		vaddr += psize;
2840 	}
2841 
2842 #ifdef __xpv
2843 	/*
2844 	 * If we're dom0, we're using a real device so we need to load
2845 	 * the cookies with MFNs instead of PFNs.
2846 	 */
2847 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2848 #else
2849 	raddr = paddr;
2850 #endif
2851 
2852 	/*
2853 	 * Setup the first cookie with the physical address of the page and the
2854 	 * size of the page (which takes into account the initial offset into
2855 	 * the page.
2856 	 */
2857 	sgl[cnt].dmac_laddress = raddr;
2858 	sgl[cnt].dmac_size = psize;
2859 	sgl[cnt].dmac_type = 0;
2860 
2861 	/*
2862 	 * Save away the buffer offset into the page. We'll need this later in
2863 	 * the copy buffer code to help figure out the page index within the
2864 	 * buffer and the offset into the current page.
2865 	 */
2866 	sglinfo->si_buf_offset = offset;
2867 
2868 	/*
2869 	 * If we are using the copy buffer for anything over the segment
2870 	 * boundary, and this page is over the segment boundary.
2871 	 *   OR
2872 	 * if the DMA engine can't reach the physical address.
2873 	 */
2874 	if (((sglinfo->si_bounce_on_seg) &&
2875 	    ((raddr + psize) > sglinfo->si_segmask)) ||
2876 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2877 		/*
2878 		 * Increase how much copy buffer we use. We always increase by
2879 		 * pagesize so we don't have to worry about converting offsets.
2880 		 * Set a flag in the cookies dmac_type to indicate that it uses
2881 		 * the copy buffer. If this isn't the last cookie, go to the
2882 		 * next cookie (since we separate each page which uses the copy
2883 		 * buffer in case the copy buffer is not physically contiguous.
2884 		 */
2885 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
2886 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2887 		if ((cnt + 1) < sglinfo->si_max_pages) {
2888 			cnt++;
2889 			sgl[cnt].dmac_laddress = 0;
2890 			sgl[cnt].dmac_size = 0;
2891 			sgl[cnt].dmac_type = 0;
2892 		}
2893 	}
2894 
2895 	/*
2896 	 * save this page's physical address so we can figure out if the next
2897 	 * page is physically contiguous. Keep decrementing size until we are
2898 	 * done with the buffer.
2899 	 */
2900 	last_page = raddr & MMU_PAGEMASK;
2901 	size -= psize;
2902 
2903 	while (size > 0) {
2904 		/* Get the size for this page (i.e. partial or full page) */
2905 		psize = MIN(size, MMU_PAGESIZE);
2906 
2907 		if (buftype == DMA_OTYP_PAGES) {
2908 			/* get the paddr from the page_t */
2909 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2910 			paddr = pfn_to_pa(pp->p_pagenum);
2911 			pp = pp->p_next;
2912 		} else if (pplist != NULL) {
2913 			/* index into the array of page_t's to get the paddr */
2914 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2915 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2916 			pcnt++;
2917 		} else {
2918 			/* call into the VM to get the paddr */
2919 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2920 			    vaddr));
2921 			vaddr += psize;
2922 		}
2923 
2924 #ifdef __xpv
2925 		/*
2926 		 * If we're dom0, we're using a real device so we need to load
2927 		 * the cookies with MFNs instead of PFNs.
2928 		 */
2929 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2930 #else
2931 		raddr = paddr;
2932 #endif
2933 
2934 		/*
2935 		 * If we are using the copy buffer for anything over the
2936 		 * segment boundary, and this page is over the segment
2937 		 * boundary.
2938 		 *   OR
2939 		 * if the DMA engine can't reach the physical address.
2940 		 */
2941 		if (((sglinfo->si_bounce_on_seg) &&
2942 		    ((raddr + psize) > sglinfo->si_segmask)) ||
2943 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2944 
2945 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
2946 
2947 			/*
2948 			 * if there is something in the current cookie, go to
2949 			 * the next one. We only want one page in a cookie which
2950 			 * uses the copybuf since the copybuf doesn't have to
2951 			 * be physically contiguous.
2952 			 */
2953 			if (sgl[cnt].dmac_size != 0) {
2954 				cnt++;
2955 			}
2956 			sgl[cnt].dmac_laddress = raddr;
2957 			sgl[cnt].dmac_size = psize;
2958 #if defined(__amd64)
2959 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2960 #else
2961 			/*
2962 			 * save the buf offset for 32-bit kernel. used in the
2963 			 * obsoleted interfaces.
2964 			 */
2965 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
2966 			    (dmar_object->dmao_size - size);
2967 #endif
2968 			/* if this isn't the last cookie, go to the next one */
2969 			if ((cnt + 1) < sglinfo->si_max_pages) {
2970 				cnt++;
2971 				sgl[cnt].dmac_laddress = 0;
2972 				sgl[cnt].dmac_size = 0;
2973 				sgl[cnt].dmac_type = 0;
2974 			}
2975 
2976 		/*
2977 		 * this page didn't need the copy buffer, if it's not physically
2978 		 * contiguous, or it would put us over a segment boundary, or it
2979 		 * puts us over the max cookie size, or the current sgl doesn't
2980 		 * have anything in it.
2981 		 */
2982 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
2983 		    !(raddr & sglinfo->si_segmask) ||
2984 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
2985 		    (sgl[cnt].dmac_size == 0)) {
2986 			/*
2987 			 * if we're not already in a new cookie, go to the next
2988 			 * cookie.
2989 			 */
2990 			if (sgl[cnt].dmac_size != 0) {
2991 				cnt++;
2992 			}
2993 
2994 			/* save the cookie information */
2995 			sgl[cnt].dmac_laddress = raddr;
2996 			sgl[cnt].dmac_size = psize;
2997 #if defined(__amd64)
2998 			sgl[cnt].dmac_type = 0;
2999 #else
3000 			/*
3001 			 * save the buf offset for 32-bit kernel. used in the
3002 			 * obsoleted interfaces.
3003 			 */
3004 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
3005 #endif
3006 
3007 		/*
3008 		 * this page didn't need the copy buffer, it is physically
3009 		 * contiguous with the last page, and it's <= the max cookie
3010 		 * size.
3011 		 */
3012 		} else {
3013 			sgl[cnt].dmac_size += psize;
3014 
3015 			/*
3016 			 * if this exactly ==  the maximum cookie size, and
3017 			 * it isn't the last cookie, go to the next cookie.
3018 			 */
3019 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3020 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3021 				cnt++;
3022 				sgl[cnt].dmac_laddress = 0;
3023 				sgl[cnt].dmac_size = 0;
3024 				sgl[cnt].dmac_type = 0;
3025 			}
3026 		}
3027 
3028 		/*
3029 		 * save this page's physical address so we can figure out if the
3030 		 * next page is physically contiguous. Keep decrementing size
3031 		 * until we are done with the buffer.
3032 		 */
3033 		last_page = raddr;
3034 		size -= psize;
3035 	}
3036 
3037 	/* we're done, save away how many cookies the sgl has */
3038 	if (sgl[cnt].dmac_size == 0) {
3039 		ASSERT(cnt < sglinfo->si_max_pages);
3040 		sglinfo->si_sgl_size = cnt;
3041 	} else {
3042 		sglinfo->si_sgl_size = cnt + 1;
3043 	}
3044 }
3045 
3046 /*
3047  * rootnex_bind_slowpath()
3048  *    Call in the bind path if the calling driver can't use the sgl without
3049  *    modifying it. We either need to use the copy buffer and/or we will end up
3050  *    with a partial bind.
3051  */
3052 static int
3053 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3054     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag)
3055 {
3056 	rootnex_sglinfo_t *sinfo;
3057 	rootnex_window_t *window;
3058 	ddi_dma_cookie_t *cookie;
3059 	size_t copybuf_used;
3060 	size_t dmac_size;
3061 	boolean_t partial;
3062 	off_t cur_offset;
3063 	page_t *cur_pp;
3064 	major_t mnum;
3065 	int e;
3066 	int i;
3067 
3068 
3069 	sinfo = &dma->dp_sglinfo;
3070 	copybuf_used = 0;
3071 	partial = B_FALSE;
3072 
3073 	/*
3074 	 * If we're using the copybuf, set the copybuf state in dma struct.
3075 	 * Needs to be first since it sets the copy buffer size.
3076 	 */
3077 	if (sinfo->si_copybuf_req != 0) {
3078 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
3079 		if (e != DDI_SUCCESS) {
3080 			return (e);
3081 		}
3082 	} else {
3083 		dma->dp_copybuf_size = 0;
3084 	}
3085 
3086 	/*
3087 	 * Figure out if we need to do a partial mapping. If so, figure out
3088 	 * if we need to trim the buffers when we munge the sgl.
3089 	 */
3090 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
3091 	    (dma->dp_dma.dmao_size > dma->dp_maxxfer) ||
3092 	    (attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
3093 		dma->dp_partial_required = B_TRUE;
3094 		if (attr->dma_attr_granular != 1) {
3095 			dma->dp_trim_required = B_TRUE;
3096 		}
3097 	} else {
3098 		dma->dp_partial_required = B_FALSE;
3099 		dma->dp_trim_required = B_FALSE;
3100 	}
3101 
3102 	/* If we need to do a partial bind, make sure the driver supports it */
3103 	if (dma->dp_partial_required &&
3104 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3105 
3106 		mnum = ddi_driver_major(dma->dp_dip);
3107 		/*
3108 		 * patchable which allows us to print one warning per major
3109 		 * number.
3110 		 */
3111 		if ((rootnex_bind_warn) &&
3112 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3113 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3114 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3115 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3116 			    "There is a small risk of data corruption in "
3117 			    "particular with large I/Os. The driver should be "
3118 			    "replaced with a corrected version for proper "
3119 			    "system operation. To disable this warning, add "
3120 			    "'set rootnex:rootnex_bind_warn=0' to "
3121 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3122 		}
3123 		return (DDI_DMA_TOOBIG);
3124 	}
3125 
3126 	/*
3127 	 * we might need multiple windows, setup state to handle them. In this
3128 	 * code path, we will have at least one window.
3129 	 */
3130 	e = rootnex_setup_windows(hp, dma, attr, kmflag);
3131 	if (e != DDI_SUCCESS) {
3132 		rootnex_teardown_copybuf(dma);
3133 		return (e);
3134 	}
3135 
3136 	window = &dma->dp_window[0];
3137 	cookie = &dma->dp_cookies[0];
3138 	cur_offset = 0;
3139 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3140 	if (dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) {
3141 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3142 	}
3143 
3144 	/* loop though all the cookies we got back from get_sgl() */
3145 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3146 		/*
3147 		 * If we're using the copy buffer, check this cookie and setup
3148 		 * its associated copy buffer state. If this cookie uses the
3149 		 * copy buffer, make sure we sync this window during dma_sync.
3150 		 */
3151 		if (dma->dp_copybuf_size > 0) {
3152 			rootnex_setup_cookie(&dmareq->dmar_object, dma, cookie,
3153 			    cur_offset, &copybuf_used, &cur_pp);
3154 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3155 				window->wd_dosync = B_TRUE;
3156 			}
3157 		}
3158 
3159 		/*
3160 		 * save away the cookie size, since it could be modified in
3161 		 * the windowing code.
3162 		 */
3163 		dmac_size = cookie->dmac_size;
3164 
3165 		/* if we went over max copybuf size */
3166 		if (dma->dp_copybuf_size &&
3167 		    (copybuf_used > dma->dp_copybuf_size)) {
3168 			partial = B_TRUE;
3169 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3170 			    cookie, cur_offset, &copybuf_used);
3171 			if (e != DDI_SUCCESS) {
3172 				rootnex_teardown_copybuf(dma);
3173 				rootnex_teardown_windows(dma);
3174 				return (e);
3175 			}
3176 
3177 			/*
3178 			 * if the coookie uses the copy buffer, make sure the
3179 			 * new window we just moved to is set to sync.
3180 			 */
3181 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3182 				window->wd_dosync = B_TRUE;
3183 			}
3184 			DTRACE_PROBE1(rootnex__copybuf__window, dev_info_t *,
3185 			    dma->dp_dip);
3186 
3187 		/* if the cookie cnt == max sgllen, move to the next window */
3188 		} else if (window->wd_cookie_cnt >= attr->dma_attr_sgllen) {
3189 			partial = B_TRUE;
3190 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3191 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3192 			    cookie, attr, cur_offset);
3193 			if (e != DDI_SUCCESS) {
3194 				rootnex_teardown_copybuf(dma);
3195 				rootnex_teardown_windows(dma);
3196 				return (e);
3197 			}
3198 
3199 			/*
3200 			 * if the coookie uses the copy buffer, make sure the
3201 			 * new window we just moved to is set to sync.
3202 			 */
3203 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3204 				window->wd_dosync = B_TRUE;
3205 			}
3206 			DTRACE_PROBE1(rootnex__sgllen__window, dev_info_t *,
3207 			    dma->dp_dip);
3208 
3209 		/* else if we will be over maxxfer */
3210 		} else if ((window->wd_size + dmac_size) >
3211 		    dma->dp_maxxfer) {
3212 			partial = B_TRUE;
3213 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3214 			    cookie);
3215 			if (e != DDI_SUCCESS) {
3216 				rootnex_teardown_copybuf(dma);
3217 				rootnex_teardown_windows(dma);
3218 				return (e);
3219 			}
3220 
3221 			/*
3222 			 * if the coookie uses the copy buffer, make sure the
3223 			 * new window we just moved to is set to sync.
3224 			 */
3225 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3226 				window->wd_dosync = B_TRUE;
3227 			}
3228 			DTRACE_PROBE1(rootnex__maxxfer__window, dev_info_t *,
3229 			    dma->dp_dip);
3230 
3231 		/* else this cookie fits in the current window */
3232 		} else {
3233 			window->wd_cookie_cnt++;
3234 			window->wd_size += dmac_size;
3235 		}
3236 
3237 		/* track our offset into the buffer, go to the next cookie */
3238 		ASSERT(dmac_size <= dma->dp_dma.dmao_size);
3239 		ASSERT(cookie->dmac_size <= dmac_size);
3240 		cur_offset += dmac_size;
3241 		cookie++;
3242 	}
3243 
3244 	/* if we ended up with a zero sized window in the end, clean it up */
3245 	if (window->wd_size == 0) {
3246 		hp->dmai_nwin--;
3247 		window--;
3248 	}
3249 
3250 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3251 
3252 	if (!partial) {
3253 		return (DDI_DMA_MAPPED);
3254 	}
3255 
3256 	ASSERT(dma->dp_partial_required);
3257 	return (DDI_DMA_PARTIAL_MAP);
3258 }
3259 
3260 
3261 /*
3262  * rootnex_setup_copybuf()
3263  *    Called in bind slowpath. Figures out if we're going to use the copy
3264  *    buffer, and if we do, sets up the basic state to handle it.
3265  */
3266 static int
3267 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3268     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3269 {
3270 	rootnex_sglinfo_t *sinfo;
3271 	ddi_dma_attr_t lattr;
3272 	size_t max_copybuf;
3273 	int cansleep;
3274 	int e;
3275 #if !defined(__amd64)
3276 	int vmflag;
3277 #endif
3278 
3279 
3280 	sinfo = &dma->dp_sglinfo;
3281 
3282 	/* read this first so it's consistent through the routine  */
3283 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3284 
3285 	/* We need to call into the rootnex on ddi_dma_sync() */
3286 	hp->dmai_rflags &= ~DMP_NOSYNC;
3287 
3288 	/* make sure the copybuf size <= the max size */
3289 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3290 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3291 
3292 #if !defined(__amd64)
3293 	/*
3294 	 * if we don't have kva space to copy to/from, allocate the KVA space
3295 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3296 	 * the 64-bit kernel.
3297 	 */
3298 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3299 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3300 
3301 		/* convert the sleep flags */
3302 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3303 			vmflag = VM_SLEEP;
3304 		} else {
3305 			vmflag = VM_NOSLEEP;
3306 		}
3307 
3308 		/* allocate Kernel VA space that we can bcopy to/from */
3309 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3310 		    vmflag);
3311 		if (dma->dp_kva == NULL) {
3312 			return (DDI_DMA_NORESOURCES);
3313 		}
3314 	}
3315 #endif
3316 
3317 	/* convert the sleep flags */
3318 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3319 		cansleep = 1;
3320 	} else {
3321 		cansleep = 0;
3322 	}
3323 
3324 	/*
3325 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3326 	 * engine limits, so we can't use kmem_alloc... We don't need
3327 	 * contiguous memory (sgllen) since we will be forcing windows on
3328 	 * sgllen anyway.
3329 	 */
3330 	lattr = *attr;
3331 	lattr.dma_attr_align = MMU_PAGESIZE;
3332 	/*
3333 	 * this should be < 0 to indicate no limit, but due to a bug in
3334 	 * the rootnex, we'll set it to the maximum positive int.
3335 	 */
3336 	lattr.dma_attr_sgllen = 0x7fffffff;
3337 	/*
3338 	 * if we're using the copy buffer because of seg, use that for our
3339 	 * upper address limit.
3340 	 */
3341 	if (sinfo->si_bounce_on_seg) {
3342 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3343 	}
3344 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3345 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3346 	if (e != DDI_SUCCESS) {
3347 #if !defined(__amd64)
3348 		if (dma->dp_kva != NULL) {
3349 			vmem_free(heap_arena, dma->dp_kva,
3350 			    dma->dp_copybuf_size);
3351 		}
3352 #endif
3353 		return (DDI_DMA_NORESOURCES);
3354 	}
3355 
3356 	DTRACE_PROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3357 	    size_t, dma->dp_copybuf_size);
3358 
3359 	return (DDI_SUCCESS);
3360 }
3361 
3362 
3363 /*
3364  * rootnex_setup_windows()
3365  *    Called in bind slowpath to setup the window state. We always have windows
3366  *    in the slowpath. Even if the window count = 1.
3367  */
3368 static int
3369 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3370     ddi_dma_attr_t *attr, int kmflag)
3371 {
3372 	rootnex_window_t *windowp;
3373 	rootnex_sglinfo_t *sinfo;
3374 	size_t copy_state_size;
3375 	size_t win_state_size;
3376 	size_t state_available;
3377 	size_t space_needed;
3378 	uint_t copybuf_win;
3379 	uint_t maxxfer_win;
3380 	size_t space_used;
3381 	uint_t sglwin;
3382 
3383 
3384 	sinfo = &dma->dp_sglinfo;
3385 
3386 	dma->dp_current_win = 0;
3387 	hp->dmai_nwin = 0;
3388 
3389 	/* If we don't need to do a partial, we only have one window */
3390 	if (!dma->dp_partial_required) {
3391 		dma->dp_max_win = 1;
3392 
3393 	/*
3394 	 * we need multiple windows, need to figure out the worse case number
3395 	 * of windows.
3396 	 */
3397 	} else {
3398 		/*
3399 		 * if we need windows because we need more copy buffer that
3400 		 * we allow, the worse case number of windows we could need
3401 		 * here would be (copybuf space required / copybuf space that
3402 		 * we have) plus one for remainder, and plus 2 to handle the
3403 		 * extra pages on the trim for the first and last pages of the
3404 		 * buffer (a page is the minimum window size so under the right
3405 		 * attr settings, you could have a window for each page).
3406 		 * The last page will only be hit here if the size is not a
3407 		 * multiple of the granularity (which theoretically shouldn't
3408 		 * be the case but never has been enforced, so we could have
3409 		 * broken things without it).
3410 		 */
3411 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3412 			ASSERT(dma->dp_copybuf_size > 0);
3413 			copybuf_win = (sinfo->si_copybuf_req /
3414 			    dma->dp_copybuf_size) + 1 + 2;
3415 		} else {
3416 			copybuf_win = 0;
3417 		}
3418 
3419 		/*
3420 		 * if we need windows because we have more cookies than the H/W
3421 		 * can handle, the number of windows we would need here would
3422 		 * be (cookie count / cookies count H/W supports minus 1[for
3423 		 * trim]) plus one for remainder.
3424 		 */
3425 		if (attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3426 			sglwin = (sinfo->si_sgl_size /
3427 			    (attr->dma_attr_sgllen - 1)) + 1;
3428 		} else {
3429 			sglwin = 0;
3430 		}
3431 
3432 		/*
3433 		 * if we need windows because we're binding more memory than the
3434 		 * H/W can transfer at once, the number of windows we would need
3435 		 * here would be (xfer count / max xfer H/W supports) plus one
3436 		 * for remainder, and plus 2 to handle the extra pages on the
3437 		 * trim (see above comment about trim)
3438 		 */
3439 		if (dma->dp_dma.dmao_size > dma->dp_maxxfer) {
3440 			maxxfer_win = (dma->dp_dma.dmao_size /
3441 			    dma->dp_maxxfer) + 1 + 2;
3442 		} else {
3443 			maxxfer_win = 0;
3444 		}
3445 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3446 		ASSERT(dma->dp_max_win > 0);
3447 	}
3448 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3449 
3450 	/*
3451 	 * Get space for window and potential copy buffer state. Before we
3452 	 * go and allocate memory, see if we can get away with using what's
3453 	 * left in the pre-allocted state or the dynamically allocated sgl.
3454 	 */
3455 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3456 	    sizeof (ddi_dma_cookie_t));
3457 
3458 	/* if we dynamically allocated space for the cookies */
3459 	if (dma->dp_need_to_free_cookie) {
3460 		/* if we have more space in the pre-allocted buffer, use it */
3461 		ASSERT(space_used <= dma->dp_cookie_size);
3462 		if ((dma->dp_cookie_size - space_used) <=
3463 		    rootnex_state->r_prealloc_size) {
3464 			state_available = rootnex_state->r_prealloc_size;
3465 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3466 
3467 		/*
3468 		 * else, we have more free space in the dynamically allocated
3469 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3470 		 * didn't need a lot of cookies.
3471 		 */
3472 		} else {
3473 			state_available = dma->dp_cookie_size - space_used;
3474 			windowp = (rootnex_window_t *)
3475 			    &dma->dp_cookies[sinfo->si_sgl_size];
3476 		}
3477 
3478 	/* we used the pre-alloced buffer */
3479 	} else {
3480 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3481 		state_available = rootnex_state->r_prealloc_size - space_used;
3482 		windowp = (rootnex_window_t *)
3483 		    &dma->dp_cookies[sinfo->si_sgl_size];
3484 	}
3485 
3486 	/*
3487 	 * figure out how much state we need to track the copy buffer. Add an
3488 	 * addition 8 bytes for pointer alignemnt later.
3489 	 */
3490 	if (dma->dp_copybuf_size > 0) {
3491 		copy_state_size = sinfo->si_max_pages *
3492 		    sizeof (rootnex_pgmap_t);
3493 	} else {
3494 		copy_state_size = 0;
3495 	}
3496 	/* add an additional 8 bytes for pointer alignment */
3497 	space_needed = win_state_size + copy_state_size + 0x8;
3498 
3499 	/* if we have enough space already, use it */
3500 	if (state_available >= space_needed) {
3501 		dma->dp_window = windowp;
3502 		dma->dp_need_to_free_window = B_FALSE;
3503 
3504 	/* not enough space, need to allocate more. */
3505 	} else {
3506 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3507 		if (dma->dp_window == NULL) {
3508 			return (DDI_DMA_NORESOURCES);
3509 		}
3510 		dma->dp_need_to_free_window = B_TRUE;
3511 		dma->dp_window_size = space_needed;
3512 		DTRACE_PROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3513 		    dma->dp_dip, size_t, space_needed);
3514 	}
3515 
3516 	/*
3517 	 * we allocate copy buffer state and window state at the same time.
3518 	 * setup our copy buffer state pointers. Make sure it's aligned.
3519 	 */
3520 	if (dma->dp_copybuf_size > 0) {
3521 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3522 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3523 
3524 #if !defined(__amd64)
3525 		/*
3526 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3527 		 * false/NULL. Should be quicker to bzero vs loop and set.
3528 		 */
3529 		bzero(dma->dp_pgmap, copy_state_size);
3530 #endif
3531 	} else {
3532 		dma->dp_pgmap = NULL;
3533 	}
3534 
3535 	return (DDI_SUCCESS);
3536 }
3537 
3538 
3539 /*
3540  * rootnex_teardown_copybuf()
3541  *    cleans up after rootnex_setup_copybuf()
3542  */
3543 static void
3544 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3545 {
3546 #if !defined(__amd64)
3547 	int i;
3548 
3549 	/*
3550 	 * if we allocated kernel heap VMEM space, go through all the pages and
3551 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3552 	 * arena. Then free the VMEM space.
3553 	 */
3554 	if (dma->dp_kva != NULL) {
3555 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3556 			if (dma->dp_pgmap[i].pm_mapped) {
3557 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3558 				    MMU_PAGESIZE, HAT_UNLOAD);
3559 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3560 			}
3561 		}
3562 
3563 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3564 	}
3565 
3566 #endif
3567 
3568 	/* if we allocated a copy buffer, free it */
3569 	if (dma->dp_cbaddr != NULL) {
3570 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3571 	}
3572 }
3573 
3574 
3575 /*
3576  * rootnex_teardown_windows()
3577  *    cleans up after rootnex_setup_windows()
3578  */
3579 static void
3580 rootnex_teardown_windows(rootnex_dma_t *dma)
3581 {
3582 	/*
3583 	 * if we had to allocate window state on the last bind (because we
3584 	 * didn't have enough pre-allocated space in the handle), free it.
3585 	 */
3586 	if (dma->dp_need_to_free_window) {
3587 		kmem_free(dma->dp_window, dma->dp_window_size);
3588 	}
3589 }
3590 
3591 
3592 /*
3593  * rootnex_init_win()
3594  *    Called in bind slow path during creation of a new window. Initializes
3595  *    window state to default values.
3596  */
3597 /*ARGSUSED*/
3598 static void
3599 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3600     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3601 {
3602 	hp->dmai_nwin++;
3603 	window->wd_dosync = B_FALSE;
3604 	window->wd_offset = cur_offset;
3605 	window->wd_size = 0;
3606 	window->wd_first_cookie = cookie;
3607 	window->wd_cookie_cnt = 0;
3608 	window->wd_trim.tr_trim_first = B_FALSE;
3609 	window->wd_trim.tr_trim_last = B_FALSE;
3610 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3611 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3612 #if !defined(__amd64)
3613 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3614 #endif
3615 }
3616 
3617 
3618 /*
3619  * rootnex_setup_cookie()
3620  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3621  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3622  *    out if it uses the copy buffer, and if it does, save away everything we'll
3623  *    need during sync.
3624  */
3625 static void
3626 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3627     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3628     page_t **cur_pp)
3629 {
3630 	boolean_t copybuf_sz_power_2;
3631 	rootnex_sglinfo_t *sinfo;
3632 	paddr_t paddr;
3633 	uint_t pidx;
3634 	uint_t pcnt;
3635 	off_t poff;
3636 #if defined(__amd64)
3637 	pfn_t pfn;
3638 #else
3639 	page_t **pplist;
3640 #endif
3641 
3642 	sinfo = &dma->dp_sglinfo;
3643 
3644 	/*
3645 	 * Calculate the page index relative to the start of the buffer. The
3646 	 * index to the current page for our buffer is the offset into the
3647 	 * first page of the buffer plus our current offset into the buffer
3648 	 * itself, shifted of course...
3649 	 */
3650 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3651 	ASSERT(pidx < sinfo->si_max_pages);
3652 
3653 	/* if this cookie uses the copy buffer */
3654 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3655 		/*
3656 		 * NOTE: we know that since this cookie uses the copy buffer, it
3657 		 * is <= MMU_PAGESIZE.
3658 		 */
3659 
3660 		/*
3661 		 * get the offset into the page. For the 64-bit kernel, get the
3662 		 * pfn which we'll use with seg kpm.
3663 		 */
3664 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3665 #if defined(__amd64)
3666 		/* mfn_to_pfn() is a NOP on i86pc */
3667 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3668 #endif /* __amd64 */
3669 
3670 		/* figure out if the copybuf size is a power of 2 */
3671 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3672 			copybuf_sz_power_2 = B_FALSE;
3673 		} else {
3674 			copybuf_sz_power_2 = B_TRUE;
3675 		}
3676 
3677 		/* This page uses the copy buffer */
3678 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3679 
3680 		/*
3681 		 * save the copy buffer KVA that we'll use with this page.
3682 		 * if we still fit within the copybuf, it's a simple add.
3683 		 * otherwise, we need to wrap over using & or % accordingly.
3684 		 */
3685 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3686 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3687 			    *copybuf_used;
3688 		} else {
3689 			if (copybuf_sz_power_2) {
3690 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3691 				    (uintptr_t)dma->dp_cbaddr +
3692 				    (*copybuf_used &
3693 				    (dma->dp_copybuf_size - 1)));
3694 			} else {
3695 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3696 				    (uintptr_t)dma->dp_cbaddr +
3697 				    (*copybuf_used % dma->dp_copybuf_size));
3698 			}
3699 		}
3700 
3701 		/*
3702 		 * over write the cookie physical address with the address of
3703 		 * the physical address of the copy buffer page that we will
3704 		 * use.
3705 		 */
3706 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3707 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3708 
3709 #ifdef __xpv
3710 		/*
3711 		 * If we're dom0, we're using a real device so we need to load
3712 		 * the cookies with MAs instead of PAs.
3713 		 */
3714 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
3715 #else
3716 		cookie->dmac_laddress = paddr;
3717 #endif
3718 
3719 		/* if we have a kernel VA, it's easy, just save that address */
3720 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3721 		    (sinfo->si_asp == &kas)) {
3722 			/*
3723 			 * save away the page aligned virtual address of the
3724 			 * driver buffer. Offsets are handled in the sync code.
3725 			 */
3726 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3727 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3728 			    & MMU_PAGEMASK);
3729 #if !defined(__amd64)
3730 			/*
3731 			 * we didn't need to, and will never need to map this
3732 			 * page.
3733 			 */
3734 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3735 #endif
3736 
3737 		/* we don't have a kernel VA. We need one for the bcopy. */
3738 		} else {
3739 #if defined(__amd64)
3740 			/*
3741 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3742 			 * get a Kernel VA for the corresponding pfn.
3743 			 */
3744 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3745 #else
3746 			/*
3747 			 * for the 32-bit kernel, this is a pain. First we'll
3748 			 * save away the page_t or user VA for this page. This
3749 			 * is needed in rootnex_dma_win() when we switch to a
3750 			 * new window which requires us to re-map the copy
3751 			 * buffer.
3752 			 */
3753 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3754 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3755 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3756 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3757 			} else if (pplist != NULL) {
3758 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3759 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3760 			} else {
3761 				dma->dp_pgmap[pidx].pm_pp = NULL;
3762 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3763 				    (((uintptr_t)
3764 				    dmar_object->dmao_obj.virt_obj.v_addr +
3765 				    cur_offset) & MMU_PAGEMASK);
3766 			}
3767 
3768 			/*
3769 			 * save away the page aligned virtual address which was
3770 			 * allocated from the kernel heap arena (taking into
3771 			 * account if we need more copy buffer than we alloced
3772 			 * and use multiple windows to handle this, i.e. &,%).
3773 			 * NOTE: there isn't and physical memory backing up this
3774 			 * virtual address space currently.
3775 			 */
3776 			if ((*copybuf_used + MMU_PAGESIZE) <=
3777 			    dma->dp_copybuf_size) {
3778 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3779 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3780 				    MMU_PAGEMASK);
3781 			} else {
3782 				if (copybuf_sz_power_2) {
3783 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3784 					    (((uintptr_t)dma->dp_kva +
3785 					    (*copybuf_used &
3786 					    (dma->dp_copybuf_size - 1))) &
3787 					    MMU_PAGEMASK);
3788 				} else {
3789 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3790 					    (((uintptr_t)dma->dp_kva +
3791 					    (*copybuf_used %
3792 					    dma->dp_copybuf_size)) &
3793 					    MMU_PAGEMASK);
3794 				}
3795 			}
3796 
3797 			/*
3798 			 * if we haven't used up the available copy buffer yet,
3799 			 * map the kva to the physical page.
3800 			 */
3801 			if (!dma->dp_cb_remaping && ((*copybuf_used +
3802 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
3803 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
3804 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
3805 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
3806 					    dma->dp_pgmap[pidx].pm_kaddr);
3807 				} else {
3808 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
3809 					    sinfo->si_asp,
3810 					    dma->dp_pgmap[pidx].pm_kaddr);
3811 				}
3812 
3813 			/*
3814 			 * we've used up the available copy buffer, this page
3815 			 * will have to be mapped during rootnex_dma_win() when
3816 			 * we switch to a new window which requires a re-map
3817 			 * the copy buffer. (32-bit kernel only)
3818 			 */
3819 			} else {
3820 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3821 			}
3822 #endif
3823 			/* go to the next page_t */
3824 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3825 				*cur_pp = (*cur_pp)->p_next;
3826 			}
3827 		}
3828 
3829 		/* add to the copy buffer count */
3830 		*copybuf_used += MMU_PAGESIZE;
3831 
3832 	/*
3833 	 * This cookie doesn't use the copy buffer. Walk through the pages this
3834 	 * cookie occupies to reflect this.
3835 	 */
3836 	} else {
3837 		/*
3838 		 * figure out how many pages the cookie occupies. We need to
3839 		 * use the original page offset of the buffer and the cookies
3840 		 * offset in the buffer to do this.
3841 		 */
3842 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
3843 		pcnt = mmu_btopr(cookie->dmac_size + poff);
3844 
3845 		while (pcnt > 0) {
3846 #if !defined(__amd64)
3847 			/*
3848 			 * the 32-bit kernel doesn't have seg kpm, so we need
3849 			 * to map in the driver buffer (if it didn't come down
3850 			 * with a kernel VA) on the fly. Since this page doesn't
3851 			 * use the copy buffer, it's not, or will it ever, have
3852 			 * to be mapped in.
3853 			 */
3854 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3855 #endif
3856 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
3857 
3858 			/*
3859 			 * we need to update pidx and cur_pp or we'll loose
3860 			 * track of where we are.
3861 			 */
3862 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3863 				*cur_pp = (*cur_pp)->p_next;
3864 			}
3865 			pidx++;
3866 			pcnt--;
3867 		}
3868 	}
3869 }
3870 
3871 
3872 /*
3873  * rootnex_sgllen_window_boundary()
3874  *    Called in the bind slow path when the next cookie causes us to exceed (in
3875  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
3876  *    length supported by the DMA H/W.
3877  */
3878 static int
3879 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3880     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
3881     off_t cur_offset)
3882 {
3883 	off_t new_offset;
3884 	size_t trim_sz;
3885 	off_t coffset;
3886 
3887 
3888 	/*
3889 	 * if we know we'll never have to trim, it's pretty easy. Just move to
3890 	 * the next window and init it. We're done.
3891 	 */
3892 	if (!dma->dp_trim_required) {
3893 		(*windowp)++;
3894 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3895 		(*windowp)->wd_cookie_cnt++;
3896 		(*windowp)->wd_size = cookie->dmac_size;
3897 		return (DDI_SUCCESS);
3898 	}
3899 
3900 	/* figure out how much we need to trim from the window */
3901 	ASSERT(attr->dma_attr_granular != 0);
3902 	if (dma->dp_granularity_power_2) {
3903 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
3904 	} else {
3905 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
3906 	}
3907 
3908 	/* The window's a whole multiple of granularity. We're done */
3909 	if (trim_sz == 0) {
3910 		(*windowp)++;
3911 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3912 		(*windowp)->wd_cookie_cnt++;
3913 		(*windowp)->wd_size = cookie->dmac_size;
3914 		return (DDI_SUCCESS);
3915 	}
3916 
3917 	/*
3918 	 * The window's not a whole multiple of granularity, since we know this
3919 	 * is due to the sgllen, we need to go back to the last cookie and trim
3920 	 * that one, add the left over part of the old cookie into the new
3921 	 * window, and then add in the new cookie into the new window.
3922 	 */
3923 
3924 	/*
3925 	 * make sure the driver isn't making us do something bad... Trimming and
3926 	 * sgllen == 1 don't go together.
3927 	 */
3928 	if (attr->dma_attr_sgllen == 1) {
3929 		return (DDI_DMA_NOMAPPING);
3930 	}
3931 
3932 	/*
3933 	 * first, setup the current window to account for the trim. Need to go
3934 	 * back to the last cookie for this.
3935 	 */
3936 	cookie--;
3937 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3938 	(*windowp)->wd_trim.tr_last_cookie = cookie;
3939 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3940 	ASSERT(cookie->dmac_size > trim_sz);
3941 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3942 	(*windowp)->wd_size -= trim_sz;
3943 
3944 	/* save the buffer offsets for the next window */
3945 	coffset = cookie->dmac_size - trim_sz;
3946 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3947 
3948 	/*
3949 	 * set this now in case this is the first window. all other cases are
3950 	 * set in dma_win()
3951 	 */
3952 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
3953 
3954 	/*
3955 	 * initialize the next window using what's left over in the previous
3956 	 * cookie.
3957 	 */
3958 	(*windowp)++;
3959 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3960 	(*windowp)->wd_cookie_cnt++;
3961 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3962 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
3963 	(*windowp)->wd_trim.tr_first_size = trim_sz;
3964 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3965 		(*windowp)->wd_dosync = B_TRUE;
3966 	}
3967 
3968 	/*
3969 	 * now go back to the current cookie and add it to the new window. set
3970 	 * the new window size to the what was left over from the previous
3971 	 * cookie and what's in the current cookie.
3972 	 */
3973 	cookie++;
3974 	(*windowp)->wd_cookie_cnt++;
3975 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
3976 
3977 	/*
3978 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
3979 	 * a max size of maxxfer). Handle that case.
3980 	 */
3981 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
3982 		/*
3983 		 * maxxfer is already a whole multiple of granularity, and this
3984 		 * trim will be <= the previous trim (since a cookie can't be
3985 		 * larger than maxxfer). Make things simple here.
3986 		 */
3987 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
3988 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3989 		(*windowp)->wd_trim.tr_last_cookie = cookie;
3990 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3991 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3992 		(*windowp)->wd_size -= trim_sz;
3993 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
3994 
3995 		/* save the buffer offsets for the next window */
3996 		coffset = cookie->dmac_size - trim_sz;
3997 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3998 
3999 		/* setup the next window */
4000 		(*windowp)++;
4001 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4002 		(*windowp)->wd_cookie_cnt++;
4003 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4004 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4005 		    coffset;
4006 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4007 	}
4008 
4009 	return (DDI_SUCCESS);
4010 }
4011 
4012 
4013 /*
4014  * rootnex_copybuf_window_boundary()
4015  *    Called in bind slowpath when we get to a window boundary because we used
4016  *    up all the copy buffer that we have.
4017  */
4018 static int
4019 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4020     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
4021     size_t *copybuf_used)
4022 {
4023 	rootnex_sglinfo_t *sinfo;
4024 	off_t new_offset;
4025 	size_t trim_sz;
4026 	paddr_t paddr;
4027 	off_t coffset;
4028 	uint_t pidx;
4029 	off_t poff;
4030 
4031 
4032 	sinfo = &dma->dp_sglinfo;
4033 
4034 	/*
4035 	 * the copy buffer should be a whole multiple of page size. We know that
4036 	 * this cookie is <= MMU_PAGESIZE.
4037 	 */
4038 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
4039 
4040 	/*
4041 	 * from now on, all new windows in this bind need to be re-mapped during
4042 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
4043 	 * space...
4044 	 */
4045 #if !defined(__amd64)
4046 	dma->dp_cb_remaping = B_TRUE;
4047 #endif
4048 
4049 	/* reset copybuf used */
4050 	*copybuf_used = 0;
4051 
4052 	/*
4053 	 * if we don't have to trim (since granularity is set to 1), go to the
4054 	 * next window and add the current cookie to it. We know the current
4055 	 * cookie uses the copy buffer since we're in this code path.
4056 	 */
4057 	if (!dma->dp_trim_required) {
4058 		(*windowp)++;
4059 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4060 
4061 		/* Add this cookie to the new window */
4062 		(*windowp)->wd_cookie_cnt++;
4063 		(*windowp)->wd_size += cookie->dmac_size;
4064 		*copybuf_used += MMU_PAGESIZE;
4065 		return (DDI_SUCCESS);
4066 	}
4067 
4068 	/*
4069 	 * *** may need to trim, figure it out.
4070 	 */
4071 
4072 	/* figure out how much we need to trim from the window */
4073 	if (dma->dp_granularity_power_2) {
4074 		trim_sz = (*windowp)->wd_size &
4075 		    (hp->dmai_attr.dma_attr_granular - 1);
4076 	} else {
4077 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
4078 	}
4079 
4080 	/*
4081 	 * if the window's a whole multiple of granularity, go to the next
4082 	 * window, init it, then add in the current cookie. We know the current
4083 	 * cookie uses the copy buffer since we're in this code path.
4084 	 */
4085 	if (trim_sz == 0) {
4086 		(*windowp)++;
4087 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4088 
4089 		/* Add this cookie to the new window */
4090 		(*windowp)->wd_cookie_cnt++;
4091 		(*windowp)->wd_size += cookie->dmac_size;
4092 		*copybuf_used += MMU_PAGESIZE;
4093 		return (DDI_SUCCESS);
4094 	}
4095 
4096 	/*
4097 	 * *** We figured it out, we definitly need to trim
4098 	 */
4099 
4100 	/*
4101 	 * make sure the driver isn't making us do something bad...
4102 	 * Trimming and sgllen == 1 don't go together.
4103 	 */
4104 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4105 		return (DDI_DMA_NOMAPPING);
4106 	}
4107 
4108 	/*
4109 	 * first, setup the current window to account for the trim. Need to go
4110 	 * back to the last cookie for this. Some of the last cookie will be in
4111 	 * the current window, and some of the last cookie will be in the new
4112 	 * window. All of the current cookie will be in the new window.
4113 	 */
4114 	cookie--;
4115 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4116 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4117 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4118 	ASSERT(cookie->dmac_size > trim_sz);
4119 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4120 	(*windowp)->wd_size -= trim_sz;
4121 
4122 	/*
4123 	 * we're trimming the last cookie (not the current cookie). So that
4124 	 * last cookie may have or may not have been using the copy buffer (
4125 	 * we know the cookie passed in uses the copy buffer since we're in
4126 	 * this code path).
4127 	 *
4128 	 * If the last cookie doesn't use the copy buffer, nothing special to
4129 	 * do. However, if it does uses the copy buffer, it will be both the
4130 	 * last page in the current window and the first page in the next
4131 	 * window. Since we are reusing the copy buffer (and KVA space on the
4132 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4133 	 * current window, and the start of the copy buffer in the next window.
4134 	 * Track that info... The cookie physical address was already set to
4135 	 * the copy buffer physical address in setup_cookie..
4136 	 */
4137 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4138 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4139 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4140 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4141 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4142 		(*windowp)->wd_trim.tr_last_cbaddr =
4143 		    dma->dp_pgmap[pidx].pm_cbaddr;
4144 #if !defined(__amd64)
4145 		(*windowp)->wd_trim.tr_last_kaddr =
4146 		    dma->dp_pgmap[pidx].pm_kaddr;
4147 #endif
4148 	}
4149 
4150 	/* save the buffer offsets for the next window */
4151 	coffset = cookie->dmac_size - trim_sz;
4152 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4153 
4154 	/*
4155 	 * set this now in case this is the first window. all other cases are
4156 	 * set in dma_win()
4157 	 */
4158 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4159 
4160 	/*
4161 	 * initialize the next window using what's left over in the previous
4162 	 * cookie.
4163 	 */
4164 	(*windowp)++;
4165 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4166 	(*windowp)->wd_cookie_cnt++;
4167 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4168 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4169 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4170 
4171 	/*
4172 	 * again, we're tracking if the last cookie uses the copy buffer.
4173 	 * read the comment above for more info on why we need to track
4174 	 * additional state.
4175 	 *
4176 	 * For the first cookie in the new window, we need reset the physical
4177 	 * address to DMA into to the start of the copy buffer plus any
4178 	 * initial page offset which may be present.
4179 	 */
4180 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4181 		(*windowp)->wd_dosync = B_TRUE;
4182 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4183 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4184 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4185 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4186 
4187 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4188 		    poff;
4189 #ifdef __xpv
4190 		/*
4191 		 * If we're dom0, we're using a real device so we need to load
4192 		 * the cookies with MAs instead of PAs.
4193 		 */
4194 		(*windowp)->wd_trim.tr_first_paddr =
4195 		    ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
4196 #else
4197 		(*windowp)->wd_trim.tr_first_paddr = paddr;
4198 #endif
4199 
4200 #if !defined(__amd64)
4201 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4202 #endif
4203 		/* account for the cookie copybuf usage in the new window */
4204 		*copybuf_used += MMU_PAGESIZE;
4205 
4206 		/*
4207 		 * every piece of code has to have a hack, and here is this
4208 		 * ones :-)
4209 		 *
4210 		 * There is a complex interaction between setup_cookie and the
4211 		 * copybuf window boundary. The complexity had to be in either
4212 		 * the maxxfer window, or the copybuf window, and I chose the
4213 		 * copybuf code.
4214 		 *
4215 		 * So in this code path, we have taken the last cookie,
4216 		 * virtually broken it in half due to the trim, and it happens
4217 		 * to use the copybuf which further complicates life. At the
4218 		 * same time, we have already setup the current cookie, which
4219 		 * is now wrong. More background info: the current cookie uses
4220 		 * the copybuf, so it is only a page long max. So we need to
4221 		 * fix the current cookies copy buffer address, physical
4222 		 * address, and kva for the 32-bit kernel. We due this by
4223 		 * bumping them by page size (of course, we can't due this on
4224 		 * the physical address since the copy buffer may not be
4225 		 * physically contiguous).
4226 		 */
4227 		cookie++;
4228 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4229 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4230 
4231 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4232 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4233 #ifdef __xpv
4234 		/*
4235 		 * If we're dom0, we're using a real device so we need to load
4236 		 * the cookies with MAs instead of PAs.
4237 		 */
4238 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
4239 #else
4240 		cookie->dmac_laddress = paddr;
4241 #endif
4242 
4243 #if !defined(__amd64)
4244 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4245 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4246 #endif
4247 	} else {
4248 		/* go back to the current cookie */
4249 		cookie++;
4250 	}
4251 
4252 	/*
4253 	 * add the current cookie to the new window. set the new window size to
4254 	 * the what was left over from the previous cookie and what's in the
4255 	 * current cookie.
4256 	 */
4257 	(*windowp)->wd_cookie_cnt++;
4258 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4259 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4260 
4261 	/*
4262 	 * we know that the cookie passed in always uses the copy buffer. We
4263 	 * wouldn't be here if it didn't.
4264 	 */
4265 	*copybuf_used += MMU_PAGESIZE;
4266 
4267 	return (DDI_SUCCESS);
4268 }
4269 
4270 
4271 /*
4272  * rootnex_maxxfer_window_boundary()
4273  *    Called in bind slowpath when we get to a window boundary because we will
4274  *    go over maxxfer.
4275  */
4276 static int
4277 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4278     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4279 {
4280 	size_t dmac_size;
4281 	off_t new_offset;
4282 	size_t trim_sz;
4283 	off_t coffset;
4284 
4285 
4286 	/*
4287 	 * calculate how much we have to trim off of the current cookie to equal
4288 	 * maxxfer. We don't have to account for granularity here since our
4289 	 * maxxfer already takes that into account.
4290 	 */
4291 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4292 	ASSERT(trim_sz <= cookie->dmac_size);
4293 	ASSERT(trim_sz <= dma->dp_maxxfer);
4294 
4295 	/* save cookie size since we need it later and we might change it */
4296 	dmac_size = cookie->dmac_size;
4297 
4298 	/*
4299 	 * if we're not trimming the entire cookie, setup the current window to
4300 	 * account for the trim.
4301 	 */
4302 	if (trim_sz < cookie->dmac_size) {
4303 		(*windowp)->wd_cookie_cnt++;
4304 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4305 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4306 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4307 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4308 		(*windowp)->wd_size = dma->dp_maxxfer;
4309 
4310 		/*
4311 		 * set the adjusted cookie size now in case this is the first
4312 		 * window. All other windows are taken care of in get win
4313 		 */
4314 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4315 	}
4316 
4317 	/*
4318 	 * coffset is the current offset within the cookie, new_offset is the
4319 	 * current offset with the entire buffer.
4320 	 */
4321 	coffset = dmac_size - trim_sz;
4322 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4323 
4324 	/* initialize the next window */
4325 	(*windowp)++;
4326 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4327 	(*windowp)->wd_cookie_cnt++;
4328 	(*windowp)->wd_size = trim_sz;
4329 	if (trim_sz < dmac_size) {
4330 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4331 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4332 		    coffset;
4333 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4334 	}
4335 
4336 	return (DDI_SUCCESS);
4337 }
4338 
4339 
4340 /*ARGSUSED*/
4341 static int
4342 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4343     off_t off, size_t len, uint_t cache_flags)
4344 {
4345 	rootnex_sglinfo_t *sinfo;
4346 	rootnex_pgmap_t *cbpage;
4347 	rootnex_window_t *win;
4348 	ddi_dma_impl_t *hp;
4349 	rootnex_dma_t *dma;
4350 	caddr_t fromaddr;
4351 	caddr_t toaddr;
4352 	uint_t psize;
4353 	off_t offset;
4354 	uint_t pidx;
4355 	size_t size;
4356 	off_t poff;
4357 	int e;
4358 
4359 
4360 	hp = (ddi_dma_impl_t *)handle;
4361 	dma = (rootnex_dma_t *)hp->dmai_private;
4362 	sinfo = &dma->dp_sglinfo;
4363 
4364 	/*
4365 	 * if we don't have any windows, we don't need to sync. A copybuf
4366 	 * will cause us to have at least one window.
4367 	 */
4368 	if (dma->dp_window == NULL) {
4369 		return (DDI_SUCCESS);
4370 	}
4371 
4372 	/* This window may not need to be sync'd */
4373 	win = &dma->dp_window[dma->dp_current_win];
4374 	if (!win->wd_dosync) {
4375 		return (DDI_SUCCESS);
4376 	}
4377 
4378 	/* handle off and len special cases */
4379 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4380 		offset = win->wd_offset;
4381 	} else {
4382 		offset = off;
4383 	}
4384 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4385 		size = win->wd_size;
4386 	} else {
4387 		size = len;
4388 	}
4389 
4390 	/* check the sync args to make sure they make a little sense */
4391 	if (rootnex_sync_check_parms) {
4392 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4393 		    cache_flags);
4394 		if (e != DDI_SUCCESS) {
4395 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4396 			return (DDI_FAILURE);
4397 		}
4398 	}
4399 
4400 	/*
4401 	 * special case the first page to handle the offset into the page. The
4402 	 * offset to the current page for our buffer is the offset into the
4403 	 * first page of the buffer plus our current offset into the buffer
4404 	 * itself, masked of course.
4405 	 */
4406 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4407 	psize = MIN((MMU_PAGESIZE - poff), size);
4408 
4409 	/* go through all the pages that we want to sync */
4410 	while (size > 0) {
4411 		/*
4412 		 * Calculate the page index relative to the start of the buffer.
4413 		 * The index to the current page for our buffer is the offset
4414 		 * into the first page of the buffer plus our current offset
4415 		 * into the buffer itself, shifted of course...
4416 		 */
4417 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4418 		ASSERT(pidx < sinfo->si_max_pages);
4419 
4420 		/*
4421 		 * if this page uses the copy buffer, we need to sync it,
4422 		 * otherwise, go on to the next page.
4423 		 */
4424 		cbpage = &dma->dp_pgmap[pidx];
4425 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4426 		    (cbpage->pm_uses_copybuf == B_FALSE));
4427 		if (cbpage->pm_uses_copybuf) {
4428 			/* cbaddr and kaddr should be page aligned */
4429 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4430 			    MMU_PAGEOFFSET) == 0);
4431 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4432 			    MMU_PAGEOFFSET) == 0);
4433 
4434 			/*
4435 			 * if we're copying for the device, we are going to
4436 			 * copy from the drivers buffer and to the rootnex
4437 			 * allocated copy buffer.
4438 			 */
4439 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4440 				fromaddr = cbpage->pm_kaddr + poff;
4441 				toaddr = cbpage->pm_cbaddr + poff;
4442 				DTRACE_PROBE2(rootnex__sync__dev,
4443 				    dev_info_t *, dma->dp_dip, size_t, psize);
4444 
4445 			/*
4446 			 * if we're copying for the cpu/kernel, we are going to
4447 			 * copy from the rootnex allocated copy buffer to the
4448 			 * drivers buffer.
4449 			 */
4450 			} else {
4451 				fromaddr = cbpage->pm_cbaddr + poff;
4452 				toaddr = cbpage->pm_kaddr + poff;
4453 				DTRACE_PROBE2(rootnex__sync__cpu,
4454 				    dev_info_t *, dma->dp_dip, size_t, psize);
4455 			}
4456 
4457 			bcopy(fromaddr, toaddr, psize);
4458 		}
4459 
4460 		/*
4461 		 * decrement size until we're done, update our offset into the
4462 		 * buffer, and get the next page size.
4463 		 */
4464 		size -= psize;
4465 		offset += psize;
4466 		psize = MIN(MMU_PAGESIZE, size);
4467 
4468 		/* page offset is zero for the rest of this loop */
4469 		poff = 0;
4470 	}
4471 
4472 	return (DDI_SUCCESS);
4473 }
4474 
4475 /*
4476  * rootnex_dma_sync()
4477  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4478  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4479  *    is set, ddi_dma_sync() returns immediately passing back success.
4480  */
4481 /*ARGSUSED*/
4482 static int
4483 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4484     off_t off, size_t len, uint_t cache_flags)
4485 {
4486 #if defined(__amd64) && !defined(__xpv)
4487 	if (IOMMU_USED(rdip)) {
4488 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4489 		    cache_flags));
4490 	}
4491 #endif
4492 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4493 	    cache_flags));
4494 }
4495 
4496 /*
4497  * rootnex_valid_sync_parms()
4498  *    checks the parameters passed to sync to verify they are correct.
4499  */
4500 static int
4501 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4502     off_t offset, size_t size, uint_t cache_flags)
4503 {
4504 	off_t woffset;
4505 
4506 
4507 	/*
4508 	 * the first part of the test to make sure the offset passed in is
4509 	 * within the window.
4510 	 */
4511 	if (offset < win->wd_offset) {
4512 		return (DDI_FAILURE);
4513 	}
4514 
4515 	/*
4516 	 * second and last part of the test to make sure the offset and length
4517 	 * passed in is within the window.
4518 	 */
4519 	woffset = offset - win->wd_offset;
4520 	if ((woffset + size) > win->wd_size) {
4521 		return (DDI_FAILURE);
4522 	}
4523 
4524 	/*
4525 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4526 	 * be set too.
4527 	 */
4528 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4529 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4530 		return (DDI_SUCCESS);
4531 	}
4532 
4533 	/*
4534 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4535 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4536 	 */
4537 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4538 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4539 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4540 		return (DDI_SUCCESS);
4541 	}
4542 
4543 	return (DDI_FAILURE);
4544 }
4545 
4546 
4547 /*ARGSUSED*/
4548 static int
4549 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4550     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4551     uint_t *ccountp)
4552 {
4553 	rootnex_window_t *window;
4554 	rootnex_trim_t *trim;
4555 	ddi_dma_impl_t *hp;
4556 	rootnex_dma_t *dma;
4557 #if !defined(__amd64)
4558 	rootnex_sglinfo_t *sinfo;
4559 	rootnex_pgmap_t *pmap;
4560 	uint_t pidx;
4561 	uint_t pcnt;
4562 	off_t poff;
4563 	int i;
4564 #endif
4565 
4566 
4567 	hp = (ddi_dma_impl_t *)handle;
4568 	dma = (rootnex_dma_t *)hp->dmai_private;
4569 #if !defined(__amd64)
4570 	sinfo = &dma->dp_sglinfo;
4571 #endif
4572 
4573 	/* If we try and get a window which doesn't exist, return failure */
4574 	if (win >= hp->dmai_nwin) {
4575 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4576 		return (DDI_FAILURE);
4577 	}
4578 
4579 	/*
4580 	 * if we don't have any windows, and they're asking for the first
4581 	 * window, setup the cookie pointer to the first cookie in the bind.
4582 	 * setup our return values, then increment the cookie since we return
4583 	 * the first cookie on the stack.
4584 	 */
4585 	if (dma->dp_window == NULL) {
4586 		if (win != 0) {
4587 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4588 			return (DDI_FAILURE);
4589 		}
4590 		hp->dmai_cookie = dma->dp_cookies;
4591 		*offp = 0;
4592 		*lenp = dma->dp_dma.dmao_size;
4593 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4594 		*cookiep = hp->dmai_cookie[0];
4595 		hp->dmai_cookie++;
4596 		return (DDI_SUCCESS);
4597 	}
4598 
4599 	/* sync the old window before moving on to the new one */
4600 	window = &dma->dp_window[dma->dp_current_win];
4601 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4602 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4603 		    DDI_DMA_SYNC_FORCPU);
4604 	}
4605 
4606 #if !defined(__amd64)
4607 	/*
4608 	 * before we move to the next window, if we need to re-map, unmap all
4609 	 * the pages in this window.
4610 	 */
4611 	if (dma->dp_cb_remaping) {
4612 		/*
4613 		 * If we switch to this window again, we'll need to map in
4614 		 * on the fly next time.
4615 		 */
4616 		window->wd_remap_copybuf = B_TRUE;
4617 
4618 		/*
4619 		 * calculate the page index into the buffer where this window
4620 		 * starts, and the number of pages this window takes up.
4621 		 */
4622 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4623 		    MMU_PAGESHIFT;
4624 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4625 		    MMU_PAGEOFFSET;
4626 		pcnt = mmu_btopr(window->wd_size + poff);
4627 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4628 
4629 		/* unmap pages which are currently mapped in this window */
4630 		for (i = 0; i < pcnt; i++) {
4631 			if (dma->dp_pgmap[pidx].pm_mapped) {
4632 				hat_unload(kas.a_hat,
4633 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4634 				    HAT_UNLOAD);
4635 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4636 			}
4637 			pidx++;
4638 		}
4639 	}
4640 #endif
4641 
4642 	/*
4643 	 * Move to the new window.
4644 	 * NOTE: current_win must be set for sync to work right
4645 	 */
4646 	dma->dp_current_win = win;
4647 	window = &dma->dp_window[win];
4648 
4649 	/* if needed, adjust the first and/or last cookies for trim */
4650 	trim = &window->wd_trim;
4651 	if (trim->tr_trim_first) {
4652 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4653 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4654 #if !defined(__amd64)
4655 		window->wd_first_cookie->dmac_type =
4656 		    (window->wd_first_cookie->dmac_type &
4657 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4658 #endif
4659 		if (trim->tr_first_copybuf_win) {
4660 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4661 			    trim->tr_first_cbaddr;
4662 #if !defined(__amd64)
4663 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4664 			    trim->tr_first_kaddr;
4665 #endif
4666 		}
4667 	}
4668 	if (trim->tr_trim_last) {
4669 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4670 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4671 		if (trim->tr_last_copybuf_win) {
4672 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4673 			    trim->tr_last_cbaddr;
4674 #if !defined(__amd64)
4675 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4676 			    trim->tr_last_kaddr;
4677 #endif
4678 		}
4679 	}
4680 
4681 	/*
4682 	 * setup the cookie pointer to the first cookie in the window. setup
4683 	 * our return values, then increment the cookie since we return the
4684 	 * first cookie on the stack.
4685 	 */
4686 	hp->dmai_cookie = window->wd_first_cookie;
4687 	*offp = window->wd_offset;
4688 	*lenp = window->wd_size;
4689 	*ccountp = window->wd_cookie_cnt;
4690 	*cookiep = hp->dmai_cookie[0];
4691 	hp->dmai_cookie++;
4692 
4693 #if !defined(__amd64)
4694 	/* re-map copybuf if required for this window */
4695 	if (dma->dp_cb_remaping) {
4696 		/*
4697 		 * calculate the page index into the buffer where this
4698 		 * window starts.
4699 		 */
4700 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4701 		    MMU_PAGESHIFT;
4702 		ASSERT(pidx < sinfo->si_max_pages);
4703 
4704 		/*
4705 		 * the first page can get unmapped if it's shared with the
4706 		 * previous window. Even if the rest of this window is already
4707 		 * mapped in, we need to still check this one.
4708 		 */
4709 		pmap = &dma->dp_pgmap[pidx];
4710 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4711 			if (pmap->pm_pp != NULL) {
4712 				pmap->pm_mapped = B_TRUE;
4713 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4714 			} else if (pmap->pm_vaddr != NULL) {
4715 				pmap->pm_mapped = B_TRUE;
4716 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4717 				    pmap->pm_kaddr);
4718 			}
4719 		}
4720 		pidx++;
4721 
4722 		/* map in the rest of the pages if required */
4723 		if (window->wd_remap_copybuf) {
4724 			window->wd_remap_copybuf = B_FALSE;
4725 
4726 			/* figure out many pages this window takes up */
4727 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4728 			    MMU_PAGEOFFSET;
4729 			pcnt = mmu_btopr(window->wd_size + poff);
4730 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4731 
4732 			/* map pages which require it */
4733 			for (i = 1; i < pcnt; i++) {
4734 				pmap = &dma->dp_pgmap[pidx];
4735 				if (pmap->pm_uses_copybuf) {
4736 					ASSERT(pmap->pm_mapped == B_FALSE);
4737 					if (pmap->pm_pp != NULL) {
4738 						pmap->pm_mapped = B_TRUE;
4739 						i86_pp_map(pmap->pm_pp,
4740 						    pmap->pm_kaddr);
4741 					} else if (pmap->pm_vaddr != NULL) {
4742 						pmap->pm_mapped = B_TRUE;
4743 						i86_va_map(pmap->pm_vaddr,
4744 						    sinfo->si_asp,
4745 						    pmap->pm_kaddr);
4746 					}
4747 				}
4748 				pidx++;
4749 			}
4750 		}
4751 	}
4752 #endif
4753 
4754 	/* if the new window uses the copy buffer, sync it for the device */
4755 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4756 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4757 		    DDI_DMA_SYNC_FORDEV);
4758 	}
4759 
4760 	return (DDI_SUCCESS);
4761 }
4762 
4763 /*
4764  * rootnex_dma_win()
4765  *    called from ddi_dma_getwin()
4766  */
4767 /*ARGSUSED*/
4768 static int
4769 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4770     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4771     uint_t *ccountp)
4772 {
4773 #if defined(__amd64) && !defined(__xpv)
4774 	if (IOMMU_USED(rdip)) {
4775 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4776 		    cookiep, ccountp));
4777 	}
4778 #endif
4779 
4780 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4781 	    cookiep, ccountp));
4782 }
4783 
4784 /*
4785  * ************************
4786  *  obsoleted dma routines
4787  * ************************
4788  */
4789 
4790 /*
4791  * rootnex_dma_map()
4792  *    called from ddi_dma_setup()
4793  * NO IOMMU in 32 bit mode. The below routines doesn't work in 64 bit mode.
4794  */
4795 /* ARGSUSED */
4796 static int
4797 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
4798     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
4799 {
4800 #if defined(__amd64)
4801 	/*
4802 	 * this interface is not supported in 64-bit x86 kernel. See comment in
4803 	 * rootnex_dma_mctl()
4804 	 */
4805 	return (DDI_DMA_NORESOURCES);
4806 
4807 #else /* 32-bit x86 kernel */
4808 	ddi_dma_handle_t *lhandlep;
4809 	ddi_dma_handle_t lhandle;
4810 	ddi_dma_cookie_t cookie;
4811 	ddi_dma_attr_t dma_attr;
4812 	ddi_dma_lim_t *dma_lim;
4813 	uint_t ccnt;
4814 	int e;
4815 
4816 
4817 	/*
4818 	 * if the driver is just testing to see if it's possible to do the bind,
4819 	 * we'll use local state. Otherwise, use the handle pointer passed in.
4820 	 */
4821 	if (handlep == NULL) {
4822 		lhandlep = &lhandle;
4823 	} else {
4824 		lhandlep = handlep;
4825 	}
4826 
4827 	/* convert the limit structure to a dma_attr one */
4828 	dma_lim = dmareq->dmar_limits;
4829 	dma_attr.dma_attr_version = DMA_ATTR_V0;
4830 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
4831 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
4832 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
4833 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
4834 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
4835 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
4836 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
4837 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
4838 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
4839 	dma_attr.dma_attr_align = MMU_PAGESIZE;
4840 	dma_attr.dma_attr_flags = 0;
4841 
4842 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
4843 	    dmareq->dmar_arg, lhandlep);
4844 	if (e != DDI_SUCCESS) {
4845 		return (e);
4846 	}
4847 
4848 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
4849 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
4850 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4851 		return (e);
4852 	}
4853 
4854 	/*
4855 	 * if the driver is just testing to see if it's possible to do the bind,
4856 	 * free up the local state and return the result.
4857 	 */
4858 	if (handlep == NULL) {
4859 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
4860 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4861 		if (e == DDI_DMA_MAPPED) {
4862 			return (DDI_DMA_MAPOK);
4863 		} else {
4864 			return (DDI_DMA_NOMAPPING);
4865 		}
4866 	}
4867 
4868 	return (e);
4869 #endif /* defined(__amd64) */
4870 }
4871 
4872 /*
4873  * rootnex_dma_mctl()
4874  *
4875  * No IOMMU in 32 bit mode. The below routine doesn't work in 64 bit mode.
4876  */
4877 /* ARGSUSED */
4878 static int
4879 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4880     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
4881     uint_t cache_flags)
4882 {
4883 #if defined(__amd64)
4884 	/*
4885 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
4886 	 * common implementation in genunix, so they no longer have x86
4887 	 * specific functionality which called into dma_ctl.
4888 	 *
4889 	 * The rest of the obsoleted interfaces were never supported in the
4890 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
4891 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
4892 	 * implementation issues.
4893 	 *
4894 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
4895 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
4896 	 * reflect that now too...
4897 	 *
4898 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
4899 	 * not going to put this functionality into the 64-bit x86 kernel now.
4900 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
4901 	 * that in a future release.
4902 	 */
4903 	return (DDI_FAILURE);
4904 
4905 #else /* 32-bit x86 kernel */
4906 	ddi_dma_cookie_t lcookie;
4907 	ddi_dma_cookie_t *cookie;
4908 	rootnex_window_t *window;
4909 	ddi_dma_impl_t *hp;
4910 	rootnex_dma_t *dma;
4911 	uint_t nwin;
4912 	uint_t ccnt;
4913 	size_t len;
4914 	off_t off;
4915 	int e;
4916 
4917 
4918 	/*
4919 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
4920 	 * hacky since were optimizing for the current interfaces and so we can
4921 	 * cleanup the mess in genunix. Hopefully we will remove the this
4922 	 * obsoleted routines someday soon.
4923 	 */
4924 
4925 	switch (request) {
4926 
4927 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
4928 		hp = (ddi_dma_impl_t *)handle;
4929 		cookie = (ddi_dma_cookie_t *)objpp;
4930 
4931 		/*
4932 		 * convert segment to cookie. We don't distinguish between the
4933 		 * two :-)
4934 		 */
4935 		*cookie = *hp->dmai_cookie;
4936 		*lenp = cookie->dmac_size;
4937 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
4938 		return (DDI_SUCCESS);
4939 
4940 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
4941 		hp = (ddi_dma_impl_t *)handle;
4942 		dma = (rootnex_dma_t *)hp->dmai_private;
4943 
4944 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
4945 			return (DDI_DMA_STALE);
4946 		}
4947 
4948 		/* handle the case where we don't have any windows */
4949 		if (dma->dp_window == NULL) {
4950 			/*
4951 			 * if seg == NULL, and we don't have any windows,
4952 			 * return the first cookie in the sgl.
4953 			 */
4954 			if (*lenp == NULL) {
4955 				dma->dp_current_cookie = 0;
4956 				hp->dmai_cookie = dma->dp_cookies;
4957 				*objpp = (caddr_t)handle;
4958 				return (DDI_SUCCESS);
4959 
4960 			/* if we have more cookies, go to the next cookie */
4961 			} else {
4962 				if ((dma->dp_current_cookie + 1) >=
4963 				    dma->dp_sglinfo.si_sgl_size) {
4964 					return (DDI_DMA_DONE);
4965 				}
4966 				dma->dp_current_cookie++;
4967 				hp->dmai_cookie++;
4968 				return (DDI_SUCCESS);
4969 			}
4970 		}
4971 
4972 		/* We have one or more windows */
4973 		window = &dma->dp_window[dma->dp_current_win];
4974 
4975 		/*
4976 		 * if seg == NULL, return the first cookie in the current
4977 		 * window
4978 		 */
4979 		if (*lenp == NULL) {
4980 			dma->dp_current_cookie = 0;
4981 			hp->dmai_cookie = window->wd_first_cookie;
4982 
4983 		/*
4984 		 * go to the next cookie in the window then see if we done with
4985 		 * this window.
4986 		 */
4987 		} else {
4988 			if ((dma->dp_current_cookie + 1) >=
4989 			    window->wd_cookie_cnt) {
4990 				return (DDI_DMA_DONE);
4991 			}
4992 			dma->dp_current_cookie++;
4993 			hp->dmai_cookie++;
4994 		}
4995 		*objpp = (caddr_t)handle;
4996 		return (DDI_SUCCESS);
4997 
4998 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
4999 		hp = (ddi_dma_impl_t *)handle;
5000 		dma = (rootnex_dma_t *)hp->dmai_private;
5001 
5002 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
5003 			return (DDI_DMA_STALE);
5004 		}
5005 
5006 		/* if win == NULL, return the first window in the bind */
5007 		if (*offp == NULL) {
5008 			nwin = 0;
5009 
5010 		/*
5011 		 * else, go to the next window then see if we're done with all
5012 		 * the windows.
5013 		 */
5014 		} else {
5015 			nwin = dma->dp_current_win + 1;
5016 			if (nwin >= hp->dmai_nwin) {
5017 				return (DDI_DMA_DONE);
5018 			}
5019 		}
5020 
5021 		/* switch to the next window */
5022 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
5023 		    &lcookie, &ccnt);
5024 		ASSERT(e == DDI_SUCCESS);
5025 		if (e != DDI_SUCCESS) {
5026 			return (DDI_DMA_STALE);
5027 		}
5028 
5029 		/* reset the cookie back to the first cookie in the window */
5030 		if (dma->dp_window != NULL) {
5031 			window = &dma->dp_window[dma->dp_current_win];
5032 			hp->dmai_cookie = window->wd_first_cookie;
5033 		} else {
5034 			hp->dmai_cookie = dma->dp_cookies;
5035 		}
5036 
5037 		*objpp = (caddr_t)handle;
5038 		return (DDI_SUCCESS);
5039 
5040 	case DDI_DMA_FREE: /* ddi_dma_free() */
5041 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
5042 		(void) rootnex_dma_freehdl(dip, rdip, handle);
5043 		if (rootnex_state->r_dvma_call_list_id) {
5044 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
5045 		}
5046 		return (DDI_SUCCESS);
5047 
5048 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
5049 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
5050 		/* should never get here, handled in genunix */
5051 		ASSERT(0);
5052 		return (DDI_FAILURE);
5053 
5054 	case DDI_DMA_KVADDR:
5055 	case DDI_DMA_GETERR:
5056 	case DDI_DMA_COFF:
5057 		return (DDI_FAILURE);
5058 	}
5059 
5060 	return (DDI_FAILURE);
5061 #endif /* defined(__amd64) */
5062 }
5063 
5064 /*
5065  * *********
5066  *  FMA Code
5067  * *********
5068  */
5069 
5070 /*
5071  * rootnex_fm_init()
5072  *    FMA init busop
5073  */
5074 /* ARGSUSED */
5075 static int
5076 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
5077     ddi_iblock_cookie_t *ibc)
5078 {
5079 	*ibc = rootnex_state->r_err_ibc;
5080 
5081 	return (ddi_system_fmcap);
5082 }
5083 
5084 /*
5085  * rootnex_dma_check()
5086  *    Function called after a dma fault occurred to find out whether the
5087  *    fault address is associated with a driver that is able to handle faults
5088  *    and recover from faults.
5089  */
5090 /* ARGSUSED */
5091 static int
5092 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
5093     const void *not_used)
5094 {
5095 	rootnex_window_t *window;
5096 	uint64_t start_addr;
5097 	uint64_t fault_addr;
5098 	ddi_dma_impl_t *hp;
5099 	rootnex_dma_t *dma;
5100 	uint64_t end_addr;
5101 	size_t csize;
5102 	int i;
5103 	int j;
5104 
5105 
5106 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5107 	hp = (ddi_dma_impl_t *)handle;
5108 	ASSERT(hp);
5109 
5110 	dma = (rootnex_dma_t *)hp->dmai_private;
5111 
5112 	/* Get the address that we need to search for */
5113 	fault_addr = *(uint64_t *)addr;
5114 
5115 	/*
5116 	 * if we don't have any windows, we can just walk through all the
5117 	 * cookies.
5118 	 */
5119 	if (dma->dp_window == NULL) {
5120 		/* for each cookie */
5121 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5122 			/*
5123 			 * if the faulted address is within the physical address
5124 			 * range of the cookie, return DDI_FM_NONFATAL.
5125 			 */
5126 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5127 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5128 			    dma->dp_cookies[i].dmac_size))) {
5129 				return (DDI_FM_NONFATAL);
5130 			}
5131 		}
5132 
5133 		/* fault_addr not within this DMA handle */
5134 		return (DDI_FM_UNKNOWN);
5135 	}
5136 
5137 	/* we have mutiple windows, walk through each window */
5138 	for (i = 0; i < hp->dmai_nwin; i++) {
5139 		window = &dma->dp_window[i];
5140 
5141 		/* Go through all the cookies in the window */
5142 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5143 
5144 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5145 			csize = window->wd_first_cookie[j].dmac_size;
5146 
5147 			/*
5148 			 * if we are trimming the first cookie in the window,
5149 			 * and this is the first cookie, adjust the start
5150 			 * address and size of the cookie to account for the
5151 			 * trim.
5152 			 */
5153 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5154 				start_addr = window->wd_trim.tr_first_paddr;
5155 				csize = window->wd_trim.tr_first_size;
5156 			}
5157 
5158 			/*
5159 			 * if we are trimming the last cookie in the window,
5160 			 * and this is the last cookie, adjust the start
5161 			 * address and size of the cookie to account for the
5162 			 * trim.
5163 			 */
5164 			if (window->wd_trim.tr_trim_last &&
5165 			    (j == (window->wd_cookie_cnt - 1))) {
5166 				start_addr = window->wd_trim.tr_last_paddr;
5167 				csize = window->wd_trim.tr_last_size;
5168 			}
5169 
5170 			end_addr = start_addr + csize;
5171 
5172 			/*
5173 			 * if the faulted address is within the physical
5174 			 * address of the cookie, return DDI_FM_NONFATAL.
5175 			 */
5176 			if ((fault_addr >= start_addr) &&
5177 			    (fault_addr <= end_addr)) {
5178 				return (DDI_FM_NONFATAL);
5179 			}
5180 		}
5181 	}
5182 
5183 	/* fault_addr not within this DMA handle */
5184 	return (DDI_FM_UNKNOWN);
5185 }
5186 
5187 /*ARGSUSED*/
5188 static int
5189 rootnex_quiesce(dev_info_t *dip)
5190 {
5191 #if defined(__amd64) && !defined(__xpv)
5192 	return (immu_quiesce());
5193 #else
5194 	return (DDI_SUCCESS);
5195 #endif
5196 }
5197 
5198 #if defined(__xpv)
5199 void
5200 immu_init(void)
5201 {
5202 	;
5203 }
5204 
5205 void
5206 immu_startup(void)
5207 {
5208 	;
5209 }
5210 /*ARGSUSED*/
5211 void
5212 immu_physmem_update(uint64_t addr, uint64_t size)
5213 {
5214 	;
5215 }
5216 #endif
5217