xref: /titanic_41/usr/src/uts/i86pc/io/rootnex.c (revision e65fcc69bb33b3f4525b0c2c9732ece17c90b196)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
26  */
27 
28 /*
29  * x86 root nexus driver
30  */
31 
32 #include <sys/sysmacros.h>
33 #include <sys/conf.h>
34 #include <sys/autoconf.h>
35 #include <sys/sysmacros.h>
36 #include <sys/debug.h>
37 #include <sys/psw.h>
38 #include <sys/ddidmareq.h>
39 #include <sys/promif.h>
40 #include <sys/devops.h>
41 #include <sys/kmem.h>
42 #include <sys/cmn_err.h>
43 #include <vm/seg.h>
44 #include <vm/seg_kmem.h>
45 #include <vm/seg_dev.h>
46 #include <sys/vmem.h>
47 #include <sys/mman.h>
48 #include <vm/hat.h>
49 #include <vm/as.h>
50 #include <vm/page.h>
51 #include <sys/avintr.h>
52 #include <sys/errno.h>
53 #include <sys/modctl.h>
54 #include <sys/ddi_impldefs.h>
55 #include <sys/sunddi.h>
56 #include <sys/sunndi.h>
57 #include <sys/mach_intr.h>
58 #include <sys/psm.h>
59 #include <sys/ontrap.h>
60 #include <sys/atomic.h>
61 #include <sys/sdt.h>
62 #include <sys/rootnex.h>
63 #include <vm/hat_i86.h>
64 #include <sys/ddifm.h>
65 #include <sys/ddi_isa.h>
66 #include <sys/apic.h>
67 
68 #ifdef __xpv
69 #include <sys/bootinfo.h>
70 #include <sys/hypervisor.h>
71 #include <sys/bootconf.h>
72 #include <vm/kboot_mmu.h>
73 #endif
74 
75 #if defined(__amd64) && !defined(__xpv)
76 #include <sys/immu.h>
77 #endif
78 
79 
80 /*
81  * enable/disable extra checking of function parameters. Useful for debugging
82  * drivers.
83  */
84 #ifdef	DEBUG
85 int rootnex_alloc_check_parms = 1;
86 int rootnex_bind_check_parms = 1;
87 int rootnex_bind_check_inuse = 1;
88 int rootnex_unbind_verify_buffer = 0;
89 int rootnex_sync_check_parms = 1;
90 #else
91 int rootnex_alloc_check_parms = 0;
92 int rootnex_bind_check_parms = 0;
93 int rootnex_bind_check_inuse = 0;
94 int rootnex_unbind_verify_buffer = 0;
95 int rootnex_sync_check_parms = 0;
96 #endif
97 
98 boolean_t rootnex_dmar_not_setup;
99 
100 /* Master Abort and Target Abort panic flag */
101 int rootnex_fm_ma_ta_panic_flag = 0;
102 
103 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
104 int rootnex_bind_fail = 1;
105 int rootnex_bind_warn = 1;
106 uint8_t *rootnex_warn_list;
107 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
108 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
109 
110 /*
111  * revert back to old broken behavior of always sync'ing entire copy buffer.
112  * This is useful if be have a buggy driver which doesn't correctly pass in
113  * the offset and size into ddi_dma_sync().
114  */
115 int rootnex_sync_ignore_params = 0;
116 
117 /*
118  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
119  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
120  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
121  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
122  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
123  * (< 8K). We will still need to allocate the copy buffer during bind though
124  * (if we need one). These can only be modified in /etc/system before rootnex
125  * attach.
126  */
127 #if defined(__amd64)
128 int rootnex_prealloc_cookies = 65;
129 int rootnex_prealloc_windows = 4;
130 int rootnex_prealloc_copybuf = 2;
131 #else
132 int rootnex_prealloc_cookies = 33;
133 int rootnex_prealloc_windows = 4;
134 int rootnex_prealloc_copybuf = 2;
135 #endif
136 
137 /* driver global state */
138 static rootnex_state_t *rootnex_state;
139 
140 #ifdef DEBUG
141 /* shortcut to rootnex counters */
142 static uint64_t *rootnex_cnt;
143 #endif
144 
145 /*
146  * XXX - does x86 even need these or are they left over from the SPARC days?
147  */
148 /* statically defined integer/boolean properties for the root node */
149 static rootnex_intprop_t rootnex_intprp[] = {
150 	{ "PAGESIZE",			PAGESIZE },
151 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
152 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
153 	{ DDI_RELATIVE_ADDRESSING,	1 },
154 };
155 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
156 
157 /*
158  * If we're dom0, we're using a real device so we need to load
159  * the cookies with MFNs instead of PFNs.
160  */
161 #ifdef __xpv
162 typedef maddr_t rootnex_addr_t;
163 #define	ROOTNEX_PADDR_TO_RBASE(pa)	\
164 	(DOMAIN_IS_INITDOMAIN(xen_info) ? pa_to_ma(pa) : (pa))
165 #else
166 typedef paddr_t rootnex_addr_t;
167 #define	ROOTNEX_PADDR_TO_RBASE(pa)	(pa)
168 #endif
169 
170 #if !defined(__xpv)
171 char _depends_on[] = "misc/iommulib misc/acpica";
172 #endif
173 
174 static struct cb_ops rootnex_cb_ops = {
175 	nodev,		/* open */
176 	nodev,		/* close */
177 	nodev,		/* strategy */
178 	nodev,		/* print */
179 	nodev,		/* dump */
180 	nodev,		/* read */
181 	nodev,		/* write */
182 	nodev,		/* ioctl */
183 	nodev,		/* devmap */
184 	nodev,		/* mmap */
185 	nodev,		/* segmap */
186 	nochpoll,	/* chpoll */
187 	ddi_prop_op,	/* cb_prop_op */
188 	NULL,		/* struct streamtab */
189 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
190 	CB_REV,		/* Rev */
191 	nodev,		/* cb_aread */
192 	nodev		/* cb_awrite */
193 };
194 
195 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
196     off_t offset, off_t len, caddr_t *vaddrp);
197 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
198     struct hat *hat, struct seg *seg, caddr_t addr,
199     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
200 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
201     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
202 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
203     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
204     ddi_dma_handle_t *handlep);
205 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
206     ddi_dma_handle_t handle);
207 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
208     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
209     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
210 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
211     ddi_dma_handle_t handle);
212 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
213     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
214 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
215     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
216     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
217 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
218     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
219     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
220 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
221     ddi_ctl_enum_t ctlop, void *arg, void *result);
222 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
223     ddi_iblock_cookie_t *ibc);
224 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
225     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
226 static int rootnex_alloc_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *,
227     void *);
228 static int rootnex_free_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *);
229 
230 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
231     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
232     ddi_dma_handle_t *handlep);
233 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
234     ddi_dma_handle_t handle);
235 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
236     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
237     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
238 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
239     ddi_dma_handle_t handle);
240 #if defined(__amd64) && !defined(__xpv)
241 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
242     ddi_dma_handle_t handle);
243 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
244     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
245 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
246     ddi_dma_cookie_t *cookiep, uint_t ccount);
247 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
248     ddi_dma_handle_t handle);
249 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
250 #endif
251 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
252     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
253 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
254     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
255     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
256 
257 #if defined(__amd64) && !defined(__xpv)
258 static int rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
259     ddi_dma_handle_t handle, void *v);
260 static void *rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
261     ddi_dma_handle_t handle);
262 #endif
263 
264 
265 static struct bus_ops rootnex_bus_ops = {
266 	BUSO_REV,
267 	rootnex_map,
268 	NULL,
269 	NULL,
270 	NULL,
271 	rootnex_map_fault,
272 	rootnex_dma_map,
273 	rootnex_dma_allochdl,
274 	rootnex_dma_freehdl,
275 	rootnex_dma_bindhdl,
276 	rootnex_dma_unbindhdl,
277 	rootnex_dma_sync,
278 	rootnex_dma_win,
279 	rootnex_dma_mctl,
280 	rootnex_ctlops,
281 	ddi_bus_prop_op,
282 	i_ddi_rootnex_get_eventcookie,
283 	i_ddi_rootnex_add_eventcall,
284 	i_ddi_rootnex_remove_eventcall,
285 	i_ddi_rootnex_post_event,
286 	0,			/* bus_intr_ctl */
287 	0,			/* bus_config */
288 	0,			/* bus_unconfig */
289 	rootnex_fm_init,	/* bus_fm_init */
290 	NULL,			/* bus_fm_fini */
291 	NULL,			/* bus_fm_access_enter */
292 	NULL,			/* bus_fm_access_exit */
293 	NULL,			/* bus_powr */
294 	rootnex_intr_ops	/* bus_intr_op */
295 };
296 
297 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
298 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
299 static int rootnex_quiesce(dev_info_t *dip);
300 
301 static struct dev_ops rootnex_ops = {
302 	DEVO_REV,
303 	0,
304 	ddi_no_info,
305 	nulldev,
306 	nulldev,
307 	rootnex_attach,
308 	rootnex_detach,
309 	nulldev,
310 	&rootnex_cb_ops,
311 	&rootnex_bus_ops,
312 	NULL,
313 	rootnex_quiesce,		/* quiesce */
314 };
315 
316 static struct modldrv rootnex_modldrv = {
317 	&mod_driverops,
318 	"i86pc root nexus",
319 	&rootnex_ops
320 };
321 
322 static struct modlinkage rootnex_modlinkage = {
323 	MODREV_1,
324 	(void *)&rootnex_modldrv,
325 	NULL
326 };
327 
328 #if defined(__amd64) && !defined(__xpv)
329 static iommulib_nexops_t iommulib_nexops = {
330 	IOMMU_NEXOPS_VERSION,
331 	"Rootnex IOMMU ops Vers 1.1",
332 	NULL,
333 	rootnex_coredma_allochdl,
334 	rootnex_coredma_freehdl,
335 	rootnex_coredma_bindhdl,
336 	rootnex_coredma_unbindhdl,
337 	rootnex_coredma_reset_cookies,
338 	rootnex_coredma_get_cookies,
339 	rootnex_coredma_set_cookies,
340 	rootnex_coredma_clear_cookies,
341 	rootnex_coredma_get_sleep_flags,
342 	rootnex_coredma_sync,
343 	rootnex_coredma_win,
344 	rootnex_dma_map,
345 	rootnex_dma_mctl,
346 	rootnex_coredma_hdl_setprivate,
347 	rootnex_coredma_hdl_getprivate
348 };
349 #endif
350 
351 /*
352  *  extern hacks
353  */
354 extern struct seg_ops segdev_ops;
355 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
356 #ifdef	DDI_MAP_DEBUG
357 extern int ddi_map_debug_flag;
358 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
359 #endif
360 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
361 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
362 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
363     psm_intr_op_t, int *);
364 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
365 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
366 
367 /*
368  * Use device arena to use for device control register mappings.
369  * Various kernel memory walkers (debugger, dtrace) need to know
370  * to avoid this address range to prevent undesired device activity.
371  */
372 extern void *device_arena_alloc(size_t size, int vm_flag);
373 extern void device_arena_free(void * vaddr, size_t size);
374 
375 
376 /*
377  *  Internal functions
378  */
379 static int rootnex_dma_init();
380 static void rootnex_add_props(dev_info_t *);
381 static int rootnex_ctl_reportdev(dev_info_t *dip);
382 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
383 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
384 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
385 static int rootnex_map_handle(ddi_map_req_t *mp);
386 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
387 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
388 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
389     ddi_dma_attr_t *attr);
390 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
391     rootnex_sglinfo_t *sglinfo);
392 static void rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object,
393     ddi_dma_cookie_t *sgl, rootnex_sglinfo_t *sglinfo);
394 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
395     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
396 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
397     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
398 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
399 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
400     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
401 static void rootnex_teardown_windows(rootnex_dma_t *dma);
402 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
403     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
404 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
405     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
406     size_t *copybuf_used, page_t **cur_pp);
407 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
408     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
409     ddi_dma_attr_t *attr, off_t cur_offset);
410 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
411     rootnex_dma_t *dma, rootnex_window_t **windowp,
412     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
413 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
414     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
415 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
416     off_t offset, size_t size, uint_t cache_flags);
417 static int rootnex_verify_buffer(rootnex_dma_t *dma);
418 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
419     const void *comp_addr, const void *not_used);
420 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
421     rootnex_sglinfo_t *sglinfo);
422 static struct as *rootnex_get_as(ddi_dma_obj_t *dmar_object);
423 
424 /*
425  * _init()
426  *
427  */
428 int
429 _init(void)
430 {
431 
432 	rootnex_state = NULL;
433 	return (mod_install(&rootnex_modlinkage));
434 }
435 
436 
437 /*
438  * _info()
439  *
440  */
441 int
442 _info(struct modinfo *modinfop)
443 {
444 	return (mod_info(&rootnex_modlinkage, modinfop));
445 }
446 
447 
448 /*
449  * _fini()
450  *
451  */
452 int
453 _fini(void)
454 {
455 	return (EBUSY);
456 }
457 
458 
459 /*
460  * rootnex_attach()
461  *
462  */
463 static int
464 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
465 {
466 	int fmcap;
467 	int e;
468 
469 	switch (cmd) {
470 	case DDI_ATTACH:
471 		break;
472 	case DDI_RESUME:
473 #if defined(__amd64) && !defined(__xpv)
474 		return (immu_unquiesce());
475 #else
476 		return (DDI_SUCCESS);
477 #endif
478 	default:
479 		return (DDI_FAILURE);
480 	}
481 
482 	/*
483 	 * We should only have one instance of rootnex. Save it away since we
484 	 * don't have an easy way to get it back later.
485 	 */
486 	ASSERT(rootnex_state == NULL);
487 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
488 
489 	rootnex_state->r_dip = dip;
490 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
491 	rootnex_state->r_reserved_msg_printed = B_FALSE;
492 #ifdef DEBUG
493 	rootnex_cnt = &rootnex_state->r_counters[0];
494 #endif
495 
496 	/*
497 	 * Set minimum fm capability level for i86pc platforms and then
498 	 * initialize error handling. Since we're the rootnex, we don't
499 	 * care what's returned in the fmcap field.
500 	 */
501 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
502 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
503 	fmcap = ddi_system_fmcap;
504 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
505 
506 	/* initialize DMA related state */
507 	e = rootnex_dma_init();
508 	if (e != DDI_SUCCESS) {
509 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
510 		return (DDI_FAILURE);
511 	}
512 
513 	/* Add static root node properties */
514 	rootnex_add_props(dip);
515 
516 	/* since we can't call ddi_report_dev() */
517 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
518 
519 	/* Initialize rootnex event handle */
520 	i_ddi_rootnex_init_events(dip);
521 
522 #if defined(__amd64) && !defined(__xpv)
523 	e = iommulib_nexus_register(dip, &iommulib_nexops,
524 	    &rootnex_state->r_iommulib_handle);
525 
526 	ASSERT(e == DDI_SUCCESS);
527 #endif
528 
529 	return (DDI_SUCCESS);
530 }
531 
532 
533 /*
534  * rootnex_detach()
535  *
536  */
537 /*ARGSUSED*/
538 static int
539 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
540 {
541 	switch (cmd) {
542 	case DDI_SUSPEND:
543 #if defined(__amd64) && !defined(__xpv)
544 		return (immu_quiesce());
545 #else
546 		return (DDI_SUCCESS);
547 #endif
548 	default:
549 		return (DDI_FAILURE);
550 	}
551 	/*NOTREACHED*/
552 
553 }
554 
555 
556 /*
557  * rootnex_dma_init()
558  *
559  */
560 /*ARGSUSED*/
561 static int
562 rootnex_dma_init()
563 {
564 	size_t bufsize;
565 
566 
567 	/*
568 	 * size of our cookie/window/copybuf state needed in dma bind that we
569 	 * pre-alloc in dma_alloc_handle
570 	 */
571 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
572 	rootnex_state->r_prealloc_size =
573 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
574 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
575 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
576 
577 	/*
578 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
579 	 * allocate 16 extra bytes for struct pointer alignment
580 	 * (p->dmai_private & dma->dp_prealloc_buffer)
581 	 */
582 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
583 	    rootnex_state->r_prealloc_size + 0x10;
584 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
585 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
586 	if (rootnex_state->r_dmahdl_cache == NULL) {
587 		return (DDI_FAILURE);
588 	}
589 
590 	/*
591 	 * allocate array to track which major numbers we have printed warnings
592 	 * for.
593 	 */
594 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
595 	    KM_SLEEP);
596 
597 	return (DDI_SUCCESS);
598 }
599 
600 
601 /*
602  * rootnex_add_props()
603  *
604  */
605 static void
606 rootnex_add_props(dev_info_t *dip)
607 {
608 	rootnex_intprop_t *rpp;
609 	int i;
610 
611 	/* Add static integer/boolean properties to the root node */
612 	rpp = rootnex_intprp;
613 	for (i = 0; i < NROOT_INTPROPS; i++) {
614 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
615 		    rpp[i].prop_name, rpp[i].prop_value);
616 	}
617 }
618 
619 
620 
621 /*
622  * *************************
623  *  ctlops related routines
624  * *************************
625  */
626 
627 /*
628  * rootnex_ctlops()
629  *
630  */
631 /*ARGSUSED*/
632 static int
633 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
634     void *arg, void *result)
635 {
636 	int n, *ptr;
637 	struct ddi_parent_private_data *pdp;
638 
639 	switch (ctlop) {
640 	case DDI_CTLOPS_DMAPMAPC:
641 		/*
642 		 * Return 'partial' to indicate that dma mapping
643 		 * has to be done in the main MMU.
644 		 */
645 		return (DDI_DMA_PARTIAL);
646 
647 	case DDI_CTLOPS_BTOP:
648 		/*
649 		 * Convert byte count input to physical page units.
650 		 * (byte counts that are not a page-size multiple
651 		 * are rounded down)
652 		 */
653 		*(ulong_t *)result = btop(*(ulong_t *)arg);
654 		return (DDI_SUCCESS);
655 
656 	case DDI_CTLOPS_PTOB:
657 		/*
658 		 * Convert size in physical pages to bytes
659 		 */
660 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
661 		return (DDI_SUCCESS);
662 
663 	case DDI_CTLOPS_BTOPR:
664 		/*
665 		 * Convert byte count input to physical page units
666 		 * (byte counts that are not a page-size multiple
667 		 * are rounded up)
668 		 */
669 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
670 		return (DDI_SUCCESS);
671 
672 	case DDI_CTLOPS_INITCHILD:
673 		return (impl_ddi_sunbus_initchild(arg));
674 
675 	case DDI_CTLOPS_UNINITCHILD:
676 		impl_ddi_sunbus_removechild(arg);
677 		return (DDI_SUCCESS);
678 
679 	case DDI_CTLOPS_REPORTDEV:
680 		return (rootnex_ctl_reportdev(rdip));
681 
682 	case DDI_CTLOPS_IOMIN:
683 		/*
684 		 * Nothing to do here but reflect back..
685 		 */
686 		return (DDI_SUCCESS);
687 
688 	case DDI_CTLOPS_REGSIZE:
689 	case DDI_CTLOPS_NREGS:
690 		break;
691 
692 	case DDI_CTLOPS_SIDDEV:
693 		if (ndi_dev_is_prom_node(rdip))
694 			return (DDI_SUCCESS);
695 		if (ndi_dev_is_persistent_node(rdip))
696 			return (DDI_SUCCESS);
697 		return (DDI_FAILURE);
698 
699 	case DDI_CTLOPS_POWER:
700 		return ((*pm_platform_power)((power_req_t *)arg));
701 
702 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
703 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
704 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
705 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
706 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
707 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
708 		if (!rootnex_state->r_reserved_msg_printed) {
709 			rootnex_state->r_reserved_msg_printed = B_TRUE;
710 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
711 			    "1 or more reserved/obsolete operations.");
712 		}
713 		return (DDI_FAILURE);
714 
715 	default:
716 		return (DDI_FAILURE);
717 	}
718 	/*
719 	 * The rest are for "hardware" properties
720 	 */
721 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
722 		return (DDI_FAILURE);
723 
724 	if (ctlop == DDI_CTLOPS_NREGS) {
725 		ptr = (int *)result;
726 		*ptr = pdp->par_nreg;
727 	} else {
728 		off_t *size = (off_t *)result;
729 
730 		ptr = (int *)arg;
731 		n = *ptr;
732 		if (n >= pdp->par_nreg) {
733 			return (DDI_FAILURE);
734 		}
735 		*size = (off_t)pdp->par_reg[n].regspec_size;
736 	}
737 	return (DDI_SUCCESS);
738 }
739 
740 
741 /*
742  * rootnex_ctl_reportdev()
743  *
744  */
745 static int
746 rootnex_ctl_reportdev(dev_info_t *dev)
747 {
748 	int i, n, len, f_len = 0;
749 	char *buf;
750 
751 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
752 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
753 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
754 	len = strlen(buf);
755 
756 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
757 
758 		struct regspec *rp = sparc_pd_getreg(dev, i);
759 
760 		if (i == 0)
761 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
762 			    ": ");
763 		else
764 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
765 			    " and ");
766 		len = strlen(buf);
767 
768 		switch (rp->regspec_bustype) {
769 
770 		case BTEISA:
771 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
772 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
773 			break;
774 
775 		case BTISA:
776 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
777 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
778 			break;
779 
780 		default:
781 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
782 			    "space %x offset %x",
783 			    rp->regspec_bustype, rp->regspec_addr);
784 			break;
785 		}
786 		len = strlen(buf);
787 	}
788 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
789 		int pri;
790 
791 		if (i != 0) {
792 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
793 			    ",");
794 			len = strlen(buf);
795 		}
796 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
797 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
798 		    " sparc ipl %d", pri);
799 		len = strlen(buf);
800 	}
801 #ifdef DEBUG
802 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
803 		cmn_err(CE_NOTE, "next message is truncated: "
804 		    "printed length 1024, real length %d", f_len);
805 	}
806 #endif /* DEBUG */
807 	cmn_err(CE_CONT, "?%s\n", buf);
808 	kmem_free(buf, REPORTDEV_BUFSIZE);
809 	return (DDI_SUCCESS);
810 }
811 
812 
813 /*
814  * ******************
815  *  map related code
816  * ******************
817  */
818 
819 /*
820  * rootnex_map()
821  *
822  */
823 static int
824 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
825     off_t len, caddr_t *vaddrp)
826 {
827 	struct regspec *rp, tmp_reg;
828 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
829 	int error;
830 
831 	mp = &mr;
832 
833 	switch (mp->map_op)  {
834 	case DDI_MO_MAP_LOCKED:
835 	case DDI_MO_UNMAP:
836 	case DDI_MO_MAP_HANDLE:
837 		break;
838 	default:
839 #ifdef	DDI_MAP_DEBUG
840 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
841 		    mp->map_op);
842 #endif	/* DDI_MAP_DEBUG */
843 		return (DDI_ME_UNIMPLEMENTED);
844 	}
845 
846 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
847 #ifdef	DDI_MAP_DEBUG
848 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
849 #endif	/* DDI_MAP_DEBUG */
850 		return (DDI_ME_UNIMPLEMENTED);
851 	}
852 
853 	/*
854 	 * First, if given an rnumber, convert it to a regspec...
855 	 * (Presumably, this is on behalf of a child of the root node?)
856 	 */
857 
858 	if (mp->map_type == DDI_MT_RNUMBER)  {
859 
860 		int rnumber = mp->map_obj.rnumber;
861 #ifdef	DDI_MAP_DEBUG
862 		static char *out_of_range =
863 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
864 #endif	/* DDI_MAP_DEBUG */
865 
866 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
867 		if (rp == NULL)  {
868 #ifdef	DDI_MAP_DEBUG
869 			cmn_err(CE_WARN, out_of_range, rnumber,
870 			    ddi_get_name(rdip));
871 #endif	/* DDI_MAP_DEBUG */
872 			return (DDI_ME_RNUMBER_RANGE);
873 		}
874 
875 		/*
876 		 * Convert the given ddi_map_req_t from rnumber to regspec...
877 		 */
878 
879 		mp->map_type = DDI_MT_REGSPEC;
880 		mp->map_obj.rp = rp;
881 	}
882 
883 	/*
884 	 * Adjust offset and length correspnding to called values...
885 	 * XXX: A non-zero length means override the one in the regspec
886 	 * XXX: (regardless of what's in the parent's range?)
887 	 */
888 
889 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
890 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
891 
892 #ifdef	DDI_MAP_DEBUG
893 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
894 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
895 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
896 	    len, mp->map_handlep);
897 #endif	/* DDI_MAP_DEBUG */
898 
899 	/*
900 	 * I/O or memory mapping:
901 	 *
902 	 *	<bustype=0, addr=x, len=x>: memory
903 	 *	<bustype=1, addr=x, len=x>: i/o
904 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
905 	 */
906 
907 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
908 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
909 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
910 		    ddi_get_name(rdip), rp->regspec_bustype,
911 		    rp->regspec_addr, rp->regspec_size);
912 		return (DDI_ME_INVAL);
913 	}
914 
915 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
916 		/*
917 		 * compatibility i/o mapping
918 		 */
919 		rp->regspec_bustype += (uint_t)offset;
920 	} else {
921 		/*
922 		 * Normal memory or i/o mapping
923 		 */
924 		rp->regspec_addr += (uint_t)offset;
925 	}
926 
927 	if (len != 0)
928 		rp->regspec_size = (uint_t)len;
929 
930 #ifdef	DDI_MAP_DEBUG
931 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
932 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
933 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
934 	    offset, len, mp->map_handlep);
935 #endif	/* DDI_MAP_DEBUG */
936 
937 	/*
938 	 * Apply any parent ranges at this level, if applicable.
939 	 * (This is where nexus specific regspec translation takes place.
940 	 * Use of this function is implicit agreement that translation is
941 	 * provided via ddi_apply_range.)
942 	 */
943 
944 #ifdef	DDI_MAP_DEBUG
945 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
946 	    ddi_get_name(dip), ddi_get_name(rdip));
947 #endif	/* DDI_MAP_DEBUG */
948 
949 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
950 		return (error);
951 
952 	switch (mp->map_op)  {
953 	case DDI_MO_MAP_LOCKED:
954 
955 		/*
956 		 * Set up the locked down kernel mapping to the regspec...
957 		 */
958 
959 		return (rootnex_map_regspec(mp, vaddrp));
960 
961 	case DDI_MO_UNMAP:
962 
963 		/*
964 		 * Release mapping...
965 		 */
966 
967 		return (rootnex_unmap_regspec(mp, vaddrp));
968 
969 	case DDI_MO_MAP_HANDLE:
970 
971 		return (rootnex_map_handle(mp));
972 
973 	default:
974 		return (DDI_ME_UNIMPLEMENTED);
975 	}
976 }
977 
978 
979 /*
980  * rootnex_map_fault()
981  *
982  *	fault in mappings for requestors
983  */
984 /*ARGSUSED*/
985 static int
986 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
987     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
988     uint_t lock)
989 {
990 
991 #ifdef	DDI_MAP_DEBUG
992 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
993 	ddi_map_debug(" Seg <%s>\n",
994 	    seg->s_ops == &segdev_ops ? "segdev" :
995 	    seg == &kvseg ? "segkmem" : "NONE!");
996 #endif	/* DDI_MAP_DEBUG */
997 
998 	/*
999 	 * This is all terribly broken, but it is a start
1000 	 *
1001 	 * XXX	Note that this test means that segdev_ops
1002 	 *	must be exported from seg_dev.c.
1003 	 * XXX	What about devices with their own segment drivers?
1004 	 */
1005 	if (seg->s_ops == &segdev_ops) {
1006 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
1007 
1008 		if (hat == NULL) {
1009 			/*
1010 			 * This is one plausible interpretation of
1011 			 * a null hat i.e. use the first hat on the
1012 			 * address space hat list which by convention is
1013 			 * the hat of the system MMU.  At alternative
1014 			 * would be to panic .. this might well be better ..
1015 			 */
1016 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
1017 			hat = seg->s_as->a_hat;
1018 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
1019 		}
1020 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
1021 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
1022 	} else if (seg == &kvseg && dp == NULL) {
1023 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
1024 		    HAT_LOAD_LOCK);
1025 	} else
1026 		return (DDI_FAILURE);
1027 	return (DDI_SUCCESS);
1028 }
1029 
1030 
1031 /*
1032  * rootnex_map_regspec()
1033  *     we don't support mapping of I/O cards above 4Gb
1034  */
1035 static int
1036 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1037 {
1038 	rootnex_addr_t rbase;
1039 	void *cvaddr;
1040 	uint_t npages, pgoffset;
1041 	struct regspec *rp;
1042 	ddi_acc_hdl_t *hp;
1043 	ddi_acc_impl_t *ap;
1044 	uint_t	hat_acc_flags;
1045 	paddr_t pbase;
1046 
1047 	rp = mp->map_obj.rp;
1048 	hp = mp->map_handlep;
1049 
1050 #ifdef	DDI_MAP_DEBUG
1051 	ddi_map_debug(
1052 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1053 	    rp->regspec_bustype, rp->regspec_addr,
1054 	    rp->regspec_size, mp->map_handlep);
1055 #endif	/* DDI_MAP_DEBUG */
1056 
1057 	/*
1058 	 * I/O or memory mapping
1059 	 *
1060 	 *	<bustype=0, addr=x, len=x>: memory
1061 	 *	<bustype=1, addr=x, len=x>: i/o
1062 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1063 	 */
1064 
1065 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1066 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1067 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1068 		    rp->regspec_addr, rp->regspec_size);
1069 		return (DDI_FAILURE);
1070 	}
1071 
1072 	if (rp->regspec_bustype != 0) {
1073 		/*
1074 		 * I/O space - needs a handle.
1075 		 */
1076 		if (hp == NULL) {
1077 			return (DDI_FAILURE);
1078 		}
1079 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1080 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1081 		impl_acc_hdl_init(hp);
1082 
1083 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1084 #ifdef  DDI_MAP_DEBUG
1085 			ddi_map_debug("rootnex_map_regspec: mmap() "
1086 			    "to I/O space is not supported.\n");
1087 #endif  /* DDI_MAP_DEBUG */
1088 			return (DDI_ME_INVAL);
1089 		} else {
1090 			/*
1091 			 * 1275-compliant vs. compatibility i/o mapping
1092 			 */
1093 			*vaddrp =
1094 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1095 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1096 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1097 #ifdef __xpv
1098 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1099 				hp->ah_pfn = xen_assign_pfn(
1100 				    mmu_btop((ulong_t)rp->regspec_addr &
1101 				    MMU_PAGEMASK));
1102 			} else {
1103 				hp->ah_pfn = mmu_btop(
1104 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1105 			}
1106 #else
1107 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1108 			    MMU_PAGEMASK);
1109 #endif
1110 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1111 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1112 		}
1113 
1114 #ifdef	DDI_MAP_DEBUG
1115 		ddi_map_debug(
1116 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1117 		    rp->regspec_size, *vaddrp);
1118 #endif	/* DDI_MAP_DEBUG */
1119 		return (DDI_SUCCESS);
1120 	}
1121 
1122 	/*
1123 	 * Memory space
1124 	 */
1125 
1126 	if (hp != NULL) {
1127 		/*
1128 		 * hat layer ignores
1129 		 * hp->ah_acc.devacc_attr_endian_flags.
1130 		 */
1131 		switch (hp->ah_acc.devacc_attr_dataorder) {
1132 		case DDI_STRICTORDER_ACC:
1133 			hat_acc_flags = HAT_STRICTORDER;
1134 			break;
1135 		case DDI_UNORDERED_OK_ACC:
1136 			hat_acc_flags = HAT_UNORDERED_OK;
1137 			break;
1138 		case DDI_MERGING_OK_ACC:
1139 			hat_acc_flags = HAT_MERGING_OK;
1140 			break;
1141 		case DDI_LOADCACHING_OK_ACC:
1142 			hat_acc_flags = HAT_LOADCACHING_OK;
1143 			break;
1144 		case DDI_STORECACHING_OK_ACC:
1145 			hat_acc_flags = HAT_STORECACHING_OK;
1146 			break;
1147 		}
1148 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1149 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1150 		impl_acc_hdl_init(hp);
1151 		hp->ah_hat_flags = hat_acc_flags;
1152 	} else {
1153 		hat_acc_flags = HAT_STRICTORDER;
1154 	}
1155 
1156 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1157 #ifdef __xpv
1158 	/*
1159 	 * If we're dom0, we're using a real device so we need to translate
1160 	 * the MA to a PA.
1161 	 */
1162 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1163 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1164 	} else {
1165 		pbase = rbase;
1166 	}
1167 #else
1168 	pbase = rbase;
1169 #endif
1170 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1171 
1172 	if (rp->regspec_size == 0) {
1173 #ifdef  DDI_MAP_DEBUG
1174 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1175 #endif  /* DDI_MAP_DEBUG */
1176 		return (DDI_ME_INVAL);
1177 	}
1178 
1179 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1180 		/* extra cast to make gcc happy */
1181 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1182 	} else {
1183 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1184 
1185 #ifdef	DDI_MAP_DEBUG
1186 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1187 		    "physical %llx", npages, pbase);
1188 #endif	/* DDI_MAP_DEBUG */
1189 
1190 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1191 		if (cvaddr == NULL)
1192 			return (DDI_ME_NORESOURCES);
1193 
1194 		/*
1195 		 * Now map in the pages we've allocated...
1196 		 */
1197 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1198 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1199 		    HAT_LOAD_LOCK);
1200 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1201 
1202 		/* save away pfn and npages for FMA */
1203 		hp = mp->map_handlep;
1204 		if (hp) {
1205 			hp->ah_pfn = mmu_btop(pbase);
1206 			hp->ah_pnum = npages;
1207 		}
1208 	}
1209 
1210 #ifdef	DDI_MAP_DEBUG
1211 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1212 #endif	/* DDI_MAP_DEBUG */
1213 	return (DDI_SUCCESS);
1214 }
1215 
1216 
1217 /*
1218  * rootnex_unmap_regspec()
1219  *
1220  */
1221 static int
1222 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1223 {
1224 	caddr_t addr = (caddr_t)*vaddrp;
1225 	uint_t npages, pgoffset;
1226 	struct regspec *rp;
1227 
1228 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1229 		return (0);
1230 
1231 	rp = mp->map_obj.rp;
1232 
1233 	if (rp->regspec_size == 0) {
1234 #ifdef  DDI_MAP_DEBUG
1235 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1236 #endif  /* DDI_MAP_DEBUG */
1237 		return (DDI_ME_INVAL);
1238 	}
1239 
1240 	/*
1241 	 * I/O or memory mapping:
1242 	 *
1243 	 *	<bustype=0, addr=x, len=x>: memory
1244 	 *	<bustype=1, addr=x, len=x>: i/o
1245 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1246 	 */
1247 	if (rp->regspec_bustype != 0) {
1248 		/*
1249 		 * This is I/O space, which requires no particular
1250 		 * processing on unmap since it isn't mapped in the
1251 		 * first place.
1252 		 */
1253 		return (DDI_SUCCESS);
1254 	}
1255 
1256 	/*
1257 	 * Memory space
1258 	 */
1259 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1260 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1261 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1262 	device_arena_free(addr - pgoffset, ptob(npages));
1263 
1264 	/*
1265 	 * Destroy the pointer - the mapping has logically gone
1266 	 */
1267 	*vaddrp = NULL;
1268 
1269 	return (DDI_SUCCESS);
1270 }
1271 
1272 
1273 /*
1274  * rootnex_map_handle()
1275  *
1276  */
1277 static int
1278 rootnex_map_handle(ddi_map_req_t *mp)
1279 {
1280 	rootnex_addr_t rbase;
1281 	ddi_acc_hdl_t *hp;
1282 	uint_t pgoffset;
1283 	struct regspec *rp;
1284 	paddr_t pbase;
1285 
1286 	rp = mp->map_obj.rp;
1287 
1288 #ifdef	DDI_MAP_DEBUG
1289 	ddi_map_debug(
1290 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1291 	    rp->regspec_bustype, rp->regspec_addr,
1292 	    rp->regspec_size, mp->map_handlep);
1293 #endif	/* DDI_MAP_DEBUG */
1294 
1295 	/*
1296 	 * I/O or memory mapping:
1297 	 *
1298 	 *	<bustype=0, addr=x, len=x>: memory
1299 	 *	<bustype=1, addr=x, len=x>: i/o
1300 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1301 	 */
1302 	if (rp->regspec_bustype != 0) {
1303 		/*
1304 		 * This refers to I/O space, and we don't support "mapping"
1305 		 * I/O space to a user.
1306 		 */
1307 		return (DDI_FAILURE);
1308 	}
1309 
1310 	/*
1311 	 * Set up the hat_flags for the mapping.
1312 	 */
1313 	hp = mp->map_handlep;
1314 
1315 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1316 	case DDI_NEVERSWAP_ACC:
1317 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1318 		break;
1319 	case DDI_STRUCTURE_LE_ACC:
1320 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1321 		break;
1322 	case DDI_STRUCTURE_BE_ACC:
1323 		return (DDI_FAILURE);
1324 	default:
1325 		return (DDI_REGS_ACC_CONFLICT);
1326 	}
1327 
1328 	switch (hp->ah_acc.devacc_attr_dataorder) {
1329 	case DDI_STRICTORDER_ACC:
1330 		break;
1331 	case DDI_UNORDERED_OK_ACC:
1332 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1333 		break;
1334 	case DDI_MERGING_OK_ACC:
1335 		hp->ah_hat_flags |= HAT_MERGING_OK;
1336 		break;
1337 	case DDI_LOADCACHING_OK_ACC:
1338 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1339 		break;
1340 	case DDI_STORECACHING_OK_ACC:
1341 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1342 		break;
1343 	default:
1344 		return (DDI_FAILURE);
1345 	}
1346 
1347 	rbase = (rootnex_addr_t)rp->regspec_addr &
1348 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1349 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1350 
1351 	if (rp->regspec_size == 0)
1352 		return (DDI_ME_INVAL);
1353 
1354 #ifdef __xpv
1355 	/*
1356 	 * If we're dom0, we're using a real device so we need to translate
1357 	 * the MA to a PA.
1358 	 */
1359 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1360 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1361 		    (rbase & MMU_PAGEOFFSET);
1362 	} else {
1363 		pbase = rbase;
1364 	}
1365 #else
1366 	pbase = rbase;
1367 #endif
1368 
1369 	hp->ah_pfn = mmu_btop(pbase);
1370 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1371 
1372 	return (DDI_SUCCESS);
1373 }
1374 
1375 
1376 
1377 /*
1378  * ************************
1379  *  interrupt related code
1380  * ************************
1381  */
1382 
1383 /*
1384  * rootnex_intr_ops()
1385  *	bus_intr_op() function for interrupt support
1386  */
1387 /* ARGSUSED */
1388 static int
1389 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1390     ddi_intr_handle_impl_t *hdlp, void *result)
1391 {
1392 	struct intrspec			*ispec;
1393 
1394 	DDI_INTR_NEXDBG((CE_CONT,
1395 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1396 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1397 
1398 	/* Process the interrupt operation */
1399 	switch (intr_op) {
1400 	case DDI_INTROP_GETCAP:
1401 		/* First check with pcplusmp */
1402 		if (psm_intr_ops == NULL)
1403 			return (DDI_FAILURE);
1404 
1405 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1406 			*(int *)result = 0;
1407 			return (DDI_FAILURE);
1408 		}
1409 		break;
1410 	case DDI_INTROP_SETCAP:
1411 		if (psm_intr_ops == NULL)
1412 			return (DDI_FAILURE);
1413 
1414 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1415 			return (DDI_FAILURE);
1416 		break;
1417 	case DDI_INTROP_ALLOC:
1418 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1419 		return (rootnex_alloc_intr_fixed(rdip, hdlp, result));
1420 	case DDI_INTROP_FREE:
1421 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1422 		return (rootnex_free_intr_fixed(rdip, hdlp));
1423 	case DDI_INTROP_GETPRI:
1424 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1425 			return (DDI_FAILURE);
1426 		*(int *)result = ispec->intrspec_pri;
1427 		break;
1428 	case DDI_INTROP_SETPRI:
1429 		/* Validate the interrupt priority passed to us */
1430 		if (*(int *)result > LOCK_LEVEL)
1431 			return (DDI_FAILURE);
1432 
1433 		/* Ensure that PSM is all initialized and ispec is ok */
1434 		if ((psm_intr_ops == NULL) ||
1435 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1436 			return (DDI_FAILURE);
1437 
1438 		/* Change the priority */
1439 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1440 		    PSM_FAILURE)
1441 			return (DDI_FAILURE);
1442 
1443 		/* update the ispec with the new priority */
1444 		ispec->intrspec_pri =  *(int *)result;
1445 		break;
1446 	case DDI_INTROP_ADDISR:
1447 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1448 			return (DDI_FAILURE);
1449 		ispec->intrspec_func = hdlp->ih_cb_func;
1450 		break;
1451 	case DDI_INTROP_REMISR:
1452 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1453 			return (DDI_FAILURE);
1454 		ispec->intrspec_func = (uint_t (*)()) 0;
1455 		break;
1456 	case DDI_INTROP_ENABLE:
1457 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1458 			return (DDI_FAILURE);
1459 
1460 		/* Call psmi to translate irq with the dip */
1461 		if (psm_intr_ops == NULL)
1462 			return (DDI_FAILURE);
1463 
1464 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1465 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1466 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1467 			return (DDI_FAILURE);
1468 
1469 		/* Add the interrupt handler */
1470 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1471 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1472 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1473 			return (DDI_FAILURE);
1474 		break;
1475 	case DDI_INTROP_DISABLE:
1476 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1477 			return (DDI_FAILURE);
1478 
1479 		/* Call psm_ops() to translate irq with the dip */
1480 		if (psm_intr_ops == NULL)
1481 			return (DDI_FAILURE);
1482 
1483 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1484 		(void) (*psm_intr_ops)(rdip, hdlp,
1485 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1486 
1487 		/* Remove the interrupt handler */
1488 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1489 		    hdlp->ih_cb_func, hdlp->ih_vector);
1490 		break;
1491 	case DDI_INTROP_SETMASK:
1492 		if (psm_intr_ops == NULL)
1493 			return (DDI_FAILURE);
1494 
1495 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1496 			return (DDI_FAILURE);
1497 		break;
1498 	case DDI_INTROP_CLRMASK:
1499 		if (psm_intr_ops == NULL)
1500 			return (DDI_FAILURE);
1501 
1502 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1503 			return (DDI_FAILURE);
1504 		break;
1505 	case DDI_INTROP_GETPENDING:
1506 		if (psm_intr_ops == NULL)
1507 			return (DDI_FAILURE);
1508 
1509 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1510 		    result)) {
1511 			*(int *)result = 0;
1512 			return (DDI_FAILURE);
1513 		}
1514 		break;
1515 	case DDI_INTROP_NAVAIL:
1516 	case DDI_INTROP_NINTRS:
1517 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1518 		if (*(int *)result == 0) {
1519 			/*
1520 			 * Special case for 'pcic' driver' only. This driver
1521 			 * driver is a child of 'isa' and 'rootnex' drivers.
1522 			 *
1523 			 * See detailed comments on this in the function
1524 			 * rootnex_get_ispec().
1525 			 *
1526 			 * Children of 'pcic' send 'NINITR' request all the
1527 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1528 			 * field may not initialized. So, we fake it here
1529 			 * to return 1 (a la what PCMCIA nexus does).
1530 			 */
1531 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1532 				*(int *)result = 1;
1533 			else
1534 				return (DDI_FAILURE);
1535 		}
1536 		break;
1537 	case DDI_INTROP_SUPPORTED_TYPES:
1538 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1539 		break;
1540 	default:
1541 		return (DDI_FAILURE);
1542 	}
1543 
1544 	return (DDI_SUCCESS);
1545 }
1546 
1547 
1548 /*
1549  * rootnex_get_ispec()
1550  *	convert an interrupt number to an interrupt specification.
1551  *	The interrupt number determines which interrupt spec will be
1552  *	returned if more than one exists.
1553  *
1554  *	Look into the parent private data area of the 'rdip' to find out
1555  *	the interrupt specification.  First check to make sure there is
1556  *	one that matchs "inumber" and then return a pointer to it.
1557  *
1558  *	Return NULL if one could not be found.
1559  *
1560  *	NOTE: This is needed for rootnex_intr_ops()
1561  */
1562 static struct intrspec *
1563 rootnex_get_ispec(dev_info_t *rdip, int inum)
1564 {
1565 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1566 
1567 	/*
1568 	 * Special case handling for drivers that provide their own
1569 	 * intrspec structures instead of relying on the DDI framework.
1570 	 *
1571 	 * A broken hardware driver in ON could potentially provide its
1572 	 * own intrspec structure, instead of relying on the hardware.
1573 	 * If these drivers are children of 'rootnex' then we need to
1574 	 * continue to provide backward compatibility to them here.
1575 	 *
1576 	 * Following check is a special case for 'pcic' driver which
1577 	 * was found to have broken hardwre andby provides its own intrspec.
1578 	 *
1579 	 * Verbatim comments from this driver are shown here:
1580 	 * "Don't use the ddi_add_intr since we don't have a
1581 	 * default intrspec in all cases."
1582 	 *
1583 	 * Since an 'ispec' may not be always created for it,
1584 	 * check for that and create one if so.
1585 	 *
1586 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1587 	 */
1588 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1589 		pdp->par_nintr = 1;
1590 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1591 		    pdp->par_nintr, KM_SLEEP);
1592 	}
1593 
1594 	/* Validate the interrupt number */
1595 	if (inum >= pdp->par_nintr)
1596 		return (NULL);
1597 
1598 	/* Get the interrupt structure pointer and return that */
1599 	return ((struct intrspec *)&pdp->par_intr[inum]);
1600 }
1601 
1602 /*
1603  * Allocate interrupt vector for FIXED (legacy) type.
1604  */
1605 static int
1606 rootnex_alloc_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp,
1607     void *result)
1608 {
1609 	struct intrspec		*ispec;
1610 	ddi_intr_handle_impl_t	info_hdl;
1611 	int			ret;
1612 	int			free_phdl = 0;
1613 	apic_get_type_t		type_info;
1614 
1615 	if (psm_intr_ops == NULL)
1616 		return (DDI_FAILURE);
1617 
1618 	if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1619 		return (DDI_FAILURE);
1620 
1621 	/*
1622 	 * If the PSM module is "APIX" then pass the request for it
1623 	 * to allocate the vector now.
1624 	 */
1625 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1626 	info_hdl.ih_private = &type_info;
1627 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1628 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1629 		if (hdlp->ih_private == NULL) { /* allocate phdl structure */
1630 			free_phdl = 1;
1631 			i_ddi_alloc_intr_phdl(hdlp);
1632 		}
1633 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1634 		ret = (*psm_intr_ops)(rdip, hdlp,
1635 		    PSM_INTR_OP_ALLOC_VECTORS, result);
1636 		if (free_phdl) { /* free up the phdl structure */
1637 			free_phdl = 0;
1638 			i_ddi_free_intr_phdl(hdlp);
1639 			hdlp->ih_private = NULL;
1640 		}
1641 	} else {
1642 		/*
1643 		 * No APIX module; fall back to the old scheme where the
1644 		 * interrupt vector is allocated during ddi_enable_intr() call.
1645 		 */
1646 		hdlp->ih_pri = ispec->intrspec_pri;
1647 		*(int *)result = hdlp->ih_scratch1;
1648 		ret = DDI_SUCCESS;
1649 	}
1650 
1651 	return (ret);
1652 }
1653 
1654 /*
1655  * Free up interrupt vector for FIXED (legacy) type.
1656  */
1657 static int
1658 rootnex_free_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp)
1659 {
1660 	struct intrspec			*ispec;
1661 	struct ddi_parent_private_data	*pdp;
1662 	ddi_intr_handle_impl_t		info_hdl;
1663 	int				ret;
1664 	apic_get_type_t			type_info;
1665 
1666 	if (psm_intr_ops == NULL)
1667 		return (DDI_FAILURE);
1668 
1669 	/*
1670 	 * If the PSM module is "APIX" then pass the request for it
1671 	 * to free up the vector now.
1672 	 */
1673 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1674 	info_hdl.ih_private = &type_info;
1675 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1676 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1677 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1678 			return (DDI_FAILURE);
1679 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1680 		ret = (*psm_intr_ops)(rdip, hdlp,
1681 		    PSM_INTR_OP_FREE_VECTORS, NULL);
1682 	} else {
1683 		/*
1684 		 * No APIX module; fall back to the old scheme where
1685 		 * the interrupt vector was already freed during
1686 		 * ddi_disable_intr() call.
1687 		 */
1688 		ret = DDI_SUCCESS;
1689 	}
1690 
1691 	pdp = ddi_get_parent_data(rdip);
1692 
1693 	/*
1694 	 * Special case for 'pcic' driver' only.
1695 	 * If an intrspec was created for it, clean it up here
1696 	 * See detailed comments on this in the function
1697 	 * rootnex_get_ispec().
1698 	 */
1699 	if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1700 		kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1701 		    pdp->par_nintr);
1702 		/*
1703 		 * Set it to zero; so that
1704 		 * DDI framework doesn't free it again
1705 		 */
1706 		pdp->par_intr = NULL;
1707 		pdp->par_nintr = 0;
1708 	}
1709 
1710 	return (ret);
1711 }
1712 
1713 
1714 /*
1715  * ******************
1716  *  dma related code
1717  * ******************
1718  */
1719 
1720 /*ARGSUSED*/
1721 static int
1722 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1723     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1724     ddi_dma_handle_t *handlep)
1725 {
1726 	uint64_t maxsegmentsize_ll;
1727 	uint_t maxsegmentsize;
1728 	ddi_dma_impl_t *hp;
1729 	rootnex_dma_t *dma;
1730 	uint64_t count_max;
1731 	uint64_t seg;
1732 	int kmflag;
1733 	int e;
1734 
1735 
1736 	/* convert our sleep flags */
1737 	if (waitfp == DDI_DMA_SLEEP) {
1738 		kmflag = KM_SLEEP;
1739 	} else {
1740 		kmflag = KM_NOSLEEP;
1741 	}
1742 
1743 	/*
1744 	 * We try to do only one memory allocation here. We'll do a little
1745 	 * pointer manipulation later. If the bind ends up taking more than
1746 	 * our prealloc's space, we'll have to allocate more memory in the
1747 	 * bind operation. Not great, but much better than before and the
1748 	 * best we can do with the current bind interfaces.
1749 	 */
1750 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1751 	if (hp == NULL)
1752 		return (DDI_DMA_NORESOURCES);
1753 
1754 	/* Do our pointer manipulation now, align the structures */
1755 	hp->dmai_private = (void *)(((uintptr_t)hp +
1756 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1757 	dma = (rootnex_dma_t *)hp->dmai_private;
1758 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1759 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1760 
1761 	/* setup the handle */
1762 	rootnex_clean_dmahdl(hp);
1763 	hp->dmai_error.err_fep = NULL;
1764 	hp->dmai_error.err_cf = NULL;
1765 	dma->dp_dip = rdip;
1766 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1767 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1768 
1769 	/*
1770 	 * The BOUNCE_ON_SEG workaround is not needed when an IOMMU
1771 	 * is being used. Set the upper limit to the seg value.
1772 	 * There will be enough DVMA space to always get addresses
1773 	 * that will match the constraints.
1774 	 */
1775 	if (IOMMU_USED(rdip) &&
1776 	    (attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG)) {
1777 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_seg;
1778 		dma->dp_sglinfo.si_flags &= ~_DDI_DMA_BOUNCE_ON_SEG;
1779 	} else
1780 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1781 
1782 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1783 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1784 	hp->dmai_rdip = rdip;
1785 	hp->dmai_attr = *attr;
1786 
1787 	if (attr->dma_attr_seg >= dma->dp_sglinfo.si_max_addr)
1788 		dma->dp_sglinfo.si_cancross = B_FALSE;
1789 	else
1790 		dma->dp_sglinfo.si_cancross = B_TRUE;
1791 
1792 	/* we don't need to worry about the SPL since we do a tryenter */
1793 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1794 
1795 	/*
1796 	 * Figure out our maximum segment size. If the segment size is greater
1797 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1798 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1799 	 * dma_attr_count_max are size-1 type values.
1800 	 *
1801 	 * Maximum segment size is the largest physically contiguous chunk of
1802 	 * memory that we can return from a bind (i.e. the maximum size of a
1803 	 * single cookie).
1804 	 */
1805 
1806 	/* handle the rollover cases */
1807 	seg = attr->dma_attr_seg + 1;
1808 	if (seg < attr->dma_attr_seg) {
1809 		seg = attr->dma_attr_seg;
1810 	}
1811 	count_max = attr->dma_attr_count_max + 1;
1812 	if (count_max < attr->dma_attr_count_max) {
1813 		count_max = attr->dma_attr_count_max;
1814 	}
1815 
1816 	/*
1817 	 * granularity may or may not be a power of two. If it isn't, we can't
1818 	 * use a simple mask.
1819 	 */
1820 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1821 		dma->dp_granularity_power_2 = B_FALSE;
1822 	} else {
1823 		dma->dp_granularity_power_2 = B_TRUE;
1824 	}
1825 
1826 	/*
1827 	 * maxxfer should be a whole multiple of granularity. If we're going to
1828 	 * break up a window because we're greater than maxxfer, we might as
1829 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1830 	 * worry about triming the window later on for this case.
1831 	 */
1832 	if (attr->dma_attr_granular > 1) {
1833 		if (dma->dp_granularity_power_2) {
1834 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1835 			    (attr->dma_attr_maxxfer &
1836 			    (attr->dma_attr_granular - 1));
1837 		} else {
1838 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1839 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1840 		}
1841 	} else {
1842 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1843 	}
1844 
1845 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1846 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1847 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1848 		maxsegmentsize = 0xFFFFFFFF;
1849 	} else {
1850 		maxsegmentsize = maxsegmentsize_ll;
1851 	}
1852 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1853 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1854 
1855 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1856 	if (rootnex_alloc_check_parms) {
1857 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1858 		if (e != DDI_SUCCESS) {
1859 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1860 			(void) rootnex_dma_freehdl(dip, rdip,
1861 			    (ddi_dma_handle_t)hp);
1862 			return (e);
1863 		}
1864 	}
1865 
1866 	*handlep = (ddi_dma_handle_t)hp;
1867 
1868 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1869 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1870 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1871 
1872 	return (DDI_SUCCESS);
1873 }
1874 
1875 
1876 /*
1877  * rootnex_dma_allochdl()
1878  *    called from ddi_dma_alloc_handle().
1879  */
1880 static int
1881 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1882     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1883 {
1884 	int retval = DDI_SUCCESS;
1885 #if defined(__amd64) && !defined(__xpv)
1886 
1887 	if (IOMMU_UNITIALIZED(rdip)) {
1888 		retval = iommulib_nex_open(dip, rdip);
1889 
1890 		if (retval != DDI_SUCCESS && retval != DDI_ENOTSUP)
1891 			return (retval);
1892 	}
1893 
1894 	if (IOMMU_UNUSED(rdip)) {
1895 		retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1896 		    handlep);
1897 	} else {
1898 		retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1899 		    waitfp, arg, handlep);
1900 	}
1901 #else
1902 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1903 	    handlep);
1904 #endif
1905 	switch (retval) {
1906 	case DDI_DMA_NORESOURCES:
1907 		if (waitfp != DDI_DMA_DONTWAIT) {
1908 			ddi_set_callback(waitfp, arg,
1909 			    &rootnex_state->r_dvma_call_list_id);
1910 		}
1911 		break;
1912 	case DDI_SUCCESS:
1913 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1914 		break;
1915 	default:
1916 		break;
1917 	}
1918 	return (retval);
1919 }
1920 
1921 /*ARGSUSED*/
1922 static int
1923 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1924     ddi_dma_handle_t handle)
1925 {
1926 	ddi_dma_impl_t *hp;
1927 	rootnex_dma_t *dma;
1928 
1929 
1930 	hp = (ddi_dma_impl_t *)handle;
1931 	dma = (rootnex_dma_t *)hp->dmai_private;
1932 
1933 	/* unbind should have been called first */
1934 	ASSERT(!dma->dp_inuse);
1935 
1936 	mutex_destroy(&dma->dp_mutex);
1937 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1938 
1939 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1940 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1941 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1942 
1943 	return (DDI_SUCCESS);
1944 }
1945 
1946 /*
1947  * rootnex_dma_freehdl()
1948  *    called from ddi_dma_free_handle().
1949  */
1950 static int
1951 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1952 {
1953 	int ret;
1954 
1955 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1956 #if defined(__amd64) && !defined(__xpv)
1957 	if (IOMMU_USED(rdip))
1958 		ret = iommulib_nexdma_freehdl(dip, rdip, handle);
1959 	else
1960 #endif
1961 	ret = rootnex_coredma_freehdl(dip, rdip, handle);
1962 
1963 	if (rootnex_state->r_dvma_call_list_id)
1964 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1965 
1966 	return (ret);
1967 }
1968 
1969 /*ARGSUSED*/
1970 static int
1971 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1972     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1973     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1974 {
1975 	rootnex_sglinfo_t *sinfo;
1976 	ddi_dma_obj_t *dmao;
1977 #if defined(__amd64) && !defined(__xpv)
1978 	struct dvmaseg *dvs;
1979 	ddi_dma_cookie_t *cookie;
1980 #endif
1981 	ddi_dma_attr_t *attr;
1982 	ddi_dma_impl_t *hp;
1983 	rootnex_dma_t *dma;
1984 	int kmflag;
1985 	int e;
1986 	uint_t ncookies;
1987 
1988 	hp = (ddi_dma_impl_t *)handle;
1989 	dma = (rootnex_dma_t *)hp->dmai_private;
1990 	dmao = &dma->dp_dma;
1991 	sinfo = &dma->dp_sglinfo;
1992 	attr = &hp->dmai_attr;
1993 
1994 	/* convert the sleep flags */
1995 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1996 		dma->dp_sleep_flags = kmflag = KM_SLEEP;
1997 	} else {
1998 		dma->dp_sleep_flags = kmflag = KM_NOSLEEP;
1999 	}
2000 
2001 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
2002 
2003 	/*
2004 	 * This is useful for debugging a driver. Not as useful in a production
2005 	 * system. The only time this will fail is if you have a driver bug.
2006 	 */
2007 	if (rootnex_bind_check_inuse) {
2008 		/*
2009 		 * No one else should ever have this lock unless someone else
2010 		 * is trying to use this handle. So contention on the lock
2011 		 * is the same as inuse being set.
2012 		 */
2013 		e = mutex_tryenter(&dma->dp_mutex);
2014 		if (e == 0) {
2015 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2016 			return (DDI_DMA_INUSE);
2017 		}
2018 		if (dma->dp_inuse) {
2019 			mutex_exit(&dma->dp_mutex);
2020 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2021 			return (DDI_DMA_INUSE);
2022 		}
2023 		dma->dp_inuse = B_TRUE;
2024 		mutex_exit(&dma->dp_mutex);
2025 	}
2026 
2027 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
2028 	if (rootnex_bind_check_parms) {
2029 		e = rootnex_valid_bind_parms(dmareq, attr);
2030 		if (e != DDI_SUCCESS) {
2031 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2032 			rootnex_clean_dmahdl(hp);
2033 			return (e);
2034 		}
2035 	}
2036 
2037 	/* save away the original bind info */
2038 	dma->dp_dma = dmareq->dmar_object;
2039 
2040 #if defined(__amd64) && !defined(__xpv)
2041 	if (IOMMU_USED(rdip)) {
2042 		dmao = &dma->dp_dvma;
2043 		e = iommulib_nexdma_mapobject(dip, rdip, handle, dmareq, dmao);
2044 		switch (e) {
2045 		case DDI_SUCCESS:
2046 			if (sinfo->si_cancross ||
2047 			    dmao->dmao_obj.dvma_obj.dv_nseg != 1 ||
2048 			    dmao->dmao_size > sinfo->si_max_cookie_size) {
2049 				dma->dp_dvma_used = B_TRUE;
2050 				break;
2051 			}
2052 			sinfo->si_sgl_size = 1;
2053 			hp->dmai_rflags |= DMP_NOSYNC;
2054 
2055 			dma->dp_dvma_used = B_TRUE;
2056 			dma->dp_need_to_free_cookie = B_FALSE;
2057 
2058 			dvs = &dmao->dmao_obj.dvma_obj.dv_seg[0];
2059 			cookie = hp->dmai_cookie = dma->dp_cookies =
2060 			    (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2061 			cookie->dmac_laddress = dvs->dvs_start +
2062 			    dmao->dmao_obj.dvma_obj.dv_off;
2063 			cookie->dmac_size = dvs->dvs_len;
2064 			cookie->dmac_type = 0;
2065 
2066 			ROOTNEX_DPROBE1(rootnex__bind__dvmafast, dev_info_t *,
2067 			    rdip);
2068 			goto fast;
2069 		case DDI_ENOTSUP:
2070 			break;
2071 		default:
2072 			rootnex_clean_dmahdl(hp);
2073 			return (e);
2074 		}
2075 	}
2076 #endif
2077 
2078 	/*
2079 	 * Figure out a rough estimate of what maximum number of pages
2080 	 * this buffer could use (a high estimate of course).
2081 	 */
2082 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
2083 
2084 	if (dma->dp_dvma_used) {
2085 		/*
2086 		 * The number of physical pages is the worst case.
2087 		 *
2088 		 * For DVMA, the worst case is the length divided
2089 		 * by the maximum cookie length, plus 1. Add to that
2090 		 * the number of segment boundaries potentially crossed, and
2091 		 * the additional number of DVMA segments that was returned.
2092 		 *
2093 		 * In the normal case, for modern devices, si_cancross will
2094 		 * be false, and dv_nseg will be 1, and the fast path will
2095 		 * have been taken above.
2096 		 */
2097 		ncookies = (dma->dp_dma.dmao_size / sinfo->si_max_cookie_size)
2098 		    + 1;
2099 		if (sinfo->si_cancross)
2100 			ncookies +=
2101 			    (dma->dp_dma.dmao_size / attr->dma_attr_seg) + 1;
2102 		ncookies += (dmao->dmao_obj.dvma_obj.dv_nseg - 1);
2103 
2104 		sinfo->si_max_pages = MIN(sinfo->si_max_pages, ncookies);
2105 	}
2106 
2107 	/*
2108 	 * We'll use the pre-allocated cookies for any bind that will *always*
2109 	 * fit (more important to be consistent, we don't want to create
2110 	 * additional degenerate cases).
2111 	 */
2112 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
2113 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2114 		dma->dp_need_to_free_cookie = B_FALSE;
2115 		ROOTNEX_DPROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
2116 		    uint_t, sinfo->si_max_pages);
2117 
2118 	/*
2119 	 * For anything larger than that, we'll go ahead and allocate the
2120 	 * maximum number of pages we expect to see. Hopefuly, we won't be
2121 	 * seeing this path in the fast path for high performance devices very
2122 	 * frequently.
2123 	 *
2124 	 * a ddi bind interface that allowed the driver to provide storage to
2125 	 * the bind interface would speed this case up.
2126 	 */
2127 	} else {
2128 		/*
2129 		 * Save away how much memory we allocated. If we're doing a
2130 		 * nosleep, the alloc could fail...
2131 		 */
2132 		dma->dp_cookie_size = sinfo->si_max_pages *
2133 		    sizeof (ddi_dma_cookie_t);
2134 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
2135 		if (dma->dp_cookies == NULL) {
2136 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2137 			rootnex_clean_dmahdl(hp);
2138 			return (DDI_DMA_NORESOURCES);
2139 		}
2140 		dma->dp_need_to_free_cookie = B_TRUE;
2141 		ROOTNEX_DPROBE2(rootnex__bind__alloc, dev_info_t *, rdip,
2142 		    uint_t, sinfo->si_max_pages);
2143 	}
2144 	hp->dmai_cookie = dma->dp_cookies;
2145 
2146 	/*
2147 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
2148 	 * looking at the constraints in the dma structure. It will then put
2149 	 * some additional state about the sgl in the dma struct (i.e. is
2150 	 * the sgl clean, or do we need to do some munging; how many pages
2151 	 * need to be copied, etc.)
2152 	 */
2153 	if (dma->dp_dvma_used)
2154 		rootnex_dvma_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2155 	else
2156 		rootnex_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2157 
2158 out:
2159 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
2160 	/* if we don't need a copy buffer, we don't need to sync */
2161 	if (sinfo->si_copybuf_req == 0) {
2162 		hp->dmai_rflags |= DMP_NOSYNC;
2163 	}
2164 
2165 	/*
2166 	 * if we don't need the copybuf and we don't need to do a partial,  we
2167 	 * hit the fast path. All the high performance devices should be trying
2168 	 * to hit this path. To hit this path, a device should be able to reach
2169 	 * all of memory, shouldn't try to bind more than it can transfer, and
2170 	 * the buffer shouldn't require more cookies than the driver/device can
2171 	 * handle [sgllen]).
2172 	 *
2173 	 * Note that negative values of dma_attr_sgllen are supposed
2174 	 * to mean unlimited, but we just cast them to mean a
2175 	 * "ridiculous large limit".  This saves some extra checks on
2176 	 * hot paths.
2177 	 */
2178 	if ((sinfo->si_copybuf_req == 0) &&
2179 	    (sinfo->si_sgl_size <= (unsigned)attr->dma_attr_sgllen) &&
2180 	    (dmao->dmao_size < dma->dp_maxxfer)) {
2181 fast:
2182 		/*
2183 		 * If the driver supports FMA, insert the handle in the FMA DMA
2184 		 * handle cache.
2185 		 */
2186 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2187 			hp->dmai_error.err_cf = rootnex_dma_check;
2188 
2189 		/*
2190 		 * copy out the first cookie and ccountp, set the cookie
2191 		 * pointer to the second cookie. The first cookie is passed
2192 		 * back on the stack. Additional cookies are accessed via
2193 		 * ddi_dma_nextcookie()
2194 		 */
2195 		*cookiep = dma->dp_cookies[0];
2196 		*ccountp = sinfo->si_sgl_size;
2197 		hp->dmai_cookie++;
2198 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2199 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2200 		ROOTNEX_DPROBE4(rootnex__bind__fast, dev_info_t *, rdip,
2201 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2202 		    uint_t, dmao->dmao_size, uint_t, *ccountp);
2203 
2204 
2205 		return (DDI_DMA_MAPPED);
2206 	}
2207 
2208 	/*
2209 	 * go to the slow path, we may need to alloc more memory, create
2210 	 * multiple windows, and munge up a sgl to make the device happy.
2211 	 */
2212 
2213 	/*
2214 	 * With the IOMMU mapobject method used, we should never hit
2215 	 * the slow path. If we do, something is seriously wrong.
2216 	 * Clean up and return an error.
2217 	 */
2218 
2219 #if defined(__amd64) && !defined(__xpv)
2220 
2221 	if (dma->dp_dvma_used) {
2222 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2223 		    &dma->dp_dvma);
2224 		e = DDI_DMA_NOMAPPING;
2225 	} else {
2226 #endif
2227 		e = rootnex_bind_slowpath(hp, dmareq, dma, attr, &dma->dp_dma,
2228 		    kmflag);
2229 #if defined(__amd64) && !defined(__xpv)
2230 	}
2231 #endif
2232 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2233 		if (dma->dp_need_to_free_cookie) {
2234 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2235 		}
2236 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2237 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2238 		return (e);
2239 	}
2240 
2241 	/*
2242 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2243 	 * cache.
2244 	 */
2245 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2246 		hp->dmai_error.err_cf = rootnex_dma_check;
2247 
2248 	/* if the first window uses the copy buffer, sync it for the device */
2249 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2250 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2251 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2252 		    DDI_DMA_SYNC_FORDEV);
2253 	}
2254 
2255 	/*
2256 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2257 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2258 	 * If we have a partial map (i.e. multiple windows), the number of
2259 	 * cookies we return is the number of cookies in the first window.
2260 	 */
2261 	if (e == DDI_DMA_MAPPED) {
2262 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2263 		*ccountp = sinfo->si_sgl_size;
2264 		hp->dmai_nwin = 1;
2265 	} else {
2266 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2267 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2268 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2269 	}
2270 	*cookiep = dma->dp_cookies[0];
2271 	hp->dmai_cookie++;
2272 
2273 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2274 	ROOTNEX_DPROBE4(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2275 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2276 	    dmao->dmao_size, uint_t, *ccountp);
2277 	return (e);
2278 }
2279 
2280 /*
2281  * rootnex_dma_bindhdl()
2282  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2283  */
2284 static int
2285 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2286     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2287     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2288 {
2289 	int ret;
2290 #if defined(__amd64) && !defined(__xpv)
2291 	if (IOMMU_USED(rdip))
2292 		ret = iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2293 		    cookiep, ccountp);
2294 	else
2295 #endif
2296 	ret = rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2297 	    cookiep, ccountp);
2298 
2299 	if (ret == DDI_DMA_NORESOURCES && dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
2300 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
2301 		    &rootnex_state->r_dvma_call_list_id);
2302 	}
2303 
2304 	return (ret);
2305 }
2306 
2307 
2308 
2309 /*ARGSUSED*/
2310 static int
2311 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2312     ddi_dma_handle_t handle)
2313 {
2314 	ddi_dma_impl_t *hp;
2315 	rootnex_dma_t *dma;
2316 	int e;
2317 
2318 	hp = (ddi_dma_impl_t *)handle;
2319 	dma = (rootnex_dma_t *)hp->dmai_private;
2320 
2321 	/* make sure the buffer wasn't free'd before calling unbind */
2322 	if (rootnex_unbind_verify_buffer) {
2323 		e = rootnex_verify_buffer(dma);
2324 		if (e != DDI_SUCCESS) {
2325 			ASSERT(0);
2326 			return (DDI_FAILURE);
2327 		}
2328 	}
2329 
2330 	/* sync the current window before unbinding the buffer */
2331 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2332 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2333 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2334 		    DDI_DMA_SYNC_FORCPU);
2335 	}
2336 
2337 	/*
2338 	 * cleanup and copy buffer or window state. if we didn't use the copy
2339 	 * buffer or windows, there won't be much to do :-)
2340 	 */
2341 	rootnex_teardown_copybuf(dma);
2342 	rootnex_teardown_windows(dma);
2343 
2344 #if defined(__amd64) && !defined(__xpv)
2345 	if (IOMMU_USED(rdip))
2346 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2347 		    &dma->dp_dvma);
2348 #endif
2349 
2350 	/*
2351 	 * If we had to allocate space to for the worse case sgl (it didn't
2352 	 * fit into our pre-allocate buffer), free that up now
2353 	 */
2354 	if (dma->dp_need_to_free_cookie) {
2355 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2356 	}
2357 
2358 	/*
2359 	 * clean up the handle so it's ready for the next bind (i.e. if the
2360 	 * handle is reused).
2361 	 */
2362 	rootnex_clean_dmahdl(hp);
2363 	hp->dmai_error.err_cf = NULL;
2364 
2365 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2366 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2367 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2368 
2369 	return (DDI_SUCCESS);
2370 }
2371 
2372 /*
2373  * rootnex_dma_unbindhdl()
2374  *    called from ddi_dma_unbind_handle()
2375  */
2376 /*ARGSUSED*/
2377 static int
2378 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2379     ddi_dma_handle_t handle)
2380 {
2381 	int ret;
2382 
2383 #if defined(__amd64) && !defined(__xpv)
2384 	if (IOMMU_USED(rdip))
2385 		ret = iommulib_nexdma_unbindhdl(dip, rdip, handle);
2386 	else
2387 #endif
2388 	ret = rootnex_coredma_unbindhdl(dip, rdip, handle);
2389 
2390 	if (rootnex_state->r_dvma_call_list_id)
2391 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2392 
2393 	return (ret);
2394 }
2395 
2396 #if defined(__amd64) && !defined(__xpv)
2397 
2398 static int
2399 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2400 {
2401 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2402 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2403 
2404 	if (dma->dp_sleep_flags != KM_SLEEP &&
2405 	    dma->dp_sleep_flags != KM_NOSLEEP)
2406 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2407 	return (dma->dp_sleep_flags);
2408 }
2409 /*ARGSUSED*/
2410 static void
2411 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2412 {
2413 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2414 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2415 	rootnex_window_t *window;
2416 
2417 	if (dma->dp_window) {
2418 		window = &dma->dp_window[dma->dp_current_win];
2419 		hp->dmai_cookie = window->wd_first_cookie;
2420 	} else {
2421 		hp->dmai_cookie = dma->dp_cookies;
2422 	}
2423 	hp->dmai_cookie++;
2424 }
2425 
2426 /*ARGSUSED*/
2427 static int
2428 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2429     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2430 {
2431 	int i;
2432 	int km_flags;
2433 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2434 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2435 	rootnex_window_t *window;
2436 	ddi_dma_cookie_t *cp;
2437 	ddi_dma_cookie_t *cookie;
2438 
2439 	ASSERT(*cookiepp == NULL);
2440 	ASSERT(*ccountp == 0);
2441 
2442 	if (dma->dp_window) {
2443 		window = &dma->dp_window[dma->dp_current_win];
2444 		cp = window->wd_first_cookie;
2445 		*ccountp = window->wd_cookie_cnt;
2446 	} else {
2447 		cp = dma->dp_cookies;
2448 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2449 	}
2450 
2451 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2452 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2453 	if (cookie == NULL) {
2454 		return (DDI_DMA_NORESOURCES);
2455 	}
2456 
2457 	for (i = 0; i < *ccountp; i++) {
2458 		cookie[i].dmac_notused = cp[i].dmac_notused;
2459 		cookie[i].dmac_type = cp[i].dmac_type;
2460 		cookie[i].dmac_address = cp[i].dmac_address;
2461 		cookie[i].dmac_size = cp[i].dmac_size;
2462 	}
2463 
2464 	*cookiepp = cookie;
2465 
2466 	return (DDI_SUCCESS);
2467 }
2468 
2469 /*ARGSUSED*/
2470 static int
2471 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2472     ddi_dma_cookie_t *cookiep, uint_t ccount)
2473 {
2474 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2475 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2476 	rootnex_window_t *window;
2477 	ddi_dma_cookie_t *cur_cookiep;
2478 
2479 	ASSERT(cookiep);
2480 	ASSERT(ccount != 0);
2481 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2482 
2483 	if (dma->dp_window) {
2484 		window = &dma->dp_window[dma->dp_current_win];
2485 		dma->dp_saved_cookies = window->wd_first_cookie;
2486 		window->wd_first_cookie = cookiep;
2487 		ASSERT(ccount == window->wd_cookie_cnt);
2488 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2489 		    + window->wd_first_cookie;
2490 	} else {
2491 		dma->dp_saved_cookies = dma->dp_cookies;
2492 		dma->dp_cookies = cookiep;
2493 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2494 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2495 		    + dma->dp_cookies;
2496 	}
2497 
2498 	dma->dp_need_to_switch_cookies = B_TRUE;
2499 	hp->dmai_cookie = cur_cookiep;
2500 
2501 	return (DDI_SUCCESS);
2502 }
2503 
2504 /*ARGSUSED*/
2505 static int
2506 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2507 {
2508 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2509 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2510 	rootnex_window_t *window;
2511 	ddi_dma_cookie_t *cur_cookiep;
2512 	ddi_dma_cookie_t *cookie_array;
2513 	uint_t ccount;
2514 
2515 	/* check if cookies have not been switched */
2516 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2517 		return (DDI_SUCCESS);
2518 
2519 	ASSERT(dma->dp_saved_cookies);
2520 
2521 	if (dma->dp_window) {
2522 		window = &dma->dp_window[dma->dp_current_win];
2523 		cookie_array = window->wd_first_cookie;
2524 		window->wd_first_cookie = dma->dp_saved_cookies;
2525 		dma->dp_saved_cookies = NULL;
2526 		ccount = window->wd_cookie_cnt;
2527 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2528 		    + window->wd_first_cookie;
2529 	} else {
2530 		cookie_array = dma->dp_cookies;
2531 		dma->dp_cookies = dma->dp_saved_cookies;
2532 		dma->dp_saved_cookies = NULL;
2533 		ccount = dma->dp_sglinfo.si_sgl_size;
2534 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2535 		    + dma->dp_cookies;
2536 	}
2537 
2538 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2539 
2540 	hp->dmai_cookie = cur_cookiep;
2541 
2542 	dma->dp_need_to_switch_cookies = B_FALSE;
2543 
2544 	return (DDI_SUCCESS);
2545 }
2546 
2547 #endif
2548 
2549 static struct as *
2550 rootnex_get_as(ddi_dma_obj_t *dmao)
2551 {
2552 	struct as *asp;
2553 
2554 	switch (dmao->dmao_type) {
2555 	case DMA_OTYP_VADDR:
2556 	case DMA_OTYP_BUFVADDR:
2557 		asp = dmao->dmao_obj.virt_obj.v_as;
2558 		if (asp == NULL)
2559 			asp = &kas;
2560 		break;
2561 	default:
2562 		asp = NULL;
2563 		break;
2564 	}
2565 	return (asp);
2566 }
2567 
2568 /*
2569  * rootnex_verify_buffer()
2570  *   verify buffer wasn't free'd
2571  */
2572 static int
2573 rootnex_verify_buffer(rootnex_dma_t *dma)
2574 {
2575 	page_t **pplist;
2576 	caddr_t vaddr;
2577 	uint_t pcnt;
2578 	uint_t poff;
2579 	page_t *pp;
2580 	char b;
2581 	int i;
2582 
2583 	/* Figure out how many pages this buffer occupies */
2584 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2585 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2586 	} else {
2587 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2588 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2589 	}
2590 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2591 
2592 	switch (dma->dp_dma.dmao_type) {
2593 	case DMA_OTYP_PAGES:
2594 		/*
2595 		 * for a linked list of pp's walk through them to make sure
2596 		 * they're locked and not free.
2597 		 */
2598 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2599 		for (i = 0; i < pcnt; i++) {
2600 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2601 				return (DDI_FAILURE);
2602 			}
2603 			pp = pp->p_next;
2604 		}
2605 		break;
2606 
2607 	case DMA_OTYP_VADDR:
2608 	case DMA_OTYP_BUFVADDR:
2609 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2610 		/*
2611 		 * for an array of pp's walk through them to make sure they're
2612 		 * not free. It's possible that they may not be locked.
2613 		 */
2614 		if (pplist) {
2615 			for (i = 0; i < pcnt; i++) {
2616 				if (PP_ISFREE(pplist[i])) {
2617 					return (DDI_FAILURE);
2618 				}
2619 			}
2620 
2621 		/* For a virtual address, try to peek at each page */
2622 		} else {
2623 			if (rootnex_get_as(&dma->dp_dma) == &kas) {
2624 				for (i = 0; i < pcnt; i++) {
2625 					if (ddi_peek8(NULL, vaddr, &b) ==
2626 					    DDI_FAILURE)
2627 						return (DDI_FAILURE);
2628 					vaddr += MMU_PAGESIZE;
2629 				}
2630 			}
2631 		}
2632 		break;
2633 
2634 	default:
2635 		cmn_err(CE_PANIC, "rootnex_verify_buffer: bad DMA object");
2636 		break;
2637 	}
2638 
2639 	return (DDI_SUCCESS);
2640 }
2641 
2642 
2643 /*
2644  * rootnex_clean_dmahdl()
2645  *    Clean the dma handle. This should be called on a handle alloc and an
2646  *    unbind handle. Set the handle state to the default settings.
2647  */
2648 static void
2649 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2650 {
2651 	rootnex_dma_t *dma;
2652 
2653 
2654 	dma = (rootnex_dma_t *)hp->dmai_private;
2655 
2656 	hp->dmai_nwin = 0;
2657 	dma->dp_current_cookie = 0;
2658 	dma->dp_copybuf_size = 0;
2659 	dma->dp_window = NULL;
2660 	dma->dp_cbaddr = NULL;
2661 	dma->dp_inuse = B_FALSE;
2662 	dma->dp_dvma_used = B_FALSE;
2663 	dma->dp_need_to_free_cookie = B_FALSE;
2664 	dma->dp_need_to_switch_cookies = B_FALSE;
2665 	dma->dp_saved_cookies = NULL;
2666 	dma->dp_sleep_flags = KM_PANIC;
2667 	dma->dp_need_to_free_window = B_FALSE;
2668 	dma->dp_partial_required = B_FALSE;
2669 	dma->dp_trim_required = B_FALSE;
2670 	dma->dp_sglinfo.si_copybuf_req = 0;
2671 #if !defined(__amd64)
2672 	dma->dp_cb_remaping = B_FALSE;
2673 	dma->dp_kva = NULL;
2674 #endif
2675 
2676 	/* FMA related initialization */
2677 	hp->dmai_fault = 0;
2678 	hp->dmai_fault_check = NULL;
2679 	hp->dmai_fault_notify = NULL;
2680 	hp->dmai_error.err_ena = 0;
2681 	hp->dmai_error.err_status = DDI_FM_OK;
2682 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2683 	hp->dmai_error.err_ontrap = NULL;
2684 }
2685 
2686 
2687 /*
2688  * rootnex_valid_alloc_parms()
2689  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2690  */
2691 static int
2692 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2693 {
2694 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2695 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2696 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2697 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2698 		return (DDI_DMA_BADATTR);
2699 	}
2700 
2701 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2702 		return (DDI_DMA_BADATTR);
2703 	}
2704 
2705 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2706 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2707 	    attr->dma_attr_sgllen == 0) {
2708 		return (DDI_DMA_BADATTR);
2709 	}
2710 
2711 	/* We should be able to DMA into every byte offset in a page */
2712 	if (maxsegmentsize < MMU_PAGESIZE) {
2713 		return (DDI_DMA_BADATTR);
2714 	}
2715 
2716 	/* if we're bouncing on seg, seg must be <= addr_hi */
2717 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2718 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2719 		return (DDI_DMA_BADATTR);
2720 	}
2721 	return (DDI_SUCCESS);
2722 }
2723 
2724 /*
2725  * rootnex_valid_bind_parms()
2726  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2727  */
2728 /* ARGSUSED */
2729 static int
2730 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2731 {
2732 #if !defined(__amd64)
2733 	/*
2734 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2735 	 * we can track the offset for the obsoleted interfaces.
2736 	 */
2737 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2738 		return (DDI_DMA_TOOBIG);
2739 	}
2740 #endif
2741 
2742 	return (DDI_SUCCESS);
2743 }
2744 
2745 
2746 /*
2747  * rootnex_need_bounce_seg()
2748  *    check to see if the buffer lives on both side of the seg.
2749  */
2750 static boolean_t
2751 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2752 {
2753 	ddi_dma_atyp_t buftype;
2754 	rootnex_addr_t raddr;
2755 	boolean_t lower_addr;
2756 	boolean_t upper_addr;
2757 	uint64_t offset;
2758 	page_t **pplist;
2759 	uint64_t paddr;
2760 	uint32_t psize;
2761 	uint32_t size;
2762 	caddr_t vaddr;
2763 	uint_t pcnt;
2764 	page_t *pp;
2765 
2766 
2767 	/* shortcuts */
2768 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2769 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2770 	buftype = dmar_object->dmao_type;
2771 	size = dmar_object->dmao_size;
2772 
2773 	lower_addr = B_FALSE;
2774 	upper_addr = B_FALSE;
2775 	pcnt = 0;
2776 
2777 	/*
2778 	 * Process the first page to handle the initial offset of the buffer.
2779 	 * We'll use the base address we get later when we loop through all
2780 	 * the pages.
2781 	 */
2782 	if (buftype == DMA_OTYP_PAGES) {
2783 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2784 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2785 		    MMU_PAGEOFFSET;
2786 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2787 		psize = MIN(size, (MMU_PAGESIZE - offset));
2788 		pp = pp->p_next;
2789 		sglinfo->si_asp = NULL;
2790 	} else if (pplist != NULL) {
2791 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2792 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2793 		if (sglinfo->si_asp == NULL) {
2794 			sglinfo->si_asp = &kas;
2795 		}
2796 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2797 		paddr += offset;
2798 		psize = MIN(size, (MMU_PAGESIZE - offset));
2799 		pcnt++;
2800 	} else {
2801 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2802 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2803 		if (sglinfo->si_asp == NULL) {
2804 			sglinfo->si_asp = &kas;
2805 		}
2806 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2807 		paddr += offset;
2808 		psize = MIN(size, (MMU_PAGESIZE - offset));
2809 		vaddr += psize;
2810 	}
2811 
2812 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2813 
2814 	if ((raddr + psize) > sglinfo->si_segmask) {
2815 		upper_addr = B_TRUE;
2816 	} else {
2817 		lower_addr = B_TRUE;
2818 	}
2819 	size -= psize;
2820 
2821 	/*
2822 	 * Walk through the rest of the pages in the buffer. Track to see
2823 	 * if we have pages on both sides of the segment boundary.
2824 	 */
2825 	while (size > 0) {
2826 		/* partial or full page */
2827 		psize = MIN(size, MMU_PAGESIZE);
2828 
2829 		if (buftype == DMA_OTYP_PAGES) {
2830 			/* get the paddr from the page_t */
2831 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2832 			paddr = pfn_to_pa(pp->p_pagenum);
2833 			pp = pp->p_next;
2834 		} else if (pplist != NULL) {
2835 			/* index into the array of page_t's to get the paddr */
2836 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2837 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2838 			pcnt++;
2839 		} else {
2840 			/* call into the VM to get the paddr */
2841 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2842 			    vaddr));
2843 			vaddr += psize;
2844 		}
2845 
2846 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2847 
2848 		if ((raddr + psize) > sglinfo->si_segmask) {
2849 			upper_addr = B_TRUE;
2850 		} else {
2851 			lower_addr = B_TRUE;
2852 		}
2853 		/*
2854 		 * if the buffer lives both above and below the segment
2855 		 * boundary, or the current page is the page immediately
2856 		 * after the segment, we will use a copy/bounce buffer for
2857 		 * all pages > seg.
2858 		 */
2859 		if ((lower_addr && upper_addr) ||
2860 		    (raddr == (sglinfo->si_segmask + 1))) {
2861 			return (B_TRUE);
2862 		}
2863 
2864 		size -= psize;
2865 	}
2866 
2867 	return (B_FALSE);
2868 }
2869 
2870 /*
2871  * rootnex_get_sgl()
2872  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2873  *    with a call to the vm layer when vm2.0 comes around...
2874  */
2875 static void
2876 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2877     rootnex_sglinfo_t *sglinfo)
2878 {
2879 	ddi_dma_atyp_t buftype;
2880 	rootnex_addr_t raddr;
2881 	uint64_t last_page;
2882 	uint64_t offset;
2883 	uint64_t addrhi;
2884 	uint64_t addrlo;
2885 	uint64_t maxseg;
2886 	page_t **pplist;
2887 	uint64_t paddr;
2888 	uint32_t psize;
2889 	uint32_t size;
2890 	caddr_t vaddr;
2891 	uint_t pcnt;
2892 	page_t *pp;
2893 	uint_t cnt;
2894 
2895 
2896 	/* shortcuts */
2897 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2898 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2899 	maxseg = sglinfo->si_max_cookie_size;
2900 	buftype = dmar_object->dmao_type;
2901 	addrhi = sglinfo->si_max_addr;
2902 	addrlo = sglinfo->si_min_addr;
2903 	size = dmar_object->dmao_size;
2904 
2905 	pcnt = 0;
2906 	cnt = 0;
2907 
2908 
2909 	/*
2910 	 * check to see if we need to use the copy buffer for pages over
2911 	 * the segment attr.
2912 	 */
2913 	sglinfo->si_bounce_on_seg = B_FALSE;
2914 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2915 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2916 		    dmar_object, sglinfo);
2917 	}
2918 
2919 	/*
2920 	 * if we were passed down a linked list of pages, i.e. pointer to
2921 	 * page_t, use this to get our physical address and buf offset.
2922 	 */
2923 	if (buftype == DMA_OTYP_PAGES) {
2924 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2925 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2926 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2927 		    MMU_PAGEOFFSET;
2928 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2929 		psize = MIN(size, (MMU_PAGESIZE - offset));
2930 		pp = pp->p_next;
2931 		sglinfo->si_asp = NULL;
2932 
2933 	/*
2934 	 * We weren't passed down a linked list of pages, but if we were passed
2935 	 * down an array of pages, use this to get our physical address and buf
2936 	 * offset.
2937 	 */
2938 	} else if (pplist != NULL) {
2939 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2940 		    (buftype == DMA_OTYP_BUFVADDR));
2941 
2942 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2943 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2944 		if (sglinfo->si_asp == NULL) {
2945 			sglinfo->si_asp = &kas;
2946 		}
2947 
2948 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2949 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2950 		paddr += offset;
2951 		psize = MIN(size, (MMU_PAGESIZE - offset));
2952 		pcnt++;
2953 
2954 	/*
2955 	 * All we have is a virtual address, we'll need to call into the VM
2956 	 * to get the physical address.
2957 	 */
2958 	} else {
2959 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2960 		    (buftype == DMA_OTYP_BUFVADDR));
2961 
2962 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2963 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2964 		if (sglinfo->si_asp == NULL) {
2965 			sglinfo->si_asp = &kas;
2966 		}
2967 
2968 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2969 		paddr += offset;
2970 		psize = MIN(size, (MMU_PAGESIZE - offset));
2971 		vaddr += psize;
2972 	}
2973 
2974 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2975 
2976 	/*
2977 	 * Setup the first cookie with the physical address of the page and the
2978 	 * size of the page (which takes into account the initial offset into
2979 	 * the page.
2980 	 */
2981 	sgl[cnt].dmac_laddress = raddr;
2982 	sgl[cnt].dmac_size = psize;
2983 	sgl[cnt].dmac_type = 0;
2984 
2985 	/*
2986 	 * Save away the buffer offset into the page. We'll need this later in
2987 	 * the copy buffer code to help figure out the page index within the
2988 	 * buffer and the offset into the current page.
2989 	 */
2990 	sglinfo->si_buf_offset = offset;
2991 
2992 	/*
2993 	 * If we are using the copy buffer for anything over the segment
2994 	 * boundary, and this page is over the segment boundary.
2995 	 *   OR
2996 	 * if the DMA engine can't reach the physical address.
2997 	 */
2998 	if (((sglinfo->si_bounce_on_seg) &&
2999 	    ((raddr + psize) > sglinfo->si_segmask)) ||
3000 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
3001 		/*
3002 		 * Increase how much copy buffer we use. We always increase by
3003 		 * pagesize so we don't have to worry about converting offsets.
3004 		 * Set a flag in the cookies dmac_type to indicate that it uses
3005 		 * the copy buffer. If this isn't the last cookie, go to the
3006 		 * next cookie (since we separate each page which uses the copy
3007 		 * buffer in case the copy buffer is not physically contiguous.
3008 		 */
3009 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
3010 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3011 		if ((cnt + 1) < sglinfo->si_max_pages) {
3012 			cnt++;
3013 			sgl[cnt].dmac_laddress = 0;
3014 			sgl[cnt].dmac_size = 0;
3015 			sgl[cnt].dmac_type = 0;
3016 		}
3017 	}
3018 
3019 	/*
3020 	 * save this page's physical address so we can figure out if the next
3021 	 * page is physically contiguous. Keep decrementing size until we are
3022 	 * done with the buffer.
3023 	 */
3024 	last_page = raddr & MMU_PAGEMASK;
3025 	size -= psize;
3026 
3027 	while (size > 0) {
3028 		/* Get the size for this page (i.e. partial or full page) */
3029 		psize = MIN(size, MMU_PAGESIZE);
3030 
3031 		if (buftype == DMA_OTYP_PAGES) {
3032 			/* get the paddr from the page_t */
3033 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
3034 			paddr = pfn_to_pa(pp->p_pagenum);
3035 			pp = pp->p_next;
3036 		} else if (pplist != NULL) {
3037 			/* index into the array of page_t's to get the paddr */
3038 			ASSERT(!PP_ISFREE(pplist[pcnt]));
3039 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
3040 			pcnt++;
3041 		} else {
3042 			/* call into the VM to get the paddr */
3043 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
3044 			    vaddr));
3045 			vaddr += psize;
3046 		}
3047 
3048 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
3049 
3050 		/*
3051 		 * If we are using the copy buffer for anything over the
3052 		 * segment boundary, and this page is over the segment
3053 		 * boundary.
3054 		 *   OR
3055 		 * if the DMA engine can't reach the physical address.
3056 		 */
3057 		if (((sglinfo->si_bounce_on_seg) &&
3058 		    ((raddr + psize) > sglinfo->si_segmask)) ||
3059 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
3060 
3061 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
3062 
3063 			/*
3064 			 * if there is something in the current cookie, go to
3065 			 * the next one. We only want one page in a cookie which
3066 			 * uses the copybuf since the copybuf doesn't have to
3067 			 * be physically contiguous.
3068 			 */
3069 			if (sgl[cnt].dmac_size != 0) {
3070 				cnt++;
3071 			}
3072 			sgl[cnt].dmac_laddress = raddr;
3073 			sgl[cnt].dmac_size = psize;
3074 #if defined(__amd64)
3075 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3076 #else
3077 			/*
3078 			 * save the buf offset for 32-bit kernel. used in the
3079 			 * obsoleted interfaces.
3080 			 */
3081 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
3082 			    (dmar_object->dmao_size - size);
3083 #endif
3084 			/* if this isn't the last cookie, go to the next one */
3085 			if ((cnt + 1) < sglinfo->si_max_pages) {
3086 				cnt++;
3087 				sgl[cnt].dmac_laddress = 0;
3088 				sgl[cnt].dmac_size = 0;
3089 				sgl[cnt].dmac_type = 0;
3090 			}
3091 
3092 		/*
3093 		 * this page didn't need the copy buffer, if it's not physically
3094 		 * contiguous, or it would put us over a segment boundary, or it
3095 		 * puts us over the max cookie size, or the current sgl doesn't
3096 		 * have anything in it.
3097 		 */
3098 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
3099 		    !(raddr & sglinfo->si_segmask) ||
3100 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3101 		    (sgl[cnt].dmac_size == 0)) {
3102 			/*
3103 			 * if we're not already in a new cookie, go to the next
3104 			 * cookie.
3105 			 */
3106 			if (sgl[cnt].dmac_size != 0) {
3107 				cnt++;
3108 			}
3109 
3110 			/* save the cookie information */
3111 			sgl[cnt].dmac_laddress = raddr;
3112 			sgl[cnt].dmac_size = psize;
3113 #if defined(__amd64)
3114 			sgl[cnt].dmac_type = 0;
3115 #else
3116 			/*
3117 			 * save the buf offset for 32-bit kernel. used in the
3118 			 * obsoleted interfaces.
3119 			 */
3120 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
3121 #endif
3122 
3123 		/*
3124 		 * this page didn't need the copy buffer, it is physically
3125 		 * contiguous with the last page, and it's <= the max cookie
3126 		 * size.
3127 		 */
3128 		} else {
3129 			sgl[cnt].dmac_size += psize;
3130 
3131 			/*
3132 			 * if this exactly ==  the maximum cookie size, and
3133 			 * it isn't the last cookie, go to the next cookie.
3134 			 */
3135 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3136 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3137 				cnt++;
3138 				sgl[cnt].dmac_laddress = 0;
3139 				sgl[cnt].dmac_size = 0;
3140 				sgl[cnt].dmac_type = 0;
3141 			}
3142 		}
3143 
3144 		/*
3145 		 * save this page's physical address so we can figure out if the
3146 		 * next page is physically contiguous. Keep decrementing size
3147 		 * until we are done with the buffer.
3148 		 */
3149 		last_page = raddr;
3150 		size -= psize;
3151 	}
3152 
3153 	/* we're done, save away how many cookies the sgl has */
3154 	if (sgl[cnt].dmac_size == 0) {
3155 		ASSERT(cnt < sglinfo->si_max_pages);
3156 		sglinfo->si_sgl_size = cnt;
3157 	} else {
3158 		sglinfo->si_sgl_size = cnt + 1;
3159 	}
3160 }
3161 
3162 static void
3163 rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
3164     rootnex_sglinfo_t *sglinfo)
3165 {
3166 	uint64_t offset;
3167 	uint64_t maxseg;
3168 	uint64_t dvaddr;
3169 	struct dvmaseg *dvs;
3170 	uint64_t paddr;
3171 	uint32_t psize, ssize;
3172 	uint32_t size;
3173 	uint_t cnt;
3174 	int physcontig;
3175 
3176 	ASSERT(dmar_object->dmao_type == DMA_OTYP_DVADDR);
3177 
3178 	/* shortcuts */
3179 	maxseg = sglinfo->si_max_cookie_size;
3180 	size = dmar_object->dmao_size;
3181 
3182 	cnt = 0;
3183 	sglinfo->si_bounce_on_seg = B_FALSE;
3184 
3185 	dvs = dmar_object->dmao_obj.dvma_obj.dv_seg;
3186 	offset = dmar_object->dmao_obj.dvma_obj.dv_off;
3187 	ssize = dvs->dvs_len;
3188 	paddr = dvs->dvs_start;
3189 	paddr += offset;
3190 	psize = MIN(ssize, (maxseg - offset));
3191 	dvaddr = paddr + psize;
3192 	ssize -= psize;
3193 
3194 	sgl[cnt].dmac_laddress = paddr;
3195 	sgl[cnt].dmac_size = psize;
3196 	sgl[cnt].dmac_type = 0;
3197 
3198 	size -= psize;
3199 	while (size > 0) {
3200 		if (ssize == 0) {
3201 			dvs++;
3202 			ssize = dvs->dvs_len;
3203 			dvaddr = dvs->dvs_start;
3204 			physcontig = 0;
3205 		} else
3206 			physcontig = 1;
3207 
3208 		paddr = dvaddr;
3209 		psize = MIN(ssize, maxseg);
3210 		dvaddr += psize;
3211 		ssize -= psize;
3212 
3213 		if (!physcontig || !(paddr & sglinfo->si_segmask) ||
3214 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3215 		    (sgl[cnt].dmac_size == 0)) {
3216 			/*
3217 			 * if we're not already in a new cookie, go to the next
3218 			 * cookie.
3219 			 */
3220 			if (sgl[cnt].dmac_size != 0) {
3221 				cnt++;
3222 			}
3223 
3224 			/* save the cookie information */
3225 			sgl[cnt].dmac_laddress = paddr;
3226 			sgl[cnt].dmac_size = psize;
3227 			sgl[cnt].dmac_type = 0;
3228 		} else {
3229 			sgl[cnt].dmac_size += psize;
3230 
3231 			/*
3232 			 * if this exactly ==  the maximum cookie size, and
3233 			 * it isn't the last cookie, go to the next cookie.
3234 			 */
3235 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3236 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3237 				cnt++;
3238 				sgl[cnt].dmac_laddress = 0;
3239 				sgl[cnt].dmac_size = 0;
3240 				sgl[cnt].dmac_type = 0;
3241 			}
3242 		}
3243 		size -= psize;
3244 	}
3245 
3246 	/* we're done, save away how many cookies the sgl has */
3247 	if (sgl[cnt].dmac_size == 0) {
3248 		sglinfo->si_sgl_size = cnt;
3249 	} else {
3250 		sglinfo->si_sgl_size = cnt + 1;
3251 	}
3252 }
3253 
3254 /*
3255  * rootnex_bind_slowpath()
3256  *    Call in the bind path if the calling driver can't use the sgl without
3257  *    modifying it. We either need to use the copy buffer and/or we will end up
3258  *    with a partial bind.
3259  */
3260 static int
3261 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3262     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3263 {
3264 	rootnex_sglinfo_t *sinfo;
3265 	rootnex_window_t *window;
3266 	ddi_dma_cookie_t *cookie;
3267 	size_t copybuf_used;
3268 	size_t dmac_size;
3269 	boolean_t partial;
3270 	off_t cur_offset;
3271 	page_t *cur_pp;
3272 	major_t mnum;
3273 	int e;
3274 	int i;
3275 
3276 
3277 	sinfo = &dma->dp_sglinfo;
3278 	copybuf_used = 0;
3279 	partial = B_FALSE;
3280 
3281 	/*
3282 	 * If we're using the copybuf, set the copybuf state in dma struct.
3283 	 * Needs to be first since it sets the copy buffer size.
3284 	 */
3285 	if (sinfo->si_copybuf_req != 0) {
3286 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
3287 		if (e != DDI_SUCCESS) {
3288 			return (e);
3289 		}
3290 	} else {
3291 		dma->dp_copybuf_size = 0;
3292 	}
3293 
3294 	/*
3295 	 * Figure out if we need to do a partial mapping. If so, figure out
3296 	 * if we need to trim the buffers when we munge the sgl.
3297 	 */
3298 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
3299 	    (dmao->dmao_size > dma->dp_maxxfer) ||
3300 	    ((unsigned)attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
3301 		dma->dp_partial_required = B_TRUE;
3302 		if (attr->dma_attr_granular != 1) {
3303 			dma->dp_trim_required = B_TRUE;
3304 		}
3305 	} else {
3306 		dma->dp_partial_required = B_FALSE;
3307 		dma->dp_trim_required = B_FALSE;
3308 	}
3309 
3310 	/* If we need to do a partial bind, make sure the driver supports it */
3311 	if (dma->dp_partial_required &&
3312 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3313 
3314 		mnum = ddi_driver_major(dma->dp_dip);
3315 		/*
3316 		 * patchable which allows us to print one warning per major
3317 		 * number.
3318 		 */
3319 		if ((rootnex_bind_warn) &&
3320 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3321 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3322 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3323 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3324 			    "There is a small risk of data corruption in "
3325 			    "particular with large I/Os. The driver should be "
3326 			    "replaced with a corrected version for proper "
3327 			    "system operation. To disable this warning, add "
3328 			    "'set rootnex:rootnex_bind_warn=0' to "
3329 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3330 		}
3331 		return (DDI_DMA_TOOBIG);
3332 	}
3333 
3334 	/*
3335 	 * we might need multiple windows, setup state to handle them. In this
3336 	 * code path, we will have at least one window.
3337 	 */
3338 	e = rootnex_setup_windows(hp, dma, attr, dmao, kmflag);
3339 	if (e != DDI_SUCCESS) {
3340 		rootnex_teardown_copybuf(dma);
3341 		return (e);
3342 	}
3343 
3344 	window = &dma->dp_window[0];
3345 	cookie = &dma->dp_cookies[0];
3346 	cur_offset = 0;
3347 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3348 	if (dmao->dmao_type == DMA_OTYP_PAGES) {
3349 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3350 	}
3351 
3352 	/* loop though all the cookies we got back from get_sgl() */
3353 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3354 		/*
3355 		 * If we're using the copy buffer, check this cookie and setup
3356 		 * its associated copy buffer state. If this cookie uses the
3357 		 * copy buffer, make sure we sync this window during dma_sync.
3358 		 */
3359 		if (dma->dp_copybuf_size > 0) {
3360 			rootnex_setup_cookie(dmao, dma, cookie,
3361 			    cur_offset, &copybuf_used, &cur_pp);
3362 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3363 				window->wd_dosync = B_TRUE;
3364 			}
3365 		}
3366 
3367 		/*
3368 		 * save away the cookie size, since it could be modified in
3369 		 * the windowing code.
3370 		 */
3371 		dmac_size = cookie->dmac_size;
3372 
3373 		/* if we went over max copybuf size */
3374 		if (dma->dp_copybuf_size &&
3375 		    (copybuf_used > dma->dp_copybuf_size)) {
3376 			partial = B_TRUE;
3377 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3378 			    cookie, cur_offset, &copybuf_used);
3379 			if (e != DDI_SUCCESS) {
3380 				rootnex_teardown_copybuf(dma);
3381 				rootnex_teardown_windows(dma);
3382 				return (e);
3383 			}
3384 
3385 			/*
3386 			 * if the coookie uses the copy buffer, make sure the
3387 			 * new window we just moved to is set to sync.
3388 			 */
3389 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3390 				window->wd_dosync = B_TRUE;
3391 			}
3392 			ROOTNEX_DPROBE1(rootnex__copybuf__window, dev_info_t *,
3393 			    dma->dp_dip);
3394 
3395 		/* if the cookie cnt == max sgllen, move to the next window */
3396 		} else if (window->wd_cookie_cnt >=
3397 		    (unsigned)attr->dma_attr_sgllen) {
3398 			partial = B_TRUE;
3399 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3400 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3401 			    cookie, attr, cur_offset);
3402 			if (e != DDI_SUCCESS) {
3403 				rootnex_teardown_copybuf(dma);
3404 				rootnex_teardown_windows(dma);
3405 				return (e);
3406 			}
3407 
3408 			/*
3409 			 * if the coookie uses the copy buffer, make sure the
3410 			 * new window we just moved to is set to sync.
3411 			 */
3412 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3413 				window->wd_dosync = B_TRUE;
3414 			}
3415 			ROOTNEX_DPROBE1(rootnex__sgllen__window, dev_info_t *,
3416 			    dma->dp_dip);
3417 
3418 		/* else if we will be over maxxfer */
3419 		} else if ((window->wd_size + dmac_size) >
3420 		    dma->dp_maxxfer) {
3421 			partial = B_TRUE;
3422 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3423 			    cookie);
3424 			if (e != DDI_SUCCESS) {
3425 				rootnex_teardown_copybuf(dma);
3426 				rootnex_teardown_windows(dma);
3427 				return (e);
3428 			}
3429 
3430 			/*
3431 			 * if the coookie uses the copy buffer, make sure the
3432 			 * new window we just moved to is set to sync.
3433 			 */
3434 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3435 				window->wd_dosync = B_TRUE;
3436 			}
3437 			ROOTNEX_DPROBE1(rootnex__maxxfer__window, dev_info_t *,
3438 			    dma->dp_dip);
3439 
3440 		/* else this cookie fits in the current window */
3441 		} else {
3442 			window->wd_cookie_cnt++;
3443 			window->wd_size += dmac_size;
3444 		}
3445 
3446 		/* track our offset into the buffer, go to the next cookie */
3447 		ASSERT(dmac_size <= dmao->dmao_size);
3448 		ASSERT(cookie->dmac_size <= dmac_size);
3449 		cur_offset += dmac_size;
3450 		cookie++;
3451 	}
3452 
3453 	/* if we ended up with a zero sized window in the end, clean it up */
3454 	if (window->wd_size == 0) {
3455 		hp->dmai_nwin--;
3456 		window--;
3457 	}
3458 
3459 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3460 
3461 	if (!partial) {
3462 		return (DDI_DMA_MAPPED);
3463 	}
3464 
3465 	ASSERT(dma->dp_partial_required);
3466 	return (DDI_DMA_PARTIAL_MAP);
3467 }
3468 
3469 /*
3470  * rootnex_setup_copybuf()
3471  *    Called in bind slowpath. Figures out if we're going to use the copy
3472  *    buffer, and if we do, sets up the basic state to handle it.
3473  */
3474 static int
3475 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3476     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3477 {
3478 	rootnex_sglinfo_t *sinfo;
3479 	ddi_dma_attr_t lattr;
3480 	size_t max_copybuf;
3481 	int cansleep;
3482 	int e;
3483 #if !defined(__amd64)
3484 	int vmflag;
3485 #endif
3486 
3487 	ASSERT(!dma->dp_dvma_used);
3488 
3489 	sinfo = &dma->dp_sglinfo;
3490 
3491 	/* read this first so it's consistent through the routine  */
3492 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3493 
3494 	/* We need to call into the rootnex on ddi_dma_sync() */
3495 	hp->dmai_rflags &= ~DMP_NOSYNC;
3496 
3497 	/* make sure the copybuf size <= the max size */
3498 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3499 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3500 
3501 #if !defined(__amd64)
3502 	/*
3503 	 * if we don't have kva space to copy to/from, allocate the KVA space
3504 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3505 	 * the 64-bit kernel.
3506 	 */
3507 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3508 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3509 
3510 		/* convert the sleep flags */
3511 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3512 			vmflag = VM_SLEEP;
3513 		} else {
3514 			vmflag = VM_NOSLEEP;
3515 		}
3516 
3517 		/* allocate Kernel VA space that we can bcopy to/from */
3518 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3519 		    vmflag);
3520 		if (dma->dp_kva == NULL) {
3521 			return (DDI_DMA_NORESOURCES);
3522 		}
3523 	}
3524 #endif
3525 
3526 	/* convert the sleep flags */
3527 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3528 		cansleep = 1;
3529 	} else {
3530 		cansleep = 0;
3531 	}
3532 
3533 	/*
3534 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3535 	 * engine limits, so we can't use kmem_alloc... We don't need
3536 	 * contiguous memory (sgllen) since we will be forcing windows on
3537 	 * sgllen anyway.
3538 	 */
3539 	lattr = *attr;
3540 	lattr.dma_attr_align = MMU_PAGESIZE;
3541 	lattr.dma_attr_sgllen = -1;	/* no limit */
3542 	/*
3543 	 * if we're using the copy buffer because of seg, use that for our
3544 	 * upper address limit.
3545 	 */
3546 	if (sinfo->si_bounce_on_seg) {
3547 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3548 	}
3549 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3550 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3551 	if (e != DDI_SUCCESS) {
3552 #if !defined(__amd64)
3553 		if (dma->dp_kva != NULL) {
3554 			vmem_free(heap_arena, dma->dp_kva,
3555 			    dma->dp_copybuf_size);
3556 		}
3557 #endif
3558 		return (DDI_DMA_NORESOURCES);
3559 	}
3560 
3561 	ROOTNEX_DPROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3562 	    size_t, dma->dp_copybuf_size);
3563 
3564 	return (DDI_SUCCESS);
3565 }
3566 
3567 
3568 /*
3569  * rootnex_setup_windows()
3570  *    Called in bind slowpath to setup the window state. We always have windows
3571  *    in the slowpath. Even if the window count = 1.
3572  */
3573 static int
3574 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3575     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3576 {
3577 	rootnex_window_t *windowp;
3578 	rootnex_sglinfo_t *sinfo;
3579 	size_t copy_state_size;
3580 	size_t win_state_size;
3581 	size_t state_available;
3582 	size_t space_needed;
3583 	uint_t copybuf_win;
3584 	uint_t maxxfer_win;
3585 	size_t space_used;
3586 	uint_t sglwin;
3587 
3588 
3589 	sinfo = &dma->dp_sglinfo;
3590 
3591 	dma->dp_current_win = 0;
3592 	hp->dmai_nwin = 0;
3593 
3594 	/* If we don't need to do a partial, we only have one window */
3595 	if (!dma->dp_partial_required) {
3596 		dma->dp_max_win = 1;
3597 
3598 	/*
3599 	 * we need multiple windows, need to figure out the worse case number
3600 	 * of windows.
3601 	 */
3602 	} else {
3603 		/*
3604 		 * if we need windows because we need more copy buffer that
3605 		 * we allow, the worse case number of windows we could need
3606 		 * here would be (copybuf space required / copybuf space that
3607 		 * we have) plus one for remainder, and plus 2 to handle the
3608 		 * extra pages on the trim for the first and last pages of the
3609 		 * buffer (a page is the minimum window size so under the right
3610 		 * attr settings, you could have a window for each page).
3611 		 * The last page will only be hit here if the size is not a
3612 		 * multiple of the granularity (which theoretically shouldn't
3613 		 * be the case but never has been enforced, so we could have
3614 		 * broken things without it).
3615 		 */
3616 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3617 			ASSERT(dma->dp_copybuf_size > 0);
3618 			copybuf_win = (sinfo->si_copybuf_req /
3619 			    dma->dp_copybuf_size) + 1 + 2;
3620 		} else {
3621 			copybuf_win = 0;
3622 		}
3623 
3624 		/*
3625 		 * if we need windows because we have more cookies than the H/W
3626 		 * can handle, the number of windows we would need here would
3627 		 * be (cookie count / cookies count H/W supports minus 1[for
3628 		 * trim]) plus one for remainder.
3629 		 */
3630 		if ((unsigned)attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3631 			sglwin = (sinfo->si_sgl_size /
3632 			    (attr->dma_attr_sgllen - 1)) + 1;
3633 		} else {
3634 			sglwin = 0;
3635 		}
3636 
3637 		/*
3638 		 * if we need windows because we're binding more memory than the
3639 		 * H/W can transfer at once, the number of windows we would need
3640 		 * here would be (xfer count / max xfer H/W supports) plus one
3641 		 * for remainder, and plus 2 to handle the extra pages on the
3642 		 * trim (see above comment about trim)
3643 		 */
3644 		if (dmao->dmao_size > dma->dp_maxxfer) {
3645 			maxxfer_win = (dmao->dmao_size /
3646 			    dma->dp_maxxfer) + 1 + 2;
3647 		} else {
3648 			maxxfer_win = 0;
3649 		}
3650 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3651 		ASSERT(dma->dp_max_win > 0);
3652 	}
3653 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3654 
3655 	/*
3656 	 * Get space for window and potential copy buffer state. Before we
3657 	 * go and allocate memory, see if we can get away with using what's
3658 	 * left in the pre-allocted state or the dynamically allocated sgl.
3659 	 */
3660 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3661 	    sizeof (ddi_dma_cookie_t));
3662 
3663 	/* if we dynamically allocated space for the cookies */
3664 	if (dma->dp_need_to_free_cookie) {
3665 		/* if we have more space in the pre-allocted buffer, use it */
3666 		ASSERT(space_used <= dma->dp_cookie_size);
3667 		if ((dma->dp_cookie_size - space_used) <=
3668 		    rootnex_state->r_prealloc_size) {
3669 			state_available = rootnex_state->r_prealloc_size;
3670 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3671 
3672 		/*
3673 		 * else, we have more free space in the dynamically allocated
3674 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3675 		 * didn't need a lot of cookies.
3676 		 */
3677 		} else {
3678 			state_available = dma->dp_cookie_size - space_used;
3679 			windowp = (rootnex_window_t *)
3680 			    &dma->dp_cookies[sinfo->si_sgl_size];
3681 		}
3682 
3683 	/* we used the pre-alloced buffer */
3684 	} else {
3685 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3686 		state_available = rootnex_state->r_prealloc_size - space_used;
3687 		windowp = (rootnex_window_t *)
3688 		    &dma->dp_cookies[sinfo->si_sgl_size];
3689 	}
3690 
3691 	/*
3692 	 * figure out how much state we need to track the copy buffer. Add an
3693 	 * addition 8 bytes for pointer alignemnt later.
3694 	 */
3695 	if (dma->dp_copybuf_size > 0) {
3696 		copy_state_size = sinfo->si_max_pages *
3697 		    sizeof (rootnex_pgmap_t);
3698 	} else {
3699 		copy_state_size = 0;
3700 	}
3701 	/* add an additional 8 bytes for pointer alignment */
3702 	space_needed = win_state_size + copy_state_size + 0x8;
3703 
3704 	/* if we have enough space already, use it */
3705 	if (state_available >= space_needed) {
3706 		dma->dp_window = windowp;
3707 		dma->dp_need_to_free_window = B_FALSE;
3708 
3709 	/* not enough space, need to allocate more. */
3710 	} else {
3711 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3712 		if (dma->dp_window == NULL) {
3713 			return (DDI_DMA_NORESOURCES);
3714 		}
3715 		dma->dp_need_to_free_window = B_TRUE;
3716 		dma->dp_window_size = space_needed;
3717 		ROOTNEX_DPROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3718 		    dma->dp_dip, size_t, space_needed);
3719 	}
3720 
3721 	/*
3722 	 * we allocate copy buffer state and window state at the same time.
3723 	 * setup our copy buffer state pointers. Make sure it's aligned.
3724 	 */
3725 	if (dma->dp_copybuf_size > 0) {
3726 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3727 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3728 
3729 #if !defined(__amd64)
3730 		/*
3731 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3732 		 * false/NULL. Should be quicker to bzero vs loop and set.
3733 		 */
3734 		bzero(dma->dp_pgmap, copy_state_size);
3735 #endif
3736 	} else {
3737 		dma->dp_pgmap = NULL;
3738 	}
3739 
3740 	return (DDI_SUCCESS);
3741 }
3742 
3743 
3744 /*
3745  * rootnex_teardown_copybuf()
3746  *    cleans up after rootnex_setup_copybuf()
3747  */
3748 static void
3749 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3750 {
3751 #if !defined(__amd64)
3752 	int i;
3753 
3754 	/*
3755 	 * if we allocated kernel heap VMEM space, go through all the pages and
3756 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3757 	 * arena. Then free the VMEM space.
3758 	 */
3759 	if (dma->dp_kva != NULL) {
3760 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3761 			if (dma->dp_pgmap[i].pm_mapped) {
3762 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3763 				    MMU_PAGESIZE, HAT_UNLOAD);
3764 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3765 			}
3766 		}
3767 
3768 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3769 	}
3770 
3771 #endif
3772 
3773 	/* if we allocated a copy buffer, free it */
3774 	if (dma->dp_cbaddr != NULL) {
3775 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3776 	}
3777 }
3778 
3779 
3780 /*
3781  * rootnex_teardown_windows()
3782  *    cleans up after rootnex_setup_windows()
3783  */
3784 static void
3785 rootnex_teardown_windows(rootnex_dma_t *dma)
3786 {
3787 	/*
3788 	 * if we had to allocate window state on the last bind (because we
3789 	 * didn't have enough pre-allocated space in the handle), free it.
3790 	 */
3791 	if (dma->dp_need_to_free_window) {
3792 		kmem_free(dma->dp_window, dma->dp_window_size);
3793 	}
3794 }
3795 
3796 
3797 /*
3798  * rootnex_init_win()
3799  *    Called in bind slow path during creation of a new window. Initializes
3800  *    window state to default values.
3801  */
3802 /*ARGSUSED*/
3803 static void
3804 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3805     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3806 {
3807 	hp->dmai_nwin++;
3808 	window->wd_dosync = B_FALSE;
3809 	window->wd_offset = cur_offset;
3810 	window->wd_size = 0;
3811 	window->wd_first_cookie = cookie;
3812 	window->wd_cookie_cnt = 0;
3813 	window->wd_trim.tr_trim_first = B_FALSE;
3814 	window->wd_trim.tr_trim_last = B_FALSE;
3815 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3816 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3817 #if !defined(__amd64)
3818 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3819 #endif
3820 }
3821 
3822 
3823 /*
3824  * rootnex_setup_cookie()
3825  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3826  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3827  *    out if it uses the copy buffer, and if it does, save away everything we'll
3828  *    need during sync.
3829  */
3830 static void
3831 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3832     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3833     page_t **cur_pp)
3834 {
3835 	boolean_t copybuf_sz_power_2;
3836 	rootnex_sglinfo_t *sinfo;
3837 	paddr_t paddr;
3838 	uint_t pidx;
3839 	uint_t pcnt;
3840 	off_t poff;
3841 #if defined(__amd64)
3842 	pfn_t pfn;
3843 #else
3844 	page_t **pplist;
3845 #endif
3846 
3847 	ASSERT(dmar_object->dmao_type != DMA_OTYP_DVADDR);
3848 
3849 	sinfo = &dma->dp_sglinfo;
3850 
3851 	/*
3852 	 * Calculate the page index relative to the start of the buffer. The
3853 	 * index to the current page for our buffer is the offset into the
3854 	 * first page of the buffer plus our current offset into the buffer
3855 	 * itself, shifted of course...
3856 	 */
3857 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3858 	ASSERT(pidx < sinfo->si_max_pages);
3859 
3860 	/* if this cookie uses the copy buffer */
3861 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3862 		/*
3863 		 * NOTE: we know that since this cookie uses the copy buffer, it
3864 		 * is <= MMU_PAGESIZE.
3865 		 */
3866 
3867 		/*
3868 		 * get the offset into the page. For the 64-bit kernel, get the
3869 		 * pfn which we'll use with seg kpm.
3870 		 */
3871 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3872 #if defined(__amd64)
3873 		/* mfn_to_pfn() is a NOP on i86pc */
3874 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3875 #endif /* __amd64 */
3876 
3877 		/* figure out if the copybuf size is a power of 2 */
3878 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3879 			copybuf_sz_power_2 = B_FALSE;
3880 		} else {
3881 			copybuf_sz_power_2 = B_TRUE;
3882 		}
3883 
3884 		/* This page uses the copy buffer */
3885 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3886 
3887 		/*
3888 		 * save the copy buffer KVA that we'll use with this page.
3889 		 * if we still fit within the copybuf, it's a simple add.
3890 		 * otherwise, we need to wrap over using & or % accordingly.
3891 		 */
3892 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3893 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3894 			    *copybuf_used;
3895 		} else {
3896 			if (copybuf_sz_power_2) {
3897 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3898 				    (uintptr_t)dma->dp_cbaddr +
3899 				    (*copybuf_used &
3900 				    (dma->dp_copybuf_size - 1)));
3901 			} else {
3902 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3903 				    (uintptr_t)dma->dp_cbaddr +
3904 				    (*copybuf_used % dma->dp_copybuf_size));
3905 			}
3906 		}
3907 
3908 		/*
3909 		 * over write the cookie physical address with the address of
3910 		 * the physical address of the copy buffer page that we will
3911 		 * use.
3912 		 */
3913 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3914 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3915 
3916 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
3917 
3918 		/* if we have a kernel VA, it's easy, just save that address */
3919 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3920 		    (sinfo->si_asp == &kas)) {
3921 			/*
3922 			 * save away the page aligned virtual address of the
3923 			 * driver buffer. Offsets are handled in the sync code.
3924 			 */
3925 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3926 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3927 			    & MMU_PAGEMASK);
3928 #if !defined(__amd64)
3929 			/*
3930 			 * we didn't need to, and will never need to map this
3931 			 * page.
3932 			 */
3933 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3934 #endif
3935 
3936 		/* we don't have a kernel VA. We need one for the bcopy. */
3937 		} else {
3938 #if defined(__amd64)
3939 			/*
3940 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3941 			 * get a Kernel VA for the corresponding pfn.
3942 			 */
3943 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3944 #else
3945 			/*
3946 			 * for the 32-bit kernel, this is a pain. First we'll
3947 			 * save away the page_t or user VA for this page. This
3948 			 * is needed in rootnex_dma_win() when we switch to a
3949 			 * new window which requires us to re-map the copy
3950 			 * buffer.
3951 			 */
3952 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3953 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3954 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3955 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3956 			} else if (pplist != NULL) {
3957 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3958 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3959 			} else {
3960 				dma->dp_pgmap[pidx].pm_pp = NULL;
3961 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3962 				    (((uintptr_t)
3963 				    dmar_object->dmao_obj.virt_obj.v_addr +
3964 				    cur_offset) & MMU_PAGEMASK);
3965 			}
3966 
3967 			/*
3968 			 * save away the page aligned virtual address which was
3969 			 * allocated from the kernel heap arena (taking into
3970 			 * account if we need more copy buffer than we alloced
3971 			 * and use multiple windows to handle this, i.e. &,%).
3972 			 * NOTE: there isn't and physical memory backing up this
3973 			 * virtual address space currently.
3974 			 */
3975 			if ((*copybuf_used + MMU_PAGESIZE) <=
3976 			    dma->dp_copybuf_size) {
3977 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3978 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3979 				    MMU_PAGEMASK);
3980 			} else {
3981 				if (copybuf_sz_power_2) {
3982 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3983 					    (((uintptr_t)dma->dp_kva +
3984 					    (*copybuf_used &
3985 					    (dma->dp_copybuf_size - 1))) &
3986 					    MMU_PAGEMASK);
3987 				} else {
3988 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3989 					    (((uintptr_t)dma->dp_kva +
3990 					    (*copybuf_used %
3991 					    dma->dp_copybuf_size)) &
3992 					    MMU_PAGEMASK);
3993 				}
3994 			}
3995 
3996 			/*
3997 			 * if we haven't used up the available copy buffer yet,
3998 			 * map the kva to the physical page.
3999 			 */
4000 			if (!dma->dp_cb_remaping && ((*copybuf_used +
4001 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
4002 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
4003 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
4004 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
4005 					    dma->dp_pgmap[pidx].pm_kaddr);
4006 				} else {
4007 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
4008 					    sinfo->si_asp,
4009 					    dma->dp_pgmap[pidx].pm_kaddr);
4010 				}
4011 
4012 			/*
4013 			 * we've used up the available copy buffer, this page
4014 			 * will have to be mapped during rootnex_dma_win() when
4015 			 * we switch to a new window which requires a re-map
4016 			 * the copy buffer. (32-bit kernel only)
4017 			 */
4018 			} else {
4019 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4020 			}
4021 #endif
4022 			/* go to the next page_t */
4023 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4024 				*cur_pp = (*cur_pp)->p_next;
4025 			}
4026 		}
4027 
4028 		/* add to the copy buffer count */
4029 		*copybuf_used += MMU_PAGESIZE;
4030 
4031 	/*
4032 	 * This cookie doesn't use the copy buffer. Walk through the pages this
4033 	 * cookie occupies to reflect this.
4034 	 */
4035 	} else {
4036 		/*
4037 		 * figure out how many pages the cookie occupies. We need to
4038 		 * use the original page offset of the buffer and the cookies
4039 		 * offset in the buffer to do this.
4040 		 */
4041 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
4042 		pcnt = mmu_btopr(cookie->dmac_size + poff);
4043 
4044 		while (pcnt > 0) {
4045 #if !defined(__amd64)
4046 			/*
4047 			 * the 32-bit kernel doesn't have seg kpm, so we need
4048 			 * to map in the driver buffer (if it didn't come down
4049 			 * with a kernel VA) on the fly. Since this page doesn't
4050 			 * use the copy buffer, it's not, or will it ever, have
4051 			 * to be mapped in.
4052 			 */
4053 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4054 #endif
4055 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
4056 
4057 			/*
4058 			 * we need to update pidx and cur_pp or we'll loose
4059 			 * track of where we are.
4060 			 */
4061 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4062 				*cur_pp = (*cur_pp)->p_next;
4063 			}
4064 			pidx++;
4065 			pcnt--;
4066 		}
4067 	}
4068 }
4069 
4070 
4071 /*
4072  * rootnex_sgllen_window_boundary()
4073  *    Called in the bind slow path when the next cookie causes us to exceed (in
4074  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
4075  *    length supported by the DMA H/W.
4076  */
4077 static int
4078 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4079     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
4080     off_t cur_offset)
4081 {
4082 	off_t new_offset;
4083 	size_t trim_sz;
4084 	off_t coffset;
4085 
4086 
4087 	/*
4088 	 * if we know we'll never have to trim, it's pretty easy. Just move to
4089 	 * the next window and init it. We're done.
4090 	 */
4091 	if (!dma->dp_trim_required) {
4092 		(*windowp)++;
4093 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4094 		(*windowp)->wd_cookie_cnt++;
4095 		(*windowp)->wd_size = cookie->dmac_size;
4096 		return (DDI_SUCCESS);
4097 	}
4098 
4099 	/* figure out how much we need to trim from the window */
4100 	ASSERT(attr->dma_attr_granular != 0);
4101 	if (dma->dp_granularity_power_2) {
4102 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
4103 	} else {
4104 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
4105 	}
4106 
4107 	/* The window's a whole multiple of granularity. We're done */
4108 	if (trim_sz == 0) {
4109 		(*windowp)++;
4110 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4111 		(*windowp)->wd_cookie_cnt++;
4112 		(*windowp)->wd_size = cookie->dmac_size;
4113 		return (DDI_SUCCESS);
4114 	}
4115 
4116 	/*
4117 	 * The window's not a whole multiple of granularity, since we know this
4118 	 * is due to the sgllen, we need to go back to the last cookie and trim
4119 	 * that one, add the left over part of the old cookie into the new
4120 	 * window, and then add in the new cookie into the new window.
4121 	 */
4122 
4123 	/*
4124 	 * make sure the driver isn't making us do something bad... Trimming and
4125 	 * sgllen == 1 don't go together.
4126 	 */
4127 	if (attr->dma_attr_sgllen == 1) {
4128 		return (DDI_DMA_NOMAPPING);
4129 	}
4130 
4131 	/*
4132 	 * first, setup the current window to account for the trim. Need to go
4133 	 * back to the last cookie for this.
4134 	 */
4135 	cookie--;
4136 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4137 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4138 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4139 	ASSERT(cookie->dmac_size > trim_sz);
4140 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4141 	(*windowp)->wd_size -= trim_sz;
4142 
4143 	/* save the buffer offsets for the next window */
4144 	coffset = cookie->dmac_size - trim_sz;
4145 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4146 
4147 	/*
4148 	 * set this now in case this is the first window. all other cases are
4149 	 * set in dma_win()
4150 	 */
4151 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4152 
4153 	/*
4154 	 * initialize the next window using what's left over in the previous
4155 	 * cookie.
4156 	 */
4157 	(*windowp)++;
4158 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4159 	(*windowp)->wd_cookie_cnt++;
4160 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4161 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4162 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4163 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4164 		(*windowp)->wd_dosync = B_TRUE;
4165 	}
4166 
4167 	/*
4168 	 * now go back to the current cookie and add it to the new window. set
4169 	 * the new window size to the what was left over from the previous
4170 	 * cookie and what's in the current cookie.
4171 	 */
4172 	cookie++;
4173 	(*windowp)->wd_cookie_cnt++;
4174 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4175 
4176 	/*
4177 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
4178 	 * a max size of maxxfer). Handle that case.
4179 	 */
4180 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
4181 		/*
4182 		 * maxxfer is already a whole multiple of granularity, and this
4183 		 * trim will be <= the previous trim (since a cookie can't be
4184 		 * larger than maxxfer). Make things simple here.
4185 		 */
4186 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
4187 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4188 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4189 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4190 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4191 		(*windowp)->wd_size -= trim_sz;
4192 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
4193 
4194 		/* save the buffer offsets for the next window */
4195 		coffset = cookie->dmac_size - trim_sz;
4196 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4197 
4198 		/* setup the next window */
4199 		(*windowp)++;
4200 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4201 		(*windowp)->wd_cookie_cnt++;
4202 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4203 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4204 		    coffset;
4205 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4206 	}
4207 
4208 	return (DDI_SUCCESS);
4209 }
4210 
4211 
4212 /*
4213  * rootnex_copybuf_window_boundary()
4214  *    Called in bind slowpath when we get to a window boundary because we used
4215  *    up all the copy buffer that we have.
4216  */
4217 static int
4218 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4219     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
4220     size_t *copybuf_used)
4221 {
4222 	rootnex_sglinfo_t *sinfo;
4223 	off_t new_offset;
4224 	size_t trim_sz;
4225 	paddr_t paddr;
4226 	off_t coffset;
4227 	uint_t pidx;
4228 	off_t poff;
4229 
4230 
4231 	sinfo = &dma->dp_sglinfo;
4232 
4233 	/*
4234 	 * the copy buffer should be a whole multiple of page size. We know that
4235 	 * this cookie is <= MMU_PAGESIZE.
4236 	 */
4237 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
4238 
4239 	/*
4240 	 * from now on, all new windows in this bind need to be re-mapped during
4241 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
4242 	 * space...
4243 	 */
4244 #if !defined(__amd64)
4245 	dma->dp_cb_remaping = B_TRUE;
4246 #endif
4247 
4248 	/* reset copybuf used */
4249 	*copybuf_used = 0;
4250 
4251 	/*
4252 	 * if we don't have to trim (since granularity is set to 1), go to the
4253 	 * next window and add the current cookie to it. We know the current
4254 	 * cookie uses the copy buffer since we're in this code path.
4255 	 */
4256 	if (!dma->dp_trim_required) {
4257 		(*windowp)++;
4258 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4259 
4260 		/* Add this cookie to the new window */
4261 		(*windowp)->wd_cookie_cnt++;
4262 		(*windowp)->wd_size += cookie->dmac_size;
4263 		*copybuf_used += MMU_PAGESIZE;
4264 		return (DDI_SUCCESS);
4265 	}
4266 
4267 	/*
4268 	 * *** may need to trim, figure it out.
4269 	 */
4270 
4271 	/* figure out how much we need to trim from the window */
4272 	if (dma->dp_granularity_power_2) {
4273 		trim_sz = (*windowp)->wd_size &
4274 		    (hp->dmai_attr.dma_attr_granular - 1);
4275 	} else {
4276 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
4277 	}
4278 
4279 	/*
4280 	 * if the window's a whole multiple of granularity, go to the next
4281 	 * window, init it, then add in the current cookie. We know the current
4282 	 * cookie uses the copy buffer since we're in this code path.
4283 	 */
4284 	if (trim_sz == 0) {
4285 		(*windowp)++;
4286 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4287 
4288 		/* Add this cookie to the new window */
4289 		(*windowp)->wd_cookie_cnt++;
4290 		(*windowp)->wd_size += cookie->dmac_size;
4291 		*copybuf_used += MMU_PAGESIZE;
4292 		return (DDI_SUCCESS);
4293 	}
4294 
4295 	/*
4296 	 * *** We figured it out, we definitly need to trim
4297 	 */
4298 
4299 	/*
4300 	 * make sure the driver isn't making us do something bad...
4301 	 * Trimming and sgllen == 1 don't go together.
4302 	 */
4303 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4304 		return (DDI_DMA_NOMAPPING);
4305 	}
4306 
4307 	/*
4308 	 * first, setup the current window to account for the trim. Need to go
4309 	 * back to the last cookie for this. Some of the last cookie will be in
4310 	 * the current window, and some of the last cookie will be in the new
4311 	 * window. All of the current cookie will be in the new window.
4312 	 */
4313 	cookie--;
4314 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4315 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4316 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4317 	ASSERT(cookie->dmac_size > trim_sz);
4318 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4319 	(*windowp)->wd_size -= trim_sz;
4320 
4321 	/*
4322 	 * we're trimming the last cookie (not the current cookie). So that
4323 	 * last cookie may have or may not have been using the copy buffer (
4324 	 * we know the cookie passed in uses the copy buffer since we're in
4325 	 * this code path).
4326 	 *
4327 	 * If the last cookie doesn't use the copy buffer, nothing special to
4328 	 * do. However, if it does uses the copy buffer, it will be both the
4329 	 * last page in the current window and the first page in the next
4330 	 * window. Since we are reusing the copy buffer (and KVA space on the
4331 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4332 	 * current window, and the start of the copy buffer in the next window.
4333 	 * Track that info... The cookie physical address was already set to
4334 	 * the copy buffer physical address in setup_cookie..
4335 	 */
4336 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4337 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4338 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4339 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4340 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4341 		(*windowp)->wd_trim.tr_last_cbaddr =
4342 		    dma->dp_pgmap[pidx].pm_cbaddr;
4343 #if !defined(__amd64)
4344 		(*windowp)->wd_trim.tr_last_kaddr =
4345 		    dma->dp_pgmap[pidx].pm_kaddr;
4346 #endif
4347 	}
4348 
4349 	/* save the buffer offsets for the next window */
4350 	coffset = cookie->dmac_size - trim_sz;
4351 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4352 
4353 	/*
4354 	 * set this now in case this is the first window. all other cases are
4355 	 * set in dma_win()
4356 	 */
4357 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4358 
4359 	/*
4360 	 * initialize the next window using what's left over in the previous
4361 	 * cookie.
4362 	 */
4363 	(*windowp)++;
4364 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4365 	(*windowp)->wd_cookie_cnt++;
4366 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4367 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4368 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4369 
4370 	/*
4371 	 * again, we're tracking if the last cookie uses the copy buffer.
4372 	 * read the comment above for more info on why we need to track
4373 	 * additional state.
4374 	 *
4375 	 * For the first cookie in the new window, we need reset the physical
4376 	 * address to DMA into to the start of the copy buffer plus any
4377 	 * initial page offset which may be present.
4378 	 */
4379 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4380 		(*windowp)->wd_dosync = B_TRUE;
4381 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4382 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4383 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4384 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4385 
4386 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4387 		    poff;
4388 		(*windowp)->wd_trim.tr_first_paddr =
4389 		    ROOTNEX_PADDR_TO_RBASE(paddr);
4390 
4391 #if !defined(__amd64)
4392 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4393 #endif
4394 		/* account for the cookie copybuf usage in the new window */
4395 		*copybuf_used += MMU_PAGESIZE;
4396 
4397 		/*
4398 		 * every piece of code has to have a hack, and here is this
4399 		 * ones :-)
4400 		 *
4401 		 * There is a complex interaction between setup_cookie and the
4402 		 * copybuf window boundary. The complexity had to be in either
4403 		 * the maxxfer window, or the copybuf window, and I chose the
4404 		 * copybuf code.
4405 		 *
4406 		 * So in this code path, we have taken the last cookie,
4407 		 * virtually broken it in half due to the trim, and it happens
4408 		 * to use the copybuf which further complicates life. At the
4409 		 * same time, we have already setup the current cookie, which
4410 		 * is now wrong. More background info: the current cookie uses
4411 		 * the copybuf, so it is only a page long max. So we need to
4412 		 * fix the current cookies copy buffer address, physical
4413 		 * address, and kva for the 32-bit kernel. We due this by
4414 		 * bumping them by page size (of course, we can't due this on
4415 		 * the physical address since the copy buffer may not be
4416 		 * physically contiguous).
4417 		 */
4418 		cookie++;
4419 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4420 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4421 
4422 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4423 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4424 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
4425 
4426 #if !defined(__amd64)
4427 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4428 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4429 #endif
4430 	} else {
4431 		/* go back to the current cookie */
4432 		cookie++;
4433 	}
4434 
4435 	/*
4436 	 * add the current cookie to the new window. set the new window size to
4437 	 * the what was left over from the previous cookie and what's in the
4438 	 * current cookie.
4439 	 */
4440 	(*windowp)->wd_cookie_cnt++;
4441 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4442 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4443 
4444 	/*
4445 	 * we know that the cookie passed in always uses the copy buffer. We
4446 	 * wouldn't be here if it didn't.
4447 	 */
4448 	*copybuf_used += MMU_PAGESIZE;
4449 
4450 	return (DDI_SUCCESS);
4451 }
4452 
4453 
4454 /*
4455  * rootnex_maxxfer_window_boundary()
4456  *    Called in bind slowpath when we get to a window boundary because we will
4457  *    go over maxxfer.
4458  */
4459 static int
4460 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4461     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4462 {
4463 	size_t dmac_size;
4464 	off_t new_offset;
4465 	size_t trim_sz;
4466 	off_t coffset;
4467 
4468 
4469 	/*
4470 	 * calculate how much we have to trim off of the current cookie to equal
4471 	 * maxxfer. We don't have to account for granularity here since our
4472 	 * maxxfer already takes that into account.
4473 	 */
4474 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4475 	ASSERT(trim_sz <= cookie->dmac_size);
4476 	ASSERT(trim_sz <= dma->dp_maxxfer);
4477 
4478 	/* save cookie size since we need it later and we might change it */
4479 	dmac_size = cookie->dmac_size;
4480 
4481 	/*
4482 	 * if we're not trimming the entire cookie, setup the current window to
4483 	 * account for the trim.
4484 	 */
4485 	if (trim_sz < cookie->dmac_size) {
4486 		(*windowp)->wd_cookie_cnt++;
4487 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4488 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4489 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4490 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4491 		(*windowp)->wd_size = dma->dp_maxxfer;
4492 
4493 		/*
4494 		 * set the adjusted cookie size now in case this is the first
4495 		 * window. All other windows are taken care of in get win
4496 		 */
4497 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4498 	}
4499 
4500 	/*
4501 	 * coffset is the current offset within the cookie, new_offset is the
4502 	 * current offset with the entire buffer.
4503 	 */
4504 	coffset = dmac_size - trim_sz;
4505 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4506 
4507 	/* initialize the next window */
4508 	(*windowp)++;
4509 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4510 	(*windowp)->wd_cookie_cnt++;
4511 	(*windowp)->wd_size = trim_sz;
4512 	if (trim_sz < dmac_size) {
4513 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4514 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4515 		    coffset;
4516 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4517 	}
4518 
4519 	return (DDI_SUCCESS);
4520 }
4521 
4522 
4523 /*ARGSUSED*/
4524 static int
4525 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4526     off_t off, size_t len, uint_t cache_flags)
4527 {
4528 	rootnex_sglinfo_t *sinfo;
4529 	rootnex_pgmap_t *cbpage;
4530 	rootnex_window_t *win;
4531 	ddi_dma_impl_t *hp;
4532 	rootnex_dma_t *dma;
4533 	caddr_t fromaddr;
4534 	caddr_t toaddr;
4535 	uint_t psize;
4536 	off_t offset;
4537 	uint_t pidx;
4538 	size_t size;
4539 	off_t poff;
4540 	int e;
4541 
4542 
4543 	hp = (ddi_dma_impl_t *)handle;
4544 	dma = (rootnex_dma_t *)hp->dmai_private;
4545 	sinfo = &dma->dp_sglinfo;
4546 
4547 	/*
4548 	 * if we don't have any windows, we don't need to sync. A copybuf
4549 	 * will cause us to have at least one window.
4550 	 */
4551 	if (dma->dp_window == NULL) {
4552 		return (DDI_SUCCESS);
4553 	}
4554 
4555 	/* This window may not need to be sync'd */
4556 	win = &dma->dp_window[dma->dp_current_win];
4557 	if (!win->wd_dosync) {
4558 		return (DDI_SUCCESS);
4559 	}
4560 
4561 	/* handle off and len special cases */
4562 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4563 		offset = win->wd_offset;
4564 	} else {
4565 		offset = off;
4566 	}
4567 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4568 		size = win->wd_size;
4569 	} else {
4570 		size = len;
4571 	}
4572 
4573 	/* check the sync args to make sure they make a little sense */
4574 	if (rootnex_sync_check_parms) {
4575 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4576 		    cache_flags);
4577 		if (e != DDI_SUCCESS) {
4578 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4579 			return (DDI_FAILURE);
4580 		}
4581 	}
4582 
4583 	/*
4584 	 * special case the first page to handle the offset into the page. The
4585 	 * offset to the current page for our buffer is the offset into the
4586 	 * first page of the buffer plus our current offset into the buffer
4587 	 * itself, masked of course.
4588 	 */
4589 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4590 	psize = MIN((MMU_PAGESIZE - poff), size);
4591 
4592 	/* go through all the pages that we want to sync */
4593 	while (size > 0) {
4594 		/*
4595 		 * Calculate the page index relative to the start of the buffer.
4596 		 * The index to the current page for our buffer is the offset
4597 		 * into the first page of the buffer plus our current offset
4598 		 * into the buffer itself, shifted of course...
4599 		 */
4600 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4601 		ASSERT(pidx < sinfo->si_max_pages);
4602 
4603 		/*
4604 		 * if this page uses the copy buffer, we need to sync it,
4605 		 * otherwise, go on to the next page.
4606 		 */
4607 		cbpage = &dma->dp_pgmap[pidx];
4608 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4609 		    (cbpage->pm_uses_copybuf == B_FALSE));
4610 		if (cbpage->pm_uses_copybuf) {
4611 			/* cbaddr and kaddr should be page aligned */
4612 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4613 			    MMU_PAGEOFFSET) == 0);
4614 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4615 			    MMU_PAGEOFFSET) == 0);
4616 
4617 			/*
4618 			 * if we're copying for the device, we are going to
4619 			 * copy from the drivers buffer and to the rootnex
4620 			 * allocated copy buffer.
4621 			 */
4622 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4623 				fromaddr = cbpage->pm_kaddr + poff;
4624 				toaddr = cbpage->pm_cbaddr + poff;
4625 				ROOTNEX_DPROBE2(rootnex__sync__dev,
4626 				    dev_info_t *, dma->dp_dip, size_t, psize);
4627 
4628 			/*
4629 			 * if we're copying for the cpu/kernel, we are going to
4630 			 * copy from the rootnex allocated copy buffer to the
4631 			 * drivers buffer.
4632 			 */
4633 			} else {
4634 				fromaddr = cbpage->pm_cbaddr + poff;
4635 				toaddr = cbpage->pm_kaddr + poff;
4636 				ROOTNEX_DPROBE2(rootnex__sync__cpu,
4637 				    dev_info_t *, dma->dp_dip, size_t, psize);
4638 			}
4639 
4640 			bcopy(fromaddr, toaddr, psize);
4641 		}
4642 
4643 		/*
4644 		 * decrement size until we're done, update our offset into the
4645 		 * buffer, and get the next page size.
4646 		 */
4647 		size -= psize;
4648 		offset += psize;
4649 		psize = MIN(MMU_PAGESIZE, size);
4650 
4651 		/* page offset is zero for the rest of this loop */
4652 		poff = 0;
4653 	}
4654 
4655 	return (DDI_SUCCESS);
4656 }
4657 
4658 /*
4659  * rootnex_dma_sync()
4660  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4661  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4662  *    is set, ddi_dma_sync() returns immediately passing back success.
4663  */
4664 /*ARGSUSED*/
4665 static int
4666 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4667     off_t off, size_t len, uint_t cache_flags)
4668 {
4669 #if defined(__amd64) && !defined(__xpv)
4670 	if (IOMMU_USED(rdip)) {
4671 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4672 		    cache_flags));
4673 	}
4674 #endif
4675 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4676 	    cache_flags));
4677 }
4678 
4679 /*
4680  * rootnex_valid_sync_parms()
4681  *    checks the parameters passed to sync to verify they are correct.
4682  */
4683 static int
4684 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4685     off_t offset, size_t size, uint_t cache_flags)
4686 {
4687 	off_t woffset;
4688 
4689 
4690 	/*
4691 	 * the first part of the test to make sure the offset passed in is
4692 	 * within the window.
4693 	 */
4694 	if (offset < win->wd_offset) {
4695 		return (DDI_FAILURE);
4696 	}
4697 
4698 	/*
4699 	 * second and last part of the test to make sure the offset and length
4700 	 * passed in is within the window.
4701 	 */
4702 	woffset = offset - win->wd_offset;
4703 	if ((woffset + size) > win->wd_size) {
4704 		return (DDI_FAILURE);
4705 	}
4706 
4707 	/*
4708 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4709 	 * be set too.
4710 	 */
4711 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4712 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4713 		return (DDI_SUCCESS);
4714 	}
4715 
4716 	/*
4717 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4718 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4719 	 */
4720 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4721 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4722 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4723 		return (DDI_SUCCESS);
4724 	}
4725 
4726 	return (DDI_FAILURE);
4727 }
4728 
4729 
4730 /*ARGSUSED*/
4731 static int
4732 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4733     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4734     uint_t *ccountp)
4735 {
4736 	rootnex_window_t *window;
4737 	rootnex_trim_t *trim;
4738 	ddi_dma_impl_t *hp;
4739 	rootnex_dma_t *dma;
4740 	ddi_dma_obj_t *dmao;
4741 #if !defined(__amd64)
4742 	rootnex_sglinfo_t *sinfo;
4743 	rootnex_pgmap_t *pmap;
4744 	uint_t pidx;
4745 	uint_t pcnt;
4746 	off_t poff;
4747 	int i;
4748 #endif
4749 
4750 
4751 	hp = (ddi_dma_impl_t *)handle;
4752 	dma = (rootnex_dma_t *)hp->dmai_private;
4753 #if !defined(__amd64)
4754 	sinfo = &dma->dp_sglinfo;
4755 #endif
4756 
4757 	/* If we try and get a window which doesn't exist, return failure */
4758 	if (win >= hp->dmai_nwin) {
4759 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4760 		return (DDI_FAILURE);
4761 	}
4762 
4763 	dmao = dma->dp_dvma_used ? &dma->dp_dvma : &dma->dp_dma;
4764 
4765 	/*
4766 	 * if we don't have any windows, and they're asking for the first
4767 	 * window, setup the cookie pointer to the first cookie in the bind.
4768 	 * setup our return values, then increment the cookie since we return
4769 	 * the first cookie on the stack.
4770 	 */
4771 	if (dma->dp_window == NULL) {
4772 		if (win != 0) {
4773 			ROOTNEX_DPROF_INC(
4774 			    &rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4775 			return (DDI_FAILURE);
4776 		}
4777 		hp->dmai_cookie = dma->dp_cookies;
4778 		*offp = 0;
4779 		*lenp = dmao->dmao_size;
4780 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4781 		*cookiep = hp->dmai_cookie[0];
4782 		hp->dmai_cookie++;
4783 		return (DDI_SUCCESS);
4784 	}
4785 
4786 	/* sync the old window before moving on to the new one */
4787 	window = &dma->dp_window[dma->dp_current_win];
4788 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4789 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4790 		    DDI_DMA_SYNC_FORCPU);
4791 	}
4792 
4793 #if !defined(__amd64)
4794 	/*
4795 	 * before we move to the next window, if we need to re-map, unmap all
4796 	 * the pages in this window.
4797 	 */
4798 	if (dma->dp_cb_remaping) {
4799 		/*
4800 		 * If we switch to this window again, we'll need to map in
4801 		 * on the fly next time.
4802 		 */
4803 		window->wd_remap_copybuf = B_TRUE;
4804 
4805 		/*
4806 		 * calculate the page index into the buffer where this window
4807 		 * starts, and the number of pages this window takes up.
4808 		 */
4809 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4810 		    MMU_PAGESHIFT;
4811 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4812 		    MMU_PAGEOFFSET;
4813 		pcnt = mmu_btopr(window->wd_size + poff);
4814 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4815 
4816 		/* unmap pages which are currently mapped in this window */
4817 		for (i = 0; i < pcnt; i++) {
4818 			if (dma->dp_pgmap[pidx].pm_mapped) {
4819 				hat_unload(kas.a_hat,
4820 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4821 				    HAT_UNLOAD);
4822 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4823 			}
4824 			pidx++;
4825 		}
4826 	}
4827 #endif
4828 
4829 	/*
4830 	 * Move to the new window.
4831 	 * NOTE: current_win must be set for sync to work right
4832 	 */
4833 	dma->dp_current_win = win;
4834 	window = &dma->dp_window[win];
4835 
4836 	/* if needed, adjust the first and/or last cookies for trim */
4837 	trim = &window->wd_trim;
4838 	if (trim->tr_trim_first) {
4839 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4840 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4841 #if !defined(__amd64)
4842 		window->wd_first_cookie->dmac_type =
4843 		    (window->wd_first_cookie->dmac_type &
4844 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4845 #endif
4846 		if (trim->tr_first_copybuf_win) {
4847 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4848 			    trim->tr_first_cbaddr;
4849 #if !defined(__amd64)
4850 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4851 			    trim->tr_first_kaddr;
4852 #endif
4853 		}
4854 	}
4855 	if (trim->tr_trim_last) {
4856 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4857 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4858 		if (trim->tr_last_copybuf_win) {
4859 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4860 			    trim->tr_last_cbaddr;
4861 #if !defined(__amd64)
4862 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4863 			    trim->tr_last_kaddr;
4864 #endif
4865 		}
4866 	}
4867 
4868 	/*
4869 	 * setup the cookie pointer to the first cookie in the window. setup
4870 	 * our return values, then increment the cookie since we return the
4871 	 * first cookie on the stack.
4872 	 */
4873 	hp->dmai_cookie = window->wd_first_cookie;
4874 	*offp = window->wd_offset;
4875 	*lenp = window->wd_size;
4876 	*ccountp = window->wd_cookie_cnt;
4877 	*cookiep = hp->dmai_cookie[0];
4878 	hp->dmai_cookie++;
4879 
4880 #if !defined(__amd64)
4881 	/* re-map copybuf if required for this window */
4882 	if (dma->dp_cb_remaping) {
4883 		/*
4884 		 * calculate the page index into the buffer where this
4885 		 * window starts.
4886 		 */
4887 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4888 		    MMU_PAGESHIFT;
4889 		ASSERT(pidx < sinfo->si_max_pages);
4890 
4891 		/*
4892 		 * the first page can get unmapped if it's shared with the
4893 		 * previous window. Even if the rest of this window is already
4894 		 * mapped in, we need to still check this one.
4895 		 */
4896 		pmap = &dma->dp_pgmap[pidx];
4897 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4898 			if (pmap->pm_pp != NULL) {
4899 				pmap->pm_mapped = B_TRUE;
4900 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4901 			} else if (pmap->pm_vaddr != NULL) {
4902 				pmap->pm_mapped = B_TRUE;
4903 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4904 				    pmap->pm_kaddr);
4905 			}
4906 		}
4907 		pidx++;
4908 
4909 		/* map in the rest of the pages if required */
4910 		if (window->wd_remap_copybuf) {
4911 			window->wd_remap_copybuf = B_FALSE;
4912 
4913 			/* figure out many pages this window takes up */
4914 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4915 			    MMU_PAGEOFFSET;
4916 			pcnt = mmu_btopr(window->wd_size + poff);
4917 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4918 
4919 			/* map pages which require it */
4920 			for (i = 1; i < pcnt; i++) {
4921 				pmap = &dma->dp_pgmap[pidx];
4922 				if (pmap->pm_uses_copybuf) {
4923 					ASSERT(pmap->pm_mapped == B_FALSE);
4924 					if (pmap->pm_pp != NULL) {
4925 						pmap->pm_mapped = B_TRUE;
4926 						i86_pp_map(pmap->pm_pp,
4927 						    pmap->pm_kaddr);
4928 					} else if (pmap->pm_vaddr != NULL) {
4929 						pmap->pm_mapped = B_TRUE;
4930 						i86_va_map(pmap->pm_vaddr,
4931 						    sinfo->si_asp,
4932 						    pmap->pm_kaddr);
4933 					}
4934 				}
4935 				pidx++;
4936 			}
4937 		}
4938 	}
4939 #endif
4940 
4941 	/* if the new window uses the copy buffer, sync it for the device */
4942 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4943 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4944 		    DDI_DMA_SYNC_FORDEV);
4945 	}
4946 
4947 	return (DDI_SUCCESS);
4948 }
4949 
4950 /*
4951  * rootnex_dma_win()
4952  *    called from ddi_dma_getwin()
4953  */
4954 /*ARGSUSED*/
4955 static int
4956 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4957     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4958     uint_t *ccountp)
4959 {
4960 #if defined(__amd64) && !defined(__xpv)
4961 	if (IOMMU_USED(rdip)) {
4962 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4963 		    cookiep, ccountp));
4964 	}
4965 #endif
4966 
4967 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4968 	    cookiep, ccountp));
4969 }
4970 
4971 #if defined(__amd64) && !defined(__xpv)
4972 /*ARGSUSED*/
4973 static int
4974 rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
4975     ddi_dma_handle_t handle, void *v)
4976 {
4977 	ddi_dma_impl_t *hp;
4978 	rootnex_dma_t *dma;
4979 
4980 	hp = (ddi_dma_impl_t *)handle;
4981 	dma = (rootnex_dma_t *)hp->dmai_private;
4982 	dma->dp_iommu_private = v;
4983 
4984 	return (DDI_SUCCESS);
4985 }
4986 
4987 /*ARGSUSED*/
4988 static void *
4989 rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
4990     ddi_dma_handle_t handle)
4991 {
4992 	ddi_dma_impl_t *hp;
4993 	rootnex_dma_t *dma;
4994 
4995 	hp = (ddi_dma_impl_t *)handle;
4996 	dma = (rootnex_dma_t *)hp->dmai_private;
4997 
4998 	return (dma->dp_iommu_private);
4999 }
5000 #endif
5001 
5002 /*
5003  * ************************
5004  *  obsoleted dma routines
5005  * ************************
5006  */
5007 
5008 /*
5009  * rootnex_dma_map()
5010  *    called from ddi_dma_setup()
5011  * NO IOMMU in 32 bit mode. The below routines doesn't work in 64 bit mode.
5012  */
5013 /* ARGSUSED */
5014 static int
5015 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
5016     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
5017 {
5018 #if defined(__amd64)
5019 	/*
5020 	 * this interface is not supported in 64-bit x86 kernel. See comment in
5021 	 * rootnex_dma_mctl()
5022 	 */
5023 	return (DDI_DMA_NORESOURCES);
5024 
5025 #else /* 32-bit x86 kernel */
5026 	ddi_dma_handle_t *lhandlep;
5027 	ddi_dma_handle_t lhandle;
5028 	ddi_dma_cookie_t cookie;
5029 	ddi_dma_attr_t dma_attr;
5030 	ddi_dma_lim_t *dma_lim;
5031 	uint_t ccnt;
5032 	int e;
5033 
5034 
5035 	/*
5036 	 * if the driver is just testing to see if it's possible to do the bind,
5037 	 * we'll use local state. Otherwise, use the handle pointer passed in.
5038 	 */
5039 	if (handlep == NULL) {
5040 		lhandlep = &lhandle;
5041 	} else {
5042 		lhandlep = handlep;
5043 	}
5044 
5045 	/* convert the limit structure to a dma_attr one */
5046 	dma_lim = dmareq->dmar_limits;
5047 	dma_attr.dma_attr_version = DMA_ATTR_V0;
5048 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
5049 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
5050 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
5051 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
5052 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
5053 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
5054 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
5055 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
5056 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
5057 	dma_attr.dma_attr_align = MMU_PAGESIZE;
5058 	dma_attr.dma_attr_flags = 0;
5059 
5060 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
5061 	    dmareq->dmar_arg, lhandlep);
5062 	if (e != DDI_SUCCESS) {
5063 		return (e);
5064 	}
5065 
5066 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
5067 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
5068 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
5069 		return (e);
5070 	}
5071 
5072 	/*
5073 	 * if the driver is just testing to see if it's possible to do the bind,
5074 	 * free up the local state and return the result.
5075 	 */
5076 	if (handlep == NULL) {
5077 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
5078 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
5079 		if (e == DDI_DMA_MAPPED) {
5080 			return (DDI_DMA_MAPOK);
5081 		} else {
5082 			return (DDI_DMA_NOMAPPING);
5083 		}
5084 	}
5085 
5086 	return (e);
5087 #endif /* defined(__amd64) */
5088 }
5089 
5090 /*
5091  * rootnex_dma_mctl()
5092  *
5093  * No IOMMU in 32 bit mode. The below routine doesn't work in 64 bit mode.
5094  */
5095 /* ARGSUSED */
5096 static int
5097 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
5098     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
5099     uint_t cache_flags)
5100 {
5101 #if defined(__amd64)
5102 	/*
5103 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
5104 	 * common implementation in genunix, so they no longer have x86
5105 	 * specific functionality which called into dma_ctl.
5106 	 *
5107 	 * The rest of the obsoleted interfaces were never supported in the
5108 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
5109 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
5110 	 * implementation issues.
5111 	 *
5112 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
5113 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
5114 	 * reflect that now too...
5115 	 *
5116 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
5117 	 * not going to put this functionality into the 64-bit x86 kernel now.
5118 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
5119 	 * that in a future release.
5120 	 */
5121 	return (DDI_FAILURE);
5122 
5123 #else /* 32-bit x86 kernel */
5124 	ddi_dma_cookie_t lcookie;
5125 	ddi_dma_cookie_t *cookie;
5126 	rootnex_window_t *window;
5127 	ddi_dma_impl_t *hp;
5128 	rootnex_dma_t *dma;
5129 	uint_t nwin;
5130 	uint_t ccnt;
5131 	size_t len;
5132 	off_t off;
5133 	int e;
5134 
5135 
5136 	/*
5137 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
5138 	 * hacky since were optimizing for the current interfaces and so we can
5139 	 * cleanup the mess in genunix. Hopefully we will remove the this
5140 	 * obsoleted routines someday soon.
5141 	 */
5142 
5143 	switch (request) {
5144 
5145 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
5146 		hp = (ddi_dma_impl_t *)handle;
5147 		cookie = (ddi_dma_cookie_t *)objpp;
5148 
5149 		/*
5150 		 * convert segment to cookie. We don't distinguish between the
5151 		 * two :-)
5152 		 */
5153 		*cookie = *hp->dmai_cookie;
5154 		*lenp = cookie->dmac_size;
5155 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
5156 		return (DDI_SUCCESS);
5157 
5158 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
5159 		hp = (ddi_dma_impl_t *)handle;
5160 		dma = (rootnex_dma_t *)hp->dmai_private;
5161 
5162 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
5163 			return (DDI_DMA_STALE);
5164 		}
5165 
5166 		/* handle the case where we don't have any windows */
5167 		if (dma->dp_window == NULL) {
5168 			/*
5169 			 * if seg == NULL, and we don't have any windows,
5170 			 * return the first cookie in the sgl.
5171 			 */
5172 			if (*lenp == NULL) {
5173 				dma->dp_current_cookie = 0;
5174 				hp->dmai_cookie = dma->dp_cookies;
5175 				*objpp = (caddr_t)handle;
5176 				return (DDI_SUCCESS);
5177 
5178 			/* if we have more cookies, go to the next cookie */
5179 			} else {
5180 				if ((dma->dp_current_cookie + 1) >=
5181 				    dma->dp_sglinfo.si_sgl_size) {
5182 					return (DDI_DMA_DONE);
5183 				}
5184 				dma->dp_current_cookie++;
5185 				hp->dmai_cookie++;
5186 				return (DDI_SUCCESS);
5187 			}
5188 		}
5189 
5190 		/* We have one or more windows */
5191 		window = &dma->dp_window[dma->dp_current_win];
5192 
5193 		/*
5194 		 * if seg == NULL, return the first cookie in the current
5195 		 * window
5196 		 */
5197 		if (*lenp == NULL) {
5198 			dma->dp_current_cookie = 0;
5199 			hp->dmai_cookie = window->wd_first_cookie;
5200 
5201 		/*
5202 		 * go to the next cookie in the window then see if we done with
5203 		 * this window.
5204 		 */
5205 		} else {
5206 			if ((dma->dp_current_cookie + 1) >=
5207 			    window->wd_cookie_cnt) {
5208 				return (DDI_DMA_DONE);
5209 			}
5210 			dma->dp_current_cookie++;
5211 			hp->dmai_cookie++;
5212 		}
5213 		*objpp = (caddr_t)handle;
5214 		return (DDI_SUCCESS);
5215 
5216 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
5217 		hp = (ddi_dma_impl_t *)handle;
5218 		dma = (rootnex_dma_t *)hp->dmai_private;
5219 
5220 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
5221 			return (DDI_DMA_STALE);
5222 		}
5223 
5224 		/* if win == NULL, return the first window in the bind */
5225 		if (*offp == NULL) {
5226 			nwin = 0;
5227 
5228 		/*
5229 		 * else, go to the next window then see if we're done with all
5230 		 * the windows.
5231 		 */
5232 		} else {
5233 			nwin = dma->dp_current_win + 1;
5234 			if (nwin >= hp->dmai_nwin) {
5235 				return (DDI_DMA_DONE);
5236 			}
5237 		}
5238 
5239 		/* switch to the next window */
5240 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
5241 		    &lcookie, &ccnt);
5242 		ASSERT(e == DDI_SUCCESS);
5243 		if (e != DDI_SUCCESS) {
5244 			return (DDI_DMA_STALE);
5245 		}
5246 
5247 		/* reset the cookie back to the first cookie in the window */
5248 		if (dma->dp_window != NULL) {
5249 			window = &dma->dp_window[dma->dp_current_win];
5250 			hp->dmai_cookie = window->wd_first_cookie;
5251 		} else {
5252 			hp->dmai_cookie = dma->dp_cookies;
5253 		}
5254 
5255 		*objpp = (caddr_t)handle;
5256 		return (DDI_SUCCESS);
5257 
5258 	case DDI_DMA_FREE: /* ddi_dma_free() */
5259 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
5260 		(void) rootnex_dma_freehdl(dip, rdip, handle);
5261 		if (rootnex_state->r_dvma_call_list_id) {
5262 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
5263 		}
5264 		return (DDI_SUCCESS);
5265 
5266 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
5267 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
5268 		/* should never get here, handled in genunix */
5269 		ASSERT(0);
5270 		return (DDI_FAILURE);
5271 
5272 	case DDI_DMA_KVADDR:
5273 	case DDI_DMA_GETERR:
5274 	case DDI_DMA_COFF:
5275 		return (DDI_FAILURE);
5276 	}
5277 
5278 	return (DDI_FAILURE);
5279 #endif /* defined(__amd64) */
5280 }
5281 
5282 /*
5283  * *********
5284  *  FMA Code
5285  * *********
5286  */
5287 
5288 /*
5289  * rootnex_fm_init()
5290  *    FMA init busop
5291  */
5292 /* ARGSUSED */
5293 static int
5294 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
5295     ddi_iblock_cookie_t *ibc)
5296 {
5297 	*ibc = rootnex_state->r_err_ibc;
5298 
5299 	return (ddi_system_fmcap);
5300 }
5301 
5302 /*
5303  * rootnex_dma_check()
5304  *    Function called after a dma fault occurred to find out whether the
5305  *    fault address is associated with a driver that is able to handle faults
5306  *    and recover from faults.
5307  */
5308 /* ARGSUSED */
5309 static int
5310 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
5311     const void *not_used)
5312 {
5313 	rootnex_window_t *window;
5314 	uint64_t start_addr;
5315 	uint64_t fault_addr;
5316 	ddi_dma_impl_t *hp;
5317 	rootnex_dma_t *dma;
5318 	uint64_t end_addr;
5319 	size_t csize;
5320 	int i;
5321 	int j;
5322 
5323 
5324 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5325 	hp = (ddi_dma_impl_t *)handle;
5326 	ASSERT(hp);
5327 
5328 	dma = (rootnex_dma_t *)hp->dmai_private;
5329 
5330 	/* Get the address that we need to search for */
5331 	fault_addr = *(uint64_t *)addr;
5332 
5333 	/*
5334 	 * if we don't have any windows, we can just walk through all the
5335 	 * cookies.
5336 	 */
5337 	if (dma->dp_window == NULL) {
5338 		/* for each cookie */
5339 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5340 			/*
5341 			 * if the faulted address is within the physical address
5342 			 * range of the cookie, return DDI_FM_NONFATAL.
5343 			 */
5344 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5345 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5346 			    dma->dp_cookies[i].dmac_size))) {
5347 				return (DDI_FM_NONFATAL);
5348 			}
5349 		}
5350 
5351 		/* fault_addr not within this DMA handle */
5352 		return (DDI_FM_UNKNOWN);
5353 	}
5354 
5355 	/* we have mutiple windows, walk through each window */
5356 	for (i = 0; i < hp->dmai_nwin; i++) {
5357 		window = &dma->dp_window[i];
5358 
5359 		/* Go through all the cookies in the window */
5360 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5361 
5362 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5363 			csize = window->wd_first_cookie[j].dmac_size;
5364 
5365 			/*
5366 			 * if we are trimming the first cookie in the window,
5367 			 * and this is the first cookie, adjust the start
5368 			 * address and size of the cookie to account for the
5369 			 * trim.
5370 			 */
5371 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5372 				start_addr = window->wd_trim.tr_first_paddr;
5373 				csize = window->wd_trim.tr_first_size;
5374 			}
5375 
5376 			/*
5377 			 * if we are trimming the last cookie in the window,
5378 			 * and this is the last cookie, adjust the start
5379 			 * address and size of the cookie to account for the
5380 			 * trim.
5381 			 */
5382 			if (window->wd_trim.tr_trim_last &&
5383 			    (j == (window->wd_cookie_cnt - 1))) {
5384 				start_addr = window->wd_trim.tr_last_paddr;
5385 				csize = window->wd_trim.tr_last_size;
5386 			}
5387 
5388 			end_addr = start_addr + csize;
5389 
5390 			/*
5391 			 * if the faulted address is within the physical
5392 			 * address of the cookie, return DDI_FM_NONFATAL.
5393 			 */
5394 			if ((fault_addr >= start_addr) &&
5395 			    (fault_addr <= end_addr)) {
5396 				return (DDI_FM_NONFATAL);
5397 			}
5398 		}
5399 	}
5400 
5401 	/* fault_addr not within this DMA handle */
5402 	return (DDI_FM_UNKNOWN);
5403 }
5404 
5405 /*ARGSUSED*/
5406 static int
5407 rootnex_quiesce(dev_info_t *dip)
5408 {
5409 #if defined(__amd64) && !defined(__xpv)
5410 	return (immu_quiesce());
5411 #else
5412 	return (DDI_SUCCESS);
5413 #endif
5414 }
5415 
5416 #if defined(__xpv)
5417 void
5418 immu_init(void)
5419 {
5420 	;
5421 }
5422 
5423 void
5424 immu_startup(void)
5425 {
5426 	;
5427 }
5428 /*ARGSUSED*/
5429 void
5430 immu_physmem_update(uint64_t addr, uint64_t size)
5431 {
5432 	;
5433 }
5434 #endif
5435