xref: /titanic_51/usr/src/uts/i86pc/io/rootnex.c (revision 15e1afcd5e908ae29b1e6018838638befdc225a2)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
26  * Copyright (c) 2011 Bayard G. Bell.  All rights reserved.
27  */
28 
29 /*
30  * x86 root nexus driver
31  */
32 
33 #include <sys/sysmacros.h>
34 #include <sys/conf.h>
35 #include <sys/autoconf.h>
36 #include <sys/sysmacros.h>
37 #include <sys/debug.h>
38 #include <sys/psw.h>
39 #include <sys/ddidmareq.h>
40 #include <sys/promif.h>
41 #include <sys/devops.h>
42 #include <sys/kmem.h>
43 #include <sys/cmn_err.h>
44 #include <vm/seg.h>
45 #include <vm/seg_kmem.h>
46 #include <vm/seg_dev.h>
47 #include <sys/vmem.h>
48 #include <sys/mman.h>
49 #include <vm/hat.h>
50 #include <vm/as.h>
51 #include <vm/page.h>
52 #include <sys/avintr.h>
53 #include <sys/errno.h>
54 #include <sys/modctl.h>
55 #include <sys/ddi_impldefs.h>
56 #include <sys/sunddi.h>
57 #include <sys/sunndi.h>
58 #include <sys/mach_intr.h>
59 #include <sys/psm.h>
60 #include <sys/ontrap.h>
61 #include <sys/atomic.h>
62 #include <sys/sdt.h>
63 #include <sys/rootnex.h>
64 #include <vm/hat_i86.h>
65 #include <sys/ddifm.h>
66 #include <sys/ddi_isa.h>
67 #include <sys/apic.h>
68 
69 #ifdef __xpv
70 #include <sys/bootinfo.h>
71 #include <sys/hypervisor.h>
72 #include <sys/bootconf.h>
73 #include <vm/kboot_mmu.h>
74 #endif
75 
76 #if defined(__amd64) && !defined(__xpv)
77 #include <sys/immu.h>
78 #endif
79 
80 
81 /*
82  * enable/disable extra checking of function parameters. Useful for debugging
83  * drivers.
84  */
85 #ifdef	DEBUG
86 int rootnex_alloc_check_parms = 1;
87 int rootnex_bind_check_parms = 1;
88 int rootnex_bind_check_inuse = 1;
89 int rootnex_unbind_verify_buffer = 0;
90 int rootnex_sync_check_parms = 1;
91 #else
92 int rootnex_alloc_check_parms = 0;
93 int rootnex_bind_check_parms = 0;
94 int rootnex_bind_check_inuse = 0;
95 int rootnex_unbind_verify_buffer = 0;
96 int rootnex_sync_check_parms = 0;
97 #endif
98 
99 boolean_t rootnex_dmar_not_setup;
100 
101 /* Master Abort and Target Abort panic flag */
102 int rootnex_fm_ma_ta_panic_flag = 0;
103 
104 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
105 int rootnex_bind_fail = 1;
106 int rootnex_bind_warn = 1;
107 uint8_t *rootnex_warn_list;
108 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
109 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
110 
111 /*
112  * revert back to old broken behavior of always sync'ing entire copy buffer.
113  * This is useful if be have a buggy driver which doesn't correctly pass in
114  * the offset and size into ddi_dma_sync().
115  */
116 int rootnex_sync_ignore_params = 0;
117 
118 /*
119  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
120  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
121  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
122  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
123  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
124  * (< 8K). We will still need to allocate the copy buffer during bind though
125  * (if we need one). These can only be modified in /etc/system before rootnex
126  * attach.
127  */
128 #if defined(__amd64)
129 int rootnex_prealloc_cookies = 65;
130 int rootnex_prealloc_windows = 4;
131 int rootnex_prealloc_copybuf = 2;
132 #else
133 int rootnex_prealloc_cookies = 33;
134 int rootnex_prealloc_windows = 4;
135 int rootnex_prealloc_copybuf = 2;
136 #endif
137 
138 /* driver global state */
139 static rootnex_state_t *rootnex_state;
140 
141 #ifdef DEBUG
142 /* shortcut to rootnex counters */
143 static uint64_t *rootnex_cnt;
144 #endif
145 
146 /*
147  * XXX - does x86 even need these or are they left over from the SPARC days?
148  */
149 /* statically defined integer/boolean properties for the root node */
150 static rootnex_intprop_t rootnex_intprp[] = {
151 	{ "PAGESIZE",			PAGESIZE },
152 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
153 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
154 	{ DDI_RELATIVE_ADDRESSING,	1 },
155 };
156 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
157 
158 /*
159  * If we're dom0, we're using a real device so we need to load
160  * the cookies with MFNs instead of PFNs.
161  */
162 #ifdef __xpv
163 typedef maddr_t rootnex_addr_t;
164 #define	ROOTNEX_PADDR_TO_RBASE(pa)	\
165 	(DOMAIN_IS_INITDOMAIN(xen_info) ? pa_to_ma(pa) : (pa))
166 #else
167 typedef paddr_t rootnex_addr_t;
168 #define	ROOTNEX_PADDR_TO_RBASE(pa)	(pa)
169 #endif
170 
171 static struct cb_ops rootnex_cb_ops = {
172 	nodev,		/* open */
173 	nodev,		/* close */
174 	nodev,		/* strategy */
175 	nodev,		/* print */
176 	nodev,		/* dump */
177 	nodev,		/* read */
178 	nodev,		/* write */
179 	nodev,		/* ioctl */
180 	nodev,		/* devmap */
181 	nodev,		/* mmap */
182 	nodev,		/* segmap */
183 	nochpoll,	/* chpoll */
184 	ddi_prop_op,	/* cb_prop_op */
185 	NULL,		/* struct streamtab */
186 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
187 	CB_REV,		/* Rev */
188 	nodev,		/* cb_aread */
189 	nodev		/* cb_awrite */
190 };
191 
192 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
193     off_t offset, off_t len, caddr_t *vaddrp);
194 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
195     struct hat *hat, struct seg *seg, caddr_t addr,
196     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
197 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
198     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
199 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
200     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
201     ddi_dma_handle_t *handlep);
202 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
203     ddi_dma_handle_t handle);
204 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
205     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
206     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
207 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
208     ddi_dma_handle_t handle);
209 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
210     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
211 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
212     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
213     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
214 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
215     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
216     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
217 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
218     ddi_ctl_enum_t ctlop, void *arg, void *result);
219 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
220     ddi_iblock_cookie_t *ibc);
221 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
222     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
223 static int rootnex_alloc_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *,
224     void *);
225 static int rootnex_free_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *);
226 
227 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
228     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
229     ddi_dma_handle_t *handlep);
230 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
231     ddi_dma_handle_t handle);
232 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
233     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
234     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
235 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
236     ddi_dma_handle_t handle);
237 #if defined(__amd64) && !defined(__xpv)
238 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
239     ddi_dma_handle_t handle);
240 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
241     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
242 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
243     ddi_dma_cookie_t *cookiep, uint_t ccount);
244 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
245     ddi_dma_handle_t handle);
246 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
247 #endif
248 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
249     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
250 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
251     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
252     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
253 
254 #if defined(__amd64) && !defined(__xpv)
255 static int rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
256     ddi_dma_handle_t handle, void *v);
257 static void *rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
258     ddi_dma_handle_t handle);
259 #endif
260 
261 
262 static struct bus_ops rootnex_bus_ops = {
263 	BUSO_REV,
264 	rootnex_map,
265 	NULL,
266 	NULL,
267 	NULL,
268 	rootnex_map_fault,
269 	rootnex_dma_map,
270 	rootnex_dma_allochdl,
271 	rootnex_dma_freehdl,
272 	rootnex_dma_bindhdl,
273 	rootnex_dma_unbindhdl,
274 	rootnex_dma_sync,
275 	rootnex_dma_win,
276 	rootnex_dma_mctl,
277 	rootnex_ctlops,
278 	ddi_bus_prop_op,
279 	i_ddi_rootnex_get_eventcookie,
280 	i_ddi_rootnex_add_eventcall,
281 	i_ddi_rootnex_remove_eventcall,
282 	i_ddi_rootnex_post_event,
283 	0,			/* bus_intr_ctl */
284 	0,			/* bus_config */
285 	0,			/* bus_unconfig */
286 	rootnex_fm_init,	/* bus_fm_init */
287 	NULL,			/* bus_fm_fini */
288 	NULL,			/* bus_fm_access_enter */
289 	NULL,			/* bus_fm_access_exit */
290 	NULL,			/* bus_powr */
291 	rootnex_intr_ops	/* bus_intr_op */
292 };
293 
294 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
295 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
296 static int rootnex_quiesce(dev_info_t *dip);
297 
298 static struct dev_ops rootnex_ops = {
299 	DEVO_REV,
300 	0,
301 	ddi_no_info,
302 	nulldev,
303 	nulldev,
304 	rootnex_attach,
305 	rootnex_detach,
306 	nulldev,
307 	&rootnex_cb_ops,
308 	&rootnex_bus_ops,
309 	NULL,
310 	rootnex_quiesce,		/* quiesce */
311 };
312 
313 static struct modldrv rootnex_modldrv = {
314 	&mod_driverops,
315 	"i86pc root nexus",
316 	&rootnex_ops
317 };
318 
319 static struct modlinkage rootnex_modlinkage = {
320 	MODREV_1,
321 	(void *)&rootnex_modldrv,
322 	NULL
323 };
324 
325 #if defined(__amd64) && !defined(__xpv)
326 static iommulib_nexops_t iommulib_nexops = {
327 	IOMMU_NEXOPS_VERSION,
328 	"Rootnex IOMMU ops Vers 1.1",
329 	NULL,
330 	rootnex_coredma_allochdl,
331 	rootnex_coredma_freehdl,
332 	rootnex_coredma_bindhdl,
333 	rootnex_coredma_unbindhdl,
334 	rootnex_coredma_reset_cookies,
335 	rootnex_coredma_get_cookies,
336 	rootnex_coredma_set_cookies,
337 	rootnex_coredma_clear_cookies,
338 	rootnex_coredma_get_sleep_flags,
339 	rootnex_coredma_sync,
340 	rootnex_coredma_win,
341 	rootnex_dma_map,
342 	rootnex_dma_mctl,
343 	rootnex_coredma_hdl_setprivate,
344 	rootnex_coredma_hdl_getprivate
345 };
346 #endif
347 
348 /*
349  *  extern hacks
350  */
351 extern struct seg_ops segdev_ops;
352 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
353 #ifdef	DDI_MAP_DEBUG
354 extern int ddi_map_debug_flag;
355 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
356 #endif
357 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
358 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
359 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
360     psm_intr_op_t, int *);
361 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
362 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
363 
364 /*
365  * Use device arena to use for device control register mappings.
366  * Various kernel memory walkers (debugger, dtrace) need to know
367  * to avoid this address range to prevent undesired device activity.
368  */
369 extern void *device_arena_alloc(size_t size, int vm_flag);
370 extern void device_arena_free(void * vaddr, size_t size);
371 
372 
373 /*
374  *  Internal functions
375  */
376 static int rootnex_dma_init();
377 static void rootnex_add_props(dev_info_t *);
378 static int rootnex_ctl_reportdev(dev_info_t *dip);
379 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
380 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
381 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
382 static int rootnex_map_handle(ddi_map_req_t *mp);
383 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
384 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
385 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
386     ddi_dma_attr_t *attr);
387 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
388     rootnex_sglinfo_t *sglinfo);
389 static void rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object,
390     ddi_dma_cookie_t *sgl, rootnex_sglinfo_t *sglinfo);
391 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
392     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
393 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
394     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
395 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
396 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
397     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
398 static void rootnex_teardown_windows(rootnex_dma_t *dma);
399 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
400     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
401 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
402     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
403     size_t *copybuf_used, page_t **cur_pp);
404 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
405     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
406     ddi_dma_attr_t *attr, off_t cur_offset);
407 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
408     rootnex_dma_t *dma, rootnex_window_t **windowp,
409     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
410 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
411     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
412 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
413     off_t offset, size_t size, uint_t cache_flags);
414 static int rootnex_verify_buffer(rootnex_dma_t *dma);
415 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
416     const void *comp_addr, const void *not_used);
417 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
418     rootnex_sglinfo_t *sglinfo);
419 static struct as *rootnex_get_as(ddi_dma_obj_t *dmar_object);
420 
421 /*
422  * _init()
423  *
424  */
425 int
426 _init(void)
427 {
428 
429 	rootnex_state = NULL;
430 	return (mod_install(&rootnex_modlinkage));
431 }
432 
433 
434 /*
435  * _info()
436  *
437  */
438 int
439 _info(struct modinfo *modinfop)
440 {
441 	return (mod_info(&rootnex_modlinkage, modinfop));
442 }
443 
444 
445 /*
446  * _fini()
447  *
448  */
449 int
450 _fini(void)
451 {
452 	return (EBUSY);
453 }
454 
455 
456 /*
457  * rootnex_attach()
458  *
459  */
460 static int
461 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
462 {
463 	int fmcap;
464 	int e;
465 
466 	switch (cmd) {
467 	case DDI_ATTACH:
468 		break;
469 	case DDI_RESUME:
470 #if defined(__amd64) && !defined(__xpv)
471 		return (immu_unquiesce());
472 #else
473 		return (DDI_SUCCESS);
474 #endif
475 	default:
476 		return (DDI_FAILURE);
477 	}
478 
479 	/*
480 	 * We should only have one instance of rootnex. Save it away since we
481 	 * don't have an easy way to get it back later.
482 	 */
483 	ASSERT(rootnex_state == NULL);
484 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
485 
486 	rootnex_state->r_dip = dip;
487 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
488 	rootnex_state->r_reserved_msg_printed = B_FALSE;
489 #ifdef DEBUG
490 	rootnex_cnt = &rootnex_state->r_counters[0];
491 #endif
492 
493 	/*
494 	 * Set minimum fm capability level for i86pc platforms and then
495 	 * initialize error handling. Since we're the rootnex, we don't
496 	 * care what's returned in the fmcap field.
497 	 */
498 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
499 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
500 	fmcap = ddi_system_fmcap;
501 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
502 
503 	/* initialize DMA related state */
504 	e = rootnex_dma_init();
505 	if (e != DDI_SUCCESS) {
506 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
507 		return (DDI_FAILURE);
508 	}
509 
510 	/* Add static root node properties */
511 	rootnex_add_props(dip);
512 
513 	/* since we can't call ddi_report_dev() */
514 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
515 
516 	/* Initialize rootnex event handle */
517 	i_ddi_rootnex_init_events(dip);
518 
519 #if defined(__amd64) && !defined(__xpv)
520 	e = iommulib_nexus_register(dip, &iommulib_nexops,
521 	    &rootnex_state->r_iommulib_handle);
522 
523 	ASSERT(e == DDI_SUCCESS);
524 #endif
525 
526 	return (DDI_SUCCESS);
527 }
528 
529 
530 /*
531  * rootnex_detach()
532  *
533  */
534 /*ARGSUSED*/
535 static int
536 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
537 {
538 	switch (cmd) {
539 	case DDI_SUSPEND:
540 #if defined(__amd64) && !defined(__xpv)
541 		return (immu_quiesce());
542 #else
543 		return (DDI_SUCCESS);
544 #endif
545 	default:
546 		return (DDI_FAILURE);
547 	}
548 	/*NOTREACHED*/
549 
550 }
551 
552 
553 /*
554  * rootnex_dma_init()
555  *
556  */
557 /*ARGSUSED*/
558 static int
559 rootnex_dma_init()
560 {
561 	size_t bufsize;
562 
563 
564 	/*
565 	 * size of our cookie/window/copybuf state needed in dma bind that we
566 	 * pre-alloc in dma_alloc_handle
567 	 */
568 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
569 	rootnex_state->r_prealloc_size =
570 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
571 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
572 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
573 
574 	/*
575 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
576 	 * allocate 16 extra bytes for struct pointer alignment
577 	 * (p->dmai_private & dma->dp_prealloc_buffer)
578 	 */
579 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
580 	    rootnex_state->r_prealloc_size + 0x10;
581 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
582 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
583 	if (rootnex_state->r_dmahdl_cache == NULL) {
584 		return (DDI_FAILURE);
585 	}
586 
587 	/*
588 	 * allocate array to track which major numbers we have printed warnings
589 	 * for.
590 	 */
591 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
592 	    KM_SLEEP);
593 
594 	return (DDI_SUCCESS);
595 }
596 
597 
598 /*
599  * rootnex_add_props()
600  *
601  */
602 static void
603 rootnex_add_props(dev_info_t *dip)
604 {
605 	rootnex_intprop_t *rpp;
606 	int i;
607 
608 	/* Add static integer/boolean properties to the root node */
609 	rpp = rootnex_intprp;
610 	for (i = 0; i < NROOT_INTPROPS; i++) {
611 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
612 		    rpp[i].prop_name, rpp[i].prop_value);
613 	}
614 }
615 
616 
617 
618 /*
619  * *************************
620  *  ctlops related routines
621  * *************************
622  */
623 
624 /*
625  * rootnex_ctlops()
626  *
627  */
628 /*ARGSUSED*/
629 static int
630 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
631     void *arg, void *result)
632 {
633 	int n, *ptr;
634 	struct ddi_parent_private_data *pdp;
635 
636 	switch (ctlop) {
637 	case DDI_CTLOPS_DMAPMAPC:
638 		/*
639 		 * Return 'partial' to indicate that dma mapping
640 		 * has to be done in the main MMU.
641 		 */
642 		return (DDI_DMA_PARTIAL);
643 
644 	case DDI_CTLOPS_BTOP:
645 		/*
646 		 * Convert byte count input to physical page units.
647 		 * (byte counts that are not a page-size multiple
648 		 * are rounded down)
649 		 */
650 		*(ulong_t *)result = btop(*(ulong_t *)arg);
651 		return (DDI_SUCCESS);
652 
653 	case DDI_CTLOPS_PTOB:
654 		/*
655 		 * Convert size in physical pages to bytes
656 		 */
657 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
658 		return (DDI_SUCCESS);
659 
660 	case DDI_CTLOPS_BTOPR:
661 		/*
662 		 * Convert byte count input to physical page units
663 		 * (byte counts that are not a page-size multiple
664 		 * are rounded up)
665 		 */
666 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
667 		return (DDI_SUCCESS);
668 
669 	case DDI_CTLOPS_INITCHILD:
670 		return (impl_ddi_sunbus_initchild(arg));
671 
672 	case DDI_CTLOPS_UNINITCHILD:
673 		impl_ddi_sunbus_removechild(arg);
674 		return (DDI_SUCCESS);
675 
676 	case DDI_CTLOPS_REPORTDEV:
677 		return (rootnex_ctl_reportdev(rdip));
678 
679 	case DDI_CTLOPS_IOMIN:
680 		/*
681 		 * Nothing to do here but reflect back..
682 		 */
683 		return (DDI_SUCCESS);
684 
685 	case DDI_CTLOPS_REGSIZE:
686 	case DDI_CTLOPS_NREGS:
687 		break;
688 
689 	case DDI_CTLOPS_SIDDEV:
690 		if (ndi_dev_is_prom_node(rdip))
691 			return (DDI_SUCCESS);
692 		if (ndi_dev_is_persistent_node(rdip))
693 			return (DDI_SUCCESS);
694 		return (DDI_FAILURE);
695 
696 	case DDI_CTLOPS_POWER:
697 		return ((*pm_platform_power)((power_req_t *)arg));
698 
699 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
700 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
701 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
702 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
703 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
704 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
705 		if (!rootnex_state->r_reserved_msg_printed) {
706 			rootnex_state->r_reserved_msg_printed = B_TRUE;
707 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
708 			    "1 or more reserved/obsolete operations.");
709 		}
710 		return (DDI_FAILURE);
711 
712 	default:
713 		return (DDI_FAILURE);
714 	}
715 	/*
716 	 * The rest are for "hardware" properties
717 	 */
718 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
719 		return (DDI_FAILURE);
720 
721 	if (ctlop == DDI_CTLOPS_NREGS) {
722 		ptr = (int *)result;
723 		*ptr = pdp->par_nreg;
724 	} else {
725 		off_t *size = (off_t *)result;
726 
727 		ptr = (int *)arg;
728 		n = *ptr;
729 		if (n >= pdp->par_nreg) {
730 			return (DDI_FAILURE);
731 		}
732 		*size = (off_t)pdp->par_reg[n].regspec_size;
733 	}
734 	return (DDI_SUCCESS);
735 }
736 
737 
738 /*
739  * rootnex_ctl_reportdev()
740  *
741  */
742 static int
743 rootnex_ctl_reportdev(dev_info_t *dev)
744 {
745 	int i, n, len, f_len = 0;
746 	char *buf;
747 
748 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
749 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
750 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
751 	len = strlen(buf);
752 
753 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
754 
755 		struct regspec *rp = sparc_pd_getreg(dev, i);
756 
757 		if (i == 0)
758 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
759 			    ": ");
760 		else
761 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
762 			    " and ");
763 		len = strlen(buf);
764 
765 		switch (rp->regspec_bustype) {
766 
767 		case BTEISA:
768 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
769 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
770 			break;
771 
772 		case BTISA:
773 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
774 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
775 			break;
776 
777 		default:
778 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
779 			    "space %x offset %x",
780 			    rp->regspec_bustype, rp->regspec_addr);
781 			break;
782 		}
783 		len = strlen(buf);
784 	}
785 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
786 		int pri;
787 
788 		if (i != 0) {
789 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
790 			    ",");
791 			len = strlen(buf);
792 		}
793 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
794 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
795 		    " sparc ipl %d", pri);
796 		len = strlen(buf);
797 	}
798 #ifdef DEBUG
799 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
800 		cmn_err(CE_NOTE, "next message is truncated: "
801 		    "printed length 1024, real length %d", f_len);
802 	}
803 #endif /* DEBUG */
804 	cmn_err(CE_CONT, "?%s\n", buf);
805 	kmem_free(buf, REPORTDEV_BUFSIZE);
806 	return (DDI_SUCCESS);
807 }
808 
809 
810 /*
811  * ******************
812  *  map related code
813  * ******************
814  */
815 
816 /*
817  * rootnex_map()
818  *
819  */
820 static int
821 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
822     off_t len, caddr_t *vaddrp)
823 {
824 	struct regspec *rp, tmp_reg;
825 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
826 	int error;
827 
828 	mp = &mr;
829 
830 	switch (mp->map_op)  {
831 	case DDI_MO_MAP_LOCKED:
832 	case DDI_MO_UNMAP:
833 	case DDI_MO_MAP_HANDLE:
834 		break;
835 	default:
836 #ifdef	DDI_MAP_DEBUG
837 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
838 		    mp->map_op);
839 #endif	/* DDI_MAP_DEBUG */
840 		return (DDI_ME_UNIMPLEMENTED);
841 	}
842 
843 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
844 #ifdef	DDI_MAP_DEBUG
845 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
846 #endif	/* DDI_MAP_DEBUG */
847 		return (DDI_ME_UNIMPLEMENTED);
848 	}
849 
850 	/*
851 	 * First, if given an rnumber, convert it to a regspec...
852 	 * (Presumably, this is on behalf of a child of the root node?)
853 	 */
854 
855 	if (mp->map_type == DDI_MT_RNUMBER)  {
856 
857 		int rnumber = mp->map_obj.rnumber;
858 #ifdef	DDI_MAP_DEBUG
859 		static char *out_of_range =
860 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
861 #endif	/* DDI_MAP_DEBUG */
862 
863 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
864 		if (rp == NULL)  {
865 #ifdef	DDI_MAP_DEBUG
866 			cmn_err(CE_WARN, out_of_range, rnumber,
867 			    ddi_get_name(rdip));
868 #endif	/* DDI_MAP_DEBUG */
869 			return (DDI_ME_RNUMBER_RANGE);
870 		}
871 
872 		/*
873 		 * Convert the given ddi_map_req_t from rnumber to regspec...
874 		 */
875 
876 		mp->map_type = DDI_MT_REGSPEC;
877 		mp->map_obj.rp = rp;
878 	}
879 
880 	/*
881 	 * Adjust offset and length correspnding to called values...
882 	 * XXX: A non-zero length means override the one in the regspec
883 	 * XXX: (regardless of what's in the parent's range?)
884 	 */
885 
886 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
887 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
888 
889 #ifdef	DDI_MAP_DEBUG
890 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
891 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
892 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
893 	    len, mp->map_handlep);
894 #endif	/* DDI_MAP_DEBUG */
895 
896 	/*
897 	 * I/O or memory mapping:
898 	 *
899 	 *	<bustype=0, addr=x, len=x>: memory
900 	 *	<bustype=1, addr=x, len=x>: i/o
901 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
902 	 */
903 
904 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
905 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
906 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
907 		    ddi_get_name(rdip), rp->regspec_bustype,
908 		    rp->regspec_addr, rp->regspec_size);
909 		return (DDI_ME_INVAL);
910 	}
911 
912 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
913 		/*
914 		 * compatibility i/o mapping
915 		 */
916 		rp->regspec_bustype += (uint_t)offset;
917 	} else {
918 		/*
919 		 * Normal memory or i/o mapping
920 		 */
921 		rp->regspec_addr += (uint_t)offset;
922 	}
923 
924 	if (len != 0)
925 		rp->regspec_size = (uint_t)len;
926 
927 #ifdef	DDI_MAP_DEBUG
928 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
929 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
930 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
931 	    offset, len, mp->map_handlep);
932 #endif	/* DDI_MAP_DEBUG */
933 
934 	/*
935 	 * Apply any parent ranges at this level, if applicable.
936 	 * (This is where nexus specific regspec translation takes place.
937 	 * Use of this function is implicit agreement that translation is
938 	 * provided via ddi_apply_range.)
939 	 */
940 
941 #ifdef	DDI_MAP_DEBUG
942 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
943 	    ddi_get_name(dip), ddi_get_name(rdip));
944 #endif	/* DDI_MAP_DEBUG */
945 
946 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
947 		return (error);
948 
949 	switch (mp->map_op)  {
950 	case DDI_MO_MAP_LOCKED:
951 
952 		/*
953 		 * Set up the locked down kernel mapping to the regspec...
954 		 */
955 
956 		return (rootnex_map_regspec(mp, vaddrp));
957 
958 	case DDI_MO_UNMAP:
959 
960 		/*
961 		 * Release mapping...
962 		 */
963 
964 		return (rootnex_unmap_regspec(mp, vaddrp));
965 
966 	case DDI_MO_MAP_HANDLE:
967 
968 		return (rootnex_map_handle(mp));
969 
970 	default:
971 		return (DDI_ME_UNIMPLEMENTED);
972 	}
973 }
974 
975 
976 /*
977  * rootnex_map_fault()
978  *
979  *	fault in mappings for requestors
980  */
981 /*ARGSUSED*/
982 static int
983 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
984     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
985     uint_t lock)
986 {
987 
988 #ifdef	DDI_MAP_DEBUG
989 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
990 	ddi_map_debug(" Seg <%s>\n",
991 	    seg->s_ops == &segdev_ops ? "segdev" :
992 	    seg == &kvseg ? "segkmem" : "NONE!");
993 #endif	/* DDI_MAP_DEBUG */
994 
995 	/*
996 	 * This is all terribly broken, but it is a start
997 	 *
998 	 * XXX	Note that this test means that segdev_ops
999 	 *	must be exported from seg_dev.c.
1000 	 * XXX	What about devices with their own segment drivers?
1001 	 */
1002 	if (seg->s_ops == &segdev_ops) {
1003 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
1004 
1005 		if (hat == NULL) {
1006 			/*
1007 			 * This is one plausible interpretation of
1008 			 * a null hat i.e. use the first hat on the
1009 			 * address space hat list which by convention is
1010 			 * the hat of the system MMU.  At alternative
1011 			 * would be to panic .. this might well be better ..
1012 			 */
1013 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
1014 			hat = seg->s_as->a_hat;
1015 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
1016 		}
1017 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
1018 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
1019 	} else if (seg == &kvseg && dp == NULL) {
1020 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
1021 		    HAT_LOAD_LOCK);
1022 	} else
1023 		return (DDI_FAILURE);
1024 	return (DDI_SUCCESS);
1025 }
1026 
1027 
1028 /*
1029  * rootnex_map_regspec()
1030  *     we don't support mapping of I/O cards above 4Gb
1031  */
1032 static int
1033 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1034 {
1035 	rootnex_addr_t rbase;
1036 	void *cvaddr;
1037 	uint_t npages, pgoffset;
1038 	struct regspec *rp;
1039 	ddi_acc_hdl_t *hp;
1040 	ddi_acc_impl_t *ap;
1041 	uint_t	hat_acc_flags;
1042 	paddr_t pbase;
1043 
1044 	rp = mp->map_obj.rp;
1045 	hp = mp->map_handlep;
1046 
1047 #ifdef	DDI_MAP_DEBUG
1048 	ddi_map_debug(
1049 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1050 	    rp->regspec_bustype, rp->regspec_addr,
1051 	    rp->regspec_size, mp->map_handlep);
1052 #endif	/* DDI_MAP_DEBUG */
1053 
1054 	/*
1055 	 * I/O or memory mapping
1056 	 *
1057 	 *	<bustype=0, addr=x, len=x>: memory
1058 	 *	<bustype=1, addr=x, len=x>: i/o
1059 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1060 	 */
1061 
1062 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1063 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1064 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1065 		    rp->regspec_addr, rp->regspec_size);
1066 		return (DDI_FAILURE);
1067 	}
1068 
1069 	if (rp->regspec_bustype != 0) {
1070 		/*
1071 		 * I/O space - needs a handle.
1072 		 */
1073 		if (hp == NULL) {
1074 			return (DDI_FAILURE);
1075 		}
1076 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1077 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1078 		impl_acc_hdl_init(hp);
1079 
1080 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1081 #ifdef  DDI_MAP_DEBUG
1082 			ddi_map_debug("rootnex_map_regspec: mmap() "
1083 			    "to I/O space is not supported.\n");
1084 #endif  /* DDI_MAP_DEBUG */
1085 			return (DDI_ME_INVAL);
1086 		} else {
1087 			/*
1088 			 * 1275-compliant vs. compatibility i/o mapping
1089 			 */
1090 			*vaddrp =
1091 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1092 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1093 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1094 #ifdef __xpv
1095 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1096 				hp->ah_pfn = xen_assign_pfn(
1097 				    mmu_btop((ulong_t)rp->regspec_addr &
1098 				    MMU_PAGEMASK));
1099 			} else {
1100 				hp->ah_pfn = mmu_btop(
1101 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1102 			}
1103 #else
1104 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1105 			    MMU_PAGEMASK);
1106 #endif
1107 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1108 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1109 		}
1110 
1111 #ifdef	DDI_MAP_DEBUG
1112 		ddi_map_debug(
1113 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1114 		    rp->regspec_size, *vaddrp);
1115 #endif	/* DDI_MAP_DEBUG */
1116 		return (DDI_SUCCESS);
1117 	}
1118 
1119 	/*
1120 	 * Memory space
1121 	 */
1122 
1123 	if (hp != NULL) {
1124 		/*
1125 		 * hat layer ignores
1126 		 * hp->ah_acc.devacc_attr_endian_flags.
1127 		 */
1128 		switch (hp->ah_acc.devacc_attr_dataorder) {
1129 		case DDI_STRICTORDER_ACC:
1130 			hat_acc_flags = HAT_STRICTORDER;
1131 			break;
1132 		case DDI_UNORDERED_OK_ACC:
1133 			hat_acc_flags = HAT_UNORDERED_OK;
1134 			break;
1135 		case DDI_MERGING_OK_ACC:
1136 			hat_acc_flags = HAT_MERGING_OK;
1137 			break;
1138 		case DDI_LOADCACHING_OK_ACC:
1139 			hat_acc_flags = HAT_LOADCACHING_OK;
1140 			break;
1141 		case DDI_STORECACHING_OK_ACC:
1142 			hat_acc_flags = HAT_STORECACHING_OK;
1143 			break;
1144 		}
1145 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1146 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1147 		impl_acc_hdl_init(hp);
1148 		hp->ah_hat_flags = hat_acc_flags;
1149 	} else {
1150 		hat_acc_flags = HAT_STRICTORDER;
1151 	}
1152 
1153 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1154 #ifdef __xpv
1155 	/*
1156 	 * If we're dom0, we're using a real device so we need to translate
1157 	 * the MA to a PA.
1158 	 */
1159 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1160 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1161 	} else {
1162 		pbase = rbase;
1163 	}
1164 #else
1165 	pbase = rbase;
1166 #endif
1167 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1168 
1169 	if (rp->regspec_size == 0) {
1170 #ifdef  DDI_MAP_DEBUG
1171 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1172 #endif  /* DDI_MAP_DEBUG */
1173 		return (DDI_ME_INVAL);
1174 	}
1175 
1176 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1177 		/* extra cast to make gcc happy */
1178 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1179 	} else {
1180 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1181 
1182 #ifdef	DDI_MAP_DEBUG
1183 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1184 		    "physical %llx", npages, pbase);
1185 #endif	/* DDI_MAP_DEBUG */
1186 
1187 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1188 		if (cvaddr == NULL)
1189 			return (DDI_ME_NORESOURCES);
1190 
1191 		/*
1192 		 * Now map in the pages we've allocated...
1193 		 */
1194 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1195 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1196 		    HAT_LOAD_LOCK);
1197 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1198 
1199 		/* save away pfn and npages for FMA */
1200 		hp = mp->map_handlep;
1201 		if (hp) {
1202 			hp->ah_pfn = mmu_btop(pbase);
1203 			hp->ah_pnum = npages;
1204 		}
1205 	}
1206 
1207 #ifdef	DDI_MAP_DEBUG
1208 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1209 #endif	/* DDI_MAP_DEBUG */
1210 	return (DDI_SUCCESS);
1211 }
1212 
1213 
1214 /*
1215  * rootnex_unmap_regspec()
1216  *
1217  */
1218 static int
1219 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1220 {
1221 	caddr_t addr = (caddr_t)*vaddrp;
1222 	uint_t npages, pgoffset;
1223 	struct regspec *rp;
1224 
1225 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1226 		return (0);
1227 
1228 	rp = mp->map_obj.rp;
1229 
1230 	if (rp->regspec_size == 0) {
1231 #ifdef  DDI_MAP_DEBUG
1232 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1233 #endif  /* DDI_MAP_DEBUG */
1234 		return (DDI_ME_INVAL);
1235 	}
1236 
1237 	/*
1238 	 * I/O or memory mapping:
1239 	 *
1240 	 *	<bustype=0, addr=x, len=x>: memory
1241 	 *	<bustype=1, addr=x, len=x>: i/o
1242 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1243 	 */
1244 	if (rp->regspec_bustype != 0) {
1245 		/*
1246 		 * This is I/O space, which requires no particular
1247 		 * processing on unmap since it isn't mapped in the
1248 		 * first place.
1249 		 */
1250 		return (DDI_SUCCESS);
1251 	}
1252 
1253 	/*
1254 	 * Memory space
1255 	 */
1256 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1257 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1258 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1259 	device_arena_free(addr - pgoffset, ptob(npages));
1260 
1261 	/*
1262 	 * Destroy the pointer - the mapping has logically gone
1263 	 */
1264 	*vaddrp = NULL;
1265 
1266 	return (DDI_SUCCESS);
1267 }
1268 
1269 
1270 /*
1271  * rootnex_map_handle()
1272  *
1273  */
1274 static int
1275 rootnex_map_handle(ddi_map_req_t *mp)
1276 {
1277 	rootnex_addr_t rbase;
1278 	ddi_acc_hdl_t *hp;
1279 	uint_t pgoffset;
1280 	struct regspec *rp;
1281 	paddr_t pbase;
1282 
1283 	rp = mp->map_obj.rp;
1284 
1285 #ifdef	DDI_MAP_DEBUG
1286 	ddi_map_debug(
1287 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1288 	    rp->regspec_bustype, rp->regspec_addr,
1289 	    rp->regspec_size, mp->map_handlep);
1290 #endif	/* DDI_MAP_DEBUG */
1291 
1292 	/*
1293 	 * I/O or memory mapping:
1294 	 *
1295 	 *	<bustype=0, addr=x, len=x>: memory
1296 	 *	<bustype=1, addr=x, len=x>: i/o
1297 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1298 	 */
1299 	if (rp->regspec_bustype != 0) {
1300 		/*
1301 		 * This refers to I/O space, and we don't support "mapping"
1302 		 * I/O space to a user.
1303 		 */
1304 		return (DDI_FAILURE);
1305 	}
1306 
1307 	/*
1308 	 * Set up the hat_flags for the mapping.
1309 	 */
1310 	hp = mp->map_handlep;
1311 
1312 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1313 	case DDI_NEVERSWAP_ACC:
1314 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1315 		break;
1316 	case DDI_STRUCTURE_LE_ACC:
1317 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1318 		break;
1319 	case DDI_STRUCTURE_BE_ACC:
1320 		return (DDI_FAILURE);
1321 	default:
1322 		return (DDI_REGS_ACC_CONFLICT);
1323 	}
1324 
1325 	switch (hp->ah_acc.devacc_attr_dataorder) {
1326 	case DDI_STRICTORDER_ACC:
1327 		break;
1328 	case DDI_UNORDERED_OK_ACC:
1329 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1330 		break;
1331 	case DDI_MERGING_OK_ACC:
1332 		hp->ah_hat_flags |= HAT_MERGING_OK;
1333 		break;
1334 	case DDI_LOADCACHING_OK_ACC:
1335 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1336 		break;
1337 	case DDI_STORECACHING_OK_ACC:
1338 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1339 		break;
1340 	default:
1341 		return (DDI_FAILURE);
1342 	}
1343 
1344 	rbase = (rootnex_addr_t)rp->regspec_addr &
1345 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1346 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1347 
1348 	if (rp->regspec_size == 0)
1349 		return (DDI_ME_INVAL);
1350 
1351 #ifdef __xpv
1352 	/*
1353 	 * If we're dom0, we're using a real device so we need to translate
1354 	 * the MA to a PA.
1355 	 */
1356 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1357 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1358 		    (rbase & MMU_PAGEOFFSET);
1359 	} else {
1360 		pbase = rbase;
1361 	}
1362 #else
1363 	pbase = rbase;
1364 #endif
1365 
1366 	hp->ah_pfn = mmu_btop(pbase);
1367 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1368 
1369 	return (DDI_SUCCESS);
1370 }
1371 
1372 
1373 
1374 /*
1375  * ************************
1376  *  interrupt related code
1377  * ************************
1378  */
1379 
1380 /*
1381  * rootnex_intr_ops()
1382  *	bus_intr_op() function for interrupt support
1383  */
1384 /* ARGSUSED */
1385 static int
1386 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1387     ddi_intr_handle_impl_t *hdlp, void *result)
1388 {
1389 	struct intrspec			*ispec;
1390 
1391 	DDI_INTR_NEXDBG((CE_CONT,
1392 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1393 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1394 
1395 	/* Process the interrupt operation */
1396 	switch (intr_op) {
1397 	case DDI_INTROP_GETCAP:
1398 		/* First check with pcplusmp */
1399 		if (psm_intr_ops == NULL)
1400 			return (DDI_FAILURE);
1401 
1402 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1403 			*(int *)result = 0;
1404 			return (DDI_FAILURE);
1405 		}
1406 		break;
1407 	case DDI_INTROP_SETCAP:
1408 		if (psm_intr_ops == NULL)
1409 			return (DDI_FAILURE);
1410 
1411 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1412 			return (DDI_FAILURE);
1413 		break;
1414 	case DDI_INTROP_ALLOC:
1415 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1416 		return (rootnex_alloc_intr_fixed(rdip, hdlp, result));
1417 	case DDI_INTROP_FREE:
1418 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1419 		return (rootnex_free_intr_fixed(rdip, hdlp));
1420 	case DDI_INTROP_GETPRI:
1421 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1422 			return (DDI_FAILURE);
1423 		*(int *)result = ispec->intrspec_pri;
1424 		break;
1425 	case DDI_INTROP_SETPRI:
1426 		/* Validate the interrupt priority passed to us */
1427 		if (*(int *)result > LOCK_LEVEL)
1428 			return (DDI_FAILURE);
1429 
1430 		/* Ensure that PSM is all initialized and ispec is ok */
1431 		if ((psm_intr_ops == NULL) ||
1432 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1433 			return (DDI_FAILURE);
1434 
1435 		/* Change the priority */
1436 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1437 		    PSM_FAILURE)
1438 			return (DDI_FAILURE);
1439 
1440 		/* update the ispec with the new priority */
1441 		ispec->intrspec_pri =  *(int *)result;
1442 		break;
1443 	case DDI_INTROP_ADDISR:
1444 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1445 			return (DDI_FAILURE);
1446 		ispec->intrspec_func = hdlp->ih_cb_func;
1447 		break;
1448 	case DDI_INTROP_REMISR:
1449 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1450 			return (DDI_FAILURE);
1451 		ispec->intrspec_func = (uint_t (*)()) 0;
1452 		break;
1453 	case DDI_INTROP_ENABLE:
1454 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1455 			return (DDI_FAILURE);
1456 
1457 		/* Call psmi to translate irq with the dip */
1458 		if (psm_intr_ops == NULL)
1459 			return (DDI_FAILURE);
1460 
1461 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1462 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1463 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1464 			return (DDI_FAILURE);
1465 
1466 		/* Add the interrupt handler */
1467 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1468 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1469 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1470 			return (DDI_FAILURE);
1471 		break;
1472 	case DDI_INTROP_DISABLE:
1473 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1474 			return (DDI_FAILURE);
1475 
1476 		/* Call psm_ops() to translate irq with the dip */
1477 		if (psm_intr_ops == NULL)
1478 			return (DDI_FAILURE);
1479 
1480 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1481 		(void) (*psm_intr_ops)(rdip, hdlp,
1482 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1483 
1484 		/* Remove the interrupt handler */
1485 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1486 		    hdlp->ih_cb_func, hdlp->ih_vector);
1487 		break;
1488 	case DDI_INTROP_SETMASK:
1489 		if (psm_intr_ops == NULL)
1490 			return (DDI_FAILURE);
1491 
1492 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1493 			return (DDI_FAILURE);
1494 		break;
1495 	case DDI_INTROP_CLRMASK:
1496 		if (psm_intr_ops == NULL)
1497 			return (DDI_FAILURE);
1498 
1499 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1500 			return (DDI_FAILURE);
1501 		break;
1502 	case DDI_INTROP_GETPENDING:
1503 		if (psm_intr_ops == NULL)
1504 			return (DDI_FAILURE);
1505 
1506 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1507 		    result)) {
1508 			*(int *)result = 0;
1509 			return (DDI_FAILURE);
1510 		}
1511 		break;
1512 	case DDI_INTROP_NAVAIL:
1513 	case DDI_INTROP_NINTRS:
1514 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1515 		if (*(int *)result == 0) {
1516 			/*
1517 			 * Special case for 'pcic' driver' only. This driver
1518 			 * driver is a child of 'isa' and 'rootnex' drivers.
1519 			 *
1520 			 * See detailed comments on this in the function
1521 			 * rootnex_get_ispec().
1522 			 *
1523 			 * Children of 'pcic' send 'NINITR' request all the
1524 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1525 			 * field may not initialized. So, we fake it here
1526 			 * to return 1 (a la what PCMCIA nexus does).
1527 			 */
1528 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1529 				*(int *)result = 1;
1530 			else
1531 				return (DDI_FAILURE);
1532 		}
1533 		break;
1534 	case DDI_INTROP_SUPPORTED_TYPES:
1535 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1536 		break;
1537 	default:
1538 		return (DDI_FAILURE);
1539 	}
1540 
1541 	return (DDI_SUCCESS);
1542 }
1543 
1544 
1545 /*
1546  * rootnex_get_ispec()
1547  *	convert an interrupt number to an interrupt specification.
1548  *	The interrupt number determines which interrupt spec will be
1549  *	returned if more than one exists.
1550  *
1551  *	Look into the parent private data area of the 'rdip' to find out
1552  *	the interrupt specification.  First check to make sure there is
1553  *	one that matchs "inumber" and then return a pointer to it.
1554  *
1555  *	Return NULL if one could not be found.
1556  *
1557  *	NOTE: This is needed for rootnex_intr_ops()
1558  */
1559 static struct intrspec *
1560 rootnex_get_ispec(dev_info_t *rdip, int inum)
1561 {
1562 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1563 
1564 	/*
1565 	 * Special case handling for drivers that provide their own
1566 	 * intrspec structures instead of relying on the DDI framework.
1567 	 *
1568 	 * A broken hardware driver in ON could potentially provide its
1569 	 * own intrspec structure, instead of relying on the hardware.
1570 	 * If these drivers are children of 'rootnex' then we need to
1571 	 * continue to provide backward compatibility to them here.
1572 	 *
1573 	 * Following check is a special case for 'pcic' driver which
1574 	 * was found to have broken hardwre andby provides its own intrspec.
1575 	 *
1576 	 * Verbatim comments from this driver are shown here:
1577 	 * "Don't use the ddi_add_intr since we don't have a
1578 	 * default intrspec in all cases."
1579 	 *
1580 	 * Since an 'ispec' may not be always created for it,
1581 	 * check for that and create one if so.
1582 	 *
1583 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1584 	 */
1585 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1586 		pdp->par_nintr = 1;
1587 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1588 		    pdp->par_nintr, KM_SLEEP);
1589 	}
1590 
1591 	/* Validate the interrupt number */
1592 	if (inum >= pdp->par_nintr)
1593 		return (NULL);
1594 
1595 	/* Get the interrupt structure pointer and return that */
1596 	return ((struct intrspec *)&pdp->par_intr[inum]);
1597 }
1598 
1599 /*
1600  * Allocate interrupt vector for FIXED (legacy) type.
1601  */
1602 static int
1603 rootnex_alloc_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp,
1604     void *result)
1605 {
1606 	struct intrspec		*ispec;
1607 	ddi_intr_handle_impl_t	info_hdl;
1608 	int			ret;
1609 	int			free_phdl = 0;
1610 	apic_get_type_t		type_info;
1611 
1612 	if (psm_intr_ops == NULL)
1613 		return (DDI_FAILURE);
1614 
1615 	if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1616 		return (DDI_FAILURE);
1617 
1618 	/*
1619 	 * If the PSM module is "APIX" then pass the request for it
1620 	 * to allocate the vector now.
1621 	 */
1622 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1623 	info_hdl.ih_private = &type_info;
1624 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1625 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1626 		if (hdlp->ih_private == NULL) { /* allocate phdl structure */
1627 			free_phdl = 1;
1628 			i_ddi_alloc_intr_phdl(hdlp);
1629 		}
1630 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1631 		ret = (*psm_intr_ops)(rdip, hdlp,
1632 		    PSM_INTR_OP_ALLOC_VECTORS, result);
1633 		if (free_phdl) { /* free up the phdl structure */
1634 			free_phdl = 0;
1635 			i_ddi_free_intr_phdl(hdlp);
1636 			hdlp->ih_private = NULL;
1637 		}
1638 	} else {
1639 		/*
1640 		 * No APIX module; fall back to the old scheme where the
1641 		 * interrupt vector is allocated during ddi_enable_intr() call.
1642 		 */
1643 		hdlp->ih_pri = ispec->intrspec_pri;
1644 		*(int *)result = hdlp->ih_scratch1;
1645 		ret = DDI_SUCCESS;
1646 	}
1647 
1648 	return (ret);
1649 }
1650 
1651 /*
1652  * Free up interrupt vector for FIXED (legacy) type.
1653  */
1654 static int
1655 rootnex_free_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp)
1656 {
1657 	struct intrspec			*ispec;
1658 	struct ddi_parent_private_data	*pdp;
1659 	ddi_intr_handle_impl_t		info_hdl;
1660 	int				ret;
1661 	apic_get_type_t			type_info;
1662 
1663 	if (psm_intr_ops == NULL)
1664 		return (DDI_FAILURE);
1665 
1666 	/*
1667 	 * If the PSM module is "APIX" then pass the request for it
1668 	 * to free up the vector now.
1669 	 */
1670 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1671 	info_hdl.ih_private = &type_info;
1672 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1673 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1674 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1675 			return (DDI_FAILURE);
1676 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1677 		ret = (*psm_intr_ops)(rdip, hdlp,
1678 		    PSM_INTR_OP_FREE_VECTORS, NULL);
1679 	} else {
1680 		/*
1681 		 * No APIX module; fall back to the old scheme where
1682 		 * the interrupt vector was already freed during
1683 		 * ddi_disable_intr() call.
1684 		 */
1685 		ret = DDI_SUCCESS;
1686 	}
1687 
1688 	pdp = ddi_get_parent_data(rdip);
1689 
1690 	/*
1691 	 * Special case for 'pcic' driver' only.
1692 	 * If an intrspec was created for it, clean it up here
1693 	 * See detailed comments on this in the function
1694 	 * rootnex_get_ispec().
1695 	 */
1696 	if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1697 		kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1698 		    pdp->par_nintr);
1699 		/*
1700 		 * Set it to zero; so that
1701 		 * DDI framework doesn't free it again
1702 		 */
1703 		pdp->par_intr = NULL;
1704 		pdp->par_nintr = 0;
1705 	}
1706 
1707 	return (ret);
1708 }
1709 
1710 
1711 /*
1712  * ******************
1713  *  dma related code
1714  * ******************
1715  */
1716 
1717 /*ARGSUSED*/
1718 static int
1719 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1720     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1721     ddi_dma_handle_t *handlep)
1722 {
1723 	uint64_t maxsegmentsize_ll;
1724 	uint_t maxsegmentsize;
1725 	ddi_dma_impl_t *hp;
1726 	rootnex_dma_t *dma;
1727 	uint64_t count_max;
1728 	uint64_t seg;
1729 	int kmflag;
1730 	int e;
1731 
1732 
1733 	/* convert our sleep flags */
1734 	if (waitfp == DDI_DMA_SLEEP) {
1735 		kmflag = KM_SLEEP;
1736 	} else {
1737 		kmflag = KM_NOSLEEP;
1738 	}
1739 
1740 	/*
1741 	 * We try to do only one memory allocation here. We'll do a little
1742 	 * pointer manipulation later. If the bind ends up taking more than
1743 	 * our prealloc's space, we'll have to allocate more memory in the
1744 	 * bind operation. Not great, but much better than before and the
1745 	 * best we can do with the current bind interfaces.
1746 	 */
1747 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1748 	if (hp == NULL)
1749 		return (DDI_DMA_NORESOURCES);
1750 
1751 	/* Do our pointer manipulation now, align the structures */
1752 	hp->dmai_private = (void *)(((uintptr_t)hp +
1753 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1754 	dma = (rootnex_dma_t *)hp->dmai_private;
1755 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1756 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1757 
1758 	/* setup the handle */
1759 	rootnex_clean_dmahdl(hp);
1760 	hp->dmai_error.err_fep = NULL;
1761 	hp->dmai_error.err_cf = NULL;
1762 	dma->dp_dip = rdip;
1763 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1764 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1765 
1766 	/*
1767 	 * The BOUNCE_ON_SEG workaround is not needed when an IOMMU
1768 	 * is being used. Set the upper limit to the seg value.
1769 	 * There will be enough DVMA space to always get addresses
1770 	 * that will match the constraints.
1771 	 */
1772 	if (IOMMU_USED(rdip) &&
1773 	    (attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG)) {
1774 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_seg;
1775 		dma->dp_sglinfo.si_flags &= ~_DDI_DMA_BOUNCE_ON_SEG;
1776 	} else
1777 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1778 
1779 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1780 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1781 	hp->dmai_rdip = rdip;
1782 	hp->dmai_attr = *attr;
1783 
1784 	if (attr->dma_attr_seg >= dma->dp_sglinfo.si_max_addr)
1785 		dma->dp_sglinfo.si_cancross = B_FALSE;
1786 	else
1787 		dma->dp_sglinfo.si_cancross = B_TRUE;
1788 
1789 	/* we don't need to worry about the SPL since we do a tryenter */
1790 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1791 
1792 	/*
1793 	 * Figure out our maximum segment size. If the segment size is greater
1794 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1795 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1796 	 * dma_attr_count_max are size-1 type values.
1797 	 *
1798 	 * Maximum segment size is the largest physically contiguous chunk of
1799 	 * memory that we can return from a bind (i.e. the maximum size of a
1800 	 * single cookie).
1801 	 */
1802 
1803 	/* handle the rollover cases */
1804 	seg = attr->dma_attr_seg + 1;
1805 	if (seg < attr->dma_attr_seg) {
1806 		seg = attr->dma_attr_seg;
1807 	}
1808 	count_max = attr->dma_attr_count_max + 1;
1809 	if (count_max < attr->dma_attr_count_max) {
1810 		count_max = attr->dma_attr_count_max;
1811 	}
1812 
1813 	/*
1814 	 * granularity may or may not be a power of two. If it isn't, we can't
1815 	 * use a simple mask.
1816 	 */
1817 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1818 		dma->dp_granularity_power_2 = B_FALSE;
1819 	} else {
1820 		dma->dp_granularity_power_2 = B_TRUE;
1821 	}
1822 
1823 	/*
1824 	 * maxxfer should be a whole multiple of granularity. If we're going to
1825 	 * break up a window because we're greater than maxxfer, we might as
1826 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1827 	 * worry about triming the window later on for this case.
1828 	 */
1829 	if (attr->dma_attr_granular > 1) {
1830 		if (dma->dp_granularity_power_2) {
1831 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1832 			    (attr->dma_attr_maxxfer &
1833 			    (attr->dma_attr_granular - 1));
1834 		} else {
1835 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1836 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1837 		}
1838 	} else {
1839 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1840 	}
1841 
1842 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1843 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1844 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1845 		maxsegmentsize = 0xFFFFFFFF;
1846 	} else {
1847 		maxsegmentsize = maxsegmentsize_ll;
1848 	}
1849 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1850 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1851 
1852 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1853 	if (rootnex_alloc_check_parms) {
1854 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1855 		if (e != DDI_SUCCESS) {
1856 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1857 			(void) rootnex_dma_freehdl(dip, rdip,
1858 			    (ddi_dma_handle_t)hp);
1859 			return (e);
1860 		}
1861 	}
1862 
1863 	*handlep = (ddi_dma_handle_t)hp;
1864 
1865 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1866 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1867 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1868 
1869 	return (DDI_SUCCESS);
1870 }
1871 
1872 
1873 /*
1874  * rootnex_dma_allochdl()
1875  *    called from ddi_dma_alloc_handle().
1876  */
1877 static int
1878 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1879     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1880 {
1881 	int retval = DDI_SUCCESS;
1882 #if defined(__amd64) && !defined(__xpv)
1883 
1884 	if (IOMMU_UNITIALIZED(rdip)) {
1885 		retval = iommulib_nex_open(dip, rdip);
1886 
1887 		if (retval != DDI_SUCCESS && retval != DDI_ENOTSUP)
1888 			return (retval);
1889 	}
1890 
1891 	if (IOMMU_UNUSED(rdip)) {
1892 		retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1893 		    handlep);
1894 	} else {
1895 		retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1896 		    waitfp, arg, handlep);
1897 	}
1898 #else
1899 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1900 	    handlep);
1901 #endif
1902 	switch (retval) {
1903 	case DDI_DMA_NORESOURCES:
1904 		if (waitfp != DDI_DMA_DONTWAIT) {
1905 			ddi_set_callback(waitfp, arg,
1906 			    &rootnex_state->r_dvma_call_list_id);
1907 		}
1908 		break;
1909 	case DDI_SUCCESS:
1910 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1911 		break;
1912 	default:
1913 		break;
1914 	}
1915 	return (retval);
1916 }
1917 
1918 /*ARGSUSED*/
1919 static int
1920 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1921     ddi_dma_handle_t handle)
1922 {
1923 	ddi_dma_impl_t *hp;
1924 	rootnex_dma_t *dma;
1925 
1926 
1927 	hp = (ddi_dma_impl_t *)handle;
1928 	dma = (rootnex_dma_t *)hp->dmai_private;
1929 
1930 	/* unbind should have been called first */
1931 	ASSERT(!dma->dp_inuse);
1932 
1933 	mutex_destroy(&dma->dp_mutex);
1934 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1935 
1936 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1937 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1938 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1939 
1940 	return (DDI_SUCCESS);
1941 }
1942 
1943 /*
1944  * rootnex_dma_freehdl()
1945  *    called from ddi_dma_free_handle().
1946  */
1947 static int
1948 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1949 {
1950 	int ret;
1951 
1952 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1953 #if defined(__amd64) && !defined(__xpv)
1954 	if (IOMMU_USED(rdip))
1955 		ret = iommulib_nexdma_freehdl(dip, rdip, handle);
1956 	else
1957 #endif
1958 	ret = rootnex_coredma_freehdl(dip, rdip, handle);
1959 
1960 	if (rootnex_state->r_dvma_call_list_id)
1961 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1962 
1963 	return (ret);
1964 }
1965 
1966 /*ARGSUSED*/
1967 static int
1968 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1969     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1970     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1971 {
1972 	rootnex_sglinfo_t *sinfo;
1973 	ddi_dma_obj_t *dmao;
1974 #if defined(__amd64) && !defined(__xpv)
1975 	struct dvmaseg *dvs;
1976 	ddi_dma_cookie_t *cookie;
1977 #endif
1978 	ddi_dma_attr_t *attr;
1979 	ddi_dma_impl_t *hp;
1980 	rootnex_dma_t *dma;
1981 	int kmflag;
1982 	int e;
1983 	uint_t ncookies;
1984 
1985 	hp = (ddi_dma_impl_t *)handle;
1986 	dma = (rootnex_dma_t *)hp->dmai_private;
1987 	dmao = &dma->dp_dma;
1988 	sinfo = &dma->dp_sglinfo;
1989 	attr = &hp->dmai_attr;
1990 
1991 	/* convert the sleep flags */
1992 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1993 		dma->dp_sleep_flags = kmflag = KM_SLEEP;
1994 	} else {
1995 		dma->dp_sleep_flags = kmflag = KM_NOSLEEP;
1996 	}
1997 
1998 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
1999 
2000 	/*
2001 	 * This is useful for debugging a driver. Not as useful in a production
2002 	 * system. The only time this will fail is if you have a driver bug.
2003 	 */
2004 	if (rootnex_bind_check_inuse) {
2005 		/*
2006 		 * No one else should ever have this lock unless someone else
2007 		 * is trying to use this handle. So contention on the lock
2008 		 * is the same as inuse being set.
2009 		 */
2010 		e = mutex_tryenter(&dma->dp_mutex);
2011 		if (e == 0) {
2012 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2013 			return (DDI_DMA_INUSE);
2014 		}
2015 		if (dma->dp_inuse) {
2016 			mutex_exit(&dma->dp_mutex);
2017 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2018 			return (DDI_DMA_INUSE);
2019 		}
2020 		dma->dp_inuse = B_TRUE;
2021 		mutex_exit(&dma->dp_mutex);
2022 	}
2023 
2024 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
2025 	if (rootnex_bind_check_parms) {
2026 		e = rootnex_valid_bind_parms(dmareq, attr);
2027 		if (e != DDI_SUCCESS) {
2028 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2029 			rootnex_clean_dmahdl(hp);
2030 			return (e);
2031 		}
2032 	}
2033 
2034 	/* save away the original bind info */
2035 	dma->dp_dma = dmareq->dmar_object;
2036 
2037 #if defined(__amd64) && !defined(__xpv)
2038 	if (IOMMU_USED(rdip)) {
2039 		dmao = &dma->dp_dvma;
2040 		e = iommulib_nexdma_mapobject(dip, rdip, handle, dmareq, dmao);
2041 		switch (e) {
2042 		case DDI_SUCCESS:
2043 			if (sinfo->si_cancross ||
2044 			    dmao->dmao_obj.dvma_obj.dv_nseg != 1 ||
2045 			    dmao->dmao_size > sinfo->si_max_cookie_size) {
2046 				dma->dp_dvma_used = B_TRUE;
2047 				break;
2048 			}
2049 			sinfo->si_sgl_size = 1;
2050 			hp->dmai_rflags |= DMP_NOSYNC;
2051 
2052 			dma->dp_dvma_used = B_TRUE;
2053 			dma->dp_need_to_free_cookie = B_FALSE;
2054 
2055 			dvs = &dmao->dmao_obj.dvma_obj.dv_seg[0];
2056 			cookie = hp->dmai_cookie = dma->dp_cookies =
2057 			    (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2058 			cookie->dmac_laddress = dvs->dvs_start +
2059 			    dmao->dmao_obj.dvma_obj.dv_off;
2060 			cookie->dmac_size = dvs->dvs_len;
2061 			cookie->dmac_type = 0;
2062 
2063 			ROOTNEX_DPROBE1(rootnex__bind__dvmafast, dev_info_t *,
2064 			    rdip);
2065 			goto fast;
2066 		case DDI_ENOTSUP:
2067 			break;
2068 		default:
2069 			rootnex_clean_dmahdl(hp);
2070 			return (e);
2071 		}
2072 	}
2073 #endif
2074 
2075 	/*
2076 	 * Figure out a rough estimate of what maximum number of pages
2077 	 * this buffer could use (a high estimate of course).
2078 	 */
2079 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
2080 
2081 	if (dma->dp_dvma_used) {
2082 		/*
2083 		 * The number of physical pages is the worst case.
2084 		 *
2085 		 * For DVMA, the worst case is the length divided
2086 		 * by the maximum cookie length, plus 1. Add to that
2087 		 * the number of segment boundaries potentially crossed, and
2088 		 * the additional number of DVMA segments that was returned.
2089 		 *
2090 		 * In the normal case, for modern devices, si_cancross will
2091 		 * be false, and dv_nseg will be 1, and the fast path will
2092 		 * have been taken above.
2093 		 */
2094 		ncookies = (dma->dp_dma.dmao_size / sinfo->si_max_cookie_size)
2095 		    + 1;
2096 		if (sinfo->si_cancross)
2097 			ncookies +=
2098 			    (dma->dp_dma.dmao_size / attr->dma_attr_seg) + 1;
2099 		ncookies += (dmao->dmao_obj.dvma_obj.dv_nseg - 1);
2100 
2101 		sinfo->si_max_pages = MIN(sinfo->si_max_pages, ncookies);
2102 	}
2103 
2104 	/*
2105 	 * We'll use the pre-allocated cookies for any bind that will *always*
2106 	 * fit (more important to be consistent, we don't want to create
2107 	 * additional degenerate cases).
2108 	 */
2109 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
2110 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2111 		dma->dp_need_to_free_cookie = B_FALSE;
2112 		ROOTNEX_DPROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
2113 		    uint_t, sinfo->si_max_pages);
2114 
2115 	/*
2116 	 * For anything larger than that, we'll go ahead and allocate the
2117 	 * maximum number of pages we expect to see. Hopefuly, we won't be
2118 	 * seeing this path in the fast path for high performance devices very
2119 	 * frequently.
2120 	 *
2121 	 * a ddi bind interface that allowed the driver to provide storage to
2122 	 * the bind interface would speed this case up.
2123 	 */
2124 	} else {
2125 		/*
2126 		 * Save away how much memory we allocated. If we're doing a
2127 		 * nosleep, the alloc could fail...
2128 		 */
2129 		dma->dp_cookie_size = sinfo->si_max_pages *
2130 		    sizeof (ddi_dma_cookie_t);
2131 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
2132 		if (dma->dp_cookies == NULL) {
2133 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2134 			rootnex_clean_dmahdl(hp);
2135 			return (DDI_DMA_NORESOURCES);
2136 		}
2137 		dma->dp_need_to_free_cookie = B_TRUE;
2138 		ROOTNEX_DPROBE2(rootnex__bind__alloc, dev_info_t *, rdip,
2139 		    uint_t, sinfo->si_max_pages);
2140 	}
2141 	hp->dmai_cookie = dma->dp_cookies;
2142 
2143 	/*
2144 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
2145 	 * looking at the constraints in the dma structure. It will then put
2146 	 * some additional state about the sgl in the dma struct (i.e. is
2147 	 * the sgl clean, or do we need to do some munging; how many pages
2148 	 * need to be copied, etc.)
2149 	 */
2150 	if (dma->dp_dvma_used)
2151 		rootnex_dvma_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2152 	else
2153 		rootnex_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2154 
2155 out:
2156 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
2157 	/* if we don't need a copy buffer, we don't need to sync */
2158 	if (sinfo->si_copybuf_req == 0) {
2159 		hp->dmai_rflags |= DMP_NOSYNC;
2160 	}
2161 
2162 	/*
2163 	 * if we don't need the copybuf and we don't need to do a partial,  we
2164 	 * hit the fast path. All the high performance devices should be trying
2165 	 * to hit this path. To hit this path, a device should be able to reach
2166 	 * all of memory, shouldn't try to bind more than it can transfer, and
2167 	 * the buffer shouldn't require more cookies than the driver/device can
2168 	 * handle [sgllen]).
2169 	 *
2170 	 * Note that negative values of dma_attr_sgllen are supposed
2171 	 * to mean unlimited, but we just cast them to mean a
2172 	 * "ridiculous large limit".  This saves some extra checks on
2173 	 * hot paths.
2174 	 */
2175 	if ((sinfo->si_copybuf_req == 0) &&
2176 	    (sinfo->si_sgl_size <= (unsigned)attr->dma_attr_sgllen) &&
2177 	    (dmao->dmao_size < dma->dp_maxxfer)) {
2178 fast:
2179 		/*
2180 		 * If the driver supports FMA, insert the handle in the FMA DMA
2181 		 * handle cache.
2182 		 */
2183 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2184 			hp->dmai_error.err_cf = rootnex_dma_check;
2185 
2186 		/*
2187 		 * copy out the first cookie and ccountp, set the cookie
2188 		 * pointer to the second cookie. The first cookie is passed
2189 		 * back on the stack. Additional cookies are accessed via
2190 		 * ddi_dma_nextcookie()
2191 		 */
2192 		*cookiep = dma->dp_cookies[0];
2193 		*ccountp = sinfo->si_sgl_size;
2194 		hp->dmai_cookie++;
2195 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2196 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2197 		ROOTNEX_DPROBE4(rootnex__bind__fast, dev_info_t *, rdip,
2198 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2199 		    uint_t, dmao->dmao_size, uint_t, *ccountp);
2200 
2201 
2202 		return (DDI_DMA_MAPPED);
2203 	}
2204 
2205 	/*
2206 	 * go to the slow path, we may need to alloc more memory, create
2207 	 * multiple windows, and munge up a sgl to make the device happy.
2208 	 */
2209 
2210 	/*
2211 	 * With the IOMMU mapobject method used, we should never hit
2212 	 * the slow path. If we do, something is seriously wrong.
2213 	 * Clean up and return an error.
2214 	 */
2215 
2216 #if defined(__amd64) && !defined(__xpv)
2217 
2218 	if (dma->dp_dvma_used) {
2219 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2220 		    &dma->dp_dvma);
2221 		e = DDI_DMA_NOMAPPING;
2222 	} else {
2223 #endif
2224 		e = rootnex_bind_slowpath(hp, dmareq, dma, attr, &dma->dp_dma,
2225 		    kmflag);
2226 #if defined(__amd64) && !defined(__xpv)
2227 	}
2228 #endif
2229 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2230 		if (dma->dp_need_to_free_cookie) {
2231 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2232 		}
2233 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2234 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2235 		return (e);
2236 	}
2237 
2238 	/*
2239 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2240 	 * cache.
2241 	 */
2242 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2243 		hp->dmai_error.err_cf = rootnex_dma_check;
2244 
2245 	/* if the first window uses the copy buffer, sync it for the device */
2246 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2247 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2248 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2249 		    DDI_DMA_SYNC_FORDEV);
2250 	}
2251 
2252 	/*
2253 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2254 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2255 	 * If we have a partial map (i.e. multiple windows), the number of
2256 	 * cookies we return is the number of cookies in the first window.
2257 	 */
2258 	if (e == DDI_DMA_MAPPED) {
2259 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2260 		*ccountp = sinfo->si_sgl_size;
2261 		hp->dmai_nwin = 1;
2262 	} else {
2263 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2264 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2265 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2266 	}
2267 	*cookiep = dma->dp_cookies[0];
2268 	hp->dmai_cookie++;
2269 
2270 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2271 	ROOTNEX_DPROBE4(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2272 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2273 	    dmao->dmao_size, uint_t, *ccountp);
2274 	return (e);
2275 }
2276 
2277 /*
2278  * rootnex_dma_bindhdl()
2279  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2280  */
2281 static int
2282 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2283     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2284     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2285 {
2286 	int ret;
2287 #if defined(__amd64) && !defined(__xpv)
2288 	if (IOMMU_USED(rdip))
2289 		ret = iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2290 		    cookiep, ccountp);
2291 	else
2292 #endif
2293 	ret = rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2294 	    cookiep, ccountp);
2295 
2296 	if (ret == DDI_DMA_NORESOURCES && dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
2297 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
2298 		    &rootnex_state->r_dvma_call_list_id);
2299 	}
2300 
2301 	return (ret);
2302 }
2303 
2304 
2305 
2306 /*ARGSUSED*/
2307 static int
2308 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2309     ddi_dma_handle_t handle)
2310 {
2311 	ddi_dma_impl_t *hp;
2312 	rootnex_dma_t *dma;
2313 	int e;
2314 
2315 	hp = (ddi_dma_impl_t *)handle;
2316 	dma = (rootnex_dma_t *)hp->dmai_private;
2317 
2318 	/* make sure the buffer wasn't free'd before calling unbind */
2319 	if (rootnex_unbind_verify_buffer) {
2320 		e = rootnex_verify_buffer(dma);
2321 		if (e != DDI_SUCCESS) {
2322 			ASSERT(0);
2323 			return (DDI_FAILURE);
2324 		}
2325 	}
2326 
2327 	/* sync the current window before unbinding the buffer */
2328 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2329 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2330 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2331 		    DDI_DMA_SYNC_FORCPU);
2332 	}
2333 
2334 	/*
2335 	 * cleanup and copy buffer or window state. if we didn't use the copy
2336 	 * buffer or windows, there won't be much to do :-)
2337 	 */
2338 	rootnex_teardown_copybuf(dma);
2339 	rootnex_teardown_windows(dma);
2340 
2341 #if defined(__amd64) && !defined(__xpv)
2342 	if (IOMMU_USED(rdip))
2343 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2344 		    &dma->dp_dvma);
2345 #endif
2346 
2347 	/*
2348 	 * If we had to allocate space to for the worse case sgl (it didn't
2349 	 * fit into our pre-allocate buffer), free that up now
2350 	 */
2351 	if (dma->dp_need_to_free_cookie) {
2352 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2353 	}
2354 
2355 	/*
2356 	 * clean up the handle so it's ready for the next bind (i.e. if the
2357 	 * handle is reused).
2358 	 */
2359 	rootnex_clean_dmahdl(hp);
2360 	hp->dmai_error.err_cf = NULL;
2361 
2362 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2363 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2364 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2365 
2366 	return (DDI_SUCCESS);
2367 }
2368 
2369 /*
2370  * rootnex_dma_unbindhdl()
2371  *    called from ddi_dma_unbind_handle()
2372  */
2373 /*ARGSUSED*/
2374 static int
2375 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2376     ddi_dma_handle_t handle)
2377 {
2378 	int ret;
2379 
2380 #if defined(__amd64) && !defined(__xpv)
2381 	if (IOMMU_USED(rdip))
2382 		ret = iommulib_nexdma_unbindhdl(dip, rdip, handle);
2383 	else
2384 #endif
2385 	ret = rootnex_coredma_unbindhdl(dip, rdip, handle);
2386 
2387 	if (rootnex_state->r_dvma_call_list_id)
2388 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2389 
2390 	return (ret);
2391 }
2392 
2393 #if defined(__amd64) && !defined(__xpv)
2394 
2395 static int
2396 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2397 {
2398 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2399 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2400 
2401 	if (dma->dp_sleep_flags != KM_SLEEP &&
2402 	    dma->dp_sleep_flags != KM_NOSLEEP)
2403 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2404 	return (dma->dp_sleep_flags);
2405 }
2406 /*ARGSUSED*/
2407 static void
2408 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2409 {
2410 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2411 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2412 	rootnex_window_t *window;
2413 
2414 	if (dma->dp_window) {
2415 		window = &dma->dp_window[dma->dp_current_win];
2416 		hp->dmai_cookie = window->wd_first_cookie;
2417 	} else {
2418 		hp->dmai_cookie = dma->dp_cookies;
2419 	}
2420 	hp->dmai_cookie++;
2421 }
2422 
2423 /*ARGSUSED*/
2424 static int
2425 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2426     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2427 {
2428 	int i;
2429 	int km_flags;
2430 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2431 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2432 	rootnex_window_t *window;
2433 	ddi_dma_cookie_t *cp;
2434 	ddi_dma_cookie_t *cookie;
2435 
2436 	ASSERT(*cookiepp == NULL);
2437 	ASSERT(*ccountp == 0);
2438 
2439 	if (dma->dp_window) {
2440 		window = &dma->dp_window[dma->dp_current_win];
2441 		cp = window->wd_first_cookie;
2442 		*ccountp = window->wd_cookie_cnt;
2443 	} else {
2444 		cp = dma->dp_cookies;
2445 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2446 	}
2447 
2448 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2449 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2450 	if (cookie == NULL) {
2451 		return (DDI_DMA_NORESOURCES);
2452 	}
2453 
2454 	for (i = 0; i < *ccountp; i++) {
2455 		cookie[i].dmac_notused = cp[i].dmac_notused;
2456 		cookie[i].dmac_type = cp[i].dmac_type;
2457 		cookie[i].dmac_address = cp[i].dmac_address;
2458 		cookie[i].dmac_size = cp[i].dmac_size;
2459 	}
2460 
2461 	*cookiepp = cookie;
2462 
2463 	return (DDI_SUCCESS);
2464 }
2465 
2466 /*ARGSUSED*/
2467 static int
2468 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2469     ddi_dma_cookie_t *cookiep, uint_t ccount)
2470 {
2471 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2472 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2473 	rootnex_window_t *window;
2474 	ddi_dma_cookie_t *cur_cookiep;
2475 
2476 	ASSERT(cookiep);
2477 	ASSERT(ccount != 0);
2478 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2479 
2480 	if (dma->dp_window) {
2481 		window = &dma->dp_window[dma->dp_current_win];
2482 		dma->dp_saved_cookies = window->wd_first_cookie;
2483 		window->wd_first_cookie = cookiep;
2484 		ASSERT(ccount == window->wd_cookie_cnt);
2485 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2486 		    + window->wd_first_cookie;
2487 	} else {
2488 		dma->dp_saved_cookies = dma->dp_cookies;
2489 		dma->dp_cookies = cookiep;
2490 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2491 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2492 		    + dma->dp_cookies;
2493 	}
2494 
2495 	dma->dp_need_to_switch_cookies = B_TRUE;
2496 	hp->dmai_cookie = cur_cookiep;
2497 
2498 	return (DDI_SUCCESS);
2499 }
2500 
2501 /*ARGSUSED*/
2502 static int
2503 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2504 {
2505 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2506 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2507 	rootnex_window_t *window;
2508 	ddi_dma_cookie_t *cur_cookiep;
2509 	ddi_dma_cookie_t *cookie_array;
2510 	uint_t ccount;
2511 
2512 	/* check if cookies have not been switched */
2513 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2514 		return (DDI_SUCCESS);
2515 
2516 	ASSERT(dma->dp_saved_cookies);
2517 
2518 	if (dma->dp_window) {
2519 		window = &dma->dp_window[dma->dp_current_win];
2520 		cookie_array = window->wd_first_cookie;
2521 		window->wd_first_cookie = dma->dp_saved_cookies;
2522 		dma->dp_saved_cookies = NULL;
2523 		ccount = window->wd_cookie_cnt;
2524 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2525 		    + window->wd_first_cookie;
2526 	} else {
2527 		cookie_array = dma->dp_cookies;
2528 		dma->dp_cookies = dma->dp_saved_cookies;
2529 		dma->dp_saved_cookies = NULL;
2530 		ccount = dma->dp_sglinfo.si_sgl_size;
2531 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2532 		    + dma->dp_cookies;
2533 	}
2534 
2535 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2536 
2537 	hp->dmai_cookie = cur_cookiep;
2538 
2539 	dma->dp_need_to_switch_cookies = B_FALSE;
2540 
2541 	return (DDI_SUCCESS);
2542 }
2543 
2544 #endif
2545 
2546 static struct as *
2547 rootnex_get_as(ddi_dma_obj_t *dmao)
2548 {
2549 	struct as *asp;
2550 
2551 	switch (dmao->dmao_type) {
2552 	case DMA_OTYP_VADDR:
2553 	case DMA_OTYP_BUFVADDR:
2554 		asp = dmao->dmao_obj.virt_obj.v_as;
2555 		if (asp == NULL)
2556 			asp = &kas;
2557 		break;
2558 	default:
2559 		asp = NULL;
2560 		break;
2561 	}
2562 	return (asp);
2563 }
2564 
2565 /*
2566  * rootnex_verify_buffer()
2567  *   verify buffer wasn't free'd
2568  */
2569 static int
2570 rootnex_verify_buffer(rootnex_dma_t *dma)
2571 {
2572 	page_t **pplist;
2573 	caddr_t vaddr;
2574 	uint_t pcnt;
2575 	uint_t poff;
2576 	page_t *pp;
2577 	char b;
2578 	int i;
2579 
2580 	/* Figure out how many pages this buffer occupies */
2581 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2582 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2583 	} else {
2584 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2585 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2586 	}
2587 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2588 
2589 	switch (dma->dp_dma.dmao_type) {
2590 	case DMA_OTYP_PAGES:
2591 		/*
2592 		 * for a linked list of pp's walk through them to make sure
2593 		 * they're locked and not free.
2594 		 */
2595 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2596 		for (i = 0; i < pcnt; i++) {
2597 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2598 				return (DDI_FAILURE);
2599 			}
2600 			pp = pp->p_next;
2601 		}
2602 		break;
2603 
2604 	case DMA_OTYP_VADDR:
2605 	case DMA_OTYP_BUFVADDR:
2606 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2607 		/*
2608 		 * for an array of pp's walk through them to make sure they're
2609 		 * not free. It's possible that they may not be locked.
2610 		 */
2611 		if (pplist) {
2612 			for (i = 0; i < pcnt; i++) {
2613 				if (PP_ISFREE(pplist[i])) {
2614 					return (DDI_FAILURE);
2615 				}
2616 			}
2617 
2618 		/* For a virtual address, try to peek at each page */
2619 		} else {
2620 			if (rootnex_get_as(&dma->dp_dma) == &kas) {
2621 				for (i = 0; i < pcnt; i++) {
2622 					if (ddi_peek8(NULL, vaddr, &b) ==
2623 					    DDI_FAILURE)
2624 						return (DDI_FAILURE);
2625 					vaddr += MMU_PAGESIZE;
2626 				}
2627 			}
2628 		}
2629 		break;
2630 
2631 	default:
2632 		cmn_err(CE_PANIC, "rootnex_verify_buffer: bad DMA object");
2633 		break;
2634 	}
2635 
2636 	return (DDI_SUCCESS);
2637 }
2638 
2639 
2640 /*
2641  * rootnex_clean_dmahdl()
2642  *    Clean the dma handle. This should be called on a handle alloc and an
2643  *    unbind handle. Set the handle state to the default settings.
2644  */
2645 static void
2646 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2647 {
2648 	rootnex_dma_t *dma;
2649 
2650 
2651 	dma = (rootnex_dma_t *)hp->dmai_private;
2652 
2653 	hp->dmai_nwin = 0;
2654 	dma->dp_current_cookie = 0;
2655 	dma->dp_copybuf_size = 0;
2656 	dma->dp_window = NULL;
2657 	dma->dp_cbaddr = NULL;
2658 	dma->dp_inuse = B_FALSE;
2659 	dma->dp_dvma_used = B_FALSE;
2660 	dma->dp_need_to_free_cookie = B_FALSE;
2661 	dma->dp_need_to_switch_cookies = B_FALSE;
2662 	dma->dp_saved_cookies = NULL;
2663 	dma->dp_sleep_flags = KM_PANIC;
2664 	dma->dp_need_to_free_window = B_FALSE;
2665 	dma->dp_partial_required = B_FALSE;
2666 	dma->dp_trim_required = B_FALSE;
2667 	dma->dp_sglinfo.si_copybuf_req = 0;
2668 #if !defined(__amd64)
2669 	dma->dp_cb_remaping = B_FALSE;
2670 	dma->dp_kva = NULL;
2671 #endif
2672 
2673 	/* FMA related initialization */
2674 	hp->dmai_fault = 0;
2675 	hp->dmai_fault_check = NULL;
2676 	hp->dmai_fault_notify = NULL;
2677 	hp->dmai_error.err_ena = 0;
2678 	hp->dmai_error.err_status = DDI_FM_OK;
2679 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2680 	hp->dmai_error.err_ontrap = NULL;
2681 }
2682 
2683 
2684 /*
2685  * rootnex_valid_alloc_parms()
2686  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2687  */
2688 static int
2689 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2690 {
2691 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2692 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2693 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2694 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2695 		return (DDI_DMA_BADATTR);
2696 	}
2697 
2698 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2699 		return (DDI_DMA_BADATTR);
2700 	}
2701 
2702 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2703 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2704 	    attr->dma_attr_sgllen == 0) {
2705 		return (DDI_DMA_BADATTR);
2706 	}
2707 
2708 	/* We should be able to DMA into every byte offset in a page */
2709 	if (maxsegmentsize < MMU_PAGESIZE) {
2710 		return (DDI_DMA_BADATTR);
2711 	}
2712 
2713 	/* if we're bouncing on seg, seg must be <= addr_hi */
2714 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2715 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2716 		return (DDI_DMA_BADATTR);
2717 	}
2718 	return (DDI_SUCCESS);
2719 }
2720 
2721 /*
2722  * rootnex_valid_bind_parms()
2723  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2724  */
2725 /* ARGSUSED */
2726 static int
2727 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2728 {
2729 #if !defined(__amd64)
2730 	/*
2731 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2732 	 * we can track the offset for the obsoleted interfaces.
2733 	 */
2734 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2735 		return (DDI_DMA_TOOBIG);
2736 	}
2737 #endif
2738 
2739 	return (DDI_SUCCESS);
2740 }
2741 
2742 
2743 /*
2744  * rootnex_need_bounce_seg()
2745  *    check to see if the buffer lives on both side of the seg.
2746  */
2747 static boolean_t
2748 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2749 {
2750 	ddi_dma_atyp_t buftype;
2751 	rootnex_addr_t raddr;
2752 	boolean_t lower_addr;
2753 	boolean_t upper_addr;
2754 	uint64_t offset;
2755 	page_t **pplist;
2756 	uint64_t paddr;
2757 	uint32_t psize;
2758 	uint32_t size;
2759 	caddr_t vaddr;
2760 	uint_t pcnt;
2761 	page_t *pp;
2762 
2763 
2764 	/* shortcuts */
2765 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2766 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2767 	buftype = dmar_object->dmao_type;
2768 	size = dmar_object->dmao_size;
2769 
2770 	lower_addr = B_FALSE;
2771 	upper_addr = B_FALSE;
2772 	pcnt = 0;
2773 
2774 	/*
2775 	 * Process the first page to handle the initial offset of the buffer.
2776 	 * We'll use the base address we get later when we loop through all
2777 	 * the pages.
2778 	 */
2779 	if (buftype == DMA_OTYP_PAGES) {
2780 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2781 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2782 		    MMU_PAGEOFFSET;
2783 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2784 		psize = MIN(size, (MMU_PAGESIZE - offset));
2785 		pp = pp->p_next;
2786 		sglinfo->si_asp = NULL;
2787 	} else if (pplist != NULL) {
2788 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2789 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2790 		if (sglinfo->si_asp == NULL) {
2791 			sglinfo->si_asp = &kas;
2792 		}
2793 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2794 		paddr += offset;
2795 		psize = MIN(size, (MMU_PAGESIZE - offset));
2796 		pcnt++;
2797 	} else {
2798 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2799 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2800 		if (sglinfo->si_asp == NULL) {
2801 			sglinfo->si_asp = &kas;
2802 		}
2803 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2804 		paddr += offset;
2805 		psize = MIN(size, (MMU_PAGESIZE - offset));
2806 		vaddr += psize;
2807 	}
2808 
2809 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2810 
2811 	if ((raddr + psize) > sglinfo->si_segmask) {
2812 		upper_addr = B_TRUE;
2813 	} else {
2814 		lower_addr = B_TRUE;
2815 	}
2816 	size -= psize;
2817 
2818 	/*
2819 	 * Walk through the rest of the pages in the buffer. Track to see
2820 	 * if we have pages on both sides of the segment boundary.
2821 	 */
2822 	while (size > 0) {
2823 		/* partial or full page */
2824 		psize = MIN(size, MMU_PAGESIZE);
2825 
2826 		if (buftype == DMA_OTYP_PAGES) {
2827 			/* get the paddr from the page_t */
2828 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2829 			paddr = pfn_to_pa(pp->p_pagenum);
2830 			pp = pp->p_next;
2831 		} else if (pplist != NULL) {
2832 			/* index into the array of page_t's to get the paddr */
2833 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2834 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2835 			pcnt++;
2836 		} else {
2837 			/* call into the VM to get the paddr */
2838 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2839 			    vaddr));
2840 			vaddr += psize;
2841 		}
2842 
2843 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2844 
2845 		if ((raddr + psize) > sglinfo->si_segmask) {
2846 			upper_addr = B_TRUE;
2847 		} else {
2848 			lower_addr = B_TRUE;
2849 		}
2850 		/*
2851 		 * if the buffer lives both above and below the segment
2852 		 * boundary, or the current page is the page immediately
2853 		 * after the segment, we will use a copy/bounce buffer for
2854 		 * all pages > seg.
2855 		 */
2856 		if ((lower_addr && upper_addr) ||
2857 		    (raddr == (sglinfo->si_segmask + 1))) {
2858 			return (B_TRUE);
2859 		}
2860 
2861 		size -= psize;
2862 	}
2863 
2864 	return (B_FALSE);
2865 }
2866 
2867 /*
2868  * rootnex_get_sgl()
2869  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2870  *    with a call to the vm layer when vm2.0 comes around...
2871  */
2872 static void
2873 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2874     rootnex_sglinfo_t *sglinfo)
2875 {
2876 	ddi_dma_atyp_t buftype;
2877 	rootnex_addr_t raddr;
2878 	uint64_t last_page;
2879 	uint64_t offset;
2880 	uint64_t addrhi;
2881 	uint64_t addrlo;
2882 	uint64_t maxseg;
2883 	page_t **pplist;
2884 	uint64_t paddr;
2885 	uint32_t psize;
2886 	uint32_t size;
2887 	caddr_t vaddr;
2888 	uint_t pcnt;
2889 	page_t *pp;
2890 	uint_t cnt;
2891 
2892 
2893 	/* shortcuts */
2894 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2895 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2896 	maxseg = sglinfo->si_max_cookie_size;
2897 	buftype = dmar_object->dmao_type;
2898 	addrhi = sglinfo->si_max_addr;
2899 	addrlo = sglinfo->si_min_addr;
2900 	size = dmar_object->dmao_size;
2901 
2902 	pcnt = 0;
2903 	cnt = 0;
2904 
2905 
2906 	/*
2907 	 * check to see if we need to use the copy buffer for pages over
2908 	 * the segment attr.
2909 	 */
2910 	sglinfo->si_bounce_on_seg = B_FALSE;
2911 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2912 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2913 		    dmar_object, sglinfo);
2914 	}
2915 
2916 	/*
2917 	 * if we were passed down a linked list of pages, i.e. pointer to
2918 	 * page_t, use this to get our physical address and buf offset.
2919 	 */
2920 	if (buftype == DMA_OTYP_PAGES) {
2921 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2922 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2923 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2924 		    MMU_PAGEOFFSET;
2925 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2926 		psize = MIN(size, (MMU_PAGESIZE - offset));
2927 		pp = pp->p_next;
2928 		sglinfo->si_asp = NULL;
2929 
2930 	/*
2931 	 * We weren't passed down a linked list of pages, but if we were passed
2932 	 * down an array of pages, use this to get our physical address and buf
2933 	 * offset.
2934 	 */
2935 	} else if (pplist != NULL) {
2936 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2937 		    (buftype == DMA_OTYP_BUFVADDR));
2938 
2939 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2940 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2941 		if (sglinfo->si_asp == NULL) {
2942 			sglinfo->si_asp = &kas;
2943 		}
2944 
2945 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2946 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2947 		paddr += offset;
2948 		psize = MIN(size, (MMU_PAGESIZE - offset));
2949 		pcnt++;
2950 
2951 	/*
2952 	 * All we have is a virtual address, we'll need to call into the VM
2953 	 * to get the physical address.
2954 	 */
2955 	} else {
2956 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2957 		    (buftype == DMA_OTYP_BUFVADDR));
2958 
2959 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2960 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2961 		if (sglinfo->si_asp == NULL) {
2962 			sglinfo->si_asp = &kas;
2963 		}
2964 
2965 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2966 		paddr += offset;
2967 		psize = MIN(size, (MMU_PAGESIZE - offset));
2968 		vaddr += psize;
2969 	}
2970 
2971 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2972 
2973 	/*
2974 	 * Setup the first cookie with the physical address of the page and the
2975 	 * size of the page (which takes into account the initial offset into
2976 	 * the page.
2977 	 */
2978 	sgl[cnt].dmac_laddress = raddr;
2979 	sgl[cnt].dmac_size = psize;
2980 	sgl[cnt].dmac_type = 0;
2981 
2982 	/*
2983 	 * Save away the buffer offset into the page. We'll need this later in
2984 	 * the copy buffer code to help figure out the page index within the
2985 	 * buffer and the offset into the current page.
2986 	 */
2987 	sglinfo->si_buf_offset = offset;
2988 
2989 	/*
2990 	 * If we are using the copy buffer for anything over the segment
2991 	 * boundary, and this page is over the segment boundary.
2992 	 *   OR
2993 	 * if the DMA engine can't reach the physical address.
2994 	 */
2995 	if (((sglinfo->si_bounce_on_seg) &&
2996 	    ((raddr + psize) > sglinfo->si_segmask)) ||
2997 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2998 		/*
2999 		 * Increase how much copy buffer we use. We always increase by
3000 		 * pagesize so we don't have to worry about converting offsets.
3001 		 * Set a flag in the cookies dmac_type to indicate that it uses
3002 		 * the copy buffer. If this isn't the last cookie, go to the
3003 		 * next cookie (since we separate each page which uses the copy
3004 		 * buffer in case the copy buffer is not physically contiguous.
3005 		 */
3006 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
3007 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3008 		if ((cnt + 1) < sglinfo->si_max_pages) {
3009 			cnt++;
3010 			sgl[cnt].dmac_laddress = 0;
3011 			sgl[cnt].dmac_size = 0;
3012 			sgl[cnt].dmac_type = 0;
3013 		}
3014 	}
3015 
3016 	/*
3017 	 * save this page's physical address so we can figure out if the next
3018 	 * page is physically contiguous. Keep decrementing size until we are
3019 	 * done with the buffer.
3020 	 */
3021 	last_page = raddr & MMU_PAGEMASK;
3022 	size -= psize;
3023 
3024 	while (size > 0) {
3025 		/* Get the size for this page (i.e. partial or full page) */
3026 		psize = MIN(size, MMU_PAGESIZE);
3027 
3028 		if (buftype == DMA_OTYP_PAGES) {
3029 			/* get the paddr from the page_t */
3030 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
3031 			paddr = pfn_to_pa(pp->p_pagenum);
3032 			pp = pp->p_next;
3033 		} else if (pplist != NULL) {
3034 			/* index into the array of page_t's to get the paddr */
3035 			ASSERT(!PP_ISFREE(pplist[pcnt]));
3036 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
3037 			pcnt++;
3038 		} else {
3039 			/* call into the VM to get the paddr */
3040 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
3041 			    vaddr));
3042 			vaddr += psize;
3043 		}
3044 
3045 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
3046 
3047 		/*
3048 		 * If we are using the copy buffer for anything over the
3049 		 * segment boundary, and this page is over the segment
3050 		 * boundary.
3051 		 *   OR
3052 		 * if the DMA engine can't reach the physical address.
3053 		 */
3054 		if (((sglinfo->si_bounce_on_seg) &&
3055 		    ((raddr + psize) > sglinfo->si_segmask)) ||
3056 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
3057 
3058 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
3059 
3060 			/*
3061 			 * if there is something in the current cookie, go to
3062 			 * the next one. We only want one page in a cookie which
3063 			 * uses the copybuf since the copybuf doesn't have to
3064 			 * be physically contiguous.
3065 			 */
3066 			if (sgl[cnt].dmac_size != 0) {
3067 				cnt++;
3068 			}
3069 			sgl[cnt].dmac_laddress = raddr;
3070 			sgl[cnt].dmac_size = psize;
3071 #if defined(__amd64)
3072 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3073 #else
3074 			/*
3075 			 * save the buf offset for 32-bit kernel. used in the
3076 			 * obsoleted interfaces.
3077 			 */
3078 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
3079 			    (dmar_object->dmao_size - size);
3080 #endif
3081 			/* if this isn't the last cookie, go to the next one */
3082 			if ((cnt + 1) < sglinfo->si_max_pages) {
3083 				cnt++;
3084 				sgl[cnt].dmac_laddress = 0;
3085 				sgl[cnt].dmac_size = 0;
3086 				sgl[cnt].dmac_type = 0;
3087 			}
3088 
3089 		/*
3090 		 * this page didn't need the copy buffer, if it's not physically
3091 		 * contiguous, or it would put us over a segment boundary, or it
3092 		 * puts us over the max cookie size, or the current sgl doesn't
3093 		 * have anything in it.
3094 		 */
3095 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
3096 		    !(raddr & sglinfo->si_segmask) ||
3097 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3098 		    (sgl[cnt].dmac_size == 0)) {
3099 			/*
3100 			 * if we're not already in a new cookie, go to the next
3101 			 * cookie.
3102 			 */
3103 			if (sgl[cnt].dmac_size != 0) {
3104 				cnt++;
3105 			}
3106 
3107 			/* save the cookie information */
3108 			sgl[cnt].dmac_laddress = raddr;
3109 			sgl[cnt].dmac_size = psize;
3110 #if defined(__amd64)
3111 			sgl[cnt].dmac_type = 0;
3112 #else
3113 			/*
3114 			 * save the buf offset for 32-bit kernel. used in the
3115 			 * obsoleted interfaces.
3116 			 */
3117 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
3118 #endif
3119 
3120 		/*
3121 		 * this page didn't need the copy buffer, it is physically
3122 		 * contiguous with the last page, and it's <= the max cookie
3123 		 * size.
3124 		 */
3125 		} else {
3126 			sgl[cnt].dmac_size += psize;
3127 
3128 			/*
3129 			 * if this exactly ==  the maximum cookie size, and
3130 			 * it isn't the last cookie, go to the next cookie.
3131 			 */
3132 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3133 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3134 				cnt++;
3135 				sgl[cnt].dmac_laddress = 0;
3136 				sgl[cnt].dmac_size = 0;
3137 				sgl[cnt].dmac_type = 0;
3138 			}
3139 		}
3140 
3141 		/*
3142 		 * save this page's physical address so we can figure out if the
3143 		 * next page is physically contiguous. Keep decrementing size
3144 		 * until we are done with the buffer.
3145 		 */
3146 		last_page = raddr;
3147 		size -= psize;
3148 	}
3149 
3150 	/* we're done, save away how many cookies the sgl has */
3151 	if (sgl[cnt].dmac_size == 0) {
3152 		ASSERT(cnt < sglinfo->si_max_pages);
3153 		sglinfo->si_sgl_size = cnt;
3154 	} else {
3155 		sglinfo->si_sgl_size = cnt + 1;
3156 	}
3157 }
3158 
3159 static void
3160 rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
3161     rootnex_sglinfo_t *sglinfo)
3162 {
3163 	uint64_t offset;
3164 	uint64_t maxseg;
3165 	uint64_t dvaddr;
3166 	struct dvmaseg *dvs;
3167 	uint64_t paddr;
3168 	uint32_t psize, ssize;
3169 	uint32_t size;
3170 	uint_t cnt;
3171 	int physcontig;
3172 
3173 	ASSERT(dmar_object->dmao_type == DMA_OTYP_DVADDR);
3174 
3175 	/* shortcuts */
3176 	maxseg = sglinfo->si_max_cookie_size;
3177 	size = dmar_object->dmao_size;
3178 
3179 	cnt = 0;
3180 	sglinfo->si_bounce_on_seg = B_FALSE;
3181 
3182 	dvs = dmar_object->dmao_obj.dvma_obj.dv_seg;
3183 	offset = dmar_object->dmao_obj.dvma_obj.dv_off;
3184 	ssize = dvs->dvs_len;
3185 	paddr = dvs->dvs_start;
3186 	paddr += offset;
3187 	psize = MIN(ssize, (maxseg - offset));
3188 	dvaddr = paddr + psize;
3189 	ssize -= psize;
3190 
3191 	sgl[cnt].dmac_laddress = paddr;
3192 	sgl[cnt].dmac_size = psize;
3193 	sgl[cnt].dmac_type = 0;
3194 
3195 	size -= psize;
3196 	while (size > 0) {
3197 		if (ssize == 0) {
3198 			dvs++;
3199 			ssize = dvs->dvs_len;
3200 			dvaddr = dvs->dvs_start;
3201 			physcontig = 0;
3202 		} else
3203 			physcontig = 1;
3204 
3205 		paddr = dvaddr;
3206 		psize = MIN(ssize, maxseg);
3207 		dvaddr += psize;
3208 		ssize -= psize;
3209 
3210 		if (!physcontig || !(paddr & sglinfo->si_segmask) ||
3211 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3212 		    (sgl[cnt].dmac_size == 0)) {
3213 			/*
3214 			 * if we're not already in a new cookie, go to the next
3215 			 * cookie.
3216 			 */
3217 			if (sgl[cnt].dmac_size != 0) {
3218 				cnt++;
3219 			}
3220 
3221 			/* save the cookie information */
3222 			sgl[cnt].dmac_laddress = paddr;
3223 			sgl[cnt].dmac_size = psize;
3224 			sgl[cnt].dmac_type = 0;
3225 		} else {
3226 			sgl[cnt].dmac_size += psize;
3227 
3228 			/*
3229 			 * if this exactly ==  the maximum cookie size, and
3230 			 * it isn't the last cookie, go to the next cookie.
3231 			 */
3232 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3233 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3234 				cnt++;
3235 				sgl[cnt].dmac_laddress = 0;
3236 				sgl[cnt].dmac_size = 0;
3237 				sgl[cnt].dmac_type = 0;
3238 			}
3239 		}
3240 		size -= psize;
3241 	}
3242 
3243 	/* we're done, save away how many cookies the sgl has */
3244 	if (sgl[cnt].dmac_size == 0) {
3245 		sglinfo->si_sgl_size = cnt;
3246 	} else {
3247 		sglinfo->si_sgl_size = cnt + 1;
3248 	}
3249 }
3250 
3251 /*
3252  * rootnex_bind_slowpath()
3253  *    Call in the bind path if the calling driver can't use the sgl without
3254  *    modifying it. We either need to use the copy buffer and/or we will end up
3255  *    with a partial bind.
3256  */
3257 static int
3258 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3259     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3260 {
3261 	rootnex_sglinfo_t *sinfo;
3262 	rootnex_window_t *window;
3263 	ddi_dma_cookie_t *cookie;
3264 	size_t copybuf_used;
3265 	size_t dmac_size;
3266 	boolean_t partial;
3267 	off_t cur_offset;
3268 	page_t *cur_pp;
3269 	major_t mnum;
3270 	int e;
3271 	int i;
3272 
3273 
3274 	sinfo = &dma->dp_sglinfo;
3275 	copybuf_used = 0;
3276 	partial = B_FALSE;
3277 
3278 	/*
3279 	 * If we're using the copybuf, set the copybuf state in dma struct.
3280 	 * Needs to be first since it sets the copy buffer size.
3281 	 */
3282 	if (sinfo->si_copybuf_req != 0) {
3283 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
3284 		if (e != DDI_SUCCESS) {
3285 			return (e);
3286 		}
3287 	} else {
3288 		dma->dp_copybuf_size = 0;
3289 	}
3290 
3291 	/*
3292 	 * Figure out if we need to do a partial mapping. If so, figure out
3293 	 * if we need to trim the buffers when we munge the sgl.
3294 	 */
3295 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
3296 	    (dmao->dmao_size > dma->dp_maxxfer) ||
3297 	    ((unsigned)attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
3298 		dma->dp_partial_required = B_TRUE;
3299 		if (attr->dma_attr_granular != 1) {
3300 			dma->dp_trim_required = B_TRUE;
3301 		}
3302 	} else {
3303 		dma->dp_partial_required = B_FALSE;
3304 		dma->dp_trim_required = B_FALSE;
3305 	}
3306 
3307 	/* If we need to do a partial bind, make sure the driver supports it */
3308 	if (dma->dp_partial_required &&
3309 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3310 
3311 		mnum = ddi_driver_major(dma->dp_dip);
3312 		/*
3313 		 * patchable which allows us to print one warning per major
3314 		 * number.
3315 		 */
3316 		if ((rootnex_bind_warn) &&
3317 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3318 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3319 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3320 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3321 			    "There is a small risk of data corruption in "
3322 			    "particular with large I/Os. The driver should be "
3323 			    "replaced with a corrected version for proper "
3324 			    "system operation. To disable this warning, add "
3325 			    "'set rootnex:rootnex_bind_warn=0' to "
3326 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3327 		}
3328 		return (DDI_DMA_TOOBIG);
3329 	}
3330 
3331 	/*
3332 	 * we might need multiple windows, setup state to handle them. In this
3333 	 * code path, we will have at least one window.
3334 	 */
3335 	e = rootnex_setup_windows(hp, dma, attr, dmao, kmflag);
3336 	if (e != DDI_SUCCESS) {
3337 		rootnex_teardown_copybuf(dma);
3338 		return (e);
3339 	}
3340 
3341 	window = &dma->dp_window[0];
3342 	cookie = &dma->dp_cookies[0];
3343 	cur_offset = 0;
3344 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3345 	if (dmao->dmao_type == DMA_OTYP_PAGES) {
3346 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3347 	}
3348 
3349 	/* loop though all the cookies we got back from get_sgl() */
3350 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3351 		/*
3352 		 * If we're using the copy buffer, check this cookie and setup
3353 		 * its associated copy buffer state. If this cookie uses the
3354 		 * copy buffer, make sure we sync this window during dma_sync.
3355 		 */
3356 		if (dma->dp_copybuf_size > 0) {
3357 			rootnex_setup_cookie(dmao, dma, cookie,
3358 			    cur_offset, &copybuf_used, &cur_pp);
3359 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3360 				window->wd_dosync = B_TRUE;
3361 			}
3362 		}
3363 
3364 		/*
3365 		 * save away the cookie size, since it could be modified in
3366 		 * the windowing code.
3367 		 */
3368 		dmac_size = cookie->dmac_size;
3369 
3370 		/* if we went over max copybuf size */
3371 		if (dma->dp_copybuf_size &&
3372 		    (copybuf_used > dma->dp_copybuf_size)) {
3373 			partial = B_TRUE;
3374 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3375 			    cookie, cur_offset, &copybuf_used);
3376 			if (e != DDI_SUCCESS) {
3377 				rootnex_teardown_copybuf(dma);
3378 				rootnex_teardown_windows(dma);
3379 				return (e);
3380 			}
3381 
3382 			/*
3383 			 * if the coookie uses the copy buffer, make sure the
3384 			 * new window we just moved to is set to sync.
3385 			 */
3386 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3387 				window->wd_dosync = B_TRUE;
3388 			}
3389 			ROOTNEX_DPROBE1(rootnex__copybuf__window, dev_info_t *,
3390 			    dma->dp_dip);
3391 
3392 		/* if the cookie cnt == max sgllen, move to the next window */
3393 		} else if (window->wd_cookie_cnt >=
3394 		    (unsigned)attr->dma_attr_sgllen) {
3395 			partial = B_TRUE;
3396 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3397 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3398 			    cookie, attr, cur_offset);
3399 			if (e != DDI_SUCCESS) {
3400 				rootnex_teardown_copybuf(dma);
3401 				rootnex_teardown_windows(dma);
3402 				return (e);
3403 			}
3404 
3405 			/*
3406 			 * if the coookie uses the copy buffer, make sure the
3407 			 * new window we just moved to is set to sync.
3408 			 */
3409 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3410 				window->wd_dosync = B_TRUE;
3411 			}
3412 			ROOTNEX_DPROBE1(rootnex__sgllen__window, dev_info_t *,
3413 			    dma->dp_dip);
3414 
3415 		/* else if we will be over maxxfer */
3416 		} else if ((window->wd_size + dmac_size) >
3417 		    dma->dp_maxxfer) {
3418 			partial = B_TRUE;
3419 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3420 			    cookie);
3421 			if (e != DDI_SUCCESS) {
3422 				rootnex_teardown_copybuf(dma);
3423 				rootnex_teardown_windows(dma);
3424 				return (e);
3425 			}
3426 
3427 			/*
3428 			 * if the coookie uses the copy buffer, make sure the
3429 			 * new window we just moved to is set to sync.
3430 			 */
3431 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3432 				window->wd_dosync = B_TRUE;
3433 			}
3434 			ROOTNEX_DPROBE1(rootnex__maxxfer__window, dev_info_t *,
3435 			    dma->dp_dip);
3436 
3437 		/* else this cookie fits in the current window */
3438 		} else {
3439 			window->wd_cookie_cnt++;
3440 			window->wd_size += dmac_size;
3441 		}
3442 
3443 		/* track our offset into the buffer, go to the next cookie */
3444 		ASSERT(dmac_size <= dmao->dmao_size);
3445 		ASSERT(cookie->dmac_size <= dmac_size);
3446 		cur_offset += dmac_size;
3447 		cookie++;
3448 	}
3449 
3450 	/* if we ended up with a zero sized window in the end, clean it up */
3451 	if (window->wd_size == 0) {
3452 		hp->dmai_nwin--;
3453 		window--;
3454 	}
3455 
3456 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3457 
3458 	if (!partial) {
3459 		return (DDI_DMA_MAPPED);
3460 	}
3461 
3462 	ASSERT(dma->dp_partial_required);
3463 	return (DDI_DMA_PARTIAL_MAP);
3464 }
3465 
3466 /*
3467  * rootnex_setup_copybuf()
3468  *    Called in bind slowpath. Figures out if we're going to use the copy
3469  *    buffer, and if we do, sets up the basic state to handle it.
3470  */
3471 static int
3472 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3473     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3474 {
3475 	rootnex_sglinfo_t *sinfo;
3476 	ddi_dma_attr_t lattr;
3477 	size_t max_copybuf;
3478 	int cansleep;
3479 	int e;
3480 #if !defined(__amd64)
3481 	int vmflag;
3482 #endif
3483 
3484 	ASSERT(!dma->dp_dvma_used);
3485 
3486 	sinfo = &dma->dp_sglinfo;
3487 
3488 	/* read this first so it's consistent through the routine  */
3489 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3490 
3491 	/* We need to call into the rootnex on ddi_dma_sync() */
3492 	hp->dmai_rflags &= ~DMP_NOSYNC;
3493 
3494 	/* make sure the copybuf size <= the max size */
3495 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3496 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3497 
3498 #if !defined(__amd64)
3499 	/*
3500 	 * if we don't have kva space to copy to/from, allocate the KVA space
3501 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3502 	 * the 64-bit kernel.
3503 	 */
3504 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3505 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3506 
3507 		/* convert the sleep flags */
3508 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3509 			vmflag = VM_SLEEP;
3510 		} else {
3511 			vmflag = VM_NOSLEEP;
3512 		}
3513 
3514 		/* allocate Kernel VA space that we can bcopy to/from */
3515 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3516 		    vmflag);
3517 		if (dma->dp_kva == NULL) {
3518 			return (DDI_DMA_NORESOURCES);
3519 		}
3520 	}
3521 #endif
3522 
3523 	/* convert the sleep flags */
3524 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3525 		cansleep = 1;
3526 	} else {
3527 		cansleep = 0;
3528 	}
3529 
3530 	/*
3531 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3532 	 * engine limits, so we can't use kmem_alloc... We don't need
3533 	 * contiguous memory (sgllen) since we will be forcing windows on
3534 	 * sgllen anyway.
3535 	 */
3536 	lattr = *attr;
3537 	lattr.dma_attr_align = MMU_PAGESIZE;
3538 	lattr.dma_attr_sgllen = -1;	/* no limit */
3539 	/*
3540 	 * if we're using the copy buffer because of seg, use that for our
3541 	 * upper address limit.
3542 	 */
3543 	if (sinfo->si_bounce_on_seg) {
3544 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3545 	}
3546 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3547 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3548 	if (e != DDI_SUCCESS) {
3549 #if !defined(__amd64)
3550 		if (dma->dp_kva != NULL) {
3551 			vmem_free(heap_arena, dma->dp_kva,
3552 			    dma->dp_copybuf_size);
3553 		}
3554 #endif
3555 		return (DDI_DMA_NORESOURCES);
3556 	}
3557 
3558 	ROOTNEX_DPROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3559 	    size_t, dma->dp_copybuf_size);
3560 
3561 	return (DDI_SUCCESS);
3562 }
3563 
3564 
3565 /*
3566  * rootnex_setup_windows()
3567  *    Called in bind slowpath to setup the window state. We always have windows
3568  *    in the slowpath. Even if the window count = 1.
3569  */
3570 static int
3571 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3572     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3573 {
3574 	rootnex_window_t *windowp;
3575 	rootnex_sglinfo_t *sinfo;
3576 	size_t copy_state_size;
3577 	size_t win_state_size;
3578 	size_t state_available;
3579 	size_t space_needed;
3580 	uint_t copybuf_win;
3581 	uint_t maxxfer_win;
3582 	size_t space_used;
3583 	uint_t sglwin;
3584 
3585 
3586 	sinfo = &dma->dp_sglinfo;
3587 
3588 	dma->dp_current_win = 0;
3589 	hp->dmai_nwin = 0;
3590 
3591 	/* If we don't need to do a partial, we only have one window */
3592 	if (!dma->dp_partial_required) {
3593 		dma->dp_max_win = 1;
3594 
3595 	/*
3596 	 * we need multiple windows, need to figure out the worse case number
3597 	 * of windows.
3598 	 */
3599 	} else {
3600 		/*
3601 		 * if we need windows because we need more copy buffer that
3602 		 * we allow, the worse case number of windows we could need
3603 		 * here would be (copybuf space required / copybuf space that
3604 		 * we have) plus one for remainder, and plus 2 to handle the
3605 		 * extra pages on the trim for the first and last pages of the
3606 		 * buffer (a page is the minimum window size so under the right
3607 		 * attr settings, you could have a window for each page).
3608 		 * The last page will only be hit here if the size is not a
3609 		 * multiple of the granularity (which theoretically shouldn't
3610 		 * be the case but never has been enforced, so we could have
3611 		 * broken things without it).
3612 		 */
3613 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3614 			ASSERT(dma->dp_copybuf_size > 0);
3615 			copybuf_win = (sinfo->si_copybuf_req /
3616 			    dma->dp_copybuf_size) + 1 + 2;
3617 		} else {
3618 			copybuf_win = 0;
3619 		}
3620 
3621 		/*
3622 		 * if we need windows because we have more cookies than the H/W
3623 		 * can handle, the number of windows we would need here would
3624 		 * be (cookie count / cookies count H/W supports minus 1[for
3625 		 * trim]) plus one for remainder.
3626 		 */
3627 		if ((unsigned)attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3628 			sglwin = (sinfo->si_sgl_size /
3629 			    (attr->dma_attr_sgllen - 1)) + 1;
3630 		} else {
3631 			sglwin = 0;
3632 		}
3633 
3634 		/*
3635 		 * if we need windows because we're binding more memory than the
3636 		 * H/W can transfer at once, the number of windows we would need
3637 		 * here would be (xfer count / max xfer H/W supports) plus one
3638 		 * for remainder, and plus 2 to handle the extra pages on the
3639 		 * trim (see above comment about trim)
3640 		 */
3641 		if (dmao->dmao_size > dma->dp_maxxfer) {
3642 			maxxfer_win = (dmao->dmao_size /
3643 			    dma->dp_maxxfer) + 1 + 2;
3644 		} else {
3645 			maxxfer_win = 0;
3646 		}
3647 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3648 		ASSERT(dma->dp_max_win > 0);
3649 	}
3650 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3651 
3652 	/*
3653 	 * Get space for window and potential copy buffer state. Before we
3654 	 * go and allocate memory, see if we can get away with using what's
3655 	 * left in the pre-allocted state or the dynamically allocated sgl.
3656 	 */
3657 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3658 	    sizeof (ddi_dma_cookie_t));
3659 
3660 	/* if we dynamically allocated space for the cookies */
3661 	if (dma->dp_need_to_free_cookie) {
3662 		/* if we have more space in the pre-allocted buffer, use it */
3663 		ASSERT(space_used <= dma->dp_cookie_size);
3664 		if ((dma->dp_cookie_size - space_used) <=
3665 		    rootnex_state->r_prealloc_size) {
3666 			state_available = rootnex_state->r_prealloc_size;
3667 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3668 
3669 		/*
3670 		 * else, we have more free space in the dynamically allocated
3671 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3672 		 * didn't need a lot of cookies.
3673 		 */
3674 		} else {
3675 			state_available = dma->dp_cookie_size - space_used;
3676 			windowp = (rootnex_window_t *)
3677 			    &dma->dp_cookies[sinfo->si_sgl_size];
3678 		}
3679 
3680 	/* we used the pre-alloced buffer */
3681 	} else {
3682 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3683 		state_available = rootnex_state->r_prealloc_size - space_used;
3684 		windowp = (rootnex_window_t *)
3685 		    &dma->dp_cookies[sinfo->si_sgl_size];
3686 	}
3687 
3688 	/*
3689 	 * figure out how much state we need to track the copy buffer. Add an
3690 	 * addition 8 bytes for pointer alignemnt later.
3691 	 */
3692 	if (dma->dp_copybuf_size > 0) {
3693 		copy_state_size = sinfo->si_max_pages *
3694 		    sizeof (rootnex_pgmap_t);
3695 	} else {
3696 		copy_state_size = 0;
3697 	}
3698 	/* add an additional 8 bytes for pointer alignment */
3699 	space_needed = win_state_size + copy_state_size + 0x8;
3700 
3701 	/* if we have enough space already, use it */
3702 	if (state_available >= space_needed) {
3703 		dma->dp_window = windowp;
3704 		dma->dp_need_to_free_window = B_FALSE;
3705 
3706 	/* not enough space, need to allocate more. */
3707 	} else {
3708 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3709 		if (dma->dp_window == NULL) {
3710 			return (DDI_DMA_NORESOURCES);
3711 		}
3712 		dma->dp_need_to_free_window = B_TRUE;
3713 		dma->dp_window_size = space_needed;
3714 		ROOTNEX_DPROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3715 		    dma->dp_dip, size_t, space_needed);
3716 	}
3717 
3718 	/*
3719 	 * we allocate copy buffer state and window state at the same time.
3720 	 * setup our copy buffer state pointers. Make sure it's aligned.
3721 	 */
3722 	if (dma->dp_copybuf_size > 0) {
3723 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3724 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3725 
3726 #if !defined(__amd64)
3727 		/*
3728 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3729 		 * false/NULL. Should be quicker to bzero vs loop and set.
3730 		 */
3731 		bzero(dma->dp_pgmap, copy_state_size);
3732 #endif
3733 	} else {
3734 		dma->dp_pgmap = NULL;
3735 	}
3736 
3737 	return (DDI_SUCCESS);
3738 }
3739 
3740 
3741 /*
3742  * rootnex_teardown_copybuf()
3743  *    cleans up after rootnex_setup_copybuf()
3744  */
3745 static void
3746 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3747 {
3748 #if !defined(__amd64)
3749 	int i;
3750 
3751 	/*
3752 	 * if we allocated kernel heap VMEM space, go through all the pages and
3753 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3754 	 * arena. Then free the VMEM space.
3755 	 */
3756 	if (dma->dp_kva != NULL) {
3757 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3758 			if (dma->dp_pgmap[i].pm_mapped) {
3759 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3760 				    MMU_PAGESIZE, HAT_UNLOAD);
3761 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3762 			}
3763 		}
3764 
3765 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3766 	}
3767 
3768 #endif
3769 
3770 	/* if we allocated a copy buffer, free it */
3771 	if (dma->dp_cbaddr != NULL) {
3772 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3773 	}
3774 }
3775 
3776 
3777 /*
3778  * rootnex_teardown_windows()
3779  *    cleans up after rootnex_setup_windows()
3780  */
3781 static void
3782 rootnex_teardown_windows(rootnex_dma_t *dma)
3783 {
3784 	/*
3785 	 * if we had to allocate window state on the last bind (because we
3786 	 * didn't have enough pre-allocated space in the handle), free it.
3787 	 */
3788 	if (dma->dp_need_to_free_window) {
3789 		kmem_free(dma->dp_window, dma->dp_window_size);
3790 	}
3791 }
3792 
3793 
3794 /*
3795  * rootnex_init_win()
3796  *    Called in bind slow path during creation of a new window. Initializes
3797  *    window state to default values.
3798  */
3799 /*ARGSUSED*/
3800 static void
3801 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3802     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3803 {
3804 	hp->dmai_nwin++;
3805 	window->wd_dosync = B_FALSE;
3806 	window->wd_offset = cur_offset;
3807 	window->wd_size = 0;
3808 	window->wd_first_cookie = cookie;
3809 	window->wd_cookie_cnt = 0;
3810 	window->wd_trim.tr_trim_first = B_FALSE;
3811 	window->wd_trim.tr_trim_last = B_FALSE;
3812 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3813 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3814 #if !defined(__amd64)
3815 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3816 #endif
3817 }
3818 
3819 
3820 /*
3821  * rootnex_setup_cookie()
3822  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3823  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3824  *    out if it uses the copy buffer, and if it does, save away everything we'll
3825  *    need during sync.
3826  */
3827 static void
3828 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3829     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3830     page_t **cur_pp)
3831 {
3832 	boolean_t copybuf_sz_power_2;
3833 	rootnex_sglinfo_t *sinfo;
3834 	paddr_t paddr;
3835 	uint_t pidx;
3836 	uint_t pcnt;
3837 	off_t poff;
3838 #if defined(__amd64)
3839 	pfn_t pfn;
3840 #else
3841 	page_t **pplist;
3842 #endif
3843 
3844 	ASSERT(dmar_object->dmao_type != DMA_OTYP_DVADDR);
3845 
3846 	sinfo = &dma->dp_sglinfo;
3847 
3848 	/*
3849 	 * Calculate the page index relative to the start of the buffer. The
3850 	 * index to the current page for our buffer is the offset into the
3851 	 * first page of the buffer plus our current offset into the buffer
3852 	 * itself, shifted of course...
3853 	 */
3854 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3855 	ASSERT(pidx < sinfo->si_max_pages);
3856 
3857 	/* if this cookie uses the copy buffer */
3858 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3859 		/*
3860 		 * NOTE: we know that since this cookie uses the copy buffer, it
3861 		 * is <= MMU_PAGESIZE.
3862 		 */
3863 
3864 		/*
3865 		 * get the offset into the page. For the 64-bit kernel, get the
3866 		 * pfn which we'll use with seg kpm.
3867 		 */
3868 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3869 #if defined(__amd64)
3870 		/* mfn_to_pfn() is a NOP on i86pc */
3871 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3872 #endif /* __amd64 */
3873 
3874 		/* figure out if the copybuf size is a power of 2 */
3875 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3876 			copybuf_sz_power_2 = B_FALSE;
3877 		} else {
3878 			copybuf_sz_power_2 = B_TRUE;
3879 		}
3880 
3881 		/* This page uses the copy buffer */
3882 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3883 
3884 		/*
3885 		 * save the copy buffer KVA that we'll use with this page.
3886 		 * if we still fit within the copybuf, it's a simple add.
3887 		 * otherwise, we need to wrap over using & or % accordingly.
3888 		 */
3889 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3890 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3891 			    *copybuf_used;
3892 		} else {
3893 			if (copybuf_sz_power_2) {
3894 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3895 				    (uintptr_t)dma->dp_cbaddr +
3896 				    (*copybuf_used &
3897 				    (dma->dp_copybuf_size - 1)));
3898 			} else {
3899 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3900 				    (uintptr_t)dma->dp_cbaddr +
3901 				    (*copybuf_used % dma->dp_copybuf_size));
3902 			}
3903 		}
3904 
3905 		/*
3906 		 * over write the cookie physical address with the address of
3907 		 * the physical address of the copy buffer page that we will
3908 		 * use.
3909 		 */
3910 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3911 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3912 
3913 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
3914 
3915 		/* if we have a kernel VA, it's easy, just save that address */
3916 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3917 		    (sinfo->si_asp == &kas)) {
3918 			/*
3919 			 * save away the page aligned virtual address of the
3920 			 * driver buffer. Offsets are handled in the sync code.
3921 			 */
3922 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3923 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3924 			    & MMU_PAGEMASK);
3925 #if !defined(__amd64)
3926 			/*
3927 			 * we didn't need to, and will never need to map this
3928 			 * page.
3929 			 */
3930 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3931 #endif
3932 
3933 		/* we don't have a kernel VA. We need one for the bcopy. */
3934 		} else {
3935 #if defined(__amd64)
3936 			/*
3937 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3938 			 * get a Kernel VA for the corresponding pfn.
3939 			 */
3940 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3941 #else
3942 			/*
3943 			 * for the 32-bit kernel, this is a pain. First we'll
3944 			 * save away the page_t or user VA for this page. This
3945 			 * is needed in rootnex_dma_win() when we switch to a
3946 			 * new window which requires us to re-map the copy
3947 			 * buffer.
3948 			 */
3949 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3950 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3951 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3952 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3953 			} else if (pplist != NULL) {
3954 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3955 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3956 			} else {
3957 				dma->dp_pgmap[pidx].pm_pp = NULL;
3958 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3959 				    (((uintptr_t)
3960 				    dmar_object->dmao_obj.virt_obj.v_addr +
3961 				    cur_offset) & MMU_PAGEMASK);
3962 			}
3963 
3964 			/*
3965 			 * save away the page aligned virtual address which was
3966 			 * allocated from the kernel heap arena (taking into
3967 			 * account if we need more copy buffer than we alloced
3968 			 * and use multiple windows to handle this, i.e. &,%).
3969 			 * NOTE: there isn't and physical memory backing up this
3970 			 * virtual address space currently.
3971 			 */
3972 			if ((*copybuf_used + MMU_PAGESIZE) <=
3973 			    dma->dp_copybuf_size) {
3974 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3975 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3976 				    MMU_PAGEMASK);
3977 			} else {
3978 				if (copybuf_sz_power_2) {
3979 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3980 					    (((uintptr_t)dma->dp_kva +
3981 					    (*copybuf_used &
3982 					    (dma->dp_copybuf_size - 1))) &
3983 					    MMU_PAGEMASK);
3984 				} else {
3985 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3986 					    (((uintptr_t)dma->dp_kva +
3987 					    (*copybuf_used %
3988 					    dma->dp_copybuf_size)) &
3989 					    MMU_PAGEMASK);
3990 				}
3991 			}
3992 
3993 			/*
3994 			 * if we haven't used up the available copy buffer yet,
3995 			 * map the kva to the physical page.
3996 			 */
3997 			if (!dma->dp_cb_remaping && ((*copybuf_used +
3998 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
3999 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
4000 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
4001 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
4002 					    dma->dp_pgmap[pidx].pm_kaddr);
4003 				} else {
4004 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
4005 					    sinfo->si_asp,
4006 					    dma->dp_pgmap[pidx].pm_kaddr);
4007 				}
4008 
4009 			/*
4010 			 * we've used up the available copy buffer, this page
4011 			 * will have to be mapped during rootnex_dma_win() when
4012 			 * we switch to a new window which requires a re-map
4013 			 * the copy buffer. (32-bit kernel only)
4014 			 */
4015 			} else {
4016 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4017 			}
4018 #endif
4019 			/* go to the next page_t */
4020 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4021 				*cur_pp = (*cur_pp)->p_next;
4022 			}
4023 		}
4024 
4025 		/* add to the copy buffer count */
4026 		*copybuf_used += MMU_PAGESIZE;
4027 
4028 	/*
4029 	 * This cookie doesn't use the copy buffer. Walk through the pages this
4030 	 * cookie occupies to reflect this.
4031 	 */
4032 	} else {
4033 		/*
4034 		 * figure out how many pages the cookie occupies. We need to
4035 		 * use the original page offset of the buffer and the cookies
4036 		 * offset in the buffer to do this.
4037 		 */
4038 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
4039 		pcnt = mmu_btopr(cookie->dmac_size + poff);
4040 
4041 		while (pcnt > 0) {
4042 #if !defined(__amd64)
4043 			/*
4044 			 * the 32-bit kernel doesn't have seg kpm, so we need
4045 			 * to map in the driver buffer (if it didn't come down
4046 			 * with a kernel VA) on the fly. Since this page doesn't
4047 			 * use the copy buffer, it's not, or will it ever, have
4048 			 * to be mapped in.
4049 			 */
4050 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4051 #endif
4052 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
4053 
4054 			/*
4055 			 * we need to update pidx and cur_pp or we'll loose
4056 			 * track of where we are.
4057 			 */
4058 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4059 				*cur_pp = (*cur_pp)->p_next;
4060 			}
4061 			pidx++;
4062 			pcnt--;
4063 		}
4064 	}
4065 }
4066 
4067 
4068 /*
4069  * rootnex_sgllen_window_boundary()
4070  *    Called in the bind slow path when the next cookie causes us to exceed (in
4071  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
4072  *    length supported by the DMA H/W.
4073  */
4074 static int
4075 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4076     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
4077     off_t cur_offset)
4078 {
4079 	off_t new_offset;
4080 	size_t trim_sz;
4081 	off_t coffset;
4082 
4083 
4084 	/*
4085 	 * if we know we'll never have to trim, it's pretty easy. Just move to
4086 	 * the next window and init it. We're done.
4087 	 */
4088 	if (!dma->dp_trim_required) {
4089 		(*windowp)++;
4090 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4091 		(*windowp)->wd_cookie_cnt++;
4092 		(*windowp)->wd_size = cookie->dmac_size;
4093 		return (DDI_SUCCESS);
4094 	}
4095 
4096 	/* figure out how much we need to trim from the window */
4097 	ASSERT(attr->dma_attr_granular != 0);
4098 	if (dma->dp_granularity_power_2) {
4099 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
4100 	} else {
4101 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
4102 	}
4103 
4104 	/* The window's a whole multiple of granularity. We're done */
4105 	if (trim_sz == 0) {
4106 		(*windowp)++;
4107 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4108 		(*windowp)->wd_cookie_cnt++;
4109 		(*windowp)->wd_size = cookie->dmac_size;
4110 		return (DDI_SUCCESS);
4111 	}
4112 
4113 	/*
4114 	 * The window's not a whole multiple of granularity, since we know this
4115 	 * is due to the sgllen, we need to go back to the last cookie and trim
4116 	 * that one, add the left over part of the old cookie into the new
4117 	 * window, and then add in the new cookie into the new window.
4118 	 */
4119 
4120 	/*
4121 	 * make sure the driver isn't making us do something bad... Trimming and
4122 	 * sgllen == 1 don't go together.
4123 	 */
4124 	if (attr->dma_attr_sgllen == 1) {
4125 		return (DDI_DMA_NOMAPPING);
4126 	}
4127 
4128 	/*
4129 	 * first, setup the current window to account for the trim. Need to go
4130 	 * back to the last cookie for this.
4131 	 */
4132 	cookie--;
4133 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4134 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4135 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4136 	ASSERT(cookie->dmac_size > trim_sz);
4137 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4138 	(*windowp)->wd_size -= trim_sz;
4139 
4140 	/* save the buffer offsets for the next window */
4141 	coffset = cookie->dmac_size - trim_sz;
4142 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4143 
4144 	/*
4145 	 * set this now in case this is the first window. all other cases are
4146 	 * set in dma_win()
4147 	 */
4148 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4149 
4150 	/*
4151 	 * initialize the next window using what's left over in the previous
4152 	 * cookie.
4153 	 */
4154 	(*windowp)++;
4155 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4156 	(*windowp)->wd_cookie_cnt++;
4157 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4158 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4159 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4160 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4161 		(*windowp)->wd_dosync = B_TRUE;
4162 	}
4163 
4164 	/*
4165 	 * now go back to the current cookie and add it to the new window. set
4166 	 * the new window size to the what was left over from the previous
4167 	 * cookie and what's in the current cookie.
4168 	 */
4169 	cookie++;
4170 	(*windowp)->wd_cookie_cnt++;
4171 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4172 
4173 	/*
4174 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
4175 	 * a max size of maxxfer). Handle that case.
4176 	 */
4177 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
4178 		/*
4179 		 * maxxfer is already a whole multiple of granularity, and this
4180 		 * trim will be <= the previous trim (since a cookie can't be
4181 		 * larger than maxxfer). Make things simple here.
4182 		 */
4183 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
4184 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4185 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4186 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4187 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4188 		(*windowp)->wd_size -= trim_sz;
4189 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
4190 
4191 		/* save the buffer offsets for the next window */
4192 		coffset = cookie->dmac_size - trim_sz;
4193 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4194 
4195 		/* setup the next window */
4196 		(*windowp)++;
4197 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4198 		(*windowp)->wd_cookie_cnt++;
4199 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4200 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4201 		    coffset;
4202 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4203 	}
4204 
4205 	return (DDI_SUCCESS);
4206 }
4207 
4208 
4209 /*
4210  * rootnex_copybuf_window_boundary()
4211  *    Called in bind slowpath when we get to a window boundary because we used
4212  *    up all the copy buffer that we have.
4213  */
4214 static int
4215 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4216     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
4217     size_t *copybuf_used)
4218 {
4219 	rootnex_sglinfo_t *sinfo;
4220 	off_t new_offset;
4221 	size_t trim_sz;
4222 	paddr_t paddr;
4223 	off_t coffset;
4224 	uint_t pidx;
4225 	off_t poff;
4226 
4227 
4228 	sinfo = &dma->dp_sglinfo;
4229 
4230 	/*
4231 	 * the copy buffer should be a whole multiple of page size. We know that
4232 	 * this cookie is <= MMU_PAGESIZE.
4233 	 */
4234 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
4235 
4236 	/*
4237 	 * from now on, all new windows in this bind need to be re-mapped during
4238 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
4239 	 * space...
4240 	 */
4241 #if !defined(__amd64)
4242 	dma->dp_cb_remaping = B_TRUE;
4243 #endif
4244 
4245 	/* reset copybuf used */
4246 	*copybuf_used = 0;
4247 
4248 	/*
4249 	 * if we don't have to trim (since granularity is set to 1), go to the
4250 	 * next window and add the current cookie to it. We know the current
4251 	 * cookie uses the copy buffer since we're in this code path.
4252 	 */
4253 	if (!dma->dp_trim_required) {
4254 		(*windowp)++;
4255 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4256 
4257 		/* Add this cookie to the new window */
4258 		(*windowp)->wd_cookie_cnt++;
4259 		(*windowp)->wd_size += cookie->dmac_size;
4260 		*copybuf_used += MMU_PAGESIZE;
4261 		return (DDI_SUCCESS);
4262 	}
4263 
4264 	/*
4265 	 * *** may need to trim, figure it out.
4266 	 */
4267 
4268 	/* figure out how much we need to trim from the window */
4269 	if (dma->dp_granularity_power_2) {
4270 		trim_sz = (*windowp)->wd_size &
4271 		    (hp->dmai_attr.dma_attr_granular - 1);
4272 	} else {
4273 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
4274 	}
4275 
4276 	/*
4277 	 * if the window's a whole multiple of granularity, go to the next
4278 	 * window, init it, then add in the current cookie. We know the current
4279 	 * cookie uses the copy buffer since we're in this code path.
4280 	 */
4281 	if (trim_sz == 0) {
4282 		(*windowp)++;
4283 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4284 
4285 		/* Add this cookie to the new window */
4286 		(*windowp)->wd_cookie_cnt++;
4287 		(*windowp)->wd_size += cookie->dmac_size;
4288 		*copybuf_used += MMU_PAGESIZE;
4289 		return (DDI_SUCCESS);
4290 	}
4291 
4292 	/*
4293 	 * *** We figured it out, we definitly need to trim
4294 	 */
4295 
4296 	/*
4297 	 * make sure the driver isn't making us do something bad...
4298 	 * Trimming and sgllen == 1 don't go together.
4299 	 */
4300 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4301 		return (DDI_DMA_NOMAPPING);
4302 	}
4303 
4304 	/*
4305 	 * first, setup the current window to account for the trim. Need to go
4306 	 * back to the last cookie for this. Some of the last cookie will be in
4307 	 * the current window, and some of the last cookie will be in the new
4308 	 * window. All of the current cookie will be in the new window.
4309 	 */
4310 	cookie--;
4311 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4312 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4313 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4314 	ASSERT(cookie->dmac_size > trim_sz);
4315 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4316 	(*windowp)->wd_size -= trim_sz;
4317 
4318 	/*
4319 	 * we're trimming the last cookie (not the current cookie). So that
4320 	 * last cookie may have or may not have been using the copy buffer (
4321 	 * we know the cookie passed in uses the copy buffer since we're in
4322 	 * this code path).
4323 	 *
4324 	 * If the last cookie doesn't use the copy buffer, nothing special to
4325 	 * do. However, if it does uses the copy buffer, it will be both the
4326 	 * last page in the current window and the first page in the next
4327 	 * window. Since we are reusing the copy buffer (and KVA space on the
4328 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4329 	 * current window, and the start of the copy buffer in the next window.
4330 	 * Track that info... The cookie physical address was already set to
4331 	 * the copy buffer physical address in setup_cookie..
4332 	 */
4333 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4334 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4335 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4336 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4337 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4338 		(*windowp)->wd_trim.tr_last_cbaddr =
4339 		    dma->dp_pgmap[pidx].pm_cbaddr;
4340 #if !defined(__amd64)
4341 		(*windowp)->wd_trim.tr_last_kaddr =
4342 		    dma->dp_pgmap[pidx].pm_kaddr;
4343 #endif
4344 	}
4345 
4346 	/* save the buffer offsets for the next window */
4347 	coffset = cookie->dmac_size - trim_sz;
4348 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4349 
4350 	/*
4351 	 * set this now in case this is the first window. all other cases are
4352 	 * set in dma_win()
4353 	 */
4354 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4355 
4356 	/*
4357 	 * initialize the next window using what's left over in the previous
4358 	 * cookie.
4359 	 */
4360 	(*windowp)++;
4361 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4362 	(*windowp)->wd_cookie_cnt++;
4363 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4364 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4365 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4366 
4367 	/*
4368 	 * again, we're tracking if the last cookie uses the copy buffer.
4369 	 * read the comment above for more info on why we need to track
4370 	 * additional state.
4371 	 *
4372 	 * For the first cookie in the new window, we need reset the physical
4373 	 * address to DMA into to the start of the copy buffer plus any
4374 	 * initial page offset which may be present.
4375 	 */
4376 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4377 		(*windowp)->wd_dosync = B_TRUE;
4378 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4379 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4380 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4381 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4382 
4383 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4384 		    poff;
4385 		(*windowp)->wd_trim.tr_first_paddr =
4386 		    ROOTNEX_PADDR_TO_RBASE(paddr);
4387 
4388 #if !defined(__amd64)
4389 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4390 #endif
4391 		/* account for the cookie copybuf usage in the new window */
4392 		*copybuf_used += MMU_PAGESIZE;
4393 
4394 		/*
4395 		 * every piece of code has to have a hack, and here is this
4396 		 * ones :-)
4397 		 *
4398 		 * There is a complex interaction between setup_cookie and the
4399 		 * copybuf window boundary. The complexity had to be in either
4400 		 * the maxxfer window, or the copybuf window, and I chose the
4401 		 * copybuf code.
4402 		 *
4403 		 * So in this code path, we have taken the last cookie,
4404 		 * virtually broken it in half due to the trim, and it happens
4405 		 * to use the copybuf which further complicates life. At the
4406 		 * same time, we have already setup the current cookie, which
4407 		 * is now wrong. More background info: the current cookie uses
4408 		 * the copybuf, so it is only a page long max. So we need to
4409 		 * fix the current cookies copy buffer address, physical
4410 		 * address, and kva for the 32-bit kernel. We due this by
4411 		 * bumping them by page size (of course, we can't due this on
4412 		 * the physical address since the copy buffer may not be
4413 		 * physically contiguous).
4414 		 */
4415 		cookie++;
4416 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4417 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4418 
4419 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4420 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4421 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
4422 
4423 #if !defined(__amd64)
4424 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4425 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4426 #endif
4427 	} else {
4428 		/* go back to the current cookie */
4429 		cookie++;
4430 	}
4431 
4432 	/*
4433 	 * add the current cookie to the new window. set the new window size to
4434 	 * the what was left over from the previous cookie and what's in the
4435 	 * current cookie.
4436 	 */
4437 	(*windowp)->wd_cookie_cnt++;
4438 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4439 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4440 
4441 	/*
4442 	 * we know that the cookie passed in always uses the copy buffer. We
4443 	 * wouldn't be here if it didn't.
4444 	 */
4445 	*copybuf_used += MMU_PAGESIZE;
4446 
4447 	return (DDI_SUCCESS);
4448 }
4449 
4450 
4451 /*
4452  * rootnex_maxxfer_window_boundary()
4453  *    Called in bind slowpath when we get to a window boundary because we will
4454  *    go over maxxfer.
4455  */
4456 static int
4457 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4458     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4459 {
4460 	size_t dmac_size;
4461 	off_t new_offset;
4462 	size_t trim_sz;
4463 	off_t coffset;
4464 
4465 
4466 	/*
4467 	 * calculate how much we have to trim off of the current cookie to equal
4468 	 * maxxfer. We don't have to account for granularity here since our
4469 	 * maxxfer already takes that into account.
4470 	 */
4471 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4472 	ASSERT(trim_sz <= cookie->dmac_size);
4473 	ASSERT(trim_sz <= dma->dp_maxxfer);
4474 
4475 	/* save cookie size since we need it later and we might change it */
4476 	dmac_size = cookie->dmac_size;
4477 
4478 	/*
4479 	 * if we're not trimming the entire cookie, setup the current window to
4480 	 * account for the trim.
4481 	 */
4482 	if (trim_sz < cookie->dmac_size) {
4483 		(*windowp)->wd_cookie_cnt++;
4484 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4485 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4486 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4487 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4488 		(*windowp)->wd_size = dma->dp_maxxfer;
4489 
4490 		/*
4491 		 * set the adjusted cookie size now in case this is the first
4492 		 * window. All other windows are taken care of in get win
4493 		 */
4494 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4495 	}
4496 
4497 	/*
4498 	 * coffset is the current offset within the cookie, new_offset is the
4499 	 * current offset with the entire buffer.
4500 	 */
4501 	coffset = dmac_size - trim_sz;
4502 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4503 
4504 	/* initialize the next window */
4505 	(*windowp)++;
4506 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4507 	(*windowp)->wd_cookie_cnt++;
4508 	(*windowp)->wd_size = trim_sz;
4509 	if (trim_sz < dmac_size) {
4510 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4511 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4512 		    coffset;
4513 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4514 	}
4515 
4516 	return (DDI_SUCCESS);
4517 }
4518 
4519 
4520 /*ARGSUSED*/
4521 static int
4522 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4523     off_t off, size_t len, uint_t cache_flags)
4524 {
4525 	rootnex_sglinfo_t *sinfo;
4526 	rootnex_pgmap_t *cbpage;
4527 	rootnex_window_t *win;
4528 	ddi_dma_impl_t *hp;
4529 	rootnex_dma_t *dma;
4530 	caddr_t fromaddr;
4531 	caddr_t toaddr;
4532 	uint_t psize;
4533 	off_t offset;
4534 	uint_t pidx;
4535 	size_t size;
4536 	off_t poff;
4537 	int e;
4538 
4539 
4540 	hp = (ddi_dma_impl_t *)handle;
4541 	dma = (rootnex_dma_t *)hp->dmai_private;
4542 	sinfo = &dma->dp_sglinfo;
4543 
4544 	/*
4545 	 * if we don't have any windows, we don't need to sync. A copybuf
4546 	 * will cause us to have at least one window.
4547 	 */
4548 	if (dma->dp_window == NULL) {
4549 		return (DDI_SUCCESS);
4550 	}
4551 
4552 	/* This window may not need to be sync'd */
4553 	win = &dma->dp_window[dma->dp_current_win];
4554 	if (!win->wd_dosync) {
4555 		return (DDI_SUCCESS);
4556 	}
4557 
4558 	/* handle off and len special cases */
4559 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4560 		offset = win->wd_offset;
4561 	} else {
4562 		offset = off;
4563 	}
4564 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4565 		size = win->wd_size;
4566 	} else {
4567 		size = len;
4568 	}
4569 
4570 	/* check the sync args to make sure they make a little sense */
4571 	if (rootnex_sync_check_parms) {
4572 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4573 		    cache_flags);
4574 		if (e != DDI_SUCCESS) {
4575 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4576 			return (DDI_FAILURE);
4577 		}
4578 	}
4579 
4580 	/*
4581 	 * special case the first page to handle the offset into the page. The
4582 	 * offset to the current page for our buffer is the offset into the
4583 	 * first page of the buffer plus our current offset into the buffer
4584 	 * itself, masked of course.
4585 	 */
4586 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4587 	psize = MIN((MMU_PAGESIZE - poff), size);
4588 
4589 	/* go through all the pages that we want to sync */
4590 	while (size > 0) {
4591 		/*
4592 		 * Calculate the page index relative to the start of the buffer.
4593 		 * The index to the current page for our buffer is the offset
4594 		 * into the first page of the buffer plus our current offset
4595 		 * into the buffer itself, shifted of course...
4596 		 */
4597 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4598 		ASSERT(pidx < sinfo->si_max_pages);
4599 
4600 		/*
4601 		 * if this page uses the copy buffer, we need to sync it,
4602 		 * otherwise, go on to the next page.
4603 		 */
4604 		cbpage = &dma->dp_pgmap[pidx];
4605 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4606 		    (cbpage->pm_uses_copybuf == B_FALSE));
4607 		if (cbpage->pm_uses_copybuf) {
4608 			/* cbaddr and kaddr should be page aligned */
4609 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4610 			    MMU_PAGEOFFSET) == 0);
4611 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4612 			    MMU_PAGEOFFSET) == 0);
4613 
4614 			/*
4615 			 * if we're copying for the device, we are going to
4616 			 * copy from the drivers buffer and to the rootnex
4617 			 * allocated copy buffer.
4618 			 */
4619 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4620 				fromaddr = cbpage->pm_kaddr + poff;
4621 				toaddr = cbpage->pm_cbaddr + poff;
4622 				ROOTNEX_DPROBE2(rootnex__sync__dev,
4623 				    dev_info_t *, dma->dp_dip, size_t, psize);
4624 
4625 			/*
4626 			 * if we're copying for the cpu/kernel, we are going to
4627 			 * copy from the rootnex allocated copy buffer to the
4628 			 * drivers buffer.
4629 			 */
4630 			} else {
4631 				fromaddr = cbpage->pm_cbaddr + poff;
4632 				toaddr = cbpage->pm_kaddr + poff;
4633 				ROOTNEX_DPROBE2(rootnex__sync__cpu,
4634 				    dev_info_t *, dma->dp_dip, size_t, psize);
4635 			}
4636 
4637 			bcopy(fromaddr, toaddr, psize);
4638 		}
4639 
4640 		/*
4641 		 * decrement size until we're done, update our offset into the
4642 		 * buffer, and get the next page size.
4643 		 */
4644 		size -= psize;
4645 		offset += psize;
4646 		psize = MIN(MMU_PAGESIZE, size);
4647 
4648 		/* page offset is zero for the rest of this loop */
4649 		poff = 0;
4650 	}
4651 
4652 	return (DDI_SUCCESS);
4653 }
4654 
4655 /*
4656  * rootnex_dma_sync()
4657  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4658  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4659  *    is set, ddi_dma_sync() returns immediately passing back success.
4660  */
4661 /*ARGSUSED*/
4662 static int
4663 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4664     off_t off, size_t len, uint_t cache_flags)
4665 {
4666 #if defined(__amd64) && !defined(__xpv)
4667 	if (IOMMU_USED(rdip)) {
4668 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4669 		    cache_flags));
4670 	}
4671 #endif
4672 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4673 	    cache_flags));
4674 }
4675 
4676 /*
4677  * rootnex_valid_sync_parms()
4678  *    checks the parameters passed to sync to verify they are correct.
4679  */
4680 static int
4681 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4682     off_t offset, size_t size, uint_t cache_flags)
4683 {
4684 	off_t woffset;
4685 
4686 
4687 	/*
4688 	 * the first part of the test to make sure the offset passed in is
4689 	 * within the window.
4690 	 */
4691 	if (offset < win->wd_offset) {
4692 		return (DDI_FAILURE);
4693 	}
4694 
4695 	/*
4696 	 * second and last part of the test to make sure the offset and length
4697 	 * passed in is within the window.
4698 	 */
4699 	woffset = offset - win->wd_offset;
4700 	if ((woffset + size) > win->wd_size) {
4701 		return (DDI_FAILURE);
4702 	}
4703 
4704 	/*
4705 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4706 	 * be set too.
4707 	 */
4708 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4709 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4710 		return (DDI_SUCCESS);
4711 	}
4712 
4713 	/*
4714 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4715 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4716 	 */
4717 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4718 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4719 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4720 		return (DDI_SUCCESS);
4721 	}
4722 
4723 	return (DDI_FAILURE);
4724 }
4725 
4726 
4727 /*ARGSUSED*/
4728 static int
4729 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4730     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4731     uint_t *ccountp)
4732 {
4733 	rootnex_window_t *window;
4734 	rootnex_trim_t *trim;
4735 	ddi_dma_impl_t *hp;
4736 	rootnex_dma_t *dma;
4737 	ddi_dma_obj_t *dmao;
4738 #if !defined(__amd64)
4739 	rootnex_sglinfo_t *sinfo;
4740 	rootnex_pgmap_t *pmap;
4741 	uint_t pidx;
4742 	uint_t pcnt;
4743 	off_t poff;
4744 	int i;
4745 #endif
4746 
4747 
4748 	hp = (ddi_dma_impl_t *)handle;
4749 	dma = (rootnex_dma_t *)hp->dmai_private;
4750 #if !defined(__amd64)
4751 	sinfo = &dma->dp_sglinfo;
4752 #endif
4753 
4754 	/* If we try and get a window which doesn't exist, return failure */
4755 	if (win >= hp->dmai_nwin) {
4756 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4757 		return (DDI_FAILURE);
4758 	}
4759 
4760 	dmao = dma->dp_dvma_used ? &dma->dp_dvma : &dma->dp_dma;
4761 
4762 	/*
4763 	 * if we don't have any windows, and they're asking for the first
4764 	 * window, setup the cookie pointer to the first cookie in the bind.
4765 	 * setup our return values, then increment the cookie since we return
4766 	 * the first cookie on the stack.
4767 	 */
4768 	if (dma->dp_window == NULL) {
4769 		if (win != 0) {
4770 			ROOTNEX_DPROF_INC(
4771 			    &rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4772 			return (DDI_FAILURE);
4773 		}
4774 		hp->dmai_cookie = dma->dp_cookies;
4775 		*offp = 0;
4776 		*lenp = dmao->dmao_size;
4777 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4778 		*cookiep = hp->dmai_cookie[0];
4779 		hp->dmai_cookie++;
4780 		return (DDI_SUCCESS);
4781 	}
4782 
4783 	/* sync the old window before moving on to the new one */
4784 	window = &dma->dp_window[dma->dp_current_win];
4785 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4786 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4787 		    DDI_DMA_SYNC_FORCPU);
4788 	}
4789 
4790 #if !defined(__amd64)
4791 	/*
4792 	 * before we move to the next window, if we need to re-map, unmap all
4793 	 * the pages in this window.
4794 	 */
4795 	if (dma->dp_cb_remaping) {
4796 		/*
4797 		 * If we switch to this window again, we'll need to map in
4798 		 * on the fly next time.
4799 		 */
4800 		window->wd_remap_copybuf = B_TRUE;
4801 
4802 		/*
4803 		 * calculate the page index into the buffer where this window
4804 		 * starts, and the number of pages this window takes up.
4805 		 */
4806 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4807 		    MMU_PAGESHIFT;
4808 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4809 		    MMU_PAGEOFFSET;
4810 		pcnt = mmu_btopr(window->wd_size + poff);
4811 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4812 
4813 		/* unmap pages which are currently mapped in this window */
4814 		for (i = 0; i < pcnt; i++) {
4815 			if (dma->dp_pgmap[pidx].pm_mapped) {
4816 				hat_unload(kas.a_hat,
4817 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4818 				    HAT_UNLOAD);
4819 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4820 			}
4821 			pidx++;
4822 		}
4823 	}
4824 #endif
4825 
4826 	/*
4827 	 * Move to the new window.
4828 	 * NOTE: current_win must be set for sync to work right
4829 	 */
4830 	dma->dp_current_win = win;
4831 	window = &dma->dp_window[win];
4832 
4833 	/* if needed, adjust the first and/or last cookies for trim */
4834 	trim = &window->wd_trim;
4835 	if (trim->tr_trim_first) {
4836 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4837 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4838 #if !defined(__amd64)
4839 		window->wd_first_cookie->dmac_type =
4840 		    (window->wd_first_cookie->dmac_type &
4841 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4842 #endif
4843 		if (trim->tr_first_copybuf_win) {
4844 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4845 			    trim->tr_first_cbaddr;
4846 #if !defined(__amd64)
4847 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4848 			    trim->tr_first_kaddr;
4849 #endif
4850 		}
4851 	}
4852 	if (trim->tr_trim_last) {
4853 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4854 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4855 		if (trim->tr_last_copybuf_win) {
4856 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4857 			    trim->tr_last_cbaddr;
4858 #if !defined(__amd64)
4859 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4860 			    trim->tr_last_kaddr;
4861 #endif
4862 		}
4863 	}
4864 
4865 	/*
4866 	 * setup the cookie pointer to the first cookie in the window. setup
4867 	 * our return values, then increment the cookie since we return the
4868 	 * first cookie on the stack.
4869 	 */
4870 	hp->dmai_cookie = window->wd_first_cookie;
4871 	*offp = window->wd_offset;
4872 	*lenp = window->wd_size;
4873 	*ccountp = window->wd_cookie_cnt;
4874 	*cookiep = hp->dmai_cookie[0];
4875 	hp->dmai_cookie++;
4876 
4877 #if !defined(__amd64)
4878 	/* re-map copybuf if required for this window */
4879 	if (dma->dp_cb_remaping) {
4880 		/*
4881 		 * calculate the page index into the buffer where this
4882 		 * window starts.
4883 		 */
4884 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4885 		    MMU_PAGESHIFT;
4886 		ASSERT(pidx < sinfo->si_max_pages);
4887 
4888 		/*
4889 		 * the first page can get unmapped if it's shared with the
4890 		 * previous window. Even if the rest of this window is already
4891 		 * mapped in, we need to still check this one.
4892 		 */
4893 		pmap = &dma->dp_pgmap[pidx];
4894 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4895 			if (pmap->pm_pp != NULL) {
4896 				pmap->pm_mapped = B_TRUE;
4897 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4898 			} else if (pmap->pm_vaddr != NULL) {
4899 				pmap->pm_mapped = B_TRUE;
4900 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4901 				    pmap->pm_kaddr);
4902 			}
4903 		}
4904 		pidx++;
4905 
4906 		/* map in the rest of the pages if required */
4907 		if (window->wd_remap_copybuf) {
4908 			window->wd_remap_copybuf = B_FALSE;
4909 
4910 			/* figure out many pages this window takes up */
4911 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4912 			    MMU_PAGEOFFSET;
4913 			pcnt = mmu_btopr(window->wd_size + poff);
4914 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4915 
4916 			/* map pages which require it */
4917 			for (i = 1; i < pcnt; i++) {
4918 				pmap = &dma->dp_pgmap[pidx];
4919 				if (pmap->pm_uses_copybuf) {
4920 					ASSERT(pmap->pm_mapped == B_FALSE);
4921 					if (pmap->pm_pp != NULL) {
4922 						pmap->pm_mapped = B_TRUE;
4923 						i86_pp_map(pmap->pm_pp,
4924 						    pmap->pm_kaddr);
4925 					} else if (pmap->pm_vaddr != NULL) {
4926 						pmap->pm_mapped = B_TRUE;
4927 						i86_va_map(pmap->pm_vaddr,
4928 						    sinfo->si_asp,
4929 						    pmap->pm_kaddr);
4930 					}
4931 				}
4932 				pidx++;
4933 			}
4934 		}
4935 	}
4936 #endif
4937 
4938 	/* if the new window uses the copy buffer, sync it for the device */
4939 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4940 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4941 		    DDI_DMA_SYNC_FORDEV);
4942 	}
4943 
4944 	return (DDI_SUCCESS);
4945 }
4946 
4947 /*
4948  * rootnex_dma_win()
4949  *    called from ddi_dma_getwin()
4950  */
4951 /*ARGSUSED*/
4952 static int
4953 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4954     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4955     uint_t *ccountp)
4956 {
4957 #if defined(__amd64) && !defined(__xpv)
4958 	if (IOMMU_USED(rdip)) {
4959 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4960 		    cookiep, ccountp));
4961 	}
4962 #endif
4963 
4964 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4965 	    cookiep, ccountp));
4966 }
4967 
4968 #if defined(__amd64) && !defined(__xpv)
4969 /*ARGSUSED*/
4970 static int
4971 rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
4972     ddi_dma_handle_t handle, void *v)
4973 {
4974 	ddi_dma_impl_t *hp;
4975 	rootnex_dma_t *dma;
4976 
4977 	hp = (ddi_dma_impl_t *)handle;
4978 	dma = (rootnex_dma_t *)hp->dmai_private;
4979 	dma->dp_iommu_private = v;
4980 
4981 	return (DDI_SUCCESS);
4982 }
4983 
4984 /*ARGSUSED*/
4985 static void *
4986 rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
4987     ddi_dma_handle_t handle)
4988 {
4989 	ddi_dma_impl_t *hp;
4990 	rootnex_dma_t *dma;
4991 
4992 	hp = (ddi_dma_impl_t *)handle;
4993 	dma = (rootnex_dma_t *)hp->dmai_private;
4994 
4995 	return (dma->dp_iommu_private);
4996 }
4997 #endif
4998 
4999 /*
5000  * ************************
5001  *  obsoleted dma routines
5002  * ************************
5003  */
5004 
5005 /*
5006  * rootnex_dma_map()
5007  *    called from ddi_dma_setup()
5008  * NO IOMMU in 32 bit mode. The below routines doesn't work in 64 bit mode.
5009  */
5010 /* ARGSUSED */
5011 static int
5012 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
5013     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
5014 {
5015 #if defined(__amd64)
5016 	/*
5017 	 * this interface is not supported in 64-bit x86 kernel. See comment in
5018 	 * rootnex_dma_mctl()
5019 	 */
5020 	return (DDI_DMA_NORESOURCES);
5021 
5022 #else /* 32-bit x86 kernel */
5023 	ddi_dma_handle_t *lhandlep;
5024 	ddi_dma_handle_t lhandle;
5025 	ddi_dma_cookie_t cookie;
5026 	ddi_dma_attr_t dma_attr;
5027 	ddi_dma_lim_t *dma_lim;
5028 	uint_t ccnt;
5029 	int e;
5030 
5031 
5032 	/*
5033 	 * if the driver is just testing to see if it's possible to do the bind,
5034 	 * we'll use local state. Otherwise, use the handle pointer passed in.
5035 	 */
5036 	if (handlep == NULL) {
5037 		lhandlep = &lhandle;
5038 	} else {
5039 		lhandlep = handlep;
5040 	}
5041 
5042 	/* convert the limit structure to a dma_attr one */
5043 	dma_lim = dmareq->dmar_limits;
5044 	dma_attr.dma_attr_version = DMA_ATTR_V0;
5045 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
5046 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
5047 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
5048 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
5049 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
5050 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
5051 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
5052 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
5053 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
5054 	dma_attr.dma_attr_align = MMU_PAGESIZE;
5055 	dma_attr.dma_attr_flags = 0;
5056 
5057 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
5058 	    dmareq->dmar_arg, lhandlep);
5059 	if (e != DDI_SUCCESS) {
5060 		return (e);
5061 	}
5062 
5063 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
5064 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
5065 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
5066 		return (e);
5067 	}
5068 
5069 	/*
5070 	 * if the driver is just testing to see if it's possible to do the bind,
5071 	 * free up the local state and return the result.
5072 	 */
5073 	if (handlep == NULL) {
5074 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
5075 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
5076 		if (e == DDI_DMA_MAPPED) {
5077 			return (DDI_DMA_MAPOK);
5078 		} else {
5079 			return (DDI_DMA_NOMAPPING);
5080 		}
5081 	}
5082 
5083 	return (e);
5084 #endif /* defined(__amd64) */
5085 }
5086 
5087 /*
5088  * rootnex_dma_mctl()
5089  *
5090  * No IOMMU in 32 bit mode. The below routine doesn't work in 64 bit mode.
5091  */
5092 /* ARGSUSED */
5093 static int
5094 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
5095     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
5096     uint_t cache_flags)
5097 {
5098 #if defined(__amd64)
5099 	/*
5100 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
5101 	 * common implementation in genunix, so they no longer have x86
5102 	 * specific functionality which called into dma_ctl.
5103 	 *
5104 	 * The rest of the obsoleted interfaces were never supported in the
5105 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
5106 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
5107 	 * implementation issues.
5108 	 *
5109 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
5110 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
5111 	 * reflect that now too...
5112 	 *
5113 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
5114 	 * not going to put this functionality into the 64-bit x86 kernel now.
5115 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
5116 	 * that in a future release.
5117 	 */
5118 	return (DDI_FAILURE);
5119 
5120 #else /* 32-bit x86 kernel */
5121 	ddi_dma_cookie_t lcookie;
5122 	ddi_dma_cookie_t *cookie;
5123 	rootnex_window_t *window;
5124 	ddi_dma_impl_t *hp;
5125 	rootnex_dma_t *dma;
5126 	uint_t nwin;
5127 	uint_t ccnt;
5128 	size_t len;
5129 	off_t off;
5130 	int e;
5131 
5132 
5133 	/*
5134 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
5135 	 * hacky since were optimizing for the current interfaces and so we can
5136 	 * cleanup the mess in genunix. Hopefully we will remove the this
5137 	 * obsoleted routines someday soon.
5138 	 */
5139 
5140 	switch (request) {
5141 
5142 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
5143 		hp = (ddi_dma_impl_t *)handle;
5144 		cookie = (ddi_dma_cookie_t *)objpp;
5145 
5146 		/*
5147 		 * convert segment to cookie. We don't distinguish between the
5148 		 * two :-)
5149 		 */
5150 		*cookie = *hp->dmai_cookie;
5151 		*lenp = cookie->dmac_size;
5152 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
5153 		return (DDI_SUCCESS);
5154 
5155 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
5156 		hp = (ddi_dma_impl_t *)handle;
5157 		dma = (rootnex_dma_t *)hp->dmai_private;
5158 
5159 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
5160 			return (DDI_DMA_STALE);
5161 		}
5162 
5163 		/* handle the case where we don't have any windows */
5164 		if (dma->dp_window == NULL) {
5165 			/*
5166 			 * if seg == NULL, and we don't have any windows,
5167 			 * return the first cookie in the sgl.
5168 			 */
5169 			if (*lenp == NULL) {
5170 				dma->dp_current_cookie = 0;
5171 				hp->dmai_cookie = dma->dp_cookies;
5172 				*objpp = (caddr_t)handle;
5173 				return (DDI_SUCCESS);
5174 
5175 			/* if we have more cookies, go to the next cookie */
5176 			} else {
5177 				if ((dma->dp_current_cookie + 1) >=
5178 				    dma->dp_sglinfo.si_sgl_size) {
5179 					return (DDI_DMA_DONE);
5180 				}
5181 				dma->dp_current_cookie++;
5182 				hp->dmai_cookie++;
5183 				return (DDI_SUCCESS);
5184 			}
5185 		}
5186 
5187 		/* We have one or more windows */
5188 		window = &dma->dp_window[dma->dp_current_win];
5189 
5190 		/*
5191 		 * if seg == NULL, return the first cookie in the current
5192 		 * window
5193 		 */
5194 		if (*lenp == NULL) {
5195 			dma->dp_current_cookie = 0;
5196 			hp->dmai_cookie = window->wd_first_cookie;
5197 
5198 		/*
5199 		 * go to the next cookie in the window then see if we done with
5200 		 * this window.
5201 		 */
5202 		} else {
5203 			if ((dma->dp_current_cookie + 1) >=
5204 			    window->wd_cookie_cnt) {
5205 				return (DDI_DMA_DONE);
5206 			}
5207 			dma->dp_current_cookie++;
5208 			hp->dmai_cookie++;
5209 		}
5210 		*objpp = (caddr_t)handle;
5211 		return (DDI_SUCCESS);
5212 
5213 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
5214 		hp = (ddi_dma_impl_t *)handle;
5215 		dma = (rootnex_dma_t *)hp->dmai_private;
5216 
5217 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
5218 			return (DDI_DMA_STALE);
5219 		}
5220 
5221 		/* if win == NULL, return the first window in the bind */
5222 		if (*offp == NULL) {
5223 			nwin = 0;
5224 
5225 		/*
5226 		 * else, go to the next window then see if we're done with all
5227 		 * the windows.
5228 		 */
5229 		} else {
5230 			nwin = dma->dp_current_win + 1;
5231 			if (nwin >= hp->dmai_nwin) {
5232 				return (DDI_DMA_DONE);
5233 			}
5234 		}
5235 
5236 		/* switch to the next window */
5237 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
5238 		    &lcookie, &ccnt);
5239 		ASSERT(e == DDI_SUCCESS);
5240 		if (e != DDI_SUCCESS) {
5241 			return (DDI_DMA_STALE);
5242 		}
5243 
5244 		/* reset the cookie back to the first cookie in the window */
5245 		if (dma->dp_window != NULL) {
5246 			window = &dma->dp_window[dma->dp_current_win];
5247 			hp->dmai_cookie = window->wd_first_cookie;
5248 		} else {
5249 			hp->dmai_cookie = dma->dp_cookies;
5250 		}
5251 
5252 		*objpp = (caddr_t)handle;
5253 		return (DDI_SUCCESS);
5254 
5255 	case DDI_DMA_FREE: /* ddi_dma_free() */
5256 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
5257 		(void) rootnex_dma_freehdl(dip, rdip, handle);
5258 		if (rootnex_state->r_dvma_call_list_id) {
5259 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
5260 		}
5261 		return (DDI_SUCCESS);
5262 
5263 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
5264 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
5265 		/* should never get here, handled in genunix */
5266 		ASSERT(0);
5267 		return (DDI_FAILURE);
5268 
5269 	case DDI_DMA_KVADDR:
5270 	case DDI_DMA_GETERR:
5271 	case DDI_DMA_COFF:
5272 		return (DDI_FAILURE);
5273 	}
5274 
5275 	return (DDI_FAILURE);
5276 #endif /* defined(__amd64) */
5277 }
5278 
5279 /*
5280  * *********
5281  *  FMA Code
5282  * *********
5283  */
5284 
5285 /*
5286  * rootnex_fm_init()
5287  *    FMA init busop
5288  */
5289 /* ARGSUSED */
5290 static int
5291 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
5292     ddi_iblock_cookie_t *ibc)
5293 {
5294 	*ibc = rootnex_state->r_err_ibc;
5295 
5296 	return (ddi_system_fmcap);
5297 }
5298 
5299 /*
5300  * rootnex_dma_check()
5301  *    Function called after a dma fault occurred to find out whether the
5302  *    fault address is associated with a driver that is able to handle faults
5303  *    and recover from faults.
5304  */
5305 /* ARGSUSED */
5306 static int
5307 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
5308     const void *not_used)
5309 {
5310 	rootnex_window_t *window;
5311 	uint64_t start_addr;
5312 	uint64_t fault_addr;
5313 	ddi_dma_impl_t *hp;
5314 	rootnex_dma_t *dma;
5315 	uint64_t end_addr;
5316 	size_t csize;
5317 	int i;
5318 	int j;
5319 
5320 
5321 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5322 	hp = (ddi_dma_impl_t *)handle;
5323 	ASSERT(hp);
5324 
5325 	dma = (rootnex_dma_t *)hp->dmai_private;
5326 
5327 	/* Get the address that we need to search for */
5328 	fault_addr = *(uint64_t *)addr;
5329 
5330 	/*
5331 	 * if we don't have any windows, we can just walk through all the
5332 	 * cookies.
5333 	 */
5334 	if (dma->dp_window == NULL) {
5335 		/* for each cookie */
5336 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5337 			/*
5338 			 * if the faulted address is within the physical address
5339 			 * range of the cookie, return DDI_FM_NONFATAL.
5340 			 */
5341 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5342 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5343 			    dma->dp_cookies[i].dmac_size))) {
5344 				return (DDI_FM_NONFATAL);
5345 			}
5346 		}
5347 
5348 		/* fault_addr not within this DMA handle */
5349 		return (DDI_FM_UNKNOWN);
5350 	}
5351 
5352 	/* we have mutiple windows, walk through each window */
5353 	for (i = 0; i < hp->dmai_nwin; i++) {
5354 		window = &dma->dp_window[i];
5355 
5356 		/* Go through all the cookies in the window */
5357 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5358 
5359 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5360 			csize = window->wd_first_cookie[j].dmac_size;
5361 
5362 			/*
5363 			 * if we are trimming the first cookie in the window,
5364 			 * and this is the first cookie, adjust the start
5365 			 * address and size of the cookie to account for the
5366 			 * trim.
5367 			 */
5368 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5369 				start_addr = window->wd_trim.tr_first_paddr;
5370 				csize = window->wd_trim.tr_first_size;
5371 			}
5372 
5373 			/*
5374 			 * if we are trimming the last cookie in the window,
5375 			 * and this is the last cookie, adjust the start
5376 			 * address and size of the cookie to account for the
5377 			 * trim.
5378 			 */
5379 			if (window->wd_trim.tr_trim_last &&
5380 			    (j == (window->wd_cookie_cnt - 1))) {
5381 				start_addr = window->wd_trim.tr_last_paddr;
5382 				csize = window->wd_trim.tr_last_size;
5383 			}
5384 
5385 			end_addr = start_addr + csize;
5386 
5387 			/*
5388 			 * if the faulted address is within the physical
5389 			 * address of the cookie, return DDI_FM_NONFATAL.
5390 			 */
5391 			if ((fault_addr >= start_addr) &&
5392 			    (fault_addr <= end_addr)) {
5393 				return (DDI_FM_NONFATAL);
5394 			}
5395 		}
5396 	}
5397 
5398 	/* fault_addr not within this DMA handle */
5399 	return (DDI_FM_UNKNOWN);
5400 }
5401 
5402 /*ARGSUSED*/
5403 static int
5404 rootnex_quiesce(dev_info_t *dip)
5405 {
5406 #if defined(__amd64) && !defined(__xpv)
5407 	return (immu_quiesce());
5408 #else
5409 	return (DDI_SUCCESS);
5410 #endif
5411 }
5412 
5413 #if defined(__xpv)
5414 void
5415 immu_init(void)
5416 {
5417 	;
5418 }
5419 
5420 void
5421 immu_startup(void)
5422 {
5423 	;
5424 }
5425 /*ARGSUSED*/
5426 void
5427 immu_physmem_update(uint64_t addr, uint64_t size)
5428 {
5429 	;
5430 }
5431 #endif
5432